Commit | Line | Data |
---|---|---|
2e1df07d JA |
1 | /* |
2 | * fio - the flexible io tester | |
3 | * | |
4 | * Copyright (C) 2005 Jens Axboe <axboe@suse.de> | |
5 | * Copyright (C) 2006-2012 Jens Axboe <axboe@kernel.dk> | |
6 | * | |
7 | * The license below covers all files distributed with fio unless otherwise | |
8 | * noted in the file itself. | |
9 | * | |
10 | * This program is free software; you can redistribute it and/or modify | |
11 | * it under the terms of the GNU General Public License version 2 as | |
12 | * published by the Free Software Foundation. | |
13 | * | |
14 | * This program is distributed in the hope that it will be useful, | |
15 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
17 | * GNU General Public License for more details. | |
18 | * | |
19 | * You should have received a copy of the GNU General Public License | |
20 | * along with this program; if not, write to the Free Software | |
21 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
22 | * | |
23 | */ | |
24 | #include <unistd.h> | |
25 | #include <fcntl.h> | |
26 | #include <string.h> | |
27 | #include <limits.h> | |
28 | #include <signal.h> | |
29 | #include <time.h> | |
30 | #include <locale.h> | |
31 | #include <assert.h> | |
32 | #include <time.h> | |
e43606c2 | 33 | #include <inttypes.h> |
2e1df07d JA |
34 | #include <sys/stat.h> |
35 | #include <sys/wait.h> | |
36 | #include <sys/ipc.h> | |
37 | #include <sys/shm.h> | |
38 | #include <sys/mman.h> | |
39 | ||
40 | #include "fio.h" | |
41 | #include "hash.h" | |
42 | #include "smalloc.h" | |
43 | #include "verify.h" | |
44 | #include "trim.h" | |
45 | #include "diskutil.h" | |
46 | #include "cgroup.h" | |
47 | #include "profile.h" | |
48 | #include "lib/rand.h" | |
49 | #include "memalign.h" | |
50 | #include "server.h" | |
51 | ||
52 | static pthread_t disk_util_thread; | |
feb41855 | 53 | static struct fio_mutex *disk_thread_mutex; |
2e1df07d JA |
54 | static struct fio_mutex *startup_mutex; |
55 | static struct fio_mutex *writeout_mutex; | |
56 | static struct flist_head *cgroup_list; | |
57 | static char *cgroup_mnt; | |
58 | static int exit_value; | |
59 | static volatile int fio_abort; | |
812409e9 JA |
60 | static unsigned int nr_process = 0; |
61 | static unsigned int nr_thread = 0; | |
2e1df07d | 62 | |
6eaf09d6 | 63 | struct io_log *agg_io_log[DDIR_RWDIR_CNT]; |
2e1df07d | 64 | |
a3efc919 JA |
65 | int groupid = 0; |
66 | unsigned int thread_number = 0; | |
a3efc919 JA |
67 | int shm_id = 0; |
68 | int temp_stall_ts; | |
69 | unsigned long done_secs = 0; | |
27357187 | 70 | volatile int disk_util_exit = 0; |
a3efc919 | 71 | |
2e1df07d | 72 | #define PAGE_ALIGN(buf) \ |
e43606c2 | 73 | (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask) |
2e1df07d JA |
74 | |
75 | #define JOB_START_TIMEOUT (5 * 1000) | |
76 | ||
77 | static void sig_int(int sig) | |
78 | { | |
79 | if (threads) { | |
80 | if (is_backend) | |
81 | fio_server_got_signal(sig); | |
82 | else { | |
83 | log_info("\nfio: terminating on signal %d\n", sig); | |
84 | fflush(stdout); | |
85 | exit_value = 128; | |
86 | } | |
87 | ||
88 | fio_terminate_threads(TERMINATE_ALL); | |
89 | } | |
90 | } | |
91 | ||
b852e7cf JA |
92 | static void sig_show_status(int sig) |
93 | { | |
94 | show_running_run_stats(); | |
95 | } | |
96 | ||
2e1df07d JA |
97 | static void set_sig_handlers(void) |
98 | { | |
99 | struct sigaction act; | |
100 | ||
101 | memset(&act, 0, sizeof(act)); | |
102 | act.sa_handler = sig_int; | |
103 | act.sa_flags = SA_RESTART; | |
104 | sigaction(SIGINT, &act, NULL); | |
105 | ||
106 | memset(&act, 0, sizeof(act)); | |
107 | act.sa_handler = sig_int; | |
108 | act.sa_flags = SA_RESTART; | |
109 | sigaction(SIGTERM, &act, NULL); | |
110 | ||
b852e7cf JA |
111 | memset(&act, 0, sizeof(act)); |
112 | act.sa_handler = sig_show_status; | |
113 | act.sa_flags = SA_RESTART; | |
114 | sigaction(SIGUSR1, &act, NULL); | |
115 | ||
2e1df07d JA |
116 | if (is_backend) { |
117 | memset(&act, 0, sizeof(act)); | |
118 | act.sa_handler = sig_int; | |
119 | act.sa_flags = SA_RESTART; | |
120 | sigaction(SIGPIPE, &act, NULL); | |
121 | } | |
122 | } | |
123 | ||
124 | /* | |
125 | * Check if we are above the minimum rate given. | |
126 | */ | |
127 | static int __check_min_rate(struct thread_data *td, struct timeval *now, | |
128 | enum fio_ddir ddir) | |
129 | { | |
130 | unsigned long long bytes = 0; | |
131 | unsigned long iops = 0; | |
132 | unsigned long spent; | |
133 | unsigned long rate; | |
134 | unsigned int ratemin = 0; | |
135 | unsigned int rate_iops = 0; | |
136 | unsigned int rate_iops_min = 0; | |
137 | ||
138 | assert(ddir_rw(ddir)); | |
139 | ||
140 | if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir]) | |
141 | return 0; | |
142 | ||
143 | /* | |
144 | * allow a 2 second settle period in the beginning | |
145 | */ | |
146 | if (mtime_since(&td->start, now) < 2000) | |
147 | return 0; | |
148 | ||
149 | iops += td->this_io_blocks[ddir]; | |
150 | bytes += td->this_io_bytes[ddir]; | |
151 | ratemin += td->o.ratemin[ddir]; | |
152 | rate_iops += td->o.rate_iops[ddir]; | |
153 | rate_iops_min += td->o.rate_iops_min[ddir]; | |
154 | ||
155 | /* | |
156 | * if rate blocks is set, sample is running | |
157 | */ | |
158 | if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) { | |
159 | spent = mtime_since(&td->lastrate[ddir], now); | |
160 | if (spent < td->o.ratecycle) | |
161 | return 0; | |
162 | ||
163 | if (td->o.rate[ddir]) { | |
164 | /* | |
165 | * check bandwidth specified rate | |
166 | */ | |
167 | if (bytes < td->rate_bytes[ddir]) { | |
168 | log_err("%s: min rate %u not met\n", td->o.name, | |
169 | ratemin); | |
170 | return 1; | |
171 | } else { | |
172 | rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent; | |
173 | if (rate < ratemin || | |
174 | bytes < td->rate_bytes[ddir]) { | |
175 | log_err("%s: min rate %u not met, got" | |
176 | " %luKB/sec\n", td->o.name, | |
177 | ratemin, rate); | |
178 | return 1; | |
179 | } | |
180 | } | |
181 | } else { | |
182 | /* | |
183 | * checks iops specified rate | |
184 | */ | |
185 | if (iops < rate_iops) { | |
186 | log_err("%s: min iops rate %u not met\n", | |
187 | td->o.name, rate_iops); | |
188 | return 1; | |
189 | } else { | |
190 | rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent; | |
191 | if (rate < rate_iops_min || | |
192 | iops < td->rate_blocks[ddir]) { | |
193 | log_err("%s: min iops rate %u not met," | |
194 | " got %lu\n", td->o.name, | |
195 | rate_iops_min, rate); | |
196 | } | |
197 | } | |
198 | } | |
199 | } | |
200 | ||
201 | td->rate_bytes[ddir] = bytes; | |
202 | td->rate_blocks[ddir] = iops; | |
203 | memcpy(&td->lastrate[ddir], now, sizeof(*now)); | |
204 | return 0; | |
205 | } | |
