2 * This file contains job initialization and setup functions.
14 #include <sys/types.h>
23 static char fio_version_string[] = "fio 1.35";
25 #define FIO_RANDSEED (0xb1899bedUL)
27 static char **ini_file;
28 static int max_jobs = MAX_JOBS;
29 static int dump_cmdline;
31 static struct thread_data def_thread;
32 struct thread_data *threads = NULL;
34 int exitall_on_terminate = 0;
37 unsigned long long mlock_size = 0;
40 char *job_section = NULL;
45 static int def_timeout;
46 static int write_lat_log;
48 static int prev_group_jobs;
50 unsigned long fio_debug = 0;
51 unsigned int fio_debug_jobno = -1;
52 unsigned int *fio_debug_jobp = NULL;
55 * Command line options. These will contain the above, plus a few
56 * extra that only pertain to fio itself and not jobs.
58 static struct option l_opts[FIO_NR_OPTIONS] = {
61 .has_arg = required_argument,
66 .has_arg = required_argument,
70 .name = "latency-log",
71 .has_arg = required_argument,
75 .name = "bandwidth-log",
76 .has_arg = required_argument,
81 .has_arg = optional_argument,
86 .has_arg = no_argument,
91 .has_arg = no_argument,
96 .has_arg = optional_argument,
101 .has_arg = no_argument,
106 .has_arg = no_argument,
111 .has_arg = required_argument,
116 .has_arg = required_argument,
121 .has_arg = required_argument,
125 .name = "alloc-size",
126 .has_arg = required_argument,
131 .has_arg = required_argument,
139 static const char *tiobench_opts[] = {
140 "buffered=0", "size=4*1024*$mb_memory", "bs=4k", "timeout=600",
141 "numjobs=4", "group_reporting", "thread", "overwrite=1",
142 "filename=.fio.tio.1:.fio.tio.2:.fio.tio.3:.fio.tio.4",
143 "name=seqwrite", "rw=write", "end_fsync=1",
144 "name=randwrite", "stonewall", "rw=randwrite", "end_fsync=1",
145 "name=seqread", "stonewall", "rw=read",
146 "name=randread", "stonewall", "rw=randread", NULL,
149 static const char **fio_prof_strings[PROFILE_END] = {
154 static const char *profiles[PROFILE_END] = {
170 * Return a free job structure.
172 static struct thread_data *get_new_job(int global, struct thread_data *parent)
174 struct thread_data *td;
178 if (thread_number >= max_jobs)
181 td = &threads[thread_number++];
184 dup_files(td, parent);
185 options_mem_dupe(td);
187 td->thread_number = thread_number;
191 static void put_job(struct thread_data *td)
193 if (td == &def_thread)
197 log_info("fio: %s\n", td->verror);
199 memset(&threads[td->thread_number - 1], 0, sizeof(*td));
203 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
205 unsigned int bs = td->o.min_bs[ddir];
206 unsigned long long rate;
207 unsigned long ios_per_msec;
209 if (td->o.rate[ddir]) {
210 rate = td->o.rate[ddir];
211 ios_per_msec = (rate * 1000LL) / bs;
213 ios_per_msec = td->o.rate_iops[ddir] * 1000UL;
216 log_err("rate lower than supported\n");
220 td->rate_usec_cycle[ddir] = 1000000000ULL / ios_per_msec;
221 td->rate_pending_usleep[ddir] = 0;
225 static int setup_rate(struct thread_data *td)
229 if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
230 ret = __setup_rate(td, DDIR_READ);
231 if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
232 ret |= __setup_rate(td, DDIR_WRITE);
237 static int fixed_block_size(struct thread_options *o)
239 return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
240 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
241 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
245 * Lazy way of fixing up options that depend on each other. We could also
246 * define option callback handlers, but this is easier.
248 static int fixup_options(struct thread_data *td)
250 struct thread_options *o = &td->o;
252 #ifndef FIO_HAVE_PSHARED_MUTEX
253 if (!td->o.use_thread) {
254 log_info("fio: this platform does not support process shared"
255 " mutexes, forcing use of threads. Use the 'thread'"
256 " option to get rid of this warning.\n");
257 td->o.use_thread = 1;
261 if (o->write_iolog_file && o->read_iolog_file) {
262 log_err("fio: read iolog overrides write_iolog\n");
263 free(o->write_iolog_file);
264 o->write_iolog_file = NULL;
268 * only really works for sequential io for now, and with 1 file
270 if (o->zone_size && td_random(td) && o->open_files == 1)
274 * Reads can do overwrites, we always need to pre-create the file
276 if (td_read(td) || td_rw(td))
279 if (!o->min_bs[DDIR_READ])
280 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
281 if (!o->max_bs[DDIR_READ])
282 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
283 if (!o->min_bs[DDIR_WRITE])
284 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
285 if (!o->max_bs[DDIR_WRITE])
286 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
288 o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
291 * For random IO, allow blockalign offset other than min_bs.
