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> |
5f350952 JA |
11 | |
12 | #include "../fio.h" | |
0f38bbef | 13 | #include "../lib/pow2.h" |
d220c761 | 14 | #include "../optgroup.h" |
2866c82d | 15 | |
ee636f3f JA |
16 | #ifndef IOCB_FLAG_HIPRI |
17 | #define IOCB_FLAG_HIPRI (1 << 2) | |
18 | #endif | |
ebec344d JA |
19 | #ifndef IOCTX_FLAG_IOPOLL |
20 | #define IOCTX_FLAG_IOPOLL (1 << 0) | |
21 | #endif | |
a391d73d | 22 | |
0fc2e103 JA |
23 | static int fio_libaio_commit(struct thread_data *td); |
24 | ||
2866c82d JA |
25 | struct libaio_data { |
26 | io_context_t aio_ctx; | |
27 | struct io_event *aio_events; | |
755200a3 | 28 | struct iocb **iocbs; |
7e77dd02 | 29 | struct io_u **io_us; |
acd45e02 JA |
30 | |
31 | /* | |
32 | * Basic ring buffer. 'head' is incremented in _queue(), and | |
33 | * 'tail' is incremented in _commit(). We keep 'queued' so | |
34 | * that we know if the ring is full or empty, when | |
35 | * 'head' == 'tail'. 'entries' is the ring size, and | |
36 | * 'is_pow2' is just an optimization to use AND instead of | |
37 | * modulus to get the remainder on ring increment. | |
38 | */ | |
39 | int is_pow2; | |
40 | unsigned int entries; | |
41 | unsigned int queued; | |
42 | unsigned int head; | |
43 | unsigned int tail; | |
2866c82d JA |
44 | }; |
45 | ||
de890a1e | 46 | struct libaio_options { |
a1f871c7 | 47 | void *pad; |
de890a1e | 48 | unsigned int userspace_reap; |
a391d73d | 49 | unsigned int hipri; |
de890a1e SL |
50 | }; |
51 | ||
52 | static struct fio_option options[] = { | |
53 | { | |
54 | .name = "userspace_reap", | |
e8b0e958 | 55 | .lname = "Libaio userspace reaping", |
de890a1e SL |
56 | .type = FIO_OPT_STR_SET, |
57 | .off1 = offsetof(struct libaio_options, userspace_reap), | |
58 | .help = "Use alternative user-space reap implementation", | |
e90a0adf | 59 | .category = FIO_OPT_C_ENGINE, |
5a3cd5f3 | 60 | .group = FIO_OPT_G_LIBAIO, |
de890a1e | 61 | }, |
a391d73d JA |
62 | { |
63 | .name = "hipri", | |
ee636f3f | 64 | .lname = "High Priority", |
a391d73d JA |
65 | .type = FIO_OPT_STR_SET, |
66 | .off1 = offsetof(struct libaio_options, hipri), | |
6acc46d8 | 67 | .help = "Use polled IO completions", |
a391d73d JA |
68 | .category = FIO_OPT_C_ENGINE, |
69 | .group = FIO_OPT_G_LIBAIO, | |
70 | }, | |
de890a1e SL |
71 | { |
72 | .name = NULL, | |
73 | }, | |
74 | }; | |
75 | ||
acd45e02 JA |
76 | static inline void ring_inc(struct libaio_data *ld, unsigned int *val, |
77 | unsigned int add) | |
78 | { | |
79 | if (ld->is_pow2) | |
80 | *val = (*val + add) & (ld->entries - 1); | |
81 | else | |
82 | *val = (*val + add) % ld->entries; | |
83 | } | |
84 | ||
7a16dd02 | 85 | static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u) |
2866c82d | 86 | { |
53cdc686 | 87 | struct fio_file *f = io_u->file; |
a391d73d | 88 | struct libaio_options *o = td->eo; |
53cdc686 | 89 | |
a391d73d | 90 | if (io_u->ddir == DDIR_READ) { |
cec6b55d | 91 | io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
a391d73d | 92 | if (o->hipri) |
