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(&ops->list);
56 INIT_FLIST_HEAD(&ops->list);
59 void register_ioengine(struct ioengine_ops *ops)
61 dprint(FD_IO, "ioengine %s registered\n", ops->name);
62 INIT_FLIST_HEAD(&ops->list);
63 flist_add_tail(&ops->list, &engine_list);
66 static struct ioengine_ops *find_ioengine(const char *name)
68 struct ioengine_ops *ops;
69 struct flist_head *entry;
71 flist_for_each(entry, &engine_list) {
72 ops = flist_entry(entry, struct ioengine_ops, list);
73 if (!strcmp(name, ops->name))
80 static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
81 const char *engine_lib)
83 struct ioengine_ops *ops;
86 dprint(FD_IO, "dload engine %s\n", engine_lib);
89 dlhandle = dlopen(engine_lib, RTLD_LAZY);
91 td_vmsg(td, -1, dlerror(), "dlopen");
96 * Unlike the included modules, external engines should have a
97 * non-static ioengine structure that we can reference.
99 ops = dlsym(dlhandle, engine_lib);
101 ops = dlsym(dlhandle, "ioengine");
104 * For some external engines (like C++ ones) it is not that trivial
105 * to provide a non-static ionengine structure that we can reference.
106 * Instead we call a method which allocates the required ioengine
110 get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
117 td_vmsg(td, -1, dlerror(), "dlsym");
122 td->io_ops_dlhandle = dlhandle;
126 struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
128 struct ioengine_ops *ops;
131 dprint(FD_IO, "load ioengine %s\n", name);
133 strncpy(engine, name, sizeof(engine) - 1);
136 * linux libaio has alias names, so convert to what we want
138 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
139 strcpy(engine, "libaio");
141 ops = find_ioengine(engine);
143 ops = dlopen_ioengine(td, name);
146 log_err("fio: engine %s not loadable\n", name);
151 * Check that the required methods are there.
153 if (check_engine_ops(ops))
160 * For cleaning up an ioengine which never made it to init().
162 void free_ioengine(struct thread_data *td)
164 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
166 if (td->eo && td->io_ops->options) {
167 options_free(td->io_ops->options, td->eo);
172 if (td->io_ops_dlhandle)
173 dlclose(td->io_ops_dlhandle);
178 void close_ioengine(struct thread_data *td)
180 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
182 if (td->io_ops->cleanup) {
183 td->io_ops->cleanup(td);
184 td->io_ops_data = NULL;
190 int td_io_prep(struct thread_data *td, struct io_u *io_u)
192 dprint_io_u(io_u, "prep");
193 fio_ro_check(td, io_u);
195 lock_file(td, io_u->file, io_u->ddir);
197 if (td->io_ops->prep) {
198 int ret = td->io_ops->prep(td, io_u);
200 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
202 unlock_file(td, io_u->file);
209 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
210 const struct timespec *t)
215 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
216 * server side gets a message from the client
217 * side that the task is finished, and
218 * td->done is set to 1 after td_io_commit(). In this case,
219 * there is no need to reap complete event in server side.
224 if (min > 0 && td->io_ops->commit) {
225 r = td->io_ops->commit(td);
229 if (max > td->cur_depth)
235 if (max && td->io_ops->getevents)
236 r = td->io_ops->getevents(td, min, max, t);
240 * Reflect that our submitted requests were retrieved with
241 * whatever OS async calls are in the underlying engine.
243 td->io_u_in_flight -= r;
244 io_u_mark_complete(td, r);
246 td_verror(td, r, "get_events");
248 dprint(FD_IO, "getevents: %d\n", r);
252 int td_io_queue(struct thread_data *td, struct io_u *io_u)
254 const enum fio_ddir ddir = acct_ddir(io_u);
255 unsigned long buflen = io_u->xfer_buflen;
258 dprint_io_u(io_u, "queue");
259 fio_ro_check(td, io_u);
261 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
262 io_u_set(io_u, IO_U_F_FLIGHT);
264 assert(fio_file_open(io_u->file));
267 * If using a write iolog, store this entry.
274 if (td->io_ops->flags & FIO_SYNCIO) {
275 if (fio_fill_issue_time(td))
276 fio_gettime(&io_u->issue_time, NULL);
279 * only used for iolog
281 if (td->o.read_iolog_file)
282 memcpy(&td->last_issue, &io_u->issue_time,
283 sizeof(struct timeval));
287 td->io_issues[ddir]++;
288 td->io_issue_bytes[ddir] += buflen;
289 td->rate_io_issue_bytes[ddir] += buflen;
292 ret = td->io_ops->queue(td, io_u);
294 unlock_file(td, io_u->file);
296 if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
297 td->io_issues[ddir]--;
298 td->io_issue_bytes[ddir] -= buflen;
299 td->rate_io_issue_bytes[ddir] -= buflen;
303 * If an error was seen and the io engine didn't propagate it
304 * back to 'td', do so.
