2 * This file contains job initialization and setup functions.
14 #include <sys/types.h>
24 static char fio_version_string[] = "fio 1.37";
26 #define FIO_RANDSEED (0xb1899bedUL)
28 static char **ini_file;
29 static int max_jobs = MAX_JOBS;
30 static int dump_cmdline;
32 static struct thread_data def_thread;
33 struct thread_data *threads = NULL;
35 int exitall_on_terminate = 0;
38 unsigned long long mlock_size = 0;
41 char *job_section = NULL;
42 char *exec_profile = NULL;
47 static int def_timeout;
48 static int write_lat_log;
50 static int prev_group_jobs;
52 unsigned long fio_debug = 0;
53 unsigned int fio_debug_jobno = -1;
54 unsigned int *fio_debug_jobp = NULL;
57 * Command line options. These will contain the above, plus a few
58 * extra that only pertain to fio itself and not jobs.
60 static struct option l_opts[FIO_NR_OPTIONS] = {
63 .has_arg = required_argument,
68 .has_arg = required_argument,
72 .name = "latency-log",
73 .has_arg = required_argument,
77 .name = "bandwidth-log",
78 .has_arg = required_argument,
83 .has_arg = optional_argument,
88 .has_arg = no_argument,
93 .has_arg = no_argument,
98 .has_arg = optional_argument,
103 .has_arg = no_argument,
108 .has_arg = no_argument,
113 .has_arg = required_argument,
118 .has_arg = required_argument,
123 .has_arg = required_argument,
127 .name = "alloc-size",
128 .has_arg = required_argument,
133 .has_arg = required_argument,
152 * Return a free job structure.
154 static struct thread_data *get_new_job(int global, struct thread_data *parent)
156 struct thread_data *td;
160 if (thread_number >= max_jobs)
163 td = &threads[thread_number++];
166 td->o.uid = td->o.gid = -1U;
168 dup_files(td, parent);
169 options_mem_dupe(td);
171 profile_add_hooks(td);
173 td->thread_number = thread_number;
177 static void put_job(struct thread_data *td)
179 if (td == &def_thread)
183 log_info("fio: %s\n", td->verror);
185 memset(&threads[td->thread_number - 1], 0, sizeof(*td));
189 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
191 unsigned int bs = td->o.min_bs[ddir];
192 unsigned long long bytes_per_sec;
194 if (td->o.rate[ddir])
195 bytes_per_sec = td->o.rate[ddir];
197 bytes_per_sec = td->o.rate_iops[ddir] * bs;
199 if (!bytes_per_sec) {
200 log_err("rate lower than supported\n");
204 td->rate_nsec_cycle[ddir] = 1000000000ULL / bytes_per_sec;
205 td->rate_pending_usleep[ddir] = 0;
209 static int setup_rate(struct thread_data *td)
213 if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
214 ret = __setup_rate(td, DDIR_READ);
215 if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
216 ret |= __setup_rate(td, DDIR_WRITE);
221 static int fixed_block_size(struct thread_options *o)
223 return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
224 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
225 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
229 * Lazy way of fixing up options that depend on each other. We could also
230 * define option callback handlers, but this is easier.
232 static int fixup_options(struct thread_data *td)
234 struct thread_options *o = &td->o;
236 #ifndef FIO_HAVE_PSHARED_MUTEX
237 if (!td->o.use_thread) {
238 log_info("fio: this platform does not support process shared"
239 " mutexes, forcing use of threads. Use the 'thread'"
240 " option to get rid of this warning.\n");
241 td->o.use_thread = 1;
245 if (o->write_iolog_file && o->read_iolog_file) {
246 log_err("fio: read iolog overrides write_iolog\n");
247 free(o->write_iolog_file);
248 o->write_iolog_file = NULL;
252 * only really works for sequential io for now, and with 1 file
254 if (o->zone_size && td_random(td) && o->open_files == 1)
258 * Reads can do overwrites, we always need to pre-create the file
260 if (td_read(td) || td_rw(td))
263 if (!o->min_bs[DDIR_READ])
264 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
265 if (!o->max_bs[DDIR_READ])
266 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
267 if (!o->min_bs[DDIR_WRITE])
268 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
269 if (!o->max_bs[DDIR_WRITE])
270 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
272 o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
275 * For random IO, allow blockalign offset other than min_bs.
