ioengine: don't account verify bytes
[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);
adf075fa 55 flist_del_init(&ops->list);
5f350952
JA
56}
57
b2fdda43 58void register_ioengine(struct ioengine_ops *ops)
5f350952 59{
ee56ad50 60 dprint(FD_IO, "ioengine %s registered\n", ops->name);
01743ee1 61 flist_add_tail(&ops->list, &engine_list);
5f350952
JA
62}
63
64static struct ioengine_ops *find_ioengine(const char *name)
65{
66 struct ioengine_ops *ops;
01743ee1 67 struct flist_head *entry;
ebac4655 68
01743ee1
JA
69 flist_for_each(entry, &engine_list) {
70 ops = flist_entry(entry, struct ioengine_ops, list);
bc5b77a8 71 if (!strcmp(name, ops->name))
5f350952
JA
72 return ops;
73 }
74
75 return NULL;
76}
77
78static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
79 const char *engine_lib)
80{
81 struct ioengine_ops *ops;
82 void *dlhandle;
83
ee56ad50
JA
84 dprint(FD_IO, "dload engine %s\n", engine_lib);
85
2866c82d
JA
86 dlerror();
87 dlhandle = dlopen(engine_lib, RTLD_LAZY);
d4dbaaa8 88 if (!dlhandle) {
e1161c32 89 td_vmsg(td, -1, dlerror(), "dlopen");
d4dbaaa8
JA
90 return NULL;
91 }
8756e4d4 92
da51c050
JA
93 /*
94 * Unlike the included modules, external engines should have a
95 * non-static ioengine structure that we can reference.
96 */
0abea0b7
DM
97 ops = dlsym(dlhandle, engine_lib);
98 if (!ops)
99 ops = dlsym(dlhandle, "ioengine");
a8075704
DG
100
101 /*
102 * For some external engines (like C++ ones) it is not that trivial
103 * to provide a non-static ionengine structure that we can reference.
104 * Instead we call a method which allocates the required ioengine
105 * structure.
106 */
107 if (!ops) {
108 get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
109
110 if (get_ioengine)
111 get_ioengine(&ops);
112 }
113
d4dbaaa8 114 if (!ops) {
e1161c32 115 td_vmsg(td, -1, dlerror(), "dlsym");
d4dbaaa8
JA
116 dlclose(dlhandle);
117 return NULL;
118 }
8756e4d4 119
565e784d 120 td->io_ops_dlhandle = dlhandle;
5f350952
JA
121 return ops;
122}
123
97bb54c9
TK
124static struct ioengine_ops *__load_ioengine(const char *name)
125{
126 char engine[64];
127
128 engine[sizeof(engine) - 1] = '\0';
129 strncpy(engine, name, sizeof(engine) - 1);
130
131 /*
132 * linux libaio has alias names, so convert to what we want
133 */
966fcbd4 134 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3)) {
135 dprint(FD_IO, "converting ioengine name: %s -> libaio\n", name);
97bb54c9 136 strcpy(engine, "libaio");
966fcbd4 137 }
97bb54c9
TK
138
139 dprint(FD_IO, "load ioengine %s\n", engine);
140 return find_ioengine(engine);
141}
142
4fedf59a 143struct ioengine_ops *load_ioengine(struct thread_data *td)
5f350952 144{
81647a9a 145 struct ioengine_ops *ops = NULL;
ba872a0b 146 const char *name;
81647a9a 147
ba872a0b
TK
148 /*
149 * Use ->ioengine_so_path if an external ioengine path is specified.
150 * In this case, ->ioengine is "external" which also means the prefix
151 * for external ioengines "external:" is properly used.
152 */
153 name = td->o.ioengine_so_path ?: td->o.ioengine;
154
155 /*
156 * Try to load ->ioengine first, and if failed try to dlopen(3) either
157 * ->ioengine or ->ioengine_so_path. This is redundant for an external
158 * ioengine with prefix, and also leaves the possibility of unexpected
159 * behavior (e.g. if the "external" ioengine exists), but we do this
160 * so as not to break job files not using the prefix.
