1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Infrastructure for migratable timers
4 *
5 * Copyright(C) 2022 linutronix GmbH
6 */
7 #include <linux/cpuhotplug.h>
8 #include <linux/slab.h>
9 #include <linux/smp.h>
10 #include <linux/spinlock.h>
11 #include <linux/timerqueue.h>
12 #include <trace/events/ipi.h>
13
14 #include "timer_migration.h"
15 #include "tick-internal.h"
16
17 #define CREATE_TRACE_POINTS
18 #include <trace/events/timer_migration.h>
19
20 /*
21 * The timer migration mechanism is built on a hierarchy of groups. The
22 * lowest level group contains CPUs, the next level groups of CPU groups
23 * and so forth. The CPU groups are kept per node so for the normal case
24 * lock contention won't happen across nodes. Depending on the number of
25 * CPUs per node even the next level might be kept as groups of CPU groups
26 * per node and only the levels above cross the node topology.
27 *
28 * Example topology for a two node system with 24 CPUs each.
29 *
30 * LVL 2 [GRP2:0]
31 * GRP1:0 = GRP1:M
32 *
33 * LVL 1 [GRP1:0] [GRP1:1]
34 * GRP0:0 - GRP0:2 GRP0:3 - GRP0:5
35 *
36 * LVL 0 [GRP0:0] [GRP0:1] [GRP0:2] [GRP0:3] [GRP0:4] [GRP0:5]
37 * CPUS 0-7 8-15 16-23 24-31 32-39 40-47
38 *
39 * The groups hold a timer queue of events sorted by expiry time. These
40 * queues are updated when CPUs go in idle. When they come out of idle
41 * ignore flag of events is set.
42 *
43 * Each group has a designated migrator CPU/group as long as a CPU/group is
44 * active in the group. This designated role is necessary to avoid that all
45 * active CPUs in a group try to migrate expired timers from other CPUs,
46 * which would result in massive lock bouncing.
47 *
48 * When a CPU is awake, it checks in it's own timer tick the group
49 * hierarchy up to the point where it is assigned the migrator role or if
50 * no CPU is active, it also checks the groups where no migrator is set
51 * (TMIGR_NONE).
52 *
53 * If it finds expired timers in one of the group queues it pulls them over
54 * from the idle CPU and runs the timer function. After that it updates the
55 * group and the parent groups if required.
56 *
57 * CPUs which go idle arm their CPU local timer hardware for the next local
58 * (pinned) timer event. If the next migratable timer expires after the
59 * next local timer or the CPU has no migratable timer pending then the
60 * CPU does not queue an event in the LVL0 group. If the next migratable
61 * timer expires before the next local timer then the CPU queues that timer
62 * in the LVL0 group. In both cases the CPU marks itself idle in the LVL0
63 * group.
64 *
65 * When CPU comes out of idle and when a group has at least a single active
66 * child, the ignore flag of the tmigr_event is set. This indicates, that
67 * the event is ignored even if it is still enqueued in the parent groups
68 * timer queue. It will be removed when touching the timer queue the next
69 * time. This spares locking in active path as the lock protects (after
70 * setup) only event information. For more information about locking,
71 * please read the section "Locking rules".
72 *
73 * If the CPU is the migrator of the group then it delegates that role to
74 * the next active CPU in the group or sets migrator to TMIGR_NONE when
75 * there is no active CPU in the group. This delegation needs to be
76 * propagated up the hierarchy so hand over from other leaves can happen at
77 * all hierarchy levels w/o doing a search.
78 *
79 * When the last CPU in the system goes idle, then it drops all migrator
80 * duties up to the top level of the hierarchy (LVL2 in the example). It
81 * then has to make sure, that it arms it's own local hardware timer for
82 * the earliest event in the system.
83 *
84 *
85 * Lifetime rules:
86 * ---------------
87 *
88 * The groups are built up at init time or when CPUs come online. They are
89 * not destroyed when a group becomes empty due to offlining. The group
90 * just won't participate in the hierarchy management anymore. Destroying
91 * groups would result in interesting race conditions which would just make
92 * the whole mechanism slow and complex.
93 *
94 *
95 * Locking rules:
96 * --------------
97 *
98 * For setting up new groups and handling events it's required to lock both
99 * child and parent group. The lock ordering is always bottom up. This also
100 * includes the per CPU locks in struct tmigr_cpu. For updating the migrator and
101 * active CPU/group information atomic_try_cmpxchg() is used instead and only
102 * the per CPU tmigr_cpu->lock is held.
103 *
104 * During the setup of groups tmigr_level_list is required. It is protected by
105 * @tmigr_mutex.
106 *
107 * When @timer_base->lock as well as tmigr related locks are required, the lock
108 * ordering is: first @timer_base->lock, afterwards tmigr related locks.
109 *
110 *
111 * Protection of the tmigr group state information:
112 * ------------------------------------------------
113 *
114 * The state information with the list of active children and migrator needs to
115 * be protected by a sequence counter. It prevents a race when updates in child
116 * groups are propagated in changed order. The state update is performed
117 * lockless and group wise. The following scenario describes what happens
118 * without updating the sequence counter:
119 *
120 * Therefore, let's take three groups and four CPUs (CPU2 and CPU3 as well
121 * as GRP0:1 will not change during the scenario):
122 *
123 * LVL 1 [GRP1:0]
124 * migrator = GRP0:1
125 * active = GRP0:0, GRP0:1
126 * / \
127 * LVL 0 [GRP0:0] [GRP0:1]
128 * migrator = CPU0 migrator = CPU2
129 * active = CPU0 active = CPU2
130 * / \ / \
131 * CPUs 0 1 2 3
132 * active idle active idle
133 *
134 *
135 * 1. CPU0 goes idle. As the update is performed group wise, in the first step
136 * only GRP0:0 is updated. The update of GRP1:0 is pending as CPU0 has to
137 * walk the hierarchy.
138 *
139 * LVL 1 [GRP1:0]
140 * migrator = GRP0:1
141 * active = GRP0:0, GRP0:1
142 * / \
143 * LVL 0 [GRP0:0] [GRP0:1]
144 * --> migrator = TMIGR_NONE migrator = CPU2
145 * --> active = active = CPU2
146 * / \ / \
147 * CPUs 0 1 2 3
148 * --> idle idle active idle
149 *
150 * 2. While CPU0 goes idle and continues to update the state, CPU1 comes out of
151 * idle. CPU1 updates GRP0:0. The update for GRP1:0 is pending as CPU1 also
152 * has to walk the hierarchy. Both CPUs (CPU0 and CPU1) now walk the
153 * hierarchy to perform the needed update from their point of view. The
154 * currently visible state looks the following:
155 *
156 * LVL 1 [GRP1:0]
157 * migrator = GRP0:1
158 * active = GRP0:0, GRP0:1
159 * / \
160 * LVL 0 [GRP0:0] [GRP0:1]
161 * --> migrator = CPU1 migrator = CPU2
162 * --> active = CPU1 active = CPU2
163 * / \ / \
164 * CPUs 0 1 2 3
165 * idle --> active active idle
166 *
167 * 3. Here is the race condition: CPU1 managed to propagate its changes (from
168 * step 2) through the hierarchy to GRP1:0 before CPU0 (step 1) did. The
169 * active members of GRP1:0 remain unchanged after the update since it is
170 * still valid from CPU1 current point of view:
171 *
172 * LVL 1 [GRP1:0]
173 * --> migrator = GRP0:1
174 * --> active = GRP0:0, GRP0:1
175 * / \
176 * LVL 0 [GRP0:0] [GRP0:1]
177 * migrator = CPU1 migrator = CPU2
178 * active = CPU1 active = CPU2
179 * / \ / \
180 * CPUs 0 1 2 3
181 * idle active active idle
182 *
183 * 4. Now CPU0 finally propagates its changes (from step 1) to GRP1:0.
184 *
185 * LVL 1 [GRP1:0]
186 * --> migrator = GRP0:1
187 * --> active = GRP0:1
188 * / \
189 * LVL 0 [GRP0:0] [GRP0:1]
190 * migrator = CPU1 migrator = CPU2
191 * active = CPU1 active = CPU2
192 * / \ / \
193 * CPUs 0 1 2 3
194 * idle active active idle
195 *
196 *
197 * The race of CPU0 vs. CPU1 led to an inconsistent state in GRP1:0. CPU1 is
198 * active and is correctly listed as active in GRP0:0. However GRP1:0 does not
199 * have GRP0:0 listed as active, which is wrong. The sequence counter has been
200 * added to avoid inconsistent states during updates. The state is updated
201 * atomically only if all members, including the sequence counter, match the
202 * expected value (compare-and-exchange).
