1c2b38b27SPaolo Bonzini /* 2c2b38b27SPaolo Bonzini * Data plane event loop 3c2b38b27SPaolo Bonzini * 4c2b38b27SPaolo Bonzini * Copyright (c) 2003-2008 Fabrice Bellard 5c2b38b27SPaolo Bonzini * Copyright (c) 2009-2017 QEMU contributors 6c2b38b27SPaolo Bonzini * 7c2b38b27SPaolo Bonzini * Permission is hereby granted, free of charge, to any person obtaining a copy 8c2b38b27SPaolo Bonzini * of this software and associated documentation files (the "Software"), to deal 9c2b38b27SPaolo Bonzini * in the Software without restriction, including without limitation the rights 10c2b38b27SPaolo Bonzini * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 11c2b38b27SPaolo Bonzini * copies of the Software, and to permit persons to whom the Software is 12c2b38b27SPaolo Bonzini * furnished to do so, subject to the following conditions: 13c2b38b27SPaolo Bonzini * 14c2b38b27SPaolo Bonzini * The above copyright notice and this permission notice shall be included in 15c2b38b27SPaolo Bonzini * all copies or substantial portions of the Software. 16c2b38b27SPaolo Bonzini * 17c2b38b27SPaolo Bonzini * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 18c2b38b27SPaolo Bonzini * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 19c2b38b27SPaolo Bonzini * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 20c2b38b27SPaolo Bonzini * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 21c2b38b27SPaolo Bonzini * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 22c2b38b27SPaolo Bonzini * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 23c2b38b27SPaolo Bonzini * THE SOFTWARE. 24c2b38b27SPaolo Bonzini */ 25c2b38b27SPaolo Bonzini 26c2b38b27SPaolo Bonzini #include "qemu/osdep.h" 27c2b38b27SPaolo Bonzini #include "qapi/error.h" 28c2b38b27SPaolo Bonzini #include "block/aio.h" 29c2b38b27SPaolo Bonzini #include "block/thread-pool.h" 30c2b38b27SPaolo Bonzini #include "qemu/main-loop.h" 31c2b38b27SPaolo Bonzini #include "qemu/atomic.h" 328c6b0356SStefan Hajnoczi #include "qemu/rcu_queue.h" 33c2b38b27SPaolo Bonzini #include "block/raw-aio.h" 340c330a73SPaolo Bonzini #include "qemu/coroutine_int.h" 350c330a73SPaolo Bonzini #include "trace.h" 36c2b38b27SPaolo Bonzini 37c2b38b27SPaolo Bonzini /***********************************************************/ 38c2b38b27SPaolo Bonzini /* bottom halves (can be seen as timers which expire ASAP) */ 39c2b38b27SPaolo Bonzini 408c6b0356SStefan Hajnoczi /* QEMUBH::flags values */ 418c6b0356SStefan Hajnoczi enum { 428c6b0356SStefan Hajnoczi /* Already enqueued and waiting for aio_bh_poll() */ 438c6b0356SStefan Hajnoczi BH_PENDING = (1 << 0), 448c6b0356SStefan Hajnoczi 458c6b0356SStefan Hajnoczi /* Invoke the callback */ 468c6b0356SStefan Hajnoczi BH_SCHEDULED = (1 << 1), 478c6b0356SStefan Hajnoczi 488c6b0356SStefan Hajnoczi /* Delete without invoking callback */ 498c6b0356SStefan Hajnoczi BH_DELETED = (1 << 2), 508c6b0356SStefan Hajnoczi 518c6b0356SStefan Hajnoczi /* Delete after invoking callback */ 528c6b0356SStefan Hajnoczi BH_ONESHOT = (1 << 3), 538c6b0356SStefan Hajnoczi 548c6b0356SStefan Hajnoczi /* Schedule periodically when the event loop is idle */ 558c6b0356SStefan Hajnoczi BH_IDLE = (1 << 4), 568c6b0356SStefan Hajnoczi }; 578c6b0356SStefan Hajnoczi 58c2b38b27SPaolo Bonzini struct QEMUBH { 59c2b38b27SPaolo Bonzini AioContext *ctx; 60c2b38b27SPaolo Bonzini QEMUBHFunc *cb; 61c2b38b27SPaolo Bonzini void *opaque; 628c6b0356SStefan Hajnoczi QSLIST_ENTRY(QEMUBH) next; 638c6b0356SStefan Hajnoczi unsigned flags; 64c2b38b27SPaolo Bonzini }; 65c2b38b27SPaolo Bonzini 668c6b0356SStefan Hajnoczi /* Called concurrently from any thread */ 678c6b0356SStefan Hajnoczi static void aio_bh_enqueue(QEMUBH *bh, unsigned new_flags) 688c6b0356SStefan Hajnoczi { 698c6b0356SStefan Hajnoczi AioContext *ctx = bh->ctx; 708c6b0356SStefan Hajnoczi unsigned old_flags; 718c6b0356SStefan Hajnoczi 728c6b0356SStefan Hajnoczi /* 73*d73415a3SStefan Hajnoczi * The memory barrier implicit in qatomic_fetch_or makes sure that: 748c6b0356SStefan Hajnoczi * 1. idle & any writes needed by the callback are done before the 758c6b0356SStefan Hajnoczi * locations are read in the aio_bh_poll. 768c6b0356SStefan Hajnoczi * 2. ctx is loaded before the callback has a chance to execute and bh 778c6b0356SStefan Hajnoczi * could be freed. 