t/io_uring: allow setting fixed files/buffers as arguments
[fio.git] / engines / io_uring.c
CommitLineData
52885fa2 1/*
bffad86f 2 * io_uring engine
52885fa2 3 *
bffad86f 4 * IO engine using the new native Linux aio io_uring interface. See:
a90cd050 5 *
bffad86f 6 * http://git.kernel.dk/cgit/linux-block/log/?h=io_uring
52885fa2
JA
7 *
8 */
9#include <stdlib.h>
10#include <unistd.h>
11#include <errno.h>
52885fa2
JA
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"
b87aa01a 19#include "../lib/fls.h"
6d975f2c 20#include "../lib/roundup.h"
52885fa2 21
bffad86f 22#ifdef ARCH_HAVE_IOURING
52885fa2 23
57fa61f0 24#include "../lib/types.h"
f3e769a4 25#include "../os/linux/io_uring.h"
9a2d78b3 26
bffad86f 27struct io_sq_ring {
e2239016
JA
28 unsigned *head;
29 unsigned *tail;
30 unsigned *ring_mask;
31 unsigned *ring_entries;
32 unsigned *flags;
33 unsigned *array;
52885fa2
JA
34};
35
bffad86f 36struct io_cq_ring {
e2239016
JA
37 unsigned *head;
38 unsigned *tail;
39 unsigned *ring_mask;
40 unsigned *ring_entries;
f0403f94 41 struct io_uring_cqe *cqes;
9a2d78b3
JA
42};
43
bffad86f 44struct ioring_mmap {
9a2d78b3
JA
45 void *ptr;
46 size_t len;
52885fa2
JA
47};
48
bffad86f 49struct ioring_data {
9a2d78b3
JA
50 int ring_fd;
51
52885fa2
JA
52 struct io_u **io_u_index;
53
5ffd5626
JA
54 int *fds;
55
bffad86f 56 struct io_sq_ring sq_ring;
f0403f94 57 struct io_uring_sqe *sqes;
9a2d78b3 58 struct iovec *iovecs;
b87aa01a 59 unsigned sq_ring_mask;
52885fa2 60
bffad86f 61 struct io_cq_ring cq_ring;
b87aa01a 62 unsigned cq_ring_mask;
52885fa2
JA
63
64 int queued;
65 int cq_ring_off;
b87aa01a 66 unsigned iodepth;
1af44196
XW
67 bool ioprio_class_set;
68 bool ioprio_set;
96563db9 69
bffad86f 70 struct ioring_mmap mmap[3];
52885fa2
JA
71};
72
bffad86f 73struct ioring_options {
52885fa2
JA
74 void *pad;
75 unsigned int hipri;
b2a432bf 76 unsigned int cmdprio_percentage;
52885fa2 77 unsigned int fixedbufs;
5ffd5626 78 unsigned int registerfiles;
3d7d00a3 79 unsigned int sqpoll_thread;
2ea53ca3
JA
80 unsigned int sqpoll_set;
81 unsigned int sqpoll_cpu;
b10b1e70 82 unsigned int nonvectored;
4a87b584 83 unsigned int uncached;
7d42e66e 84 unsigned int nowait;
52885fa2
JA
85};
86
b10b1e70
JA
87static const int ddir_to_op[2][2] = {
88 { IORING_OP_READV, IORING_OP_READ },
89 { IORING_OP_WRITEV, IORING_OP_WRITE }
90};
91
3f1e3af7
KB
92static const int fixed_ddir_to_op[2] = {
93 IORING_OP_READ_FIXED,
94 IORING_OP_WRITE_FIXED
95};
96
2ea53ca3 97static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val)
a90cd050 98{
bffad86f 99 struct ioring_options *o = data;
a90cd050 100
2ea53ca3
JA
101 o->sqpoll_cpu = *val;
102 o->sqpoll_set = 1;
a90cd050
JA
103 return 0;
104}
105
52885fa2
JA
106static struct fio_option options[] = {
107 {
108 .name = "hipri",
109 .lname = "High Priority",
110 .type = FIO_OPT_STR_SET,
bffad86f 111 .off1 = offsetof(struct ioring_options, hipri),
52885fa2
JA
112 .help = "Use polled IO completions",
113 .category = FIO_OPT_C_ENGINE,
27f436d9 114 .group = FIO_OPT_G_IOURING,