161 */
162 ops = __load_ioengine(td->o.ioengine);
163 if (!ops)
5f350952
JA
164 ops = dlopen_ioengine(td, name);
165
ba872a0b
TK
166 /*
167 * If ops is NULL, we failed to load ->ioengine, and also failed to
168 * dlopen(3) either ->ioengine or ->ioengine_so_path as a path.
169 */
5f350952
JA
170 if (!ops) {
171 log_err("fio: engine %s not loadable\n", name);
b902ceb5
JA
172 return NULL;
173 }
174
8c16d840
JA
175 /*
176 * Check that the required methods are there.
177 */
5f350952 178 if (check_engine_ops(ops))
8c16d840 179 return NULL;
8c16d840 180
565e784d 181 return ops;
8756e4d4
JA
182}
183
de890a1e
SL
184/*
185 * For cleaning up an ioengine which never made it to init().
186 */
187void free_ioengine(struct thread_data *td)
8756e4d4 188{
de890a1e 189 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
ee56ad50 190
de890a1e
SL
191 if (td->eo && td->io_ops->options) {
192 options_free(td->io_ops->options, td->eo);
193 free(td->eo);
194 td->eo = NULL;
2992b059 195 }
b990b5c0 196
565e784d
TK
197 if (td->io_ops_dlhandle)
198 dlclose(td->io_ops_dlhandle);
5f350952 199
84585003 200 td->io_ops = NULL;
b990b5c0 201}
10ba535a 202
de890a1e
SL
203void close_ioengine(struct thread_data *td)
204{
205 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
206
207 if (td->io_ops->cleanup) {
208 td->io_ops->cleanup(td);
565e784d 209 td->io_ops_data = NULL;
de890a1e
SL
210 }
211
212 free_ioengine(td);
213}
214
10ba535a
JA
215int td_io_prep(struct thread_data *td, struct io_u *io_u)
216{
ee56ad50 217 dprint_io_u(io_u, "prep");
7101d9c2
JA
218 fio_ro_check(td, io_u);
219
4d4e80f2 220 lock_file(td, io_u->file, io_u->ddir);
b2bd2bd9 221
2ba1c290
JA
222 if (td->io_ops->prep) {
223 int ret = td->io_ops->prep(td, io_u);
224
225 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
b2bd2bd9 226 if (ret)
4d4e80f2 227 unlock_file(td, io_u->file);
2ba1c290
JA
228 return ret;
229 }
10ba535a
JA
230
231 return 0;
232}
233
e7d2e616 234int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
1f440ece 235 const struct timespec *t)
10ba535a 236{
ee56ad50 237 int r = 0;
face81b2 238
a05d62b2
YR
239 /*
240 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
241 * server side gets a message from the client
242 * side that the task is finished, and
243 * td->done is set to 1 after td_io_commit(). In this case,
244 * there is no need to reap complete event in server side.
245 */
246 if (td->done)
247 return 0;
248
ee56ad50
JA
249 if (min > 0 && td->io_ops->commit) {
250 r = td->io_ops->commit(td);
face81b2 251 if (r < 0)
ee56ad50 252 goto out;
face81b2 253 }
4950421a
JA
254 if (max > td->cur_depth)
255 max = td->cur_depth;
256 if (min > max)
257 max = min;
36167d82 258
ee56ad50 259 r = 0;
4950421a 260 if (max && td->io_ops->getevents)
ee56ad50
JA
261 r = td->io_ops->getevents(td, min, max, t);
262out:
422f9e4b
RM
263 if (r >= 0) {
264 /*
3fd9efbc 265 * Reflect that our submitted requests were retrieved with
422f9e4b
RM
266 * whatever OS async calls are in the underlying engine.
