1 /* 2 * Linux native AIO support. 3 * 4 * Copyright (C) 2009 IBM, Corp. 5 * Copyright (C) 2009 Red Hat, Inc. 6 * 7 * This work is licensed under the terms of the GNU GPL, version 2 or later. 8 * See the COPYING file in the top-level directory. 9 */ 10 #include "qemu/osdep.h" 11 #include "block/aio.h" 12 #include "qemu/queue.h" 13 #include "block/block.h" 14 #include "block/raw-aio.h" 15 #include "qemu/event_notifier.h" 16 #include "qemu/coroutine.h" 17 #include "qapi/error.h" 18 19 #include <libaio.h> 20 21 /* 22 * Queue size (per-device). 23 * 24 * XXX: eventually we need to communicate this to the guest and/or make it 25 * tunable by the guest. If we get more outstanding requests at a time 26 * than this we will get EAGAIN from io_submit which is communicated to 27 * the guest as an I/O error. 28 */ 29 #define MAX_EVENTS 1024 30 31 /* Maximum number of requests in a batch. (default value) */ 32 #define DEFAULT_MAX_BATCH 32 33 34 struct qemu_laiocb { 35 Coroutine *co; 36 LinuxAioState *ctx; 37 struct iocb iocb; 38 ssize_t ret; 39 size_t nbytes; 40 QEMUIOVector *qiov; 41 bool is_read; 42 QSIMPLEQ_ENTRY(qemu_laiocb) next; 43 }; 44 45 typedef struct { 46 int plugged; 47 unsigned int in_queue; 48 unsigned int in_flight; 49 bool blocked; 50 QSIMPLEQ_HEAD(, qemu_laiocb) pending; 51 } LaioQueue; 52 53 struct LinuxAioState { 54 AioContext *aio_context; 55 56 io_context_t ctx; 57 EventNotifier e; 58 59 /* io queue for submit at batch. Protected by AioContext lock. */ 60 LaioQueue io_q; 61 62 /* I/O completion processing. Only runs in I/O thread. */ 63 QEMUBH *completion_bh; 64 int event_idx; 65 int event_max; 66 }; 67 68 static void ioq_submit(LinuxAioState *s); 69 70 static inline ssize_t io_event_ret(struct io_event *ev) 71 { 72 return (ssize_t)(((uint64_t)ev->res2 << 32) | ev->res); 73 } 74 75 /* 76 * Completes an AIO request. 77 */ 78 static void qemu_laio_process_completion(struct qemu_laiocb *laiocb) 79 { 80 int ret; 81 82 ret = laiocb->ret; 83 if (ret != -ECANCELED) { 84 if (ret == laiocb->nbytes) { 85 ret = 0; 86 } else if (ret >= 0) { 87 /* Short reads mean EOF, pad with zeros. */ 88 if (laiocb->is_read) { 89 qemu_iovec_memset(laiocb->qiov, ret, 0, 90 laiocb->qiov->size - ret); 91 } else { 92 ret = -ENOSPC; 93 } 94 } 95 } 96 97 laiocb->ret = ret; 98 99 /* 100 * If the coroutine is already entered it must be in ioq_submit() and 101 * will notice laio->ret has been filled in when it eventually runs 102 * later. Coroutines cannot be entered recursively so avoid doing 103 * that! 104 */ 105 if (!qemu_coroutine_entered(laiocb->co)) { 106 aio_co_wake(laiocb->co); 107 } 108 } 109 110 /** 111 * aio_ring buffer which is shared between userspace and kernel. 112 * 113 * This copied from linux/fs/aio.c, common header does not exist 114 * but AIO exists for ages so we assume ABI is stable. 115 */ 116 struct aio_ring { 117 unsigned id; /* kernel internal index number */ 118 unsigned nr; /* number of io_events */ 119 unsigned head; /* Written to by userland or by kernel. */ 120 unsigned tail; 121 122 unsigned magic; 123 unsigned compat_features; 124 unsigned incompat_features; 125 unsigned header_length; /* size of aio_ring */ 126 127 struct io_event io_events[]; 128 }; 129 130 /** 131 * io_getevents_peek: 132 * @ctx: AIO context 133 * @events: pointer on events array, output value 134 135 * Returns the number of completed events and sets a pointer 136 * on events array. This function does not update the internal 137 * ring buffer, only reads head and tail. When @events has been 138 * processed io_getevents_commit() must be called. 