52885fa2 115 },
b2a432bf
PC
116#ifdef FIO_HAVE_IOPRIO_CLASS
117 {
118 .name = "cmdprio_percentage",
119 .lname = "high priority percentage",
120 .type = FIO_OPT_INT,
121 .off1 = offsetof(struct ioring_options, cmdprio_percentage),
122 .minval = 1,
123 .maxval = 100,
124 .help = "Send high priority I/O this percentage of the time",
125 .category = FIO_OPT_C_ENGINE,
126 .group = FIO_OPT_G_IOURING,
127 },
128#else
129 {
130 .name = "cmdprio_percentage",
131 .lname = "high priority percentage",
132 .type = FIO_OPT_UNSUPPORTED,
133 .help = "Your platform does not support I/O priority classes",
134 },
135#endif
52885fa2
JA
136 {
137 .name = "fixedbufs",
138 .lname = "Fixed (pre-mapped) IO buffers",
139 .type = FIO_OPT_STR_SET,
bffad86f 140 .off1 = offsetof(struct ioring_options, fixedbufs),
52885fa2
JA
141 .help = "Pre map IO buffers",
142 .category = FIO_OPT_C_ENGINE,
27f436d9 143 .group = FIO_OPT_G_IOURING,
52885fa2 144 },
5ffd5626
JA
145 {
146 .name = "registerfiles",
147 .lname = "Register file set",
148 .type = FIO_OPT_STR_SET,
149 .off1 = offsetof(struct ioring_options, registerfiles),
150 .help = "Pre-open/register files",
151 .category = FIO_OPT_C_ENGINE,
27f436d9 152 .group = FIO_OPT_G_IOURING,
5ffd5626 153 },
771c9901
JA
154 {
155 .name = "sqthread_poll",
3d7d00a3
JA
156 .lname = "Kernel SQ thread polling",
157 .type = FIO_OPT_INT,
158 .off1 = offsetof(struct ioring_options, sqpoll_thread),
159 .help = "Offload submission/completion to kernel thread",
160 .category = FIO_OPT_C_ENGINE,
27f436d9 161 .group = FIO_OPT_G_IOURING,
3d7d00a3
JA
162 },
163 {
164 .name = "sqthread_poll_cpu",
165 .lname = "SQ Thread Poll CPU",
2ea53ca3
JA
166 .type = FIO_OPT_INT,
167 .cb = fio_ioring_sqpoll_cb,
3d7d00a3 168 .help = "What CPU to run SQ thread polling on",
a90cd050 169 .category = FIO_OPT_C_ENGINE,
27f436d9 170 .group = FIO_OPT_G_IOURING,
a90cd050 171 },
b10b1e70
JA
172 {
173 .name = "nonvectored",
174 .lname = "Non-vectored",
175 .type = FIO_OPT_INT,
176 .off1 = offsetof(struct ioring_options, nonvectored),
556d8415 177 .def = "-1",
b10b1e70
JA
178 .help = "Use non-vectored read/write commands",
179 .category = FIO_OPT_C_ENGINE,
180 .group = FIO_OPT_G_IOURING,
181 },
4a87b584
JA
182 {
183 .name = "uncached",
184 .lname = "Uncached",
185 .type = FIO_OPT_INT,
186 .off1 = offsetof(struct ioring_options, uncached),
187 .help = "Use RWF_UNCACHED for buffered read/writes",
188 .category = FIO_OPT_C_ENGINE,
189 .group = FIO_OPT_G_IOURING,
190 },
7d42e66e
KK
191 {
192 .name = "nowait",
193 .lname = "RWF_NOWAIT",
194 .type = FIO_OPT_BOOL,
195 .off1 = offsetof(struct ioring_options, nowait),
196 .help = "Use RWF_NOWAIT for reads/writes",
197 .category = FIO_OPT_C_ENGINE,
198 .group = FIO_OPT_G_IOURING,
199 },
52885fa2
JA
200 {
201 .name = NULL,
202 },
203};
204
bffad86f 205static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
52885fa2
JA
206 unsigned int min_complete, unsigned int flags)
207{
bfed648c 208 return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit,
