libaio,io_uring: introduce cmdprio_class and cmdprio options
[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 {
a1f871c7 59 void *pad;
de890a1e 60 unsigned int userspace_reap;
e9f6567a 61 struct cmdprio cmdprio;
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
DLM
80 .off1 = offsetof(struct libaio_options,
81 cmdprio.percentage[DDIR_READ]),
82 .off2 = offsetof(struct libaio_options,
83 cmdprio.percentage[DDIR_WRITE]),
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,
95 cmdprio.class[DDIR_READ]),
96 .off2 = offsetof(struct libaio_options,
97 cmdprio.class[DDIR_WRITE]),
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,
110 cmdprio.level[DDIR_READ]),
111 .off2 = offsetof(struct libaio_options,
112 cmdprio.level[DDIR_WRITE]),
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 },
b2a432bf
PC
120#else
121 {
122 .name = "cmdprio_percentage",
123 .lname = "high priority percentage",
124 .type = FIO_OPT_UNSUPPORTED,
125 .help = "Your platform does not support I/O priority classes",
126 },
12f9d54a
DLM
127 {
128 .name = "cmdprio_class",
129 .lname = "Asynchronous I/O priority class",
130 .type = FIO_OPT_UNSUPPORTED,
131 .help = "Your platform does not support I/O priority classes",
132 },
133 {
134 .name = "cmdprio",
135 .lname = "Asynchronous I/O priority level",
136 .type = FIO_OPT_UNSUPPORTED,
137 .help = "Your platform does not support I/O priority classes",
138 },
b2a432bf 139#endif
7d42e66e
KK
140 {
141 .name = "nowait",
142 .lname = "RWF_NOWAIT",
143 .type = FIO_OPT_BOOL,
144 .off1 = offsetof(struct libaio_options, nowait),
145 .help = "Set RWF_NOWAIT for reads/writes",
146 .category = FIO_OPT_C_ENGINE,
147 .group = FIO_OPT_G_LIBAIO,
148 },
de890a1e
SL
149 {
150 .name = NULL,
151 },
152};
153
acd45e02
JA
154static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
155 unsigned int add)
156{
157 if (ld->is_pow2)
158 *val = (*val + add) & (ld->entries - 1);
159 else
160 *val = (*val + add) % ld->entries;
161}
162
7d42e66e 163static int fio_libaio_prep(struct thread_data *td, struct io_u *io_u)
2866c82d 164{
7d42e66e 165 struct libaio_options *o = td->eo;
53cdc686 166 struct fio_file *f = io_u->file;
ad4e9298 167 struct iocb *iocb = &io_u->iocb;
53cdc686 168
a391d73d 169 if (io_u->ddir == DDIR_READ) {
702906e9 170 io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
7d42e66e
KK
171 if (o->nowait)
172 iocb->aio_rw_flags |= RWF_NOWAIT;
a391d73d 173 } else if (io_u->ddir == DDIR_WRITE) {
702906e9 174 io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
7d42e66e
KK
175 if (o->nowait)
176 iocb->aio_rw_flags |= RWF_NOWAIT;
a391d73d 177 } else if (ddir_sync(io_u->ddir))
3c3168e9 178 io_prep_fsync(iocb, f->fd);
2866c82d
JA
179
180 return 0;
181}
182
b2a432bf
PC
183static void fio_libaio_prio_prep(struct thread_data *td, struct io_u *io_u)
184{
185 struct libaio_options *o = td->eo;
e9f6567a 186 struct cmdprio *cmdprio = &o->cmdprio;
12f9d54a
DLM
187 enum fio_ddir ddir = io_u->ddir;
188 unsigned int p = cmdprio->percentage[ddir];
e9f6567a
DLM
189
190 if (p && rand_between(&td->prio_state, 0, 99) < p) {
12f9d54a
DLM
191 io_u->iocb.aio_reqprio =
192 ioprio_value(cmdprio->class[ddir], cmdprio->level[ddir]);
b2a432bf
PC
193 io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO;
194 io_u->flags |= IO_U_F_PRIORITY;
195 }
b2a432bf
PC
196}
197
2866c82d
