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));
155 void close_ioengine(struct thread_data *td)
157 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
159 if (td->io_ops->cleanup) {
160 td->io_ops->cleanup(td);
161 td->io_ops->data = NULL;
164 if (td->io_ops->dlhandle)
165 dlclose(td->io_ops->dlhandle);
171 int td_io_prep(struct thread_data *td, struct io_u *io_u)
173 dprint_io_u(io_u, "prep");
174 fio_ro_check(td, io_u);
176 lock_file(td, io_u->file, io_u->ddir);
178 if (td->io_ops->prep) {
179 int ret = td->io_ops->prep(td, io_u);
181 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
183 unlock_file(td, io_u->file);
190 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
195 if (min > 0 && td->io_ops->commit) {
196 r = td->io_ops->commit(td);
200 if (max > td->cur_depth)
206 if (max && td->io_ops->getevents)
207 r = td->io_ops->getevents(td, min, max, t);
210 io_u_mark_complete(td, r);
212 td_verror(td, r, "get_events");
214 dprint(FD_IO, "getevents: %d\n", r);
218 int td_io_queue(struct thread_data *td, struct io_u *io_u)
222 dprint_io_u(io_u, "queue");
223 fio_ro_check(td, io_u);
225 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
226 io_u->flags |= IO_U_F_FLIGHT;
228 assert(fio_file_open(io_u->file));
233 if (td->io_ops->flags & FIO_SYNCIO) {
234 if (fio_fill_issue_time(td))
235 fio_gettime(&io_u->issue_time, NULL);
238 * only used for iolog
240 if (td->o.read_iolog_file)
241 memcpy(&td->last_issue, &io_u->issue_time,
242 sizeof(struct timeval));
245 if (ddir_rw(io_u->ddir))
246 td->io_issues[io_u->ddir]++;
248 ret = td->io_ops->queue(td, io_u);
250 unlock_file(td, io_u->file);
253 * Add warning for O_DIRECT so that users have an easier time
254 * spotting potentially bad alignment. If this triggers for the first
255 * IO, then it's likely an alignment problem or because the host fs
256 * does not support O_DIRECT
258 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
260 log_info("fio: first direct IO errored. File system may not "
261 "support direct IO, or iomem_align= is bad.\n");
264 if (!td->io_ops->commit) {
265 io_u_mark_submit(td, 1);
266 io_u_mark_complete(td, 1);
269 if (ret == FIO_Q_COMPLETED) {
270 if (ddir_rw(io_u->ddir)) {
271 io_u_mark_depth(td, 1);
272 td->ts.total_io_u[io_u->ddir]++;
273 } else if (io_u->ddir == DDIR_TRIM)
274 td->ts.total_io_u[2]++;
275 } else if (ret == FIO_Q_QUEUED) {
278 if (ddir_rw(io_u->ddir)) {
280 td->ts.total_io_u[io_u->ddir]++;
283 if (td->io_u_queued >= td->o.iodepth_batch) {
284 r = td_io_commit(td);
290 if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
291 if (fio_fill_issue_time(td))
292 fio_gettime(&io_u->issue_time, NULL);
295 * only used for iolog
297 if (td->o.read_iolog_file)
298 memcpy(&td->last_issue, &io_u->issue_time,
299 sizeof(struct timeval));
305 int td_io_init(struct thread_data *td)
309 if (td->io_ops->init) {
310 ret = td->io_ops->init(td);
311 if (ret && td->o.iodepth > 1) {
312 log_err("fio: io engine init failed. Perhaps try"
313 " reducing io depth?\n");
322 int td_io_commit(struct thread_data *td)
326 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
328 if (!td->cur_depth || !td->io_u_queued)
331 io_u_mark_depth(td, td->io_u_queued);
334 if (td->io_ops->commit) {
335 ret = td->io_ops->commit(td);
337 td_verror(td, -ret, "io commit");
343 int td_io_open_file(struct thread_data *td, struct fio_file *f)
345 assert(!fio_file_open(f));
348 if (td->io_ops->open_file(td, f)) {
349 if (td->error == EINVAL && td->o.odirect)
350 log_err("fio: destination does not support O_DIRECT\n");
351 if (td->error == EMFILE) {
352 log_err("fio: try reducing/setting openfiles (failed"
353 " at %u of %u)\n", td->nr_open_files,
358 assert(!fio_file_open(f));
363 fio_file_set_open(f);
364 fio_file_clear_closing(f);
365 disk_util_inc(f->du);
370 if (f->filetype == FIO_TYPE_PIPE) {
372 log_err("fio: can't seek on pipes (no random io)\n");
377 if (td->io_ops->flags & FIO_DISKLESSIO)
380 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
383 if (td->o.fadvise_hint &&
384 (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
388 flags = POSIX_FADV_RANDOM;
390 flags = POSIX_FADV_SEQUENTIAL;
392 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
393 td_verror(td, errno, "fadvise");
398 #ifdef FIO_OS_DIRECTIO
400 * Some OS's have a distinct call to mark the file non-buffered,
401 * instead of using O_DIRECT (Solaris)
404 int ret = fio_set_odirect(f->fd);
407 td_verror(td, ret, "fio_set_odirect");
408 log_err("fio: the file system does not seem to support direct IO\n");
415 log_file(td, f, FIO_LOG_OPEN_FILE);
418 disk_util_dec(f->du);
419 if (td->io_ops->close_file)
420 td->io_ops->close_file(td, f);
424 int td_io_close_file(struct thread_data *td, struct fio_file *f)
426 if (!fio_file_closing(f))
427 log_file(td, f, FIO_LOG_CLOSE_FILE);
430 * mark as closing, do real close when last io on it has completed
432 fio_file_set_closing(f);
434 disk_util_dec(f->du);
435 unlock_file_all(td, f);
437 return put_file(td, f);
440 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
442 if (!td->io_ops->get_file_size)
445 return td->io_ops->get_file_size(td, f);
448 static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
450 off64_t offset, nbytes;
452 offset = f->first_write;
453 nbytes = f->last_write - f->first_write;
458 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
461 int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
465 if (io_u->ddir == DDIR_SYNC) {
466 ret = fsync(io_u->file->fd);
467 } else if (io_u->ddir == DDIR_DATASYNC) {
468 #ifdef FIO_HAVE_FDATASYNC
469 ret = fdatasync(io_u->file->fd);
471 ret = io_u->xfer_buflen;
472 io_u->error = EINVAL;
474 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
475 ret = do_sync_file_range(td, io_u->file);
477 ret = io_u->xfer_buflen;
478 io_u->error = EINVAL;
487 int do_io_u_trim(struct thread_data *td, struct io_u *io_u)
489 #ifndef FIO_HAVE_TRIM
490 io_u->error = EINVAL;
493 struct fio_file *f = io_u->file;
496 ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
498 return io_u->xfer_buflen;;