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 io_uring_sqe *sqe;
408 /* nvme_uring_cmd case */
409 if (o->cmd_type == FIO_URING_CMD_NVME) {
410 struct nvme_uring_cmd *cmd;
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 ret = fio_nvme_uring_cmd_prep(cmd, io_u,
439 o->nonvectored ? NULL : &ld->iovecs[io_u->index]);
446 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
448 struct ioring_data *ld = td->io_ops_data;
449 struct io_uring_cqe *cqe;
453 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
455 cqe = &ld->cq_ring.cqes[index];
456 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
458 if (cqe->res != io_u->xfer_buflen) {
459 if (cqe->res > io_u->xfer_buflen)
460 io_u->error = -cqe->res;
462 io_u->resid = io_u->xfer_buflen - cqe->res;
469 static struct io_u *fio_ioring_cmd_event(struct thread_data *td, int event)
471 struct ioring_data *ld = td->io_ops_data;
472 struct ioring_options *o = td->eo;
473 struct io_uring_cqe *cqe;
477 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
478 if (o->cmd_type == FIO_URING_CMD_NVME)
481 cqe = &ld->cq_ring.cqes[index];
482 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
485 io_u->error = -cqe->res;
492 static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
495 struct ioring_data *ld = td->io_ops_data;
496 struct io_cq_ring *ring = &ld->cq_ring;
497 unsigned head, reaped = 0;
501 if (head == atomic_load_acquire(ring->tail))
505 } while (reaped + events < max);
508 atomic_store_release(ring->head, head);
513 static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
514 unsigned int max, const struct timespec *t)
516 struct ioring_data *ld = td->io_ops_data;
517 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
518 struct ioring_options *o = td->eo;
519 struct io_cq_ring *ring = &ld->cq_ring;
523 ld->cq_ring_off = *ring->head;
525 r = fio_ioring_cqring_reap(td, events, max);
533 if (!o->sqpoll_thread) {
534 r = io_uring_enter(ld, 0, actual_min,
535 IORING_ENTER_GETEVENTS);
537 if (errno == EAGAIN || errno == EINTR)
539 td_verror(td, errno, "io_uring_enter");
543 } while (events < min);
545 return r < 0 ? r : events;
548 static inline void fio_ioring_cmdprio_prep(struct thread_data *td,
551 struct ioring_data *ld = td->io_ops_data;
552 struct cmdprio *cmdprio = &ld->cmdprio;
554 if (fio_cmdprio_set_ioprio(td, cmdprio, io_u))
555 ld->sqes[io_u->index].ioprio = io_u->ioprio;
558 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
561 struct ioring_data *ld = td->io_ops_data;
562 struct io_sq_ring *ring = &ld->sq_ring;
563 unsigned tail, next_tail;
565 fio_ro_check(td, io_u);
567 if (ld->queued == ld->iodepth)
570 if (io_u->ddir == DDIR_TRIM) {
574 do_io_u_trim(td, io_u);
575 io_u_mark_submit(td, 1);
576 io_u_mark_complete(td, 1);
577 return FIO_Q_COMPLETED;
581 next_tail = tail + 1;
582 if (next_tail == atomic_load_acquire(ring->head))
585 if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
586 fio_ioring_cmdprio_prep(td, io_u);
588 ring->array[tail & ld->sq_ring_mask] = io_u->index;
589 atomic_store_release(ring->tail, next_tail);
595 static void fio_ioring_queued(struct thread_data *td, int start, int nr)
597 struct ioring_data *ld = td->io_ops_data;
600 if (!fio_fill_issue_time(td))
603 fio_gettime(&now, NULL);
606 struct io_sq_ring *ring = &ld->sq_ring;
607 int index = ring->array[start & ld->sq_ring_mask];
608 struct io_u *io_u = ld->io_u_index[index];
610 memcpy(&io_u->issue_time, &now, sizeof(now));
611 io_u_queued(td, io_u);
617 static int fio_ioring_commit(struct thread_data *td)
619 struct ioring_data *ld = td->io_ops_data;
620 struct ioring_options *o = td->eo;
627 * Kernel side does submission. just need to check if the ring is
628 * flagged as needing a kick, if so, call io_uring_enter(). This
629 * only happens if we've been idle too long.