267 */
268 td->io_u_in_flight -= r;
838bc709 269 io_u_mark_complete(td, r);
422f9e4b 270 } else
7c639b14 271 td_verror(td, r, "get_events");
f3e11d05 272
ee56ad50
JA
273 dprint(FD_IO, "getevents: %d\n", r);
274 return r;
10ba535a
JA
275}
276
277int td_io_queue(struct thread_data *td, struct io_u *io_u)
278{
a9da8ab2
JA
279 const enum fio_ddir ddir = acct_ddir(io_u);
280 unsigned long buflen = io_u->xfer_buflen;
7e77dd02
JA
281 int ret;
282
ee56ad50 283 dprint_io_u(io_u, "queue");
7101d9c2
JA
284 fio_ro_check(td, io_u);
285
0c6e7517 286 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
1651e431 287 io_u_set(td, io_u, IO_U_F_FLIGHT);
0c6e7517 288
d6aed795 289 assert(fio_file_open(io_u->file));
3d7b485f 290
bcd5abfa
JA
291 /*
292 * If using a write iolog, store this entry.
293 */
294 log_io_u(td, io_u);
295
11786802
JA
296 io_u->error = 0;
297 io_u->resid = 0;
298
9b87f09b 299 if (td_ioengine_flagged(td, FIO_SYNCIO)) {
12d9d841 300 if (fio_fill_issue_time(td))
9520ebb9 301 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
302
303 /*
304 * only used for iolog
305 */
306 if (td->o.read_iolog_file)
307 memcpy(&td->last_issue, &io_u->issue_time,
8b6a404c 308 sizeof(io_u->issue_time));
433afcb4
JA
309 }
310
a9da8ab2 311 if (ddir_rw(ddir)) {
a7d01f02
JA
312 if (!(io_u->flags & IO_U_F_VER_LIST)) {
313 td->io_issues[ddir]++;
314 td->io_issue_bytes[ddir] += buflen;
315 }
50a8ce86 316 td->rate_io_issue_bytes[ddir] += buflen;
74d6277f 317 }
755200a3 318
7e77dd02 319 ret = td->io_ops->queue(td, io_u);
5aeb77df 320
4d4e80f2 321 unlock_file(td, io_u->file);
b2bd2bd9 322
a9da8ab2
JA
323 if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
324 td->io_issues[ddir]--;
325 td->io_issue_bytes[ddir] -= buflen;
50a8ce86 326 td->rate_io_issue_bytes[ddir] -= buflen;
871467d9 327 io_u_clear(td, io_u, IO_U_F_FLIGHT);
cd8a19e6
JA
328 }
329
39a43d34
JA
330 /*
331 * If an error was seen and the io engine didn't propagate it
332 * back to 'td', do so.
333 */
334 if (io_u->error && !td->error)
335 td_verror(td, io_u->error, "td_io_queue");
336
cb211682
JA
337 /*
338 * Add warning for O_DIRECT so that users have an easier time
339 * spotting potentially bad alignment. If this triggers for the first
340 * IO, then it's likely an alignment problem or because the host fs
341 * does not support O_DIRECT
342 */
ff58fced 343 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
cb211682 344 td->o.odirect) {
214ac7e0 345
cb211682 346 log_info("fio: first direct IO errored. File system may not "
78d4a262
TK
347 "support direct IO, or iomem_align= is bad, or "
348 "invalid block size. Try setting direct=0.\n");
cb211682
JA
349 }
350
77b451cc 351 if (!td->io_ops->commit || io_u->ddir == DDIR_TRIM) {
838bc709
JA
352 io_u_mark_submit(td, 1);
353 io_u_mark_complete(td, 1);
354 }
355
d8005759 356 if (ret == FIO_Q_COMPLETED) {
ff58fced 357 if (ddir_rw(io_u->ddir)) {
d8005759
JA
358 io_u_mark_depth(td, 1);
359 td->ts.total_io_u[io_u->ddir]++;
6eaf09d6 360 }
d8005759 361 } else if (ret == FIO_Q_QUEUED) {
eb7c8ae2
JA
362 int r;
363
e6727cbd
JA
364 td->io_u_queued++;
365
366 if (ddir_rw(io_u->ddir))
d8005759 367 td->ts.total_io_u[io_u->ddir]++;
d8005759
JA
368
369 if (td->io_u_queued >= td->o.iodepth_batch) {
eb7c8ae2
JA
370 r = td_io_commit(td);
371 if (r < 0)
372 return r;
373 }
374 }
cb5ab512 375
9b87f09b 376 if (!td_ioengine_flagged(td, FIO_SYNCIO)) {
12d9d841 377 if (fio_fill_issue_time(td))
9520ebb9 378 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
379
380 /*
381 * only used for iolog
382 */
383 if (td->o.read_iolog_file)
384 memcpy(&td->last_issue, &io_u->issue_time,
8b6a404c 385 sizeof(io_u->issue_time));
433afcb4
JA
386 }
387
7e77dd02 388 return ret;
10ba535a 389}
8c16d840
JA
390
391int td_io_init(struct thread_data *td)
392{
eeb12160 393 int ret = 0;
8c16d840 394
eeb12160
JA
395 if (td->io_ops->init) {
396 ret = td->io_ops->init(td);
356ef1a1
TK
397 if (ret)
398 log_err("fio: io engine %s init failed.%s\n",
399 td->io_ops->name,
400 td->o.iodepth > 1 ?
