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 | |
fa07eaa6 | 21 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA. |
2e1df07d JA |
22 | * |
23 | */ | |
24 | #include <unistd.h> | |
2e1df07d | 25 | #include <string.h> |
2e1df07d | 26 | #include <signal.h> |
2e1df07d | 27 | #include <assert.h> |
e43606c2 | 28 | #include <inttypes.h> |
2e1df07d JA |
29 | #include <sys/stat.h> |
30 | #include <sys/wait.h> | |
ff6bb260 | 31 | #include <math.h> |
2e1df07d JA |
32 | |
33 | #include "fio.h" | |
2e1df07d JA |
34 | #include "smalloc.h" |
35 | #include "verify.h" | |
2e1df07d JA |
36 | #include "diskutil.h" |
37 | #include "cgroup.h" | |
38 | #include "profile.h" | |
39 | #include "lib/rand.h" | |
f7690c4a | 40 | #include "lib/memalign.h" |
2e1df07d | 41 | #include "server.h" |
44404c5a | 42 | #include "lib/getrusage.h" |
f2a2ce0e | 43 | #include "idletime.h" |
002fe734 | 44 | #include "err.h" |
a9da8ab2 | 45 | #include "workqueue.h" |
e81ecca3 | 46 | #include "lib/mountcheck.h" |
40511301 | 47 | #include "rate-submit.h" |
a39fb9ea | 48 | #include "helper_thread.h" |
ae626d4e | 49 | #include "pshared.h" |
2e1df07d | 50 | |
971caeb1 | 51 | static struct fio_sem *startup_sem; |
2e1df07d JA |
52 | static struct flist_head *cgroup_list; |
53 | static char *cgroup_mnt; | |
54 | static int exit_value; | |
55 | static volatile int fio_abort; | |
3a5f6bde JA |
56 | static unsigned int nr_process = 0; |
57 | static unsigned int nr_thread = 0; | |
2e1df07d | 58 | |
6eaf09d6 | 59 | struct io_log *agg_io_log[DDIR_RWDIR_CNT]; |
2e1df07d | 60 | |
a3efc919 JA |
61 | int groupid = 0; |
62 | unsigned int thread_number = 0; | |
108fea77 | 63 | unsigned int stat_number = 0; |
a3efc919 JA |
64 | int shm_id = 0; |
65 | int temp_stall_ts; | |
66 | unsigned long done_secs = 0; | |
a47591e4 | 67 | |
2e1df07d JA |
68 | #define JOB_START_TIMEOUT (5 * 1000) |
69 | ||
70 | static void sig_int(int sig) | |
71 | { | |
72 | if (threads) { | |
73 | if (is_backend) | |
74 | fio_server_got_signal(sig); | |
75 | else { | |
76 | log_info("\nfio: terminating on signal %d\n", sig); | |
e411c301 | 77 | log_info_flush(); |
2e1df07d JA |
78 | exit_value = 128; |
79 | } | |
80 | ||
81 | fio_terminate_threads(TERMINATE_ALL); | |
82 | } | |
83 | } | |
84 | ||
d2235e56 | 85 | void sig_show_status(int sig) |
4c6d91e8 JA |
86 | { |
87 | show_running_run_stats(); | |
88 | } | |
89 | ||
2e1df07d JA |
90 | static void set_sig_handlers(void) |
91 | { | |
92 | struct sigaction act; | |
93 | ||
94 | memset(&act, 0, sizeof(act)); | |
95 | act.sa_handler = sig_int; | |
96 | act.sa_flags = SA_RESTART; | |
97 | sigaction(SIGINT, &act, NULL); | |
98 | ||
99 | memset(&act, 0, sizeof(act)); | |
100 | act.sa_handler = sig_int; | |
101 | act.sa_flags = SA_RESTART; | |
102 | sigaction(SIGTERM, &act, NULL); | |
103 | ||
2f694507 BC |
104 | /* Windows uses SIGBREAK as a quit signal from other applications */ |
105 | #ifdef WIN32 | |
106 | memset(&act, 0, sizeof(act)); | |
107 | act.sa_handler = sig_int; | |
108 | act.sa_flags = SA_RESTART; | |
109 | sigaction(SIGBREAK, &act, NULL); | |
110 | #endif | |
111 | ||
4c6d91e8 JA |
112 | memset(&act, 0, sizeof(act)); |
113 | act.sa_handler = sig_show_status; | |
114 | act.sa_flags = SA_RESTART; | |
115 | sigaction(SIGUSR1, &act, NULL); | |
116 | ||
2e1df07d JA |
117 | if (is_backend) { |
118 | memset(&act, 0, sizeof(act)); | |
119 | act.sa_handler = sig_int; | |
120 | act.sa_flags = SA_RESTART; | |
121 | sigaction(SIGPIPE, &act, NULL); | |
122 | } | |
123 | } | |
124 | ||
125 | /* | |
126 | * Check if we are above the minimum rate given. | |
127 | */ | |
8b6a404c | 128 | static bool __check_min_rate(struct thread_data *td, struct timespec *now, |
e39c0676 | 129 | enum fio_ddir ddir) |
2e1df07d JA |
130 | { |
131 | unsigned long long bytes = 0; | |
132 | unsigned long iops = 0; | |
133 | unsigned long spent; | |
134 | unsigned long rate; | |
135 | unsigned int ratemin = 0; | |
136 | unsigned int rate_iops = 0; | |
137 | unsigned int rate_iops_min = 0; | |
138 | ||
139 | assert(ddir_rw(ddir)); | |
140 | ||
141 | if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir]) | |
e39c0676 | 142 | return false; |
2e1df07d JA |
143 | |
144 | /* | |
145 | * allow a 2 second settle period in the beginning | |
146 | */ | |
147 | if (mtime_since(&td->start, now) < 2000) | |
e39c0676 | 148 | return false; |
2e1df07d JA |
149 | |
150 | iops += td->this_io_blocks[ddir]; | |
151 | bytes += td->this_io_bytes[ddir]; | |
152 | ratemin += td->o.ratemin[ddir]; | |
153 | rate_iops += td->o.rate_iops[ddir]; | |
154 | rate_iops_min += td->o.rate_iops_min[ddir]; | |
155 | ||
156 | /* | |
157 | * if rate blocks is set, sample is running | |
158 | */ | |
159 | if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) { | |
160 | spent = mtime_since(&td->lastrate[ddir], now); | |
161 | if (spent < td->o.ratecycle) | |
e39c0676 | 162 | return false; |
2e1df07d | 163 | |
20e37e4a | 164 | if (td->o.rate[ddir] || td->o.ratemin[ddir]) { |
2e1df07d JA |
165 | /* |
166 | * check bandwidth specified rate | |
167 | */ | |
168 | if (bytes < td->rate_bytes[ddir]) { | |
420b104a RE |
169 | log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n", |
170 | td->o.name, ratemin, bytes); | |
e39c0676 | 171 | return true; |
2e1df07d | 172 | } else { |
49cba9b3 JA |
173 | if (spent) |
174 | rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent; | |
175 | else | |
176 | rate = 0; | |
177 | ||
2e1df07d JA |
178 | if (rate < ratemin || |
179 | bytes < td->rate_bytes[ddir]) { | |
420b104a RE |
180 | log_err("%s: rate_min=%uB/s not met, got %luB/s\n", |
181 | td->o.name, ratemin, rate); | |
e39c0676 | 182 | return true; |
2e1df07d JA |
183 | } |
184 | } | |
185 | } else { | |
186 | /* | |
187 | * checks iops specified rate | |
188 | */ | |
189 | if (iops < rate_iops) { | |
420b104a RE |
190 | log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n", |
191 | td->o.name, rate_iops, iops); | |
e39c0676 | 192 | return true; |
2e1df07d | 193 | } else { |
4c707a3b JA |
194 | if (spent) |
195 | rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent; | |
196 | else | |
197 | rate = 0; | |
198 | ||
2e1df07d JA |
199 | if (rate < rate_iops_min || |
200 | iops < td->rate_blocks[ddir]) { | |
420b104a RE |
201 | log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n", |
202 | td->o.name, rate_iops_min, rate); | |
e39c0676 | 203 | return true; |
2e1df07d JA |
204 | } |
205 | } | |
206 | } | |
207 | } | |
208 | ||
209 | td->rate_bytes[ddir] = bytes; | |
210 | td->rate_blocks[ddir] = iops; | |
211 | memcpy(&td->lastrate[ddir], now, sizeof(*now)); | |
e39c0676 | 212 | return false; |
2e1df07d JA |
213 | } |
214 | ||
8b6a404c | 215 | static bool check_min_rate(struct thread_data *td, struct timespec *now) |
2e1df07d | 216 | { |
e39c0676 | 217 | bool ret = false; |
2e1df07d | 218 | |
55312f9f | 219 | if (td->bytes_done[DDIR_READ]) |
6eaf09d6 | 220 | ret |= __check_min_rate(td, now, DDIR_READ); |
55312f9f | 221 | if (td->bytes_done[DDIR_WRITE]) |
6eaf09d6 | 222 | ret |= __check_min_rate(td, now, DDIR_WRITE); |
55312f9f | 223 | if (td->bytes_done[DDIR_TRIM]) |
6eaf09d6 | 224 | ret |= __check_min_rate(td, now, DDIR_TRIM); |
2e1df07d JA |
225 | |
226 | return ret; | |
227 | } | |
228 | ||
229 | /* | |
230 | * When job exits, we can cancel the in-flight IO if we are using async | |
231 | * io. Attempt to do so. | |
232 | */ | |
233 | static void cleanup_pending_aio(struct thread_data *td) | |
234 | { | |
2e1df07d JA |
235 | int r; |
236 | ||
237 | /* | |
238 | * get immediately available events, if any | |
239 | */ | |
55312f9f | 240 | r = io_u_queued_complete(td, 0); |
2e1df07d JA |
241 | if (r < 0) |
242 | return; | |
243 | ||
244 | /* | |
245 | * now cancel remaining active events | |
246 | */ | |
247 | if (td->io_ops->cancel) { | |
2ae0b204 JA |
248 | struct io_u *io_u; |
249 | int i; | |
2e1df07d | 250 | |
2ae0b204 JA |
251 | io_u_qiter(&td->io_u_all, io_u, i) { |
252 | if (io_u->flags & IO_U_F_FLIGHT) { | |
2e1df07d JA |
253 | r = td->io_ops->cancel(td, io_u); |
254 | if (!r) | |
255 | put_io_u(td, io_u); | |
256 | } | |
257 | } | |
258 | } | |
259 | ||
260 | if (td->cur_depth) | |
55312f9f | 261 | r = io_u_queued_complete(td, td->cur_depth); |
2e1df07d JA |
262 | } |
263 | ||
264 | /* | |
265 | * Helper to handle the final sync of a file. Works just like the normal | |
266 | * io path, just does everything sync. | |
267 | */ | |
e39c0676 | 268 | static bool fio_io_sync(struct thread_data *td, struct fio_file *f) |
2e1df07d JA |
269 | { |
270 | struct io_u *io_u = __get_io_u(td); | |
271 | int ret; | |
272 | ||
273 | if (!io_u) | |
e39c0676 | 274 | return true; |
2e1df07d JA |
275 | |
276 | io_u->ddir = DDIR_SYNC; | |
277 | io_u->file = f; | |
278 | ||
279 | if (td_io_prep(td, io_u)) { | |
280 | put_io_u(td, io_u); | |
e39c0676 | 281 | return true; |
2e1df07d JA |
282 | } |
283 | ||
284 | requeue: | |
285 | ret = td_io_queue(td, io_u); | |
286 | if (ret < 0) { | |
287 | td_verror(td, io_u->error, "td_io_queue"); | |
288 | put_io_u(td, io_u); | |
e39c0676 | 289 | return true; |
2e1df07d | 290 | } else if (ret == FIO_Q_QUEUED) { |
42ff945c JA |
291 | if (td_io_commit(td)) |
292 | return true; | |
55312f9f | 293 | if (io_u_queued_complete(td, 1) < 0) |
e39c0676 | 294 | return true; |
2e1df07d JA |
295 | } else if (ret == FIO_Q_COMPLETED) { |
296 | if (io_u->error) { | |
297 | td_verror(td, io_u->error, "td_io_queue"); | |
e39c0676 | 298 | return true; |
2e1df07d JA |
299 | } |
300 | ||
55312f9f | 301 | if (io_u_sync_complete(td, io_u) < 0) |
e39c0676 | 302 | return true; |
2e1df07d JA |
303 | } else if (ret == FIO_Q_BUSY) { |
304 | if (td_io_commit(td)) | |
e39c0676 | 305 | return true; |
2e1df07d JA |
306 | goto requeue; |
307 | } | |
308 | ||
e39c0676 | 309 | return false; |
2e1df07d | 310 | } |
a3efc919 | 311 | |
61ee0f86 JA |
312 | static int fio_file_fsync(struct thread_data *td, struct fio_file *f) |
313 | { | |
314 | int ret; | |
315 | ||
316 | if (fio_file_open(f)) | |
317 | return fio_io_sync(td, f); | |
318 | ||
319 | if (td_io_open_file(td, f)) | |
320 | return 1; | |
321 | ||
322 | ret = fio_io_sync(td, f); | |
323 | td_io_close_file(td, f); | |
324 | return ret; | |
325 | } | |
326 | ||
8b6a404c | 327 | static inline void __update_ts_cache(struct thread_data *td) |
2e1df07d | 328 | { |
8b6a404c | 329 | fio_gettime(&td->ts_cache, NULL); |
2e1df07d JA |
330 | } |
331 | ||
8b6a404c | 332 | static inline void update_ts_cache(struct thread_data *td) |
2e1df07d | 333 | { |
8b6a404c VF |
334 | if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask) |
335 | __update_ts_cache(td); | |
2e1df07d JA |
336 | } |
337 | ||
8b6a404c | 338 | static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t) |
