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> | |
2e1df07d | 37 | #include <sys/mman.h> |
ff6bb260 | 38 | #include <math.h> |
2e1df07d JA |
39 | |
40 | #include "fio.h" | |
a5e0ee11 O |
41 | #ifndef FIO_NO_HAVE_SHM_H |
42 | #include <sys/shm.h> | |
43 | #endif | |
2e1df07d JA |
44 | #include "hash.h" |
45 | #include "smalloc.h" | |
46 | #include "verify.h" | |
47 | #include "trim.h" | |
48 | #include "diskutil.h" | |
49 | #include "cgroup.h" | |
50 | #include "profile.h" | |
51 | #include "lib/rand.h" | |
f7690c4a | 52 | #include "lib/memalign.h" |
2e1df07d | 53 | #include "server.h" |
44404c5a | 54 | #include "lib/getrusage.h" |
f2a2ce0e | 55 | #include "idletime.h" |
002fe734 | 56 | #include "err.h" |
a9da8ab2 | 57 | #include "workqueue.h" |
e81ecca3 | 58 | #include "lib/mountcheck.h" |
40511301 | 59 | #include "rate-submit.h" |
a39fb9ea | 60 | #include "helper_thread.h" |
2e1df07d | 61 | |
2e1df07d | 62 | static struct fio_mutex *startup_mutex; |
2e1df07d JA |
63 | static struct flist_head *cgroup_list; |
64 | static char *cgroup_mnt; | |
65 | static int exit_value; | |
66 | static volatile int fio_abort; | |
3a5f6bde JA |
67 | static unsigned int nr_process = 0; |
68 | static unsigned int nr_thread = 0; | |
2e1df07d | 69 | |
6eaf09d6 | 70 | struct io_log *agg_io_log[DDIR_RWDIR_CNT]; |
2e1df07d | 71 | |
a3efc919 JA |
72 | int groupid = 0; |
73 | unsigned int thread_number = 0; | |
108fea77 | 74 | unsigned int stat_number = 0; |
a3efc919 JA |
75 | int shm_id = 0; |
76 | int temp_stall_ts; | |
77 | unsigned long done_secs = 0; | |
a47591e4 | 78 | |
2e1df07d JA |
79 | #define JOB_START_TIMEOUT (5 * 1000) |
80 | ||
81 | static void sig_int(int sig) | |
82 | { | |
83 | if (threads) { | |
84 | if (is_backend) | |
85 | fio_server_got_signal(sig); | |
86 | else { | |
87 | log_info("\nfio: terminating on signal %d\n", sig); | |
e411c301 | 88 | log_info_flush(); |
2e1df07d JA |
89 | exit_value = 128; |
90 | } | |
91 | ||
92 | fio_terminate_threads(TERMINATE_ALL); | |
93 | } | |
94 | } | |
95 | ||
d2235e56 | 96 | void sig_show_status(int sig) |
4c6d91e8 JA |
97 | { |
98 | show_running_run_stats(); | |
99 | } | |
100 | ||
2e1df07d JA |
101 | static void set_sig_handlers(void) |
102 | { | |
103 | struct sigaction act; | |
104 | ||
105 | memset(&act, 0, sizeof(act)); | |
106 | act.sa_handler = sig_int; | |
107 | act.sa_flags = SA_RESTART; | |
108 | sigaction(SIGINT, &act, NULL); | |
109 | ||
110 | memset(&act, 0, sizeof(act)); | |
111 | act.sa_handler = sig_int; | |
112 | act.sa_flags = SA_RESTART; | |
113 | sigaction(SIGTERM, &act, NULL); | |
114 | ||
2f694507 BC |
115 | /* Windows uses SIGBREAK as a quit signal from other applications */ |
116 | #ifdef WIN32 | |
117 | memset(&act, 0, sizeof(act)); | |
118 | act.sa_handler = sig_int; | |
119 | act.sa_flags = SA_RESTART; | |
120 | sigaction(SIGBREAK, &act, NULL); | |
121 | #endif | |
122 | ||
4c6d91e8 JA |
123 | memset(&act, 0, sizeof(act)); |
124 | act.sa_handler = sig_show_status; | |
125 | act.sa_flags = SA_RESTART; | |
126 | sigaction(SIGUSR1, &act, NULL); | |
127 | ||
2e1df07d JA |
128 | if (is_backend) { |
129 | memset(&act, 0, sizeof(act)); | |
130 | act.sa_handler = sig_int; | |
131 | act.sa_flags = SA_RESTART; | |
132 | sigaction(SIGPIPE, &act, NULL); | |
133 | } | |
134 | } | |
135 | ||
136 | /* | |
137 | * Check if we are above the minimum rate given. | |
138 | */ | |
8b6a404c | 139 | static bool __check_min_rate(struct thread_data *td, struct timespec *now, |
e39c0676 | 140 | enum fio_ddir ddir) |
2e1df07d JA |
141 | { |
142 | unsigned long long bytes = 0; | |
143 | unsigned long iops = 0; | |
144 | unsigned long spent; | |
145 | unsigned long rate; | |
146 | unsigned int ratemin = 0; | |
147 | unsigned int rate_iops = 0; | |
148 | unsigned int rate_iops_min = 0; | |
149 | ||
150 | assert(ddir_rw(ddir)); | |
151 | ||
152 | if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir]) | |
e39c0676 | 153 | return false; |
2e1df07d JA |
154 | |
155 | /* | |
156 | * allow a 2 second settle period in the beginning | |
157 | */ | |
158 | if (mtime_since(&td->start, now) < 2000) | |
e39c0676 | 159 | return false; |
2e1df07d JA |
160 | |
161 | iops += td->this_io_blocks[ddir]; | |
162 | bytes += td->this_io_bytes[ddir]; | |
163 | ratemin += td->o.ratemin[ddir]; | |
164 | rate_iops += td->o.rate_iops[ddir]; | |
165 | rate_iops_min += td->o.rate_iops_min[ddir]; | |
166 | ||
167 | /* | |
168 | * if rate blocks is set, sample is running | |
169 | */ | |
170 | if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) { | |
171 | spent = mtime_since(&td->lastrate[ddir], now); | |
172 | if (spent < td->o.ratecycle) | |
e39c0676 | 173 | return false; |
2e1df07d | 174 | |
20e37e4a | 175 | if (td->o.rate[ddir] || td->o.ratemin[ddir]) { |
2e1df07d JA |
176 | /* |
177 | * check bandwidth specified rate | |
178 | */ | |
179 | if (bytes < td->rate_bytes[ddir]) { | |
420b104a RE |
180 | log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n", |
181 | td->o.name, ratemin, bytes); | |
e39c0676 | 182 | return true; |
2e1df07d | 183 | } else { |
49cba9b3 JA |
184 | if (spent) |
185 | rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent; | |
186 | else | |
187 | rate = 0; | |
188 | ||
2e1df07d JA |
189 | if (rate < ratemin || |
190 | bytes < td->rate_bytes[ddir]) { | |
420b104a RE |
191 | log_err("%s: rate_min=%uB/s not met, got %luB/s\n", |
192 | td->o.name, ratemin, rate); | |
e39c0676 | 193 | return true; |
2e1df07d JA |
194 | } |
195 | } | |
196 | } else { | |
197 | /* | |
198 | * checks iops specified rate | |
199 | */ | |
200 | if (iops < rate_iops) { | |
420b104a RE |
201 | log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n", |
202 | td->o.name, rate_iops, iops); | |
e39c0676 | 203 | return true; |
2e1df07d | 204 | } else { |
4c707a3b JA |
205 | if (spent) |
206 | rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent; | |
207 | else | |
208 | rate = 0; | |
209 | ||
2e1df07d JA |
210 | if (rate < rate_iops_min || |
211 | iops < td->rate_blocks[ddir]) { | |
420b104a RE |
212 | log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n", |
213 | td->o.name, rate_iops_min, rate); | |
e39c0676 | 214 | return true; |
2e1df07d JA |
215 | } |
216 | } | |
217 | } | |
218 | } | |
219 | ||
220 | td->rate_bytes[ddir] = bytes; | |
221 | td->rate_blocks[ddir] = iops; | |
222 | memcpy(&td->lastrate[ddir], now, sizeof(*now)); | |
e39c0676 | 223 | return false; |
2e1df07d JA |
224 | } |
225 | ||
8b6a404c | 226 | static bool check_min_rate(struct thread_data *td, struct timespec *now) |
2e1df07d | 227 | { |
e39c0676 | 228 | bool ret = false; |
2e1df07d | 229 | |
55312f9f | 230 | if (td->bytes_done[DDIR_READ]) |
6eaf09d6 | 231 | ret |= __check_min_rate(td, now, DDIR_READ); |
55312f9f | 232 | if (td->bytes_done[DDIR_WRITE]) |
6eaf09d6 | 233 | ret |= __check_min_rate(td, now, DDIR_WRITE); |
55312f9f | 234 | if (td->bytes_done[DDIR_TRIM]) |
6eaf09d6 | 235 | ret |= __check_min_rate(td, now, DDIR_TRIM); |
2e1df07d JA |
236 | |
237 | return ret; | |
238 | } | |
239 | ||
240 | /* | |
241 | * When job exits, we can cancel the in-flight IO if we are using async | |
242 | * io. Attempt to do so. | |
243 | */ | |
244 | static void cleanup_pending_aio(struct thread_data *td) | |
245 | { | |
2e1df07d JA |
246 | int r; |
247 | ||
248 | /* | |
249 | * get immediately available events, if any | |
250 | */ | |
55312f9f | 251 | r = io_u_queued_complete(td, 0); |
2e1df07d JA |
252 | if (r < 0) |
253 | return; | |
254 | ||
255 | /* | |
256 | * now cancel remaining active events | |
257 | */ | |
258 | if (td->io_ops->cancel) { | |
2ae0b204 JA |
259 | struct io_u *io_u; |
260 | int i; | |
2e1df07d | 261 | |
2ae0b204 JA |
262 | io_u_qiter(&td->io_u_all, io_u, i) { |
263 | if (io_u->flags & IO_U_F_FLIGHT) { | |
2e1df07d JA |
264 | r = td->io_ops->cancel(td, io_u); |
265 | if (!r) | |
266 | put_io_u(td, io_u); | |
267 | } | |
268 | } | |
269 | } | |
270 | ||
271 | if (td->cur_depth) | |
55312f9f | 272 | r = io_u_queued_complete(td, td->cur_depth); |
2e1df07d JA |
273 | } |
274 | ||
275 | /* | |
276 | * Helper to handle the final sync of a file. Works just like the normal | |
277 | * io path, just does everything sync. | |
278 | */ | |
e39c0676 | 279 | static bool fio_io_sync(struct thread_data *td, struct fio_file *f) |
2e1df07d JA |
280 | { |
281 | struct io_u *io_u = __get_io_u(td); | |
282 | int ret; | |
283 | ||
284 | if (!io_u) | |
e39c0676 | 285 | return true; |
2e1df07d JA |
286 | |
287 | io_u->ddir = DDIR_SYNC; | |
288 | io_u->file = f; | |
289 | ||
290 | if (td_io_prep(td, io_u)) { | |
291 | put_io_u(td, io_u); | |
e39c0676 | 292 | return true; |
2e1df07d JA |
293 | } |
294 | ||
295 | requeue: | |
296 | ret = td_io_queue(td, io_u); | |
297 | if (ret < 0) { | |
298 | td_verror(td, io_u->error, "td_io_queue"); | |
299 | put_io_u(td, io_u); | |
e39c0676 | 300 | return true; |
2e1df07d | 301 | } else if (ret == FIO_Q_QUEUED) { |
42ff945c JA |
302 | if (td_io_commit(td)) |
303 | return true; | |
55312f9f | 304 | if (io_u_queued_complete(td, 1) < 0) |
e39c0676 | 305 | return true; |
2e1df07d JA |
306 | } else if (ret == FIO_Q_COMPLETED) { |
307 | if (io_u->error) { | |
308 | td_verror(td, io_u->error, "td_io_queue"); | |
e39c0676 | 309 | return true; |
2e1df07d JA |
310 | } |
311 | ||
55312f9f | 312 | if (io_u_sync_complete(td, io_u) < 0) |
e39c0676 | 313 | return true; |
2e1df07d JA |
314 | } else if (ret == FIO_Q_BUSY) { |
315 | if (td_io_commit(td)) | |
e39c0676 | 316 | return true; |
2e1df07d JA |
317 | goto requeue; |
318 | } | |
319 | ||
e39c0676 | 320 | return false; |
2e1df07d | 321 | } |
a3efc919 | 322 | |
61ee0f86 JA |
323 | static int fio_file_fsync(struct thread_data *td, struct fio_file *f) |
324 | { | |
325 | int ret; | |
326 | ||
327 | if (fio_file_open(f)) | |
328 | return fio_io_sync(td, f); | |
329 | ||
330 | if (td_io_open_file(td, f)) | |
331 | return 1; | |
332 | ||
333 | ret = fio_io_sync(td, f); | |
334 | td_io_close_file(td, f); | |
335 | return ret; | |
336 | } | |
337 | ||
8b6a404c | 338 | static inline void __update_ts_cache(struct thread_data *td) |
2e1df07d | 339 | { |
8b6a404c | 340 | fio_gettime(&td->ts_cache, NULL); |
2e1df07d JA |
341 | } |
342 | ||
8b6a404c | 343 | static inline void update_ts_cache(struct thread_data *td) |
