Always update rusage before grabbing stat_mutex
[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 */
0abea0b7
DM
107 ops = dlsym(dlhandle, engine_lib);
108 if (!ops)
109 ops = dlsym(dlhandle, "ioengine");
a8075704
DG
110
111 /*
112 * For some external engines (like C++ ones) it is not that trivial
113 * to provide a non-static ionengine structure that we can reference.
114 * Instead we call a method which allocates the required ioengine
115 * structure.
116 */
117 if (!ops) {
118 get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
119
120 if (get_ioengine)
121 get_ioengine(&ops);
122 }
123
d4dbaaa8 124 if (!ops) {
e1161c32 125 td_vmsg(td, -1, dlerror(), "dlsym");
d4dbaaa8
JA
126 dlclose(dlhandle);
127 return NULL;
128 }
8756e4d4 129
5f350952
JA
130 ops->dlhandle = dlhandle;
131 return ops;
132}
133
134struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
135{
136 struct ioengine_ops *ops, *ret;
137 char engine[16];
138
ee56ad50
JA
139 dprint(FD_IO, "load ioengine %s\n", name);
140
5f350952
JA
141 strncpy(engine, name, sizeof(engine) - 1);
142
143 /*
144 * linux libaio has alias names, so convert to what we want
145 */
146 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
147 strcpy(engine, "libaio");
148
149 ops = find_ioengine(engine);
150 if (!ops)
151 ops = dlopen_ioengine(td, name);
152
153 if (!ops) {
154 log_err("fio: engine %s not loadable\n", name);
b902ceb5
JA
155 return NULL;
156 }
157
8c16d840
JA
158 /*
159 * Check that the required methods are there.
160 */
5f350952 161 if (check_engine_ops(ops))
8c16d840 162 return NULL;
8c16d840 163
84585003
JA
164 ret = malloc(sizeof(*ret));
165 memcpy(ret, ops, sizeof(*ret));
166 ret->data = NULL;
84585003
JA
167
168 return ret;
8756e4d4
JA
169}
170
de890a1e
SL
171/*
172 * For cleaning up an ioengine which never made it to init().
173 */
174void free_ioengine(struct thread_data *td)
8756e4d4 175{
de890a1e 176 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
ee56ad50 177
de890a1e
SL
178 if (td->eo && td->io_ops->options) {
179 options_free(td->io_ops->options, td->eo);
180 free(td->eo);
181 td->eo = NULL;
2992b059 182 }
b990b5c0 183
5f350952
JA
184 if (td->io_ops->dlhandle)
185 dlclose(td->io_ops->dlhandle);
186
84585003
JA
187 free(td->io_ops);
188 td->io_ops = NULL;
b990b5c0 189}
10ba535a 190
de890a1e
SL
191void close_ioengine(struct thread_data *td)
192{
193 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
194
195 if (td->io_ops->cleanup) {
196 td->io_ops->cleanup(td);
197 td->io_ops->data = NULL;
198 }
199
200 free_ioengine(td);
201}
202
10ba535a
JA
203int td_io_prep(struct thread_data *td, struct io_u *io_u)
204{
ee56ad50 205 dprint_io_u(io_u, "prep");
7101d9c2
JA
206 fio_ro_check(td, io_u);
207
4d4e80f2 208 lock_file(td, io_u->file, io_u->ddir);
b2bd2bd9 209
2ba1c290
JA
210 if (td->io_ops->prep) {
211 int ret = td->io_ops->prep(td, io_u);
212
213 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
b2bd2bd9 214 if (ret)
4d4e80f2 215 unlock_file(td, io_u->file);
2ba1c290
JA
216 return ret;
217 }
10ba535a
JA
218
219 return 0;
220}
221
e7d2e616 222int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
1f440ece 223 const struct timespec *t)
10ba535a 224{
ee56ad50 225 int r = 0;
face81b2 226
a05d62b2
YR
227 /*
228 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
229 * server side gets a message from the client
230 * side that the task is finished, and
231 * td->done is set to 1 after td_io_commit(). In this case,
232 * there is no need to reap complete event in server side.
