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 static bool aio_remove_fd_handler(AioContext *ctx, AioHandler *node)
 204 {
 205     /* If the GSource is in the process of being destroyed then
 206      * g_source_remove_poll() causes an assertion failure.  Skip
 207      * removal in that case, because glib cleans up its state during
 208      * destruction anyway.
 209      */
 210     if (!g_source_is_destroyed(&ctx->source)) {
 211         g_source_remove_poll(&ctx->source, &node->pfd);
 212     }
 213
 214     /* If a read is in progress, just mark the node as deleted */
 215     if (qemu_lockcnt_count(&ctx->list_lock)) {
 216         node->deleted = 1;
 217         node->pfd.revents = 0;
 218         return false;
 219     }
 220     /* Otherwise, delete it for real.  We can't just mark it as
 221      * deleted because deleted nodes are only cleaned up while
 222      * no one is walking the handlers list.
 223      */
 224     QLIST_REMOVE(node, node);
 225     return true;
 226 }
 227
 228 void aio_set_fd_handler(AioContext *ctx,
 229                         int fd,
 230                         bool is_external,
 231                         IOHandler *io_read,
 232                         IOHandler *io_write,
 233                         AioPollFn *io_poll,
 234                         void *opaque)
 235 {
 236     AioHandler *node;
 237     AioHandler *new_node = NULL;
 238     bool is_new = false;
 239     bool deleted = false;
 240     int poll_disable_change;
 241
 242     qemu_lockcnt_lock(&ctx->list_lock);
 243
 244     node = find_aio_handler(ctx, fd);
 245
 246     /* Are we deleting the fd handler? */
 247     if (!io_read && !io_write && !io_poll) {
 248         if (node == NULL) {
 249             qemu_lockcnt_unlock(&ctx->list_lock);
 250             return;
 251         }
 252         /* Clean events in order to unregister fd from the ctx epoll. */
 253         node->pfd.events = 0;
 254
 255         poll_disable_change = -!node->io_poll;
 256     } else {
 257         poll_disable_change = !io_poll - (node && !node->io_poll);
 258         if (node == NULL) {
 259             is_new = true;
 260         }
 261         /* Alloc and insert if it's not already there */
 262         new_node = g_new0(AioHandler, 1);
 263
 264         /* Update handler with latest information */
 265         new_node->io_read = io_read;
 266         new_node->io_write = io_write;
 267         new_node->io_poll = io_poll;
 268         new_node->opaque = opaque;
 269         new_node->is_external = is_external;
 270
 271         if (is_new) {
 272             new_node->pfd.fd = fd;
 273         } else {
 274             new_node->pfd = node->pfd;
 275         }
 276         g_source_add_poll(&ctx->source, &new_node->pfd);
 277
 278         new_node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
 279         new_node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
 280
 281         QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, new_node, node);
 282     }
 283     if (node) {
 284         deleted = aio_remove_fd_handler(ctx, node);
 285     }
 286
 287     /* No need to order poll_disable_cnt writes against other updates;
 288      * the counter is only used to avoid wasting time and latency on
 289      * iterated polling when the system call will be ultimately necessary.
 290      * Changing handlers is a rare event, and a little wasted polling until
 291      * the aio_notify below is not an issue.
