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/main-loop.h" 22 #include "qemu/queue.h" 23 #include "qemu/sockets.h" 24 #include "qapi/error.h" 25 #include "qemu/rcu_queue.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 bool is_external; 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 bool is_external, 68 IOHandler *io_read, 69 IOHandler *io_write, 70 AioPollFn *io_poll, 71 IOHandler *io_poll_ready, 72 void *opaque) 73 { 74 /* fd is a SOCKET in our case */ 75 AioHandler *old_node; 76 AioHandler *node = NULL; 77 78 qemu_lockcnt_lock(&ctx->list_lock); 79 QLIST_FOREACH(old_node, &ctx->aio_handlers, node) { 80 if (old_node->pfd.fd == fd && !old_node->deleted) { 81 break; 82 } 83 } 84 85 if (io_read || io_write) { 86 HANDLE event; 87 long bitmask = 0; 88 89 /* Alloc and insert if it's not already there */ 90 node = g_new0(AioHandler, 1); 91 node->pfd.fd = fd; 92 93 node->pfd.events = 0; 94 if (node->io_read) { 95 node->pfd.events |= G_IO_IN; 96 } 97 if (node->io_write) { 98 node->pfd.events |= G_IO_OUT; 99 } 100 101 node->e = &ctx->notifier; 102 103 /* Update handler with latest information */ 104 node->opaque = opaque; 105 node->io_read = io_read; 106 node->io_write = io_write; 107 node->is_external = is_external; 108 109 if (io_read) { 110 bitmask |= FD_READ | FD_ACCEPT | FD_CLOSE; 111 } 112 113 if (io_write) { 114 bitmask |= FD_WRITE | FD_CONNECT; 115 } 116 117 QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node); 118 event = event_notifier_get_handle(&ctx->notifier); 119 WSAEventSelect(node->pfd.fd, event, bitmask); 120 } 121 if (old_node) { 122 aio_remove_fd_handler(ctx, old_node); 123 } 124 125 qemu_lockcnt_unlock(&ctx->list_lock); 126 aio_notify(ctx); 127 } 128 129 void aio_set_fd_poll(AioContext *ctx, int fd, 130 IOHandler *io_poll_begin, 131 IOHandler *io_poll_end) 132 { 133 /* Not implemented */ 134 } 135 136 void aio_set_event_notifier(AioContext *ctx, 137 EventNotifier *e, 138 bool is_external, 139 EventNotifierHandler *io_notify, 140 AioPollFn *io_poll, 141 EventNotifierHandler *io_poll_ready) 142 { 143 AioHandler *node; 144 145 qemu_lockcnt_lock(&ctx->list_lock); 146 QLIST_FOREACH(node, &ctx->aio_handlers, node) { 147 if (node->e == e && !node->deleted) { 148 break; 149 } 150 } 151 152 /* Are we deleting the fd handler? */ 153 if (!io_notify) { 154 if (node) { 155 aio_remove_fd_handler(ctx, node); 156 } 157 } else { 158 if (node == NULL) { 159 /* Alloc and insert if it's not already there */ 160 node = g_new0(AioHandler, 1); 161 node->e = e; 162 node->pfd.fd = (uintptr_t)event_notifier_get_handle(e); 163 node->pfd.events = G_IO_IN; 164 node->is_external = is_external; 165 QLIST_INSERT_HEAD_RCU(&ctx->aio_handlers, node, node); 166 167 g_source_add_poll(&ctx->source, &node->pfd); 168 } 169 /* Update handler with latest information */ 170 node->io_notify = io_notify; 171 } 172 173 qemu_lockcnt_unlock(&ctx->list_lock); 174 aio_notify(ctx); 175 } 176 177 void aio_set_event_notifier_poll(AioContext *ctx, 178 EventNotifier *notifier, 179 EventNotifierHandler *io_poll_begin, 180 EventNotifierHandler *io_poll_end) 181 { 182 /* Not implemented */ 183 } 184 185 bool aio_prepare(AioContext *ctx) 186 { 187 static struct timeval tv0; 188 AioHandler *node; 189 bool have_select_revents = false; 190 fd_set rfds, wfds; 191 192 /* 193 * We have to walk very carefully in case aio_set_fd_handler is 194 * called while we're walking. 