aio-win32.c

   1 /*
   2  * QEMU aio implementation
   3  *
   4  * Copyright IBM Corp., 2008
   5  * Copyright Red Hat Inc., 2012
   6  *
   7  * Authors:
   8  *  Anthony Liguori   <aliguori@us.ibm.com>
   9  *  Paolo Bonzini     <pbonzini@redhat.com>
  10  *
  11  * This work is licensed under the terms of the GNU GPL, version 2.  See
  12  * the COPYING file in the top-level directory.
  13  *
  14  * Contributions after 2012-01-13 are licensed under the terms of the
  15  * GNU GPL, version 2 or (at your option) any later version.
  16  */
  17
  18 #include "qemu/osdep.h"
  19 #include "qemu-common.h"
  20 #include "block/block.h"
  21 #include "qemu/queue.h"
  22 #include "qemu/sockets.h"
  23 #include "qapi/error.h"
  24
  25 struct AioHandler {
  26     EventNotifier *e;
  27     IOHandler *io_read;
  28     IOHandler *io_write;
  29     EventNotifierHandler *io_notify;
  30     GPollFD pfd;
  31     int deleted;
  32     void *opaque;
  33     bool is_external;
  34     QLIST_ENTRY(AioHandler) node;
  35 };
  36
  37 void aio_set_fd_handler(AioContext *ctx,
  38                         int fd,
  39                         bool is_external,
  40                         IOHandler *io_read,
  41                         IOHandler *io_write,
  42                         AioPollFn *io_poll,
  43                         void *opaque)
  44 {
  45     /* fd is a SOCKET in our case */
  46     AioHandler *node;
  47
  48     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
  49         if (node->pfd.fd == fd && !node->deleted) {
  50             break;
  51         }
  52     }
  53
  54     /* Are we deleting the fd handler? */
  55     if (!io_read && !io_write) {
  56         if (node) {
  57             /* If the lock is held, just mark the node as deleted */
  58             if (ctx->walking_handlers) {
  59                 node->deleted = 1;
  60                 node->pfd.revents = 0;
  61             } else {
  62                 /* Otherwise, delete it for real.  We can't just mark it as
  63                  * deleted because deleted nodes are only cleaned up after
  64                  * releasing the walking_handlers lock.
  65                  */
  66                 QLIST_REMOVE(node, node);
  67                 g_free(node);
  68             }
  69         }
  70     } else {
  71         HANDLE event;
  72
  73         if (node == NULL) {
  74             /* Alloc and insert if it's not already there */
  75             node = g_new0(AioHandler, 1);
  76             node->pfd.fd = fd;
  77             QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node);
  78         }
  79
  80         node->pfd.events = 0;
  81         if (node->io_read) {
  82             node->pfd.events |= G_IO_IN;
  83         }
  84         if (node->io_write) {
  85             node->pfd.events |= G_IO_OUT;
  86         }
  87
  88         node->e = &ctx->notifier;
  89
  90         /* Update handler with latest information */
  91         node->opaque = opaque;
  92         node->io_read = io_read;
  93         node->io_write = io_write;
  94         node->is_external = is_external;
  95
  96         event = event_notifier_get_handle(&ctx->notifier);
  97         WSAEventSelect(node->pfd.fd, event,
  98                        FD_READ | FD_ACCEPT | FD_CLOSE |
  99                        FD_CONNECT | FD_WRITE | FD_OOB);
 100     }
 101
 102     aio_notify(ctx);
 103 }
 104
 105 void aio_set_event_notifier(AioContext *ctx,
 106                             EventNotifier *e,
 107                             bool is_external,
 108                             EventNotifierHandler *io_notify,
 109                             AioPollFn *io_poll)
 110 {
 111     AioHandler *node;
 112
 113     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 114         if (node->e == e && !node->deleted) {
 115             break;
 116         }
 117     }
 118
 119     /* Are we deleting the fd handler? */
 120     if (!io_notify) {
 121         if (node) {
 122             g_source_remove_poll(&ctx->source, &node->pfd);
 123
 124             /* If the lock is held, just mark the node as deleted */
 125             if (ctx->walking_handlers) {
 126                 node->deleted = 1;
 127                 node->pfd.revents = 0;
 128             } else {
 129                 /* Otherwise, delete it for real.  We can't just mark it as
 130                  * deleted because deleted nodes are only cleaned up after
 131                  * releasing the walking_handlers lock.
