projects / fio.git / blame_incremental
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (non-incremental) | history | HEAD
syslet engine: submission latency numbers
[fio.git] / init.c
... / ...
CommitLineData
1/*
2 * This file contains job initialization and setup functions.
3 */
4#include <stdio.h>
5#include <stdlib.h>
6#include <unistd.h>
7#include <fcntl.h>
8#include <ctype.h>
9#include <string.h>
10#include <errno.h>
11#include <getopt.h>
12#include <assert.h>
13#include <sys/ipc.h>
14#include <sys/shm.h>
15#include <sys/types.h>
16#include <sys/stat.h>
17
18#include "fio.h"
19#include "parse.h"
20
21#define FIO_RANDSEED (0xb1899bedUL)
22
23#define td_var_offset(var) ((size_t) &((struct thread_data *)0)->var)
24
25static int str_ioengine_cb(void *, const char *);
26static int str_mem_cb(void *, const char *);
27static int str_lockmem_cb(void *, unsigned long *);
28#ifdef FIO_HAVE_IOPRIO
29static int str_prio_cb(void *, unsigned int *);
30static int str_prioclass_cb(void *, unsigned int *);
31#endif
32static int str_exitall_cb(void);
33static int str_cpumask_cb(void *, unsigned int *);
34
35#define __stringify_1(x) #x
36#define __stringify(x) __stringify_1(x)
37
38/*
39 * Map of job/command line options
40 */
41static struct fio_option options[] = {
42 {
43 .name = "description",
44 .type = FIO_OPT_STR_STORE,
45 .off1 = td_var_offset(description),
46 .help = "Text job description",
47 },
48 {
49 .name = "name",
50 .type = FIO_OPT_STR_STORE,
51 .off1 = td_var_offset(name),
52 .help = "Name of this job",
53 },
54 {
55 .name = "directory",
56 .type = FIO_OPT_STR_STORE,
57 .off1 = td_var_offset(directory),
58 .help = "Directory to store files in",
59 },
60 {
61 .name = "filename",
62 .type = FIO_OPT_STR_STORE,
63 .off1 = td_var_offset(filename),
64 .help = "Force the use of a specific file",
65 },
66 {
67 .name = "rw",
68 .type = FIO_OPT_STR,
69 .off1 = td_var_offset(td_ddir),
70 .help = "IO direction",
71 .def = "read",
72 .posval = {
73 { .ival = "read", .oval = TD_DDIR_READ },
74 { .ival = "write", .oval = TD_DDIR_WRITE },
75 { .ival = "randread", .oval = TD_DDIR_RANDREAD },
76 { .ival = "randwrite", .oval = TD_DDIR_RANDWRITE },
77 { .ival = "rw", .oval = TD_DDIR_RW },
78 { .ival = "randrw", .oval = TD_DDIR_RANDRW },
79 },
80 },
81 {
82 .name = "ioengine",
83 .type = FIO_OPT_STR,
84 .cb = str_ioengine_cb,
85 .help = "IO engine to use",
86 .def = "sync",
87 .posval = {
88 { .ival = "sync", },
89#ifdef FIO_HAVE_LIBAIO
90 { .ival = "libaio", },
91#endif
92#ifdef FIO_HAVE_POSIXAIO
93 { .ival = "posixaio", },
94#endif
95 { .ival = "mmap", },
96#ifdef FIO_HAVE_SPLICE
97 { .ival = "splice", },
98#endif
99#ifdef FIO_HAVE_SGIO
100 { .ival = "sg", },
101#endif
102 { .ival = "null", }, { .ival = "net", },
103#ifdef FIO_HAVE_SYSLET
104 { .ival = "syslet-rw", },
105#endif
106 },
107 },
108 {
109 .name = "iodepth",
110 .type = FIO_OPT_INT,
111 .off1 = td_var_offset(iodepth),
112 .help = "Amount of IO buffers to keep in flight",
113 .def = "1",
114 },
115 {
116 .name = "iodepth_batch",
117 .type = FIO_OPT_INT,
118 .off1 = td_var_offset(iodepth_batch),
119 .help = "Number of IO to submit in one go",
120 },
121 {
122 .name = "iodepth_low",
123 .type = FIO_OPT_INT,
124 .off1 = td_var_offset(iodepth_low),
125 .help = "Low water mark for queuing depth",
126 },
127 {
128 .name = "size",
129 .type = FIO_OPT_STR_VAL,
130 .off1 = td_var_offset(total_file_size),
131 .help = "Size of device or file",
132 },
133 {
134 .name = "bs",
135 .type = FIO_OPT_STR_VAL_INT,
136 .off1 = td_var_offset(bs[DDIR_READ]),
137 .off2 = td_var_offset(bs[DDIR_WRITE]),
138 .help = "Block size unit",
139 .def = "4k",
140 },
141 {
142 .name = "bsrange",
143 .type = FIO_OPT_RANGE,
144 .off1 = td_var_offset(min_bs[DDIR_READ]),
145 .off2 = td_var_offset(max_bs[DDIR_READ]),
146 .off3 = td_var_offset(min_bs[DDIR_WRITE]),
147 .off4 = td_var_offset(max_bs[DDIR_WRITE]),
148 .help = "Set block size range (in more detail than bs)",