233 */
234 if (td->done)
235 return 0;
236
ee56ad50
JA
237 if (min > 0 && td->io_ops->commit) {
238 r = td->io_ops->commit(td);
face81b2 239 if (r < 0)
ee56ad50 240 goto out;
face81b2 241 }
4950421a
JA
242 if (max > td->cur_depth)
243 max = td->cur_depth;
244 if (min > max)
245 max = min;
36167d82 246
ee56ad50 247 r = 0;
4950421a 248 if (max && td->io_ops->getevents)
ee56ad50
JA
249 r = td->io_ops->getevents(td, min, max, t);
250out:
422f9e4b
RM
251 if (r >= 0) {
252 /*
3fd9efbc 253 * Reflect that our submitted requests were retrieved with
422f9e4b
RM
254 * whatever OS async calls are in the underlying engine.
255 */
256 td->io_u_in_flight -= r;
838bc709 257 io_u_mark_complete(td, r);
422f9e4b 258 } else
7c639b14 259 td_verror(td, r, "get_events");
f3e11d05 260
ee56ad50
JA
261 dprint(FD_IO, "getevents: %d\n", r);
262 return r;
10ba535a
JA
263}
264
265int td_io_queue(struct thread_data *td, struct io_u *io_u)
266{
7e77dd02
JA
267 int ret;
268
ee56ad50 269 dprint_io_u(io_u, "queue");
7101d9c2
JA
270 fio_ro_check(td, io_u);
271
0c6e7517
JA
272 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
273 io_u->flags |= IO_U_F_FLIGHT;
274
d6aed795 275 assert(fio_file_open(io_u->file));
3d7b485f 276
bcd5abfa
JA
277 /*
278 * If using a write iolog, store this entry.
279 */
280 log_io_u(td, io_u);
281
11786802
JA
282 io_u->error = 0;
283 io_u->resid = 0;
284
433afcb4 285 if (td->io_ops->flags & FIO_SYNCIO) {
12d9d841 286 if (fio_fill_issue_time(td))
9520ebb9 287 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
288
289 /*
290 * only used for iolog
291 */
292 if (td->o.read_iolog_file)
293 memcpy(&td->last_issue, &io_u->issue_time,
5ec10eaa 294 sizeof(struct timeval));
433afcb4
JA
295 }
296
74d6277f 297 if (ddir_rw(acct_ddir(io_u))) {
bcd5abfa 298 td->io_issues[acct_ddir(io_u)]++;
74d6277f
JA
299 td->io_issue_bytes[acct_ddir(io_u)] += io_u->xfer_buflen;
300 }
755200a3 301
7e77dd02 302 ret = td->io_ops->queue(td, io_u);
5aeb77df 303
4d4e80f2 304 unlock_file(td, io_u->file);
b2bd2bd9 305
cd8a19e6
JA
306 if (ret == FIO_Q_BUSY && ddir_rw(acct_ddir(io_u))) {
307 td->io_issues[acct_ddir(io_u)]--;
308 td->io_issue_bytes[acct_ddir(io_u)] -= io_u->xfer_buflen;
309 }
310
39a43d34
JA
311 /*
312 * If an error was seen and the io engine didn't propagate it
313 * back to 'td', do so.
314 */
315 if (io_u->error && !td->error)
316 td_verror(td, io_u->error, "td_io_queue");
317
cb211682
JA
318 /*
319 * Add warning for O_DIRECT so that users have an easier time
320 * spotting potentially bad alignment. If this triggers for the first
321 * IO, then it's likely an alignment problem or because the host fs
322 * does not support O_DIRECT
323 */
ff58fced 324 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
cb211682 325 td->o.odirect) {
214ac7e0 326
cb211682
JA
327 log_info("fio: first direct IO errored. File system may not "
328 "support direct IO, or iomem_align= is bad.\n");
329 }
330
77b451cc 331 if (!td->io_ops->commit || io_u->ddir == DDIR_TRIM) {
838bc709
JA
332 io_u_mark_submit(td, 1);
333 io_u_mark_complete(td, 1);
334 }
335
d8005759 336 if (ret == FIO_Q_COMPLETED) {
ff58fced 337 if (ddir_rw(io_u->ddir)) {
d8005759
JA
338 io_u_mark_depth(td, 1);
339 td->ts.total_io_u[io_u->ddir]++;
6eaf09d6 340 }
d8005759 341 } else if (ret == FIO_Q_QUEUED) {
eb7c8ae2
JA
342 int r;
343
ff58fced 344 if (ddir_rw(io_u->ddir)) {
d8005759
JA
345 td->io_u_queued++;
346 td->ts.total_io_u[io_u->ddir]++;
347 }
348
349 if (td->io_u_queued >= td->o.iodepth_batch) {
eb7c8ae2
JA
350 r = td_io_commit(td);
351 if (r < 0)
352 return r;