521164fa 209 min_complete, flags, NULL, 0);
52885fa2
JA
210}
211
bffad86f 212static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
52885fa2 213{
bffad86f 214 struct ioring_data *ld = td->io_ops_data;
cfcc8564 215 struct ioring_options *o = td->eo;
52885fa2 216 struct fio_file *f = io_u->file;
f0403f94 217 struct io_uring_sqe *sqe;
52885fa2 218
f0403f94 219 sqe = &ld->sqes[io_u->index];
34d6090e 220
5ffd5626
JA
221 if (o->registerfiles) {
222 sqe->fd = f->engine_pos;
223 sqe->flags = IOSQE_FIXED_FILE;
224 } else {
225 sqe->fd = f->fd;
5ffd5626 226 }
52885fa2 227
e3970057 228 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
f0403f94 229 if (o->fixedbufs) {
3f1e3af7 230 sqe->opcode = fixed_ddir_to_op[io_u->ddir];
919850d2 231 sqe->addr = (unsigned long) io_u->xfer_buf;
f0403f94 232 sqe->len = io_u->xfer_buflen;
2ea53ca3 233 sqe->buf_index = io_u->index;
cfcc8564 234 } else {
832faaaf
JA
235 struct iovec *iov = &ld->iovecs[io_u->index];
236
237 /*
238 * Update based on actual io_u, requeue could have
239 * adjusted these
240 */
241 iov->iov_base = io_u->xfer_buf;
242 iov->iov_len = io_u->xfer_buflen;
243
3f1e3af7 244 sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored];
b10b1e70 245 if (o->nonvectored) {
832faaaf
JA
246 sqe->addr = (unsigned long) iov->iov_base;
247 sqe->len = iov->iov_len;
b10b1e70 248 } else {
832faaaf 249 sqe->addr = (unsigned long) iov;
b10b1e70
JA
250 sqe->len = 1;
251 }
cfcc8564 252 }
4a87b584
JA
253 if (!td->o.odirect && o->uncached)
254 sqe->rw_flags = RWF_UNCACHED;
7d42e66e
KK
255 if (o->nowait)
256 sqe->rw_flags |= RWF_NOWAIT;
1af44196 257 if (ld->ioprio_class_set)
b7ed2a86 258 sqe->ioprio = td->o.ioprio_class << 13;
1af44196 259 if (ld->ioprio_set)
b7ed2a86 260 sqe->ioprio |= td->o.ioprio;
f0403f94 261 sqe->off = io_u->offset;
7c70f506 262 sqe->rw_flags = 0;
48e698fa 263 } else if (ddir_sync(io_u->ddir)) {
7c70f506 264 sqe->ioprio = 0;
01387bfe
AF
265 if (io_u->ddir == DDIR_SYNC_FILE_RANGE) {
266 sqe->off = f->first_write;
267 sqe->len = f->last_write - f->first_write;
268 sqe->sync_range_flags = td->o.sync_file_range;
269 sqe->opcode = IORING_OP_SYNC_FILE_RANGE;
270 } else {
7c70f506
JA
271 sqe->off = 0;
272 sqe->addr = 0;
273 sqe->len = 0;
01387bfe
AF
274 if (io_u->ddir == DDIR_DATASYNC)
275 sqe->fsync_flags |= IORING_FSYNC_DATASYNC;
276 sqe->opcode = IORING_OP_FSYNC;
277 }
48e698fa 278 }
52885fa2 279
48e698fa 280 sqe->user_data = (unsigned long) io_u;
52885fa2
JA
281 return 0;
282}
283
bffad86f 284static struct io_u *fio_ioring_event(struct thread_data *td, int event)
52885fa2 285{
bffad86f 286 struct ioring_data *ld = td->io_ops_data;
f0403f94 287 struct io_uring_cqe *cqe;
52885fa2 288 struct io_u *io_u;
b87aa01a 289 unsigned index;
52885fa2 290
b87aa01a 291 index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
52885fa2 292
f0403f94 293 cqe = &ld->cq_ring.cqes[index];
e3466352 294 io_u = (struct io_u *) (uintptr_t) cqe->user_data;
52885fa2 295
f0403f94
JA
296 if (cqe->res != io_u->xfer_buflen) {
297 if (cqe->res > io_u->xfer_buflen)
