Rename FIO_TYPE_BD to FIO_TYPE_BLOCK
[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
1cc954ba 25static bool check_engine_ops(struct ioengine_ops *ops)
8c16d840 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);
1cc954ba 30 return true;
5f350952
JA
31 }
32
36167d82
JA
33 if (!ops->queue) {
34 log_err("%s: no queue handler\n", ops->name);
1cc954ba 35 return true;
36167d82
JA
36 }
37
38 /*
39 * sync engines only need a ->queue()
40 */
41 if (ops->flags & FIO_SYNCIO)
1cc954ba 42 return false;
5ec10eaa 43
1cc954ba
JA
44 if (!ops->event || !ops->getevents) {
45 log_err("%s: no event/getevents handler\n", ops->name);
46 return true;
8c16d840 47 }
5ec10eaa 48
1cc954ba 49 return false;
8c16d840
JA
50}
51
5f350952 52void unregister_ioengine(struct ioengine_ops *ops)
ebac4655 53{
ee56ad50 54 dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
01743ee1
JA
55 flist_del(&ops->list);
56 INIT_FLIST_HEAD(&ops->list);
5f350952
JA
57}
58
b2fdda43 59void register_ioengine(struct ioengine_ops *ops)
5f350952 60{
ee56ad50 61 dprint(FD_IO, "ioengine %s registered\n", ops->name);
01743ee1
JA
62 INIT_FLIST_HEAD(&ops->list);
63 flist_add_tail(&ops->list, &engine_list);
5f350952
JA
64}
65
66static struct ioengine_ops *find_ioengine(const char *name)
67{
68 struct ioengine_ops *ops;
01743ee1 69 struct flist_head *entry;
ebac4655 70
01743ee1
JA
71 flist_for_each(entry, &engine_list) {
72 ops = flist_entry(entry, struct ioengine_ops, list);
bc5b77a8 73 if (!strcmp(name, ops->name))
5f350952
JA
74 return ops;
75 }
76
77 return NULL;
78}
79
80static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
81 const char *engine_lib)
82{
83 struct ioengine_ops *ops;
84 void *dlhandle;
85
ee56ad50
JA
86 dprint(FD_IO, "dload engine %s\n", engine_lib);
87
2866c82d
JA
88 dlerror();
89 dlhandle = dlopen(engine_lib, RTLD_LAZY);
d4dbaaa8 90 if (!dlhandle) {
e1161c32 91 td_vmsg(td, -1, dlerror(), "dlopen");
d4dbaaa8
JA
92 return NULL;
93 }
8756e4d4 94
da51c050
JA
95 /*
96 * Unlike the included modules, external engines should have a
97 * non-static ioengine structure that we can reference.
98 */
0abea0b7
DM
99 ops = dlsym(dlhandle, engine_lib);
100 if (!ops)
101 ops = dlsym(dlhandle, "ioengine");
a8075704
DG
102
103 /*
104 * For some external engines (like C++ ones) it is not that trivial
105 * to provide a non-static ionengine structure that we can reference.
106 * Instead we call a method which allocates the required ioengine
107 * structure.
108 */
109 if (!ops) {
110 get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
111
112 if (get_ioengine)
113 get_ioengine(&ops);
114 }
115
d4dbaaa8 116 if (!ops) {
e1161c32 117 td_vmsg(td, -1, dlerror(), "dlsym");
d4dbaaa8
JA
118 dlclose(dlhandle);
119 return NULL;
120 }
8756e4d4 121
565e784d 122 td->io_ops_dlhandle = dlhandle;
5f350952
JA
123 return ops;
124}
125
126struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
127{
565e784d 128 struct ioengine_ops *ops;
aa2b823c 129 char engine[64];
5f350952 130
ee56ad50
JA
131 dprint(FD_IO, "load ioengine %s\n", name);
132
087d0ed0 133 engine[sizeof(engine) - 1] = '\0';
5f350952
JA
134 strncpy(engine, name, sizeof(engine) - 1);
135
136 /*
137 * linux libaio has alias names, so convert to what we want
138 */
139 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
140 strcpy(engine, "libaio");
141
142 ops = find_ioengine(engine);
143 if (!ops)
144 ops = dlopen_ioengine(td, name);
145
146 if (!ops) {
147 log_err("fio: engine %s not loadable\n", name);
b902ceb5
JA
148 return NULL;
149 }
150
8c16d840
JA
151 /*
152 * Check that the required methods are there.
