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 &&
282 log_info("fio: first direct IO errored. File system may not "
283 "support direct IO, or iomem_align= is bad.\n");
286 if (!td->io_ops->commit) {
287 io_u_mark_submit(td, 1);
288 io_u_mark_complete(td, 1);
291 if (ret == FIO_Q_COMPLETED) {
292 if (ddir_rw(io_u->ddir)) {
293 io_u_mark_depth(td, 1);
294 td->ts.total_io_u[io_u->ddir]++;
295 } else if (io_u->ddir == DDIR_TRIM)
296 td->ts.total_io_u[2]++;
297 } else if (ret == FIO_Q_QUEUED) {
300 if (ddir_rw(io_u->ddir)) {
302 td->ts.total_io_u[io_u->ddir]++;
305 if (td->io_u_queued >= td->o.iodepth_batch) {
306 r = td_io_commit(td);
312 if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
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(struct timeval));
327 int td_io_init(struct thread_data *td)
331 if (td->io_ops->init) {
332 ret = td->io_ops->init(td);
333 if (ret && td->o.iodepth > 1) {
334 log_err("fio: io engine init failed. Perhaps try"
335 " reducing io depth?\n");
344 int td_io_commit(struct thread_data *td)
348 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
350 if (!td->cur_depth || !td->io_u_queued)
353 io_u_mark_depth(td, td->io_u_queued);
355 if (td->io_ops->commit) {
356 ret = td->io_ops->commit(td);
358 td_verror(td, -ret, "io commit");
362 * Reflect that events were submitted as async IO requests.
364 td->io_u_in_flight += td->io_u_queued;
370 int td_io_open_file(struct thread_data *td, struct fio_file *f)
372 assert(!fio_file_open(f));
375 if (td->io_ops->open_file(td, f)) {
376 if (td->error == EINVAL && td->o.odirect)
377 log_err("fio: destination does not support O_DIRECT\n");
378 if (td->error == EMFILE) {
379 log_err("fio: try reducing/setting openfiles (failed"
380 " at %u of %u)\n", td->nr_open_files,
385 assert(!fio_file_open(f));
390 fio_file_set_open(f);
391 fio_file_clear_closing(f);
392 disk_util_inc(f->du);
397 if (f->filetype == FIO_TYPE_PIPE) {
399 log_err("fio: can't seek on pipes (no random io)\n");
404 if (td->io_ops->flags & FIO_DISKLESSIO)
407 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
410 if (td->o.fadvise_hint &&
411 (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
415 flags = POSIX_FADV_RANDOM;
417 flags = POSIX_FADV_SEQUENTIAL;
419 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
420 td_verror(td, errno, "fadvise");
425 #ifdef FIO_OS_DIRECTIO
427 * Some OS's have a distinct call to mark the file non-buffered,
428 * instead of using O_DIRECT (Solaris)
431 int ret = fio_set_odirect(f->fd);
434 td_verror(td, ret, "fio_set_odirect");
435 log_err("fio: the file system does not seem to support direct IO\n");
442 log_file(td, f, FIO_LOG_OPEN_FILE);
445 disk_util_dec(f->du);
446 if (td->io_ops->close_file)
447 td->io_ops->close_file(td, f);
451 int td_io_close_file(struct thread_data *td, struct fio_file *f)
453 if (!fio_file_closing(f))
454 log_file(td, f, FIO_LOG_CLOSE_FILE);
457 * mark as closing, do real close when last io on it has completed
459 fio_file_set_closing(f);
461 disk_util_dec(f->du);
462 unlock_file_all(td, f);
464 return put_file(td, f);
467 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
469 if (!td->io_ops->get_file_size)
472 return td->io_ops->get_file_size(td, f);
475 static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
477 off64_t offset, nbytes;
479 offset = f->first_write;
480 nbytes = f->last_write - f->first_write;
485 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
488 int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
492 if (io_u->ddir == DDIR_SYNC) {
493 ret = fsync(io_u->file->fd);
494 } else if (io_u->ddir == DDIR_DATASYNC) {
495 #ifdef FIO_HAVE_FDATASYNC
496 ret = fdatasync(io_u->file->fd);
498 ret = io_u->xfer_buflen;
499 io_u->error = EINVAL;
501 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
502 ret = do_sync_file_range(td, io_u->file);
504 ret = io_u->xfer_buflen;
505 io_u->error = EINVAL;
514 int do_io_u_trim(struct thread_data *td, struct io_u *io_u)
516 #ifndef FIO_HAVE_TRIM
517 io_u->error = EINVAL;
520 struct fio_file *f = io_u->file;
523 ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
525 return io_u->xfer_buflen;;
532 int fio_show_ioengine_help(const char *engine)
534 struct flist_head *entry;
535 struct thread_data td;
539 if (!engine || !*engine) {
540 log_info("Available IO engines:\n");
541 flist_for_each(entry, &engine_list) {
542 td.io_ops = flist_entry(entry, struct ioengine_ops,
544 log_info("\t%s\n", td.io_ops->name);
548 sep = strchr(engine, ',');
554 memset(&td, 0, sizeof(td));
556 td.io_ops = load_ioengine(&td, engine);
558 log_info("IO engine %s not found\n", engine);
562 if (td.io_ops->options)
563 ret = show_cmd_help(td.io_ops->options, sep);
565 log_info("IO engine %s has no options\n", td.io_ops->name);