hash: cleanups
[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 <stdlib.h>
13#include <unistd.h>
14#include <string.h>
15#include <dlfcn.h>
16#include <fcntl.h>
17#include <assert.h>
18#include <sys/types.h>
19#include <dirent.h>
20
21#include "fio.h"
22#include "diskutil.h"
23#include "zbd.h"
24
25static FLIST_HEAD(engine_list);
26
27static inline bool async_ioengine_sync_trim(struct thread_data *td,
28 struct io_u *io_u)
29{
30 return td_ioengine_flagged(td, FIO_ASYNCIO_SYNC_TRIM) &&
31 io_u->ddir == DDIR_TRIM;
32}
33
34static bool check_engine_ops(struct thread_data *td, struct ioengine_ops *ops)
35{
36 if (ops->version != FIO_IOOPS_VERSION) {
37 log_err("bad ioops version %d (want %d)\n", ops->version,
38 FIO_IOOPS_VERSION);
39 return true;
40 }
41
42 if (!ops->queue) {
43 log_err("%s: no queue handler\n", ops->name);
44 return true;
45 }
46
47 /*
48 * sync engines only need a ->queue()
49 */
50 if (ops->flags & FIO_SYNCIO)
51 return false;
52
53 /*
54 * async engines aren't reliable with offload
55 */
56 if ((td->o.io_submit_mode == IO_MODE_OFFLOAD) &&
57 (ops->flags & FIO_NO_OFFLOAD)) {
58 log_err("%s: can't be used with offloaded submit. Use a sync "
59 "engine\n", ops->name);
60 return true;
61 }
62
63 if (!ops->event || !ops->getevents) {
64 log_err("%s: no event/getevents handler\n", ops->name);
65 return true;
66 }
67
68 return false;
69}
70
71void unregister_ioengine(struct ioengine_ops *ops)
72{
73 dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
74 flist_del_init(&ops->list);
75}
76
77void register_ioengine(struct ioengine_ops *ops)
78{
79 dprint(FD_IO, "ioengine %s registered\n", ops->name);
80 flist_add_tail(&ops->list, &engine_list);
81}
82
83static struct ioengine_ops *find_ioengine(const char *name)
84{
85 struct ioengine_ops *ops;
86 struct flist_head *entry;
87
88 flist_for_each(entry, &engine_list) {
89 ops = flist_entry(entry, struct ioengine_ops, list);
90 if (!strcmp(name, ops->name))
91 return ops;
92 }
93
94 return NULL;
95}
96
97#ifdef CONFIG_DYNAMIC_ENGINES
98static void *dlopen_external(struct thread_data *td, const char *engine)
99{
100 char engine_path[PATH_MAX];
101 void *dlhandle;
102
103 sprintf(engine_path, "%s/fio-%s.so", FIO_EXT_ENG_DIR, engine);
104
105 dprint(FD_IO, "dlopen external %s\n", engine_path);
106 dlhandle = dlopen(engine_path, RTLD_LAZY);
107 if (!dlhandle)
108 log_info("Engine %s not found; Either name is invalid, was not built, or fio-engine-%s package is missing.\n",
109 engine, engine);
110
111 return dlhandle;
112}
113#else
114#define dlopen_external(td, engine) (NULL)
115#endif
116
117static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
118 const char *engine_lib)
119{
120 struct ioengine_ops *ops;
121 void *dlhandle;
122
123 if (!strncmp(engine_lib, "linuxaio", 8) ||
124 !strncmp(engine_lib, "aio", 3))
125 engine_lib = "libaio";
126
127 dprint(FD_IO, "dlopen engine %s\n", engine_lib);
128
129 dlerror();
130 dlhandle = dlopen(engine_lib, RTLD_LAZY);
131 if (!dlhandle) {
132 dlhandle = dlopen_external(td, engine_lib);
133 if (!dlhandle) {
134 td_vmsg(td, -1, dlerror(), "dlopen");
135 return NULL;
136 }
137 }
138
139 /*
140 * Unlike the included modules, external engines should have a
141 * non-static ioengine structure that we can reference.
142 */
143 ops = dlsym(dlhandle, engine_lib);
144 if (!ops)
145 ops = dlsym(dlhandle, "ioengine");
146
147 /*
148 * For some external engines (like C++ ones) it is not that trivial
149 * to provide a non-static ionengine structure that we can reference.
