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 "qemu/main-loop.h"
  31 #include "qemu/atomic.h"
  32 #include "qemu/rcu_queue.h"
  33 #include "block/raw-aio.h"
  34 #include "qemu/coroutine_int.h"
  35 #include "trace.h"
  36
  37 /***********************************************************/
  38 /* bottom halves (can be seen as timers which expire ASAP) */
  39
  40 /* QEMUBH::flags values */
  41 enum {
  42     /* Already enqueued and waiting for aio_bh_poll() */
  43     BH_PENDING   = (1 << 0),
  44
  45     /* Invoke the callback */
  46     BH_SCHEDULED = (1 << 1),
  47
  48     /* Delete without invoking callback */
  49     BH_DELETED   = (1 << 2),
  50
  51     /* Delete after invoking callback */
  52     BH_ONESHOT   = (1 << 3),
  53
  54     /* Schedule periodically when the event loop is idle */
  55     BH_IDLE      = (1 << 4),
  56 };
  57
  58 struct QEMUBH {
  59     AioContext *ctx;
  60     QEMUBHFunc *cb;
  61     void *opaque;
  62     QSLIST_ENTRY(QEMUBH) next;
  63     unsigned flags;
  64 };
  65
  66 /* Called concurrently from any thread */
  67 static void aio_bh_enqueue(QEMUBH *bh, unsigned new_flags)
  68 {
  69     AioContext *ctx = bh->ctx;
  70     unsigned old_flags;
  71
  72     /*
  73      * The memory barrier implicit in atomic_fetch_or makes sure that:
  74      * 1. idle & any writes needed by the callback are done before the
  75      *    locations are read in the aio_bh_poll.
  76      * 2. ctx is loaded before the callback has a chance to execute and bh
  77      *    could be freed.
  78      */
  79     old_flags = atomic_fetch_or(&bh->flags, BH_PENDING | new_flags);
  80     if (!(old_flags & BH_PENDING)) {
  81         QSLIST_INSERT_HEAD_ATOMIC(&ctx->bh_list, bh, next);
  82     }
  83
  84     aio_notify(ctx);
  85 }
  86
  87 /* Only called from aio_bh_poll() and aio_ctx_finalize() */
  88 static QEMUBH *aio_bh_dequeue(BHList *head, unsigned *flags)
  89 {
  90     QEMUBH *bh = QSLIST_FIRST_RCU(head);
  91
  92     if (!bh) {
  93         return NULL;
  94     }
  95
  96     QSLIST_REMOVE_HEAD(head, next);
  97
  98     /*
  99      * The atomic_and is paired with aio_bh_enqueue().  The implicit memory
 100      * barrier ensures that the callback sees all writes done by the scheduling
 101      * thread.  It also ensures that the scheduling thread sees the cleared
 102      * flag before bh->cb has run, and thus will call aio_notify again if
 103      * necessary.