 292      */
 293     atomic_set(&ctx->poll_disable_cnt,
 294                atomic_read(&ctx->poll_disable_cnt) + poll_disable_change);
 295
 296     if (new_node) {
 297         aio_epoll_update(ctx, new_node, is_new);
 298     } else if (node) {
 299         /* Unregister deleted fd_handler */
 300         aio_epoll_update(ctx, node, false);
 301     }
 302     qemu_lockcnt_unlock(&ctx->list_lock);
 303     aio_notify(ctx);
 304
 305     if (deleted) {
 306         g_free(node);
 307     }
 308 }
 309
 310 void aio_set_fd_poll(AioContext *ctx, int fd,
 311                      IOHandler *io_poll_begin,
 312                      IOHandler *io_poll_end)
 313 {
 314     AioHandler *node = find_aio_handler(ctx, fd);
 315
 316     if (!node) {
 317         return;
 318     }
 319
 320     node->io_poll_begin = io_poll_begin;
 321     node->io_poll_end = io_poll_end;
 322 }
 323
 324 void aio_set_event_notifier(AioContext *ctx,
 325                             EventNotifier *notifier,
 326                             bool is_external,
 327                             EventNotifierHandler *io_read,
 328                             AioPollFn *io_poll)
 329 {
 330     aio_set_fd_handler(ctx, event_notifier_get_fd(notifier), is_external,
 331                        (IOHandler *)io_read, NULL, io_poll, notifier);
 332 }
 333
 334 void aio_set_event_notifier_poll(AioContext *ctx,
 335                                  EventNotifier *notifier,
 336                                  EventNotifierHandler *io_poll_begin,
 337                                  EventNotifierHandler *io_poll_end)
 338 {
 339     aio_set_fd_poll(ctx, event_notifier_get_fd(notifier),
 340                     (IOHandler *)io_poll_begin,
 341                     (IOHandler *)io_poll_end);
 342 }
 343
 344 static void poll_set_started(AioContext *ctx, bool started)
 345 {
 346     AioHandler *node;
 347
 348     if (started == ctx->poll_started) {
 349         return;
 350     }
 351
 352     ctx->poll_started = started;
 353
 354     qemu_lockcnt_inc(&ctx->list_lock);
 355     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
 356         IOHandler *fn;
 357
 358         if (node->deleted) {
 359             continue;
 360         }
 361
 362         if (started) {
 363             fn = node->io_poll_begin;
 364         } else {
 365             fn = node->io_poll_end;
 366         }
 367
 368         if (fn) {
 369             fn(node->opaque);
 370         }
 371     }
 372     qemu_lockcnt_dec(&ctx->list_lock);
 373 }
 374
 375
 376 bool aio_prepare(AioContext *ctx)
 377 {
 378     /* Poll mode cannot be used with glib's event loop, disable it. */
 379     poll_set_started(ctx, false);
 380
 381     return false;
 382 }
 383
 384 bool aio_pending(AioContext *ctx)
 385 {
 386     AioHandler *node;
 387     bool result = false;
 388
 389     /*
 390      * We have to walk very carefully in case aio_set_fd_handler is
 391      * called while we're walking.
 392      */
 393     qemu_lockcnt_inc(&ctx->list_lock);
 394
 395     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
 396         int revents;
 397
 398         revents = node->pfd.revents & node->pfd.events;
 399         if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read &&
 400             aio_node_check(ctx, node->is_external)) {
 401             result = true;
 402             break;
 403         }
 404         if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write &&
 405             aio_node_check(ctx, node->is_external)) {
 406             result = true;
 407             break;
 408         }
 409     }
 410     qemu_lockcnt_dec(&ctx->list_lock);
 411
 412     return result;
 413 }
 414
 415 static bool aio_dispatch_handlers(AioContext *ctx)
 416 {
 417     AioHandler *node, *tmp;
 418     bool progress = false;
 419
 420     QLIST_FOREACH_SAFE_RCU(node, &ctx->aio_handlers, node, tmp) {
 421         int revents;
 422
 423         revents = node->pfd.revents & node->pfd.events;
 424         node->pfd.revents = 0;
 425
 426         if (!node->deleted &&
 427             (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
 428             aio_node_check(ctx, node->is_external) &&
 429             node->io_read) {
 430             node->io_read(node->opaque);
 431
 432             /* aio_notify() does not count as progress */
 433             if (node->opaque != &ctx->notifier) {
 434                 progress = true;
 435             }
 436         }
 437         if (!node->deleted &&
 438             (revents & (G_IO_OUT | G_IO_ERR)) &&
 439             aio_node_check(ctx, node->is_external) &&
 440             node->io_write) {
 441             node->io_write(node->opaque);
 442             progress = true;
 443         }
 444
 445         if (node->deleted) {
 446             if (qemu_lockcnt_dec_if_lock(&ctx->list_lock)) {
 447                 QLIST_REMOVE(node, node);
 448                 g_free(node);
 449                 qemu_lockcnt_inc_and_unlock(&ctx->list_lock);
 450             }
 451         }
 452     }
 453
 454     return progress;
 455 }
 456
 457 void aio_dispatch(AioContext *ctx)
 458 {
 459     qemu_lockcnt_inc(&ctx->list_lock);
 460     aio_bh_poll(ctx);
 461     aio_dispatch_handlers(ctx);
 462     qemu_lockcnt_dec(&ctx->list_lock);
 463
 464     timerlistgroup_run_timers(&ctx->tlg);
 465 }
 466
 467 /* These thread-local variables are used only in a small part of aio_poll
 468  * around the call to the poll() system call.  In particular they are not
 469  * used while aio_poll is performing callbacks, which makes it much easier
 470  * to think about reentrancy!