277 if (!o->ba[DDIR_READ] || !td_random(td))
278 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
279 if (!o->ba[DDIR_WRITE] || !td_random(td))
280 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
282 if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
283 o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
284 !td->o.norandommap) {
285 log_err("fio: Any use of blockalign= turns off randommap\n");
286 td->o.norandommap = 1;
289 if (!o->file_size_high)
290 o->file_size_high = o->file_size_low;
292 if (o->norandommap && o->verify != VERIFY_NONE
293 && !fixed_block_size(o)) {
294 log_err("fio: norandommap given for variable block sizes, "
295 "verify disabled\n");
296 o->verify = VERIFY_NONE;
298 if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
299 log_err("fio: bs_unaligned may not work with raw io\n");
302 * thinktime_spin must be less than thinktime
304 if (o->thinktime_spin > o->thinktime)
305 o->thinktime_spin = o->thinktime;
308 * The low water mark cannot be bigger than the iodepth
310 if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
312 * syslet work around - if the workload is sequential,
313 * we want to let the queue drain all the way down to
314 * avoid seeking between async threads
316 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
319 o->iodepth_low = o->iodepth;
323 * If batch number isn't set, default to the same as iodepth
325 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
326 o->iodepth_batch = o->iodepth;
328 if (o->nr_files > td->files_index)
329 o->nr_files = td->files_index;
331 if (o->open_files > o->nr_files || !o->open_files)
332 o->open_files = o->nr_files;
334 if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
335 ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
336 o->rate_iops_min[1]))) {
337 log_err("fio: rate and rate_iops are mutually exclusive\n");
340 if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
341 (o->rate_iops[0] < o->rate_iops_min[0]) ||
342 (o->rate_iops[1] < o->rate_iops_min[1])) {
343 log_err("fio: minimum rate exceeds rate\n");
347 if (!o->timeout && o->time_based) {
348 log_err("fio: time_based requires a runtime/timeout setting\n");
352 if (o->fill_device && !o->size)
355 if (td_rw(td) && td->o.verify != VERIFY_NONE)
356 log_info("fio: mixed read/write workload with verify. May not "
357 "work as expected, unless you pre-populated the file\n");
359 if (td->o.verify != VERIFY_NONE)
360 td->o.refill_buffers = 1;
362 if (td->o.pre_read) {
363 td->o.invalidate_cache = 0;
364 if (td->io_ops->flags & FIO_PIPEIO)
365 log_info("fio: cannot pre-read files with an IO engine"
366 " that isn't seekable. Pre-read disabled.\n");
369 #ifndef FIO_HAVE_FDATASYNC
370 if (td->o.fdatasync_blocks) {
371 log_info("fio: this platform does not support fdatasync()"
372 " falling back to using fsync(). Use the 'fsync'"
373 " option instead of 'fdatasync' to get rid of"
375 td->o.fsync_blocks = td->o.fdatasync_blocks;
376 td->o.fdatasync_blocks = 0;
384 * This function leaks the buffer
386 static char *to_kmg(unsigned int val)
388 char *buf = malloc(32);
389 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
400 snprintf(buf, 31, "%u%c", val, *p);
404 /* External engines are specified by "external:name.o") */
405 static const char *get_engine_name(const char *str)
407 char *p = strstr(str, ":");
413 strip_blank_front(&p);
418 static int exists_and_not_file(const char *filename)
422 if (lstat(filename, &sb) == -1)
425 if (S_ISREG(sb.st_mode))
431 void td_fill_rand_seeds(struct thread_data *td)
433 os_random_seed(td->rand_seeds[0], &td->bsrange_state);
434 os_random_seed(td->rand_seeds[1], &td->verify_state);
435 os_random_seed(td->rand_seeds[2], &td->rwmix_state);
437 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
438 os_random_seed(td->rand_seeds[3], &td->next_file_state);
440 os_random_seed(td->rand_seeds[5], &td->file_size_state);
445 if (td->o.rand_repeatable)
446 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
448 os_random_seed(td->rand_seeds[4], &td->random_state);
452 * Initialize the various random states we need (random io, block size ranges,
453 * read/write mix, etc).