788c6b0356SStefan Hajnoczi */ 79*d73415a3SStefan Hajnoczi old_flags = qatomic_fetch_or(&bh->flags, BH_PENDING | new_flags); 808c6b0356SStefan Hajnoczi if (!(old_flags & BH_PENDING)) { 818c6b0356SStefan Hajnoczi QSLIST_INSERT_HEAD_ATOMIC(&ctx->bh_list, bh, next); 828c6b0356SStefan Hajnoczi } 838c6b0356SStefan Hajnoczi 848c6b0356SStefan Hajnoczi aio_notify(ctx); 858c6b0356SStefan Hajnoczi } 868c6b0356SStefan Hajnoczi 878c6b0356SStefan Hajnoczi /* Only called from aio_bh_poll() and aio_ctx_finalize() */ 888c6b0356SStefan Hajnoczi static QEMUBH *aio_bh_dequeue(BHList *head, unsigned *flags) 898c6b0356SStefan Hajnoczi { 908c6b0356SStefan Hajnoczi QEMUBH *bh = QSLIST_FIRST_RCU(head); 918c6b0356SStefan Hajnoczi 928c6b0356SStefan Hajnoczi if (!bh) { 938c6b0356SStefan Hajnoczi return NULL; 948c6b0356SStefan Hajnoczi } 958c6b0356SStefan Hajnoczi 968c6b0356SStefan Hajnoczi QSLIST_REMOVE_HEAD(head, next); 978c6b0356SStefan Hajnoczi 988c6b0356SStefan Hajnoczi /* 99*d73415a3SStefan Hajnoczi * The qatomic_and is paired with aio_bh_enqueue(). The implicit memory 1008c6b0356SStefan Hajnoczi * barrier ensures that the callback sees all writes done by the scheduling 1018c6b0356SStefan Hajnoczi * thread. It also ensures that the scheduling thread sees the cleared 1028c6b0356SStefan Hajnoczi * flag before bh->cb has run, and thus will call aio_notify again if 1038c6b0356SStefan Hajnoczi * necessary. 1048c6b0356SStefan Hajnoczi */ 105*d73415a3SStefan Hajnoczi *flags = qatomic_fetch_and(&bh->flags, 1068c6b0356SStefan Hajnoczi ~(BH_PENDING | BH_SCHEDULED | BH_IDLE)); 1078c6b0356SStefan Hajnoczi return bh; 1088c6b0356SStefan Hajnoczi } 1098c6b0356SStefan Hajnoczi 110c2b38b27SPaolo Bonzini void aio_bh_schedule_oneshot(AioContext *ctx, QEMUBHFunc *cb, void *opaque) 111c2b38b27SPaolo Bonzini { 112c2b38b27SPaolo Bonzini QEMUBH *bh; 113c2b38b27SPaolo Bonzini bh = g_new(QEMUBH, 1); 114c2b38b27SPaolo Bonzini *bh = (QEMUBH){ 115c2b38b27SPaolo Bonzini .ctx = ctx, 116c2b38b27SPaolo Bonzini .cb = cb, 117c2b38b27SPaolo Bonzini .opaque = opaque, 118c2b38b27SPaolo Bonzini }; 1198c6b0356SStefan Hajnoczi aio_bh_enqueue(bh, BH_SCHEDULED | BH_ONESHOT); 120c2b38b27SPaolo Bonzini } 121c2b38b27SPaolo Bonzini 122c2b38b27SPaolo Bonzini QEMUBH *aio_bh_new(AioContext *ctx, QEMUBHFunc *cb, void *opaque) 123c2b38b27SPaolo Bonzini { 124c2b38b27SPaolo Bonzini QEMUBH *bh; 125c2b38b27SPaolo Bonzini bh = g_new(QEMUBH, 1); 126c2b38b27SPaolo Bonzini *bh = (QEMUBH){ 127c2b38b27SPaolo Bonzini .ctx = ctx, 128c2b38b27SPaolo Bonzini .cb = cb, 129c2b38b27SPaolo Bonzini .opaque = opaque, 130c2b38b27SPaolo Bonzini }; 131c2b38b27SPaolo Bonzini return bh; 132c2b38b27SPaolo Bonzini } 133c2b38b27SPaolo Bonzini 134c2b38b27SPaolo Bonzini void aio_bh_call(QEMUBH *bh) 135c2b38b27SPaolo Bonzini { 136c2b38b27SPaolo Bonzini bh->cb(bh->opaque); 137c2b38b27SPaolo Bonzini } 138c2b38b27SPaolo Bonzini 1398c6b0356SStefan Hajnoczi /* Multiple occurrences of aio_bh_poll cannot be called concurrently. */ 140c2b38b27SPaolo Bonzini int aio_bh_poll(AioContext *ctx) 141c2b38b27SPaolo Bonzini { 1428c6b0356SStefan Hajnoczi BHListSlice slice; 1438c6b0356SStefan Hajnoczi BHListSlice *s; 1448c6b0356SStefan Hajnoczi int ret = 0; 