150 * Instead we call a method which allocates the required ioengine
151 * structure.
152 */
153 if (!ops) {
154 get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
155
156 if (get_ioengine)
157 get_ioengine(&ops);
158 }
159
160 if (!ops) {
161 td_vmsg(td, -1, dlerror(), "dlsym");
162 dlclose(dlhandle);
163 return NULL;
164 }
165
166 ops->dlhandle = dlhandle;
167 return ops;
168}
169
170static struct ioengine_ops *__load_ioengine(const char *engine)
171{
172 /*
173 * linux libaio has alias names, so convert to what we want
174 */
175 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3)) {
176 dprint(FD_IO, "converting ioengine name: %s -> libaio\n",
177 engine);
178 engine = "libaio";
179 }
180
181 dprint(FD_IO, "load ioengine %s\n", engine);
182 return find_ioengine(engine);
183}
184
185struct ioengine_ops *load_ioengine(struct thread_data *td)
186{
187 struct ioengine_ops *ops = NULL;
188 const char *name;
189
190 /*
191 * Use ->ioengine_so_path if an external ioengine path is specified.
192 * In this case, ->ioengine is "external" which also means the prefix
193 * for external ioengines "external:" is properly used.
194 */
195 name = td->o.ioengine_so_path ?: td->o.ioengine;
196
197 /*
198 * Try to load ->ioengine first, and if failed try to dlopen(3) either
199 * ->ioengine or ->ioengine_so_path. This is redundant for an external
200 * ioengine with prefix, and also leaves the possibility of unexpected
201 * behavior (e.g. if the "external" ioengine exists), but we do this
202 * so as not to break job files not using the prefix.
203 */
204 ops = __load_ioengine(td->o.ioengine);
205
206 /* We do re-dlopen existing handles, for reference counting */
207 if (!ops || ops->dlhandle)
208 ops = dlopen_ioengine(td, name);
209
210 /*
211 * If ops is NULL, we failed to load ->ioengine, and also failed to
212 * dlopen(3) either ->ioengine or ->ioengine_so_path as a path.
213 */
214 if (!ops) {
215 log_err("fio: engine %s not loadable\n", name);
216 return NULL;
217 }
218
219 /*
220 * Check that the required methods are there.
221 */
222 if (check_engine_ops(td, ops))
223 return NULL;
224
225 return ops;
226}
227
228/*
229 * For cleaning up an ioengine which never made it to init().
230 */
231void free_ioengine(struct thread_data *td)
232{
233 assert(td != NULL && td->io_ops != NULL);
234
235 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
236
237 if (td->eo && td->io_ops->options) {
238 options_free(td->io_ops->options, td->eo);
239 free(td->eo);
240 td->eo = NULL;
241 }
242
243 if (td->io_ops->dlhandle) {
244 dprint(FD_IO, "dlclose ioengine %s\n", td->io_ops->name);
245 dlclose(td->io_ops->dlhandle);
246 }
247
248 td->io_ops = NULL;
249}
250
251void close_ioengine(struct thread_data *td)
252{
253 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
254
255 if (td->io_ops->cleanup) {
256 td->io_ops->cleanup(td);
257 td->io_ops_data = NULL;
258 }
259
260 free_ioengine(td);
261}
262
263int td_io_prep(struct thread_data *td, struct io_u *io_u)
264{
265 dprint_io_u(io_u, "prep");
266 fio_ro_check(td, io_u);
267
268 lock_file(td, io_u->file, io_u->ddir);
269
270 if (td->io_ops->prep) {
271 int ret = td->io_ops->prep(td, io_u);
272
273 dprint(FD_IO, "prep: io_u %p: ret=%d\n", io_u, ret);
274
275 if (ret)
276 unlock_file(td, io_u->file);
277 return ret;
278 }
279
280 return 0;
281}
282
283int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
284 const struct timespec *t)
285{
286 int r = 0;
287
288 /*
289 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
290 * server side gets a message from the client
291 * side that the task is finished, and
292 * td->done is set to 1 after td_io_commit(). In this case,
293 * there is no need to reap complete event in server side.