206 | ||
207 | static int check_min_rate(struct thread_data *td, struct timeval *now, | |
208 | unsigned long *bytes_done) | |
209 | { | |
210 | int ret = 0; | |
211 | ||
6eaf09d6 SL |
212 | if (bytes_done[DDIR_READ]) |
213 | ret |= __check_min_rate(td, now, DDIR_READ); | |
214 | if (bytes_done[DDIR_WRITE]) | |
215 | ret |= __check_min_rate(td, now, DDIR_WRITE); | |
216 | if (bytes_done[DDIR_TRIM]) | |
217 | ret |= __check_min_rate(td, now, DDIR_TRIM); | |
2e1df07d JA |
218 | |
219 | return ret; | |
220 | } | |
221 | ||
222 | /* | |
223 | * When job exits, we can cancel the in-flight IO if we are using async | |
224 | * io. Attempt to do so. | |
225 | */ | |
226 | static void cleanup_pending_aio(struct thread_data *td) | |
227 | { | |
228 | struct flist_head *entry, *n; | |
229 | struct io_u *io_u; | |
230 | int r; | |
231 | ||
232 | /* | |
233 | * get immediately available events, if any | |
234 | */ | |
235 | r = io_u_queued_complete(td, 0, NULL); | |
236 | if (r < 0) | |
237 | return; | |
238 | ||
239 | /* | |
240 | * now cancel remaining active events | |
241 | */ | |
242 | if (td->io_ops->cancel) { | |
243 | flist_for_each_safe(entry, n, &td->io_u_busylist) { | |
244 | io_u = flist_entry(entry, struct io_u, list); | |
245 | ||
246 | /* | |
247 | * if the io_u isn't in flight, then that generally | |
248 | * means someone leaked an io_u. complain but fix | |
249 | * it up, so we don't stall here. | |
250 | */ | |
251 | if ((io_u->flags & IO_U_F_FLIGHT) == 0) { | |
252 | log_err("fio: non-busy IO on busy list\n"); | |
253 | put_io_u(td, io_u); | |
254 | } else { | |
255 | r = td->io_ops->cancel(td, io_u); | |
256 | if (!r) | |
257 | put_io_u(td, io_u); | |
258 | } | |
259 | } | |
260 | } | |
261 | ||
262 | if (td->cur_depth) | |
263 | r = io_u_queued_complete(td, td->cur_depth, NULL); | |
264 | } | |
265 | ||
266 | /* | |
267 | * Helper to handle the final sync of a file. Works just like the normal | |
268 | * io path, just does everything sync. | |
269 | */ | |
270 | static int fio_io_sync(struct thread_data *td, struct fio_file *f) | |
271 | { | |
272 | struct io_u *io_u = __get_io_u(td); | |
273 | int ret; | |
274 | ||
275 | if (!io_u) | |
276 | return 1; | |
277 | ||
278 | io_u->ddir = DDIR_SYNC; | |
279 | io_u->file = f; | |
280 | ||
281 | if (td_io_prep(td, io_u)) { | |
282 | put_io_u(td, io_u); | |
283 | return 1; | |
284 | } | |
285 | ||
286 | requeue: | |
287 | ret = td_io_queue(td, io_u); | |
288 | if (ret < 0) { | |
289 | td_verror(td, io_u->error, "td_io_queue"); | |
290 | put_io_u(td, io_u); | |
291 | return 1; | |
292 | } else if (ret == FIO_Q_QUEUED) { | |
293 | if (io_u_queued_complete(td, 1, NULL) < 0) | |
294 | return 1; | |
295 | } else if (ret == FIO_Q_COMPLETED) { | |
296 | if (io_u->error) { | |
297 | td_verror(td, io_u->error, "td_io_queue"); | |
298 | return 1; | |
299 | } | |
300 | ||
301 | if (io_u_sync_complete(td, io_u, NULL) < 0) | |
302 | return 1; | |
303 | } else if (ret == FIO_Q_BUSY) { | |
304 | if (td_io_commit(td)) | |
305 | return 1; | |
306 | goto requeue; | |
307 | } | |
308 | ||
309 | return 0; | |
310 | } | |
a3efc919 | 311 | |
2e1df07d JA |
312 | static inline void __update_tv_cache(struct thread_data *td) |
313 | { | |
314 | fio_gettime(&td->tv_cache, NULL); | |
315 | } | |
316 | ||
317 | static inline void update_tv_cache(struct thread_data *td) | |
318 | { | |
319 | if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask) | |
320 | __update_tv_cache(td); | |
321 | } | |
322 | ||
323 | static inline int runtime_exceeded(struct thread_data *td, struct timeval *t) | |
324 | { | |
325 | if (in_ramp_time(td)) | |
326 | return 0; | |
327 | if (!td->o.timeout) | |
328 | return 0; | |
329 | if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000) | |
330 | return 1; | |
331 | ||
332 | return 0; | |
333 | } | |
334 | ||
335 | static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir, | |
336 | int *retptr) | |
337 | { | |
338 | int ret = *retptr; | |
339 | ||
340 | if (ret < 0 || td->error) { | |
8b28bd41 DM |
341 | int err = td->error; |
342 | enum error_type_bit eb; | |
2e1df07d JA |
343 | |
344 | if (ret < 0) | |
345 | err = -ret; | |
2e1df07d | 346 | |
8b28bd41 DM |
347 | eb = td_error_type(ddir, err); |
348 | if (!(td->o.continue_on_error & (1 << eb))) | |
2e1df07d JA |
349 | return 1; |
350 | ||
8b28bd41 | 351 | if (td_non_fatal_error(td, eb, err)) { |
2e1df07d JA |
352 | /* |
353 | * Continue with the I/Os in case of | |
354 | * a non fatal error. | |
355 | */ | |
356 | update_error_count(td, err); | |
357 | td_clear_error(td); | |
358 | *retptr = 0; | |
359 | return 0; | |
360 | } else if (td->o.fill_device && err == ENOSPC) { | |
361 | /* | |
362 | * We expect to hit this error if | |
363 | * fill_device option is set. | |
364 | */ | |
365 | td_clear_error(td); | |
366 | td->terminate = 1; | |
367 | return 1; | |
368 | } else { | |
369 | /* | |
370 | * Stop the I/O in case of a fatal | |
371 | * error. | |
372 | */ | |
373 | update_error_count(td, err); | |
374 | return 1; | |
375 | } | |
376 | } | |
377 | ||
378 | return 0; | |
379 | } | |
380 | ||
2e1df07d JA |
381 | /* |
382 | * The main verify engine. Runs over the writes we previously submitted, | |
383 | * reads the blocks back in, and checks the crc/md5 of the data. | |
384 | */ | |
385 | static void do_verify(struct thread_data *td) | |
386 | { | |
387 | struct fio_file *f; | |
388 | struct io_u *io_u; | |
389 | int ret, min_events; | |
390 | unsigned int i; | |
391 | ||
392 | dprint(FD_VERIFY, "starting loop\n"); | |
393 | ||
394 | /* | |
395 | * sync io first and invalidate cache, to make sure we really | |
396 | * read from disk. | |
397 | */ | |
398 | for_each_file(td, f, i) { | |
399 | if (!fio_file_open(f)) | |
400 | continue; | |
401 | if (fio_io_sync(td, f)) | |
402 | break; | |
403 | if (file_invalidate_cache(td, f)) | |
404 | break; | |
405 | } | |
406 | ||
407 | if (td->error) | |
408 | return; | |
409 | ||
410 | td_set_runstate(td, TD_VERIFYING); | |
411 | ||
412 | io_u = NULL; | |
413 | while (!td->terminate) { | |
414 | int ret2, full; | |
415 | ||
416 | update_tv_cache(td); | |
417 | ||
418 | if (runtime_exceeded(td, &td->tv_cache)) { | |
419 | __update_tv_cache(td); | |
420 | if (runtime_exceeded(td, &td->tv_cache)) { | |
421 | td->terminate = 1; | |
422 | break; | |
423 | } | |
424 | } | |
425 | ||
9e684a49 DE |
426 | if (flow_threshold_exceeded(td)) |
427 | continue; | |
428 | ||
2e1df07d JA |
429 | io_u = __get_io_u(td); |
430 | if (!io_u) | |
431 | break; | |
432 | ||
433 | if (get_next_verify(td, io_u)) { | |
434 | put_io_u(td, io_u); | |
435 | break; | |
436 | } | |
437 | ||
438 | if (td_io_prep(td, io_u)) { | |
439 | put_io_u(td, io_u); | |
440 | break; | |
441 | } | |
442 | ||
