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 int 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)
45 log_err("%s: no event handler\n", ops->name);
48 if (!ops->getevents) {
49 log_err("%s: no getevents handler\n", ops->name);
53 log_err("%s: no queue handler\n", ops->name);
60 void unregister_ioengine(struct ioengine_ops *ops)
62 dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
63 flist_del(&ops->list);
64 INIT_FLIST_HEAD(&ops->list);
67 void register_ioengine(struct ioengine_ops *ops)
69 dprint(FD_IO, "ioengine %s registered\n", ops->name);
70 INIT_FLIST_HEAD(&ops->list);
71 flist_add_tail(&ops->list, &engine_list);
74 static struct ioengine_ops *find_ioengine(const char *name)
76 struct ioengine_ops *ops;
77 struct flist_head *entry;
79 flist_for_each(entry, &engine_list) {
80 ops = flist_entry(entry, struct ioengine_ops, list);
81 if (!strcmp(name, ops->name))
88 static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
89 const char *engine_lib)
91 struct ioengine_ops *ops;
94 dprint(FD_IO, "dload engine %s\n", engine_lib);
97 dlhandle = dlopen(engine_lib, RTLD_LAZY);
99 td_vmsg(td, -1, dlerror(), "dlopen");
104 * Unlike the included modules, external engines should have a
105 * non-static ioengine structure that we can reference.
107 ops = dlsym(dlhandle, "ioengine");
109 td_vmsg(td, -1, dlerror(), "dlsym");
114 ops->dlhandle = dlhandle;
118 struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
120 struct ioengine_ops *ops, *ret;
123 dprint(FD_IO, "load ioengine %s\n", name);
125 strncpy(engine, name, sizeof(engine) - 1);
128 * linux libaio has alias names, so convert to what we want
130 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
131 strcpy(engine, "libaio");
133 ops = find_ioengine(engine);
135 ops = dlopen_ioengine(td, name);
138 log_err("fio: engine %s not loadable\n", name);
143 * Check that the required methods are there.
145 if (check_engine_ops(ops))
148 ret = malloc(sizeof(*ret));
149 memcpy(ret, ops, sizeof(*ret));
156 * For cleaning up an ioengine which never made it to init().
158 void free_ioengine(struct thread_data *td)
160 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
162 if (td->eo && td->io_ops->options) {
163 options_free(td->io_ops->options, td->eo);
168 if (td->io_ops->dlhandle)
169 dlclose(td->io_ops->dlhandle);
175 void close_ioengine(struct thread_data *td)
177 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
179 if (td->io_ops->cleanup) {
180 td->io_ops->cleanup(td);
181 td->io_ops->data = NULL;
187 int td_io_prep(struct thread_data *td, struct io_u *io_u)
189 dprint_io_u(io_u, "prep");
190 fio_ro_check(td, io_u);
192 lock_file(td, io_u->file, io_u->ddir);
194 if (td->io_ops->prep) {
195 int ret = td->io_ops->prep(td, io_u);
197 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
199 unlock_file(td, io_u->file);
206 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
211 if (min > 0 && td->io_ops->commit) {
212 r = td->io_ops->commit(td);
216 if (max > td->cur_depth)
222 if (max && td->io_ops->getevents)
223 r = td->io_ops->getevents(td, min, max, t);
227 * Reflect that our submitted requests were retrieved with
228 * whatever OS async calls are in the underlying engine.
230 td->io_u_in_flight -= r;
231 io_u_mark_complete(td, r);
233 td_verror(td, r, "get_events");
235 dprint(FD_IO, "getevents: %d\n", r);
239 int td_io_queue(struct thread_data *td, struct io_u *io_u)
243 dprint_io_u(io_u, "queue");
244 fio_ro_check(td, io_u);
246 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
247 io_u->flags |= IO_U_F_FLIGHT;
249 assert(fio_file_open(io_u->file));
254 if (td->io_ops->flags & FIO_SYNCIO) {
255 if (fio_fill_issue_time(td))
256 fio_gettime(&io_u->issue_time, NULL);
259 * only used for iolog
261 if (td->o.read_iolog_file)
262 memcpy(&td->last_issue, &io_u->issue_time,
263 sizeof(struct timeval));
266 if (ddir_rw(io_u->ddir))
267 td->io_issues[io_u->ddir]++;
269 ret = td->io_ops->queue(td, io_u);
271 unlock_file(td, io_u->file);
274 * Add warning for O_DIRECT so that users have an easier time
275 * spotting potentially bad alignment. If this triggers for the first
276 * IO, then it's likely an alignment problem or because the host fs
277 * does not support O_DIRECT
279 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
281 log_info("fio: first direct IO errored. File system may not "
282 "support direct IO, or iomem_align= is bad.\n");
285 if (!td->io_ops->commit) {
286 io_u_mark_submit(td, 1);
287 io_u_mark_complete(td, 1);
290 if (ret == FIO_Q_COMPLETED) {
291 if (ddir_rw(io_u->ddir)) {
292 io_u_mark_depth(td, 1);
293 td->ts.total_io_u[io_u->ddir]++;
294 } else if (io_u->ddir == DDIR_TRIM)
295 td->ts.total_io_u[2]++;
296 } else if (ret == FIO_Q_QUEUED) {
299 if (ddir_rw(io_u->ddir)) {
301 td->ts.total_io_u[io_u->ddir]++;
304 if (td->io_u_queued >= td->o.iodepth_batch) {
305 r = td_io_commit(td);
311 if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
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(struct timeval));
326 int td_io_init(struct thread_data *td)
330 if (td->io_ops->init) {
331 ret = td->io_ops->init(td);
332 if (ret && td->o.iodepth > 1) {
333 log_err("fio: io engine init failed. Perhaps try"
334 " reducing io depth?\n");
343 int td_io_commit(struct thread_data *td)
347 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
349 if (!td->cur_depth || !td->io_u_queued)
352 io_u_mark_depth(td, td->io_u_queued);
354 if (td->io_ops->commit) {
355 ret = td->io_ops->commit(td);
357 td_verror(td, -ret, "io commit");
361 * Reflect that events were submitted as async IO requests.
