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 "block/block.h"
  18 #include "qemu/main-loop.h"
  19 #include "qemu/rcu.h"
  20 #include "qemu/rcu_queue.h"
  21 #include "qemu/sockets.h"
  22 #include "qemu/cutils.h"
  23 #include "trace.h"
  24 #include "aio-posix.h"
  25
  26 /*
  27  * G_IO_IN and G_IO_OUT are not appropriate revents values for polling, since
  28  * the handler may not need to access the file descriptor. For example, the
  29  * handler doesn't need to read from an EventNotifier if it polled a memory
  30  * location and a read syscall would be slow. Define our own unique revents
  31  * value to indicate that polling determined this AioHandler is ready.
  32  */
  33 #define REVENTS_POLL_READY 0
  34
  35 /* Stop userspace polling on a handler if it isn't active for some time */
  36 #define POLL_IDLE_INTERVAL_NS (7 * NANOSECONDS_PER_SECOND)
  37
  38 bool aio_poll_disabled(AioContext *ctx)
  39 {
  40     return qatomic_read(&ctx->poll_disable_cnt);
  41 }
  42
  43 void aio_add_ready_handler(AioHandlerList *ready_list,
  44                            AioHandler *node,
  45                            int revents)
  46 {
  47     QLIST_SAFE_REMOVE(node, node_ready); /* remove from nested parent's list */
  48     node->pfd.revents = revents;
  49     QLIST_INSERT_HEAD(ready_list, node, node_ready);
  50 }
  51
  52 static AioHandler *find_aio_handler(AioContext *ctx, int fd)
  53 {
  54     AioHandler *node;
  55
  56     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
  57         if (node->pfd.fd == fd) {
  58             if (!QLIST_IS_INSERTED(node, node_deleted)) {
  59                 return node;
  60             }
  61         }
  62     }
  63
  64     return NULL;
  65 }
  66
  67 static bool aio_remove_fd_handler(AioContext *ctx, AioHandler *node)
  68 {
  69     /* If the GSource is in the process of being destroyed then
  70      * g_source_remove_poll() causes an assertion failure.  Skip
  71      * removal in that case, because glib cleans up its state during
  72      * destruction anyway.
  73      */
  74     if (!g_source_is_destroyed(&ctx->source)) {
  75         g_source_remove_poll(&ctx->source, &node->pfd);
  76     }
  77
  78     node->pfd.revents = 0;
  79
  80     /* If the fd monitor has already marked it deleted, leave it alone */
  81     if (QLIST_IS_INSERTED(node, node_deleted)) {
  82         return false;
  83     }
  84
  85     /* If a read is in progress, just mark the node as deleted */
  86     if (qemu_lockcnt_count(&ctx->list_lock)) {
  87         QLIST_INSERT_HEAD_RCU(&ctx->deleted_aio_handlers, node, node_deleted);
  88         return false;
  89     }
  90     /* Otherwise, delete it for real.  We can't just mark it as
  91      * deleted because deleted nodes are only cleaned up while
  92      * no one is walking the handlers list.
