projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
examples: uring-cmd-zoned: expand the reasoning behind QD1
[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 if (o->fixedbufs) {
437 sqe->uring_cmd_flags = IORING_URING_CMD_FIXED;
438 sqe->buf_index = io_u->index;
439 }
440
441 cmd = (struct nvme_uring_cmd *)sqe->cmd;
442 return fio_nvme_uring_cmd_prep(cmd, io_u,
443 o->nonvectored ? NULL : &ld->iovecs[io_u->index]);
444}
445
446static struct io_u *fio_ioring_event(struct thread_data *td, int event)
447{
448 struct ioring_data *ld = td->io_ops_data;
449 struct io_uring_cqe *cqe;
450 struct io_u *io_u;
451 unsigned index;
452
453 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
454
455 cqe = &ld->cq_ring.cqes[index];
456 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
457
458 if (cqe->res != io_u->xfer_buflen) {
459 if (cqe->res > io_u->xfer_buflen)
460 io_u->error = -cqe->res;
461 else
462 io_u->resid = io_u->xfer_buflen - cqe->res;
463 } else
464 io_u->error = 0;
465
466 return io_u;
467}
468
469static struct io_u *fio_ioring_cmd_event(struct thread_data *td, int event)
470{
471 struct ioring_data *ld = td->io_ops_data;
472 struct ioring_options *o = td->eo;
473 struct io_uring_cqe *cqe;
474 struct io_u *io_u;
475 unsigned index;
476
477 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
478 if (o->cmd_type == FIO_URING_CMD_NVME)
479 index <<= 1;
480
481 cqe = &ld->cq_ring.cqes[index];
482 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
483
484 if (cqe->res != 0)
485 io_u->error = -cqe->res;
486 else
487 io_u->error = 0;
488
489 return io_u;
490}
491
492static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
493 unsigned int max)
494{
495 struct ioring_data *ld = td->io_ops_data;
496 struct io_cq_ring *ring = &ld->cq_ring;
497 unsigned head, reaped = 0;
498
499 head = *ring->head;
500 do {
501 if (head == atomic_load_acquire(ring->tail))
502 break;
503 reaped++;
504 head++;
505 } while (reaped + events < max);
506
507 if (reaped)
508 atomic_store_release(ring->head, head);
509
510 return reaped;
511}
512
513static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
514 unsigned int max, const struct timespec *t)
515{
516 struct ioring_data *ld = td->io_ops_data;
517 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
518 struct ioring_options *o = td->eo;
519 struct io_cq_ring *ring = &ld->cq_ring;
520 unsigned events = 0;
521 int r;
522
523 ld->cq_ring_off = *ring->head;
524 do {
525 r = fio_ioring_cqring_reap(td, events, max);
526 if (r) {
527 events += r;
528 if (actual_min != 0)
529 actual_min -= r;
530 continue;
531 }
532
533 if (!o->sqpoll_thread) {
534 r = io_uring_enter(ld, 0, actual_min,
535 IORING_ENTER_GETEVENTS);
536 if (r < 0) {
537 if (errno == EAGAIN || errno == EINTR)
538 continue;
539 r = -errno;
540 td_verror(td, errno, "io_uring_enter");
541 break;
542 }
543 }
544 } while (events < min);
545
546 return r < 0 ? r : events;
547}
548
549static inline void fio_ioring_cmdprio_prep(struct thread_data *td,
550 struct io_u *io_u)
551{
552 struct ioring_data *ld = td->io_ops_data;
553 struct cmdprio *cmdprio = &ld->cmdprio;
554
555 if (fio_cmdprio_set_ioprio(td, cmdprio, io_u))
556 ld->sqes[io_u->index].ioprio = io_u->ioprio;
557}
558
559static enum fio_q_status fio_ioring_queue(struct thread_data *td,
