Commit | Line | Data |
---|---|---|
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" |
2866c82d | 18 | |
b2a432bf PC |
19 | /* Should be defined in newest aio_abi.h */ |
20 | #ifndef IOCB_FLAG_IOPRIO | |
21 | #define IOCB_FLAG_IOPRIO (1 << 1) | |
22 | #endif | |
23 | ||
7d42e66e KK |
24 | /* Hack for libaio < 0.3.111 */ |
25 | #ifndef CONFIG_LIBAIO_RW_FLAGS | |
26 | #define aio_rw_flags __pad2 | |
27 | #endif | |
28 | ||
0fc2e103 | 29 | static int fio_libaio_commit(struct thread_data *td); |
b2a432bf | 30 | static int fio_libaio_init(struct thread_data *td); |
0fc2e103 | 31 | |
2866c82d JA |
32 | struct libaio_data { |
33 | io_context_t aio_ctx; | |
34 | struct io_event *aio_events; | |
755200a3 | 35 | struct iocb **iocbs; |
7e77dd02 | 36 | struct io_u **io_us; |
acd45e02 | 37 | |
3c3168e9 JA |
38 | struct io_u **io_u_index; |
39 | ||
acd45e02 JA |
40 | /* |
41 | * Basic ring buffer. 'head' is incremented in _queue(), and | |
42 | * 'tail' is incremented in _commit(). We keep 'queued' so | |
43 | * that we know if the ring is full or empty, when | |
44 | * 'head' == 'tail'. 'entries' is the ring size, and | |
45 | * 'is_pow2' is just an optimization to use AND instead of | |
46 | * modulus to get the remainder on ring increment. | |
47 | */ | |
48 | int is_pow2; | |
49 | unsigned int entries; | |
50 | unsigned int queued; | |
51 | unsigned int head; | |
52 | unsigned int tail; | |
2866c82d JA |
53 | }; |
54 | ||
de890a1e | 55 | struct libaio_options { |
a1f871c7 | 56 | void *pad; |
de890a1e | 57 | unsigned int userspace_reap; |
b2a432bf | 58 | unsigned int cmdprio_percentage; |
7d42e66e | 59 | unsigned int nowait; |
de890a1e SL |
60 | }; |
61 | ||
62 | static struct fio_option options[] = { | |
63 | { | |
64 | .name = "userspace_reap", | |
e8b0e958 | 65 | .lname = "Libaio userspace reaping", |
de890a1e SL |
66 | .type = FIO_OPT_STR_SET, |
67 | .off1 = offsetof(struct libaio_options, userspace_reap), | |
68 | .help = "Use alternative user-space reap implementation", | |
e90a0adf | 69 | .category = FIO_OPT_C_ENGINE, |
5a3cd5f3 | 70 | .group = FIO_OPT_G_LIBAIO, |
de890a1e | 71 | }, |
b2a432bf PC |
72 | #ifdef FIO_HAVE_IOPRIO_CLASS |
73 | { | |
74 | .name = "cmdprio_percentage", | |
75 | .lname = "high priority percentage", | |
76 | .type = FIO_OPT_INT, | |
77 | .off1 = offsetof(struct libaio_options, cmdprio_percentage), | |
78 | .minval = 1, | |
79 | .maxval = 100, | |
80 | .help = "Send high priority I/O this percentage of the time", | |
81 | .category = FIO_OPT_C_ENGINE, | |
82 | .group = FIO_OPT_G_LIBAIO, | |
83 | }, | |
84 | #else | |
85 | { | |
86 | .name = "cmdprio_percentage", | |
87 | .lname = "high priority percentage", | |
88 | .type = FIO_OPT_UNSUPPORTED, | |
89 | .help = "Your platform does not support I/O priority classes", | |
90 | }, | |
91 | #endif | |
7d42e66e KK |
92 | { |
93 | .name = "nowait", | |
94 | .lname = "RWF_NOWAIT", | |
95 | .type = FIO_OPT_BOOL, | |
96 | .off1 = offsetof(struct libaio_options, nowait), | |
97 | .help = "Set RWF_NOWAIT for reads/writes", | |
98 | .category = FIO_OPT_C_ENGINE, | |
99 | .group = FIO_OPT_G_LIBAIO, | |
100 | }, | |
de890a1e SL |
101 | { |
102 | .name = NULL, | |
103 | }, | |
104 | }; | |
105 | ||
acd45e02 JA |
106 | static inline void ring_inc(struct libaio_data *ld, unsigned int *val, |