443 | if (td->o.verify_async) | |
444 | io_u->end_io = verify_io_u_async; | |
445 | else | |
446 | io_u->end_io = verify_io_u; | |
447 | ||
448 | ret = td_io_queue(td, io_u); | |
449 | switch (ret) { | |
450 | case FIO_Q_COMPLETED: | |
451 | if (io_u->error) { | |
452 | ret = -io_u->error; | |
453 | clear_io_u(td, io_u); | |
454 | } else if (io_u->resid) { | |
455 | int bytes = io_u->xfer_buflen - io_u->resid; | |
456 | ||
457 | /* | |
458 | * zero read, fail | |
459 | */ | |
460 | if (!bytes) { | |
461 | td_verror(td, EIO, "full resid"); | |
462 | put_io_u(td, io_u); | |
463 | break; | |
464 | } | |
465 | ||
466 | io_u->xfer_buflen = io_u->resid; | |
467 | io_u->xfer_buf += bytes; | |
468 | io_u->offset += bytes; | |
469 | ||
470 | if (ddir_rw(io_u->ddir)) | |
471 | td->ts.short_io_u[io_u->ddir]++; | |
472 | ||
473 | f = io_u->file; | |
474 | if (io_u->offset == f->real_file_size) | |
475 | goto sync_done; | |
476 | ||
477 | requeue_io_u(td, &io_u); | |
478 | } else { | |
479 | sync_done: | |
480 | ret = io_u_sync_complete(td, io_u, NULL); | |
481 | if (ret < 0) | |
482 | break; | |
483 | } | |
484 | continue; | |
485 | case FIO_Q_QUEUED: | |
486 | break; | |
487 | case FIO_Q_BUSY: | |
488 | requeue_io_u(td, &io_u); | |
489 | ret2 = td_io_commit(td); | |
490 | if (ret2 < 0) | |
491 | ret = ret2; | |
492 | break; | |
493 | default: | |
494 | assert(ret < 0); | |
495 | td_verror(td, -ret, "td_io_queue"); | |
496 | break; | |
497 | } | |
498 | ||
499 | if (break_on_this_error(td, io_u->ddir, &ret)) | |
500 | break; | |
501 | ||
502 | /* | |
503 | * if we can queue more, do so. but check if there are | |
504 | * completed io_u's first. Note that we can get BUSY even | |
505 | * without IO queued, if the system is resource starved. | |
506 | */ | |
507 | full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth); | |
508 | if (full || !td->o.iodepth_batch_complete) { | |
509 | min_events = min(td->o.iodepth_batch_complete, | |
510 | td->cur_depth); | |
8a74b56d JA |
511 | /* |
512 | * if the queue is full, we MUST reap at least 1 event | |
513 | */ | |
514 | if (full && !min_events) | |
2e1df07d JA |
515 | min_events = 1; |
516 | ||
517 | do { | |
518 | /* | |
519 | * Reap required number of io units, if any, | |
520 | * and do the verification on them through | |
521 | * the callback handler | |
522 | */ | |
523 | if (io_u_queued_complete(td, min_events, NULL) < 0) { | |
524 | ret = -1; | |
525 | break; | |
526 | } | |
527 | } while (full && (td->cur_depth > td->o.iodepth_low)); | |
528 | } | |
529 | if (ret < 0) | |
530 | break; | |
531 | } | |
532 | ||
533 | if (!td->error) { | |
534 | min_events = td->cur_depth; | |
535 | ||
536 | if (min_events) | |
537 | ret = io_u_queued_complete(td, min_events, NULL); | |
538 | } else | |
539 | cleanup_pending_aio(td); | |
540 | ||
541 | td_set_runstate(td, TD_RUNNING); | |
542 | ||
543 | dprint(FD_VERIFY, "exiting loop\n"); | |
544 | } | |
545 | ||
f7078f7b JA |
546 | static int io_bytes_exceeded(struct thread_data *td) |
547 | { | |
548 | unsigned long long bytes; | |
549 | ||
550 | if (td_rw(td)) | |
6eaf09d6 | 551 | bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE]; |
f7078f7b | 552 | else if (td_write(td)) |
6eaf09d6 SL |
553 | bytes = td->this_io_bytes[DDIR_WRITE]; |
554 | else if (td_read(td)) | |
555 | bytes = td->this_io_bytes[DDIR_READ]; | |
f7078f7b | 556 | else |
6eaf09d6 | 557 | bytes = td->this_io_bytes[DDIR_TRIM]; |
f7078f7b JA |
558 | |
559 | return bytes >= td->o.size; | |
560 | } | |
561 | ||
2e1df07d JA |
562 | /* |
563 | * Main IO worker function. It retrieves io_u's to process and queues | |
564 | * and reaps them, checking for rate and errors along the way. | |
565 | */ | |
566 | static void do_io(struct thread_data *td) | |
567 | { | |
568 | unsigned int i; | |
569 | int ret = 0; | |
570 | ||
571 | if (in_ramp_time(td)) | |
572 | td_set_runstate(td, TD_RAMP); | |
573 | else | |
574 | td_set_runstate(td, TD_RUNNING); | |
575 | ||
f7078f7b | 576 | while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || |
c04e4661 DE |
577 | (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) || |
578 | td->o.time_based) { | |
2e1df07d | 579 | struct timeval comp_time; |
6eaf09d6 | 580 | unsigned long bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 }; |
2e1df07d JA |
581 | int min_evts = 0; |
582 | struct io_u *io_u; | |
583 | int ret2, full; | |
584 | enum fio_ddir ddir; | |
585 | ||
586 | if (td->terminate) | |
587 | break; | |
588 | ||
589 | update_tv_cache(td); | |
590 | ||
591 | if (runtime_exceeded(td, &td->tv_cache)) { | |
592 | __update_tv_cache(td); | |
593 | if (runtime_exceeded(td, &td->tv_cache)) { | |
594 | td->terminate = 1; | |
595 | break; | |
596 | } | |
597 | } | |
598 | ||
9e684a49 DE |
599 | if (flow_threshold_exceeded(td)) |
600 | continue; | |
601 | ||
2e1df07d JA |
602 | io_u = get_io_u(td); |
603 | if (!io_u) | |
604 | break; | |
605 | ||
606 | ddir = io_u->ddir; | |
607 | ||
608 | /* | |
82af2a7c JA |
609 | * Add verification end_io handler if: |
610 | * - Asked to verify (!td_rw(td)) | |
611 | * - Or the io_u is from our verify list (mixed write/ver) | |
2e1df07d JA |
612 | */ |
613 | if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ && | |
82af2a7c | 614 | ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) { |
2e1df07d JA |
615 | if (td->o.verify_async) |
616 | io_u->end_io = verify_io_u_async; | |
617 | else | |
618 | io_u->end_io = verify_io_u; | |
619 | td_set_runstate(td, TD_VERIFYING); | |
620 | } else if (in_ramp_time(td)) | |
621 | td_set_runstate(td, TD_RAMP); | |
622 | else | |
623 | td_set_runstate(td, TD_RUNNING); | |
624 | ||
625 | ret = td_io_queue(td, io_u); | |
626 | switch (ret) { | |
627 | case FIO_Q_COMPLETED: | |
628 | if (io_u->error) { | |
629 | ret = -io_u->error; | |
630 | clear_io_u(td, io_u); | |
631 | } else if (io_u->resid) { | |
632 | int bytes = io_u->xfer_buflen - io_u->resid; | |
633 | struct fio_file *f = io_u->file; | |
634 | ||
635 | /* | |
636 | * zero read, fail | |
637 | */ | |
638 | if (!bytes) { | |
639 | td_verror(td, EIO, "full resid"); | |
640 | put_io_u(td, io_u); | |
641 | break; | |
642 | } | |
643 | ||
644 | io_u->xfer_buflen = io_u->resid; | |
645 | io_u->xfer_buf += bytes; | |
646 | io_u->offset += bytes; | |
647 | ||
648 | if (ddir_rw(io_u->ddir)) | |
649 | td->ts.short_io_u[io_u->ddir]++; | |
650 | ||
651 | if (io_u->offset == f->real_file_size) | |
652 | goto sync_done; | |
653 | ||
654 | requeue_io_u(td, &io_u); | |
655 | } else { | |
656 | sync_done: | |
6eaf09d6 SL |
657 | if (__should_check_rate(td, DDIR_READ) || |
658 | __should_check_rate(td, DDIR_WRITE) || | |
659 | __should_check_rate(td, DDIR_TRIM)) | |
2e1df07d JA |
660 | fio_gettime(&comp_time, NULL); |
661 | ||
