[PATCH] Correct command line parsing
[fio.git] / fio.c
CommitLineData
ebac4655
JA
1/*
2 * fio - the flexible io tester
3 *
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19 *
20 */
ebac4655
JA
21#include <unistd.h>
22#include <fcntl.h>
23#include <string.h>
ebac4655
JA
24#include <signal.h>
25#include <time.h>
ebac4655 26#include <assert.h>
ebac4655
JA
27#include <sys/stat.h>
28#include <sys/wait.h>
29#include <sys/ipc.h>
30#include <sys/shm.h>
31#include <sys/ioctl.h>
32#include <sys/mman.h>
33
34#include "fio.h"
35#include "os.h"
36
37#define MASK (4095)
38
39#define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
40
41int groupid = 0;
42int thread_number = 0;
43static char run_str[MAX_JOBS + 1];
44int shm_id = 0;
5289b847 45static struct timeval genesis;
88c6ed80 46static int temp_stall_ts;
ebac4655 47
ebac4655
JA
48static void print_thread_status(void);
49
c04f7ec3
JA
50extern unsigned long long mlock_size;
51
ebac4655 52/*
79809113
JA
53 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
54 * will never back again. It may cycle between running/verififying/fsyncing.
55 * Once the thread reaches TD_EXITED, it is just waiting for the core to
56 * reap it.
ebac4655
JA
57 */
58enum {
59 TD_NOT_CREATED = 0,
60 TD_CREATED,
75154845 61 TD_INITIALIZED,
ebac4655
JA
62 TD_RUNNING,
63 TD_VERIFYING,
5853e5a8 64 TD_FSYNCING,
ebac4655
JA
65 TD_EXITED,
66 TD_REAPED,
67};
68
3d60d1ed 69#define should_fsync(td) ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
ebac4655 70
bbfd6b00 71static volatile int startup_sem;
ebac4655
JA
72
73#define TERMINATE_ALL (-1)
75154845 74#define JOB_START_TIMEOUT (5 * 1000)
ebac4655
JA
75
76static void terminate_threads(int group_id)
77{
78 int i;
79
80 for (i = 0; i < thread_number; i++) {
81 struct thread_data *td = &threads[i];
82
83 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
84 td->terminate = 1;
85 td->start_delay = 0;
86 }
87 }
88}
89
90static void sig_handler(int sig)
91{
92 switch (sig) {
93 case SIGALRM:
94 update_io_ticks();
95 disk_util_timer_arm();
96 print_thread_status();
97 break;
98 default:
99 printf("\nfio: terminating on signal\n");
100 fflush(stdout);
101 terminate_threads(TERMINATE_ALL);
102 break;
103 }
104}
105
906c8d75
JA
106/*
107 * The ->file_map[] contains a map of blocks we have or have not done io
108 * to yet. Used to make sure we cover the entire range in a fair fashion.
109 */
ebac4655
JA
110static int random_map_free(struct thread_data *td, unsigned long long block)
111{
112 unsigned int idx = RAND_MAP_IDX(td, block);
113 unsigned int bit = RAND_MAP_BIT(td, block);
114
115 return (td->file_map[idx] & (1UL << bit)) == 0;
116}
117
906c8d75
JA
118/*
119 * Return the next free block in the map.
120 */
ebac4655
JA
121static int get_next_free_block(struct thread_data *td, unsigned long long *b)
122{
123 int i;
124
125 *b = 0;
126 i = 0;
127 while ((*b) * td->min_bs < td->io_size) {
128 if (td->file_map[i] != -1UL) {
129 *b += ffz(td->file_map[i]);
130 return 0;
131 }
132
133 *b += BLOCKS_PER_MAP;
134 i++;
135 }
136
137 return 1;
138}
139
906c8d75
JA
140/*
141 * Mark a given offset as used in the map.
142 */
ebac4655
JA
143static void mark_random_map(struct thread_data *td, struct io_u *io_u)
144{
200bc855 145 unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
ebac4655
JA
146 unsigned int blocks = 0;
147
148 while (blocks < (io_u->buflen / td->min_bs)) {
149 unsigned int idx, bit;
150
151 if (!random_map_free(td, block))
152 break;
153
154 idx = RAND_MAP_IDX(td, block);
155 bit = RAND_MAP_BIT(td, block);
156
157 assert(idx < td->num_maps);
158
159 td->file_map[idx] |= (1UL << bit);
160 block++;
161 blocks++;
162 }
163
164 if ((blocks * td->min_bs) < io_u->buflen)
165 io_u->buflen = blocks * td->min_bs;
166}
167
906c8d75
JA
168/*
169 * For random io, generate a random new block and see if it's used. Repeat
170 * until we find a free one. For sequential io, just return the end of
171 * the last io issued.
172 */
20dc95c4
JA
173static int get_next_offset(struct thread_data *td, unsigned long long *offset)
174{
175 unsigned long long b, rb;
176 long r;
177
178 if (!td->sequential) {
085227ad 179 unsigned long long max_blocks = td->io_size / td->min_bs;
20dc95c4
JA
180 int loops = 50;
181
182 do {
6dfd46b9 183 r = os_random_long(&td->random_state);
085227ad 184 b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
20dc95c4
JA
185 rb = b + (td->file_offset / td->min_bs);
186 loops--;
187 } while (!random_map_free(td, rb) && loops);
188
189 if (!loops) {
190 if (get_next_free_block(td, &b))
191 return 1;
192 }
193 } else
194 b = td->last_pos / td->min_bs;
195
196 *offset = (b * td->min_bs) + td->file_offset;
838a3cd3 197 if (*offset > td->real_file_size)
20dc95c4
JA
198 return 1;
199
200 return 0;
201}
202
203static unsigned int get_next_buflen(struct thread_data *td)
204{
205 unsigned int buflen;
206 long r;
207
208 if (td->min_bs == td->max_bs)
209 buflen = td->min_bs;
210 else {
6dfd46b9 211 r = os_random_long(&td->bsrange_state);
20dc95c4
JA
212 buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
213 buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
214 }
215
216 if (buflen > td->io_size - td->this_io_bytes[td->ddir])
217 buflen = td->io_size - td->this_io_bytes[td->ddir];
218
219 return buflen;
220}
221
906c8d75
JA
222/*
223 * Check if we are above the minimum rate given.
