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/io_uring.h"
30 unsigned *ring_entries;
39 unsigned *ring_entries;
40 struct io_uring_cqe *cqes;
52 struct io_u **io_u_index;
54 struct io_sq_ring sq_ring;
55 struct io_uring_sqe *sqes;
57 unsigned sq_ring_mask;
59 struct io_cq_ring cq_ring;
60 unsigned cq_ring_mask;
69 struct ioring_mmap mmap[3];
72 struct ioring_options {
75 unsigned int fixedbufs;
76 unsigned int sqthread;
77 unsigned int sqthread_set;
78 unsigned int sqthread_poll;
82 static int fio_ioring_sqthread_cb(void *data, unsigned long long *val)
84 struct ioring_options *o = data;
91 static struct fio_option options[] = {
94 .lname = "High Priority",
95 .type = FIO_OPT_STR_SET,
96 .off1 = offsetof(struct ioring_options, hipri),
97 .help = "Use polled IO completions",
98 .category = FIO_OPT_C_ENGINE,
99 .group = FIO_OPT_G_LIBAIO,
103 .lname = "Fixed (pre-mapped) IO buffers",
104 .type = FIO_OPT_STR_SET,
105 .off1 = offsetof(struct ioring_options, fixedbufs),
106 .help = "Pre map IO buffers",
107 .category = FIO_OPT_C_ENGINE,
108 .group = FIO_OPT_G_LIBAIO,
112 .lname = "Use kernel SQ thread on this CPU",
114 .cb = fio_ioring_sqthread_cb,
115 .help = "Offload submission to kernel thread",
116 .category = FIO_OPT_C_ENGINE,
117 .group = FIO_OPT_G_LIBAIO,
120 .name = "sqthread_poll",
121 .lname = "Kernel SQ thread should poll",
122 .type = FIO_OPT_STR_SET,
123 .off1 = offsetof(struct ioring_options, sqthread_poll),
124 .help = "Used with sqthread, enables kernel side polling",
125 .category = FIO_OPT_C_ENGINE,
126 .group = FIO_OPT_G_LIBAIO,
130 .lname = "Offload submission to kernel workqueue",
131 .type = FIO_OPT_STR_SET,
132 .off1 = offsetof(struct ioring_options, sqwq),
133 .help = "Offload submission to kernel workqueue",
134 .category = FIO_OPT_C_ENGINE,
135 .group = FIO_OPT_G_LIBAIO,
142 static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
143 unsigned int min_complete, unsigned int flags)
145 return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
146 min_complete, flags);
149 static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
151 struct ioring_data *ld = td->io_ops_data;
152 struct ioring_options *o = td->eo;
153 struct fio_file *f = io_u->file;
154 struct io_uring_sqe *sqe;
156 sqe = &ld->sqes[io_u->index];
161 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
163 if (io_u->ddir == DDIR_READ)
164 sqe->opcode = IORING_OP_READ_FIXED;
166 sqe->opcode = IORING_OP_WRITE_FIXED;
167 sqe->addr = io_u->xfer_buf;
168 sqe->len = io_u->xfer_buflen;
170 if (io_u->ddir == DDIR_READ)
171 sqe->opcode = IORING_OP_READV;
173 sqe->opcode = IORING_OP_WRITEV;
174 sqe->addr = &ld->iovecs[io_u->index];
177 sqe->off = io_u->offset;
178 } else if (ddir_sync(io_u->ddir))
179 sqe->opcode = IORING_OP_FSYNC;
184 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
186 struct ioring_data *ld = td->io_ops_data;
187 struct io_uring_cqe *cqe;
191 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
193 cqe = &ld->cq_ring.cqes[index];
194 io_u = ld->io_u_index[cqe->index];
196 if (cqe->res != io_u->xfer_buflen) {
197 if (cqe->res > io_u->xfer_buflen)
198 io_u->error = -cqe->res;
200 io_u->resid = io_u->xfer_buflen - cqe->res;
204 if (io_u->ddir == DDIR_READ) {
205 if (cqe->flags & IOCQE_FLAG_CACHEHIT)
214 static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
217 struct ioring_data *ld = td->io_ops_data;
