util/aio-posix.c

   1 /*
   2  * QEMU aio implementation
   3  *
   4  * Copyright IBM, Corp. 2008
   5  *
   6  * Authors:
   7  *  Anthony Liguori   <aliguori@us.ibm.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  * Contributions after 2012-01-13 are licensed under the terms of the
  13  * GNU GPL, version 2 or (at your option) any later version.
  14  */
  15
  16 #include "qemu/osdep.h"
  17 #include "qemu-common.h"
  18 #include "block/block.h"
  19 #include "qemu/rcu_queue.h"
  20 #include "qemu/sockets.h"
  21 #include "qemu/cutils.h"
  22 #include "trace.h"
  23 #ifdef CONFIG_EPOLL_CREATE1
  24 #include <sys/epoll.h>
  25 #endif
  26
  27 struct AioHandler
  28 {
  29     GPollFD pfd;
  30     IOHandler *io_read;
  31     IOHandler *io_write;
  32     AioPollFn *io_poll;
  33     IOHandler *io_poll_begin;
  34     IOHandler *io_poll_end;
  35     int deleted;
  36     void *opaque;
  37     bool is_external;
  38     QLIST_ENTRY(AioHandler) node;
  39 };
  40
  41 #ifdef CONFIG_EPOLL_CREATE1
  42
  43 /* The fd number threshold to switch to epoll */
  44 #define EPOLL_ENABLE_THRESHOLD 64
  45
  46 static void aio_epoll_disable(AioContext *ctx)
  47 {
  48     ctx->epoll_enabled = false;
  49     if (!ctx->epoll_available) {
  50         return;
  51     }
  52     ctx->epoll_available = false;
  53     close(ctx->epollfd);
  54 }
  55
  56 static inline int epoll_events_from_pfd(int pfd_events)
  57 {
  58     return (pfd_events & G_IO_IN ? EPOLLIN : 0) |
  59            (pfd_events & G_IO_OUT ? EPOLLOUT : 0) |
  60            (pfd_events & G_IO_HUP ? EPOLLHUP : 0) |
  61            (pfd_events & G_IO_ERR ? EPOLLERR : 0);
  62 }
  63
  64 static bool aio_epoll_try_enable(AioContext *ctx)
  65 {
  66     AioHandler *node;
  67     struct epoll_event event;
  68
  69     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
  70         int r;
  71         if (node->deleted || !node->pfd.events) {
  72             continue;
  73         }
  74         event.events = epoll_events_from_pfd(node->pfd.events);
  75         event.data.ptr = node;
  76         r = epoll_ctl(ctx->epollfd, EPOLL_CTL_ADD, node->pfd.fd, &event);
  77         if (r) {
  78             return false;
  79         }
  80     }
  81     ctx->epoll_enabled = true;
  82     return true;
  83 }
  84
  85 static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
  86 {
  87     struct epoll_event event;
  88     int r;
  89     int ctl;
  90
  91     if (!ctx->epoll_enabled) {
  92         return;
  93     }
  94     if (!node->pfd.events) {
  95         ctl = EPOLL_CTL_DEL;
  96     } else {
  97         event.data.ptr = node;
  98         event.events = epoll_events_from_pfd(node->pfd.events);
  99         ctl = is_new ? EPOLL_CTL_ADD : EPOLL_CTL_MOD;
 100     }
 101
 102     r = epoll_ctl(ctx->epollfd, ctl, node->pfd.fd, &event);
 103     if (r) {
 104         aio_epoll_disable(ctx);
 105     }
 106 }
 107
 108 static int aio_epoll(AioContext *ctx, GPollFD *pfds,
 109                      unsigned npfd, int64_t timeout)
 110 {
 111     AioHandler *node;
 112     int i, ret = 0;
 113     struct epoll_event events[128];
 114
 115     assert(npfd == 1);
 116     assert(pfds[0].fd == ctx->epollfd);
 117     if (timeout > 0) {
 118         ret = qemu_poll_ns(pfds, npfd, timeout);
 119     }
 120     if (timeout <= 0 || ret > 0) {
 121         ret = epoll_wait(ctx->epollfd, events,
 122                          ARRAY_SIZE(events),
 123                          timeout);
 124         if (ret <= 0) {
 125             goto out;
 126         }
 127         for (i = 0; i < ret; i++) {
 128             int ev = events[i].events;
 129             node = events[i].data.ptr;
 130             node->pfd.revents = (ev & EPOLLIN ? G_IO_IN : 0) |
 131                 (ev & EPOLLOUT ? G_IO_OUT : 0) |
 132                 (ev & EPOLLHUP ? G_IO_HUP : 0) |
 133                 (ev & EPOLLERR ? G_IO_ERR : 0);
 134         }
 135     }
 136 out:
 137     return ret;
 138 }
 139
 140 static bool aio_epoll_enabled(AioContext *ctx)
 141 {
 142     /* Fall back to ppoll when external clients are disabled. */
 143     return !aio_external_disabled(ctx) && ctx->epoll_enabled;
 144 }
 145
 146 static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