224 */
ebac4655
JA
225static int check_min_rate(struct thread_data *td, struct timeval *now)
226{
227 unsigned long spent;
228 unsigned long rate;
229 int ddir = td->ddir;
230
231 /*
232 * allow a 2 second settle period in the beginning
233 */
234 if (mtime_since(&td->start, now) < 2000)
235 return 0;
236
237 /*
238 * if rate blocks is set, sample is running
239 */
240 if (td->rate_bytes) {
241 spent = mtime_since(&td->lastrate, now);
242 if (spent < td->ratecycle)
243 return 0;
244
245 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
246 if (rate < td->ratemin) {
eb8bbf48 247 fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
ebac4655
JA
248 if (rate_quit)
249 terminate_threads(td->groupid);
250 return 1;
251 }
252 }
253
254 td->rate_bytes = td->this_io_bytes[ddir];
255 memcpy(&td->lastrate, now, sizeof(*now));
256 return 0;
257}
258
259static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
260{
261 if (!td->timeout)
262 return 0;
263 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
264 return 1;
265
266 return 0;
267}
268
269static void fill_random_bytes(struct thread_data *td,
270 unsigned char *p, unsigned int len)
271{
272 unsigned int todo;
273 double r;
274
275 while (len) {
6dfd46b9 276 r = os_random_double(&td->verify_state);
ebac4655
JA
277
278 /*
279 * lrand48_r seems to be broken and only fill the bottom
280 * 32-bits, even on 64-bit archs with 64-bit longs
281 */
282 todo = sizeof(r);
283 if (todo > len)
284 todo = len;
285
286 memcpy(p, &r, todo);
287
288 len -= todo;
289 p += todo;
290 }
291}
292
293static void hexdump(void *buffer, int len)
294{
295 unsigned char *p = buffer;
296 int i;
297
298 for (i = 0; i < len; i++)
eb8bbf48
JA
299 fprintf(f_out, "%02x", p[i]);
300 fprintf(f_out, "\n");
ebac4655
JA
301}
302
303static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
304{
305 unsigned char *p = (unsigned char *) io_u->buf;
306 unsigned long c;
ebac4655
JA
307
308 p += sizeof(*hdr);
309 c = crc32(p, hdr->len - sizeof(*hdr));
ebac4655 310
22f78b32 311 if (c != hdr->crc32) {
3b70d7e5
JA
312 log_err("crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
313 log_err("crc32: wanted %lx, got %lx\n", hdr->crc32, c);
22f78b32 314 return 1;
ebac4655
JA
315 }
316
22f78b32 317 return 0;
ebac4655
JA
318}
319
320static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
321{
322 unsigned char *p = (unsigned char *) io_u->buf;
323 struct md5_ctx md5_ctx;
ebac4655
JA
324
325 memset(&md5_ctx, 0, sizeof(md5_ctx));
326 p += sizeof(*hdr);
327 md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
328
22f78b32 329 if (memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash))) {
3b70d7e5 330 log_err("md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen);
ebac4655
JA
331 hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
332 hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
22f78b32 333 return 1;
ebac4655
JA
334 }
335
22f78b32 336 return 0;
ebac4655
JA
337}
338
339static int verify_io_u(struct io_u *io_u)
340{
341 struct verify_header *hdr = (struct verify_header *) io_u->buf;
342 int ret;
343
344 if (hdr->fio_magic != FIO_HDR_MAGIC)
345 return 1;
346
347 if (hdr->verify_type == VERIFY_MD5)
348 ret = verify_io_u_md5(hdr, io_u);
349 else if (hdr->verify_type == VERIFY_CRC32)
350 ret = verify_io_u_crc32(hdr, io_u);
351 else {
3b70d7e5 352 log_err("Bad verify type %d\n", hdr->verify_type);
ebac4655
JA
353 ret = 1;
354 }
355
356 return ret;
357}
358
359static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
360{
361 hdr->crc32 = crc32(p, len);
362}
363
364static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
365{
366 struct md5_ctx md5_ctx;
367
368 memset(&md5_ctx, 0, sizeof(md5_ctx));
369 md5_update(&md5_ctx, p, len);
370 memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
371}
372
906c8d75
JA
373/*
374 * Return the data direction for the next io_u. If the job is a
375 * mixed read/write workload, check the rwmix cycle and switch if
376 * necessary.
377 */
a6ccc7be 378static int get_rw_ddir(struct thread_data *td)
3d60d1ed 379{
a6ccc7be
JA
380 if (td_rw(td)) {
381 struct timeval now;
382 unsigned long elapsed;
3d60d1ed 383
a6ccc7be
JA
384 gettimeofday(&now, NULL);
385 elapsed = mtime_since_now(&td->rwmix_switch);
3d60d1ed 386
a6ccc7be
JA
387 /*
388 * Check if it's time to seed a new data direction.
389 */
390 if (elapsed >= td->rwmixcycle) {
c1ee2ca4 391 int v;
a6ccc7be 392 long r;
3d60d1ed 393
c1ee2ca4
JA
394 r = os_random_long(&td->rwmix_state);
395 v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
a6ccc7be
JA
396 if (v < td->rwmixread)
397 td->rwmix_ddir = DDIR_READ;
398 else
399 td->rwmix_ddir = DDIR_WRITE;
400 memcpy(&td->rwmix_switch, &now, sizeof(now));
401 }
402 return td->rwmix_ddir;
403 } else if (td_read(td))
3d60d1ed
JA
404 return DDIR_READ;
405 else
406 return DDIR_WRITE;
407}
408
ebac4655
JA
409/*
410 * fill body of io_u->buf with random data and add a header with the
22f78b32 411 * crc32 or md5 sum of that data.
ebac4655
JA
412 */
413static void populate_io_u(struct thread_data *td, struct io_u *io_u)
414{
415 unsigned char *p = (unsigned char *) io_u->buf;
416 struct verify_header hdr;
417
418 hdr.fio_magic = FIO_HDR_MAGIC;
419 hdr.len = io_u->buflen;
420 p += sizeof(hdr);
421 fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
422
423 if (td->verify == VERIFY_MD5) {
424 fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
425 hdr.verify_type = VERIFY_MD5;
426 } else {
427 fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
428 hdr.verify_type = VERIFY_CRC32;
429 }
430
431 memcpy(io_u->buf, &hdr, sizeof(hdr));
432}
433
aea47d44 434static int td_io_prep(struct thread_data *td, struct io_u *io_u)
20dc95c4 435{
20dc95c4
JA
436 if (td->io_prep && td->io_prep(td, io_u))
437 return 1;
438
439 return 0;
440}
441
b1ff3403 442void put_io_u(struct thread_data *td, struct io_u *io_u)
ebac4655
JA
443{
444 list_del(&io_u->list);
445 list_add(&io_u->list, &td->io_u_freelist);
446 td->cur_depth--;
447}
448
843a7413
JA
449static int fill_io_u(struct thread_data *td, struct io_u *io_u)
450{
451 /*
452 * If using an iolog, grab next piece if any available.
453 */
454 if (td->read_iolog)
455 return read_iolog_get(td, io_u);
456
aea47d44
JA
457 /*
458 * No log, let the seq/rand engine retrieve the next position.
459 */
460 if (!get_next_offset(td, &io_u->offset)) {
461 io_u->buflen = get_next_buflen(td);
462
463 if (io_u->buflen) {
464 io_u->ddir = get_rw_ddir(td);
843a7413
JA
465
466 /*
467 * If using a write iolog, store this entry.
468 */
469 if (td->write_iolog)
470 write_iolog_put(td, io_u);
471
aea47d44
JA
472 return 0;
473 }
474 }
475
476 return 1;
477}
478
22f78b32 479#define queue_full(td) list_empty(&(td)->io_u_freelist)
ebac4655 480
b1ff3403 481struct io_u *__get_io_u(struct thread_data *td)
ebac4655 482{
22f78b32 483 struct io_u *io_u = NULL;
ebac4655 484
22f78b32
JA
485 if (!queue_full(td)) {
486 io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
487
488 io_u->error = 0;
489 io_u->resid = 0;
490 list_del(&io_u->list);
491 list_add(&io_u->list, &td->io_u_busylist);
492 td->cur_depth++;
493 }
ebac4655 494
ebac4655
JA
495 return io_u;
496}
497
906c8d75
JA
498/*
499 * Return an io_u to be processed. Gets a buflen and offset, sets direction,
500 * etc. The returned io_u is fully ready to be prepped and submitted.
