xref: /xnu-10002.61.3/osfmk/kern/machine.c (revision 0f4c859e951fba394238ab619495c4e1d54d0f34)
1 /*
2  * Copyright (c) 2000-2019 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:	kern/machine.c
60  *	Author:	Avadis Tevanian, Jr.
61  *	Date:	1987
62  *
63  *	Support for machine independent machine abstraction.
64  */
65 
66 #include <string.h>
67 
68 #include <mach/mach_types.h>
69 #include <mach/boolean.h>
70 #include <mach/kern_return.h>
71 #include <mach/machine.h>
72 #include <mach/host_info.h>
73 #include <mach/host_reboot.h>
74 #include <mach/host_priv_server.h>
75 #include <mach/processor_server.h>
76 #include <mach/sdt.h>
77 
78 #include <kern/kern_types.h>
79 #include <kern/cpu_data.h>
80 #include <kern/ipc_host.h>
81 #include <kern/host.h>
82 #include <kern/machine.h>
83 #include <kern/misc_protos.h>
84 #include <kern/percpu.h>
85 #include <kern/processor.h>
86 #include <kern/queue.h>
87 #include <kern/sched.h>
88 #include <kern/startup.h>
89 #include <kern/task.h>
90 #include <kern/thread.h>
91 #include <kern/iotrace.h>
92 
93 #include <libkern/OSDebug.h>
94 #if ML_IO_TIMEOUTS_ENABLED
95 #include <libkern/tree.h>
96 #endif
97 
98 #include <pexpert/device_tree.h>
99 
100 #include <machine/commpage.h>
101 #include <machine/machine_routines.h>
102 
103 #if HIBERNATION
104 #include <IOKit/IOHibernatePrivate.h>
105 #endif
106 #include <IOKit/IOPlatformExpert.h>
107 
108 #if CONFIG_DTRACE
109 extern void (*dtrace_cpu_state_changed_hook)(int, boolean_t);
110 #endif
111 
112 #if defined(__arm64__)
113 extern void wait_while_mp_kdp_trap(bool check_SIGPdebug);
114 #include <arm/pmap/pmap_data.h>
115 #endif
116 
117 #if defined(__x86_64__)
118 #include <i386/panic_notify.h>
119 #endif
120 
121 /*
122  *	Exported variables:
123  */
124 
125 struct machine_info     machine_info;
126 
127 /* Forwards */
128 static void
129 processor_doshutdown(processor_t processor);
130 
131 static void
132 processor_offline(void * parameter, __unused wait_result_t result);
133 
134 static void
135 processor_offline_intstack(processor_t processor) __dead2;
136 
137 static void
processor_up_update_counts(processor_t processor)138 processor_up_update_counts(processor_t processor)
139 {
140 	ml_cpu_up_update_counts(processor->cpu_id);
141 
142 	os_atomic_inc(&processor_avail_count, relaxed);
143 	if (processor->is_recommended) {
144 		os_atomic_inc(&processor_avail_count_user, relaxed);
145 	}
146 	if (processor->processor_primary == processor) {
147 		os_atomic_inc(&primary_processor_avail_count, relaxed);
148 		if (processor->is_recommended) {
149 			os_atomic_inc(&primary_processor_avail_count_user, relaxed);
150 		}
151 	}
152 	commpage_update_active_cpus();
153 }
154 
155 /*
156  *	processor_up:
157  *
158  *	Flag processor as up and running, and available
159  *	for scheduling.
160  */
161 void
processor_up(processor_t processor)162 processor_up(
163 	processor_t                     processor)
164 {
165 	processor_set_t         pset;
166 	spl_t                           s;
167 
168 	s = splsched();
169 	init_ast_check(processor);
170 
171 #if defined(__arm64__)
172 	/*
173 	 * A processor coming online won't have received a SIGPdebug signal
174 	 * to cause it to spin while a stackshot or panic is taking place,
175 	 * so spin here on mp_kdp_trap.
176 	 *
177 	 * However, since cpu_signal() is not yet enabled for this processor,
178 	 * there is a race if we have just passed this when a cpu_signal()
179 	 * is attempted.  The sender will assume the cpu is offline, so it will
180 	 * not end up spinning anywhere.  See processor_offline() for the fix
181 	 * for this race.
182 	 */
183 	wait_while_mp_kdp_trap(false);
184 #endif
185 
186 	pset = processor->processor_set;
187 	simple_lock(&sched_available_cores_lock, LCK_GRP_NULL);
188 	pset_lock(pset);
189 
190 	++pset->online_processor_count;
191 	simple_lock(&processor->start_state_lock, LCK_GRP_NULL);
192 	pset_update_processor_state(pset, processor, PROCESSOR_RUNNING);
193 	simple_unlock(&processor->start_state_lock);
194 	bool temporary = processor->shutdown_temporary;
195 	if (temporary) {
196 		processor->shutdown_temporary = false;
197 	} else {
198 		processor_up_update_counts(processor);
199 	}
200 	if (processor->is_recommended) {
201 		SCHED(pset_made_schedulable)(processor, pset, false);
202 	}
203 	pset_unlock(pset);
204 	ml_cpu_up();
205 	smr_cpu_up(processor, SMR_CPU_REASON_OFFLINE);
206 	sched_mark_processor_online_locked(processor, processor->last_startup_reason);
207 	simple_unlock(&sched_available_cores_lock);
208 	splx(s);
209 
210 	thread_wakeup((event_t)&processor->state);
211 
212 #if CONFIG_DTRACE
213 	if (dtrace_cpu_state_changed_hook) {
214 		(*dtrace_cpu_state_changed_hook)(processor->cpu_id, TRUE);
215 	}
216 #endif
217 }
218 #include <atm/atm_internal.h>
219 
220 kern_return_t
host_reboot(host_priv_t host_priv,int options)221 host_reboot(
222 	host_priv_t             host_priv,
223 	int                             options)
224 {
225 	if (host_priv == HOST_PRIV_NULL) {
226 		return KERN_INVALID_HOST;
227 	}
228 
229 #if DEVELOPMENT || DEBUG
230 	if (options & HOST_REBOOT_DEBUGGER) {
231 		Debugger("Debugger");
232 		return KERN_SUCCESS;
233 	}
234 #endif
235 
236 	if (options & HOST_REBOOT_UPSDELAY) {
237 		// UPS power cutoff path
238 		PEHaltRestart( kPEUPSDelayHaltCPU );
239 	} else {
240 		halt_all_cpus(!(options & HOST_REBOOT_HALT));
241 	}
242 
243 	return KERN_SUCCESS;
244 }
245 
246 kern_return_t
processor_assign(__unused processor_t processor,__unused processor_set_t new_pset,__unused boolean_t wait)247 processor_assign(
248 	__unused processor_t            processor,
249 	__unused processor_set_t        new_pset,
250 	__unused boolean_t              wait)
251 {
252 	return KERN_FAILURE;
253 }
254 
255 static void
processor_down_update_counts(processor_t processor)256 processor_down_update_counts(processor_t processor)
257 {
258 	ml_cpu_down_update_counts(processor->cpu_id);
259 
260 	os_atomic_dec(&processor_avail_count, relaxed);
261 	if (processor->is_recommended) {
262 		os_atomic_dec(&processor_avail_count_user, relaxed);
263 	}
264 	if (processor->processor_primary == processor) {
265 		os_atomic_dec(&primary_processor_avail_count, relaxed);
266 		if (processor->is_recommended) {
267 			os_atomic_dec(&primary_processor_avail_count_user, relaxed);
268 		}
269 	}
270 	commpage_update_active_cpus();
271 }
272 
273 extern lck_mtx_t processor_updown_lock;
274 
275 kern_return_t
processor_shutdown(processor_t processor,processor_reason_t reason,uint32_t flags)276 processor_shutdown(
277 	processor_t                     processor,
278 	processor_reason_t              reason,
279 	uint32_t                        flags)
280 {
281 	if (!ml_cpu_can_exit(processor->cpu_id, reason)) {
282 		/*
283 		 * Failure if disallowed by arch code.
284 		 */
285 		return KERN_NOT_SUPPORTED;
286 	}
287 
288 	lck_mtx_lock(&processor_updown_lock);
289 
290 	spl_t s = splsched();
291 	processor_set_t pset = processor->processor_set;
292 
293 	pset_lock(pset);
294 
295 	if (processor->state == PROCESSOR_START) {
296 		pset_unlock(pset);
297 		splx(s);
298 
299 		processor_wait_for_start(processor);
300 
301 		s = splsched();
302 		pset_lock(pset);
303 	}
304 
305 	/*
306 	 * If the processor is dispatching, let it finish.
