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;
70 struct ioring_mmap mmap[3];
73 struct ioring_options {
76 unsigned int cmdprio_percentage;
77 unsigned int fixedbufs;
78 unsigned int registerfiles;
79 unsigned int sqpoll_thread;
80 unsigned int sqpoll_set;
81 unsigned int sqpoll_cpu;
82 unsigned int nonvectored;
83 unsigned int uncached;
87 static const int ddir_to_op[2][2] = {
88 { IORING_OP_READV, IORING_OP_READ },
89 { IORING_OP_WRITEV, IORING_OP_WRITE }
92 static const int fixed_ddir_to_op[2] = {
97 static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val)
99 struct ioring_options *o = data;
101 o->sqpoll_cpu = *val;
106 static struct fio_option options[] = {
109 .lname = "High Priority",
110 .type = FIO_OPT_STR_SET,
111 .off1 = offsetof(struct ioring_options, hipri),
112 .help = "Use polled IO completions",
113 .category = FIO_OPT_C_ENGINE,
114 .group = FIO_OPT_G_IOURING,
116 #ifdef FIO_HAVE_IOPRIO_CLASS
118 .name = "cmdprio_percentage",
119 .lname = "high priority percentage",
121 .off1 = offsetof(struct ioring_options, cmdprio_percentage),
124 .help = "Send high priority I/O this percentage of the time",
125 .category = FIO_OPT_C_ENGINE,
126 .group = FIO_OPT_G_IOURING,
130 .name = "cmdprio_percentage",
131 .lname = "high priority percentage",
132 .type = FIO_OPT_UNSUPPORTED,
133 .help = "Your platform does not support I/O priority classes",
138 .lname = "Fixed (pre-mapped) IO buffers",
139 .type = FIO_OPT_STR_SET,
140 .off1 = offsetof(struct ioring_options, fixedbufs),
141 .help = "Pre map IO buffers",
142 .category = FIO_OPT_C_ENGINE,
143 .group = FIO_OPT_G_IOURING,
146 .name = "registerfiles",
147 .lname = "Register file set",
148 .type = FIO_OPT_STR_SET,
149 .off1 = offsetof(struct ioring_options, registerfiles),
150 .help = "Pre-open/register files",
151 .category = FIO_OPT_C_ENGINE,
152 .group = FIO_OPT_G_IOURING,
155 .name = "sqthread_poll",
156 .lname = "Kernel SQ thread polling",
158 .off1 = offsetof(struct ioring_options, sqpoll_thread),
159 .help = "Offload submission/completion to kernel thread",
160 .category = FIO_OPT_C_ENGINE,
161 .group = FIO_OPT_G_IOURING,
164 .name = "sqthread_poll_cpu",
165 .lname = "SQ Thread Poll CPU",
167 .cb = fio_ioring_sqpoll_cb,
168 .help = "What CPU to run SQ thread polling on",
169 .category = FIO_OPT_C_ENGINE,
170 .group = FIO_OPT_G_IOURING,
173 .name = "nonvectored",
174 .lname = "Non-vectored",
176 .off1 = offsetof(struct ioring_options, nonvectored),
178 .help = "Use non-vectored read/write commands",
179 .category = FIO_OPT_C_ENGINE,
180 .group = FIO_OPT_G_IOURING,
186 .off1 = offsetof(struct ioring_options, uncached),
187 .help = "Use RWF_UNCACHED for buffered read/writes",
188 .category = FIO_OPT_C_ENGINE,
189 .group = FIO_OPT_G_IOURING,
193 .lname = "RWF_NOWAIT",
194 .type = FIO_OPT_BOOL,
195 .off1 = offsetof(struct ioring_options, nowait),
196 .help = "Use RWF_NOWAIT for reads/writes",
197 .category = FIO_OPT_C_ENGINE,
198 .group = FIO_OPT_G_IOURING,
205 static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
206 unsigned int min_complete, unsigned int flags)
208 return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit,
209 min_complete, flags, NULL, 0);
212 static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
214 struct ioring_data *ld = td->io_ops_data;
215 struct ioring_options *o = td->eo;
216 struct fio_file *f = io_u->file;
217 struct io_uring_sqe *sqe;
219 sqe = &ld->sqes[io_u->index];
221 /* zero out fields not used in this submission */
222 memset(sqe, 0, sizeof(*sqe));
224 if (o->registerfiles) {
225 sqe->fd = f->engine_pos;
226 sqe->flags = IOSQE_FIXED_FILE;
231 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
233 sqe->opcode = fixed_ddir_to_op[io_u->ddir];
234 sqe->addr = (unsigned long) io_u->xfer_buf;
235 sqe->len = io_u->xfer_buflen;
236 sqe->buf_index = io_u->index;
238 struct iovec *iov = &ld->iovecs[io_u->index];
241 * Update based on actual io_u, requeue could have
244 iov->iov_base = io_u->xfer_buf;
245 iov->iov_len = io_u->xfer_buflen;
