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 engine[sizeof(engine) - 1] = '\0';
129 strncpy(engine, name, sizeof(engine) - 1);
132 * linux libaio has alias names, so convert to what we want
134 if (!strncmp(engine, "linuxaio", 8)) {
135 dprint(FD_IO, "converting ioengine name: %s -> libaio\n", name);
136 strcpy(engine, "libaio");
139 dprint(FD_IO, "load ioengine %s\n", engine);
140 return find_ioengine(engine);
143 struct ioengine_ops *load_ioengine(struct thread_data *td)
145 struct ioengine_ops *ops = NULL;
149 * Use ->ioengine_so_path if an external ioengine path is specified.
150 * In this case, ->ioengine is "external" which also means the prefix
151 * for external ioengines "external:" is properly used.
153 name = td->o.ioengine_so_path ?: td->o.ioengine;
156 * Try to load ->ioengine first, and if failed try to dlopen(3) either
157 * ->ioengine or ->ioengine_so_path. This is redundant for an external
158 * ioengine with prefix, and also leaves the possibility of unexpected
159 * behavior (e.g. if the "external" ioengine exists), but we do this
160 * so as not to break job files not using the prefix.
162 ops = __load_ioengine(td->o.ioengine);
164 ops = dlopen_ioengine(td, name);
167 * If ops is NULL, we failed to load ->ioengine, and also failed to
168 * dlopen(3) either ->ioengine or ->ioengine_so_path as a path.
171 log_err("fio: engine %s not loadable\n", name);
176 * Check that the required methods are there.
178 if (check_engine_ops(ops))
185 * For cleaning up an ioengine which never made it to init().
187 void free_ioengine(struct thread_data *td)
189 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
191 if (td->eo && td->io_ops->options) {
192 options_free(td->io_ops->options, td->eo);
197 if (td->io_ops_dlhandle) {
198 dlclose(td->io_ops_dlhandle);
199 td->io_ops_dlhandle = NULL;
205 void close_ioengine(struct thread_data *td)
207 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
209 if (td->io_ops->cleanup) {
210 td->io_ops->cleanup(td);
211 td->io_ops_data = NULL;
217 int td_io_prep(struct thread_data *td, struct io_u *io_u)
219 dprint_io_u(io_u, "prep");
220 fio_ro_check(td, io_u);
222 lock_file(td, io_u->file, io_u->ddir);
224 if (td->io_ops->prep) {
225 int ret = td->io_ops->prep(td, io_u);
227 dprint(FD_IO, "prep: io_u %p: ret=%d\n", io_u, ret);
230 unlock_file(td, io_u->file);
237 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
238 const struct timespec *t)
243 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
244 * server side gets a message from the client
245 * side that the task is finished, and
246 * td->done is set to 1 after td_io_commit(). In this case,
247 * there is no need to reap complete event in server side.
252 if (min > 0 && td->io_ops->commit) {
253 r = td->io_ops->commit(td);
257 if (max > td->cur_depth)
263 if (max && td->io_ops->getevents)
264 r = td->io_ops->getevents(td, min, max, t);
268 * Reflect that our submitted requests were retrieved with
269 * whatever OS async calls are in the underlying engine.
271 td->io_u_in_flight -= r;
272 io_u_mark_complete(td, r);
274 td_verror(td, r, "get_events");
276 dprint(FD_IO, "getevents: %d\n", r);
280 enum fio_q_status td_io_queue(struct thread_data *td, struct io_u *io_u)
282 const enum fio_ddir ddir = acct_ddir(io_u);
283 unsigned long long buflen = io_u->xfer_buflen;
284 enum fio_q_status ret;
286 dprint_io_u(io_u, "queue");
287 fio_ro_check(td, io_u);
289 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
290 io_u_set(td, io_u, IO_U_F_FLIGHT);
293 * If overlap checking was enabled in offload mode we
294 * can release this lock that was acquired when we
295 * started the overlap check because the IO_U_F_FLIGHT
298 if (td_offload_overlap(td))
299 pthread_mutex_unlock(&overlap_check);
301 assert(fio_file_open(io_u->file));
304 * If using a write iolog, store this entry.
311 if (td_ioengine_flagged(td, FIO_SYNCIO)) {
312 if (fio_fill_issue_time(td))
313 fio_gettime(&io_u->issue_time, NULL);
316 * only used for iolog
318 if (td->o.read_iolog_file)
319 memcpy(&td->last_issue, &io_u->issue_time,
320 sizeof(io_u->issue_time));
324 if (!(io_u->flags & IO_U_F_VER_LIST)) {
325 td->io_issues[ddir]++;
326 td->io_issue_bytes[ddir] += buflen;
328 td->rate_io_issue_bytes[ddir] += buflen;
331 ret = td->io_ops->queue(td, io_u);
332 zbd_queue_io_u(io_u, ret);
334 unlock_file(td, io_u->file);
336 if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
337 td->io_issues[ddir]--;
338 td->io_issue_bytes[ddir] -= buflen;
339 td->rate_io_issue_bytes[ddir] -= buflen;
340 io_u_clear(td, io_u, IO_U_F_FLIGHT);
344 * If an error was seen and the io engine didn't propagate it
345 * back to 'td', do so.