 471  *
 472  * Stack-allocated arrays would be perfect but they have size limitations;
 473  * heap allocation is expensive enough that we want to reuse arrays across
 474  * calls to aio_poll().  And because poll() has to be called without holding
 475  * any lock, the arrays cannot be stored in AioContext.  Thread-local data
 476  * has none of the disadvantages of these three options.
 477  */
 478 static __thread GPollFD *pollfds;
 479 static __thread AioHandler **nodes;
 480 static __thread unsigned npfd, nalloc;
 481 static __thread Notifier pollfds_cleanup_notifier;
 482
 483 static void pollfds_cleanup(Notifier *n, void *unused)
 484 {
 485     g_assert(npfd == 0);
 486     g_free(pollfds);
 487     g_free(nodes);
 488     nalloc = 0;
 489 }
 490
 491 static void add_pollfd(AioHandler *node)
 492 {
 493     if (npfd == nalloc) {
 494         if (nalloc == 0) {
 495             pollfds_cleanup_notifier.notify = pollfds_cleanup;
 496             qemu_thread_atexit_add(&pollfds_cleanup_notifier);
 497             nalloc = 8;
 498         } else {
 499             g_assert(nalloc <= INT_MAX);
 500             nalloc *= 2;
 501         }
 502         pollfds = g_renew(GPollFD, pollfds, nalloc);
 503         nodes = g_renew(AioHandler *, nodes, nalloc);
 504     }
 505     nodes[npfd] = node;
 506     pollfds[npfd] = (GPollFD) {
 507         .fd = node->pfd.fd,
 508         .events = node->pfd.events,
 509     };
 510     npfd++;
 511 }
 512
 513 static bool run_poll_handlers_once(AioContext *ctx, int64_t *timeout)
 514 {
 515     bool progress = false;
 516     AioHandler *node;
 517
 518     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
 519         if (!node->deleted && node->io_poll &&
 520             aio_node_check(ctx, node->is_external) &&
 521             node->io_poll(node->opaque)) {
 522             *timeout = 0;
 523             if (node->opaque != &ctx->notifier) {
 524                 progress = true;
 525             }
 526         }
 527
 528         /* Caller handles freeing deleted nodes.  Don't do it here. */
 529     }
 530
 531     return progress;
 532 }
 533
 534 /* run_poll_handlers:
 535  * @ctx: the AioContext
 536  * @max_ns: maximum time to poll for, in nanoseconds
 537  *
 538  * Polls for a given time.
 539  *
 540  * Note that ctx->notify_me must be non-zero so this function can detect
 541  * aio_notify().
 542  *
 543  * Note that the caller must have incremented ctx->list_lock.
