xref: /xnu-8796.101.5/osfmk/vm/vm_pageout.c (revision aca3beaa3dfbd42498b42c5e5ce20a938e6554e5)
1 /*
2  * Copyright (c) 2000-2020 Apple Inc. All rights reserved.
3  *
4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5  *
6  * This file contains Original Code and/or Modifications of Original Code
7  * as defined in and that are subject to the Apple Public Source License
8  * Version 2.0 (the 'License'). You may not use this file except in
9  * compliance with the License. The rights granted to you under the License
10  * may not be used to create, or enable the creation or redistribution of,
11  * unlawful or unlicensed copies of an Apple operating system, or to
12  * circumvent, violate, or enable the circumvention or violation of, any
13  * terms of an Apple operating system software license agreement.
14  *
15  * Please obtain a copy of the License at
16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
17  *
18  * The Original Code and all software distributed under the License are
19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23  * Please see the License for the specific language governing rights and
24  * limitations under the License.
25  *
26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27  */
28 /*
29  * @OSF_COPYRIGHT@
30  */
31 /*
32  * Mach Operating System
33  * Copyright (c) 1991,1990,1989,1988,1987 Carnegie Mellon University
34  * All Rights Reserved.
35  *
36  * Permission to use, copy, modify and distribute this software and its
37  * documentation is hereby granted, provided that both the copyright
38  * notice and this permission notice appear in all copies of the
39  * software, derivative works or modified versions, and any portions
40  * thereof, and that both notices appear in supporting documentation.
41  *
42  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
43  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND FOR
44  * ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
45  *
46  * Carnegie Mellon requests users of this software to return to
47  *
48  *  Software Distribution Coordinator  or  [email protected]
49  *  School of Computer Science
50  *  Carnegie Mellon University
51  *  Pittsburgh PA 15213-3890
52  *
53  * any improvements or extensions that they make and grant Carnegie Mellon
54  * the rights to redistribute these changes.
55  */
56 /*
57  */
58 /*
59  *	File:	vm/vm_pageout.c
60  *	Author:	Avadis Tevanian, Jr., Michael Wayne Young
61  *	Date:	1985
62  *
63  *	The proverbial page-out daemon.
64  */
65 
66 #include <stdint.h>
67 #include <ptrauth.h>
68 
69 #include <debug.h>
70 
71 #include <mach/mach_types.h>
72 #include <mach/memory_object.h>
73 #include <mach/mach_host_server.h>
74 #include <mach/upl.h>
75 #include <mach/vm_map.h>
76 #include <mach/vm_param.h>
77 #include <mach/vm_statistics.h>
78 #include <mach/sdt.h>
79 
80 #include <kern/kern_types.h>
81 #include <kern/counter.h>
82 #include <kern/host_statistics.h>
83 #include <kern/machine.h>
84 #include <kern/misc_protos.h>
85 #include <kern/sched.h>
86 #include <kern/thread.h>
87 #include <kern/kalloc.h>
88 #include <kern/zalloc_internal.h>
89 #include <kern/policy_internal.h>
90 #include <kern/thread_group.h>
91 
92 #include <os/log.h>
93 
94 #include <sys/kdebug_triage.h>
95 
96 #include <machine/vm_tuning.h>
97 #include <machine/commpage.h>
98 
99 #include <vm/pmap.h>
100 #include <vm/vm_compressor_pager.h>
101 #include <vm/vm_fault.h>
102 #include <vm/vm_map_internal.h>
103 #include <vm/vm_object.h>
104 #include <vm/vm_page.h>
105 #include <vm/vm_pageout.h>
106 #include <vm/vm_protos.h> /* must be last */
107 #include <vm/memory_object.h>
108 #include <vm/vm_purgeable_internal.h>
109 #include <vm/vm_shared_region.h>
110 #include <vm/vm_compressor.h>
111 
112 #include <san/kasan.h>
113 
114 #if CONFIG_PHANTOM_CACHE
115 #include <vm/vm_phantom_cache.h>
116 #endif
117 
118 #if UPL_DEBUG
119 #include <libkern/OSDebug.h>
120 #endif
121 
122 extern int cs_debug;
123 
124 extern void mbuf_drain(boolean_t);
125 
126 #if VM_PRESSURE_EVENTS
127 #if CONFIG_JETSAM
128 extern unsigned int memorystatus_available_pages;
129 extern unsigned int memorystatus_available_pages_pressure;
130 extern unsigned int memorystatus_available_pages_critical;
131 #else /* CONFIG_JETSAM */
132 extern uint64_t memorystatus_available_pages;
133 extern uint64_t memorystatus_available_pages_pressure;
134 extern uint64_t memorystatus_available_pages_critical;
135 #endif /* CONFIG_JETSAM */
136 
137 extern unsigned int memorystatus_frozen_count;
138 extern unsigned int memorystatus_suspended_count;
139 extern vm_pressure_level_t memorystatus_vm_pressure_level;
140 
141 extern lck_mtx_t memorystatus_jetsam_fg_band_lock;
142 extern uint32_t memorystatus_jetsam_fg_band_waiters;
143 
144 void vm_pressure_response(void);
145 extern void consider_vm_pressure_events(void);
146 
147 #define MEMORYSTATUS_SUSPENDED_THRESHOLD  4
148 #endif /* VM_PRESSURE_EVENTS */
149 
150 SECURITY_READ_ONLY_LATE(thread_t) vm_pageout_scan_thread;
151 SECURITY_READ_ONLY_LATE(thread_t) vm_pageout_gc_thread;
152 boolean_t vps_dynamic_priority_enabled = FALSE;
153 boolean_t vps_yield_for_pgqlockwaiters = TRUE;
154 
155 #ifndef VM_PAGEOUT_BURST_INACTIVE_THROTTLE  /* maximum iterations of the inactive queue w/o stealing/cleaning a page */
156 #if !XNU_TARGET_OS_OSX
157 #define VM_PAGEOUT_BURST_INACTIVE_THROTTLE 1024
158 #else /* !XNU_TARGET_OS_OSX */
159 #define VM_PAGEOUT_BURST_INACTIVE_THROTTLE 4096
160 #endif /* !XNU_TARGET_OS_OSX */
161 #endif
162 
163 #ifndef VM_PAGEOUT_DEADLOCK_RELIEF
164 #define VM_PAGEOUT_DEADLOCK_RELIEF 100  /* number of pages to move to break deadlock */
165 #endif
166 
167 #ifndef VM_PAGE_LAUNDRY_MAX
168 #define VM_PAGE_LAUNDRY_MAX     128UL   /* maximum pageouts on a given pageout queue */
169 #endif  /* VM_PAGEOUT_LAUNDRY_MAX */
170 
171 #ifndef VM_PAGEOUT_BURST_WAIT
172 #define VM_PAGEOUT_BURST_WAIT   1       /* milliseconds */
173 #endif  /* VM_PAGEOUT_BURST_WAIT */
174 
175 #ifndef VM_PAGEOUT_EMPTY_WAIT
176 #define VM_PAGEOUT_EMPTY_WAIT   50      /* milliseconds */
177 #endif  /* VM_PAGEOUT_EMPTY_WAIT */
178 
179 #ifndef VM_PAGEOUT_DEADLOCK_WAIT
180 #define VM_PAGEOUT_DEADLOCK_WAIT 100    /* milliseconds */
181 #endif  /* VM_PAGEOUT_DEADLOCK_WAIT */
182 
183 #ifndef VM_PAGEOUT_IDLE_WAIT
184 #define VM_PAGEOUT_IDLE_WAIT    10      /* milliseconds */
185 #endif  /* VM_PAGEOUT_IDLE_WAIT */
186 
187 #ifndef VM_PAGEOUT_SWAP_WAIT
188 #define VM_PAGEOUT_SWAP_WAIT    10      /* milliseconds */
189 #endif  /* VM_PAGEOUT_SWAP_WAIT */
190 
191 
192 #ifndef VM_PAGE_SPECULATIVE_TARGET
193 #define VM_PAGE_SPECULATIVE_TARGET(total) ((total) * 1 / (100 / vm_pageout_state.vm_page_speculative_percentage))
194 #endif /* VM_PAGE_SPECULATIVE_TARGET */
195 
196 
197 /*
198  *	To obtain a reasonable LRU approximation, the inactive queue
199  *	needs to be large enough to give pages on it a chance to be
200  *	referenced a second time.  This macro defines the fraction
201  *	of active+inactive pages that should be inactive.
202  *	The pageout daemon uses it to update vm_page_inactive_target.
203  *
204  *	If vm_page_free_count falls below vm_page_free_target and
205  *	vm_page_inactive_count is below vm_page_inactive_target,
206  *	then the pageout daemon starts running.
207  */
208 
209 #ifndef VM_PAGE_INACTIVE_TARGET
210 #define VM_PAGE_INACTIVE_TARGET(avail)  ((avail) * 1 / 2)
211 #endif  /* VM_PAGE_INACTIVE_TARGET */
212 
213 /*
214  *	Once the pageout daemon starts running, it keeps going
215  *	until vm_page_free_count meets or exceeds vm_page_free_target.
216  */
217 
218 #ifndef VM_PAGE_FREE_TARGET
219 #if !XNU_TARGET_OS_OSX
220 #define VM_PAGE_FREE_TARGET(free)       (15 + (free) / 100)
221 #else /* !XNU_TARGET_OS_OSX */
222 #define VM_PAGE_FREE_TARGET(free)       (15 + (free) / 80)
223 #endif /* !XNU_TARGET_OS_OSX */
224 #endif  /* VM_PAGE_FREE_TARGET */
225 
226 
227 /*
228  *	The pageout daemon always starts running once vm_page_free_count
229  *	falls below vm_page_free_min.
230  */
231 
232 #ifndef VM_PAGE_FREE_MIN
233 #if !XNU_TARGET_OS_OSX
234 #define VM_PAGE_FREE_MIN(free)          (10 + (free) / 200)
235 #else /* !XNU_TARGET_OS_OSX */
236 #define VM_PAGE_FREE_MIN(free)          (10 + (free) / 100)
237 #endif /* !XNU_TARGET_OS_OSX */
238 #endif  /* VM_PAGE_FREE_MIN */
239 
240 #if !XNU_TARGET_OS_OSX
241 #define VM_PAGE_FREE_RESERVED_LIMIT     100
242 #define VM_PAGE_FREE_MIN_LIMIT          1500
243 #define VM_PAGE_FREE_TARGET_LIMIT       2000
244 #else /* !XNU_TARGET_OS_OSX */
245 #define VM_PAGE_FREE_RESERVED_LIMIT     1700
246 #define VM_PAGE_FREE_MIN_LIMIT          3500
247 #define VM_PAGE_FREE_TARGET_LIMIT       4000
248 #endif /* !XNU_TARGET_OS_OSX */
249 
250 /*
251  *	When vm_page_free_count falls below vm_page_free_reserved,
252  *	only vm-privileged threads can allocate pages.  vm-privilege
253  *	allows the pageout daemon and default pager (and any other
254  *	associated threads needed for default pageout) to continue
255  *	operation by dipping into the reserved pool of pages.
256  */
257 
258 #ifndef VM_PAGE_FREE_RESERVED
259 #define VM_PAGE_FREE_RESERVED(n)        \
260 	((unsigned) (6 * VM_PAGE_LAUNDRY_MAX) + (n))
261 #endif  /* VM_PAGE_FREE_RESERVED */
262 
263 /*
264  *	When we dequeue pages from the inactive list, they are
265  *	reactivated (ie, put back on the active queue) if referenced.
266  *	However, it is possible to starve the free list if other
267  *	processors are referencing pages faster than we can turn off
268  *	the referenced bit.  So we limit the number of reactivations
269  *	we will make per call of vm_pageout_scan().
270  */
271 #define VM_PAGE_REACTIVATE_LIMIT_MAX 20000
272 
273 #ifndef VM_PAGE_REACTIVATE_LIMIT
274 #if !XNU_TARGET_OS_OSX
275 #define VM_PAGE_REACTIVATE_LIMIT(avail) (VM_PAGE_INACTIVE_TARGET(avail) / 2)
276 #else /* !XNU_TARGET_OS_OSX */
277 #define VM_PAGE_REACTIVATE_LIMIT(avail) (MAX((avail) * 1 / 20,VM_PAGE_REACTIVATE_LIMIT_MAX))
278 #endif /* !XNU_TARGET_OS_OSX */
279 #endif  /* VM_PAGE_REACTIVATE_LIMIT */
280 #define VM_PAGEOUT_INACTIVE_FORCE_RECLAIM       1000
281 
282 int vm_pageout_protect_realtime = true;
283 
284 extern boolean_t hibernate_cleaning_in_progress;
285 
286 struct pgo_iothread_state pgo_iothread_internal_state[MAX_COMPRESSOR_THREAD_COUNT];
287 struct pgo_iothread_state pgo_iothread_external_state;
288 
289 #if VM_PRESSURE_EVENTS
290 void vm_pressure_thread(void);
291 
292 boolean_t VM_PRESSURE_NORMAL_TO_WARNING(void);
293 boolean_t VM_PRESSURE_WARNING_TO_CRITICAL(void);
294 
295 boolean_t VM_PRESSURE_WARNING_TO_NORMAL(void);
296 boolean_t VM_PRESSURE_CRITICAL_TO_WARNING(void);
297 #endif
298 
299 static void vm_pageout_iothread_external(struct pgo_iothread_state *, wait_result_t);
300 static void vm_pageout_iothread_internal(struct pgo_iothread_state *, wait_result_t);
301 static void vm_pageout_adjust_eq_iothrottle(struct pgo_iothread_state *, boolean_t);
302 
303 extern void vm_pageout_continue(void);
304 extern void vm_pageout_scan(void);
305 
306 boolean_t vm_pageout_running = FALSE;
307 
308 uint32_t vm_page_upl_tainted = 0;
309 uint32_t vm_page_iopl_tainted = 0;
310 
311 #if XNU_TARGET_OS_OSX
312 static boolean_t vm_pageout_waiter  = FALSE;
313 #endif /* XNU_TARGET_OS_OSX */
314 
315 
316 #if DEVELOPMENT || DEBUG
317 struct vm_pageout_debug vm_pageout_debug;
318 #endif
319 struct vm_pageout_vminfo vm_pageout_vminfo;
320 struct vm_pageout_state  vm_pageout_state;
321 struct vm_config         vm_config;
322 
323 struct  vm_pageout_queue vm_pageout_queue_internal VM_PAGE_PACKED_ALIGNED;
324 struct  vm_pageout_queue vm_pageout_queue_external VM_PAGE_PACKED_ALIGNED;
325 #if DEVELOPMENT || DEBUG
326 struct vm_pageout_queue vm_pageout_queue_benchmark VM_PAGE_PACKED_ALIGNED;
327 #endif /* DEVELOPMENT || DEBUG */
328 
329 int         vm_upl_wait_for_pages = 0;
330 vm_object_t vm_pageout_scan_wants_object = VM_OBJECT_NULL;
331 
332 boolean_t(*volatile consider_buffer_cache_collect)(int) = NULL;
333 
334 int     vm_debug_events = 0;
335 
336 LCK_GRP_DECLARE(vm_pageout_lck_grp, "vm_pageout");
337 
338 #if CONFIG_MEMORYSTATUS
339 extern boolean_t memorystatus_kill_on_VM_page_shortage(void);
340 
341 uint32_t vm_pageout_memorystatus_fb_factor_nr = 5;
342 uint32_t vm_pageout_memorystatus_fb_factor_dr = 2;
343 
344 #endif
345 
346 #if __AMP__
347 
348 // bind compressor threads e-cores
349 #define VM_COMPRESSOR_EBOUND_DEFAULT 1
350 
351 TUNABLE(bool, vm_compressor_ebound, "vmcomp_ecluster", VM_COMPRESSOR_EBOUND_DEFAULT);
352 int vm_pgo_pbound = 0;
353 extern void thread_bind_cluster_type(thread_t, char, bool);
354 
355 #endif /* __AMP__ */
356 
357 
358 /*
359  *	Routine:	vm_pageout_object_terminate
360  *	Purpose:
361  *		Destroy the pageout_object, and perform all of the
362  *		required cleanup actions.
363  *
364  *	In/Out conditions:
365  *		The object must be locked, and will be returned locked.
366  */
367 void
vm_pageout_object_terminate(vm_object_t object)368 vm_pageout_object_terminate(
369 	vm_object_t     object)
370 {
371 	vm_object_t     shadow_object;
372 
373 	/*
374 	 * Deal with the deallocation (last reference) of a pageout object
375 	 * (used for cleaning-in-place) by dropping the paging references/
376 	 * freeing pages in the original object.
377 	 */
378 
379 	assert(object->pageout);
380 	shadow_object = object->shadow;
381 	vm_object_lock(shadow_object);
382 
383 	while (!vm_page_queue_empty(&object->memq)) {
384 		vm_page_t               p, m;
385 		vm_object_offset_t      offset;
386 
387 		p = (vm_page_t) vm_page_queue_first(&object->memq);
388 
389 		assert(p->vmp_private);
390 		assert(p->vmp_free_when_done);
391 		p->vmp_free_when_done = FALSE;
392 		assert(!p->vmp_cleaning);
393 		assert(!p->vmp_laundry);
394 
395 		offset = p->vmp_offset;
396 		VM_PAGE_FREE(p);
397 		p = VM_PAGE_NULL;
398 
399 		m = vm_page_lookup(shadow_object,
400 		    offset + object->vo_shadow_offset);
401 
402 		if (m == VM_PAGE_NULL) {
403 			continue;
404 		}
405 
406 		assert((m->vmp_dirty) || (m->vmp_precious) ||
407 		    (m->vmp_busy && m->vmp_cleaning));
408 
409 		/*
410 		 * Handle the trusted pager throttle.
411 		 * Also decrement the burst throttle (if external).
412 		 */
413 		vm_page_lock_queues();
414 		if (m->vmp_q_state == VM_PAGE_ON_PAGEOUT_Q) {
415 			vm_pageout_throttle_up(m);
416 		}
417 
418 		/*
419 		 * Handle the "target" page(s). These pages are to be freed if
420 		 * successfully cleaned. Target pages are always busy, and are
421 		 * wired exactly once. The initial target pages are not mapped,
422 		 * (so cannot be referenced or modified) but converted target
423 		 * pages may have been modified between the selection as an
424 		 * adjacent page and conversion to a target.
425 		 */
426 		if (m->vmp_free_when_done) {
427 			assert(m->vmp_busy);
428 			assert(m->vmp_q_state == VM_PAGE_IS_WIRED);
429 			assert(m->vmp_wire_count == 1);
430 			m->vmp_cleaning = FALSE;
431 			m->vmp_free_when_done = FALSE;
432 			/*
433 			 * Revoke all access to the page. Since the object is
434 			 * locked, and the page is busy, this prevents the page
435 			 * from being dirtied after the pmap_disconnect() call
436 			 * returns.
437 			 *
438 			 * Since the page is left "dirty" but "not modifed", we
439 			 * can detect whether the page was redirtied during
440 			 * pageout by checking the modify state.
441 			 */
442 			if (pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m)) & VM_MEM_MODIFIED) {
443 				SET_PAGE_DIRTY(m, FALSE);
444 			} else {
445 				m->vmp_dirty = FALSE;
446 			}
447 
448 			if (m->vmp_dirty) {
449 				vm_page_unwire(m, TRUE);        /* reactivates */
450 				counter_inc(&vm_statistics_reactivations);
451 				PAGE_WAKEUP_DONE(m);
452 			} else {
453 				vm_page_free(m);  /* clears busy, etc. */
454 			}
455 			vm_page_unlock_queues();
456 			continue;
457 		}
458 		/*
459 		 * Handle the "adjacent" pages. These pages were cleaned in
460 		 * place, and should be left alone.
461 		 * If prep_pin_count is nonzero, then someone is using the
462 		 * page, so make it active.
463 		 */
464 		if ((m->vmp_q_state == VM_PAGE_NOT_ON_Q) && !m->vmp_private) {
465 			if (m->vmp_reference) {
466 				vm_page_activate(m);
467 			} else {
468 				vm_page_deactivate(m);
469 			}
470 		}
471 		if (m->vmp_overwriting) {
472 			/*
473 			 * the (COPY_OUT_FROM == FALSE) request_page_list case
474 			 */
475 			if (m->vmp_busy) {
476 				/*
477 				 * We do not re-set m->vmp_dirty !
478 				 * The page was busy so no extraneous activity
479 				 * could have occurred. COPY_INTO is a read into the
480 				 * new pages. CLEAN_IN_PLACE does actually write
481 				 * out the pages but handling outside of this code
482 				 * will take care of resetting dirty. We clear the
483 				 * modify however for the Programmed I/O case.
484 				 */
485 				pmap_clear_modify(VM_PAGE_GET_PHYS_PAGE(m));
486 
487 				m->vmp_busy = FALSE;
488 				m->vmp_absent = FALSE;
489 			} else {
490 				/*
491 				 * alternate (COPY_OUT_FROM == FALSE) request_page_list case
492 				 * Occurs when the original page was wired
493 				 * at the time of the list request
494 				 */
495 				assert(VM_PAGE_WIRED(m));
496 				vm_page_unwire(m, TRUE);        /* reactivates */
497 			}
498 			m->vmp_overwriting = FALSE;
499 		} else {
500 			m->vmp_dirty = FALSE;
501 		}
502 		m->vmp_cleaning = FALSE;
503 
504 		/*
505 		 * Wakeup any thread waiting for the page to be un-cleaning.
506 		 */
507 		PAGE_WAKEUP(m);
508 		vm_page_unlock_queues();
509 	}
510 	/*
511 	 * Account for the paging reference taken in vm_paging_object_allocate.
512 	 */
513 	vm_object_activity_end(shadow_object);
514 	vm_object_unlock(shadow_object);
515 
516 	assert(object->ref_count == 0);
517 	assert(object->paging_in_progress == 0);
518 	assert(object->activity_in_progress == 0);
519 	assert(object->resident_page_count == 0);
520 	return;
521 }
522 
523 /*
524  * Routine:	vm_pageclean_setup
525  *
526  * Purpose:	setup a page to be cleaned (made non-dirty), but not
527  *		necessarily flushed from the VM page cache.
528  *		This is accomplished by cleaning in place.
529  *
530  *		The page must not be busy, and new_object
531  *		must be locked.
532  *
533  */
534 static void
vm_pageclean_setup(vm_page_t m,vm_page_t new_m,vm_object_t new_object,vm_object_offset_t new_offset)535 vm_pageclean_setup(
536 	vm_page_t               m,
537 	vm_page_t               new_m,
538 	vm_object_t             new_object,
539 	vm_object_offset_t      new_offset)
540 {
541 	assert(!m->vmp_busy);
542 #if 0
543 	assert(!m->vmp_cleaning);
544 #endif
545 
546 	pmap_clear_modify(VM_PAGE_GET_PHYS_PAGE(m));
547 
548 	/*
549 	 * Mark original page as cleaning in place.
550 	 */
551 	m->vmp_cleaning = TRUE;
552 	SET_PAGE_DIRTY(m, FALSE);
553 	m->vmp_precious = FALSE;
554 
555 	/*
556 	 * Convert the fictitious page to a private shadow of
557 	 * the real page.
558 	 */
559 	assert(new_m->vmp_fictitious);
560 	assert(VM_PAGE_GET_PHYS_PAGE(new_m) == vm_page_fictitious_addr);
561 	new_m->vmp_fictitious = FALSE;
562 	new_m->vmp_private = TRUE;
563 	new_m->vmp_free_when_done = TRUE;
564 	VM_PAGE_SET_PHYS_PAGE(new_m, VM_PAGE_GET_PHYS_PAGE(m));
565 
566 	vm_page_lockspin_queues();
567 	vm_page_wire(new_m, VM_KERN_MEMORY_NONE, TRUE);
568 	vm_page_unlock_queues();
569 
570 	vm_page_insert_wired(new_m, new_object, new_offset, VM_KERN_MEMORY_NONE);
571 	assert(!new_m->vmp_wanted);
572 	new_m->vmp_busy = FALSE;
573 }
574 
575 /*
576  *	Routine:	vm_pageout_initialize_page
577  *	Purpose:
578  *		Causes the specified page to be initialized in
579  *		the appropriate memory object. This routine is used to push
580  *		pages into a copy-object when they are modified in the
581  *		permanent object.
582  *
583  *		The page is moved to a temporary object and paged out.
584  *
585  *	In/out conditions:
586  *		The page in question must not be on any pageout queues.
587  *		The object to which it belongs must be locked.
588  *		The page must be busy, but not hold a paging reference.
589  *
590  *	Implementation:
591  *		Move this page to a completely new object.
592  */
593 void
vm_pageout_initialize_page(vm_page_t m)594 vm_pageout_initialize_page(
595 	vm_page_t       m)
596 {
597 	vm_object_t             object;
598 	vm_object_offset_t      paging_offset;
599 	memory_object_t         pager;
600 
601 	assert(VM_CONFIG_COMPRESSOR_IS_PRESENT);
602 
603 	object = VM_PAGE_OBJECT(m);
604 
605 	assert(m->vmp_busy);
606 	assert(object->internal);
607 
608 	/*
609 	 *	Verify that we really want to clean this page
610 	 */
611 	assert(!m->vmp_absent);
612 	assert(m->vmp_dirty);
613 
614 	/*
615 	 *	Create a paging reference to let us play with the object.
616 	 */
617 	paging_offset = m->vmp_offset + object->paging_offset;
618 
619 	if (m->vmp_absent || VMP_ERROR_GET(m) || m->vmp_restart || (!m->vmp_dirty && !m->vmp_precious)) {
620 		panic("reservation without pageout?"); /* alan */
621 
622 		VM_PAGE_FREE(m);
623 		vm_object_unlock(object);
624 
625 		return;
626 	}
627 
628 	/*
629 	 * If there's no pager, then we can't clean the page.  This should
630 	 * never happen since this should be a copy object and therefore not
631 	 * an external object, so the pager should always be there.
632 	 */
633 
634 	pager = object->pager;
635 
636 	if (pager == MEMORY_OBJECT_NULL) {
637 		panic("missing pager for copy object");
638 
639 		VM_PAGE_FREE(m);
640 		return;
641 	}
642 
643 	/*
644 	 * set the page for future call to vm_fault_list_request
645 	 */
646 	pmap_clear_modify(VM_PAGE_GET_PHYS_PAGE(m));
647 	SET_PAGE_DIRTY(m, FALSE);
648 
649 	/*
650 	 * keep the object from collapsing or terminating
651 	 */
652 	vm_object_paging_begin(object);
653 	vm_object_unlock(object);
654 
655 	/*
656 	 *	Write the data to its pager.
657 	 *	Note that the data is passed by naming the new object,
658 	 *	not a virtual address; the pager interface has been
659 	 *	manipulated to use the "internal memory" data type.
660 	 *	[The object reference from its allocation is donated
661 	 *	to the eventual recipient.]
662 	 */
663 	memory_object_data_initialize(pager, paging_offset, PAGE_SIZE);
664 
665 	vm_object_lock(object);
666 	vm_object_paging_end(object);
667 }
668 
669 
670 /*
671  * vm_pageout_cluster:
672  *
673  * Given a page, queue it to the appropriate I/O thread,
674  * which will page it out and attempt to clean adjacent pages
675  * in the same operation.
676  *
677  * The object and queues must be locked. We will take a
678  * paging reference to prevent deallocation or collapse when we
679  * release the object lock back at the call site.  The I/O thread
680  * is responsible for consuming this reference
681  *
682  * The page must not be on any pageout queue.
683  */
684 #if DEVELOPMENT || DEBUG
685 vmct_stats_t vmct_stats;
686 
687 int32_t vmct_active = 0;
688 uint64_t vm_compressor_epoch_start = 0;
689 uint64_t vm_compressor_epoch_stop = 0;
690 
691 typedef enum vmct_state_t {
692 	VMCT_IDLE,
693 	VMCT_AWAKENED,
694 	VMCT_ACTIVE,
695 } vmct_state_t;
696 vmct_state_t vmct_state[MAX_COMPRESSOR_THREAD_COUNT];
697 #endif
698 
699 
700 
701 static void
vm_pageout_cluster_to_queue(vm_page_t m,struct vm_pageout_queue * q)702 vm_pageout_cluster_to_queue(vm_page_t m, struct vm_pageout_queue *q)
703 {
704 	vm_object_t object = VM_PAGE_OBJECT(m);
705 
706 	VM_PAGE_CHECK(m);
707 	LCK_MTX_ASSERT(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED);
708 	vm_object_lock_assert_exclusive(object);
709 
710 	/*
711 	 * Make sure it's OK to page this out.
712 	 */
713 	assert((m->vmp_dirty || m->vmp_precious) && (!VM_PAGE_WIRED(m)));
714 	assert(!m->vmp_cleaning && !m->vmp_laundry);
715 	assert(m->vmp_q_state == VM_PAGE_NOT_ON_Q);
716 
717 	/*
718 	 * protect the object from collapse or termination
719 	 */
720 	vm_object_activity_begin(object);
721 
722 
723 	/*
724 	 * pgo_laundry count is tied to the laundry bit
725 	 */
726 	m->vmp_laundry = TRUE;
727 	q->pgo_laundry++;
728 
729 	m->vmp_q_state = VM_PAGE_ON_PAGEOUT_Q;
730 	vm_page_queue_enter(&q->pgo_pending, m, vmp_pageq);
731 
732 	// the benchmark queue will be woken up independently by the benchmark itself
733 	if (
734 		object->internal == TRUE
735 #if DEVELOPMENT || DEBUG
736 		&& q != &vm_pageout_queue_benchmark
737 #endif
738 		) {
739 		assert(VM_CONFIG_COMPRESSOR_IS_PRESENT);
740 		m->vmp_busy = TRUE;
741 		// Wake up the first compressor thread. It will wake subsequent threads if necessary.
742 		sched_cond_signal(&pgo_iothread_internal_state[0].pgo_wakeup, pgo_iothread_internal_state[0].pgo_iothread);
743 	} else {
744 		sched_cond_signal(&pgo_iothread_external_state.pgo_wakeup, pgo_iothread_external_state.pgo_iothread);
745 	}
746 	VM_PAGE_CHECK(m);
747 }
748 
749 void
vm_pageout_cluster(vm_page_t m)750 vm_pageout_cluster(vm_page_t m)
751 {
752 	struct          vm_pageout_queue *q;
753 	vm_object_t     object = VM_PAGE_OBJECT(m);
754 	if (object->internal) {
755 		q = &vm_pageout_queue_internal;
756 	} else {
757 		q = &vm_pageout_queue_external;
758 	}
759 	vm_pageout_cluster_to_queue(m, q);
760 }
761 
762 
763 /*
764  * A page is back from laundry or we are stealing it back from
765  * the laundering state.  See if there are some pages waiting to
766  * go to laundry and if we can let some of them go now.
767  *
768  * Object and page queues must be locked.
769  */
770 void
vm_pageout_throttle_up(vm_page_t m)771 vm_pageout_throttle_up(
772 	vm_page_t       m)
773 {
774 	struct vm_pageout_queue *q;
775 	vm_object_t      m_object;
776 
777 	m_object = VM_PAGE_OBJECT(m);
778 
779 	assert(m_object != VM_OBJECT_NULL);
780 	assert(m_object != kernel_object);
781 
782 	LCK_MTX_ASSERT(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED);
783 	vm_object_lock_assert_exclusive(m_object);
784 
785 	if (m_object->internal == TRUE) {
786 		q = &vm_pageout_queue_internal;
787 	} else {
788 		q = &vm_pageout_queue_external;
789 	}
790 
791 	if (m->vmp_q_state == VM_PAGE_ON_PAGEOUT_Q) {
792 		vm_page_queue_remove(&q->pgo_pending, m, vmp_pageq);
793 		m->vmp_q_state = VM_PAGE_NOT_ON_Q;
794 
795 		VM_PAGE_ZERO_PAGEQ_ENTRY(m);
796 
797 		vm_object_activity_end(m_object);
798 
799 		VM_PAGEOUT_DEBUG(vm_page_steal_pageout_page, 1);
800 	}
801 	if (m->vmp_laundry == TRUE) {
802 		m->vmp_laundry = FALSE;
803 		q->pgo_laundry--;
804 
805 		if (q->pgo_throttled == TRUE) {
806 			q->pgo_throttled = FALSE;
807 			thread_wakeup((event_t) &q->pgo_laundry);
808 		}
809 		if (q->pgo_draining == TRUE && q->pgo_laundry == 0) {
810 			q->pgo_draining = FALSE;
811 			thread_wakeup((event_t) (&q->pgo_laundry + 1));
812 		}
813 		VM_PAGEOUT_DEBUG(vm_pageout_throttle_up_count, 1);
814 	}
815 }
816 
817 
818 static void
vm_pageout_throttle_up_batch(struct vm_pageout_queue * q,int batch_cnt)819 vm_pageout_throttle_up_batch(
820 	struct vm_pageout_queue *q,
821 	int             batch_cnt)
822 {
823 	LCK_MTX_ASSERT(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED);
824 
825 	VM_PAGEOUT_DEBUG(vm_pageout_throttle_up_count, batch_cnt);
826 
827 	q->pgo_laundry -= batch_cnt;
828 
829 	if (q->pgo_throttled == TRUE) {
830 		q->pgo_throttled = FALSE;
831 		thread_wakeup((event_t) &q->pgo_laundry);
832 	}
833 	if (q->pgo_draining == TRUE && q->pgo_laundry == 0) {
834 		q->pgo_draining = FALSE;
835 		thread_wakeup((event_t) (&q->pgo_laundry + 1));
836 	}
837 }
838 
839 
840 
841 /*
842  * VM memory pressure monitoring.
843  *
844  * vm_pageout_scan() keeps track of the number of pages it considers and
845  * reclaims, in the currently active vm_pageout_stat[vm_pageout_stat_now].
846  *
847  * compute_memory_pressure() is called every second from compute_averages()
848  * and moves "vm_pageout_stat_now" forward, to start accumulating the number
849  * of recalimed pages in a new vm_pageout_stat[] bucket.
850  *
851  * mach_vm_pressure_monitor() collects past statistics about memory pressure.
852  * The caller provides the number of seconds ("nsecs") worth of statistics
853  * it wants, up to 30 seconds.
854  * It computes the number of pages reclaimed in the past "nsecs" seconds and
855  * also returns the number of pages the system still needs to reclaim at this
856  * moment in time.
857  */
858 #if DEVELOPMENT || DEBUG
859 #define VM_PAGEOUT_STAT_SIZE    (30 * 8) + 1
860 #else
861 #define VM_PAGEOUT_STAT_SIZE    (1 * 8) + 1
862 #endif
863 struct vm_pageout_stat {
864 	unsigned long vm_page_active_count;
865 	unsigned long vm_page_speculative_count;
866 	unsigned long vm_page_inactive_count;
867 	unsigned long vm_page_anonymous_count;
868 
869 	unsigned long vm_page_free_count;
870 	unsigned long vm_page_wire_count;
871 	unsigned long vm_page_compressor_count;
872 
873 	unsigned long vm_page_pages_compressed;
874 	unsigned long vm_page_pageable_internal_count;
875 	unsigned long vm_page_pageable_external_count;
876 	unsigned long vm_page_xpmapped_external_count;
877 
878 	unsigned int pages_grabbed;
879 	unsigned int pages_freed;
880 
881 	unsigned int pages_compressed;
882 	unsigned int pages_grabbed_by_compressor;
883 	unsigned int failed_compressions;
884 
885 	unsigned int pages_evicted;
886 	unsigned int pages_purged;
887 
888 	unsigned int considered;
889 	unsigned int considered_bq_internal;
890 	unsigned int considered_bq_external;
891 
892 	unsigned int skipped_external;
893 	unsigned int skipped_internal;
894 	unsigned int filecache_min_reactivations;
895 
896 	unsigned int freed_speculative;
897 	unsigned int freed_cleaned;
898 	unsigned int freed_internal;
899 	unsigned int freed_external;
900 
901 	unsigned int cleaned_dirty_external;
902 	unsigned int cleaned_dirty_internal;
903 
904 	unsigned int inactive_referenced;
905 	unsigned int inactive_nolock;
906 	unsigned int reactivation_limit_exceeded;
907 	unsigned int forced_inactive_reclaim;
908 
909 	unsigned int throttled_internal_q;
910 	unsigned int throttled_external_q;
911 
912 	unsigned int phantom_ghosts_found;
913 	unsigned int phantom_ghosts_added;
914 
915 	unsigned int vm_page_realtime_count;
916 	unsigned int forcereclaimed_sharedcache;
917 	unsigned int forcereclaimed_realtime;
918 	unsigned int protected_sharedcache;
919 	unsigned int protected_realtime;
920 } vm_pageout_stats[VM_PAGEOUT_STAT_SIZE];
921 
922 unsigned int vm_pageout_stat_now = 0;
923 
924 #define VM_PAGEOUT_STAT_BEFORE(i) \
925 	(((i) == 0) ? VM_PAGEOUT_STAT_SIZE - 1 : (i) - 1)
926 #define VM_PAGEOUT_STAT_AFTER(i) \
927 	(((i) == VM_PAGEOUT_STAT_SIZE - 1) ? 0 : (i) + 1)
928 
929 #if VM_PAGE_BUCKETS_CHECK
930 int vm_page_buckets_check_interval = 80; /* in eighths of a second */
931 #endif /* VM_PAGE_BUCKETS_CHECK */
932 
933 
934 void
935 record_memory_pressure(void);
936 void
record_memory_pressure(void)937 record_memory_pressure(void)
938 {
939 	unsigned int vm_pageout_next;
940 
941 #if VM_PAGE_BUCKETS_CHECK
942 	/* check the consistency of VM page buckets at regular interval */
943 	static int counter = 0;
944 	if ((++counter % vm_page_buckets_check_interval) == 0) {
945 		vm_page_buckets_check();
946 	}
947 #endif /* VM_PAGE_BUCKETS_CHECK */
948 
949 	vm_pageout_state.vm_memory_pressure =
950 	    vm_pageout_stats[VM_PAGEOUT_STAT_BEFORE(vm_pageout_stat_now)].freed_speculative +
951 	    vm_pageout_stats[VM_PAGEOUT_STAT_BEFORE(vm_pageout_stat_now)].freed_cleaned +
952 	    vm_pageout_stats[VM_PAGEOUT_STAT_BEFORE(vm_pageout_stat_now)].freed_internal +
953 	    vm_pageout_stats[VM_PAGEOUT_STAT_BEFORE(vm_pageout_stat_now)].freed_external;
954 
955 	commpage_set_memory_pressure((unsigned int)vm_pageout_state.vm_memory_pressure );
956 
957 	/* move "now" forward */
958 	vm_pageout_next = VM_PAGEOUT_STAT_AFTER(vm_pageout_stat_now);
959 
960 	bzero(&vm_pageout_stats[vm_pageout_next], sizeof(struct vm_pageout_stat));
961 
962 	vm_pageout_stat_now = vm_pageout_next;
963 }
964 
965 
966 /*
967  * IMPORTANT
968  * mach_vm_ctl_page_free_wanted() is called indirectly, via
969  * mach_vm_pressure_monitor(), when taking a stackshot. Therefore,
970  * it must be safe in the restricted stackshot context. Locks and/or
971  * blocking are not allowable.
972  */
973 unsigned int
mach_vm_ctl_page_free_wanted(void)974 mach_vm_ctl_page_free_wanted(void)
975 {
976 	unsigned int page_free_target, page_free_count, page_free_wanted;
977 
978 	page_free_target = vm_page_free_target;
979 	page_free_count = vm_page_free_count;
980 	if (page_free_target > page_free_count) {
981 		page_free_wanted = page_free_target - page_free_count;
982 	} else {
983 		page_free_wanted = 0;
984 	}
985 
986 	return page_free_wanted;
987 }
988 
989 
990 /*
991  * IMPORTANT:
992  * mach_vm_pressure_monitor() is called when taking a stackshot, with
993  * wait_for_pressure FALSE, so that code path must remain safe in the
994  * restricted stackshot context. No blocking or locks are allowable.
995  * on that code path.
996  */
997 
998 kern_return_t
mach_vm_pressure_monitor(boolean_t wait_for_pressure,unsigned int nsecs_monitored,unsigned int * pages_reclaimed_p,unsigned int * pages_wanted_p)999 mach_vm_pressure_monitor(
1000 	boolean_t       wait_for_pressure,
1001 	unsigned int    nsecs_monitored,
1002 	unsigned int    *pages_reclaimed_p,
1003 	unsigned int    *pages_wanted_p)
1004 {
1005 	wait_result_t   wr;
1006 	unsigned int    vm_pageout_then, vm_pageout_now;
1007 	unsigned int    pages_reclaimed;
1008 	unsigned int    units_of_monitor;
1009 
1010 	units_of_monitor = 8 * nsecs_monitored;
1011 	/*
1012 	 * We don't take the vm_page_queue_lock here because we don't want
1013 	 * vm_pressure_monitor() to get in the way of the vm_pageout_scan()
1014 	 * thread when it's trying to reclaim memory.  We don't need fully
1015 	 * accurate monitoring anyway...
1016 	 */
1017 
1018 	if (wait_for_pressure) {
1019 		/* wait until there's memory pressure */
1020 		while (vm_page_free_count >= vm_page_free_target) {
1021 			wr = assert_wait((event_t) &vm_page_free_wanted,
1022 			    THREAD_INTERRUPTIBLE);
1023 			if (wr == THREAD_WAITING) {
1024 				wr = thread_block(THREAD_CONTINUE_NULL);
1025 			}
1026 			if (wr == THREAD_INTERRUPTED) {
1027 				return KERN_ABORTED;
1028 			}
1029 			if (wr == THREAD_AWAKENED) {
1030 				/*
1031 				 * The memory pressure might have already
1032 				 * been relieved but let's not block again
1033 				 * and let's report that there was memory
1034 				 * pressure at some point.
1035 				 */
1036 				break;
1037 			}
1038 		}
1039 	}
1040 
1041 	/* provide the number of pages the system wants to reclaim */
1042 	if (pages_wanted_p != NULL) {
1043 		*pages_wanted_p = mach_vm_ctl_page_free_wanted();
1044 	}
1045 
1046 	if (pages_reclaimed_p == NULL) {
1047 		return KERN_SUCCESS;
1048 	}
1049 
1050 	/* provide number of pages reclaimed in the last "nsecs_monitored" */
1051 	vm_pageout_now = vm_pageout_stat_now;
1052 	pages_reclaimed = 0;
1053 	for (vm_pageout_then =
1054 	    VM_PAGEOUT_STAT_BEFORE(vm_pageout_now);
1055 	    vm_pageout_then != vm_pageout_now &&
1056 	    units_of_monitor-- != 0;
1057 	    vm_pageout_then =
1058 	    VM_PAGEOUT_STAT_BEFORE(vm_pageout_then)) {
1059 		pages_reclaimed += vm_pageout_stats[vm_pageout_then].freed_speculative;
1060 		pages_reclaimed += vm_pageout_stats[vm_pageout_then].freed_cleaned;
1061 		pages_reclaimed += vm_pageout_stats[vm_pageout_then].freed_internal;
1062 		pages_reclaimed += vm_pageout_stats[vm_pageout_then].freed_external;
1063 	}
1064 	*pages_reclaimed_p = pages_reclaimed;
1065 
1066 	return KERN_SUCCESS;
1067 }
1068 
1069 
1070 
1071 #if DEVELOPMENT || DEBUG
1072 
1073 static void
1074 vm_pageout_disconnect_all_pages_in_queue(vm_page_queue_head_t *, int);
1075 
1076 /*
1077  * condition variable used to make sure there is
1078  * only a single sweep going on at a time
1079  */
1080 boolean_t       vm_pageout_disconnect_all_pages_active = FALSE;
1081 
1082 
1083 void
vm_pageout_disconnect_all_pages()1084 vm_pageout_disconnect_all_pages()
1085 {
1086 	vm_page_lock_queues();
1087 
1088 	if (vm_pageout_disconnect_all_pages_active == TRUE) {
1089 		vm_page_unlock_queues();
1090 		return;
1091 	}
1092 	vm_pageout_disconnect_all_pages_active = TRUE;
1093 	vm_page_unlock_queues();
1094 
1095 	vm_pageout_disconnect_all_pages_in_queue(&vm_page_queue_throttled, vm_page_throttled_count);
1096 	vm_pageout_disconnect_all_pages_in_queue(&vm_page_queue_anonymous, vm_page_anonymous_count);
1097 	vm_pageout_disconnect_all_pages_in_queue(&vm_page_queue_active, vm_page_active_count);
1098 
1099 	vm_pageout_disconnect_all_pages_active = FALSE;
1100 }
1101 
1102 
1103 void
vm_pageout_disconnect_all_pages_in_queue(vm_page_queue_head_t * q,int qcount)1104 vm_pageout_disconnect_all_pages_in_queue(vm_page_queue_head_t *q, int qcount)
1105 {
1106 	vm_page_t       m;
1107 	vm_object_t     t_object = NULL;
1108 	vm_object_t     l_object = NULL;
1109 	vm_object_t     m_object = NULL;
1110 	int             delayed_unlock = 0;
1111 	int             try_failed_count = 0;
1112 	int             disconnected_count = 0;
1113 	int             paused_count = 0;
1114 	int             object_locked_count = 0;
1115 
1116 	KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, (MACHDBG_CODE(DBG_MACH_WORKINGSET, VM_DISCONNECT_ALL_PAGE_MAPPINGS)) | DBG_FUNC_START,
1117 	    q, qcount, 0, 0, 0);
1118 
1119 	vm_page_lock_queues();
1120 
1121 	while (qcount && !vm_page_queue_empty(q)) {
1122 		LCK_MTX_ASSERT(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED);
1123 
1124 		m = (vm_page_t) vm_page_queue_first(q);
1125 		m_object = VM_PAGE_OBJECT(m);
1126 
1127 		/*
1128 		 * check to see if we currently are working
1129 		 * with the same object... if so, we've
1130 		 * already got the lock
1131 		 */
1132 		if (m_object != l_object) {
1133 			/*
1134 			 * the object associated with candidate page is
1135 			 * different from the one we were just working
1136 			 * with... dump the lock if we still own it
1137 			 */
1138 			if (l_object != NULL) {
1139 				vm_object_unlock(l_object);
1140 				l_object = NULL;
1141 			}
1142 			if (m_object != t_object) {
1143 				try_failed_count = 0;
1144 			}
1145 
1146 			/*
1147 			 * Try to lock object; since we've alread got the
1148 			 * page queues lock, we can only 'try' for this one.
1149 			 * if the 'try' fails, we need to do a mutex_pause
1150 			 * to allow the owner of the object lock a chance to
1151 			 * run...
1152 			 */
1153 			if (!vm_object_lock_try_scan(m_object)) {
1154 				if (try_failed_count > 20) {
1155 					goto reenter_pg_on_q;
1156 				}
1157 				vm_page_unlock_queues();
1158 				mutex_pause(try_failed_count++);
1159 				vm_page_lock_queues();
1160 				delayed_unlock = 0;
1161 
1162 				paused_count++;
1163 
1164 				t_object = m_object;
1165 				continue;
1166 			}
1167 			object_locked_count++;
1168 
1169 			l_object = m_object;
1170 		}
1171 		if (!m_object->alive || m->vmp_cleaning || m->vmp_laundry || m->vmp_busy || m->vmp_absent || VMP_ERROR_GET(m) || m->vmp_free_when_done) {
1172 			/*
1173 			 * put it back on the head of its queue
1174 			 */
1175 			goto reenter_pg_on_q;
1176 		}
1177 		if (m->vmp_pmapped == TRUE) {
1178 			pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m));
1179 
1180 			disconnected_count++;
1181 		}
1182 reenter_pg_on_q:
1183 		vm_page_queue_remove(q, m, vmp_pageq);
1184 		vm_page_queue_enter(q, m, vmp_pageq);
1185 
1186 		qcount--;
1187 		try_failed_count = 0;
1188 
1189 		if (delayed_unlock++ > 128) {
1190 			if (l_object != NULL) {
1191 				vm_object_unlock(l_object);
1192 				l_object = NULL;
1193 			}
1194 			lck_mtx_yield(&vm_page_queue_lock);
1195 			delayed_unlock = 0;
1196 		}
1197 	}
1198 	if (l_object != NULL) {
1199 		vm_object_unlock(l_object);
1200 		l_object = NULL;
1201 	}
1202 	vm_page_unlock_queues();
1203 
1204 	KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE, (MACHDBG_CODE(DBG_MACH_WORKINGSET, VM_DISCONNECT_ALL_PAGE_MAPPINGS)) | DBG_FUNC_END,
1205 	    q, disconnected_count, object_locked_count, paused_count, 0);
1206 }
1207 
1208 extern char* proc_best_name(struct proc* proc);
1209 
1210 int
vm_toggle_task_selfdonate_pages(task_t task)1211 vm_toggle_task_selfdonate_pages(task_t task)
1212 {
1213 	int state = 0;
1214 	if (vm_page_donate_mode == VM_PAGE_DONATE_DISABLED) {
1215 		printf("VM Donation mode is OFF on the system\n");
1216 		return state;
1217 	}
1218 	if (task != kernel_task) {
1219 		task_lock(task);
1220 		if (!task->donates_own_pages) {
1221 			printf("SELF DONATE for %s ON\n", proc_best_name(get_bsdtask_info(task)));
1222 			task->donates_own_pages = true;
1223 			state = 1;
1224 		} else if (task->donates_own_pages) {
1225 			printf("SELF DONATE for %s OFF\n", proc_best_name(get_bsdtask_info(task)));
1226 			task->donates_own_pages = false;
1227 			state = 0;
1228 		}
1229 		task_unlock(task);
1230 	}
1231 	return state;
1232 }
1233 #endif /* DEVELOPMENT || DEBUG */
1234 
1235 void
vm_task_set_selfdonate_pages(task_t task,bool donate)1236 vm_task_set_selfdonate_pages(task_t task, bool donate)
1237 {
1238 	assert(vm_page_donate_mode != VM_PAGE_DONATE_DISABLED);
1239 	assert(task != kernel_task);
1240 
1241 	task_lock(task);
1242 	task->donates_own_pages = donate;
1243 	task_unlock(task);
1244 }
1245 
1246 
1247 
1248 static size_t
1249 vm_pageout_page_queue(vm_page_queue_head_t *, size_t, bool);
1250 
1251 /*
1252  * condition variable used to make sure there is
1253  * only a single sweep going on at a time
1254  */
1255 boolean_t       vm_pageout_anonymous_pages_active = FALSE;
1256 
1257 
1258 void
vm_pageout_anonymous_pages()1259 vm_pageout_anonymous_pages()
1260 {
1261 	if (VM_CONFIG_COMPRESSOR_IS_PRESENT) {
1262 		vm_page_lock_queues();
1263 
1264 		if (vm_pageout_anonymous_pages_active == TRUE) {
1265 			vm_page_unlock_queues();
1266 			return;
1267 		}
1268 		vm_pageout_anonymous_pages_active = TRUE;
1269 		vm_page_unlock_queues();
1270 
1271 		vm_pageout_page_queue(&vm_page_queue_throttled, vm_page_throttled_count, false);
1272 		vm_pageout_page_queue(&vm_page_queue_anonymous, vm_page_anonymous_count, false);
1273 		vm_pageout_page_queue(&vm_page_queue_active, vm_page_active_count, false);
1274 
1275 		if (VM_CONFIG_SWAP_IS_PRESENT) {
1276 			vm_consider_swapping();
1277 		}
1278 
1279 		vm_page_lock_queues();
1280 		vm_pageout_anonymous_pages_active = FALSE;
1281 		vm_page_unlock_queues();
1282 	}
1283 }
1284 
1285 
1286 size_t
vm_pageout_page_queue(vm_page_queue_head_t * q,size_t qcount,bool perf_test)1287 vm_pageout_page_queue(vm_page_queue_head_t *q, size_t qcount, bool perf_test)
1288 {
1289 	vm_page_t       m;
1290 	vm_object_t     t_object = NULL;
1291 	vm_object_t     l_object = NULL;
1292 	vm_object_t     m_object = NULL;
1293 	int             delayed_unlock = 0;
1294 	int             try_failed_count = 0;
1295 	int             refmod_state;
1296 	int             pmap_options;
1297 	struct          vm_pageout_queue *iq;
1298 	ppnum_t         phys_page;
1299 	size_t          pages_moved = 0;
1300 
1301 
1302 	iq = &vm_pageout_queue_internal;
1303 
1304 	vm_page_lock_queues();
1305 
1306 #if DEVELOPMENT || DEBUG
1307 	if (perf_test) {
1308 		iq = &vm_pageout_queue_benchmark;
1309 		// ensure the benchmark queue isn't throttled
1310 		iq->pgo_maxlaundry = (unsigned int) qcount;
1311 	}
1312 #endif /* DEVELOPMENT ||DEBUG */
1313 
1314 	while (qcount && !vm_page_queue_empty(q)) {
1315 		LCK_MTX_ASSERT(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED);
1316 
1317 		if (VM_PAGE_Q_THROTTLED(iq)) {
1318 			if (l_object != NULL) {
1319 				vm_object_unlock(l_object);
1320 				l_object = NULL;
1321 			}
1322 			iq->pgo_draining = TRUE;
1323 
1324 			assert_wait((event_t) (&iq->pgo_laundry + 1), THREAD_INTERRUPTIBLE);
1325 			vm_page_unlock_queues();
1326 
1327 			thread_block(THREAD_CONTINUE_NULL);
1328 
1329 			vm_page_lock_queues();
1330 			delayed_unlock = 0;
1331 			continue;
1332 		}
1333 		m = (vm_page_t) vm_page_queue_first(q);
1334 		m_object = VM_PAGE_OBJECT(m);
1335 
1336 		/*
1337 		 * check to see if we currently are working
1338 		 * with the same object... if so, we've
1339 		 * already got the lock
1340 		 */
1341 		if (m_object != l_object) {
1342 			if (!m_object->internal) {
1343 				goto reenter_pg_on_q;
1344 			}
1345 
1346 			/*
1347 			 * the object associated with candidate page is
1348 			 * different from the one we were just working
1349 			 * with... dump the lock if we still own it
1350 			 */
1351 			if (l_object != NULL) {
1352 				vm_object_unlock(l_object);
1353 				l_object = NULL;
1354 			}
1355 			if (m_object != t_object) {
1356 				try_failed_count = 0;
1357 			}
1358 
1359 			/*
1360 			 * Try to lock object; since we've alread got the
1361 			 * page queues lock, we can only 'try' for this one.
1362 			 * if the 'try' fails, we need to do a mutex_pause
1363 			 * to allow the owner of the object lock a chance to
1364 			 * run...
1365 			 */
1366 			if (!vm_object_lock_try_scan(m_object)) {
1367 				if (try_failed_count > 20) {
1368 					goto reenter_pg_on_q;
1369 				}
1370 				vm_page_unlock_queues();
1371 				mutex_pause(try_failed_count++);
1372 				vm_page_lock_queues();
1373 				delayed_unlock = 0;
1374 
1375 				t_object = m_object;
1376 				continue;
1377 			}
1378 			l_object = m_object;
1379 		}
1380 		if (!m_object->alive || m->vmp_cleaning || m->vmp_laundry || m->vmp_busy || m->vmp_absent || VMP_ERROR_GET(m) || m->vmp_free_when_done) {
1381 			/*
1382 			 * page is not to be cleaned
1383 			 * put it back on the head of its queue
1384 			 */
1385 			goto reenter_pg_on_q;
1386 		}
1387 		phys_page = VM_PAGE_GET_PHYS_PAGE(m);
1388 
1389 		if (m->vmp_reference == FALSE && m->vmp_pmapped == TRUE) {
1390 			refmod_state = pmap_get_refmod(phys_page);
1391 
1392 			if (refmod_state & VM_MEM_REFERENCED) {
1393 				m->vmp_reference = TRUE;
1394 			}
1395 			if (refmod_state & VM_MEM_MODIFIED) {
1396 				SET_PAGE_DIRTY(m, FALSE);
1397 			}
1398 		}
1399 		if (m->vmp_reference == TRUE) {
1400 			m->vmp_reference = FALSE;
1401 			pmap_clear_refmod_options(phys_page, VM_MEM_REFERENCED, PMAP_OPTIONS_NOFLUSH, (void *)NULL);
1402 			goto reenter_pg_on_q;
1403 		}
1404 		if (m->vmp_pmapped == TRUE) {
1405 			if (m->vmp_dirty || m->vmp_precious) {
1406 				pmap_options = PMAP_OPTIONS_COMPRESSOR;
1407 			} else {
1408 				pmap_options = PMAP_OPTIONS_COMPRESSOR_IFF_MODIFIED;
1409 			}
1410 			refmod_state = pmap_disconnect_options(phys_page, pmap_options, NULL);
1411 			if (refmod_state & VM_MEM_MODIFIED) {
1412 				SET_PAGE_DIRTY(m, FALSE);
1413 			}
1414 		}
1415 
1416 		if (!m->vmp_dirty && !m->vmp_precious) {
1417 			vm_page_unlock_queues();
1418 			VM_PAGE_FREE(m);
1419 			vm_page_lock_queues();
1420 			delayed_unlock = 0;
1421 
1422 			goto next_pg;
1423 		}
1424 		if (!m_object->pager_initialized || m_object->pager == MEMORY_OBJECT_NULL) {
1425 			if (!m_object->pager_initialized) {
1426 				vm_page_unlock_queues();
1427 
1428 				vm_object_collapse(m_object, (vm_object_offset_t) 0, TRUE);
1429 
1430 				if (!m_object->pager_initialized) {
1431 					vm_object_compressor_pager_create(m_object);
1432 				}
1433 
1434 				vm_page_lock_queues();
1435 				delayed_unlock = 0;
1436 			}
1437 			if (!m_object->pager_initialized || m_object->pager == MEMORY_OBJECT_NULL) {
1438 				goto reenter_pg_on_q;
1439 			}
1440 			/*
1441 			 * vm_object_compressor_pager_create will drop the object lock
1442 			 * which means 'm' may no longer be valid to use
1443 			 */
1444 			continue;
1445 		}
1446 
1447 		if (!perf_test) {
1448 			/*
1449 			 * we've already factored out pages in the laundry which
1450 			 * means this page can't be on the pageout queue so it's
1451 			 * safe to do the vm_page_queues_remove
1452 			 */
1453 			bool donate = (m->vmp_on_specialq == VM_PAGE_SPECIAL_Q_DONATE);
1454 			vm_page_queues_remove(m, TRUE);
1455 			if (donate) {
1456 				/*
1457 				 * The compressor needs to see this bit to know
1458 				 * where this page needs to land. Also if stolen,
1459 				 * this bit helps put the page back in the right
1460 				 * special queue where it belongs.
