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 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_LIBAIO,
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_LIBAIO,
109 .name = "sqthread_poll",
110 .lname = "Kernel SQ thread should poll",
112 .cb = fio_ioring_sqpoll_cb,
113 .help = "Offload submission to kernel thread",
114 .category = FIO_OPT_C_ENGINE,
115 .group = FIO_OPT_G_LIBAIO,
122 static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
123 unsigned int min_complete, unsigned int flags)
125 return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
126 min_complete, flags);
129 static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
131 struct ioring_data *ld = td->io_ops_data;
132 struct ioring_options *o = td->eo;
133 struct fio_file *f = io_u->file;
134 struct io_uring_sqe *sqe;
136 sqe = &ld->sqes[io_u->index];
142 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
143 if (io_u->ddir == DDIR_READ)
144 sqe->opcode = IORING_OP_READV;
146 sqe->opcode = IORING_OP_WRITEV;
149 sqe->flags |= IOSQE_FIXED_BUFFER;
150 sqe->addr = io_u->xfer_buf;
151 sqe->len = io_u->xfer_buflen;
152 sqe->buf_index = io_u->index;
154 sqe->addr = &ld->iovecs[io_u->index];
157 sqe->off = io_u->offset;
158 } else if (ddir_sync(io_u->ddir))
159 sqe->opcode = IORING_OP_FSYNC;
161 sqe->data = (unsigned long) io_u;
165 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
167 struct ioring_data *ld = td->io_ops_data;
168 struct io_uring_cqe *cqe;
172 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
174 cqe = &ld->cq_ring.cqes[index];
175 io_u = (struct io_u *) cqe->data;
177 if (cqe->res != io_u->xfer_buflen) {
178 if (cqe->res > io_u->xfer_buflen)
179 io_u->error = -cqe->res;
181 io_u->resid = io_u->xfer_buflen - cqe->res;
185 if (io_u->ddir == DDIR_READ) {
186 if (cqe->flags & IOCQE_FLAG_CACHEHIT)
195 static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
198 struct ioring_data *ld = td->io_ops_data;
199 struct io_cq_ring *ring = &ld->cq_ring;
200 unsigned head, reaped = 0;
205 if (head == *ring->tail)
209 } while (reaped + events < max);
216 static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
217 unsigned int max, const struct timespec *t)
219 struct ioring_data *ld = td->io_ops_data;
220 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
221 struct ioring_options *o = td->eo;
222 struct io_cq_ring *ring = &ld->cq_ring;
226 ld->cq_ring_off = *ring->head;
228 r = fio_ioring_cqring_reap(td, events, max);
234 if (!o->sqpoll_set) {
235 r = io_uring_enter(ld, 0, actual_min,
236 IORING_ENTER_GETEVENTS);
240 td_verror(td, errno, "io_uring_enter");
244 } while (events < min);
246 return r < 0 ? r : events;
249 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
252 struct ioring_data *ld = td->io_ops_data;
253 struct io_sq_ring *ring = &ld->sq_ring;
254 unsigned tail, next_tail;
256 fio_ro_check(td, io_u);
258 if (ld->queued == ld->iodepth)
261 if (io_u->ddir == DDIR_TRIM) {
265 do_io_u_trim(td, io_u);
266 io_u_mark_submit(td, 1);
267 io_u_mark_complete(td, 1);
268 return FIO_Q_COMPLETED;
272 next_tail = tail + 1;
274 if (next_tail == *ring->head)
277 ring->array[tail & ld->sq_ring_mask] = io_u->index;
278 *ring->tail = next_tail;
285 static void fio_ioring_queued(struct thread_data *td, int start, int nr)
287 struct ioring_data *ld = td->io_ops_data;
290 if (!fio_fill_issue_time(td))
293 fio_gettime(&now, NULL);
296 struct io_sq_ring *ring = &ld->sq_ring;
297 int index = ring->array[start & ld->sq_ring_mask];
298 struct io_u *io_u = ld->io_u_index[index];
300 memcpy(&io_u->issue_time, &now, sizeof(now));
301 io_u_queued(td, io_u);
307 static int fio_ioring_commit(struct thread_data *td)
309 struct ioring_data *ld = td->io_ops_data;
310 struct ioring_options *o = td->eo;
318 struct io_sq_ring *ring = &ld->sq_ring;
321 if (*ring->flags & IORING_SQ_NEED_WAKEUP)
322 io_uring_enter(ld, ld->queued, 0, 0);
328 unsigned start = *ld->sq_ring.head;
329 long nr = ld->queued;
331 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
333 fio_ioring_queued(td, start, ret);
334 io_u_mark_submit(td, ret);
339 io_u_mark_submit(td, ret);
342 if (errno == EAGAIN) {
343 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
346 /* Shouldn't happen */
350 td_verror(td, errno, "io_uring_enter submit");
353 } while (ld->queued);