145c2b38b27SPaolo Bonzini 1468c6b0356SStefan Hajnoczi QSLIST_MOVE_ATOMIC(&slice.bh_list, &ctx->bh_list); 1478c6b0356SStefan Hajnoczi QSIMPLEQ_INSERT_TAIL(&ctx->bh_slice_list, &slice, next); 1488c6b0356SStefan Hajnoczi 1498c6b0356SStefan Hajnoczi while ((s = QSIMPLEQ_FIRST(&ctx->bh_slice_list))) { 1508c6b0356SStefan Hajnoczi QEMUBH *bh; 1518c6b0356SStefan Hajnoczi unsigned flags; 1528c6b0356SStefan Hajnoczi 1538c6b0356SStefan Hajnoczi bh = aio_bh_dequeue(&s->bh_list, &flags); 1548c6b0356SStefan Hajnoczi if (!bh) { 1558c6b0356SStefan Hajnoczi QSIMPLEQ_REMOVE_HEAD(&ctx->bh_slice_list, next); 1568c6b0356SStefan Hajnoczi continue; 1578c6b0356SStefan Hajnoczi } 1588c6b0356SStefan Hajnoczi 1598c6b0356SStefan Hajnoczi if ((flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) { 160c2b38b27SPaolo Bonzini /* Idle BHs don't count as progress */ 1618c6b0356SStefan Hajnoczi if (!(flags & BH_IDLE)) { 162c2b38b27SPaolo Bonzini ret = 1; 163c2b38b27SPaolo Bonzini } 164c2b38b27SPaolo Bonzini aio_bh_call(bh); 165c2b38b27SPaolo Bonzini } 1668c6b0356SStefan Hajnoczi if (flags & (BH_DELETED | BH_ONESHOT)) { 167c2b38b27SPaolo Bonzini g_free(bh); 168c2b38b27SPaolo Bonzini } 169c2b38b27SPaolo Bonzini } 1708c6b0356SStefan Hajnoczi 171c2b38b27SPaolo Bonzini return ret; 172c2b38b27SPaolo Bonzini } 173c2b38b27SPaolo Bonzini 174c2b38b27SPaolo Bonzini void qemu_bh_schedule_idle(QEMUBH *bh) 175c2b38b27SPaolo Bonzini { 1768c6b0356SStefan Hajnoczi aio_bh_enqueue(bh, BH_SCHEDULED | BH_IDLE); 177c2b38b27SPaolo Bonzini } 178c2b38b27SPaolo Bonzini 179c2b38b27SPaolo Bonzini void qemu_bh_schedule(QEMUBH *bh) 180c2b38b27SPaolo Bonzini { 1818c6b0356SStefan Hajnoczi aio_bh_enqueue(bh, BH_SCHEDULED); 182c2b38b27SPaolo Bonzini } 183c2b38b27SPaolo Bonzini 184c2b38b27SPaolo Bonzini /* This func is async. 185c2b38b27SPaolo Bonzini */ 186c2b38b27SPaolo Bonzini void qemu_bh_cancel(QEMUBH *bh) 187c2b38b27SPaolo Bonzini { 188*d73415a3SStefan Hajnoczi qatomic_and(&bh->flags, ~BH_SCHEDULED); 189c2b38b27SPaolo Bonzini } 190c2b38b27SPaolo Bonzini 191c2b38b27SPaolo Bonzini /* This func is async.The bottom half will do the delete action at the finial 192c2b38b27SPaolo Bonzini * end. 193c2b38b27SPaolo Bonzini */ 194c2b38b27SPaolo Bonzini void qemu_bh_delete(QEMUBH *bh) 195c2b38b27SPaolo Bonzini { 1968c6b0356SStefan Hajnoczi aio_bh_enqueue(bh, BH_DELETED); 197c2b38b27SPaolo Bonzini } 198c2b38b27SPaolo Bonzini 1998c6b0356SStefan Hajnoczi static int64_t aio_compute_bh_timeout(BHList *head, int timeout) 200c2b38b27SPaolo Bonzini { 201c2b38b27SPaolo Bonzini QEMUBH *bh; 202c2b38b27SPaolo Bonzini 2038c6b0356SStefan Hajnoczi QSLIST_FOREACH_RCU(bh, head, next) { 2048c6b0356SStefan Hajnoczi if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) { 2058c6b0356SStefan Hajnoczi if (bh->flags & BH_IDLE) { 206c2b38b27SPaolo Bonzini /* idle bottom halves will be polled at least 207c2b38b27SPaolo Bonzini * every 10ms */ 208c2b38b27SPaolo Bonzini timeout = 10000000; 209c2b38b27SPaolo Bonzini } else { 210c2b38b27SPaolo Bonzini /* non-idle bottom halves will be executed 211c2b38b27SPaolo Bonzini * immediately */ 212c2b38b27SPaolo Bonzini return 0; 213c2b38b27SPaolo Bonzini } 214c2b38b27SPaolo Bonzini } 215c2b38b27SPaolo Bonzini } 216c2b38b27SPaolo Bonzini 2178c6b0356SStefan Hajnoczi return timeout; 2188c6b0356SStefan Hajnoczi } 2198c6b0356SStefan Hajnoczi 2208c6b0356SStefan Hajnoczi int64_t 2218c6b0356SStefan Hajnoczi aio_compute_timeout(AioContext *ctx) 2228c6b0356SStefan Hajnoczi { 2238c6b0356SStefan Hajnoczi BHListSlice *s; 2248c6b0356SStefan Hajnoczi int64_t deadline; 2258c6b0356SStefan Hajnoczi int timeout = -1; 2268c6b0356SStefan Hajnoczi 2278c6b0356SStefan Hajnoczi timeout = aio_compute_bh_timeout(&ctx->bh_list, timeout); 2288c6b0356SStefan