2 * This file contains job initialization and setup functions.
12 #include <sys/types.h>
16 #ifndef FIO_NO_HAVE_SHM_H
28 #include "lib/getopt.h"
29 #include "lib/strcasestr.h"
31 const char fio_version_string[] = FIO_VERSION;
33 #define FIO_RANDSEED (0xb1899bedUL)
35 static char **ini_file;
36 static int max_jobs = FIO_MAX_JOBS;
37 static int dump_cmdline;
38 static int def_timeout;
39 static int parse_only;
41 static struct thread_data def_thread;
42 struct thread_data *threads = NULL;
44 int exitall_on_terminate = 0;
45 int output_format = FIO_OUTPUT_NORMAL;
46 int eta_print = FIO_ETA_AUTO;
50 char **job_sections = NULL;
51 int nr_job_sections = 0;
52 char *exec_profile = NULL;
53 int warnings_fatal = 0;
54 int terse_version = 3;
61 int status_interval = 0;
63 static int write_lat_log;
65 static int prev_group_jobs;
67 unsigned long fio_debug = 0;
68 unsigned int fio_debug_jobno = -1;
69 unsigned int *fio_debug_jobp = NULL;
71 static char cmd_optstr[256];
74 #define FIO_CLIENT_FLAG (1 << 16)
77 * Command line options. These will contain the above, plus a few
78 * extra that only pertain to fio itself and not jobs.
80 static struct option l_opts[FIO_NR_OPTIONS] = {
82 .name = (char *) "output",
83 .has_arg = required_argument,
84 .val = 'o' | FIO_CLIENT_FLAG,
87 .name = (char *) "timeout",
88 .has_arg = required_argument,
89 .val = 't' | FIO_CLIENT_FLAG,
92 .name = (char *) "latency-log",
93 .has_arg = required_argument,
94 .val = 'l' | FIO_CLIENT_FLAG,
97 .name = (char *) "bandwidth-log",
98 .has_arg = required_argument,
99 .val = 'b' | FIO_CLIENT_FLAG,
102 .name = (char *) "minimal",
103 .has_arg = optional_argument,
104 .val = 'm' | FIO_CLIENT_FLAG,
107 .name = (char *) "output-format",
108 .has_arg = optional_argument,
109 .val = 'F' | FIO_CLIENT_FLAG,
112 .name = (char *) "version",
113 .has_arg = no_argument,
114 .val = 'v' | FIO_CLIENT_FLAG,
117 .name = (char *) "help",
118 .has_arg = no_argument,
119 .val = 'h' | FIO_CLIENT_FLAG,
122 .name = (char *) "cmdhelp",
123 .has_arg = optional_argument,
124 .val = 'c' | FIO_CLIENT_FLAG,
127 .name = (char *) "enghelp",
128 .has_arg = optional_argument,
129 .val = 'i' | FIO_CLIENT_FLAG,
132 .name = (char *) "showcmd",
133 .has_arg = no_argument,
134 .val = 's' | FIO_CLIENT_FLAG,
137 .name = (char *) "readonly",
138 .has_arg = no_argument,
139 .val = 'r' | FIO_CLIENT_FLAG,
142 .name = (char *) "eta",
143 .has_arg = required_argument,
144 .val = 'e' | FIO_CLIENT_FLAG,
147 .name = (char *) "eta-newline",
148 .has_arg = required_argument,
149 .val = 'E' | FIO_CLIENT_FLAG,
152 .name = (char *) "debug",
153 .has_arg = required_argument,
154 .val = 'd' | FIO_CLIENT_FLAG,
157 .name = (char *) "parse-only",
158 .has_arg = no_argument,
159 .val = 'P' | FIO_CLIENT_FLAG,
162 .name = (char *) "section",
163 .has_arg = required_argument,
164 .val = 'x' | FIO_CLIENT_FLAG,
167 .name = (char *) "alloc-size",
168 .has_arg = required_argument,
169 .val = 'a' | FIO_CLIENT_FLAG,
172 .name = (char *) "profile",
173 .has_arg = required_argument,
174 .val = 'p' | FIO_CLIENT_FLAG,
177 .name = (char *) "warnings-fatal",
178 .has_arg = no_argument,
179 .val = 'w' | FIO_CLIENT_FLAG,
182 .name = (char *) "max-jobs",
183 .has_arg = required_argument,
184 .val = 'j' | FIO_CLIENT_FLAG,
187 .name = (char *) "terse-version",
188 .has_arg = required_argument,
189 .val = 'V' | FIO_CLIENT_FLAG,
192 .name = (char *) "server",
193 .has_arg = optional_argument,
196 { .name = (char *) "daemonize",
197 .has_arg = required_argument,
201 .name = (char *) "client",
202 .has_arg = required_argument,
206 .name = (char *) "cpuclock-test",
207 .has_arg = no_argument,
211 .name = (char *) "idle-prof",
212 .has_arg = required_argument,
216 .name = (char *) "status-interval",
217 .has_arg = required_argument,
225 void free_threads_shm(void)
227 struct shmid_ds sbuf;
234 shmctl(shm_id, IPC_RMID, &sbuf);
244 fio_debug_jobp = NULL;
252 * The thread area is shared between the main process and the job
253 * threads/processes. So setup a shared memory segment that will hold
254 * all the job info. We use the end of the region for keeping track of
255 * open files across jobs, for file sharing.
