2 * This file contains job initialization and setup functions.
14 #include <sys/types.h>
22 static char fio_version_string[] = "fio 1.20";
24 #define FIO_RANDSEED (0xb1899bedUL)
26 static char **ini_file;
27 static int max_jobs = MAX_JOBS;
28 static int dump_cmdline;
30 struct thread_data def_thread;
31 struct thread_data *threads = NULL;
33 int exitall_on_terminate = 0;
36 unsigned long long mlock_size = 0;
39 char *job_section = NULL;
44 static int def_timeout;
45 static int write_lat_log;
47 static int prev_group_jobs;
49 unsigned long fio_debug = 0;
50 unsigned int fio_debug_jobno = -1;
51 unsigned int *fio_debug_jobp = NULL;
54 * Command line options. These will contain the above, plus a few
55 * extra that only pertain to fio itself and not jobs.
57 static struct option l_opts[FIO_NR_OPTIONS] = {
60 .has_arg = required_argument,
65 .has_arg = required_argument,
69 .name = "latency-log",
70 .has_arg = required_argument,
74 .name = "bandwidth-log",
75 .has_arg = required_argument,
80 .has_arg = optional_argument,
85 .has_arg = no_argument,
90 .has_arg = no_argument,
95 .has_arg = optional_argument,
100 .has_arg = no_argument,
105 .has_arg = no_argument,
110 .has_arg = required_argument,
115 .has_arg = required_argument,
120 .has_arg = required_argument,
124 .name = "alloc-size",
125 .has_arg = required_argument,
144 * Return a free job structure.
146 static struct thread_data *get_new_job(int global, struct thread_data *parent)
148 struct thread_data *td;
152 if (thread_number >= max_jobs)
155 td = &threads[thread_number++];
158 dup_files(td, parent);
159 options_mem_dupe(td);
161 td->thread_number = thread_number;
165 static void put_job(struct thread_data *td)
167 if (td == &def_thread)
171 log_info("fio: %s\n", td->verror);
173 memset(&threads[td->thread_number - 1], 0, sizeof(*td));
177 static int setup_rate(struct thread_data *td)
179 unsigned long nr_reads_per_msec;
180 unsigned long long rate;
183 if (!td->o.rate && !td->o.rate_iops)
187 bs = td->o.rw_min_bs;
188 else if (td_read(td))
189 bs = td->o.min_bs[DDIR_READ];
191 bs = td->o.min_bs[DDIR_WRITE];
195 nr_reads_per_msec = (rate * 1024 * 1000LL) / bs;
197 nr_reads_per_msec = td->o.rate_iops * 1000UL;
199 if (!nr_reads_per_msec) {
200 log_err("rate lower than supported\n");
204 td->rate_usec_cycle = 1000000000ULL / nr_reads_per_msec;
205 td->rate_pending_usleep = 0;
210 * Lazy way of fixing up options that depend on each other. We could also
211 * define option callback handlers, but this is easier.
213 static int fixup_options(struct thread_data *td)
215 struct thread_options *o = &td->o;
217 if (read_only && td_write(td)) {
218 log_err("fio: job <%s> has write bit set, but fio is in"
219 " read-only mode\n", td->o.name);
223 if (o->rwmix[DDIR_READ] + o->rwmix[DDIR_WRITE] > 100)
224 o->rwmix[DDIR_WRITE] = 100 - o->rwmix[DDIR_READ];
226 if (o->write_iolog_file && o->read_iolog_file) {
227 log_err("fio: read iolog overrides write_iolog\n");
228 free(o->write_iolog_file);
229 o->write_iolog_file = NULL;
233 * only really works for sequential io for now, and with 1 file
235 if (o->zone_size && td_random(td) && o->open_files == 1)
239 * Reads can do overwrites, we always need to pre-create the file
241 if (td_read(td) || td_rw(td))
244 if (!o->min_bs[DDIR_READ])
245 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
246 if (!o->max_bs[DDIR_READ])
247 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
248 if (!o->min_bs[DDIR_WRITE])
249 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
250 if (!o->max_bs[DDIR_WRITE])
251 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
253 o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
255 if (!o->file_size_high)
256 o->file_size_high = o->file_size_low;
258 if (o->norandommap && o->verify != VERIFY_NONE) {
259 log_err("fio: norandommap given, verify disabled\n");
260 o->verify = VERIFY_NONE;
262 if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
263 log_err("fio: bs_unaligned may not work with raw io\n");
266 * thinktime_spin must be less than thinktime
268 if (o->thinktime_spin > o->thinktime)
269 o->thinktime_spin = o->thinktime;
272 * The low water mark cannot be bigger than the iodepth
274 if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
276 * syslet work around - if the workload is sequential,
277 * we want to let the queue drain all the way down to
278 * avoid seeking between async threads
280 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
283 o->iodepth_low = o->iodepth;
287 * If batch number isn't set, default to the same as iodepth
289 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
290 o->iodepth_batch = o->iodepth;
292 if (o->nr_files > td->files_index)
293 o->nr_files = td->files_index;
295 if (o->open_files > o->nr_files || !o->open_files)
296 o->open_files = o->nr_files;
298 if ((o->rate && o->rate_iops) || (o->ratemin && o->rate_iops_min)) {