2e1df07d | 344 | { |
8b6a404c VF |
345 | if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask) |
346 | __update_ts_cache(td); | |
2e1df07d JA |
347 | } |
348 | ||
8b6a404c | 349 | static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t) |
2e1df07d JA |
350 | { |
351 | if (in_ramp_time(td)) | |
e39c0676 | 352 | return false; |
2e1df07d | 353 | if (!td->o.timeout) |
e39c0676 | 354 | return false; |
0de5b26f | 355 | if (utime_since(&td->epoch, t) >= td->o.timeout) |
e39c0676 | 356 | return true; |
2e1df07d | 357 | |
e39c0676 | 358 | return false; |
2e1df07d JA |
359 | } |
360 | ||
95603b74 BF |
361 | /* |
362 | * We need to update the runtime consistently in ms, but keep a running | |
363 | * tally of the current elapsed time in microseconds for sub millisecond | |
364 | * updates. | |
365 | */ | |
366 | static inline void update_runtime(struct thread_data *td, | |
367 | unsigned long long *elapsed_us, | |
368 | const enum fio_ddir ddir) | |
369 | { | |
d5c8ea29 JA |
370 | if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only) |
371 | return; | |
372 | ||
95603b74 BF |
373 | td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000; |
374 | elapsed_us[ddir] += utime_since_now(&td->start); | |
375 | td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000; | |
376 | } | |
377 | ||
e39c0676 JA |
378 | static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir, |
379 | int *retptr) | |
2e1df07d JA |
380 | { |
381 | int ret = *retptr; | |
382 | ||
383 | if (ret < 0 || td->error) { | |
8b28bd41 DM |
384 | int err = td->error; |
385 | enum error_type_bit eb; | |
2e1df07d JA |
386 | |
387 | if (ret < 0) | |
388 | err = -ret; | |
2e1df07d | 389 | |
8b28bd41 DM |
390 | eb = td_error_type(ddir, err); |
391 | if (!(td->o.continue_on_error & (1 << eb))) | |
e39c0676 | 392 | return true; |
2e1df07d | 393 | |
8b28bd41 | 394 | if (td_non_fatal_error(td, eb, err)) { |
2e1df07d JA |
395 | /* |
396 | * Continue with the I/Os in case of | |
397 | * a non fatal error. | |
398 | */ | |
399 | update_error_count(td, err); | |
400 | td_clear_error(td); | |
401 | *retptr = 0; | |
e39c0676 | 402 | return false; |
2e1df07d JA |
403 | } else if (td->o.fill_device && err == ENOSPC) { |
404 | /* | |
405 | * We expect to hit this error if | |
406 | * fill_device option is set. | |
407 | */ | |
408 | td_clear_error(td); | |
ebea2133 | 409 | fio_mark_td_terminate(td); |
e39c0676 | 410 | return true; |
2e1df07d JA |
411 | } else { |
412 | /* | |
413 | * Stop the I/O in case of a fatal | |
414 | * error. | |
415 | */ | |
416 | update_error_count(td, err); | |
e39c0676 | 417 | return true; |
2e1df07d JA |
418 | } |
419 | } | |
420 | ||
e39c0676 | 421 | return false; |
2e1df07d JA |
422 | } |
423 | ||
c97f1ad6 JA |
424 | static void check_update_rusage(struct thread_data *td) |
425 | { | |
426 | if (td->update_rusage) { | |
427 | td->update_rusage = 0; | |
428 | update_rusage_stat(td); | |
429 | fio_mutex_up(td->rusage_sem); | |
430 | } | |
431 | } | |
432 | ||
8b6a404c | 433 | static int wait_for_completions(struct thread_data *td, struct timespec *time) |
69fea81e JA |
434 | { |
435 | const int full = queue_full(td); | |
436 | int min_evts = 0; | |
437 | int ret; | |
438 | ||
6be06c46 JA |
439 | if (td->flags & TD_F_REGROW_LOGS) |
440 | return io_u_quiesce(td); | |
1fed2080 | 441 | |
69fea81e JA |
442 | /* |
443 | * if the queue is full, we MUST reap at least 1 event | |
444 | */ | |
82407585 | 445 | min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth); |
4b157ac6 | 446 | if ((full && !min_evts) || !td->o.iodepth_batch_complete_min) |
69fea81e JA |
447 | min_evts = 1; |
448 | ||
66b6c5ef | 449 | if (time && (__should_check_rate(td, DDIR_READ) || |
69fea81e JA |
450 | __should_check_rate(td, DDIR_WRITE) || |
451 | __should_check_rate(td, DDIR_TRIM))) | |
452 | fio_gettime(time, NULL); | |
453 | ||
454 | do { | |
55312f9f | 455 | ret = io_u_queued_complete(td, min_evts); |
69fea81e JA |
456 | if (ret < 0) |
457 | break; | |
458 | } while (full && (td->cur_depth > td->o.iodepth_low)); | |
459 | ||
460 | return ret; | |
461 | } | |
462 | ||
e9d512d8 JA |
463 | int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret, |
464 | enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify, | |
8b6a404c | 465 | struct timespec *comp_time) |
e9d512d8 JA |
466 | { |
467 | int ret2; | |
468 | ||
469 | switch (*ret) { | |
470 | case FIO_Q_COMPLETED: | |
471 | if (io_u->error) { | |
472 | *ret = -io_u->error; | |
473 | clear_io_u(td, io_u); | |
474 | } else if (io_u->resid) { | |
475 | int bytes = io_u->xfer_buflen - io_u->resid; | |
476 | struct fio_file *f = io_u->file; | |
477 | ||
478 | if (bytes_issued) | |
479 | *bytes_issued += bytes; | |
480 | ||
481 | if (!from_verify) | |
482 | trim_io_piece(td, io_u); | |
483 | ||
484 | /* | |
485 | * zero read, fail | |
486 | */ | |
487 | if (!bytes) { | |
488 | if (!from_verify) | |
489 | unlog_io_piece(td, io_u); | |
490 | td_verror(td, EIO, "full resid"); | |
491 | put_io_u(td, io_u); | |
492 | break; | |
493 | } | |
494 | ||
495 | io_u->xfer_buflen = io_u->resid; | |
496 | io_u->xfer_buf += bytes; | |
497 | io_u->offset += bytes; | |
498 | ||
499 | if (ddir_rw(io_u->ddir)) | |
500 | td->ts.short_io_u[io_u->ddir]++; | |
501 | ||
502 | f = io_u->file; | |
503 | if (io_u->offset == f->real_file_size) | |
504 | goto sync_done; | |
505 | ||
506 | requeue_io_u(td, &io_u); | |
507 | } else { | |
508 | sync_done: | |
509 | if (comp_time && (__should_check_rate(td, DDIR_READ) || | |
510 | __should_check_rate(td, DDIR_WRITE) || | |
511 | __should_check_rate(td, DDIR_TRIM))) | |
512 | fio_gettime(comp_time, NULL); | |
513 | ||
514 | *ret = io_u_sync_complete(td, io_u); | |
515 | if (*ret < 0) | |
516 | break; | |
517 | } | |
a0e3e5a6 | 518 | |
cf2c8d52 JA |
519 | if (td->flags & TD_F_REGROW_LOGS) |
520 | regrow_logs(td); | |
521 | ||
a0e3e5a6 EW |
522 | /* |
523 | * when doing I/O (not when verifying), | |
524 | * check for any errors that are to be ignored | |
525 | */ | |
526 | if (!from_verify) | |
527 | break; | |
528 | ||
e9d512d8 JA |
529 | return 0; |
530 | case FIO_Q_QUEUED: | |
531 | /* | |
532 | * if the engine doesn't have a commit hook, | |
533 | * the io_u is really queued. if it does have such | |
534 | * a hook, it has to call io_u_queued() itself. | |
535 | */ | |
536 | if (td->io_ops->commit == NULL) | |
537 | io_u_queued(td, io_u); | |
538 | if (bytes_issued) | |
539 | *bytes_issued += io_u->xfer_buflen; | |
540 | break; | |
541 | case FIO_Q_BUSY: | |
542 | if (!from_verify) | |
543 | unlog_io_piece(td, io_u); | |
544 | requeue_io_u(td, &io_u); | |
545 | ret2 = td_io_commit(td); | |
546 | if (ret2 < 0) | |
547 | *ret = ret2; | |
548 | break; | |
549 | default: | |
b0775325 | 550 | assert(*ret < 0); |
e9d512d8 JA |
551 | td_verror(td, -(*ret), "td_io_queue"); |
552 | break; | |
553 | } | |
554 | ||
555 | if (break_on_this_error(td, ddir, ret)) | |
556 | return 1; | |
557 | ||
558 | return 0; | |
559 | } | |
560 | ||
e39c0676 | 561 | static inline bool io_in_polling(struct thread_data *td) |
82407585 RP |
562 | { |
563 | return !td->o.iodepth_batch_complete_min && | |
564 | !td->o.iodepth_batch_complete_max; | |
565 | } | |
39c1c323 | 566 | /* |
567 | * Unlinks files from thread data fio_file structure | |
568 | */ | |
569 | static int unlink_all_files(struct thread_data *td) | |
570 | { | |
571 | struct fio_file *f; | |
572 | unsigned int i; | |
573 | int ret = 0; | |
574 | ||
575 | for_each_file(td, f, i) { | |
576 | if (f->filetype != FIO_TYPE_FILE) | |
577 | continue; | |
578 | ret = td_io_unlink_file(td, f); | |
579 | if (ret) | |
580 | break; | |
581 | } | |
582 | ||
583 | if (ret) | |
584 | td_verror(td, ret, "unlink_all_files"); | |
585 | ||
586 | return ret; | |
587 | } | |
82407585 | 588 | |
2e1df07d JA |
589 | /* |
590 | * The main verify engine. Runs over the writes we previously submitted, | |
591 | * reads the blocks back in, and checks the crc/md5 of the data. | |
592 | */ | |
100f49f1 | 593 | static void do_verify(struct thread_data *td, uint64_t verify_bytes) |
2e1df07d JA |
594 | { |
595 | struct fio_file *f; | |
596 | struct io_u *io_u; | |
597 | int ret, min_events; | |
598 | unsigned int i; | |
599 | ||
600 | dprint(FD_VERIFY, "starting loop\n"); | |
601 | ||
602 | /* | |
603 | * sync io first and invalidate cache, to make sure we really | |
604 | * read from disk. | |
605 | */ | |
606 | for_each_file(td, f, i) { | |
607 | if (!fio_file_open(f)) | |
608 | continue; | |
609 | if (fio_io_sync(td, f)) | |
610 | break; | |
611 | if (file_invalidate_cache(td, f)) | |
612 | break; | |
613 | } | |
614 | ||
c97f1ad6 JA |
615 | check_update_rusage(td); |
616 | ||
2e1df07d JA |
617 | if (td->error) |
618 | return; | |
619 | ||
8f380de2 CJ |
620 | /* |
621 | * verify_state needs to be reset before verification | |
622 | * proceeds so that expected random seeds match actual | |
623 | * random seeds in headers. The main loop will reset | |
624 | * all random number generators if randrepeat is set. | |
625 | */ | |
626 | if (!td->o.rand_repeatable) | |
627 | td_fill_verify_state_seed(td); | |
628 | ||
2e1df07d JA |
629 | td_set_runstate(td, TD_VERIFYING); |
630 | ||
631 | io_u = NULL; | |
632 | while (!td->terminate) { | |
fbccf46c | 633 | enum fio_ddir ddir; |
e9d512d8 | 634 | int full; |
2e1df07d | 635 | |
8b6a404c | 636 | update_ts_cache(td); |
c97f1ad6 | 637 | check_update_rusage(td); |
2e1df07d | 638 | |
8b6a404c VF |
639 | if (runtime_exceeded(td, &td->ts_cache)) { |
640 | __update_ts_cache(td); | |
641 | if (runtime_exceeded(td, &td->ts_cache)) { | |
ebea2133 | 642 | fio_mark_td_terminate(td); |
2e1df07d JA |
643 | break; |
644 | } | |
645 | } | |
646 | ||
9e684a49 DE |
647 | if (flow_threshold_exceeded(td)) |
648 | continue; | |
649 | ||
44cbc6da JA |
650 | if (!td->o.experimental_verify) { |
651 | io_u = __get_io_u(td); | |
652 | if (!io_u) | |
653 | break; | |
2e1df07d | 654 | |
44cbc6da JA |
655 | if (get_next_verify(td, io_u)) { |
656 | put_io_u(td, io_u); | |
657 | break; | |
658 | } | |
2e1df07d | 659 | |
44cbc6da JA |
660 | if (td_io_prep(td, io_u)) { |
661 | put_io_u(td, io_u); | |
662 | break; | |
663 | } | |
664 | } else { | |
55312f9f | 665 | if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes) |
100f49f1 JA |