Hajnoczi if (timeout == 0) { 2298c6b0356SStefan Hajnoczi return 0; 2308c6b0356SStefan Hajnoczi } 2318c6b0356SStefan Hajnoczi 2328c6b0356SStefan Hajnoczi QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) { 2338c6b0356SStefan Hajnoczi timeout = aio_compute_bh_timeout(&s->bh_list, timeout); 2348c6b0356SStefan Hajnoczi if (timeout == 0) { 2358c6b0356SStefan Hajnoczi return 0; 2368c6b0356SStefan Hajnoczi } 2378c6b0356SStefan Hajnoczi } 2388c6b0356SStefan Hajnoczi 239c2b38b27SPaolo Bonzini deadline = timerlistgroup_deadline_ns(&ctx->tlg); 240c2b38b27SPaolo Bonzini if (deadline == 0) { 241c2b38b27SPaolo Bonzini return 0; 242c2b38b27SPaolo Bonzini } else { 243c2b38b27SPaolo Bonzini return qemu_soonest_timeout(timeout, deadline); 244c2b38b27SPaolo Bonzini } 245c2b38b27SPaolo Bonzini } 246c2b38b27SPaolo Bonzini 247c2b38b27SPaolo Bonzini static gboolean 248c2b38b27SPaolo Bonzini aio_ctx_prepare(GSource *source, gint *timeout) 249c2b38b27SPaolo Bonzini { 250c2b38b27SPaolo Bonzini AioContext *ctx = (AioContext *) source; 251c2b38b27SPaolo Bonzini 252*d73415a3SStefan Hajnoczi qatomic_set(&ctx->notify_me, qatomic_read(&ctx->notify_me) | 1); 2535710a3e0SPaolo Bonzini 2545710a3e0SPaolo Bonzini /* 2555710a3e0SPaolo Bonzini * Write ctx->notify_me before computing the timeout 2565710a3e0SPaolo Bonzini * (reading bottom half flags, etc.). Pairs with 2575710a3e0SPaolo Bonzini * smp_mb in aio_notify(). 2585710a3e0SPaolo Bonzini */ 2595710a3e0SPaolo Bonzini smp_mb(); 260c2b38b27SPaolo Bonzini 261c2b38b27SPaolo Bonzini /* We assume there is no timeout already supplied */ 262c2b38b27SPaolo Bonzini *timeout = qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)); 263c2b38b27SPaolo Bonzini 264c2b38b27SPaolo Bonzini if (aio_prepare(ctx)) { 265c2b38b27SPaolo Bonzini *timeout = 0; 266c2b38b27SPaolo Bonzini } 267c2b38b27SPaolo Bonzini 268c2b38b27SPaolo Bonzini return *timeout == 0; 269c2b38b27SPaolo Bonzini } 270c2b38b27SPaolo Bonzini 271c2b38b27SPaolo Bonzini static gboolean 272c2b38b27SPaolo Bonzini aio_ctx_check(GSource *source) 273c2b38b27SPaolo Bonzini { 274c2b38b27SPaolo Bonzini AioContext *ctx = (AioContext *) source; 275c2b38b27SPaolo Bonzini QEMUBH *bh; 2768c6b0356SStefan Hajnoczi BHListSlice *s; 277c2b38b27SPaolo Bonzini 2785710a3e0SPaolo Bonzini /* Finish computing the timeout before clearing the flag. */ 279*d73415a3SStefan Hajnoczi qatomic_store_release(&ctx->notify_me, qatomic_read(&ctx->notify_me) & ~1); 280c2b38b27SPaolo Bonzini aio_notify_accept(ctx); 281c2b38b27SPaolo Bonzini 2828c6b0356SStefan Hajnoczi QSLIST_FOREACH_RCU(bh, &ctx->bh_list, next) { 2838c6b0356SStefan Hajnoczi if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) { 284c2b38b27SPaolo Bonzini return true; 285c2b38b27SPaolo Bonzini } 286c2b38b27SPaolo Bonzini } 2878c6b0356SStefan Hajnoczi 2888c6b0356SStefan Hajnoczi QSIMPLEQ_FOREACH(s, &ctx->bh_slice_list, next) { 2898c6b0356SStefan Hajnoczi QSLIST_FOREACH_RCU(bh, &s->bh_list, next) { 2908c6b0356SStefan Hajnoczi if ((bh->flags & (BH_SCHEDULED | BH_DELETED)) == BH_SCHEDULED) { 2918c6b0356SStefan Hajnoczi return true; 2928c6b0356SStefan Hajnoczi } 2938c6b0356SStefan Hajnoczi } 2948c6b0356SStefan Hajnoczi } 295c2b38b27SPaolo Bonzini return aio_pending(ctx) || (timerlistgroup_deadline_ns(&ctx->tlg) == 0); 296c2b38b27SPaolo Bonzini } 297c2b38b27SPaolo Bonzini 298c2b38b27SPaolo Bonzini static gboolean 299c2b38b27SPaolo Bonzini aio_ctx_dispatch(GSource *source, 300c2b38b27SPaolo Bonzini GSourceFunc callback, 301c2b38b27SPaolo Bonzini gpointer user_data) 302c2b38b27SPaolo Bonzini { 303c2b38b27SPaolo Bonzini AioContext *ctx = (AioContext *) source; 304c2b38b27SPaolo Bonzini 305c2b38b27SPaolo Bonzini assert(callback == NULL); 306a153bf52SPaolo Bonzini aio_dispatch(ctx); 