247 sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored];
248 if (o->nonvectored) {
249 sqe->addr = (unsigned long) iov->iov_base;
250 sqe->len = iov->iov_len;
252 sqe->addr = (unsigned long) iov;
256 if (!td->o.odirect && o->uncached)
257 sqe->rw_flags = RWF_UNCACHED;
259 sqe->rw_flags |= RWF_NOWAIT;
260 if (ld->ioprio_class_set)
261 sqe->ioprio = td->o.ioprio_class << 13;
263 sqe->ioprio |= td->o.ioprio;
264 sqe->off = io_u->offset;
265 } else if (ddir_sync(io_u->ddir)) {
266 if (io_u->ddir == DDIR_SYNC_FILE_RANGE) {
267 sqe->off = f->first_write;
268 sqe->len = f->last_write - f->first_write;
269 sqe->sync_range_flags = td->o.sync_file_range;
270 sqe->opcode = IORING_OP_SYNC_FILE_RANGE;
272 if (io_u->ddir == DDIR_DATASYNC)
273 sqe->fsync_flags |= IORING_FSYNC_DATASYNC;
274 sqe->opcode = IORING_OP_FSYNC;
278 sqe->user_data = (unsigned long) io_u;
282 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
284 struct ioring_data *ld = td->io_ops_data;
285 struct io_uring_cqe *cqe;
289 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
291 cqe = &ld->cq_ring.cqes[index];
292 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
294 if (cqe->res != io_u->xfer_buflen) {
295 if (cqe->res > io_u->xfer_buflen)
296 io_u->error = -cqe->res;
298 io_u->resid = io_u->xfer_buflen - cqe->res;
305 static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
308 struct ioring_data *ld = td->io_ops_data;
309 struct io_cq_ring *ring = &ld->cq_ring;
310 unsigned head, reaped = 0;
314 if (head == atomic_load_acquire(ring->tail))
318 } while (reaped + events < max);
321 atomic_store_release(ring->head, head);
326 static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
327 unsigned int max, const struct timespec *t)
329 struct ioring_data *ld = td->io_ops_data;
330 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
331 struct ioring_options *o = td->eo;
332 struct io_cq_ring *ring = &ld->cq_ring;
336 ld->cq_ring_off = *ring->head;
338 r = fio_ioring_cqring_reap(td, events, max);
346 if (!o->sqpoll_thread) {
347 r = io_uring_enter(ld, 0, actual_min,
348 IORING_ENTER_GETEVENTS);
350 if (errno == EAGAIN || errno == EINTR)
352 td_verror(td, errno, "io_uring_enter");
356 } while (events < min);
358 return r < 0 ? r : events;
361 static void fio_ioring_prio_prep(struct thread_data *td, struct io_u *io_u)
363 struct ioring_options *o = td->eo;
364 struct ioring_data *ld = td->io_ops_data;
365 if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) {
366 ld->sqes[io_u->index].ioprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT;
367 io_u->flags |= IO_U_F_PRIORITY;
372 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
375 struct ioring_data *ld = td->io_ops_data;
376 struct io_sq_ring *ring = &ld->sq_ring;
377 struct ioring_options *o = td->eo;
378 unsigned tail, next_tail;
380 fio_ro_check(td, io_u);
382 if (ld->queued == ld->iodepth)
385 if (io_u->ddir == DDIR_TRIM) {
389 do_io_u_trim(td, io_u);
390 io_u_mark_submit(td, 1);
391 io_u_mark_complete(td, 1);
392 return FIO_Q_COMPLETED;
396 next_tail = tail + 1;
397 if (next_tail == atomic_load_acquire(ring->head))
400 if (o->cmdprio_percentage)
401 fio_ioring_prio_prep(td, io_u);
402 ring->array[tail & ld->sq_ring_mask] = io_u->index;
403 atomic_store_release(ring->tail, next_tail);
409 static void fio_ioring_queued(struct thread_data *td, int start, int nr)
411 struct ioring_data *ld = td->io_ops_data;
414 if (!fio_fill_issue_time(td))
417 fio_gettime(&now, NULL);
420 struct io_sq_ring *ring = &ld->sq_ring;
421 int index = ring->array[start & ld->sq_ring_mask];
422 struct io_u *io_u = ld->io_u_index[index];
424 memcpy(&io_u->issue_time, &now, sizeof(now));
425 io_u_queued(td, io_u);
431 static int fio_ioring_commit(struct thread_data *td)
433 struct ioring_data *ld = td->io_ops_data;
434 struct ioring_options *o = td->eo;
441 * Kernel side does submission. just need to check if the ring is
442 * flagged as needing a kick, if so, call io_uring_enter(). This
443 * only happens if we've been idle too long.
