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 dprint(FD_IO, "converting ioengine name: %s -> libaio\n", name);
138 strcpy(engine, "libaio");
141 dprint(FD_IO, "load ioengine %s\n", engine);
142 return find_ioengine(engine);
145 struct ioengine_ops *load_ioengine(struct thread_data *td)
147 struct ioengine_ops *ops = NULL;
151 * Use ->ioengine_so_path if an external ioengine path is specified.
152 * In this case, ->ioengine is "external" which also means the prefix
153 * for external ioengines "external:" is properly used.
155 name = td->o.ioengine_so_path ?: td->o.ioengine;
158 * Try to load ->ioengine first, and if failed try to dlopen(3) either
159 * ->ioengine or ->ioengine_so_path. This is redundant for an external
160 * ioengine with prefix, and also leaves the possibility of unexpected
161 * behavior (e.g. if the "external" ioengine exists), but we do this
162 * so as not to break job files not using the prefix.
164 ops = __load_ioengine(td->o.ioengine);
166 ops = dlopen_ioengine(td, name);
169 * If ops is NULL, we failed to load ->ioengine, and also failed to
170 * dlopen(3) either ->ioengine or ->ioengine_so_path as a path.
173 log_err("fio: engine %s not loadable\n", name);
178 * Check that the required methods are there.
180 if (check_engine_ops(ops))
187 * For cleaning up an ioengine which never made it to init().
189 void free_ioengine(struct thread_data *td)
191 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
193 if (td->eo && td->io_ops->options) {
194 options_free(td->io_ops->options, td->eo);
199 if (td->io_ops_dlhandle)
200 dlclose(td->io_ops_dlhandle);
205 void close_ioengine(struct thread_data *td)
207 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
209 if (td->io_ops->cleanup) {
210 td->io_ops->cleanup(td);
211 td->io_ops_data = NULL;
217 int td_io_prep(struct thread_data *td, struct io_u *io_u)
219 dprint_io_u(io_u, "prep");
220 fio_ro_check(td, io_u);
222 lock_file(td, io_u->file, io_u->ddir);
224 if (td->io_ops->prep) {
225 int ret = td->io_ops->prep(td, io_u);
227 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
229 unlock_file(td, io_u->file);
236 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
237 const struct timespec *t)
242 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
243 * server side gets a message from the client
244 * side that the task is finished, and
245 * td->done is set to 1 after td_io_commit(). In this case,
246 * there is no need to reap complete event in server side.
251 if (min > 0 && td->io_ops->commit) {
252 r = td->io_ops->commit(td);
256 if (max > td->cur_depth)
262 if (max && td->io_ops->getevents)
263 r = td->io_ops->getevents(td, min, max, t);
267 * Reflect that our submitted requests were retrieved with
268 * whatever OS async calls are in the underlying engine.
270 td->io_u_in_flight -= r;
271 io_u_mark_complete(td, r);
273 td_verror(td, r, "get_events");
275 dprint(FD_IO, "getevents: %d\n", r);
279 int td_io_queue(struct thread_data *td, struct io_u *io_u)
281 const enum fio_ddir ddir = acct_ddir(io_u);
282 unsigned long buflen = io_u->xfer_buflen;
285 dprint_io_u(io_u, "queue");
286 fio_ro_check(td, io_u);
288 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
289 io_u_set(td, io_u, IO_U_F_FLIGHT);
291 assert(fio_file_open(io_u->file));
294 * If using a write iolog, store this entry.
301 if (td_ioengine_flagged(td, FIO_SYNCIO)) {
302 if (fio_fill_issue_time(td))
303 fio_gettime(&io_u->issue_time, NULL);
306 * only used for iolog
308 if (td->o.read_iolog_file)
309 memcpy(&td->last_issue, &io_u->issue_time,
310 sizeof(io_u->issue_time));
314 td->io_issues[ddir]++;
315 td->io_issue_bytes[ddir] += buflen;
316 td->rate_io_issue_bytes[ddir] += buflen;
319 ret = td->io_ops->queue(td, io_u);
321 unlock_file(td, io_u->file);
323 if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
324 td->io_issues[ddir]--;
325 td->io_issue_bytes[ddir] -= buflen;
326 td->rate_io_issue_bytes[ddir] -= buflen;
327 io_u_clear(td, io_u, IO_U_F_FLIGHT);
331 * If an error was seen and the io engine didn't propagate it
332 * back to 'td', do so.