257 static int setup_thread_area(void)
265 * 1024 is too much on some machines, scale max_jobs if
266 * we get a failure that looks like too large a shm segment
269 size_t size = max_jobs * sizeof(struct thread_data);
271 size += file_hash_size;
272 size += sizeof(unsigned int);
274 shm_id = shmget(0, size, IPC_CREAT | 0600);
277 if (errno != EINVAL && errno != ENOMEM && errno != ENOSPC) {
288 threads = shmat(shm_id, NULL, 0);
289 if (threads == (void *) -1) {
294 memset(threads, 0, max_jobs * sizeof(struct thread_data));
295 hash = (void *) threads + max_jobs * sizeof(struct thread_data);
296 fio_debug_jobp = (void *) hash + file_hash_size;
297 *fio_debug_jobp = -1;
298 file_hash_init(hash);
306 * Return a free job structure.
308 static struct thread_data *get_new_job(int global, struct thread_data *parent,
311 struct thread_data *td;
315 if (setup_thread_area()) {
316 log_err("error: failed to setup shm segment\n");
319 if (thread_number >= max_jobs) {
320 log_err("error: maximum number of jobs (%d) reached.\n",
325 td = &threads[thread_number++];
332 td->o.uid = td->o.gid = -1U;
334 dup_files(td, parent);
335 fio_options_mem_dupe(td);
337 profile_add_hooks(td);
339 td->thread_number = thread_number;
341 if (!parent || !parent->o.group_reporting)
347 static void put_job(struct thread_data *td)
349 if (td == &def_thread)
356 log_info("fio: %s\n", td->verror);
358 fio_options_free(td);
362 memset(&threads[td->thread_number - 1], 0, sizeof(*td));
366 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
368 unsigned int bs = td->o.min_bs[ddir];
370 assert(ddir_rw(ddir));
372 if (td->o.rate[ddir])
373 td->rate_bps[ddir] = td->o.rate[ddir];
375 td->rate_bps[ddir] = td->o.rate_iops[ddir] * bs;
377 if (!td->rate_bps[ddir]) {
378 log_err("rate lower than supported\n");
382 td->rate_pending_usleep[ddir] = 0;
386 static int setup_rate(struct thread_data *td)
390 if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
391 ret = __setup_rate(td, DDIR_READ);
392 if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
393 ret |= __setup_rate(td, DDIR_WRITE);
394 if (td->o.rate[DDIR_TRIM] || td->o.rate_iops[DDIR_TRIM])
395 ret |= __setup_rate(td, DDIR_TRIM);
400 static int fixed_block_size(struct thread_options *o)
402 return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
403 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
404 o->min_bs[DDIR_TRIM] == o->max_bs[DDIR_TRIM] &&
405 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE] &&
406 o->min_bs[DDIR_READ] == o->min_bs[DDIR_TRIM];
410 * Lazy way of fixing up options that depend on each other. We could also
411 * define option callback handlers, but this is easier.
413 static int fixup_options(struct thread_data *td)
415 struct thread_options *o = &td->o;
418 #ifndef FIO_HAVE_PSHARED_MUTEX
419 if (!o->use_thread) {
420 log_info("fio: this platform does not support process shared"
421 " mutexes, forcing use of threads. Use the 'thread'"
422 " option to get rid of this warning.\n");
424 ret = warnings_fatal;
428 if (o->write_iolog_file && o->read_iolog_file) {
429 log_err("fio: read iolog overrides write_iolog\n");
430 free(o->write_iolog_file);
431 o->write_iolog_file = NULL;
432 ret = warnings_fatal;
436 * only really works with 1 file
438 if (o->zone_size && o->open_files > 1)
442 * If zone_range isn't specified, backward compatibility dictates it
443 * should be made equal to zone_size.
445 if (o->zone_size && !o->zone_range)
446 o->zone_range = o->zone_size;
449 * Reads can do overwrites, we always need to pre-create the file
451 if (td_read(td) || td_rw(td))
454 if (!o->min_bs[DDIR_READ])
455 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
456 if (!o->max_bs[DDIR_READ])
457 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
458 if (!o->min_bs[DDIR_WRITE])
459 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
460 if (!o->max_bs[DDIR_WRITE])
461 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
462 if (!o->min_bs[DDIR_TRIM])
463 o->min_bs[DDIR_TRIM] = o->bs[DDIR_TRIM];
464 if (!o->max_bs[DDIR_TRIM])
465 o->max_bs[DDIR_TRIM] = o->bs[DDIR_TRIM];
468 o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
469 o->rw_min_bs = min(o->min_bs[DDIR_TRIM], o->rw_min_bs);
472 * For random IO, allow blockalign offset other than min_bs.
