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"
32 unsigned *ring_entries;
41 unsigned *ring_entries;
42 struct io_uring_cqe *cqes;
53 struct io_u **io_u_index;
57 struct io_sq_ring sq_ring;
58 struct io_uring_sqe *sqes;
60 unsigned sq_ring_mask;
62 struct io_cq_ring cq_ring;
63 unsigned cq_ring_mask;
70 struct ioring_mmap mmap[3];
75 struct ioring_options {
76 struct thread_data *td;
78 struct cmdprio cmdprio;
79 unsigned int fixedbufs;
80 unsigned int registerfiles;
81 unsigned int sqpoll_thread;
82 unsigned int sqpoll_set;
83 unsigned int sqpoll_cpu;
84 unsigned int nonvectored;
85 unsigned int uncached;
87 unsigned int force_async;
90 static const int ddir_to_op[2][2] = {
91 { IORING_OP_READV, IORING_OP_READ },
92 { IORING_OP_WRITEV, IORING_OP_WRITE }
95 static const int fixed_ddir_to_op[2] = {
100 static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val)
102 struct ioring_options *o = data;
104 o->sqpoll_cpu = *val;
109 static int str_cmdprio_bssplit_cb(void *data, const char *input)
111 struct ioring_options *o = data;
112 struct thread_data *td = o->td;
113 struct cmdprio *cmdprio = &o->cmdprio;
115 return fio_cmdprio_bssplit_parse(td, input, cmdprio);
118 static struct fio_option options[] = {
121 .lname = "High Priority",
122 .type = FIO_OPT_STR_SET,
123 .off1 = offsetof(struct ioring_options, hipri),
124 .help = "Use polled IO completions",
125 .category = FIO_OPT_C_ENGINE,
126 .group = FIO_OPT_G_IOURING,
128 #ifdef FIO_HAVE_IOPRIO_CLASS
130 .name = "cmdprio_percentage",
131 .lname = "high priority percentage",
133 .off1 = offsetof(struct ioring_options,
134 cmdprio.percentage[DDIR_READ]),
135 .off2 = offsetof(struct ioring_options,
136 cmdprio.percentage[DDIR_WRITE]),
139 .help = "Send high priority I/O this percentage of the time",
140 .category = FIO_OPT_C_ENGINE,
141 .group = FIO_OPT_G_IOURING,
144 .name = "cmdprio_class",
145 .lname = "Asynchronous I/O priority class",
147 .off1 = offsetof(struct ioring_options,
148 cmdprio.class[DDIR_READ]),
149 .off2 = offsetof(struct ioring_options,
150 cmdprio.class[DDIR_WRITE]),
151 .help = "Set asynchronous IO priority class",
152 .minval = IOPRIO_MIN_PRIO_CLASS + 1,
153 .maxval = IOPRIO_MAX_PRIO_CLASS,
155 .category = FIO_OPT_C_ENGINE,
156 .group = FIO_OPT_G_IOURING,
160 .lname = "Asynchronous I/O priority level",
162 .off1 = offsetof(struct ioring_options,
163 cmdprio.level[DDIR_READ]),
164 .off2 = offsetof(struct ioring_options,
165 cmdprio.level[DDIR_WRITE]),
166 .help = "Set asynchronous IO priority level",
167 .minval = IOPRIO_MIN_PRIO,
168 .maxval = IOPRIO_MAX_PRIO,
170 .category = FIO_OPT_C_ENGINE,
171 .group = FIO_OPT_G_IOURING,
174 .name = "cmdprio_bssplit",
175 .lname = "Priority percentage block size split",
176 .type = FIO_OPT_STR_ULL,
177 .cb = str_cmdprio_bssplit_cb,
178 .off1 = offsetof(struct ioring_options, cmdprio.bssplit),
179 .help = "Set priority percentages for different block sizes",
180 .category = FIO_OPT_C_ENGINE,
181 .group = FIO_OPT_G_IOURING,
185 .name = "cmdprio_percentage",
186 .lname = "high priority percentage",
187 .type = FIO_OPT_UNSUPPORTED,
188 .help = "Your platform does not support I/O priority classes",
191 .name = "cmdprio_class",
192 .lname = "Asynchronous I/O priority class",
193 .type = FIO_OPT_UNSUPPORTED,
194 .help = "Your platform does not support I/O priority classes",
198 .lname = "Asynchronous I/O priority level",
199 .type = FIO_OPT_UNSUPPORTED,
200 .help = "Your platform does not support I/O priority classes",
203 .name = "cmdprio_bssplit",
204 .lname = "Priority percentage block size split",
205 .type = FIO_OPT_UNSUPPORTED,
