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> |
c06379a6 | 32 | #include <pthread.h> |
2e1df07d JA |
33 | |
34 | #include "fio.h" | |
2e1df07d JA |
35 | #include "smalloc.h" |
36 | #include "verify.h" | |
2e1df07d JA |
37 | #include "diskutil.h" |
38 | #include "cgroup.h" | |
39 | #include "profile.h" | |
40 | #include "lib/rand.h" | |
f7690c4a | 41 | #include "lib/memalign.h" |
2e1df07d | 42 | #include "server.h" |
44404c5a | 43 | #include "lib/getrusage.h" |
f2a2ce0e | 44 | #include "idletime.h" |
002fe734 | 45 | #include "err.h" |
a9da8ab2 | 46 | #include "workqueue.h" |
e81ecca3 | 47 | #include "lib/mountcheck.h" |
40511301 | 48 | #include "rate-submit.h" |
a39fb9ea | 49 | #include "helper_thread.h" |
ae626d4e | 50 | #include "pshared.h" |
07776127 | 51 | #include "zone-dist.h" |
2e1df07d | 52 | |
971caeb1 | 53 | static struct fio_sem *startup_sem; |
2e1df07d | 54 | static struct flist_head *cgroup_list; |
c3dc516a | 55 | static struct cgroup_mnt *cgroup_mnt; |
2e1df07d | 56 | static int exit_value; |
f1867a7f | 57 | static volatile bool fio_abort; |
3a5f6bde JA |
58 | static unsigned int nr_process = 0; |
59 | static unsigned int nr_thread = 0; | |
2e1df07d | 60 | |
6eaf09d6 | 61 | struct io_log *agg_io_log[DDIR_RWDIR_CNT]; |
2e1df07d | 62 | |
a3efc919 JA |
63 | int groupid = 0; |
64 | unsigned int thread_number = 0; | |
108fea77 | 65 | unsigned int stat_number = 0; |
a3efc919 JA |
66 | int temp_stall_ts; |
67 | unsigned long done_secs = 0; | |
a38c70a8 BVA |
68 | #ifdef PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP |
69 | pthread_mutex_t overlap_check = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP; | |
70 | #else | |
c06379a6 | 71 | pthread_mutex_t overlap_check = PTHREAD_MUTEX_INITIALIZER; |
a38c70a8 | 72 | #endif |
a47591e4 | 73 | |
2e1df07d JA |
74 | #define JOB_START_TIMEOUT (5 * 1000) |
75 | ||
76 | static void sig_int(int sig) | |
77 | { | |
d41647e6 | 78 | if (segments[0].threads) { |
2e1df07d JA |
79 | if (is_backend) |
80 | fio_server_got_signal(sig); | |
81 | else { | |
82 | log_info("\nfio: terminating on signal %d\n", sig); | |
e411c301 | 83 | log_info_flush(); |
2e1df07d JA |
84 | exit_value = 128; |
85 | } | |
86 | ||
64402a8a | 87 | fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL); |
2e1df07d JA |
88 | } |
89 | } | |
90 | ||
d2235e56 | 91 | void sig_show_status(int sig) |
4c6d91e8 JA |
92 | { |
93 | show_running_run_stats(); | |
94 | } | |
95 | ||
2e1df07d JA |
96 | static void set_sig_handlers(void) |
97 | { | |
98 | struct sigaction act; | |
99 | ||
100 | memset(&act, 0, sizeof(act)); | |
101 | act.sa_handler = sig_int; | |
102 | act.sa_flags = SA_RESTART; | |
103 | sigaction(SIGINT, &act, NULL); | |
104 | ||
105 | memset(&act, 0, sizeof(act)); | |
106 | act.sa_handler = sig_int; | |
107 | act.sa_flags = SA_RESTART; | |
108 | sigaction(SIGTERM, &act, NULL); | |
109 | ||
2f694507 BC |
110 | /* Windows uses SIGBREAK as a quit signal from other applications */ |
111 | #ifdef WIN32 | |
112 | memset(&act, 0, sizeof(act)); | |
113 | act.sa_handler = sig_int; | |
114 | act.sa_flags = SA_RESTART; | |
115 | sigaction(SIGBREAK, &act, NULL); | |
116 | #endif | |
117 | ||
4c6d91e8 JA |
118 | memset(&act, 0, sizeof(act)); |
119 | act.sa_handler = sig_show_status; | |
120 | act.sa_flags = SA_RESTART; | |
121 | sigaction(SIGUSR1, &act, NULL); | |
122 | ||
2e1df07d JA |
123 | if (is_backend) { |
124 | memset(&act, 0, sizeof(act)); | |
125 | act.sa_handler = sig_int; | |
126 | act.sa_flags = SA_RESTART; | |
127 | sigaction(SIGPIPE, &act, NULL); | |
128 | } | |
129 | } | |
130 | ||
131 | /* | |
132 | * Check if we are above the minimum rate given. | |
133 | */ | |
8b6a404c | 134 | static bool __check_min_rate(struct thread_data *td, struct timespec *now, |
e39c0676 | 135 | enum fio_ddir ddir) |
2e1df07d JA |
136 | { |
137 | unsigned long long bytes = 0; | |
138 | unsigned long iops = 0; | |
139 | unsigned long spent; | |
140a6888 SB |
140 | unsigned long long rate; |
141 | unsigned long long ratemin = 0; | |
2e1df07d JA |
142 | unsigned int rate_iops = 0; |
143 | unsigned int rate_iops_min = 0; | |
144 | ||
145 | assert(ddir_rw(ddir)); | |
146 | ||
147 | if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir]) | |
e39c0676 | 148 | return false; |
2e1df07d JA |
149 | |
150 | /* | |
151 | * allow a 2 second settle period in the beginning | |
152 | */ | |
153 | if (mtime_since(&td->start, now) < 2000) | |
e39c0676 | 154 | return false; |
2e1df07d JA |
155 | |
156 | iops += td->this_io_blocks[ddir]; | |
157 | bytes += td->this_io_bytes[ddir]; | |
158 | ratemin += td->o.ratemin[ddir]; | |
159 | rate_iops += td->o.rate_iops[ddir]; | |
160 | rate_iops_min += td->o.rate_iops_min[ddir]; | |
161 | ||
162 | /* | |
163 | * if rate blocks is set, sample is running | |
164 | */ | |
165 | if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) { | |
166 | spent = mtime_since(&td->lastrate[ddir], now); | |
167 | if (spent < td->o.ratecycle) | |
e39c0676 | 168 | return false; |
2e1df07d | 169 | |
20e37e4a | 170 | if (td->o.rate[ddir] || td->o.ratemin[ddir]) { |
2e1df07d JA |
171 | /* |
172 | * check bandwidth specified rate | |
173 | */ | |
174 | if (bytes < td->rate_bytes[ddir]) { | |
140a6888 | 175 | log_err("%s: rate_min=%lluB/s not met, only transferred %lluB\n", |
420b104a | 176 | td->o.name, ratemin, bytes); |
e39c0676 | 177 | return true; |
2e1df07d | 178 | } else { |
49cba9b3 JA |
179 | if (spent) |
180 | rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent; | |
181 | else | |
182 | rate = 0; | |
183 | ||
2e1df07d JA |
184 | if (rate < ratemin || |
185 | bytes < td->rate_bytes[ddir]) { | |
140a6888 | 186 | log_err("%s: rate_min=%lluB/s not met, got %lluB/s\n", |
420b104a | 187 | td->o.name, ratemin, rate); |
e39c0676 | 188 | return true; |
2e1df07d JA |
189 | } |
190 | } | |
191 | } else { | |
192 | /* | |
193 | * checks iops specified rate | |
194 | */ | |
195 | if (iops < rate_iops) { | |
420b104a RE |
196 | log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n", |
197 | td->o.name, rate_iops, iops); | |
e39c0676 | 198 | return true; |
2e1df07d | 199 | } else { |
4c707a3b JA |
200 | if (spent) |
201 | rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent; | |
202 | else | |
203 | rate = 0; | |
204 | ||
2e1df07d JA |
205 | if (rate < rate_iops_min || |
206 | iops < td->rate_blocks[ddir]) { | |
140a6888 | 207 | log_err("%s: rate_iops_min=%u not met, got %llu IOPS\n", |
420b104a | 208 | td->o.name, rate_iops_min, rate); |
e39c0676 | 209 | return true; |
2e1df07d JA |
210 | } |
211 | } | |
212 | } | |
213 | } | |
214 | ||
215 | td->rate_bytes[ddir] = bytes; | |
216 | td->rate_blocks[ddir] = iops; | |
217 | memcpy(&td->lastrate[ddir], now, sizeof(*now)); | |
e39c0676 | 218 | return false; |
2e1df07d JA |
219 | } |
220 | ||
8b6a404c | 221 | static bool check_min_rate(struct thread_data *td, struct timespec *now) |
2e1df07d | 222 | { |
e39c0676 | 223 | bool ret = false; |
2e1df07d | 224 | |
dde39f8c AD |
225 | for_each_rw_ddir(ddir) { |
226 | if (td->bytes_done[ddir]) | |
227 | ret |= __check_min_rate(td, now, ddir); | |
228 | } | |
2e1df07d JA |
229 | |
230 | return ret; | |
231 | } | |
232 | ||
233 | /* | |
234 | * When job exits, we can cancel the in-flight IO if we are using async | |
235 | * io. Attempt to do so. | |
236 | */ | |
237 | static void cleanup_pending_aio(struct thread_data *td) | |
238 | { | |
2e1df07d JA |
239 | int r; |
240 | ||
241 | /* | |
242 | * get immediately available events, if any | |
243 | */ | |
55312f9f | 244 | r = io_u_queued_complete(td, 0); |
2e1df07d JA |
245 | |
246 | /* | |
247 | * now cancel remaining active events | |
248 | */ | |
249 | if (td->io_ops->cancel) { | |
2ae0b204 JA |
250 | struct io_u *io_u; |
251 | int i; | |
2e1df07d | 252 | |
2ae0b204 JA |
253 | io_u_qiter(&td->io_u_all, io_u, i) { |
254 | if (io_u->flags & IO_U_F_FLIGHT) { | |
2e1df07d JA |
255 | r = td->io_ops->cancel(td, io_u); |
256 | if (!r) | |
257 | put_io_u(td, io_u); | |
258 | } | |
259 | } | |
260 | } | |
261 | ||
262 | if (td->cur_depth) | |
55312f9f | 263 | r = io_u_queued_complete(td, td->cur_depth); |
2e1df07d JA |
264 | } |
265 | ||
266 | /* | |
267 | * Helper to handle the final sync of a file. Works just like the normal | |
268 | * io path, just does everything sync. | |
269 | */ | |
e39c0676 | 270 | static bool fio_io_sync(struct thread_data *td, struct fio_file *f) |
2e1df07d JA |
271 | { |
272 | struct io_u *io_u = __get_io_u(td); | |
d3b07186 | 273 | enum fio_q_status ret; |
2e1df07d JA |
274 | |
275 | if (!io_u) | |
e39c0676 | 276 | return true; |
2e1df07d JA |
277 | |
278 | io_u->ddir = DDIR_SYNC; | |
279 | io_u->file = f; | |
00615bfb | 280 | io_u_set(td, io_u, IO_U_F_NO_FILE_PUT); |
2e1df07d JA |
281 | |
282 | if (td_io_prep(td, io_u)) { | |
283 | put_io_u(td, io_u); | |
e39c0676 | 284 | return true; |
2e1df07d JA |
285 | } |
286 | ||
287 | requeue: | |
288 | ret = td_io_queue(td, io_u); | |
18f6165e BVA |
289 | switch (ret) { |
290 | case FIO_Q_QUEUED: | |
a80cb54b | 291 | td_io_commit(td); |
55312f9f | 292 | if (io_u_queued_complete(td, 1) < 0) |
e39c0676 | 293 | return true; |
18f6165e BVA |
294 | break; |
295 | case FIO_Q_COMPLETED: | |
2e1df07d JA |
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; |
18f6165e BVA |
303 | break; |
304 | case FIO_Q_BUSY: | |
a80cb54b | 305 | td_io_commit(td); |
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 | { | |
00615bfb | 314 | int ret, ret2; |
61ee0f86 JA |
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); | |
00615bfb SW |
323 | ret2 = 0; |
324 | if (fio_file_open(f)) | |
325 | ret2 = td_io_close_file(td, f); | |
326 | return (ret || ret2); | |
61ee0f86 JA |
327 | } |
328 | ||
8b6a404c | 329 | static inline void __update_ts_cache(struct thread_data *td) |
2e1df07d | 330 | { |
8b6a404c | 331 | fio_gettime(&td->ts_cache, NULL); |
2e1df07d JA |
332 | } |
333 | ||
8b6a404c | 334 | static inline void update_ts_cache(struct thread_data *td) |
2e1df07d | 335 | { |
8b6a404c VF |
336 | if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask) |
337 | __update_ts_cache(td); | |
2e1df07d JA |
338 | } |
339 | ||
8b6a404c | 340 | static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t) |
2e1df07d JA |
341 | { |
342 | if (in_ramp_time(td)) | |
e39c0676 | 343 | return false; |
2e1df07d | 344 | if (!td->o.timeout) |
e39c0676 | 345 | return false; |
0de5b26f | 346 | if (utime_since(&td->epoch, t) >= td->o.timeout) |
e39c0676 | 347 | return true; |
2e1df07d | 348 | |
e39c0676 | 349 | return false; |
2e1df07d JA |
350 | } |
351 | ||
95603b74 BF |
352 | /* |
353 | * We need to update the runtime consistently in ms, but keep a running | |
