libaio: pass in 0 for queue depth if we can
[fio.git] / ioengines.c
CommitLineData
ebac4655
JA
1/*
2 * The io parts of the fio tool, includes workers for sync and mmap'ed
3 * io, as well as both posix and linux libaio support.
4 *
5 * sync io is implemented on top of aio.
6 *
7 * This is not really specific to fio, if the get_io_u/put_io_u and
8 * structures was pulled into this as well it would be a perfectly
9 * generic io engine that could be used for other projects.
10 *
11 */
12#include <stdio.h>
13#include <stdlib.h>
14#include <unistd.h>
5c4e1dbc 15#include <string.h>
2866c82d 16#include <dlfcn.h>
ecc314ba 17#include <fcntl.h>
0c6e7517 18#include <assert.h>
8c16d840 19
ebac4655 20#include "fio.h"
7c9b1bce 21#include "diskutil.h"
ebac4655 22
01743ee1 23static FLIST_HEAD(engine_list);
5f350952 24
8c16d840
JA
25static int check_engine_ops(struct ioengine_ops *ops)
26{
5f350952 27 if (ops->version != FIO_IOOPS_VERSION) {
5ec10eaa
JA
28 log_err("bad ioops version %d (want %d)\n", ops->version,
29 FIO_IOOPS_VERSION);
5f350952
JA
30 return 1;
31 }
32
36167d82
JA
33 if (!ops->queue) {
34 log_err("%s: no queue handler\n", ops->name);
35 return 1;
36 }
37
38 /*
39 * sync engines only need a ->queue()
40 */
41 if (ops->flags & FIO_SYNCIO)
42 return 0;
5ec10eaa 43
8c16d840 44 if (!ops->event) {
36167d82 45 log_err("%s: no event handler\n", ops->name);
8c16d840
JA
46 return 1;
47 }
48 if (!ops->getevents) {
36167d82 49 log_err("%s: no getevents handler\n", ops->name);
8c16d840
JA
50 return 1;
51 }
52 if (!ops->queue) {
36167d82 53 log_err("%s: no queue handler\n", ops->name);
8c16d840
JA
54 return 1;
55 }
5ec10eaa 56
8c16d840
JA
57 return 0;
58}
59
5f350952 60void unregister_ioengine(struct ioengine_ops *ops)
ebac4655 61{
ee56ad50 62 dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
01743ee1
JA
63 flist_del(&ops->list);
64 INIT_FLIST_HEAD(&ops->list);
5f350952
JA
65}
66
b2fdda43 67void register_ioengine(struct ioengine_ops *ops)
5f350952 68{
ee56ad50 69 dprint(FD_IO, "ioengine %s registered\n", ops->name);
01743ee1
JA
70 INIT_FLIST_HEAD(&ops->list);
71 flist_add_tail(&ops->list, &engine_list);
5f350952
JA
72}
73
74static struct ioengine_ops *find_ioengine(const char *name)
75{
76 struct ioengine_ops *ops;
01743ee1 77 struct flist_head *entry;
ebac4655 78
01743ee1
JA
79 flist_for_each(entry, &engine_list) {
80 ops = flist_entry(entry, struct ioengine_ops, list);
bc5b77a8 81 if (!strcmp(name, ops->name))
5f350952
JA
82 return ops;
83 }
84
85 return NULL;
86}
87
88static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
89 const char *engine_lib)
90{
91 struct ioengine_ops *ops;
92 void *dlhandle;
93
ee56ad50
JA
94 dprint(FD_IO, "dload engine %s\n", engine_lib);
95
2866c82d
JA
96 dlerror();
97 dlhandle = dlopen(engine_lib, RTLD_LAZY);
d4dbaaa8 98 if (!dlhandle) {
e1161c32 99 td_vmsg(td, -1, dlerror(), "dlopen");
d4dbaaa8
JA
100 return NULL;
101 }
8756e4d4 102
da51c050
JA
103 /*
104 * Unlike the included modules, external engines should have a
105 * non-static ioengine structure that we can reference.
