1 /*
2 * Copyright (c) 1993-2008 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 * Timer interrupt callout module.
30 */
31
32 #include <mach/mach_types.h>
33
34 #include <kern/clock.h>
35 #include <kern/counter.h>
36 #include <kern/smp.h>
37 #include <kern/processor.h>
38 #include <kern/timer_call.h>
39 #include <kern/timer_queue.h>
40 #include <kern/thread.h>
41 #include <kern/thread_group.h>
42 #include <kern/policy_internal.h>
43
44 #include <sys/kdebug.h>
45
46 #if CONFIG_DTRACE
47 #include <mach/sdt.h>
48 #endif
49
50
51 #if DEBUG
52 #define TIMER_ASSERT 1
53 #endif
54
55 //#define TIMER_ASSERT 1
56 //#define TIMER_DBG 1
57
58 #if TIMER_DBG
59 #define DBG(x...) kprintf("DBG: " x);
60 #else
61 #define DBG(x...)
62 #endif
63
64 #if TIMER_TRACE
65 #define TIMER_KDEBUG_TRACE KERNEL_DEBUG_CONSTANT_IST
66 #else
67 #define TIMER_KDEBUG_TRACE(x...)
68 #endif
69
70 LCK_GRP_DECLARE(timer_call_lck_grp, "timer_call");
71 LCK_GRP_DECLARE(timer_longterm_lck_grp, "timer_longterm");
72 LCK_GRP_DECLARE(timer_queue_lck_grp, "timer_queue");
73
74 /* Timer queue lock must be acquired with interrupts disabled (under splclock()) */
75 #define timer_queue_lock_spin(queue) lck_ticket_lock(&(queue)->lock_data, &timer_queue_lck_grp)
76 #define timer_queue_unlock(queue) lck_ticket_unlock(&(queue)->lock_data)
77
78 /*
79 * The longterm timer object is a global structure holding all timers
80 * beyond the short-term, local timer queue threshold. The boot processor
81 * is responsible for moving each timer to its local timer queue
82 * if and when that timer becomes due within the threshold.
83 */
84
85 /* Sentinel for "no time set": */
86 #define TIMER_LONGTERM_NONE EndOfAllTime
87 /* The default threadhold is the delta above which a timer is "long-term" */
88 #if defined(__x86_64__)
89 #define TIMER_LONGTERM_THRESHOLD (1ULL * NSEC_PER_SEC) /* 1 sec */
90 #else
91 #define TIMER_LONGTERM_THRESHOLD TIMER_LONGTERM_NONE /* disabled */
92 #endif
93
94 /*
95 * The scan_limit throttles processing of the longterm queue.
96 * If the scan time exceeds this limit, we terminate, unlock
97 * and defer for scan_interval. This prevents unbounded holding of
98 * timer queue locks with interrupts masked.
99 */
100 #define TIMER_LONGTERM_SCAN_LIMIT (100ULL * NSEC_PER_USEC) /* 100 us */
101 #define TIMER_LONGTERM_SCAN_INTERVAL (100ULL * NSEC_PER_USEC) /* 100 us */
102 /* Sentinel for "scan limit exceeded": */
103 #define TIMER_LONGTERM_SCAN_AGAIN 0
104
105 /*
106 * In a similar way to the longterm queue's scan limit, the following bounds the
107 * amount of time spent processing regular timers.
108 */
109 TUNABLE_WRITEABLE(uint64_t, timer_scan_limit_us, "timer_scan_limit_us", 400);
110 TUNABLE_WRITEABLE(uint64_t, timer_scan_interval_us, "timer_scan_interval_us", 40);
111 static uint64_t timer_scan_limit_abs = 0;
112 static uint64_t timer_scan_interval_abs = 0;
113
114 /*
115 * Count of times scanning the queue was aborted early (to avoid long
116 * scan times).
117 */
118 SCALABLE_COUNTER_DEFINE(timer_scan_pauses_cnt);
119
120 /*
121 * Count of times scanning the queue was aborted early resulting in a
122 * postponed hard deadline.
123 */
124 SCALABLE_COUNTER_DEFINE(timer_scan_postpones_cnt);
125
126 #define MAX_TIMER_SCAN_LIMIT (30000ULL * NSEC_PER_USEC) /* 30 ms */
127 #define MIN_TIMER_SCAN_LIMIT ( 50ULL * NSEC_PER_USEC) /* 50 us */
128 #define MAX_TIMER_SCAN_INTERVAL ( 2000ULL * NSEC_PER_USEC) /* 2 ms */
129 #define MIN_TIMER_SCAN_INTERVAL ( 20ULL * NSEC_PER_USEC) /* 20 us */
130
131 typedef struct {
132 uint64_t interval; /* longterm timer interval */
133 uint64_t margin; /* fudge factor (10% of interval */
134 uint64_t deadline; /* first/soonest longterm deadline */
135 uint64_t preempted; /* sooner timer has pre-empted */
136 timer_call_t call; /* first/soonest longterm timer call */
137 uint64_t deadline_set; /* next timer set */
138 timer_call_data_t timer; /* timer used by threshold management */
139 /* Stats: */
140 uint64_t scans; /* num threshold timer scans */
141 uint64_t preempts; /* num threshold reductions */
142 uint64_t latency; /* average threshold latency */
143 uint64_t latency_min; /* minimum threshold latency */
144 uint64_t latency_max; /* maximum threshold latency */
145 } threshold_t;
146
147 typedef struct {
148 mpqueue_head_t queue; /* longterm timer list */
149 uint64_t enqueues; /* num timers queued */
150 uint64_t dequeues; /* num timers dequeued */
151 uint64_t escalates; /* num timers becoming shortterm */
152 uint64_t scan_time; /* last time the list was scanned */
153 threshold_t threshold; /* longterm timer threshold */
154 uint64_t scan_limit; /* maximum scan time */
155 uint64_t scan_interval; /* interval between LT "escalation" scans */
156 uint64_t scan_pauses; /* num scans exceeding time limit */
157 } timer_longterm_t;
158
159 timer_longterm_t timer_longterm = {
160 .scan_limit = TIMER_LONGTERM_SCAN_LIMIT,
161 .scan_interval = TIMER_LONGTERM_SCAN_INTERVAL,
162 };
163
164 static mpqueue_head_t *timer_longterm_queue = NULL;
165
166 static void timer_longterm_init(void);
167 static void timer_longterm_callout(
168 timer_call_param_t p0,
169 timer_call_param_t p1);
170 extern void timer_longterm_scan(
171 timer_longterm_t *tlp,
172 uint64_t now);
173 static void timer_longterm_update(
174 timer_longterm_t *tlp);
175 static void timer_longterm_update_locked(
176 timer_longterm_t *tlp);
177 static mpqueue_head_t * timer_longterm_enqueue_unlocked(
178 timer_call_t call,
179 uint64_t now,
180 uint64_t deadline,
181 mpqueue_head_t ** old_queue,
182 uint64_t soft_deadline,
183 uint64_t ttd,
184 timer_call_param_t param1,
185 uint32_t callout_flags);
186 static void timer_longterm_dequeued_locked(
187 timer_call_t call);
188
189 uint64_t past_deadline_timers;
190 uint64_t past_deadline_deltas;
191 uint64_t past_deadline_longest;
192 uint64_t past_deadline_shortest = ~0ULL;
193 enum {PAST_DEADLINE_TIMER_ADJUSTMENT_NS = 10 * 1000};
194
195 uint64_t past_deadline_timer_adjustment;
196
197 static boolean_t timer_call_enter_internal(timer_call_t call, timer_call_param_t param1, uint64_t deadline, uint64_t leeway, uint32_t flags, boolean_t ratelimited);
198 boolean_t mach_timer_coalescing_enabled = TRUE;
199
200 mpqueue_head_t *timer_call_enqueue_deadline_unlocked(
201 timer_call_t call,
202 mpqueue_head_t *queue,
203 uint64_t deadline,
204 uint64_t soft_deadline,
205 uint64_t ttd,
206 timer_call_param_t param1,
207 uint32_t flags);
208
209 mpqueue_head_t *timer_call_dequeue_unlocked(
210 timer_call_t call);
211
212 timer_coalescing_priority_params_t tcoal_prio_params;
213
214 #if TCOAL_PRIO_STATS
215 int32_t nc_tcl, rt_tcl, bg_tcl, kt_tcl, fp_tcl, ts_tcl, qos_tcl;
216 #define TCOAL_PRIO_STAT(x) (x++)
217 #else
218 #define TCOAL_PRIO_STAT(x)
219 #endif
220
221 static void
timer_call_init_abstime(void)222 timer_call_init_abstime(void)
223 {
224 int i;
225 uint64_t result;
226 timer_coalescing_priority_params_ns_t * tcoal_prio_params_init = timer_call_get_priority_params();
227 nanoseconds_to_absolutetime(PAST_DEADLINE_TIMER_ADJUSTMENT_NS, &past_deadline_timer_adjustment);
228 nanoseconds_to_absolutetime(tcoal_prio_params_init->idle_entry_timer_processing_hdeadline_threshold_ns, &result);
229 tcoal_prio_params.idle_entry_timer_processing_hdeadline_threshold_abstime = (uint32_t)result;
230 nanoseconds_to_absolutetime(tcoal_prio_params_init->interrupt_timer_coalescing_ilat_threshold_ns, &result);
231 tcoal_prio_params.interrupt_timer_coalescing_ilat_threshold_abstime = (uint32_t)result;
232 nanoseconds_to_absolutetime(tcoal_prio_params_init->timer_resort_threshold_ns, &result);
233 tcoal_prio_params.timer_resort_threshold_abstime = (uint32_t)result;
234 tcoal_prio_params.timer_coalesce_rt_shift = tcoal_prio_params_init->timer_coalesce_rt_shift;
235 tcoal_prio_params.timer_coalesce_bg_shift = tcoal_prio_params_init->timer_coalesce_bg_shift;
236 tcoal_prio_params.timer_coalesce_kt_shift = tcoal_prio_params_init->timer_coalesce_kt_shift;
237 tcoal_prio_params.timer_coalesce_fp_shift = tcoal_prio_params_init->timer_coalesce_fp_shift;
238 tcoal_prio_params.timer_coalesce_ts_shift = tcoal_prio_params_init->timer_coalesce_ts_shift;
239
240 nanoseconds_to_absolutetime(tcoal_prio_params_init->timer_coalesce_rt_ns_max,
241 &tcoal_prio_params.timer_coalesce_rt_abstime_max);
242 nanoseconds_to_absolutetime(tcoal_prio_params_init->timer_coalesce_bg_ns_max,
243 &tcoal_prio_params.timer_coalesce_bg_abstime_max);
244 nanoseconds_to_absolutetime(tcoal_prio_params_init->timer_coalesce_kt_ns_max,
245 &tcoal_prio_params.timer_coalesce_kt_abstime_max);
246 nanoseconds_to_absolutetime(tcoal_prio_params_init->timer_coalesce_fp_ns_max,
247 &tcoal_prio_params.timer_coalesce_fp_abstime_max);
248 nanoseconds_to_absolutetime(tcoal_prio_params_init->timer_coalesce_ts_ns_max,
249 &tcoal_prio_params.timer_coalesce_ts_abstime_max);
250
251 for (i = 0; i < NUM_LATENCY_QOS_TIERS; i++) {
252 tcoal_prio_params.latency_qos_scale[i] = tcoal_prio_params_init->latency_qos_scale[i];
253 nanoseconds_to_absolutetime(tcoal_prio_params_init->latency_qos_ns_max[i],
254 &tcoal_prio_params.latency_qos_abstime_max[i]);
255 tcoal_prio_params.latency_tier_rate_limited[i] = tcoal_prio_params_init->latency_tier_rate_limited[i];
256 }
257
258 nanoseconds_to_absolutetime(timer_scan_limit_us * NSEC_PER_USEC, &timer_scan_limit_abs);
259 nanoseconds_to_absolutetime(timer_scan_interval_us * NSEC_PER_USEC, &timer_scan_interval_abs);
260 }
261
262
263 void
timer_call_init(void)264 timer_call_init(void)
265 {
266 timer_longterm_init();
267 timer_call_init_abstime();
268 }
269
270
271 void
timer_call_queue_init(mpqueue_head_t * queue)272 timer_call_queue_init(mpqueue_head_t *queue)
273 {
274 DBG("timer_call_queue_init(%p)\n", queue);
275 mpqueue_init(queue, &timer_call_lck_grp, LCK_ATTR_NULL);
276 }
277
278
279 void
timer_call_setup(timer_call_t call,timer_call_func_t func,timer_call_param_t param0)280 timer_call_setup(
281 timer_call_t call,
282 timer_call_func_t func,
283 timer_call_param_t param0)
284 {
285 DBG("timer_call_setup(%p,%p,%p)\n", call, func, param0);
286
287 *call = (struct timer_call) {
288 .tc_func = func,
289 .tc_param0 = param0,
290 .tc_async_dequeue = false,
291 };
292
293 simple_lock_init(&(call)->tc_lock, 0);
294 }
295
296 timer_call_t
timer_call_alloc(timer_call_func_t func,timer_call_param_t param0)297 timer_call_alloc(
298 timer_call_func_t func,
299 timer_call_param_t param0)
300 {
301 timer_call_t call;
302
303 call = kalloc_type(struct timer_call, Z_ZERO | Z_WAITOK | Z_NOFAIL);
304 timer_call_setup(call, func, param0);
305 return call;
306 }
307
308 void
timer_call_free(timer_call_t call)309 timer_call_free(
310 timer_call_t call)
311 {
312 kfree_type(struct timer_call, call);
313 }
314
315 static mpqueue_head_t*
mpqueue_for_timer_call(timer_call_t entry)316 mpqueue_for_timer_call(timer_call_t entry)
317 {
318 queue_t queue_entry_is_on = entry->tc_queue;
319 /* 'cast' the queue back to the orignal mpqueue */
320 return __container_of(queue_entry_is_on, struct mpqueue_head, head);
321 }
322
323
324 static __inline__ mpqueue_head_t *
timer_call_entry_dequeue(timer_call_t entry)325 timer_call_entry_dequeue(
326 timer_call_t entry)
327 {
328 mpqueue_head_t *old_mpqueue = mpqueue_for_timer_call(entry);
329
330 /* The entry was always on a queue */
331 assert(old_mpqueue != NULL);
332
333 #if TIMER_ASSERT
334 if (!hw_lock_held((hw_lock_t)&entry->tc_lock)) {
335 panic("_call_entry_dequeue() "
336 "entry %p is not locked\n", entry);
337 }
338
339 /*
340 * XXX The queue lock is actually a mutex in spin mode
341 * but there's no way to test for it being held
342 * so we pretend it's a spinlock!