307 	 */
308 	while (processor->state == PROCESSOR_DISPATCHING) {
309 		pset_unlock(pset);
310 		splx(s);
311 		delay(1);
312 		s = splsched();
313 		pset_lock(pset);
314 	}
315 	pset_unlock(pset);
316 	splx(s);
317 
318 	kern_return_t mark_ret = sched_mark_processor_offline(processor, reason);
319 	if (mark_ret != KERN_SUCCESS) {
320 		/* Must fail or we deadlock */
321 		lck_mtx_unlock(&processor_updown_lock);
322 		return KERN_FAILURE;
323 	}
324 
325 	ml_cpu_begin_state_transition(processor->cpu_id);
326 	s = splsched();
327 
328 	pset_lock(pset);
329 	if (processor->state == PROCESSOR_OFF_LINE) {
330 		/*
331 		 * Success if already shutdown.
332 		 */
333 		if (processor->shutdown_temporary && !(flags & SHUTDOWN_TEMPORARY)) {
334 			/* Convert a temporary shutdown into a permanent shutdown */
335 			processor->shutdown_temporary = false;
336 			processor_down_update_counts(processor);
337 		}
338 		pset_unlock(pset);
339 		splx(s);
340 		ml_cpu_end_state_transition(processor->cpu_id);
341 
342 		lck_mtx_unlock(&processor_updown_lock);
343 		return KERN_SUCCESS;
344 	}
345 
346 	if (processor->shutdown_locked && (reason != REASON_SYSTEM)) {
347 		/*
348 		 * Failure if processor is locked against shutdown.
349 		 */
350 		pset_unlock(pset);
351 		splx(s);
352 
353 		lck_mtx_unlock(&processor_updown_lock);
354 		return KERN_FAILURE;
355 	}
356 
357 	/*
358 	 * If the processor is dispatching, let it finish.
359 	 */
360 	while (processor->state == PROCESSOR_DISPATCHING) {
361 		pset_unlock(pset);
362 		splx(s);
363 		delay(1);
364 		s = splsched();
365 		pset_lock(pset);
366 	}
367 
368 	/*
369 	 * Success if already being shutdown with matching SHUTDOWN_TEMPORARY flag.
370 	 */
371 	if ((processor->state == PROCESSOR_SHUTDOWN) || (processor->state == PROCESSOR_PENDING_OFFLINE)) {
372 		bool success = (flags & SHUTDOWN_TEMPORARY) ? processor->shutdown_temporary : !processor->shutdown_temporary;
373 
374 		pset_unlock(pset);
375 		splx(s);
376 		ml_cpu_end_state_transition(processor->cpu_id);
377 
378 		lck_mtx_unlock(&processor_updown_lock);
379 		return success ? KERN_SUCCESS : KERN_FAILURE;
380 	}
381 
382 	ml_broadcast_cpu_event(CPU_EXIT_REQUESTED, processor->cpu_id);
383 	pset_update_processor_state(pset, processor, PROCESSOR_SHUTDOWN);
384 	processor->last_shutdown_reason = reason;
385 	if (flags & SHUTDOWN_TEMPORARY) {
386 		processor->shutdown_temporary = true;
387 	}
388 	pset_unlock(pset);
389 
390 	processor_doshutdown(processor);
391 	splx(s);
392 
393 	cpu_exit_wait(processor->cpu_id);
394 
395 	if (processor != master_processor) {
396 		s = splsched();
397 		pset_lock(pset);
398 		pset_update_processor_state(pset, processor, PROCESSOR_OFF_LINE);
399 		pset_unlock(pset);
400 		splx(s);
401 	}
402 
403 	ml_cpu_end_state_transition(processor->cpu_id);
404 	ml_broadcast_cpu_event(CPU_EXITED, processor->cpu_id);
405 	ml_cpu_power_disable(processor->cpu_id);
406 
407 	lck_mtx_unlock(&processor_updown_lock);
408 	return KERN_SUCCESS;
409 }
410 
411 /*
412  * Called with interrupts disabled.
413  */
414 static void
processor_doshutdown(processor_t processor)415 processor_doshutdown(
416 	processor_t processor)
417 {
418 	thread_t self = current_thread();
419 
420 	/*
421 	 *	Get onto the processor to shutdown
422 	 */
423 	processor_t prev = thread_bind(processor);
424 	thread_block(THREAD_CONTINUE_NULL);
425 
426 	/* interrupts still disabled */
427 	assert(ml_get_interrupts_enabled() == FALSE);
428 
429 	assert(processor == current_processor());
430 	assert(processor->state == PROCESSOR_SHUTDOWN);
431 
432 #if CONFIG_DTRACE
433 	if (dtrace_cpu_state_changed_hook) {
434 		(*dtrace_cpu_state_changed_hook)(processor->cpu_id, FALSE);
435 	}
436 #endif
437 
438 #if defined(__arm64__)
439 	/*
440 	 * Catch a processor going offline
441 	 * while a panic or stackshot is in progress, as it won't
442 	 * receive a SIGPdebug now that interrupts are disabled.
443 	 */
444 	wait_while_mp_kdp_trap(false);
445 #endif
446 
447 	smr_cpu_down(processor, SMR_CPU_REASON_OFFLINE);
448 	ml_cpu_down();
449 
450 #if HIBERNATION
451 	if (processor_avail_count < 2) {
452 		hibernate_vm_lock();
453 		hibernate_vm_unlock();
454 	}
455 #endif
456 
457 	processor_set_t pset = processor->processor_set;
458 
459 	pset_lock(pset);
460 	pset_update_processor_state(pset, processor, PROCESSOR_PENDING_OFFLINE);
461 	--pset->online_processor_count;
462 	if (!processor->shutdown_temporary) {
463 		processor_down_update_counts(processor);
464 	}
465 	SCHED(processor_queue_shutdown)(processor);
466 	/* pset lock dropped */
467 	SCHED(rt_queue_shutdown)(processor);
468 
469 	thread_bind(prev);
470 
471 	/* interrupts still disabled */
472 
473 	/*
474 	 * Continue processor shutdown on the processor's idle thread.
475 	 * The handoff won't fail because the idle thread has a reserved stack.
476 	 * Switching to the idle thread leaves interrupts disabled,
477 	 * so we can't accidentally take an interrupt after the context switch.
478 	 */
479 	thread_t shutdown_thread = processor->idle_thread;
480 	shutdown_thread->continuation = processor_offline;
481 	shutdown_thread->parameter = processor;
482 
483 	thread_run(self, NULL, NULL, shutdown_thread);
484 }
485 
486 /*
487  * Called in the context of the idle thread to shut down the processor
488  *
489  * A shut-down processor looks like it's 'running' the idle thread parked
490  * in this routine, but it's actually been powered off and has no hardware state.
491  */
492 static void
processor_offline(void * parameter,__unused wait_result_t result)493 processor_offline(
494 	void * parameter,
495 	__unused wait_result_t result)
496 {
497 	processor_t processor = (processor_t) parameter;
498 	thread_t self = current_thread();
499 	__assert_only thread_t old_thread = THREAD_NULL;
500 
501 	assert(processor == current_processor());
502 	assert(self->state & TH_IDLE);
503 	assert(processor->idle_thread == self);
504 	assert(ml_get_interrupts_enabled() == FALSE);
505 	assert(self->continuation == NULL);
506 	assert(processor->processor_offlined == false);
507 	assert(processor->running_timers_active == false);
508 
509 	bool enforce_quiesce_safety = gEnforcePlatformActionSafety;
510 
511 	/*
512 	 * Scheduling is now disabled for this processor.
513 	 * Ensure that primitives that need scheduling (like mutexes) know this.
514 	 */
515 	if (enforce_quiesce_safety) {
516 		disable_preemption_without_measurements();
517 	}
518 
519 	/* convince slave_main to come back here */
520 	processor->processor_offlined = true;
521 
522 	/*
523 	 * Switch to the interrupt stack and shut down the processor.
524 	 *
525 	 * When the processor comes back, it will eventually call load_context which
526 	 * restores the context saved by machine_processor_shutdown, returning here.
