projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
HOWTO: improve description of latency measures
[fio.git] / engines / io_uring.c
... / ...
CommitLineData
1/*
2 * io_uring engine
3 *
4 * IO engine using the new native Linux aio io_uring interface. See:
5 *
6 * http://git.kernel.dk/cgit/linux-block/log/?h=io_uring
7 *
8 */
9#include <stdlib.h>
10#include <unistd.h>
11#include <errno.h>
12#include <sys/time.h>
13#include <sys/resource.h>
14
15#include "../fio.h"
16#include "../lib/pow2.h"
17#include "../optgroup.h"
18#include "../lib/memalign.h"
19#include "../lib/fls.h"
20#include "../lib/roundup.h"
21
22#ifdef ARCH_HAVE_IOURING
23
24#include "../lib/types.h"
25#include "../os/linux/io_uring.h"
26#include "cmdprio.h"
27#include "nvme.h"
28
29#include <sys/stat.h>
30
31enum uring_cmd_type {
32 FIO_URING_CMD_NVME = 1,
33};
34
35struct io_sq_ring {
36 unsigned *head;
37 unsigned *tail;
38 unsigned *ring_mask;
39 unsigned *ring_entries;
40 unsigned *flags;
41 unsigned *array;
42};
43
44struct io_cq_ring {
45 unsigned *head;
46 unsigned *tail;
47 unsigned *ring_mask;
48 unsigned *ring_entries;
49 struct io_uring_cqe *cqes;
50};
51
52struct ioring_mmap {
53 void *ptr;
54 size_t len;
55};
56
57struct ioring_data {
58 int ring_fd;
59
60 struct io_u **io_u_index;
61
62 int *fds;
63
64 struct io_sq_ring sq_ring;
65 struct io_uring_sqe *sqes;
66 struct iovec *iovecs;
67 unsigned sq_ring_mask;
68
69 struct io_cq_ring cq_ring;
70 unsigned cq_ring_mask;
71
72 int queued;
73 int cq_ring_off;
74 unsigned iodepth;
75 int prepped;
76
77 struct ioring_mmap mmap[3];
78
79 struct cmdprio cmdprio;
80};
81
82struct ioring_options {
83 struct thread_data *td;
84 unsigned int hipri;
85 struct cmdprio_options cmdprio_options;
86 unsigned int fixedbufs;
87 unsigned int registerfiles;
88 unsigned int sqpoll_thread;
89 unsigned int sqpoll_set;
90 unsigned int sqpoll_cpu;
91 unsigned int nonvectored;
92 unsigned int uncached;
93 unsigned int nowait;
94 unsigned int force_async;
95 enum uring_cmd_type cmd_type;
96};
97
98static const int ddir_to_op[2][2] = {
99 { IORING_OP_READV, IORING_OP_READ },
100 { IORING_OP_WRITEV, IORING_OP_WRITE }
101};
102
103static const int fixed_ddir_to_op[2] = {
104 IORING_OP_READ_FIXED,
105 IORING_OP_WRITE_FIXED
106};
107
108static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val)
109{
110 struct ioring_options *o = data;
111
112 o->sqpoll_cpu = *val;
113 o->sqpoll_set = 1;
114 return 0;
115}
116
117static struct fio_option options[] = {
118 {
119 .name = "hipri",
120 .lname = "High Priority",
121 .type = FIO_OPT_STR_SET,
122 .off1 = offsetof(struct ioring_options, hipri),
123 .help = "Use polled IO completions",
124 .category = FIO_OPT_C_ENGINE,
125 .group = FIO_OPT_G_IOURING,
126 },
127#ifdef FIO_HAVE_IOPRIO_CLASS
128 {
129 .name = "cmdprio_percentage",
130 .lname = "high priority percentage",
131 .type = FIO_OPT_INT,
132 .off1 = offsetof(struct ioring_options,
133 cmdprio_options.percentage[DDIR_READ]),
134 .off2 = offsetof(struct ioring_options,
135 cmdprio_options.percentage[DDIR_WRITE]),
136 .minval = 0,
137 .maxval = 100,
138 .help = "Send high priority I/O this percentage of the time",
139 .category = FIO_OPT_C_ENGINE,
140 .group = FIO_OPT_G_IOURING,
141 },
142 {
143 .name = "cmdprio_class",
144 .lname = "Asynchronous I/O priority class",
145 .type = FIO_OPT_INT,
146 .off1 = offsetof(struct ioring_options,
147 cmdprio_options.class[DDIR_READ]),
148 .off2 = offsetof(struct ioring_options,
149 cmdprio_options.class[DDIR_WRITE]),
150 .help = "Set asynchronous IO priority class",
151 .minval = IOPRIO_MIN_PRIO_CLASS + 1,
152 .maxval = IOPRIO_MAX_PRIO_CLASS,
153 .interval = 1,
154 .category = FIO_OPT_C_ENGINE,
155 .group = FIO_OPT_G_IOURING,
156 },
157 {
158 .name = "cmdprio",
159 .lname = "Asynchronous I/O priority level",
160 .type = FIO_OPT_INT,
161 .off1 = offsetof(struct ioring_options,
162 cmdprio_options.level[DDIR_READ]),
163 .off2 = offsetof(struct ioring_options,
164 cmdprio_options.level[DDIR_WRITE]),
165 .help = "Set asynchronous IO priority level",
166 .minval = IOPRIO_MIN_PRIO,
167 .maxval = IOPRIO_MAX_PRIO,
168 .interval = 1,