  93      */
  94     QLIST_SAFE_REMOVE(node, node_poll);
  95     QLIST_REMOVE(node, node);
  96     return true;
  97 }
  98
  99 void aio_set_fd_handler(AioContext *ctx,
 100                         int fd,
 101                         bool is_external,
 102                         IOHandler *io_read,
 103                         IOHandler *io_write,
 104                         AioPollFn *io_poll,
 105                         IOHandler *io_poll_ready,
 106                         void *opaque)
 107 {
 108     AioHandler *node;
 109     AioHandler *new_node = NULL;
 110     bool is_new = false;
 111     bool deleted = false;
 112     int poll_disable_change;
 113
 114     if (io_poll && !io_poll_ready) {
 115         io_poll = NULL; /* polling only makes sense if there is a handler */
 116     }
 117
 118     qemu_lockcnt_lock(&ctx->list_lock);
 119
 120     node = find_aio_handler(ctx, fd);
 121
 122     /* Are we deleting the fd handler? */
 123     if (!io_read && !io_write && !io_poll) {
 124         if (node == NULL) {
 125             qemu_lockcnt_unlock(&ctx->list_lock);
 126             return;
 127         }
 128         /* Clean events in order to unregister fd from the ctx epoll. */
 129         node->pfd.events = 0;
 130
 131         poll_disable_change = -!node->io_poll;
 132     } else {
 133         poll_disable_change = !io_poll - (node && !node->io_poll);
 134         if (node == NULL) {
 135             is_new = true;
 136         }
 137         /* Alloc and insert if it's not already there */
 138         new_node = g_new0(AioHandler, 1);
 139
 140         /* Update handler with latest information */
 141         new_node->io_read = io_read;
 142         new_node->io_write = io_write;
 143         new_node->io_poll = io_poll;
 144         new_node->io_poll_ready = io_poll_ready;
 145         new_node->opaque = opaque;
 146         new_node->is_external = is_external;
 147
 148         if (is_new) {
 149             new_node->pfd.fd = fd;
 150         } else {
 151             new_node->pfd = node->pfd;
 152         }
 153         g_source_add_poll(&ctx->source, &new_node->pfd);
 154
 155         new_node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
 156         new_node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
 157
 158         QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, new_node, node);
 159     }
 160
 161     /* No need to order poll_disable_cnt writes against other updates;
 162      * the counter is only used to avoid wasting time and latency on
 163      * iterated polling when the system call will be ultimately necessary.
 164      * Changing handlers is a rare event, and a little wasted polling until
 165      * the aio_notify below is not an issue.
 166      */
 167     qatomic_set(&ctx->poll_disable_cnt,
 168                qatomic_read(&ctx->poll_disable_cnt) + poll_disable_change);
 169
 170     ctx->fdmon_ops->update(ctx, node, new_node);
 171     if (node) {
 172         deleted = aio_remove_fd_handler(ctx, node);
 173     }
 174     qemu_lockcnt_unlock(&ctx->list_lock);
 175     aio_notify(ctx);
 176
 177     if (deleted) {
 178         g_free(node);
 179     }
 180 }
 181
 182 void aio_set_fd_poll(AioContext *ctx, int fd,
 183                      IOHandler *io_poll_begin,
 184                      IOHandler *io_poll_end)
 185 {
 186     AioHandler *node = find_aio_handler(ctx, fd);
 187
 188     if (!node) {
 189         return;
 190     }
 191
 192     node->io_poll_begin = io_poll_begin;
 193     node->io_poll_end = io_poll_end;
 194 }
 195
 196 void aio_set_event_notifier(AioContext *ctx,
 197                             EventNotifier *notifier,
 198                             bool is_external,
 199                             EventNotifierHandler *io_read,
 200                             AioPollFn *io_poll,
 201                             EventNotifierHandler *io_poll_ready)
 202 {
 203     aio_set_fd_handler(ctx, event_notifier_get_fd(notifier), is_external,
 204                        (IOHandler *)io_read, NULL, io_poll,
 205                        (IOHandler *)io_poll_ready, notifier);
 206 }
 207
 208 void aio_set_event_notifier_poll(AioContext *ctx,
 209                                  EventNotifier *notifier,
 210                                  EventNotifierHandler *io_poll_begin,
 211                                  EventNotifierHandler *io_poll_end)
 212 {
 213     aio_set_fd_poll(ctx, event_notifier_get_fd(notifier),
 214                     (IOHandler *)io_poll_begin,
