Merge branch 'master' of https://github.com/bvanassche/fio
[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;
e9f6567a
DLM
54
55 bool use_cmdprio;
2866c82d
JA
56};
57
de890a1e 58struct libaio_options {
a48f0cc7 59 struct thread_data *td;
de890a1e 60 unsigned int userspace_reap;
e9f6567a 61 struct cmdprio cmdprio;
7d42e66e 62 unsigned int nowait;
de890a1e
SL
63};
64
a48f0cc7
DLM
65static int str_cmdprio_bssplit_cb(void *data, const char *input)
66{
67 struct libaio_options *o = data;
68 struct thread_data *td = o->td;
69 struct cmdprio *cmdprio = &o->cmdprio;
70
71 return fio_cmdprio_bssplit_parse(td, input, cmdprio);
72}
73
de890a1e
SL
74static struct fio_option options[] = {
75 {
76 .name = "userspace_reap",
e8b0e958 77 .lname = "Libaio userspace reaping",
de890a1e
SL
78 .type = FIO_OPT_STR_SET,
79 .off1 = offsetof(struct libaio_options, userspace_reap),
80 .help = "Use alternative user-space reap implementation",
e90a0adf 81 .category = FIO_OPT_C_ENGINE,
5a3cd5f3 82 .group = FIO_OPT_G_LIBAIO,
de890a1e 83 },
b2a432bf
PC
84#ifdef FIO_HAVE_IOPRIO_CLASS
85 {
86 .name = "cmdprio_percentage",
87 .lname = "high priority percentage",
88 .type = FIO_OPT_INT,
e9f6567a
DLM
89 .off1 = offsetof(struct libaio_options,
90 cmdprio.percentage[DDIR_READ]),
91 .off2 = offsetof(struct libaio_options,
92 cmdprio.percentage[DDIR_WRITE]),
93 .minval = 0,
b2a432bf
PC
94 .maxval = 100,
95 .help = "Send high priority I/O this percentage of the time",
96 .category = FIO_OPT_C_ENGINE,
97 .group = FIO_OPT_G_LIBAIO,
98 },
12f9d54a
DLM
99 {
100 .name = "cmdprio_class",
101 .lname = "Asynchronous I/O priority class",
102 .type = FIO_OPT_INT,
103 .off1 = offsetof(struct libaio_options,
104 cmdprio.class[DDIR_READ]),
105 .off2 = offsetof(struct libaio_options,
106 cmdprio.class[DDIR_WRITE]),
107 .help = "Set asynchronous IO priority class",
108 .minval = IOPRIO_MIN_PRIO_CLASS + 1,
109 .maxval = IOPRIO_MAX_PRIO_CLASS,
110 .interval = 1,
111 .category = FIO_OPT_C_ENGINE,
112 .group = FIO_OPT_G_LIBAIO,
113 },
114 {
115 .name = "cmdprio",
116 .lname = "Asynchronous I/O priority level",
117 .type = FIO_OPT_INT,
118 .off1 = offsetof(struct libaio_options,
119 cmdprio.level[DDIR_READ]),
120 .off2 = offsetof(struct libaio_options,
121 cmdprio.level[DDIR_WRITE]),
122 .help = "Set asynchronous IO priority level",
123 .minval = IOPRIO_MIN_PRIO,
124 .maxval = IOPRIO_MAX_PRIO,
125 .interval = 1,
126 .category = FIO_OPT_C_ENGINE,
127 .group = FIO_OPT_G_LIBAIO,
128 },
a48f0cc7
DLM
129 {
130 .name = "cmdprio_bssplit",
131 .lname = "Priority percentage block size split",
132 .type = FIO_OPT_STR_ULL,
133 .cb = str_cmdprio_bssplit_cb,
134 .off1 = offsetof(struct libaio_options, cmdprio.bssplit),
135 .help = "Set priority percentages for different block sizes",
136 .category = FIO_OPT_C_ENGINE,
137 .group = FIO_OPT_G_LIBAIO,
138 },
b2a432bf
PC
139#else
140 {
141 .name = "cmdprio_percentage",
142 .lname = "high priority percentage",
143 .type = FIO_OPT_UNSUPPORTED,
144 .help = "Your platform does not support I/O priority classes",
