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];
72 struct cmdprio cmdprio;
75 struct ioring_options {
76 struct thread_data *td;
78 struct cmdprio_options cmdprio_options;
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 struct fio_option options[] = {
112 .lname = "High Priority",
113 .type = FIO_OPT_STR_SET,
114 .off1 = offsetof(struct ioring_options, hipri),
115 .help = "Use polled IO completions",
116 .category = FIO_OPT_C_ENGINE,
117 .group = FIO_OPT_G_IOURING,
119 #ifdef FIO_HAVE_IOPRIO_CLASS
121 .name = "cmdprio_percentage",
122 .lname = "high priority percentage",
124 .off1 = offsetof(struct ioring_options,
125 cmdprio_options.percentage[DDIR_READ]),
126 .off2 = offsetof(struct ioring_options,
127 cmdprio_options.percentage[DDIR_WRITE]),
130 .help = "Send high priority I/O this percentage of the time",
131 .category = FIO_OPT_C_ENGINE,
132 .group = FIO_OPT_G_IOURING,
135 .name = "cmdprio_class",
136 .lname = "Asynchronous I/O priority class",
138 .off1 = offsetof(struct ioring_options,
139 cmdprio_options.class[DDIR_READ]),
140 .off2 = offsetof(struct ioring_options,
141 cmdprio_options.class[DDIR_WRITE]),
142 .help = "Set asynchronous IO priority class",
143 .minval = IOPRIO_MIN_PRIO_CLASS + 1,
144 .maxval = IOPRIO_MAX_PRIO_CLASS,
146 .category = FIO_OPT_C_ENGINE,
147 .group = FIO_OPT_G_IOURING,
151 .lname = "Asynchronous I/O priority level",
153 .off1 = offsetof(struct ioring_options,
154 cmdprio_options.level[DDIR_READ]),
155 .off2 = offsetof(struct ioring_options,
156 cmdprio_options.level[DDIR_WRITE]),
157 .help = "Set asynchronous IO priority level",
158 .minval = IOPRIO_MIN_PRIO,
159 .maxval = IOPRIO_MAX_PRIO,
161 .category = FIO_OPT_C_ENGINE,
162 .group = FIO_OPT_G_IOURING,
165 .name = "cmdprio_bssplit",
166 .lname = "Priority percentage block size split",
167 .type = FIO_OPT_STR_STORE,
168 .off1 = offsetof(struct ioring_options,
169 cmdprio_options.bssplit_str),
170 .help = "Set priority percentages for different block sizes",
171 .category = FIO_OPT_C_ENGINE,
172 .group = FIO_OPT_G_IOURING,
176 .name = "cmdprio_percentage",
177 .lname = "high priority percentage",
178 .type = FIO_OPT_UNSUPPORTED,
179 .help = "Your platform does not support I/O priority classes",
182 .name = "cmdprio_class",
183 .lname = "Asynchronous I/O priority class",
184 .type = FIO_OPT_UNSUPPORTED,
185 .help = "Your platform does not support I/O priority classes",
189 .lname = "Asynchronous I/O priority level",
190 .type = FIO_OPT_UNSUPPORTED,
191 .help = "Your platform does not support I/O priority classes",
194 .name = "cmdprio_bssplit",
195 .lname = "Priority percentage block size split",
196 .type = FIO_OPT_UNSUPPORTED,
197 .help = "Your platform does not support I/O priority classes",
202 .lname = "Fixed (pre-mapped) IO buffers",
203 .type = FIO_OPT_STR_SET,
204 .off1 = offsetof(struct ioring_options, fixedbufs),
205 .help = "Pre map IO buffers",
206 .category = FIO_OPT_C_ENGINE,
207 .group = FIO_OPT_G_IOURING,
210 .name = "registerfiles",
211 .lname = "Register file set",
212 .type = FIO_OPT_STR_SET,
213 .off1 = offsetof(struct ioring_options, registerfiles),
214 .help = "Pre-open/register files",
215 .category = FIO_OPT_C_ENGINE,