2e1df07d JA |
339 | { |
340 | if (in_ramp_time(td)) | |
e39c0676 | 341 | return false; |
2e1df07d | 342 | if (!td->o.timeout) |
e39c0676 | 343 | return false; |
0de5b26f | 344 | if (utime_since(&td->epoch, t) >= td->o.timeout) |
e39c0676 | 345 | return true; |
2e1df07d | 346 | |
e39c0676 | 347 | return false; |
2e1df07d JA |
348 | } |
349 | ||
95603b74 BF |
350 | /* |
351 | * We need to update the runtime consistently in ms, but keep a running | |
352 | * tally of the current elapsed time in microseconds for sub millisecond | |
353 | * updates. | |
354 | */ | |
355 | static inline void update_runtime(struct thread_data *td, | |
356 | unsigned long long *elapsed_us, | |
357 | const enum fio_ddir ddir) | |
358 | { | |
d5c8ea29 JA |
359 | if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only) |
360 | return; | |
361 | ||
95603b74 BF |
362 | td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000; |
363 | elapsed_us[ddir] += utime_since_now(&td->start); | |
364 | td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000; | |
365 | } | |
366 | ||
e39c0676 JA |
367 | static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir, |
368 | int *retptr) | |
2e1df07d JA |
369 | { |
370 | int ret = *retptr; | |
371 | ||
372 | if (ret < 0 || td->error) { | |
8b28bd41 DM |
373 | int err = td->error; |
374 | enum error_type_bit eb; | |
2e1df07d JA |
375 | |
376 | if (ret < 0) | |
377 | err = -ret; | |
2e1df07d | 378 | |
8b28bd41 DM |
379 | eb = td_error_type(ddir, err); |
380 | if (!(td->o.continue_on_error & (1 << eb))) | |
e39c0676 | 381 | return true; |
2e1df07d | 382 | |
8b28bd41 | 383 | if (td_non_fatal_error(td, eb, err)) { |
2e1df07d JA |
384 | /* |
385 | * Continue with the I/Os in case of | |
386 | * a non fatal error. | |
387 | */ | |
388 | update_error_count(td, err); | |
389 | td_clear_error(td); | |
390 | *retptr = 0; | |
e39c0676 | 391 | return false; |
2e1df07d JA |
392 | } else if (td->o.fill_device && err == ENOSPC) { |
393 | /* | |
394 | * We expect to hit this error if | |
395 | * fill_device option is set. | |
396 | */ | |
397 | td_clear_error(td); | |
ebea2133 | 398 | fio_mark_td_terminate(td); |
e39c0676 | 399 | return true; |
2e1df07d JA |
400 | } else { |
401 | /* | |
402 | * Stop the I/O in case of a fatal | |
403 | * error. | |
404 | */ | |
405 | update_error_count(td, err); | |
e39c0676 | 406 | return true; |
2e1df07d JA |
407 | } |
408 | } | |
409 | ||
e39c0676 | 410 | return false; |
2e1df07d JA |
411 | } |
412 | ||
c97f1ad6 JA |
413 | static void check_update_rusage(struct thread_data *td) |
414 | { | |
415 | if (td->update_rusage) { | |
416 | td->update_rusage = 0; | |
417 | update_rusage_stat(td); | |
971caeb1 | 418 | fio_sem_up(td->rusage_sem); |
c97f1ad6 JA |
419 | } |
420 | } | |
421 | ||
8b6a404c | 422 | static int wait_for_completions(struct thread_data *td, struct timespec *time) |
69fea81e JA |
423 | { |
424 | const int full = queue_full(td); | |
425 | int min_evts = 0; | |
426 | int ret; | |
427 | ||
6be06c46 JA |
428 | if (td->flags & TD_F_REGROW_LOGS) |
429 | return io_u_quiesce(td); | |
1fed2080 | 430 | |
69fea81e JA |
431 | /* |
432 | * if the queue is full, we MUST reap at least 1 event | |
433 | */ | |
82407585 | 434 | min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth); |
4b157ac6 | 435 | if ((full && !min_evts) || !td->o.iodepth_batch_complete_min) |
69fea81e JA |
436 | min_evts = 1; |
437 | ||
66b6c5ef | 438 | if (time && (__should_check_rate(td, DDIR_READ) || |
69fea81e JA |
439 | __should_check_rate(td, DDIR_WRITE) || |
440 | __should_check_rate(td, DDIR_TRIM))) | |
441 | fio_gettime(time, NULL); | |
442 | ||
443 | do { | |
55312f9f | 444 | ret = io_u_queued_complete(td, min_evts); |
69fea81e JA |
445 | if (ret < 0) |
446 | break; | |
447 | } while (full && (td->cur_depth > td->o.iodepth_low)); | |
448 | ||
449 | return ret; | |
450 | } | |
451 | ||
e9d512d8 JA |
452 | int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret, |
453 | enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify, | |
8b6a404c | 454 | struct timespec *comp_time) |
e9d512d8 JA |
455 | { |
456 | int ret2; | |
457 | ||
458 | switch (*ret) { | |
459 | case FIO_Q_COMPLETED: | |
460 | if (io_u->error) { | |
461 | *ret = -io_u->error; | |
462 | clear_io_u(td, io_u); | |
463 | } else if (io_u->resid) { | |
464 | int bytes = io_u->xfer_buflen - io_u->resid; | |
465 | struct fio_file *f = io_u->file; | |
466 | ||
467 | if (bytes_issued) | |
468 | *bytes_issued += bytes; | |
469 | ||
470 | if (!from_verify) | |
471 | trim_io_piece(td, io_u); | |
472 | ||
473 | /* | |
474 | * zero read, fail | |
475 | */ | |
476 | if (!bytes) { | |
477 | if (!from_verify) | |
478 | unlog_io_piece(td, io_u); | |
479 | td_verror(td, EIO, "full resid"); | |
480 | put_io_u(td, io_u); | |
481 | break; | |
482 | } | |
483 | ||
484 | io_u->xfer_buflen = io_u->resid; | |
485 | io_u->xfer_buf += bytes; | |
486 | io_u->offset += bytes; | |
487 | ||
488 | if (ddir_rw(io_u->ddir)) | |
489 | td->ts.short_io_u[io_u->ddir]++; | |
490 | ||
e9d512d8 JA |
491 | if (io_u->offset == f->real_file_size) |
492 | goto sync_done; | |
493 | ||
494 | requeue_io_u(td, &io_u); | |
495 | } else { | |
496 | sync_done: | |
497 | if (comp_time && (__should_check_rate(td, DDIR_READ) || | |
498 | __should_check_rate(td, DDIR_WRITE) || | |
499 | __should_check_rate(td, DDIR_TRIM))) | |
500 | fio_gettime(comp_time, NULL); | |
501 | ||
502 | *ret = io_u_sync_complete(td, io_u); | |
503 | if (*ret < 0) | |
504 | break; | |
505 | } | |
a0e3e5a6 | 506 | |
cf2c8d52 JA |
507 | if (td->flags & TD_F_REGROW_LOGS) |
508 | regrow_logs(td); | |
509 | ||
a0e3e5a6 EW |
510 | /* |
511 | * when doing I/O (not when verifying), | |
512 | * check for any errors that are to be ignored | |
513 | */ | |
514 | if (!from_verify) | |
515 | break; | |
516 | ||
e9d512d8 JA |
517 | return 0; |
518 | case FIO_Q_QUEUED: | |
519 | /* | |
520 | * if the engine doesn't have a commit hook, | |
521 | * the io_u is really queued. if it does have such | |
522 | * a hook, it has to call io_u_queued() itself. | |
523 | */ | |
524 | if (td->io_ops->commit == NULL) | |
525 | io_u_queued(td, io_u); | |
526 | if (bytes_issued) | |
527 | *bytes_issued += io_u->xfer_buflen; | |
528 | break; | |
529 | case FIO_Q_BUSY: | |
530 | if (!from_verify) | |
531 | unlog_io_piece(td, io_u); | |
532 | requeue_io_u(td, &io_u); | |
533 | ret2 = td_io_commit(td); | |
534 | if (ret2 < 0) | |
535 | *ret = ret2; | |
536 | break; | |
537 | default: | |
b0775325 | 538 | assert(*ret < 0); |
e9d512d8 JA |
539 | td_verror(td, -(*ret), "td_io_queue"); |
540 | break; | |
541 | } | |
542 | ||
543 | if (break_on_this_error(td, ddir, ret)) | |
544 | return 1; | |
545 | ||
546 | return 0; | |
547 | } | |
548 | ||
e39c0676 | 549 | static inline bool io_in_polling(struct thread_data *td) |
82407585 RP |
550 | { |
551 | return !td->o.iodepth_batch_complete_min && | |
552 | !td->o.iodepth_batch_complete_max; | |
553 | } | |
39c1c323 | 554 | /* |
555 | * Unlinks files from thread data fio_file structure | |
556 | */ | |
557 | static int unlink_all_files(struct thread_data *td) | |
558 | { | |
559 | struct fio_file *f; | |
560 | unsigned int i; | |
561 | int ret = 0; | |
562 | ||
563 | for_each_file(td, f, i) { | |
564 | if (f->filetype != FIO_TYPE_FILE) | |
565 | continue; | |
566 | ret = td_io_unlink_file(td, f); | |
567 | if (ret) | |
568 | break; | |
569 | } | |
570 | ||
571 | if (ret) | |
572 | td_verror(td, ret, "unlink_all_files"); | |
573 | ||
574 | return ret; | |
575 | } | |
82407585 | 576 | |
997b5680 SW |
577 | /* |
578 | * Check if io_u will overlap an in-flight IO in the queue | |
579 | */ | |
580 | static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u) | |
581 | { | |
582 | bool overlap; | |
583 | struct io_u *check_io_u; | |
584 | unsigned long long x1, x2, y1, y2; | |
585 | int i; | |
586 | ||
587 | x1 = io_u->offset; | |
588 | x2 = io_u->offset + io_u->buflen; | |
589 | overlap = false; | |
590 | io_u_qiter(q, check_io_u, i) { | |
591 | if (check_io_u->flags & IO_U_F_FLIGHT) { | |
592 | y1 = check_io_u->offset; | |
593 | y2 = check_io_u->offset + check_io_u->buflen; | |
594 | ||
595 | if (x1 < y2 && y1 < x2) { | |
596 | overlap = true; | |
597 | dprint(FD_IO, "in-flight overlap: %llu/%lu, %llu/%lu\n", | |
598 | x1, io_u->buflen, | |
599 | y1, check_io_u->buflen); | |
600 | break; | |
601 | } | |
602 | } | |
603 | } | |
604 | ||
605 | return overlap; | |
606 | } | |
607 | ||
08b2b386 JA |
608 | static int io_u_submit(struct thread_data *td, struct io_u *io_u) |
609 | { | |
610 | /* | |
611 | * Check for overlap if the user asked us to, and we have | |
612 | * at least one IO in flight besides this one. | |
613 | */ | |
614 | if (td->o.serialize_overlap && td->cur_depth > 1 && | |
615 | in_flight_overlap(&td->io_u_all, io_u)) | |
616 | return FIO_Q_BUSY; | |
617 | ||
618 | return td_io_queue(td, io_u); | |
619 | } | |
620 | ||
2e1df07d JA |
621 | /* |
622 | * The main verify engine. Runs over the writes we previously submitted, | |
623 | * reads the blocks back in, and checks the crc/md5 of the data. | |
624 | */ | |
100f49f1 | 625 | static void do_verify(struct thread_data *td, uint64_t verify_bytes) |
2e1df07d JA |
626 | { |
627 | struct fio_file *f; | |
628 | struct io_u *io_u; | |
629 | int ret, min_events; | |
630 | unsigned int i; | |
631 | ||
632 | dprint(FD_VERIFY, "starting loop\n"); | |
633 | ||
634 | /* | |
635 | * sync io first and invalidate cache, to make sure we really | |
636 | * read from disk. | |
637 | */ | |
638 | for_each_file(td, f, i) { | |
639 | if (!fio_file_open(f)) | |
640 | continue; | |
641 | if (fio_io_sync(td, f)) | |
642 | break; | |
643 | if (file_invalidate_cache(td, f)) | |
644 | break; | |
645 | } | |
646 | ||
c97f1ad6 JA |
647 | check_update_rusage(td); |
648 | ||
2e1df07d JA |
649 | if (td->error) |
650 | return; | |
651 | ||
8f380de2 CJ |
652 | /* |
653 | * verify_state needs to be reset before verification | |
654 | * proceeds so that expected random seeds match actual | |
655 | * random seeds in headers. The main loop will reset | |
656 | * all random number generators if randrepeat is set. | |
657 | */ | |
658 | if (!td->o.rand_repeatable) | |
659 | td_fill_verify_state_seed(td); | |
660 | ||
2e1df07d JA |
661 | td_set_runstate(td, TD_VERIFYING); |
662 | ||
663 | io_u = NULL; | |
664 | while (!td->terminate) { | |
fbccf46c | 665 | enum fio_ddir ddir; |
e9d512d8 | 666 | int full; |
2e1df07d | 667 | |
8b6a404c | 668 | update_ts_cache(td); |
c97f1ad6 | 669 | check_update_rusage(td); |
2e1df07d | 670 | |
8b6a404c VF |
671 | if (runtime_exceeded(td, &td->ts_cache)) { |
672 | __update_ts_cache(td); | |
673 | if (runtime_exceeded(td, &td->ts_cache)) { | |
