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 static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
79 const char *engine_lib)
81 struct ioengine_ops *ops;
84 dprint(FD_IO, "dload engine %s\n", engine_lib);
87 dlhandle = dlopen(engine_lib, RTLD_LAZY);
89 td_vmsg(td, -1, dlerror(), "dlopen");
94 * Unlike the included modules, external engines should have a
95 * non-static ioengine structure that we can reference.
97 ops = dlsym(dlhandle, engine_lib);
99 ops = dlsym(dlhandle, "ioengine");
102 * For some external engines (like C++ ones) it is not that trivial
103 * to provide a non-static ionengine structure that we can reference.
104 * Instead we call a method which allocates the required ioengine
108 get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
115 td_vmsg(td, -1, dlerror(), "dlsym");
120 td->io_ops_dlhandle = dlhandle;
124 static struct ioengine_ops *__load_ioengine(const char *name)
128 snprintf(engine, sizeof(engine), "%s", name);
131 * linux libaio has alias names, so convert to what we want
133 if (!strncmp(engine, "linuxaio", 8)) {
134 dprint(FD_IO, "converting ioengine name: %s -> libaio\n", name);
135 strcpy(engine, "libaio");
138 dprint(FD_IO, "load ioengine %s\n", engine);
139 return find_ioengine(engine);
142 struct ioengine_ops *load_ioengine(struct thread_data *td)
144 struct ioengine_ops *ops = NULL;
148 * Use ->ioengine_so_path if an external ioengine path is specified.
149 * In this case, ->ioengine is "external" which also means the prefix
150 * for external ioengines "external:" is properly used.
152 name = td->o.ioengine_so_path ?: td->o.ioengine;
155 * Try to load ->ioengine first, and if failed try to dlopen(3) either
156 * ->ioengine or ->ioengine_so_path. This is redundant for an external
157 * ioengine with prefix, and also leaves the possibility of unexpected
158 * behavior (e.g. if the "external" ioengine exists), but we do this
159 * so as not to break job files not using the prefix.
161 ops = __load_ioengine(td->o.ioengine);
163 ops = dlopen_ioengine(td, name);
166 * If ops is NULL, we failed to load ->ioengine, and also failed to
167 * dlopen(3) either ->ioengine or ->ioengine_so_path as a path.
170 log_err("fio: engine %s not loadable\n", name);
175 * Check that the required methods are there.
177 if (check_engine_ops(ops))
184 * For cleaning up an ioengine which never made it to init().
186 void free_ioengine(struct thread_data *td)
188 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
190 if (td->eo && td->io_ops->options) {
191 options_free(td->io_ops->options, td->eo);
196 if (td->io_ops_dlhandle) {
197 dlclose(td->io_ops_dlhandle);
198 td->io_ops_dlhandle = NULL;
204 void close_ioengine(struct thread_data *td)
206 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
208 if (td->io_ops->cleanup) {
209 td->io_ops->cleanup(td);
210 td->io_ops_data = NULL;
216 int td_io_prep(struct thread_data *td, struct io_u *io_u)
218 dprint_io_u(io_u, "prep");
219 fio_ro_check(td, io_u);
221 lock_file(td, io_u->file, io_u->ddir);
223 if (td->io_ops->prep) {
224 int ret = td->io_ops->prep(td, io_u);
226 dprint(FD_IO, "prep: io_u %p: ret=%d\n", io_u, ret);
229 unlock_file(td, io_u->file);
236 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
237 const struct timespec *t)
242 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
243 * server side gets a message from the client
244 * side that the task is finished, and
245 * td->done is set to 1 after td_io_commit(). In this case,
246 * there is no need to reap complete event in server side.
251 if (min > 0 && td->io_ops->commit) {
252 r = td->io_ops->commit(td);
256 if (max > td->cur_depth)
262 if (max && td->io_ops->getevents)
263 r = td->io_ops->getevents(td, min, max, t);
267 * Reflect that our submitted requests were retrieved with
268 * whatever OS async calls are in the underlying engine.
