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.h"
  18 #include "qemu-queue.h"
  19 #include "qemu_socket.h"
  20
  21 struct AioHandler
  22 {
  23     GPollFD pfd;
  24     IOHandler *io_read;
  25     IOHandler *io_write;
  26     AioFlushHandler *io_flush;
  27     int deleted;
  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                         AioFlushHandler *io_flush,
  50                         void *opaque)
  51 {
  52     AioHandler *node;
  53
  54     node = find_aio_handler(ctx, fd);
  55
  56     /* Are we deleting the fd handler? */
  57     if (!io_read && !io_write) {
  58         if (node) {
  59             /* If the lock is held, just mark the node as deleted */
  60             if (ctx->walking_handlers) {
  61                 node->deleted = 1;
  62                 node->pfd.revents = 0;
  63             } else {
  64                 /* Otherwise, delete it for real.  We can't just mark it as
  65                  * deleted because deleted nodes are only cleaned up after
  66                  * releasing the walking_handlers lock.
  67                  */
  68                 QLIST_REMOVE(node, node);
  69                 g_free(node);
  70             }
  71         }
  72     } else {
  73         if (node == NULL) {
  74             /* Alloc and insert if it's not already there */
  75             node = g_malloc0(sizeof(AioHandler));
  76             node->pfd.fd = fd;
  77             QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node);
  78         }
  79         /* Update handler with latest information */
  80         node->io_read = io_read;
  81         node->io_write = io_write;
  82         node->io_flush = io_flush;
  83         node->opaque = opaque;
  84
  85         node->pfd.events = (io_read ? G_IO_IN | G_IO_HUP : 0);
  86         node->pfd.events |= (io_write ? G_IO_OUT : 0);
  87     }
  88 }
  89
  90 void aio_set_event_notifier(AioContext *ctx,
  91                             EventNotifier *notifier,
  92                             EventNotifierHandler *io_read,
  93                             AioFlushEventNotifierHandler *io_flush)
  94 {
  95     aio_set_fd_handler(ctx, event_notifier_get_fd(notifier),
  96                        (IOHandler *)io_read, NULL,
  97                        (AioFlushHandler *)io_flush, notifier);
  98 }
  99
 100 bool aio_pending(AioContext *ctx)
 101 {
 102     AioHandler *node;
 103
 104     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 105         int revents;
 106
 107         /*
 108          * FIXME: right now we cannot get G_IO_HUP and G_IO_ERR because
 109          * main-loop.c is still select based (due to the slirp legacy).
 110          * If main-loop.c ever switches to poll, G_IO_ERR should be
 111          * tested too.  Dispatching G_IO_ERR to both handlers should be
 112          * okay, since handlers need to be ready for spurious wakeups.
 113          */
 114         revents = node->pfd.revents & node->pfd.events;
 115         if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read) {
 116             return true;
 117         }
 118         if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write) {
 119             return true;
 120         }
 121     }
 122
 123     return false;
 124 }
 125
 126 bool aio_poll(AioContext *ctx, bool blocking)
 127 {
 128     static struct timeval tv0;
 129     AioHandler *node;
 130     fd_set rdfds, wrfds;
 131     int max_fd = -1;
 132     int ret;
 133     bool busy, progress;
 134
 135     progress = false;
 136
 137     /*
 138      * If there are callbacks left that have been queued, we need to call then.
 139      * Do not call select in this case, because it is possible that the caller
 140      * does not need a complete flush (as is the case for qemu_aio_wait loops).
 141      */
 142     if (aio_bh_poll(ctx)) {
 143         blocking = false;
 144         progress = true;
 145     }
 146
 147     /*
 148      * Then dispatch any pending callbacks from the GSource.
 149      *
 150      * We have to walk very carefully in case qemu_aio_set_fd_handler is
 151      * called while we're walking.
 152      */
 153     node = QLIST_FIRST(&ctx->aio_handlers);
 154     while (node) {
 155         AioHandler *tmp;
 156         int revents;
 157
 158         ctx->walking_handlers++;
 159
 160         revents = node->pfd.revents & node->pfd.events;
 161         node->pfd.revents = 0;
 162
 163         /* See comment in aio_pending.  */
 164         if (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR) && node->io_read) {
 165             node->io_read(node->opaque);
 166             progress = true;
 167         }
 168         if (revents & (G_IO_OUT | G_IO_ERR) && node->io_write) {
 169             node->io_write(node->opaque);
 170             progress = true;
 171         }
 172
 173         tmp = node;
 174         node = QLIST_NEXT(node, node);
 175
 176         ctx->walking_handlers--;
 177
 178         if (!ctx->walking_handlers && tmp->deleted) {
 179             QLIST_REMOVE(tmp, node);
 180             g_free(tmp);
 181         }
 182     }
 183
 184     if (progress && !blocking) {
 185         return true;
 186     }
 187
 188     ctx->walking_handlers++;
 189
 190     FD_ZERO(&rdfds);
 191     FD_ZERO(&wrfds);
 192
 193     /* fill fd sets */
 194     busy = false;
 195     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 196         /* If there aren't pending AIO operations, don't invoke callbacks.
 197          * Otherwise, if there are no AIO requests, qemu_aio_wait() would
 198          * wait indefinitely.
 199          */
 200         if (!node->deleted && node->io_flush) {
 201             if (node->io_flush(node->opaque) == 0) {
 202                 continue;
 203             }
 204             busy = true;
 205         }
 206         if (!node->deleted && node->io_read) {
 207             FD_SET(node->pfd.fd, &rdfds);
 208             max_fd = MAX(max_fd, node->pfd.fd + 1);
 209         }
 210         if (!node->deleted && node->io_write) {
 211             FD_SET(node->pfd.fd, &wrfds);
 212             max_fd = MAX(max_fd, node->pfd.fd + 1);
 213         }
 214     }
 215
 216     ctx->walking_handlers--;
 217
 218     /* No AIO operations?  Get us out of here */
 219     if (!busy) {
 220         return progress;
 221     }
 222
 223     /* wait until next event */
 224     ret = select(max_fd, &rdfds, &wrfds, NULL, blocking ? NULL : &tv0);
 225
 226     /* if we have any readable fds, dispatch event */
 227     if (ret > 0) {
 228         /* we have to walk very carefully in case
 229          * qemu_aio_set_fd_handler is called while we're walking */
 230         node = QLIST_FIRST(&ctx->aio_handlers);
 231         while (node) {
 232             AioHandler *tmp;
 233
 234             ctx->walking_handlers++;
 235
 236             if (!node->deleted &&
 237                 FD_ISSET(node->pfd.fd, &rdfds) &&
 238                 node->io_read) {
 239                 node->io_read(node->opaque);
 240                 progress = true;
 241             }
 242             if (!node->deleted &&
 243                 FD_ISSET(node->pfd.fd, &wrfds) &&
 244                 node->io_write) {
 245                 node->io_write(node->opaque);
 246                 progress = true;
 247             }
 248
 249             tmp = node;
 250             node = QLIST_NEXT(node, node);
 251
 252             ctx->walking_handlers--;
 253
 254             if (!ctx->walking_handlers && tmp->deleted) {
 255                 QLIST_REMOVE(tmp, node);
 256                 g_free(tmp);
 257             }
 258         }
 259     }
 260
 261     return progress;
 262 }