206 .help = "Your platform does not support I/O priority classes",
211 .lname = "Fixed (pre-mapped) IO buffers",
212 .type = FIO_OPT_STR_SET,
213 .off1 = offsetof(struct ioring_options, fixedbufs),
214 .help = "Pre map IO buffers",
215 .category = FIO_OPT_C_ENGINE,
216 .group = FIO_OPT_G_IOURING,
219 .name = "registerfiles",
220 .lname = "Register file set",
221 .type = FIO_OPT_STR_SET,
222 .off1 = offsetof(struct ioring_options, registerfiles),
223 .help = "Pre-open/register files",
224 .category = FIO_OPT_C_ENGINE,
225 .group = FIO_OPT_G_IOURING,
228 .name = "sqthread_poll",
229 .lname = "Kernel SQ thread polling",
231 .off1 = offsetof(struct ioring_options, sqpoll_thread),
232 .help = "Offload submission/completion to kernel thread",
233 .category = FIO_OPT_C_ENGINE,
234 .group = FIO_OPT_G_IOURING,
237 .name = "sqthread_poll_cpu",
238 .lname = "SQ Thread Poll CPU",
240 .cb = fio_ioring_sqpoll_cb,
241 .help = "What CPU to run SQ thread polling on",
242 .category = FIO_OPT_C_ENGINE,
243 .group = FIO_OPT_G_IOURING,
246 .name = "nonvectored",
247 .lname = "Non-vectored",
249 .off1 = offsetof(struct ioring_options, nonvectored),
251 .help = "Use non-vectored read/write commands",
252 .category = FIO_OPT_C_ENGINE,
253 .group = FIO_OPT_G_IOURING,
259 .off1 = offsetof(struct ioring_options, uncached),
260 .help = "Use RWF_UNCACHED for buffered read/writes",
261 .category = FIO_OPT_C_ENGINE,
262 .group = FIO_OPT_G_IOURING,
266 .lname = "RWF_NOWAIT",
267 .type = FIO_OPT_BOOL,
268 .off1 = offsetof(struct ioring_options, nowait),
269 .help = "Use RWF_NOWAIT for reads/writes",
270 .category = FIO_OPT_C_ENGINE,
271 .group = FIO_OPT_G_IOURING,
274 .name = "force_async",
275 .lname = "Force async",
277 .off1 = offsetof(struct ioring_options, force_async),
278 .help = "Set IOSQE_ASYNC every N requests",
279 .category = FIO_OPT_C_ENGINE,
280 .group = FIO_OPT_G_IOURING,
287 static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
288 unsigned int min_complete, unsigned int flags)
290 return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit,
291 min_complete, flags, NULL, 0);
294 static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
296 struct ioring_data *ld = td->io_ops_data;
297 struct ioring_options *o = td->eo;
298 struct fio_file *f = io_u->file;
299 struct io_uring_sqe *sqe;
301 sqe = &ld->sqes[io_u->index];
303 if (o->registerfiles) {
304 sqe->fd = f->engine_pos;
305 sqe->flags = IOSQE_FIXED_FILE;
311 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
313 sqe->opcode = fixed_ddir_to_op[io_u->ddir];
314 sqe->addr = (unsigned long) io_u->xfer_buf;
315 sqe->len = io_u->xfer_buflen;
316 sqe->buf_index = io_u->index;
318 struct iovec *iov = &ld->iovecs[io_u->index];
321 * Update based on actual io_u, requeue could have
324 iov->iov_base = io_u->xfer_buf;
325 iov->iov_len = io_u->xfer_buflen;
327 sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored];
328 if (o->nonvectored) {
329 sqe->addr = (unsigned long) iov->iov_base;
330 sqe->len = iov->iov_len;
332 sqe->addr = (unsigned long) iov;
337 if (!td->o.odirect && o->uncached)
338 sqe->rw_flags |= RWF_UNCACHED;
340 sqe->rw_flags |= RWF_NOWAIT;
343 * Since io_uring can have a submission context (sqthread_poll)
344 * that is different from the process context, we cannot rely on
345 * the IO priority set by ioprio_set() (option prio/prioclass)
347 * td->ioprio will have the value of the "default prio", so set
348 * this unconditionally. This value might get overridden by
349 * fio_ioring_prio_prep() if the option cmdprio_percentage or
350 * cmdprio_bssplit is used.