ee636f3f | 93 | io_u->iocb.u.c.flags |= IOCB_FLAG_HIPRI; |
a391d73d | 94 | } else if (io_u->ddir == DDIR_WRITE) { |
cec6b55d | 95 | io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
a391d73d | 96 | if (o->hipri) |
ee636f3f | 97 | io_u->iocb.u.c.flags |= IOCB_FLAG_HIPRI; |
a391d73d | 98 | } else if (ddir_sync(io_u->ddir)) |
87dc1ab1 | 99 | io_prep_fsync(&io_u->iocb, f->fd); |
2866c82d JA |
100 | |
101 | return 0; | |
102 | } | |
103 | ||
104 | static struct io_u *fio_libaio_event(struct thread_data *td, int event) | |
105 | { | |
565e784d | 106 | struct libaio_data *ld = td->io_ops_data; |
f423479d JA |
107 | struct io_event *ev; |
108 | struct io_u *io_u; | |
2866c82d | 109 | |
f423479d | 110 | ev = ld->aio_events + event; |
0cc6ee59 | 111 | io_u = container_of(ev->obj, struct io_u, iocb); |
f423479d JA |
112 | |
113 | if (ev->res != io_u->xfer_buflen) { | |
114 | if (ev->res > io_u->xfer_buflen) | |
115 | io_u->error = -ev->res; | |
116 | else | |
117 | io_u->resid = io_u->xfer_buflen - ev->res; | |
118 | } else | |
119 | io_u->error = 0; | |
120 | ||
121 | return io_u; | |
2866c82d JA |
122 | } |
123 | ||
675012f0 DE |
124 | struct aio_ring { |
125 | unsigned id; /** kernel internal index number */ | |
126 | unsigned nr; /** number of io_events */ | |
127 | unsigned head; | |
128 | unsigned tail; | |
c44b1ff5 | 129 | |
675012f0 DE |
130 | unsigned magic; |
131 | unsigned compat_features; | |
132 | unsigned incompat_features; | |
133 | unsigned header_length; /** size of aio_ring */ | |
134 | ||
135 | struct io_event events[0]; | |
136 | }; | |
137 | ||
138 | #define AIO_RING_MAGIC 0xa10a10a1 | |
139 | ||
140 | static int user_io_getevents(io_context_t aio_ctx, unsigned int max, | |
c44b1ff5 | 141 | struct io_event *events) |
675012f0 DE |
142 | { |
143 | long i = 0; | |
144 | unsigned head; | |
c44b1ff5 | 145 | struct aio_ring *ring = (struct aio_ring*) aio_ctx; |
675012f0 DE |
146 | |
147 | while (i < max) { | |
148 | head = ring->head; | |
149 | ||
150 | if (head == ring->tail) { | |
151 | /* There are no more completions */ | |
152 | break; | |
153 | } else { | |
154 | /* There is another completion to reap */ | |
155 | events[i] = ring->events[head]; | |
156 | read_barrier(); | |
c44b1ff5 | 157 | ring->head = (head + 1) % ring->nr; |
675012f0 DE |
158 | i++; |
159 | } | |
160 | } | |
161 | ||
162 | return i; | |
163 | } | |
164 | ||
e7d2e616 | 165 | static int fio_libaio_getevents(struct thread_data *td, unsigned int min, |
1f440ece | 166 | unsigned int max, const struct timespec *t) |
2866c82d | 167 | { |
565e784d | 168 | struct libaio_data *ld = td->io_ops_data; |
de890a1e | 169 | struct libaio_options *o = td->eo; |
82407585 | 170 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
1f440ece | 171 | struct timespec __lt, *lt = NULL; |
0b7fdba7 | 172 | int r, events = 0; |
2866c82d | 173 | |
1f440ece JA |
174 | if (t) { |
175 | __lt = *t; | |
176 | lt = &__lt; | |
177 | } | |
178 | ||
2866c82d | 179 | do { |
de890a1e | 180 | if (o->userspace_reap == 1 |
675012f0 DE |
181 | && actual_min == 0 |