 544  *
 545  * Returns: true if progress was made, false otherwise
 546  */
 547 static bool run_poll_handlers(AioContext *ctx, int64_t max_ns, int64_t *timeout)
 548 {
 549     bool progress;
 550     int64_t start_time, elapsed_time;
 551
 552     assert(ctx->notify_me);
 553     assert(qemu_lockcnt_count(&ctx->list_lock) > 0);
 554
 555     trace_run_poll_handlers_begin(ctx, max_ns, *timeout);
 556
 557     start_time = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 558     do {
 559         progress = run_poll_handlers_once(ctx, timeout);
 560         elapsed_time = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - start_time;
 561     } while (!progress && elapsed_time < max_ns
 562              && !atomic_read(&ctx->poll_disable_cnt));
 563
 564     /* If time has passed with no successful polling, adjust *timeout to
 565      * keep the same ending time.
 566      */
 567     if (*timeout != -1) {
 568         *timeout -= MIN(*timeout, elapsed_time);
 569     }
 570
 571     trace_run_poll_handlers_end(ctx, progress, *timeout);
 572     return progress;
 573 }
 574
 575 /* try_poll_mode:
 576  * @ctx: the AioContext
 577  * @timeout: timeout for blocking wait, computed by the caller and updated if
 578  *    polling succeeds.
 579  *
 580  * ctx->notify_me must be non-zero so this function can detect aio_notify().
 581  *
 582  * Note that the caller must have incremented ctx->list_lock.
 583  *
 584  * Returns: true if progress was made, false otherwise
 585  */
 586 static bool try_poll_mode(AioContext *ctx, int64_t *timeout)
 587 {
 588     /* See qemu_soonest_timeout() uint64_t hack */
 589     int64_t max_ns = MIN((uint64_t)*timeout, (uint64_t)ctx->poll_ns);
 590
 591     if (max_ns && !atomic_read(&ctx->poll_disable_cnt)) {
 592         poll_set_started(ctx, true);
 593
 594         if (run_poll_handlers(ctx, max_ns, timeout)) {
 595             return true;
 596         }
 597     }
 598
 599     poll_set_started(ctx, false);
 600
 601     /* Even if we don't run busy polling, try polling once in case it can make
 602      * progress and the caller will be able to avoid ppoll(2)/epoll_wait(2).
 603      */
 604     return run_poll_handlers_once(ctx, timeout);
 605 }
 606
 607 bool aio_poll(AioContext *ctx, bool blocking)
 608 {
 609     AioHandler *node;
 610     int i;
 611     int ret = 0;
 612     bool progress;
 613     int64_t timeout;
 614     int64_t start = 0;
 615
 616     assert(in_aio_context_home_thread(ctx));
 617
 618     /* aio_notify can avoid the expensive event_notifier_set if
 619      * everything (file descriptors, bottom halves, timers) will
 620      * be re-evaluated before the next blocking poll().  This is
 621      * already true when aio_poll is called with blocking == false;
 622      * if blocking == true, it is only true after poll() returns,
 623      * so disable the optimization now.
 624      */
 625     if (blocking) {
 626         atomic_add(&ctx->notify_me, 2);
 627     }
 628
 629     qemu_lockcnt_inc(&ctx->list_lock);
 630
 631     if (ctx->poll_max_ns) {
 632         start = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 633     }
 634
 635     timeout = blocking ? aio_compute_timeout(ctx) : 0;
 636     progress = try_poll_mode(ctx, &timeout);
 637     assert(!(timeout && progress));
 638
 639     /* If polling is allowed, non-blocking aio_poll does not need the
 640      * system call---a single round of run_poll_handlers_once suffices.
