ww_mutex.h

  1/* SPDX-License-Identifier: GPL-2.0-only */
  2
  3#ifndef WW_RT
  4
  5#define MUTEX		mutex
  6#define MUTEX_WAITER	mutex_waiter
  7
  8static inline struct mutex_waiter *
  9__ww_waiter_first(struct mutex *lock)
 10{
 11	struct mutex_waiter *w;
 12
 13	w = list_first_entry(&lock->wait_list, struct mutex_waiter, list);
 14	if (list_entry_is_head(w, &lock->wait_list, list))
 15		return NULL;
 16
 17	return w;
 18}
 19
 20static inline struct mutex_waiter *
 21__ww_waiter_next(struct mutex *lock, struct mutex_waiter *w)
 22{
 23	w = list_next_entry(w, list);
 24	if (list_entry_is_head(w, &lock->wait_list, list))
 25		return NULL;
 26
 27	return w;
 28}
 29
 30static inline struct mutex_waiter *
 31__ww_waiter_prev(struct mutex *lock, struct mutex_waiter *w)
 32{
 33	w = list_prev_entry(w, list);
 34	if (list_entry_is_head(w, &lock->wait_list, list))
 35		return NULL;
 36
 37	return w;
 38}
 39
 40static inline struct mutex_waiter *
 41__ww_waiter_last(struct mutex *lock)
 42{
 43	struct mutex_waiter *w;
 44
 45	w = list_last_entry(&lock->wait_list, struct mutex_waiter, list);
 46	if (list_entry_is_head(w, &lock->wait_list, list))
 47		return NULL;
 48
 49	return w;
 50}
 51
 52static inline void
 53__ww_waiter_add(struct mutex *lock, struct mutex_waiter *waiter, struct mutex_waiter *pos)
 54{
 55	struct list_head *p = &lock->wait_list;
 56	if (pos)
 57		p = &pos->list;
 58	__mutex_add_waiter(lock, waiter, p);
 59}
 60
 61static inline struct task_struct *
 62__ww_mutex_owner(struct mutex *lock)
 63{
 64	return __mutex_owner(lock);
 65}
 66
 67static inline bool
 68__ww_mutex_has_waiters(struct mutex *lock)
 69{
 70	return atomic_long_read(&lock->owner) & MUTEX_FLAG_WAITERS;
 71}
 72
 73static inline void lock_wait_lock(struct mutex *lock, unsigned long *flags)
 74{
 75	raw_spin_lock_irqsave(&lock->wait_lock, *flags);
 76}
 77
 78static inline void unlock_wait_lock(struct mutex *lock, unsigned long *flags)
 79{
 80	raw_spin_unlock_irqrestore(&lock->wait_lock, *flags);
 81}
 82
 83static inline void lockdep_assert_wait_lock_held(struct mutex *lock)
 84{
 85	lockdep_assert_held(&lock->wait_lock);
 86}
 87
 88#else /* WW_RT */
 89
 90#define MUTEX		rt_mutex
 91#define MUTEX_WAITER	rt_mutex_waiter
 92
 93static inline struct rt_mutex_waiter *
 94__ww_waiter_first(struct rt_mutex *lock)
 95{
 96	struct rb_node *n = rb_first(&lock->rtmutex.waiters.rb_root);
 97	if (!n)
 98		return NULL;
 99	return rb_entry(n, struct rt_mutex_waiter, tree.entry);
100}
101
102static inline struct rt_mutex_waiter *
103__ww_waiter_next(struct rt_mutex *lock, struct rt_mutex_waiter *w)
104{
105	struct rb_node *n = rb_next(&w->tree.entry);
106	if (!n)
107		return NULL;
108	return rb_entry(n, struct rt_mutex_waiter, tree.entry);
109}
110
111static inline struct rt_mutex_waiter *
112__ww_waiter_prev(struct rt_mutex *lock, struct rt_mutex_waiter *w)
113{
114	struct rb_node *n = rb_prev(&w->tree.entry);
115	if (!n)
116		return NULL;
117	return rb_entry(n, struct rt_mutex_waiter, tree.entry);
118}
119
120static inline struct rt_mutex_waiter *
121__ww_waiter_last(struct rt_mutex *lock)
122{
123	struct rb_node *n = rb_last(&lock->rtmutex.waiters.rb_root);
124	if (!n)