169 .category = FIO_OPT_C_ENGINE,
170 .group = FIO_OPT_G_IOURING,
171 },
172 {
173 .name = "cmdprio_bssplit",
174 .lname = "Priority percentage block size split",
175 .type = FIO_OPT_STR_STORE,
176 .off1 = offsetof(struct ioring_options,
177 cmdprio_options.bssplit_str),
178 .help = "Set priority percentages for different block sizes",
179 .category = FIO_OPT_C_ENGINE,
180 .group = FIO_OPT_G_IOURING,
181 },
182#else
183 {
184 .name = "cmdprio_percentage",
185 .lname = "high priority percentage",
186 .type = FIO_OPT_UNSUPPORTED,
187 .help = "Your platform does not support I/O priority classes",
188 },
189 {
190 .name = "cmdprio_class",
191 .lname = "Asynchronous I/O priority class",
192 .type = FIO_OPT_UNSUPPORTED,
193 .help = "Your platform does not support I/O priority classes",
194 },
195 {
196 .name = "cmdprio",
197 .lname = "Asynchronous I/O priority level",
198 .type = FIO_OPT_UNSUPPORTED,
199 .help = "Your platform does not support I/O priority classes",
200 },
201 {
202 .name = "cmdprio_bssplit",
203 .lname = "Priority percentage block size split",
204 .type = FIO_OPT_UNSUPPORTED,
205 .help = "Your platform does not support I/O priority classes",
206 },
207#endif
208 {
209 .name = "fixedbufs",
210 .lname = "Fixed (pre-mapped) IO buffers",
211 .type = FIO_OPT_STR_SET,
212 .off1 = offsetof(struct ioring_options, fixedbufs),
213 .help = "Pre map IO buffers",
214 .category = FIO_OPT_C_ENGINE,
215 .group = FIO_OPT_G_IOURING,
216 },
217 {
218 .name = "registerfiles",
219 .lname = "Register file set",
220 .type = FIO_OPT_STR_SET,
221 .off1 = offsetof(struct ioring_options, registerfiles),
222 .help = "Pre-open/register files",
223 .category = FIO_OPT_C_ENGINE,
224 .group = FIO_OPT_G_IOURING,
225 },
226 {
227 .name = "sqthread_poll",
228 .lname = "Kernel SQ thread polling",
229 .type = FIO_OPT_INT,
230 .off1 = offsetof(struct ioring_options, sqpoll_thread),
231 .help = "Offload submission/completion to kernel thread",
232 .category = FIO_OPT_C_ENGINE,
233 .group = FIO_OPT_G_IOURING,
234 },
235 {
236 .name = "sqthread_poll_cpu",
237 .lname = "SQ Thread Poll CPU",
238 .type = FIO_OPT_INT,
239 .cb = fio_ioring_sqpoll_cb,
240 .help = "What CPU to run SQ thread polling on",
241 .category = FIO_OPT_C_ENGINE,
242 .group = FIO_OPT_G_IOURING,
243 },
244 {
245 .name = "nonvectored",
246 .lname = "Non-vectored",
247 .type = FIO_OPT_INT,
248 .off1 = offsetof(struct ioring_options, nonvectored),
249 .def = "-1",
250 .help = "Use non-vectored read/write commands",
251 .category = FIO_OPT_C_ENGINE,
252 .group = FIO_OPT_G_IOURING,
253 },
254 {
255 .name = "uncached",
256 .lname = "Uncached",
257 .type = FIO_OPT_INT,
258 .off1 = offsetof(struct ioring_options, uncached),
259 .help = "Use RWF_UNCACHED for buffered read/writes",
260 .category = FIO_OPT_C_ENGINE,
261 .group = FIO_OPT_G_IOURING,
262 },
263 {
264 .name = "nowait",
265 .lname = "RWF_NOWAIT",
266 .type = FIO_OPT_BOOL,
267 .off1 = offsetof(struct ioring_options, nowait),
268 .help = "Use RWF_NOWAIT for reads/writes",
269 .category = FIO_OPT_C_ENGINE,
270 .group = FIO_OPT_G_IOURING,
271 },
272 {
273 .name = "force_async",
274 .lname = "Force async",
275 .type = FIO_OPT_INT,
276 .off1 = offsetof(struct ioring_options, force_async),
277 .help = "Set IOSQE_ASYNC every N requests",
278 .category = FIO_OPT_C_ENGINE,
279 .group = FIO_OPT_G_IOURING,
280 },
281 {
282 .name = "cmd_type",
283 .lname = "Uring cmd type",
284 .type = FIO_OPT_STR,
285 .off1 = offsetof(struct ioring_options, cmd_type),
286 .help = "Specify uring-cmd type",
287 .def = "nvme",
288 .posval = {
289 { .ival = "nvme",
290 .oval = FIO_URING_CMD_NVME,
291 .help = "Issue nvme-uring-cmd",
292 },
293 },
294 .category = FIO_OPT_C_ENGINE,
295 .group = FIO_OPT_G_IOURING,
296 },
297 {
298 .name = NULL,
299 },
300};
301
302static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
303 unsigned int min_complete, unsigned int flags)
304{
305#ifdef FIO_ARCH_HAS_SYSCALL
306 return __do_syscall6(__NR_io_uring_enter, ld->ring_fd, to_submit,
307 min_complete, flags, NULL, 0);
308#else
309 return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit,
310 min_complete, flags, NULL, 0);
311#endif
