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/queue.h" 22 #include "qemu/sockets.h" 23 #include "qapi/error.h" 24 #include "qemu/rcu_queue.h" 25 #include "qemu/error-report.h" 26 27 struct AioHandler { 28 EventNotifier *e; 29 IOHandler *io_read; 30 IOHandler *io_write; 31 EventNotifierHandler *io_notify; 32 GPollFD pfd; 33 int deleted; 34 void *opaque; 35 QLIST_ENTRY(AioHandler) node; 36 }; 37 38 static void aio_remove_fd_handler(AioContext *ctx, AioHandler *node) 39 { 40 /* 41 * If the GSource is in the process of being destroyed then 42 * g_source_remove_poll() causes an assertion failure. Skip 43 * removal in that case, because glib cleans up its state during 44 * destruction anyway. 45 */ 46 if (!g_source_is_destroyed(&ctx->source)) { 47 g_source_remove_poll(&ctx->source, &node->pfd); 48 } 49 50 /* If aio_poll is in progress, just mark the node as deleted */ 51 if (qemu_lockcnt_count(&ctx->list_lock)) { 52 node->deleted = 1; 53 node->pfd.revents = 0; 54 } else { 55 /* Otherwise, delete it for real. We can't just mark it as 56 * deleted because deleted nodes are only cleaned up after 57 * releasing the list_lock. 58 */ 59 QLIST_REMOVE(node, node); 60 g_free(node); 61 } 62 } 63 64 void aio_set_fd_handler(AioContext *ctx, 65 int fd, 66 IOHandler *io_read, 67 IOHandler *io_write, 68 AioPollFn *io_poll, 69 IOHandler *io_poll_ready, 70 void *opaque) 71 { 72 AioHandler *old_node; 73 AioHandler *node = NULL; 74 SOCKET s; 75 76 if (!fd_is_socket(fd)) { 77 error_report("fd=%d is not a socket, AIO implementation is missing", fd); 78 return; 79 } 80 81 s = _get_osfhandle(fd); 82 83 qemu_lockcnt_lock(&ctx->list_lock); 84 QLIST_FOREACH(old_node, &ctx->aio_handlers, node) { 85 if (old_node->pfd.fd == s && !old_node->deleted) { 86 break; 87 } 88 } 89 90 if (io_read || io_write) { 91 HANDLE event; 92 long bitmask = 0; 93 94 /* Alloc and insert if it's not already there */ 95 node = g_new0(AioHandler, 1); 96 node->pfd.fd = s; 97 98 node->pfd.events = 0; 99 if (node->io_read) { 100 node->pfd.events |= G_IO_IN; 101 } 102 if (node->io_write) { 103 node->pfd.events |= G_IO_OUT; 104 } 105 106 node->e = &ctx->notifier; 107 108 /* Update handler with latest information */ 109 node->opaque = opaque; 110 node->io_read = io_read; 111 node->io_write = io_write; 112 113 if (io_read) { 114 bitmask |= FD_READ | FD_ACCEPT | FD_CLOSE; 115 } 116 117 if (io_write) { 118 bitmask |= FD_WRITE | FD_CONNECT; 119 } 120 121 QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node); 122 event = event_notifier_get_handle(&ctx->notifier); 123 qemu_socket_select(fd, event, bitmask, NULL); 124 } 125 if (old_node) { 126 aio_remove_fd_handler(ctx, old_node); 127 } 128 129 qemu_lockcnt_unlock(&ctx->list_lock); 130 aio_notify(ctx); 131 } 132 133 void aio_set_event_notifier(AioContext *ctx, 134 EventNotifier *e, 135 EventNotifierHandler *io_notify, 136 AioPollFn *io_poll, 137 EventNotifierHandler *io_poll_ready) 138 { 139 AioHandler *node; 140 141 qemu_lockcnt_lock(&ctx->list_lock); 142 QLIST_FOREACH(node, &ctx->aio_handlers, node) { 143 if (node->e == e && !node->deleted) { 144 break; 145 } 146 } 147 148 /* Are we deleting the fd handler? */ 149 if (!io_notify) { 150 if (node) { 151 aio_remove_fd_handler(ctx, node); 152 } 153 } else { 154 if (node == NULL) { 155 /* Alloc and insert if it's not already there */ 156 node = g_new0(AioHandler, 1); 157 node->e = e; 158 node->pfd.fd = (uintptr_t)event_notifier_get_handle(e); 159 node->pfd.events = G_IO_IN; 160 QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node); 161 162 g_source_add_poll(&ctx->source, &node->pfd); 163 } 164 /* Update handler with latest information */ 165 node->io_notify = io_notify; 