 147                                  unsigned npfd, int64_t timeout)
 148 {
 149     if (!ctx->epoll_available) {
 150         return false;
 151     }
 152     if (aio_epoll_enabled(ctx)) {
 153         return true;
 154     }
 155     if (npfd >= EPOLL_ENABLE_THRESHOLD) {
 156         if (aio_epoll_try_enable(ctx)) {
 157             return true;
 158         } else {
 159             aio_epoll_disable(ctx);
 160         }
 161     }
 162     return false;
 163 }
 164
 165 #else
 166
 167 static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
 168 {
 169 }
 170
 171 static int aio_epoll(AioContext *ctx, GPollFD *pfds,
 172                      unsigned npfd, int64_t timeout)
 173 {
 174     assert(false);
 175 }
 176
 177 static bool aio_epoll_enabled(AioContext *ctx)
 178 {
 179     return false;
 180 }
 181
 182 static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
 183                           unsigned npfd, int64_t timeout)
 184 {
 185     return false;
 186 }
 187
 188 #endif
 189
 190 static AioHandler *find_aio_handler(AioContext *ctx, int fd)
 191 {
 192     AioHandler *node;
 193
 194     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 195         if (node->pfd.fd == fd)
 196             if (!node->deleted)
 197                 return node;
 198     }
 199
 200     return NULL;
 201 }
 202
 203 void aio_set_fd_handler(AioContext *ctx,
 204                         int fd,
 205                         bool is_external,
 206                         IOHandler *io_read,
 207                         IOHandler *io_write,
 208                         AioPollFn *io_poll,
 209                         void *opaque)
 210 {
 211     AioHandler *node;
 212     bool is_new = false;
 213     bool deleted = false;
 214     int poll_disable_change;
 215
 216     qemu_lockcnt_lock(&ctx->list_lock);
 217
 218     node = find_aio_handler(ctx, fd);
 219
 220     /* Are we deleting the fd handler? */
 221     if (!io_read && !io_write && !io_poll) {
 222         if (node == NULL) {
 223             qemu_lockcnt_unlock(&ctx->list_lock);
 224             return;
 225         }
 226
 227         /* If the GSource is in the process of being destroyed then
 228          * g_source_remove_poll() causes an assertion failure.  Skip
 229          * removal in that case, because glib cleans up its state during
 230          * destruction anyway.
 231          */
 232         if (!g_source_is_destroyed(&ctx->source)) {
 233             g_source_remove_poll(&ctx->source, &node->pfd);
 234         }
 235
 236         /* If a read is in progress, just mark the node as deleted */
 237         if (qemu_lockcnt_count(&ctx->list_lock)) {
 238             node->deleted = 1;
 239             node->pfd.revents = 0;
 240         } else {
 241             /* Otherwise, delete it for real.  We can't just mark it as
 242              * deleted because deleted nodes are only cleaned up while
 243              * no one is walking the handlers list.
 244              */
 245             QLIST_REMOVE(node, node);
 246             deleted = true;
 247         }
 248         poll_disable_change = -!node->io_poll;
 249     } else {
 250         poll_disable_change = !io_poll - (node && !node->io_poll);
 251         if (node == NULL) {
 252             /* Alloc and insert if it's not already there */
 253             node = g_new0(AioHandler, 1);
 254             node->pfd.fd = fd;
 255             QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node);
 256
 257             g_source_add_poll(&ctx->source, &node->pfd);
 258             is_new = true;
 259         }
 260
 261         /* Update handler with latest information */
 262         node->io_read = io_read;
 263         node->io_write = io_write;
 264         node->io_poll = io_poll;
 265         node->opaque = opaque;
 266         node->is_external = is_external;
 267
 268         node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
 269         node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
 270     }
 271
 272     /* No need to order poll_disable_cnt writes against other updates;
 273      * the counter is only used to avoid wasting time and latency on
 274      * iterated polling when the system call will be ultimately necessary.
