1 /* 2 * Linux io_uring support. 3 * 4 * Copyright (C) 2009 IBM, Corp. 5 * Copyright (C) 2009 Red Hat, Inc. 6 * Copyright (C) 2019 Aarushi Mehta 7 * 8 * This work is licensed under the terms of the GNU GPL, version 2 or later. 9 * See the COPYING file in the top-level directory. 10 */ 11 #include "qemu/osdep.h" 12 #include <liburing.h> 13 #include "block/aio.h" 14 #include "qemu/queue.h" 15 #include "block/block.h" 16 #include "block/raw-aio.h" 17 #include "qemu/coroutine.h" 18 #include "qapi/error.h" 19 #include "trace.h" 20 21 22 /* io_uring ring size */ 23 #define MAX_ENTRIES 128 24 25 typedef struct LuringAIOCB { 26 Coroutine *co; 27 struct io_uring_sqe sqeq; 28 ssize_t ret; 29 QEMUIOVector *qiov; 30 bool is_read; 31 QSIMPLEQ_ENTRY(LuringAIOCB) next; 32 33 /* 34 * Buffered reads may require resubmission, see 35 * luring_resubmit_short_read(). 36 */ 37 int total_read; 38 QEMUIOVector resubmit_qiov; 39 } LuringAIOCB; 40 41 typedef struct LuringQueue { 42 int plugged; 43 unsigned int in_queue; 44 unsigned int in_flight; 45 bool blocked; 46 QSIMPLEQ_HEAD(, LuringAIOCB) submit_queue; 47 } LuringQueue; 48 49 typedef struct LuringState { 50 AioContext *aio_context; 51 52 struct io_uring ring; 53 54 /* io queue for submit at batch. Protected by AioContext lock. */ 55 LuringQueue io_q; 56 57 /* I/O completion processing. Only runs in I/O thread. */ 58 QEMUBH *completion_bh; 59 } LuringState; 60 61 /** 62 * luring_resubmit: 63 * 64 * Resubmit a request by appending it to submit_queue. The caller must ensure 65 * that ioq_submit() is called later so that submit_queue requests are started. 66 */ 67 static void luring_resubmit(LuringState *s, LuringAIOCB *luringcb) 68 { 69 QSIMPLEQ_INSERT_TAIL(&s->io_q.submit_queue, luringcb, next); 70 s->io_q.in_queue++; 71 } 72 73 /** 74 * luring_resubmit_short_read: 75 * 76 * Before Linux commit 9d93a3f5a0c ("io_uring: punt short reads to async 77 * context") a buffered I/O request with the start of the file range in the 78 * page cache could result in a short read. Applications need to resubmit the 79 * remaining read request. 80 * 81 * This is a slow path but recent kernels never take it. 82 */ 83 static void luring_resubmit_short_read(LuringState *s, LuringAIOCB *luringcb, 84 int nread) 85 { 86 QEMUIOVector *resubmit_qiov; 87 size_t remaining; 88 89 trace_luring_resubmit_short_read(s, luringcb, nread); 90 91 /* Update read position */ 92 luringcb->total_read = nread; 93 remaining = luringcb->qiov->size - luringcb->total_read; 94 95 /* Shorten qiov */ 96 resubmit_qiov = &luringcb->resubmit_qiov; 97 if (resubmit_qiov->iov == NULL) { 98 qemu_iovec_init(resubmit_qiov, luringcb->qiov->niov); 99 } else { 100 qemu_iovec_reset(resubmit_qiov); 101 } 102 qemu_iovec_concat(resubmit_qiov, luringcb->qiov, luringcb->total_read, 103 remaining); 104 105 /* Update sqe */ 106 luringcb->sqeq.off = nread; 107 luringcb->sqeq.addr = (__u64)(uintptr_t)luringcb->resubmit_qiov.iov; 108 luringcb->sqeq.len = luringcb->resubmit_qiov.niov; 109 110 luring_resubmit(s, luringcb); 111 } 112 113 /** 114 * luring_process_completions: 115 * @s: AIO state 116 * 117 * Fetches completed I/O requests, consumes cqes and invokes their callbacks 118 * The function is somewhat tricky because it supports nested event loops, for 119 * example when a request callback invokes aio_poll(). 