149 },
150 {
151 .name = "bs_unaligned",
152 .type = FIO_OPT_STR_SET,
153 .off1 = td_var_offset(bs_unaligned),
154 .help = "Don't sector align IO buffer sizes",
155 },
156 {
157 .name = "offset",
158 .type = FIO_OPT_STR_VAL,
159 .off1 = td_var_offset(start_offset),
160 .help = "Start IO from this offset",
161 .def = "0",
162 },
163 {
164 .name = "randrepeat",
165 .type = FIO_OPT_BOOL,
166 .off1 = td_var_offset(rand_repeatable),
167 .help = "Use repeatable random IO pattern",
168 .def = "1",
169 },
170 {
171 .name = "norandommap",
172 .type = FIO_OPT_STR_SET,
173 .off1 = td_var_offset(norandommap),
174 .help = "Accept potential duplicate random blocks",
175 },
176 {
177 .name = "nrfiles",
178 .type = FIO_OPT_INT,
179 .off1 = td_var_offset(nr_files),
180 .help = "Split job workload between this number of files",
181 .def = "1",
182 },
183 {
184 .name = "file_service_type",
185 .type = FIO_OPT_STR,
186 .off1 = td_var_offset(file_service_type),
187 .help = "How to select which file to service next",
188 .def = "roundrobin",
189 .posval = {
190 { .ival = "random", .oval = FIO_FSERVICE_RANDOM },
191 { .ival = "roundrobin", .oval = FIO_FSERVICE_RR },
192 },
193 },
194 {
195 .name = "fsync",
196 .type = FIO_OPT_INT,
197 .off1 = td_var_offset(fsync_blocks),
198 .help = "Issue fsync for writes every given number of blocks",
199 .def = "0",
200 },
201 {
202 .name = "direct",
203 .type = FIO_OPT_BOOL,
204 .off1 = td_var_offset(odirect),
205 .help = "Use O_DIRECT IO (negates buffered)",
206 .def = "0",
207 },
208 {
209 .name = "buffered",
210 .type = FIO_OPT_BOOL,
211 .off1 = td_var_offset(odirect),
212 .neg = 1,
213 .help = "Use buffered IO (negates direct)",
214 .def = "1",
215 },
216 {
217 .name = "overwrite",
218 .type = FIO_OPT_BOOL,
219 .off1 = td_var_offset(overwrite),
220 .help = "When writing, set whether to overwrite current data",
221 .def = "0",
222 },
223 {
224 .name = "loops",
225 .type = FIO_OPT_INT,
226 .off1 = td_var_offset(loops),
227 .help = "Number of times to run the job",
228 .def = "1",
229 },
230 {
231 .name = "numjobs",
232 .type = FIO_OPT_INT,
233 .off1 = td_var_offset(numjobs),
234 .help = "Duplicate this job this many times",
235 .def = "1",
236 },
237 {
238 .name = "startdelay",
239 .type = FIO_OPT_INT,
240 .off1 = td_var_offset(start_delay),
241 .help = "Only start job when this period has passed",
242 .def = "0",
243 },
244 {
245 .name = "runtime",
246 .alias = "timeout",
247 .type = FIO_OPT_STR_VAL_TIME,
248 .off1 = td_var_offset(timeout),
249 .help = "Stop workload when this amount of time has passed",
250 .def = "0",
251 },
252 {
253 .name = "mem",
254 .type = FIO_OPT_STR,
255 .cb = str_mem_cb,
256 .off1 = td_var_offset(mem_type),
257 .help = "Backing type for IO buffers",
258 .def = "malloc",
259 .posval = {
260 { .ival = "malloc", .oval = MEM_MALLOC },
261 { .ival = "shm", .oval = MEM_SHM },
262#ifdef FIO_HAVE_HUGETLB
263 { .ival = "shmhuge", .oval = MEM_SHMHUGE },
264#endif
265 { .ival = "mmap", .oval = MEM_MMAP },
266#ifdef FIO_HAVE_HUGETLB
267 { .ival = "mmaphuge", .oval = MEM_MMAPHUGE },
268#endif
269 },
270 },
271 {
272 .name = "verify",
273 .type = FIO_OPT_STR,
274 .off1 = td_var_offset(verify),
275 .help = "Verify sum function",
276 .def = "0",
277 .posval = {
278 { .ival = "0", .oval = VERIFY_NONE },
279 { .ival = "crc32", .oval = VERIFY_CRC32 },
280 { .ival = "md5", .oval = VERIFY_MD5 },
281 },
282 },
283 {
284 .name = "write_iolog",
285 .type = FIO_OPT_STR_STORE,
286 .off1 = td_var_offset(write_iolog_file),
287 .help = "Store IO pattern to file",
288 },
289 {
290 .name = "read_iolog",
291 .type = FIO_OPT_STR_STORE,
292 .off1 = td_var_offset(read_iolog_file),
293 .help = "Playback IO pattern from file",
294 },
295 {
296 .name = "exec_prerun",
297 .type = FIO_OPT_STR_STORE,
298 .off1 = td_var_offset(exec_prerun),
299 .help = "Execute this file prior to running job",
300 },
301 {
302 .name = "exec_postrun",
303 .type = FIO_OPT_STR_STORE,
