xref: /qemu/util/qemu-coroutine-lock.c (revision 370ed600)
1 /*
2  * coroutine queues and locks
3  *
4  * Copyright (c) 2011 Kevin Wolf <kwolf@redhat.com>
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to deal
8  * in the Software without restriction, including without limitation the rights
9  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
10  * copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22  * THE SOFTWARE.
23  *
24  * The lock-free mutex implementation is based on OSv
25  * (core/lfmutex.cc, include/lockfree/mutex.hh).
26  * Copyright (C) 2013 Cloudius Systems, Ltd.
27  */
28 
29 #include "qemu/osdep.h"
30 #include "qemu/coroutine_int.h"
31 #include "qemu/processor.h"
32 #include "qemu/queue.h"
33 #include "block/aio.h"
34 #include "trace.h"
35 
36 void qemu_co_queue_init(CoQueue *queue)
37 {
38     QSIMPLEQ_INIT(&queue->entries);
39 }
40 
41 void coroutine_fn qemu_co_queue_wait_impl(CoQueue *queue, QemuLockable *lock,
42                                           CoQueueWaitFlags flags)
43 {
44     Coroutine *self = qemu_coroutine_self();
45     if (flags & CO_QUEUE_WAIT_FRONT) {
46         QSIMPLEQ_INSERT_HEAD(&queue->entries, self, co_queue_next);
47     } else {
48         QSIMPLEQ_INSERT_TAIL(&queue->entries, self, co_queue_next);
49     }
50 
51     if (lock) {
52         qemu_lockable_unlock(lock);
53     }
54 
55     /* There is no race condition here.  Other threads will call
56      * aio_co_schedule on our AioContext, which can reenter this
57      * coroutine but only after this yield and after the main loop
58      * has gone through the next iteration.
59      */
60     qemu_coroutine_yield();
61     assert(qemu_in_coroutine());
62 
63     /* TODO: OSv implements wait morphing here, where the wakeup
64      * primitive automatically places the woken coroutine on the
65      * mutex's queue.  This avoids the thundering herd effect.
66      * This could be implemented for CoMutexes, but not really for
67      * other cases of QemuLockable.
68      */
69     if (lock) {
70         qemu_lockable_lock(lock);
71     }
72 }
73 
74 bool qemu_co_enter_next_impl(CoQueue *queue, QemuLockable *lock)
75 {
76     Coroutine *next;
77 
78     next = QSIMPLEQ_FIRST(&queue->entries);
79     if (!next) {
80         return false;
81     }
82 
83     QSIMPLEQ_REMOVE_HEAD(&queue->entries, co_queue_next);
84     if (lock) {
85         qemu_lockable_unlock(lock);
86     }
87     aio_co_wake(next);
88     if (lock) {
89         qemu_lockable_lock(lock);
90     }
91     return true;
92 }
93 
94 bool coroutine_fn qemu_co_queue_next(CoQueue *queue)
95 {
96     /* No unlock/lock needed in coroutine context.  */
97     return qemu_co_enter_next_impl(queue, NULL);
98 }
99 
100 void qemu_co_enter_all_impl(CoQueue *queue, QemuLockable *lock)
101 {
102     while (qemu_co_enter_next_impl(queue, lock)) {
103         /* just loop */
104     }
105 }
106 
107 void coroutine_fn qemu_co_queue_restart_all(CoQueue *queue)
108 {
109     /* No unlock/lock needed in coroutine context.  */
110     qemu_co_enter_all_impl(queue, NULL);
111 }
112 
113 bool qemu_co_queue_empty(CoQueue *queue)
114 {
115     return QSIMPLEQ_FIRST(&queue->entries) == NULL;
116 }
117 
118 /* The wait records are handled with a multiple-producer, single-consumer
119  * lock-free queue.  There cannot be two concurrent pop_waiter() calls
120  * because pop_waiter() can only be called while mutex->handoff is zero.
121  * This can happen in three cases:
122  * - in qemu_co_mutex_unlock, before the hand-off protocol has started.
123  *   In this case, qemu_co_mutex_lock will see mutex->handoff == 0 and
124  *   not take part in the handoff.