666 | break; |
667 | ||
bcd5abfa | 668 | while ((io_u = get_io_u(td)) != NULL) { |
059b61f2 | 669 | if (IS_ERR_OR_NULL(io_u)) { |
002fe734 JA |
670 | io_u = NULL; |
671 | ret = FIO_Q_BUSY; | |
672 | goto reap; | |
673 | } | |
674 | ||
bcd5abfa JA |
675 | /* |
676 | * We are only interested in the places where | |
677 | * we wrote or trimmed IOs. Turn those into | |
678 | * reads for verification purposes. | |
679 | */ | |
680 | if (io_u->ddir == DDIR_READ) { | |
681 | /* | |
682 | * Pretend we issued it for rwmix | |
683 | * accounting | |
684 | */ | |
685 | td->io_issues[DDIR_READ]++; | |
686 | put_io_u(td, io_u); | |
687 | continue; | |
688 | } else if (io_u->ddir == DDIR_TRIM) { | |
689 | io_u->ddir = DDIR_READ; | |
1651e431 | 690 | io_u_set(td, io_u, IO_U_F_TRIMMED); |
bcd5abfa JA |
691 | break; |
692 | } else if (io_u->ddir == DDIR_WRITE) { | |
693 | io_u->ddir = DDIR_READ; | |
694 | break; | |
695 | } else { | |
696 | put_io_u(td, io_u); | |
697 | continue; | |
698 | } | |
699 | } | |
44cbc6da | 700 | |
bcd5abfa | 701 | if (!io_u) |
44cbc6da | 702 | break; |
2e1df07d JA |
703 | } |
704 | ||
ca09be4b JA |
705 | if (verify_state_should_stop(td, io_u)) { |
706 | put_io_u(td, io_u); | |
707 | break; | |
708 | } | |
709 | ||
2e1df07d JA |
710 | if (td->o.verify_async) |
711 | io_u->end_io = verify_io_u_async; | |
712 | else | |
713 | io_u->end_io = verify_io_u; | |
714 | ||
fbccf46c | 715 | ddir = io_u->ddir; |
10adc4a7 GG |
716 | if (!td->o.disable_slat) |
717 | fio_gettime(&io_u->start_time, NULL); | |
fbccf46c | 718 | |
2e1df07d | 719 | ret = td_io_queue(td, io_u); |
2e1df07d | 720 | |
e9d512d8 | 721 | if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL)) |
2e1df07d JA |
722 | break; |
723 | ||
724 | /* | |
725 | * if we can queue more, do so. but check if there are | |
726 | * completed io_u's first. Note that we can get BUSY even | |
727 | * without IO queued, if the system is resource starved. | |
728 | */ | |
002fe734 | 729 | reap: |
2e1df07d | 730 | full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth); |
82407585 | 731 | if (full || io_in_polling(td)) |
55312f9f | 732 | ret = wait_for_completions(td, NULL); |
2e1df07d | 733 | |
2e1df07d JA |
734 | if (ret < 0) |
735 | break; | |
736 | } | |
737 | ||
c97f1ad6 JA |
738 | check_update_rusage(td); |
739 | ||
2e1df07d JA |
740 | if (!td->error) { |
741 | min_events = td->cur_depth; | |
742 | ||
743 | if (min_events) | |
55312f9f | 744 | ret = io_u_queued_complete(td, min_events); |
2e1df07d JA |
745 | } else |
746 | cleanup_pending_aio(td); | |
747 | ||
748 | td_set_runstate(td, TD_RUNNING); | |
749 | ||
750 | dprint(FD_VERIFY, "exiting loop\n"); | |
751 | } | |
752 | ||
e39c0676 | 753 | static bool exceeds_number_ios(struct thread_data *td) |
3939fe85 JA |
754 | { |
755 | unsigned long long number_ios; | |
756 | ||
757 | if (!td->o.number_ios) | |
e39c0676 | 758 | return false; |
3939fe85 | 759 | |
cf8a46a7 | 760 | number_ios = ddir_rw_sum(td->io_blocks); |
3939fe85 JA |
761 | number_ios += td->io_u_queued + td->io_u_in_flight; |
762 | ||
cf8a46a7 | 763 | return number_ios >= (td->o.number_ios * td->loops); |
3939fe85 JA |
764 | } |
765 | ||
926098e0 | 766 | static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes) |
f7078f7b | 767 | { |
77731b29 | 768 | unsigned long long bytes, limit; |
f7078f7b JA |
769 | |
770 | if (td_rw(td)) | |
926098e0 | 771 | bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE]; |
f7078f7b | 772 | else if (td_write(td)) |
926098e0 | 773 | bytes = this_bytes[DDIR_WRITE]; |
6eaf09d6 | 774 | else if (td_read(td)) |
926098e0 | 775 | bytes = this_bytes[DDIR_READ]; |
f7078f7b | 776 | else |
926098e0 | 777 | bytes = this_bytes[DDIR_TRIM]; |
f7078f7b | 778 | |
5be9bf09 TK |
779 | if (td->o.io_size) |
780 | limit = td->o.io_size; | |
77731b29 JA |
781 | else |
782 | limit = td->o.size; | |
783 | ||
cf8a46a7 | 784 | limit *= td->loops; |
77731b29 | 785 | return bytes >= limit || exceeds_number_ios(td); |
f7078f7b JA |
786 | } |
787 | ||
926098e0 | 788 | static bool io_issue_bytes_exceeded(struct thread_data *td) |
e28dd2cf | 789 | { |
926098e0 JA |
790 | return io_bytes_exceeded(td, td->io_issue_bytes); |
791 | } | |
e28dd2cf | 792 | |
926098e0 JA |
793 | static bool io_complete_bytes_exceeded(struct thread_data *td) |
794 | { | |
795 | return io_bytes_exceeded(td, td->this_io_bytes); | |
e28dd2cf JE |
796 | } |
797 | ||
50a8ce86 D |
798 | /* |
799 | * used to calculate the next io time for rate control | |
800 | * | |
801 | */ | |
802 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) | |
803 | { | |
ff6bb260 | 804 | uint64_t secs, remainder, bps, bytes, iops; |
50a8ce86 D |
805 | |
806 | assert(!(td->flags & TD_F_CHILD)); | |
807 | bytes = td->rate_io_issue_bytes[ddir]; | |
808 | bps = td->rate_bps[ddir]; | |
ff6bb260 | 809 | |
6de65959 | 810 | if (td->o.rate_process == RATE_PROCESS_POISSON) { |
a77d139b | 811 | uint64_t val; |
ff6bb260 | 812 | iops = bps / td->o.bs[ddir]; |
a77d139b | 813 | val = (int64_t) (1000000 / iops) * |
6c8611c7 | 814 | -logf(__rand_0_1(&td->poisson_state[ddir])); |
a77d139b | 815 | if (val) { |
98e51d7c JA |
816 | dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n", |
817 | (unsigned long long) 1000000 / val, | |
818 | ddir); | |
a77d139b | 819 | } |
6c8611c7 JA |
820 | td->last_usec[ddir] += val; |
821 | return td->last_usec[ddir]; | |
ff6bb260 | 822 | } else if (bps) { |
50a8ce86 D |
823 | secs = bytes / bps; |
824 | remainder = bytes % bps; | |
825 | return remainder * 1000000 / bps + secs * 1000000; | |
e7b24047 JA |
826 | } |
827 | ||
828 | return 0; | |
50a8ce86 D |
829 | } |
830 | ||
2e1df07d JA |
831 | /* |
832 | * Main IO worker function. It retrieves io_u's to process and queues | |
833 | * and reaps them, checking for rate and errors along the way. | |
100f49f1 JA |
834 | * |
835 | * Returns number of bytes written and trimmed. | |
2e1df07d | 836 | */ |
95b03be0 | 837 | static void do_io(struct thread_data *td, uint64_t *bytes_done) |
2e1df07d JA |
838 | { |
839 | unsigned int i; | |
840 | int ret = 0; | |
c2703bf3 | 841 | uint64_t total_bytes, bytes_issued = 0; |
70c68076 | 842 | |
95b03be0 JA |
843 | for (i = 0; i < DDIR_RWDIR_CNT; i++) |
844 | bytes_done[i] = td->bytes_done[i]; | |
2e1df07d JA |
845 | |
846 | if (in_ramp_time(td)) | |
847 | td_set_runstate(td, TD_RAMP); | |
848 | else | |
849 | td_set_runstate(td, TD_RUNNING); | |
850 | ||
3e260a46 JA |
851 | lat_target_init(td); |
852 | ||
1e564979 JE |
853 | total_bytes = td->o.size; |
854 | /* | |
5be9bf09 | 855 | * Allow random overwrite workloads to write up to io_size |
1e564979 JE |
856 | * before starting verification phase as 'size' doesn't apply. |
857 | */ | |
858 | if (td_write(td) && td_random(td) && td->o.norandommap) | |
5be9bf09 | 859 | total_bytes = max(total_bytes, (uint64_t) td->o.io_size); |
78a6469c JA |
860 | /* |
861 | * If verify_backlog is enabled, we'll run the verify in this | |
862 | * handler as well. For that case, we may need up to twice the | |
863 | * amount of bytes. | |
864 | */ | |
78a6469c JA |
865 | if (td->o.verify != VERIFY_NONE && |
866 | (td_write(td) && td->o.verify_backlog)) | |
c2703bf3 JA |
867 | total_bytes += td->o.size; |
868 | ||
82a90686 | 869 | /* In trimwrite mode, each byte is trimmed and then written, so |
0e4dd95c | 870 | * allow total_bytes to be twice as big */ |
82a90686 | 871 | if (td_trimwrite(td)) |
0e4dd95c DE |
872 | total_bytes += td->total_io_size; |
873 | ||
f7078f7b | 874 | while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || |
e28dd2cf | 875 | (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) || |
c04e4661 | 876 | td->o.time_based) { |
8b6a404c | 877 | struct timespec comp_time; |
2e1df07d | 878 | struct io_u *io_u; |
e9d512d8 | 879 | int full; |
2e1df07d JA |
880 | enum fio_ddir ddir; |
881 | ||
c97f1ad6 JA |
882 | check_update_rusage(td); |
883 | ||
7d7803fa | 884 | if (td->terminate || td->done) |
2e1df07d JA |
885 | break; |
886 | ||
8b6a404c | 887 | update_ts_cache(td); |
2e1df07d | 888 | |
8b6a404c VF |
889 | if (runtime_exceeded(td, &td->ts_cache)) { |
890 | __update_ts_cache(td); | |
891 | if (runtime_exceeded(td, &td->ts_cache)) { | |
ebea2133 | 892 | fio_mark_td_terminate(td); |
2e1df07d JA |
893 | break; |
894 | } | |
895 | } | |
896 | ||
9e684a49 DE |
897 | if (flow_threshold_exceeded(td)) |
898 | continue; | |
899 | ||
f1a32461 JA |
900 | /* |
901 | * Break if we exceeded the bytes. The exception is time | |
902 | * based runs, but we still need to break out of the loop | |
903 | * for those to run verification, if enabled. | |
904 | */ | |
905 | if (bytes_issued >= total_bytes && | |
906 | (!td->o.time_based || | |
907 | (td->o.time_based && td->o.verify != VERIFY_NONE))) | |
20876c53 JC |
908 | break; |
909 | ||
2e1df07d | 910 | io_u = get_io_u(td); |
002fe734 JA |
911 | if (IS_ERR_OR_NULL(io_u)) { |
912 | int err = PTR_ERR(io_u); | |
913 | ||
914 | io_u = NULL; | |
915 | if (err == -EBUSY) { | |
916 | ret = FIO_Q_BUSY; | |
917 | goto reap; | |
918 | } | |
3e260a46 JA |
919 | if (td->o.latency_target) |
920 | goto reap; | |
2e1df07d | 921 | break; |
3e260a46 | 922 | } |
2e1df07d JA |
923 | |
924 | ddir = io_u->ddir; | |
925 | ||
926 | /* | |
82af2a7c JA |
927 | * Add verification end_io handler if: |
928 | * - Asked to verify (!td_rw(td)) | |
929 | * - Or the io_u is from our verify list (mixed write/ver) | |
2e1df07d JA |
930 | */ |
931 | if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ && | |
82af2a7c | 932 | ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) { |
c4b6117b PV |
933 | |
934 | if (!td->o.verify_pattern_bytes) { | |
d6b72507 | 935 | io_u->rand_seed = __rand(&td->verify_state); |
c4b6117b | 936 | if (sizeof(int) != sizeof(long *)) |
d6b72507 | 937 | io_u->rand_seed *= __rand(&td->verify_state); |
c4b6117b PV |
938 | } |
939 | ||
ca09be4b JA |
940 | if (verify_state_should_stop(td, io_u)) { |
941 | put_io_u(td, io_u); | |
942 | break; | |
943 | } | |
944 | ||
2e1df07d JA |
945 | if (td->o.verify_async) |
946 | io_u->end_io = verify_io_u_async; | |
947 | else | |
948 | io_u->end_io = verify_io_u; | |
949 | td_set_runstate(td, TD_VERIFYING); | |
950 | } else if (in_ramp_time(td)) | |
951 | td_set_runstate(td, TD_RAMP); | |
952 | else | |
953 | td_set_runstate(td, TD_RUNNING); | |