294 */
295 if (td->done)
296 return 0;
297
298 if (min > 0 && td->io_ops->commit) {
299 r = td->io_ops->commit(td);
300 if (r < 0)
301 goto out;
302 }
303 if (max > td->cur_depth)
304 max = td->cur_depth;
305 if (min > max)
306 max = min;
307
308 r = 0;
309 if (max && td->io_ops->getevents)
310 r = td->io_ops->getevents(td, min, max, t);
311out:
312 if (r >= 0) {
313 /*
314 * Reflect that our submitted requests were retrieved with
315 * whatever OS async calls are in the underlying engine.
316 */
317 td->io_u_in_flight -= r;
318 io_u_mark_complete(td, r);
319 } else
320 td_verror(td, r, "get_events");
321
322 dprint(FD_IO, "getevents: %d\n", r);
323 return r;
324}
325
326enum fio_q_status td_io_queue(struct thread_data *td, struct io_u *io_u)
327{
328 const enum fio_ddir ddir = acct_ddir(io_u);
329 unsigned long long buflen = io_u->xfer_buflen;
330 enum fio_q_status ret;
331
332 dprint_io_u(io_u, "queue");
333 fio_ro_check(td, io_u);
334
335 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
336 io_u_set(td, io_u, IO_U_F_FLIGHT);
337
338 /*
339 * If overlap checking was enabled in offload mode we
340 * can release this lock that was acquired when we
341 * started the overlap check because the IO_U_F_FLIGHT
342 * flag is now set
343 */
344 if (td_offload_overlap(td)) {
345 int res = pthread_mutex_unlock(&overlap_check);
346 assert(res == 0);
347 }
348
349 assert(fio_file_open(io_u->file));
350
351 /*
352 * If using a write iolog, store this entry.
353 */
354 log_io_u(td, io_u);
355
356 io_u->error = 0;
357 io_u->resid = 0;
358
359 if (td_ioengine_flagged(td, FIO_SYNCIO) ||
360 async_ioengine_sync_trim(td, io_u)) {
361 if (fio_fill_issue_time(td)) {
362 fio_gettime(&io_u->issue_time, NULL);
363
364 /*
365 * only used for iolog
366 */
367 if (td->o.read_iolog_file)
368 memcpy(&td->last_issue, &io_u->issue_time,
369 sizeof(io_u->issue_time));
370 }
371 }
372
373
374 if (ddir_rw(ddir)) {
375 if (!(io_u->flags & IO_U_F_VER_LIST)) {
376 td->io_issues[ddir]++;
377 td->io_issue_bytes[ddir] += buflen;
378 }
379 td->rate_io_issue_bytes[ddir] += buflen;
380 }
381
382 ret = td->io_ops->queue(td, io_u);
383 zbd_queue_io_u(td, io_u, ret);
384
385 unlock_file(td, io_u->file);
386
387 if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
388 td->io_issues[ddir]--;
389 td->io_issue_bytes[ddir] -= buflen;
390 td->rate_io_issue_bytes[ddir] -= buflen;
391 io_u_clear(td, io_u, IO_U_F_FLIGHT);
392 }
393
394 /*
395 * If an error was seen and the io engine didn't propagate it
396 * back to 'td', do so.