 215                     (IOHandler *)io_poll_end);
 216 }
 217
 218 static bool poll_set_started(AioContext *ctx, AioHandlerList *ready_list,
 219                              bool started)
 220 {
 221     AioHandler *node;
 222     bool progress = false;
 223
 224     if (started == ctx->poll_started) {
 225         return false;
 226     }
 227
 228     ctx->poll_started = started;
 229
 230     qemu_lockcnt_inc(&ctx->list_lock);
 231     QLIST_FOREACH(node, &ctx->poll_aio_handlers, node_poll) {
 232         IOHandler *fn;
 233
 234         if (QLIST_IS_INSERTED(node, node_deleted)) {
 235             continue;
 236         }
 237
 238         if (started) {
 239             fn = node->io_poll_begin;
 240         } else {
 241             fn = node->io_poll_end;
 242         }
 243
 244         if (fn) {
 245             fn(node->opaque);
 246         }
 247
 248         /* Poll one last time in case ->io_poll_end() raced with the event */
 249         if (!started && node->io_poll(node->opaque)) {
 250             aio_add_ready_handler(ready_list, node, REVENTS_POLL_READY);
 251             progress = true;
 252         }
 253     }
 254     qemu_lockcnt_dec(&ctx->list_lock);
 255
 256     return progress;
 257 }
 258
 259
 260 bool aio_prepare(AioContext *ctx)
 261 {
 262     AioHandlerList ready_list = QLIST_HEAD_INITIALIZER(ready_list);
 263
 264     /* Poll mode cannot be used with glib's event loop, disable it. */
 265     poll_set_started(ctx, &ready_list, false);
 266     /* TODO what to do with this list? */
 267
 268     return false;
 269 }
 270
 271 bool aio_pending(AioContext *ctx)
 272 {
 273     AioHandler *node;
 274     bool result = false;
 275
 276     /*
 277      * We have to walk very carefully in case aio_set_fd_handler is
 278      * called while we're walking.
 279      */
 280     qemu_lockcnt_inc(&ctx->list_lock);
 281
 282     QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) {
 283         int revents;
 284
 285         revents = node->pfd.revents & node->pfd.events;
 286         if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read &&
 287             aio_node_check(ctx, node->is_external)) {
 288             result = true;
 289             break;
 290         }
 291         if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write &&
 292             aio_node_check(ctx, node->is_external)) {
 293             result = true;
 294             break;
 295         }
 296     }
 297     qemu_lockcnt_dec(&ctx->list_lock);
 298
 299     return result;
 300 }
 301
 302 static void aio_free_deleted_handlers(AioContext *ctx)
 303 {
 304     AioHandler *node;
 305
 306     if (QLIST_EMPTY_RCU(&ctx->deleted_aio_handlers)) {
 307         return;
 308     }
 309     if (!qemu_lockcnt_dec_if_lock(&ctx->list_lock)) {
 310         return; /* we are nested, let the parent do the freeing */
 311     }
 312
 313     while ((node = QLIST_FIRST_RCU(&ctx->deleted_aio_handlers))) {
 314         QLIST_REMOVE(node, node);
 315         QLIST_REMOVE(node, node_deleted);
 316         QLIST_SAFE_REMOVE(node, node_poll);
 317         g_free(node);
 318     }
 319
 320     qemu_lockcnt_inc_and_unlock(&ctx->list_lock);
 321 }
 322
 323 static bool aio_dispatch_handler(AioContext *ctx, AioHandler *node)
 324 {
 325     bool progress = false;
 326     int revents;
 327
 328     revents = node->pfd.revents & node->pfd.events;
 329     node->pfd.revents = 0;
 330
 331     /*
 332      * Start polling AioHandlers when they become ready because activity is
 333      * likely to continue.  Note that starvation is theoretically possible when
 334      * fdmon_supports_polling(), but only until the fd fires for the first
 335      * time.
 336      */
 337     if (!QLIST_IS_INSERTED(node, node_deleted) &&
 338         !QLIST_IS_INSERTED(node, node_poll) &&
 339         node->io_poll) {
 340         trace_poll_add(ctx, node, node->pfd.fd, revents);
 341         if (ctx->poll_started && node->io_poll_begin) {
 342             node->io_poll_begin(node->opaque);
 343         }
 344         QLIST_INSERT_HEAD(&ctx->poll_aio_handlers, node, node_poll);
 345     }
 346     if (!QLIST_IS_INSERTED(node, node_deleted) &&
 347         revents == 0 &&
 348         aio_node_check(ctx, node->is_external) &&
 349         node->io_poll_ready) {
 350         node->io_poll_ready(node->opaque);
 351
 352         /*
 353          * Return early since revents was zero. aio_notify() does not count as
 354          * progress.