527 	 */
528 	old_thread = machine_processor_shutdown(self, processor_offline_intstack, processor);
529 
530 	/* old_thread should be NULL because we got here through Load_context */
531 	assert(old_thread == THREAD_NULL);
532 
533 	assert(processor == current_processor());
534 	assert(processor->idle_thread == current_thread());
535 
536 	assert(ml_get_interrupts_enabled() == FALSE);
537 	assert(self->continuation == NULL);
538 
539 	/* Extract the machine_param value stashed by slave_main */
540 	void * machine_param = self->parameter;
541 	self->parameter = NULL;
542 
543 	/* Re-initialize the processor */
544 	slave_machine_init(machine_param);
545 
546 	assert(processor->processor_offlined == true);
547 	processor->processor_offlined = false;
548 
549 	if (enforce_quiesce_safety) {
550 		enable_preemption();
551 	}
552 
553 #if defined(__arm64__)
554 	/*
555 	 * See the comments for DebuggerLock in processor_up().
556 	 *
557 	 * SIGPdisabled is cleared (to enable cpu_signal() to succeed with this processor)
558 	 * the first time we take an IPI.  This is triggered by slave_machine_init(), above,
559 	 * which calls cpu_machine_init()->PE_cpu_machine_init()->PE_cpu_signal() which sends
560 	 * a self-IPI to ensure that happens when we enable interrupts.  So enable interrupts
561 	 * here so that cpu_signal() can succeed before we spin on mp_kdp_trap.
562 	 */
563 	ml_set_interrupts_enabled(TRUE);
564 
565 	ml_set_interrupts_enabled(FALSE);
566 
567 	wait_while_mp_kdp_trap(true);
568 
569 	/*
570 	 * At this point,
571 	 * if a stackshot or panic is in progress, we either spin on mp_kdp_trap
572 	 * or we sucessfully received a SIGPdebug signal which will cause us to
573 	 * break out of the spin on mp_kdp_trap and instead
574 	 * spin next time interrupts are enabled in idle_thread().
575 	 */
576 #endif
577 
578 	/*
579 	 * Now that the processor is back, invoke the idle thread to find out what to do next.
580 	 * idle_thread will enable interrupts.
581 	 */
582 	thread_block(idle_thread);
583 	/*NOTREACHED*/
584 }
585 
586 /*
587  * Complete the shutdown and place the processor offline.
588  *
589  * Called at splsched in the shutdown context
590  * (i.e. on the idle thread, on the interrupt stack)
591  *
592  * The onlining half of this is done in load_context().
593  */
594 static void
processor_offline_intstack(processor_t processor)595 processor_offline_intstack(
596 	processor_t processor)
597 {
598 	assert(processor == current_processor());
599 	assert(processor->active_thread == current_thread());
600 
601 	struct recount_snap snap = { 0 };
602 	recount_snapshot(&snap);
603 	recount_processor_idle(&processor->pr_recount, &snap);
604 
605 	smr_cpu_leave(processor, processor->last_dispatch);
606 
607 	PMAP_DEACTIVATE_KERNEL(processor->cpu_id);
608 
609 	cpu_sleep();
610 	panic("zombie processor");
611 	/*NOTREACHED*/
612 }
613 
614 kern_return_t
host_get_boot_info(host_priv_t host_priv,kernel_boot_info_t boot_info)615 host_get_boot_info(
616 	host_priv_t         host_priv,
617 	kernel_boot_info_t  boot_info)
618 {
619 	const char *src = "";
620 	if (host_priv == HOST_PRIV_NULL) {
621 		return KERN_INVALID_HOST;
622 	}
623 
624 	/*
625 	 * Copy first operator string terminated by '\0' followed by
626 	 *	standardized strings generated from boot string.
627 	 */
628 	src = machine_boot_info(boot_info, KERNEL_BOOT_INFO_MAX);
629 	if (src != boot_info) {
630 		(void) strncpy(boot_info, src, KERNEL_BOOT_INFO_MAX);
631 	}
632 
633 	return KERN_SUCCESS;
634 }
635 
636 // These are configured through sysctls.
637 #if DEVELOPMENT || DEBUG
638 uint32_t phy_read_panic = 1;
639 uint32_t phy_write_panic = 1;
640 uint64_t simulate_stretched_io = 0;
641 #else
642 uint32_t phy_read_panic = 0;
643 uint32_t phy_write_panic = 0;
644 #endif
645 
646 #if !defined(__x86_64__)
647 
648 #if DEVELOPMENT || DEBUG
649 static const uint64_t TIMEBASE_TICKS_PER_USEC = 24000000ULL / USEC_PER_SEC;
650 static const uint64_t DEFAULT_TRACE_PHY_TIMEOUT = 100 * TIMEBASE_TICKS_PER_USEC;
651 #else
652 static const uint64_t DEFAULT_TRACE_PHY_TIMEOUT = 0;
653 #endif
654 
655 // The MACHINE_TIMEOUT facility only exists on ARM.
656 MACHINE_TIMEOUT_DEV_WRITEABLE(report_phy_read_delay_to, "report-phy-read-delay", 0, MACHINE_TIMEOUT_UNIT_TIMEBASE, NULL);
657 MACHINE_TIMEOUT_DEV_WRITEABLE(report_phy_write_delay_to, "report-phy-write-delay", 0, MACHINE_TIMEOUT_UNIT_TIMEBASE, NULL);
658 MACHINE_TIMEOUT_DEV_WRITEABLE(trace_phy_read_delay_to, "trace-phy-read-delay", DEFAULT_TRACE_PHY_TIMEOUT, MACHINE_TIMEOUT_UNIT_TIMEBASE, NULL);
659 MACHINE_TIMEOUT_DEV_WRITEABLE(trace_phy_write_delay_to, "trace-phy-write-delay", DEFAULT_TRACE_PHY_TIMEOUT, MACHINE_TIMEOUT_UNIT_TIMEBASE, NULL);
660 
661 #if SCHED_HYGIENE_DEBUG
662 /*
663  * Note: The interrupt-masked timeout goes through two initializations - one
664  * early in boot and one later. Thus this function is also called twice and
665  * can't be marked '__startup_func'.
666  */
667 static void
ml_io_init_timeouts(void)668 ml_io_init_timeouts(void)
669 {
670 	/*
671 	 * The timeouts may be completely disabled via an override. Check that
672 	 * last and set the timeouts to zero (disabling) if that's the case.
673 	 */
674 	if (kern_feature_override(KF_IO_TIMEOUT_OVRD)) {
675 		os_atomic_store(&report_phy_write_delay_to, 0, relaxed);
676 		os_atomic_store(&report_phy_read_delay_to, 0, relaxed);
677 	}
678 }
679 
680 /*
681  * It's important that this happens after machine timeouts have initialized so
682  * the correct timeouts can be inherited.