125		return NULL;
126	return rb_entry(n, struct rt_mutex_waiter, tree.entry);
127}
128
129static inline void
130__ww_waiter_add(struct rt_mutex *lock, struct rt_mutex_waiter *waiter, struct rt_mutex_waiter *pos)
131{
132	/* RT unconditionally adds the waiter first and then removes it on error */
133}
134
135static inline struct task_struct *
136__ww_mutex_owner(struct rt_mutex *lock)
137{
138	return rt_mutex_owner(&lock->rtmutex);
139}
140
141static inline bool
142__ww_mutex_has_waiters(struct rt_mutex *lock)
143{
144	return rt_mutex_has_waiters(&lock->rtmutex);
145}
146
147static inline void lock_wait_lock(struct rt_mutex *lock, unsigned long *flags)
148{
149	raw_spin_lock_irqsave(&lock->rtmutex.wait_lock, *flags);
150}
151
152static inline void unlock_wait_lock(struct rt_mutex *lock, unsigned long *flags)
153{
154	raw_spin_unlock_irqrestore(&lock->rtmutex.wait_lock, *flags);
155}
156
157static inline void lockdep_assert_wait_lock_held(struct rt_mutex *lock)
158{
159	lockdep_assert_held(&lock->rtmutex.wait_lock);
160}
161
162#endif /* WW_RT */
163
164/*
165 * Wait-Die:
166 *   The newer transactions are killed when:
167 *     It (the new transaction) makes a request for a lock being held
168 *     by an older transaction.
169 *
170 * Wound-Wait:
171 *   The newer transactions are wounded when:
172 *     An older transaction makes a request for a lock being held by
173 *     the newer transaction.
174 */
175
176/*
177 * Associate the ww_mutex @ww with the context @ww_ctx under which we acquired
178 * it.
179 */
180static __always_inline void
181ww_mutex_lock_acquired(struct ww_mutex *ww, struct ww_acquire_ctx *ww_ctx)
182{
183#ifdef DEBUG_WW_MUTEXES
184	/*
185	 * If this WARN_ON triggers, you used ww_mutex_lock to acquire,
186	 * but released with a normal mutex_unlock in this call.
187	 *
188	 * This should never happen, always use ww_mutex_unlock.
189	 */
190	DEBUG_LOCKS_WARN_ON(ww->ctx);
191
192	/*
193	 * Not quite done after calling ww_acquire_done() ?
194	 */
195	DEBUG_LOCKS_WARN_ON(ww_ctx->done_acquire);
196
197	if (ww_ctx->contending_lock) {
198		/*
199		 * After -EDEADLK you tried to
200		 * acquire a different ww_mutex? Bad!
201		 */
202		DEBUG_LOCKS_WARN_ON(ww_ctx->contending_lock != ww);
203
204		/*
205		 * You called ww_mutex_lock after receiving -EDEADLK,
206		 * but 'forgot' to unlock everything else first?
207		 */
208		DEBUG_LOCKS_WARN_ON(ww_ctx->acquired > 0);
209		ww_ctx->contending_lock = NULL;
210	}
211
212	/*
213	 * Naughty, using a different class will lead to undefined behavior!
214	 */
215	DEBUG_LOCKS_WARN_ON(ww_ctx->ww_class != ww->ww_class);
216#endif
217	ww_ctx->acquired++;
218	ww->ctx = ww_ctx;
219}
220
221/*
222 * Determine if @a is 'less' than @b. IOW, either @a is a lower priority task
223 * or, when of equal priority, a younger transaction than @b.
224 *
225 * Depending on the algorithm, @a will either need to wait for @b, or die.
226 */
227static inline bool
228__ww_ctx_less(struct ww_acquire_ctx *a, struct ww_acquire_ctx *b)
229{
230/*
231 * Can only do the RT prio for WW_RT, because task->prio isn't stable due to PI,
232 * so the wait_list ordering will go wobbly. rt_mutex re-queues the waiter and
233 * isn't affected by this.
