| 1 | /* |
| 2 | * io_uring engine |
| 3 | * |
| 4 | * IO engine using the new native Linux aio io_uring interface. See: |
| 5 | * |
| 6 | * http://git.kernel.dk/cgit/linux-block/log/?h=io_uring |
| 7 | * |
| 8 | */ |
| 9 | #include <stdlib.h> |
| 10 | #include <unistd.h> |
| 11 | #include <errno.h> |
| 12 | #include <sys/time.h> |
| 13 | #include <sys/resource.h> |
| 14 | |
| 15 | #include "../fio.h" |
| 16 | #include "../lib/pow2.h" |
| 17 | #include "../optgroup.h" |
| 18 | #include "../lib/memalign.h" |
| 19 | #include "../lib/fls.h" |
| 20 | #include "../lib/roundup.h" |
| 21 | |
| 22 | #ifdef ARCH_HAVE_IOURING |
| 23 | |
| 24 | #include "../lib/types.h" |
| 25 | #include "../os/linux/io_uring.h" |
| 26 | |
| 27 | struct io_sq_ring { |
| 28 | unsigned *head; |
| 29 | unsigned *tail; |
| 30 | unsigned *ring_mask; |
| 31 | unsigned *ring_entries; |
| 32 | unsigned *flags; |
| 33 | unsigned *array; |
| 34 | }; |
| 35 | |
| 36 | struct io_cq_ring { |
| 37 | unsigned *head; |
| 38 | unsigned *tail; |
| 39 | unsigned *ring_mask; |
| 40 | unsigned *ring_entries; |
| 41 | struct io_uring_cqe *cqes; |
| 42 | }; |
| 43 | |
| 44 | struct ioring_mmap { |
| 45 | void *ptr; |
| 46 | size_t len; |
| 47 | }; |
| 48 | |
| 49 | struct ioring_data { |
| 50 | int ring_fd; |
| 51 | |
| 52 | struct io_u **io_u_index; |
| 53 | |
| 54 | int *fds; |
| 55 | |
| 56 | struct io_sq_ring sq_ring; |
| 57 | struct io_uring_sqe *sqes; |
| 58 | struct iovec *iovecs; |
| 59 | unsigned sq_ring_mask; |
| 60 | |
| 61 | struct io_cq_ring cq_ring; |
| 62 | unsigned cq_ring_mask; |
| 63 | |
| 64 | int queued; |
| 65 | int cq_ring_off; |
| 66 | unsigned iodepth; |
| 67 | bool ioprio_class_set; |
| 68 | bool ioprio_set; |
| 69 | int prepped; |
| 70 | |
| 71 | struct ioring_mmap mmap[3]; |
| 72 | }; |
| 73 | |
| 74 | struct ioring_options { |
| 75 | void *pad; |
| 76 | unsigned int hipri; |
| 77 | unsigned int cmdprio_percentage; |
| 78 | unsigned int fixedbufs; |
| 79 | unsigned int registerfiles; |
| 80 | unsigned int sqpoll_thread; |
| 81 | unsigned int sqpoll_set; |
| 82 | unsigned int sqpoll_cpu; |
| 83 | unsigned int nonvectored; |
| 84 | unsigned int uncached; |
| 85 | unsigned int nowait; |
| 86 | unsigned int force_async; |
| 87 | }; |
| 88 | |
| 89 | static const int ddir_to_op[2][2] = { |
| 90 | { IORING_OP_READV, IORING_OP_READ }, |
| 91 | { IORING_OP_WRITEV, IORING_OP_WRITE } |
| 92 | }; |
| 93 | |
| 94 | static const int fixed_ddir_to_op[2] = { |
| 95 | IORING_OP_READ_FIXED, |
| 96 | IORING_OP_WRITE_FIXED |
| 97 | }; |
| 98 | |
| 99 | static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val) |
| 100 | { |
| 101 | struct ioring_options *o = data; |
| 102 | |
| 103 | o->sqpoll_cpu = *val; |
| 104 | o->sqpoll_set = 1; |
| 105 | return 0; |
| 106 | } |
| 107 | |
| 108 | static struct fio_option options[] = { |
| 109 | { |
| 110 | .name = "hipri", |
| 111 | .lname = "High Priority", |
| 112 | .type = FIO_OPT_STR_SET, |
| 113 | .off1 = offsetof(struct ioring_options, hipri), |
| 114 | .help = "Use polled IO completions", |
| 115 | .category = FIO_OPT_C_ENGINE, |
| 116 | .group = FIO_OPT_G_IOURING, |
| 117 | }, |
| 118 | #ifdef FIO_HAVE_IOPRIO_CLASS |
| 119 | { |
| 120 | .name = "cmdprio_percentage", |
| 121 | .lname = "high priority percentage", |