560 struct io_u *io_u)
561{
562 struct ioring_data *ld = td->io_ops_data;
563 struct io_sq_ring *ring = &ld->sq_ring;
564 unsigned tail, next_tail;
565
566 fio_ro_check(td, io_u);
567
568 if (ld->queued == ld->iodepth)
569 return FIO_Q_BUSY;
570
571 if (io_u->ddir == DDIR_TRIM) {
572 if (ld->queued)
573 return FIO_Q_BUSY;
574
575 do_io_u_trim(td, io_u);
576 io_u_mark_submit(td, 1);
577 io_u_mark_complete(td, 1);
578 return FIO_Q_COMPLETED;
579 }
580
581 tail = *ring->tail;
582 next_tail = tail + 1;
583 if (next_tail == atomic_load_acquire(ring->head))
584 return FIO_Q_BUSY;
585
586 if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
587 fio_ioring_cmdprio_prep(td, io_u);
588
589 ring->array[tail & ld->sq_ring_mask] = io_u->index;
590 atomic_store_release(ring->tail, next_tail);
591
592 ld->queued++;
593 return FIO_Q_QUEUED;
594}
595
596static void fio_ioring_queued(struct thread_data *td, int start, int nr)
597{
598 struct ioring_data *ld = td->io_ops_data;
599 struct timespec now;
600
601 if (!fio_fill_issue_time(td))
602 return;
603
604 fio_gettime(&now, NULL);
605
606 while (nr--) {
607 struct io_sq_ring *ring = &ld->sq_ring;
608 int index = ring->array[start & ld->sq_ring_mask];
609 struct io_u *io_u = ld->io_u_index[index];
610
611 memcpy(&io_u->issue_time, &now, sizeof(now));
612 io_u_queued(td, io_u);
613
614 start++;
615 }
616
617 /*
618 * only used for iolog
619 */
620 if (td->o.read_iolog_file)
621 memcpy(&td->last_issue, &now, sizeof(now));
622}
623
624static int fio_ioring_commit(struct thread_data *td)
625{
626 struct ioring_data *ld = td->io_ops_data;
627 struct ioring_options *o = td->eo;
628 int ret;
629
630 if (!ld->queued)
631 return 0;
632
633 /*
634 * Kernel side does submission. just need to check if the ring is
635 * flagged as needing a kick, if so, call io_uring_enter(). This
636 * only happens if we've been idle too long.
637 */
638 if (o->sqpoll_thread) {
639 struct io_sq_ring *ring = &ld->sq_ring;
640 unsigned flags;
641
642 flags = atomic_load_acquire(ring->flags);
643 if (flags & IORING_SQ_NEED_WAKEUP)
644 io_uring_enter(ld, ld->queued, 0,
645 IORING_ENTER_SQ_WAKEUP);
646 ld->queued = 0;
647 return 0;
648 }
649
650 do {
651 unsigned start = *ld->sq_ring.head;
652 long nr = ld->queued;
653
654 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
655 if (ret > 0) {
656 fio_ioring_queued(td, start, ret);
657 io_u_mark_submit(td, ret);
658
659 ld->queued -= ret;
660 ret = 0;
661 } else if (!ret) {
662 io_u_mark_submit(td, ret);
663 continue;
664 } else {
665 if (errno == EAGAIN || errno == EINTR) {
666 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
667 if (ret)
668 continue;
669 /* Shouldn't happen */
670 usleep(1);
671 continue;
672 }
673 ret = -errno;
674 td_verror(td, errno, "io_uring_enter submit");
675 break;
676 }
677 } while (ld->queued);
678
679 return ret;
680}
681
682static void fio_ioring_unmap(struct ioring_data *ld)
683{
684 int i;
685
686 for (i = 0; i < FIO_ARRAY_SIZE(ld->mmap); i++)
687 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
688 close(ld->ring_fd);
689}
690
691static void fio_ioring_cleanup(struct thread_data *td)
692{
693 struct ioring_data *ld = td->io_ops_data;
694
695 if (ld) {
696 if (!(td->flags & TD_F_CHILD))
697 fio_ioring_unmap(ld);
698
699 fio_cmdprio_cleanup(&ld->cmdprio);
