Call path below SIGALRM isn't safe
[fio.git] / ioengines.c
... / ...
CommitLineData
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>
15#include <string.h>
16#include <dlfcn.h>
17#include <assert.h>
18
19#include "fio.h"
20#include "diskutil.h"
21
22static FLIST_HEAD(engine_list);
23
24static int check_engine_ops(struct ioengine_ops *ops)
25{
26 if (ops->version != FIO_IOOPS_VERSION) {
27 log_err("bad ioops version %d (want %d)\n", ops->version,
28 FIO_IOOPS_VERSION);
29 return 1;
30 }
31
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;
42
43 if (!ops->event) {
44 log_err("%s: no event handler\n", ops->name);
45 return 1;
46 }
47 if (!ops->getevents) {
48 log_err("%s: no getevents handler\n", ops->name);
49 return 1;
50 }
51 if (!ops->queue) {
52 log_err("%s: no queue handler\n", ops->name);
53 return 1;
54 }
55
56 return 0;
57}
58
59void unregister_ioengine(struct ioengine_ops *ops)
60{
61 dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
62 flist_del(&ops->list);
63 INIT_FLIST_HEAD(&ops->list);
64}
65
66void register_ioengine(struct ioengine_ops *ops)
67{
68 dprint(FD_IO, "ioengine %s registered\n", ops->name);
69 INIT_FLIST_HEAD(&ops->list);
70 flist_add_tail(&ops->list, &engine_list);
71}
72
73static struct ioengine_ops *find_ioengine(const char *name)
74{
75 struct ioengine_ops *ops;
76 struct flist_head *entry;
77
78 flist_for_each(entry, &engine_list) {
79 ops = flist_entry(entry, struct ioengine_ops, list);
80 if (!strcmp(name, ops->name))
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
93 dprint(FD_IO, "dload engine %s\n", engine_lib);
94
95 dlerror();
96 dlhandle = dlopen(engine_lib, RTLD_LAZY);
97 if (!dlhandle) {
98 td_vmsg(td, -1, dlerror(), "dlopen");
99 return NULL;
100 }
101
102 /*
103 * Unlike the included modules, external engines should have a
104 * non-static ioengine structure that we can reference.
105 */
106 ops = dlsym(dlhandle, "ioengine");
107 if (!ops) {
108 td_vmsg(td, -1, dlerror(), "dlsym");
109 dlclose(dlhandle);
110 return NULL;
111 }
112
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
122 dprint(FD_IO, "load ioengine %s\n", name);
123
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);
138 return NULL;
139 }
140
141 /*
142 * Check that the required methods are there.
143 */
144 if (check_engine_ops(ops))
145 return NULL;
146
147 ret = malloc(sizeof(*ret));
148 memcpy(ret, ops, sizeof(*ret));
149 ret->data = NULL;
150
151 return ret;
152}
153
154void close_ioengine(struct thread_data *td)
155{
156 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
157
158 if (td->io_ops->cleanup) {
159 td->io_ops->cleanup(td);
160 td->io_ops->data = NULL;
161 }
162
163 if (td->io_ops->dlhandle)
164 dlclose(td->io_ops->dlhandle);
165
166 free(td->io_ops);
167 td->io_ops = NULL;
168}
169
170int td_io_prep(struct thread_data *td, struct io_u *io_u)
171{
172 dprint_io_u(io_u, "prep");
173 fio_ro_check(td, io_u);
174
175 lock_file(td, io_u->file, io_u->ddir);
176
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);
181 if (ret)
182 unlock_file(td, io_u->file);
183 return ret;
184 }
185
186 return 0;
187}
188
189int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
190 struct timespec *t)
191{
192 int r = 0;
193
194 if (min > 0 && td->io_ops->commit) {
195 r = td->io_ops->commit(td);
196 if (r < 0)
197 goto out;
198 }
199 if (max > td->cur_depth)
200 max = td->cur_depth;
201 if (min > max)
202 max = min;
203
204 r = 0;
205 if (max && td->io_ops->getevents)
206 r = td->io_ops->getevents(td, min, max, t);
207out:
208 if (r >= 0)
209 io_u_mark_complete(td, r);
210 else
211 td_verror(td, r, "get_events");
212
213 dprint(FD_IO, "getevents: %d\n", r);
214 return r;
215}
216
217int td_io_queue(struct thread_data *td, struct io_u *io_u)
218{
219 int ret;
220
221 dprint_io_u(io_u, "queue");
222 fio_ro_check(td, io_u);
223
224 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
225 io_u->flags |= IO_U_F_FLIGHT;
226
227 assert(fio_file_open(io_u->file));
228
229 io_u->error = 0;
230 io_u->resid = 0;
231
232 if (td->io_ops->flags & FIO_SYNCIO) {
233 if (fio_fill_issue_time(td))
234 fio_gettime(&io_u->issue_time, NULL);
235
236 /*
237 * only used for iolog
238 */
239 if (td->o.read_iolog_file)
240 memcpy(&td->last_issue, &io_u->issue_time,
241 sizeof(struct timeval));
242 }
243
244 if (ddir_rw(io_u->ddir))
245 td->io_issues[io_u->ddir]++;
246
247 ret = td->io_ops->queue(td, io_u);
248
249 unlock_file(td, io_u->file);
250
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] == 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
263 if (!td->io_ops->commit) {
264 io_u_mark_submit(td, 1);
265 io_u_mark_complete(td, 1);
266 }
267
268 if (ret == FIO_Q_COMPLETED) {
269 if (ddir_rw(io_u->ddir)) {
270 io_u_mark_depth(td, 1);
271 td->ts.total_io_u[io_u->ddir]++;
272 } else if (io_u->ddir == DDIR_TRIM)
273 td->ts.total_io_u[2]++;