216 .group = FIO_OPT_G_IOURING,
219 .name = "sqthread_poll",
220 .lname = "Kernel SQ thread polling",
222 .off1 = offsetof(struct ioring_options, sqpoll_thread),
223 .help = "Offload submission/completion to kernel thread",
224 .category = FIO_OPT_C_ENGINE,
225 .group = FIO_OPT_G_IOURING,
228 .name = "sqthread_poll_cpu",
229 .lname = "SQ Thread Poll CPU",
231 .cb = fio_ioring_sqpoll_cb,
232 .help = "What CPU to run SQ thread polling on",
233 .category = FIO_OPT_C_ENGINE,
234 .group = FIO_OPT_G_IOURING,
237 .name = "nonvectored",
238 .lname = "Non-vectored",
240 .off1 = offsetof(struct ioring_options, nonvectored),
242 .help = "Use non-vectored read/write commands",
243 .category = FIO_OPT_C_ENGINE,
244 .group = FIO_OPT_G_IOURING,
250 .off1 = offsetof(struct ioring_options, uncached),
251 .help = "Use RWF_UNCACHED for buffered read/writes",
252 .category = FIO_OPT_C_ENGINE,
253 .group = FIO_OPT_G_IOURING,
257 .lname = "RWF_NOWAIT",
258 .type = FIO_OPT_BOOL,
259 .off1 = offsetof(struct ioring_options, nowait),
260 .help = "Use RWF_NOWAIT for reads/writes",
261 .category = FIO_OPT_C_ENGINE,
262 .group = FIO_OPT_G_IOURING,
265 .name = "force_async",
266 .lname = "Force async",
268 .off1 = offsetof(struct ioring_options, force_async),
269 .help = "Set IOSQE_ASYNC every N requests",
270 .category = FIO_OPT_C_ENGINE,
271 .group = FIO_OPT_G_IOURING,
278 static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
279 unsigned int min_complete, unsigned int flags)
281 #ifdef FIO_ARCH_HAS_SYSCALL
282 return __do_syscall6(__NR_io_uring_enter, ld->ring_fd, to_submit,
283 min_complete, flags, NULL, 0);
285 return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit,
286 min_complete, flags, NULL, 0);
290 static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
292 struct ioring_data *ld = td->io_ops_data;
293 struct ioring_options *o = td->eo;
294 struct fio_file *f = io_u->file;
295 struct io_uring_sqe *sqe;
297 sqe = &ld->sqes[io_u->index];
299 if (o->registerfiles) {
300 sqe->fd = f->engine_pos;
301 sqe->flags = IOSQE_FIXED_FILE;
307 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
309 sqe->opcode = fixed_ddir_to_op[io_u->ddir];
310 sqe->addr = (unsigned long) io_u->xfer_buf;
311 sqe->len = io_u->xfer_buflen;
312 sqe->buf_index = io_u->index;
314 struct iovec *iov = &ld->iovecs[io_u->index];
317 * Update based on actual io_u, requeue could have
320 iov->iov_base = io_u->xfer_buf;
321 iov->iov_len = io_u->xfer_buflen;
323 sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored];
324 if (o->nonvectored) {
325 sqe->addr = (unsigned long) iov->iov_base;
326 sqe->len = iov->iov_len;
328 sqe->addr = (unsigned long) iov;
333 if (!td->o.odirect && o->uncached)
334 sqe->rw_flags |= RWF_UNCACHED;
336 sqe->rw_flags |= RWF_NOWAIT;
339 * Since io_uring can have a submission context (sqthread_poll)
340 * that is different from the process context, we cannot rely on
341 * the IO priority set by ioprio_set() (option prio/prioclass)
343 * td->ioprio will have the value of the "default prio", so set
344 * this unconditionally. This value might get overridden by
345 * fio_ioring_cmdprio_prep() if the option cmdprio_percentage or
346 * cmdprio_bssplit is used.