501 */
ebac4655
JA
502static struct io_u *get_io_u(struct thread_data *td)
503{
504 struct io_u *io_u;
505
506 io_u = __get_io_u(td);
507 if (!io_u)
508 return NULL;
509
20dc95c4
JA
510 if (td->zone_bytes >= td->zone_size) {
511 td->zone_bytes = 0;
512 td->last_pos += td->zone_skip;
513 }
514
aea47d44 515 if (fill_io_u(td, io_u)) {
ebac4655
JA
516 put_io_u(td, io_u);
517 return NULL;
518 }
519
838a3cd3
JA
520 if (io_u->buflen + io_u->offset > td->real_file_size)
521 io_u->buflen = td->real_file_size - io_u->offset;
ebac4655
JA
522
523 if (!io_u->buflen) {
524 put_io_u(td, io_u);
525 return NULL;
526 }
527
843a7413 528 if (!td->read_iolog && !td->sequential)
ebac4655
JA
529 mark_random_map(td, io_u);
530
20dc95c4 531 td->last_pos += io_u->buflen;
ebac4655
JA
532
533 if (td->verify != VERIFY_NONE)
534 populate_io_u(td, io_u);
535
aea47d44 536 if (td_io_prep(td, io_u)) {
ebac4655
JA
537 put_io_u(td, io_u);
538 return NULL;
539 }
540
541 gettimeofday(&io_u->start_time, NULL);
542 return io_u;
543}
544
545static inline void td_set_runstate(struct thread_data *td, int runstate)
546{
ebac4655
JA
547 td->runstate = runstate;
548}
549
aea47d44 550static int get_next_verify(struct thread_data *td, struct io_u *io_u)
ebac4655
JA
551{
552 struct io_piece *ipo;
553
22f78b32
JA
554 if (!list_empty(&td->io_hist_list)) {
555 ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
ebac4655 556
22f78b32 557 list_del(&ipo->list);
ebac4655 558
22f78b32
JA
559 io_u->offset = ipo->offset;
560 io_u->buflen = ipo->len;
561 io_u->ddir = DDIR_READ;
562 free(ipo);
563 return 0;
564 }
565
566 return 1;
ebac4655
JA
567}
568
ebac4655
JA
569static int sync_td(struct thread_data *td)
570{
571 if (td->io_sync)
572 return td->io_sync(td);
573
574 return 0;
575}
576
577static int io_u_getevents(struct thread_data *td, int min, int max,
578 struct timespec *t)
579{
580 return td->io_getevents(td, min, max, t);
581}
582
583static int io_u_queue(struct thread_data *td, struct io_u *io_u)
584{
585 gettimeofday(&io_u->issue_time, NULL);
586
587 return td->io_queue(td, io_u);
588}
589
590#define iocb_time(iocb) ((unsigned long) (iocb)->data)
591
592static void io_completed(struct thread_data *td, struct io_u *io_u,
593 struct io_completion_data *icd)
594{
595 struct timeval e;
596 unsigned long msec;
597
598 gettimeofday(&e, NULL);
599
600 if (!io_u->error) {
20dc95c4 601 unsigned int bytes = io_u->buflen - io_u->resid;
3d60d1ed 602 const int idx = io_u->ddir;
ebac4655
JA
603
604 td->io_blocks[idx]++;
20dc95c4
JA
605 td->io_bytes[idx] += bytes;
606 td->zone_bytes += bytes;
607 td->this_io_bytes[idx] += bytes;
ebac4655
JA
608
609 msec = mtime_since(&io_u->issue_time, &e);
610
3d60d1ed
JA
611 add_clat_sample(td, idx, msec);
612 add_bw_sample(td, idx);
ebac4655 613
3d60d1ed 614 if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
ebac4655
JA
615 log_io_piece(td, io_u);
616
20dc95c4 617 icd->bytes_done[idx] += bytes;
ebac4655
JA
618 } else
619 icd->error = io_u->error;
620}
621
622static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
623{
624 struct io_u *io_u;
625 int i;
626
627 icd->error = 0;
628 icd->bytes_done[0] = icd->bytes_done[1] = 0;
629
630 for (i = 0; i < icd->nr; i++) {
631 io_u = td->io_event(td, i);
632
633 io_completed(td, io_u, icd);
634 put_io_u(td, io_u);
635 }
636}
637
906c8d75
JA
638/*
639 * When job exits, we can cancel the in-flight IO if we are using async
640 * io. Attempt to do so.
641 */
ebac4655
JA
642static void cleanup_pending_aio(struct thread_data *td)
643{
644 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
645 struct list_head *entry, *n;
646 struct io_completion_data icd;
647 struct io_u *io_u;
648 int r;
649
650 /*
651 * get immediately available events, if any
652 */
653 r = io_u_getevents(td, 0, td->cur_depth, &ts);
654 if (r > 0) {
655 icd.nr = r;
656 ios_completed(td, &icd);
657 }
658
659 /*
660 * now cancel remaining active events
661 */
662 if (td->io_cancel) {
663 list_for_each_safe(entry, n, &td->io_u_busylist) {
664 io_u = list_entry(entry, struct io_u, list);
665
666 r = td->io_cancel(td, io_u);
667 if (!r)
668 put_io_u(td, io_u);
669 }
670 }
671
672 if (td->cur_depth) {
673 r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
674 if (r > 0) {
675 icd.nr = r;
676 ios_completed(td, &icd);
677 }
678 }
679}
680
681static int do_io_u_verify(struct thread_data *td, struct io_u **io_u)
682{
683 struct io_u *v_io_u = *io_u;
684 int ret = 0;
685
686 if (v_io_u) {
687 ret = verify_io_u(v_io_u);
688 put_io_u(td, v_io_u);
689 *io_u = NULL;
690 }
691
692 return ret;
693}
694
906c8d75
JA
695/*
696 * The main verify engine. Runs over the writes we previusly submitted,
697 * reads the blocks back in, and checks the crc/md5 of the data.
698 */
ebac4655
JA
699static void do_verify(struct thread_data *td)
700{
701 struct timeval t;
702 struct io_u *io_u, *v_io_u = NULL;
703 struct io_completion_data icd;
704 int ret;
705
706 td_set_runstate(td, TD_VERIFYING);
707
708 do {
709 if (td->terminate)
710 break;
711
712 gettimeofday(&t, NULL);
713 if (runtime_exceeded(td, &t))
714 break;
715
716 io_u = __get_io_u(td);
717 if (!io_u)
718 break;
719
aea47d44 720 if (get_next_verify(td, io_u)) {
ebac4655
JA
721 put_io_u(td, io_u);
722 break;
723 }
724
aea47d44 725 if (td_io_prep(td, io_u)) {
ebac4655
JA
726 put_io_u(td, io_u);
727 break;
728 }
729
730 ret = io_u_queue(td, io_u);
731 if (ret) {
732 put_io_u(td, io_u);
733 td_verror(td, ret);
734 break;
735 }
736
737 /*
738 * we have one pending to verify, do that while
739 * we are doing io on the next one
740 */
741 if (do_io_u_verify(td, &v_io_u))
742 break;
743
744 ret = io_u_getevents(td, 1, 1, NULL);
745 if (ret != 1) {
746 if (ret < 0)
747 td_verror(td, ret);
748 break;
749 }
750
751 v_io_u = td->io_event(td, 0);
752 icd.nr = 1;
753 icd.error = 0;
754 io_completed(td, v_io_u, &icd);
755
756 if (icd.error) {
757 td_verror(td, icd.error);
758 put_io_u(td, v_io_u);
759 v_io_u = NULL;
760 break;
761 }
762
ebac4655
JA
763 /*
764 * if we can't submit more io, we need to verify now
765 */
766 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
767 break;
768
769 } while (1);
770
771 do_io_u_verify(td, &v_io_u);
772
773 if (td->cur_depth)
774 cleanup_pending_aio(td);
775
776 td_set_runstate(td, TD_RUNNING);
777}
778
32cd46a0 779/*
906c8d75 780 * Main IO worker function. It retrieves io_u's to process and queues
32cd46a0
JA
781 * and reaps them, checking for rate and errors along the way.
