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, engine_lib);
109 ops = dlsym(dlhandle, "ioengine");
111 td_vmsg(td, -1, dlerror(), "dlsym");
116 ops->dlhandle = dlhandle;
120 struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
122 struct ioengine_ops *ops, *ret;
125 dprint(FD_IO, "load ioengine %s\n", name);
127 strncpy(engine, name, sizeof(engine) - 1);
130 * linux libaio has alias names, so convert to what we want
132 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
133 strcpy(engine, "libaio");
135 ops = find_ioengine(engine);
137 ops = dlopen_ioengine(td, name);
140 log_err("fio: engine %s not loadable\n", name);
145 * Check that the required methods are there.
147 if (check_engine_ops(ops))
150 ret = malloc(sizeof(*ret));
151 memcpy(ret, ops, sizeof(*ret));
158 * For cleaning up an ioengine which never made it to init().
160 void free_ioengine(struct thread_data *td)
162 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
164 if (td->eo && td->io_ops->options) {
165 options_free(td->io_ops->options, td->eo);
170 if (td->io_ops->dlhandle)
171 dlclose(td->io_ops->dlhandle);
177 void close_ioengine(struct thread_data *td)
179 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
181 if (td->io_ops->cleanup) {
182 td->io_ops->cleanup(td);
183 td->io_ops->data = NULL;
189 int td_io_prep(struct thread_data *td, struct io_u *io_u)
191 dprint_io_u(io_u, "prep");
192 fio_ro_check(td, io_u);
194 lock_file(td, io_u->file, io_u->ddir);
196 if (td->io_ops->prep) {
197 int ret = td->io_ops->prep(td, io_u);
199 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
201 unlock_file(td, io_u->file);
208 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
214 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
215 * server side gets a message from the client
216 * side that the task is finished, and
217 * td->done is set to 1 after td_io_commit(). In this case,
218 * there is no need to reap complete event in server side.
223 if (min > 0 && td->io_ops->commit) {
224 r = td->io_ops->commit(td);
228 if (max > td->cur_depth)
234 if (max && td->io_ops->getevents)
235 r = td->io_ops->getevents(td, min, max, t);
239 * Reflect that our submitted requests were retrieved with
240 * whatever OS async calls are in the underlying engine.
242 td->io_u_in_flight -= r;
243 io_u_mark_complete(td, r);
245 td_verror(td, r, "get_events");
247 dprint(FD_IO, "getevents: %d\n", r);
251 int td_io_queue(struct thread_data *td, struct io_u *io_u)
255 dprint_io_u(io_u, "queue");
256 fio_ro_check(td, io_u);
258 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
259 io_u->flags |= IO_U_F_FLIGHT;
261 assert(fio_file_open(io_u->file));
264 * If using a write iolog, store this entry.
271 if (td->io_ops->flags & FIO_SYNCIO) {
272 if (fio_fill_issue_time(td))
273 fio_gettime(&io_u->issue_time, NULL);
276 * only used for iolog
278 if (td->o.read_iolog_file)
279 memcpy(&td->last_issue, &io_u->issue_time,
280 sizeof(struct timeval));
283 if (ddir_rw(acct_ddir(io_u)))
284 td->io_issues[acct_ddir(io_u)]++;
286 ret = td->io_ops->queue(td, io_u);
288 unlock_file(td, io_u->file);
291 * Add warning for O_DIRECT so that users have an easier time
292 * spotting potentially bad alignment. If this triggers for the first
293 * IO, then it's likely an alignment problem or because the host fs
294 * does not support O_DIRECT
296 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
299 log_info("fio: first direct IO errored. File system may not "
300 "support direct IO, or iomem_align= is bad.\n");
303 if (!td->io_ops->commit || ddir_trim(io_u->ddir)) {
304 io_u_mark_submit(td, 1);
305 io_u_mark_complete(td, 1);
308 if (ret == FIO_Q_COMPLETED) {
309 if (ddir_rw(io_u->ddir)) {
310 io_u_mark_depth(td, 1);
311 td->ts.total_io_u[io_u->ddir]++;
313 } else if (ret == FIO_Q_QUEUED) {
316 if (ddir_rw(io_u->ddir)) {
318 td->ts.total_io_u[io_u->ddir]++;
321 if (td->io_u_queued >= td->o.iodepth_batch) {
322 r = td_io_commit(td);
328 if ((td->io_ops->flags & FIO_SYNCIO) == 0) {
329 if (fio_fill_issue_time(td))
330 fio_gettime(&io_u->issue_time, NULL);
333 * only used for iolog
335 if (td->o.read_iolog_file)
336 memcpy(&td->last_issue, &io_u->issue_time,
337 sizeof(struct timeval));
343 int td_io_init(struct thread_data *td)
347 if (td->io_ops->init) {
348 ret = td->io_ops->init(td);
349 if (ret && td->o.iodepth > 1) {
350 log_err("fio: io engine init failed. Perhaps try"
351 " reducing io depth?\n");
360 int td_io_commit(struct thread_data *td)
364 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
366 if (!td->cur_depth || !td->io_u_queued)
369 io_u_mark_depth(td, td->io_u_queued);
371 if (td->io_ops->commit) {
372 ret = td->io_ops->commit(td);
374 td_verror(td, -ret, "io commit");
378 * Reflect that events were submitted as async IO requests.