 104      */
 105     *flags = atomic_fetch_and(&bh->flags,
 106                               ~(BH_PENDING | BH_SCHEDULED | BH_IDLE));
 107     return bh;
 108 }
 109
 110 void aio_bh_schedule_oneshot(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
 111 {
 112     QEMUBH *bh;
 113     bh = g_new(QEMUBH, 1);
 114     *bh = (QEMUBH){
 115         .ctx = ctx,
 116         .cb = cb,
 117         .opaque = opaque,
 118     };
 119     aio_bh_enqueue(bh, BH_SCHEDULED | BH_ONESHOT);
 120 }
 121
 122 QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque)
 123 {
 124     QEMUBH *bh;
 125     bh = g_new(QEMUBH, 1);
 126     *bh = (QEMUBH){
 127         .ctx = ctx,
 128         .cb = cb,
 129         .opaque = opaque,
 130     };
 131     return bh;
 132 }
 133
 134 void aio_bh_call(QEMUBH *bh)
 135 {
 136     bh->cb(bh->opaque);
 137 }
 138
 139 /* Multiple occurrences of aio_bh_poll cannot be called concurrently. */
 140 int aio_bh_poll(AioContext *ctx)
 141 {
 142     BHListSlice slice;
 143     BHListSlice *s;
 144     int ret = 0;
 145
 146     QSLIST_MOVE_ATOMIC(&slice.bh_list, &ctx->bh_list);
 147     QSIMPLEQ_INSERT_TAIL(&ctx->bh_slice_list, &slice, next);
 148
 149     while ((s = QSIMPLEQ_FIRST(&ctx->bh_slice_list))) {
 150         QEMUBH *bh;
 151         unsigned flags;
 152
 153         bh = aio_bh_dequeue(&s->bh_list, &flags);
 154         if (!bh) {
 155             QSIMPLEQ_REMOVE_HEAD(&ctx->bh_slice_list, next);
 156             continue;
 157         }
 158
 159         if ((flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
 160             /* Idle BHs don't count as progress */
 161             if (!(flags & BH_IDLE)) {
 162                 ret = 1;
 163             }
 164             aio_bh_call(bh);
 165         }
 166         if (flags & (BH_DELETED | BH_ONESHOT)) {
 167             g_free(bh);
 168         }
 169     }
 170
 171     return ret;
 172 }
 173
 174 void qemu_bh_schedule_idle(QEMUBH *bh)
 175 {
 176     aio_bh_enqueue(bh, BH_SCHEDULED | BH_IDLE);
 177 }
 178
 179 void qemu_bh_schedule(QEMUBH *bh)
 180 {
 181     aio_bh_enqueue(bh, BH_SCHEDULED);
 182 }
 183
 184 /* This func is async.
 185  */
 186 void qemu_bh_cancel(QEMUBH *bh)
 187 {
 188     atomic_and(&bh->flags, ~BH_SCHEDULED);
 189 }
 190
 191 /* This func is async.The bottom half will do the delete action at the finial
 192  * end.
 193  */
 194 void qemu_bh_delete(QEMUBH *bh)
 195 {
 196     aio_bh_enqueue(bh, BH_DELETED);
 197 }
 198
 199 static int64_t aio_compute_bh_timeout(BHList *head, int timeout)
 200 {
 201     QEMUBH *bh;
 202
 203     QSLIST_FOREACH_RCU(bh, head, next) {
 204         if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
 205             if (bh->flags & BH_IDLE) {
 206                 /* idle bottom halves will be polled at least
 207                  * every 10ms */
 208                 timeout = 10000000;
 209             } else {
 210                 /* non-idle bottom halves will be executed
 211                  * immediately */
 212                 return 0;
 213             }
 214         }
 215     }
 216
 217     return timeout;
 218 }
 219
 220 int64_t
 221 aio_compute_timeout(AioContext *ctx)
 222 {
 223     BHListSlice *s;
 224     int64_t deadline;
 225     int timeout = -1;
 226
 227     timeout = aio_compute_bh_timeout(&ctx->bh_list, timeout);
 228     if (timeout == 0) {
 229         return 0;
 230     }
 231
 232     QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) {
 233         timeout = aio_compute_bh_timeout(&s->bh_list, timeout);
 234         if (timeout == 0) {
 235             return 0;
 236         }
 237     }
 238
 239     deadline = timerlistgroup_deadline_ns(&ctx->tlg);
 240     if (deadline == 0) {
 241         return 0;
 242     } else {
 243         return qemu_soonest_timeout(timeout, deadline);
 244     }
 245 }
 246
 247 static gboolean
 248 aio_ctx_prepare(GSource *source, gint    *timeout)
 249 {
 250     AioContext *ctx = (AioContext *) source;
 251
 252     atomic_set(&ctx->notify_me, atomic_read(&ctx->notify_me) | 1);
 253
 254     /*
 255      * Write ctx->notify_me before computing the timeout
 256      * (reading bottom half flags, etc.).  Pairs with
 257      * smp_mb in aio_notify().
