2 * This file contains job initialization and setup functions.
14 #include <sys/types.h>
22 static char fio_version_string[] = "fio 1.25.1";
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;
209 static int fixed_block_size(struct thread_options *o)
211 return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
212 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
213 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
217 * Lazy way of fixing up options that depend on each other. We could also
218 * define option callback handlers, but this is easier.
220 static int fixup_options(struct thread_data *td)
222 struct thread_options *o = &td->o;
224 #ifndef FIO_HAVE_PSHARED_MUTEX
225 if (!td->o.use_thread) {
226 log_info("fio: this platform does not support process shared"
227 " mutexes, forcing use of threads. Use the 'thread'"
228 " option to get rid of this warning.\n");
229 td->o.use_thread = 1;
233 #ifndef FIO_HAVE_CPU_AFFINITY
234 if (td->o.gtod_cpu) {
235 log_err("fio: platform must support CPU affinity for"
236 "gettimeofday() offloading\n");
241 if (read_only && td_write(td)) {
242 log_err("fio: job <%s> has write bit set, but fio is in"
243 " read-only mode\n", td->o.name);
247 if (o->write_iolog_file && o->read_iolog_file) {
248 log_err("fio: read iolog overrides write_iolog\n");
249 free(o->write_iolog_file);
250 o->write_iolog_file = NULL;
254 * only really works for sequential io for now, and with 1 file
256 if (o->zone_size && td_random(td) && o->open_files == 1)
260 * Reads can do overwrites, we always need to pre-create the file
262 if (td_read(td) || td_rw(td))
265 if (!o->min_bs[DDIR_READ])
266 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
267 if (!o->max_bs[DDIR_READ])
268 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
269 if (!o->min_bs[DDIR_WRITE])
270 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
271 if (!o->max_bs[DDIR_WRITE])
272 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
274 o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
277 * For random IO, allow blockalign offset other than min_bs.
279 if (!o->ba[DDIR_READ] || !td_random(td))
280 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
281 if (!o->ba[DDIR_WRITE] || !td_random(td))
282 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
284 if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
285 o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
286 !td->o.norandommap) {
287 log_err("fio: Any use of blockalign= turns off randommap\n");
288 td->o.norandommap = 1;
291 if (!o->file_size_high)
292 o->file_size_high = o->file_size_low;
294 if (o->norandommap && o->verify != VERIFY_NONE
295 && !fixed_block_size(o)) {
296 log_err("fio: norandommap given for variable block sizes, "
297 "verify disabled\n");
298 o->verify = VERIFY_NONE;
300 if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
301 log_err("fio: bs_unaligned may not work with raw io\n");
304 * thinktime_spin must be less than thinktime
306 if (o->thinktime_spin > o->thinktime)
307 o->thinktime_spin = o->thinktime;
310 * The low water mark cannot be bigger than the iodepth
312 if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
314 * syslet work around - if the workload is sequential,
315 * we want to let the queue drain all the way down to
316 * avoid seeking between async threads
318 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
321 o->iodepth_low = o->iodepth;
325 * If batch number isn't set, default to the same as iodepth
327 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
328 o->iodepth_batch = o->iodepth;
330 if (o->nr_files > td->files_index)
331 o->nr_files = td->files_index;
333 if (o->open_files > o->nr_files || !o->open_files)
334 o->open_files = o->nr_files;
336 if ((o->rate && o->rate_iops) || (o->ratemin && o->rate_iops_min)) {
337 log_err("fio: rate and rate_iops are mutually exclusive\n");
340 if ((o->rate < o->ratemin) || (o->rate_iops < o->rate_iops_min)) {
341 log_err("fio: minimum rate exceeds rate\n");
345 if (!o->timeout && o->time_based) {
346 log_err("fio: time_based requires a runtime/timeout setting\n");
350 if (o->fill_device && !o->size)
353 if (td_rw(td) && td->o.verify != VERIFY_NONE)
354 log_info("fio: mixed read/write workload with verify. May not "
355 "work as expected, unless you pre-populated the file\n");
357 if (td->o.verify != VERIFY_NONE)
358 td->o.refill_buffers = 1;
364 * This function leaks the buffer
366 static char *to_kmg(unsigned int val)
368 char *buf = malloc(32);
369 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
380 snprintf(buf, 31, "%u%c", val, *p);
384 /* External engines are specified by "external:name.o") */
385 static const char *get_engine_name(const char *str)
387 char *p = strstr(str, ":");
393 strip_blank_front(&p);
398 static int exists_and_not_file(const char *filename)
402 if (lstat(filename, &sb) == -1)
405 if (S_ISREG(sb.st_mode))
411 void td_fill_rand_seeds(struct thread_data *td)
413 os_random_seed(td->rand_seeds[0], &td->bsrange_state);
414 os_random_seed(td->rand_seeds[1], &td->verify_state);
415 os_random_seed(td->rand_seeds[2], &td->rwmix_state);
417 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
418 os_random_seed(td->rand_seeds[3], &td->next_file_state);
420 os_random_seed(td->rand_seeds[5], &td->file_size_state);
425 if (td->o.rand_repeatable)
426 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
428 os_random_seed(td->rand_seeds[4], &td->random_state);
432 * Initialize the various random states we need (random io, block size ranges,
433 * read/write mix, etc).
