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.
5 * sync io is implemented on top of aio.
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.
18 #include <sys/types.h>
25 static FLIST_HEAD(engine_list);
27 static inline bool async_ioengine_sync_trim(struct thread_data *td,
30 return td_ioengine_flagged(td, FIO_ASYNCIO_SYNC_TRIM) &&
31 io_u->ddir == DDIR_TRIM;
34 static bool check_engine_ops(struct thread_data *td, struct ioengine_ops *ops)
36 if (ops->version != FIO_IOOPS_VERSION) {
37 log_err("bad ioops version %d (want %d)\n", ops->version,
43 log_err("%s: no queue handler\n", ops->name);
48 * sync engines only need a ->queue()
50 if (ops->flags & FIO_SYNCIO)
54 * async engines aren't reliable with offload
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);
63 if (!ops->event || !ops->getevents) {
64 log_err("%s: no event/getevents handler\n", ops->name);
71 void unregister_ioengine(struct ioengine_ops *ops)
73 dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
74 flist_del_init(&ops->list);
77 void register_ioengine(struct ioengine_ops *ops)
79 dprint(FD_IO, "ioengine %s registered\n", ops->name);
80 flist_add_tail(&ops->list, &engine_list);
83 static struct ioengine_ops *find_ioengine(const char *name)
85 struct ioengine_ops *ops;
86 struct flist_head *entry;
88 flist_for_each(entry, &engine_list) {
89 ops = flist_entry(entry, struct ioengine_ops, list);
90 if (!strcmp(name, ops->name))
97 #ifdef CONFIG_DYNAMIC_ENGINES
98 static void *dlopen_external(struct thread_data *td, const char *engine)
100 char engine_path[PATH_MAX];
103 sprintf(engine_path, "%s/fio-%s.so", FIO_EXT_ENG_DIR, engine);
105 dprint(FD_IO, "dlopen external %s\n", engine_path);
106 dlhandle = dlopen(engine_path, RTLD_LAZY);
108 log_info("Engine %s not found; Either name is invalid, was not built, or fio-engine-%s package is missing.\n",
114 #define dlopen_external(td, engine) (NULL)
117 static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
118 const char *engine_lib)
120 struct ioengine_ops *ops;
123 if (!strncmp(engine_lib, "linuxaio", 8) ||
124 !strncmp(engine_lib, "aio", 3))
125 engine_lib = "libaio";
127 dprint(FD_IO, "dlopen engine %s\n", engine_lib);
130 dlhandle = dlopen(engine_lib, RTLD_LAZY);
132 dlhandle = dlopen_external(td, engine_lib);
134 td_vmsg(td, -1, dlerror(), "dlopen");
140 * Unlike the included modules, external engines should have a
141 * non-static ioengine structure that we can reference.
143 ops = dlsym(dlhandle, engine_lib);
145 ops = dlsym(dlhandle, "ioengine");
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
154 get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
161 td_vmsg(td, -1, dlerror(), "dlsym");
166 ops->dlhandle = dlhandle;
170 static struct ioengine_ops *__load_ioengine(const char *engine)
173 * linux libaio has alias names, so convert to what we want
175 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3)) {
176 dprint(FD_IO, "converting ioengine name: %s -> libaio\n",
181 dprint(FD_IO, "load ioengine %s\n", engine);
182 return find_ioengine(engine);
185 struct ioengine_ops *load_ioengine(struct thread_data *td)
187 struct ioengine_ops *ops = NULL;
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.
195 name = td->o.ioengine_so_path ?: td->o.ioengine;
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.
204 ops = __load_ioengine(td->o.ioengine);
206 /* We do re-dlopen existing handles, for reference counting */
207 if (!ops || ops->dlhandle)
208 ops = dlopen_ioengine(td, name);
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.
215 log_err("fio: engine %s not loadable\n", name);
220 * Check that the required methods are there.
222 if (check_engine_ops(td, ops))
229 * For cleaning up an ioengine which never made it to init().