203 *
204 * Looking back at the previous example with the addition of the sequence
205 * counter: The update as performed by CPU0 in step 4 will fail. CPU1 changed
206 * the sequence number during the update in step 3 so the expected old value (as
207 * seen by CPU0 before starting the walk) does not match.
208 *
209 * Prevent race between new event and last CPU going inactive
210 * ----------------------------------------------------------
211 *
212 * When the last CPU is going idle and there is a concurrent update of a new
213 * first global timer of an idle CPU, the group and child states have to be read
214 * while holding the lock in tmigr_update_events(). The following scenario shows
215 * what happens, when this is not done.
216 *
217 * 1. Only CPU2 is active:
218 *
219 * LVL 1 [GRP1:0]
220 * migrator = GRP0:1
221 * active = GRP0:1
222 * next_expiry = KTIME_MAX
223 * / \
224 * LVL 0 [GRP0:0] [GRP0:1]
225 * migrator = TMIGR_NONE migrator = CPU2
226 * active = active = CPU2
227 * next_expiry = KTIME_MAX next_expiry = KTIME_MAX
228 * / \ / \
229 * CPUs 0 1 2 3
230 * idle idle active idle
231 *
232 * 2. Now CPU 2 goes idle (and has no global timer, that has to be handled) and
233 * propagates that to GRP0:1:
234 *
235 * LVL 1 [GRP1:0]
236 * migrator = GRP0:1
237 * active = GRP0:1
238 * next_expiry = KTIME_MAX
239 * / \
240 * LVL 0 [GRP0:0] [GRP0:1]
241 * migrator = TMIGR_NONE --> migrator = TMIGR_NONE
242 * active = --> active =
243 * next_expiry = KTIME_MAX next_expiry = KTIME_MAX
244 * / \ / \
245 * CPUs 0 1 2 3
246 * idle idle --> idle idle
247 *
248 * 3. Now the idle state is propagated up to GRP1:0. As this is now the last
249 * child going idle in top level group, the expiry of the next group event
250 * has to be handed back to make sure no event is lost. As there is no event
251 * enqueued, KTIME_MAX is handed back to CPU2.
252 *
253 * LVL 1 [GRP1:0]
254 * --> migrator = TMIGR_NONE
255 * --> active =
256 * next_expiry = KTIME_MAX
257 * / \
258 * LVL 0 [GRP0:0] [GRP0:1]
259 * migrator = TMIGR_NONE migrator = TMIGR_NONE
260 * active = active =
261 * next_expiry = KTIME_MAX next_expiry = KTIME_MAX
262 * / \ / \
263 * CPUs 0 1 2 3
264 * idle idle --> idle idle
265 *
266 * 4. CPU 0 has a new timer queued from idle and it expires at TIMER0. CPU0
267 * propagates that to GRP0:0:
268 *
269 * LVL 1 [GRP1:0]
270 * migrator = TMIGR_NONE
271 * active =
272 * next_expiry = KTIME_MAX
273 * / \
274 * LVL 0 [GRP0:0] [GRP0:1]
275 * migrator = TMIGR_NONE migrator = TMIGR_NONE
276 * active = active =
277 * --> next_expiry = TIMER0 next_expiry = KTIME_MAX
278 * / \ / \
279 * CPUs 0 1 2 3
280 * idle idle idle idle
281 *
282 * 5. GRP0:0 is not active, so the new timer has to be propagated to
283 * GRP1:0. Therefore the GRP1:0 state has to be read. When the stalled value
284 * (from step 2) is read, the timer is enqueued into GRP1:0, but nothing is
285 * handed back to CPU0, as it seems that there is still an active child in
286 * top level group.
287 *
288 * LVL 1 [GRP1:0]
289 * migrator = TMIGR_NONE
290 * active =
291 * --> next_expiry = TIMER0
292 * / \
293 * LVL 0 [GRP0:0] [GRP0:1]
294 * migrator = TMIGR_NONE migrator = TMIGR_NONE
295 * active = active =
296 * next_expiry = TIMER0 next_expiry = KTIME_MAX
297 * / \ / \
298 * CPUs 0 1 2 3
299 * idle idle idle idle
300 *
301 * This is prevented by reading the state when holding the lock (when a new
302 * timer has to be propagated from idle path)::
303 *
304 * CPU2 (tmigr_inactive_up()) CPU0 (tmigr_new_timer_up())
305 * -------------------------- ---------------------------
306 * // step 3:
307 * cmpxchg(&GRP1:0->state);
308 * tmigr_update_events() {
309 * spin_lock(&GRP1:0->lock);
310 * // ... update events ...
311 * // hand back first expiry when GRP1:0 is idle
312 * spin_unlock(&GRP1:0->lock);
313 * // ^^^ release state modification
314 * }
315 * tmigr_update_events() {
316 * spin_lock(&GRP1:0->lock)
317 * // ^^^ acquire state modification
318 * group_state = atomic_read(&GRP1:0->state)
319 * // .... update events ...
320 * // hand back first expiry when GRP1:0 is idle
321 * spin_unlock(&GRP1:0->lock) <3>
322 * // ^^^ makes state visible for other
323 * // callers of tmigr_new_timer_up()
324 * }
325 *
326 * When CPU0 grabs the lock directly after cmpxchg, the first timer is reported
327 * back to CPU0 and also later on to CPU2. So no timer is missed. A concurrent
328 * update of the group state from active path is no problem, as the upcoming CPU
329 * will take care of the group events.
330 *
331 * Required event and timerqueue update after a remote expiry:
332 * -----------------------------------------------------------
333 *
334 * After expiring timers of a remote CPU, a walk through the hierarchy and
335 * update of events and timerqueues is required. It is obviously needed if there
336 * is a 'new' global timer but also if there is no new global timer but the
337 * remote CPU is still idle.
338 *
339 * 1. CPU0 and CPU1 are idle and have both a global timer expiring at the same
340 * time. So both have an event enqueued in the timerqueue of GRP0:0. CPU3 is
341 * also idle and has no global timer pending. CPU2 is the only active CPU and
342 * thus also the migrator:
343 *
344 * LVL 1 [GRP1:0]
345 * migrator = GRP0:1
346 * active = GRP0:1
347 * --> timerqueue = evt-GRP0:0
348 * / \
349 * LVL 0 [GRP0:0] [GRP0:1]
350 * migrator = TMIGR_NONE migrator = CPU2
351 * active = active = CPU2
352 * groupevt.ignore = false groupevt.ignore = true
353 * groupevt.cpu = CPU0 groupevt.cpu =
354 * timerqueue = evt-CPU0, timerqueue =
355 * evt-CPU1
356 * / \ / \
357 * CPUs 0 1 2 3
358 * idle idle active idle
359 *
360 * 2. CPU2 starts to expire remote timers. It starts with LVL0 group
361 * GRP0:1. There is no event queued in the timerqueue, so CPU2 continues with
362 * the parent of GRP0:1: GRP1:0. In GRP1:0 it dequeues the first event. It
363 * looks at tmigr_event::cpu struct member and expires the pending timer(s)
364 * of CPU0.
365 *
366 * LVL 1 [GRP1:0]
367 * migrator = GRP0:1
368 * active = GRP0:1
369 * --> timerqueue =
370 * / \
371 * LVL 0 [GRP0:0] [GRP0:1]
372 * migrator = TMIGR_NONE migrator = CPU2
373 * active = active = CPU2
374 * groupevt.ignore = false groupevt.ignore = true
375 * --> groupevt.cpu = CPU0 groupevt.cpu =
376 * timerqueue = evt-CPU0, timerqueue =
377 * evt-CPU1
378 * / \ / \
379 * CPUs 0 1 2 3
380 * idle idle active idle
381 *
382 * 3. Some work has to be done after expiring the timers of CPU0. If we stop
383 * here, then CPU1's pending global timer(s) will not expire in time and the
384 * timerqueue of GRP0:0 has still an event for CPU0 enqueued which has just
385 * been processed. So it is required to walk the hierarchy from CPU0's point
386 * of view and update it accordingly. CPU0's event will be removed from the
387 * timerqueue because it has no pending timer. If CPU0 would have a timer
388 * pending then it has to expire after CPU1's first timer because all timers
389 * from this period were just expired. Either way CPU1's event will be first
390 * in GRP0:0's timerqueue and therefore set in the CPU field of the group
391 * event which is then enqueued in GRP1:0's timerqueue as GRP0:0 is still not
392 * active:
393 *
394 * LVL 1 [GRP1:0]
395 * migrator = GRP0:1
396 * active = GRP0:1
397 * --> timerqueue = evt-GRP0:0
398 * / \
399 * LVL 0 [GRP0:0] [GRP0:1]
400 * migrator = TMIGR_NONE migrator = CPU2
401 * active = active = CPU2
402 * groupevt.ignore = false groupevt.ignore = true
403 * --> groupevt.cpu = CPU1 groupevt.cpu =
404 * --> timerqueue = evt-CPU1 timerqueue =
405 * / \ / \
406 * CPUs 0 1 2 3
407 * idle idle active idle
408 *
409 * Now CPU2 (migrator) will continue step 2 at GRP1:0 and will expire the
410 * timer(s) of CPU1.