631 if (o->sqpoll_thread) {
632 struct io_sq_ring *ring = &ld->sq_ring;
635 flags = atomic_load_acquire(ring->flags);
636 if (flags & IORING_SQ_NEED_WAKEUP)
637 io_uring_enter(ld, ld->queued, 0,
638 IORING_ENTER_SQ_WAKEUP);
644 unsigned start = *ld->sq_ring.head;
645 long nr = ld->queued;
647 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
649 fio_ioring_queued(td, start, ret);
650 io_u_mark_submit(td, ret);
655 io_u_mark_submit(td, ret);
658 if (errno == EAGAIN || errno == EINTR) {
659 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
662 /* Shouldn't happen */
666 td_verror(td, errno, "io_uring_enter submit");
669 } while (ld->queued);
674 static void fio_ioring_unmap(struct ioring_data *ld)
678 for (i = 0; i < FIO_ARRAY_SIZE(ld->mmap); i++)
679 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
683 static void fio_ioring_cleanup(struct thread_data *td)
685 struct ioring_data *ld = td->io_ops_data;
688 if (!(td->flags & TD_F_CHILD))
689 fio_ioring_unmap(ld);
691 fio_cmdprio_cleanup(&ld->cmdprio);
692 free(ld->io_u_index);
699 static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
701 struct io_sq_ring *sring = &ld->sq_ring;
702 struct io_cq_ring *cring = &ld->cq_ring;
705 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
706 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
707 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
709 ld->mmap[0].ptr = ptr;
710 sring->head = ptr + p->sq_off.head;
711 sring->tail = ptr + p->sq_off.tail;
712 sring->ring_mask = ptr + p->sq_off.ring_mask;
713 sring->ring_entries = ptr + p->sq_off.ring_entries;
714 sring->flags = ptr + p->sq_off.flags;
715 sring->array = ptr + p->sq_off.array;
716 ld->sq_ring_mask = *sring->ring_mask;
718 if (p->flags & IORING_SETUP_SQE128)
719 ld->mmap[1].len = 2 * p->sq_entries * sizeof(struct io_uring_sqe);
721 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
722 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
723 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
725 ld->mmap[1].ptr = ld->sqes;
727 if (p->flags & IORING_SETUP_CQE32) {
728 ld->mmap[2].len = p->cq_off.cqes +
729 2 * p->cq_entries * sizeof(struct io_uring_cqe);
731 ld->mmap[2].len = p->cq_off.cqes +
732 p->cq_entries * sizeof(struct io_uring_cqe);
734 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
735 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
737 ld->mmap[2].ptr = ptr;
738 cring->head = ptr + p->cq_off.head;
739 cring->tail = ptr + p->cq_off.tail;
740 cring->ring_mask = ptr + p->cq_off.ring_mask;
741 cring->ring_entries = ptr + p->cq_off.ring_entries;
742 cring->cqes = ptr + p->cq_off.cqes;
743 ld->cq_ring_mask = *cring->ring_mask;
747 static void fio_ioring_probe(struct thread_data *td)
749 struct ioring_data *ld = td->io_ops_data;
750 struct ioring_options *o = td->eo;
751 struct io_uring_probe *p;
754 /* already set by user, don't touch */
755 if (o->nonvectored != -1)
758 /* default to off, as that's always safe */
761 p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
765 memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
766 ret = syscall(__NR_io_uring_register, ld->ring_fd,
767 IORING_REGISTER_PROBE, p, 256);
771 if (IORING_OP_WRITE > p->ops_len)