312}
313
314static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
315{
316 struct ioring_data *ld = td->io_ops_data;
317 struct ioring_options *o = td->eo;
318 struct fio_file *f = io_u->file;
319 struct io_uring_sqe *sqe;
320
321 sqe = &ld->sqes[io_u->index];
322
323 if (o->registerfiles) {
324 sqe->fd = f->engine_pos;
325 sqe->flags = IOSQE_FIXED_FILE;
326 } else {
327 sqe->fd = f->fd;
328 sqe->flags = 0;
329 }
330
331 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
332 if (o->fixedbufs) {
333 sqe->opcode = fixed_ddir_to_op[io_u->ddir];
334 sqe->addr = (unsigned long) io_u->xfer_buf;
335 sqe->len = io_u->xfer_buflen;
336 sqe->buf_index = io_u->index;
337 } else {
338 struct iovec *iov = &ld->iovecs[io_u->index];
339
340 /*
341 * Update based on actual io_u, requeue could have
342 * adjusted these
343 */
344 iov->iov_base = io_u->xfer_buf;
345 iov->iov_len = io_u->xfer_buflen;
346
347 sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored];
348 if (o->nonvectored) {
349 sqe->addr = (unsigned long) iov->iov_base;
350 sqe->len = iov->iov_len;
351 } else {
352 sqe->addr = (unsigned long) iov;
353 sqe->len = 1;
354 }
355 }
356 sqe->rw_flags = 0;
357 if (!td->o.odirect && o->uncached)
358 sqe->rw_flags |= RWF_UNCACHED;
359 if (o->nowait)
360 sqe->rw_flags |= RWF_NOWAIT;
361
362 /*
363 * Since io_uring can have a submission context (sqthread_poll)
364 * that is different from the process context, we cannot rely on
365 * the IO priority set by ioprio_set() (option prio/prioclass)
366 * to be inherited.
367 * td->ioprio will have the value of the "default prio", so set
368 * this unconditionally. This value might get overridden by
369 * fio_ioring_cmdprio_prep() if the option cmdprio_percentage or
370 * cmdprio_bssplit is used.
371 */
372 sqe->ioprio = td->ioprio;
373 sqe->off = io_u->offset;
374 } else if (ddir_sync(io_u->ddir)) {
375 sqe->ioprio = 0;
376 if (io_u->ddir == DDIR_SYNC_FILE_RANGE) {
377 sqe->off = f->first_write;
378 sqe->len = f->last_write - f->first_write;
379 sqe->sync_range_flags = td->o.sync_file_range;
380 sqe->opcode = IORING_OP_SYNC_FILE_RANGE;
381 } else {
382 sqe->off = 0;
383 sqe->addr = 0;
384 sqe->len = 0;
385 if (io_u->ddir == DDIR_DATASYNC)
386 sqe->fsync_flags |= IORING_FSYNC_DATASYNC;
387 sqe->opcode = IORING_OP_FSYNC;
388 }
389 }
390
391 if (o->force_async && ++ld->prepped == o->force_async) {
392 ld->prepped = 0;
393 sqe->flags |= IOSQE_ASYNC;
394 }
395
396 sqe->user_data = (unsigned long) io_u;
397 return 0;
398}
399
400static int fio_ioring_cmd_prep(struct thread_data *td, struct io_u *io_u)
401{
402 struct ioring_data *ld = td->io_ops_data;
403 struct ioring_options *o = td->eo;
404 struct fio_file *f = io_u->file;
405 struct nvme_uring_cmd *cmd;
406 struct io_uring_sqe *sqe;
407
408 /* only supports nvme_uring_cmd */
409 if (o->cmd_type != FIO_URING_CMD_NVME)
410 return -EINVAL;
411
412 sqe = &ld->sqes[(io_u->index) << 1];
413
414 if (o->registerfiles) {
415 sqe->fd = f->engine_pos;
416 sqe->flags = IOSQE_FIXED_FILE;
417 } else {
418 sqe->fd = f->fd;
419 }
420 sqe->rw_flags = 0;
421 if (!td->o.odirect && o->uncached)
422 sqe->rw_flags |= RWF_UNCACHED;
423 if (o->nowait)
424 sqe->rw_flags |= RWF_NOWAIT;
425
426 sqe->opcode = IORING_OP_URING_CMD;
427 sqe->user_data = (unsigned long) io_u;
428 if (o->nonvectored)
429 sqe->cmd_op = NVME_URING_CMD_IO;
430 else
431 sqe->cmd_op = NVME_URING_CMD_IO_VEC;
432 if (o->force_async && ++ld->prepped == o->force_async) {
433 ld->prepped = 0;
434 sqe->flags |= IOSQE_ASYNC;
435 }
436
437 cmd = (struct nvme_uring_cmd *)sqe->cmd;
438 return fio_nvme_uring_cmd_prep(cmd, io_u,
439 o->nonvectored ? NULL : &ld->iovecs[io_u->index]);
440}
441
442static struct io_u *fio_ioring_event(struct thread_data *td, int event)
443{
444 struct ioring_data *ld = td->io_ops_data;
445 struct io_uring_cqe *cqe;
446 struct io_u *io_u;
447 unsigned index;
448
449 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
450
451 cqe = &ld->cq_ring.cqes[index];
452 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
453
454 if (cqe->res != io_u->xfer_buflen) {
455 if (cqe->res > io_u->xfer_buflen)
456 io_u->error = -cqe->res;