107 | unsigned int add) | |
108 | { | |
109 | if (ld->is_pow2) | |
110 | *val = (*val + add) & (ld->entries - 1); | |
111 | else | |
112 | *val = (*val + add) % ld->entries; | |
113 | } | |
114 | ||
7d42e66e | 115 | static int fio_libaio_prep(struct thread_data *td, struct io_u *io_u) |
2866c82d | 116 | { |
7d42e66e | 117 | struct libaio_options *o = td->eo; |
53cdc686 | 118 | struct fio_file *f = io_u->file; |
ad4e9298 | 119 | struct iocb *iocb = &io_u->iocb; |
53cdc686 | 120 | |
a391d73d | 121 | if (io_u->ddir == DDIR_READ) { |
702906e9 | 122 | io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
7d42e66e KK |
123 | if (o->nowait) |
124 | iocb->aio_rw_flags |= RWF_NOWAIT; | |
a391d73d | 125 | } else if (io_u->ddir == DDIR_WRITE) { |
702906e9 | 126 | io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
7d42e66e KK |
127 | if (o->nowait) |
128 | iocb->aio_rw_flags |= RWF_NOWAIT; | |
a391d73d | 129 | } else if (ddir_sync(io_u->ddir)) |
3c3168e9 | 130 | io_prep_fsync(iocb, f->fd); |
2866c82d JA |
131 | |
132 | return 0; | |
133 | } | |
134 | ||
b2a432bf PC |
135 | static void fio_libaio_prio_prep(struct thread_data *td, struct io_u *io_u) |
136 | { | |
137 | struct libaio_options *o = td->eo; | |
138 | if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) { | |
139 | io_u->iocb.aio_reqprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT; | |
140 | io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO; | |
141 | io_u->flags |= IO_U_F_PRIORITY; | |
142 | } | |
143 | return; | |
144 | } | |
145 | ||
2866c82d JA |
146 | static struct io_u *fio_libaio_event(struct thread_data *td, int event) |
147 | { | |
565e784d | 148 | struct libaio_data *ld = td->io_ops_data; |
f423479d JA |
149 | struct io_event *ev; |
150 | struct io_u *io_u; | |
2866c82d | 151 | |
f423479d | 152 | ev = ld->aio_events + event; |
702906e9 | 153 | io_u = container_of(ev->obj, struct io_u, iocb); |
f423479d JA |
154 | |
155 | if (ev->res != io_u->xfer_buflen) { | |
156 | if (ev->res > io_u->xfer_buflen) | |
157 | io_u->error = -ev->res; | |
158 | else | |
159 | io_u->resid = io_u->xfer_buflen - ev->res; | |
160 | } else | |
161 | io_u->error = 0; | |
162 | ||
163 | return io_u; | |
2866c82d JA |
164 | } |
165 | ||
675012f0 DE |
166 | struct aio_ring { |
167 | unsigned id; /** kernel internal index number */ | |
168 | unsigned nr; /** number of io_events */ | |
169 | unsigned head; | |
170 | unsigned tail; | |
c44b1ff5 | 171 | |
675012f0 DE |
172 | unsigned magic; |
173 | unsigned compat_features; | |
174 | unsigned incompat_features; | |
175 | unsigned header_length; /** size of aio_ring */ | |
176 | ||
177 | struct io_event events[0]; | |
178 | }; | |
179 | ||
180 | #define AIO_RING_MAGIC 0xa10a10a1 | |
181 | ||
182 | static int user_io_getevents(io_context_t aio_ctx, unsigned int max, | |
c44b1ff5 | 183 | struct io_event *events) |
675012f0 DE |
184 | { |
185 | long i = 0; | |
186 | unsigned head; | |
c44b1ff5 | 187 | struct aio_ring *ring = (struct aio_ring*) aio_ctx; |
675012f0 DE |
188 | |
189 | while (i < max) { | |
190 | head = ring->head; | |
191 | ||
192 | if (head == ring->tail) { | |
193 | /* There are no more completions */ | |
194 | break; | |
195 | } else { | |
196 | /* There is another completion to reap */ | |
197 | events[i] = ring->events[head]; | |
