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 | |
3c3168e9 JA |
22 | #ifndef IOCTX_FLAG_USERIOCB |
23 | #define IOCTX_FLAG_USERIOCB (1 << 1) | |
24 | #endif | |
25 | ||
a391d73d | 26 | |
0fc2e103 JA |
27 | static int fio_libaio_commit(struct thread_data *td); |
28 | ||
2866c82d JA |
29 | struct libaio_data { |
30 | io_context_t aio_ctx; | |
31 | struct io_event *aio_events; | |
755200a3 | 32 | struct iocb **iocbs; |
7e77dd02 | 33 | struct io_u **io_us; |
acd45e02 | 34 | |
3c3168e9 JA |
35 | struct iocb *user_iocbs; |
36 | struct io_u **io_u_index; | |
37 | ||
acd45e02 JA |
38 | /* |
39 | * Basic ring buffer. 'head' is incremented in _queue(), and | |
40 | * 'tail' is incremented in _commit(). We keep 'queued' so | |
41 | * that we know if the ring is full or empty, when | |
42 | * 'head' == 'tail'. 'entries' is the ring size, and | |
43 | * 'is_pow2' is just an optimization to use AND instead of | |
44 | * modulus to get the remainder on ring increment. | |
45 | */ | |
46 | int is_pow2; | |
47 | unsigned int entries; | |
48 | unsigned int queued; | |
49 | unsigned int head; | |
50 | unsigned int tail; | |
2866c82d JA |
51 | }; |
52 | ||
de890a1e | 53 | struct libaio_options { |
a1f871c7 | 54 | void *pad; |
de890a1e | 55 | unsigned int userspace_reap; |
a391d73d | 56 | unsigned int hipri; |
3c3168e9 | 57 | unsigned int useriocb; |
de890a1e SL |
58 | }; |
59 | ||
60 | static struct fio_option options[] = { | |
61 | { | |
62 | .name = "userspace_reap", | |
e8b0e958 | 63 | .lname = "Libaio userspace reaping", |
de890a1e SL |
64 | .type = FIO_OPT_STR_SET, |
65 | .off1 = offsetof(struct libaio_options, userspace_reap), | |
66 | .help = "Use alternative user-space reap implementation", | |
e90a0adf | 67 | .category = FIO_OPT_C_ENGINE, |
5a3cd5f3 | 68 | .group = FIO_OPT_G_LIBAIO, |
de890a1e | 69 | }, |
a391d73d JA |
70 | { |
71 | .name = "hipri", | |
ee636f3f | 72 | .lname = "High Priority", |
a391d73d JA |
73 | .type = FIO_OPT_STR_SET, |
74 | .off1 = offsetof(struct libaio_options, hipri), | |
6acc46d8 | 75 | .help = "Use polled IO completions", |
a391d73d JA |
76 | .category = FIO_OPT_C_ENGINE, |
77 | .group = FIO_OPT_G_LIBAIO, | |
78 | }, | |
3c3168e9 JA |
79 | { |
80 | .name = "useriocb", | |
81 | .lname = "User IOCBs", | |
82 | .type = FIO_OPT_STR_SET, | |
83 | .off1 = offsetof(struct libaio_options, useriocb), | |
84 | .help = "Use user mapped IOCBs", | |
85 | .category = FIO_OPT_C_ENGINE, | |
86 | .group = FIO_OPT_G_LIBAIO, | |
87 | }, | |
de890a1e SL |
88 | { |
89 | .name = NULL, | |
90 | }, | |
91 | }; | |
92 | ||
acd45e02 JA |
93 | static inline void ring_inc(struct libaio_data *ld, unsigned int *val, |
94 | unsigned int add) | |
95 | { | |
96 | if (ld->is_pow2) | |
97 | *val = (*val + add) & (ld->entries - 1); | |
98 | else | |
99 | *val = (*val + add) % ld->entries; | |
100 | } | |
101 | ||
7a16dd02 | 102 | static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u) |
2866c82d | 103 | { |
3c3168e9 | 104 | struct libaio_data *ld = td->io_ops_data; |
53cdc686 | 105 | struct fio_file *f = io_u->file; |
a391d73d | 106 | struct libaio_options *o = td->eo; |
3c3168e9 JA |
107 | struct iocb *iocb; |
108 | ||
109 | if (o->useriocb) | |
110 | iocb = &ld->user_iocbs[io_u->index]; | |
111 | else | |
112 | iocb = &io_u->iocb; | |
53cdc686 | 113 | |
a391d73d | 114 | if (io_u->ddir == DDIR_READ) { |
3c3168e9 | 115 | io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
a391d73d | 116 | if (o->hipri) |