 275      * Changing handlers is a rare event, and a little wasted polling until
 276      * the aio_notify below is not an issue.
 277      */
 278     atomic_set(&ctx->poll_disable_cnt,
 279                atomic_read(&ctx->poll_disable_cnt) + poll_disable_change);
 280
 281     aio_epoll_update(ctx, node, is_new);
 282     qemu_lockcnt_unlock(&ctx->list_lock);
 283     aio_notify(ctx);
 284
 285     if (deleted) {
 286         g_free(node);
 287     }
 288 }
 289
 290 void aio_set_fd_poll(AioContext *ctx, int fd,
 291                      IOHandler *io_poll_begin,
 292                      IOHandler *io_poll_end)
 293 {
 294     AioHandler *node = find_aio_handler(ctx, fd);
 295
 296     if (!node) {
 297         return;
 298     }
 299
 300     node->io_poll_begin = io_poll_begin;
 301     node->io_poll_end = io_poll_end;
 302 }
 303
 304 void aio_set_event_notifier(AioContext *ctx,
 305                             EventNotifier *notifier,
 306                             bool is_external,
 307                             EventNotifierHandler *io_read,
 308                             AioPollFn *io_poll)
 309 {
 310     aio_set_fd_handler(ctx, event_notifier_get_fd(notifier), is_external,
 311                        (IOHandler *)io_read, NULL, io_poll, notifier);
 312 }
 313
 314 void aio_set_event_notifier_poll(AioContext *ctx,
 315                                  EventNotifier *notifier,
 316                                  EventNotifierHandler *io_poll_begin,
 317                                  EventNotifierHandler *io_poll_end)
 318 {
 319     aio_set_fd_poll(ctx, event_notifier_get_fd(notifier),
 320                     (IOHandler *)io_poll_begin,
 321                     (IOHandler *)io_poll_end);
 322 }
 323
 324 static void poll_set_started(AioContext *ctx, bool started)
 325 {
 326     AioHandler *node;
 327
 328     if (started == ctx->poll_started) {
 329         return;
 330     }
 331
 332     ctx->poll_started = started;
 333
 334     qemu_lockcnt_inc(&ctx->list_lock);
 335     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
 336         IOHandler *fn;
 337
 338         if (node->deleted) {
 339             continue;
 340         }
 341
 342         if (started) {
 343             fn = node->io_poll_begin;
 344         } else {
 345             fn = node->io_poll_end;
 346         }
 347
 348         if (fn) {
 349             fn(node->opaque);
 350         }
 351     }
 352     qemu_lockcnt_dec(&ctx->list_lock);
 353 }
 354
 355
 356 bool aio_prepare(AioContext *ctx)
 357 {
 358     /* Poll mode cannot be used with glib's event loop, disable it. */
 359     poll_set_started(ctx, false);
 360
 361     return false;
 362 }
 363
 364 bool aio_pending(AioContext *ctx)
 365 {
 366     AioHandler *node;
 367     bool result = false;
 368
 369     /*
 370      * We have to walk very carefully in case aio_set_fd_handler is
 371      * called while we're walking.