683  */
684 STARTUP(TIMEOUTS, STARTUP_RANK_SECOND, ml_io_init_timeouts);
685 #endif /* SCHED_HYGIENE_DEBUG */
686 
687 extern pmap_paddr_t kvtophys(vm_offset_t va);
688 #endif
689 
690 #if ML_IO_TIMEOUTS_ENABLED
691 
692 static LCK_GRP_DECLARE(io_timeout_override_lock_grp, "io_timeout_override");
693 static LCK_SPIN_DECLARE(io_timeout_override_lock, &io_timeout_override_lock_grp);
694 
695 struct io_timeout_override_entry {
696 	RB_ENTRY(io_timeout_override_entry) tree;
697 
698 	uintptr_t iovaddr_base;
699 	unsigned int size;
700 	uint32_t read_timeout;
701 	uint32_t write_timeout;
702 };
703 
704 static inline int
io_timeout_override_cmp(const struct io_timeout_override_entry * a,const struct io_timeout_override_entry * b)705 io_timeout_override_cmp(const struct io_timeout_override_entry *a, const struct io_timeout_override_entry *b)
706 {
707 	if (a->iovaddr_base < b->iovaddr_base) {
708 		return -1;
709 	} else if (a->iovaddr_base > b->iovaddr_base) {
710 		return 1;
711 	} else {
712 		return 0;
713 	}
714 }
715 
716 static RB_HEAD(io_timeout_override, io_timeout_override_entry) io_timeout_override_root;
717 RB_PROTOTYPE_PREV(io_timeout_override, io_timeout_override_entry, tree, io_timeout_override_cmp);
718 RB_GENERATE_PREV(io_timeout_override, io_timeout_override_entry, tree, io_timeout_override_cmp);
719 
720 #endif /* ML_IO_TIMEOUTS_ENABLED */
721 
722 int
ml_io_increase_timeouts(uintptr_t iovaddr_base,unsigned int size,uint32_t read_timeout_us,uint32_t write_timeout_us)723 ml_io_increase_timeouts(uintptr_t iovaddr_base, unsigned int size, uint32_t read_timeout_us, uint32_t write_timeout_us)
724 {
725 #if ML_IO_TIMEOUTS_ENABLED
726 	const size_t MAX_SIZE = 4096;
727 	const uint64_t MAX_TIMEOUT_ABS = UINT32_MAX;
728 
729 	assert(preemption_enabled());
730 
731 	int ret = KERN_SUCCESS;
732 
733 	if (size == 0) {
734 		return KERN_INVALID_ARGUMENT;
735 	}
736 
737 	uintptr_t iovaddr_end;
738 	if (size > MAX_SIZE || os_add_overflow(iovaddr_base, size - 1, &iovaddr_end)) {
739 		return KERN_INVALID_ARGUMENT;
740 	}
741 
742 	uint64_t read_timeout_abs, write_timeout_abs;
743 	nanoseconds_to_absolutetime(NSEC_PER_USEC * read_timeout_us, &read_timeout_abs);
744 	nanoseconds_to_absolutetime(NSEC_PER_USEC * write_timeout_us, &write_timeout_abs);
745 	if (read_timeout_abs > MAX_TIMEOUT_ABS || write_timeout_abs > MAX_TIMEOUT_ABS) {
746 		return KERN_INVALID_ARGUMENT;
747 	}
748 
749 	struct io_timeout_override_entry *node = kalloc_type(struct io_timeout_override_entry, Z_WAITOK | Z_ZERO | Z_NOFAIL);
750 	node->iovaddr_base = iovaddr_base;
751 	node->size = size;
752 	node->read_timeout = (uint32_t)read_timeout_abs;
753 	node->write_timeout = (uint32_t)write_timeout_abs;
754 
755 	/*
756 	 * Interrupt handlers are allowed to call ml_io_{read,write}*, so
757 	 * interrupts must be disabled any time io_timeout_override_lock is
758 	 * held.  Otherwise the CPU could take an interrupt while holding the
759 	 * lock, invoke an ISR that calls ml_io_{read,write}*, and deadlock
760 	 * trying to acquire the lock again.
761 	 */
762 	boolean_t istate = ml_set_interrupts_enabled(FALSE);
763 	lck_spin_lock(&io_timeout_override_lock);
764 	if (RB_INSERT(io_timeout_override, &io_timeout_override_root, node)) {
765 		ret = KERN_INVALID_ARGUMENT;
766 		goto out;
767 	}
768 
769 	/* Check that this didn't create any new overlaps */
770 	struct io_timeout_override_entry *prev = RB_PREV(io_timeout_override, &io_timeout_override_root, node);
771 	if (prev && (prev->iovaddr_base + prev->size) > node->iovaddr_base) {
772 		RB_REMOVE(io_timeout_override, &io_timeout_override_root, node);
773 		ret = KERN_INVALID_ARGUMENT;
774 		goto out;
775 	}
776 	struct io_timeout_override_entry *next = RB_NEXT(io_timeout_override, &io_timeout_override_root, node);
777 	if (next && (node->iovaddr_base + node->size) > next->iovaddr_base) {
778 		RB_REMOVE(io_timeout_override, &io_timeout_override_root, node);
779 		ret = KERN_INVALID_ARGUMENT;
780 		goto out;
781 	}
782 
783 out:
784 	lck_spin_unlock(&io_timeout_override_lock);
785 	ml_set_interrupts_enabled(istate);
786 	if (ret != KERN_SUCCESS) {
787 		kfree_type(struct io_timeout_override_entry, node);
788 	}
789 	return ret;
790 #else /* !ML_IO_TIMEOUTS_ENABLED */
791 #pragma unused(iovaddr_base, size, read_timeout_us, write_timeout_us)
792 	return KERN_SUCCESS;
793 #endif
794 }
795 
796 int
ml_io_reset_timeouts(uintptr_t iovaddr_base,unsigned int size)797 ml_io_reset_timeouts(uintptr_t iovaddr_base, unsigned int size)
798 {
799 #if ML_IO_TIMEOUTS_ENABLED
800 	assert(preemption_enabled());
801 
802 	struct io_timeout_override_entry key = { .iovaddr_base = iovaddr_base };
803 
804 	boolean_t istate = ml_set_interrupts_enabled(FALSE);
805 	lck_spin_lock(&io_timeout_override_lock);
806 	struct io_timeout_override_entry *node = RB_FIND(io_timeout_override, &io_timeout_override_root, &key);
807 	if (node) {
808 		if (node->size == size) {
809 			RB_REMOVE(io_timeout_override, &io_timeout_override_root, node);
810 		} else {
811 			node = NULL;
812 		}
813 	}
814 	lck_spin_unlock(&io_timeout_override_lock);
815 	ml_set_interrupts_enabled(istate);
816 
817 	if (!node) {
818 		return KERN_NOT_FOUND;
819 	}
820 
821 	kfree_type(struct io_timeout_override_entry, node);
822 #else /* !ML_IO_TIMEOUTS_ENABLED */
823 #pragma unused(iovaddr_base, size)
824 #endif
825 	return KERN_SUCCESS;
826 }
827 
828 #if ML_IO_TIMEOUTS_ENABLED
829 
830 static bool
override_io_timeouts_va(uintptr_t vaddr,uint64_t * read_timeout,uint64_t * write_timeout)831 override_io_timeouts_va(uintptr_t vaddr, uint64_t *read_timeout, uint64_t *write_timeout)
832 {
833 	assert(!ml_get_interrupts_enabled());
834 
835 	struct io_timeout_override_entry *node = RB_ROOT(&io_timeout_override_root);
836 
837 	lck_spin_lock(&io_timeout_override_lock);
838 	/* RB_FIND() doesn't support custom cmp functions, so we have to open-code our own */
839 	while (node) {
840 		if (node->iovaddr_base <= vaddr && vaddr < node->iovaddr_base + node->size) {
841 			if (read_timeout) {
842 				*read_timeout = node->read_timeout;
843 			}
844 			if (write_timeout) {
845 				*write_timeout = node->write_timeout;
846 			}
847 			lck_spin_unlock(&io_timeout_override_lock);
848 			return true;
849 		} else if (vaddr < node->iovaddr_base) {
850 			node = RB_LEFT(node, tree);
851 		} else {
852 			node = RB_RIGHT(node, tree);
853 		}
854 	}
855 	lck_spin_unlock(&io_timeout_override_lock);
856 
857 	return false;
858 }
859 
860 static bool
override_io_timeouts_pa(uint64_t paddr,uint64_t * read_timeout,uint64_t * write_timeout)861 override_io_timeouts_pa(uint64_t paddr, uint64_t *read_timeout, uint64_t *write_timeout)
862 {
863 #if defined(__arm64__)
864 	/*
865 	 * PCIe regions are marked with PMAP_IO_RANGE_STRONG_SYNC. Apply a
866 	 * timeout greater than the PCIe completion timeout (50ms). In some
867 	 * cases those timeouts can stack so make the timeout significantly
868 	 * higher.
