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 | |
0fc2e103 JA |
19 | static int fio_libaio_commit(struct thread_data *td); |
20 | ||
2866c82d JA |
21 | struct libaio_data { |
22 | io_context_t aio_ctx; | |
23 | struct io_event *aio_events; | |
755200a3 | 24 | struct iocb **iocbs; |
7e77dd02 | 25 | struct io_u **io_us; |
acd45e02 | 26 | |
3c3168e9 JA |
27 | struct io_u **io_u_index; |
28 | ||
acd45e02 JA |
29 | /* |
30 | * Basic ring buffer. 'head' is incremented in _queue(), and | |
31 | * 'tail' is incremented in _commit(). We keep 'queued' so | |
32 | * that we know if the ring is full or empty, when | |
33 | * 'head' == 'tail'. 'entries' is the ring size, and | |
34 | * 'is_pow2' is just an optimization to use AND instead of | |
35 | * modulus to get the remainder on ring increment. | |
36 | */ | |
37 | int is_pow2; | |
38 | unsigned int entries; | |
39 | unsigned int queued; | |
40 | unsigned int head; | |
41 | unsigned int tail; | |
2866c82d JA |
42 | }; |
43 | ||
de890a1e | 44 | struct libaio_options { |
a1f871c7 | 45 | void *pad; |
de890a1e SL |
46 | unsigned int userspace_reap; |
47 | }; | |
48 | ||
49 | static struct fio_option options[] = { | |
50 | { | |
51 | .name = "userspace_reap", | |
e8b0e958 | 52 | .lname = "Libaio userspace reaping", |
de890a1e SL |
53 | .type = FIO_OPT_STR_SET, |
54 | .off1 = offsetof(struct libaio_options, userspace_reap), | |
55 | .help = "Use alternative user-space reap implementation", | |
e90a0adf | 56 | .category = FIO_OPT_C_ENGINE, |
5a3cd5f3 | 57 | .group = FIO_OPT_G_LIBAIO, |
de890a1e SL |
58 | }, |
59 | { | |
60 | .name = NULL, | |
61 | }, | |
62 | }; | |
63 | ||
acd45e02 JA |
64 | static inline void ring_inc(struct libaio_data *ld, unsigned int *val, |
65 | unsigned int add) | |
66 | { | |
67 | if (ld->is_pow2) | |
68 | *val = (*val + add) & (ld->entries - 1); | |
69 | else | |
70 | *val = (*val + add) % ld->entries; | |
71 | } | |
72 | ||
7a16dd02 | 73 | static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u) |
2866c82d | 74 | { |
53cdc686 | 75 | struct fio_file *f = io_u->file; |
ad4e9298 | 76 | struct iocb *iocb = &io_u->iocb; |
53cdc686 | 77 | |
a391d73d | 78 | if (io_u->ddir == DDIR_READ) { |
702906e9 | 79 | io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
a391d73d | 80 | } else if (io_u->ddir == DDIR_WRITE) { |
702906e9 | 81 | io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); |
a391d73d | 82 | } else if (ddir_sync(io_u->ddir)) |
3c3168e9 | 83 | io_prep_fsync(iocb, f->fd); |
2866c82d JA |
84 | |
85 | return 0; | |
86 | } | |
87 | ||
88 | static struct io_u *fio_libaio_event(struct thread_data *td, int event) | |
89 | { | |
565e784d | 90 | struct libaio_data *ld = td->io_ops_data; |
f423479d JA |
91 | struct io_event *ev; |
92 | struct io_u *io_u; | |
2866c82d | 93 | |
f423479d | 94 | ev = ld->aio_events + event; |
702906e9 | 95 | io_u = container_of(ev->obj, struct io_u, iocb); |
f423479d JA |
96 | |
97 | if (ev->res != io_u->xfer_buflen) { | |
98 | if (ev->res > io_u->xfer_buflen) | |
99 | io_u->error = -ev->res; | |
100 | else | |
101 | io_u->resid = io_u->xfer_buflen - ev->res; | |
102 | } else | |
103 | io_u->error = 0; | |
104 | ||
105 | return io_u; | |
2866c82d JA |
106 | } |
107 | ||
675012f0 DE |
108 | struct aio_ring { |
109 | unsigned id; /** kernel internal index number */ | |
110 | unsigned nr; /** number of io_events */ | |
