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 FIO_URING_CMD_NVME = 1,
39 unsigned *ring_entries;
48 unsigned *ring_entries;
49 struct io_uring_cqe *cqes;
60 struct io_u **io_u_index;
64 struct io_sq_ring sq_ring;
65 struct io_uring_sqe *sqes;
67 unsigned sq_ring_mask;
69 struct io_cq_ring cq_ring;
70 unsigned cq_ring_mask;
77 struct ioring_mmap mmap[3];
79 struct cmdprio cmdprio;
82 struct ioring_options {
83 struct thread_data *td;
85 struct cmdprio_options cmdprio_options;
86 unsigned int fixedbufs;
87 unsigned int registerfiles;
88 unsigned int sqpoll_thread;
89 unsigned int sqpoll_set;
90 unsigned int sqpoll_cpu;
91 unsigned int nonvectored;
92 unsigned int uncached;
94 unsigned int force_async;
95 enum uring_cmd_type cmd_type;
98 static const int ddir_to_op[2][2] = {
99 { IORING_OP_READV, IORING_OP_READ },
100 { IORING_OP_WRITEV, IORING_OP_WRITE }
103 static const int fixed_ddir_to_op[2] = {
104 IORING_OP_READ_FIXED,
105 IORING_OP_WRITE_FIXED
108 static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val)
110 struct ioring_options *o = data;
112 o->sqpoll_cpu = *val;
117 static struct fio_option options[] = {
120 .lname = "High Priority",
121 .type = FIO_OPT_STR_SET,
122 .off1 = offsetof(struct ioring_options, hipri),
123 .help = "Use polled IO completions",
124 .category = FIO_OPT_C_ENGINE,
125 .group = FIO_OPT_G_IOURING,
127 #ifdef FIO_HAVE_IOPRIO_CLASS
129 .name = "cmdprio_percentage",
130 .lname = "high priority percentage",
132 .off1 = offsetof(struct ioring_options,
133 cmdprio_options.percentage[DDIR_READ]),
134 .off2 = offsetof(struct ioring_options,
135 cmdprio_options.percentage[DDIR_WRITE]),
138 .help = "Send high priority I/O this percentage of the time",
139 .category = FIO_OPT_C_ENGINE,
140 .group = FIO_OPT_G_IOURING,
143 .name = "cmdprio_class",
144 .lname = "Asynchronous I/O priority class",
146 .off1 = offsetof(struct ioring_options,
147 cmdprio_options.class[DDIR_READ]),
148 .off2 = offsetof(struct ioring_options,
149 cmdprio_options.class[DDIR_WRITE]),
150 .help = "Set asynchronous IO priority class",
151 .minval = IOPRIO_MIN_PRIO_CLASS + 1,
152 .maxval = IOPRIO_MAX_PRIO_CLASS,
154 .category = FIO_OPT_C_ENGINE,
155 .group = FIO_OPT_G_IOURING,
159 .lname = "Asynchronous I/O priority level",
161 .off1 = offsetof(struct ioring_options,
162 cmdprio_options.level[DDIR_READ]),
163 .off2 = offsetof(struct ioring_options,
164 cmdprio_options.level[DDIR_WRITE]),
165 .help = "Set asynchronous IO priority level",
166 .minval = IOPRIO_MIN_PRIO,
167 .maxval = IOPRIO_MAX_PRIO,
169 .category = FIO_OPT_C_ENGINE,
170 .group = FIO_OPT_G_IOURING,
173 .name = "cmdprio_bssplit",
174 .lname = "Priority percentage block size split",
175 .type = FIO_OPT_STR_STORE,
176 .off1 = offsetof(struct ioring_options,
177 cmdprio_options.bssplit_str),
178 .help = "Set priority percentages for different block sizes",
179 .category = FIO_OPT_C_ENGINE,
180 .group = FIO_OPT_G_IOURING,
184 .name = "cmdprio_percentage",
185 .lname = "high priority percentage",
186 .type = FIO_OPT_UNSUPPORTED,
187 .help = "Your platform does not support I/O priority classes",
190 .name = "cmdprio_class",
191 .lname = "Asynchronous I/O priority class",