352 sqe->ioprio = td->ioprio;
353 sqe->off = io_u->offset;
354 } else if (ddir_sync(io_u->ddir)) {
356 if (io_u->ddir == DDIR_SYNC_FILE_RANGE) {
357 sqe->off = f->first_write;
358 sqe->len = f->last_write - f->first_write;
359 sqe->sync_range_flags = td->o.sync_file_range;
360 sqe->opcode = IORING_OP_SYNC_FILE_RANGE;
365 if (io_u->ddir == DDIR_DATASYNC)
366 sqe->fsync_flags |= IORING_FSYNC_DATASYNC;
367 sqe->opcode = IORING_OP_FSYNC;
371 if (o->force_async && ++ld->prepped == o->force_async) {
373 sqe->flags |= IOSQE_ASYNC;
376 sqe->user_data = (unsigned long) io_u;
380 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
382 struct ioring_data *ld = td->io_ops_data;
383 struct io_uring_cqe *cqe;
387 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
389 cqe = &ld->cq_ring.cqes[index];
390 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
392 if (cqe->res != io_u->xfer_buflen) {
393 if (cqe->res > io_u->xfer_buflen)
394 io_u->error = -cqe->res;
396 io_u->resid = io_u->xfer_buflen - cqe->res;
403 static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
406 struct ioring_data *ld = td->io_ops_data;
407 struct io_cq_ring *ring = &ld->cq_ring;
408 unsigned head, reaped = 0;
412 if (head == atomic_load_acquire(ring->tail))
416 } while (reaped + events < max);
419 atomic_store_release(ring->head, head);
424 static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
425 unsigned int max, const struct timespec *t)
427 struct ioring_data *ld = td->io_ops_data;
428 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
429 struct ioring_options *o = td->eo;
430 struct io_cq_ring *ring = &ld->cq_ring;
434 ld->cq_ring_off = *ring->head;
436 r = fio_ioring_cqring_reap(td, events, max);
444 if (!o->sqpoll_thread) {
445 r = io_uring_enter(ld, 0, actual_min,
446 IORING_ENTER_GETEVENTS);
448 if (errno == EAGAIN || errno == EINTR)
450 td_verror(td, errno, "io_uring_enter");
454 } while (events < min);
456 return r < 0 ? r : events;
459 static void fio_ioring_prio_prep(struct thread_data *td, struct io_u *io_u)
461 struct ioring_options *o = td->eo;
462 struct ioring_data *ld = td->io_ops_data;
463 struct io_uring_sqe *sqe = &ld->sqes[io_u->index];
464 struct cmdprio *cmdprio = &o->cmdprio;
465 enum fio_ddir ddir = io_u->ddir;
466 unsigned int p = fio_cmdprio_percentage(cmdprio, io_u);
467 unsigned int cmdprio_value =
468 ioprio_value(cmdprio->class[ddir], cmdprio->level[ddir]);
470 if (p && rand_between(&td->prio_state, 0, 99) < p) {
471 io_u->ioprio = cmdprio_value;
472 sqe->ioprio = cmdprio_value;
473 if (!td->ioprio || cmdprio_value < td->ioprio) {
475 * The async IO priority is higher (has a lower value)
476 * than the default priority (which is either 0 or the
477 * value set by "prio" and "prioclass" options).
479 io_u->flags |= IO_U_F_HIGH_PRIO;
481 } else if (td->ioprio && td->ioprio < cmdprio_value) {
483 * The IO will be executed with the default priority (which is
484 * either 0 or the value set by "prio" and "prioclass options),
485 * and this priority is higher (has a lower value) than the
488 io_u->flags |= IO_U_F_HIGH_PRIO;
492 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
495 struct ioring_data *ld = td->io_ops_data;
496 struct io_sq_ring *ring = &ld->sq_ring;
497 unsigned tail, next_tail;
499 fio_ro_check(td, io_u);
501 if (ld->queued == ld->iodepth)
504 if (io_u->ddir == DDIR_TRIM) {
508 do_io_u_trim(td, io_u);
509 io_u_mark_submit(td, 1);
510 io_u_mark_complete(td, 1);
511 return FIO_Q_COMPLETED;
515 next_tail = tail + 1;
516 if (next_tail == atomic_load_acquire(ring->head))
520 fio_ioring_prio_prep(td, io_u);
521 ring->array[tail & ld->sq_ring_mask] = io_u->index;
522 atomic_store_release(ring->tail, next_tail);
528 static void fio_ioring_queued(struct thread_data *td, int start, int nr)
530 struct ioring_data *ld = td->io_ops_data;
533 if (!fio_fill_issue_time(td))
536 fio_gettime(&now, NULL);
539 struct io_sq_ring *ring = &ld->sq_ring;
540 int index = ring->array[start & ld->sq_ring_mask];
541 struct io_u *io_u = ld->io_u_index[index];
543 memcpy(&io_u->issue_time, &now, sizeof(now));
544 io_u_queued(td, io_u);
550 static int fio_ioring_commit(struct thread_data *td)
552 struct ioring_data *ld = td->io_ops_data;
553 struct ioring_options *o = td->eo;
560 * Kernel side does submission. just need to check if the ring is
561 * flagged as needing a kick, if so, call io_uring_enter(). This
562 * only happens if we've been idle too long.