| 122 | .type = FIO_OPT_INT, |
| 123 | .off1 = offsetof(struct ioring_options, cmdprio_percentage), |
| 124 | .minval = 1, |
| 125 | .maxval = 100, |
| 126 | .help = "Send high priority I/O this percentage of the time", |
| 127 | .category = FIO_OPT_C_ENGINE, |
| 128 | .group = FIO_OPT_G_IOURING, |
| 129 | }, |
| 130 | #else |
| 131 | { |
| 132 | .name = "cmdprio_percentage", |
| 133 | .lname = "high priority percentage", |
| 134 | .type = FIO_OPT_UNSUPPORTED, |
| 135 | .help = "Your platform does not support I/O priority classes", |
| 136 | }, |
| 137 | #endif |
| 138 | { |
| 139 | .name = "fixedbufs", |
| 140 | .lname = "Fixed (pre-mapped) IO buffers", |
| 141 | .type = FIO_OPT_STR_SET, |
| 142 | .off1 = offsetof(struct ioring_options, fixedbufs), |
| 143 | .help = "Pre map IO buffers", |
| 144 | .category = FIO_OPT_C_ENGINE, |
| 145 | .group = FIO_OPT_G_IOURING, |
| 146 | }, |
| 147 | { |
| 148 | .name = "registerfiles", |
| 149 | .lname = "Register file set", |
| 150 | .type = FIO_OPT_STR_SET, |
| 151 | .off1 = offsetof(struct ioring_options, registerfiles), |
| 152 | .help = "Pre-open/register files", |
| 153 | .category = FIO_OPT_C_ENGINE, |
| 154 | .group = FIO_OPT_G_IOURING, |
| 155 | }, |
| 156 | { |
| 157 | .name = "sqthread_poll", |
| 158 | .lname = "Kernel SQ thread polling", |
| 159 | .type = FIO_OPT_INT, |
| 160 | .off1 = offsetof(struct ioring_options, sqpoll_thread), |
| 161 | .help = "Offload submission/completion to kernel thread", |
| 162 | .category = FIO_OPT_C_ENGINE, |
| 163 | .group = FIO_OPT_G_IOURING, |
| 164 | }, |
| 165 | { |
| 166 | .name = "sqthread_poll_cpu", |
| 167 | .lname = "SQ Thread Poll CPU", |
| 168 | .type = FIO_OPT_INT, |
| 169 | .cb = fio_ioring_sqpoll_cb, |
| 170 | .help = "What CPU to run SQ thread polling on", |
| 171 | .category = FIO_OPT_C_ENGINE, |
| 172 | .group = FIO_OPT_G_IOURING, |
| 173 | }, |
| 174 | { |
| 175 | .name = "nonvectored", |
| 176 | .lname = "Non-vectored", |
| 177 | .type = FIO_OPT_INT, |
| 178 | .off1 = offsetof(struct ioring_options, nonvectored), |
| 179 | .def = "-1", |
| 180 | .help = "Use non-vectored read/write commands", |
| 181 | .category = FIO_OPT_C_ENGINE, |
| 182 | .group = FIO_OPT_G_IOURING, |
| 183 | }, |
| 184 | { |
| 185 | .name = "uncached", |
| 186 | .lname = "Uncached", |
| 187 | .type = FIO_OPT_INT, |
| 188 | .off1 = offsetof(struct ioring_options, uncached), |
| 189 | .help = "Use RWF_UNCACHED for buffered read/writes", |
| 190 | .category = FIO_OPT_C_ENGINE, |
| 191 | .group = FIO_OPT_G_IOURING, |
| 192 | }, |
| 193 | { |
| 194 | .name = "nowait", |
| 195 | .lname = "RWF_NOWAIT", |
| 196 | .type = FIO_OPT_BOOL, |
| 197 | .off1 = offsetof(struct ioring_options, nowait), |
| 198 | .help = "Use RWF_NOWAIT for reads/writes", |
| 199 | .category = FIO_OPT_C_ENGINE, |
| 200 | .group = FIO_OPT_G_IOURING, |
| 201 | }, |
| 202 | { |
| 203 | .name = "force_async", |
| 204 | .lname = "Force async", |
| 205 | .type = FIO_OPT_INT, |
| 206 | .off1 = offsetof(struct ioring_options, force_async), |
| 207 | .help = "Set IOSQE_ASYNC every N requests", |
| 208 | .category = FIO_OPT_C_ENGINE, |
| 209 | .group = FIO_OPT_G_IOURING, |
| 210 | }, |
| 211 | { |
| 212 | .name = NULL, |
| 213 | }, |
| 214 | }; |
| 215 | |
| 216 | static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit, |