145 },
12f9d54a
DLM
146 {
147 .name = "cmdprio_class",
148 .lname = "Asynchronous I/O priority class",
149 .type = FIO_OPT_UNSUPPORTED,
150 .help = "Your platform does not support I/O priority classes",
151 },
152 {
153 .name = "cmdprio",
154 .lname = "Asynchronous I/O priority level",
155 .type = FIO_OPT_UNSUPPORTED,
156 .help = "Your platform does not support I/O priority classes",
157 },
a48f0cc7
DLM
158 {
159 .name = "cmdprio_bssplit",
160 .lname = "Priority percentage block size split",
161 .type = FIO_OPT_UNSUPPORTED,
162 .help = "Your platform does not support I/O priority classes",
163 },
b2a432bf 164#endif
7d42e66e
KK
165 {
166 .name = "nowait",
167 .lname = "RWF_NOWAIT",
168 .type = FIO_OPT_BOOL,
169 .off1 = offsetof(struct libaio_options, nowait),
170 .help = "Set RWF_NOWAIT for reads/writes",
171 .category = FIO_OPT_C_ENGINE,
172 .group = FIO_OPT_G_LIBAIO,
173 },
de890a1e
SL
174 {
175 .name = NULL,
176 },
177};
178
acd45e02
JA
179static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
180 unsigned int add)
181{
182 if (ld->is_pow2)
183 *val = (*val + add) & (ld->entries - 1);
184 else
185 *val = (*val + add) % ld->entries;
186}
187
7d42e66e 188static int fio_libaio_prep(struct thread_data *td, struct io_u *io_u)
2866c82d 189{
7d42e66e 190 struct libaio_options *o = td->eo;
53cdc686 191 struct fio_file *f = io_u->file;
ad4e9298 192 struct iocb *iocb = &io_u->iocb;
53cdc686 193
a391d73d 194 if (io_u->ddir == DDIR_READ) {
702906e9 195 io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
7d42e66e
KK
196 if (o->nowait)
197 iocb->aio_rw_flags |= RWF_NOWAIT;
a391d73d 198 } else if (io_u->ddir == DDIR_WRITE) {
702906e9 199 io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
7d42e66e
KK
200 if (o->nowait)
201 iocb->aio_rw_flags |= RWF_NOWAIT;
a391d73d 202 } else if (ddir_sync(io_u->ddir))
3c3168e9 203 io_prep_fsync(iocb, f->fd);
2866c82d
JA
204
205 return 0;
206}
207
b2a432bf
PC
208static void fio_libaio_prio_prep(struct thread_data *td, struct io_u *io_u)
209{
210 struct libaio_options *o = td->eo;
e9f6567a 211 struct cmdprio *cmdprio = &o->cmdprio;
12f9d54a 212 enum fio_ddir ddir = io_u->ddir;
a48f0cc7 213 unsigned int p = fio_cmdprio_percentage(cmdprio, io_u);
1437d635
DLM
214 unsigned int cmdprio_value =
215 ioprio_value(cmdprio->class[ddir], cmdprio->level[ddir]);
e9f6567a
DLM
216
217 if (p && rand_between(&td->prio_state, 0, 99) < p) {
03ec570f 218 io_u->ioprio = cmdprio_value;
1437d635 219 io_u->iocb.aio_reqprio = cmdprio_value;
b2a432bf 220 io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO;
1437d635
DLM
221 if (!td->ioprio || cmdprio_value < td->ioprio) {
222 /*
223 * The async IO priority is higher (has a lower value)
224 * than the default context priority.
225 */
03ec570f 226 io_u->flags |= IO_U_F_HIGH_PRIO;
1437d635
DLM
227 }
228 } else if (td->ioprio && td->ioprio < cmdprio_value) {
229 /*
230 * The IO will be executed with the default context priority,
231 * and this priority is higher (has a lower value) than the
232 * async IO priority.