411 *
412 * The hierarchy walk in step 3 can be skipped if the migrator notices that a
413 * CPU of GRP0:0 is active again. The CPU will mark GRP0:0 active and take care
414 * of the group as migrator and any needed updates within the hierarchy.
415 */
416
417 static DEFINE_MUTEX(tmigr_mutex);
418 static struct list_head *tmigr_level_list __read_mostly;
419
420 static unsigned int tmigr_hierarchy_levels __read_mostly;
421 static unsigned int tmigr_crossnode_level __read_mostly;
422
423 static DEFINE_PER_CPU(struct tmigr_cpu, tmigr_cpu);
424
425 #define TMIGR_NONE 0xFF
426 #define BIT_CNT 8
427
428 static inline bool tmigr_is_not_available(struct tmigr_cpu *tmc)
429 {
430 return !(tmc->tmgroup && tmc->online);
431 }
432
433 /*
434 * Returns true, when @childmask corresponds to the group migrator or when the
435 * group is not active - so no migrator is set.
436 */
437 static bool tmigr_check_migrator(struct tmigr_group *group, u8 childmask)
438 {
439 union tmigr_state s;
440
441 s.state = atomic_read(&group->migr_state);
442
443 if ((s.migrator == childmask) || (s.migrator == TMIGR_NONE))
444 return true;
445
446 return false;
447 }
448
449 static bool tmigr_check_migrator_and_lonely(struct tmigr_group *group, u8 childmask)
450 {
451 bool lonely, migrator = false;
452 unsigned long active;
453 union tmigr_state s;
454
455 s.state = atomic_read(&group->migr_state);
456
457 if ((s.migrator == childmask) || (s.migrator == TMIGR_NONE))
458 migrator = true;
459
460 active = s.active;
461 lonely = bitmap_weight(&active, BIT_CNT) <= 1;
462
463 return (migrator && lonely);
464 }
465
466 static bool tmigr_check_lonely(struct tmigr_group *group)
467 {
468 unsigned long active;
469 union tmigr_state s;
470
471 s.state = atomic_read(&group->migr_state);
472
473 active = s.active;
474
475 return bitmap_weight(&active, BIT_CNT) <= 1;
476 }
477
478 /**
479 * struct tmigr_walk - data required for walking the hierarchy
480 * @nextexp: Next CPU event expiry information which is handed into
481 * the timer migration code by the timer code
482 * (get_next_timer_interrupt())
483 * @firstexp: Contains the first event expiry information when
484 * hierarchy is completely idle. When CPU itself was the
485 * last going idle, information makes sure, that CPU will
486 * be back in time. When using this value in the remote
487 * expiry case, firstexp is stored in the per CPU tmigr_cpu
488 * struct of CPU which expires remote timers. It is updated
489 * in top level group only. Be aware, there could occur a
490 * new top level of the hierarchy between the 'top level
491 * call' in tmigr_update_events() and the check for the
492 * parent group in walk_groups(). Then @firstexp might
493 * contain a value != KTIME_MAX even if it was not the
494 * final top level. This is not a problem, as the worst
495 * outcome is a CPU which might wake up a little early.
496 * @evt: Pointer to tmigr_event which needs to be queued (of idle
497 * child group)
498 * @childmask: groupmask of child group
499 * @remote: Is set, when the new timer path is executed in
500 * tmigr_handle_remote_cpu()
501 * @basej: timer base in jiffies
502 * @now: timer base monotonic
503 * @check: is set if there is the need to handle remote timers;
504 * required in tmigr_requires_handle_remote() only
505 * @tmc_active: this flag indicates, whether the CPU which triggers
506 * the hierarchy walk is !idle in the timer migration
507 * hierarchy. When the CPU is idle and the whole hierarchy is
508 * idle, only the first event of the top level has to be
509 * considered.
510 */
511 struct tmigr_walk {
512 u64 nextexp;
513 u64 firstexp;
514 struct tmigr_event *evt;
515 u8 childmask;
516 bool remote;
517 unsigned long basej;
518 u64 now;
519 bool check;
520 bool tmc_active;
521 };
522
523 typedef bool (*up_f)(struct tmigr_group *, struct tmigr_group *, struct tmigr_walk *);
524
525 static void __walk_groups(up_f up, struct tmigr_walk *data,
526 struct tmigr_cpu *tmc)
527 {
528 struct tmigr_group *child = NULL, *group = tmc->tmgroup;
529
530 do {
531 WARN_ON_ONCE(group->level >= tmigr_hierarchy_levels);
532
533 if (up(group, child, data))
534 break;
535
536 child = group;
537 group = group->parent;
538 data->childmask = child->groupmask;
539 } while (group);
540 }
541
542 static void walk_groups(up_f up, struct tmigr_walk *data, struct tmigr_cpu *tmc)
543 {
544 lockdep_assert_held(&tmc->lock);
545
546 __walk_groups(up, data, tmc);
547 }
548
549 /*
550 * Returns the next event of the timerqueue @group->events
551 *
552 * Removes timers with ignore flag and update next_expiry of the group. Values
553 * of the group event are updated in tmigr_update_events() only.
554 */
555 static struct tmigr_event *tmigr_next_groupevt(struct tmigr_group *group)
556 {
557 struct timerqueue_node *node = NULL;
558 struct tmigr_event *evt = NULL;
559
560 lockdep_assert_held(&group->lock);
561
562 WRITE_ONCE(group->next_expiry, KTIME_MAX);
563
564 while ((node = timerqueue_getnext(&group->events))) {
565 evt = container_of(node, struct tmigr_event, nextevt);
566
567 if (!evt->ignore) {
568 WRITE_ONCE(group->next_expiry, evt->nextevt.expires);
569 return evt;
570 }
571
572 /*
573 * Remove next timers with ignore flag, because the group lock
574 * is held anyway
575 */
576 if (!timerqueue_del(&group->events, node))
577 break;
578 }
579
580 return NULL;
581 }
582
583 /*
584 * Return the next event (with the expiry equal or before @now)
585 *
586 * Event, which is returned, is also removed from the queue.
587 */
588 static struct tmigr_event *tmigr_next_expired_groupevt(struct tmigr_group *group,
589 u64 now)
590 {
591 struct tmigr_event *evt = tmigr_next_groupevt(group);
592
593 if (!evt || now < evt->nextevt.expires)
594 return NULL;
595
596 /*
597 * The event is ready to expire. Remove it and update next group event.
598 */
599 timerqueue_del(&group->events, &evt->nextevt);
600 tmigr_next_groupevt(group);
601
602 return evt;
603 }
604
605 static u64 tmigr_next_groupevt_expires(struct tmigr_group *group)
606 {
607 struct tmigr_event *evt;
608
609 evt = tmigr_next_groupevt(group);
610
611 if (!evt)
612 return KTIME_MAX;
613 else
614 return evt->nextevt.expires;
615 }
616
617 static bool tmigr_active_up(struct tmigr_group *group,
618 struct tmigr_group *child,
619 struct tmigr_walk *data)
620 {
621 union tmigr_state curstate, newstate;
622 bool walk_done;
623 u8 childmask;
624
625 childmask = data->childmask;
626 /*
627 * No memory barrier is required here in contrast to
628 * tmigr_inactive_up(), as the group state change does not depend on the
629 * child state.
