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-common.h"
  17 #include "block/block.h"
  18 #include "qemu/queue.h"
  19 #include "qemu/sockets.h"
  20
  21 struct AioHandler
  22 {
  23     GPollFD pfd;
  24     IOHandler *io_read;
  25     IOHandler *io_write;
  26     int deleted;
  27     int pollfds_idx;
  28     void *opaque;
  29     QLIST_ENTRY(AioHandler) node;
  30 };
  31
  32 static AioHandler *find_aio_handler(AioContext *ctx, int fd)
  33 {
  34     AioHandler *node;
  35
  36     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
  37         if (node->pfd.fd == fd)
  38             if (!node->deleted)
  39                 return node;
  40     }
  41
  42     return NULL;
  43 }
  44
  45 void aio_set_fd_handler(AioContext *ctx,
  46                         int fd,
  47                         IOHandler *io_read,
  48                         IOHandler *io_write,
  49                         void *opaque)
  50 {
  51     AioHandler *node;
  52
  53     node = find_aio_handler(ctx, fd);
  54
  55     /* Are we deleting the fd handler? */
  56     if (!io_read && !io_write) {
  57         if (node) {
  58             g_source_remove_poll(&ctx->source, &node->pfd);
  59
  60             /* If the lock is held, just mark the node as deleted */
  61             if (ctx->walking_handlers) {
  62                 node->deleted = 1;
  63                 node->pfd.revents = 0;
  64             } else {
  65                 /* Otherwise, delete it for real.  We can't just mark it as
  66                  * deleted because deleted nodes are only cleaned up after
  67                  * releasing the walking_handlers lock.
  68                  */
  69                 QLIST_REMOVE(node, node);
  70                 g_free(node);
  71             }
  72         }
  73     } else {
  74         if (node == NULL) {
  75             /* Alloc and insert if it's not already there */
  76             node = g_malloc0(sizeof(AioHandler));
  77             node->pfd.fd = fd;
  78             QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node);
  79
  80             g_source_add_poll(&ctx->source, &node->pfd);
  81         }
  82         /* Update handler with latest information */
  83         node->io_read = io_read;
  84         node->io_write = io_write;
  85         node->opaque = opaque;
  86         node->pollfds_idx = -1;
  87
  88         node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP | G_IO_ERR : 0);
  89         node->pfd.events |= (io_write ? G_IO_OUT | G_IO_ERR : 0);
  90     }
  91
  92     aio_notify(ctx);
  93 }
  94
  95 void aio_set_event_notifier(AioContext *ctx,
  96                             EventNotifier *notifier,
  97                             EventNotifierHandler *io_read)
  98 {
  99     aio_set_fd_handler(ctx, event_notifier_get_fd(notifier),
 100                        (IOHandler *)io_read, NULL, notifier);
 101 }
 102
 103 bool aio_prepare(AioContext *ctx)
 104 {
 105     return false;
 106 }
 107
 108 bool aio_pending(AioContext *ctx)
 109 {
 110     AioHandler *node;
 111
 112     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 113         int revents;
 114
 115         revents = node->pfd.revents & node->pfd.events;
 116         if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read) {
 117             return true;
 118         }
 119         if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write) {
 120             return true;
 121         }
 122     }
 123
 124     return false;
 125 }
 126
 127 bool aio_dispatch(AioContext *ctx)
 128 {
 129     AioHandler *node;
 130     bool progress = false;
 131
 132     /*
 133      * If there are callbacks left that have been queued, we need to call them.
 134      * Do not call select in this case, because it is possible that the caller
 135      * does not need a complete flush (as is the case for aio_poll loops).
 136      */
 137     if (aio_bh_poll(ctx)) {
 138         progress = true;
 139     }
 140
 141     /*
 142      * We have to walk very carefully in case aio_set_fd_handler is
 143      * called while we're walking.
 144      */
 145     node = QLIST_FIRST(&ctx->aio_handlers);
 146     while (node) {
 147         AioHandler *tmp;
 148         int revents;
 149
 150         ctx->walking_handlers++;
 151
 152         revents = node->pfd.revents & node->pfd.events;
 153         node->pfd.revents = 0;
 154
 155         if (!node->deleted &&
 156             (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) &&
 157             node->io_read) {
 158             node->io_read(node->opaque);
 159
 160             /* aio_notify() does not count as progress */
 161             if (node->opaque != &ctx->notifier) {
 162                 progress = true;
 163             }
 164         }
 165         if (!node->deleted &&
 166             (revents & (G_IO_OUT | G_IO_ERR)) &&
 167             node->io_write) {
 168             node->io_write(node->opaque);
 169             progress = true;
 170         }
 171
 172         tmp = node;
 173         node = QLIST_NEXT(node, node);
 174
 175         ctx->walking_handlers--;
 176
 177         if (!ctx->walking_handlers && tmp->deleted) {
 178             QLIST_REMOVE(tmp, node);
 179             g_free(tmp);
 180         }
 181     }
 182
 183     /* Run our timers */
 184     progress |= timerlistgroup_run_timers(&ctx->tlg);
 185
 186     return progress;
 187 }
 188
 189 bool aio_poll(AioContext *ctx, bool blocking)
 190 {
 191     AioHandler *node;
 192     bool was_dispatching;
 193     int ret;
 194     bool progress;
 195
 196     was_dispatching = ctx->dispatching;
 197     progress = false;
 198
 199     /* aio_notify can avoid the expensive event_notifier_set if
 200      * everything (file descriptors, bottom halves, timers) will
 201      * be re-evaluated before the next blocking poll().  This is
 202      * already true when aio_poll is called with blocking == false;
 203      * if blocking == true, it is only true after poll() returns.
 204      *
 205      * If we're in a nested event loop, ctx->dispatching might be true.
 206      * In that case we can restore it just before returning, but we
 207      * have to clear it now.
 208      */
 209     aio_set_dispatching(ctx, !blocking);
 210
 211     ctx->walking_handlers++;
 212
 213     g_array_set_size(ctx->pollfds, 0);
 214
 215     /* fill pollfds */
 216     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 217         node->pollfds_idx = -1;
 218         if (!node->deleted && node->pfd.events) {
 219             GPollFD pfd = {
 220                 .fd = node->pfd.fd,
 221                 .events = node->pfd.events,
 222             };
 223             node->pollfds_idx = ctx->pollfds->len;
 224             g_array_append_val(ctx->pollfds, pfd);
 225         }
 226     }
 227
 228     ctx->walking_handlers--;
 229
 230     /* wait until next event */
 231     ret = qemu_poll_ns((GPollFD *)ctx->pollfds->data,
 232                          ctx->pollfds->len,
 233                          blocking ? aio_compute_timeout(ctx) : 0);
 234
 235     /* if we have any readable fds, dispatch event */
 236     if (ret > 0) {
 237         QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 238             if (node->pollfds_idx != -1) {
 239                 GPollFD *pfd = &g_array_index(ctx->pollfds, GPollFD,
 240                                               node->pollfds_idx);
 241                 node->pfd.revents = pfd->revents;
 242             }
 243         }
 244     }
 245
 246     /* Run dispatch even if there were no readable fds to run timers */
 247     aio_set_dispatching(ctx, true);
 248     if (aio_dispatch(ctx)) {
 249         progress = true;
 250     }
 251
 252     aio_set_dispatching(ctx, was_dispatching);
 253     return progress;
 254 }