782 */
ebac4655
JA
783static void do_io(struct thread_data *td)
784{
785 struct io_completion_data icd;
786 struct timeval s, e;
787 unsigned long usec;
788
5853e5a8
JA
789 td_set_runstate(td, TD_RUNNING);
790
ebac4655
JA
791 while (td->this_io_bytes[td->ddir] < td->io_size) {
792 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
793 struct timespec *timeout;
794 int ret, min_evts = 0;
795 struct io_u *io_u;
796
797 if (td->terminate)
798 break;
799
800 io_u = get_io_u(td);
801 if (!io_u)
802 break;
803
804 memcpy(&s, &io_u->start_time, sizeof(s));
805
806 ret = io_u_queue(td, io_u);
807 if (ret) {
808 put_io_u(td, io_u);
809 td_verror(td, ret);
810 break;
811 }
812
813 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
814
815 if (td->cur_depth < td->iodepth) {
816 timeout = &ts;
817 min_evts = 0;
818 } else {
819 timeout = NULL;
820 min_evts = 1;
821 }
822
823 ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
824 if (ret < 0) {
825 td_verror(td, ret);
826 break;
827 } else if (!ret)
828 continue;
829
830 icd.nr = ret;
831 ios_completed(td, &icd);
832 if (icd.error) {
833 td_verror(td, icd.error);
834 break;
835 }
836
837 /*
838 * the rate is batched for now, it should work for batches
839 * of completions except the very first one which may look
840 * a little bursty
841 */
842 gettimeofday(&e, NULL);
843 usec = utime_since(&s, &e);
844
845 rate_throttle(td, usec, icd.bytes_done[td->ddir]);
846
847 if (check_min_rate(td, &e)) {
848 td_verror(td, ENOMEM);
849 break;
850 }
851
852 if (runtime_exceeded(td, &e))
853 break;
854
855 if (td->thinktime)
856 usec_sleep(td, td->thinktime);
857
858 if (should_fsync(td) && td->fsync_blocks &&
859 (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
860 sync_td(td);
861 }
862
863 if (td->cur_depth)
864 cleanup_pending_aio(td);
865
5853e5a8
JA
866 if (should_fsync(td) && td->end_fsync) {
867 td_set_runstate(td, TD_FSYNCING);
ebac4655 868 sync_td(td);
5853e5a8 869 }
ebac4655
JA
870}
871
872static void cleanup_io(struct thread_data *td)
873{
874 if (td->io_cleanup)
875 td->io_cleanup(td);
876}
877
878static int init_io(struct thread_data *td)
879{
880 if (td->io_engine == FIO_SYNCIO)
881 return fio_syncio_init(td);
882 else if (td->io_engine == FIO_MMAPIO)
883 return fio_mmapio_init(td);
884 else if (td->io_engine == FIO_LIBAIO)
885 return fio_libaio_init(td);
886 else if (td->io_engine == FIO_POSIXAIO)
887 return fio_posixaio_init(td);
888 else if (td->io_engine == FIO_SGIO)
889 return fio_sgio_init(td);
8756e4d4
JA
890 else if (td->io_engine == FIO_SPLICEIO)
891 return fio_spliceio_init(td);
ebac4655 892 else {
3b70d7e5 893 log_err("bad io_engine %d\n", td->io_engine);
ebac4655
JA
894 return 1;
895 }
896}
897
898static void cleanup_io_u(struct thread_data *td)
899{
900 struct list_head *entry, *n;
901 struct io_u *io_u;
902
903 list_for_each_safe(entry, n, &td->io_u_freelist) {
904 io_u = list_entry(entry, struct io_u, list);
905
906 list_del(&io_u->list);
907 free(io_u);
908 }
909
910 if (td->mem_type == MEM_MALLOC)
911 free(td->orig_buffer);
912 else if (td->mem_type == MEM_SHM) {
913 struct shmid_ds sbuf;
914
915 shmdt(td->orig_buffer);
916 shmctl(td->shm_id, IPC_RMID, &sbuf);
917 } else if (td->mem_type == MEM_MMAP)
918 munmap(td->orig_buffer, td->orig_buffer_size);
919 else
3b70d7e5 920 log_err("Bad memory type %d\n", td->mem_type);
ebac4655
JA
921
922 td->orig_buffer = NULL;
923}
924
925static int init_io_u(struct thread_data *td)
926{
927 struct io_u *io_u;
928 int i, max_units;
929 char *p;
930
931 if (td->io_engine & FIO_SYNCIO)
932 max_units = 1;
933 else
934 max_units = td->iodepth;
935
936 td->orig_buffer_size = td->max_bs * max_units + MASK;
937
938 if (td->mem_type == MEM_MALLOC)
939 td->orig_buffer = malloc(td->orig_buffer_size);
940 else if (td->mem_type == MEM_SHM) {
941 td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
942 if (td->shm_id < 0) {
943 td_verror(td, errno);
944 perror("shmget");
945 return 1;
946 }
947
948 td->orig_buffer = shmat(td->shm_id, NULL, 0);
949 if (td->orig_buffer == (void *) -1) {
950 td_verror(td, errno);
951 perror("shmat");
952 td->orig_buffer = NULL;
953 return 1;
954 }
955 } else if (td->mem_type == MEM_MMAP) {
956 td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
957 if (td->orig_buffer == MAP_FAILED) {
958 td_verror(td, errno);
959 perror("mmap");
960 td->orig_buffer = NULL;
961 return 1;
962 }
963 }
964
ebac4655
JA
965 p = ALIGN(td->orig_buffer);
966 for (i = 0; i < max_units; i++) {
967 io_u = malloc(sizeof(*io_u));
968 memset(io_u, 0, sizeof(*io_u));
969 INIT_LIST_HEAD(&io_u->list);
970
971 io_u->buf = p + td->max_bs * i;
b1ff3403 972 io_u->index = i;
ebac4655
JA
973 list_add(&io_u->list, &td->io_u_freelist);
974 }
975
976 return 0;
977}
978
9cc935a1 979static int create_file(struct thread_data *td, unsigned long long size)
ebac4655
JA
980{
981 unsigned long long left;
982 unsigned int bs;
ebac4655 983 char *b;
9cc935a1 984 int r;
ebac4655
JA
985
986 /*
987 * unless specifically asked for overwrite, let normal io extend it
988 */
9cc935a1 989 if (!td->overwrite) {
c20d25b4 990 td->real_file_size = size;
ebac4655 991 return 0;
c20d25b4 992 }
ebac4655
JA
993
994 if (!size) {
3b70d7e5 995 log_err("Need size for create\n");
ebac4655
JA
996 td_verror(td, EINVAL);
997 return 1;
998 }
999
88c6ed80 1000 temp_stall_ts = 1;
9cc935a1 1001 fprintf(f_out, "%s: Laying out IO file (%LuMiB)\n",td->name,size >> 20);
88c6ed80 1002
9cc935a1 1003 td->fd = open(td->file_name, O_WRONLY | O_CREAT | O_TRUNC, 0644);
ebac4655
JA
1004 if (td->fd < 0) {
1005 td_verror(td, errno);
c20d25b4 1006 goto done_noclose;
ebac4655
JA
1007 }
1008
9cc935a1 1009 if (ftruncate(td->fd, td->file_size) == -1) {
ebac4655 1010 td_verror(td, errno);
c20d25b4 1011 goto done;
ebac4655
JA
1012 }
1013
1014 td->io_size = td->file_size;
1015 b = malloc(td->max_bs);
1016 memset(b, 0, td->max_bs);
