client: use temp buffer for single output flush for json/disk util
[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 */
ebac4655
JA
12#include <stdlib.h>
13#include <unistd.h>
5c4e1dbc 14#include <string.h>
2866c82d 15#include <dlfcn.h>
ecc314ba 16#include <fcntl.h>
0c6e7517 17#include <assert.h>
8c16d840 18
ebac4655 19#include "fio.h"
7c9b1bce 20#include "diskutil.h"
bfbdd35b 21#include "zbd.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
9b50942e 197 if (td->io_ops_dlhandle) {
565e784d 198 dlclose(td->io_ops_dlhandle);
9b50942e
JA
199 td->io_ops_dlhandle = NULL;
200 }
5f350952 201
84585003 202 td->io_ops = NULL;
b990b5c0 203}
10ba535a 204
de890a1e
SL
205void close_ioengine(struct thread_data *td)
206{
207 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
208
209 if (td->io_ops->cleanup) {
210 td->io_ops->cleanup(td);
565e784d 211 td->io_ops_data = NULL;
de890a1e
SL
212 }
213
214 free_ioengine(td);
215}
216
10ba535a
JA
217int td_io_prep(struct thread_data *td, struct io_u *io_u)
218{
ee56ad50 219 dprint_io_u(io_u, "prep");
7101d9c2
JA
220 fio_ro_check(td, io_u);
221
4d4e80f2 222 lock_file(td, io_u->file, io_u->ddir);
b2bd2bd9 223
2ba1c290
JA
224 if (td->io_ops->prep) {
225 int ret = td->io_ops->prep(td, io_u);
226
e5f9a813
RE
227 dprint(FD_IO, "prep: io_u %p: ret=%d\n", io_u, ret);
228
b2bd2bd9 229 if (ret)
4d4e80f2 230 unlock_file(td, io_u->file);
2ba1c290
JA
231 return ret;
232 }
10ba535a
JA
233
234 return 0;
235}
236
e7d2e616 237int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
1f440ece 238 const struct timespec *t)
10ba535a 239{
ee56ad50 240 int r = 0;
face81b2 241
a05d62b2
YR
242 /*
243 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
244 * server side gets a message from the client
245 * side that the task is finished, and
246 * td->done is set to 1 after td_io_commit(). In this case,
247 * there is no need to reap complete event in server side.
248 */
249 if (td->done)
250 return 0;
251
ee56ad50
JA
252 if (min > 0 && td->io_ops->commit) {
253 r = td->io_ops->commit(td);
face81b2 254 if (r < 0)
ee56ad50 255 goto out;
face81b2 256 }
4950421a
JA
257 if (max > td->cur_depth)
258 max = td->cur_depth;
259 if (min > max)
260 max = min;
36167d82 261
ee56ad50 262 r = 0;
4950421a 263 if (max && td->io_ops->getevents)
ee56ad50
JA
264 r = td->io_ops->getevents(td, min, max, t);
265out:
422f9e4b
RM
266 if (r >= 0) {
267 /*
3fd9efbc 268 * Reflect that our submitted requests were retrieved with
422f9e4b
RM
269 * whatever OS async calls are in the underlying engine.
270 */
271 td->io_u_in_flight -= r;
838bc709 272 io_u_mark_complete(td, r);
422f9e4b 273 } else
7c639b14 274 td_verror(td, r, "get_events");
f3e11d05 275
ee56ad50
JA
276 dprint(FD_IO, "getevents: %d\n", r);
277 return r;
10ba535a
JA
278}
279
d3b07186 280enum fio_q_status td_io_queue(struct thread_data *td, struct io_u *io_u)
10ba535a 281{
a9da8ab2 282 const enum fio_ddir ddir = acct_ddir(io_u);
5fff9543 283 unsigned long long buflen = io_u->xfer_buflen;
d3b07186 284 enum fio_q_status ret;
7e77dd02 285
ee56ad50 286 dprint_io_u(io_u, "queue");
7101d9c2
JA
287 fio_ro_check(td, io_u);
288
0c6e7517 289 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
1651e431 290 io_u_set(td, io_u, IO_U_F_FLIGHT);
0c6e7517 291
d6aed795 292 assert(fio_file_open(io_u->file));
3d7b485f 293
bcd5abfa
JA
294 /*
295 * If using a write iolog, store this entry.