401 " Perhaps try reducing io depth?" : "");
402 else
403 td->io_ops_init = 1;
7c973896
JA
404 if (!td->error)
405 td->error = ret;
eeb12160
JA
406 }
407
408 return ret;
8c16d840 409}
755200a3
JA
410
411int td_io_commit(struct thread_data *td)
412{
f3e11d05
JA
413 int ret;
414
ee56ad50
JA
415 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
416
d8005759 417 if (!td->cur_depth || !td->io_u_queued)
e1161c32 418 return 0;
cb5ab512 419
3fd9efbc 420 io_u_mark_depth(td, td->io_u_queued);
d8005759 421
f3e11d05
JA
422 if (td->io_ops->commit) {
423 ret = td->io_ops->commit(td);
424 if (ret)
425 td_verror(td, -ret, "io commit");
426 }
3fd9efbc 427
422f9e4b
RM
428 /*
429 * Reflect that events were submitted as async IO requests.
430 */
431 td->io_u_in_flight += td->io_u_queued;
432 td->io_u_queued = 0;
755200a3
JA
433
434 return 0;
435}
b5af8293
JA
436
437int td_io_open_file(struct thread_data *td, struct fio_file *f)
438{
22a57ba8
JA
439 assert(!fio_file_open(f));
440 assert(f->fd == -1);
966fcbd4 441 assert(td->io_ops->open_file);
22a57ba8 442
413d6693
JA
443 if (td->io_ops->open_file(td, f)) {
444 if (td->error == EINVAL && td->o.odirect)
445 log_err("fio: destination does not support O_DIRECT\n");
5ec10eaa
JA
446 if (td->error == EMFILE) {
447 log_err("fio: try reducing/setting openfiles (failed"
448 " at %u of %u)\n", td->nr_open_files,
449 td->o.nr_files);
450 }
413d6693 451
22a57ba8
JA
452 assert(f->fd == -1);
453 assert(!fio_file_open(f));
413d6693
JA
454 return 1;
455 }
456
33c48814 457 fio_file_reset(td, f);
d6aed795
JA
458 fio_file_set_open(f);
459 fio_file_clear_closing(f);
c97bd0fa 460 disk_util_inc(f->du);
d5707a35
JA
461
462 td->nr_open_files++;
463 get_file(f);
464
66159828
JA
465 if (f->filetype == FIO_TYPE_PIPE) {
466 if (td_random(td)) {
467 log_err("fio: can't seek on pipes (no random io)\n");
468 goto err;
469 }
470 }
471
9b87f09b 472 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
413d6693
JA
473 goto done;
474
475 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
476 goto err;
477
ecb2083d 478 if (td->o.fadvise_hint != F_ADV_NONE &&
686fbd31 479 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
413d6693
JA
480 int flags;
481
ecb2083d
JA
482 if (td->o.fadvise_hint == F_ADV_TYPE) {
483 if (td_random(td))
484 flags = POSIX_FADV_RANDOM;
485 else
486 flags = POSIX_FADV_SEQUENTIAL;
487 } else if (td->o.fadvise_hint == F_ADV_RANDOM)
413d6693 488 flags = POSIX_FADV_RANDOM;
ecb2083d 489 else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
413d6693 490 flags = POSIX_FADV_SEQUENTIAL;
ecb2083d
JA
491 else {
492 log_err("fio: unknown fadvise type %d\n",
493 td->o.fadvise_hint);
494 flags = POSIX_FADV_NORMAL;
495 }
413d6693 496
ecc314ba 497 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
413d6693
JA
498 td_verror(td, errno, "fadvise");
499 goto err;
500 }
7bb48f84 501 }
ae8e559e
JA
502#ifdef FIO_HAVE_WRITE_HINT
503 if (fio_option_is_set(&td->o, write_hint) &&
686fbd31 504 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
ae8e559e 505 uint64_t hint = td->o.write_hint;
bd553af6 506 int cmd;
37659335 507
bd553af6
JA
508 /*
509 * For direct IO, we just need/want to set the hint on
510 * the file descriptor. For buffered IO, we need to set
511 * it on the inode.
