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