347 if (io_u->error && !td->error)
348 td_verror(td, io_u->error, "td_io_queue");
351 * Add warning for O_DIRECT so that users have an easier time
352 * spotting potentially bad alignment. If this triggers for the first
353 * IO, then it's likely an alignment problem or because the host fs
354 * does not support O_DIRECT
356 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
359 log_info("fio: first direct IO errored. File system may not "
360 "support direct IO, or iomem_align= is bad, or "
361 "invalid block size. Try setting direct=0.\n");
364 if (zbd_unaligned_write(io_u->error) &&
365 td->io_issues[io_u->ddir & 1] == 1 &&
366 td->o.zone_mode != ZONE_MODE_ZBD) {
367 log_info("fio: first I/O failed. If %s is a zoned block device, consider --zonemode=zbd\n",
368 io_u->file->file_name);
371 if (!td->io_ops->commit) {
372 io_u_mark_submit(td, 1);
373 io_u_mark_complete(td, 1);
377 if (ret == FIO_Q_COMPLETED) {
378 if (ddir_rw(io_u->ddir) || ddir_sync(io_u->ddir)) {
379 io_u_mark_depth(td, 1);
380 td->ts.total_io_u[io_u->ddir]++;
382 } else if (ret == FIO_Q_QUEUED) {
385 if (ddir_rw(io_u->ddir) || ddir_sync(io_u->ddir))
386 td->ts.total_io_u[io_u->ddir]++;
388 if (td->io_u_queued >= td->o.iodepth_batch)
392 if (!td_ioengine_flagged(td, FIO_SYNCIO)) {
393 if (fio_fill_issue_time(td))
394 fio_gettime(&io_u->issue_time, NULL);
397 * only used for iolog
399 if (td->o.read_iolog_file)
400 memcpy(&td->last_issue, &io_u->issue_time,
401 sizeof(io_u->issue_time));
407 int td_io_init(struct thread_data *td)
411 if (td->io_ops->init) {
412 ret = td->io_ops->init(td);
414 log_err("fio: io engine %s init failed.%s\n",
417 " Perhaps try reducing io depth?" : "");
427 void td_io_commit(struct thread_data *td)
431 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
433 if (!td->cur_depth || !td->io_u_queued)
436 io_u_mark_depth(td, td->io_u_queued);
438 if (td->io_ops->commit) {
439 ret = td->io_ops->commit(td);
441 td_verror(td, -ret, "io commit");
445 * Reflect that events were submitted as async IO requests.
447 td->io_u_in_flight += td->io_u_queued;
451 int td_io_open_file(struct thread_data *td, struct fio_file *f)
453 if (fio_file_closing(f)) {
455 * Open translates to undo closing.
457 fio_file_clear_closing(f);
461 assert(!fio_file_open(f));
463 assert(td->io_ops->open_file);
465 if (td->io_ops->open_file(td, f)) {
466 if (td->error == EINVAL && td->o.odirect)
467 log_err("fio: destination does not support O_DIRECT\n");
468 if (td->error == EMFILE) {
469 log_err("fio: try reducing/setting openfiles (failed"
470 " at %u of %u)\n", td->nr_open_files,
475 assert(!fio_file_open(f));
479 fio_file_reset(td, f);
480 fio_file_set_open(f);
481 fio_file_clear_closing(f);
482 disk_util_inc(f->du);
487 if (f->filetype == FIO_TYPE_PIPE) {
489 log_err("fio: can't seek on pipes (no random io)\n");
494 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
497 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
500 if (td->o.fadvise_hint != F_ADV_NONE &&
501 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
504 if (td->o.fadvise_hint == F_ADV_TYPE) {
506 flags = POSIX_FADV_RANDOM;
508 flags = POSIX_FADV_SEQUENTIAL;
509 } else if (td->o.fadvise_hint == F_ADV_RANDOM)
510 flags = POSIX_FADV_RANDOM;
511 else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
512 flags = POSIX_FADV_SEQUENTIAL;
514 log_err("fio: unknown fadvise type %d\n",
516 flags = POSIX_FADV_NORMAL;
519 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
520 if (!fio_did_warn(FIO_WARN_FADVISE))
521 log_err("fio: fadvise hint failed\n");
524 #ifdef FIO_HAVE_WRITE_HINT
525 if (fio_option_is_set(&td->o, write_hint) &&
526 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
527 uint64_t hint = td->o.write_hint;
531 * For direct IO, we just need/want to set the hint on
532 * the file descriptor. For buffered IO, we need to set
536 cmd = F_SET_FILE_RW_HINT;
540 if (fcntl(f->fd, cmd, &hint) < 0) {
541 td_verror(td, errno, "fcntl write hint");
547 if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
551 log_file(td, f, FIO_LOG_OPEN_FILE);
554 disk_util_dec(f->du);
555 if (td->io_ops->close_file)
556 td->io_ops->close_file(td, f);
560 int td_io_close_file(struct thread_data *td, struct fio_file *f)
562 if (!fio_file_closing(f))
563 log_file(td, f, FIO_LOG_CLOSE_FILE);
566 * mark as closing, do real close when last io on it has completed
568 fio_file_set_closing(f);
570 return put_file(td, f);
573 int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
575 if (td->io_ops->unlink_file)
576 return td->io_ops->unlink_file(td, f);
580 ret = unlink(f->file_name);
588 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
590 if (!td->io_ops->get_file_size)
593 return td->io_ops->get_file_size(td, f);
596 int fio_show_ioengine_help(const char *engine)
598 struct flist_head *entry;
599 struct thread_data td;
600 struct ioengine_ops *io_ops;
604 if (!engine || !*engine) {
605 log_info("Available IO engines:\n");
606 flist_for_each(entry, &engine_list) {
607 io_ops = flist_entry(entry, struct ioengine_ops, list);
608 log_info("\t%s\n", io_ops->name);
612 sep = strchr(engine, ',');
618 memset(&td, 0, sizeof(struct thread_data));
619 td.o.ioengine = (char *)engine;
620 io_ops = load_ioengine(&td);
623 log_info("IO engine %s not found\n", engine);
628 ret = show_cmd_help(io_ops->options, sep);
630 log_info("IO engine %s has no options\n", io_ops->name);