192 .type = FIO_OPT_UNSUPPORTED,
193 .help = "Your platform does not support I/O priority classes",
197 .lname = "Asynchronous I/O priority level",
198 .type = FIO_OPT_UNSUPPORTED,
199 .help = "Your platform does not support I/O priority classes",
202 .name = "cmdprio_bssplit",
203 .lname = "Priority percentage block size split",
204 .type = FIO_OPT_UNSUPPORTED,
205 .help = "Your platform does not support I/O priority classes",
210 .lname = "Fixed (pre-mapped) IO buffers",
211 .type = FIO_OPT_STR_SET,
212 .off1 = offsetof(struct ioring_options, fixedbufs),
213 .help = "Pre map IO buffers",
214 .category = FIO_OPT_C_ENGINE,
215 .group = FIO_OPT_G_IOURING,
218 .name = "registerfiles",
219 .lname = "Register file set",
220 .type = FIO_OPT_STR_SET,
221 .off1 = offsetof(struct ioring_options, registerfiles),
222 .help = "Pre-open/register files",
223 .category = FIO_OPT_C_ENGINE,
224 .group = FIO_OPT_G_IOURING,
227 .name = "sqthread_poll",
228 .lname = "Kernel SQ thread polling",
230 .off1 = offsetof(struct ioring_options, sqpoll_thread),
231 .help = "Offload submission/completion to kernel thread",
232 .category = FIO_OPT_C_ENGINE,
233 .group = FIO_OPT_G_IOURING,
236 .name = "sqthread_poll_cpu",
237 .lname = "SQ Thread Poll CPU",
239 .cb = fio_ioring_sqpoll_cb,
240 .help = "What CPU to run SQ thread polling on",
241 .category = FIO_OPT_C_ENGINE,
242 .group = FIO_OPT_G_IOURING,
245 .name = "nonvectored",
246 .lname = "Non-vectored",
248 .off1 = offsetof(struct ioring_options, nonvectored),
250 .help = "Use non-vectored read/write commands",
251 .category = FIO_OPT_C_ENGINE,
252 .group = FIO_OPT_G_IOURING,
258 .off1 = offsetof(struct ioring_options, uncached),
259 .help = "Use RWF_UNCACHED for buffered read/writes",
260 .category = FIO_OPT_C_ENGINE,
261 .group = FIO_OPT_G_IOURING,
265 .lname = "RWF_NOWAIT",
266 .type = FIO_OPT_BOOL,
267 .off1 = offsetof(struct ioring_options, nowait),
268 .help = "Use RWF_NOWAIT for reads/writes",
269 .category = FIO_OPT_C_ENGINE,
270 .group = FIO_OPT_G_IOURING,
273 .name = "force_async",
274 .lname = "Force async",
276 .off1 = offsetof(struct ioring_options, force_async),
277 .help = "Set IOSQE_ASYNC every N requests",
278 .category = FIO_OPT_C_ENGINE,
279 .group = FIO_OPT_G_IOURING,
283 .lname = "Uring cmd type",
285 .off1 = offsetof(struct ioring_options, cmd_type),
286 .help = "Specify uring-cmd type",
290 .oval = FIO_URING_CMD_NVME,
291 .help = "Issue nvme-uring-cmd",
294 .category = FIO_OPT_C_ENGINE,
295 .group = FIO_OPT_G_IOURING,
302 static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
303 unsigned int min_complete, unsigned int flags)
305 #ifdef FIO_ARCH_HAS_SYSCALL
306 return __do_syscall6(__NR_io_uring_enter, ld->ring_fd, to_submit,
307 min_complete, flags, NULL, 0);
309 return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit,
310 min_complete, flags, NULL, 0);
314 static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
316 struct ioring_data *ld = td->io_ops_data;
317 struct ioring_options *o = td->eo;
318 struct fio_file *f = io_u->file;
319 struct io_uring_sqe *sqe;
321 sqe = &ld->sqes[io_u->index];
323 if (o->registerfiles) {
