2 * This file contains job initialization and setup functions.
14 #include <sys/types.h>
22 static char fio_version_string[] = "fio 1.23";
24 #define FIO_RANDSEED (0xb1899bedUL)
26 static char **ini_file;
27 static int max_jobs = MAX_JOBS;
28 static int dump_cmdline;
30 static 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 #ifndef FIO_HAVE_CPU_AFFINITY
218 if (td->o.gtod_cpu) {
219 log_err("fio: platform must support CPU affinity for"
220 "gettimeofday() offloading\n");
225 if (read_only && td_write(td)) {
226 log_err("fio: job <%s> has write bit set, but fio is in"
227 " read-only mode\n", td->o.name);
231 if (o->write_iolog_file && o->read_iolog_file) {
232 log_err("fio: read iolog overrides write_iolog\n");
233 free(o->write_iolog_file);
234 o->write_iolog_file = NULL;
238 * only really works for sequential io for now, and with 1 file
240 if (o->zone_size && td_random(td) && o->open_files == 1)
244 * Reads can do overwrites, we always need to pre-create the file
246 if (td_read(td) || td_rw(td))
249 if (!o->min_bs[DDIR_READ])
250 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
251 if (!o->max_bs[DDIR_READ])
252 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
253 if (!o->min_bs[DDIR_WRITE])
254 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
255 if (!o->max_bs[DDIR_WRITE])
256 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
258 o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
260 if (!o->file_size_high)
261 o->file_size_high = o->file_size_low;
263 if (o->norandommap && o->verify != VERIFY_NONE) {
264 log_err("fio: norandommap given, verify disabled\n");
265 o->verify = VERIFY_NONE;
267 if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
268 log_err("fio: bs_unaligned may not work with raw io\n");
271 * thinktime_spin must be less than thinktime
273 if (o->thinktime_spin > o->thinktime)
274 o->thinktime_spin = o->thinktime;
277 * The low water mark cannot be bigger than the iodepth
279 if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
281 * syslet work around - if the workload is sequential,
282 * we want to let the queue drain all the way down to
283 * avoid seeking between async threads
285 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
288 o->iodepth_low = o->iodepth;
292 * If batch number isn't set, default to the same as iodepth
294 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
295 o->iodepth_batch = o->iodepth;
297 if (o->nr_files > td->files_index)
298 o->nr_files = td->files_index;
300 if (o->open_files > o->nr_files || !o->open_files)
301 o->open_files = o->nr_files;
303 if ((o->rate && o->rate_iops) || (o->ratemin && o->rate_iops_min)) {
304 log_err("fio: rate and rate_iops are mutually exclusive\n");
307 if ((o->rate < o->ratemin) || (o->rate_iops < o->rate_iops_min)) {
308 log_err("fio: minimum rate exceeds rate\n");
312 if (!o->timeout && o->time_based) {
313 log_err("fio: time_based requires a runtime/timeout setting\n");
317 if (o->fill_device && !o->size)
320 if (td_rw(td) && td->o.verify != VERIFY_NONE)
321 log_info("fio: mixed read/write workload with verify. May not "
322 "work as expected, unless you pre-populated the file\n");
324 if (td->o.verify != VERIFY_NONE)
325 td->o.refill_buffers = 1;
331 * This function leaks the buffer
333 static char *to_kmg(unsigned int val)
335 char *buf = malloc(32);
336 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
347 snprintf(buf, 31, "%u%c", val, *p);
351 /* External engines are specified by "external:name.o") */
352 static const char *get_engine_name(const char *str)
354 char *p = strstr(str, ":");
360 strip_blank_front(&p);
365 static int exists_and_not_file(const char *filename)
369 if (lstat(filename, &sb) == -1)
372 if (S_ISREG(sb.st_mode))
379 * Initialize the various random states we need (random io, block size ranges,
380 * read/write mix, etc).