1461 				 */
1462 				m->vmp_on_specialq = VM_PAGE_SPECIAL_Q_DONATE;
1463 			}
1464 		} else {
1465 			vm_page_queue_remove(q, m, vmp_pageq);
1466 		}
1467 
1468 		LCK_MTX_ASSERT(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED);
1469 
1470 		vm_pageout_cluster_to_queue(m, iq);
1471 
1472 		pages_moved++;
1473 		goto next_pg;
1474 
1475 reenter_pg_on_q:
1476 		vm_page_queue_remove(q, m, vmp_pageq);
1477 		vm_page_queue_enter(q, m, vmp_pageq);
1478 next_pg:
1479 		qcount--;
1480 		try_failed_count = 0;
1481 
1482 		if (delayed_unlock++ > 128) {
1483 			if (l_object != NULL) {
1484 				vm_object_unlock(l_object);
1485 				l_object = NULL;
1486 			}
1487 			lck_mtx_yield(&vm_page_queue_lock);
1488 			delayed_unlock = 0;
1489 		}
1490 	}
1491 	if (l_object != NULL) {
1492 		vm_object_unlock(l_object);
1493 		l_object = NULL;
1494 	}
1495 	vm_page_unlock_queues();
1496 	return pages_moved;
1497 }
1498 
1499 
1500 
1501 /*
1502  * function in BSD to apply I/O throttle to the pageout thread
1503  */
1504 extern void vm_pageout_io_throttle(void);
1505 
1506 #define VM_PAGEOUT_SCAN_HANDLE_REUSABLE_PAGE(m, obj)                    \
1507 	MACRO_BEGIN                                                     \
1508 	/* \
1509 	 * If a "reusable" page somehow made it back into \
1510 	 * the active queue, it's been re-used and is not \
1511 	 * quite re-usable. \
1512 	 * If the VM object was "all_reusable", consider it \
1513 	 * as "all re-used" instead of converting it to \
1514 	 * "partially re-used", which could be expensive. \
1515 	 */                                                             \
1516 	assert(VM_PAGE_OBJECT((m)) == (obj));                           \
1517 	if ((m)->vmp_reusable ||                                        \
1518 	    (obj)->all_reusable) {                                      \
1519 	        vm_object_reuse_pages((obj),                            \
1520 	                              (m)->vmp_offset,                  \
1521 	                              (m)->vmp_offset + PAGE_SIZE_64,   \
1522 	                              FALSE);                           \
1523 	}                                                               \
1524 	MACRO_END
1525 
1526 
1527 #define VM_PAGEOUT_DELAYED_UNLOCK_LIMIT         64
1528 #define VM_PAGEOUT_DELAYED_UNLOCK_LIMIT_MAX     1024
1529 
1530 #define FCS_IDLE                0
1531 #define FCS_DELAYED             1
1532 #define FCS_DEADLOCK_DETECTED   2
1533 
1534 struct flow_control {
1535 	int             state;
1536 	mach_timespec_t ts;
1537 };
1538 
1539 
1540 uint64_t vm_pageout_rejected_bq_internal = 0;
1541 uint64_t vm_pageout_rejected_bq_external = 0;
1542 uint64_t vm_pageout_skipped_bq_internal = 0;
1543 uint64_t vm_pageout_skipped_bq_external = 0;
1544 
1545 #define ANONS_GRABBED_LIMIT     2
1546 
1547 
1548 #if 0
1549 static void vm_pageout_delayed_unlock(int *, int *, vm_page_t *);
1550 #endif
1551 static void vm_pageout_prepare_to_block(vm_object_t *, int *, vm_page_t *, int *, int);
1552 
1553 #define VM_PAGEOUT_PB_NO_ACTION                         0
1554 #define VM_PAGEOUT_PB_CONSIDER_WAKING_COMPACTOR_SWAPPER 1
1555 #define VM_PAGEOUT_PB_THREAD_YIELD                      2
1556 
1557 
1558 #if 0
1559 static void
1560 vm_pageout_delayed_unlock(int *delayed_unlock, int *local_freed, vm_page_t *local_freeq)
1561 {
1562 	if (*local_freeq) {
1563 		vm_page_unlock_queues();
1564 
1565 		VM_DEBUG_CONSTANT_EVENT(
1566 			vm_pageout_freelist, VM_PAGEOUT_FREELIST, DBG_FUNC_START,
1567 			vm_page_free_count, 0, 0, 1);
1568 
1569 		vm_page_free_list(*local_freeq, TRUE);
1570 
1571 		VM_DEBUG_CONSTANT_EVENT(vm_pageout_freelist, VM_PAGEOUT_FREELIST, DBG_FUNC_END,
1572 		    vm_page_free_count, *local_freed, 0, 1);
1573 
1574 		*local_freeq = NULL;
1575 		*local_freed = 0;
1576 
1577 		vm_page_lock_queues();
1578 	} else {
1579 		lck_mtx_yield(&vm_page_queue_lock);
1580 	}
1581 	*delayed_unlock = 1;
1582 }
1583 #endif
1584 
1585 
1586 static void
vm_pageout_prepare_to_block(vm_object_t * object,int * delayed_unlock,vm_page_t * local_freeq,int * local_freed,int action)1587 vm_pageout_prepare_to_block(vm_object_t *object, int *delayed_unlock,
1588     vm_page_t *local_freeq, int *local_freed, int action)
1589 {
1590 	vm_page_unlock_queues();
1591 
1592 	if (*object != NULL) {
1593 		vm_object_unlock(*object);
1594 		*object = NULL;
1595 	}
1596 	if (*local_freeq) {
1597 		vm_page_free_list(*local_freeq, TRUE);
1598 
1599 		*local_freeq = NULL;
1600 		*local_freed = 0;
1601 	}
1602 	*delayed_unlock = 1;
1603 
1604 	switch (action) {
1605 	case VM_PAGEOUT_PB_CONSIDER_WAKING_COMPACTOR_SWAPPER:
1606 		vm_consider_waking_compactor_swapper();
1607 		break;
1608 	case VM_PAGEOUT_PB_THREAD_YIELD:
1609 		thread_yield_internal(1);
1610 		break;
1611 	case VM_PAGEOUT_PB_NO_ACTION:
1612 	default:
1613 		break;
1614 	}
1615 	vm_page_lock_queues();
1616 }
1617 
1618 
1619 static struct vm_pageout_vminfo last;
1620 
1621 uint64_t last_vm_page_pages_grabbed = 0;
1622 
1623 extern  uint32_t c_segment_pages_compressed;
1624 
1625 extern uint64_t shared_region_pager_reclaimed;
1626 extern struct memory_object_pager_ops shared_region_pager_ops;
1627 
1628 void
update_vm_info(void)1629 update_vm_info(void)
1630 {
1631 	unsigned long tmp;
1632 	uint64_t tmp64;
1633 
1634 	vm_pageout_stats[vm_pageout_stat_now].vm_page_active_count = vm_page_active_count;
1635 	vm_pageout_stats[vm_pageout_stat_now].vm_page_speculative_count = vm_page_speculative_count;
1636 	vm_pageout_stats[vm_pageout_stat_now].vm_page_inactive_count = vm_page_inactive_count;
1637 	vm_pageout_stats[vm_pageout_stat_now].vm_page_anonymous_count = vm_page_anonymous_count;
1638 
1639 	vm_pageout_stats[vm_pageout_stat_now].vm_page_free_count = vm_page_free_count;
1640 	vm_pageout_stats[vm_pageout_stat_now].vm_page_wire_count = vm_page_wire_count;
1641 	vm_pageout_stats[vm_pageout_stat_now].vm_page_compressor_count = VM_PAGE_COMPRESSOR_COUNT;
1642 
1643 	vm_pageout_stats[vm_pageout_stat_now].vm_page_pages_compressed = c_segment_pages_compressed;
1644 	vm_pageout_stats[vm_pageout_stat_now].vm_page_pageable_internal_count = vm_page_pageable_internal_count;
1645 	vm_pageout_stats[vm_pageout_stat_now].vm_page_pageable_external_count = vm_page_pageable_external_count;
1646 	vm_pageout_stats[vm_pageout_stat_now].vm_page_xpmapped_external_count = vm_page_xpmapped_external_count;
1647 	vm_pageout_stats[vm_pageout_stat_now].vm_page_realtime_count = vm_page_realtime_count;
1648 
1649 	tmp = vm_pageout_vminfo.vm_pageout_considered_page;
1650 	vm_pageout_stats[vm_pageout_stat_now].considered = (unsigned int)(tmp - last.vm_pageout_considered_page);
1651 	last.vm_pageout_considered_page = tmp;
1652 
1653 	tmp64 = vm_pageout_vminfo.vm_pageout_compressions;
1654 	vm_pageout_stats[vm_pageout_stat_now].pages_compressed = (unsigned int)(tmp64 - last.vm_pageout_compressions);
1655 	last.vm_pageout_compressions = tmp64;
1656 
1657 	tmp = vm_pageout_vminfo.vm_compressor_failed;
1658 	vm_pageout_stats[vm_pageout_stat_now].failed_compressions = (unsigned int)(tmp - last.vm_compressor_failed);
1659 	last.vm_compressor_failed = tmp;
1660 
1661 	tmp64 = vm_pageout_vminfo.vm_compressor_pages_grabbed;
1662 	vm_pageout_stats[vm_pageout_stat_now].pages_grabbed_by_compressor = (unsigned int)(tmp64 - last.vm_compressor_pages_grabbed);
1663 	last.vm_compressor_pages_grabbed = tmp64;
1664 
1665 	tmp = vm_pageout_vminfo.vm_phantom_cache_found_ghost;
1666 	vm_pageout_stats[vm_pageout_stat_now].phantom_ghosts_found = (unsigned int)(tmp - last.vm_phantom_cache_found_ghost);
1667 	last.vm_phantom_cache_found_ghost = tmp;
1668 
1669 	tmp = vm_pageout_vminfo.vm_phantom_cache_added_ghost;
1670 	vm_pageout_stats[vm_pageout_stat_now].phantom_ghosts_added = (unsigned int)(tmp - last.vm_phantom_cache_added_ghost);
1671 	last.vm_phantom_cache_added_ghost = tmp;
1672 
1673 	tmp64 = counter_load(&vm_page_grab_count);
1674 	vm_pageout_stats[vm_pageout_stat_now].pages_grabbed = (unsigned int)(tmp64 - last_vm_page_pages_grabbed);
1675 	last_vm_page_pages_grabbed = tmp64;
1676 
1677 	tmp = vm_pageout_vminfo.vm_page_pages_freed;
1678 	vm_pageout_stats[vm_pageout_stat_now].pages_freed = (unsigned int)(tmp - last.vm_page_pages_freed);
1679 	last.vm_page_pages_freed = tmp;
1680 
1681 	if (vm_pageout_stats[vm_pageout_stat_now].considered) {
1682 		tmp = vm_pageout_vminfo.vm_pageout_pages_evicted;
1683 		vm_pageout_stats[vm_pageout_stat_now].pages_evicted = (unsigned int)(tmp - last.vm_pageout_pages_evicted);
1684 		last.vm_pageout_pages_evicted = tmp;
1685 
1686 		tmp = vm_pageout_vminfo.vm_pageout_pages_purged;
1687 		vm_pageout_stats[vm_pageout_stat_now].pages_purged = (unsigned int)(tmp - last.vm_pageout_pages_purged);
1688 		last.vm_pageout_pages_purged = tmp;
1689 
1690 		tmp = vm_pageout_vminfo.vm_pageout_freed_speculative;
1691 		vm_pageout_stats[vm_pageout_stat_now].freed_speculative = (unsigned int)(tmp - last.vm_pageout_freed_speculative);
1692 		last.vm_pageout_freed_speculative = tmp;
1693 
1694 		tmp = vm_pageout_vminfo.vm_pageout_freed_external;
1695 		vm_pageout_stats[vm_pageout_stat_now].freed_external = (unsigned int)(tmp - last.vm_pageout_freed_external);
1696 		last.vm_pageout_freed_external = tmp;
1697 
1698 		tmp = vm_pageout_vminfo.vm_pageout_inactive_referenced;
1699 		vm_pageout_stats[vm_pageout_stat_now].inactive_referenced = (unsigned int)(tmp - last.vm_pageout_inactive_referenced);
1700 		last.vm_pageout_inactive_referenced = tmp;
1701 
1702 		tmp = vm_pageout_vminfo.vm_pageout_scan_inactive_throttled_external;
1703 		vm_pageout_stats[vm_pageout_stat_now].throttled_external_q = (unsigned int)(tmp - last.vm_pageout_scan_inactive_throttled_external);
1704 		last.vm_pageout_scan_inactive_throttled_external = tmp;
1705 
1706 		tmp = vm_pageout_vminfo.vm_pageout_inactive_dirty_external;
1707 		vm_pageout_stats[vm_pageout_stat_now].cleaned_dirty_external = (unsigned int)(tmp - last.vm_pageout_inactive_dirty_external);
1708 		last.vm_pageout_inactive_dirty_external = tmp;
1709 
1710 		tmp = vm_pageout_vminfo.vm_pageout_freed_cleaned;
1711 		vm_pageout_stats[vm_pageout_stat_now].freed_cleaned = (unsigned int)(tmp - last.vm_pageout_freed_cleaned);
1712 		last.vm_pageout_freed_cleaned = tmp;
1713 
1714 		tmp = vm_pageout_vminfo.vm_pageout_inactive_nolock;
1715 		vm_pageout_stats[vm_pageout_stat_now].inactive_nolock = (unsigned int)(tmp - last.vm_pageout_inactive_nolock);
1716 		last.vm_pageout_inactive_nolock = tmp;
1717 
1718 		tmp = vm_pageout_vminfo.vm_pageout_scan_inactive_throttled_internal;
1719 		vm_pageout_stats[vm_pageout_stat_now].throttled_internal_q = (unsigned int)(tmp - last.vm_pageout_scan_inactive_throttled_internal);
1720 		last.vm_pageout_scan_inactive_throttled_internal = tmp;
1721 
1722 		tmp = vm_pageout_vminfo.vm_pageout_skipped_external;
1723 		vm_pageout_stats[vm_pageout_stat_now].skipped_external = (unsigned int)(tmp - last.vm_pageout_skipped_external);
1724 		last.vm_pageout_skipped_external = tmp;
1725 
1726 		tmp = vm_pageout_vminfo.vm_pageout_skipped_internal;
1727 		vm_pageout_stats[vm_pageout_stat_now].skipped_internal = (unsigned int)(tmp - last.vm_pageout_skipped_internal);
1728 		last.vm_pageout_skipped_internal = tmp;
1729 
1730 		tmp = vm_pageout_vminfo.vm_pageout_reactivation_limit_exceeded;
1731 		vm_pageout_stats[vm_pageout_stat_now].reactivation_limit_exceeded = (unsigned int)(tmp - last.vm_pageout_reactivation_limit_exceeded);
1732 		last.vm_pageout_reactivation_limit_exceeded = tmp;
1733 
1734 		tmp = vm_pageout_vminfo.vm_pageout_inactive_force_reclaim;
1735 		vm_pageout_stats[vm_pageout_stat_now].forced_inactive_reclaim = (unsigned int)(tmp - last.vm_pageout_inactive_force_reclaim);
1736 		last.vm_pageout_inactive_force_reclaim = tmp;
1737 
1738 		tmp = vm_pageout_vminfo.vm_pageout_freed_internal;
1739 		vm_pageout_stats[vm_pageout_stat_now].freed_internal = (unsigned int)(tmp - last.vm_pageout_freed_internal);
1740 		last.vm_pageout_freed_internal = tmp;
1741 
1742 		tmp = vm_pageout_vminfo.vm_pageout_considered_bq_internal;
1743 		vm_pageout_stats[vm_pageout_stat_now].considered_bq_internal = (unsigned int)(tmp - last.vm_pageout_considered_bq_internal);
1744 		last.vm_pageout_considered_bq_internal = tmp;
1745 
1746 		tmp = vm_pageout_vminfo.vm_pageout_considered_bq_external;
1747 		vm_pageout_stats[vm_pageout_stat_now].considered_bq_external = (unsigned int)(tmp - last.vm_pageout_considered_bq_external);
1748 		last.vm_pageout_considered_bq_external = tmp;
1749 
1750 		tmp = vm_pageout_vminfo.vm_pageout_filecache_min_reactivated;
1751 		vm_pageout_stats[vm_pageout_stat_now].filecache_min_reactivations = (unsigned int)(tmp - last.vm_pageout_filecache_min_reactivated);
1752 		last.vm_pageout_filecache_min_reactivated = tmp;
1753 
1754 		tmp = vm_pageout_vminfo.vm_pageout_inactive_dirty_internal;
1755 		vm_pageout_stats[vm_pageout_stat_now].cleaned_dirty_internal = (unsigned int)(tmp - last.vm_pageout_inactive_dirty_internal);
1756 		last.vm_pageout_inactive_dirty_internal = tmp;
1757 
1758 		tmp = vm_pageout_vminfo.vm_pageout_forcereclaimed_sharedcache;
1759 		vm_pageout_stats[vm_pageout_stat_now].forcereclaimed_sharedcache = (unsigned int)(tmp - last.vm_pageout_forcereclaimed_sharedcache);
1760 		last.vm_pageout_forcereclaimed_sharedcache = tmp;
1761 
1762 		tmp = vm_pageout_vminfo.vm_pageout_forcereclaimed_realtime;
1763 		vm_pageout_stats[vm_pageout_stat_now].forcereclaimed_realtime = (unsigned int)(tmp - last.vm_pageout_forcereclaimed_realtime);
1764 		last.vm_pageout_forcereclaimed_realtime = tmp;
1765 
1766 		tmp = vm_pageout_vminfo.vm_pageout_protected_sharedcache;
1767 		vm_pageout_stats[vm_pageout_stat_now].protected_sharedcache = (unsigned int)(tmp - last.vm_pageout_protected_sharedcache);
1768 		last.vm_pageout_protected_sharedcache = tmp;
1769 
1770 		tmp = vm_pageout_vminfo.vm_pageout_protected_realtime;
1771 		vm_pageout_stats[vm_pageout_stat_now].protected_realtime = (unsigned int)(tmp - last.vm_pageout_protected_realtime);
1772 		last.vm_pageout_protected_realtime = tmp;
1773 	}
1774 
1775 	KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_INFO1)) | DBG_FUNC_NONE,
1776 	    vm_pageout_stats[vm_pageout_stat_now].vm_page_active_count,
1777 	    vm_pageout_stats[vm_pageout_stat_now].vm_page_speculative_count,
1778 	    vm_pageout_stats[vm_pageout_stat_now].vm_page_inactive_count,
1779 	    vm_pageout_stats[vm_pageout_stat_now].vm_page_anonymous_count,
1780 	    0);
1781 
1782 	KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_INFO2)) | DBG_FUNC_NONE,
1783 	    vm_pageout_stats[vm_pageout_stat_now].vm_page_free_count,
1784 	    vm_pageout_stats[vm_pageout_stat_now].vm_page_wire_count,
1785 	    vm_pageout_stats[vm_pageout_stat_now].vm_page_compressor_count,
1786 	    0,
1787 	    0);
1788 
1789 	KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_INFO3)) | DBG_FUNC_NONE,
1790 	    vm_pageout_stats[vm_pageout_stat_now].vm_page_pages_compressed,
1791 	    vm_pageout_stats[vm_pageout_stat_now].vm_page_pageable_internal_count,
1792 	    vm_pageout_stats[vm_pageout_stat_now].vm_page_pageable_external_count,
1793 	    vm_pageout_stats[vm_pageout_stat_now].vm_page_xpmapped_external_count,
1794 	    0);
1795 
1796 	if (vm_pageout_stats[vm_pageout_stat_now].considered ||
1797 	    vm_pageout_stats[vm_pageout_stat_now].pages_compressed ||
1798 	    vm_pageout_stats[vm_pageout_stat_now].failed_compressions) {
1799 		KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_INFO4)) | DBG_FUNC_NONE,
1800 		    vm_pageout_stats[vm_pageout_stat_now].considered,
1801 		    vm_pageout_stats[vm_pageout_stat_now].freed_speculative,
1802 		    vm_pageout_stats[vm_pageout_stat_now].freed_external,
1803 		    vm_pageout_stats[vm_pageout_stat_now].inactive_referenced,
1804 		    0);
1805 
1806 		KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_INFO5)) | DBG_FUNC_NONE,
1807 		    vm_pageout_stats[vm_pageout_stat_now].throttled_external_q,
1808 		    vm_pageout_stats[vm_pageout_stat_now].cleaned_dirty_external,
1809 		    vm_pageout_stats[vm_pageout_stat_now].freed_cleaned,
1810 		    vm_pageout_stats[vm_pageout_stat_now].inactive_nolock,
1811 		    0);
1812 
1813 		KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_INFO6)) | DBG_FUNC_NONE,
1814 		    vm_pageout_stats[vm_pageout_stat_now].throttled_internal_q,
1815 		    vm_pageout_stats[vm_pageout_stat_now].pages_compressed,
1816 		    vm_pageout_stats[vm_pageout_stat_now].pages_grabbed_by_compressor,
1817 		    vm_pageout_stats[vm_pageout_stat_now].skipped_external,
1818 		    0);
1819 
1820 		KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_INFO7)) | DBG_FUNC_NONE,
1821 		    vm_pageout_stats[vm_pageout_stat_now].reactivation_limit_exceeded,
1822 		    vm_pageout_stats[vm_pageout_stat_now].forced_inactive_reclaim,
1823 		    vm_pageout_stats[vm_pageout_stat_now].failed_compressions,
1824 		    vm_pageout_stats[vm_pageout_stat_now].freed_internal,
1825 		    0);
1826 
1827 		KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_INFO8)) | DBG_FUNC_NONE,
1828 		    vm_pageout_stats[vm_pageout_stat_now].considered_bq_internal,
1829 		    vm_pageout_stats[vm_pageout_stat_now].considered_bq_external,
1830 		    vm_pageout_stats[vm_pageout_stat_now].filecache_min_reactivations,
1831 		    vm_pageout_stats[vm_pageout_stat_now].cleaned_dirty_internal,
1832 		    0);
1833 
1834 		KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_INFO10)) | DBG_FUNC_NONE,
1835 		    vm_pageout_stats[vm_pageout_stat_now].forcereclaimed_sharedcache,
1836 		    vm_pageout_stats[vm_pageout_stat_now].forcereclaimed_realtime,
1837 		    vm_pageout_stats[vm_pageout_stat_now].protected_sharedcache,
1838 		    vm_pageout_stats[vm_pageout_stat_now].protected_realtime,
1839 		    0);
1840 	}
1841 	KERNEL_DEBUG_CONSTANT((MACHDBG_CODE(DBG_MACH_VM, VM_INFO9)) | DBG_FUNC_NONE,
1842 	    vm_pageout_stats[vm_pageout_stat_now].pages_grabbed,
1843 	    vm_pageout_stats[vm_pageout_stat_now].pages_freed,
1844 	    vm_pageout_stats[vm_pageout_stat_now].phantom_ghosts_found,
1845 	    vm_pageout_stats[vm_pageout_stat_now].phantom_ghosts_added,
1846 	    0);
1847 
1848 	record_memory_pressure();
1849 }
1850 
1851 extern boolean_t hibernation_vmqueues_inspection;
1852 
1853 /*
1854  * Return values for functions called by vm_pageout_scan
1855  * that control its flow.
1856  *
1857  * PROCEED -- vm_pageout_scan will keep making forward progress.
1858  * DONE_RETURN -- page demand satisfied, work is done -> vm_pageout_scan returns.
1859  * NEXT_ITERATION -- restart the 'for' loop in vm_pageout_scan aka continue.
1860  */
1861 
1862 #define VM_PAGEOUT_SCAN_PROCEED                 (0)
1863 #define VM_PAGEOUT_SCAN_DONE_RETURN             (1)
1864 #define VM_PAGEOUT_SCAN_NEXT_ITERATION          (2)
1865 
1866 /*
1867  * This function is called only from vm_pageout_scan and
1868  * it moves overflow secluded pages (one-at-a-time) to the
1869  * batched 'local' free Q or active Q.
1870  */
1871 static void
vps_deal_with_secluded_page_overflow(vm_page_t * local_freeq,int * local_freed)1872 vps_deal_with_secluded_page_overflow(vm_page_t *local_freeq, int *local_freed)
1873 {
1874 #if CONFIG_SECLUDED_MEMORY
1875 	/*
1876 	 * Deal with secluded_q overflow.
1877 	 */
1878 	if (vm_page_secluded_count > vm_page_secluded_target) {
1879 		vm_page_t secluded_page;
1880 
1881 		/*
1882 		 * SECLUDED_AGING_BEFORE_ACTIVE:
1883 		 * Excess secluded pages go to the active queue and
1884 		 * will later go to the inactive queue.
1885 		 */
1886 		assert((vm_page_secluded_count_free +
1887 		    vm_page_secluded_count_inuse) ==
1888 		    vm_page_secluded_count);
1889 		secluded_page = (vm_page_t)vm_page_queue_first(&vm_page_queue_secluded);
1890 		assert(secluded_page->vmp_q_state == VM_PAGE_ON_SECLUDED_Q);
1891 
1892 		vm_page_queues_remove(secluded_page, FALSE);
1893 		assert(!secluded_page->vmp_fictitious);
1894 		assert(!VM_PAGE_WIRED(secluded_page));
1895 
1896 		if (secluded_page->vmp_object == 0) {
1897 			/* transfer to free queue */
1898 			assert(secluded_page->vmp_busy);
1899 			secluded_page->vmp_snext = *local_freeq;
1900 			*local_freeq = secluded_page;
1901 			*local_freed += 1;
1902 		} else {
1903 			/* transfer to head of active queue */
1904 			vm_page_enqueue_active(secluded_page, FALSE);
1905 			secluded_page = VM_PAGE_NULL;
1906 		}
1907 	}
1908 #else /* CONFIG_SECLUDED_MEMORY */
1909 
1910 #pragma unused(local_freeq)
1911 #pragma unused(local_freed)
1912 
1913 	return;
1914 
1915 #endif /* CONFIG_SECLUDED_MEMORY */
1916 }
1917 
1918 /*
1919  * This function is called only from vm_pageout_scan and
1920  * it initializes the loop targets for vm_pageout_scan().
1921  */
1922 static void
vps_init_page_targets(void)1923 vps_init_page_targets(void)
1924 {
1925 	/*
1926 	 * LD TODO: Other page targets should be calculated here too.
1927 	 */
1928 	vm_page_anonymous_min = vm_page_inactive_target / 20;
1929 
1930 	if (vm_pageout_state.vm_page_speculative_percentage > 50) {
1931 		vm_pageout_state.vm_page_speculative_percentage = 50;
1932 	} else if (vm_pageout_state.vm_page_speculative_percentage <= 0) {
1933 		vm_pageout_state.vm_page_speculative_percentage = 1;
1934 	}
1935 
1936 	vm_pageout_state.vm_page_speculative_target = VM_PAGE_SPECULATIVE_TARGET(vm_page_active_count +
1937 	    vm_page_inactive_count);
1938 }
1939 
1940 /*
1941  * This function is called only from vm_pageout_scan and
1942  * it purges a single VM object at-a-time and will either
1943  * make vm_pageout_scan() restart the loop or keeping moving forward.
1944  */
1945 static int
vps_purge_object()1946 vps_purge_object()
1947 {
1948 	int             force_purge;
1949 
1950 	assert(available_for_purge >= 0);
1951 	force_purge = 0; /* no force-purging */
1952 
1953 #if VM_PRESSURE_EVENTS
1954 	vm_pressure_level_t pressure_level;
1955 
1956 	pressure_level = memorystatus_vm_pressure_level;
1957 
1958 	if (pressure_level > kVMPressureNormal) {
1959 		if (pressure_level >= kVMPressureCritical) {
1960 			force_purge = vm_pageout_state.memorystatus_purge_on_critical;
1961 		} else if (pressure_level >= kVMPressureUrgent) {
1962 			force_purge = vm_pageout_state.memorystatus_purge_on_urgent;
1963 		} else if (pressure_level >= kVMPressureWarning) {
1964 			force_purge = vm_pageout_state.memorystatus_purge_on_warning;
1965 		}
1966 	}
1967 #endif /* VM_PRESSURE_EVENTS */
1968 
1969 	if (available_for_purge || force_purge) {
1970 		memoryshot(VM_PAGEOUT_PURGEONE, DBG_FUNC_START);
1971 
1972 		VM_DEBUG_EVENT(vm_pageout_purgeone, VM_PAGEOUT_PURGEONE, DBG_FUNC_START, vm_page_free_count, 0, 0, 0);
1973 		if (vm_purgeable_object_purge_one(force_purge, C_DONT_BLOCK)) {
1974 			VM_PAGEOUT_DEBUG(vm_pageout_purged_objects, 1);
1975 			VM_DEBUG_EVENT(vm_pageout_purgeone, VM_PAGEOUT_PURGEONE, DBG_FUNC_END, vm_page_free_count, 0, 0, 0);
1976 			memoryshot(VM_PAGEOUT_PURGEONE, DBG_FUNC_END);
1977 
1978 			return VM_PAGEOUT_SCAN_NEXT_ITERATION;
1979 		}
1980 		VM_DEBUG_EVENT(vm_pageout_purgeone, VM_PAGEOUT_PURGEONE, DBG_FUNC_END, 0, 0, 0, -1);
1981 		memoryshot(VM_PAGEOUT_PURGEONE, DBG_FUNC_END);
1982 	}
1983 
1984 	return VM_PAGEOUT_SCAN_PROCEED;
1985 }
1986 
1987 /*
1988  * This function is called only from vm_pageout_scan and
1989  * it will try to age the next speculative Q if the oldest
1990  * one is empty.
1991  */
1992 static int
vps_age_speculative_queue(boolean_t force_speculative_aging)1993 vps_age_speculative_queue(boolean_t force_speculative_aging)
1994 {
1995 #define DELAY_SPECULATIVE_AGE   1000
1996 
1997 	/*
1998 	 * try to pull pages from the aging bins...
1999 	 * see vm_page.h for an explanation of how
2000 	 * this mechanism works
2001 	 */
2002 	boolean_t                       can_steal = FALSE;
2003 	int                             num_scanned_queues;
2004 	static int                      delay_speculative_age = 0; /* depends the # of times we go through the main pageout_scan loop.*/
2005 	mach_timespec_t                 ts;
2006 	struct vm_speculative_age_q     *aq;
2007 	struct vm_speculative_age_q     *sq;
2008 
2009 	sq = &vm_page_queue_speculative[VM_PAGE_SPECULATIVE_AGED_Q];
2010 
2011 	aq = &vm_page_queue_speculative[speculative_steal_index];
2012 
2013 	num_scanned_queues = 0;
2014 	while (vm_page_queue_empty(&aq->age_q) &&
2015 	    num_scanned_queues++ != VM_PAGE_MAX_SPECULATIVE_AGE_Q) {
2016 		speculative_steal_index++;
2017 
2018 		if (speculative_steal_index > VM_PAGE_MAX_SPECULATIVE_AGE_Q) {
2019 			speculative_steal_index = VM_PAGE_MIN_SPECULATIVE_AGE_Q;
2020 		}
2021 
2022 		aq = &vm_page_queue_speculative[speculative_steal_index];
2023 	}
2024 
2025 	if (num_scanned_queues == VM_PAGE_MAX_SPECULATIVE_AGE_Q + 1) {
2026 		/*
2027 		 * XXX We've scanned all the speculative
2028 		 * queues but still haven't found one
2029 		 * that is not empty, even though
2030 		 * vm_page_speculative_count is not 0.
2031 		 */
2032 		if (!vm_page_queue_empty(&sq->age_q)) {
2033 			return VM_PAGEOUT_SCAN_NEXT_ITERATION;
2034 		}
2035 #if DEVELOPMENT || DEBUG
2036 		panic("vm_pageout_scan: vm_page_speculative_count=%d but queues are empty", vm_page_speculative_count);
2037 #endif
2038 		/* readjust... */
2039 		vm_page_speculative_count = 0;
2040 		/* ... and continue */
2041 		return VM_PAGEOUT_SCAN_NEXT_ITERATION;
2042 	}
2043 
2044 	if (vm_page_speculative_count > vm_pageout_state.vm_page_speculative_target || force_speculative_aging == TRUE) {
2045 		can_steal = TRUE;
2046 	} else {
2047 		if (!delay_speculative_age) {
2048 			mach_timespec_t ts_fully_aged;
2049 
2050 			ts_fully_aged.tv_sec = (VM_PAGE_MAX_SPECULATIVE_AGE_Q * vm_pageout_state.vm_page_speculative_q_age_ms) / 1000;
2051 			ts_fully_aged.tv_nsec = ((VM_PAGE_MAX_SPECULATIVE_AGE_Q * vm_pageout_state.vm_page_speculative_q_age_ms) % 1000)
2052 			    * 1000 * NSEC_PER_USEC;
2053 
2054 			ADD_MACH_TIMESPEC(&ts_fully_aged, &aq->age_ts);
2055 
2056 			clock_sec_t sec;
2057 			clock_nsec_t nsec;
2058 			clock_get_system_nanotime(&sec, &nsec);
2059 			ts.tv_sec = (unsigned int) sec;
2060 			ts.tv_nsec = nsec;
2061 
2062 			if (CMP_MACH_TIMESPEC(&ts, &ts_fully_aged) >= 0) {
2063 				can_steal = TRUE;
2064 			} else {
2065 				delay_speculative_age++;
2066 			}
2067 		} else {
2068 			delay_speculative_age++;
2069 			if (delay_speculative_age == DELAY_SPECULATIVE_AGE) {
2070 				delay_speculative_age = 0;
2071 			}
2072 		}
2073 	}
2074 	if (can_steal == TRUE) {
2075 		vm_page_speculate_ageit(aq);
2076 	}
2077 
2078 	return VM_PAGEOUT_SCAN_PROCEED;
2079 }
2080 
2081 /*
2082  * This function is called only from vm_pageout_scan and
2083  * it evicts a single VM object from the cache.
2084  */
2085 static int inline
vps_object_cache_evict(vm_object_t * object_to_unlock)2086 vps_object_cache_evict(vm_object_t *object_to_unlock)
2087 {
2088 	static int                      cache_evict_throttle = 0;
2089 	struct vm_speculative_age_q     *sq;
2090 
2091 	sq = &vm_page_queue_speculative[VM_PAGE_SPECULATIVE_AGED_Q];
2092 
2093 	if (vm_page_queue_empty(&sq->age_q) && cache_evict_throttle == 0) {
2094 		int     pages_evicted;
2095 
2096 		if (*object_to_unlock != NULL) {
2097 			vm_object_unlock(*object_to_unlock);
2098 			*object_to_unlock = NULL;
2099 		}
2100 		KERNEL_DEBUG_CONSTANT(0x13001ec | DBG_FUNC_START, 0, 0, 0, 0, 0);
2101 
2102 		pages_evicted = vm_object_cache_evict(100, 10);
2103 
2104 		KERNEL_DEBUG_CONSTANT(0x13001ec | DBG_FUNC_END, pages_evicted, 0, 0, 0, 0);
2105 
2106 		if (pages_evicted) {
2107 			vm_pageout_vminfo.vm_pageout_pages_evicted += pages_evicted;
2108 
2109 			VM_DEBUG_EVENT(vm_pageout_cache_evict, VM_PAGEOUT_CACHE_EVICT, DBG_FUNC_NONE,
2110 			    vm_page_free_count, pages_evicted, vm_pageout_vminfo.vm_pageout_pages_evicted, 0);
2111 			memoryshot(VM_PAGEOUT_CACHE_EVICT, DBG_FUNC_NONE);
2112 
2113 			/*
2114 			 * we just freed up to 100 pages,
2115 			 * so go back to the top of the main loop
2116 			 * and re-evaulate the memory situation
2117 			 */
2118 			return VM_PAGEOUT_SCAN_NEXT_ITERATION;
2119 		} else {
2120 			cache_evict_throttle = 1000;
2121 		}
2122 	}
2123 	if (cache_evict_throttle) {
2124 		cache_evict_throttle--;
2125 	}
2126 
2127 	return VM_PAGEOUT_SCAN_PROCEED;
2128 }
2129 
2130 
2131 /*
2132  * This function is called only from vm_pageout_scan and
2133  * it calculates the filecache min. that needs to be maintained
2134  * as we start to steal pages.
2135  */
2136 static void
vps_calculate_filecache_min(void)2137 vps_calculate_filecache_min(void)
2138 {
2139 	int divisor = vm_pageout_state.vm_page_filecache_min_divisor;
2140 
2141 #if CONFIG_JETSAM
2142 	/*
2143 	 * don't let the filecache_min fall below 15% of available memory
2144 	 * on systems with an active compressor that isn't nearing its
2145 	 * limits w/r to accepting new data
2146 	 *
2147 	 * on systems w/o the compressor/swapper, the filecache is always
2148 	 * a very large percentage of the AVAILABLE_NON_COMPRESSED_MEMORY
2149 	 * since most (if not all) of the anonymous pages are in the
2150 	 * throttled queue (which isn't counted as available) which
2151 	 * effectively disables this filter
2152 	 */
2153 	if (vm_compressor_low_on_space() || divisor == 0) {
2154 		vm_pageout_state.vm_page_filecache_min = 0;
2155 	} else {
2156 		vm_pageout_state.vm_page_filecache_min =
2157 		    ((AVAILABLE_NON_COMPRESSED_MEMORY) * 10) / divisor;
2158 	}
2159 #else
2160 	if (vm_compressor_out_of_space() || divisor == 0) {
2161 		vm_pageout_state.vm_page_filecache_min = 0;
2162 	} else {
2163 		/*
2164 		 * don't let the filecache_min fall below the specified critical level
2165 		 */
2166 		vm_pageout_state.vm_page_filecache_min =
2167 		    ((AVAILABLE_NON_COMPRESSED_MEMORY) * 10) / divisor;
2168 	}
2169 #endif
2170 	if (vm_page_free_count < (vm_page_free_reserved / 4)) {
2171 		vm_pageout_state.vm_page_filecache_min = 0;
2172 	}
2173 }
2174 
2175 /*
2176  * This function is called only from vm_pageout_scan and
2177  * it updates the flow control time to detect if VM pageoutscan
2178  * isn't making progress.
2179  */
2180 static void
vps_flow_control_reset_deadlock_timer(struct flow_control * flow_control)2181 vps_flow_control_reset_deadlock_timer(struct flow_control *flow_control)
2182 {
2183 	mach_timespec_t ts;
2184 	clock_sec_t sec;
2185 	clock_nsec_t nsec;
2186 
2187 	ts.tv_sec = vm_pageout_state.vm_pageout_deadlock_wait / 1000;
2188 	ts.tv_nsec = (vm_pageout_state.vm_pageout_deadlock_wait % 1000) * 1000 * NSEC_PER_USEC;
2189 	clock_get_system_nanotime(&sec, &nsec);
2190 	flow_control->ts.tv_sec = (unsigned int) sec;
2191 	flow_control->ts.tv_nsec = nsec;
2192 	ADD_MACH_TIMESPEC(&flow_control->ts, &ts);
2193 
2194 	flow_control->state = FCS_DELAYED;
2195 
2196 	vm_pageout_vminfo.vm_pageout_scan_inactive_throttled_internal++;
2197 }
2198 
2199 /*
2200  * This function is called only from vm_pageout_scan and
2201  * it is the flow control logic of VM pageout scan which
2202  * controls if it should block and for how long.
2203  * Any blocking of vm_pageout_scan happens ONLY in this function.
2204  */
2205 static int
vps_flow_control(struct flow_control * flow_control,int * anons_grabbed,vm_object_t * object,int * delayed_unlock,vm_page_t * local_freeq,int * local_freed,int * vm_pageout_deadlock_target,unsigned int inactive_burst_count)2206 vps_flow_control(struct flow_control *flow_control, int *anons_grabbed, vm_object_t *object, int *delayed_unlock,
2207     vm_page_t *local_freeq, int *local_freed, int *vm_pageout_deadlock_target, unsigned int inactive_burst_count)
2208 {
2209 	boolean_t       exceeded_burst_throttle = FALSE;
2210 	unsigned int    msecs = 0;
2211 	uint32_t        inactive_external_count;
2212 	mach_timespec_t ts;
2213 	struct  vm_pageout_queue *iq;
2214 	struct  vm_pageout_queue *eq;
2215 	struct  vm_speculative_age_q *sq;
2216 
2217 	iq = &vm_pageout_queue_internal;
2218 	eq = &vm_pageout_queue_external;
2219 	sq = &vm_page_queue_speculative[VM_PAGE_SPECULATIVE_AGED_Q];
2220 
2221 	/*
2222 	 * Sometimes we have to pause:
2223 	 *	1) No inactive pages - nothing to do.
2224 	 *	2) Loop control - no acceptable pages found on the inactive queue
2225 	 *         within the last vm_pageout_burst_inactive_throttle iterations
2226 	 *	3) Flow control - default pageout queue is full
2227 	 */
2228 	if (vm_page_queue_empty(&vm_page_queue_inactive) &&
2229 	    vm_page_queue_empty(&vm_page_queue_anonymous) &&
2230 	    vm_page_queue_empty(&vm_page_queue_cleaned) &&
2231 	    vm_page_queue_empty(&sq->age_q)) {
2232 		VM_PAGEOUT_DEBUG(vm_pageout_scan_empty_throttle, 1);
2233 		msecs = vm_pageout_state.vm_pageout_empty_wait;
2234 	} else if (inactive_burst_count >=
2235 	    MIN(vm_pageout_state.vm_pageout_burst_inactive_throttle,
2236 	    (vm_page_inactive_count +
2237 	    vm_page_speculative_count))) {
2238 		VM_PAGEOUT_DEBUG(vm_pageout_scan_burst_throttle, 1);
2239 		msecs = vm_pageout_state.vm_pageout_burst_wait;
2240 
2241 		exceeded_burst_throttle = TRUE;
2242 	} else if (VM_PAGE_Q_THROTTLED(iq) &&
2243 	    VM_DYNAMIC_PAGING_ENABLED()) {
2244 		clock_sec_t sec;
2245 		clock_nsec_t nsec;
2246 
2247 		switch (flow_control->state) {
2248 		case FCS_IDLE:
2249 			if ((vm_page_free_count + *local_freed) < vm_page_free_target &&
2250 			    vm_pageout_state.vm_restricted_to_single_processor == FALSE) {
2251 				/*
2252 				 * since the compressor is running independently of vm_pageout_scan
2253 				 * let's not wait for it just yet... as long as we have a healthy supply
2254 				 * of filecache pages to work with, let's keep stealing those.
2255 				 */
2256 				inactive_external_count = vm_page_inactive_count - vm_page_anonymous_count;
2257 
2258 				if (vm_page_pageable_external_count > vm_pageout_state.vm_page_filecache_min &&
2259 				    (inactive_external_count >= VM_PAGE_INACTIVE_TARGET(vm_page_pageable_external_count))) {
2260 					*anons_grabbed = ANONS_GRABBED_LIMIT;
2261 					VM_PAGEOUT_DEBUG(vm_pageout_scan_throttle_deferred, 1);
2262 					return VM_PAGEOUT_SCAN_PROCEED;
2263 				}
2264 			}
2265 
2266 			vps_flow_control_reset_deadlock_timer(flow_control);
2267 			msecs = vm_pageout_state.vm_pageout_deadlock_wait;
2268 
2269 			break;
2270 
2271 		case FCS_DELAYED:
2272 			clock_get_system_nanotime(&sec, &nsec);
2273 			ts.tv_sec = (unsigned int) sec;
2274 			ts.tv_nsec = nsec;
2275 
2276 			if (CMP_MACH_TIMESPEC(&ts, &flow_control->ts) >= 0) {
2277 				/*
2278 				 * the pageout thread for the default pager is potentially
2279 				 * deadlocked since the
2280 				 * default pager queue has been throttled for more than the
2281 				 * allowable time... we need to move some clean pages or dirty
2282 				 * pages belonging to the external pagers if they aren't throttled
2283 				 * vm_page_free_wanted represents the number of threads currently
2284 				 * blocked waiting for pages... we'll move one page for each of
2285 				 * these plus a fixed amount to break the logjam... once we're done
2286 				 * moving this number of pages, we'll re-enter the FSC_DELAYED state
2287 				 * with a new timeout target since we have no way of knowing
2288 				 * whether we've broken the deadlock except through observation
2289 				 * of the queue associated with the default pager... we need to
2290 				 * stop moving pages and allow the system to run to see what
2291 				 * state it settles into.
2292 				 */
2293 
2294 				*vm_pageout_deadlock_target = vm_pageout_state.vm_pageout_deadlock_relief +
2295 				    vm_page_free_wanted + vm_page_free_wanted_privileged;
2296 				VM_PAGEOUT_DEBUG(vm_pageout_scan_deadlock_detected, 1);
2297 				flow_control->state = FCS_DEADLOCK_DETECTED;
2298 				thread_wakeup(VM_PAGEOUT_GC_EVENT);
2299 				return VM_PAGEOUT_SCAN_PROCEED;
2300 			}
2301 			/*
2302 			 * just resniff instead of trying
2303 			 * to compute a new delay time... we're going to be
2304 			 * awakened immediately upon a laundry completion,
2305 			 * so we won't wait any longer than necessary
2306 			 */
2307 			msecs = vm_pageout_state.vm_pageout_idle_wait;
2308 			break;
2309 
2310 		case FCS_DEADLOCK_DETECTED:
2311 			if (*vm_pageout_deadlock_target) {
2312 				return VM_PAGEOUT_SCAN_PROCEED;
2313 			}
2314 
2315 			vps_flow_control_reset_deadlock_timer(flow_control);
2316 			msecs = vm_pageout_state.vm_pageout_deadlock_wait;
2317 
2318 			break;
2319 		}
2320 	} else {
2321 		/*
2322 		 * No need to pause...
2323 		 */
2324 		return VM_PAGEOUT_SCAN_PROCEED;
2325 	}
2326 
2327 	vm_pageout_scan_wants_object = VM_OBJECT_NULL;
2328 
2329 	vm_pageout_prepare_to_block(object, delayed_unlock, local_freeq, local_freed,
2330 	    VM_PAGEOUT_PB_CONSIDER_WAKING_COMPACTOR_SWAPPER);
2331 
2332 	if (vm_page_free_count >= vm_page_free_target) {
2333 		/*
2334 		 * we're here because
2335 		 *  1) someone else freed up some pages while we had
2336 		 *     the queues unlocked above
2337 		 * and we've hit one of the 3 conditions that
2338 		 * cause us to pause the pageout scan thread
2339 		 *
2340 		 * since we already have enough free pages,
2341 		 * let's avoid stalling and return normally
2342 		 *
2343 		 * before we return, make sure the pageout I/O threads
2344 		 * are running throttled in case there are still requests
2345 		 * in the laundry... since we have enough free pages
2346 		 * we don't need the laundry to be cleaned in a timely
2347 		 * fashion... so let's avoid interfering with foreground
2348 		 * activity
2349 		 *
2350 		 * we don't want to hold vm_page_queue_free_lock when
2351 		 * calling vm_pageout_adjust_eq_iothrottle (since it
2352 		 * may cause other locks to be taken), we do the intitial
2353 		 * check outside of the lock.  Once we take the lock,
2354 		 * we recheck the condition since it may have changed.
2355 		 * if it has, no problem, we will make the threads
2356 		 * non-throttled before actually blocking
2357 		 */
2358 		vm_pageout_adjust_eq_iothrottle(&pgo_iothread_external_state, TRUE);
2359 	}
2360 	vm_free_page_lock();
2361 
2362 	if (vm_page_free_count >= vm_page_free_target &&
2363 	    (vm_page_free_wanted == 0) && (vm_page_free_wanted_privileged == 0)) {
2364 		return VM_PAGEOUT_SCAN_DONE_RETURN;
2365 	}
2366 	vm_free_page_unlock();
2367 
2368 	if ((vm_page_free_count + vm_page_cleaned_count) < vm_page_free_target) {
2369 		/*
2370 		 * we're most likely about to block due to one of
2371 		 * the 3 conditions that cause vm_pageout_scan to
2372 		 * not be able to make forward progress w/r
2373 		 * to providing new pages to the free queue,
2374 		 * so unthrottle the I/O threads in case we
2375 		 * have laundry to be cleaned... it needs
2376 		 * to be completed ASAP.
2377 		 *
2378 		 * even if we don't block, we want the io threads
2379 		 * running unthrottled since the sum of free +
2380 		 * clean pages is still under our free target
2381 		 */
2382 		vm_pageout_adjust_eq_iothrottle(&pgo_iothread_external_state, FALSE);
2383 	}
2384 	if (vm_page_cleaned_count > 0 && exceeded_burst_throttle == FALSE) {
2385 		/*
2386 		 * if we get here we're below our free target and
2387 		 * we're stalling due to a full laundry queue or
2388 		 * we don't have any inactive pages other then
2389 		 * those in the clean queue...
2390 		 * however, we have pages on the clean queue that
2391 		 * can be moved to the free queue, so let's not
2392 		 * stall the pageout scan
2393 		 */
2394 		flow_control->state = FCS_IDLE;
2395 		return VM_PAGEOUT_SCAN_PROCEED;
2396 	}
2397 	if (flow_control->state == FCS_DELAYED && !VM_PAGE_Q_THROTTLED(iq)) {
2398 		flow_control->state = FCS_IDLE;
2399 		return VM_PAGEOUT_SCAN_PROCEED;
2400 	}
2401 
2402 	VM_CHECK_MEMORYSTATUS;
2403 
2404 	if (flow_control->state != FCS_IDLE) {
2405 		VM_PAGEOUT_DEBUG(vm_pageout_scan_throttle, 1);
2406 	}
2407 
2408 	iq->pgo_throttled = TRUE;
2409 	assert_wait_timeout((event_t) &iq->pgo_laundry, THREAD_INTERRUPTIBLE, msecs, 1000 * NSEC_PER_USEC);
2410 
2411 	vm_page_unlock_queues();
2412 
2413 	assert(vm_pageout_scan_wants_object == VM_OBJECT_NULL);
2414 
2415 	VM_DEBUG_EVENT(vm_pageout_thread_block, VM_PAGEOUT_THREAD_BLOCK, DBG_FUNC_START,
2416 	    iq->pgo_laundry, iq->pgo_maxlaundry, msecs, 0);
2417 	memoryshot(VM_PAGEOUT_THREAD_BLOCK, DBG_FUNC_START);
2418 
2419 	thread_block(THREAD_CONTINUE_NULL);
2420 
2421 	VM_DEBUG_EVENT(vm_pageout_thread_block, VM_PAGEOUT_THREAD_BLOCK, DBG_FUNC_END,
2422 	    iq->pgo_laundry, iq->pgo_maxlaundry, msecs, 0);
2423 	memoryshot(VM_PAGEOUT_THREAD_BLOCK, DBG_FUNC_END);
2424 
2425 	vm_page_lock_queues();
2426 
2427 	iq->pgo_throttled = FALSE;
2428 
2429 	vps_init_page_targets();
2430 
2431 	return VM_PAGEOUT_SCAN_NEXT_ITERATION;
2432 }
2433 
2434 extern boolean_t vm_darkwake_mode;
2435 /*
2436  * This function is called only from vm_pageout_scan and
2437  * it will find and return the most appropriate page to be
2438  * reclaimed.
2439  */
2440 static int
vps_choose_victim_page(vm_page_t * victim_page,int * anons_grabbed,boolean_t * grab_anonymous,boolean_t force_anonymous,boolean_t * is_page_from_bg_q,unsigned int * reactivated_this_call)2441 vps_choose_victim_page(vm_page_t *victim_page, int *anons_grabbed, boolean_t *grab_anonymous, boolean_t force_anonymous,
2442     boolean_t *is_page_from_bg_q, unsigned int *reactivated_this_call)
2443 {
2444 	vm_page_t                       m = NULL;
2445 	vm_object_t                     m_object = VM_OBJECT_NULL;
2446 	uint32_t                        inactive_external_count;
2447 	struct vm_speculative_age_q     *sq;
2448 	struct vm_pageout_queue         *iq;
2449 	int                             retval = VM_PAGEOUT_SCAN_PROCEED;
2450 
2451 	sq = &vm_page_queue_speculative[VM_PAGE_SPECULATIVE_AGED_Q];
2452 	iq = &vm_pageout_queue_internal;
2453 
2454 	*is_page_from_bg_q = FALSE;
2455 
2456 	m = NULL;
2457 	m_object = VM_OBJECT_NULL;
2458 
2459 	if (VM_DYNAMIC_PAGING_ENABLED()) {
2460 		assert(vm_page_throttled_count == 0);
2461 		assert(vm_page_queue_empty(&vm_page_queue_throttled));
2462 	}
2463 
2464 	/*
2465 	 * Try for a clean-queue inactive page.
2466 	 * These are pages that vm_pageout_scan tried to steal earlier, but
2467 	 * were dirty and had to be cleaned.  Pick them up now that they are clean.
2468 	 */
2469 	if (!vm_page_queue_empty(&vm_page_queue_cleaned)) {
2470 		m = (vm_page_t) vm_page_queue_first(&vm_page_queue_cleaned);
2471 
2472 		assert(m->vmp_q_state == VM_PAGE_ON_INACTIVE_CLEANED_Q);
2473 
2474 		goto found_page;
2475 	}
2476 
2477 	/*
2478 	 * The next most eligible pages are ones we paged in speculatively,
2479 	 * but which have not yet been touched and have been aged out.
2480 	 */
2481 	if (!vm_page_queue_empty(&sq->age_q)) {
2482 		m = (vm_page_t) vm_page_queue_first(&sq->age_q);
2483 
2484 		assert(m->vmp_q_state == VM_PAGE_ON_SPECULATIVE_Q);
2485 
2486 		if (!m->vmp_dirty || force_anonymous == FALSE) {
2487 			goto found_page;
2488 		} else {
2489 			m = NULL;
2490 		}
2491 	}
2492 
2493 #if !CONFIG_JETSAM
2494 	if (vm_page_donate_mode != VM_PAGE_DONATE_DISABLED) {
2495 		if (vm_page_donate_queue_ripe && !vm_page_queue_empty(&vm_page_queue_donate)) {
2496 			m = (vm_page_t) vm_page_queue_first(&vm_page_queue_donate);
2497 			assert(m->vmp_on_specialq == VM_PAGE_SPECIAL_Q_DONATE);
2498 			goto found_page;
2499 		}
2500 	}
2501 #endif /* !CONFIG_JETSAM */
2502 
2503 	if (vm_page_background_mode != VM_PAGE_BG_DISABLED && (vm_page_background_count > vm_page_background_target)) {
2504 		vm_object_t     bg_m_object = NULL;
2505 
2506 		m = (vm_page_t) vm_page_queue_first(&vm_page_queue_background);
2507 
2508 		bg_m_object = VM_PAGE_OBJECT(m);
2509 
2510 		if (!VM_PAGE_PAGEABLE(m) || (vm_darkwake_mode && m->vmp_busy)) {
2511 			/*
2512 			 * This page is on the background queue
2513 			 * but not on a pageable queue OR is busy during
2514 			 * darkwake mode when the target is artificially lowered.
2515 			 * If it is busy during darkwake mode, and we don't skip it,
2516 			 * we will just swing back around and try again with the same
2517 			 * queue and might hit the same page or its neighbor in a
2518 			 * similar state. Both of these are transient states and will
2519 			 * get resolved, but, at this point let's ignore this page.
2520 			 */
2521 			if (vm_darkwake_mode && m->vmp_busy) {
2522 				if (bg_m_object->internal) {
2523 					vm_pageout_skipped_bq_internal++;
2524 				} else {
2525 					vm_pageout_skipped_bq_external++;
2526 				}
2527 			}
2528 		} else if (force_anonymous == FALSE || bg_m_object->internal) {
2529 			if (bg_m_object->internal &&
2530 			    (VM_PAGE_Q_THROTTLED(iq) ||
2531 			    vm_compressor_out_of_space() == TRUE ||
2532 			    vm_page_free_count < (vm_page_free_reserved / 4))) {
2533 				vm_pageout_skipped_bq_internal++;
2534 			} else {
2535 				*is_page_from_bg_q = TRUE;
2536 
2537 				if (bg_m_object->internal) {
2538 					vm_pageout_vminfo.vm_pageout_considered_bq_internal++;
2539 				} else {
2540 					vm_pageout_vminfo.vm_pageout_considered_bq_external++;
2541 				}
2542 				goto found_page;
2543 			}
2544 		}
2545 	}
2546 
2547 	inactive_external_count = vm_page_inactive_count - vm_page_anonymous_count;
2548 
2549 	if ((vm_page_pageable_external_count < vm_pageout_state.vm_page_filecache_min || force_anonymous == TRUE) ||
2550 	    (inactive_external_count < VM_PAGE_INACTIVE_TARGET(vm_page_pageable_external_count))) {
2551 		*grab_anonymous = TRUE;
2552 		*anons_grabbed = 0;
2553 
2554 		if (VM_CONFIG_SWAP_IS_ACTIVE) {
2555 			vm_pageout_vminfo.vm_pageout_skipped_external++;
2556 		} else {
2557 			if (vm_page_free_count < (COMPRESSOR_FREE_RESERVED_LIMIT * 2)) {
2558 				/*
2559 				 * No swap and we are in dangerously low levels of free memory.
2560 				 * If we keep going ahead with anonymous pages, we are going to run into a situation
2561 				 * where the compressor will be stuck waiting for free pages (if it isn't already).
2562 				 *
2563 				 * So, pick a file backed page...
2564 				 */
2565 				*grab_anonymous = FALSE;
2566 				*anons_grabbed = ANONS_GRABBED_LIMIT;
2567 				vm_pageout_vminfo.vm_pageout_skipped_internal++;
2568 			}
2569 		}
2570 		goto want_anonymous;
2571 	}
2572 	*grab_anonymous = (vm_page_anonymous_count > vm_page_anonymous_min);
2573 
2574 #if CONFIG_JETSAM
2575 	/* If the file-backed pool has accumulated
2576 	 * significantly more pages than the jetsam
2577 	 * threshold, prefer to reclaim those
2578 	 * inline to minimise compute overhead of reclaiming
2579 	 * anonymous pages.
2580 	 * This calculation does not account for the CPU local
2581 	 * external page queues, as those are expected to be
2582 	 * much smaller relative to the global pools.
2583 	 */
2584 
2585 	struct vm_pageout_queue *eq = &vm_pageout_queue_external;
2586 
2587 	if (*grab_anonymous == TRUE && !VM_PAGE_Q_THROTTLED(eq)) {
2588 		if (vm_page_pageable_external_count >
2589 		    vm_pageout_state.vm_page_filecache_min) {
2590 			if ((vm_page_pageable_external_count *
2591 			    vm_pageout_memorystatus_fb_factor_dr) >
2592 			    (memorystatus_available_pages_critical *
2593 			    vm_pageout_memorystatus_fb_factor_nr)) {
2594 				*grab_anonymous = FALSE;
2595 
2596 				VM_PAGEOUT_DEBUG(vm_grab_anon_overrides, 1);
2597 			}
2598 		}
2599 		if (*grab_anonymous) {
2600 			VM_PAGEOUT_DEBUG(vm_grab_anon_nops, 1);
2601 		}
2602 	}
2603 #endif /* CONFIG_JETSAM */
2604 
2605 want_anonymous:
2606 	if (*grab_anonymous == FALSE || *anons_grabbed >= ANONS_GRABBED_LIMIT || vm_page_queue_empty(&vm_page_queue_anonymous)) {
2607 		if (!vm_page_queue_empty(&vm_page_queue_inactive)) {
2608 			m = (vm_page_t) vm_page_queue_first(&vm_page_queue_inactive);
2609 
2610 			assert(m->vmp_q_state == VM_PAGE_ON_INACTIVE_EXTERNAL_Q);
2611 			*anons_grabbed = 0;
2612 
2613 			if (vm_page_pageable_external_count < vm_pageout_state.vm_page_filecache_min) {
2614 				if (!vm_page_queue_empty(&vm_page_queue_anonymous)) {
2615 					if ((++(*reactivated_this_call) % 100)) {
2616 						vm_pageout_vminfo.vm_pageout_filecache_min_reactivated++;
2617 
2618 						vm_page_activate(m);
2619 						counter_inc(&vm_statistics_reactivations);
2620 #if DEVELOPMENT || DEBUG
2621 						if (*is_page_from_bg_q == TRUE) {
2622 							if (m_object->internal) {
2623 								vm_pageout_rejected_bq_internal++;
2624 							} else {
2625 								vm_pageout_rejected_bq_external++;
2626 							}
2627 						}
2628 #endif /* DEVELOPMENT || DEBUG */
2629 						vm_pageout_state.vm_pageout_inactive_used++;
2630 
2631 						m = NULL;
2632 						retval = VM_PAGEOUT_SCAN_NEXT_ITERATION;
2633 
2634 						goto found_page;
2635 					}
2636 
2637 					/*
2638 					 * steal 1 of the file backed pages even if
2639 					 * we are under the limit that has been set
2640 					 * for a healthy filecache
2641 					 */
2642 				}
2643 			}
2644 			goto found_page;
2645 		}
2646 	}
2647 	if (!vm_page_queue_empty(&vm_page_queue_anonymous)) {
2648 		m = (vm_page_t) vm_page_queue_first(&vm_page_queue_anonymous);
2649 
2650 		assert(m->vmp_q_state == VM_PAGE_ON_INACTIVE_INTERNAL_Q);
2651 		*anons_grabbed += 1;
2652 
2653 		goto found_page;
2654 	}
2655 
2656 	m = NULL;
2657 
2658 found_page:
2659 	*victim_page = m;
2660 
2661 	return retval;
2662 }
2663 
2664 /*
2665  * This function is called only from vm_pageout_scan and
2666  * it will put a page back on the active/inactive queue
2667  * if we can't reclaim it for some reason.
2668  */
2669 static void
vps_requeue_page(vm_page_t m,int page_prev_q_state,__unused boolean_t page_from_bg_q)2670 vps_requeue_page(vm_page_t m, int page_prev_q_state, __unused boolean_t page_from_bg_q)
2671 {
2672 	if (page_prev_q_state == VM_PAGE_ON_SPECULATIVE_Q) {
2673 		vm_page_enqueue_inactive(m, FALSE);
2674 	} else {
2675 		vm_page_activate(m);
2676 	}
2677 
2678 #if DEVELOPMENT || DEBUG
2679 	vm_object_t m_object = VM_PAGE_OBJECT(m);
2680 
2681 	if (page_from_bg_q == TRUE) {
2682 		if (m_object->internal) {
2683 			vm_pageout_rejected_bq_internal++;
2684 		} else {
2685 			vm_pageout_rejected_bq_external++;
2686 		}
2687 	}
2688 #endif /* DEVELOPMENT || DEBUG */
2689 }
2690 
2691 /*
2692  * This function is called only from vm_pageout_scan and
2693  * it will try to grab the victim page's VM object (m_object)
2694  * which differs from the previous victim page's object (object).
2695  */
2696 static int
vps_switch_object(vm_page_t m,vm_object_t m_object,vm_object_t * object,int page_prev_q_state,boolean_t avoid_anon_pages,boolean_t page_from_bg_q)2697 vps_switch_object(vm_page_t m, vm_object_t m_object, vm_object_t *object, int page_prev_q_state, boolean_t avoid_anon_pages, boolean_t page_from_bg_q)
2698 {
2699 	struct vm_speculative_age_q *sq;
2700 
2701 	sq = &vm_page_queue_speculative[VM_PAGE_SPECULATIVE_AGED_Q];
2702 
2703 	/*
2704 	 * the object associated with candidate page is
2705 	 * different from the one we were just working
2706 	 * with... dump the lock if we still own it
2707 	 */
2708 	if (*object != NULL) {
2709 		vm_object_unlock(*object);
2710 		*object = NULL;
2711 	}
2712 	/*
2713 	 * Try to lock object; since we've alread got the
2714 	 * page queues lock, we can only 'try' for this one.
2715 	 * if the 'try' fails, we need to do a mutex_pause
2716 	 * to allow the owner of the object lock a chance to
2717 	 * run... otherwise, we're likely to trip over this
2718 	 * object in the same state as we work our way through
2719 	 * the queue... clumps of pages associated with the same
2720 	 * object are fairly typical on the inactive and active queues
2721 	 */
2722 	if (!vm_object_lock_try_scan(m_object)) {
2723 		vm_page_t m_want = NULL;
2724 
2725 		vm_pageout_vminfo.vm_pageout_inactive_nolock++;
2726 
2727 		if (page_prev_q_state == VM_PAGE_ON_INACTIVE_CLEANED_Q) {
2728 			VM_PAGEOUT_DEBUG(vm_pageout_cleaned_nolock, 1);
2729 		}
2730 
2731 		pmap_clear_reference(VM_PAGE_GET_PHYS_PAGE(m));
2732 
2733 		m->vmp_reference = FALSE;
2734 
2735 		if (!m_object->object_is_shared_cache) {
2736 			/*
2737 			 * don't apply this optimization if this is the shared cache
2738 			 * object, it's too easy to get rid of very hot and important
2739 			 * pages...
2740 			 * m->vmp_object must be stable since we hold the page queues lock...
2741 			 * we can update the scan_collisions field sans the object lock
2742 			 * since it is a separate field and this is the only spot that does
2743 			 * a read-modify-write operation and it is never executed concurrently...