299 log_err("fio: rate and rate_iops are mutually exclusive\n");
302 if ((o->rate < o->ratemin) || (o->rate_iops < o->rate_iops_min)) {
303 log_err("fio: minimum rate exceeds rate\n");
307 if (!o->timeout && o->time_based) {
308 log_err("fio: time_based requires a runtime/timeout setting\n");
312 if (o->fill_device && !o->size)
313 o->size = ULONG_LONG_MAX;
315 if (td_rw(td) && td->o.verify != VERIFY_NONE)
316 log_info("fio: mixed read/write workload with verify. May not "
317 "work as expected, unless you pre-populated the file\n");
319 if (td->o.verify != VERIFY_NONE)
320 td->o.refill_buffers = 1;
326 * This function leaks the buffer
328 static char *to_kmg(unsigned int val)
330 char *buf = malloc(32);
331 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
342 snprintf(buf, 31, "%u%c", val, *p);
346 /* External engines are specified by "external:name.o") */
347 static const char *get_engine_name(const char *str)
349 char *p = strstr(str, ":");
355 strip_blank_front(&p);
360 static int exists_and_not_file(const char *filename)
364 if (lstat(filename, &sb) == -1)
367 if (S_ISREG(sb.st_mode))
374 * Initialize the various random states we need (random io, block size ranges,
375 * read/write mix, etc).
377 static int init_random_state(struct thread_data *td)
379 unsigned long seeds[6];
382 fd = open("/dev/urandom", O_RDONLY);
384 td_verror(td, errno, "open");
388 if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
389 td_verror(td, EIO, "read");
396 os_random_seed(seeds[0], &td->bsrange_state);
397 os_random_seed(seeds[1], &td->verify_state);
398 os_random_seed(seeds[2], &td->rwmix_state);
400 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
401 os_random_seed(seeds[3], &td->next_file_state);
403 os_random_seed(seeds[5], &td->file_size_state);
408 if (td->o.rand_repeatable)
409 seeds[4] = FIO_RANDSEED * td->thread_number;
411 os_random_seed(seeds[4], &td->random_state);
416 * Adds a job to the list of things todo. Sanitizes the various options
417 * to make sure we don't have conflicts, and initializes various
420 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
422 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
423 "randread", "randwrite", "randrw" };
426 char fname[PATH_MAX];
427 int numjobs, file_alloced;
430 * the def_thread is just for options, it's not a real job
432 if (td == &def_thread)
436 * if we are just dumping the output command line, don't add the job
443 engine = get_engine_name(td->o.ioengine);
444 td->io_ops = load_ioengine(td, engine);
446 log_err("fio: failed to load engine %s\n", engine);
450 if (td->o.use_thread)
456 td->io_ops->flags |= FIO_RAWIO;
459 if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
462 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
463 add_file(td, jobname);
465 for (i = 0; i < td->o.nr_files; i++) {
466 sprintf(fname, "%s.%d.%d", jobname,
467 td->thread_number, i);
473 if (fixup_options(td))
476 if (td->io_ops->flags & FIO_DISKLESSIO) {
479 for_each_file(td, f, i)
480 f->real_file_size = -1ULL;
483 td->mutex = fio_mutex_init(0);
485 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
486 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
487 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
488 td->ddir_nr = td->o.ddir_nr;
490 if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
491 && prev_group_jobs) {
496 td->groupid = groupid;
499 if (init_random_state(td))
505 if (td->o.write_lat_log) {
506 setup_log(&td->ts.slat_log);
507 setup_log(&td->ts.clat_log);
509 if (td->o.write_bw_log)
510 setup_log(&td->ts.bw_log);
513 td->o.name = strdup(jobname);
517 if (!strcmp(td->io_ops->name, "cpuio")) {
518 log_info("%s: ioengine=cpu, cpuload=%u,"
519 " cpucycle=%u\n", td->o.name,
523 char *c1, *c2, *c3, *c4;
525 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
526 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
527 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
528 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
530 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
531 " ioengine=%s, iodepth=%u\n",
532 td->o.name, td->groupid,
533 ddir_str[td->o.td_ddir],
543 } else if (job_add_num == 1)
548 * recurse add identical jobs, clear numjobs and stonewall options
549 * as they don't apply to sub-jobs
551 numjobs = td->o.numjobs;
553 struct thread_data *td_new = get_new_job(0, td);
558 td_new->o.numjobs = 1;
559 td_new->o.stonewall = 0;
560 td_new->o.new_group = 0;
563 td_new->o.filename = NULL;
564 td_new->files_index = 0;
565 td_new->files = NULL;
568 job_add_num = numjobs - 1;
570 if (add_job(td_new, jobname, job_add_num))
580 static int skip_this_section(const char *name)
584 if (!strncmp(name, "global", 6))
587 return strncmp(job_section, name, strlen(job_section));
590 static int is_empty_or_comment(char *line)
594 for (i = 0; i < strlen(line); i++) {
599 if (!isspace(line[i]) && !iscntrl(line[i]))
607 * This is our [ini] type file parser.