457 else
458 io_u->resid = io_u->xfer_buflen - cqe->res;
459 } else
460 io_u->error = 0;
461
462 return io_u;
463}
464
465static struct io_u *fio_ioring_cmd_event(struct thread_data *td, int event)
466{
467 struct ioring_data *ld = td->io_ops_data;
468 struct ioring_options *o = td->eo;
469 struct io_uring_cqe *cqe;
470 struct io_u *io_u;
471 unsigned index;
472
473 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
474 if (o->cmd_type == FIO_URING_CMD_NVME)
475 index <<= 1;
476
477 cqe = &ld->cq_ring.cqes[index];
478 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
479
480 if (cqe->res != 0)
481 io_u->error = -cqe->res;
482 else
483 io_u->error = 0;
484
485 return io_u;
486}
487
488static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
489 unsigned int max)
490{
491 struct ioring_data *ld = td->io_ops_data;
492 struct io_cq_ring *ring = &ld->cq_ring;
493 unsigned head, reaped = 0;
494
495 head = *ring->head;
496 do {
497 if (head == atomic_load_acquire(ring->tail))
498 break;
499 reaped++;
500 head++;
501 } while (reaped + events < max);
502
503 if (reaped)
504 atomic_store_release(ring->head, head);
505
506 return reaped;
507}
508
509static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
510 unsigned int max, const struct timespec *t)
511{
512 struct ioring_data *ld = td->io_ops_data;
513 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
514 struct ioring_options *o = td->eo;
515 struct io_cq_ring *ring = &ld->cq_ring;
516 unsigned events = 0;
517 int r;
518
519 ld->cq_ring_off = *ring->head;
520 do {
521 r = fio_ioring_cqring_reap(td, events, max);
522 if (r) {
523 events += r;
524 if (actual_min != 0)
525 actual_min -= r;
526 continue;
527 }
528
529 if (!o->sqpoll_thread) {
530 r = io_uring_enter(ld, 0, actual_min,
531 IORING_ENTER_GETEVENTS);
532 if (r < 0) {
533 if (errno == EAGAIN || errno == EINTR)
534 continue;
535 td_verror(td, errno, "io_uring_enter");
536 break;
537 }
538 }
539 } while (events < min);
540
541 return r < 0 ? r : events;
542}
543
544static inline void fio_ioring_cmdprio_prep(struct thread_data *td,
545 struct io_u *io_u)
546{
547 struct ioring_data *ld = td->io_ops_data;
548 struct cmdprio *cmdprio = &ld->cmdprio;
549
550 if (fio_cmdprio_set_ioprio(td, cmdprio, io_u))
551 ld->sqes[io_u->index].ioprio = io_u->ioprio;
552}
553
554static enum fio_q_status fio_ioring_queue(struct thread_data *td,
555 struct io_u *io_u)
556{
557 struct ioring_data *ld = td->io_ops_data;
558 struct io_sq_ring *ring = &ld->sq_ring;
559 unsigned tail, next_tail;
560
561 fio_ro_check(td, io_u);
562
563 if (ld->queued == ld->iodepth)
564 return FIO_Q_BUSY;
565
566 if (io_u->ddir == DDIR_TRIM) {
567 if (ld->queued)
568 return FIO_Q_BUSY;
569
570 do_io_u_trim(td, io_u);
571 io_u_mark_submit(td, 1);
572 io_u_mark_complete(td, 1);
573 return FIO_Q_COMPLETED;
574 }
575
576 tail = *ring->tail;
577 next_tail = tail + 1;
578 if (next_tail == atomic_load_acquire(ring->head))
579 return FIO_Q_BUSY;
580
581 if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
582 fio_ioring_cmdprio_prep(td, io_u);
583
584 ring->array[tail & ld->sq_ring_mask] = io_u->index;
585 atomic_store_release(ring->tail, next_tail);
586
587 ld->queued++;
588 return FIO_Q_QUEUED;
589}
590
591static void fio_ioring_queued(struct thread_data *td, int start, int nr)
592{
593 struct ioring_data *ld = td->io_ops_data;
594 struct timespec now;
595
596 if (!fio_fill_issue_time(td))
597 return;
598
599 fio_gettime(&now, NULL);
600
601 while (nr--) {
602 struct io_sq_ring *ring = &ld->sq_ring;
603 int index = ring->array[start & ld->sq_ring_mask];
604 struct io_u *io_u = ld->io_u_index[index];
605
606 memcpy(&io_u->issue_time, &now, sizeof(now));
607 io_u_queued(td, io_u);
608
609 start++;
610 }
611}
612
613static int fio_ioring_commit(struct thread_data *td)
614{
615 struct ioring_data *ld = td->io_ops_data;
616 struct ioring_options *o = td->eo;
617 int ret;
618
619 if (!ld->queued)
620 return 0;
621
622 /*
623 * Kernel side does submission. just need to check if the ring is
624 * flagged as needing a kick, if so, call io_uring_enter(). This
625 * only happens if we've been idle too long.
