1 /* 2 * Data plane event loop 3 * 4 * Copyright (c) 2003-2008 Fabrice Bellard 5 * Copyright (c) 2009-2017 QEMU contributors 6 * 7 * Permission is hereby granted, free of charge, to any person obtaining a copy 8 * of this software and associated documentation files (the "Software"), to deal 9 * in the Software without restriction, including without limitation the rights 10 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 11 * copies of the Software, and to permit persons to whom the Software is 12 * furnished to do so, subject to the following conditions: 13 * 14 * The above copyright notice and this permission notice shall be included in 15 * all copies or substantial portions of the Software. 16 * 17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 23 * THE SOFTWARE. 24 */ 25 26 #include "qemu/osdep.h" 27 #include "qapi/error.h" 28 #include "qemu-common.h" 29 #include "block/aio.h" 30 #include "block/thread-pool.h" 31 #include "qemu/main-loop.h" 32 #include "qemu/atomic.h" 33 #include "block/raw-aio.h" 34 #include "qemu/coroutine_int.h" 35 #include "trace.h" 36 37 /***********************************************************/ 38 /* bottom halves (can be seen as timers which expire ASAP) */ 39 40 struct QEMUBH { 41 AioContext *ctx; 42 QEMUBHFunc *cb; 43 void *opaque; 44 QEMUBH *next; 45 bool scheduled; 46 bool idle; 47 bool deleted; 48 }; 49 50 void aio_bh_schedule_oneshot(AioContext *ctx, QEMUBHFunc *cb, void *opaque) 51 { 52 QEMUBH *bh; 53 bh = g_new(QEMUBH, 1); 54 *bh = (QEMUBH){ 55 .ctx = ctx, 56 .cb = cb, 57 .opaque = opaque, 58 }; 59 qemu_lockcnt_lock(&ctx->list_lock); 60 bh->next = ctx->first_bh; 61 bh->scheduled = 1; 62 bh->deleted = 1; 63 /* Make sure that the members are ready before putting bh into list */ 64 smp_wmb(); 65 ctx->first_bh = bh; 66 qemu_lockcnt_unlock(&ctx->list_lock); 67 aio_notify(ctx); 68 } 69 70 QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque) 71 { 72 QEMUBH *bh; 73 bh = g_new(QEMUBH, 1); 74 *bh = (QEMUBH){ 75 .ctx = ctx, 76 .cb = cb, 77 .opaque = opaque, 78 }; 79 qemu_lockcnt_lock(&ctx->list_lock); 80 bh->next = ctx->first_bh; 81 /* Make sure that the members are ready before putting bh into list */ 82 smp_wmb(); 83 ctx->first_bh = bh; 84 qemu_lockcnt_unlock(&ctx->list_lock); 85 return bh; 86 } 87 88 void aio_bh_call(QEMUBH *bh) 89 { 90 bh->cb(bh->opaque); 91 } 92 93 /* Multiple occurrences of aio_bh_poll cannot be called concurrently. 94 * The count in ctx->list_lock is incremented before the call, and is 95 * not affected by the call. 96 */ 97 int aio_bh_poll(AioContext *ctx) 98 { 99 QEMUBH *bh, **bhp, *next; 100 int ret; 101 bool deleted = false; 102 103 ret = 0; 104 for (bh = atomic_rcu_read(&ctx->first_bh); bh; bh = next) { 105 next = atomic_rcu_read(&bh->next); 106 /* The atomic_xchg is paired with the one in qemu_bh_schedule. The 107 * implicit memory barrier ensures that the callback sees all writes 108 * done by the scheduling thread. It also ensures that the scheduling 109 * thread sees the zero before bh->cb has run, and thus will call 110 * aio_notify again if necessary. 