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.
23 static FLIST_HEAD(engine_list);
25 static bool check_engine_ops(struct ioengine_ops *ops)
27 if (ops->version != FIO_IOOPS_VERSION) {
28 log_err("bad ioops version %d (want %d)\n", ops->version,
34 log_err("%s: no queue handler\n", ops->name);
39 * sync engines only need a ->queue()
41 if (ops->flags & FIO_SYNCIO)
44 if (!ops->event || !ops->getevents) {
45 log_err("%s: no event/getevents handler\n", ops->name);
52 void unregister_ioengine(struct ioengine_ops *ops)
54 dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
55 flist_del_init(&ops->list);
58 void register_ioengine(struct ioengine_ops *ops)
60 dprint(FD_IO, "ioengine %s registered\n", ops->name);
61 flist_add_tail(&ops->list, &engine_list);
64 static struct ioengine_ops *find_ioengine(const char *name)
66 struct ioengine_ops *ops;
67 struct flist_head *entry;
69 flist_for_each(entry, &engine_list) {
70 ops = flist_entry(entry, struct ioengine_ops, list);
71 if (!strcmp(name, ops->name))
78 #ifdef CONFIG_DYNAMIC_ENGINES
79 static void *dlopen_external(struct thread_data *td, const char *engine)
81 char engine_path[PATH_MAX];
84 sprintf(engine_path, "%s/lib%s.so", FIO_EXT_ENG_DIR, engine);
86 dlhandle = dlopen(engine_path, RTLD_LAZY);
88 log_info("Engine %s not found; Either name is invalid, was not built, or fio-engine-%s package is missing.\n",
94 #define dlopen_external(td, engine) (NULL)
97 static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
98 const char *engine_lib)
100 struct ioengine_ops *ops;
103 dprint(FD_IO, "dload engine %s\n", engine_lib);
106 dlhandle = dlopen(engine_lib, RTLD_LAZY);
108 dlhandle = dlopen_external(td, engine_lib);
110 td_vmsg(td, -1, dlerror(), "dlopen");
116 * Unlike the included modules, external engines should have a
117 * non-static ioengine structure that we can reference.
119 ops = dlsym(dlhandle, engine_lib);
121 ops = dlsym(dlhandle, "ioengine");
124 * For some external engines (like C++ ones) it is not that trivial
125 * to provide a non-static ionengine structure that we can reference.
126 * Instead we call a method which allocates the required ioengine
130 get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
137 td_vmsg(td, -1, dlerror(), "dlsym");
142 td->io_ops_dlhandle = dlhandle;
146 static struct ioengine_ops *__load_ioengine(const char *engine)
149 * linux libaio has alias names, so convert to what we want
151 if (!strncmp(engine, "linuxaio", 8)) {
152 dprint(FD_IO, "converting ioengine name: %s -> libaio\n",
157 dprint(FD_IO, "load ioengine %s\n", engine);
158 return find_ioengine(engine);
161 struct ioengine_ops *load_ioengine(struct thread_data *td)
163 struct ioengine_ops *ops = NULL;
167 * Use ->ioengine_so_path if an external ioengine path is specified.
168 * In this case, ->ioengine is "external" which also means the prefix
169 * for external ioengines "external:" is properly used.
171 name = td->o.ioengine_so_path ?: td->o.ioengine;
174 * Try to load ->ioengine first, and if failed try to dlopen(3) either
175 * ->ioengine or ->ioengine_so_path. This is redundant for an external
176 * ioengine with prefix, and also leaves the possibility of unexpected
177 * behavior (e.g. if the "external" ioengine exists), but we do this
178 * so as not to break job files not using the prefix.
180 ops = __load_ioengine(td->o.ioengine);
182 ops = dlopen_ioengine(td, name);
185 * If ops is NULL, we failed to load ->ioengine, and also failed to
186 * dlopen(3) either ->ioengine or ->ioengine_so_path as a path.
189 log_err("fio: engine %s not loadable\n", name);
194 * Check that the required methods are there.
196 if (check_engine_ops(ops))
203 * For cleaning up an ioengine which never made it to init().
205 void free_ioengine(struct thread_data *td)
207 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
209 if (td->eo && td->io_ops->options) {
210 options_free(td->io_ops->options, td->eo);
215 if (td->io_ops_dlhandle) {
216 dlclose(td->io_ops_dlhandle);
217 td->io_ops_dlhandle = NULL;
223 void close_ioengine(struct thread_data *td)
225 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
227 if (td->io_ops->cleanup) {
228 td->io_ops->cleanup(td);
229 td->io_ops_data = NULL;
235 int td_io_prep(struct thread_data *td, struct io_u *io_u)
237 dprint_io_u(io_u, "prep");
238 fio_ro_check(td, io_u);
240 lock_file(td, io_u->file, io_u->ddir);
242 if (td->io_ops->prep) {
243 int ret = td->io_ops->prep(td, io_u);
245 dprint(FD_IO, "prep: io_u %p: ret=%d\n", io_u, ret);
248 unlock_file(td, io_u->file);
255 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
256 const struct timespec *t)
261 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
262 * server side gets a message from the client
263 * side that the task is finished, and
264 * td->done is set to 1 after td_io_commit(). In this case,
265 * there is no need to reap complete event in server side.