106 */
2866c82d 107 ops = dlsym(dlhandle, "ioengine");
d4dbaaa8 108 if (!ops) {
e1161c32 109 td_vmsg(td, -1, dlerror(), "dlsym");
d4dbaaa8
JA
110 dlclose(dlhandle);
111 return NULL;
112 }
8756e4d4 113
5f350952
JA
114 ops->dlhandle = dlhandle;
115 return ops;
116}
117
118struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
119{
120 struct ioengine_ops *ops, *ret;
121 char engine[16];
122
ee56ad50
JA
123 dprint(FD_IO, "load ioengine %s\n", name);
124
5f350952
JA
125 strncpy(engine, name, sizeof(engine) - 1);
126
127 /*
128 * linux libaio has alias names, so convert to what we want
129 */
130 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
131 strcpy(engine, "libaio");
132
133 ops = find_ioengine(engine);
134 if (!ops)
135 ops = dlopen_ioengine(td, name);
136
137 if (!ops) {
138 log_err("fio: engine %s not loadable\n", name);
b902ceb5
JA
139 return NULL;
140 }
141
8c16d840
JA
142 /*
143 * Check that the required methods are there.
144 */
5f350952 145 if (check_engine_ops(ops))
8c16d840 146 return NULL;
8c16d840 147
84585003
JA
148 ret = malloc(sizeof(*ret));
149 memcpy(ret, ops, sizeof(*ret));
150 ret->data = NULL;
84585003
JA
151
152 return ret;
8756e4d4
JA
153}
154
de890a1e
SL
155/*
156 * For cleaning up an ioengine which never made it to init().
157 */
158void free_ioengine(struct thread_data *td)
8756e4d4 159{
de890a1e 160 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
ee56ad50 161
de890a1e
SL
162 if (td->eo && td->io_ops->options) {
163 options_free(td->io_ops->options, td->eo);
164 free(td->eo);
165 td->eo = NULL;
2992b059 166 }
b990b5c0 167
5f350952
JA
168 if (td->io_ops->dlhandle)
169 dlclose(td->io_ops->dlhandle);
170
84585003
JA
171 free(td->io_ops);
172 td->io_ops = NULL;
b990b5c0 173}
10ba535a 174
de890a1e
SL
175void close_ioengine(struct thread_data *td)
176{
177 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
178
179 if (td->io_ops->cleanup) {
180 td->io_ops->cleanup(td);
181 td->io_ops->data = NULL;
182 }
183
184 free_ioengine(td);
185}
186
10ba535a
JA
187int td_io_prep(struct thread_data *td, struct io_u *io_u)
188{
ee56ad50 189 dprint_io_u(io_u, "prep");
7101d9c2
JA
190 fio_ro_check(td, io_u);
191
4d4e80f2 192 lock_file(td, io_u->file, io_u->ddir);
b2bd2bd9 193
2ba1c290
JA
194 if (td->io_ops->prep) {
195 int ret = td->io_ops->prep(td, io_u);
196
197 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
b2bd2bd9 198 if (ret)
4d4e80f2 199 unlock_file(td, io_u->file);
2ba1c290
JA
200 return ret;
201 }
10ba535a
JA
202
203 return 0;
204}
205
e7d2e616 206int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
10ba535a
JA
207 struct timespec *t)
208{
ee56ad50 209 int r = 0;
face81b2 210
a05d62b2
YR
211 /*
212 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
213 * server side gets a message from the client
214 * side that the task is finished, and
215 * td->done is set to 1 after td_io_commit(). In this case,
216 * there is no need to reap complete event in server side.
217 */
218 if (td->done)
219 return 0;
220
ee56ad50
JA
221 if (min > 0 && td->io_ops->commit) {
222 r = td->io_ops->commit(td);
face81b2 223 if (r < 0)
ee56ad50 224 goto out;
face81b2 225 }
4950421a
JA
226 if (max > td->cur_depth)
227 max = td->cur_depth;
228 if (min > max)
229 max = min;
36167d82 230
ee56ad50 231 r = 0;
4950421a 232 if (max && td->io_ops->getevents)
ee56ad50
JA
233 r = td->io_ops->getevents(td, min, max, t);
234out:
422f9e4b
RM
235 if (r >= 0) {
236 /*
237 * Reflect that our submitted requests were retrieved with
238 * whatever OS async calls are in the underlying engine.
