| 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 "block/block.h" |
| 20 | #include "qemu/main-loop.h" |
| 21 | #include "qemu/lockcnt.h" |
| 22 | #include "qemu/queue.h" |
| 23 | #include "qemu/sockets.h" |
| 24 | #include "qapi/error.h" |
| 25 | #include "qemu/rcu_queue.h" |
| 26 | #include "qemu/error-report.h" |
| 27 | |
| 28 | struct AioHandler { |
| 29 | EventNotifier *e; |
| 30 | IOHandler *io_read; |
| 31 | IOHandler *io_write; |
| 32 | EventNotifierHandler *io_notify; |
| 33 | GPollFD pfd; |
| 34 | int deleted; |
| 35 | void *opaque; |
| 36 | QLIST_ENTRY(AioHandler) node; |
| 37 | }; |
| 38 | |
| 39 | static void aio_remove_fd_handler(AioContext *ctx, AioHandler *node) |
| 40 | { |
| 41 | /* |
| 42 | * If the GSource is in the process of being destroyed then |
| 43 | * g_source_remove_poll() causes an assertion failure. Skip |
| 44 | * removal in that case, because glib cleans up its state during |
| 45 | * destruction anyway. |
| 46 | */ |
| 47 | if (!g_source_is_destroyed(&ctx->source)) { |
| 48 | g_source_remove_poll(&ctx->source, &node->pfd); |
| 49 | } |
| 50 | |
| 51 | /* If aio_poll is in progress, just mark the node as deleted */ |
| 52 | if (qemu_lockcnt_count(&ctx->list_lock)) { |
| 53 | node->deleted = 1; |
| 54 | node->pfd.revents = 0; |
| 55 | } else { |
| 56 | /* Otherwise, delete it for real. We can't just mark it as |
| 57 | * deleted because deleted nodes are only cleaned up after |
| 58 | * releasing the list_lock. |
| 59 | */ |
| 60 | QLIST_REMOVE(node, node); |
| 61 | g_free(node); |
| 62 | } |
| 63 | } |
| 64 | |
| 65 | void aio_set_fd_handler(AioContext *ctx, |
| 66 | int fd, |
| 67 | IOHandler *io_read, |
| 68 | IOHandler *io_write, |
| 69 | AioPollFn *io_poll, |
| 70 | IOHandler *io_poll_ready, |
| 71 | void *opaque) |
| 72 | { |
| 73 | AioHandler *old_node; |
| 74 | AioHandler *node = NULL; |
| 75 | SOCKET s; |
| 76 | |
| 77 | if (!fd_is_socket(fd)) { |
| 78 | error_report("fd=%d is not a socket, AIO implementation is missing", fd); |
| 79 | return; |
| 80 | } |
| 81 | |
| 82 | s = _get_osfhandle(fd); |
| 83 | |
| 84 | qemu_lockcnt_lock(&ctx->list_lock); |
| 85 | QLIST_FOREACH(old_node, &ctx->aio_handlers, node) { |
| 86 | if (old_node->pfd.fd == s && !old_node->deleted) { |
| 87 | break; |
| 88 | } |
| 89 | } |
| 90 | |
| 91 | if (io_read || io_write) { |
| 92 | HANDLE event; |
| 93 | long bitmask = 0; |
| 94 | |
| 95 | /* Alloc and insert if it's not already there */ |
| 96 | node = g_new0(AioHandler, 1); |
| 97 | node->pfd.fd = s; |
| 98 | |
| 99 | node->pfd.events = 0; |
| 100 | if (node->io_read) { |
| 101 | node->pfd.events |= G_IO_IN; |
| 102 | } |
| 103 | if (node->io_write) { |
| 104 | node->pfd.events |= G_IO_OUT; |
| 105 | } |
| 106 | |
| 107 | node->e = &ctx->notifier; |
| 108 | |
| 109 | /* Update handler with latest information */ |
| 110 | node->opaque = opaque; |
| 111 | node->io_read = io_read; |
| 112 | node->io_write = io_write; |
| 113 | |
| 114 | if (io_read) { |
| 115 | bitmask |= FD_READ | FD_ACCEPT | FD_CLOSE; |
| 116 | } |
| 117 | |
| 118 | if (io_write) { |
| 119 | bitmask |= FD_WRITE | FD_CONNECT; |
| 120 | } |
| 121 | |
| 122 | QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node); |
| 123 | event = event_notifier_get_handle(&ctx->notifier); |
| 124 | qemu_socket_select_nofail(fd, event, bitmask); |
| 125 | } |
| 126 | if (old_node) { |
| 127 | aio_remove_fd_handler(ctx, old_node); |