609 static int parse_jobs_ini(char *file, int stonewall_flag)
612 struct thread_data *td;
616 int ret = 0, stonewall;
621 if (!strcmp(file, "-"))
624 f = fopen(file, "r");
627 perror("fopen job file");
631 string = malloc(4096);
634 * it's really 256 + small bit, 280 should suffice
637 memset(name, 0, 280);
639 stonewall = stonewall_flag;
642 * if skip_fgets is set, we already have loaded a line we
646 p = fgets(string, 4095, f);
652 strip_blank_front(&p);
655 if (is_empty_or_comment(p))
657 if (sscanf(p, "[%255s]", name) != 1) {
660 log_err("fio: option <%s> outside of [] job section\n",
665 if (skip_this_section(name)) {
671 global = !strncmp(name, "global", 6);
673 name[strlen(name) - 1] = '\0';
679 log_info("--name=%s ", name);
683 td = get_new_job(global, &def_thread);
690 * Seperate multiple job files by a stonewall
692 if (!global && stonewall) {
693 td->o.stonewall = stonewall;
697 while ((p = fgets(string, 4096, f)) != NULL) {
698 if (is_empty_or_comment(p))
701 strip_blank_front(&p);
704 * new section, break out and make sure we don't
705 * fgets() a new line at the top.
715 * Don't break here, continue parsing options so we
716 * dump all the bad ones. Makes trial/error fixups
717 * easier on the user.
719 ret |= fio_option_parse(td, p);
720 if (!ret && dump_cmdline)
721 log_info("--%s ", p);
725 ret = add_job(td, name, 0);
727 log_err("fio: job %s dropped\n", name);
742 static int fill_def_thread(void)
744 memset(&def_thread, 0, sizeof(def_thread));
746 fio_getaffinity(getpid(), &def_thread.o.cpumask);
749 * fill default options
751 fio_fill_default_options(&def_thread);
753 def_thread.o.timeout = def_timeout;
754 def_thread.o.write_bw_log = write_bw_log;
755 def_thread.o.write_lat_log = write_lat_log;
760 static void free_shm(void)
762 struct shmid_ds sbuf;
769 fio_debug_jobp = NULL;
771 shmctl(shm_id, IPC_RMID, &sbuf);
778 * The thread area is shared between the main process and the job
779 * threads/processes. So setup a shared memory segment that will hold
780 * all the job info. We use the end of the region for keeping track of
781 * open files across jobs, for file sharing.