353 }
354 }
cb5ab512 355
433afcb4 356 if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
12d9d841 357 if (fio_fill_issue_time(td))
9520ebb9 358 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
359
360 /*
361 * only used for iolog
362 */
363 if (td->o.read_iolog_file)
364 memcpy(&td->last_issue, &io_u->issue_time,
365 sizeof(struct timeval));
433afcb4
JA
366 }
367
7e77dd02 368 return ret;
10ba535a 369}
8c16d840
JA
370
371int td_io_init(struct thread_data *td)
372{
eeb12160 373 int ret = 0;
8c16d840 374
eeb12160
JA
375 if (td->io_ops->init) {
376 ret = td->io_ops->init(td);
5ec10eaa
JA
377 if (ret && td->o.iodepth > 1) {
378 log_err("fio: io engine init failed. Perhaps try"
379 " reducing io depth?\n");
380 }
7c973896
JA
381 if (!td->error)
382 td->error = ret;
eeb12160
JA
383 }
384
046395d7
JA
385 if (!ret && (td->io_ops->flags & FIO_NOIO))
386 td->flags |= TD_F_NOIO;
387
eeb12160 388 return ret;
8c16d840 389}
755200a3
JA
390
391int td_io_commit(struct thread_data *td)
392{
f3e11d05
JA
393 int ret;
394
ee56ad50
JA
395 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
396
d8005759 397 if (!td->cur_depth || !td->io_u_queued)
e1161c32 398 return 0;
cb5ab512 399
3fd9efbc 400 io_u_mark_depth(td, td->io_u_queued);
d8005759 401
f3e11d05
JA
402 if (td->io_ops->commit) {
403 ret = td->io_ops->commit(td);
404 if (ret)
405 td_verror(td, -ret, "io commit");
406 }
3fd9efbc 407
422f9e4b
RM
408 /*
409 * Reflect that events were submitted as async IO requests.
410 */
411 td->io_u_in_flight += td->io_u_queued;
412 td->io_u_queued = 0;
755200a3
JA
413
414 return 0;
415}
b5af8293
JA
416
417int td_io_open_file(struct thread_data *td, struct fio_file *f)
418{
22a57ba8
JA
419 assert(!fio_file_open(f));
420 assert(f->fd == -1);
421
413d6693
JA
422 if (td->io_ops->open_file(td, f)) {
423 if (td->error == EINVAL && td->o.odirect)
424 log_err("fio: destination does not support O_DIRECT\n");
5ec10eaa
JA
425 if (td->error == EMFILE) {
426 log_err("fio: try reducing/setting openfiles (failed"
427 " at %u of %u)\n", td->nr_open_files,
428 td->o.nr_files);
429 }
413d6693 430
22a57ba8
JA
431 assert(f->fd == -1);
432 assert(!fio_file_open(f));
413d6693
JA
433 return 1;
434 }
435
33c48814 436 fio_file_reset(td, f);
d6aed795
JA
437 fio_file_set_open(f);
438 fio_file_clear_closing(f);
c97bd0fa 439 disk_util_inc(f->du);
d5707a35
JA
440
441 td->nr_open_files++;
442 get_file(f);
443
66159828
JA
444 if (f->filetype == FIO_TYPE_PIPE) {
445 if (td_random(td)) {
446 log_err("fio: can't seek on pipes (no random io)\n");
447 goto err;
448 }
449 }
450
413d6693
JA
451 if (td->io_ops->flags & FIO_DISKLESSIO)
452 goto done;
453
454 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
455 goto err;
456
66159828
JA
457 if (td->o.fadvise_hint &&
458 (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
413d6693
JA
459 int flags;
460
461 if (td_random(td))
462 flags = POSIX_FADV_RANDOM;
463 else
464 flags = POSIX_FADV_SEQUENTIAL;
465
ecc314ba 466 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
413d6693
JA
467 td_verror(td, errno, "fadvise");
468 goto err;
469 }
7bb48f84 470 }
a978ba68 471
e116f2b9
JA
472#ifdef FIO_OS_DIRECTIO
473 /*
474 * Some OS's have a distinct call to mark the file non-buffered,
475 * instead of using O_DIRECT (Solaris)
476 */
477 if (td->o.odirect) {
478 int ret = fio_set_odirect(f->fd);
479
480 if (ret) {
481 td_verror(td, ret, "fio_set_odirect");
78e51c72 482 log_err("fio: the file system does not seem to support direct IO\n");
e116f2b9
JA
483 goto err;
484 }
485 }
486#endif
487