348 sqe->ioprio = td->ioprio;
349 sqe->off = io_u->offset;
350 } else if (ddir_sync(io_u->ddir)) {
352 if (io_u->ddir == DDIR_SYNC_FILE_RANGE) {
353 sqe->off = f->first_write;
354 sqe->len = f->last_write - f->first_write;
355 sqe->sync_range_flags = td->o.sync_file_range;
356 sqe->opcode = IORING_OP_SYNC_FILE_RANGE;
361 if (io_u->ddir == DDIR_DATASYNC)
362 sqe->fsync_flags |= IORING_FSYNC_DATASYNC;
363 sqe->opcode = IORING_OP_FSYNC;
367 if (o->force_async && ++ld->prepped == o->force_async) {
369 sqe->flags |= IOSQE_ASYNC;
372 sqe->user_data = (unsigned long) io_u;
376 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
378 struct ioring_data *ld = td->io_ops_data;
379 struct io_uring_cqe *cqe;
383 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
385 cqe = &ld->cq_ring.cqes[index];
386 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
388 if (cqe->res != io_u->xfer_buflen) {
389 if (cqe->res > io_u->xfer_buflen)
390 io_u->error = -cqe->res;
392 io_u->resid = io_u->xfer_buflen - cqe->res;
399 static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
402 struct ioring_data *ld = td->io_ops_data;
403 struct io_cq_ring *ring = &ld->cq_ring;
404 unsigned head, reaped = 0;
408 if (head == atomic_load_acquire(ring->tail))
412 } while (reaped + events < max);
415 atomic_store_release(ring->head, head);
420 static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
421 unsigned int max, const struct timespec *t)
423 struct ioring_data *ld = td->io_ops_data;
424 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
425 struct ioring_options *o = td->eo;
426 struct io_cq_ring *ring = &ld->cq_ring;
430 ld->cq_ring_off = *ring->head;
432 r = fio_ioring_cqring_reap(td, events, max);
440 if (!o->sqpoll_thread) {
441 r = io_uring_enter(ld, 0, actual_min,
442 IORING_ENTER_GETEVENTS);
444 if (errno == EAGAIN || errno == EINTR)
446 td_verror(td, errno, "io_uring_enter");
450 } while (events < min);
452 return r < 0 ? r : events;
455 static inline void fio_ioring_cmdprio_prep(struct thread_data *td,
458 struct ioring_data *ld = td->io_ops_data;
459 struct cmdprio *cmdprio = &ld->cmdprio;
461 if (fio_cmdprio_set_ioprio(td, cmdprio, io_u))
462 ld->sqes[io_u->index].ioprio = io_u->ioprio;
465 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
468 struct ioring_data *ld = td->io_ops_data;
469 struct io_sq_ring *ring = &ld->sq_ring;
470 unsigned tail, next_tail;
472 fio_ro_check(td, io_u);
474 if (ld->queued == ld->iodepth)
477 if (io_u->ddir == DDIR_TRIM) {
481 do_io_u_trim(td, io_u);
482 io_u_mark_submit(td, 1);
483 io_u_mark_complete(td, 1);
484 return FIO_Q_COMPLETED;
488 next_tail = tail + 1;
489 if (next_tail == atomic_load_acquire(ring->head))
492 if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
493 fio_ioring_cmdprio_prep(td, io_u);
495 ring->array[tail & ld->sq_ring_mask] = io_u->index;
496 atomic_store_release(ring->tail, next_tail);
502 static void fio_ioring_queued(struct thread_data *td, int start, int nr)
504 struct ioring_data *ld = td->io_ops_data;
507 if (!fio_fill_issue_time(td))
510 fio_gettime(&now, NULL);
513 struct io_sq_ring *ring = &ld->sq_ring;
514 int index = ring->array[start & ld->sq_ring_mask];
515 struct io_u *io_u = ld->io_u_index[index];
517 memcpy(&io_u->issue_time, &now, sizeof(now));
518 io_u_queued(td, io_u);
524 static int fio_ioring_commit(struct thread_data *td)
526 struct ioring_data *ld = td->io_ops_data;
527 struct ioring_options *o = td->eo;
534 * Kernel side does submission. just need to check if the ring is
535 * flagged as needing a kick, if so, call io_uring_enter(). This
536 * only happens if we've been idle too long.
