xref: /linux/kernel/locking/ww_mutex.h (revision f9aec1648df09d55436a0e3a94acff1df507751f)
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 
8 static 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 
20 static 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 
30 static 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 
40 static 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 
52 static 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 
61 static inline struct task_struct *
62 __ww_mutex_owner(struct mutex *lock)
63 {
64 	return __mutex_owner(lock);
65 }
66 
67 static inline bool
68 __ww_mutex_has_waiters(struct mutex *lock)
69 {
70 	return atomic_long_read(&lock->owner) & MUTEX_FLAG_WAITERS;
71 }
72 
73 static inline void lock_wait_lock(struct mutex *lock)
74 {
75 	raw_spin_lock(&lock->wait_lock);
76 }
77 
78 static inline void unlock_wait_lock(struct mutex *lock)
79 {
80 	raw_spin_unlock(&lock->wait_lock);
81 }
82 
83 static 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 
93 static 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 
102 static 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 
111 static 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 
120 static 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 
129 static 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 
135 static inline struct task_struct *
136 __ww_mutex_owner(struct rt_mutex *lock)
137 {
138 	return rt_mutex_owner(&lock->rtmutex);
139 }
140 
141 static inline bool
142 __ww_mutex_has_waiters(struct rt_mutex *lock)
143 {
144 	return rt_mutex_has_waiters(&lock->rtmutex);
145 }
146 
147 static inline void lock_wait_lock(struct rt_mutex *lock)
148 {
149 	raw_spin_lock(&lock->rtmutex.wait_lock);
150 }
151 
152 static inline void unlock_wait_lock(struct rt_mutex *lock)
153 {
154 	raw_spin_unlock(&lock->rtmutex.wait_lock);
155 }
156 
157 static 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  */
180 static __always_inline void
181 ww_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  */
227 static 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_prio(a_prio) || rt_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  */
276 static bool
277 __ww_mutex_die(struct MUTEX *lock, struct MUTEX_WAITER *waiter,
278 	       struct ww_acquire_ctx *ww_ctx)
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 		wake_up_process(waiter->task);
288 	}
289 
290 	return true;
291 }
292 
293 /*
294  * Wound-Wait; wound a lesser @hold_ctx if it holds the lock.
295  *
296  * Wound the lock holder if there are waiters with more important transactions
297  * than the lock holders. Even if multiple waiters may wound the lock holder,
298  * it's sufficient that only one does.
299  */
300 static bool __ww_mutex_wound(struct MUTEX *lock,
301 			     struct ww_acquire_ctx *ww_ctx,
302 			     struct ww_acquire_ctx *hold_ctx)
303 {
304 	struct task_struct *owner = __ww_mutex_owner(lock);
305 
306 	lockdep_assert_wait_lock_held(lock);
307 
308 	/*
309 	 * Possible through __ww_mutex_add_waiter() when we race with
310 	 * ww_mutex_set_context_fastpath(). In that case we'll get here again
311 	 * through __ww_mutex_check_waiters().
312 	 */
313 	if (!hold_ctx)
314 		return false;
315 
316 	/*
317 	 * Can have !owner because of __mutex_unlock_slowpath(), but if owner,
318 	 * it cannot go away because we'll have FLAG_WAITERS set and hold
319 	 * wait_lock.
320 	 */
321 	if (!owner)
322 		return false;
323 
324 	if (ww_ctx->acquired > 0 && __ww_ctx_less(hold_ctx, ww_ctx)) {
325 		hold_ctx->wounded = 1;
326 
327 		/*
328 		 * wake_up_process() paired with set_current_state()
329 		 * inserts sufficient barriers to make sure @owner either sees
330 		 * it's wounded in __ww_mutex_check_kill() or has a
331 		 * wakeup pending to re-read the wounded state.
332 		 */
333 		if (owner != current)
334 			wake_up_process(owner);
335 
336 		return true;
337 	}
338 
339 	return false;
340 }
341 
342 /*
343  * We just acquired @lock under @ww_ctx, if there are more important contexts
344  * waiting behind us on the wait-list, check if they need to die, or wound us.
345  *
346  * See __ww_mutex_add_waiter() for the list-order construction; basically the
347  * list is ordered by stamp, smallest (oldest) first.