1017
1018 left = size;
1019 while (left && !td->terminate) {
1020 bs = td->max_bs;
1021 if (bs > left)
1022 bs = left;
1023
1024 r = write(td->fd, b, bs);
1025
1026 if (r == (int) bs) {
1027 left -= bs;
1028 continue;
1029 } else {
1030 if (r < 0)
1031 td_verror(td, errno);
1032 else
1033 td_verror(td, EIO);
1034
1035 break;
1036 }
1037 }
1038
1039 if (td->terminate)
1040 unlink(td->file_name);
1041 else if (td->create_fsync)
1042 fsync(td->fd);
1043
c20d25b4
JA
1044 free(b);
1045done:
ebac4655
JA
1046 close(td->fd);
1047 td->fd = -1;
c20d25b4
JA
1048done_noclose:
1049 temp_stall_ts = 0;
ebac4655
JA
1050 return 0;
1051}
1052
1053static int file_size(struct thread_data *td)
1054{
1055 struct stat st;
1056
c20d25b4
JA
1057 if (td->overwrite) {
1058 if (fstat(td->fd, &st) == -1) {
1059 td_verror(td, errno);
1060 return 1;
1061 }
ebac4655 1062
c20d25b4 1063 td->real_file_size = st.st_size;
838a3cd3 1064
c20d25b4
JA
1065 if (!td->file_size || td->file_size > td->real_file_size)
1066 td->file_size = td->real_file_size;
1067 }
ebac4655 1068
e8e387c1 1069 td->file_size -= td->file_offset;
ebac4655
JA
1070 return 0;
1071}
1072
1073static int bdev_size(struct thread_data *td)
1074{
9104f874 1075 unsigned long long bytes;
ebac4655
JA
1076 int r;
1077
1078 r = blockdev_size(td->fd, &bytes);
1079 if (r) {
1080 td_verror(td, r);
1081 return 1;
1082 }
1083
838a3cd3
JA
1084 td->real_file_size = bytes;
1085
ebac4655
JA
1086 /*
1087 * no extend possibilities, so limit size to device size if too large
1088 */
838a3cd3
JA
1089 if (!td->file_size || td->file_size > td->real_file_size)
1090 td->file_size = td->real_file_size;
ebac4655 1091
e8e387c1 1092 td->file_size -= td->file_offset;
ebac4655
JA
1093 return 0;
1094}
1095
1096static int get_file_size(struct thread_data *td)
1097{
0af7b542 1098 int ret = 0;
ebac4655
JA
1099
1100 if (td->filetype == FIO_TYPE_FILE)
1101 ret = file_size(td);
0af7b542 1102 else if (td->filetype == FIO_TYPE_BD)
ebac4655 1103 ret = bdev_size(td);
0af7b542
JA
1104 else
1105 td->real_file_size = -1;
ebac4655
JA
1106
1107 if (ret)
1108 return ret;
1109
e8e387c1 1110 if (td->file_offset > td->real_file_size) {
3b70d7e5 1111 log_err("%s: offset extends end (%Lu > %Lu)\n", td->name, td->file_offset, td->real_file_size);
ebac4655
JA
1112 return 1;
1113 }
1114
838a3cd3 1115 td->io_size = td->file_size;
ebac4655 1116 if (td->io_size == 0) {
3b70d7e5 1117 log_err("%s: no io blocks\n", td->name);
ebac4655
JA
1118 td_verror(td, EINVAL);
1119 return 1;
1120 }
1121
20dc95c4
JA
1122 if (!td->zone_size)
1123 td->zone_size = td->io_size;
1124
ebac4655
JA
1125 td->total_io_size = td->io_size * td->loops;
1126 return 0;
1127}
1128
1129static int setup_file_mmap(struct thread_data *td)
1130{
1131 int flags;
1132
3d60d1ed
JA
1133 if (td_rw(td))
1134 flags = PROT_READ | PROT_WRITE;
1135 else if (td_write(td)) {
ebac4655
JA
1136 flags = PROT_WRITE;
1137
1138 if (td->verify != VERIFY_NONE)
1139 flags |= PROT_READ;
3d60d1ed
JA
1140 } else
1141 flags = PROT_READ;
ebac4655
JA
1142
1143 td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
1144 if (td->mmap == MAP_FAILED) {
1145 td->mmap = NULL;
1146 td_verror(td, errno);
1147 return 1;
1148 }
1149
1150 if (td->invalidate_cache) {
1151 if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
1152 td_verror(td, errno);
1153 return 1;
1154 }
1155 }
1156
1157 if (td->sequential) {
1158 if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
1159 td_verror(td, errno);
1160 return 1;
1161 }
1162 } else {
1163 if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
1164 td_verror(td, errno);
1165 return 1;
1166 }
1167 }
1168
1169 return 0;
1170}
1171
1172static int setup_file_plain(struct thread_data *td)
1173{
1174 if (td->invalidate_cache) {
1175 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
1176 td_verror(td, errno);
1177 return 1;
1178 }
1179 }
1180
1181 if (td->sequential) {
1182 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
1183 td_verror(td, errno);
1184 return 1;
1185 }
1186 } else {
1187 if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
1188 td_verror(td, errno);
1189 return 1;
1190 }
1191 }
1192
1193 return 0;
1194}
1195
1196static int setup_file(struct thread_data *td)
1197{
1198 struct stat st;
1199 int flags = 0;
1200
1201 if (stat(td->file_name, &st) == -1) {
1202 if (errno != ENOENT) {
1203 td_verror(td, errno);
1204 return 1;
1205 }
1206 if (!td->create_file) {
1207 td_verror(td, ENOENT);
1208 return 1;
1209 }
9cc935a1
JA
1210 if (create_file(td, td->file_size))
1211 return 1;
1212 } else if (td->filetype == FIO_TYPE_FILE &&
1213 st.st_size < (off_t) td->file_size) {
1214 if (create_file(td, td->file_size))
ebac4655 1215 return 1;
ebac4655
JA
1216 }
1217
1218 if (td->odirect)
2c0ecd28 1219 flags |= OS_O_DIRECT;
ebac4655 1220
3d60d1ed 1221 if (td_write(td) || td_rw(td)) {
ebac4655
JA
1222 if (td->filetype == FIO_TYPE_FILE) {
1223 if (!td->overwrite)
1224 flags |= O_TRUNC;
1225
1226 flags |= O_CREAT;
1227 }
1228 if (td->sync_io)
1229 flags |= O_SYNC;
1230
1231 flags |= O_RDWR;
1232
1233 td->fd = open(td->file_name, flags, 0600);
3d60d1ed
JA
1234 } else {
1235 if (td->filetype == FIO_TYPE_CHAR)
1236 flags |= O_RDWR;
1237 else
1238 flags |= O_RDONLY;
1239
1240 td->fd = open(td->file_name, flags);
ebac4655
JA
1241 }
1242
1243 if (td->fd == -1) {
1244 td_verror(td, errno);
1245 return 1;
1246 }
1247
1248 if (get_file_size(td))
1249 return 1;
1250
1251 if (td->io_engine != FIO_MMAPIO)
1252 return setup_file_plain(td);
1253 else
1254 return setup_file_mmap(td);
1255}
1256
da86774e
JA
1257static int switch_ioscheduler(struct thread_data *td)
1258{
1259 char tmp[256], tmp2[128];
1260 FILE *f;
1261 int ret;
1262
1263 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
1264
1265 f = fopen(tmp, "r+");
1266 if (!f) {
1267 td_verror(td, errno);
1268 return 1;
1269 }
1270
1271 /*
1272 * Set io scheduler.