445 if (o->sqpoll_thread) {
446 struct io_sq_ring *ring = &ld->sq_ring;
449 if (*ring->flags & IORING_SQ_NEED_WAKEUP)
450 io_uring_enter(ld, ld->queued, 0,
451 IORING_ENTER_SQ_WAKEUP);
457 unsigned start = *ld->sq_ring.head;
458 long nr = ld->queued;
460 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
462 fio_ioring_queued(td, start, ret);
463 io_u_mark_submit(td, ret);
468 io_u_mark_submit(td, ret);
471 if (errno == EAGAIN || errno == EINTR) {
472 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
475 /* Shouldn't happen */
479 td_verror(td, errno, "io_uring_enter submit");
482 } while (ld->queued);
487 static void fio_ioring_unmap(struct ioring_data *ld)
491 for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
492 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
496 static void fio_ioring_cleanup(struct thread_data *td)
498 struct ioring_data *ld = td->io_ops_data;
501 if (!(td->flags & TD_F_CHILD))
502 fio_ioring_unmap(ld);
504 free(ld->io_u_index);
511 static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
513 struct io_sq_ring *sring = &ld->sq_ring;
514 struct io_cq_ring *cring = &ld->cq_ring;
517 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
518 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
519 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
521 ld->mmap[0].ptr = ptr;
522 sring->head = ptr + p->sq_off.head;
523 sring->tail = ptr + p->sq_off.tail;
524 sring->ring_mask = ptr + p->sq_off.ring_mask;
525 sring->ring_entries = ptr + p->sq_off.ring_entries;
526 sring->flags = ptr + p->sq_off.flags;
527 sring->array = ptr + p->sq_off.array;
528 ld->sq_ring_mask = *sring->ring_mask;
530 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
531 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
532 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
534 ld->mmap[1].ptr = ld->sqes;
536 ld->mmap[2].len = p->cq_off.cqes +
537 p->cq_entries * sizeof(struct io_uring_cqe);
538 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
539 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
541 ld->mmap[2].ptr = ptr;
542 cring->head = ptr + p->cq_off.head;
543 cring->tail = ptr + p->cq_off.tail;
544 cring->ring_mask = ptr + p->cq_off.ring_mask;
545 cring->ring_entries = ptr + p->cq_off.ring_entries;
546 cring->cqes = ptr + p->cq_off.cqes;
547 ld->cq_ring_mask = *cring->ring_mask;
551 static void fio_ioring_probe(struct thread_data *td)
553 struct ioring_data *ld = td->io_ops_data;
554 struct ioring_options *o = td->eo;
555 struct io_uring_probe *p;
558 /* already set by user, don't touch */
559 if (o->nonvectored != -1)
562 /* default to off, as that's always safe */
565 p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
569 memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
570 ret = syscall(__NR_io_uring_register, ld->ring_fd,
571 IORING_REGISTER_PROBE, p, 256);
575 if (IORING_OP_WRITE > p->ops_len)
578 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) &&
579 (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED))
585 static int fio_ioring_queue_init(struct thread_data *td)
587 struct ioring_data *ld = td->io_ops_data;
588 struct ioring_options *o = td->eo;
589 int depth = td->o.iodepth;
590 struct io_uring_params p;
593 memset(&p, 0, sizeof(p));
596 p.flags |= IORING_SETUP_IOPOLL;
597 if (o->sqpoll_thread) {
598 p.flags |= IORING_SETUP_SQPOLL;
600 p.flags |= IORING_SETUP_SQ_AFF;
601 p.sq_thread_cpu = o->sqpoll_cpu;
605 ret = syscall(__NR_io_uring_setup, depth, &p);
611 fio_ioring_probe(td);
614 ret = syscall(__NR_io_uring_register, ld->ring_fd,
615 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