296 */
297 log_io_u(td, io_u);
298
11786802
JA
299 io_u->error = 0;
300 io_u->resid = 0;
301
9b87f09b 302 if (td_ioengine_flagged(td, FIO_SYNCIO)) {
12d9d841 303 if (fio_fill_issue_time(td))
9520ebb9 304 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
305
306 /*
307 * only used for iolog
308 */
309 if (td->o.read_iolog_file)
310 memcpy(&td->last_issue, &io_u->issue_time,
8b6a404c 311 sizeof(io_u->issue_time));
433afcb4
JA
312 }
313
a9da8ab2 314 if (ddir_rw(ddir)) {
a7d01f02
JA
315 if (!(io_u->flags & IO_U_F_VER_LIST)) {
316 td->io_issues[ddir]++;
317 td->io_issue_bytes[ddir] += buflen;
318 }
50a8ce86 319 td->rate_io_issue_bytes[ddir] += buflen;
74d6277f 320 }
755200a3 321
7e77dd02 322 ret = td->io_ops->queue(td, io_u);
99952ca7
BVA
323 if (ret != FIO_Q_BUSY && io_u->post_submit) {
324 io_u->post_submit(io_u, io_u->error == 0);
325 io_u->post_submit = NULL;
326 }
5aeb77df 327
4d4e80f2 328 unlock_file(td, io_u->file);
b2bd2bd9 329
a9da8ab2
JA
330 if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
331 td->io_issues[ddir]--;
332 td->io_issue_bytes[ddir] -= buflen;
50a8ce86 333 td->rate_io_issue_bytes[ddir] -= buflen;
871467d9 334 io_u_clear(td, io_u, IO_U_F_FLIGHT);
cd8a19e6
JA
335 }
336
39a43d34
JA
337 /*
338 * If an error was seen and the io engine didn't propagate it
339 * back to 'td', do so.
340 */
341 if (io_u->error && !td->error)
342 td_verror(td, io_u->error, "td_io_queue");
343
cb211682
JA
344 /*
345 * Add warning for O_DIRECT so that users have an easier time
346 * spotting potentially bad alignment. If this triggers for the first
347 * IO, then it's likely an alignment problem or because the host fs
348 * does not support O_DIRECT
349 */
ff58fced 350 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
cb211682 351 td->o.odirect) {
214ac7e0 352
cb211682 353 log_info("fio: first direct IO errored. File system may not "
78d4a262
TK
354 "support direct IO, or iomem_align= is bad, or "
355 "invalid block size. Try setting direct=0.\n");
cb211682
JA
356 }
357
bfbdd35b
BVA
358 if (zbd_unaligned_write(io_u->error) &&
359 td->io_issues[io_u->ddir & 1] == 1 &&
360 td->o.zone_mode != ZONE_MODE_ZBD) {
361 log_info("fio: first I/O failed. If %s is a zoned block device, consider --zonemode=zbd\n",
362 io_u->file->file_name);
363 }
364
c0681c9d 365 if (!td->io_ops->commit) {
838bc709
JA
366 io_u_mark_submit(td, 1);
367 io_u_mark_complete(td, 1);
368 }
369
d8005759 370 if (ret == FIO_Q_COMPLETED) {
7f3ecee2 371 if (ddir_rw(io_u->ddir) || ddir_sync(io_u->ddir)) {
d8005759
JA
372 io_u_mark_depth(td, 1);
373 td->ts.total_io_u[io_u->ddir]++;
6eaf09d6 374 }
d8005759 375 } else if (ret == FIO_Q_QUEUED) {
e6727cbd
JA
376 td->io_u_queued++;
377
7f3ecee2 378 if (ddir_rw(io_u->ddir) || ddir_sync(io_u->ddir))
d8005759 379 td->ts.total_io_u[io_u->ddir]++;
d8005759 380
a80cb54b
BVA
381 if (td->io_u_queued >= td->o.iodepth_batch)
382 td_io_commit(td);
eb7c8ae2 383 }
cb5ab512 384
9b87f09b 385 if (!td_ioengine_flagged(td, FIO_SYNCIO)) {
12d9d841 386 if (fio_fill_issue_time(td))
9520ebb9 387 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
388
389 /*
390 * only used for iolog
391 */
392 if (td->o.read_iolog_file)
393 memcpy(&td->last_issue, &io_u->issue_time,
8b6a404c 394 sizeof(io_u->issue_time));
433afcb4
JA
395 }
396
7e77dd02 397 return ret;
10ba535a 398}
8c16d840
JA
399
400int td_io_init(struct thread_data *td)
401{
eeb12160 402 int ret = 0;
8c16d840 403
eeb12160
JA
404 if (td->io_ops->init) {
405 ret = td->io_ops->init(td);
356ef1a1
TK
406 if (ret)
407 log_err("fio: io engine %s init failed.%s\n",
408 td->io_ops->name,
409 td->o.iodepth > 1 ?