474 if (!o->ba[DDIR_READ] || !td_random(td))
475 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
476 if (!o->ba[DDIR_WRITE] || !td_random(td))
477 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
478 if (!o->ba[DDIR_TRIM] || !td_random(td))
479 o->ba[DDIR_TRIM] = o->min_bs[DDIR_TRIM];
481 if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
482 o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE] ||
483 o->ba[DDIR_TRIM] != o->min_bs[DDIR_TRIM]) &&
485 log_err("fio: Any use of blockalign= turns off randommap\n");
487 ret = warnings_fatal;
490 if (!o->file_size_high)
491 o->file_size_high = o->file_size_low;
493 if (o->norandommap && o->verify != VERIFY_NONE
494 && !fixed_block_size(o)) {
495 log_err("fio: norandommap given for variable block sizes, "
496 "verify disabled\n");
497 o->verify = VERIFY_NONE;
498 ret = warnings_fatal;
500 if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
501 log_err("fio: bs_unaligned may not work with raw io\n");
504 * thinktime_spin must be less than thinktime
506 if (o->thinktime_spin > o->thinktime)
507 o->thinktime_spin = o->thinktime;
510 * The low water mark cannot be bigger than the iodepth
512 if (o->iodepth_low > o->iodepth || !o->iodepth_low)
513 o->iodepth_low = o->iodepth;
516 * If batch number isn't set, default to the same as iodepth
518 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
519 o->iodepth_batch = o->iodepth;
521 if (o->nr_files > td->files_index)
522 o->nr_files = td->files_index;
524 if (o->open_files > o->nr_files || !o->open_files)
525 o->open_files = o->nr_files;
527 if (((o->rate[DDIR_READ] + o->rate[DDIR_WRITE] + o->rate[DDIR_TRIM]) &&
528 (o->rate_iops[DDIR_READ] + o->rate_iops[DDIR_WRITE] + o->rate_iops[DDIR_TRIM])) ||
529 ((o->ratemin[DDIR_READ] + o->ratemin[DDIR_WRITE] + o->ratemin[DDIR_TRIM]) &&
530 (o->rate_iops_min[DDIR_READ] + o->rate_iops_min[DDIR_WRITE] + o->rate_iops_min[DDIR_TRIM]))) {
531 log_err("fio: rate and rate_iops are mutually exclusive\n");
534 if ((o->rate[DDIR_READ] < o->ratemin[DDIR_READ]) ||
535 (o->rate[DDIR_WRITE] < o->ratemin[DDIR_WRITE]) ||
536 (o->rate[DDIR_TRIM] < o->ratemin[DDIR_TRIM]) ||
537 (o->rate_iops[DDIR_READ] < o->rate_iops_min[DDIR_READ]) ||
538 (o->rate_iops[DDIR_WRITE] < o->rate_iops_min[DDIR_WRITE]) ||
539 (o->rate_iops[DDIR_TRIM] < o->rate_iops_min[DDIR_TRIM])) {
540 log_err("fio: minimum rate exceeds rate\n");
544 if (!o->timeout && o->time_based) {
545 log_err("fio: time_based requires a runtime/timeout setting\n");
547 ret = warnings_fatal;
550 if (o->fill_device && !o->size)
553 if (o->verify != VERIFY_NONE) {
554 if (td_write(td) && o->do_verify && o->numjobs > 1) {
555 log_info("Multiple writers may overwrite blocks that "
556 "belong to other jobs. This can cause "
557 "verification failures.\n");
558 ret = warnings_fatal;
561 o->refill_buffers = 1;
562 if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
564 o->verify_interval = o->min_bs[DDIR_WRITE];
568 o->invalidate_cache = 0;
569 if (td->io_ops->flags & FIO_PIPEIO) {
570 log_info("fio: cannot pre-read files with an IO engine"
571 " that isn't seekable. Pre-read disabled.\n");
572 ret = warnings_fatal;
577 if (td->io_ops->flags & FIO_BIT_BASED)
583 #ifndef CONFIG_FDATASYNC
584 if (o->fdatasync_blocks) {
585 log_info("fio: this platform does not support fdatasync()"
586 " falling back to using fsync(). Use the 'fsync'"
587 " option instead of 'fdatasync' to get rid of"
589 o->fsync_blocks = o->fdatasync_blocks;
590 o->fdatasync_blocks = 0;
591 ret = warnings_fatal;
597 * Windows doesn't support O_DIRECT or O_SYNC with the _open interface,
598 * so fail if we're passed those flags
600 if ((td->io_ops->flags & FIO_SYNCIO) && (td->o.odirect || td->o.sync_io)) {
601 log_err("fio: Windows does not support direct or non-buffered io with"
602 " the synchronous ioengines. Use the 'windowsaio' ioengine"
603 " with 'direct=1' and 'iodepth=1' instead.\n");
609 * For fully compressible data, just zero them at init time.
610 * It's faster than repeatedly filling it.
612 if (td->o.compress_percentage == 100) {
613 td->o.zero_buffers = 1;
614 td->o.compress_percentage = 0;
618 * Using a non-uniform random distribution excludes usage of
621 if (td->o.random_distribution != FIO_RAND_DIST_RANDOM)
622 td->o.norandommap = 1;
625 * If size is set but less than the min block size, complain
627 if (o->size && o->size < td_min_bs(td)) {
628 log_err("fio: size too small, must be larger than the IO size: %llu\n", (unsigned long long) o->size);
633 * O_ATOMIC implies O_DIRECT
639 * If randseed is set, that overrides randrepeat
642 td->o.rand_repeatable = 0;
648 * This function leaks the buffer
650 char *fio_uint_to_kmg(unsigned int val)
652 char *buf = malloc(32);