626 */
627 if (o->sqpoll_thread) {
628 struct io_sq_ring *ring = &ld->sq_ring;
629 unsigned flags;
630
631 flags = atomic_load_acquire(ring->flags);
632 if (flags & IORING_SQ_NEED_WAKEUP)
633 io_uring_enter(ld, ld->queued, 0,
634 IORING_ENTER_SQ_WAKEUP);
635 ld->queued = 0;
636 return 0;
637 }
638
639 do {
640 unsigned start = *ld->sq_ring.head;
641 long nr = ld->queued;
642
643 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
644 if (ret > 0) {
645 fio_ioring_queued(td, start, ret);
646 io_u_mark_submit(td, ret);
647
648 ld->queued -= ret;
649 ret = 0;
650 } else if (!ret) {
651 io_u_mark_submit(td, ret);
652 continue;
653 } else {
654 if (errno == EAGAIN || errno == EINTR) {
655 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
656 if (ret)
657 continue;
658 /* Shouldn't happen */
659 usleep(1);
660 continue;
661 }
662 td_verror(td, errno, "io_uring_enter submit");
663 break;
664 }
665 } while (ld->queued);
666
667 return ret;
668}
669
670static void fio_ioring_unmap(struct ioring_data *ld)
671{
672 int i;
673
674 for (i = 0; i < FIO_ARRAY_SIZE(ld->mmap); i++)
675 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
676 close(ld->ring_fd);
677}
678
679static void fio_ioring_cleanup(struct thread_data *td)
680{
681 struct ioring_data *ld = td->io_ops_data;
682
683 if (ld) {
684 if (!(td->flags & TD_F_CHILD))
685 fio_ioring_unmap(ld);
686
687 fio_cmdprio_cleanup(&ld->cmdprio);
688 free(ld->io_u_index);
689 free(ld->iovecs);
690 free(ld->fds);
691 free(ld);
692 }
693}
694
695static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
696{
697 struct io_sq_ring *sring = &ld->sq_ring;
698 struct io_cq_ring *cring = &ld->cq_ring;
699 void *ptr;
700
701 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
702 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
703 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
704 IORING_OFF_SQ_RING);
705 ld->mmap[0].ptr = ptr;
706 sring->head = ptr + p->sq_off.head;
707 sring->tail = ptr + p->sq_off.tail;
708 sring->ring_mask = ptr + p->sq_off.ring_mask;
709 sring->ring_entries = ptr + p->sq_off.ring_entries;
710 sring->flags = ptr + p->sq_off.flags;
711 sring->array = ptr + p->sq_off.array;
712 ld->sq_ring_mask = *sring->ring_mask;
713
714 if (p->flags & IORING_SETUP_SQE128)
715 ld->mmap[1].len = 2 * p->sq_entries * sizeof(struct io_uring_sqe);
716 else
717 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
718 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
719 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
720 IORING_OFF_SQES);
721 ld->mmap[1].ptr = ld->sqes;
722
723 if (p->flags & IORING_SETUP_CQE32) {
724 ld->mmap[2].len = p->cq_off.cqes +
725 2 * p->cq_entries * sizeof(struct io_uring_cqe);
726 } else {
727 ld->mmap[2].len = p->cq_off.cqes +
728 p->cq_entries * sizeof(struct io_uring_cqe);
729 }
730 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
731 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
732 IORING_OFF_CQ_RING);
733 ld->mmap[2].ptr = ptr;
734 cring->head = ptr + p->cq_off.head;
735 cring->tail = ptr + p->cq_off.tail;
736 cring->ring_mask = ptr + p->cq_off.ring_mask;
737 cring->ring_entries = ptr + p->cq_off.ring_entries;
738 cring->cqes = ptr + p->cq_off.cqes;
739 ld->cq_ring_mask = *cring->ring_mask;
740 return 0;
741}
742
743static void fio_ioring_probe(struct thread_data *td)
744{
745 struct ioring_data *ld = td->io_ops_data;
746 struct ioring_options *o = td->eo;
747 struct io_uring_probe *p;
748 int ret;
749
750 /* already set by user, don't touch */
751 if (o->nonvectored != -1)
752 return;
753
754 /* default to off, as that's always safe */
755 o->nonvectored = 0;
756
757 p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
758 if (!p)
759 return;
760
761 memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
762 ret = syscall(__NR_io_uring_register, ld->ring_fd,
763 IORING_REGISTER_PROBE, p, 256);
764 if (ret < 0)
765 goto out;
766
767 if (IORING_OP_WRITE > p->ops_len)
768 goto out;
769
770 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) &&
771 (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED))
772 o->nonvectored = 1;
773out:
774 free(p);
775}
776
777static int fio_ioring_queue_init(struct thread_data *td)
778{
779 struct ioring_data *ld = td->io_ops_data;
780 struct ioring_options *o = td->eo;
781 int depth = td->o.iodepth;
782 struct io_uring_params p;
783 int ret;
784
785 memset(&p, 0, sizeof(p));
786
787 if (o->hipri)
788 p.flags |= IORING_SETUP_IOPOLL;
789 if (o->sqpoll_thread) {
790 p.flags |= IORING_SETUP_SQPOLL;
791 if (o->sqpoll_set) {
792 p.flags |= IORING_SETUP_SQ_AFF;
793 p.sq_thread_cpu = o->sqpoll_cpu;
794 }
795 }
796
797 /*
798 * Clamp CQ ring size at our SQ ring size, we don't need more entries
799 * than that.