195 */ 196 qemu_lockcnt_inc(&ctx->list_lock); 197 198 /* fill fd sets */ 199 FD_ZERO(&rfds); 200 FD_ZERO(&wfds); 201 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { 202 if (node->io_read) { 203 FD_SET ((SOCKET)node->pfd.fd, &rfds); 204 } 205 if (node->io_write) { 206 FD_SET ((SOCKET)node->pfd.fd, &wfds); 207 } 208 } 209 210 if (select(0, &rfds, &wfds, NULL, &tv0) > 0) { 211 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { 212 node->pfd.revents = 0; 213 if (FD_ISSET(node->pfd.fd, &rfds)) { 214 node->pfd.revents |= G_IO_IN; 215 have_select_revents = true; 216 } 217 218 if (FD_ISSET(node->pfd.fd, &wfds)) { 219 node->pfd.revents |= G_IO_OUT; 220 have_select_revents = true; 221 } 222 } 223 } 224 225 qemu_lockcnt_dec(&ctx->list_lock); 226 return have_select_revents; 227 } 228 229 bool aio_pending(AioContext *ctx) 230 { 231 AioHandler *node; 232 bool result = false; 233 234 /* 235 * We have to walk very carefully in case aio_set_fd_handler is 236 * called while we're walking. 237 */ 238 qemu_lockcnt_inc(&ctx->list_lock); 239 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { 240 if (node->pfd.revents && node->io_notify) { 241 result = true; 242 break; 243 } 244 245 if ((node->pfd.revents & G_IO_IN) && node->io_read) { 246 result = true; 247 break; 248 } 249 if ((node->pfd.revents & G_IO_OUT) && node->io_write) { 250 result = true; 251 break; 252 } 253 } 254 255 qemu_lockcnt_dec(&ctx->list_lock); 256 return result; 257 } 258 259 static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event) 260 { 261 AioHandler *node; 262 bool progress = false; 263 AioHandler *tmp; 264 265 /* 266 * We have to walk very carefully in case aio_set_fd_handler is 267 * called while we're walking. 268 */ 269 QLIST_FOREACH_SAFE_RCU(node, &ctx->aio_handlers, node, tmp) { 270 int revents = node->pfd.revents; 271 272 if (!node->deleted && 273 (revents || event_notifier_get_handle(node->e) == event) && 274 node->io_notify) { 275 node->pfd.revents = 0; 276 node->io_notify(node->e); 277 278 /* aio_notify() does not count as progress */ 279 if (node->e != &ctx->notifier) { 280 progress = true; 281 } 282 } 283 284 if (!node->deleted && 285 (node->io_read || node->io_write)) { 286 node->pfd.revents = 0; 287 if ((revents & G_IO_IN) && node->io_read) { 288 node->io_read(node->opaque); 289 progress = true; 290 } 291 if ((revents & G_IO_OUT) && node->io_write) { 292 node->io_write(node->opaque); 293 progress = true; 294 } 295 296 /* if the next select() will return an event, we have progressed */ 297 if (event == event_notifier_get_handle(&ctx->notifier)) { 298 WSANETWORKEVENTS ev; 299 WSAEnumNetworkEvents(node->pfd.fd, event, &ev); 300 if (ev.lNetworkEvents) { 301 progress = true; 302 } 303 } 304 } 305 306 if (node->deleted) { 307 if (qemu_lockcnt_dec_if_lock(&ctx->list_lock)) { 308 QLIST_REMOVE(node, node); 309 g_free(node); 310 qemu_lockcnt_inc_and_unlock(&ctx->list_lock); 311 } 312 } 313 } 314 315 return progress; 316 } 317 318 void aio_dispatch(AioContext *ctx) 319 { 320 qemu_lockcnt_inc(&ctx->list_lock); 321 aio_bh_poll(ctx); 322 aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE); 323 qemu_lockcnt_dec(&ctx->list_lock); 324 timerlistgroup_run_timers(&ctx->tlg); 325 } 326 327 bool aio_poll(AioContext *ctx, bool blocking) 328 { 329 AioHandler *node; 330 HANDLE events[MAXIMUM_WAIT_OBJECTS + 1]; 331 bool progress, have_select_revents, first; 332 int count; 333 int timeout; 334 335 /* 336 * There cannot be two concurrent aio_poll calls for the same AioContext (or 337 * an aio_poll concurrent with a GSource prepare/check/dispatch callback). 