397 */
398 if (io_u->error && !td->error)
399 td_verror(td, io_u->error, "td_io_queue");
400
401 /*
402 * Add warning for O_DIRECT so that users have an easier time
403 * spotting potentially bad alignment. If this triggers for the first
404 * IO, then it's likely an alignment problem or because the host fs
405 * does not support O_DIRECT
406 */
407 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
408 td->o.odirect) {
409
410 log_info("fio: first direct IO errored. File system may not "
411 "support direct IO, or iomem_align= is bad, or "
412 "invalid block size. Try setting direct=0.\n");
413 }
414
415 if (zbd_unaligned_write(io_u->error) &&
416 td->io_issues[io_u->ddir & 1] == 1 &&
417 td->o.zone_mode != ZONE_MODE_ZBD) {
418 log_info("fio: first I/O failed. If %s is a zoned block device, consider --zonemode=zbd\n",
419 io_u->file->file_name);
420 }
421
422 if (!td->io_ops->commit) {
423 io_u_mark_submit(td, 1);
424 io_u_mark_complete(td, 1);
425 }
426
427 if (ret == FIO_Q_COMPLETED) {
428 if (ddir_rw(io_u->ddir) ||
429 (ddir_sync(io_u->ddir) && td->runstate != TD_FSYNCING)) {
430 io_u_mark_depth(td, 1);
431 td->ts.total_io_u[io_u->ddir]++;
432 }
433 } else if (ret == FIO_Q_QUEUED) {
434 td->io_u_queued++;
435
436 if (ddir_rw(io_u->ddir) ||
437 (ddir_sync(io_u->ddir) && td->runstate != TD_FSYNCING))
438 td->ts.total_io_u[io_u->ddir]++;
439
440 if (td->io_u_queued >= td->o.iodepth_batch)
441 td_io_commit(td);
442 }
443
444 if (!td_ioengine_flagged(td, FIO_SYNCIO) &&
445 !async_ioengine_sync_trim(td, io_u)) {
446 if (fio_fill_issue_time(td) &&
447 !td_ioengine_flagged(td, FIO_ASYNCIO_SETS_ISSUE_TIME)) {
448 fio_gettime(&io_u->issue_time, NULL);
449
450 /*
451 * only used for iolog
452 */
453 if (td->o.read_iolog_file)
454 memcpy(&td->last_issue, &io_u->issue_time,
455 sizeof(io_u->issue_time));
456 }
457 }
458
459 return ret;
460}
461
462int td_io_init(struct thread_data *td)
463{
464 int ret = 0;
465
466 if (td->io_ops->init) {
467 ret = td->io_ops->init(td);
468 if (ret)
469 log_err("fio: io engine %s init failed.%s\n",
470 td->io_ops->name,
471 td->o.iodepth > 1 ?
472 " Perhaps try reducing io depth?" : "");
473 else
474 td->io_ops_init = 1;
475 if (!td->error)
476 td->error = ret;
477 }
478
479 return ret;
480}
481
482void td_io_commit(struct thread_data *td)
483{
484 int ret;
485
486 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
487
488 if (!td->cur_depth || !td->io_u_queued)
489 return;
490
491 io_u_mark_depth(td, td->io_u_queued);
492
493 if (td->io_ops->commit) {
494 ret = td->io_ops->commit(td);
495 if (ret)
496 td_verror(td, -ret, "io commit");
497 }
498
499 /*
500 * Reflect that events were submitted as async IO requests.
501 */
502 td->io_u_in_flight += td->io_u_queued;
503 td->io_u_queued = 0;
504}
505
506int td_io_open_file(struct thread_data *td, struct fio_file *f)
507{
508 if (fio_file_closing(f)) {
509 /*
510 * Open translates to undo closing.
511 */
512 fio_file_clear_closing(f);
513 get_file(f);
514 return 0;
515 }
516 assert(!fio_file_open(f));
517 assert(f->fd == -1);
518 assert(td->io_ops->open_file);
519
520 if (td->io_ops->open_file(td, f)) {
521 if (td->error == EINVAL && td->o.odirect)
522 log_err("fio: destination does not support O_DIRECT\n");
523 if (td->error == EMFILE) {
524 log_err("fio: try reducing/setting openfiles (failed"
525 " at %u of %u)\n", td->nr_open_files,
526 td->o.nr_files);
527 }
528
529 assert(f->fd == -1);
530 assert(!fio_file_open(f));
531 return 1;
532 }
533
534 fio_file_reset(td, f);
535 fio_file_set_open(f);
536 fio_file_clear_closing(f);
537 disk_util_inc(f->du);
538
539 td->nr_open_files++;
540 get_file(f);
541
542 if (f->filetype == FIO_TYPE_PIPE) {
543 if (td_random(td)) {
544 log_err("fio: can't seek on pipes (no random io)\n");
545 goto err;
546 }
547 }
548
549 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
550 goto done;
551
552 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
553 goto err;
554
555 if (td->o.fadvise_hint != F_ADV_NONE &&
556 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
557 int flags;
558
559 if (td->o.fadvise_hint == F_ADV_TYPE) {
560 if (td_random(td))
561 flags = POSIX_FADV_RANDOM;
562 else
563 flags = POSIX_FADV_SEQUENTIAL;
564 } else if (td->o.fadvise_hint == F_ADV_RANDOM)
565 flags = POSIX_FADV_RANDOM;
566 else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
567 flags = POSIX_FADV_SEQUENTIAL;
568 else {
569 log_err("fio: unknown fadvise type %d\n",
570 td->o.fadvise_hint);
571 flags = POSIX_FADV_NORMAL;
572 }
573
574 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
575 if (!fio_did_warn(FIO_WARN_FADVISE))
576 log_err("fio: fadvise hint failed\n");
577 }
578 }
579#ifdef FIO_HAVE_WRITE_HINT
580 if (fio_option_is_set(&td->o, write_hint) &&
581 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
582 uint64_t hint = td->o.write_hint;
583 int cmd;
584
585 /*
586 * For direct IO, we just need/want to set the hint on
587 * the file descriptor. For buffered IO, we need to set
588 * it on the inode.