343 */
344 if (!hw_lock_held((hw_lock_t)&old_mpqueue->lock_data)) {
345 panic("_call_entry_dequeue() "
346 "queue %p is not locked\n", old_mpqueue);
347 }
348 #endif /* TIMER_ASSERT */
349
350 if (old_mpqueue != timer_longterm_queue) {
351 priority_queue_remove(&old_mpqueue->mpq_pqhead,
352 &entry->tc_pqlink);
353 }
354
355 remqueue(&entry->tc_qlink);
356
357 entry->tc_queue = NULL;
358
359 old_mpqueue->count--;
360
361 return old_mpqueue;
362 }
363
364 static __inline__ mpqueue_head_t *
timer_call_entry_enqueue_deadline(timer_call_t entry,mpqueue_head_t * new_mpqueue,uint64_t deadline)365 timer_call_entry_enqueue_deadline(
366 timer_call_t entry,
367 mpqueue_head_t *new_mpqueue,
368 uint64_t deadline)
369 {
370 mpqueue_head_t *old_mpqueue = mpqueue_for_timer_call(entry);
371
372 #if TIMER_ASSERT
373 if (!hw_lock_held((hw_lock_t)&entry->tc_lock)) {
374 panic("_call_entry_enqueue_deadline() "
375 "entry %p is not locked\n", entry);
376 }
377
378 /* XXX More lock pretense: */
379 if (!hw_lock_held((hw_lock_t)&new_mpqueue->lock_data)) {
380 panic("_call_entry_enqueue_deadline() "
381 "queue %p is not locked\n", new_mpqueue);
382 }
383
384 if (old_mpqueue != NULL && old_mpqueue != new_mpqueue) {
385 panic("_call_entry_enqueue_deadline() "
386 "old_mpqueue %p != new_mpqueue", old_mpqueue);
387 }
388 #endif /* TIMER_ASSERT */
389
390 /* no longterm queue involved */
391 assert(new_mpqueue != timer_longterm_queue);
392 assert(old_mpqueue != timer_longterm_queue);
393
394 if (old_mpqueue == new_mpqueue) {
395 /* optimize the same-queue case to avoid a full re-insert */
396 uint64_t old_deadline = entry->tc_pqlink.deadline;
397 entry->tc_pqlink.deadline = deadline;
398
399 if (old_deadline < deadline) {
400 priority_queue_entry_increased(&new_mpqueue->mpq_pqhead,
401 &entry->tc_pqlink);
402 } else {
403 priority_queue_entry_decreased(&new_mpqueue->mpq_pqhead,
404 &entry->tc_pqlink);
405 }
406 } else {
407 if (old_mpqueue != NULL) {
408 priority_queue_remove(&old_mpqueue->mpq_pqhead,
409 &entry->tc_pqlink);
410
411 re_queue_tail(&new_mpqueue->head, &entry->tc_qlink);
412 } else {
413 enqueue_tail(&new_mpqueue->head, &entry->tc_qlink);
414 }
415
416 entry->tc_queue = &new_mpqueue->head;
417 entry->tc_pqlink.deadline = deadline;
418
419 priority_queue_insert(&new_mpqueue->mpq_pqhead, &entry->tc_pqlink);
420 }
421
422
423 /* For efficiency, track the earliest soft deadline on the queue,
424 * so that fuzzy decisions can be made without lock acquisitions.
425 */
426
427 timer_call_t thead = priority_queue_min(&new_mpqueue->mpq_pqhead, struct timer_call, tc_pqlink);
428
429 new_mpqueue->earliest_soft_deadline = thead->tc_flags & TIMER_CALL_RATELIMITED ? thead->tc_pqlink.deadline : thead->tc_soft_deadline;
430
431 if (old_mpqueue) {
432 old_mpqueue->count--;
433 }
434 new_mpqueue->count++;
435
436 return old_mpqueue;
437 }
438
439 static __inline__ void
timer_call_entry_enqueue_tail(timer_call_t entry,mpqueue_head_t * queue)440 timer_call_entry_enqueue_tail(
441 timer_call_t entry,
442 mpqueue_head_t *queue)
443 {
444 /* entry is always dequeued before this call */
445 assert(entry->tc_queue == NULL);
446
447 /*
448 * this is only used for timer_longterm_queue, which is unordered
449 * and thus needs no priority queueing
450 */
451 assert(queue == timer_longterm_queue);
452
453 enqueue_tail(&queue->head, &entry->tc_qlink);
454
455 entry->tc_queue = &queue->head;
456
457 queue->count++;
458 return;
459 }
460
461 /*
462 * Remove timer entry from its queue but don't change the queue pointer
463 * and set the async_dequeue flag. This is locking case 2b.
464 */
465 static __inline__ void
timer_call_entry_dequeue_async(timer_call_t entry)466 timer_call_entry_dequeue_async(
467 timer_call_t entry)
468 {
469 mpqueue_head_t *old_mpqueue = mpqueue_for_timer_call(entry);
470 if (old_mpqueue) {
471 old_mpqueue->count--;
472
473 if (old_mpqueue != timer_longterm_queue) {
474 priority_queue_remove(&old_mpqueue->mpq_pqhead,
475 &entry->tc_pqlink);
476 }
477
478 remqueue(&entry->tc_qlink);
479 entry->tc_async_dequeue = true;
480 }
481 return;
482 }
483
484 #if TIMER_ASSERT
485 unsigned timer_call_enqueue_deadline_unlocked_async1;
486 unsigned timer_call_enqueue_deadline_unlocked_async2;
487 #endif
488 /*
489 * Assumes call_entry and queues unlocked, interrupts disabled.