234 */
235#ifdef WW_RT
236	/* kernel prio; less is more */
237	int a_prio = a->task->prio;
238	int b_prio = b->task->prio;
239
240	if (rt_or_dl_prio(a_prio) || rt_or_dl_prio(b_prio)) {
241
242		if (a_prio > b_prio)
243			return true;
244
245		if (a_prio < b_prio)
246			return false;
247
248		/* equal static prio */
249
250		if (dl_prio(a_prio)) {
251			if (dl_time_before(b->task->dl.deadline,
252					   a->task->dl.deadline))
253				return true;
254
255			if (dl_time_before(a->task->dl.deadline,
256					   b->task->dl.deadline))
257				return false;
258		}
259
260		/* equal prio */
261	}
262#endif
263
264	/* FIFO order tie break -- bigger is younger */
265	return (signed long)(a->stamp - b->stamp) > 0;
266}
267
268/*
269 * Wait-Die; wake a lesser waiter context (when locks held) such that it can
270 * die.
271 *
272 * Among waiters with context, only the first one can have other locks acquired
273 * already (ctx->acquired > 0), because __ww_mutex_add_waiter() and
274 * __ww_mutex_check_kill() wake any but the earliest context.
275 */
276static bool
277__ww_mutex_die(struct MUTEX *lock, struct MUTEX_WAITER *waiter,
278	       struct ww_acquire_ctx *ww_ctx, struct wake_q_head *wake_q)
279{
280	if (!ww_ctx->is_wait_die)
281		return false;
282
283	if (waiter->ww_ctx->acquired > 0 && __ww_ctx_less(waiter->ww_ctx, ww_ctx)) {
284#ifndef WW_RT
285		debug_mutex_wake_waiter(lock, waiter);
286#endif
287		/*
288		 * When waking up the task to die, be sure to clear the
289		 * blocked_on pointer. Otherwise we can see circular
290		 * blocked_on relationships that can't resolve.
291		 */
292		__clear_task_blocked_on(waiter->task, lock);
293		wake_q_add(wake_q, waiter->task);
294	}
295
296	return true;
297}
298
299/*
300 * Wound-Wait; wound a lesser @hold_ctx if it holds the lock.
301 *
302 * Wound the lock holder if there are waiters with more important transactions
303 * than the lock holders. Even if multiple waiters may wound the lock holder,
304 * it's sufficient that only one does.
305 */
306static bool __ww_mutex_wound(struct MUTEX *lock,
307			     struct ww_acquire_ctx *ww_ctx,
308			     struct ww_acquire_ctx *hold_ctx,
309			     struct wake_q_head *wake_q)
310{
311	struct task_struct *owner = __ww_mutex_owner(lock);
312
313	lockdep_assert_wait_lock_held(lock);
314
315	/*
316	 * Possible through __ww_mutex_add_waiter() when we race with
317	 * ww_mutex_set_context_fastpath(). In that case we'll get here again
318	 * through __ww_mutex_check_waiters().
319	 */
320	if (!hold_ctx)
321		return false;
322
323	/*
324	 * Can have !owner because of __mutex_unlock_slowpath(), but if owner,
325	 * it cannot go away because we'll have FLAG_WAITERS set and hold
326	 * wait_lock.
327	 */
328	if (!owner)
329		return false;
330
331	if (ww_ctx->acquired > 0 && __ww_ctx_less(hold_ctx, ww_ctx)) {
332		hold_ctx->wounded = 1;
333
334		/*
335		 * wake_up_process() paired with set_current_state()
336		 * inserts sufficient barriers to make sure @owner either sees
337		 * it's wounded in __ww_mutex_check_kill() or has a
338		 * wakeup pending to re-read the wounded state.
339		 */
340		if (owner != current) {
341			/*
342			 * When waking up the task to wound, be sure to clear the
343			 * blocked_on pointer. Otherwise we can see circular
344			 * blocked_on relationships that can't resolve.
345			 *
346			 * NOTE: We pass NULL here instead of lock, because we
347			 * are waking the mutex owner, who may be currently
348			 * blocked on a different mutex.
