xref: /openbmc/qemu/util/qemu-coroutine-lock.c (revision 64552b6b)
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.h"
31 #include "qemu/coroutine_int.h"
32 #include "qemu/processor.h"
33 #include "qemu/queue.h"
34 #include "block/aio.h"
35 #include "trace.h"
36 
37 void qemu_co_queue_init(CoQueue *queue)
38 {
39     QSIMPLEQ_INIT(&queue->entries);
40 }
41 
42 void coroutine_fn qemu_co_queue_wait_impl(CoQueue *queue, QemuLockable *lock)
43 {
44     Coroutine *self = qemu_coroutine_self();
45     QSIMPLEQ_INSERT_TAIL(&queue->entries, self, co_queue_next);
46 
47     if (lock) {
48         qemu_lockable_unlock(lock);
49     }
50 
51     /* There is no race condition here.  Other threads will call
52      * aio_co_schedule on our AioContext, which can reenter this
53      * coroutine but only after this yield and after the main loop
54      * has gone through the next iteration.
55      */
56     qemu_coroutine_yield();
57     assert(qemu_in_coroutine());
58 
59     /* TODO: OSv implements wait morphing here, where the wakeup
60      * primitive automatically places the woken coroutine on the
61      * mutex's queue.  This avoids the thundering herd effect.
62      * This could be implemented for CoMutexes, but not really for
63      * other cases of QemuLockable.
64      */
65     if (lock) {
66         qemu_lockable_lock(lock);
67     }
68 }
69 
70 static bool qemu_co_queue_do_restart(CoQueue *queue, bool single)
71 {
72     Coroutine *next;
73 
74     if (QSIMPLEQ_EMPTY(&queue->entries)) {
75         return false;
76     }
77 
78     while ((next = QSIMPLEQ_FIRST(&queue->entries)) != NULL) {
79         QSIMPLEQ_REMOVE_HEAD(&queue->entries, co_queue_next);
80         aio_co_wake(next);
81         if (single) {
82             break;
83         }
84     }
85     return true;
86 }
87 
88 bool coroutine_fn qemu_co_queue_next(CoQueue *queue)
89 {
90     assert(qemu_in_coroutine());
91     return qemu_co_queue_do_restart(queue, true);
92 }
93 
94 void coroutine_fn qemu_co_queue_restart_all(CoQueue *queue)
95 {
96     assert(qemu_in_coroutine());
97     qemu_co_queue_do_restart(queue, false);
98 }
99 
100 bool qemu_co_enter_next_impl(CoQueue *queue, QemuLockable *lock)
101 {
102     Coroutine *next;
103 
104     next = QSIMPLEQ_FIRST(&queue->entries);
105     if (!next) {
106         return false;
107     }
108 
109     QSIMPLEQ_REMOVE_HEAD(&queue->entries, co_queue_next);
110     if (lock) {
111         qemu_lockable_unlock(lock);
112     }
113     aio_co_wake(next);
114     if (lock) {
115         qemu_lockable_lock(lock);
116     }
117     return true;
118 }
119 
120 bool qemu_co_queue_empty(CoQueue *queue)
121 {
122     return QSIMPLEQ_FIRST(&queue->entries) == NULL;
123 }
124 
125 /* The wait records are handled with a multiple-producer, single-consumer
126  * lock-free queue.  There cannot be two concurrent pop_waiter() calls
127  * because pop_waiter() can only be called while mutex->handoff is zero.
128  * This can happen in three cases:
129  * - in qemu_co_mutex_unlock, before the hand-off protocol has started.
130  *   In this case, qemu_co_mutex_lock will see mutex->handoff == 0 and
131  *   not take part in the handoff.
132  * - in qemu_co_mutex_lock, if it steals the hand-off responsibility from
133  *   qemu_co_mutex_unlock.  In this case, qemu_co_mutex_unlock will fail
134  *   the cmpxchg (it will see either 0 or the next sequence value) and
135  *   exit.  The next hand-off cannot begin until qemu_co_mutex_lock has
136  *   woken up someone.
137  * - in qemu_co_mutex_unlock, if it takes the hand-off token itself.
138  *   In this case another iteration starts with mutex->handoff == 0;
139  *   a concurrent qemu_co_mutex_lock will fail the cmpxchg, and
140  *   qemu_co_mutex_unlock will go back to case (1).
141  *
142  * The following functions manage this queue.
