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" |
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 | 30 | static int fio_libaio_commit(struct thread_data *td); |
b2a432bf | 31 | static int fio_libaio_init(struct thread_data *td); |
0fc2e103 | 32 | |
2866c82d JA |
33 | struct 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 | 58 | struct 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 | ||
65 | static 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 | }, | |
90 | #else | |
91 | { | |
92 | .name = "cmdprio_percentage", | |
93 | .lname = "high priority percentage", | |
94 | .type = FIO_OPT_UNSUPPORTED, | |
95 | .help = "Your platform does not support I/O priority classes", | |
96 | }, | |
97 | #endif | |
7d42e66e KK |
98 | { |
99 | .name = "nowait", | |
100 | .lname = "RWF_NOWAIT", | |
101 | .type = FIO_OPT_BOOL, | |
102 | .off1 = offsetof(struct libaio_options, nowait), | |
103 | .help = "Set RWF_NOWAIT for reads/writes", | |
104 | .category = FIO_OPT_C_ENGINE, | |
105 | .group = FIO_OPT_G_LIBAIO, | |
106 | }, | |
de890a1e SL |
107 | { |
108 | .name = NULL, | |
109 | }, | |
110 | }; | |
111 | ||
acd45e02 JA |
112 | static inline void ring_inc(struct libaio_data *ld, unsigned int *val, |
113 | unsigned int add) | |
114 | { | |
115 | if (ld->is_pow2) | |
116 | *val = (*val + add) & (ld->entries - 1); | |
117 | else | |
118 | *val = (*val + add) % ld->entries; | |
119 | } | |
120 | ||
7d42e66e | 121 | static int fio_libaio_prep(struct thread_data *td, struct io_u *io_u) |
2866c82d | 122 | { |
7d42e66e | 123 | struct libaio_options *o = td->eo; |
53cdc686 | 124 | struct fio_file *f = io_u->file; |
ad4e9298 | 125 | struct iocb *iocb = &io_u->iocb; |
53cdc686 | 126 | |
a391d73d | 127 | if (io_u->ddir == DDIR_READ) { |
702906e9 | 128 | io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
7d42e66e KK |
129 | if (o->nowait) |
130 | iocb->aio_rw_flags |= RWF_NOWAIT; | |
a391d73d | 131 | } else if (io_u->ddir == DDIR_WRITE) { |
702906e9 | 132 | io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
7d42e66e KK |
133 | if (o->nowait) |
134 | iocb->aio_rw_flags |= RWF_NOWAIT; | |
a391d73d | 135 | } else if (ddir_sync(io_u->ddir)) |
3c3168e9 | 136 | io_prep_fsync(iocb, f->fd); |
2866c82d JA |
137 | |
138 | return 0; | |
139 | } | |
140 | ||
b2a432bf PC |
141 | static void fio_libaio_prio_prep(struct thread_data *td, struct io_u *io_u) |
142 | { | |
143 | struct libaio_options *o = td->eo; | |
e9f6567a DLM |
144 | struct cmdprio *cmdprio = &o->cmdprio; |
145 | unsigned int p = cmdprio->percentage[io_u->ddir]; | |
146 | ||
147 | if (p && rand_between(&td->prio_state, 0, 99) < p) { | |
8d6e8d99 | 148 | io_u->iocb.aio_reqprio = ioprio_value(IOPRIO_CLASS_RT, 0); |
b2a432bf PC |
149 | io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO; |
150 | io_u->flags |= IO_U_F_PRIORITY; | |
151 | } | |
b2a432bf PC |
152 | } |
153 | ||
2866c82d JA |
154 | static struct io_u *fio_libaio_event(struct thread_data *td, int event) |
155 | { | |
565e784d | 156 | struct libaio_data *ld = td->io_ops_data; |
f423479d JA |
157 | struct io_event *ev; |
158 | struct io_u *io_u; | |
2866c82d | 159 | |
f423479d | 160 | ev = ld->aio_events + event; |
702906e9 | 161 | io_u = container_of(ev->obj, struct io_u, iocb); |
f423479d JA |
162 | |
163 | if (ev->res != io_u->xfer_buflen) { | |
164 | if (ev->res > io_u->xfer_buflen) | |
165 | io_u->error = -ev->res; | |
166 | else | |
167 | io_u->resid = io_u->xfer_buflen - ev->res; | |
168 | } else | |
169 | io_u->error = 0; | |
170 | ||
171 | return io_u; | |
2866c82d JA |
172 | } |
173 | ||
