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/queue.h"
  20 #include "qemu/sockets.h"
  21 #ifdef CONFIG_EPOLL_CREATE1
  22 #include <sys/epoll.h>
  23 #endif
  24
  25 struct AioHandler
  26 {
  27     GPollFD pfd;
  28     IOHandler *io_read;
  29     IOHandler *io_write;
  30     int deleted;
  31     void *opaque;
  32     bool is_external;
  33     QLIST_ENTRY(AioHandler) node;
  34 };
  35
  36 #ifdef CONFIG_EPOLL_CREATE1
  37
  38 /* The fd number threashold to switch to epoll */
  39 #define EPOLL_ENABLE_THRESHOLD 64
  40
  41 static void aio_epoll_disable(AioContext *ctx)
  42 {
  43     ctx->epoll_available = false;
  44     if (!ctx->epoll_enabled) {
  45         return;
  46     }
  47     ctx->epoll_enabled = false;
  48     close(ctx->epollfd);
  49 }
  50
  51 static inline int epoll_events_from_pfd(int pfd_events)
  52 {
  53     return (pfd_events & G_IO_IN ? EPOLLIN : 0) |
  54            (pfd_events & G_IO_OUT ? EPOLLOUT : 0) |
  55            (pfd_events & G_IO_HUP ? EPOLLHUP : 0) |
  56            (pfd_events & G_IO_ERR ? EPOLLERR : 0);
  57 }
  58
  59 static bool aio_epoll_try_enable(AioContext *ctx)
  60 {
  61     AioHandler *node;
  62     struct epoll_event event;
  63
  64     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
  65         int r;
  66         if (node->deleted || !node->pfd.events) {
  67             continue;
  68         }
  69         event.events = epoll_events_from_pfd(node->pfd.events);
  70         event.data.ptr = node;
  71         r = epoll_ctl(ctx->epollfd, EPOLL_CTL_ADD, node->pfd.fd, &event);
  72         if (r) {
  73             return false;
  74         }
  75     }
  76     ctx->epoll_enabled = true;
  77     return true;
  78 }
  79
  80 static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
  81 {
  82     struct epoll_event event;
  83     int r;
  84
  85     if (!ctx->epoll_enabled) {
  86         return;
  87     }
  88     if (!node->pfd.events) {
  89         r = epoll_ctl(ctx->epollfd, EPOLL_CTL_DEL, node->pfd.fd, &event);
  90         if (r) {
  91             aio_epoll_disable(ctx);
  92         }
  93     } else {
  94         event.data.ptr = node;
  95         event.events = epoll_events_from_pfd(node->pfd.events);
  96         if (is_new) {
  97             r = epoll_ctl(ctx->epollfd, EPOLL_CTL_ADD, node->pfd.fd, &event);
  98             if (r) {
  99                 aio_epoll_disable(ctx);
 100             }
 101         } else {
 102             r = epoll_ctl(ctx->epollfd, EPOLL_CTL_MOD, node->pfd.fd, &event);
 103             if (r) {
 104                 aio_epoll_disable(ctx);
 105             }
 106         }
 107     }
 108 }
 109
 110 static int aio_epoll(AioContext *ctx, GPollFD *pfds,
 111                      unsigned npfd, int64_t timeout)
 112 {
 113     AioHandler *node;
 114     int i, ret = 0;
 115     struct epoll_event events[128];
 116
 117     assert(npfd == 1);
 118     assert(pfds[0].fd == ctx->epollfd);
 119     if (timeout > 0) {
 120         ret = qemu_poll_ns(pfds, npfd, timeout);
 121     }
 122     if (timeout <= 0 || ret > 0) {
 123         ret = epoll_wait(ctx->epollfd, events,
 124                          sizeof(events) / sizeof(events[0]),