410 " Perhaps try reducing io depth?" : "");
411 else
412 td->io_ops_init = 1;
7c973896
JA
413 if (!td->error)
414 td->error = ret;
eeb12160
JA
415 }
416
417 return ret;
8c16d840 418}
755200a3 419
a80cb54b 420void td_io_commit(struct thread_data *td)
755200a3 421{
f3e11d05
JA
422 int ret;
423
ee56ad50
JA
424 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
425
d8005759 426 if (!td->cur_depth || !td->io_u_queued)
a80cb54b 427 return;
cb5ab512 428
3fd9efbc 429 io_u_mark_depth(td, td->io_u_queued);
d8005759 430
f3e11d05
JA
431 if (td->io_ops->commit) {
432 ret = td->io_ops->commit(td);
433 if (ret)
434 td_verror(td, -ret, "io commit");
435 }
3fd9efbc 436
422f9e4b
RM
437 /*
438 * Reflect that events were submitted as async IO requests.
439 */
440 td->io_u_in_flight += td->io_u_queued;
441 td->io_u_queued = 0;
755200a3 442}
b5af8293
JA
443
444int td_io_open_file(struct thread_data *td, struct fio_file *f)
445{
230f33fb
AK
446 if (fio_file_closing(f)) {
447 /*
448 * Open translates to undo closing.
449 */
450 fio_file_clear_closing(f);
451 get_file(f);
452 return 0;
453 }
22a57ba8
JA
454 assert(!fio_file_open(f));
455 assert(f->fd == -1);
966fcbd4 456 assert(td->io_ops->open_file);
22a57ba8 457
413d6693
JA
458 if (td->io_ops->open_file(td, f)) {
459 if (td->error == EINVAL && td->o.odirect)
460 log_err("fio: destination does not support O_DIRECT\n");
5ec10eaa
JA
461 if (td->error == EMFILE) {
462 log_err("fio: try reducing/setting openfiles (failed"
463 " at %u of %u)\n", td->nr_open_files,
464 td->o.nr_files);
465 }
413d6693 466
22a57ba8
JA
467 assert(f->fd == -1);
468 assert(!fio_file_open(f));
413d6693
JA
469 return 1;
470 }
471
33c48814 472 fio_file_reset(td, f);
d6aed795
JA
473 fio_file_set_open(f);
474 fio_file_clear_closing(f);
c97bd0fa 475 disk_util_inc(f->du);
d5707a35
JA
476
477 td->nr_open_files++;
478 get_file(f);
479
66159828
JA
480 if (f->filetype == FIO_TYPE_PIPE) {
481 if (td_random(td)) {
482 log_err("fio: can't seek on pipes (no random io)\n");
483 goto err;
484 }
485 }
486
9b87f09b 487 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
413d6693
JA
488 goto done;
489
490 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
491 goto err;
492
ecb2083d 493 if (td->o.fadvise_hint != F_ADV_NONE &&
686fbd31 494 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
413d6693
JA
495 int flags;
496
ecb2083d
JA
497 if (td->o.fadvise_hint == F_ADV_TYPE) {
498 if (td_random(td))
499 flags = POSIX_FADV_RANDOM;
500 else
501 flags = POSIX_FADV_SEQUENTIAL;
502 } else if (td->o.fadvise_hint == F_ADV_RANDOM)
413d6693 503 flags = POSIX_FADV_RANDOM;
ecb2083d 504 else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
413d6693 505 flags = POSIX_FADV_SEQUENTIAL;
ecb2083d
JA
506 else {
507 log_err("fio: unknown fadvise type %d\n",
508 td->o.fadvise_hint);
509 flags = POSIX_FADV_NORMAL;
510 }
413d6693 511
ecc314ba 512 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
825b0a75
JA
513 if (!fio_did_warn(FIO_WARN_FADVISE))
514 log_err("fio: fadvise hint failed\n");
413d6693 515 }
7bb48f84 516 }
ae8e559e
JA
517#ifdef FIO_HAVE_WRITE_HINT
518 if (fio_option_is_set(&td->o, write_hint) &&
686fbd31 519 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
ae8e559e 520 uint64_t hint = td->o.write_hint;
bd553af6 521 int cmd;
37659335 522
bd553af6
JA
523 /*
524 * For direct IO, we just need/want to set the hint on
525 * the file descriptor. For buffered IO, we need to set
526 * it on the inode.