270 if (min > 0 && td->io_ops->commit) {
271 r = td->io_ops->commit(td);
275 if (max > td->cur_depth)
281 if (max && td->io_ops->getevents)
282 r = td->io_ops->getevents(td, min, max, t);
286 * Reflect that our submitted requests were retrieved with
287 * whatever OS async calls are in the underlying engine.
289 td->io_u_in_flight -= r;
290 io_u_mark_complete(td, r);
292 td_verror(td, r, "get_events");
294 dprint(FD_IO, "getevents: %d\n", r);
298 enum fio_q_status td_io_queue(struct thread_data *td, struct io_u *io_u)
300 const enum fio_ddir ddir = acct_ddir(io_u);
301 unsigned long long buflen = io_u->xfer_buflen;
302 enum fio_q_status ret;
304 dprint_io_u(io_u, "queue");
305 fio_ro_check(td, io_u);
307 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
308 io_u_set(td, io_u, IO_U_F_FLIGHT);
311 * If overlap checking was enabled in offload mode we
312 * can release this lock that was acquired when we
313 * started the overlap check because the IO_U_F_FLIGHT
316 if (td_offload_overlap(td))
317 pthread_mutex_unlock(&overlap_check);
319 assert(fio_file_open(io_u->file));
322 * If using a write iolog, store this entry.
329 if (td_ioengine_flagged(td, FIO_SYNCIO) ||
330 (td_ioengine_flagged(td, FIO_ASYNCIO_SYNC_TRIM) &&
331 io_u->ddir == DDIR_TRIM)) {
332 if (fio_fill_issue_time(td))
333 fio_gettime(&io_u->issue_time, NULL);
336 * only used for iolog
338 if (td->o.read_iolog_file)
339 memcpy(&td->last_issue, &io_u->issue_time,
340 sizeof(io_u->issue_time));
345 if (!(io_u->flags & IO_U_F_VER_LIST)) {
346 td->io_issues[ddir]++;
347 td->io_issue_bytes[ddir] += buflen;
349 td->rate_io_issue_bytes[ddir] += buflen;
352 ret = td->io_ops->queue(td, io_u);
353 zbd_queue_io_u(io_u, ret);
355 unlock_file(td, io_u->file);
357 if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
358 td->io_issues[ddir]--;
359 td->io_issue_bytes[ddir] -= buflen;
360 td->rate_io_issue_bytes[ddir] -= buflen;
361 io_u_clear(td, io_u, IO_U_F_FLIGHT);
365 * If an error was seen and the io engine didn't propagate it
366 * back to 'td', do so.
368 if (io_u->error && !td->error)
369 td_verror(td, io_u->error, "td_io_queue");
372 * Add warning for O_DIRECT so that users have an easier time
373 * spotting potentially bad alignment. If this triggers for the first
374 * IO, then it's likely an alignment problem or because the host fs
375 * does not support O_DIRECT
377 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
380 log_info("fio: first direct IO errored. File system may not "
381 "support direct IO, or iomem_align= is bad, or "
382 "invalid block size. Try setting direct=0.\n");
385 if (zbd_unaligned_write(io_u->error) &&
386 td->io_issues[io_u->ddir & 1] == 1 &&
387 td->o.zone_mode != ZONE_MODE_ZBD) {
388 log_info("fio: first I/O failed. If %s is a zoned block device, consider --zonemode=zbd\n",
389 io_u->file->file_name);
392 if (!td->io_ops->commit) {
393 io_u_mark_submit(td, 1);
394 io_u_mark_complete(td, 1);
398 if (ret == FIO_Q_COMPLETED) {
399 if (ddir_rw(io_u->ddir) ||
400 (ddir_sync(io_u->ddir) && td->runstate != TD_FSYNCING)) {
401 io_u_mark_depth(td, 1);
402 td->ts.total_io_u[io_u->ddir]++;
404 } else if (ret == FIO_Q_QUEUED) {
407 if (ddir_rw(io_u->ddir) ||
408 (ddir_sync(io_u->ddir) && td->runstate != TD_FSYNCING))
409 td->ts.total_io_u[io_u->ddir]++;
411 if (td->io_u_queued >= td->o.iodepth_batch)
415 if (!td_ioengine_flagged(td, FIO_SYNCIO) &&
416 (!td_ioengine_flagged(td, FIO_ASYNCIO_SYNC_TRIM) ||
417 io_u->ddir != DDIR_TRIM)) {
418 if (fio_fill_issue_time(td))
419 fio_gettime(&io_u->issue_time, NULL);
422 * only used for iolog
424 if (td->o.read_iolog_file)
425 memcpy(&td->last_issue, &io_u->issue_time,
426 sizeof(io_u->issue_time));
432 int td_io_init(struct thread_data *td)
436 if (td->io_ops->init) {
437 ret = td->io_ops->init(td);
439 log_err("fio: io engine %s init failed.%s\n",
442 " Perhaps try reducing io depth?" : "");
452 void td_io_commit(struct thread_data *td)
456 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
458 if (!td->cur_depth || !td->io_u_queued)
461 io_u_mark_depth(td, td->io_u_queued);
463 if (td->io_ops->commit) {
464 ret = td->io_ops->commit(td);
466 td_verror(td, -ret, "io commit");
470 * Reflect that events were submitted as async IO requests.