d473a06d BVA |
198 | atomic_store_release(&ring->head, |
199 | (head + 1) % ring->nr); | |
675012f0 DE |
200 | i++; |
201 | } | |
202 | } | |
203 | ||
204 | return i; | |
205 | } | |
206 | ||
e7d2e616 | 207 | static int fio_libaio_getevents(struct thread_data *td, unsigned int min, |
1f440ece | 208 | unsigned int max, const struct timespec *t) |
2866c82d | 209 | { |
565e784d | 210 | struct libaio_data *ld = td->io_ops_data; |
de890a1e | 211 | struct libaio_options *o = td->eo; |
82407585 | 212 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
1f440ece | 213 | struct timespec __lt, *lt = NULL; |
0b7fdba7 | 214 | int r, events = 0; |
2866c82d | 215 | |
1f440ece JA |
216 | if (t) { |
217 | __lt = *t; | |
218 | lt = &__lt; | |
219 | } | |
220 | ||
2866c82d | 221 | do { |
de890a1e | 222 | if (o->userspace_reap == 1 |
675012f0 DE |
223 | && actual_min == 0 |
224 | && ((struct aio_ring *)(ld->aio_ctx))->magic | |
225 | == AIO_RING_MAGIC) { | |
226 | r = user_io_getevents(ld->aio_ctx, max, | |
227 | ld->aio_events + events); | |
228 | } else { | |
229 | r = io_getevents(ld->aio_ctx, actual_min, | |
1f440ece | 230 | max, ld->aio_events + events, lt); |
675012f0 | 231 | } |
3441a52d | 232 | if (r > 0) |
0b7fdba7 | 233 | events += r; |
3441a52d | 234 | else if ((min && r == 0) || r == -EAGAIN) { |
0fc2e103 | 235 | fio_libaio_commit(td); |
6347e43d JA |
236 | if (actual_min) |
237 | usleep(10); | |
0fc2e103 | 238 | } else if (r != -EINTR) |
a31dc2dd | 239 | break; |
0b7fdba7 DE |
240 | } while (events < min); |
241 | ||
242 | return r < 0 ? r : events; | |
2866c82d JA |
243 | } |
244 | ||
2e4ef4fb JA |
245 | static enum fio_q_status fio_libaio_queue(struct thread_data *td, |
246 | struct io_u *io_u) | |
2866c82d | 247 | { |
565e784d | 248 | struct libaio_data *ld = td->io_ops_data; |
b2a432bf | 249 | struct libaio_options *o = td->eo; |
2866c82d | 250 | |
7101d9c2 JA |
251 | fio_ro_check(td, io_u); |
252 | ||
acd45e02 | 253 | if (ld->queued == td->o.iodepth) |
755200a3 JA |
254 | return FIO_Q_BUSY; |
255 | ||
256 | /* | |
257 | * fsync is tricky, since it can fail and we need to do it | |
258 | * serialized with other io. the reason is that linux doesn't | |
259 | * support aio fsync yet. So return busy for the case where we | |
260 | * have pending io, to let fio complete those first. | |
261 | */ | |
f011531e | 262 | if (ddir_sync(io_u->ddir)) { |
acd45e02 | 263 | if (ld->queued) |
755200a3 | 264 | return FIO_Q_BUSY; |
5f9099ea | 265 | |
f011531e | 266 | do_io_u_sync(td, io_u); |
755200a3 JA |
267 | return FIO_Q_COMPLETED; |
268 | } | |
269 | ||
a5f3027c | 270 | if (io_u->ddir == DDIR_TRIM) { |
acd45e02 | 271 | if (ld->queued) |
a5f3027c JA |
272 | return FIO_Q_BUSY; |
273 | ||
274 | do_io_u_trim(td, io_u); | |
c0681c9d VF |
275 | io_u_mark_submit(td, 1); |
276 | io_u_mark_complete(td, 1); | |
a5f3027c JA |
277 | return FIO_Q_COMPLETED; |
278 | } | |
279 | ||
b2a432bf PC |
280 | if (o->cmdprio_percentage) |
281 | fio_libaio_prio_prep(td, io_u); | |
282 | ||
702906e9 | 283 | ld->iocbs[ld->head] = &io_u->iocb; |
acd45e02 JA |
284 | ld->io_us[ld->head] = io_u; |
285 | ring_inc(ld, &ld->head, 1); | |
286 | ld->queued++; | |
755200a3 JA |
287 | return FIO_Q_QUEUED; |
288 | } | |
289 | ||
7e77dd02 JA |
290 | static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us, |