153 */
5f350952 154 if (check_engine_ops(ops))
8c16d840 155 return NULL;
8c16d840 156
565e784d 157 return ops;
8756e4d4
JA
158}
159
de890a1e
SL
160/*
161 * For cleaning up an ioengine which never made it to init().
162 */
163void free_ioengine(struct thread_data *td)
8756e4d4 164{
de890a1e 165 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
ee56ad50 166
de890a1e
SL
167 if (td->eo && td->io_ops->options) {
168 options_free(td->io_ops->options, td->eo);
169 free(td->eo);
170 td->eo = NULL;
2992b059 171 }
b990b5c0 172
565e784d
TK
173 if (td->io_ops_dlhandle)
174 dlclose(td->io_ops_dlhandle);
5f350952 175
84585003 176 td->io_ops = NULL;
b990b5c0 177}
10ba535a 178
de890a1e
SL
179void close_ioengine(struct thread_data *td)
180{
181 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
182
183 if (td->io_ops->cleanup) {
184 td->io_ops->cleanup(td);
565e784d 185 td->io_ops_data = NULL;
de890a1e
SL
186 }
187
188 free_ioengine(td);
189}
190
10ba535a
JA
191int td_io_prep(struct thread_data *td, struct io_u *io_u)
192{
ee56ad50 193 dprint_io_u(io_u, "prep");
7101d9c2
JA
194 fio_ro_check(td, io_u);
195
4d4e80f2 196 lock_file(td, io_u->file, io_u->ddir);
b2bd2bd9 197
2ba1c290
JA
198 if (td->io_ops->prep) {
199 int ret = td->io_ops->prep(td, io_u);
200
201 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
b2bd2bd9 202 if (ret)
4d4e80f2 203 unlock_file(td, io_u->file);
2ba1c290
JA
204 return ret;
205 }
10ba535a
JA
206
207 return 0;
208}
209
e7d2e616 210int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
1f440ece 211 const struct timespec *t)
10ba535a 212{
ee56ad50 213 int r = 0;
face81b2 214
a05d62b2
YR
215 /*
216 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
217 * server side gets a message from the client
218 * side that the task is finished, and
219 * td->done is set to 1 after td_io_commit(). In this case,
220 * there is no need to reap complete event in server side.
221 */
222 if (td->done)
223 return 0;
224
ee56ad50
JA
225 if (min > 0 && td->io_ops->commit) {
226 r = td->io_ops->commit(td);
face81b2 227 if (r < 0)
ee56ad50 228 goto out;
face81b2 229 }
4950421a
JA
230 if (max > td->cur_depth)
231 max = td->cur_depth;
232 if (min > max)
233 max = min;
36167d82 234
ee56ad50 235 r = 0;
4950421a 236 if (max && td->io_ops->getevents)
ee56ad50
JA
237 r = td->io_ops->getevents(td, min, max, t);
238out:
422f9e4b
RM
239 if (r >= 0) {
240 /*
3fd9efbc 241 * Reflect that our submitted requests were retrieved with
422f9e4b
RM
242 * whatever OS async calls are in the underlying engine.
243 */
244 td->io_u_in_flight -= r;
838bc709 245 io_u_mark_complete(td, r);
422f9e4b 246 } else
7c639b14 247 td_verror(td, r, "get_events");
f3e11d05 248
ee56ad50
JA
249 dprint(FD_IO, "getevents: %d\n", r);
250 return r;
10ba535a
JA
251}
252
253int td_io_queue(struct thread_data *td, struct io_u *io_u)
254{
a9da8ab2
JA
255 const enum fio_ddir ddir = acct_ddir(io_u);
256 unsigned long buflen = io_u->xfer_buflen;
7e77dd02
JA
257 int ret;
258
ee56ad50 259 dprint_io_u(io_u, "queue");
7101d9c2
JA
260 fio_ro_check(td, io_u);
261
0c6e7517 262 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
1651e431 263 io_u_set(td, io_u, IO_U_F_FLIGHT);
0c6e7517 264
d6aed795 265 assert(fio_file_open(io_u->file));
3d7b485f 266
bcd5abfa
JA
267 /*
268 * If using a write iolog, store this entry.