620 return fio_ioring_mmap(ld, &p);
623 static int fio_ioring_register_files(struct thread_data *td)
625 struct ioring_data *ld = td->io_ops_data;
630 ld->fds = calloc(td->o.nr_files, sizeof(int));
632 for_each_file(td, f, i) {
633 ret = generic_open_file(td, f);
640 ret = syscall(__NR_io_uring_register, ld->ring_fd,
641 IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
649 * Pretend the file is closed again, and really close it if we hit
652 for_each_file(td, f, i) {
655 ret2 = generic_close_file(td, f);
663 static int fio_ioring_post_init(struct thread_data *td)
665 struct ioring_data *ld = td->io_ops_data;
666 struct ioring_options *o = td->eo;
670 for (i = 0; i < td->o.iodepth; i++) {
671 struct iovec *iov = &ld->iovecs[i];
673 io_u = ld->io_u_index[i];
674 iov->iov_base = io_u->buf;
675 iov->iov_len = td_max_bs(td);
678 err = fio_ioring_queue_init(td);
680 td_verror(td, errno, "io_queue_init");
684 if (o->registerfiles) {
685 err = fio_ioring_register_files(td);
687 td_verror(td, errno, "ioring_register_files");
695 static int fio_ioring_init(struct thread_data *td)
697 struct ioring_options *o = td->eo;
698 struct ioring_data *ld;
699 struct thread_options *to = &td->o;
701 /* sqthread submission requires registered files */
702 if (o->sqpoll_thread)
703 o->registerfiles = 1;
705 if (o->registerfiles && td->o.nr_files != td->o.open_files) {
706 log_err("fio: io_uring registered files require nr_files to "
707 "be identical to open_files\n");
711 ld = calloc(1, sizeof(*ld));
713 /* ring depth must be a power-of-2 */
714 ld->iodepth = td->o.iodepth;
715 td->o.iodepth = roundup_pow2(td->o.iodepth);
718 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
719 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
721 td->io_ops_data = ld;
724 * Check for option conflicts
726 if ((fio_option_is_set(to, ioprio) || fio_option_is_set(to, ioprio_class)) &&
727 o->cmdprio_percentage != 0) {
728 log_err("%s: cmdprio_percentage option and mutually exclusive "
729 "prio or prioclass option is set, exiting\n", to->name);
730 td_verror(td, EINVAL, "fio_io_uring_init");
734 if (fio_option_is_set(&td->o, ioprio_class))
735 ld->ioprio_class_set = true;
736 if (fio_option_is_set(&td->o, ioprio))
737 ld->ioprio_set = true;
742 static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
744 struct ioring_data *ld = td->io_ops_data;
746 ld->io_u_index[io_u->index] = io_u;
750 static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
752 struct ioring_data *ld = td->io_ops_data;
753 struct ioring_options *o = td->eo;
755 if (!ld || !o->registerfiles)
756 return generic_open_file(td, f);
758 f->fd = ld->fds[f->engine_pos];
762 static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
764 struct ioring_data *ld = td->io_ops_data;
765 struct ioring_options *o = td->eo;
767 if (!ld || !o->registerfiles)
768 return generic_close_file(td, f);
774 static struct ioengine_ops ioengine = {
776 .version = FIO_IOOPS_VERSION,
777 .flags = FIO_ASYNCIO_SYNC_TRIM,
778 .init = fio_ioring_init,
779 .post_init = fio_ioring_post_init,
780 .io_u_init = fio_ioring_io_u_init,
781 .prep = fio_ioring_prep,
782 .queue = fio_ioring_queue,
783 .commit = fio_ioring_commit,
784 .getevents = fio_ioring_getevents,
785 .event = fio_ioring_event,
786 .cleanup = fio_ioring_cleanup,
787 .open_file = fio_ioring_open_file,
788 .close_file = fio_ioring_close_file,
789 .get_file_size = generic_get_file_size,
791 .option_struct_size = sizeof(struct ioring_options),
794 static void fio_init fio_ioring_register(void)
796 register_ioengine(&ioengine);
799 static void fio_exit fio_ioring_unregister(void)
801 unregister_ioengine(&ioengine);