 125                          timeout);
 126         if (ret <= 0) {
 127             goto out;
 128         }
 129         for (i = 0; i < ret; i++) {
 130             int ev = events[i].events;
 131             node = events[i].data.ptr;
 132             node->pfd.revents = (ev & EPOLLIN ? G_IO_IN : 0) |
 133                 (ev & EPOLLOUT ? G_IO_OUT : 0) |
 134                 (ev & EPOLLHUP ? G_IO_HUP : 0) |
 135                 (ev & EPOLLERR ? G_IO_ERR : 0);
 136         }
 137     }
 138 out:
 139     return ret;
 140 }
 141
 142 static bool aio_epoll_enabled(AioContext *ctx)
 143 {
 144     /* Fall back to ppoll when external clients are disabled. */
 145     return !aio_external_disabled(ctx) && ctx->epoll_enabled;
 146 }
 147
 148 static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
 149                                  unsigned npfd, int64_t timeout)
 150 {
 151     if (!ctx->epoll_available) {
 152         return false;
 153     }
 154     if (aio_epoll_enabled(ctx)) {
 155         return true;
 156     }
 157     if (npfd >= EPOLL_ENABLE_THRESHOLD) {
 158         if (aio_epoll_try_enable(ctx)) {
 159             return true;
 160         } else {
 161             aio_epoll_disable(ctx);
 162         }
 163     }
 164     return false;
 165 }
 166
 167 #else
 168
 169 static void aio_epoll_update(AioContext *ctx, AioHandler *node, bool is_new)
 170 {
 171 }
 172
 173 static int aio_epoll(AioContext *ctx, GPollFD *pfds,
 174                      unsigned npfd, int64_t timeout)
 175 {
 176     assert(false);
 177 }
 178
 179 static bool aio_epoll_enabled(AioContext *ctx)
 180 {
 181     return false;
 182 }
 183
 184 static bool aio_epoll_check_poll(AioContext *ctx, GPollFD *pfds,
 185                           unsigned npfd, int64_t timeout)
 186 {
 187     return false;
 188 }
 189
 190 #endif
 191
 192 static AioHandler *find_aio_handler(AioContext *ctx, int fd)
 193 {
 194     AioHandler *node;
 195
 196     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 197         if (node->pfd.fd == fd)
 198             if (!node->deleted)
 199                 return node;
 200     }
 201
 202     return NULL;
 203 }
 204
 205 void aio_set_fd_handler(AioContext *ctx,
 206                         int fd,
 207                         bool is_external,
 208                         IOHandler *io_read,
 209                         IOHandler *io_write,
 210                         void *opaque)
 211 {
 212     AioHandler *node;
 213     bool is_new = false;
 214     bool deleted = false;
 215
 216     node = find_aio_handler(ctx, fd);
 217
 218     /* Are we deleting the fd handler? */
 219     if (!io_read && !io_write) {
 220         if (node) {
 221             g_source_remove_poll(&ctx->source, &node->pfd);
 222
 223             /* If the lock is held, just mark the node as deleted */
 224             if (ctx->walking_handlers) {
 225                 node->deleted = 1;
 226                 node->pfd.revents = 0;
 227             } else {
 228                 /* Otherwise, delete it for real.  We can't just mark it as
 229                  * deleted because deleted nodes are only cleaned up after
 230                  * releasing the walking_handlers lock.