354 | * tally of the current elapsed time in microseconds for sub millisecond | |
355 | * updates. | |
356 | */ | |
357 | static inline void update_runtime(struct thread_data *td, | |
358 | unsigned long long *elapsed_us, | |
359 | const enum fio_ddir ddir) | |
360 | { | |
d5c8ea29 JA |
361 | if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only) |
362 | return; | |
363 | ||
95603b74 BF |
364 | td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000; |
365 | elapsed_us[ddir] += utime_since_now(&td->start); | |
366 | td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000; | |
367 | } | |
368 | ||
e39c0676 JA |
369 | static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir, |
370 | int *retptr) | |
2e1df07d JA |
371 | { |
372 | int ret = *retptr; | |
373 | ||
374 | if (ret < 0 || td->error) { | |
8b28bd41 DM |
375 | int err = td->error; |
376 | enum error_type_bit eb; | |
2e1df07d JA |
377 | |
378 | if (ret < 0) | |
379 | err = -ret; | |
2e1df07d | 380 | |
8b28bd41 DM |
381 | eb = td_error_type(ddir, err); |
382 | if (!(td->o.continue_on_error & (1 << eb))) | |
e39c0676 | 383 | return true; |
2e1df07d | 384 | |
8b28bd41 | 385 | if (td_non_fatal_error(td, eb, err)) { |
2e1df07d JA |
386 | /* |
387 | * Continue with the I/Os in case of | |
388 | * a non fatal error. | |
389 | */ | |
390 | update_error_count(td, err); | |
391 | td_clear_error(td); | |
392 | *retptr = 0; | |
e39c0676 | 393 | return false; |
2e1df07d JA |
394 | } else if (td->o.fill_device && err == ENOSPC) { |
395 | /* | |
396 | * We expect to hit this error if | |
397 | * fill_device option is set. | |
398 | */ | |
399 | td_clear_error(td); | |
ebea2133 | 400 | fio_mark_td_terminate(td); |
e39c0676 | 401 | return true; |
2e1df07d JA |
402 | } else { |
403 | /* | |
404 | * Stop the I/O in case of a fatal | |
405 | * error. | |
406 | */ | |
407 | update_error_count(td, err); | |
e39c0676 | 408 | return true; |
2e1df07d JA |
409 | } |
410 | } | |
411 | ||
e39c0676 | 412 | return false; |
2e1df07d JA |
413 | } |
414 | ||
c97f1ad6 JA |
415 | static void check_update_rusage(struct thread_data *td) |
416 | { | |
417 | if (td->update_rusage) { | |
418 | td->update_rusage = 0; | |
419 | update_rusage_stat(td); | |
971caeb1 | 420 | fio_sem_up(td->rusage_sem); |
c97f1ad6 JA |
421 | } |
422 | } | |
423 | ||
8b6a404c | 424 | static int wait_for_completions(struct thread_data *td, struct timespec *time) |
69fea81e JA |
425 | { |
426 | const int full = queue_full(td); | |
427 | int min_evts = 0; | |
428 | int ret; | |
429 | ||
6be06c46 JA |
430 | if (td->flags & TD_F_REGROW_LOGS) |
431 | return io_u_quiesce(td); | |
1fed2080 | 432 | |
69fea81e JA |
433 | /* |
434 | * if the queue is full, we MUST reap at least 1 event | |
435 | */ | |
82407585 | 436 | min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth); |
4b157ac6 | 437 | if ((full && !min_evts) || !td->o.iodepth_batch_complete_min) |
69fea81e JA |
438 | min_evts = 1; |
439 | ||
46960d0d | 440 | if (time && __should_check_rate(td)) |
69fea81e JA |
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 | 455 | { |
e9d512d8 JA |
456 | switch (*ret) { |
457 | case FIO_Q_COMPLETED: | |
458 | if (io_u->error) { | |
459 | *ret = -io_u->error; | |
460 | clear_io_u(td, io_u); | |
461 | } else if (io_u->resid) { | |
5fff9543 | 462 | long long bytes = io_u->xfer_buflen - io_u->resid; |
e9d512d8 JA |
463 | struct fio_file *f = io_u->file; |
464 | ||
465 | if (bytes_issued) | |
466 | *bytes_issued += bytes; | |
467 | ||
468 | if (!from_verify) | |
94a78d52 | 469 | trim_io_piece(io_u); |
e9d512d8 JA |
470 | |
471 | /* | |
472 | * zero read, fail | |
473 | */ | |
474 | if (!bytes) { | |
475 | if (!from_verify) | |
476 | unlog_io_piece(td, io_u); | |
477 | td_verror(td, EIO, "full resid"); | |
478 | put_io_u(td, io_u); | |
479 | break; | |
480 | } | |
481 | ||
482 | io_u->xfer_buflen = io_u->resid; | |
483 | io_u->xfer_buf += bytes; | |
484 | io_u->offset += bytes; | |
485 | ||
486 | if (ddir_rw(io_u->ddir)) | |
487 | td->ts.short_io_u[io_u->ddir]++; | |
488 | ||
e9d512d8 JA |
489 | if (io_u->offset == f->real_file_size) |
490 | goto sync_done; | |
491 | ||
492 | requeue_io_u(td, &io_u); | |
493 | } else { | |
494 | sync_done: | |
46960d0d | 495 | if (comp_time && __should_check_rate(td)) |
e9d512d8 JA |
496 | fio_gettime(comp_time, NULL); |
497 | ||
498 | *ret = io_u_sync_complete(td, io_u); | |
499 | if (*ret < 0) | |
500 | break; | |
501 | } | |
a0e3e5a6 | 502 | |
cf2c8d52 JA |
503 | if (td->flags & TD_F_REGROW_LOGS) |
504 | regrow_logs(td); | |
505 | ||
a0e3e5a6 EW |
506 | /* |
507 | * when doing I/O (not when verifying), | |
508 | * check for any errors that are to be ignored | |
509 | */ | |
510 | if (!from_verify) | |
511 | break; | |
512 | ||
e9d512d8 JA |
513 | return 0; |
514 | case FIO_Q_QUEUED: | |
515 | /* | |
516 | * if the engine doesn't have a commit hook, | |
517 | * the io_u is really queued. if it does have such | |
518 | * a hook, it has to call io_u_queued() itself. | |
519 | */ | |
520 | if (td->io_ops->commit == NULL) | |
521 | io_u_queued(td, io_u); | |
522 | if (bytes_issued) | |
523 | *bytes_issued += io_u->xfer_buflen; | |
524 | break; | |
525 | case FIO_Q_BUSY: | |
526 | if (!from_verify) | |
527 | unlog_io_piece(td, io_u); | |
528 | requeue_io_u(td, &io_u); | |
a80cb54b | 529 | td_io_commit(td); |
e9d512d8 JA |
530 | break; |
531 | default: | |
b0775325 | 532 | assert(*ret < 0); |
e9d512d8 JA |
533 | td_verror(td, -(*ret), "td_io_queue"); |
534 | break; | |
535 | } | |
536 | ||
537 | if (break_on_this_error(td, ddir, ret)) | |
538 | return 1; | |
539 | ||
540 | return 0; | |
541 | } | |
542 | ||
e39c0676 | 543 | static inline bool io_in_polling(struct thread_data *td) |
82407585 RP |
544 | { |
545 | return !td->o.iodepth_batch_complete_min && | |
546 | !td->o.iodepth_batch_complete_max; | |
547 | } | |
39c1c323 | 548 | /* |
549 | * Unlinks files from thread data fio_file structure | |
550 | */ | |
551 | static int unlink_all_files(struct thread_data *td) | |
552 | { | |
553 | struct fio_file *f; | |
554 | unsigned int i; | |
555 | int ret = 0; | |
556 | ||
557 | for_each_file(td, f, i) { | |
558 | if (f->filetype != FIO_TYPE_FILE) | |
559 | continue; | |
560 | ret = td_io_unlink_file(td, f); | |
561 | if (ret) | |
562 | break; | |
563 | } | |
564 | ||
565 | if (ret) | |
566 | td_verror(td, ret, "unlink_all_files"); | |
567 | ||
568 | return ret; | |
569 | } | |
82407585 | 570 | |
997b5680 SW |
571 | /* |
572 | * Check if io_u will overlap an in-flight IO in the queue | |
573 | */ | |
c06379a6 | 574 | bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u) |
997b5680 SW |
575 | { |
576 | bool overlap; | |
577 | struct io_u *check_io_u; | |
578 | unsigned long long x1, x2, y1, y2; | |
579 | int i; | |
580 | ||
581 | x1 = io_u->offset; | |
582 | x2 = io_u->offset + io_u->buflen; | |
583 | overlap = false; | |
584 | io_u_qiter(q, check_io_u, i) { | |
585 | if (check_io_u->flags & IO_U_F_FLIGHT) { | |
586 | y1 = check_io_u->offset; | |
587 | y2 = check_io_u->offset + check_io_u->buflen; | |
588 | ||
589 | if (x1 < y2 && y1 < x2) { | |
590 | overlap = true; | |
5fff9543 | 591 | dprint(FD_IO, "in-flight overlap: %llu/%llu, %llu/%llu\n", |
997b5680 SW |
592 | x1, io_u->buflen, |
593 | y1, check_io_u->buflen); | |
594 | break; | |
595 | } | |
596 | } | |
597 | } | |
598 | ||
599 | return overlap; | |
600 | } | |
601 | ||
d3b07186 | 602 | static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u) |
08b2b386 JA |
603 | { |
604 | /* | |
605 | * Check for overlap if the user asked us to, and we have | |
606 | * at least one IO in flight besides this one. | |
607 | */ | |
608 | if (td->o.serialize_overlap && td->cur_depth > 1 && | |
609 | in_flight_overlap(&td->io_u_all, io_u)) | |
610 | return FIO_Q_BUSY; | |
611 | ||
612 | return td_io_queue(td, io_u); | |
613 | } | |
614 | ||
2e1df07d JA |
615 | /* |
616 | * The main verify engine. Runs over the writes we previously submitted, | |
617 | * reads the blocks back in, and checks the crc/md5 of the data. | |
618 | */ | |
100f49f1 | 619 | static void do_verify(struct thread_data *td, uint64_t verify_bytes) |
2e1df07d JA |
620 | { |
621 | struct fio_file *f; | |
622 | struct io_u *io_u; | |
623 | int ret, min_events; | |
624 | unsigned int i; | |
625 | ||
626 | dprint(FD_VERIFY, "starting loop\n"); | |
627 | ||
628 | /* | |
629 | * sync io first and invalidate cache, to make sure we really | |
630 | * read from disk. | |
631 | */ | |
632 | for_each_file(td, f, i) { | |
633 | if (!fio_file_open(f)) | |
634 | continue; | |
635 | if (fio_io_sync(td, f)) | |
636 | break; | |
637 | if (file_invalidate_cache(td, f)) | |
638 | break; | |
639 | } | |
640 | ||
c97f1ad6 JA |
641 | check_update_rusage(td); |
642 | ||
2e1df07d JA |
643 | if (td->error) |
644 | return; | |
645 | ||
8f380de2 CJ |
646 | /* |
647 | * verify_state needs to be reset before verification | |
648 | * proceeds so that expected random seeds match actual | |
649 | * random seeds in headers. The main loop will reset | |
650 | * all random number generators if randrepeat is set. | |
651 | */ | |
652 | if (!td->o.rand_repeatable) | |
653 | td_fill_verify_state_seed(td); | |
654 | ||
2e1df07d JA |
655 | td_set_runstate(td, TD_VERIFYING); |
656 | ||
657 | io_u = NULL; | |
658 | while (!td->terminate) { | |
fbccf46c | 659 | enum fio_ddir ddir; |
e9d512d8 | 660 | int full; |
2e1df07d | 661 | |
8b6a404c | 662 | update_ts_cache(td); |
c97f1ad6 | 663 | check_update_rusage(td); |
2e1df07d | 664 | |
8b6a404c VF |
665 | if (runtime_exceeded(td, &td->ts_cache)) { |
666 | __update_ts_cache(td); | |
667 | if (runtime_exceeded(td, &td->ts_cache)) { | |
ebea2133 | 668 | fio_mark_td_terminate(td); |
2e1df07d JA |
669 | break; |
670 | } | |
671 | } | |
672 | ||
9e684a49 DE |
673 | if (flow_threshold_exceeded(td)) |
674 | continue; | |
675 | ||
44cbc6da JA |
676 | if (!td->o.experimental_verify) { |
677 | io_u = __get_io_u(td); | |
678 | if (!io_u) | |
679 | break; | |
2e1df07d | 680 | |
44cbc6da JA |
681 | if (get_next_verify(td, io_u)) { |
682 | put_io_u(td, io_u); | |
683 | break; | |
684 | } | |
2e1df07d | 685 | |
44cbc6da JA |
686 | if (td_io_prep(td, io_u)) { |
687 | put_io_u(td, io_u); | |
688 | break; | |
689 | } | |
690 | } else { | |
55312f9f | 691 | if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes) |
100f49f1 JA |
692 | break; |
693 | ||
bcd5abfa | 694 | while ((io_u = get_io_u(td)) != NULL) { |
059b61f2 | 695 | if (IS_ERR_OR_NULL(io_u)) { |
002fe734 JA |
696 | io_u = NULL; |
697 | ret = FIO_Q_BUSY; | |
698 | goto reap; | |
699 | } | |
700 | ||
bcd5abfa JA |
701 | /* |
702 | * We are only interested in the places where | |
