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"
21 #ifdef ARCH_HAVE_IOURING
23 #include "../lib/types.h"
24 #include "../os/linux/io_uring.h"
30 unsigned *ring_entries;
39 unsigned *ring_entries;
40 struct io_uring_cqe *cqes;
51 struct io_u **io_u_index;
55 struct io_sq_ring sq_ring;
56 struct io_uring_sqe *sqes;
58 unsigned sq_ring_mask;
60 struct io_cq_ring cq_ring;
61 unsigned cq_ring_mask;
67 struct ioring_mmap mmap[3];
70 struct ioring_options {
73 unsigned int fixedbufs;
74 unsigned int registerfiles;
75 unsigned int sqpoll_thread;
76 unsigned int sqpoll_set;
77 unsigned int sqpoll_cpu;
78 unsigned int nonvectored;
79 unsigned int uncached;
82 static const int ddir_to_op[2][2] = {
83 { IORING_OP_READV, IORING_OP_READ },
84 { IORING_OP_WRITEV, IORING_OP_WRITE }
87 static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val)
89 struct ioring_options *o = data;
96 static struct fio_option options[] = {
99 .lname = "High Priority",
100 .type = FIO_OPT_STR_SET,
101 .off1 = offsetof(struct ioring_options, hipri),
102 .help = "Use polled IO completions",
103 .category = FIO_OPT_C_ENGINE,
104 .group = FIO_OPT_G_IOURING,
108 .lname = "Fixed (pre-mapped) IO buffers",
109 .type = FIO_OPT_STR_SET,
110 .off1 = offsetof(struct ioring_options, fixedbufs),
111 .help = "Pre map IO buffers",
112 .category = FIO_OPT_C_ENGINE,
113 .group = FIO_OPT_G_IOURING,
116 .name = "registerfiles",
117 .lname = "Register file set",
118 .type = FIO_OPT_STR_SET,
119 .off1 = offsetof(struct ioring_options, registerfiles),
120 .help = "Pre-open/register files",
121 .category = FIO_OPT_C_ENGINE,
122 .group = FIO_OPT_G_IOURING,
125 .name = "sqthread_poll",
126 .lname = "Kernel SQ thread polling",
128 .off1 = offsetof(struct ioring_options, sqpoll_thread),
129 .help = "Offload submission/completion to kernel thread",
130 .category = FIO_OPT_C_ENGINE,
131 .group = FIO_OPT_G_IOURING,
134 .name = "sqthread_poll_cpu",
135 .lname = "SQ Thread Poll CPU",
137 .cb = fio_ioring_sqpoll_cb,
138 .help = "What CPU to run SQ thread polling on",
139 .category = FIO_OPT_C_ENGINE,
140 .group = FIO_OPT_G_IOURING,
143 .name = "nonvectored",
144 .lname = "Non-vectored",
146 .off1 = offsetof(struct ioring_options, nonvectored),
147 .help = "Use non-vectored read/write commands",
148 .category = FIO_OPT_C_ENGINE,
149 .group = FIO_OPT_G_IOURING,
155 .off1 = offsetof(struct ioring_options, uncached),
156 .help = "Use RWF_UNCACHED for buffered read/writes",
157 .category = FIO_OPT_C_ENGINE,
158 .group = FIO_OPT_G_IOURING,
165 static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
166 unsigned int min_complete, unsigned int flags)
168 return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
169 min_complete, flags, NULL, 0);
172 static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
174 struct ioring_data *ld = td->io_ops_data;
175 struct ioring_options *o = td->eo;
176 struct fio_file *f = io_u->file;
177 struct io_uring_sqe *sqe;
179 sqe = &ld->sqes[io_u->index];
181 /* zero out fields not used in this submission */
182 memset(sqe, 0, sizeof(*sqe));
184 if (o->registerfiles) {
185 sqe->fd = f->engine_pos;
186 sqe->flags = IOSQE_FIXED_FILE;
191 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
192 sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored];
194 sqe->addr = (unsigned long) io_u->xfer_buf;
195 sqe->len = io_u->xfer_buflen;
196 sqe->buf_index = io_u->index;
198 if (o->nonvectored) {
199 sqe->addr = (unsigned long)
200 ld->iovecs[io_u->index].iov_base;
201 sqe->len = ld->iovecs[io_u->index].iov_len;
203 sqe->addr = (unsigned long) &ld->iovecs[io_u->index];
207 if (!td->o.odirect && o->uncached)
208 sqe->rw_flags = RWF_UNCACHED;
209 sqe->off = io_u->offset;
210 } else if (ddir_sync(io_u->ddir)) {
211 if (io_u->ddir == DDIR_SYNC_FILE_RANGE) {
212 sqe->off = f->first_write;
213 sqe->len = f->last_write - f->first_write;
214 sqe->sync_range_flags = td->o.sync_file_range;
215 sqe->opcode = IORING_OP_SYNC_FILE_RANGE;
217 if (io_u->ddir == DDIR_DATASYNC)
218 sqe->fsync_flags |= IORING_FSYNC_DATASYNC;
219 sqe->opcode = IORING_OP_FSYNC;
223 sqe->user_data = (unsigned long) io_u;
227 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
229 struct ioring_data *ld = td->io_ops_data;
230 struct io_uring_cqe *cqe;
234 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
236 cqe = &ld->cq_ring.cqes[index];