 372      */
 373     qemu_lockcnt_inc(&ctx->list_lock);
 374
 375     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
 376         int revents;
 377
 378         revents = node->pfd.revents & node->pfd.events;
 379         if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read &&
 380             aio_node_check(ctx, node->is_external)) {
 381             result = true;
 382             break;
 383         }
 384         if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write &&
 385             aio_node_check(ctx, node->is_external)) {
 386             result = true;
 387             break;
 388         }
 389     }
 390     qemu_lockcnt_dec(&ctx->list_lock);
 391
 392     return result;
 393 }
 394
 395 static bool aio_dispatch_handlers(AioContext *ctx)
 396 {
 397     AioHandler *node, *tmp;
 398     bool progress = false;
 399
 400     QLIST_FOREACH_SAFE_RCU(node, &ctx->aio_handlers, node, tmp) {
 401         int revents;
 402
 403         revents = node->pfd.revents & node->pfd.events;
 404         node->pfd.revents = 0;
 405
 406         if (!node->deleted &&
 407             (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
 408             aio_node_check(ctx, node->is_external) &&
 409             node->io_read) {
 410             node->io_read(node->opaque);
 411
 412             /* aio_notify() does not count as progress */
 413             if (node->opaque != &ctx->notifier) {
 414                 progress = true;
 415             }
 416         }
 417         if (!node->deleted &&
 418             (revents & (G_IO_OUT | G_IO_ERR)) &&
 419             aio_node_check(ctx, node->is_external) &&
 420             node->io_write) {
 421             node->io_write(node->opaque);
 422             progress = true;
 423         }
 424
 425         if (node->deleted) {
 426             if (qemu_lockcnt_dec_if_lock(&ctx->list_lock)) {
 427                 QLIST_REMOVE(node, node);
 428                 g_free(node);
 429                 qemu_lockcnt_inc_and_unlock(&ctx->list_lock);
 430             }
 431         }
 432     }
 433
 434     return progress;
 435 }
 436
 437 void aio_dispatch(AioContext *ctx)
 438 {
 439     qemu_lockcnt_inc(&ctx->list_lock);
 440     aio_bh_poll(ctx);
 441     aio_dispatch_handlers(ctx);
 442     qemu_lockcnt_dec(&ctx->list_lock);
 443
 444     timerlistgroup_run_timers(&ctx->tlg);
 445 }
 446
 447 /* These thread-local variables are used only in a small part of aio_poll
 448  * around the call to the poll() system call.  In particular they are not
 449  * used while aio_poll is performing callbacks, which makes it much easier
 450  * to think about reentrancy!
 451  *
 452  * Stack-allocated arrays would be perfect but they have size limitations;
 453  * heap allocation is expensive enough that we want to reuse arrays across
 454  * calls to aio_poll().  And because poll() has to be called without holding
 455  * any lock, the arrays cannot be stored in AioContext.  Thread-local data
 456  * has none of the disadvantages of these three options.
 457  */
 458 static __thread GPollFD *pollfds;
 459 static __thread AioHandler **nodes;
 460 static __thread unsigned npfd, nalloc;
 461 static __thread Notifier pollfds_cleanup_notifier;
 462
 463 static void pollfds_cleanup(Notifier *n, void *unused)
 464 {
 465     g_assert(npfd == 0);
 466     g_free(pollfds);
 467     g_free(nodes);
 468     nalloc = 0;
 469 }
 470
 471 static void add_pollfd(AioHandler *node)
 472 {
 473     if (npfd == nalloc) {
 474         if (nalloc == 0) {
 475             pollfds_cleanup_notifier.notify = pollfds_cleanup;
 476             qemu_thread_atexit_add(&pollfds_cleanup_notifier);
 477             nalloc = 8;
 478         } else {
 479             g_assert(nalloc <= INT_MAX);
 480             nalloc *= 2;
 481         }
 482         pollfds = g_renew(GPollFD, pollfds, nalloc);
 483         nodes = g_renew(AioHandler *, nodes, nalloc);
 484     }
 485     nodes[npfd] = node;
 486     pollfds[npfd] = (GPollFD) {
 487         .fd = node->pfd.fd,
 488         .events = node->pfd.events,
 489     };
 490     npfd++;
 491 }
 492
 493 static bool run_poll_handlers_once(AioContext *ctx, int64_t *timeout)
 494 {
 495     bool progress = false;
 496     AioHandler *node;
 497
 498     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
 499         if (!node->deleted && node->io_poll &&
 500             aio_node_check(ctx, node->is_external) &&
 501             node->io_poll(node->opaque)) {
 502             *timeout = 0;
 503             if (node->opaque != &ctx->notifier) {
 504                 progress = true;
 505             }
 506         }
 507
 508         /* Caller handles freeing deleted nodes.  Don't do it here. */
 509     }
 510
 511     return progress;
 512 }
 513
 514 /* run_poll_handlers:
 515  * @ctx: the AioContext
 516  * @max_ns: maximum time to poll for, in nanoseconds
 517  *
 518  * Polls for a given time.