512 */
513 if (td->o.odirect)
514 cmd = F_SET_FILE_RW_HINT;
515 else
516 cmd = F_SET_RW_HINT;
517
518 if (fcntl(f->fd, cmd, &hint) < 0) {
ae8e559e 519 td_verror(td, errno, "fcntl write hint");
37659335
JA
520 goto err;
521 }
522 }
523#endif
a978ba68 524
47534cda
TK
525 if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
526 goto err;
e116f2b9 527
413d6693 528done:
f29b25a3 529 log_file(td, f, FIO_LOG_OPEN_FILE);
413d6693
JA
530 return 0;
531err:
c97bd0fa 532 disk_util_dec(f->du);
b284075a
JA
533 if (td->io_ops->close_file)
534 td->io_ops->close_file(td, f);
7bb48f84 535 return 1;
b5af8293
JA
536}
537
6977bcd0 538int td_io_close_file(struct thread_data *td, struct fio_file *f)
b5af8293 539{
d6aed795 540 if (!fio_file_closing(f))
f29b25a3
JA
541 log_file(td, f, FIO_LOG_CLOSE_FILE);
542
0ad920e7
JA
543 /*
544 * mark as closing, do real close when last io on it has completed
545 */
d6aed795 546 fio_file_set_closing(f);
0ad920e7 547
c97bd0fa 548 disk_util_dec(f->du);
cba5243b
JA
549
550 if (td->o.file_lock_mode != FILE_LOCK_NONE)
551 unlock_file_all(td, f);
29c1349f 552
6977bcd0 553 return put_file(td, f);
b5af8293 554}
df9c26b1 555
38ef9c90
CF
556int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
557{
558 if (td->io_ops->unlink_file)
559 return td->io_ops->unlink_file(td, f);
2442c935
JA
560 else {
561 int ret;
562
563 ret = unlink(f->file_name);
564 if (ret < 0)
565 return errno;
566
567 return 0;
568 }
38ef9c90
CF
569}
570
df9c26b1
JA
571int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
572{
573 if (!td->io_ops->get_file_size)
574 return 0;
575
576 return td->io_ops->get_file_size(td, f);
577}
44f29692 578
de890a1e
SL
579int fio_show_ioengine_help(const char *engine)
580{
581 struct flist_head *entry;
755dcbbd 582 struct ioengine_ops *io_ops;
de890a1e
SL
583 char *sep;
584 int ret = 1;
585
586 if (!engine || !*engine) {
587 log_info("Available IO engines:\n");
588 flist_for_each(entry, &engine_list) {
755dcbbd
TK
589 io_ops = flist_entry(entry, struct ioengine_ops, list);
590 log_info("\t%s\n", io_ops->name);
de890a1e
SL
591 }
592 return 0;
593 }
594 sep = strchr(engine, ',');
595 if (sep) {
596 *sep = 0;
597 sep++;
598 }
599
97bb54c9 600 io_ops = __load_ioengine(engine);
755dcbbd 601 if (!io_ops) {
de890a1e
SL
602 log_info("IO engine %s not found\n", engine);
603 return 1;
604 }
605
755dcbbd
TK
606 if (io_ops->options)
607 ret = show_cmd_help(io_ops->options, sep);
de890a1e 608 else
755dcbbd 609 log_info("IO engine %s has no options\n", io_ops->name);
de890a1e 610
de890a1e
SL
611 return ret;
612}