 231                  */
 232                 QLIST_REMOVE(node, node);
 233                 deleted = true;
 234             }
 235         }
 236     } else {
 237         if (node == NULL) {
 238             /* Alloc and insert if it's not already there */
 239             node = g_new0(AioHandler, 1);
 240             node->pfd.fd = fd;
 241             QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node);
 242
 243             g_source_add_poll(&ctx->source, &node->pfd);
 244             is_new = true;
 245         }
 246         /* Update handler with latest information */
 247         node->io_read = io_read;
 248         node->io_write = io_write;
 249         node->opaque = opaque;
 250         node->is_external = is_external;
 251
 252         node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
 253         node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
 254     }
 255
 256     aio_epoll_update(ctx, node, is_new);
 257     aio_notify(ctx);
 258     if (deleted) {
 259         g_free(node);
 260     }
 261 }
 262
 263 void aio_set_event_notifier(AioContext *ctx,
 264                             EventNotifier *notifier,
 265                             bool is_external,
 266                             EventNotifierHandler *io_read)
 267 {
 268     aio_set_fd_handler(ctx, event_notifier_get_fd(notifier),
 269                        is_external, (IOHandler *)io_read, NULL, notifier);
 270 }
 271
 272 bool aio_prepare(AioContext *ctx)
 273 {
 274     return false;
 275 }
 276
 277 bool aio_pending(AioContext *ctx)
 278 {
 279     AioHandler *node;
 280
 281     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 282         int revents;
 283
 284         revents = node->pfd.revents & node->pfd.events;
 285         if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read) {
 286             return true;
 287         }
 288         if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write) {
 289             return true;
 290         }
 291     }
 292
 293     return false;
 294 }
 295
 296 bool aio_dispatch(AioContext *ctx)
 297 {
 298     AioHandler *node;
 299     bool progress = false;
 300
 301     /*
 302      * If there are callbacks left that have been queued, we need to call them.
 303      * Do not call select in this case, because it is possible that the caller
 304      * does not need a complete flush (as is the case for aio_poll loops).
 305      */
 306     if (aio_bh_poll(ctx)) {
 307         progress = true;
 308     }
 309
 310     /*
 311      * We have to walk very carefully in case aio_set_fd_handler is
 312      * called while we're walking.
 313      */
 314     node = QLIST_FIRST(&ctx->aio_handlers);
 315     while (node) {
 316         AioHandler *tmp;
 317         int revents;
 318
 319         ctx->walking_handlers++;
 320
 321         revents = node->pfd.revents & node->pfd.events;
 322         node->pfd.revents = 0;
 323
 324         if (!node->deleted &&
 325             (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
 326             node->io_read) {
 327             node->io_read(node->opaque);
 328
 329             /* aio_notify() does not count as progress */
 330             if (node->opaque != &ctx->notifier) {
 331                 progress = true;
 332             }
 333         }
 334         if (!node->deleted &&
 335             (revents & (G_IO_OUT | G_IO_ERR)) &&
 336             node->io_write) {
 337             node->io_write(node->opaque);
 338             progress = true;
 339         }
 340
 341         tmp = node;
 342         node = QLIST_NEXT(node, node);
 343
 344         ctx->walking_handlers--;
 345
 346         if (!ctx->walking_handlers && tmp->deleted) {
 347             QLIST_REMOVE(tmp, node);
 348             g_free(tmp);
 349         }
 350     }
 351
 352     /* Run our timers */
 353     progress |= timerlistgroup_run_timers(&ctx->tlg);
 354
 355     return progress;
 356 }
 357
 358 /* These thread-local variables are used only in a small part of aio_poll
 359  * around the call to the poll() system call.  In particular they are not
 360  * used while aio_poll is performing callbacks, which makes it much easier
 361  * to think about reentrancy!
 362  *
 363  * Stack-allocated arrays would be perfect but they have size limitations;
 364  * heap allocation is expensive enough that we want to reuse arrays across
 365  * calls to aio_poll().  And because poll() has to be called without holding
 366  * any lock, the arrays cannot be stored in AioContext.  Thread-local data
 367  * has none of the disadvantages of these three options.