954 | ||
9a50c5c5 | 955 | /* |
f9401285 JA |
956 | * Always log IO before it's issued, so we know the specific |
957 | * order of it. The logged unit will track when the IO has | |
958 | * completed. | |
9a50c5c5 | 959 | */ |
c4b6117b PV |
960 | if (td_write(td) && io_u->ddir == DDIR_WRITE && |
961 | td->o.do_verify && | |
962 | td->o.verify != VERIFY_NONE && | |
f9401285 | 963 | !td->o.experimental_verify) |
c4b6117b PV |
964 | log_io_piece(td, io_u); |
965 | ||
a9da8ab2 | 966 | if (td->o.io_submit_mode == IO_MODE_OFFLOAD) { |
0c5df5f9 JA |
967 | const unsigned long blen = io_u->xfer_buflen; |
968 | const enum fio_ddir ddir = acct_ddir(io_u); | |
969 | ||
a9da8ab2 JA |
970 | if (td->error) |
971 | break; | |
0c5df5f9 | 972 | |
26de50cf JA |
973 | workqueue_enqueue(&td->io_wq, &io_u->work); |
974 | ret = FIO_Q_QUEUED; | |
50a8ce86 | 975 | |
26de50cf | 976 | if (ddir_rw(ddir)) { |
0c5df5f9 JA |
977 | td->io_issues[ddir]++; |
978 | td->io_issue_bytes[ddir] += blen; | |
979 | td->rate_io_issue_bytes[ddir] += blen; | |
980 | } | |
981 | ||
50a8ce86 D |
982 | if (should_check_rate(td)) |
983 | td->rate_next_io_time[ddir] = usec_for_io(td, ddir); | |
984 | ||
a9da8ab2 JA |
985 | } else { |
986 | ret = td_io_queue(td, io_u); | |
2e1df07d | 987 | |
50a8ce86 D |
988 | if (should_check_rate(td)) |
989 | td->rate_next_io_time[ddir] = usec_for_io(td, ddir); | |
990 | ||
fd727d9d | 991 | if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time)) |
a9da8ab2 | 992 | break; |
2e1df07d | 993 | |
a9da8ab2 JA |
994 | /* |
995 | * See if we need to complete some commands. Note that | |
996 | * we can get BUSY even without IO queued, if the | |
997 | * system is resource starved. | |
998 | */ | |
3e260a46 | 999 | reap: |
a9da8ab2 JA |
1000 | full = queue_full(td) || |
1001 | (ret == FIO_Q_BUSY && td->cur_depth); | |
82407585 | 1002 | if (full || io_in_polling(td)) |
a9da8ab2 JA |
1003 | ret = wait_for_completions(td, &comp_time); |
1004 | } | |
2e1df07d JA |
1005 | if (ret < 0) |
1006 | break; | |
55312f9f | 1007 | if (!ddir_rw_sum(td->bytes_done) && |
9b87f09b | 1008 | !td_ioengine_flagged(td, FIO_NOIO)) |
2e1df07d JA |
1009 | continue; |
1010 | ||
55312f9f JA |
1011 | if (!in_ramp_time(td) && should_check_rate(td)) { |
1012 | if (check_min_rate(td, &comp_time)) { | |
f9cafb12 | 1013 | if (exitall_on_terminate || td->o.exitall_error) |
2e1df07d JA |
1014 | fio_terminate_threads(td->groupid); |
1015 | td_verror(td, EIO, "check_min_rate"); | |
1016 | break; | |
1017 | } | |
1018 | } | |
3e260a46 JA |
1019 | if (!in_ramp_time(td) && td->o.latency_target) |
1020 | lat_target_check(td); | |
e155cb64 | 1021 | |
2e1df07d JA |
1022 | if (td->o.thinktime) { |
1023 | unsigned long long b; | |
1024 | ||
342f4be4 | 1025 | b = ddir_rw_sum(td->io_blocks); |
2e1df07d JA |
1026 | if (!(b % td->o.thinktime_blocks)) { |
1027 | int left; | |
1028 | ||
002e7183 JA |
1029 | io_u_quiesce(td); |
1030 | ||
2e1df07d JA |
1031 | if (td->o.thinktime_spin) |
1032 | usec_spin(td->o.thinktime_spin); | |
1033 | ||
1034 | left = td->o.thinktime - td->o.thinktime_spin; | |
1035 | if (left) | |
1036 | usec_sleep(td, left); | |
1037 | } | |
1038 | } | |
1039 | } | |
1040 | ||
c97f1ad6 JA |
1041 | check_update_rusage(td); |
1042 | ||
2e1df07d | 1043 | if (td->trim_entries) |
4e0a8fa2 | 1044 | log_err("fio: %lu trim entries leaked?\n", td->trim_entries); |
2e1df07d JA |
1045 | |
1046 | if (td->o.fill_device && td->error == ENOSPC) { | |
1047 | td->error = 0; | |
ebea2133 | 1048 | fio_mark_td_terminate(td); |
2e1df07d JA |
1049 | } |
1050 | if (!td->error) { | |
1051 | struct fio_file *f; | |
1052 | ||
a9da8ab2 JA |
1053 | if (td->o.io_submit_mode == IO_MODE_OFFLOAD) { |
1054 | workqueue_flush(&td->io_wq); | |
1055 | i = 0; | |
1056 | } else | |
1057 | i = td->cur_depth; | |
1058 | ||
2e1df07d | 1059 | if (i) { |
55312f9f | 1060 | ret = io_u_queued_complete(td, i); |
2e1df07d JA |
1061 | if (td->o.fill_device && td->error == ENOSPC) |
1062 | td->error = 0; | |
1063 | } | |
1064 | ||
1065 | if (should_fsync(td) && td->o.end_fsync) { | |
1066 | td_set_runstate(td, TD_FSYNCING); | |
1067 | ||
1068 | for_each_file(td, f, i) { | |
61ee0f86 | 1069 | if (!fio_file_fsync(td, f)) |
2e1df07d | 1070 | continue; |
61ee0f86 JA |
1071 | |
1072 | log_err("fio: end_fsync failed for file %s\n", | |
1073 | f->file_name); | |
2e1df07d JA |
1074 | } |
1075 | } | |
1076 | } else | |
1077 | cleanup_pending_aio(td); | |
1078 | ||
1079 | /* | |
1080 | * stop job if we failed doing any IO | |
1081 | */ | |
342f4be4 | 1082 | if (!ddir_rw_sum(td->this_io_bytes)) |
2e1df07d | 1083 | td->done = 1; |
100f49f1 | 1084 | |
95b03be0 JA |
1085 | for (i = 0; i < DDIR_RWDIR_CNT; i++) |
1086 | bytes_done[i] = td->bytes_done[i] - bytes_done[i]; | |
2e1df07d JA |
1087 | } |
1088 | ||
94a6e1bb JA |
1089 | static void free_file_completion_logging(struct thread_data *td) |
1090 | { | |
1091 | struct fio_file *f; | |
1092 | unsigned int i; | |
1093 | ||
1094 | for_each_file(td, f, i) { | |
1095 | if (!f->last_write_comp) | |
1096 | break; | |
1097 | sfree(f->last_write_comp); | |
1098 | } | |
1099 | } | |
1100 | ||
1101 | static int init_file_completion_logging(struct thread_data *td, | |
1102 | unsigned int depth) | |
1103 | { | |
1104 | struct fio_file *f; | |
1105 | unsigned int i; | |
1106 | ||
1107 | if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save) | |
1108 | return 0; | |
1109 | ||
1110 | for_each_file(td, f, i) { | |
1111 | f->last_write_comp = scalloc(depth, sizeof(uint64_t)); | |
1112 | if (!f->last_write_comp) | |
1113 | goto cleanup; | |
1114 | } | |
1115 | ||
1116 | return 0; | |
1117 | ||
1118 | cleanup: | |
1119 | free_file_completion_logging(td); | |
1120 | log_err("fio: failed to alloc write comp data\n"); | |
1121 | return 1; | |
1122 | } | |
1123 | ||
2e1df07d JA |
1124 | static void cleanup_io_u(struct thread_data *td) |
1125 | { | |
2e1df07d JA |
1126 | struct io_u *io_u; |
1127 | ||
2ae0b204 | 1128 | while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) { |
c73ed246 JA |
1129 | |
1130 | if (td->io_ops->io_u_free) | |
1131 | td->io_ops->io_u_free(td, io_u); | |
1132 | ||
2e1df07d JA |
1133 | fio_memfree(io_u, sizeof(*io_u)); |
1134 | } | |
1135 | ||
1136 | free_io_mem(td); | |
2ae0b204 JA |
1137 | |
1138 | io_u_rexit(&td->io_u_requeues); | |
1139 | io_u_qexit(&td->io_u_freelist); | |
1140 | io_u_qexit(&td->io_u_all); | |
ca09be4b | 1141 | |
94a6e1bb | 1142 | free_file_completion_logging(td); |
2e1df07d JA |
1143 | } |
1144 | ||
1145 | static int init_io_u(struct thread_data *td) | |
1146 | { | |
1147 | struct io_u *io_u; | |
9c42684e | 1148 | unsigned int max_bs, min_write; |
2e1df07d | 1149 | int cl_align, i, max_units; |
2ae0b204 | 1150 | int data_xfer = 1, err; |
2e1df07d JA |
1151 | char *p; |
1152 | ||
1153 | max_units = td->o.iodepth; | |
74f4b020 | 1154 | max_bs = td_max_bs(td); |
9c42684e | 1155 | min_write = td->o.min_bs[DDIR_WRITE]; |
2e1df07d JA |
1156 | td->orig_buffer_size = (unsigned long long) max_bs |
1157 | * (unsigned long long) max_units; | |
1158 | ||
9b87f09b | 1159 | if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td))) |
59d8d0f5 JA |
1160 | data_xfer = 0; |
1161 | ||
2ae0b204 JA |
1162 | err = 0; |
1163 | err += io_u_rinit(&td->io_u_requeues, td->o.iodepth); | |
1164 | err += io_u_qinit(&td->io_u_freelist, td->o.iodepth); | |
1165 | err += io_u_qinit(&td->io_u_all, td->o.iodepth); | |
1166 | ||
1167 | if (err) { | |
1168 | log_err("fio: failed setting up IO queues\n"); | |
1169 | return 1; | |
1170 | } | |
1171 | ||
fd8a09b8 | 1172 | /* |
1173 | * if we may later need to do address alignment, then add any | |
1174 | * possible adjustment here so that we don't cause a buffer | |
1175 | * overflow later. this adjustment may be too much if we get | |
1176 | * lucky and the allocator gives us an aligned address. | |
1177 | */ | |
d01612f3 | 1178 | if (td->o.odirect || td->o.mem_align || td->o.oatomic || |
9b87f09b | 1179 | td_ioengine_flagged(td, FIO_RAWIO)) |
fd8a09b8 | 1180 | td->orig_buffer_size += page_mask + td->o.mem_align; |
1181 | ||
2e1df07d JA |
1182 | if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) { |
1183 | unsigned long bs; | |
1184 | ||
1185 | bs = td->orig_buffer_size + td->o.hugepage_size - 1; | |
1186 | td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1); | |
1187 | } | |
1188 | ||
1189 | if (td->orig_buffer_size != (size_t) td->orig_buffer_size) { | |
1190 | log_err("fio: IO memory too large. Reduce max_bs or iodepth\n"); | |
1191 | return 1; | |
1192 | } | |
1193 | ||
59d8d0f5 | 1194 | if (data_xfer && allocate_io_mem(td)) |
2e1df07d JA |
1195 | return 1; |
1196 | ||
d01612f3 | 1197 | if (td->o.odirect || td->o.mem_align || td->o.oatomic || |
9b87f09b | 1198 | td_ioengine_flagged(td, FIO_RAWIO)) |
248c9436 | 1199 | p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align; |
2e1df07d JA |
1200 | else |
1201 | p = td->orig_buffer; | |
1202 | ||
1203 | cl_align = os_cache_line_size(); | |
1204 | ||
1205 | for (i = 0; i < max_units; i++) { | |
1206 | void *ptr; | |
1207 | ||
1208 | if (td->terminate) | |
1209 | return 1; | |
1210 | ||
1211 | ptr = fio_memalign(cl_align, sizeof(*io_u)); | |
1212 | if (!ptr) { | |
1213 | log_err("fio: unable to allocate aligned memory\n"); | |
1214 | break; | |
1215 | } | |
1216 | ||
1217 | io_u = ptr; | |
1218 | memset(io_u, 0, sizeof(*io_u)); | |
2ae0b204 | 1219 | INIT_FLIST_HEAD(&io_u->verify_list); |
2e1df07d JA |
1220 | dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i); |
1221 | ||
59d8d0f5 | 1222 | if (data_xfer) { |
2e1df07d JA |
1223 | io_u->buf = p; |
1224 | dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf); | |
1225 | ||
1226 | if (td_write(td)) | |
9c42684e | 1227 | io_u_fill_buffer(td, io_u, min_write, max_bs); |
2e1df07d JA |
1228 | if (td_write(td) && td->o.verify_pattern_bytes) { |
1229 | /* | |
1230 | * Fill the buffer with the pattern if we are | |
1231 | * going to be doing writes. | |
1232 | */ | |
ce35b1ec | 1233 | fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0); |
2e1df07d JA |
1234 | } |
1235 | } | |
1236 | ||
1237 | io_u->index = i; | |
1238 | io_u->flags = IO_U_F_FREE; | |
2ae0b204 JA |
1239 | io_u_qpush(&td->io_u_freelist, io_u); |
1240 | ||
1241 | /* | |
1242 | * io_u never leaves this stack, used for iteration of all | |