304 .off1 = td_var_offset(exec_postrun),
305 .help = "Execute this file after running job",
306 },
307#ifdef FIO_HAVE_IOSCHED_SWITCH
308 {
309 .name = "ioscheduler",
310 .type = FIO_OPT_STR_STORE,
311 .off1 = td_var_offset(ioscheduler),
312 .help = "Use this IO scheduler on the backing device",
313 },
314#endif
315 {
316 .name = "zonesize",
317 .type = FIO_OPT_STR_VAL,
318 .off1 = td_var_offset(zone_size),
319 .help = "Give size of an IO zone",
320 .def = "0",
321 },
322 {
323 .name = "zoneskip",
324 .type = FIO_OPT_STR_VAL,
325 .off1 = td_var_offset(zone_skip),
326 .help = "Space between IO zones",
327 .def = "0",
328 },
329 {
330 .name = "lockmem",
331 .type = FIO_OPT_STR_VAL,
332 .cb = str_lockmem_cb,
333 .help = "Lock down this amount of memory",
334 .def = "0",
335 },
336 {
337 .name = "rwmixcycle",
338 .type = FIO_OPT_INT,
339 .off1 = td_var_offset(rwmixcycle),
340 .help = "Cycle period for mixed read/write workloads (msec)",
341 .def = "500",
342 },
343 {
344 .name = "rwmixread",
345 .type = FIO_OPT_INT,
346 .off1 = td_var_offset(rwmixread),
347 .maxval = 100,
348 .help = "Percentage of mixed workload that is reads",
349 .def = "50",
350 },
351 {
352 .name = "rwmixwrite",
353 .type = FIO_OPT_INT,
354 .off1 = td_var_offset(rwmixwrite),
355 .maxval = 100,
356 .help = "Percentage of mixed workload that is writes",
357 .def = "50",
358 },
359 {
360 .name = "nice",
361 .type = FIO_OPT_INT,
362 .off1 = td_var_offset(nice),
363 .help = "Set job CPU nice value",
364 .minval = -19,
365 .maxval = 20,
366 .def = "0",
367 },
368#ifdef FIO_HAVE_IOPRIO
369 {
370 .name = "prio",
371 .type = FIO_OPT_INT,
372 .cb = str_prio_cb,
373 .help = "Set job IO priority value",
374 .minval = 0,
375 .maxval = 7,
376 },
377 {
378 .name = "prioclass",
379 .type = FIO_OPT_INT,
380 .cb = str_prioclass_cb,
381 .help = "Set job IO priority class",
382 .minval = 0,
383 .maxval = 3,
384 },
385#endif
386 {
387 .name = "thinktime",
388 .type = FIO_OPT_INT,
389 .off1 = td_var_offset(thinktime),
390 .help = "Idle time between IO buffers (usec)",
391 .def = "0",
392 },
393 {
394 .name = "thinktime_spin",
395 .type = FIO_OPT_INT,
396 .off1 = td_var_offset(thinktime_spin),
397 .help = "Start thinktime by spinning this amount (usec)",
398 .def = "0",
399 },
400 {
401 .name = "thinktime_blocks",
402 .type = FIO_OPT_INT,
403 .off1 = td_var_offset(thinktime_blocks),
404 .help = "IO buffer period between 'thinktime'",
405 .def = "1",
406 },
407 {
408 .name = "rate",
409 .type = FIO_OPT_INT,
410 .off1 = td_var_offset(rate),
411 .help = "Set bandwidth rate",
412 },
413 {
414 .name = "ratemin",
415 .type = FIO_OPT_INT,
416 .off1 = td_var_offset(ratemin),
417 .help = "The bottom limit accepted",
418 },
419 {
420 .name = "ratecycle",
421 .type = FIO_OPT_INT,
422 .off1 = td_var_offset(ratecycle),
423 .help = "Window average for rate limits (msec)",
424 .def = "1000",
425 },
426 {
427 .name = "invalidate",
428 .type = FIO_OPT_BOOL,
429 .off1 = td_var_offset(invalidate_cache),
430 .help = "Invalidate buffer/page cache prior to running job",
431 .def = "1",
432 },
433 {
434 .name = "sync",
435 .type = FIO_OPT_BOOL,
436 .off1 = td_var_offset(sync_io),
437 .help = "Use O_SYNC for buffered writes",
438 .def = "0",
439 },
440 {
441 .name = "bwavgtime",
442 .type = FIO_OPT_INT,
443 .off1 = td_var_offset(bw_avg_time),
444 .help = "Time window over which to calculate bandwidth (msec)",
445 .def = "500",
446 },
447 {
448 .name = "create_serialize",
449 .type = FIO_OPT_BOOL,
450 .off1 = td_var_offset(create_serialize),
451 .help = "Serialize creating of job files",
452 .def = "1",
453 },
454 {
455 .name = "create_fsync",
456 .type = FIO_OPT_BOOL,
457 .off1 = td_var_offset(create_fsync),
458 .help = "Fsync file after creation",
459 .def = "1",
460 },
461 {
462 .name = "cpuload",
463 .type = FIO_OPT_INT,
464 .off1 = td_var_offset(cpuload),
465 .help = "Use this percentage of CPU",
466 },
467 {
468 .name = "cpuchunks",
