util/async.c

   1 /*
   2  * Data plane event loop
   3  *
   4  * Copyright (c) 2003-2008 Fabrice Bellard
   5  * Copyright (c) 2009-2017 QEMU contributors
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a copy
   8  * of this software and associated documentation files (the "Software"), to deal
   9  * in the Software without restriction, including without limitation the rights
  10  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  11  * copies of the Software, and to permit persons to whom the Software is
  12  * furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice shall be included in
  15  * all copies or substantial portions of the Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  23  * THE SOFTWARE.
  24  */
  25
  26 #include "qemu/osdep.h"
  27 #include "qapi/error.h"
  28 #include "block/aio.h"
  29 #include "block/thread-pool.h"
  30 #include "block/graph-lock.h"
  31 #include "qemu/main-loop.h"
  32 #include "qemu/atomic.h"
  33 #include "qemu/rcu_queue.h"
  34 #include "block/raw-aio.h"
  35 #include "qemu/coroutine_int.h"
  36 #include "qemu/coroutine-tls.h"
  37 #include "sysemu/cpu-timers.h"
  38 #include "trace.h"
  39
  40 /***********************************************************/
  41 /* bottom halves (can be seen as timers which expire ASAP) */
  42
  43 /* QEMUBH::flags values */
  44 enum {
  45     /* Already enqueued and waiting for aio_bh_poll() */
  46     BH_PENDING   = (1 << 0),
  47
  48     /* Invoke the callback */
  49     BH_SCHEDULED = (1 << 1),
  50
  51     /* Delete without invoking callback */
  52     BH_DELETED   = (1 << 2),
  53
  54     /* Delete after invoking callback */
  55     BH_ONESHOT   = (1 << 3),
  56
  57     /* Schedule periodically when the event loop is idle */
  58     BH_IDLE      = (1 << 4),
  59 };
  60
  61 struct QEMUBH {
  62     AioContext *ctx;
  63     const char *name;
  64     QEMUBHFunc *cb;
  65     void *opaque;
  66     QSLIST_ENTRY(QEMUBH) next;
  67     unsigned flags;
  68     MemReentrancyGuard *reentrancy_guard;
  69 };
  70
  71 /* Called concurrently from any thread */
  72 static void aio_bh_enqueue(QEMUBH *bh, unsigned new_flags)
  73 {
  74     AioContext *ctx = bh->ctx;
  75     unsigned old_flags;
  76
  77     /*
  78      * Synchronizes with atomic_fetch_and() in aio_bh_dequeue(), ensuring that
  79      * insertion starts after BH_PENDING is set.
  80      */
  81     old_flags = qatomic_fetch_or(&bh->flags, BH_PENDING | new_flags);
  82
  83     if (!(old_flags & BH_PENDING)) {
  84         /*
  85          * At this point the bottom half becomes visible to aio_bh_poll().
  86          * This insertion thus synchronizes with QSLIST_MOVE_ATOMIC in
  87          * aio_bh_poll(), ensuring that:
  88          * 1. any writes needed by the callback are visible from the callback
  89          *    after aio_bh_dequeue() returns bh.
  90          * 2. ctx is loaded before the callback has a chance to execute and bh
  91          *    could be freed.
  92          */
  93         QSLIST_INSERT_HEAD_ATOMIC(&ctx->bh_list, bh, next);
  94     }
  95
  96     aio_notify(ctx);
  97     /*
  98      * Workaround for record/replay.
  99      * vCPU execution should be suspended when new BH is set.
 100      * This is needed to avoid guest timeouts caused
 101      * by the long cycles of the execution.
 102      */
 103     icount_notify_exit();
 104 }
 105
 106 /* Only called from aio_bh_poll() and aio_ctx_finalize() */
 107 static QEMUBH *aio_bh_dequeue(BHList *head, unsigned *flags)
 108 {
 109     QEMUBH *bh = QSLIST_FIRST_RCU(head);
 110
 111     if (!bh) {
 112         return NULL;
 113     }
 114
 115     QSLIST_REMOVE_HEAD(head, next);
 116
 117     /*
 118      * Synchronizes with qatomic_fetch_or() in aio_bh_enqueue(), ensuring that
 119      * the removal finishes before BH_PENDING is reset.