125  * - in qemu_co_mutex_lock, if it steals the hand-off responsibility from
126  *   qemu_co_mutex_unlock.  In this case, qemu_co_mutex_unlock will fail
127  *   the cmpxchg (it will see either 0 or the next sequence value) and
128  *   exit.  The next hand-off cannot begin until qemu_co_mutex_lock has
129  *   woken up someone.
130  * - in qemu_co_mutex_unlock, if it takes the hand-off token itself.
131  *   In this case another iteration starts with mutex->handoff == 0;
132  *   a concurrent qemu_co_mutex_lock will fail the cmpxchg, and
133  *   qemu_co_mutex_unlock will go back to case (1).
134  *
135  * The following functions manage this queue.
136  */
137 typedef struct CoWaitRecord {
138     Coroutine *co;
139     QSLIST_ENTRY(CoWaitRecord) next;
140 } CoWaitRecord;
141 
142 static void coroutine_fn push_waiter(CoMutex *mutex, CoWaitRecord *w)
143 {
144     w->co = qemu_coroutine_self();
145     QSLIST_INSERT_HEAD_ATOMIC(&mutex->from_push, w, next);
146 }
147 
148 static void move_waiters(CoMutex *mutex)
149 {
150     QSLIST_HEAD(, CoWaitRecord) reversed;
151     QSLIST_MOVE_ATOMIC(&reversed, &mutex->from_push);
152     while (!QSLIST_EMPTY(&reversed)) {
153         CoWaitRecord *w = QSLIST_FIRST(&reversed);
154         QSLIST_REMOVE_HEAD(&reversed, next);
155         QSLIST_INSERT_HEAD(&mutex->to_pop, w, next);
156     }
157 }
158 
159 static CoWaitRecord *pop_waiter(CoMutex *mutex)
160 {
161     CoWaitRecord *w;
162 
163     if (QSLIST_EMPTY(&mutex->to_pop)) {
164         move_waiters(mutex);
165         if (QSLIST_EMPTY(&mutex->to_pop)) {
166             return NULL;
167         }
168     }
169     w = QSLIST_FIRST(&mutex->to_pop);
170     QSLIST_REMOVE_HEAD(&mutex->to_pop, next);
171     return w;
172 }
173 
174 static bool has_waiters(CoMutex *mutex)
175 {
176     return QSLIST_EMPTY(&mutex->to_pop) || QSLIST_EMPTY(&mutex->from_push);
177 }
178 
179 void qemu_co_mutex_init(CoMutex *mutex)
180 {
181     memset(mutex, 0, sizeof(*mutex));
182 }
183 
184 static void coroutine_fn qemu_co_mutex_wake(CoMutex *mutex, Coroutine *co)
185 {
186     /* Read co before co->ctx; pairs with smp_wmb() in
187      * qemu_coroutine_enter().
188      */
189     smp_read_barrier_depends();
190     mutex->ctx = co->ctx;
191     aio_co_wake(co);
192 }
193 
194 static void coroutine_fn qemu_co_mutex_lock_slowpath(AioContext *ctx,
195                                                      CoMutex *mutex)
196 {
197     Coroutine *self = qemu_coroutine_self();
198     CoWaitRecord w;
199     unsigned old_handoff;
200 
201     trace_qemu_co_mutex_lock_entry(mutex, self);
202     push_waiter(mutex, &w);
203 
204     /*
205      * Add waiter before reading mutex->handoff.  Pairs with qatomic_mb_set
206      * in qemu_co_mutex_unlock.
207      */
208     smp_mb__after_rmw();
209 
210     /* This is the "Responsibility Hand-Off" protocol; a lock() picks from
211      * a concurrent unlock() the responsibility of waking somebody up.
212      */
213     old_handoff = qatomic_read(&mutex->handoff);
214     if (old_handoff &&
215         has_waiters(mutex) &&
216         qatomic_cmpxchg(&mutex->handoff, old_handoff, 0) == old_handoff) {
217         /* There can be no concurrent pops, because there can be only
218          * one active handoff at a time.