675012f0 DE |
174 | struct aio_ring { |
175 | unsigned id; /** kernel internal index number */ | |
176 | unsigned nr; /** number of io_events */ | |
177 | unsigned head; | |
178 | unsigned tail; | |
c44b1ff5 | 179 | |
675012f0 DE |
180 | unsigned magic; |
181 | unsigned compat_features; | |
182 | unsigned incompat_features; | |
183 | unsigned header_length; /** size of aio_ring */ | |
184 | ||
185 | struct io_event events[0]; | |
186 | }; | |
187 | ||
188 | #define AIO_RING_MAGIC 0xa10a10a1 | |
189 | ||
190 | static int user_io_getevents(io_context_t aio_ctx, unsigned int max, | |
c44b1ff5 | 191 | struct io_event *events) |
675012f0 DE |
192 | { |
193 | long i = 0; | |
194 | unsigned head; | |
c44b1ff5 | 195 | struct aio_ring *ring = (struct aio_ring*) aio_ctx; |
675012f0 DE |
196 | |
197 | while (i < max) { | |
198 | head = ring->head; | |
199 | ||
200 | if (head == ring->tail) { | |
201 | /* There are no more completions */ | |
202 | break; | |
203 | } else { | |
204 | /* There is another completion to reap */ | |
205 | events[i] = ring->events[head]; | |
d473a06d BVA |
206 | atomic_store_release(&ring->head, |
207 | (head + 1) % ring->nr); | |
675012f0 DE |
208 | i++; |
209 | } | |
210 | } | |
211 | ||
212 | return i; | |
213 | } | |
214 | ||
e7d2e616 | 215 | static int fio_libaio_getevents(struct thread_data *td, unsigned int min, |
1f440ece | 216 | unsigned int max, const struct timespec *t) |
2866c82d | 217 | { |
565e784d | 218 | struct libaio_data *ld = td->io_ops_data; |
de890a1e | 219 | struct libaio_options *o = td->eo; |
82407585 | 220 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
1f440ece | 221 | struct timespec __lt, *lt = NULL; |
0b7fdba7 | 222 | int r, events = 0; |
2866c82d | 223 | |
1f440ece JA |
224 | if (t) { |
225 | __lt = *t; | |
226 | lt = &__lt; | |
227 | } | |
228 | ||
2866c82d | 229 | do { |
de890a1e | 230 | if (o->userspace_reap == 1 |
675012f0 DE |
231 | && actual_min == 0 |
232 | && ((struct aio_ring *)(ld->aio_ctx))->magic | |
233 | == AIO_RING_MAGIC) { | |
234 | r = user_io_getevents(ld->aio_ctx, max, | |
235 | ld->aio_events + events); | |
236 | } else { | |
237 | r = io_getevents(ld->aio_ctx, actual_min, | |
1f440ece | 238 | max, ld->aio_events + events, lt); |
675012f0 | 239 | } |
3441a52d | 240 | if (r > 0) |
0b7fdba7 | 241 | events += r; |
3441a52d | 242 | else if ((min && r == 0) || r == -EAGAIN) { |
0fc2e103 | 243 | fio_libaio_commit(td); |
6347e43d JA |
244 | if (actual_min) |
245 | usleep(10); | |
0fc2e103 | 246 | } else if (r != -EINTR) |
a31dc2dd | 247 | break; |
0b7fdba7 DE |
248 | } while (events < min); |
249 | ||
250 | return r < 0 ? r : events; | |
2866c82d JA |
251 | } |
252 | ||
2e4ef4fb JA |
253 | static enum fio_q_status fio_libaio_queue(struct thread_data *td, |
254 | struct io_u *io_u) | |
2866c82d | 255 | { |
565e784d | 256 | struct libaio_data *ld = td->io_ops_data; |
2866c82d | 257 | |
7101d9c2 JA |
258 | fio_ro_check(td, io_u); |
259 | ||
acd45e02 | 260 | if (ld->queued == td->o.iodepth) |
755200a3 JA |
261 | return FIO_Q_BUSY; |
262 | ||
263 | /* | |
264 | * fsync is tricky, since it can fail and we need to do it | |
265 | * serialized with other io. the reason is that linux doesn't | |
266 | * support aio fsync yet. So return busy for the case where we | |
267 | * have pending io, to let fio complete those first. | |
268 | */ | |
f011531e | 269 | if (ddir_sync(io_u->ddir)) { |
acd45e02 | 270 | if (ld->queued) |
755200a3 | 271 | return FIO_Q_BUSY; |
5f9099ea | 272 | |
f011531e | 273 | do_io_u_sync(td, io_u); |
755200a3 JA |
274 | return FIO_Q_COMPLETED; |
275 | } | |