869 	 */
870 	#define STRONG_SYNC_TIMEOUT 1800000 /* 75ms */
871 
872 	pmap_io_range_t *range = pmap_find_io_attr(paddr);
873 	if (range != NULL && (range->wimg & PMAP_IO_RANGE_STRONG_SYNC) != 0) {
874 		if (read_timeout) {
875 			*read_timeout = STRONG_SYNC_TIMEOUT;
876 		}
877 		if (write_timeout) {
878 			*write_timeout = STRONG_SYNC_TIMEOUT;
879 		}
880 
881 		return true;
882 	}
883 #else
884 	(void)paddr;
885 	(void)read_timeout;
886 	(void)write_timeout;
887 #endif /* __arm64__ */
888 	return false;
889 }
890 
891 void
override_io_timeouts(uintptr_t vaddr,uint64_t paddr,uint64_t * read_timeout,uint64_t * write_timeout)892 override_io_timeouts(uintptr_t vaddr, uint64_t paddr, uint64_t *read_timeout, uint64_t *write_timeout)
893 {
894 	if (vaddr != 0 &&
895 	    override_io_timeouts_va(vaddr, read_timeout, write_timeout)) {
896 		return;
897 	}
898 
899 	if (paddr != 0 &&
900 	    override_io_timeouts_pa(paddr, read_timeout, write_timeout)) {
901 		return;
902 	}
903 }
904 #endif /* ML_IO_TIMEOUTS_ENABLED */
905 
906 unsigned long long
ml_io_read(uintptr_t vaddr,int size)907 ml_io_read(uintptr_t vaddr, int size)
908 {
909 	unsigned long long result = 0;
910 	unsigned char s1;
911 	unsigned short s2;
912 
913 #ifdef ML_IO_VERIFY_UNCACHEABLE
914 	uintptr_t const paddr = pmap_verify_noncacheable(vaddr);
915 #elif defined(ML_IO_TIMEOUTS_ENABLED)
916 	uintptr_t const paddr = kvtophys(vaddr);
917 #endif
918 
919 #ifdef ML_IO_TIMEOUTS_ENABLED
920 	uint64_t sabs, eabs;
921 	boolean_t istate, timeread = FALSE;
922 	uint64_t report_read_delay;
923 #if __x86_64__
924 	report_read_delay = report_phy_read_delay;
925 #else
926 	report_read_delay = os_atomic_load(&report_phy_read_delay_to, relaxed);
927 	uint64_t const trace_phy_read_delay = os_atomic_load(&trace_phy_read_delay_to, relaxed);
928 #endif /* __x86_64__ */
929 
930 	if (__improbable(report_read_delay != 0)) {
931 		istate = ml_set_interrupts_enabled(FALSE);
932 		sabs = mach_absolute_time();
933 		timeread = TRUE;
934 	}
935 
936 #ifdef ML_IO_SIMULATE_STRETCHED_ENABLED
937 	if (__improbable(timeread && simulate_stretched_io)) {
938 		sabs -= simulate_stretched_io;
939 	}
940 #endif /* ML_IO_SIMULATE_STRETCHED_ENABLED */
941 #endif /* ML_IO_TIMEOUTS_ENABLED */
942 
943 #if DEVELOPMENT || DEBUG
944 	boolean_t use_fences = !kern_feature_override(KF_IO_TIMEOUT_OVRD);
945 	if (use_fences) {
946 		ml_timebase_to_memory_fence();
947 	}
948 #endif
949 
950 	switch (size) {
951 	case 1:
952 		s1 = *(volatile unsigned char *)vaddr;
953 		result = s1;
954 		break;
955 	case 2:
956 		s2 = *(volatile unsigned short *)vaddr;
957 		result = s2;
958 		break;
959 	case 4:
960 		result = *(volatile unsigned int *)vaddr;
961 		break;
962 	case 8:
963 		result = *(volatile unsigned long long *)vaddr;
964 		break;
965 	default:
966 		panic("Invalid size %d for ml_io_read(%p)", size, (void *)vaddr);
967 		break;
968 	}
969 
970 #if DEVELOPMENT || DEBUG
971 	if (use_fences) {
972 		ml_memory_to_timebase_fence();
973 	}
974 #endif
975 
976 #ifdef ML_IO_TIMEOUTS_ENABLED
977 	if (__improbable(timeread == TRUE)) {
978 		eabs = mach_absolute_time();
979 
980 		/* Prevent the processor from calling iotrace during its
981 		 * initialization procedure. */
982 		if (current_processor()->state == PROCESSOR_RUNNING) {
983 			iotrace(IOTRACE_IO_READ, vaddr, paddr, size, result, sabs, eabs - sabs);
984 		}
985 
986 		if (__improbable((eabs - sabs) > report_read_delay)) {
987 			DTRACE_PHYSLAT5(physioread, uint64_t, (eabs - sabs),
988 			    uint64_t, vaddr, uint32_t, size, uint64_t, paddr, uint64_t, result);
989 
990 			uint64_t override = 0;
991 			override_io_timeouts(vaddr, paddr, &override, NULL);
992 
993 			if (override != 0) {
994 #if SCHED_HYGIENE_DEBUG
995 				/*
996 				 * The IO timeout was overridden. As interrupts are disabled in
997 				 * order to accurately measure IO time this can cause the
998 				 * interrupt masked timeout threshold to be exceeded.  If the
999 				 * interrupt masked debug mode is set to panic, abandon the
1000 				 * measurement. If in trace mode leave it as-is for
1001 				 * observability.
1002 				 */
1003 				if (interrupt_masked_debug_mode == SCHED_HYGIENE_MODE_PANIC) {
1004 					ml_spin_debug_clear(current_thread());
1005 					ml_irq_debug_abandon();
1006 				}
1007 #endif
1008 				report_read_delay = override;
1009 			}
1010 		}
1011 
1012 		if (__improbable((eabs - sabs) > report_read_delay)) {
1013 			if (phy_read_panic && (machine_timeout_suspended() == FALSE)) {
1014 #if defined(__x86_64__)
1015 				panic_notify();
1016 #endif /* defined(__x86_64__) */
1017 				uint64_t nsec = 0;
1018 				absolutetime_to_nanoseconds(eabs - sabs, &nsec);
1019 				panic("Read from IO vaddr 0x%lx paddr 0x%lx took %llu ns, "
1020 				    "result: 0x%llx (start: %llu, end: %llu), ceiling: %llu",
1021 				    vaddr, paddr, nsec, result, sabs, eabs,
1022 				    report_read_delay);
1023 			}
1024 		}
1025 
1026 		if (__improbable(trace_phy_read_delay > 0 && (eabs - sabs) > trace_phy_read_delay)) {
1027 			KDBG(MACHDBG_CODE(DBG_MACH_IO, DBC_MACH_IO_MMIO_READ),
1028 			    (eabs - sabs), VM_KERNEL_UNSLIDE_OR_PERM(vaddr), paddr, result);
1029 		}
1030 
1031 		(void)ml_set_interrupts_enabled(istate);
1032 	}
1033 #endif /*  ML_IO_TIMEOUTS_ENABLED */
1034 	return result;
1035 }
1036 
1037 unsigned int
ml_io_read8(uintptr_t vaddr)1038 ml_io_read8(uintptr_t vaddr)
1039 {
1040 	return (unsigned) ml_io_read(vaddr, 1);
1041 }
1042 
1043 unsigned int
ml_io_read16(uintptr_t vaddr)1044 ml_io_read16(uintptr_t vaddr)
1045 {
1046 	return (unsigned) ml_io_read(vaddr, 2);
1047 }
1048 
1049 unsigned int
ml_io_read32(uintptr_t vaddr)1050 ml_io_read32(uintptr_t vaddr)
1051 {
1052 	return (unsigned) ml_io_read(vaddr, 4);
1053 }
1054 
1055 unsigned long long
ml_io_read64(uintptr_t vaddr)1056 ml_io_read64(uintptr_t vaddr)
1057 {
1058 	return ml_io_read(vaddr, 8);
1059 }
1060 
1061 /* ml_io_write* */
1062 
1063 void
ml_io_write(uintptr_t vaddr,uint64_t val,int size)1064 ml_io_write(uintptr_t vaddr, uint64_t val, int size)
1065 {
1066 #ifdef ML_IO_VERIFY_UNCACHEABLE
1067 	uintptr_t const paddr = pmap_verify_noncacheable(vaddr);
1068 #elif defined(ML_IO_TIMEOUTS_ENABLED)
1069 	uintptr_t const paddr = kvtophys(vaddr);
1070 #endif
1071 
1072 #ifdef ML_IO_TIMEOUTS_ENABLED
1073 	uint64_t sabs, eabs;
1074 	boolean_t istate, timewrite = FALSE;
1075 	uint64_t report_write_delay;
1076 #if __x86_64__
1077 	report_write_delay = report_phy_write_delay;
1078 #else