358 static void fio_ioring_unmap(struct ioring_data *ld)
362 for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
363 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
367 static void fio_ioring_cleanup(struct thread_data *td)
369 struct ioring_data *ld = td->io_ops_data;
372 td->ts.cachehit += ld->cachehit;
373 td->ts.cachemiss += ld->cachemiss;
375 if (!(td->flags & TD_F_CHILD))
376 fio_ioring_unmap(ld);
378 free(ld->io_u_index);
385 static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
387 struct io_sq_ring *sring = &ld->sq_ring;
388 struct io_cq_ring *cring = &ld->cq_ring;
391 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
392 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
393 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
395 ld->mmap[0].ptr = ptr;
396 sring->head = ptr + p->sq_off.head;
397 sring->tail = ptr + p->sq_off.tail;
398 sring->ring_mask = ptr + p->sq_off.ring_mask;
399 sring->ring_entries = ptr + p->sq_off.ring_entries;
400 sring->flags = ptr + p->sq_off.flags;
401 sring->array = ptr + p->sq_off.array;
402 ld->sq_ring_mask = *sring->ring_mask;
404 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
405 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
406 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
408 ld->mmap[1].ptr = ld->sqes;
410 ld->mmap[2].len = p->cq_off.cqes +
411 p->cq_entries * sizeof(struct io_uring_cqe);
412 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
413 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
415 ld->mmap[2].ptr = ptr;
416 cring->head = ptr + p->cq_off.head;
417 cring->tail = ptr + p->cq_off.tail;
418 cring->ring_mask = ptr + p->cq_off.ring_mask;
419 cring->ring_entries = ptr + p->cq_off.ring_entries;
420 cring->cqes = ptr + p->cq_off.cqes;
421 ld->cq_ring_mask = *cring->ring_mask;
425 static int fio_ioring_queue_init(struct thread_data *td)
427 struct ioring_data *ld = td->io_ops_data;
428 struct ioring_options *o = td->eo;
429 int depth = td->o.iodepth;
430 struct io_uring_params p;
433 memset(&p, 0, sizeof(p));
436 p.flags |= IORING_SETUP_IOPOLL;
438 p.flags |= IORING_SETUP_SQPOLL | IORING_SETUP_SQ_AFF;
439 p.sq_thread_cpu = o->sqpoll_cpu;
443 struct rlimit rlim = {
444 .rlim_cur = RLIM_INFINITY,
445 .rlim_max = RLIM_INFINITY,
448 setrlimit(RLIMIT_MEMLOCK, &rlim);
451 ret = syscall(__NR_sys_io_uring_setup, depth, &p);
458 struct io_uring_register_buffers reg = {
459 .iovecs = ld->iovecs,
463 ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
464 IORING_REGISTER_BUFFERS, ®);
469 return fio_ioring_mmap(ld, &p);
472 static int fio_ioring_post_init(struct thread_data *td)
474 struct ioring_data *ld = td->io_ops_data;
478 for (i = 0; i < td->o.iodepth; i++) {
479 struct iovec *iov = &ld->iovecs[i];
481 io_u = ld->io_u_index[i];
482 iov->iov_base = io_u->buf;
483 iov->iov_len = td_max_bs(td);
486 err = fio_ioring_queue_init(td);
488 td_verror(td, errno, "io_queue_init");
495 static unsigned roundup_pow2(unsigned depth)
497 return 1UL << __fls(depth - 1);
500 static int fio_ioring_init(struct thread_data *td)
502 struct ioring_data *ld;
504 ld = calloc(1, sizeof(*ld));
506 /* ring depth must be a power-of-2 */
507 ld->iodepth = td->o.iodepth;
508 td->o.iodepth = roundup_pow2(td->o.iodepth);
511 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
512 ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
513 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
515 td->io_ops_data = ld;
519 static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
521 struct ioring_data *ld = td->io_ops_data;
523 ld->io_u_index[io_u->index] = io_u;
527 static struct ioengine_ops ioengine = {
529 .version = FIO_IOOPS_VERSION,
530 .init = fio_ioring_init,
531 .post_init = fio_ioring_post_init,
532 .io_u_init = fio_ioring_io_u_init,
533 .prep = fio_ioring_prep,
534 .queue = fio_ioring_queue,
535 .commit = fio_ioring_commit,
536 .getevents = fio_ioring_getevents,
537 .event = fio_ioring_event,
538 .cleanup = fio_ioring_cleanup,
539 .open_file = generic_open_file,
540 .close_file = generic_close_file,
541 .get_file_size = generic_get_file_size,
543 .option_struct_size = sizeof(struct ioring_options),
546 static void fio_init fio_ioring_register(void)
548 register_ioengine(&ioengine);
551 static void fio_exit fio_ioring_unregister(void)
553 unregister_ioengine(&ioengine);