1243 | * io_u buffers. | |
1244 | */ | |
1245 | io_u_qpush(&td->io_u_all, io_u); | |
c73ed246 JA |
1246 | |
1247 | if (td->io_ops->io_u_init) { | |
1248 | int ret = td->io_ops->io_u_init(td, io_u); | |
1249 | ||
1250 | if (ret) { | |
1251 | log_err("fio: failed to init engine data: %d\n", ret); | |
1252 | return 1; | |
1253 | } | |
1254 | } | |
1255 | ||
2e1df07d JA |
1256 | p += max_bs; |
1257 | } | |
1258 | ||
94a6e1bb JA |
1259 | if (init_file_completion_logging(td, max_units)) |
1260 | return 1; | |
ca09be4b | 1261 | |
2e1df07d JA |
1262 | return 0; |
1263 | } | |
1264 | ||
bd4d9bdc TK |
1265 | /* |
1266 | * This function is Linux specific. | |
1267 | * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux. | |
1268 | */ | |
2e1df07d JA |
1269 | static int switch_ioscheduler(struct thread_data *td) |
1270 | { | |
db815fb3 | 1271 | #ifdef FIO_HAVE_IOSCHED_SWITCH |
2e1df07d JA |
1272 | char tmp[256], tmp2[128]; |
1273 | FILE *f; | |
1274 | int ret; | |
1275 | ||
9b87f09b | 1276 | if (td_ioengine_flagged(td, FIO_DISKLESSIO)) |
2e1df07d JA |
1277 | return 0; |
1278 | ||
bd4d9bdc TK |
1279 | assert(td->files && td->files[0]); |
1280 | sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root); | |
2e1df07d JA |
1281 | |
1282 | f = fopen(tmp, "r+"); | |
1283 | if (!f) { | |
1284 | if (errno == ENOENT) { | |
1285 | log_err("fio: os or kernel doesn't support IO scheduler" | |
1286 | " switching\n"); | |
1287 | return 0; | |
1288 | } | |
1289 | td_verror(td, errno, "fopen iosched"); | |
1290 | return 1; | |
1291 | } | |
1292 | ||
1293 | /* | |
1294 | * Set io scheduler. | |
1295 | */ | |
1296 | ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f); | |
1297 | if (ferror(f) || ret != 1) { | |
1298 | td_verror(td, errno, "fwrite"); | |
1299 | fclose(f); | |
1300 | return 1; | |
1301 | } | |
1302 | ||
1303 | rewind(f); | |
1304 | ||
1305 | /* | |
1306 | * Read back and check that the selected scheduler is now the default. | |
1307 | */ | |
b44b9e45 | 1308 | memset(tmp, 0, sizeof(tmp)); |
49c6f33d | 1309 | ret = fread(tmp, sizeof(tmp), 1, f); |
2e1df07d JA |
1310 | if (ferror(f) || ret < 0) { |
1311 | td_verror(td, errno, "fread"); | |
1312 | fclose(f); | |
1313 | return 1; | |
1314 | } | |
b44b9e45 TK |
1315 | /* |
1316 | * either a list of io schedulers or "none\n" is expected. | |
1317 | */ | |
1318 | tmp[strlen(tmp) - 1] = '\0'; | |
49c6f33d | 1319 | |
67f28a8c TK |
1320 | /* |
1321 | * Write to "none" entry doesn't fail, so check the result here. | |
1322 | */ | |
1323 | if (!strcmp(tmp, "none")) { | |
1324 | log_err("fio: io scheduler is not tunable\n"); | |
1325 | fclose(f); | |
1326 | return 0; | |
1327 | } | |
2e1df07d JA |
1328 | |
1329 | sprintf(tmp2, "[%s]", td->o.ioscheduler); | |
1330 | if (!strstr(tmp, tmp2)) { | |
1331 | log_err("fio: io scheduler %s not found\n", td->o.ioscheduler); | |
1332 | td_verror(td, EINVAL, "iosched_switch"); | |
1333 | fclose(f); | |
1334 | return 1; | |
1335 | } | |
1336 | ||
1337 | fclose(f); | |
1338 | return 0; | |
db815fb3 TK |
1339 | #else |
1340 | return 0; | |
1341 | #endif | |
2e1df07d JA |
1342 | } |
1343 | ||
e39c0676 | 1344 | static bool keep_running(struct thread_data *td) |
2e1df07d | 1345 | { |
77731b29 JA |
1346 | unsigned long long limit; |
1347 | ||
2e1df07d | 1348 | if (td->done) |
e39c0676 | 1349 | return false; |
2e1df07d | 1350 | if (td->o.time_based) |
e39c0676 | 1351 | return true; |
2e1df07d JA |
1352 | if (td->o.loops) { |
1353 | td->o.loops--; | |
e39c0676 | 1354 | return true; |
2e1df07d | 1355 | } |
3939fe85 | 1356 | if (exceeds_number_ios(td)) |
e39c0676 | 1357 | return false; |
26251d8d | 1358 | |
5be9bf09 TK |
1359 | if (td->o.io_size) |
1360 | limit = td->o.io_size; | |
77731b29 JA |
1361 | else |
1362 | limit = td->o.size; | |
1363 | ||
1364 | if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) { | |
5bd5f71a JA |
1365 | uint64_t diff; |
1366 | ||
1367 | /* | |
589d5d4f | 1368 | * If the difference is less than the maximum IO size, we |
5bd5f71a JA |
1369 | * are done. |
1370 | */ | |
77731b29 | 1371 | diff = limit - ddir_rw_sum(td->io_bytes); |
74f4b020 | 1372 | if (diff < td_max_bs(td)) |
e39c0676 | 1373 | return false; |
5bd5f71a | 1374 | |
5be9bf09 | 1375 | if (fio_files_done(td) && !td->o.io_size) |
e39c0676 | 1376 | return false; |
002fe734 | 1377 | |
e39c0676 | 1378 | return true; |
5bd5f71a | 1379 | } |
2e1df07d | 1380 | |
e39c0676 | 1381 | return false; |
2e1df07d JA |
1382 | } |
1383 | ||
ce486495 | 1384 | static int exec_string(struct thread_options *o, const char *string, const char *mode) |
2e1df07d | 1385 | { |
6c9ce469 | 1386 | size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1; |
f491a907 | 1387 | int ret; |
2e1df07d JA |
1388 | char *str; |
1389 | ||
1390 | str = malloc(newlen); | |
ce486495 | 1391 | sprintf(str, "%s &> %s.%s.txt", string, o->name, mode); |
2e1df07d | 1392 | |
ce486495 | 1393 | log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode); |
2e1df07d JA |
1394 | ret = system(str); |
1395 | if (ret == -1) | |
1396 | log_err("fio: exec of cmd <%s> failed\n", str); | |
1397 | ||
1398 | free(str); | |
1399 | return ret; | |
1400 | } | |
1401 | ||
62167762 JC |
1402 | /* |
1403 | * Dry run to compute correct state of numberio for verification. | |
1404 | */ | |
1405 | static uint64_t do_dry_run(struct thread_data *td) | |
1406 | { | |
62167762 JC |
1407 | td_set_runstate(td, TD_RUNNING); |
1408 | ||
1409 | while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || | |
e28dd2cf | 1410 | (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) { |
62167762 JC |
1411 | struct io_u *io_u; |
1412 | int ret; | |
1413 | ||
1414 | if (td->terminate || td->done) | |
1415 | break; | |
1416 | ||
1417 | io_u = get_io_u(td); | |
99955d3d | 1418 | if (IS_ERR_OR_NULL(io_u)) |
62167762 JC |
1419 | break; |
1420 | ||
1651e431 | 1421 | io_u_set(td, io_u, IO_U_F_FLIGHT); |
62167762 JC |
1422 | io_u->error = 0; |
1423 | io_u->resid = 0; | |
1424 | if (ddir_rw(acct_ddir(io_u))) | |
1425 | td->io_issues[acct_ddir(io_u)]++; | |
1426 | if (ddir_rw(io_u->ddir)) { | |
1427 | io_u_mark_depth(td, 1); | |
1428 | td->ts.total_io_u[io_u->ddir]++; | |
1429 | } | |
1430 | ||
2e63e96b PV |
1431 | if (td_write(td) && io_u->ddir == DDIR_WRITE && |
1432 | td->o.do_verify && | |
1433 | td->o.verify != VERIFY_NONE && | |
1434 | !td->o.experimental_verify) | |
1435 | log_io_piece(td, io_u); | |
1436 | ||
55312f9f | 1437 | ret = io_u_sync_complete(td, io_u); |
62167762 JC |
1438 | (void) ret; |
1439 | } | |
1440 | ||
55312f9f | 1441 | return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM]; |
62167762 JC |
1442 | } |
1443 | ||
24660963 JA |
1444 | struct fork_data { |
1445 | struct thread_data *td; | |
1446 | struct sk_out *sk_out; | |
1447 | }; | |
1448 | ||
2e1df07d JA |
1449 | /* |
1450 | * Entry point for the thread based jobs. The process based jobs end up | |
1451 | * here as well, after a little setup. | |
1452 | */ | |
1453 | static void *thread_main(void *data) | |
1454 | { | |
24660963 | 1455 | struct fork_data *fd = data; |
95603b74 | 1456 | unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, }; |
24660963 | 1457 | struct thread_data *td = fd->td; |
4896473e | 1458 | struct thread_options *o = &td->o; |
24660963 | 1459 | struct sk_out *sk_out = fd->sk_out; |
48366f3a | 1460 | uint64_t bytes_done[DDIR_RWDIR_CNT]; |
ca0122d8 | 1461 | int deadlock_loop_cnt; |
2e1df07d | 1462 | int clear_state; |
28727df7 | 1463 | int ret; |
2e1df07d | 1464 | |
24660963 JA |
1465 | sk_out_assign(sk_out); |
1466 | free(fd); | |
1467 | ||
4896473e | 1468 | if (!o->use_thread) { |
2e1df07d JA |
1469 | setsid(); |
1470 | td->pid = getpid(); | |
1471 | } else | |
1472 | td->pid = gettid(); | |
1473 | ||
4896473e | 1474 | fio_local_clock_init(o->use_thread); |
5d879392 | 1475 | |
2e1df07d JA |
1476 | dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid); |
1477 | ||
122c7725 JA |
1478 | if (is_backend) |
1479 | fio_server_send_start(td); | |
1480 | ||
2e1df07d JA |
1481 | INIT_FLIST_HEAD(&td->io_log_list); |
1482 | INIT_FLIST_HEAD(&td->io_hist_list); | |
1483 | INIT_FLIST_HEAD(&td->verify_list); | |
1484 | INIT_FLIST_HEAD(&td->trim_list); | |
1ae83d45 | 1485 | INIT_FLIST_HEAD(&td->next_rand_list); |
2e1df07d JA |
1486 | td->io_hist_tree = RB_ROOT; |
1487 | ||
34febb23 | 1488 | ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond); |
f9e5b5ee | 1489 | if (ret) { |
34febb23 | 1490 | td_verror(td, ret, "mutex_cond_init_pshared"); |
f9e5b5ee JK |
1491 | goto err; |
1492 | } | |
34febb23 | 1493 | ret = cond_init_pshared(&td->verify_cond); |
f9e5b5ee | 1494 | if (ret) { |
34febb23 | 1495 | td_verror(td, ret, "mutex_cond_pshared"); |
f9e5b5ee JK |
1496 | goto err; |
1497 | } | |
2e1df07d JA |
1498 | |
1499 | td_set_runstate(td, TD_INITIALIZED); | |
1500 | dprint(FD_MUTEX, "up startup_mutex\n"); | |
1501 | fio_mutex_up(startup_mutex); | |
1502 | dprint(FD_MUTEX, "wait on td->mutex\n"); | |
1503 | fio_mutex_down(td->mutex); | |
1504 | dprint(FD_MUTEX, "done waiting on td->mutex\n"); | |
1505 | ||
2e1df07d JA |
1506 | /* |
1507 | * A new gid requires privilege, so we need to do this before setting | |
1508 | * the uid. | |
1509 | */ | |
4896473e | 1510 | if (o->gid != -1U && setgid(o->gid)) { |
2e1df07d JA |
1511 | td_verror(td, errno, "setgid"); |
1512 | goto err; | |
1513 | } | |
4896473e | 1514 | if (o->uid != -1U && setuid(o->uid)) { |
2e1df07d JA |
1515 | td_verror(td, errno, "setuid"); |
1516 | goto err; | |
1517 | } | |
1518 | ||
2354d810 JA |
1519 | /* |
1520 | * Do this early, we don't want the compress threads to be limited | |
1521 | * to the same CPUs as the IO workers. So do this before we set | |
1522 | * any potential CPU affinity | |
1523 | */ | |
1524 | if (iolog_compress_init(td, sk_out)) | |
1525 | goto err; | |
1526 | ||
2e1df07d JA |
1527 | /* |
1528 | * If we have a gettimeofday() thread, make sure we exclude that | |
1529 | * thread from this job | |
1530 | */ | |
4896473e JA |
1531 | if (o->gtod_cpu) |
1532 | fio_cpu_clear(&o->cpumask, o->gtod_cpu); | |
2e1df07d JA |
1533 | |
1534 | /* | |
1535 | * Set affinity first, in case it has an impact on the memory | |
1536 | * allocations. | |
1537 | */ | |
b2a9e649 | 1538 | if (fio_option_is_set(o, cpumask)) { |