 132                  */
 133                 QLIST_REMOVE(node, node);
 134                 g_free(node);
 135             }
 136         }
 137     } else {
 138         if (node == NULL) {
 139             /* Alloc and insert if it's not already there */
 140             node = g_new0(AioHandler, 1);
 141             node->e = e;
 142             node->pfd.fd = (uintptr_t)event_notifier_get_handle(e);
 143             node->pfd.events = G_IO_IN;
 144             node->is_external = is_external;
 145             QLIST_INSERT_HEAD(&ctx->aio_handlers, node, node);
 146
 147             g_source_add_poll(&ctx->source, &node->pfd);
 148         }
 149         /* Update handler with latest information */
 150         node->io_notify = io_notify;
 151     }
 152
 153     aio_notify(ctx);
 154 }
 155
 156 bool aio_prepare(AioContext *ctx)
 157 {
 158     static struct timeval tv0;
 159     AioHandler *node;
 160     bool have_select_revents = false;
 161     fd_set rfds, wfds;
 162
 163     /* fill fd sets */
 164     FD_ZERO(&rfds);
 165     FD_ZERO(&wfds);
 166     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 167         if (node->io_read) {
 168             FD_SET ((SOCKET)node->pfd.fd, &rfds);
 169         }
 170         if (node->io_write) {
 171             FD_SET ((SOCKET)node->pfd.fd, &wfds);
 172         }
 173     }
 174
 175     if (select(0, &rfds, &wfds, NULL, &tv0) > 0) {
 176         QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 177             node->pfd.revents = 0;
 178             if (FD_ISSET(node->pfd.fd, &rfds)) {
 179                 node->pfd.revents |= G_IO_IN;
 180                 have_select_revents = true;
 181             }
 182
 183             if (FD_ISSET(node->pfd.fd, &wfds)) {
 184                 node->pfd.revents |= G_IO_OUT;
 185                 have_select_revents = true;
 186             }
 187         }
 188     }
 189
 190     return have_select_revents;
 191 }
 192
 193 bool aio_pending(AioContext *ctx)
 194 {
 195     AioHandler *node;
 196
 197     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 198         if (node->pfd.revents && node->io_notify) {
 199             return true;
 200         }
 201
 202         if ((node->pfd.revents & G_IO_IN) && node->io_read) {
 203             return true;
 204         }
 205         if ((node->pfd.revents & G_IO_OUT) && node->io_write) {
 206             return true;
 207         }
 208     }
 209
 210     return false;
 211 }
 212
 213 static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event)
 214 {
 215     AioHandler *node;
 216     bool progress = false;
 217
 218     /*
 219      * We have to walk very carefully in case aio_set_fd_handler is
 220      * called while we're walking.