139 */ 140 static inline unsigned int io_getevents_peek(io_context_t ctx, 141 struct io_event **events) 142 { 143 struct aio_ring *ring = (struct aio_ring *)ctx; 144 unsigned int head = ring->head, tail = ring->tail; 145 unsigned int nr; 146 147 nr = tail >= head ? tail - head : ring->nr - head; 148 *events = ring->io_events + head; 149 /* To avoid speculative loads of s->events[i] before observing tail. 150 Paired with smp_wmb() inside linux/fs/aio.c: aio_complete(). */ 151 smp_rmb(); 152 153 return nr; 154 } 155 156 /** 157 * io_getevents_commit: 158 * @ctx: AIO context 159 * @nr: the number of events on which head should be advanced 160 * 161 * Advances head of a ring buffer. 162 */ 163 static inline void io_getevents_commit(io_context_t ctx, unsigned int nr) 164 { 165 struct aio_ring *ring = (struct aio_ring *)ctx; 166 167 if (nr) { 168 ring->head = (ring->head + nr) % ring->nr; 169 } 170 } 171 172 /** 173 * io_getevents_advance_and_peek: 174 * @ctx: AIO context 175 * @events: pointer on events array, output value 176 * @nr: the number of events on which head should be advanced 177 * 178 * Advances head of a ring buffer and returns number of elements left. 179 */ 180 static inline unsigned int 181 io_getevents_advance_and_peek(io_context_t ctx, 182 struct io_event **events, 183 unsigned int nr) 184 { 185 io_getevents_commit(ctx, nr); 186 return io_getevents_peek(ctx, events); 187 } 188 189 /** 190 * qemu_laio_process_completions: 191 * @s: AIO state 192 * 193 * Fetches completed I/O requests and invokes their callbacks. 194 * 195 * The function is somewhat tricky because it supports nested event loops, for 196 * example when a request callback invokes aio_poll(). In order to do this, 197 * indices are kept in LinuxAioState. Function schedules BH completion so it 198 * can be called again in a nested event loop. When there are no events left 199 * to complete the BH is being canceled. 200 */ 201 static void qemu_laio_process_completions(LinuxAioState *s) 202 { 203 struct io_event *events; 204 205 /* Reschedule so nested event loops see currently pending completions */ 206 qemu_bh_schedule(s->completion_bh); 207 208 while ((s->event_max = io_getevents_advance_and_peek(s->ctx, &events, 209 s->event_idx))) { 210 for (s->event_idx = 0; s->event_idx < s->event_max; ) { 211 struct iocb *iocb = events[s->event_idx].obj; 212 struct qemu_laiocb *laiocb = 213 container_of(iocb, struct qemu_laiocb, iocb); 214 215 laiocb->ret = io_event_ret(&events[s->event_idx]); 216 217 /* Change counters one-by-one because we can be nested. */ 218 s->io_q.in_flight--; 219 s->event_idx++; 220 qemu_laio_process_completion(laiocb); 221 } 222 } 223 224 qemu_bh_cancel(s->completion_bh); 225 226 /* If we are nested we have to notify the level above that we are done 227 * by setting event_max to zero, upper level will then jump out of it's 228 * own `for` loop. If we are the last all counters droped to zero. */ 229 s->event_max = 0; 230 s->event_idx = 0; 231 } 232 233 static void qemu_laio_process_completions_and_submit(LinuxAioState *s) 234 { 235 aio_context_acquire(s->aio_context); 236 qemu_laio_process_completions(s); 237 238 if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) { 239 ioq_submit(s); 240 } 241 aio_context_release(s->aio_context); 242 } 243 244 static void