c2acfbac | 1539 | if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) { |
30cb4c65 | 1540 | ret = fio_cpus_split(&o->cpumask, td->thread_number - 1); |
c2acfbac JA |
1541 | if (!ret) { |
1542 | log_err("fio: no CPUs set\n"); | |
1543 | log_err("fio: Try increasing number of available CPUs\n"); | |
1544 | td_verror(td, EINVAL, "cpus_split"); | |
1545 | goto err; | |
1546 | } | |
1547 | } | |
28727df7 JA |
1548 | ret = fio_setaffinity(td->pid, o->cpumask); |
1549 | if (ret == -1) { | |
4896473e JA |
1550 | td_verror(td, errno, "cpu_set_affinity"); |
1551 | goto err; | |
1552 | } | |
2e1df07d JA |
1553 | } |
1554 | ||
67bf9823 | 1555 | #ifdef CONFIG_LIBNUMA |
d0b937ed | 1556 | /* numa node setup */ |
b2a9e649 JA |
1557 | if (fio_option_is_set(o, numa_cpunodes) || |
1558 | fio_option_is_set(o, numa_memnodes)) { | |
43522848 | 1559 | struct bitmask *mask; |
d0b937ed YR |
1560 | |
1561 | if (numa_available() < 0) { | |
1562 | td_verror(td, errno, "Does not support NUMA API\n"); | |
1563 | goto err; | |
1564 | } | |
1565 | ||
b2a9e649 | 1566 | if (fio_option_is_set(o, numa_cpunodes)) { |
43522848 DG |
1567 | mask = numa_parse_nodestring(o->numa_cpunodes); |
1568 | ret = numa_run_on_node_mask(mask); | |
1569 | numa_free_nodemask(mask); | |
d0b937ed YR |
1570 | if (ret == -1) { |
1571 | td_verror(td, errno, \ | |
1572 | "numa_run_on_node_mask failed\n"); | |
1573 | goto err; | |
1574 | } | |
1575 | } | |
1576 | ||
b2a9e649 | 1577 | if (fio_option_is_set(o, numa_memnodes)) { |
43522848 DG |
1578 | mask = NULL; |
1579 | if (o->numa_memnodes) | |
1580 | mask = numa_parse_nodestring(o->numa_memnodes); | |
1581 | ||
4896473e | 1582 | switch (o->numa_mem_mode) { |
d0b937ed | 1583 | case MPOL_INTERLEAVE: |
43522848 | 1584 | numa_set_interleave_mask(mask); |
d0b937ed YR |
1585 | break; |
1586 | case MPOL_BIND: | |
43522848 | 1587 | numa_set_membind(mask); |
d0b937ed YR |
1588 | break; |
1589 | case MPOL_LOCAL: | |
1590 | numa_set_localalloc(); | |
1591 | break; | |
1592 | case MPOL_PREFERRED: | |
4896473e | 1593 | numa_set_preferred(o->numa_mem_prefer_node); |
d0b937ed YR |
1594 | break; |
1595 | case MPOL_DEFAULT: | |
1596 | default: | |
1597 | break; | |
1598 | } | |
1599 | ||
43522848 DG |
1600 | if (mask) |
1601 | numa_free_nodemask(mask); | |
1602 | ||
d0b937ed YR |
1603 | } |
1604 | } | |
1605 | #endif | |
1606 | ||
9a3f1100 JA |
1607 | if (fio_pin_memory(td)) |
1608 | goto err; | |
1609 | ||
2e1df07d JA |
1610 | /* |
1611 | * May alter parameters that init_io_u() will use, so we need to | |
1612 | * do this first. | |
1613 | */ | |
1614 | if (init_iolog(td)) | |
1615 | goto err; | |
1616 | ||
1617 | if (init_io_u(td)) | |
1618 | goto err; | |
1619 | ||
4896473e | 1620 | if (o->verify_async && verify_async_init(td)) |
2e1df07d JA |
1621 | goto err; |
1622 | ||
27d74836 JA |
1623 | if (fio_option_is_set(o, ioprio) || |
1624 | fio_option_is_set(o, ioprio_class)) { | |
28727df7 JA |
1625 | ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio); |
1626 | if (ret == -1) { | |
2e1df07d JA |
1627 | td_verror(td, errno, "ioprio_set"); |
1628 | goto err; | |
1629 | } | |
1630 | } | |
1631 | ||
4896473e | 1632 | if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt)) |
2e1df07d JA |
1633 | goto err; |
1634 | ||
649c10c5 | 1635 | errno = 0; |
4896473e | 1636 | if (nice(o->nice) == -1 && errno != 0) { |
2e1df07d JA |
1637 | td_verror(td, errno, "nice"); |
1638 | goto err; | |
1639 | } | |
1640 | ||
4896473e | 1641 | if (o->ioscheduler && switch_ioscheduler(td)) |
2e1df07d JA |
1642 | goto err; |
1643 | ||
4896473e | 1644 | if (!o->create_serialize && setup_files(td)) |
2e1df07d JA |
1645 | goto err; |
1646 | ||
1647 | if (td_io_init(td)) | |
1648 | goto err; | |
1649 | ||
1650 | if (init_random_map(td)) | |
1651 | goto err; | |
1652 | ||
ce486495 | 1653 | if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun")) |
4896473e | 1654 | goto err; |
2e1df07d | 1655 | |
4896473e | 1656 | if (o->pre_read) { |
2e1df07d JA |
1657 | if (pre_read_files(td) < 0) |
1658 | goto err; | |
1659 | } | |
1660 | ||
dc5bfbb2 JA |
1661 | fio_verify_init(td); |
1662 | ||
24660963 | 1663 | if (rate_submit_init(td, sk_out)) |
a9da8ab2 JA |
1664 | goto err; |
1665 | ||
3aea75b1 | 1666 | set_epoch_time(td, o->log_unix_epoch); |
44404c5a | 1667 | fio_getrusage(&td->ru_start); |
ac28d905 JA |
1668 | memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch)); |
1669 | memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch)); | |
16e56d25 | 1670 | memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch)); |
ac28d905 JA |
1671 | |
1672 | if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] || | |
1673 | o->ratemin[DDIR_TRIM]) { | |
1674 | memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time, | |
1675 | sizeof(td->bw_sample_time)); | |
1676 | memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time, | |
1677 | sizeof(td->bw_sample_time)); | |
1678 | memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time, | |
1679 | sizeof(td->bw_sample_time)); | |
1680 | } | |
1681 | ||
48366f3a | 1682 | memset(bytes_done, 0, sizeof(bytes_done)); |
2e1df07d | 1683 | clear_state = 0; |
48366f3a | 1684 | |
2e1df07d | 1685 | while (keep_running(td)) { |
100f49f1 JA |
1686 | uint64_t verify_bytes; |
1687 | ||
2e1df07d | 1688 | fio_gettime(&td->start, NULL); |
8b6a404c | 1689 | memcpy(&td->ts_cache, &td->start, sizeof(td->start)); |
2e1df07d | 1690 | |
39c1c323 | 1691 | if (clear_state) { |
ac28d905 | 1692 | clear_io_state(td, 0); |
2e1df07d | 1693 | |
39c1c323 | 1694 | if (o->unlink_each_loop && unlink_all_files(td)) |
1695 | break; | |
1696 | } | |
1697 | ||
2e1df07d JA |
1698 | prune_io_piece_log(td); |
1699 | ||
dcf9844e | 1700 | if (td->o.verify_only && td_write(td)) |
62167762 | 1701 | verify_bytes = do_dry_run(td); |
095196b1 | 1702 | else { |
95b03be0 JA |
1703 | do_io(td, bytes_done); |
1704 | ||
1705 | if (!ddir_rw_sum(bytes_done)) { | |
095196b1 | 1706 | fio_mark_td_terminate(td); |
95b03be0 JA |
1707 | verify_bytes = 0; |
1708 | } else { | |
1709 | verify_bytes = bytes_done[DDIR_WRITE] + | |
1710 | bytes_done[DDIR_TRIM]; | |
1711 | } | |
095196b1 | 1712 | } |
2e1df07d | 1713 | |
49f0b8e0 JA |
1714 | /* |
1715 | * If we took too long to shut down, the main thread could | |
1716 | * already consider us reaped/exited. If that happens, break | |
1717 | * out and clean up. | |
1718 | */ | |
1719 | if (td->runstate >= TD_EXITED) | |
1720 | break; | |
1721 | ||
2e1df07d JA |
1722 | clear_state = 1; |
1723 | ||
40c666c8 JA |
1724 | /* |
1725 | * Make sure we've successfully updated the rusage stats | |
1726 | * before waiting on the stat mutex. Otherwise we could have | |
1727 | * the stat thread holding stat mutex and waiting for | |
1728 | * the rusage_sem, which would never get upped because | |
1729 | * this thread is waiting for the stat mutex. | |
1730 | */ | |
ca0122d8 | 1731 | deadlock_loop_cnt = 0; |
39d13e67 JA |
1732 | do { |
1733 | check_update_rusage(td); | |
1734 | if (!fio_mutex_down_trylock(stat_mutex)) | |
1735 | break; | |
1736 | usleep(1000); | |
ca0122d8 TT |
1737 | if (deadlock_loop_cnt++ > 5000) { |
1738 | log_err("fio seems to be stuck grabbing stat_mutex, forcibly exiting\n"); | |
43f248c4 | 1739 | td->error = EDEADLK; |
ca0122d8 TT |
1740 | goto err; |
1741 | } | |
39d13e67 | 1742 | } while (1); |
40c666c8 | 1743 | |
95603b74 BF |
1744 | if (td_read(td) && td->io_bytes[DDIR_READ]) |
1745 | update_runtime(td, elapsed_us, DDIR_READ); | |
1746 | if (td_write(td) && td->io_bytes[DDIR_WRITE]) | |
1747 | update_runtime(td, elapsed_us, DDIR_WRITE); | |
1748 | if (td_trim(td) && td->io_bytes[DDIR_TRIM]) | |
1749 | update_runtime(td, elapsed_us, DDIR_TRIM); | |
e5437a07 VT |
1750 | fio_gettime(&td->start, NULL); |
1751 | fio_mutex_up(stat_mutex); | |
2e1df07d JA |
1752 | |
1753 | if (td->error || td->terminate) | |
1754 | break; | |
1755 | ||
4896473e JA |
1756 | if (!o->do_verify || |
1757 | o->verify == VERIFY_NONE || | |
9b87f09b | 1758 | td_ioengine_flagged(td, FIO_UNIDIR)) |
2e1df07d JA |
1759 | continue; |
1760 | ||
ac28d905 | 1761 | clear_io_state(td, 0); |
2e1df07d JA |
1762 | |
1763 | fio_gettime(&td->start, NULL); | |
1764 | ||
100f49f1 | 1765 | do_verify(td, verify_bytes); |
2e1df07d | 1766 | |
40c666c8 JA |
1767 | /* |
1768 | * See comment further up for why this is done here. | |
1769 | */ | |
1770 | check_update_rusage(td); | |
1771 | ||
e5437a07 | 1772 | fio_mutex_down(stat_mutex); |
95603b74 | 1773 | update_runtime(td, elapsed_us, DDIR_READ); |
e5437a07 VT |
1774 | fio_gettime(&td->start, NULL); |
1775 | fio_mutex_up(stat_mutex); | |
2e1df07d JA |
1776 | |
1777 | if (td->error || td->terminate) | |
1778 | break; | |
1779 | } | |
1780 | ||
48366f3a TK |
1781 | /* |
1782 | * If td ended up with no I/O when it should have had, | |
1783 | * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO. | |
1784 | * (Are we not missing other flags that can be ignored ?) | |
1785 | */ | |
1786 | if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) && | |
1787 | !(td_ioengine_flagged(td, FIO_NOIO) || | |
1788 | td_ioengine_flagged(td, FIO_DISKLESSIO))) | |
1789 | log_err("%s: No I/O performed by %s, " | |
1790 | "perhaps try --debug=io option for details?\n", | |
1791 | td->o.name, td->io_ops->name); | |
1792 | ||
2cea0b4c JA |
1793 | td_set_runstate(td, TD_FINISHING); |
1794 | ||
2e1df07d | 1795 | update_rusage_stat(td); |
2e1df07d | 1796 | td->ts.total_run_time = mtime_since_now(&td->epoch); |
6eaf09d6 SL |
1797 | td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ]; |
1798 | td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE]; | |
1799 | td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM]; | |
2e1df07d | 1800 | |
ca09be4b | 1801 | if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) && |
d264264a JA |
1802 | (td->o.verify != VERIFY_NONE && td_write(td))) |
1803 | verify_save_state(td->thread_number); | |
ca09be4b | 1804 | |
9a3f1100 JA |
1805 | fio_unpin_memory(td); |
1806 | ||
a47591e4 | 1807 | td_writeout_logs(td, true); |
2e1df07d | 1808 | |
155f2f02 | 1809 | iolog_compress_exit(td); |
103b174e | 1810 | rate_submit_exit(td); |
aee2ab67 | 1811 | |
4896473e | 1812 | if (o->exec_postrun) |
ce486495 | 1813 | exec_string(o, o->exec_postrun, (const char *)"postrun"); |
