4 * IO engine using the new native Linux aio io_uring interface. See:
6 * http://git.kernel.dk/cgit/linux-block/log/?h=io_uring
13 #include <sys/resource.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"
22 #ifdef ARCH_HAVE_IOURING
24 #include "../lib/types.h"
25 #include "../os/linux/io_uring.h"
31 unsigned *ring_entries;
40 unsigned *ring_entries;
41 struct io_uring_cqe *cqes;
52 struct io_u **io_u_index;
56 struct io_sq_ring sq_ring;
57 struct io_uring_sqe *sqes;
59 unsigned sq_ring_mask;
61 struct io_cq_ring cq_ring;
62 unsigned cq_ring_mask;
67 bool ioprio_class_set;
71 struct ioring_mmap mmap[3];
74 struct ioring_options {
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;
86 unsigned int force_async;
89 static const int ddir_to_op[2][2] = {
90 { IORING_OP_READV, IORING_OP_READ },
91 { IORING_OP_WRITEV, IORING_OP_WRITE }
94 static const int fixed_ddir_to_op[2] = {
99 static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val)
101 struct ioring_options *o = data;
103 o->sqpoll_cpu = *val;
108 static struct fio_option options[] = {
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,
118 #ifdef FIO_HAVE_IOPRIO_CLASS
120 .name = "cmdprio_percentage",
121 .lname = "high priority percentage",
123 .off1 = offsetof(struct ioring_options, cmdprio_percentage),
126 .help = "Send high priority I/O this percentage of the time",
127 .category = FIO_OPT_C_ENGINE,
128 .group = FIO_OPT_G_IOURING,
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",
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,
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,
157 .name = "sqthread_poll",
158 .lname = "Kernel SQ thread polling",
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,
166 .name = "sqthread_poll_cpu",
167 .lname = "SQ Thread Poll CPU",
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,
175 .name = "nonvectored",
176 .lname = "Non-vectored",
178 .off1 = offsetof(struct ioring_options, nonvectored),
180 .help = "Use non-vectored read/write commands",
181 .category = FIO_OPT_C_ENGINE,
182 .group = FIO_OPT_G_IOURING,
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,
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,
203 .name = "force_async",
204 .lname = "Force async",
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,
216 static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
217 unsigned int min_complete, unsigned int flags)
219 return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit,
220 min_complete, flags, NULL, 0);
223 static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
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;
230 sqe = &ld->sqes[io_u->index];
232 if (o->registerfiles) {
233 sqe->fd = f->engine_pos;
234 sqe->flags = IOSQE_FIXED_FILE;
240 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
242 sqe->opcode = fixed_ddir_to_op[io_u->ddir];
243 sqe->addr = (unsigned long) io_u->xfer_buf;
244 sqe->len = io_u->xfer_buflen;
245 sqe->buf_index = io_u->index;
247 struct iovec *iov = &ld->iovecs[io_u->index];
250 * Update based on actual io_u, requeue could have
253 iov->iov_base = io_u->xfer_buf;
254 iov->iov_len = io_u->xfer_buflen;
256 sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored];
257 if (o->nonvectored) {
258 sqe->addr = (unsigned long) iov->iov_base;
259 sqe->len = iov->iov_len;
261 sqe->addr = (unsigned long) iov;
266 if (!td->o.odirect && o->uncached)
267 sqe->rw_flags |= RWF_UNCACHED;
269 sqe->rw_flags |= RWF_NOWAIT;
270 if (ld->ioprio_class_set)
271 sqe->ioprio = td->o.ioprio_class << 13;
273 sqe->ioprio |= td->o.ioprio;
274 sqe->off = io_u->offset;
275 } else if (ddir_sync(io_u->ddir)) {