| 128 | } |
| 129 | |
| 130 | qemu_lockcnt_unlock(&ctx->list_lock); |
| 131 | aio_notify(ctx); |
| 132 | } |
| 133 | |
| 134 | void aio_set_event_notifier(AioContext *ctx, |
| 135 | EventNotifier *e, |
| 136 | EventNotifierHandler *io_notify, |
| 137 | AioPollFn *io_poll, |
| 138 | EventNotifierHandler *io_poll_ready) |
| 139 | { |
| 140 | AioHandler *node; |
| 141 | |
| 142 | qemu_lockcnt_lock(&ctx->list_lock); |
| 143 | QLIST_FOREACH(node, &ctx->aio_handlers, node) { |
| 144 | if (node->e == e && !node->deleted) { |
| 145 | break; |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | /* Are we deleting the fd handler? */ |
| 150 | if (!io_notify) { |
| 151 | if (node) { |
| 152 | aio_remove_fd_handler(ctx, node); |
| 153 | } |
| 154 | } else { |
| 155 | if (node == NULL) { |
| 156 | /* Alloc and insert if it's not already there */ |
| 157 | node = g_new0(AioHandler, 1); |
| 158 | node->e = e; |
| 159 | node->pfd.fd = (uintptr_t)event_notifier_get_handle(e); |
| 160 | node->pfd.events = G_IO_IN; |
| 161 | QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node); |
| 162 | |
| 163 | g_source_add_poll(&ctx->source, &node->pfd); |
| 164 | } |
| 165 | /* Update handler with latest information */ |
| 166 | node->io_notify = io_notify; |
| 167 | } |
| 168 | |
| 169 | qemu_lockcnt_unlock(&ctx->list_lock); |
| 170 | aio_notify(ctx); |
| 171 | } |
| 172 | |
| 173 | void aio_set_event_notifier_poll(AioContext *ctx, |
| 174 | EventNotifier *notifier, |
| 175 | EventNotifierHandler *io_poll_begin, |
| 176 | EventNotifierHandler *io_poll_end) |
| 177 | { |
| 178 | /* Not implemented */ |
| 179 | } |
| 180 | |
| 181 | bool aio_prepare(AioContext *ctx) |
| 182 | { |
| 183 | static struct timeval tv0; |
| 184 | AioHandler *node; |
| 185 | bool have_select_revents = false; |
| 186 | fd_set rfds, wfds; |
| 187 | |
| 188 | /* |
| 189 | * We have to walk very carefully in case aio_set_fd_handler is |
| 190 | * called while we're walking. |
| 191 | */ |
| 192 | qemu_lockcnt_inc(&ctx->list_lock); |
| 193 | |
| 194 | /* fill fd sets */ |
| 195 | FD_ZERO(&rfds); |
| 196 | FD_ZERO(&wfds); |
| 197 | QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { |
| 198 | if (node->io_read) { |
| 199 | FD_SET ((SOCKET)node->pfd.fd, &rfds); |
| 200 | } |
| 201 | if (node->io_write) { |
| 202 | FD_SET ((SOCKET)node->pfd.fd, &wfds); |
| 203 | } |
| 204 | } |
| 205 | |
| 206 | if (select(0, &rfds, &wfds, NULL, &tv0) > 0) { |
| 207 | QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { |
| 208 | node->pfd.revents = 0; |
| 209 | if (FD_ISSET(node->pfd.fd, &rfds)) { |
| 210 | node->pfd.revents |= G_IO_IN; |
| 211 | have_select_revents = true; |
| 212 | } |
| 213 | |
| 214 | if (FD_ISSET(node->pfd.fd, &wfds)) { |
| 215 | node->pfd.revents |= G_IO_OUT; |
| 216 | have_select_revents = true; |
| 217 | } |
| 218 | } |
| 219 | } |
| 220 | |
| 221 | qemu_lockcnt_dec(&ctx->list_lock); |
| 222 | return have_select_revents; |
| 223 | } |
| 224 | |
| 225 | bool aio_pending(AioContext *ctx) |
| 226 | { |
| 227 | AioHandler *node; |
| 228 | bool result = false; |
| 229 | |
| 230 | /* |
| 231 | * We have to walk very carefully in case aio_set_fd_handler is |
| 232 | * called while we're walking. |
| 233 | */ |
| 234 | qemu_lockcnt_inc(&ctx->list_lock); |
| 235 | QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { |
| 236 | if (node->pfd.revents && node->io_notify) { |
| 237 | result = true; |
| 238 | break; |