538 if (o->sqpoll_thread) {
539 struct io_sq_ring *ring = &ld->sq_ring;
542 flags = atomic_load_acquire(ring->flags);
543 if (flags & IORING_SQ_NEED_WAKEUP)
544 io_uring_enter(ld, ld->queued, 0,
545 IORING_ENTER_SQ_WAKEUP);
551 unsigned start = *ld->sq_ring.head;
552 long nr = ld->queued;
554 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
556 fio_ioring_queued(td, start, ret);
557 io_u_mark_submit(td, ret);
562 io_u_mark_submit(td, ret);
565 if (errno == EAGAIN || errno == EINTR) {
566 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
569 /* Shouldn't happen */
573 td_verror(td, errno, "io_uring_enter submit");
576 } while (ld->queued);
581 static void fio_ioring_unmap(struct ioring_data *ld)
585 for (i = 0; i < FIO_ARRAY_SIZE(ld->mmap); i++)
586 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
590 static void fio_ioring_cleanup(struct thread_data *td)
592 struct ioring_data *ld = td->io_ops_data;
595 if (!(td->flags & TD_F_CHILD))
596 fio_ioring_unmap(ld);
598 fio_cmdprio_cleanup(&ld->cmdprio);
599 free(ld->io_u_index);
606 static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
608 struct io_sq_ring *sring = &ld->sq_ring;
609 struct io_cq_ring *cring = &ld->cq_ring;
612 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
613 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
614 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
616 ld->mmap[0].ptr = ptr;
617 sring->head = ptr + p->sq_off.head;
618 sring->tail = ptr + p->sq_off.tail;
619 sring->ring_mask = ptr + p->sq_off.ring_mask;
620 sring->ring_entries = ptr + p->sq_off.ring_entries;
621 sring->flags = ptr + p->sq_off.flags;
622 sring->array = ptr + p->sq_off.array;
623 ld->sq_ring_mask = *sring->ring_mask;
625 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
626 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
627 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
629 ld->mmap[1].ptr = ld->sqes;
631 ld->mmap[2].len = p->cq_off.cqes +
632 p->cq_entries * sizeof(struct io_uring_cqe);
633 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
634 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
636 ld->mmap[2].ptr = ptr;
637 cring->head = ptr + p->cq_off.head;
638 cring->tail = ptr + p->cq_off.tail;
639 cring->ring_mask = ptr + p->cq_off.ring_mask;
640 cring->ring_entries = ptr + p->cq_off.ring_entries;
641 cring->cqes = ptr + p->cq_off.cqes;
642 ld->cq_ring_mask = *cring->ring_mask;
646 static void fio_ioring_probe(struct thread_data *td)
648 struct ioring_data *ld = td->io_ops_data;
649 struct ioring_options *o = td->eo;
650 struct io_uring_probe *p;
653 /* already set by user, don't touch */
654 if (o->nonvectored != -1)
657 /* default to off, as that's always safe */
660 p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
664 memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
665 ret = syscall(__NR_io_uring_register, ld->ring_fd,
666 IORING_REGISTER_PROBE, p, 256);
670 if (IORING_OP_WRITE > p->ops_len)
673 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) &&
674 (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED))
680 static int fio_ioring_queue_init(struct thread_data *td)
682 struct ioring_data *ld = td->io_ops_data;
683 struct ioring_options *o = td->eo;
684 int depth = td->o.iodepth;
685 struct io_uring_params p;
688 memset(&p, 0, sizeof(p));
691 p.flags |= IORING_SETUP_IOPOLL;
692 if (o->sqpoll_thread) {
693 p.flags |= IORING_SETUP_SQPOLL;
695 p.flags |= IORING_SETUP_SQ_AFF;
696 p.sq_thread_cpu = o->sqpoll_cpu;
701 * Clamp CQ ring size at our SQ ring size, we don't need more entries
704 p.flags |= IORING_SETUP_CQSIZE;
705 p.cq_entries = depth;
708 ret = syscall(__NR_io_uring_setup, depth, &p);