277 if (io_u->ddir == DDIR_SYNC_FILE_RANGE) {
278 sqe->off = f->first_write;
279 sqe->len = f->last_write - f->first_write;
280 sqe->sync_range_flags = td->o.sync_file_range;
281 sqe->opcode = IORING_OP_SYNC_FILE_RANGE;
286 if (io_u->ddir == DDIR_DATASYNC)
287 sqe->fsync_flags |= IORING_FSYNC_DATASYNC;
288 sqe->opcode = IORING_OP_FSYNC;
292 if (o->force_async && ++ld->prepped == o->force_async) {
294 sqe->flags |= IOSQE_ASYNC;
297 sqe->user_data = (unsigned long) io_u;
301 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
303 struct ioring_data *ld = td->io_ops_data;
304 struct io_uring_cqe *cqe;
308 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
310 cqe = &ld->cq_ring.cqes[index];
311 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
313 if (cqe->res != io_u->xfer_buflen) {
314 if (cqe->res > io_u->xfer_buflen)
315 io_u->error = -cqe->res;
317 io_u->resid = io_u->xfer_buflen - cqe->res;
324 static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
327 struct ioring_data *ld = td->io_ops_data;
328 struct io_cq_ring *ring = &ld->cq_ring;
329 unsigned head, reaped = 0;
333 if (head == atomic_load_acquire(ring->tail))
337 } while (reaped + events < max);
340 atomic_store_release(ring->head, head);
345 static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
346 unsigned int max, const struct timespec *t)
348 struct ioring_data *ld = td->io_ops_data;
349 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
350 struct ioring_options *o = td->eo;
351 struct io_cq_ring *ring = &ld->cq_ring;
355 ld->cq_ring_off = *ring->head;
357 r = fio_ioring_cqring_reap(td, events, max);
365 if (!o->sqpoll_thread) {
366 r = io_uring_enter(ld, 0, actual_min,
367 IORING_ENTER_GETEVENTS);
369 if (errno == EAGAIN || errno == EINTR)
371 td_verror(td, errno, "io_uring_enter");
375 } while (events < min);
377 return r < 0 ? r : events;
380 static void fio_ioring_prio_prep(struct thread_data *td, struct io_u *io_u)
382 struct ioring_options *o = td->eo;
383 struct ioring_data *ld = td->io_ops_data;
384 if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) {
385 ld->sqes[io_u->index].ioprio = ioprio_value(IOPRIO_CLASS_RT, 0);
386 io_u->flags |= IO_U_F_PRIORITY;
388 ld->sqes[io_u->index].ioprio = 0;
393 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
396 struct ioring_data *ld = td->io_ops_data;
397 struct io_sq_ring *ring = &ld->sq_ring;
398 struct ioring_options *o = td->eo;
399 unsigned tail, next_tail;
401 fio_ro_check(td, io_u);
403 if (ld->queued == ld->iodepth)
406 if (io_u->ddir == DDIR_TRIM) {
410 do_io_u_trim(td, io_u);
411 io_u_mark_submit(td, 1);
412 io_u_mark_complete(td, 1);
413 return FIO_Q_COMPLETED;
417 next_tail = tail + 1;
418 if (next_tail == atomic_load_acquire(ring->head))
421 if (o->cmdprio_percentage)
422 fio_ioring_prio_prep(td, io_u);
423 ring->array[tail & ld->sq_ring_mask] = io_u->index;
424 atomic_store_release(ring->tail, next_tail);
430 static void fio_ioring_queued(struct thread_data *td, int start, int nr)
432 struct ioring_data *ld = td->io_ops_data;
435 if (!fio_fill_issue_time(td))
438 fio_gettime(&now, NULL);
441 struct io_sq_ring *ring = &ld->sq_ring;
442 int index = ring->array[start & ld->sq_ring_mask];
443 struct io_u *io_u = ld->io_u_index[index];
445 memcpy(&io_u->issue_time, &now, sizeof(now));
446 io_u_queued(td, io_u);
452 static int fio_ioring_commit(struct thread_data *td)
454 struct ioring_data *ld = td->io_ops_data;
455 struct ioring_options *o = td->eo;
462 * Kernel side does submission. just need to check if the ring is
463 * flagged as needing a kick, if so, call io_uring_enter(). This
464 * only happens if we've been idle too long.