630 */
631 curstate.state = atomic_read(&group->migr_state);
632
633 do {
634 newstate = curstate;
635 walk_done = true;
636
637 if (newstate.migrator == TMIGR_NONE) {
638 newstate.migrator = childmask;
639
640 /* Changes need to be propagated */
641 walk_done = false;
642 }
643
644 newstate.active |= childmask;
645 newstate.seq++;
646
647 } while (!atomic_try_cmpxchg(&group->migr_state, &curstate.state, newstate.state));
648
649 trace_tmigr_group_set_cpu_active(group, newstate, childmask);
650
651 /*
652 * The group is active (again). The group event might be still queued
653 * into the parent group's timerqueue but can now be handled by the
654 * migrator of this group. Therefore the ignore flag for the group event
655 * is updated to reflect this.
656 *
657 * The update of the ignore flag in the active path is done lockless. In
658 * worst case the migrator of the parent group observes the change too
659 * late and expires remotely all events belonging to this group. The
660 * lock is held while updating the ignore flag in idle path. So this
661 * state change will not be lost.
662 */
663 group->groupevt.ignore = true;
664
665 return walk_done;
666 }
667
668 static void __tmigr_cpu_activate(struct tmigr_cpu *tmc)
669 {
670 struct tmigr_walk data;
671
672 data.childmask = tmc->groupmask;
673
674 trace_tmigr_cpu_active(tmc);
675
676 tmc->cpuevt.ignore = true;
677 WRITE_ONCE(tmc->wakeup, KTIME_MAX);
678
679 walk_groups(&tmigr_active_up, &data, tmc);
680 }
681
682 /**
683 * tmigr_cpu_activate() - set this CPU active in timer migration hierarchy
684 *
685 * Call site timer_clear_idle() is called with interrupts disabled.
686 */
687 void tmigr_cpu_activate(void)
688 {
689 struct tmigr_cpu *tmc = this_cpu_ptr(&tmigr_cpu);
690
691 if (tmigr_is_not_available(tmc))
692 return;
693
694 if (WARN_ON_ONCE(!tmc->idle))
695 return;
696
697 raw_spin_lock(&tmc->lock);
698 tmc->idle = false;
699 __tmigr_cpu_activate(tmc);
700 raw_spin_unlock(&tmc->lock);
701 }
702
703 /*
704 * Returns true, if there is nothing to be propagated to the next level
705 *
706 * @data->firstexp is set to expiry of first gobal event of the (top level of
707 * the) hierarchy, but only when hierarchy is completely idle.
708 *
709 * The child and group states need to be read under the lock, to prevent a race
710 * against a concurrent tmigr_inactive_up() run when the last CPU goes idle. See
711 * also section "Prevent race between new event and last CPU going inactive" in
712 * the documentation at the top.
713 *
714 * This is the only place where the group event expiry value is set.
715 */
716 static
717 bool tmigr_update_events(struct tmigr_group *group, struct tmigr_group *child,
718 struct tmigr_walk *data)
719 {
720 struct tmigr_event *evt, *first_childevt;
721 union tmigr_state childstate, groupstate;
722 bool remote = data->remote;
723 bool walk_done = false;
724 u64 nextexp;
725
726 if (child) {
727 raw_spin_lock(&child->lock);
728 raw_spin_lock_nested(&group->lock, SINGLE_DEPTH_NESTING);
729
730 childstate.state = atomic_read(&child->migr_state);
731 groupstate.state = atomic_read(&group->migr_state);
732
733 if (childstate.active) {
734 walk_done = true;
735 goto unlock;
736 }
737
738 first_childevt = tmigr_next_groupevt(child);
739 nextexp = child->next_expiry;
740 evt = &child->groupevt;
741
742 evt->ignore = (nextexp == KTIME_MAX) ? true : false;
743 } else {
744 nextexp = data->nextexp;
745
746 first_childevt = evt = data->evt;
747
748 /*
749 * Walking the hierarchy is required in any case when a
750 * remote expiry was done before. This ensures to not lose
751 * already queued events in non active groups (see section
752 * "Required event and timerqueue update after a remote
753 * expiry" in the documentation at the top).
754 *
755 * The two call sites which are executed without a remote expiry
756 * before, are not prevented from propagating changes through
757 * the hierarchy by the return:
758 * - When entering this path by tmigr_new_timer(), @evt->ignore
759 * is never set.
760 * - tmigr_inactive_up() takes care of the propagation by
761 * itself and ignores the return value. But an immediate
762 * return is possible if there is a parent, sparing group
763 * locking at this level, because the upper walking call to
764 * the parent will take care about removing this event from
765 * within the group and update next_expiry accordingly.
766 *
767 * However if there is no parent, ie: the hierarchy has only a
768 * single level so @group is the top level group, make sure the
769 * first event information of the group is updated properly and
770 * also handled properly, so skip this fast return path.
771 */
772 if (evt->ignore && !remote && group->parent)
773 return true;
774
775 raw_spin_lock(&group->lock);
776
777 childstate.state = 0;
778 groupstate.state = atomic_read(&group->migr_state);
779 }
780
781 /*
782 * If the child event is already queued in the group, remove it from the
783 * queue when the expiry time changed only or when it could be ignored.
784 */
785 if (timerqueue_node_queued(&evt->nextevt)) {
786 if ((evt->nextevt.expires == nextexp) && !evt->ignore) {
787 /* Make sure not to miss a new CPU event with the same expiry */
788 evt->cpu = first_childevt->cpu;
789 goto check_toplvl;
790 }
791
792 if (!timerqueue_del(&group->events, &evt->nextevt))
793 WRITE_ONCE(group->next_expiry, KTIME_MAX);
794 }
795
796 if (evt->ignore) {
797 /*
798 * When the next child event could be ignored (nextexp is
799 * KTIME_MAX) and there was no remote timer handling before or
800 * the group is already active, there is no need to walk the
801 * hierarchy even if there is a parent group.
802 *
803 * The other way round: even if the event could be ignored, but
804 * if a remote timer handling was executed before and the group
805 * is not active, walking the hierarchy is required to not miss
806 * an enqueued timer in the non active group. The enqueued timer
807 * of the group needs to be propagated to a higher level to
808 * ensure it is handled.
809 */
810 if (!remote || groupstate.active)
811 walk_done = true;
812 } else {
813 evt->nextevt.expires = nextexp;
814 evt->cpu = first_childevt->cpu;
815
816 if (timerqueue_add(&group->events, &evt->nextevt))
817 WRITE_ONCE(group->next_expiry, nextexp);
818 }
819
820 check_toplvl:
821 if (!group->parent && (groupstate.migrator == TMIGR_NONE)) {
822 walk_done = true;
823
824 /*
825 * Nothing to do when update was done during remote timer
826 * handling. First timer in top level group which needs to be
827 * handled when top level group is not active, is calculated
828 * directly in tmigr_handle_remote_up().
829 */
830 if (remote)
831 goto unlock;
832
833 /*
834 * The top level group is idle and it has to be ensured the
835 * global timers are handled in time. (This could be optimized
836 * by keeping track of the last global scheduled event and only
837 * arming it on the CPU if the new event is earlier. Not sure if
838 * its worth the complexity.)
839 */
840 data->firstexp = tmigr_next_groupevt_expires(group);
841 }
842
843 trace_tmigr_update_events(child, group, childstate, groupstate,
844 nextexp);
845
846 unlock:
847 raw_spin_unlock(&group->lock);
848
849 if (child)
850 raw_spin_unlock(&child->lock);
851
852 return walk_done;
853 }
854
855 static bool tmigr_new_timer_up(struct tmigr_group *group,
856 struct tmigr_group *child,
857 struct tmigr_walk *data)
858 {
859 return tmigr_update_events(group, child, data);
860 }
861
862 /*
863 * Returns the expiry of the next timer that needs to be handled. KTIME_MAX is
864 * returned, if an active CPU will handle all the timer migration hierarchy
865 * timers.
866 */
867 static u64 tmigr_new_timer(struct tmigr_cpu *tmc, u64 nextexp)
868 {
869 struct tmigr_walk data = { .nextexp = nextexp,
870 .firstexp = KTIME_MAX,
871 .evt = &tmc->cpuevt };
872
873 lockdep_assert_held(&tmc->lock);
874
875 if (tmc->remote)
876 return KTIME_MAX;
877
878 trace_tmigr_cpu_new_timer(tmc);
879
880 tmc->cpuevt.ignore = false;
881 data.remote = false;
882
883 walk_groups(&tmigr_new_timer_up, &data, tmc);
884
885 /* If there is a new first global event, make sure it is handled */
886 return data.firstexp;
887 }
888
889 static void tmigr_handle_remote_cpu(unsigned int cpu, u64 now,
890 unsigned long jif)
891 {
892 struct timer_events tevt;
893 struct tmigr_walk data;
894 struct tmigr_cpu *tmc;
895
896 tmc = per_cpu_ptr(&tmigr_cpu, cpu);
897
898 raw_spin_lock_irq(&tmc->lock);
899
900 /*
901 * If the remote CPU is offline then the timers have been migrated to
902 * another CPU.