 258      */
 259     smp_mb();
 260
 261     /* We assume there is no timeout already supplied */
 262     *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx));
 263
 264     if (aio_prepare(ctx)) {
 265         *timeout = 0;
 266     }
 267
 268     return *timeout == 0;
 269 }
 270
 271 static gboolean
 272 aio_ctx_check(GSource *source)
 273 {
 274     AioContext *ctx = (AioContext *) source;
 275     QEMUBH *bh;
 276     BHListSlice *s;
 277
 278     /* Finish computing the timeout before clearing the flag.  */
 279     atomic_store_release(&ctx->notify_me, atomic_read(&ctx->notify_me) & ~1);
 280     aio_notify_accept(ctx);
 281
 282     QSLIST_FOREACH_RCU(bh, &ctx->bh_list, next) {
 283         if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
 284             return true;
 285         }
 286     }
 287
 288     QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) {
 289         QSLIST_FOREACH_RCU(bh, &s->bh_list, next) {
 290             if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) {
 291                 return true;
 292             }
 293         }
 294     }
 295     return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0);
 296 }
 297
 298 static gboolean
 299 aio_ctx_dispatch(GSource     *source,
 300                  GSourceFunc  callback,
 301                  gpointer     user_data)
 302 {
 303     AioContext *ctx = (AioContext *) source;
 304
 305     assert(callback == NULL);
 306     aio_dispatch(ctx);
 307     return true;
 308 }
 309
 310 static void
 311 aio_ctx_finalize(GSource     *source)
 312 {
 313     AioContext *ctx = (AioContext *) source;
 314     QEMUBH *bh;
 315     unsigned flags;
 316
 317     thread_pool_free(ctx->thread_pool);
 318
 319 #ifdef CONFIG_LINUX_AIO
 320     if (ctx->linux_aio) {
 321         laio_detach_aio_context(ctx->linux_aio, ctx);
 322         laio_cleanup(ctx->linux_aio);
 323         ctx->linux_aio = NULL;
 324     }
 325 #endif
 326
 327 #ifdef CONFIG_LINUX_IO_URING
 328     if (ctx->linux_io_uring) {
 329         luring_detach_aio_context(ctx->linux_io_uring, ctx);
 330         luring_cleanup(ctx->linux_io_uring);
 331         ctx->linux_io_uring = NULL;
 332     }
 333 #endif
 334
 335     assert(QSLIST_EMPTY(&ctx->scheduled_coroutines));
 336     qemu_bh_delete(ctx->co_schedule_bh);
 337
 338     /* There must be no aio_bh_poll() calls going on */
 339     assert(QSIMPLEQ_EMPTY(&ctx->bh_slice_list));
 340
 341     while ((bh = aio_bh_dequeue(&ctx->bh_list, &flags))) {
 342         /* qemu_bh_delete() must have been called on BHs in this AioContext */
 343         assert(flags & BH_DELETED);
 344
 345         g_free(bh);
 346     }
 347
 348     aio_set_event_notifier(ctx, &ctx->notifier, false, NULL, NULL);
 349     event_notifier_cleanup(&ctx->notifier);
 350     qemu_rec_mutex_destroy(&ctx->lock);
 351     qemu_lockcnt_destroy(&ctx->list_lock);
 352     timerlistgroup_deinit(&ctx->tlg);
 353     aio_context_destroy(ctx);
 354 }
 355
 356 static GSourceFuncs aio_source_funcs = {
 357     aio_ctx_prepare,
 358     aio_ctx_check,
 359     aio_ctx_dispatch,
 360     aio_ctx_finalize
 361 };
 362
 363 GSource *aio_get_g_source(AioContext *ctx)
 364 {
 365     g_source_ref(&ctx->source);
 366     return &ctx->source;
 367 }
 368
 369 ThreadPool *aio_get_thread_pool(AioContext *ctx)
 370 {
 371     if (!ctx->thread_pool) {
 372         ctx->thread_pool = thread_pool_new(ctx);
 373     }
 374     return ctx->thread_pool;