| 217 | unsigned int min_complete, unsigned int flags) |
| 218 | { |
| 219 | return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit, |
| 220 | min_complete, flags, NULL, 0); |
| 221 | } |
| 222 | |
| 223 | static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u) |
| 224 | { |
| 225 | struct ioring_data *ld = td->io_ops_data; |
| 226 | struct ioring_options *o = td->eo; |
| 227 | struct fio_file *f = io_u->file; |
| 228 | struct io_uring_sqe *sqe; |
| 229 | |
| 230 | sqe = &ld->sqes[io_u->index]; |
| 231 | |
| 232 | if (o->registerfiles) { |
| 233 | sqe->fd = f->engine_pos; |
| 234 | sqe->flags = IOSQE_FIXED_FILE; |
| 235 | } else { |
| 236 | sqe->fd = f->fd; |
| 237 | } |
| 238 | |
| 239 | if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) { |
| 240 | if (o->fixedbufs) { |
| 241 | sqe->opcode = fixed_ddir_to_op[io_u->ddir]; |
| 242 | sqe->addr = (unsigned long) io_u->xfer_buf; |
| 243 | sqe->len = io_u->xfer_buflen; |
| 244 | sqe->buf_index = io_u->index; |
| 245 | } else { |
| 246 | struct iovec *iov = &ld->iovecs[io_u->index]; |
| 247 | |
| 248 | /* |
| 249 | * Update based on actual io_u, requeue could have |
| 250 | * adjusted these |
| 251 | */ |
| 252 | iov->iov_base = io_u->xfer_buf; |
| 253 | iov->iov_len = io_u->xfer_buflen; |
| 254 | |
| 255 | sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored]; |
| 256 | if (o->nonvectored) { |
| 257 | sqe->addr = (unsigned long) iov->iov_base; |
| 258 | sqe->len = iov->iov_len; |
| 259 | } else { |
| 260 | sqe->addr = (unsigned long) iov; |
| 261 | sqe->len = 1; |
| 262 | } |
| 263 | } |
| 264 | if (!td->o.odirect && o->uncached) |
| 265 | sqe->rw_flags = RWF_UNCACHED; |
| 266 | if (o->nowait) |
| 267 | sqe->rw_flags |= RWF_NOWAIT; |
| 268 | if (ld->ioprio_class_set) |
| 269 | sqe->ioprio = td->o.ioprio_class << 13; |
| 270 | if (ld->ioprio_set) |
| 271 | sqe->ioprio |= td->o.ioprio; |
| 272 | sqe->off = io_u->offset; |
| 273 | sqe->rw_flags = 0; |
| 274 | } else if (ddir_sync(io_u->ddir)) { |
| 275 | sqe->ioprio = 0; |
| 276 | if (io_u->ddir == DDIR_SYNC_FILE_RANGE) { |
| 277 | sqe->off = f->first_write; |
| 278 | sqe->len = f->last_write - f->first_write; |
| 279 | sqe->sync_range_flags = td->o.sync_file_range; |
| 280 | sqe->opcode = IORING_OP_SYNC_FILE_RANGE; |
| 281 | } else { |
| 282 | sqe->off = 0; |
| 283 | sqe->addr = 0; |
| 284 | sqe->len = 0; |
| 285 | if (io_u->ddir == DDIR_DATASYNC) |
| 286 | sqe->fsync_flags |= IORING_FSYNC_DATASYNC; |
| 287 | sqe->opcode = IORING_OP_FSYNC; |
| 288 | } |
| 289 | } |
| 290 | |
| 291 | if (o->force_async && ++ld->prepped == o->force_async) { |
| 292 | ld->prepped = 0; |
| 293 | sqe->flags |= IOSQE_ASYNC; |
| 294 | } |
| 295 | |
| 296 | sqe->user_data = (unsigned long) io_u; |
| 297 | return 0; |
| 298 | } |
| 299 | |
| 300 | static struct io_u *fio_ioring_event(struct thread_data *td, int event) |
| 301 | { |
| 302 | struct ioring_data *ld = td->io_ops_data; |
| 303 | struct io_uring_cqe *cqe; |
| 304 | struct io_u *io_u; |
| 305 | unsigned index; |
| 306 | |
| 307 | index = (event + ld->cq_ring_off) & ld->cq_ring_mask; |
| 308 | |
| 309 | cqe = &ld->cq_ring.cqes[index]; |
| 310 | io_u = (struct io_u *) (uintptr_t) cqe->user_data; |
| 311 | |
| 312 | if (cqe->res != io_u->xfer_buflen) { |