274 } else if (ret == FIO_Q_QUEUED) {
275 int r;
276
277 if (ddir_rw(io_u->ddir)) {
278 td->io_u_queued++;
279 td->ts.total_io_u[io_u->ddir]++;
280 }
281
282 if (td->io_u_queued >= td->o.iodepth_batch) {
283 r = td_io_commit(td);
284 if (r < 0)
285 return r;
286 }
287 }
288
289 if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
290 if (fio_fill_issue_time(td))
291 fio_gettime(&io_u->issue_time, NULL);
292
293 /*
294 * only used for iolog
295 */
296 if (td->o.read_iolog_file)
297 memcpy(&td->last_issue, &io_u->issue_time,
298 sizeof(struct timeval));
299 }
300
301 return ret;
302}
303
304int td_io_init(struct thread_data *td)
305{
306 int ret = 0;
307
308 if (td->io_ops->init) {
309 ret = td->io_ops->init(td);
310 if (ret && td->o.iodepth > 1) {
311 log_err("fio: io engine init failed. Perhaps try"
312 " reducing io depth?\n");
313 }
314 }
315
316 return ret;
317}
318
319int td_io_commit(struct thread_data *td)
320{
321 int ret;
322
323 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
324
325 if (!td->cur_depth || !td->io_u_queued)
326 return 0;
327
328 io_u_mark_depth(td, td->io_u_queued);
329 td->io_u_queued = 0;
330
331 if (td->io_ops->commit) {
332 ret = td->io_ops->commit(td);
333 if (ret)
334 td_verror(td, -ret, "io commit");
335 }
336
337 return 0;
338}
339
340int td_io_open_file(struct thread_data *td, struct fio_file *f)
341{
342 assert(!fio_file_open(f));
343 assert(f->fd == -1);
344
345 if (td->io_ops->open_file(td, f)) {
346 if (td->error == EINVAL && td->o.odirect)
347 log_err("fio: destination does not support O_DIRECT\n");
348 if (td->error == EMFILE) {
349 log_err("fio: try reducing/setting openfiles (failed"
350 " at %u of %u)\n", td->nr_open_files,
351 td->o.nr_files);
352 }
353
354 assert(f->fd == -1);
355 assert(!fio_file_open(f));
356 return 1;
357 }
358
359 fio_file_reset(f);
360 fio_file_set_open(f);
361 fio_file_clear_closing(f);
362 disk_util_inc(f->du);
363
364 td->nr_open_files++;
365 get_file(f);
366
367 if (f->filetype == FIO_TYPE_PIPE) {
368 if (td_random(td)) {
369 log_err("fio: can't seek on pipes (no random io)\n");
370 goto err;
371 }
372 }
373
374 if (td->io_ops->flags & FIO_DISKLESSIO)
375 goto done;
376
377 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
378 goto err;
379
380 if (td->o.fadvise_hint &&
381 (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
382 int flags;
383
384 if (td_random(td))
385 flags = POSIX_FADV_RANDOM;
386 else
387 flags = POSIX_FADV_SEQUENTIAL;
388
389 if (fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
390 td_verror(td, errno, "fadvise");
391 goto err;
392 }
393 }
394
395#ifdef FIO_OS_DIRECTIO
396 /*
397 * Some OS's have a distinct call to mark the file non-buffered,
398 * instead of using O_DIRECT (Solaris)
399 */
400 if (td->o.odirect) {
401 int ret = fio_set_odirect(f->fd);
402
403 if (ret) {
404 td_verror(td, ret, "fio_set_odirect");
405 goto err;
406 }
407 }
408#endif
409
410done:
411 log_file(td, f, FIO_LOG_OPEN_FILE);
412 return 0;
413err:
414 disk_util_dec(f->du);
415 if (td->io_ops->close_file)
416 td->io_ops->close_file(td, f);
417 return 1;
418}
419
420int td_io_close_file(struct thread_data *td, struct fio_file *f)
421{
422 if (!fio_file_closing(f))
423 log_file(td, f, FIO_LOG_CLOSE_FILE);
424
425 /*
426 * mark as closing, do real close when last io on it has completed
427 */
428 fio_file_set_closing(f);
429
430 disk_util_dec(f->du);
431 unlock_file_all(td, f);
432
433 return put_file(td, f);
434}
435
436int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
437{
438 if (!td->io_ops->get_file_size)
439 return 0;
440
441 return td->io_ops->get_file_size(td, f);
442}
443
444static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
445{
446 off64_t offset, nbytes;
447
448 offset = f->first_write;
449 nbytes = f->last_write - f->first_write;
450
451 if (!nbytes)
452 return 0;
453
454 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
455}
456
457int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
458{
459 int ret;
460
461 if (io_u->ddir == DDIR_SYNC) {
462 ret = fsync(io_u->file->fd);
463 } else if (io_u->ddir == DDIR_DATASYNC) {
464#ifdef FIO_HAVE_FDATASYNC
465 ret = fdatasync(io_u->file->fd);
466#else
467 ret = io_u->xfer_buflen;
468 io_u->error = EINVAL;
469#endif
470 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
471 ret = do_sync_file_range(td, io_u->file);
472 else {
473 ret = io_u->xfer_buflen;
474 io_u->error = EINVAL;
475 }
476
477 if (ret < 0)
478 io_u->error = errno;
479
480 return ret;
481}
482
483int do_io_u_trim(struct thread_data *td, struct io_u *io_u)
484{
485#ifndef FIO_HAVE_TRIM
486 io_u->error = EINVAL;
487 return 0;
488#else
489 struct fio_file *f = io_u->file;
490 int ret;
491
492 ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
493 if (!ret)
494 return io_u->xfer_buflen;;
495
496 io_u->error = ret;
497 return 0;
498#endif
499}