 120      */
 121     *flags = qatomic_fetch_and(&bh->flags,
 122                               ~(BH_PENDING | BH_SCHEDULED | BH_IDLE));
 123     return bh;
 124 }
 125
 126 void aio_bh_schedule_oneshot_full(AioContext *ctx, QEMUBHFunc *cb,
 127                                   void *opaque, const char *name)
 128 {
 129     QEMUBH *bh;
 130     bh = g_new(QEMUBH, 1);
 131     *bh = (QEMUBH){
 132         .ctx = ctx,
 133         .cb = cb,
 134         .opaque = opaque,
 135         .name = name,
 136     };
 137     aio_bh_enqueue(bh, BH_SCHEDULED | BH_ONESHOT);
 138 }
 139
 140 QEMUBH *aio_bh_new_full(AioContext *ctx, QEMUBHFunc *cb, void *opaque,
 141                         const char *name, MemReentrancyGuard *reentrancy_guard)
 142 {
 143     QEMUBH *bh;
 144     bh = g_new(QEMUBH, 1);
 145     *bh = (QEMUBH){
 146         .ctx = ctx,
 147         .cb = cb,
 148         .opaque = opaque,
 149         .name = name,
 150         .reentrancy_guard = reentrancy_guard,
 151     };
 152     return bh;
 153 }
 154
 155 void aio_bh_call(QEMUBH *bh)
 156 {
 157     bool last_engaged_in_io = false;
 158
 159     /* Make a copy of the guard-pointer as cb may free the bh */
 160     MemReentrancyGuard *reentrancy_guard = bh->reentrancy_guard;
 161     if (reentrancy_guard) {
 162         last_engaged_in_io = reentrancy_guard->engaged_in_io;
 163         if (reentrancy_guard->engaged_in_io) {
 164             trace_reentrant_aio(bh->ctx, bh->name);
 165         }
 166         reentrancy_guard->engaged_in_io = true;
 167     }
 168
 169     bh->cb(bh->opaque);
 170
 171     if (reentrancy_guard) {
 172         reentrancy_guard->engaged_in_io = last_engaged_in_io;
 173     }
 174 }
 175
 176 /* Multiple occurrences of aio_bh_poll cannot be called concurrently. */
 177 int aio_bh_poll(AioContext *ctx)
 178 {
 179     BHListSlice slice;
 180     BHListSlice *s;
 181     int ret = 0;
 182
 183     /* Synchronizes with QSLIST_INSERT_HEAD_ATOMIC in aio_bh_enqueue().  */
 184     QSLIST_MOVE_ATOMIC(&slice.bh_list, &ctx->bh_list);
 185
 186     /*
 187      * GCC13 [-Werror=dangling-pointer=] complains that the local variable
 188      * 'slice' is being stored in the global 'ctx->bh_slice_list' but the
 189      * list is emptied before this function returns.