469 .type = FIO_OPT_INT,
470 .off1 = td_var_offset(cpucycle),
471 .help = "Length of the CPU burn cycles",
472 },
473#ifdef FIO_HAVE_CPU_AFFINITY
474 {
475 .name = "cpumask",
476 .type = FIO_OPT_INT,
477 .cb = str_cpumask_cb,
478 .help = "CPU affinity mask",
479 },
480#endif
481 {
482 .name = "end_fsync",
483 .type = FIO_OPT_BOOL,
484 .off1 = td_var_offset(end_fsync),
485 .help = "Include fsync at the end of job",
486 .def = "0",
487 },
488 {
489 .name = "unlink",
490 .type = FIO_OPT_BOOL,
491 .off1 = td_var_offset(unlink),
492 .help = "Unlink created files after job has completed",
493 .def = "0",
494 },
495 {
496 .name = "exitall",
497 .type = FIO_OPT_STR_SET,
498 .cb = str_exitall_cb,
499 .help = "Terminate all jobs when one exits",
500 },
501 {
502 .name = "stonewall",
503 .type = FIO_OPT_STR_SET,
504 .off1 = td_var_offset(stonewall),
505 .help = "Insert a hard barrier between this job and previous",
506 },
507 {
508 .name = "thread",
509 .type = FIO_OPT_STR_SET,
510 .off1 = td_var_offset(use_thread),
511 .help = "Use threads instead of forks",
512 },
513 {
514 .name = "write_bw_log",
515 .type = FIO_OPT_STR_SET,
516 .off1 = td_var_offset(write_bw_log),
517 .help = "Write log of bandwidth during run",
518 },
519 {
520 .name = "write_lat_log",
521 .type = FIO_OPT_STR_SET,
522 .off1 = td_var_offset(write_lat_log),
523 .help = "Write log of latency during run",
524 },
525 {
526 .name = "hugepage-size",
527 .type = FIO_OPT_STR_VAL,
528 .off1 = td_var_offset(hugepage_size),
529 .help = "When using hugepages, specify size of each page",
530 .def = __stringify(FIO_HUGE_PAGE),
531 },
532 {
533 .name = NULL,
534 },
535};
536
537#define FIO_JOB_OPTS (sizeof(options) / sizeof(struct fio_option))
538#define FIO_CMD_OPTS (16)
539#define FIO_GETOPT_JOB (0x89988998)
540
541/*
542 * Command line options. These will contain the above, plus a few
543 * extra that only pertain to fio itself and not jobs.
544 */
545static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
546 {
547 .name = "output",
548 .has_arg = required_argument,
549 .val = 'o',
550 },
551 {
552 .name = "timeout",
553 .has_arg = required_argument,
554 .val = 't',
555 },
556 {
557 .name = "latency-log",
558 .has_arg = required_argument,
559 .val = 'l',
560 },
561 {
562 .name = "bandwidth-log",
563 .has_arg = required_argument,
564 .val = 'b',
565 },
566 {
567 .name = "minimal",
568 .has_arg = optional_argument,
569 .val = 'm',
570 },
571 {
572 .name = "version",
573 .has_arg = no_argument,
574 .val = 'v',
575 },
576 {
577 .name = "help",
578 .has_arg = no_argument,
579 .val = 'h',
580 },
581 {
582 .name = "cmdhelp",
583 .has_arg = required_argument,
584 .val = 'c',
585 },
586 {
587 .name = NULL,
588 },
589};
590
591static int def_timeout = 0;
592
593static char fio_version_string[] = "fio 1.11";
594
595static char **ini_file;
596static int max_jobs = MAX_JOBS;
597
598struct thread_data def_thread;
599struct thread_data *threads = NULL;
600
601int exitall_on_terminate = 0;
602int terse_output = 0;
603unsigned long long mlock_size = 0;
604FILE *f_out = NULL;
605FILE *f_err = NULL;
606
607static int write_lat_log = 0;
608int write_bw_log = 0;
609
610/*
611 * Return a free job structure.
612 */
613static struct thread_data *get_new_job(int global, struct thread_data *parent)
614{
615 struct thread_data *td;
616
617 if (global)
618 return &def_thread;
619 if (thread_number >= max_jobs)
620 return NULL;
621
622 td = &threads[thread_number++];
623 *td = *parent;
624
625 td->thread_number = thread_number;
626 return td;
627}
628
629static void put_job(struct thread_data *td)
630{
631 if (td == &def_thread)
632 return;
633
634 if (td->error)
635 fprintf(f_out, "fio: %s\n", td->verror);
636
637 memset(&threads[td->thread_number - 1], 0, sizeof(*td));
638 thread_number--;
639}
640
641/*
642 * Lazy way of fixing up options that depend on each other. We could also
643 * define option callback handlers, but this is easier.