 519  *
 520  * Note that ctx->notify_me must be non-zero so this function can detect
 521  * aio_notify().
 522  *
 523  * Note that the caller must have incremented ctx->list_lock.
 524  *
 525  * Returns: true if progress was made, false otherwise
 526  */
 527 static bool run_poll_handlers(AioContext *ctx, int64_t max_ns, int64_t *timeout)
 528 {
 529     bool progress;
 530     int64_t start_time, elapsed_time;
 531
 532     assert(ctx->notify_me);
 533     assert(qemu_lockcnt_count(&ctx->list_lock) > 0);
 534
 535     trace_run_poll_handlers_begin(ctx, max_ns, *timeout);
 536
 537     start_time = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 538     do {
 539         progress = run_poll_handlers_once(ctx, timeout);
 540         elapsed_time = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - start_time;
 541     } while (!progress && elapsed_time < max_ns
 542              && !atomic_read(&ctx->poll_disable_cnt));
 543
 544     /* If time has passed with no successful polling, adjust *timeout to
 545      * keep the same ending time.
 546      */
 547     if (*timeout != -1) {
 548         *timeout -= MIN(*timeout, elapsed_time);
 549     }
 550
 551     trace_run_poll_handlers_end(ctx, progress, *timeout);
 552     return progress;
 553 }
 554
 555 /* try_poll_mode:
 556  * @ctx: the AioContext
 557  * @timeout: timeout for blocking wait, computed by the caller and updated if
 558  *    polling succeeds.
 559  *
 560  * ctx->notify_me must be non-zero so this function can detect aio_notify().
 561  *
 562  * Note that the caller must have incremented ctx->list_lock.
 563  *
 564  * Returns: true if progress was made, false otherwise
 565  */
 566 static bool try_poll_mode(AioContext *ctx, int64_t *timeout)
 567 {
 568     /* See qemu_soonest_timeout() uint64_t hack */
 569     int64_t max_ns = MIN((uint64_t)*timeout, (uint64_t)ctx->poll_ns);
 570
 571     if (max_ns && !atomic_read(&ctx->poll_disable_cnt)) {
 572         poll_set_started(ctx, true);
 573
 574         if (run_poll_handlers(ctx, max_ns, timeout)) {
 575             return true;
 576         }
 577     }
 578
 579     poll_set_started(ctx, false);
 580
 581     /* Even if we don't run busy polling, try polling once in case it can make
 582      * progress and the caller will be able to avoid ppoll(2)/epoll_wait(2).
 583      */
 584     return run_poll_handlers_once(ctx, timeout);
 585 }
 586
 587 bool aio_poll(AioContext *ctx, bool blocking)
 588 {
 589     AioHandler *node;
 590     int i;
 591     int ret = 0;
 592     bool progress;
 593     int64_t timeout;
 594     int64_t start = 0;
 595
 596     /* aio_notify can avoid the expensive event_notifier_set if
 597      * everything (file descriptors, bottom halves, timers) will
 598      * be re-evaluated before the next blocking poll().  This is
 599      * already true when aio_poll is called with blocking == false;
 600      * if blocking == true, it is only true after poll() returns,
 601      * so disable the optimization now.
 602      */
 603     if (blocking) {
 604         assert(in_aio_context_home_thread(ctx));
 605         atomic_add(&ctx->notify_me, 2);
 606     }
 607
 608     qemu_lockcnt_inc(&ctx->list_lock);
 609
 610     if (ctx->poll_max_ns) {
 611         start = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 612     }
 613
 614     timeout = blocking ? aio_compute_timeout(ctx) : 0;
 615     progress = try_poll_mode(ctx, &timeout);
 616     assert(!(timeout && progress));
 617
 618     /* If polling is allowed, non-blocking aio_poll does not need the
 619      * system call---a single round of run_poll_handlers_once suffices.