653 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
664 snprintf(buf, 32, "%u%c", val, *p);
668 /* External engines are specified by "external:name.o") */
669 static const char *get_engine_name(const char *str)
671 char *p = strstr(str, ":");
677 strip_blank_front(&p);
682 static int exists_and_not_file(const char *filename)
686 if (lstat(filename, &sb) == -1)
689 /* \\.\ is the device namespace in Windows, where every file
690 * is a device node */
691 if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
697 static void td_fill_rand_seeds_os(struct thread_data *td)
699 os_random_seed(td->rand_seeds[FIO_RAND_BS_OFF], &td->bsrange_state);
700 os_random_seed(td->rand_seeds[FIO_RAND_VER_OFF], &td->verify_state);
701 os_random_seed(td->rand_seeds[FIO_RAND_MIX_OFF], &td->rwmix_state);
703 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
704 os_random_seed(td->rand_seeds[FIO_RAND_FILE_OFF], &td->next_file_state);
706 os_random_seed(td->rand_seeds[FIO_RAND_FILE_SIZE_OFF], &td->file_size_state);
707 os_random_seed(td->rand_seeds[FIO_RAND_TRIM_OFF], &td->trim_state);
712 if (td->o.rand_repeatable)
713 td->rand_seeds[FIO_RAND_BLOCK_OFF] = FIO_RANDSEED * td->thread_number;
715 os_random_seed(td->rand_seeds[FIO_RAND_BLOCK_OFF], &td->random_state);
717 os_random_seed(td->rand_seeds[FIO_RAND_SEQ_RAND_READ_OFF], &td->seq_rand_state[DDIR_READ]);
718 os_random_seed(td->rand_seeds[FIO_RAND_SEQ_RAND_WRITE_OFF], &td->seq_rand_state[DDIR_WRITE]);
719 os_random_seed(td->rand_seeds[FIO_RAND_SEQ_RAND_TRIM_OFF], &td->seq_rand_state[DDIR_TRIM]);
722 static void td_fill_rand_seeds_internal(struct thread_data *td)
724 init_rand_seed(&td->__bsrange_state, td->rand_seeds[FIO_RAND_BS_OFF]);
725 init_rand_seed(&td->__verify_state, td->rand_seeds[FIO_RAND_VER_OFF]);
726 init_rand_seed(&td->__rwmix_state, td->rand_seeds[FIO_RAND_MIX_OFF]);
728 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
729 init_rand_seed(&td->__next_file_state, td->rand_seeds[FIO_RAND_FILE_OFF]);
731 init_rand_seed(&td->__file_size_state, td->rand_seeds[FIO_RAND_FILE_SIZE_OFF]);
732 init_rand_seed(&td->__trim_state, td->rand_seeds[FIO_RAND_TRIM_OFF]);
737 if (td->o.rand_repeatable)
738 td->rand_seeds[FIO_RAND_BLOCK_OFF] = FIO_RANDSEED * td->thread_number;
740 init_rand_seed(&td->__random_state, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
741 init_rand_seed(&td->__seq_rand_state[DDIR_READ], td->rand_seeds[FIO_RAND_SEQ_RAND_READ_OFF]);
742 init_rand_seed(&td->__seq_rand_state[DDIR_WRITE], td->rand_seeds[FIO_RAND_SEQ_RAND_WRITE_OFF]);
743 init_rand_seed(&td->__seq_rand_state[DDIR_TRIM], td->rand_seeds[FIO_RAND_SEQ_RAND_TRIM_OFF]);
746 void td_fill_rand_seeds(struct thread_data *td)
748 if (td->o.use_os_rand)
749 td_fill_rand_seeds_os(td);
751 td_fill_rand_seeds_internal(td);
753 init_rand_seed(&td->buf_state, td->rand_seeds[FIO_RAND_BUF_OFF]);
757 * Initializes the ioengine configured for a job, if it has not been done so
760 int ioengine_load(struct thread_data *td)
765 * Engine has already been loaded.
769 if (!td->o.ioengine) {
770 log_err("fio: internal fault, no IO engine specified\n");
774 engine = get_engine_name(td->o.ioengine);
775 td->io_ops = load_ioengine(td, engine);
777 log_err("fio: failed to load engine %s\n", engine);
781 if (td->io_ops->option_struct_size && td->io_ops->options) {
783 * In cases where td->eo is set, clone it for a child thread.
784 * This requires that the parent thread has the same ioengine,
785 * but that requirement must be enforced by the code which
788 void *origeo = td->eo;
790 * Otherwise use the default thread options.
792 if (!origeo && td != &def_thread && def_thread.eo &&
793 def_thread.io_ops->options == td->io_ops->options)
794 origeo = def_thread.eo;
796 options_init(td->io_ops->options);
797 td->eo = malloc(td->io_ops->option_struct_size);
799 * Use the default thread as an option template if this uses the
800 * same options structure and there are non-default options
804 memcpy(td->eo, origeo, td->io_ops->option_struct_size);
805 options_mem_dupe(td->eo, td->io_ops->options);
807 memset(td->eo, 0, td->io_ops->option_struct_size);
808 fill_default_options(td->eo, td->io_ops->options);
810 *(struct thread_data **)td->eo = td;
816 static void init_flags(struct thread_data *td)
818 struct thread_options *o = &td->o;
820 if (o->verify_backlog)
821 td->flags |= TD_F_VER_BACKLOG;
823 td->flags |= TD_F_TRIM_BACKLOG;
824 if (o->read_iolog_file)
825 td->flags |= TD_F_READ_IOLOG;
826 if (o->refill_buffers)
827 td->flags |= TD_F_REFILL_BUFFERS;
828 if (o->scramble_buffers)
829 td->flags |= TD_F_SCRAMBLE_BUFFERS;
830 if (o->verify != VERIFY_NONE)
831 td->flags |= TD_F_VER_NONE;
834 static int setup_random_seeds(struct thread_data *td)