2744 			 * we can asynchronously set this field to 0 when creating a UPL, so it
2745 			 * is possible for the value to be a bit non-determistic, but that's ok
2746 			 * since it's only used as a hint
2747 			 */
2748 			m_object->scan_collisions = 1;
2749 		}
2750 		if (page_from_bg_q) {
2751 			m_want = (vm_page_t) vm_page_queue_first(&vm_page_queue_background);
2752 		} else if (!vm_page_queue_empty(&vm_page_queue_cleaned)) {
2753 			m_want = (vm_page_t) vm_page_queue_first(&vm_page_queue_cleaned);
2754 		} else if (!vm_page_queue_empty(&sq->age_q)) {
2755 			m_want = (vm_page_t) vm_page_queue_first(&sq->age_q);
2756 		} else if ((avoid_anon_pages || vm_page_queue_empty(&vm_page_queue_anonymous)) &&
2757 		    !vm_page_queue_empty(&vm_page_queue_inactive)) {
2758 			m_want = (vm_page_t) vm_page_queue_first(&vm_page_queue_inactive);
2759 		} else if (!vm_page_queue_empty(&vm_page_queue_anonymous)) {
2760 			m_want = (vm_page_t) vm_page_queue_first(&vm_page_queue_anonymous);
2761 		}
2762 
2763 		/*
2764 		 * this is the next object we're going to be interested in
2765 		 * try to make sure its available after the mutex_pause
2766 		 * returns control
2767 		 */
2768 		if (m_want) {
2769 			vm_pageout_scan_wants_object = VM_PAGE_OBJECT(m_want);
2770 		}
2771 
2772 		vps_requeue_page(m, page_prev_q_state, page_from_bg_q);
2773 
2774 		return VM_PAGEOUT_SCAN_NEXT_ITERATION;
2775 	} else {
2776 		*object = m_object;
2777 		vm_pageout_scan_wants_object = VM_OBJECT_NULL;
2778 	}
2779 
2780 	return VM_PAGEOUT_SCAN_PROCEED;
2781 }
2782 
2783 /*
2784  * This function is called only from vm_pageout_scan and
2785  * it notices that pageout scan may be rendered ineffective
2786  * due to a FS deadlock and will jetsam a process if possible.
2787  * If jetsam isn't supported, it'll move the page to the active
2788  * queue to try and get some different pages pushed onwards so
2789  * we can try to get out of this scenario.
2790  */
2791 static void
vps_deal_with_throttled_queues(vm_page_t m,vm_object_t * object,uint32_t * vm_pageout_inactive_external_forced_reactivate_limit,int * delayed_unlock,boolean_t * force_anonymous,__unused boolean_t is_page_from_bg_q)2792 vps_deal_with_throttled_queues(vm_page_t m, vm_object_t *object, uint32_t *vm_pageout_inactive_external_forced_reactivate_limit,
2793     int *delayed_unlock, boolean_t *force_anonymous, __unused boolean_t is_page_from_bg_q)
2794 {
2795 	struct  vm_pageout_queue *eq;
2796 	vm_object_t cur_object = VM_OBJECT_NULL;
2797 
2798 	cur_object = *object;
2799 
2800 	eq = &vm_pageout_queue_external;
2801 
2802 	if (cur_object->internal == FALSE) {
2803 		/*
2804 		 * we need to break up the following potential deadlock case...
2805 		 *  a) The external pageout thread is stuck on the truncate lock for a file that is being extended i.e. written.
2806 		 *  b) The thread doing the writing is waiting for pages while holding the truncate lock
2807 		 *  c) Most of the pages in the inactive queue belong to this file.
2808 		 *
2809 		 * we are potentially in this deadlock because...
2810 		 *  a) the external pageout queue is throttled
2811 		 *  b) we're done with the active queue and moved on to the inactive queue
2812 		 *  c) we've got a dirty external page
2813 		 *
2814 		 * since we don't know the reason for the external pageout queue being throttled we
2815 		 * must suspect that we are deadlocked, so move the current page onto the active queue
2816 		 * in an effort to cause a page from the active queue to 'age' to the inactive queue
2817 		 *
2818 		 * if we don't have jetsam configured (i.e. we have a dynamic pager), set
2819 		 * 'force_anonymous' to TRUE to cause us to grab a page from the cleaned/anonymous
2820 		 * pool the next time we select a victim page... if we can make enough new free pages,
2821 		 * the deadlock will break, the external pageout queue will empty and it will no longer
2822 		 * be throttled
2823 		 *
2824 		 * if we have jetsam configured, keep a count of the pages reactivated this way so
2825 		 * that we can try to find clean pages in the active/inactive queues before
2826 		 * deciding to jetsam a process
2827 		 */
2828 		vm_pageout_vminfo.vm_pageout_scan_inactive_throttled_external++;
2829 
2830 		vm_page_check_pageable_safe(m);
2831 		assert(m->vmp_q_state == VM_PAGE_NOT_ON_Q);
2832 		vm_page_queue_enter(&vm_page_queue_active, m, vmp_pageq);
2833 		m->vmp_q_state = VM_PAGE_ON_ACTIVE_Q;
2834 		vm_page_active_count++;
2835 		vm_page_pageable_external_count++;
2836 
2837 		vm_pageout_adjust_eq_iothrottle(&pgo_iothread_external_state, FALSE);
2838 
2839 #if CONFIG_MEMORYSTATUS && CONFIG_JETSAM
2840 
2841 #pragma unused(force_anonymous)
2842 
2843 		*vm_pageout_inactive_external_forced_reactivate_limit -= 1;
2844 
2845 		if (*vm_pageout_inactive_external_forced_reactivate_limit <= 0) {
2846 			*vm_pageout_inactive_external_forced_reactivate_limit = vm_page_active_count + vm_page_inactive_count;
2847 			/*
2848 			 * Possible deadlock scenario so request jetsam action
2849 			 */
2850 
2851 			assert(cur_object);
2852 			vm_object_unlock(cur_object);
2853 
2854 			cur_object = VM_OBJECT_NULL;
2855 
2856 			/*
2857 			 * VM pageout scan needs to know we have dropped this lock and so set the
2858 			 * object variable we got passed in to NULL.
2859 			 */
2860 			*object = VM_OBJECT_NULL;
2861 
2862 			vm_page_unlock_queues();
2863 
2864 			VM_DEBUG_CONSTANT_EVENT(vm_pageout_jetsam, VM_PAGEOUT_JETSAM, DBG_FUNC_START,
2865 			    vm_page_active_count, vm_page_inactive_count, vm_page_free_count, vm_page_free_count);
2866 
2867 			/* Kill first suitable process. If this call returned FALSE, we might have simply purged a process instead. */
2868 			if (memorystatus_kill_on_VM_page_shortage() == TRUE) {
2869 				VM_PAGEOUT_DEBUG(vm_pageout_inactive_external_forced_jetsam_count, 1);
2870 			}
2871 
2872 			VM_DEBUG_CONSTANT_EVENT(vm_pageout_jetsam, VM_PAGEOUT_JETSAM, DBG_FUNC_END,
2873 			    vm_page_active_count, vm_page_inactive_count, vm_page_free_count, vm_page_free_count);
2874 
2875 			vm_page_lock_queues();
2876 			*delayed_unlock = 1;
2877 		}
2878 #else /* CONFIG_MEMORYSTATUS && CONFIG_JETSAM */
2879 
2880 #pragma unused(vm_pageout_inactive_external_forced_reactivate_limit)
2881 #pragma unused(delayed_unlock)
2882 
2883 		*force_anonymous = TRUE;
2884 #endif /* CONFIG_MEMORYSTATUS && CONFIG_JETSAM */
2885 	} else {
2886 		vm_page_activate(m);
2887 		counter_inc(&vm_statistics_reactivations);
2888 
2889 #if DEVELOPMENT || DEBUG
2890 		if (is_page_from_bg_q == TRUE) {
2891 			if (cur_object->internal) {
2892 				vm_pageout_rejected_bq_internal++;
2893 			} else {
2894 				vm_pageout_rejected_bq_external++;
2895 			}
2896 		}
2897 #endif /* DEVELOPMENT || DEBUG */
2898 
2899 		vm_pageout_state.vm_pageout_inactive_used++;
2900 	}
2901 }
2902 
2903 
2904 void
vm_page_balance_inactive(int max_to_move)2905 vm_page_balance_inactive(int max_to_move)
2906 {
2907 	vm_page_t m;
2908 
2909 	LCK_MTX_ASSERT(&vm_page_queue_lock, LCK_MTX_ASSERT_OWNED);
2910 
2911 	if (hibernation_vmqueues_inspection || hibernate_cleaning_in_progress) {
2912 		/*
2913 		 * It is likely that the hibernation code path is
2914 		 * dealing with these very queues as we are about
2915 		 * to move pages around in/from them and completely
2916 		 * change the linkage of the pages.
2917 		 *
2918 		 * And so we skip the rebalancing of these queues.
2919 		 */
2920 		return;
2921 	}
2922 	vm_page_inactive_target = VM_PAGE_INACTIVE_TARGET(vm_page_active_count +
2923 	    vm_page_inactive_count +
2924 	    vm_page_speculative_count);
2925 
2926 	while (max_to_move-- && (vm_page_inactive_count + vm_page_speculative_count) < vm_page_inactive_target) {
2927 		VM_PAGEOUT_DEBUG(vm_pageout_balanced, 1);
2928 
2929 		m = (vm_page_t) vm_page_queue_first(&vm_page_queue_active);
2930 
2931 		assert(m->vmp_q_state == VM_PAGE_ON_ACTIVE_Q);
2932 		assert(!m->vmp_laundry);
2933 		assert(VM_PAGE_OBJECT(m) != kernel_object);
2934 		assert(VM_PAGE_GET_PHYS_PAGE(m) != vm_page_guard_addr);
2935 
2936 		DTRACE_VM2(scan, int, 1, (uint64_t *), NULL);
2937 
2938 		/*
2939 		 * by not passing in a pmap_flush_context we will forgo any TLB flushing, local or otherwise...
2940 		 *
2941 		 * a TLB flush isn't really needed here since at worst we'll miss the reference bit being
2942 		 * updated in the PTE if a remote processor still has this mapping cached in its TLB when the
2943 		 * new reference happens. If no futher references happen on the page after that remote TLB flushes
2944 		 * we'll see a clean, non-referenced page when it eventually gets pulled out of the inactive queue
2945 		 * by pageout_scan, which is just fine since the last reference would have happened quite far
2946 		 * in the past (TLB caches don't hang around for very long), and of course could just as easily
2947 		 * have happened before we moved the page
2948 		 */
2949 		if (m->vmp_pmapped == TRUE) {
2950 			/*
2951 			 * We might be holding the page queue lock as a
2952 			 * spin lock and clearing the "referenced" bit could
2953 			 * take a while if there are lots of mappings of
2954 			 * that page, so make sure we acquire the lock as
2955 			 * as mutex to avoid a spinlock timeout.
2956 			 */
2957 			vm_page_lockconvert_queues();
2958 			pmap_clear_refmod_options(VM_PAGE_GET_PHYS_PAGE(m), VM_MEM_REFERENCED, PMAP_OPTIONS_NOFLUSH, (void *)NULL);
2959 		}
2960 
2961 		/*
2962 		 * The page might be absent or busy,
2963 		 * but vm_page_deactivate can handle that.
2964 		 * FALSE indicates that we don't want a H/W clear reference
2965 		 */
2966 		vm_page_deactivate_internal(m, FALSE);
2967 	}
2968 }
2969 
2970 /*
2971  *	vm_pageout_scan does the dirty work for the pageout daemon.
2972  *	It returns with both vm_page_queue_free_lock and vm_page_queue_lock
2973  *	held and vm_page_free_wanted == 0.
2974  */
2975 void
vm_pageout_scan(void)2976 vm_pageout_scan(void)
2977 {
2978 	unsigned int loop_count = 0;
2979 	unsigned int inactive_burst_count = 0;
2980 	unsigned int reactivated_this_call;
2981 	unsigned int reactivate_limit;
2982 	vm_page_t   local_freeq = NULL;
2983 	int         local_freed = 0;
2984 	int         delayed_unlock;
2985 	int         delayed_unlock_limit = 0;
2986 	int         refmod_state = 0;
2987 	int     vm_pageout_deadlock_target = 0;
2988 	struct  vm_pageout_queue *iq;
2989 	struct  vm_pageout_queue *eq;
2990 	struct  vm_speculative_age_q *sq;
2991 	struct  flow_control    flow_control = { .state = 0, .ts = { .tv_sec = 0, .tv_nsec = 0 } };
2992 	boolean_t inactive_throttled = FALSE;
2993 	vm_object_t     object = NULL;
2994 	uint32_t        inactive_reclaim_run;
2995 	boolean_t       grab_anonymous = FALSE;
2996 	boolean_t       force_anonymous = FALSE;
2997 	boolean_t       force_speculative_aging = FALSE;
2998 	int             anons_grabbed = 0;
2999 	int             page_prev_q_state = 0;
3000 	boolean_t       page_from_bg_q = FALSE;
3001 	uint32_t        vm_pageout_inactive_external_forced_reactivate_limit = 0;
3002 	vm_object_t     m_object = VM_OBJECT_NULL;
3003 	int             retval = 0;
3004 	boolean_t       lock_yield_check = FALSE;
3005 
3006 
3007 	VM_DEBUG_CONSTANT_EVENT(vm_pageout_scan, VM_PAGEOUT_SCAN, DBG_FUNC_START,
3008 	    vm_pageout_vminfo.vm_pageout_freed_speculative,
3009 	    vm_pageout_state.vm_pageout_inactive_clean,
3010 	    vm_pageout_vminfo.vm_pageout_inactive_dirty_internal,
3011 	    vm_pageout_vminfo.vm_pageout_inactive_dirty_external);
3012 
3013 	flow_control.state = FCS_IDLE;
3014 	iq = &vm_pageout_queue_internal;
3015 	eq = &vm_pageout_queue_external;
3016 	sq = &vm_page_queue_speculative[VM_PAGE_SPECULATIVE_AGED_Q];
3017 
3018 	/* Ask the pmap layer to return any pages it no longer needs. */
3019 	pmap_release_pages_fast();
3020 
3021 	vm_page_lock_queues();
3022 
3023 	delayed_unlock = 1;
3024 
3025 	/*
3026 	 *	Calculate the max number of referenced pages on the inactive
3027 	 *	queue that we will reactivate.
3028 	 */
3029 	reactivated_this_call = 0;
3030 	reactivate_limit = VM_PAGE_REACTIVATE_LIMIT(vm_page_active_count +
3031 	    vm_page_inactive_count);
3032 	inactive_reclaim_run = 0;
3033 
3034 	vm_pageout_inactive_external_forced_reactivate_limit = vm_page_active_count + vm_page_inactive_count;
3035 
3036 	/*
3037 	 *	We must limit the rate at which we send pages to the pagers
3038 	 *	so that we don't tie up too many pages in the I/O queues.
3039 	 *	We implement a throttling mechanism using the laundry count
3040 	 *      to limit the number of pages outstanding to the default
3041 	 *	and external pagers.  We can bypass the throttles and look
3042 	 *	for clean pages if the pageout queues don't drain in a timely
3043 	 *	fashion since this may indicate that the pageout paths are
3044 	 *	stalled waiting for memory, which only we can provide.
3045 	 */
3046 
3047 	vps_init_page_targets();
3048 	assert(object == NULL);
3049 	assert(delayed_unlock != 0);
3050 
3051 	for (;;) {
3052 		vm_page_t m;
3053 
3054 		DTRACE_VM2(rev, int, 1, (uint64_t *), NULL);
3055 
3056 		if (lock_yield_check) {
3057 			lock_yield_check = FALSE;
3058 
3059 			if (delayed_unlock++ > delayed_unlock_limit) {
3060 				vm_pageout_prepare_to_block(&object, &delayed_unlock, &local_freeq, &local_freed,
3061 				    VM_PAGEOUT_PB_CONSIDER_WAKING_COMPACTOR_SWAPPER);
3062 			} else if (vm_pageout_scan_wants_object) {
3063 				vm_page_unlock_queues();
3064 				mutex_pause(0);
3065 				vm_page_lock_queues();
3066 			} else if (vps_yield_for_pgqlockwaiters && lck_mtx_yield(&vm_page_queue_lock)) {
3067 				VM_PAGEOUT_DEBUG(vm_pageout_yield_for_free_pages, 1);
3068 			}
3069 		}
3070 
3071 		if (vm_upl_wait_for_pages < 0) {
3072 			vm_upl_wait_for_pages = 0;
3073 		}
3074 
3075 		delayed_unlock_limit = VM_PAGEOUT_DELAYED_UNLOCK_LIMIT + vm_upl_wait_for_pages;
3076 
3077 		if (delayed_unlock_limit > VM_PAGEOUT_DELAYED_UNLOCK_LIMIT_MAX) {
3078 			delayed_unlock_limit = VM_PAGEOUT_DELAYED_UNLOCK_LIMIT_MAX;
3079 		}
3080 
3081 		vps_deal_with_secluded_page_overflow(&local_freeq, &local_freed);
3082 
3083 		assert(delayed_unlock);
3084 
3085 		/*
3086 		 * maintain our balance
3087 		 */
3088 		vm_page_balance_inactive(1);
3089 
3090 
3091 		/**********************************************************************
3092 		* above this point we're playing with the active and secluded queues
3093 		* below this point we're playing with the throttling mechanisms
3094 		* and the inactive queue
3095 		**********************************************************************/
3096 
3097 		if (vm_page_free_count + local_freed >= vm_page_free_target) {
3098 			vm_pageout_scan_wants_object = VM_OBJECT_NULL;
3099 
3100 			vm_pageout_prepare_to_block(&object, &delayed_unlock, &local_freeq, &local_freed,
3101 			    VM_PAGEOUT_PB_CONSIDER_WAKING_COMPACTOR_SWAPPER);
3102 			/*
3103 			 * make sure the pageout I/O threads are running
3104 			 * throttled in case there are still requests
3105 			 * in the laundry... since we have met our targets
3106 			 * we don't need the laundry to be cleaned in a timely
3107 			 * fashion... so let's avoid interfering with foreground
3108 			 * activity
3109 			 */
3110 			vm_pageout_adjust_eq_iothrottle(&pgo_iothread_external_state, TRUE);
3111 
3112 			vm_free_page_lock();
3113 
3114 			if ((vm_page_free_count >= vm_page_free_target) &&
3115 			    (vm_page_free_wanted == 0) && (vm_page_free_wanted_privileged == 0)) {
3116 				/*
3117 				 * done - we have met our target *and*
3118 				 * there is no one waiting for a page.
3119 				 */
3120 return_from_scan:
3121 				assert(vm_pageout_scan_wants_object == VM_OBJECT_NULL);
3122 
3123 				VM_DEBUG_CONSTANT_EVENT(vm_pageout_scan, VM_PAGEOUT_SCAN, DBG_FUNC_NONE,
3124 				    vm_pageout_state.vm_pageout_inactive,
3125 				    vm_pageout_state.vm_pageout_inactive_used, 0, 0);
3126 				VM_DEBUG_CONSTANT_EVENT(vm_pageout_scan, VM_PAGEOUT_SCAN, DBG_FUNC_END,
3127 				    vm_pageout_vminfo.vm_pageout_freed_speculative,
3128 				    vm_pageout_state.vm_pageout_inactive_clean,
3129 				    vm_pageout_vminfo.vm_pageout_inactive_dirty_internal,
3130 				    vm_pageout_vminfo.vm_pageout_inactive_dirty_external);
3131 
3132 				return;
3133 			}
3134 			vm_free_page_unlock();
3135 		}
3136 
3137 		/*
3138 		 * Before anything, we check if we have any ripe volatile
3139 		 * objects around. If so, try to purge the first object.
3140 		 * If the purge fails, fall through to reclaim a page instead.
3141 		 * If the purge succeeds, go back to the top and reevalute
3142 		 * the new memory situation.
3143 		 */
3144 		retval = vps_purge_object();
3145 
3146 		if (retval == VM_PAGEOUT_SCAN_NEXT_ITERATION) {
3147 			/*
3148 			 * Success
3149 			 */
3150 			if (object != NULL) {
3151 				vm_object_unlock(object);
3152 				object = NULL;
3153 			}
3154 
3155 			lock_yield_check = FALSE;
3156 			continue;
3157 		}
3158 
3159 		/*
3160 		 * If our 'aged' queue is empty and we have some speculative pages
3161 		 * in the other queues, let's go through and see if we need to age
3162 		 * them.
3163 		 *
3164 		 * If we succeeded in aging a speculative Q or just that everything
3165 		 * looks normal w.r.t queue age and queue counts, we keep going onward.
3166 		 *
3167 		 * If, for some reason, we seem to have a mismatch between the spec.
3168 		 * page count and the page queues, we reset those variables and
3169 		 * restart the loop (LD TODO: Track this better?).
3170 		 */
3171 		if (vm_page_queue_empty(&sq->age_q) && vm_page_speculative_count) {
3172 			retval = vps_age_speculative_queue(force_speculative_aging);
3173 
3174 			if (retval == VM_PAGEOUT_SCAN_NEXT_ITERATION) {
3175 				lock_yield_check = FALSE;
3176 				continue;
3177 			}
3178 		}
3179 		force_speculative_aging = FALSE;
3180 
3181 		/*
3182 		 * Check to see if we need to evict objects from the cache.
3183 		 *
3184 		 * Note: 'object' here doesn't have anything to do with
3185 		 * the eviction part. We just need to make sure we have dropped
3186 		 * any object lock we might be holding if we need to go down
3187 		 * into the eviction logic.
3188 		 */
3189 		retval = vps_object_cache_evict(&object);
3190 
3191 		if (retval == VM_PAGEOUT_SCAN_NEXT_ITERATION) {
3192 			lock_yield_check = FALSE;
3193 			continue;
3194 		}
3195 
3196 
3197 		/*
3198 		 * Calculate our filecache_min that will affect the loop
3199 		 * going forward.
3200 		 */
3201 		vps_calculate_filecache_min();
3202 
3203 		/*
3204 		 * LD TODO: Use a structure to hold all state variables for a single
3205 		 * vm_pageout_scan iteration and pass that structure to this function instead.
3206 		 */
3207 		retval = vps_flow_control(&flow_control, &anons_grabbed, &object,
3208 		    &delayed_unlock, &local_freeq, &local_freed,
3209 		    &vm_pageout_deadlock_target, inactive_burst_count);
3210 
3211 		if (retval == VM_PAGEOUT_SCAN_NEXT_ITERATION) {
3212 			if (loop_count >= vm_page_inactive_count) {
3213 				loop_count = 0;
3214 			}
3215 
3216 			inactive_burst_count = 0;
3217 
3218 			assert(object == NULL);
3219 			assert(delayed_unlock != 0);
3220 
3221 			lock_yield_check = FALSE;
3222 			continue;
3223 		} else if (retval == VM_PAGEOUT_SCAN_DONE_RETURN) {
3224 			goto return_from_scan;
3225 		}
3226 
3227 		flow_control.state = FCS_IDLE;
3228 
3229 		vm_pageout_inactive_external_forced_reactivate_limit = MIN((vm_page_active_count + vm_page_inactive_count),
3230 		    vm_pageout_inactive_external_forced_reactivate_limit);
3231 		loop_count++;
3232 		inactive_burst_count++;
3233 		vm_pageout_state.vm_pageout_inactive++;
3234 
3235 		/*
3236 		 * Choose a victim.
3237 		 */
3238 
3239 		m = NULL;
3240 		retval = vps_choose_victim_page(&m, &anons_grabbed, &grab_anonymous, force_anonymous, &page_from_bg_q, &reactivated_this_call);
3241 
3242 		if (m == NULL) {
3243 			if (retval == VM_PAGEOUT_SCAN_NEXT_ITERATION) {
3244 				inactive_burst_count = 0;
3245 
3246 				if (page_prev_q_state == VM_PAGE_ON_INACTIVE_CLEANED_Q) {
3247 					VM_PAGEOUT_DEBUG(vm_pageout_cleaned_reactivated, 1);
3248 				}
3249 
3250 				lock_yield_check = TRUE;
3251 				continue;
3252 			}
3253 
3254 			/*
3255 			 * if we've gotten here, we have no victim page.
3256 			 * check to see if we've not finished balancing the queues
3257 			 * or we have a page on the aged speculative queue that we
3258 			 * skipped due to force_anonymous == TRUE.. or we have
3259 			 * speculative  pages that we can prematurely age... if
3260 			 * one of these cases we'll keep going, else panic
3261 			 */
3262 			force_anonymous = FALSE;
3263 			VM_PAGEOUT_DEBUG(vm_pageout_no_victim, 1);
3264 
3265 			if (!vm_page_queue_empty(&sq->age_q)) {
3266 				lock_yield_check = TRUE;
3267 				continue;
3268 			}
3269 
3270 			if (vm_page_speculative_count) {
3271 				force_speculative_aging = TRUE;
3272 				lock_yield_check = TRUE;
3273 				continue;
3274 			}
3275 			panic("vm_pageout: no victim");
3276 
3277 			/* NOTREACHED */
3278 		}
3279 
3280 		assert(VM_PAGE_PAGEABLE(m));
3281 		m_object = VM_PAGE_OBJECT(m);
3282 		force_anonymous = FALSE;
3283 
3284 		page_prev_q_state = m->vmp_q_state;
3285 		/*
3286 		 * we just found this page on one of our queues...
3287 		 * it can't also be on the pageout queue, so safe
3288 		 * to call vm_page_queues_remove
3289 		 */
3290 		bool donate = (m->vmp_on_specialq == VM_PAGE_SPECIAL_Q_DONATE);
3291 		vm_page_queues_remove(m, TRUE);
3292 		if (donate) {
3293 			/*
3294 			 * The compressor needs to see this bit to know
3295 			 * where this page needs to land. Also if stolen,
3296 			 * this bit helps put the page back in the right
3297 			 * special queue where it belongs.
3298 			 */
3299 			m->vmp_on_specialq = VM_PAGE_SPECIAL_Q_DONATE;
3300 		}
3301 
3302 		assert(!m->vmp_laundry);
3303 		assert(!m->vmp_private);
3304 		assert(!m->vmp_fictitious);
3305 		assert(m_object != kernel_object);
3306 		assert(VM_PAGE_GET_PHYS_PAGE(m) != vm_page_guard_addr);
3307 
3308 		vm_pageout_vminfo.vm_pageout_considered_page++;
3309 
3310 		DTRACE_VM2(scan, int, 1, (uint64_t *), NULL);
3311 
3312 		/*
3313 		 * check to see if we currently are working
3314 		 * with the same object... if so, we've
3315 		 * already got the lock
3316 		 */
3317 		if (m_object != object) {
3318 			boolean_t avoid_anon_pages = (grab_anonymous == FALSE || anons_grabbed >= ANONS_GRABBED_LIMIT);
3319 
3320 			/*
3321 			 * vps_switch_object() will always drop the 'object' lock first
3322 			 * and then try to acquire the 'm_object' lock. So 'object' has to point to
3323 			 * either 'm_object' or NULL.
3324 			 */
3325 			retval = vps_switch_object(m, m_object, &object, page_prev_q_state, avoid_anon_pages, page_from_bg_q);
3326 
3327 			if (retval == VM_PAGEOUT_SCAN_NEXT_ITERATION) {
3328 				lock_yield_check = TRUE;
3329 				continue;
3330 			}
3331 		}
3332 		assert(m_object == object);
3333 		assert(VM_PAGE_OBJECT(m) == m_object);
3334 
3335 		if (m->vmp_busy) {
3336 			/*
3337 			 *	Somebody is already playing with this page.
3338 			 *	Put it back on the appropriate queue
3339 			 *
3340 			 */
3341 			VM_PAGEOUT_DEBUG(vm_pageout_inactive_busy, 1);
3342 
3343 			if (page_prev_q_state == VM_PAGE_ON_INACTIVE_CLEANED_Q) {
3344 				VM_PAGEOUT_DEBUG(vm_pageout_cleaned_busy, 1);
3345 			}
3346 
3347 			vps_requeue_page(m, page_prev_q_state, page_from_bg_q);
3348 
3349 			lock_yield_check = TRUE;
3350 			continue;
3351 		}
3352 
3353 		/*
3354 		 *   if (m->vmp_cleaning && !m->vmp_free_when_done)
3355 		 *	If already cleaning this page in place
3356 		 *	just leave if off the paging queues.
3357 		 *	We can leave the page mapped, and upl_commit_range
3358 		 *	will put it on the clean queue.
3359 		 *
3360 		 *   if (m->vmp_free_when_done && !m->vmp_cleaning)
3361 		 *	an msync INVALIDATE is in progress...
3362 		 *	this page has been marked for destruction
3363 		 *      after it has been cleaned,
3364 		 *      but not yet gathered into a UPL
3365 		 *	where 'cleaning' will be set...
3366 		 *	just leave it off the paging queues
3367 		 *
3368 		 *   if (m->vmp_free_when_done && m->vmp_clenaing)
3369 		 *	an msync INVALIDATE is in progress
3370 		 *	and the UPL has already gathered this page...
3371 		 *	just leave it off the paging queues
3372 		 */
3373 		if (m->vmp_free_when_done || m->vmp_cleaning) {
3374 			lock_yield_check = TRUE;
3375 			continue;
3376 		}
3377 
3378 
3379 		/*
3380 		 *	If it's absent, in error or the object is no longer alive,
3381 		 *	we can reclaim the page... in the no longer alive case,
3382 		 *	there are 2 states the page can be in that preclude us
3383 		 *	from reclaiming it - busy or cleaning - that we've already
3384 		 *	dealt with
3385 		 */
3386 		if (m->vmp_absent || VMP_ERROR_GET(m) || !object->alive ||
3387 		    (!object->internal && object->pager == MEMORY_OBJECT_NULL)) {
3388 			if (m->vmp_absent) {
3389 				VM_PAGEOUT_DEBUG(vm_pageout_inactive_absent, 1);
3390 			} else if (!object->alive ||
3391 			    (!object->internal &&
3392 			    object->pager == MEMORY_OBJECT_NULL)) {
3393 				VM_PAGEOUT_DEBUG(vm_pageout_inactive_notalive, 1);
3394 			} else {
3395 				VM_PAGEOUT_DEBUG(vm_pageout_inactive_error, 1);
3396 			}
3397 reclaim_page:
3398 			if (vm_pageout_deadlock_target) {
3399 				VM_PAGEOUT_DEBUG(vm_pageout_scan_inactive_throttle_success, 1);
3400 				vm_pageout_deadlock_target--;
3401 			}
3402 
3403 			DTRACE_VM2(dfree, int, 1, (uint64_t *), NULL);
3404 
3405 			if (object->internal) {
3406 				DTRACE_VM2(anonfree, int, 1, (uint64_t *), NULL);
3407 			} else {
3408 				DTRACE_VM2(fsfree, int, 1, (uint64_t *), NULL);
3409 			}
3410 			assert(!m->vmp_cleaning);
3411 			assert(!m->vmp_laundry);
3412 
3413 			if (!object->internal &&
3414 			    object->pager != NULL &&
3415 			    object->pager->mo_pager_ops == &shared_region_pager_ops) {
3416 				shared_region_pager_reclaimed++;
3417 			}
3418 
3419 			m->vmp_busy = TRUE;
3420 
3421 			/*
3422 			 * remove page from object here since we're already
3423 			 * behind the object lock... defer the rest of the work
3424 			 * we'd normally do in vm_page_free_prepare_object
3425 			 * until 'vm_page_free_list' is called
3426 			 */
3427 			if (m->vmp_tabled) {
3428 				vm_page_remove(m, TRUE);
3429 			}
3430 
3431 			assert(m->vmp_pageq.next == 0 && m->vmp_pageq.prev == 0);
3432 			m->vmp_snext = local_freeq;
3433 			local_freeq = m;
3434 			local_freed++;
3435 
3436 			if (page_prev_q_state == VM_PAGE_ON_SPECULATIVE_Q) {
3437 				vm_pageout_vminfo.vm_pageout_freed_speculative++;
3438 			} else if (page_prev_q_state == VM_PAGE_ON_INACTIVE_CLEANED_Q) {
3439 				vm_pageout_vminfo.vm_pageout_freed_cleaned++;
3440 			} else if (page_prev_q_state == VM_PAGE_ON_INACTIVE_INTERNAL_Q) {
3441 				vm_pageout_vminfo.vm_pageout_freed_internal++;
3442 			} else {
3443 				vm_pageout_vminfo.vm_pageout_freed_external++;
3444 			}
3445 
3446 			inactive_burst_count = 0;
3447 
3448 			lock_yield_check = TRUE;
3449 			continue;
3450 		}
3451 		if (object->copy == VM_OBJECT_NULL) {
3452 			/*
3453 			 * No one else can have any interest in this page.
3454 			 * If this is an empty purgable object, the page can be
3455 			 * reclaimed even if dirty.
3456 			 * If the page belongs to a volatile purgable object, we
3457 			 * reactivate it if the compressor isn't active.
3458 			 */
3459 			if (object->purgable == VM_PURGABLE_EMPTY) {
3460 				if (m->vmp_pmapped == TRUE) {
3461 					/* unmap the page */
3462 					refmod_state = pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m));
3463 					if (refmod_state & VM_MEM_MODIFIED) {
3464 						SET_PAGE_DIRTY(m, FALSE);
3465 					}
3466 				}
3467 				if (m->vmp_dirty || m->vmp_precious) {
3468 					/* we saved the cost of cleaning this page ! */
3469 					vm_page_purged_count++;
3470 				}
3471 				goto reclaim_page;
3472 			}
3473 
3474 			if (VM_CONFIG_COMPRESSOR_IS_ACTIVE) {
3475 				/*
3476 				 * With the VM compressor, the cost of
3477 				 * reclaiming a page is much lower (no I/O),
3478 				 * so if we find a "volatile" page, it's better
3479 				 * to let it get compressed rather than letting
3480 				 * it occupy a full page until it gets purged.
3481 				 * So no need to check for "volatile" here.
3482 				 */
3483 			} else if (object->purgable == VM_PURGABLE_VOLATILE) {
3484 				/*
3485 				 * Avoid cleaning a "volatile" page which might
3486 				 * be purged soon.
3487 				 */
3488 
3489 				/* if it's wired, we can't put it on our queue */
3490 				assert(!VM_PAGE_WIRED(m));
3491 
3492 				/* just stick it back on! */
3493 				reactivated_this_call++;
3494 
3495 				if (page_prev_q_state == VM_PAGE_ON_INACTIVE_CLEANED_Q) {
3496 					VM_PAGEOUT_DEBUG(vm_pageout_cleaned_volatile_reactivated, 1);
3497 				}
3498 
3499 				goto reactivate_page;
3500 			}
3501 		}
3502 		/*
3503 		 *	If it's being used, reactivate.
3504 		 *	(Fictitious pages are either busy or absent.)
3505 		 *	First, update the reference and dirty bits
3506 		 *	to make sure the page is unreferenced.
3507 		 */
3508 		refmod_state = -1;
3509 
3510 		if (m->vmp_reference == FALSE && m->vmp_pmapped == TRUE) {
3511 			refmod_state = pmap_get_refmod(VM_PAGE_GET_PHYS_PAGE(m));
3512 
3513 			if (refmod_state & VM_MEM_REFERENCED) {
3514 				m->vmp_reference = TRUE;
3515 			}
3516 			if (refmod_state & VM_MEM_MODIFIED) {
3517 				SET_PAGE_DIRTY(m, FALSE);
3518 			}
3519 		}
3520 
3521 		if (m->vmp_reference || m->vmp_dirty) {
3522 			/* deal with a rogue "reusable" page */
3523 			VM_PAGEOUT_SCAN_HANDLE_REUSABLE_PAGE(m, m_object);
3524 		}
3525 
3526 		if (vm_pageout_state.vm_page_xpmapped_min_divisor == 0) {
3527 			vm_pageout_state.vm_page_xpmapped_min = 0;
3528 		} else {
3529 			vm_pageout_state.vm_page_xpmapped_min = (vm_page_external_count * 10) / vm_pageout_state.vm_page_xpmapped_min_divisor;
3530 		}
3531 
3532 		if (!m->vmp_no_cache &&
3533 		    page_from_bg_q == FALSE &&
3534 		    (m->vmp_reference || (m->vmp_xpmapped && !object->internal &&
3535 		    (vm_page_xpmapped_external_count < vm_pageout_state.vm_page_xpmapped_min)))) {
3536 			/*
3537 			 * The page we pulled off the inactive list has
3538 			 * been referenced.  It is possible for other
3539 			 * processors to be touching pages faster than we
3540 			 * can clear the referenced bit and traverse the
3541 			 * inactive queue, so we limit the number of
3542 			 * reactivations.
3543 			 */
3544 			if (++reactivated_this_call >= reactivate_limit &&
3545 			    !object->object_is_shared_cache &&
3546 			    !((m->vmp_realtime ||
3547 			    object->for_realtime) &&
3548 			    vm_pageout_protect_realtime)) {
3549 				vm_pageout_vminfo.vm_pageout_reactivation_limit_exceeded++;
3550 			} else if (++inactive_reclaim_run >= VM_PAGEOUT_INACTIVE_FORCE_RECLAIM) {
3551 				vm_pageout_vminfo.vm_pageout_inactive_force_reclaim++;
3552 				if (object->object_is_shared_cache) {
3553 					vm_pageout_vminfo.vm_pageout_forcereclaimed_sharedcache++;
3554 				} else if (m->vmp_realtime ||
3555 				    object->for_realtime) {
3556 					vm_pageout_vminfo.vm_pageout_forcereclaimed_realtime++;
3557 				}
3558 			} else {
3559 				uint32_t isinuse;
3560 
3561 				if (reactivated_this_call >= reactivate_limit) {
3562 					if (object->object_is_shared_cache) {
3563 						vm_pageout_vminfo.vm_pageout_protected_sharedcache++;
3564 					} else if ((m->vmp_realtime ||
3565 					    object->for_realtime) &&
3566 					    vm_pageout_protect_realtime) {
3567 						vm_pageout_vminfo.vm_pageout_protected_realtime++;
3568 					}
3569 				}
3570 				if (page_prev_q_state == VM_PAGE_ON_INACTIVE_CLEANED_Q) {
3571 					VM_PAGEOUT_DEBUG(vm_pageout_cleaned_reference_reactivated, 1);
3572 				}
3573 
3574 				vm_pageout_vminfo.vm_pageout_inactive_referenced++;
3575 reactivate_page:
3576 				if (!object->internal && object->pager != MEMORY_OBJECT_NULL &&
3577 				    vnode_pager_get_isinuse(object->pager, &isinuse) == KERN_SUCCESS && !isinuse) {
3578 					/*
3579 					 * no explict mappings of this object exist
3580 					 * and it's not open via the filesystem
3581 					 */
3582 					vm_page_deactivate(m);
3583 					VM_PAGEOUT_DEBUG(vm_pageout_inactive_deactivated, 1);
3584 				} else {
3585 					/*
3586 					 * The page was/is being used, so put back on active list.
3587 					 */
3588 					vm_page_activate(m);
3589 					counter_inc(&vm_statistics_reactivations);
3590 					inactive_burst_count = 0;
3591 				}
3592 #if DEVELOPMENT || DEBUG
3593 				if (page_from_bg_q == TRUE) {
3594 					if (m_object->internal) {
3595 						vm_pageout_rejected_bq_internal++;
3596 					} else {
3597 						vm_pageout_rejected_bq_external++;
3598 					}
3599 				}
3600 #endif /* DEVELOPMENT || DEBUG */
3601 
3602 				if (page_prev_q_state == VM_PAGE_ON_INACTIVE_CLEANED_Q) {
3603 					VM_PAGEOUT_DEBUG(vm_pageout_cleaned_reactivated, 1);
3604 				}
3605 				vm_pageout_state.vm_pageout_inactive_used++;
3606 
3607 				lock_yield_check = TRUE;
3608 				continue;
3609 			}
3610 			/*
3611 			 * Make sure we call pmap_get_refmod() if it
3612 			 * wasn't already called just above, to update
3613 			 * the dirty bit.
3614 			 */
3615 			if ((refmod_state == -1) && !m->vmp_dirty && m->vmp_pmapped) {
3616 				refmod_state = pmap_get_refmod(VM_PAGE_GET_PHYS_PAGE(m));
3617 				if (refmod_state & VM_MEM_MODIFIED) {
3618 					SET_PAGE_DIRTY(m, FALSE);
3619 				}
3620 			}
3621 		}
3622 
3623 		/*
3624 		 * we've got a candidate page to steal...
3625 		 *
3626 		 * m->vmp_dirty is up to date courtesy of the
3627 		 * preceding check for m->vmp_reference... if
3628 		 * we get here, then m->vmp_reference had to be
3629 		 * FALSE (or possibly "reactivate_limit" was
3630 		 * exceeded), but in either case we called
3631 		 * pmap_get_refmod() and updated both
3632 		 * m->vmp_reference and m->vmp_dirty
3633 		 *
3634 		 * if it's dirty or precious we need to
3635 		 * see if the target queue is throtttled
3636 		 * it if is, we need to skip over it by moving it back
3637 		 * to the end of the inactive queue
3638 		 */
3639 
3640 		inactive_throttled = FALSE;
3641 
3642 		if (m->vmp_dirty || m->vmp_precious) {
3643 			if (object->internal) {
3644 				if (VM_PAGE_Q_THROTTLED(iq)) {
3645 					inactive_throttled = TRUE;
3646 				}
3647 			} else if (VM_PAGE_Q_THROTTLED(eq)) {
3648 				inactive_throttled = TRUE;
3649 			}
3650 		}
3651 throttle_inactive:
3652 		if (!VM_DYNAMIC_PAGING_ENABLED() &&
3653 		    object->internal && m->vmp_dirty &&
3654 		    (object->purgable == VM_PURGABLE_DENY ||
3655 		    object->purgable == VM_PURGABLE_NONVOLATILE ||
3656 		    object->purgable == VM_PURGABLE_VOLATILE)) {
3657 			vm_page_check_pageable_safe(m);
3658 			assert(m->vmp_q_state == VM_PAGE_NOT_ON_Q);
3659 			vm_page_queue_enter(&vm_page_queue_throttled, m, vmp_pageq);
3660 			m->vmp_q_state = VM_PAGE_ON_THROTTLED_Q;
3661 			vm_page_throttled_count++;
3662 
3663 			VM_PAGEOUT_DEBUG(vm_pageout_scan_reclaimed_throttled, 1);
3664 
3665 			inactive_burst_count = 0;
3666 
3667 			lock_yield_check = TRUE;
3668 			continue;
3669 		}
3670 		if (inactive_throttled == TRUE) {
3671 			vps_deal_with_throttled_queues(m, &object, &vm_pageout_inactive_external_forced_reactivate_limit,
3672 			    &delayed_unlock, &force_anonymous, page_from_bg_q);
3673 
3674 			inactive_burst_count = 0;
3675 
3676 			if (page_prev_q_state == VM_PAGE_ON_INACTIVE_CLEANED_Q) {
3677 				VM_PAGEOUT_DEBUG(vm_pageout_cleaned_reactivated, 1);
3678 			}
3679 
3680 			lock_yield_check = TRUE;
3681 			continue;
3682 		}
3683 
3684 		/*
3685 		 * we've got a page that we can steal...
3686 		 * eliminate all mappings and make sure
3687 		 * we have the up-to-date modified state
3688 		 *
3689 		 * if we need to do a pmap_disconnect then we
3690 		 * need to re-evaluate m->vmp_dirty since the pmap_disconnect
3691 		 * provides the true state atomically... the
3692 		 * page was still mapped up to the pmap_disconnect
3693 		 * and may have been dirtied at the last microsecond
3694 		 *
3695 		 * Note that if 'pmapped' is FALSE then the page is not
3696 		 * and has not been in any map, so there is no point calling
3697 		 * pmap_disconnect().  m->vmp_dirty could have been set in anticipation
3698 		 * of likely usage of the page.
3699 		 */
3700 		if (m->vmp_pmapped == TRUE) {
3701 			int pmap_options;
3702 
3703 			/*
3704 			 * Don't count this page as going into the compressor
3705 			 * if any of these are true:
3706 			 * 1) compressed pager isn't enabled
3707 			 * 2) Freezer enabled device with compressed pager
3708 			 *    backend (exclusive use) i.e. most of the VM system
3709 			 *    (including vm_pageout_scan) has no knowledge of
3710 			 *    the compressor
3711 			 * 3) This page belongs to a file and hence will not be
3712 			 *    sent into the compressor
3713 			 */
3714 			if (!VM_CONFIG_COMPRESSOR_IS_ACTIVE ||
3715 			    object->internal == FALSE) {
3716 				pmap_options = 0;
3717 			} else if (m->vmp_dirty || m->vmp_precious) {
3718 				/*
3719 				 * VM knows that this page is dirty (or
3720 				 * precious) and needs to be compressed
3721 				 * rather than freed.
3722 				 * Tell the pmap layer to count this page
3723 				 * as "compressed".
3724 				 */
3725 				pmap_options = PMAP_OPTIONS_COMPRESSOR;
3726 			} else {
3727 				/*
3728 				 * VM does not know if the page needs to
3729 				 * be preserved but the pmap layer might tell
3730 				 * us if any mapping has "modified" it.
3731 				 * Let's the pmap layer to count this page
3732 				 * as compressed if and only if it has been
3733 				 * modified.
3734 				 */
3735 				pmap_options =
3736 				    PMAP_OPTIONS_COMPRESSOR_IFF_MODIFIED;
3737 			}
3738 			refmod_state = pmap_disconnect_options(VM_PAGE_GET_PHYS_PAGE(m),
3739 			    pmap_options,
3740 			    NULL);
3741 			if (refmod_state & VM_MEM_MODIFIED) {
3742 				SET_PAGE_DIRTY(m, FALSE);
3743 			}
3744 		}
3745 
3746 		/*
3747 		 * reset our count of pages that have been reclaimed
3748 		 * since the last page was 'stolen'
3749 		 */
3750 		inactive_reclaim_run = 0;
3751 
3752 		/*
3753 		 *	If it's clean and not precious, we can free the page.
3754 		 */
3755 		if (!m->vmp_dirty && !m->vmp_precious) {
3756 			vm_pageout_state.vm_pageout_inactive_clean++;
3757 
3758 			/*
3759 			 * OK, at this point we have found a page we are going to free.
3760 			 */
3761 #if CONFIG_PHANTOM_CACHE
3762 			if (!object->internal) {
3763 				vm_phantom_cache_add_ghost(m);
3764 			}
3765 #endif
3766 			goto reclaim_page;
3767 		}
3768 
3769 		/*
3770 		 * The page may have been dirtied since the last check
3771 		 * for a throttled target queue (which may have been skipped
3772 		 * if the page was clean then).  With the dirty page
3773 		 * disconnected here, we can make one final check.
3774 		 */
3775 		if (object->internal) {
3776 			if (VM_PAGE_Q_THROTTLED(iq)) {
3777 				inactive_throttled = TRUE;
3778 			}
3779 		} else if (VM_PAGE_Q_THROTTLED(eq)) {
3780 			inactive_throttled = TRUE;
3781 		}
3782 
3783 		if (inactive_throttled == TRUE) {
3784 			goto throttle_inactive;
3785 		}
3786 
3787 #if VM_PRESSURE_EVENTS
3788 #if CONFIG_JETSAM
3789 
3790 		/*
3791 		 * If Jetsam is enabled, then the sending
3792 		 * of memory pressure notifications is handled
3793 		 * from the same thread that takes care of high-water
3794 		 * and other jetsams i.e. the memorystatus_thread.
3795 		 */
3796 
3797 #else /* CONFIG_JETSAM */
3798 
3799 		vm_pressure_response();
3800 
3801 #endif /* CONFIG_JETSAM */
3802 #endif /* VM_PRESSURE_EVENTS */
3803 
3804 		if (page_prev_q_state == VM_PAGE_ON_SPECULATIVE_Q) {
3805 			VM_PAGEOUT_DEBUG(vm_pageout_speculative_dirty, 1);
3806 		}
3807 
3808 		if (object->internal) {
3809 			vm_pageout_vminfo.vm_pageout_inactive_dirty_internal++;
3810 		} else {
3811 			vm_pageout_vminfo.vm_pageout_inactive_dirty_external++;
3812 		}
3813 
3814 		/*
3815 		 * internal pages will go to the compressor...
3816 		 * external pages will go to the appropriate pager to be cleaned
3817 		 * and upon completion will end up on 'vm_page_queue_cleaned' which
3818 		 * is a preferred queue to steal from
3819 		 */
3820 		vm_pageout_cluster(m);
3821 		inactive_burst_count = 0;
3822 
3823 		/*
3824 		 * back to top of pageout scan loop
3825 		 */
3826 	}
3827 }
3828 
3829 
3830 void
vm_page_free_reserve(int pages)3831 vm_page_free_reserve(
3832 	int pages)
3833 {
3834 	int             free_after_reserve;
3835 
3836 	if (VM_CONFIG_COMPRESSOR_IS_PRESENT) {
3837 		if ((vm_page_free_reserved + pages + COMPRESSOR_FREE_RESERVED_LIMIT) >= (VM_PAGE_FREE_RESERVED_LIMIT + COMPRESSOR_FREE_RESERVED_LIMIT)) {
3838 			vm_page_free_reserved = VM_PAGE_FREE_RESERVED_LIMIT + COMPRESSOR_FREE_RESERVED_LIMIT;
3839 		} else {
3840 			vm_page_free_reserved += (pages + COMPRESSOR_FREE_RESERVED_LIMIT);
3841 		}
3842 	} else {
3843 		if ((vm_page_free_reserved + pages) >= VM_PAGE_FREE_RESERVED_LIMIT) {
3844 			vm_page_free_reserved = VM_PAGE_FREE_RESERVED_LIMIT;
3845 		} else {
3846 			vm_page_free_reserved += pages;
3847 		}
3848 	}
3849 	free_after_reserve = vm_pageout_state.vm_page_free_count_init - vm_page_free_reserved;
3850 
3851 	vm_page_free_min = vm_page_free_reserved +
3852 	    VM_PAGE_FREE_MIN(free_after_reserve);
3853 
3854 	if (vm_page_free_min > VM_PAGE_FREE_MIN_LIMIT) {
3855 		vm_page_free_min = VM_PAGE_FREE_MIN_LIMIT;
3856 	}
3857 
3858 	vm_page_free_target = vm_page_free_reserved +
3859 	    VM_PAGE_FREE_TARGET(free_after_reserve);
3860 
3861 	if (vm_page_free_target > VM_PAGE_FREE_TARGET_LIMIT) {
3862 		vm_page_free_target = VM_PAGE_FREE_TARGET_LIMIT;
3863 	}
3864 
3865 	if (vm_page_free_target < vm_page_free_min + 5) {
3866 		vm_page_free_target = vm_page_free_min + 5;
3867 	}
3868 
3869 	vm_page_throttle_limit = vm_page_free_target - (vm_page_free_target / 2);
3870 }
3871 
3872 /*
3873  *	vm_pageout is the high level pageout daemon.
3874  */
3875 
3876 void
vm_pageout_continue(void)3877 vm_pageout_continue(void)
3878 {
3879 	DTRACE_VM2(pgrrun, int, 1, (uint64_t *), NULL);
3880 	VM_PAGEOUT_DEBUG(vm_pageout_scan_event_counter, 1);
3881 
3882 	vm_free_page_lock();
3883 	vm_pageout_running = TRUE;
3884 	vm_free_page_unlock();
3885 
3886 	vm_pageout_scan();
3887 	/*
3888 	 * we hold both the vm_page_queue_free_lock
3889 	 * and the vm_page_queues_lock at this point
3890 	 */
3891 	assert(vm_page_free_wanted == 0);
3892 	assert(vm_page_free_wanted_privileged == 0);
3893 	assert_wait((event_t) &vm_page_free_wanted, THREAD_UNINT);
3894 
3895 	vm_pageout_running = FALSE;
3896 #if XNU_TARGET_OS_OSX
3897 	if (vm_pageout_waiter) {
3898 		vm_pageout_waiter = FALSE;
3899 		thread_wakeup((event_t)&vm_pageout_waiter);
3900 	}
3901 #endif /* XNU_TARGET_OS_OSX */
3902 
3903 	vm_free_page_unlock();
3904 	vm_page_unlock_queues();
3905 
3906 	thread_block((thread_continue_t)vm_pageout_continue);
3907 	/*NOTREACHED*/
3908 }
3909 
3910 #if XNU_TARGET_OS_OSX
3911 kern_return_t
vm_pageout_wait(uint64_t deadline)3912 vm_pageout_wait(uint64_t deadline)
3913 {
3914 	kern_return_t kr;
3915 
3916 	vm_free_page_lock();
3917 	for (kr = KERN_SUCCESS; vm_pageout_running && (KERN_SUCCESS == kr);) {
3918 		vm_pageout_waiter = TRUE;
3919 		if (THREAD_AWAKENED != lck_mtx_sleep_deadline(
3920 			    &vm_page_queue_free_lock, LCK_SLEEP_DEFAULT,
3921 			    (event_t) &vm_pageout_waiter, THREAD_UNINT, deadline)) {
3922 			kr = KERN_OPERATION_TIMED_OUT;
3923 		}
3924 	}
3925 	vm_free_page_unlock();
3926 
3927 	return kr;
3928 }
3929 #endif /* XNU_TARGET_OS_OSX */
3930 
3931 OS_NORETURN
3932 static void
vm_pageout_iothread_external_continue(struct pgo_iothread_state * ethr,__unused wait_result_t w)3933 vm_pageout_iothread_external_continue(struct pgo_iothread_state *ethr, __unused wait_result_t w)
3934 {
3935 	vm_page_t       m = NULL;
3936 	vm_object_t     object;
3937 	vm_object_offset_t offset;
3938 	memory_object_t pager;
3939 	struct vm_pageout_queue *q = ethr->q;
3940 
3941 	/* On systems with a compressor, the external IO thread clears its
3942 	 * VM privileged bit to accommodate large allocations (e.g. bulk UPL
3943 	 * creation)
3944 	 */
3945 	if (VM_CONFIG_COMPRESSOR_IS_PRESENT) {
3946 		current_thread()->options &= ~TH_OPT_VMPRIV;
3947 	}
3948 
3949 	sched_cond_ack(&(ethr->pgo_wakeup));
3950 
3951 	while (true) {
3952 		vm_page_lockspin_queues();
3953 
3954 		while (!vm_page_queue_empty(&q->pgo_pending)) {
3955 			q->pgo_busy = TRUE;
3956 			vm_page_queue_remove_first(&q->pgo_pending, m, vmp_pageq);
3957 
3958 			assert(m->vmp_q_state == VM_PAGE_ON_PAGEOUT_Q);
3959 			VM_PAGE_CHECK(m);
3960 			/*
3961 			 * grab a snapshot of the object and offset this
3962 			 * page is tabled in so that we can relookup this
3963 			 * page after we've taken the object lock - these
3964 			 * fields are stable while we hold the page queues lock
3965 			 * but as soon as we drop it, there is nothing to keep
3966 			 * this page in this object... we hold an activity_in_progress
3967 			 * on this object which will keep it from terminating
3968 			 */
3969 			object = VM_PAGE_OBJECT(m);
3970 			offset = m->vmp_offset;
3971 
3972 			m->vmp_q_state = VM_PAGE_NOT_ON_Q;
3973 			VM_PAGE_ZERO_PAGEQ_ENTRY(m);
3974 
3975 			vm_page_unlock_queues();
3976 
3977 			vm_object_lock(object);
3978 
3979 			m = vm_page_lookup(object, offset);
3980 
3981 			if (m == NULL || m->vmp_busy || m->vmp_cleaning ||
3982 			    !m->vmp_laundry || (m->vmp_q_state != VM_PAGE_NOT_ON_Q)) {
3983 				/*
3984 				 * it's either the same page that someone else has
3985 				 * started cleaning (or it's finished cleaning or
3986 				 * been put back on the pageout queue), or
3987 				 * the page has been freed or we have found a
3988 				 * new page at this offset... in all of these cases
3989 				 * we merely need to release the activity_in_progress
3990 				 * we took when we put the page on the pageout queue
3991 				 */
3992 				vm_object_activity_end(object);
3993 				vm_object_unlock(object);
3994 
3995 				vm_page_lockspin_queues();
3996 				continue;
3997 			}
3998 			pager = object->pager;
3999 
4000 			if (pager == MEMORY_OBJECT_NULL) {
4001 				/*
4002 				 * This pager has been destroyed by either
4003 				 * memory_object_destroy or vm_object_destroy, and
4004 				 * so there is nowhere for the page to go.
4005 				 */
4006 				if (m->vmp_free_when_done) {
4007 					/*
4008 					 * Just free the page... VM_PAGE_FREE takes
4009 					 * care of cleaning up all the state...
4010 					 * including doing the vm_pageout_throttle_up
4011 					 */
4012 					VM_PAGE_FREE(m);
4013 				} else {
4014 					vm_page_lockspin_queues();
4015 
4016 					vm_pageout_throttle_up(m);
4017 					vm_page_activate(m);
4018 
4019 					vm_page_unlock_queues();
4020 
4021 					/*
4022 					 *	And we are done with it.
4023 					 */
4024 				}
4025 				vm_object_activity_end(object);
4026 				vm_object_unlock(object);
4027 
4028 				vm_page_lockspin_queues();
4029 				continue;
4030 			}
4031 	#if 0
4032 			/*
4033 			 * we don't hold the page queue lock
4034 			 * so this check isn't safe to make
4035 			 */
4036 			VM_PAGE_CHECK(m);
4037 	#endif
4038 			/*
4039 			 * give back the activity_in_progress reference we
4040 			 * took when we queued up this page and replace it
4041 			 * it with a paging_in_progress reference that will
4042 			 * also hold the paging offset from changing and
4043 			 * prevent the object from terminating
4044 			 */
4045 			vm_object_activity_end(object);
4046 			vm_object_paging_begin(object);
4047 			vm_object_unlock(object);
4048 
4049 			/*
4050 			 * Send the data to the pager.
4051 			 * any pageout clustering happens there
4052 			 */
4053 			memory_object_data_return(pager,
4054 			    m->vmp_offset + object->paging_offset,
4055 			    PAGE_SIZE,
4056 			    NULL,
4057 			    NULL,
4058 			    FALSE,
4059 			    FALSE,
4060 			    0);
4061 
4062 			vm_object_lock(object);
4063 			vm_object_paging_end(object);
4064 			vm_object_unlock(object);
4065 
4066 			vm_pageout_io_throttle();
4067 
4068 			vm_page_lockspin_queues();
4069 		}
4070 		q->pgo_busy = FALSE;
4071 
4072 		vm_page_unlock_queues();
4073 		sched_cond_wait_parameter(&(ethr->pgo_wakeup), THREAD_UNINT, (thread_continue_t)vm_pageout_iothread_external_continue, ethr);
4074 	}
4075 	/*NOTREACHED*/
4076 }
4077 
4078 
4079 #define         MAX_FREE_BATCH          32
4080 uint32_t vm_compressor_time_thread; /* Set via sysctl to record time accrued by
4081                                      * this thread.
4082                                      */
4083 
4084 
4085 OS_NORETURN
4086 static void
vm_pageout_iothread_internal_continue(struct pgo_iothread_state * cq,__unused wait_result_t w)4087 vm_pageout_iothread_internal_continue(struct pgo_iothread_state *cq, __unused wait_result_t w)
4088 {
4089 	struct vm_pageout_queue *q;
4090 	vm_page_t       m = NULL;
4091 	boolean_t       pgo_draining;
4092 	vm_page_t   local_q;
4093 	int         local_cnt;
4094 	vm_page_t   local_freeq = NULL;
4095 	int         local_freed = 0;
4096 	int         local_batch_size;
4097 #if DEVELOPMENT || DEBUG
4098 	int       ncomps = 0;
4099 	boolean_t marked_active = FALSE;
4100 	int       num_pages_processed = 0;
4101 #endif
4102 	void *chead = NULL;
4103 
4104 	KERNEL_DEBUG(0xe040000c | DBG_FUNC_END, 0, 0, 0, 0, 0);
4105 
4106 	sched_cond_ack(&(cq->pgo_wakeup));
4107 
4108 	q = cq->q;
4109 
4110 	while (true) {
4111 #if DEVELOPMENT || DEBUG
4112 		bool benchmark_accounting = false;
4113 		/*
4114 		 * If we're running the compressor perf test, only process the benchmark pages.