ebea2133 | 674 | fio_mark_td_terminate(td); |
2e1df07d JA |
675 | break; |
676 | } | |
677 | } | |
678 | ||
9e684a49 DE |
679 | if (flow_threshold_exceeded(td)) |
680 | continue; | |
681 | ||
44cbc6da JA |
682 | if (!td->o.experimental_verify) { |
683 | io_u = __get_io_u(td); | |
684 | if (!io_u) | |
685 | break; | |
2e1df07d | 686 | |
44cbc6da JA |
687 | if (get_next_verify(td, io_u)) { |
688 | put_io_u(td, io_u); | |
689 | break; | |
690 | } | |
2e1df07d | 691 | |
44cbc6da JA |
692 | if (td_io_prep(td, io_u)) { |
693 | put_io_u(td, io_u); | |
694 | break; | |
695 | } | |
696 | } else { | |
55312f9f | 697 | if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes) |
100f49f1 JA |
698 | break; |
699 | ||
bcd5abfa | 700 | while ((io_u = get_io_u(td)) != NULL) { |
059b61f2 | 701 | if (IS_ERR_OR_NULL(io_u)) { |
002fe734 JA |
702 | io_u = NULL; |
703 | ret = FIO_Q_BUSY; | |
704 | goto reap; | |
705 | } | |
706 | ||
bcd5abfa JA |
707 | /* |
708 | * We are only interested in the places where | |
709 | * we wrote or trimmed IOs. Turn those into | |
710 | * reads for verification purposes. | |
711 | */ | |
712 | if (io_u->ddir == DDIR_READ) { | |
713 | /* | |
714 | * Pretend we issued it for rwmix | |
715 | * accounting | |
716 | */ | |
717 | td->io_issues[DDIR_READ]++; | |
718 | put_io_u(td, io_u); | |
719 | continue; | |
720 | } else if (io_u->ddir == DDIR_TRIM) { | |
721 | io_u->ddir = DDIR_READ; | |
1651e431 | 722 | io_u_set(td, io_u, IO_U_F_TRIMMED); |
bcd5abfa JA |
723 | break; |
724 | } else if (io_u->ddir == DDIR_WRITE) { | |
725 | io_u->ddir = DDIR_READ; | |
5c5c33c1 | 726 | populate_verify_io_u(td, io_u); |
bcd5abfa JA |
727 | break; |
728 | } else { | |
729 | put_io_u(td, io_u); | |
730 | continue; | |
731 | } | |
732 | } | |
44cbc6da | 733 | |
bcd5abfa | 734 | if (!io_u) |
44cbc6da | 735 | break; |
2e1df07d JA |
736 | } |
737 | ||
ca09be4b JA |
738 | if (verify_state_should_stop(td, io_u)) { |
739 | put_io_u(td, io_u); | |
740 | break; | |
741 | } | |
742 | ||
2e1df07d JA |
743 | if (td->o.verify_async) |
744 | io_u->end_io = verify_io_u_async; | |
745 | else | |
746 | io_u->end_io = verify_io_u; | |
747 | ||
fbccf46c | 748 | ddir = io_u->ddir; |
10adc4a7 GG |
749 | if (!td->o.disable_slat) |
750 | fio_gettime(&io_u->start_time, NULL); | |
fbccf46c | 751 | |
08b2b386 | 752 | ret = io_u_submit(td, io_u); |
2e1df07d | 753 | |
e9d512d8 | 754 | if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL)) |
2e1df07d JA |
755 | break; |
756 | ||
757 | /* | |
758 | * if we can queue more, do so. but check if there are | |
759 | * completed io_u's first. Note that we can get BUSY even | |
760 | * without IO queued, if the system is resource starved. | |
761 | */ | |
002fe734 | 762 | reap: |
2e1df07d | 763 | full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth); |
82407585 | 764 | if (full || io_in_polling(td)) |
55312f9f | 765 | ret = wait_for_completions(td, NULL); |
2e1df07d | 766 | |
2e1df07d JA |
767 | if (ret < 0) |
768 | break; | |
769 | } | |
770 | ||
c97f1ad6 JA |
771 | check_update_rusage(td); |
772 | ||
2e1df07d JA |
773 | if (!td->error) { |
774 | min_events = td->cur_depth; | |
775 | ||
776 | if (min_events) | |
55312f9f | 777 | ret = io_u_queued_complete(td, min_events); |
2e1df07d JA |
778 | } else |
779 | cleanup_pending_aio(td); | |
780 | ||
781 | td_set_runstate(td, TD_RUNNING); | |
782 | ||
783 | dprint(FD_VERIFY, "exiting loop\n"); | |
784 | } | |
785 | ||
e39c0676 | 786 | static bool exceeds_number_ios(struct thread_data *td) |
3939fe85 JA |
787 | { |
788 | unsigned long long number_ios; | |
789 | ||
790 | if (!td->o.number_ios) | |
e39c0676 | 791 | return false; |
3939fe85 | 792 | |
cf8a46a7 | 793 | number_ios = ddir_rw_sum(td->io_blocks); |
3939fe85 JA |
794 | number_ios += td->io_u_queued + td->io_u_in_flight; |
795 | ||
cf8a46a7 | 796 | return number_ios >= (td->o.number_ios * td->loops); |
3939fe85 JA |
797 | } |
798 | ||
926098e0 | 799 | static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes) |
f7078f7b | 800 | { |
77731b29 | 801 | unsigned long long bytes, limit; |
f7078f7b JA |
802 | |
803 | if (td_rw(td)) | |
926098e0 | 804 | bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE]; |
f7078f7b | 805 | else if (td_write(td)) |
926098e0 | 806 | bytes = this_bytes[DDIR_WRITE]; |
6eaf09d6 | 807 | else if (td_read(td)) |
926098e0 | 808 | bytes = this_bytes[DDIR_READ]; |
f7078f7b | 809 | else |
926098e0 | 810 | bytes = this_bytes[DDIR_TRIM]; |
f7078f7b | 811 | |
5be9bf09 TK |
812 | if (td->o.io_size) |
813 | limit = td->o.io_size; | |
77731b29 JA |
814 | else |
815 | limit = td->o.size; | |
816 | ||
cf8a46a7 | 817 | limit *= td->loops; |
77731b29 | 818 | return bytes >= limit || exceeds_number_ios(td); |
f7078f7b JA |
819 | } |
820 | ||
926098e0 | 821 | static bool io_issue_bytes_exceeded(struct thread_data *td) |
e28dd2cf | 822 | { |
926098e0 JA |
823 | return io_bytes_exceeded(td, td->io_issue_bytes); |
824 | } | |
e28dd2cf | 825 | |
926098e0 JA |
826 | static bool io_complete_bytes_exceeded(struct thread_data *td) |
827 | { | |
828 | return io_bytes_exceeded(td, td->this_io_bytes); | |
e28dd2cf JE |
829 | } |
830 | ||
50a8ce86 D |
831 | /* |
832 | * used to calculate the next io time for rate control | |
833 | * | |
834 | */ | |
835 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) | |
836 | { | |
1a9bf814 | 837 | uint64_t bps = td->rate_bps[ddir]; |
50a8ce86 D |
838 | |
839 | assert(!(td->flags & TD_F_CHILD)); | |
ff6bb260 | 840 | |
6de65959 | 841 | if (td->o.rate_process == RATE_PROCESS_POISSON) { |
1a9bf814 JA |
842 | uint64_t val, iops; |
843 | ||
ff6bb260 | 844 | iops = bps / td->o.bs[ddir]; |
a77d139b | 845 | val = (int64_t) (1000000 / iops) * |
6c8611c7 | 846 | -logf(__rand_0_1(&td->poisson_state[ddir])); |
a77d139b | 847 | if (val) { |
98e51d7c JA |
848 | dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n", |
849 | (unsigned long long) 1000000 / val, | |
850 | ddir); | |
a77d139b | 851 | } |
6c8611c7 JA |
852 | td->last_usec[ddir] += val; |
853 | return td->last_usec[ddir]; | |
ff6bb260 | 854 | } else if (bps) { |
1a9bf814 JA |
855 | uint64_t bytes = td->rate_io_issue_bytes[ddir]; |
856 | uint64_t secs = bytes / bps; | |
857 | uint64_t remainder = bytes % bps; | |
858 | ||
50a8ce86 | 859 | return remainder * 1000000 / bps + secs * 1000000; |
e7b24047 JA |
860 | } |
861 | ||
862 | return 0; | |
50a8ce86 D |
863 | } |
864 | ||
1a9bf814 JA |
865 | static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir) |
866 | { | |
867 | unsigned long long b; | |
868 | uint64_t total; | |
869 | int left; | |
870 | ||
871 | b = ddir_rw_sum(td->io_blocks); | |
872 | if (b % td->o.thinktime_blocks) | |
873 | return; | |
874 | ||
875 | io_u_quiesce(td); | |
876 | ||
877 | total = 0; | |
878 | if (td->o.thinktime_spin) | |
879 | total = usec_spin(td->o.thinktime_spin); | |
880 | ||
881 | left = td->o.thinktime - total; | |
882 | if (left) | |
883 | total += usec_sleep(td, left); | |
884 | ||
885 | /* | |
886 | * If we're ignoring thinktime for the rate, add the number of bytes | |
1aa39b0c JA |
887 | * we would have done while sleeping, minus one block to ensure we |
888 | * start issuing immediately after the sleep. | |
1a9bf814 | 889 | */ |
1aa39b0c | 890 | if (total && td->rate_bps[ddir] && td->o.rate_ign_think) { |
f2aedb17 | 891 | uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL; |
62b858f6 JA |
892 | uint64_t bs = td->o.min_bs[ddir]; |
893 | uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir]; | |
f2aedb17 JA |
894 | uint64_t over; |
895 | ||
62b858f6 JA |
896 | if (usperop <= total) |
897 | over = bs; | |
f2aedb17 | 898 | else |
62b858f6 | 899 | over = (usperop - total) / usperop * -bs; |
f2aedb17 JA |
900 | |
901 | td->rate_io_issue_bytes[ddir] += (missed - over); | |
1aa39b0c | 902 | } |
1a9bf814 JA |
903 | } |
904 | ||
2e1df07d JA |
905 | /* |
906 | * Main IO worker function. It retrieves io_u's to process and queues | |
907 | * and reaps them, checking for rate and errors along the way. | |
100f49f1 JA |
908 | * |
909 | * Returns number of bytes written and trimmed. | |
2e1df07d | 910 | */ |
95b03be0 | 911 | static void do_io(struct thread_data *td, uint64_t *bytes_done) |
2e1df07d JA |
912 | { |
913 | unsigned int i; | |
914 | int ret = 0; | |
c2703bf3 | 915 | uint64_t total_bytes, bytes_issued = 0; |
70c68076 | 916 | |
95b03be0 JA |
917 | for (i = 0; i < DDIR_RWDIR_CNT; i++) |
918 | bytes_done[i] = td->bytes_done[i]; | |
2e1df07d JA |
919 | |
920 | if (in_ramp_time(td)) | |
921 | td_set_runstate(td, TD_RAMP); | |
922 | else | |
923 | td_set_runstate(td, TD_RUNNING); | |
924 | ||
3e260a46 JA |
925 | lat_target_init(td); |
926 | ||
1e564979 JE |
927 | total_bytes = td->o.size; |
928 | /* | |
5be9bf09 | 929 | * Allow random overwrite workloads to write up to io_size |
1e564979 JE |
930 | * before starting verification phase as 'size' doesn't apply. |
931 | */ | |
932 | if (td_write(td) && td_random(td) && td->o.norandommap) | |
5be9bf09 | 933 | total_bytes = max(total_bytes, (uint64_t) td->o.io_size); |
78a6469c JA |
934 | /* |
935 | * If verify_backlog is enabled, we'll run the verify in this | |
936 | * handler as well. For that case, we may need up to twice the | |
937 | * amount of bytes. | |
938 | */ | |
78a6469c JA |
939 | if (td->o.verify != VERIFY_NONE && |
940 | (td_write(td) && td->o.verify_backlog)) | |
c2703bf3 JA |
941 | total_bytes += td->o.size; |
942 | ||
82a90686 | 943 | /* In trimwrite mode, each byte is trimmed and then written, so |
0e4dd95c | 944 | * allow total_bytes to be twice as big */ |
82a90686 | 945 | if (td_trimwrite(td)) |
0e4dd95c DE |
946 | total_bytes += td->total_io_size; |
947 | ||
f7078f7b | 948 | while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || |
e28dd2cf | 949 | (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) || |
c04e4661 | 950 | td->o.time_based) { |
8b6a404c | 951 | struct timespec comp_time; |
2e1df07d | 952 | struct io_u *io_u; |
e9d512d8 | 953 | int full; |
2e1df07d JA |
954 | enum fio_ddir ddir; |
955 | ||
c97f1ad6 JA |
956 | check_update_rusage(td); |
957 | ||
7d7803fa | 958 | if (td->terminate || td->done) |
2e1df07d JA |
959 | break; |
960 | ||
8b6a404c | 961 | update_ts_cache(td); |
2e1df07d | 962 | |
8b6a404c VF |
963 | if (runtime_exceeded(td, &td->ts_cache)) { |
964 | __update_ts_cache(td); | |
965 | if (runtime_exceeded(td, &td->ts_cache)) { | |
ebea2133 | 966 | fio_mark_td_terminate(td); |
2e1df07d JA |
967 | break; |
968 | } | |
969 | } | |
970 | ||
9e684a49 DE |
971 | if (flow_threshold_exceeded(td)) |
972 | continue; | |
973 | ||
f1a32461 JA |
974 | /* |
975 | * Break if we exceeded the bytes. The exception is time | |
976 | * based runs, but we still need to break out of the loop | |