 375 }
 376
 377 #ifdef CONFIG_LINUX_AIO
 378 LinuxAioState *aio_setup_linux_aio(AioContext *ctx, Error **errp)
 379 {
 380     if (!ctx->linux_aio) {
 381         ctx->linux_aio = laio_init(errp);
 382         if (ctx->linux_aio) {
 383             laio_attach_aio_context(ctx->linux_aio, ctx);
 384         }
 385     }
 386     return ctx->linux_aio;
 387 }
 388
 389 LinuxAioState *aio_get_linux_aio(AioContext *ctx)
 390 {
 391     assert(ctx->linux_aio);
 392     return ctx->linux_aio;
 393 }
 394 #endif
 395
 396 #ifdef CONFIG_LINUX_IO_URING
 397 LuringState *aio_setup_linux_io_uring(AioContext *ctx, Error **errp)
 398 {
 399     if (ctx->linux_io_uring) {
 400         return ctx->linux_io_uring;
 401     }
 402
 403     ctx->linux_io_uring = luring_init(errp);
 404     if (!ctx->linux_io_uring) {
 405         return NULL;
 406     }
 407
 408     luring_attach_aio_context(ctx->linux_io_uring, ctx);
 409     return ctx->linux_io_uring;
 410 }
 411
 412 LuringState *aio_get_linux_io_uring(AioContext *ctx)
 413 {
 414     assert(ctx->linux_io_uring);
 415     return ctx->linux_io_uring;
 416 }
 417 #endif
 418
 419 void aio_notify(AioContext *ctx)
 420 {
 421     /* Write e.g. bh->scheduled before reading ctx->notify_me.  Pairs
 422      * with smp_mb in aio_ctx_prepare or aio_poll.
 423      */
 424     smp_mb();
 425     if (atomic_read(&ctx->notify_me)) {
 426         event_notifier_set(&ctx->notifier);
 427         atomic_mb_set(&ctx->notified, true);
 428     }
 429 }
 430
 431 void aio_notify_accept(AioContext *ctx)
 432 {
 433     if (atomic_xchg(&ctx->notified, false)
 434 #ifdef WIN32
 435         || true
 436 #endif
 437     ) {
 438         event_notifier_test_and_clear(&ctx->notifier);
 439     }
 440 }
 441
 442 static void aio_timerlist_notify(void *opaque, QEMUClockType type)
 443 {
 444     aio_notify(opaque);
 445 }
 446
 447 static void event_notifier_dummy_cb(EventNotifier *e)
 448 {
 449 }
 450
 451 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */
 452 static bool event_notifier_poll(void *opaque)
 453 {
 454     EventNotifier *e = opaque;
 455     AioContext *ctx = container_of(e, AioContext, notifier);
 456
 457     return atomic_read(&ctx->notified);
 458 }
 459
 460 static void co_schedule_bh_cb(void *opaque)
 461 {
 462     AioContext *ctx = opaque;
 463     QSLIST_HEAD(, Coroutine) straight, reversed;
 464
 465     QSLIST_MOVE_ATOMIC(&reversed, &ctx->scheduled_coroutines);
 466     QSLIST_INIT(&straight);
 467
 468     while (!QSLIST_EMPTY(&reversed)) {
 469         Coroutine *co = QSLIST_FIRST(&reversed);
 470         QSLIST_REMOVE_HEAD(&reversed, co_scheduled_next);
 471         QSLIST_INSERT_HEAD(&straight, co, co_scheduled_next);
 472     }
 473
 474     while (!QSLIST_EMPTY(&straight)) {
 475         Coroutine *co = QSLIST_FIRST(&straight);
 476         QSLIST_REMOVE_HEAD(&straight, co_scheduled_next);
 477         trace_aio_co_schedule_bh_cb(ctx, co);
 478         aio_context_acquire(ctx);
 479
 480         /* Protected by write barrier in qemu_aio_coroutine_enter */
 481         atomic_set(&co->scheduled, NULL);
 482         qemu_aio_coroutine_enter(ctx, co);
 483         aio_context_release(ctx);
 484     }
 485 }
 486
 487 AioContext *aio_context_new(Error **errp)
 488 {
 489     int ret;
 490     AioContext *ctx;
 491
 492     ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext));
 493     QSLIST_INIT(&ctx->bh_list);