703 | * we wrote or trimmed IOs. Turn those into | |
704 | * reads for verification purposes. | |
705 | */ | |
706 | if (io_u->ddir == DDIR_READ) { | |
707 | /* | |
708 | * Pretend we issued it for rwmix | |
709 | * accounting | |
710 | */ | |
711 | td->io_issues[DDIR_READ]++; | |
712 | put_io_u(td, io_u); | |
713 | continue; | |
714 | } else if (io_u->ddir == DDIR_TRIM) { | |
715 | io_u->ddir = DDIR_READ; | |
1651e431 | 716 | io_u_set(td, io_u, IO_U_F_TRIMMED); |
bcd5abfa JA |
717 | break; |
718 | } else if (io_u->ddir == DDIR_WRITE) { | |
719 | io_u->ddir = DDIR_READ; | |
5c5c33c1 | 720 | populate_verify_io_u(td, io_u); |
bcd5abfa JA |
721 | break; |
722 | } else { | |
723 | put_io_u(td, io_u); | |
724 | continue; | |
725 | } | |
726 | } | |
44cbc6da | 727 | |
bcd5abfa | 728 | if (!io_u) |
44cbc6da | 729 | break; |
2e1df07d JA |
730 | } |
731 | ||
ca09be4b JA |
732 | if (verify_state_should_stop(td, io_u)) { |
733 | put_io_u(td, io_u); | |
734 | break; | |
735 | } | |
736 | ||
2e1df07d JA |
737 | if (td->o.verify_async) |
738 | io_u->end_io = verify_io_u_async; | |
739 | else | |
740 | io_u->end_io = verify_io_u; | |
741 | ||
fbccf46c | 742 | ddir = io_u->ddir; |
10adc4a7 GG |
743 | if (!td->o.disable_slat) |
744 | fio_gettime(&io_u->start_time, NULL); | |
fbccf46c | 745 | |
08b2b386 | 746 | ret = io_u_submit(td, io_u); |
2e1df07d | 747 | |
e9d512d8 | 748 | if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL)) |
2e1df07d JA |
749 | break; |
750 | ||
751 | /* | |
752 | * if we can queue more, do so. but check if there are | |
753 | * completed io_u's first. Note that we can get BUSY even | |
754 | * without IO queued, if the system is resource starved. | |
755 | */ | |
002fe734 | 756 | reap: |
2e1df07d | 757 | full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth); |
82407585 | 758 | if (full || io_in_polling(td)) |
55312f9f | 759 | ret = wait_for_completions(td, NULL); |
2e1df07d | 760 | |
2e1df07d JA |
761 | if (ret < 0) |
762 | break; | |
763 | } | |
764 | ||
c97f1ad6 JA |
765 | check_update_rusage(td); |
766 | ||
2e1df07d JA |
767 | if (!td->error) { |
768 | min_events = td->cur_depth; | |
769 | ||
770 | if (min_events) | |
55312f9f | 771 | ret = io_u_queued_complete(td, min_events); |
2e1df07d JA |
772 | } else |
773 | cleanup_pending_aio(td); | |
774 | ||
775 | td_set_runstate(td, TD_RUNNING); | |
776 | ||
777 | dprint(FD_VERIFY, "exiting loop\n"); | |
778 | } | |
779 | ||
e39c0676 | 780 | static bool exceeds_number_ios(struct thread_data *td) |
3939fe85 JA |
781 | { |
782 | unsigned long long number_ios; | |
783 | ||
784 | if (!td->o.number_ios) | |
e39c0676 | 785 | return false; |
3939fe85 | 786 | |
cf8a46a7 | 787 | number_ios = ddir_rw_sum(td->io_blocks); |
3939fe85 JA |
788 | number_ios += td->io_u_queued + td->io_u_in_flight; |
789 | ||
cf8a46a7 | 790 | return number_ios >= (td->o.number_ios * td->loops); |
3939fe85 JA |
791 | } |
792 | ||
926098e0 | 793 | static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes) |
f7078f7b | 794 | { |
77731b29 | 795 | unsigned long long bytes, limit; |
f7078f7b JA |
796 | |
797 | if (td_rw(td)) | |
926098e0 | 798 | bytes = this_bytes[DDIR_READ] + this_bytes[DDIR_WRITE]; |
f7078f7b | 799 | else if (td_write(td)) |
926098e0 | 800 | bytes = this_bytes[DDIR_WRITE]; |
6eaf09d6 | 801 | else if (td_read(td)) |
926098e0 | 802 | bytes = this_bytes[DDIR_READ]; |
f7078f7b | 803 | else |
926098e0 | 804 | bytes = this_bytes[DDIR_TRIM]; |
f7078f7b | 805 | |
5be9bf09 TK |
806 | if (td->o.io_size) |
807 | limit = td->o.io_size; | |
77731b29 JA |
808 | else |
809 | limit = td->o.size; | |
810 | ||
cf8a46a7 | 811 | limit *= td->loops; |
77731b29 | 812 | return bytes >= limit || exceeds_number_ios(td); |
f7078f7b JA |
813 | } |
814 | ||
926098e0 | 815 | static bool io_issue_bytes_exceeded(struct thread_data *td) |
e28dd2cf | 816 | { |
926098e0 JA |
817 | return io_bytes_exceeded(td, td->io_issue_bytes); |
818 | } | |
e28dd2cf | 819 | |
926098e0 JA |
820 | static bool io_complete_bytes_exceeded(struct thread_data *td) |
821 | { | |
822 | return io_bytes_exceeded(td, td->this_io_bytes); | |
e28dd2cf JE |
823 | } |
824 | ||
50a8ce86 D |
825 | /* |
826 | * used to calculate the next io time for rate control | |
827 | * | |
828 | */ | |
829 | static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) | |
830 | { | |
1a9bf814 | 831 | uint64_t bps = td->rate_bps[ddir]; |
50a8ce86 D |
832 | |
833 | assert(!(td->flags & TD_F_CHILD)); | |
ff6bb260 | 834 | |
6de65959 | 835 | if (td->o.rate_process == RATE_PROCESS_POISSON) { |
1a9bf814 JA |
836 | uint64_t val, iops; |
837 | ||
ff6bb260 | 838 | iops = bps / td->o.bs[ddir]; |
a77d139b | 839 | val = (int64_t) (1000000 / iops) * |
6c8611c7 | 840 | -logf(__rand_0_1(&td->poisson_state[ddir])); |
a77d139b | 841 | if (val) { |
98e51d7c JA |
842 | dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n", |
843 | (unsigned long long) 1000000 / val, | |
844 | ddir); | |
a77d139b | 845 | } |
6c8611c7 JA |
846 | td->last_usec[ddir] += val; |
847 | return td->last_usec[ddir]; | |
ff6bb260 | 848 | } else if (bps) { |
1a9bf814 JA |
849 | uint64_t bytes = td->rate_io_issue_bytes[ddir]; |
850 | uint64_t secs = bytes / bps; | |
851 | uint64_t remainder = bytes % bps; | |
852 | ||
50a8ce86 | 853 | return remainder * 1000000 / bps + secs * 1000000; |
e7b24047 JA |
854 | } |
855 | ||
856 | return 0; | |
50a8ce86 D |
857 | } |
858 | ||
1a9bf814 JA |
859 | static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir) |
860 | { | |
861 | unsigned long long b; | |
862 | uint64_t total; | |
863 | int left; | |
864 | ||
865 | b = ddir_rw_sum(td->io_blocks); | |
866 | if (b % td->o.thinktime_blocks) | |
867 | return; | |
868 | ||
869 | io_u_quiesce(td); | |
870 | ||
871 | total = 0; | |
872 | if (td->o.thinktime_spin) | |
873 | total = usec_spin(td->o.thinktime_spin); | |
874 | ||
875 | left = td->o.thinktime - total; | |
876 | if (left) | |
877 | total += usec_sleep(td, left); | |
878 | ||
879 | /* | |
880 | * If we're ignoring thinktime for the rate, add the number of bytes | |
1aa39b0c JA |
881 | * we would have done while sleeping, minus one block to ensure we |
882 | * start issuing immediately after the sleep. | |
1a9bf814 | 883 | */ |
1aa39b0c | 884 | if (total && td->rate_bps[ddir] && td->o.rate_ign_think) { |
f2aedb17 | 885 | uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL; |
62b858f6 JA |
886 | uint64_t bs = td->o.min_bs[ddir]; |
887 | uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir]; | |
f2aedb17 JA |
888 | uint64_t over; |
889 | ||
62b858f6 JA |
890 | if (usperop <= total) |
891 | over = bs; | |
f2aedb17 | 892 | else |
62b858f6 | 893 | over = (usperop - total) / usperop * -bs; |
f2aedb17 JA |
894 | |
895 | td->rate_io_issue_bytes[ddir] += (missed - over); | |
182cab24 JA |
896 | /* adjust for rate_process=poisson */ |
897 | td->last_usec[ddir] += total; | |
1aa39b0c | 898 | } |
1a9bf814 JA |
899 | } |
900 | ||
2e1df07d JA |
901 | /* |
902 | * Main IO worker function. It retrieves io_u's to process and queues | |
903 | * and reaps them, checking for rate and errors along the way. | |
100f49f1 JA |
904 | * |
905 | * Returns number of bytes written and trimmed. | |
2e1df07d | 906 | */ |
95b03be0 | 907 | static void do_io(struct thread_data *td, uint64_t *bytes_done) |
2e1df07d JA |
908 | { |
909 | unsigned int i; | |
910 | int ret = 0; | |
c2703bf3 | 911 | uint64_t total_bytes, bytes_issued = 0; |
70c68076 | 912 | |
95b03be0 JA |
913 | for (i = 0; i < DDIR_RWDIR_CNT; i++) |
914 | bytes_done[i] = td->bytes_done[i]; | |
2e1df07d JA |
915 | |
916 | if (in_ramp_time(td)) | |
917 | td_set_runstate(td, TD_RAMP); | |
918 | else | |
919 | td_set_runstate(td, TD_RUNNING); | |
920 | ||
3e260a46 JA |
921 | lat_target_init(td); |
922 | ||
1e564979 JE |
923 | total_bytes = td->o.size; |
924 | /* | |
5be9bf09 | 925 | * Allow random overwrite workloads to write up to io_size |
1e564979 JE |
926 | * before starting verification phase as 'size' doesn't apply. |
927 | */ | |
928 | if (td_write(td) && td_random(td) && td->o.norandommap) | |
5be9bf09 | 929 | total_bytes = max(total_bytes, (uint64_t) td->o.io_size); |
78a6469c JA |
930 | /* |
931 | * If verify_backlog is enabled, we'll run the verify in this | |
932 | * handler as well. For that case, we may need up to twice the | |
933 | * amount of bytes. | |
934 | */ | |
78a6469c JA |
935 | if (td->o.verify != VERIFY_NONE && |
936 | (td_write(td) && td->o.verify_backlog)) | |
c2703bf3 JA |
937 | total_bytes += td->o.size; |
938 | ||
82a90686 | 939 | /* In trimwrite mode, each byte is trimmed and then written, so |
0e4dd95c | 940 | * allow total_bytes to be twice as big */ |
82a90686 | 941 | if (td_trimwrite(td)) |
0e4dd95c DE |
942 | total_bytes += td->total_io_size; |
943 | ||
f7078f7b | 944 | while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || |
e28dd2cf | 945 | (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) || |
c04e4661 | 946 | td->o.time_based) { |
8b6a404c | 947 | struct timespec comp_time; |
2e1df07d | 948 | struct io_u *io_u; |
e9d512d8 | 949 | int full; |
2e1df07d JA |
950 | enum fio_ddir ddir; |
951 | ||
c97f1ad6 JA |
952 | check_update_rusage(td); |
953 | ||
7d7803fa | 954 | if (td->terminate || td->done) |
2e1df07d JA |
955 | break; |
956 | ||
8b6a404c | 957 | update_ts_cache(td); |
2e1df07d | 958 | |
8b6a404c VF |
959 | if (runtime_exceeded(td, &td->ts_cache)) { |
960 | __update_ts_cache(td); | |
961 | if (runtime_exceeded(td, &td->ts_cache)) { | |
ebea2133 | 962 | fio_mark_td_terminate(td); |
2e1df07d JA |
963 | break; |
964 | } | |
965 | } | |
966 | ||
9e684a49 DE |
967 | if (flow_threshold_exceeded(td)) |
968 | continue; | |
969 | ||
f1a32461 JA |
970 | /* |
971 | * Break if we exceeded the bytes. The exception is time | |
972 | * based runs, but we still need to break out of the loop | |
973 | * for those to run verification, if enabled. | |
98e7161c | 974 | * Jobs read from iolog do not use this stop condition. |
f1a32461 JA |
975 | */ |
976 | if (bytes_issued >= total_bytes && | |
98e7161c | 977 | !td->o.read_iolog_file && |
f1a32461 JA |
978 | (!td->o.time_based || |
979 | (td->o.time_based && td->o.verify != VERIFY_NONE))) | |
20876c53 JC |
980 | break; |
981 | ||
2e1df07d | 982 | io_u = get_io_u(td); |
002fe734 JA |
983 | if (IS_ERR_OR_NULL(io_u)) { |
984 | int err = PTR_ERR(io_u); | |
985 | ||
986 | io_u = NULL; | |
1a9bf814 | 987 | ddir = DDIR_INVAL; |
002fe734 JA |
988 | if (err == -EBUSY) { |
989 | ret = FIO_Q_BUSY; | |
990 | goto reap; | |
991 | } | |
3e260a46 JA |
992 | if (td->o.latency_target) |
993 | goto reap; | |
2e1df07d | 994 | break; |
3e260a46 | 995 | } |
2e1df07d | 996 | |
5c5c33c1 BVA |
