engines/libblkio: Share a single blkio instance among threads in same process
[fio.git] / engines / libaio.c
CommitLineData
2866c82d 1/*
da751ca9
JA
2 * libaio engine
3 *
4 * IO engine using the Linux native aio interface.
2866c82d
JA
5 *
6 */
2866c82d
JA
7#include <stdlib.h>
8#include <unistd.h>
9#include <errno.h>
67bf9823 10#include <libaio.h>
46961748
JA
11#include <sys/time.h>
12#include <sys/resource.h>
5f350952
JA
13
14#include "../fio.h"
0f38bbef 15#include "../lib/pow2.h"
d220c761 16#include "../optgroup.h"
90a8c9b2 17#include "../lib/memalign.h"
e9f6567a 18#include "cmdprio.h"
2866c82d 19
b2a432bf
PC
20/* Should be defined in newest aio_abi.h */
21#ifndef IOCB_FLAG_IOPRIO
22#define IOCB_FLAG_IOPRIO (1 << 1)
23#endif
24
7d42e66e
KK
25/* Hack for libaio < 0.3.111 */
26#ifndef CONFIG_LIBAIO_RW_FLAGS
27#define aio_rw_flags __pad2
28#endif
29
0fc2e103 30static int fio_libaio_commit(struct thread_data *td);
b2a432bf 31static int fio_libaio_init(struct thread_data *td);
0fc2e103 32
2866c82d
JA
33struct libaio_data {
34 io_context_t aio_ctx;
35 struct io_event *aio_events;
755200a3 36 struct iocb **iocbs;
7e77dd02 37 struct io_u **io_us;
acd45e02 38
3c3168e9
JA
39 struct io_u **io_u_index;
40
acd45e02
JA
41 /*
42 * Basic ring buffer. 'head' is incremented in _queue(), and
43 * 'tail' is incremented in _commit(). We keep 'queued' so
44 * that we know if the ring is full or empty, when
45 * 'head' == 'tail'. 'entries' is the ring size, and
46 * 'is_pow2' is just an optimization to use AND instead of
47 * modulus to get the remainder on ring increment.
48 */
49 int is_pow2;
50 unsigned int entries;
51 unsigned int queued;
52 unsigned int head;
53 unsigned int tail;
d6cbeab4
NC
54
55 struct cmdprio cmdprio;
2866c82d
JA
56};
57
de890a1e 58struct libaio_options {
a48f0cc7 59 struct thread_data *td;
de890a1e 60 unsigned int userspace_reap;
d6cbeab4 61 struct cmdprio_options cmdprio_options;
7d42e66e 62 unsigned int nowait;
de890a1e
SL
63};
64
65static struct fio_option options[] = {
66 {
67 .name = "userspace_reap",
e8b0e958 68 .lname = "Libaio userspace reaping",
de890a1e
SL
69 .type = FIO_OPT_STR_SET,
70 .off1 = offsetof(struct libaio_options, userspace_reap),
71 .help = "Use alternative user-space reap implementation",
e90a0adf 72 .category = FIO_OPT_C_ENGINE,
5a3cd5f3 73 .group = FIO_OPT_G_LIBAIO,
de890a1e 74 },
b2a432bf
PC
75#ifdef FIO_HAVE_IOPRIO_CLASS
76 {
77 .name = "cmdprio_percentage",
78 .lname = "high priority percentage",
79 .type = FIO_OPT_INT,
e9f6567a 80 .off1 = offsetof(struct libaio_options,
d6cbeab4 81 cmdprio_options.percentage[DDIR_READ]),
e9f6567a 82 .off2 = offsetof(struct libaio_options,
d6cbeab4 83 cmdprio_options.percentage[DDIR_WRITE]),
e9f6567a 84 .minval = 0,
b2a432bf
PC
85 .maxval = 100,
86 .help = "Send high priority I/O this percentage of the time",
87 .category = FIO_OPT_C_ENGINE,
88 .group = FIO_OPT_G_LIBAIO,
89 },
12f9d54a
DLM
90 {
91 .name = "cmdprio_class",
92 .lname = "Asynchronous I/O priority class",
93 .type = FIO_OPT_INT,
94 .off1 = offsetof(struct libaio_options,
d6cbeab4 95 cmdprio_options.class[DDIR_READ]),
12f9d54a 96 .off2 = offsetof(struct libaio_options,
d6cbeab4 97 cmdprio_options.class[DDIR_WRITE]),
12f9d54a
DLM
98 .help = "Set asynchronous IO priority class",
99 .minval = IOPRIO_MIN_PRIO_CLASS + 1,
100 .maxval = IOPRIO_MAX_PRIO_CLASS,
101 .interval = 1,
102 .category = FIO_OPT_C_ENGINE,