455 static int init_random_state(struct thread_data *td)
459 fd = open("/dev/urandom", O_RDONLY);
461 td_verror(td, errno, "open");
465 if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
466 (int) sizeof(td->rand_seeds)) {
467 td_verror(td, EIO, "read");
473 td_fill_rand_seeds(td);
478 * Adds a job to the list of things todo. Sanitizes the various options
479 * to make sure we don't have conflicts, and initializes various
482 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
484 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
485 "randread", "randwrite", "randrw" };
488 char fname[PATH_MAX];
489 int numjobs, file_alloced;
492 * the def_thread is just for options, it's not a real job
494 if (td == &def_thread)
498 * if we are just dumping the output command line, don't add the job
505 engine = get_engine_name(td->o.ioengine);
506 td->io_ops = load_ioengine(td, engine);
508 log_err("fio: failed to load engine %s\n", engine);
512 if (td->o.use_thread)
518 td->io_ops->flags |= FIO_RAWIO;
521 if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
524 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
525 add_file(td, jobname);
527 for (i = 0; i < td->o.nr_files; i++) {
528 sprintf(fname, "%s.%d.%d", jobname,
529 td->thread_number, i);
535 if (fixup_options(td))
538 if (td->io_ops->flags & FIO_DISKLESSIO) {
541 for_each_file(td, f, i)
542 f->real_file_size = -1ULL;
545 td->mutex = fio_mutex_init(0);
547 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
548 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
549 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
550 td->ddir_nr = td->o.ddir_nr;
552 if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
557 td->groupid = groupid;
560 if (init_random_state(td))
566 if (td->o.write_lat_log) {
567 setup_log(&td->ts.slat_log);
568 setup_log(&td->ts.clat_log);
570 if (td->o.write_bw_log)
571 setup_log(&td->ts.bw_log);
574 td->o.name = strdup(jobname);
578 if (!strcmp(td->io_ops->name, "cpuio")) {
579 log_info("%s: ioengine=cpu, cpuload=%u,"
580 " cpucycle=%u\n", td->o.name,
584 char *c1, *c2, *c3, *c4;
586 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
587 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
588 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
589 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
591 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
592 " ioengine=%s, iodepth=%u\n",
593 td->o.name, td->groupid,
594 ddir_str[td->o.td_ddir],
604 } else if (job_add_num == 1)
609 * recurse add identical jobs, clear numjobs and stonewall options
610 * as they don't apply to sub-jobs
612 numjobs = td->o.numjobs;
614 struct thread_data *td_new = get_new_job(0, td);
619 td_new->o.numjobs = 1;
620 td_new->o.stonewall = 0;
621 td_new->o.new_group = 0;
624 td_new->o.filename = NULL;
625 td_new->files_index = 0;
626 td_new->files_size = 0;
627 td_new->files = NULL;
630 job_add_num = numjobs - 1;
632 if (add_job(td_new, jobname, job_add_num))
643 * Parse as if 'o' was a command line
645 void add_job_opts(const char **o)
647 struct thread_data *td, *td_parent;
648 int i, in_global = 1;
652 td_parent = td = NULL;
654 if (!strncmp(o[i], "name", 4)) {
657 add_job(td, jobname, 0);
659 sprintf(jobname, "%s", o[i] + 5);
661 if (in_global && !td_parent)
662 td_parent = get_new_job(1, &def_thread);
663 else if (!in_global && !td) {
665 td_parent = &def_thread;
666 td = get_new_job(0, td_parent);
669 fio_options_parse(td_parent, (char **) &o[i], 1);
671 fio_options_parse(td, (char **) &o[i], 1);
676 add_job(td, jobname, 0);
679 static int skip_this_section(const char *name)
683 if (!strncmp(name, "global", 6))
686 return strcmp(job_section, name);
689 static int is_empty_or_comment(char *line)
693 for (i = 0; i < strlen(line); i++) {
698 if (!isspace(line[i]) && !iscntrl(line[i]))
706 * This is our [ini] type file parser.