269 */
270 log_io_u(td, io_u);
271
11786802
JA
272 io_u->error = 0;
273 io_u->resid = 0;
274
9b87f09b 275 if (td_ioengine_flagged(td, FIO_SYNCIO)) {
12d9d841 276 if (fio_fill_issue_time(td))
9520ebb9 277 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
278
279 /*
280 * only used for iolog
281 */
282 if (td->o.read_iolog_file)
283 memcpy(&td->last_issue, &io_u->issue_time,
5ec10eaa 284 sizeof(struct timeval));
433afcb4
JA
285 }
286
a9da8ab2
JA
287 if (ddir_rw(ddir)) {
288 td->io_issues[ddir]++;
289 td->io_issue_bytes[ddir] += buflen;
50a8ce86 290 td->rate_io_issue_bytes[ddir] += buflen;
74d6277f 291 }
755200a3 292
7e77dd02 293 ret = td->io_ops->queue(td, io_u);
5aeb77df 294
4d4e80f2 295 unlock_file(td, io_u->file);
b2bd2bd9 296
a9da8ab2
JA
297 if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
298 td->io_issues[ddir]--;
299 td->io_issue_bytes[ddir] -= buflen;
50a8ce86 300 td->rate_io_issue_bytes[ddir] -= buflen;
871467d9 301 io_u_clear(td, io_u, IO_U_F_FLIGHT);
cd8a19e6
JA
302 }
303
39a43d34
JA
304 /*
305 * If an error was seen and the io engine didn't propagate it
306 * back to 'td', do so.
307 */
308 if (io_u->error && !td->error)
309 td_verror(td, io_u->error, "td_io_queue");
310
cb211682
JA
311 /*
312 * Add warning for O_DIRECT so that users have an easier time
313 * spotting potentially bad alignment. If this triggers for the first
314 * IO, then it's likely an alignment problem or because the host fs
315 * does not support O_DIRECT
316 */
ff58fced 317 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
cb211682 318 td->o.odirect) {
214ac7e0 319
cb211682 320 log_info("fio: first direct IO errored. File system may not "
44cc9c8f
JA
321 "support direct IO, or iomem_align= is bad. Try "
322 "setting direct=0.\n");
cb211682
JA
323 }
324
77b451cc 325 if (!td->io_ops->commit || io_u->ddir == DDIR_TRIM) {
838bc709
JA
326 io_u_mark_submit(td, 1);
327 io_u_mark_complete(td, 1);
328 }
329
d8005759 330 if (ret == FIO_Q_COMPLETED) {
ff58fced 331 if (ddir_rw(io_u->ddir)) {
d8005759
JA
332 io_u_mark_depth(td, 1);
333 td->ts.total_io_u[io_u->ddir]++;
6eaf09d6 334 }
d8005759 335 } else if (ret == FIO_Q_QUEUED) {
eb7c8ae2
JA
336 int r;
337
e6727cbd
JA
338 td->io_u_queued++;
339
340 if (ddir_rw(io_u->ddir))
d8005759 341 td->ts.total_io_u[io_u->ddir]++;
d8005759
JA
342
343 if (td->io_u_queued >= td->o.iodepth_batch) {
eb7c8ae2
JA
344 r = td_io_commit(td);
345 if (r < 0)
346 return r;
347 }
348 }
cb5ab512 349
9b87f09b 350 if (!td_ioengine_flagged(td, FIO_SYNCIO)) {
12d9d841 351 if (fio_fill_issue_time(td))
9520ebb9 352 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
353
354 /*
355 * only used for iolog
356 */
357 if (td->o.read_iolog_file)
358 memcpy(&td->last_issue, &io_u->issue_time,
359 sizeof(struct timeval));
433afcb4
JA
360 }
361
7e77dd02 362 return ret;
10ba535a 363}
8c16d840
JA
364
365int td_io_init(struct thread_data *td)
366{
eeb12160 367 int ret = 0;
8c16d840 368
eeb12160
JA
369 if (td->io_ops->init) {
370 ret = td->io_ops->init(td);
5ec10eaa
JA
371 if (ret && td->o.iodepth > 1) {
372 log_err("fio: io engine init failed. Perhaps try"
373 " reducing io depth?\n");
374 }
7c973896
JA
375 if (!td->error)
376 td->error = ret;
eeb12160
JA
377 }
378
9b87f09b 379 if (!ret && td_ioengine_flagged(td, FIO_NOIO))
046395d7
JA
380 td->flags |= TD_F_NOIO;
381
eeb12160 382 return ret;
8c16d840 383}
755200a3
JA
384
385int td_io_commit(struct thread_data *td)
386{
f3e11d05
JA
387 int ret;
388
ee56ad50
JA
389 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
390
d8005759 391 if (!td->cur_depth || !td->io_u_queued)
e1161c32 392 return 0;
cb5ab512 393
3fd9efbc 394 io_u_mark_depth(td, td->io_u_queued);
d8005759 395
f3e11d05
JA
396 if (td->io_ops->commit) {
397 ret = td->io_ops->commit(td);
398 if (ret)
399 td_verror(td, -ret, "io commit");
400 }
3fd9efbc 401
422f9e4b
RM
402 /*
403 * Reflect that events were submitted as async IO requests.