334 if (io_u->error && !td->error)
335 td_verror(td, io_u->error, "td_io_queue");
338 * Add warning for O_DIRECT so that users have an easier time
339 * spotting potentially bad alignment. If this triggers for the first
340 * IO, then it's likely an alignment problem or because the host fs
341 * does not support O_DIRECT
343 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
346 log_info("fio: first direct IO errored. File system may not "
347 "support direct IO, or iomem_align= is bad, or "
348 "invalid block size. Try setting direct=0.\n");
351 if (!td->io_ops->commit || io_u->ddir == DDIR_TRIM) {
352 io_u_mark_submit(td, 1);
353 io_u_mark_complete(td, 1);
356 if (ret == FIO_Q_COMPLETED) {
357 if (ddir_rw(io_u->ddir)) {
358 io_u_mark_depth(td, 1);
359 td->ts.total_io_u[io_u->ddir]++;
361 } else if (ret == FIO_Q_QUEUED) {
366 if (ddir_rw(io_u->ddir))
367 td->ts.total_io_u[io_u->ddir]++;
369 if (td->io_u_queued >= td->o.iodepth_batch) {
370 r = td_io_commit(td);
376 if (!td_ioengine_flagged(td, FIO_SYNCIO)) {
377 if (fio_fill_issue_time(td))
378 fio_gettime(&io_u->issue_time, NULL);
381 * only used for iolog
383 if (td->o.read_iolog_file)
384 memcpy(&td->last_issue, &io_u->issue_time,
385 sizeof(io_u->issue_time));
391 int td_io_init(struct thread_data *td)
395 if (td->io_ops->init) {
396 ret = td->io_ops->init(td);
398 log_err("fio: io engine %s init failed.%s\n",
401 " Perhaps try reducing io depth?" : "");
411 int td_io_commit(struct thread_data *td)
415 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
417 if (!td->cur_depth || !td->io_u_queued)
420 io_u_mark_depth(td, td->io_u_queued);
422 if (td->io_ops->commit) {
423 ret = td->io_ops->commit(td);
425 td_verror(td, -ret, "io commit");
429 * Reflect that events were submitted as async IO requests.
431 td->io_u_in_flight += td->io_u_queued;
437 int td_io_open_file(struct thread_data *td, struct fio_file *f)
439 assert(!fio_file_open(f));
441 assert(td->io_ops->open_file);
443 if (td->io_ops->open_file(td, f)) {
444 if (td->error == EINVAL && td->o.odirect)
445 log_err("fio: destination does not support O_DIRECT\n");
446 if (td->error == EMFILE) {
447 log_err("fio: try reducing/setting openfiles (failed"
448 " at %u of %u)\n", td->nr_open_files,
453 assert(!fio_file_open(f));
457 fio_file_reset(td, f);
458 fio_file_set_open(f);
459 fio_file_clear_closing(f);
460 disk_util_inc(f->du);
465 if (f->filetype == FIO_TYPE_PIPE) {
467 log_err("fio: can't seek on pipes (no random io)\n");
472 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
475 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
478 if (td->o.fadvise_hint != F_ADV_NONE &&
479 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
482 if (td->o.fadvise_hint == F_ADV_TYPE) {
484 flags = POSIX_FADV_RANDOM;
486 flags = POSIX_FADV_SEQUENTIAL;
487 } else if (td->o.fadvise_hint == F_ADV_RANDOM)
488 flags = POSIX_FADV_RANDOM;
489 else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
490 flags = POSIX_FADV_SEQUENTIAL;
492 log_err("fio: unknown fadvise type %d\n",
494 flags = POSIX_FADV_NORMAL;
497 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
498 td_verror(td, errno, "fadvise");
502 #ifdef FIO_HAVE_WRITE_HINT
503 if (fio_option_is_set(&td->o, write_hint) &&
504 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
505 uint64_t hint = td->o.write_hint;
509 * For direct IO, we just need/want to set the hint on
510 * the file descriptor. For buffered IO, we need to set
514 cmd = F_SET_FILE_RW_HINT;
518 if (fcntl(f->fd, cmd, &hint) < 0) {
519 td_verror(td, errno, "fcntl write hint");
525 if (td->o.odirect && !OS_O_DIRECT && fio_set_directio(td, f))
529 log_file(td, f, FIO_LOG_OPEN_FILE);
532 disk_util_dec(f->du);
533 if (td->io_ops->close_file)
534 td->io_ops->close_file(td, f);
538 int td_io_close_file(struct thread_data *td, struct fio_file *f)
540 if (!fio_file_closing(f))
541 log_file(td, f, FIO_LOG_CLOSE_FILE);
544 * mark as closing, do real close when last io on it has completed
546 fio_file_set_closing(f);
548 disk_util_dec(f->du);
550 if (td->o.file_lock_mode != FILE_LOCK_NONE)
551 unlock_file_all(td, f);
553 return put_file(td, f);
556 int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
558 if (td->io_ops->unlink_file)
559 return td->io_ops->unlink_file(td, f);
563 ret = unlink(f->file_name);
571 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
573 if (!td->io_ops->get_file_size)
576 return td->io_ops->get_file_size(td, f);
579 int fio_show_ioengine_help(const char *engine)
581 struct flist_head *entry;
582 struct ioengine_ops *io_ops;
586 if (!engine || !*engine) {
587 log_info("Available IO engines:\n");
588 flist_for_each(entry, &engine_list) {
589 io_ops = flist_entry(entry, struct ioengine_ops, list);
590 log_info("\t%s\n", io_ops->name);
594 sep = strchr(engine, ',');
600 io_ops = __load_ioengine(engine);
602 log_info("IO engine %s not found\n", engine);
607 ret = show_cmd_help(io_ops->options, sep);
609 log_info("IO engine %s has no options\n", io_ops->name);