237 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
239 if (cqe->res != io_u->xfer_buflen) {
240 if (cqe->res > io_u->xfer_buflen)
241 io_u->error = -cqe->res;
243 io_u->resid = io_u->xfer_buflen - cqe->res;
250 static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
253 struct ioring_data *ld = td->io_ops_data;
254 struct io_cq_ring *ring = &ld->cq_ring;
255 unsigned head, reaped = 0;
260 if (head == *ring->tail)
264 } while (reaped + events < max);
271 static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
272 unsigned int max, const struct timespec *t)
274 struct ioring_data *ld = td->io_ops_data;
275 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
276 struct ioring_options *o = td->eo;
277 struct io_cq_ring *ring = &ld->cq_ring;
281 ld->cq_ring_off = *ring->head;
283 r = fio_ioring_cqring_reap(td, events, max);
291 if (!o->sqpoll_thread) {
292 r = io_uring_enter(ld, 0, actual_min,
293 IORING_ENTER_GETEVENTS);
295 if (errno == EAGAIN || errno == EINTR)
297 td_verror(td, errno, "io_uring_enter");
301 } while (events < min);
303 return r < 0 ? r : events;
306 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
309 struct ioring_data *ld = td->io_ops_data;
310 struct io_sq_ring *ring = &ld->sq_ring;
311 unsigned tail, next_tail;
313 fio_ro_check(td, io_u);
315 if (ld->queued == ld->iodepth)
318 if (io_u->ddir == DDIR_TRIM) {
322 do_io_u_trim(td, io_u);
323 io_u_mark_submit(td, 1);
324 io_u_mark_complete(td, 1);
325 return FIO_Q_COMPLETED;
329 next_tail = tail + 1;
331 if (next_tail == *ring->head)
334 /* ensure sqe stores are ordered with tail update */
336 ring->array[tail & ld->sq_ring_mask] = io_u->index;
337 *ring->tail = next_tail;
344 static void fio_ioring_queued(struct thread_data *td, int start, int nr)
346 struct ioring_data *ld = td->io_ops_data;
349 if (!fio_fill_issue_time(td))
352 fio_gettime(&now, NULL);
355 struct io_sq_ring *ring = &ld->sq_ring;
356 int index = ring->array[start & ld->sq_ring_mask];
357 struct io_u *io_u = ld->io_u_index[index];
359 memcpy(&io_u->issue_time, &now, sizeof(now));
360 io_u_queued(td, io_u);
366 static int fio_ioring_commit(struct thread_data *td)
368 struct ioring_data *ld = td->io_ops_data;
369 struct ioring_options *o = td->eo;
376 * Kernel side does submission. just need to check if the ring is
377 * flagged as needing a kick, if so, call io_uring_enter(). This
378 * only happens if we've been idle too long.
380 if (o->sqpoll_thread) {
381 struct io_sq_ring *ring = &ld->sq_ring;
384 if (*ring->flags & IORING_SQ_NEED_WAKEUP)
385 io_uring_enter(ld, ld->queued, 0,
386 IORING_ENTER_SQ_WAKEUP);
392 unsigned start = *ld->sq_ring.head;
393 long nr = ld->queued;
395 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
397 fio_ioring_queued(td, start, ret);
398 io_u_mark_submit(td, ret);
403 io_u_mark_submit(td, ret);
406 if (errno == EAGAIN || errno == EINTR) {
407 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
410 /* Shouldn't happen */
414 td_verror(td, errno, "io_uring_enter submit");
417 } while (ld->queued);
422 static void fio_ioring_unmap(struct ioring_data *ld)
426 for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
427 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
431 static void fio_ioring_cleanup(struct thread_data *td)
433 struct ioring_data *ld = td->io_ops_data;
436 if (!(td->flags & TD_F_CHILD))
437 fio_ioring_unmap(ld);
439 free(ld->io_u_index);
446 static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
448 struct io_sq_ring *sring = &ld->sq_ring;
449 struct io_cq_ring *cring = &ld->cq_ring;
452 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
453 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
454 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
456 ld->mmap[0].ptr = ptr;
457 sring->head = ptr + p->sq_off.head;
458 sring->tail = ptr + p->sq_off.tail;
459 sring->ring_mask = ptr + p->sq_off.ring_mask;
460 sring->ring_entries = ptr + p->sq_off.ring_entries;
461 sring->flags = ptr + p->sq_off.flags;
462 sring->array = ptr + p->sq_off.array;
463 ld->sq_ring_mask = *sring->ring_mask;
465 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
466 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
467 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
469 ld->mmap[1].ptr = ld->sqes;
471 ld->mmap[2].len = p->cq_off.cqes +
472 p->cq_entries * sizeof(struct io_uring_cqe);