997 | if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY) |
998 | populate_verify_io_u(td, io_u); | |
999 | ||
2e1df07d JA |
1000 | ddir = io_u->ddir; |
1001 | ||
1002 | /* | |
82af2a7c JA |
1003 | * Add verification end_io handler if: |
1004 | * - Asked to verify (!td_rw(td)) | |
1005 | * - Or the io_u is from our verify list (mixed write/ver) | |
2e1df07d JA |
1006 | */ |
1007 | if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ && | |
82af2a7c | 1008 | ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) { |
c4b6117b | 1009 | |
ca09be4b JA |
1010 | if (verify_state_should_stop(td, io_u)) { |
1011 | put_io_u(td, io_u); | |
1012 | break; | |
1013 | } | |
1014 | ||
2e1df07d JA |
1015 | if (td->o.verify_async) |
1016 | io_u->end_io = verify_io_u_async; | |
1017 | else | |
1018 | io_u->end_io = verify_io_u; | |
1019 | td_set_runstate(td, TD_VERIFYING); | |
1020 | } else if (in_ramp_time(td)) | |
1021 | td_set_runstate(td, TD_RAMP); | |
1022 | else | |
1023 | td_set_runstate(td, TD_RUNNING); | |
1024 | ||
9a50c5c5 | 1025 | /* |
f9401285 JA |
1026 | * Always log IO before it's issued, so we know the specific |
1027 | * order of it. The logged unit will track when the IO has | |
1028 | * completed. | |
9a50c5c5 | 1029 | */ |
c4b6117b PV |
1030 | if (td_write(td) && io_u->ddir == DDIR_WRITE && |
1031 | td->o.do_verify && | |
1032 | td->o.verify != VERIFY_NONE && | |
f9401285 | 1033 | !td->o.experimental_verify) |
c4b6117b PV |
1034 | log_io_piece(td, io_u); |
1035 | ||
a9da8ab2 | 1036 | if (td->o.io_submit_mode == IO_MODE_OFFLOAD) { |
5fff9543 | 1037 | const unsigned long long blen = io_u->xfer_buflen; |
a43f4461 | 1038 | const enum fio_ddir __ddir = acct_ddir(io_u); |
0c5df5f9 | 1039 | |
a9da8ab2 JA |
1040 | if (td->error) |
1041 | break; | |
0c5df5f9 | 1042 | |
26de50cf JA |
1043 | workqueue_enqueue(&td->io_wq, &io_u->work); |
1044 | ret = FIO_Q_QUEUED; | |
50a8ce86 | 1045 | |
a43f4461 JA |
1046 | if (ddir_rw(__ddir)) { |
1047 | td->io_issues[__ddir]++; | |
1048 | td->io_issue_bytes[__ddir] += blen; | |
1049 | td->rate_io_issue_bytes[__ddir] += blen; | |
0c5df5f9 JA |
1050 | } |
1051 | ||
50a8ce86 | 1052 | if (should_check_rate(td)) |
a43f4461 | 1053 | td->rate_next_io_time[__ddir] = usec_for_io(td, __ddir); |
50a8ce86 | 1054 | |
a9da8ab2 | 1055 | } else { |
08b2b386 | 1056 | ret = io_u_submit(td, io_u); |
2e1df07d | 1057 | |
50a8ce86 D |
1058 | if (should_check_rate(td)) |
1059 | td->rate_next_io_time[ddir] = usec_for_io(td, ddir); | |
1060 | ||
fd727d9d | 1061 | if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time)) |
a9da8ab2 | 1062 | break; |
2e1df07d | 1063 | |
a9da8ab2 JA |
1064 | /* |
1065 | * See if we need to complete some commands. Note that | |
1066 | * we can get BUSY even without IO queued, if the | |
1067 | * system is resource starved. | |
1068 | */ | |
3e260a46 | 1069 | reap: |
a9da8ab2 JA |
1070 | full = queue_full(td) || |
1071 | (ret == FIO_Q_BUSY && td->cur_depth); | |
82407585 | 1072 | if (full || io_in_polling(td)) |
a9da8ab2 JA |
1073 | ret = wait_for_completions(td, &comp_time); |
1074 | } | |
2e1df07d JA |
1075 | if (ret < 0) |
1076 | break; | |
55312f9f | 1077 | if (!ddir_rw_sum(td->bytes_done) && |
9b87f09b | 1078 | !td_ioengine_flagged(td, FIO_NOIO)) |
2e1df07d JA |
1079 | continue; |
1080 | ||
55312f9f JA |
1081 | if (!in_ramp_time(td) && should_check_rate(td)) { |
1082 | if (check_min_rate(td, &comp_time)) { | |
f9cafb12 | 1083 | if (exitall_on_terminate || td->o.exitall_error) |
64402a8a | 1084 | fio_terminate_threads(td->groupid, td->o.exit_what); |
2e1df07d JA |
1085 | td_verror(td, EIO, "check_min_rate"); |
1086 | break; | |
1087 | } | |
1088 | } | |
3e260a46 JA |
1089 | if (!in_ramp_time(td) && td->o.latency_target) |
1090 | lat_target_check(td); | |
e155cb64 | 1091 | |
1a9bf814 JA |
1092 | if (ddir_rw(ddir) && td->o.thinktime) |
1093 | handle_thinktime(td, ddir); | |
2e1df07d JA |
1094 | } |
1095 | ||
c97f1ad6 JA |
1096 | check_update_rusage(td); |
1097 | ||
2e1df07d | 1098 | if (td->trim_entries) |
4e0a8fa2 | 1099 | log_err("fio: %lu trim entries leaked?\n", td->trim_entries); |
2e1df07d JA |
1100 | |
1101 | if (td->o.fill_device && td->error == ENOSPC) { | |
1102 | td->error = 0; | |
ebea2133 | 1103 | fio_mark_td_terminate(td); |
2e1df07d JA |
1104 | } |
1105 | if (!td->error) { | |
1106 | struct fio_file *f; | |
1107 | ||
a9da8ab2 JA |
1108 | if (td->o.io_submit_mode == IO_MODE_OFFLOAD) { |
1109 | workqueue_flush(&td->io_wq); | |
1110 | i = 0; | |
1111 | } else | |
1112 | i = td->cur_depth; | |
1113 | ||
2e1df07d | 1114 | if (i) { |
55312f9f | 1115 | ret = io_u_queued_complete(td, i); |
2e1df07d JA |
1116 | if (td->o.fill_device && td->error == ENOSPC) |
1117 | td->error = 0; | |
1118 | } | |
1119 | ||
00615bfb | 1120 | if (should_fsync(td) && (td->o.end_fsync || td->o.fsync_on_close)) { |
2e1df07d JA |
1121 | td_set_runstate(td, TD_FSYNCING); |
1122 | ||
1123 | for_each_file(td, f, i) { | |
61ee0f86 | 1124 | if (!fio_file_fsync(td, f)) |
2e1df07d | 1125 | continue; |
61ee0f86 JA |
1126 | |
1127 | log_err("fio: end_fsync failed for file %s\n", | |
1128 | f->file_name); | |
2e1df07d JA |
1129 | } |
1130 | } | |
1131 | } else | |
1132 | cleanup_pending_aio(td); | |
1133 | ||
1134 | /* | |
1135 | * stop job if we failed doing any IO | |
1136 | */ | |
342f4be4 | 1137 | if (!ddir_rw_sum(td->this_io_bytes)) |
2e1df07d | 1138 | td->done = 1; |
100f49f1 | 1139 | |
95b03be0 JA |
1140 | for (i = 0; i < DDIR_RWDIR_CNT; i++) |
1141 | bytes_done[i] = td->bytes_done[i] - bytes_done[i]; | |
2e1df07d JA |
1142 | } |
1143 | ||
94a6e1bb JA |
1144 | static void free_file_completion_logging(struct thread_data *td) |
1145 | { | |
1146 | struct fio_file *f; | |
1147 | unsigned int i; | |
1148 | ||
1149 | for_each_file(td, f, i) { | |
1150 | if (!f->last_write_comp) | |
1151 | break; | |
1152 | sfree(f->last_write_comp); | |
1153 | } | |
1154 | } | |
1155 | ||
1156 | static int init_file_completion_logging(struct thread_data *td, | |
1157 | unsigned int depth) | |
1158 | { | |
1159 | struct fio_file *f; | |
1160 | unsigned int i; | |
1161 | ||
1162 | if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save) | |
1163 | return 0; | |
1164 | ||
1165 | for_each_file(td, f, i) { | |
1166 | f->last_write_comp = scalloc(depth, sizeof(uint64_t)); | |
1167 | if (!f->last_write_comp) | |
1168 | goto cleanup; | |
1169 | } | |
1170 | ||
1171 | return 0; | |
1172 | ||
1173 | cleanup: | |
1174 | free_file_completion_logging(td); | |
1175 | log_err("fio: failed to alloc write comp data\n"); | |
1176 | return 1; | |
1177 | } | |
1178 | ||
2e1df07d JA |
1179 | static void cleanup_io_u(struct thread_data *td) |
1180 | { | |
2e1df07d JA |
1181 | struct io_u *io_u; |
1182 | ||
2ae0b204 | 1183 | while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) { |
c73ed246 JA |
1184 | |
1185 | if (td->io_ops->io_u_free) | |
1186 | td->io_ops->io_u_free(td, io_u); | |
1187 | ||
3114b675 | 1188 | fio_memfree(io_u, sizeof(*io_u), td_offload_overlap(td)); |
2e1df07d JA |
1189 | } |
1190 | ||
1191 | free_io_mem(td); | |
2ae0b204 JA |
1192 | |
1193 | io_u_rexit(&td->io_u_requeues); | |
3114b675 VF |
1194 | io_u_qexit(&td->io_u_freelist, false); |
1195 | io_u_qexit(&td->io_u_all, td_offload_overlap(td)); | |
ca09be4b | 1196 | |
94a6e1bb | 1197 | free_file_completion_logging(td); |
2e1df07d JA |
1198 | } |
1199 | ||
1200 | static int init_io_u(struct thread_data *td) | |
1201 | { | |
1202 | struct io_u *io_u; | |
2e1df07d | 1203 | int cl_align, i, max_units; |
71e6e5a2 | 1204 | int err; |
2e1df07d JA |
1205 | |
1206 | max_units = td->o.iodepth; | |
59d8d0f5 | 1207 | |
2ae0b204 | 1208 | err = 0; |
34851ad5 | 1209 | err += !io_u_rinit(&td->io_u_requeues, td->o.iodepth); |
3114b675 VF |
1210 | err += !io_u_qinit(&td->io_u_freelist, td->o.iodepth, false); |
1211 | err += !io_u_qinit(&td->io_u_all, td->o.iodepth, td_offload_overlap(td)); | |
2ae0b204 JA |
1212 | |
1213 | if (err) { | |
1214 | log_err("fio: failed setting up IO queues\n"); | |
1215 | return 1; | |
1216 | } | |
1217 | ||
71e6e5a2 AK |
1218 | cl_align = os_cache_line_size(); |
1219 | ||
1220 | for (i = 0; i < max_units; i++) { | |
1221 | void *ptr; | |
1222 | ||
1223 | if (td->terminate) | |
1224 | return 1; | |
1225 | ||
3114b675 | 1226 | ptr = fio_memalign(cl_align, sizeof(*io_u), td_offload_overlap(td)); |
71e6e5a2 AK |
1227 | if (!ptr) { |
1228 | log_err("fio: unable to allocate aligned memory\n"); | |
ab43b8cb | 1229 | return 1; |
71e6e5a2 AK |
1230 | } |
1231 | ||
1232 | io_u = ptr; | |
1233 | memset(io_u, 0, sizeof(*io_u)); | |
1234 | INIT_FLIST_HEAD(&io_u->verify_list); | |
1235 | dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i); | |
1236 | ||
1237 | io_u->index = i; | |
1238 | io_u->flags = IO_U_F_FREE; | |
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); | |
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 | } | |
1256 | ||
1257 | init_io_u_buffers(td); | |
1258 | ||
1259 | if (init_file_completion_logging(td, max_units)) | |
1260 | return 1; | |
1261 | ||
1262 | return 0; | |
1263 | } | |
1264 | ||
1265 | int init_io_u_buffers(struct thread_data *td) | |
1266 | { | |
1267 | struct io_u *io_u; | |
1268 | unsigned long long max_bs, min_write; | |
1269 | int i, max_units; | |
1270 | int data_xfer = 1; | |
1271 | char *p; | |
1272 | ||
1273 | max_units = td->o.iodepth; | |
1274 | max_bs = td_max_bs(td); | |
1275 | min_write = td->o.min_bs[DDIR_WRITE]; | |
1276 | td->orig_buffer_size = (unsigned long long) max_bs | |
1277 | * (unsigned long long) max_units; | |
1278 | ||
1279 | if (td_ioengine_flagged(td, FIO_NOIO) || !(td_read(td) || td_write(td))) | |
1280 | data_xfer = 0; | |
1281 | ||
fd8a09b8 | 1282 | /* |
1283 | * if we may later need to do address alignment, then add any | |
1284 | * possible adjustment here so that we don't cause a buffer | |
1285 | * overflow later. this adjustment may be too much if we get | |
1286 | * lucky and the allocator gives us an aligned address. | |
1287 | */ | |
d01612f3 | 1288 | if (td->o.odirect || td->o.mem_align || td->o.oatomic || |
9b87f09b | 1289 | td_ioengine_flagged(td, FIO_RAWIO)) |
fd8a09b8 | 1290 | td->orig_buffer_size += page_mask + td->o.mem_align; |
1291 | ||
2e1df07d | 1292 | if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) { |
5fff9543 | 1293 | unsigned long long bs; |
2e1df07d JA |
1294 | |
1295 | bs = td->orig_buffer_size + td->o.hugepage_size - 1; | |
1296 | td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1); | |
1297 | } | |
1298 | ||
1299 | if (td->orig_buffer_size != (size_t) td->orig_buffer_size) { | |
1300 | log_err("fio: IO memory too large. Reduce max_bs or iodepth\n"); | |
1301 | return 1; | |