564 if (o->sqpoll_thread) {
565 struct io_sq_ring *ring = &ld->sq_ring;
568 flags = atomic_load_acquire(ring->flags);
569 if (flags & IORING_SQ_NEED_WAKEUP)
570 io_uring_enter(ld, ld->queued, 0,
571 IORING_ENTER_SQ_WAKEUP);
577 unsigned start = *ld->sq_ring.head;
578 long nr = ld->queued;
580 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
582 fio_ioring_queued(td, start, ret);
583 io_u_mark_submit(td, ret);
588 io_u_mark_submit(td, ret);
591 if (errno == EAGAIN || errno == EINTR) {
592 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
595 /* Shouldn't happen */
599 td_verror(td, errno, "io_uring_enter submit");
602 } while (ld->queued);
607 static void fio_ioring_unmap(struct ioring_data *ld)
611 for (i = 0; i < FIO_ARRAY_SIZE(ld->mmap); i++)
612 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
616 static void fio_ioring_cleanup(struct thread_data *td)
618 struct ioring_data *ld = td->io_ops_data;
621 if (!(td->flags & TD_F_CHILD))
622 fio_ioring_unmap(ld);
624 free(ld->io_u_index);
631 static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
633 struct io_sq_ring *sring = &ld->sq_ring;
634 struct io_cq_ring *cring = &ld->cq_ring;
637 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
638 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
639 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
641 ld->mmap[0].ptr = ptr;
642 sring->head = ptr + p->sq_off.head;
643 sring->tail = ptr + p->sq_off.tail;
644 sring->ring_mask = ptr + p->sq_off.ring_mask;
645 sring->ring_entries = ptr + p->sq_off.ring_entries;
646 sring->flags = ptr + p->sq_off.flags;
647 sring->array = ptr + p->sq_off.array;
648 ld->sq_ring_mask = *sring->ring_mask;
650 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
651 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
652 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
654 ld->mmap[1].ptr = ld->sqes;
656 ld->mmap[2].len = p->cq_off.cqes +
657 p->cq_entries * sizeof(struct io_uring_cqe);
658 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
659 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
661 ld->mmap[2].ptr = ptr;
662 cring->head = ptr + p->cq_off.head;
663 cring->tail = ptr + p->cq_off.tail;
664 cring->ring_mask = ptr + p->cq_off.ring_mask;
665 cring->ring_entries = ptr + p->cq_off.ring_entries;
666 cring->cqes = ptr + p->cq_off.cqes;
667 ld->cq_ring_mask = *cring->ring_mask;
671 static void fio_ioring_probe(struct thread_data *td)
673 struct ioring_data *ld = td->io_ops_data;
674 struct ioring_options *o = td->eo;
675 struct io_uring_probe *p;
678 /* already set by user, don't touch */
679 if (o->nonvectored != -1)
682 /* default to off, as that's always safe */
685 p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
689 memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
690 ret = syscall(__NR_io_uring_register, ld->ring_fd,
691 IORING_REGISTER_PROBE, p, 256);
695 if (IORING_OP_WRITE > p->ops_len)
698 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) &&
699 (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED))
705 static int fio_ioring_queue_init(struct thread_data *td)
707 struct ioring_data *ld = td->io_ops_data;
708 struct ioring_options *o = td->eo;
709 int depth = td->o.iodepth;
710 struct io_uring_params p;
713 memset(&p, 0, sizeof(p));
716 p.flags |= IORING_SETUP_IOPOLL;
717 if (o->sqpoll_thread) {
718 p.flags |= IORING_SETUP_SQPOLL;
720 p.flags |= IORING_SETUP_SQ_AFF;
721 p.sq_thread_cpu = o->sqpoll_cpu;
725 ret = syscall(__NR_io_uring_setup, depth, &p);
731 fio_ioring_probe(td);
734 ret = syscall(__NR_io_uring_register, ld->ring_fd,