qemu_laio_completion_bh(void *opaque) 245 { 246 LinuxAioState *s = opaque; 247 248 qemu_laio_process_completions_and_submit(s); 249 } 250 251 static void qemu_laio_completion_cb(EventNotifier *e) 252 { 253 LinuxAioState *s = container_of(e, LinuxAioState, e); 254 255 if (event_notifier_test_and_clear(&s->e)) { 256 qemu_laio_process_completions_and_submit(s); 257 } 258 } 259 260 static bool qemu_laio_poll_cb(void *opaque) 261 { 262 EventNotifier *e = opaque; 263 LinuxAioState *s = container_of(e, LinuxAioState, e); 264 struct io_event *events; 265 266 return io_getevents_peek(s->ctx, &events); 267 } 268 269 static void qemu_laio_poll_ready(EventNotifier *opaque) 270 { 271 EventNotifier *e = opaque; 272 LinuxAioState *s = container_of(e, LinuxAioState, e); 273 274 qemu_laio_process_completions_and_submit(s); 275 } 276 277 static void ioq_init(LaioQueue *io_q) 278 { 279 QSIMPLEQ_INIT(&io_q->pending); 280 io_q->plugged = 0; 281 io_q->in_queue = 0; 282 io_q->in_flight = 0; 283 io_q->blocked = false; 284 } 285 286 static void ioq_submit(LinuxAioState *s) 287 { 288 int ret, len; 289 struct qemu_laiocb *aiocb; 290 struct iocb *iocbs[MAX_EVENTS]; 291 QSIMPLEQ_HEAD(, qemu_laiocb) completed; 292 293 do { 294 if (s->io_q.in_flight >= MAX_EVENTS) { 295 break; 296 } 297 len = 0; 298 QSIMPLEQ_FOREACH(aiocb, &s->io_q.pending, next) { 299 iocbs[len++] = &aiocb->iocb; 300 if (s->io_q.in_flight + len >= MAX_EVENTS) { 301 break; 302 } 303 } 304 305 ret = io_submit(s->ctx, len, iocbs); 306 if (ret == -EAGAIN) { 307 break; 308 } 309 if (ret < 0) { 310 /* Fail the first request, retry the rest */ 311 aiocb = QSIMPLEQ_FIRST(&s->io_q.pending); 312 QSIMPLEQ_REMOVE_HEAD(&s->io_q.pending, next); 313 s->io_q.in_queue--; 314 aiocb->ret = ret; 315 qemu_laio_process_completion(aiocb); 316 continue; 317 } 318 319 s->io_q.in_flight += ret; 320 s->io_q.in_queue -= ret; 321 aiocb = container_of(iocbs[ret - 1], struct qemu_laiocb, iocb); 322 QSIMPLEQ_SPLIT_AFTER(&s->io_q.pending, aiocb, next, &completed); 323 } while (ret == len && !QSIMPLEQ_EMPTY(&s->io_q.pending)); 324 s->io_q.blocked = (s->io_q.in_queue > 0); 325 326 if (s->io_q.in_flight) { 327 /* We can try to complete something just right away if there are 328 * still requests in-flight. */ 329 qemu_laio_process_completions(s); 330 /* 331 * Even we have completed everything (in_flight == 0), the queue can 332 * have still pended requests (in_queue > 0). We do not attempt to 333 * repeat submission to avoid IO hang. The reason is simple: s->e is 334 * still set and completion callback will be called shortly and all 335 * pended requests will be submitted from there. 336 */ 337 } 338 } 339 340 static uint64_t laio_max_batch(LinuxAioState *s, uint64_t dev_max_batch) 341 { 342 uint64_t max_batch = s->aio_context->aio_max_batch ?: DEFAULT_MAX_BATCH; 343 344 /* 345 * AIO context can be shared between multiple block devices, so 346 * `dev_max_batch` allows reducing the batch size for latency-sensitive 347 * devices. 348 */ 349 max_batch = MIN_NON_ZERO(dev_max_batch, max_batch); 350 351 /* limit the batch with the number of available events */ 352 max_batch = MIN_NON_ZERO(MAX_EVENTS - s->io_q.in_flight, max_batch); 353 354 return max_batch; 355 } 356 357 void laio_io_plug(BlockDriverState *bs, LinuxAioState *s) 358 { 359 s->io_q.plugged++; 360 } 361 362 void laio_io_unplug(BlockDriverState *bs, LinuxAioState *s, 363 uint64_t dev_max_batch) 364 { 365 assert(s->io_q.plugged); 366 s->io_q.plugged--; 367 368 /* 369 * Why max batch checking is performed here: 370 * Another BDS may have queued requests with a higher dev_max_batch and 371 * therefore in_queue could now exceed our dev_max_batch. Re-check the max 372 * batch so we can honor our device's dev_max_batch. 