 221      */
 222     node = QLIST_FIRST(&ctx->aio_handlers);
 223     while (node) {
 224         AioHandler *tmp;
 225         int revents = node->pfd.revents;
 226
 227         ctx->walking_handlers++;
 228
 229         if (!node->deleted &&
 230             (revents || event_notifier_get_handle(node->e) == event) &&
 231             node->io_notify) {
 232             node->pfd.revents = 0;
 233             node->io_notify(node->e);
 234
 235             /* aio_notify() does not count as progress */
 236             if (node->e != &ctx->notifier) {
 237                 progress = true;
 238             }
 239         }
 240
 241         if (!node->deleted &&
 242             (node->io_read || node->io_write)) {
 243             node->pfd.revents = 0;
 244             if ((revents & G_IO_IN) && node->io_read) {
 245                 node->io_read(node->opaque);
 246                 progress = true;
 247             }
 248             if ((revents & G_IO_OUT) && node->io_write) {
 249                 node->io_write(node->opaque);
 250                 progress = true;
 251             }
 252
 253             /* if the next select() will return an event, we have progressed */
 254             if (event == event_notifier_get_handle(&ctx->notifier)) {
 255                 WSANETWORKEVENTS ev;
 256                 WSAEnumNetworkEvents(node->pfd.fd, event, &ev);
 257                 if (ev.lNetworkEvents) {
 258                     progress = true;
 259                 }
 260             }
 261         }
 262
 263         tmp = node;
 264         node = QLIST_NEXT(node, node);
 265
 266         ctx->walking_handlers--;
 267
 268         if (!ctx->walking_handlers && tmp->deleted) {
 269             QLIST_REMOVE(tmp, node);
 270             g_free(tmp);
 271         }
 272     }
 273
 274     return progress;
 275 }
 276
 277 bool aio_dispatch(AioContext *ctx, bool dispatch_fds)
 278 {
 279     bool progress;
 280
 281     progress = aio_bh_poll(ctx);
 282     if (dispatch_fds) {
 283         progress |= aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE);
 284     }
 285     progress |= timerlistgroup_run_timers(&ctx->tlg);
 286     return progress;
 287 }
 288
 289 bool aio_poll(AioContext *ctx, bool blocking)
 290 {
 291     AioHandler *node;
 292     HANDLE events[MAXIMUM_WAIT_OBJECTS + 1];
 293     bool progress, have_select_revents, first;
 294     int count;
 295     int timeout;
 296
 297     aio_context_acquire(ctx);
 298     progress = false;
 299
 300     /* aio_notify can avoid the expensive event_notifier_set if
 301      * everything (file descriptors, bottom halves, timers) will
 302      * be re-evaluated before the next blocking poll().  This is
 303      * already true when aio_poll is called with blocking == false;
 304      * if blocking == true, it is only true after poll() returns,
 305      * so disable the optimization now.
 306      */
 307     if (blocking) {
 308         atomic_add(&ctx->notify_me, 2);
 309     }
 310
 311     have_select_revents = aio_prepare(ctx);
 312
 313     ctx->walking_handlers++;
 314
 315     /* fill fd sets */
 316     count = 0;
 317     QLIST_FOREACH(node, &ctx->aio_handlers, node) {
 318         if (!node->deleted && node->io_notify
 319             && aio_node_check(ctx, node->is_external)) {
 320             events[count++] = event_notifier_get_handle(node->e);
 321         }
 322     }
 323
 324     ctx->walking_handlers--;
 325     first = true;
 326
 327     /* ctx->notifier is always registered.  */
 328     assert(count > 0);
 329
 330     /* Multiple iterations, all of them non-blocking except the first,
 331      * may be necessary to process all pending events.  After the first
 332      * WaitForMultipleObjects call ctx->notify_me will be decremented.
 333      */
 334     do {
 335         HANDLE event;
 336         int ret;
 337
 338         timeout = blocking && !have_select_revents
 339             ? qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)) : 0;
 340         if (timeout) {
 341             aio_context_release(ctx);
 342         }
 343         ret = WaitForMultipleObjects(count, events, FALSE, timeout);
 344         if (blocking) {
 345             assert(first);
 346             atomic_sub(&ctx->notify_me, 2);
 347         }
 348         if (timeout) {
 349             aio_context_acquire(ctx);
 350         }
 351
 352         if (first) {
 353             aio_notify_accept(ctx);
 354             progress |= aio_bh_poll(ctx);
 355             first = false;
 356         }
 357
 358         /* if we have any signaled events, dispatch event */
 359         event = NULL;
 360         if ((DWORD) (ret - WAIT_OBJECT_0) < count) {
 361             event = events[ret - WAIT_OBJECT_0];
 362             events[ret - WAIT_OBJECT_0] = events[--count];
 363         } else if (!have_select_revents) {
 364             break;
 365         }
 366
 367         have_select_revents = false;
 368         blocking = false;
 369
 370         progress |= aio_dispatch_handlers(ctx, event);
 371     } while (count > 0);
 372
 373     progress |= timerlistgroup_run_timers(&ctx->tlg);
 374
 375     aio_context_release(ctx);
 376     return progress;
 377 }
 378
 379 void aio_context_setup(AioContext *ctx)
 380 {
 381 }
 382
 383 void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns, Error **errp)
 384 {
 385     error_setg(errp, "AioContext polling is not implemented on Windows");
 386 }