800 */
801 p.flags |= IORING_SETUP_CQSIZE;
802 p.cq_entries = depth;
803
804retry:
805 ret = syscall(__NR_io_uring_setup, depth, &p);
806 if (ret < 0) {
807 if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) {
808 p.flags &= ~IORING_SETUP_CQSIZE;
809 goto retry;
810 }
811 return ret;
812 }
813
814 ld->ring_fd = ret;
815
816 fio_ioring_probe(td);
817
818 if (o->fixedbufs) {
819 ret = syscall(__NR_io_uring_register, ld->ring_fd,
820 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
821 if (ret < 0)
822 return ret;
823 }
824
825 return fio_ioring_mmap(ld, &p);
826}
827
828static int fio_ioring_cmd_queue_init(struct thread_data *td)
829{
830 struct ioring_data *ld = td->io_ops_data;
831 struct ioring_options *o = td->eo;
832 int depth = td->o.iodepth;
833 struct io_uring_params p;
834 int ret;
835
836 memset(&p, 0, sizeof(p));
837
838 if (o->hipri)
839 p.flags |= IORING_SETUP_IOPOLL;
840 if (o->sqpoll_thread) {
841 p.flags |= IORING_SETUP_SQPOLL;
842 if (o->sqpoll_set) {
843 p.flags |= IORING_SETUP_SQ_AFF;
844 p.sq_thread_cpu = o->sqpoll_cpu;
845 }
846 }
847 if (o->cmd_type == FIO_URING_CMD_NVME) {
848 p.flags |= IORING_SETUP_SQE128;
849 p.flags |= IORING_SETUP_CQE32;
850 }
851
852 /*
853 * Clamp CQ ring size at our SQ ring size, we don't need more entries
854 * than that.
855 */
856 p.flags |= IORING_SETUP_CQSIZE;
857 p.cq_entries = depth;
858
859retry:
860 ret = syscall(__NR_io_uring_setup, depth, &p);
861 if (ret < 0) {
862 if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) {
863 p.flags &= ~IORING_SETUP_CQSIZE;
864 goto retry;
865 }
866 return ret;
867 }
868
869 ld->ring_fd = ret;
870
871 fio_ioring_probe(td);
872
873 if (o->fixedbufs) {
874 ret = syscall(__NR_io_uring_register, ld->ring_fd,
875 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
876 if (ret < 0)
877 return ret;
878 }
879
880 return fio_ioring_mmap(ld, &p);
881}
882
883static int fio_ioring_register_files(struct thread_data *td)
884{
885 struct ioring_data *ld = td->io_ops_data;
886 struct fio_file *f;
887 unsigned int i;
888 int ret;
889
890 ld->fds = calloc(td->o.nr_files, sizeof(int));
891
892 for_each_file(td, f, i) {
893 ret = generic_open_file(td, f);
894 if (ret)
895 goto err;
896 ld->fds[i] = f->fd;
897 f->engine_pos = i;
898 }
899
900 ret = syscall(__NR_io_uring_register, ld->ring_fd,
901 IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
902 if (ret) {
903err:
904 free(ld->fds);
905 ld->fds = NULL;
906 }
907
908 /*
909 * Pretend the file is closed again, and really close it if we hit
910 * an error.