103 .group = FIO_OPT_G_LIBAIO,
104 },
105 {
106 .name = "cmdprio",
107 .lname = "Asynchronous I/O priority level",
108 .type = FIO_OPT_INT,
109 .off1 = offsetof(struct libaio_options,
d6cbeab4 110 cmdprio_options.level[DDIR_READ]),
12f9d54a 111 .off2 = offsetof(struct libaio_options,
d6cbeab4 112 cmdprio_options.level[DDIR_WRITE]),
12f9d54a
DLM
113 .help = "Set asynchronous IO priority level",
114 .minval = IOPRIO_MIN_PRIO,
115 .maxval = IOPRIO_MAX_PRIO,
116 .interval = 1,
117 .category = FIO_OPT_C_ENGINE,
118 .group = FIO_OPT_G_LIBAIO,
119 },
a48f0cc7
DLM
120 {
121 .name = "cmdprio_bssplit",
122 .lname = "Priority percentage block size split",
d6cbeab4
NC
123 .type = FIO_OPT_STR_STORE,
124 .off1 = offsetof(struct libaio_options,
125 cmdprio_options.bssplit_str),
a48f0cc7
DLM
126 .help = "Set priority percentages for different block sizes",
127 .category = FIO_OPT_C_ENGINE,
128 .group = FIO_OPT_G_LIBAIO,
129 },
b2a432bf
PC
130#else
131 {
132 .name = "cmdprio_percentage",
133 .lname = "high priority percentage",
134 .type = FIO_OPT_UNSUPPORTED,
135 .help = "Your platform does not support I/O priority classes",
136 },
12f9d54a
DLM
137 {
138 .name = "cmdprio_class",
139 .lname = "Asynchronous I/O priority class",
140 .type = FIO_OPT_UNSUPPORTED,
141 .help = "Your platform does not support I/O priority classes",
142 },
143 {
144 .name = "cmdprio",
145 .lname = "Asynchronous I/O priority level",
146 .type = FIO_OPT_UNSUPPORTED,
147 .help = "Your platform does not support I/O priority classes",
148 },
a48f0cc7
DLM
149 {
150 .name = "cmdprio_bssplit",
151 .lname = "Priority percentage block size split",
152 .type = FIO_OPT_UNSUPPORTED,
153 .help = "Your platform does not support I/O priority classes",
154 },
b2a432bf 155#endif
7d42e66e
KK
156 {
157 .name = "nowait",
158 .lname = "RWF_NOWAIT",
159 .type = FIO_OPT_BOOL,
160 .off1 = offsetof(struct libaio_options, nowait),
161 .help = "Set RWF_NOWAIT for reads/writes",
162 .category = FIO_OPT_C_ENGINE,
163 .group = FIO_OPT_G_LIBAIO,
164 },
de890a1e
SL
165 {
166 .name = NULL,
167 },
168};
169
acd45e02
JA
170static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
171 unsigned int add)
172{
173 if (ld->is_pow2)
174 *val = (*val + add) & (ld->entries - 1);
175 else
176 *val = (*val + add) % ld->entries;
177}
178
7d42e66e 179static int fio_libaio_prep(struct thread_data *td, struct io_u *io_u)
2866c82d 180{
7d42e66e 181 struct libaio_options *o = td->eo;
53cdc686 182 struct fio_file *f = io_u->file;
ad4e9298 183 struct iocb *iocb = &io_u->iocb;
53cdc686 184
a391d73d 185 if (io_u->ddir == DDIR_READ) {
702906e9 186 io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
7d42e66e
KK
187 if (o->nowait)
188 iocb->aio_rw_flags |= RWF_NOWAIT;
a391d73d 189 } else if (io_u->ddir == DDIR_WRITE) {
702906e9 190 io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
7d42e66e
KK
191 if (o->nowait)
192 iocb->aio_rw_flags |= RWF_NOWAIT;
a391d73d 193 } else if (ddir_sync(io_u->ddir))
3c3168e9 194 io_prep_fsync(iocb, f->fd);
2866c82d
JA
195
196 return 0;
197}
198
127715b6
NC
199static inline void fio_libaio_cmdprio_prep(struct thread_data *td,
200 struct io_u *io_u)
b2a432bf 201{
d6cbeab4
NC
202 struct libaio_data *ld = td->io_ops_data;
203 struct cmdprio *cmdprio = &ld->cmdprio;
127715b6
NC
204
205 if (fio_cmdprio_set_ioprio(td, cmdprio, io_u)) {
206 io_u->iocb.aio_reqprio = io_u->ioprio;
b2a432bf 207 io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO;