4115 		 * We'll get back to our regular queue once the benchmark is done
4116 		 */
4117 		if (compressor_running_perf_test) {
4118 			q = cq->benchmark_q;
4119 			if (!vm_page_queue_empty(&q->pgo_pending)) {
4120 				benchmark_accounting = true;
4121 			} else {
4122 				q = cq->q;
4123 				benchmark_accounting = false;
4124 			}
4125 		}
4126 #endif /* DEVELOPMENT || DEBUG */
4127 
4128 #if __AMP__
4129 		if (vm_compressor_ebound && (vm_pageout_state.vm_compressor_thread_count > 1)) {
4130 			local_batch_size = (q->pgo_maxlaundry >> 3);
4131 			local_batch_size = MAX(local_batch_size, 16);
4132 		} else {
4133 			local_batch_size = q->pgo_maxlaundry / (vm_pageout_state.vm_compressor_thread_count * 2);
4134 		}
4135 #else
4136 		local_batch_size = q->pgo_maxlaundry / (vm_pageout_state.vm_compressor_thread_count * 2);
4137 #endif
4138 
4139 #if RECORD_THE_COMPRESSED_DATA
4140 		if (q->pgo_laundry) {
4141 			c_compressed_record_init();
4142 		}
4143 #endif
4144 		while (true) {
4145 			int     pages_left_on_q = 0;
4146 
4147 			local_cnt = 0;
4148 			local_q = NULL;
4149 
4150 			KERNEL_DEBUG(0xe0400014 | DBG_FUNC_START, 0, 0, 0, 0, 0);
4151 
4152 			vm_page_lock_queues();
4153 #if DEVELOPMENT || DEBUG
4154 			if (marked_active == FALSE) {
4155 				vmct_active++;
4156 				vmct_state[cq->id] = VMCT_ACTIVE;
4157 				marked_active = TRUE;
4158 				if (vmct_active == 1) {
4159 					vm_compressor_epoch_start = mach_absolute_time();
4160 				}
4161 			}
4162 #endif
4163 			KERNEL_DEBUG(0xe0400014 | DBG_FUNC_END, 0, 0, 0, 0, 0);
4164 
4165 			KERNEL_DEBUG(0xe0400018 | DBG_FUNC_START, q->pgo_laundry, 0, 0, 0, 0);
4166 
4167 			while (!vm_page_queue_empty(&q->pgo_pending) && local_cnt < local_batch_size) {
4168 				vm_page_queue_remove_first(&q->pgo_pending, m, vmp_pageq);
4169 				assert(m->vmp_q_state == VM_PAGE_ON_PAGEOUT_Q);
4170 				VM_PAGE_CHECK(m);
4171 
4172 				m->vmp_q_state = VM_PAGE_NOT_ON_Q;
4173 				VM_PAGE_ZERO_PAGEQ_ENTRY(m);
4174 				m->vmp_laundry = FALSE;
4175 
4176 				m->vmp_snext = local_q;
4177 				local_q = m;
4178 				local_cnt++;
4179 			}
4180 			if (local_q == NULL) {
4181 				break;
4182 			}
4183 
4184 			q->pgo_busy = TRUE;
4185 
4186 			if ((pgo_draining = q->pgo_draining) == FALSE) {
4187 				vm_pageout_throttle_up_batch(q, local_cnt);
4188 				pages_left_on_q = q->pgo_laundry;
4189 			} else {
4190 				pages_left_on_q = q->pgo_laundry - local_cnt;
4191 			}
4192 
4193 			vm_page_unlock_queues();
4194 
4195 #if !RECORD_THE_COMPRESSED_DATA
4196 			if (pages_left_on_q >= local_batch_size && cq->id < (vm_pageout_state.vm_compressor_thread_count - 1)) {
4197 				// wake up the next compressor thread
4198 				sched_cond_signal(&pgo_iothread_internal_state[cq->id + 1].pgo_wakeup,
4199 				    pgo_iothread_internal_state[cq->id + 1].pgo_iothread);
4200 			}
4201 #endif
4202 			KERNEL_DEBUG(0xe0400018 | DBG_FUNC_END, q->pgo_laundry, 0, 0, 0, 0);
4203 
4204 			while (local_q) {
4205 				KERNEL_DEBUG(0xe0400024 | DBG_FUNC_START, local_cnt, 0, 0, 0, 0);
4206 
4207 				m = local_q;
4208 				local_q = m->vmp_snext;
4209 				m->vmp_snext = NULL;
4210 
4211 				/*
4212 				 * Technically we need the pageq locks to manipulate this field.
4213 				 * However, this page has been removed from all queues and is only
4214 				 * known to this compressor thread dealing with this local queue.
4215 				 *
4216 				 * TODO LIONEL: Add a second localq that is the early localq and
4217 				 * put special pages like this one on that queue in the block above
4218 				 * under the pageq lock to avoid this 'works but not clean' logic.
4219 				 */
4220 				void *donate_queue_head;
4221 #if XNU_TARGET_OS_OSX
4222 				donate_queue_head = &cq->current_early_swapout_chead;
4223 #else /* XNU_TARGET_OS_OSX */
4224 				donate_queue_head = &cq->current_late_swapout_chead;
4225 #endif /* XNU_TARGET_OS_OSX */
4226 				if (m->vmp_on_specialq == VM_PAGE_SPECIAL_Q_DONATE) {
4227 					m->vmp_on_specialq = VM_PAGE_SPECIAL_Q_EMPTY;
4228 					chead = donate_queue_head;
4229 				} else {
4230 					chead = &cq->current_regular_swapout_chead;
4231 				}
4232 
4233 				if (vm_pageout_compress_page(chead, cq->scratch_buf, m) == KERN_SUCCESS) {
4234 #if DEVELOPMENT || DEBUG
4235 					ncomps++;
4236 #endif
4237 					KERNEL_DEBUG(0xe0400024 | DBG_FUNC_END, local_cnt, 0, 0, 0, 0);
4238 
4239 					m->vmp_snext = local_freeq;
4240 					local_freeq = m;
4241 					local_freed++;
4242 
4243 					if (local_freed >= MAX_FREE_BATCH) {
4244 						OSAddAtomic64(local_freed, &vm_pageout_vminfo.vm_pageout_compressions);
4245 
4246 						vm_page_free_list(local_freeq, TRUE);
4247 
4248 						local_freeq = NULL;
4249 						local_freed = 0;
4250 					}
4251 				}
4252 #if DEVELOPMENT || DEBUG
4253 				num_pages_processed++;
4254 #endif /* DEVELOPMENT || DEBUG */
4255 #if !CONFIG_JETSAM
4256 				while (vm_page_free_count < COMPRESSOR_FREE_RESERVED_LIMIT) {
4257 					kern_return_t   wait_result;
4258 					int             need_wakeup = 0;
4259 
4260 					if (local_freeq) {
4261 						OSAddAtomic64(local_freed, &vm_pageout_vminfo.vm_pageout_compressions);
4262 
4263 						vm_page_free_list(local_freeq, TRUE);
4264 						local_freeq = NULL;
4265 						local_freed = 0;
4266 
4267 						continue;
4268 					}
4269 					vm_free_page_lock_spin();
4270 
4271 					if (vm_page_free_count < COMPRESSOR_FREE_RESERVED_LIMIT) {
4272 						if (vm_page_free_wanted_privileged++ == 0) {
4273 							need_wakeup = 1;
4274 						}
4275 						wait_result = assert_wait((event_t)&vm_page_free_wanted_privileged, THREAD_UNINT);
4276 
4277 						vm_free_page_unlock();
4278 
4279 						if (need_wakeup) {
4280 							thread_wakeup((event_t)&vm_page_free_wanted);
4281 						}
4282 
4283 						if (wait_result == THREAD_WAITING) {
4284 							thread_block(THREAD_CONTINUE_NULL);
4285 						}
4286 					} else {
4287 						vm_free_page_unlock();
4288 					}
4289 				}
4290 #endif
4291 			}
4292 			if (local_freeq) {
4293 				OSAddAtomic64(local_freed, &vm_pageout_vminfo.vm_pageout_compressions);
4294 
4295 				vm_page_free_list(local_freeq, TRUE);
4296 				local_freeq = NULL;
4297 				local_freed = 0;
4298 			}
4299 			if (pgo_draining == TRUE) {
4300 				vm_page_lockspin_queues();
4301 				vm_pageout_throttle_up_batch(q, local_cnt);
4302 				vm_page_unlock_queues();
4303 			}
4304 		}
4305 		KERNEL_DEBUG(0xe040000c | DBG_FUNC_START, 0, 0, 0, 0, 0);
4306 
4307 		/*
4308 		 * queue lock is held and our q is empty
4309 		 */
4310 		q->pgo_busy = FALSE;
4311 #if DEVELOPMENT || DEBUG
4312 		if (marked_active == TRUE) {
4313 			vmct_active--;
4314 			vmct_state[cq->id] = VMCT_IDLE;
4315 
4316 			if (vmct_active == 0) {
4317 				vm_compressor_epoch_stop = mach_absolute_time();
4318 				assertf(vm_compressor_epoch_stop >= vm_compressor_epoch_start,
4319 				    "Compressor epoch non-monotonic: 0x%llx -> 0x%llx",
4320 				    vm_compressor_epoch_start, vm_compressor_epoch_stop);
4321 				/* This interval includes intervals where one or more
4322 				 * compressor threads were pre-empted
4323 				 */
4324 				vmct_stats.vmct_cthreads_total += vm_compressor_epoch_stop - vm_compressor_epoch_start;
4325 			}
4326 		}
4327 		if (compressor_running_perf_test && benchmark_accounting) {
4328 			/*
4329 			 * We could turn ON compressor_running_perf_test while still processing
4330 			 * regular non-benchmark pages. We shouldn't count them here else we
4331 			 * could overshoot. We might also still be populating that benchmark Q
4332 			 * and be under pressure. So we will go back to the regular queues. And
4333 			 * benchmark accounting will be off for that case too.
4334 			 */
4335 			compressor_perf_test_pages_processed += num_pages_processed;
4336 			thread_wakeup(&compressor_perf_test_pages_processed);
4337 		}
4338 #endif
4339 		vm_page_unlock_queues();
4340 #if DEVELOPMENT || DEBUG
4341 		if (__improbable(vm_compressor_time_thread)) {
4342 			vmct_stats.vmct_runtimes[cq->id] = thread_get_runtime_self();
4343 			vmct_stats.vmct_pages[cq->id] += ncomps;
4344 			vmct_stats.vmct_iterations[cq->id]++;
4345 			if (ncomps > vmct_stats.vmct_maxpages[cq->id]) {
4346 				vmct_stats.vmct_maxpages[cq->id] = ncomps;
4347 			}
4348 			if (ncomps < vmct_stats.vmct_minpages[cq->id]) {
4349 				vmct_stats.vmct_minpages[cq->id] = ncomps;
4350 			}
4351 		}
4352 #endif
4353 
4354 		KERNEL_DEBUG(0xe0400018 | DBG_FUNC_END, 0, 0, 0, 0, 0);
4355 #if DEVELOPMENT || DEBUG
4356 		if (compressor_running_perf_test && benchmark_accounting) {
4357 			/*
4358 			 * We've been exclusively compressing pages from the benchmark queue,
4359 			 * do 1 pass over the internal queue before blocking.
4360 			 */
4361 			continue;
4362 		}
4363 #endif
4364 
4365 		sched_cond_wait_parameter(&(cq->pgo_wakeup), THREAD_UNINT, (thread_continue_t)vm_pageout_iothread_internal_continue, (void *) cq);
4366 	}
4367 	/*NOTREACHED*/
4368 }
4369 
4370 
4371 kern_return_t
vm_pageout_compress_page(void ** current_chead,char * scratch_buf,vm_page_t m)4372 vm_pageout_compress_page(void **current_chead, char *scratch_buf, vm_page_t m)
4373 {
4374 	vm_object_t     object;
4375 	memory_object_t pager;
4376 	int             compressed_count_delta;
4377 	kern_return_t   retval;
4378 
4379 	object = VM_PAGE_OBJECT(m);
4380 
4381 	assert(!m->vmp_free_when_done);
4382 	assert(!m->vmp_laundry);
4383 
4384 	pager = object->pager;
4385 
4386 	if (!object->pager_initialized || pager == MEMORY_OBJECT_NULL) {
4387 		KERNEL_DEBUG(0xe0400010 | DBG_FUNC_START, object, pager, 0, 0, 0);
4388 
4389 		vm_object_lock(object);
4390 
4391 		/*
4392 		 * If there is no memory object for the page, create
4393 		 * one and hand it to the compression pager.
4394 		 */
4395 
4396 		if (!object->pager_initialized) {
4397 			vm_object_collapse(object, (vm_object_offset_t) 0, TRUE);
4398 		}
4399 		if (!object->pager_initialized) {
4400 			vm_object_compressor_pager_create(object);
4401 		}
4402 
4403 		pager = object->pager;
4404 
4405 		if (!object->pager_initialized || pager == MEMORY_OBJECT_NULL) {
4406 			/*
4407 			 * Still no pager for the object,
4408 			 * or the pager has been destroyed.
4409 			 * Reactivate the page.
4410 			 *
4411 			 * Should only happen if there is no
4412 			 * compression pager
4413 			 */
4414 			PAGE_WAKEUP_DONE(m);
4415 
4416 			vm_page_lockspin_queues();
4417 			vm_page_activate(m);
4418 			VM_PAGEOUT_DEBUG(vm_pageout_dirty_no_pager, 1);
4419 			vm_page_unlock_queues();
4420 
4421 			/*
4422 			 *	And we are done with it.
4423 			 */
4424 			vm_object_activity_end(object);
4425 			vm_object_unlock(object);
4426 
4427 			return KERN_FAILURE;
4428 		}
4429 		vm_object_unlock(object);
4430 
4431 		KERNEL_DEBUG(0xe0400010 | DBG_FUNC_END, object, pager, 0, 0, 0);
4432 	}
4433 	assert(object->pager_initialized && pager != MEMORY_OBJECT_NULL);
4434 	assert(object->activity_in_progress > 0);
4435 
4436 	retval = vm_compressor_pager_put(
4437 		pager,
4438 		m->vmp_offset + object->paging_offset,
4439 		VM_PAGE_GET_PHYS_PAGE(m),
4440 		current_chead,
4441 		scratch_buf,
4442 		&compressed_count_delta);
4443 
4444 	vm_object_lock(object);
4445 
4446 	assert(object->activity_in_progress > 0);
4447 	assert(VM_PAGE_OBJECT(m) == object);
4448 	assert( !VM_PAGE_WIRED(m));
4449 
4450 	vm_compressor_pager_count(pager,
4451 	    compressed_count_delta,
4452 	    FALSE,                       /* shared_lock */
4453 	    object);
4454 
4455 	if (retval == KERN_SUCCESS) {
4456 		/*
4457 		 * If the object is purgeable, its owner's
4458 		 * purgeable ledgers will be updated in
4459 		 * vm_page_remove() but the page still
4460 		 * contributes to the owner's memory footprint,
4461 		 * so account for it as such.
4462 		 */
4463 		if ((object->purgable != VM_PURGABLE_DENY ||
4464 		    object->vo_ledger_tag) &&
4465 		    object->vo_owner != NULL) {
4466 			/* one more compressed purgeable/tagged page */
4467 			vm_object_owner_compressed_update(object,
4468 			    +1);
4469 		}
4470 		counter_inc(&vm_statistics_compressions);
4471 
4472 		if (m->vmp_tabled) {
4473 			vm_page_remove(m, TRUE);
4474 		}
4475 	} else {
4476 		PAGE_WAKEUP_DONE(m);
4477 
4478 		vm_page_lockspin_queues();
4479 
4480 		vm_page_activate(m);
4481 		vm_pageout_vminfo.vm_compressor_failed++;
4482 
4483 		vm_page_unlock_queues();
4484 	}
4485 	vm_object_activity_end(object);
4486 	vm_object_unlock(object);
4487 
4488 	return retval;
4489 }
4490 
4491 
4492 static void
vm_pageout_adjust_eq_iothrottle(struct pgo_iothread_state * ethr,boolean_t req_lowpriority)4493 vm_pageout_adjust_eq_iothrottle(struct pgo_iothread_state *ethr, boolean_t req_lowpriority)
4494 {
4495 	uint32_t        policy;
4496 
4497 	if (hibernate_cleaning_in_progress == TRUE) {
4498 		req_lowpriority = FALSE;
4499 	}
4500 
4501 	if (ethr->q->pgo_inited == TRUE && ethr->q->pgo_lowpriority != req_lowpriority) {
4502 		vm_page_unlock_queues();
4503 
4504 		if (req_lowpriority == TRUE) {
4505 			policy = THROTTLE_LEVEL_PAGEOUT_THROTTLED;
4506 			DTRACE_VM(laundrythrottle);
4507 		} else {
4508 			policy = THROTTLE_LEVEL_PAGEOUT_UNTHROTTLED;
4509 			DTRACE_VM(laundryunthrottle);
4510 		}
4511 		proc_set_thread_policy(ethr->pgo_iothread,
4512 		    TASK_POLICY_EXTERNAL, TASK_POLICY_IO, policy);
4513 
4514 		vm_page_lock_queues();
4515 		ethr->q->pgo_lowpriority = req_lowpriority;
4516 	}
4517 }
4518 
4519 OS_NORETURN
4520 static void
vm_pageout_iothread_external(struct pgo_iothread_state * ethr,__unused wait_result_t w)4521 vm_pageout_iothread_external(struct pgo_iothread_state *ethr, __unused wait_result_t w)
4522 {
4523 	thread_t        self = current_thread();
4524 
4525 	self->options |= TH_OPT_VMPRIV;
4526 
4527 	DTRACE_VM2(laundrythrottle, int, 1, (uint64_t *), NULL);
4528 
4529 	proc_set_thread_policy(self, TASK_POLICY_EXTERNAL,
4530 	    TASK_POLICY_IO, THROTTLE_LEVEL_PAGEOUT_THROTTLED);
4531 
4532 	vm_page_lock_queues();
4533 
4534 	vm_pageout_queue_external.pgo_lowpriority = TRUE;
4535 	vm_pageout_queue_external.pgo_inited = TRUE;
4536 
4537 	vm_page_unlock_queues();
4538 
4539 #if CONFIG_THREAD_GROUPS
4540 	thread_group_vm_add();
4541 #endif /* CONFIG_THREAD_GROUPS */
4542 
4543 	vm_pageout_iothread_external_continue(ethr, 0);
4544 	/*NOTREACHED*/
4545 }
4546 
4547 
4548 OS_NORETURN
4549 static void
vm_pageout_iothread_internal(struct pgo_iothread_state * cthr,__unused wait_result_t w)4550 vm_pageout_iothread_internal(struct pgo_iothread_state *cthr, __unused wait_result_t w)
4551 {
4552 	thread_t        self = current_thread();
4553 
4554 	self->options |= TH_OPT_VMPRIV;
4555 
4556 	vm_page_lock_queues();
4557 
4558 	vm_pageout_queue_internal.pgo_lowpriority = TRUE;
4559 	vm_pageout_queue_internal.pgo_inited = TRUE;
4560 
4561 #if DEVELOPMENT || DEBUG
4562 	vm_pageout_queue_benchmark.pgo_lowpriority = vm_pageout_queue_internal.pgo_lowpriority;
4563 	vm_pageout_queue_benchmark.pgo_inited = vm_pageout_queue_internal.pgo_inited;
4564 	vm_pageout_queue_benchmark.pgo_busy = FALSE;
4565 #endif /* DEVELOPMENT || DEBUG */
4566 
4567 	vm_page_unlock_queues();
4568 
4569 	if (vm_pageout_state.vm_restricted_to_single_processor == TRUE) {
4570 		thread_vm_bind_group_add();
4571 	}
4572 
4573 #if CONFIG_THREAD_GROUPS
4574 	thread_group_vm_add();
4575 #endif /* CONFIG_THREAD_GROUPS */
4576 
4577 #if __AMP__
4578 	if (vm_compressor_ebound) {
4579 		/*
4580 		 * Use the soft bound option for vm_compressor to allow it to run on
4581 		 * P-cores if E-cluster is unavailable.
4582 		 */
4583 		thread_bind_cluster_type(self, 'E', true);
4584 	}
4585 #endif /* __AMP__ */
4586 
4587 	thread_set_thread_name(current_thread(), "VM_compressor");
4588 #if DEVELOPMENT || DEBUG
4589 	vmct_stats.vmct_minpages[cthr->id] = INT32_MAX;
4590 #endif
4591 	vm_pageout_iothread_internal_continue(cthr, 0);
4592 
4593 	/*NOTREACHED*/
4594 }
4595 
4596 kern_return_t
vm_set_buffer_cleanup_callout(boolean_t (* func)(int))4597 vm_set_buffer_cleanup_callout(boolean_t (*func)(int))
4598 {
4599 	if (OSCompareAndSwapPtr(NULL, ptrauth_nop_cast(void *, func), (void * volatile *) &consider_buffer_cache_collect)) {
4600 		return KERN_SUCCESS;
4601 	} else {
4602 		return KERN_FAILURE; /* Already set */
4603 	}
4604 }
4605 
4606 extern boolean_t        memorystatus_manual_testing_on;
4607 extern unsigned int     memorystatus_level;
4608 
4609 
4610 #if VM_PRESSURE_EVENTS
4611 
4612 boolean_t vm_pressure_events_enabled = FALSE;
4613 
4614 extern uint64_t next_warning_notification_sent_at_ts;
4615 extern uint64_t next_critical_notification_sent_at_ts;
4616 
4617 #define PRESSURE_LEVEL_STUCK_THRESHOLD_MINS    (30)    /* 30 minutes. */
4618 
4619 /*
4620  * The last time there was change in pressure level OR we forced a check
4621  * because the system is stuck in a non-normal pressure level.
4622  */
4623 uint64_t  vm_pressure_last_level_transition_abs = 0;
4624 
4625 /*
4626  *  This is how the long the system waits 'stuck' in an unchanged non-normal pressure
4627  * level before resending out notifications for that level again.
4628  */
4629 int  vm_pressure_level_transition_threshold = PRESSURE_LEVEL_STUCK_THRESHOLD_MINS;
4630 
4631 void
vm_pressure_response(void)4632 vm_pressure_response(void)
4633 {
4634 	vm_pressure_level_t     old_level = kVMPressureNormal;
4635 	int                     new_level = -1;
4636 	unsigned int            total_pages;
4637 	uint64_t                available_memory = 0;
4638 	uint64_t                curr_ts, abs_time_since_level_transition, time_in_ns;
4639 	bool                    force_check = false;
4640 	int                     time_in_mins;
4641 
4642 
4643 	if (vm_pressure_events_enabled == FALSE) {
4644 		return;
4645 	}
4646 
4647 #if !XNU_TARGET_OS_OSX
4648 
4649 	available_memory = (uint64_t) memorystatus_available_pages;
4650 
4651 #else /* !XNU_TARGET_OS_OSX */
4652 
4653 	available_memory = (uint64_t) AVAILABLE_NON_COMPRESSED_MEMORY;
4654 	memorystatus_available_pages = (uint64_t) AVAILABLE_NON_COMPRESSED_MEMORY;
4655 
4656 #endif /* !XNU_TARGET_OS_OSX */
4657 
4658 	total_pages = (unsigned int) atop_64(max_mem);
4659 #if CONFIG_SECLUDED_MEMORY
4660 	total_pages -= vm_page_secluded_count;
4661 #endif /* CONFIG_SECLUDED_MEMORY */
4662 	memorystatus_level = (unsigned int) ((available_memory * 100) / total_pages);
4663 
4664 	if (memorystatus_manual_testing_on) {
4665 		return;
4666 	}
4667 
4668 	curr_ts = mach_absolute_time();
4669 	abs_time_since_level_transition = curr_ts - vm_pressure_last_level_transition_abs;
4670 
4671 	absolutetime_to_nanoseconds(abs_time_since_level_transition, &time_in_ns);
4672 	time_in_mins = (int) ((time_in_ns / NSEC_PER_SEC) / 60);
4673 	force_check = (time_in_mins >= vm_pressure_level_transition_threshold);
4674 
4675 	old_level = memorystatus_vm_pressure_level;
4676 
4677 	switch (memorystatus_vm_pressure_level) {
4678 	case kVMPressureNormal:
4679 	{
4680 		if (VM_PRESSURE_WARNING_TO_CRITICAL()) {
4681 			new_level = kVMPressureCritical;
4682 		} else if (VM_PRESSURE_NORMAL_TO_WARNING()) {
4683 			new_level = kVMPressureWarning;
4684 		}
4685 		break;
4686 	}
4687 
4688 	case kVMPressureWarning:
4689 	case kVMPressureUrgent:
4690 	{
4691 		if (VM_PRESSURE_WARNING_TO_NORMAL()) {
4692 			new_level = kVMPressureNormal;
4693 		} else if (VM_PRESSURE_WARNING_TO_CRITICAL()) {
4694 			new_level = kVMPressureCritical;
4695 		} else if (force_check) {
4696 			new_level = kVMPressureWarning;
4697 			next_warning_notification_sent_at_ts = curr_ts;
4698 		}
4699 		break;
4700 	}
4701 
4702 	case kVMPressureCritical:
4703 	{
4704 		if (VM_PRESSURE_WARNING_TO_NORMAL()) {
4705 			new_level = kVMPressureNormal;
4706 		} else if (VM_PRESSURE_CRITICAL_TO_WARNING()) {
4707 			new_level = kVMPressureWarning;
4708 		} else if (force_check) {
4709 			new_level = kVMPressureCritical;
4710 			next_critical_notification_sent_at_ts = curr_ts;
4711 		}
4712 		break;
4713 	}
4714 
4715 	default:
4716 		return;
4717 	}
4718 
4719 	if (new_level != -1 || force_check) {
4720 		if (new_level != -1) {
4721 			memorystatus_vm_pressure_level = (vm_pressure_level_t) new_level;
4722 
4723 			if (new_level != (int) old_level) {
4724 				VM_DEBUG_CONSTANT_EVENT(vm_pressure_level_change, VM_PRESSURE_LEVEL_CHANGE, DBG_FUNC_NONE,
4725 				    new_level, old_level, 0, 0);
4726 			}
4727 		} else {
4728 			VM_DEBUG_CONSTANT_EVENT(vm_pressure_level_change, VM_PRESSURE_LEVEL_CHANGE, DBG_FUNC_NONE,
4729 			    new_level, old_level, force_check, 0);
4730 		}
4731 
4732 		if (hibernation_vmqueues_inspection || hibernate_cleaning_in_progress) {
4733 			/*
4734 			 * We don't want to schedule a wakeup while hibernation is in progress
4735 			 * because that could collide with checks for non-monotonicity in the scheduler.
4736 			 * We do however do all the updates to memorystatus_vm_pressure_level because
4737 			 * we _might_ want to use that for decisions regarding which pages or how
4738 			 * many pages we want to dump in hibernation.
4739 			 */
4740 			return;
4741 		}
4742 
4743 		if ((memorystatus_vm_pressure_level != kVMPressureNormal) || (old_level != memorystatus_vm_pressure_level) || force_check) {
4744 			if (vm_pageout_state.vm_pressure_thread_running == FALSE) {
4745 				thread_wakeup(&vm_pressure_thread);
4746 			}
4747 
4748 			if (old_level != memorystatus_vm_pressure_level) {
4749 				thread_wakeup(&vm_pageout_state.vm_pressure_changed);
4750 			}
4751 			vm_pressure_last_level_transition_abs = curr_ts; /* renew the window of observation for a stuck pressure level */
4752 		}
4753 	}
4754 }
4755 #endif /* VM_PRESSURE_EVENTS */
4756 
4757 /*
4758  * Function called by a kernel thread to either get the current pressure level or
4759  * wait until memory pressure changes from a given level.
4760  */
4761 kern_return_t
mach_vm_pressure_level_monitor(__unused boolean_t wait_for_pressure,__unused unsigned int * pressure_level)4762 mach_vm_pressure_level_monitor(__unused boolean_t wait_for_pressure, __unused unsigned int *pressure_level)
4763 {
4764 #if !VM_PRESSURE_EVENTS
4765 
4766 	return KERN_FAILURE;
4767 
4768 #else /* VM_PRESSURE_EVENTS */
4769 
4770 	wait_result_t       wr = 0;
4771 	vm_pressure_level_t old_level = memorystatus_vm_pressure_level;
4772 
4773 	if (pressure_level == NULL) {
4774 		return KERN_INVALID_ARGUMENT;
4775 	}
4776 
4777 	if (*pressure_level == kVMPressureJetsam) {
4778 		if (!wait_for_pressure) {
4779 			return KERN_INVALID_ARGUMENT;
4780 		}
4781 
4782 		lck_mtx_lock(&memorystatus_jetsam_fg_band_lock);
4783 		wr = assert_wait((event_t)&memorystatus_jetsam_fg_band_waiters,
4784 		    THREAD_INTERRUPTIBLE);
4785 		if (wr == THREAD_WAITING) {
4786 			++memorystatus_jetsam_fg_band_waiters;
4787 			lck_mtx_unlock(&memorystatus_jetsam_fg_band_lock);
4788 			wr = thread_block(THREAD_CONTINUE_NULL);
4789 		} else {
4790 			lck_mtx_unlock(&memorystatus_jetsam_fg_band_lock);
4791 		}
4792 		if (wr != THREAD_AWAKENED) {
4793 			return KERN_ABORTED;
4794 		}
4795 		*pressure_level = kVMPressureJetsam;
4796 		return KERN_SUCCESS;
4797 	}
4798 
4799 	if (wait_for_pressure == TRUE) {
4800 		while (old_level == *pressure_level) {
4801 			wr = assert_wait((event_t) &vm_pageout_state.vm_pressure_changed,
4802 			    THREAD_INTERRUPTIBLE);
4803 			if (wr == THREAD_WAITING) {
4804 				wr = thread_block(THREAD_CONTINUE_NULL);
4805 			}
4806 			if (wr == THREAD_INTERRUPTED) {
4807 				return KERN_ABORTED;
4808 			}
4809 
4810 			if (wr == THREAD_AWAKENED) {
4811 				old_level = memorystatus_vm_pressure_level;
4812 			}
4813 		}
4814 	}
4815 
4816 	*pressure_level = old_level;
4817 	return KERN_SUCCESS;
4818 #endif /* VM_PRESSURE_EVENTS */
4819 }
4820 
4821 #if VM_PRESSURE_EVENTS
4822 void
vm_pressure_thread(void)4823 vm_pressure_thread(void)
4824 {
4825 	static boolean_t thread_initialized = FALSE;
4826 
4827 	if (thread_initialized == TRUE) {
4828 		vm_pageout_state.vm_pressure_thread_running = TRUE;
4829 		consider_vm_pressure_events();
4830 		vm_pageout_state.vm_pressure_thread_running = FALSE;
4831 	}
4832 
4833 #if CONFIG_THREAD_GROUPS
4834 	thread_group_vm_add();
4835 #endif /* CONFIG_THREAD_GROUPS */
4836 
4837 	thread_set_thread_name(current_thread(), "VM_pressure");
4838 	thread_initialized = TRUE;
4839 	assert_wait((event_t) &vm_pressure_thread, THREAD_UNINT);
4840 	thread_block((thread_continue_t)vm_pressure_thread);
4841 }
4842 #endif /* VM_PRESSURE_EVENTS */
4843 
4844 
4845 /*
4846  * called once per-second via "compute_averages"
4847  */
4848 void
compute_pageout_gc_throttle(__unused void * arg)4849 compute_pageout_gc_throttle(__unused void *arg)
4850 {
4851 	if (vm_pageout_vminfo.vm_pageout_considered_page != vm_pageout_state.vm_pageout_considered_page_last) {
4852 		vm_pageout_state.vm_pageout_considered_page_last = vm_pageout_vminfo.vm_pageout_considered_page;
4853 
4854 		thread_wakeup(VM_PAGEOUT_GC_EVENT);
4855 	}
4856 }
4857 
4858 /*
4859  * vm_pageout_garbage_collect can also be called when the zone allocator needs
4860  * to call zone_gc on a different thread in order to trigger zone-map-exhaustion
4861  * jetsams. We need to check if the zone map size is above its jetsam limit to
4862  * decide if this was indeed the case.
4863  *
4864  * We need to do this on a different thread because of the following reasons:
4865  *
4866  * 1. In the case of synchronous jetsams, the leaking process can try to jetsam
4867  * itself causing the system to hang. We perform synchronous jetsams if we're
4868  * leaking in the VM map entries zone, so the leaking process could be doing a
4869  * zalloc for a VM map entry while holding its vm_map lock, when it decides to
4870  * jetsam itself. We also need the vm_map lock on the process termination path,
4871  * which would now lead the dying process to deadlock against itself.
4872  *
4873  * 2. The jetsam path might need to allocate zone memory itself. We could try
4874  * using the non-blocking variant of zalloc for this path, but we can still
4875  * end up trying to do a kmem_alloc when the zone maps are almost full.
4876  */
4877 __dead2
4878 void
vm_pageout_garbage_collect(void * step,wait_result_t wr __unused)4879 vm_pageout_garbage_collect(void *step, wait_result_t wr __unused)
4880 {
4881 	assert(step == VM_PAGEOUT_GC_INIT || step == VM_PAGEOUT_GC_COLLECT);
4882 
4883 	if (step == VM_PAGEOUT_GC_INIT) {
4884 		/* first time being called is not about GC */
4885 #if CONFIG_THREAD_GROUPS
4886 		thread_group_vm_add();
4887 #endif /* CONFIG_THREAD_GROUPS */
4888 	} else if (zone_map_nearing_exhaustion()) {
4889 		/*
4890 		 * Woken up by the zone allocator for zone-map-exhaustion jetsams.
4891 		 *
4892 		 * Bail out after calling zone_gc (which triggers the
4893 		 * zone-map-exhaustion jetsams). If we fall through, the subsequent
4894 		 * operations that clear out a bunch of caches might allocate zone
4895 		 * memory themselves (for eg. vm_map operations would need VM map
4896 		 * entries). Since the zone map is almost full at this point, we
4897 		 * could end up with a panic. We just need to quickly jetsam a
4898 		 * process and exit here.
4899 		 *
4900 		 * It could so happen that we were woken up to relieve memory
4901 		 * pressure and the zone map also happened to be near its limit at
4902 		 * the time, in which case we'll skip out early. But that should be
4903 		 * ok; if memory pressure persists, the thread will simply be woken
4904 		 * up again.
4905 		 */
4906 		zone_gc(ZONE_GC_JETSAM);
4907 	} else {
4908 		/* Woken up by vm_pageout_scan or compute_pageout_gc_throttle. */
4909 		boolean_t buf_large_zfree = FALSE;
4910 		boolean_t first_try = TRUE;
4911 
4912 		stack_collect();
4913 
4914 		consider_machine_collect();
4915 		mbuf_drain(FALSE);
4916 
4917 		do {
4918 			if (consider_buffer_cache_collect != NULL) {
4919 				buf_large_zfree = (*consider_buffer_cache_collect)(0);
4920 			}
4921 			if (first_try == TRUE || buf_large_zfree == TRUE) {
4922 				/*
4923 				 * zone_gc should be last, because the other operations
4924 				 * might return memory to zones.
4925 				 */
4926 				zone_gc(ZONE_GC_TRIM);
4927 			}
4928 			first_try = FALSE;
4929 		} while (buf_large_zfree == TRUE && vm_page_free_count < vm_page_free_target);
4930 
4931 		consider_machine_adjust();
4932 	}
4933 
4934 	assert_wait(VM_PAGEOUT_GC_EVENT, THREAD_UNINT);
4935 
4936 	thread_block_parameter(vm_pageout_garbage_collect, VM_PAGEOUT_GC_COLLECT);
4937 	__builtin_unreachable();
4938 }
4939 
4940 
4941 #if VM_PAGE_BUCKETS_CHECK
4942 #if VM_PAGE_FAKE_BUCKETS
4943 extern vm_map_offset_t vm_page_fake_buckets_start, vm_page_fake_buckets_end;
4944 #endif /* VM_PAGE_FAKE_BUCKETS */
4945 #endif /* VM_PAGE_BUCKETS_CHECK */
4946 
4947 
4948 
4949 void
vm_set_restrictions(unsigned int num_cpus)4950 vm_set_restrictions(unsigned int num_cpus)
4951 {
4952 	int vm_restricted_to_single_processor = 0;
4953 
4954 	if (PE_parse_boot_argn("vm_restricted_to_single_processor", &vm_restricted_to_single_processor, sizeof(vm_restricted_to_single_processor))) {
4955 		kprintf("Overriding vm_restricted_to_single_processor to %d\n", vm_restricted_to_single_processor);
4956 		vm_pageout_state.vm_restricted_to_single_processor = (vm_restricted_to_single_processor ? TRUE : FALSE);
4957 	} else {
4958 		assert(num_cpus > 0);
4959 
4960 		if (num_cpus <= 3) {
4961 			/*
4962 			 * on systems with a limited number of CPUS, bind the
4963 			 * 4 major threads that can free memory and that tend to use
4964 			 * a fair bit of CPU under pressured conditions to a single processor.
4965 			 * This insures that these threads don't hog all of the available CPUs
4966 			 * (important for camera launch), while allowing them to run independently
4967 			 * w/r to locks... the 4 threads are
4968 			 * vm_pageout_scan,  vm_pageout_iothread_internal (compressor),
4969 			 * vm_compressor_swap_trigger_thread (minor and major compactions),
4970 			 * memorystatus_thread (jetsams).
4971 			 *
4972 			 * the first time the thread is run, it is responsible for checking the
4973 			 * state of vm_restricted_to_single_processor, and if TRUE it calls
4974 			 * thread_bind_master...  someday this should be replaced with a group
4975 			 * scheduling mechanism and KPI.
4976 			 */
4977 			vm_pageout_state.vm_restricted_to_single_processor = TRUE;
4978 		} else {
4979 			vm_pageout_state.vm_restricted_to_single_processor = FALSE;
4980 		}
4981 	}
4982 }
4983 
4984 /*
4985  * Set up vm_config based on the vm_compressor_mode.
4986  * Must run BEFORE the pageout thread starts up.
4987  */
4988 __startup_func
4989 void
vm_config_init(void)4990 vm_config_init(void)
4991 {
4992 	bzero(&vm_config, sizeof(vm_config));
4993 
4994 	switch (vm_compressor_mode) {
4995 	case VM_PAGER_DEFAULT:
4996 		printf("mapping deprecated VM_PAGER_DEFAULT to VM_PAGER_COMPRESSOR_WITH_SWAP\n");
4997 		OS_FALLTHROUGH;
4998 
4999 	case VM_PAGER_COMPRESSOR_WITH_SWAP:
5000 		vm_config.compressor_is_present = TRUE;
5001 		vm_config.swap_is_present = TRUE;
5002 		vm_config.compressor_is_active = TRUE;
5003 		vm_config.swap_is_active = TRUE;
5004 		break;
5005 
5006 	case VM_PAGER_COMPRESSOR_NO_SWAP:
5007 		vm_config.compressor_is_present = TRUE;
5008 		vm_config.swap_is_present = TRUE;
5009 		vm_config.compressor_is_active = TRUE;
5010 		break;
5011 
5012 	case VM_PAGER_FREEZER_DEFAULT:
5013 		printf("mapping deprecated VM_PAGER_FREEZER_DEFAULT to VM_PAGER_FREEZER_COMPRESSOR_NO_SWAP\n");
5014 		OS_FALLTHROUGH;
5015 
5016 	case VM_PAGER_FREEZER_COMPRESSOR_NO_SWAP:
5017 		vm_config.compressor_is_present = TRUE;
5018 		vm_config.swap_is_present = TRUE;
5019 		break;
5020 
5021 	case VM_PAGER_COMPRESSOR_NO_SWAP_PLUS_FREEZER_COMPRESSOR_WITH_SWAP:
5022 		vm_config.compressor_is_present = TRUE;
5023 		vm_config.swap_is_present = TRUE;
5024 		vm_config.compressor_is_active = TRUE;
5025 		vm_config.freezer_swap_is_active = TRUE;
5026 		break;
5027 
5028 	case VM_PAGER_NOT_CONFIGURED:
5029 		break;
5030 
5031 	default:
5032 		printf("unknown compressor mode - %x\n", vm_compressor_mode);
5033 		break;
5034 	}
5035 }
5036 
5037 __startup_func
5038 static void
vm_pageout_create_gc_thread(void)5039 vm_pageout_create_gc_thread(void)
5040 {
5041 	thread_t thread;
5042 
5043 	if (kernel_thread_create(vm_pageout_garbage_collect,
5044 	    VM_PAGEOUT_GC_INIT, BASEPRI_DEFAULT, &thread) != KERN_SUCCESS) {
5045 		panic("vm_pageout_garbage_collect: create failed");
5046 	}
5047 	thread_set_thread_name(thread, "VM_pageout_garbage_collect");
5048 	if (thread->reserved_stack == 0) {
5049 		assert(thread->kernel_stack);
5050 		thread->reserved_stack = thread->kernel_stack;
5051 	}
5052 
5053 	/* thread is started in vm_pageout() */
5054 	vm_pageout_gc_thread = thread;
5055 }
5056 STARTUP(EARLY_BOOT, STARTUP_RANK_MIDDLE, vm_pageout_create_gc_thread);
5057 
5058 void
vm_pageout(void)5059 vm_pageout(void)
5060 {
5061 	thread_t        self = current_thread();
5062 	thread_t        thread;
5063 	kern_return_t   result;
5064 	spl_t           s;
5065 
5066 	/*
5067 	 * Set thread privileges.
5068 	 */
5069 	s = splsched();
5070 
5071 #if CONFIG_VPS_DYNAMIC_PRIO
5072 
5073 	int             vps_dynprio_bootarg = 0;
5074 
5075 	if (PE_parse_boot_argn("vps_dynamic_priority_enabled", &vps_dynprio_bootarg, sizeof(vps_dynprio_bootarg))) {
5076 		vps_dynamic_priority_enabled = (vps_dynprio_bootarg ? TRUE : FALSE);
5077 		kprintf("Overriding vps_dynamic_priority_enabled to %d\n", vps_dynamic_priority_enabled);
5078 	} else {
5079 		if (vm_pageout_state.vm_restricted_to_single_processor == TRUE) {
5080 			vps_dynamic_priority_enabled = TRUE;
5081 		} else {
5082 			vps_dynamic_priority_enabled = FALSE;
5083 		}
5084 	}
5085 
5086 	if (vps_dynamic_priority_enabled) {
5087 		sched_set_kernel_thread_priority(self, MAXPRI_THROTTLE);
5088 		thread_set_eager_preempt(self);
5089 	} else {
5090 		sched_set_kernel_thread_priority(self, BASEPRI_VM);
5091 	}
5092 
5093 #else /* CONFIG_VPS_DYNAMIC_PRIO */
5094 
5095 	vps_dynamic_priority_enabled = FALSE;
5096 	sched_set_kernel_thread_priority(self, BASEPRI_VM);
5097 
5098 #endif /* CONFIG_VPS_DYNAMIC_PRIO */
5099 
5100 	thread_lock(self);
5101 	self->options |= TH_OPT_VMPRIV;
5102 	thread_unlock(self);
5103 
5104 	if (!self->reserved_stack) {
5105 		self->reserved_stack = self->kernel_stack;
5106 	}
5107 
5108 	if (vm_pageout_state.vm_restricted_to_single_processor == TRUE &&
5109 	    vps_dynamic_priority_enabled == FALSE) {
5110 		thread_vm_bind_group_add();
5111 	}
5112 
5113 
5114 #if CONFIG_THREAD_GROUPS
5115 	thread_group_vm_add();
5116 #endif /* CONFIG_THREAD_GROUPS */
5117 
5118 #if __AMP__
5119 	PE_parse_boot_argn("vmpgo_pcluster", &vm_pgo_pbound, sizeof(vm_pgo_pbound));
5120 	if (vm_pgo_pbound) {
5121 		/*
5122 		 * Use the soft bound option for vm pageout to allow it to run on
5123 		 * E-cores if P-cluster is unavailable.
5124 		 */
5125 		thread_bind_cluster_type(self, 'P', true);
5126 	}
5127 #endif /* __AMP__ */
5128 
5129 	PE_parse_boot_argn("vmpgo_protect_realtime",
5130 	    &vm_pageout_protect_realtime,
5131 	    sizeof(vm_pageout_protect_realtime));
5132 	splx(s);
5133 
5134 	thread_set_thread_name(current_thread(), "VM_pageout_scan");
5135 
5136 	/*
5137 	 *	Initialize some paging parameters.
5138 	 */
5139 
5140 	vm_pageout_state.vm_pressure_thread_running = FALSE;
5141 	vm_pageout_state.vm_pressure_changed = FALSE;
5142 	vm_pageout_state.memorystatus_purge_on_warning = 2;
5143 	vm_pageout_state.memorystatus_purge_on_urgent = 5;
5144 	vm_pageout_state.memorystatus_purge_on_critical = 8;
5145 	vm_pageout_state.vm_page_speculative_q_age_ms = VM_PAGE_SPECULATIVE_Q_AGE_MS;
5146 	vm_pageout_state.vm_page_speculative_percentage = 5;
5147 	vm_pageout_state.vm_page_speculative_target = 0;
5148 
5149 	vm_pageout_state.vm_pageout_swap_wait = 0;
5150 	vm_pageout_state.vm_pageout_idle_wait = 0;
5151 	vm_pageout_state.vm_pageout_empty_wait = 0;
5152 	vm_pageout_state.vm_pageout_burst_wait = 0;
5153 	vm_pageout_state.vm_pageout_deadlock_wait = 0;
5154 	vm_pageout_state.vm_pageout_deadlock_relief = 0;
5155 	vm_pageout_state.vm_pageout_burst_inactive_throttle = 0;
5156 
5157 	vm_pageout_state.vm_pageout_inactive = 0;
5158 	vm_pageout_state.vm_pageout_inactive_used = 0;
5159 	vm_pageout_state.vm_pageout_inactive_clean = 0;
5160 
5161 	vm_pageout_state.vm_memory_pressure = 0;
5162 	vm_pageout_state.vm_page_filecache_min = 0;
5163 #if CONFIG_JETSAM
5164 	vm_pageout_state.vm_page_filecache_min_divisor = 70;
5165 	vm_pageout_state.vm_page_xpmapped_min_divisor = 40;
5166 #else
5167 	vm_pageout_state.vm_page_filecache_min_divisor = 27;
5168 	vm_pageout_state.vm_page_xpmapped_min_divisor = 36;
5169 #endif
5170 	vm_pageout_state.vm_page_free_count_init = vm_page_free_count;
5171 
5172 	vm_pageout_state.vm_pageout_considered_page_last = 0;
5173 
5174 	if (vm_pageout_state.vm_pageout_swap_wait == 0) {
5175 		vm_pageout_state.vm_pageout_swap_wait = VM_PAGEOUT_SWAP_WAIT;
5176 	}
5177 
5178 	if (vm_pageout_state.vm_pageout_idle_wait == 0) {
5179 		vm_pageout_state.vm_pageout_idle_wait = VM_PAGEOUT_IDLE_WAIT;
5180 	}
5181 
5182 	if (vm_pageout_state.vm_pageout_burst_wait == 0) {
5183 		vm_pageout_state.vm_pageout_burst_wait = VM_PAGEOUT_BURST_WAIT;
5184 	}
5185 
5186 	if (vm_pageout_state.vm_pageout_empty_wait == 0) {
5187 		vm_pageout_state.vm_pageout_empty_wait = VM_PAGEOUT_EMPTY_WAIT;
5188 	}
5189 
5190 	if (vm_pageout_state.vm_pageout_deadlock_wait == 0) {
5191 		vm_pageout_state.vm_pageout_deadlock_wait = VM_PAGEOUT_DEADLOCK_WAIT;
5192 	}
5193 
5194 	if (vm_pageout_state.vm_pageout_deadlock_relief == 0) {
5195 		vm_pageout_state.vm_pageout_deadlock_relief = VM_PAGEOUT_DEADLOCK_RELIEF;
5196 	}
5197 
5198 	if (vm_pageout_state.vm_pageout_burst_inactive_throttle == 0) {
5199 		vm_pageout_state.vm_pageout_burst_inactive_throttle = VM_PAGEOUT_BURST_INACTIVE_THROTTLE;
5200 	}
5201 	/*
5202 	 * even if we've already called vm_page_free_reserve
5203 	 * call it again here to insure that the targets are
5204 	 * accurately calculated (it uses vm_page_free_count_init)
5205 	 * calling it with an arg of 0 will not change the reserve
5206 	 * but will re-calculate free_min and free_target
5207 	 */
5208 	if (vm_page_free_reserved < VM_PAGE_FREE_RESERVED(processor_count)) {
5209 		vm_page_free_reserve((VM_PAGE_FREE_RESERVED(processor_count)) - vm_page_free_reserved);
5210 	} else {
5211 		vm_page_free_reserve(0);
5212 	}
5213 
5214 	bzero(&vm_pageout_queue_external, sizeof(struct vm_pageout_queue));
5215 	bzero(&vm_pageout_queue_internal, sizeof(struct vm_pageout_queue));
5216 
5217 	vm_page_queue_init(&vm_pageout_queue_external.pgo_pending);
5218 	vm_pageout_queue_external.pgo_maxlaundry = VM_PAGE_LAUNDRY_MAX;
5219 
5220 	vm_page_queue_init(&vm_pageout_queue_internal.pgo_pending);
5221 
5222 #if DEVELOPMENT || DEBUG
5223 	bzero(&vm_pageout_queue_benchmark, sizeof(struct vm_pageout_queue));
5224 	vm_page_queue_init(&vm_pageout_queue_benchmark.pgo_pending);
5225 #endif /* DEVELOPMENT || DEBUG */
5226 
5227 
5228 	/* internal pageout thread started when default pager registered first time */
5229 	/* external pageout and garbage collection threads started here */
5230 	struct pgo_iothread_state *ethr = &pgo_iothread_external_state;
5231 	ethr->id = 0;
5232 	ethr->q = &vm_pageout_queue_external;
5233 	ethr->current_early_swapout_chead = NULL;
5234 	ethr->current_regular_swapout_chead = NULL;
5235 	ethr->current_late_swapout_chead = NULL;
5236 	ethr->scratch_buf = NULL;
5237 #if DEVELOPMENT || DEBUG
5238 	ethr->benchmark_q = NULL;
5239 #endif /* DEVELOPMENT || DEBUG */
5240 	sched_cond_init(&(ethr->pgo_wakeup));
5241 
5242 	result = kernel_thread_start_priority((thread_continue_t)vm_pageout_iothread_external,
5243 	    (void *)ethr, BASEPRI_VM,
5244 	    &(ethr->pgo_iothread));
5245 	if (result != KERN_SUCCESS) {
5246 		panic("vm_pageout: Unable to create external thread (%d)\n", result);
5247 	}
5248 	thread_set_thread_name(ethr->pgo_iothread, "VM_pageout_external_iothread");
5249 
5250 	thread_mtx_lock(vm_pageout_gc_thread );
5251 	thread_start(vm_pageout_gc_thread );
5252 	thread_mtx_unlock(vm_pageout_gc_thread);
5253 
5254 #if VM_PRESSURE_EVENTS
5255 	result = kernel_thread_start_priority((thread_continue_t)vm_pressure_thread, NULL,
5256 	    BASEPRI_DEFAULT,
5257 	    &thread);
5258 
5259 	if (result != KERN_SUCCESS) {
5260 		panic("vm_pressure_thread: create failed");
5261 	}
5262 
5263 	thread_deallocate(thread);
5264 #endif
5265 
5266 	vm_object_reaper_init();
5267 
5268 
5269 	if (VM_CONFIG_COMPRESSOR_IS_PRESENT) {
5270 		vm_compressor_init();
5271 	}
5272 
5273 #if VM_PRESSURE_EVENTS
5274 	vm_pressure_events_enabled = TRUE;
5275 #endif /* VM_PRESSURE_EVENTS */
5276 
5277 #if CONFIG_PHANTOM_CACHE
5278 	vm_phantom_cache_init();
5279 #endif
5280 #if VM_PAGE_BUCKETS_CHECK
5281 #if VM_PAGE_FAKE_BUCKETS
5282 	printf("**** DEBUG: protecting fake buckets [0x%llx:0x%llx]\n",
5283 	    (uint64_t) vm_page_fake_buckets_start,
5284 	    (uint64_t) vm_page_fake_buckets_end);
5285 	pmap_protect(kernel_pmap,
5286 	    vm_page_fake_buckets_start,
5287 	    vm_page_fake_buckets_end,
5288 	    VM_PROT_READ);
5289 //	*(char *) vm_page_fake_buckets_start = 'x';	/* panic! */
5290 #endif /* VM_PAGE_FAKE_BUCKETS */
5291 #endif /* VM_PAGE_BUCKETS_CHECK */
5292 
5293 #if VM_OBJECT_TRACKING
5294 	vm_object_tracking_init();
5295 #endif /* VM_OBJECT_TRACKING */
5296 
5297 #if __arm64__
5298 //	vm_tests();
5299 #endif /* __arm64__ */
5300 
5301 	vm_pageout_continue();
5302 
5303 	/*
5304 	 * Unreached code!
5305 	 *
5306 	 * The vm_pageout_continue() call above never returns, so the code below is never
5307 	 * executed.  We take advantage of this to declare several DTrace VM related probe
5308 	 * points that our kernel doesn't have an analog for.  These are probe points that
5309 	 * exist in Solaris and are in the DTrace documentation, so people may have written
5310 	 * scripts that use them.  Declaring the probe points here means their scripts will
5311 	 * compile and execute which we want for portability of the scripts, but since this
5312 	 * section of code is never reached, the probe points will simply never fire.  Yes,
5313 	 * this is basically a hack.  The problem is the DTrace probe points were chosen with
5314 	 * Solaris specific VM events in mind, not portability to different VM implementations.
5315 	 */
5316 
5317 	DTRACE_VM2(execfree, int, 1, (uint64_t *), NULL);
5318 	DTRACE_VM2(execpgin, int, 1, (uint64_t *), NULL);
5319 	DTRACE_VM2(execpgout, int, 1, (uint64_t *), NULL);
5320 	DTRACE_VM2(pgswapin, int, 1, (uint64_t *), NULL);
5321 	DTRACE_VM2(pgswapout, int, 1, (uint64_t *), NULL);
5322 	DTRACE_VM2(swapin, int, 1, (uint64_t *), NULL);
5323 	DTRACE_VM2(swapout, int, 1, (uint64_t *), NULL);
5324 	/*NOTREACHED*/
5325 }
5326 
5327 
5328 
5329 kern_return_t
vm_pageout_internal_start(void)5330 vm_pageout_internal_start(void)
5331 {
5332 	kern_return_t   result = KERN_SUCCESS;
5333 	host_basic_info_data_t hinfo;
5334 	vm_offset_t     buf, bufsize;
5335 
5336 	assert(VM_CONFIG_COMPRESSOR_IS_PRESENT);
5337 
5338 	mach_msg_type_number_t count = HOST_BASIC_INFO_COUNT;
5339 #define BSD_HOST 1
5340 	host_info((host_t)BSD_HOST, HOST_BASIC_INFO, (host_info_t)&hinfo, &count);
5341 
5342 	assert(hinfo.max_cpus > 0);
5343 
5344 #if !XNU_TARGET_OS_OSX
5345 	vm_pageout_state.vm_compressor_thread_count = 1;
5346 #else /* !XNU_TARGET_OS_OSX */
5347 	if (hinfo.max_cpus > 4) {
5348 		vm_pageout_state.vm_compressor_thread_count = 2;
5349 	} else {
5350 		vm_pageout_state.vm_compressor_thread_count = 1;
5351 	}
5352 #endif /* !XNU_TARGET_OS_OSX */
5353 #if     __AMP__
5354 	if (vm_compressor_ebound) {
5355 		vm_pageout_state.vm_compressor_thread_count = 2;
5356 	}
5357 #endif
5358 	PE_parse_boot_argn("vmcomp_threads", &vm_pageout_state.vm_compressor_thread_count,
5359 	    sizeof(vm_pageout_state.vm_compressor_thread_count));
5360 
5361 	if (vm_pageout_state.vm_compressor_thread_count >= hinfo.max_cpus) {
5362 		vm_pageout_state.vm_compressor_thread_count = hinfo.max_cpus - 1;
5363 	}
5364 	if (vm_pageout_state.vm_compressor_thread_count <= 0) {
5365 		vm_pageout_state.vm_compressor_thread_count = 1;
5366 	} else if (vm_pageout_state.vm_compressor_thread_count > MAX_COMPRESSOR_THREAD_COUNT) {
5367 		vm_pageout_state.vm_compressor_thread_count = MAX_COMPRESSOR_THREAD_COUNT;
5368 	}
5369 
5370 	vm_pageout_queue_internal.pgo_maxlaundry =
5371 	    (vm_pageout_state.vm_compressor_thread_count * 4) * VM_PAGE_LAUNDRY_MAX;
5372 
5373 	PE_parse_boot_argn("vmpgoi_maxlaundry",
5374 	    &vm_pageout_queue_internal.pgo_maxlaundry,
5375 	    sizeof(vm_pageout_queue_internal.pgo_maxlaundry));
5376 
5377 #if DEVELOPMENT || DEBUG
5378 	// Note: this will be modified at enqueue-time such that the benchmark queue is never throttled
5379 	vm_pageout_queue_benchmark.pgo_maxlaundry = vm_pageout_queue_internal.pgo_maxlaundry;
5380 #endif /* DEVELOPMENT || DEBUG */
5381 
5382 	bufsize = COMPRESSOR_SCRATCH_BUF_SIZE;
5383 
5384 	kmem_alloc(kernel_map, &buf,
5385 	    bufsize * vm_pageout_state.vm_compressor_thread_count,
5386 	    KMA_DATA | KMA_NOFAIL | KMA_KOBJECT | KMA_PERMANENT,
5387 	    VM_KERN_MEMORY_COMPRESSOR);
5388 
5389 	for (int i = 0; i < vm_pageout_state.vm_compressor_thread_count; i++) {
5390 		struct pgo_iothread_state *iq = &pgo_iothread_internal_state[i];
5391 		iq->id = i;
5392 		iq->q = &vm_pageout_queue_internal;
5393 		iq->current_early_swapout_chead = NULL;
5394 		iq->current_regular_swapout_chead = NULL;
5395 		iq->current_late_swapout_chead = NULL;
5396 		iq->scratch_buf = (char *)(buf + i * bufsize);
5397 #if DEVELOPMENT || DEBUG
5398 		iq->benchmark_q = &vm_pageout_queue_benchmark;
5399 #endif /* DEVELOPMENT || DEBUG */
5400 		sched_cond_init(&(iq->pgo_wakeup));
5401 		result = kernel_thread_start_priority((thread_continue_t)vm_pageout_iothread_internal,
5402 		    (void *)iq, BASEPRI_VM,
5403 		    &(iq->pgo_iothread));
5404 
5405 		if (result != KERN_SUCCESS) {
5406 			panic("vm_pageout: Unable to create compressor thread no. %d (%d)\n", i, result);
5407 		}
5408 	}
5409 	return result;
5410 }
5411 
5412 #if CONFIG_IOSCHED
5413 /*
5414  * To support I/O Expedite for compressed files we mark the upls with special flags.
5415  * The way decmpfs works is that we create a big upl which marks all the pages needed to
5416  * represent the compressed file as busy. We tag this upl with the flag UPL_DECMP_REQ. Decmpfs
5417  * then issues smaller I/Os for compressed I/Os, deflates them and puts the data into the pages
5418  * being held in the big original UPL. We mark each of these smaller UPLs with the flag
5419  * UPL_DECMP_REAL_IO. Any outstanding real I/O UPL is tracked by the big req upl using the
5420  * decmp_io_upl field (in the upl structure). This link is protected in the forward direction
5421  * by the req upl lock (the reverse link doesnt need synch. since we never inspect this link
5422  * unless the real I/O upl is being destroyed).
5423  */
5424 
5425 
5426 static void
upl_set_decmp_info(upl_t upl,upl_t src_upl)5427 upl_set_decmp_info(upl_t upl, upl_t src_upl)
5428 {
5429 	assert((src_upl->flags & UPL_DECMP_REQ) != 0);
5430 
5431 	upl_lock(src_upl);
5432 	if (src_upl->decmp_io_upl) {
5433 		/*
5434 		 * If there is already an alive real I/O UPL, ignore this new UPL.
5435 		 * This case should rarely happen and even if it does, it just means
5436 		 * that we might issue a spurious expedite which the driver is expected
5437 		 * to handle.
5438 		 */
5439 		upl_unlock(src_upl);
5440 		return;
5441 	}
5442 	src_upl->decmp_io_upl = (void *)upl;
5443 	src_upl->ref_count++;
5444 
5445 	upl->flags |= UPL_DECMP_REAL_IO;
5446 	upl->decmp_io_upl = (void *)src_upl;
5447 	upl_unlock(src_upl);
5448 }
5449 #endif /* CONFIG_IOSCHED */
5450 
5451 #if UPL_DEBUG
5452 int     upl_debug_enabled = 1;
5453 #else
5454 int     upl_debug_enabled = 0;
5455 #endif
5456 
5457 static upl_t
upl_create(int type,int flags,upl_size_t size)5458 upl_create(int type, int flags, upl_size_t size)
5459 {
5460 	uint32_t pages = (uint32_t)atop(round_page_32(size));
5461 	upl_t    upl;
5462 
5463 	assert(page_aligned(size));
5464 
5465 	/*
5466 	 * FIXME: this code assumes the allocation always succeeds,
5467 	 *        however `pages` can be up to MAX_UPL_SIZE.
5468 	 *
5469 	 *        The allocation size is above 32k (resp. 128k)
5470 	 *        on 16k pages (resp. 4k), which kalloc might fail
5471 	 *        to allocate.