774 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) &&
775 (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED))
781 static int fio_ioring_queue_init(struct thread_data *td)
783 struct ioring_data *ld = td->io_ops_data;
784 struct ioring_options *o = td->eo;
785 int depth = td->o.iodepth;
786 struct io_uring_params p;
789 memset(&p, 0, sizeof(p));
792 p.flags |= IORING_SETUP_IOPOLL;
793 if (o->sqpoll_thread) {
794 p.flags |= IORING_SETUP_SQPOLL;
796 p.flags |= IORING_SETUP_SQ_AFF;
797 p.sq_thread_cpu = o->sqpoll_cpu;
802 * Clamp CQ ring size at our SQ ring size, we don't need more entries
805 p.flags |= IORING_SETUP_CQSIZE;
806 p.cq_entries = depth;
809 ret = syscall(__NR_io_uring_setup, depth, &p);
811 if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) {
812 p.flags &= ~IORING_SETUP_CQSIZE;
820 fio_ioring_probe(td);
823 ret = syscall(__NR_io_uring_register, ld->ring_fd,
824 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
829 return fio_ioring_mmap(ld, &p);
832 static int fio_ioring_cmd_queue_init(struct thread_data *td)
834 struct ioring_data *ld = td->io_ops_data;
835 struct ioring_options *o = td->eo;
836 int depth = td->o.iodepth;
837 struct io_uring_params p;
840 memset(&p, 0, sizeof(p));
843 p.flags |= IORING_SETUP_IOPOLL;
844 if (o->sqpoll_thread) {
845 p.flags |= IORING_SETUP_SQPOLL;
847 p.flags |= IORING_SETUP_SQ_AFF;
848 p.sq_thread_cpu = o->sqpoll_cpu;
851 if (o->cmd_type == FIO_URING_CMD_NVME) {
852 p.flags |= IORING_SETUP_SQE128;
853 p.flags |= IORING_SETUP_CQE32;
857 * Clamp CQ ring size at our SQ ring size, we don't need more entries
860 p.flags |= IORING_SETUP_CQSIZE;
861 p.cq_entries = depth;
864 ret = syscall(__NR_io_uring_setup, depth, &p);
866 if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) {
867 p.flags &= ~IORING_SETUP_CQSIZE;
875 fio_ioring_probe(td);
878 ret = syscall(__NR_io_uring_register, ld->ring_fd,
879 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
884 return fio_ioring_mmap(ld, &p);
887 static int fio_ioring_register_files(struct thread_data *td)
889 struct ioring_data *ld = td->io_ops_data;
894 ld->fds = calloc(td->o.nr_files, sizeof(int));
896 for_each_file(td, f, i) {
897 ret = generic_open_file(td, f);
904 ret = syscall(__NR_io_uring_register, ld->ring_fd,
905 IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
913 * Pretend the file is closed again, and really close it if we hit
916 for_each_file(td, f, i) {
919 ret2 = generic_close_file(td, f);
927 static int fio_ioring_post_init(struct thread_data *td)
929 struct ioring_data *ld = td->io_ops_data;
930 struct ioring_options *o = td->eo;
934 for (i = 0; i < td->o.iodepth; i++) {
935 struct iovec *iov = &ld->iovecs[i];
937 io_u = ld->io_u_index[i];
938 iov->iov_base = io_u->buf;
939 iov->iov_len = td_max_bs(td);
942 err = fio_ioring_queue_init(td);
944 int init_err = errno;
946 if (init_err == ENOSYS)
947 log_err("fio: your kernel doesn't support io_uring\n");
948 td_verror(td, init_err, "io_queue_init");
952 for (i = 0; i < td->o.iodepth; i++) {
953 struct io_uring_sqe *sqe;
956 memset(sqe, 0, sizeof(*sqe));
959 if (o->registerfiles) {
960 err = fio_ioring_register_files(td);
962 td_verror(td, errno, "ioring_register_files");
970 static int fio_ioring_cmd_post_init(struct thread_data *td)