435 static int init_random_state(struct thread_data *td)
439 fd = open("/dev/urandom", O_RDONLY);
441 td_verror(td, errno, "open");
445 if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
446 (int) sizeof(td->rand_seeds)) {
447 td_verror(td, EIO, "read");
453 td_fill_rand_seeds(td);
458 * Adds a job to the list of things todo. Sanitizes the various options
459 * to make sure we don't have conflicts, and initializes various
462 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
464 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
465 "randread", "randwrite", "randrw" };
468 char fname[PATH_MAX];
469 int numjobs, file_alloced;
472 * the def_thread is just for options, it's not a real job
474 if (td == &def_thread)
478 * if we are just dumping the output command line, don't add the job
485 engine = get_engine_name(td->o.ioengine);
486 td->io_ops = load_ioengine(td, engine);
488 log_err("fio: failed to load engine %s\n", engine);
492 if (td->o.use_thread)
498 td->io_ops->flags |= FIO_RAWIO;
501 if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
504 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
505 add_file(td, jobname);
507 for (i = 0; i < td->o.nr_files; i++) {
508 sprintf(fname, "%s.%d.%d", jobname,
509 td->thread_number, i);
515 if (fixup_options(td))
518 if (td->io_ops->flags & FIO_DISKLESSIO) {
521 for_each_file(td, f, i)
522 f->real_file_size = -1ULL;
525 td->mutex = fio_mutex_init(0);
527 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
528 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
529 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
530 td->ddir_nr = td->o.ddir_nr;
532 if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
533 && prev_group_jobs) {
538 td->groupid = groupid;
541 if (init_random_state(td))
547 if (td->o.write_lat_log) {
548 setup_log(&td->ts.slat_log);
549 setup_log(&td->ts.clat_log);
551 if (td->o.write_bw_log)
552 setup_log(&td->ts.bw_log);
555 td->o.name = strdup(jobname);
559 if (!strcmp(td->io_ops->name, "cpuio")) {
560 log_info("%s: ioengine=cpu, cpuload=%u,"
561 " cpucycle=%u\n", td->o.name,
565 char *c1, *c2, *c3, *c4;
567 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
568 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
569 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
570 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
572 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
573 " ioengine=%s, iodepth=%u\n",
574 td->o.name, td->groupid,
575 ddir_str[td->o.td_ddir],
585 } else if (job_add_num == 1)
590 * recurse add identical jobs, clear numjobs and stonewall options
591 * as they don't apply to sub-jobs
593 numjobs = td->o.numjobs;
595 struct thread_data *td_new = get_new_job(0, td);
600 td_new->o.numjobs = 1;
601 td_new->o.stonewall = 0;
602 td_new->o.new_group = 0;
605 td_new->o.filename = NULL;
606 td_new->files_index = 0;
607 td_new->files_size = 0;
608 td_new->files = NULL;
611 job_add_num = numjobs - 1;
613 if (add_job(td_new, jobname, job_add_num))
623 static int skip_this_section(const char *name)
627 if (!strncmp(name, "global", 6))
630 return strcmp(job_section, name);
633 static int is_empty_or_comment(char *line)
637 for (i = 0; i < strlen(line); i++) {
642 if (!isspace(line[i]) && !iscntrl(line[i]))
650 * This is our [ini] type file parser.