 494     QSIMPLEQ_INIT(&ctx->bh_slice_list);
 495     aio_context_setup(ctx);
 496
 497     ret = event_notifier_init(&ctx->notifier, false);
 498     if (ret < 0) {
 499         error_setg_errno(errp, -ret, "Failed to initialize event notifier");
 500         goto fail;
 501     }
 502     g_source_set_can_recurse(&ctx->source, true);
 503     qemu_lockcnt_init(&ctx->list_lock);
 504
 505     ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx);
 506     QSLIST_INIT(&ctx->scheduled_coroutines);
 507
 508     aio_set_event_notifier(ctx, &ctx->notifier,
 509                            false,
 510                            event_notifier_dummy_cb,
 511                            event_notifier_poll);
 512 #ifdef CONFIG_LINUX_AIO
 513     ctx->linux_aio = NULL;
 514 #endif
 515
 516 #ifdef CONFIG_LINUX_IO_URING
 517     ctx->linux_io_uring = NULL;
 518 #endif
 519
 520     ctx->thread_pool = NULL;
 521     qemu_rec_mutex_init(&ctx->lock);
 522     timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx);
 523
 524     ctx->poll_ns = 0;
 525     ctx->poll_max_ns = 0;
 526     ctx->poll_grow = 0;
 527     ctx->poll_shrink = 0;
 528
 529     return ctx;
 530 fail:
 531     g_source_destroy(&ctx->source);
 532     return NULL;
 533 }
 534
 535 void aio_co_schedule(AioContext *ctx, Coroutine *co)
 536 {
 537     trace_aio_co_schedule(ctx, co);
 538     const char *scheduled = atomic_cmpxchg(&co->scheduled, NULL,
 539                                            __func__);
 540
 541     if (scheduled) {
 542         fprintf(stderr,
 543                 "%s: Co-routine was already scheduled in '%s'\n",
 544                 __func__, scheduled);
 545         abort();
 546     }
 547
 548     /* The coroutine might run and release the last ctx reference before we
 549      * invoke qemu_bh_schedule().  Take a reference to keep ctx alive until
 550      * we're done.
 551      */
 552     aio_context_ref(ctx);
 553
 554     QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines,
 555                               co, co_scheduled_next);
 556     qemu_bh_schedule(ctx->co_schedule_bh);
 557
 558     aio_context_unref(ctx);
 559 }
 560
 561 void aio_co_wake(struct Coroutine *co)
 562 {
 563     AioContext *ctx;
 564
 565     /* Read coroutine before co->ctx.  Matches smp_wmb in
 566      * qemu_coroutine_enter.
 567      */
 568     smp_read_barrier_depends();
 569     ctx = atomic_read(&co->ctx);
 570
 571     aio_co_enter(ctx, co);
 572 }
 573
 574 void aio_co_enter(AioContext *ctx, struct Coroutine *co)
 575 {
 576     if (ctx != qemu_get_current_aio_context()) {
 577         aio_co_schedule(ctx, co);
 578         return;
 579     }
 580
 581     if (qemu_in_coroutine()) {
 582         Coroutine *self = qemu_coroutine_self();
 583         assert(self != co);
 584         QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, co, co_queue_next);
 585     } else {
 586         aio_context_acquire(ctx);
 587         qemu_aio_coroutine_enter(ctx, co);
 588         aio_context_release(ctx);
 589     }
 590 }
 591
 592 void aio_context_ref(AioContext *ctx)
 593 {
 594     g_source_ref(&ctx->source);
 595 }
 596
 597 void aio_context_unref(AioContext *ctx)
 598 {
 599     g_source_unref(&ctx->source);
 600 }
 601
 602 void aio_context_acquire(AioContext *ctx)
 603 {
 604     qemu_rec_mutex_lock(&ctx->lock);
 605 }
 606
 607 void aio_context_release(AioContext *ctx)
 608 {
 609     qemu_rec_mutex_unlock(&ctx->lock);
 610 }