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 static struct ioengine_ops *__load_ioengine(const char *name)
130 engine[sizeof(engine) - 1] = '\0';
131 strncpy(engine, name, sizeof(engine) - 1);
134 * linux libaio has alias names, so convert to what we want
136 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
137 strcpy(engine, "libaio");
139 dprint(FD_IO, "load ioengine %s\n", engine);
140 return find_ioengine(engine);
143 struct ioengine_ops *load_ioengine(struct thread_data *td)
145 struct ioengine_ops *ops = NULL;
149 * Use ->ioengine_so_path if an external ioengine path is specified.
150 * In this case, ->ioengine is "external" which also means the prefix
151 * for external ioengines "external:" is properly used.
153 name = td->o.ioengine_so_path ?: td->o.ioengine;
156 * Try to load ->ioengine first, and if failed try to dlopen(3) either
157 * ->ioengine or ->ioengine_so_path. This is redundant for an external
158 * ioengine with prefix, and also leaves the possibility of unexpected
159 * behavior (e.g. if the "external" ioengine exists), but we do this
160 * so as not to break job files not using the prefix.
162 ops = __load_ioengine(td->o.ioengine);
164 ops = dlopen_ioengine(td, name);
167 * If ops is NULL, we failed to load ->ioengine, and also failed to
168 * dlopen(3) either ->ioengine or ->ioengine_so_path as a path.
171 log_err("fio: engine %s not loadable\n", name);
176 * Check that the required methods are there.
178 if (check_engine_ops(ops))
185 * For cleaning up an ioengine which never made it to init().
187 void free_ioengine(struct thread_data *td)
189 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
191 if (td->eo && td->io_ops->options) {
192 options_free(td->io_ops->options, td->eo);
197 if (td->io_ops_dlhandle)
198 dlclose(td->io_ops_dlhandle);
203 void close_ioengine(struct thread_data *td)
205 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
207 if (td->io_ops->cleanup) {
208 td->io_ops->cleanup(td);
209 td->io_ops_data = NULL;
215 int td_io_prep(struct thread_data *td, struct io_u *io_u)
217 dprint_io_u(io_u, "prep");
218 fio_ro_check(td, io_u);
220 lock_file(td, io_u->file, io_u->ddir);
222 if (td->io_ops->prep) {
223 int ret = td->io_ops->prep(td, io_u);
225 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
227 unlock_file(td, io_u->file);
234 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
235 const struct timespec *t)
240 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
241 * server side gets a message from the client
242 * side that the task is finished, and
243 * td->done is set to 1 after td_io_commit(). In this case,
244 * there is no need to reap complete event in server side.
249 if (min > 0 && td->io_ops->commit) {
250 r = td->io_ops->commit(td);
254 if (max > td->cur_depth)
260 if (max && td->io_ops->getevents)
261 r = td->io_ops->getevents(td, min, max, t);
265 * Reflect that our submitted requests were retrieved with
266 * whatever OS async calls are in the underlying engine.
268 td->io_u_in_flight -= r;
269 io_u_mark_complete(td, r);
271 td_verror(td, r, "get_events");
273 dprint(FD_IO, "getevents: %d\n", r);
277 int td_io_queue(struct thread_data *td, struct io_u *io_u)
279 const enum fio_ddir ddir = acct_ddir(io_u);
280 unsigned long buflen = io_u->xfer_buflen;
283 dprint_io_u(io_u, "queue");
284 fio_ro_check(td, io_u);
286 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
287 io_u_set(td, io_u, IO_U_F_FLIGHT);
289 assert(fio_file_open(io_u->file));
292 * If using a write iolog, store this entry.
299 if (td_ioengine_flagged(td, FIO_SYNCIO)) {
300 if (fio_fill_issue_time(td))
301 fio_gettime(&io_u->issue_time, NULL);
304 * only used for iolog
306 if (td->o.read_iolog_file)
307 memcpy(&td->last_issue, &io_u->issue_time,
308 sizeof(io_u->issue_time));
312 td->io_issues[ddir]++;
313 td->io_issue_bytes[ddir] += buflen;
314 td->rate_io_issue_bytes[ddir] += buflen;
317 ret = td->io_ops->queue(td, io_u);
319 unlock_file(td, io_u->file);
321 if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
322 td->io_issues[ddir]--;
323 td->io_issue_bytes[ddir] -= buflen;
324 td->rate_io_issue_bytes[ddir] -= buflen;
325 io_u_clear(td, io_u, IO_U_F_FLIGHT);
329 * If an error was seen and the io engine didn't propagate it
330 * back to 'td', do so.