 355          */
 356         return node->opaque != &ctx->notifier;
 357     }
 358
 359     if (!QLIST_IS_INSERTED(node, node_deleted) &&
 360         (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
 361         aio_node_check(ctx, node->is_external) &&
 362         node->io_read) {
 363         node->io_read(node->opaque);
 364
 365         /* aio_notify() does not count as progress */
 366         if (node->opaque != &ctx->notifier) {
 367             progress = true;
 368         }
 369     }
 370     if (!QLIST_IS_INSERTED(node, node_deleted) &&
 371         (revents & (G_IO_OUT | G_IO_ERR)) &&
 372         aio_node_check(ctx, node->is_external) &&
 373         node->io_write) {
 374         node->io_write(node->opaque);
 375         progress = true;
 376     }
 377
 378     return progress;
 379 }
 380
 381 /*
 382  * If we have a list of ready handlers then this is more efficient than
 383  * scanning all handlers with aio_dispatch_handlers().
 384  */
 385 static bool aio_dispatch_ready_handlers(AioContext *ctx,
 386                                         AioHandlerList *ready_list)
 387 {
 388     bool progress = false;
 389     AioHandler *node;
 390
 391     while ((node = QLIST_FIRST(ready_list))) {
 392         QLIST_REMOVE(node, node_ready);
 393         progress = aio_dispatch_handler(ctx, node) || progress;
 394     }
 395
 396     return progress;
 397 }
 398
 399 /* Slower than aio_dispatch_ready_handlers() but only used via glib */
 400 static bool aio_dispatch_handlers(AioContext *ctx)
 401 {
 402     AioHandler *node, *tmp;
 403     bool progress = false;
 404
 405     QLIST_FOREACH_SAFE_RCU(node, &ctx->aio_handlers, node, tmp) {
 406         progress = aio_dispatch_handler(ctx, node) || progress;
 407     }
 408
 409     return progress;
 410 }
 411
 412 void aio_dispatch(AioContext *ctx)
 413 {
 414     qemu_lockcnt_inc(&ctx->list_lock);
 415     aio_bh_poll(ctx);
 416     aio_dispatch_handlers(ctx);
 417     aio_free_deleted_handlers(ctx);
 418     qemu_lockcnt_dec(&ctx->list_lock);
 419
 420     timerlistgroup_run_timers(&ctx->tlg);
 421 }
 422
 423 static bool run_poll_handlers_once(AioContext *ctx,
 424                                    AioHandlerList *ready_list,
 425                                    int64_t now,
 426                                    int64_t *timeout)
 427 {
 428     bool progress = false;
 429     AioHandler *node;
 430     AioHandler *tmp;
 431
 432     QLIST_FOREACH_SAFE(node, &ctx->poll_aio_handlers, node_poll, tmp) {
 433         if (aio_node_check(ctx, node->is_external) &&
 434             node->io_poll(node->opaque)) {
 435             aio_add_ready_handler(ready_list, node, REVENTS_POLL_READY);
 436
 437             node->poll_idle_timeout = now + POLL_IDLE_INTERVAL_NS;
 438
 439             /*
 440              * Polling was successful, exit try_poll_mode immediately
 441              * to adjust the next polling time.
 442              */
 443             *timeout = 0;
 444             if (node->opaque != &ctx->notifier) {
 445                 progress = true;
 446             }
 447         }
 448
 449         /* Caller handles freeing deleted nodes.  Don't do it here. */
 450     }
 451
 452     return progress;
 453 }
 454
 455 static bool fdmon_supports_polling(AioContext *ctx)
 456 {
 457     return ctx->fdmon_ops->need_wait != aio_poll_disabled;
 458 }
 459
 460 static bool remove_idle_poll_handlers(AioContext *ctx,
 461                                       AioHandlerList *ready_list,
 462                                       int64_t now)
 463 {
 464     AioHandler *node;
 465     AioHandler *tmp;
 466     bool progress = false;
 467
 468     /*
 469      * File descriptor monitoring implementations without userspace polling
 470      * support suffer from starvation when a subset of handlers is polled
 471      * because fds will not be processed in a timely fashion.  Don't remove
 472      * idle poll handlers.