5472 	 */
5473 	upl = kalloc_type(struct upl, struct upl_page_info,
5474 	    (type & UPL_CREATE_INTERNAL) ? pages : 0, Z_WAITOK | Z_ZERO);
5475 	if (type & UPL_CREATE_INTERNAL) {
5476 		flags |= UPL_INTERNAL;
5477 	}
5478 
5479 	if (type & UPL_CREATE_LITE) {
5480 		flags |= UPL_LITE;
5481 		if (pages) {
5482 			upl->lite_list = bitmap_alloc(pages);
5483 		}
5484 	}
5485 
5486 	upl->flags = flags;
5487 	upl->ref_count = 1;
5488 	upl_lock_init(upl);
5489 #if CONFIG_IOSCHED
5490 	if (type & UPL_CREATE_IO_TRACKING) {
5491 		upl->upl_priority = proc_get_effective_thread_policy(current_thread(), TASK_POLICY_IO);
5492 	}
5493 
5494 	if ((type & UPL_CREATE_INTERNAL) && (type & UPL_CREATE_EXPEDITE_SUP)) {
5495 		/* Only support expedite on internal UPLs */
5496 		thread_t        curthread = current_thread();
5497 		upl->upl_reprio_info = kalloc_data(sizeof(uint64_t) * pages,
5498 		    Z_WAITOK | Z_ZERO);
5499 		upl->flags |= UPL_EXPEDITE_SUPPORTED;
5500 		if (curthread->decmp_upl != NULL) {
5501 			upl_set_decmp_info(upl, curthread->decmp_upl);
5502 		}
5503 	}
5504 #endif
5505 #if CONFIG_IOSCHED || UPL_DEBUG
5506 	if ((type & UPL_CREATE_IO_TRACKING) || upl_debug_enabled) {
5507 		upl->upl_creator = current_thread();
5508 		upl->flags |= UPL_TRACKED_BY_OBJECT;
5509 	}
5510 #endif
5511 
5512 #if UPL_DEBUG
5513 	upl->uple_create_btref = btref_get(__builtin_frame_address(0), 0);
5514 #endif /* UPL_DEBUG */
5515 
5516 	return upl;
5517 }
5518 
5519 static void
upl_destroy(upl_t upl)5520 upl_destroy(upl_t upl)
5521 {
5522 	uint32_t pages;
5523 
5524 //	DEBUG4K_UPL("upl %p (u_offset 0x%llx u_size 0x%llx) object %p\n", upl, (uint64_t)upl->u_offset, (uint64_t)upl->u_size, upl->map_object);
5525 
5526 	if (upl->ext_ref_count) {
5527 		panic("upl(%p) ext_ref_count", upl);
5528 	}
5529 
5530 #if CONFIG_IOSCHED
5531 	if ((upl->flags & UPL_DECMP_REAL_IO) && upl->decmp_io_upl) {
5532 		upl_t src_upl;
5533 		src_upl = upl->decmp_io_upl;
5534 		assert((src_upl->flags & UPL_DECMP_REQ) != 0);
5535 		upl_lock(src_upl);
5536 		src_upl->decmp_io_upl = NULL;
5537 		upl_unlock(src_upl);
5538 		upl_deallocate(src_upl);
5539 	}
5540 #endif /* CONFIG_IOSCHED */
5541 
5542 #if CONFIG_IOSCHED || UPL_DEBUG
5543 	if (((upl->flags & UPL_TRACKED_BY_OBJECT) || upl_debug_enabled) &&
5544 	    !(upl->flags & UPL_VECTOR)) {
5545 		vm_object_t     object;
5546 
5547 		if (upl->flags & UPL_SHADOWED) {
5548 			object = upl->map_object->shadow;
5549 		} else {
5550 			object = upl->map_object;
5551 		}
5552 
5553 		vm_object_lock(object);
5554 		queue_remove(&object->uplq, upl, upl_t, uplq);
5555 		vm_object_activity_end(object);
5556 		vm_object_collapse(object, 0, TRUE);
5557 		vm_object_unlock(object);
5558 	}
5559 #endif
5560 	/*
5561 	 * drop a reference on the map_object whether or
5562 	 * not a pageout object is inserted
5563 	 */
5564 	if (upl->flags & UPL_SHADOWED) {
5565 		vm_object_deallocate(upl->map_object);
5566 	}
5567 
5568 	if (upl->flags & UPL_DEVICE_MEMORY) {
5569 		pages = 1;
5570 	} else {
5571 		pages = (uint32_t)atop(upl_adjusted_size(upl, PAGE_MASK));
5572 	}
5573 
5574 	upl_lock_destroy(upl);
5575 
5576 #if CONFIG_IOSCHED
5577 	if (upl->flags & UPL_EXPEDITE_SUPPORTED) {
5578 		kfree_data(upl->upl_reprio_info, sizeof(uint64_t) * pages);
5579 	}
5580 #endif
5581 
5582 #if UPL_DEBUG
5583 	for (int i = 0; i < upl->upl_commit_index; i++) {
5584 		btref_put(upl->upl_commit_records[i].c_btref);
5585 	}
5586 	btref_put(upl->uple_create_btref);
5587 #endif /* UPL_DEBUG */
5588 
5589 	if ((upl->flags & UPL_LITE) && pages) {
5590 		bitmap_free(upl->lite_list, pages);
5591 	}
5592 	kfree_type(struct upl, struct upl_page_info,
5593 	    (upl->flags & UPL_INTERNAL) ? pages : 0, upl);
5594 }
5595 
5596 void
upl_deallocate(upl_t upl)5597 upl_deallocate(upl_t upl)
5598 {
5599 	upl_lock(upl);
5600 
5601 	if (--upl->ref_count == 0) {
5602 		if (vector_upl_is_valid(upl)) {
5603 			vector_upl_deallocate(upl);
5604 		}
5605 		upl_unlock(upl);
5606 
5607 		if (upl->upl_iodone) {
5608 			upl_callout_iodone(upl);
5609 		}
5610 
5611 		upl_destroy(upl);
5612 	} else {
5613 		upl_unlock(upl);
5614 	}
5615 }
5616 
5617 #if CONFIG_IOSCHED
5618 void
upl_mark_decmp(upl_t upl)5619 upl_mark_decmp(upl_t upl)
5620 {
5621 	if (upl->flags & UPL_TRACKED_BY_OBJECT) {
5622 		upl->flags |= UPL_DECMP_REQ;
5623 		upl->upl_creator->decmp_upl = (void *)upl;
5624 	}
5625 }
5626 
5627 void
upl_unmark_decmp(upl_t upl)5628 upl_unmark_decmp(upl_t upl)
5629 {
5630 	if (upl && (upl->flags & UPL_DECMP_REQ)) {
5631 		upl->upl_creator->decmp_upl = NULL;
5632 	}
5633 }
5634 
5635 #endif /* CONFIG_IOSCHED */
5636 
5637 #define VM_PAGE_Q_BACKING_UP(q)         \
5638 	((q)->pgo_laundry >= (((q)->pgo_maxlaundry * 8) / 10))
5639 
5640 boolean_t must_throttle_writes(void);
5641 
5642 boolean_t
must_throttle_writes()5643 must_throttle_writes()
5644 {
5645 	if (VM_PAGE_Q_BACKING_UP(&vm_pageout_queue_external) &&
5646 	    vm_page_pageable_external_count > (AVAILABLE_NON_COMPRESSED_MEMORY * 6) / 10) {
5647 		return TRUE;
5648 	}
5649 
5650 	return FALSE;
5651 }
5652 
5653 int vm_page_delayed_work_ctx_needed = 0;
5654 KALLOC_TYPE_DEFINE(dw_ctx_zone, struct vm_page_delayed_work_ctx, KT_PRIV_ACCT);
5655 
5656 __startup_func
5657 static void
vm_page_delayed_work_init_ctx(void)5658 vm_page_delayed_work_init_ctx(void)
5659 {
5660 	uint16_t min_delayed_work_ctx_allocated = 16;
5661 
5662 	/*
5663 	 * try really hard to always keep NCPU elements around in the zone
5664 	 * in order for the UPL code to almost always get an element.
5665 	 */
5666 	if (min_delayed_work_ctx_allocated < zpercpu_count()) {
5667 		min_delayed_work_ctx_allocated = (uint16_t)zpercpu_count();
5668 	}
5669 
5670 	zone_raise_reserve(dw_ctx_zone, min_delayed_work_ctx_allocated);
5671 }
5672 STARTUP(ZALLOC, STARTUP_RANK_LAST, vm_page_delayed_work_init_ctx);
5673 
5674 struct vm_page_delayed_work*
vm_page_delayed_work_get_ctx(void)5675 vm_page_delayed_work_get_ctx(void)
5676 {
5677 	struct vm_page_delayed_work_ctx * dw_ctx = NULL;
5678 
5679 	dw_ctx = zalloc_flags(dw_ctx_zone, Z_ZERO | Z_NOWAIT);
5680 
5681 	if (__probable(dw_ctx)) {
5682 		dw_ctx->delayed_owner = current_thread();
5683 	} else {
5684 		vm_page_delayed_work_ctx_needed++;
5685 	}
5686 	return dw_ctx ? dw_ctx->dwp : NULL;
5687 }
5688 
5689 void
vm_page_delayed_work_finish_ctx(struct vm_page_delayed_work * dwp)5690 vm_page_delayed_work_finish_ctx(struct vm_page_delayed_work* dwp)
5691 {
5692 	struct  vm_page_delayed_work_ctx *ldw_ctx;
5693 
5694 	ldw_ctx = (struct vm_page_delayed_work_ctx *)dwp;
5695 	ldw_ctx->delayed_owner = NULL;
5696 
5697 	zfree(dw_ctx_zone, ldw_ctx);
5698 }
5699 
5700 /*
5701  *	Routine:	vm_object_upl_request
5702  *	Purpose:
5703  *		Cause the population of a portion of a vm_object.
5704  *		Depending on the nature of the request, the pages
5705  *		returned may be contain valid data or be uninitialized.
5706  *		A page list structure, listing the physical pages
5707  *		will be returned upon request.
5708  *		This function is called by the file system or any other
5709  *		supplier of backing store to a pager.
5710  *		IMPORTANT NOTE: The caller must still respect the relationship
5711  *		between the vm_object and its backing memory object.  The
5712  *		caller MUST NOT substitute changes in the backing file
5713  *		without first doing a memory_object_lock_request on the
5714  *		target range unless it is know that the pages are not
5715  *		shared with another entity at the pager level.
5716  *		Copy_in_to:
5717  *			if a page list structure is present
5718  *			return the mapped physical pages, where a
5719  *			page is not present, return a non-initialized
5720  *			one.  If the no_sync bit is turned on, don't
5721  *			call the pager unlock to synchronize with other
5722  *			possible copies of the page. Leave pages busy
5723  *			in the original object, if a page list structure
5724  *			was specified.  When a commit of the page list
5725  *			pages is done, the dirty bit will be set for each one.
5726  *		Copy_out_from:
5727  *			If a page list structure is present, return
5728  *			all mapped pages.  Where a page does not exist
5729  *			map a zero filled one. Leave pages busy in
5730  *			the original object.  If a page list structure
5731  *			is not specified, this call is a no-op.
5732  *
5733  *		Note:  access of default pager objects has a rather interesting
5734  *		twist.  The caller of this routine, presumably the file system
5735  *		page cache handling code, will never actually make a request
5736  *		against a default pager backed object.  Only the default
5737  *		pager will make requests on backing store related vm_objects
5738  *		In this way the default pager can maintain the relationship
5739  *		between backing store files (abstract memory objects) and
5740  *		the vm_objects (cache objects), they support.
5741  *
5742  */
5743 
5744 __private_extern__ kern_return_t
vm_object_upl_request(vm_object_t object,vm_object_offset_t offset,upl_size_t size,upl_t * upl_ptr,upl_page_info_array_t user_page_list,unsigned int * page_list_count,upl_control_flags_t cntrl_flags,vm_tag_t tag)5745 vm_object_upl_request(
5746 	vm_object_t             object,
5747 	vm_object_offset_t      offset,
5748 	upl_size_t              size,
5749 	upl_t                   *upl_ptr,
5750 	upl_page_info_array_t   user_page_list,
5751 	unsigned int            *page_list_count,
5752 	upl_control_flags_t     cntrl_flags,
5753 	vm_tag_t                tag)
5754 {
5755 	vm_page_t               dst_page = VM_PAGE_NULL;
5756 	vm_object_offset_t      dst_offset;
5757 	upl_size_t              xfer_size;
5758 	unsigned int            size_in_pages;
5759 	boolean_t               dirty;
5760 	boolean_t               hw_dirty;
5761 	upl_t                   upl = NULL;
5762 	unsigned int            entry;
5763 	vm_page_t               alias_page = NULL;
5764 	int                     refmod_state = 0;
5765 	vm_object_t             last_copy_object;
5766 	struct  vm_page_delayed_work    dw_array;
5767 	struct  vm_page_delayed_work    *dwp, *dwp_start;
5768 	bool                    dwp_finish_ctx = TRUE;
5769 	int                     dw_count;
5770 	int                     dw_limit;
5771 	int                     io_tracking_flag = 0;
5772 	int                     grab_options;
5773 	int                     page_grab_count = 0;
5774 	ppnum_t                 phys_page;
5775 	pmap_flush_context      pmap_flush_context_storage;
5776 	boolean_t               pmap_flushes_delayed = FALSE;
5777 #if DEVELOPMENT || DEBUG
5778 	task_t                  task = current_task();
5779 #endif /* DEVELOPMENT || DEBUG */
5780 
5781 	dwp_start = dwp = NULL;
5782 
5783 	if (cntrl_flags & ~UPL_VALID_FLAGS) {
5784 		/*
5785 		 * For forward compatibility's sake,
5786 		 * reject any unknown flag.
5787 		 */
5788 		return KERN_INVALID_VALUE;
5789 	}
5790 	if ((!object->internal) && (object->paging_offset != 0)) {
5791 		panic("vm_object_upl_request: external object with non-zero paging offset");
5792 	}
5793 	if (object->phys_contiguous) {
5794 		panic("vm_object_upl_request: contiguous object specified");
5795 	}
5796 
5797 	assertf(page_aligned(offset) && page_aligned(size),
5798 	    "offset 0x%llx size 0x%x",
5799 	    offset, size);
5800 
5801 	VM_DEBUG_CONSTANT_EVENT(vm_object_upl_request, VM_UPL_REQUEST, DBG_FUNC_START, size, cntrl_flags, 0, 0);
5802 
5803 	dw_count = 0;
5804 	dw_limit = DELAYED_WORK_LIMIT(DEFAULT_DELAYED_WORK_LIMIT);
5805 	dwp_start = vm_page_delayed_work_get_ctx();
5806 	if (dwp_start == NULL) {
5807 		dwp_start = &dw_array;
5808 		dw_limit = 1;
5809 		dwp_finish_ctx = FALSE;
5810 	}
5811 
5812 	dwp = dwp_start;
5813 
5814 	if (size > MAX_UPL_SIZE_BYTES) {
5815 		size = MAX_UPL_SIZE_BYTES;
5816 	}
5817 
5818 	if ((cntrl_flags & UPL_SET_INTERNAL) && page_list_count != NULL) {
5819 		*page_list_count = MAX_UPL_SIZE_BYTES >> PAGE_SHIFT;
5820 	}
5821 
5822 #if CONFIG_IOSCHED || UPL_DEBUG
5823 	if (object->io_tracking || upl_debug_enabled) {
5824 		io_tracking_flag |= UPL_CREATE_IO_TRACKING;
5825 	}
5826 #endif
5827 #if CONFIG_IOSCHED
5828 	if (object->io_tracking) {
5829 		io_tracking_flag |= UPL_CREATE_EXPEDITE_SUP;
5830 	}
5831 #endif
5832 
5833 	if (cntrl_flags & UPL_SET_INTERNAL) {
5834 		if (cntrl_flags & UPL_SET_LITE) {
5835 			upl = upl_create(UPL_CREATE_INTERNAL | UPL_CREATE_LITE | io_tracking_flag, 0, size);
5836 		} else {
5837 			upl = upl_create(UPL_CREATE_INTERNAL | io_tracking_flag, 0, size);
5838 		}
5839 		user_page_list = size ? upl->page_list : NULL;
5840 	} else {
5841 		if (cntrl_flags & UPL_SET_LITE) {
5842 			upl = upl_create(UPL_CREATE_EXTERNAL | UPL_CREATE_LITE | io_tracking_flag, 0, size);
5843 		} else {
5844 			upl = upl_create(UPL_CREATE_EXTERNAL | io_tracking_flag, 0, size);
5845 		}
5846 	}
5847 	*upl_ptr = upl;
5848 
5849 	if (user_page_list) {
5850 		user_page_list[0].device = FALSE;
5851 	}
5852 
5853 	if (cntrl_flags & UPL_SET_LITE) {
5854 		upl->map_object = object;
5855 	} else {
5856 		upl->map_object = vm_object_allocate(size);
5857 		/*
5858 		 * No neeed to lock the new object: nobody else knows
5859 		 * about it yet, so it's all ours so far.
5860 		 */
5861 		upl->map_object->shadow = object;
5862 		upl->map_object->pageout = TRUE;
5863 		upl->map_object->can_persist = FALSE;
5864 		upl->map_object->copy_strategy = MEMORY_OBJECT_COPY_NONE;
5865 		upl->map_object->vo_shadow_offset = offset;
5866 		upl->map_object->wimg_bits = object->wimg_bits;
5867 		assertf(page_aligned(upl->map_object->vo_shadow_offset),
5868 		    "object %p shadow_offset 0x%llx",
5869 		    upl->map_object, upl->map_object->vo_shadow_offset);
5870 
5871 		alias_page = vm_page_grab_fictitious(TRUE);
5872 
5873 		upl->flags |= UPL_SHADOWED;
5874 	}
5875 	if (cntrl_flags & UPL_FOR_PAGEOUT) {
5876 		upl->flags |= UPL_PAGEOUT;
5877 	}
5878 
5879 	vm_object_lock(object);
5880 	vm_object_activity_begin(object);
5881 
5882 	grab_options = 0;
5883 #if CONFIG_SECLUDED_MEMORY
5884 	if (object->can_grab_secluded) {
5885 		grab_options |= VM_PAGE_GRAB_SECLUDED;
5886 	}
5887 #endif /* CONFIG_SECLUDED_MEMORY */
5888 
5889 	/*
5890 	 * we can lock in the paging_offset once paging_in_progress is set
5891 	 */
5892 	upl->u_size = size;
5893 	upl->u_offset = offset + object->paging_offset;
5894 
5895 #if CONFIG_IOSCHED || UPL_DEBUG
5896 	if (object->io_tracking || upl_debug_enabled) {
5897 		vm_object_activity_begin(object);
5898 		queue_enter(&object->uplq, upl, upl_t, uplq);
5899 	}
5900 #endif
5901 	if ((cntrl_flags & UPL_WILL_MODIFY) && object->copy != VM_OBJECT_NULL) {
5902 		/*
5903 		 * Honor copy-on-write obligations
5904 		 *
5905 		 * The caller is gathering these pages and
5906 		 * might modify their contents.  We need to
5907 		 * make sure that the copy object has its own
5908 		 * private copies of these pages before we let
5909 		 * the caller modify them.
5910 		 */
5911 		vm_object_update(object,
5912 		    offset,
5913 		    size,
5914 		    NULL,
5915 		    NULL,
5916 		    FALSE,              /* should_return */
5917 		    MEMORY_OBJECT_COPY_SYNC,
5918 		    VM_PROT_NO_CHANGE);
5919 
5920 		VM_PAGEOUT_DEBUG(upl_cow, 1);
5921 		VM_PAGEOUT_DEBUG(upl_cow_pages, (size >> PAGE_SHIFT));
5922 	}
5923 	/*
5924 	 * remember which copy object we synchronized with
5925 	 */
5926 	last_copy_object = object->copy;
5927 	entry = 0;
5928 
5929 	xfer_size = size;
5930 	dst_offset = offset;
5931 	size_in_pages = size / PAGE_SIZE;
5932 
5933 	if (vm_page_free_count > (vm_page_free_target + size_in_pages) ||
5934 	    object->resident_page_count < ((MAX_UPL_SIZE_BYTES * 2) >> PAGE_SHIFT)) {
5935 		object->scan_collisions = 0;
5936 	}
5937 
5938 	if ((cntrl_flags & UPL_WILL_MODIFY) && must_throttle_writes() == TRUE) {
5939 		boolean_t       isSSD = FALSE;
5940 
5941 #if !XNU_TARGET_OS_OSX
5942 		isSSD = TRUE;
5943 #else /* !XNU_TARGET_OS_OSX */
5944 		vnode_pager_get_isSSD(object->pager, &isSSD);
5945 #endif /* !XNU_TARGET_OS_OSX */
5946 		vm_object_unlock(object);
5947 
5948 		OSAddAtomic(size_in_pages, &vm_upl_wait_for_pages);
5949 
5950 		if (isSSD == TRUE) {
5951 			delay(1000 * size_in_pages);
5952 		} else {
5953 			delay(5000 * size_in_pages);
5954 		}
5955 		OSAddAtomic(-size_in_pages, &vm_upl_wait_for_pages);
5956 
5957 		vm_object_lock(object);
5958 	}
5959 
5960 	while (xfer_size) {
5961 		dwp->dw_mask = 0;
5962 
5963 		if ((alias_page == NULL) && !(cntrl_flags & UPL_SET_LITE)) {
5964 			vm_object_unlock(object);
5965 			alias_page = vm_page_grab_fictitious(TRUE);
5966 			vm_object_lock(object);
5967 		}
5968 		if (cntrl_flags & UPL_COPYOUT_FROM) {
5969 			upl->flags |= UPL_PAGE_SYNC_DONE;
5970 
5971 			if (((dst_page = vm_page_lookup(object, dst_offset)) == VM_PAGE_NULL) ||
5972 			    dst_page->vmp_fictitious ||
5973 			    dst_page->vmp_absent ||
5974 			    VMP_ERROR_GET(dst_page) ||
5975 			    dst_page->vmp_cleaning ||
5976 			    (VM_PAGE_WIRED(dst_page))) {
5977 				if (user_page_list) {
5978 					user_page_list[entry].phys_addr = 0;
5979 				}
5980 
5981 				goto try_next_page;
5982 			}
5983 			phys_page = VM_PAGE_GET_PHYS_PAGE(dst_page);
5984 
5985 			/*
5986 			 * grab this up front...
5987 			 * a high percentange of the time we're going to
5988 			 * need the hardware modification state a bit later
5989 			 * anyway... so we can eliminate an extra call into
5990 			 * the pmap layer by grabbing it here and recording it
5991 			 */
5992 			if (dst_page->vmp_pmapped) {
5993 				refmod_state = pmap_get_refmod(phys_page);
5994 			} else {
5995 				refmod_state = 0;
5996 			}
5997 
5998 			if ((refmod_state & VM_MEM_REFERENCED) && VM_PAGE_INACTIVE(dst_page)) {
5999 				/*
6000 				 * page is on inactive list and referenced...
6001 				 * reactivate it now... this gets it out of the
6002 				 * way of vm_pageout_scan which would have to
6003 				 * reactivate it upon tripping over it
6004 				 */
6005 				dwp->dw_mask |= DW_vm_page_activate;
6006 			}
6007 			if (cntrl_flags & UPL_RET_ONLY_DIRTY) {
6008 				/*
6009 				 * we're only asking for DIRTY pages to be returned
6010 				 */
6011 				if (dst_page->vmp_laundry || !(cntrl_flags & UPL_FOR_PAGEOUT)) {
6012 					/*
6013 					 * if we were the page stolen by vm_pageout_scan to be
6014 					 * cleaned (as opposed to a buddy being clustered in
6015 					 * or this request is not being driven by a PAGEOUT cluster
6016 					 * then we only need to check for the page being dirty or
6017 					 * precious to decide whether to return it
6018 					 */
6019 					if (dst_page->vmp_dirty || dst_page->vmp_precious || (refmod_state & VM_MEM_MODIFIED)) {
6020 						goto check_busy;
6021 					}
6022 					goto dont_return;
6023 				}
6024 				/*
6025 				 * this is a request for a PAGEOUT cluster and this page
6026 				 * is merely along for the ride as a 'buddy'... not only
6027 				 * does it have to be dirty to be returned, but it also
6028 				 * can't have been referenced recently...
6029 				 */
6030 				if ((hibernate_cleaning_in_progress == TRUE ||
6031 				    (!((refmod_state & VM_MEM_REFERENCED) || dst_page->vmp_reference) ||
6032 				    (dst_page->vmp_q_state == VM_PAGE_ON_THROTTLED_Q))) &&
6033 				    ((refmod_state & VM_MEM_MODIFIED) || dst_page->vmp_dirty || dst_page->vmp_precious)) {
6034 					goto check_busy;
6035 				}
6036 dont_return:
6037 				/*
6038 				 * if we reach here, we're not to return
6039 				 * the page... go on to the next one
6040 				 */
6041 				if (dst_page->vmp_laundry == TRUE) {
6042 					/*
6043 					 * if we get here, the page is not 'cleaning' (filtered out above).
6044 					 * since it has been referenced, remove it from the laundry
6045 					 * so we don't pay the cost of an I/O to clean a page
6046 					 * we're just going to take back
6047 					 */
6048 					vm_page_lockspin_queues();
6049 
6050 					vm_pageout_steal_laundry(dst_page, TRUE);
6051 					vm_page_activate(dst_page);
6052 
6053 					vm_page_unlock_queues();
6054 				}
6055 				if (user_page_list) {
6056 					user_page_list[entry].phys_addr = 0;
6057 				}
6058 
6059 				goto try_next_page;
6060 			}
6061 check_busy:
6062 			if (dst_page->vmp_busy) {
6063 				if (cntrl_flags & UPL_NOBLOCK) {
6064 					if (user_page_list) {
6065 						user_page_list[entry].phys_addr = 0;
6066 					}
6067 					dwp->dw_mask = 0;
6068 
6069 					goto try_next_page;
6070 				}
6071 				/*
6072 				 * someone else is playing with the
6073 				 * page.  We will have to wait.
6074 				 */
6075 				PAGE_SLEEP(object, dst_page, THREAD_UNINT);
6076 
6077 				continue;
6078 			}
6079 			if (dst_page->vmp_q_state == VM_PAGE_ON_PAGEOUT_Q) {
6080 				vm_page_lockspin_queues();
6081 
6082 				if (dst_page->vmp_q_state == VM_PAGE_ON_PAGEOUT_Q) {
6083 					/*
6084 					 * we've buddied up a page for a clustered pageout
6085 					 * that has already been moved to the pageout
6086 					 * queue by pageout_scan... we need to remove
6087 					 * it from the queue and drop the laundry count
6088 					 * on that queue
6089 					 */
6090 					vm_pageout_throttle_up(dst_page);
6091 				}
6092 				vm_page_unlock_queues();
6093 			}
6094 			hw_dirty = refmod_state & VM_MEM_MODIFIED;
6095 			dirty = hw_dirty ? TRUE : dst_page->vmp_dirty;
6096 
6097 			if (phys_page > upl->highest_page) {
6098 				upl->highest_page = phys_page;
6099 			}
6100 
6101 			assert(!pmap_is_noencrypt(phys_page));
6102 
6103 			if (cntrl_flags & UPL_SET_LITE) {
6104 				unsigned int    pg_num;
6105 
6106 				pg_num = (unsigned int) ((dst_offset - offset) / PAGE_SIZE);
6107 				assert(pg_num == (dst_offset - offset) / PAGE_SIZE);
6108 				bitmap_set(upl->lite_list, pg_num);
6109 
6110 				if (hw_dirty) {
6111 					if (pmap_flushes_delayed == FALSE) {
6112 						pmap_flush_context_init(&pmap_flush_context_storage);
6113 						pmap_flushes_delayed = TRUE;
6114 					}
6115 					pmap_clear_refmod_options(phys_page,
6116 					    VM_MEM_MODIFIED,
6117 					    PMAP_OPTIONS_NOFLUSH | PMAP_OPTIONS_CLEAR_WRITE,
6118 					    &pmap_flush_context_storage);
6119 				}
6120 
6121 				/*
6122 				 * Mark original page as cleaning
6123 				 * in place.
6124 				 */
6125 				dst_page->vmp_cleaning = TRUE;
6126 				dst_page->vmp_precious = FALSE;
6127 			} else {
6128 				/*
6129 				 * use pageclean setup, it is more
6130 				 * convenient even for the pageout
6131 				 * cases here
6132 				 */
6133 				vm_object_lock(upl->map_object);
6134 				vm_pageclean_setup(dst_page, alias_page, upl->map_object, size - xfer_size);
6135 				vm_object_unlock(upl->map_object);
6136 
6137 				alias_page->vmp_absent = FALSE;
6138 				alias_page = NULL;
6139 			}
6140 			if (dirty) {
6141 				SET_PAGE_DIRTY(dst_page, FALSE);
6142 			} else {
6143 				dst_page->vmp_dirty = FALSE;
6144 			}
6145 
6146 			if (!dirty) {
6147 				dst_page->vmp_precious = TRUE;
6148 			}
6149 
6150 			if (!(cntrl_flags & UPL_CLEAN_IN_PLACE)) {
6151 				if (!VM_PAGE_WIRED(dst_page)) {
6152 					dst_page->vmp_free_when_done = TRUE;
6153 				}
6154 			}
6155 		} else {
6156 			if ((cntrl_flags & UPL_WILL_MODIFY) && object->copy != last_copy_object) {
6157 				/*
6158 				 * Honor copy-on-write obligations
6159 				 *
6160 				 * The copy object has changed since we
6161 				 * last synchronized for copy-on-write.
6162 				 * Another copy object might have been
6163 				 * inserted while we released the object's
6164 				 * lock.  Since someone could have seen the
6165 				 * original contents of the remaining pages
6166 				 * through that new object, we have to
6167 				 * synchronize with it again for the remaining
6168 				 * pages only.  The previous pages are "busy"
6169 				 * so they can not be seen through the new
6170 				 * mapping.  The new mapping will see our
6171 				 * upcoming changes for those previous pages,
6172 				 * but that's OK since they couldn't see what
6173 				 * was there before.  It's just a race anyway
6174 				 * and there's no guarantee of consistency or
6175 				 * atomicity.  We just don't want new mappings
6176 				 * to see both the *before* and *after* pages.
6177 				 */
6178 				if (object->copy != VM_OBJECT_NULL) {
6179 					vm_object_update(
6180 						object,
6181 						dst_offset,/* current offset */
6182 						xfer_size, /* remaining size */
6183 						NULL,
6184 						NULL,
6185 						FALSE,     /* should_return */
6186 						MEMORY_OBJECT_COPY_SYNC,
6187 						VM_PROT_NO_CHANGE);
6188 
6189 					VM_PAGEOUT_DEBUG(upl_cow_again, 1);
6190 					VM_PAGEOUT_DEBUG(upl_cow_again_pages, (xfer_size >> PAGE_SHIFT));
6191 				}
6192 				/*
6193 				 * remember the copy object we synced with
6194 				 */
6195 				last_copy_object = object->copy;
6196 			}
6197 			dst_page = vm_page_lookup(object, dst_offset);
6198 
6199 			if (dst_page != VM_PAGE_NULL) {
6200 				if ((cntrl_flags & UPL_RET_ONLY_ABSENT)) {
6201 					/*
6202 					 * skip over pages already present in the cache
6203 					 */
6204 					if (user_page_list) {
6205 						user_page_list[entry].phys_addr = 0;
6206 					}
6207 
6208 					goto try_next_page;
6209 				}
6210 				if (dst_page->vmp_fictitious) {
6211 					panic("need corner case for fictitious page");
6212 				}
6213 
6214 				if (dst_page->vmp_busy || dst_page->vmp_cleaning) {
6215 					/*
6216 					 * someone else is playing with the
6217 					 * page.  We will have to wait.
6218 					 */
6219 					PAGE_SLEEP(object, dst_page, THREAD_UNINT);
6220 
6221 					continue;
6222 				}
6223 				if (dst_page->vmp_laundry) {
6224 					vm_pageout_steal_laundry(dst_page, FALSE);
6225 				}
6226 			} else {
6227 				if (object->private) {
6228 					/*
6229 					 * This is a nasty wrinkle for users
6230 					 * of upl who encounter device or
6231 					 * private memory however, it is
6232 					 * unavoidable, only a fault can
6233 					 * resolve the actual backing
6234 					 * physical page by asking the
6235 					 * backing device.
6236 					 */
6237 					if (user_page_list) {
6238 						user_page_list[entry].phys_addr = 0;
6239 					}
6240 
6241 					goto try_next_page;
6242 				}
6243 				if (object->scan_collisions) {
6244 					/*
6245 					 * the pageout_scan thread is trying to steal
6246 					 * pages from this object, but has run into our
6247 					 * lock... grab 2 pages from the head of the object...
6248 					 * the first is freed on behalf of pageout_scan, the
6249 					 * 2nd is for our own use... we use vm_object_page_grab
6250 					 * in both cases to avoid taking pages from the free
6251 					 * list since we are under memory pressure and our
6252 					 * lock on this object is getting in the way of
6253 					 * relieving it
6254 					 */
6255 					dst_page = vm_object_page_grab(object);
6256 
6257 					if (dst_page != VM_PAGE_NULL) {
6258 						vm_page_release(dst_page,
6259 						    FALSE);
6260 					}
6261 
6262 					dst_page = vm_object_page_grab(object);
6263 				}
6264 				if (dst_page == VM_PAGE_NULL) {
6265 					/*
6266 					 * need to allocate a page
6267 					 */
6268 					dst_page = vm_page_grab_options(grab_options);
6269 					if (dst_page != VM_PAGE_NULL) {
6270 						page_grab_count++;
6271 					}
6272 				}
6273 				if (dst_page == VM_PAGE_NULL) {
6274 					if ((cntrl_flags & (UPL_RET_ONLY_ABSENT | UPL_NOBLOCK)) == (UPL_RET_ONLY_ABSENT | UPL_NOBLOCK)) {
6275 						/*
6276 						 * we don't want to stall waiting for pages to come onto the free list
6277 						 * while we're already holding absent pages in this UPL
6278 						 * the caller will deal with the empty slots
6279 						 */
6280 						if (user_page_list) {
6281 							user_page_list[entry].phys_addr = 0;
6282 						}
6283 
6284 						goto try_next_page;
6285 					}
6286 					/*
6287 					 * no pages available... wait
6288 					 * then try again for the same
6289 					 * offset...
6290 					 */
6291 					vm_object_unlock(object);
6292 
6293 					OSAddAtomic(size_in_pages, &vm_upl_wait_for_pages);
6294 
6295 					VM_DEBUG_EVENT(vm_upl_page_wait, VM_UPL_PAGE_WAIT, DBG_FUNC_START, vm_upl_wait_for_pages, 0, 0, 0);
6296 
6297 					VM_PAGE_WAIT();
6298 					OSAddAtomic(-size_in_pages, &vm_upl_wait_for_pages);
6299 
6300 					VM_DEBUG_EVENT(vm_upl_page_wait, VM_UPL_PAGE_WAIT, DBG_FUNC_END, vm_upl_wait_for_pages, 0, 0, 0);
6301 
6302 					vm_object_lock(object);
6303 
6304 					continue;
6305 				}
6306 				vm_page_insert(dst_page, object, dst_offset);
6307 
6308 				dst_page->vmp_absent = TRUE;
6309 				dst_page->vmp_busy = FALSE;
6310 
6311 				if (cntrl_flags & UPL_RET_ONLY_ABSENT) {
6312 					/*
6313 					 * if UPL_RET_ONLY_ABSENT was specified,
6314 					 * than we're definitely setting up a
6315 					 * upl for a clustered read/pagein
6316 					 * operation... mark the pages as clustered
6317 					 * so upl_commit_range can put them on the
6318 					 * speculative list
6319 					 */
6320 					dst_page->vmp_clustered = TRUE;
6321 
6322 					if (!(cntrl_flags & UPL_FILE_IO)) {
6323 						counter_inc(&vm_statistics_pageins);
6324 					}
6325 				}
6326 			}
6327 			phys_page = VM_PAGE_GET_PHYS_PAGE(dst_page);
6328 
6329 			dst_page->vmp_overwriting = TRUE;
6330 
6331 			if (dst_page->vmp_pmapped) {
6332 				if (!(cntrl_flags & UPL_FILE_IO)) {
6333 					/*
6334 					 * eliminate all mappings from the
6335 					 * original object and its prodigy
6336 					 */
6337 					refmod_state = pmap_disconnect(phys_page);
6338 				} else {
6339 					refmod_state = pmap_get_refmod(phys_page);
6340 				}
6341 			} else {
6342 				refmod_state = 0;
6343 			}
6344 
6345 			hw_dirty = refmod_state & VM_MEM_MODIFIED;
6346 			dirty = hw_dirty ? TRUE : dst_page->vmp_dirty;
6347 
6348 			if (cntrl_flags & UPL_SET_LITE) {
6349 				unsigned int    pg_num;
6350 
6351 				pg_num = (unsigned int) ((dst_offset - offset) / PAGE_SIZE);
6352 				assert(pg_num == (dst_offset - offset) / PAGE_SIZE);
6353 				bitmap_set(upl->lite_list, pg_num);
6354 
6355 				if (hw_dirty) {
6356 					pmap_clear_modify(phys_page);
6357 				}
6358 
6359 				/*
6360 				 * Mark original page as cleaning
6361 				 * in place.
6362 				 */
6363 				dst_page->vmp_cleaning = TRUE;
6364 				dst_page->vmp_precious = FALSE;
6365 			} else {
6366 				/*
6367 				 * use pageclean setup, it is more
6368 				 * convenient even for the pageout
6369 				 * cases here
6370 				 */
6371 				vm_object_lock(upl->map_object);
6372 				vm_pageclean_setup(dst_page, alias_page, upl->map_object, size - xfer_size);
6373 				vm_object_unlock(upl->map_object);
6374 
6375 				alias_page->vmp_absent = FALSE;
6376 				alias_page = NULL;
6377 			}
6378 
6379 			if (cntrl_flags & UPL_REQUEST_SET_DIRTY) {
6380 				upl->flags &= ~UPL_CLEAR_DIRTY;
6381 				upl->flags |= UPL_SET_DIRTY;
6382 				dirty = TRUE;
6383 				/*
6384 				 * Page belonging to a code-signed object is about to
6385 				 * be written. Mark it tainted and disconnect it from
6386 				 * all pmaps so processes have to fault it back in and
6387 				 * deal with the tainted bit.
6388 				 */
6389 				if (object->code_signed && dst_page->vmp_cs_tainted != VMP_CS_ALL_TRUE) {
6390 					dst_page->vmp_cs_tainted = VMP_CS_ALL_TRUE;
6391 					vm_page_upl_tainted++;
6392 					if (dst_page->vmp_pmapped) {
6393 						refmod_state = pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(dst_page));
6394 						if (refmod_state & VM_MEM_REFERENCED) {
6395 							dst_page->vmp_reference = TRUE;
6396 						}
6397 					}
6398 				}
6399 			} else if (cntrl_flags & UPL_CLEAN_IN_PLACE) {
6400 				/*
6401 				 * clean in place for read implies
6402 				 * that a write will be done on all
6403 				 * the pages that are dirty before
6404 				 * a upl commit is done.  The caller
6405 				 * is obligated to preserve the
6406 				 * contents of all pages marked dirty
6407 				 */
6408 				upl->flags |= UPL_CLEAR_DIRTY;
6409 			}
6410 			dst_page->vmp_dirty = dirty;
6411 
6412 			if (!dirty) {
6413 				dst_page->vmp_precious = TRUE;
6414 			}
6415 
6416 			if (!VM_PAGE_WIRED(dst_page)) {
6417 				/*
6418 				 * deny access to the target page while
6419 				 * it is being worked on
6420 				 */
6421 				dst_page->vmp_busy = TRUE;
6422 			} else {
6423 				dwp->dw_mask |= DW_vm_page_wire;
6424 			}
6425 
6426 			/*
6427 			 * We might be about to satisfy a fault which has been
6428 			 * requested. So no need for the "restart" bit.
6429 			 */
6430 			dst_page->vmp_restart = FALSE;
6431 			if (!dst_page->vmp_absent && !(cntrl_flags & UPL_WILL_MODIFY)) {
6432 				/*
6433 				 * expect the page to be used
6434 				 */
6435 				dwp->dw_mask |= DW_set_reference;
6436 			}
6437 			if (cntrl_flags & UPL_PRECIOUS) {
6438 				if (object->internal) {
6439 					SET_PAGE_DIRTY(dst_page, FALSE);
6440 					dst_page->vmp_precious = FALSE;
6441 				} else {
6442 					dst_page->vmp_precious = TRUE;
6443 				}
6444 			} else {
6445 				dst_page->vmp_precious = FALSE;
6446 			}
6447 		}
6448 		if (dst_page->vmp_busy) {
6449 			upl->flags |= UPL_HAS_BUSY;
6450 		}
6451 
6452 		if (phys_page > upl->highest_page) {
6453 			upl->highest_page = phys_page;
6454 		}
6455 		assert(!pmap_is_noencrypt(phys_page));
6456 		if (user_page_list) {
6457 			user_page_list[entry].phys_addr = phys_page;
6458 			user_page_list[entry].free_when_done    = dst_page->vmp_free_when_done;
6459 			user_page_list[entry].absent    = dst_page->vmp_absent;
6460 			user_page_list[entry].dirty     = dst_page->vmp_dirty;
6461 			user_page_list[entry].precious  = dst_page->vmp_precious;
6462 			user_page_list[entry].device    = FALSE;
6463 			user_page_list[entry].needed    = FALSE;
6464 			if (dst_page->vmp_clustered == TRUE) {
6465 				user_page_list[entry].speculative = (dst_page->vmp_q_state == VM_PAGE_ON_SPECULATIVE_Q) ? TRUE : FALSE;
6466 			} else {
6467 				user_page_list[entry].speculative = FALSE;
6468 			}
6469 			user_page_list[entry].cs_validated = dst_page->vmp_cs_validated;
6470 			user_page_list[entry].cs_tainted = dst_page->vmp_cs_tainted;
6471 			user_page_list[entry].cs_nx = dst_page->vmp_cs_nx;
6472 			user_page_list[entry].mark      = FALSE;
6473 		}
6474 		/*
6475 		 * if UPL_RET_ONLY_ABSENT is set, then
6476 		 * we are working with a fresh page and we've
6477 		 * just set the clustered flag on it to
6478 		 * indicate that it was drug in as part of a
6479 		 * speculative cluster... so leave it alone
6480 		 */
6481 		if (!(cntrl_flags & UPL_RET_ONLY_ABSENT)) {
6482 			/*
6483 			 * someone is explicitly grabbing this page...
6484 			 * update clustered and speculative state
6485 			 *
6486 			 */
6487 			if (dst_page->vmp_clustered) {
6488 				VM_PAGE_CONSUME_CLUSTERED(dst_page);
6489 			}
6490 		}
6491 try_next_page:
6492 		if (dwp->dw_mask) {
6493 			if (dwp->dw_mask & DW_vm_page_activate) {
6494 				counter_inc(&vm_statistics_reactivations);
6495 			}
6496 
6497 			VM_PAGE_ADD_DELAYED_WORK(dwp, dst_page, dw_count);
6498 
6499 			if (dw_count >= dw_limit) {
6500 				vm_page_do_delayed_work(object, tag, dwp_start, dw_count);
6501 
6502 				dwp = dwp_start;
6503 				dw_count = 0;
6504 			}
6505 		}
6506 		entry++;
6507 		dst_offset += PAGE_SIZE_64;
6508 		xfer_size -= PAGE_SIZE;
6509 	}
6510 	if (dw_count) {
6511 		vm_page_do_delayed_work(object, tag, dwp_start, dw_count);
6512 		dwp = dwp_start;
6513 		dw_count = 0;
6514 	}
6515 
6516 	if (alias_page != NULL) {
6517 		VM_PAGE_FREE(alias_page);
6518 	}
6519 	if (pmap_flushes_delayed == TRUE) {
6520 		pmap_flush(&pmap_flush_context_storage);
6521 	}
6522 
6523 	if (page_list_count != NULL) {
6524 		if (upl->flags & UPL_INTERNAL) {
6525 			*page_list_count = 0;
6526 		} else if (*page_list_count > entry) {
6527 			*page_list_count = entry;
6528 		}
6529 	}
6530 #if UPL_DEBUG
6531 	upl->upl_state = 1;
6532 #endif
6533 	vm_object_unlock(object);
6534 
6535 	VM_DEBUG_CONSTANT_EVENT(vm_object_upl_request, VM_UPL_REQUEST, DBG_FUNC_END, page_grab_count, 0, 0, 0);
6536 #if DEVELOPMENT || DEBUG
6537 	if (task != NULL) {
6538 		ledger_credit(task->ledger, task_ledgers.pages_grabbed_upl, page_grab_count);
6539 	}
6540 #endif /* DEVELOPMENT || DEBUG */
6541 
6542 	if (dwp_start && dwp_finish_ctx) {
6543 		vm_page_delayed_work_finish_ctx(dwp_start);
6544 		dwp_start = dwp = NULL;
6545 	}
6546 
6547 	return KERN_SUCCESS;
6548 }
6549 
6550 /*
6551  *	Routine:	vm_object_super_upl_request
6552  *	Purpose:
6553  *		Cause the population of a portion of a vm_object
6554  *		in much the same way as memory_object_upl_request.
6555  *		Depending on the nature of the request, the pages
6556  *		returned may be contain valid data or be uninitialized.
6557  *		However, the region may be expanded up to the super
6558  *		cluster size provided.
6559  */
6560 
6561 __private_extern__ kern_return_t
vm_object_super_upl_request(vm_object_t object,vm_object_offset_t offset,upl_size_t size,upl_size_t super_cluster,upl_t * upl,upl_page_info_t * user_page_list,unsigned int * page_list_count,upl_control_flags_t cntrl_flags,vm_tag_t tag)6562 vm_object_super_upl_request(
6563 	vm_object_t object,
6564 	vm_object_offset_t      offset,
6565 	upl_size_t              size,
6566 	upl_size_t              super_cluster,
6567 	upl_t                   *upl,
6568 	upl_page_info_t         *user_page_list,
6569 	unsigned int            *page_list_count,
6570 	upl_control_flags_t     cntrl_flags,
6571 	vm_tag_t                tag)
6572 {
6573 	if (object->paging_offset > offset || ((cntrl_flags & UPL_VECTOR) == UPL_VECTOR)) {
6574 		return KERN_FAILURE;
6575 	}
6576 
6577 	assert(object->paging_in_progress);
6578 	offset = offset - object->paging_offset;
6579 
6580 	if (super_cluster > size) {
6581 		vm_object_offset_t      base_offset;
6582 		upl_size_t              super_size;
6583 		vm_object_size_t        super_size_64;
6584 
6585 		base_offset = (offset & ~((vm_object_offset_t) super_cluster - 1));
6586 		super_size = (offset + size) > (base_offset + super_cluster) ? super_cluster << 1 : super_cluster;
6587 		super_size_64 = ((base_offset + super_size) > object->vo_size) ? (object->vo_size - base_offset) : super_size;
6588 		super_size = (upl_size_t) super_size_64;
6589 		assert(super_size == super_size_64);
6590 
6591 		if (offset > (base_offset + super_size)) {
6592 			panic("vm_object_super_upl_request: Missed target pageout"
6593 			    " %#llx,%#llx, %#x, %#x, %#x, %#llx\n",
6594 			    offset, base_offset, super_size, super_cluster,
6595 			    size, object->paging_offset);
6596 		}
6597 		/*
6598 		 * apparently there is a case where the vm requests a
6599 		 * page to be written out who's offset is beyond the
6600 		 * object size
6601 		 */
6602 		if ((offset + size) > (base_offset + super_size)) {
6603 			super_size_64 = (offset + size) - base_offset;
6604 			super_size = (upl_size_t) super_size_64;
6605 			assert(super_size == super_size_64);
6606 		}
6607 
6608 		offset = base_offset;
6609 		size = super_size;
6610 	}
6611 	return vm_object_upl_request(object, offset, size, upl, user_page_list, page_list_count, cntrl_flags, tag);
6612 }
6613 
6614 int cs_executable_create_upl = 0;
6615 extern int proc_selfpid(void);
6616 extern char *proc_name_address(void *p);
6617 
6618 kern_return_t
vm_map_create_upl(vm_map_t map,vm_map_address_t offset,upl_size_t * upl_size,upl_t * upl,upl_page_info_array_t page_list,unsigned int * count,upl_control_flags_t * flags,vm_tag_t tag)6619 vm_map_create_upl(
6620 	vm_map_t                map,
6621 	vm_map_address_t        offset,
6622 	upl_size_t              *upl_size,
6623 	upl_t                   *upl,
6624 	upl_page_info_array_t   page_list,
6625 	unsigned int            *count,
6626 	upl_control_flags_t     *flags,
6627 	vm_tag_t                tag)
6628 {
6629 	vm_map_entry_t          entry;
6630 	upl_control_flags_t     caller_flags;
6631 	int                     force_data_sync;
6632 	int                     sync_cow_data;
6633 	vm_object_t             local_object;
6634 	vm_map_offset_t         local_offset;
6635 	vm_map_offset_t         local_start;
6636 	kern_return_t           ret;
6637 	vm_map_address_t        original_offset;
6638 	vm_map_size_t           original_size, adjusted_size;
6639 	vm_map_offset_t         local_entry_start;
6640 	vm_object_offset_t      local_entry_offset;
6641 	vm_object_offset_t      offset_in_mapped_page;
6642 	boolean_t               release_map = FALSE;
6643 
6644 start_with_map:
6645 
6646 	original_offset = offset;
6647 	original_size = *upl_size;
6648 	adjusted_size = original_size;
6649 
6650 	caller_flags = *flags;
6651 
6652 	if (caller_flags & ~UPL_VALID_FLAGS) {
6653 		/*
6654 		 * For forward compatibility's sake,
6655 		 * reject any unknown flag.
6656 		 */
6657 		ret = KERN_INVALID_VALUE;
6658 		goto done;
6659 	}
6660 	force_data_sync = (caller_flags & UPL_FORCE_DATA_SYNC);
6661 	sync_cow_data = !(caller_flags & UPL_COPYOUT_FROM);
6662 
6663 	if (upl == NULL) {
6664 		ret = KERN_INVALID_ARGUMENT;
6665 		goto done;
6666 	}
6667 
6668 REDISCOVER_ENTRY:
6669 	vm_map_lock_read(map);
6670 
6671 	if (!vm_map_lookup_entry(map, offset, &entry)) {
6672 		vm_map_unlock_read(map);
6673 		ret = KERN_FAILURE;
6674 		goto done;
6675 	}
6676 
6677 	local_entry_start = entry->vme_start;
6678 	local_entry_offset = VME_OFFSET(entry);
6679 
6680 	if (VM_MAP_PAGE_SHIFT(map) < PAGE_SHIFT) {
6681 		DEBUG4K_UPL("map %p (%d) offset 0x%llx size 0x%x flags 0x%llx\n", map, VM_MAP_PAGE_SHIFT(map), (uint64_t)offset, *upl_size, *flags);
6682 	}
6683 
6684 	if (entry->vme_end - original_offset < adjusted_size) {
6685 		adjusted_size = entry->vme_end - original_offset;
6686 		assert(adjusted_size > 0);
6687 		*upl_size = (upl_size_t) adjusted_size;
6688 		assert(*upl_size == adjusted_size);
6689 	}
6690 
6691 	if (caller_flags & UPL_QUERY_OBJECT_TYPE) {
6692 		*flags = 0;
6693 
6694 		if (!entry->is_sub_map &&
6695 		    VME_OBJECT(entry) != VM_OBJECT_NULL) {
6696 			if (VME_OBJECT(entry)->private) {
6697 				*flags = UPL_DEV_MEMORY;
6698 			}
6699 
6700 			if (VME_OBJECT(entry)->phys_contiguous) {
6701 				*flags |= UPL_PHYS_CONTIG;
6702 			}
6703 		}
6704 		vm_map_unlock_read(map);
6705 		ret = KERN_SUCCESS;
6706 		goto done;
6707 	}
6708 
6709 	offset_in_mapped_page = 0;
6710 	if (VM_MAP_PAGE_SIZE(map) < PAGE_SIZE) {
6711 		offset = vm_map_trunc_page(original_offset, VM_MAP_PAGE_MASK(map));
6712 		*upl_size = (upl_size_t)
6713 		    (vm_map_round_page(original_offset + adjusted_size,
6714 		    VM_MAP_PAGE_MASK(map))
6715 		    - offset);
6716 
6717 		offset_in_mapped_page = original_offset - offset;
6718 		assert(offset_in_mapped_page < VM_MAP_PAGE_SIZE(map));
6719 
6720 		DEBUG4K_UPL("map %p (%d) offset 0x%llx size 0x%llx flags 0x%llx -> offset 0x%llx adjusted_size 0x%llx *upl_size 0x%x offset_in_mapped_page 0x%llx\n", map, VM_MAP_PAGE_SHIFT(map), (uint64_t)original_offset, (uint64_t)original_size, *flags, (uint64_t)offset, (uint64_t)adjusted_size, *upl_size, offset_in_mapped_page);
6721 	}
6722 
6723 	if (!entry->is_sub_map) {
6724 		if (VME_OBJECT(entry) == VM_OBJECT_NULL ||
6725 		    !VME_OBJECT(entry)->phys_contiguous) {
6726 			if (*upl_size > MAX_UPL_SIZE_BYTES) {
6727 				*upl_size = MAX_UPL_SIZE_BYTES;
6728 			}
6729 		}
6730 
6731 		/*
6732 		 *      Create an object if necessary.
6733 		 */
6734 		if (VME_OBJECT(entry) == VM_OBJECT_NULL) {
6735 			if (vm_map_lock_read_to_write(map)) {
6736 				goto REDISCOVER_ENTRY;
6737 			}
6738 
6739 			VME_OBJECT_SET(entry,
6740 			    vm_object_allocate((vm_size_t)
6741 			    vm_object_round_page((entry->vme_end - entry->vme_start))),
6742 			    false, 0);
6743 			VME_OFFSET_SET(entry, 0);
6744 			assert(entry->use_pmap);
6745 
6746 			vm_map_lock_write_to_read(map);
6747 		}
6748 
6749 		if (!(caller_flags & UPL_COPYOUT_FROM) &&
6750 		    !(entry->protection & VM_PROT_WRITE)) {
6751 			vm_map_unlock_read(map);
6752 			ret = KERN_PROTECTION_FAILURE;
6753 			goto done;
6754 		}
6755 	}
6756 
6757 #if !XNU_TARGET_OS_OSX
6758 	if (map->pmap != kernel_pmap &&
6759 	    (caller_flags & UPL_COPYOUT_FROM) &&
6760 	    (entry->protection & VM_PROT_EXECUTE) &&
6761 	    !(entry->protection & VM_PROT_WRITE)) {
6762 		vm_offset_t     kaddr;
6763 		vm_size_t       ksize;
6764 
6765 		/*
6766 		 * We're about to create a read-only UPL backed by
6767 		 * memory from an executable mapping.
6768 		 * Wiring the pages would result in the pages being copied
6769 		 * (due to the "MAP_PRIVATE" mapping) and no longer
6770 		 * code-signed, so no longer eligible for execution.
6771 		 * Instead, let's copy the data into a kernel buffer and
6772 		 * create the UPL from this kernel buffer.
6773 		 * The kernel buffer is then freed, leaving the UPL holding
6774 		 * the last reference on the VM object, so the memory will
6775 		 * be released when the UPL is committed.
6776 		 */
6777 
6778 		vm_map_unlock_read(map);
6779 		entry = VM_MAP_ENTRY_NULL;
6780 		/* allocate kernel buffer */
6781 		ksize = round_page(*upl_size);
6782 		kaddr = 0;
6783 		ret = kmem_alloc(kernel_map, &kaddr, ksize,
6784 		    KMA_PAGEABLE | KMA_DATA, tag);
6785 		if (ret == KERN_SUCCESS) {
6786 			/* copyin the user data */
6787 			ret = copyinmap(map, offset, (void *)kaddr, *upl_size);
6788 		}
6789 		if (ret == KERN_SUCCESS) {
6790 			if (ksize > *upl_size) {
6791 				/* zero out the extra space in kernel buffer */
6792 				memset((void *)(kaddr + *upl_size),
6793 				    0,
6794 				    ksize - *upl_size);
6795 			}
6796 			/* create the UPL from the kernel buffer */
6797 			vm_object_offset_t      offset_in_object;
6798 			vm_object_offset_t      offset_in_object_page;
6799 
6800 			offset_in_object = offset - local_entry_start + local_entry_offset;
6801 			offset_in_object_page = offset_in_object - vm_object_trunc_page(offset_in_object);
6802 			assert(offset_in_object_page < PAGE_SIZE);
6803 			assert(offset_in_object_page + offset_in_mapped_page < PAGE_SIZE);
6804 			*upl_size -= offset_in_object_page + offset_in_mapped_page;
6805 			ret = vm_map_create_upl(kernel_map,
6806 			    (vm_map_address_t)(kaddr + offset_in_object_page + offset_in_mapped_page),
6807 			    upl_size, upl, page_list, count, flags, tag);
6808 		}
6809 		if (kaddr != 0) {
6810 			/* free the kernel buffer */
6811 			kmem_free(kernel_map, kaddr, ksize);
6812 			kaddr = 0;
6813 			ksize = 0;
6814 		}
6815 #if DEVELOPMENT || DEBUG
6816 		DTRACE_VM4(create_upl_from_executable,
6817 		    vm_map_t, map,
6818 		    vm_map_address_t, offset,
6819 		    upl_size_t, *upl_size,
6820 		    kern_return_t, ret);
6821 #endif /* DEVELOPMENT || DEBUG */
6822 		goto done;
6823 	}
6824 #endif /* !XNU_TARGET_OS_OSX */
6825 
6826 	if (!entry->is_sub_map) {
6827 		local_object = VME_OBJECT(entry);
6828 		assert(local_object != VM_OBJECT_NULL);
6829 	}
6830 
6831 	if (!entry->is_sub_map &&
6832 	    !entry->needs_copy &&
6833 	    *upl_size != 0 &&
6834 	    local_object->vo_size > *upl_size && /* partial UPL */
6835 	    entry->wired_count == 0 && /* No COW for entries that are wired */
6836 	    (map->pmap != kernel_pmap) && /* alias checks */
6837 	    (vm_map_entry_should_cow_for_true_share(entry) /* case 1 */
6838 	    ||
6839 	    ( /* case 2 */
6840 		    local_object->internal &&
6841 		    (local_object->copy_strategy == MEMORY_OBJECT_COPY_SYMMETRIC) &&
6842 		    local_object->ref_count > 1))) {
6843 		vm_prot_t       prot;
6844 
6845 		/*
6846 		 * Case 1:
6847 		 * Set up the targeted range for copy-on-write to avoid
6848 		 * applying true_share/copy_delay to the entire object.
6849 		 *
6850 		 * Case 2:
6851 		 * This map entry covers only part of an internal
6852 		 * object.  There could be other map entries covering
6853 		 * other areas of this object and some of these map
6854 		 * entries could be marked as "needs_copy", which
6855 		 * assumes that the object is COPY_SYMMETRIC.
6856 		 * To avoid marking this object as COPY_DELAY and
6857 		 * "true_share", let's shadow it and mark the new
6858 		 * (smaller) object as "true_share" and COPY_DELAY.
6859 		 */
6860 
6861 		if (vm_map_lock_read_to_write(map)) {
6862 			goto REDISCOVER_ENTRY;
6863 		}
6864 		vm_map_lock_assert_exclusive(map);
6865 		assert(VME_OBJECT(entry) == local_object);
6866 
6867 		vm_map_clip_start(map,
6868 		    entry,
6869 		    vm_map_trunc_page(offset,
6870 		    VM_MAP_PAGE_MASK(map)));
6871 		vm_map_clip_end(map,
6872 		    entry,
6873 		    vm_map_round_page(offset + *upl_size,
6874 		    VM_MAP_PAGE_MASK(map)));
6875 		if ((entry->vme_end - offset) < *upl_size) {
6876 			*upl_size = (upl_size_t) (entry->vme_end - offset);
6877 			assert(*upl_size == entry->vme_end - offset);
6878 		}
6879 
6880 		prot = entry->protection & ~VM_PROT_WRITE;
6881 		if (override_nx(map, VME_ALIAS(entry)) && prot) {
6882 			prot |= VM_PROT_EXECUTE;
6883 		}
6884 		vm_object_pmap_protect(local_object,
6885 		    VME_OFFSET(entry),
6886 		    entry->vme_end - entry->vme_start,
6887 		    ((entry->is_shared ||
6888 		    map->mapped_in_other_pmaps)
6889 		    ? PMAP_NULL
6890 		    : map->pmap),
6891 		    VM_MAP_PAGE_SIZE(map),
6892 		    entry->vme_start,
6893 		    prot);
6894 
6895 		assert(entry->wired_count == 0);
6896 
6897 		/*
6898 		 * Lock the VM object and re-check its status: if it's mapped
6899 		 * in another address space, we could still be racing with
6900 		 * another thread holding that other VM map exclusively.
6901 		 */
6902 		vm_object_lock(local_object);
6903 		if (local_object->true_share) {
6904 			/* object is already in proper state: no COW needed */
6905 			assert(local_object->copy_strategy !=
6906 			    MEMORY_OBJECT_COPY_SYMMETRIC);
6907 		} else {
6908 			/* not true_share: ask for copy-on-write below */
6909 			assert(local_object->copy_strategy ==
6910 			    MEMORY_OBJECT_COPY_SYMMETRIC);
6911 			entry->needs_copy = TRUE;
6912 		}
6913 		vm_object_unlock(local_object);
6914 
6915 		vm_map_lock_write_to_read(map);
6916 	}
6917 
6918 	if (entry->needs_copy) {
6919 		/*
6920 		 * Honor copy-on-write for COPY_SYMMETRIC
6921 		 * strategy.