662 | ret = io_u_sync_complete(td, io_u, bytes_done); | |
663 | if (ret < 0) | |
664 | break; | |
665 | } | |
666 | break; | |
667 | case FIO_Q_QUEUED: | |
668 | /* | |
669 | * if the engine doesn't have a commit hook, | |
670 | * the io_u is really queued. if it does have such | |
671 | * a hook, it has to call io_u_queued() itself. | |
672 | */ | |
673 | if (td->io_ops->commit == NULL) | |
674 | io_u_queued(td, io_u); | |
675 | break; | |
676 | case FIO_Q_BUSY: | |
677 | requeue_io_u(td, &io_u); | |
678 | ret2 = td_io_commit(td); | |
679 | if (ret2 < 0) | |
680 | ret = ret2; | |
681 | break; | |
682 | default: | |
683 | assert(ret < 0); | |
684 | put_io_u(td, io_u); | |
685 | break; | |
686 | } | |
687 | ||
688 | if (break_on_this_error(td, ddir, &ret)) | |
689 | break; | |
690 | ||
691 | /* | |
692 | * See if we need to complete some commands. Note that we | |
693 | * can get BUSY even without IO queued, if the system is | |
694 | * resource starved. | |
695 | */ | |
696 | full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth); | |
697 | if (full || !td->o.iodepth_batch_complete) { | |
698 | min_evts = min(td->o.iodepth_batch_complete, | |
699 | td->cur_depth); | |
8a74b56d JA |
700 | /* |
701 | * if the queue is full, we MUST reap at least 1 event | |
702 | */ | |
703 | if (full && !min_evts) | |
2e1df07d JA |
704 | min_evts = 1; |
705 | ||
6eaf09d6 SL |
706 | if (__should_check_rate(td, DDIR_READ) || |
707 | __should_check_rate(td, DDIR_WRITE) || | |
708 | __should_check_rate(td, DDIR_TRIM)) | |
2e1df07d JA |
709 | fio_gettime(&comp_time, NULL); |
710 | ||
711 | do { | |
712 | ret = io_u_queued_complete(td, min_evts, bytes_done); | |
713 | if (ret < 0) | |
714 | break; | |
715 | ||
716 | } while (full && (td->cur_depth > td->o.iodepth_low)); | |
717 | } | |
718 | ||
719 | if (ret < 0) | |
720 | break; | |
342f4be4 | 721 | if (!ddir_rw_sum(bytes_done)) |
2e1df07d JA |
722 | continue; |
723 | ||
724 | if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) { | |
725 | if (check_min_rate(td, &comp_time, bytes_done)) { | |
726 | if (exitall_on_terminate) | |
727 | fio_terminate_threads(td->groupid); | |
728 | td_verror(td, EIO, "check_min_rate"); | |
729 | break; | |
730 | } | |
731 | } | |
732 | ||
733 | if (td->o.thinktime) { | |
734 | unsigned long long b; | |
735 | ||
342f4be4 | 736 | b = ddir_rw_sum(td->io_blocks); |
2e1df07d JA |
737 | if (!(b % td->o.thinktime_blocks)) { |
738 | int left; | |
739 | ||
740 | if (td->o.thinktime_spin) | |
741 | usec_spin(td->o.thinktime_spin); | |
742 | ||
743 | left = td->o.thinktime - td->o.thinktime_spin; | |
744 | if (left) | |
745 | usec_sleep(td, left); | |
746 | } | |
747 | } | |
748 | } | |
749 | ||
750 | if (td->trim_entries) | |
751 | log_err("fio: %d trim entries leaked?\n", td->trim_entries); | |
752 | ||
753 | if (td->o.fill_device && td->error == ENOSPC) { | |
754 | td->error = 0; | |
755 | td->terminate = 1; | |
756 | } | |
757 | if (!td->error) { | |
758 | struct fio_file *f; | |
759 | ||
760 | i = td->cur_depth; | |
761 | if (i) { | |
762 | ret = io_u_queued_complete(td, i, NULL); | |
763 | if (td->o.fill_device && td->error == ENOSPC) | |
764 | td->error = 0; | |
765 | } | |
766 | ||
767 | if (should_fsync(td) && td->o.end_fsync) { | |
768 | td_set_runstate(td, TD_FSYNCING); | |
769 | ||
770 | for_each_file(td, f, i) { | |
771 | if (!fio_file_open(f)) | |
772 | continue; | |
773 | fio_io_sync(td, f); | |
774 | } | |
775 | } | |
776 | } else | |
777 | cleanup_pending_aio(td); | |
778 | ||
779 | /* | |
780 | * stop job if we failed doing any IO | |
781 | */ | |
342f4be4 | 782 | if (!ddir_rw_sum(td->this_io_bytes)) |
2e1df07d JA |
783 | td->done = 1; |
784 | } | |
785 | ||
786 | static void cleanup_io_u(struct thread_data *td) | |
787 | { | |
788 | struct flist_head *entry, *n; | |
789 | struct io_u *io_u; | |
790 | ||
791 | flist_for_each_safe(entry, n, &td->io_u_freelist) { | |
792 | io_u = flist_entry(entry, struct io_u, list); | |
793 | ||
794 | flist_del(&io_u->list); | |
795 | fio_memfree(io_u, sizeof(*io_u)); | |
796 | } | |
797 | ||
798 | free_io_mem(td); | |
799 | } | |
800 | ||
801 | static int init_io_u(struct thread_data *td) | |
802 | { | |
803 | struct io_u *io_u; | |
9c42684e | 804 | unsigned int max_bs, min_write; |
2e1df07d | 805 | int cl_align, i, max_units; |
59d8d0f5 | 806 | int data_xfer = 1; |
2e1df07d JA |
807 | char *p; |
808 | ||
809 | max_units = td->o.iodepth; | |
810 | max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]); | |
6eaf09d6 | 811 | max_bs = max(td->o.max_bs[DDIR_TRIM], max_bs); |
9c42684e | 812 | min_write = td->o.min_bs[DDIR_WRITE]; |
2e1df07d JA |
813 | td->orig_buffer_size = (unsigned long long) max_bs |
814 | * (unsigned long long) max_units; | |
815 | ||
88045e04 | 816 | if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td))) |
59d8d0f5 JA |
817 | data_xfer = 0; |
818 | ||
2e1df07d JA |
819 | if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) { |
820 | unsigned long bs; | |
821 | ||
822 | bs = td->orig_buffer_size + td->o.hugepage_size - 1; | |
823 | td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1); | |
824 | } | |
825 | ||
826 | if (td->orig_buffer_size != (size_t) td->orig_buffer_size) { | |
827 | log_err("fio: IO memory too large. Reduce max_bs or iodepth\n"); | |
828 | return 1; | |
829 | } | |
830 | ||
59d8d0f5 | 831 | if (data_xfer && allocate_io_mem(td)) |
2e1df07d JA |
832 | return 1; |
833 | ||
834 | if (td->o.odirect || td->o.mem_align || | |
835 | (td->io_ops->flags & FIO_RAWIO)) | |
836 | p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align; | |
837 | else | |
838 | p = td->orig_buffer; | |
839 | ||
840 | cl_align = os_cache_line_size(); | |
841 | ||
842 | for (i = 0; i < max_units; i++) { | |
843 | void *ptr; | |
844 | ||
845 | if (td->terminate) | |
846 | return 1; | |
847 | ||
848 | ptr = fio_memalign(cl_align, sizeof(*io_u)); | |
849 | if (!ptr) { | |
850 | log_err("fio: unable to allocate aligned memory\n"); | |
851 | break; | |
852 | } | |
853 | ||
854 | io_u = ptr; | |
855 | memset(io_u, 0, sizeof(*io_u)); | |
856 | INIT_FLIST_HEAD(&io_u->list); | |
857 | dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i); | |
858 | ||
59d8d0f5 | 859 | if (data_xfer) { |
2e1df07d JA |
860 | io_u->buf = p; |
861 | dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf); | |
862 | ||
863 | if (td_write(td)) | |
9c42684e | 864 | io_u_fill_buffer(td, io_u, min_write, max_bs); |
2e1df07d JA |
865 | if (td_write(td) && td->o.verify_pattern_bytes) { |
866 | /* | |
867 | * Fill the buffer with the pattern if we are | |
868 | * going to be doing writes. | |
869 | */ | |
870 | fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0); | |
871 | } | |
872 | } | |
873 | ||
874 | io_u->index = i; | |