270 td->io_u_in_flight -= r;
271 io_u_mark_complete(td, r);
273 td_verror(td, r, "get_events");
275 dprint(FD_IO, "getevents: %d\n", r);
279 enum fio_q_status td_io_queue(struct thread_data *td, struct io_u *io_u)
281 const enum fio_ddir ddir = acct_ddir(io_u);
282 unsigned long long buflen = io_u->xfer_buflen;
283 enum fio_q_status ret;
285 dprint_io_u(io_u, "queue");
286 fio_ro_check(td, io_u);
288 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
289 io_u_set(td, io_u, IO_U_F_FLIGHT);
292 * If overlap checking was enabled in offload mode we
293 * can release this lock that was acquired when we
294 * started the overlap check because the IO_U_F_FLIGHT
297 if (td_offload_overlap(td))
298 pthread_mutex_unlock(&overlap_check);
300 assert(fio_file_open(io_u->file));
303 * If using a write iolog, store this entry.
310 if (td_ioengine_flagged(td, FIO_SYNCIO) ||
311 (td_ioengine_flagged(td, FIO_ASYNCIO_SYNC_TRIM) &&
312 io_u->ddir == DDIR_TRIM)) {
313 if (fio_fill_issue_time(td))
314 fio_gettime(&io_u->issue_time, NULL);
317 * only used for iolog
319 if (td->o.read_iolog_file)
320 memcpy(&td->last_issue, &io_u->issue_time,
321 sizeof(io_u->issue_time));
325 if (!(io_u->flags & IO_U_F_VER_LIST)) {
326 td->io_issues[ddir]++;
327 td->io_issue_bytes[ddir] += buflen;
329 td->rate_io_issue_bytes[ddir] += buflen;
332 ret = td->io_ops->queue(td, io_u);
333 zbd_queue_io_u(io_u, ret);
335 unlock_file(td, io_u->file);
337 if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
338 td->io_issues[ddir]--;
339 td->io_issue_bytes[ddir] -= buflen;
340 td->rate_io_issue_bytes[ddir] -= buflen;
341 io_u_clear(td, io_u, IO_U_F_FLIGHT);
345 * If an error was seen and the io engine didn't propagate it
346 * back to 'td', do so.
348 if (io_u->error && !td->error)
349 td_verror(td, io_u->error, "td_io_queue");
352 * Add warning for O_DIRECT so that users have an easier time
353 * spotting potentially bad alignment. If this triggers for the first
354 * IO, then it's likely an alignment problem or because the host fs
355 * does not support O_DIRECT
357 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
360 log_info("fio: first direct IO errored. File system may not "
361 "support direct IO, or iomem_align= is bad, or "
362 "invalid block size. Try setting direct=0.\n");
365 if (zbd_unaligned_write(io_u->error) &&
366 td->io_issues[io_u->ddir & 1] == 1 &&
367 td->o.zone_mode != ZONE_MODE_ZBD) {
368 log_info("fio: first I/O failed. If %s is a zoned block device, consider --zonemode=zbd\n",
369 io_u->file->file_name);
372 if (!td->io_ops->commit) {
373 io_u_mark_submit(td, 1);
374 io_u_mark_complete(td, 1);
378 if (ret == FIO_Q_COMPLETED) {
379 if (ddir_rw(io_u->ddir) || ddir_sync(io_u->ddir)) {
380 io_u_mark_depth(td, 1);
381 td->ts.total_io_u[io_u->ddir]++;
383 } else if (ret == FIO_Q_QUEUED) {
386 if (ddir_rw(io_u->ddir) || ddir_sync(io_u->ddir))
387 td->ts.total_io_u[io_u->ddir]++;
389 if (td->io_u_queued >= td->o.iodepth_batch)
393 if (!td_ioengine_flagged(td, FIO_SYNCIO) &&
394 (!td_ioengine_flagged(td, FIO_ASYNCIO_SYNC_TRIM) ||
395 io_u->ddir != DDIR_TRIM)) {
396 if (fio_fill_issue_time(td))
397 fio_gettime(&io_u->issue_time, NULL);
400 * only used for iolog
402 if (td->o.read_iolog_file)
403 memcpy(&td->last_issue, &io_u->issue_time,
404 sizeof(io_u->issue_time));
410 int td_io_init(struct thread_data *td)
414 if (td->io_ops->init) {
415 ret = td->io_ops->init(td);
417 log_err("fio: io engine %s init failed.%s\n",
420 " Perhaps try reducing io depth?" : "");
430 void td_io_commit(struct thread_data *td)
434 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
436 if (!td->cur_depth || !td->io_u_queued)
439 io_u_mark_depth(td, td->io_u_queued);
441 if (td->io_ops->commit) {
442 ret = td->io_ops->commit(td);
444 td_verror(td, -ret, "io commit");
448 * Reflect that events were submitted as async IO requests.