324 sqe->fd = f->engine_pos;
325 sqe->flags = IOSQE_FIXED_FILE;
331 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
333 sqe->opcode = fixed_ddir_to_op[io_u->ddir];
334 sqe->addr = (unsigned long) io_u->xfer_buf;
335 sqe->len = io_u->xfer_buflen;
336 sqe->buf_index = io_u->index;
338 struct iovec *iov = &ld->iovecs[io_u->index];
341 * Update based on actual io_u, requeue could have
344 iov->iov_base = io_u->xfer_buf;
345 iov->iov_len = io_u->xfer_buflen;
347 sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored];
348 if (o->nonvectored) {
349 sqe->addr = (unsigned long) iov->iov_base;
350 sqe->len = iov->iov_len;
352 sqe->addr = (unsigned long) iov;
357 if (!td->o.odirect && o->uncached)
358 sqe->rw_flags |= RWF_UNCACHED;
360 sqe->rw_flags |= RWF_NOWAIT;
363 * Since io_uring can have a submission context (sqthread_poll)
364 * that is different from the process context, we cannot rely on
365 * the IO priority set by ioprio_set() (option prio/prioclass)
367 * td->ioprio will have the value of the "default prio", so set
368 * this unconditionally. This value might get overridden by
369 * fio_ioring_cmdprio_prep() if the option cmdprio_percentage or
370 * cmdprio_bssplit is used.
372 sqe->ioprio = td->ioprio;
373 sqe->off = io_u->offset;
374 } else if (ddir_sync(io_u->ddir)) {
376 if (io_u->ddir == DDIR_SYNC_FILE_RANGE) {
377 sqe->off = f->first_write;
378 sqe->len = f->last_write - f->first_write;
379 sqe->sync_range_flags = td->o.sync_file_range;
380 sqe->opcode = IORING_OP_SYNC_FILE_RANGE;
385 if (io_u->ddir == DDIR_DATASYNC)
386 sqe->fsync_flags |= IORING_FSYNC_DATASYNC;
387 sqe->opcode = IORING_OP_FSYNC;
391 if (o->force_async && ++ld->prepped == o->force_async) {
393 sqe->flags |= IOSQE_ASYNC;
396 sqe->user_data = (unsigned long) io_u;
400 static int fio_ioring_cmd_prep(struct thread_data *td, struct io_u *io_u)
402 struct ioring_data *ld = td->io_ops_data;
403 struct ioring_options *o = td->eo;
404 struct fio_file *f = io_u->file;
405 struct nvme_uring_cmd *cmd;
406 struct io_uring_sqe *sqe;
408 /* only supports nvme_uring_cmd */
409 if (o->cmd_type != FIO_URING_CMD_NVME)
412 sqe = &ld->sqes[(io_u->index) << 1];
414 if (o->registerfiles) {
415 sqe->fd = f->engine_pos;
416 sqe->flags = IOSQE_FIXED_FILE;
421 if (!td->o.odirect && o->uncached)
422 sqe->rw_flags |= RWF_UNCACHED;
424 sqe->rw_flags |= RWF_NOWAIT;
426 sqe->opcode = IORING_OP_URING_CMD;
427 sqe->user_data = (unsigned long) io_u;
429 sqe->cmd_op = NVME_URING_CMD_IO;
431 sqe->cmd_op = NVME_URING_CMD_IO_VEC;
432 if (o->force_async && ++ld->prepped == o->force_async) {
434 sqe->flags |= IOSQE_ASYNC;
437 cmd = (struct nvme_uring_cmd *)sqe->cmd;
438 return fio_nvme_uring_cmd_prep(cmd, io_u,
439 o->nonvectored ? NULL : &ld->iovecs[io_u->index]);
442 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
444 struct ioring_data *ld = td->io_ops_data;
445 struct io_uring_cqe *cqe;
449 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
451 cqe = &ld->cq_ring.cqes[index];
452 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
454 if (cqe->res != io_u->xfer_buflen) {
455 if (cqe->res > io_u->xfer_buflen)
456 io_u->error = -cqe->res;
