OSX should use off_t for off64_t
[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>
0c6e7517 17#include <assert.h>
8c16d840 18
ebac4655 19#include "fio.h"
7c9b1bce 20#include "diskutil.h"
ebac4655 21
01743ee1 22static FLIST_HEAD(engine_list);
5f350952 23
8c16d840
JA
24static int check_engine_ops(struct ioengine_ops *ops)
25{
5f350952 26 if (ops->version != FIO_IOOPS_VERSION) {
5ec10eaa
JA
27 log_err("bad ioops version %d (want %d)\n", ops->version,
28 FIO_IOOPS_VERSION);
5f350952
JA
29 return 1;
30 }
31
36167d82
JA
32 if (!ops->queue) {
33 log_err("%s: no queue handler\n", ops->name);
34 return 1;
35 }
36
37 /*
38 * sync engines only need a ->queue()
39 */
40 if (ops->flags & FIO_SYNCIO)
41 return 0;
5ec10eaa 42
8c16d840 43 if (!ops->event) {
36167d82 44 log_err("%s: no event handler\n", ops->name);
8c16d840
JA
45 return 1;
46 }
47 if (!ops->getevents) {
36167d82 48 log_err("%s: no getevents handler\n", ops->name);
8c16d840
JA
49 return 1;
50 }
51 if (!ops->queue) {
36167d82 52 log_err("%s: no queue handler\n", ops->name);
8c16d840
JA
53 return 1;
54 }
5ec10eaa 55
8c16d840
JA
56 return 0;
57}
58
5f350952 59void unregister_ioengine(struct ioengine_ops *ops)
ebac4655 60{
ee56ad50 61 dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
01743ee1
JA
62 flist_del(&ops->list);
63 INIT_FLIST_HEAD(&ops->list);
5f350952
JA
64}
65
b2fdda43 66void register_ioengine(struct ioengine_ops *ops)
5f350952 67{
ee56ad50 68 dprint(FD_IO, "ioengine %s registered\n", ops->name);
01743ee1
JA
69 INIT_FLIST_HEAD(&ops->list);
70 flist_add_tail(&ops->list, &engine_list);
5f350952
JA
71}
72
73static struct ioengine_ops *find_ioengine(const char *name)
74{
75 struct ioengine_ops *ops;
01743ee1 76 struct flist_head *entry;
ebac4655 77
01743ee1
JA
78 flist_for_each(entry, &engine_list) {
79 ops = flist_entry(entry, struct ioengine_ops, list);
bc5b77a8 80 if (!strcmp(name, ops->name))
5f350952
JA
81 return ops;
82 }
83
84 return NULL;
85}
86
87static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
88 const char *engine_lib)
89{
90 struct ioengine_ops *ops;
91 void *dlhandle;
92
ee56ad50
JA
93 dprint(FD_IO, "dload engine %s\n", engine_lib);
94
2866c82d
JA
95 dlerror();
96 dlhandle = dlopen(engine_lib, RTLD_LAZY);
d4dbaaa8 97 if (!dlhandle) {
e1161c32 98 td_vmsg(td, -1, dlerror(), "dlopen");
d4dbaaa8
JA
99 return NULL;
100 }
8756e4d4 101
da51c050
JA
102 /*
103 * Unlike the included modules, external engines should have a
104 * non-static ioengine structure that we can reference.
105 */
2866c82d 106 ops = dlsym(dlhandle, "ioengine");
d4dbaaa8 107 if (!ops) {
e1161c32 108 td_vmsg(td, -1, dlerror(), "dlsym");
d4dbaaa8
JA
109 dlclose(dlhandle);
110 return NULL;
111 }
8756e4d4 112
5f350952
JA
113 ops->dlhandle = dlhandle;
114 return ops;
115}
116
117struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
118{
119 struct ioengine_ops *ops, *ret;
120 char engine[16];
121
ee56ad50
JA
122 dprint(FD_IO, "load ioengine %s\n", name);
123
5f350952
JA
124 strncpy(engine, name, sizeof(engine) - 1);
125
126 /*
127 * linux libaio has alias names, so convert to what we want
128 */
129 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
130 strcpy(engine, "libaio");
131
132 ops = find_ioengine(engine);
133 if (!ops)
134 ops = dlopen_ioengine(td, name);
135
136 if (!ops) {
137 log_err("fio: engine %s not loadable\n", name);
b902ceb5
JA
138 return NULL;
139 }
140
8c16d840
JA
141 /*
142 * Check that the required methods are there.