911 */
912 for_each_file(td, f, i) {
913 if (ret) {
914 int fio_unused ret2;
915 ret2 = generic_close_file(td, f);
916 } else
917 f->fd = -1;
918 }
919
920 return ret;
921}
922
923static int fio_ioring_post_init(struct thread_data *td)
924{
925 struct ioring_data *ld = td->io_ops_data;
926 struct ioring_options *o = td->eo;
927 struct io_u *io_u;
928 int err, i;
929
930 for (i = 0; i < td->o.iodepth; i++) {
931 struct iovec *iov = &ld->iovecs[i];
932
933 io_u = ld->io_u_index[i];
934 iov->iov_base = io_u->buf;
935 iov->iov_len = td_max_bs(td);
936 }
937
938 err = fio_ioring_queue_init(td);
939 if (err) {
940 int init_err = errno;
941
942 if (init_err == ENOSYS)
943 log_err("fio: your kernel doesn't support io_uring\n");
944 td_verror(td, init_err, "io_queue_init");
945 return 1;
946 }
947
948 for (i = 0; i < td->o.iodepth; i++) {
949 struct io_uring_sqe *sqe;
950
951 sqe = &ld->sqes[i];
952 memset(sqe, 0, sizeof(*sqe));
953 }
954
955 if (o->registerfiles) {
956 err = fio_ioring_register_files(td);
957 if (err) {
958 td_verror(td, errno, "ioring_register_files");
959 return 1;
960 }
961 }
962
963 return 0;
964}
965
966static int fio_ioring_cmd_post_init(struct thread_data *td)
967{
968 struct ioring_data *ld = td->io_ops_data;
969 struct ioring_options *o = td->eo;
970 struct io_u *io_u;
971 int err, i;
972
973 for (i = 0; i < td->o.iodepth; i++) {
974 struct iovec *iov = &ld->iovecs[i];
975
976 io_u = ld->io_u_index[i];
977 iov->iov_base = io_u->buf;
978 iov->iov_len = td_max_bs(td);
979 }
980
981 err = fio_ioring_cmd_queue_init(td);
982 if (err) {
983 int init_err = errno;
984
985 td_verror(td, init_err, "io_queue_init");
986 return 1;
987 }
988
989 for (i = 0; i < td->o.iodepth; i++) {
990 struct io_uring_sqe *sqe;
991
992 if (o->cmd_type == FIO_URING_CMD_NVME) {
993 sqe = &ld->sqes[i << 1];
994 memset(sqe, 0, 2 * sizeof(*sqe));
995 } else {
996 sqe = &ld->sqes[i];
997 memset(sqe, 0, sizeof(*sqe));
998 }
999 }
1000
1001 if (o->registerfiles) {
1002 err = fio_ioring_register_files(td);
1003 if (err) {
1004 td_verror(td, errno, "ioring_register_files");
1005 return 1;
1006 }
1007 }
1008
1009 return 0;
1010}
1011
1012static int fio_ioring_init(struct thread_data *td)
1013{
1014 struct ioring_options *o = td->eo;
1015 struct ioring_data *ld;
1016 int ret;
1017
1018 /* sqthread submission requires registered files */
1019 if (o->sqpoll_thread)
1020 o->registerfiles = 1;
1021
1022 if (o->registerfiles && td->o.nr_files != td->o.open_files) {
1023 log_err("fio: io_uring registered files require nr_files to "
1024 "be identical to open_files\n");
1025 return 1;
1026 }
1027
1028 ld = calloc(1, sizeof(*ld));
1029
1030 /* ring depth must be a power-of-2 */
1031 ld->iodepth = td->o.iodepth;
1032 td->o.iodepth = roundup_pow2(td->o.iodepth);
1033
1034 /* io_u index */
1035 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
1036 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
1037
1038 td->io_ops_data = ld;
1039
1040 ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options);
1041 if (ret) {
1042 td_verror(td, EINVAL, "fio_ioring_init");
1043 return 1;
1044 }
1045
1046 return 0;
1047}
1048
1049static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
1050{
1051 struct ioring_data *ld = td->io_ops_data;
1052
1053 ld->io_u_index[io_u->index] = io_u;
1054 return 0;
1055}
1056
1057static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
1058{
1059 struct ioring_data *ld = td->io_ops_data;
1060 struct ioring_options *o = td->eo;
1061
1062 if (!ld || !o->registerfiles)
1063 return generic_open_file(td, f);
1064
1065 f->fd = ld->fds[f->engine_pos];
1066 return 0;
1067}
1068
1069static int fio_ioring_cmd_open_file(struct thread_data *td, struct fio_file *f)
1070{
1071 struct ioring_data *ld = td->io_ops_data;
1072 struct ioring_options *o = td->eo;
1073
1074 if (o->cmd_type == FIO_URING_CMD_NVME) {
1075 struct nvme_data *data = NULL;
1076 unsigned int nsid, lba_size = 0;
1077 unsigned long long nlba = 0;
1078 int ret;
1079
1080 /* Store the namespace-id and lba size. */
1081 data = FILE_ENG_DATA(f);
1082 if (data == NULL) {
1083 ret = fio_nvme_get_info(f, &nsid, &lba_size, &nlba);
1084 if (ret)
1085 return ret;
1086
1087 data = calloc(1, sizeof(struct nvme_data));
1088 data->nsid = nsid;
1089 data->lba_shift = ilog2(lba_size);
1090
1091 FILE_SET_ENG_DATA(f, data);
1092 }
1093 }
1094 if (!ld || !o->registerfiles)
1095 return generic_open_file(td, f);
1096
1097 f->fd = ld->fds[f->engine_pos];
1098 return 0;