903 *
904 * If tmigr_cpu::remote is set, at the moment another CPU already
905 * expires the timers of the remote CPU.
906 *
907 * If tmigr_event::ignore is set, then the CPU returns from idle and
908 * takes care of its timers.
909 *
910 * If the next event expires in the future, then the event has been
911 * updated and there are no timers to expire right now. The CPU which
912 * updated the event takes care when hierarchy is completely
913 * idle. Otherwise the migrator does it as the event is enqueued.
914 */
915 if (!tmc->online || tmc->remote || tmc->cpuevt.ignore ||
916 now < tmc->cpuevt.nextevt.expires) {
917 raw_spin_unlock_irq(&tmc->lock);
918 return;
919 }
920
921 trace_tmigr_handle_remote_cpu(tmc);
922
923 tmc->remote = true;
924 WRITE_ONCE(tmc->wakeup, KTIME_MAX);
925
926 /* Drop the lock to allow the remote CPU to exit idle */
927 raw_spin_unlock_irq(&tmc->lock);
928
929 if (cpu != smp_processor_id())
930 timer_expire_remote(cpu);
931
932 /*
933 * Lock ordering needs to be preserved - timer_base locks before tmigr
934 * related locks (see section "Locking rules" in the documentation at
935 * the top). During fetching the next timer interrupt, also tmc->lock
936 * needs to be held. Otherwise there is a possible race window against
937 * the CPU itself when it comes out of idle, updates the first timer in
938 * the hierarchy and goes back to idle.
939 *
940 * timer base locks are dropped as fast as possible: After checking
941 * whether the remote CPU went offline in the meantime and after
942 * fetching the next remote timer interrupt. Dropping the locks as fast
943 * as possible keeps the locking region small and prevents holding
944 * several (unnecessary) locks during walking the hierarchy for updating
945 * the timerqueue and group events.
946 */
947 local_irq_disable();
948 timer_lock_remote_bases(cpu);
949 raw_spin_lock(&tmc->lock);
950
951 /*
952 * When the CPU went offline in the meantime, no hierarchy walk has to
953 * be done for updating the queued events, because the walk was
954 * already done during marking the CPU offline in the hierarchy.
955 *
956 * When the CPU is no longer idle, the CPU takes care of the timers and
957 * also of the timers in the hierarchy.
958 *
959 * (See also section "Required event and timerqueue update after a
960 * remote expiry" in the documentation at the top)
961 */
962 if (!tmc->online || !tmc->idle) {
963 timer_unlock_remote_bases(cpu);
964 goto unlock;
965 }
966
967 /* next event of CPU */
968 fetch_next_timer_interrupt_remote(jif, now, &tevt, cpu);
969 timer_unlock_remote_bases(cpu);
970
971 data.nextexp = tevt.global;
972 data.firstexp = KTIME_MAX;
973 data.evt = &tmc->cpuevt;
974 data.remote = true;
975
976 /*
977 * The update is done even when there is no 'new' global timer pending
978 * on the remote CPU (see section "Required event and timerqueue update
979 * after a remote expiry" in the documentation at the top)
980 */
981 walk_groups(&tmigr_new_timer_up, &data, tmc);
982
983 unlock:
984 tmc->remote = false;
985 raw_spin_unlock_irq(&tmc->lock);
986 }
987
988 static bool tmigr_handle_remote_up(struct tmigr_group *group,
989 struct tmigr_group *child,
990 struct tmigr_walk *data)
991 {
992 struct tmigr_event *evt;
993 unsigned long jif;
994 u8 childmask;
995 u64 now;
996
997 jif = data->basej;
998 now = data->now;
999
1000 childmask = data->childmask;
1001
1002 trace_tmigr_handle_remote(group);
1003 again:
1004 /*
1005 * Handle the group only if @childmask is the migrator or if the
1006 * group has no migrator. Otherwise the group is active and is
1007 * handled by its own migrator.
1008 */
1009 if (!tmigr_check_migrator(group, childmask))
1010 return true;
1011
1012 raw_spin_lock_irq(&group->lock);
1013
1014 evt = tmigr_next_expired_groupevt(group, now);
1015
1016 if (evt) {
1017 unsigned int remote_cpu = evt->cpu;
1018
1019 raw_spin_unlock_irq(&group->lock);
1020
1021 tmigr_handle_remote_cpu(remote_cpu, now, jif);
1022
1023 /* check if there is another event, that needs to be handled */
1024 goto again;
1025 }
1026
1027 /*
1028 * Keep track of the expiry of the first event that needs to be handled
1029 * (group->next_expiry was updated by tmigr_next_expired_groupevt(),
1030 * next was set by tmigr_handle_remote_cpu()).
1031 */
1032 data->firstexp = group->next_expiry;
1033
1034 raw_spin_unlock_irq(&group->lock);
1035
1036 return false;
1037 }
1038
1039 /**
1040 * tmigr_handle_remote() - Handle global timers of remote idle CPUs
1041 *
1042 * Called from the timer soft interrupt with interrupts enabled.
1043 */
1044 void tmigr_handle_remote(void)
1045 {
1046 struct tmigr_cpu *tmc = this_cpu_ptr(&tmigr_cpu);
1047 struct tmigr_walk data;
1048
1049 if (tmigr_is_not_available(tmc))
1050 return;
1051
1052 data.childmask = tmc->groupmask;
1053 data.firstexp = KTIME_MAX;
1054
1055 /*
1056 * NOTE: This is a doubled check because the migrator test will be done
1057 * in tmigr_handle_remote_up() anyway. Keep this check to speed up the
1058 * return when nothing has to be done.
1059 */
1060 if (!tmigr_check_migrator(tmc->tmgroup, tmc->groupmask)) {
1061 /*
1062 * If this CPU was an idle migrator, make sure to clear its wakeup
1063 * value so it won't chase timers that have already expired elsewhere.
1064 * This avoids endless requeue from tmigr_new_timer().
1065 */
1066 if (READ_ONCE(tmc->wakeup) == KTIME_MAX)
1067 return;
1068 }
1069
1070 data.now = get_jiffies_update(&data.basej);
1071
1072 /*
1073 * Update @tmc->wakeup only at the end and do not reset @tmc->wakeup to
1074 * KTIME_MAX. Even if tmc->lock is not held during the whole remote
1075 * handling, tmc->wakeup is fine to be stale as it is called in
1076 * interrupt context and tick_nohz_next_event() is executed in interrupt
1077 * exit path only after processing the last pending interrupt.
1078 */
1079
1080 __walk_groups(&tmigr_handle_remote_up, &data, tmc);
1081
1082 raw_spin_lock_irq(&tmc->lock);
1083 WRITE_ONCE(tmc->wakeup, data.firstexp);
1084 raw_spin_unlock_irq(&tmc->lock);
1085 }
1086
1087 static bool tmigr_requires_handle_remote_up(struct tmigr_group *group,
1088 struct tmigr_group *child,
1089 struct tmigr_walk *data)
1090 {
1091 u8 childmask;
1092
1093 childmask = data->childmask;
1094
1095 /*
1096 * Handle the group only if the child is the migrator or if the group
1097 * has no migrator. Otherwise the group is active and is handled by its
1098 * own migrator.
1099 */
1100 if (!tmigr_check_migrator(group, childmask))
1101 return true;
1102
1103 /*
1104 * When there is a parent group and the CPU which triggered the
1105 * hierarchy walk is not active, proceed the walk to reach the top level
1106 * group before reading the next_expiry value.
1107 */
1108 if (group->parent && !data->tmc_active)
1109 return false;
1110
1111 /*
1112 * The lock is required on 32bit architectures to read the variable
1113 * consistently with a concurrent writer. On 64bit the lock is not
1114 * required because the read operation is not split and so it is always
1115 * consistent.