306 if (io_u->error && !td->error)
307 td_verror(td, io_u->error, "td_io_queue");
310 * Add warning for O_DIRECT so that users have an easier time
311 * spotting potentially bad alignment. If this triggers for the first
312 * IO, then it's likely an alignment problem or because the host fs
313 * does not support O_DIRECT
315 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
318 log_info("fio: first direct IO errored. File system may not "
319 "support direct IO, or iomem_align= is bad. Try "
320 "setting direct=0.\n");
323 if (!td->io_ops->commit || io_u->ddir == DDIR_TRIM) {
324 io_u_mark_submit(td, 1);
325 io_u_mark_complete(td, 1);
328 if (ret == FIO_Q_COMPLETED) {
329 if (ddir_rw(io_u->ddir)) {
330 io_u_mark_depth(td, 1);
331 td->ts.total_io_u[io_u->ddir]++;
333 } else if (ret == FIO_Q_QUEUED) {
338 if (ddir_rw(io_u->ddir))
339 td->ts.total_io_u[io_u->ddir]++;
341 if (td->io_u_queued >= td->o.iodepth_batch) {
342 r = td_io_commit(td);
348 if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
349 if (fio_fill_issue_time(td))
350 fio_gettime(&io_u->issue_time, NULL);
353 * only used for iolog
355 if (td->o.read_iolog_file)
356 memcpy(&td->last_issue, &io_u->issue_time,
357 sizeof(struct timeval));
363 int td_io_init(struct thread_data *td)
367 if (td->io_ops->init) {
368 ret = td->io_ops->init(td);
369 if (ret && td->o.iodepth > 1) {
370 log_err("fio: io engine init failed. Perhaps try"
371 " reducing io depth?\n");
377 if (!ret && (td->io_ops->flags & FIO_NOIO))
378 td->flags |= TD_F_NOIO;
383 int td_io_commit(struct thread_data *td)
387 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
389 if (!td->cur_depth || !td->io_u_queued)
392 io_u_mark_depth(td, td->io_u_queued);
394 if (td->io_ops->commit) {
395 ret = td->io_ops->commit(td);
397 td_verror(td, -ret, "io commit");
401 * Reflect that events were submitted as async IO requests.
403 td->io_u_in_flight += td->io_u_queued;
409 int td_io_open_file(struct thread_data *td, struct fio_file *f)
411 assert(!fio_file_open(f));
414 if (td->io_ops->open_file(td, f)) {
415 if (td->error == EINVAL && td->o.odirect)
416 log_err("fio: destination does not support O_DIRECT\n");
417 if (td->error == EMFILE) {
418 log_err("fio: try reducing/setting openfiles (failed"
419 " at %u of %u)\n", td->nr_open_files,
424 assert(!fio_file_open(f));
428 fio_file_reset(td, f);
429 fio_file_set_open(f);
430 fio_file_clear_closing(f);
431 disk_util_inc(f->du);
436 if (f->filetype == FIO_TYPE_PIPE) {
438 log_err("fio: can't seek on pipes (no random io)\n");
443 if (td->io_ops->flags & FIO_DISKLESSIO)
446 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
449 if (td->o.fadvise_hint &&
450 (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
454 flags = POSIX_FADV_RANDOM;
456 flags = POSIX_FADV_SEQUENTIAL;
458 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
459 td_verror(td, errno, "fadvise");
463 #ifdef FIO_HAVE_STREAMID
464 if (td->o.fadvise_stream &&
465 (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
466 off_t stream = td->o.fadvise_stream;
468 if (posix_fadvise(f->fd, stream, f->io_size, POSIX_FADV_STREAMID) < 0) {
469 td_verror(td, errno, "fadvise streamid");
475 #ifdef FIO_OS_DIRECTIO
477 * Some OS's have a distinct call to mark the file non-buffered,
478 * instead of using O_DIRECT (Solaris)
481 int ret = fio_set_odirect(f->fd);
484 td_verror(td, ret, "fio_set_odirect");
485 log_err("fio: the file system does not seem to support direct IO\n");
492 log_file(td, f, FIO_LOG_OPEN_FILE);
495 disk_util_dec(f->du);
496 if (td->io_ops->close_file)
497 td->io_ops->close_file(td, f);
501 int td_io_close_file(struct thread_data *td, struct fio_file *f)
503 if (!fio_file_closing(f))
504 log_file(td, f, FIO_LOG_CLOSE_FILE);
507 * mark as closing, do real close when last io on it has completed
509 fio_file_set_closing(f);
511 disk_util_dec(f->du);
513 if (td->o.file_lock_mode != FILE_LOCK_NONE)
514 unlock_file_all(td, f);
516 return put_file(td, f);
519 int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
521 if (td->io_ops->unlink_file)
522 return td->io_ops->unlink_file(td, f);
524 return unlink(f->file_name);
527 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
529 if (!td->io_ops->get_file_size)
532 return td->io_ops->get_file_size(td, f);
535 static int do_sync_file_range(const struct thread_data *td,
538 off64_t offset, nbytes;
540 offset = f->first_write;
541 nbytes = f->last_write - f->first_write;
546 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
549 int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
553 if (io_u->ddir == DDIR_SYNC) {
554 ret = fsync(io_u->file->fd);
555 } else if (io_u->ddir == DDIR_DATASYNC) {
556 #ifdef CONFIG_FDATASYNC
557 ret = fdatasync(io_u->file->fd);
559 ret = io_u->xfer_buflen;
560 io_u->error = EINVAL;
562 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
563 ret = do_sync_file_range(td, io_u->file);
565 ret = io_u->xfer_buflen;
566 io_u->error = EINVAL;
575 int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
577 #ifndef FIO_HAVE_TRIM
578 io_u->error = EINVAL;
581 struct fio_file *f = io_u->file;
584 ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
586 return io_u->xfer_buflen;
593 int fio_show_ioengine_help(const char *engine)
595 struct flist_head *entry;
596 struct thread_data td;
600 if (!engine || !*engine) {
601 log_info("Available IO engines:\n");
602 flist_for_each(entry, &engine_list) {
603 td.io_ops = flist_entry(entry, struct ioengine_ops,
605 log_info("\t%s\n", td.io_ops->name);
609 sep = strchr(engine, ',');
615 memset(&td, 0, sizeof(td));
617 td.io_ops = load_ioengine(&td, engine);
619 log_info("IO engine %s not found\n", engine);
623 if (td.io_ops->options)
624 ret = show_cmd_help(td.io_ops->options, sep);
626 log_info("IO engine %s has no options\n", td.io_ops->name);