xref: /openbmc/qemu/util/qemu-coroutine-lock.c (revision 80bdd58ae41ee7f65081c33e5f96cd6033da8337)
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  
qemu_co_queue_init(CoQueue * queue)36  void qemu_co_queue_init(CoQueue *queue)
37  {
38      QSIMPLEQ_INIT(&queue->entries);
39  }
40  
qemu_co_queue_wait_impl(CoQueue * queue,QemuLockable * lock,CoQueueWaitFlags flags)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  
qemu_co_enter_next_impl(CoQueue * queue,QemuLockable * lock)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  
qemu_co_queue_next(CoQueue * queue)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  
qemu_co_enter_all_impl(CoQueue * queue,QemuLockable * lock)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  
qemu_co_queue_restart_all(CoQueue * queue)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  
qemu_co_queue_empty(CoQueue * queue)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  
push_waiter(CoMutex * mutex,CoWaitRecord * w)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  
move_waiters(CoMutex * mutex)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  
pop_waiter(CoMutex * mutex)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  
has_waiters(CoMutex * mutex)174  static bool has_waiters(CoMutex *mutex)
175  {
176      return QSLIST_EMPTY(&mutex->to_pop) || QSLIST_EMPTY(&mutex->from_push);
177  }
178  
qemu_co_mutex_init(CoMutex * mutex)179  void qemu_co_mutex_init(CoMutex *mutex)
180  {
181      memset(mutex, 0, sizeof(*mutex));
182  }
183  
qemu_co_mutex_wake(CoMutex * mutex,Coroutine * co)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  
qemu_co_mutex_lock_slowpath(AioContext * ctx,CoMutex * mutex)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_set_mb
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  
qemu_co_mutex_lock(CoMutex * mutex)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  
qemu_co_mutex_unlock(CoMutex * mutex)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_set_mb(&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  
qemu_co_rwlock_init(CoRwlock * lock)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.  */
qemu_co_rwlock_maybe_wake_one(CoRwlock * lock)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  
qemu_co_rwlock_rdlock(CoRwlock * lock)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  
qemu_co_rwlock_unlock(CoRwlock * lock)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  
qemu_co_rwlock_downgrade(CoRwlock * lock)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  
qemu_co_rwlock_wrlock(CoRwlock * lock)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  
qemu_co_rwlock_upgrade(CoRwlock * lock)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