349			 */
350			__clear_task_blocked_on(owner, NULL);
351			wake_q_add(wake_q, owner);
352		}
353		return true;
354	}
355
356	return false;
357}
358
359/*
360 * We just acquired @lock under @ww_ctx, if there are more important contexts
361 * waiting behind us on the wait-list, check if they need to die, or wound us.
362 *
363 * See __ww_mutex_add_waiter() for the list-order construction; basically the
364 * list is ordered by stamp, smallest (oldest) first.
365 *
366 * This relies on never mixing wait-die/wound-wait on the same wait-list;
367 * which is currently ensured by that being a ww_class property.
368 *
369 * The current task must not be on the wait list.
370 */
371static void
372__ww_mutex_check_waiters(struct MUTEX *lock, struct ww_acquire_ctx *ww_ctx,
373			 struct wake_q_head *wake_q)
374{
375	struct MUTEX_WAITER *cur;
376
377	lockdep_assert_wait_lock_held(lock);
378
379	for (cur = __ww_waiter_first(lock); cur;
380	     cur = __ww_waiter_next(lock, cur)) {
381
382		if (!cur->ww_ctx)
383			continue;
384
385		if (__ww_mutex_die(lock, cur, ww_ctx, wake_q) ||
386		    __ww_mutex_wound(lock, cur->ww_ctx, ww_ctx, wake_q))
387			break;
388	}
389}
390
391/*
392 * After acquiring lock with fastpath, where we do not hold wait_lock, set ctx
393 * and wake up any waiters so they can recheck.
394 */
395static __always_inline void
396ww_mutex_set_context_fastpath(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
397{
398	DEFINE_WAKE_Q(wake_q);
399	unsigned long flags;
400
401	ww_mutex_lock_acquired(lock, ctx);
402
403	/*
404	 * The lock->ctx update should be visible on all cores before
405	 * the WAITERS check is done, otherwise contended waiters might be
406	 * missed. The contended waiters will either see ww_ctx == NULL
407	 * and keep spinning, or it will acquire wait_lock, add itself
408	 * to waiter list and sleep.
409	 */
410	smp_mb(); /* See comments above and below. */
411
412	/*
413	 * [W] ww->ctx = ctx	    [W] MUTEX_FLAG_WAITERS
414	 *     MB		        MB
415	 * [R] MUTEX_FLAG_WAITERS   [R] ww->ctx
416	 *
417	 * The memory barrier above pairs with the memory barrier in
418	 * __ww_mutex_add_waiter() and makes sure we either observe ww->ctx
419	 * and/or !empty list.
420	 */
421	if (likely(!__ww_mutex_has_waiters(&lock->base)))
422		return;
423
424	/*
425	 * Uh oh, we raced in fastpath, check if any of the waiters need to
426	 * die or wound us.
427	 */
428	lock_wait_lock(&lock->base, &flags);
429	__ww_mutex_check_waiters(&lock->base, ctx, &wake_q);
430	preempt_disable();
431	unlock_wait_lock(&lock->base, &flags);
432	wake_up_q(&wake_q);
433	preempt_enable();
434}
435
436static __always_inline int
437__ww_mutex_kill(struct MUTEX *lock, struct ww_acquire_ctx *ww_ctx)
438{
439	if (ww_ctx->acquired > 0) {
440#ifdef DEBUG_WW_MUTEXES
441		struct ww_mutex *ww;
442
443		ww = container_of(lock, struct ww_mutex, base);
444		DEBUG_LOCKS_WARN_ON(ww_ctx->contending_lock);
445		ww_ctx->contending_lock = ww;
446#endif
447		return -EDEADLK;
448	}
449
450	return 0;
451}
452
453/*
454 * Check the wound condition for the current lock acquire.
455 *
456 * Wound-Wait: If we're wounded, kill ourself.
457 *
458 * Wait-Die: If we're trying to acquire a lock already held by an older
459 *           context, kill ourselves.
460 *
461 * Since __ww_mutex_add_waiter() orders the wait-list on stamp, we only have to
462 * look at waiters before us in the wait-list.