 473      */
 474     if (!fdmon_supports_polling(ctx)) {
 475         return false;
 476     }
 477
 478     QLIST_FOREACH_SAFE(node, &ctx->poll_aio_handlers, node_poll, tmp) {
 479         if (node->poll_idle_timeout == 0LL) {
 480             node->poll_idle_timeout = now + POLL_IDLE_INTERVAL_NS;
 481         } else if (now >= node->poll_idle_timeout) {
 482             trace_poll_remove(ctx, node, node->pfd.fd);
 483             node->poll_idle_timeout = 0LL;
 484             QLIST_SAFE_REMOVE(node, node_poll);
 485             if (ctx->poll_started && node->io_poll_end) {
 486                 node->io_poll_end(node->opaque);
 487
 488                 /*
 489                  * Final poll in case ->io_poll_end() races with an event.
 490                  * Nevermind about re-adding the handler in the rare case where
 491                  * this causes progress.
 492                  */
 493                 if (node->io_poll(node->opaque)) {
 494                     aio_add_ready_handler(ready_list, node,
 495                                           REVENTS_POLL_READY);
 496                     progress = true;
 497                 }
 498             }
 499         }
 500     }
 501
 502     return progress;
 503 }
 504
 505 /* run_poll_handlers:
 506  * @ctx: the AioContext
 507  * @ready_list: the list to place ready handlers on
 508  * @max_ns: maximum time to poll for, in nanoseconds
 509  *
 510  * Polls for a given time.
 511  *
 512  * Note that the caller must have incremented ctx->list_lock.
 513  *
 514  * Returns: true if progress was made, false otherwise
 515  */
 516 static bool run_poll_handlers(AioContext *ctx, AioHandlerList *ready_list,
 517                               int64_t max_ns, int64_t *timeout)
 518 {
 519     bool progress;
 520     int64_t start_time, elapsed_time;
 521
 522     assert(qemu_lockcnt_count(&ctx->list_lock) > 0);
 523
 524     trace_run_poll_handlers_begin(ctx, max_ns, *timeout);
 525
 526     /*
 527      * Optimization: ->io_poll() handlers often contain RCU read critical
 528      * sections and we therefore see many rcu_read_lock() -> rcu_read_unlock()
 529      * -> rcu_read_lock() -> ... sequences with expensive memory
 530      * synchronization primitives.  Make the entire polling loop an RCU
 531      * critical section because nested rcu_read_lock()/rcu_read_unlock() calls
 532      * are cheap.
 533      */
 534     RCU_READ_LOCK_GUARD();
 535
 536     start_time = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 537     do {
 538         progress = run_poll_handlers_once(ctx, ready_list,
 539                                           start_time, timeout);
 540         elapsed_time = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - start_time;
 541         max_ns = qemu_soonest_timeout(*timeout, max_ns);
 542         assert(!(max_ns && progress));
 543     } while (elapsed_time < max_ns && !ctx->fdmon_ops->need_wait(ctx));
 544
 545     if (remove_idle_poll_handlers(ctx, ready_list,
 546                                   start_time + elapsed_time)) {
 547         *timeout = 0;
 548         progress = true;
 549     }
 550
 551     /* If time has passed with no successful polling, adjust *timeout to
 552      * keep the same ending time.
 553      */
 554     if (*timeout != -1) {
 555         *timeout -= MIN(*timeout, elapsed_time);
 556     }
 557
 558     trace_run_poll_handlers_end(ctx, progress, *timeout);
 559     return progress;
 560 }
 561
 562 /* try_poll_mode:
 563  * @ctx: the AioContext
 564  * @ready_list: list to add handlers that need to be run
 565  * @timeout: timeout for blocking wait, computed by the caller and updated if
 566  *    polling succeeds.