783 static int setup_thread_area(void)
788 * 1024 is too much on some machines, scale max_jobs if
789 * we get a failure that looks like too large a shm segment
792 size_t size = max_jobs * sizeof(struct thread_data);
794 size += file_hash_size;
795 size += sizeof(unsigned int);
797 shm_id = shmget(0, size, IPC_CREAT | 0600);
800 if (errno != EINVAL) {
811 threads = shmat(shm_id, NULL, 0);
812 if (threads == (void *) -1) {
817 memset(threads, 0, max_jobs * sizeof(struct thread_data));
818 hash = (void *) threads + max_jobs * sizeof(struct thread_data);
819 fio_debug_jobp = (void *) hash + file_hash_size;
820 *fio_debug_jobp = -1;
821 file_hash_init(hash);
826 static void usage(const char *name)
828 printf("%s\n", fio_version_string);
829 printf("%s [options] [job options] <job file(s)>\n", name);
830 printf("\t--debug=options\tEnable debug logging\n");
831 printf("\t--output\tWrite output to file\n");
832 printf("\t--timeout\tRuntime in seconds\n");
833 printf("\t--latency-log\tGenerate per-job latency logs\n");
834 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
835 printf("\t--minimal\tMinimal (terse) output\n");
836 printf("\t--version\tPrint version info and exit\n");
837 printf("\t--help\t\tPrint this page\n");
838 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
840 printf("\t--showcmd\tTurn a job file into command line options\n");
841 printf("\t--eta=when\tWhen ETA estimate should be printed\n");
842 printf("\t \tMay be \"always\", \"never\" or \"auto\"\n");
843 printf("\t--readonly\tTurn on safety read-only checks, preventing"
845 printf("\t--section=name\tOnly run specified section in job file\n");
846 printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
851 struct debug_level debug_levels[] = {
852 { .name = "process", .shift = FD_PROCESS, },
853 { .name = "file", .shift = FD_FILE, },
854 { .name = "io", .shift = FD_IO, },
855 { .name = "mem", .shift = FD_MEM, },
856 { .name = "blktrace", .shift = FD_BLKTRACE },
857 { .name = "verify", .shift = FD_VERIFY },
858 { .name = "random", .shift = FD_RANDOM },
859 { .name = "parse", .shift = FD_PARSE },
860 { .name = "diskutil", .shift = FD_DISKUTIL },
861 { .name = "job", .shift = FD_JOB },
865 static int set_debug(const char *string)
867 struct debug_level *dl;
868 char *p = (char *) string;
872 if (!strcmp(string, "?") || !strcmp(string, "help")) {
875 log_info("fio: dumping debug options:");
876 for (i = 0; debug_levels[i].name; i++) {
877 dl = &debug_levels[i];
878 log_info("%s,", dl->name);
884 while ((opt = strsep(&p, ",")) != NULL) {
887 if (!strncmp(opt, "all", 3)) {
888 log_info("fio: set all debug options\n");
893 for (i = 0; debug_levels[i].name; i++) {
894 dl = &debug_levels[i];
895 found = !strncmp(opt, dl->name, strlen(dl->name));
899 if (dl->shift == FD_JOB) {
900 opt = strchr(opt, ':');
902 log_err("fio: missing job number\n");
906 fio_debug_jobno = atoi(opt);
907 log_info("fio: set debug jobno %d\n",
910 log_info("fio: set debug option %s\n", opt);
911 fio_debug |= (1UL << dl->shift);
917 log_err("fio: debug mask %s not found\n", opt);
922 static void set_debug(const char *string)
924 log_err("fio: debug tracing not included in build\n");
929 static int parse_cmd_line(int argc, char *argv[])
931 struct thread_data *td = NULL;
932 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
934 while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
937 smalloc_pool_size = atoi(optarg);
940 def_timeout = atoi(optarg);
949 f_out = fopen(optarg, "w+");
951 perror("fopen output");
963 exit(fio_show_option_help(optarg));
971 printf("%s\n", fio_version_string);
974 if (!strcmp("always", optarg))
975 eta_print = FIO_ETA_ALWAYS;
976 else if (!strcmp("never", optarg))
977 eta_print = FIO_ETA_NEVER;
980 if (set_debug(optarg))
984 if (!strcmp(optarg, "global")) {
985 log_err("fio: can't use global as only "
993 job_section = strdup(optarg);
995 case FIO_GETOPT_JOB: {
996 const char *opt = l_opts[lidx].name;
999 if (!strncmp(opt, "name", 4) && td) {
1000 ret = add_job(td, td->o.name ?: "fio", 0);
1008 int is_section = !strncmp(opt, "name", 4);
1011 if (!is_section || !strncmp(val, "global", 6))
1014 if (is_section && skip_this_section(val))
1017 td = get_new_job(global, &def_thread);
1022 ret = fio_cmd_option_parse(td, opt, val);
1037 ret = add_job(td, td->o.name ?: "fio", 0);
1042 while (optind < argc) {
1044 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1045 ini_file[ini_idx - 1] = strdup(argv[optind]);
1052 int parse_options(int argc, char *argv[])
1059 fio_options_dup_and_init(l_opts);
1061 if (setup_thread_area())
1063 if (fill_def_thread())
1066 job_files = parse_cmd_line(argc, argv);
1068 for (i = 0; i < job_files; i++) {
1069 if (fill_def_thread())
1071 if (parse_jobs_ini(ini_file[i], i))
1077 options_mem_free(&def_thread);
1079 if (!thread_number) {
1083 log_err("No jobs defined(s)\n");