166 } 167 168 qemu_lockcnt_unlock(&ctx->list_lock); 169 aio_notify(ctx); 170 } 171 172 void aio_set_event_notifier_poll(AioContext *ctx, 173 EventNotifier *notifier, 174 EventNotifierHandler *io_poll_begin, 175 EventNotifierHandler *io_poll_end) 176 { 177 /* Not implemented */ 178 } 179 180 bool aio_prepare(AioContext *ctx) 181 { 182 static struct timeval tv0; 183 AioHandler *node; 184 bool have_select_revents = false; 185 fd_set rfds, wfds; 186 187 /* 188 * We have to walk very carefully in case aio_set_fd_handler is 189 * called while we're walking. 190 */ 191 qemu_lockcnt_inc(&ctx->list_lock); 192 193 /* fill fd sets */ 194 FD_ZERO(&rfds); 195 FD_ZERO(&wfds); 196 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { 197 if (node->io_read) { 198 FD_SET ((SOCKET)node->pfd.fd, &rfds); 199 } 200 if (node->io_write) { 201 FD_SET ((SOCKET)node->pfd.fd, &wfds); 202 } 203 } 204 205 if (select(0, &rfds, &wfds, NULL, &tv0) > 0) { 206 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { 207 node->pfd.revents = 0; 208 if (FD_ISSET(node->pfd.fd, &rfds)) { 209 node->pfd.revents |= G_IO_IN; 210 have_select_revents = true; 211 } 212 213 if (FD_ISSET(node->pfd.fd, &wfds)) { 214 node->pfd.revents |= G_IO_OUT; 215 have_select_revents = true; 216 } 217 } 218 } 219 220 qemu_lockcnt_dec(&ctx->list_lock); 221 return have_select_revents; 222 } 223 224 bool aio_pending(AioContext *ctx) 225 { 226 AioHandler *node; 227 bool result = false; 228 229 /* 230 * We have to walk very carefully in case aio_set_fd_handler is 231 * called while we're walking. 232 */ 233 qemu_lockcnt_inc(&ctx->list_lock); 234 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { 235 if (node->pfd.revents && node->io_notify) { 236 result = true; 237 break; 238 } 239 240 if ((node->pfd.revents & G_IO_IN) && node->io_read) { 241 result = true; 242 break; 243 } 244 if ((node->pfd.revents & G_IO_OUT) && node->io_write) { 245 result = true; 246 break; 247 } 248 } 249 250 qemu_lockcnt_dec(&ctx->list_lock); 251 return result; 252 } 253 254 static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event) 255 { 256 AioHandler *node; 257 bool progress = false; 258 AioHandler *tmp; 259 260 /* 261 * We have to walk very carefully in case aio_set_fd_handler is 262 * called while we're walking. 263 */ 264 QLIST_FOREACH_SAFE_RCU(node, &ctx->aio_handlers, node, tmp) { 265 int revents = node->pfd.revents; 266 267 if (!node->deleted && 268 (revents || event_notifier_get_handle(node->e) == event) && 269 node->io_notify) { 270 node->pfd.revents = 0; 271 node->io_notify(node->e); 272 273 /* aio_notify() does not count as progress */ 274 if (node->e != &ctx->notifier) { 275 progress = true; 276 } 277 } 278 279 if (!node->deleted && 280 (node->io_read || node->io_write)) { 281 node->pfd.revents = 0; 282 if ((revents & G_IO_IN) && node->io_read) { 283 node->io_read(node->opaque); 284 progress = true; 285 } 286 if ((revents & G_IO_OUT) && node->io_write) { 287 node->io_write(node->opaque); 288 progress = true; 289 } 290 291 /* if the next select() will return an event, we have progressed */ 292 if (event == event_notifier_get_handle(&ctx->notifier)) { 293 WSANETWORKEVENTS ev; 294 WSAEnumNetworkEvents(node->pfd.fd, event, &ev); 295 if (ev.lNetworkEvents) { 296 progress = true; 297 } 298 } 299 } 300 301 if (node->deleted) { 302 if (qemu_lockcnt_dec_if_lock(&ctx->list_lock)) { 303 QLIST_REMOVE(node, node); 304 g_free(node); 305 qemu_lockcnt_inc_and_unlock(&ctx->list_lock); 306 } 307 } 308 } 309 310 return progress; 311 } 312 313 void aio_dispatch(AioContext *ctx) 314 { 315 qemu_lockcnt_inc(&ctx->list_lock); 316 aio_bh_poll(ctx); 317 aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE); 318 qemu_lockcnt_dec(&ctx->list_lock); 319 timerlistgroup_run_timers(&ctx->tlg); 320 } 321 322 bool aio_poll(AioContext *ctx, bool blocking) 323 { 324 AioHandler *node; 325 HANDLE events[MAXIMUM_WAIT_OBJECTS]; 326 bool progress, have_select_revents, first; 327 unsigned count; 328 int timeout; 329 330 /* 331 * There cannot be two concurrent aio_poll calls for the same AioContext (or 332 * an aio_poll concurrent with a GSource prepare/check/dispatch callback). 