182 | && ((struct aio_ring *)(ld->aio_ctx))->magic | |
183 | == AIO_RING_MAGIC) { | |
184 | r = user_io_getevents(ld->aio_ctx, max, | |
185 | ld->aio_events + events); | |
186 | } else { | |
187 | r = io_getevents(ld->aio_ctx, actual_min, | |
1f440ece | 188 | max, ld->aio_events + events, lt); |
675012f0 | 189 | } |
3441a52d | 190 | if (r > 0) |
0b7fdba7 | 191 | events += r; |
3441a52d | 192 | else if ((min && r == 0) || r == -EAGAIN) { |
0fc2e103 | 193 | fio_libaio_commit(td); |
6347e43d JA |
194 | if (actual_min) |
195 | usleep(10); | |
0fc2e103 | 196 | } else if (r != -EINTR) |
a31dc2dd | 197 | break; |
0b7fdba7 DE |
198 | } while (events < min); |
199 | ||
200 | return r < 0 ? r : events; | |
2866c82d JA |
201 | } |
202 | ||
2e4ef4fb JA |
203 | static enum fio_q_status fio_libaio_queue(struct thread_data *td, |
204 | struct io_u *io_u) | |
2866c82d | 205 | { |
565e784d | 206 | struct libaio_data *ld = td->io_ops_data; |
2866c82d | 207 | |
7101d9c2 JA |
208 | fio_ro_check(td, io_u); |
209 | ||
acd45e02 | 210 | if (ld->queued == td->o.iodepth) |
755200a3 JA |
211 | return FIO_Q_BUSY; |
212 | ||
213 | /* | |
214 | * fsync is tricky, since it can fail and we need to do it | |
215 | * serialized with other io. the reason is that linux doesn't | |
216 | * support aio fsync yet. So return busy for the case where we | |
217 | * have pending io, to let fio complete those first. | |
218 | */ | |
f011531e | 219 | if (ddir_sync(io_u->ddir)) { |
acd45e02 | 220 | if (ld->queued) |
755200a3 | 221 | return FIO_Q_BUSY; |
5f9099ea | 222 | |
f011531e | 223 | do_io_u_sync(td, io_u); |
755200a3 JA |
224 | return FIO_Q_COMPLETED; |
225 | } | |
226 | ||
a5f3027c | 227 | if (io_u->ddir == DDIR_TRIM) { |
acd45e02 | 228 | if (ld->queued) |
a5f3027c JA |
229 | return FIO_Q_BUSY; |
230 | ||
231 | do_io_u_trim(td, io_u); | |
c0681c9d VF |
232 | io_u_mark_submit(td, 1); |
233 | io_u_mark_complete(td, 1); | |
a5f3027c JA |
234 | return FIO_Q_COMPLETED; |
235 | } | |
236 | ||
acd45e02 JA |
237 | ld->iocbs[ld->head] = &io_u->iocb; |
238 | ld->io_us[ld->head] = io_u; | |
239 | ring_inc(ld, &ld->head, 1); | |
240 | ld->queued++; | |
755200a3 JA |
241 | return FIO_Q_QUEUED; |
242 | } | |
243 | ||
7e77dd02 JA |
244 | static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us, |
245 | unsigned int nr) | |
246 | { | |
8b6a404c | 247 | struct timespec now; |
7e77dd02 JA |
248 | unsigned int i; |
249 | ||
12d9d841 JA |
250 | if (!fio_fill_issue_time(td)) |
251 | return; | |
252 | ||
7e77dd02 JA |
253 | fio_gettime(&now, NULL); |
254 | ||
255 | for (i = 0; i < nr; i++) { | |
256 | struct io_u *io_u = io_us[i]; | |
257 | ||
258 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
259 | io_u_queued(td, io_u); | |
260 | } | |
261 | } | |
262 | ||
755200a3 JA |
263 | static int fio_libaio_commit(struct thread_data *td) |
264 | { | |
565e784d | 265 | struct libaio_data *ld = td->io_ops_data; |
755200a3 | 266 | struct iocb **iocbs; |
7e77dd02 | 267 | struct io_u **io_us; |
8b6a404c | 268 | struct timespec ts; |
a120ca7f | 269 | int ret, wait_start = 0; |