490 */
491 __inline__ mpqueue_head_t *
timer_call_enqueue_deadline_unlocked(timer_call_t call,mpqueue_head_t * queue,uint64_t deadline,uint64_t soft_deadline,uint64_t ttd,timer_call_param_t param1,uint32_t callout_flags)492 timer_call_enqueue_deadline_unlocked(
493 timer_call_t call,
494 mpqueue_head_t *queue,
495 uint64_t deadline,
496 uint64_t soft_deadline,
497 uint64_t ttd,
498 timer_call_param_t param1,
499 uint32_t callout_flags)
500 {
501 DBG("timer_call_enqueue_deadline_unlocked(%p,%p,)\n", call, queue);
502
503 simple_lock(&call->tc_lock, LCK_GRP_NULL);
504
505 mpqueue_head_t *old_queue = mpqueue_for_timer_call(call);
506
507 if (old_queue != NULL) {
508 timer_queue_lock_spin(old_queue);
509 if (call->tc_async_dequeue) {
510 /* collision (1c): timer already dequeued, clear flag */
511 #if TIMER_ASSERT
512 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
513 DECR_TIMER_ASYNC_DEQ | DBG_FUNC_NONE,
514 VM_KERNEL_UNSLIDE_OR_PERM(call),
515 call->tc_async_dequeue,
516 VM_KERNEL_UNSLIDE_OR_PERM(call->tc_queue),
517 0x1c, 0);
518 timer_call_enqueue_deadline_unlocked_async1++;
519 #endif
520 call->tc_async_dequeue = false;
521 call->tc_queue = NULL;
522 } else if (old_queue != queue) {
523 timer_call_entry_dequeue(call);
524 #if TIMER_ASSERT
525 timer_call_enqueue_deadline_unlocked_async2++;
526 #endif
527 }
528 if (old_queue == timer_longterm_queue) {
529 timer_longterm_dequeued_locked(call);
530 }
531 if (old_queue != queue) {
532 timer_queue_unlock(old_queue);
533 timer_queue_lock_spin(queue);
534 }
535 } else {
536 timer_queue_lock_spin(queue);
537 }
538
539 call->tc_soft_deadline = soft_deadline;
540 call->tc_flags = callout_flags;
541 call->tc_param1 = param1;
542 call->tc_ttd = ttd;
543
544 timer_call_entry_enqueue_deadline(call, queue, deadline);
545 timer_queue_unlock(queue);
546 simple_unlock(&call->tc_lock);
547
548 return old_queue;
549 }
550
551 #if TIMER_ASSERT
552 unsigned timer_call_dequeue_unlocked_async1;
553 unsigned timer_call_dequeue_unlocked_async2;
554 #endif
555 mpqueue_head_t *
timer_call_dequeue_unlocked(timer_call_t call)556 timer_call_dequeue_unlocked(
557 timer_call_t call)
558 {
559 DBG("timer_call_dequeue_unlocked(%p)\n", call);
560
561 simple_lock(&call->tc_lock, LCK_GRP_NULL);
562
563 mpqueue_head_t *old_queue = mpqueue_for_timer_call(call);
564
565 #if TIMER_ASSERT
566 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
567 DECR_TIMER_ASYNC_DEQ | DBG_FUNC_NONE,
568 VM_KERNEL_UNSLIDE_OR_PERM(call),
569 call->tc_async_dequeue,
570 VM_KERNEL_UNSLIDE_OR_PERM(call->tc_queue),
571 0, 0);
572 #endif
573 if (old_queue != NULL) {
574 timer_queue_lock_spin(old_queue);
575 if (call->tc_async_dequeue) {
576 /* collision (1c): timer already dequeued, clear flag */
577 #if TIMER_ASSERT
578 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
579 DECR_TIMER_ASYNC_DEQ | DBG_FUNC_NONE,
580 VM_KERNEL_UNSLIDE_OR_PERM(call),
581 call->tc_async_dequeue,
582 VM_KERNEL_UNSLIDE_OR_PERM(call->tc_queue),
583 0x1c, 0);
584 timer_call_dequeue_unlocked_async1++;
585 #endif
586 call->tc_async_dequeue = false;
587 call->tc_queue = NULL;
588 } else {
589 timer_call_entry_dequeue(call);
590 }
591 if (old_queue == timer_longterm_queue) {
592 timer_longterm_dequeued_locked(call);
593 }
594 timer_queue_unlock(old_queue);
595 }
596 simple_unlock(&call->tc_lock);
597 return old_queue;
598 }
599
600 uint64_t
timer_call_past_deadline_timer_handle(uint64_t deadline,uint64_t ctime)601 timer_call_past_deadline_timer_handle(uint64_t deadline, uint64_t ctime)
602 {
603 uint64_t delta = (ctime - deadline);
604
605 past_deadline_timers++;
606 past_deadline_deltas += delta;
607 if (delta > past_deadline_longest) {
608 past_deadline_longest = deadline;
609 }
610 if (delta < past_deadline_shortest) {
611 past_deadline_shortest = delta;
612 }
613
614 return ctime + past_deadline_timer_adjustment;
615 }
616
617 /*
618 * Timer call entry locking model
619 * ==============================
620 *
621 * Timer call entries are linked on per-cpu timer queues which are protected
622 * by the queue lock and the call entry lock. The locking protocol is:
623 *
624 * 0) The canonical locking order is timer call entry followed by queue.
625 *
626 * 1) With only the entry lock held, entry.queue is valid:
627 * 1a) NULL: the entry is not queued, or
628 * 1b) non-NULL: this queue must be locked before the entry is modified.
629 * After locking the queue, the call.async_dequeue flag must be checked:
630 * 1c) TRUE: the entry was removed from the queue by another thread
631 * and we must NULL the entry.queue and reset this flag, or
632 * 1d) FALSE: (ie. queued), the entry can be manipulated.
633 *
634 * 2) If a queue lock is obtained first, the queue is stable:
635 * 2a) If a try-lock of a queued entry succeeds, the call can be operated on
636 * and dequeued.
637 * 2b) If a try-lock fails, it indicates that another thread is attempting
638 * to change the entry and move it to a different position in this queue
639 * or to different queue. The entry can be dequeued but it should not be
640 * operated upon since it is being changed. Furthermore, we don't null
641 * the entry.queue pointer (protected by the entry lock we don't own).
642 * Instead, we set the async_dequeue flag -- see (1c).
643 * 2c) Same as 2b but occurring when a longterm timer is matured.
644 * 3) A callout's parameters (deadline, flags, parameters, soft deadline &c.)
645 * should be manipulated with the appropriate timer queue lock held,
646 * to prevent queue traversal observations from observing inconsistent
647 * updates to an in-flight callout.
648 */
649
650 /*
651 * In the debug case, we assert that the timer call locking protocol
652 * is being obeyed.
653 */
654
655 static boolean_t
timer_call_enter_internal(timer_call_t call,timer_call_param_t param1,uint64_t deadline,uint64_t leeway,uint32_t flags,boolean_t ratelimited)656 timer_call_enter_internal(
657 timer_call_t call,
658 timer_call_param_t param1,
659 uint64_t deadline,
660 uint64_t leeway,
661 uint32_t flags,
662 boolean_t ratelimited)
663 {
664 mpqueue_head_t *queue = NULL;
665 mpqueue_head_t *old_queue;
666 spl_t s;
667 uint64_t slop;
668 uint32_t urgency;
669 uint64_t sdeadline, ttd;
670
671 assert(call->tc_func != NULL);
672 s = splclock();
673
674 sdeadline = deadline;
675 uint64_t ctime = mach_absolute_time();
676
677 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
678 DECR_TIMER_ENTER | DBG_FUNC_START,
679 VM_KERNEL_UNSLIDE_OR_PERM(call),
680 VM_KERNEL_ADDRHIDE(param1), deadline, flags, 0);
681
682 urgency = (flags & TIMER_CALL_URGENCY_MASK);
683
684 boolean_t slop_ratelimited = FALSE;
685 slop = timer_call_slop(deadline, ctime, urgency, current_thread(), &slop_ratelimited);
686
687 if ((flags & TIMER_CALL_LEEWAY) != 0 && leeway > slop) {
688 slop = leeway;
689 }
690
691 if (UINT64_MAX - deadline <= slop) {
692 deadline = UINT64_MAX;
693 } else {
694 deadline += slop;
695 }
696
697 if (__improbable(deadline < ctime)) {
698 deadline = timer_call_past_deadline_timer_handle(deadline, ctime);
699 sdeadline = deadline;
700 }
701
702 if (ratelimited || slop_ratelimited) {
703 flags |= TIMER_CALL_RATELIMITED;
704 } else {
705 flags &= ~TIMER_CALL_RATELIMITED;
706 }
707
708 ttd = sdeadline - ctime;
709 #if CONFIG_DTRACE
710 DTRACE_TMR7(callout__create, timer_call_func_t, call->tc_func,
711 timer_call_param_t, call->tc_param0, uint32_t, flags,
712 (deadline - sdeadline),
713 (ttd >> 32), (unsigned) (ttd & 0xFFFFFFFF), call);
714 #endif
715
716 /* Program timer callout parameters under the appropriate per-CPU or
717 * longterm queue lock. The callout may have been previously enqueued
718 * and in-flight on this or another timer queue.
719 */
720 if (!ratelimited && !slop_ratelimited) {
721 queue = timer_longterm_enqueue_unlocked(call, ctime, deadline, &old_queue, sdeadline, ttd, param1, flags);
722 }
723
724 if (queue == NULL) {
725 queue = timer_queue_assign(deadline);
726 old_queue = timer_call_enqueue_deadline_unlocked(call, queue, deadline, sdeadline, ttd, param1, flags);
727 }
728
729 #if TIMER_TRACE
730 call->tc_entry_time = ctime;
731 #endif
732
733 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
734 DECR_TIMER_ENTER | DBG_FUNC_END,
735 VM_KERNEL_UNSLIDE_OR_PERM(call),
736 (old_queue != NULL), deadline, queue->count, 0);
737
738 splx(s);
739
740 return old_queue != NULL;
741 }
742
743 /*
744 * timer_call_*()
745 * return boolean indicating whether the call was previously queued.
746 */
747 boolean_t
timer_call_enter(timer_call_t call,uint64_t deadline,uint32_t flags)748 timer_call_enter(
749 timer_call_t call,
750 uint64_t deadline,
751 uint32_t flags)
752 {
753 return timer_call_enter_internal(call, NULL, deadline, 0, flags, FALSE);
754 }
755
756 boolean_t
timer_call_enter1(timer_call_t call,timer_call_param_t param1,uint64_t deadline,uint32_t flags)757 timer_call_enter1(
758 timer_call_t call,
759 timer_call_param_t param1,
760 uint64_t deadline,
761 uint32_t flags)
762 {
763 return timer_call_enter_internal(call, param1, deadline, 0, flags, FALSE);
764 }
765
766 boolean_t
timer_call_enter_with_leeway(timer_call_t call,timer_call_param_t param1,uint64_t deadline,uint64_t leeway,uint32_t flags,boolean_t ratelimited)767 timer_call_enter_with_leeway(
768 timer_call_t call,
769 timer_call_param_t param1,
770 uint64_t deadline,
771 uint64_t leeway,
772 uint32_t flags,
773 boolean_t ratelimited)
774 {
775 return timer_call_enter_internal(call, param1, deadline, leeway, flags, ratelimited);
776 }
777
778 boolean_t
timer_call_cancel(timer_call_t call)779 timer_call_cancel(
780 timer_call_t call)
781 {
782 mpqueue_head_t *old_queue;
783 spl_t s;
784
785 s = splclock();
786
787 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
788 DECR_TIMER_CANCEL | DBG_FUNC_START,
789 VM_KERNEL_UNSLIDE_OR_PERM(call),
790 call->tc_pqlink.deadline, call->tc_soft_deadline, call->tc_flags, 0);
791
792 old_queue = timer_call_dequeue_unlocked(call);
793
794 if (old_queue != NULL) {
795 timer_queue_lock_spin(old_queue);
796
797 timer_call_t new_head = priority_queue_min(&old_queue->mpq_pqhead, struct timer_call, tc_pqlink);
798
799 if (new_head) {
800 timer_queue_cancel(old_queue, call->tc_pqlink.deadline, new_head->tc_pqlink.deadline);
801 old_queue->earliest_soft_deadline = new_head->tc_flags & TIMER_CALL_RATELIMITED ? new_head->tc_pqlink.deadline : new_head->tc_soft_deadline;
802 } else {
803 timer_queue_cancel(old_queue, call->tc_pqlink.deadline, UINT64_MAX);
804 old_queue->earliest_soft_deadline = UINT64_MAX;
805 }
806
807 timer_queue_unlock(old_queue);
808 }
809 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
810 DECR_TIMER_CANCEL | DBG_FUNC_END,
811 VM_KERNEL_UNSLIDE_OR_PERM(call),
812 VM_KERNEL_UNSLIDE_OR_PERM(old_queue),
813 call->tc_pqlink.deadline - mach_absolute_time(),
814 call->tc_pqlink.deadline - call->tc_entry_time, 0);
815 splx(s);
816
817 #if CONFIG_DTRACE
818 DTRACE_TMR6(callout__cancel, timer_call_func_t, call->tc_func,
819 timer_call_param_t, call->tc_param0, uint32_t, call->tc_flags, 0,
820 (call->tc_ttd >> 32), (unsigned) (call->tc_ttd & 0xFFFFFFFF));
821 #endif /* CONFIG_DTRACE */
822
823 return old_queue != NULL;
824 }
825
826 static uint32_t timer_queue_shutdown_lock_skips;
827 static uint32_t timer_queue_shutdown_discarded;
828
829 void
timer_queue_shutdown(mpqueue_head_t * queue)830 timer_queue_shutdown(
831 mpqueue_head_t *queue)
832 {
833 timer_call_t call;
834 mpqueue_head_t *new_queue;
835 spl_t s;
836
837
838 DBG("timer_queue_shutdown(%p)\n", queue);
839
840 s = splclock();
841
842 while (TRUE) {
843 timer_queue_lock_spin(queue);
844
845 call = qe_queue_first(&queue->head, struct timer_call, tc_qlink);
846
847 if (call == NULL) {
848 break;
849 }
850
851 if (!simple_lock_try(&call->tc_lock, LCK_GRP_NULL)) {
852 /*
853 * case (2b) lock order inversion, dequeue and skip
854 * Don't change the call_entry queue back-pointer
855 * but set the async_dequeue field.