231 void free_ioengine(struct thread_data *td)
233 assert(td != NULL && td->io_ops != NULL);
235 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
237 if (td->eo && td->io_ops->options) {
238 options_free(td->io_ops->options, td->eo);
243 if (td->io_ops->dlhandle) {
244 dprint(FD_IO, "dlclose ioengine %s\n", td->io_ops->name);
245 dlclose(td->io_ops->dlhandle);
251 void close_ioengine(struct thread_data *td)
253 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
255 if (td->io_ops->cleanup) {
256 td->io_ops->cleanup(td);
257 td->io_ops_data = NULL;
263 int td_io_prep(struct thread_data *td, struct io_u *io_u)
265 dprint_io_u(io_u, "prep");
266 fio_ro_check(td, io_u);
268 lock_file(td, io_u->file, io_u->ddir);
270 if (td->io_ops->prep) {
271 int ret = td->io_ops->prep(td, io_u);
273 dprint(FD_IO, "prep: io_u %p: ret=%d\n", io_u, ret);
276 unlock_file(td, io_u->file);
283 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
284 const struct timespec *t)
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.
298 if (min > 0 && td->io_ops->commit) {
299 r = td->io_ops->commit(td);
303 if (max > td->cur_depth)
309 if (max && td->io_ops->getevents)
310 r = td->io_ops->getevents(td, min, max, t);
314 * Reflect that our submitted requests were retrieved with
315 * whatever OS async calls are in the underlying engine.
317 td->io_u_in_flight -= r;
318 io_u_mark_complete(td, r);
320 td_verror(td, r, "get_events");
322 dprint(FD_IO, "getevents: %d\n", r);
326 enum fio_q_status td_io_queue(struct thread_data *td, struct io_u *io_u)
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;
332 dprint_io_u(io_u, "queue");
333 fio_ro_check(td, io_u);
335 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
336 io_u_set(td, io_u, IO_U_F_FLIGHT);
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
344 if (td_offload_overlap(td)) {
345 int res = pthread_mutex_unlock(&overlap_check);
349 assert(fio_file_open(io_u->file));
352 * If using a write iolog, store this entry.
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);
365 * only used for iolog
367 if (td->o.read_iolog_file)
368 memcpy(&td->last_issue, &io_u->issue_time,
369 sizeof(io_u->issue_time));
375 if (!(io_u->flags & IO_U_F_VER_LIST)) {
376 td->io_issues[ddir]++;
377 td->io_issue_bytes[ddir] += buflen;
379 td->rate_io_issue_bytes[ddir] += buflen;
382 ret = td->io_ops->queue(td, io_u);
383 zbd_queue_io_u(td, io_u, ret);
385 unlock_file(td, io_u->file);
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);
395 * If an error was seen and the io engine didn't propagate it
396 * back to 'td', do so.
398 if (io_u->error && !td->error)
399 td_verror(td, io_u->error, "td_io_queue");
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
407 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
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");
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);
422 if (!td->io_ops->commit) {
423 io_u_mark_submit(td, 1);
424 io_u_mark_complete(td, 1);
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]++;
433 } else if (ret == FIO_Q_QUEUED) {
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]++;
440 if (td->io_u_queued >= td->o.iodepth_batch)
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);
451 * only used for iolog
453 if (td->o.read_iolog_file)
454 memcpy(&td->last_issue, &io_u->issue_time,
455 sizeof(io_u->issue_time));
462 int td_io_init(struct thread_data *td)
466 if (td->io_ops->init) {
467 ret = td->io_ops->init(td);
469 log_err("fio: io engine %s init failed.%s\n",
472 " Perhaps try reducing io depth?" : "");
482 void td_io_commit(struct thread_data *td)
486 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
488 if (!td->cur_depth || !td->io_u_queued)
491 io_u_mark_depth(td, td->io_u_queued);
493 if (td->io_ops->commit) {
494 ret = td->io_ops->commit(td);
496 td_verror(td, -ret, "io commit");
500 * Reflect that events were submitted as async IO requests.
502 td->io_u_in_flight += td->io_u_queued;
506 int td_io_open_file(struct thread_data *td, struct fio_file *f)
508 if (fio_file_closing(f)) {
510 * Open translates to undo closing.
