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;
80 static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val)
82 struct ioring_options *o = data;
89 static struct fio_option options[] = {
92 .lname = "High Priority",
93 .type = FIO_OPT_STR_SET,
94 .off1 = offsetof(struct ioring_options, hipri),
95 .help = "Use polled IO completions",
96 .category = FIO_OPT_C_ENGINE,
97 .group = FIO_OPT_G_IOURING,
101 .lname = "Fixed (pre-mapped) IO buffers",
102 .type = FIO_OPT_STR_SET,
103 .off1 = offsetof(struct ioring_options, fixedbufs),
104 .help = "Pre map IO buffers",
105 .category = FIO_OPT_C_ENGINE,
106 .group = FIO_OPT_G_IOURING,
109 .name = "registerfiles",
110 .lname = "Register file set",
111 .type = FIO_OPT_STR_SET,
112 .off1 = offsetof(struct ioring_options, registerfiles),
113 .help = "Pre-open/register files",
114 .category = FIO_OPT_C_ENGINE,
115 .group = FIO_OPT_G_IOURING,
118 .name = "sqthread_poll",
119 .lname = "Kernel SQ thread polling",
121 .off1 = offsetof(struct ioring_options, sqpoll_thread),
122 .help = "Offload submission/completion to kernel thread",
123 .category = FIO_OPT_C_ENGINE,
124 .group = FIO_OPT_G_IOURING,
127 .name = "sqthread_poll_cpu",
128 .lname = "SQ Thread Poll CPU",
130 .cb = fio_ioring_sqpoll_cb,
131 .help = "What CPU to run SQ thread polling on",
132 .category = FIO_OPT_C_ENGINE,
133 .group = FIO_OPT_G_IOURING,
140 static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
141 unsigned int min_complete, unsigned int flags)
143 return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
144 min_complete, flags, NULL, 0);
147 static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
149 struct ioring_data *ld = td->io_ops_data;
150 struct ioring_options *o = td->eo;
151 struct fio_file *f = io_u->file;
152 struct io_uring_sqe *sqe;
154 sqe = &ld->sqes[io_u->index];
155 if (o->registerfiles) {
156 sqe->fd = f->engine_pos;
157 sqe->flags = IOSQE_FIXED_FILE;
165 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
167 if (io_u->ddir == DDIR_READ)
168 sqe->opcode = IORING_OP_READ_FIXED;
170 sqe->opcode = IORING_OP_WRITE_FIXED;
171 sqe->addr = (unsigned long) io_u->xfer_buf;
172 sqe->len = io_u->xfer_buflen;
173 sqe->buf_index = io_u->index;
175 if (io_u->ddir == DDIR_READ)
176 sqe->opcode = IORING_OP_READV;
178 sqe->opcode = IORING_OP_WRITEV;
179 sqe->addr = (unsigned long) &ld->iovecs[io_u->index];
182 sqe->off = io_u->offset;
183 } else if (ddir_sync(io_u->ddir)) {
184 sqe->fsync_flags = 0;
185 if (io_u->ddir == DDIR_DATASYNC)
186 sqe->fsync_flags |= IORING_FSYNC_DATASYNC;
187 sqe->opcode = IORING_OP_FSYNC;
190 sqe->user_data = (unsigned long) io_u;
194 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
196 struct ioring_data *ld = td->io_ops_data;
197 struct io_uring_cqe *cqe;
201 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
203 cqe = &ld->cq_ring.cqes[index];
204 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
206 if (cqe->res != io_u->xfer_buflen) {
207 if (cqe->res > io_u->xfer_buflen)
208 io_u->error = -cqe->res;
210 io_u->resid = io_u->xfer_buflen - cqe->res;
217 static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
220 struct ioring_data *ld = td->io_ops_data;
221 struct io_cq_ring *ring = &ld->cq_ring;
222 unsigned head, reaped = 0;
227 if (head == *ring->tail)
231 } while (reaped + events < max);
238 static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
239 unsigned int max, const struct timespec *t)
241 struct ioring_data *ld = td->io_ops_data;
242 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
243 struct ioring_options *o = td->eo;
244 struct io_cq_ring *ring = &ld->cq_ring;
248 ld->cq_ring_off = *ring->head;
250 r = fio_ioring_cqring_reap(td, events, max);
258 if (!o->sqpoll_thread) {
259 r = io_uring_enter(ld, 0, actual_min,
260 IORING_ENTER_GETEVENTS);