298 io_u->error = -cqe->res;
52885fa2 299 else
f0403f94 300 io_u->resid = io_u->xfer_buflen - cqe->res;
52885fa2
JA
301 } else
302 io_u->error = 0;
303
304 return io_u;
305}
306
bffad86f 307static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
52885fa2
JA
308 unsigned int max)
309{
bffad86f
JA
310 struct ioring_data *ld = td->io_ops_data;
311 struct io_cq_ring *ring = &ld->cq_ring;
e2239016 312 unsigned head, reaped = 0;
52885fa2 313
9a2d78b3 314 head = *ring->head;
52885fa2 315 do {
9e26aff9 316 if (head == atomic_load_acquire(ring->tail))
52885fa2
JA
317 break;
318 reaped++;
319 head++;
52885fa2
JA
320 } while (reaped + events < max);
321
76ce63dd
AB
322 if (reaped)
323 atomic_store_release(ring->head, head);
324
52885fa2
JA
325 return reaped;
326}
327
bffad86f
JA
328static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
329 unsigned int max, const struct timespec *t)
52885fa2 330{
bffad86f 331 struct ioring_data *ld = td->io_ops_data;
52885fa2 332 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
bffad86f
JA
333 struct ioring_options *o = td->eo;
334 struct io_cq_ring *ring = &ld->cq_ring;
b87aa01a
JA
335 unsigned events = 0;
336 int r;
52885fa2 337
9a2d78b3 338 ld->cq_ring_off = *ring->head;
52885fa2 339 do {
bffad86f 340 r = fio_ioring_cqring_reap(td, events, max);
52885fa2
JA
341 if (r) {
342 events += r;
f7cbbbf8
ST
343 if (actual_min != 0)
344 actual_min -= r;
52885fa2
JA
345 continue;
346 }
347
3d7d00a3 348 if (!o->sqpoll_thread) {
9a2d78b3
JA
349 r = io_uring_enter(ld, 0, actual_min,
350 IORING_ENTER_GETEVENTS);
771c9901 351 if (r < 0) {
f6abd731 352 if (errno == EAGAIN || errno == EINTR)
771c9901 353 continue;
9a2d78b3 354 td_verror(td, errno, "io_uring_enter");
771c9901
JA
355 break;
356 }
52885fa2
JA
357 }
358 } while (events < min);
359
360 return r < 0 ? r : events;
361}
362
b2a432bf
PC
363static void fio_ioring_prio_prep(struct thread_data *td, struct io_u *io_u)
364{
365 struct ioring_options *o = td->eo;
366 struct ioring_data *ld = td->io_ops_data;
367 if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) {
368 ld->sqes[io_u->index].ioprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT;
369 io_u->flags |= IO_U_F_PRIORITY;
370 }
371 return;
372}
373
bffad86f
JA
374static enum fio_q_status fio_ioring_queue(struct thread_data *td,
375 struct io_u *io_u)
52885fa2 376{
bffad86f
JA
377 struct ioring_data *ld = td->io_ops_data;
378 struct io_sq_ring *ring = &ld->sq_ring;
b2a432bf 379 struct ioring_options *o = td->eo;
52885fa2
JA
380 unsigned tail, next_tail;
381
382 fio_ro_check(td, io_u);
383
b87aa01a 384 if (ld->queued == ld->iodepth)
52885fa2
JA
385 return FIO_Q_BUSY;
386
52885fa2
JA
387 if (io_u->ddir == DDIR_TRIM) {
388 if (ld->queued)
389 return FIO_Q_BUSY;
390
391 do_io_u_trim(td, io_u);
392 io_u_mark_submit(td, 1);
393 io_u_mark_complete(td, 1);
394 return FIO_Q_COMPLETED;
395 }
396
9a2d78b3 397 tail = *ring->tail;
52885fa2 398 next_tail = tail + 1;
9e26aff9 399 if (next_tail == atomic_load_acquire(ring->head))
52885fa2
JA