218 struct io_cq_ring *ring = &ld->cq_ring;
219 unsigned head, reaped = 0;
224 if (head == *ring->tail)
228 } while (reaped + events < max);
235 static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
236 unsigned int max, const struct timespec *t)
238 struct ioring_data *ld = td->io_ops_data;
239 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
240 struct ioring_options *o = td->eo;
241 struct io_cq_ring *ring = &ld->cq_ring;
245 ld->cq_ring_off = *ring->head;
247 r = fio_ioring_cqring_reap(td, events, max);
253 if (!o->sqthread_poll) {
254 r = io_uring_enter(ld, 0, actual_min,
255 IORING_ENTER_GETEVENTS);
259 td_verror(td, errno, "io_uring_enter");
263 } while (events < min);
265 return r < 0 ? r : events;
268 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
271 struct ioring_data *ld = td->io_ops_data;
272 struct io_sq_ring *ring = &ld->sq_ring;
273 unsigned tail, next_tail;
275 fio_ro_check(td, io_u);
277 if (ld->queued == ld->iodepth)
280 if (io_u->ddir == DDIR_TRIM) {
284 do_io_u_trim(td, io_u);
285 io_u_mark_submit(td, 1);
286 io_u_mark_complete(td, 1);
287 return FIO_Q_COMPLETED;
291 next_tail = tail + 1;
293 if (next_tail == *ring->head)
296 ring->array[tail & ld->sq_ring_mask] = io_u->index;
297 *ring->tail = next_tail;
304 static void fio_ioring_queued(struct thread_data *td, int start, int nr)
306 struct ioring_data *ld = td->io_ops_data;
309 if (!fio_fill_issue_time(td))
312 fio_gettime(&now, NULL);
315 struct io_sq_ring *ring = &ld->sq_ring;
316 int index = ring->array[start & ld->sq_ring_mask];
317 struct io_u *io_u = ld->io_u_index[index];
319 memcpy(&io_u->issue_time, &now, sizeof(now));
320 io_u_queued(td, io_u);
326 static int fio_ioring_commit(struct thread_data *td)
328 struct ioring_data *ld = td->io_ops_data;
329 struct ioring_options *o = td->eo;
336 if (o->sqthread_poll) {
337 struct io_sq_ring *ring = &ld->sq_ring;
339 if (*ring->flags & IORING_SQ_NEED_WAKEUP)
340 io_uring_enter(ld, ld->queued, 0, 0);
346 unsigned start = *ld->sq_ring.head;
347 long nr = ld->queued;
349 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
351 fio_ioring_queued(td, start, ret);
352 io_u_mark_submit(td, ret);
357 io_u_mark_submit(td, ret);
360 if (errno == EAGAIN) {
361 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
364 /* Shouldn't happen */
368 td_verror(td, errno, "io_uring_enter submit");
371 } while (ld->queued);
376 static void fio_ioring_unmap(struct ioring_data *ld)
380 for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
381 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
385 static void fio_ioring_cleanup(struct thread_data *td)
387 struct ioring_data *ld = td->io_ops_data;
390 td->ts.cachehit += ld->cachehit;
391 td->ts.cachemiss += ld->cachemiss;
393 if (!(td->flags & TD_F_CHILD))
394 fio_ioring_unmap(ld);
396 free(ld->io_u_index);
403 static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
405 struct io_sq_ring *sring = &ld->sq_ring;
406 struct io_cq_ring *cring = &ld->cq_ring;
409 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
410 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
411 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
413 ld->mmap[0].ptr = ptr;
414 sring->head = ptr + p->sq_off.head;
415 sring->tail = ptr + p->sq_off.tail;
416 sring->ring_mask = ptr + p->sq_off.ring_mask;