700 free(ld->io_u_index);
701 free(ld->iovecs);
702 free(ld->fds);
703 free(ld);
704 }
705}
706
707static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
708{
709 struct io_sq_ring *sring = &ld->sq_ring;
710 struct io_cq_ring *cring = &ld->cq_ring;
711 void *ptr;
712
713 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
714 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
715 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
716 IORING_OFF_SQ_RING);
717 ld->mmap[0].ptr = ptr;
718 sring->head = ptr + p->sq_off.head;
719 sring->tail = ptr + p->sq_off.tail;
720 sring->ring_mask = ptr + p->sq_off.ring_mask;
721 sring->ring_entries = ptr + p->sq_off.ring_entries;
722 sring->flags = ptr + p->sq_off.flags;
723 sring->array = ptr + p->sq_off.array;
724 ld->sq_ring_mask = *sring->ring_mask;
725
726 if (p->flags & IORING_SETUP_SQE128)
727 ld->mmap[1].len = 2 * p->sq_entries * sizeof(struct io_uring_sqe);
728 else
729 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
730 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
731 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
732 IORING_OFF_SQES);
733 ld->mmap[1].ptr = ld->sqes;
734
735 if (p->flags & IORING_SETUP_CQE32) {
736 ld->mmap[2].len = p->cq_off.cqes +
737 2 * p->cq_entries * sizeof(struct io_uring_cqe);
738 } else {
739 ld->mmap[2].len = p->cq_off.cqes +
740 p->cq_entries * sizeof(struct io_uring_cqe);
741 }
742 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
743 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
744 IORING_OFF_CQ_RING);
745 ld->mmap[2].ptr = ptr;
746 cring->head = ptr + p->cq_off.head;
747 cring->tail = ptr + p->cq_off.tail;
748 cring->ring_mask = ptr + p->cq_off.ring_mask;
749 cring->ring_entries = ptr + p->cq_off.ring_entries;
750 cring->cqes = ptr + p->cq_off.cqes;
751 ld->cq_ring_mask = *cring->ring_mask;
752 return 0;
753}
754
755static void fio_ioring_probe(struct thread_data *td)
756{
757 struct ioring_data *ld = td->io_ops_data;
758 struct ioring_options *o = td->eo;
759 struct io_uring_probe *p;
760 int ret;
761
762 /* already set by user, don't touch */
763 if (o->nonvectored != -1)
764 return;
765
766 /* default to off, as that's always safe */
767 o->nonvectored = 0;
768
769 p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
770 if (!p)
771 return;
772
773 memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
774 ret = syscall(__NR_io_uring_register, ld->ring_fd,
775 IORING_REGISTER_PROBE, p, 256);
776 if (ret < 0)
777 goto out;
778
779 if (IORING_OP_WRITE > p->ops_len)
780 goto out;
781
782 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) &&
783 (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED))
784 o->nonvectored = 1;
785out:
786 free(p);
787}
788
789static int fio_ioring_queue_init(struct thread_data *td)
790{
791 struct ioring_data *ld = td->io_ops_data;
792 struct ioring_options *o = td->eo;
793 int depth = td->o.iodepth;
794 struct io_uring_params p;
795 int ret;
796
797 memset(&p, 0, sizeof(p));
798
799 if (o->hipri)
800 p.flags |= IORING_SETUP_IOPOLL;
801 if (o->sqpoll_thread) {
802 p.flags |= IORING_SETUP_SQPOLL;
803 if (o->sqpoll_set) {
804 p.flags |= IORING_SETUP_SQ_AFF;
805 p.sq_thread_cpu = o->sqpoll_cpu;
806 }
807 }
808
809 /*
810 * Clamp CQ ring size at our SQ ring size, we don't need more entries
811 * than that.