400 return FIO_Q_BUSY;
401
b2a432bf
PC
402 if (o->cmdprio_percentage)
403 fio_ioring_prio_prep(td, io_u);
b87aa01a 404 ring->array[tail & ld->sq_ring_mask] = io_u->index;
9e26aff9 405 atomic_store_release(ring->tail, next_tail);
52885fa2
JA
406
407 ld->queued++;
408 return FIO_Q_QUEUED;
409}
410
bffad86f 411static void fio_ioring_queued(struct thread_data *td, int start, int nr)
52885fa2 412{
bffad86f 413 struct ioring_data *ld = td->io_ops_data;
52885fa2
JA
414 struct timespec now;
415
416 if (!fio_fill_issue_time(td))
417 return;
418
419 fio_gettime(&now, NULL);
420
421 while (nr--) {
bffad86f 422 struct io_sq_ring *ring = &ld->sq_ring;
9a2d78b3 423 int index = ring->array[start & ld->sq_ring_mask];
f8289afc 424 struct io_u *io_u = ld->io_u_index[index];
52885fa2
JA
425
426 memcpy(&io_u->issue_time, &now, sizeof(now));
427 io_u_queued(td, io_u);
428
429 start++;
52885fa2
JA
430 }
431}
432
bffad86f 433static int fio_ioring_commit(struct thread_data *td)
52885fa2 434{
bffad86f
JA
435 struct ioring_data *ld = td->io_ops_data;
436 struct ioring_options *o = td->eo;
52885fa2
JA
437 int ret;
438
439 if (!ld->queued)
440 return 0;
441
3d7d00a3
JA
442 /*
443 * Kernel side does submission. just need to check if the ring is
444 * flagged as needing a kick, if so, call io_uring_enter(). This
445 * only happens if we've been idle too long.
446 */
447 if (o->sqpoll_thread) {
bffad86f 448 struct io_sq_ring *ring = &ld->sq_ring;
4cdbc048 449
2ea53ca3 450 read_barrier();
9a2d78b3 451 if (*ring->flags & IORING_SQ_NEED_WAKEUP)
b532dd6d
JA
452 io_uring_enter(ld, ld->queued, 0,
453 IORING_ENTER_SQ_WAKEUP);
771c9901
JA
454 ld->queued = 0;
455 return 0;
456 }
457
52885fa2 458 do {
9a2d78b3 459 unsigned start = *ld->sq_ring.head;
52885fa2
JA
460 long nr = ld->queued;
461
9a2d78b3 462 ret = io_uring_enter(ld, nr, 0, IORING_ENTER_GETEVENTS);
52885fa2 463 if (ret > 0) {
bffad86f 464 fio_ioring_queued(td, start, ret);
52885fa2
JA
465 io_u_mark_submit(td, ret);
466
467 ld->queued -= ret;
468 ret = 0;
a90cd050
JA
469 } else if (!ret) {
470 io_u_mark_submit(td, ret);
52885fa2 471 continue;
a90cd050 472 } else {
f6abd731 473 if (errno == EAGAIN || errno == EINTR) {
bffad86f 474 ret = fio_ioring_cqring_reap(td, 0, ld->queued);
a90cd050
JA
475 if (ret)
476 continue;
477 /* Shouldn't happen */
478 usleep(1);
479 continue;
52885fa2 480 }
9a2d78b3 481 td_verror(td, errno, "io_uring_enter submit");
52885fa2 482 break;
a90cd050 483 }
52885fa2
JA
484 } while (ld->queued);
485
486 return ret;
487}
488
bffad86f 489static void fio_ioring_unmap(struct ioring_data *ld)
52885fa2 490{
9a2d78b3 491 int i;
52885fa2 492
9a2d78b3
JA
493 for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
494 munmap(ld->mmap[i].ptr, ld->mmap[i].len);
495 close(ld->ring_fd);
b87aa01a
JA
496}
497
bffad86f 498static void fio_ioring_cleanup(struct thread_data *td)
52885fa2 499{
bffad86f 500 struct ioring_data *ld = td->io_ops_data;
52885fa2
JA
501
502 if (ld) {
52885fa2 503 if (!(td->flags & TD_F_CHILD))
bffad86f 504 fio_ioring_unmap(ld);