839 if (!td->o.rand_repeatable && !td->o.rand_seed)
840 return init_random_state(td, td->rand_seeds, sizeof(td->rand_seeds));
842 if (!td->o.rand_seed)
845 seed = td->o.rand_seed;
847 for (i = 0; i < 4; i++)
848 seed *= 0x9e370001UL;
850 for (i = 0; i < FIO_RAND_NR_OFFS; i++) {
851 td->rand_seeds[i] = seed;
852 seed *= 0x9e370001UL;
855 td_fill_rand_seeds(td);
866 static struct fpre_keyword {
870 } fpre_keywords[] = {
871 { .keyword = "$jobname", .key = FPRE_JOBNAME, },
872 { .keyword = "$jobnum", .key = FPRE_JOBNUM, },
873 { .keyword = "$filenum", .key = FPRE_FILENUM, },
874 { .keyword = NULL, },
877 static char *make_filename(char *buf, struct thread_options *o,
878 const char *jobname, int jobnum, int filenum)
880 struct fpre_keyword *f;
883 if (!o->filename_format || !strlen(o->filename_format)) {
884 sprintf(buf, "%s.%d.%d", jobname, jobnum, filenum);
888 for (f = &fpre_keywords[0]; f->keyword; f++)
889 f->strlen = strlen(f->keyword);
891 strcpy(buf, o->filename_format);
892 memset(copy, 0, sizeof(copy));
893 for (f = &fpre_keywords[0]; f->keyword; f++) {
895 size_t pre_len, post_start = 0;
896 char *str, *dst = copy;
898 str = strcasestr(buf, f->keyword);
903 if (strlen(str) != f->strlen)
904 post_start = pre_len + f->strlen;
907 strncpy(dst, buf, pre_len);
913 dst += sprintf(dst, "%s", jobname);
916 dst += sprintf(dst, "%d", jobnum);
919 dst += sprintf(dst, "%d", filenum);
927 strcpy(dst, buf + post_start);
936 int parse_dryrun(void)
938 return dump_cmdline || parse_only;
942 * Adds a job to the list of things todo. Sanitizes the various options
943 * to make sure we don't have conflicts, and initializes various
946 static int add_job(struct thread_data *td, const char *jobname, int job_add_num,
947 int recursed, int client_type)
950 char fname[PATH_MAX];
951 int numjobs, file_alloced;
952 struct thread_options *o = &td->o;
955 * the def_thread is just for options, it's not a real job
957 if (td == &def_thread)
963 * if we are just dumping the output command line, don't add the job
965 if (parse_dryrun()) {
970 td->client_type = client_type;
972 if (profile_td_init(td))
975 if (ioengine_load(td))
979 td->io_ops->flags |= FIO_RAWIO;
982 if (!o->filename && !td->files_index && !o->read_iolog_file) {
985 if (o->nr_files == 1 && exists_and_not_file(jobname))
986 add_file(td, jobname);
988 for (i = 0; i < o->nr_files; i++)
989 add_file(td, make_filename(fname, o, jobname, td->thread_number, i));
993 if (fixup_options(td))
999 * IO engines only need this for option callbacks, and the address may
1000 * change in subprocesses.
1003 *(struct thread_data **)td->eo = NULL;
1005 if (td->io_ops->flags & FIO_DISKLESSIO) {
1008 for_each_file(td, f, i)
1009 f->real_file_size = -1ULL;
1012 td->mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
1014 td->ts.clat_percentiles = o->clat_percentiles;
1015 td->ts.percentile_precision = o->percentile_precision;
1016 memcpy(td->ts.percentile_list, o->percentile_list, sizeof(o->percentile_list));
1018 for (i = 0; i < DDIR_RWDIR_CNT; i++) {
1019 td->ts.clat_stat[i].min_val = ULONG_MAX;
1020 td->ts.slat_stat[i].min_val = ULONG_MAX;
1021 td->ts.lat_stat[i].min_val = ULONG_MAX;
1022 td->ts.bw_stat[i].min_val = ULONG_MAX;
1024 td->ddir_seq_nr = o->ddir_seq_nr;
1026 if ((o->stonewall || o->new_group) && prev_group_jobs) {
1027 prev_group_jobs = 0;
1031 td->groupid = groupid;
1034 if (setup_random_seeds(td)) {
1035 td_verror(td, errno, "init_random_state");
1042 if (o->lat_log_file) {
1043 setup_log(&td->lat_log, o->log_avg_msec, IO_LOG_TYPE_LAT);
1044 setup_log(&td->slat_log, o->log_avg_msec, IO_LOG_TYPE_SLAT);
1045 setup_log(&td->clat_log, o->log_avg_msec, IO_LOG_TYPE_CLAT);
1048 setup_log(&td->bw_log, o->log_avg_msec, IO_LOG_TYPE_BW);
1049 if (o->iops_log_file)
1050 setup_log(&td->iops_log, o->log_avg_msec, IO_LOG_TYPE_IOPS);
1053 o->name = strdup(jobname);
1055 if (output_format == FIO_OUTPUT_NORMAL) {
1057 if (is_backend && !recursed)
1058 fio_server_send_add_job(td);
1060 if (!(td->io_ops->flags & FIO_NOIO)) {
1061 char *c1, *c2, *c3, *c4;
1062 char *c5 = NULL, *c6 = NULL;
1064 c1 = fio_uint_to_kmg(o->min_bs[DDIR_READ]);
1065 c2 = fio_uint_to_kmg(o->max_bs[DDIR_READ]);
1066 c3 = fio_uint_to_kmg(o->min_bs[DDIR_WRITE]);
1067 c4 = fio_uint_to_kmg(o->max_bs[DDIR_WRITE]);
1069 if (!o->bs_is_seq_rand) {
1070 c5 = fio_uint_to_kmg(o->min_bs[DDIR_TRIM]);
1071 c6 = fio_uint_to_kmg(o->max_bs[DDIR_TRIM]);
1074 log_info("%s: (g=%d): rw=%s, ", td->o.name,
1076 ddir_str(o->td_ddir));
1078 if (o->bs_is_seq_rand)
1079 log_info("bs(seq/rand)=%s-%s/%s-%s, ",
1082 log_info("bs=%s-%s/%s-%s/%s-%s, ",
1083 c1, c2, c3, c4, c5, c6);
1085 log_info("ioengine=%s, iodepth=%u\n",