812 */
813 p.flags |= IORING_SETUP_CQSIZE;
814 p.cq_entries = depth;
815
816 /*
817 * Setup COOP_TASKRUN as we don't need to get IPI interrupted for
818 * completing IO operations.
819 */
820 p.flags |= IORING_SETUP_COOP_TASKRUN;
821
822 /*
823 * io_uring is always a single issuer, and we can defer task_work
824 * runs until we reap events.
825 */
826 p.flags |= IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN;
827
828retry:
829 ret = syscall(__NR_io_uring_setup, depth, &p);
830 if (ret < 0) {
831 if (errno == EINVAL && p.flags & IORING_SETUP_DEFER_TASKRUN) {
832 p.flags &= ~IORING_SETUP_DEFER_TASKRUN;
833 p.flags &= ~IORING_SETUP_SINGLE_ISSUER;
834 goto retry;
835 }
836 if (errno == EINVAL && p.flags & IORING_SETUP_COOP_TASKRUN) {
837 p.flags &= ~IORING_SETUP_COOP_TASKRUN;
838 goto retry;
839 }
840 if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) {
841 p.flags &= ~IORING_SETUP_CQSIZE;
842 goto retry;
843 }
844 return ret;
845 }
846
847 ld->ring_fd = ret;
848
849 fio_ioring_probe(td);
850
851 if (o->fixedbufs) {
852 ret = syscall(__NR_io_uring_register, ld->ring_fd,
853 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
854 if (ret < 0)
855 return ret;
856 }
857
858 return fio_ioring_mmap(ld, &p);
859}
860
861static int fio_ioring_cmd_queue_init(struct thread_data *td)
862{
863 struct ioring_data *ld = td->io_ops_data;
864 struct ioring_options *o = td->eo;
865 int depth = td->o.iodepth;
866 struct io_uring_params p;
867 int ret;
868
869 memset(&p, 0, sizeof(p));
870
871 if (o->hipri)
872 p.flags |= IORING_SETUP_IOPOLL;
873 if (o->sqpoll_thread) {
874 p.flags |= IORING_SETUP_SQPOLL;
875 if (o->sqpoll_set) {
876 p.flags |= IORING_SETUP_SQ_AFF;
877 p.sq_thread_cpu = o->sqpoll_cpu;
878 }
879 }
880 if (o->cmd_type == FIO_URING_CMD_NVME) {
881 p.flags |= IORING_SETUP_SQE128;
882 p.flags |= IORING_SETUP_CQE32;
883 }
884
885 /*
886 * Clamp CQ ring size at our SQ ring size, we don't need more entries
887 * than that.
888 */
889 p.flags |= IORING_SETUP_CQSIZE;
890 p.cq_entries = depth;
891
892retry:
893 ret = syscall(__NR_io_uring_setup, depth, &p);
894 if (ret < 0) {
895 if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) {
896 p.flags &= ~IORING_SETUP_CQSIZE;
897 goto retry;
898 }
899 return ret;
900 }
901
902 ld->ring_fd = ret;
903
904 fio_ioring_probe(td);
905
906 if (o->fixedbufs) {
907 ret = syscall(__NR_io_uring_register, ld->ring_fd,
908 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
909 if (ret < 0)
910 return ret;
911 }
912
913 return fio_ioring_mmap(ld, &p);
914}
915
916static int fio_ioring_register_files(struct thread_data *td)
917{
918 struct ioring_data *ld = td->io_ops_data;
919 struct fio_file *f;
920 unsigned int i;
921 int ret;
922
923 ld->fds = calloc(td->o.nr_files, sizeof(int));
924
925 for_each_file(td, f, i) {
926 ret = generic_open_file(td, f);
927 if (ret)
928 goto err;
929 ld->fds[i] = f->fd;
930 f->engine_pos = i;
931 }
932
933 ret = syscall(__NR_io_uring_register, ld->ring_fd,
934 IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
935 if (ret) {
936err:
937 free(ld->fds);
938 ld->fds = NULL;
939 }
940
941 /*
942 * Pretend the file is closed again, and really close it if we hit
943 * an error.