9a2d78b3 505
52885fa2 506 free(ld->io_u_index);
9a2d78b3 507 free(ld->iovecs);
5ffd5626 508 free(ld->fds);
52885fa2
JA
509 free(ld);
510 }
511}
512
bffad86f 513static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
9a2d78b3 514{
bffad86f
JA
515 struct io_sq_ring *sring = &ld->sq_ring;
516 struct io_cq_ring *cring = &ld->cq_ring;
9a2d78b3
JA
517 void *ptr;
518
e2239016 519 ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
9a2d78b3
JA
520 ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
521 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
522 IORING_OFF_SQ_RING);
523 ld->mmap[0].ptr = ptr;
524 sring->head = ptr + p->sq_off.head;
525 sring->tail = ptr + p->sq_off.tail;
526 sring->ring_mask = ptr + p->sq_off.ring_mask;
527 sring->ring_entries = ptr + p->sq_off.ring_entries;
528 sring->flags = ptr + p->sq_off.flags;
ac122fea 529 sring->array = ptr + p->sq_off.array;
9a2d78b3
JA
530 ld->sq_ring_mask = *sring->ring_mask;
531
f0403f94
JA
532 ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
533 ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
9a2d78b3 534 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
f0403f94
JA
535 IORING_OFF_SQES);
536 ld->mmap[1].ptr = ld->sqes;
9a2d78b3 537
f0403f94
JA
538 ld->mmap[2].len = p->cq_off.cqes +
539 p->cq_entries * sizeof(struct io_uring_cqe);
9a2d78b3
JA
540 ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
541 MAP_SHARED | MAP_POPULATE, ld->ring_fd,
542 IORING_OFF_CQ_RING);
543 ld->mmap[2].ptr = ptr;
544 cring->head = ptr + p->cq_off.head;
545 cring->tail = ptr + p->cq_off.tail;
546 cring->ring_mask = ptr + p->cq_off.ring_mask;
547 cring->ring_entries = ptr + p->cq_off.ring_entries;
f0403f94 548 cring->cqes = ptr + p->cq_off.cqes;
9a2d78b3
JA
549 ld->cq_ring_mask = *cring->ring_mask;
550 return 0;
551}
552
556d8415
JA
553static void fio_ioring_probe(struct thread_data *td)
554{
555 struct ioring_data *ld = td->io_ops_data;
556 struct ioring_options *o = td->eo;
557 struct io_uring_probe *p;
558 int ret;
559
560 /* already set by user, don't touch */
561 if (o->nonvectored != -1)
562 return;
563
564 /* default to off, as that's always safe */
565 o->nonvectored = 0;
566
567 p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
568 if (!p)
569 return;
570
571 memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
572 ret = syscall(__NR_io_uring_register, ld->ring_fd,
573 IORING_REGISTER_PROBE, p, 256);
574 if (ret < 0)
575 goto out;
576
577 if (IORING_OP_WRITE > p->ops_len)
578 goto out;
579
580 if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) &&
581 (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED))
582 o->nonvectored = 1;
583out:
584 free(p);
585}
586
bffad86f 587static int fio_ioring_queue_init(struct thread_data *td)
52885fa2 588{
bffad86f
JA
589 struct ioring_data *ld = td->io_ops_data;
590 struct ioring_options *o = td->eo;
52885fa2 591 int depth = td->o.iodepth;
bffad86f 592 struct io_uring_params p;
9a2d78b3
JA
593 int ret;
594
595 memset(&p, 0, sizeof(p));
52885fa2
JA
596