143 */
5f350952 144 if (check_engine_ops(ops))
8c16d840 145 return NULL;
8c16d840 146
84585003
JA
147 ret = malloc(sizeof(*ret));
148 memcpy(ret, ops, sizeof(*ret));
149 ret->data = NULL;
84585003
JA
150
151 return ret;
8756e4d4
JA
152}
153
2866c82d 154void close_ioengine(struct thread_data *td)
8756e4d4 155{
ee56ad50
JA
156 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
157
2992b059 158 if (td->io_ops->cleanup) {
2866c82d 159 td->io_ops->cleanup(td);
2992b059
JA
160 td->io_ops->data = NULL;
161 }
b990b5c0 162
5f350952
JA
163 if (td->io_ops->dlhandle)
164 dlclose(td->io_ops->dlhandle);
165
84585003
JA
166 free(td->io_ops);
167 td->io_ops = NULL;
b990b5c0 168}
10ba535a
JA
169
170int td_io_prep(struct thread_data *td, struct io_u *io_u)
171{
ee56ad50 172 dprint_io_u(io_u, "prep");
7101d9c2
JA
173 fio_ro_check(td, io_u);
174
4d4e80f2 175 lock_file(td, io_u->file, io_u->ddir);
b2bd2bd9 176
2ba1c290
JA
177 if (td->io_ops->prep) {
178 int ret = td->io_ops->prep(td, io_u);
179
180 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
b2bd2bd9 181 if (ret)
4d4e80f2 182 unlock_file(td, io_u->file);
2ba1c290
JA
183 return ret;
184 }
10ba535a
JA
185
186 return 0;
187}
188
e7d2e616 189int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
10ba535a
JA
190 struct timespec *t)
191{
ee56ad50 192 int r = 0;
face81b2 193
ee56ad50
JA
194 if (min > 0 && td->io_ops->commit) {
195 r = td->io_ops->commit(td);
face81b2 196 if (r < 0)
ee56ad50 197 goto out;
face81b2 198 }
4950421a
JA
199 if (max > td->cur_depth)
200 max = td->cur_depth;
201 if (min > max)
202 max = min;
36167d82 203
ee56ad50 204 r = 0;
4950421a 205 if (max && td->io_ops->getevents)
ee56ad50
JA
206 r = td->io_ops->getevents(td, min, max, t);
207out:
838bc709
JA
208 if (r >= 0)
209 io_u_mark_complete(td, r);
f3e11d05 210 else
7c639b14 211 td_verror(td, r, "get_events");
f3e11d05 212
ee56ad50
JA
213 dprint(FD_IO, "getevents: %d\n", r);
214 return r;
10ba535a
JA
215}
216
217int td_io_queue(struct thread_data *td, struct io_u *io_u)
218{
7e77dd02
JA
219 int ret;
220
ee56ad50 221 dprint_io_u(io_u, "queue");
7101d9c2
JA
222 fio_ro_check(td, io_u);
223
0c6e7517
JA
224 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
225 io_u->flags |= IO_U_F_FLIGHT;
226
d6aed795 227 assert(fio_file_open(io_u->file));
3d7b485f 228
11786802
JA
229 io_u->error = 0;
230 io_u->resid = 0;
231
433afcb4 232 if (td->io_ops->flags & FIO_SYNCIO) {
12d9d841 233 if (fio_fill_issue_time(td))
9520ebb9 234 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
235
236 /*
237 * only used for iolog
238 */
239 if (td->o.read_iolog_file)
240 memcpy(&td->last_issue, &io_u->issue_time,
5ec10eaa 241 sizeof(struct timeval));
433afcb4
JA
242 }
243
5f9099ea 244 if (!ddir_sync(io_u->ddir))
755200a3
JA
245 td->io_issues[io_u->ddir]++;
246
7e77dd02 247 ret = td->io_ops->queue(td, io_u);
5aeb77df 248
4d4e80f2 249 unlock_file(td, io_u->file);
b2bd2bd9 250
cb211682
JA
251 /*
252 * Add warning for O_DIRECT so that users have an easier time