291 | unsigned int nr) | |
292 | { | |
8b6a404c | 293 | struct timespec now; |
7e77dd02 JA |
294 | unsigned int i; |
295 | ||
12d9d841 JA |
296 | if (!fio_fill_issue_time(td)) |
297 | return; | |
298 | ||
7e77dd02 JA |
299 | fio_gettime(&now, NULL); |
300 | ||
301 | for (i = 0; i < nr; i++) { | |
302 | struct io_u *io_u = io_us[i]; | |
303 | ||
304 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
305 | io_u_queued(td, io_u); | |
306 | } | |
307 | } | |
308 | ||
755200a3 JA |
309 | static int fio_libaio_commit(struct thread_data *td) |
310 | { | |
565e784d | 311 | struct libaio_data *ld = td->io_ops_data; |
755200a3 | 312 | struct iocb **iocbs; |
7e77dd02 | 313 | struct io_u **io_us; |
8b6a404c | 314 | struct timespec ts; |
a120ca7f | 315 | int ret, wait_start = 0; |
755200a3 | 316 | |
acd45e02 | 317 | if (!ld->queued) |
755200a3 JA |
318 | return 0; |
319 | ||
2866c82d | 320 | do { |
acd45e02 JA |
321 | long nr = ld->queued; |
322 | ||
323 | nr = min((unsigned int) nr, ld->entries - ld->tail); | |
324 | io_us = ld->io_us + ld->tail; | |
325 | iocbs = ld->iocbs + ld->tail; | |
326 | ||
327 | ret = io_submit(ld->aio_ctx, nr, iocbs); | |
5e00c2c4 | 328 | if (ret > 0) { |
7e77dd02 | 329 | fio_libaio_queued(td, io_us, ret); |
838bc709 | 330 | io_u_mark_submit(td, ret); |
acd45e02 JA |
331 | |
332 | ld->queued -= ret; | |
333 | ring_inc(ld, &ld->tail, ret); | |
5e00c2c4 | 334 | ret = 0; |
e3b4e568 | 335 | wait_start = 0; |
acd45e02 | 336 | } else if (ret == -EINTR || !ret) { |
838bc709 JA |
337 | if (!ret) |
338 | io_u_mark_submit(td, ret); | |
e3b4e568 | 339 | wait_start = 0; |
2866c82d | 340 | continue; |
acd45e02 JA |
341 | } else if (ret == -EAGAIN) { |
342 | /* | |
343 | * If we get EAGAIN, we should break out without | |
344 | * error and let the upper layer reap some | |
a120ca7f JA |
345 | * events for us. If we have no queued IO, we |
346 | * must loop here. If we loop for more than 30s, | |
347 | * just error out, something must be buggy in the | |
348 | * IO path. | |
acd45e02 | 349 | */ |
a120ca7f JA |
350 | if (ld->queued) { |
351 | ret = 0; | |
352 | break; | |
353 | } | |
354 | if (!wait_start) { | |
8b6a404c | 355 | fio_gettime(&ts, NULL); |
d36b072d | 356 | wait_start = 1; |
8b6a404c | 357 | } else if (mtime_since_now(&ts) > 30000) { |
a120ca7f JA |
358 | log_err("fio: aio appears to be stalled, giving up\n"); |
359 | break; | |
360 | } | |
361 | usleep(1); | |
362 | continue; | |
a31dc2dd JA |
363 | } else if (ret == -ENOMEM) { |
364 | /* | |
365 | * If we get -ENOMEM, reap events if we can. If | |
366 | * we cannot, treat it as a fatal event since there's | |
367 | * nothing we can do about it. | |
368 | */ | |
369 | if (ld->queued) | |
370 | ret = 0; | |
371 | break; | |
838bc709 | 372 | } else |
2866c82d | 373 | break; |
2c3a4ae9 | 374 | } while (ld->queued); |
2866c82d | 375 | |
36167d82 | 376 | return ret; |
2866c82d JA |
377 | } |
378 | ||
379 | static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u) | |
380 | { | |
565e784d | 381 | struct libaio_data *ld = td->io_ops_data; |
2866c82d JA |
382 | |
383 | return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events); | |
384 | } | |
385 | ||
386 | static void fio_libaio_cleanup(struct thread_data *td) | |
387 | { | |
565e784d | 388 | struct libaio_data *ld = td->io_ops_data; |