875 | io_u->flags = IO_U_F_FREE; | |
876 | flist_add(&io_u->list, &td->io_u_freelist); | |
877 | p += max_bs; | |
878 | } | |
879 | ||
880 | return 0; | |
881 | } | |
882 | ||
883 | static int switch_ioscheduler(struct thread_data *td) | |
884 | { | |
885 | char tmp[256], tmp2[128]; | |
886 | FILE *f; | |
887 | int ret; | |
888 | ||
889 | if (td->io_ops->flags & FIO_DISKLESSIO) | |
890 | return 0; | |
891 | ||
892 | sprintf(tmp, "%s/queue/scheduler", td->sysfs_root); | |
893 | ||
894 | f = fopen(tmp, "r+"); | |
895 | if (!f) { | |
896 | if (errno == ENOENT) { | |
897 | log_err("fio: os or kernel doesn't support IO scheduler" | |
898 | " switching\n"); | |
899 | return 0; | |
900 | } | |
901 | td_verror(td, errno, "fopen iosched"); | |
902 | return 1; | |
903 | } | |
904 | ||
905 | /* | |
906 | * Set io scheduler. | |
907 | */ | |
908 | ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f); | |
909 | if (ferror(f) || ret != 1) { | |
910 | td_verror(td, errno, "fwrite"); | |
911 | fclose(f); | |
912 | return 1; | |
913 | } | |
914 | ||
915 | rewind(f); | |
916 | ||
917 | /* | |
918 | * Read back and check that the selected scheduler is now the default. | |
919 | */ | |
920 | ret = fread(tmp, 1, sizeof(tmp), f); | |
921 | if (ferror(f) || ret < 0) { | |
922 | td_verror(td, errno, "fread"); | |
923 | fclose(f); | |
924 | return 1; | |
925 | } | |
926 | ||
927 | sprintf(tmp2, "[%s]", td->o.ioscheduler); | |
928 | if (!strstr(tmp, tmp2)) { | |
929 | log_err("fio: io scheduler %s not found\n", td->o.ioscheduler); | |
930 | td_verror(td, EINVAL, "iosched_switch"); | |
931 | fclose(f); | |
932 | return 1; | |
933 | } | |
934 | ||
935 | fclose(f); | |
936 | return 0; | |
937 | } | |
938 | ||
939 | static int keep_running(struct thread_data *td) | |
940 | { | |
2e1df07d JA |
941 | if (td->done) |
942 | return 0; | |
943 | if (td->o.time_based) | |
944 | return 1; | |
945 | if (td->o.loops) { | |
946 | td->o.loops--; | |
947 | return 1; | |
948 | } | |
949 | ||
342f4be4 | 950 | if (ddir_rw_sum(td->io_bytes) < td->o.size) |
2e1df07d JA |
951 | return 1; |
952 | ||
953 | return 0; | |
954 | } | |
955 | ||
956 | static int exec_string(const char *string) | |
957 | { | |
958 | int ret, newlen = strlen(string) + 1 + 8; | |
959 | char *str; | |
960 | ||
961 | str = malloc(newlen); | |
962 | sprintf(str, "sh -c %s", string); | |
963 | ||
964 | ret = system(str); | |
965 | if (ret == -1) | |
966 | log_err("fio: exec of cmd <%s> failed\n", str); | |
967 | ||
968 | free(str); | |
969 | return ret; | |
970 | } | |
971 | ||
972 | /* | |
973 | * Entry point for the thread based jobs. The process based jobs end up | |
974 | * here as well, after a little setup. | |
975 | */ | |
976 | static void *thread_main(void *data) | |
977 | { | |
978 | unsigned long long elapsed; | |
979 | struct thread_data *td = data; | |
28727df7 | 980 | struct thread_options *o = &td->o; |
2e1df07d JA |
981 | pthread_condattr_t attr; |
982 | int clear_state; | |
28727df7 | 983 | int ret; |
2e1df07d | 984 | |
28727df7 | 985 | if (!o->use_thread) { |
2e1df07d JA |
986 | setsid(); |
987 | td->pid = getpid(); | |
988 | } else | |
989 | td->pid = gettid(); | |
990 | ||
991 | dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid); | |
992 | ||
122c7725 JA |
993 | if (is_backend) |
994 | fio_server_send_start(td); | |
995 | ||
2e1df07d JA |
996 | INIT_FLIST_HEAD(&td->io_u_freelist); |
997 | INIT_FLIST_HEAD(&td->io_u_busylist); | |
998 | INIT_FLIST_HEAD(&td->io_u_requeues); | |
999 | INIT_FLIST_HEAD(&td->io_log_list); | |
1000 | INIT_FLIST_HEAD(&td->io_hist_list); | |
1001 | INIT_FLIST_HEAD(&td->verify_list); | |
1002 | INIT_FLIST_HEAD(&td->trim_list); | |
1003 | pthread_mutex_init(&td->io_u_lock, NULL); | |
1004 | td->io_hist_tree = RB_ROOT; | |
1005 | ||
1006 | pthread_condattr_init(&attr); | |
1007 | pthread_cond_init(&td->verify_cond, &attr); | |
1008 | pthread_cond_init(&td->free_cond, &attr); | |
1009 | ||
1010 | td_set_runstate(td, TD_INITIALIZED); | |
1011 | dprint(FD_MUTEX, "up startup_mutex\n"); | |
1012 | fio_mutex_up(startup_mutex); | |
1013 | dprint(FD_MUTEX, "wait on td->mutex\n"); | |
1014 | fio_mutex_down(td->mutex); | |
1015 | dprint(FD_MUTEX, "done waiting on td->mutex\n"); | |
1016 | ||
1017 | /* | |
1018 | * the ->mutex mutex is now no longer used, close it to avoid | |
1019 | * eating a file descriptor | |
1020 | */ | |
1021 | fio_mutex_remove(td->mutex); | |
812409e9 | 1022 | td->mutex = NULL; |
2e1df07d JA |
1023 | |
1024 | /* | |
1025 | * A new gid requires privilege, so we need to do this before setting | |
1026 | * the uid. | |
1027 | */ | |
28727df7 | 1028 | if (o->gid != -1U && setgid(o->gid)) { |
2e1df07d JA |
1029 | td_verror(td, errno, "setgid"); |
1030 | goto err; | |
1031 | } | |
28727df7 | 1032 | if (o->uid != -1U && setuid(o->uid)) { |
2e1df07d JA |
1033 | td_verror(td, errno, "setuid"); |
1034 | goto err; | |
1035 | } | |
1036 | ||
1037 | /* | |
1038 | * If we have a gettimeofday() thread, make sure we exclude that | |
1039 | * thread from this job | |
1040 | */ | |
28727df7 JA |
1041 | if (o->gtod_cpu) |
1042 | fio_cpu_clear(&o->cpumask, o->gtod_cpu); | |
2e1df07d JA |
1043 | |
1044 | /* | |
1045 | * Set affinity first, in case it has an impact on the memory | |
1046 | * allocations. | |
1047 | */ | |
28727df7 JA |
1048 | if (o->cpumask_set) { |
1049 | ret = fio_setaffinity(td->pid, o->cpumask); | |
1050 | if (ret == -1) { | |
1051 | td_verror(td, errno, "cpu_set_affinity"); | |
1052 | goto err; | |
1053 | } | |
2e1df07d JA |
1054 | } |
1055 | ||
1b79a070 JA |
1056 | if (fio_pin_memory(td)) |
1057 | goto err; | |
1058 | ||
2e1df07d JA |
1059 | /* |
1060 | * May alter parameters that init_io_u() will use, so we need to | |
1061 | * do this first. | |
1062 | */ | |
1063 | if (init_iolog(td)) | |
1064 | goto err; | |
1065 | ||
1066 | if (init_io_u(td)) | |
1067 | goto err; | |
1068 | ||
28727df7 | 1069 | if (o->verify_async && verify_async_init(td)) |
2e1df07d JA |
1070 | goto err; |
1071 | ||
28727df7 JA |
1072 | if (o->ioprio) { |
1073 | ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio); | |
1074 | if (ret == -1) { | |
2e1df07d JA |
1075 | td_verror(td, errno, "ioprio_set"); |
1076 | goto err; | |
1077 | } | |
1078 | } | |
1079 | ||
5d89ff79 | 1080 | if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt)) |
2e1df07d JA |
1081 | goto err; |
1082 | ||
649c10c5 | 1083 | errno = 0; |
28727df7 | 1084 | if (nice(o->nice) == -1 && errno != 0) { |
2e1df07d JA |
1085 | td_verror(td, errno, "nice"); |
1086 | goto err; | |
1087 | } | |
1088 | ||
28727df7 | 1089 | if (o->ioscheduler && switch_ioscheduler(td)) |
2e1df07d JA |
1090 | goto err; |
1091 | ||
28727df7 | 1092 | if (!o->create_serialize && setup_files(td)) |
2e1df07d JA |