 368  */
 369 static __thread GPollFD *pollfds;
 370 static __thread AioHandler **nodes;
 371 static __thread unsigned npfd, nalloc;
 372 static __thread Notifier pollfds_cleanup_notifier;
 373
 374 static void pollfds_cleanup(Notifier *n, void *unused)
 375 {
 376     g_assert(npfd == 0);
 377     g_free(pollfds);
 378     g_free(nodes);
 379     nalloc = 0;
 380 }
 381
 382 static void add_pollfd(AioHandler *node)
 383 {
 384     if (npfd == nalloc) {
 385         if (nalloc == 0) {
 386             pollfds_cleanup_notifier.notify = pollfds_cleanup;
 387             qemu_thread_atexit_add(&pollfds_cleanup_notifier);
 388             nalloc = 8;
 389         } else {
 390             g_assert(nalloc <= INT_MAX);
 391             nalloc *= 2;
 392         }
 393         pollfds = g_renew(GPollFD, pollfds, nalloc);
 394         nodes = g_renew(AioHandler *, nodes, nalloc);
 395     }
 396     nodes[npfd] = node;
 397     pollfds[npfd] = (GPollFD) {
 398         .fd = node->pfd.fd,
 399         .events = node->pfd.events,
 400     };
 401     npfd++;
 402 }
 403
 404 bool aio_poll(AioContext *ctx, bool blocking)
 405 {
 406     AioHandler *node;
 407     int i, ret;
 408     bool progress;
 409     int64_t timeout;
 410
 411     aio_context_acquire(ctx);
 412     progress = false;
 413
 414     /* aio_notify can avoid the expensive event_notifier_set if
 415      * everything (file descriptors, bottom halves, timers) will
 416      * be re-evaluated before the next blocking poll().  This is
 417      * already true when aio_poll is called with blocking == false;
 418      * if blocking == true, it is only true after poll() returns,
 419      * so disable the optimization now.
 420      */
 421     if (blocking) {
 422         atomic_add(&ctx->notify_me, 2);
 423     }
 424
 425     ctx->walking_handlers++;
 426
 427     assert(npfd == 0);
 428
 429     /* fill pollfds */
 430     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 431         if (!node->deleted && node->pfd.events
 432             && !aio_epoll_enabled(ctx)
 433             && aio_node_check(ctx, node->is_external)) {
 434             add_pollfd(node);
 435         }
 436     }
 437
 438     timeout = blocking ? aio_compute_timeout(ctx) : 0;
 439
 440     /* wait until next event */
 441     if (timeout) {
 442         aio_context_release(ctx);
 443     }
 444     if (aio_epoll_check_poll(ctx, pollfds, npfd, timeout)) {
 445         AioHandler epoll_handler;
 446
 447         epoll_handler.pfd.fd = ctx->epollfd;
 448         epoll_handler.pfd.events = G_IO_IN | G_IO_OUT | G_IO_HUP | G_IO_ERR;
 449         npfd = 0;
 450         add_pollfd(&epoll_handler);
 451         ret = aio_epoll(ctx, pollfds, npfd, timeout);
 452     } else  {
 453         ret = qemu_poll_ns(pollfds, npfd, timeout);
 454     }
 455     if (blocking) {
 456         atomic_sub(&ctx->notify_me, 2);
 457     }
 458     if (timeout) {
 459         aio_context_acquire(ctx);
 460     }
 461
 462     aio_notify_accept(ctx);
 463
 464     /* if we have any readable fds, dispatch event */
 465     if (ret > 0) {
 466         for (i = 0; i < npfd; i++) {
 467             nodes[i]->pfd.revents = pollfds[i].revents;
 468         }
 469     }
 470
 471     npfd = 0;
 472     ctx->walking_handlers--;
 473
 474     /* Run dispatch even if there were no readable fds to run timers */
 475     if (aio_dispatch(ctx)) {
 476         progress = true;
 477     }
 478
 479     aio_context_release(ctx);
 480
 481     return progress;
 482 }
 483
 484 void aio_context_setup(AioContext *ctx, Error **errp)
 485 {
 486 #ifdef CONFIG_EPOLL_CREATE1
 487     assert(!ctx->epollfd);
 488     ctx->epollfd = epoll_create1(EPOLL_CLOEXEC);
 489     if (ctx->epollfd == -1) {
 490         ctx->epoll_available = false;
 491     } else {
 492         ctx->epoll_available = true;
 493     }
 494 #endif
 495 }