348  *
349  * This relies on never mixing wait-die/wound-wait on the same wait-list;
350  * which is currently ensured by that being a ww_class property.
351  *
352  * The current task must not be on the wait list.
353  */
354 static void
355 __ww_mutex_check_waiters(struct MUTEX *lock, struct ww_acquire_ctx *ww_ctx)
356 {
357 	struct MUTEX_WAITER *cur;
358 
359 	lockdep_assert_wait_lock_held(lock);
360 
361 	for (cur = __ww_waiter_first(lock); cur;
362 	     cur = __ww_waiter_next(lock, cur)) {
363 
364 		if (!cur->ww_ctx)
365 			continue;
366 
367 		if (__ww_mutex_die(lock, cur, ww_ctx) ||
368 		    __ww_mutex_wound(lock, cur->ww_ctx, ww_ctx))
369 			break;
370 	}
371 }
372 
373 /*
374  * After acquiring lock with fastpath, where we do not hold wait_lock, set ctx
375  * and wake up any waiters so they can recheck.
376  */
377 static __always_inline void
378 ww_mutex_set_context_fastpath(struct ww_mutex *lock, struct ww_acquire_ctx *ctx)
379 {
380 	ww_mutex_lock_acquired(lock, ctx);
381 
382 	/*
383 	 * The lock->ctx update should be visible on all cores before
384 	 * the WAITERS check is done, otherwise contended waiters might be
385 	 * missed. The contended waiters will either see ww_ctx == NULL
386 	 * and keep spinning, or it will acquire wait_lock, add itself
387 	 * to waiter list and sleep.
388 	 */
389 	smp_mb(); /* See comments above and below. */
390 
391 	/*
392 	 * [W] ww->ctx = ctx	    [W] MUTEX_FLAG_WAITERS
393 	 *     MB		        MB
394 	 * [R] MUTEX_FLAG_WAITERS   [R] ww->ctx
395 	 *
396 	 * The memory barrier above pairs with the memory barrier in
397 	 * __ww_mutex_add_waiter() and makes sure we either observe ww->ctx
398 	 * and/or !empty list.
399 	 */
400 	if (likely(!__ww_mutex_has_waiters(&lock->base)))
401 		return;
402 
403 	/*
404 	 * Uh oh, we raced in fastpath, check if any of the waiters need to
405 	 * die or wound us.
406 	 */
407 	lock_wait_lock(&lock->base);
408 	__ww_mutex_check_waiters(&lock->base, ctx);
409 	unlock_wait_lock(&lock->base);
410 }
411 
412 static __always_inline int
413 __ww_mutex_kill(struct MUTEX *lock, struct ww_acquire_ctx *ww_ctx)
414 {
415 	if (ww_ctx->acquired > 0) {
416 #ifdef DEBUG_WW_MUTEXES
417 		struct ww_mutex *ww;
418 
419 		ww = container_of(lock, struct ww_mutex, base);
420 		DEBUG_LOCKS_WARN_ON(ww_ctx->contending_lock);
421 		ww_ctx->contending_lock = ww;
422 #endif
423 		return -EDEADLK;
424 	}
425 
426 	return 0;
427 }
428 
429 /*
430  * Check the wound condition for the current lock acquire.
431  *
432  * Wound-Wait: If we're wounded, kill ourself.
433  *
434  * Wait-Die: If we're trying to acquire a lock already held by an older
435  *           context, kill ourselves.
436  *
437  * Since __ww_mutex_add_waiter() orders the wait-list on stamp, we only have to
438  * look at waiters before us in the wait-list.
439  */
440 static inline int
441 __ww_mutex_check_kill(struct MUTEX *lock, struct MUTEX_WAITER *waiter,
442 		      struct ww_acquire_ctx *ctx)
443 {
444 	struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
445 	struct ww_acquire_ctx *hold_ctx = READ_ONCE(ww->ctx);
446 	struct MUTEX_WAITER *cur;
447 
448 	if (ctx->acquired == 0)
449 		return 0;
450 
451 	if (!ctx->is_wait_die) {
452 		if (ctx->wounded)
453 			return __ww_mutex_kill(lock, ctx);
454 
455 		return 0;
456 	}
457 
458 	if (hold_ctx && __ww_ctx_less(ctx, hold_ctx))
459 		return __ww_mutex_kill(lock, ctx);
460 
461 	/*
462 	 * If there is a waiter in front of us that has a context, then its
463 	 * stamp is earlier than ours and we must kill ourself.