307c2b38b27SPaolo Bonzini return true; 308c2b38b27SPaolo Bonzini } 309c2b38b27SPaolo Bonzini 310c2b38b27SPaolo Bonzini static void 311c2b38b27SPaolo Bonzini aio_ctx_finalize(GSource *source) 312c2b38b27SPaolo Bonzini { 313c2b38b27SPaolo Bonzini AioContext *ctx = (AioContext *) source; 3148c6b0356SStefan Hajnoczi QEMUBH *bh; 3158c6b0356SStefan Hajnoczi unsigned flags; 316c2b38b27SPaolo Bonzini 317c2b38b27SPaolo Bonzini thread_pool_free(ctx->thread_pool); 318c2b38b27SPaolo Bonzini 319c2b38b27SPaolo Bonzini #ifdef CONFIG_LINUX_AIO 320c2b38b27SPaolo Bonzini if (ctx->linux_aio) { 321c2b38b27SPaolo Bonzini laio_detach_aio_context(ctx->linux_aio, ctx); 322c2b38b27SPaolo Bonzini laio_cleanup(ctx->linux_aio); 323c2b38b27SPaolo Bonzini ctx->linux_aio = NULL; 324c2b38b27SPaolo Bonzini } 325c2b38b27SPaolo Bonzini #endif 326c2b38b27SPaolo Bonzini 327fcb7a4a4SAarushi Mehta #ifdef CONFIG_LINUX_IO_URING 328fcb7a4a4SAarushi Mehta if (ctx->linux_io_uring) { 329fcb7a4a4SAarushi Mehta luring_detach_aio_context(ctx->linux_io_uring, ctx); 330fcb7a4a4SAarushi Mehta luring_cleanup(ctx->linux_io_uring); 331fcb7a4a4SAarushi Mehta ctx->linux_io_uring = NULL; 332fcb7a4a4SAarushi Mehta } 333fcb7a4a4SAarushi Mehta #endif 334fcb7a4a4SAarushi Mehta 3350c330a73SPaolo Bonzini assert(QSLIST_EMPTY(&ctx->scheduled_coroutines)); 3360c330a73SPaolo Bonzini qemu_bh_delete(ctx->co_schedule_bh); 3370c330a73SPaolo Bonzini 3388c6b0356SStefan Hajnoczi /* There must be no aio_bh_poll() calls going on */ 3398c6b0356SStefan Hajnoczi assert(QSIMPLEQ_EMPTY(&ctx->bh_slice_list)); 340c2b38b27SPaolo Bonzini 3418c6b0356SStefan Hajnoczi while ((bh = aio_bh_dequeue(&ctx->bh_list, &flags))) { 342c2b38b27SPaolo Bonzini /* qemu_bh_delete() must have been called on BHs in this AioContext */ 3438c6b0356SStefan Hajnoczi assert(flags & BH_DELETED); 344c2b38b27SPaolo Bonzini 3458c6b0356SStefan Hajnoczi g_free(bh); 346c2b38b27SPaolo Bonzini } 347c2b38b27SPaolo Bonzini 348c2b38b27SPaolo Bonzini aio_set_event_notifier(ctx, &ctx->notifier, false, NULL, NULL); 349c2b38b27SPaolo Bonzini event_notifier_cleanup(&ctx->notifier); 350c2b38b27SPaolo Bonzini qemu_rec_mutex_destroy(&ctx->lock); 351c2b38b27SPaolo Bonzini qemu_lockcnt_destroy(&ctx->list_lock); 352c2b38b27SPaolo Bonzini timerlistgroup_deinit(&ctx->tlg); 353cd0a6d2bSJie Wang aio_context_destroy(ctx); 354c2b38b27SPaolo Bonzini } 355c2b38b27SPaolo Bonzini 356c2b38b27SPaolo Bonzini static GSourceFuncs aio_source_funcs = { 357c2b38b27SPaolo Bonzini aio_ctx_prepare, 358c2b38b27SPaolo Bonzini aio_ctx_check, 359c2b38b27SPaolo Bonzini aio_ctx_dispatch, 360c2b38b27SPaolo Bonzini aio_ctx_finalize 361c2b38b27SPaolo Bonzini }; 362c2b38b27SPaolo Bonzini 363c2b38b27SPaolo Bonzini GSource *aio_get_g_source(AioContext *ctx) 364c2b38b27SPaolo Bonzini { 365ba607ca8SStefan Hajnoczi aio_context_use_g_source(ctx); 366c2b38b27SPaolo Bonzini g_source_ref(&ctx->source); 367c2b38b27SPaolo Bonzini return &ctx->source; 368c2b38b27SPaolo Bonzini } 369c2b38b27SPaolo Bonzini 370c2b38b27SPaolo Bonzini ThreadPool *aio_get_thread_pool(AioContext *ctx) 371c2b38b27SPaolo Bonzini { 372c2b38b27SPaolo Bonzini if (!ctx->thread_pool) { 373c2b38b27SPaolo Bonzini ctx->thread_pool = thread_pool_new(ctx); 374c2b38b27SPaolo Bonzini } 375c2b38b27SPaolo Bonzini return ctx->thread_pool; 376c2b38b27SPaolo Bonzini } 377c2b38b27SPaolo Bonzini 378c2b38b27SPaolo Bonzini #ifdef CONFIG_LINUX_AIO 379ed6e2161SNishanth Aravamudan LinuxAioState *aio_setup_linux_aio(AioContext *ctx, Error **errp) 380c2b38b27SPaolo Bonzini { 381c2b38b27SPaolo Bonzini if (!ctx->linux_aio) { 382ed6e2161SNishanth Aravamudan ctx->linux_aio = laio_init(errp); 383ed6e2161SNishanth Aravamudan if (ctx->linux_aio) { 384c2b38b27SPaolo Bonzini