 190      */
 191 #if !defined(__clang__)
 192 #pragma GCC diagnostic push
 193 #pragma GCC diagnostic ignored "-Wpragmas"
 194 #pragma GCC diagnostic ignored "-Wdangling-pointer="
 195 #endif
 196     QSIMPLEQ_INSERT_TAIL(&ctx->bh_slice_list, &slice, next);
 197 #if !defined(__clang__)
 198 #pragma GCC diagnostic pop
 199 #endif
 200
 201     while ((s = QSIMPLEQ_FIRST(&ctx->bh_slice_list))) {
 202         QEMUBH *bh;
 203         unsigned flags;
 204
 205         bh = aio_bh_dequeue(&s->bh_list, &flags);
 206         if (!bh) {
 207             QSIMPLEQ_REMOVE_HEAD(&ctx->bh_slice_list, next);
 208             continue;
 209         }
 210
 211         if ((flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
 212             /* Idle BHs don't count as progress */
 213             if (!(flags & BH_IDLE)) {
 214                 ret = 1;
 215             }
 216             aio_bh_call(bh);
 217         }
 218         if (flags & (BH_DELETED | BH_ONESHOT)) {
 219             g_free(bh);
 220         }
 221     }
 222
 223     return ret;
 224 }
 225
 226 void qemu_bh_schedule_idle(QEMUBH *bh)
 227 {
 228     aio_bh_enqueue(bh, BH_SCHEDULED | BH_IDLE);
 229 }
 230
 231 void qemu_bh_schedule(QEMUBH *bh)
 232 {
 233     aio_bh_enqueue(bh, BH_SCHEDULED);
 234 }
 235
 236 /* This func is async.
 237  */
 238 void qemu_bh_cancel(QEMUBH *bh)
 239 {
 240     qatomic_and(&bh->flags, ~BH_SCHEDULED);
 241 }
 242
 243 /* This func is async.The bottom half will do the delete action at the finial
 244  * end.
 245  */
 246 void qemu_bh_delete(QEMUBH *bh)
 247 {
 248     aio_bh_enqueue(bh, BH_DELETED);
 249 }
 250
 251 static int64_t aio_compute_bh_timeout(BHList *head, int timeout)
 252 {
 253     QEMUBH *bh;
 254
 255     QSLIST_FOREACH_RCU(bh, head, next) {
 256         if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
 257             if (bh->flags & BH_IDLE) {
 258                 /* idle bottom halves will be polled at least
 259                  * every 10ms */
 260                 timeout = 10000000;
 261             } else {
 262                 /* non-idle bottom halves will be executed
 263                  * immediately */
 264                 return 0;
 265             }
 266         }
 267     }
 268
 269     return timeout;
 270 }
 271
 272 int64_t
 273 aio_compute_timeout(AioContext *ctx)
 274 {
 275     BHListSlice *s;
 276     int64_t deadline;
 277     int timeout = -1;
 278
 279     timeout = aio_compute_bh_timeout(&ctx->bh_list, timeout);
 280     if (timeout == 0) {
 281         return 0;
 282     }
 283
 284     QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) {
 285         timeout = aio_compute_bh_timeout(&s->bh_list, timeout);
 286         if (timeout == 0) {
 287             return 0;
 288         }
 289     }
 290
 291     deadline = timerlistgroup_deadline_ns(&ctx->tlg);
 292     if (deadline == 0) {
 293         return 0;
 294     } else {
 295         return qemu_soonest_timeout(timeout, deadline);
 296     }
 297 }
 298
 299 static gboolean
 300 aio_ctx_prepare(GSource *source, gint    *timeout)
 301 {
 302     AioContext *ctx = (AioContext *) source;
 303
 304     qatomic_set(&ctx->notify_me, qatomic_read(&ctx->notify_me) | 1);
 305
 306     /*
 307      * Write ctx->notify_me before computing the timeout
 308      * (reading bottom half flags, etc.).  Pairs with
 309      * smp_mb in aio_notify().