1116 */
1117 if (IS_ENABLED(CONFIG_64BIT)) {
1118 data->firstexp = READ_ONCE(group->next_expiry);
1119 if (data->now >= data->firstexp) {
1120 data->check = true;
1121 return true;
1122 }
1123 } else {
1124 raw_spin_lock(&group->lock);
1125 data->firstexp = group->next_expiry;
1126 if (data->now >= group->next_expiry) {
1127 data->check = true;
1128 raw_spin_unlock(&group->lock);
1129 return true;
1130 }
1131 raw_spin_unlock(&group->lock);
1132 }
1133
1134 return false;
1135 }
1136
1137 /**
1138 * tmigr_requires_handle_remote() - Check the need of remote timer handling
1139 *
1140 * Must be called with interrupts disabled.
1141 */
1142 bool tmigr_requires_handle_remote(void)
1143 {
1144 struct tmigr_cpu *tmc = this_cpu_ptr(&tmigr_cpu);
1145 struct tmigr_walk data;
1146 unsigned long jif;
1147 bool ret = false;
1148
1149 if (tmigr_is_not_available(tmc))
1150 return ret;
1151
1152 data.now = get_jiffies_update(&jif);
1153 data.childmask = tmc->groupmask;
1154 data.firstexp = KTIME_MAX;
1155 data.tmc_active = !tmc->idle;
1156 data.check = false;
1157
1158 /*
1159 * If the CPU is active, walk the hierarchy to check whether a remote
1160 * expiry is required.
1161 *
1162 * Check is done lockless as interrupts are disabled and @tmc->idle is
1163 * set only by the local CPU.
1164 */
1165 if (!tmc->idle) {
1166 __walk_groups(&tmigr_requires_handle_remote_up, &data, tmc);
1167
1168 return data.check;
1169 }
1170
1171 /*
1172 * When the CPU is idle, compare @tmc->wakeup with @data.now. The lock
1173 * is required on 32bit architectures to read the variable consistently
1174 * with a concurrent writer. On 64bit the lock is not required because
1175 * the read operation is not split and so it is always consistent.
1176 */
1177 if (IS_ENABLED(CONFIG_64BIT)) {
1178 if (data.now >= READ_ONCE(tmc->wakeup))
1179 return true;
1180 } else {
1181 raw_spin_lock(&tmc->lock);
1182 if (data.now >= tmc->wakeup)
1183 ret = true;
1184 raw_spin_unlock(&tmc->lock);
1185 }
1186
1187 return ret;
1188 }
1189
1190 /**
1191 * tmigr_cpu_new_timer() - enqueue next global timer into hierarchy (idle tmc)
1192 * @nextexp: Next expiry of global timer (or KTIME_MAX if not)
1193 *
1194 * The CPU is already deactivated in the timer migration
1195 * hierarchy. tick_nohz_get_sleep_length() calls tick_nohz_next_event()
1196 * and thereby the timer idle path is executed once more. @tmc->wakeup
1197 * holds the first timer, when the timer migration hierarchy is
1198 * completely idle.
1199 *
1200 * Returns the first timer that needs to be handled by this CPU or KTIME_MAX if
1201 * nothing needs to be done.
1202 */
1203 u64 tmigr_cpu_new_timer(u64 nextexp)
1204 {
1205 struct tmigr_cpu *tmc = this_cpu_ptr(&tmigr_cpu);
1206 u64 ret;
1207
1208 if (tmigr_is_not_available(tmc))
1209 return nextexp;
1210
1211 raw_spin_lock(&tmc->lock);
1212
1213 ret = READ_ONCE(tmc->wakeup);
1214 if (nextexp != KTIME_MAX) {
1215 if (nextexp != tmc->cpuevt.nextevt.expires ||
1216 tmc->cpuevt.ignore) {
1217 ret = tmigr_new_timer(tmc, nextexp);
1218 /*
1219 * Make sure the reevaluation of timers in idle path
1220 * will not miss an event.
1221 */
1222 WRITE_ONCE(tmc->wakeup, ret);
1223 }
1224 }
1225 trace_tmigr_cpu_new_timer_idle(tmc, nextexp);
1226 raw_spin_unlock(&tmc->lock);
1227 return ret;
1228 }
1229
1230 static bool tmigr_inactive_up(struct tmigr_group *group,
1231 struct tmigr_group *child,
1232 struct tmigr_walk *data)
1233 {
1234 union tmigr_state curstate, newstate, childstate;
1235 bool walk_done;
1236 u8 childmask;
1237
1238 childmask = data->childmask;
1239 childstate.state = 0;
1240
1241 /*
1242 * The memory barrier is paired with the cmpxchg() in tmigr_active_up()
1243 * to make sure the updates of child and group states are ordered. The
1244 * ordering is mandatory, as the group state change depends on the child
1245 * state.
1246 */
1247 curstate.state = atomic_read_acquire(&group->migr_state);
1248
1249 for (;;) {
1250 if (child)
1251 childstate.state = atomic_read(&child->migr_state);
1252
1253 newstate = curstate;
1254 walk_done = true;
1255
1256 /* Reset active bit when the child is no longer active */
1257 if (!childstate.active)
1258 newstate.active &= ~childmask;
1259
1260 if (newstate.migrator == childmask) {
1261 /*
1262 * Find a new migrator for the group, because the child
1263 * group is idle!
1264 */
1265 if (!childstate.active) {
1266 unsigned long new_migr_bit, active = newstate.active;
1267
1268 new_migr_bit = find_first_bit(&active, BIT_CNT);
1269
1270 if (new_migr_bit != BIT_CNT) {
1271 newstate.migrator = BIT(new_migr_bit);
1272 } else {
1273 newstate.migrator = TMIGR_NONE;
1274
1275 /* Changes need to be propagated */
1276 walk_done = false;
1277 }
1278 }
1279 }
1280
1281 newstate.seq++;
1282
1283 WARN_ON_ONCE((newstate.migrator != TMIGR_NONE) && !(newstate.active));
1284
1285 if (atomic_try_cmpxchg(&group->migr_state, &curstate.state, newstate.state)) {
1286 trace_tmigr_group_set_cpu_inactive(group, newstate, childmask);
1287 break;
1288 }
1289
1290 /*
1291 * The memory barrier is paired with the cmpxchg() in
1292 * tmigr_active_up() to make sure the updates of child and group
1293 * states are ordered. It is required only when the above
1294 * try_cmpxchg() fails.
1295 */
1296 smp_mb__after_atomic();
1297 }
1298
1299 data->remote = false;
1300
1301 /* Event Handling */
1302 tmigr_update_events(group, child, data);
1303
1304 return walk_done;
1305 }
1306
1307 static u64 __tmigr_cpu_deactivate(struct tmigr_cpu *tmc, u64 nextexp)
1308 {
1309 struct tmigr_walk data = { .nextexp = nextexp,
1310 .firstexp = KTIME_MAX,
1311 .evt = &tmc->cpuevt,
1312 .childmask = tmc->groupmask };
1313
1314 /*
1315 * If nextexp is KTIME_MAX, the CPU event will be ignored because the
1316 * local timer expires before the global timer, no global timer is set
1317 * or CPU goes offline.
1318 */
1319 if (nextexp != KTIME_MAX)
1320 tmc->cpuevt.ignore = false;
1321
1322 walk_groups(&tmigr_inactive_up, &data, tmc);
1323 return data.firstexp;
1324 }
1325
1326 /**
1327 * tmigr_cpu_deactivate() - Put current CPU into inactive state
1328 * @nextexp: The next global timer expiry of the current CPU
1329 *
1330 * Must be called with interrupts disabled.
1331 *
1332 * Return: the next event expiry of the current CPU or the next event expiry
1333 * from the hierarchy if this CPU is the top level migrator or the hierarchy is
1334 * completely idle.
1335 */
1336 u64 tmigr_cpu_deactivate(u64 nextexp)
1337 {
1338 struct tmigr_cpu *tmc = this_cpu_ptr(&tmigr_cpu);
1339 u64 ret;
1340
1341 if (tmigr_is_not_available(tmc))
1342 return nextexp;
1343
1344 raw_spin_lock(&tmc->lock);
1345
1346 ret = __tmigr_cpu_deactivate(tmc, nextexp);
1347
1348 tmc->idle = true;
1349
1350 /*
1351 * Make sure the reevaluation of timers in idle path will not miss an
1352 * event.
