splice: use ->o.memalign to indicate we want aligned buffers
[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
2866c82d 155void close_ioengine(struct thread_data *td)
8756e4d4 156{
ee56ad50
JA
157 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
158
2992b059 159 if (td->io_ops->cleanup) {
2866c82d 160 td->io_ops->cleanup(td);
2992b059
JA
161 td->io_ops->data = NULL;
162 }
b990b5c0 163
5f350952
JA
164 if (td->io_ops->dlhandle)
165 dlclose(td->io_ops->dlhandle);
166
84585003
JA
167 free(td->io_ops);
168 td->io_ops = NULL;
b990b5c0 169}
10ba535a
JA
170
171int td_io_prep(struct thread_data *td, struct io_u *io_u)
172{
ee56ad50 173 dprint_io_u(io_u, "prep");
7101d9c2
JA
174 fio_ro_check(td, io_u);
175
4d4e80f2 176 lock_file(td, io_u->file, io_u->ddir);
b2bd2bd9 177
2ba1c290
JA
178 if (td->io_ops->prep) {
179 int ret = td->io_ops->prep(td, io_u);
180
181 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
b2bd2bd9 182 if (ret)
4d4e80f2 183 unlock_file(td, io_u->file);
2ba1c290
JA
184 return ret;
185 }
10ba535a
JA
186
187 return 0;
188}
189
e7d2e616 190int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
10ba535a
JA
191 struct timespec *t)
192{
ee56ad50 193 int r = 0;
face81b2 194
ee56ad50
JA
195 if (min > 0 && td->io_ops->commit) {
196 r = td->io_ops->commit(td);
face81b2 197 if (r < 0)
ee56ad50 198 goto out;
face81b2 199 }
4950421a
JA
200 if (max > td->cur_depth)
201 max = td->cur_depth;
202 if (min > max)
203 max = min;
36167d82 204
ee56ad50 205 r = 0;
4950421a 206 if (max && td->io_ops->getevents)
ee56ad50
JA
207 r = td->io_ops->getevents(td, min, max, t);
208out:
838bc709
JA
209 if (r >= 0)
210 io_u_mark_complete(td, r);
f3e11d05 211 else
7c639b14 212 td_verror(td, r, "get_events");
f3e11d05 213
ee56ad50
JA
214 dprint(FD_IO, "getevents: %d\n", r);
215 return r;
10ba535a
JA
216}
217
218int td_io_queue(struct thread_data *td, struct io_u *io_u)
219{
7e77dd02
JA
220 int ret;
221
ee56ad50 222 dprint_io_u(io_u, "queue");
7101d9c2
JA
223 fio_ro_check(td, io_u);
224
0c6e7517
JA
225 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
226 io_u->flags |= IO_U_F_FLIGHT;
227
d6aed795 228 assert(fio_file_open(io_u->file));
3d7b485f 229
11786802
JA
230 io_u->error = 0;
231 io_u->resid = 0;
232
433afcb4 233 if (td->io_ops->flags & FIO_SYNCIO) {
12d9d841 234 if (fio_fill_issue_time(td))
9520ebb9 235 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
236
237 /*
238 * only used for iolog
239 */
240 if (td->o.read_iolog_file)
241 memcpy(&td->last_issue, &io_u->issue_time,
5ec10eaa 242 sizeof(struct timeval));
433afcb4
JA
243 }
244
ff58fced 245 if (ddir_rw(io_u->ddir))
755200a3
JA
246 td->io_issues[io_u->ddir]++;
247
7e77dd02 248 ret = td->io_ops->queue(td, io_u);
5aeb77df 249
4d4e80f2 250 unlock_file(td, io_u->file);
b2bd2bd9 251
cb211682
JA
252 /*
253 * Add warning for O_DIRECT so that users have an easier time
254 * spotting potentially bad alignment. If this triggers for the first
255 * IO, then it's likely an alignment problem or because the host fs
256 * does not support O_DIRECT
257 */
ff58fced 258 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
cb211682
JA
259 td->o.odirect) {
260 log_info("fio: first direct IO errored. File system may not "