589 */
590 if (td->o.odirect)
591 cmd = F_SET_FILE_RW_HINT;
592 else
593 cmd = F_SET_RW_HINT;
594
595 if (fcntl(f->fd, cmd, &hint) < 0) {
596 td_verror(td, errno, "fcntl write hint");
597 goto err;
598 }
599 }
600#endif
601
602 if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
603 goto err;
604
605done:
606 log_file(td, f, FIO_LOG_OPEN_FILE);
607 return 0;
608err:
609 disk_util_dec(f->du);
610 if (td->io_ops->close_file)
611 td->io_ops->close_file(td, f);
612 return 1;
613}
614
615int td_io_close_file(struct thread_data *td, struct fio_file *f)
616{
617 if (!fio_file_closing(f))
618 log_file(td, f, FIO_LOG_CLOSE_FILE);
619
620 /*
621 * mark as closing, do real close when last io on it has completed
622 */
623 fio_file_set_closing(f);
624
625 return put_file(td, f);
626}
627
628int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
629{
630 if (td->io_ops->unlink_file)
631 return td->io_ops->unlink_file(td, f);
632 else {
633 int ret;
634
635 ret = unlink(f->file_name);
636 if (ret < 0)
637 return errno;
638
639 return 0;
640 }
641}
642
643int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
644{
645 if (!td->io_ops->get_file_size)
646 return 0;
647
648 return td->io_ops->get_file_size(td, f);
649}
650
651#ifdef CONFIG_DYNAMIC_ENGINES
652/* Load all dynamic engines in FIO_EXT_ENG_DIR for enghelp command */
653static void
654fio_load_dynamic_engines(struct thread_data *td)
655{
656 DIR *dirhandle = NULL;
657 struct dirent *dirent = NULL;
658 char engine_path[PATH_MAX];
659
660 dirhandle = opendir(FIO_EXT_ENG_DIR);
661 if (!dirhandle)
662 return;
663
664 while ((dirent = readdir(dirhandle)) != NULL) {
665 if (!strcmp(dirent->d_name, ".") ||
666 !strcmp(dirent->d_name, ".."))
667 continue;
668
669 sprintf(engine_path, "%s/%s", FIO_EXT_ENG_DIR, dirent->d_name);
670 dlopen_ioengine(td, engine_path);
671 }
672
673 closedir(dirhandle);
674}
675#else
676#define fio_load_dynamic_engines(td) do { } while (0)
677#endif
678
679int fio_show_ioengine_help(const char *engine)
680{
681 struct flist_head *entry;
682 struct thread_data td;
683 struct ioengine_ops *io_ops;
684 char *sep;
685 int ret = 1;
686
687 memset(&td, 0, sizeof(struct thread_data));
688
689 if (!engine || !*engine) {
690 log_info("Available IO engines:\n");
691 fio_load_dynamic_engines(&td);
692 flist_for_each(entry, &engine_list) {
693 io_ops = flist_entry(entry, struct ioengine_ops, list);
694 log_info("\t%s\n", io_ops->name);
695 }
696 return 0;
697 }
698 sep = strchr(engine, ',');
699 if (sep) {
700 *sep = 0;
701 sep++;
702 }
703
704 td.o.ioengine = (char *)engine;
705 td.io_ops = load_ioengine(&td);
706
707 if (!td.io_ops) {
708 log_info("IO engine %s not found\n", engine);
709 return 1;
710 }
711
712 if (td.io_ops->options)
713 ret = show_cmd_help(td.io_ops->options, sep);
714 else
715 log_info("IO engine %s has no options\n", td.io_ops->name);
716
717 free_ioengine(&td);
718 return ret;
719}