111 | unsigned head; | |
112 | unsigned tail; | |
c44b1ff5 | 113 | |
675012f0 DE |
114 | unsigned magic; |
115 | unsigned compat_features; | |
116 | unsigned incompat_features; | |
117 | unsigned header_length; /** size of aio_ring */ | |
118 | ||
119 | struct io_event events[0]; | |
120 | }; | |
121 | ||
122 | #define AIO_RING_MAGIC 0xa10a10a1 | |
123 | ||
124 | static int user_io_getevents(io_context_t aio_ctx, unsigned int max, | |
c44b1ff5 | 125 | struct io_event *events) |
675012f0 DE |
126 | { |
127 | long i = 0; | |
128 | unsigned head; | |
c44b1ff5 | 129 | struct aio_ring *ring = (struct aio_ring*) aio_ctx; |
675012f0 DE |
130 | |
131 | while (i < max) { | |
132 | head = ring->head; | |
133 | ||
134 | if (head == ring->tail) { | |
135 | /* There are no more completions */ | |
136 | break; | |
137 | } else { | |
138 | /* There is another completion to reap */ | |
139 | events[i] = ring->events[head]; | |
140 | read_barrier(); | |
c44b1ff5 | 141 | ring->head = (head + 1) % ring->nr; |
675012f0 DE |
142 | i++; |
143 | } | |
144 | } | |
145 | ||
146 | return i; | |
147 | } | |
148 | ||
e7d2e616 | 149 | static int fio_libaio_getevents(struct thread_data *td, unsigned int min, |
1f440ece | 150 | unsigned int max, const struct timespec *t) |
2866c82d | 151 | { |
565e784d | 152 | struct libaio_data *ld = td->io_ops_data; |
de890a1e | 153 | struct libaio_options *o = td->eo; |
82407585 | 154 | unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min; |
1f440ece | 155 | struct timespec __lt, *lt = NULL; |
0b7fdba7 | 156 | int r, events = 0; |
2866c82d | 157 | |
1f440ece JA |
158 | if (t) { |
159 | __lt = *t; | |
160 | lt = &__lt; | |
161 | } | |
162 | ||
2866c82d | 163 | do { |
de890a1e | 164 | if (o->userspace_reap == 1 |
675012f0 DE |
165 | && actual_min == 0 |
166 | && ((struct aio_ring *)(ld->aio_ctx))->magic | |
167 | == AIO_RING_MAGIC) { | |
168 | r = user_io_getevents(ld->aio_ctx, max, | |
169 | ld->aio_events + events); | |
170 | } else { | |
171 | r = io_getevents(ld->aio_ctx, actual_min, | |
1f440ece | 172 | max, ld->aio_events + events, lt); |
675012f0 | 173 | } |
3441a52d | 174 | if (r > 0) |
0b7fdba7 | 175 | events += r; |
3441a52d | 176 | else if ((min && r == 0) || r == -EAGAIN) { |
0fc2e103 | 177 | fio_libaio_commit(td); |
6347e43d JA |
178 | if (actual_min) |
179 | usleep(10); | |
0fc2e103 | 180 | } else if (r != -EINTR) |
a31dc2dd | 181 | break; |
0b7fdba7 DE |
182 | } while (events < min); |
183 | ||
184 | return r < 0 ? r : events; | |
2866c82d JA |
185 | } |
186 | ||
2e4ef4fb JA |
187 | static enum fio_q_status fio_libaio_queue(struct thread_data *td, |
188 | struct io_u *io_u) | |
2866c82d | 189 | { |
565e784d | 190 | struct libaio_data *ld = td->io_ops_data; |
2866c82d | 191 | |
7101d9c2 JA |
192 | fio_ro_check(td, io_u); |
193 | ||
acd45e02 | 194 | if (ld->queued == td->o.iodepth) |
755200a3 JA |
195 | return FIO_Q_BUSY; |
196 | ||
197 | /* | |
198 | * fsync is tricky, since it can fail and we need to do it | |
199 | * serialized with other io. the reason is that linux doesn't | |
200 | * support aio fsync yet. So return busy for the case where we | |
201 | * have pending io, to let fio complete those first. | |
202 | */ | |
f011531e | 203 | if (ddir_sync(io_u->ddir)) { |
acd45e02 | 204 | if (ld->queued) |
755200a3 | 205 | return FIO_Q_BUSY; |
5f9099ea | 206 | |
f011531e | 207 | do_io_u_sync(td, io_u); |