1086 td->io_ops->name, o->iodepth);
1095 } else if (job_add_num == 1)
1100 * recurse add identical jobs, clear numjobs and stonewall options
1101 * as they don't apply to sub-jobs
1103 numjobs = o->numjobs;
1105 struct thread_data *td_new = get_new_job(0, td, 1);
1110 td_new->o.numjobs = 1;
1111 td_new->o.stonewall = 0;
1112 td_new->o.new_group = 0;
1115 td_new->o.filename = NULL;
1116 td_new->files_index = 0;
1117 td_new->files_size = 0;
1118 td_new->files = NULL;
1121 job_add_num = numjobs - 1;
1123 if (add_job(td_new, jobname, job_add_num, 1, client_type))
1134 * Parse as if 'o' was a command line
1136 void add_job_opts(const char **o, int client_type)
1138 struct thread_data *td, *td_parent;
1139 int i, in_global = 1;
1143 td_parent = td = NULL;
1145 if (!strncmp(o[i], "name", 4)) {
1148 add_job(td, jobname, 0, 0, client_type);
1150 sprintf(jobname, "%s", o[i] + 5);
1152 if (in_global && !td_parent)
1153 td_parent = get_new_job(1, &def_thread, 0);
1154 else if (!in_global && !td) {
1156 td_parent = &def_thread;
1157 td = get_new_job(0, td_parent, 0);
1160 fio_options_parse(td_parent, (char **) &o[i], 1, 0);
1162 fio_options_parse(td, (char **) &o[i], 1, 0);
1167 add_job(td, jobname, 0, 0, client_type);
1170 static int skip_this_section(const char *name)
1174 if (!nr_job_sections)
1176 if (!strncmp(name, "global", 6))
1179 for (i = 0; i < nr_job_sections; i++)
1180 if (!strcmp(job_sections[i], name))
1186 static int is_empty_or_comment(char *line)
1190 for (i = 0; i < strlen(line); i++) {
1195 if (!isspace((int) line[i]) && !iscntrl((int) line[i]))
1203 * This is our [ini] type file parser.
1205 int parse_jobs_ini(char *file, int is_buf, int stonewall_flag, int type)
1207 unsigned int global;
1208 struct thread_data *td;
1209 char *string, *name;
1212 int ret = 0, stonewall;
1215 int inside_skip = 0;
1217 int i, alloc_opts, num_opts;
1222 if (!strcmp(file, "-"))
1225 f = fopen(file, "r");
1228 perror("fopen job file");
1233 string = malloc(4096);
1236 * it's really 256 + small bit, 280 should suffice
1239 memset(name, 0, 280);
1242 opts = malloc(sizeof(char *) * alloc_opts);
1245 stonewall = stonewall_flag;
1248 * if skip_fgets is set, we already have loaded a line we
1253 p = strsep(&file, "\n");
1255 p = fgets(string, 4096, f);
1261 strip_blank_front(&p);
1264 if (is_empty_or_comment(p))
1266 if (sscanf(p, "[%255[^\n]]", name) != 1) {
1269 log_err("fio: option <%s> outside of [] job section\n",
1274 name[strlen(name) - 1] = '\0';
1276 if (skip_this_section(name)) {
1282 global = !strncmp(name, "global", 6);
1288 log_info("--name=%s ", name);
1292 td = get_new_job(global, &def_thread, 0);
1299 * Separate multiple job files by a stonewall
1301 if (!global && stonewall) {
1302 td->o.stonewall = stonewall;
1307 memset(opts, 0, alloc_opts * sizeof(char *));
1311 p = strsep(&file, "\n");
1313 p = fgets(string, 4096, f);
1317 if (is_empty_or_comment(p))
1320 strip_blank_front(&p);
1323 * new section, break out and make sure we don't
1324 * fgets() a new line at the top.
1333 if (num_opts == alloc_opts) {
1335 opts = realloc(opts,
1336 alloc_opts * sizeof(char *));
1339 opts[num_opts] = strdup(p);
1343 ret = fio_options_parse(td, opts, num_opts, dump_cmdline);
1345 ret = add_job(td, name, 0, 0, type);
1347 log_err("fio: job %s dropped\n", name);
1351 for (i = 0; i < num_opts; i++)
1360 while (i < nr_job_sections) {
1361 free(job_sections[i]);
1365 for (i = 0; i < num_opts; i++)
1371 if (!is_buf && f != stdin)
1376 static int fill_def_thread(void)
1378 memset(&def_thread, 0, sizeof(def_thread));
1380 fio_getaffinity(getpid(), &def_thread.o.cpumask);
1381 def_thread.o.timeout = def_timeout;
1382 def_thread.o.error_dump = 1;
1384 * fill default options
1386 fio_fill_default_options(&def_thread);
1390 static void usage(const char *name)
1392 printf("%s\n", fio_version_string);
1393 printf("%s [options] [job options] <job file(s)>\n", name);
1394 printf(" --debug=options\tEnable debug logging. May be one/more of:\n"
1395 "\t\t\tprocess,file,io,mem,blktrace,verify,random,parse,\n"
1396 "\t\t\tdiskutil,job,mutex,profile,time,net,rate\n");
1397 printf(" --parse-only\t\tParse options only, don't start any IO\n");
1398 printf(" --output\t\tWrite output to file\n");
1399 printf(" --runtime\t\tRuntime in seconds\n");
1400 printf(" --latency-log\t\tGenerate per-job latency logs\n");
1401 printf(" --bandwidth-log\tGenerate per-job bandwidth logs\n");
1402 printf(" --minimal\t\tMinimal (terse) output\n");
1403 printf(" --output-format=x\tOutput format (terse,json,normal)\n");
1404 printf(" --terse-version=x\tSet terse version output format to 'x'\n");
1405 printf(" --version\t\tPrint version info and exit\n");
1406 printf(" --help\t\tPrint this page\n");