1302 | } | |
1303 | ||
59d8d0f5 | 1304 | if (data_xfer && allocate_io_mem(td)) |
2e1df07d JA |
1305 | return 1; |
1306 | ||
d01612f3 | 1307 | if (td->o.odirect || td->o.mem_align || td->o.oatomic || |
9b87f09b | 1308 | td_ioengine_flagged(td, FIO_RAWIO)) |
248c9436 | 1309 | p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align; |
2e1df07d JA |
1310 | else |
1311 | p = td->orig_buffer; | |
1312 | ||
2e1df07d | 1313 | for (i = 0; i < max_units; i++) { |
71e6e5a2 | 1314 | io_u = td->io_u_all.io_us[i]; |
2e1df07d JA |
1315 | dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i); |
1316 | ||
59d8d0f5 | 1317 | if (data_xfer) { |
2e1df07d JA |
1318 | io_u->buf = p; |
1319 | dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf); | |
1320 | ||
1321 | if (td_write(td)) | |
9c42684e | 1322 | io_u_fill_buffer(td, io_u, min_write, max_bs); |
2e1df07d JA |
1323 | if (td_write(td) && td->o.verify_pattern_bytes) { |
1324 | /* | |
1325 | * Fill the buffer with the pattern if we are | |
1326 | * going to be doing writes. | |
1327 | */ | |
ce35b1ec | 1328 | fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0); |
2e1df07d JA |
1329 | } |
1330 | } | |
2e1df07d JA |
1331 | p += max_bs; |
1332 | } | |
1333 | ||
1334 | return 0; | |
1335 | } | |
1336 | ||
bd4d9bdc TK |
1337 | /* |
1338 | * This function is Linux specific. | |
1339 | * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux. | |
1340 | */ | |
2e1df07d JA |
1341 | static int switch_ioscheduler(struct thread_data *td) |
1342 | { | |
db815fb3 | 1343 | #ifdef FIO_HAVE_IOSCHED_SWITCH |
178ef27a | 1344 | char tmp[256], tmp2[128], *p; |
2e1df07d JA |
1345 | FILE *f; |
1346 | int ret; | |
1347 | ||
9b87f09b | 1348 | if (td_ioengine_flagged(td, FIO_DISKLESSIO)) |
2e1df07d JA |
1349 | return 0; |
1350 | ||
bd4d9bdc TK |
1351 | assert(td->files && td->files[0]); |
1352 | sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root); | |
2e1df07d JA |
1353 | |
1354 | f = fopen(tmp, "r+"); | |
1355 | if (!f) { | |
1356 | if (errno == ENOENT) { | |
1357 | log_err("fio: os or kernel doesn't support IO scheduler" | |
1358 | " switching\n"); | |
1359 | return 0; | |
1360 | } | |
1361 | td_verror(td, errno, "fopen iosched"); | |
1362 | return 1; | |
1363 | } | |
1364 | ||
1365 | /* | |
1366 | * Set io scheduler. | |
1367 | */ | |
1368 | ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f); | |
1369 | if (ferror(f) || ret != 1) { | |
1370 | td_verror(td, errno, "fwrite"); | |
1371 | fclose(f); | |
1372 | return 1; | |
1373 | } | |
1374 | ||
1375 | rewind(f); | |
1376 | ||
1377 | /* | |
1378 | * Read back and check that the selected scheduler is now the default. | |
1379 | */ | |
0e23d456 | 1380 | ret = fread(tmp, 1, sizeof(tmp) - 1, f); |
2e1df07d JA |
1381 | if (ferror(f) || ret < 0) { |
1382 | td_verror(td, errno, "fread"); | |
1383 | fclose(f); | |
1384 | return 1; | |
1385 | } | |
0e23d456 | 1386 | tmp[ret] = '\0'; |
b44b9e45 | 1387 | /* |
178ef27a BVA |
1388 | * either a list of io schedulers or "none\n" is expected. Strip the |
1389 | * trailing newline. | |
b44b9e45 | 1390 | */ |
178ef27a BVA |
1391 | p = tmp; |
1392 | strsep(&p, "\n"); | |
49c6f33d | 1393 | |
67f28a8c TK |
1394 | /* |
1395 | * Write to "none" entry doesn't fail, so check the result here. | |
1396 | */ | |
1397 | if (!strcmp(tmp, "none")) { | |
1398 | log_err("fio: io scheduler is not tunable\n"); | |
1399 | fclose(f); | |
1400 | return 0; | |
1401 | } | |
2e1df07d JA |
1402 | |
1403 | sprintf(tmp2, "[%s]", td->o.ioscheduler); | |
1404 | if (!strstr(tmp, tmp2)) { | |
1405 | log_err("fio: io scheduler %s not found\n", td->o.ioscheduler); | |
1406 | td_verror(td, EINVAL, "iosched_switch"); | |
1407 | fclose(f); | |
1408 | return 1; | |
1409 | } | |
1410 | ||
1411 | fclose(f); | |
1412 | return 0; | |
db815fb3 TK |
1413 | #else |
1414 | return 0; | |
1415 | #endif | |
2e1df07d JA |
1416 | } |
1417 | ||
e39c0676 | 1418 | static bool keep_running(struct thread_data *td) |
2e1df07d | 1419 | { |
77731b29 JA |
1420 | unsigned long long limit; |
1421 | ||
2e1df07d | 1422 | if (td->done) |
e39c0676 | 1423 | return false; |
98413be7 SW |
1424 | if (td->terminate) |
1425 | return false; | |
2e1df07d | 1426 | if (td->o.time_based) |
e39c0676 | 1427 | return true; |
2e1df07d JA |
1428 | if (td->o.loops) { |
1429 | td->o.loops--; | |
e39c0676 | 1430 | return true; |
2e1df07d | 1431 | } |
3939fe85 | 1432 | if (exceeds_number_ios(td)) |
e39c0676 | 1433 | return false; |
26251d8d | 1434 | |
5be9bf09 TK |
1435 | if (td->o.io_size) |
1436 | limit = td->o.io_size; | |
77731b29 JA |
1437 | else |
1438 | limit = td->o.size; | |
1439 | ||
1440 | if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) { | |
5bd5f71a JA |
1441 | uint64_t diff; |
1442 | ||
1443 | /* | |
589d5d4f | 1444 | * If the difference is less than the maximum IO size, we |
5bd5f71a JA |
1445 | * are done. |
1446 | */ | |
77731b29 | 1447 | diff = limit - ddir_rw_sum(td->io_bytes); |
74f4b020 | 1448 | if (diff < td_max_bs(td)) |
e39c0676 | 1449 | return false; |
5bd5f71a | 1450 | |
5be9bf09 | 1451 | if (fio_files_done(td) && !td->o.io_size) |
e39c0676 | 1452 | return false; |
002fe734 | 1453 | |
e39c0676 | 1454 | return true; |
5bd5f71a | 1455 | } |
2e1df07d | 1456 | |
e39c0676 | 1457 | return false; |
2e1df07d JA |
1458 | } |
1459 | ||
b7e4a53c BVA |
1460 | static int exec_string(struct thread_options *o, const char *string, |
1461 | const char *mode) | |
2e1df07d | 1462 | { |
f491a907 | 1463 | int ret; |
2e1df07d JA |
1464 | char *str; |
1465 | ||
b7e4a53c BVA |
1466 | if (asprintf(&str, "%s > %s.%s.txt 2>&1", string, o->name, mode) < 0) |
1467 | return -1; | |
2e1df07d | 1468 | |
b7e4a53c BVA |
1469 | log_info("%s : Saving output of %s in %s.%s.txt\n", o->name, mode, |
1470 | o->name, mode); | |
2e1df07d JA |
1471 | ret = system(str); |
1472 | if (ret == -1) | |
1473 | log_err("fio: exec of cmd <%s> failed\n", str); | |
1474 | ||
1475 | free(str); | |
1476 | return ret; | |
1477 | } | |
1478 | ||
62167762 JC |
1479 | /* |
1480 | * Dry run to compute correct state of numberio for verification. | |
1481 | */ | |
1482 | static uint64_t do_dry_run(struct thread_data *td) | |
1483 | { | |
62167762 JC |
1484 | td_set_runstate(td, TD_RUNNING); |
1485 | ||
1486 | while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || | |
e28dd2cf | 1487 | (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) { |
62167762 JC |
1488 | struct io_u *io_u; |
1489 | int ret; | |
1490 | ||
1491 | if (td->terminate || td->done) | |
1492 | break; | |
1493 | ||
1494 | io_u = get_io_u(td); | |
99955d3d | 1495 | if (IS_ERR_OR_NULL(io_u)) |
62167762 JC |
1496 | break; |
1497 | ||
1651e431 | 1498 | io_u_set(td, io_u, IO_U_F_FLIGHT); |
62167762 JC |
1499 | io_u->error = 0; |
1500 | io_u->resid = 0; | |
1501 | if (ddir_rw(acct_ddir(io_u))) | |
1502 | td->io_issues[acct_ddir(io_u)]++; | |
1503 | if (ddir_rw(io_u->ddir)) { | |
1504 | io_u_mark_depth(td, 1); | |
1505 | td->ts.total_io_u[io_u->ddir]++; | |
1506 | } | |
1507 | ||
2e63e96b PV |
1508 | if (td_write(td) && io_u->ddir == DDIR_WRITE && |
1509 | td->o.do_verify && | |
1510 | td->o.verify != VERIFY_NONE && | |
1511 | !td->o.experimental_verify) | |
1512 | log_io_piece(td, io_u); | |
1513 | ||
55312f9f | 1514 | ret = io_u_sync_complete(td, io_u); |
62167762 JC |
1515 | (void) ret; |
1516 | } | |
1517 | ||
55312f9f | 1518 | return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM]; |
62167762 JC |
1519 | } |
1520 | ||
24660963 JA |
1521 | struct fork_data { |
1522 | struct thread_data *td; | |
1523 | struct sk_out *sk_out; | |
1524 | }; | |
1525 | ||
2e1df07d JA |
1526 | /* |
1527 | * Entry point for the thread based jobs. The process based jobs end up | |
1528 | * here as well, after a little setup. | |
1529 | */ | |
1530 | static void *thread_main(void *data) | |
1531 | { | |
24660963 | 1532 | struct fork_data *fd = data; |
95603b74 | 1533 | unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, }; |
24660963 | 1534 | struct thread_data *td = fd->td; |
4896473e | 1535 | struct thread_options *o = &td->o; |
24660963 | 1536 | struct sk_out *sk_out = fd->sk_out; |
48366f3a | 1537 | uint64_t bytes_done[DDIR_RWDIR_CNT]; |
ca0122d8 | 1538 | int deadlock_loop_cnt; |
ff3aa922 | 1539 | bool clear_state; |
a38c70a8 | 1540 | int res, ret; |
2e1df07d | 1541 | |
24660963 JA |
1542 | sk_out_assign(sk_out); |
1543 | free(fd); | |
1544 | ||
4896473e | 1545 | if (!o->use_thread) { |
2e1df07d JA |
1546 | setsid(); |
1547 | td->pid = getpid(); | |
1548 | } else | |
1549 | td->pid = gettid(); | |
1550 | ||
a0eba820 | 1551 | fio_local_clock_init(); |
5d879392 | 1552 | |
2e1df07d JA |
1553 | dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid); |
1554 | ||
122c7725 JA |
1555 | if (is_backend) |
1556 | fio_server_send_start(td); | |
1557 | ||
2e1df07d JA |
1558 | INIT_FLIST_HEAD(&td->io_log_list); |
1559 | INIT_FLIST_HEAD(&td->io_hist_list); | |
1560 | INIT_FLIST_HEAD(&td->verify_list); | |
1561 | INIT_FLIST_HEAD(&td->trim_list); | |
2e1df07d JA |
1562 | td->io_hist_tree = RB_ROOT; |
1563 | ||
34febb23 | 1564 | ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond); |
f9e5b5ee | 1565 | if (ret) { |
34febb23 | 1566 | td_verror(td, ret, "mutex_cond_init_pshared"); |
f9e5b5ee JK |
1567 | goto err; |
1568 | } | |
34febb23 | 1569 | ret = cond_init_pshared(&td->verify_cond); |
f9e5b5ee | 1570 | if (ret) { |
34febb23 | 1571 | td_verror(td, ret, "mutex_cond_pshared"); |
f9e5b5ee JK |
1572 | goto err; |
1573 | } | |
2e1df07d JA |
1574 | |
1575 | td_set_runstate(td, TD_INITIALIZED); | |
971caeb1 BVA |
1576 | dprint(FD_MUTEX, "up startup_sem\n"); |
1577 | fio_sem_up(startup_sem); | |
1578 | dprint(FD_MUTEX, "wait on td->sem\n"); | |
1579 | fio_sem_down(td->sem); | |
1580 | dprint(FD_MUTEX, "done waiting on td->sem\n"); | |
2e1df07d | 1581 | |
2e1df07d JA |
1582 | /* |
1583 | * A new gid requires privilege, so we need to do this before setting | |
1584 | * the uid. | |
1585 | */ | |
4896473e | 1586 | if (o->gid != -1U && setgid(o->gid)) { |
2e1df07d JA |
1587 | td_verror(td, errno, "setgid"); |
1588 | goto err; | |
1589 | } | |
4896473e | 1590 | if (o->uid != -1U && setuid(o->uid)) { |
2e1df07d JA |
1591 | td_verror(td, errno, "setuid"); |
1592 | goto err; | |
1593 | } | |
1594 | ||
07776127 JA |
1595 | td_zone_gen_index(td); |
1596 | ||
2354d810 JA |
1597 | /* |
1598 | * Do this early, we don't want the compress threads to be limited | |
1599 | * to the same CPUs as the IO workers. So do this before we set | |
1600 | * any potential CPU affinity | |
1601 | */ | |
1602 | if (iolog_compress_init(td, sk_out)) | |
1603 | goto err; | |
1604 | ||
2e1df07d JA |
1605 | /* |