253 * spotting potentially bad alignment. If this triggers for the first
254 * IO, then it's likely an alignment problem or because the host fs
255 * does not support O_DIRECT
256 */
257 if (io_u->error == EINVAL && td->io_issues[io_u->ddir] == 1 &&
258 td->o.odirect) {
259 log_info("fio: first direct IO errored. File system may not "
260 "support direct IO, or iomem_align= is bad.\n");
261 }
262
838bc709
JA
263 if (!td->io_ops->commit) {
264 io_u_mark_submit(td, 1);
265 io_u_mark_complete(td, 1);
266 }
267
d8005759 268 if (ret == FIO_Q_COMPLETED) {
5f9099ea 269 if (!ddir_sync(io_u->ddir)) {
d8005759
JA
270 io_u_mark_depth(td, 1);
271 td->ts.total_io_u[io_u->ddir]++;
272 }
273 } else if (ret == FIO_Q_QUEUED) {
eb7c8ae2
JA
274 int r;
275
5f9099ea 276 if (!ddir_sync(io_u->ddir)) {
d8005759
JA
277 td->io_u_queued++;
278 td->ts.total_io_u[io_u->ddir]++;
279 }
280
281 if (td->io_u_queued >= td->o.iodepth_batch) {
eb7c8ae2
JA
282 r = td_io_commit(td);
283 if (r < 0)
284 return r;
285 }
286 }
cb5ab512 287
433afcb4 288 if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
12d9d841 289 if (fio_fill_issue_time(td))
9520ebb9 290 fio_gettime(&io_u->issue_time, NULL);
d0c15328
JA
291
292 /*
293 * only used for iolog
294 */
295 if (td->o.read_iolog_file)
296 memcpy(&td->last_issue, &io_u->issue_time,
297 sizeof(struct timeval));
433afcb4
JA
298 }
299
7e77dd02 300 return ret;
10ba535a 301}
8c16d840
JA
302
303int td_io_init(struct thread_data *td)
304{
eeb12160 305 int ret = 0;
8c16d840 306
eeb12160
JA
307 if (td->io_ops->init) {
308 ret = td->io_ops->init(td);
5ec10eaa
JA
309 if (ret && td->o.iodepth > 1) {
310 log_err("fio: io engine init failed. Perhaps try"
311 " reducing io depth?\n");
312 }
eeb12160
JA
313 }
314
315 return ret;
8c16d840 316}
755200a3
JA
317
318int td_io_commit(struct thread_data *td)
319{
f3e11d05
JA
320 int ret;
321
ee56ad50
JA
322 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
323
d8005759 324 if (!td->cur_depth || !td->io_u_queued)
e1161c32 325 return 0;
cb5ab512 326
d8005759 327 io_u_mark_depth(td, td->io_u_queued);
cb5ab512 328 td->io_u_queued = 0;
d8005759 329
f3e11d05
JA
330 if (td->io_ops->commit) {
331 ret = td->io_ops->commit(td);
332 if (ret)
333 td_verror(td, -ret, "io commit");
334 }
755200a3
JA
335
336 return 0;
337}
b5af8293
JA
338
339int td_io_open_file(struct thread_data *td, struct fio_file *f)
340{
22a57ba8
JA
341 assert(!fio_file_open(f));
342 assert(f->fd == -1);
343
413d6693
JA
344 if (td->io_ops->open_file(td, f)) {
345 if (td->error == EINVAL && td->o.odirect)
346 log_err("fio: destination does not support O_DIRECT\n");
5ec10eaa
JA
347 if (td->error == EMFILE) {
348 log_err("fio: try reducing/setting openfiles (failed"
349 " at %u of %u)\n", td->nr_open_files,
350 td->o.nr_files);
351 }
413d6693 352
22a57ba8
JA
353 assert(f->fd == -1);
354 assert(!fio_file_open(f));
413d6693
JA
355 return 1;
356 }
357
d5707a35 358 fio_file_reset(f);
d6aed795
JA
359 fio_file_set_open(f);
360 fio_file_clear_closing(f);