 310      */
 311     smp_mb();
 312
 313     /* We assume there is no timeout already supplied */
 314     *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx));
 315
 316     if (aio_prepare(ctx)) {
 317         *timeout = 0;
 318     }
 319
 320     return *timeout == 0;
 321 }
 322
 323 static gboolean
 324 aio_ctx_check(GSource *source)
 325 {
 326     AioContext *ctx = (AioContext *) source;
 327     QEMUBH *bh;
 328     BHListSlice *s;
 329
 330     /* Finish computing the timeout before clearing the flag.  */
 331     qatomic_store_release(&ctx->notify_me, qatomic_read(&ctx->notify_me) & ~1);
 332     aio_notify_accept(ctx);
 333
 334     QSLIST_FOREACH_RCU(bh, &ctx->bh_list, next) {
 335         if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
 336             return true;
 337         }
 338     }
 339
 340     QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) {
 341         QSLIST_FOREACH_RCU(bh, &s->bh_list, next) {
 342             if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
 343                 return true;
 344             }
 345         }
 346     }
 347     return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0);
 348 }
 349
 350 static gboolean
 351 aio_ctx_dispatch(GSource     *source,
 352                  GSourceFunc  callback,
 353                  gpointer     user_data)
 354 {
 355     AioContext *ctx = (AioContext *) source;
 356
 357     assert(callback == NULL);
 358     aio_dispatch(ctx);
 359     return true;
 360 }
 361
 362 static void
 363 aio_ctx_finalize(GSource     *source)
 364 {
 365     AioContext *ctx = (AioContext *) source;
 366     QEMUBH *bh;
 367     unsigned flags;
 368
 369     thread_pool_free(ctx->thread_pool);
 370
 371 #ifdef CONFIG_LINUX_AIO
 372     if (ctx->linux_aio) {
 373         laio_detach_aio_context(ctx->linux_aio, ctx);
 374         laio_cleanup(ctx->linux_aio);
 375         ctx->linux_aio = NULL;
 376     }
 377 #endif
 378
 379 #ifdef CONFIG_LINUX_IO_URING
 380     if (ctx->linux_io_uring) {
 381         luring_detach_aio_context(ctx->linux_io_uring, ctx);
 382         luring_cleanup(ctx->linux_io_uring);
 383         ctx->linux_io_uring = NULL;
 384     }
 385 #endif
 386
 387     assert(QSLIST_EMPTY(&ctx->scheduled_coroutines));
 388     qemu_bh_delete(ctx->co_schedule_bh);
 389
 390     /* There must be no aio_bh_poll() calls going on */
 391     assert(QSIMPLEQ_EMPTY(&ctx->bh_slice_list));
 392
 393     while ((bh = aio_bh_dequeue(&ctx->bh_list, &flags))) {
 394         /*
 395          * qemu_bh_delete() must have been called on BHs in this AioContext. In
 396          * many cases memory leaks, hangs, or inconsistent state occur when a
 397          * BH is leaked because something still expects it to run.
 398          *
 399          * If you hit this, fix the lifecycle of the BH so that
 400          * qemu_bh_delete() and any associated cleanup is called before the
 401          * AioContext is finalized.
 402          */
 403         if (unlikely(!(flags & BH_DELETED))) {
 404             fprintf(stderr, "%s: BH '%s' leaked, aborting...\n",
 405                     __func__, bh->name);
 406             abort();
 407         }
 408
 409         g_free(bh);
 410     }
 411
 412     aio_set_event_notifier(ctx, &ctx->notifier, NULL, NULL, NULL);
 413     event_notifier_cleanup(&ctx->notifier);
 414     qemu_rec_mutex_destroy(&ctx->lock);
 415     qemu_lockcnt_destroy(&ctx->list_lock);