382 static int init_random_state(struct thread_data *td)
384 unsigned long seeds[6];
387 fd = open("/dev/urandom", O_RDONLY);
389 td_verror(td, errno, "open");
393 if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
394 td_verror(td, EIO, "read");
401 os_random_seed(seeds[0], &td->bsrange_state);
402 os_random_seed(seeds[1], &td->verify_state);
403 os_random_seed(seeds[2], &td->rwmix_state);
405 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
406 os_random_seed(seeds[3], &td->next_file_state);
408 os_random_seed(seeds[5], &td->file_size_state);
413 if (td->o.rand_repeatable)
414 seeds[4] = FIO_RANDSEED * td->thread_number;
416 os_random_seed(seeds[4], &td->random_state);
421 * Adds a job to the list of things todo. Sanitizes the various options
422 * to make sure we don't have conflicts, and initializes various
425 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
427 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
428 "randread", "randwrite", "randrw" };
431 char fname[PATH_MAX];
432 int numjobs, file_alloced;
435 * the def_thread is just for options, it's not a real job
437 if (td == &def_thread)
441 * if we are just dumping the output command line, don't add the job
448 engine = get_engine_name(td->o.ioengine);
449 td->io_ops = load_ioengine(td, engine);
451 log_err("fio: failed to load engine %s\n", engine);
455 if (td->o.use_thread)
461 td->io_ops->flags |= FIO_RAWIO;
464 if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
467 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
468 add_file(td, jobname);
470 for (i = 0; i < td->o.nr_files; i++) {
471 sprintf(fname, "%s.%d.%d", jobname,
472 td->thread_number, i);
478 if (fixup_options(td))
481 if (td->io_ops->flags & FIO_DISKLESSIO) {
484 for_each_file(td, f, i)
485 f->real_file_size = -1ULL;
488 td->mutex = fio_mutex_init(0);
490 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
491 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
492 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
493 td->ddir_nr = td->o.ddir_nr;
495 if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
496 && prev_group_jobs) {
501 td->groupid = groupid;
504 if (init_random_state(td))
510 if (td->o.write_lat_log) {
511 setup_log(&td->ts.slat_log);
512 setup_log(&td->ts.clat_log);
514 if (td->o.write_bw_log)
515 setup_log(&td->ts.bw_log);
518 td->o.name = strdup(jobname);
522 if (!strcmp(td->io_ops->name, "cpuio")) {
523 log_info("%s: ioengine=cpu, cpuload=%u,"
524 " cpucycle=%u\n", td->o.name,
528 char *c1, *c2, *c3, *c4;
530 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
531 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
532 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
533 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
535 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
536 " ioengine=%s, iodepth=%u\n",
537 td->o.name, td->groupid,
538 ddir_str[td->o.td_ddir],
548 } else if (job_add_num == 1)
553 * recurse add identical jobs, clear numjobs and stonewall options
554 * as they don't apply to sub-jobs
556 numjobs = td->o.numjobs;
558 struct thread_data *td_new = get_new_job(0, td);
563 td_new->o.numjobs = 1;
564 td_new->o.stonewall = 0;
565 td_new->o.new_group = 0;
568 td_new->o.filename = NULL;
569 td_new->files_index = 0;
570 td_new->files = NULL;
573 job_add_num = numjobs - 1;
575 if (add_job(td_new, jobname, job_add_num))
585 static int skip_this_section(const char *name)
589 if (!strncmp(name, "global", 6))
592 return strncmp(job_section, name, strlen(job_section));
595 static int is_empty_or_comment(char *line)
599 for (i = 0; i < strlen(line); i++) {
604 if (!isspace(line[i]) && !iscntrl(line[i]))
612 * This is our [ini] type file parser.