708 static int parse_jobs_ini(char *file, int stonewall_flag)
711 struct thread_data *td;
715 int ret = 0, stonewall;
720 int i, alloc_opts, num_opts;
722 if (!strcmp(file, "-"))
725 f = fopen(file, "r");
728 perror("fopen job file");
732 string = malloc(4096);
735 * it's really 256 + small bit, 280 should suffice
738 memset(name, 0, 280);
741 opts = malloc(sizeof(char *) * alloc_opts);
744 stonewall = stonewall_flag;
747 * if skip_fgets is set, we already have loaded a line we
751 p = fgets(string, 4095, f);
757 strip_blank_front(&p);
760 if (is_empty_or_comment(p))
762 if (sscanf(p, "[%255s]", name) != 1) {
765 log_err("fio: option <%s> outside of [] job section\n",
770 if (skip_this_section(name)) {
776 global = !strncmp(name, "global", 6);
778 name[strlen(name) - 1] = '\0';
784 log_info("--name=%s ", name);
788 td = get_new_job(global, &def_thread);
795 * Seperate multiple job files by a stonewall
797 if (!global && stonewall) {
798 td->o.stonewall = stonewall;
803 memset(opts, 0, alloc_opts * sizeof(char *));
805 while ((p = fgets(string, 4096, f)) != NULL) {
806 if (is_empty_or_comment(p))
809 strip_blank_front(&p);
812 * new section, break out and make sure we don't
813 * fgets() a new line at the top.
822 if (num_opts == alloc_opts) {
825 alloc_opts * sizeof(char *));
828 opts[num_opts] = strdup(p);
832 ret = fio_options_parse(td, opts, num_opts);
835 for (i = 0; i < num_opts; i++)
836 log_info("--%s ", opts[i]);
838 ret = add_job(td, name, 0);
840 log_err("fio: job %s dropped\n", name);
844 for (i = 0; i < num_opts; i++)
852 for (i = 0; i < num_opts; i++)
863 static int fill_def_thread(void)
865 memset(&def_thread, 0, sizeof(def_thread));
867 fio_getaffinity(getpid(), &def_thread.o.cpumask);
870 * fill default options
872 fio_fill_default_options(&def_thread);
874 def_thread.o.timeout = def_timeout;
875 def_thread.o.write_bw_log = write_bw_log;
876 def_thread.o.write_lat_log = write_lat_log;
881 static void free_shm(void)
883 struct shmid_ds sbuf;
890 fio_debug_jobp = NULL;
892 shmctl(shm_id, IPC_RMID, &sbuf);
899 * The thread area is shared between the main process and the job
900 * threads/processes. So setup a shared memory segment that will hold
901 * all the job info. We use the end of the region for keeping track of
902 * open files across jobs, for file sharing.
904 static int setup_thread_area(void)
909 * 1024 is too much on some machines, scale max_jobs if
910 * we get a failure that looks like too large a shm segment
913 size_t size = max_jobs * sizeof(struct thread_data);
915 size += file_hash_size;
916 size += sizeof(unsigned int);
918 shm_id = shmget(0, size, IPC_CREAT | 0600);
921 if (errno != EINVAL) {
932 threads = shmat(shm_id, NULL, 0);
933 if (threads == (void *) -1) {
938 memset(threads, 0, max_jobs * sizeof(struct thread_data));
939 hash = (void *) threads + max_jobs * sizeof(struct thread_data);
940 fio_debug_jobp = (void *) hash + file_hash_size;
941 *fio_debug_jobp = -1;
942 file_hash_init(hash);
947 static void usage(const char *name)
949 printf("%s\n", fio_version_string);
950 printf("%s [options] [job options] <job file(s)>\n", name);
951 printf("\t--debug=options\tEnable debug logging\n");
952 printf("\t--output\tWrite output to file\n");
953 printf("\t--timeout\tRuntime in seconds\n");
954 printf("\t--latency-log\tGenerate per-job latency logs\n");
955 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
956 printf("\t--minimal\tMinimal (terse) output\n");
957 printf("\t--version\tPrint version info and exit\n");
958 printf("\t--help\t\tPrint this page\n");
959 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
961 printf("\t--showcmd\tTurn a job file into command line options\n");
962 printf("\t--eta=when\tWhen ETA estimate should be printed\n");