373 */ 374 if (s->io_q.in_queue >= laio_max_batch(s, dev_max_batch) || 375 (!s->io_q.plugged && 376 !s->io_q.blocked && !QSIMPLEQ_EMPTY(&s->io_q.pending))) { 377 ioq_submit(s); 378 } 379 } 380 381 static int laio_do_submit(int fd, struct qemu_laiocb *laiocb, off_t offset, 382 int type, uint64_t dev_max_batch) 383 { 384 LinuxAioState *s = laiocb->ctx; 385 struct iocb *iocbs = &laiocb->iocb; 386 QEMUIOVector *qiov = laiocb->qiov; 387 388 switch (type) { 389 case QEMU_AIO_WRITE: 390 io_prep_pwritev(iocbs, fd, qiov->iov, qiov->niov, offset); 391 break; 392 case QEMU_AIO_READ: 393 io_prep_preadv(iocbs, fd, qiov->iov, qiov->niov, offset); 394 break; 395 /* Currently Linux kernel does not support other operations */ 396 default: 397 fprintf(stderr, "%s: invalid AIO request type 0x%x.\n", 398 __func__, type); 399 return -EIO; 400 } 401 io_set_eventfd(&laiocb->iocb, event_notifier_get_fd(&s->e)); 402 403 QSIMPLEQ_INSERT_TAIL(&s->io_q.pending, laiocb, next); 404 s->io_q.in_queue++; 405 if (!s->io_q.blocked && 406 (!s->io_q.plugged || 407 s->io_q.in_queue >= laio_max_batch(s, dev_max_batch))) { 408 ioq_submit(s); 409 } 410 411 return 0; 412 } 413 414 int coroutine_fn laio_co_submit(BlockDriverState *bs, LinuxAioState *s, int fd, 415 uint64_t offset, QEMUIOVector *qiov, int type, 416 uint64_t dev_max_batch) 417 { 418 int ret; 419 struct qemu_laiocb laiocb = { 420 .co = qemu_coroutine_self(), 421 .nbytes = qiov->size, 422 .ctx = s, 423 .ret = -EINPROGRESS, 424 .is_read = (type == QEMU_AIO_READ), 425 .qiov = qiov, 426 }; 427 428 ret = laio_do_submit(fd, &laiocb, offset, type, dev_max_batch); 429 if (ret < 0) { 430 return ret; 431 } 432 433 if (laiocb.ret == -EINPROGRESS) { 434 qemu_coroutine_yield(); 435 } 436 return laiocb.ret; 437 } 438 439 void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context) 440 { 441 aio_set_event_notifier(old_context, &s->e, false, NULL, NULL, NULL); 442 qemu_bh_delete(s->completion_bh); 443 s->aio_context = NULL; 444 } 445 446 void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context) 447 { 448 s->aio_context = new_context; 449 s->completion_bh = aio_bh_new(new_context, qemu_laio_completion_bh, s); 450 aio_set_event_notifier(new_context, &s->e, false, 451 qemu_laio_completion_cb, 452 qemu_laio_poll_cb, 453 qemu_laio_poll_ready); 454 } 455 456 LinuxAioState *laio_init(Error **errp) 457 { 458 int rc; 459 LinuxAioState *s; 460 461 s = g_malloc0(sizeof(*s)); 462 rc = event_notifier_init(&s->e, false); 463 if (rc < 0) { 464 error_setg_errno(errp, -rc, "failed to initialize event notifier"); 465 goto out_free_state; 466 } 467 468 rc = io_setup(MAX_EVENTS, &s->ctx); 469 if (rc < 0) { 470 error_setg_errno(errp, -rc, "failed to create linux AIO context"); 471 goto out_close_efd; 472 } 473 474 ioq_init(&s->io_q); 475 476 return s; 477 478 out_close_efd: 479 event_notifier_cleanup(&s->e); 480 out_free_state: 481 g_free(s); 482 return NULL; 483 } 484 485 void laio_cleanup(LinuxAioState *s) 486 { 487 event_notifier_cleanup(&s->e); 488 489 if (io_destroy(s->ctx) != 0) { 490 fprintf(stderr, "%s: destroy AIO context %p failed\n", 491 __func__, &s->ctx); 492 } 493 g_free(s); 494 } 495