1093 | goto err; |
1094 | ||
1095 | if (td_io_init(td)) | |
1096 | goto err; | |
1097 | ||
1098 | if (init_random_map(td)) | |
1099 | goto err; | |
1100 | ||
28727df7 JA |
1101 | if (o->exec_prerun && exec_string(o->exec_prerun)) |
1102 | goto err; | |
2e1df07d | 1103 | |
28727df7 | 1104 | if (o->pre_read) { |
2e1df07d JA |
1105 | if (pre_read_files(td) < 0) |
1106 | goto err; | |
1107 | } | |
1108 | ||
83ea422a JA |
1109 | fio_verify_init(td); |
1110 | ||
2e1df07d JA |
1111 | fio_gettime(&td->epoch, NULL); |
1112 | getrusage(RUSAGE_SELF, &td->ru_start); | |
1113 | ||
1114 | clear_state = 0; | |
1115 | while (keep_running(td)) { | |
1116 | fio_gettime(&td->start, NULL); | |
1117 | memcpy(&td->bw_sample_time, &td->start, sizeof(td->start)); | |
1118 | memcpy(&td->iops_sample_time, &td->start, sizeof(td->start)); | |
1119 | memcpy(&td->tv_cache, &td->start, sizeof(td->start)); | |
1120 | ||
6eaf09d6 SL |
1121 | if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] || |
1122 | td->o.ratemin[DDIR_TRIM]) { | |
1123 | memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time, | |
2e1df07d | 1124 | sizeof(td->bw_sample_time)); |
6eaf09d6 | 1125 | memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time, |
2e1df07d | 1126 | sizeof(td->bw_sample_time)); |
6eaf09d6 | 1127 | memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time, |
2e1df07d JA |
1128 | sizeof(td->bw_sample_time)); |
1129 | } | |
1130 | ||
1131 | if (clear_state) | |
1132 | clear_io_state(td); | |
1133 | ||
1134 | prune_io_piece_log(td); | |
1135 | ||
1136 | do_io(td); | |
1137 | ||
1138 | clear_state = 1; | |
1139 | ||
1140 | if (td_read(td) && td->io_bytes[DDIR_READ]) { | |
1141 | elapsed = utime_since_now(&td->start); | |
1142 | td->ts.runtime[DDIR_READ] += elapsed; | |
1143 | } | |
1144 | if (td_write(td) && td->io_bytes[DDIR_WRITE]) { | |
1145 | elapsed = utime_since_now(&td->start); | |
1146 | td->ts.runtime[DDIR_WRITE] += elapsed; | |
1147 | } | |
6eaf09d6 SL |
1148 | if (td_trim(td) && td->io_bytes[DDIR_TRIM]) { |
1149 | elapsed = utime_since_now(&td->start); | |
1150 | td->ts.runtime[DDIR_TRIM] += elapsed; | |
1151 | } | |
2e1df07d JA |
1152 | |
1153 | if (td->error || td->terminate) | |
1154 | break; | |
1155 | ||
1156 | if (!td->o.do_verify || | |
1157 | td->o.verify == VERIFY_NONE || | |
1158 | (td->io_ops->flags & FIO_UNIDIR)) | |
1159 | continue; | |
1160 | ||
1161 | clear_io_state(td); | |
1162 | ||
1163 | fio_gettime(&td->start, NULL); | |
1164 | ||
1165 | do_verify(td); | |
1166 | ||
1167 | td->ts.runtime[DDIR_READ] += utime_since_now(&td->start); | |
1168 | ||
1169 | if (td->error || td->terminate) | |
1170 | break; | |
1171 | } | |
1172 | ||
1173 | update_rusage_stat(td); | |
6eaf09d6 SL |
1174 | td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000; |
1175 | td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000; | |
1176 | td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000; | |
2e1df07d | 1177 | td->ts.total_run_time = mtime_since_now(&td->epoch); |
6eaf09d6 SL |
1178 | td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ]; |
1179 | td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE]; | |
1180 | td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM]; | |
2e1df07d | 1181 | |
1b79a070 JA |
1182 | fio_unpin_memory(td); |
1183 | ||
2e1df07d JA |
1184 | fio_mutex_down(writeout_mutex); |
1185 | if (td->bw_log) { | |
1186 | if (td->o.bw_log_file) { | |
1187 | finish_log_named(td, td->bw_log, | |
1188 | td->o.bw_log_file, "bw"); | |
1189 | } else | |
1190 | finish_log(td, td->bw_log, "bw"); | |
1191 | } | |
1192 | if (td->lat_log) { | |
1193 | if (td->o.lat_log_file) { | |
1194 | finish_log_named(td, td->lat_log, | |
1195 | td->o.lat_log_file, "lat"); | |
1196 | } else | |
1197 | finish_log(td, td->lat_log, "lat"); | |
1198 | } | |
1199 | if (td->slat_log) { | |
1200 | if (td->o.lat_log_file) { | |
1201 | finish_log_named(td, td->slat_log, | |
1202 | td->o.lat_log_file, "slat"); | |
1203 | } else | |
1204 | finish_log(td, td->slat_log, "slat"); | |
1205 | } | |
1206 | if (td->clat_log) { | |
1207 | if (td->o.lat_log_file) { | |
1208 | finish_log_named(td, td->clat_log, | |
1209 | td->o.lat_log_file, "clat"); | |
1210 | } else | |
1211 | finish_log(td, td->clat_log, "clat"); | |
1212 | } | |
1213 | if (td->iops_log) { | |
1214 | if (td->o.iops_log_file) { | |
1215 | finish_log_named(td, td->iops_log, | |
1216 | td->o.iops_log_file, "iops"); | |
1217 | } else | |
1218 | finish_log(td, td->iops_log, "iops"); | |
1219 | } | |
1220 | ||
1221 | fio_mutex_up(writeout_mutex); | |
1222 | if (td->o.exec_postrun) | |
1223 | exec_string(td->o.exec_postrun); | |
1224 | ||
1225 | if (exitall_on_terminate) | |
1226 | fio_terminate_threads(td->groupid); | |
1227 | ||
1228 | err: | |
1229 | if (td->error) | |
1230 | log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error, | |
1231 | td->verror); | |
1232 | ||
1233 | if (td->o.verify_async) | |
1234 | verify_async_exit(td); | |
1235 | ||
1236 | close_and_free_files(td); | |
1237 | close_ioengine(td); | |
1238 | cleanup_io_u(td); | |
1239 | cgroup_shutdown(td, &cgroup_mnt); | |
1240 | ||
28727df7 JA |
1241 | if (o->cpumask_set) { |
1242 | int ret = fio_cpuset_exit(&o->cpumask); | |
2e1df07d JA |
1243 | |
1244 | td_verror(td, ret, "fio_cpuset_exit"); | |
1245 | } | |
1246 | ||
1247 | /* | |
1248 | * do this very late, it will log file closing as well | |
1249 | */ | |
1250 | if (td->o.write_iolog_file) | |
1251 | write_iolog_close(td); | |
1252 | ||
1253 | td_set_runstate(td, TD_EXITED); | |
e43606c2 | 1254 | return (void *) (uintptr_t) td->error; |
2e1df07d JA |
1255 | } |
1256 | ||
1257 | ||
1258 | /* | |
1259 | * We cannot pass the td data into a forked process, so attach the td and | |
1260 | * pass it to the thread worker. | |
1261 | */ | |
1262 | static int fork_main(int shmid, int offset) | |
1263 | { | |
1264 | struct thread_data *td; | |
1265 | void *data, *ret; | |
1266 | ||
1267 | #ifndef __hpux | |
1268 | data = shmat(shmid, NULL, 0); | |
1269 | if (data == (void *) -1) { | |
1270 | int __err = errno; | |
1271 | ||
1272 | perror("shmat"); | |
1273 | return __err; | |
1274 | } | |
1275 | #else | |
1276 | /* | |
1277 | * HP-UX inherits shm mappings? | |
1278 | */ | |
1279 | data = threads; | |
1280 | #endif | |
1281 | ||
1282 | td = data + offset * sizeof(struct thread_data); | |
1283 | ret = thread_main(td); | |
1284 | shmdt(data); | |
e43606c2 | 1285 | return (int) (uintptr_t) ret; |
2e1df07d JA |
1286 | } |
1287 | ||
1288 | /* | |
1289 | * Run over the job map and reap the threads that have exited, if any. | |
1290 | */ | |
1291 | static void reap_threads(unsigned int *nr_running, unsigned int *t_rate, | |
1292 | unsigned int *m_rate) | |
1293 | { | |
1294 | struct thread_data *td; | |