laio_attach_aio_context(ctx->linux_aio, ctx); 385c2b38b27SPaolo Bonzini } 386ed6e2161SNishanth Aravamudan } 387ed6e2161SNishanth Aravamudan return ctx->linux_aio; 388ed6e2161SNishanth Aravamudan } 389ed6e2161SNishanth Aravamudan 390ed6e2161SNishanth Aravamudan LinuxAioState *aio_get_linux_aio(AioContext *ctx) 391ed6e2161SNishanth Aravamudan { 392ed6e2161SNishanth Aravamudan assert(ctx->linux_aio); 393c2b38b27SPaolo Bonzini return ctx->linux_aio; 394c2b38b27SPaolo Bonzini } 395c2b38b27SPaolo Bonzini #endif 396c2b38b27SPaolo Bonzini 397fcb7a4a4SAarushi Mehta #ifdef CONFIG_LINUX_IO_URING 398fcb7a4a4SAarushi Mehta LuringState *aio_setup_linux_io_uring(AioContext *ctx, Error **errp) 399fcb7a4a4SAarushi Mehta { 400fcb7a4a4SAarushi Mehta if (ctx->linux_io_uring) { 401fcb7a4a4SAarushi Mehta return ctx->linux_io_uring; 402fcb7a4a4SAarushi Mehta } 403fcb7a4a4SAarushi Mehta 404fcb7a4a4SAarushi Mehta ctx->linux_io_uring = luring_init(errp); 405fcb7a4a4SAarushi Mehta if (!ctx->linux_io_uring) { 406fcb7a4a4SAarushi Mehta return NULL; 407fcb7a4a4SAarushi Mehta } 408fcb7a4a4SAarushi Mehta 409fcb7a4a4SAarushi Mehta luring_attach_aio_context(ctx->linux_io_uring, ctx); 410fcb7a4a4SAarushi Mehta return ctx->linux_io_uring; 411fcb7a4a4SAarushi Mehta } 412fcb7a4a4SAarushi Mehta 413fcb7a4a4SAarushi Mehta LuringState *aio_get_linux_io_uring(AioContext *ctx) 414fcb7a4a4SAarushi Mehta { 415fcb7a4a4SAarushi Mehta assert(ctx->linux_io_uring); 416fcb7a4a4SAarushi Mehta return ctx->linux_io_uring; 417fcb7a4a4SAarushi Mehta } 418fcb7a4a4SAarushi Mehta #endif 419fcb7a4a4SAarushi Mehta 420c2b38b27SPaolo Bonzini void aio_notify(AioContext *ctx) 421c2b38b27SPaolo Bonzini { 422601829f8SStefan Hajnoczi /* 423601829f8SStefan Hajnoczi * Write e.g. bh->flags before writing ctx->notified. Pairs with smp_mb in 424601829f8SStefan Hajnoczi * aio_notify_accept. 425601829f8SStefan Hajnoczi */ 426601829f8SStefan Hajnoczi smp_wmb(); 427*d73415a3SStefan Hajnoczi qatomic_set(&ctx->notified, true); 428601829f8SStefan Hajnoczi 429601829f8SStefan Hajnoczi /* 430601829f8SStefan Hajnoczi * Write ctx->notified before reading ctx->notify_me. Pairs 4315710a3e0SPaolo Bonzini * with smp_mb in aio_ctx_prepare or aio_poll. 432c2b38b27SPaolo Bonzini */ 433c2b38b27SPaolo Bonzini smp_mb(); 434*d73415a3SStefan Hajnoczi if (qatomic_read(&ctx->notify_me)) { 435c2b38b27SPaolo Bonzini event_notifier_set(&ctx->notifier); 436c2b38b27SPaolo Bonzini } 437c2b38b27SPaolo Bonzini } 438c2b38b27SPaolo Bonzini 439c2b38b27SPaolo Bonzini void aio_notify_accept(AioContext *ctx) 440c2b38b27SPaolo Bonzini { 441*d73415a3SStefan Hajnoczi qatomic_set(&ctx->notified, false); 442601829f8SStefan Hajnoczi 443601829f8SStefan Hajnoczi /* 444601829f8SStefan Hajnoczi * Write ctx->notified before reading e.g. bh->flags. Pairs with smp_wmb 445601829f8SStefan Hajnoczi * in aio_notify. 446601829f8SStefan Hajnoczi */ 447601829f8SStefan Hajnoczi smp_mb(); 448c2b38b27SPaolo Bonzini } 449c2b38b27SPaolo Bonzini 4503f53bc61SPaolo Bonzini static void aio_timerlist_notify(void *opaque, QEMUClockType type) 451c2b38b27SPaolo Bonzini { 452c2b38b27SPaolo Bonzini aio_notify(opaque); 453c2b38b27SPaolo Bonzini } 454c2b38b27SPaolo Bonzini 455601829f8SStefan Hajnoczi static void aio_context_notifier_cb(EventNotifier *e) 456c2b38b27SPaolo Bonzini { 457601829f8SStefan Hajnoczi AioContext *ctx = container_of(e, AioContext, notifier); 458601829f8SStefan Hajnoczi 459601829f8SStefan Hajnoczi event_notifier_test_and_clear(&ctx->notifier); 460c2b38b27SPaolo Bonzini } 461c2b38b27SPaolo Bonzini 462c2b38b27SPaolo Bonzini /* Returns true if aio_notify() was called (e.g. a BH was scheduled) */ 463c13be5a1SStefan Hajnoczi static bool aio_context_notifier_poll(void *opaque) 464c2b38b27SPaolo Bonzini { 465c2b38b27SPaolo