413d6693 488done:
f29b25a3 489 log_file(td, f, FIO_LOG_OPEN_FILE);
413d6693
JA
490 return 0;
491err:
c97bd0fa 492 disk_util_dec(f->du);
b284075a
JA
493 if (td->io_ops->close_file)
494 td->io_ops->close_file(td, f);
7bb48f84 495 return 1;
b5af8293
JA
496}
497
6977bcd0 498int td_io_close_file(struct thread_data *td, struct fio_file *f)
b5af8293 499{
d6aed795 500 if (!fio_file_closing(f))
f29b25a3
JA
501 log_file(td, f, FIO_LOG_CLOSE_FILE);
502
0ad920e7
JA
503 /*
504 * mark as closing, do real close when last io on it has completed
505 */
d6aed795 506 fio_file_set_closing(f);
0ad920e7 507
c97bd0fa 508 disk_util_dec(f->du);
cba5243b
JA
509
510 if (td->o.file_lock_mode != FILE_LOCK_NONE)
511 unlock_file_all(td, f);
29c1349f 512
6977bcd0 513 return put_file(td, f);
b5af8293 514}
df9c26b1 515
38ef9c90
CF
516int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
517{
518 if (td->io_ops->unlink_file)
519 return td->io_ops->unlink_file(td, f);
520 else
521 return unlink(f->file_name);
522}
523
df9c26b1
JA
524int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
525{
526 if (!td->io_ops->get_file_size)
527 return 0;
528
529 return td->io_ops->get_file_size(td, f);
530}
44f29692 531
effd6ff0
JA
532static int do_sync_file_range(const struct thread_data *td,
533 struct fio_file *f)
44f29692
JA
534{
535 off64_t offset, nbytes;
536
537 offset = f->first_write;
538 nbytes = f->last_write - f->first_write;
539
3843deb3
JA
540 if (!nbytes)
541 return 0;
44f29692 542
3843deb3 543 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
44f29692 544}
0a28ecda 545
effd6ff0 546int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
0a28ecda
JA
547{
548 int ret;
549
550 if (io_u->ddir == DDIR_SYNC) {
551 ret = fsync(io_u->file->fd);
552 } else if (io_u->ddir == DDIR_DATASYNC) {
67bf9823 553#ifdef CONFIG_FDATASYNC
0a28ecda
JA
554 ret = fdatasync(io_u->file->fd);
555#else
556 ret = io_u->xfer_buflen;
557 io_u->error = EINVAL;
558#endif
559 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
560 ret = do_sync_file_range(td, io_u->file);
561 else {
562 ret = io_u->xfer_buflen;
563 io_u->error = EINVAL;
564 }
565
cb849a79
JA
566 if (ret < 0)
567 io_u->error = errno;
568
0a28ecda
JA
569 return ret;
570}
a5f3027c 571
effd6ff0 572int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
a5f3027c
JA
573{
574#ifndef FIO_HAVE_TRIM
575 io_u->error = EINVAL;
ff58fced 576 return 0;
a5f3027c
JA
577#else
578 struct fio_file *f = io_u->file;
579 int ret;
580
c6404a44 581 ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
a5f3027c 582 if (!ret)
3fd9efbc 583 return io_u->xfer_buflen;
a5f3027c 584
ff58fced
JA
585 io_u->error = ret;
586 return 0;
a5f3027c
JA
587#endif
588}
de890a1e
SL
589
590int fio_show_ioengine_help(const char *engine)
591{
592 struct flist_head *entry;
593 struct thread_data td;
594 char *sep;
595 int ret = 1;
596
597 if (!engine || !*engine) {
598 log_info("Available IO engines:\n");
599 flist_for_each(entry, &engine_list) {
600 td.io_ops = flist_entry(entry, struct ioengine_ops,
601 list);
602 log_info("\t%s\n", td.io_ops->name);
603 }
604 return 0;
605 }
606 sep = strchr(engine, ',');
607 if (sep) {
608 *sep = 0;
609 sep++;
610 }
611
612 memset(&td, 0, sizeof(td));
613
614 td.io_ops = load_ioengine(&td, engine);
615 if (!td.io_ops) {
616 log_info("IO engine %s not found\n", engine);
617 return 1;
618 }
619
620 if (td.io_ops->options)
621 ret = show_cmd_help(td.io_ops->options, sep);
622 else
623 log_info("IO engine %s has no options\n", td.io_ops->name);
624
625 free_ioengine(&td);
626
627 return ret;
628}