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_event *events;
52 struct io_u **io_u_index;
54 struct io_sq_ring sq_ring;
55 struct io_uring_iocb *iocbs;
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 fio_file *f = io_u->file;
153 struct io_uring_iocb *iocb;
155 iocb = &ld->iocbs[io_u->index];
160 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
161 if (io_u->ddir == DDIR_READ)
162 iocb->opcode = IORING_OP_READ;
164 iocb->opcode = IORING_OP_WRITE;
165 iocb->off = io_u->offset;
166 iocb->addr = io_u->xfer_buf;
167 iocb->len = io_u->xfer_buflen;
168 } else if (ddir_sync(io_u->ddir))
169 iocb->opcode = IORING_OP_FSYNC;
174 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
176 struct ioring_data *ld = td->io_ops_data;
177 struct io_uring_event *ev;
181 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
183 ev = &ld->cq_ring.events[index];
184 io_u = ld->io_u_index[ev->index];
186 if (ev->res != io_u->xfer_buflen) {
187 if (ev->res > io_u->xfer_buflen)
188 io_u->error = -ev->res;
190 io_u->resid = io_u->xfer_buflen - ev->res;
194 if (io_u->ddir == DDIR_READ) {
195 if (ev->flags & IOEV_FLAG_CACHEHIT)
204 static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
207 struct ioring_data *ld = td->io_ops_data;
208 struct io_cq_ring *ring = &ld->cq_ring;
209 unsigned head, reaped = 0;
214 if (head == *ring->tail)
218 } while (reaped + events < max);
225 static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
226 unsigned int max, const struct timespec *t)
228 struct ioring_data *ld = td->io_ops_data;
229 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
230 struct ioring_options *o = td->eo;
231 struct io_cq_ring *ring = &ld->cq_ring;
235 ld->cq_ring_off = *ring->head;
237 r = fio_ioring_cqring_reap(td, events, max);
243 if (!o->sqthread_poll) {
244 r = io_uring_enter(ld, 0, actual_min,
245 IORING_ENTER_GETEVENTS);
249 td_verror(td, errno, "io_uring_enter");
253 } while (events < min);
255 return r < 0 ? r : events;
258 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
261 struct ioring_data *ld = td->io_ops_data;
262 struct io_sq_ring *ring = &ld->sq_ring;
263 unsigned tail, next_tail;
265 fio_ro_check(td, io_u);
267 if (ld->queued == ld->iodepth)
270 if (io_u->ddir == DDIR_TRIM) {
274 do_io_u_trim(td, io_u);
275 io_u_mark_submit(td, 1);
276 io_u_mark_complete(td, 1);
277 return FIO_Q_COMPLETED;
281 next_tail = tail + 1;
283 if (next_tail == *ring->head)
286 ring->array[tail & ld->sq_ring_mask] = io_u->index;
287 *ring->tail = next_tail;
294 static void fio_ioring_queued(struct thread_data *td, int start, int nr)
296 struct ioring_data *ld = td->io_ops_data;
299 if (!fio_fill_issue_time(td))
302 fio_gettime(&now, NULL);
305 struct io_sq_ring *ring = &ld->sq_ring;
306 int index = ring->array[start & ld->sq_ring_mask];
307 struct io_u *io_u = ld->io_u_index[index];
309 memcpy(&io_u->issue_time, &now, sizeof(now));
310 io_u_queued(td, io_u);
316 static int fio_ioring_commit(struct thread_data *td)
318 struct ioring_data *ld = td->io_ops_data;
319 struct ioring_options *o = td->eo;
326 if (o->sqthread_poll) {
327 struct io_sq_ring *ring = &ld->sq_ring;
329 if (*ring->flags & IORING_SQ_NEED_WAKEUP)
330 io_uring_enter(ld, ld->queued, 0, 0);
336 unsigned start = *ld->sq_ring.head;
337 long nr = ld->queued;
339 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
341 fio_ioring_queued(td, start, ret);
342 io_u_mark_submit(td, ret);
347 io_u_mark_submit(td, ret);
350 if (errno == EAGAIN) {
351 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
354 /* Shouldn't happen */
358 td_verror(td, errno, "io_uring_enter submit");
361 } while (ld->queued);
366 static void fio_ioring_unmap(struct ioring_data *ld)
370 for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
371 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
375 static void fio_ioring_cleanup(struct thread_data *td)
377 struct ioring_data *ld = td->io_ops_data;
380 td->ts.cachehit += ld->cachehit;