Bonzini EventNotifier *e = opaque; 466c2b38b27SPaolo Bonzini AioContext *ctx = container_of(e, AioContext, notifier); 467c2b38b27SPaolo Bonzini 468*d73415a3SStefan Hajnoczi return qatomic_read(&ctx->notified); 469c2b38b27SPaolo Bonzini } 470c2b38b27SPaolo Bonzini 4710c330a73SPaolo Bonzini static void co_schedule_bh_cb(void *opaque) 4720c330a73SPaolo Bonzini { 4730c330a73SPaolo Bonzini AioContext *ctx = opaque; 4740c330a73SPaolo Bonzini QSLIST_HEAD(, Coroutine) straight, reversed; 4750c330a73SPaolo Bonzini 4760c330a73SPaolo Bonzini QSLIST_MOVE_ATOMIC(&reversed, &ctx->scheduled_coroutines); 4770c330a73SPaolo Bonzini QSLIST_INIT(&straight); 4780c330a73SPaolo Bonzini 4790c330a73SPaolo Bonzini while (!QSLIST_EMPTY(&reversed)) { 4800c330a73SPaolo Bonzini Coroutine *co = QSLIST_FIRST(&reversed); 4810c330a73SPaolo Bonzini QSLIST_REMOVE_HEAD(&reversed, co_scheduled_next); 4820c330a73SPaolo Bonzini QSLIST_INSERT_HEAD(&straight, co, co_scheduled_next); 4830c330a73SPaolo Bonzini } 4840c330a73SPaolo Bonzini 4850c330a73SPaolo Bonzini while (!QSLIST_EMPTY(&straight)) { 4860c330a73SPaolo Bonzini Coroutine *co = QSLIST_FIRST(&straight); 4870c330a73SPaolo Bonzini QSLIST_REMOVE_HEAD(&straight, co_scheduled_next); 4880c330a73SPaolo Bonzini trace_aio_co_schedule_bh_cb(ctx, co); 4891919631eSPaolo Bonzini aio_context_acquire(ctx); 4906133b39fSJeff Cody 4916133b39fSJeff Cody /* Protected by write barrier in qemu_aio_coroutine_enter */ 492*d73415a3SStefan Hajnoczi qatomic_set(&co->scheduled, NULL); 4936808ae04SSergio Lopez qemu_aio_coroutine_enter(ctx, co); 4941919631eSPaolo Bonzini aio_context_release(ctx); 4950c330a73SPaolo Bonzini } 4960c330a73SPaolo Bonzini } 4970c330a73SPaolo Bonzini 498c2b38b27SPaolo Bonzini AioContext *aio_context_new(Error **errp) 499c2b38b27SPaolo Bonzini { 500c2b38b27SPaolo Bonzini int ret; 501c2b38b27SPaolo Bonzini AioContext *ctx; 502c2b38b27SPaolo Bonzini 503c2b38b27SPaolo Bonzini ctx = (AioContext *) g_source_new(&aio_source_funcs, sizeof(AioContext)); 5048c6b0356SStefan Hajnoczi QSLIST_INIT(&ctx->bh_list); 5058c6b0356SStefan Hajnoczi QSIMPLEQ_INIT(&ctx->bh_slice_list); 506c2b38b27SPaolo Bonzini aio_context_setup(ctx); 507c2b38b27SPaolo Bonzini 508c2b38b27SPaolo Bonzini ret = event_notifier_init(&ctx->notifier, false); 509c2b38b27SPaolo Bonzini if (ret < 0) { 510c2b38b27SPaolo Bonzini error_setg_errno(errp, -ret, "Failed to initialize event notifier"); 511c2b38b27SPaolo Bonzini goto fail; 512c2b38b27SPaolo Bonzini } 513c2b38b27SPaolo Bonzini g_source_set_can_recurse(&ctx->source, true); 514c2b38b27SPaolo Bonzini qemu_lockcnt_init(&ctx->list_lock); 5150c330a73SPaolo Bonzini 5160c330a73SPaolo Bonzini ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx); 5170c330a73SPaolo Bonzini QSLIST_INIT(&ctx->scheduled_coroutines); 5180c330a73SPaolo Bonzini 519c2b38b27SPaolo Bonzini aio_set_event_notifier(ctx, &ctx->notifier, 520c2b38b27SPaolo Bonzini false, 521601829f8SStefan Hajnoczi aio_context_notifier_cb, 522c13be5a1SStefan Hajnoczi aio_context_notifier_poll); 523c2b38b27SPaolo Bonzini #ifdef CONFIG_LINUX_AIO 524c2b38b27SPaolo Bonzini ctx->linux_aio = NULL; 525c2b38b27SPaolo Bonzini #endif 526fcb7a4a4SAarushi Mehta 527fcb7a4a4SAarushi Mehta #ifdef CONFIG_LINUX_IO_URING 528fcb7a4a4SAarushi Mehta ctx->linux_io_uring = NULL; 529fcb7a4a4SAarushi Mehta #endif 530fcb7a4a4SAarushi Mehta 531c2b38b27SPaolo Bonzini ctx->thread_pool = NULL; 532c2b38b27SPaolo Bonzini qemu_rec_mutex_init(&ctx->lock); 533c2b38b27SPaolo Bonzini timerlistgroup_init(&ctx->tlg, aio_timerlist_notify, ctx); 534c2b38b27SPaolo Bonzini 535c2b38b27SPaolo Bonzini ctx->poll_ns = 0; 536c2b38b27SPaolo Bonzini ctx->poll_max_ns = 0; 537c2b38b27SPaolo Bonzini ctx->poll_grow = 0; 538c2b38b27SPaolo Bonzini