463 */
464static inline int
465__ww_mutex_check_kill(struct MUTEX *lock, struct MUTEX_WAITER *waiter,
466		      struct ww_acquire_ctx *ctx)
467{
468	struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
469	struct ww_acquire_ctx *hold_ctx = READ_ONCE(ww->ctx);
470	struct MUTEX_WAITER *cur;
471
472	if (ctx->acquired == 0)
473		return 0;
474
475	if (!ctx->is_wait_die) {
476		if (ctx->wounded)
477			return __ww_mutex_kill(lock, ctx);
478
479		return 0;
480	}
481
482	if (hold_ctx && __ww_ctx_less(ctx, hold_ctx))
483		return __ww_mutex_kill(lock, ctx);
484
485	/*
486	 * If there is a waiter in front of us that has a context, then its
487	 * stamp is earlier than ours and we must kill ourself.
488	 */
489	for (cur = __ww_waiter_prev(lock, waiter); cur;
490	     cur = __ww_waiter_prev(lock, cur)) {
491
492		if (!cur->ww_ctx)
493			continue;
494
495		return __ww_mutex_kill(lock, ctx);
496	}
497
498	return 0;
499}
500
501/*
502 * Add @waiter to the wait-list, keep the wait-list ordered by stamp, smallest
503 * first. Such that older contexts are preferred to acquire the lock over
504 * younger contexts.
505 *
506 * Waiters without context are interspersed in FIFO order.
507 *
508 * Furthermore, for Wait-Die kill ourself immediately when possible (there are
509 * older contexts already waiting) to avoid unnecessary waiting and for
510 * Wound-Wait ensure we wound the owning context when it is younger.
511 */
512static inline int
513__ww_mutex_add_waiter(struct MUTEX_WAITER *waiter,
514		      struct MUTEX *lock,
515		      struct ww_acquire_ctx *ww_ctx,
516		      struct wake_q_head *wake_q)
517{
518	struct MUTEX_WAITER *cur, *pos = NULL;
519	bool is_wait_die;
520
521	if (!ww_ctx) {
522		__ww_waiter_add(lock, waiter, NULL);
523		return 0;
524	}
525
526	is_wait_die = ww_ctx->is_wait_die;
527
528	/*
529	 * Add the waiter before the first waiter with a higher stamp.
530	 * Waiters without a context are skipped to avoid starving
531	 * them. Wait-Die waiters may die here. Wound-Wait waiters
532	 * never die here, but they are sorted in stamp order and
533	 * may wound the lock holder.
534	 */
535	for (cur = __ww_waiter_last(lock); cur;
536	     cur = __ww_waiter_prev(lock, cur)) {
537
538		if (!cur->ww_ctx)
539			continue;
540
541		if (__ww_ctx_less(ww_ctx, cur->ww_ctx)) {
542			/*
543			 * Wait-Die: if we find an older context waiting, there
544			 * is no point in queueing behind it, as we'd have to
545			 * die the moment it would acquire the lock.
546			 */
547			if (is_wait_die) {
548				int ret = __ww_mutex_kill(lock, ww_ctx);
549
550				if (ret)
551					return ret;
552			}
553
554			break;
555		}
556
557		pos = cur;
558
559		/* Wait-Die: ensure younger waiters die. */
560		__ww_mutex_die(lock, cur, ww_ctx, wake_q);
561	}
562
563	__ww_waiter_add(lock, waiter, pos);
564
565	/*
566	 * Wound-Wait: if we're blocking on a mutex owned by a younger context,
567	 * wound that such that we might proceed.
568	 */
569	if (!is_wait_die) {
570		struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
571
572		/*
573		 * See ww_mutex_set_context_fastpath(). Orders setting
574		 * MUTEX_FLAG_WAITERS vs the ww->ctx load,
575		 * such that either we or the fastpath will wound @ww->ctx.
576		 */
577		smp_mb();
578		__ww_mutex_wound(lock, ww_ctx, ww->ctx, wake_q);
579	}
580
581	return 0;
582}
583
584static inline void __ww_mutex_unlock(struct ww_mutex *lock)
585{
586	if (lock->ctx) {
587#ifdef DEBUG_WW_MUTEXES
588		DEBUG_LOCKS_WARN_ON(!lock->ctx->acquired);
589#endif
590		if (lock->ctx->acquired > 0)
591			lock->ctx->acquired--;
592		lock->ctx = NULL;
593	}
594}