| 239 | } |
| 240 | |
| 241 | if ((node->pfd.revents & G_IO_IN) && node->io_read) { |
| 242 | result = true; |
| 243 | break; |
| 244 | } |
| 245 | if ((node->pfd.revents & G_IO_OUT) && node->io_write) { |
| 246 | result = true; |
| 247 | break; |
| 248 | } |
| 249 | } |
| 250 | |
| 251 | qemu_lockcnt_dec(&ctx->list_lock); |
| 252 | return result; |
| 253 | } |
| 254 | |
| 255 | static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event) |
| 256 | { |
| 257 | AioHandler *node; |
| 258 | bool progress = false; |
| 259 | AioHandler *tmp; |
| 260 | |
| 261 | /* |
| 262 | * We have to walk very carefully in case aio_set_fd_handler is |
| 263 | * called while we're walking. |
| 264 | */ |
| 265 | QLIST_FOREACH_SAFE_RCU(node, &ctx->aio_handlers, node, tmp) { |
| 266 | int revents = node->pfd.revents; |
| 267 | |
| 268 | if (!node->deleted && |
| 269 | (revents || event_notifier_get_handle(node->e) == event) && |
| 270 | node->io_notify) { |
| 271 | node->pfd.revents = 0; |
| 272 | node->io_notify(node->e); |
| 273 | |
| 274 | /* aio_notify() does not count as progress */ |
| 275 | if (node->e != &ctx->notifier) { |
| 276 | progress = true; |
| 277 | } |
| 278 | } |
| 279 | |
| 280 | if (!node->deleted && |
| 281 | (node->io_read || node->io_write)) { |
| 282 | node->pfd.revents = 0; |
| 283 | if ((revents & G_IO_IN) && node->io_read) { |
| 284 | node->io_read(node->opaque); |
| 285 | progress = true; |
| 286 | } |
| 287 | if ((revents & G_IO_OUT) && node->io_write) { |
| 288 | node->io_write(node->opaque); |
| 289 | progress = true; |
| 290 | } |
| 291 | |
| 292 | /* if the next select() will return an event, we have progressed */ |
| 293 | if (event == event_notifier_get_handle(&ctx->notifier)) { |
| 294 | WSANETWORKEVENTS ev; |
| 295 | WSAEnumNetworkEvents(node->pfd.fd, event, &ev); |
| 296 | if (ev.lNetworkEvents) { |
| 297 | progress = true; |
| 298 | } |
| 299 | } |
| 300 | } |
| 301 | |
| 302 | if (node->deleted) { |
| 303 | if (qemu_lockcnt_dec_if_lock(&ctx->list_lock)) { |
| 304 | QLIST_REMOVE(node, node); |
| 305 | g_free(node); |
| 306 | qemu_lockcnt_inc_and_unlock(&ctx->list_lock); |
| 307 | } |
| 308 | } |
| 309 | } |
| 310 | |
| 311 | return progress; |
| 312 | } |
| 313 | |
| 314 | void aio_dispatch(AioContext *ctx) |
| 315 | { |
| 316 | qemu_lockcnt_inc(&ctx->list_lock); |
| 317 | aio_bh_poll(ctx); |
| 318 | aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE); |
| 319 | qemu_lockcnt_dec(&ctx->list_lock); |
| 320 | timerlistgroup_run_timers(&ctx->tlg); |
| 321 | } |
| 322 | |
| 323 | bool aio_poll(AioContext *ctx, bool blocking) |
| 324 | { |
| 325 | AioHandler *node; |
| 326 | HANDLE events[MAXIMUM_WAIT_OBJECTS]; |
| 327 | bool progress, have_select_revents, first; |
| 328 | unsigned count; |
| 329 | int timeout; |
| 330 | |
| 331 | /* |
| 332 | * There cannot be two concurrent aio_poll calls for the same AioContext (or |
| 333 | * an aio_poll concurrent with a GSource prepare/check/dispatch callback). |
| 334 | * We rely on this below to avoid slow locked accesses to ctx->notify_me. |
| 335 | * |
| 336 | * aio_poll() may only be called in the AioContext's thread. iohandler_ctx |
| 337 | * is special in that it runs in the main thread, but that thread's context |
| 338 | * is qemu_aio_context. |
| 339 | */ |
| 340 | assert(in_aio_context_home_thread(ctx == iohandler_get_aio_context() ? |
| 341 | qemu_get_aio_context() : ctx)); |