735 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
740 return fio_ioring_mmap(ld, &p);
743 static int fio_ioring_register_files(struct thread_data *td)
745 struct ioring_data *ld = td->io_ops_data;
750 ld->fds = calloc(td->o.nr_files, sizeof(int));
752 for_each_file(td, f, i) {
753 ret = generic_open_file(td, f);
760 ret = syscall(__NR_io_uring_register, ld->ring_fd,
761 IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
769 * Pretend the file is closed again, and really close it if we hit
772 for_each_file(td, f, i) {
775 ret2 = generic_close_file(td, f);
783 static int fio_ioring_post_init(struct thread_data *td)
785 struct ioring_data *ld = td->io_ops_data;
786 struct ioring_options *o = td->eo;
790 for (i = 0; i < td->o.iodepth; i++) {
791 struct iovec *iov = &ld->iovecs[i];
793 io_u = ld->io_u_index[i];
794 iov->iov_base = io_u->buf;
795 iov->iov_len = td_max_bs(td);
798 err = fio_ioring_queue_init(td);
800 int init_err = errno;
802 if (init_err == ENOSYS)
803 log_err("fio: your kernel doesn't support io_uring\n");
804 td_verror(td, init_err, "io_queue_init");
808 for (i = 0; i < td->o.iodepth; i++) {
809 struct io_uring_sqe *sqe;
812 memset(sqe, 0, sizeof(*sqe));
815 if (o->registerfiles) {
816 err = fio_ioring_register_files(td);
818 td_verror(td, errno, "ioring_register_files");
826 static int fio_ioring_init(struct thread_data *td)
828 struct ioring_options *o = td->eo;
829 struct ioring_data *ld;
830 struct cmdprio *cmdprio = &o->cmdprio;
833 /* sqthread submission requires registered files */
834 if (o->sqpoll_thread)
835 o->registerfiles = 1;
837 if (o->registerfiles && td->o.nr_files != td->o.open_files) {
838 log_err("fio: io_uring registered files require nr_files to "
839 "be identical to open_files\n");
843 ld = calloc(1, sizeof(*ld));
845 /* ring depth must be a power-of-2 */
846 ld->iodepth = td->o.iodepth;
847 td->o.iodepth = roundup_pow2(td->o.iodepth);
850 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
851 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
853 td->io_ops_data = ld;
855 ret = fio_cmdprio_init(td, cmdprio, &ld->use_cmdprio);
857 td_verror(td, EINVAL, "fio_ioring_init");
864 static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
866 struct ioring_data *ld = td->io_ops_data;
868 ld->io_u_index[io_u->index] = io_u;
872 static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
874 struct ioring_data *ld = td->io_ops_data;
875 struct ioring_options *o = td->eo;
877 if (!ld || !o->registerfiles)
878 return generic_open_file(td, f);
880 f->fd = ld->fds[f->engine_pos];
884 static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
886 struct ioring_data *ld = td->io_ops_data;
887 struct ioring_options *o = td->eo;
889 if (!ld || !o->registerfiles)
890 return generic_close_file(td, f);
896 static struct ioengine_ops ioengine = {
898 .version = FIO_IOOPS_VERSION,
899 .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD,
900 .init = fio_ioring_init,
901 .post_init = fio_ioring_post_init,
902 .io_u_init = fio_ioring_io_u_init,
903 .prep = fio_ioring_prep,
904 .queue = fio_ioring_queue,
905 .commit = fio_ioring_commit,
906 .getevents = fio_ioring_getevents,
907 .event = fio_ioring_event,
908 .cleanup = fio_ioring_cleanup,
909 .open_file = fio_ioring_open_file,
910 .close_file = fio_ioring_close_file,
911 .get_file_size = generic_get_file_size,
913 .option_struct_size = sizeof(struct ioring_options),
916 static void fio_init fio_ioring_register(void)
918 register_ioengine(&ioengine);
921 static void fio_exit fio_ioring_unregister(void)
923 unregister_ioengine(&ioengine);