6922 		 */
6923 		vm_map_t                local_map;
6924 		vm_object_t             object;
6925 		vm_object_offset_t      new_offset;
6926 		vm_prot_t               prot;
6927 		boolean_t               wired;
6928 		vm_map_version_t        version;
6929 		vm_map_t                real_map;
6930 		vm_prot_t               fault_type;
6931 
6932 		local_map = map;
6933 
6934 		if (caller_flags & UPL_COPYOUT_FROM) {
6935 			fault_type = VM_PROT_READ | VM_PROT_COPY;
6936 			vm_counters.create_upl_extra_cow++;
6937 			vm_counters.create_upl_extra_cow_pages +=
6938 			    (entry->vme_end - entry->vme_start) / PAGE_SIZE;
6939 		} else {
6940 			fault_type = VM_PROT_WRITE;
6941 		}
6942 		if (vm_map_lookup_and_lock_object(&local_map,
6943 		    offset, fault_type,
6944 		    OBJECT_LOCK_EXCLUSIVE,
6945 		    &version, &object,
6946 		    &new_offset, &prot, &wired,
6947 		    NULL,
6948 		    &real_map, NULL) != KERN_SUCCESS) {
6949 			if (fault_type == VM_PROT_WRITE) {
6950 				vm_counters.create_upl_lookup_failure_write++;
6951 			} else {
6952 				vm_counters.create_upl_lookup_failure_copy++;
6953 			}
6954 			vm_map_unlock_read(local_map);
6955 			ret = KERN_FAILURE;
6956 			goto done;
6957 		}
6958 		if (real_map != local_map) {
6959 			vm_map_unlock(real_map);
6960 		}
6961 		vm_map_unlock_read(local_map);
6962 
6963 		vm_object_unlock(object);
6964 
6965 		goto REDISCOVER_ENTRY;
6966 	}
6967 
6968 	if (entry->is_sub_map) {
6969 		vm_map_t        submap;
6970 
6971 		submap = VME_SUBMAP(entry);
6972 		local_start = entry->vme_start;
6973 		local_offset = (vm_map_offset_t)VME_OFFSET(entry);
6974 
6975 		vm_map_reference(submap);
6976 		vm_map_unlock_read(map);
6977 
6978 		DEBUG4K_UPL("map %p offset 0x%llx (0x%llx) size 0x%x (adjusted 0x%llx original 0x%llx) offset_in_mapped_page 0x%llx submap %p\n", map, (uint64_t)offset, (uint64_t)original_offset, *upl_size, (uint64_t)adjusted_size, (uint64_t)original_size, offset_in_mapped_page, submap);
6979 		offset += offset_in_mapped_page;
6980 		*upl_size -= offset_in_mapped_page;
6981 
6982 		if (release_map) {
6983 			vm_map_deallocate(map);
6984 		}
6985 		map = submap;
6986 		release_map = TRUE;
6987 		offset = local_offset + (offset - local_start);
6988 		goto start_with_map;
6989 	}
6990 
6991 	if (sync_cow_data &&
6992 	    (VME_OBJECT(entry)->shadow ||
6993 	    VME_OBJECT(entry)->copy)) {
6994 		local_object = VME_OBJECT(entry);
6995 		local_start = entry->vme_start;
6996 		local_offset = (vm_map_offset_t)VME_OFFSET(entry);
6997 
6998 		vm_object_reference(local_object);
6999 		vm_map_unlock_read(map);
7000 
7001 		if (local_object->shadow && local_object->copy) {
7002 			vm_object_lock_request(local_object->shadow,
7003 			    ((vm_object_offset_t)
7004 			    ((offset - local_start) +
7005 			    local_offset) +
7006 			    local_object->vo_shadow_offset),
7007 			    *upl_size, FALSE,
7008 			    MEMORY_OBJECT_DATA_SYNC,
7009 			    VM_PROT_NO_CHANGE);
7010 		}
7011 		sync_cow_data = FALSE;
7012 		vm_object_deallocate(local_object);
7013 
7014 		goto REDISCOVER_ENTRY;
7015 	}
7016 	if (force_data_sync) {
7017 		local_object = VME_OBJECT(entry);
7018 		local_start = entry->vme_start;
7019 		local_offset = (vm_map_offset_t)VME_OFFSET(entry);
7020 
7021 		vm_object_reference(local_object);
7022 		vm_map_unlock_read(map);
7023 
7024 		vm_object_lock_request(local_object,
7025 		    ((vm_object_offset_t)
7026 		    ((offset - local_start) +
7027 		    local_offset)),
7028 		    (vm_object_size_t)*upl_size,
7029 		    FALSE,
7030 		    MEMORY_OBJECT_DATA_SYNC,
7031 		    VM_PROT_NO_CHANGE);
7032 
7033 		force_data_sync = FALSE;
7034 		vm_object_deallocate(local_object);
7035 
7036 		goto REDISCOVER_ENTRY;
7037 	}
7038 	if (VME_OBJECT(entry)->private) {
7039 		*flags = UPL_DEV_MEMORY;
7040 	} else {
7041 		*flags = 0;
7042 	}
7043 
7044 	if (VME_OBJECT(entry)->phys_contiguous) {
7045 		*flags |= UPL_PHYS_CONTIG;
7046 	}
7047 
7048 	local_object = VME_OBJECT(entry);
7049 	local_offset = (vm_map_offset_t)VME_OFFSET(entry);
7050 	local_start = entry->vme_start;
7051 
7052 	/*
7053 	 * Wiring will copy the pages to the shadow object.
7054 	 * The shadow object will not be code-signed so
7055 	 * attempting to execute code from these copied pages
7056 	 * would trigger a code-signing violation.
7057 	 */
7058 	if (entry->protection & VM_PROT_EXECUTE) {
7059 #if MACH_ASSERT
7060 		printf("pid %d[%s] create_upl out of executable range from "
7061 		    "0x%llx to 0x%llx: side effects may include "
7062 		    "code-signing violations later on\n",
7063 		    proc_selfpid(),
7064 		    (get_bsdtask_info(current_task())
7065 		    ? proc_name_address(get_bsdtask_info(current_task()))
7066 		    : "?"),
7067 		    (uint64_t) entry->vme_start,
7068 		    (uint64_t) entry->vme_end);
7069 #endif /* MACH_ASSERT */
7070 		DTRACE_VM2(cs_executable_create_upl,
7071 		    uint64_t, (uint64_t)entry->vme_start,
7072 		    uint64_t, (uint64_t)entry->vme_end);
7073 		cs_executable_create_upl++;
7074 	}
7075 
7076 	vm_object_lock(local_object);
7077 
7078 	/*
7079 	 * Ensure that this object is "true_share" and "copy_delay" now,
7080 	 * while we're still holding the VM map lock.  After we unlock the map,
7081 	 * anything could happen to that mapping, including some copy-on-write
7082 	 * activity.  We need to make sure that the IOPL will point at the
7083 	 * same memory as the mapping.
7084 	 */
7085 	if (local_object->true_share) {
7086 		assert(local_object->copy_strategy !=
7087 		    MEMORY_OBJECT_COPY_SYMMETRIC);
7088 	} else if (local_object != kernel_object &&
7089 	    local_object != compressor_object &&
7090 	    !local_object->phys_contiguous) {
7091 #if VM_OBJECT_TRACKING_OP_TRUESHARE
7092 		if (!local_object->true_share &&
7093 		    vm_object_tracking_btlog) {
7094 			btlog_record(vm_object_tracking_btlog, local_object,
7095 			    VM_OBJECT_TRACKING_OP_TRUESHARE,
7096 			    btref_get(__builtin_frame_address(0), 0));
7097 		}
7098 #endif /* VM_OBJECT_TRACKING_OP_TRUESHARE */
7099 		local_object->true_share = TRUE;
7100 		if (local_object->copy_strategy ==
7101 		    MEMORY_OBJECT_COPY_SYMMETRIC) {
7102 			local_object->copy_strategy = MEMORY_OBJECT_COPY_DELAY;
7103 		}
7104 	}
7105 
7106 	vm_object_reference_locked(local_object);
7107 	vm_object_unlock(local_object);
7108 
7109 	vm_map_unlock_read(map);
7110 
7111 	offset += offset_in_mapped_page;
7112 	assert(*upl_size > offset_in_mapped_page);
7113 	*upl_size -= offset_in_mapped_page;
7114 
7115 	ret = vm_object_iopl_request(local_object,
7116 	    ((vm_object_offset_t)
7117 	    ((offset - local_start) + local_offset)),
7118 	    *upl_size,
7119 	    upl,
7120 	    page_list,
7121 	    count,
7122 	    caller_flags,
7123 	    tag);
7124 	vm_object_deallocate(local_object);
7125 
7126 done:
7127 	if (release_map) {
7128 		vm_map_deallocate(map);
7129 	}
7130 
7131 	return ret;
7132 }
7133 
7134 /*
7135  * Internal routine to enter a UPL into a VM map.
7136  *
7137  * JMM - This should just be doable through the standard
7138  * vm_map_enter() API.
7139  */
7140 kern_return_t
vm_map_enter_upl_range(vm_map_t map,upl_t upl,vm_object_offset_t offset_to_map,upl_size_t size_to_map,vm_prot_t prot_to_map,vm_map_offset_t * dst_addr)7141 vm_map_enter_upl_range(
7142 	vm_map_t                map,
7143 	upl_t                   upl,
7144 	vm_object_offset_t      offset_to_map,
7145 	upl_size_t              size_to_map,
7146 	vm_prot_t               prot_to_map,
7147 	vm_map_offset_t         *dst_addr)
7148 {
7149 	vm_map_size_t           size;
7150 	vm_object_offset_t      offset;
7151 	vm_map_offset_t         addr;
7152 	vm_page_t               m;
7153 	kern_return_t           kr;
7154 	int                     isVectorUPL = 0, curr_upl = 0;
7155 	upl_t                   vector_upl = NULL;
7156 	mach_vm_offset_t        vector_upl_dst_addr = 0;
7157 	vm_map_t                vector_upl_submap = NULL;
7158 	upl_offset_t            subupl_offset = 0;
7159 	upl_size_t              subupl_size = 0;
7160 
7161 	if (upl == UPL_NULL) {
7162 		return KERN_INVALID_ARGUMENT;
7163 	}
7164 
7165 	DEBUG4K_UPL("map %p upl %p flags 0x%x object %p offset 0x%llx (uploff: 0x%llx) size 0x%x (uplsz: 0x%x) \n", map, upl, upl->flags, upl->map_object, offset_to_map, upl->u_offset, size_to_map, upl->u_size);
7166 	assert(map == kernel_map);
7167 
7168 	if ((isVectorUPL = vector_upl_is_valid(upl))) {
7169 		int mapped = 0, valid_upls = 0;
7170 		vector_upl = upl;
7171 
7172 		upl_lock(vector_upl);
7173 		for (curr_upl = 0; curr_upl < vector_upl_max_upls(vector_upl); curr_upl++) {
7174 			upl =  vector_upl_subupl_byindex(vector_upl, curr_upl );
7175 			if (upl == NULL) {
7176 				continue;
7177 			}
7178 			valid_upls++;
7179 			if (UPL_PAGE_LIST_MAPPED & upl->flags) {
7180 				mapped++;
7181 			}
7182 		}
7183 
7184 		if (mapped) {
7185 			if (mapped != valid_upls) {
7186 				panic("Only %d of the %d sub-upls within the Vector UPL are alread mapped", mapped, valid_upls);
7187 			} else {
7188 				upl_unlock(vector_upl);
7189 				return KERN_FAILURE;
7190 			}
7191 		}
7192 
7193 		if (VM_MAP_PAGE_MASK(map) < PAGE_MASK) {
7194 			panic("TODO4K: vector UPL not implemented");
7195 		}
7196 
7197 		vector_upl_submap = kmem_suballoc(map, &vector_upl_dst_addr,
7198 		    vector_upl->u_size, VM_MAP_CREATE_DEFAULT,
7199 		    VM_FLAGS_ANYWHERE, KMS_NOFAIL | KMS_DATA,
7200 		    VM_KERN_MEMORY_NONE).kmr_submap;
7201 		map = vector_upl_submap;
7202 		vector_upl_set_submap(vector_upl, vector_upl_submap, vector_upl_dst_addr);
7203 		curr_upl = 0;
7204 	} else {
7205 		upl_lock(upl);
7206 	}
7207 
7208 process_upl_to_enter:
7209 	if (isVectorUPL) {
7210 		if (curr_upl == vector_upl_max_upls(vector_upl)) {
7211 			*dst_addr = vector_upl_dst_addr;
7212 			upl_unlock(vector_upl);
7213 			return KERN_SUCCESS;
7214 		}
7215 		upl =  vector_upl_subupl_byindex(vector_upl, curr_upl++ );
7216 		if (upl == NULL) {
7217 			goto process_upl_to_enter;
7218 		}
7219 
7220 		vector_upl_get_iostate(vector_upl, upl, &subupl_offset, &subupl_size);
7221 		*dst_addr = (vm_map_offset_t)(vector_upl_dst_addr + (vm_map_offset_t)subupl_offset);
7222 	} else {
7223 		/*
7224 		 * check to see if already mapped
7225 		 */
7226 		if (UPL_PAGE_LIST_MAPPED & upl->flags) {
7227 			upl_unlock(upl);
7228 			return KERN_FAILURE;
7229 		}
7230 	}
7231 
7232 	if ((!(upl->flags & UPL_SHADOWED)) &&
7233 	    ((upl->flags & UPL_HAS_BUSY) ||
7234 	    !((upl->flags & (UPL_DEVICE_MEMORY | UPL_IO_WIRE)) || (upl->map_object->phys_contiguous)))) {
7235 		vm_object_t             object;
7236 		vm_page_t               alias_page;
7237 		vm_object_offset_t      new_offset;
7238 		unsigned int            pg_num;
7239 
7240 		size = upl_adjusted_size(upl, VM_MAP_PAGE_MASK(map));
7241 		object = upl->map_object;
7242 		upl->map_object = vm_object_allocate(vm_object_round_page(size));
7243 
7244 		vm_object_lock(upl->map_object);
7245 
7246 		upl->map_object->shadow = object;
7247 		upl->map_object->pageout = TRUE;
7248 		upl->map_object->can_persist = FALSE;
7249 		upl->map_object->copy_strategy = MEMORY_OBJECT_COPY_NONE;
7250 		upl->map_object->vo_shadow_offset = upl_adjusted_offset(upl, PAGE_MASK) - object->paging_offset;
7251 		assertf(page_aligned(upl->map_object->vo_shadow_offset),
7252 		    "object %p shadow_offset 0x%llx",
7253 		    upl->map_object,
7254 		    (uint64_t)upl->map_object->vo_shadow_offset);
7255 		upl->map_object->wimg_bits = object->wimg_bits;
7256 		offset = upl->map_object->vo_shadow_offset;
7257 		new_offset = 0;
7258 
7259 		upl->flags |= UPL_SHADOWED;
7260 
7261 		while (size) {
7262 			pg_num = (unsigned int) (new_offset / PAGE_SIZE);
7263 			assert(pg_num == new_offset / PAGE_SIZE);
7264 
7265 			if (bitmap_test(upl->lite_list, pg_num)) {
7266 				alias_page = vm_page_grab_fictitious(TRUE);
7267 
7268 				vm_object_lock(object);
7269 
7270 				m = vm_page_lookup(object, offset);
7271 				if (m == VM_PAGE_NULL) {
7272 					panic("vm_upl_map: page missing");
7273 				}
7274 
7275 				/*
7276 				 * Convert the fictitious page to a private
7277 				 * shadow of the real page.
7278 				 */
7279 				assert(alias_page->vmp_fictitious);
7280 				alias_page->vmp_fictitious = FALSE;
7281 				alias_page->vmp_private = TRUE;
7282 				alias_page->vmp_free_when_done = TRUE;
7283 				/*
7284 				 * since m is a page in the upl it must
7285 				 * already be wired or BUSY, so it's
7286 				 * safe to assign the underlying physical
7287 				 * page to the alias
7288 				 */
7289 				VM_PAGE_SET_PHYS_PAGE(alias_page, VM_PAGE_GET_PHYS_PAGE(m));
7290 
7291 				vm_object_unlock(object);
7292 
7293 				vm_page_lockspin_queues();
7294 				vm_page_wire(alias_page, VM_KERN_MEMORY_NONE, TRUE);
7295 				vm_page_unlock_queues();
7296 
7297 				vm_page_insert_wired(alias_page, upl->map_object, new_offset, VM_KERN_MEMORY_NONE);
7298 
7299 				assert(!alias_page->vmp_wanted);
7300 				alias_page->vmp_busy = FALSE;
7301 				alias_page->vmp_absent = FALSE;
7302 			}
7303 			size -= PAGE_SIZE;
7304 			offset += PAGE_SIZE_64;
7305 			new_offset += PAGE_SIZE_64;
7306 		}
7307 		vm_object_unlock(upl->map_object);
7308 	}
7309 	if (upl->flags & UPL_SHADOWED) {
7310 		if (isVectorUPL) {
7311 			offset = 0;
7312 		} else {
7313 			offset = offset_to_map;
7314 		}
7315 	} else {
7316 		offset = upl_adjusted_offset(upl, VM_MAP_PAGE_MASK(map)) - upl->map_object->paging_offset;
7317 		if (!isVectorUPL) {
7318 			offset += offset_to_map;
7319 		}
7320 	}
7321 
7322 	if (isVectorUPL) {
7323 		size = upl_adjusted_size(upl, VM_MAP_PAGE_MASK(map));
7324 	} else {
7325 		size = MIN(upl_adjusted_size(upl, VM_MAP_PAGE_MASK(map)), size_to_map);
7326 	}
7327 
7328 	vm_object_reference(upl->map_object);
7329 
7330 	if (!isVectorUPL) {
7331 		*dst_addr = 0;
7332 		/*
7333 		 * NEED A UPL_MAP ALIAS
7334 		 */
7335 		kr = vm_map_enter(map, dst_addr, (vm_map_size_t)size, (vm_map_offset_t) 0,
7336 		    VM_MAP_KERNEL_FLAGS_DATA_ANYWHERE(.vm_tag = VM_KERN_MEMORY_OSFMK),
7337 		    upl->map_object, offset, FALSE,
7338 		    prot_to_map, VM_PROT_ALL, VM_INHERIT_DEFAULT);
7339 
7340 		if (kr != KERN_SUCCESS) {
7341 			vm_object_deallocate(upl->map_object);
7342 			upl_unlock(upl);
7343 			return kr;
7344 		}
7345 	} else {
7346 		kr = vm_map_enter(map, dst_addr, (vm_map_size_t)size, (vm_map_offset_t) 0,
7347 		    VM_MAP_KERNEL_FLAGS_FIXED(.vm_tag = VM_KERN_MEMORY_OSFMK),
7348 		    upl->map_object, offset, FALSE,
7349 		    prot_to_map, VM_PROT_ALL, VM_INHERIT_DEFAULT);
7350 		if (kr) {
7351 			panic("vm_map_enter failed for a Vector UPL");
7352 		}
7353 	}
7354 	upl->u_mapped_size = (upl_size_t) size; /* When we allow multiple submappings of the UPL */
7355 	                                        /* this will have to be an increment rather than */
7356 	                                        /* an assignment. */
7357 	vm_object_lock(upl->map_object);
7358 
7359 	for (addr = *dst_addr; size > 0; size -= PAGE_SIZE, addr += PAGE_SIZE) {
7360 		m = vm_page_lookup(upl->map_object, offset);
7361 
7362 		if (m) {
7363 			m->vmp_pmapped = TRUE;
7364 
7365 			/* CODE SIGNING ENFORCEMENT: page has been wpmapped,
7366 			 * but only in kernel space. If this was on a user map,
7367 			 * we'd have to set the wpmapped bit. */
7368 			/* m->vmp_wpmapped = TRUE; */
7369 			assert(map->pmap == kernel_pmap);
7370 
7371 			PMAP_ENTER(map->pmap, addr, m, prot_to_map, VM_PROT_NONE, 0, TRUE, kr);
7372 
7373 			assert(kr == KERN_SUCCESS);
7374 #if KASAN
7375 			kasan_notify_address(addr, PAGE_SIZE_64);
7376 #endif
7377 		}
7378 		offset += PAGE_SIZE_64;
7379 	}
7380 	vm_object_unlock(upl->map_object);
7381 
7382 	/*
7383 	 * hold a reference for the mapping
7384 	 */
7385 	upl->ref_count++;
7386 	upl->flags |= UPL_PAGE_LIST_MAPPED;
7387 	upl->kaddr = (vm_offset_t) *dst_addr;
7388 	assert(upl->kaddr == *dst_addr);
7389 
7390 	if (isVectorUPL) {
7391 		goto process_upl_to_enter;
7392 	}
7393 
7394 	if (!isVectorUPL) {
7395 		vm_map_offset_t addr_adjustment;
7396 
7397 		addr_adjustment = (vm_map_offset_t)(upl->u_offset - upl_adjusted_offset(upl, VM_MAP_PAGE_MASK(map)));
7398 		if (addr_adjustment) {
7399 			assert(VM_MAP_PAGE_MASK(map) != PAGE_MASK);
7400 			DEBUG4K_UPL("dst_addr 0x%llx (+ 0x%llx) -> 0x%llx\n", (uint64_t)*dst_addr, (uint64_t)addr_adjustment, (uint64_t)(*dst_addr + addr_adjustment));
7401 			*dst_addr += addr_adjustment;
7402 		}
7403 	}
7404 
7405 	upl_unlock(upl);
7406 
7407 	return KERN_SUCCESS;
7408 }
7409 
7410 kern_return_t
vm_map_enter_upl(vm_map_t map,upl_t upl,vm_map_offset_t * dst_addr)7411 vm_map_enter_upl(
7412 	vm_map_t                map,
7413 	upl_t                   upl,
7414 	vm_map_offset_t         *dst_addr)
7415 {
7416 	upl_size_t upl_size = upl_adjusted_size(upl, VM_MAP_PAGE_MASK(map));
7417 	return vm_map_enter_upl_range(map, upl, 0, upl_size, VM_PROT_DEFAULT, dst_addr);
7418 }
7419 
7420 /*
7421  * Internal routine to remove a UPL mapping from a VM map.
7422  *
7423  * XXX - This should just be doable through a standard
7424  * vm_map_remove() operation.  Otherwise, implicit clean-up
7425  * of the target map won't be able to correctly remove
7426  * these (and release the reference on the UPL).  Having
7427  * to do this means we can't map these into user-space
7428  * maps yet.
7429  */
7430 kern_return_t
vm_map_remove_upl_range(vm_map_t map,upl_t upl,__unused vm_object_offset_t offset_to_unmap,__unused upl_size_t size_to_unmap)7431 vm_map_remove_upl_range(
7432 	vm_map_t        map,
7433 	upl_t           upl,
7434 	__unused vm_object_offset_t    offset_to_unmap,
7435 	__unused upl_size_t      size_to_unmap)
7436 {
7437 	vm_address_t    addr;
7438 	upl_size_t      size;
7439 	int             isVectorUPL = 0, curr_upl = 0;
7440 	upl_t           vector_upl = NULL;
7441 
7442 	if (upl == UPL_NULL) {
7443 		return KERN_INVALID_ARGUMENT;
7444 	}
7445 
7446 	if ((isVectorUPL = vector_upl_is_valid(upl))) {
7447 		int     unmapped = 0, valid_upls = 0;
7448 		vector_upl = upl;
7449 		upl_lock(vector_upl);
7450 		for (curr_upl = 0; curr_upl < vector_upl_max_upls(vector_upl); curr_upl++) {
7451 			upl =  vector_upl_subupl_byindex(vector_upl, curr_upl );
7452 			if (upl == NULL) {
7453 				continue;
7454 			}
7455 			valid_upls++;
7456 			if (!(UPL_PAGE_LIST_MAPPED & upl->flags)) {
7457 				unmapped++;
7458 			}
7459 		}
7460 
7461 		if (unmapped) {
7462 			if (unmapped != valid_upls) {
7463 				panic("%d of the %d sub-upls within the Vector UPL is/are not mapped", unmapped, valid_upls);
7464 			} else {
7465 				upl_unlock(vector_upl);
7466 				return KERN_FAILURE;
7467 			}
7468 		}
7469 		curr_upl = 0;
7470 	} else {
7471 		upl_lock(upl);
7472 	}
7473 
7474 process_upl_to_remove:
7475 	if (isVectorUPL) {
7476 		if (curr_upl == vector_upl_max_upls(vector_upl)) {
7477 			vm_map_t v_upl_submap;
7478 			vm_offset_t v_upl_submap_dst_addr;
7479 			vector_upl_get_submap(vector_upl, &v_upl_submap, &v_upl_submap_dst_addr);
7480 
7481 			kmem_free_guard(map, v_upl_submap_dst_addr,
7482 			    vector_upl->u_size, KMF_NONE, KMEM_GUARD_SUBMAP);
7483 			vm_map_deallocate(v_upl_submap);
7484 			upl_unlock(vector_upl);
7485 			return KERN_SUCCESS;
7486 		}
7487 
7488 		upl =  vector_upl_subupl_byindex(vector_upl, curr_upl++ );
7489 		if (upl == NULL) {
7490 			goto process_upl_to_remove;
7491 		}
7492 	}
7493 
7494 	if (upl->flags & UPL_PAGE_LIST_MAPPED) {
7495 		addr = upl->kaddr;
7496 		size = upl->u_mapped_size;
7497 
7498 		assert(upl->ref_count > 1);
7499 		upl->ref_count--;               /* removing mapping ref */
7500 
7501 		upl->flags &= ~UPL_PAGE_LIST_MAPPED;
7502 		upl->kaddr = (vm_offset_t) 0;
7503 		upl->u_mapped_size = 0;
7504 
7505 		if (isVectorUPL) {
7506 			/*
7507 			 * If it's a Vectored UPL, we'll be removing the entire
7508 			 * submap anyways, so no need to remove individual UPL
7509 			 * element mappings from within the submap
7510 			 */
7511 			goto process_upl_to_remove;
7512 		}
7513 
7514 		upl_unlock(upl);
7515 
7516 		vm_map_remove(map,
7517 		    vm_map_trunc_page(addr, VM_MAP_PAGE_MASK(map)),
7518 		    vm_map_round_page(addr + size, VM_MAP_PAGE_MASK(map)));
7519 		return KERN_SUCCESS;
7520 	}
7521 	upl_unlock(upl);
7522 
7523 	return KERN_FAILURE;
7524 }
7525 
7526 kern_return_t
vm_map_remove_upl(vm_map_t map,upl_t upl)7527 vm_map_remove_upl(
7528 	vm_map_t        map,
7529 	upl_t           upl)
7530 {
7531 	upl_size_t upl_size = upl_adjusted_size(upl, VM_MAP_PAGE_MASK(map));
7532 	return vm_map_remove_upl_range(map, upl, 0, upl_size);
7533 }
7534 
7535 kern_return_t
upl_commit_range(upl_t upl,upl_offset_t offset,upl_size_t size,int flags,upl_page_info_t * page_list,mach_msg_type_number_t count,boolean_t * empty)7536 upl_commit_range(
7537 	upl_t                   upl,
7538 	upl_offset_t            offset,
7539 	upl_size_t              size,
7540 	int                     flags,
7541 	upl_page_info_t         *page_list,
7542 	mach_msg_type_number_t  count,
7543 	boolean_t               *empty)
7544 {
7545 	upl_size_t              xfer_size, subupl_size;
7546 	vm_object_t             shadow_object;
7547 	vm_object_t             object;
7548 	vm_object_t             m_object;
7549 	vm_object_offset_t      target_offset;
7550 	upl_offset_t            subupl_offset = offset;
7551 	int                     entry;
7552 	int                     occupied;
7553 	int                     clear_refmod = 0;
7554 	int                     pgpgout_count = 0;
7555 	struct  vm_page_delayed_work    dw_array;
7556 	struct  vm_page_delayed_work    *dwp, *dwp_start;
7557 	bool                    dwp_finish_ctx = TRUE;
7558 	int                     dw_count;
7559 	int                     dw_limit;
7560 	int                     isVectorUPL = 0;
7561 	upl_t                   vector_upl = NULL;
7562 	boolean_t               should_be_throttled = FALSE;
7563 
7564 	vm_page_t               nxt_page = VM_PAGE_NULL;
7565 	int                     fast_path_possible = 0;
7566 	int                     fast_path_full_commit = 0;
7567 	int                     throttle_page = 0;
7568 	int                     unwired_count = 0;
7569 	int                     local_queue_count = 0;
7570 	vm_page_t               first_local, last_local;
7571 	vm_object_offset_t      obj_start, obj_end, obj_offset;
7572 	kern_return_t           kr = KERN_SUCCESS;
7573 
7574 //	DEBUG4K_UPL("upl %p (u_offset 0x%llx u_size 0x%llx) object %p offset 0x%llx size 0x%llx flags 0x%x\n", upl, (uint64_t)upl->u_offset, (uint64_t)upl->u_size, upl->map_object, (uint64_t)offset, (uint64_t)size, flags);
7575 
7576 	dwp_start = dwp = NULL;
7577 
7578 	subupl_size = size;
7579 	*empty = FALSE;
7580 
7581 	if (upl == UPL_NULL) {
7582 		return KERN_INVALID_ARGUMENT;
7583 	}
7584 
7585 	dw_count = 0;
7586 	dw_limit = DELAYED_WORK_LIMIT(DEFAULT_DELAYED_WORK_LIMIT);
7587 	dwp_start = vm_page_delayed_work_get_ctx();
7588 	if (dwp_start == NULL) {
7589 		dwp_start = &dw_array;
7590 		dw_limit = 1;
7591 		dwp_finish_ctx = FALSE;
7592 	}
7593 
7594 	dwp = dwp_start;
7595 
7596 	if (count == 0) {
7597 		page_list = NULL;
7598 	}
7599 
7600 	if ((isVectorUPL = vector_upl_is_valid(upl))) {
7601 		vector_upl = upl;
7602 		upl_lock(vector_upl);
7603 	} else {
7604 		upl_lock(upl);
7605 	}
7606 
7607 process_upl_to_commit:
7608 
7609 	if (isVectorUPL) {
7610 		size = subupl_size;
7611 		offset = subupl_offset;
7612 		if (size == 0) {
7613 			upl_unlock(vector_upl);
7614 			kr = KERN_SUCCESS;
7615 			goto done;
7616 		}
7617 		upl =  vector_upl_subupl_byoffset(vector_upl, &offset, &size);
7618 		if (upl == NULL) {
7619 			upl_unlock(vector_upl);
7620 			kr = KERN_FAILURE;
7621 			goto done;
7622 		}
7623 		page_list = upl->page_list;
7624 		subupl_size -= size;
7625 		subupl_offset += size;
7626 	}
7627 
7628 #if UPL_DEBUG
7629 	if (upl->upl_commit_index < UPL_DEBUG_COMMIT_RECORDS) {
7630 		upl->upl_commit_records[upl->upl_commit_index].c_btref = btref_get(__builtin_frame_address(0), 0);
7631 		upl->upl_commit_records[upl->upl_commit_index].c_beg = offset;
7632 		upl->upl_commit_records[upl->upl_commit_index].c_end = (offset + size);
7633 
7634 		upl->upl_commit_index++;
7635 	}
7636 #endif
7637 	if (upl->flags & UPL_DEVICE_MEMORY) {
7638 		xfer_size = 0;
7639 	} else if ((offset + size) <= upl_adjusted_size(upl, PAGE_MASK)) {
7640 		xfer_size = size;
7641 	} else {
7642 		if (!isVectorUPL) {
7643 			upl_unlock(upl);
7644 		} else {
7645 			upl_unlock(vector_upl);
7646 		}
7647 		DEBUG4K_ERROR("upl %p (u_offset 0x%llx u_size 0x%x) offset 0x%x size 0x%x\n", upl, upl->u_offset, upl->u_size, offset, size);
7648 		kr = KERN_FAILURE;
7649 		goto done;
7650 	}
7651 	if (upl->flags & UPL_SET_DIRTY) {
7652 		flags |= UPL_COMMIT_SET_DIRTY;
7653 	}
7654 	if (upl->flags & UPL_CLEAR_DIRTY) {
7655 		flags |= UPL_COMMIT_CLEAR_DIRTY;
7656 	}
7657 
7658 	object = upl->map_object;
7659 
7660 	if (upl->flags & UPL_SHADOWED) {
7661 		vm_object_lock(object);
7662 		shadow_object = object->shadow;
7663 	} else {
7664 		shadow_object = object;
7665 	}
7666 	entry = offset / PAGE_SIZE;
7667 	target_offset = (vm_object_offset_t)offset;
7668 
7669 	if (upl->flags & UPL_KERNEL_OBJECT) {
7670 		vm_object_lock_shared(shadow_object);
7671 	} else {
7672 		vm_object_lock(shadow_object);
7673 	}
7674 
7675 	VM_OBJECT_WIRED_PAGE_UPDATE_START(shadow_object);
7676 
7677 	if (upl->flags & UPL_ACCESS_BLOCKED) {
7678 		assert(shadow_object->blocked_access);
7679 		shadow_object->blocked_access = FALSE;
7680 		vm_object_wakeup(object, VM_OBJECT_EVENT_UNBLOCKED);
7681 	}
7682 
7683 	if (shadow_object->code_signed) {
7684 		/*
7685 		 * CODE SIGNING:
7686 		 * If the object is code-signed, do not let this UPL tell
7687 		 * us if the pages are valid or not.  Let the pages be
7688 		 * validated by VM the normal way (when they get mapped or
7689 		 * copied).
7690 		 */
7691 		flags &= ~UPL_COMMIT_CS_VALIDATED;
7692 	}
7693 	if (!page_list) {
7694 		/*
7695 		 * No page list to get the code-signing info from !?
7696 		 */
7697 		flags &= ~UPL_COMMIT_CS_VALIDATED;
7698 	}
7699 	if (!VM_DYNAMIC_PAGING_ENABLED() && shadow_object->internal) {
7700 		should_be_throttled = TRUE;
7701 	}
7702 
7703 	if ((upl->flags & UPL_IO_WIRE) &&
7704 	    !(flags & UPL_COMMIT_FREE_ABSENT) &&
7705 	    !isVectorUPL &&
7706 	    shadow_object->purgable != VM_PURGABLE_VOLATILE &&
7707 	    shadow_object->purgable != VM_PURGABLE_EMPTY) {
7708 		if (!vm_page_queue_empty(&shadow_object->memq)) {
7709 			if (size == shadow_object->vo_size) {
7710 				nxt_page = (vm_page_t)vm_page_queue_first(&shadow_object->memq);
7711 				fast_path_full_commit = 1;
7712 			}
7713 			fast_path_possible = 1;
7714 
7715 			if (!VM_DYNAMIC_PAGING_ENABLED() && shadow_object->internal &&
7716 			    (shadow_object->purgable == VM_PURGABLE_DENY ||
7717 			    shadow_object->purgable == VM_PURGABLE_NONVOLATILE ||
7718 			    shadow_object->purgable == VM_PURGABLE_VOLATILE)) {
7719 				throttle_page = 1;
7720 			}
7721 		}
7722 	}
7723 	first_local = VM_PAGE_NULL;
7724 	last_local = VM_PAGE_NULL;
7725 
7726 	obj_start = target_offset + upl->u_offset - shadow_object->paging_offset;
7727 	obj_end = obj_start + xfer_size;
7728 	obj_start = vm_object_trunc_page(obj_start);
7729 	obj_end = vm_object_round_page(obj_end);
7730 	for (obj_offset = obj_start;
7731 	    obj_offset < obj_end;
7732 	    obj_offset += PAGE_SIZE) {
7733 		vm_page_t       t, m;
7734 
7735 		dwp->dw_mask = 0;
7736 		clear_refmod = 0;
7737 
7738 		m = VM_PAGE_NULL;
7739 
7740 		if (upl->flags & UPL_LITE) {
7741 			unsigned int    pg_num;
7742 
7743 			if (nxt_page != VM_PAGE_NULL) {
7744 				m = nxt_page;
7745 				nxt_page = (vm_page_t)vm_page_queue_next(&nxt_page->vmp_listq);
7746 				target_offset = m->vmp_offset;
7747 			}
7748 			pg_num = (unsigned int) (target_offset / PAGE_SIZE);
7749 			assert(pg_num == target_offset / PAGE_SIZE);
7750 
7751 			if (bitmap_test(upl->lite_list, pg_num)) {
7752 				bitmap_clear(upl->lite_list, pg_num);
7753 
7754 				if (!(upl->flags & UPL_KERNEL_OBJECT) && m == VM_PAGE_NULL) {
7755 					m = vm_page_lookup(shadow_object, obj_offset);
7756 				}
7757 			} else {
7758 				m = NULL;
7759 			}
7760 		}
7761 		if (upl->flags & UPL_SHADOWED) {
7762 			if ((t = vm_page_lookup(object, target_offset)) != VM_PAGE_NULL) {
7763 				t->vmp_free_when_done = FALSE;
7764 
7765 				VM_PAGE_FREE(t);
7766 
7767 				if (!(upl->flags & UPL_KERNEL_OBJECT) && m == VM_PAGE_NULL) {
7768 					m = vm_page_lookup(shadow_object, target_offset + object->vo_shadow_offset);
7769 				}
7770 			}
7771 		}
7772 		if (m == VM_PAGE_NULL) {
7773 			goto commit_next_page;
7774 		}
7775 
7776 		m_object = VM_PAGE_OBJECT(m);
7777 
7778 		if (m->vmp_q_state == VM_PAGE_USED_BY_COMPRESSOR) {
7779 			assert(m->vmp_busy);
7780 
7781 			dwp->dw_mask |= (DW_clear_busy | DW_PAGE_WAKEUP);
7782 			goto commit_next_page;
7783 		}
7784 
7785 		if (flags & UPL_COMMIT_CS_VALIDATED) {
7786 			/*
7787 			 * CODE SIGNING:
7788 			 * Set the code signing bits according to
7789 			 * what the UPL says they should be.
7790 			 */
7791 			m->vmp_cs_validated |= page_list[entry].cs_validated;
7792 			m->vmp_cs_tainted |= page_list[entry].cs_tainted;
7793 			m->vmp_cs_nx |= page_list[entry].cs_nx;
7794 		}
7795 		if (flags & UPL_COMMIT_WRITTEN_BY_KERNEL) {
7796 			m->vmp_written_by_kernel = TRUE;
7797 		}
7798 
7799 		if (upl->flags & UPL_IO_WIRE) {
7800 			if (page_list) {
7801 				page_list[entry].phys_addr = 0;
7802 			}
7803 
7804 			if (flags & UPL_COMMIT_SET_DIRTY) {
7805 				SET_PAGE_DIRTY(m, FALSE);
7806 			} else if (flags & UPL_COMMIT_CLEAR_DIRTY) {
7807 				m->vmp_dirty = FALSE;
7808 
7809 				if (!(flags & UPL_COMMIT_CS_VALIDATED) &&
7810 				    m->vmp_cs_validated &&
7811 				    m->vmp_cs_tainted != VMP_CS_ALL_TRUE) {
7812 					/*
7813 					 * CODE SIGNING:
7814 					 * This page is no longer dirty
7815 					 * but could have been modified,
7816 					 * so it will need to be
7817 					 * re-validated.
7818 					 */
7819 					m->vmp_cs_validated = VMP_CS_ALL_FALSE;
7820 
7821 					VM_PAGEOUT_DEBUG(vm_cs_validated_resets, 1);
7822 
7823 					pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m));
7824 				}
7825 				clear_refmod |= VM_MEM_MODIFIED;
7826 			}
7827 			if (upl->flags & UPL_ACCESS_BLOCKED) {
7828 				/*
7829 				 * We blocked access to the pages in this UPL.
7830 				 * Clear the "busy" bit and wake up any waiter
7831 				 * for this page.
7832 				 */
7833 				dwp->dw_mask |= (DW_clear_busy | DW_PAGE_WAKEUP);
7834 			}
7835 			if (fast_path_possible) {
7836 				assert(m_object->purgable != VM_PURGABLE_EMPTY);
7837 				assert(m_object->purgable != VM_PURGABLE_VOLATILE);
7838 				if (m->vmp_absent) {
7839 					assert(m->vmp_q_state == VM_PAGE_NOT_ON_Q);
7840 					assert(m->vmp_wire_count == 0);
7841 					assert(m->vmp_busy);
7842 
7843 					m->vmp_absent = FALSE;
7844 					dwp->dw_mask |= (DW_clear_busy | DW_PAGE_WAKEUP);
7845 				} else {
7846 					if (m->vmp_wire_count == 0) {
7847 						panic("wire_count == 0, m = %p, obj = %p", m, shadow_object);
7848 					}
7849 					assert(m->vmp_q_state == VM_PAGE_IS_WIRED);
7850 
7851 					/*
7852 					 * XXX FBDP need to update some other
7853 					 * counters here (purgeable_wired_count)
7854 					 * (ledgers), ...
7855 					 */
7856 					assert(m->vmp_wire_count > 0);
7857 					m->vmp_wire_count--;
7858 
7859 					if (m->vmp_wire_count == 0) {
7860 						m->vmp_q_state = VM_PAGE_NOT_ON_Q;
7861 						unwired_count++;
7862 					}
7863 				}
7864 				if (m->vmp_wire_count == 0) {
7865 					assert(m->vmp_pageq.next == 0 && m->vmp_pageq.prev == 0);
7866 
7867 					if (last_local == VM_PAGE_NULL) {
7868 						assert(first_local == VM_PAGE_NULL);
7869 
7870 						last_local = m;
7871 						first_local = m;
7872 					} else {
7873 						assert(first_local != VM_PAGE_NULL);
7874 
7875 						m->vmp_pageq.next = VM_PAGE_CONVERT_TO_QUEUE_ENTRY(first_local);
7876 						first_local->vmp_pageq.prev = VM_PAGE_CONVERT_TO_QUEUE_ENTRY(m);
7877 						first_local = m;
7878 					}
7879 					local_queue_count++;
7880 
7881 					if (throttle_page) {
7882 						m->vmp_q_state = VM_PAGE_ON_THROTTLED_Q;
7883 					} else {
7884 						if (flags & UPL_COMMIT_INACTIVATE) {
7885 							if (shadow_object->internal) {
7886 								m->vmp_q_state = VM_PAGE_ON_INACTIVE_INTERNAL_Q;
7887 							} else {
7888 								m->vmp_q_state = VM_PAGE_ON_INACTIVE_EXTERNAL_Q;
7889 							}
7890 						} else {
7891 							m->vmp_q_state = VM_PAGE_ON_ACTIVE_Q;
7892 						}
7893 					}
7894 				}
7895 			} else {
7896 				if (flags & UPL_COMMIT_INACTIVATE) {
7897 					dwp->dw_mask |= DW_vm_page_deactivate_internal;
7898 					clear_refmod |= VM_MEM_REFERENCED;
7899 				}
7900 				if (m->vmp_absent) {
7901 					if (flags & UPL_COMMIT_FREE_ABSENT) {
7902 						dwp->dw_mask |= DW_vm_page_free;
7903 					} else {
7904 						m->vmp_absent = FALSE;
7905 						dwp->dw_mask |= (DW_clear_busy | DW_PAGE_WAKEUP);
7906 
7907 						if (!(dwp->dw_mask & DW_vm_page_deactivate_internal)) {
7908 							dwp->dw_mask |= DW_vm_page_activate;
7909 						}
7910 					}
7911 				} else {
7912 					dwp->dw_mask |= DW_vm_page_unwire;
7913 				}
7914 			}
7915 			goto commit_next_page;
7916 		}
7917 		assert(m->vmp_q_state != VM_PAGE_USED_BY_COMPRESSOR);
7918 
7919 		if (page_list) {
7920 			page_list[entry].phys_addr = 0;
7921 		}
7922 
7923 		/*
7924 		 * make sure to clear the hardware
7925 		 * modify or reference bits before
7926 		 * releasing the BUSY bit on this page
7927 		 * otherwise we risk losing a legitimate
7928 		 * change of state
7929 		 */
7930 		if (flags & UPL_COMMIT_CLEAR_DIRTY) {
7931 			m->vmp_dirty = FALSE;
7932 
7933 			clear_refmod |= VM_MEM_MODIFIED;
7934 		}
7935 		if (m->vmp_laundry) {
7936 			dwp->dw_mask |= DW_vm_pageout_throttle_up;
7937 		}
7938 
7939 		if (VM_PAGE_WIRED(m)) {
7940 			m->vmp_free_when_done = FALSE;
7941 		}
7942 
7943 		if (!(flags & UPL_COMMIT_CS_VALIDATED) &&
7944 		    m->vmp_cs_validated &&
7945 		    m->vmp_cs_tainted != VMP_CS_ALL_TRUE) {
7946 			/*
7947 			 * CODE SIGNING:
7948 			 * This page is no longer dirty
7949 			 * but could have been modified,
7950 			 * so it will need to be
7951 			 * re-validated.
7952 			 */
7953 			m->vmp_cs_validated = VMP_CS_ALL_FALSE;
7954 
7955 			VM_PAGEOUT_DEBUG(vm_cs_validated_resets, 1);
7956 
7957 			pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m));
7958 		}
7959 		if (m->vmp_overwriting) {
7960 			/*
7961 			 * the (COPY_OUT_FROM == FALSE) request_page_list case
7962 			 */
7963 			if (m->vmp_busy) {
7964 #if CONFIG_PHANTOM_CACHE
7965 				if (m->vmp_absent && !m_object->internal) {
7966 					dwp->dw_mask |= DW_vm_phantom_cache_update;
7967 				}
7968 #endif
7969 				m->vmp_absent = FALSE;
7970 
7971 				dwp->dw_mask |= DW_clear_busy;
7972 			} else {
7973 				/*
7974 				 * alternate (COPY_OUT_FROM == FALSE) page_list case
7975 				 * Occurs when the original page was wired
7976 				 * at the time of the list request
7977 				 */
7978 				assert(VM_PAGE_WIRED(m));
7979 
7980 				dwp->dw_mask |= DW_vm_page_unwire; /* reactivates */
7981 			}
7982 			m->vmp_overwriting = FALSE;
7983 		}
7984 		m->vmp_cleaning = FALSE;
7985 
7986 		if (m->vmp_free_when_done) {
7987 			/*
7988 			 * With the clean queue enabled, UPL_PAGEOUT should
7989 			 * no longer set the pageout bit. Its pages now go
7990 			 * to the clean queue.
7991 			 *
7992 			 * We don't use the cleaned Q anymore and so this
7993 			 * assert isn't correct. The code for the clean Q
7994 			 * still exists and might be used in the future. If we
7995 			 * go back to the cleaned Q, we will re-enable this
7996 			 * assert.
7997 			 *
7998 			 * assert(!(upl->flags & UPL_PAGEOUT));
7999 			 */
8000 			assert(!m_object->internal);
8001 
8002 			m->vmp_free_when_done = FALSE;
8003 
8004 			if ((flags & UPL_COMMIT_SET_DIRTY) ||
8005 			    (m->vmp_pmapped && (pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m)) & VM_MEM_MODIFIED))) {
8006 				/*
8007 				 * page was re-dirtied after we started
8008 				 * the pageout... reactivate it since
8009 				 * we don't know whether the on-disk
8010 				 * copy matches what is now in memory
8011 				 */
8012 				SET_PAGE_DIRTY(m, FALSE);
8013 
8014 				dwp->dw_mask |= DW_vm_page_activate | DW_PAGE_WAKEUP;
8015 
8016 				if (upl->flags & UPL_PAGEOUT) {
8017 					counter_inc(&vm_statistics_reactivations);
8018 					DTRACE_VM2(pgrec, int, 1, (uint64_t *), NULL);
8019 				}
8020 			} else if (m->vmp_busy && !(upl->flags & UPL_HAS_BUSY)) {
8021 				/*
8022 				 * Someone else might still be handling this
8023 				 * page (vm_fault() for example), so let's not
8024 				 * free it or "un-busy" it!
8025 				 * Put that page in the "speculative" queue
8026 				 * for now (since we would otherwise have freed
8027 				 * it) and let whoever is keeping the page
8028 				 * "busy" move it if needed when they're done
8029 				 * with it.
8030 				 */
8031 				dwp->dw_mask |= DW_vm_page_speculate;
8032 			} else {
8033 				/*
8034 				 * page has been successfully cleaned
8035 				 * go ahead and free it for other use
8036 				 */
8037 				if (m_object->internal) {
8038 					DTRACE_VM2(anonpgout, int, 1, (uint64_t *), NULL);
8039 				} else {
8040 					DTRACE_VM2(fspgout, int, 1, (uint64_t *), NULL);
8041 				}
8042 				m->vmp_dirty = FALSE;
8043 				if (!(upl->flags & UPL_HAS_BUSY)) {
8044 					assert(!m->vmp_busy);
8045 				}
8046 				m->vmp_busy = TRUE;
8047 
8048 				dwp->dw_mask |= DW_vm_page_free;
8049 			}
8050 			goto commit_next_page;
8051 		}
8052 		/*
8053 		 * It is a part of the semantic of COPYOUT_FROM
8054 		 * UPLs that a commit implies cache sync
8055 		 * between the vm page and the backing store
8056 		 * this can be used to strip the precious bit
8057 		 * as well as clean
8058 		 */
8059 		if ((upl->flags & UPL_PAGE_SYNC_DONE) || (flags & UPL_COMMIT_CLEAR_PRECIOUS)) {
8060 			m->vmp_precious = FALSE;
8061 		}
8062 
8063 		if (flags & UPL_COMMIT_SET_DIRTY) {
8064 			SET_PAGE_DIRTY(m, FALSE);
8065 		} else {
8066 			m->vmp_dirty = FALSE;
8067 		}
8068 
8069 		/* with the clean queue on, move *all* cleaned pages to the clean queue */
8070 		if (hibernate_cleaning_in_progress == FALSE && !m->vmp_dirty && (upl->flags & UPL_PAGEOUT)) {
8071 			pgpgout_count++;
8072 
8073 			counter_inc(&vm_statistics_pageouts);
8074 			DTRACE_VM2(pgout, int, 1, (uint64_t *), NULL);
8075 
8076 			dwp->dw_mask |= DW_enqueue_cleaned;
8077 		} else if (should_be_throttled == TRUE && (m->vmp_q_state == VM_PAGE_NOT_ON_Q)) {
8078 			/*
8079 			 * page coming back in from being 'frozen'...
8080 			 * it was dirty before it was frozen, so keep it so
8081 			 * the vm_page_activate will notice that it really belongs
8082 			 * on the throttle queue and put it there
8083 			 */
8084 			SET_PAGE_DIRTY(m, FALSE);
8085 			dwp->dw_mask |= DW_vm_page_activate;
8086 		} else {
8087 			if ((flags & UPL_COMMIT_INACTIVATE) && !m->vmp_clustered && (m->vmp_q_state != VM_PAGE_ON_SPECULATIVE_Q)) {
8088 				dwp->dw_mask |= DW_vm_page_deactivate_internal;
8089 				clear_refmod |= VM_MEM_REFERENCED;
8090 			} else if (!VM_PAGE_PAGEABLE(m)) {
8091 				if (m->vmp_clustered || (flags & UPL_COMMIT_SPECULATE)) {
8092 					dwp->dw_mask |= DW_vm_page_speculate;
8093 				} else if (m->vmp_reference) {
8094 					dwp->dw_mask |= DW_vm_page_activate;
8095 				} else {
8096 					dwp->dw_mask |= DW_vm_page_deactivate_internal;
8097 					clear_refmod |= VM_MEM_REFERENCED;
8098 				}
8099 			}
8100 		}
8101 		if (upl->flags & UPL_ACCESS_BLOCKED) {
8102 			/*
8103 			 * We blocked access to the pages in this URL.
8104 			 * Clear the "busy" bit on this page before we
8105 			 * wake up any waiter.
8106 			 */
8107 			dwp->dw_mask |= DW_clear_busy;
8108 		}
8109 		/*
8110 		 * Wakeup any thread waiting for the page to be un-cleaning.
8111 		 */
8112 		dwp->dw_mask |= DW_PAGE_WAKEUP;
8113 
8114 commit_next_page:
8115 		if (clear_refmod) {
8116 			pmap_clear_refmod(VM_PAGE_GET_PHYS_PAGE(m), clear_refmod);
8117 		}
8118 
8119 		target_offset += PAGE_SIZE_64;
8120 		xfer_size -= PAGE_SIZE;
8121 		entry++;
8122 
8123 		if (dwp->dw_mask) {
8124 			if (dwp->dw_mask & ~(DW_clear_busy | DW_PAGE_WAKEUP)) {
8125 				VM_PAGE_ADD_DELAYED_WORK(dwp, m, dw_count);
8126 
8127 				if (dw_count >= dw_limit) {
8128 					vm_page_do_delayed_work(shadow_object, VM_KERN_MEMORY_NONE, dwp_start, dw_count);
8129 
8130 					dwp = dwp_start;
8131 					dw_count = 0;
8132 				}
8133 			} else {
8134 				if (dwp->dw_mask & DW_clear_busy) {
8135 					m->vmp_busy = FALSE;
8136 				}
8137 
8138 				if (dwp->dw_mask & DW_PAGE_WAKEUP) {
8139 					PAGE_WAKEUP(m);
8140 				}
8141 			}
8142 		}
8143 	}
8144 	if (dw_count) {
8145 		vm_page_do_delayed_work(shadow_object, VM_KERN_MEMORY_NONE, dwp_start, dw_count);
8146 		dwp = dwp_start;
8147 		dw_count = 0;
8148 	}
8149 
8150 	if (fast_path_possible) {
8151 		assert(shadow_object->purgable != VM_PURGABLE_VOLATILE);
8152 		assert(shadow_object->purgable != VM_PURGABLE_EMPTY);
8153 
8154 		if (local_queue_count || unwired_count) {
8155 			if (local_queue_count) {
8156 				vm_page_t       first_target;
8157 				vm_page_queue_head_t    *target_queue;
8158 
8159 				if (throttle_page) {
8160 					target_queue = &vm_page_queue_throttled;
8161 				} else {
8162 					if (flags & UPL_COMMIT_INACTIVATE) {
8163 						if (shadow_object->internal) {
8164 							target_queue = &vm_page_queue_anonymous;
8165 						} else {
8166 							target_queue = &vm_page_queue_inactive;
8167 						}
8168 					} else {
8169 						target_queue = &vm_page_queue_active;
8170 					}
8171 				}
8172 				/*
8173 				 * Transfer the entire local queue to a regular LRU page queues.
8174 				 */
8175 				vm_page_lockspin_queues();
8176 
8177 				first_target = (vm_page_t) vm_page_queue_first(target_queue);
8178 
8179 				if (vm_page_queue_empty(target_queue)) {
8180 					target_queue->prev = VM_PAGE_CONVERT_TO_QUEUE_ENTRY(last_local);
8181 				} else {
8182 					first_target->vmp_pageq.prev = VM_PAGE_CONVERT_TO_QUEUE_ENTRY(last_local);
8183 				}
8184 
8185 				target_queue->next = VM_PAGE_CONVERT_TO_QUEUE_ENTRY(first_local);
8186 				first_local->vmp_pageq.prev = VM_PAGE_CONVERT_TO_QUEUE_ENTRY(target_queue);
8187 				last_local->vmp_pageq.next = VM_PAGE_CONVERT_TO_QUEUE_ENTRY(first_target);
8188 
8189 				/*
8190 				 * Adjust the global page counts.
8191 				 */
8192 				if (throttle_page) {
8193 					vm_page_throttled_count += local_queue_count;
8194 				} else {
8195 					if (flags & UPL_COMMIT_INACTIVATE) {
8196 						if (shadow_object->internal) {
8197 							vm_page_anonymous_count += local_queue_count;
8198 						}
8199 						vm_page_inactive_count += local_queue_count;
8200 
8201 						token_new_pagecount += local_queue_count;
8202 					} else {
8203 						vm_page_active_count += local_queue_count;
8204 					}
8205 
8206 					if (shadow_object->internal) {
8207 						vm_page_pageable_internal_count += local_queue_count;
8208 					} else {
8209 						vm_page_pageable_external_count += local_queue_count;
8210 					}
8211 				}
8212 			} else {
8213 				vm_page_lockspin_queues();
8214 			}
8215 			if (unwired_count) {
8216 				vm_page_wire_count -= unwired_count;
8217 				VM_CHECK_MEMORYSTATUS;
8218 			}
8219 			vm_page_unlock_queues();
8220 
8221 			VM_OBJECT_WIRED_PAGE_COUNT(shadow_object, -unwired_count);
8222 		}
8223 	}
8224 
8225 	if (upl->flags & UPL_DEVICE_MEMORY) {
8226 		occupied = 0;
8227 	} else if (upl->flags & UPL_LITE) {
8228 		uint32_t pages = (uint32_t)atop(upl_adjusted_size(upl, PAGE_MASK));
8229 
8230 		occupied = !fast_path_full_commit &&
8231 		    !bitmap_is_empty(upl->lite_list, pages);
8232 	} else {
8233 		occupied = !vm_page_queue_empty(&upl->map_object->memq);
8234 	}
8235 	if (occupied == 0) {
8236 		/*
8237 		 * If this UPL element belongs to a Vector UPL and is
8238 		 * empty, then this is the right function to deallocate
8239 		 * it. So go ahead set the *empty variable. The flag
8240 		 * UPL_COMMIT_NOTIFY_EMPTY, from the caller's point of view
8241 		 * should be considered relevant for the Vector UPL and not
8242 		 * the internal UPLs.
8243 		 */
8244 		if ((upl->flags & UPL_COMMIT_NOTIFY_EMPTY) || isVectorUPL) {
8245 			*empty = TRUE;
8246 		}
8247 
8248 		if (object == shadow_object && !(upl->flags & UPL_KERNEL_OBJECT)) {
8249 			/*
8250 			 * this is not a paging object
8251 			 * so we need to drop the paging reference
8252 			 * that was taken when we created the UPL
8253 			 * against this object
8254 			 */
8255 			vm_object_activity_end(shadow_object);
8256 			vm_object_collapse(shadow_object, 0, TRUE);
8257 		} else {
8258 			/*
8259 			 * we dontated the paging reference to
8260 			 * the map object... vm_pageout_object_terminate
8261 			 * will drop this reference
8262 			 */
8263 		}
8264 	}
8265 	VM_OBJECT_WIRED_PAGE_UPDATE_END(shadow_object, shadow_object->wire_tag);
8266 	vm_object_unlock(shadow_object);
8267 	if (object != shadow_object) {
8268 		vm_object_unlock(object);
8269 	}
8270 
8271 	if (!isVectorUPL) {
8272 		upl_unlock(upl);
8273 	} else {
8274 		/*
8275 		 * If we completed our operations on an UPL that is
8276 		 * part of a Vectored UPL and if empty is TRUE, then
8277 		 * we should go ahead and deallocate this UPL element.
8278 		 * Then we check if this was the last of the UPL elements
8279 		 * within that Vectored UPL. If so, set empty to TRUE
8280 		 * so that in ubc_upl_commit_range or ubc_upl_commit, we
8281 		 * can go ahead and deallocate the Vector UPL too.
8282 		 */
8283 		if (*empty == TRUE) {
8284 			*empty = vector_upl_set_subupl(vector_upl, upl, 0);
8285 			upl_deallocate(upl);
8286 		}
8287 		goto process_upl_to_commit;
8288 	}
8289 	if (pgpgout_count) {
8290 		DTRACE_VM2(pgpgout, int, pgpgout_count, (uint64_t *), NULL);
8291 	}
8292 
8293 	kr = KERN_SUCCESS;
8294 done:
8295 	if (dwp_start && dwp_finish_ctx) {
8296 		vm_page_delayed_work_finish_ctx(dwp_start);
8297 		dwp_start = dwp = NULL;
8298 	}
8299 
8300 	return kr;
8301 }
8302 
8303 kern_return_t
upl_abort_range(upl_t upl,upl_offset_t offset,upl_size_t size,int error,boolean_t * empty)8304 upl_abort_range(
8305 	upl_t                   upl,
8306 	upl_offset_t            offset,
8307 	upl_size_t              size,
8308 	int                     error,
8309 	boolean_t               *empty)
8310 {
8311 	upl_size_t              xfer_size, subupl_size;
8312 	vm_object_t             shadow_object;
8313 	vm_object_t             object;
8314 	vm_object_offset_t      target_offset;
8315 	upl_offset_t            subupl_offset = offset;
8316 	int                     occupied;
8317 	struct  vm_page_delayed_work    dw_array;
8318 	struct  vm_page_delayed_work    *dwp, *dwp_start;
8319 	bool                    dwp_finish_ctx = TRUE;
8320 	int                     dw_count;
8321 	int                     dw_limit;
8322 	int                     isVectorUPL = 0;
8323 	upl_t                   vector_upl = NULL;
8324 	vm_object_offset_t      obj_start, obj_end, obj_offset;
8325 	kern_return_t           kr = KERN_SUCCESS;
8326 
8327 //	DEBUG4K_UPL("upl %p (u_offset 0x%llx u_size 0x%llx) object %p offset 0x%llx size 0x%llx error 0x%x\n", upl, (uint64_t)upl->u_offset, (uint64_t)upl->u_size, upl->map_object, (uint64_t)offset, (uint64_t)size, error);
8328 
8329 	dwp_start = dwp = NULL;
8330 
8331 	subupl_size = size;
8332 	*empty = FALSE;
8333 
8334 	if (upl == UPL_NULL) {
8335 		return KERN_INVALID_ARGUMENT;
8336 	}
8337 
8338 	if ((upl->flags & UPL_IO_WIRE) && !(error & UPL_ABORT_DUMP_PAGES)) {
8339 		return upl_commit_range(upl, offset, size, UPL_COMMIT_FREE_ABSENT, NULL, 0, empty);
8340 	}
8341 
8342 	dw_count = 0;
8343 	dw_limit = DELAYED_WORK_LIMIT(DEFAULT_DELAYED_WORK_LIMIT);
8344 	dwp_start = vm_page_delayed_work_get_ctx();
8345 	if (dwp_start == NULL) {
8346 		dwp_start = &dw_array;
8347 		dw_limit = 1;
8348 		dwp_finish_ctx = FALSE;
8349 	}
8350 
8351 	dwp = dwp_start;
8352 
8353 	if ((isVectorUPL = vector_upl_is_valid(upl))) {
8354 		vector_upl = upl;
8355 		upl_lock(vector_upl);
8356 	} else {
8357 		upl_lock(upl);
8358 	}
8359 
8360 process_upl_to_abort:
8361 	if (isVectorUPL) {
8362 		size = subupl_size;
8363 		offset = subupl_offset;
8364 		if (size == 0) {
8365 			upl_unlock(vector_upl);
8366 			kr = KERN_SUCCESS;
8367 			goto done;
8368 		}
8369 		upl =  vector_upl_subupl_byoffset(vector_upl, &offset, &size);
8370 		if (upl == NULL) {
8371 			upl_unlock(vector_upl);
8372 			kr = KERN_FAILURE;
8373 			goto done;
8374 		}
8375 		subupl_size -= size;
8376 		subupl_offset += size;
8377 	}
8378 
8379 	*empty = FALSE;
8380 
8381 #if UPL_DEBUG
8382 	if (upl->upl_commit_index < UPL_DEBUG_COMMIT_RECORDS) {
8383 		upl->upl_commit_records[upl->upl_commit_index].c_btref = btref_get(__builtin_frame_address(0), 0);
8384 		upl->upl_commit_records[upl->upl_commit_index].c_beg = offset;
8385 		upl->upl_commit_records[upl->upl_commit_index].c_end = (offset + size);
8386 		upl->upl_commit_records[upl->upl_commit_index].c_aborted = 1;
8387 
8388 		upl->upl_commit_index++;
8389 	}
8390 #endif
8391 	if (upl->flags & UPL_DEVICE_MEMORY) {
8392 		xfer_size = 0;
8393 	} else if ((offset + size) <= upl_adjusted_size(upl, PAGE_MASK)) {
8394 		xfer_size = size;
8395 	} else {
8396 		if (!isVectorUPL) {
8397 			upl_unlock(upl);
8398 		} else {
8399 			upl_unlock(vector_upl);
8400 		}
8401 		DEBUG4K_ERROR("upl %p (u_offset 0x%llx u_size 0x%x) offset 0x%x size 0x%x\n", upl, upl->u_offset, upl->u_size, offset, size);
8402 		kr = KERN_FAILURE;
8403 		goto done;
8404 	}
8405 	object = upl->map_object;
8406 
8407 	if (upl->flags & UPL_SHADOWED) {
8408 		vm_object_lock(object);
8409 		shadow_object = object->shadow;
8410 	} else {
8411 		shadow_object = object;
8412 	}
8413 
8414 	target_offset = (vm_object_offset_t)offset;
8415 
8416 	if (upl->flags & UPL_KERNEL_OBJECT) {
8417 		vm_object_lock_shared(shadow_object);
8418 	} else {
8419 		vm_object_lock(shadow_object);
8420 	}
8421 
8422 	if (upl->flags & UPL_ACCESS_BLOCKED) {
8423 		assert(shadow_object->blocked_access);
8424 		shadow_object->blocked_access = FALSE;
8425 		vm_object_wakeup(object, VM_OBJECT_EVENT_UNBLOCKED);
8426 	}
8427 
8428 	if ((error & UPL_ABORT_DUMP_PAGES) && (upl->flags & UPL_KERNEL_OBJECT)) {
8429 		panic("upl_abort_range: kernel_object being DUMPED");
8430 	}
8431 
8432 	obj_start = target_offset + upl->u_offset - shadow_object->paging_offset;
8433 	obj_end = obj_start + xfer_size;
8434 	obj_start = vm_object_trunc_page(obj_start);
8435 	obj_end = vm_object_round_page(obj_end);
8436 	for (obj_offset = obj_start;
8437 	    obj_offset < obj_end;
8438 	    obj_offset += PAGE_SIZE) {
8439 		vm_page_t       t, m;
8440 		unsigned int    pg_num;
8441 		boolean_t       needed;
8442 
8443 		pg_num = (unsigned int) (target_offset / PAGE_SIZE);
8444 		assert(pg_num == target_offset / PAGE_SIZE);
8445 
8446 		needed = FALSE;
8447 
8448 		if (upl->flags & UPL_INTERNAL) {
8449 			needed = upl->page_list[pg_num].needed;
8450 		}
8451 
8452 		dwp->dw_mask = 0;
8453 		m = VM_PAGE_NULL;
8454 
8455 		if (upl->flags & UPL_LITE) {
8456 			if (bitmap_test(upl->lite_list, pg_num)) {
8457 				bitmap_clear(upl->lite_list, pg_num);
8458 
8459 				if (!(upl->flags & UPL_KERNEL_OBJECT)) {
8460 					m = vm_page_lookup(shadow_object, obj_offset);
8461 				}
8462 			}
8463 		}
8464 		if (upl->flags & UPL_SHADOWED) {
8465 			if ((t = vm_page_lookup(object, target_offset)) != VM_PAGE_NULL) {
8466 				t->vmp_free_when_done = FALSE;
8467 
8468 				VM_PAGE_FREE(t);
8469 
8470 				if (m == VM_PAGE_NULL) {
8471 					m = vm_page_lookup(shadow_object, target_offset + object->vo_shadow_offset);
8472 				}
8473 			}
8474 		}
8475 		if ((upl->flags & UPL_KERNEL_OBJECT)) {
8476 			goto abort_next_page;
8477 		}
8478 
8479 		if (m != VM_PAGE_NULL) {
8480 			assert(m->vmp_q_state != VM_PAGE_USED_BY_COMPRESSOR);
8481 
8482 			if (m->vmp_absent) {
8483 				boolean_t must_free = TRUE;
8484 
8485 				/*
8486 				 * COPYOUT = FALSE case
8487 				 * check for error conditions which must
8488 				 * be passed back to the pages customer
8489 				 */
8490 				if (error & UPL_ABORT_RESTART) {
8491 					m->vmp_restart = TRUE;
8492 					m->vmp_absent = FALSE;
8493 					m->vmp_unusual = TRUE;
8494 					must_free = FALSE;
8495 				} else if (error & UPL_ABORT_UNAVAILABLE) {
8496 					m->vmp_restart = FALSE;
8497 					m->vmp_unusual = TRUE;
8498 					must_free = FALSE;
8499 				} else if (error & UPL_ABORT_ERROR) {
8500 					m->vmp_restart = FALSE;
8501 					m->vmp_absent = FALSE;
8502 					m->vmp_error = TRUE;
8503 					m->vmp_unusual = TRUE;
8504 					must_free = FALSE;
8505 				}
8506 				if (m->vmp_clustered && needed == FALSE) {
8507 					/*
8508 					 * This page was a part of a speculative
8509 					 * read-ahead initiated by the kernel
8510 					 * itself.  No one is expecting this
8511 					 * page and no one will clean up its
8512 					 * error state if it ever becomes valid
8513 					 * in the future.