381 td->ts.cachemiss += ld->cachemiss;
383 if (!(td->flags & TD_F_CHILD))
384 fio_ioring_unmap(ld);
386 free(ld->io_u_index);
393 static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
395 struct io_sq_ring *sring = &ld->sq_ring;
396 struct io_cq_ring *cring = &ld->cq_ring;
399 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
400 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
401 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
403 ld->mmap[0].ptr = ptr;
404 sring->head = ptr + p->sq_off.head;
405 sring->tail = ptr + p->sq_off.tail;
406 sring->ring_mask = ptr + p->sq_off.ring_mask;
407 sring->ring_entries = ptr + p->sq_off.ring_entries;
408 sring->flags = ptr + p->sq_off.flags;
409 sring->array = ptr + p->sq_off.array;
410 ld->sq_ring_mask = *sring->ring_mask;
412 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_iocb);
413 ld->iocbs = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
414 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
416 ld->mmap[1].ptr = ld->iocbs;
418 ld->mmap[2].len = p->cq_off.events +
419 p->cq_entries * sizeof(struct io_uring_event);
420 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
421 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
423 ld->mmap[2].ptr = ptr;
424 cring->head = ptr + p->cq_off.head;
425 cring->tail = ptr + p->cq_off.tail;
426 cring->ring_mask = ptr + p->cq_off.ring_mask;
427 cring->ring_entries = ptr + p->cq_off.ring_entries;
428 cring->events = ptr + p->cq_off.events;
429 ld->cq_ring_mask = *cring->ring_mask;
433 static int fio_ioring_queue_init(struct thread_data *td)
435 struct ioring_data *ld = td->io_ops_data;
436 struct ioring_options *o = td->eo;
437 int depth = td->o.iodepth;
438 struct io_uring_params p;
441 memset(&p, 0, sizeof(p));
444 p.flags |= IORING_SETUP_IOPOLL;
445 if (o->sqthread_set) {
446 p.sq_thread_cpu = o->sqthread;
447 p.flags |= IORING_SETUP_SQTHREAD;
448 if (o->sqthread_poll)
449 p.flags |= IORING_SETUP_SQPOLL;
452 p.flags |= IORING_SETUP_SQWQ;
455 struct rlimit rlim = {
456 .rlim_cur = RLIM_INFINITY,
457 .rlim_max = RLIM_INFINITY,
460 setrlimit(RLIMIT_MEMLOCK, &rlim);
461 p.flags |= IORING_SETUP_FIXEDBUFS;
464 ret = syscall(__NR_sys_io_uring_setup, depth, ld->iovecs, &p);
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;
475 struct ioring_options *o = td->eo;
482 for (i = 0; i < td->o.iodepth; i++) {
483 struct iovec *iov = &ld->iovecs[i];
485 io_u = ld->io_u_index[i];
486 iov->iov_base = io_u->buf;
487 iov->iov_len = td_max_bs(td);
491 err = fio_ioring_queue_init(td);
493 td_verror(td, errno, "io_queue_init");
500 static unsigned roundup_pow2(unsigned depth)
502 return 1UL << __fls(depth - 1);
505 static int fio_ioring_init(struct thread_data *td)
507 struct ioring_data *ld;
509 ld = calloc(1, sizeof(*ld));
511 /* ring depth must be a power-of-2 */
512 ld->iodepth = td->o.iodepth;
513 td->o.iodepth = roundup_pow2(td->o.iodepth);
516 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
517 ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
519 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
521 td->io_ops_data = ld;
525 static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
527 struct ioring_data *ld = td->io_ops_data;
529 ld->io_u_index[io_u->index] = io_u;
533 static struct ioengine_ops ioengine = {
535 .version = FIO_IOOPS_VERSION,
536 .init = fio_ioring_init,
537 .post_init = fio_ioring_post_init,
538 .io_u_init = fio_ioring_io_u_init,
539 .prep = fio_ioring_prep,
540 .queue = fio_ioring_queue,
541 .commit = fio_ioring_commit,
542 .getevents = fio_ioring_getevents,
543 .event = fio_ioring_event,
544 .cleanup = fio_ioring_cleanup,
545 .open_file = generic_open_file,
546 .close_file = generic_close_file,
547 .get_file_size = generic_get_file_size,
549 .option_struct_size = sizeof(struct ioring_options),
552 static void fio_init fio_ioring_register(void)
554 register_ioengine(&ioengine);
557 static void fio_exit fio_ioring_unregister(void)
559 unregister_ioengine(&ioengine);