3c3168e9 | 117 | iocb->u.c.flags |= IOCB_FLAG_HIPRI; |
a391d73d | 118 | } else if (io_u->ddir == DDIR_WRITE) { |
3c3168e9 | 119 | io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
a391d73d | 120 | if (o->hipri) |
3c3168e9 | 121 | iocb->u.c.flags |= IOCB_FLAG_HIPRI; |
a391d73d | 122 | } else if (ddir_sync(io_u->ddir)) |
3c3168e9 | 123 | io_prep_fsync(iocb, f->fd); |
2866c82d JA |
124 | |
125 | return 0; | |
126 | } | |
127 | ||
128 | static struct io_u *fio_libaio_event(struct thread_data *td, int event) | |
129 | { | |
565e784d | 130 | struct libaio_data *ld = td->io_ops_data; |
3c3168e9 | 131 | struct libaio_options *o = td->eo; |
f423479d JA |
132 | struct io_event *ev; |
133 | struct io_u *io_u; | |
2866c82d | 134 | |
f423479d | 135 | ev = ld->aio_events + event; |
3c3168e9 JA |
136 | if (o->useriocb) { |
137 | int index = (int) (uintptr_t) ev->obj; | |
138 | io_u = ld->io_u_index[index]; | |
139 | } else | |
140 | io_u = container_of(ev->obj, struct io_u, iocb); | |
f423479d JA |
141 | |
142 | if (ev->res != io_u->xfer_buflen) { | |
143 | if (ev->res > io_u->xfer_buflen) | |
144 | io_u->error = -ev->res; | |
145 | else | |
146 | io_u->resid = io_u->xfer_buflen - ev->res; | |
147 | } else | |
148 | io_u->error = 0; | |
149 | ||
150 | return io_u; | |
2866c82d JA |
151 | } |
152 | ||
675012f0 DE |
153 | struct aio_ring { |
154 | unsigned id; /** kernel internal index number */ | |
155 | unsigned nr; /** number of io_events */ | |
156 | unsigned head; | |
157 | unsigned tail; | |
c44b1ff5 | 158 | |
675012f0 DE |
159 | unsigned magic; |
160 | unsigned compat_features; | |
161 | unsigned incompat_features; | |
162 | unsigned header_length; /** size of aio_ring */ | |
163 | ||
164 | struct io_event events[0]; | |
165 | }; | |
166 | ||
167 | #define AIO_RING_MAGIC 0xa10a10a1 | |
168 | ||
169 | static int user_io_getevents(io_context_t aio_ctx, unsigned int max, | |
c44b1ff5 | 170 | struct io_event *events) |
675012f0 DE |
171 | { |
172 | long i = 0; | |
173 | unsigned head; | |
c44b1ff5 | 174 | struct aio_ring *ring = (struct aio_ring*) aio_ctx; |
675012f0 DE |
175 | |
176 | while (i < max) { | |
177 | head = ring->head; | |
178 | ||
179 | if (head == ring->tail) { | |
180 | /* There are no more completions */ | |
181 | break; | |
182 | } else { | |
183 | /* There is another completion to reap */ | |
184 | events[i] = ring->events[head]; | |
185 | read_barrier(); | |
c44b1ff5 | 186 | ring->head = (head + 1) % ring->nr; |
675012f0 DE |
187 | i++; |
188 | } | |
189 | } | |
190 | ||
191 | return i; | |
192 | } | |
193 | ||
e7d2e616 | 194 | static int fio_libaio_getevents(struct thread_data *td, unsigned int min, |
1f440ece | 195 | unsigned int max, const struct timespec *t) |
2866c82d | 196 | { |
565e784d | 197 | struct libaio_data *ld = td->io_ops_data; |
de890a1e | 198 | struct libaio_options *o = td->eo; |
82407585 | 199 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
1f440ece | 200 | struct timespec __lt, *lt = NULL; |
0b7fdba7 | 201 | int r, events = 0; |
2866c82d | 202 | |
1f440ece JA |
203 | if (t) { |
204 | __lt = *t; | |
205 | lt = &__lt; | |
206 | } | |
207 | ||
2866c82d | 208 | do { |
de890a1e | 209 | if (o->userspace_reap == 1 |
675012f0 DE |
210 | && actual_min == 0 |
211 | && ((struct aio_ring *)(ld->aio_ctx))->magic | |
212 | == AIO_RING_MAGIC) { | |
213 | r = user_io_getevents(ld->aio_ctx, max, | |
214 | ld->aio_events + events); | |
215 | } else { | |