644 */
645static void fixup_options(struct thread_data *td)
646{
647 if (!td->rwmixread && td->rwmixwrite)
648 td->rwmixread = 100 - td->rwmixwrite;
649
650 if (td->write_iolog_file && td->read_iolog_file) {
651 log_err("fio: read iolog overrides write_iolog\n");
652 free(td->write_iolog_file);
653 td->write_iolog_file = NULL;
654 }
655
656 if (td->io_ops->flags & FIO_SYNCIO)
657 td->iodepth = 1;
658 else {
659 if (!td->iodepth)
660 td->iodepth = td->nr_files;
661 }
662
663 /*
664 * only really works for sequential io for now, and with 1 file
665 */
666 if (td->zone_size && td_random(td) && td->nr_files == 1)
667 td->zone_size = 0;
668
669 /*
670 * Reads can do overwrites, we always need to pre-create the file
671 */
672 if (td_read(td) || td_rw(td))
673 td->overwrite = 1;
674
675 if (!td->min_bs[DDIR_READ])
676 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
677 if (!td->max_bs[DDIR_READ])
678 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
679 if (!td->min_bs[DDIR_WRITE])
680 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
681 if (!td->max_bs[DDIR_WRITE])
682 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
683
684 td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
685
686 if (td_read(td) && !td_rw(td))
687 td->verify = 0;
688
689 if (td->norandommap && td->verify != VERIFY_NONE) {
690 log_err("fio: norandommap given, verify disabled\n");
691 td->verify = VERIFY_NONE;
692 }
693 if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
694 log_err("fio: bs_unaligned may not work with raw io\n");
695
696 /*
697 * O_DIRECT and char doesn't mix, clear that flag if necessary.
698 */
699 if (td->filetype == FIO_TYPE_CHAR && td->odirect)
700 td->odirect = 0;
701
702 /*
703 * thinktime_spin must be less than thinktime
704 */
705 if (td->thinktime_spin > td->thinktime)
706 td->thinktime_spin = td->thinktime;
707
708 /*
709 * The low water mark cannot be bigger than the iodepth
710 */
711 if (td->iodepth_low > td->iodepth || !td->iodepth_low)
712 td->iodepth_low = td->iodepth;
713
714 /*
715 * If batch number isn't set, default to the same as iodepth
716 */
717 if (td->iodepth_batch > td->iodepth || !td->iodepth_batch)
718 td->iodepth_batch = td->iodepth;
719}
720
721/*
722 * This function leaks the buffer
723 */
724static char *to_kmg(unsigned int val)
725{
726 char *buf = malloc(32);
727 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
728 char *p = post;
729
730 do {
731 if (val & 1023)
732 break;
733
734 val >>= 10;
735 p++;
736 } while (*p);
737
738 snprintf(buf, 31, "%u%c", val, *p);
739 return buf;
740}
741
742/*
743 * Adds a job to the list of things todo. Sanitizes the various options
744 * to make sure we don't have conflicts, and initializes various
745 * members of td.
746 */
747static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
748{
749 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
750 "randread", "randwrite", "randrw" };
751 struct stat sb;
752 int numjobs, i;
753 struct fio_file *f;
754
755 /*
756 * the def_thread is just for options, it's not a real job
757 */
758 if (td == &def_thread)
759 return 0;
760
761 assert(td->io_ops);
762
763 if (td->odirect)
764 td->io_ops->flags |= FIO_RAWIO;
765
766 td->filetype = FIO_TYPE_FILE;
767 if (td->filename && !lstat(td->filename, &sb)) {
768 if (S_ISBLK(sb.st_mode))
769 td->filetype = FIO_TYPE_BD;
770 else if (S_ISCHR(sb.st_mode))
771 td->filetype = FIO_TYPE_CHAR;
772 }
773
774 fixup_options(td);
775
776 if (td->filename)
777 td->nr_uniq_files = 1;
778 else
779 td->nr_uniq_files = td->nr_files;
780
781 if (td->filetype == FIO_TYPE_FILE || td->filename) {
782 char tmp[PATH_MAX];
783 int len = 0;
784
785 if (td->directory && td->directory[0] != '\0') {
786 if (lstat(td->directory, &sb) < 0) {
787 log_err("fio: %s is not a directory\n", td->directory);
788 td_verror(td, errno, "lstat");
789 return 1;
790 }
791 if (!S_ISDIR(sb.st_mode)) {
792 log_err("fio: %s is not a directory\n", td->directory);
793 return 1;
794 }
795 len = sprintf(tmp, "%s/", td->directory);
796 }
797
798 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
799
800 for_each_file(td, f, i) {
801 memset(f, 0, sizeof(*f));
802 f->fd = -1;
803
804 if (td->filename)
805 sprintf(tmp + len, "%s", td->filename);