 641      */
 642     if (timeout || atomic_read(&ctx->poll_disable_cnt)) {
 643         assert(npfd == 0);
 644
 645         /* fill pollfds */
 646
 647         if (!aio_epoll_enabled(ctx)) {
 648             QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
 649                 if (!node->deleted && node->pfd.events
 650                     && aio_node_check(ctx, node->is_external)) {
 651                     add_pollfd(node);
 652                 }
 653             }
 654         }
 655
 656         /* wait until next event */
 657         if (aio_epoll_check_poll(ctx, pollfds, npfd, timeout)) {
 658             AioHandler epoll_handler;
 659
 660             epoll_handler.pfd.fd = ctx->epollfd;
 661             epoll_handler.pfd.events = G_IO_IN | G_IO_OUT | G_IO_HUP | G_IO_ERR;
 662             npfd = 0;
 663             add_pollfd(&epoll_handler);
 664             ret = aio_epoll(ctx, pollfds, npfd, timeout);
 665         } else  {
 666             ret = qemu_poll_ns(pollfds, npfd, timeout);
 667         }
 668     }
 669
 670     if (blocking) {
 671         atomic_sub(&ctx->notify_me, 2);
 672         aio_notify_accept(ctx);
 673     }
 674
 675     /* Adjust polling time */
 676     if (ctx->poll_max_ns) {
 677         int64_t block_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - start;
 678
 679         if (block_ns <= ctx->poll_ns) {
 680             /* This is the sweet spot, no adjustment needed */
 681         } else if (block_ns > ctx->poll_max_ns) {
 682             /* We'd have to poll for too long, poll less */
 683             int64_t old = ctx->poll_ns;
 684
 685             if (ctx->poll_shrink) {
 686                 ctx->poll_ns /= ctx->poll_shrink;
 687             } else {
 688                 ctx->poll_ns = 0;
 689             }
 690
 691             trace_poll_shrink(ctx, old, ctx->poll_ns);
 692         } else if (ctx->poll_ns < ctx->poll_max_ns &&
 693                    block_ns < ctx->poll_max_ns) {
 694             /* There is room to grow, poll longer */
 695             int64_t old = ctx->poll_ns;
 696             int64_t grow = ctx->poll_grow;
 697
 698             if (grow == 0) {
 699                 grow = 2;
 700             }
 701
 702             if (ctx->poll_ns) {
 703                 ctx->poll_ns *= grow;
 704             } else {
 705                 ctx->poll_ns = 4000; /* start polling at 4 microseconds */
 706             }
 707
 708             if (ctx->poll_ns > ctx->poll_max_ns) {
 709                 ctx->poll_ns = ctx->poll_max_ns;
 710             }
 711
 712             trace_poll_grow(ctx, old, ctx->poll_ns);
 713         }
 714     }
 715
 716     /* if we have any readable fds, dispatch event */
 717     if (ret > 0) {
 718         for (i = 0; i < npfd; i++) {
 719             nodes[i]->pfd.revents = pollfds[i].revents;
 720         }
 721     }
 722
 723     npfd = 0;
 724
 725     progress |= aio_bh_poll(ctx);
 726
 727     if (ret > 0) {
 728         progress |= aio_dispatch_handlers(ctx);
 729     }
 730
 731     qemu_lockcnt_dec(&ctx->list_lock);
 732
 733     progress |= timerlistgroup_run_timers(&ctx->tlg);
 734
 735     return progress;
 736 }
 737
 738 void aio_context_setup(AioContext *ctx)
 739 {
 740 #ifdef CONFIG_EPOLL_CREATE1
 741     assert(!ctx->epollfd);
 742     ctx->epollfd = epoll_create1(EPOLL_CLOEXEC);
 743     if (ctx->epollfd == -1) {
 744         fprintf(stderr, "Failed to create epoll instance: %s", strerror(errno));
 745         ctx->epoll_available = false;
 746     } else {
 747         ctx->epoll_available = true;
 748     }
 749 #endif
 750 }
 751
 752 void aio_context_destroy(AioContext *ctx)
 753 {
 754 #ifdef CONFIG_EPOLL_CREATE1
 755     aio_epoll_disable(ctx);
 756 #endif
 757 }
 758
 759 void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns,
 760                                  int64_t grow, int64_t shrink, Error **errp)
 761 {
 762     /* No thread synchronization here, it doesn't matter if an incorrect value
 763      * is used once.
 764      */
 765     ctx->poll_max_ns = max_ns;
 766     ctx->poll_ns = 0;
 767     ctx->poll_grow = grow;
 768     ctx->poll_shrink = shrink;
 769
 770     aio_notify(ctx);
 771 }