| 342 | progress = false; |
| 343 | |
| 344 | /* aio_notify can avoid the expensive event_notifier_set if |
| 345 | * everything (file descriptors, bottom halves, timers) will |
| 346 | * be re-evaluated before the next blocking poll(). This is |
| 347 | * already true when aio_poll is called with blocking == false; |
| 348 | * if blocking == true, it is only true after poll() returns, |
| 349 | * so disable the optimization now. |
| 350 | */ |
| 351 | if (blocking) { |
| 352 | qatomic_set(&ctx->notify_me, qatomic_read(&ctx->notify_me) + 2); |
| 353 | /* |
| 354 | * Write ctx->notify_me before computing the timeout |
| 355 | * (reading bottom half flags, etc.). Pairs with |
| 356 | * smp_mb in aio_notify(). |
| 357 | */ |
| 358 | smp_mb(); |
| 359 | } |
| 360 | |
| 361 | qemu_lockcnt_inc(&ctx->list_lock); |
| 362 | have_select_revents = aio_prepare(ctx); |
| 363 | |
| 364 | /* fill fd sets */ |
| 365 | count = 0; |
| 366 | QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { |
| 367 | if (!node->deleted && node->io_notify) { |
| 368 | assert(count < MAXIMUM_WAIT_OBJECTS); |
| 369 | events[count++] = event_notifier_get_handle(node->e); |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | first = true; |
| 374 | |
| 375 | /* ctx->notifier is always registered. */ |
| 376 | assert(count > 0); |
| 377 | |
| 378 | /* Multiple iterations, all of them non-blocking except the first, |
| 379 | * may be necessary to process all pending events. After the first |
| 380 | * WaitForMultipleObjects call ctx->notify_me will be decremented. |
| 381 | */ |
| 382 | do { |
| 383 | HANDLE event; |
| 384 | int ret; |
| 385 | |
| 386 | timeout = blocking && !have_select_revents |
| 387 | ? qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)) : 0; |
| 388 | ret = WaitForMultipleObjects(count, events, FALSE, timeout); |
| 389 | if (blocking) { |
| 390 | assert(first); |
| 391 | qatomic_store_release(&ctx->notify_me, |
| 392 | qatomic_read(&ctx->notify_me) - 2); |
| 393 | aio_notify_accept(ctx); |
| 394 | } |
| 395 | |
| 396 | if (first) { |
| 397 | progress |= aio_bh_poll(ctx); |
| 398 | first = false; |
| 399 | } |
| 400 | |
| 401 | /* if we have any signaled events, dispatch event */ |
| 402 | event = NULL; |
| 403 | if ((DWORD) (ret - WAIT_OBJECT_0) < count) { |
| 404 | event = events[ret - WAIT_OBJECT_0]; |
| 405 | events[ret - WAIT_OBJECT_0] = events[--count]; |
| 406 | } else if (!have_select_revents) { |
| 407 | break; |
| 408 | } |
| 409 | |
| 410 | have_select_revents = false; |
| 411 | blocking = false; |
| 412 | |
| 413 | progress |= aio_dispatch_handlers(ctx, event); |
| 414 | } while (count > 0); |
| 415 | |
| 416 | qemu_lockcnt_dec(&ctx->list_lock); |
| 417 | |
| 418 | progress |= timerlistgroup_run_timers(&ctx->tlg); |
| 419 | return progress; |
| 420 | } |
| 421 | |
| 422 | bool aio_context_setup(AioContext *ctx, Error **errp) |
| 423 | { |
| 424 | return true; |
| 425 | } |
| 426 | |
| 427 | void aio_context_destroy(AioContext *ctx) |
| 428 | { |
| 429 | } |
| 430 | |
| 431 | void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns, |
| 432 | int64_t grow, int64_t shrink, |
| 433 | int64_t weight, Error **errp) |
| 434 | { |
| 435 | if (max_ns) { |
| 436 | error_setg(errp, "AioContext polling is not implemented on Windows"); |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | void aio_context_set_aio_params(AioContext *ctx, int64_t max_batch) |
| 441 | { |
| 442 | } |