458 io_u->resid = io_u->xfer_buflen - cqe->res;
465 static struct io_u *fio_ioring_cmd_event(struct thread_data *td, int event)
467 struct ioring_data *ld = td->io_ops_data;
468 struct ioring_options *o = td->eo;
469 struct io_uring_cqe *cqe;
473 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
474 if (o->cmd_type == FIO_URING_CMD_NVME)
477 cqe = &ld->cq_ring.cqes[index];
478 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
481 io_u->error = -cqe->res;
488 static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
491 struct ioring_data *ld = td->io_ops_data;
492 struct io_cq_ring *ring = &ld->cq_ring;
493 unsigned head, reaped = 0;
497 if (head == atomic_load_acquire(ring->tail))
501 } while (reaped + events < max);
504 atomic_store_release(ring->head, head);
509 static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
510 unsigned int max, const struct timespec *t)
512 struct ioring_data *ld = td->io_ops_data;
513 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
514 struct ioring_options *o = td->eo;
515 struct io_cq_ring *ring = &ld->cq_ring;
519 ld->cq_ring_off = *ring->head;
521 r = fio_ioring_cqring_reap(td, events, max);
529 if (!o->sqpoll_thread) {
530 r = io_uring_enter(ld, 0, actual_min,
531 IORING_ENTER_GETEVENTS);
533 if (errno == EAGAIN || errno == EINTR)
535 td_verror(td, errno, "io_uring_enter");
539 } while (events < min);
541 return r < 0 ? r : events;
544 static inline void fio_ioring_cmdprio_prep(struct thread_data *td,
547 struct ioring_data *ld = td->io_ops_data;
548 struct cmdprio *cmdprio = &ld->cmdprio;
550 if (fio_cmdprio_set_ioprio(td, cmdprio, io_u))
551 ld->sqes[io_u->index].ioprio = io_u->ioprio;
554 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
557 struct ioring_data *ld = td->io_ops_data;
558 struct io_sq_ring *ring = &ld->sq_ring;
559 unsigned tail, next_tail;
561 fio_ro_check(td, io_u);
563 if (ld->queued == ld->iodepth)
566 if (io_u->ddir == DDIR_TRIM) {
570 do_io_u_trim(td, io_u);
571 io_u_mark_submit(td, 1);
572 io_u_mark_complete(td, 1);
573 return FIO_Q_COMPLETED;
577 next_tail = tail + 1;
578 if (next_tail == atomic_load_acquire(ring->head))
581 if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
582 fio_ioring_cmdprio_prep(td, io_u);
584 ring->array[tail & ld->sq_ring_mask] = io_u->index;
585 atomic_store_release(ring->tail, next_tail);
591 static void fio_ioring_queued(struct thread_data *td, int start, int nr)
593 struct ioring_data *ld = td->io_ops_data;
596 if (!fio_fill_issue_time(td))
599 fio_gettime(&now, NULL);
602 struct io_sq_ring *ring = &ld->sq_ring;
603 int index = ring->array[start & ld->sq_ring_mask];
604 struct io_u *io_u = ld->io_u_index[index];
606 memcpy(&io_u->issue_time, &now, sizeof(now));
607 io_u_queued(td, io_u);
613 * only used for iolog
615 if (td->o.read_iolog_file)
616 memcpy(&td->last_issue, &now, sizeof(now));
619 static int fio_ioring_commit(struct thread_data *td)
621 struct ioring_data *ld = td->io_ops_data;
622 struct ioring_options *o = td->eo;
629 * Kernel side does submission. just need to check if the ring is
630 * flagged as needing a kick, if so, call io_uring_enter(). This
631 * only happens if we've been idle too long.
