1 /*
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.
4 *
5 * sync io is implemented on top of aio.
6 *
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.
10 *
11 */
12 #include <stdio.h>
13 #include <stdlib.h>
14 #include <unistd.h>
15 #include <string.h>
16 #include <dlfcn.h>
17 #include <fcntl.h>
18 #include <assert.h>
19
20 #include "fio.h"
21 #include "diskutil.h"
22
23 static FLIST_HEAD(engine_list);
24
check_engine_ops(struct ioengine_ops * ops)25 static bool check_engine_ops(struct ioengine_ops *ops)
26 {
27 if (ops->version != FIO_IOOPS_VERSION) {
28 log_err("bad ioops version %d (want %d)\n", ops->version,
29 FIO_IOOPS_VERSION);
30 return true;
31 }
32
33 if (!ops->queue) {
34 log_err("%s: no queue handler\n", ops->name);
35 return true;
36 }
37
38 /*
39 * sync engines only need a ->queue()
40 */
41 if (ops->flags & FIO_SYNCIO)
42 return false;
43
44 if (!ops->event || !ops->getevents) {
45 log_err("%s: no event/getevents handler\n", ops->name);
46 return true;
47 }
48
49 return false;
50 }
51
unregister_ioengine(struct ioengine_ops * ops)52 void unregister_ioengine(struct ioengine_ops *ops)
53 {
54 dprint(FD_IO, "ioengine %s unregistered\n", ops->name);
55 flist_del(&ops->list);
56 INIT_FLIST_HEAD(&ops->list);
57 }
58
register_ioengine(struct ioengine_ops * ops)59 void register_ioengine(struct ioengine_ops *ops)
60 {
61 dprint(FD_IO, "ioengine %s registered\n", ops->name);
62 INIT_FLIST_HEAD(&ops->list);
63 flist_add_tail(&ops->list, &engine_list);
64 }
65
find_ioengine(const char * name)66 static struct ioengine_ops *find_ioengine(const char *name)
67 {
68 struct ioengine_ops *ops;
69 struct flist_head *entry;
70
71 flist_for_each(entry, &engine_list) {
72 ops = flist_entry(entry, struct ioengine_ops, list);
73 if (!strcmp(name, ops->name))
74 return ops;
75 }
76
77 return NULL;
78 }
79
dlopen_ioengine(struct thread_data * td,const char * engine_lib)80 static struct ioengine_ops *dlopen_ioengine(struct thread_data *td,
81 const char *engine_lib)
82 {
83 struct ioengine_ops *ops;
84 void *dlhandle;
85
86 dprint(FD_IO, "dload engine %s\n", engine_lib);
87
88 dlerror();
89 dlhandle = dlopen(engine_lib, RTLD_LAZY);
90 if (!dlhandle) {
91 td_vmsg(td, -1, dlerror(), "dlopen");
92 return NULL;
93 }
94
95 /*
96 * Unlike the included modules, external engines should have a
97 * non-static ioengine structure that we can reference.
98 */
99 ops = dlsym(dlhandle, engine_lib);
100 if (!ops)
101 ops = dlsym(dlhandle, "ioengine");
102
103 /*
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
107 * structure.
108 */
109 if (!ops) {
110 get_ioengine_t get_ioengine = dlsym(dlhandle, "get_ioengine");
111
112 if (get_ioengine)
113 get_ioengine(&ops);
114 }
115
116 if (!ops) {
117 td_vmsg(td, -1, dlerror(), "dlsym");
118 dlclose(dlhandle);
119 return NULL;
120 }
121
122 td->io_ops_dlhandle = dlhandle;
123 return ops;
124 }
125
load_ioengine(struct thread_data * td,const char * name)126 struct ioengine_ops *load_ioengine(struct thread_data *td, const char *name)
127 {
128 struct ioengine_ops *ops;
129 char engine[64];
130
131 dprint(FD_IO, "load ioengine %s\n", name);
132
133 engine[sizeof(engine) - 1] = '\0';
134 strncpy(engine, name, sizeof(engine) - 1);
135
136 /*
137 * linux libaio has alias names, so convert to what we want
138 */
139 if (!strncmp(engine, "linuxaio", 8) || !strncmp(engine, "aio", 3))
140 strcpy(engine, "libaio");
141
142 ops = find_ioengine(engine);
143 if (!ops)
144 ops = dlopen_ioengine(td, name);
145
146 if (!ops) {
147 log_err("fio: engine %s not loadable\n", name);
148 return NULL;
149 }
150
151 /*
152 * Check that the required methods are there.
153 */
154 if (check_engine_ops(ops))
155 return NULL;
156
157 return ops;
158 }
159
160 /*
161 * For cleaning up an ioengine which never made it to init().