276 | ||
a5f3027c | 277 | if (io_u->ddir == DDIR_TRIM) { |
acd45e02 | 278 | if (ld->queued) |
a5f3027c JA |
279 | return FIO_Q_BUSY; |
280 | ||
281 | do_io_u_trim(td, io_u); | |
c0681c9d VF |
282 | io_u_mark_submit(td, 1); |
283 | io_u_mark_complete(td, 1); | |
a5f3027c JA |
284 | return FIO_Q_COMPLETED; |
285 | } | |
286 | ||
e9f6567a | 287 | if (ld->use_cmdprio) |
b2a432bf PC |
288 | fio_libaio_prio_prep(td, io_u); |
289 | ||
702906e9 | 290 | ld->iocbs[ld->head] = &io_u->iocb; |
acd45e02 JA |
291 | ld->io_us[ld->head] = io_u; |
292 | ring_inc(ld, &ld->head, 1); | |
293 | ld->queued++; | |
755200a3 JA |
294 | return FIO_Q_QUEUED; |
295 | } | |
296 | ||
7e77dd02 JA |
297 | static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us, |
298 | unsigned int nr) | |
299 | { | |
8b6a404c | 300 | struct timespec now; |
7e77dd02 JA |
301 | unsigned int i; |
302 | ||
12d9d841 JA |
303 | if (!fio_fill_issue_time(td)) |
304 | return; | |
305 | ||
7e77dd02 JA |
306 | fio_gettime(&now, NULL); |
307 | ||
308 | for (i = 0; i < nr; i++) { | |
309 | struct io_u *io_u = io_us[i]; | |
310 | ||
311 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
312 | io_u_queued(td, io_u); | |
313 | } | |
314 | } | |
315 | ||
755200a3 JA |
316 | static int fio_libaio_commit(struct thread_data *td) |
317 | { | |
565e784d | 318 | struct libaio_data *ld = td->io_ops_data; |
755200a3 | 319 | struct iocb **iocbs; |
7e77dd02 | 320 | struct io_u **io_us; |
8b6a404c | 321 | struct timespec ts; |
a120ca7f | 322 | int ret, wait_start = 0; |
755200a3 | 323 | |
acd45e02 | 324 | if (!ld->queued) |
755200a3 JA |
325 | return 0; |
326 | ||
2866c82d | 327 | do { |
acd45e02 JA |
328 | long nr = ld->queued; |
329 | ||
330 | nr = min((unsigned int) nr, ld->entries - ld->tail); | |
331 | io_us = ld->io_us + ld->tail; | |
332 | iocbs = ld->iocbs + ld->tail; | |
333 | ||
334 | ret = io_submit(ld->aio_ctx, nr, iocbs); | |
5e00c2c4 | 335 | if (ret > 0) { |
7e77dd02 | 336 | fio_libaio_queued(td, io_us, ret); |
838bc709 | 337 | io_u_mark_submit(td, ret); |
acd45e02 JA |
338 | |
339 | ld->queued -= ret; | |
340 | ring_inc(ld, &ld->tail, ret); | |
5e00c2c4 | 341 | ret = 0; |
e3b4e568 | 342 | wait_start = 0; |
acd45e02 | 343 | } else if (ret == -EINTR || !ret) { |
838bc709 JA |
344 | if (!ret) |
345 | io_u_mark_submit(td, ret); | |
e3b4e568 | 346 | wait_start = 0; |
2866c82d | 347 | continue; |
acd45e02 JA |
348 | } else if (ret == -EAGAIN) { |
349 | /* | |
350 | * If we get EAGAIN, we should break out without | |
351 | * error and let the upper layer reap some | |
a120ca7f JA |
352 | * events for us. If we have no queued IO, we |
353 | * must loop here. If we loop for more than 30s, | |
354 | * just error out, something must be buggy in the | |
355 | * IO path. | |
acd45e02 | 356 | */ |
a120ca7f JA |
357 | if (ld->queued) { |
358 | ret = 0; | |
359 | break; | |
360 | } | |
361 | if (!wait_start) { | |
8b6a404c | 362 | fio_gettime(&ts, NULL); |
d36b072d | 363 | wait_start = 1; |
8b6a404c | 364 | } else if (mtime_since_now(&ts) > 30000) { |
a120ca7f JA |
365 | log_err("fio: aio appears to be stalled, giving up\n"); |
366 | break; | |
367 | } | |
368 | usleep(1); | |
369 | continue; | |
a31dc2dd JA |
370 | } else if (ret == -ENOMEM) { |
371 | /* | |
372 | * If we get -ENOMEM, reap events if we can. If | |
373 | * we cannot, treat it as a fatal event since there's | |
374 | * nothing we can do about it. | |
375 | */ | |
376 | if (ld->queued) | |
377 | ret = 0; | |
378 | break; | |
838bc709 | 379 | } else |
2866c82d | 380 | break; |
2c3a4ae9 | 381 | } while (ld->queued); |