b2a432bf 208 }
b2a432bf
PC
209}
210
2866c82d
JA
211static struct io_u *fio_libaio_event(struct thread_data *td, int event)
212{
565e784d 213 struct libaio_data *ld = td->io_ops_data;
f423479d
JA
214 struct io_event *ev;
215 struct io_u *io_u;
2866c82d 216
f423479d 217 ev = ld->aio_events + event;
702906e9 218 io_u = container_of(ev->obj, struct io_u, iocb);
f423479d
JA
219
220 if (ev->res != io_u->xfer_buflen) {
221 if (ev->res > io_u->xfer_buflen)
222 io_u->error = -ev->res;
223 else
224 io_u->resid = io_u->xfer_buflen - ev->res;
225 } else
226 io_u->error = 0;
227
228 return io_u;
2866c82d
JA
229}
230
675012f0
DE
231struct aio_ring {
232 unsigned id; /** kernel internal index number */
233 unsigned nr; /** number of io_events */
234 unsigned head;
235 unsigned tail;
c44b1ff5 236
675012f0
DE
237 unsigned magic;
238 unsigned compat_features;
239 unsigned incompat_features;
240 unsigned header_length; /** size of aio_ring */
241
242 struct io_event events[0];
243};
244
245#define AIO_RING_MAGIC 0xa10a10a1
246
247static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
c44b1ff5 248 struct io_event *events)
675012f0
DE
249{
250 long i = 0;
251 unsigned head;
c44b1ff5 252 struct aio_ring *ring = (struct aio_ring*) aio_ctx;
675012f0
DE
253
254 while (i < max) {
255 head = ring->head;
256
257 if (head == ring->tail) {
258 /* There are no more completions */
259 break;
260 } else {
261 /* There is another completion to reap */
262 events[i] = ring->events[head];
d473a06d
BVA
263 atomic_store_release(&ring->head,
264 (head + 1) % ring->nr);
675012f0
DE
265 i++;
266 }
267 }
268
269 return i;
270}
271
e7d2e616 272static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
1f440ece 273 unsigned int max, const struct timespec *t)
2866c82d 274{
565e784d 275 struct libaio_data *ld = td->io_ops_data;
de890a1e 276 struct libaio_options *o = td->eo;
82407585 277 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
1f440ece 278 struct timespec __lt, *lt = NULL;
0b7fdba7 279 int r, events = 0;
2866c82d 280
1f440ece
JA
281 if (t) {
282 __lt = *t;
283 lt = &__lt;
284 }
285
2866c82d 286 do {
de890a1e 287 if (o->userspace_reap == 1
675012f0
DE
288 && actual_min == 0
289 && ((struct aio_ring *)(ld->aio_ctx))->magic
290 == AIO_RING_MAGIC) {
291 r = user_io_getevents(ld->aio_ctx, max,
292 ld->aio_events + events);
293 } else {
294 r = io_getevents(ld->aio_ctx, actual_min,
1f440ece 295 max, ld->aio_events + events, lt);
675012f0 296 }
3441a52d 297 if (r > 0)
0b7fdba7 298 events += r;
3441a52d 299 else if ((min && r == 0) || r == -EAGAIN) {
0fc2e103 300 fio_libaio_commit(td);
6347e43d
JA
301 if (actual_min)
302 usleep(10);
0fc2e103 303 } else if (r != -EINTR)
a31dc2dd 304 break;
0b7fdba7
DE
305 } while (events < min);
306
307 return r < 0 ? r : events;
2866c82d
JA
308}
309
2e4ef4fb
JA
310static enum fio_q_status fio_libaio_queue(struct thread_data *td,
311 struct io_u *io_u)
2866c82d 312{
565e784d 313 struct libaio_data *ld = td->io_ops_data;
2866c82d 314
7101d9c2
JA
315 fio_ro_check(td, io_u);
316
acd45e02 317 if (ld->queued == td->o.iodepth)
755200a3
JA
318 return FIO_Q_BUSY;
319
320 /*
321 * fsync is tricky, since it can fail and we need to do it
322 * serialized with other io. the reason is that linux doesn't
323 * support aio fsync yet. So return busy for the case where we
324 * have pending io, to let fio complete those first.