219          */
220         CoWaitRecord *to_wake = pop_waiter(mutex);
221         Coroutine *co = to_wake->co;
222         if (co == self) {
223             /* We got the lock ourselves!  */
224             assert(to_wake == &w);
225             mutex->ctx = ctx;
226             return;
227         }
228 
229         qemu_co_mutex_wake(mutex, co);
230     }
231 
232     qemu_coroutine_yield();
233     trace_qemu_co_mutex_lock_return(mutex, self);
234 }
235 
236 void coroutine_fn qemu_co_mutex_lock(CoMutex *mutex)
237 {
238     AioContext *ctx = qemu_get_current_aio_context();
239     Coroutine *self = qemu_coroutine_self();
240     int waiters, i;
241 
242     /* Running a very small critical section on pthread_mutex_t and CoMutex
243      * shows that pthread_mutex_t is much faster because it doesn't actually
244      * go to sleep.  What happens is that the critical section is shorter
245      * than the latency of entering the kernel and thus FUTEX_WAIT always
246      * fails.  With CoMutex there is no such latency but you still want to
247      * avoid wait and wakeup.  So introduce it artificially.
248      */
249     i = 0;
250 retry_fast_path:
251     waiters = qatomic_cmpxchg(&mutex->locked, 0, 1);
252     if (waiters != 0) {
253         while (waiters == 1 && ++i < 1000) {
254             if (qatomic_read(&mutex->ctx) == ctx) {
255                 break;
256             }
257             if (qatomic_read(&mutex->locked) == 0) {
258                 goto retry_fast_path;
259             }
260             cpu_relax();
261         }
262         waiters = qatomic_fetch_inc(&mutex->locked);
263     }
264 
265     if (waiters == 0) {
266         /* Uncontended.  */
267         trace_qemu_co_mutex_lock_uncontended(mutex, self);
268         mutex->ctx = ctx;
269     } else {
270         qemu_co_mutex_lock_slowpath(ctx, mutex);
271     }
272     mutex->holder = self;
273     self->locks_held++;
274 }
275 
276 void coroutine_fn qemu_co_mutex_unlock(CoMutex *mutex)
277 {
278     Coroutine *self = qemu_coroutine_self();
279 
280     trace_qemu_co_mutex_unlock_entry(mutex, self);
281 
282     assert(mutex->locked);
283     assert(mutex->holder == self);
284     assert(qemu_in_coroutine());
285 
286     mutex->ctx = NULL;
287     mutex->holder = NULL;
288     self->locks_held--;
289     if (qatomic_fetch_dec(&mutex->locked) == 1) {
290         /* No waiting qemu_co_mutex_lock().  Pfew, that was easy!  */
291         return;
292     }
293 
294     for (;;) {
295         CoWaitRecord *to_wake = pop_waiter(mutex);
296         unsigned our_handoff;
297 
298         if (to_wake) {
299             qemu_co_mutex_wake(mutex, to_wake->co);
300             break;
301         }
302 
303         /* Some concurrent lock() is in progress (we know this because
304          * mutex->locked was >1) but it hasn't yet put itself on the wait
305          * queue.  Pick a sequence number for the handoff protocol (not 0).
306          */
307         if (++mutex->sequence == 0) {
308             mutex->sequence = 1;
309         }
310 
311         our_handoff = mutex->sequence;
312         /* Set handoff before checking for waiters.  */
313         qatomic_mb_set(&mutex->handoff, our_handoff);
314         if (!has_waiters(mutex)) {
315             /* The concurrent lock has not added itself yet, so it
316              * will be able to pick our handoff.
317              */
318             break;
319         }
320 
321         /* Try to do the handoff protocol ourselves; if somebody else has
322          * already taken it, however, we're done and they're responsible.
323          */
324         if (qatomic_cmpxchg(&mutex->handoff, our_handoff, 0) != our_handoff) {
325             break;
326         }
327     }
328 
329     trace_qemu_co_mutex_unlock_return(mutex, self);
330 }
331 
332 struct CoRwTicket {
333     bool read;
334     Coroutine *co;
335     QSIMPLEQ_ENTRY(CoRwTicket) next;
336 };
337 
338 void qemu_co_rwlock_init(CoRwlock *lock)
339 {
340     qemu_co_mutex_init(&lock->mutex);
341     lock->owners = 0;
342     QSIMPLEQ_INIT(&lock->tickets);
343 }
344 
345 /* Releases the internal CoMutex.  */
346 static void coroutine_fn qemu_co_rwlock_maybe_wake_one(CoRwlock *lock)
347 {
348     CoRwTicket *tkt = QSIMPLEQ_FIRST(&lock->tickets);
349     Coroutine *co = NULL;
350 
351     /*
352      * Setting lock->owners here prevents rdlock and wrlock from
353      * sneaking in between unlock and wake.