472 td->io_u_in_flight += td->io_u_queued;
476 int td_io_open_file(struct thread_data *td, struct fio_file *f)
478 if (fio_file_closing(f)) {
480 * Open translates to undo closing.
482 fio_file_clear_closing(f);
486 assert(!fio_file_open(f));
488 assert(td->io_ops->open_file);
490 if (td->io_ops->open_file(td, f)) {
491 if (td->error == EINVAL && td->o.odirect)
492 log_err("fio: destination does not support O_DIRECT\n");
493 if (td->error == EMFILE) {
494 log_err("fio: try reducing/setting openfiles (failed"
495 " at %u of %u)\n", td->nr_open_files,
500 assert(!fio_file_open(f));
504 fio_file_reset(td, f);
505 fio_file_set_open(f);
506 fio_file_clear_closing(f);
507 disk_util_inc(f->du);
512 if (f->filetype == FIO_TYPE_PIPE) {
514 log_err("fio: can't seek on pipes (no random io)\n");
519 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
522 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
525 if (td->o.fadvise_hint != F_ADV_NONE &&
526 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
529 if (td->o.fadvise_hint == F_ADV_TYPE) {
531 flags = POSIX_FADV_RANDOM;
533 flags = POSIX_FADV_SEQUENTIAL;
534 } else if (td->o.fadvise_hint == F_ADV_RANDOM)
535 flags = POSIX_FADV_RANDOM;
536 else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
537 flags = POSIX_FADV_SEQUENTIAL;
539 log_err("fio: unknown fadvise type %d\n",
541 flags = POSIX_FADV_NORMAL;
544 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
545 if (!fio_did_warn(FIO_WARN_FADVISE))
546 log_err("fio: fadvise hint failed\n");
549 #ifdef FIO_HAVE_WRITE_HINT
550 if (fio_option_is_set(&td->o, write_hint) &&
551 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
552 uint64_t hint = td->o.write_hint;
556 * For direct IO, we just need/want to set the hint on
557 * the file descriptor. For buffered IO, we need to set
561 cmd = F_SET_FILE_RW_HINT;
565 if (fcntl(f->fd, cmd, &hint) < 0) {
566 td_verror(td, errno, "fcntl write hint");
572 if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
576 log_file(td, f, FIO_LOG_OPEN_FILE);
579 disk_util_dec(f->du);
580 if (td->io_ops->close_file)
581 td->io_ops->close_file(td, f);
585 int td_io_close_file(struct thread_data *td, struct fio_file *f)
587 if (!fio_file_closing(f))
588 log_file(td, f, FIO_LOG_CLOSE_FILE);
591 * mark as closing, do real close when last io on it has completed
593 fio_file_set_closing(f);
595 return put_file(td, f);
598 int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
600 if (td->io_ops->unlink_file)
601 return td->io_ops->unlink_file(td, f);
605 ret = unlink(f->file_name);
613 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
615 if (!td->io_ops->get_file_size)
618 return td->io_ops->get_file_size(td, f);
621 int fio_show_ioengine_help(const char *engine)
623 struct flist_head *entry;
624 struct thread_data td;
625 struct ioengine_ops *io_ops;
629 if (!engine || !*engine) {
630 log_info("Available IO engines:\n");
631 flist_for_each(entry, &engine_list) {
632 io_ops = flist_entry(entry, struct ioengine_ops, list);
633 log_info("\t%s\n", io_ops->name);
637 sep = strchr(engine, ',');
643 memset(&td, 0, sizeof(struct thread_data));
644 td.o.ioengine = (char *)engine;
645 io_ops = load_ioengine(&td);
648 log_info("IO engine %s not found\n", engine);
653 ret = show_cmd_help(io_ops->options, sep);
655 log_info("IO engine %s has no options\n", io_ops->name);