466 if (o->sqpoll_thread) {
467 struct io_sq_ring *ring = &ld->sq_ring;
470 flags = atomic_load_acquire(ring->flags);
471 if (flags & IORING_SQ_NEED_WAKEUP)
472 io_uring_enter(ld, ld->queued, 0,
473 IORING_ENTER_SQ_WAKEUP);
479 unsigned start = *ld->sq_ring.head;
480 long nr = ld->queued;
482 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
484 fio_ioring_queued(td, start, ret);
485 io_u_mark_submit(td, ret);
490 io_u_mark_submit(td, ret);
493 if (errno == EAGAIN || errno == EINTR) {
494 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
497 /* Shouldn't happen */
501 td_verror(td, errno, "io_uring_enter submit");
504 } while (ld->queued);
509 static void fio_ioring_unmap(struct ioring_data *ld)
513 for (i = 0; i < FIO_ARRAY_SIZE(ld->mmap); i++)
514 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
518 static void fio_ioring_cleanup(struct thread_data *td)
520 struct ioring_data *ld = td->io_ops_data;
523 if (!(td->flags & TD_F_CHILD))
524 fio_ioring_unmap(ld);
526 free(ld->io_u_index);
533 static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
535 struct io_sq_ring *sring = &ld->sq_ring;
536 struct io_cq_ring *cring = &ld->cq_ring;
539 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
540 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
541 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
543 ld->mmap[0].ptr = ptr;
544 sring->head = ptr + p->sq_off.head;
545 sring->tail = ptr + p->sq_off.tail;
546 sring->ring_mask = ptr + p->sq_off.ring_mask;
547 sring->ring_entries = ptr + p->sq_off.ring_entries;
548 sring->flags = ptr + p->sq_off.flags;
549 sring->array = ptr + p->sq_off.array;
550 ld->sq_ring_mask = *sring->ring_mask;
552 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
553 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
554 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
556 ld->mmap[1].ptr = ld->sqes;
558 ld->mmap[2].len = p->cq_off.cqes +
559 p->cq_entries * sizeof(struct io_uring_cqe);
560 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
561 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
563 ld->mmap[2].ptr = ptr;
564 cring->head = ptr + p->cq_off.head;
565 cring->tail = ptr + p->cq_off.tail;
566 cring->ring_mask = ptr + p->cq_off.ring_mask;
567 cring->ring_entries = ptr + p->cq_off.ring_entries;
568 cring->cqes = ptr + p->cq_off.cqes;
569 ld->cq_ring_mask = *cring->ring_mask;
573 static void fio_ioring_probe(struct thread_data *td)
575 struct ioring_data *ld = td->io_ops_data;
576 struct ioring_options *o = td->eo;
577 struct io_uring_probe *p;
580 /* already set by user, don't touch */
581 if (o->nonvectored != -1)
584 /* default to off, as that's always safe */
587 p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
591 memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
592 ret = syscall(__NR_io_uring_register, ld->ring_fd,
593 IORING_REGISTER_PROBE, p, 256);
597 if (IORING_OP_WRITE > p->ops_len)
600 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) &&
601 (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED))
607 static int fio_ioring_queue_init(struct thread_data *td)
609 struct ioring_data *ld = td->io_ops_data;
610 struct ioring_options *o = td->eo;
611 int depth = td->o.iodepth;
612 struct io_uring_params p;
615 memset(&p, 0, sizeof(p));
618 p.flags |= IORING_SETUP_IOPOLL;
619 if (o->sqpoll_thread) {
620 p.flags |= IORING_SETUP_SQPOLL;
622 p.flags |= IORING_SETUP_SQ_AFF;
623 p.sq_thread_cpu = o->sqpoll_cpu;
627 ret = syscall(__NR_io_uring_setup, depth, &p);
633 fio_ioring_probe(td);
636 ret = syscall(__NR_io_uring_register, ld->ring_fd,
637 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
642 return fio_ioring_mmap(ld, &p);
645 static int fio_ioring_register_files(struct thread_data *td)
647 struct ioring_data *ld = td->io_ops_data;