162 */
free_ioengine(struct thread_data * td)163 void free_ioengine(struct thread_data *td)
164 {
165 dprint(FD_IO, "free ioengine %s\n", td->io_ops->name);
166
167 if (td->eo && td->io_ops->options) {
168 options_free(td->io_ops->options, td->eo);
169 free(td->eo);
170 td->eo = NULL;
171 }
172
173 if (td->io_ops_dlhandle)
174 dlclose(td->io_ops_dlhandle);
175
176 td->io_ops = NULL;
177 }
178
close_ioengine(struct thread_data * td)179 void close_ioengine(struct thread_data *td)
180 {
181 dprint(FD_IO, "close ioengine %s\n", td->io_ops->name);
182
183 if (td->io_ops->cleanup) {
184 td->io_ops->cleanup(td);
185 td->io_ops_data = NULL;
186 }
187
188 free_ioengine(td);
189 }
190
td_io_prep(struct thread_data * td,struct io_u * io_u)191 int td_io_prep(struct thread_data *td, struct io_u *io_u)
192 {
193 dprint_io_u(io_u, "prep");
194 fio_ro_check(td, io_u);
195
196 lock_file(td, io_u->file, io_u->ddir);
197
198 if (td->io_ops->prep) {
199 int ret = td->io_ops->prep(td, io_u);
200
201 dprint(FD_IO, "->prep(%p)=%d\n", io_u, ret);
202 if (ret)
203 unlock_file(td, io_u->file);
204 return ret;
205 }
206
207 return 0;
208 }
209
td_io_getevents(struct thread_data * td,unsigned int min,unsigned int max,const struct timespec * t)210 int td_io_getevents(struct thread_data *td, unsigned int min, unsigned int max,
211 const struct timespec *t)
212 {
213 int r = 0;
214
215 /*
216 * For ioengine=rdma one side operation RDMA_WRITE or RDMA_READ,
217 * server side gets a message from the client
218 * side that the task is finished, and
219 * td->done is set to 1 after td_io_commit(). In this case,
220 * there is no need to reap complete event in server side.
221 */
222 if (td->done)
223 return 0;
224
225 if (min > 0 && td->io_ops->commit) {
226 r = td->io_ops->commit(td);
227 if (r < 0)
228 goto out;
229 }
230 if (max > td->cur_depth)
231 max = td->cur_depth;
232 if (min > max)
233 max = min;
234
235 r = 0;
236 if (max && td->io_ops->getevents)
237 r = td->io_ops->getevents(td, min, max, t);
238 out:
239 if (r >= 0) {
240 /*
241 * Reflect that our submitted requests were retrieved with
242 * whatever OS async calls are in the underlying engine.
243 */
244 td->io_u_in_flight -= r;
245 io_u_mark_complete(td, r);
246 } else
247 td_verror(td, r, "get_events");
248
249 dprint(FD_IO, "getevents: %d\n", r);
250 return r;
251 }
252
td_io_queue(struct thread_data * td,struct io_u * io_u)253 int td_io_queue(struct thread_data *td, struct io_u *io_u)
254 {
255 const enum fio_ddir ddir = acct_ddir(io_u);
256 unsigned long buflen = io_u->xfer_buflen;
257 int ret;
258
259 dprint_io_u(io_u, "queue");
260 fio_ro_check(td, io_u);
261
262 assert((io_u->flags & IO_U_F_FLIGHT) == 0);
263 io_u_set(td, io_u, IO_U_F_FLIGHT);
264
265 assert(fio_file_open(io_u->file));
266
267 /*
268 * If using a write iolog, store this entry.
269 */
270 log_io_u(td, io_u);
271
272 io_u->error = 0;
273 io_u->resid = 0;
274
275 if (td_ioengine_flagged(td, FIO_SYNCIO)) {
276 if (fio_fill_issue_time(td))
277 fio_gettime(&io_u->issue_time, NULL);
278
279 /*
280 * only used for iolog
281 */
282 if (td->o.read_iolog_file)
283 memcpy(&td->last_issue, &io_u->issue_time,
284 sizeof(struct timeval));
285 }
286
287 if (ddir_rw(ddir)) {
288 td->io_issues[ddir]++;
289 td->io_issue_bytes[ddir] += buflen;
290 td->rate_io_issue_bytes[ddir] += buflen;
291 }
292
293 ret = td->io_ops->queue(td, io_u);
294
295 unlock_file(td, io_u->file);
296
297 if (ret == FIO_Q_BUSY && ddir_rw(ddir)) {
298 td->io_issues[ddir]--;
299 td->io_issue_bytes[ddir] -= buflen;
300 td->rate_io_issue_bytes[ddir] -= buflen;
301 io_u_clear(td, io_u, IO_U_F_FLIGHT);
302 }
303
304 /*
305 * If an error was seen and the io engine didn't propagate it
306 * back to 'td', do so.