c97bd0fa 361 disk_util_inc(f->du);
d5707a35
JA
362
363 td->nr_open_files++;
364 get_file(f);
365
66159828
JA
366 if (f->filetype == FIO_TYPE_PIPE) {
367 if (td_random(td)) {
368 log_err("fio: can't seek on pipes (no random io)\n");
369 goto err;
370 }
371 }
372
413d6693
JA
373 if (td->io_ops->flags & FIO_DISKLESSIO)
374 goto done;
375
376 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
377 goto err;
378
66159828
JA
379 if (td->o.fadvise_hint &&
380 (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
413d6693
JA
381 int flags;
382
383 if (td_random(td))
384 flags = POSIX_FADV_RANDOM;
385 else
386 flags = POSIX_FADV_SEQUENTIAL;
387
388 if (fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
389 td_verror(td, errno, "fadvise");
390 goto err;
391 }
7bb48f84 392 }
a978ba68 393
e116f2b9
JA
394#ifdef FIO_OS_DIRECTIO
395 /*
396 * Some OS's have a distinct call to mark the file non-buffered,
397 * instead of using O_DIRECT (Solaris)
398 */
399 if (td->o.odirect) {
400 int ret = fio_set_odirect(f->fd);
401
402 if (ret) {
403 td_verror(td, ret, "fio_set_odirect");
404 goto err;
405 }
406 }
407#endif
408
413d6693 409done:
f29b25a3 410 log_file(td, f, FIO_LOG_OPEN_FILE);
413d6693
JA
411 return 0;
412err:
c97bd0fa 413 disk_util_dec(f->du);
b284075a
JA
414 if (td->io_ops->close_file)
415 td->io_ops->close_file(td, f);
7bb48f84 416 return 1;
b5af8293
JA
417}
418
6977bcd0 419int td_io_close_file(struct thread_data *td, struct fio_file *f)
b5af8293 420{
d6aed795 421 if (!fio_file_closing(f))
f29b25a3
JA
422 log_file(td, f, FIO_LOG_CLOSE_FILE);
423
0ad920e7
JA
424 /*
425 * mark as closing, do real close when last io on it has completed
426 */
d6aed795 427 fio_file_set_closing(f);
0ad920e7 428
c97bd0fa 429 disk_util_dec(f->du);
4d4e80f2 430 unlock_file_all(td, f);
29c1349f 431
6977bcd0 432 return put_file(td, f);
b5af8293 433}
df9c26b1
JA
434
435int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
436{
437 if (!td->io_ops->get_file_size)
438 return 0;
439
440 return td->io_ops->get_file_size(td, f);
441}
44f29692 442
0a28ecda 443static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
44f29692
JA
444{
445 off64_t offset, nbytes;
446
447 offset = f->first_write;
448 nbytes = f->last_write - f->first_write;
449
3843deb3
JA
450 if (!nbytes)
451 return 0;
44f29692 452
3843deb3 453 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
44f29692 454}
0a28ecda
JA
455
456int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
457{
458 int ret;
459
460 if (io_u->ddir == DDIR_SYNC) {
461 ret = fsync(io_u->file->fd);
462 } else if (io_u->ddir == DDIR_DATASYNC) {
463#ifdef FIO_HAVE_FDATASYNC
464 ret = fdatasync(io_u->file->fd);
465#else
466 ret = io_u->xfer_buflen;
467 io_u->error = EINVAL;
468#endif
469 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
470 ret = do_sync_file_range(td, io_u->file);
471 else {
472 ret = io_u->xfer_buflen;
473 io_u->error = EINVAL;
474 }
475
cb849a79
JA
476 if (ret < 0)
477 io_u->error = errno;
478
0a28ecda
JA
479 return ret;
480}