977 | * for those to run verification, if enabled. | |
978 | */ | |
979 | if (bytes_issued >= total_bytes && | |
980 | (!td->o.time_based || | |
981 | (td->o.time_based && td->o.verify != VERIFY_NONE))) | |
20876c53 JC |
982 | break; |
983 | ||
2e1df07d | 984 | io_u = get_io_u(td); |
002fe734 JA |
985 | if (IS_ERR_OR_NULL(io_u)) { |
986 | int err = PTR_ERR(io_u); | |
987 | ||
988 | io_u = NULL; | |
1a9bf814 | 989 | ddir = DDIR_INVAL; |
002fe734 JA |
990 | if (err == -EBUSY) { |
991 | ret = FIO_Q_BUSY; | |
992 | goto reap; | |
993 | } | |
3e260a46 JA |
994 | if (td->o.latency_target) |
995 | goto reap; | |
2e1df07d | 996 | break; |
3e260a46 | 997 | } |
2e1df07d | 998 | |
5c5c33c1 BVA |
999 | if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) |
1000 | populate_verify_io_u(td, io_u); | |
1001 | ||
2e1df07d JA |
1002 | ddir = io_u->ddir; |
1003 | ||
1004 | /* | |
82af2a7c JA |
1005 | * Add verification end_io handler if: |
1006 | * - Asked to verify (!td_rw(td)) | |
1007 | * - Or the io_u is from our verify list (mixed write/ver) | |
2e1df07d JA |
1008 | */ |
1009 | if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ && | |
82af2a7c | 1010 | ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) { |
c4b6117b PV |
1011 | |
1012 | if (!td->o.verify_pattern_bytes) { | |
d6b72507 | 1013 | io_u->rand_seed = __rand(&td->verify_state); |
c4b6117b | 1014 | if (sizeof(int) != sizeof(long *)) |
d6b72507 | 1015 | io_u->rand_seed *= __rand(&td->verify_state); |
c4b6117b PV |
1016 | } |
1017 | ||
ca09be4b JA |
1018 | if (verify_state_should_stop(td, io_u)) { |
1019 | put_io_u(td, io_u); | |
1020 | break; | |
1021 | } | |
1022 | ||
2e1df07d JA |
1023 | if (td->o.verify_async) |
1024 | io_u->end_io = verify_io_u_async; | |
1025 | else | |
1026 | io_u->end_io = verify_io_u; | |
1027 | td_set_runstate(td, TD_VERIFYING); | |
1028 | } else if (in_ramp_time(td)) | |
1029 | td_set_runstate(td, TD_RAMP); | |
1030 | else | |
1031 | td_set_runstate(td, TD_RUNNING); | |
1032 | ||
9a50c5c5 | 1033 | /* |
f9401285 JA |
1034 | * Always log IO before it's issued, so we know the specific |
1035 | * order of it. The logged unit will track when the IO has | |
1036 | * completed. | |
9a50c5c5 | 1037 | */ |
c4b6117b PV |
1038 | if (td_write(td) && io_u->ddir == DDIR_WRITE && |
1039 | td->o.do_verify && | |
1040 | td->o.verify != VERIFY_NONE && | |
f9401285 | 1041 | !td->o.experimental_verify) |
c4b6117b PV |
1042 | log_io_piece(td, io_u); |
1043 | ||
a9da8ab2 | 1044 | if (td->o.io_submit_mode == IO_MODE_OFFLOAD) { |
0c5df5f9 JA |
1045 | const unsigned long blen = io_u->xfer_buflen; |
1046 | const enum fio_ddir ddir = acct_ddir(io_u); | |
1047 | ||
a9da8ab2 JA |
1048 | if (td->error) |
1049 | break; | |
0c5df5f9 | 1050 | |
26de50cf JA |
1051 | workqueue_enqueue(&td->io_wq, &io_u->work); |
1052 | ret = FIO_Q_QUEUED; | |
50a8ce86 | 1053 | |
26de50cf | 1054 | if (ddir_rw(ddir)) { |
0c5df5f9 JA |
1055 | td->io_issues[ddir]++; |
1056 | td->io_issue_bytes[ddir] += blen; | |
1057 | td->rate_io_issue_bytes[ddir] += blen; | |
1058 | } | |
1059 | ||
50a8ce86 D |
1060 | if (should_check_rate(td)) |
1061 | td->rate_next_io_time[ddir] = usec_for_io(td, ddir); | |
1062 | ||
a9da8ab2 | 1063 | } else { |
08b2b386 | 1064 | ret = io_u_submit(td, io_u); |
2e1df07d | 1065 | |
50a8ce86 D |
1066 | if (should_check_rate(td)) |
1067 | td->rate_next_io_time[ddir] = usec_for_io(td, ddir); | |
1068 | ||
fd727d9d | 1069 | if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time)) |
a9da8ab2 | 1070 | break; |
2e1df07d | 1071 | |
a9da8ab2 JA |
1072 | /* |
1073 | * See if we need to complete some commands. Note that | |
1074 | * we can get BUSY even without IO queued, if the | |
1075 | * system is resource starved. | |
1076 | */ | |
3e260a46 | 1077 | reap: |
a9da8ab2 JA |
1078 | full = queue_full(td) || |
1079 | (ret == FIO_Q_BUSY && td->cur_depth); | |
82407585 | 1080 | if (full || io_in_polling(td)) |
a9da8ab2 JA |
1081 | ret = wait_for_completions(td, &comp_time); |
1082 | } | |
2e1df07d JA |
1083 | if (ret < 0) |
1084 | break; | |
55312f9f | 1085 | if (!ddir_rw_sum(td->bytes_done) && |
9b87f09b | 1086 | !td_ioengine_flagged(td, FIO_NOIO)) |
2e1df07d JA |
1087 | continue; |
1088 | ||
55312f9f JA |
1089 | if (!in_ramp_time(td) && should_check_rate(td)) { |
1090 | if (check_min_rate(td, &comp_time)) { | |
f9cafb12 | 1091 | if (exitall_on_terminate || td->o.exitall_error) |
2e1df07d JA |
1092 | fio_terminate_threads(td->groupid); |
1093 | td_verror(td, EIO, "check_min_rate"); | |
1094 | break; | |
1095 | } | |
1096 | } | |
3e260a46 JA |
1097 | if (!in_ramp_time(td) && td->o.latency_target) |
1098 | lat_target_check(td); | |
e155cb64 | 1099 | |
1a9bf814 JA |
1100 | if (ddir_rw(ddir) && td->o.thinktime) |
1101 | handle_thinktime(td, ddir); | |
2e1df07d JA |
1102 | } |
1103 | ||
c97f1ad6 JA |
1104 | check_update_rusage(td); |
1105 | ||
2e1df07d | 1106 | if (td->trim_entries) |
4e0a8fa2 | 1107 | log_err("fio: %lu trim entries leaked?\n", td->trim_entries); |
2e1df07d JA |
1108 | |
1109 | if (td->o.fill_device && td->error == ENOSPC) { | |
1110 | td->error = 0; | |
ebea2133 | 1111 | fio_mark_td_terminate(td); |
2e1df07d JA |
1112 | } |
1113 | if (!td->error) { | |
1114 | struct fio_file *f; | |
1115 | ||
a9da8ab2 JA |
1116 | if (td->o.io_submit_mode == IO_MODE_OFFLOAD) { |
1117 | workqueue_flush(&td->io_wq); | |
1118 | i = 0; | |
1119 | } else | |
1120 | i = td->cur_depth; | |
1121 | ||
2e1df07d | 1122 | if (i) { |
55312f9f | 1123 | ret = io_u_queued_complete(td, i); |
2e1df07d JA |
1124 | if (td->o.fill_device && td->error == ENOSPC) |
1125 | td->error = 0; | |
1126 | } | |
1127 | ||
1128 | if (should_fsync(td) && td->o.end_fsync) { | |
1129 | td_set_runstate(td, TD_FSYNCING); | |
1130 | ||
1131 | for_each_file(td, f, i) { | |
61ee0f86 | 1132 | if (!fio_file_fsync(td, f)) |
2e1df07d | 1133 | continue; |
61ee0f86 JA |
1134 | |
1135 | log_err("fio: end_fsync failed for file %s\n", | |
1136 | f->file_name); | |
2e1df07d JA |
1137 | } |
1138 | } | |
1139 | } else | |
1140 | cleanup_pending_aio(td); | |
1141 | ||
1142 | /* | |
1143 | * stop job if we failed doing any IO | |
1144 | */ | |
342f4be4 | 1145 | if (!ddir_rw_sum(td->this_io_bytes)) |
2e1df07d | 1146 | td->done = 1; |
100f49f1 | 1147 | |
95b03be0 JA |
1148 | for (i = 0; i < DDIR_RWDIR_CNT; i++) |
1149 | bytes_done[i] = td->bytes_done[i] - bytes_done[i]; | |
2e1df07d JA |
1150 | } |
1151 | ||
94a6e1bb JA |
1152 | static void free_file_completion_logging(struct thread_data *td) |
1153 | { | |
1154 | struct fio_file *f; | |
1155 | unsigned int i; | |
1156 | ||
1157 | for_each_file(td, f, i) { | |
1158 | if (!f->last_write_comp) | |
1159 | break; | |
1160 | sfree(f->last_write_comp); | |
1161 | } | |
1162 | } | |
1163 | ||
1164 | static int init_file_completion_logging(struct thread_data *td, | |
1165 | unsigned int depth) | |
1166 | { | |
1167 | struct fio_file *f; | |
1168 | unsigned int i; | |
1169 | ||
1170 | if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save) | |
1171 | return 0; | |
1172 | ||
1173 | for_each_file(td, f, i) { | |
1174 | f->last_write_comp = scalloc(depth, sizeof(uint64_t)); | |
1175 | if (!f->last_write_comp) | |
1176 | goto cleanup; | |
1177 | } | |
1178 | ||
1179 | return 0; | |
1180 | ||
1181 | cleanup: | |
1182 | free_file_completion_logging(td); | |
1183 | log_err("fio: failed to alloc write comp data\n"); | |
1184 | return 1; | |
1185 | } | |
1186 | ||
2e1df07d JA |
1187 | static void cleanup_io_u(struct thread_data *td) |
1188 | { | |
2e1df07d JA |
1189 | struct io_u *io_u; |
1190 | ||
2ae0b204 | 1191 | while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) { |
c73ed246 JA |
1192 | |
1193 | if (td->io_ops->io_u_free) | |
1194 | td->io_ops->io_u_free(td, io_u); | |
1195 | ||
2e1df07d JA |
1196 | fio_memfree(io_u, sizeof(*io_u)); |
1197 | } | |
1198 | ||
1199 | free_io_mem(td); | |
2ae0b204 JA |
1200 | |
1201 | io_u_rexit(&td->io_u_requeues); | |
1202 | io_u_qexit(&td->io_u_freelist); | |
1203 | io_u_qexit(&td->io_u_all); | |
ca09be4b | 1204 | |
94a6e1bb | 1205 | free_file_completion_logging(td); |
2e1df07d JA |
1206 | } |
1207 | ||
1208 | static int init_io_u(struct thread_data *td) | |
1209 | { | |
1210 | struct io_u *io_u; | |
9c42684e | 1211 | unsigned int max_bs, min_write; |
2e1df07d | 1212 | int cl_align, i, max_units; |
2ae0b204 | 1213 | int data_xfer = 1, err; |
2e1df07d JA |
1214 | char *p; |
1215 | ||
1216 | max_units = td->o.iodepth; | |
74f4b020 | 1217 | max_bs = td_max_bs(td); |
9c42684e | 1218 | min_write = td->o.min_bs[DDIR_WRITE]; |
2e1df07d JA |
1219 | td->orig_buffer_size = (unsigned long long) max_bs |
1220 | * (unsigned long long) max_units; | |
1221 | ||
9b87f09b | 1222 | if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td))) |
59d8d0f5 JA |
1223 | data_xfer = 0; |
1224 | ||
2ae0b204 | 1225 | err = 0; |
34851ad5 JA |
1226 | err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth); |
1227 | err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth); | |
1228 | err += !io_u_qinit(&td->io_u_all, td->o.iodepth); | |
2ae0b204 JA |
1229 | |
1230 | if (err) { | |
1231 | log_err("fio: failed setting up IO queues\n"); | |
1232 | return 1; | |
1233 | } | |
1234 | ||
fd8a09b8 | 1235 | /* |
1236 | * if we may later need to do address alignment, then add any | |
1237 | * possible adjustment here so that we don't cause a buffer | |
1238 | * overflow later. this adjustment may be too much if we get | |
1239 | * lucky and the allocator gives us an aligned address. | |
1240 | */ | |
d01612f3 | 1241 | if (td->o.odirect || td->o.mem_align || td->o.oatomic || |
9b87f09b | 1242 | td_ioengine_flagged(td, FIO_RAWIO)) |
fd8a09b8 | 1243 | td->orig_buffer_size += page_mask + td->o.mem_align; |
1244 | ||
2e1df07d JA |
1245 | if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) { |
1246 | unsigned long bs; | |
1247 | ||
1248 | bs = td->orig_buffer_size + td->o.hugepage_size - 1; | |
1249 | td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1); | |
1250 | } | |
1251 | ||
1252 | if (td->orig_buffer_size != (size_t) td->orig_buffer_size) { | |
1253 | log_err("fio: IO memory too large. Reduce max_bs or iodepth\n"); | |
1254 | return 1; | |
1255 | } | |
1256 | ||
59d8d0f5 | 1257 | if (data_xfer && allocate_io_mem(td)) |
2e1df07d JA |
1258 | return 1; |
1259 | ||
d01612f3 | 1260 | if (td->o.odirect || td->o.mem_align || td->o.oatomic || |
9b87f09b | 1261 | td_ioengine_flagged(td, FIO_RAWIO)) |
248c9436 | 1262 | p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align; |
2e1df07d JA |
1263 | else |
1264 | p = td->orig_buffer; | |
1265 | ||
1266 | cl_align = os_cache_line_size(); | |
1267 | ||
1268 | for (i = 0; i < max_units; i++) { | |