417 sring->ring_entries = ptr + p->sq_off.ring_entries;
418 sring->flags = ptr + p->sq_off.flags;
419 sring->array = ptr + p->sq_off.array;
420 ld->sq_ring_mask = *sring->ring_mask;
422 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
423 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
424 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
426 ld->mmap[1].ptr = ld->sqes;
428 ld->mmap[2].len = p->cq_off.cqes +
429 p->cq_entries * sizeof(struct io_uring_cqe);
430 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
431 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
433 ld->mmap[2].ptr = ptr;
434 cring->head = ptr + p->cq_off.head;
435 cring->tail = ptr + p->cq_off.tail;
436 cring->ring_mask = ptr + p->cq_off.ring_mask;
437 cring->ring_entries = ptr + p->cq_off.ring_entries;
438 cring->cqes = ptr + p->cq_off.cqes;
439 ld->cq_ring_mask = *cring->ring_mask;
443 static int fio_ioring_queue_init(struct thread_data *td)
445 struct ioring_data *ld = td->io_ops_data;
446 struct ioring_options *o = td->eo;
447 int depth = td->o.iodepth;
448 struct iovec *vecs = NULL;
449 struct io_uring_params p;
452 memset(&p, 0, sizeof(p));
455 p.flags |= IORING_SETUP_IOPOLL;
456 if (o->sqthread_set) {
457 p.sq_thread_cpu = o->sqthread;
458 p.flags |= IORING_SETUP_SQTHREAD;
459 if (o->sqthread_poll)
460 p.flags |= IORING_SETUP_SQPOLL;
463 p.flags |= IORING_SETUP_SQWQ;
466 struct rlimit rlim = {
467 .rlim_cur = RLIM_INFINITY,
468 .rlim_max = RLIM_INFINITY,
471 setrlimit(RLIMIT_MEMLOCK, &rlim);
475 ret = syscall(__NR_sys_io_uring_setup, depth, vecs, &p);
480 return fio_ioring_mmap(ld, &p);
483 static int fio_ioring_post_init(struct thread_data *td)
485 struct ioring_data *ld = td->io_ops_data;
489 for (i = 0; i < td->o.iodepth; i++) {
490 struct iovec *iov = &ld->iovecs[i];
492 io_u = ld->io_u_index[i];
493 iov->iov_base = io_u->buf;
494 iov->iov_len = td_max_bs(td);
497 err = fio_ioring_queue_init(td);
499 td_verror(td, errno, "io_queue_init");
506 static unsigned roundup_pow2(unsigned depth)
508 return 1UL << __fls(depth - 1);
511 static int fio_ioring_init(struct thread_data *td)
513 struct ioring_data *ld;
515 ld = calloc(1, sizeof(*ld));
517 /* ring depth must be a power-of-2 */
518 ld->iodepth = td->o.iodepth;
519 td->o.iodepth = roundup_pow2(td->o.iodepth);
522 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
523 ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
524 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
526 td->io_ops_data = ld;
530 static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
532 struct ioring_data *ld = td->io_ops_data;
534 ld->io_u_index[io_u->index] = io_u;
538 static struct ioengine_ops ioengine = {
540 .version = FIO_IOOPS_VERSION,
541 .init = fio_ioring_init,
542 .post_init = fio_ioring_post_init,
543 .io_u_init = fio_ioring_io_u_init,
544 .prep = fio_ioring_prep,
545 .queue = fio_ioring_queue,
546 .commit = fio_ioring_commit,
547 .getevents = fio_ioring_getevents,
548 .event = fio_ioring_event,
549 .cleanup = fio_ioring_cleanup,
550 .open_file = generic_open_file,
551 .close_file = generic_close_file,
552 .get_file_size = generic_get_file_size,
554 .option_struct_size = sizeof(struct ioring_options),
557 static void fio_init fio_ioring_register(void)
559 register_ioengine(&ioengine);
562 static void fio_exit fio_ioring_unregister(void)
564 unregister_ioengine(&ioengine);