338 * We rely on this below to avoid slow locked accesses to ctx->notify_me. 339 * 340 * aio_poll() may only be called in the AioContext's thread. iohandler_ctx 341 * is special in that it runs in the main thread, but that thread's context 342 * is qemu_aio_context. 343 */ 344 assert(in_aio_context_home_thread(ctx == iohandler_get_aio_context() ? 345 qemu_get_aio_context() : ctx)); 346 progress = false; 347 348 /* aio_notify can avoid the expensive event_notifier_set if 349 * everything (file descriptors, bottom halves, timers) will 350 * be re-evaluated before the next blocking poll(). This is 351 * already true when aio_poll is called with blocking == false; 352 * if blocking == true, it is only true after poll() returns, 353 * so disable the optimization now. 354 */ 355 if (blocking) { 356 qatomic_set(&ctx->notify_me, qatomic_read(&ctx->notify_me) + 2); 357 /* 358 * Write ctx->notify_me before computing the timeout 359 * (reading bottom half flags, etc.). Pairs with 360 * smp_mb in aio_notify(). 361 */ 362 smp_mb(); 363 } 364 365 qemu_lockcnt_inc(&ctx->list_lock); 366 have_select_revents = aio_prepare(ctx); 367 368 /* fill fd sets */ 369 count = 0; 370 QLIST_FOREACH_RCU(node, &ctx->aio_handlers, node) { 371 if (!node->deleted && node->io_notify 372 && aio_node_check(ctx, node->is_external)) { 373 events[count++] = event_notifier_get_handle(node->e); 374 } 375 } 376 377 first = true; 378 379 /* ctx->notifier is always registered. */ 380 assert(count > 0); 381 382 /* Multiple iterations, all of them non-blocking except the first, 383 * may be necessary to process all pending events. After the first 384 * WaitForMultipleObjects call ctx->notify_me will be decremented. 385 */ 386 do { 387 HANDLE event; 388 int ret; 389 390 timeout = blocking && !have_select_revents 391 ? qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)) : 0; 392 ret = WaitForMultipleObjects(count, events, FALSE, timeout); 393 if (blocking) { 394 assert(first); 395 qatomic_store_release(&ctx->notify_me, 396 qatomic_read(&ctx->notify_me) - 2); 397 aio_notify_accept(ctx); 398 } 399 400 if (first) { 401 progress |= aio_bh_poll(ctx); 402 first = false; 403 } 404 405 /* if we have any signaled events, dispatch event */ 406 event = NULL; 407 if ((DWORD) (ret - WAIT_OBJECT_0) < count) { 408 event = events[ret - WAIT_OBJECT_0]; 409 events[ret - WAIT_OBJECT_0] = events[--count]; 410 } else if (!have_select_revents) { 411 break; 412 } 413 414 have_select_revents = false; 415 blocking = false; 416 417 progress |= aio_dispatch_handlers(ctx, event); 418 } while (count > 0); 419 420 qemu_lockcnt_dec(&ctx->list_lock); 421 422 progress |= timerlistgroup_run_timers(&ctx->tlg); 423 return progress; 424 } 425 426 void aio_context_setup(AioContext *ctx) 427 { 428 } 429 430 void aio_context_destroy(AioContext *ctx) 431 { 432 } 433 434 void aio_context_use_g_source(AioContext *ctx) 435 { 436 } 437 438 void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns, 439 int64_t grow, int64_t shrink, Error **errp) 440 { 441 if (max_ns) { 442 error_setg(errp, "AioContext polling is not implemented on Windows"); 443 } 444 } 445 446 void aio_context_set_aio_params(AioContext *ctx, int64_t max_batch, 447 Error **errp) 448 { 449 } 450