856 */
857 timer_queue_shutdown_lock_skips++;
858 timer_call_entry_dequeue_async(call);
859 #if TIMER_ASSERT
860 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
861 DECR_TIMER_ASYNC_DEQ | DBG_FUNC_NONE,
862 VM_KERNEL_UNSLIDE_OR_PERM(call),
863 call->tc_async_dequeue,
864 VM_KERNEL_UNSLIDE_OR_PERM(call->tc_queue),
865 0x2b, 0);
866 #endif
867 timer_queue_unlock(queue);
868 continue;
869 }
870
871 boolean_t call_local = ((call->tc_flags & TIMER_CALL_LOCAL) != 0);
872
873 /* remove entry from old queue */
874 timer_call_entry_dequeue(call);
875 timer_queue_unlock(queue);
876
877 if (call_local == FALSE) {
878 /* and queue it on new, discarding LOCAL timers */
879 new_queue = timer_queue_assign(call->tc_pqlink.deadline);
880 timer_queue_lock_spin(new_queue);
881 timer_call_entry_enqueue_deadline(
882 call, new_queue, call->tc_pqlink.deadline);
883 timer_queue_unlock(new_queue);
884 } else {
885 timer_queue_shutdown_discarded++;
886 }
887
888 assert(call_local == FALSE);
889 simple_unlock(&call->tc_lock);
890 }
891
892 timer_queue_unlock(queue);
893 splx(s);
894 }
895
896
897 static uint32_t timer_queue_expire_lock_skips;
898 uint64_t
timer_queue_expire_with_options(mpqueue_head_t * queue,uint64_t deadline,boolean_t rescan)899 timer_queue_expire_with_options(
900 mpqueue_head_t *queue,
901 uint64_t deadline,
902 boolean_t rescan)
903 {
904 timer_call_t call = NULL;
905 uint32_t tc_iterations = 0;
906 DBG("timer_queue_expire(%p,)\n", queue);
907
908 /* 'rescan' means look at every timer in the list, instead of
909 * early-exiting when the head of the list expires in the future.
910 * when 'rescan' is true, iterate by linked list instead of priority queue.
911 *
912 * TODO: if we keep a deadline ordered and soft-deadline ordered
913 * priority queue, then it's no longer necessary to do that
914 */
915
916 uint64_t cur_deadline = deadline;
917
918 /* Force an early return if this time limit is hit. */
919 const uint64_t time_limit = deadline + timer_scan_limit_abs;
920
921 /* Next deadline if the time limit is hit */
922 uint64_t time_limit_deadline = 0;
923
924 timer_queue_lock_spin(queue);
925
926 while (!queue_empty(&queue->head)) {
927 if (++tc_iterations > 1) {
928 const uint64_t now = mach_absolute_time();
929
930 /*
931 * Abort the scan if it's taking too long to avoid long
932 * periods with interrupts disabled.
933 * Scanning will restart after a short pause
934 * (timer_scan_interval_abs) if there's a hard deadline
935 * pending.
936 */
937 if (rescan == FALSE && now > time_limit) {
938 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
939 DECR_TIMER_PAUSE | DBG_FUNC_NONE,
940 queue->count, tc_iterations - 1,
941 0, 0, 0);
942
943 counter_inc(&timer_scan_pauses_cnt);
944 time_limit_deadline = now + timer_scan_interval_abs;
945 break;
946 }
947
948 /*
949 * Upon processing one or more timer calls, refresh the
950 * deadline to account for time elapsed in the callout
951 */
952 cur_deadline = now;
953 }
954
955 if (call == NULL) {
956 if (rescan == FALSE) {
957 call = priority_queue_min(&queue->mpq_pqhead, struct timer_call, tc_pqlink);
958 } else {
959 call = qe_queue_first(&queue->head, struct timer_call, tc_qlink);
960 }
961 }
962
963 if (call->tc_soft_deadline <= cur_deadline) {
964 timer_call_func_t func;
965 timer_call_param_t param0, param1;
966
967 TCOAL_DEBUG(0xDDDD0000, queue->earliest_soft_deadline, call->tc_soft_deadline, 0, 0, 0);
968 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
969 DECR_TIMER_EXPIRE | DBG_FUNC_NONE,
970 VM_KERNEL_UNSLIDE_OR_PERM(call),
971 call->tc_soft_deadline,
972 call->tc_pqlink.deadline,
973 call->tc_entry_time, 0);
974
975 if ((call->tc_flags & TIMER_CALL_RATELIMITED) &&
976 (call->tc_pqlink.deadline > cur_deadline)) {
977 if (rescan == FALSE) {
978 break;
979 }
980 }
981
982 if (!simple_lock_try(&call->tc_lock, LCK_GRP_NULL)) {
983 /* case (2b) lock inversion, dequeue and skip */
984 timer_queue_expire_lock_skips++;
985 timer_call_entry_dequeue_async(call);
986 call = NULL;
987 continue;
988 }
989
990 timer_call_entry_dequeue(call);
991
992 func = call->tc_func;
993 param0 = call->tc_param0;
994 param1 = call->tc_param1;
995
996 simple_unlock(&call->tc_lock);
997 timer_queue_unlock(queue);
998
999 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
1000 DECR_TIMER_CALLOUT | DBG_FUNC_START,
1001 VM_KERNEL_UNSLIDE_OR_PERM(call), VM_KERNEL_UNSLIDE(func),
1002 VM_KERNEL_ADDRHIDE(param0),
1003 VM_KERNEL_ADDRHIDE(param1),
1004 0);
1005
1006 #if CONFIG_DTRACE
1007 DTRACE_TMR7(callout__start, timer_call_func_t, func,
1008 timer_call_param_t, param0, unsigned, call->tc_flags,
1009 0, (call->tc_ttd >> 32),
1010 (unsigned) (call->tc_ttd & 0xFFFFFFFF), call);
1011 #endif
1012 /* Maintain time-to-deadline in per-processor data
1013 * structure for thread wakeup deadline statistics.
1014 */
1015 uint64_t *ttdp = ¤t_processor()->timer_call_ttd;
1016 *ttdp = call->tc_ttd;
1017 (*func)(param0, param1);
1018 *ttdp = 0;
1019 #if CONFIG_DTRACE
1020 DTRACE_TMR4(callout__end, timer_call_func_t, func,
1021 param0, param1, call);
1022 #endif
1023
1024 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
1025 DECR_TIMER_CALLOUT | DBG_FUNC_END,
1026 VM_KERNEL_UNSLIDE_OR_PERM(call), VM_KERNEL_UNSLIDE(func),
1027 VM_KERNEL_ADDRHIDE(param0),
1028 VM_KERNEL_ADDRHIDE(param1),
1029 0);
1030 call = NULL;
1031 timer_queue_lock_spin(queue);
1032 } else {
1033 if (__probable(rescan == FALSE)) {
1034 break;
1035 } else {
1036 int64_t skew = call->tc_pqlink.deadline - call->tc_soft_deadline;
1037 assert(call->tc_pqlink.deadline >= call->tc_soft_deadline);
1038
1039 /* DRK: On a latency quality-of-service level change,
1040 * re-sort potentially rate-limited timers. The platform
1041 * layer determines which timers require
1042 * this. In the absence of the per-callout
1043 * synchronization requirement, a global resort could
1044 * be more efficient. The re-sort effectively
1045 * annuls all timer adjustments, i.e. the "soft
1046 * deadline" is the sort key.
1047 */
1048
1049 if (timer_resort_threshold(skew)) {
1050 if (__probable(simple_lock_try(&call->tc_lock, LCK_GRP_NULL))) {
1051 /* TODO: don't need to dequeue before enqueue */
1052 timer_call_entry_dequeue(call);
1053 timer_call_entry_enqueue_deadline(call, queue, call->tc_soft_deadline);
1054 simple_unlock(&call->tc_lock);
1055 call = NULL;
1056 }
1057 }
1058 if (call) {
1059 call = qe_queue_next(&queue->head, call, struct timer_call, tc_qlink);
1060
1061 if (call == NULL) {
1062 break;
1063 }
1064 }
1065 }
1066 }
1067 }
1068
1069 call = priority_queue_min(&queue->mpq_pqhead, struct timer_call, tc_pqlink);
1070
1071 if (call) {
1072 /*
1073 * Even if the time limit has been hit, it doesn't mean a hard
1074 * deadline will be missed - the next hard deadline may be in
1075 * future.
1076 */
1077 if (time_limit_deadline > call->tc_pqlink.deadline) {
1078 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
1079 DECR_TIMER_POSTPONE | DBG_FUNC_NONE,
1080 VM_KERNEL_UNSLIDE_OR_PERM(call),
1081 call->tc_pqlink.deadline,
1082 time_limit_deadline,
1083 0, 0);
1084 counter_inc(&timer_scan_postpones_cnt);
1085 cur_deadline = time_limit_deadline;
1086 } else {
1087 cur_deadline = call->tc_pqlink.deadline;
1088 }
1089 queue->earliest_soft_deadline = (call->tc_flags & TIMER_CALL_RATELIMITED) ? call->tc_pqlink.deadline: call->tc_soft_deadline;
1090 } else {
1091 queue->earliest_soft_deadline = cur_deadline = UINT64_MAX;
1092 }
1093
1094 timer_queue_unlock(queue);
1095
1096 return cur_deadline;
1097 }
1098
1099 uint64_t
timer_queue_expire(mpqueue_head_t * queue,uint64_t deadline)1100 timer_queue_expire(
1101 mpqueue_head_t *queue,
1102 uint64_t deadline)
1103 {
1104 return timer_queue_expire_with_options(queue, deadline, FALSE);
1105 }
1106
1107 extern int serverperfmode;
1108 static uint32_t timer_queue_migrate_lock_skips;
1109 /*
1110 * timer_queue_migrate() is called by timer_queue_migrate_cpu()
1111 * to move timer requests from the local processor (queue_from)
1112 * to a target processor's (queue_to).
1113 */
1114 int
timer_queue_migrate(mpqueue_head_t * queue_from,mpqueue_head_t * queue_to)1115 timer_queue_migrate(mpqueue_head_t *queue_from, mpqueue_head_t *queue_to)
1116 {
1117 timer_call_t call;
1118 timer_call_t head_to;
1119 int timers_migrated = 0;
1120
1121 DBG("timer_queue_migrate(%p,%p)\n", queue_from, queue_to);
1122
1123 assert(!ml_get_interrupts_enabled());
1124 assert(queue_from != queue_to);
1125
1126 if (serverperfmode) {
1127 /*
1128 * if we're running a high end server
1129 * avoid migrations... they add latency
1130 * and don't save us power under typical
1131 * server workloads
1132 */
1133 return -4;
1134 }
1135
1136 /*
1137 * Take both local (from) and target (to) timer queue locks while
1138 * moving the timers from the local queue to the target processor.