972 struct ioring_data *ld = td->io_ops_data;
973 struct ioring_options *o = td->eo;
977 for (i = 0; i < td->o.iodepth; i++) {
978 struct iovec *iov = &ld->iovecs[i];
980 io_u = ld->io_u_index[i];
981 iov->iov_base = io_u->buf;
982 iov->iov_len = td_max_bs(td);
985 err = fio_ioring_cmd_queue_init(td);
987 int init_err = errno;
989 td_verror(td, init_err, "io_queue_init");
993 for (i = 0; i < td->o.iodepth; i++) {
994 struct io_uring_sqe *sqe;
996 if (o->cmd_type == FIO_URING_CMD_NVME) {
997 sqe = &ld->sqes[i << 1];
998 memset(sqe, 0, 2 * sizeof(*sqe));
1001 memset(sqe, 0, sizeof(*sqe));
1005 if (o->registerfiles) {
1006 err = fio_ioring_register_files(td);
1008 td_verror(td, errno, "ioring_register_files");
1016 static int fio_ioring_init(struct thread_data *td)
1018 struct ioring_options *o = td->eo;
1019 struct ioring_data *ld;
1022 /* sqthread submission requires registered files */
1023 if (o->sqpoll_thread)
1024 o->registerfiles = 1;
1026 if (o->registerfiles && td->o.nr_files != td->o.open_files) {
1027 log_err("fio: io_uring registered files require nr_files to "
1028 "be identical to open_files\n");
1032 ld = calloc(1, sizeof(*ld));
1034 /* ring depth must be a power-of-2 */
1035 ld->iodepth = td->o.iodepth;
1036 td->o.iodepth = roundup_pow2(td->o.iodepth);
1039 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
1040 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
1042 td->io_ops_data = ld;
1044 ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options);
1046 td_verror(td, EINVAL, "fio_ioring_init");
1053 static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
1055 struct ioring_data *ld = td->io_ops_data;
1057 ld->io_u_index[io_u->index] = io_u;
1061 static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
1063 struct ioring_data *ld = td->io_ops_data;
1064 struct ioring_options *o = td->eo;
1066 if (!ld || !o->registerfiles)
1067 return generic_open_file(td, f);
1069 f->fd = ld->fds[f->engine_pos];
1073 static int fio_ioring_cmd_open_file(struct thread_data *td, struct fio_file *f)
1075 struct ioring_data *ld = td->io_ops_data;
1076 struct ioring_options *o = td->eo;
1078 if (o->cmd_type == FIO_URING_CMD_NVME) {
1079 struct nvme_data *data = NULL;
1080 unsigned int nsid, lba_size = 0;
1081 unsigned long long nlba = 0;
1084 /* Store the namespace-id and lba size. */
1085 data = FILE_ENG_DATA(f);
1087 ret = fio_nvme_get_info(f, &nsid, &lba_size, &nlba);
1091 data = calloc(1, sizeof(struct nvme_data));
1093 data->lba_shift = ilog2(lba_size);
1095 FILE_SET_ENG_DATA(f, data);
1098 if (!ld || !o->registerfiles)
1099 return generic_open_file(td, f);
1101 f->fd = ld->fds[f->engine_pos];
1105 static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
1107 struct ioring_data *ld = td->io_ops_data;
1108 struct ioring_options *o = td->eo;
1110 if (!ld || !o->registerfiles)
1111 return generic_close_file(td, f);
1117 static int fio_ioring_cmd_close_file(struct thread_data *td,
1120 struct ioring_data *ld = td->io_ops_data;
1121 struct ioring_options *o = td->eo;
1123 if (o->cmd_type == FIO_URING_CMD_NVME) {
1124 struct nvme_data *data = FILE_ENG_DATA(f);
1126 FILE_SET_ENG_DATA(f, NULL);
1129 if (!ld || !o->registerfiles)