404 */
405 td->io_u_in_flight += td->io_u_queued;
406 td->io_u_queued = 0;
755200a3
JA
407
408 return 0;
409}
b5af8293
JA
410
411int td_io_open_file(struct thread_data *td, struct fio_file *f)
412{
22a57ba8
JA
413 assert(!fio_file_open(f));
414 assert(f->fd == -1);
415
413d6693
JA
416 if (td->io_ops->open_file(td, f)) {
417 if (td->error == EINVAL && td->o.odirect)
418 log_err("fio: destination does not support O_DIRECT\n");
5ec10eaa
JA
419 if (td->error == EMFILE) {
420 log_err("fio: try reducing/setting openfiles (failed"
421 " at %u of %u)\n", td->nr_open_files,
422 td->o.nr_files);
423 }
413d6693 424
22a57ba8
JA
425 assert(f->fd == -1);
426 assert(!fio_file_open(f));
413d6693
JA
427 return 1;
428 }
429
33c48814 430 fio_file_reset(td, f);
d6aed795
JA
431 fio_file_set_open(f);
432 fio_file_clear_closing(f);
c97bd0fa 433 disk_util_inc(f->du);
d5707a35
JA
434
435 td->nr_open_files++;
436 get_file(f);
437
66159828
JA
438 if (f->filetype == FIO_TYPE_PIPE) {
439 if (td_random(td)) {
440 log_err("fio: can't seek on pipes (no random io)\n");
441 goto err;
442 }
443 }
444
9b87f09b 445 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
413d6693
JA
446 goto done;
447
448 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
449 goto err;
450
ecb2083d 451 if (td->o.fadvise_hint != F_ADV_NONE &&
686fbd31 452 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
413d6693
JA
453 int flags;
454
ecb2083d
JA
455 if (td->o.fadvise_hint == F_ADV_TYPE) {
456 if (td_random(td))
457 flags = POSIX_FADV_RANDOM;
458 else
459 flags = POSIX_FADV_SEQUENTIAL;
460 } else if (td->o.fadvise_hint == F_ADV_RANDOM)
413d6693 461 flags = POSIX_FADV_RANDOM;
ecb2083d 462 else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
413d6693 463 flags = POSIX_FADV_SEQUENTIAL;
ecb2083d
JA
464 else {
465 log_err("fio: unknown fadvise type %d\n",
466 td->o.fadvise_hint);
467 flags = POSIX_FADV_NORMAL;
468 }
413d6693 469
ecc314ba 470 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
413d6693
JA
471 td_verror(td, errno, "fadvise");
472 goto err;
473 }
7bb48f84 474 }
37659335
JA
475#ifdef FIO_HAVE_STREAMID
476 if (td->o.fadvise_stream &&
686fbd31 477 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
37659335
JA
478 off_t stream = td->o.fadvise_stream;
479
480 if (posix_fadvise(f->fd, stream, f->io_size, POSIX_FADV_STREAMID) < 0) {
481 td_verror(td, errno, "fadvise streamid");
482 goto err;
483 }
484 }
485#endif
a978ba68 486
e116f2b9
JA
487#ifdef FIO_OS_DIRECTIO
488 /*
489 * Some OS's have a distinct call to mark the file non-buffered,
490 * instead of using O_DIRECT (Solaris)
491 */
492 if (td->o.odirect) {
493 int ret = fio_set_odirect(f->fd);
494
495 if (ret) {
496 td_verror(td, ret, "fio_set_odirect");
42f1ee68
SZ
497 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
498 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
499 } else {
500 log_err("fio: the file system does not seem to support direct IO\n");
501 }
502
e116f2b9
JA
503 goto err;
504 }
505 }
506#endif
507
413d6693 508done:
f29b25a3 509 log_file(td, f, FIO_LOG_OPEN_FILE);
413d6693
JA
510 return 0;
511err:
c97bd0fa 512 disk_util_dec(f->du);
b284075a
JA
513 if (td->io_ops->close_file)
514 td->io_ops->close_file(td, f);
7bb48f84 515 return 1;
b5af8293
JA
516}
517
6977bcd0 518int td_io_close_file(struct thread_data *td, struct fio_file *f)