755200a3 | 270 | |
acd45e02 | 271 | if (!ld->queued) |
755200a3 JA |
272 | return 0; |
273 | ||
2866c82d | 274 | do { |
acd45e02 JA |
275 | long nr = ld->queued; |
276 | ||
277 | nr = min((unsigned int) nr, ld->entries - ld->tail); | |
278 | io_us = ld->io_us + ld->tail; | |
279 | iocbs = ld->iocbs + ld->tail; | |
280 | ||
281 | ret = io_submit(ld->aio_ctx, nr, iocbs); | |
5e00c2c4 | 282 | if (ret > 0) { |
7e77dd02 | 283 | fio_libaio_queued(td, io_us, ret); |
838bc709 | 284 | io_u_mark_submit(td, ret); |
acd45e02 JA |
285 | |
286 | ld->queued -= ret; | |
287 | ring_inc(ld, &ld->tail, ret); | |
5e00c2c4 | 288 | ret = 0; |
e3b4e568 | 289 | wait_start = 0; |
acd45e02 | 290 | } else if (ret == -EINTR || !ret) { |
838bc709 JA |
291 | if (!ret) |
292 | io_u_mark_submit(td, ret); | |
e3b4e568 | 293 | wait_start = 0; |
2866c82d | 294 | continue; |
acd45e02 JA |
295 | } else if (ret == -EAGAIN) { |
296 | /* | |
297 | * If we get EAGAIN, we should break out without | |
298 | * error and let the upper layer reap some | |
a120ca7f JA |
299 | * events for us. If we have no queued IO, we |
300 | * must loop here. If we loop for more than 30s, | |
301 | * just error out, something must be buggy in the | |
302 | * IO path. | |
acd45e02 | 303 | */ |
a120ca7f JA |
304 | if (ld->queued) { |
305 | ret = 0; | |
306 | break; | |
307 | } | |
308 | if (!wait_start) { | |
8b6a404c | 309 | fio_gettime(&ts, NULL); |
d36b072d | 310 | wait_start = 1; |
8b6a404c | 311 | } else if (mtime_since_now(&ts) > 30000) { |
a120ca7f JA |
312 | log_err("fio: aio appears to be stalled, giving up\n"); |
313 | break; | |
314 | } | |
315 | usleep(1); | |
316 | continue; | |
a31dc2dd JA |
317 | } else if (ret == -ENOMEM) { |
318 | /* | |
319 | * If we get -ENOMEM, reap events if we can. If | |
320 | * we cannot, treat it as a fatal event since there's | |
321 | * nothing we can do about it. | |
322 | */ | |
323 | if (ld->queued) | |
324 | ret = 0; | |
325 | break; | |
838bc709 | 326 | } else |
2866c82d | 327 | break; |
2c3a4ae9 | 328 | } while (ld->queued); |
2866c82d | 329 | |
36167d82 | 330 | return ret; |
2866c82d JA |
331 | } |
332 | ||
333 | static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u) | |
334 | { | |
565e784d | 335 | struct libaio_data *ld = td->io_ops_data; |
2866c82d JA |
336 | |
337 | return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events); | |
338 | } | |
339 | ||
340 | static void fio_libaio_cleanup(struct thread_data *td) | |
341 | { | |
565e784d | 342 | struct libaio_data *ld = td->io_ops_data; |
2866c82d JA |
343 | |
344 | if (ld) { | |
f24c2649 JA |
345 | /* |
346 | * Work-around to avoid huge RCU stalls at exit time. If we | |
347 | * don't do this here, then it'll be torn down by exit_aio(). | |
348 | * But for that case we can parallellize the freeing, thus | |
349 | * speeding it up a lot. | |
350 | */ | |
351 | if (!(td->flags & TD_F_CHILD)) | |
352 | io_destroy(ld->aio_ctx); | |
7e77dd02 JA |
353 | free(ld->aio_events); |
354 | free(ld->iocbs); | |
355 | free(ld->io_us); | |