307 */
308 if (io_u->error && !td->error)
309 td_verror(td, io_u->error, "td_io_queue");
310
311 /*
312 * Add warning for O_DIRECT so that users have an easier time
313 * spotting potentially bad alignment. If this triggers for the first
314 * IO, then it's likely an alignment problem or because the host fs
315 * does not support O_DIRECT
316 */
317 if (io_u->error == EINVAL && td->io_issues[io_u->ddir & 1] == 1 &&
318 td->o.odirect) {
319
320 log_info("fio: first direct IO errored. File system may not "
321 "support direct IO, or iomem_align= is bad. Try "
322 "setting direct=0.\n");
323 }
324
325 if (!td->io_ops->commit || io_u->ddir == DDIR_TRIM) {
326 io_u_mark_submit(td, 1);
327 io_u_mark_complete(td, 1);
328 }
329
330 if (ret == FIO_Q_COMPLETED) {
331 if (ddir_rw(io_u->ddir)) {
332 io_u_mark_depth(td, 1);
333 td->ts.total_io_u[io_u->ddir]++;
334 }
335 } else if (ret == FIO_Q_QUEUED) {
336 int r;
337
338 td->io_u_queued++;
339
340 if (ddir_rw(io_u->ddir))
341 td->ts.total_io_u[io_u->ddir]++;
342
343 if (td->io_u_queued >= td->o.iodepth_batch) {
344 r = td_io_commit(td);
345 if (r < 0)
346 return r;
347 }
348 }
349
350 if (!td_ioengine_flagged(td, FIO_SYNCIO)) {
351 if (fio_fill_issue_time(td))
352 fio_gettime(&io_u->issue_time, NULL);
353
354 /*
355 * only used for iolog
356 */
357 if (td->o.read_iolog_file)
358 memcpy(&td->last_issue, &io_u->issue_time,
359 sizeof(struct timeval));
360 }
361
362 return ret;
363 }
364
td_io_init(struct thread_data * td)365 int td_io_init(struct thread_data *td)
366 {
367 int ret = 0;
368
369 if (td->io_ops->init) {
370 ret = td->io_ops->init(td);
371 if (ret)
372 log_err("fio: io engine %s init failed.%s\n",
373 td->io_ops->name,
374 td->o.iodepth > 1 ?
375 " Perhaps try reducing io depth?" : "");
376 else
377 td->io_ops_init = 1;
378 if (!td->error)
379 td->error = ret;
380 }
381
382 return ret;
383 }
384
td_io_commit(struct thread_data * td)385 int td_io_commit(struct thread_data *td)
386 {
387 int ret;
388
389 dprint(FD_IO, "calling ->commit(), depth %d\n", td->cur_depth);
390
391 if (!td->cur_depth || !td->io_u_queued)
392 return 0;
393
394 io_u_mark_depth(td, td->io_u_queued);
395
396 if (td->io_ops->commit) {
397 ret = td->io_ops->commit(td);
398 if (ret)
399 td_verror(td, -ret, "io commit");
400 }
401
402 /*
403 * Reflect that events were submitted as async IO requests.
404 */
405 td->io_u_in_flight += td->io_u_queued;
406 td->io_u_queued = 0;
407
408 return 0;
409 }
410
td_io_open_file(struct thread_data * td,struct fio_file * f)411 int td_io_open_file(struct thread_data *td, struct fio_file *f)
412 {
413 assert(!fio_file_open(f));
414 assert(f->fd == -1);
415
416 if (td->io_ops->open_file(td, f)) {
417 if (td->error == EINVAL && td->o.odirect)
418 log_err("fio: destination does not support O_DIRECT\n");
419 if (td->error == EMFILE) {
420 log_err("fio: try reducing/setting openfiles (failed"
421 " at %u of %u)\n", td->nr_open_files,
422 td->o.nr_files);
423 }
424
425 assert(f->fd == -1);
426 assert(!fio_file_open(f));
427 return 1;
428 }
429
430 fio_file_reset(td, f);
431 fio_file_set_open(f);
432 fio_file_clear_closing(f);
433 disk_util_inc(f->du);
434
435 td->nr_open_files++;
436 get_file(f);
437
438 if (f->filetype == FIO_TYPE_PIPE) {
439 if (td_random(td)) {
440 log_err("fio: can't seek on pipes (no random io)\n");
441 goto err;
442 }
443 }
444
445 if (td_ioengine_flagged(td, FIO_DISKLESSIO))
446 goto done;
447
448 if (td->o.invalidate_cache && file_invalidate_cache(td, f))
449 goto err;
450
451 if (td->o.fadvise_hint != F_ADV_NONE &&
452 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
453 int flags;
454