| 313 | if (cqe->res > io_u->xfer_buflen) |
| 314 | io_u->error = -cqe->res; |
| 315 | else |
| 316 | io_u->resid = io_u->xfer_buflen - cqe->res; |
| 317 | } else |
| 318 | io_u->error = 0; |
| 319 | |
| 320 | return io_u; |
| 321 | } |
| 322 | |
| 323 | static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events, |
| 324 | unsigned int max) |
| 325 | { |
| 326 | struct ioring_data *ld = td->io_ops_data; |
| 327 | struct io_cq_ring *ring = &ld->cq_ring; |
| 328 | unsigned head, reaped = 0; |
| 329 | |
| 330 | head = *ring->head; |
| 331 | do { |
| 332 | if (head == atomic_load_acquire(ring->tail)) |
| 333 | break; |
| 334 | reaped++; |
| 335 | head++; |
| 336 | } while (reaped + events < max); |
| 337 | |
| 338 | if (reaped) |
| 339 | atomic_store_release(ring->head, head); |
| 340 | |
| 341 | return reaped; |
| 342 | } |
| 343 | |
| 344 | static int fio_ioring_getevents(struct thread_data *td, unsigned int min, |
| 345 | unsigned int max, const struct timespec *t) |
| 346 | { |
| 347 | struct ioring_data *ld = td->io_ops_data; |
| 348 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
| 349 | struct ioring_options *o = td->eo; |
| 350 | struct io_cq_ring *ring = &ld->cq_ring; |
| 351 | unsigned events = 0; |
| 352 | int r; |
| 353 | |
| 354 | ld->cq_ring_off = *ring->head; |
| 355 | do { |
| 356 | r = fio_ioring_cqring_reap(td, events, max); |
| 357 | if (r) { |
| 358 | events += r; |
| 359 | if (actual_min != 0) |
| 360 | actual_min -= r; |
| 361 | continue; |
| 362 | } |
| 363 | |
| 364 | if (!o->sqpoll_thread) { |
| 365 | r = io_uring_enter(ld, 0, actual_min, |
| 366 | IORING_ENTER_GETEVENTS); |
| 367 | if (r < 0) { |
| 368 | if (errno == EAGAIN || errno == EINTR) |
| 369 | continue; |
| 370 | td_verror(td, errno, "io_uring_enter"); |
| 371 | break; |
| 372 | } |
| 373 | } |
| 374 | } while (events < min); |
| 375 | |
| 376 | return r < 0 ? r : events; |
| 377 | } |
| 378 | |
| 379 | static void fio_ioring_prio_prep(struct thread_data *td, struct io_u *io_u) |
| 380 | { |
| 381 | struct ioring_options *o = td->eo; |
| 382 | struct ioring_data *ld = td->io_ops_data; |
| 383 | if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) { |
| 384 | ld->sqes[io_u->index].ioprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT; |
| 385 | io_u->flags |= IO_U_F_PRIORITY; |
| 386 | } |
| 387 | return; |
| 388 | } |
| 389 | |
| 390 | static enum fio_q_status fio_ioring_queue(struct thread_data *td, |
| 391 | struct io_u *io_u) |
| 392 | { |
| 393 | struct ioring_data *ld = td->io_ops_data; |
| 394 | struct io_sq_ring *ring = &ld->sq_ring; |
| 395 | struct ioring_options *o = td->eo; |
| 396 | unsigned tail, next_tail; |
| 397 | |
| 398 | fio_ro_check(td, io_u); |
| 399 | |
| 400 | if (ld->queued == ld->iodepth) |
| 401 | return FIO_Q_BUSY; |
| 402 | |
| 403 | if (io_u->ddir == DDIR_TRIM) { |
| 404 | if (ld->queued) |
| 405 | return FIO_Q_BUSY; |
| 406 | |
| 407 | do_io_u_trim(td, io_u); |
| 408 | io_u_mark_submit(td, 1); |
| 409 | io_u_mark_complete(td, 1); |
| 410 | return FIO_Q_COMPLETED; |
| 411 | } |
| 412 | |
| 413 | tail = *ring->tail; |
| 414 | next_tail = tail + 1; |
| 415 | if (next_tail == atomic_load_acquire(ring->head)) |
| 416 | return FIO_Q_BUSY; |
| 417 | |
| 418 | if (o->cmdprio_percentage) |