633 if (o->sqpoll_thread) {
634 struct io_sq_ring *ring = &ld->sq_ring;
637 flags = atomic_load_acquire(ring->flags);
638 if (flags & IORING_SQ_NEED_WAKEUP)
639 io_uring_enter(ld, ld->queued, 0,
640 IORING_ENTER_SQ_WAKEUP);
646 unsigned start = *ld->sq_ring.head;
647 long nr = ld->queued;
649 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
651 fio_ioring_queued(td, start, ret);
652 io_u_mark_submit(td, ret);
657 io_u_mark_submit(td, ret);
660 if (errno == EAGAIN || errno == EINTR) {
661 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
664 /* Shouldn't happen */
668 td_verror(td, errno, "io_uring_enter submit");
671 } while (ld->queued);
676 static void fio_ioring_unmap(struct ioring_data *ld)
680 for (i = 0; i < FIO_ARRAY_SIZE(ld->mmap); i++)
681 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
685 static void fio_ioring_cleanup(struct thread_data *td)
687 struct ioring_data *ld = td->io_ops_data;
690 if (!(td->flags & TD_F_CHILD))
691 fio_ioring_unmap(ld);
693 fio_cmdprio_cleanup(&ld->cmdprio);
694 free(ld->io_u_index);
701 static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
703 struct io_sq_ring *sring = &ld->sq_ring;
704 struct io_cq_ring *cring = &ld->cq_ring;
707 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
708 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
709 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
711 ld->mmap[0].ptr = ptr;
712 sring->head = ptr + p->sq_off.head;
713 sring->tail = ptr + p->sq_off.tail;
714 sring->ring_mask = ptr + p->sq_off.ring_mask;
715 sring->ring_entries = ptr + p->sq_off.ring_entries;
716 sring->flags = ptr + p->sq_off.flags;
717 sring->array = ptr + p->sq_off.array;
718 ld->sq_ring_mask = *sring->ring_mask;
720 if (p->flags & IORING_SETUP_SQE128)
721 ld->mmap[1].len = 2 * p->sq_entries * sizeof(struct io_uring_sqe);
723 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
724 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
725 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
727 ld->mmap[1].ptr = ld->sqes;
729 if (p->flags & IORING_SETUP_CQE32) {
730 ld->mmap[2].len = p->cq_off.cqes +
731 2 * p->cq_entries * sizeof(struct io_uring_cqe);
733 ld->mmap[2].len = p->cq_off.cqes +
734 p->cq_entries * sizeof(struct io_uring_cqe);
736 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
737 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
739 ld->mmap[2].ptr = ptr;
740 cring->head = ptr + p->cq_off.head;
741 cring->tail = ptr + p->cq_off.tail;
742 cring->ring_mask = ptr + p->cq_off.ring_mask;
743 cring->ring_entries = ptr + p->cq_off.ring_entries;
744 cring->cqes = ptr + p->cq_off.cqes;
745 ld->cq_ring_mask = *cring->ring_mask;
749 static void fio_ioring_probe(struct thread_data *td)
751 struct ioring_data *ld = td->io_ops_data;
752 struct ioring_options *o = td->eo;
753 struct io_uring_probe *p;
756 /* already set by user, don't touch */
757 if (o->nonvectored != -1)
760 /* default to off, as that's always safe */
763 p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
767 memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
768 ret = syscall(__NR_io_uring_register, ld->ring_fd,
769 IORING_REGISTER_PROBE, p, 256);
773 if (IORING_OP_WRITE > p->ops_len)
776 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) &&