527 */
528 if (td->o.odirect)
529 cmd = F_SET_FILE_RW_HINT;
530 else
531 cmd = F_SET_RW_HINT;
532
533 if (fcntl(f->fd, cmd, &hint) < 0) {
ae8e559e 534 td_verror(td, errno, "fcntl write hint");
37659335
JA
535 goto err;
536 }
537 }
538#endif
a978ba68 539
47534cda
TK
540 if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
541 goto err;
e116f2b9 542
413d6693 543done:
f29b25a3 544 log_file(td, f, FIO_LOG_OPEN_FILE);
413d6693
JA
545 return 0;
546err:
c97bd0fa 547 disk_util_dec(f->du);
b284075a
JA
548 if (td->io_ops->close_file)
549 td->io_ops->close_file(td, f);
7bb48f84 550 return 1;
b5af8293
JA
551}
552
6977bcd0 553int td_io_close_file(struct thread_data *td, struct fio_file *f)
b5af8293 554{
d6aed795 555 if (!fio_file_closing(f))
f29b25a3
JA
556 log_file(td, f, FIO_LOG_CLOSE_FILE);
557
0ad920e7
JA
558 /*
559 * mark as closing, do real close when last io on it has completed
560 */
d6aed795 561 fio_file_set_closing(f);
0ad920e7 562
6977bcd0 563 return put_file(td, f);
b5af8293 564}
df9c26b1 565
38ef9c90
CF
566int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
567{
568 if (td->io_ops->unlink_file)
569 return td->io_ops->unlink_file(td, f);
2442c935
JA
570 else {
571 int ret;
572
573 ret = unlink(f->file_name);
574 if (ret < 0)
575 return errno;
576
577 return 0;
578 }
38ef9c90
CF
579}
580
df9c26b1
JA
581int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
582{
583 if (!td->io_ops->get_file_size)
584 return 0;
585
586 return td->io_ops->get_file_size(td, f);
587}
44f29692 588
de890a1e
SL
589int fio_show_ioengine_help(const char *engine)
590{
591 struct flist_head *entry;
d04f1d5b 592 struct thread_data td;
755dcbbd 593 struct ioengine_ops *io_ops;
de890a1e
SL
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) {
755dcbbd
TK
600 io_ops = flist_entry(entry, struct ioengine_ops, list);
601 log_info("\t%s\n", io_ops->name);
de890a1e
SL
602 }
603 return 0;
604 }
605 sep = strchr(engine, ',');
606 if (sep) {
607 *sep = 0;
608 sep++;
609 }
610
d04f1d5b
JM
611 memset(&td, 0, sizeof(struct thread_data));
612 td.o.ioengine = (char *)engine;
613 io_ops = load_ioengine(&td);
614
755dcbbd 615 if (!io_ops) {
de890a1e
SL
616 log_info("IO engine %s not found\n", engine);
617 return 1;
618 }
619
755dcbbd
TK
620 if (io_ops->options)
621 ret = show_cmd_help(io_ops->options, sep);
de890a1e 622 else
755dcbbd 623 log_info("IO engine %s has no options\n", io_ops->name);
de890a1e 624
d04f1d5b 625 free_ioengine(&td);
de890a1e
SL
626 return ret;
627}