1079 	report_write_delay = os_atomic_load(&report_phy_write_delay_to, relaxed);
1080 	uint64_t trace_phy_write_delay = os_atomic_load(&trace_phy_write_delay_to, relaxed);
1081 #endif /* !defined(__x86_64__) */
1082 	if (__improbable(report_write_delay != 0)) {
1083 		istate = ml_set_interrupts_enabled(FALSE);
1084 		sabs = mach_absolute_time();
1085 		timewrite = TRUE;
1086 	}
1087 
1088 #ifdef ML_IO_SIMULATE_STRETCHED_ENABLED
1089 	if (__improbable(timewrite && simulate_stretched_io)) {
1090 		sabs -= simulate_stretched_io;
1091 	}
1092 #endif /* DEVELOPMENT || DEBUG */
1093 #endif /* ML_IO_TIMEOUTS_ENABLED */
1094 
1095 #if DEVELOPMENT || DEBUG
1096 	boolean_t use_fences = !kern_feature_override(KF_IO_TIMEOUT_OVRD);
1097 	if (use_fences) {
1098 		ml_timebase_to_memory_fence();
1099 	}
1100 #endif
1101 
1102 	switch (size) {
1103 	case 1:
1104 		*(volatile uint8_t *)vaddr = (uint8_t)val;
1105 		break;
1106 	case 2:
1107 		*(volatile uint16_t *)vaddr = (uint16_t)val;
1108 		break;
1109 	case 4:
1110 		*(volatile uint32_t *)vaddr = (uint32_t)val;
1111 		break;
1112 	case 8:
1113 		*(volatile uint64_t *)vaddr = (uint64_t)val;
1114 		break;
1115 	default:
1116 		panic("Invalid size %d for ml_io_write(%p, 0x%llx)", size, (void *)vaddr, val);
1117 		break;
1118 	}
1119 
1120 #if DEVELOPMENT || DEBUG
1121 	if (use_fences) {
1122 		ml_memory_to_timebase_fence();
1123 	}
1124 #endif
1125 
1126 #ifdef ML_IO_TIMEOUTS_ENABLED
1127 	if (__improbable(timewrite == TRUE)) {
1128 		eabs = mach_absolute_time();
1129 
1130 
1131 		/* Prevent the processor from calling iotrace during its
1132 		 * initialization procedure. */
1133 		if (current_processor()->state == PROCESSOR_RUNNING) {
1134 			iotrace(IOTRACE_IO_WRITE, vaddr, paddr, size, val, sabs, eabs - sabs);
1135 		}
1136 
1137 
1138 		if (__improbable((eabs - sabs) > report_write_delay)) {
1139 			DTRACE_PHYSLAT5(physiowrite, uint64_t, (eabs - sabs),
1140 			    uint64_t, vaddr, uint32_t, size, uint64_t, paddr, uint64_t, val);
1141 
1142 			uint64_t override = 0;
1143 			override_io_timeouts(vaddr, paddr, NULL, &override);
1144 
1145 			if (override != 0) {
1146 #if SCHED_HYGIENE_DEBUG
1147 				/*
1148 				 * The IO timeout was overridden. As interrupts are disabled in
1149 				 * order to accurately measure IO time this can cause the
1150 				 * interrupt masked timeout threshold to be exceeded.  If the
1151 				 * interrupt masked debug mode is set to panic, abandon the
1152 				 * measurement. If in trace mode leave it as-is for
1153 				 * observability.
1154 				 */
1155 				if (interrupt_masked_debug_mode == SCHED_HYGIENE_MODE_PANIC) {
1156 					ml_spin_debug_clear(current_thread());
1157 					ml_irq_debug_abandon();
1158 				}
1159 #endif
1160 				report_write_delay = override;
1161 			}
1162 		}
1163 
1164 		if (__improbable((eabs - sabs) > report_write_delay)) {
1165 			if (phy_write_panic && (machine_timeout_suspended() == FALSE)) {
1166 #if defined(__x86_64__)
1167 				panic_notify();
1168 #endif /*  defined(__x86_64__) */
1169 
1170 				uint64_t nsec = 0;
1171 				absolutetime_to_nanoseconds(eabs - sabs, &nsec);
1172 				panic("Write to IO vaddr %p paddr %p val 0x%llx took %llu ns,"
1173 				    " (start: %llu, end: %llu), ceiling: %llu",
1174 				    (void *)vaddr, (void *)paddr, val, nsec, sabs, eabs,
1175 				    report_write_delay);
1176 			}
1177 		}
1178 
1179 		if (__improbable(trace_phy_write_delay > 0 && (eabs - sabs) > trace_phy_write_delay)) {
1180 			KDBG(MACHDBG_CODE(DBG_MACH_IO, DBC_MACH_IO_MMIO_WRITE),
1181 			    (eabs - sabs), VM_KERNEL_UNSLIDE_OR_PERM(vaddr), paddr, val);
1182 		}
1183 
1184 		(void)ml_set_interrupts_enabled(istate);
1185 	}
1186 #endif /* ML_IO_TIMEOUTS_ENABLED */
1187 }
1188 
1189 void
ml_io_write8(uintptr_t vaddr,uint8_t val)1190 ml_io_write8(uintptr_t vaddr, uint8_t val)
1191 {
1192 	ml_io_write(vaddr, val, 1);
1193 }
1194 
1195 void
ml_io_write16(uintptr_t vaddr,uint16_t val)1196 ml_io_write16(uintptr_t vaddr, uint16_t val)
1197 {
1198 	ml_io_write(vaddr, val, 2);
1199 }
1200 
1201 void
ml_io_write32(uintptr_t vaddr,uint32_t val)1202 ml_io_write32(uintptr_t vaddr, uint32_t val)
1203 {
1204 	ml_io_write(vaddr, val, 4);
1205 }
1206 
1207 void
ml_io_write64(uintptr_t vaddr,uint64_t val)1208 ml_io_write64(uintptr_t vaddr, uint64_t val)
1209 {
1210 	ml_io_write(vaddr, val, 8);
1211 }
1212 
1213 struct cpu_callback_chain_elem {
1214 	cpu_callback_t                  fn;
1215 	void                            *param;
1216 	struct cpu_callback_chain_elem  *next;
1217 };
1218 
1219 static struct cpu_callback_chain_elem *cpu_callback_chain;
1220 static LCK_GRP_DECLARE(cpu_callback_chain_lock_grp, "cpu_callback_chain");
1221 static LCK_SPIN_DECLARE(cpu_callback_chain_lock, &cpu_callback_chain_lock_grp);
1222 
1223 void
cpu_event_register_callback(cpu_callback_t fn,void * param)1224 cpu_event_register_callback(cpu_callback_t fn, void *param)
1225 {
1226 	struct cpu_callback_chain_elem *new_elem;
1227 
1228 	new_elem = zalloc_permanent_type(struct cpu_callback_chain_elem);
1229 	if (!new_elem) {
1230 		panic("can't allocate cpu_callback_chain_elem");
1231 	}
1232 
1233 	lck_spin_lock(&cpu_callback_chain_lock);
1234 	new_elem->next = cpu_callback_chain;
1235 	new_elem->fn = fn;
1236 	new_elem->param = param;
1237 	os_atomic_store(&cpu_callback_chain, new_elem, release);
1238 	lck_spin_unlock(&cpu_callback_chain_lock);
1239 }
1240 
1241 __attribute__((noreturn))
1242 void
cpu_event_unregister_callback(__unused cpu_callback_t fn)1243 cpu_event_unregister_callback(__unused cpu_callback_t fn)
1244 {
1245 	panic("Unfortunately, cpu_event_unregister_callback is unimplemented.");
1246 }
1247 
1248 void
ml_broadcast_cpu_event(enum cpu_event event,unsigned int cpu_or_cluster)1249 ml_broadcast_cpu_event(enum cpu_event event, unsigned int cpu_or_cluster)
1250 {
1251 	struct cpu_callback_chain_elem *cursor;
1252 
1253 	cursor = os_atomic_load(&cpu_callback_chain, dependency);
1254 	for (; cursor != NULL; cursor = cursor->next) {
1255 		cursor->fn(cursor->param, event, cpu_or_cluster);
1256 	}
1257 }
1258 
1259 // Initialize Machine Timeouts (see the MACHINE_TIMEOUT macro
1260 // definition)
1261 
1262 void
machine_timeout_init_with_suffix(const struct machine_timeout_spec * spec,char const * suffix)1263 machine_timeout_init_with_suffix(const struct machine_timeout_spec *spec, char const *suffix)
1264 {
1265 	if (spec->skip_predicate != NULL && spec->skip_predicate(spec)) {
1266 		// This timeout should be disabled.