2866c82d JA |
389 | |
390 | if (ld) { | |
f24c2649 JA |
391 | /* |
392 | * Work-around to avoid huge RCU stalls at exit time. If we | |
393 | * don't do this here, then it'll be torn down by exit_aio(). | |
394 | * But for that case we can parallellize the freeing, thus | |
395 | * speeding it up a lot. | |
396 | */ | |
397 | if (!(td->flags & TD_F_CHILD)) | |
398 | io_destroy(ld->aio_ctx); | |
7e77dd02 JA |
399 | free(ld->aio_events); |
400 | free(ld->iocbs); | |
401 | free(ld->io_us); | |
2866c82d | 402 | free(ld); |
2866c82d JA |
403 | } |
404 | } | |
405 | ||
2041bd34 JA |
406 | static int fio_libaio_post_init(struct thread_data *td) |
407 | { | |
408 | struct libaio_data *ld = td->io_ops_data; | |
ad4e9298 | 409 | int err; |
2041bd34 | 410 | |
ad4e9298 | 411 | err = io_queue_init(td->o.iodepth, &ld->aio_ctx); |
2041bd34 JA |
412 | if (err) { |
413 | td_verror(td, -err, "io_queue_init"); | |
414 | return 1; | |
415 | } | |
416 | ||
417 | return 0; | |
ebec344d JA |
418 | } |
419 | ||
2866c82d JA |
420 | static int fio_libaio_init(struct thread_data *td) |
421 | { | |
acd45e02 | 422 | struct libaio_data *ld; |
b2a432bf PC |
423 | struct thread_options *to = &td->o; |
424 | struct libaio_options *o = td->eo; | |
2866c82d | 425 | |
acd45e02 | 426 | ld = calloc(1, sizeof(*ld)); |
c1db2dce | 427 | |
acd45e02 JA |
428 | ld->entries = td->o.iodepth; |
429 | ld->is_pow2 = is_power_of_2(ld->entries); | |
430 | ld->aio_events = calloc(ld->entries, sizeof(struct io_event)); | |
431 | ld->iocbs = calloc(ld->entries, sizeof(struct iocb *)); | |
432 | ld->io_us = calloc(ld->entries, sizeof(struct io_u *)); | |
755200a3 | 433 | |
565e784d | 434 | td->io_ops_data = ld; |
b2a432bf PC |
435 | /* |
436 | * Check for option conflicts | |
437 | */ | |
438 | if ((fio_option_is_set(to, ioprio) || fio_option_is_set(to, ioprio_class)) && | |
439 | o->cmdprio_percentage != 0) { | |
440 | log_err("%s: cmdprio_percentage option and mutually exclusive " | |
441 | "prio or prioclass option is set, exiting\n", to->name); | |
442 | td_verror(td, EINVAL, "fio_libaio_init"); | |
443 | return 1; | |
444 | } | |
2866c82d JA |
445 | return 0; |
446 | } | |
447 | ||
5a8a6a03 | 448 | FIO_STATIC struct ioengine_ops ioengine = { |
de890a1e SL |
449 | .name = "libaio", |
450 | .version = FIO_IOOPS_VERSION, | |
04ba61df | 451 | .flags = FIO_ASYNCIO_SYNC_TRIM, |
de890a1e | 452 | .init = fio_libaio_init, |
2041bd34 | 453 | .post_init = fio_libaio_post_init, |
de890a1e SL |
454 | .prep = fio_libaio_prep, |
455 | .queue = fio_libaio_queue, | |
456 | .commit = fio_libaio_commit, | |
457 | .cancel = fio_libaio_cancel, | |
458 | .getevents = fio_libaio_getevents, | |
459 | .event = fio_libaio_event, | |
460 | .cleanup = fio_libaio_cleanup, | |
461 | .open_file = generic_open_file, | |
462 | .close_file = generic_close_file, | |
463 | .get_file_size = generic_get_file_size, | |
464 | .options = options, | |
465 | .option_struct_size = sizeof(struct libaio_options), | |
2866c82d | 466 | }; |
34cfcdaf | 467 | |
5f350952 JA |
468 | static void fio_init fio_libaio_register(void) |
469 | { | |
470 | register_ioengine(&ioengine); | |
471 | } | |
472 | ||
473 | static void fio_exit fio_libaio_unregister(void) | |
474 | { | |
475 | unregister_ioengine(&ioengine); | |
476 | } |