 567  *
 568  * Note that the caller must have incremented ctx->list_lock.
 569  *
 570  * Returns: true if progress was made, false otherwise
 571  */
 572 static bool try_poll_mode(AioContext *ctx, AioHandlerList *ready_list,
 573                           int64_t *timeout)
 574 {
 575     int64_t max_ns;
 576
 577     if (QLIST_EMPTY_RCU(&ctx->poll_aio_handlers)) {
 578         return false;
 579     }
 580
 581     max_ns = qemu_soonest_timeout(*timeout, ctx->poll_ns);
 582     if (max_ns && !ctx->fdmon_ops->need_wait(ctx)) {
 583         poll_set_started(ctx, ready_list, true);
 584
 585         if (run_poll_handlers(ctx, ready_list, max_ns, timeout)) {
 586             return true;
 587         }
 588     }
 589
 590     if (poll_set_started(ctx, ready_list, false)) {
 591         *timeout = 0;
 592         return true;
 593     }
 594
 595     return false;
 596 }
 597
 598 bool aio_poll(AioContext *ctx, bool blocking)
 599 {
 600     AioHandlerList ready_list = QLIST_HEAD_INITIALIZER(ready_list);
 601     bool progress;
 602     bool use_notify_me;
 603     int64_t timeout;
 604     int64_t start = 0;
 605
 606     /*
 607      * There cannot be two concurrent aio_poll calls for the same AioContext (or
 608      * an aio_poll concurrent with a GSource prepare/check/dispatch callback).
 609      * We rely on this below to avoid slow locked accesses to ctx->notify_me.
 610      *
 611      * aio_poll() may only be called in the AioContext's thread. iohandler_ctx
 612      * is special in that it runs in the main thread, but that thread's context
 613      * is qemu_aio_context.
 614      */
 615     assert(in_aio_context_home_thread(ctx == iohandler_get_aio_context() ?
 616                                       qemu_get_aio_context() : ctx));
 617
 618     qemu_lockcnt_inc(&ctx->list_lock);
 619
 620     if (ctx->poll_max_ns) {
 621         start = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 622     }
 623
 624     timeout = blocking ? aio_compute_timeout(ctx) : 0;
 625     progress = try_poll_mode(ctx, &ready_list, &timeout);
 626     assert(!(timeout && progress));
 627
 628     /*
 629      * aio_notify can avoid the expensive event_notifier_set if
 630      * everything (file descriptors, bottom halves, timers) will
 631      * be re-evaluated before the next blocking poll().  This is
 632      * already true when aio_poll is called with blocking == false;
 633      * if blocking == true, it is only true after poll() returns,
 634      * so disable the optimization now.
 635      */
 636     use_notify_me = timeout != 0;
 637     if (use_notify_me) {
 638         qatomic_set(&ctx->notify_me, qatomic_read(&ctx->notify_me) + 2);
 639         /*
 640          * Write ctx->notify_me before reading ctx->notified.  Pairs with
 641          * smp_mb in aio_notify().
 642          */
 643         smp_mb();
 644
 645         /* Don't block if aio_notify() was called */
 646         if (qatomic_read(&ctx->notified)) {
 647             timeout = 0;
 648         }
 649     }
 650
 651     /* If polling is allowed, non-blocking aio_poll does not need the
 652      * system call---a single round of run_poll_handlers_once suffices.