| 419 | fio_ioring_prio_prep(td, io_u); |
| 420 | ring->array[tail & ld->sq_ring_mask] = io_u->index; |
| 421 | atomic_store_release(ring->tail, next_tail); |
| 422 | |
| 423 | ld->queued++; |
| 424 | return FIO_Q_QUEUED; |
| 425 | } |
| 426 | |
| 427 | static void fio_ioring_queued(struct thread_data *td, int start, int nr) |
| 428 | { |
| 429 | struct ioring_data *ld = td->io_ops_data; |
| 430 | struct timespec now; |
| 431 | |
| 432 | if (!fio_fill_issue_time(td)) |
| 433 | return; |
| 434 | |
| 435 | fio_gettime(&now, NULL); |
| 436 | |
| 437 | while (nr--) { |
| 438 | struct io_sq_ring *ring = &ld->sq_ring; |
| 439 | int index = ring->array[start & ld->sq_ring_mask]; |
| 440 | struct io_u *io_u = ld->io_u_index[index]; |
| 441 | |
| 442 | memcpy(&io_u->issue_time, &now, sizeof(now)); |
| 443 | io_u_queued(td, io_u); |
| 444 | |
| 445 | start++; |
| 446 | } |
| 447 | } |
| 448 | |
| 449 | static int fio_ioring_commit(struct thread_data *td) |
| 450 | { |
| 451 | struct ioring_data *ld = td->io_ops_data; |
| 452 | struct ioring_options *o = td->eo; |
| 453 | int ret; |
| 454 | |
| 455 | if (!ld->queued) |
| 456 | return 0; |
| 457 | |
| 458 | /* |
| 459 | * Kernel side does submission. just need to check if the ring is |
| 460 | * flagged as needing a kick, if so, call io_uring_enter(). This |
| 461 | * only happens if we've been idle too long. |
| 462 | */ |
| 463 | if (o->sqpoll_thread) { |
| 464 | struct io_sq_ring *ring = &ld->sq_ring; |
| 465 | unsigned flags; |
| 466 | |
| 467 | flags = atomic_load_acquire(ring->flags); |
| 468 | if (flags & IORING_SQ_NEED_WAKEUP) |
| 469 | io_uring_enter(ld, ld->queued, 0, |
| 470 | IORING_ENTER_SQ_WAKEUP); |
| 471 | ld->queued = 0; |
| 472 | return 0; |
| 473 | } |
| 474 | |
| 475 | do { |
| 476 | unsigned start = *ld->sq_ring.head; |
| 477 | long nr = ld->queued; |
| 478 | |
| 479 | ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS); |
| 480 | if (ret > 0) { |
| 481 | fio_ioring_queued(td, start, ret); |
| 482 | io_u_mark_submit(td, ret); |
| 483 | |
| 484 | ld->queued -= ret; |
| 485 | ret = 0; |
| 486 | } else if (!ret) { |
| 487 | io_u_mark_submit(td, ret); |
| 488 | continue; |
| 489 | } else { |
| 490 | if (errno == EAGAIN || errno == EINTR) { |
| 491 | ret = fio_ioring_cqring_reap(td, 0, ld->queued); |
| 492 | if (ret) |
| 493 | continue; |
| 494 | /* Shouldn't happen */ |
| 495 | usleep(1); |
| 496 | continue; |
| 497 | } |
| 498 | td_verror(td, errno, "io_uring_enter submit"); |
| 499 | break; |
| 500 | } |
| 501 | } while (ld->queued); |
| 502 | |
| 503 | return ret; |
| 504 | } |
| 505 | |
| 506 | static void fio_ioring_unmap(struct ioring_data *ld) |
| 507 | { |
| 508 | int i; |
| 509 | |
| 510 | for (i = 0; i < ARRAY_SIZE(ld->mmap); i++) |
| 511 | munmap(ld->mmap[i].ptr, ld->mmap[i].len); |
| 512 | close(ld->ring_fd); |
| 513 | } |
| 514 | |
| 515 | static void fio_ioring_cleanup(struct thread_data *td) |
| 516 | { |
| 517 | struct ioring_data *ld = td->io_ops_data; |
| 518 | |
| 519 | if (ld) { |
| 520 | if (!(td->flags & TD_F_CHILD)) |
| 521 | fio_ioring_unmap(ld); |
| 522 | |
| 523 | free(ld->io_u_index); |
| 524 | free(ld->iovecs); |
| 525 | free(ld->fds); |
| 526 | free(ld); |
| 527 | } |
| 528 | } |
| 529 | |
| 530 | static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p) |