1139 * We assume that the target is always the boot processor.
1140 * But only move if all of the following is true:
1141 * - the target queue is non-empty
1142 * - the local queue is non-empty
1143 * - the local queue's first deadline is later than the target's
1144 * - the local queue contains no non-migrateable "local" call
1145 * so that we need not have the target resync.
1146 */
1147
1148 timer_queue_lock_spin(queue_to);
1149
1150 head_to = priority_queue_min(&queue_to->mpq_pqhead, struct timer_call, tc_pqlink);
1151
1152 if (head_to == NULL) {
1153 timers_migrated = -1;
1154 goto abort1;
1155 }
1156
1157 timer_queue_lock_spin(queue_from);
1158
1159 call = priority_queue_min(&queue_from->mpq_pqhead, struct timer_call, tc_pqlink);
1160
1161 if (call == NULL) {
1162 timers_migrated = -2;
1163 goto abort2;
1164 }
1165
1166 if (call->tc_pqlink.deadline < head_to->tc_pqlink.deadline) {
1167 timers_migrated = 0;
1168 goto abort2;
1169 }
1170
1171 /* perform scan for non-migratable timers */
1172 qe_foreach_element(call, &queue_from->head, tc_qlink) {
1173 if (call->tc_flags & TIMER_CALL_LOCAL) {
1174 timers_migrated = -3;
1175 goto abort2;
1176 }
1177 }
1178
1179 /* migration loop itself -- both queues are locked */
1180 qe_foreach_element_safe(call, &queue_from->head, tc_qlink) {
1181 if (!simple_lock_try(&call->tc_lock, LCK_GRP_NULL)) {
1182 /* case (2b) lock order inversion, dequeue only */
1183 #ifdef TIMER_ASSERT
1184 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
1185 DECR_TIMER_ASYNC_DEQ | DBG_FUNC_NONE,
1186 VM_KERNEL_UNSLIDE_OR_PERM(call),
1187 VM_KERNEL_UNSLIDE_OR_PERM(call->tc_queue),
1188 0,
1189 0x2b, 0);
1190 #endif
1191 timer_queue_migrate_lock_skips++;
1192 timer_call_entry_dequeue_async(call);
1193 continue;
1194 }
1195 timer_call_entry_dequeue(call);
1196 timer_call_entry_enqueue_deadline(
1197 call, queue_to, call->tc_pqlink.deadline);
1198 timers_migrated++;
1199 simple_unlock(&call->tc_lock);
1200 }
1201 queue_from->earliest_soft_deadline = UINT64_MAX;
1202 abort2:
1203 timer_queue_unlock(queue_from);
1204 abort1:
1205 timer_queue_unlock(queue_to);
1206
1207 return timers_migrated;
1208 }
1209
1210 void
timer_queue_trace_cpu(int ncpu)1211 timer_queue_trace_cpu(int ncpu)
1212 {
1213 timer_call_nosync_cpu(
1214 ncpu,
1215 (void (*)(void *))timer_queue_trace,
1216 (void*) timer_queue_cpu(ncpu));
1217 }
1218
1219 void
timer_queue_trace(mpqueue_head_t * queue)1220 timer_queue_trace(
1221 mpqueue_head_t *queue)
1222 {
1223 timer_call_t call;
1224 spl_t s;
1225
1226 if (!kdebug_enable) {
1227 return;
1228 }
1229
1230 s = splclock();
1231 timer_queue_lock_spin(queue);
1232
1233 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
1234 DECR_TIMER_QUEUE | DBG_FUNC_START,
1235 queue->count, mach_absolute_time(), 0, 0, 0);
1236
1237 qe_foreach_element(call, &queue->head, tc_qlink) {
1238 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
1239 DECR_TIMER_QUEUE | DBG_FUNC_NONE,
1240 call->tc_soft_deadline,
1241 call->tc_pqlink.deadline,
1242 call->tc_entry_time,
1243 VM_KERNEL_UNSLIDE(call->tc_func),
1244 0);
1245 }
1246
1247 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
1248 DECR_TIMER_QUEUE | DBG_FUNC_END,
1249 queue->count, mach_absolute_time(), 0, 0, 0);
1250
1251 timer_queue_unlock(queue);
1252 splx(s);
1253 }
1254
1255 void
timer_longterm_dequeued_locked(timer_call_t call)1256 timer_longterm_dequeued_locked(timer_call_t call)
1257 {
1258 timer_longterm_t *tlp = &timer_longterm;
1259
1260 tlp->dequeues++;
1261 if (call == tlp->threshold.call) {
1262 tlp->threshold.call = NULL;
1263 }
1264 }
1265
1266 /*
1267 * Place a timer call in the longterm list
1268 * and adjust the next timer callout deadline if the new timer is first.
1269 */
1270 mpqueue_head_t *
timer_longterm_enqueue_unlocked(timer_call_t call,uint64_t now,uint64_t deadline,mpqueue_head_t ** old_queue,uint64_t soft_deadline,uint64_t ttd,timer_call_param_t param1,uint32_t callout_flags)1271 timer_longterm_enqueue_unlocked(timer_call_t call,
1272 uint64_t now,
1273 uint64_t deadline,
1274 mpqueue_head_t **old_queue,
1275 uint64_t soft_deadline,
1276 uint64_t ttd,
1277 timer_call_param_t param1,
1278 uint32_t callout_flags)
1279 {
1280 timer_longterm_t *tlp = &timer_longterm;
1281 boolean_t update_required = FALSE;
1282 uint64_t longterm_threshold;
1283
1284 longterm_threshold = now + tlp->threshold.interval;
1285
1286 /*
1287 * Return NULL without doing anything if:
1288 * - this timer is local, or
1289 * - the longterm mechanism is disabled, or
1290 * - this deadline is too short.
1291 */
1292 if ((callout_flags & TIMER_CALL_LOCAL) != 0 ||
1293 (tlp->threshold.interval == TIMER_LONGTERM_NONE) ||
1294 (deadline <= longterm_threshold)) {
1295 return NULL;
1296 }
1297
1298 /*
1299 * Remove timer from its current queue, if any.
1300 */
1301 *old_queue = timer_call_dequeue_unlocked(call);
1302
1303 /*
1304 * Lock the longterm queue, queue timer and determine
1305 * whether an update is necessary.
1306 */
1307 assert(!ml_get_interrupts_enabled());
1308 simple_lock(&call->tc_lock, LCK_GRP_NULL);
1309 timer_queue_lock_spin(timer_longterm_queue);
1310 call->tc_pqlink.deadline = deadline;
1311 call->tc_param1 = param1;
1312 call->tc_ttd = ttd;
1313 call->tc_soft_deadline = soft_deadline;
1314 call->tc_flags = callout_flags;
1315 timer_call_entry_enqueue_tail(call, timer_longterm_queue);
1316
1317 tlp->enqueues++;
1318
1319 /*
1320 * We'll need to update the currently set threshold timer
1321 * if the new deadline is sooner and no sooner update is in flight.
1322 */
1323 if (deadline < tlp->threshold.deadline &&
1324 deadline < tlp->threshold.preempted) {
1325 tlp->threshold.preempted = deadline;
1326 tlp->threshold.call = call;
1327 update_required = TRUE;
1328 }
1329 timer_queue_unlock(timer_longterm_queue);
1330 simple_unlock(&call->tc_lock);
1331
1332 if (update_required) {
1333 /*
1334 * Note: this call expects that calling the master cpu
1335 * alone does not involve locking the topo lock.
1336 */
1337 timer_call_nosync_cpu(
1338 master_cpu,
1339 (void (*)(void *))timer_longterm_update,
1340 (void *)tlp);
1341 }
1342
1343 return timer_longterm_queue;
1344 }
1345
1346 /*
1347 * Scan for timers below the longterm threshold.
1348 * Move these to the local timer queue (of the boot processor on which the
1349 * calling thread is running).
1350 * Both the local (boot) queue and the longterm queue are locked.
1351 * The scan is similar to the timer migrate sequence but is performed by
1352 * successively examining each timer on the longterm queue:
1353 * - if within the short-term threshold
1354 * - enter on the local queue (unless being deleted),
1355 * - otherwise:
1356 * - if sooner, deadline becomes the next threshold deadline.
1357 * The total scan time is limited to TIMER_LONGTERM_SCAN_LIMIT. Should this be
1358 * exceeded, we abort and reschedule again so that we don't shut others from
1359 * the timer queues. Longterm timers firing late is not critical.
1360 */
1361 void
timer_longterm_scan(timer_longterm_t * tlp,uint64_t time_start)1362 timer_longterm_scan(timer_longterm_t *tlp,
1363 uint64_t time_start)
1364 {
1365 timer_call_t call;
1366 uint64_t threshold = TIMER_LONGTERM_NONE;
1367 uint64_t deadline;
1368 uint64_t time_limit = time_start + tlp->scan_limit;
1369 mpqueue_head_t *timer_master_queue;
1370
1371 assert(!ml_get_interrupts_enabled());
1372 assert(cpu_number() == master_cpu);
1373
1374 if (tlp->threshold.interval != TIMER_LONGTERM_NONE) {
1375 threshold = time_start + tlp->threshold.interval;
1376 }
1377
1378 tlp->threshold.deadline = TIMER_LONGTERM_NONE;
1379 tlp->threshold.call = NULL;
1380
1381 if (queue_empty(&timer_longterm_queue->head)) {
1382 return;
1383 }
1384
1385 timer_master_queue = timer_queue_cpu(master_cpu);
1386 timer_queue_lock_spin(timer_master_queue);
1387
1388 qe_foreach_element_safe(call, &timer_longterm_queue->head, tc_qlink) {
1389 deadline = call->tc_soft_deadline;
1390 if (!simple_lock_try(&call->tc_lock, LCK_GRP_NULL)) {
1391 /* case (2c) lock order inversion, dequeue only */
1392 #ifdef TIMER_ASSERT
1393 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
1394 DECR_TIMER_ASYNC_DEQ | DBG_FUNC_NONE,
1395 VM_KERNEL_UNSLIDE_OR_PERM(call),
1396 VM_KERNEL_UNSLIDE_OR_PERM(call->tc_queue),
1397 0,
1398 0x2c, 0);
1399 #endif
1400 timer_call_entry_dequeue_async(call);
1401 continue;
1402 }
1403 if (deadline < threshold) {
1404 /*
1405 * This timer needs moving (escalating)
1406 * to the local (boot) processor's queue.
1407 */
1408 #ifdef TIMER_ASSERT
1409 if (deadline < time_start) {
1410 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
1411 DECR_TIMER_OVERDUE | DBG_FUNC_NONE,
1412 VM_KERNEL_UNSLIDE_OR_PERM(call),
1413 deadline,
1414 time_start,
1415 threshold,
1416 0);
1417 }
1418 #endif
1419 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
1420 DECR_TIMER_ESCALATE | DBG_FUNC_NONE,
1421 VM_KERNEL_UNSLIDE_OR_PERM(call),
1422 call->tc_pqlink.deadline,
1423 call->tc_entry_time,
1424 VM_KERNEL_UNSLIDE(call->tc_func),
1425 0);
1426 tlp->escalates++;
1427 timer_call_entry_dequeue(call);
1428 timer_call_entry_enqueue_deadline(
1429 call, timer_master_queue, call->tc_pqlink.deadline);
1430 /*
1431 * A side-effect of the following call is to update
1432 * the actual hardware deadline if required.