1269 | void *ptr; | |
1270 | ||
1271 | if (td->terminate) | |
1272 | return 1; | |
1273 | ||
1274 | ptr = fio_memalign(cl_align, sizeof(*io_u)); | |
1275 | if (!ptr) { | |
1276 | log_err("fio: unable to allocate aligned memory\n"); | |
1277 | break; | |
1278 | } | |
1279 | ||
1280 | io_u = ptr; | |
1281 | memset(io_u, 0, sizeof(*io_u)); | |
2ae0b204 | 1282 | INIT_FLIST_HEAD(&io_u->verify_list); |
2e1df07d JA |
1283 | dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i); |
1284 | ||
59d8d0f5 | 1285 | if (data_xfer) { |
2e1df07d JA |
1286 | io_u->buf = p; |
1287 | dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf); | |
1288 | ||
1289 | if (td_write(td)) | |
9c42684e | 1290 | io_u_fill_buffer(td, io_u, min_write, max_bs); |
2e1df07d JA |
1291 | if (td_write(td) && td->o.verify_pattern_bytes) { |
1292 | /* | |
1293 | * Fill the buffer with the pattern if we are | |
1294 | * going to be doing writes. | |
1295 | */ | |
ce35b1ec | 1296 | fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0); |
2e1df07d JA |
1297 | } |
1298 | } | |
1299 | ||
1300 | io_u->index = i; | |
1301 | io_u->flags = IO_U_F_FREE; | |
2ae0b204 JA |
1302 | io_u_qpush(&td->io_u_freelist, io_u); |
1303 | ||
1304 | /* | |
1305 | * io_u never leaves this stack, used for iteration of all | |
1306 | * io_u buffers. | |
1307 | */ | |
1308 | io_u_qpush(&td->io_u_all, io_u); | |
c73ed246 JA |
1309 | |
1310 | if (td->io_ops->io_u_init) { | |
1311 | int ret = td->io_ops->io_u_init(td, io_u); | |
1312 | ||
1313 | if (ret) { | |
1314 | log_err("fio: failed to init engine data: %d\n", ret); | |
1315 | return 1; | |
1316 | } | |
1317 | } | |
1318 | ||
2e1df07d JA |
1319 | p += max_bs; |
1320 | } | |
1321 | ||
94a6e1bb JA |
1322 | if (init_file_completion_logging(td, max_units)) |
1323 | return 1; | |
ca09be4b | 1324 | |
2e1df07d JA |
1325 | return 0; |
1326 | } | |
1327 | ||
bd4d9bdc TK |
1328 | /* |
1329 | * This function is Linux specific. | |
1330 | * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux. | |
1331 | */ | |
2e1df07d JA |
1332 | static int switch_ioscheduler(struct thread_data *td) |
1333 | { | |
db815fb3 | 1334 | #ifdef FIO_HAVE_IOSCHED_SWITCH |
178ef27a | 1335 | char tmp[256], tmp2[128], *p; |
2e1df07d JA |
1336 | FILE *f; |
1337 | int ret; | |
1338 | ||
9b87f09b | 1339 | if (td_ioengine_flagged(td, FIO_DISKLESSIO)) |
2e1df07d JA |
1340 | return 0; |
1341 | ||
bd4d9bdc TK |
1342 | assert(td->files && td->files[0]); |
1343 | sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root); | |
2e1df07d JA |
1344 | |
1345 | f = fopen(tmp, "r+"); | |
1346 | if (!f) { | |
1347 | if (errno == ENOENT) { | |
1348 | log_err("fio: os or kernel doesn't support IO scheduler" | |
1349 | " switching\n"); | |
1350 | return 0; | |
1351 | } | |
1352 | td_verror(td, errno, "fopen iosched"); | |
1353 | return 1; | |
1354 | } | |
1355 | ||
1356 | /* | |
1357 | * Set io scheduler. | |
1358 | */ | |
1359 | ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f); | |
1360 | if (ferror(f) || ret != 1) { | |
1361 | td_verror(td, errno, "fwrite"); | |
1362 | fclose(f); | |
1363 | return 1; | |
1364 | } | |
1365 | ||
1366 | rewind(f); | |
1367 | ||
1368 | /* | |
1369 | * Read back and check that the selected scheduler is now the default. | |
1370 | */ | |
0e23d456 | 1371 | ret = fread(tmp, 1, sizeof(tmp) - 1, f); |
2e1df07d JA |
1372 | if (ferror(f) || ret < 0) { |
1373 | td_verror(td, errno, "fread"); | |
1374 | fclose(f); | |
1375 | return 1; | |
1376 | } | |
0e23d456 | 1377 | tmp[ret] = '\0'; |
b44b9e45 | 1378 | /* |
178ef27a BVA |
1379 | * either a list of io schedulers or "none\n" is expected. Strip the |
1380 | * trailing newline. | |
b44b9e45 | 1381 | */ |
178ef27a BVA |
1382 | p = tmp; |
1383 | strsep(&p, "\n"); | |
49c6f33d | 1384 | |
67f28a8c TK |
1385 | /* |
1386 | * Write to "none" entry doesn't fail, so check the result here. | |
1387 | */ | |
1388 | if (!strcmp(tmp, "none")) { | |
1389 | log_err("fio: io scheduler is not tunable\n"); | |
1390 | fclose(f); | |
1391 | return 0; | |
1392 | } | |
2e1df07d JA |
1393 | |
1394 | sprintf(tmp2, "[%s]", td->o.ioscheduler); | |
1395 | if (!strstr(tmp, tmp2)) { | |
1396 | log_err("fio: io scheduler %s not found\n", td->o.ioscheduler); | |
1397 | td_verror(td, EINVAL, "iosched_switch"); | |
1398 | fclose(f); | |
1399 | return 1; | |
1400 | } | |
1401 | ||
1402 | fclose(f); | |
1403 | return 0; | |
db815fb3 TK |
1404 | #else |
1405 | return 0; | |
1406 | #endif | |
2e1df07d JA |
1407 | } |
1408 | ||
e39c0676 | 1409 | static bool keep_running(struct thread_data *td) |
2e1df07d | 1410 | { |
77731b29 JA |
1411 | unsigned long long limit; |
1412 | ||
2e1df07d | 1413 | if (td->done) |
e39c0676 | 1414 | return false; |
98413be7 SW |
1415 | if (td->terminate) |
1416 | return false; | |
2e1df07d | 1417 | if (td->o.time_based) |
e39c0676 | 1418 | return true; |
2e1df07d JA |
1419 | if (td->o.loops) { |
1420 | td->o.loops--; | |
e39c0676 | 1421 | return true; |
2e1df07d | 1422 | } |
3939fe85 | 1423 | if (exceeds_number_ios(td)) |
e39c0676 | 1424 | return false; |
26251d8d | 1425 | |
5be9bf09 TK |
1426 | if (td->o.io_size) |
1427 | limit = td->o.io_size; | |
77731b29 JA |
1428 | else |
1429 | limit = td->o.size; | |
1430 | ||
1431 | if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) { | |
5bd5f71a JA |
1432 | uint64_t diff; |
1433 | ||
1434 | /* | |
589d5d4f | 1435 | * If the difference is less than the maximum IO size, we |
5bd5f71a JA |
1436 | * are done. |
1437 | */ | |
77731b29 | 1438 | diff = limit - ddir_rw_sum(td->io_bytes); |
74f4b020 | 1439 | if (diff < td_max_bs(td)) |
e39c0676 | 1440 | return false; |
5bd5f71a | 1441 | |
5be9bf09 | 1442 | if (fio_files_done(td) && !td->o.io_size) |
e39c0676 | 1443 | return false; |
002fe734 | 1444 | |
e39c0676 | 1445 | return true; |
5bd5f71a | 1446 | } |
2e1df07d | 1447 | |
e39c0676 | 1448 | return false; |
2e1df07d JA |
1449 | } |
1450 | ||
ce486495 | 1451 | static int exec_string(struct thread_options *o, const char *string, const char *mode) |
2e1df07d | 1452 | { |
6c9ce469 | 1453 | size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1; |
f491a907 | 1454 | int ret; |
2e1df07d JA |
1455 | char *str; |
1456 | ||
1457 | str = malloc(newlen); | |
ce486495 | 1458 | sprintf(str, "%s &> %s.%s.txt", string, o->name, mode); |
2e1df07d | 1459 | |
ce486495 | 1460 | log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode); |
2e1df07d JA |
1461 | ret = system(str); |
1462 | if (ret == -1) | |
1463 | log_err("fio: exec of cmd <%s> failed\n", str); | |
1464 | ||
1465 | free(str); | |
1466 | return ret; | |
1467 | } | |
1468 | ||
62167762 JC |
1469 | /* |
1470 | * Dry run to compute correct state of numberio for verification. | |
1471 | */ | |
1472 | static uint64_t do_dry_run(struct thread_data *td) | |
1473 | { | |
62167762 JC |
1474 | td_set_runstate(td, TD_RUNNING); |
1475 | ||
1476 | while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || | |
e28dd2cf | 1477 | (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) { |
62167762 JC |
1478 | struct io_u *io_u; |
1479 | int ret; | |
1480 | ||
1481 | if (td->terminate || td->done) | |
1482 | break; | |
1483 | ||
1484 | io_u = get_io_u(td); | |
99955d3d | 1485 | if (IS_ERR_OR_NULL(io_u)) |
62167762 JC |
1486 | break; |
1487 | ||
1651e431 | 1488 | io_u_set(td, io_u, IO_U_F_FLIGHT); |
62167762 JC |
1489 | io_u->error = 0; |
1490 | io_u->resid = 0; | |
1491 | if (ddir_rw(acct_ddir(io_u))) | |
1492 | td->io_issues[acct_ddir(io_u)]++; | |
1493 | if (ddir_rw(io_u->ddir)) { | |
1494 | io_u_mark_depth(td, 1); | |
1495 | td->ts.total_io_u[io_u->ddir]++; | |
1496 | } | |
1497 | ||
2e63e96b PV |
1498 | if (td_write(td) && io_u->ddir == DDIR_WRITE && |
1499 | td->o.do_verify && | |
1500 | td->o.verify != VERIFY_NONE && | |
1501 | !td->o.experimental_verify) | |
1502 | log_io_piece(td, io_u); | |
1503 | ||
55312f9f | 1504 | ret = io_u_sync_complete(td, io_u); |
62167762 JC |
1505 | (void) ret; |
1506 | } | |
1507 | ||
55312f9f | 1508 | return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM]; |
62167762 JC |
1509 | } |
1510 | ||
24660963 JA |
1511 | struct fork_data { |
1512 | struct thread_data *td; | |
1513 | struct sk_out *sk_out; | |
1514 | }; | |
1515 | ||
2e1df07d JA |
1516 | /* |
1517 | * Entry point for the thread based jobs. The process based jobs end up | |
1518 | * here as well, after a little setup. | |
1519 | */ | |
1520 | static void *thread_main(void *data) | |
1521 | { | |
24660963 | 1522 | struct fork_data *fd = data; |
95603b74 | 1523 | unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, }; |
24660963 | 1524 | struct thread_data *td = fd->td; |
4896473e | 1525 | struct thread_options *o = &td->o; |
24660963 | 1526 | struct sk_out *sk_out = fd->sk_out; |
48366f3a | 1527 | uint64_t bytes_done[DDIR_RWDIR_CNT]; |
ca0122d8 | 1528 | int deadlock_loop_cnt; |
c16035aa | 1529 | bool clear_state, did_some_io; |
28727df7 | 1530 | int ret; |
2e1df07d | 1531 | |
24660963 JA |
1532 | sk_out_assign(sk_out); |
1533 | free(fd); | |
1534 | ||
4896473e | 1535 | if (!o->use_thread) { |
2e1df07d JA |
1536 | setsid(); |
1537 | td->pid = getpid(); | |
1538 | } else | |
1539 | td->pid = gettid(); | |
1540 | ||
4896473e | 1541 | fio_local_clock_init(o->use_thread); |
5d879392 | 1542 | |
2e1df07d JA |
1543 | dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid); |
1544 | ||
122c7725 JA |
1545 | if (is_backend) |
1546 | fio_server_send_start(td); | |
1547 | ||
2e1df07d JA |
1548 | INIT_FLIST_HEAD(&td->io_log_list); |
1549 | INIT_FLIST_HEAD(&td->io_hist_list); | |
1550 | INIT_FLIST_HEAD(&td->verify_list); | |
1551 | INIT_FLIST_HEAD(&td->trim_list); | |
1ae83d45 | 1552 | INIT_FLIST_HEAD(&td->next_rand_list); |
2e1df07d JA |
1553 | td->io_hist_tree = RB_ROOT; |
1554 | ||
34febb23 | 1555 | ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond); |
f9e5b5ee | 1556 | if (ret) { |
34febb23 | 1557 | td_verror(td, ret, "mutex_cond_init_pshared"); |
f9e5b5ee JK |
1558 | goto err; |
1559 | } | |
34febb23 | 1560 | ret = cond_init_pshared(&td->verify_cond); |
f9e5b5ee | 1561 | if (ret) { |
34febb23 | 1562 | td_verror(td, ret, "mutex_cond_pshared"); |
f9e5b5ee JK |
1563 | goto err; |
1564 | } | |
2e1df07d JA |
1565 | |
1566 | td_set_runstate(td, TD_INITIALIZED); | |
971caeb1 BVA |
1567 | dprint(FD_MUTEX, "up startup_sem\n"); |
1568 | fio_sem_up(startup_sem); | |
1569 | dprint(FD_MUTEX, "wait on td->sem\n"); | |
1570 | fio_sem_down(td->sem); | |
1571 | dprint(FD_MUTEX, "done waiting on td->sem\n"); | |
2e1df07d | 1572 | |
2e1df07d JA |
1573 | /* |
1574 | * A new gid requires privilege, so we need to do this before setting | |
1575 | * the uid. | |
1576 | */ | |
4896473e | 1577 | if (o->gid != -1U && setgid(o->gid)) { |
2e1df07d JA |