111 */ 112 if (atomic_xchg(&bh->scheduled, 0)) { 113 /* Idle BHs don't count as progress */ 114 if (!bh->idle) { 115 ret = 1; 116 } 117 bh->idle = 0; 118 aio_bh_call(bh); 119 } 120 if (bh->deleted) { 121 deleted = true; 122 } 123 } 124 125 /* remove deleted bhs */ 126 if (!deleted) { 127 return ret; 128 } 129 130 if (qemu_lockcnt_dec_if_lock(&ctx->list_lock)) { 131 bhp = &ctx->first_bh; 132 while (*bhp) { 133 bh = *bhp; 134 if (bh->deleted && !bh->scheduled) { 135 *bhp = bh->next; 136 g_free(bh); 137 } else { 138 bhp = &bh->next; 139 } 140 } 141 qemu_lockcnt_inc_and_unlock(&ctx->list_lock); 142 } 143 return ret; 144 } 145 146 void qemu_bh_schedule_idle(QEMUBH *bh) 147 { 148 bh->idle = 1; 149 /* Make sure that idle & any writes needed by the callback are done 150 * before the locations are read in the aio_bh_poll. 151 */ 152 atomic_mb_set(&bh->scheduled, 1); 153 } 154 155 void qemu_bh_schedule(QEMUBH *bh) 156 { 157 AioContext *ctx; 158 159 ctx = bh->ctx; 160 bh->idle = 0; 161 /* The memory barrier implicit in atomic_xchg makes sure that: 162 * 1. idle & any writes needed by the callback are done before the 163 * locations are read in the aio_bh_poll. 164 * 2. ctx is loaded before scheduled is set and the callback has a chance 165 * to execute. 166 */ 167 if (atomic_xchg(&bh->scheduled, 1) == 0) { 168 aio_notify(ctx); 169 } 170 } 171 172 173 /* This func is async. 174 */ 175 void qemu_bh_cancel(QEMUBH *bh) 176 { 177 atomic_mb_set(&bh->scheduled, 0); 178 } 179 180 /* This func is async.The bottom half will do the delete action at the finial 181 * end. 182 */ 183 void qemu_bh_delete(QEMUBH *bh) 184 { 185 bh->scheduled = 0; 186 bh->deleted = 1; 187 } 188 189 int64_t 190 aio_compute_timeout(AioContext *ctx) 191 { 192 int64_t deadline; 193 int timeout = -1; 194 QEMUBH *bh; 195 196 for (bh = atomic_rcu_read(&ctx->first_bh); bh; 197 bh = atomic_rcu_read(&bh->next)) { 198 if (bh->scheduled) { 199 if (bh->idle) { 200 /* idle bottom halves will be polled at least 201 * every 10ms */ 202 timeout = 10000000; 203 } else { 204 /* non-idle bottom halves will be executed 205 * immediately */ 206 return 0; 207 } 208 } 209 } 210 211 deadline = timerlistgroup_deadline_ns(&ctx->tlg); 212 if (deadline == 0) { 213 return 0; 214 } else { 215 return qemu_soonest_timeout(timeout, deadline); 216 } 217 } 218 219 static gboolean 220 aio_ctx_prepare(GSource *source, gint *timeout) 221 { 222 AioContext *ctx = (AioContext *) source; 223 224 atomic_or(&ctx->notify_me, 1); 225 226 /* We assume there is no timeout already supplied */ 227 *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)); 228 229 if (aio_prepare(ctx)) { 230 *timeout = 0; 231 } 232 233 return *timeout == 0; 234 } 235 236 static gboolean 237 aio_ctx_check(GSource *source) 238 { 239 AioContext *ctx = (AioContext *) source; 240 QEMUBH *bh; 241 242 atomic_and(&ctx->notify_me, ~1); 243 aio_notify_accept(ctx); 244 245 for (bh = ctx->first_bh; bh; bh = bh->next) { 246 if (bh->scheduled) { 247 return true; 248 } 249 } 250 return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0); 251 } 252 253 static gboolean 254 aio_ctx_dispatch(GSource *source, 255 GSourceFunc callback, 256 gpointer user_data) 257 { 258 AioContext *ctx = (AioContext *) source; 259 260 assert(callback == NULL); 261 aio_dispatch(ctx); 262 return true; 263 } 264 265 static