363 td->io_u_in_flight += td->io_u_queued;
369 int td_io_open_file(struct thread_data *td, struct fio_file *f)
371 assert(!fio_file_open(f));
374 if (td->io_ops->open_file(td, f)) {
375 if (td->error == EINVAL && td->o.odirect)
376 log_err("fio: destination does not support O_DIRECT\n");
377 if (td->error == EMFILE) {
378 log_err("fio: try reducing/setting openfiles (failed"
379 " at %u of %u)\n", td->nr_open_files,
384 assert(!fio_file_open(f));
389 fio_file_set_open(f);
390 fio_file_clear_closing(f);
391 disk_util_inc(f->du);
396 if (f->filetype == FIO_TYPE_PIPE) {
398 log_err("fio: can't seek on pipes (no random io)\n");
403 if (td->io_ops->flags & FIO_DISKLESSIO)
406 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
409 if (td->o.fadvise_hint &&
410 (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
414 flags = POSIX_FADV_RANDOM;
416 flags = POSIX_FADV_SEQUENTIAL;
418 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
419 td_verror(td, errno, "fadvise");
424 #ifdef FIO_OS_DIRECTIO
426 * Some OS's have a distinct call to mark the file non-buffered,
427 * instead of using O_DIRECT (Solaris)
430 int ret = fio_set_odirect(f->fd);
433 td_verror(td, ret, "fio_set_odirect");
434 log_err("fio: the file system does not seem to support direct IO\n");
441 log_file(td, f, FIO_LOG_OPEN_FILE);
444 disk_util_dec(f->du);
445 if (td->io_ops->close_file)
446 td->io_ops->close_file(td, f);
450 int td_io_close_file(struct thread_data *td, struct fio_file *f)
452 if (!fio_file_closing(f))
453 log_file(td, f, FIO_LOG_CLOSE_FILE);
456 * mark as closing, do real close when last io on it has completed
458 fio_file_set_closing(f);
460 disk_util_dec(f->du);
461 unlock_file_all(td, f);
463 return put_file(td, f);
466 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
468 if (!td->io_ops->get_file_size)
471 return td->io_ops->get_file_size(td, f);
474 static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
476 off64_t offset, nbytes;
478 offset = f->first_write;
479 nbytes = f->last_write - f->first_write;
484 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
487 int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
491 if (io_u->ddir == DDIR_SYNC) {
492 ret = fsync(io_u->file->fd);
493 } else if (io_u->ddir == DDIR_DATASYNC) {
494 #ifdef FIO_HAVE_FDATASYNC
495 ret = fdatasync(io_u->file->fd);
497 ret = io_u->xfer_buflen;
498 io_u->error = EINVAL;
500 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
501 ret = do_sync_file_range(td, io_u->file);
503 ret = io_u->xfer_buflen;
504 io_u->error = EINVAL;
513 int do_io_u_trim(struct thread_data *td, struct io_u *io_u)
515 #ifndef FIO_HAVE_TRIM
516 io_u->error = EINVAL;
519 struct fio_file *f = io_u->file;
522 ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
524 return io_u->xfer_buflen;;
531 int fio_show_ioengine_help(const char *engine)
533 struct flist_head *entry;
534 struct thread_data td;
538 if (!engine || !*engine) {
539 log_info("Available IO engines:\n");
540 flist_for_each(entry, &engine_list) {
541 td.io_ops = flist_entry(entry, struct ioengine_ops,
543 log_info("\t%s\n", td.io_ops->name);
547 sep = strchr(engine, ',');
553 memset(&td, 0, sizeof(td));
555 td.io_ops = load_ioengine(&td, engine);
557 log_info("IO engine %s not found\n", engine);
561 if (td.io_ops->options)
562 ret = show_cmd_help(td.io_ops->options, sep);
564 log_info("IO engine %s has no options\n", td.io_ops->name);