264 td_verror(td, errno, "io_uring_enter");
268 } while (events < min);
270 return r < 0 ? r : events;
273 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
276 struct ioring_data *ld = td->io_ops_data;
277 struct io_sq_ring *ring = &ld->sq_ring;
278 unsigned tail, next_tail;
280 fio_ro_check(td, io_u);
282 if (ld->queued == ld->iodepth)
285 if (io_u->ddir == DDIR_TRIM) {
289 do_io_u_trim(td, io_u);
290 io_u_mark_submit(td, 1);
291 io_u_mark_complete(td, 1);
292 return FIO_Q_COMPLETED;
296 next_tail = tail + 1;
298 if (next_tail == *ring->head)
301 /* ensure sqe stores are ordered with tail update */
303 ring->array[tail & ld->sq_ring_mask] = io_u->index;
304 *ring->tail = next_tail;
311 static void fio_ioring_queued(struct thread_data *td, int start, int nr)
313 struct ioring_data *ld = td->io_ops_data;
316 if (!fio_fill_issue_time(td))
319 fio_gettime(&now, NULL);
322 struct io_sq_ring *ring = &ld->sq_ring;
323 int index = ring->array[start & ld->sq_ring_mask];
324 struct io_u *io_u = ld->io_u_index[index];
326 memcpy(&io_u->issue_time, &now, sizeof(now));
327 io_u_queued(td, io_u);
333 static int fio_ioring_commit(struct thread_data *td)
335 struct ioring_data *ld = td->io_ops_data;
336 struct ioring_options *o = td->eo;
343 * Kernel side does submission. just need to check if the ring is
344 * flagged as needing a kick, if so, call io_uring_enter(). This
345 * only happens if we've been idle too long.
347 if (o->sqpoll_thread) {
348 struct io_sq_ring *ring = &ld->sq_ring;
351 if (*ring->flags & IORING_SQ_NEED_WAKEUP)
352 io_uring_enter(ld, ld->queued, 0,
353 IORING_ENTER_SQ_WAKEUP);
359 unsigned start = *ld->sq_ring.head;
360 long nr = ld->queued;
362 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
364 fio_ioring_queued(td, start, ret);
365 io_u_mark_submit(td, ret);
370 io_u_mark_submit(td, ret);
373 if (errno == EAGAIN) {
374 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
377 /* Shouldn't happen */
381 td_verror(td, errno, "io_uring_enter submit");
384 } while (ld->queued);
389 static void fio_ioring_unmap(struct ioring_data *ld)
393 for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
394 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
398 static void fio_ioring_cleanup(struct thread_data *td)
400 struct ioring_data *ld = td->io_ops_data;
403 if (!(td->flags & TD_F_CHILD))
404 fio_ioring_unmap(ld);
406 free(ld->io_u_index);
413 static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
415 struct io_sq_ring *sring = &ld->sq_ring;
416 struct io_cq_ring *cring = &ld->cq_ring;
419 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
420 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
421 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
423 ld->mmap[0].ptr = ptr;
424 sring->head = ptr + p->sq_off.head;
425 sring->tail = ptr + p->sq_off.tail;
426 sring->ring_mask = ptr + p->sq_off.ring_mask;
427 sring->ring_entries = ptr + p->sq_off.ring_entries;
428 sring->flags = ptr + p->sq_off.flags;
429 sring->array = ptr + p->sq_off.array;
430 ld->sq_ring_mask = *sring->ring_mask;
432 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
433 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
434 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
436 ld->mmap[1].ptr = ld->sqes;
438 ld->mmap[2].len = p->cq_off.cqes +
439 p->cq_entries * sizeof(struct io_uring_cqe);
440 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
441 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
443 ld->mmap[2].ptr = ptr;
444 cring->head = ptr + p->cq_off.head;
445 cring->tail = ptr + p->cq_off.tail;
446 cring->ring_mask = ptr + p->cq_off.ring_mask;
447 cring->ring_entries = ptr + p->cq_off.ring_entries;
448 cring->cqes = ptr + p->cq_off.cqes;
449 ld->cq_ring_mask = *cring->ring_mask;
453 static int fio_ioring_queue_init(struct thread_data *td)