806 else
807 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
808 f->file_name = strdup(tmp);
809 }
810 } else {
811 td->nr_files = 1;
812 td->files = malloc(sizeof(struct fio_file));
813 f = &td->files[0];
814
815 memset(f, 0, sizeof(*f));
816 f->fd = -1;
817 f->file_name = strdup(jobname);
818 }
819
820 for_each_file(td, f, i) {
821 f->file_size = td->total_file_size / td->nr_files;
822 f->file_offset = td->start_offset;
823 }
824
825 fio_sem_init(&td->mutex, 0);
826
827 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
828 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
829 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
830
831 if (td->stonewall && td->thread_number > 1)
832 groupid++;
833
834 td->groupid = groupid;
835
836 if (setup_rate(td))
837 goto err;
838
839 if (td->write_lat_log) {
840 setup_log(&td->ts.slat_log);
841 setup_log(&td->ts.clat_log);
842 }
843 if (td->write_bw_log)
844 setup_log(&td->ts.bw_log);
845
846 if (!td->name)
847 td->name = strdup(jobname);
848
849 if (!terse_output) {
850 if (!job_add_num) {
851 if (td->io_ops->flags & FIO_CPUIO)
852 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
853 else {
854 char *c1, *c2, *c3, *c4;
855
856 c1 = to_kmg(td->min_bs[DDIR_READ]);
857 c2 = to_kmg(td->max_bs[DDIR_READ]);
858 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
859 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
860
861 fprintf(f_out, "%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[td->td_ddir], c1, c2, c3, c4, td->io_ops->name, td->iodepth);
862
863 free(c1);
864 free(c2);
865 free(c3);
866 free(c4);
867 }
868 } else if (job_add_num == 1)
869 fprintf(f_out, "...\n");
870 }
871
872 /*
873 * recurse add identical jobs, clear numjobs and stonewall options
874 * as they don't apply to sub-jobs
875 */
876 numjobs = td->numjobs;
877 while (--numjobs) {
878 struct thread_data *td_new = get_new_job(0, td);
879
880 if (!td_new)
881 goto err;
882
883 td_new->numjobs = 1;
884 td_new->stonewall = 0;
885 job_add_num = numjobs - 1;
886
887 if (add_job(td_new, jobname, job_add_num))
888 goto err;
889 }
890 return 0;
891err:
892 put_job(td);
893 return -1;
894}
895
896/*
897 * Initialize the various random states we need (random io, block size ranges,
898 * read/write mix, etc).
899 */
900int init_random_state(struct thread_data *td)
901{
902 unsigned long seeds[5];
903 int fd, num_maps, blocks, i;
904 struct fio_file *f;
905
906 if (td->io_ops->flags & FIO_CPUIO)
907 return 0;
908
909 fd = open("/dev/urandom", O_RDONLY);
910 if (fd == -1) {
911 td_verror(td, errno, "open");
912 return 1;
913 }
914
915 if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
916 td_verror(td, EIO, "read");
917 close(fd);
918 return 1;
919 }
920
921 close(fd);
922
923 os_random_seed(seeds[0], &td->bsrange_state);
924 os_random_seed(seeds[1], &td->verify_state);
925 os_random_seed(seeds[2], &td->rwmix_state);
926
927 if (td->file_service_type == FIO_FSERVICE_RANDOM)
928 os_random_seed(seeds[3], &td->next_file_state);
929
930 if (!td_random(td))
931 return 0;
932
933 if (td->rand_repeatable)
934 seeds[4] = FIO_RANDSEED * td->thread_number;
935
936 if (!td->norandommap) {
937 for_each_file(td, f, i) {
938 blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
939 num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
940 f->file_map = malloc(num_maps * sizeof(long));
941 f->num_maps = num_maps;
942 memset(f->file_map, 0, num_maps * sizeof(long));
943 }
944 }
945
946 os_random_seed(seeds[4], &td->random_state);
947 return 0;
948}
949
950static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
951{
952#ifdef FIO_HAVE_CPU_AFFINITY
953 unsigned int i;
954
955 CPU_ZERO(&cpumask);
956
957 for (i = 0; i < sizeof(int) * 8; i++) {
958 if ((1 << i) & cpu)
959 CPU_SET(i, &cpumask);
960 }
961#endif
962}
963
964static int is_empty_or_comment(char *line)
965{
966 unsigned int i;
967
968 for (i = 0; i < strlen(line); i++) {
969 if (line[i] == ';')
970 return 1;
971 if (line[i] == '#')
972 return 1;
973 if (!isspace(line[i]) && !iscntrl(line[i]))
974 return 0;
975 }
976
977 return 1;
978}
979
980/*
981 * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
982 */
983static char *get_mmap_file(const char *str)
984{
985 char *p = strstr(str, ":");