1353 */
1354 WRITE_ONCE(tmc->wakeup, ret);
1355
1356 trace_tmigr_cpu_idle(tmc, nextexp);
1357 raw_spin_unlock(&tmc->lock);
1358 return ret;
1359 }
1360
1361 /**
1362 * tmigr_quick_check() - Quick forecast of next tmigr event when CPU wants to
1363 * go idle
1364 * @nextevt: The next global timer expiry of the current CPU
1365 *
1366 * Return:
1367 * * KTIME_MAX - when it is probable that nothing has to be done (not
1368 * the only one in the level 0 group; and if it is the
1369 * only one in level 0 group, but there are more than a
1370 * single group active on the way to top level)
1371 * * nextevt - when CPU is offline and has to handle timer on its own
1372 * or when on the way to top in every group only a single
1373 * child is active but @nextevt is before the lowest
1374 * next_expiry encountered while walking up to top level.
1375 * * next_expiry - value of lowest expiry encountered while walking groups
1376 * if only a single child is active on each and @nextevt
1377 * is after this lowest expiry.
1378 */
1379 u64 tmigr_quick_check(u64 nextevt)
1380 {
1381 struct tmigr_cpu *tmc = this_cpu_ptr(&tmigr_cpu);
1382 struct tmigr_group *group = tmc->tmgroup;
1383
1384 if (tmigr_is_not_available(tmc))
1385 return nextevt;
1386
1387 if (WARN_ON_ONCE(tmc->idle))
1388 return nextevt;
1389
1390 if (!tmigr_check_migrator_and_lonely(tmc->tmgroup, tmc->groupmask))
1391 return KTIME_MAX;
1392
1393 do {
1394 if (!tmigr_check_lonely(group)) {
1395 return KTIME_MAX;
1396 } else {
1397 /*
1398 * Since current CPU is active, events may not be sorted
1399 * from bottom to the top because the CPU's event is ignored
1400 * up to the top and its sibling's events not propagated upwards.
1401 * Thus keep track of the lowest observed expiry.
1402 */
1403 nextevt = min_t(u64, nextevt, READ_ONCE(group->next_expiry));
1404 if (!group->parent)
1405 return nextevt;
1406 }
1407 group = group->parent;
1408 } while (group);
1409
1410 return KTIME_MAX;
1411 }
1412
1413 /*
1414 * tmigr_trigger_active() - trigger a CPU to become active again
1415 *
1416 * This function is executed on a CPU which is part of cpu_online_mask, when the
1417 * last active CPU in the hierarchy is offlining. With this, it is ensured that
1418 * the other CPU is active and takes over the migrator duty.
1419 */
1420 static long tmigr_trigger_active(void *unused)
1421 {
1422 struct tmigr_cpu *tmc = this_cpu_ptr(&tmigr_cpu);
1423
1424 WARN_ON_ONCE(!tmc->online || tmc->idle);
1425
1426 return 0;
1427 }
1428
1429 static int tmigr_cpu_offline(unsigned int cpu)
1430 {
1431 struct tmigr_cpu *tmc = this_cpu_ptr(&tmigr_cpu);
1432 int migrator;
1433 u64 firstexp;
1434
1435 raw_spin_lock_irq(&tmc->lock);
1436 tmc->online = false;
1437 WRITE_ONCE(tmc->wakeup, KTIME_MAX);
1438
1439 /*
1440 * CPU has to handle the local events on his own, when on the way to
1441 * offline; Therefore nextevt value is set to KTIME_MAX
1442 */
1443 firstexp = __tmigr_cpu_deactivate(tmc, KTIME_MAX);
1444 trace_tmigr_cpu_offline(tmc);
1445 raw_spin_unlock_irq(&tmc->lock);
1446
1447 if (firstexp != KTIME_MAX) {
1448 migrator = cpumask_any_but(cpu_online_mask, cpu);
1449 work_on_cpu(migrator, tmigr_trigger_active, NULL);
1450 }
1451
1452 return 0;
1453 }
1454
1455 static int tmigr_cpu_online(unsigned int cpu)
1456 {
1457 struct tmigr_cpu *tmc = this_cpu_ptr(&tmigr_cpu);
1458
1459 /* Check whether CPU data was successfully initialized */
1460 if (WARN_ON_ONCE(!tmc->tmgroup))
1461 return -EINVAL;
1462
1463 raw_spin_lock_irq(&tmc->lock);
1464 trace_tmigr_cpu_online(tmc);
1465 tmc->idle = timer_base_is_idle();
1466 if (!tmc->idle)
1467 __tmigr_cpu_activate(tmc);
1468 tmc->online = true;
1469 raw_spin_unlock_irq(&tmc->lock);
1470 return 0;
1471 }
1472
1473 static void tmigr_init_group(struct tmigr_group *group, unsigned int lvl,
1474 int node)
1475 {
1476 union tmigr_state s;
1477
1478 raw_spin_lock_init(&group->lock);
1479
1480 group->level = lvl;
1481 group->numa_node = lvl < tmigr_crossnode_level ? node : NUMA_NO_NODE;
1482
1483 group->num_children = 0;
1484
1485 s.migrator = TMIGR_NONE;
1486 s.active = 0;
1487 s.seq = 0;
1488 atomic_set(&group->migr_state, s.state);
1489
1490 timerqueue_init_head(&group->events);
1491 timerqueue_init(&group->groupevt.nextevt);
1492 group->groupevt.nextevt.expires = KTIME_MAX;
1493 WRITE_ONCE(group->next_expiry, KTIME_MAX);
1494 group->groupevt.ignore = true;
1495 }
1496
1497 static struct tmigr_group *tmigr_get_group(unsigned int cpu, int node,
1498 unsigned int lvl)
1499 {
1500 struct tmigr_group *tmp, *group = NULL;
1501
1502 lockdep_assert_held(&tmigr_mutex);
1503
1504 /* Try to attach to an existing group first */
1505 list_for_each_entry(tmp, &tmigr_level_list[lvl], list) {
1506 /*
1507 * If @lvl is below the cross NUMA node level, check whether
1508 * this group belongs to the same NUMA node.
1509 */
1510 if (lvl < tmigr_crossnode_level && tmp->numa_node != node)
1511 continue;
1512
1513 /* Capacity left? */
1514 if (tmp->num_children >= TMIGR_CHILDREN_PER_GROUP)
1515 continue;
1516
1517 /*
1518 * TODO: A possible further improvement: Make sure that all CPU
1519 * siblings end up in the same group of the lowest level of the
1520 * hierarchy. Rely on the topology sibling mask would be a
1521 * reasonable solution.
1522 */
1523
1524 group = tmp;
1525 break;
1526 }
1527
1528 if (group)
1529 return group;
1530
1531 /* Allocate and set up a new group */
1532 group = kzalloc_node(sizeof(*group), GFP_KERNEL, node);
1533 if (!group)
1534 return ERR_PTR(-ENOMEM);
1535
1536 tmigr_init_group(group, lvl, node);
1537
1538 /* Setup successful. Add it to the hierarchy */
1539 list_add(&group->list, &tmigr_level_list[lvl]);
1540 trace_tmigr_group_set(group);
1541 return group;
1542 }
1543
1544 static void tmigr_connect_child_parent(struct tmigr_group *child,
1545 struct tmigr_group *parent,
1546 bool activate)
1547 {
1548 struct tmigr_walk data;
1549
1550 raw_spin_lock_irq(&child->lock);
1551 raw_spin_lock_nested(&parent->lock, SINGLE_DEPTH_NESTING);
1552
1553 child->parent = parent;
1554 child->groupmask = BIT(parent->num_children++);
1555
1556 raw_spin_unlock(&parent->lock);
1557 raw_spin_unlock_irq(&child->lock);
1558
1559 trace_tmigr_connect_child_parent(child);
1560
1561 if (!activate)
1562 return;
1563
1564 /*
1565 * To prevent inconsistent states, active children need to be active in
1566 * the new parent as well. Inactive children are already marked inactive
1567 * in the parent group:
1568 *
1569 * * When new groups were created by tmigr_setup_groups() starting from
1570 * the lowest level (and not higher then one level below the current
1571 * top level), then they are not active. They will be set active when
1572 * the new online CPU comes active.
1573 *
1574 * * But if a new group above the current top level is required, it is
1575 * mandatory to propagate the active state of the already existing
1576 * child to the new parent. So tmigr_connect_child_parent() is
1577 * executed with the formerly top level group (child) and the newly
1578 * created group (parent).
1579 *
1580 * * It is ensured that the child is active, as this setup path is
1581 * executed in hotplug prepare callback. This is exectued by an
1582 * already connected and !idle CPU. Even if all other CPUs go idle,
1583 * the CPU executing the setup will be responsible up to current top
1584 * level group. And the next time it goes inactive, it will release
1585 * the new childmask and parent to subsequent walkers through this
1586 * @child. Therefore propagate active state unconditionally.