1407 printf(" --cpuclock-test\tPerform test/validation of CPU clock\n");
1408 printf(" --cmdhelp=cmd\t\tPrint command help, \"all\" for all of"
1410 printf(" --enghelp=engine\tPrint ioengine help, or list"
1411 " available ioengines\n");
1412 printf(" --enghelp=engine,cmd\tPrint help for an ioengine"
1414 printf(" --showcmd\t\tTurn a job file into command line options\n");
1415 printf(" --eta=when\t\tWhen ETA estimate should be printed\n");
1416 printf(" \t\tMay be \"always\", \"never\" or \"auto\"\n");
1417 printf(" --eta-newline=time\tForce a new line for every 'time'");
1418 printf(" period passed\n");
1419 printf(" --status-interval=t\tForce full status dump every");
1420 printf(" 't' period passed\n");
1421 printf(" --readonly\t\tTurn on safety read-only checks, preventing"
1423 printf(" --section=name\tOnly run specified section in job file\n");
1424 printf(" --alloc-size=kb\tSet smalloc pool to this size in kb"
1426 printf(" --warnings-fatal\tFio parser warnings are fatal\n");
1427 printf(" --max-jobs=nr\t\tMaximum number of threads/processes to support\n");
1428 printf(" --server=args\t\tStart a backend fio server\n");
1429 printf(" --daemonize=pidfile\tBackground fio server, write pid to file\n");
1430 printf(" --client=hostname\tTalk to remote backend fio server at hostname\n");
1431 printf(" --idle-prof=option\tReport cpu idleness on a system or percpu basis\n"
1432 "\t\t\t(option=system,percpu) or run unit work\n"
1433 "\t\t\tcalibration only (option=calibrate)\n");
1434 printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
1435 printf("\n Jens Axboe <jaxboe@fusionio.com>\n");
1438 #ifdef FIO_INC_DEBUG
1439 struct debug_level debug_levels[] = {
1440 { .name = "process",
1441 .help = "Process creation/exit logging",
1442 .shift = FD_PROCESS,
1445 .help = "File related action logging",
1449 .help = "IO and IO engine action logging (offsets, queue, completions, etc)",
1453 .help = "Memory allocation/freeing logging",
1456 { .name = "blktrace",
1457 .help = "blktrace action logging",
1458 .shift = FD_BLKTRACE,
1461 .help = "IO verification action logging",
1465 .help = "Random generation logging",
1469 .help = "Parser logging",
1472 { .name = "diskutil",
1473 .help = "Disk utility logging actions",
1474 .shift = FD_DISKUTIL,
1477 .help = "Logging related to creating/destroying jobs",
1481 .help = "Mutex logging",
1484 { .name = "profile",
1485 .help = "Logging related to profiles",
1486 .shift = FD_PROFILE,
1489 .help = "Logging related to time keeping functions",
1493 .help = "Network logging",
1497 .help = "Rate logging",
1503 static int set_debug(const char *string)
1505 struct debug_level *dl;
1506 char *p = (char *) string;
1510 if (!strcmp(string, "?") || !strcmp(string, "help")) {
1511 log_info("fio: dumping debug options:");
1512 for (i = 0; debug_levels[i].name; i++) {
1513 dl = &debug_levels[i];
1514 log_info("%s,", dl->name);
1520 while ((opt = strsep(&p, ",")) != NULL) {
1523 if (!strncmp(opt, "all", 3)) {
1524 log_info("fio: set all debug options\n");
1529 for (i = 0; debug_levels[i].name; i++) {
1530 dl = &debug_levels[i];
1531 found = !strncmp(opt, dl->name, strlen(dl->name));
1535 if (dl->shift == FD_JOB) {
1536 opt = strchr(opt, ':');
1538 log_err("fio: missing job number\n");
1542 fio_debug_jobno = atoi(opt);
1543 log_info("fio: set debug jobno %d\n",
1546 log_info("fio: set debug option %s\n", opt);
1547 fio_debug |= (1UL << dl->shift);
1553 log_err("fio: debug mask %s not found\n", opt);
1558 static int set_debug(const char *string)
1560 log_err("fio: debug tracing not included in build\n");
1565 static void fio_options_fill_optstring(void)
1567 char *ostr = cmd_optstr;
1571 while (l_opts[i].name) {
1572 ostr[c++] = l_opts[i].val;
1573 if (l_opts[i].has_arg == required_argument)
1575 else if (l_opts[i].has_arg == optional_argument) {
1584 static int client_flag_set(char c)
1589 while (l_opts[i].name) {
1590 int val = l_opts[i].val;
1592 if (c == (val & 0xff))
1593 return (val & FIO_CLIENT_FLAG);
1601 void parse_cmd_client(void *client, char *opt)
1603 fio_client_add_cmd_option(client, opt);
1606 int parse_cmd_line(int argc, char *argv[], int client_type)
1608 struct thread_data *td = NULL;
1609 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1610 char *ostr = cmd_optstr;
1611 void *pid_file = NULL;
1612 void *cur_client = NULL;
1616 * Reset optind handling, since we may call this multiple times
1621 while ((c = getopt_long_only(argc, argv, ostr, l_opts, &lidx)) != -1) {
1624 if ((c & FIO_CLIENT_FLAG) || client_flag_set(c)) {
1625 parse_cmd_client(cur_client, argv[optind - 1]);
1626 c &= ~FIO_CLIENT_FLAG;
1631 smalloc_pool_size = atoi(optarg);
1634 def_timeout = atoi(optarg);