216 | r = io_getevents(ld->aio_ctx, actual_min, | |
1f440ece | 217 | max, ld->aio_events + events, lt); |
675012f0 | 218 | } |
3441a52d | 219 | if (r > 0) |
0b7fdba7 | 220 | events += r; |
3441a52d | 221 | else if ((min && r == 0) || r == -EAGAIN) { |
0fc2e103 | 222 | fio_libaio_commit(td); |
6347e43d JA |
223 | if (actual_min) |
224 | usleep(10); | |
0fc2e103 | 225 | } else if (r != -EINTR) |
a31dc2dd | 226 | break; |
0b7fdba7 DE |
227 | } while (events < min); |
228 | ||
229 | return r < 0 ? r : events; | |
2866c82d JA |
230 | } |
231 | ||
2e4ef4fb JA |
232 | static enum fio_q_status fio_libaio_queue(struct thread_data *td, |
233 | struct io_u *io_u) | |
2866c82d | 234 | { |
565e784d | 235 | struct libaio_data *ld = td->io_ops_data; |
3c3168e9 | 236 | struct libaio_options *o = td->eo; |
2866c82d | 237 | |
7101d9c2 JA |
238 | fio_ro_check(td, io_u); |
239 | ||
acd45e02 | 240 | if (ld->queued == td->o.iodepth) |
755200a3 JA |
241 | return FIO_Q_BUSY; |
242 | ||
243 | /* | |
244 | * fsync is tricky, since it can fail and we need to do it | |
245 | * serialized with other io. the reason is that linux doesn't | |
246 | * support aio fsync yet. So return busy for the case where we | |
247 | * have pending io, to let fio complete those first. | |
248 | */ | |
f011531e | 249 | if (ddir_sync(io_u->ddir)) { |
acd45e02 | 250 | if (ld->queued) |
755200a3 | 251 | return FIO_Q_BUSY; |
5f9099ea | 252 | |
f011531e | 253 | do_io_u_sync(td, io_u); |
755200a3 JA |
254 | return FIO_Q_COMPLETED; |
255 | } | |
256 | ||
a5f3027c | 257 | if (io_u->ddir == DDIR_TRIM) { |
acd45e02 | 258 | if (ld->queued) |
a5f3027c JA |
259 | return FIO_Q_BUSY; |
260 | ||
261 | do_io_u_trim(td, io_u); | |
c0681c9d VF |
262 | io_u_mark_submit(td, 1); |
263 | io_u_mark_complete(td, 1); | |
a5f3027c JA |
264 | return FIO_Q_COMPLETED; |
265 | } | |
266 | ||
3c3168e9 JA |
267 | if (o->useriocb) |
268 | ld->iocbs[ld->head] = (struct iocb *) (uintptr_t) io_u->index; | |
269 | else | |
270 | ld->iocbs[ld->head] = &io_u->iocb; | |
271 | ||
acd45e02 JA |
272 | ld->io_us[ld->head] = io_u; |
273 | ring_inc(ld, &ld->head, 1); | |
274 | ld->queued++; | |
755200a3 JA |
275 | return FIO_Q_QUEUED; |
276 | } | |
277 | ||
7e77dd02 JA |
278 | static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us, |
279 | unsigned int nr) | |
280 | { | |
8b6a404c | 281 | struct timespec now; |
7e77dd02 JA |
282 | unsigned int i; |
283 | ||
12d9d841 JA |
284 | if (!fio_fill_issue_time(td)) |
285 | return; | |
286 | ||
7e77dd02 JA |
287 | fio_gettime(&now, NULL); |
288 | ||
289 | for (i = 0; i < nr; i++) { | |
290 | struct io_u *io_u = io_us[i]; | |
291 | ||
292 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
293 | io_u_queued(td, io_u); | |
294 | } | |
295 | } | |
296 | ||
755200a3 JA |
297 | static int fio_libaio_commit(struct thread_data *td) |
298 | { | |
565e784d | 299 | struct libaio_data *ld = td->io_ops_data; |
755200a3 | 300 | struct iocb **iocbs; |
7e77dd02 | 301 | struct io_u **io_us; |
8b6a404c | 302 | struct timespec ts; |
a120ca7f | 303 | int ret, wait_start = 0; |
755200a3 | 304 | |
acd45e02 | 305 | if (!ld->queued) |
755200a3 JA |
306 | return 0; |
307 | ||
2866c82d | 308 | do { |
acd45e02 JA |
309 | long nr = ld->queued; |
310 | ||
311 | nr = min((unsigned int) nr, ld->entries - ld->tail); | |
312 | io_us = ld->io_us + ld->tail; | |
313 | iocbs = ld->iocbs + ld->tail; | |
314 | ||
315 | ret = io_submit(ld->aio_ctx, nr, iocbs); | |