464 	 */
465 	for (cur = __ww_waiter_prev(lock, waiter); cur;
466 	     cur = __ww_waiter_prev(lock, cur)) {
467 
468 		if (!cur->ww_ctx)
469 			continue;
470 
471 		return __ww_mutex_kill(lock, ctx);
472 	}
473 
474 	return 0;
475 }
476 
477 /*
478  * Add @waiter to the wait-list, keep the wait-list ordered by stamp, smallest
479  * first. Such that older contexts are preferred to acquire the lock over
480  * younger contexts.
481  *
482  * Waiters without context are interspersed in FIFO order.
483  *
484  * Furthermore, for Wait-Die kill ourself immediately when possible (there are
485  * older contexts already waiting) to avoid unnecessary waiting and for
486  * Wound-Wait ensure we wound the owning context when it is younger.
487  */
488 static inline int
489 __ww_mutex_add_waiter(struct MUTEX_WAITER *waiter,
490 		      struct MUTEX *lock,
491 		      struct ww_acquire_ctx *ww_ctx)
492 {
493 	struct MUTEX_WAITER *cur, *pos = NULL;
494 	bool is_wait_die;
495 
496 	if (!ww_ctx) {
497 		__ww_waiter_add(lock, waiter, NULL);
498 		return 0;
499 	}
500 
501 	is_wait_die = ww_ctx->is_wait_die;
502 
503 	/*
504 	 * Add the waiter before the first waiter with a higher stamp.
505 	 * Waiters without a context are skipped to avoid starving
506 	 * them. Wait-Die waiters may die here. Wound-Wait waiters
507 	 * never die here, but they are sorted in stamp order and
508 	 * may wound the lock holder.
509 	 */
510 	for (cur = __ww_waiter_last(lock); cur;
511 	     cur = __ww_waiter_prev(lock, cur)) {
512 
513 		if (!cur->ww_ctx)
514 			continue;
515 
516 		if (__ww_ctx_less(ww_ctx, cur->ww_ctx)) {
517 			/*
518 			 * Wait-Die: if we find an older context waiting, there
519 			 * is no point in queueing behind it, as we'd have to
520 			 * die the moment it would acquire the lock.
521 			 */
522 			if (is_wait_die) {
523 				int ret = __ww_mutex_kill(lock, ww_ctx);
524 
525 				if (ret)
526 					return ret;
527 			}
528 
529 			break;
530 		}
531 
532 		pos = cur;
533 
534 		/* Wait-Die: ensure younger waiters die. */
535 		__ww_mutex_die(lock, cur, ww_ctx);
536 	}
537 
538 	__ww_waiter_add(lock, waiter, pos);
539 
540 	/*
541 	 * Wound-Wait: if we're blocking on a mutex owned by a younger context,
542 	 * wound that such that we might proceed.
543 	 */
544 	if (!is_wait_die) {
545 		struct ww_mutex *ww = container_of(lock, struct ww_mutex, base);
546 
547 		/*
548 		 * See ww_mutex_set_context_fastpath(). Orders setting
549 		 * MUTEX_FLAG_WAITERS vs the ww->ctx load,
550 		 * such that either we or the fastpath will wound @ww->ctx.
551 		 */
552 		smp_mb();
553 		__ww_mutex_wound(lock, ww_ctx, ww->ctx);
554 	}
555 
556 	return 0;
557 }
558 
559 static inline void __ww_mutex_unlock(struct ww_mutex *lock)
560 {
561 	if (lock->ctx) {
562 #ifdef DEBUG_WW_MUTEXES
563 		DEBUG_LOCKS_WARN_ON(!lock->ctx->acquired);
564 #endif
565 		if (lock->ctx->acquired > 0)
566 			lock->ctx->acquired--;
567 		lock->ctx = NULL;
568 	}
569 }
570