void 266 aio_ctx_finalize(GSource *source) 267 { 268 AioContext *ctx = (AioContext *) source; 269 270 thread_pool_free(ctx->thread_pool); 271 272 #ifdef CONFIG_LINUX_AIO 273 if (ctx->linux_aio) { 274 laio_detach_aio_context(ctx->linux_aio, ctx); 275 laio_cleanup(ctx->linux_aio); 276 ctx->linux_aio = NULL; 277 } 278 #endif 279 280 assert(QSLIST_EMPTY(&ctx->scheduled_coroutines)); 281 qemu_bh_delete(ctx->co_schedule_bh); 282 283 qemu_lockcnt_lock(&ctx->list_lock); 284 assert(!qemu_lockcnt_count(&ctx->list_lock)); 285 while (ctx->first_bh) { 286 QEMUBH *next = ctx->first_bh->next; 287 288 /* qemu_bh_delete() must have been called on BHs in this AioContext */ 289 assert(ctx->first_bh->deleted); 290 291 g_free(ctx->first_bh); 292 ctx->first_bh = next; 293 } 294 qemu_lockcnt_unlock(&ctx->list_lock); 295 296 aio_set_event_notifier(ctx, &ctx->notifier, false, NULL, NULL); 297 event_notifier_cleanup(&ctx->notifier); 298 qemu_rec_mutex_destroy(&ctx->lock); 299 qemu_lockcnt_destroy(&ctx->list_lock); 300 timerlistgroup_deinit(&ctx->tlg); 301 aio_context_destroy(ctx); 302 } 303 304 static GSourceFuncs aio_source_funcs = { 305 aio_ctx_prepare, 306 aio_ctx_check, 307 aio_ctx_dispatch, 308 aio_ctx_finalize 309 }; 310 311 GSource *aio_get_g_source(AioContext *ctx) 312 { 313 g_source_ref(&ctx->source); 314 return &ctx->source; 315 } 316 317 ThreadPool *aio_get_thread_pool(AioContext *ctx) 318 { 319 if (!ctx->thread_pool) { 320 ctx->thread_pool = thread_pool_new(ctx); 321 } 322 return ctx->thread_pool; 323 } 324 325 #ifdef CONFIG_LINUX_AIO 326 LinuxAioState *aio_setup_linux_aio(AioContext *ctx, Error **errp) 327 { 328 if (!ctx->linux_aio) { 329 ctx->linux_aio = laio_init(errp); 330 if (ctx->linux_aio) { 331 laio_attach_aio_context(ctx->linux_aio, ctx); 332 } 333 } 334 return ctx->linux_aio; 335 } 336 337 LinuxAioState *aio_get_linux_aio(AioContext *ctx) 338 { 339 assert(ctx->linux_aio); 340 return ctx->linux_aio; 341 } 342 #endif 343 344 void aio_notify(AioContext *ctx) 345 { 346 /* Write e.g. bh->scheduled before reading ctx->notify_me. Pairs 347 * with atomic_or in aio_ctx_prepare or atomic_add in aio_poll. 348 */ 349 smp_mb(); 350 if (ctx->notify_me) { 351 event_notifier_set(&ctx->notifier); 352 atomic_mb_set(&ctx->notified, true); 353 } 354 } 355 356 void aio_notify_accept(AioContext *ctx) 357 { 358 if (atomic_xchg(&ctx->notified, false)) { 359 event_notifier_test_and_clear(&ctx->notifier); 360 } 361 } 362 363 static void aio_timerlist_notify(void *opaque, QEMUClockType type) 364 { 365 aio_notify(opaque); 366 } 367 368 static void event_notifier_dummy_cb(EventNotifier *e) 369 { 370 } 371 372 /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */ 373 static bool event_notifier_poll(void *opaque) 374 { 375 EventNotifier *e = opaque; 376 AioContext *ctx = container_of(e, AioContext, notifier); 377 378 return atomic_read(&ctx->notified); 379 } 380 381 static void co_schedule_bh_cb(void *opaque) 382 { 383 AioContext *ctx = opaque; 384 QSLIST_HEAD(, Coroutine) straight, reversed; 385 386 QSLIST_MOVE_ATOMIC(&reversed, &ctx->scheduled_coroutines); 387 QSLIST_INIT(&straight); 388 389 while (!QSLIST_EMPTY(&reversed)) { 390 Coroutine *co = QSLIST_FIRST(&reversed); 391 