aio.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 typedef struct AioHandler AioHandler;
  22
  23 /* The list of registered AIO handlers */
  24 static QLIST_HEAD(, AioHandler) aio_handlers;
  25
  26 /* This is a simple lock used to protect the aio_handlers list.  Specifically,
  27  * it's used to ensure that no callbacks are removed while we're walking and
  28  * dispatching callbacks.
  29  */
  30 static int walking_handlers;
  31
  32 struct AioHandler
  33 {
  34     int fd;
  35     IOHandler *io_read;
  36     IOHandler *io_write;
  37     AioFlushHandler *io_flush;
  38     int deleted;
  39     void *opaque;
  40     QLIST_ENTRY(AioHandler) node;
  41 };
  42
  43 static AioHandler *find_aio_handler(int fd)
  44 {
  45     AioHandler *node;
  46
  47     QLIST_FOREACH(node, &aio_handlers, node) {
  48         if (node->fd == fd)
  49             if (!node->deleted)
  50                 return node;
  51     }
  52
  53     return NULL;
  54 }
  55
  56 int qemu_aio_set_fd_handler(int fd,
  57                             IOHandler *io_read,
  58                             IOHandler *io_write,
  59                             AioFlushHandler *io_flush,
  60                             void *opaque)
  61 {
  62     AioHandler *node;
  63
  64     node = find_aio_handler(fd);
  65
  66     /* Are we deleting the fd handler? */
  67     if (!io_read && !io_write) {
  68         if (node) {
  69             /* If the lock is held, just mark the node as deleted */
  70             if (walking_handlers)
  71                 node->deleted = 1;
  72             else {
  73                 /* Otherwise, delete it for real.  We can't just mark it as
  74                  * deleted because deleted nodes are only cleaned up after
  75                  * releasing the walking_handlers lock.
  76                  */
  77                 QLIST_REMOVE(node, node);
  78                 g_free(node);
  79             }
  80         }
  81     } else {
  82         if (node == NULL) {
  83             /* Alloc and insert if it's not already there */
  84             node = g_malloc0(sizeof(AioHandler));
  85             node->fd = fd;
  86             QLIST_INSERT_HEAD(&aio_handlers, node, node);
  87         }
  88         /* Update handler with latest information */
  89         node->io_read = io_read;
  90         node->io_write = io_write;
  91         node->io_flush = io_flush;
  92         node->opaque = opaque;
  93     }
  94
  95     qemu_set_fd_handler2(fd, NULL, io_read, io_write, opaque);
  96
  97     return 0;
  98 }
  99
 100 void qemu_aio_flush(void)
 101 {
 102     while (qemu_aio_wait());
 103 }
 104
 105 bool qemu_aio_wait(void)
 106 {
 107     int ret;
 108
 109     /*
 110      * If there are callbacks left that have been queued, we need to call then.
 111      * Do not call select in this case, because it is possible that the caller
 112      * does not need a complete flush (as is the case for qemu_aio_wait loops).
 113      */
 114     if (qemu_bh_poll()) {
 115         return true;
 116     }
 117
 118     do {
 119         AioHandler *node;
 120         fd_set rdfds, wrfds;
 121         bool busy;
 122         int max_fd = -1;
 123
 124         walking_handlers = 1;
 125
 126         FD_ZERO(&rdfds);
 127         FD_ZERO(&wrfds);
 128
 129         /* fill fd sets */
 130         busy = false;
 131         QLIST_FOREACH(node, &aio_handlers, node) {
 132             /* If there aren't pending AIO operations, don't invoke callbacks.
 133              * Otherwise, if there are no AIO requests, qemu_aio_wait() would
 134              * wait indefinitely.
 135              */
 136             if (node->io_flush) {
 137                 if (node->io_flush(node->opaque) == 0) {
 138                     continue;
 139                 }
 140                 busy = true;
 141             }
 142             if (!node->deleted && node->io_read) {
 143                 FD_SET(node->fd, &rdfds);
 144                 max_fd = MAX(max_fd, node->fd + 1);
 145             }
 146             if (!node->deleted && node->io_write) {
 147                 FD_SET(node->fd, &wrfds);
 148                 max_fd = MAX(max_fd, node->fd + 1);
 149             }
 150         }
 151
 152         walking_handlers = 0;
 153
 154         /* No AIO operations?  Get us out of here */
 155         if (!busy) {
 156             return false;
 157         }
 158
 159         /* wait until next event */
 160         ret = select(max_fd, &rdfds, &wrfds, NULL, NULL);
 161         if (ret == -1 && errno == EINTR)
 162             continue;
 163
 164         /* if we have any readable fds, dispatch event */
 165         if (ret > 0) {
 166             walking_handlers = 1;
 167
 168             /* we have to walk very carefully in case
 169              * qemu_aio_set_fd_handler is called while we're walking */
 170             node = QLIST_FIRST(&aio_handlers);
 171             while (node) {
 172                 AioHandler *tmp;
 173
 174                 if (!node->deleted &&
 175                     FD_ISSET(node->fd, &rdfds) &&
 176                     node->io_read) {
 177                     node->io_read(node->opaque);
 178                 }
 179                 if (!node->deleted &&
 180                     FD_ISSET(node->fd, &wrfds) &&
 181                     node->io_write) {
 182                     node->io_write(node->opaque);
 183                 }
 184
 185                 tmp = node;
 186                 node = QLIST_NEXT(node, node);
 187
 188                 if (tmp->deleted) {
 189                     QLIST_REMOVE(tmp, node);
 190                     g_free(tmp);
 191                 }
 192             }
 193
 194             walking_handlers = 0;
 195         }
 196     } while (ret == 0);
 197
 198     return true;
 199 }