325 */
f011531e 326 if (ddir_sync(io_u->ddir)) {
acd45e02 327 if (ld->queued)
755200a3 328 return FIO_Q_BUSY;
5f9099ea 329
f011531e 330 do_io_u_sync(td, io_u);
755200a3
JA
331 return FIO_Q_COMPLETED;
332 }
333
a5f3027c 334 if (io_u->ddir == DDIR_TRIM) {
acd45e02 335 if (ld->queued)
a5f3027c
JA
336 return FIO_Q_BUSY;
337
338 do_io_u_trim(td, io_u);
c0681c9d
VF
339 io_u_mark_submit(td, 1);
340 io_u_mark_complete(td, 1);
a5f3027c
JA
341 return FIO_Q_COMPLETED;
342 }
343
d6cbeab4 344 if (ld->cmdprio.mode != CMDPRIO_MODE_NONE)
ff00f247 345 fio_libaio_cmdprio_prep(td, io_u);
b2a432bf 346
702906e9 347 ld->iocbs[ld->head] = &io_u->iocb;
acd45e02
JA
348 ld->io_us[ld->head] = io_u;
349 ring_inc(ld, &ld->head, 1);
350 ld->queued++;
755200a3
JA
351 return FIO_Q_QUEUED;
352}
353
7e77dd02
JA
354static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
355 unsigned int nr)
356{
8b6a404c 357 struct timespec now;
7e77dd02
JA
358 unsigned int i;
359
12d9d841
JA
360 if (!fio_fill_issue_time(td))
361 return;
362
7e77dd02
JA
363 fio_gettime(&now, NULL);
364
365 for (i = 0; i < nr; i++) {
366 struct io_u *io_u = io_us[i];
367
368 memcpy(&io_u->issue_time, &now, sizeof(now));
369 io_u_queued(td, io_u);
370 }
39f56400
VF
371
372 /*
373 * only used for iolog
374 */
375 if (td->o.read_iolog_file)
376 memcpy(&td->last_issue, &now, sizeof(now));
7e77dd02
JA
377}
378
755200a3
JA
379static int fio_libaio_commit(struct thread_data *td)
380{
565e784d 381 struct libaio_data *ld = td->io_ops_data;
755200a3 382 struct iocb **iocbs;
7e77dd02 383 struct io_u **io_us;
8b6a404c 384 struct timespec ts;
a120ca7f 385 int ret, wait_start = 0;
755200a3 386
acd45e02 387 if (!ld->queued)
755200a3
JA
388 return 0;
389
2866c82d 390 do {
acd45e02
JA
391 long nr = ld->queued;
392
393 nr = min((unsigned int) nr, ld->entries - ld->tail);
394 io_us = ld->io_us + ld->tail;
395 iocbs = ld->iocbs + ld->tail;
396
397 ret = io_submit(ld->aio_ctx, nr, iocbs);
5e00c2c4 398 if (ret > 0) {
7e77dd02 399 fio_libaio_queued(td, io_us, ret);
838bc709 400 io_u_mark_submit(td, ret);
acd45e02
JA
401
402 ld->queued -= ret;
403 ring_inc(ld, &ld->tail, ret);
5e00c2c4 404 ret = 0;
e3b4e568 405 wait_start = 0;
acd45e02 406 } else if (ret == -EINTR || !ret) {
838bc709
JA
407 if (!ret)
408 io_u_mark_submit(td, ret);
e3b4e568 409 wait_start = 0;
2866c82d 410 continue;
acd45e02
JA
411 } else if (ret == -EAGAIN) {
412 /*
413 * If we get EAGAIN, we should break out without
414 * error and let the upper layer reap some
a120ca7f
JA
415 * events for us. If we have no queued IO, we
416 * must loop here. If we loop for more than 30s,
417 * just error out, something must be buggy in the
418 * IO path.
acd45e02 419 */
a120ca7f
JA
420 if (ld->queued) {
421 ret = 0;
422 break;
423 }
424 if (!wait_start) {
8b6a404c 425 fio_gettime(&ts, NULL);
d36b072d 426 wait_start = 1;
8b6a404c 427 } else if (mtime_since_now(&ts) > 30000) {
a120ca7f
JA
428 log_err("fio: aio appears to be stalled, giving up\n");
429 break;
430 }
431 usleep(1);
432 continue;
a31dc2dd
JA
433 } else if (ret == -ENOMEM) {
434 /*
435 * If we get -ENOMEM, reap events if we can. If
436 * we cannot, treat it as a fatal event since there's
437 * nothing we can do about it.