 416     timerlistgroup_deinit(&ctx->tlg);
 417     unregister_aiocontext(ctx);
 418     aio_context_destroy(ctx);
 419 }
 420
 421 static GSourceFuncs aio_source_funcs = {
 422     aio_ctx_prepare,
 423     aio_ctx_check,
 424     aio_ctx_dispatch,
 425     aio_ctx_finalize
 426 };
 427
 428 GSource *aio_get_g_source(AioContext *ctx)
 429 {
 430     aio_context_use_g_source(ctx);
 431     g_source_ref(&ctx->source);
 432     return &ctx->source;
 433 }
 434
 435 ThreadPool *aio_get_thread_pool(AioContext *ctx)
 436 {
 437     if (!ctx->thread_pool) {
 438         ctx->thread_pool = thread_pool_new(ctx);
 439     }
 440     return ctx->thread_pool;
 441 }
 442
 443 #ifdef CONFIG_LINUX_AIO
 444 LinuxAioState *aio_setup_linux_aio(AioContext *ctx, Error **errp)
 445 {
 446     if (!ctx->linux_aio) {
 447         ctx->linux_aio = laio_init(errp);
 448         if (ctx->linux_aio) {
 449             laio_attach_aio_context(ctx->linux_aio, ctx);
 450         }
 451     }
 452     return ctx->linux_aio;
 453 }
 454
 455 LinuxAioState *aio_get_linux_aio(AioContext *ctx)
 456 {
 457     assert(ctx->linux_aio);
 458     return ctx->linux_aio;
 459 }
 460 #endif
 461
 462 #ifdef CONFIG_LINUX_IO_URING
 463 LuringState *aio_setup_linux_io_uring(AioContext *ctx, Error **errp)
 464 {
 465     if (ctx->linux_io_uring) {
 466         return ctx->linux_io_uring;
 467     }
 468
 469     ctx->linux_io_uring = luring_init(errp);
 470     if (!ctx->linux_io_uring) {
 471         return NULL;
 472     }
 473
 474     luring_attach_aio_context(ctx->linux_io_uring, ctx);
 475     return ctx->linux_io_uring;
 476 }
 477
 478 LuringState *aio_get_linux_io_uring(AioContext *ctx)
 479 {
 480     assert(ctx->linux_io_uring);
 481     return ctx->linux_io_uring;
 482 }
 483 #endif
 484
 485 void aio_notify(AioContext *ctx)
 486 {
 487     /*
 488      * Write e.g. ctx->bh_list before writing ctx->notified.  Pairs with
 489      * smp_mb() in aio_notify_accept().
 490      */
 491     smp_wmb();
 492     qatomic_set(&ctx->notified, true);
 493
 494     /*
 495      * Write ctx->notified (and also ctx->bh_list) before reading ctx->notify_me.
 496      * Pairs with smp_mb() in aio_ctx_prepare or aio_poll.
 497      */
 498     smp_mb();
 499     if (qatomic_read(&ctx->notify_me)) {
 500         event_notifier_set(&ctx->notifier);
 501     }
 502 }
 503
 504 void aio_notify_accept(AioContext *ctx)
 505 {
 506     qatomic_set(&ctx->notified, false);
 507
 508     /*
 509      * Order reads of ctx->notified (in aio_context_notifier_poll()) and the
 510      * above clearing of ctx->notified before reads of e.g. bh->flags.  Pairs
 511      * with smp_wmb() in aio_notify.
 512      */
 513     smp_mb();
 514 }
 515
 516 static void aio_timerlist_notify(void *opaque, QEMUClockType type)
 517 {
 518     aio_notify(opaque);
 519 }
 520
 521 static void aio_context_notifier_cb(EventNotifier *e)
 522 {
 523     AioContext *ctx = container_of(e, AioContext, notifier);
 524
 525     event_notifier_test_and_clear(&ctx->notifier);
 526 }
 527
 528 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
 529 static bool aio_context_notifier_poll(void *opaque)
 530 {
 531     EventNotifier *e = opaque;
 532     AioContext *ctx = container_of(e, AioContext, notifier);
 533
 534     /*
 535      * No need for load-acquire because we just want to kick the
 536      * event loop.  aio_notify_accept() takes care of synchronizing
 537      * the event loop with the producers.