710 if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) {
711 p.flags &= ~IORING_SETUP_CQSIZE;
719 fio_ioring_probe(td);
722 ret = syscall(__NR_io_uring_register, ld->ring_fd,
723 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
728 return fio_ioring_mmap(ld, &p);
731 static int fio_ioring_register_files(struct thread_data *td)
733 struct ioring_data *ld = td->io_ops_data;
738 ld->fds = calloc(td->o.nr_files, sizeof(int));
740 for_each_file(td, f, i) {
741 ret = generic_open_file(td, f);
748 ret = syscall(__NR_io_uring_register, ld->ring_fd,
749 IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
757 * Pretend the file is closed again, and really close it if we hit
760 for_each_file(td, f, i) {
763 ret2 = generic_close_file(td, f);
771 static int fio_ioring_post_init(struct thread_data *td)
773 struct ioring_data *ld = td->io_ops_data;
774 struct ioring_options *o = td->eo;
778 for (i = 0; i < td->o.iodepth; i++) {
779 struct iovec *iov = &ld->iovecs[i];
781 io_u = ld->io_u_index[i];
782 iov->iov_base = io_u->buf;
783 iov->iov_len = td_max_bs(td);
786 err = fio_ioring_queue_init(td);
788 int init_err = errno;
790 if (init_err == ENOSYS)
791 log_err("fio: your kernel doesn't support io_uring\n");
792 td_verror(td, init_err, "io_queue_init");
796 for (i = 0; i < td->o.iodepth; i++) {
797 struct io_uring_sqe *sqe;
800 memset(sqe, 0, sizeof(*sqe));
803 if (o->registerfiles) {
804 err = fio_ioring_register_files(td);
806 td_verror(td, errno, "ioring_register_files");
814 static int fio_ioring_init(struct thread_data *td)
816 struct ioring_options *o = td->eo;
817 struct ioring_data *ld;
820 /* sqthread submission requires registered files */
821 if (o->sqpoll_thread)
822 o->registerfiles = 1;
824 if (o->registerfiles && td->o.nr_files != td->o.open_files) {
825 log_err("fio: io_uring registered files require nr_files to "
826 "be identical to open_files\n");
830 ld = calloc(1, sizeof(*ld));
832 /* ring depth must be a power-of-2 */
833 ld->iodepth = td->o.iodepth;
834 td->o.iodepth = roundup_pow2(td->o.iodepth);
837 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
838 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
840 td->io_ops_data = ld;
842 ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options);
844 td_verror(td, EINVAL, "fio_ioring_init");
851 static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
853 struct ioring_data *ld = td->io_ops_data;
855 ld->io_u_index[io_u->index] = io_u;
859 static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
861 struct ioring_data *ld = td->io_ops_data;
862 struct ioring_options *o = td->eo;
864 if (!ld || !o->registerfiles)
865 return generic_open_file(td, f);
867 f->fd = ld->fds[f->engine_pos];
871 static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
873 struct ioring_data *ld = td->io_ops_data;
874 struct ioring_options *o = td->eo;
876 if (!ld || !o->registerfiles)
877 return generic_close_file(td, f);
883 static struct ioengine_ops ioengine = {
885 .version = FIO_IOOPS_VERSION,
886 .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD,
887 .init = fio_ioring_init,
888 .post_init = fio_ioring_post_init,
889 .io_u_init = fio_ioring_io_u_init,
890 .prep = fio_ioring_prep,
891 .queue = fio_ioring_queue,
892 .commit = fio_ioring_commit,
893 .getevents = fio_ioring_getevents,
894 .event = fio_ioring_event,
895 .cleanup = fio_ioring_cleanup,
896 .open_file = fio_ioring_open_file,
897 .close_file = fio_ioring_close_file,
898 .get_file_size = generic_get_file_size,
900 .option_struct_size = sizeof(struct ioring_options),
903 static void fio_init fio_ioring_register(void)
905 register_ioengine(&ioengine);
908 static void fio_exit fio_ioring_unregister(void)
910 unregister_ioengine(&ioengine);