1273 */
1274 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
1275 if (ferror(f) || ret != 1) {
1276 td_verror(td, errno);
1277 fclose(f);
1278 return 1;
1279 }
1280
1281 rewind(f);
1282
1283 /*
1284 * Read back and check that the selected scheduler is now the default.
1285 */
1286 ret = fread(tmp, 1, sizeof(tmp), f);
1287 if (ferror(f) || ret < 0) {
1288 td_verror(td, errno);
1289 fclose(f);
1290 return 1;
1291 }
1292
1293 sprintf(tmp2, "[%s]", td->ioscheduler);
1294 if (!strstr(tmp, tmp2)) {
3b70d7e5 1295 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
da86774e
JA
1296 td_verror(td, EINVAL);
1297 fclose(f);
1298 return 1;
1299 }
1300
1301 fclose(f);
1302 return 0;
1303}
1304
ebac4655
JA
1305static void clear_io_state(struct thread_data *td)
1306{
1307 if (td->io_engine == FIO_SYNCIO)
1308 lseek(td->fd, SEEK_SET, 0);
1309
20dc95c4 1310 td->last_pos = 0;
ebac4655
JA
1311 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
1312 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
20dc95c4 1313 td->zone_bytes = 0;
ebac4655
JA
1314
1315 if (td->file_map)
1316 memset(td->file_map, 0, td->num_maps * sizeof(long));
1317}
1318
906c8d75
JA
1319/*
1320 * Entry point for the thread based jobs. The process based jobs end up
1321 * here as well, after a little setup.
1322 */
ebac4655
JA
1323static void *thread_main(void *data)
1324{
1325 struct thread_data *td = data;
ebac4655
JA
1326
1327 if (!td->use_thread)
1328 setsid();
1329
1330 td->pid = getpid();
1331
aea47d44
JA
1332 INIT_LIST_HEAD(&td->io_u_freelist);
1333 INIT_LIST_HEAD(&td->io_u_busylist);
1334 INIT_LIST_HEAD(&td->io_hist_list);
1335 INIT_LIST_HEAD(&td->io_log_list);
1336
ebac4655
JA
1337 if (init_io_u(td))
1338 goto err;
1339
1340 if (fio_setaffinity(td) == -1) {
1341 td_verror(td, errno);
1342 goto err;
1343 }
1344
1345 if (init_io(td))
1346 goto err;
1347
aea47d44
JA
1348 if (init_iolog(td))
1349 goto err;
1350
ebac4655
JA
1351 if (td->ioprio) {
1352 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
1353 td_verror(td, errno);
1354 goto err;
1355 }
1356 }
1357
1056eaad 1358 if (nice(td->nice) == -1) {
b6f4d880
JA
1359 td_verror(td, errno);
1360 goto err;
1361 }
1362
75154845
JA
1363 if (init_random_state(td))
1364 goto err;
1365
da86774e
JA
1366 if (td->ioscheduler && switch_ioscheduler(td))
1367 goto err;
1368
75154845 1369 td_set_runstate(td, TD_INITIALIZED);
bbfd6b00
JA
1370 fio_sem_up(&startup_sem);
1371 fio_sem_down(&td->mutex);
ebac4655
JA
1372
1373 if (!td->create_serialize && setup_file(td))
1374 goto err;
1375
ebac4655
JA
1376 gettimeofday(&td->epoch, NULL);
1377
4e0ba8af
JA
1378 if (td->exec_prerun)
1379 system(td->exec_prerun);
1380
ebac4655
JA
1381 while (td->loops--) {
1382 getrusage(RUSAGE_SELF, &td->ru_start);
1383 gettimeofday(&td->start, NULL);
1384 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
1385
1386 if (td->ratemin)
1387 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
1388
1389 clear_io_state(td);
1390 prune_io_piece_log(td);
1391
1392 do_io(td);
1393
1394 td->runtime[td->ddir] += mtime_since_now(&td->start);
aea47d44 1395 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
3d60d1ed
JA
1396 td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
1397
ebac4655
JA
1398 update_rusage_stat(td);
1399
1400 if (td->error || td->terminate)
1401 break;
1402
1403 if (td->verify == VERIFY_NONE)
1404 continue;
1405
1406 clear_io_state(td);
1407 gettimeofday(&td->start, NULL);
1408
1409 do_verify(td);
1410
1411 td->runtime[DDIR_READ] += mtime_since_now(&td->start);
1412
1413 if (td->error || td->terminate)
1414 break;
1415 }
1416
ebac4655
JA
1417 if (td->bw_log)
1418 finish_log(td, td->bw_log, "bw");
1419 if (td->slat_log)
1420 finish_log(td, td->slat_log, "slat");
1421 if (td->clat_log)
1422 finish_log(td, td->clat_log, "clat");
843a7413
JA
1423 if (td->write_iolog)
1424 write_iolog_close(td);
4e0ba8af
JA
1425 if (td->exec_postrun)
1426 system(td->exec_postrun);
ebac4655
JA
1427
1428 if (exitall_on_terminate)
1429 terminate_threads(td->groupid);
1430
1431err:
1432 if (td->fd != -1) {
1433 close(td->fd);
1434 td->fd = -1;
1435 }
1436 if (td->mmap)
1437 munmap(td->mmap, td->file_size);
1438 cleanup_io(td);
1439 cleanup_io_u(td);
ebac4655
JA
1440 td_set_runstate(td, TD_EXITED);
1441 return NULL;
1442
1443}
1444
906c8d75
JA
1445/*
1446 * We cannot pass the td data into a forked process, so attach the td and
1447 * pass it to the thread worker.
1448 */
ebac4655
JA
1449static void *fork_main(int shmid, int offset)
1450{
1451 struct thread_data *td;
1452 void *data;
1453
1454 data = shmat(shmid, NULL, 0);
1455 if (data == (void *) -1) {
1456 perror("shmat");
1457 return NULL;
1458 }
1459
1460 td = data + offset * sizeof(struct thread_data);
1461 thread_main(td);
1462 shmdt(data);
1463 return NULL;
1464}
1465
906c8d75
JA
1466/*
1467 * Sets the status of the 'td' in the printed status map.
1468 */
ebac4655
JA
1469static void check_str_update(struct thread_data *td)
1470{
1471 char c = run_str[td->thread_number - 1];
1472
ebac4655
JA
1473 switch (td->runstate) {
1474 case TD_REAPED:
1475 c = '_';
1476 break;
1477 case TD_EXITED:
1478 c = 'E';
1479 break;
1480 case TD_RUNNING:
3d60d1ed
JA
1481 if (td_rw(td)) {
1482 if (td->sequential)
1483 c = 'M';
1484 else
1485 c = 'm';
1486 } else if (td_read(td)) {
ebac4655
JA
1487 if (td->sequential)
1488 c = 'R';
1489 else
1490 c = 'r';
1491 } else {
1492 if (td->sequential)
1493 c = 'W';
1494 else
1495 c = 'w';
1496 }
1497 break;
1498 case TD_VERIFYING:
1499 c = 'V';
1500 break;
5853e5a8
JA
1501 case TD_FSYNCING:
1502 c = 'F';
1503 break;
ebac4655
JA
1504 case TD_CREATED:
1505 c = 'C';
1506 break;
75154845
JA
1507 case TD_INITIALIZED:
1508 c = 'I';
1509 break;
ebac4655
JA
1510 case TD_NOT_CREATED:
1511 c = 'P';
1512 break;
1513 default:
3b70d7e5 1514 log_err("state %d\n", td->runstate);
ebac4655
JA
1515 }
1516
1517 run_str[td->thread_number - 1] = c;
ebac4655
JA
1518}
1519
906c8d75
JA
1520/*
1521 * Convert seconds to a printable string.