 538      */
 539     return qatomic_read(&ctx->notified);
 540 }
 541
 542 static void aio_context_notifier_poll_ready(EventNotifier *e)
 543 {
 544     /* Do nothing, we just wanted to kick the event loop */
 545 }
 546
 547 static void co_schedule_bh_cb(void *opaque)
 548 {
 549     AioContext *ctx = opaque;
 550     QSLIST_HEAD(, Coroutine) straight, reversed;
 551
 552     QSLIST_MOVE_ATOMIC(&reversed, &ctx->scheduled_coroutines);
 553     QSLIST_INIT(&straight);
 554
 555     while (!QSLIST_EMPTY(&reversed)) {
 556         Coroutine *co = QSLIST_FIRST(&reversed);
 557         QSLIST_REMOVE_HEAD(&reversed, co_scheduled_next);
 558         QSLIST_INSERT_HEAD(&straight, co, co_scheduled_next);
 559     }
 560
 561     while (!QSLIST_EMPTY(&straight)) {
 562         Coroutine *co = QSLIST_FIRST(&straight);
 563         QSLIST_REMOVE_HEAD(&straight, co_scheduled_next);
 564         trace_aio_co_schedule_bh_cb(ctx, co);
 565         aio_context_acquire(ctx);
 566
 567         /* Protected by write barrier in qemu_aio_coroutine_enter */
 568         qatomic_set(&co->scheduled, NULL);
 569         qemu_aio_coroutine_enter(ctx, co);
 570         aio_context_release(ctx);
 571     }
 572 }
 573
 574 AioContext *aio_context_new(Error **errp)
 575 {
 576     int ret;
 577     AioContext *ctx;
 578
 579     ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
 580     QSLIST_INIT(&ctx->bh_list);
 581     QSIMPLEQ_INIT(&ctx->bh_slice_list);
 582     aio_context_setup(ctx);
 583
 584     ret = event_notifier_init(&ctx->notifier, false);
 585     if (ret < 0) {
 586         error_setg_errno(errp, -ret, "Failed to initialize event notifier");
 587         goto fail;
 588     }
 589     g_source_set_can_recurse(&ctx->source, true);
 590     qemu_lockcnt_init(&ctx->list_lock);
 591
 592     ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx);
 593     QSLIST_INIT(&ctx->scheduled_coroutines);
 594
 595     aio_set_event_notifier(ctx, &ctx->notifier,
 596                            aio_context_notifier_cb,
 597                            aio_context_notifier_poll,
 598                            aio_context_notifier_poll_ready);
 599 #ifdef CONFIG_LINUX_AIO
 600     ctx->linux_aio = NULL;
 601 #endif
 602
 603 #ifdef CONFIG_LINUX_IO_URING
 604     ctx->linux_io_uring = NULL;
 605 #endif
 606
 607     ctx->thread_pool = NULL;
 608     qemu_rec_mutex_init(&ctx->lock);
 609     timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx);
 610
 611     ctx->poll_ns = 0;
 612     ctx->poll_max_ns = 0;
 613     ctx->poll_grow = 0;
 614     ctx->poll_shrink = 0;
 615
 616     ctx->aio_max_batch = 0;
 617
 618     ctx->thread_pool_min = 0;
 619     ctx->thread_pool_max = THREAD_POOL_MAX_THREADS_DEFAULT;
 620
 621     register_aiocontext(ctx);
 622
 623     return ctx;
 624 fail:
 625     g_source_destroy(&ctx->source);
 626     return NULL;
 627 }
 628
 629 void aio_co_schedule(AioContext *ctx, Coroutine *co)
 630 {
 631     trace_aio_co_schedule(ctx, co);
 632     const char *scheduled = qatomic_cmpxchg(&co->scheduled, NULL,
 633                                            __func__);
 634
 635     if (scheduled) {
 636         fprintf(stderr,
 637                 "%s: Co-routine was already scheduled in '%s'\n",
 638                 __func__, scheduled);
 639         abort();
 640     }
 641
 642     /* The coroutine might run and release the last ctx reference before we
 643      * invoke qemu_bh_schedule().  Take a reference to keep ctx alive until
 644      * we're done.