473 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
474 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
476 ld->mmap[2].ptr = ptr;
477 cring->head = ptr + p->cq_off.head;
478 cring->tail = ptr + p->cq_off.tail;
479 cring->ring_mask = ptr + p->cq_off.ring_mask;
480 cring->ring_entries = ptr + p->cq_off.ring_entries;
481 cring->cqes = ptr + p->cq_off.cqes;
482 ld->cq_ring_mask = *cring->ring_mask;
486 static int fio_ioring_queue_init(struct thread_data *td)
488 struct ioring_data *ld = td->io_ops_data;
489 struct ioring_options *o = td->eo;
490 int depth = td->o.iodepth;
491 struct io_uring_params p;
494 memset(&p, 0, sizeof(p));
497 p.flags |= IORING_SETUP_IOPOLL;
498 if (o->sqpoll_thread) {
499 p.flags |= IORING_SETUP_SQPOLL;
501 p.flags |= IORING_SETUP_SQ_AFF;
502 p.sq_thread_cpu = o->sqpoll_cpu;
506 ret = syscall(__NR_sys_io_uring_setup, depth, &p);
513 struct rlimit rlim = {
514 .rlim_cur = RLIM_INFINITY,
515 .rlim_max = RLIM_INFINITY,
518 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0)
521 ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
522 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
527 return fio_ioring_mmap(ld, &p);
530 static int fio_ioring_register_files(struct thread_data *td)
532 struct ioring_data *ld = td->io_ops_data;
537 ld->fds = calloc(td->o.nr_files, sizeof(int));
539 for_each_file(td, f, i) {
540 ret = generic_open_file(td, f);
547 ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
548 IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
556 * Pretend the file is closed again, and really close it if we hit
559 for_each_file(td, f, i) {
562 ret2 = generic_close_file(td, f);
570 static int fio_ioring_post_init(struct thread_data *td)
572 struct ioring_data *ld = td->io_ops_data;
573 struct ioring_options *o = td->eo;
577 for (i = 0; i < td->o.iodepth; i++) {
578 struct iovec *iov = &ld->iovecs[i];
580 io_u = ld->io_u_index[i];
581 iov->iov_base = io_u->buf;
582 iov->iov_len = td_max_bs(td);
585 err = fio_ioring_queue_init(td);
587 td_verror(td, errno, "io_queue_init");
591 if (o->registerfiles) {
592 err = fio_ioring_register_files(td);
594 td_verror(td, errno, "ioring_register_files");
602 static unsigned roundup_pow2(unsigned depth)
604 return 1UL << __fls(depth - 1);
607 static int fio_ioring_init(struct thread_data *td)
609 struct ioring_options *o = td->eo;
610 struct ioring_data *ld;
612 /* sqthread submission requires registered files */
613 if (o->sqpoll_thread)
614 o->registerfiles = 1;
616 if (o->registerfiles && td->o.nr_files != td->o.open_files) {
617 log_err("fio: io_uring registered files require nr_files to "
618 "be identical to open_files\n");
622 ld = calloc(1, sizeof(*ld));
624 /* ring depth must be a power-of-2 */
625 ld->iodepth = td->o.iodepth;
626 td->o.iodepth = roundup_pow2(td->o.iodepth);
629 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
630 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
632 td->io_ops_data = ld;
636 static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
638 struct ioring_data *ld = td->io_ops_data;
640 ld->io_u_index[io_u->index] = io_u;
644 static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
646 struct ioring_data *ld = td->io_ops_data;
647 struct ioring_options *o = td->eo;
649 if (!ld || !o->registerfiles)
650 return generic_open_file(td, f);
652 f->fd = ld->fds[f->engine_pos];
656 static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
658 struct ioring_data *ld = td->io_ops_data;
659 struct ioring_options *o = td->eo;
661 if (!ld || !o->registerfiles)
662 return generic_close_file(td, f);
668 static struct ioengine_ops ioengine = {
670 .version = FIO_IOOPS_VERSION,
671 .flags = FIO_ASYNCIO_SYNC_TRIM,
672 .init = fio_ioring_init,
673 .post_init = fio_ioring_post_init,
674 .io_u_init = fio_ioring_io_u_init,
675 .prep = fio_ioring_prep,
676 .queue = fio_ioring_queue,
677 .commit = fio_ioring_commit,
678 .getevents = fio_ioring_getevents,
679 .event = fio_ioring_event,
680 .cleanup = fio_ioring_cleanup,
681 .open_file = fio_ioring_open_file,
682 .close_file = fio_ioring_close_file,
683 .get_file_size = generic_get_file_size,
685 .option_struct_size = sizeof(struct ioring_options),
688 static void fio_init fio_ioring_register(void)
690 register_ioengine(&ioengine);
693 static void fio_exit fio_ioring_unregister(void)
695 unregister_ioengine(&ioengine);