1578 | td_verror(td, errno, "setgid"); |
1579 | goto err; | |
1580 | } | |
4896473e | 1581 | if (o->uid != -1U && setuid(o->uid)) { |
2e1df07d JA |
1582 | td_verror(td, errno, "setuid"); |
1583 | goto err; | |
1584 | } | |
1585 | ||
2354d810 JA |
1586 | /* |
1587 | * Do this early, we don't want the compress threads to be limited | |
1588 | * to the same CPUs as the IO workers. So do this before we set | |
1589 | * any potential CPU affinity | |
1590 | */ | |
1591 | if (iolog_compress_init(td, sk_out)) | |
1592 | goto err; | |
1593 | ||
2e1df07d JA |
1594 | /* |
1595 | * If we have a gettimeofday() thread, make sure we exclude that | |
1596 | * thread from this job | |
1597 | */ | |
4896473e JA |
1598 | if (o->gtod_cpu) |
1599 | fio_cpu_clear(&o->cpumask, o->gtod_cpu); | |
2e1df07d JA |
1600 | |
1601 | /* | |
1602 | * Set affinity first, in case it has an impact on the memory | |
1603 | * allocations. | |
1604 | */ | |
b2a9e649 | 1605 | if (fio_option_is_set(o, cpumask)) { |
c2acfbac | 1606 | if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) { |
30cb4c65 | 1607 | ret = fio_cpus_split(&o->cpumask, td->thread_number - 1); |
c2acfbac JA |
1608 | if (!ret) { |
1609 | log_err("fio: no CPUs set\n"); | |
1610 | log_err("fio: Try increasing number of available CPUs\n"); | |
1611 | td_verror(td, EINVAL, "cpus_split"); | |
1612 | goto err; | |
1613 | } | |
1614 | } | |
28727df7 JA |
1615 | ret = fio_setaffinity(td->pid, o->cpumask); |
1616 | if (ret == -1) { | |
4896473e JA |
1617 | td_verror(td, errno, "cpu_set_affinity"); |
1618 | goto err; | |
1619 | } | |
2e1df07d JA |
1620 | } |
1621 | ||
67bf9823 | 1622 | #ifdef CONFIG_LIBNUMA |
d0b937ed | 1623 | /* numa node setup */ |
b2a9e649 JA |
1624 | if (fio_option_is_set(o, numa_cpunodes) || |
1625 | fio_option_is_set(o, numa_memnodes)) { | |
43522848 | 1626 | struct bitmask *mask; |
d0b937ed YR |
1627 | |
1628 | if (numa_available() < 0) { | |
1629 | td_verror(td, errno, "Does not support NUMA API\n"); | |
1630 | goto err; | |
1631 | } | |
1632 | ||
b2a9e649 | 1633 | if (fio_option_is_set(o, numa_cpunodes)) { |
43522848 DG |
1634 | mask = numa_parse_nodestring(o->numa_cpunodes); |
1635 | ret = numa_run_on_node_mask(mask); | |
1636 | numa_free_nodemask(mask); | |
d0b937ed YR |
1637 | if (ret == -1) { |
1638 | td_verror(td, errno, \ | |
1639 | "numa_run_on_node_mask failed\n"); | |
1640 | goto err; | |
1641 | } | |
1642 | } | |
1643 | ||
b2a9e649 | 1644 | if (fio_option_is_set(o, numa_memnodes)) { |
43522848 DG |
1645 | mask = NULL; |
1646 | if (o->numa_memnodes) | |
1647 | mask = numa_parse_nodestring(o->numa_memnodes); | |
1648 | ||
4896473e | 1649 | switch (o->numa_mem_mode) { |
d0b937ed | 1650 | case MPOL_INTERLEAVE: |
43522848 | 1651 | numa_set_interleave_mask(mask); |
d0b937ed YR |
1652 | break; |
1653 | case MPOL_BIND: | |
43522848 | 1654 | numa_set_membind(mask); |
d0b937ed YR |
1655 | break; |
1656 | case MPOL_LOCAL: | |
1657 | numa_set_localalloc(); | |
1658 | break; | |
1659 | case MPOL_PREFERRED: | |
4896473e | 1660 | numa_set_preferred(o->numa_mem_prefer_node); |
d0b937ed YR |
1661 | break; |
1662 | case MPOL_DEFAULT: | |
1663 | default: | |
1664 | break; | |
1665 | } | |
1666 | ||
43522848 DG |
1667 | if (mask) |
1668 | numa_free_nodemask(mask); | |
1669 | ||
d0b937ed YR |
1670 | } |
1671 | } | |
1672 | #endif | |
1673 | ||
9a3f1100 JA |
1674 | if (fio_pin_memory(td)) |
1675 | goto err; | |
1676 | ||
2e1df07d JA |
1677 | /* |
1678 | * May alter parameters that init_io_u() will use, so we need to | |
1679 | * do this first. | |
1680 | */ | |
1681 | if (init_iolog(td)) | |
1682 | goto err; | |
1683 | ||
1684 | if (init_io_u(td)) | |
1685 | goto err; | |
1686 | ||
4896473e | 1687 | if (o->verify_async && verify_async_init(td)) |
2e1df07d JA |
1688 | goto err; |
1689 | ||
27d74836 JA |
1690 | if (fio_option_is_set(o, ioprio) || |
1691 | fio_option_is_set(o, ioprio_class)) { | |
28727df7 JA |
1692 | ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio); |
1693 | if (ret == -1) { | |
2e1df07d JA |
1694 | td_verror(td, errno, "ioprio_set"); |
1695 | goto err; | |
1696 | } | |
1697 | } | |
1698 | ||
4896473e | 1699 | if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt)) |
2e1df07d JA |
1700 | goto err; |
1701 | ||
649c10c5 | 1702 | errno = 0; |
4896473e | 1703 | if (nice(o->nice) == -1 && errno != 0) { |
2e1df07d JA |
1704 | td_verror(td, errno, "nice"); |
1705 | goto err; | |
1706 | } | |
1707 | ||
4896473e | 1708 | if (o->ioscheduler && switch_ioscheduler(td)) |
2e1df07d JA |
1709 | goto err; |
1710 | ||
4896473e | 1711 | if (!o->create_serialize && setup_files(td)) |
2e1df07d JA |
1712 | goto err; |
1713 | ||
1714 | if (td_io_init(td)) | |
1715 | goto err; | |
1716 | ||
4d832322 | 1717 | if (!init_random_map(td)) |
2e1df07d JA |
1718 | goto err; |
1719 | ||
ce486495 | 1720 | if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun")) |
4896473e | 1721 | goto err; |
2e1df07d | 1722 | |
c139f3b4 JA |
1723 | if (o->pre_read && !pre_read_files(td)) |
1724 | goto err; | |
2e1df07d | 1725 | |
dc5bfbb2 JA |
1726 | fio_verify_init(td); |
1727 | ||
24660963 | 1728 | if (rate_submit_init(td, sk_out)) |
a9da8ab2 JA |
1729 | goto err; |
1730 | ||
3aea75b1 | 1731 | set_epoch_time(td, o->log_unix_epoch); |
44404c5a | 1732 | fio_getrusage(&td->ru_start); |
ac28d905 JA |
1733 | memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch)); |
1734 | memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch)); | |
16e56d25 | 1735 | memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch)); |
ac28d905 JA |
1736 | |
1737 | if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] || | |
1738 | o->ratemin[DDIR_TRIM]) { | |
1739 | memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time, | |
1740 | sizeof(td->bw_sample_time)); | |
1741 | memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time, | |
1742 | sizeof(td->bw_sample_time)); | |
1743 | memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time, | |
1744 | sizeof(td->bw_sample_time)); | |
1745 | } | |
1746 | ||
48366f3a | 1747 | memset(bytes_done, 0, sizeof(bytes_done)); |
c16035aa JA |
1748 | clear_state = false; |
1749 | did_some_io = false; | |
48366f3a | 1750 | |
2e1df07d | 1751 | while (keep_running(td)) { |
100f49f1 JA |
1752 | uint64_t verify_bytes; |
1753 | ||
2e1df07d | 1754 | fio_gettime(&td->start, NULL); |
8b6a404c | 1755 | memcpy(&td->ts_cache, &td->start, sizeof(td->start)); |
2e1df07d | 1756 | |
39c1c323 | 1757 | if (clear_state) { |
ac28d905 | 1758 | clear_io_state(td, 0); |
2e1df07d | 1759 | |
39c1c323 | 1760 | if (o->unlink_each_loop && unlink_all_files(td)) |
1761 | break; | |
1762 | } | |
1763 | ||
2e1df07d JA |
1764 | prune_io_piece_log(td); |
1765 | ||
dcf9844e | 1766 | if (td->o.verify_only && td_write(td)) |
62167762 | 1767 | verify_bytes = do_dry_run(td); |
095196b1 | 1768 | else { |
95b03be0 JA |
1769 | do_io(td, bytes_done); |
1770 | ||
1771 | if (!ddir_rw_sum(bytes_done)) { | |
095196b1 | 1772 | fio_mark_td_terminate(td); |
95b03be0 JA |
1773 | verify_bytes = 0; |
1774 | } else { | |
1775 | verify_bytes = bytes_done[DDIR_WRITE] + | |
1776 | bytes_done[DDIR_TRIM]; | |
1777 | } | |
095196b1 | 1778 | } |
2e1df07d | 1779 | |
49f0b8e0 JA |
1780 | /* |
1781 | * If we took too long to shut down, the main thread could | |
1782 | * already consider us reaped/exited. If that happens, break | |
1783 | * out and clean up. | |
1784 | */ | |
1785 | if (td->runstate >= TD_EXITED) | |
1786 | break; | |
1787 | ||
c16035aa | 1788 | clear_state = true; |
2e1df07d | 1789 | |
40c666c8 JA |
1790 | /* |
1791 | * Make sure we've successfully updated the rusage stats | |
1792 | * before waiting on the stat mutex. Otherwise we could have | |
1793 | * the stat thread holding stat mutex and waiting for | |
1794 | * the rusage_sem, which would never get upped because | |
1795 | * this thread is waiting for the stat mutex. | |
1796 | */ | |
ca0122d8 | 1797 | deadlock_loop_cnt = 0; |
39d13e67 JA |
1798 | do { |
1799 | check_update_rusage(td); | |
971caeb1 | 1800 | if (!fio_sem_down_trylock(stat_sem)) |
39d13e67 JA |
1801 | break; |
1802 | usleep(1000); | |
ca0122d8 | 1803 | if (deadlock_loop_cnt++ > 5000) { |
971caeb1 | 1804 | log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n"); |
43f248c4 | 1805 | td->error = EDEADLK; |
ca0122d8 TT |
1806 | goto err; |
1807 | } | |
39d13e67 | 1808 | } while (1); |
40c666c8 | 1809 | |
95603b74 BF |
1810 | if (td_read(td) && td->io_bytes[DDIR_READ]) |
1811 | update_runtime(td, elapsed_us, DDIR_READ); | |
1812 | if (td_write(td) && td->io_bytes[DDIR_WRITE]) | |
1813 | update_runtime(td, elapsed_us, DDIR_WRITE); | |
1814 | if (td_trim(td) && td->io_bytes[DDIR_TRIM]) | |
1815 | update_runtime(td, elapsed_us, DDIR_TRIM); | |
e5437a07 | 1816 | fio_gettime(&td->start, NULL); |
971caeb1 | 1817 | fio_sem_up(stat_sem); |
2e1df07d JA |
1818 | |
1819 | if (td->error || td->terminate) | |
1820 | break; | |
1821 | ||
4896473e JA |
1822 | if (!o->do_verify || |
1823 | o->verify == VERIFY_NONE || | |
9b87f09b | 1824 | td_ioengine_flagged(td, FIO_UNIDIR)) |
2e1df07d JA |
1825 | continue; |
1826 | ||
c16035aa JA |
1827 | if (ddir_rw_sum(bytes_done)) |
1828 | did_some_io = true; | |
1829 | ||
ac28d905 | 1830 | clear_io_state(td, 0); |
2e1df07d JA |
1831 | |
1832 | fio_gettime(&td->start, NULL); | |
1833 | ||
100f49f1 | 1834 | do_verify(td, verify_bytes); |
2e1df07d | 1835 | |
40c666c8 JA |
1836 | /* |
1837 | * See comment further up for why this is done here. | |
1838 | */ | |
1839 | check_update_rusage(td); | |
1840 | ||
971caeb1 | 1841 | fio_sem_down(stat_sem); |
95603b74 | 1842 | update_runtime(td, elapsed_us, DDIR_READ); |
e5437a07 | 1843 | fio_gettime(&td->start, NULL); |
971caeb1 | 1844 | fio_sem_up(stat_sem); |
2e1df07d JA |
1845 | |
1846 | if (td->error || td->terminate) | |
1847 | break; | |
1848 | } | |
1849 | ||
48366f3a TK |
1850 | /* |
1851 | * If td ended up with no I/O when it should have had, | |
1852 | * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO. | |
1853 | * (Are we not missing other flags that can be ignored ?) | |
1854 | */ | |
1855 | if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) && | |
f7c9bfd5 | 1856 | !did_some_io && !td->o.create_only && |
48366f3a TK |
1857 | !(td_ioengine_flagged(td, FIO_NOIO) || |
1858 | td_ioengine_flagged(td, FIO_DISKLESSIO))) | |
1859 | log_err("%s: No I/O performed by %s, " | |
1860 | "perhaps try --debug=io option for details?\n", | |
1861 | td->o.name, td->io_ops->name); | |
1862 | ||
2cea0b4c JA |
1863 | td_set_runstate(td, TD_FINISHING); |
1864 | ||
2e1df07d | 1865 | update_rusage_stat(td); |
2e1df07d | 1866 | td->ts.total_run_time = mtime_since_now(&td->epoch); |
6eaf09d6 SL |