1130 return generic_close_file(td, f);
1136 static int fio_ioring_cmd_get_file_size(struct thread_data *td,
1139 struct ioring_options *o = td->eo;
1141 if (fio_file_size_known(f))
1144 if (o->cmd_type == FIO_URING_CMD_NVME) {
1145 struct nvme_data *data = NULL;
1146 unsigned int nsid, lba_size = 0;
1147 unsigned long long nlba = 0;
1150 ret = fio_nvme_get_info(f, &nsid, &lba_size, &nlba);
1154 data = calloc(1, sizeof(struct nvme_data));
1156 data->lba_shift = ilog2(lba_size);
1158 f->real_file_size = lba_size * nlba;
1159 fio_file_set_size_known(f);
1161 FILE_SET_ENG_DATA(f, data);
1164 return generic_get_file_size(td, f);
1167 static int fio_ioring_cmd_get_zoned_model(struct thread_data *td,
1169 enum zbd_zoned_model *model)
1171 return fio_nvme_get_zoned_model(td, f, model);
1174 static int fio_ioring_cmd_report_zones(struct thread_data *td,
1175 struct fio_file *f, uint64_t offset,
1176 struct zbd_zone *zbdz,
1177 unsigned int nr_zones)
1179 return fio_nvme_report_zones(td, f, offset, zbdz, nr_zones);
1182 static int fio_ioring_cmd_reset_wp(struct thread_data *td, struct fio_file *f,
1183 uint64_t offset, uint64_t length)
1185 return fio_nvme_reset_wp(td, f, offset, length);
1188 static int fio_ioring_cmd_get_max_open_zones(struct thread_data *td,
1190 unsigned int *max_open_zones)
1192 return fio_nvme_get_max_open_zones(td, f, max_open_zones);
1195 static struct ioengine_ops ioengine_uring = {
1197 .version = FIO_IOOPS_VERSION,
1198 .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD,
1199 .init = fio_ioring_init,
1200 .post_init = fio_ioring_post_init,
1201 .io_u_init = fio_ioring_io_u_init,
1202 .prep = fio_ioring_prep,
1203 .queue = fio_ioring_queue,
1204 .commit = fio_ioring_commit,
1205 .getevents = fio_ioring_getevents,
1206 .event = fio_ioring_event,
1207 .cleanup = fio_ioring_cleanup,
1208 .open_file = fio_ioring_open_file,
1209 .close_file = fio_ioring_close_file,
1210 .get_file_size = generic_get_file_size,
1212 .option_struct_size = sizeof(struct ioring_options),
1215 static struct ioengine_ops ioengine_uring_cmd = {
1216 .name = "io_uring_cmd",
1217 .version = FIO_IOOPS_VERSION,
1218 .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD | FIO_MEMALIGN | FIO_RAWIO,
1219 .init = fio_ioring_init,
1220 .post_init = fio_ioring_cmd_post_init,
1221 .io_u_init = fio_ioring_io_u_init,
1222 .prep = fio_ioring_cmd_prep,
1223 .queue = fio_ioring_queue,
1224 .commit = fio_ioring_commit,
1225 .getevents = fio_ioring_getevents,
1226 .event = fio_ioring_cmd_event,
1227 .cleanup = fio_ioring_cleanup,
1228 .open_file = fio_ioring_cmd_open_file,
1229 .close_file = fio_ioring_cmd_close_file,
1230 .get_file_size = fio_ioring_cmd_get_file_size,
1231 .get_zoned_model = fio_ioring_cmd_get_zoned_model,
1232 .report_zones = fio_ioring_cmd_report_zones,
1233 .reset_wp = fio_ioring_cmd_reset_wp,
1234 .get_max_open_zones = fio_ioring_cmd_get_max_open_zones,
1236 .option_struct_size = sizeof(struct ioring_options),
1239 static void fio_init fio_ioring_register(void)
1241 register_ioengine(&ioengine_uring);
1242 register_ioengine(&ioengine_uring_cmd);
1245 static void fio_exit fio_ioring_unregister(void)
1247 unregister_ioengine(&ioengine_uring);
1248 unregister_ioengine(&ioengine_uring_cmd);