1643 f_out = fopen(optarg, "w+");
1645 perror("fopen output");
1651 output_format = FIO_OUTPUT_TERSE;
1654 if (!strcmp(optarg, "minimal") ||
1655 !strcmp(optarg, "terse") ||
1656 !strcmp(optarg, "csv"))
1657 output_format = FIO_OUTPUT_TERSE;
1658 else if (!strcmp(optarg, "json"))
1659 output_format = FIO_OUTPUT_JSON;
1661 output_format = FIO_OUTPUT_NORMAL;
1671 fio_show_option_help(optarg);
1677 fio_show_ioengine_help(optarg);
1689 log_info("%s\n", fio_version_string);
1694 terse_version = atoi(optarg);
1695 if (!(terse_version == 2 || terse_version == 3 ||
1696 terse_version == 4)) {
1697 log_err("fio: bad terse version format\n");
1703 if (!strcmp("always", optarg))
1704 eta_print = FIO_ETA_ALWAYS;
1705 else if (!strcmp("never", optarg))
1706 eta_print = FIO_ETA_NEVER;
1711 if (str_to_decimal(optarg, &t, 0, NULL)) {
1712 log_err("fio: failed parsing eta time %s\n", optarg);
1720 if (set_debug(optarg))
1729 if (!strcmp(optarg, "global")) {
1730 log_err("fio: can't use global as only "
1736 new_size = (nr_job_sections + 1) * sizeof(char *);
1737 job_sections = realloc(job_sections, new_size);
1738 job_sections[nr_job_sections] = strdup(optarg);
1743 exec_profile = strdup(optarg);
1745 case FIO_GETOPT_JOB: {
1746 const char *opt = l_opts[lidx].name;
1749 if (!strncmp(opt, "name", 4) && td) {
1750 ret = add_job(td, td->o.name ?: "fio", 0, 0, client_type);
1756 int is_section = !strncmp(opt, "name", 4);
1759 if (!is_section || !strncmp(val, "global", 6))
1762 if (is_section && skip_this_section(val))
1765 td = get_new_job(global, &def_thread, 1);
1766 if (!td || ioengine_load(td))
1768 fio_options_set_ioengine_opts(l_opts, td);
1771 if ((!val || !strlen(val)) &&
1772 l_opts[lidx].has_arg == required_argument) {
1773 log_err("fio: option %s requires an argument\n", opt);
1776 ret = fio_cmd_option_parse(td, opt, val);
1786 if (!ret && !strcmp(opt, "ioengine")) {
1788 if (ioengine_load(td))
1790 fio_options_set_ioengine_opts(l_opts, td);
1794 case FIO_GETOPT_IOENGINE: {
1795 const char *opt = l_opts[lidx].name;
1797 ret = fio_cmd_ioengine_option_parse(td, opt, val);
1804 max_jobs = atoi(optarg);
1805 if (!max_jobs || max_jobs > REAL_MAX_JOBS) {
1806 log_err("fio: invalid max jobs: %d\n", max_jobs);
1813 log_err("fio: can't be both client and server\n");
1819 fio_server_set_arg(optarg);
1824 pid_file = strdup(optarg);
1827 if ((ret = fio_idle_prof_parse_opt(optarg))) {
1828 /* exit on error and calibration only */
1836 log_err("fio: can't be both client and server\n");
1841 if (fio_client_add(&fio_client_ops, optarg, &cur_client)) {
1842 log_err("fio: failed adding client %s\n", optarg);
1848 * If the next argument exists and isn't an option,
1849 * assume it's a job file for this client only.
1851 while (optind < argc) {
1852 if (!strncmp(argv[optind], "--", 2) ||
1853 !strncmp(argv[optind], "-", 1))
1856 fio_client_add_ini_file(cur_client, argv[optind]);
1862 exit_val = fio_monotonic_clocktest();
1867 if (check_str_time(optarg, &val)) {
1868 log_err("fio: failed parsing time %s\n", optarg);
1873 status_interval = val * 1000;
1877 log_err("%s: unrecognized option '%s'\n", argv[0],
1888 if (do_exit && !(is_backend || nr_clients))
1891 if (nr_clients && fio_clients_connect()) {
1897 if (is_backend && backend)
1898 return fio_start_server(pid_file);
1902 ret = add_job(td, td->o.name ?: "fio", 0, 0, client_type);
1905 while (!ret && optind < argc) {
1907 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1908 ini_file[ini_idx - 1] = strdup(argv[optind]);
1915 int fio_init_options(void)
1920 fio_options_fill_optstring();
1921 fio_options_dup_and_init(l_opts);
1925 if (fill_def_thread())
1931 extern int fio_check_options(struct thread_options *);
1933 int parse_options(int argc, char *argv[])
1935 const int type = FIO_CLIENT_TYPE_CLI;
1938 if (fio_init_options())
1940 if (fio_test_cconv(&def_thread.o))
1941 log_err("fio: failed internal cconv test\n");
1943 job_files = parse_cmd_line(argc, argv, type);
1945 if (job_files > 0) {
1946 for (i = 0; i < job_files; i++) {
1947 if (fill_def_thread())
1950 if (fio_clients_send_ini(ini_file[i]))
1953 } else if (!is_backend) {
1954 if (parse_jobs_ini(ini_file[i], 0, i, type))
1959 } else if (nr_clients) {
1960 if (fill_def_thread())
1962 if (fio_clients_send_ini(NULL))
1967 fio_options_free(&def_thread);
1969 if (!thread_number) {
1974 if (is_backend || nr_clients)
1979 log_err("No jobs(s) defined\n\n");
1989 if (def_thread.o.gtod_offload) {
1991 fio_gtod_offload = 1;
1992 fio_gtod_cpu = def_thread.o.gtod_cpu;
1995 if (output_format == FIO_OUTPUT_NORMAL)
1996 log_info("%s\n", fio_version_string);
2001 void options_default_fill(struct thread_options *o)
2003 memcpy(o, &def_thread.o, sizeof(*o));