614 static int parse_jobs_ini(char *file, int stonewall_flag)
617 struct thread_data *td;
621 int ret = 0, stonewall;
626 int i, alloc_opts, num_opts;
628 if (!strcmp(file, "-"))
631 f = fopen(file, "r");
634 perror("fopen job file");
638 string = malloc(4096);
641 * it's really 256 + small bit, 280 should suffice
644 memset(name, 0, 280);
647 opts = malloc(sizeof(char *) * alloc_opts);
650 stonewall = stonewall_flag;
653 * if skip_fgets is set, we already have loaded a line we
657 p = fgets(string, 4095, f);
663 strip_blank_front(&p);
666 if (is_empty_or_comment(p))
668 if (sscanf(p, "[%255s]", name) != 1) {
671 log_err("fio: option <%s> outside of [] job section\n",
676 if (skip_this_section(name)) {
682 global = !strncmp(name, "global", 6);
684 name[strlen(name) - 1] = '\0';
690 log_info("--name=%s ", name);
694 td = get_new_job(global, &def_thread);
701 * Seperate multiple job files by a stonewall
703 if (!global && stonewall) {
704 td->o.stonewall = stonewall;
709 memset(opts, 0, alloc_opts * sizeof(char *));
711 while ((p = fgets(string, 4096, f)) != NULL) {
712 if (is_empty_or_comment(p))
715 strip_blank_front(&p);
718 * new section, break out and make sure we don't
719 * fgets() a new line at the top.
728 if (num_opts == alloc_opts) {
731 alloc_opts * sizeof(char *));
734 opts[num_opts] = strdup(p);
738 ret = fio_options_parse(td, opts, num_opts);
741 for (i = 0; i < num_opts; i++)
742 log_info("--%s ", opts[i]);
744 ret = add_job(td, name, 0);
746 log_err("fio: job %s dropped\n", name);
750 for (i = 0; i < num_opts; i++)
758 for (i = 0; i < num_opts; i++)
769 static int fill_def_thread(void)
771 memset(&def_thread, 0, sizeof(def_thread));
773 fio_getaffinity(getpid(), &def_thread.o.cpumask);
776 * fill default options
778 fio_fill_default_options(&def_thread);
780 def_thread.o.timeout = def_timeout;
781 def_thread.o.write_bw_log = write_bw_log;
782 def_thread.o.write_lat_log = write_lat_log;
787 static void free_shm(void)
789 struct shmid_ds sbuf;
796 fio_debug_jobp = NULL;
798 shmctl(shm_id, IPC_RMID, &sbuf);
805 * The thread area is shared between the main process and the job
806 * threads/processes. So setup a shared memory segment that will hold
807 * all the job info. We use the end of the region for keeping track of
808 * open files across jobs, for file sharing.
810 static int setup_thread_area(void)
815 * 1024 is too much on some machines, scale max_jobs if
816 * we get a failure that looks like too large a shm segment
819 size_t size = max_jobs * sizeof(struct thread_data);
821 size += file_hash_size;
822 size += sizeof(unsigned int);
824 shm_id = shmget(0, size, IPC_CREAT | 0600);
827 if (errno != EINVAL) {
838 threads = shmat(shm_id, NULL, 0);
839 if (threads == (void *) -1) {
844 memset(threads, 0, max_jobs * sizeof(struct thread_data));
845 hash = (void *) threads + max_jobs * sizeof(struct thread_data);
846 fio_debug_jobp = (void *) hash + file_hash_size;
847 *fio_debug_jobp = -1;
848 file_hash_init(hash);
853 static void usage(const char *name)
855 printf("%s\n", fio_version_string);
856 printf("%s [options] [job options] <job file(s)>\n", name);
857 printf("\t--debug=options\tEnable debug logging\n");
858 printf("\t--output\tWrite output to file\n");
859 printf("\t--timeout\tRuntime in seconds\n");
860 printf("\t--latency-log\tGenerate per-job latency logs\n");
861 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
862 printf("\t--minimal\tMinimal (terse) output\n");
863 printf("\t--version\tPrint version info and exit\n");
864 printf("\t--help\t\tPrint this page\n");
865 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