944 */
945 for_each_file(td, f, i) {
946 if (ret) {
947 int fio_unused ret2;
948 ret2 = generic_close_file(td, f);
949 } else
950 f->fd = -1;
951 }
952
953 return ret;
954}
955
956static int fio_ioring_post_init(struct thread_data *td)
957{
958 struct ioring_data *ld = td->io_ops_data;
959 struct ioring_options *o = td->eo;
960 struct io_u *io_u;
961 int err, i;
962
963 for (i = 0; i < td->o.iodepth; i++) {
964 struct iovec *iov = &ld->iovecs[i];
965
966 io_u = ld->io_u_index[i];
967 iov->iov_base = io_u->buf;
968 iov->iov_len = td_max_bs(td);
969 }
970
971 err = fio_ioring_queue_init(td);
972 if (err) {
973 int init_err = errno;
974
975 if (init_err == ENOSYS)
976 log_err("fio: your kernel doesn't support io_uring\n");
977 td_verror(td, init_err, "io_queue_init");
978 return 1;
979 }
980
981 for (i = 0; i < td->o.iodepth; i++) {
982 struct io_uring_sqe *sqe;
983
984 sqe = &ld->sqes[i];
985 memset(sqe, 0, sizeof(*sqe));
986 }
987
988 if (o->registerfiles) {
989 err = fio_ioring_register_files(td);
990 if (err) {
991 td_verror(td, errno, "ioring_register_files");
992 return 1;
993 }
994 }
995
996 return 0;
997}
998
999static int fio_ioring_cmd_post_init(struct thread_data *td)
1000{
1001 struct ioring_data *ld = td->io_ops_data;
1002 struct ioring_options *o = td->eo;
1003 struct io_u *io_u;
1004 int err, i;
1005
1006 for (i = 0; i < td->o.iodepth; i++) {
1007 struct iovec *iov = &ld->iovecs[i];
1008
1009 io_u = ld->io_u_index[i];
1010 iov->iov_base = io_u->buf;
1011 iov->iov_len = td_max_bs(td);
1012 }
1013
1014 err = fio_ioring_cmd_queue_init(td);
1015 if (err) {
1016 int init_err = errno;
1017
1018 td_verror(td, init_err, "io_queue_init");
1019 return 1;
1020 }
1021
1022 for (i = 0; i < td->o.iodepth; i++) {
1023 struct io_uring_sqe *sqe;
1024
1025 if (o->cmd_type == FIO_URING_CMD_NVME) {
1026 sqe = &ld->sqes[i << 1];
1027 memset(sqe, 0, 2 * sizeof(*sqe));
1028 } else {
1029 sqe = &ld->sqes[i];
1030 memset(sqe, 0, sizeof(*sqe));
1031 }
1032 }
1033
1034 if (o->registerfiles) {
1035 err = fio_ioring_register_files(td);
1036 if (err) {
1037 td_verror(td, errno, "ioring_register_files");
1038 return 1;
1039 }
1040 }
1041
1042 return 0;
1043}
1044
1045static int fio_ioring_init(struct thread_data *td)
1046{
1047 struct ioring_options *o = td->eo;
1048 struct ioring_data *ld;
1049 int ret;
1050
1051 /* sqthread submission requires registered files */
1052 if (o->sqpoll_thread)
1053 o->registerfiles = 1;
1054
1055 if (o->registerfiles && td->o.nr_files != td->o.open_files) {
1056 log_err("fio: io_uring registered files require nr_files to "
1057 "be identical to open_files\n");
1058 return 1;
1059 }
1060
1061 ld = calloc(1, sizeof(*ld));
1062
1063 /* ring depth must be a power-of-2 */
1064 ld->iodepth = td->o.iodepth;
1065 td->o.iodepth = roundup_pow2(td->o.iodepth);
1066
1067 /* io_u index */
1068 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
1069 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
1070
1071 td->io_ops_data = ld;
1072