986
987 if (!p)
988 return NULL;
989
990 p++;
991 strip_blank_front(&p);
992 strip_blank_end(p);
993 return strdup(p);
994}
995
996static int str_mem_cb(void *data, const char *mem)
997{
998 struct thread_data *td = data;
999
1000 if (td->mem_type == MEM_MMAPHUGE || td->mem_type == MEM_MMAP) {
1001 td->mmapfile = get_mmap_file(mem);
1002 if (td->mem_type == MEM_MMAPHUGE && !td->mmapfile) {
1003 log_err("fio: mmaphuge:/path/to/file\n");
1004 return 1;
1005 }
1006 }
1007
1008 return 0;
1009}
1010
1011static int str_ioengine_cb(void *data, const char *str)
1012{
1013 struct thread_data *td = data;
1014
1015 td->io_ops = load_ioengine(td, str);
1016 if (td->io_ops)
1017 return 0;
1018
1019 return 1;
1020}
1021
1022static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1023{
1024 mlock_size = *val;
1025 return 0;
1026}
1027
1028#ifdef FIO_HAVE_IOPRIO
1029static int str_prioclass_cb(void *data, unsigned int *val)
1030{
1031 struct thread_data *td = data;
1032
1033 td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1034 return 0;
1035}
1036
1037static int str_prio_cb(void *data, unsigned int *val)
1038{
1039 struct thread_data *td = data;
1040
1041 td->ioprio |= *val;
1042 return 0;
1043}
1044#endif
1045
1046static int str_exitall_cb(void)
1047{
1048 exitall_on_terminate = 1;
1049 return 0;
1050}
1051
1052static int str_cpumask_cb(void *data, unsigned int *val)
1053{
1054 struct thread_data *td = data;
1055
1056 fill_cpu_mask(td->cpumask, *val);
1057 return 0;
1058}
1059
1060/*
1061 * This is our [ini] type file parser.
1062 */
1063static int parse_jobs_ini(char *file, int stonewall_flag)
1064{
1065 unsigned int global;
1066 struct thread_data *td;
1067 char *string, *name;
1068 fpos_t off;
1069 FILE *f;
1070 char *p;
1071 int ret = 0, stonewall;
1072
1073 f = fopen(file, "r");
1074 if (!f) {
1075 perror("fopen job file");
1076 return 1;
1077 }
1078
1079 string = malloc(4096);
1080 name = malloc(256);
1081 memset(name, 0, 256);
1082
1083 stonewall = stonewall_flag;
1084 do {
1085 p = fgets(string, 4095, f);
1086 if (!p)
1087 break;
1088 if (is_empty_or_comment(p))
1089 continue;
1090 if (sscanf(p, "[%255s]", name) != 1)
1091 continue;
1092
1093 global = !strncmp(name, "global", 6);
1094
1095 name[strlen(name) - 1] = '\0';
1096
1097 td = get_new_job(global, &def_thread);
1098 if (!td) {
1099 ret = 1;
1100 break;
1101 }
1102
1103 /*
1104 * Seperate multiple job files by a stonewall
1105 */
1106 if (!global && stonewall) {
1107 td->stonewall = stonewall;
1108 stonewall = 0;
1109 }
1110
1111 fgetpos(f, &off);
1112 while ((p = fgets(string, 4096, f)) != NULL) {
1113 if (is_empty_or_comment(p))
1114 continue;
1115
1116 strip_blank_front(&p);
1117
1118 if (p[0] == '[')
1119 break;
1120
1121 strip_blank_end(p);
1122
1123 fgetpos(f, &off);
1124
1125 /*
1126 * Don't break here, continue parsing options so we
1127 * dump all the bad ones. Makes trial/error fixups
1128 * easier on the user.
1129 */
1130 ret |= parse_option(p, options, td);
1131 }
1132
1133 if (!ret) {
1134 fsetpos(f, &off);
1135 ret = add_job(td, name, 0);
1136 } else {
1137 log_err("fio: job %s dropped\n", name);
1138 put_job(td);
1139 }
1140 } while (!ret);
1141
1142 free(string);
1143 free(name);
1144 fclose(f);
1145 return ret;
1146}
1147
1148static int fill_def_thread(void)
1149{
1150 memset(&def_thread, 0, sizeof(def_thread));
1151
1152 if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1153 perror("sched_getaffinity");
1154 return 1;
1155 }
1156
1157 /*
1158 * fill default options
1159 */
1160 fill_default_options(&def_thread, options);
1161
1162 def_thread.timeout = def_timeout;
1163 def_thread.write_bw_log = write_bw_log;
1164 def_thread.write_lat_log = write_lat_log;
1165
1166#ifdef FIO_HAVE_DISK_UTIL
1167 def_thread.do_disk_util = 1;
1168#endif
1169
1170 return 0;
1171}
1172
1173static void usage(void)
1174{
1175 printf("%s\n", fio_version_string);
1176 printf("\t--output\tWrite output to file\n");
1177 printf("\t--timeout\tRuntime in seconds\n");
1178 printf("\t--latency-log\tGenerate per-job latency logs\n");
1179 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1180 printf("\t--minimal\tMinimal (terse) output\n");
1181 printf("\t--version\tPrint version info and exit\n");