 653      */
 654     if (timeout || ctx->fdmon_ops->need_wait(ctx)) {
 655         ctx->fdmon_ops->wait(ctx, &ready_list, timeout);
 656     }
 657
 658     if (use_notify_me) {
 659         /* Finish the poll before clearing the flag.  */
 660         qatomic_store_release(&ctx->notify_me,
 661                              qatomic_read(&ctx->notify_me) - 2);
 662     }
 663
 664     aio_notify_accept(ctx);
 665
 666     /* Adjust polling time */
 667     if (ctx->poll_max_ns) {
 668         int64_t block_ns = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) - start;
 669
 670         if (block_ns <= ctx->poll_ns) {
 671             /* This is the sweet spot, no adjustment needed */
 672         } else if (block_ns > ctx->poll_max_ns) {
 673             /* We'd have to poll for too long, poll less */
 674             int64_t old = ctx->poll_ns;
 675
 676             if (ctx->poll_shrink) {
 677                 ctx->poll_ns /= ctx->poll_shrink;
 678             } else {
 679                 ctx->poll_ns = 0;
 680             }
 681
 682             trace_poll_shrink(ctx, old, ctx->poll_ns);
 683         } else if (ctx->poll_ns < ctx->poll_max_ns &&
 684                    block_ns < ctx->poll_max_ns) {
 685             /* There is room to grow, poll longer */
 686             int64_t old = ctx->poll_ns;
 687             int64_t grow = ctx->poll_grow;
 688
 689             if (grow == 0) {
 690                 grow = 2;
 691             }
 692
 693             if (ctx->poll_ns) {
 694                 ctx->poll_ns *= grow;
 695             } else {
 696                 ctx->poll_ns = 4000; /* start polling at 4 microseconds */
 697             }
 698
 699             if (ctx->poll_ns > ctx->poll_max_ns) {
 700                 ctx->poll_ns = ctx->poll_max_ns;
 701             }
 702
 703             trace_poll_grow(ctx, old, ctx->poll_ns);
 704         }
 705     }
 706
 707     progress |= aio_bh_poll(ctx);
 708     progress |= aio_dispatch_ready_handlers(ctx, &ready_list);
 709
 710     aio_free_deleted_handlers(ctx);
 711
 712     qemu_lockcnt_dec(&ctx->list_lock);
 713
 714     progress |= timerlistgroup_run_timers(&ctx->tlg);
 715
 716     return progress;
 717 }
 718
 719 void aio_context_setup(AioContext *ctx)
 720 {
 721     ctx->fdmon_ops = &fdmon_poll_ops;
 722     ctx->epollfd = -1;
 723
 724     /* Use the fastest fd monitoring implementation if available */
 725     if (fdmon_io_uring_setup(ctx)) {
 726         return;
 727     }
 728
 729     fdmon_epoll_setup(ctx);
 730 }
 731
 732 void aio_context_destroy(AioContext *ctx)
 733 {
 734     fdmon_io_uring_destroy(ctx);
 735     fdmon_epoll_disable(ctx);
 736     aio_free_deleted_handlers(ctx);
 737 }
 738
 739 void aio_context_use_g_source(AioContext *ctx)
 740 {
 741     /*
 742      * Disable io_uring when the glib main loop is used because it doesn't
 743      * support mixed glib/aio_poll() usage. It relies on aio_poll() being
 744      * called regularly so that changes to the monitored file descriptors are
 745      * submitted, otherwise a list of pending fd handlers builds up.
 746      */
 747     fdmon_io_uring_destroy(ctx);
 748     aio_free_deleted_handlers(ctx);
 749 }
 750
 751 void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns,
 752                                  int64_t grow, int64_t shrink, Error **errp)
 753 {
 754     /* No thread synchronization here, it doesn't matter if an incorrect value
 755      * is used once.
 756      */
 757     ctx->poll_max_ns = max_ns;
 758     ctx->poll_ns = 0;
 759     ctx->poll_grow = grow;
 760     ctx->poll_shrink = shrink;
 761
 762     aio_notify(ctx);
 763 }
 764
 765 void aio_context_set_aio_params(AioContext *ctx, int64_t max_batch,
 766                                 Error **errp)
 767 {
 768     /*
 769      * No thread synchronization here, it doesn't matter if an incorrect value
 770      * is used once.
 771      */
 772     ctx->aio_max_batch = max_batch;
 773
 774     aio_notify(ctx);
 775 }