597 if (o->hipri)
bffad86f 598 p.flags |= IORING_SETUP_IOPOLL;
3d7d00a3
JA
599 if (o->sqpoll_thread) {
600 p.flags |= IORING_SETUP_SQPOLL;
601 if (o->sqpoll_set) {
602 p.flags |= IORING_SETUP_SQ_AFF;
603 p.sq_thread_cpu = o->sqpoll_cpu;
604 }
f635f1fb 605 }
a90cd050 606
bfed648c 607 ret = syscall(__NR_io_uring_setup, depth, &p);
9a2d78b3
JA
608 if (ret < 0)
609 return ret;
610
611 ld->ring_fd = ret;
2ea53ca3 612
556d8415
JA
613 fio_ioring_probe(td);
614
2ea53ca3 615 if (o->fixedbufs) {
bfed648c 616 ret = syscall(__NR_io_uring_register, ld->ring_fd,
919850d2 617 IORING_REGISTER_BUFFERS, ld->iovecs, depth);
2ea53ca3
JA
618 if (ret < 0)
619 return ret;
620 }
621
bffad86f 622 return fio_ioring_mmap(ld, &p);
52885fa2
JA
623}
624
5ffd5626
JA
625static int fio_ioring_register_files(struct thread_data *td)
626{
627 struct ioring_data *ld = td->io_ops_data;
628 struct fio_file *f;
629 unsigned int i;
630 int ret;
631
632 ld->fds = calloc(td->o.nr_files, sizeof(int));
633
634 for_each_file(td, f, i) {
635 ret = generic_open_file(td, f);
636 if (ret)
637 goto err;
638 ld->fds[i] = f->fd;
639 f->engine_pos = i;
640 }
641
bfed648c 642 ret = syscall(__NR_io_uring_register, ld->ring_fd,
5ffd5626
JA
643 IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
644 if (ret) {
645err:
646 free(ld->fds);
647 ld->fds = NULL;
648 }
649
650 /*
651 * Pretend the file is closed again, and really close it if we hit
652 * an error.
653 */
654 for_each_file(td, f, i) {
655 if (ret) {
656 int fio_unused ret2;
657 ret2 = generic_close_file(td, f);
658 } else
659 f->fd = -1;
660 }
661
662 return ret;
663}
664
bffad86f 665static int fio_ioring_post_init(struct thread_data *td)
52885fa2 666{
bffad86f 667 struct ioring_data *ld = td->io_ops_data;
5ffd5626 668 struct ioring_options *o = td->eo;
52885fa2 669 struct io_u *io_u;
650346e1 670 int err, i;
52885fa2 671
650346e1
JA
672 for (i = 0; i < td->o.iodepth; i++) {
673 struct iovec *iov = &ld->iovecs[i];
9a2d78b3 674
650346e1
JA
675 io_u = ld->io_u_index[i];
676 iov->iov_base = io_u->buf;
677 iov->iov_len = td_max_bs(td);
52885fa2
JA
678 }
679
bffad86f 680 err = fio_ioring_queue_init(td);
52885fa2 681 if (err) {
d63a472d 682 td_verror(td, errno, "io_queue_init");
52885fa2
JA
683 return 1;
684 }
685
7c70f506
JA
686 for (i = 0; i < td->o.iodepth; i++) {
687 struct io_uring_sqe *sqe;
688
689 sqe = &ld->sqes[i];
690 memset(sqe, 0, sizeof(*sqe));
691 }
692
5ffd5626
JA
693 if (o->registerfiles) {
694 err = fio_ioring_register_files(td);
695 if (err) {
696 td_verror(td, errno, "ioring_register_files");
697 return 1;
698 }
699 }
700
52885fa2
JA
701 return 0;
702}
703
bffad86f 704static int fio_ioring_init(struct thread_data *td)
52885fa2 705{
5ffd5626 706 struct ioring_options *o = td->eo;
bffad86f 707 struct ioring_data *ld;
b2a432bf 708 struct thread_options *to = &td->o;
52885fa2 709
5ffd5626
JA
710 /* sqthread submission requires registered files */
711 if (o->sqpoll_thread)
712 o->registerfiles = 1;
713
714 if (o->registerfiles && td->o.nr_files != td->o.open_files) {