2866c82d | 356 | free(ld); |
2866c82d JA |
357 | } |
358 | } | |
359 | ||
ebec344d JA |
360 | static int fio_libaio_queue_init(struct libaio_data *ld, unsigned int depth, |
361 | bool hipri) | |
362 | { | |
363 | #ifdef __NR_sys_io_setup2 | |
364 | int err, flags = 0; | |
365 | ||
366 | if (hipri) | |
367 | flags = IOCTX_FLAG_IOPOLL; | |
368 | ||
369 | return syscall(__NR_sys_io_setup2, depth, flags, &ld->aio_ctx); | |
370 | #else | |
371 | if (hipri) { | |
372 | log_err("fio: polled aio not available on your platform\n"); | |
373 | return 1; | |
374 | } | |
375 | return io_queue_init(depth, &ld->aio_ctx); | |
376 | #endif | |
377 | } | |
378 | ||
2866c82d JA |
379 | static int fio_libaio_init(struct thread_data *td) |
380 | { | |
c90e1017 | 381 | struct libaio_options *o = td->eo; |
acd45e02 | 382 | struct libaio_data *ld; |
c90e1017 | 383 | int err = 0; |
2866c82d | 384 | |
acd45e02 | 385 | ld = calloc(1, sizeof(*ld)); |
c1db2dce | 386 | |
c90e1017 JA |
387 | /* |
388 | * First try passing in 0 for queue depth, since we don't | |
389 | * care about the user ring. If that fails, the kernel is too old | |
390 | * and we need the right depth. | |
391 | */ | |
392 | if (!o->userspace_reap) | |
ebec344d | 393 | err = fio_libaio_queue_init(ld, INT_MAX, o->hipri); |
c90e1017 | 394 | if (o->userspace_reap || err == -EINVAL) |
ebec344d | 395 | err = fio_libaio_queue_init(ld, td->o.iodepth, o->hipri); |
c1db2dce JA |
396 | if (err) { |
397 | td_verror(td, -err, "io_queue_init"); | |
75de55ac | 398 | log_err("fio: check /proc/sys/fs/aio-max-nr\n"); |
cb781c75 | 399 | free(ld); |
2866c82d JA |
400 | return 1; |
401 | } | |
402 | ||
acd45e02 JA |
403 | ld->entries = td->o.iodepth; |
404 | ld->is_pow2 = is_power_of_2(ld->entries); | |
405 | ld->aio_events = calloc(ld->entries, sizeof(struct io_event)); | |
406 | ld->iocbs = calloc(ld->entries, sizeof(struct iocb *)); | |
407 | ld->io_us = calloc(ld->entries, sizeof(struct io_u *)); | |
755200a3 | 408 | |
565e784d | 409 | td->io_ops_data = ld; |
2866c82d JA |
410 | return 0; |
411 | } | |
412 | ||
5f350952 | 413 | static struct ioengine_ops ioengine = { |
de890a1e SL |
414 | .name = "libaio", |
415 | .version = FIO_IOOPS_VERSION, | |
416 | .init = fio_libaio_init, | |
417 | .prep = fio_libaio_prep, | |
418 | .queue = fio_libaio_queue, | |
419 | .commit = fio_libaio_commit, | |
420 | .cancel = fio_libaio_cancel, | |
421 | .getevents = fio_libaio_getevents, | |
422 | .event = fio_libaio_event, | |
423 | .cleanup = fio_libaio_cleanup, | |
424 | .open_file = generic_open_file, | |
425 | .close_file = generic_close_file, | |
426 | .get_file_size = generic_get_file_size, | |
427 | .options = options, | |
428 | .option_struct_size = sizeof(struct libaio_options), | |
2866c82d | 429 | }; |
34cfcdaf | 430 | |
5f350952 JA |
431 | static void fio_init fio_libaio_register(void) |
432 | { | |
433 | register_ioengine(&ioengine); | |
434 | } | |
435 | ||
436 | static void fio_exit fio_libaio_unregister(void) | |
437 | { | |
438 | unregister_ioengine(&ioengine); | |
439 | } |