777 (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED))
783 static int fio_ioring_queue_init(struct thread_data *td)
785 struct ioring_data *ld = td->io_ops_data;
786 struct ioring_options *o = td->eo;
787 int depth = td->o.iodepth;
788 struct io_uring_params p;
791 memset(&p, 0, sizeof(p));
794 p.flags |= IORING_SETUP_IOPOLL;
795 if (o->sqpoll_thread) {
796 p.flags |= IORING_SETUP_SQPOLL;
798 p.flags |= IORING_SETUP_SQ_AFF;
799 p.sq_thread_cpu = o->sqpoll_cpu;
804 * Clamp CQ ring size at our SQ ring size, we don't need more entries
807 p.flags |= IORING_SETUP_CQSIZE;
808 p.cq_entries = depth;
811 ret = syscall(__NR_io_uring_setup, depth, &p);
813 if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) {
814 p.flags &= ~IORING_SETUP_CQSIZE;
822 fio_ioring_probe(td);
825 ret = syscall(__NR_io_uring_register, ld->ring_fd,
826 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
831 return fio_ioring_mmap(ld, &p);
834 static int fio_ioring_cmd_queue_init(struct thread_data *td)
836 struct ioring_data *ld = td->io_ops_data;
837 struct ioring_options *o = td->eo;
838 int depth = td->o.iodepth;
839 struct io_uring_params p;
842 memset(&p, 0, sizeof(p));
845 p.flags |= IORING_SETUP_IOPOLL;
846 if (o->sqpoll_thread) {
847 p.flags |= IORING_SETUP_SQPOLL;
849 p.flags |= IORING_SETUP_SQ_AFF;
850 p.sq_thread_cpu = o->sqpoll_cpu;
853 if (o->cmd_type == FIO_URING_CMD_NVME) {
854 p.flags |= IORING_SETUP_SQE128;
855 p.flags |= IORING_SETUP_CQE32;
859 * Clamp CQ ring size at our SQ ring size, we don't need more entries
862 p.flags |= IORING_SETUP_CQSIZE;
863 p.cq_entries = depth;
866 ret = syscall(__NR_io_uring_setup, depth, &p);
868 if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) {
869 p.flags &= ~IORING_SETUP_CQSIZE;
877 fio_ioring_probe(td);
880 ret = syscall(__NR_io_uring_register, ld->ring_fd,
881 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
886 return fio_ioring_mmap(ld, &p);
889 static int fio_ioring_register_files(struct thread_data *td)
891 struct ioring_data *ld = td->io_ops_data;
896 ld->fds = calloc(td->o.nr_files, sizeof(int));
898 for_each_file(td, f, i) {
899 ret = generic_open_file(td, f);
906 ret = syscall(__NR_io_uring_register, ld->ring_fd,
907 IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
915 * Pretend the file is closed again, and really close it if we hit
918 for_each_file(td, f, i) {
921 ret2 = generic_close_file(td, f);
929 static int fio_ioring_post_init(struct thread_data *td)
931 struct ioring_data *ld = td->io_ops_data;
932 struct ioring_options *o = td->eo;
936 for (i = 0; i < td->o.iodepth; i++) {
937 struct iovec *iov = &ld->iovecs[i];
939 io_u = ld->io_u_index[i];
940 iov->iov_base = io_u->buf;
941 iov->iov_len = td_max_bs(td);
944 err = fio_ioring_queue_init(td);
946 int init_err = errno;
948 if (init_err == ENOSYS)
949 log_err("fio: your kernel doesn't support io_uring\n");
950 td_verror(td, init_err, "io_queue_init");
954 for (i = 0; i < td->o.iodepth; i++) {
955 struct io_uring_sqe *sqe;
958 memset(sqe, 0, sizeof(*sqe));
961 if (o->registerfiles) {
962 err = fio_ioring_register_files(td);
964 td_verror(td, errno, "ioring_register_files");
972 static int fio_ioring_cmd_post_init(struct thread_data *td)