1267 		os_atomic_store_wide((uint64_t*)spec->ptr, 0, relaxed);
1268 		return;
1269 	}
1270 
1271 	assert(suffix != NULL);
1272 	assert(strlen(spec->name) <= MACHINE_TIMEOUT_MAX_NAME_LEN);
1273 
1274 	size_t const suffix_len = strlen(suffix);
1275 
1276 	size_t const dt_name_size = MACHINE_TIMEOUT_MAX_NAME_LEN + suffix_len + 1;
1277 	char dt_name[dt_name_size];
1278 
1279 	strlcpy(dt_name, spec->name, dt_name_size);
1280 	strlcat(dt_name, suffix, dt_name_size);
1281 
1282 	size_t const scale_name_size = MACHINE_TIMEOUT_MAX_NAME_LEN + suffix_len + strlen("-scale") + 1;
1283 	char scale_name[scale_name_size];
1284 
1285 	strlcpy(scale_name, spec->name, scale_name_size);
1286 	strlcat(scale_name, suffix, scale_name_size);
1287 	strlcat(scale_name, "-scale", scale_name_size);
1288 
1289 	size_t const boot_arg_name_size = MACHINE_TIMEOUT_MAX_NAME_LEN + strlen("ml-timeout-") + suffix_len + 1;
1290 	char boot_arg_name[boot_arg_name_size];
1291 
1292 	strlcpy(boot_arg_name, "ml-timeout-", boot_arg_name_size);
1293 	strlcat(boot_arg_name, spec->name, boot_arg_name_size);
1294 	strlcat(boot_arg_name, suffix, boot_arg_name_size);
1295 
1296 	size_t const boot_arg_scale_name_size = MACHINE_TIMEOUT_MAX_NAME_LEN +
1297 	    strlen("ml-timeout-") + strlen("-scale") + suffix_len + 1;
1298 	char boot_arg_scale_name[boot_arg_scale_name_size];
1299 
1300 	strlcpy(boot_arg_scale_name, "ml-timeout-", boot_arg_scale_name_size);
1301 	strlcat(boot_arg_scale_name, spec->name, boot_arg_scale_name_size);
1302 	strlcat(boot_arg_scale_name, suffix, boot_arg_name_size);
1303 	strlcat(boot_arg_scale_name, "-scale", boot_arg_scale_name_size);
1304 
1305 
1306 	/*
1307 	 * Determine base value from DT and boot-args.
1308 	 */
1309 
1310 	DTEntry base, chosen;
1311 
1312 	if (SecureDTLookupEntry(NULL, "/machine-timeouts", &base) != kSuccess) {
1313 		base = NULL;
1314 	}
1315 
1316 	if (SecureDTLookupEntry(NULL, "/chosen/machine-timeouts", &chosen) != kSuccess) {
1317 		chosen = NULL;
1318 	}
1319 
1320 	uint64_t timeout = spec->default_value;
1321 	bool found = false;
1322 
1323 	uint64_t const *data = NULL;
1324 	unsigned int data_size = sizeof(*data);
1325 
1326 	/* First look in /machine-timeouts/<name> */
1327 	if (base != NULL && SecureDTGetProperty(base, dt_name, (const void **)&data, &data_size) == kSuccess) {
1328 		if (data_size != sizeof(*data)) {
1329 			panic("%s: unexpected machine timeout data_size %u for /machine-timeouts/%s", __func__, data_size, dt_name);
1330 		}
1331 
1332 		timeout = *data;
1333 		found = true;
1334 	}
1335 
1336 	/* A value in /chosen/machine-timeouts/<name> overrides */
1337 	if (chosen != NULL && SecureDTGetProperty(chosen, dt_name, (const void **)&data, &data_size) == kSuccess) {
1338 		if (data_size != sizeof(*data)) {
1339 			panic("%s: unexpected machine timeout data_size %u for /chosen/machine-timeouts/%s", __func__, data_size, dt_name);
1340 		}
1341 
1342 		timeout = *data;
1343 		found = true;
1344 	}
1345 
1346 	/* A boot-arg ml-timeout-<name> overrides */
1347 	uint64_t boot_arg = 0;
1348 
1349 	if (PE_parse_boot_argn(boot_arg_name, &boot_arg, sizeof(boot_arg))) {
1350 		timeout = boot_arg;
1351 		found = true;
1352 	}
1353 
1354 
1355 	/*
1356 	 * Determine scale value from DT and boot-args.
1357 	 */
1358 
1359 	uint64_t scale = 1;
1360 	uint32_t const *scale_data;
1361 	unsigned int scale_size = sizeof(scale_data);
1362 
1363 	/* If there is a scale factor /machine-timeouts/<name>-scale, apply it. */
1364 	if (base != NULL && SecureDTGetProperty(base, scale_name, (const void **)&scale_data, &scale_size) == kSuccess) {
1365 		if (scale_size != sizeof(*scale_data)) {
1366 			panic("%s: unexpected machine timeout data_size %u for /machine-timeouts/%s-scale", __func__, scale_size, dt_name);
1367 		}
1368 
1369 		scale = *scale_data;
1370 	}
1371 
1372 	/* If there is a scale factor /chosen/machine-timeouts/<name>-scale, use that. */
1373 	if (chosen != NULL && SecureDTGetProperty(chosen, scale_name, (const void **)&scale_data, &scale_size) == kSuccess) {
1374 		if (scale_size != sizeof(*scale_data)) {
1375 			panic("%s: unexpected machine timeout data_size %u for /chosen/machine-timeouts/%s-scale", __func__,
1376 			    scale_size, dt_name);
1377 		}
1378 
1379 		scale = *scale_data;
1380 	}
1381 
1382 	/* Finally, a boot-arg ml-timeout-<name>-scale takes precedence. */
1383 	if (PE_parse_boot_argn(boot_arg_scale_name, &boot_arg, sizeof(boot_arg))) {
1384 		scale = boot_arg;
1385 	}
1386 
1387 	static bool global_scale_set;
1388 	static uint64_t global_scale;
1389 
1390 	if (!global_scale_set) {
1391 		/* Apply /machine-timeouts/global-scale if present */
1392 		if (SecureDTGetProperty(base, "global-scale", (const void **)&scale_data, &scale_size) == kSuccess) {
1393 			if (scale_size != sizeof(*scale_data)) {
1394 				panic("%s: unexpected machine timeout data_size %u for /machine-timeouts/global-scale", __func__,
1395 				    scale_size);
1396 			}
1397 
1398 			global_scale = *scale_data;
1399 			global_scale_set = true;
1400 		}
1401 
1402 		/* Use /chosen/machine-timeouts/global-scale if present */
1403 		if (SecureDTGetProperty(chosen, "global-scale", (const void **)&scale_data, &scale_size) == kSuccess) {
1404 			if (scale_size != sizeof(*scale_data)) {
1405 				panic("%s: unexpected machine timeout data_size %u for /chosen/machine-timeouts/global-scale", __func__,
1406 				    scale_size);
1407 			}
1408 
1409 			global_scale = *scale_data;
1410 			global_scale_set = true;
1411 		}
1412 
1413 		/* Finally, the boot-arg ml-timeout-global-scale takes precedence. */
1414 		if (PE_parse_boot_argn("ml-timeout-global-scale", &boot_arg, sizeof(boot_arg))) {
1415 			global_scale = boot_arg;
1416 			global_scale_set = true;
1417 		}
1418 	}
1419 
1420 	if (global_scale_set) {
1421 		scale *= global_scale;
1422 	}
1423 
1424 	/* Compute the final timeout, and done. */
1425 	if (found && timeout > 0) {
1426 		/* Only apply inherent unit scale if the value came in
1427 		 * externally. */
1428 
1429 		if (spec->unit_scale == MACHINE_TIMEOUT_UNIT_TIMEBASE) {
1430 			uint64_t nanoseconds = timeout / 1000;
1431 			nanoseconds_to_absolutetime(nanoseconds, &timeout);
1432 		} else {
1433 			timeout /= spec->unit_scale;
1434 		}
1435 
1436 		if (timeout == 0) {
1437 			/* Ensure unit scaling did not disable the timeout. */
1438 			timeout = 1;
1439 		}
1440 	}
1441 
1442 	if (os_mul_overflow(timeout, scale, &timeout)) {
1443 		timeout = UINT64_MAX; // clamp
1444 	}
1445 
1446 	os_atomic_store_wide((uint64_t*)spec->ptr, timeout, relaxed);
1447 }
1448 
1449 void
machine_timeout_init(const struct machine_timeout_spec * spec)1450 machine_timeout_init(const struct machine_timeout_spec *spec)
1451 {
1452 	machine_timeout_init_with_suffix(spec, "");
1453 }
1454 
1455 #if DEVELOPMENT || DEBUG
1456 /*
1457  * Late timeout (re-)initialization, at the end of bsd_init()
1458  */
1459 void
machine_timeout_bsd_init(void)1460 machine_timeout_bsd_init(void)
1461 {
1462 	char const * const __unused mt_suffix = "-b";
1463 #if SCHED_HYGIENE_DEBUG
1464 	machine_timeout_init_with_suffix(MACHINE_TIMEOUT_SPEC_REF(interrupt_masked_timeout), mt_suffix);
1465 	machine_timeout_init_with_suffix(MACHINE_TIMEOUT_SPEC_REF(sched_preemption_disable_threshold_mt), mt_suffix);
1466 
1467 	/*
1468 	 * The io timeouts can inherit from interrupt_masked_timeout.
