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.
22 static FLIST_HEAD(engine_list);
24 static int check_engine_ops(struct ioengine_ops *ops)
26 if (ops->version != FIO_IOOPS_VERSION) {
27 log_err("bad ioops version %d (want %d)\n", ops->version,
33 log_err("%s: no queue handler\n", ops->name);
38 * sync engines only need a ->queue()
40 if (ops->flags & FIO_SYNCIO)
44 log_err("%s: no event handler\n", ops->name);
47 if (!ops->getevents) {
48 log_err("%s: no getevents handler\n", ops->name);
52 log_err("%s: no queue handler\n", ops->name);
59 void unregister_ioengine(struct ioengine_ops *ops)
61 dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
62 flist_del(&ops->list);
63 INIT_FLIST_HEAD(&ops->list);
66 void register_ioengine(struct ioengine_ops *ops)
68 dprint(FD_IO, "ioengine %s registered\n", ops->name);
69 INIT_FLIST_HEAD(&ops->list);
70 flist_add_tail(&ops->list, &engine_list);
73 static struct ioengine_ops *find_ioengine(const char *name)
75 struct ioengine_ops *ops;
76 struct flist_head *entry;
78 flist_for_each(entry, &engine_list) {
79 ops = flist_entry(entry, struct ioengine_ops, list);
80 if (!strcmp(name, ops->name))
87 static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
88 const char *engine_lib)
90 struct ioengine_ops *ops;
93 dprint(FD_IO, "dload engine %s\n", engine_lib);
96 dlhandle = dlopen(engine_lib, RTLD_LAZY);
98 td_vmsg(td, -1, dlerror(), "dlopen");
103 * Unlike the included modules, external engines should have a
104 * non-static ioengine structure that we can reference.
106 ops = dlsym(dlhandle, "ioengine");
108 td_vmsg(td, -1, dlerror(), "dlsym");
113 ops->dlhandle = dlhandle;
117 struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
119 struct ioengine_ops *ops, *ret;
122 dprint(FD_IO, "load ioengine %s\n", name);
124 strncpy(engine, name, sizeof(engine) - 1);
127 * linux libaio has alias names, so convert to what we want
129 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
130 strcpy(engine, "libaio");
132 ops = find_ioengine(engine);
134 ops = dlopen_ioengine(td, name);
137 log_err("fio: engine %s not loadable\n", name);
142 * Check that the required methods are there.
144 if (check_engine_ops(ops))
147 ret = malloc(sizeof(*ret));
148 memcpy(ret, ops, sizeof(*ret));
154 void close_ioengine(struct thread_data *td)
156 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
158 if (td->io_ops->cleanup) {
159 td->io_ops->cleanup(td);
160 td->io_ops->data = NULL;
163 if (td->io_ops->dlhandle)
164 dlclose(td->io_ops->dlhandle);
170 int td_io_prep(struct thread_data *td, struct io_u *io_u)
172 dprint_io_u(io_u, "prep");
173 fio_ro_check(td, io_u);
175 lock_file(td, io_u->file, io_u->ddir);
177 if (td->io_ops->prep) {
178 int ret = td->io_ops->prep(td, io_u);
180 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
182 unlock_file(td, io_u->file);
189 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
194 if (min > 0 && td->io_ops->commit) {
195 r = td->io_ops->commit(td);
199 if (max > td->cur_depth)
205 if (max && td->io_ops->getevents)
206 r = td->io_ops->getevents(td, min, max, t);
209 io_u_mark_complete(td, r);
211 td_verror(td, r, "get_events");
213 dprint(FD_IO, "getevents: %d\n", r);
217 int td_io_queue(struct thread_data *td, struct io_u *io_u)
221 dprint_io_u(io_u, "queue");
222 fio_ro_check(td, io_u);
224 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
225 io_u->flags |= IO_U_F_FLIGHT;
227 assert(fio_file_open(io_u->file));
232 if (td->io_ops->flags & FIO_SYNCIO) {
233 if (fio_fill_issue_time(td))
234 fio_gettime(&io_u->issue_time, NULL);
237 * only used for iolog
239 if (td->o.read_iolog_file)
240 memcpy(&td->last_issue, &io_u->issue_time,
241 sizeof(struct timeval));
244 if (!ddir_sync(io_u->ddir))
245 td->io_issues[io_u->ddir]++;
247 ret = td->io_ops->queue(td, io_u);
249 unlock_file(td, io_u->file);
252 * Add warning for O_DIRECT so that users have an easier time
253 * spotting potentially bad alignment. If this triggers for the first
254 * IO, then it's likely an alignment problem or because the host fs
255 * does not support O_DIRECT