233 */
03ec570f 234 io_u->flags |= IO_U_F_HIGH_PRIO;
b2a432bf 235 }
b2a432bf
PC
236}
237
2866c82d
JA
238static struct io_u *fio_libaio_event(struct thread_data *td, int event)
239{
565e784d 240 struct libaio_data *ld = td->io_ops_data;
f423479d
JA
241 struct io_event *ev;
242 struct io_u *io_u;
2866c82d 243
f423479d 244 ev = ld->aio_events + event;
702906e9 245 io_u = container_of(ev->obj, struct io_u, iocb);
f423479d
JA
246
247 if (ev->res != io_u->xfer_buflen) {
248 if (ev->res > io_u->xfer_buflen)
249 io_u->error = -ev->res;
250 else
251 io_u->resid = io_u->xfer_buflen - ev->res;
252 } else
253 io_u->error = 0;
254
255 return io_u;
2866c82d
JA
256}
257
675012f0
DE
258struct aio_ring {
259 unsigned id; /** kernel internal index number */
260 unsigned nr; /** number of io_events */
261 unsigned head;
262 unsigned tail;
c44b1ff5 263
675012f0
DE
264 unsigned magic;
265 unsigned compat_features;
266 unsigned incompat_features;
267 unsigned header_length; /** size of aio_ring */
268
269 struct io_event events[0];
270};
271
272#define AIO_RING_MAGIC 0xa10a10a1
273
274static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
c44b1ff5 275 struct io_event *events)
675012f0
DE
276{
277 long i = 0;
278 unsigned head;
c44b1ff5 279 struct aio_ring *ring = (struct aio_ring*) aio_ctx;
675012f0
DE
280
281 while (i < max) {
282 head = ring->head;
283
284 if (head == ring->tail) {
285 /* There are no more completions */
286 break;
287 } else {
288 /* There is another completion to reap */
289 events[i] = ring->events[head];
d473a06d
BVA
290 atomic_store_release(&ring->head,
291 (head + 1) % ring->nr);
675012f0
DE
292 i++;
293 }
294 }
295
296 return i;
297}
298
e7d2e616 299static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
1f440ece 300 unsigned int max, const struct timespec *t)
2866c82d 301{
565e784d 302 struct libaio_data *ld = td->io_ops_data;
de890a1e 303 struct libaio_options *o = td->eo;
82407585 304 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
1f440ece 305 struct timespec __lt, *lt = NULL;
0b7fdba7 306 int r, events = 0;
2866c82d 307
1f440ece
JA
308 if (t) {
309 __lt = *t;
310 lt = &__lt;
311 }
312
2866c82d 313 do {
de890a1e 314 if (o->userspace_reap == 1
675012f0
DE
315 && actual_min == 0
316 && ((struct aio_ring *)(ld->aio_ctx))->magic
317 == AIO_RING_MAGIC) {
318 r = user_io_getevents(ld->aio_ctx, max,
319 ld->aio_events + events);
320 } else {
321 r = io_getevents(ld->aio_ctx, actual_min,
1f440ece 322 max, ld->aio_events + events, lt);
675012f0 323 }
3441a52d 324 if (r > 0)
0b7fdba7 325 events += r;
3441a52d 326 else if ((min && r == 0) || r == -EAGAIN) {
0fc2e103 327 fio_libaio_commit(td);
6347e43d
JA
328 if (actual_min)
329 usleep(10);
0fc2e103 330 } else if (r != -EINTR)
a31dc2dd 331 break;
0b7fdba7
DE
332 } while (events < min);
333
334 return r < 0 ? r : events;
2866c82d
JA
335}
336
2e4ef4fb
JA
337static enum fio_q_status fio_libaio_queue(struct thread_data *td,
338 struct io_u *io_u)
2866c82d 339{
565e784d 340 struct libaio_data *ld = td->io_ops_data;
2866c82d 341
7101d9c2
JA
342 fio_ro_check(td, io_u);
343
acd45e02 344 if (ld->queued == td->o.iodepth)
755200a3
JA
345 return FIO_Q_BUSY;
346
347 /*
348 * fsync is tricky, since it can fail and we need to do it
349 * serialized with other io. the reason is that linux doesn't
350 * support aio fsync yet. So return busy for the case where we
351 * have pending io, to let fio complete those first.