652 static int parse_jobs_ini(char *file, int stonewall_flag)
655 struct thread_data *td;
659 int ret = 0, stonewall;
664 int i, alloc_opts, num_opts;
666 if (!strcmp(file, "-"))
669 f = fopen(file, "r");
672 perror("fopen job file");
676 string = malloc(4096);
679 * it's really 256 + small bit, 280 should suffice
682 memset(name, 0, 280);
685 opts = malloc(sizeof(char *) * alloc_opts);
688 stonewall = stonewall_flag;
691 * if skip_fgets is set, we already have loaded a line we
695 p = fgets(string, 4095, f);
701 strip_blank_front(&p);
704 if (is_empty_or_comment(p))
706 if (sscanf(p, "[%255s]", name) != 1) {
709 log_err("fio: option <%s> outside of [] job section\n",
714 if (skip_this_section(name)) {
720 global = !strncmp(name, "global", 6);
722 name[strlen(name) - 1] = '\0';
728 log_info("--name=%s ", name);
732 td = get_new_job(global, &def_thread);
739 * Seperate multiple job files by a stonewall
741 if (!global && stonewall) {
742 td->o.stonewall = stonewall;
747 memset(opts, 0, alloc_opts * sizeof(char *));
749 while ((p = fgets(string, 4096, f)) != NULL) {
750 if (is_empty_or_comment(p))
753 strip_blank_front(&p);
756 * new section, break out and make sure we don't
757 * fgets() a new line at the top.
766 if (num_opts == alloc_opts) {
769 alloc_opts * sizeof(char *));
772 opts[num_opts] = strdup(p);
776 ret = fio_options_parse(td, opts, num_opts);
779 for (i = 0; i < num_opts; i++)
780 log_info("--%s ", opts[i]);
782 ret = add_job(td, name, 0);
784 log_err("fio: job %s dropped\n", name);
788 for (i = 0; i < num_opts; i++)
796 for (i = 0; i < num_opts; i++)
807 static int fill_def_thread(void)
809 memset(&def_thread, 0, sizeof(def_thread));
811 fio_getaffinity(getpid(), &def_thread.o.cpumask);
814 * fill default options
816 fio_fill_default_options(&def_thread);
818 def_thread.o.timeout = def_timeout;
819 def_thread.o.write_bw_log = write_bw_log;
820 def_thread.o.write_lat_log = write_lat_log;
825 static void free_shm(void)
827 struct shmid_ds sbuf;
834 fio_debug_jobp = NULL;
836 shmctl(shm_id, IPC_RMID, &sbuf);
843 * The thread area is shared between the main process and the job
844 * threads/processes. So setup a shared memory segment that will hold
845 * all the job info. We use the end of the region for keeping track of
846 * open files across jobs, for file sharing.
848 static int setup_thread_area(void)
853 * 1024 is too much on some machines, scale max_jobs if
854 * we get a failure that looks like too large a shm segment
857 size_t size = max_jobs * sizeof(struct thread_data);
859 size += file_hash_size;
860 size += sizeof(unsigned int);
862 shm_id = shmget(0, size, IPC_CREAT | 0600);
865 if (errno != EINVAL) {
876 threads = shmat(shm_id, NULL, 0);
877 if (threads == (void *) -1) {
882 memset(threads, 0, max_jobs * sizeof(struct thread_data));
883 hash = (void *) threads + max_jobs * sizeof(struct thread_data);
884 fio_debug_jobp = (void *) hash + file_hash_size;
885 *fio_debug_jobp = -1;
886 file_hash_init(hash);
891 static void usage(const char *name)
893 printf("%s\n", fio_version_string);
894 printf("%s [options] [job options] <job file(s)>\n", name);
895 printf("\t--debug=options\tEnable debug logging\n");
896 printf("\t--output\tWrite output to file\n");
897 printf("\t--timeout\tRuntime in seconds\n");
898 printf("\t--latency-log\tGenerate per-job latency logs\n");
899 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
900 printf("\t--minimal\tMinimal (terse) output\n");
901 printf("\t--version\tPrint version info and exit\n");
902 printf("\t--help\t\tPrint this page\n");
903 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
905 printf("\t--showcmd\tTurn a job file into command line options\n");