293 if (!o->ba[DDIR_READ] || !td_random(td))
294 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
295 if (!o->ba[DDIR_WRITE] || !td_random(td))
296 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
298 if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
299 o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
300 !td->o.norandommap) {
301 log_err("fio: Any use of blockalign= turns off randommap\n");
302 td->o.norandommap = 1;
305 if (!o->file_size_high)
306 o->file_size_high = o->file_size_low;
308 if (o->norandommap && o->verify != VERIFY_NONE
309 && !fixed_block_size(o)) {
310 log_err("fio: norandommap given for variable block sizes, "
311 "verify disabled\n");
312 o->verify = VERIFY_NONE;
314 if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
315 log_err("fio: bs_unaligned may not work with raw io\n");
318 * thinktime_spin must be less than thinktime
320 if (o->thinktime_spin > o->thinktime)
321 o->thinktime_spin = o->thinktime;
324 * The low water mark cannot be bigger than the iodepth
326 if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
328 * syslet work around - if the workload is sequential,
329 * we want to let the queue drain all the way down to
330 * avoid seeking between async threads
332 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
335 o->iodepth_low = o->iodepth;
339 * If batch number isn't set, default to the same as iodepth
341 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
342 o->iodepth_batch = o->iodepth;
344 if (o->nr_files > td->files_index)
345 o->nr_files = td->files_index;
347 if (o->open_files > o->nr_files || !o->open_files)
348 o->open_files = o->nr_files;
350 if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
351 ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
352 o->rate_iops_min[1]))) {
353 log_err("fio: rate and rate_iops are mutually exclusive\n");
356 if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
357 (o->rate_iops[0] < o->rate_iops_min[0]) ||
358 (o->rate_iops[1] < o->rate_iops_min[1])) {
359 log_err("fio: minimum rate exceeds rate\n");
363 if (!o->timeout && o->time_based) {
364 log_err("fio: time_based requires a runtime/timeout setting\n");
368 if (o->fill_device && !o->size)
371 if (td_rw(td) && td->o.verify != VERIFY_NONE)
372 log_info("fio: mixed read/write workload with verify. May not "
373 "work as expected, unless you pre-populated the file\n");
375 if (td->o.verify != VERIFY_NONE)
376 td->o.refill_buffers = 1;
378 if (td->o.pre_read) {
379 td->o.invalidate_cache = 0;
380 if (td->io_ops->flags & FIO_PIPEIO)
381 log_info("fio: cannot pre-read files with an IO engine"
382 " that isn't seekable. Pre-read disabled.\n");
389 * This function leaks the buffer
391 static char *to_kmg(unsigned int val)
393 char *buf = malloc(32);
394 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
405 snprintf(buf, 31, "%u%c", val, *p);
409 /* External engines are specified by "external:name.o") */
410 static const char *get_engine_name(const char *str)
412 char *p = strstr(str, ":");
418 strip_blank_front(&p);
423 static int exists_and_not_file(const char *filename)
427 if (lstat(filename, &sb) == -1)
430 if (S_ISREG(sb.st_mode))
436 void td_fill_rand_seeds(struct thread_data *td)
438 os_random_seed(td->rand_seeds[0], &td->bsrange_state);
439 os_random_seed(td->rand_seeds[1], &td->verify_state);
440 os_random_seed(td->rand_seeds[2], &td->rwmix_state);
442 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
443 os_random_seed(td->rand_seeds[3], &td->next_file_state);
445 os_random_seed(td->rand_seeds[5], &td->file_size_state);
450 if (td->o.rand_repeatable)
451 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
453 os_random_seed(td->rand_seeds[4], &td->random_state);
457 * Initialize the various random states we need (random io, block size ranges,
458 * read/write mix, etc).