352 */
f011531e 353 if (ddir_sync(io_u->ddir)) {
acd45e02 354 if (ld->queued)
755200a3 355 return FIO_Q_BUSY;
5f9099ea 356
f011531e 357 do_io_u_sync(td, io_u);
755200a3
JA
358 return FIO_Q_COMPLETED;
359 }
360
a5f3027c 361 if (io_u->ddir == DDIR_TRIM) {
acd45e02 362 if (ld->queued)
a5f3027c
JA
363 return FIO_Q_BUSY;
364
365 do_io_u_trim(td, io_u);
c0681c9d
VF
366 io_u_mark_submit(td, 1);
367 io_u_mark_complete(td, 1);
a5f3027c
JA
368 return FIO_Q_COMPLETED;
369 }
370
e9f6567a 371 if (ld->use_cmdprio)
b2a432bf
PC
372 fio_libaio_prio_prep(td, io_u);
373
702906e9 374 ld->iocbs[ld->head] = &io_u->iocb;
acd45e02
JA
375 ld->io_us[ld->head] = io_u;
376 ring_inc(ld, &ld->head, 1);
377 ld->queued++;
755200a3
JA
378 return FIO_Q_QUEUED;
379}
380
7e77dd02
JA
381static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
382 unsigned int nr)
383{
8b6a404c 384 struct timespec now;
7e77dd02
JA
385 unsigned int i;
386
12d9d841
JA
387 if (!fio_fill_issue_time(td))
388 return;
389
7e77dd02
JA
390 fio_gettime(&now, NULL);
391
392 for (i = 0; i < nr; i++) {
393 struct io_u *io_u = io_us[i];
394
395 memcpy(&io_u->issue_time, &now, sizeof(now));
396 io_u_queued(td, io_u);
397 }
398}
399
755200a3
JA
400static int fio_libaio_commit(struct thread_data *td)
401{
565e784d 402 struct libaio_data *ld = td->io_ops_data;
755200a3 403 struct iocb **iocbs;
7e77dd02 404 struct io_u **io_us;
8b6a404c 405 struct timespec ts;
a120ca7f 406 int ret, wait_start = 0;
755200a3 407
acd45e02 408 if (!ld->queued)
755200a3
JA
409 return 0;
410
2866c82d 411 do {
acd45e02
JA
412 long nr = ld->queued;
413
414 nr = min((unsigned int) nr, ld->entries - ld->tail);
415 io_us = ld->io_us + ld->tail;
416 iocbs = ld->iocbs + ld->tail;
417
418 ret = io_submit(ld->aio_ctx, nr, iocbs);
5e00c2c4 419 if (ret > 0) {
7e77dd02 420 fio_libaio_queued(td, io_us, ret);
838bc709 421 io_u_mark_submit(td, ret);
acd45e02
JA
422
423 ld->queued -= ret;
424 ring_inc(ld, &ld->tail, ret);
5e00c2c4 425 ret = 0;
e3b4e568 426 wait_start = 0;
acd45e02 427 } else if (ret == -EINTR || !ret) {
838bc709
JA
428 if (!ret)
429 io_u_mark_submit(td, ret);
e3b4e568 430 wait_start = 0;
2866c82d 431 continue;
acd45e02
JA
432 } else if (ret == -EAGAIN) {
433 /*
434 * If we get EAGAIN, we should break out without
435 * error and let the upper layer reap some
a120ca7f
JA
436 * events for us. If we have no queued IO, we
437 * must loop here. If we loop for more than 30s,
438 * just error out, something must be buggy in the
439 * IO path.
acd45e02 440 */
a120ca7f
JA
441 if (ld->queued) {
442 ret = 0;
443 break;
444 }
445 if (!wait_start) {
8b6a404c 446 fio_gettime(&ts, NULL);
d36b072d 447 wait_start = 1;
8b6a404c 448 } else if (mtime_since_now(&ts) > 30000) {
a120ca7f
JA
449 log_err("fio: aio appears to be stalled, giving up\n");
450 break;
451 }
452 usleep(1);
453 continue;
a31dc2dd
JA
454 } else if (ret == -ENOMEM) {
455 /*
456 * If we get -ENOMEM, reap events if we can. If
457 * we cannot, treat it as a fatal event since there's
458 * nothing we can do about it.