1073 ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options);
1074 if (ret) {
1075 td_verror(td, EINVAL, "fio_ioring_init");
1076 return 1;
1077 }
1078
1079 return 0;
1080}
1081
1082static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
1083{
1084 struct ioring_data *ld = td->io_ops_data;
1085
1086 ld->io_u_index[io_u->index] = io_u;
1087 return 0;
1088}
1089
1090static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
1091{
1092 struct ioring_data *ld = td->io_ops_data;
1093 struct ioring_options *o = td->eo;
1094
1095 if (!ld || !o->registerfiles)
1096 return generic_open_file(td, f);
1097
1098 f->fd = ld->fds[f->engine_pos];
1099 return 0;
1100}
1101
1102static int fio_ioring_cmd_open_file(struct thread_data *td, struct fio_file *f)
1103{
1104 struct ioring_data *ld = td->io_ops_data;
1105 struct ioring_options *o = td->eo;
1106
1107 if (o->cmd_type == FIO_URING_CMD_NVME) {
1108 struct nvme_data *data = NULL;
1109 unsigned int nsid, lba_size = 0;
1110 unsigned long long nlba = 0;
1111 int ret;
1112
1113 /* Store the namespace-id and lba size. */
1114 data = FILE_ENG_DATA(f);
1115 if (data == NULL) {
1116 ret = fio_nvme_get_info(f, &nsid, &lba_size, &nlba);
1117 if (ret)
1118 return ret;
1119
1120 data = calloc(1, sizeof(struct nvme_data));
1121 data->nsid = nsid;
1122 data->lba_shift = ilog2(lba_size);
1123
1124 FILE_SET_ENG_DATA(f, data);
1125 }
1126 }
1127 if (!ld || !o->registerfiles)
1128 return generic_open_file(td, f);
1129
1130 f->fd = ld->fds[f->engine_pos];
1131 return 0;
1132}
1133
1134static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
1135{
1136 struct ioring_data *ld = td->io_ops_data;
1137 struct ioring_options *o = td->eo;
1138
1139 if (!ld || !o->registerfiles)
1140 return generic_close_file(td, f);
1141
1142 f->fd = -1;
1143 return 0;
1144}
1145
1146static int fio_ioring_cmd_close_file(struct thread_data *td,
1147 struct fio_file *f)
1148{
1149 struct ioring_data *ld = td->io_ops_data;
1150 struct ioring_options *o = td->eo;
1151
1152 if (o->cmd_type == FIO_URING_CMD_NVME) {
1153 struct nvme_data *data = FILE_ENG_DATA(f);
1154
1155 FILE_SET_ENG_DATA(f, NULL);
1156 free(data);
1157 }
1158 if (!ld || !o->registerfiles)
1159 return generic_close_file(td, f);
1160
1161 f->fd = -1;
1162 return 0;
1163}
1164
1165static int fio_ioring_cmd_get_file_size(struct thread_data *td,
1166 struct fio_file *f)
1167{
1168 struct ioring_options *o = td->eo;
1169
1170 if (fio_file_size_known(f))
1171 return 0;
1172
1173 if (o->cmd_type == FIO_URING_CMD_NVME) {
1174 struct nvme_data *data = NULL;
1175 unsigned int nsid, lba_size = 0;
1176 unsigned long long nlba = 0;
1177 int ret;
1178
1179 ret = fio_nvme_get_info(f, &nsid, &lba_size, &nlba);
1180 if (ret)
1181 return ret;
1182
1183 data = calloc(1, sizeof(struct nvme_data));
1184 data->nsid = nsid;
1185 data->lba_shift = ilog2(lba_size);
1186
1187 f->real_file_size = lba_size * nlba;
1188 fio_file_set_size_known(f);
1189
1190 FILE_SET_ENG_DATA(f, data);
1191 return 0;
1192 }