963 printf("\t \tMay be \"always\", \"never\" or \"auto\"\n");
964 printf("\t--readonly\tTurn on safety read-only checks, preventing"
966 printf("\t--section=name\tOnly run specified section in job file\n");
967 printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
969 printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>\n");
973 struct debug_level debug_levels[] = {
974 { .name = "process", .shift = FD_PROCESS, },
975 { .name = "file", .shift = FD_FILE, },
976 { .name = "io", .shift = FD_IO, },
977 { .name = "mem", .shift = FD_MEM, },
978 { .name = "blktrace", .shift = FD_BLKTRACE },
979 { .name = "verify", .shift = FD_VERIFY },
980 { .name = "random", .shift = FD_RANDOM },
981 { .name = "parse", .shift = FD_PARSE },
982 { .name = "diskutil", .shift = FD_DISKUTIL },
983 { .name = "job", .shift = FD_JOB },
984 { .name = "mutex", .shift = FD_MUTEX },
985 { .name = "profile", .shift = FD_PROFILE },
989 static int set_debug(const char *string)
991 struct debug_level *dl;
992 char *p = (char *) string;
996 if (!strcmp(string, "?") || !strcmp(string, "help")) {
999 log_info("fio: dumping debug options:");
1000 for (i = 0; debug_levels[i].name; i++) {
1001 dl = &debug_levels[i];
1002 log_info("%s,", dl->name);
1008 while ((opt = strsep(&p, ",")) != NULL) {
1011 if (!strncmp(opt, "all", 3)) {
1012 log_info("fio: set all debug options\n");
1017 for (i = 0; debug_levels[i].name; i++) {
1018 dl = &debug_levels[i];
1019 found = !strncmp(opt, dl->name, strlen(dl->name));
1023 if (dl->shift == FD_JOB) {
1024 opt = strchr(opt, ':');
1026 log_err("fio: missing job number\n");
1030 fio_debug_jobno = atoi(opt);
1031 log_info("fio: set debug jobno %d\n",
1034 log_info("fio: set debug option %s\n", opt);
1035 fio_debug |= (1UL << dl->shift);
1041 log_err("fio: debug mask %s not found\n", opt);
1046 static int set_debug(const char *string)
1048 log_err("fio: debug tracing not included in build\n");
1053 static int parse_cmd_line(int argc, char *argv[])
1055 struct thread_data *td = NULL;
1056 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1058 while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1061 smalloc_pool_size = atoi(optarg);
1064 def_timeout = atoi(optarg);
1073 f_out = fopen(optarg, "w+");
1075 perror("fopen output");
1087 exit(fio_show_option_help(optarg));
1095 printf("%s\n", fio_version_string);
1098 if (!strcmp("always", optarg))
1099 eta_print = FIO_ETA_ALWAYS;
1100 else if (!strcmp("never", optarg))
1101 eta_print = FIO_ETA_NEVER;
1104 if (set_debug(optarg))
1108 if (!strcmp(optarg, "global")) {
1109 log_err("fio: can't use global as only "
1117 job_section = strdup(optarg);
1120 exec_profile = strdup(optarg);
1122 case FIO_GETOPT_JOB: {
1123 const char *opt = l_opts[lidx].name;
1126 if (!strncmp(opt, "name", 4) && td) {
1127 ret = add_job(td, td->o.name ?: "fio", 0);
1135 int is_section = !strncmp(opt, "name", 4);
1138 if (!is_section || !strncmp(val, "global", 6))
1141 if (is_section && skip_this_section(val))
1144 td = get_new_job(global, &def_thread);
1149 ret = fio_cmd_option_parse(td, opt, val);
1164 ret = add_job(td, td->o.name ?: "fio", 0);
1169 while (optind < argc) {
1171 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1172 ini_file[ini_idx - 1] = strdup(argv[optind]);
1179 int parse_options(int argc, char *argv[])
1186 fio_options_dup_and_init(l_opts);
1188 if (setup_thread_area())
1190 if (fill_def_thread())
1193 job_files = parse_cmd_line(argc, argv);
1195 for (i = 0; i < job_files; i++) {
1196 if (fill_def_thread())
1198 if (parse_jobs_ini(ini_file[i], i))
1204 options_mem_free(&def_thread);
1206 if (!thread_number) {
1212 log_err("No jobs defined(s)\n\n");
1217 if (def_thread.o.gtod_offload) {
1219 fio_gtod_offload = 1;
1220 fio_gtod_cpu = def_thread.o.gtod_cpu;