1867 | td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ]; |
1868 | td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE]; | |
1869 | td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM]; | |
2e1df07d | 1870 | |
ca09be4b | 1871 | if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) && |
d264264a JA |
1872 | (td->o.verify != VERIFY_NONE && td_write(td))) |
1873 | verify_save_state(td->thread_number); | |
ca09be4b | 1874 | |
9a3f1100 JA |
1875 | fio_unpin_memory(td); |
1876 | ||
a47591e4 | 1877 | td_writeout_logs(td, true); |
2e1df07d | 1878 | |
155f2f02 | 1879 | iolog_compress_exit(td); |
103b174e | 1880 | rate_submit_exit(td); |
aee2ab67 | 1881 | |
4896473e | 1882 | if (o->exec_postrun) |
ce486495 | 1883 | exec_string(o, o->exec_postrun, (const char *)"postrun"); |
2e1df07d | 1884 | |
f9cafb12 | 1885 | if (exitall_on_terminate || (o->exitall_error && td->error)) |
2e1df07d JA |
1886 | fio_terminate_threads(td->groupid); |
1887 | ||
1888 | err: | |
1889 | if (td->error) | |
1890 | log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error, | |
1891 | td->verror); | |
1892 | ||
4896473e | 1893 | if (o->verify_async) |
2e1df07d JA |
1894 | verify_async_exit(td); |
1895 | ||
1896 | close_and_free_files(td); | |
2e1df07d | 1897 | cleanup_io_u(td); |
32dbca2c | 1898 | close_ioengine(td); |
2e1df07d | 1899 | cgroup_shutdown(td, &cgroup_mnt); |
ca09be4b | 1900 | verify_free_state(td); |
2e1df07d | 1901 | |
7c4d0fb7 JA |
1902 | if (td->zone_state_index) { |
1903 | int i; | |
1904 | ||
1905 | for (i = 0; i < DDIR_RWDIR_CNT; i++) | |
1906 | free(td->zone_state_index[i]); | |
1907 | free(td->zone_state_index); | |
1908 | td->zone_state_index = NULL; | |
1909 | } | |
1910 | ||
b2a9e649 | 1911 | if (fio_option_is_set(o, cpumask)) { |
8a1db9a1 JA |
1912 | ret = fio_cpuset_exit(&o->cpumask); |
1913 | if (ret) | |
1914 | td_verror(td, ret, "fio_cpuset_exit"); | |
2e1df07d JA |
1915 | } |
1916 | ||
1917 | /* | |
1918 | * do this very late, it will log file closing as well | |
1919 | */ | |
4896473e | 1920 | if (o->write_iolog_file) |
2e1df07d JA |
1921 | write_iolog_close(td); |
1922 | ||
1923 | td_set_runstate(td, TD_EXITED); | |
fda2cfac JA |
1924 | |
1925 | /* | |
1926 | * Do this last after setting our runstate to exited, so we | |
1927 | * know that the stat thread is signaled. | |
1928 | */ | |
1929 | check_update_rusage(td); | |
1930 | ||
24660963 | 1931 | sk_out_drop(); |
e43606c2 | 1932 | return (void *) (uintptr_t) td->error; |
2e1df07d JA |
1933 | } |
1934 | ||
2e1df07d JA |
1935 | /* |
1936 | * Run over the job map and reap the threads that have exited, if any. | |
1937 | */ | |
90eff1c9 SW |
1938 | static void reap_threads(unsigned int *nr_running, uint64_t *t_rate, |
1939 | uint64_t *m_rate) | |
2e1df07d JA |
1940 | { |
1941 | struct thread_data *td; | |
1942 | unsigned int cputhreads, realthreads, pending; | |
1943 | int i, status, ret; | |
1944 | ||
1945 | /* | |
1946 | * reap exited threads (TD_EXITED -> TD_REAPED) | |
1947 | */ | |
1948 | realthreads = pending = cputhreads = 0; | |
1949 | for_each_td(td, i) { | |
1950 | int flags = 0; | |
1951 | ||
8847ae4c | 1952 | if (!strcmp(td->o.ioengine, "cpuio")) |
2e1df07d JA |
1953 | cputhreads++; |
1954 | else | |
1955 | realthreads++; | |
1956 | ||
1957 | if (!td->pid) { | |
1958 | pending++; | |
1959 | continue; | |
1960 | } | |
1961 | if (td->runstate == TD_REAPED) | |
1962 | continue; | |
1963 | if (td->o.use_thread) { | |
1964 | if (td->runstate == TD_EXITED) { | |
1965 | td_set_runstate(td, TD_REAPED); | |
1966 | goto reaped; | |
1967 | } | |
1968 | continue; | |
1969 | } | |
1970 | ||
1971 | flags = WNOHANG; | |
1972 | if (td->runstate == TD_EXITED) | |
1973 | flags = 0; | |
1974 | ||
1975 | /* | |
1976 | * check if someone quit or got killed in an unusual way | |
1977 | */ | |
1978 | ret = waitpid(td->pid, &status, flags); | |
1979 | if (ret < 0) { | |
1980 | if (errno == ECHILD) { | |
1981 | log_err("fio: pid=%d disappeared %d\n", | |
1982 | (int) td->pid, td->runstate); | |
a5e371a6 | 1983 | td->sig = ECHILD; |
2e1df07d JA |
1984 | td_set_runstate(td, TD_REAPED); |
1985 | goto reaped; | |
1986 | } | |
1987 | perror("waitpid"); | |
1988 | } else if (ret == td->pid) { | |
1989 | if (WIFSIGNALED(status)) { | |
1990 | int sig = WTERMSIG(status); | |
1991 | ||
36d80bc7 | 1992 | if (sig != SIGTERM && sig != SIGUSR2) |
2e1df07d JA |
1993 | log_err("fio: pid=%d, got signal=%d\n", |
1994 | (int) td->pid, sig); | |
a5e371a6 | 1995 | td->sig = sig; |
2e1df07d JA |
1996 | td_set_runstate(td, TD_REAPED); |
1997 | goto reaped; | |
1998 | } | |
1999 | if (WIFEXITED(status)) { | |
2000 | if (WEXITSTATUS(status) && !td->error) | |
2001 | td->error = WEXITSTATUS(status); | |
2002 | ||
2003 | td_set_runstate(td, TD_REAPED); | |
2004 | goto reaped; | |
2005 | } | |
2006 | } | |
2007 | ||
cba5460c JA |
2008 | /* |
2009 | * If the job is stuck, do a forceful timeout of it and | |
2010 | * move on. | |
2011 | */ | |
2012 | if (td->terminate && | |
2c45a4ac | 2013 | td->runstate < TD_FSYNCING && |
cba5460c | 2014 | time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) { |
86310a1f TK |
2015 | log_err("fio: job '%s' (state=%d) hasn't exited in " |
2016 | "%lu seconds, it appears to be stuck. Doing " | |
2017 | "forceful exit of this job.\n", | |
2018 | td->o.name, td->runstate, | |
2019 | (unsigned long) time_since_now(&td->terminate_time)); | |
cba5460c JA |
2020 | td_set_runstate(td, TD_REAPED); |
2021 | goto reaped; | |
2022 | } | |
2023 | ||
2e1df07d JA |
2024 | /* |
2025 | * thread is not dead, continue | |
2026 | */ | |
2027 | pending++; | |
2028 | continue; | |
2029 | reaped: | |
2030 | (*nr_running)--; | |
342f4be4 JA |
2031 | (*m_rate) -= ddir_rw_sum(td->o.ratemin); |
2032 | (*t_rate) -= ddir_rw_sum(td->o.rate); | |
2e1df07d JA |
2033 | if (!td->pid) |
2034 | pending--; | |
2035 | ||
2036 | if (td->error) | |
2037 | exit_value++; | |
2038 | ||
2039 | done_secs += mtime_since_now(&td->epoch) / 1000; | |
4a88752a | 2040 | profile_td_exit(td); |
2e1df07d JA |
2041 | } |
2042 | ||
2043 | if (*nr_running == cputhreads && !pending && realthreads) | |
2044 | fio_terminate_threads(TERMINATE_ALL); | |
2045 | } | |
2046 | ||
e39c0676 | 2047 | static bool __check_trigger_file(void) |
ca09be4b JA |
2048 | { |
2049 | struct stat sb; | |
2050 | ||
2051 | if (!trigger_file) | |
e39c0676 | 2052 | return false; |
ca09be4b JA |
2053 | |
2054 | if (stat(trigger_file, &sb)) | |
e39c0676 | 2055 | return false; |
ca09be4b JA |
2056 | |
2057 | if (unlink(trigger_file) < 0) | |
2058 | log_err("fio: failed to unlink %s: %s\n", trigger_file, | |
2059 | strerror(errno)); | |
2060 | ||
e39c0676 | 2061 | return true; |
ca09be4b JA |
2062 | } |
2063 | ||
e39c0676 | 2064 | static bool trigger_timedout(void) |
ca09be4b JA |
2065 | { |
2066 | if (trigger_timeout) | |
98413be7 SW |
2067 | if (time_since_genesis() >= trigger_timeout) { |
2068 | trigger_timeout = 0; | |
2069 | return true; | |
2070 | } | |
ca09be4b | 2071 | |
e39c0676 | 2072 | return false; |
ca09be4b JA |
2073 | } |
2074 | ||
2075 | void exec_trigger(const char *cmd) | |
2076 | { | |
2077 | int ret; | |
2078 | ||
98413be7 | 2079 | if (!cmd || cmd[0] == '\0') |
ca09be4b JA |
2080 | return; |
2081 | ||
2082 | ret = system(cmd); | |
2083 | if (ret == -1) | |
2084 | log_err("fio: failed executing %s trigger\n", cmd); | |
2085 | } | |
2086 | ||
2087 | void check_trigger_file(void) | |
2088 | { | |
2089 | if (__check_trigger_file() || trigger_timedout()) { | |
796fb3ce JA |
2090 | if (nr_clients) |
2091 | fio_clients_send_trigger(trigger_remote_cmd); | |
2092 | else { | |
d264264a | 2093 | verify_save_state(IO_LIST_ALL); |
ca09be4b JA |
2094 | fio_terminate_threads(TERMINATE_ALL); |
2095 | exec_trigger(trigger_cmd); | |
2096 | } | |
2097 | } | |
2098 | } | |
2099 | ||
2100 | static int fio_verify_load_state(struct thread_data *td) | |
2101 | { | |
2102 | int ret; | |
2103 | ||
2104 | if (!td->o.verify_state) | |
2105 | return 0; | |
2106 | ||
2107 | if (is_backend) { | |
2108 | void *data; | |
2109 | ||
2110 | ret = fio_server_get_verify_state(td->o.name, | |
94a6e1bb | 2111 | td->thread_number - 1, &data); |
ca09be4b | 2112 | if (!ret) |
94a6e1bb | 2113 | verify_assign_state(td, data); |
ca09be4b JA |
2114 | } else |
2115 | ret = verify_load_state(td, "local"); | |
2116 | ||
2117 | return ret; | |
2118 | } | |
2119 | ||
06464907 JA |
2120 | static void do_usleep(unsigned int usecs) |
2121 | { | |
2122 | check_for_running_stats(); | |
ca09be4b | 2123 | check_trigger_file(); |
06464907 JA |
2124 | usleep(usecs); |
2125 | } | |
2126 | ||
e39c0676 | 2127 | static bool check_mount_writes(struct thread_data *td) |
e81ecca3 JA |
2128 | { |
2129 | struct fio_file *f; | |
2130 | unsigned int i; | |
2131 | ||
2132 | if (!td_write(td) || td->o.allow_mounted_write) | |
e39c0676 | 2133 | return false; |
e81ecca3 | 2134 | |
5e81f9c1 TK |
2135 | /* |
2136 | * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs | |
2137 | * are mkfs'd and mounted. | |
2138 | */ | |
e81ecca3 | 2139 | for_each_file(td, f, i) { |
5e81f9c1 TK |
2140 | #ifdef FIO_HAVE_CHARDEV_SIZE |
2141 | if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR) | |
2142 | #else | |
686fbd31 | 2143 | if (f->filetype != FIO_TYPE_BLOCK) |
5e81f9c1 | 2144 | #endif |
e81ecca3 JA |
2145 | continue; |
2146 | if (device_is_mounted(f->file_name)) | |
2147 | goto mounted; | |
2148 | } | |
2149 | ||
e39c0676 | 2150 | return false; |
e81ecca3 | 2151 | mounted: |
c5839011 | 2152 | log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name); |
e39c0676 | 2153 | return true; |
e81ecca3 JA |
2154 | } |
2155 | ||
9cc8cb91 AK |
2156 | static bool waitee_running(struct thread_data *me) |
2157 | { | |
2158 | const char *waitee = me->o.wait_for; | |
2159 | const char *self = me->o.name; | |
2160 | struct thread_data *td; | |
2161 | int i; | |
2162 | ||
2163 | if (!waitee) | |
2164 | return false; | |
2165 | ||
2166 | for_each_td(td, i) { | |
2167 | if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee)) | |
2168 | continue; | |
2169 | ||
2170 | if (td->runstate < TD_EXITED) { | |
2171 | dprint(FD_PROCESS, "%s fenced by %s(%s)\n", | |
2172 | self, td->o.name, | |
2173 | runstate_to_name(td->runstate)); | |
2174 | return true; | |
2175 | } | |
2176 | } | |
2177 | ||
2178 | dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee); | |
2179 | return false; | |
2180 | } | |
2181 | ||
2e1df07d JA |
2182 | /* |
2183 | * Main function for kicking off and reaping jobs, as needed. | |
2184 | */ | |
24660963 | 2185 | static void run_threads(struct sk_out *sk_out) |
2e1df07d JA |
2186 | { |
2187 | struct thread_data *td; | |
90eff1c9 SW |
2188 | unsigned int i, todo, nr_running, nr_started; |
2189 | uint64_t m_rate, t_rate; | |
