xref: /qemu/util/qemu-coroutine-lock.c (revision b25f23e7)
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-common.h"
31 #include "qemu/coroutine.h"
32 #include "qemu/coroutine_int.h"
33 #include "qemu/processor.h"
34 #include "qemu/queue.h"
35 #include "block/aio.h"
36 #include "trace.h"
37 
38 void qemu_co_queue_init(CoQueue *queue)
39 {
40     QSIMPLEQ_INIT(&queue->entries);
41 }
42 
43 void coroutine_fn qemu_co_queue_wait(CoQueue *queue, CoMutex *mutex)
44 {
45     Coroutine *self = qemu_coroutine_self();
46     QSIMPLEQ_INSERT_TAIL(&queue->entries, self, co_queue_next);
47 
48     if (mutex) {
49         qemu_co_mutex_unlock(mutex);
50     }
51 
52     /* There is no race condition here.  Other threads will call
53      * aio_co_schedule on our AioContext, which can reenter this
54      * coroutine but only after this yield and after the main loop
55      * has gone through the next iteration.
56      */
57     qemu_coroutine_yield();
58     assert(qemu_in_coroutine());
59 
60     /* TODO: OSv implements wait morphing here, where the wakeup
61      * primitive automatically places the woken coroutine on the
62      * mutex's queue.  This avoids the thundering herd effect.
63      */
64     if (mutex) {
65         qemu_co_mutex_lock(mutex);
66     }
67 }
68 
69 /**
70  * qemu_co_queue_run_restart:
71  *
72  * Enter each coroutine that was previously marked for restart by
73  * qemu_co_queue_next() or qemu_co_queue_restart_all().  This function is
74  * invoked by the core coroutine code when the current coroutine yields or
75  * terminates.
76  */
77 void qemu_co_queue_run_restart(Coroutine *co)
78 {
79     Coroutine *next;
80 
81     trace_qemu_co_queue_run_restart(co);
82     while ((next = QSIMPLEQ_FIRST(&co->co_queue_wakeup))) {
83         QSIMPLEQ_REMOVE_HEAD(&co->co_queue_wakeup, co_queue_next);
84         qemu_coroutine_enter(next);
85     }
86 }
87 
88 static bool qemu_co_queue_do_restart(CoQueue *queue, bool single)
89 {
90     Coroutine *next;
91 
92     if (QSIMPLEQ_EMPTY(&queue->entries)) {
93         return false;
94     }
95 
96     while ((next = QSIMPLEQ_FIRST(&queue->entries)) != NULL) {
97         QSIMPLEQ_REMOVE_HEAD(&queue->entries, co_queue_next);
98         aio_co_wake(next);
99         if (single) {
100             break;
101         }
102     }
103     return true;
104 }
105 
106 bool coroutine_fn qemu_co_queue_next(CoQueue *queue)
107 {
108     assert(qemu_in_coroutine());
109     return qemu_co_queue_do_restart(queue, true);
110 }
111 
112 void coroutine_fn qemu_co_queue_restart_all(CoQueue *queue)
113 {
114     assert(qemu_in_coroutine());
115     qemu_co_queue_do_restart(queue, false);
116 }
117 
118 bool qemu_co_enter_next(CoQueue *queue)
119 {
120     Coroutine *next;
121 
122     next = QSIMPLEQ_FIRST(&queue->entries);
123     if (!next) {
124         return false;
125     }
126 
127     QSIMPLEQ_REMOVE_HEAD(&queue->entries, co_queue_next);
128     qemu_coroutine_enter(next);
129     return true;
130 }
131 
132 bool qemu_co_queue_empty(CoQueue *queue)
133 {
134     return QSIMPLEQ_FIRST(&queue->entries) == NULL;
135 }
136 
137 /* The wait records are handled with a multiple-producer, single-consumer
138  * lock-free queue.  There cannot be two concurrent pop_waiter() calls
139  * because pop_waiter() can only be called while mutex->handoff is zero.
140  * This can happen in three cases:
141  * - in qemu_co_mutex_unlock, before the hand-off protocol has started.
142  *   In this case, qemu_co_mutex_lock will see mutex->handoff == 0 and
143  *   not take part in the handoff.
144  * - in qemu_co_mutex_lock, if it steals the hand-off responsibility from
145  *   qemu_co_mutex_unlock.  In this case, qemu_co_mutex_unlock will fail
146  *   the cmpxchg (it will see either 0 or the next sequence value) and
147  *   exit.  The next hand-off cannot begin until qemu_co_mutex_lock has
148  *   woken up someone.
149  * - in qemu_co_mutex_unlock, if it takes the hand-off token itself.