755200a3 JA |
208 | return FIO_Q_COMPLETED; |
209 | } | |
210 | ||
a5f3027c | 211 | if (io_u->ddir == DDIR_TRIM) { |
acd45e02 | 212 | if (ld->queued) |
a5f3027c JA |
213 | return FIO_Q_BUSY; |
214 | ||
215 | do_io_u_trim(td, io_u); | |
c0681c9d VF |
216 | io_u_mark_submit(td, 1); |
217 | io_u_mark_complete(td, 1); | |
a5f3027c JA |
218 | return FIO_Q_COMPLETED; |
219 | } | |
220 | ||
702906e9 | 221 | ld->iocbs[ld->head] = &io_u->iocb; |
acd45e02 JA |
222 | ld->io_us[ld->head] = io_u; |
223 | ring_inc(ld, &ld->head, 1); | |
224 | ld->queued++; | |
755200a3 JA |
225 | return FIO_Q_QUEUED; |
226 | } | |
227 | ||
7e77dd02 JA |
228 | static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us, |
229 | unsigned int nr) | |
230 | { | |
8b6a404c | 231 | struct timespec now; |
7e77dd02 JA |
232 | unsigned int i; |
233 | ||
12d9d841 JA |
234 | if (!fio_fill_issue_time(td)) |
235 | return; | |
236 | ||
7e77dd02 JA |
237 | fio_gettime(&now, NULL); |
238 | ||
239 | for (i = 0; i < nr; i++) { | |
240 | struct io_u *io_u = io_us[i]; | |
241 | ||
242 | memcpy(&io_u->issue_time, &now, sizeof(now)); | |
243 | io_u_queued(td, io_u); | |
244 | } | |
245 | } | |
246 | ||
755200a3 JA |
247 | static int fio_libaio_commit(struct thread_data *td) |
248 | { | |
565e784d | 249 | struct libaio_data *ld = td->io_ops_data; |
755200a3 | 250 | struct iocb **iocbs; |
7e77dd02 | 251 | struct io_u **io_us; |
8b6a404c | 252 | struct timespec ts; |
a120ca7f | 253 | int ret, wait_start = 0; |
755200a3 | 254 | |
acd45e02 | 255 | if (!ld->queued) |
755200a3 JA |
256 | return 0; |
257 | ||
2866c82d | 258 | do { |
acd45e02 JA |
259 | long nr = ld->queued; |
260 | ||
261 | nr = min((unsigned int) nr, ld->entries - ld->tail); | |
262 | io_us = ld->io_us + ld->tail; | |
263 | iocbs = ld->iocbs + ld->tail; | |
264 | ||
265 | ret = io_submit(ld->aio_ctx, nr, iocbs); | |
5e00c2c4 | 266 | if (ret > 0) { |
7e77dd02 | 267 | fio_libaio_queued(td, io_us, ret); |
838bc709 | 268 | io_u_mark_submit(td, ret); |
acd45e02 JA |
269 | |
270 | ld->queued -= ret; | |
271 | ring_inc(ld, &ld->tail, ret); | |
5e00c2c4 | 272 | ret = 0; |
e3b4e568 | 273 | wait_start = 0; |
acd45e02 | 274 | } else if (ret == -EINTR || !ret) { |
838bc709 JA |
275 | if (!ret) |
276 | io_u_mark_submit(td, ret); | |
e3b4e568 | 277 | wait_start = 0; |
2866c82d | 278 | continue; |
acd45e02 JA |
279 | } else if (ret == -EAGAIN) { |
280 | /* | |
281 | * If we get EAGAIN, we should break out without | |
282 | * error and let the upper layer reap some | |
a120ca7f JA |
283 | * events for us. If we have no queued IO, we |
284 | * must loop here. If we loop for more than 30s, | |
285 | * just error out, something must be buggy in the | |
286 | * IO path. | |
acd45e02 | 287 | */ |
a120ca7f JA |
288 | if (ld->queued) { |
289 | ret = 0; | |
290 | break; | |
291 | } | |
292 | if (!wait_start) { | |
8b6a404c | 293 | fio_gettime(&ts, NULL); |
d36b072d | 294 | wait_start = 1; |
8b6a404c | 295 | } else if (mtime_since_now(&ts) > 30000) { |
a120ca7f JA |
296 | log_err("fio: aio appears to be stalled, giving up\n"); |
297 | break; | |
298 | } | |
299 | usleep(1); | |
300 | continue; | |
a31dc2dd JA |
301 | } else if (ret == -ENOMEM) { |
302 | /* | |
303 | * If we get -ENOMEM, reap events if we can. If | |
304 | * we cannot, treat it as a fatal event since there's | |
305 | * nothing we can do about it. | |
306 | */ | |