652 ld->fds = calloc(td->o.nr_files, sizeof(int));
654 for_each_file(td, f, i) {
655 ret = generic_open_file(td, f);
662 ret = syscall(__NR_io_uring_register, ld->ring_fd,
663 IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
671 * Pretend the file is closed again, and really close it if we hit
674 for_each_file(td, f, i) {
677 ret2 = generic_close_file(td, f);
685 static int fio_ioring_post_init(struct thread_data *td)
687 struct ioring_data *ld = td->io_ops_data;
688 struct ioring_options *o = td->eo;
692 for (i = 0; i < td->o.iodepth; i++) {
693 struct iovec *iov = &ld->iovecs[i];
695 io_u = ld->io_u_index[i];
696 iov->iov_base = io_u->buf;
697 iov->iov_len = td_max_bs(td);
700 err = fio_ioring_queue_init(td);
702 int init_err = errno;
704 if (init_err == ENOSYS)
705 log_err("fio: your kernel doesn't support io_uring\n");
706 td_verror(td, init_err, "io_queue_init");
710 for (i = 0; i < td->o.iodepth; i++) {
711 struct io_uring_sqe *sqe;
714 memset(sqe, 0, sizeof(*sqe));
717 if (o->registerfiles) {
718 err = fio_ioring_register_files(td);
720 td_verror(td, errno, "ioring_register_files");
728 static int fio_ioring_init(struct thread_data *td)
730 struct ioring_options *o = td->eo;
731 struct ioring_data *ld;
732 struct thread_options *to = &td->o;
734 /* sqthread submission requires registered files */
735 if (o->sqpoll_thread)
736 o->registerfiles = 1;
738 if (o->registerfiles && td->o.nr_files != td->o.open_files) {
739 log_err("fio: io_uring registered files require nr_files to "
740 "be identical to open_files\n");
744 ld = calloc(1, sizeof(*ld));
746 /* ring depth must be a power-of-2 */
747 ld->iodepth = td->o.iodepth;
748 td->o.iodepth = roundup_pow2(td->o.iodepth);
751 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
752 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
754 td->io_ops_data = ld;
757 * Check for option conflicts
759 if ((fio_option_is_set(to, ioprio) || fio_option_is_set(to, ioprio_class)) &&
760 o->cmdprio_percentage != 0) {
761 log_err("%s: cmdprio_percentage option and mutually exclusive "
762 "prio or prioclass option is set, exiting\n", to->name);
763 td_verror(td, EINVAL, "fio_io_uring_init");
767 if (fio_option_is_set(&td->o, ioprio_class))
768 ld->ioprio_class_set = true;
769 if (fio_option_is_set(&td->o, ioprio))
770 ld->ioprio_set = true;
775 static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
777 struct ioring_data *ld = td->io_ops_data;
779 ld->io_u_index[io_u->index] = io_u;
783 static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
785 struct ioring_data *ld = td->io_ops_data;
786 struct ioring_options *o = td->eo;
788 if (!ld || !o->registerfiles)
789 return generic_open_file(td, f);
791 f->fd = ld->fds[f->engine_pos];
795 static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
797 struct ioring_data *ld = td->io_ops_data;
798 struct ioring_options *o = td->eo;
800 if (!ld || !o->registerfiles)
801 return generic_close_file(td, f);
807 static struct ioengine_ops ioengine = {
809 .version = FIO_IOOPS_VERSION,
810 .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD,
811 .init = fio_ioring_init,
812 .post_init = fio_ioring_post_init,
813 .io_u_init = fio_ioring_io_u_init,
814 .prep = fio_ioring_prep,
815 .queue = fio_ioring_queue,
816 .commit = fio_ioring_commit,
817 .getevents = fio_ioring_getevents,
818 .event = fio_ioring_event,
819 .cleanup = fio_ioring_cleanup,
820 .open_file = fio_ioring_open_file,
821 .close_file = fio_ioring_close_file,
822 .get_file_size = generic_get_file_size,
824 .option_struct_size = sizeof(struct ioring_options),
827 static void fio_init fio_ioring_register(void)
829 register_ioengine(&ioengine);
832 static void fio_exit fio_ioring_unregister(void)
834 unregister_ioengine(&ioengine);