257 if (io_u->error == EINVAL && td->io_issues[io_u->ddir] == 1 &&
259 log_info("fio: first direct IO errored. File system may not "
260 "support direct IO, or iomem_align= is bad.\n");
263 if (!td->io_ops->commit) {
264 io_u_mark_submit(td, 1);
265 io_u_mark_complete(td, 1);
268 if (ret == FIO_Q_COMPLETED) {
269 if (!ddir_sync(io_u->ddir)) {
270 io_u_mark_depth(td, 1);
271 td->ts.total_io_u[io_u->ddir]++;
273 } else if (ret == FIO_Q_QUEUED) {
276 if (!ddir_sync(io_u->ddir)) {
278 td->ts.total_io_u[io_u->ddir]++;
281 if (td->io_u_queued >= td->o.iodepth_batch) {
282 r = td_io_commit(td);
288 if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
289 if (fio_fill_issue_time(td))
290 fio_gettime(&io_u->issue_time, NULL);
293 * only used for iolog
295 if (td->o.read_iolog_file)
296 memcpy(&td->last_issue, &io_u->issue_time,
297 sizeof(struct timeval));
303 int td_io_init(struct thread_data *td)
307 if (td->io_ops->init) {
308 ret = td->io_ops->init(td);
309 if (ret && td->o.iodepth > 1) {
310 log_err("fio: io engine init failed. Perhaps try"
311 " reducing io depth?\n");
318 int td_io_commit(struct thread_data *td)
322 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
324 if (!td->cur_depth || !td->io_u_queued)
327 io_u_mark_depth(td, td->io_u_queued);
330 if (td->io_ops->commit) {
331 ret = td->io_ops->commit(td);
333 td_verror(td, -ret, "io commit");
339 int td_io_open_file(struct thread_data *td, struct fio_file *f)
341 assert(!fio_file_open(f));
344 if (td->io_ops->open_file(td, f)) {
345 if (td->error == EINVAL && td->o.odirect)
346 log_err("fio: destination does not support O_DIRECT\n");
347 if (td->error == EMFILE) {
348 log_err("fio: try reducing/setting openfiles (failed"
349 " at %u of %u)\n", td->nr_open_files,
354 assert(!fio_file_open(f));
359 fio_file_set_open(f);
360 fio_file_clear_closing(f);
361 disk_util_inc(f->du);
366 if (f->filetype == FIO_TYPE_PIPE) {
368 log_err("fio: can't seek on pipes (no random io)\n");
373 if (td->io_ops->flags & FIO_DISKLESSIO)
376 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
379 if (td->o.fadvise_hint &&
380 (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
384 flags = POSIX_FADV_RANDOM;
386 flags = POSIX_FADV_SEQUENTIAL;
388 if (fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
389 td_verror(td, errno, "fadvise");
394 #ifdef FIO_OS_DIRECTIO
396 * Some OS's have a distinct call to mark the file non-buffered,
397 * instead of using O_DIRECT (Solaris)
400 int ret = fio_set_odirect(f->fd);
403 td_verror(td, ret, "fio_set_odirect");
410 log_file(td, f, FIO_LOG_OPEN_FILE);
413 disk_util_dec(f->du);
414 if (td->io_ops->close_file)
415 td->io_ops->close_file(td, f);
419 int td_io_close_file(struct thread_data *td, struct fio_file *f)
421 if (!fio_file_closing(f))
422 log_file(td, f, FIO_LOG_CLOSE_FILE);
425 * mark as closing, do real close when last io on it has completed
427 fio_file_set_closing(f);
429 disk_util_dec(f->du);
430 unlock_file_all(td, f);
432 return put_file(td, f);
435 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
437 if (!td->io_ops->get_file_size)
440 return td->io_ops->get_file_size(td, f);
443 static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
445 off64_t offset, nbytes;
447 offset = f->first_write;
448 nbytes = f->last_write - f->first_write;
453 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
456 int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
460 if (io_u->ddir == DDIR_SYNC) {
461 ret = fsync(io_u->file->fd);
462 } else if (io_u->ddir == DDIR_DATASYNC) {
463 #ifdef FIO_HAVE_FDATASYNC
464 ret = fdatasync(io_u->file->fd);
466 ret = io_u->xfer_buflen;
467 io_u->error = EINVAL;
469 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
470 ret = do_sync_file_range(td, io_u->file);
472 ret = io_u->xfer_buflen;
473 io_u->error = EINVAL;
482 int do_io_u_trim(struct thread_data *td, struct io_u *io_u)
484 #ifndef FIO_HAVE_TRIM
485 io_u->error = EINVAL;
486 return io_u->xfer_buflen;
488 struct fio_file *f = io_u->file;
491 ret = os_trim(f->fd, io_u->offset + f->file_offset, io_u->xfer_buflen);
496 return io_u->xfer_buflen;