1099}
1100
1101static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
1102{
1103 struct ioring_data *ld = td->io_ops_data;
1104 struct ioring_options *o = td->eo;
1105
1106 if (!ld || !o->registerfiles)
1107 return generic_close_file(td, f);
1108
1109 f->fd = -1;
1110 return 0;
1111}
1112
1113static int fio_ioring_cmd_close_file(struct thread_data *td,
1114 struct fio_file *f)
1115{
1116 struct ioring_data *ld = td->io_ops_data;
1117 struct ioring_options *o = td->eo;
1118
1119 if (o->cmd_type == FIO_URING_CMD_NVME) {
1120 struct nvme_data *data = FILE_ENG_DATA(f);
1121
1122 FILE_SET_ENG_DATA(f, NULL);
1123 free(data);
1124 }
1125 if (!ld || !o->registerfiles)
1126 return generic_close_file(td, f);
1127
1128 f->fd = -1;
1129 return 0;
1130}
1131
1132static int fio_ioring_cmd_get_file_size(struct thread_data *td,
1133 struct fio_file *f)
1134{
1135 struct ioring_options *o = td->eo;
1136
1137 if (fio_file_size_known(f))
1138 return 0;
1139
1140 if (o->cmd_type == FIO_URING_CMD_NVME) {
1141 struct nvme_data *data = NULL;
1142 unsigned int nsid, lba_size = 0;
1143 unsigned long long nlba = 0;
1144 int ret;
1145
1146 ret = fio_nvme_get_info(f, &nsid, &lba_size, &nlba);
1147 if (ret)
1148 return ret;
1149
1150 data = calloc(1, sizeof(struct nvme_data));
1151 data->nsid = nsid;
1152 data->lba_shift = ilog2(lba_size);
1153
1154 f->real_file_size = lba_size * nlba;
1155 fio_file_set_size_known(f);
1156
1157 FILE_SET_ENG_DATA(f, data);
1158 return 0;
1159 }
1160 return generic_get_file_size(td, f);
1161}
1162
1163static int fio_ioring_cmd_get_zoned_model(struct thread_data *td,
1164 struct fio_file *f,
1165 enum zbd_zoned_model *model)
1166{
1167 return fio_nvme_get_zoned_model(td, f, model);
1168}
1169
1170static int fio_ioring_cmd_report_zones(struct thread_data *td,
1171 struct fio_file *f, uint64_t offset,
1172 struct zbd_zone *zbdz,
1173 unsigned int nr_zones)
1174{
1175 return fio_nvme_report_zones(td, f, offset, zbdz, nr_zones);
1176}
1177
1178static int fio_ioring_cmd_reset_wp(struct thread_data *td, struct fio_file *f,
1179 uint64_t offset, uint64_t length)
1180{
1181 return fio_nvme_reset_wp(td, f, offset, length);
1182}
1183
1184static int fio_ioring_cmd_get_max_open_zones(struct thread_data *td,
1185 struct fio_file *f,
1186 unsigned int *max_open_zones)
1187{
1188 return fio_nvme_get_max_open_zones(td, f, max_open_zones);
1189}
1190
1191static struct ioengine_ops ioengine_uring = {
1192 .name = "io_uring",
1193 .version = FIO_IOOPS_VERSION,
1194 .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD |
1195 FIO_ASYNCIO_SETS_ISSUE_TIME,
1196 .init = fio_ioring_init,
1197 .post_init = fio_ioring_post_init,
1198 .io_u_init = fio_ioring_io_u_init,
1199 .prep = fio_ioring_prep,
1200 .queue = fio_ioring_queue,
1201 .commit = fio_ioring_commit,
1202 .getevents = fio_ioring_getevents,
1203 .event = fio_ioring_event,
1204 .cleanup = fio_ioring_cleanup,
1205 .open_file = fio_ioring_open_file,
1206 .close_file = fio_ioring_close_file,
1207 .get_file_size = generic_get_file_size,
1208 .options = options,
1209 .option_struct_size = sizeof(struct ioring_options),
1210};
1211
1212static struct ioengine_ops ioengine_uring_cmd = {
1213 .name = "io_uring_cmd",
1214 .version = FIO_IOOPS_VERSION,
1215 .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD |
1216 FIO_MEMALIGN | FIO_RAWIO |
1217 FIO_ASYNCIO_SETS_ISSUE_TIME,
1218 .init = fio_ioring_init,
1219 .post_init = fio_ioring_cmd_post_init,
1220 .io_u_init = fio_ioring_io_u_init,
1221 .prep = fio_ioring_cmd_prep,
1222 .queue = fio_ioring_queue,
1223 .commit = fio_ioring_commit,
1224 .getevents = fio_ioring_getevents,
1225 .event = fio_ioring_cmd_event,
1226 .cleanup = fio_ioring_cleanup,
1227 .open_file = fio_ioring_cmd_open_file,
1228 .close_file = fio_ioring_cmd_close_file,
1229 .get_file_size = fio_ioring_cmd_get_file_size,
1230 .get_zoned_model = fio_ioring_cmd_get_zoned_model,
1231 .report_zones = fio_ioring_cmd_report_zones,
1232 .reset_wp = fio_ioring_cmd_reset_wp,
1233 .get_max_open_zones = fio_ioring_cmd_get_max_open_zones,
1234 .options = options,
1235 .option_struct_size = sizeof(struct ioring_options),
1236};
1237
1238static void fio_init fio_ioring_register(void)
1239{
1240 register_ioengine(&ioengine_uring);
1241 register_ioengine(&ioengine_uring_cmd);
1242}
1243
1244static void fio_exit fio_ioring_unregister(void)
1245{
1246 unregister_ioengine(&ioengine_uring);
1247 unregister_ioengine(&ioengine_uring_cmd);
1248}
1249#endif
fio - Flexible IO Tester