1587 */
1588 data.childmask = child->groupmask;
1589
1590 /*
1591 * There is only one new level per time (which is protected by
1592 * tmigr_mutex). When connecting the child and the parent and set the
1593 * child active when the parent is inactive, the parent needs to be the
1594 * uppermost level. Otherwise there went something wrong!
1595 */
1596 WARN_ON(!tmigr_active_up(parent, child, &data) && parent->parent);
1597 }
1598
1599 static int tmigr_setup_groups(unsigned int cpu, unsigned int node)
1600 {
1601 struct tmigr_group *group, *child, **stack;
1602 int top = 0, err = 0, i = 0;
1603 struct list_head *lvllist;
1604
1605 stack = kcalloc(tmigr_hierarchy_levels, sizeof(*stack), GFP_KERNEL);
1606 if (!stack)
1607 return -ENOMEM;
1608
1609 do {
1610 group = tmigr_get_group(cpu, node, i);
1611 if (IS_ERR(group)) {
1612 err = PTR_ERR(group);
1613 break;
1614 }
1615
1616 top = i;
1617 stack[i++] = group;
1618
1619 /*
1620 * When booting only less CPUs of a system than CPUs are
1621 * available, not all calculated hierarchy levels are required.
1622 *
1623 * The loop is aborted as soon as the highest level, which might
1624 * be different from tmigr_hierarchy_levels, contains only a
1625 * single group.
1626 */
1627 if (group->parent || i == tmigr_hierarchy_levels ||
1628 (list_empty(&tmigr_level_list[i]) &&
1629 list_is_singular(&tmigr_level_list[i - 1])))
1630 break;
1631
1632 } while (i < tmigr_hierarchy_levels);
1633
1634 while (i > 0) {
1635 group = stack[--i];
1636
1637 if (err < 0) {
1638 list_del(&group->list);
1639 kfree(group);
1640 continue;
1641 }
1642
1643 WARN_ON_ONCE(i != group->level);
1644
1645 /*
1646 * Update tmc -> group / child -> group connection
1647 */
1648 if (i == 0) {
1649 struct tmigr_cpu *tmc = per_cpu_ptr(&tmigr_cpu, cpu);
1650
1651 raw_spin_lock_irq(&group->lock);
1652
1653 tmc->tmgroup = group;
1654 tmc->groupmask = BIT(group->num_children++);
1655
1656 raw_spin_unlock_irq(&group->lock);
1657
1658 trace_tmigr_connect_cpu_parent(tmc);
1659
1660 /* There are no children that need to be connected */
1661 continue;
1662 } else {
1663 child = stack[i - 1];
1664 /* Will be activated at online time */
1665 tmigr_connect_child_parent(child, group, false);
1666 }
1667
1668 /* check if uppermost level was newly created */
1669 if (top != i)
1670 continue;
1671
1672 WARN_ON_ONCE(top == 0);
1673
1674 lvllist = &tmigr_level_list[top];
1675 if (group->num_children == 1 && list_is_singular(lvllist)) {
1676 /*
1677 * The target CPU must never do the prepare work, except
1678 * on early boot when the boot CPU is the target. Otherwise
1679 * it may spuriously activate the old top level group inside
1680 * the new one (nevertheless whether old top level group is
1681 * active or not) and/or release an uninitialized childmask.
1682 */
1683 WARN_ON_ONCE(cpu == raw_smp_processor_id());
1684
1685 lvllist = &tmigr_level_list[top - 1];
1686 list_for_each_entry(child, lvllist, list) {
1687 if (child->parent)
1688 continue;
1689
1690 tmigr_connect_child_parent(child, group, true);
1691 }
1692 }
1693 }
1694
1695 kfree(stack);
1696
1697 return err;
1698 }
1699
1700 static int tmigr_add_cpu(unsigned int cpu)
1701 {
1702 int node = cpu_to_node(cpu);
1703 int ret;
1704
1705 mutex_lock(&tmigr_mutex);
1706 ret = tmigr_setup_groups(cpu, node);
1707 mutex_unlock(&tmigr_mutex);
1708
1709 return ret;
1710 }
1711
1712 static int tmigr_cpu_prepare(unsigned int cpu)
1713 {
1714 struct tmigr_cpu *tmc = per_cpu_ptr(&tmigr_cpu, cpu);
1715 int ret = 0;
1716
1717 /* Not first online attempt? */
1718 if (tmc->tmgroup)
1719 return ret;
1720
1721 raw_spin_lock_init(&tmc->lock);
1722 timerqueue_init(&tmc->cpuevt.nextevt);
1723 tmc->cpuevt.nextevt.expires = KTIME_MAX;
1724 tmc->cpuevt.ignore = true;
1725 tmc->cpuevt.cpu = cpu;
1726 tmc->remote = false;
1727 WRITE_ONCE(tmc->wakeup, KTIME_MAX);
1728
1729 ret = tmigr_add_cpu(cpu);
1730 if (ret < 0)
1731 return ret;
1732
1733 if (tmc->groupmask == 0)
1734 return -EINVAL;
1735
1736 return ret;
1737 }
1738
1739 static int __init tmigr_init(void)
1740 {
1741 unsigned int cpulvl, nodelvl, cpus_per_node, i;
1742 unsigned int nnodes = num_possible_nodes();
1743 unsigned int ncpus = num_possible_cpus();
1744 int ret = -ENOMEM;
1745
1746 BUILD_BUG_ON_NOT_POWER_OF_2(TMIGR_CHILDREN_PER_GROUP);
1747
1748 /* Nothing to do if running on UP */
1749 if (ncpus == 1)
1750 return 0;
1751
1752 /*
1753 * Calculate the required hierarchy levels. Unfortunately there is no
1754 * reliable information available, unless all possible CPUs have been
1755 * brought up and all NUMA nodes are populated.
1756 *
1757 * Estimate the number of levels with the number of possible nodes and
1758 * the number of possible CPUs. Assume CPUs are spread evenly across
1759 * nodes. We cannot rely on cpumask_of_node() because it only works for
1760 * online CPUs.
1761 */
1762 cpus_per_node = DIV_ROUND_UP(ncpus, nnodes);
1763
1764 /* Calc the hierarchy levels required to hold the CPUs of a node */
1765 cpulvl = DIV_ROUND_UP(order_base_2(cpus_per_node),
1766 ilog2(TMIGR_CHILDREN_PER_GROUP));
1767
1768 /* Calculate the extra levels to connect all nodes */
1769 nodelvl = DIV_ROUND_UP(order_base_2(nnodes),
1770 ilog2(TMIGR_CHILDREN_PER_GROUP));
1771
1772 tmigr_hierarchy_levels = cpulvl + nodelvl;
1773
1774 /*
1775 * If a NUMA node spawns more than one CPU level group then the next
1776 * level(s) of the hierarchy contains groups which handle all CPU groups
1777 * of the same NUMA node. The level above goes across NUMA nodes. Store
1778 * this information for the setup code to decide in which level node
1779 * matching is no longer required.
1780 */
1781 tmigr_crossnode_level = cpulvl;
1782
1783 tmigr_level_list = kcalloc(tmigr_hierarchy_levels, sizeof(struct list_head), GFP_KERNEL);
1784 if (!tmigr_level_list)
1785 goto err;
1786
1787 for (i = 0; i < tmigr_hierarchy_levels; i++)
1788 INIT_LIST_HEAD(&tmigr_level_list[i]);
1789
1790 pr_info("Timer migration: %d hierarchy levels; %d children per group;"
1791 " %d crossnode level\n",
1792 tmigr_hierarchy_levels, TMIGR_CHILDREN_PER_GROUP,
1793 tmigr_crossnode_level);
1794
1795 ret = cpuhp_setup_state(CPUHP_TMIGR_PREPARE, "tmigr:prepare",
1796 tmigr_cpu_prepare, NULL);
1797 if (ret)
1798 goto err;
1799
1800 ret = cpuhp_setup_state(CPUHP_AP_TMIGR_ONLINE, "tmigr:online",
1801 tmigr_cpu_online, tmigr_cpu_offline);
1802 if (ret)
1803 goto err;
1804
1805 return 0;
1806
1807 err:
1808 pr_err("Timer migration setup failed\n");
1809 return ret;
1810 }
1811 early_initcall(tmigr_init);
1812
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.