438 */
439 if (ld->queued)
440 ret = 0;
441 break;
838bc709 442 } else
2866c82d 443 break;
2c3a4ae9 444 } while (ld->queued);
2866c82d 445
36167d82 446 return ret;
2866c82d
JA
447}
448
449static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
450{
565e784d 451 struct libaio_data *ld = td->io_ops_data;
2866c82d
JA
452
453 return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
454}
455
456static void fio_libaio_cleanup(struct thread_data *td)
457{
565e784d 458 struct libaio_data *ld = td->io_ops_data;
2866c82d
JA
459
460 if (ld) {
f24c2649
JA
461 /*
462 * Work-around to avoid huge RCU stalls at exit time. If we
463 * don't do this here, then it'll be torn down by exit_aio().
464 * But for that case we can parallellize the freeing, thus
465 * speeding it up a lot.
466 */
467 if (!(td->flags & TD_F_CHILD))
468 io_destroy(ld->aio_ctx);
d6cbeab4
NC
469
470 fio_cmdprio_cleanup(&ld->cmdprio);
7e77dd02
JA
471 free(ld->aio_events);
472 free(ld->iocbs);
473 free(ld->io_us);
2866c82d 474 free(ld);
2866c82d
JA
475 }
476}
477
2041bd34
JA
478static int fio_libaio_post_init(struct thread_data *td)
479{
480 struct libaio_data *ld = td->io_ops_data;
ad4e9298 481 int err;
2041bd34 482
ad4e9298 483 err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
2041bd34
JA
484 if (err) {
485 td_verror(td, -err, "io_queue_init");
486 return 1;
487 }
488
489 return 0;
ebec344d
JA
490}
491
2866c82d
JA
492static int fio_libaio_init(struct thread_data *td)
493{
acd45e02 494 struct libaio_data *ld;
b2a432bf 495 struct libaio_options *o = td->eo;
e9f6567a 496 int ret;
2866c82d 497
acd45e02 498 ld = calloc(1, sizeof(*ld));
c1db2dce 499
acd45e02
JA
500 ld->entries = td->o.iodepth;
501 ld->is_pow2 = is_power_of_2(ld->entries);
502 ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
503 ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
504 ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
755200a3 505
565e784d 506 td->io_ops_data = ld;
e9f6567a 507
d6cbeab4 508 ret = fio_cmdprio_init(td, &ld->cmdprio, &o->cmdprio_options);
e9f6567a 509 if (ret) {
b2a432bf
PC
510 td_verror(td, EINVAL, "fio_libaio_init");
511 return 1;
512 }
e9f6567a 513
2866c82d
JA
514 return 0;
515}
516
5a8a6a03 517FIO_STATIC struct ioengine_ops ioengine = {
de890a1e
SL
518 .name = "libaio",
519 .version = FIO_IOOPS_VERSION,
4e7e7898
VF
520 .flags = FIO_ASYNCIO_SYNC_TRIM |
521 FIO_ASYNCIO_SETS_ISSUE_TIME,
de890a1e 522 .init = fio_libaio_init,
2041bd34 523 .post_init = fio_libaio_post_init,
de890a1e
SL
524 .prep = fio_libaio_prep,
525 .queue = fio_libaio_queue,
526 .commit = fio_libaio_commit,
527 .cancel = fio_libaio_cancel,
528 .getevents = fio_libaio_getevents,
529 .event = fio_libaio_event,
530 .cleanup = fio_libaio_cleanup,
531 .open_file = generic_open_file,
532 .close_file = generic_close_file,
533 .get_file_size = generic_get_file_size,
534 .options = options,
535 .option_struct_size = sizeof(struct libaio_options),
2866c82d 536};
34cfcdaf 537
5f350952
JA
538static void fio_init fio_libaio_register(void)
539{
540 register_ioengine(&ioengine);
541}
542
543static void fio_exit fio_libaio_unregister(void)
544{
545 unregister_ioengine(&ioengine);
546}