143  */
144 typedef struct CoWaitRecord {
145     Coroutine *co;
146     QSLIST_ENTRY(CoWaitRecord) next;
147 } CoWaitRecord;
148 
149 static void push_waiter(CoMutex *mutex, CoWaitRecord *w)
150 {
151     w->co = qemu_coroutine_self();
152     QSLIST_INSERT_HEAD_ATOMIC(&mutex->from_push, w, next);
153 }
154 
155 static void move_waiters(CoMutex *mutex)
156 {
157     QSLIST_HEAD(, CoWaitRecord) reversed;
158     QSLIST_MOVE_ATOMIC(&reversed, &mutex->from_push);
159     while (!QSLIST_EMPTY(&reversed)) {
160         CoWaitRecord *w = QSLIST_FIRST(&reversed);
161         QSLIST_REMOVE_HEAD(&reversed, next);
162         QSLIST_INSERT_HEAD(&mutex->to_pop, w, next);
163     }
164 }
165 
166 static CoWaitRecord *pop_waiter(CoMutex *mutex)
167 {
168     CoWaitRecord *w;
169 
170     if (QSLIST_EMPTY(&mutex->to_pop)) {
171         move_waiters(mutex);
172         if (QSLIST_EMPTY(&mutex->to_pop)) {
173             return NULL;
174         }
175     }
176     w = QSLIST_FIRST(&mutex->to_pop);
177     QSLIST_REMOVE_HEAD(&mutex->to_pop, next);
178     return w;
179 }
180 
181 static bool has_waiters(CoMutex *mutex)
182 {
183     return QSLIST_EMPTY(&mutex->to_pop) || QSLIST_EMPTY(&mutex->from_push);
184 }
185 
186 void qemu_co_mutex_init(CoMutex *mutex)
187 {
188     memset(mutex, 0, sizeof(*mutex));
189 }
190 
191 static void coroutine_fn qemu_co_mutex_wake(CoMutex *mutex, Coroutine *co)
192 {
193     /* Read co before co->ctx; pairs with smp_wmb() in
194      * qemu_coroutine_enter().
195      */
196     smp_read_barrier_depends();
197     mutex->ctx = co->ctx;
198     aio_co_wake(co);
199 }
200 
201 static void coroutine_fn qemu_co_mutex_lock_slowpath(AioContext *ctx,
202                                                      CoMutex *mutex)
203 {
204     Coroutine *self = qemu_coroutine_self();
205     CoWaitRecord w;
206     unsigned old_handoff;
207 
208     trace_qemu_co_mutex_lock_entry(mutex, self);
209     w.co = self;
210     push_waiter(mutex, &w);
211 
212     /* This is the "Responsibility Hand-Off" protocol; a lock() picks from
213      * a concurrent unlock() the responsibility of waking somebody up.
214      */
215     old_handoff = atomic_mb_read(&mutex->handoff);
216     if (old_handoff &&
217         has_waiters(mutex) &&
218         atomic_cmpxchg(&mutex->handoff, old_handoff, 0) == old_handoff) {
219         /* There can be no concurrent pops, because there can be only
220          * one active handoff at a time.
221          */
222         CoWaitRecord *to_wake = pop_waiter(mutex);
223         Coroutine *co = to_wake->co;
224         if (co == self) {
225             /* We got the lock ourselves!  */
226             assert(to_wake == &w);
227             mutex->ctx = ctx;
228             return;
229         }
230 
231         qemu_co_mutex_wake(mutex, co);
232     }
233 
234     qemu_coroutine_yield();
235     trace_qemu_co_mutex_lock_return(mutex, self);
236 }
237 
238 void coroutine_fn qemu_co_mutex_lock(CoMutex *mutex)
239 {
240     AioContext *ctx = qemu_get_current_aio_context();
241     Coroutine *self = qemu_coroutine_self();
242     int waiters, i;
243 
244     /* Running a very small critical section on pthread_mutex_t and CoMutex
245      * shows that pthread_mutex_t is much faster because it doesn't actually
246      * go to sleep.  What happens is that the critical section is shorter
247      * than the latency of entering the kernel and thus FUTEX_WAIT always
248      * fails.  With CoMutex there is no such latency but you still want to
249      * avoid wait and wakeup.  So introduce it artificially.
250      */
251     i = 0;
252 retry_fast_path:
253     waiters = atomic_cmpxchg(&mutex->locked, 0, 1);
254     if (waiters != 0) {
255         while (waiters == 1 && ++i < 1000) {
256             if (atomic_read(&mutex->ctx) == ctx) {
257                 break;
258             }
259             if (atomic_read(&mutex->locked) == 0) {
260                 goto retry_fast_path;
261             }
262             cpu_relax();
263         }
264         waiters = atomic_fetch_inc(&mutex->locked);
265     }
266 
267     if (waiters == 0) {
268         /* Uncontended.  */
269         trace_qemu_co_mutex_lock_uncontended(mutex, self);
270         mutex->ctx = ctx;
271     } else {
272         qemu_co_mutex_lock_slowpath(ctx, mutex);
273     }
274     mutex->holder = self;
275     self->locks_held++;
276 }
277 
278 void coroutine_fn qemu_co_mutex_unlock(CoMutex *mutex)
279 {
280     Coroutine *self = qemu_coroutine_self();
281 
282     trace_qemu_co_mutex_unlock_entry(mutex, self);
283 
284     assert(mutex->locked);
285     assert(mutex->holder == self);
286     assert(qemu_in_coroutine());
287 
288     mutex->ctx = NULL;
289     mutex->holder = NULL;
290     self->locks_held--;
291     if (atomic_fetch_dec(&mutex->locked) == 1) {
292         /* No waiting qemu_co_mutex_lock().  Pfew, that was easy!  */
293         return;
294     }
295 
296     for (;;) {
297         CoWaitRecord *to_wake = pop_waiter(mutex);
298         unsigned our_handoff;
299 
300         if (to_wake) {
301             qemu_co_mutex_wake(mutex, to_wake->co);
302             break;
303         }
304 
305         /* Some concurrent lock() is in progress (we know this because
306          * mutex->locked was >1) but it hasn't yet put itself on the wait
307          * queue.  Pick a sequence number for the handoff protocol (not 0).