459 */
460 if (ld->queued)
461 ret = 0;
462 break;
838bc709 463 } else
2866c82d 464 break;
2c3a4ae9 465 } while (ld->queued);
2866c82d 466
36167d82 467 return ret;
2866c82d
JA
468}
469
470static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
471{
565e784d 472 struct libaio_data *ld = td->io_ops_data;
2866c82d
JA
473
474 return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
475}
476
477static void fio_libaio_cleanup(struct thread_data *td)
478{
565e784d 479 struct libaio_data *ld = td->io_ops_data;
2866c82d
JA
480
481 if (ld) {
f24c2649
JA
482 /*
483 * Work-around to avoid huge RCU stalls at exit time. If we
484 * don't do this here, then it'll be torn down by exit_aio().
485 * But for that case we can parallellize the freeing, thus
486 * speeding it up a lot.
487 */
488 if (!(td->flags & TD_F_CHILD))
489 io_destroy(ld->aio_ctx);
7e77dd02
JA
490 free(ld->aio_events);
491 free(ld->iocbs);
492 free(ld->io_us);
2866c82d 493 free(ld);
2866c82d
JA
494 }
495}
496
2041bd34
JA
497static int fio_libaio_post_init(struct thread_data *td)
498{
499 struct libaio_data *ld = td->io_ops_data;
ad4e9298 500 int err;
2041bd34 501
ad4e9298 502 err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
2041bd34
JA
503 if (err) {
504 td_verror(td, -err, "io_queue_init");
505 return 1;
506 }
507
508 return 0;
ebec344d
JA
509}
510
2866c82d
JA
511static int fio_libaio_init(struct thread_data *td)
512{
acd45e02 513 struct libaio_data *ld;
b2a432bf 514 struct libaio_options *o = td->eo;
e9f6567a
DLM
515 struct cmdprio *cmdprio = &o->cmdprio;
516 int ret;
2866c82d 517
acd45e02 518 ld = calloc(1, sizeof(*ld));
c1db2dce 519
acd45e02
JA
520 ld->entries = td->o.iodepth;
521 ld->is_pow2 = is_power_of_2(ld->entries);
522 ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
523 ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
524 ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
755200a3 525
565e784d 526 td->io_ops_data = ld;
e9f6567a
DLM
527
528 ret = fio_cmdprio_init(td, cmdprio, &ld->use_cmdprio);
529 if (ret) {
b2a432bf
PC
530 td_verror(td, EINVAL, "fio_libaio_init");
531 return 1;
532 }
e9f6567a 533
2866c82d
JA
534 return 0;
535}
536
5a8a6a03 537FIO_STATIC struct ioengine_ops ioengine = {
de890a1e
SL
538 .name = "libaio",
539 .version = FIO_IOOPS_VERSION,
04ba61df 540 .flags = FIO_ASYNCIO_SYNC_TRIM,
de890a1e 541 .init = fio_libaio_init,
2041bd34 542 .post_init = fio_libaio_post_init,
de890a1e
SL
543 .prep = fio_libaio_prep,
544 .queue = fio_libaio_queue,
545 .commit = fio_libaio_commit,
546 .cancel = fio_libaio_cancel,
547 .getevents = fio_libaio_getevents,
548 .event = fio_libaio_event,
549 .cleanup = fio_libaio_cleanup,
550 .open_file = generic_open_file,
551 .close_file = generic_close_file,
552 .get_file_size = generic_get_file_size,
553 .options = options,
554 .option_struct_size = sizeof(struct libaio_options),
2866c82d 555};
34cfcdaf 556
5f350952
JA
557static void fio_init fio_libaio_register(void)
558{
559 register_ioengine(&ioengine);
560}
561
562static void fio_exit fio_libaio_unregister(void)
563{
564 unregister_ioengine(&ioengine);
565}