1433 */
1434 (void) timer_queue_assign(deadline);
1435 } else {
1436 if (deadline < tlp->threshold.deadline) {
1437 tlp->threshold.deadline = deadline;
1438 tlp->threshold.call = call;
1439 }
1440 }
1441 simple_unlock(&call->tc_lock);
1442
1443 /* Abort scan if we're taking too long. */
1444 if (mach_absolute_time() > time_limit) {
1445 tlp->threshold.deadline = TIMER_LONGTERM_SCAN_AGAIN;
1446 tlp->scan_pauses++;
1447 DBG("timer_longterm_scan() paused %llu, qlen: %llu\n",
1448 time_limit, tlp->queue.count);
1449 break;
1450 }
1451 }
1452
1453 timer_queue_unlock(timer_master_queue);
1454 }
1455
1456 void
timer_longterm_callout(timer_call_param_t p0,__unused timer_call_param_t p1)1457 timer_longterm_callout(timer_call_param_t p0, __unused timer_call_param_t p1)
1458 {
1459 timer_longterm_t *tlp = (timer_longterm_t *) p0;
1460
1461 timer_longterm_update(tlp);
1462 }
1463
1464 void
timer_longterm_update_locked(timer_longterm_t * tlp)1465 timer_longterm_update_locked(timer_longterm_t *tlp)
1466 {
1467 uint64_t latency;
1468
1469 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
1470 DECR_TIMER_UPDATE | DBG_FUNC_START,
1471 VM_KERNEL_UNSLIDE_OR_PERM(&tlp->queue),
1472 tlp->threshold.deadline,
1473 tlp->threshold.preempted,
1474 tlp->queue.count, 0);
1475
1476 tlp->scan_time = mach_absolute_time();
1477 if (tlp->threshold.preempted != TIMER_LONGTERM_NONE) {
1478 tlp->threshold.preempts++;
1479 tlp->threshold.deadline = tlp->threshold.preempted;
1480 tlp->threshold.preempted = TIMER_LONGTERM_NONE;
1481 /*
1482 * Note: in the unlikely event that a pre-empted timer has
1483 * itself been cancelled, we'll simply re-scan later at the
1484 * time of the preempted/cancelled timer.
1485 */
1486 } else {
1487 tlp->threshold.scans++;
1488
1489 /*
1490 * Maintain a moving average of our wakeup latency.
1491 * Clamp latency to 0 and ignore above threshold interval.
1492 */
1493 if (tlp->scan_time > tlp->threshold.deadline_set) {
1494 latency = tlp->scan_time - tlp->threshold.deadline_set;
1495 } else {
1496 latency = 0;
1497 }
1498 if (latency < tlp->threshold.interval) {
1499 tlp->threshold.latency_min =
1500 MIN(tlp->threshold.latency_min, latency);
1501 tlp->threshold.latency_max =
1502 MAX(tlp->threshold.latency_max, latency);
1503 tlp->threshold.latency =
1504 (tlp->threshold.latency * 99 + latency) / 100;
1505 }
1506
1507 timer_longterm_scan(tlp, tlp->scan_time);
1508 }
1509
1510 tlp->threshold.deadline_set = tlp->threshold.deadline;
1511 /* The next deadline timer to be set is adjusted */
1512 if (tlp->threshold.deadline != TIMER_LONGTERM_NONE &&
1513 tlp->threshold.deadline != TIMER_LONGTERM_SCAN_AGAIN) {
1514 tlp->threshold.deadline_set -= tlp->threshold.margin;
1515 tlp->threshold.deadline_set -= tlp->threshold.latency;
1516 }
1517
1518 /* Throttle next scan time */
1519 uint64_t scan_clamp = mach_absolute_time() + tlp->scan_interval;
1520 if (tlp->threshold.deadline_set < scan_clamp) {
1521 tlp->threshold.deadline_set = scan_clamp;
1522 }
1523
1524 TIMER_KDEBUG_TRACE(KDEBUG_TRACE,
1525 DECR_TIMER_UPDATE | DBG_FUNC_END,
1526 VM_KERNEL_UNSLIDE_OR_PERM(&tlp->queue),
1527 tlp->threshold.deadline,
1528 tlp->threshold.scans,
1529 tlp->queue.count, 0);
1530 }
1531
1532 void
timer_longterm_update(timer_longterm_t * tlp)1533 timer_longterm_update(timer_longterm_t *tlp)
1534 {
1535 spl_t s = splclock();
1536
1537 timer_queue_lock_spin(timer_longterm_queue);
1538
1539 if (cpu_number() != master_cpu) {
1540 panic("timer_longterm_update_master() on non-boot cpu");
1541 }
1542
1543 timer_longterm_update_locked(tlp);
1544
1545 if (tlp->threshold.deadline != TIMER_LONGTERM_NONE) {
1546 timer_call_enter(
1547 &tlp->threshold.timer,
1548 tlp->threshold.deadline_set,
1549 TIMER_CALL_LOCAL | TIMER_CALL_SYS_CRITICAL);
1550 }
1551
1552 timer_queue_unlock(timer_longterm_queue);
1553 splx(s);
1554 }
1555
1556 void
timer_longterm_init(void)1557 timer_longterm_init(void)
1558 {
1559 uint32_t longterm;
1560 timer_longterm_t *tlp = &timer_longterm;
1561
1562 DBG("timer_longterm_init() tlp: %p, queue: %p\n", tlp, &tlp->queue);
1563
1564 /*
1565 * Set the longterm timer threshold. Defaults to TIMER_LONGTERM_THRESHOLD
1566 * or TIMER_LONGTERM_NONE (disabled) for server;
1567 * overridden longterm boot-arg
1568 */
1569 tlp->threshold.interval = serverperfmode ? TIMER_LONGTERM_NONE
1570 : TIMER_LONGTERM_THRESHOLD;
1571 if (PE_parse_boot_argn("longterm", &longterm, sizeof(longterm))) {
1572 tlp->threshold.interval = (longterm == 0) ?
1573 TIMER_LONGTERM_NONE :
1574 longterm * NSEC_PER_MSEC;
1575 }
1576 if (tlp->threshold.interval != TIMER_LONGTERM_NONE) {
1577 printf("Longterm timer threshold: %llu ms\n",
1578 tlp->threshold.interval / NSEC_PER_MSEC);
1579 kprintf("Longterm timer threshold: %llu ms\n",
1580 tlp->threshold.interval / NSEC_PER_MSEC);
1581 nanoseconds_to_absolutetime(tlp->threshold.interval,
1582 &tlp->threshold.interval);
1583 tlp->threshold.margin = tlp->threshold.interval / 10;
1584 tlp->threshold.latency_min = EndOfAllTime;
1585 tlp->threshold.latency_max = 0;
1586 }
1587
1588 tlp->threshold.preempted = TIMER_LONGTERM_NONE;
1589 tlp->threshold.deadline = TIMER_LONGTERM_NONE;
1590
1591 mpqueue_init(&tlp->queue, &timer_longterm_lck_grp, LCK_ATTR_NULL);
1592
1593 timer_call_setup(&tlp->threshold.timer,
1594 timer_longterm_callout, (timer_call_param_t) tlp);
1595
1596 timer_longterm_queue = &tlp->queue;
1597 }
1598
1599 enum {
1600 THRESHOLD, QCOUNT,
1601 ENQUEUES, DEQUEUES, ESCALATES, SCANS, PREEMPTS,
1602 LATENCY, LATENCY_MIN, LATENCY_MAX, LONG_TERM_SCAN_LIMIT,
1603 LONG_TERM_SCAN_INTERVAL, LONG_TERM_SCAN_PAUSES,
1604 SCAN_LIMIT, SCAN_INTERVAL, SCAN_PAUSES, SCAN_POSTPONES,
1605 };
1606 uint64_t
timer_sysctl_get(int oid)1607 timer_sysctl_get(int oid)
1608 {
1609 timer_longterm_t *tlp = &timer_longterm;
1610
1611 switch (oid) {
1612 case THRESHOLD:
1613 return (tlp->threshold.interval == TIMER_LONGTERM_NONE) ?
1614 0 : tlp->threshold.interval / NSEC_PER_MSEC;
1615 case QCOUNT:
1616 return tlp->queue.count;
1617 case ENQUEUES:
1618 return tlp->enqueues;
1619 case DEQUEUES:
1620 return tlp->dequeues;
1621 case ESCALATES:
1622 return tlp->escalates;
1623 case SCANS:
1624 return tlp->threshold.scans;
1625 case PREEMPTS:
1626 return tlp->threshold.preempts;
1627 case LATENCY:
1628 return tlp->threshold.latency;
1629 case LATENCY_MIN:
1630 return tlp->threshold.latency_min;
1631 case LATENCY_MAX:
1632 return tlp->threshold.latency_max;
1633 case LONG_TERM_SCAN_LIMIT:
1634 return tlp->scan_limit;
1635 case LONG_TERM_SCAN_INTERVAL:
1636 return tlp->scan_interval;
1637 case LONG_TERM_SCAN_PAUSES:
1638 return tlp->scan_pauses;
1639 case SCAN_LIMIT:
1640 return timer_scan_limit_us * NSEC_PER_USEC;
1641 case SCAN_INTERVAL:
1642 return timer_scan_interval_us * NSEC_PER_USEC;
1643 case SCAN_PAUSES:
1644 return counter_load(&timer_scan_pauses_cnt);
1645 case SCAN_POSTPONES:
1646 return counter_load(&timer_scan_postpones_cnt);
1647
1648 default:
1649 return 0;
1650 }
1651 }
1652
1653 /*
1654 * timer_master_scan() is the inverse of timer_longterm_scan()
1655 * since it un-escalates timers to the longterm queue.
1656 */
1657 static void
timer_master_scan(timer_longterm_t * tlp,uint64_t now)1658 timer_master_scan(timer_longterm_t *tlp,
1659 uint64_t now)
1660 {
1661 timer_call_t call;
1662 uint64_t threshold;
1663 uint64_t deadline;
1664 mpqueue_head_t *timer_master_queue;
1665
1666 if (tlp->threshold.interval != TIMER_LONGTERM_NONE) {
1667 threshold = now + tlp->threshold.interval;
1668 } else {
1669 threshold = TIMER_LONGTERM_NONE;
1670 }
1671
1672 timer_master_queue = timer_queue_cpu(master_cpu);
1673 timer_queue_lock_spin(timer_master_queue);
1674
1675 qe_foreach_element_safe(call, &timer_master_queue->head, tc_qlink) {
1676 deadline = call->tc_pqlink.deadline;
1677 if ((call->tc_flags & TIMER_CALL_LOCAL) != 0) {
1678 continue;
1679 }
1680 if (!simple_lock_try(&call->tc_lock, LCK_GRP_NULL)) {
1681 /* case (2c) lock order inversion, dequeue only */
1682 timer_call_entry_dequeue_async(call);
1683 continue;
1684 }
1685 if (deadline > threshold) {
1686 /* move from master to longterm */
1687 timer_call_entry_dequeue(call);
1688 timer_call_entry_enqueue_tail(call, timer_longterm_queue);
1689 if (deadline < tlp->threshold.deadline) {
1690 tlp->threshold.deadline = deadline;
1691 tlp->threshold.call = call;
1692 }
1693 }
1694 simple_unlock(&call->tc_lock);
1695 }
1696 timer_queue_unlock(timer_master_queue);
1697 }
1698
1699 static void
timer_sysctl_set_threshold(void * valp)1700 timer_sysctl_set_threshold(void* valp)
1701 {
1702 uint64_t value = (uint64_t)valp;
1703 timer_longterm_t *tlp = &timer_longterm;
1704 spl_t s = splclock();
1705 boolean_t threshold_increase;
1706
1707 timer_queue_lock_spin(timer_longterm_queue);
1708
1709 timer_call_cancel(&tlp->threshold.timer);
1710
1711 /*
1712 * Set the new threshold and note whther it's increasing.