 620      */
 621     if (timeout || atomic_read(&ctx->poll_disable_cnt)) {
 622         assert(npfd == 0);
 623
 624         /* fill pollfds */
 625
 626         if (!aio_epoll_enabled(ctx)) {
 627             QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
 628                 if (!node->deleted && node->pfd.events
 629                     && aio_node_check(ctx, node->is_external)) {
 630                     add_pollfd(node);
 631                 }
 632             }
 633         }
 634
 635         /* wait until next event */
 636         if (aio_epoll_check_poll(ctx, pollfds, npfd, timeout)) {
 637             AioHandler epoll_handler;
 638
 639             epoll_handler.pfd.fd = ctx->epollfd;
 640             epoll_handler.pfd.events = G_IO_IN | G_IO_OUT | G_IO_HUP | G_IO_ERR;
 641             npfd = 0;
 642             add_pollfd(&epoll_handler);
 643             ret = aio_epoll(ctx, pollfds, npfd, timeout);
 644         } else  {
 645             ret = qemu_poll_ns(pollfds, npfd, timeout);
 646         }
 647     }
 648
 649     if (blocking) {
 650         atomic_sub(&ctx->notify_me, 2);
 651         aio_notify_accept(ctx);
 652     }
 653
 654     /* Adjust polling time */
 655     if (ctx->poll_max_ns) {
 656         int64_t block_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - start;
 657
 658         if (block_ns <= ctx->poll_ns) {
 659             /* This is the sweet spot, no adjustment needed */
 660         } else if (block_ns > ctx->poll_max_ns) {
 661             /* We'd have to poll for too long, poll less */
 662             int64_t old = ctx->poll_ns;
 663
 664             if (ctx->poll_shrink) {
 665                 ctx->poll_ns /= ctx->poll_shrink;
 666             } else {
 667                 ctx->poll_ns = 0;
 668             }
 669
 670             trace_poll_shrink(ctx, old, ctx->poll_ns);
 671         } else if (ctx->poll_ns < ctx->poll_max_ns &&
 672                    block_ns < ctx->poll_max_ns) {
 673             /* There is room to grow, poll longer */
 674             int64_t old = ctx->poll_ns;
 675             int64_t grow = ctx->poll_grow;
 676
 677             if (grow == 0) {
 678                 grow = 2;
 679             }
 680
 681             if (ctx->poll_ns) {
 682                 ctx->poll_ns *= grow;
 683             } else {
 684                 ctx->poll_ns = 4000; /* start polling at 4 microseconds */
 685             }
 686
 687             if (ctx->poll_ns > ctx->poll_max_ns) {
 688                 ctx->poll_ns = ctx->poll_max_ns;
 689             }
 690
 691             trace_poll_grow(ctx, old, ctx->poll_ns);
 692         }
 693     }
 694
 695     /* if we have any readable fds, dispatch event */
 696     if (ret > 0) {
 697         for (i = 0; i < npfd; i++) {
 698             nodes[i]->pfd.revents = pollfds[i].revents;
 699         }
 700     }
 701
 702     npfd = 0;
 703
 704     progress |= aio_bh_poll(ctx);
 705
 706     if (ret > 0) {
 707         progress |= aio_dispatch_handlers(ctx);
 708     }
 709
 710     qemu_lockcnt_dec(&ctx->list_lock);
 711
 712     progress |= timerlistgroup_run_timers(&ctx->tlg);
 713
 714     return progress;
 715 }
 716
 717 void aio_context_setup(AioContext *ctx)
 718 {
 719 #ifdef CONFIG_EPOLL_CREATE1
 720     assert(!ctx->epollfd);
 721     ctx->epollfd = epoll_create1(EPOLL_CLOEXEC);
 722     if (ctx->epollfd == -1) {
 723         fprintf(stderr, "Failed to create epoll instance: %s", strerror(errno));
 724         ctx->epoll_available = false;
 725     } else {
 726         ctx->epoll_available = true;
 727     }
 728 #endif
 729 }
 730
 731 void aio_context_destroy(AioContext *ctx)
 732 {
 733 #ifdef CONFIG_EPOLL_CREATE1
 734     aio_epoll_disable(ctx);
 735 #endif
 736 }
 737
 738 void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns,
 739                                  int64_t grow, int64_t shrink, Error **errp)
 740 {
 741     /* No thread synchronization here, it doesn't matter if an incorrect value
 742      * is used once.
 743      */
 744     ctx->poll_max_ns = max_ns;
 745     ctx->poll_ns = 0;
 746     ctx->poll_grow = grow;
 747     ctx->poll_shrink = shrink;
 748
 749     aio_notify(ctx);
 750 }