1295 | unsigned int cputhreads, realthreads, pending; | |
1296 | int i, status, ret; | |
1297 | ||
1298 | /* | |
1299 | * reap exited threads (TD_EXITED -> TD_REAPED) | |
1300 | */ | |
1301 | realthreads = pending = cputhreads = 0; | |
1302 | for_each_td(td, i) { | |
1303 | int flags = 0; | |
1304 | ||
1305 | /* | |
1306 | * ->io_ops is NULL for a thread that has closed its | |
1307 | * io engine | |
1308 | */ | |
1309 | if (td->io_ops && !strcmp(td->io_ops->name, "cpuio")) | |
1310 | cputhreads++; | |
1311 | else | |
1312 | realthreads++; | |
1313 | ||
1314 | if (!td->pid) { | |
1315 | pending++; | |
1316 | continue; | |
1317 | } | |
1318 | if (td->runstate == TD_REAPED) | |
1319 | continue; | |
1320 | if (td->o.use_thread) { | |
1321 | if (td->runstate == TD_EXITED) { | |
1322 | td_set_runstate(td, TD_REAPED); | |
1323 | goto reaped; | |
1324 | } | |
1325 | continue; | |
1326 | } | |
1327 | ||
1328 | flags = WNOHANG; | |
1329 | if (td->runstate == TD_EXITED) | |
1330 | flags = 0; | |
1331 | ||
1332 | /* | |
1333 | * check if someone quit or got killed in an unusual way | |
1334 | */ | |
1335 | ret = waitpid(td->pid, &status, flags); | |
1336 | if (ret < 0) { | |
1337 | if (errno == ECHILD) { | |
1338 | log_err("fio: pid=%d disappeared %d\n", | |
1339 | (int) td->pid, td->runstate); | |
a5e371a6 | 1340 | td->sig = ECHILD; |
2e1df07d JA |
1341 | td_set_runstate(td, TD_REAPED); |
1342 | goto reaped; | |
1343 | } | |
1344 | perror("waitpid"); | |
1345 | } else if (ret == td->pid) { | |
1346 | if (WIFSIGNALED(status)) { | |
1347 | int sig = WTERMSIG(status); | |
1348 | ||
1349 | if (sig != SIGTERM) | |
1350 | log_err("fio: pid=%d, got signal=%d\n", | |
1351 | (int) td->pid, sig); | |
a5e371a6 | 1352 | td->sig = sig; |
2e1df07d JA |
1353 | td_set_runstate(td, TD_REAPED); |
1354 | goto reaped; | |
1355 | } | |
1356 | if (WIFEXITED(status)) { | |
1357 | if (WEXITSTATUS(status) && !td->error) | |
1358 | td->error = WEXITSTATUS(status); | |
1359 | ||
1360 | td_set_runstate(td, TD_REAPED); | |
1361 | goto reaped; | |
1362 | } | |
1363 | } | |
1364 | ||
1365 | /* | |
1366 | * thread is not dead, continue | |
1367 | */ | |
1368 | pending++; | |
1369 | continue; | |
1370 | reaped: | |
1371 | (*nr_running)--; | |
342f4be4 JA |
1372 | (*m_rate) -= ddir_rw_sum(td->o.ratemin); |
1373 | (*t_rate) -= ddir_rw_sum(td->o.rate); | |
2e1df07d JA |
1374 | if (!td->pid) |
1375 | pending--; | |
1376 | ||
1377 | if (td->error) | |
1378 | exit_value++; | |
1379 | ||
1380 | done_secs += mtime_since_now(&td->epoch) / 1000; | |
1381 | } | |
1382 | ||
1383 | if (*nr_running == cputhreads && !pending && realthreads) | |
1384 | fio_terminate_threads(TERMINATE_ALL); | |
1385 | } | |
1386 | ||
2e1df07d JA |
1387 | /* |
1388 | * Main function for kicking off and reaping jobs, as needed. | |
1389 | */ | |
1390 | static void run_threads(void) | |
1391 | { | |
1392 | struct thread_data *td; | |
1393 | unsigned long spent; | |
1394 | unsigned int i, todo, nr_running, m_rate, t_rate, nr_started; | |
1395 | ||
2e1df07d JA |
1396 | if (fio_gtod_offload && fio_start_gtod_thread()) |
1397 | return; | |
1398 | ||
1399 | set_sig_handlers(); | |
1400 | ||
812409e9 JA |
1401 | nr_thread = nr_process = 0; |
1402 | for_each_td(td, i) { | |
1403 | if (td->o.use_thread) | |
1404 | nr_thread++; | |
1405 | else | |
1406 | nr_process++; | |
1407 | } | |
1408 | ||
f3afa57e | 1409 | if (output_format == FIO_OUTPUT_NORMAL) { |
2e1df07d JA |
1410 | log_info("Starting "); |
1411 | if (nr_thread) | |
1412 | log_info("%d thread%s", nr_thread, | |
1413 | nr_thread > 1 ? "s" : ""); | |
1414 | if (nr_process) { | |
1415 | if (nr_thread) | |
1416 | log_info(" and "); | |
1417 | log_info("%d process%s", nr_process, | |
1418 | nr_process > 1 ? "es" : ""); | |
1419 | } | |
1420 | log_info("\n"); | |
1421 | fflush(stdout); | |
1422 | } | |
1423 | ||
1424 | todo = thread_number; | |
1425 | nr_running = 0; | |
1426 | nr_started = 0; | |
1427 | m_rate = t_rate = 0; | |
1428 | ||
1429 | for_each_td(td, i) { | |
1430 | print_status_init(td->thread_number - 1); | |
1431 | ||
1432 | if (!td->o.create_serialize) | |
1433 | continue; | |
1434 | ||
1435 | /* | |
1436 | * do file setup here so it happens sequentially, | |
1437 | * we don't want X number of threads getting their | |
1438 | * client data interspersed on disk | |
1439 | */ | |
1440 | if (setup_files(td)) { | |
1441 | exit_value++; | |
1442 | if (td->error) | |
1443 | log_err("fio: pid=%d, err=%d/%s\n", | |
1444 | (int) td->pid, td->error, td->verror); | |
1445 | td_set_runstate(td, TD_REAPED); | |
1446 | todo--; | |
1447 | } else { | |
1448 | struct fio_file *f; | |
1449 | unsigned int j; | |
1450 | ||
1451 | /* | |
1452 | * for sharing to work, each job must always open | |
1453 | * its own files. so close them, if we opened them | |
1454 | * for creation | |
1455 | */ | |
1456 | for_each_file(td, f, j) { | |
1457 | if (fio_file_open(f)) | |
1458 | td_io_close_file(td, f); | |
1459 | } | |
1460 | } | |
1461 | } | |
1462 | ||
1463 | set_genesis_time(); | |
1464 | ||
1465 | while (todo) { | |
1466 | struct thread_data *map[REAL_MAX_JOBS]; | |
1467 | struct timeval this_start; | |
1468 | int this_jobs = 0, left; | |
1469 | ||
1470 | /* | |
1471 | * create threads (TD_NOT_CREATED -> TD_CREATED) | |
1472 | */ | |
1473 | for_each_td(td, i) { | |
1474 | if (td->runstate != TD_NOT_CREATED) | |
1475 | continue; | |
1476 | ||
1477 | /* | |
1478 | * never got a chance to start, killed by other | |
1479 | * thread for some reason | |
1480 | */ | |
1481 | if (td->terminate) { | |
1482 | todo--; | |
1483 | continue; | |
1484 | } | |
1485 | ||
1486 | if (td->o.start_delay) { | |
1487 | spent = mtime_since_genesis(); | |
1488 | ||
1489 | if (td->o.start_delay * 1000 > spent) | |
1490 | continue; | |
1491 | } | |
1492 | ||
1493 | if (td->o.stonewall && (nr_started || nr_running)) { | |
1494 | dprint(FD_PROCESS, "%s: stonewall wait\n", | |
1495 | td->o.name); | |
1496 | break; | |
1497 | } | |
1498 | ||
1499 | init_disk_util(td); | |
1500 | ||
1501 | /* | |
1502 | * Set state to created. Thread will transition | |
1503 | * to TD_INITIALIZED when it's done setting up. | |
1504 | */ | |
1505 | td_set_runstate(td, TD_CREATED); | |
1506 | map[this_jobs++] = td; | |
1507 | nr_started++; | |
1508 | ||
1509 | if (td->o.use_thread) { | |
1510 | int ret; | |
1511 | ||
1512 | dprint(FD_PROCESS, "will pthread_create\n"); | |
1513 | ret = pthread_create(&td->thread, NULL, | |
1514 | thread_main, td); | |
1515 | if (ret) { | |
1516 | log_err("pthread_create: %s\n", | |
1517 | strerror(ret)); | |
1518 | nr_started--; | |
1519 | break; | |
1520 | } | |
1521 | ret = pthread_detach(td->thread); | |
1522 | if (ret) | |
1523 | log_err("pthread_detach: %s", | |