239 */
240 td->io_u_in_flight -= r;
838bc709 241 io_u_mark_complete(td, r);
422f9e4b 242 } else
7c639b14 243 td_verror(td, r, "get_events");
f3e11d05 244
ee56ad50
JA
245 dprint(FD_IO, "getevents: %d\n", r);
246 return r;
10ba535a
JA
247}
248
249int td_io_queue(struct thread_data *td, struct io_u *io_u)
250{
7e77dd02
JA
251 int ret;
252
ee56ad50 253 dprint_io_u(io_u, "queue");
7101d9c2
JA
254 fio_ro_check(td, io_u);
255
0c6e7517
JA
256 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
257 io_u->flags |= IO_U_F_FLIGHT;
258
d6aed795 259 assert(fio_file_open(io_u->file));
3d7b485f 260
11786802
JA
261 io_u->error = 0;
262 io_u->resid = 0;
263
433afcb4 264 if (td->io_ops->flags & FIO_SYNCIO) {
12d9d841 265 if (fio_fill_issue_time(td))
9520ebb9 266 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
267
268 /*
269 * only used for iolog
270 */
271 if (td->o.read_iolog_file)
272 memcpy(&td->last_issue, &io_u->issue_time,
5ec10eaa 273 sizeof(struct timeval));
433afcb4
JA
274 }
275
ff58fced 276 if (ddir_rw(io_u->ddir))
755200a3
JA
277 td->io_issues[io_u->ddir]++;
278
7e77dd02 279 ret = td->io_ops->queue(td, io_u);
5aeb77df 280
4d4e80f2 281 unlock_file(td, io_u->file);
b2bd2bd9 282
cb211682
JA
283 /*
284 * Add warning for O_DIRECT so that users have an easier time
285 * spotting potentially bad alignment. If this triggers for the first
286 * IO, then it's likely an alignment problem or because the host fs
287 * does not support O_DIRECT
288 */
ff58fced 289 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
cb211682 290 td->o.odirect) {
214ac7e0 291
cb211682
JA
292 log_info("fio: first direct IO errored. File system may not "
293 "support direct IO, or iomem_align= is bad.\n");
294 }
295
838bc709
JA
296 if (!td->io_ops->commit) {
297 io_u_mark_submit(td, 1);
298 io_u_mark_complete(td, 1);
299 }
300
d8005759 301 if (ret == FIO_Q_COMPLETED) {
ff58fced 302 if (ddir_rw(io_u->ddir)) {
d8005759
JA
303 io_u_mark_depth(td, 1);
304 td->ts.total_io_u[io_u->ddir]++;
0d29de83
JA
305 } else if (io_u->ddir == DDIR_TRIM)
306 td->ts.total_io_u[2]++;
d8005759 307 } else if (ret == FIO_Q_QUEUED) {
eb7c8ae2
JA
308 int r;
309
ff58fced 310 if (ddir_rw(io_u->ddir)) {
d8005759
JA
311 td->io_u_queued++;
312 td->ts.total_io_u[io_u->ddir]++;
313 }
314
315 if (td->io_u_queued >= td->o.iodepth_batch) {
eb7c8ae2
JA
316 r = td_io_commit(td);
317 if (r < 0)
318 return r;
319 }
320 }
cb5ab512 321
433afcb4 322 if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
12d9d841 323 if (fio_fill_issue_time(td))
9520ebb9 324 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
325
326 /*
327 * only used for iolog
328 */
329 if (td->o.read_iolog_file)
330 memcpy(&td->last_issue, &io_u->issue_time,
331 sizeof(struct timeval));
433afcb4
JA
332 }
333
7e77dd02 334 return ret;
10ba535a 335}
8c16d840
JA
336
337int td_io_init(struct thread_data *td)
338{
eeb12160 339 int ret = 0;
8c16d840 340
eeb12160
JA
341 if (td->io_ops->init) {
342 ret = td->io_ops->init(td);
5ec10eaa
JA
343 if (ret && td->o.iodepth > 1) {