906 printf("\t--eta=when\tWhen ETA estimate should be printed\n");
907 printf("\t \tMay be \"always\", \"never\" or \"auto\"\n");
908 printf("\t--readonly\tTurn on safety read-only checks, preventing"
910 printf("\t--section=name\tOnly run specified section in job file\n");
911 printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
916 struct debug_level debug_levels[] = {
917 { .name = "process", .shift = FD_PROCESS, },
918 { .name = "file", .shift = FD_FILE, },
919 { .name = "io", .shift = FD_IO, },
920 { .name = "mem", .shift = FD_MEM, },
921 { .name = "blktrace", .shift = FD_BLKTRACE },
922 { .name = "verify", .shift = FD_VERIFY },
923 { .name = "random", .shift = FD_RANDOM },
924 { .name = "parse", .shift = FD_PARSE },
925 { .name = "diskutil", .shift = FD_DISKUTIL },
926 { .name = "job", .shift = FD_JOB },
927 { .name = "mutex", .shift = FD_MUTEX },
931 static int set_debug(const char *string)
933 struct debug_level *dl;
934 char *p = (char *) string;
938 if (!strcmp(string, "?") || !strcmp(string, "help")) {
941 log_info("fio: dumping debug options:");
942 for (i = 0; debug_levels[i].name; i++) {
943 dl = &debug_levels[i];
944 log_info("%s,", dl->name);
950 while ((opt = strsep(&p, ",")) != NULL) {
953 if (!strncmp(opt, "all", 3)) {
954 log_info("fio: set all debug options\n");
959 for (i = 0; debug_levels[i].name; i++) {
960 dl = &debug_levels[i];
961 found = !strncmp(opt, dl->name, strlen(dl->name));
965 if (dl->shift == FD_JOB) {
966 opt = strchr(opt, ':');
968 log_err("fio: missing job number\n");
972 fio_debug_jobno = atoi(opt);
973 log_info("fio: set debug jobno %d\n",
976 log_info("fio: set debug option %s\n", opt);
977 fio_debug |= (1UL << dl->shift);
983 log_err("fio: debug mask %s not found\n", opt);
988 static int set_debug(const char *string)
990 log_err("fio: debug tracing not included in build\n");
995 static int parse_cmd_line(int argc, char *argv[])
997 struct thread_data *td = NULL;
998 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1000 while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1003 smalloc_pool_size = atoi(optarg);
1006 def_timeout = atoi(optarg);
1015 f_out = fopen(optarg, "w+");
1017 perror("fopen output");
1029 exit(fio_show_option_help(optarg));
1037 printf("%s\n", fio_version_string);
1040 if (!strcmp("always", optarg))
1041 eta_print = FIO_ETA_ALWAYS;
1042 else if (!strcmp("never", optarg))
1043 eta_print = FIO_ETA_NEVER;
1046 if (set_debug(optarg))
1050 if (!strcmp(optarg, "global")) {
1051 log_err("fio: can't use global as only "
1059 job_section = strdup(optarg);
1061 case FIO_GETOPT_JOB: {
1062 const char *opt = l_opts[lidx].name;
1065 if (!strncmp(opt, "name", 4) && td) {
1066 ret = add_job(td, td->o.name ?: "fio", 0);
1074 int is_section = !strncmp(opt, "name", 4);
1077 if (!is_section || !strncmp(val, "global", 6))
1080 if (is_section && skip_this_section(val))
1083 td = get_new_job(global, &def_thread);
1088 ret = fio_cmd_option_parse(td, opt, val);
1103 ret = add_job(td, td->o.name ?: "fio", 0);
1108 while (optind < argc) {
1110 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1111 ini_file[ini_idx - 1] = strdup(argv[optind]);
1118 int parse_options(int argc, char *argv[])
1125 fio_options_dup_and_init(l_opts);
1127 if (setup_thread_area())
1129 if (fill_def_thread())
1132 job_files = parse_cmd_line(argc, argv);
1134 for (i = 0; i < job_files; i++) {
1135 if (fill_def_thread())
1137 if (parse_jobs_ini(ini_file[i], i))
1143 options_mem_free(&def_thread);
1145 if (!thread_number) {
1149 log_err("No jobs defined(s)\n");
1154 if (def_thread.o.gtod_offload) {
1156 fio_gtod_offload = 1;
1157 fio_gtod_cpu = def_thread.o.gtod_cpu;