1606 | * If we have a gettimeofday() thread, make sure we exclude that | |
1607 | * thread from this job | |
1608 | */ | |
4896473e JA |
1609 | if (o->gtod_cpu) |
1610 | fio_cpu_clear(&o->cpumask, o->gtod_cpu); | |
2e1df07d JA |
1611 | |
1612 | /* | |
1613 | * Set affinity first, in case it has an impact on the memory | |
1614 | * allocations. | |
1615 | */ | |
b2a9e649 | 1616 | if (fio_option_is_set(o, cpumask)) { |
c2acfbac | 1617 | if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) { |
30cb4c65 | 1618 | ret = fio_cpus_split(&o->cpumask, td->thread_number - 1); |
c2acfbac JA |
1619 | if (!ret) { |
1620 | log_err("fio: no CPUs set\n"); | |
1621 | log_err("fio: Try increasing number of available CPUs\n"); | |
1622 | td_verror(td, EINVAL, "cpus_split"); | |
1623 | goto err; | |
1624 | } | |
1625 | } | |
28727df7 JA |
1626 | ret = fio_setaffinity(td->pid, o->cpumask); |
1627 | if (ret == -1) { | |
4896473e JA |
1628 | td_verror(td, errno, "cpu_set_affinity"); |
1629 | goto err; | |
1630 | } | |
2e1df07d JA |
1631 | } |
1632 | ||
67bf9823 | 1633 | #ifdef CONFIG_LIBNUMA |
d0b937ed | 1634 | /* numa node setup */ |
b2a9e649 JA |
1635 | if (fio_option_is_set(o, numa_cpunodes) || |
1636 | fio_option_is_set(o, numa_memnodes)) { | |
43522848 | 1637 | struct bitmask *mask; |
d0b937ed YR |
1638 | |
1639 | if (numa_available() < 0) { | |
1640 | td_verror(td, errno, "Does not support NUMA API\n"); | |
1641 | goto err; | |
1642 | } | |
1643 | ||
b2a9e649 | 1644 | if (fio_option_is_set(o, numa_cpunodes)) { |
43522848 DG |
1645 | mask = numa_parse_nodestring(o->numa_cpunodes); |
1646 | ret = numa_run_on_node_mask(mask); | |
1647 | numa_free_nodemask(mask); | |
d0b937ed YR |
1648 | if (ret == -1) { |
1649 | td_verror(td, errno, \ | |
1650 | "numa_run_on_node_mask failed\n"); | |
1651 | goto err; | |
1652 | } | |
1653 | } | |
1654 | ||
b2a9e649 | 1655 | if (fio_option_is_set(o, numa_memnodes)) { |
43522848 DG |
1656 | mask = NULL; |
1657 | if (o->numa_memnodes) | |
1658 | mask = numa_parse_nodestring(o->numa_memnodes); | |
1659 | ||
4896473e | 1660 | switch (o->numa_mem_mode) { |
d0b937ed | 1661 | case MPOL_INTERLEAVE: |
43522848 | 1662 | numa_set_interleave_mask(mask); |
d0b937ed YR |
1663 | break; |
1664 | case MPOL_BIND: | |
43522848 | 1665 | numa_set_membind(mask); |
d0b937ed YR |
1666 | break; |
1667 | case MPOL_LOCAL: | |
1668 | numa_set_localalloc(); | |
1669 | break; | |
1670 | case MPOL_PREFERRED: | |
4896473e | 1671 | numa_set_preferred(o->numa_mem_prefer_node); |
d0b937ed YR |
1672 | break; |
1673 | case MPOL_DEFAULT: | |
1674 | default: | |
1675 | break; | |
1676 | } | |
1677 | ||
43522848 DG |
1678 | if (mask) |
1679 | numa_free_nodemask(mask); | |
1680 | ||
d0b937ed YR |
1681 | } |
1682 | } | |
1683 | #endif | |
1684 | ||
9a3f1100 JA |
1685 | if (fio_pin_memory(td)) |
1686 | goto err; | |
1687 | ||
2e1df07d JA |
1688 | /* |
1689 | * May alter parameters that init_io_u() will use, so we need to | |
1690 | * do this first. | |
1691 | */ | |
b153f94a | 1692 | if (!init_iolog(td)) |
2e1df07d JA |
1693 | goto err; |
1694 | ||
08dc3bd5 JA |
1695 | if (td_io_init(td)) |
1696 | goto err; | |
1697 | ||
2e1df07d JA |
1698 | if (init_io_u(td)) |
1699 | goto err; | |
1700 | ||
2041bd34 JA |
1701 | if (td->io_ops->post_init && td->io_ops->post_init(td)) |
1702 | goto err; | |
1703 | ||
4896473e | 1704 | if (o->verify_async && verify_async_init(td)) |
2e1df07d JA |
1705 | goto err; |
1706 | ||
27d74836 JA |
1707 | if (fio_option_is_set(o, ioprio) || |
1708 | fio_option_is_set(o, ioprio_class)) { | |
28727df7 JA |
1709 | ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio); |
1710 | if (ret == -1) { | |
2e1df07d JA |
1711 | td_verror(td, errno, "ioprio_set"); |
1712 | goto err; | |
1713 | } | |
1714 | } | |
1715 | ||
4896473e | 1716 | if (o->cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt)) |
2e1df07d JA |
1717 | goto err; |
1718 | ||
649c10c5 | 1719 | errno = 0; |
4896473e | 1720 | if (nice(o->nice) == -1 && errno != 0) { |
2e1df07d JA |
1721 | td_verror(td, errno, "nice"); |
1722 | goto err; | |
1723 | } | |
1724 | ||
4896473e | 1725 | if (o->ioscheduler && switch_ioscheduler(td)) |
2e1df07d JA |
1726 | goto err; |
1727 | ||
4896473e | 1728 | if (!o->create_serialize && setup_files(td)) |
2e1df07d JA |
1729 | goto err; |
1730 | ||
4d832322 | 1731 | if (!init_random_map(td)) |
2e1df07d JA |
1732 | goto err; |
1733 | ||
0518f09f | 1734 | if (o->exec_prerun && exec_string(o, o->exec_prerun, "prerun")) |
4896473e | 1735 | goto err; |
2e1df07d | 1736 | |
c139f3b4 JA |
1737 | if (o->pre_read && !pre_read_files(td)) |
1738 | goto err; | |
2e1df07d | 1739 | |
dc5bfbb2 JA |
1740 | fio_verify_init(td); |
1741 | ||
24660963 | 1742 | if (rate_submit_init(td, sk_out)) |
a9da8ab2 JA |
1743 | goto err; |
1744 | ||
3aea75b1 | 1745 | set_epoch_time(td, o->log_unix_epoch); |
44404c5a | 1746 | fio_getrusage(&td->ru_start); |
ac28d905 JA |
1747 | memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch)); |
1748 | memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch)); | |
16e56d25 | 1749 | memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch)); |
ac28d905 JA |
1750 | |
1751 | if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] || | |
1752 | o->ratemin[DDIR_TRIM]) { | |
1753 | memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time, | |
1754 | sizeof(td->bw_sample_time)); | |
1755 | memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time, | |
1756 | sizeof(td->bw_sample_time)); | |
1757 | memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time, | |
1758 | sizeof(td->bw_sample_time)); | |
1759 | } | |
1760 | ||
48366f3a | 1761 | memset(bytes_done, 0, sizeof(bytes_done)); |
c16035aa | 1762 | clear_state = false; |
48366f3a | 1763 | |
2e1df07d | 1764 | while (keep_running(td)) { |
100f49f1 JA |
1765 | uint64_t verify_bytes; |
1766 | ||
2e1df07d | 1767 | fio_gettime(&td->start, NULL); |
8b6a404c | 1768 | memcpy(&td->ts_cache, &td->start, sizeof(td->start)); |
2e1df07d | 1769 | |
39c1c323 | 1770 | if (clear_state) { |
ac28d905 | 1771 | clear_io_state(td, 0); |
2e1df07d | 1772 | |
39c1c323 | 1773 | if (o->unlink_each_loop && unlink_all_files(td)) |
1774 | break; | |
1775 | } | |
1776 | ||
2e1df07d JA |
1777 | prune_io_piece_log(td); |
1778 | ||
dcf9844e | 1779 | if (td->o.verify_only && td_write(td)) |
62167762 | 1780 | verify_bytes = do_dry_run(td); |
095196b1 | 1781 | else { |
95b03be0 JA |
1782 | do_io(td, bytes_done); |
1783 | ||
1784 | if (!ddir_rw_sum(bytes_done)) { | |
095196b1 | 1785 | fio_mark_td_terminate(td); |
95b03be0 JA |
1786 | verify_bytes = 0; |
1787 | } else { | |
1788 | verify_bytes = bytes_done[DDIR_WRITE] + | |
1789 | bytes_done[DDIR_TRIM]; | |
1790 | } | |
095196b1 | 1791 | } |
2e1df07d | 1792 | |
49f0b8e0 JA |
1793 | /* |
1794 | * If we took too long to shut down, the main thread could | |
1795 | * already consider us reaped/exited. If that happens, break | |
1796 | * out and clean up. | |
1797 | */ | |
1798 | if (td->runstate >= TD_EXITED) | |
1799 | break; | |
1800 | ||
c16035aa | 1801 | clear_state = true; |
2e1df07d | 1802 | |
40c666c8 JA |
1803 | /* |
1804 | * Make sure we've successfully updated the rusage stats | |
1805 | * before waiting on the stat mutex. Otherwise we could have | |
1806 | * the stat thread holding stat mutex and waiting for | |
1807 | * the rusage_sem, which would never get upped because | |
1808 | * this thread is waiting for the stat mutex. | |
1809 | */ | |
ca0122d8 | 1810 | deadlock_loop_cnt = 0; |
39d13e67 JA |
1811 | do { |
1812 | check_update_rusage(td); | |
971caeb1 | 1813 | if (!fio_sem_down_trylock(stat_sem)) |
39d13e67 JA |
1814 | break; |
1815 | usleep(1000); | |
ca0122d8 | 1816 | if (deadlock_loop_cnt++ > 5000) { |
971caeb1 | 1817 | log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n"); |
43f248c4 | 1818 | td->error = EDEADLK; |
ca0122d8 TT |
1819 | goto err; |
1820 | } | |
39d13e67 | 1821 | } while (1); |
40c666c8 | 1822 | |
95603b74 BF |
1823 | if (td_read(td) && td->io_bytes[DDIR_READ]) |
1824 | update_runtime(td, elapsed_us, DDIR_READ); | |
1825 | if (td_write(td) && td->io_bytes[DDIR_WRITE]) | |
1826 | update_runtime(td, elapsed_us, DDIR_WRITE); | |
1827 | if (td_trim(td) && td->io_bytes[DDIR_TRIM]) | |
1828 | update_runtime(td, elapsed_us, DDIR_TRIM); | |
e5437a07 | 1829 | fio_gettime(&td->start, NULL); |
971caeb1 | 1830 | fio_sem_up(stat_sem); |
2e1df07d JA |
1831 | |
1832 | if (td->error || td->terminate) | |
1833 | break; | |
1834 | ||
4896473e JA |
1835 | if (!o->do_verify || |
1836 | o->verify == VERIFY_NONE || | |
9b87f09b | 1837 | td_ioengine_flagged(td, FIO_UNIDIR)) |
2e1df07d JA |
1838 | continue; |
1839 | ||
ac28d905 | 1840 | clear_io_state(td, 0); |
2e1df07d JA |
1841 | |
1842 | fio_gettime(&td->start, NULL); | |
1843 | ||
100f49f1 | 1844 | do_verify(td, verify_bytes); |
2e1df07d | 1845 | |
40c666c8 JA |
1846 | /* |
1847 | * See comment further up for why this is done here. | |
1848 | */ | |
1849 | check_update_rusage(td); | |
1850 | ||
971caeb1 | 1851 | fio_sem_down(stat_sem); |
95603b74 | 1852 | update_runtime(td, elapsed_us, DDIR_READ); |
e5437a07 | 1853 | fio_gettime(&td->start, NULL); |
971caeb1 | 1854 | fio_sem_up(stat_sem); |
2e1df07d JA |
1855 | |
1856 | if (td->error || td->terminate) | |
1857 | break; | |
1858 | } | |
1859 | ||
817ae977 VF |
1860 | /* |
1861 | * Acquire this lock if we were doing overlap checking in | |
1862 | * offload mode so that we don't clean up this job while | |
1863 | * another thread is checking its io_u's for overlap | |
1864 | */ | |
a38c70a8 BVA |
1865 | if (td_offload_overlap(td)) { |
1866 | int res = pthread_mutex_lock(&overlap_check); | |
1867 | assert(res == 0); | |
1868 | } | |
2cea0b4c | 1869 | td_set_runstate(td, TD_FINISHING); |
a38c70a8 BVA |
1870 | if (td_offload_overlap(td)) { |
1871 | res = pthread_mutex_unlock(&overlap_check); | |
1872 | assert(res == 0); | |
1873 | } | |
2cea0b4c | 1874 | |
2e1df07d | 1875 | update_rusage_stat(td); |
2e1df07d | 1876 | td->ts.total_run_time = mtime_since_now(&td->epoch); |
dde39f8c AD |
1877 | for_each_rw_ddir(ddir) { |
1878 | td->ts.io_bytes[ddir] = td->io_bytes[ddir]; | |
1879 | } | |
2e1df07d | 1880 | |
ca09be4b | 1881 | if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) && |
d264264a JA |
1882 | (td->o.verify != VERIFY_NONE && td_write(td))) |
1883 | verify_save_state(td->thread_number); | |
ca09be4b | 1884 | |
9a3f1100 JA |
1885 | fio_unpin_memory(td); |
1886 | ||
a47591e4 | 1887 | td_writeout_logs(td, true); |
2e1df07d | 1888 | |
155f2f02 | 1889 | iolog_compress_exit(td); |
103b174e | 1890 | rate_submit_exit(td); |
aee2ab67 | 1891 | |
4896473e | 1892 | if (o->exec_postrun) |