QSLIST_REMOVE_HEAD(&reversed, co_scheduled_next); 392 QSLIST_INSERT_HEAD(&straight, co, co_scheduled_next); 393 } 394 395 while (!QSLIST_EMPTY(&straight)) { 396 Coroutine *co = QSLIST_FIRST(&straight); 397 QSLIST_REMOVE_HEAD(&straight, co_scheduled_next); 398 trace_aio_co_schedule_bh_cb(ctx, co); 399 aio_context_acquire(ctx); 400 401 /* Protected by write barrier in qemu_aio_coroutine_enter */ 402 atomic_set(&co->scheduled, NULL); 403 qemu_aio_coroutine_enter(ctx, co); 404 aio_context_release(ctx); 405 } 406 } 407 408 AioContext *aio_context_new(Error **errp) 409 { 410 int ret; 411 AioContext *ctx; 412 413 ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext)); 414 aio_context_setup(ctx); 415 416 ret = event_notifier_init(&ctx->notifier, false); 417 if (ret < 0) { 418 error_setg_errno(errp, -ret, "Failed to initialize event notifier"); 419 goto fail; 420 } 421 g_source_set_can_recurse(&ctx->source, true); 422 qemu_lockcnt_init(&ctx->list_lock); 423 424 ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx); 425 QSLIST_INIT(&ctx->scheduled_coroutines); 426 427 aio_set_event_notifier(ctx, &ctx->notifier, 428 false, 429 (EventNotifierHandler *) 430 event_notifier_dummy_cb, 431 event_notifier_poll); 432 #ifdef CONFIG_LINUX_AIO 433 ctx->linux_aio = NULL; 434 #endif 435 ctx->thread_pool = NULL; 436 qemu_rec_mutex_init(&ctx->lock); 437 timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx); 438 439 ctx->poll_ns = 0; 440 ctx->poll_max_ns = 0; 441 ctx->poll_grow = 0; 442 ctx->poll_shrink = 0; 443 444 return ctx; 445 fail: 446 g_source_destroy(&ctx->source); 447 return NULL; 448 } 449 450 void aio_co_schedule(AioContext *ctx, Coroutine *co) 451 { 452 trace_aio_co_schedule(ctx, co); 453 const char *scheduled = atomic_cmpxchg(&co->scheduled, NULL, 454 __func__); 455 456 if (scheduled) { 457 fprintf(stderr, 458 "%s: Co-routine was already scheduled in '%s'\n", 459 __func__, scheduled); 460 abort(); 461 } 462 463 QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines, 464 co, co_scheduled_next); 465 qemu_bh_schedule(ctx->co_schedule_bh); 466 } 467 468 void aio_co_wake(struct Coroutine *co) 469 { 470 AioContext *ctx; 471 472 /* Read coroutine before co->ctx. Matches smp_wmb in 473 * qemu_coroutine_enter. 474 */ 475 smp_read_barrier_depends(); 476 ctx = atomic_read(&co->ctx); 477 478 aio_co_enter(ctx, co); 479 } 480 481 void aio_co_enter(AioContext *ctx, struct Coroutine *co) 482 { 483 if (ctx != qemu_get_current_aio_context()) { 484 aio_co_schedule(ctx, co); 485 return; 486 } 487 488 if (qemu_in_coroutine()) { 489 Coroutine *self = qemu_coroutine_self(); 490 assert(self != co); 491 QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, co, co_queue_next); 492 } else { 493 aio_context_acquire(ctx); 494 qemu_aio_coroutine_enter(ctx, co); 495 aio_context_release(ctx); 496 } 497 } 498 499 void aio_context_ref(AioContext *ctx) 500 { 501 g_source_ref(&ctx->source); 502 } 503 504 void aio_context_unref(AioContext *ctx) 505 { 506 g_source_unref(&ctx->source); 507 } 508 509 void aio_context_acquire(AioContext *ctx) 510 { 511 qemu_rec_mutex_lock(&ctx->lock); 512 } 513 514 void aio_context_release(AioContext *ctx) 515 { 516 qemu_rec_mutex_unlock(&ctx->lock); 517 } 518