| 531 | { |
| 532 | struct io_sq_ring *sring = &ld->sq_ring; |
| 533 | struct io_cq_ring *cring = &ld->cq_ring; |
| 534 | void *ptr; |
| 535 | |
| 536 | ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32); |
| 537 | ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE, |
| 538 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, |
| 539 | IORING_OFF_SQ_RING); |
| 540 | ld->mmap[0].ptr = ptr; |
| 541 | sring->head = ptr + p->sq_off.head; |
| 542 | sring->tail = ptr + p->sq_off.tail; |
| 543 | sring->ring_mask = ptr + p->sq_off.ring_mask; |
| 544 | sring->ring_entries = ptr + p->sq_off.ring_entries; |
| 545 | sring->flags = ptr + p->sq_off.flags; |
| 546 | sring->array = ptr + p->sq_off.array; |
| 547 | ld->sq_ring_mask = *sring->ring_mask; |
| 548 | |
| 549 | ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe); |
| 550 | ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE, |
| 551 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, |
| 552 | IORING_OFF_SQES); |
| 553 | ld->mmap[1].ptr = ld->sqes; |
| 554 | |
| 555 | ld->mmap[2].len = p->cq_off.cqes + |
| 556 | p->cq_entries * sizeof(struct io_uring_cqe); |
| 557 | ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE, |
| 558 | MAP_SHARED | MAP_POPULATE, ld->ring_fd, |
| 559 | IORING_OFF_CQ_RING); |
| 560 | ld->mmap[2].ptr = ptr; |
| 561 | cring->head = ptr + p->cq_off.head; |
| 562 | cring->tail = ptr + p->cq_off.tail; |
| 563 | cring->ring_mask = ptr + p->cq_off.ring_mask; |
| 564 | cring->ring_entries = ptr + p->cq_off.ring_entries; |
| 565 | cring->cqes = ptr + p->cq_off.cqes; |
| 566 | ld->cq_ring_mask = *cring->ring_mask; |
| 567 | return 0; |
| 568 | } |
| 569 | |
| 570 | static void fio_ioring_probe(struct thread_data *td) |
| 571 | { |
| 572 | struct ioring_data *ld = td->io_ops_data; |
| 573 | struct ioring_options *o = td->eo; |
| 574 | struct io_uring_probe *p; |
| 575 | int ret; |
| 576 | |
| 577 | /* already set by user, don't touch */ |
| 578 | if (o->nonvectored != -1) |
| 579 | return; |
| 580 | |
| 581 | /* default to off, as that's always safe */ |
| 582 | o->nonvectored = 0; |
| 583 | |
| 584 | p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op)); |
| 585 | if (!p) |
| 586 | return; |
| 587 | |
| 588 | memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op)); |
| 589 | ret = syscall(__NR_io_uring_register, ld->ring_fd, |
| 590 | IORING_REGISTER_PROBE, p, 256); |
| 591 | if (ret < 0) |
| 592 | goto out; |
| 593 | |
| 594 | if (IORING_OP_WRITE > p->ops_len) |
| 595 | goto out; |
| 596 | |
| 597 | if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) && |
| 598 | (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED)) |
| 599 | o->nonvectored = 1; |
| 600 | out: |
| 601 | free(p); |
| 602 | } |
| 603 | |
| 604 | static int fio_ioring_queue_init(struct thread_data *td) |
| 605 | { |
| 606 | struct ioring_data *ld = td->io_ops_data; |
| 607 | struct ioring_options *o = td->eo; |
| 608 | int depth = td->o.iodepth; |
| 609 | struct io_uring_params p; |
| 610 | int ret; |
| 611 | |
| 612 | memset(&p, 0, sizeof(p)); |
| 613 | |
| 614 | if (o->hipri) |
| 615 | p.flags |= IORING_SETUP_IOPOLL; |
| 616 | if (o->sqpoll_thread) { |
| 617 | p.flags |= IORING_SETUP_SQPOLL; |
| 618 | if (o->sqpoll_set) { |
| 619 | p.flags |= IORING_SETUP_SQ_AFF; |