512 fio_file_clear_closing(f);
516 assert(!fio_file_open(f));
518 assert(td->io_ops->open_file);
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,
530 assert(!fio_file_open(f));
534 fio_file_reset(td, f);
535 fio_file_set_open(f);
536 fio_file_clear_closing(f);
537 disk_util_inc(f->du);
542 if (f->filetype == FIO_TYPE_PIPE) {
544 log_err("fio: can't seek on pipes (no random io)\n");
549 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
552 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
555 if (td->o.fadvise_hint != F_ADV_NONE &&
556 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
559 if (td->o.fadvise_hint == F_ADV_TYPE) {
561 flags = POSIX_FADV_RANDOM;
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 #ifdef POSIX_FADV_NOREUSE
569 else if (td->o.fadvise_hint == F_ADV_NOREUSE)
570 flags = POSIX_FADV_NOREUSE;
573 log_err("fio: unknown fadvise type %d\n",
575 flags = POSIX_FADV_NORMAL;
578 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
579 if (!fio_did_warn(FIO_WARN_FADVISE))
580 log_err("fio: fadvise hint failed\n");
583 #ifdef FIO_HAVE_WRITE_HINT
584 if (fio_option_is_set(&td->o, write_hint) &&
585 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
586 uint64_t hint = td->o.write_hint;
590 * For direct IO, we just need/want to set the hint on
591 * the file descriptor. For buffered IO, we need to set
595 cmd = F_SET_FILE_RW_HINT;
599 if (fcntl(f->fd, cmd, &hint) < 0) {
600 td_verror(td, errno, "fcntl write hint");
606 if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
610 log_file(td, f, FIO_LOG_OPEN_FILE);
613 disk_util_dec(f->du);
614 if (td->io_ops->close_file)
615 td->io_ops->close_file(td, f);
619 int td_io_close_file(struct thread_data *td, struct fio_file *f)
621 if (!fio_file_closing(f))
622 log_file(td, f, FIO_LOG_CLOSE_FILE);
625 * mark as closing, do real close when last io on it has completed
627 fio_file_set_closing(f);
629 return put_file(td, f);
632 int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
634 if (td->io_ops->unlink_file)
635 return td->io_ops->unlink_file(td, f);
639 ret = unlink(f->file_name);
647 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
649 if (!td->io_ops->get_file_size)
652 return td->io_ops->get_file_size(td, f);
655 #ifdef CONFIG_DYNAMIC_ENGINES
656 /* Load all dynamic engines in FIO_EXT_ENG_DIR for enghelp command */
658 fio_load_dynamic_engines(struct thread_data *td)
660 DIR *dirhandle = NULL;
661 struct dirent *dirent = NULL;
662 char engine_path[PATH_MAX];
664 dirhandle = opendir(FIO_EXT_ENG_DIR);
668 while ((dirent = readdir(dirhandle)) != NULL) {
669 if (!strcmp(dirent->d_name, ".") ||
670 !strcmp(dirent->d_name, ".."))
673 sprintf(engine_path, "%s/%s", FIO_EXT_ENG_DIR, dirent->d_name);
674 dlopen_ioengine(td, engine_path);
680 #define fio_load_dynamic_engines(td) do { } while (0)
683 int fio_show_ioengine_help(const char *engine)
685 struct flist_head *entry;
686 struct thread_data td;
687 struct ioengine_ops *io_ops;
691 memset(&td, 0, sizeof(struct thread_data));
693 if (!engine || !*engine) {
694 log_info("Available IO engines:\n");
695 fio_load_dynamic_engines(&td);
696 flist_for_each(entry, &engine_list) {
697 io_ops = flist_entry(entry, struct ioengine_ops, list);
698 log_info("\t%s\n", io_ops->name);
702 sep = strchr(engine, ',');
708 td.o.ioengine = (char *)engine;
709 td.io_ops = load_ioengine(&td);
712 log_info("IO engine %s not found\n", engine);
716 if (td.io_ops->options)
717 ret = show_cmd_help(td.io_ops->options, sep);
719 log_info("IO engine %s has no options\n", td.io_ops->name);