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