344 log_err("fio: io engine init failed. Perhaps try"
345 " reducing io depth?\n");
346 }
7c973896
JA
347 if (!td->error)
348 td->error = ret;
eeb12160
JA
349 }
350
351 return ret;
8c16d840 352}
755200a3
JA
353
354int td_io_commit(struct thread_data *td)
355{
f3e11d05
JA
356 int ret;
357
ee56ad50
JA
358 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
359
d8005759 360 if (!td->cur_depth || !td->io_u_queued)
e1161c32 361 return 0;
cb5ab512 362
422f9e4b 363 io_u_mark_depth(td, td->io_u_queued);
d8005759 364
f3e11d05
JA
365 if (td->io_ops->commit) {
366 ret = td->io_ops->commit(td);
367 if (ret)
368 td_verror(td, -ret, "io commit");
369 }
422f9e4b
RM
370
371 /*
372 * Reflect that events were submitted as async IO requests.
373 */
374 td->io_u_in_flight += td->io_u_queued;
375 td->io_u_queued = 0;
755200a3
JA
376
377 return 0;
378}
b5af8293
JA
379
380int td_io_open_file(struct thread_data *td, struct fio_file *f)
381{
22a57ba8
JA
382 assert(!fio_file_open(f));
383 assert(f->fd == -1);
384
413d6693
JA
385 if (td->io_ops->open_file(td, f)) {
386 if (td->error == EINVAL && td->o.odirect)
387 log_err("fio: destination does not support O_DIRECT\n");
5ec10eaa
JA
388 if (td->error == EMFILE) {
389 log_err("fio: try reducing/setting openfiles (failed"
390 " at %u of %u)\n", td->nr_open_files,
391 td->o.nr_files);
392 }
413d6693 393
22a57ba8
JA
394 assert(f->fd == -1);
395 assert(!fio_file_open(f));
413d6693
JA
396 return 1;
397 }
398
d5707a35 399 fio_file_reset(f);
d6aed795
JA
400 fio_file_set_open(f);
401 fio_file_clear_closing(f);
c97bd0fa 402 disk_util_inc(f->du);
d5707a35
JA
403
404 td->nr_open_files++;
405 get_file(f);
406
66159828
JA
407 if (f->filetype == FIO_TYPE_PIPE) {
408 if (td_random(td)) {
409 log_err("fio: can't seek on pipes (no random io)\n");
410 goto err;
411 }
412 }
413
413d6693
JA
414 if (td->io_ops->flags & FIO_DISKLESSIO)
415 goto done;
416
417 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
418 goto err;
419
66159828
JA
420 if (td->o.fadvise_hint &&
421 (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
413d6693
JA
422 int flags;
423
424 if (td_random(td))
425 flags = POSIX_FADV_RANDOM;
426 else
427 flags = POSIX_FADV_SEQUENTIAL;
428
ecc314ba 429 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
413d6693
JA
430 td_verror(td, errno, "fadvise");
431 goto err;
432 }
7bb48f84 433 }
a978ba68 434
e116f2b9
JA
435#ifdef FIO_OS_DIRECTIO
436 /*
437 * Some OS's have a distinct call to mark the file non-buffered,
438 * instead of using O_DIRECT (Solaris)
439 */
440 if (td->o.odirect) {
441 int ret = fio_set_odirect(f->fd);
442
443 if (ret) {
444 td_verror(td, ret, "fio_set_odirect");
78e51c72 445 log_err("fio: the file system does not seem to support direct IO\n");
e116f2b9
JA
446 goto err;
447 }
448 }
449#endif
450
413d6693 451done:
f29b25a3 452 log_file(td, f, FIO_LOG_OPEN_FILE);
413d6693
JA
453 return 0;
454err:
c97bd0fa 455 disk_util_dec(f->du);
b284075a
JA
456 if (td->io_ops->close_file)
457 td->io_ops->close_file(td, f);