150  *   In this case another iteration starts with mutex->handoff == 0;
151  *   a concurrent qemu_co_mutex_lock will fail the cmpxchg, and
152  *   qemu_co_mutex_unlock will go back to case (1).
153  *
154  * The following functions manage this queue.
155  */
156 typedef struct CoWaitRecord {
157     Coroutine *co;
158     QSLIST_ENTRY(CoWaitRecord) next;
159 } CoWaitRecord;
160 
161 static void push_waiter(CoMutex *mutex, CoWaitRecord *w)
162 {
163     w->co = qemu_coroutine_self();
164     QSLIST_INSERT_HEAD_ATOMIC(&mutex->from_push, w, next);
165 }
166 
167 static void move_waiters(CoMutex *mutex)
168 {
169     QSLIST_HEAD(, CoWaitRecord) reversed;
170     QSLIST_MOVE_ATOMIC(&reversed, &mutex->from_push);
171     while (!QSLIST_EMPTY(&reversed)) {
172         CoWaitRecord *w = QSLIST_FIRST(&reversed);
173         QSLIST_REMOVE_HEAD(&reversed, next);
174         QSLIST_INSERT_HEAD(&mutex->to_pop, w, next);
175     }
176 }
177 
178 static CoWaitRecord *pop_waiter(CoMutex *mutex)
179 {
180     CoWaitRecord *w;
181 
182     if (QSLIST_EMPTY(&mutex->to_pop)) {
183         move_waiters(mutex);
184         if (QSLIST_EMPTY(&mutex->to_pop)) {
185             return NULL;
186         }
187     }
188     w = QSLIST_FIRST(&mutex->to_pop);
189     QSLIST_REMOVE_HEAD(&mutex->to_pop, next);
190     return w;
191 }
192 
193 static bool has_waiters(CoMutex *mutex)
194 {
195     return QSLIST_EMPTY(&mutex->to_pop) || QSLIST_EMPTY(&mutex->from_push);
196 }
197 
198 void qemu_co_mutex_init(CoMutex *mutex)
199 {
200     memset(mutex, 0, sizeof(*mutex));
201 }
202 
203 static void coroutine_fn qemu_co_mutex_wake(CoMutex *mutex, Coroutine *co)
204 {
205     /* Read co before co->ctx; pairs with smp_wmb() in
206      * qemu_coroutine_enter().
207      */
208     smp_read_barrier_depends();
209     mutex->ctx = co->ctx;
210     aio_co_wake(co);
211 }
212 
213 static void coroutine_fn qemu_co_mutex_lock_slowpath(AioContext *ctx,
214                                                      CoMutex *mutex)
215 {
216     Coroutine *self = qemu_coroutine_self();
217     CoWaitRecord w;
218     unsigned old_handoff;
219 
220     trace_qemu_co_mutex_lock_entry(mutex, self);
221     w.co = self;
222     push_waiter(mutex, &w);
223 
224     /* This is the "Responsibility Hand-Off" protocol; a lock() picks from
225      * a concurrent unlock() the responsibility of waking somebody up.
226      */
227     old_handoff = atomic_mb_read(&mutex->handoff);
228     if (old_handoff &&
229         has_waiters(mutex) &&
230         atomic_cmpxchg(&mutex->handoff, old_handoff, 0) == old_handoff) {
231         /* There can be no concurrent pops, because there can be only
232          * one active handoff at a time.
233          */
234         CoWaitRecord *to_wake = pop_waiter(mutex);
235         Coroutine *co = to_wake->co;
236         if (co == self) {
237             /* We got the lock ourselves!  */
238             assert(to_wake == &w);
239             mutex->ctx = ctx;
240             return;
241         }
242 
243         qemu_co_mutex_wake(mutex, co);
244     }
245 
246     qemu_coroutine_yield();
247     trace_qemu_co_mutex_lock_return(mutex, self);
248 }
249 
250 void coroutine_fn qemu_co_mutex_lock(CoMutex *mutex)
251 {
252     AioContext *ctx = qemu_get_current_aio_context();
253     Coroutine *self = qemu_coroutine_self();
254     int waiters, i;
255 
256     /* Running a very small critical section on pthread_mutex_t and CoMutex
257      * shows that pthread_mutex_t is much faster because it doesn't actually
258      * go to sleep.  What happens is that the critical section is shorter
259      * than the latency of entering the kernel and thus FUTEX_WAIT always
260      * fails.  With CoMutex there is no such latency but you still want to
261      * avoid wait and wakeup.  So introduce it artificially.