308          */
309         if (++mutex->sequence == 0) {
310             mutex->sequence = 1;
311         }
312 
313         our_handoff = mutex->sequence;
314         atomic_mb_set(&mutex->handoff, our_handoff);
315         if (!has_waiters(mutex)) {
316             /* The concurrent lock has not added itself yet, so it
317              * will be able to pick our handoff.
318              */
319             break;
320         }
321 
322         /* Try to do the handoff protocol ourselves; if somebody else has
323          * already taken it, however, we're done and they're responsible.
324          */
325         if (atomic_cmpxchg(&mutex->handoff, our_handoff, 0) != our_handoff) {
326             break;
327         }
328     }
329 
330     trace_qemu_co_mutex_unlock_return(mutex, self);
331 }
332 
333 void qemu_co_rwlock_init(CoRwlock *lock)
334 {
335     memset(lock, 0, sizeof(*lock));
336     qemu_co_queue_init(&lock->queue);
337     qemu_co_mutex_init(&lock->mutex);
338 }
339 
340 void qemu_co_rwlock_rdlock(CoRwlock *lock)
341 {
342     Coroutine *self = qemu_coroutine_self();
343 
344     qemu_co_mutex_lock(&lock->mutex);
345     /* For fairness, wait if a writer is in line.  */
346     while (lock->pending_writer) {
347         qemu_co_queue_wait(&lock->queue, &lock->mutex);
348     }
349     lock->reader++;
350     qemu_co_mutex_unlock(&lock->mutex);
351 
352     /* The rest of the read-side critical section is run without the mutex.  */
353     self->locks_held++;
354 }
355 
356 void qemu_co_rwlock_unlock(CoRwlock *lock)
357 {
358     Coroutine *self = qemu_coroutine_self();
359 
360     assert(qemu_in_coroutine());
361     if (!lock->reader) {
362         /* The critical section started in qemu_co_rwlock_wrlock.  */
363         qemu_co_queue_restart_all(&lock->queue);
364     } else {
365         self->locks_held--;
366 
367         qemu_co_mutex_lock(&lock->mutex);
368         lock->reader--;
369         assert(lock->reader >= 0);
370         /* Wakeup only one waiting writer */
371         if (!lock->reader) {
372             qemu_co_queue_next(&lock->queue);
373         }
374     }
375     qemu_co_mutex_unlock(&lock->mutex);
376 }
377 
378 void qemu_co_rwlock_downgrade(CoRwlock *lock)
379 {
380     Coroutine *self = qemu_coroutine_self();
381 
382     /* lock->mutex critical section started in qemu_co_rwlock_wrlock or
383      * qemu_co_rwlock_upgrade.
384      */
385     assert(lock->reader == 0);
386     lock->reader++;
387     qemu_co_mutex_unlock(&lock->mutex);
388 
389     /* The rest of the read-side critical section is run without the mutex.  */
390     self->locks_held++;
391 }
392 
393 void qemu_co_rwlock_wrlock(CoRwlock *lock)
394 {
395     qemu_co_mutex_lock(&lock->mutex);
396     lock->pending_writer++;
397     while (lock->reader) {
398         qemu_co_queue_wait(&lock->queue, &lock->mutex);
399     }
400     lock->pending_writer--;
401 
402     /* The rest of the write-side critical section is run with
403      * the mutex taken, so that lock->reader remains zero.
404      * There is no need to update self->locks_held.
405      */
406 }
407 
408 void qemu_co_rwlock_upgrade(CoRwlock *lock)
409 {
410     Coroutine *self = qemu_coroutine_self();
411 
412     qemu_co_mutex_lock(&lock->mutex);
413     assert(lock->reader > 0);
414     lock->reader--;
415     lock->pending_writer++;
416     while (lock->reader) {
417         qemu_co_queue_wait(&lock->queue, &lock->mutex);
418     }
419     lock->pending_writer--;
420 
421     /* The rest of the write-side critical section is run with
422      * the mutex taken, similar to qemu_co_rwlock_wrlock.  Do
423      * not account for the lock twice in self->locks_held.
424      */
425     self->locks_held--;
426 }
427