1522 */
5289b847
JA
1523static void eta_to_str(char *str, int eta_sec)
1524{
1525 unsigned int d, h, m, s;
1526 static int always_d, always_h;
1527
1528 d = h = m = s = 0;
1529
1530 s = eta_sec % 60;
1531 eta_sec /= 60;
1532 m = eta_sec % 60;
1533 eta_sec /= 60;
1534 h = eta_sec % 24;
1535 eta_sec /= 24;
1536 d = eta_sec;
1537
1538 if (d || always_d) {
1539 always_d = 1;
1540 str += sprintf(str, "%02dd:", d);
1541 }
1542 if (h || always_h) {
1543 always_h = 1;
1544 str += sprintf(str, "%02dh:", h);
1545 }
1546
1547 str += sprintf(str, "%02dm:", m);
1548 str += sprintf(str, "%02ds", s);
1549}
1550
906c8d75
JA
1551/*
1552 * Best effort calculation of the estimated pending runtime of a job.
1553 */
6a0106a0
JA
1554static int thread_eta(struct thread_data *td, unsigned long elapsed)
1555{
1556 unsigned long long bytes_total, bytes_done;
1557 unsigned int eta_sec = 0;
1558
1559 bytes_total = td->total_io_size;
3d60d1ed
JA
1560
1561 /*
1562 * if writing, bytes_total will be twice the size. If mixing,
1563 * assume a 50/50 split and thus bytes_total will be 50% larger.
1564 */
1565 if (td->verify) {
1566 if (td_rw(td))
1567 bytes_total = bytes_total * 3 / 2;
1568 else
1569 bytes_total <<= 1;
1570 }
6a0106a0
JA
1571 if (td->zone_size && td->zone_skip)
1572 bytes_total /= (td->zone_skip / td->zone_size);
1573
1574 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) {
1575 double perc;
8b611c34 1576
6a0106a0
JA
1577 bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE];
1578 perc = (double) bytes_done / (double) bytes_total;
1579 if (perc > 1.0)
1580 perc = 1.0;
1581
1582 eta_sec = (elapsed * (1.0 / perc)) - elapsed;
1583
8b611c34 1584 if (td->timeout && eta_sec > (td->timeout - elapsed))
6a0106a0 1585 eta_sec = td->timeout - elapsed;
75154845
JA
1586 } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
1587 || td->runstate == TD_INITIALIZED) {
6a0106a0
JA
1588 int t_eta = 0, r_eta = 0;
1589
1590 /*
1591 * We can only guess - assume it'll run the full timeout
1592 * if given, otherwise assume it'll run at the specified rate.
1593 */
1594 if (td->timeout)
1595 t_eta = td->timeout + td->start_delay - elapsed;
1596 if (td->rate) {
1597 r_eta = (bytes_total / 1024) / td->rate;
1598 r_eta += td->start_delay - elapsed;
1599 }
1600
1601 if (r_eta && t_eta)
1602 eta_sec = min(r_eta, t_eta);
1603 else if (r_eta)
1604 eta_sec = r_eta;
1605 else if (t_eta)
1606 eta_sec = t_eta;
1607 else
972cfd25 1608 eta_sec = 0;
6a0106a0
JA
1609 } else {
1610 /*
5853e5a8 1611 * thread is already done or waiting for fsync
6a0106a0
JA
1612 */
1613 eta_sec = 0;
1614 }
1615
1616 return eta_sec;
1617}
1618
906c8d75
JA
1619/*
1620 * Print status of the jobs we know about. This includes rate estimates,
1621 * ETA, thread state, etc.
1622 */
ebac4655
JA
1623static void print_thread_status(void)
1624{
6a0106a0 1625 unsigned long elapsed = time_since_now(&genesis);
71a751ce 1626 int i, nr_running, nr_pending, t_rate, m_rate, *eta_secs, eta_sec;
5289b847 1627 char eta_str[32];
6a0106a0
JA
1628 double perc = 0.0;
1629
c6ae0a5b 1630 if (temp_stall_ts || terse_output)
88c6ed80
JA
1631 return;
1632
6a0106a0
JA
1633 eta_secs = malloc(thread_number * sizeof(int));
1634 memset(eta_secs, 0, thread_number * sizeof(int));
ebac4655 1635
71a751ce 1636 nr_pending = nr_running = t_rate = m_rate = 0;
ebac4655
JA
1637 for (i = 0; i < thread_number; i++) {
1638 struct thread_data *td = &threads[i];
1639
5853e5a8
JA
1640 if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING||
1641 td->runstate == TD_FSYNCING) {
ebac4655
JA
1642 nr_running++;
1643 t_rate += td->rate;
1644 m_rate += td->ratemin;
71a751ce
JA
1645 } else if (td->runstate < TD_RUNNING)
1646 nr_pending++;
ebac4655 1647
6a0106a0
JA
1648 if (elapsed >= 3)
1649 eta_secs[i] = thread_eta(td, elapsed);
8b611c34
JA
1650 else
1651 eta_secs[i] = INT_MAX;
ebac4655
JA
1652
1653 check_str_update(td);
1654 }
1655
6a0106a0
JA
1656 if (exitall_on_terminate)
1657 eta_sec = INT_MAX;
1658 else
1659 eta_sec = 0;
5289b847 1660
6a0106a0
JA
1661 for (i = 0; i < thread_number; i++) {
1662 if (exitall_on_terminate) {
1663 if (eta_secs[i] < eta_sec)
1664 eta_sec = eta_secs[i];
1665 } else {
1666 if (eta_secs[i] > eta_sec)
1667 eta_sec = eta_secs[i];
5289b847 1668 }
6a0106a0 1669 }
5289b847 1670
8b611c34 1671 if (eta_sec != INT_MAX && elapsed) {
6a0106a0
JA
1672 perc = (double) elapsed / (double) (elapsed + eta_sec);
1673 eta_to_str(eta_str, eta_sec);
ebac4655
JA
1674 }
1675
71a751ce 1676 if (!nr_running && !nr_pending)
f44c1d46
JA
1677 return;
1678
eb8bbf48 1679 printf("Threads running: %d", nr_running);
ebac4655
JA
1680 if (m_rate || t_rate)
1681 printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
30521453 1682 if (eta_sec != INT_MAX && nr_running) {
6a0106a0
JA
1683 perc *= 100.0;
1684 printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str);
1685 }
5289b847 1686 printf("\r");
ebac4655 1687 fflush(stdout);
6a0106a0 1688 free(eta_secs);
ebac4655
JA
1689}
1690
906c8d75
JA
1691/*
1692 * Run over the job map and reap the threads that have exited, if any.