8514 					 * We have to free it here.
8515 					 */
8516 					must_free = TRUE;
8517 				}
8518 				m->vmp_cleaning = FALSE;
8519 
8520 				if (m->vmp_overwriting && !m->vmp_busy) {
8521 					/*
8522 					 * this shouldn't happen since
8523 					 * this is an 'absent' page, but
8524 					 * it doesn't hurt to check for
8525 					 * the 'alternate' method of
8526 					 * stabilizing the page...
8527 					 * we will mark 'busy' to be cleared
8528 					 * in the following code which will
8529 					 * take care of the primary stabilzation
8530 					 * method (i.e. setting 'busy' to TRUE)
8531 					 */
8532 					dwp->dw_mask |= DW_vm_page_unwire;
8533 				}
8534 				m->vmp_overwriting = FALSE;
8535 
8536 				dwp->dw_mask |= (DW_clear_busy | DW_PAGE_WAKEUP);
8537 
8538 				if (must_free == TRUE) {
8539 					dwp->dw_mask |= DW_vm_page_free;
8540 				} else {
8541 					dwp->dw_mask |= DW_vm_page_activate;
8542 				}
8543 			} else {
8544 				/*
8545 				 * Handle the trusted pager throttle.
8546 				 */
8547 				if (m->vmp_laundry) {
8548 					dwp->dw_mask |= DW_vm_pageout_throttle_up;
8549 				}
8550 
8551 				if (upl->flags & UPL_ACCESS_BLOCKED) {
8552 					/*
8553 					 * We blocked access to the pages in this UPL.
8554 					 * Clear the "busy" bit and wake up any waiter
8555 					 * for this page.
8556 					 */
8557 					dwp->dw_mask |= DW_clear_busy;
8558 				}
8559 				if (m->vmp_overwriting) {
8560 					if (m->vmp_busy) {
8561 						dwp->dw_mask |= DW_clear_busy;
8562 					} else {
8563 						/*
8564 						 * deal with the 'alternate' method
8565 						 * of stabilizing the page...
8566 						 * we will either free the page
8567 						 * or mark 'busy' to be cleared
8568 						 * in the following code which will
8569 						 * take care of the primary stabilzation
8570 						 * method (i.e. setting 'busy' to TRUE)
8571 						 */
8572 						dwp->dw_mask |= DW_vm_page_unwire;
8573 					}
8574 					m->vmp_overwriting = FALSE;
8575 				}
8576 				m->vmp_free_when_done = FALSE;
8577 				m->vmp_cleaning = FALSE;
8578 
8579 				if (error & UPL_ABORT_DUMP_PAGES) {
8580 					pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(m));
8581 
8582 					dwp->dw_mask |= DW_vm_page_free;
8583 				} else {
8584 					if (!(dwp->dw_mask & DW_vm_page_unwire)) {
8585 						if (error & UPL_ABORT_REFERENCE) {
8586 							/*
8587 							 * we've been told to explictly
8588 							 * reference this page... for
8589 							 * file I/O, this is done by
8590 							 * implementing an LRU on the inactive q
8591 							 */
8592 							dwp->dw_mask |= DW_vm_page_lru;
8593 						} else if (!VM_PAGE_PAGEABLE(m)) {
8594 							dwp->dw_mask |= DW_vm_page_deactivate_internal;
8595 						}
8596 					}
8597 					dwp->dw_mask |= DW_PAGE_WAKEUP;
8598 				}
8599 			}
8600 		}
8601 abort_next_page:
8602 		target_offset += PAGE_SIZE_64;
8603 		xfer_size -= PAGE_SIZE;
8604 
8605 		if (dwp->dw_mask) {
8606 			if (dwp->dw_mask & ~(DW_clear_busy | DW_PAGE_WAKEUP)) {
8607 				VM_PAGE_ADD_DELAYED_WORK(dwp, m, dw_count);
8608 
8609 				if (dw_count >= dw_limit) {
8610 					vm_page_do_delayed_work(shadow_object, VM_KERN_MEMORY_NONE, dwp_start, dw_count);
8611 
8612 					dwp = dwp_start;
8613 					dw_count = 0;
8614 				}
8615 			} else {
8616 				if (dwp->dw_mask & DW_clear_busy) {
8617 					m->vmp_busy = FALSE;
8618 				}
8619 
8620 				if (dwp->dw_mask & DW_PAGE_WAKEUP) {
8621 					PAGE_WAKEUP(m);
8622 				}
8623 			}
8624 		}
8625 	}
8626 	if (dw_count) {
8627 		vm_page_do_delayed_work(shadow_object, VM_KERN_MEMORY_NONE, dwp_start, dw_count);
8628 		dwp = dwp_start;
8629 		dw_count = 0;
8630 	}
8631 
8632 	if (upl->flags & UPL_DEVICE_MEMORY) {
8633 		occupied = 0;
8634 	} else if (upl->flags & UPL_LITE) {
8635 		uint32_t pages = (uint32_t)atop(upl_adjusted_size(upl, PAGE_MASK));
8636 
8637 		occupied = !bitmap_is_empty(upl->lite_list, pages);
8638 	} else {
8639 		occupied = !vm_page_queue_empty(&upl->map_object->memq);
8640 	}
8641 	if (occupied == 0) {
8642 		/*
8643 		 * If this UPL element belongs to a Vector UPL and is
8644 		 * empty, then this is the right function to deallocate
8645 		 * it. So go ahead set the *empty variable. The flag
8646 		 * UPL_COMMIT_NOTIFY_EMPTY, from the caller's point of view
8647 		 * should be considered relevant for the Vector UPL and
8648 		 * not the internal UPLs.
8649 		 */
8650 		if ((upl->flags & UPL_COMMIT_NOTIFY_EMPTY) || isVectorUPL) {
8651 			*empty = TRUE;
8652 		}
8653 
8654 		if (object == shadow_object && !(upl->flags & UPL_KERNEL_OBJECT)) {
8655 			/*
8656 			 * this is not a paging object
8657 			 * so we need to drop the paging reference
8658 			 * that was taken when we created the UPL
8659 			 * against this object
8660 			 */
8661 			vm_object_activity_end(shadow_object);
8662 			vm_object_collapse(shadow_object, 0, TRUE);
8663 		} else {
8664 			/*
8665 			 * we dontated the paging reference to
8666 			 * the map object... vm_pageout_object_terminate
8667 			 * will drop this reference
8668 			 */
8669 		}
8670 	}
8671 	vm_object_unlock(shadow_object);
8672 	if (object != shadow_object) {
8673 		vm_object_unlock(object);
8674 	}
8675 
8676 	if (!isVectorUPL) {
8677 		upl_unlock(upl);
8678 	} else {
8679 		/*
8680 		 * If we completed our operations on an UPL that is
8681 		 * part of a Vectored UPL and if empty is TRUE, then
8682 		 * we should go ahead and deallocate this UPL element.
8683 		 * Then we check if this was the last of the UPL elements
8684 		 * within that Vectored UPL. If so, set empty to TRUE
8685 		 * so that in ubc_upl_abort_range or ubc_upl_abort, we
8686 		 * can go ahead and deallocate the Vector UPL too.
8687 		 */
8688 		if (*empty == TRUE) {
8689 			*empty = vector_upl_set_subupl(vector_upl, upl, 0);
8690 			upl_deallocate(upl);
8691 		}
8692 		goto process_upl_to_abort;
8693 	}
8694 
8695 	kr = KERN_SUCCESS;
8696 
8697 done:
8698 	if (dwp_start && dwp_finish_ctx) {
8699 		vm_page_delayed_work_finish_ctx(dwp_start);
8700 		dwp_start = dwp = NULL;
8701 	}
8702 
8703 	return kr;
8704 }
8705 
8706 
8707 kern_return_t
upl_abort(upl_t upl,int error)8708 upl_abort(
8709 	upl_t   upl,
8710 	int     error)
8711 {
8712 	boolean_t       empty;
8713 
8714 	if (upl == UPL_NULL) {
8715 		return KERN_INVALID_ARGUMENT;
8716 	}
8717 
8718 	return upl_abort_range(upl, 0, upl->u_size, error, &empty);
8719 }
8720 
8721 
8722 /* an option on commit should be wire */
8723 kern_return_t
upl_commit(upl_t upl,upl_page_info_t * page_list,mach_msg_type_number_t count)8724 upl_commit(
8725 	upl_t                   upl,
8726 	upl_page_info_t         *page_list,
8727 	mach_msg_type_number_t  count)
8728 {
8729 	boolean_t       empty;
8730 
8731 	if (upl == UPL_NULL) {
8732 		return KERN_INVALID_ARGUMENT;
8733 	}
8734 
8735 	return upl_commit_range(upl, 0, upl->u_size, 0,
8736 	           page_list, count, &empty);
8737 }
8738 
8739 
8740 void
iopl_valid_data(upl_t upl,vm_tag_t tag)8741 iopl_valid_data(
8742 	upl_t    upl,
8743 	vm_tag_t tag)
8744 {
8745 	vm_object_t     object;
8746 	vm_offset_t     offset;
8747 	vm_page_t       m, nxt_page = VM_PAGE_NULL;
8748 	upl_size_t      size;
8749 	int             wired_count = 0;
8750 
8751 	if (upl == NULL) {
8752 		panic("iopl_valid_data: NULL upl");
8753 	}
8754 	if (vector_upl_is_valid(upl)) {
8755 		panic("iopl_valid_data: vector upl");
8756 	}
8757 	if ((upl->flags & (UPL_DEVICE_MEMORY | UPL_SHADOWED | UPL_ACCESS_BLOCKED | UPL_IO_WIRE | UPL_INTERNAL)) != UPL_IO_WIRE) {
8758 		panic("iopl_valid_data: unsupported upl, flags = %x", upl->flags);
8759 	}
8760 
8761 	object = upl->map_object;
8762 
8763 	if (object == kernel_object || object == compressor_object) {
8764 		panic("iopl_valid_data: object == kernel or compressor");
8765 	}
8766 
8767 	if (object->purgable == VM_PURGABLE_VOLATILE ||
8768 	    object->purgable == VM_PURGABLE_EMPTY) {
8769 		panic("iopl_valid_data: object %p purgable %d",
8770 		    object, object->purgable);
8771 	}
8772 
8773 	size = upl_adjusted_size(upl, PAGE_MASK);
8774 
8775 	vm_object_lock(object);
8776 	VM_OBJECT_WIRED_PAGE_UPDATE_START(object);
8777 
8778 	bool whole_object;
8779 
8780 	if (object->vo_size == size && object->resident_page_count == (size / PAGE_SIZE)) {
8781 		nxt_page = (vm_page_t)vm_page_queue_first(&object->memq);
8782 		whole_object = true;
8783 	} else {
8784 		offset = (vm_offset_t)(upl_adjusted_offset(upl, PAGE_MASK) - object->paging_offset);
8785 		whole_object = false;
8786 	}
8787 
8788 	while (size) {
8789 		if (whole_object) {
8790 			if (nxt_page != VM_PAGE_NULL) {
8791 				m = nxt_page;
8792 				nxt_page = (vm_page_t)vm_page_queue_next(&nxt_page->vmp_listq);
8793 			}
8794 		} else {
8795 			m = vm_page_lookup(object, offset);
8796 			offset += PAGE_SIZE;
8797 
8798 			if (m == VM_PAGE_NULL) {
8799 				panic("iopl_valid_data: missing expected page at offset %lx", (long)offset);
8800 			}
8801 		}
8802 		if (m->vmp_busy) {
8803 			if (!m->vmp_absent) {
8804 				panic("iopl_valid_data: busy page w/o absent");
8805 			}
8806 
8807 			if (m->vmp_pageq.next || m->vmp_pageq.prev) {
8808 				panic("iopl_valid_data: busy+absent page on page queue");
8809 			}
8810 			if (m->vmp_reusable) {
8811 				panic("iopl_valid_data: %p is reusable", m);
8812 			}
8813 
8814 			m->vmp_absent = FALSE;
8815 			m->vmp_dirty = TRUE;
8816 			assert(m->vmp_q_state == VM_PAGE_NOT_ON_Q);
8817 			assert(m->vmp_wire_count == 0);
8818 			m->vmp_wire_count++;
8819 			assert(m->vmp_wire_count);
8820 			if (m->vmp_wire_count == 1) {
8821 				m->vmp_q_state = VM_PAGE_IS_WIRED;
8822 				wired_count++;
8823 			} else {
8824 				panic("iopl_valid_data: %p already wired", m);
8825 			}
8826 
8827 			PAGE_WAKEUP_DONE(m);
8828 		}
8829 		size -= PAGE_SIZE;
8830 	}
8831 	if (wired_count) {
8832 		VM_OBJECT_WIRED_PAGE_COUNT(object, wired_count);
8833 		assert(object->resident_page_count >= object->wired_page_count);
8834 
8835 		/* no need to adjust purgeable accounting for this object: */
8836 		assert(object->purgable != VM_PURGABLE_VOLATILE);
8837 		assert(object->purgable != VM_PURGABLE_EMPTY);
8838 
8839 		vm_page_lockspin_queues();
8840 		vm_page_wire_count += wired_count;
8841 		vm_page_unlock_queues();
8842 	}
8843 	VM_OBJECT_WIRED_PAGE_UPDATE_END(object, tag);
8844 	vm_object_unlock(object);
8845 }
8846 
8847 
8848 void
vm_object_set_pmap_cache_attr(vm_object_t object,upl_page_info_array_t user_page_list,unsigned int num_pages,boolean_t batch_pmap_op)8849 vm_object_set_pmap_cache_attr(
8850 	vm_object_t             object,
8851 	upl_page_info_array_t   user_page_list,
8852 	unsigned int            num_pages,
8853 	boolean_t               batch_pmap_op)
8854 {
8855 	unsigned int    cache_attr = 0;
8856 
8857 	cache_attr = object->wimg_bits & VM_WIMG_MASK;
8858 	assert(user_page_list);
8859 	if (cache_attr != VM_WIMG_USE_DEFAULT) {
8860 		PMAP_BATCH_SET_CACHE_ATTR(object, user_page_list, cache_attr, num_pages, batch_pmap_op);
8861 	}
8862 }
8863 
8864 
8865 static bool
vm_object_iopl_wire_full(vm_object_t object,upl_t upl,upl_page_info_array_t user_page_list,upl_control_flags_t cntrl_flags,vm_tag_t tag)8866 vm_object_iopl_wire_full(
8867 	vm_object_t             object,
8868 	upl_t                   upl,
8869 	upl_page_info_array_t   user_page_list,
8870 	upl_control_flags_t     cntrl_flags,
8871 	vm_tag_t                tag)
8872 {
8873 	vm_page_t       dst_page;
8874 	unsigned int    entry;
8875 	int             page_count;
8876 	int             delayed_unlock = 0;
8877 	boolean_t       retval = TRUE;
8878 	ppnum_t         phys_page;
8879 
8880 	vm_object_lock_assert_exclusive(object);
8881 	assert(object->purgable != VM_PURGABLE_VOLATILE);
8882 	assert(object->purgable != VM_PURGABLE_EMPTY);
8883 	assert(object->pager == NULL);
8884 	assert(object->copy == NULL);
8885 	assert(object->shadow == NULL);
8886 
8887 	page_count = object->resident_page_count;
8888 	dst_page = (vm_page_t)vm_page_queue_first(&object->memq);
8889 
8890 	vm_page_lock_queues();
8891 
8892 	while (page_count--) {
8893 		if (dst_page->vmp_busy ||
8894 		    dst_page->vmp_fictitious ||
8895 		    dst_page->vmp_absent ||
8896 		    VMP_ERROR_GET(dst_page) ||
8897 		    dst_page->vmp_cleaning ||
8898 		    dst_page->vmp_restart ||
8899 		    dst_page->vmp_laundry) {
8900 			retval = FALSE;
8901 			goto done;
8902 		}
8903 		if ((cntrl_flags & UPL_REQUEST_FORCE_COHERENCY) && dst_page->vmp_written_by_kernel == TRUE) {
8904 			retval = FALSE;
8905 			goto done;
8906 		}
8907 		dst_page->vmp_reference = TRUE;
8908 
8909 		vm_page_wire(dst_page, tag, FALSE);
8910 
8911 		if (!(cntrl_flags & UPL_COPYOUT_FROM)) {
8912 			SET_PAGE_DIRTY(dst_page, FALSE);
8913 		}
8914 		entry = (unsigned int)(dst_page->vmp_offset / PAGE_SIZE);
8915 		assert(entry >= 0 && entry < object->resident_page_count);
8916 		bitmap_set(upl->lite_list, entry);
8917 
8918 		phys_page = VM_PAGE_GET_PHYS_PAGE(dst_page);
8919 
8920 		if (phys_page > upl->highest_page) {
8921 			upl->highest_page = phys_page;
8922 		}
8923 
8924 		if (user_page_list) {
8925 			user_page_list[entry].phys_addr = phys_page;
8926 			user_page_list[entry].absent    = dst_page->vmp_absent;
8927 			user_page_list[entry].dirty     = dst_page->vmp_dirty;
8928 			user_page_list[entry].free_when_done   = dst_page->vmp_free_when_done;
8929 			user_page_list[entry].precious  = dst_page->vmp_precious;
8930 			user_page_list[entry].device    = FALSE;
8931 			user_page_list[entry].speculative = FALSE;
8932 			user_page_list[entry].cs_validated = FALSE;
8933 			user_page_list[entry].cs_tainted = FALSE;
8934 			user_page_list[entry].cs_nx     = FALSE;
8935 			user_page_list[entry].needed    = FALSE;
8936 			user_page_list[entry].mark      = FALSE;
8937 		}
8938 		if (delayed_unlock++ > 256) {
8939 			delayed_unlock = 0;
8940 			lck_mtx_yield(&vm_page_queue_lock);
8941 
8942 			VM_CHECK_MEMORYSTATUS;
8943 		}
8944 		dst_page = (vm_page_t)vm_page_queue_next(&dst_page->vmp_listq);
8945 	}
8946 done:
8947 	vm_page_unlock_queues();
8948 
8949 	VM_CHECK_MEMORYSTATUS;
8950 
8951 	return retval;
8952 }
8953 
8954 
8955 static kern_return_t
vm_object_iopl_wire_empty(vm_object_t object,upl_t upl,upl_page_info_array_t user_page_list,upl_control_flags_t cntrl_flags,vm_tag_t tag,vm_object_offset_t * dst_offset,int page_count,int * page_grab_count)8956 vm_object_iopl_wire_empty(
8957 	vm_object_t             object,
8958 	upl_t                   upl,
8959 	upl_page_info_array_t   user_page_list,
8960 	upl_control_flags_t     cntrl_flags,
8961 	vm_tag_t                tag,
8962 	vm_object_offset_t     *dst_offset,
8963 	int                     page_count,
8964 	int                    *page_grab_count)
8965 {
8966 	vm_page_t       dst_page;
8967 	boolean_t       no_zero_fill = FALSE;
8968 	int             interruptible;
8969 	int             pages_wired = 0;
8970 	int             pages_inserted = 0;
8971 	int             entry = 0;
8972 	uint64_t        delayed_ledger_update = 0;
8973 	kern_return_t   ret = KERN_SUCCESS;
8974 	int             grab_options;
8975 	ppnum_t         phys_page;
8976 
8977 	vm_object_lock_assert_exclusive(object);
8978 	assert(object->purgable != VM_PURGABLE_VOLATILE);
8979 	assert(object->purgable != VM_PURGABLE_EMPTY);
8980 	assert(object->pager == NULL);
8981 	assert(object->copy == NULL);
8982 	assert(object->shadow == NULL);
8983 
8984 	if (cntrl_flags & UPL_SET_INTERRUPTIBLE) {
8985 		interruptible = THREAD_ABORTSAFE;
8986 	} else {
8987 		interruptible = THREAD_UNINT;
8988 	}
8989 
8990 	if (cntrl_flags & (UPL_NOZEROFILL | UPL_NOZEROFILLIO)) {
8991 		no_zero_fill = TRUE;
8992 	}
8993 
8994 	grab_options = 0;
8995 #if CONFIG_SECLUDED_MEMORY
8996 	if (object->can_grab_secluded) {
8997 		grab_options |= VM_PAGE_GRAB_SECLUDED;
8998 	}
8999 #endif /* CONFIG_SECLUDED_MEMORY */
9000 
9001 	while (page_count--) {
9002 		while ((dst_page = vm_page_grab_options(grab_options))
9003 		    == VM_PAGE_NULL) {
9004 			OSAddAtomic(page_count, &vm_upl_wait_for_pages);
9005 
9006 			VM_DEBUG_EVENT(vm_iopl_page_wait, VM_IOPL_PAGE_WAIT, DBG_FUNC_START, vm_upl_wait_for_pages, 0, 0, 0);
9007 
9008 			if (vm_page_wait(interruptible) == FALSE) {
9009 				/*
9010 				 * interrupted case
9011 				 */
9012 				OSAddAtomic(-page_count, &vm_upl_wait_for_pages);
9013 
9014 				VM_DEBUG_EVENT(vm_iopl_page_wait, VM_IOPL_PAGE_WAIT, DBG_FUNC_END, vm_upl_wait_for_pages, 0, 0, -1);
9015 
9016 				ret = MACH_SEND_INTERRUPTED;
9017 				goto done;
9018 			}
9019 			OSAddAtomic(-page_count, &vm_upl_wait_for_pages);
9020 
9021 			VM_DEBUG_EVENT(vm_iopl_page_wait, VM_IOPL_PAGE_WAIT, DBG_FUNC_END, vm_upl_wait_for_pages, 0, 0, 0);
9022 		}
9023 		if (no_zero_fill == FALSE) {
9024 			vm_page_zero_fill(dst_page);
9025 		} else {
9026 			dst_page->vmp_absent = TRUE;
9027 		}
9028 
9029 		dst_page->vmp_reference = TRUE;
9030 
9031 		if (!(cntrl_flags & UPL_COPYOUT_FROM)) {
9032 			SET_PAGE_DIRTY(dst_page, FALSE);
9033 		}
9034 		if (dst_page->vmp_absent == FALSE) {
9035 			assert(dst_page->vmp_q_state == VM_PAGE_NOT_ON_Q);
9036 			assert(dst_page->vmp_wire_count == 0);
9037 			dst_page->vmp_wire_count++;
9038 			dst_page->vmp_q_state = VM_PAGE_IS_WIRED;
9039 			assert(dst_page->vmp_wire_count);
9040 			pages_wired++;
9041 			PAGE_WAKEUP_DONE(dst_page);
9042 		}
9043 		pages_inserted++;
9044 
9045 		vm_page_insert_internal(dst_page, object, *dst_offset, tag, FALSE, TRUE, TRUE, TRUE, &delayed_ledger_update);
9046 
9047 		bitmap_set(upl->lite_list, entry);
9048 
9049 		phys_page = VM_PAGE_GET_PHYS_PAGE(dst_page);
9050 
9051 		if (phys_page > upl->highest_page) {
9052 			upl->highest_page = phys_page;
9053 		}
9054 
9055 		if (user_page_list) {
9056 			user_page_list[entry].phys_addr = phys_page;
9057 			user_page_list[entry].absent    = dst_page->vmp_absent;
9058 			user_page_list[entry].dirty     = dst_page->vmp_dirty;
9059 			user_page_list[entry].free_when_done    = FALSE;
9060 			user_page_list[entry].precious  = FALSE;
9061 			user_page_list[entry].device    = FALSE;
9062 			user_page_list[entry].speculative = FALSE;
9063 			user_page_list[entry].cs_validated = FALSE;
9064 			user_page_list[entry].cs_tainted = FALSE;
9065 			user_page_list[entry].cs_nx     = FALSE;
9066 			user_page_list[entry].needed    = FALSE;
9067 			user_page_list[entry].mark      = FALSE;
9068 		}
9069 		entry++;
9070 		*dst_offset += PAGE_SIZE_64;
9071 	}
9072 done:
9073 	if (pages_wired) {
9074 		vm_page_lockspin_queues();
9075 		vm_page_wire_count += pages_wired;
9076 		vm_page_unlock_queues();
9077 	}
9078 	if (pages_inserted) {
9079 		if (object->internal) {
9080 			OSAddAtomic(pages_inserted, &vm_page_internal_count);
9081 		} else {
9082 			OSAddAtomic(pages_inserted, &vm_page_external_count);
9083 		}
9084 	}
9085 	if (delayed_ledger_update) {
9086 		task_t          owner;
9087 		int             ledger_idx_volatile;
9088 		int             ledger_idx_nonvolatile;
9089 		int             ledger_idx_volatile_compressed;
9090 		int             ledger_idx_nonvolatile_compressed;
9091 		boolean_t       do_footprint;
9092 
9093 		owner = VM_OBJECT_OWNER(object);
9094 		assert(owner);
9095 
9096 		vm_object_ledger_tag_ledgers(object,
9097 		    &ledger_idx_volatile,
9098 		    &ledger_idx_nonvolatile,
9099 		    &ledger_idx_volatile_compressed,
9100 		    &ledger_idx_nonvolatile_compressed,
9101 		    &do_footprint);
9102 
9103 		/* more non-volatile bytes */
9104 		ledger_credit(owner->ledger,
9105 		    ledger_idx_nonvolatile,
9106 		    delayed_ledger_update);
9107 		if (do_footprint) {
9108 			/* more footprint */
9109 			ledger_credit(owner->ledger,
9110 			    task_ledgers.phys_footprint,
9111 			    delayed_ledger_update);
9112 		}
9113 	}
9114 
9115 	assert(page_grab_count);
9116 	*page_grab_count = pages_inserted;
9117 
9118 	return ret;
9119 }
9120 
9121 
9122 
9123 kern_return_t
vm_object_iopl_request(vm_object_t object,vm_object_offset_t offset,upl_size_t size,upl_t * upl_ptr,upl_page_info_array_t user_page_list,unsigned int * page_list_count,upl_control_flags_t cntrl_flags,vm_tag_t tag)9124 vm_object_iopl_request(
9125 	vm_object_t             object,
9126 	vm_object_offset_t      offset,
9127 	upl_size_t              size,
9128 	upl_t                   *upl_ptr,
9129 	upl_page_info_array_t   user_page_list,
9130 	unsigned int            *page_list_count,
9131 	upl_control_flags_t     cntrl_flags,
9132 	vm_tag_t                tag)
9133 {
9134 	vm_page_t               dst_page;
9135 	vm_object_offset_t      dst_offset;
9136 	upl_size_t              xfer_size;
9137 	upl_t                   upl = NULL;
9138 	unsigned int            entry;
9139 	int                     no_zero_fill = FALSE;
9140 	unsigned int            size_in_pages;
9141 	int                     page_grab_count = 0;
9142 	u_int32_t               psize;
9143 	kern_return_t           ret;
9144 	vm_prot_t               prot;
9145 	struct vm_object_fault_info fault_info = {};
9146 	struct  vm_page_delayed_work    dw_array;
9147 	struct  vm_page_delayed_work    *dwp, *dwp_start;
9148 	bool                    dwp_finish_ctx = TRUE;
9149 	int                     dw_count;
9150 	int                     dw_limit;
9151 	int                     dw_index;
9152 	boolean_t               caller_lookup;
9153 	int                     io_tracking_flag = 0;
9154 	int                     interruptible;
9155 	ppnum_t                 phys_page;
9156 
9157 	boolean_t               set_cache_attr_needed = FALSE;
9158 	boolean_t               free_wired_pages = FALSE;
9159 	boolean_t               fast_path_empty_req = FALSE;
9160 	boolean_t               fast_path_full_req = FALSE;
9161 
9162 #if DEVELOPMENT || DEBUG
9163 	task_t                  task = current_task();
9164 #endif /* DEVELOPMENT || DEBUG */
9165 
9166 	dwp_start = dwp = NULL;
9167 
9168 	vm_object_offset_t original_offset = offset;
9169 	upl_size_t original_size = size;
9170 
9171 //	DEBUG4K_UPL("object %p offset 0x%llx size 0x%llx cntrl_flags 0x%llx\n", object, (uint64_t)offset, (uint64_t)size, cntrl_flags);
9172 
9173 	size = (upl_size_t)(vm_object_round_page(offset + size) - vm_object_trunc_page(offset));
9174 	offset = vm_object_trunc_page(offset);
9175 	if (size != original_size || offset != original_offset) {
9176 		DEBUG4K_IOKIT("flags 0x%llx object %p offset 0x%llx size 0x%x -> offset 0x%llx size 0x%x\n", cntrl_flags, object, original_offset, original_size, offset, size);
9177 	}
9178 
9179 	if (cntrl_flags & ~UPL_VALID_FLAGS) {
9180 		/*
9181 		 * For forward compatibility's sake,
9182 		 * reject any unknown flag.
9183 		 */
9184 		return KERN_INVALID_VALUE;
9185 	}
9186 	if (vm_lopage_needed == FALSE) {
9187 		cntrl_flags &= ~UPL_NEED_32BIT_ADDR;
9188 	}
9189 
9190 	if (cntrl_flags & UPL_NEED_32BIT_ADDR) {
9191 		if ((cntrl_flags & (UPL_SET_IO_WIRE | UPL_SET_LITE)) != (UPL_SET_IO_WIRE | UPL_SET_LITE)) {
9192 			return KERN_INVALID_VALUE;
9193 		}
9194 
9195 		if (object->phys_contiguous) {
9196 			if ((offset + object->vo_shadow_offset) >= (vm_object_offset_t)max_valid_dma_address) {
9197 				return KERN_INVALID_ADDRESS;
9198 			}
9199 
9200 			if (((offset + object->vo_shadow_offset) + size) >= (vm_object_offset_t)max_valid_dma_address) {
9201 				return KERN_INVALID_ADDRESS;
9202 			}
9203 		}
9204 	}
9205 	if (cntrl_flags & (UPL_NOZEROFILL | UPL_NOZEROFILLIO)) {
9206 		no_zero_fill = TRUE;
9207 	}
9208 
9209 	if (cntrl_flags & UPL_COPYOUT_FROM) {
9210 		prot = VM_PROT_READ;
9211 	} else {
9212 		prot = VM_PROT_READ | VM_PROT_WRITE;
9213 	}
9214 
9215 	if ((!object->internal) && (object->paging_offset != 0)) {
9216 		panic("vm_object_iopl_request: external object with non-zero paging offset");
9217 	}
9218 
9219 
9220 	VM_DEBUG_CONSTANT_EVENT(vm_object_iopl_request, VM_IOPL_REQUEST, DBG_FUNC_START, size, cntrl_flags, prot, 0);
9221 
9222 #if CONFIG_IOSCHED || UPL_DEBUG
9223 	if ((object->io_tracking && object != kernel_object) || upl_debug_enabled) {
9224 		io_tracking_flag |= UPL_CREATE_IO_TRACKING;
9225 	}
9226 #endif
9227 
9228 #if CONFIG_IOSCHED
9229 	if (object->io_tracking) {
9230 		/* Check if we're dealing with the kernel object. We do not support expedite on kernel object UPLs */
9231 		if (object != kernel_object) {
9232 			io_tracking_flag |= UPL_CREATE_EXPEDITE_SUP;
9233 		}
9234 	}
9235 #endif
9236 
9237 	if (object->phys_contiguous) {
9238 		psize = PAGE_SIZE;
9239 	} else {
9240 		psize = size;
9241 
9242 		dw_count = 0;
9243 		dw_limit = DELAYED_WORK_LIMIT(DEFAULT_DELAYED_WORK_LIMIT);
9244 		dwp_start = vm_page_delayed_work_get_ctx();
9245 		if (dwp_start == NULL) {
9246 			dwp_start = &dw_array;
9247 			dw_limit = 1;
9248 			dwp_finish_ctx = FALSE;
9249 		}
9250 
9251 		dwp = dwp_start;
9252 	}
9253 
9254 	if (cntrl_flags & UPL_SET_INTERNAL) {
9255 		upl = upl_create(UPL_CREATE_INTERNAL | UPL_CREATE_LITE | io_tracking_flag, UPL_IO_WIRE, psize);
9256 		user_page_list = size ? upl->page_list : NULL;
9257 	} else {
9258 		upl = upl_create(UPL_CREATE_LITE | io_tracking_flag, UPL_IO_WIRE, psize);
9259 	}
9260 	if (user_page_list) {
9261 		user_page_list[0].device = FALSE;
9262 	}
9263 	*upl_ptr = upl;
9264 
9265 	if (cntrl_flags & UPL_NOZEROFILLIO) {
9266 		DTRACE_VM4(upl_nozerofillio,
9267 		    vm_object_t, object,
9268 		    vm_object_offset_t, offset,
9269 		    upl_size_t, size,
9270 		    upl_t, upl);
9271 	}
9272 
9273 	upl->map_object = object;
9274 	upl->u_offset = original_offset;
9275 	upl->u_size = original_size;
9276 
9277 	size_in_pages = size / PAGE_SIZE;
9278 
9279 	if (object == kernel_object &&
9280 	    !(cntrl_flags & (UPL_NEED_32BIT_ADDR | UPL_BLOCK_ACCESS))) {
9281 		upl->flags |= UPL_KERNEL_OBJECT;
9282 #if UPL_DEBUG
9283 		vm_object_lock(object);
9284 #else
9285 		vm_object_lock_shared(object);
9286 #endif
9287 	} else {
9288 		vm_object_lock(object);
9289 		vm_object_activity_begin(object);
9290 	}
9291 	/*
9292 	 * paging in progress also protects the paging_offset
9293 	 */
9294 	upl->u_offset = original_offset + object->paging_offset;
9295 
9296 	if (cntrl_flags & UPL_BLOCK_ACCESS) {
9297 		/*
9298 		 * The user requested that access to the pages in this UPL
9299 		 * be blocked until the UPL is commited or aborted.
9300 		 */
9301 		upl->flags |= UPL_ACCESS_BLOCKED;
9302 	}
9303 
9304 #if CONFIG_IOSCHED || UPL_DEBUG
9305 	if ((upl->flags & UPL_TRACKED_BY_OBJECT) || upl_debug_enabled) {
9306 		vm_object_activity_begin(object);
9307 		queue_enter(&object->uplq, upl, upl_t, uplq);
9308 	}
9309 #endif
9310 
9311 	if (object->phys_contiguous) {
9312 		if (upl->flags & UPL_ACCESS_BLOCKED) {
9313 			assert(!object->blocked_access);
9314 			object->blocked_access = TRUE;
9315 		}
9316 
9317 		vm_object_unlock(object);
9318 
9319 		/*
9320 		 * don't need any shadow mappings for this one
9321 		 * since it is already I/O memory
9322 		 */
9323 		upl->flags |= UPL_DEVICE_MEMORY;
9324 
9325 		upl->highest_page = (ppnum_t) ((offset + object->vo_shadow_offset + size - 1) >> PAGE_SHIFT);
9326 
9327 		if (user_page_list) {
9328 			user_page_list[0].phys_addr = (ppnum_t) ((offset + object->vo_shadow_offset) >> PAGE_SHIFT);
9329 			user_page_list[0].device = TRUE;
9330 		}
9331 		if (page_list_count != NULL) {
9332 			if (upl->flags & UPL_INTERNAL) {
9333 				*page_list_count = 0;
9334 			} else {
9335 				*page_list_count = 1;
9336 			}
9337 		}
9338 
9339 		VM_DEBUG_CONSTANT_EVENT(vm_object_iopl_request, VM_IOPL_REQUEST, DBG_FUNC_END, page_grab_count, KERN_SUCCESS, 0, 0);
9340 #if DEVELOPMENT || DEBUG
9341 		if (task != NULL) {
9342 			ledger_credit(task->ledger, task_ledgers.pages_grabbed_iopl, page_grab_count);
9343 		}
9344 #endif /* DEVELOPMENT || DEBUG */
9345 		return KERN_SUCCESS;
9346 	}
9347 	if (object != kernel_object && object != compressor_object) {
9348 		/*
9349 		 * Protect user space from future COW operations
9350 		 */
9351 #if VM_OBJECT_TRACKING_OP_TRUESHARE
9352 		if (!object->true_share &&
9353 		    vm_object_tracking_btlog) {
9354 			btlog_record(vm_object_tracking_btlog, object,
9355 			    VM_OBJECT_TRACKING_OP_TRUESHARE,
9356 			    btref_get(__builtin_frame_address(0), 0));
9357 		}
9358 #endif /* VM_OBJECT_TRACKING_OP_TRUESHARE */
9359 
9360 		vm_object_lock_assert_exclusive(object);
9361 		object->true_share = TRUE;
9362 
9363 		if (object->copy_strategy == MEMORY_OBJECT_COPY_SYMMETRIC) {
9364 			object->copy_strategy = MEMORY_OBJECT_COPY_DELAY;
9365 		}
9366 	}
9367 
9368 	if (!(cntrl_flags & UPL_COPYOUT_FROM) &&
9369 	    object->copy != VM_OBJECT_NULL) {
9370 		/*
9371 		 * Honor copy-on-write obligations
9372 		 *
9373 		 * The caller is gathering these pages and
9374 		 * might modify their contents.  We need to
9375 		 * make sure that the copy object has its own
9376 		 * private copies of these pages before we let
9377 		 * the caller modify them.
9378 		 *
9379 		 * NOTE: someone else could map the original object
9380 		 * after we've done this copy-on-write here, and they
9381 		 * could then see an inconsistent picture of the memory
9382 		 * while it's being modified via the UPL.  To prevent this,
9383 		 * we would have to block access to these pages until the
9384 		 * UPL is released.  We could use the UPL_BLOCK_ACCESS
9385 		 * code path for that...
9386 		 */
9387 		vm_object_update(object,
9388 		    offset,
9389 		    size,
9390 		    NULL,
9391 		    NULL,
9392 		    FALSE,              /* should_return */
9393 		    MEMORY_OBJECT_COPY_SYNC,
9394 		    VM_PROT_NO_CHANGE);
9395 		VM_PAGEOUT_DEBUG(iopl_cow, 1);
9396 		VM_PAGEOUT_DEBUG(iopl_cow_pages, (size >> PAGE_SHIFT));
9397 	}
9398 	if (!(cntrl_flags & (UPL_NEED_32BIT_ADDR | UPL_BLOCK_ACCESS)) &&
9399 	    object->purgable != VM_PURGABLE_VOLATILE &&
9400 	    object->purgable != VM_PURGABLE_EMPTY &&
9401 	    object->copy == NULL &&
9402 	    size == object->vo_size &&
9403 	    offset == 0 &&
9404 	    object->shadow == NULL &&
9405 	    object->pager == NULL) {
9406 		if (object->resident_page_count == size_in_pages) {
9407 			assert(object != compressor_object);
9408 			assert(object != kernel_object);
9409 			fast_path_full_req = TRUE;
9410 		} else if (object->resident_page_count == 0) {
9411 			assert(object != compressor_object);
9412 			assert(object != kernel_object);
9413 			fast_path_empty_req = TRUE;
9414 			set_cache_attr_needed = TRUE;
9415 		}
9416 	}
9417 
9418 	if (cntrl_flags & UPL_SET_INTERRUPTIBLE) {
9419 		interruptible = THREAD_ABORTSAFE;
9420 	} else {
9421 		interruptible = THREAD_UNINT;
9422 	}
9423 
9424 	entry = 0;
9425 
9426 	xfer_size = size;
9427 	dst_offset = offset;
9428 
9429 	if (fast_path_full_req) {
9430 		if (vm_object_iopl_wire_full(object, upl, user_page_list, cntrl_flags, tag) == TRUE) {
9431 			goto finish;
9432 		}
9433 		/*
9434 		 * we couldn't complete the processing of this request on the fast path
9435 		 * so fall through to the slow path and finish up
9436 		 */
9437 	} else if (fast_path_empty_req) {
9438 		if (cntrl_flags & UPL_REQUEST_NO_FAULT) {
9439 			ret = KERN_MEMORY_ERROR;
9440 			goto return_err;
9441 		}
9442 		ret = vm_object_iopl_wire_empty(object, upl, user_page_list,
9443 		    cntrl_flags, tag, &dst_offset, size_in_pages, &page_grab_count);
9444 
9445 		if (ret) {
9446 			free_wired_pages = TRUE;
9447 			goto return_err;
9448 		}
9449 		goto finish;
9450 	}
9451 
9452 	fault_info.behavior = VM_BEHAVIOR_SEQUENTIAL;
9453 	fault_info.lo_offset = offset;
9454 	fault_info.hi_offset = offset + xfer_size;
9455 	fault_info.mark_zf_absent = TRUE;
9456 	fault_info.interruptible = interruptible;
9457 	fault_info.batch_pmap_op = TRUE;
9458 
9459 	while (xfer_size) {
9460 		vm_fault_return_t       result;
9461 
9462 		dwp->dw_mask = 0;
9463 
9464 		if (fast_path_full_req) {
9465 			/*
9466 			 * if we get here, it means that we ran into a page
9467 			 * state we couldn't handle in the fast path and
9468 			 * bailed out to the slow path... since the order
9469 			 * we look at pages is different between the 2 paths,
9470 			 * the following check is needed to determine whether
9471 			 * this page was already processed in the fast path
9472 			 */
9473 			if (bitmap_test(upl->lite_list, entry)) {
9474 				goto skip_page;
9475 			}
9476 		}
9477 		dst_page = vm_page_lookup(object, dst_offset);
9478 
9479 		if (dst_page == VM_PAGE_NULL ||
9480 		    dst_page->vmp_busy ||
9481 		    VMP_ERROR_GET(dst_page) ||
9482 		    dst_page->vmp_restart ||
9483 		    dst_page->vmp_absent ||
9484 		    dst_page->vmp_fictitious) {
9485 			if (object == kernel_object) {
9486 				panic("vm_object_iopl_request: missing/bad page in kernel object");
9487 			}
9488 			if (object == compressor_object) {
9489 				panic("vm_object_iopl_request: missing/bad page in compressor object");
9490 			}
9491 
9492 			if (cntrl_flags & UPL_REQUEST_NO_FAULT) {
9493 				ret = KERN_MEMORY_ERROR;
9494 				goto return_err;
9495 			}
9496 			set_cache_attr_needed = TRUE;
9497 
9498 			/*
9499 			 * We just looked up the page and the result remains valid
9500 			 * until the object lock is release, so send it to
9501 			 * vm_fault_page() (as "dst_page"), to avoid having to
9502 			 * look it up again there.
9503 			 */
9504 			caller_lookup = TRUE;
9505 
9506 			do {
9507 				vm_page_t       top_page;
9508 				kern_return_t   error_code;
9509 
9510 				fault_info.cluster_size = xfer_size;
9511 
9512 				vm_object_paging_begin(object);
9513 
9514 				result = vm_fault_page(object, dst_offset,
9515 				    prot | VM_PROT_WRITE, FALSE,
9516 				    caller_lookup,
9517 				    &prot, &dst_page, &top_page,
9518 				    (int *)0,
9519 				    &error_code, no_zero_fill,
9520 				    &fault_info);
9521 
9522 				/* our lookup is no longer valid at this point */
9523 				caller_lookup = FALSE;
9524 
9525 				switch (result) {
9526 				case VM_FAULT_SUCCESS:
9527 					page_grab_count++;
9528 
9529 					if (!dst_page->vmp_absent) {
9530 						PAGE_WAKEUP_DONE(dst_page);
9531 					} else {
9532 						/*
9533 						 * we only get back an absent page if we
9534 						 * requested that it not be zero-filled
9535 						 * because we are about to fill it via I/O
9536 						 *
9537 						 * absent pages should be left BUSY
9538 						 * to prevent them from being faulted
9539 						 * into an address space before we've
9540 						 * had a chance to complete the I/O on
9541 						 * them since they may contain info that
9542 						 * shouldn't be seen by the faulting task
9543 						 */
9544 					}
9545 					/*
9546 					 *	Release paging references and
9547 					 *	top-level placeholder page, if any.
9548 					 */
9549 					if (top_page != VM_PAGE_NULL) {
9550 						vm_object_t local_object;
9551 
9552 						local_object = VM_PAGE_OBJECT(top_page);
9553 
9554 						/*
9555 						 * comparing 2 packed pointers
9556 						 */
9557 						if (top_page->vmp_object != dst_page->vmp_object) {
9558 							vm_object_lock(local_object);
9559 							VM_PAGE_FREE(top_page);
9560 							vm_object_paging_end(local_object);
9561 							vm_object_unlock(local_object);
9562 						} else {
9563 							VM_PAGE_FREE(top_page);
9564 							vm_object_paging_end(local_object);
9565 						}
9566 					}
9567 					vm_object_paging_end(object);
9568 					break;
9569 
9570 				case VM_FAULT_RETRY:
9571 					vm_object_lock(object);
9572 					break;
9573 
9574 				case VM_FAULT_MEMORY_SHORTAGE:
9575 					OSAddAtomic((size_in_pages - entry), &vm_upl_wait_for_pages);
9576 
9577 					VM_DEBUG_EVENT(vm_iopl_page_wait, VM_IOPL_PAGE_WAIT, DBG_FUNC_START, vm_upl_wait_for_pages, 0, 0, 0);
9578 
9579 					if (vm_page_wait(interruptible)) {
9580 						OSAddAtomic(-(size_in_pages - entry), &vm_upl_wait_for_pages);
9581 
9582 						VM_DEBUG_EVENT(vm_iopl_page_wait, VM_IOPL_PAGE_WAIT, DBG_FUNC_END, vm_upl_wait_for_pages, 0, 0, 0);
9583 						vm_object_lock(object);
9584 
9585 						break;
9586 					}
9587 					OSAddAtomic(-(size_in_pages - entry), &vm_upl_wait_for_pages);
9588 
9589 					VM_DEBUG_EVENT(vm_iopl_page_wait, VM_IOPL_PAGE_WAIT, DBG_FUNC_END, vm_upl_wait_for_pages, 0, 0, -1);
9590 					ktriage_record(thread_tid(current_thread()), KDBG_TRIAGE_EVENTID(KDBG_TRIAGE_SUBSYS_VM, KDBG_TRIAGE_RESERVED, KDBG_TRIAGE_VM_FAULT_OBJIOPLREQ_MEMORY_SHORTAGE), 0 /* arg */);
9591 					OS_FALLTHROUGH;
9592 
9593 				case VM_FAULT_INTERRUPTED:
9594 					error_code = MACH_SEND_INTERRUPTED;
9595 					OS_FALLTHROUGH;
9596 				case VM_FAULT_MEMORY_ERROR:
9597 memory_error:
9598 					ret = (error_code ? error_code: KERN_MEMORY_ERROR);
9599 
9600 					vm_object_lock(object);
9601 					goto return_err;
9602 
9603 				case VM_FAULT_SUCCESS_NO_VM_PAGE:
9604 					/* success but no page: fail */
9605 					vm_object_paging_end(object);
9606 					vm_object_unlock(object);
9607 					goto memory_error;
9608 
9609 				default:
9610 					panic("vm_object_iopl_request: unexpected error"
9611 					    " 0x%x from vm_fault_page()\n", result);
9612 				}
9613 			} while (result != VM_FAULT_SUCCESS);
9614 		}
9615 		phys_page = VM_PAGE_GET_PHYS_PAGE(dst_page);
9616 
9617 		if (upl->flags & UPL_KERNEL_OBJECT) {
9618 			goto record_phys_addr;
9619 		}
9620 
9621 		if (dst_page->vmp_q_state == VM_PAGE_USED_BY_COMPRESSOR) {
9622 			dst_page->vmp_busy = TRUE;
9623 			goto record_phys_addr;
9624 		}
9625 
9626 		if (dst_page->vmp_cleaning) {
9627 			/*
9628 			 * Someone else is cleaning this page in place.
9629 			 * In theory, we should be able to  proceed and use this
9630 			 * page but they'll probably end up clearing the "busy"
9631 			 * bit on it in upl_commit_range() but they didn't set
9632 			 * it, so they would clear our "busy" bit and open
9633 			 * us to race conditions.
9634 			 * We'd better wait for the cleaning to complete and
9635 			 * then try again.
9636 			 */
9637 			VM_PAGEOUT_DEBUG(vm_object_iopl_request_sleep_for_cleaning, 1);
9638 			PAGE_SLEEP(object, dst_page, THREAD_UNINT);
9639 			continue;
9640 		}
9641 		if (dst_page->vmp_laundry) {
9642 			vm_pageout_steal_laundry(dst_page, FALSE);
9643 		}
9644 
9645 		if ((cntrl_flags & UPL_NEED_32BIT_ADDR) &&
9646 		    phys_page >= (max_valid_dma_address >> PAGE_SHIFT)) {
9647 			vm_page_t       low_page;
9648 			int             refmod;
9649 
9650 			/*
9651 			 * support devices that can't DMA above 32 bits
9652 			 * by substituting pages from a pool of low address
9653 			 * memory for any pages we find above the 4G mark
9654 			 * can't substitute if the page is already wired because
9655 			 * we don't know whether that physical address has been
9656 			 * handed out to some other 64 bit capable DMA device to use
9657 			 */
9658 			if (VM_PAGE_WIRED(dst_page)) {
9659 				ret = KERN_PROTECTION_FAILURE;
9660 				goto return_err;
9661 			}
9662 			low_page = vm_page_grablo();
9663 
9664 			if (low_page == VM_PAGE_NULL) {
9665 				ret = KERN_RESOURCE_SHORTAGE;
9666 				goto return_err;
9667 			}
9668 			/*
9669 			 * from here until the vm_page_replace completes
9670 			 * we musn't drop the object lock... we don't
9671 			 * want anyone refaulting this page in and using
9672 			 * it after we disconnect it... we want the fault
9673 			 * to find the new page being substituted.
9674 			 */
9675 			if (dst_page->vmp_pmapped) {
9676 				refmod = pmap_disconnect(phys_page);
9677 			} else {
9678 				refmod = 0;
9679 			}
9680 
9681 			if (!dst_page->vmp_absent) {
9682 				vm_page_copy(dst_page, low_page);
9683 			}
9684 
9685 			low_page->vmp_reference = dst_page->vmp_reference;
9686 			low_page->vmp_dirty     = dst_page->vmp_dirty;
9687 			low_page->vmp_absent    = dst_page->vmp_absent;
9688 
9689 			if (refmod & VM_MEM_REFERENCED) {
9690 				low_page->vmp_reference = TRUE;
9691 			}
9692 			if (refmod & VM_MEM_MODIFIED) {
9693 				SET_PAGE_DIRTY(low_page, FALSE);
9694 			}
9695 
9696 			vm_page_replace(low_page, object, dst_offset);
9697 
9698 			dst_page = low_page;
9699 			/*
9700 			 * vm_page_grablo returned the page marked
9701 			 * BUSY... we don't need a PAGE_WAKEUP_DONE
9702 			 * here, because we've never dropped the object lock
9703 			 */
9704 			if (!dst_page->vmp_absent) {
9705 				dst_page->vmp_busy = FALSE;
9706 			}
9707 
9708 			phys_page = VM_PAGE_GET_PHYS_PAGE(dst_page);
9709 		}
9710 		if (!dst_page->vmp_busy) {
9711 			dwp->dw_mask |= DW_vm_page_wire;
9712 		}
9713 
9714 		if (cntrl_flags & UPL_BLOCK_ACCESS) {
9715 			/*
9716 			 * Mark the page "busy" to block any future page fault
9717 			 * on this page in addition to wiring it.
9718 			 * We'll also remove the mapping
9719 			 * of all these pages before leaving this routine.
9720 			 */
9721 			assert(!dst_page->vmp_fictitious);
9722 			dst_page->vmp_busy = TRUE;
9723 		}
9724 		/*
9725 		 * expect the page to be used
9726 		 * page queues lock must be held to set 'reference'
9727 		 */
9728 		dwp->dw_mask |= DW_set_reference;
9729 
9730 		if (!(cntrl_flags & UPL_COPYOUT_FROM)) {
9731 			SET_PAGE_DIRTY(dst_page, TRUE);
9732 			/*
9733 			 * Page belonging to a code-signed object is about to
9734 			 * be written. Mark it tainted and disconnect it from
9735 			 * all pmaps so processes have to fault it back in and
9736 			 * deal with the tainted bit.
9737 			 */
9738 			if (object->code_signed && dst_page->vmp_cs_tainted != VMP_CS_ALL_TRUE) {
9739 				dst_page->vmp_cs_tainted = VMP_CS_ALL_TRUE;
9740 				vm_page_iopl_tainted++;
9741 				if (dst_page->vmp_pmapped) {
9742 					int refmod = pmap_disconnect(VM_PAGE_GET_PHYS_PAGE(dst_page));
9743 					if (refmod & VM_MEM_REFERENCED) {
9744 						dst_page->vmp_reference = TRUE;
9745 					}
9746 				}
9747 			}
9748 		}
9749 		if ((cntrl_flags & UPL_REQUEST_FORCE_COHERENCY) && dst_page->vmp_written_by_kernel == TRUE) {
9750 			pmap_sync_page_attributes_phys(phys_page);
9751 			dst_page->vmp_written_by_kernel = FALSE;
9752 		}
9753 
9754 record_phys_addr:
9755 		if (dst_page->vmp_busy) {
9756 			upl->flags |= UPL_HAS_BUSY;
9757 		}
9758 
9759 		bitmap_set(upl->lite_list, entry);
9760 
9761 		if (phys_page > upl->highest_page) {
9762 			upl->highest_page = phys_page;
9763 		}
9764 
9765 		if (user_page_list) {
9766 			user_page_list[entry].phys_addr = phys_page;
9767 			user_page_list[entry].free_when_done    = dst_page->vmp_free_when_done;
9768 			user_page_list[entry].absent    = dst_page->vmp_absent;
9769 			user_page_list[entry].dirty     = dst_page->vmp_dirty;
9770 			user_page_list[entry].precious  = dst_page->vmp_precious;
9771 			user_page_list[entry].device    = FALSE;
9772 			user_page_list[entry].needed    = FALSE;
9773 			if (dst_page->vmp_clustered == TRUE) {
9774 				user_page_list[entry].speculative = (dst_page->vmp_q_state == VM_PAGE_ON_SPECULATIVE_Q) ? TRUE : FALSE;
9775 			} else {
9776 				user_page_list[entry].speculative = FALSE;
9777 			}
9778 			user_page_list[entry].cs_validated = dst_page->vmp_cs_validated;
9779 			user_page_list[entry].cs_tainted = dst_page->vmp_cs_tainted;
9780 			user_page_list[entry].cs_nx = dst_page->vmp_cs_nx;
9781 			user_page_list[entry].mark      = FALSE;
9782 		}
9783 		if (object != kernel_object && object != compressor_object) {
9784 			/*
9785 			 * someone is explicitly grabbing this page...
9786 			 * update clustered and speculative state
9787 			 *
9788 			 */
9789 			if (dst_page->vmp_clustered) {
9790 				VM_PAGE_CONSUME_CLUSTERED(dst_page);
9791 			}
9792 		}
9793 skip_page:
9794 		entry++;
9795 		dst_offset += PAGE_SIZE_64;
9796 		xfer_size -= PAGE_SIZE;
9797 
9798 		if (dwp->dw_mask) {
9799 			VM_PAGE_ADD_DELAYED_WORK(dwp, dst_page, dw_count);
9800 
9801 			if (dw_count >= dw_limit) {
9802 				vm_page_do_delayed_work(object, tag, dwp_start, dw_count);
9803 
9804 				dwp = dwp_start;
9805 				dw_count = 0;
9806 			}
9807 		}
9808 	}
9809 	assert(entry == size_in_pages);
9810 
9811 	if (dw_count) {
9812 		vm_page_do_delayed_work(object, tag, dwp_start, dw_count);
9813 		dwp = dwp_start;
9814 		dw_count = 0;
9815 	}
9816 finish:
9817 	if (user_page_list && set_cache_attr_needed == TRUE) {
9818 		vm_object_set_pmap_cache_attr(object, user_page_list, size_in_pages, TRUE);
9819 	}
9820 
9821 	if (page_list_count != NULL) {
9822 		if (upl->flags & UPL_INTERNAL) {
9823 			*page_list_count = 0;
9824 		} else if (*page_list_count > size_in_pages) {
9825 			*page_list_count = size_in_pages;
9826 		}
9827 	}
9828 	vm_object_unlock(object);
9829 
9830 	if (cntrl_flags & UPL_BLOCK_ACCESS) {
9831 		/*
9832 		 * We've marked all the pages "busy" so that future
9833 		 * page faults will block.
9834 		 * Now remove the mapping for these pages, so that they
9835 		 * can't be accessed without causing a page fault.