| 620 | p.sq_thread_cpu = o->sqpoll_cpu; |
| 621 | } |
| 622 | } |
| 623 | |
| 624 | ret = syscall(__NR_io_uring_setup, depth, &p); |
| 625 | if (ret < 0) |
| 626 | return ret; |
| 627 | |
| 628 | ld->ring_fd = ret; |
| 629 | |
| 630 | fio_ioring_probe(td); |
| 631 | |
| 632 | if (o->fixedbufs) { |
| 633 | ret = syscall(__NR_io_uring_register, ld->ring_fd, |
| 634 | IORING_REGISTER_BUFFERS, ld->iovecs, depth); |
| 635 | if (ret < 0) |
| 636 | return ret; |
| 637 | } |
| 638 | |
| 639 | return fio_ioring_mmap(ld, &p); |
| 640 | } |
| 641 | |
| 642 | static int fio_ioring_register_files(struct thread_data *td) |
| 643 | { |
| 644 | struct ioring_data *ld = td->io_ops_data; |
| 645 | struct fio_file *f; |
| 646 | unsigned int i; |
| 647 | int ret; |
| 648 | |
| 649 | ld->fds = calloc(td->o.nr_files, sizeof(int)); |
| 650 | |
| 651 | for_each_file(td, f, i) { |
| 652 | ret = generic_open_file(td, f); |
| 653 | if (ret) |
| 654 | goto err; |
| 655 | ld->fds[i] = f->fd; |
| 656 | f->engine_pos = i; |
| 657 | } |
| 658 | |
| 659 | ret = syscall(__NR_io_uring_register, ld->ring_fd, |
| 660 | IORING_REGISTER_FILES, ld->fds, td->o.nr_files); |
| 661 | if (ret) { |
| 662 | err: |
| 663 | free(ld->fds); |
| 664 | ld->fds = NULL; |
| 665 | } |
| 666 | |
| 667 | /* |
| 668 | * Pretend the file is closed again, and really close it if we hit |
| 669 | * an error. |
| 670 | */ |
| 671 | for_each_file(td, f, i) { |
| 672 | if (ret) { |
| 673 | int fio_unused ret2; |
| 674 | ret2 = generic_close_file(td, f); |
| 675 | } else |
| 676 | f->fd = -1; |
| 677 | } |
| 678 | |
| 679 | return ret; |
| 680 | } |
| 681 | |
| 682 | static int fio_ioring_post_init(struct thread_data *td) |
| 683 | { |
| 684 | struct ioring_data *ld = td->io_ops_data; |
| 685 | struct ioring_options *o = td->eo; |
| 686 | struct io_u *io_u; |
| 687 | int err, i; |
| 688 | |
| 689 | for (i = 0; i < td->o.iodepth; i++) { |
| 690 | struct iovec *iov = &ld->iovecs[i]; |
| 691 | |
| 692 | io_u = ld->io_u_index[i]; |
| 693 | iov->iov_base = io_u->buf; |
| 694 | iov->iov_len = td_max_bs(td); |
| 695 | } |
| 696 | |
| 697 | err = fio_ioring_queue_init(td); |
| 698 | if (err) { |
| 699 | td_verror(td, errno, "io_queue_init"); |
| 700 | return 1; |
| 701 | } |
| 702 | |
| 703 | for (i = 0; i < td->o.iodepth; i++) { |
| 704 | struct io_uring_sqe *sqe; |
| 705 | |
| 706 | sqe = &ld->sqes[i]; |
| 707 | memset(sqe, 0, sizeof(*sqe)); |
| 708 | } |
| 709 | |
| 710 | if (o->registerfiles) { |
| 711 | err = fio_ioring_register_files(td); |
| 712 | if (err) { |
| 713 | td_verror(td, errno, "ioring_register_files"); |
| 714 | return 1; |
| 715 | } |
| 716 | } |
| 717 | |
| 718 | return 0; |
| 719 | } |
| 720 | |
| 721 | static int fio_ioring_init(struct thread_data *td) |
| 722 | { |
| 723 | struct ioring_options *o = td->eo; |
| 724 | struct ioring_data *ld; |
| 725 | struct thread_options *to = &td->o; |
| 726 | |
| 727 | if (to->io_submit_mode == IO_MODE_OFFLOAD) { |
| 728 | log_err("fio: io_submit_mode=offload is not compatible (or " |
| 729 | "useful) with io_uring\n"); |
| 730 | return 1; |
| 731 | } |
| 732 | |
| 733 | /* sqthread submission requires registered files */ |
| 734 | if (o->sqpoll_thread) |
| 735 | o->registerfiles = 1; |
| 736 | |
| 737 | if (o->registerfiles && td->o.nr_files != td->o.open_files) { |