1693 */
ebac4655
JA
1694static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
1695{
1696 int i;
1697
1698 /*
1699 * reap exited threads (TD_EXITED -> TD_REAPED)
1700 */
1701 for (i = 0; i < thread_number; i++) {
1702 struct thread_data *td = &threads[i];
1703
1704 if (td->runstate != TD_EXITED)
1705 continue;
1706
1707 td_set_runstate(td, TD_REAPED);
1708
1709 if (td->use_thread) {
1710 long ret;
1711
1712 if (pthread_join(td->thread, (void *) &ret))
1713 perror("thread_join");
1714 } else
1715 waitpid(td->pid, NULL, 0);
1716
1717 (*nr_running)--;
1718 (*m_rate) -= td->ratemin;
1719 (*t_rate) -= td->rate;
1720 }
1721}
1722
fcb6ade2
JA
1723static void fio_unpin_memory(void *pinned)
1724{
1725 if (pinned) {
1726 if (munlock(pinned, mlock_size) < 0)
1727 perror("munlock");
1728 munmap(pinned, mlock_size);
1729 }
1730}
1731
1732static void *fio_pin_memory(void)
1733{
32cd46a0 1734 unsigned long long phys_mem;
fcb6ade2
JA
1735 void *ptr;
1736
1737 if (!mlock_size)
1738 return NULL;
1739
1740 /*
1741 * Don't allow mlock of more than real_mem-128MB
1742 */
32cd46a0
JA
1743 phys_mem = os_phys_mem();
1744 if (phys_mem) {
1745 if ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
1746 mlock_size = phys_mem - 128 * 1024 * 1024;
eb8bbf48 1747 fprintf(f_out, "fio: limiting mlocked memory to %lluMiB\n", mlock_size >> 20);
fcb6ade2
JA
1748 }
1749 }
1750
1751 ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
1752 if (!ptr) {
1753 perror("malloc locked mem");
1754 return NULL;
1755 }
1756 if (mlock(ptr, mlock_size) < 0) {
1757 munmap(ptr, mlock_size);
1758 perror("mlock");
1759 return NULL;
1760 }
1761
1762 return ptr;
1763}
1764
906c8d75
JA
1765/*
1766 * Main function for kicking off and reaping jobs, as needed.
1767 */
ebac4655
JA
1768static void run_threads(void)
1769{
ebac4655
JA
1770 struct thread_data *td;
1771 unsigned long spent;
1772 int i, todo, nr_running, m_rate, t_rate, nr_started;
fcb6ade2
JA
1773 void *mlocked_mem;
1774
1775 mlocked_mem = fio_pin_memory();
ebac4655 1776
c6ae0a5b
JA
1777 if (!terse_output) {
1778 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
1779 fflush(stdout);
1780 }
c04f7ec3 1781
4efa970e
JA
1782 signal(SIGINT, sig_handler);
1783 signal(SIGALRM, sig_handler);
1784
ebac4655
JA
1785 todo = thread_number;
1786 nr_running = 0;
1787 nr_started = 0;
1788 m_rate = t_rate = 0;
1789
1790 for (i = 0; i < thread_number; i++) {
1791 td = &threads[i];
1792
1793 run_str[td->thread_number - 1] = 'P';
1794
1795 init_disk_util(td);
1796
1797 if (!td->create_serialize)
1798 continue;
1799
1800 /*
1801 * do file setup here so it happens sequentially,
1802 * we don't want X number of threads getting their
1803 * client data interspersed on disk
1804 */
1805 if (setup_file(td)) {
1806 td_set_runstate(td, TD_REAPED);
1807 todo--;
1808 }
1809 }
1810
1811 gettimeofday(&genesis, NULL);
1812
1813 while (todo) {
75154845
JA
1814 struct thread_data *map[MAX_JOBS];
1815 struct timeval this_start;
1816 int this_jobs = 0, left;
1817
ebac4655
JA
1818 /*
1819 * create threads (TD_NOT_CREATED -> TD_CREATED)
1820 */
1821 for (i = 0; i < thread_number; i++) {
1822 td = &threads[i];
1823
1824 if (td->runstate != TD_NOT_CREATED)
1825 continue;
1826
1827 /*
1828 * never got a chance to start, killed by other
1829 * thread for some reason
1830 */
1831 if (td->terminate) {
1832 todo--;
1833 continue;
1834 }
1835
1836 if (td->start_delay) {
1837 spent = mtime_since_now(&genesis);
1838
1839 if (td->start_delay * 1000 > spent)
1840 continue;
1841 }
1842
1843 if (td->stonewall && (nr_started || nr_running))
1844 break;
1845
75154845
JA
1846 /*
1847 * Set state to created. Thread will transition
1848 * to TD_INITIALIZED when it's done setting up.
1849 */
ebac4655 1850 td_set_runstate(td, TD_CREATED);
75154845 1851 map[this_jobs++] = td;
bbfd6b00 1852 fio_sem_init(&startup_sem, 1);
ebac4655
JA
1853 nr_started++;
1854
1855 if (td->use_thread) {
1856 if (pthread_create(&td->thread, NULL, thread_main, td)) {
1857 perror("thread_create");
1858 nr_started--;
1859 }
1860 } else {
1861 if (fork())
bbfd6b00 1862 fio_sem_down(&startup_sem);
ebac4655
JA
1863 else {
1864 fork_main(shm_id, i);
1865 exit(0);
1866 }
1867 }
1868 }
1869
1870 /*
75154845
JA
1871 * Wait for the started threads to transition to
1872 * TD_INITIALIZED.
ebac4655 1873 */
75154845
JA
1874 gettimeofday(&this_start, NULL);
1875 left = this_jobs;
1876 while (left) {
1877 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
1878 break;
1879
1880 usleep(100000);
1881
1882 for (i = 0; i < this_jobs; i++) {
1883 td = map[i];
1884 if (!td)
1885 continue;
b6f4d880 1886 if (td->runstate == TD_INITIALIZED) {
75154845
JA
1887 map[i] = NULL;
1888 left--;
b6f4d880
JA
1889 } else if (td->runstate >= TD_EXITED) {
1890 map[i] = NULL;
1891 left--;
1892 todo--;
1893 nr_running++; /* work-around... */
75154845
JA
1894 }
1895 }
1896 }
1897
1898 if (left) {
3b70d7e5 1899 log_err("fio: %d jobs failed to start\n", left);
75154845
JA
1900 for (i = 0; i < this_jobs; i++) {
1901 td = map[i];
1902 if (!td)
1903 continue;
1904 kill(td->pid, SIGTERM);
1905 }
1906 break;
1907 }
1908
1909 /*
b6f4d880 1910 * start created threads (TD_INITIALIZED -> TD_RUNNING).
75154845 1911 */
ebac4655
JA
1912 for (i = 0; i < thread_number; i++) {
1913 td = &threads[i];
1914
75154845 1915 if (td->runstate != TD_INITIALIZED)
ebac4655
JA
1916 continue;
1917
1918 td_set_runstate(td, TD_RUNNING);
1919 nr_running++;
1920 nr_started--;
1921 m_rate += td->ratemin;
1922 t_rate += td->rate;
75154845 1923 todo--;
bbfd6b00 1924 fio_sem_up(&td->mutex);
ebac4655
JA
1925 }
1926
1927 reap_threads(&nr_running, &t_rate, &m_rate);
1928
1929 if (todo)
1930 usleep(100000);
1931 }
1932
1933 while (nr_running) {
1934 reap_threads(&nr_running, &t_rate, &m_rate);
1935 usleep(10000);
1936 }
1937
1938 update_io_ticks();
fcb6ade2 1939 fio_unpin_memory(mlocked_mem);
ebac4655
JA
1940}
1941
ebac4655
JA
1942int main(int argc, char *argv[])
1943{
1944 if (parse_options(argc, argv))
1945 return 1;
1946
1947 if (!thread_number) {
3b70d7e5 1948 log_err("Nothing to do\n");
ebac4655
JA
1949 return 1;
1950 }
1951
1952 disk_util_timer_arm();
1953
1954 run_threads();
1955 show_run_stats();
1956
1957 return 0;
1958}