380 td->io_u_in_flight += td->io_u_queued;
386 int td_io_open_file(struct thread_data *td, struct fio_file *f)
388 assert(!fio_file_open(f));
391 if (td->io_ops->open_file(td, f)) {
392 if (td->error == EINVAL && td->o.odirect)
393 log_err("fio: destination does not support O_DIRECT\n");
394 if (td->error == EMFILE) {
395 log_err("fio: try reducing/setting openfiles (failed"
396 " at %u of %u)\n", td->nr_open_files,
401 assert(!fio_file_open(f));
405 fio_file_reset(td, f);
406 fio_file_set_open(f);
407 fio_file_clear_closing(f);
408 disk_util_inc(f->du);
413 if (f->filetype == FIO_TYPE_PIPE) {
415 log_err("fio: can't seek on pipes (no random io)\n");
420 if (td->io_ops->flags & FIO_DISKLESSIO)
423 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
426 if (td->o.fadvise_hint &&
427 (f->filetype == FIO_TYPE_BD || f->filetype == FIO_TYPE_FILE)) {
431 flags = POSIX_FADV_RANDOM;
433 flags = POSIX_FADV_SEQUENTIAL;
435 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
436 td_verror(td, errno, "fadvise");
441 #ifdef FIO_OS_DIRECTIO
443 * Some OS's have a distinct call to mark the file non-buffered,
444 * instead of using O_DIRECT (Solaris)
447 int ret = fio_set_odirect(f->fd);
450 td_verror(td, ret, "fio_set_odirect");
451 log_err("fio: the file system does not seem to support direct IO\n");
458 log_file(td, f, FIO_LOG_OPEN_FILE);
461 disk_util_dec(f->du);
462 if (td->io_ops->close_file)
463 td->io_ops->close_file(td, f);
467 int td_io_close_file(struct thread_data *td, struct fio_file *f)
469 if (!fio_file_closing(f))
470 log_file(td, f, FIO_LOG_CLOSE_FILE);
473 * mark as closing, do real close when last io on it has completed
475 fio_file_set_closing(f);
477 disk_util_dec(f->du);
478 unlock_file_all(td, f);
480 return put_file(td, f);
483 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
485 if (!td->io_ops->get_file_size)
488 return td->io_ops->get_file_size(td, f);
491 static int do_sync_file_range(struct thread_data *td, struct fio_file *f)
493 off64_t offset, nbytes;
495 offset = f->first_write;
496 nbytes = f->last_write - f->first_write;
501 return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
504 int do_io_u_sync(struct thread_data *td, struct io_u *io_u)
508 if (io_u->ddir == DDIR_SYNC) {
509 ret = fsync(io_u->file->fd);
510 } else if (io_u->ddir == DDIR_DATASYNC) {
511 #ifdef CONFIG_FDATASYNC
512 ret = fdatasync(io_u->file->fd);
514 ret = io_u->xfer_buflen;
515 io_u->error = EINVAL;
517 } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
518 ret = do_sync_file_range(td, io_u->file);
520 ret = io_u->xfer_buflen;
521 io_u->error = EINVAL;
530 int do_io_u_trim(struct thread_data *td, struct io_u *io_u)
532 #ifndef FIO_HAVE_TRIM
533 io_u->error = EINVAL;
536 struct fio_file *f = io_u->file;
539 ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
541 return io_u->xfer_buflen;;
548 int fio_show_ioengine_help(const char *engine)
550 struct flist_head *entry;
551 struct thread_data td;
555 if (!engine || !*engine) {
556 log_info("Available IO engines:\n");
557 flist_for_each(entry, &engine_list) {
558 td.io_ops = flist_entry(entry, struct ioengine_ops,
560 log_info("\t%s\n", td.io_ops->name);
564 sep = strchr(engine, ',');
570 memset(&td, 0, sizeof(td));
572 td.io_ops = load_ioengine(&td, engine);
574 log_info("IO engine %s not found\n", engine);
578 if (td.io_ops->options)
579 ret = show_cmd_help(td.io_ops->options, sep);
581 log_info("IO engine %s has no options\n", td.io_ops->name);