2e1df07d | 1814 | |
f9cafb12 | 1815 | if (exitall_on_terminate || (o->exitall_error && td->error)) |
2e1df07d JA |
1816 | fio_terminate_threads(td->groupid); |
1817 | ||
1818 | err: | |
1819 | if (td->error) | |
1820 | log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error, | |
1821 | td->verror); | |
1822 | ||
4896473e | 1823 | if (o->verify_async) |
2e1df07d JA |
1824 | verify_async_exit(td); |
1825 | ||
1826 | close_and_free_files(td); | |
2e1df07d | 1827 | cleanup_io_u(td); |
32dbca2c | 1828 | close_ioengine(td); |
2e1df07d | 1829 | cgroup_shutdown(td, &cgroup_mnt); |
ca09be4b | 1830 | verify_free_state(td); |
2e1df07d | 1831 | |
7c4d0fb7 JA |
1832 | if (td->zone_state_index) { |
1833 | int i; | |
1834 | ||
1835 | for (i = 0; i < DDIR_RWDIR_CNT; i++) | |
1836 | free(td->zone_state_index[i]); | |
1837 | free(td->zone_state_index); | |
1838 | td->zone_state_index = NULL; | |
1839 | } | |
1840 | ||
b2a9e649 | 1841 | if (fio_option_is_set(o, cpumask)) { |
8a1db9a1 JA |
1842 | ret = fio_cpuset_exit(&o->cpumask); |
1843 | if (ret) | |
1844 | td_verror(td, ret, "fio_cpuset_exit"); | |
2e1df07d JA |
1845 | } |
1846 | ||
1847 | /* | |
1848 | * do this very late, it will log file closing as well | |
1849 | */ | |
4896473e | 1850 | if (o->write_iolog_file) |
2e1df07d JA |
1851 | write_iolog_close(td); |
1852 | ||
1853 | td_set_runstate(td, TD_EXITED); | |
fda2cfac JA |
1854 | |
1855 | /* | |
1856 | * Do this last after setting our runstate to exited, so we | |
1857 | * know that the stat thread is signaled. | |
1858 | */ | |
1859 | check_update_rusage(td); | |
1860 | ||
24660963 | 1861 | sk_out_drop(); |
e43606c2 | 1862 | return (void *) (uintptr_t) td->error; |
2e1df07d JA |
1863 | } |
1864 | ||
2e1df07d JA |
1865 | /* |
1866 | * Run over the job map and reap the threads that have exited, if any. | |
1867 | */ | |
90eff1c9 SW |
1868 | static void reap_threads(unsigned int *nr_running, uint64_t *t_rate, |
1869 | uint64_t *m_rate) | |
2e1df07d JA |
1870 | { |
1871 | struct thread_data *td; | |
1872 | unsigned int cputhreads, realthreads, pending; | |
1873 | int i, status, ret; | |
1874 | ||
1875 | /* | |
1876 | * reap exited threads (TD_EXITED -> TD_REAPED) | |
1877 | */ | |
1878 | realthreads = pending = cputhreads = 0; | |
1879 | for_each_td(td, i) { | |
1880 | int flags = 0; | |
1881 | ||
1882 | /* | |
1883 | * ->io_ops is NULL for a thread that has closed its | |
1884 | * io engine | |
1885 | */ | |
1886 | if (td->io_ops && !strcmp(td->io_ops->name, "cpuio")) | |
1887 | cputhreads++; | |
1888 | else | |
1889 | realthreads++; | |
1890 | ||
1891 | if (!td->pid) { | |
1892 | pending++; | |
1893 | continue; | |
1894 | } | |
1895 | if (td->runstate == TD_REAPED) | |
1896 | continue; | |
1897 | if (td->o.use_thread) { | |
1898 | if (td->runstate == TD_EXITED) { | |
1899 | td_set_runstate(td, TD_REAPED); | |
1900 | goto reaped; | |
1901 | } | |
1902 | continue; | |
1903 | } | |
1904 | ||
1905 | flags = WNOHANG; | |
1906 | if (td->runstate == TD_EXITED) | |
1907 | flags = 0; | |
1908 | ||
1909 | /* | |
1910 | * check if someone quit or got killed in an unusual way | |
1911 | */ | |
1912 | ret = waitpid(td->pid, &status, flags); | |
1913 | if (ret < 0) { | |
1914 | if (errno == ECHILD) { | |
1915 | log_err("fio: pid=%d disappeared %d\n", | |
1916 | (int) td->pid, td->runstate); | |
a5e371a6 | 1917 | td->sig = ECHILD; |
2e1df07d JA |
1918 | td_set_runstate(td, TD_REAPED); |
1919 | goto reaped; | |
1920 | } | |
1921 | perror("waitpid"); | |
1922 | } else if (ret == td->pid) { | |
1923 | if (WIFSIGNALED(status)) { | |
1924 | int sig = WTERMSIG(status); | |
1925 | ||
36d80bc7 | 1926 | if (sig != SIGTERM && sig != SIGUSR2) |
2e1df07d JA |
1927 | log_err("fio: pid=%d, got signal=%d\n", |
1928 | (int) td->pid, sig); | |
a5e371a6 | 1929 | td->sig = sig; |
2e1df07d JA |
1930 | td_set_runstate(td, TD_REAPED); |
1931 | goto reaped; | |
1932 | } | |
1933 | if (WIFEXITED(status)) { | |
1934 | if (WEXITSTATUS(status) && !td->error) | |
1935 | td->error = WEXITSTATUS(status); | |
1936 | ||
1937 | td_set_runstate(td, TD_REAPED); | |
1938 | goto reaped; | |
1939 | } | |
1940 | } | |
1941 | ||
cba5460c JA |
1942 | /* |
1943 | * If the job is stuck, do a forceful timeout of it and | |
1944 | * move on. | |
1945 | */ | |
1946 | if (td->terminate && | |
2c45a4ac | 1947 | td->runstate < TD_FSYNCING && |
cba5460c | 1948 | time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) { |
86310a1f TK |
1949 | log_err("fio: job '%s' (state=%d) hasn't exited in " |
1950 | "%lu seconds, it appears to be stuck. Doing " | |
1951 | "forceful exit of this job.\n", | |
1952 | td->o.name, td->runstate, | |
1953 | (unsigned long) time_since_now(&td->terminate_time)); | |
cba5460c JA |
1954 | td_set_runstate(td, TD_REAPED); |
1955 | goto reaped; | |
1956 | } | |
1957 | ||
2e1df07d JA |
1958 | /* |
1959 | * thread is not dead, continue | |
1960 | */ | |
1961 | pending++; | |
1962 | continue; | |
1963 | reaped: | |
1964 | (*nr_running)--; | |
342f4be4 JA |
1965 | (*m_rate) -= ddir_rw_sum(td->o.ratemin); |
1966 | (*t_rate) -= ddir_rw_sum(td->o.rate); | |
2e1df07d JA |
1967 | if (!td->pid) |
1968 | pending--; | |
1969 | ||
1970 | if (td->error) | |
1971 | exit_value++; | |
1972 | ||
1973 | done_secs += mtime_since_now(&td->epoch) / 1000; | |
4a88752a | 1974 | profile_td_exit(td); |
2e1df07d JA |
1975 | } |
1976 | ||
1977 | if (*nr_running == cputhreads && !pending && realthreads) | |
1978 | fio_terminate_threads(TERMINATE_ALL); | |
1979 | } | |
1980 | ||
e39c0676 | 1981 | static bool __check_trigger_file(void) |
ca09be4b JA |
1982 | { |
1983 | struct stat sb; | |
1984 | ||
1985 | if (!trigger_file) | |
e39c0676 | 1986 | return false; |
ca09be4b JA |
1987 | |
1988 | if (stat(trigger_file, &sb)) | |
e39c0676 | 1989 | return false; |
ca09be4b JA |
1990 | |
1991 | if (unlink(trigger_file) < 0) | |
1992 | log_err("fio: failed to unlink %s: %s\n", trigger_file, | |
1993 | strerror(errno)); | |
1994 | ||
e39c0676 | 1995 | return true; |
ca09be4b JA |
1996 | } |
1997 | ||
e39c0676 | 1998 | static bool trigger_timedout(void) |
ca09be4b JA |
1999 | { |
2000 | if (trigger_timeout) | |
2001 | return time_since_genesis() >= trigger_timeout; | |
2002 | ||
e39c0676 | 2003 | return false; |
ca09be4b JA |
2004 | } |
2005 | ||
2006 | void exec_trigger(const char *cmd) | |
2007 | { | |
2008 | int ret; | |
2009 | ||
2010 | if (!cmd) | |
2011 | return; | |
2012 | ||
2013 | ret = system(cmd); | |
2014 | if (ret == -1) | |
2015 | log_err("fio: failed executing %s trigger\n", cmd); | |
2016 | } | |
2017 | ||
2018 | void check_trigger_file(void) | |
2019 | { | |
2020 | if (__check_trigger_file() || trigger_timedout()) { | |
796fb3ce JA |
2021 | if (nr_clients) |
2022 | fio_clients_send_trigger(trigger_remote_cmd); | |
2023 | else { | |
d264264a | 2024 | verify_save_state(IO_LIST_ALL); |
ca09be4b JA |
2025 | fio_terminate_threads(TERMINATE_ALL); |
2026 | exec_trigger(trigger_cmd); | |
2027 | } | |
2028 | } | |
2029 | } | |
2030 | ||
2031 | static int fio_verify_load_state(struct thread_data *td) | |
2032 | { | |
2033 | int ret; | |
2034 | ||
2035 | if (!td->o.verify_state) | |
2036 | return 0; | |
2037 | ||
2038 | if (is_backend) { | |
2039 | void *data; | |
2040 | ||
2041 | ret = fio_server_get_verify_state(td->o.name, | |
94a6e1bb | 2042 | td->thread_number - 1, &data); |
ca09be4b | 2043 | if (!ret) |
94a6e1bb | 2044 | verify_assign_state(td, data); |
ca09be4b JA |
2045 | } else |
2046 | ret = verify_load_state(td, "local"); | |
2047 | ||
2048 | return ret; | |
2049 | } | |
2050 | ||
06464907 JA |
2051 | static void do_usleep(unsigned int usecs) |
2052 | { | |
2053 | check_for_running_stats(); | |
ca09be4b | 2054 | check_trigger_file(); |
06464907 JA |
2055 | usleep(usecs); |
2056 | } | |
2057 | ||
e39c0676 | 2058 | static bool check_mount_writes(struct thread_data *td) |
e81ecca3 JA |
2059 | { |
2060 | struct fio_file *f; | |
2061 | unsigned int i; | |
2062 | ||
2063 | if (!td_write(td) || td->o.allow_mounted_write) | |
e39c0676 | 2064 | return false; |
e81ecca3 | 2065 | |
5e81f9c1 TK |
2066 | /* |
2067 | * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs | |
2068 | * are mkfs'd and mounted. | |
2069 | */ | |
e81ecca3 | 2070 | for_each_file(td, f, i) { |
5e81f9c1 TK |
2071 | #ifdef FIO_HAVE_CHARDEV_SIZE |
2072 | if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR) | |
2073 | #else | |
686fbd31 | 2074 | if (f->filetype != FIO_TYPE_BLOCK) |
5e81f9c1 | 2075 | #endif |
e81ecca3 JA |
2076 | continue; |
2077 | if (device_is_mounted(f->file_name)) | |
2078 | goto mounted; | |
2079 | } | |
2080 | ||
e39c0676 | 2081 | return false; |
e81ecca3 | 2082 | mounted: |
c5839011 | 2083 | log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name); |
e39c0676 | 2084 | return true; |
e81ecca3 JA |
2085 | } |
2086 | ||
9cc8cb91 AK |
2087 | static bool waitee_running(struct thread_data *me) |
2088 | { | |
2089 | const char *waitee = me->o.wait_for; | |
2090 | const char *self = me->o.name; | |
2091 | struct thread_data *td; | |
2092 | int i; | |
2093 | ||
2094 | if (!waitee) | |
2095 | return false; | |
2096 | ||
2097 | for_each_td(td, i) { | |
2098 | if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee)) | |
2099 | continue; | |
2100 | ||
2101 | if (td->runstate < TD_EXITED) { | |
2102 | dprint(FD_PROCESS, "%s fenced by %s(%s)\n", | |
2103 | self, td->o.name, | |
2104 | runstate_to_name(td->runstate)); | |
2105 | return true; | |
2106 | } | |
2107 | } | |
2108 | ||
2109 | dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee); | |
2110 | return false; | |
2111 | } | |
2112 | ||
2e1df07d JA |
2113 | /* |
2114 | * Main function for kicking off and reaping jobs, as needed. | |
2115 | */ | |
24660963 | 2116 | static void run_threads(struct sk_out *sk_out) |
2e1df07d JA |
2117 | { |
2118 | struct thread_data *td; | |
90eff1c9 SW |
2119 | unsigned int i, todo, nr_running, nr_started; |
2120 | uint64_t m_rate, t_rate; | |
0de5b26f | 2121 | uint64_t spent; |
2e1df07d | 2122 | |
2e1df07d JA |
2123 | if (fio_gtod_offload && fio_start_gtod_thread()) |
2124 | return; | |
334185e9 | 2125 | |
f2a2ce0e | 2126 | fio_idle_prof_init(); |
2e1df07d JA |
2127 | |
2128 | set_sig_handlers(); | |
2129 | ||
3a5f6bde JA |
2130 | nr_thread = nr_process = 0; |