7bb48f84 458 return 1;
b5af8293
JA
459}
460
6977bcd0 461int td_io_close_file(struct thread_data *td, struct fio_file *f)
b5af8293 462{
d6aed795 463 if (!fio_file_closing(f))
f29b25a3
JA
464 log_file(td, f, FIO_LOG_CLOSE_FILE);
465
0ad920e7
JA
466 /*
467 * mark as closing, do real close when last io on it has completed
468 */
d6aed795 469 fio_file_set_closing(f);
0ad920e7 470
c97bd0fa 471 disk_util_dec(f->du);
4d4e80f2 472 unlock_file_all(td, f);
29c1349f 473
6977bcd0 474 return put_file(td, f);
b5af8293 475}
df9c26b1
JA
476
477int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
478{
479 if (!td->io_ops->get_file_size)
480 return 0;
481
482 return td->io_ops->get_file_size(td, f);
483}
44f29692 484
0a28ecda 485static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
44f29692
JA
486{
487 off64_t offset, nbytes;
488
489 offset = f->first_write;
490 nbytes = f->last_write - f->first_write;
491
3843deb3
JA
492 if (!nbytes)
493 return 0;
44f29692 494
3843deb3 495 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
44f29692 496}
0a28ecda
JA
497
498int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
499{
500 int ret;
501
502 if (io_u->ddir == DDIR_SYNC) {
503 ret = fsync(io_u->file->fd);
504 } else if (io_u->ddir == DDIR_DATASYNC) {
505#ifdef FIO_HAVE_FDATASYNC
506 ret = fdatasync(io_u->file->fd);
507#else
508 ret = io_u->xfer_buflen;
509 io_u->error = EINVAL;
510#endif
511 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
512 ret = do_sync_file_range(td, io_u->file);
513 else {
514 ret = io_u->xfer_buflen;
515 io_u->error = EINVAL;
516 }
517
cb849a79
JA
518 if (ret < 0)
519 io_u->error = errno;
520
0a28ecda
JA
521 return ret;
522}
a5f3027c
JA
523
524int do_io_u_trim(struct thread_data *td, struct io_u *io_u)
525{
526#ifndef FIO_HAVE_TRIM
527 io_u->error = EINVAL;
ff58fced 528 return 0;
a5f3027c
JA
529#else
530 struct fio_file *f = io_u->file;
531 int ret;
532
c6404a44 533 ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
a5f3027c 534 if (!ret)
ff58fced 535 return io_u->xfer_buflen;;
a5f3027c 536
ff58fced
JA
537 io_u->error = ret;
538 return 0;
a5f3027c
JA
539#endif
540}
de890a1e
SL
541
542int fio_show_ioengine_help(const char *engine)
543{
544 struct flist_head *entry;
545 struct thread_data td;
546 char *sep;
547 int ret = 1;
548
549 if (!engine || !*engine) {
550 log_info("Available IO engines:\n");
551 flist_for_each(entry, &engine_list) {
552 td.io_ops = flist_entry(entry, struct ioengine_ops,
553 list);
554 log_info("\t%s\n", td.io_ops->name);
555 }
556 return 0;
557 }
558 sep = strchr(engine, ',');
559 if (sep) {
560 *sep = 0;
561 sep++;
562 }
563
564 memset(&td, 0, sizeof(td));
565
566 td.io_ops = load_ioengine(&td, engine);
567 if (!td.io_ops) {
568 log_info("IO engine %s not found\n", engine);
569 return 1;
570 }
571
572 if (td.io_ops->options)
573 ret = show_cmd_help(td.io_ops->options, sep);
574 else
575 log_info("IO engine %s has no options\n", td.io_ops->name);
576
577 free_ioengine(&td);
578
579 return ret;
580}