1182 printf("\t--help\t\tPrint this page\n");
1183 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1184}
1185
1186static int parse_cmd_line(int argc, char *argv[])
1187{
1188 struct thread_data *td = NULL;
1189 int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
1190
1191 while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1192 switch (c) {
1193 case 't':
1194 def_timeout = atoi(optarg);
1195 break;
1196 case 'l':
1197 write_lat_log = 1;
1198 break;
1199 case 'w':
1200 write_bw_log = 1;
1201 break;
1202 case 'o':
1203 f_out = fopen(optarg, "w+");
1204 if (!f_out) {
1205 perror("fopen output");
1206 exit(1);
1207 }
1208 f_err = f_out;
1209 break;
1210 case 'm':
1211 terse_output = 1;
1212 break;
1213 case 'h':
1214 usage();
1215 exit(0);
1216 case 'c':
1217 ret = show_cmd_help(options, optarg);
1218 exit(ret);
1219 case 'v':
1220 printf("%s\n", fio_version_string);
1221 exit(0);
1222 case FIO_GETOPT_JOB: {
1223 const char *opt = long_options[lidx].name;
1224 char *val = optarg;
1225
1226 if (!strncmp(opt, "name", 4) && td) {
1227 ret = add_job(td, td->name ?: "fio", 0);
1228 if (ret) {
1229 put_job(td);
1230 return 0;
1231 }
1232 td = NULL;
1233 }
1234 if (!td) {
1235 int global = !strncmp(val, "global", 6);
1236
1237 td = get_new_job(global, &def_thread);
1238 if (!td)
1239 return 0;
1240 }
1241
1242 ret = parse_cmd_option(opt, val, options, td);
1243 if (ret)
1244 dont_add_job = 1;
1245 break;
1246 }
1247 default:
1248 break;
1249 }
1250 }
1251
1252 if (td) {
1253 if (dont_add_job)
1254 put_job(td);
1255 else {
1256 ret = add_job(td, td->name ?: "fio", 0);
1257 if (ret)
1258 put_job(td);
1259 }
1260 }
1261
1262 while (optind < argc) {
1263 ini_idx++;
1264 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1265 ini_file[ini_idx - 1] = strdup(argv[optind]);
1266 optind++;
1267 }
1268
1269 return ini_idx;
1270}
1271
1272static void free_shm(void)
1273{
1274 struct shmid_ds sbuf;
1275
1276 if (threads) {
1277 shmdt((void *) threads);
1278 threads = NULL;
1279 shmctl(shm_id, IPC_RMID, &sbuf);
1280 }
1281}
1282
1283/*
1284 * The thread area is shared between the main process and the job
1285 * threads/processes. So setup a shared memory segment that will hold
1286 * all the job info.
1287 */
1288static int setup_thread_area(void)
1289{
1290 /*
1291 * 1024 is too much on some machines, scale max_jobs if
1292 * we get a failure that looks like too large a shm segment
1293 */
1294 do {
1295 size_t size = max_jobs * sizeof(struct thread_data);
1296
1297 shm_id = shmget(0, size, IPC_CREAT | 0600);
1298 if (shm_id != -1)
1299 break;
1300 if (errno != EINVAL) {
1301 perror("shmget");
1302 break;
1303 }
1304
1305 max_jobs >>= 1;
1306 } while (max_jobs);
1307
1308 if (shm_id == -1)
1309 return 1;
1310
1311 threads = shmat(shm_id, NULL, 0);
1312 if (threads == (void *) -1) {
1313 perror("shmat");
1314 return 1;
1315 }
1316
1317 atexit(free_shm);
1318 return 0;
1319}
1320
1321/*
1322 * Copy the fio options into the long options map, so we mirror
1323 * job and cmd line options.
1324 */
1325static void dupe_job_options(void)
1326{
1327 struct fio_option *o;
1328 unsigned int i;
1329
1330 i = 0;
1331 while (long_options[i].name)
1332 i++;
1333
1334 o = &options[0];
1335 while (o->name) {
1336 long_options[i].name = o->name;
1337 long_options[i].val = FIO_GETOPT_JOB;
1338 if (o->type == FIO_OPT_STR_SET)
1339 long_options[i].has_arg = no_argument;
1340 else
1341 long_options[i].has_arg = required_argument;
1342
1343 i++;
1344 o++;
1345 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1346 }
1347}
1348
1349int parse_options(int argc, char *argv[])
1350{
1351 int job_files, i;
1352
1353 f_out = stdout;
1354 f_err = stderr;
1355
1356 options_init(options);
1357
1358 dupe_job_options();
1359
1360 if (setup_thread_area())
1361 return 1;
1362 if (fill_def_thread())
1363 return 1;
1364
1365 job_files = parse_cmd_line(argc, argv);
1366
1367 for (i = 0; i < job_files; i++) {
1368 if (fill_def_thread())
1369 return 1;
1370 if (parse_jobs_ini(ini_file[i], i))
1371 return 1;
1372 free(ini_file[i]);
1373 }
1374
1375 free(ini_file);
1376
1377 if (!thread_number) {
1378 log_err("No jobs defined(s)\n");
1379 return 1;
1380 }
1381
1382 return 0;
1383}
fio - Flexible IO Tester