867 printf("\t--showcmd\tTurn a job file into command line options\n");
868 printf("\t--eta=when\tWhen ETA estimate should be printed\n");
869 printf("\t \tMay be \"always\", \"never\" or \"auto\"\n");
870 printf("\t--readonly\tTurn on safety read-only checks, preventing"
872 printf("\t--section=name\tOnly run specified section in job file\n");
873 printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
878 struct debug_level debug_levels[] = {
879 { .name = "process", .shift = FD_PROCESS, },
880 { .name = "file", .shift = FD_FILE, },
881 { .name = "io", .shift = FD_IO, },
882 { .name = "mem", .shift = FD_MEM, },
883 { .name = "blktrace", .shift = FD_BLKTRACE },
884 { .name = "verify", .shift = FD_VERIFY },
885 { .name = "random", .shift = FD_RANDOM },
886 { .name = "parse", .shift = FD_PARSE },
887 { .name = "diskutil", .shift = FD_DISKUTIL },
888 { .name = "job", .shift = FD_JOB },
892 static int set_debug(const char *string)
894 struct debug_level *dl;
895 char *p = (char *) string;
899 if (!strcmp(string, "?") || !strcmp(string, "help")) {
902 log_info("fio: dumping debug options:");
903 for (i = 0; debug_levels[i].name; i++) {
904 dl = &debug_levels[i];
905 log_info("%s,", dl->name);
911 while ((opt = strsep(&p, ",")) != NULL) {
914 if (!strncmp(opt, "all", 3)) {
915 log_info("fio: set all debug options\n");
920 for (i = 0; debug_levels[i].name; i++) {
921 dl = &debug_levels[i];
922 found = !strncmp(opt, dl->name, strlen(dl->name));
926 if (dl->shift == FD_JOB) {
927 opt = strchr(opt, ':');
929 log_err("fio: missing job number\n");
933 fio_debug_jobno = atoi(opt);
934 log_info("fio: set debug jobno %d\n",
937 log_info("fio: set debug option %s\n", opt);
938 fio_debug |= (1UL << dl->shift);
944 log_err("fio: debug mask %s not found\n", opt);
949 static int set_debug(const char *string)
951 log_err("fio: debug tracing not included in build\n");
956 static int parse_cmd_line(int argc, char *argv[])
958 struct thread_data *td = NULL;
959 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
961 while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
964 smalloc_pool_size = atoi(optarg);
967 def_timeout = atoi(optarg);
976 f_out = fopen(optarg, "w+");
978 perror("fopen output");
990 exit(fio_show_option_help(optarg));
998 printf("%s\n", fio_version_string);
1001 if (!strcmp("always", optarg))
1002 eta_print = FIO_ETA_ALWAYS;
1003 else if (!strcmp("never", optarg))
1004 eta_print = FIO_ETA_NEVER;
1007 if (set_debug(optarg))
1011 if (!strcmp(optarg, "global")) {
1012 log_err("fio: can't use global as only "
1020 job_section = strdup(optarg);
1022 case FIO_GETOPT_JOB: {
1023 const char *opt = l_opts[lidx].name;
1026 if (!strncmp(opt, "name", 4) && td) {
1027 ret = add_job(td, td->o.name ?: "fio", 0);
1035 int is_section = !strncmp(opt, "name", 4);
1038 if (!is_section || !strncmp(val, "global", 6))
1041 if (is_section && skip_this_section(val))
1044 td = get_new_job(global, &def_thread);
1049 ret = fio_cmd_option_parse(td, opt, val);
1064 ret = add_job(td, td->o.name ?: "fio", 0);
1069 while (optind < argc) {
1071 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1072 ini_file[ini_idx - 1] = strdup(argv[optind]);
1079 int parse_options(int argc, char *argv[])
1086 fio_options_dup_and_init(l_opts);
1088 if (setup_thread_area())
1090 if (fill_def_thread())
1093 job_files = parse_cmd_line(argc, argv);
1095 for (i = 0; i < job_files; i++) {
1096 if (fill_def_thread())
1098 if (parse_jobs_ini(ini_file[i], i))
1104 options_mem_free(&def_thread);
1106 if (!thread_number) {
1110 log_err("No jobs defined(s)\n");
1115 if (def_thread.o.gtod_offload) {
1117 fio_gtod_offload = 1;
1118 fio_gtod_cpu = def_thread.o.gtod_cpu;