 645      */
 646     aio_context_ref(ctx);
 647
 648     QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines,
 649                               co, co_scheduled_next);
 650     qemu_bh_schedule(ctx->co_schedule_bh);
 651
 652     aio_context_unref(ctx);
 653 }
 654
 655 typedef struct AioCoRescheduleSelf {
 656     Coroutine *co;
 657     AioContext *new_ctx;
 658 } AioCoRescheduleSelf;
 659
 660 static void aio_co_reschedule_self_bh(void *opaque)
 661 {
 662     AioCoRescheduleSelf *data = opaque;
 663     aio_co_schedule(data->new_ctx, data->co);
 664 }
 665
 666 void coroutine_fn aio_co_reschedule_self(AioContext *new_ctx)
 667 {
 668     AioContext *old_ctx = qemu_get_current_aio_context();
 669
 670     if (old_ctx != new_ctx) {
 671         AioCoRescheduleSelf data = {
 672             .co = qemu_coroutine_self(),
 673             .new_ctx = new_ctx,
 674         };
 675         /*
 676          * We can't directly schedule the coroutine in the target context
 677          * because this would be racy: The other thread could try to enter the
 678          * coroutine before it has yielded in this one.
 679          */
 680         aio_bh_schedule_oneshot(old_ctx, aio_co_reschedule_self_bh, &data);
 681         qemu_coroutine_yield();
 682     }
 683 }
 684
 685 void aio_co_wake(Coroutine *co)
 686 {
 687     AioContext *ctx;
 688
 689     /* Read coroutine before co->ctx.  Matches smp_wmb in
 690      * qemu_coroutine_enter.
 691      */
 692     smp_read_barrier_depends();
 693     ctx = qatomic_read(&co->ctx);
 694
 695     aio_co_enter(ctx, co);
 696 }
 697
 698 void aio_co_enter(AioContext *ctx, Coroutine *co)
 699 {
 700     if (ctx != qemu_get_current_aio_context()) {
 701         aio_co_schedule(ctx, co);
 702         return;
 703     }
 704
 705     if (qemu_in_coroutine()) {
 706         Coroutine *self = qemu_coroutine_self();
 707         assert(self != co);
 708         QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, co, co_queue_next);
 709     } else {
 710         aio_context_acquire(ctx);
 711         qemu_aio_coroutine_enter(ctx, co);
 712         aio_context_release(ctx);
 713     }
 714 }
 715
 716 void aio_context_ref(AioContext *ctx)
 717 {
 718     g_source_ref(&ctx->source);
 719 }
 720
 721 void aio_context_unref(AioContext *ctx)
 722 {
 723     g_source_unref(&ctx->source);
 724 }
 725
 726 void aio_context_acquire(AioContext *ctx)
 727 {
 728     qemu_rec_mutex_lock(&ctx->lock);
 729 }
 730
 731 void aio_context_release(AioContext *ctx)
 732 {
 733     qemu_rec_mutex_unlock(&ctx->lock);
 734 }
 735
 736 QEMU_DEFINE_STATIC_CO_TLS(AioContext *, my_aiocontext)
 737
 738 AioContext *qemu_get_current_aio_context(void)
 739 {
 740     AioContext *ctx = get_my_aiocontext();
 741     if (ctx) {
 742         return ctx;
 743     }
 744     if (qemu_mutex_iothread_locked()) {
 745         /* Possibly in a vCPU thread.  */
 746         return qemu_get_aio_context();
 747     }
 748     return NULL;
 749 }
 750
 751 void qemu_set_current_aio_context(AioContext *ctx)
 752 {
 753     assert(!get_my_aiocontext());
 754     set_my_aiocontext(ctx);
 755 }
 756
 757 void aio_context_set_thread_pool_params(AioContext *ctx, int64_t min,
 758                                         int64_t max, Error **errp)
 759 {
 760
 761     if (min > max || !max || min > INT_MAX || max > INT_MAX) {
 762         error_setg(errp, "bad thread-pool-min/thread-pool-max values");
 763         return;
 764     }
 765
 766     ctx->thread_pool_min = min;
 767     ctx->thread_pool_max = max;
 768
 769     if (ctx->thread_pool) {
 770         thread_pool_update_params(ctx->thread_pool, ctx);
 771     }
 772 }