261 "support direct IO, or iomem_align= is bad.\n");
262 }
263
838bc709
JA
264 if (!td->io_ops->commit) {
265 io_u_mark_submit(td, 1);
266 io_u_mark_complete(td, 1);
267 }
268
d8005759 269 if (ret == FIO_Q_COMPLETED) {
ff58fced 270 if (ddir_rw(io_u->ddir)) {
d8005759
JA
271 io_u_mark_depth(td, 1);
272 td->ts.total_io_u[io_u->ddir]++;
0d29de83
JA
273 } else if (io_u->ddir == DDIR_TRIM)
274 td->ts.total_io_u[2]++;
d8005759 275 } else if (ret == FIO_Q_QUEUED) {
eb7c8ae2
JA
276 int r;
277
ff58fced 278 if (ddir_rw(io_u->ddir)) {
d8005759
JA
279 td->io_u_queued++;
280 td->ts.total_io_u[io_u->ddir]++;
281 }
282
283 if (td->io_u_queued >= td->o.iodepth_batch) {
eb7c8ae2
JA
284 r = td_io_commit(td);
285 if (r < 0)
286 return r;
287 }
288 }
cb5ab512 289
433afcb4 290 if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
12d9d841 291 if (fio_fill_issue_time(td))
9520ebb9 292 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
293
294 /*
295 * only used for iolog
296 */
297 if (td->o.read_iolog_file)
298 memcpy(&td->last_issue, &io_u->issue_time,
299 sizeof(struct timeval));
433afcb4
JA
300 }
301
7e77dd02 302 return ret;
10ba535a 303}
8c16d840
JA
304
305int td_io_init(struct thread_data *td)
306{
eeb12160 307 int ret = 0;
8c16d840 308
eeb12160
JA
309 if (td->io_ops->init) {
310 ret = td->io_ops->init(td);
5ec10eaa
JA
311 if (ret && td->o.iodepth > 1) {
312 log_err("fio: io engine init failed. Perhaps try"
313 " reducing io depth?\n");
314 }
7c973896
JA
315 if (!td->error)
316 td->error = ret;
eeb12160
JA
317 }
318
319 return ret;
8c16d840 320}
755200a3
JA
321
322int td_io_commit(struct thread_data *td)
323{
f3e11d05
JA
324 int ret;
325
ee56ad50
JA
326 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
327
d8005759 328 if (!td->cur_depth || !td->io_u_queued)
e1161c32 329 return 0;
cb5ab512 330
d8005759 331 io_u_mark_depth(td, td->io_u_queued);
cb5ab512 332 td->io_u_queued = 0;
d8005759 333
f3e11d05
JA
334 if (td->io_ops->commit) {
335 ret = td->io_ops->commit(td);
336 if (ret)
337 td_verror(td, -ret, "io commit");
338 }
755200a3
JA
339
340 return 0;
341}
b5af8293
JA
342
343int td_io_open_file(struct thread_data *td, struct fio_file *f)
344{
22a57ba8
JA
345 assert(!fio_file_open(f));
346 assert(f->fd == -1);
347
413d6693
JA
348 if (td->io_ops->open_file(td, f)) {
349 if (td->error == EINVAL && td->o.odirect)
350 log_err("fio: destination does not support O_DIRECT\n");
5ec10eaa
JA
351 if (td->error == EMFILE) {
352 log_err("fio: try reducing/setting openfiles (failed"
353 " at %u of %u)\n", td->nr_open_files,
354 td->o.nr_files);
355 }
413d6693 356
22a57ba8
JA
357 assert(f->fd == -1);
358 assert(!fio_file_open(f));
413d6693
JA
359 return 1;
360 }
361
d5707a35 362 fio_file_reset(f);
d6aed795
JA
363 fio_file_set_open(f);
364 fio_file_clear_closing(f);
c97bd0fa 365 disk_util_inc(f->du);
d5707a35
JA
366
367 td->nr_open_files++;
368 get_file(f);
369
66159828
JA
370 if (f->filetype == FIO_TYPE_PIPE) {
371 if (td_random(td)) {
372 log_err("fio: can't seek on pipes (no random io)\n");
373 goto err;
374 }
375 }
376
413d6693
JA
377 if (td->io_ops->flags & FIO_DISKLESSIO)
378 goto done;
379
380 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