120 * 121 * Function schedules BH completion so it can be called again in a nested 122 * event loop. When there are no events left to complete the BH is being 123 * canceled. 124 * 125 */ 126 static void luring_process_completions(LuringState *s) 127 { 128 struct io_uring_cqe *cqes; 129 int total_bytes; 130 /* 131 * Request completion callbacks can run the nested event loop. 132 * Schedule ourselves so the nested event loop will "see" remaining 133 * completed requests and process them. Without this, completion 134 * callbacks that wait for other requests using a nested event loop 135 * would hang forever. 136 * 137 * This workaround is needed because io_uring uses poll_wait, which 138 * is woken up when new events are added to the uring, thus polling on 139 * the same uring fd will block unless more events are received. 140 * 141 * Other leaf block drivers (drivers that access the data themselves) 142 * are networking based, so they poll sockets for data and run the 143 * correct coroutine. 144 */ 145 qemu_bh_schedule(s->completion_bh); 146 147 while (io_uring_peek_cqe(&s->ring, &cqes) == 0) { 148 LuringAIOCB *luringcb; 149 int ret; 150 151 if (!cqes) { 152 break; 153 } 154 155 luringcb = io_uring_cqe_get_data(cqes); 156 ret = cqes->res; 157 io_uring_cqe_seen(&s->ring, cqes); 158 cqes = NULL; 159 160 /* Change counters one-by-one because we can be nested. */ 161 s->io_q.in_flight--; 162 trace_luring_process_completion(s, luringcb, ret); 163 164 /* total_read is non-zero only for resubmitted read requests */ 165 total_bytes = ret + luringcb->total_read; 166 167 if (ret < 0) { 168 /* 169 * Only writev/readv/fsync requests on regular files or host block 170 * devices are submitted. Therefore -EAGAIN is not expected but it's 171 * known to happen sometimes with Linux SCSI. Submit again and hope 172 * the request completes successfully. 173 * 174 * For more information, see: 175 * https://lore.kernel.org/io-uring/20210727165811.284510-3-axboe@kernel.dk/T/#u 176 * 177 * If the code is changed to submit other types of requests in the 178 * future, then this workaround may need to be extended to deal with 179 * genuine -EAGAIN results that should not be resubmitted 180 * immediately. 181 */ 182 if (ret == -EINTR || ret == -EAGAIN) { 183 luring_resubmit(s, luringcb); 184 continue; 185 } 186 } else if (!luringcb->qiov) { 187 goto end; 188 } else if (total_bytes == luringcb->qiov->size) { 189 ret = 0; 190 /* Only read/write */ 191 } else { 192 /* Short Read/Write */ 193 if (luringcb->is_read) { 194 if (ret > 0) { 195 luring_resubmit_short_read(s, luringcb, ret); 196 continue; 197 } else { 198 /* Pad with zeroes */ 199 qemu_iovec_memset(luringcb->qiov, total_bytes, 0, 200 luringcb->qiov->size - total_bytes); 201 ret = 0; 202 } 203 } else { 204 ret = -ENOSPC; 205 } 206 } 207 end: 208 luringcb->ret = ret; 209 qemu_iovec_destroy(&luringcb->resubmit_qiov); 210 211 /* 212 * If the coroutine is already entered it must be in ioq_submit() 213 * and will notice luringcb->ret has been filled in when it 214 * eventually runs later. Coroutines cannot be entered recursively 215 * so avoid doing that! 216 */ 217 if (!qemu_coroutine_entered(luringcb->co)) { 218 aio_co_wake(luringcb->co); 219 } 220 } 221 qemu_bh_cancel(s->completion_bh); 222 } 223 224 static int ioq_submit(LuringState *s) 225 { 226 int ret = 0; 227 LuringAIOCB *luringcb, *luringcb_next; 228 229 while (s->io_q.in_queue > 0) { 230 /* 231 * Try to fetch sqes from the ring for requests waiting in 232 * the overflow queue 233 */ 234 QSIMPLEQ_FOREACH_SAFE(luringcb, &s->io_q.submit_queue, next, 235 luringcb_next) { 236 struct io_uring_sqe *sqes = io_uring_get_sqe(&s->ring); 237 if (!sqes) { 238 break; 239 } 240 /* Prep sqe for submission */ 241 *sqes = luringcb->sqeq; 242 QSIMPLEQ_REMOVE_HEAD(&s->io_q.submit_queue, next); 243 } 244 ret = io_uring_submit(&s->ring); 245 trace_luring_io_uring_submit(s, ret); 246 /* Prevent infinite loop if submission is refused */ 247 if (ret <= 0) { 248 if (ret == -EAGAIN || ret == -EINTR) { 249 continue; 250 } 251 break; 252 } 253 s->io_q.in_flight += ret; 254 s->io_q.in_queue -= ret; 255 } 256 s->io_q.blocked = (s->io_q.in_queue > 0); 257 258 if (s->io_q.in_flight) { 259 /* 260 * We can try to complete something just right away if there are 261 * still requests in-flight. 262 */ 263 luring_process_completions(s); 264 } 265 return ret; 266 } 267 268 static void luring_process_completions_and_submit(LuringState *s) 269 { 270 aio_context_acquire(s->aio_context); 271 luring_process_completions(s); 272 273 if (!s->io_q.plugged && s->io_q.in_queue > 0) { 274 ioq_submit(s); 275 } 276 aio_context_release(s->aio_context); 277 } 278 279 static void qemu_luring_completion_bh(void *opaque) 280 { 281 LuringState *s = opaque; 282 luring_process_completions_and_submit(s); 283 } 284 285 static void qemu_luring_completion_cb(void *opaque) 286 { 287 LuringState *s = opaque; 288 luring_process_completions_and_submit(s); 289 } 290 291 static bool qemu_luring_poll_cb(void *opaque) 292 { 293 LuringState *s = opaque; 294 295 return io_uring_cq_ready(&s->ring); 296 } 297 298 static void qemu_luring_poll_ready(void *opaque) 299 { 300 LuringState *s = opaque; 301 302 luring_process_completions_and_submit(s); 303 } 304 305 static void ioq_init(LuringQueue *io_q) 306 { 307 QSIMPLEQ_INIT(&io_q->submit_queue); 308 io_q->plugged = 0; 309 io_q->in_queue = 0; 310 io_q->in_flight = 0; 311 io_q->blocked = false; 312 } 313 314 void luring_io_plug(BlockDriverState *bs, LuringState *s) 315 { 316 trace_luring_io_plug(s); 317 s->io_q.plugged++; 318 } 319 320 void luring_io_unplug(BlockDriverState *bs, LuringState *s) 321 { 322 assert(s->io_q.plugged); 323 trace_luring_io_unplug(s, s->io_q.blocked, s->io_q.plugged, 324 s->io_q.in_queue, s->io_q.in_flight); 325 if (--s->io_q.plugged == 0 && 326 !s->io_q.blocked && s->io_q.in_queue > 0) { 327 ioq_submit(s); 328 } 329 } 330 331 /** 332 * luring_do_submit: 333 * @fd: file descriptor for I/O 334 * @luringcb: AIO control block 335 * @s: AIO state 336 * @offset: offset for request 337 * @type: type of request 338 * 339 * Fetches sqes from ring, adds to pending queue and preps them 340 * 341 */ 342 static int luring_do_submit(int fd, LuringAIOCB *luringcb, LuringState *s, 343 uint64_t offset, int type) 344 { 345 int ret; 346 struct io_uring_sqe *sqes = &luringcb->sqeq; 347 348 switch (type) { 349 case QEMU_AIO_WRITE: 350 io_uring_prep_writev(sqes, fd, luringcb->qiov->iov, 351 luringcb->qiov->niov, offset); 352 break; 353 case QEMU_AIO_READ: 354 io_uring_prep_readv(sqes, fd, luringcb->qiov->iov, 355 luringcb->qiov->niov, offset); 356 break; 357 case QEMU_AIO_FLUSH: 358 io_uring_prep_fsync(sqes, fd, IORING_FSYNC_DATASYNC); 359 break; 360 default: 361 fprintf(stderr, "%s: invalid AIO request type, aborting 0x%x.\n", 362 __func__, type); 363 abort(); 364 } 365 io_uring_sqe_set_data(sqes, luringcb); 366 367 QSIMPLEQ_INSERT_TAIL(&s->io_q.submit_queue, luringcb, next); 368 s->io_q.in_queue++; 369 trace_luring_do_submit(s, s->io_q.blocked, s->io_q.plugged, 370 s->io_q.in_queue, s->io_q.in_flight); 371 if (!s->io_q.blocked && 372 (!s->io_q.plugged || 373 s->io_q.in_flight + s->io_q.in_queue >= MAX_ENTRIES)) { 374 ret = ioq_submit(s); 375 trace_luring_do_submit_done(s, ret); 376 return ret; 377 } 378 return 0; 379 } 380 381 int coroutine_fn luring_co_submit(BlockDriverState *bs, LuringState *s, int fd, 382 uint64_t offset, QEMUIOVector *qiov, int type) 383 { 384 int ret; 385 LuringAIOCB luringcb = { 386 .co = qemu_coroutine_self(), 387 .ret = -EINPROGRESS, 388 .qiov = qiov, 389 .is_read = (type == QEMU_AIO_READ), 390 }; 391 trace_luring_co_submit(bs, s, &luringcb, fd, offset, qiov ? qiov->size : 0, 392 type); 393 ret = luring_do_submit(fd, &luringcb, s, offset, type); 394 395 if (ret < 0) { 396 return ret; 397 } 398 399 if (luringcb.ret == -EINPROGRESS) { 400 qemu_coroutine_yield(); 401 } 402 return luringcb.ret; 403 } 404 405 void luring_detach_aio_context(LuringState *s, AioContext *old_context) 406 { 407 aio_set_fd_handler(old_context, s->ring.ring_fd, false, 408 NULL, NULL, NULL, NULL, s); 409 qemu_bh_delete(s->completion_bh); 410 s->aio_context = NULL; 411 } 412 413 void luring_attach_aio_context(LuringState *s, AioContext *new_context) 414 { 415 s->aio_context = new_context; 416 s->completion_bh = aio_bh_new(new_context, qemu_luring_completion_bh, s); 417 aio_set_fd_handler(s->aio_context, s->ring.ring_fd, false, 418 qemu_luring_completion_cb, NULL, 419 qemu_luring_poll_cb, qemu_luring_poll_ready, s); 420 } 421 422 LuringState *luring_init(Error **errp) 423 { 424 int rc; 425 LuringState *s = g_new0(LuringState, 1); 426 struct io_uring *ring = &s->ring; 427 428 trace_luring_init_state(s, sizeof(*s)); 429 430 rc = io_uring_queue_init(MAX_ENTRIES, ring, 0); 431 if (rc < 0) { 432 error_setg_errno(errp, errno, "failed to init linux io_uring ring"); 433 g_free(s); 434 return NULL; 435 } 436 437 ioq_init(&s->io_q); 438 #ifdef CONFIG_LIBURING_REGISTER_RING_FD 439 if (io_uring_register_ring_fd(&s->ring) < 0) { 440 /* 441 * Only warn about this error: we will fallback to the non-optimized 442 * io_uring operations. 443 */ 444 warn_report("failed to register linux io_uring ring file descriptor"); 445 } 446 #endif 447 448 return s; 449 } 450 451 void luring_cleanup(LuringState *s) 452 { 453 io_uring_queue_exit(&s->ring); 454 trace_luring_cleanup_state(s); 455 g_free(s); 456 } 457