974 struct ioring_data *ld = td->io_ops_data;
975 struct ioring_options *o = td->eo;
979 for (i = 0; i < td->o.iodepth; i++) {
980 struct iovec *iov = &ld->iovecs[i];
982 io_u = ld->io_u_index[i];
983 iov->iov_base = io_u->buf;
984 iov->iov_len = td_max_bs(td);
987 err = fio_ioring_cmd_queue_init(td);
989 int init_err = errno;
991 td_verror(td, init_err, "io_queue_init");
995 for (i = 0; i < td->o.iodepth; i++) {
996 struct io_uring_sqe *sqe;
998 if (o->cmd_type == FIO_URING_CMD_NVME) {
999 sqe = &ld->sqes[i << 1];
1000 memset(sqe, 0, 2 * sizeof(*sqe));
1003 memset(sqe, 0, sizeof(*sqe));
1007 if (o->registerfiles) {
1008 err = fio_ioring_register_files(td);
1010 td_verror(td, errno, "ioring_register_files");
1018 static int fio_ioring_init(struct thread_data *td)
1020 struct ioring_options *o = td->eo;
1021 struct ioring_data *ld;
1024 /* sqthread submission requires registered files */
1025 if (o->sqpoll_thread)
1026 o->registerfiles = 1;
1028 if (o->registerfiles && td->o.nr_files != td->o.open_files) {
1029 log_err("fio: io_uring registered files require nr_files to "
1030 "be identical to open_files\n");
1034 ld = calloc(1, sizeof(*ld));
1036 /* ring depth must be a power-of-2 */
1037 ld->iodepth = td->o.iodepth;
1038 td->o.iodepth = roundup_pow2(td->o.iodepth);
1041 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
1042 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
1044 td->io_ops_data = ld;
1046 ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options);
1048 td_verror(td, EINVAL, "fio_ioring_init");
1055 static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
1057 struct ioring_data *ld = td->io_ops_data;
1059 ld->io_u_index[io_u->index] = io_u;
1063 static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
1065 struct ioring_data *ld = td->io_ops_data;
1066 struct ioring_options *o = td->eo;
1068 if (!ld || !o->registerfiles)
1069 return generic_open_file(td, f);
1071 f->fd = ld->fds[f->engine_pos];
1075 static int fio_ioring_cmd_open_file(struct thread_data *td, struct fio_file *f)
1077 struct ioring_data *ld = td->io_ops_data;
1078 struct ioring_options *o = td->eo;
1080 if (o->cmd_type == FIO_URING_CMD_NVME) {
1081 struct nvme_data *data = NULL;
1082 unsigned int nsid, lba_size = 0;
1083 unsigned long long nlba = 0;
1086 /* Store the namespace-id and lba size. */
1087 data = FILE_ENG_DATA(f);
1089 ret = fio_nvme_get_info(f, &nsid, &lba_size, &nlba);
1093 data = calloc(1, sizeof(struct nvme_data));
1095 data->lba_shift = ilog2(lba_size);
1097 FILE_SET_ENG_DATA(f, data);
1100 if (!ld || !o->registerfiles)
1101 return generic_open_file(td, f);
1103 f->fd = ld->fds[f->engine_pos];
1107 static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
1109 struct ioring_data *ld = td->io_ops_data;
1110 struct ioring_options *o = td->eo;
1112 if (!ld || !o->registerfiles)
1113 return generic_close_file(td, f);
1119 static int fio_ioring_cmd_close_file(struct thread_data *td,
1122 struct ioring_data *ld = td->io_ops_data;
1123 struct ioring_options *o = td->eo;
1125 if (o->cmd_type == FIO_URING_CMD_NVME) {
1126 struct nvme_data *data = FILE_ENG_DATA(f);
1128 FILE_SET_ENG_DATA(f, NULL);
1131 if (!ld || !o->registerfiles)
1132 return generic_close_file(td, f);