0518f09f | 1893 | exec_string(o, o->exec_postrun, "postrun"); |
2e1df07d | 1894 | |
f9cafb12 | 1895 | if (exitall_on_terminate || (o->exitall_error && td->error)) |
64402a8a | 1896 | fio_terminate_threads(td->groupid, td->o.exit_what); |
2e1df07d JA |
1897 | |
1898 | err: | |
1899 | if (td->error) | |
1900 | log_info("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error, | |
1901 | td->verror); | |
1902 | ||
4896473e | 1903 | if (o->verify_async) |
2e1df07d JA |
1904 | verify_async_exit(td); |
1905 | ||
1906 | close_and_free_files(td); | |
2e1df07d | 1907 | cleanup_io_u(td); |
32dbca2c | 1908 | close_ioengine(td); |
c3dc516a | 1909 | cgroup_shutdown(td, cgroup_mnt); |
ca09be4b | 1910 | verify_free_state(td); |
07776127 | 1911 | td_zone_free_index(td); |
7c4d0fb7 | 1912 | |
b2a9e649 | 1913 | if (fio_option_is_set(o, cpumask)) { |
8a1db9a1 JA |
1914 | ret = fio_cpuset_exit(&o->cpumask); |
1915 | if (ret) | |
1916 | td_verror(td, ret, "fio_cpuset_exit"); | |
2e1df07d JA |
1917 | } |
1918 | ||
1919 | /* | |
1920 | * do this very late, it will log file closing as well | |
1921 | */ | |
4896473e | 1922 | if (o->write_iolog_file) |
2e1df07d | 1923 | write_iolog_close(td); |
98e7161c AK |
1924 | if (td->io_log_rfile) |
1925 | fclose(td->io_log_rfile); | |
2e1df07d JA |
1926 | |
1927 | td_set_runstate(td, TD_EXITED); | |
fda2cfac JA |
1928 | |
1929 | /* | |
1930 | * Do this last after setting our runstate to exited, so we | |
1931 | * know that the stat thread is signaled. | |
1932 | */ | |
1933 | check_update_rusage(td); | |
1934 | ||
24660963 | 1935 | sk_out_drop(); |
e43606c2 | 1936 | return (void *) (uintptr_t) td->error; |
2e1df07d JA |
1937 | } |
1938 | ||
2e1df07d JA |
1939 | /* |
1940 | * Run over the job map and reap the threads that have exited, if any. | |
1941 | */ | |
90eff1c9 SW |
1942 | static void reap_threads(unsigned int *nr_running, uint64_t *t_rate, |
1943 | uint64_t *m_rate) | |
2e1df07d JA |
1944 | { |
1945 | struct thread_data *td; | |
1946 | unsigned int cputhreads, realthreads, pending; | |
1947 | int i, status, ret; | |
1948 | ||
1949 | /* | |
1950 | * reap exited threads (TD_EXITED -> TD_REAPED) | |
1951 | */ | |
1952 | realthreads = pending = cputhreads = 0; | |
1953 | for_each_td(td, i) { | |
1954 | int flags = 0; | |
1955 | ||
8847ae4c | 1956 | if (!strcmp(td->o.ioengine, "cpuio")) |
2e1df07d JA |
1957 | cputhreads++; |
1958 | else | |
1959 | realthreads++; | |
1960 | ||
1961 | if (!td->pid) { | |
1962 | pending++; | |
1963 | continue; | |
1964 | } | |
1965 | if (td->runstate == TD_REAPED) | |
1966 | continue; | |
1967 | if (td->o.use_thread) { | |
1968 | if (td->runstate == TD_EXITED) { | |
1969 | td_set_runstate(td, TD_REAPED); | |
1970 | goto reaped; | |
1971 | } | |
1972 | continue; | |
1973 | } | |
1974 | ||
1975 | flags = WNOHANG; | |
1976 | if (td->runstate == TD_EXITED) | |
1977 | flags = 0; | |
1978 | ||
1979 | /* | |
1980 | * check if someone quit or got killed in an unusual way | |
1981 | */ | |
1982 | ret = waitpid(td->pid, &status, flags); | |
1983 | if (ret < 0) { | |
1984 | if (errno == ECHILD) { | |
1985 | log_err("fio: pid=%d disappeared %d\n", | |
1986 | (int) td->pid, td->runstate); | |
a5e371a6 | 1987 | td->sig = ECHILD; |
2e1df07d JA |
1988 | td_set_runstate(td, TD_REAPED); |
1989 | goto reaped; | |
1990 | } | |
1991 | perror("waitpid"); | |
1992 | } else if (ret == td->pid) { | |
1993 | if (WIFSIGNALED(status)) { | |
1994 | int sig = WTERMSIG(status); | |
1995 | ||
36d80bc7 | 1996 | if (sig != SIGTERM && sig != SIGUSR2) |
2e1df07d JA |
1997 | log_err("fio: pid=%d, got signal=%d\n", |
1998 | (int) td->pid, sig); | |
a5e371a6 | 1999 | td->sig = sig; |
2e1df07d JA |
2000 | td_set_runstate(td, TD_REAPED); |
2001 | goto reaped; | |
2002 | } | |
2003 | if (WIFEXITED(status)) { | |
2004 | if (WEXITSTATUS(status) && !td->error) | |
2005 | td->error = WEXITSTATUS(status); | |
2006 | ||
2007 | td_set_runstate(td, TD_REAPED); | |
2008 | goto reaped; | |
2009 | } | |
2010 | } | |
2011 | ||
cba5460c JA |
2012 | /* |
2013 | * If the job is stuck, do a forceful timeout of it and | |
2014 | * move on. | |
2015 | */ | |
2016 | if (td->terminate && | |
2c45a4ac | 2017 | td->runstate < TD_FSYNCING && |
cba5460c | 2018 | time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) { |
86310a1f TK |
2019 | log_err("fio: job '%s' (state=%d) hasn't exited in " |
2020 | "%lu seconds, it appears to be stuck. Doing " | |
2021 | "forceful exit of this job.\n", | |
2022 | td->o.name, td->runstate, | |
2023 | (unsigned long) time_since_now(&td->terminate_time)); | |
cba5460c JA |
2024 | td_set_runstate(td, TD_REAPED); |
2025 | goto reaped; | |
2026 | } | |
2027 | ||
2e1df07d JA |
2028 | /* |
2029 | * thread is not dead, continue | |
2030 | */ | |
2031 | pending++; | |
2032 | continue; | |
2033 | reaped: | |
2034 | (*nr_running)--; | |
342f4be4 JA |
2035 | (*m_rate) -= ddir_rw_sum(td->o.ratemin); |
2036 | (*t_rate) -= ddir_rw_sum(td->o.rate); | |
2e1df07d JA |
2037 | if (!td->pid) |
2038 | pending--; | |
2039 | ||
2040 | if (td->error) | |
2041 | exit_value++; | |
2042 | ||
2043 | done_secs += mtime_since_now(&td->epoch) / 1000; | |
4a88752a | 2044 | profile_td_exit(td); |
42e26e94 | 2045 | flow_exit_job(td); |
2e1df07d JA |
2046 | } |
2047 | ||
2048 | if (*nr_running == cputhreads && !pending && realthreads) | |
64402a8a | 2049 | fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL); |
2e1df07d JA |
2050 | } |
2051 | ||
e39c0676 | 2052 | static bool __check_trigger_file(void) |
ca09be4b JA |
2053 | { |
2054 | struct stat sb; | |
2055 | ||
2056 | if (!trigger_file) | |
e39c0676 | 2057 | return false; |
ca09be4b JA |
2058 | |
2059 | if (stat(trigger_file, &sb)) | |
e39c0676 | 2060 | return false; |
ca09be4b JA |
2061 | |
2062 | if (unlink(trigger_file) < 0) | |
2063 | log_err("fio: failed to unlink %s: %s\n", trigger_file, | |
2064 | strerror(errno)); | |
2065 | ||
e39c0676 | 2066 | return true; |
ca09be4b JA |
2067 | } |
2068 | ||
e39c0676 | 2069 | static bool trigger_timedout(void) |
ca09be4b JA |
2070 | { |
2071 | if (trigger_timeout) | |
98413be7 SW |
2072 | if (time_since_genesis() >= trigger_timeout) { |
2073 | trigger_timeout = 0; | |
2074 | return true; | |
2075 | } | |
ca09be4b | 2076 | |
e39c0676 | 2077 | return false; |
ca09be4b JA |
2078 | } |
2079 | ||
2080 | void exec_trigger(const char *cmd) | |
2081 | { | |
2082 | int ret; | |
2083 | ||
98413be7 | 2084 | if (!cmd || cmd[0] == '\0') |
ca09be4b JA |
2085 | return; |
2086 | ||
2087 | ret = system(cmd); | |
2088 | if (ret == -1) | |
2089 | log_err("fio: failed executing %s trigger\n", cmd); | |
2090 | } | |
2091 | ||
2092 | void check_trigger_file(void) | |
2093 | { | |
2094 | if (__check_trigger_file() || trigger_timedout()) { | |
796fb3ce JA |
2095 | if (nr_clients) |
2096 | fio_clients_send_trigger(trigger_remote_cmd); | |
2097 | else { | |
d264264a | 2098 | verify_save_state(IO_LIST_ALL); |
64402a8a | 2099 | fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL); |
ca09be4b JA |
2100 | exec_trigger(trigger_cmd); |
2101 | } | |
2102 | } | |
2103 | } | |
2104 | ||
2105 | static int fio_verify_load_state(struct thread_data *td) | |
2106 | { | |
2107 | int ret; | |
2108 | ||
2109 | if (!td->o.verify_state) | |
2110 | return 0; | |
2111 | ||
2112 | if (is_backend) { | |
2113 | void *data; | |
2114 | ||
2115 | ret = fio_server_get_verify_state(td->o.name, | |
94a6e1bb | 2116 | td->thread_number - 1, &data); |
ca09be4b | 2117 | if (!ret) |
94a6e1bb | 2118 | verify_assign_state(td, data); |
67710881 AK |
2119 | } else { |
2120 | char prefix[PATH_MAX]; | |
2121 | ||
2122 | if (aux_path) | |
2123 | sprintf(prefix, "%s%clocal", aux_path, | |
2124 | FIO_OS_PATH_SEPARATOR); | |
2125 | else | |
2126 | strcpy(prefix, "local"); | |
2127 | ret = verify_load_state(td, prefix); | |
2128 | } | |
ca09be4b JA |
2129 | |
2130 | return ret; | |
2131 | } | |
2132 | ||
06464907 JA |
2133 | static void do_usleep(unsigned int usecs) |
2134 | { | |
2135 | check_for_running_stats(); | |
ca09be4b | 2136 | check_trigger_file(); |
06464907 JA |
2137 | usleep(usecs); |
2138 | } | |
2139 | ||
e39c0676 | 2140 | static bool check_mount_writes(struct thread_data *td) |
e81ecca3 JA |
2141 | { |
2142 | struct fio_file *f; | |
2143 | unsigned int i; | |
2144 | ||
2145 | if (!td_write(td) || td->o.allow_mounted_write) | |
e39c0676 | 2146 | return false; |
e81ecca3 | 2147 | |
5e81f9c1 TK |
2148 | /* |
2149 | * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs | |
2150 | * are mkfs'd and mounted. | |
2151 | */ | |
e81ecca3 | 2152 | for_each_file(td, f, i) { |
5e81f9c1 TK |
2153 | #ifdef FIO_HAVE_CHARDEV_SIZE |
2154 | if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR) | |
2155 | #else | |
686fbd31 | 2156 | if (f->filetype != FIO_TYPE_BLOCK) |
5e81f9c1 | 2157 | #endif |
e81ecca3 JA |
2158 | continue; |
2159 | if (device_is_mounted(f->file_name)) | |
2160 | goto mounted; | |
2161 | } | |
2162 | ||
e39c0676 | 2163 | return false; |
e81ecca3 | 2164 | mounted: |
c5839011 | 2165 | log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name); |
e39c0676 | 2166 | return true; |
e81ecca3 JA |
2167 | } |
2168 | ||
9cc8cb91 AK |
2169 | static bool waitee_running(struct thread_data *me) |
2170 | { | |
2171 | const char *waitee = me->o.wait_for; | |
2172 | const char *self = me->o.name; | |
2173 | struct thread_data *td; | |
2174 | int i; | |
2175 | ||
2176 | if (!waitee) | |
2177 | return false; | |
2178 | ||
2179 | for_each_td(td, i) { | |
2180 | if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee)) | |
2181 | continue; | |
2182 | ||
2183 | if (td->runstate < TD_EXITED) { | |
2184 | dprint(FD_PROCESS, "%s fenced by %s(%s)\n", | |
2185 | self, td->o.name, | |
2186 | runstate_to_name(td->runstate)); | |
2187 | return true; | |
2188 | } | |
2189 | } | |
2190 | ||
2191 | dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee); | |
2192 | return false; | |
2193 | } | |
2194 | ||
2e1df07d JA |
2195 | /* |
2196 | * Main function for kicking off and reaping jobs, as needed. | |
2197 | */ | |
24660963 | 2198 | static void run_threads(struct sk_out *sk_out) |
2e1df07d JA |
2199 | { |
2200 | struct thread_data *td; | |
90eff1c9 SW |
2201 | unsigned int i, todo, nr_running, nr_started; |
2202 | uint64_t m_rate, t_rate; | |
0de5b26f | 2203 | uint64_t spent; |
2e1df07d | 2204 | |
2e1df07d JA |
2205 | if (fio_gtod_offload && fio_start_gtod_thread()) |
2206 | return; | |
334185e9 | 2207 | |
f2a2ce0e | 2208 | fio_idle_prof_init(); |
2e1df07d JA |
2209 | |
2210 | set_sig_handlers(); | |
2211 | ||
3a5f6bde JA |
2212 | nr_thread = nr_process = 0; |
2213 | for_each_td(td, i) { | |
e81ecca3 JA |
2214 | if (check_mount_writes(td)) |
2215 | return; | |