1713 */
1714 if (value == 0) {
1715 tlp->threshold.interval = TIMER_LONGTERM_NONE;
1716 threshold_increase = TRUE;
1717 timer_call_cancel(&tlp->threshold.timer);
1718 } else {
1719 uint64_t old_interval = tlp->threshold.interval;
1720 tlp->threshold.interval = value * NSEC_PER_MSEC;
1721 nanoseconds_to_absolutetime(tlp->threshold.interval,
1722 &tlp->threshold.interval);
1723 tlp->threshold.margin = tlp->threshold.interval / 10;
1724 if (old_interval == TIMER_LONGTERM_NONE) {
1725 threshold_increase = FALSE;
1726 } else {
1727 threshold_increase = (tlp->threshold.interval > old_interval);
1728 }
1729 }
1730
1731 if (threshold_increase /* or removal */) {
1732 /* Escalate timers from the longterm queue */
1733 timer_longterm_scan(tlp, mach_absolute_time());
1734 } else { /* decrease or addition */
1735 /*
1736 * We scan the local/master queue for timers now longterm.
1737 * To be strictly correct, we should scan all processor queues
1738 * but timer migration results in most timers gravitating to the
1739 * master processor in any case.
1740 */
1741 timer_master_scan(tlp, mach_absolute_time());
1742 }
1743
1744 /* Set new timer accordingly */
1745 tlp->threshold.deadline_set = tlp->threshold.deadline;
1746 if (tlp->threshold.deadline != TIMER_LONGTERM_NONE) {
1747 tlp->threshold.deadline_set -= tlp->threshold.margin;
1748 tlp->threshold.deadline_set -= tlp->threshold.latency;
1749 timer_call_enter(
1750 &tlp->threshold.timer,
1751 tlp->threshold.deadline_set,
1752 TIMER_CALL_LOCAL | TIMER_CALL_SYS_CRITICAL);
1753 }
1754
1755 /* Reset stats */
1756 tlp->enqueues = 0;
1757 tlp->dequeues = 0;
1758 tlp->escalates = 0;
1759 tlp->scan_pauses = 0;
1760 tlp->threshold.scans = 0;
1761 tlp->threshold.preempts = 0;
1762 tlp->threshold.latency = 0;
1763 tlp->threshold.latency_min = EndOfAllTime;
1764 tlp->threshold.latency_max = 0;
1765
1766 timer_queue_unlock(timer_longterm_queue);
1767 splx(s);
1768 }
1769
1770 int
timer_sysctl_set(int oid,uint64_t value)1771 timer_sysctl_set(int oid, uint64_t value)
1772 {
1773 switch (oid) {
1774 case THRESHOLD:
1775 timer_call_cpu(
1776 master_cpu,
1777 timer_sysctl_set_threshold,
1778 (void *) value);
1779 return KERN_SUCCESS;
1780 case LONG_TERM_SCAN_LIMIT:
1781 timer_longterm.scan_limit = value;
1782 return KERN_SUCCESS;
1783 case LONG_TERM_SCAN_INTERVAL:
1784 timer_longterm.scan_interval = value;
1785 return KERN_SUCCESS;
1786 case SCAN_LIMIT:
1787 if (value > MAX_TIMER_SCAN_LIMIT ||
1788 value < MIN_TIMER_SCAN_LIMIT) {
1789 return KERN_INVALID_ARGUMENT;
1790 }
1791 timer_scan_limit_us = value / NSEC_PER_USEC;
1792 nanoseconds_to_absolutetime(timer_scan_limit_us * NSEC_PER_USEC,
1793 &timer_scan_limit_abs);
1794 return KERN_SUCCESS;
1795 case SCAN_INTERVAL:
1796 if (value > MAX_TIMER_SCAN_INTERVAL ||
1797 value < MIN_TIMER_SCAN_INTERVAL) {
1798 return KERN_INVALID_ARGUMENT;
1799 }
1800 timer_scan_interval_us = value / NSEC_PER_USEC;
1801 nanoseconds_to_absolutetime(timer_scan_interval_us * NSEC_PER_USEC,
1802 &timer_scan_interval_abs);
1803 return KERN_SUCCESS;
1804 default:
1805 return KERN_INVALID_ARGUMENT;
1806 }
1807 }
1808
1809
1810 /* Select timer coalescing window based on per-task quality-of-service hints */
1811 static boolean_t
tcoal_qos_adjust(thread_t t,int32_t * tshift,uint64_t * tmax_abstime,boolean_t * pratelimited)1812 tcoal_qos_adjust(thread_t t, int32_t *tshift, uint64_t *tmax_abstime, boolean_t *pratelimited)
1813 {
1814 uint32_t latency_qos;
1815 boolean_t adjusted = FALSE;
1816 task_t ctask = get_threadtask(t);
1817
1818 if (ctask) {
1819 latency_qos = proc_get_effective_thread_policy(t, TASK_POLICY_LATENCY_QOS);
1820
1821 assert(latency_qos <= NUM_LATENCY_QOS_TIERS);
1822
1823 if (latency_qos) {
1824 *tshift = tcoal_prio_params.latency_qos_scale[latency_qos - 1];
1825 *tmax_abstime = tcoal_prio_params.latency_qos_abstime_max[latency_qos - 1];
1826 *pratelimited = tcoal_prio_params.latency_tier_rate_limited[latency_qos - 1];
1827 adjusted = TRUE;
1828 }
1829 }
1830 return adjusted;
1831 }
1832
1833
1834 /* Adjust timer deadlines based on priority of the thread and the
1835 * urgency value provided at timeout establishment. With this mechanism,
1836 * timers are no longer necessarily sorted in order of soft deadline
1837 * on a given timer queue, i.e. they may be differentially skewed.
1838 * In the current scheme, this could lead to fewer pending timers
1839 * processed than is technically possible when the HW deadline arrives.
1840 */
1841 static void
timer_compute_leeway(thread_t cthread,int32_t urgency,int32_t * tshift,uint64_t * tmax_abstime,boolean_t * pratelimited)1842 timer_compute_leeway(thread_t cthread, int32_t urgency, int32_t *tshift, uint64_t *tmax_abstime, boolean_t *pratelimited)
1843 {
1844 int16_t tpri = cthread->sched_pri;
1845 if ((urgency & TIMER_CALL_USER_MASK) != 0) {
1846 bool tg_critical = false;
1847 #if CONFIG_THREAD_GROUPS
1848 uint32_t tg_flags = thread_group_get_flags(thread_group_get(cthread));
1849 tg_critical = tg_flags & (THREAD_GROUP_FLAGS_CRITICAL | THREAD_GROUP_FLAGS_STRICT_TIMERS);
1850 #endif /* CONFIG_THREAD_GROUPS */
1851 bool timer_critical = (tpri >= BASEPRI_RTQUEUES) ||
1852 (urgency == TIMER_CALL_USER_CRITICAL) ||
1853 tg_critical;
1854 if (timer_critical) {
1855 *tshift = tcoal_prio_params.timer_coalesce_rt_shift;
1856 *tmax_abstime = tcoal_prio_params.timer_coalesce_rt_abstime_max;
1857 TCOAL_PRIO_STAT(rt_tcl);
1858 } else if (proc_get_effective_thread_policy(cthread, TASK_POLICY_DARWIN_BG) ||
1859 (urgency == TIMER_CALL_USER_BACKGROUND)) {
1860 /* Determine if timer should be subjected to a lower QoS */
1861 if (tcoal_qos_adjust(cthread, tshift, tmax_abstime, pratelimited)) {
1862 if (*tmax_abstime > tcoal_prio_params.timer_coalesce_bg_abstime_max) {
1863 return;
1864 } else {
1865 *pratelimited = FALSE;
1866 }
1867 }
1868 *tshift = tcoal_prio_params.timer_coalesce_bg_shift;
1869 *tmax_abstime = tcoal_prio_params.timer_coalesce_bg_abstime_max;
1870 TCOAL_PRIO_STAT(bg_tcl);
1871 } else if (tpri >= MINPRI_KERNEL) {
1872 *tshift = tcoal_prio_params.timer_coalesce_kt_shift;
1873 *tmax_abstime = tcoal_prio_params.timer_coalesce_kt_abstime_max;
1874 TCOAL_PRIO_STAT(kt_tcl);
1875 } else if (cthread->sched_mode == TH_MODE_FIXED) {
1876 *tshift = tcoal_prio_params.timer_coalesce_fp_shift;
1877 *tmax_abstime = tcoal_prio_params.timer_coalesce_fp_abstime_max;
1878 TCOAL_PRIO_STAT(fp_tcl);
1879 } else if (tcoal_qos_adjust(cthread, tshift, tmax_abstime, pratelimited)) {
1880 TCOAL_PRIO_STAT(qos_tcl);
1881 } else if (cthread->sched_mode == TH_MODE_TIMESHARE) {
1882 *tshift = tcoal_prio_params.timer_coalesce_ts_shift;
1883 *tmax_abstime = tcoal_prio_params.timer_coalesce_ts_abstime_max;
1884 TCOAL_PRIO_STAT(ts_tcl);
1885 } else {
1886 TCOAL_PRIO_STAT(nc_tcl);
1887 }
1888 } else if (urgency == TIMER_CALL_SYS_BACKGROUND) {
1889 *tshift = tcoal_prio_params.timer_coalesce_bg_shift;
1890 *tmax_abstime = tcoal_prio_params.timer_coalesce_bg_abstime_max;
1891 TCOAL_PRIO_STAT(bg_tcl);
1892 } else {
1893 *tshift = tcoal_prio_params.timer_coalesce_kt_shift;
1894 *tmax_abstime = tcoal_prio_params.timer_coalesce_kt_abstime_max;
1895 TCOAL_PRIO_STAT(kt_tcl);
1896 }
1897 }
1898
1899
1900 int timer_user_idle_level;
1901
1902 uint64_t
timer_call_slop(uint64_t deadline,uint64_t now,uint32_t flags,thread_t cthread,boolean_t * pratelimited)1903 timer_call_slop(uint64_t deadline, uint64_t now, uint32_t flags, thread_t cthread, boolean_t *pratelimited)
1904 {
1905 int32_t tcs_shift = 0;
1906 uint64_t tcs_max_abstime = 0;
1907 uint64_t adjval;
1908 uint32_t urgency = (flags & TIMER_CALL_URGENCY_MASK);
1909
1910 if (mach_timer_coalescing_enabled &&
1911 (deadline > now) && (urgency != TIMER_CALL_SYS_CRITICAL)) {
1912 timer_compute_leeway(cthread, urgency, &tcs_shift, &tcs_max_abstime, pratelimited);
1913
1914 if (tcs_shift >= 0) {
1915 adjval = MIN((deadline - now) >> tcs_shift, tcs_max_abstime);
1916 } else {
1917 adjval = MIN((deadline - now) << (-tcs_shift), tcs_max_abstime);
1918 }
1919 /* Apply adjustments derived from "user idle level" heuristic */
1920 adjval += (adjval * timer_user_idle_level) >> 7;
1921 return adjval;
1922 } else {
1923 return 0;
1924 }
1925 }
1926
1927 int
timer_get_user_idle_level(void)1928 timer_get_user_idle_level(void)
1929 {
1930 return timer_user_idle_level;
1931 }
1932
1933 kern_return_t
timer_set_user_idle_level(int ilevel)1934 timer_set_user_idle_level(int ilevel)
1935 {
1936 boolean_t do_reeval = FALSE;
1937
1938 if ((ilevel < 0) || (ilevel > 128)) {
1939 return KERN_INVALID_ARGUMENT;
1940 }
1941
1942 if (ilevel < timer_user_idle_level) {
1943 do_reeval = TRUE;
1944 }
1945
1946 timer_user_idle_level = ilevel;
1947
1948 if (do_reeval) {
1949 ml_timer_evaluate();
1950 }
1951
1952 return KERN_SUCCESS;
1953 }
1954
1955 #pragma mark - running timers
1956
1957 #define RUNNING_TIMER_FAKE_FLAGS (TIMER_CALL_SYS_CRITICAL | \
1958 TIMER_CALL_LOCAL)
1959
1960 /*
1961 * timer_call_trace_* functions mimic the tracing behavior from the normal
1962 * timer_call subsystem, so tools continue to function.