333 * We rely on this below to avoid slow locked accesses to ctx->notify_me. 334 * 335 * aio_poll() may only be called in the AioContext's thread. iohandler_ctx 336 * is special in that it runs in the main thread, but that thread's context 337 * is qemu_aio_context. 338 */ 339 assert(in_aio_context_home_thread(ctx == iohandler_get_aio_context() ? 340 qemu_get_aio_context() : ctx)); 341 progress = false; 342 343 /* aio_notify can avoid the expensive event_notifier_set if 344 * everything (file descriptors, bottom halves, timers) will 345 * be re-evaluated before the next blocking poll(). This is 346 * already true when aio_poll is called with blocking == false; 347 * if blocking == true, it is only true after poll() returns, 348 * so disable the optimization now. 349 */ 350 if (blocking) { 351 qatomic_set(&ctx->notify_me, qatomic_read(&ctx->notify_me) + 2); 352 /* 353 * Write ctx->notify_me before computing the timeout 354 * (reading bottom half flags, etc.). Pairs with 355 * smp_mb in aio_notify(). 356 */ 357 smp_mb(); 358 } 359 360 qemu_lockcnt_inc(&ctx->list_lock); 361 have_select_revents = aio_prepare(ctx); 362 363 /* fill fd sets */ 364 count = 0; 365 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { 366 if (!node->deleted && node->io_notify) { 367 assert(count < MAXIMUM_WAIT_OBJECTS); 368 events[count++] = event_notifier_get_handle(node->e); 369 } 370 } 371 372 first = true; 373 374 /* ctx->notifier is always registered. */ 375 assert(count > 0); 376 377 /* Multiple iterations, all of them non-blocking except the first, 378 * may be necessary to process all pending events. After the first 379 * WaitForMultipleObjects call ctx->notify_me will be decremented. 380 */ 381 do { 382 HANDLE event; 383 int ret; 384 385 timeout = blocking && !have_select_revents 386 ? qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)) : 0; 387 ret = WaitForMultipleObjects(count, events, FALSE, timeout); 388 if (blocking) { 389 assert(first); 390 qatomic_store_release(&ctx->notify_me, 391 qatomic_read(&ctx->notify_me) - 2); 392 aio_notify_accept(ctx); 393 } 394 395 if (first) { 396 progress |= aio_bh_poll(ctx); 397 first = false; 398 } 399 400 /* if we have any signaled events, dispatch event */ 401 event = NULL; 402 if ((DWORD) (ret - WAIT_OBJECT_0) < count) { 403 event = events[ret - WAIT_OBJECT_0]; 404 events[ret - WAIT_OBJECT_0] = events[--count]; 405 } else if (!have_select_revents) { 406 break; 407 } 408 409 have_select_revents = false; 410 blocking = false; 411 412 progress |= aio_dispatch_handlers(ctx, event); 413 } while (count > 0); 414 415 qemu_lockcnt_dec(&ctx->list_lock); 416 417 progress |= timerlistgroup_run_timers(&ctx->tlg); 418 return progress; 419 } 420 421 void aio_context_setup(AioContext *ctx) 422 { 423 } 424 425 void aio_context_destroy(AioContext *ctx) 426 { 427 } 428 429 void aio_context_use_g_source(AioContext *ctx) 430 { 431 } 432 433 void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns, 434 int64_t grow, int64_t shrink, Error **errp) 435 { 436 if (max_ns) { 437 error_setg(errp, "AioContext polling is not implemented on Windows"); 438 } 439 } 440 441 void aio_context_set_aio_params(AioContext *ctx, int64_t max_batch) 442 { 443 } 444