1524 | strerror(ret)); | |
1525 | } else { | |
1526 | pid_t pid; | |
1527 | dprint(FD_PROCESS, "will fork\n"); | |
1528 | pid = fork(); | |
1529 | if (!pid) { | |
1530 | int ret = fork_main(shm_id, i); | |
1531 | ||
1532 | _exit(ret); | |
1533 | } else if (i == fio_debug_jobno) | |
1534 | *fio_debug_jobp = pid; | |
1535 | } | |
1536 | dprint(FD_MUTEX, "wait on startup_mutex\n"); | |
1537 | if (fio_mutex_down_timeout(startup_mutex, 10)) { | |
1538 | log_err("fio: job startup hung? exiting.\n"); | |
1539 | fio_terminate_threads(TERMINATE_ALL); | |
1540 | fio_abort = 1; | |
1541 | nr_started--; | |
1542 | break; | |
1543 | } | |
1544 | dprint(FD_MUTEX, "done waiting on startup_mutex\n"); | |
1545 | } | |
1546 | ||
1547 | /* | |
1548 | * Wait for the started threads to transition to | |
1549 | * TD_INITIALIZED. | |
1550 | */ | |
1551 | fio_gettime(&this_start, NULL); | |
1552 | left = this_jobs; | |
1553 | while (left && !fio_abort) { | |
1554 | if (mtime_since_now(&this_start) > JOB_START_TIMEOUT) | |
1555 | break; | |
1556 | ||
1557 | usleep(100000); | |
1558 | ||
1559 | for (i = 0; i < this_jobs; i++) { | |
1560 | td = map[i]; | |
1561 | if (!td) | |
1562 | continue; | |
1563 | if (td->runstate == TD_INITIALIZED) { | |
1564 | map[i] = NULL; | |
1565 | left--; | |
1566 | } else if (td->runstate >= TD_EXITED) { | |
1567 | map[i] = NULL; | |
1568 | left--; | |
1569 | todo--; | |
1570 | nr_running++; /* work-around... */ | |
1571 | } | |
1572 | } | |
1573 | } | |
1574 | ||
1575 | if (left) { | |
4e87c37a JA |
1576 | log_err("fio: %d job%s failed to start\n", left, |
1577 | left > 1 ? "s" : ""); | |
2e1df07d JA |
1578 | for (i = 0; i < this_jobs; i++) { |
1579 | td = map[i]; | |
1580 | if (!td) | |
1581 | continue; | |
1582 | kill(td->pid, SIGTERM); | |
1583 | } | |
1584 | break; | |
1585 | } | |
1586 | ||
1587 | /* | |
1588 | * start created threads (TD_INITIALIZED -> TD_RUNNING). | |
1589 | */ | |
1590 | for_each_td(td, i) { | |
1591 | if (td->runstate != TD_INITIALIZED) | |
1592 | continue; | |
1593 | ||
1594 | if (in_ramp_time(td)) | |
1595 | td_set_runstate(td, TD_RAMP); | |
1596 | else | |
1597 | td_set_runstate(td, TD_RUNNING); | |
1598 | nr_running++; | |
1599 | nr_started--; | |
342f4be4 JA |
1600 | m_rate += ddir_rw_sum(td->o.ratemin); |
1601 | t_rate += ddir_rw_sum(td->o.rate); | |
2e1df07d JA |
1602 | todo--; |
1603 | fio_mutex_up(td->mutex); | |
1604 | } | |
1605 | ||
1606 | reap_threads(&nr_running, &t_rate, &m_rate); | |
1607 | ||
122c7725 JA |
1608 | if (todo) |
1609 | usleep(100000); | |
2e1df07d JA |
1610 | } |
1611 | ||
1612 | while (nr_running) { | |
1613 | reap_threads(&nr_running, &t_rate, &m_rate); | |
122c7725 | 1614 | usleep(10000); |
2e1df07d JA |
1615 | } |
1616 | ||
1617 | update_io_ticks(); | |
2e1df07d JA |
1618 | } |
1619 | ||
feb41855 JA |
1620 | void wait_for_disk_thread_exit(void) |
1621 | { | |
1622 | fio_mutex_down(disk_thread_mutex); | |
1623 | } | |
1624 | ||
27357187 JA |
1625 | static void free_disk_util(void) |
1626 | { | |
1627 | disk_util_start_exit(); | |
1628 | wait_for_disk_thread_exit(); | |
1629 | disk_util_prune_entries(); | |
1630 | } | |
1631 | ||
2e1df07d JA |
1632 | static void *disk_thread_main(void *data) |
1633 | { | |
feb41855 JA |
1634 | int ret = 0; |
1635 | ||
2e1df07d JA |
1636 | fio_mutex_up(startup_mutex); |
1637 | ||
feb41855 | 1638 | while (threads && !ret) { |
2e1df07d JA |
1639 | usleep(DISK_UTIL_MSEC * 1000); |
1640 | if (!threads) | |
1641 | break; | |
feb41855 | 1642 | ret = update_io_ticks(); |
2e1df07d JA |
1643 | |
1644 | if (!is_backend) | |
1645 | print_thread_status(); | |
1646 | } | |
1647 | ||
feb41855 | 1648 | fio_mutex_up(disk_thread_mutex); |
2e1df07d JA |
1649 | return NULL; |
1650 | } | |
1651 | ||
1652 | static int create_disk_util_thread(void) | |
1653 | { | |
1654 | int ret; | |
1655 | ||
feb41855 JA |
1656 | setup_disk_util(); |
1657 | ||
521da527 | 1658 | disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED); |
feb41855 | 1659 | |
2e1df07d JA |
1660 | ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL); |
1661 | if (ret) { | |
feb41855 | 1662 | fio_mutex_remove(disk_thread_mutex); |
2e1df07d JA |
1663 | log_err("Can't create disk util thread: %s\n", strerror(ret)); |
1664 | return 1; | |
1665 | } | |
1666 | ||
1667 | ret = pthread_detach(disk_util_thread); | |
1668 | if (ret) { | |
feb41855 | 1669 | fio_mutex_remove(disk_thread_mutex); |
2e1df07d JA |
1670 | log_err("Can't detatch disk util thread: %s\n", strerror(ret)); |
1671 | return 1; | |
1672 | } | |
1673 | ||
1674 | dprint(FD_MUTEX, "wait on startup_mutex\n"); | |
1675 | fio_mutex_down(startup_mutex); | |
1676 | dprint(FD_MUTEX, "done waiting on startup_mutex\n"); | |
1677 | return 0; | |
1678 | } | |
1679 | ||
2e1df07d JA |
1680 | int fio_backend(void) |
1681 | { | |
1682 | struct thread_data *td; | |
1683 | int i; | |
1684 | ||
1685 | if (exec_profile) { | |
1686 | if (load_profile(exec_profile)) | |
1687 | return 1; | |
1688 | free(exec_profile); | |
1689 | exec_profile = NULL; | |
1690 | } | |
1691 | if (!thread_number) | |
1692 | return 0; | |
1693 | ||
1694 | if (write_bw_log) { | |
ea51b956 JA |
1695 | setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW); |
1696 | setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW); | |
d79db122 | 1697 | setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW); |
2e1df07d JA |
1698 | } |
1699 | ||
521da527 | 1700 | startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED); |
2e1df07d JA |
1701 | if (startup_mutex == NULL) |
1702 | return 1; | |
521da527 | 1703 | writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED); |
2e1df07d JA |
1704 | if (writeout_mutex == NULL) |
1705 | return 1; | |
1706 | ||
1707 | set_genesis_time(); | |
1708 | create_disk_util_thread(); | |
1709 | ||
1710 | cgroup_list = smalloc(sizeof(*cgroup_list)); | |
1711 | INIT_FLIST_HEAD(cgroup_list); | |
1712 | ||
1713 | run_threads(); | |
1714 | ||
1715 | if (!fio_abort) { | |
1716 | show_run_stats(); | |
1717 | if (write_bw_log) { | |
1718 | __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log"); | |
1719 | __finish_log(agg_io_log[DDIR_WRITE], | |
1720 | "agg-write_bw.log"); | |
6eaf09d6 SL |
1721 | __finish_log(agg_io_log[DDIR_TRIM], |
1722 | "agg-write_bw.log"); | |
2e1df07d JA |
1723 | } |
1724 | } | |
1725 | ||
1726 | for_each_td(td, i) | |
1727 | fio_options_free(td); | |
1728 | ||
d44256b0 | 1729 | free_disk_util(); |
2e1df07d JA |
1730 | cgroup_kill(cgroup_list); |
1731 | sfree(cgroup_list); | |
1732 | sfree(cgroup_mnt); | |
1733 | ||
1734 | fio_mutex_remove(startup_mutex); | |
1735 | fio_mutex_remove(writeout_mutex); | |
feb41855 | 1736 | fio_mutex_remove(disk_thread_mutex); |
2e1df07d JA |
1737 | return exit_value; |
1738 | } |