3a5f6bde JA |
2216 | if (td->o.use_thread) |
2217 | nr_thread++; | |
2218 | else | |
2219 | nr_process++; | |
2220 | } | |
2221 | ||
01cfefcc | 2222 | if (output_format & FIO_OUTPUT_NORMAL) { |
a41fc529 JA |
2223 | struct buf_output out; |
2224 | ||
2225 | buf_output_init(&out); | |
c2fcb6bc | 2226 | __log_buf(&out, "Starting "); |
2e1df07d | 2227 | if (nr_thread) |
c2fcb6bc | 2228 | __log_buf(&out, "%d thread%s", nr_thread, |
2e1df07d JA |
2229 | nr_thread > 1 ? "s" : ""); |
2230 | if (nr_process) { | |
2231 | if (nr_thread) | |
c2fcb6bc JA |
2232 | __log_buf(&out, " and "); |
2233 | __log_buf(&out, "%d process%s", nr_process, | |
2e1df07d JA |
2234 | nr_process > 1 ? "es" : ""); |
2235 | } | |
c2fcb6bc | 2236 | __log_buf(&out, "\n"); |
a41fc529 | 2237 | log_info_buf(out.buf, out.buflen); |
d5bb0a89 | 2238 | buf_output_free(&out); |
2e1df07d JA |
2239 | } |
2240 | ||
2241 | todo = thread_number; | |
2242 | nr_running = 0; | |
2243 | nr_started = 0; | |
2244 | m_rate = t_rate = 0; | |
2245 | ||
2246 | for_each_td(td, i) { | |
2247 | print_status_init(td->thread_number - 1); | |
2248 | ||
2249 | if (!td->o.create_serialize) | |
2250 | continue; | |
2251 | ||
ca09be4b JA |
2252 | if (fio_verify_load_state(td)) |
2253 | goto reap; | |
2254 | ||
2e1df07d JA |
2255 | /* |
2256 | * do file setup here so it happens sequentially, | |
2257 | * we don't want X number of threads getting their | |
2258 | * client data interspersed on disk | |
2259 | */ | |
2260 | if (setup_files(td)) { | |
ca09be4b | 2261 | reap: |
2e1df07d JA |
2262 | exit_value++; |
2263 | if (td->error) | |
2264 | log_err("fio: pid=%d, err=%d/%s\n", | |
2265 | (int) td->pid, td->error, td->verror); | |
2266 | td_set_runstate(td, TD_REAPED); | |
2267 | todo--; | |
2268 | } else { | |
2269 | struct fio_file *f; | |
2270 | unsigned int j; | |
2271 | ||
2272 | /* | |
2273 | * for sharing to work, each job must always open | |
2274 | * its own files. so close them, if we opened them | |
2275 | * for creation | |
2276 | */ | |
2277 | for_each_file(td, f, j) { | |
2278 | if (fio_file_open(f)) | |
2279 | td_io_close_file(td, f); | |
2280 | } | |
2281 | } | |
2282 | } | |
2283 | ||
f2a2ce0e HL |
2284 | /* start idle threads before io threads start to run */ |
2285 | fio_idle_prof_start(); | |
2286 | ||
2e1df07d JA |
2287 | set_genesis_time(); |
2288 | ||
2289 | while (todo) { | |
2290 | struct thread_data *map[REAL_MAX_JOBS]; | |
8b6a404c | 2291 | struct timespec this_start; |
2e1df07d | 2292 | int this_jobs = 0, left; |
ce8b6139 | 2293 | struct fork_data *fd; |
2e1df07d JA |
2294 | |
2295 | /* | |
2296 | * create threads (TD_NOT_CREATED -> TD_CREATED) | |
2297 | */ | |
2298 | for_each_td(td, i) { | |
2299 | if (td->runstate != TD_NOT_CREATED) | |
2300 | continue; | |
2301 | ||
2302 | /* | |
2303 | * never got a chance to start, killed by other | |
2304 | * thread for some reason | |
2305 | */ | |
2306 | if (td->terminate) { | |
2307 | todo--; | |
2308 | continue; | |
2309 | } | |
2310 | ||
2311 | if (td->o.start_delay) { | |
0de5b26f | 2312 | spent = utime_since_genesis(); |
2e1df07d | 2313 | |
74454ce4 | 2314 | if (td->o.start_delay > spent) |
2e1df07d JA |
2315 | continue; |
2316 | } | |
2317 | ||
2318 | if (td->o.stonewall && (nr_started || nr_running)) { | |
2319 | dprint(FD_PROCESS, "%s: stonewall wait\n", | |
2320 | td->o.name); | |
2321 | break; | |
2322 | } | |
2323 | ||
9cc8cb91 AK |
2324 | if (waitee_running(td)) { |
2325 | dprint(FD_PROCESS, "%s: waiting for %s\n", | |
2326 | td->o.name, td->o.wait_for); | |
5c74fc76 | 2327 | continue; |
9cc8cb91 AK |
2328 | } |
2329 | ||
2e1df07d JA |
2330 | init_disk_util(td); |
2331 | ||
971caeb1 | 2332 | td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED); |
c97f1ad6 JA |
2333 | td->update_rusage = 0; |
2334 | ||
2e1df07d JA |
2335 | /* |
2336 | * Set state to created. Thread will transition | |
2337 | * to TD_INITIALIZED when it's done setting up. | |
2338 | */ | |
2339 | td_set_runstate(td, TD_CREATED); | |
2340 | map[this_jobs++] = td; | |
2341 | nr_started++; | |
2342 | ||
ce8b6139 JA |
2343 | fd = calloc(1, sizeof(*fd)); |
2344 | fd->td = td; | |
2345 | fd->sk_out = sk_out; | |
2346 | ||
2e1df07d JA |
2347 | if (td->o.use_thread) { |
2348 | int ret; | |
2349 | ||
2350 | dprint(FD_PROCESS, "will pthread_create\n"); | |
2351 | ret = pthread_create(&td->thread, NULL, | |
24660963 | 2352 | thread_main, fd); |
2e1df07d JA |
2353 | if (ret) { |
2354 | log_err("pthread_create: %s\n", | |
2355 | strerror(ret)); | |
24660963 | 2356 | free(fd); |
2e1df07d JA |
2357 | nr_started--; |
2358 | break; | |
2359 | } | |
493dd6c1 | 2360 | fd = NULL; |
2e1df07d JA |
2361 | ret = pthread_detach(td->thread); |
2362 | if (ret) | |
2363 | log_err("pthread_detach: %s", | |
2364 | strerror(ret)); | |
2365 | } else { | |
2366 | pid_t pid; | |
2367 | dprint(FD_PROCESS, "will fork\n"); | |
2368 | pid = fork(); | |
2369 | if (!pid) { | |
ce8b6139 | 2370 | int ret; |
2e1df07d | 2371 | |
ce8b6139 | 2372 | ret = (int)(uintptr_t)thread_main(fd); |
2e1df07d JA |
2373 | _exit(ret); |
2374 | } else if (i == fio_debug_jobno) | |
2375 | *fio_debug_jobp = pid; | |
2376 | } | |
971caeb1 BVA |
2377 | dprint(FD_MUTEX, "wait on startup_sem\n"); |
2378 | if (fio_sem_down_timeout(startup_sem, 10000)) { | |
2e1df07d | 2379 | log_err("fio: job startup hung? exiting.\n"); |
64402a8a | 2380 | fio_terminate_threads(TERMINATE_ALL, TERMINATE_ALL); |
f1867a7f | 2381 | fio_abort = true; |
2e1df07d | 2382 | nr_started--; |
6541f4e0 | 2383 | free(fd); |
2e1df07d JA |
2384 | break; |
2385 | } | |
971caeb1 | 2386 | dprint(FD_MUTEX, "done waiting on startup_sem\n"); |
2e1df07d JA |
2387 | } |
2388 | ||
2389 | /* | |
2390 | * Wait for the started threads to transition to | |
2391 | * TD_INITIALIZED. | |
2392 | */ | |
2393 | fio_gettime(&this_start, NULL); | |
2394 | left = this_jobs; | |
2395 | while (left && !fio_abort) { | |
2396 | if (mtime_since_now(&this_start) > JOB_START_TIMEOUT) | |
2397 | break; | |
2398 | ||
06464907 | 2399 | do_usleep(100000); |
2e1df07d JA |
2400 | |
2401 | for (i = 0; i < this_jobs; i++) { | |
2402 | td = map[i]; | |
2403 | if (!td) | |
2404 | continue; | |
2405 | if (td->runstate == TD_INITIALIZED) { | |
2406 | map[i] = NULL; | |
2407 | left--; | |
2408 | } else if (td->runstate >= TD_EXITED) { | |
2409 | map[i] = NULL; | |
2410 | left--; | |
2411 | todo--; | |
2412 | nr_running++; /* work-around... */ | |
2413 | } | |
2414 | } | |
2415 | } | |
2416 | ||
2417 | if (left) { | |
4e87c37a JA |
2418 | log_err("fio: %d job%s failed to start\n", left, |
2419 | left > 1 ? "s" : ""); | |
2e1df07d JA |
2420 | for (i = 0; i < this_jobs; i++) { |
2421 | td = map[i]; | |
2422 | if (!td) | |
2423 | continue; | |
2424 | kill(td->pid, SIGTERM); | |
2425 | } | |
2426 | break; | |
2427 | } | |
2428 | ||
2429 | /* | |
2430 | * start created threads (TD_INITIALIZED -> TD_RUNNING). | |
2431 | */ | |
2432 | for_each_td(td, i) { | |
2433 | if (td->runstate != TD_INITIALIZED) | |
2434 | continue; | |
2435 | ||
2436 | if (in_ramp_time(td)) | |
2437 | td_set_runstate(td, TD_RAMP); | |
2438 | else | |
2439 | td_set_runstate(td, TD_RUNNING); | |
2440 | nr_running++; | |
2441 | nr_started--; | |
342f4be4 JA |
2442 | m_rate += ddir_rw_sum(td->o.ratemin); |
2443 | t_rate += ddir_rw_sum(td->o.rate); | |
2e1df07d | 2444 | todo--; |
971caeb1 | 2445 | fio_sem_up(td->sem); |
2e1df07d JA |
2446 | } |
2447 | ||
2448 | reap_threads(&nr_running, &t_rate, &m_rate); | |
2449 | ||
122c7725 | 2450 | if (todo) |
06464907 | 2451 | do_usleep(100000); |
2e1df07d JA |
2452 | } |
2453 | ||
2454 | while (nr_running) { | |
2455 | reap_threads(&nr_running, &t_rate, &m_rate); | |
06464907 | 2456 | do_usleep(10000); |
2e1df07d JA |
2457 | } |
2458 | ||
f2a2ce0e HL |
2459 | fio_idle_prof_stop(); |
2460 | ||
2e1df07d | 2461 | update_io_ticks(); |
2e1df07d JA |
2462 | } |
2463 | ||
27357187 JA |
2464 | static void free_disk_util(void) |
2465 | { | |
27357187 | 2466 | disk_util_prune_entries(); |
a39fb9ea | 2467 | helper_thread_destroy(); |
2e1df07d JA |
2468 | } |
2469 | ||
24660963 | 2470 | int fio_backend(struct sk_out *sk_out) |
2e1df07d JA |
2471 | { |
2472 | struct thread_data *td; | |
2473 | int i; | |
2474 | ||
2475 | if (exec_profile) { | |
2476 | if (load_profile(exec_profile)) | |
2477 | return 1; | |
2478 | free(exec_profile); | |
2479 | exec_profile = NULL; | |
2480 | } | |
2481 | if (!thread_number) | |
2482 | return 0; | |
2483 | ||
2484 | if (write_bw_log) { | |
aee2ab67 JA |
2485 | struct log_params p = { |
2486 | .log_type = IO_LOG_TYPE_BW, | |
2487 | }; | |
2488 | ||
2489 | setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log"); | |
2490 | setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log"); | |
2491 | setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log"); | |
2e1df07d JA |
2492 | } |
2493 | ||
971caeb1 | 2494 | startup_sem = fio_sem_init(FIO_SEM_LOCKED); |
24a97130 LC |
2495 | if (!sk_out) |
2496 | is_local_backend = true; | |
971caeb1 | 2497 | if (startup_sem == NULL) |
2e1df07d | 2498 | return 1; |
2e1df07d JA |
2499 | |
2500 | set_genesis_time(); | |
cef9175e | 2501 | stat_init(); |
998e9ebb BVA |
2502 | if (helper_thread_create(startup_sem, sk_out)) |
2503 | log_err("fio: failed to create helper thread\n"); | |
2e1df07d JA |
2504 | |
2505 | cgroup_list = smalloc(sizeof(*cgroup_list)); | |
b2e6494c JA |
2506 | if (cgroup_list) |
2507 | INIT_FLIST_HEAD(cgroup_list); | |
2e1df07d | 2508 | |
24660963 | 2509 | run_threads(sk_out); |
2e1df07d | 2510 | |
a39fb9ea | 2511 | helper_thread_exit(); |
8aab824f | 2512 | |
2e1df07d | 2513 | if (!fio_abort) { |
83f7b64e | 2514 | __show_run_stats(); |
2e1df07d | 2515 | if (write_bw_log) { |
cb7e0ace JA |
2516 | for (i = 0; i < DDIR_RWDIR_CNT; i++) { |
2517 | struct io_log *log = agg_io_log[i]; | |
2518 | ||
60a25727 | 2519 | flush_log(log, false); |
518dac09 | 2520 | free_log(log); |
cb7e0ace | 2521 | } |
2e1df07d JA |
2522 | } |
2523 | } | |
2524 | ||
fda2cfac | 2525 | for_each_td(td, i) { |
c27cc65f | 2526 | steadystate_free(td); |
2e1df07d | 2527 | fio_options_free(td); |
8049adc1 | 2528 | if (td->rusage_sem) { |
971caeb1 | 2529 | fio_sem_remove(td->rusage_sem); |
8049adc1 JA |
2530 | td->rusage_sem = NULL; |
2531 | } | |
971caeb1 BVA |
2532 | fio_sem_remove(td->sem); |
2533 | td->sem = NULL; | |
fda2cfac | 2534 | } |
2e1df07d | 2535 | |
a462baef | 2536 | free_disk_util(); |
b2e6494c JA |
2537 | if (cgroup_list) { |
2538 | cgroup_kill(cgroup_list); | |
2539 | sfree(cgroup_list); | |
2540 | } | |
2e1df07d | 2541 | |
971caeb1 | 2542 | fio_sem_remove(startup_sem); |
cef9175e | 2543 | stat_exit(); |
2e1df07d JA |
2544 | return exit_value; |
2545 | } |