715 log_err("fio: io_uring registered files require nr_files to "
716 "be identical to open_files\n");
717 return 1;
718 }
719
52885fa2
JA
720 ld = calloc(1, sizeof(*ld));
721
b87aa01a
JA
722 /* ring depth must be a power-of-2 */
723 ld->iodepth = td->o.iodepth;
724 td->o.iodepth = roundup_pow2(td->o.iodepth);
725
52885fa2
JA
726 /* io_u index */
727 ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
650346e1 728 ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
52885fa2
JA
729
730 td->io_ops_data = ld;
b2a432bf
PC
731
732 /*
733 * Check for option conflicts
734 */
735 if ((fio_option_is_set(to, ioprio) || fio_option_is_set(to, ioprio_class)) &&
736 o->cmdprio_percentage != 0) {
737 log_err("%s: cmdprio_percentage option and mutually exclusive "
738 "prio or prioclass option is set, exiting\n", to->name);
739 td_verror(td, EINVAL, "fio_io_uring_init");
740 return 1;
741 }
1af44196
XW
742
743 if (fio_option_is_set(&td->o, ioprio_class))
744 ld->ioprio_class_set = true;
745 if (fio_option_is_set(&td->o, ioprio))
746 ld->ioprio_set = true;
747
52885fa2
JA
748 return 0;
749}
750
bffad86f 751static int fio_ioring_io_u_init(struct thread_data *td, struct io_u *io_u)
52885fa2 752{
bffad86f 753 struct ioring_data *ld = td->io_ops_data;
52885fa2
JA
754
755 ld->io_u_index[io_u->index] = io_u;
756 return 0;
757}
758
5ffd5626
JA
759static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
760{
761 struct ioring_data *ld = td->io_ops_data;
762 struct ioring_options *o = td->eo;
763
17318cf6 764 if (!ld || !o->registerfiles)
5ffd5626
JA
765 return generic_open_file(td, f);
766
767 f->fd = ld->fds[f->engine_pos];
768 return 0;
769}
770
771static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
772{
17318cf6 773 struct ioring_data *ld = td->io_ops_data;
5ffd5626
JA
774 struct ioring_options *o = td->eo;
775
17318cf6 776 if (!ld || !o->registerfiles)
5ffd5626
JA
777 return generic_close_file(td, f);
778
779 f->fd = -1;
780 return 0;
781}
782
52885fa2 783static struct ioengine_ops ioengine = {
bffad86f 784 .name = "io_uring",
52885fa2 785 .version = FIO_IOOPS_VERSION,
04ba61df 786 .flags = FIO_ASYNCIO_SYNC_TRIM,
bffad86f
JA
787 .init = fio_ioring_init,
788 .post_init = fio_ioring_post_init,
789 .io_u_init = fio_ioring_io_u_init,
790 .prep = fio_ioring_prep,
791 .queue = fio_ioring_queue,
792 .commit = fio_ioring_commit,
793 .getevents = fio_ioring_getevents,
794 .event = fio_ioring_event,
795 .cleanup = fio_ioring_cleanup,
5ffd5626
JA
796 .open_file = fio_ioring_open_file,
797 .close_file = fio_ioring_close_file,
52885fa2
JA
798 .get_file_size = generic_get_file_size,
799 .options = options,
bffad86f 800 .option_struct_size = sizeof(struct ioring_options),
52885fa2
JA
801};
802
bffad86f 803static void fio_init fio_ioring_register(void)
52885fa2 804{
52885fa2 805 register_ioengine(&ioengine);
52885fa2
JA
806}
807
bffad86f 808static void fio_exit fio_ioring_unregister(void)
52885fa2 809{
52885fa2 810 unregister_ioengine(&ioengine);
52885fa2 811}
1f90e9bb 812#endif