2866c82d | 382 | |
36167d82 | 383 | return ret; |
2866c82d JA |
384 | } |
385 | ||
386 | static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u) | |
387 | { | |
565e784d | 388 | struct libaio_data *ld = td->io_ops_data; |
2866c82d JA |
389 | |
390 | return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events); | |
391 | } | |
392 | ||
393 | static void fio_libaio_cleanup(struct thread_data *td) | |
394 | { | |
565e784d | 395 | struct libaio_data *ld = td->io_ops_data; |
2866c82d JA |
396 | |
397 | if (ld) { | |
f24c2649 JA |
398 | /* |
399 | * Work-around to avoid huge RCU stalls at exit time. If we | |
400 | * don't do this here, then it'll be torn down by exit_aio(). | |
401 | * But for that case we can parallellize the freeing, thus | |
402 | * speeding it up a lot. | |
403 | */ | |
404 | if (!(td->flags & TD_F_CHILD)) | |
405 | io_destroy(ld->aio_ctx); | |
7e77dd02 JA |
406 | free(ld->aio_events); |
407 | free(ld->iocbs); | |
408 | free(ld->io_us); | |
2866c82d | 409 | free(ld); |
2866c82d JA |
410 | } |
411 | } | |
412 | ||
2041bd34 JA |
413 | static int fio_libaio_post_init(struct thread_data *td) |
414 | { | |
415 | struct libaio_data *ld = td->io_ops_data; | |
ad4e9298 | 416 | int err; |
2041bd34 | 417 | |
ad4e9298 | 418 | err = io_queue_init(td->o.iodepth, &ld->aio_ctx); |
2041bd34 JA |
419 | if (err) { |
420 | td_verror(td, -err, "io_queue_init"); | |
421 | return 1; | |
422 | } | |
423 | ||
424 | return 0; | |
ebec344d JA |
425 | } |
426 | ||
2866c82d JA |
427 | static int fio_libaio_init(struct thread_data *td) |
428 | { | |
acd45e02 | 429 | struct libaio_data *ld; |
b2a432bf | 430 | struct libaio_options *o = td->eo; |
e9f6567a DLM |
431 | struct cmdprio *cmdprio = &o->cmdprio; |
432 | int ret; | |
2866c82d | 433 | |
acd45e02 | 434 | ld = calloc(1, sizeof(*ld)); |
c1db2dce | 435 | |
acd45e02 JA |
436 | ld->entries = td->o.iodepth; |
437 | ld->is_pow2 = is_power_of_2(ld->entries); | |
438 | ld->aio_events = calloc(ld->entries, sizeof(struct io_event)); | |
439 | ld->iocbs = calloc(ld->entries, sizeof(struct iocb *)); | |
440 | ld->io_us = calloc(ld->entries, sizeof(struct io_u *)); | |
755200a3 | 441 | |
565e784d | 442 | td->io_ops_data = ld; |
e9f6567a DLM |
443 | |
444 | ret = fio_cmdprio_init(td, cmdprio, &ld->use_cmdprio); | |
445 | if (ret) { | |
b2a432bf PC |
446 | td_verror(td, EINVAL, "fio_libaio_init"); |
447 | return 1; | |
448 | } | |
e9f6567a | 449 | |
2866c82d JA |
450 | return 0; |
451 | } | |
452 | ||
5a8a6a03 | 453 | FIO_STATIC struct ioengine_ops ioengine = { |
de890a1e SL |
454 | .name = "libaio", |
455 | .version = FIO_IOOPS_VERSION, | |
04ba61df | 456 | .flags = FIO_ASYNCIO_SYNC_TRIM, |
de890a1e | 457 | .init = fio_libaio_init, |
2041bd34 | 458 | .post_init = fio_libaio_post_init, |
de890a1e SL |
459 | .prep = fio_libaio_prep, |
460 | .queue = fio_libaio_queue, | |
461 | .commit = fio_libaio_commit, | |
462 | .cancel = fio_libaio_cancel, | |
463 | .getevents = fio_libaio_getevents, | |
464 | .event = fio_libaio_event, | |
465 | .cleanup = fio_libaio_cleanup, | |
466 | .open_file = generic_open_file, | |
467 | .close_file = generic_close_file, | |
468 | .get_file_size = generic_get_file_size, | |
469 | .options = options, | |
470 | .option_struct_size = sizeof(struct libaio_options), | |
2866c82d | 471 | }; |
34cfcdaf | 472 | |
5f350952 JA |
473 | static void fio_init fio_libaio_register(void) |
474 | { | |
475 | register_ioengine(&ioengine); | |
476 | } | |
477 | ||
478 | static void fio_exit fio_libaio_unregister(void) | |
479 | { | |
480 | unregister_ioengine(&ioengine); | |
481 | } |