354      */
355 
356     if (tkt) {
357         if (tkt->read) {
358             if (lock->owners >= 0) {
359                 lock->owners++;
360                 co = tkt->co;
361             }
362         } else {
363             if (lock->owners == 0) {
364                 lock->owners = -1;
365                 co = tkt->co;
366             }
367         }
368     }
369 
370     if (co) {
371         QSIMPLEQ_REMOVE_HEAD(&lock->tickets, next);
372         qemu_co_mutex_unlock(&lock->mutex);
373         aio_co_wake(co);
374     } else {
375         qemu_co_mutex_unlock(&lock->mutex);
376     }
377 }
378 
379 void coroutine_fn qemu_co_rwlock_rdlock(CoRwlock *lock)
380 {
381     Coroutine *self = qemu_coroutine_self();
382 
383     qemu_co_mutex_lock(&lock->mutex);
384     /* For fairness, wait if a writer is in line.  */
385     if (lock->owners == 0 || (lock->owners > 0 && QSIMPLEQ_EMPTY(&lock->tickets))) {
386         lock->owners++;
387         qemu_co_mutex_unlock(&lock->mutex);
388     } else {
389         CoRwTicket my_ticket = { true, self };
390 
391         QSIMPLEQ_INSERT_TAIL(&lock->tickets, &my_ticket, next);
392         qemu_co_mutex_unlock(&lock->mutex);
393         qemu_coroutine_yield();
394         assert(lock->owners >= 1);
395 
396         /* Possibly wake another reader, which will wake the next in line.  */
397         qemu_co_mutex_lock(&lock->mutex);
398         qemu_co_rwlock_maybe_wake_one(lock);
399     }
400 
401     self->locks_held++;
402 }
403 
404 void coroutine_fn qemu_co_rwlock_unlock(CoRwlock *lock)
405 {
406     Coroutine *self = qemu_coroutine_self();
407 
408     assert(qemu_in_coroutine());
409     self->locks_held--;
410 
411     qemu_co_mutex_lock(&lock->mutex);
412     if (lock->owners > 0) {
413         lock->owners--;
414     } else {
415         assert(lock->owners == -1);
416         lock->owners = 0;
417     }
418 
419     qemu_co_rwlock_maybe_wake_one(lock);
420 }
421 
422 void coroutine_fn qemu_co_rwlock_downgrade(CoRwlock *lock)
423 {
424     qemu_co_mutex_lock(&lock->mutex);
425     assert(lock->owners == -1);
426     lock->owners = 1;
427 
428     /* Possibly wake another reader, which will wake the next in line.  */
429     qemu_co_rwlock_maybe_wake_one(lock);
430 }
431 
432 void coroutine_fn qemu_co_rwlock_wrlock(CoRwlock *lock)
433 {
434     Coroutine *self = qemu_coroutine_self();
435 
436     qemu_co_mutex_lock(&lock->mutex);
437     if (lock->owners == 0) {
438         lock->owners = -1;
439         qemu_co_mutex_unlock(&lock->mutex);
440     } else {
441         CoRwTicket my_ticket = { false, qemu_coroutine_self() };
442 
443         QSIMPLEQ_INSERT_TAIL(&lock->tickets, &my_ticket, next);
444         qemu_co_mutex_unlock(&lock->mutex);
445         qemu_coroutine_yield();
446         assert(lock->owners == -1);
447     }
448 
449     self->locks_held++;
450 }
451 
452 void coroutine_fn qemu_co_rwlock_upgrade(CoRwlock *lock)
453 {
454     qemu_co_mutex_lock(&lock->mutex);
455     assert(lock->owners > 0);
456     /* For fairness, wait if a writer is in line.  */
457     if (lock->owners == 1 && QSIMPLEQ_EMPTY(&lock->tickets)) {
458         lock->owners = -1;
459         qemu_co_mutex_unlock(&lock->mutex);
460     } else {
461         CoRwTicket my_ticket = { false, qemu_coroutine_self() };
462 
463         lock->owners--;
464         QSIMPLEQ_INSERT_TAIL(&lock->tickets, &my_ticket, next);
465         qemu_co_rwlock_maybe_wake_one(lock);
466         qemu_coroutine_yield();
467         assert(lock->owners == -1);
468     }
469 }
470