262      */
263     i = 0;
264 retry_fast_path:
265     waiters = atomic_cmpxchg(&mutex->locked, 0, 1);
266     if (waiters != 0) {
267         while (waiters == 1 && ++i < 1000) {
268             if (atomic_read(&mutex->ctx) == ctx) {
269                 break;
270             }
271             if (atomic_read(&mutex->locked) == 0) {
272                 goto retry_fast_path;
273             }
274             cpu_relax();
275         }
276         waiters = atomic_fetch_inc(&mutex->locked);
277     }
278 
279     if (waiters == 0) {
280         /* Uncontended.  */
281         trace_qemu_co_mutex_lock_uncontended(mutex, self);
282         mutex->ctx = ctx;
283     } else {
284         qemu_co_mutex_lock_slowpath(ctx, mutex);
285     }
286     mutex->holder = self;
287     self->locks_held++;
288 }
289 
290 void coroutine_fn qemu_co_mutex_unlock(CoMutex *mutex)
291 {
292     Coroutine *self = qemu_coroutine_self();
293 
294     trace_qemu_co_mutex_unlock_entry(mutex, self);
295 
296     assert(mutex->locked);
297     assert(mutex->holder == self);
298     assert(qemu_in_coroutine());
299 
300     mutex->ctx = NULL;
301     mutex->holder = NULL;
302     self->locks_held--;
303     if (atomic_fetch_dec(&mutex->locked) == 1) {
304         /* No waiting qemu_co_mutex_lock().  Pfew, that was easy!  */
305         return;
306     }
307 
308     for (;;) {
309         CoWaitRecord *to_wake = pop_waiter(mutex);
310         unsigned our_handoff;
311 
312         if (to_wake) {
313             qemu_co_mutex_wake(mutex, to_wake->co);
314             break;
315         }
316 
317         /* Some concurrent lock() is in progress (we know this because
318          * mutex->locked was >1) but it hasn't yet put itself on the wait
319          * queue.  Pick a sequence number for the handoff protocol (not 0).
320          */
321         if (++mutex->sequence == 0) {
322             mutex->sequence = 1;
323         }
324 
325         our_handoff = mutex->sequence;
326         atomic_mb_set(&mutex->handoff, our_handoff);
327         if (!has_waiters(mutex)) {
328             /* The concurrent lock has not added itself yet, so it
329              * will be able to pick our handoff.
330              */
331             break;
332         }
333 
334         /* Try to do the handoff protocol ourselves; if somebody else has
335          * already taken it, however, we're done and they're responsible.
336          */
337         if (atomic_cmpxchg(&mutex->handoff, our_handoff, 0) != our_handoff) {
338             break;
339         }
340     }
341 
342     trace_qemu_co_mutex_unlock_return(mutex, self);
343 }
344 
345 void qemu_co_rwlock_init(CoRwlock *lock)
346 {
347     memset(lock, 0, sizeof(*lock));
348     qemu_co_queue_init(&lock->queue);
349     qemu_co_mutex_init(&lock->mutex);
350 }
351 
352 void qemu_co_rwlock_rdlock(CoRwlock *lock)
353 {
354     Coroutine *self = qemu_coroutine_self();
355 
356     qemu_co_mutex_lock(&lock->mutex);
357     /* For fairness, wait if a writer is in line.  */
358     while (lock->pending_writer) {
359         qemu_co_queue_wait(&lock->queue, &lock->mutex);
360     }
361     lock->reader++;
362     qemu_co_mutex_unlock(&lock->mutex);
363 
364     /* The rest of the read-side critical section is run without the mutex.  */
365     self->locks_held++;
366 }
367 
368 void qemu_co_rwlock_unlock(CoRwlock *lock)
369 {
370     Coroutine *self = qemu_coroutine_self();
371 
372     assert(qemu_in_coroutine());
373     if (!lock->reader) {
374         /* The critical section started in qemu_co_rwlock_wrlock.  */
375         qemu_co_queue_restart_all(&lock->queue);
376     } else {
377         self->locks_held--;
378 
379         qemu_co_mutex_lock(&lock->mutex);
380         lock->reader--;
381         assert(lock->reader >= 0);
382         /* Wakeup only one waiting writer */
383         if (!lock->reader) {
384             qemu_co_queue_next(&lock->queue);
385         }
386     }
387     qemu_co_mutex_unlock(&lock->mutex);
388 }
389 
390 void qemu_co_rwlock_wrlock(CoRwlock *lock)
391 {
392     qemu_co_mutex_lock(&lock->mutex);
393     lock->pending_writer++;
394     while (lock->reader) {
395         qemu_co_queue_wait(&lock->queue, &lock->mutex);
396     }
397     lock->pending_writer--;
398 
399     /* The rest of the write-side critical section is run with
400      * the mutex taken, so that lock->reader remains zero.
401      * There is no need to update self->locks_held.
402      */
403 }
404