1469 	 * Re-initialize, as interrupt_masked_timeout may have changed.
1470 	 */
1471 	ml_io_init_timeouts();
1472 
1473 	extern void preemption_disable_reset_max_durations(void);
1474 	/*
1475 	 * Reset the preemption disable stats, so that they are not
1476 	 * polluted by long early boot code.
1477 	 */
1478 	preemption_disable_reset_max_durations();
1479 #endif /* SCHED_HYGIENE_DEBUG */
1480 }
1481 #endif /* DEVELOPMENT || DEBUG */
1482 
1483 #if ML_IO_TIMEOUTS_ENABLED && CONFIG_XNUPOST
1484 #include <tests/xnupost.h>
1485 
1486 extern kern_return_t ml_io_timeout_test(void);
1487 
1488 static inline void
ml_io_timeout_test_get_timeouts(uintptr_t vaddr,uint64_t * read_timeout,uint64_t * write_timeout)1489 ml_io_timeout_test_get_timeouts(uintptr_t vaddr, uint64_t *read_timeout, uint64_t *write_timeout)
1490 {
1491 	*read_timeout = 0;
1492 	*write_timeout = 0;
1493 
1494 	boolean_t istate = ml_set_interrupts_enabled(FALSE);
1495 	override_io_timeouts(vaddr, 0, read_timeout, write_timeout);
1496 	ml_set_interrupts_enabled(istate);
1497 }
1498 
1499 kern_return_t
ml_io_timeout_test(void)1500 ml_io_timeout_test(void)
1501 {
1502 	const size_t SIZE = 16;
1503 	uintptr_t iovaddr_base1 = (uintptr_t)&ml_io_timeout_test;
1504 	uintptr_t iovaddr_base2 = iovaddr_base1 + SIZE;
1505 	uintptr_t vaddr1 = iovaddr_base1 + SIZE / 2;
1506 	uintptr_t vaddr2 = iovaddr_base2 + SIZE / 2;
1507 
1508 	const uint64_t READ_TIMEOUT1_US = 50000, WRITE_TIMEOUT1_US = 50001;
1509 	const uint64_t READ_TIMEOUT2_US = 50002, WRITE_TIMEOUT2_US = 50003;
1510 	uint64_t read_timeout1_abs, write_timeout1_abs;
1511 	uint64_t read_timeout2_abs, write_timeout2_abs;
1512 	nanoseconds_to_absolutetime(NSEC_PER_USEC * READ_TIMEOUT1_US, &read_timeout1_abs);
1513 	nanoseconds_to_absolutetime(NSEC_PER_USEC * WRITE_TIMEOUT1_US, &write_timeout1_abs);
1514 	nanoseconds_to_absolutetime(NSEC_PER_USEC * READ_TIMEOUT2_US, &read_timeout2_abs);
1515 	nanoseconds_to_absolutetime(NSEC_PER_USEC * WRITE_TIMEOUT2_US, &write_timeout2_abs);
1516 
1517 	int err = ml_io_increase_timeouts(iovaddr_base1, 0, READ_TIMEOUT1_US, WRITE_TIMEOUT1_US);
1518 	T_EXPECT_EQ_INT(err, KERN_INVALID_ARGUMENT, "Can't set timeout for empty region");
1519 
1520 	err = ml_io_increase_timeouts(iovaddr_base1, 4097, READ_TIMEOUT1_US, WRITE_TIMEOUT1_US);
1521 	T_EXPECT_EQ_INT(err, KERN_INVALID_ARGUMENT, "Can't set timeout for region > 4096 bytes");
1522 
1523 	err = ml_io_increase_timeouts(UINTPTR_MAX, SIZE, READ_TIMEOUT1_US, WRITE_TIMEOUT1_US);
1524 	T_EXPECT_EQ_INT(err, KERN_INVALID_ARGUMENT, "Can't set timeout for overflowed region");
1525 
1526 	err = ml_io_increase_timeouts(iovaddr_base1, SIZE, READ_TIMEOUT1_US, WRITE_TIMEOUT1_US);
1527 	T_EXPECT_EQ_INT(err, KERN_SUCCESS, "Setting timeout for first VA region should succeed");
1528 
1529 	err = ml_io_increase_timeouts(iovaddr_base2, SIZE, READ_TIMEOUT2_US, WRITE_TIMEOUT2_US);
1530 	T_EXPECT_EQ_INT(err, KERN_SUCCESS, "Setting timeout for second VA region should succeed");
1531 
1532 	err = ml_io_increase_timeouts(iovaddr_base1, SIZE, READ_TIMEOUT1_US, WRITE_TIMEOUT1_US);
1533 	T_EXPECT_EQ_INT(err, KERN_INVALID_ARGUMENT, "Can't set timeout for same region twice");
1534 
1535 	err = ml_io_increase_timeouts(vaddr1, (uint32_t)(vaddr2 - vaddr1), READ_TIMEOUT1_US, WRITE_TIMEOUT1_US);
1536 	T_EXPECT_EQ_INT(err, KERN_INVALID_ARGUMENT, "Can't set timeout for overlapping regions");
1537 
1538 	uint64_t read_timeout, write_timeout;
1539 	ml_io_timeout_test_get_timeouts(vaddr1, &read_timeout, &write_timeout);
1540 	T_EXPECT_EQ_ULLONG(read_timeout, read_timeout1_abs, "Read timeout for first region");
1541 	T_EXPECT_EQ_ULLONG(write_timeout, write_timeout1_abs, "Write timeout for first region");
1542 
1543 	ml_io_timeout_test_get_timeouts(vaddr2, &read_timeout, &write_timeout);
1544 	T_EXPECT_EQ_ULLONG(read_timeout, read_timeout2_abs, "Read timeout for first region");
1545 	T_EXPECT_EQ_ULLONG(write_timeout, write_timeout2_abs, "Write timeout for first region");
1546 
1547 	ml_io_timeout_test_get_timeouts(iovaddr_base2 + SIZE, &read_timeout, &write_timeout);
1548 	T_EXPECT_EQ_ULLONG(read_timeout, 0, "Read timeout without override");
1549 	T_EXPECT_EQ_ULLONG(write_timeout, 0, "Write timeout without override");
1550 
1551 	err = ml_io_reset_timeouts(iovaddr_base1 + 1, SIZE - 1);
1552 	T_EXPECT_EQ_INT(err, KERN_NOT_FOUND, "Can't reset timeout for subregion");
1553 
1554 	err = ml_io_reset_timeouts(iovaddr_base2 + SIZE, SIZE);
1555 	T_EXPECT_EQ_INT(err, KERN_NOT_FOUND, "Can't reset timeout for non-existent region");
1556 
1557 	err = ml_io_reset_timeouts(iovaddr_base1, SIZE);
1558 	T_EXPECT_EQ_INT(err, KERN_SUCCESS, "Resetting timeout for first VA region should succeed");
1559 
1560 	ml_io_timeout_test_get_timeouts(vaddr1, &read_timeout, &write_timeout);
1561 	T_EXPECT_EQ_ULLONG(read_timeout, 0, "Read timeout for reset region");
1562 	T_EXPECT_EQ_ULLONG(write_timeout, 0, "Write timeout for reset region");
1563 
1564 	err = ml_io_reset_timeouts(iovaddr_base1, SIZE);
1565 	T_EXPECT_EQ_INT(err, KERN_NOT_FOUND, "Can't reset timeout for same region twice");
1566 
1567 	err = ml_io_reset_timeouts(iovaddr_base2, SIZE);
1568 	T_EXPECT_EQ_INT(err, KERN_SUCCESS, "Resetting timeout for second VA region should succeed");
1569 
1570 	return KERN_SUCCESS;
1571 }
1572 #endif /* CONFIG_XNUPOST */
1573