450 td->io_u_in_flight += td->io_u_queued;
454 int td_io_open_file(struct thread_data *td, struct fio_file *f)
456 if (fio_file_closing(f)) {
458 * Open translates to undo closing.
460 fio_file_clear_closing(f);
464 assert(!fio_file_open(f));
466 assert(td->io_ops->open_file);
468 if (td->io_ops->open_file(td, f)) {
469 if (td->error == EINVAL && td->o.odirect)
470 log_err("fio: destination does not support O_DIRECT\n");
471 if (td->error == EMFILE) {
472 log_err("fio: try reducing/setting openfiles (failed"
473 " at %u of %u)\n", td->nr_open_files,
478 assert(!fio_file_open(f));
482 fio_file_reset(td, f);
483 fio_file_set_open(f);
484 fio_file_clear_closing(f);
485 disk_util_inc(f->du);
490 if (f->filetype == FIO_TYPE_PIPE) {
492 log_err("fio: can't seek on pipes (no random io)\n");
497 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
500 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
503 if (td->o.fadvise_hint != F_ADV_NONE &&
504 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
507 if (td->o.fadvise_hint == F_ADV_TYPE) {
509 flags = POSIX_FADV_RANDOM;
511 flags = POSIX_FADV_SEQUENTIAL;
512 } else if (td->o.fadvise_hint == F_ADV_RANDOM)
513 flags = POSIX_FADV_RANDOM;
514 else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
515 flags = POSIX_FADV_SEQUENTIAL;
517 log_err("fio: unknown fadvise type %d\n",
519 flags = POSIX_FADV_NORMAL;
522 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
523 if (!fio_did_warn(FIO_WARN_FADVISE))
524 log_err("fio: fadvise hint failed\n");
527 #ifdef FIO_HAVE_WRITE_HINT
528 if (fio_option_is_set(&td->o, write_hint) &&
529 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
530 uint64_t hint = td->o.write_hint;
534 * For direct IO, we just need/want to set the hint on
535 * the file descriptor. For buffered IO, we need to set
539 cmd = F_SET_FILE_RW_HINT;
543 if (fcntl(f->fd, cmd, &hint) < 0) {
544 td_verror(td, errno, "fcntl write hint");
550 if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
554 log_file(td, f, FIO_LOG_OPEN_FILE);
557 disk_util_dec(f->du);
558 if (td->io_ops->close_file)
559 td->io_ops->close_file(td, f);
563 int td_io_close_file(struct thread_data *td, struct fio_file *f)
565 if (!fio_file_closing(f))
566 log_file(td, f, FIO_LOG_CLOSE_FILE);
569 * mark as closing, do real close when last io on it has completed
571 fio_file_set_closing(f);
573 return put_file(td, f);
576 int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
578 if (td->io_ops->unlink_file)
579 return td->io_ops->unlink_file(td, f);
583 ret = unlink(f->file_name);
591 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
593 if (!td->io_ops->get_file_size)
596 return td->io_ops->get_file_size(td, f);
599 int fio_show_ioengine_help(const char *engine)
601 struct flist_head *entry;
602 struct thread_data td;
603 struct ioengine_ops *io_ops;
607 if (!engine || !*engine) {
608 log_info("Available IO engines:\n");
609 flist_for_each(entry, &engine_list) {
610 io_ops = flist_entry(entry, struct ioengine_ops, list);
611 log_info("\t%s\n", io_ops->name);
615 sep = strchr(engine, ',');
621 memset(&td, 0, sizeof(struct thread_data));
622 td.o.ioengine = (char *)engine;
623 io_ops = load_ioengine(&td);
626 log_info("IO engine %s not found\n", engine);
631 ret = show_cmd_help(io_ops->options, sep);
633 log_info("IO engine %s has no options\n", io_ops->name);