455 struct ioring_data *ld = td->io_ops_data;
456 struct ioring_options *o = td->eo;
457 int depth = td->o.iodepth;
458 struct io_uring_params p;
461 memset(&p, 0, sizeof(p));
464 p.flags |= IORING_SETUP_IOPOLL;
465 if (o->sqpoll_thread) {
466 p.flags |= IORING_SETUP_SQPOLL;
468 p.flags |= IORING_SETUP_SQ_AFF;
469 p.sq_thread_cpu = o->sqpoll_cpu;
473 ret = syscall(__NR_sys_io_uring_setup, depth, &p);
480 struct rlimit rlim = {
481 .rlim_cur = RLIM_INFINITY,
482 .rlim_max = RLIM_INFINITY,
485 if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0)
488 ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
489 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
494 return fio_ioring_mmap(ld, &p);
497 static int fio_ioring_register_files(struct thread_data *td)
499 struct ioring_data *ld = td->io_ops_data;
504 ld->fds = calloc(td->o.nr_files, sizeof(int));
506 for_each_file(td, f, i) {
507 ret = generic_open_file(td, f);
514 ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
515 IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
523 * Pretend the file is closed again, and really close it if we hit
526 for_each_file(td, f, i) {
529 ret2 = generic_close_file(td, f);
537 static int fio_ioring_post_init(struct thread_data *td)
539 struct ioring_data *ld = td->io_ops_data;
540 struct ioring_options *o = td->eo;
544 for (i = 0; i < td->o.iodepth; i++) {
545 struct iovec *iov = &ld->iovecs[i];
547 io_u = ld->io_u_index[i];
548 iov->iov_base = io_u->buf;
549 iov->iov_len = td_max_bs(td);
552 err = fio_ioring_queue_init(td);
554 td_verror(td, errno, "io_queue_init");
558 if (o->registerfiles) {
559 err = fio_ioring_register_files(td);
561 td_verror(td, errno, "ioring_register_files");
569 static unsigned roundup_pow2(unsigned depth)
571 return 1UL << __fls(depth - 1);
574 static int fio_ioring_init(struct thread_data *td)
576 struct ioring_options *o = td->eo;
577 struct ioring_data *ld;
579 /* sqthread submission requires registered files */
580 if (o->sqpoll_thread)
581 o->registerfiles = 1;
583 if (o->registerfiles && td->o.nr_files != td->o.open_files) {
584 log_err("fio: io_uring registered files require nr_files to "
585 "be identical to open_files\n");
589 ld = calloc(1, sizeof(*ld));
591 /* ring depth must be a power-of-2 */
592 ld->iodepth = td->o.iodepth;
593 td->o.iodepth = roundup_pow2(td->o.iodepth);
596 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
597 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
599 td->io_ops_data = ld;
603 static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
605 struct ioring_data *ld = td->io_ops_data;
607 ld->io_u_index[io_u->index] = io_u;
611 static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
613 struct ioring_data *ld = td->io_ops_data;
614 struct ioring_options *o = td->eo;
616 if (!o->registerfiles)
617 return generic_open_file(td, f);
619 f->fd = ld->fds[f->engine_pos];
623 static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
625 struct ioring_options *o = td->eo;
627 if (!o->registerfiles)
628 return generic_close_file(td, f);
634 static struct ioengine_ops ioengine = {
636 .version = FIO_IOOPS_VERSION,
637 .flags = FIO_ASYNCIO_SYNC_TRIM,
638 .init = fio_ioring_init,
639 .post_init = fio_ioring_post_init,
640 .io_u_init = fio_ioring_io_u_init,
641 .prep = fio_ioring_prep,
642 .queue = fio_ioring_queue,
643 .commit = fio_ioring_commit,
644 .getevents = fio_ioring_getevents,
645 .event = fio_ioring_event,
646 .cleanup = fio_ioring_cleanup,
647 .open_file = fio_ioring_open_file,
648 .close_file = fio_ioring_close_file,
649 .get_file_size = generic_get_file_size,
651 .option_struct_size = sizeof(struct ioring_options),
654 static void fio_init fio_ioring_register(void)
656 register_ioengine(&ioengine);
659 static void fio_exit fio_ioring_unregister(void)
661 unregister_ioengine(&ioengine);