1963 */
1964
1965 static void
timer_call_trace_enter_before(struct timer_call * call,uint64_t deadline,uint32_t flags,uint64_t now)1966 timer_call_trace_enter_before(struct timer_call *call, uint64_t deadline,
1967 uint32_t flags, uint64_t now)
1968 {
1969 #pragma unused(call, deadline, flags, now)
1970 TIMER_KDEBUG_TRACE(KDEBUG_TRACE, DECR_TIMER_ENTER | DBG_FUNC_START,
1971 VM_KERNEL_UNSLIDE_OR_PERM(call), VM_KERNEL_ADDRHIDE(call->tc_param1),
1972 deadline, flags, 0);
1973 #if CONFIG_DTRACE
1974 uint64_t ttd = deadline - now;
1975 DTRACE_TMR7(callout__create, timer_call_func_t, call->tc_func,
1976 timer_call_param_t, call->tc_param0, uint32_t, flags, 0,
1977 (ttd >> 32), (unsigned int)(ttd & 0xFFFFFFFF), NULL);
1978 #endif /* CONFIG_DTRACE */
1979 TIMER_KDEBUG_TRACE(KDEBUG_TRACE, DECR_TIMER_ENTER | DBG_FUNC_END,
1980 VM_KERNEL_UNSLIDE_OR_PERM(call), 0, deadline, 0, 0);
1981 }
1982
1983 static void
timer_call_trace_enter_after(struct timer_call * call,uint64_t deadline)1984 timer_call_trace_enter_after(struct timer_call *call, uint64_t deadline)
1985 {
1986 #pragma unused(call, deadline)
1987 TIMER_KDEBUG_TRACE(KDEBUG_TRACE, DECR_TIMER_ENTER | DBG_FUNC_END,
1988 VM_KERNEL_UNSLIDE_OR_PERM(call), 0, deadline, 0, 0);
1989 }
1990
1991 static void
timer_call_trace_cancel(struct timer_call * call)1992 timer_call_trace_cancel(struct timer_call *call)
1993 {
1994 #pragma unused(call)
1995 __unused uint64_t deadline = call->tc_pqlink.deadline;
1996 TIMER_KDEBUG_TRACE(KDEBUG_TRACE, DECR_TIMER_CANCEL | DBG_FUNC_START,
1997 VM_KERNEL_UNSLIDE_OR_PERM(call), deadline, 0,
1998 call->tc_flags, 0);
1999 TIMER_KDEBUG_TRACE(KDEBUG_TRACE, DECR_TIMER_CANCEL | DBG_FUNC_END,
2000 VM_KERNEL_UNSLIDE_OR_PERM(call), 0, deadline - mach_absolute_time(),
2001 deadline - call->tc_entry_time, 0);
2002 #if CONFIG_DTRACE
2003 #if TIMER_TRACE
2004 uint64_t ttd = deadline - call->tc_entry_time;
2005 #else
2006 uint64_t ttd = UINT64_MAX;
2007 #endif /* TIMER_TRACE */
2008 DTRACE_TMR6(callout__cancel, timer_call_func_t, call->tc_func,
2009 timer_call_param_t, call->tc_param0, uint32_t, call->tc_flags, 0,
2010 (ttd >> 32), (unsigned int)(ttd & 0xFFFFFFFF));
2011 #endif /* CONFIG_DTRACE */
2012 }
2013
2014 static void
timer_call_trace_expire_entry(struct timer_call * call)2015 timer_call_trace_expire_entry(struct timer_call *call)
2016 {
2017 #pragma unused(call)
2018 TIMER_KDEBUG_TRACE(KDEBUG_TRACE, DECR_TIMER_CALLOUT | DBG_FUNC_START,
2019 VM_KERNEL_UNSLIDE_OR_PERM(call), VM_KERNEL_UNSLIDE(call->tc_func),
2020 VM_KERNEL_ADDRHIDE(call->tc_param0),
2021 VM_KERNEL_ADDRHIDE(call->tc_param1),
2022 0);
2023 #if CONFIG_DTRACE
2024 #if TIMER_TRACE
2025 uint64_t ttd = call->tc_pqlink.deadline - call->tc_entry_time;
2026 #else /* TIMER_TRACE */
2027 uint64_t ttd = UINT64_MAX;
2028 #endif /* TIMER_TRACE */
2029 DTRACE_TMR7(callout__start, timer_call_func_t, call->tc_func,
2030 timer_call_param_t, call->tc_param0, unsigned, call->tc_flags,
2031 0, (ttd >> 32), (unsigned int)(ttd & 0xFFFFFFFF), NULL);
2032 #endif /* CONFIG_DTRACE */
2033 }
2034
2035 static void
timer_call_trace_expire_return(struct timer_call * call)2036 timer_call_trace_expire_return(struct timer_call *call)
2037 {
2038 #pragma unused(call)
2039 #if CONFIG_DTRACE
2040 DTRACE_TMR4(callout__end, timer_call_func_t, call->tc_func,
2041 call->tc_param0, call->tc_param1, NULL);
2042 #endif /* CONFIG_DTRACE */
2043 TIMER_KDEBUG_TRACE(KDEBUG_TRACE, DECR_TIMER_CALLOUT | DBG_FUNC_END,
2044 VM_KERNEL_UNSLIDE_OR_PERM(call),
2045 VM_KERNEL_UNSLIDE(call->tc_func),
2046 VM_KERNEL_ADDRHIDE(call->tc_param0),
2047 VM_KERNEL_ADDRHIDE(call->tc_param1),
2048 0);
2049 }
2050
2051 /*
2052 * Set a new deadline for a running timer on this processor.
2053 */
2054 void
running_timer_setup(processor_t processor,enum running_timer timer,void * param,uint64_t deadline,uint64_t now)2055 running_timer_setup(processor_t processor, enum running_timer timer,
2056 void *param, uint64_t deadline, uint64_t now)
2057 {
2058 assert(timer < RUNNING_TIMER_MAX);
2059 assert(ml_get_interrupts_enabled() == FALSE);
2060
2061 struct timer_call *call = &processor->running_timers[timer];
2062
2063 timer_call_trace_enter_before(call, deadline, RUNNING_TIMER_FAKE_FLAGS,
2064 now);
2065
2066 if (__improbable(deadline < now)) {
2067 deadline = timer_call_past_deadline_timer_handle(deadline, now);
2068 }
2069
2070 call->tc_pqlink.deadline = deadline;
2071 #if TIMER_TRACE
2072 call->tc_entry_time = now;
2073 #endif /* TIMER_TRACE */
2074 call->tc_param1 = param;
2075
2076 timer_call_trace_enter_after(call, deadline);
2077 }
2078
2079 void
running_timers_sync(void)2080 running_timers_sync(void)
2081 {
2082 timer_resync_deadlines();
2083 }
2084
2085 void
running_timer_enter(processor_t processor,unsigned int timer,void * param,uint64_t deadline,uint64_t now)2086 running_timer_enter(processor_t processor, unsigned int timer,
2087 void *param, uint64_t deadline, uint64_t now)
2088 {
2089 running_timer_setup(processor, timer, param, deadline, now);
2090 running_timers_sync();
2091 }
2092
2093 /*
2094 * Call the callback for any running timers that fired for this processor.
2095 * Returns true if any timers were past their deadline.
2096 */
2097 bool
running_timers_expire(processor_t processor,uint64_t now)2098 running_timers_expire(processor_t processor, uint64_t now)
2099 {
2100 bool expired = false;
2101
2102 if (!processor->running_timers_active) {
2103 return expired;
2104 }
2105
2106 for (int i = 0; i < RUNNING_TIMER_MAX; i++) {
2107 struct timer_call *call = &processor->running_timers[i];
2108
2109 uint64_t deadline = call->tc_pqlink.deadline;
2110 if (deadline > now) {
2111 continue;
2112 }
2113
2114 expired = true;
2115 timer_call_trace_expire_entry(call);
2116 call->tc_func(call->tc_param0, call->tc_param1);
2117 timer_call_trace_expire_return(call);
2118 }
2119
2120 return expired;
2121 }
2122
2123 void
running_timer_clear(processor_t processor,enum running_timer timer)2124 running_timer_clear(processor_t processor, enum running_timer timer)
2125 {
2126 struct timer_call *call = &processor->running_timers[timer];
2127 uint64_t deadline = call->tc_pqlink.deadline;
2128 if (deadline == EndOfAllTime) {
2129 return;
2130 }
2131
2132 call->tc_pqlink.deadline = EndOfAllTime;
2133 #if TIMER_TRACE
2134 call->tc_entry_time = 0;
2135 #endif /* TIMER_TRACE */
2136 timer_call_trace_cancel(call);
2137 }
2138
2139 void
running_timer_cancel(processor_t processor,unsigned int timer)2140 running_timer_cancel(processor_t processor, unsigned int timer)
2141 {
2142 running_timer_clear(processor, timer);
2143 running_timers_sync();
2144 }
2145
2146 uint64_t
running_timers_deadline(processor_t processor)2147 running_timers_deadline(processor_t processor)
2148 {
2149 if (!processor->running_timers_active) {
2150 return EndOfAllTime;
2151 }
2152
2153 uint64_t deadline = EndOfAllTime;
2154 for (int i = 0; i < RUNNING_TIMER_MAX; i++) {
2155 uint64_t candidate =
2156 processor->running_timers[i].tc_pqlink.deadline;
2157 if (candidate != 0 && candidate < deadline) {
2158 deadline = candidate;
2159 }
2160 }
2161
2162 return deadline;
2163 }
2164
2165 void
running_timers_activate(processor_t processor)2166 running_timers_activate(processor_t processor)
2167 {
2168 processor->running_timers_active = true;
2169 running_timers_sync();
2170 }
2171
2172 void
running_timers_deactivate(processor_t processor)2173 running_timers_deactivate(processor_t processor)
2174 {
2175 assert(processor->running_timers_active == true);
2176 processor->running_timers_active = false;
2177 running_timers_sync();
2178 }
2179