1193 return generic_get_file_size(td, f);
1194}
1195
1196static int fio_ioring_cmd_get_zoned_model(struct thread_data *td,
1197 struct fio_file *f,
1198 enum zbd_zoned_model *model)
1199{
1200 return fio_nvme_get_zoned_model(td, f, model);
1201}
1202
1203static int fio_ioring_cmd_report_zones(struct thread_data *td,
1204 struct fio_file *f, uint64_t offset,
1205 struct zbd_zone *zbdz,
1206 unsigned int nr_zones)
1207{
1208 return fio_nvme_report_zones(td, f, offset, zbdz, nr_zones);
1209}
1210
1211static int fio_ioring_cmd_reset_wp(struct thread_data *td, struct fio_file *f,
1212 uint64_t offset, uint64_t length)
1213{
1214 return fio_nvme_reset_wp(td, f, offset, length);
1215}
1216
1217static int fio_ioring_cmd_get_max_open_zones(struct thread_data *td,
1218 struct fio_file *f,
1219 unsigned int *max_open_zones)
1220{
1221 return fio_nvme_get_max_open_zones(td, f, max_open_zones);
1222}
1223
1224static struct ioengine_ops ioengine_uring = {
1225 .name = "io_uring",
1226 .version = FIO_IOOPS_VERSION,
1227 .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD |
1228 FIO_ASYNCIO_SETS_ISSUE_TIME,
1229 .init = fio_ioring_init,
1230 .post_init = fio_ioring_post_init,
1231 .io_u_init = fio_ioring_io_u_init,
1232 .prep = fio_ioring_prep,
1233 .queue = fio_ioring_queue,
1234 .commit = fio_ioring_commit,
1235 .getevents = fio_ioring_getevents,
1236 .event = fio_ioring_event,
1237 .cleanup = fio_ioring_cleanup,
1238 .open_file = fio_ioring_open_file,
1239 .close_file = fio_ioring_close_file,
1240 .get_file_size = generic_get_file_size,
1241 .options = options,
1242 .option_struct_size = sizeof(struct ioring_options),
1243};
1244
1245static struct ioengine_ops ioengine_uring_cmd = {
1246 .name = "io_uring_cmd",
1247 .version = FIO_IOOPS_VERSION,
1248 .flags = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD |
1249 FIO_MEMALIGN | FIO_RAWIO |
1250 FIO_ASYNCIO_SETS_ISSUE_TIME,
1251 .init = fio_ioring_init,
1252 .post_init = fio_ioring_cmd_post_init,
1253 .io_u_init = fio_ioring_io_u_init,
1254 .prep = fio_ioring_cmd_prep,
1255 .queue = fio_ioring_queue,
1256 .commit = fio_ioring_commit,
1257 .getevents = fio_ioring_getevents,
1258 .event = fio_ioring_cmd_event,
1259 .cleanup = fio_ioring_cleanup,
1260 .open_file = fio_ioring_cmd_open_file,
1261 .close_file = fio_ioring_cmd_close_file,
1262 .get_file_size = fio_ioring_cmd_get_file_size,
1263 .get_zoned_model = fio_ioring_cmd_get_zoned_model,
1264 .report_zones = fio_ioring_cmd_report_zones,
1265 .reset_wp = fio_ioring_cmd_reset_wp,
1266 .get_max_open_zones = fio_ioring_cmd_get_max_open_zones,
1267 .options = options,
1268 .option_struct_size = sizeof(struct ioring_options),
1269};
1270
1271static void fio_init fio_ioring_register(void)
1272{
1273 register_ioengine(&ioengine_uring);
1274 register_ioengine(&ioengine_uring_cmd);
1275}
1276
1277static void fio_exit fio_ioring_unregister(void)
1278{
1279 unregister_ioengine(&ioengine_uring);
1280 unregister_ioengine(&ioengine_uring_cmd);
1281}
1282#endif
fio - Flexible IO Tester