5e00c2c4 | 316 | if (ret > 0) { |
7e77dd02 | 317 | fio_libaio_queued(td, io_us, ret); |
838bc709 | 318 | io_u_mark_submit(td, ret); |
acd45e02 JA |
319 | |
320 | ld->queued -= ret; | |
321 | ring_inc(ld, &ld->tail, ret); | |
5e00c2c4 | 322 | ret = 0; |
e3b4e568 | 323 | wait_start = 0; |
acd45e02 | 324 | } else if (ret == -EINTR || !ret) { |
838bc709 JA |
325 | if (!ret) |
326 | io_u_mark_submit(td, ret); | |
e3b4e568 | 327 | wait_start = 0; |
2866c82d | 328 | continue; |
acd45e02 JA |
329 | } else if (ret == -EAGAIN) { |
330 | /* | |
331 | * If we get EAGAIN, we should break out without | |
332 | * error and let the upper layer reap some | |
a120ca7f JA |
333 | * events for us. If we have no queued IO, we |
334 | * must loop here. If we loop for more than 30s, | |
335 | * just error out, something must be buggy in the | |
336 | * IO path. | |
acd45e02 | 337 | */ |
a120ca7f JA |
338 | if (ld->queued) { |
339 | ret = 0; | |
340 | break; | |
341 | } | |
342 | if (!wait_start) { | |
8b6a404c | 343 | fio_gettime(&ts, NULL); |
d36b072d | 344 | wait_start = 1; |
8b6a404c | 345 | } else if (mtime_since_now(&ts) > 30000) { |
a120ca7f JA |
346 | log_err("fio: aio appears to be stalled, giving up\n"); |
347 | break; | |
348 | } | |
349 | usleep(1); | |
350 | continue; | |
a31dc2dd JA |
351 | } else if (ret == -ENOMEM) { |
352 | /* | |
353 | * If we get -ENOMEM, reap events if we can. If | |
354 | * we cannot, treat it as a fatal event since there's | |
355 | * nothing we can do about it. | |
356 | */ | |
357 | if (ld->queued) | |
358 | ret = 0; | |
359 | break; | |
838bc709 | 360 | } else |
2866c82d | 361 | break; |
2c3a4ae9 | 362 | } while (ld->queued); |
2866c82d | 363 | |
36167d82 | 364 | return ret; |
2866c82d JA |
365 | } |
366 | ||
367 | static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u) | |
368 | { | |
565e784d | 369 | struct libaio_data *ld = td->io_ops_data; |
2866c82d JA |
370 | |
371 | return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events); | |
372 | } | |
373 | ||
374 | static void fio_libaio_cleanup(struct thread_data *td) | |
375 | { | |
565e784d | 376 | struct libaio_data *ld = td->io_ops_data; |
2866c82d JA |
377 | |
378 | if (ld) { | |
f24c2649 JA |
379 | /* |
380 | * Work-around to avoid huge RCU stalls at exit time. If we | |
381 | * don't do this here, then it'll be torn down by exit_aio(). | |
382 | * But for that case we can parallellize the freeing, thus | |
383 | * speeding it up a lot. | |
384 | */ | |
385 | if (!(td->flags & TD_F_CHILD)) | |
386 | io_destroy(ld->aio_ctx); | |
7e77dd02 JA |
387 | free(ld->aio_events); |
388 | free(ld->iocbs); | |
389 | free(ld->io_us); | |
3c3168e9 JA |
390 | if (ld->user_iocbs) |
391 | free(ld->user_iocbs); | |
2866c82d | 392 | free(ld); |
2866c82d JA |
393 | } |
394 | } | |
395 | ||
ebec344d | 396 | static int fio_libaio_queue_init(struct libaio_data *ld, unsigned int depth, |
3c3168e9 | 397 | bool hipri, bool useriocb) |
ebec344d JA |
398 | { |
399 | #ifdef __NR_sys_io_setup2 | |
a1b006fe | 400 | int flags = 0; |
ebec344d JA |
401 | |
402 | if (hipri) | |
3c3168e9 JA |
403 | flags |= IOCTX_FLAG_IOPOLL; |
404 | if (useriocb) | |
405 | flags |= IOCTX_FLAG_USERIOCB; | |
ebec344d | 406 | |
3c3168e9 | 407 | return syscall(__NR_sys_io_setup2, depth, flags, ld->user_iocbs, &ld->aio_ctx); |
ebec344d JA |
408 | #else |
409 | if (hipri) { | |
410 | log_err("fio: polled aio not available on your platform\n"); | |
411 | return 1; | |