b5af8293 519{
d6aed795 520 if (!fio_file_closing(f))
f29b25a3
JA
521 log_file(td, f, FIO_LOG_CLOSE_FILE);
522
0ad920e7
JA
523 /*
524 * mark as closing, do real close when last io on it has completed
525 */
d6aed795 526 fio_file_set_closing(f);
0ad920e7 527
c97bd0fa 528 disk_util_dec(f->du);
cba5243b
JA
529
530 if (td->o.file_lock_mode != FILE_LOCK_NONE)
531 unlock_file_all(td, f);
29c1349f 532
6977bcd0 533 return put_file(td, f);
b5af8293 534}
df9c26b1 535
38ef9c90
CF
536int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
537{
538 if (td->io_ops->unlink_file)
539 return td->io_ops->unlink_file(td, f);
2442c935
JA
540 else {
541 int ret;
542
543 ret = unlink(f->file_name);
544 if (ret < 0)
545 return errno;
546
547 return 0;
548 }
38ef9c90
CF
549}
550
df9c26b1
JA
551int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
552{
553 if (!td->io_ops->get_file_size)
554 return 0;
555
556 return td->io_ops->get_file_size(td, f);
557}
44f29692 558
effd6ff0
JA
559static int do_sync_file_range(const struct thread_data *td,
560 struct fio_file *f)
44f29692
JA
561{
562 off64_t offset, nbytes;
563
564 offset = f->first_write;
565 nbytes = f->last_write - f->first_write;
566
3843deb3
JA
567 if (!nbytes)
568 return 0;
44f29692 569
3843deb3 570 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
44f29692 571}
0a28ecda 572
effd6ff0 573int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
0a28ecda
JA
574{
575 int ret;
576
577 if (io_u->ddir == DDIR_SYNC) {
578 ret = fsync(io_u->file->fd);
579 } else if (io_u->ddir == DDIR_DATASYNC) {
67bf9823 580#ifdef CONFIG_FDATASYNC
0a28ecda
JA
581 ret = fdatasync(io_u->file->fd);
582#else
583 ret = io_u->xfer_buflen;
584 io_u->error = EINVAL;
585#endif
586 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
587 ret = do_sync_file_range(td, io_u->file);
588 else {
589 ret = io_u->xfer_buflen;
590 io_u->error = EINVAL;
591 }
592
cb849a79
JA
593 if (ret < 0)
594 io_u->error = errno;
595
0a28ecda
JA
596 return ret;
597}
a5f3027c 598
effd6ff0 599int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
a5f3027c
JA
600{
601#ifndef FIO_HAVE_TRIM
602 io_u->error = EINVAL;
ff58fced 603 return 0;
a5f3027c
JA
604#else
605 struct fio_file *f = io_u->file;
606 int ret;
607
c6404a44 608 ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
a5f3027c 609 if (!ret)
3fd9efbc 610 return io_u->xfer_buflen;
a5f3027c 611
ff58fced
JA
612 io_u->error = ret;
613 return 0;
a5f3027c
JA
614#endif
615}
de890a1e
SL
616
617int fio_show_ioengine_help(const char *engine)
618{
619 struct flist_head *entry;
620 struct thread_data td;
621 char *sep;
622 int ret = 1;
623
624 if (!engine || !*engine) {
625 log_info("Available IO engines:\n");
626 flist_for_each(entry, &engine_list) {
627 td.io_ops = flist_entry(entry, struct ioengine_ops,
628 list);
629 log_info("\t%s\n", td.io_ops->name);
630 }
631 return 0;
632 }
633 sep = strchr(engine, ',');
634 if (sep) {
635 *sep = 0;
636 sep++;
637 }
638
639 memset(&td, 0, sizeof(td));
640
641 td.io_ops = load_ioengine(&td, engine);
642 if (!td.io_ops) {
643 log_info("IO engine %s not found\n", engine);
644 return 1;
645 }
646
647 if (td.io_ops->options)
648 ret = show_cmd_help(td.io_ops->options, sep);
649 else
650 log_info("IO engine %s has no options\n", td.io_ops->name);
651
652 free_ioengine(&td);
653
654 return ret;
655}