JA
198static struct io_u *fio_libaio_event(struct thread_data *td, int event)
199{
565e784d 200 struct libaio_data *ld = td->io_ops_data;
f423479d
JA
201 struct io_event *ev;
202 struct io_u *io_u;
2866c82d 203
f423479d 204 ev = ld->aio_events + event;
702906e9 205 io_u = container_of(ev->obj, struct io_u, iocb);
f423479d
JA
206
207 if (ev->res != io_u->xfer_buflen) {
208 if (ev->res > io_u->xfer_buflen)
209 io_u->error = -ev->res;
210 else
211 io_u->resid = io_u->xfer_buflen - ev->res;
212 } else
213 io_u->error = 0;
214
215 return io_u;
2866c82d
JA
216}
217
675012f0
DE
218struct aio_ring {
219 unsigned id; /** kernel internal index number */
220 unsigned nr; /** number of io_events */
221 unsigned head;
222 unsigned tail;
c44b1ff5 223
675012f0
DE
224 unsigned magic;
225 unsigned compat_features;
226 unsigned incompat_features;
227 unsigned header_length; /** size of aio_ring */
228
229 struct io_event events[0];
230};
231
232#define AIO_RING_MAGIC 0xa10a10a1
233
234static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
c44b1ff5 235 struct io_event *events)
675012f0
DE
236{
237 long i = 0;
238 unsigned head;
c44b1ff5 239 struct aio_ring *ring = (struct aio_ring*) aio_ctx;
675012f0
DE
240
241 while (i < max) {
242 head = ring->head;
243
244 if (head == ring->tail) {
245 /* There are no more completions */
246 break;
247 } else {
248 /* There is another completion to reap */
249 events[i] = ring->events[head];
d473a06d
BVA
250 atomic_store_release(&ring->head,
251 (head + 1) % ring->nr);
675012f0
DE
252 i++;
253 }
254 }
255
256 return i;
257}
258
e7d2e616 259static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
1f440ece 260 unsigned int max, const struct timespec *t)
2866c82d 261{
565e784d 262 struct libaio_data *ld = td->io_ops_data;
de890a1e 263 struct libaio_options *o = td->eo;
82407585 264 unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
1f440ece 265 struct timespec __lt, *lt = NULL;
0b7fdba7 266 int r, events = 0;
2866c82d 267
1f440ece
JA
268 if (t) {
269 __lt = *t;
270 lt = &__lt;
271 }
272
2866c82d 273 do {
de890a1e 274 if (o->userspace_reap == 1
675012f0
DE
275 && actual_min == 0
276 && ((struct aio_ring *)(ld->aio_ctx))->magic
277 == AIO_RING_MAGIC) {
278 r = user_io_getevents(ld->aio_ctx, max,
279 ld->aio_events + events);
280 } else {
281 r = io_getevents(ld->aio_ctx, actual_min,
1f440ece 282 max, ld->aio_events + events, lt);
675012f0 283 }
3441a52d 284 if (r > 0)
0b7fdba7 285 events += r;
3441a52d 286 else if ((min && r == 0) || r == -EAGAIN) {
0fc2e103 287 fio_libaio_commit(td);
6347e43d
JA
288 if (actual_min)
289 usleep(10);
0fc2e103 290 } else if (r != -EINTR)
a31dc2dd 291 break;
0b7fdba7
DE
292 } while (events < min);
293
294 return r < 0 ? r : events;
2866c82d
JA
295}
296
2e4ef4fb
JA
297static enum fio_q_status fio_libaio_queue(struct thread_data *td,
298 struct io_u *io_u)
2866c82d 299{
565e784d 300 struct libaio_data *ld = td->io_ops_data;
2866c82d 301
7101d9c2
JA
302 fio_ro_check(td, io_u);
303
acd45e02 304 if (ld->queued == td->o.iodepth)
755200a3
JA
305 return FIO_Q_BUSY;
306
307 /*
308 * fsync is tricky, since it can fail and we need to do it
309 * serialized with other io. the reason is that linux doesn't
310 * support aio fsync yet. So return busy for the case where we
311 * have pending io, to let fio complete those first.