ctx->poll_shrink = 0; 539c2b38b27SPaolo Bonzini 540c2b38b27SPaolo Bonzini return ctx; 541c2b38b27SPaolo Bonzini fail: 542c2b38b27SPaolo Bonzini g_source_destroy(&ctx->source); 543c2b38b27SPaolo Bonzini return NULL; 544c2b38b27SPaolo Bonzini } 545c2b38b27SPaolo Bonzini 5460c330a73SPaolo Bonzini void aio_co_schedule(AioContext *ctx, Coroutine *co) 5470c330a73SPaolo Bonzini { 5480c330a73SPaolo Bonzini trace_aio_co_schedule(ctx, co); 549*d73415a3SStefan Hajnoczi const char *scheduled = qatomic_cmpxchg(&co->scheduled, NULL, 5506133b39fSJeff Cody __func__); 5516133b39fSJeff Cody 5526133b39fSJeff Cody if (scheduled) { 5536133b39fSJeff Cody fprintf(stderr, 5546133b39fSJeff Cody "%s: Co-routine was already scheduled in '%s'\n", 5556133b39fSJeff Cody __func__, scheduled); 5566133b39fSJeff Cody abort(); 5576133b39fSJeff Cody } 5586133b39fSJeff Cody 559f0f81002SStefan Hajnoczi /* The coroutine might run and release the last ctx reference before we 560f0f81002SStefan Hajnoczi * invoke qemu_bh_schedule(). Take a reference to keep ctx alive until 561f0f81002SStefan Hajnoczi * we're done. 562f0f81002SStefan Hajnoczi */ 563f0f81002SStefan Hajnoczi aio_context_ref(ctx); 564f0f81002SStefan Hajnoczi 5650c330a73SPaolo Bonzini QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines, 5660c330a73SPaolo Bonzini co, co_scheduled_next); 5670c330a73SPaolo Bonzini qemu_bh_schedule(ctx->co_schedule_bh); 568f0f81002SStefan Hajnoczi 569f0f81002SStefan Hajnoczi aio_context_unref(ctx); 5700c330a73SPaolo Bonzini } 5710c330a73SPaolo Bonzini 5720c330a73SPaolo Bonzini void aio_co_wake(struct Coroutine *co) 5730c330a73SPaolo Bonzini { 5740c330a73SPaolo Bonzini AioContext *ctx; 5750c330a73SPaolo Bonzini 5760c330a73SPaolo Bonzini /* Read coroutine before co->ctx. Matches smp_wmb in 5770c330a73SPaolo Bonzini * qemu_coroutine_enter. 5780c330a73SPaolo Bonzini */ 5790c330a73SPaolo Bonzini smp_read_barrier_depends(); 580*d73415a3SStefan Hajnoczi ctx = qatomic_read(&co->ctx); 5810c330a73SPaolo Bonzini 5828865852eSFam Zheng aio_co_enter(ctx, co); 5838865852eSFam Zheng } 5848865852eSFam Zheng 5858865852eSFam Zheng void aio_co_enter(AioContext *ctx, struct Coroutine *co) 5868865852eSFam Zheng { 5870c330a73SPaolo Bonzini if (ctx != qemu_get_current_aio_context()) { 5880c330a73SPaolo Bonzini aio_co_schedule(ctx, co); 5890c330a73SPaolo Bonzini return; 5900c330a73SPaolo Bonzini } 5910c330a73SPaolo Bonzini 5920c330a73SPaolo Bonzini if (qemu_in_coroutine()) { 5930c330a73SPaolo Bonzini Coroutine *self = qemu_coroutine_self(); 5940c330a73SPaolo Bonzini assert(self != co); 5950c330a73SPaolo Bonzini QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, co, co_queue_next); 5960c330a73SPaolo Bonzini } else { 5970c330a73SPaolo Bonzini aio_context_acquire(ctx); 5988865852eSFam Zheng qemu_aio_coroutine_enter(ctx, co); 5990c330a73SPaolo Bonzini aio_context_release(ctx); 6000c330a73SPaolo Bonzini } 6010c330a73SPaolo Bonzini } 6020c330a73SPaolo Bonzini 603c2b38b27SPaolo Bonzini void aio_context_ref(AioContext *ctx) 604c2b38b27SPaolo Bonzini { 605c2b38b27SPaolo Bonzini g_source_ref(&ctx->source); 606c2b38b27SPaolo Bonzini } 607c2b38b27SPaolo Bonzini 608c2b38b27SPaolo Bonzini void aio_context_unref(AioContext *ctx) 609c2b38b27SPaolo Bonzini { 610c2b38b27SPaolo Bonzini g_source_unref(&ctx->source); 611c2b38b27SPaolo Bonzini } 612c2b38b27SPaolo Bonzini 613c2b38b27SPaolo Bonzini void aio_context_acquire(AioContext *ctx) 614c2b38b27SPaolo Bonzini { 615c2b38b27SPaolo Bonzini qemu_rec_mutex_lock(&ctx->lock); 616c2b38b27SPaolo Bonzini } 617c2b38b27SPaolo Bonzini 618c2b38b27SPaolo Bonzini void aio_context_release(AioContext *ctx) 619c2b38b27SPaolo Bonzini { 620c2b38b27SPaolo Bonzini qemu_rec_mutex_unlock(&ctx->lock); 621c2b38b27SPaolo Bonzini } 622