455 if (td->o.fadvise_hint == F_ADV_TYPE) {
456 if (td_random(td))
457 flags = POSIX_FADV_RANDOM;
458 else
459 flags = POSIX_FADV_SEQUENTIAL;
460 } else if (td->o.fadvise_hint == F_ADV_RANDOM)
461 flags = POSIX_FADV_RANDOM;
462 else if (td->o.fadvise_hint == F_ADV_SEQUENTIAL)
463 flags = POSIX_FADV_SEQUENTIAL;
464 else {
465 log_err("fio: unknown fadvise type %d\n",
466 td->o.fadvise_hint);
467 flags = POSIX_FADV_NORMAL;
468 }
469
470 if (posix_fadvise(f->fd, f->file_offset, f->io_size, flags) < 0) {
471 td_verror(td, errno, "fadvise");
472 goto err;
473 }
474 }
475 #ifdef FIO_HAVE_STREAMID
476 if (td->o.fadvise_stream &&
477 (f->filetype == FIO_TYPE_BLOCK || f->filetype == FIO_TYPE_FILE)) {
478 off_t stream = td->o.fadvise_stream;
479
480 if (posix_fadvise(f->fd, stream, f->io_size, POSIX_FADV_STREAMID) < 0) {
481 td_verror(td, errno, "fadvise streamid");
482 goto err;
483 }
484 }
485 #endif
486
487 #ifdef FIO_OS_DIRECTIO
488 /*
489 * Some OS's have a distinct call to mark the file non-buffered,
490 * instead of using O_DIRECT (Solaris)
491 */
492 if (td->o.odirect) {
493 int ret = fio_set_odirect(f->fd);
494
495 if (ret) {
496 td_verror(td, ret, "fio_set_odirect");
497 if (ret == ENOTTY) { /* ENOTTY suggests RAW device or ZFS */
498 log_err("fio: doing directIO to RAW devices or ZFS not supported\n");
499 } else {
500 log_err("fio: the file system does not seem to support direct IO\n");
501 }
502
503 goto err;
504 }
505 }
506 #endif
507
508 done:
509 log_file(td, f, FIO_LOG_OPEN_FILE);
510 return 0;
511 err:
512 disk_util_dec(f->du);
513 if (td->io_ops->close_file)
514 td->io_ops->close_file(td, f);
515 return 1;
516 }
517
td_io_close_file(struct thread_data * td,struct fio_file * f)518 int td_io_close_file(struct thread_data *td, struct fio_file *f)
519 {
520 if (!fio_file_closing(f))
521 log_file(td, f, FIO_LOG_CLOSE_FILE);
522
523 /*
524 * mark as closing, do real close when last io on it has completed
525 */
526 fio_file_set_closing(f);
527
528 disk_util_dec(f->du);
529
530 if (td->o.file_lock_mode != FILE_LOCK_NONE)
531 unlock_file_all(td, f);
532
533 return put_file(td, f);
534 }
535
td_io_unlink_file(struct thread_data * td,struct fio_file * f)536 int td_io_unlink_file(struct thread_data *td, struct fio_file *f)
537 {
538 if (td->io_ops->unlink_file)
539 return td->io_ops->unlink_file(td, f);
540 else {
541 int ret;
542
543 ret = unlink(f->file_name);
544 if (ret < 0)
545 return errno;
546
547 return 0;
548 }
549 }
550
td_io_get_file_size(struct thread_data * td,struct fio_file * f)551 int td_io_get_file_size(struct thread_data *td, struct fio_file *f)
552 {
553 if (!td->io_ops->get_file_size)
554 return 0;
555
556 return td->io_ops->get_file_size(td, f);
557 }
558
fio_show_ioengine_help(const char * engine)559 int fio_show_ioengine_help(const char *engine)
560 {
561 struct flist_head *entry;
562 struct thread_data td;
563 struct ioengine_ops *io_ops;
564 char *sep;
565 int ret = 1;
566
567 if (!engine || !*engine) {
568 log_info("Available IO engines:\n");
569 flist_for_each(entry, &engine_list) {
570 io_ops = flist_entry(entry, struct ioengine_ops, list);
571 log_info("\t%s\n", io_ops->name);
572 }
573 return 0;
574 }
575 sep = strchr(engine, ',');
576 if (sep) {
577 *sep = 0;
578 sep++;
579 }
580
581 memset(&td, 0, sizeof(td));
582
583 io_ops = load_ioengine(&td, engine);
584 if (!io_ops) {
585 log_info("IO engine %s not found\n", engine);
586 return 1;
587 }
588
589 if (io_ops->options)
590 ret = show_cmd_help(io_ops->options, sep);
591 else
592 log_info("IO engine %s has no options\n", io_ops->name);
593
594 free_ioengine(&td);
595
596 return ret;
597 }
598