312 */
f011531e 313 if (ddir_sync(io_u->ddir)) {
acd45e02 314 if (ld->queued)
755200a3 315 return FIO_Q_BUSY;
5f9099ea 316
f011531e 317 do_io_u_sync(td, io_u);
755200a3
JA
318 return FIO_Q_COMPLETED;
319 }
320
a5f3027c 321 if (io_u->ddir == DDIR_TRIM) {
acd45e02 322 if (ld->queued)
a5f3027c
JA
323 return FIO_Q_BUSY;
324
325 do_io_u_trim(td, io_u);
c0681c9d
VF
326 io_u_mark_submit(td, 1);
327 io_u_mark_complete(td, 1);
a5f3027c
JA
328 return FIO_Q_COMPLETED;
329 }
330
e9f6567a 331 if (ld->use_cmdprio)
b2a432bf
PC
332 fio_libaio_prio_prep(td, io_u);
333
702906e9 334 ld->iocbs[ld->head] = &io_u->iocb;
acd45e02
JA
335 ld->io_us[ld->head] = io_u;
336 ring_inc(ld, &ld->head, 1);
337 ld->queued++;
755200a3
JA
338 return FIO_Q_QUEUED;
339}
340
7e77dd02
JA
341static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us,
342 unsigned int nr)
343{
8b6a404c 344 struct timespec now;
7e77dd02
JA
345 unsigned int i;
346
12d9d841
JA
347 if (!fio_fill_issue_time(td))
348 return;
349
7e77dd02
JA
350 fio_gettime(&now, NULL);
351
352 for (i = 0; i < nr; i++) {
353 struct io_u *io_u = io_us[i];
354
355 memcpy(&io_u->issue_time, &now, sizeof(now));
356 io_u_queued(td, io_u);
357 }
358}
359
755200a3
JA
360static int fio_libaio_commit(struct thread_data *td)
361{
565e784d 362 struct libaio_data *ld = td->io_ops_data;
755200a3 363 struct iocb **iocbs;
7e77dd02 364 struct io_u **io_us;
8b6a404c 365 struct timespec ts;
a120ca7f 366 int ret, wait_start = 0;
755200a3 367
acd45e02 368 if (!ld->queued)
755200a3
JA
369 return 0;
370
2866c82d 371 do {
acd45e02
JA
372 long nr = ld->queued;
373
374 nr = min((unsigned int) nr, ld->entries - ld->tail);
375 io_us = ld->io_us + ld->tail;
376 iocbs = ld->iocbs + ld->tail;
377
378 ret = io_submit(ld->aio_ctx, nr, iocbs);
5e00c2c4 379 if (ret > 0) {
7e77dd02 380 fio_libaio_queued(td, io_us, ret);
838bc709 381 io_u_mark_submit(td, ret);
acd45e02
JA
382
383 ld->queued -= ret;
384 ring_inc(ld, &ld->tail, ret);
5e00c2c4 385 ret = 0;
e3b4e568 386 wait_start = 0;
acd45e02 387 } else if (ret == -EINTR || !ret) {
838bc709
JA
388 if (!ret)
389 io_u_mark_submit(td, ret);
e3b4e568 390 wait_start = 0;
2866c82d 391 continue;
acd45e02
JA
392 } else if (ret == -EAGAIN) {
393 /*
394 * If we get EAGAIN, we should break out without
395 * error and let the upper layer reap some
a120ca7f
JA
396 * events for us. If we have no queued IO, we
397 * must loop here. If we loop for more than 30s,
398 * just error out, something must be buggy in the
399 * IO path.
acd45e02 400 */
a120ca7f
JA
401 if (ld->queued) {
402 ret = 0;
403 break;
404 }
405 if (!wait_start) {
8b6a404c 406 fio_gettime(&ts, NULL);
d36b072d 407 wait_start = 1;
8b6a404c 408 } else if (mtime_since_now(&ts) > 30000) {
a120ca7f
JA
409 log_err("fio: aio appears to be stalled, giving up\n");
410 break;
411 }
412 usleep(1);
413 continue;
a31dc2dd
JA
414 } else if (ret == -ENOMEM) {
415 /*
416 * If we get -ENOMEM, reap events if we can. If
417 * we cannot, treat it as a fatal event since there's
418 * nothing we can do about it.