1138 static int fio_ioring_cmd_get_file_size(struct thread_data *td,
1141 struct ioring_options *o = td->eo;
1143 if (fio_file_size_known(f))
1146 if (o->cmd_type == FIO_URING_CMD_NVME) {
1147 struct nvme_data *data = NULL;
1148 unsigned int nsid, lba_size = 0;
1149 unsigned long long nlba = 0;
1152 ret = fio_nvme_get_info(f, &nsid, &lba_size, &nlba);
1156 data = calloc(1, sizeof(struct nvme_data));
1158 data->lba_shift = ilog2(lba_size);
1160 f->real_file_size = lba_size * nlba;
1161 fio_file_set_size_known(f);
1163 FILE_SET_ENG_DATA(f, data);
1166 return generic_get_file_size(td, f);
1169 static int fio_ioring_cmd_get_zoned_model(struct thread_data *td,
1171 enum zbd_zoned_model *model)
1173 return fio_nvme_get_zoned_model(td, f, model);
1176 static int fio_ioring_cmd_report_zones(struct thread_data *td,
1177 struct fio_file *f, uint64_t offset,
1178 struct zbd_zone *zbdz,
1179 unsigned int nr_zones)
1181 return fio_nvme_report_zones(td, f, offset, zbdz, nr_zones);
1184 static int fio_ioring_cmd_reset_wp(struct thread_data *td, struct fio_file *f,
1185 uint64_t offset, uint64_t length)
1187 return fio_nvme_reset_wp(td, f, offset, length);
1190 static int fio_ioring_cmd_get_max_open_zones(struct thread_data *td,
1192 unsigned int *max_open_zones)
1194 return fio_nvme_get_max_open_zones(td, f, max_open_zones);
1197 static struct ioengine_ops ioengine_uring = {
1199 .version = FIO_IOOPS_VERSION,
1200 .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD |
1201 FIO_ASYNCIO_SETS_ISSUE_TIME,
1202 .init = fio_ioring_init,
1203 .post_init = fio_ioring_post_init,
1204 .io_u_init = fio_ioring_io_u_init,
1205 .prep = fio_ioring_prep,
1206 .queue = fio_ioring_queue,
1207 .commit = fio_ioring_commit,
1208 .getevents = fio_ioring_getevents,
1209 .event = fio_ioring_event,
1210 .cleanup = fio_ioring_cleanup,
1211 .open_file = fio_ioring_open_file,
1212 .close_file = fio_ioring_close_file,
1213 .get_file_size = generic_get_file_size,
1215 .option_struct_size = sizeof(struct ioring_options),
1218 static struct ioengine_ops ioengine_uring_cmd = {
1219 .name = "io_uring_cmd",
1220 .version = FIO_IOOPS_VERSION,
1221 .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD |
1222 FIO_MEMALIGN | FIO_RAWIO |
1223 FIO_ASYNCIO_SETS_ISSUE_TIME,
1224 .init = fio_ioring_init,
1225 .post_init = fio_ioring_cmd_post_init,
1226 .io_u_init = fio_ioring_io_u_init,
1227 .prep = fio_ioring_cmd_prep,
1228 .queue = fio_ioring_queue,
1229 .commit = fio_ioring_commit,
1230 .getevents = fio_ioring_getevents,
1231 .event = fio_ioring_cmd_event,
1232 .cleanup = fio_ioring_cleanup,
1233 .open_file = fio_ioring_cmd_open_file,
1234 .close_file = fio_ioring_cmd_close_file,
1235 .get_file_size = fio_ioring_cmd_get_file_size,
1236 .get_zoned_model = fio_ioring_cmd_get_zoned_model,
1237 .report_zones = fio_ioring_cmd_report_zones,
1238 .reset_wp = fio_ioring_cmd_reset_wp,
1239 .get_max_open_zones = fio_ioring_cmd_get_max_open_zones,
1241 .option_struct_size = sizeof(struct ioring_options),
1244 static void fio_init fio_ioring_register(void)
1246 register_ioengine(&ioengine_uring);
1247 register_ioengine(&ioengine_uring_cmd);
1250 static void fio_exit fio_ioring_unregister(void)
1252 unregister_ioengine(&ioengine_uring);
1253 unregister_ioengine(&ioengine_uring_cmd);