0de5b26f | 2190 | uint64_t spent; |
2e1df07d | 2191 | |
2e1df07d JA |
2192 | if (fio_gtod_offload && fio_start_gtod_thread()) |
2193 | return; | |
334185e9 | 2194 | |
f2a2ce0e | 2195 | fio_idle_prof_init(); |
2e1df07d JA |
2196 | |
2197 | set_sig_handlers(); | |
2198 | ||
3a5f6bde JA |
2199 | nr_thread = nr_process = 0; |
2200 | for_each_td(td, i) { | |
e81ecca3 JA |
2201 | if (check_mount_writes(td)) |
2202 | return; | |
3a5f6bde JA |
2203 | if (td->o.use_thread) |
2204 | nr_thread++; | |
2205 | else | |
2206 | nr_process++; | |
2207 | } | |
2208 | ||
01cfefcc | 2209 | if (output_format & FIO_OUTPUT_NORMAL) { |
2e1df07d JA |
2210 | log_info("Starting "); |
2211 | if (nr_thread) | |
2212 | log_info("%d thread%s", nr_thread, | |
2213 | nr_thread > 1 ? "s" : ""); | |
2214 | if (nr_process) { | |
2215 | if (nr_thread) | |
2216 | log_info(" and "); | |
2217 | log_info("%d process%s", nr_process, | |
2218 | nr_process > 1 ? "es" : ""); | |
2219 | } | |
2220 | log_info("\n"); | |
e411c301 | 2221 | log_info_flush(); |
2e1df07d JA |
2222 | } |
2223 | ||
2224 | todo = thread_number; | |
2225 | nr_running = 0; | |
2226 | nr_started = 0; | |
2227 | m_rate = t_rate = 0; | |
2228 | ||
2229 | for_each_td(td, i) { | |
2230 | print_status_init(td->thread_number - 1); | |
2231 | ||
2232 | if (!td->o.create_serialize) | |
2233 | continue; | |
2234 | ||
ca09be4b JA |
2235 | if (fio_verify_load_state(td)) |
2236 | goto reap; | |
2237 | ||
2e1df07d JA |
2238 | /* |
2239 | * do file setup here so it happens sequentially, | |
2240 | * we don't want X number of threads getting their | |
2241 | * client data interspersed on disk | |
2242 | */ | |
2243 | if (setup_files(td)) { | |
ca09be4b | 2244 | reap: |
2e1df07d JA |
2245 | exit_value++; |
2246 | if (td->error) | |
2247 | log_err("fio: pid=%d, err=%d/%s\n", | |
2248 | (int) td->pid, td->error, td->verror); | |
2249 | td_set_runstate(td, TD_REAPED); | |
2250 | todo--; | |
2251 | } else { | |
2252 | struct fio_file *f; | |
2253 | unsigned int j; | |
2254 | ||
2255 | /* | |
2256 | * for sharing to work, each job must always open | |
2257 | * its own files. so close them, if we opened them | |
2258 | * for creation | |
2259 | */ | |
2260 | for_each_file(td, f, j) { | |
2261 | if (fio_file_open(f)) | |
2262 | td_io_close_file(td, f); | |
2263 | } | |
2264 | } | |
2265 | } | |
2266 | ||
f2a2ce0e HL |
2267 | /* start idle threads before io threads start to run */ |
2268 | fio_idle_prof_start(); | |
2269 | ||
2e1df07d JA |
2270 | set_genesis_time(); |
2271 | ||
2272 | while (todo) { | |
2273 | struct thread_data *map[REAL_MAX_JOBS]; | |
8b6a404c | 2274 | struct timespec this_start; |
2e1df07d | 2275 | int this_jobs = 0, left; |
ce8b6139 | 2276 | struct fork_data *fd; |
2e1df07d JA |
2277 | |
2278 | /* | |
2279 | * create threads (TD_NOT_CREATED -> TD_CREATED) | |
2280 | */ | |
2281 | for_each_td(td, i) { | |
2282 | if (td->runstate != TD_NOT_CREATED) | |
2283 | continue; | |
2284 | ||
2285 | /* | |
2286 | * never got a chance to start, killed by other | |
2287 | * thread for some reason | |
2288 | */ | |
2289 | if (td->terminate) { | |
2290 | todo--; | |
2291 | continue; | |
2292 | } | |
2293 | ||
2294 | if (td->o.start_delay) { | |
0de5b26f | 2295 | spent = utime_since_genesis(); |
2e1df07d | 2296 | |
74454ce4 | 2297 | if (td->o.start_delay > spent) |
2e1df07d JA |
2298 | continue; |
2299 | } | |
2300 | ||
2301 | if (td->o.stonewall && (nr_started || nr_running)) { | |
2302 | dprint(FD_PROCESS, "%s: stonewall wait\n", | |
2303 | td->o.name); | |
2304 | break; | |
2305 | } | |
2306 | ||
9cc8cb91 AK |
2307 | if (waitee_running(td)) { |
2308 | dprint(FD_PROCESS, "%s: waiting for %s\n", | |
2309 | td->o.name, td->o.wait_for); | |
5c74fc76 | 2310 | continue; |
9cc8cb91 AK |
2311 | } |
2312 | ||
2e1df07d JA |
2313 | init_disk_util(td); |
2314 | ||
971caeb1 | 2315 | td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED); |
c97f1ad6 JA |
2316 | td->update_rusage = 0; |
2317 | ||
2e1df07d JA |
2318 | /* |
2319 | * Set state to created. Thread will transition | |
2320 | * to TD_INITIALIZED when it's done setting up. | |
2321 | */ | |
2322 | td_set_runstate(td, TD_CREATED); | |
2323 | map[this_jobs++] = td; | |
2324 | nr_started++; | |
2325 | ||
ce8b6139 JA |
2326 | fd = calloc(1, sizeof(*fd)); |
2327 | fd->td = td; | |
2328 | fd->sk_out = sk_out; | |
2329 | ||
2e1df07d JA |
2330 | if (td->o.use_thread) { |
2331 | int ret; | |
2332 | ||
2333 | dprint(FD_PROCESS, "will pthread_create\n"); | |
2334 | ret = pthread_create(&td->thread, NULL, | |
24660963 | 2335 | thread_main, fd); |
2e1df07d JA |
2336 | if (ret) { |
2337 | log_err("pthread_create: %s\n", | |
2338 | strerror(ret)); | |
24660963 | 2339 | free(fd); |
2e1df07d JA |
2340 | nr_started--; |
2341 | break; | |
2342 | } | |
493dd6c1 | 2343 | fd = NULL; |
2e1df07d JA |
2344 | ret = pthread_detach(td->thread); |
2345 | if (ret) | |
2346 | log_err("pthread_detach: %s", | |
2347 | strerror(ret)); | |
2348 | } else { | |
2349 | pid_t pid; | |
2350 | dprint(FD_PROCESS, "will fork\n"); | |
2351 | pid = fork(); | |
2352 | if (!pid) { | |
ce8b6139 | 2353 | int ret; |
2e1df07d | 2354 | |
ce8b6139 | 2355 | ret = (int)(uintptr_t)thread_main(fd); |
2e1df07d JA |
2356 | _exit(ret); |
2357 | } else if (i == fio_debug_jobno) | |
2358 | *fio_debug_jobp = pid; | |
2359 | } | |
971caeb1 BVA |
2360 | dprint(FD_MUTEX, "wait on startup_sem\n"); |
2361 | if (fio_sem_down_timeout(startup_sem, 10000)) { | |
2e1df07d JA |
2362 | log_err("fio: job startup hung? exiting.\n"); |
2363 | fio_terminate_threads(TERMINATE_ALL); | |
2364 | fio_abort = 1; | |
2365 | nr_started--; | |
6541f4e0 | 2366 | free(fd); |
2e1df07d JA |
2367 | break; |
2368 | } | |
971caeb1 | 2369 | dprint(FD_MUTEX, "done waiting on startup_sem\n"); |
2e1df07d JA |
2370 | } |
2371 | ||
2372 | /* | |
2373 | * Wait for the started threads to transition to | |
2374 | * TD_INITIALIZED. | |
2375 | */ | |
2376 | fio_gettime(&this_start, NULL); | |
2377 | left = this_jobs; | |
2378 | while (left && !fio_abort) { | |
2379 | if (mtime_since_now(&this_start) > JOB_START_TIMEOUT) | |
2380 | break; | |
2381 | ||
06464907 | 2382 | do_usleep(100000); |
2e1df07d JA |
2383 | |
2384 | for (i = 0; i < this_jobs; i++) { | |
2385 | td = map[i]; | |
2386 | if (!td) | |
2387 | continue; | |
2388 | if (td->runstate == TD_INITIALIZED) { | |
2389 | map[i] = NULL; | |
2390 | left--; | |
2391 | } else if (td->runstate >= TD_EXITED) { | |
2392 | map[i] = NULL; | |
2393 | left--; | |
2394 | todo--; | |
2395 | nr_running++; /* work-around... */ | |
2396 | } | |
2397 | } | |
2398 | } | |
2399 | ||
2400 | if (left) { | |
4e87c37a JA |
2401 | log_err("fio: %d job%s failed to start\n", left, |
2402 | left > 1 ? "s" : ""); | |
2e1df07d JA |
2403 | for (i = 0; i < this_jobs; i++) { |
2404 | td = map[i]; | |
2405 | if (!td) | |
2406 | continue; | |
2407 | kill(td->pid, SIGTERM); | |
2408 | } | |
2409 | break; | |
2410 | } | |
2411 | ||
2412 | /* | |
2413 | * start created threads (TD_INITIALIZED -> TD_RUNNING). | |
2414 | */ | |
2415 | for_each_td(td, i) { | |
2416 | if (td->runstate != TD_INITIALIZED) | |
2417 | continue; | |
2418 | ||
2419 | if (in_ramp_time(td)) | |
2420 | td_set_runstate(td, TD_RAMP); | |
2421 | else | |
2422 | td_set_runstate(td, TD_RUNNING); | |
2423 | nr_running++; | |
2424 | nr_started--; | |
342f4be4 JA |
2425 | m_rate += ddir_rw_sum(td->o.ratemin); |
2426 | t_rate += ddir_rw_sum(td->o.rate); | |
2e1df07d | 2427 | todo--; |
971caeb1 | 2428 | fio_sem_up(td->sem); |
2e1df07d JA |
2429 | } |
2430 | ||
2431 | reap_threads(&nr_running, &t_rate, &m_rate); | |
2432 | ||
122c7725 | 2433 | if (todo) |
06464907 | 2434 | do_usleep(100000); |
2e1df07d JA |
2435 | } |
2436 | ||
2437 | while (nr_running) { | |
2438 | reap_threads(&nr_running, &t_rate, &m_rate); | |
06464907 | 2439 | do_usleep(10000); |
2e1df07d JA |
2440 | } |
2441 | ||
f2a2ce0e HL |
2442 | fio_idle_prof_stop(); |
2443 | ||
2e1df07d | 2444 | update_io_ticks(); |
2e1df07d JA |
2445 | } |
2446 | ||
27357187 JA |
2447 | static void free_disk_util(void) |
2448 | { | |
27357187 | 2449 | disk_util_prune_entries(); |
a39fb9ea | 2450 | helper_thread_destroy(); |
2e1df07d JA |
2451 | } |
2452 | ||
24660963 | 2453 | int fio_backend(struct sk_out *sk_out) |
2e1df07d JA |
2454 | { |
2455 | struct thread_data *td; | |
2456 | int i; | |
2457 | ||
2458 | if (exec_profile) { | |
2459 | if (load_profile(exec_profile)) | |
2460 | return 1; | |
2461 | free(exec_profile); | |
2462 | exec_profile = NULL; | |
2463 | } | |
2464 | if (!thread_number) | |
2465 | return 0; | |
2466 | ||
2467 | if (write_bw_log) { | |
aee2ab67 JA |
2468 | struct log_params p = { |
2469 | .log_type = IO_LOG_TYPE_BW, | |
2470 | }; | |
2471 | ||
2472 | setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log"); | |
2473 | setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log"); | |
2474 | setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log"); | |
2e1df07d JA |
2475 | } |
2476 | ||
971caeb1 BVA |
2477 | startup_sem = fio_sem_init(FIO_SEM_LOCKED); |
2478 | if (startup_sem == NULL) | |
2e1df07d | 2479 | return 1; |
2e1df07d JA |
2480 | |
2481 | set_genesis_time(); | |
cef9175e | 2482 | stat_init(); |
971caeb1 | 2483 | helper_thread_create(startup_sem, sk_out); |
2e1df07d JA |
2484 | |
2485 | cgroup_list = smalloc(sizeof(*cgroup_list)); | |
b2e6494c JA |
2486 | if (cgroup_list) |
2487 | INIT_FLIST_HEAD(cgroup_list); | |
2e1df07d | 2488 | |
24660963 | 2489 | run_threads(sk_out); |
2e1df07d | 2490 | |
a39fb9ea | 2491 | helper_thread_exit(); |
8aab824f | 2492 | |
2e1df07d | 2493 | if (!fio_abort) { |
83f7b64e | 2494 | __show_run_stats(); |
2e1df07d | 2495 | if (write_bw_log) { |
cb7e0ace JA |
2496 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
2497 | struct io_log *log = agg_io_log[i]; | |
2498 | ||
60a25727 | 2499 | flush_log(log, false); |
518dac09 | 2500 | free_log(log); |
cb7e0ace | 2501 | } |
2e1df07d JA |
2502 | } |
2503 | } | |
2504 | ||
fda2cfac | 2505 | for_each_td(td, i) { |
c27cc65f | 2506 | steadystate_free(td); |
2e1df07d | 2507 | fio_options_free(td); |
8049adc1 | 2508 | if (td->rusage_sem) { |
971caeb1 | 2509 | fio_sem_remove(td->rusage_sem); |
8049adc1 JA |
2510 | td->rusage_sem = NULL; |
2511 | } | |
971caeb1 BVA |
2512 | fio_sem_remove(td->sem); |
2513 | td->sem = NULL; | |
fda2cfac | 2514 | } |
2e1df07d | 2515 | |
a462baef | 2516 | free_disk_util(); |
b2e6494c JA |
2517 | if (cgroup_list) { |
2518 | cgroup_kill(cgroup_list); | |
2519 | sfree(cgroup_list); | |
2520 | } | |
2e1df07d JA |
2521 | sfree(cgroup_mnt); |
2522 | ||
971caeb1 | 2523 | fio_sem_remove(startup_sem); |
cef9175e | 2524 | stat_exit(); |
2e1df07d JA |
2525 | return exit_value; |
2526 | } |