419 */
420 if (ld->queued)
421 ret = 0;
422 break;
838bc709 423 } else
2866c82d 424 break;
2c3a4ae9 425 } while (ld->queued);
2866c82d 426
36167d82 427 return ret;
2866c82d
JA
428}
429
430static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u)
431{
565e784d 432 struct libaio_data *ld = td->io_ops_data;
2866c82d
JA
433
434 return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events);
435}
436
437static void fio_libaio_cleanup(struct thread_data *td)
438{
565e784d 439 struct libaio_data *ld = td->io_ops_data;
2866c82d
JA
440
441 if (ld) {
f24c2649
JA
442 /*
443 * Work-around to avoid huge RCU stalls at exit time. If we
444 * don't do this here, then it'll be torn down by exit_aio().
445 * But for that case we can parallellize the freeing, thus
446 * speeding it up a lot.
447 */
448 if (!(td->flags & TD_F_CHILD))
449 io_destroy(ld->aio_ctx);
7e77dd02
JA
450 free(ld->aio_events);
451 free(ld->iocbs);
452 free(ld->io_us);
2866c82d 453 free(ld);
2866c82d
JA
454 }
455}
456
2041bd34
JA
457static int fio_libaio_post_init(struct thread_data *td)
458{
459 struct libaio_data *ld = td->io_ops_data;
ad4e9298 460 int err;
2041bd34 461
ad4e9298 462 err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
2041bd34
JA
463 if (err) {
464 td_verror(td, -err, "io_queue_init");
465 return 1;
466 }
467
468 return 0;
ebec344d
JA
469}
470
2866c82d
JA
471static int fio_libaio_init(struct thread_data *td)
472{
acd45e02 473 struct libaio_data *ld;
b2a432bf 474 struct libaio_options *o = td->eo;
e9f6567a
DLM
475 struct cmdprio *cmdprio = &o->cmdprio;
476 int ret;
2866c82d 477
acd45e02 478 ld = calloc(1, sizeof(*ld));
c1db2dce 479
acd45e02
JA
480 ld->entries = td->o.iodepth;
481 ld->is_pow2 = is_power_of_2(ld->entries);
482 ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
483 ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
484 ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
755200a3 485
565e784d 486 td->io_ops_data = ld;
e9f6567a
DLM
487
488 ret = fio_cmdprio_init(td, cmdprio, &ld->use_cmdprio);
489 if (ret) {
b2a432bf
PC
490 td_verror(td, EINVAL, "fio_libaio_init");
491 return 1;
492 }
e9f6567a 493
2866c82d
JA
494 return 0;
495}
496
5a8a6a03 497FIO_STATIC struct ioengine_ops ioengine = {
de890a1e
SL
498 .name = "libaio",
499 .version = FIO_IOOPS_VERSION,
04ba61df 500 .flags = FIO_ASYNCIO_SYNC_TRIM,
de890a1e 501 .init = fio_libaio_init,
2041bd34 502 .post_init = fio_libaio_post_init,
de890a1e
SL
503 .prep = fio_libaio_prep,
504 .queue = fio_libaio_queue,
505 .commit = fio_libaio_commit,
506 .cancel = fio_libaio_cancel,
507 .getevents = fio_libaio_getevents,
508 .event = fio_libaio_event,
509 .cleanup = fio_libaio_cleanup,
510 .open_file = generic_open_file,
511 .close_file = generic_close_file,
512 .get_file_size = generic_get_file_size,
513 .options = options,
514 .option_struct_size = sizeof(struct libaio_options),
2866c82d 515};
34cfcdaf 516
5f350952
JA
517static void fio_init fio_libaio_register(void)
518{
519 register_ioengine(&ioengine);
520}
521
522static void fio_exit fio_libaio_unregister(void)
523{
524 unregister_ioengine(&ioengine);
525}