2131 | for_each_td(td, i) { | |
e81ecca3 JA |
2132 | if (check_mount_writes(td)) |
2133 | return; | |
3a5f6bde JA |
2134 | if (td->o.use_thread) |
2135 | nr_thread++; | |
2136 | else | |
2137 | nr_process++; | |
2138 | } | |
2139 | ||
01cfefcc | 2140 | if (output_format & FIO_OUTPUT_NORMAL) { |
2e1df07d JA |
2141 | log_info("Starting "); |
2142 | if (nr_thread) | |
2143 | log_info("%d thread%s", nr_thread, | |
2144 | nr_thread > 1 ? "s" : ""); | |
2145 | if (nr_process) { | |
2146 | if (nr_thread) | |
2147 | log_info(" and "); | |
2148 | log_info("%d process%s", nr_process, | |
2149 | nr_process > 1 ? "es" : ""); | |
2150 | } | |
2151 | log_info("\n"); | |
e411c301 | 2152 | log_info_flush(); |
2e1df07d JA |
2153 | } |
2154 | ||
2155 | todo = thread_number; | |
2156 | nr_running = 0; | |
2157 | nr_started = 0; | |
2158 | m_rate = t_rate = 0; | |
2159 | ||
2160 | for_each_td(td, i) { | |
2161 | print_status_init(td->thread_number - 1); | |
2162 | ||
2163 | if (!td->o.create_serialize) | |
2164 | continue; | |
2165 | ||
ca09be4b JA |
2166 | if (fio_verify_load_state(td)) |
2167 | goto reap; | |
2168 | ||
2e1df07d JA |
2169 | /* |
2170 | * do file setup here so it happens sequentially, | |
2171 | * we don't want X number of threads getting their | |
2172 | * client data interspersed on disk | |
2173 | */ | |
2174 | if (setup_files(td)) { | |
ca09be4b | 2175 | reap: |
2e1df07d JA |
2176 | exit_value++; |
2177 | if (td->error) | |
2178 | log_err("fio: pid=%d, err=%d/%s\n", | |
2179 | (int) td->pid, td->error, td->verror); | |
2180 | td_set_runstate(td, TD_REAPED); | |
2181 | todo--; | |
2182 | } else { | |
2183 | struct fio_file *f; | |
2184 | unsigned int j; | |
2185 | ||
2186 | /* | |
2187 | * for sharing to work, each job must always open | |
2188 | * its own files. so close them, if we opened them | |
2189 | * for creation | |
2190 | */ | |
2191 | for_each_file(td, f, j) { | |
2192 | if (fio_file_open(f)) | |
2193 | td_io_close_file(td, f); | |
2194 | } | |
2195 | } | |
2196 | } | |
2197 | ||
f2a2ce0e HL |
2198 | /* start idle threads before io threads start to run */ |
2199 | fio_idle_prof_start(); | |
2200 | ||
2e1df07d JA |
2201 | set_genesis_time(); |
2202 | ||
2203 | while (todo) { | |
2204 | struct thread_data *map[REAL_MAX_JOBS]; | |
8b6a404c | 2205 | struct timespec this_start; |
2e1df07d | 2206 | int this_jobs = 0, left; |
ce8b6139 | 2207 | struct fork_data *fd; |
2e1df07d JA |
2208 | |
2209 | /* | |
2210 | * create threads (TD_NOT_CREATED -> TD_CREATED) | |
2211 | */ | |
2212 | for_each_td(td, i) { | |
2213 | if (td->runstate != TD_NOT_CREATED) | |
2214 | continue; | |
2215 | ||
2216 | /* | |
2217 | * never got a chance to start, killed by other | |
2218 | * thread for some reason | |
2219 | */ | |
2220 | if (td->terminate) { | |
2221 | todo--; | |
2222 | continue; | |
2223 | } | |
2224 | ||
2225 | if (td->o.start_delay) { | |
0de5b26f | 2226 | spent = utime_since_genesis(); |
2e1df07d | 2227 | |
74454ce4 | 2228 | if (td->o.start_delay > spent) |
2e1df07d JA |
2229 | continue; |
2230 | } | |
2231 | ||
2232 | if (td->o.stonewall && (nr_started || nr_running)) { | |
2233 | dprint(FD_PROCESS, "%s: stonewall wait\n", | |
2234 | td->o.name); | |
2235 | break; | |
2236 | } | |
2237 | ||
9cc8cb91 AK |
2238 | if (waitee_running(td)) { |
2239 | dprint(FD_PROCESS, "%s: waiting for %s\n", | |
2240 | td->o.name, td->o.wait_for); | |
5c74fc76 | 2241 | continue; |
9cc8cb91 AK |
2242 | } |
2243 | ||
2e1df07d JA |
2244 | init_disk_util(td); |
2245 | ||
c97f1ad6 JA |
2246 | td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED); |
2247 | td->update_rusage = 0; | |
2248 | ||
2e1df07d JA |
2249 | /* |
2250 | * Set state to created. Thread will transition | |
2251 | * to TD_INITIALIZED when it's done setting up. | |
2252 | */ | |
2253 | td_set_runstate(td, TD_CREATED); | |
2254 | map[this_jobs++] = td; | |
2255 | nr_started++; | |
2256 | ||
ce8b6139 JA |
2257 | fd = calloc(1, sizeof(*fd)); |
2258 | fd->td = td; | |
2259 | fd->sk_out = sk_out; | |
2260 | ||
2e1df07d JA |
2261 | if (td->o.use_thread) { |
2262 | int ret; | |
2263 | ||
2264 | dprint(FD_PROCESS, "will pthread_create\n"); | |
2265 | ret = pthread_create(&td->thread, NULL, | |
24660963 | 2266 | thread_main, fd); |
2e1df07d JA |
2267 | if (ret) { |
2268 | log_err("pthread_create: %s\n", | |
2269 | strerror(ret)); | |
24660963 | 2270 | free(fd); |
2e1df07d JA |
2271 | nr_started--; |
2272 | break; | |
2273 | } | |
2274 | ret = pthread_detach(td->thread); | |
2275 | if (ret) | |
2276 | log_err("pthread_detach: %s", | |
2277 | strerror(ret)); | |
2278 | } else { | |
2279 | pid_t pid; | |
2280 | dprint(FD_PROCESS, "will fork\n"); | |
2281 | pid = fork(); | |
2282 | if (!pid) { | |
ce8b6139 | 2283 | int ret; |
2e1df07d | 2284 | |
ce8b6139 | 2285 | ret = (int)(uintptr_t)thread_main(fd); |
2e1df07d JA |
2286 | _exit(ret); |
2287 | } else if (i == fio_debug_jobno) | |
2288 | *fio_debug_jobp = pid; | |
2289 | } | |
2290 | dprint(FD_MUTEX, "wait on startup_mutex\n"); | |
09400a60 | 2291 | if (fio_mutex_down_timeout(startup_mutex, 10000)) { |
2e1df07d JA |
2292 | log_err("fio: job startup hung? exiting.\n"); |
2293 | fio_terminate_threads(TERMINATE_ALL); | |
2294 | fio_abort = 1; | |
2295 | nr_started--; | |
2296 | break; | |
2297 | } | |
2298 | dprint(FD_MUTEX, "done waiting on startup_mutex\n"); | |
2299 | } | |
2300 | ||
2301 | /* | |
2302 | * Wait for the started threads to transition to | |
2303 | * TD_INITIALIZED. | |
2304 | */ | |
2305 | fio_gettime(&this_start, NULL); | |
2306 | left = this_jobs; | |
2307 | while (left && !fio_abort) { | |
2308 | if (mtime_since_now(&this_start) > JOB_START_TIMEOUT) | |
2309 | break; | |
2310 | ||
06464907 | 2311 | do_usleep(100000); |
2e1df07d JA |
2312 | |
2313 | for (i = 0; i < this_jobs; i++) { | |
2314 | td = map[i]; | |
2315 | if (!td) | |
2316 | continue; | |
2317 | if (td->runstate == TD_INITIALIZED) { | |
2318 | map[i] = NULL; | |
2319 | left--; | |
2320 | } else if (td->runstate >= TD_EXITED) { | |
2321 | map[i] = NULL; | |
2322 | left--; | |
2323 | todo--; | |
2324 | nr_running++; /* work-around... */ | |
2325 | } | |
2326 | } | |
2327 | } | |
2328 | ||
2329 | if (left) { | |
4e87c37a JA |
2330 | log_err("fio: %d job%s failed to start\n", left, |
2331 | left > 1 ? "s" : ""); | |
2e1df07d JA |
2332 | for (i = 0; i < this_jobs; i++) { |
2333 | td = map[i]; | |
2334 | if (!td) | |
2335 | continue; | |
2336 | kill(td->pid, SIGTERM); | |
2337 | } | |
2338 | break; | |
2339 | } | |
2340 | ||
2341 | /* | |
2342 | * start created threads (TD_INITIALIZED -> TD_RUNNING). | |
2343 | */ | |
2344 | for_each_td(td, i) { | |
2345 | if (td->runstate != TD_INITIALIZED) | |
2346 | continue; | |
2347 | ||
2348 | if (in_ramp_time(td)) | |
2349 | td_set_runstate(td, TD_RAMP); | |
2350 | else | |
2351 | td_set_runstate(td, TD_RUNNING); | |
2352 | nr_running++; | |
2353 | nr_started--; | |
342f4be4 JA |
2354 | m_rate += ddir_rw_sum(td->o.ratemin); |
2355 | t_rate += ddir_rw_sum(td->o.rate); | |
2e1df07d JA |
2356 | todo--; |
2357 | fio_mutex_up(td->mutex); | |
2358 | } | |
2359 | ||
2360 | reap_threads(&nr_running, &t_rate, &m_rate); | |
2361 | ||
122c7725 | 2362 | if (todo) |
06464907 | 2363 | do_usleep(100000); |
2e1df07d JA |
2364 | } |
2365 | ||
2366 | while (nr_running) { | |
2367 | reap_threads(&nr_running, &t_rate, &m_rate); | |
06464907 | 2368 | do_usleep(10000); |
2e1df07d JA |
2369 | } |
2370 | ||
f2a2ce0e HL |
2371 | fio_idle_prof_stop(); |
2372 | ||
2e1df07d | 2373 | update_io_ticks(); |
2e1df07d JA |
2374 | } |
2375 | ||
27357187 JA |
2376 | static void free_disk_util(void) |
2377 | { | |
27357187 | 2378 | disk_util_prune_entries(); |
a39fb9ea | 2379 | helper_thread_destroy(); |
2e1df07d JA |
2380 | } |
2381 | ||
24660963 | 2382 | int fio_backend(struct sk_out *sk_out) |
2e1df07d JA |
2383 | { |
2384 | struct thread_data *td; | |
2385 | int i; | |
2386 | ||
2387 | if (exec_profile) { | |
2388 | if (load_profile(exec_profile)) | |
2389 | return 1; | |
2390 | free(exec_profile); | |
2391 | exec_profile = NULL; | |
2392 | } | |
2393 | if (!thread_number) | |
2394 | return 0; | |
2395 | ||
2396 | if (write_bw_log) { | |
aee2ab67 JA |
2397 | struct log_params p = { |
2398 | .log_type = IO_LOG_TYPE_BW, | |
2399 | }; | |
2400 | ||
2401 | setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log"); | |
2402 | setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log"); | |
2403 | setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log"); | |
2e1df07d JA |
2404 | } |
2405 | ||
521da527 | 2406 | startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED); |
2e1df07d JA |
2407 | if (startup_mutex == NULL) |
2408 | return 1; | |
2e1df07d JA |
2409 | |
2410 | set_genesis_time(); | |
cef9175e | 2411 | stat_init(); |
a39fb9ea | 2412 | helper_thread_create(startup_mutex, sk_out); |
2e1df07d JA |
2413 | |
2414 | cgroup_list = smalloc(sizeof(*cgroup_list)); | |
2415 | INIT_FLIST_HEAD(cgroup_list); | |
2416 | ||
24660963 | 2417 | run_threads(sk_out); |
2e1df07d | 2418 | |
a39fb9ea | 2419 | helper_thread_exit(); |
8aab824f | 2420 | |
2e1df07d | 2421 | if (!fio_abort) { |
83f7b64e | 2422 | __show_run_stats(); |
2e1df07d | 2423 | if (write_bw_log) { |
cb7e0ace JA |
2424 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
2425 | struct io_log *log = agg_io_log[i]; | |
2426 | ||
60a25727 | 2427 | flush_log(log, false); |
518dac09 | 2428 | free_log(log); |
cb7e0ace | 2429 | } |
2e1df07d JA |
2430 | } |
2431 | } | |
2432 | ||
fda2cfac | 2433 | for_each_td(td, i) { |
ba8fb6f6 VF |
2434 | if (td->ss.dur) { |
2435 | if (td->ss.iops_data != NULL) { | |
2436 | free(td->ss.iops_data); | |
2437 | free(td->ss.bw_data); | |
2438 | } | |
2439 | } | |
2e1df07d | 2440 | fio_options_free(td); |
8049adc1 JA |
2441 | if (td->rusage_sem) { |
2442 | fio_mutex_remove(td->rusage_sem); | |
2443 | td->rusage_sem = NULL; | |
2444 | } | |
5ba01402 JA |
2445 | fio_mutex_remove(td->mutex); |
2446 | td->mutex = NULL; | |
fda2cfac | 2447 | } |
2e1df07d | 2448 | |
a462baef | 2449 | free_disk_util(); |
2e1df07d JA |
2450 | cgroup_kill(cgroup_list); |
2451 | sfree(cgroup_list); | |
2452 | sfree(cgroup_mnt); | |
2453 | ||
2454 | fio_mutex_remove(startup_mutex); | |
cef9175e | 2455 | stat_exit(); |
2e1df07d JA |
2456 | return exit_value; |
2457 | } |