412 | } | |
3c3168e9 JA |
413 | if (useriocb) { |
414 | log_err("fio: user mapped iocbs not available on your platform\n"); | |
415 | return 1; | |
416 | } | |
417 | ||
ebec344d JA |
418 | return io_queue_init(depth, &ld->aio_ctx); |
419 | #endif | |
420 | } | |
421 | ||
2866c82d JA |
422 | static int fio_libaio_init(struct thread_data *td) |
423 | { | |
c90e1017 | 424 | struct libaio_options *o = td->eo; |
acd45e02 | 425 | struct libaio_data *ld; |
c90e1017 | 426 | int err = 0; |
2866c82d | 427 | |
acd45e02 | 428 | ld = calloc(1, sizeof(*ld)); |
c1db2dce | 429 | |
3c3168e9 JA |
430 | if (o->useriocb) { |
431 | size_t size; | |
432 | void *p; | |
433 | ||
434 | ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *)); | |
435 | size = td->o.iodepth * sizeof(struct iocb); | |
436 | if (posix_memalign(&p, page_size, size)) { | |
437 | log_err("fio: libaio iocb allocation failure\n"); | |
438 | free(ld); | |
439 | return 1; | |
440 | } | |
441 | ld->user_iocbs = p; | |
442 | } | |
443 | ||
c90e1017 JA |
444 | /* |
445 | * First try passing in 0 for queue depth, since we don't | |
446 | * care about the user ring. If that fails, the kernel is too old | |
447 | * and we need the right depth. | |
448 | */ | |
3c3168e9 | 449 | err = fio_libaio_queue_init(ld, td->o.iodepth, o->hipri, o->useriocb); |
c1db2dce JA |
450 | if (err) { |
451 | td_verror(td, -err, "io_queue_init"); | |
75de55ac | 452 | log_err("fio: check /proc/sys/fs/aio-max-nr\n"); |
3c3168e9 JA |
453 | if (ld->user_iocbs) |
454 | free(ld->user_iocbs); | |
cb781c75 | 455 | free(ld); |
2866c82d JA |
456 | return 1; |
457 | } | |
458 | ||
acd45e02 JA |
459 | ld->entries = td->o.iodepth; |
460 | ld->is_pow2 = is_power_of_2(ld->entries); | |
461 | ld->aio_events = calloc(ld->entries, sizeof(struct io_event)); | |
462 | ld->iocbs = calloc(ld->entries, sizeof(struct iocb *)); | |
463 | ld->io_us = calloc(ld->entries, sizeof(struct io_u *)); | |
755200a3 | 464 | |
565e784d | 465 | td->io_ops_data = ld; |
2866c82d JA |
466 | return 0; |
467 | } | |
468 | ||
3c3168e9 JA |
469 | static int fio_libaio_io_u_init(struct thread_data *td, struct io_u *io_u) |
470 | { | |
471 | struct libaio_options *o = td->eo; | |
472 | ||
473 | if (o->useriocb) { | |
474 | struct libaio_data *ld = td->io_ops_data; | |
475 | ||
476 | ld->io_u_index[io_u->index] = io_u; | |
477 | } | |
478 | ||
479 | return 0; | |
480 | } | |
481 | ||
5f350952 | 482 | static struct ioengine_ops ioengine = { |
de890a1e SL |
483 | .name = "libaio", |
484 | .version = FIO_IOOPS_VERSION, | |
485 | .init = fio_libaio_init, | |
3c3168e9 | 486 | .io_u_init = fio_libaio_io_u_init, |
de890a1e SL |
487 | .prep = fio_libaio_prep, |
488 | .queue = fio_libaio_queue, | |
489 | .commit = fio_libaio_commit, | |
490 | .cancel = fio_libaio_cancel, | |
491 | .getevents = fio_libaio_getevents, | |
492 | .event = fio_libaio_event, | |
493 | .cleanup = fio_libaio_cleanup, | |
494 | .open_file = generic_open_file, | |
495 | .close_file = generic_close_file, | |
496 | .get_file_size = generic_get_file_size, | |
497 | .options = options, | |
498 | .option_struct_size = sizeof(struct libaio_options), | |
2866c82d | 499 | }; |
34cfcdaf | 500 | |
5f350952 JA |
501 | static void fio_init fio_libaio_register(void) |
502 | { | |
503 | register_ioengine(&ioengine); | |
504 | } | |
505 | ||
506 | static void fio_exit fio_libaio_unregister(void) | |
507 | { | |
508 | unregister_ioengine(&ioengine); | |
509 | } |