| 738 | log_err("fio: io_uring registered files require nr_files to " |
| 739 | "be identical to open_files\n"); |
| 740 | return 1; |
| 741 | } |
| 742 | |
| 743 | ld = calloc(1, sizeof(*ld)); |
| 744 | |
| 745 | /* ring depth must be a power-of-2 */ |
| 746 | ld->iodepth = td->o.iodepth; |
| 747 | td->o.iodepth = roundup_pow2(td->o.iodepth); |
| 748 | |
| 749 | /* io_u index */ |
| 750 | ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *)); |
| 751 | ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec)); |
| 752 | |
| 753 | td->io_ops_data = ld; |
| 754 | |
| 755 | /* |
| 756 | * Check for option conflicts |
| 757 | */ |
| 758 | if ((fio_option_is_set(to, ioprio) || fio_option_is_set(to, ioprio_class)) && |
| 759 | o->cmdprio_percentage != 0) { |
| 760 | log_err("%s: cmdprio_percentage option and mutually exclusive " |
| 761 | "prio or prioclass option is set, exiting\n", to->name); |
| 762 | td_verror(td, EINVAL, "fio_io_uring_init"); |
| 763 | return 1; |
| 764 | } |
| 765 | |
| 766 | if (fio_option_is_set(&td->o, ioprio_class)) |
| 767 | ld->ioprio_class_set = true; |
| 768 | if (fio_option_is_set(&td->o, ioprio)) |
| 769 | ld->ioprio_set = true; |
| 770 | |
| 771 | return 0; |
| 772 | } |
| 773 | |
| 774 | static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u) |
| 775 | { |
| 776 | struct ioring_data *ld = td->io_ops_data; |
| 777 | |
| 778 | ld->io_u_index[io_u->index] = io_u; |
| 779 | return 0; |
| 780 | } |
| 781 | |
| 782 | static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f) |
| 783 | { |
| 784 | struct ioring_data *ld = td->io_ops_data; |
| 785 | struct ioring_options *o = td->eo; |
| 786 | |
| 787 | if (!ld || !o->registerfiles) |
| 788 | return generic_open_file(td, f); |
| 789 | |
| 790 | f->fd = ld->fds[f->engine_pos]; |
| 791 | return 0; |
| 792 | } |
| 793 | |
| 794 | static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f) |
| 795 | { |
| 796 | struct ioring_data *ld = td->io_ops_data; |
| 797 | struct ioring_options *o = td->eo; |
| 798 | |
| 799 | if (!ld || !o->registerfiles) |
| 800 | return generic_close_file(td, f); |
| 801 | |
| 802 | f->fd = -1; |
| 803 | return 0; |
| 804 | } |
| 805 | |
| 806 | static struct ioengine_ops ioengine = { |
| 807 | .name = "io_uring", |
| 808 | .version = FIO_IOOPS_VERSION, |
| 809 | .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD, |
| 810 | .init = fio_ioring_init, |
| 811 | .post_init = fio_ioring_post_init, |
| 812 | .io_u_init = fio_ioring_io_u_init, |
| 813 | .prep = fio_ioring_prep, |
| 814 | .queue = fio_ioring_queue, |
| 815 | .commit = fio_ioring_commit, |
| 816 | .getevents = fio_ioring_getevents, |
| 817 | .event = fio_ioring_event, |
| 818 | .cleanup = fio_ioring_cleanup, |
| 819 | .open_file = fio_ioring_open_file, |
| 820 | .close_file = fio_ioring_close_file, |
| 821 | .get_file_size = generic_get_file_size, |
| 822 | .options = options, |
| 823 | .option_struct_size = sizeof(struct ioring_options), |
| 824 | }; |
| 825 | |
| 826 | static void fio_init fio_ioring_register(void) |
| 827 | { |
| 828 | register_ioengine(&ioengine); |
| 829 | } |
| 830 | |
| 831 | static void fio_exit fio_ioring_unregister(void) |
| 832 | { |
| 833 | unregister_ioengine(&ioengine); |
| 834 | } |
| 835 | #endif |