307 | if (ld->queued) | |
308 | ret = 0; | |
309 | break; | |
838bc709 | 310 | } else |
2866c82d | 311 | break; |
2c3a4ae9 | 312 | } while (ld->queued); |
2866c82d | 313 | |
36167d82 | 314 | return ret; |
2866c82d JA |
315 | } |
316 | ||
317 | static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u) | |
318 | { | |
565e784d | 319 | struct libaio_data *ld = td->io_ops_data; |
2866c82d JA |
320 | |
321 | return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events); | |
322 | } | |
323 | ||
324 | static void fio_libaio_cleanup(struct thread_data *td) | |
325 | { | |
565e784d | 326 | struct libaio_data *ld = td->io_ops_data; |
2866c82d JA |
327 | |
328 | if (ld) { | |
f24c2649 JA |
329 | /* |
330 | * Work-around to avoid huge RCU stalls at exit time. If we | |
331 | * don't do this here, then it'll be torn down by exit_aio(). | |
332 | * But for that case we can parallellize the freeing, thus | |
333 | * speeding it up a lot. | |
334 | */ | |
335 | if (!(td->flags & TD_F_CHILD)) | |
336 | io_destroy(ld->aio_ctx); | |
7e77dd02 JA |
337 | free(ld->aio_events); |
338 | free(ld->iocbs); | |
339 | free(ld->io_us); | |
2866c82d | 340 | free(ld); |
2866c82d JA |
341 | } |
342 | } | |
343 | ||
2041bd34 JA |
344 | static int fio_libaio_post_init(struct thread_data *td) |
345 | { | |
346 | struct libaio_data *ld = td->io_ops_data; | |
ad4e9298 | 347 | int err; |
2041bd34 | 348 | |
ad4e9298 | 349 | err = io_queue_init(td->o.iodepth, &ld->aio_ctx); |
2041bd34 JA |
350 | if (err) { |
351 | td_verror(td, -err, "io_queue_init"); | |
352 | return 1; | |
353 | } | |
354 | ||
355 | return 0; | |
ebec344d JA |
356 | } |
357 | ||
2866c82d JA |
358 | static int fio_libaio_init(struct thread_data *td) |
359 | { | |
acd45e02 | 360 | struct libaio_data *ld; |
2866c82d | 361 | |
acd45e02 | 362 | ld = calloc(1, sizeof(*ld)); |
c1db2dce | 363 | |
acd45e02 JA |
364 | ld->entries = td->o.iodepth; |
365 | ld->is_pow2 = is_power_of_2(ld->entries); | |
366 | ld->aio_events = calloc(ld->entries, sizeof(struct io_event)); | |
367 | ld->iocbs = calloc(ld->entries, sizeof(struct iocb *)); | |
368 | ld->io_us = calloc(ld->entries, sizeof(struct io_u *)); | |
755200a3 | 369 | |
565e784d | 370 | td->io_ops_data = ld; |
2866c82d JA |
371 | return 0; |
372 | } | |
373 | ||
5f350952 | 374 | static struct ioengine_ops ioengine = { |
de890a1e SL |
375 | .name = "libaio", |
376 | .version = FIO_IOOPS_VERSION, | |
04ba61df | 377 | .flags = FIO_ASYNCIO_SYNC_TRIM, |
de890a1e | 378 | .init = fio_libaio_init, |
2041bd34 | 379 | .post_init = fio_libaio_post_init, |
de890a1e SL |
380 | .prep = fio_libaio_prep, |
381 | .queue = fio_libaio_queue, | |
382 | .commit = fio_libaio_commit, | |
383 | .cancel = fio_libaio_cancel, | |
384 | .getevents = fio_libaio_getevents, | |
385 | .event = fio_libaio_event, | |
386 | .cleanup = fio_libaio_cleanup, | |
387 | .open_file = generic_open_file, | |
388 | .close_file = generic_close_file, | |
389 | .get_file_size = generic_get_file_size, | |
390 | .options = options, | |
391 | .option_struct_size = sizeof(struct libaio_options), | |
2866c82d | 392 | }; |
34cfcdaf | 393 | |
5f350952 JA |
394 | static void fio_init fio_libaio_register(void) |
395 | { | |
396 | register_ioengine(&ioengine); | |
397 | } | |
398 | ||
399 | static void fio_exit fio_libaio_unregister(void) | |
400 | { | |
401 | unregister_ioengine(&ioengine); | |
402 | } |