381 goto err;
382
66159828
JA
383 if (td->o.fadvise_hint &&
384 (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
413d6693
JA
385 int flags;
386
387 if (td_random(td))
388 flags = POSIX_FADV_RANDOM;
389 else
390 flags = POSIX_FADV_SEQUENTIAL;
391
ecc314ba 392 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
413d6693
JA
393 td_verror(td, errno, "fadvise");
394 goto err;
395 }
7bb48f84 396 }
a978ba68 397
e116f2b9
JA
398#ifdef FIO_OS_DIRECTIO
399 /*
400 * Some OS's have a distinct call to mark the file non-buffered,
401 * instead of using O_DIRECT (Solaris)
402 */
403 if (td->o.odirect) {
404 int ret = fio_set_odirect(f->fd);
405
406 if (ret) {
407 td_verror(td, ret, "fio_set_odirect");
78e51c72 408 log_err("fio: the file system does not seem to support direct IO\n");
e116f2b9
JA
409 goto err;
410 }
411 }
412#endif
413
413d6693 414done:
f29b25a3 415 log_file(td, f, FIO_LOG_OPEN_FILE);
413d6693
JA
416 return 0;
417err:
c97bd0fa 418 disk_util_dec(f->du);
b284075a
JA
419 if (td->io_ops->close_file)
420 td->io_ops->close_file(td, f);
7bb48f84 421 return 1;
b5af8293
JA
422}
423
6977bcd0 424int td_io_close_file(struct thread_data *td, struct fio_file *f)
b5af8293 425{
d6aed795 426 if (!fio_file_closing(f))
f29b25a3
JA
427 log_file(td, f, FIO_LOG_CLOSE_FILE);
428
0ad920e7
JA
429 /*
430 * mark as closing, do real close when last io on it has completed
431 */
d6aed795 432 fio_file_set_closing(f);
0ad920e7 433
c97bd0fa 434 disk_util_dec(f->du);
4d4e80f2 435 unlock_file_all(td, f);
29c1349f 436
6977bcd0 437 return put_file(td, f);
b5af8293 438}
df9c26b1
JA
439
440int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
441{
442 if (!td->io_ops->get_file_size)
443 return 0;
444
445 return td->io_ops->get_file_size(td, f);
446}
44f29692 447
0a28ecda 448static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
44f29692
JA
449{
450 off64_t offset, nbytes;
451
452 offset = f->first_write;
453 nbytes = f->last_write - f->first_write;
454
3843deb3
JA
455 if (!nbytes)
456 return 0;
44f29692 457
3843deb3 458 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
44f29692 459}
0a28ecda
JA
460
461int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
462{
463 int ret;
464
465 if (io_u->ddir == DDIR_SYNC) {
466 ret = fsync(io_u->file->fd);
467 } else if (io_u->ddir == DDIR_DATASYNC) {
468#ifdef FIO_HAVE_FDATASYNC
469 ret = fdatasync(io_u->file->fd);
470#else
471 ret = io_u->xfer_buflen;
472 io_u->error = EINVAL;
473#endif
474 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
475 ret = do_sync_file_range(td, io_u->file);
476 else {
477 ret = io_u->xfer_buflen;
478 io_u->error = EINVAL;
479 }
480
cb849a79
JA
481 if (ret < 0)
482 io_u->error = errno;
483
0a28ecda
JA
484 return ret;
485}
a5f3027c
JA
486
487int do_io_u_trim(struct thread_data *td, struct io_u *io_u)
488{
489#ifndef FIO_HAVE_TRIM
490 io_u->error = EINVAL;
ff58fced 491 return 0;
a5f3027c
JA
492#else
493 struct fio_file *f = io_u->file;
494 int ret;
495
c6404a44 496 ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
a5f3027c 497 if (!ret)
ff58fced 498 return io_u->xfer_buflen;;
a5f3027c 499
ff58fced
JA
500 io_u->error = ret;
501 return 0;
a5f3027c
JA
502#endif
503}