9836 		 */
9837 		vm_object_pmap_protect(object, offset, (vm_object_size_t)size,
9838 		    PMAP_NULL,
9839 		    PAGE_SIZE,
9840 		    0, VM_PROT_NONE);
9841 		assert(!object->blocked_access);
9842 		object->blocked_access = TRUE;
9843 	}
9844 
9845 	VM_DEBUG_CONSTANT_EVENT(vm_object_iopl_request, VM_IOPL_REQUEST, DBG_FUNC_END, page_grab_count, KERN_SUCCESS, 0, 0);
9846 #if DEVELOPMENT || DEBUG
9847 	if (task != NULL) {
9848 		ledger_credit(task->ledger, task_ledgers.pages_grabbed_iopl, page_grab_count);
9849 	}
9850 #endif /* DEVELOPMENT || DEBUG */
9851 
9852 	if (dwp_start && dwp_finish_ctx) {
9853 		vm_page_delayed_work_finish_ctx(dwp_start);
9854 		dwp_start = dwp = NULL;
9855 	}
9856 
9857 	return KERN_SUCCESS;
9858 
9859 return_err:
9860 	dw_index = 0;
9861 
9862 	for (; offset < dst_offset; offset += PAGE_SIZE) {
9863 		boolean_t need_unwire;
9864 
9865 		dst_page = vm_page_lookup(object, offset);
9866 
9867 		if (dst_page == VM_PAGE_NULL) {
9868 			panic("vm_object_iopl_request: Wired page missing.");
9869 		}
9870 
9871 		/*
9872 		 * if we've already processed this page in an earlier
9873 		 * dw_do_work, we need to undo the wiring... we will
9874 		 * leave the dirty and reference bits on if they
9875 		 * were set, since we don't have a good way of knowing
9876 		 * what the previous state was and we won't get here
9877 		 * under any normal circumstances...  we will always
9878 		 * clear BUSY and wakeup any waiters via vm_page_free
9879 		 * or PAGE_WAKEUP_DONE
9880 		 */
9881 		need_unwire = TRUE;
9882 
9883 		if (dw_count) {
9884 			if ((dwp_start)[dw_index].dw_m == dst_page) {
9885 				/*
9886 				 * still in the deferred work list
9887 				 * which means we haven't yet called
9888 				 * vm_page_wire on this page
9889 				 */
9890 				need_unwire = FALSE;
9891 
9892 				dw_index++;
9893 				dw_count--;
9894 			}
9895 		}
9896 		vm_page_lock_queues();
9897 
9898 		if (dst_page->vmp_absent || free_wired_pages == TRUE) {
9899 			vm_page_free(dst_page);
9900 
9901 			need_unwire = FALSE;
9902 		} else {
9903 			if (need_unwire == TRUE) {
9904 				vm_page_unwire(dst_page, TRUE);
9905 			}
9906 
9907 			PAGE_WAKEUP_DONE(dst_page);
9908 		}
9909 		vm_page_unlock_queues();
9910 
9911 		if (need_unwire == TRUE) {
9912 			counter_inc(&vm_statistics_reactivations);
9913 		}
9914 	}
9915 #if UPL_DEBUG
9916 	upl->upl_state = 2;
9917 #endif
9918 	if (!(upl->flags & UPL_KERNEL_OBJECT)) {
9919 		vm_object_activity_end(object);
9920 		vm_object_collapse(object, 0, TRUE);
9921 	}
9922 	vm_object_unlock(object);
9923 	upl_destroy(upl);
9924 
9925 	VM_DEBUG_CONSTANT_EVENT(vm_object_iopl_request, VM_IOPL_REQUEST, DBG_FUNC_END, page_grab_count, ret, 0, 0);
9926 #if DEVELOPMENT || DEBUG
9927 	if (task != NULL) {
9928 		ledger_credit(task->ledger, task_ledgers.pages_grabbed_iopl, page_grab_count);
9929 	}
9930 #endif /* DEVELOPMENT || DEBUG */
9931 
9932 	if (dwp_start && dwp_finish_ctx) {
9933 		vm_page_delayed_work_finish_ctx(dwp_start);
9934 		dwp_start = dwp = NULL;
9935 	}
9936 	return ret;
9937 }
9938 
9939 kern_return_t
upl_transpose(upl_t upl1,upl_t upl2)9940 upl_transpose(
9941 	upl_t           upl1,
9942 	upl_t           upl2)
9943 {
9944 	kern_return_t           retval;
9945 	boolean_t               upls_locked;
9946 	vm_object_t             object1, object2;
9947 
9948 	/* LD: Should mapped UPLs be eligible for a transpose? */
9949 	if (upl1 == UPL_NULL || upl2 == UPL_NULL || upl1 == upl2 || ((upl1->flags & UPL_VECTOR) == UPL_VECTOR) || ((upl2->flags & UPL_VECTOR) == UPL_VECTOR)) {
9950 		return KERN_INVALID_ARGUMENT;
9951 	}
9952 
9953 	upls_locked = FALSE;
9954 
9955 	/*
9956 	 * Since we need to lock both UPLs at the same time,
9957 	 * avoid deadlocks by always taking locks in the same order.
9958 	 */
9959 	if (upl1 < upl2) {
9960 		upl_lock(upl1);
9961 		upl_lock(upl2);
9962 	} else {
9963 		upl_lock(upl2);
9964 		upl_lock(upl1);
9965 	}
9966 	upls_locked = TRUE;     /* the UPLs will need to be unlocked */
9967 
9968 	object1 = upl1->map_object;
9969 	object2 = upl2->map_object;
9970 
9971 	if (upl1->u_offset != 0 || upl2->u_offset != 0 ||
9972 	    upl1->u_size != upl2->u_size) {
9973 		/*
9974 		 * We deal only with full objects, not subsets.
9975 		 * That's because we exchange the entire backing store info
9976 		 * for the objects: pager, resident pages, etc...  We can't do
9977 		 * only part of it.
9978 		 */
9979 		retval = KERN_INVALID_VALUE;
9980 		goto done;
9981 	}
9982 
9983 	/*
9984 	 * Tranpose the VM objects' backing store.
9985 	 */
9986 	retval = vm_object_transpose(object1, object2,
9987 	    upl_adjusted_size(upl1, PAGE_MASK));
9988 
9989 	if (retval == KERN_SUCCESS) {
9990 		/*
9991 		 * Make each UPL point to the correct VM object, i.e. the
9992 		 * object holding the pages that the UPL refers to...
9993 		 */
9994 #if CONFIG_IOSCHED || UPL_DEBUG
9995 		if ((upl1->flags & UPL_TRACKED_BY_OBJECT) || (upl2->flags & UPL_TRACKED_BY_OBJECT)) {
9996 			vm_object_lock(object1);
9997 			vm_object_lock(object2);
9998 		}
9999 		if ((upl1->flags & UPL_TRACKED_BY_OBJECT) || upl_debug_enabled) {
10000 			queue_remove(&object1->uplq, upl1, upl_t, uplq);
10001 		}
10002 		if ((upl2->flags & UPL_TRACKED_BY_OBJECT) || upl_debug_enabled) {
10003 			queue_remove(&object2->uplq, upl2, upl_t, uplq);
10004 		}
10005 #endif
10006 		upl1->map_object = object2;
10007 		upl2->map_object = object1;
10008 
10009 #if CONFIG_IOSCHED || UPL_DEBUG
10010 		if ((upl1->flags & UPL_TRACKED_BY_OBJECT) || upl_debug_enabled) {
10011 			queue_enter(&object2->uplq, upl1, upl_t, uplq);
10012 		}
10013 		if ((upl2->flags & UPL_TRACKED_BY_OBJECT) || upl_debug_enabled) {
10014 			queue_enter(&object1->uplq, upl2, upl_t, uplq);
10015 		}
10016 		if ((upl1->flags & UPL_TRACKED_BY_OBJECT) || (upl2->flags & UPL_TRACKED_BY_OBJECT)) {
10017 			vm_object_unlock(object2);
10018 			vm_object_unlock(object1);
10019 		}
10020 #endif
10021 	}
10022 
10023 done:
10024 	/*
10025 	 * Cleanup.
10026 	 */
10027 	if (upls_locked) {
10028 		upl_unlock(upl1);
10029 		upl_unlock(upl2);
10030 		upls_locked = FALSE;
10031 	}
10032 
10033 	return retval;
10034 }
10035 
10036 void
upl_range_needed(upl_t upl,int index,int count)10037 upl_range_needed(
10038 	upl_t           upl,
10039 	int             index,
10040 	int             count)
10041 {
10042 	int             size_in_pages;
10043 
10044 	if (!(upl->flags & UPL_INTERNAL) || count <= 0) {
10045 		return;
10046 	}
10047 
10048 	size_in_pages = upl_adjusted_size(upl, PAGE_MASK) / PAGE_SIZE;
10049 
10050 	while (count-- && index < size_in_pages) {
10051 		upl->page_list[index++].needed = TRUE;
10052 	}
10053 }
10054 
10055 
10056 /*
10057  * Reserve of virtual addresses in the kernel address space.
10058  * We need to map the physical pages in the kernel, so that we
10059  * can call the code-signing or slide routines with a kernel
10060  * virtual address.  We keep this pool of pre-allocated kernel
10061  * virtual addresses so that we don't have to scan the kernel's
10062  * virtaul address space each time we need to work with
10063  * a physical page.
10064  */
10065 SIMPLE_LOCK_DECLARE(vm_paging_lock, 0);
10066 #define VM_PAGING_NUM_PAGES     64
10067 SECURITY_READ_ONLY_LATE(vm_offset_t) vm_paging_base_address = 0;
10068 bool            vm_paging_page_inuse[VM_PAGING_NUM_PAGES] = { FALSE, };
10069 int             vm_paging_max_index = 0;
10070 int             vm_paging_page_waiter = 0;
10071 int             vm_paging_page_waiter_total = 0;
10072 
10073 unsigned long   vm_paging_no_kernel_page = 0;
10074 unsigned long   vm_paging_objects_mapped = 0;
10075 unsigned long   vm_paging_pages_mapped = 0;
10076 unsigned long   vm_paging_objects_mapped_slow = 0;
10077 unsigned long   vm_paging_pages_mapped_slow = 0;
10078 
10079 __startup_func
10080 static void
vm_paging_map_init(void)10081 vm_paging_map_init(void)
10082 {
10083 	kmem_alloc(kernel_map, &vm_paging_base_address,
10084 	    ptoa(VM_PAGING_NUM_PAGES),
10085 	    KMA_DATA | KMA_NOFAIL | KMA_KOBJECT | KMA_PERMANENT | KMA_PAGEABLE,
10086 	    VM_KERN_MEMORY_NONE);
10087 }
10088 STARTUP(ZALLOC, STARTUP_RANK_LAST, vm_paging_map_init);
10089 
10090 /*
10091  * vm_paging_map_object:
10092  *	Maps part of a VM object's pages in the kernel
10093  *      virtual address space, using the pre-allocated
10094  *	kernel virtual addresses, if possible.
10095  * Context:
10096  *      The VM object is locked.  This lock will get
10097  *      dropped and re-acquired though, so the caller
10098  *      must make sure the VM object is kept alive
10099  *	(by holding a VM map that has a reference
10100  *      on it, for example, or taking an extra reference).
10101  *      The page should also be kept busy to prevent
10102  *	it from being reclaimed.
10103  */
10104 kern_return_t
vm_paging_map_object(vm_page_t page,vm_object_t object,vm_object_offset_t offset,vm_prot_t protection,boolean_t can_unlock_object,vm_map_size_t * size,vm_map_offset_t * address,boolean_t * need_unmap)10105 vm_paging_map_object(
10106 	vm_page_t               page,
10107 	vm_object_t             object,
10108 	vm_object_offset_t      offset,
10109 	vm_prot_t               protection,
10110 	boolean_t               can_unlock_object,
10111 	vm_map_size_t           *size,          /* IN/OUT */
10112 	vm_map_offset_t         *address,       /* OUT */
10113 	boolean_t               *need_unmap)    /* OUT */
10114 {
10115 	kern_return_t           kr;
10116 	vm_map_offset_t         page_map_offset;
10117 	vm_map_size_t           map_size;
10118 	vm_object_offset_t      object_offset;
10119 	int                     i;
10120 
10121 	if (page != VM_PAGE_NULL && *size == PAGE_SIZE) {
10122 		/* use permanent 1-to-1 kernel mapping of physical memory ? */
10123 		*address = (vm_map_offset_t)
10124 		    phystokv((pmap_paddr_t)VM_PAGE_GET_PHYS_PAGE(page) << PAGE_SHIFT);
10125 		*need_unmap = FALSE;
10126 		return KERN_SUCCESS;
10127 
10128 		assert(page->vmp_busy);
10129 		/*
10130 		 * Use one of the pre-allocated kernel virtual addresses
10131 		 * and just enter the VM page in the kernel address space
10132 		 * at that virtual address.
10133 		 */
10134 		simple_lock(&vm_paging_lock, &vm_pageout_lck_grp);
10135 
10136 		/*
10137 		 * Try and find an available kernel virtual address
10138 		 * from our pre-allocated pool.
10139 		 */
10140 		page_map_offset = 0;
10141 		for (;;) {
10142 			for (i = 0; i < VM_PAGING_NUM_PAGES; i++) {
10143 				if (vm_paging_page_inuse[i] == FALSE) {
10144 					page_map_offset =
10145 					    vm_paging_base_address +
10146 					    (i * PAGE_SIZE);
10147 					break;
10148 				}
10149 			}
10150 			if (page_map_offset != 0) {
10151 				/* found a space to map our page ! */
10152 				break;
10153 			}
10154 
10155 			if (can_unlock_object) {
10156 				/*
10157 				 * If we can afford to unlock the VM object,
10158 				 * let's take the slow path now...
10159 				 */
10160 				break;
10161 			}
10162 			/*
10163 			 * We can't afford to unlock the VM object, so
10164 			 * let's wait for a space to become available...
10165 			 */
10166 			vm_paging_page_waiter_total++;
10167 			vm_paging_page_waiter++;
10168 			kr = assert_wait((event_t)&vm_paging_page_waiter, THREAD_UNINT);
10169 			if (kr == THREAD_WAITING) {
10170 				simple_unlock(&vm_paging_lock);
10171 				kr = thread_block(THREAD_CONTINUE_NULL);
10172 				simple_lock(&vm_paging_lock, &vm_pageout_lck_grp);
10173 			}
10174 			vm_paging_page_waiter--;
10175 			/* ... and try again */
10176 		}
10177 
10178 		if (page_map_offset != 0) {
10179 			/*
10180 			 * We found a kernel virtual address;
10181 			 * map the physical page to that virtual address.
10182 			 */
10183 			if (i > vm_paging_max_index) {
10184 				vm_paging_max_index = i;
10185 			}
10186 			vm_paging_page_inuse[i] = TRUE;
10187 			simple_unlock(&vm_paging_lock);
10188 
10189 			page->vmp_pmapped = TRUE;
10190 
10191 			/*
10192 			 * Keep the VM object locked over the PMAP_ENTER
10193 			 * and the actual use of the page by the kernel,
10194 			 * or this pmap mapping might get undone by a
10195 			 * vm_object_pmap_protect() call...
10196 			 */
10197 			PMAP_ENTER(kernel_pmap,
10198 			    page_map_offset,
10199 			    page,
10200 			    protection,
10201 			    VM_PROT_NONE,
10202 			    0,
10203 			    TRUE,
10204 			    kr);
10205 			assert(kr == KERN_SUCCESS);
10206 			vm_paging_objects_mapped++;
10207 			vm_paging_pages_mapped++;
10208 			*address = page_map_offset;
10209 			*need_unmap = TRUE;
10210 
10211 #if KASAN
10212 			kasan_notify_address(page_map_offset, PAGE_SIZE);
10213 #endif
10214 
10215 			/* all done and mapped, ready to use ! */
10216 			return KERN_SUCCESS;
10217 		}
10218 
10219 		/*
10220 		 * We ran out of pre-allocated kernel virtual
10221 		 * addresses.  Just map the page in the kernel
10222 		 * the slow and regular way.
10223 		 */
10224 		vm_paging_no_kernel_page++;
10225 		simple_unlock(&vm_paging_lock);
10226 	}
10227 
10228 	if (!can_unlock_object) {
10229 		*address = 0;
10230 		*size = 0;
10231 		*need_unmap = FALSE;
10232 		return KERN_NOT_SUPPORTED;
10233 	}
10234 
10235 	object_offset = vm_object_trunc_page(offset);
10236 	map_size = vm_map_round_page(*size,
10237 	    VM_MAP_PAGE_MASK(kernel_map));
10238 
10239 	/*
10240 	 * Try and map the required range of the object
10241 	 * in the kernel_map. Given that allocation is
10242 	 * for pageable memory, it shouldn't contain
10243 	 * pointers and is mapped into the data range.
10244 	 */
10245 
10246 	vm_object_reference_locked(object);     /* for the map entry */
10247 	vm_object_unlock(object);
10248 
10249 	kr = vm_map_enter(kernel_map,
10250 	    address,
10251 	    map_size,
10252 	    0,
10253 	    VM_MAP_KERNEL_FLAGS_DATA_ANYWHERE(),
10254 	    object,
10255 	    object_offset,
10256 	    FALSE,
10257 	    protection,
10258 	    VM_PROT_ALL,
10259 	    VM_INHERIT_NONE);
10260 	if (kr != KERN_SUCCESS) {
10261 		*address = 0;
10262 		*size = 0;
10263 		*need_unmap = FALSE;
10264 		vm_object_deallocate(object);   /* for the map entry */
10265 		vm_object_lock(object);
10266 		return kr;
10267 	}
10268 
10269 	*size = map_size;
10270 
10271 	/*
10272 	 * Enter the mapped pages in the page table now.
10273 	 */
10274 	vm_object_lock(object);
10275 	/*
10276 	 * VM object must be kept locked from before PMAP_ENTER()
10277 	 * until after the kernel is done accessing the page(s).
10278 	 * Otherwise, the pmap mappings in the kernel could be
10279 	 * undone by a call to vm_object_pmap_protect().
10280 	 */
10281 
10282 	for (page_map_offset = 0;
10283 	    map_size != 0;
10284 	    map_size -= PAGE_SIZE_64, page_map_offset += PAGE_SIZE_64) {
10285 		page = vm_page_lookup(object, offset + page_map_offset);
10286 		if (page == VM_PAGE_NULL) {
10287 			printf("vm_paging_map_object: no page !?");
10288 			vm_object_unlock(object);
10289 			vm_map_remove(kernel_map, *address, *size);
10290 			*address = 0;
10291 			*size = 0;
10292 			*need_unmap = FALSE;
10293 			vm_object_lock(object);
10294 			return KERN_MEMORY_ERROR;
10295 		}
10296 		page->vmp_pmapped = TRUE;
10297 
10298 		PMAP_ENTER(kernel_pmap,
10299 		    *address + page_map_offset,
10300 		    page,
10301 		    protection,
10302 		    VM_PROT_NONE,
10303 		    0,
10304 		    TRUE,
10305 		    kr);
10306 		assert(kr == KERN_SUCCESS);
10307 #if KASAN
10308 		kasan_notify_address(*address + page_map_offset, PAGE_SIZE);
10309 #endif
10310 	}
10311 
10312 	vm_paging_objects_mapped_slow++;
10313 	vm_paging_pages_mapped_slow += (unsigned long) (map_size / PAGE_SIZE_64);
10314 
10315 	*need_unmap = TRUE;
10316 
10317 	return KERN_SUCCESS;
10318 }
10319 
10320 /*
10321  * vm_paging_unmap_object:
10322  *	Unmaps part of a VM object's pages from the kernel
10323  *      virtual address space.
10324  * Context:
10325  *      The VM object is locked.  This lock will get
10326  *      dropped and re-acquired though.
10327  */
10328 void
vm_paging_unmap_object(vm_object_t object,vm_map_offset_t start,vm_map_offset_t end)10329 vm_paging_unmap_object(
10330 	vm_object_t     object,
10331 	vm_map_offset_t start,
10332 	vm_map_offset_t end)
10333 {
10334 	int             i;
10335 
10336 	if ((vm_paging_base_address == 0) ||
10337 	    (start < vm_paging_base_address) ||
10338 	    (end > (vm_paging_base_address
10339 	    + (VM_PAGING_NUM_PAGES * PAGE_SIZE)))) {
10340 		/*
10341 		 * We didn't use our pre-allocated pool of
10342 		 * kernel virtual address.  Deallocate the
10343 		 * virtual memory.
10344 		 */
10345 		if (object != VM_OBJECT_NULL) {
10346 			vm_object_unlock(object);
10347 		}
10348 		vm_map_remove(kernel_map, start, end);
10349 		if (object != VM_OBJECT_NULL) {
10350 			vm_object_lock(object);
10351 		}
10352 	} else {
10353 		/*
10354 		 * We used a kernel virtual address from our
10355 		 * pre-allocated pool.  Put it back in the pool
10356 		 * for next time.
10357 		 */
10358 		assert(end - start == PAGE_SIZE);
10359 		i = (int) ((start - vm_paging_base_address) >> PAGE_SHIFT);
10360 		assert(i >= 0 && i < VM_PAGING_NUM_PAGES);
10361 
10362 		/* undo the pmap mapping */
10363 		pmap_remove(kernel_pmap, start, end);
10364 
10365 		simple_lock(&vm_paging_lock, &vm_pageout_lck_grp);
10366 		vm_paging_page_inuse[i] = FALSE;
10367 		if (vm_paging_page_waiter) {
10368 			thread_wakeup(&vm_paging_page_waiter);
10369 		}
10370 		simple_unlock(&vm_paging_lock);
10371 	}
10372 }
10373 
10374 
10375 /*
10376  * page->vmp_object must be locked
10377  */
10378 void
vm_pageout_steal_laundry(vm_page_t page,boolean_t queues_locked)10379 vm_pageout_steal_laundry(vm_page_t page, boolean_t queues_locked)
10380 {
10381 	if (!queues_locked) {
10382 		vm_page_lockspin_queues();
10383 	}
10384 
10385 	page->vmp_free_when_done = FALSE;
10386 	/*
10387 	 * need to drop the laundry count...
10388 	 * we may also need to remove it
10389 	 * from the I/O paging queue...
10390 	 * vm_pageout_throttle_up handles both cases
10391 	 *
10392 	 * the laundry and pageout_queue flags are cleared...
10393 	 */
10394 	vm_pageout_throttle_up(page);
10395 
10396 	if (!queues_locked) {
10397 		vm_page_unlock_queues();
10398 	}
10399 }
10400 
10401 #define VECTOR_UPL_ELEMENTS_UPPER_LIMIT 64
10402 
10403 upl_t
vector_upl_create(vm_offset_t upl_offset,uint32_t max_upls)10404 vector_upl_create(vm_offset_t upl_offset, uint32_t max_upls)
10405 {
10406 	int i = 0;
10407 	upl_t   upl;
10408 
10409 	assert(max_upls > 0);
10410 	if (max_upls == 0) {
10411 		return NULL;
10412 	}
10413 
10414 	if (max_upls > VECTOR_UPL_ELEMENTS_UPPER_LIMIT) {
10415 		max_upls = VECTOR_UPL_ELEMENTS_UPPER_LIMIT;
10416 	}
10417 	vector_upl_t vector_upl = kalloc_type(struct _vector_upl, typeof(vector_upl->upls[0]), max_upls, Z_WAITOK | Z_NOFAIL);
10418 
10419 	upl = upl_create(0, UPL_VECTOR, 0);
10420 	upl->vector_upl = vector_upl;
10421 	upl->u_offset = upl_offset;
10422 	vector_upl->size = 0;
10423 	vector_upl->offset = upl_offset;
10424 	vector_upl->invalid_upls = 0;
10425 	vector_upl->num_upls = 0;
10426 	vector_upl->pagelist = NULL;
10427 	vector_upl->max_upls = max_upls;
10428 
10429 	for (i = 0; i < max_upls; i++) {
10430 		vector_upl->upls[i].iostate.size = 0;
10431 		vector_upl->upls[i].iostate.offset = 0;
10432 	}
10433 	return upl;
10434 }
10435 
10436 uint32_t
vector_upl_max_upls(const upl_t upl)10437 vector_upl_max_upls(const upl_t upl)
10438 {
10439 	if (!vector_upl_is_valid(upl)) {
10440 		return 0;
10441 	}
10442 	return ((vector_upl_t)(upl->vector_upl))->max_upls;
10443 }
10444 
10445 void
vector_upl_deallocate(upl_t upl)10446 vector_upl_deallocate(upl_t upl)
10447 {
10448 	vector_upl_t vector_upl = upl->vector_upl;
10449 
10450 	assert(vector_upl_is_valid(upl));
10451 
10452 	if (vector_upl->invalid_upls != vector_upl->num_upls) {
10453 		panic("Deallocating non-empty Vectored UPL");
10454 	}
10455 	uint32_t max_upls = vector_upl->max_upls;
10456 	kfree_type(struct upl_page_info, atop(vector_upl->size), vector_upl->pagelist);
10457 	kfree_type(struct _vector_upl, typeof(vector_upl->upls[0]), max_upls, vector_upl);
10458 	upl->vector_upl = NULL;
10459 }
10460 
10461 boolean_t
vector_upl_is_valid(upl_t upl)10462 vector_upl_is_valid(upl_t upl)
10463 {
10464 	return upl && (upl->flags & UPL_VECTOR) && upl->vector_upl;
10465 }
10466 
10467 boolean_t
vector_upl_set_subupl(upl_t upl,upl_t subupl,uint32_t io_size)10468 vector_upl_set_subupl(upl_t upl, upl_t subupl, uint32_t io_size)
10469 {
10470 	if (vector_upl_is_valid(upl)) {
10471 		vector_upl_t vector_upl = upl->vector_upl;
10472 
10473 		if (vector_upl) {
10474 			if (subupl) {
10475 				if (io_size) {
10476 					if (io_size < PAGE_SIZE) {
10477 						io_size = PAGE_SIZE;
10478 					}
10479 					subupl->vector_upl = (void*)vector_upl;
10480 					vector_upl->upls[vector_upl->num_upls++].elem = subupl;
10481 					vector_upl->size += io_size;
10482 					upl->u_size += io_size;
10483 				} else {
10484 					uint32_t i = 0, invalid_upls = 0;
10485 					for (i = 0; i < vector_upl->num_upls; i++) {
10486 						if (vector_upl->upls[i].elem == subupl) {
10487 							break;
10488 						}
10489 					}
10490 					if (i == vector_upl->num_upls) {
10491 						panic("Trying to remove sub-upl when none exists");
10492 					}
10493 
10494 					vector_upl->upls[i].elem = NULL;
10495 					invalid_upls = os_atomic_inc(&(vector_upl)->invalid_upls,
10496 					    relaxed);
10497 					if (invalid_upls == vector_upl->num_upls) {
10498 						return TRUE;
10499 					} else {
10500 						return FALSE;
10501 					}
10502 				}
10503 			} else {
10504 				panic("vector_upl_set_subupl was passed a NULL upl element");
10505 			}
10506 		} else {
10507 			panic("vector_upl_set_subupl was passed a non-vectored upl");
10508 		}
10509 	} else {
10510 		panic("vector_upl_set_subupl was passed a NULL upl");
10511 	}
10512 
10513 	return FALSE;
10514 }
10515 
10516 void
vector_upl_set_pagelist(upl_t upl)10517 vector_upl_set_pagelist(upl_t upl)
10518 {
10519 	if (vector_upl_is_valid(upl)) {
10520 		uint32_t i = 0;
10521 		vector_upl_t vector_upl = upl->vector_upl;
10522 
10523 		if (vector_upl) {
10524 			vm_offset_t pagelist_size = 0, cur_upl_pagelist_size = 0;
10525 
10526 			vector_upl->pagelist = kalloc_type(struct upl_page_info,
10527 			    atop(vector_upl->size), Z_WAITOK);
10528 
10529 			for (i = 0; i < vector_upl->num_upls; i++) {
10530 				cur_upl_pagelist_size = sizeof(struct upl_page_info) * upl_adjusted_size(vector_upl->upls[i].elem, PAGE_MASK) / PAGE_SIZE;
10531 				bcopy(vector_upl->upls[i].elem->page_list, (char*)vector_upl->pagelist + pagelist_size, cur_upl_pagelist_size);
10532 				pagelist_size += cur_upl_pagelist_size;
10533 				if (vector_upl->upls[i].elem->highest_page > upl->highest_page) {
10534 					upl->highest_page = vector_upl->upls[i].elem->highest_page;
10535 				}
10536 			}
10537 			assert( pagelist_size == (sizeof(struct upl_page_info) * (vector_upl->size / PAGE_SIZE)));
10538 		} else {
10539 			panic("vector_upl_set_pagelist was passed a non-vectored upl");
10540 		}
10541 	} else {
10542 		panic("vector_upl_set_pagelist was passed a NULL upl");
10543 	}
10544 }
10545 
10546 upl_t
vector_upl_subupl_byindex(upl_t upl,uint32_t index)10547 vector_upl_subupl_byindex(upl_t upl, uint32_t index)
10548 {
10549 	if (vector_upl_is_valid(upl)) {
10550 		vector_upl_t vector_upl = upl->vector_upl;
10551 		if (vector_upl) {
10552 			if (index < vector_upl->num_upls) {
10553 				return vector_upl->upls[index].elem;
10554 			}
10555 		} else {
10556 			panic("vector_upl_subupl_byindex was passed a non-vectored upl");
10557 		}
10558 	}
10559 	return NULL;
10560 }
10561 
10562 upl_t
vector_upl_subupl_byoffset(upl_t upl,upl_offset_t * upl_offset,upl_size_t * upl_size)10563 vector_upl_subupl_byoffset(upl_t upl, upl_offset_t *upl_offset, upl_size_t *upl_size)
10564 {
10565 	if (vector_upl_is_valid(upl)) {
10566 		uint32_t i = 0;
10567 		vector_upl_t vector_upl = upl->vector_upl;
10568 
10569 		if (vector_upl) {
10570 			upl_t subupl = NULL;
10571 			vector_upl_iostates_t subupl_state;
10572 
10573 			for (i = 0; i < vector_upl->num_upls; i++) {
10574 				subupl = vector_upl->upls[i].elem;
10575 				subupl_state = vector_upl->upls[i].iostate;
10576 				if (*upl_offset <= (subupl_state.offset + subupl_state.size - 1)) {
10577 					/* We could have been passed an offset/size pair that belongs
10578 					 * to an UPL element that has already been committed/aborted.
10579 					 * If so, return NULL.
10580 					 */
10581 					if (subupl == NULL) {
10582 						return NULL;
10583 					}
10584 					if ((subupl_state.offset + subupl_state.size) < (*upl_offset + *upl_size)) {
10585 						*upl_size = (subupl_state.offset + subupl_state.size) - *upl_offset;
10586 						if (*upl_size > subupl_state.size) {
10587 							*upl_size = subupl_state.size;
10588 						}
10589 					}
10590 					if (*upl_offset >= subupl_state.offset) {
10591 						*upl_offset -= subupl_state.offset;
10592 					} else if (i) {
10593 						panic("Vector UPL offset miscalculation");
10594 					}
10595 					return subupl;
10596 				}
10597 			}
10598 		} else {
10599 			panic("vector_upl_subupl_byoffset was passed a non-vectored UPL");
10600 		}
10601 	}
10602 	return NULL;
10603 }
10604 
10605 void
vector_upl_get_submap(upl_t upl,vm_map_t * v_upl_submap,vm_offset_t * submap_dst_addr)10606 vector_upl_get_submap(upl_t upl, vm_map_t *v_upl_submap, vm_offset_t *submap_dst_addr)
10607 {
10608 	*v_upl_submap = NULL;
10609 
10610 	if (vector_upl_is_valid(upl)) {
10611 		vector_upl_t vector_upl = upl->vector_upl;
10612 		if (vector_upl) {
10613 			*v_upl_submap = vector_upl->submap;
10614 			*submap_dst_addr = vector_upl->submap_dst_addr;
10615 		} else {
10616 			panic("vector_upl_get_submap was passed a non-vectored UPL");
10617 		}
10618 	} else {
10619 		panic("vector_upl_get_submap was passed a null UPL");
10620 	}
10621 }
10622 
10623 void
vector_upl_set_submap(upl_t upl,vm_map_t submap,vm_offset_t submap_dst_addr)10624 vector_upl_set_submap(upl_t upl, vm_map_t submap, vm_offset_t submap_dst_addr)
10625 {
10626 	if (vector_upl_is_valid(upl)) {
10627 		vector_upl_t vector_upl = upl->vector_upl;
10628 		if (vector_upl) {
10629 			vector_upl->submap = submap;
10630 			vector_upl->submap_dst_addr = submap_dst_addr;
10631 		} else {
10632 			panic("vector_upl_get_submap was passed a non-vectored UPL");
10633 		}
10634 	} else {
10635 		panic("vector_upl_get_submap was passed a NULL UPL");
10636 	}
10637 }
10638 
10639 void
vector_upl_set_iostate(upl_t upl,upl_t subupl,upl_offset_t offset,upl_size_t size)10640 vector_upl_set_iostate(upl_t upl, upl_t subupl, upl_offset_t offset, upl_size_t size)
10641 {
10642 	if (vector_upl_is_valid(upl)) {
10643 		uint32_t i = 0;
10644 		vector_upl_t vector_upl = upl->vector_upl;
10645 
10646 		if (vector_upl) {
10647 			for (i = 0; i < vector_upl->num_upls; i++) {
10648 				if (vector_upl->upls[i].elem == subupl) {
10649 					break;
10650 				}
10651 			}
10652 
10653 			if (i == vector_upl->num_upls) {
10654 				panic("setting sub-upl iostate when none exists");
10655 			}
10656 
10657 			vector_upl->upls[i].iostate.offset = offset;
10658 			if (size < PAGE_SIZE) {
10659 				size = PAGE_SIZE;
10660 			}
10661 			vector_upl->upls[i].iostate.size = size;
10662 		} else {
10663 			panic("vector_upl_set_iostate was passed a non-vectored UPL");
10664 		}
10665 	} else {
10666 		panic("vector_upl_set_iostate was passed a NULL UPL");
10667 	}
10668 }
10669 
10670 void
vector_upl_get_iostate(upl_t upl,upl_t subupl,upl_offset_t * offset,upl_size_t * size)10671 vector_upl_get_iostate(upl_t upl, upl_t subupl, upl_offset_t *offset, upl_size_t *size)
10672 {
10673 	if (vector_upl_is_valid(upl)) {
10674 		uint32_t i = 0;
10675 		vector_upl_t vector_upl = upl->vector_upl;
10676 
10677 		if (vector_upl) {
10678 			for (i = 0; i < vector_upl->num_upls; i++) {
10679 				if (vector_upl->upls[i].elem == subupl) {
10680 					break;
10681 				}
10682 			}
10683 
10684 			if (i == vector_upl->num_upls) {
10685 				panic("getting sub-upl iostate when none exists");
10686 			}
10687 
10688 			*offset = vector_upl->upls[i].iostate.offset;
10689 			*size = vector_upl->upls[i].iostate.size;
10690 		} else {
10691 			panic("vector_upl_get_iostate was passed a non-vectored UPL");
10692 		}
10693 	} else {
10694 		panic("vector_upl_get_iostate was passed a NULL UPL");
10695 	}
10696 }
10697 
10698 void
vector_upl_get_iostate_byindex(upl_t upl,uint32_t index,upl_offset_t * offset,upl_size_t * size)10699 vector_upl_get_iostate_byindex(upl_t upl, uint32_t index, upl_offset_t *offset, upl_size_t *size)
10700 {
10701 	if (vector_upl_is_valid(upl)) {
10702 		vector_upl_t vector_upl = upl->vector_upl;
10703 		if (vector_upl) {
10704 			if (index < vector_upl->num_upls) {
10705 				*offset = vector_upl->upls[index].iostate.offset;
10706 				*size = vector_upl->upls[index].iostate.size;
10707 			} else {
10708 				*offset = *size = 0;
10709 			}
10710 		} else {
10711 			panic("vector_upl_get_iostate_byindex was passed a non-vectored UPL");
10712 		}
10713 	} else {
10714 		panic("vector_upl_get_iostate_byindex was passed a NULL UPL");
10715 	}
10716 }
10717 
10718 void *
upl_get_internal_vectorupl(upl_t upl)10719 upl_get_internal_vectorupl(upl_t upl)
10720 {
10721 	return upl->vector_upl;
10722 }
10723 
10724 upl_page_info_t *
upl_get_internal_vectorupl_pagelist(upl_t upl)10725 upl_get_internal_vectorupl_pagelist(upl_t upl)
10726 {
10727 	return upl->vector_upl->pagelist;
10728 }
10729 
10730 upl_page_info_t *
upl_get_internal_page_list(upl_t upl)10731 upl_get_internal_page_list(upl_t upl)
10732 {
10733 	return upl->vector_upl ? upl->vector_upl->pagelist : upl->page_list;
10734 }
10735 
10736 void
upl_clear_dirty(upl_t upl,boolean_t value)10737 upl_clear_dirty(
10738 	upl_t           upl,
10739 	boolean_t       value)
10740 {
10741 	if (value) {
10742 		upl->flags |= UPL_CLEAR_DIRTY;
10743 	} else {
10744 		upl->flags &= ~UPL_CLEAR_DIRTY;
10745 	}
10746 }
10747 
10748 void
upl_set_referenced(upl_t upl,boolean_t value)10749 upl_set_referenced(
10750 	upl_t           upl,
10751 	boolean_t       value)
10752 {
10753 	upl_lock(upl);
10754 	if (value) {
10755 		upl->ext_ref_count++;
10756 	} else {
10757 		if (!upl->ext_ref_count) {
10758 			panic("upl_set_referenced not %p", upl);
10759 		}
10760 		upl->ext_ref_count--;
10761 	}
10762 	upl_unlock(upl);
10763 }
10764 
10765 #if CONFIG_IOSCHED
10766 void
upl_set_blkno(upl_t upl,vm_offset_t upl_offset,int io_size,int64_t blkno)10767 upl_set_blkno(
10768 	upl_t           upl,
10769 	vm_offset_t     upl_offset,
10770 	int             io_size,
10771 	int64_t         blkno)
10772 {
10773 	int i, j;
10774 	if ((upl->flags & UPL_EXPEDITE_SUPPORTED) == 0) {
10775 		return;
10776 	}
10777 
10778 	assert(upl->upl_reprio_info != 0);
10779 	for (i = (int)(upl_offset / PAGE_SIZE), j = 0; j < io_size; i++, j += PAGE_SIZE) {
10780 		UPL_SET_REPRIO_INFO(upl, i, blkno, io_size);
10781 	}
10782 }
10783 #endif
10784 
10785 void inline
memoryshot(unsigned int event,unsigned int control)10786 memoryshot(unsigned int event, unsigned int control)
10787 {
10788 	if (vm_debug_events) {
10789 		KERNEL_DEBUG_CONSTANT1((MACHDBG_CODE(DBG_MACH_VM_PRESSURE, event)) | control,
10790 		    vm_page_active_count, vm_page_inactive_count,
10791 		    vm_page_free_count, vm_page_speculative_count,
10792 		    vm_page_throttled_count);
10793 	} else {
10794 		(void) event;
10795 		(void) control;
10796 	}
10797 }
10798 
10799 #ifdef MACH_BSD
10800 
10801 boolean_t
upl_device_page(upl_page_info_t * upl)10802 upl_device_page(upl_page_info_t *upl)
10803 {
10804 	return UPL_DEVICE_PAGE(upl);
10805 }
10806 boolean_t
upl_page_present(upl_page_info_t * upl,int index)10807 upl_page_present(upl_page_info_t *upl, int index)
10808 {
10809 	return UPL_PAGE_PRESENT(upl, index);
10810 }
10811 boolean_t
upl_speculative_page(upl_page_info_t * upl,int index)10812 upl_speculative_page(upl_page_info_t *upl, int index)
10813 {
10814 	return UPL_SPECULATIVE_PAGE(upl, index);
10815 }
10816 boolean_t
upl_dirty_page(upl_page_info_t * upl,int index)10817 upl_dirty_page(upl_page_info_t *upl, int index)
10818 {
10819 	return UPL_DIRTY_PAGE(upl, index);
10820 }
10821 boolean_t
upl_valid_page(upl_page_info_t * upl,int index)10822 upl_valid_page(upl_page_info_t *upl, int index)
10823 {
10824 	return UPL_VALID_PAGE(upl, index);
10825 }
10826 ppnum_t
upl_phys_page(upl_page_info_t * upl,int index)10827 upl_phys_page(upl_page_info_t *upl, int index)
10828 {
10829 	return UPL_PHYS_PAGE(upl, index);
10830 }
10831 
10832 void
upl_page_set_mark(upl_page_info_t * upl,int index,boolean_t v)10833 upl_page_set_mark(upl_page_info_t *upl, int index, boolean_t v)
10834 {
10835 	upl[index].mark = v;
10836 }
10837 
10838 boolean_t
upl_page_get_mark(upl_page_info_t * upl,int index)10839 upl_page_get_mark(upl_page_info_t *upl, int index)
10840 {
10841 	return upl[index].mark;
10842 }
10843 
10844 void
vm_countdirtypages(void)10845 vm_countdirtypages(void)
10846 {
10847 	vm_page_t m;
10848 	int dpages;
10849 	int pgopages;
10850 	int precpages;
10851 
10852 
10853 	dpages = 0;
10854 	pgopages = 0;
10855 	precpages = 0;
10856 
10857 	vm_page_lock_queues();
10858 	m = (vm_page_t) vm_page_queue_first(&vm_page_queue_inactive);
10859 	do {
10860 		if (m == (vm_page_t)0) {
10861 			break;
10862 		}
10863 
10864 		if (m->vmp_dirty) {
10865 			dpages++;
10866 		}
10867 		if (m->vmp_free_when_done) {
10868 			pgopages++;
10869 		}
10870 		if (m->vmp_precious) {
10871 			precpages++;
10872 		}
10873 
10874 		assert(VM_PAGE_OBJECT(m) != kernel_object);
10875 		m = (vm_page_t) vm_page_queue_next(&m->vmp_pageq);
10876 		if (m == (vm_page_t)0) {
10877 			break;
10878 		}
10879 	} while (!vm_page_queue_end(&vm_page_queue_inactive, (vm_page_queue_entry_t) m));
10880 	vm_page_unlock_queues();
10881 
10882 	vm_page_lock_queues();
10883 	m = (vm_page_t) vm_page_queue_first(&vm_page_queue_throttled);
10884 	do {
10885 		if (m == (vm_page_t)0) {
10886 			break;
10887 		}
10888 
10889 		dpages++;
10890 		assert(m->vmp_dirty);
10891 		assert(!m->vmp_free_when_done);
10892 		assert(VM_PAGE_OBJECT(m) != kernel_object);
10893 		m = (vm_page_t) vm_page_queue_next(&m->vmp_pageq);
10894 		if (m == (vm_page_t)0) {
10895 			break;
10896 		}
10897 	} while (!vm_page_queue_end(&vm_page_queue_throttled, (vm_page_queue_entry_t) m));
10898 	vm_page_unlock_queues();
10899 
10900 	vm_page_lock_queues();
10901 	m = (vm_page_t) vm_page_queue_first(&vm_page_queue_anonymous);
10902 	do {
10903 		if (m == (vm_page_t)0) {
10904 			break;
10905 		}
10906 
10907 		if (m->vmp_dirty) {
10908 			dpages++;
10909 		}
10910 		if (m->vmp_free_when_done) {
10911 			pgopages++;
10912 		}
10913 		if (m->vmp_precious) {
10914 			precpages++;
10915 		}
10916 
10917 		assert(VM_PAGE_OBJECT(m) != kernel_object);
10918 		m = (vm_page_t) vm_page_queue_next(&m->vmp_pageq);
10919 		if (m == (vm_page_t)0) {
10920 			break;
10921 		}
10922 	} while (!vm_page_queue_end(&vm_page_queue_anonymous, (vm_page_queue_entry_t) m));
10923 	vm_page_unlock_queues();
10924 
10925 	printf("IN Q: %d : %d : %d\n", dpages, pgopages, precpages);
10926 
10927 	dpages = 0;
10928 	pgopages = 0;
10929 	precpages = 0;
10930 
10931 	vm_page_lock_queues();
10932 	m = (vm_page_t) vm_page_queue_first(&vm_page_queue_active);
10933 
10934 	do {
10935 		if (m == (vm_page_t)0) {
10936 			break;
10937 		}
10938 		if (m->vmp_dirty) {
10939 			dpages++;
10940 		}
10941 		if (m->vmp_free_when_done) {
10942 			pgopages++;
10943 		}
10944 		if (m->vmp_precious) {
10945 			precpages++;
10946 		}
10947 
10948 		assert(VM_PAGE_OBJECT(m) != kernel_object);
10949 		m = (vm_page_t) vm_page_queue_next(&m->vmp_pageq);
10950 		if (m == (vm_page_t)0) {
10951 			break;
10952 		}
10953 	} while (!vm_page_queue_end(&vm_page_queue_active, (vm_page_queue_entry_t) m));
10954 	vm_page_unlock_queues();
10955 
10956 	printf("AC Q: %d : %d : %d\n", dpages, pgopages, precpages);
10957 }
10958 #endif /* MACH_BSD */
10959 
10960 
10961 #if CONFIG_IOSCHED
10962 int
upl_get_cached_tier(upl_t upl)10963 upl_get_cached_tier(upl_t  upl)
10964 {
10965 	assert(upl);
10966 	if (upl->flags & UPL_TRACKED_BY_OBJECT) {
10967 		return upl->upl_priority;
10968 	}
10969 	return -1;
10970 }
10971 #endif /* CONFIG_IOSCHED */
10972 
10973 
10974 void
upl_callout_iodone(upl_t upl)10975 upl_callout_iodone(upl_t upl)
10976 {
10977 	struct upl_io_completion *upl_ctx = upl->upl_iodone;
10978 
10979 	if (upl_ctx) {
10980 		void    (*iodone_func)(void *, int) = upl_ctx->io_done;
10981 
10982 		assert(upl_ctx->io_done);
10983 
10984 		(*iodone_func)(upl_ctx->io_context, upl_ctx->io_error);
10985 	}
10986 }
10987 
10988 void
upl_set_iodone(upl_t upl,void * upl_iodone)10989 upl_set_iodone(upl_t upl, void *upl_iodone)
10990 {
10991 	upl->upl_iodone = (struct upl_io_completion *)upl_iodone;
10992 }
10993 
10994 void
upl_set_iodone_error(upl_t upl,int error)10995 upl_set_iodone_error(upl_t upl, int error)
10996 {
10997 	struct upl_io_completion *upl_ctx = upl->upl_iodone;
10998 
10999 	if (upl_ctx) {
11000 		upl_ctx->io_error = error;
11001 	}
11002 }
11003 
11004 
11005 ppnum_t
upl_get_highest_page(upl_t upl)11006 upl_get_highest_page(
11007 	upl_t                      upl)
11008 {
11009 	return upl->highest_page;
11010 }
11011 
11012 upl_size_t
upl_get_size(upl_t upl)11013 upl_get_size(
11014 	upl_t                      upl)
11015 {
11016 	return upl_adjusted_size(upl, PAGE_MASK);
11017 }
11018 
11019 upl_size_t
upl_adjusted_size(upl_t upl,vm_map_offset_t pgmask)11020 upl_adjusted_size(
11021 	upl_t upl,
11022 	vm_map_offset_t pgmask)
11023 {
11024 	vm_object_offset_t start_offset, end_offset;
11025 
11026 	start_offset = trunc_page_mask_64(upl->u_offset, pgmask);
11027 	end_offset = round_page_mask_64(upl->u_offset + upl->u_size, pgmask);
11028 
11029 	return (upl_size_t)(end_offset - start_offset);
11030 }
11031 
11032 vm_object_offset_t
upl_adjusted_offset(upl_t upl,vm_map_offset_t pgmask)11033 upl_adjusted_offset(
11034 	upl_t upl,
11035 	vm_map_offset_t pgmask)
11036 {
11037 	return trunc_page_mask_64(upl->u_offset, pgmask);
11038 }
11039 
11040 vm_object_offset_t
upl_get_data_offset(upl_t upl)11041 upl_get_data_offset(
11042 	upl_t upl)
11043 {
11044 	return upl->u_offset - upl_adjusted_offset(upl, PAGE_MASK);
11045 }
11046 
11047 upl_t
upl_associated_upl(upl_t upl)11048 upl_associated_upl(upl_t upl)
11049 {
11050 	return upl->associated_upl;
11051 }
11052 
11053 void
upl_set_associated_upl(upl_t upl,upl_t associated_upl)11054 upl_set_associated_upl(upl_t upl, upl_t associated_upl)
11055 {
11056 	upl->associated_upl = associated_upl;
11057 }
11058 
11059 struct vnode *
upl_lookup_vnode(upl_t upl)11060 upl_lookup_vnode(upl_t upl)
11061 {
11062 	if (!upl->map_object->internal) {
11063 		return vnode_pager_lookup_vnode(upl->map_object->pager);
11064 	} else {
11065 		return NULL;
11066 	}
11067 }
11068 
11069 #if UPL_DEBUG
11070 kern_return_t
upl_ubc_alias_set(upl_t upl,uintptr_t alias1,uintptr_t alias2)11071 upl_ubc_alias_set(upl_t upl, uintptr_t alias1, uintptr_t alias2)
11072 {
11073 	upl->ubc_alias1 = alias1;
11074 	upl->ubc_alias2 = alias2;
11075 	return KERN_SUCCESS;
11076 }
11077 int
upl_ubc_alias_get(upl_t upl,uintptr_t * al,uintptr_t * al2)11078 upl_ubc_alias_get(upl_t upl, uintptr_t * al, uintptr_t * al2)
11079 {
11080 	if (al) {
11081 		*al = upl->ubc_alias1;
11082 	}
11083 	if (al2) {
11084 		*al2 = upl->ubc_alias2;
11085 	}
11086 	return KERN_SUCCESS;
11087 }
11088 #endif /* UPL_DEBUG */
11089 
11090 #if VM_PRESSURE_EVENTS
11091 /*
11092  * Upward trajectory.
11093  */
11094 extern boolean_t vm_compressor_low_on_space(void);
11095 
11096 boolean_t
VM_PRESSURE_NORMAL_TO_WARNING(void)11097 VM_PRESSURE_NORMAL_TO_WARNING(void)
11098 {
11099 	if (!VM_CONFIG_COMPRESSOR_IS_ACTIVE) {
11100 		/* Available pages below our threshold */
11101 		if (memorystatus_available_pages < memorystatus_available_pages_pressure) {
11102 			/* No frozen processes to kill */
11103 			if (memorystatus_frozen_count == 0) {
11104 				/* Not enough suspended processes available. */
11105 				if (memorystatus_suspended_count < MEMORYSTATUS_SUSPENDED_THRESHOLD) {
11106 					return TRUE;
11107 				}
11108 			}
11109 		}
11110 		return FALSE;
11111 	} else {
11112 		return (AVAILABLE_NON_COMPRESSED_MEMORY < VM_PAGE_COMPRESSOR_COMPACT_THRESHOLD) ? 1 : 0;
11113 	}
11114 }
11115 
11116 boolean_t
VM_PRESSURE_WARNING_TO_CRITICAL(void)11117 VM_PRESSURE_WARNING_TO_CRITICAL(void)
11118 {
11119 	if (!VM_CONFIG_COMPRESSOR_IS_ACTIVE) {
11120 		/* Available pages below our threshold */
11121 		if (memorystatus_available_pages < memorystatus_available_pages_critical) {
11122 			return TRUE;
11123 		}
11124 		return FALSE;
11125 	} else {
11126 		return vm_compressor_low_on_space() || (AVAILABLE_NON_COMPRESSED_MEMORY < ((12 * VM_PAGE_COMPRESSOR_SWAP_UNTHROTTLE_THRESHOLD) / 10)) ? 1 : 0;
11127 	}
11128 }
11129 
11130 /*
11131  * Downward trajectory.
11132  */
11133 boolean_t
VM_PRESSURE_WARNING_TO_NORMAL(void)11134 VM_PRESSURE_WARNING_TO_NORMAL(void)
11135 {
11136 	if (!VM_CONFIG_COMPRESSOR_IS_ACTIVE) {
11137 		/* Available pages above our threshold */
11138 		unsigned int target_threshold = (unsigned int) (memorystatus_available_pages_pressure + ((15 * memorystatus_available_pages_pressure) / 100));
11139 		if (memorystatus_available_pages > target_threshold) {
11140 			return TRUE;
11141 		}
11142 		return FALSE;
11143 	} else {
11144 		return (AVAILABLE_NON_COMPRESSED_MEMORY > ((12 * VM_PAGE_COMPRESSOR_COMPACT_THRESHOLD) / 10)) ? 1 : 0;
11145 	}
11146 }
11147 
11148 boolean_t
VM_PRESSURE_CRITICAL_TO_WARNING(void)11149 VM_PRESSURE_CRITICAL_TO_WARNING(void)
11150 {
11151 	if (!VM_CONFIG_COMPRESSOR_IS_ACTIVE) {
11152 		/* Available pages above our threshold */
11153 		unsigned int target_threshold = (unsigned int)(memorystatus_available_pages_critical + ((15 * memorystatus_available_pages_critical) / 100));
11154 		if (memorystatus_available_pages > target_threshold) {
11155 			return TRUE;
11156 		}
11157 		return FALSE;
11158 	} else {
11159 		return (AVAILABLE_NON_COMPRESSED_MEMORY > ((14 * VM_PAGE_COMPRESSOR_SWAP_UNTHROTTLE_THRESHOLD) / 10)) ? 1 : 0;
11160 	}
11161 }
11162 #endif /* VM_PRESSURE_EVENTS */
11163 
11164 #if DEVELOPMENT || DEBUG
11165 bool compressor_running_perf_test;
11166 uint64_t compressor_perf_test_pages_processed;
11167 
11168 kern_return_t
11169 run_compressor_perf_test(
11170 	user_addr_t buf,
11171 	size_t buffer_size,
11172 	uint64_t *time,
11173 	uint64_t *bytes_compressed,
11174 	uint64_t *compressor_growth);
11175 
11176 static kern_return_t
move_pages_to_queue(vm_map_t map,user_addr_t start_addr,size_t buffer_size,vm_page_queue_head_t * queue,size_t * pages_moved)11177 move_pages_to_queue(
11178 	vm_map_t map,
11179 	user_addr_t start_addr,
11180 	size_t buffer_size,
11181 	vm_page_queue_head_t *queue,
11182 	size_t *pages_moved)
11183 {
11184 	kern_return_t err = KERN_SUCCESS;
11185 	vm_map_entry_t curr_entry = VM_MAP_ENTRY_NULL;
11186 	boolean_t addr_in_map = FALSE;
11187 	user_addr_t end_addr = USER_ADDR_NULL, curr_addr = USER_ADDR_NULL;
11188 	vm_object_t curr_object = VM_OBJECT_NULL;
11189 	*pages_moved = 0;
11190 
11191 
11192 	if (VM_MAP_PAGE_SIZE(map) != PAGE_SIZE_64) {
11193 		/*
11194 		 * We don't currently support benchmarking maps with a different page size
11195 		 * than the kernel.
11196 		 */
11197 		return KERN_INVALID_ARGUMENT;
11198 	}
11199 
11200 	if (os_add_overflow(start_addr, buffer_size, &end_addr)) {
11201 		return KERN_INVALID_ARGUMENT;
11202 	}
11203 
11204 	vm_map_lock_read(map);
11205 	curr_addr = vm_map_trunc_page_mask(start_addr, VM_MAP_PAGE_MASK(map));
11206 	end_addr = vm_map_round_page_mask(start_addr + buffer_size, VM_MAP_PAGE_MASK(map));
11207 
11208 
11209 	while (curr_addr < end_addr) {
11210 		addr_in_map = vm_map_lookup_entry(map, curr_addr, &curr_entry);
11211 		if (!addr_in_map) {
11212 			err = KERN_INVALID_ARGUMENT;
11213 			break;
11214 		}
11215 		curr_object = VME_OBJECT(curr_entry);
11216 		if (curr_object) {
11217 			vm_object_lock(curr_object);
11218 			/* We really only want anonymous memory that's in the top level map and object here. */
11219 			if (curr_entry->is_sub_map || curr_entry->wired_count != 0 ||
11220 			    curr_object->shadow != VM_OBJECT_NULL || !curr_object->internal) {
11221 				err = KERN_INVALID_ARGUMENT;
11222 				vm_object_unlock(curr_object);
11223 				break;
11224 			}
11225 			vm_map_offset_t start_offset = (curr_addr - curr_entry->vme_start) + VME_OFFSET(curr_entry);
11226 			vm_map_offset_t end_offset = MIN(curr_entry->vme_end, end_addr) -
11227 			    (curr_entry->vme_start + VME_OFFSET(curr_entry));
11228 			vm_map_offset_t curr_offset = start_offset;
11229 			vm_page_t curr_page;
11230 			while (curr_offset < end_offset) {
11231 				curr_page = vm_page_lookup(curr_object, vm_object_trunc_page(curr_offset));
11232 				if (curr_page != VM_PAGE_NULL) {
11233 					vm_page_lock_queues();
11234 					if (curr_page->vmp_laundry) {
11235 						vm_pageout_steal_laundry(curr_page, TRUE);
11236 					}
11237 					/*
11238 					 * we've already factored out pages in the laundry which
11239 					 * means this page can't be on the pageout queue so it's
11240 					 * safe to do the vm_page_queues_remove
11241 					 */
11242 					bool donate = (curr_page->vmp_on_specialq == VM_PAGE_SPECIAL_Q_DONATE);
11243 					vm_page_queues_remove(curr_page, TRUE);
11244 					if (donate) {
11245 						/*
11246 						 * The compressor needs to see this bit to know
11247 						 * where this page needs to land. Also if stolen,
11248 						 * this bit helps put the page back in the right
11249 						 * special queue where it belongs.
11250 						 */
11251 						curr_page->vmp_on_specialq = VM_PAGE_SPECIAL_Q_DONATE;
11252 					}
11253 					// Clear the referenced bit so we ensure this gets paged out
11254 					curr_page->vmp_reference = false;
11255 					if (curr_page->vmp_pmapped) {
11256 						pmap_clear_refmod_options(VM_PAGE_GET_PHYS_PAGE(curr_page),
11257 						    VM_MEM_REFERENCED, PMAP_OPTIONS_NOFLUSH, (void*)NULL);
11258 					}
11259 					vm_page_queue_enter(queue, curr_page, vmp_pageq);
11260 					vm_page_unlock_queues();
11261 					*pages_moved += 1;
11262 				}
11263 				curr_offset += PAGE_SIZE_64;
11264 				curr_addr += PAGE_SIZE_64;
11265 			}
11266 		}
11267 		vm_object_unlock(curr_object);
11268 	}
11269 	vm_map_unlock_read(map);
11270 	return err;
11271 }
11272 
11273 /*
11274  * Local queue for processing benchmark pages.
11275  * Can't be allocated on the stack because the pointer has to
11276  * be packable.
11277  */
11278 vm_page_queue_head_t compressor_perf_test_queue VM_PAGE_PACKED_ALIGNED;
11279 kern_return_t
run_compressor_perf_test(user_addr_t buf,size_t buffer_size,uint64_t * time,uint64_t * bytes_compressed,uint64_t * compressor_growth)11280 run_compressor_perf_test(
11281 	user_addr_t buf,
11282 	size_t buffer_size,
11283 	uint64_t *time,
11284 	uint64_t *bytes_compressed,
11285 	uint64_t *compressor_growth)
11286 {
11287 	kern_return_t err = KERN_SUCCESS;
11288 	if (!VM_CONFIG_COMPRESSOR_IS_ACTIVE) {
11289 		return KERN_NOT_SUPPORTED;
11290 	}
11291 	if (current_task() == kernel_task) {
11292 		return KERN_INVALID_ARGUMENT;
11293 	}
11294 	vm_page_lock_queues();
11295 	if (compressor_running_perf_test) {
11296 		/* Only run one instance of the benchmark at a time. */
11297 		vm_page_unlock_queues();
11298 		return KERN_RESOURCE_SHORTAGE;
11299 	}
11300 	vm_page_unlock_queues();
11301 	size_t page_count = 0;
11302 	vm_map_t map;
11303 	vm_page_t p, next;
11304 	uint64_t compressor_perf_test_start = 0, compressor_perf_test_end = 0;
11305 	uint64_t compressed_bytes_start = 0, compressed_bytes_end = 0;
11306 	*bytes_compressed = *compressor_growth = 0;
11307 
11308 	vm_page_queue_init(&compressor_perf_test_queue);
11309 	map = current_task()->map;
11310 	err = move_pages_to_queue(map, buf, buffer_size, &compressor_perf_test_queue, &page_count);
11311 	if (err != KERN_SUCCESS) {
11312 		goto out;
11313 	}
11314 
11315 	vm_page_lock_queues();
11316 	compressor_running_perf_test = true;
11317 	compressor_perf_test_pages_processed = 0;
11318 	/*
11319 	 * At this point the compressor threads should only process the benchmark queue
11320 	 * so we can look at the difference in c_segment_compressed_bytes while the perf test is running
11321 	 * to determine how many compressed bytes we ended up using.
11322 	 */
11323 	compressed_bytes_start = c_segment_compressed_bytes;
11324 	vm_page_unlock_queues();
11325 
11326 	page_count = vm_pageout_page_queue(&compressor_perf_test_queue, page_count, true);
11327 
11328 	vm_page_lock_queues();
11329 	compressor_perf_test_start = mach_absolute_time();
11330 
11331 	// Wake up the compressor thread(s)
11332 	sched_cond_signal(&pgo_iothread_internal_state[0].pgo_wakeup,
11333 	    pgo_iothread_internal_state[0].pgo_iothread);
11334 
11335 	/*
11336 	 * Depending on when this test is run we could overshoot or be right on the mark
11337 	 * with our page_count. So the comparison is of the _less than_ variety.
11338 	 */
11339 	while (compressor_perf_test_pages_processed < page_count) {
11340 		assert_wait((event_t) &compressor_perf_test_pages_processed, THREAD_UNINT);
11341 		vm_page_unlock_queues();
11342 		thread_block(THREAD_CONTINUE_NULL);
11343 		vm_page_lock_queues();
11344 	}
11345 	compressor_perf_test_end = mach_absolute_time();
11346 	compressed_bytes_end = c_segment_compressed_bytes;
11347 	vm_page_unlock_queues();
11348 
11349 
11350 out:
11351 	/*
11352 	 * If we errored out above, then we could still have some pages
11353 	 * on the local queue. Make sure to put them back on the active queue before
11354 	 * returning so they're not orphaned.
11355 	 */
11356 	vm_page_lock_queues();
11357 	absolutetime_to_nanoseconds(compressor_perf_test_end - compressor_perf_test_start, time);
11358 	p = (vm_page_t) vm_page_queue_first(&compressor_perf_test_queue);
11359 	while (p && !vm_page_queue_end(&compressor_perf_test_queue, (vm_page_queue_entry_t)p)) {
11360 		next = (vm_page_t)VM_PAGE_UNPACK_PTR(p->vmp_pageq.next);
11361 
11362 		vm_page_enqueue_active(p, FALSE);
11363 		p = next;
11364 	}
11365 
11366 	compressor_running_perf_test = false;
11367 	vm_page_unlock_queues();
11368 	if (err == KERN_SUCCESS) {
11369 		*bytes_compressed = page_count * PAGE_SIZE_64;
11370 		*compressor_growth = compressed_bytes_end - compressed_bytes_start;
11371 	}
11372 
11373 	/*
11374 	 * pageout_scan will consider waking the compactor swapper
11375 	 * before it blocks. Do the same thing here before we return
11376 	 * to ensure that back to back benchmark runs can't overly fragment the
11377 	 * compressor pool.
11378 	 */
11379 	vm_consider_waking_compactor_swapper();
11380 	return err;
11381 }
11382 #endif /* DEVELOPMENT || DEBUG */
11383