332 if (io_u->error && !td->error)
333 td_verror(td, io_u->error, "td_io_queue");
336 * Add warning for O_DIRECT so that users have an easier time
337 * spotting potentially bad alignment. If this triggers for the first
338 * IO, then it's likely an alignment problem or because the host fs
339 * does not support O_DIRECT
341 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
344 log_info("fio: first direct IO errored. File system may not "
345 "support direct IO, or iomem_align= is bad. Try "
346 "setting direct=0.\n");
349 if (!td->io_ops->commit || io_u->ddir == DDIR_TRIM) {
350 io_u_mark_submit(td, 1);
351 io_u_mark_complete(td, 1);
354 if (ret == FIO_Q_COMPLETED) {
355 if (ddir_rw(io_u->ddir)) {
356 io_u_mark_depth(td, 1);
357 td->ts.total_io_u[io_u->ddir]++;
359 } else if (ret == FIO_Q_QUEUED) {
364 if (ddir_rw(io_u->ddir))
365 td->ts.total_io_u[io_u->ddir]++;
367 if (td->io_u_queued >= td->o.iodepth_batch) {
368 r = td_io_commit(td);
374 if (!td_ioengine_flagged(td, FIO_SYNCIO)) {
375 if (fio_fill_issue_time(td))
376 fio_gettime(&io_u->issue_time, NULL);
379 * only used for iolog
381 if (td->o.read_iolog_file)
382 memcpy(&td->last_issue, &io_u->issue_time,
383 sizeof(io_u->issue_time));
389 int td_io_init(struct thread_data *td)
393 if (td->io_ops->init) {
394 ret = td->io_ops->init(td);
396 log_err("fio: io engine %s init failed.%s\n",
399 " Perhaps try reducing io depth?" : "");
409 int td_io_commit(struct thread_data *td)
413 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
415 if (!td->cur_depth || !td->io_u_queued)
418 io_u_mark_depth(td, td->io_u_queued);
420 if (td->io_ops->commit) {
421 ret = td->io_ops->commit(td);
423 td_verror(td, -ret, "io commit");
427 * Reflect that events were submitted as async IO requests.
429 td->io_u_in_flight += td->io_u_queued;
435 int td_io_open_file(struct thread_data *td, struct fio_file *f)
437 assert(!fio_file_open(f));
440 if (td->io_ops->open_file(td, f)) {
441 if (td->error == EINVAL && td->o.odirect)
442 log_err("fio: destination does not support O_DIRECT\n");
443 if (td->error == EMFILE) {
444 log_err("fio: try reducing/setting openfiles (failed"
445 " at %u of %u)\n", td->nr_open_files,
450 assert(!fio_file_open(f));
454 fio_file_reset(td, f);
455 fio_file_set_open(f);
456 fio_file_clear_closing(f);
457 disk_util_inc(f->du);
462 if (f->filetype == FIO_TYPE_PIPE) {
464 log_err("fio: can't seek on pipes (no random io)\n");
469 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
472 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
475 if (td->o.fadvise_hint != F_ADV_NONE &&
476 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
479 if (td->o.fadvise_hint == F_ADV_TYPE) {
481 flags = POSIX_FADV_RANDOM;
483 flags = POSIX_FADV_SEQUENTIAL;
484 } else if (td->o.fadvise_hint == F_ADV_RANDOM)
485 flags = POSIX_FADV_RANDOM;
486 else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
487 flags = POSIX_FADV_SEQUENTIAL;
489 log_err("fio: unknown fadvise type %d\n",
491 flags = POSIX_FADV_NORMAL;
494 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
495 td_verror(td, errno, "fadvise");
499 #ifdef FIO_HAVE_WRITE_HINT
500 if (fio_option_is_set(&td->o, write_hint) &&
501 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
502 uint64_t hint = td->o.write_hint;
506 * For direct IO, we just need/want to set the hint on
507 * the file descriptor. For buffered IO, we need to set
511 cmd = F_SET_FILE_RW_HINT;
515 if (fcntl(f->fd, cmd, &hint) < 0) {
516 td_verror(td, errno, "fcntl write hint");
522 if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
526 log_file(td, f, FIO_LOG_OPEN_FILE);
529 disk_util_dec(f->du);
530 if (td->io_ops->close_file)
531 td->io_ops->close_file(td, f);
535 int td_io_close_file(struct thread_data *td, struct fio_file *f)
537 if (!fio_file_closing(f))
538 log_file(td, f, FIO_LOG_CLOSE_FILE);
541 * mark as closing, do real close when last io on it has completed
543 fio_file_set_closing(f);
545 disk_util_dec(f->du);
547 if (td->o.file_lock_mode != FILE_LOCK_NONE)
548 unlock_file_all(td, f);
550 return put_file(td, f);
553 int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
555 if (td->io_ops->unlink_file)
556 return td->io_ops->unlink_file(td, f);
560 ret = unlink(f->file_name);
568 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
570 if (!td->io_ops->get_file_size)
573 return td->io_ops->get_file_size(td, f);
576 int fio_show_ioengine_help(const char *engine)
578 struct flist_head *entry;
579 struct ioengine_ops *io_ops;
583 if (!engine || !*engine) {
584 log_info("Available IO engines:\n");
585 flist_for_each(entry, &engine_list) {
586 io_ops = flist_entry(entry, struct ioengine_ops, list);
587 log_info("\t%s\n", io_ops->name);
591 sep = strchr(engine, ',');
597 io_ops = __load_ioengine(engine);
599 log_info("IO engine %s not found\n", engine);
604 ret = show_cmd_help(io_ops->options, sep);
606 log_info("IO engine %s has no options\n", io_ops->name);