460 static int init_random_state(struct thread_data *td)
464 fd = open("/dev/urandom", O_RDONLY);
466 td_verror(td, errno, "open");
470 if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
471 (int) sizeof(td->rand_seeds)) {
472 td_verror(td, EIO, "read");
478 td_fill_rand_seeds(td);
483 * Adds a job to the list of things todo. Sanitizes the various options
484 * to make sure we don't have conflicts, and initializes various
487 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
489 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
490 "randread", "randwrite", "randrw" };
493 char fname[PATH_MAX];
494 int numjobs, file_alloced;
497 * the def_thread is just for options, it's not a real job
499 if (td == &def_thread)
503 * if we are just dumping the output command line, don't add the job
510 engine = get_engine_name(td->o.ioengine);
511 td->io_ops = load_ioengine(td, engine);
513 log_err("fio: failed to load engine %s\n", engine);
517 if (td->o.use_thread)
523 td->io_ops->flags |= FIO_RAWIO;
526 if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
529 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
530 add_file(td, jobname);
532 for (i = 0; i < td->o.nr_files; i++) {
533 sprintf(fname, "%s.%d.%d", jobname,
534 td->thread_number, i);
540 if (fixup_options(td))
543 if (td->io_ops->flags & FIO_DISKLESSIO) {
546 for_each_file(td, f, i)
547 f->real_file_size = -1ULL;
550 td->mutex = fio_mutex_init(0);
552 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
553 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
554 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
555 td->ddir_nr = td->o.ddir_nr;
557 if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
562 td->groupid = groupid;
565 if (init_random_state(td))
571 if (td->o.write_lat_log) {
572 setup_log(&td->ts.slat_log);
573 setup_log(&td->ts.clat_log);
575 if (td->o.write_bw_log)
576 setup_log(&td->ts.bw_log);
579 td->o.name = strdup(jobname);
583 if (!strcmp(td->io_ops->name, "cpuio")) {
584 log_info("%s: ioengine=cpu, cpuload=%u,"
585 " cpucycle=%u\n", td->o.name,
589 char *c1, *c2, *c3, *c4;
591 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
592 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
593 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
594 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
596 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
597 " ioengine=%s, iodepth=%u\n",
598 td->o.name, td->groupid,
599 ddir_str[td->o.td_ddir],
609 } else if (job_add_num == 1)
614 * recurse add identical jobs, clear numjobs and stonewall options
615 * as they don't apply to sub-jobs
617 numjobs = td->o.numjobs;
619 struct thread_data *td_new = get_new_job(0, td);
624 td_new->o.numjobs = 1;
625 td_new->o.stonewall = 0;
626 td_new->o.new_group = 0;
629 td_new->o.filename = NULL;
630 td_new->files_index = 0;
631 td_new->files_size = 0;
632 td_new->files = NULL;
635 job_add_num = numjobs - 1;
637 if (add_job(td_new, jobname, job_add_num))
647 static int skip_this_section(const char *name)
651 if (!strncmp(name, "global", 6))
654 return strcmp(job_section, name);
657 static int is_empty_or_comment(char *line)
661 for (i = 0; i < strlen(line); i++) {
666 if (!isspace(line[i]) && !iscntrl(line[i]))
674 * This is our [ini] type file parser.
676 static int parse_jobs_ini(char *file, int stonewall_flag)
679 struct thread_data *td;
683 int ret = 0, stonewall;
688 int i, alloc_opts, num_opts;
690 if (!strcmp(file, "-"))
693 f = fopen(file, "r");
696 perror("fopen job file");
700 string = malloc(4096);
703 * it's really 256 + small bit, 280 should suffice
706 memset(name, 0, 280);
709 opts = malloc(sizeof(char *) * alloc_opts);
712 stonewall = stonewall_flag;
715 * if skip_fgets is set, we already have loaded a line we
719 p = fgets(string, 4095, f);
725 strip_blank_front(&p);
728 if (is_empty_or_comment(p))
730 if (sscanf(p, "[%255s]", name) != 1) {
733 log_err("fio: option <%s> outside of [] job section\n",
738 if (skip_this_section(name)) {
744 global = !strncmp(name, "global", 6);
746 name[strlen(name) - 1] = '\0';
752 log_info("--name=%s ", name);
756 td = get_new_job(global, &def_thread);
763 * Seperate multiple job files by a stonewall
765 if (!global && stonewall) {
766 td->o.stonewall = stonewall;
771 memset(opts, 0, alloc_opts * sizeof(char *));
773 while ((p = fgets(string, 4096, f)) != NULL) {
774 if (is_empty_or_comment(p))
777 strip_blank_front(&p);
780 * new section, break out and make sure we don't
781 * fgets() a new line at the top.
790 if (num_opts == alloc_opts) {
793 alloc_opts * sizeof(char *));
796 opts[num_opts] = strdup(p);
800 ret = fio_options_parse(td, opts, num_opts);
803 for (i = 0; i < num_opts; i++)
804 log_info("--%s ", opts[i]);
806 ret = add_job(td, name, 0);
808 log_err("fio: job %s dropped\n", name);
812 for (i = 0; i < num_opts; i++)
820 for (i = 0; i < num_opts; i++)
831 static int fill_def_thread(void)
833 memset(&def_thread, 0, sizeof(def_thread));
835 fio_getaffinity(getpid(), &def_thread.o.cpumask);
838 * fill default options
840 fio_fill_default_options(&def_thread);
842 def_thread.o.timeout = def_timeout;
843 def_thread.o.write_bw_log = write_bw_log;
844 def_thread.o.write_lat_log = write_lat_log;
849 static void free_shm(void)
851 struct shmid_ds sbuf;
858 fio_debug_jobp = NULL;
860 shmctl(shm_id, IPC_RMID, &sbuf);
867 * The thread area is shared between the main process and the job
868 * threads/processes. So setup a shared memory segment that will hold
869 * all the job info. We use the end of the region for keeping track of
870 * open files across jobs, for file sharing.
872 static int setup_thread_area(void)
877 * 1024 is too much on some machines, scale max_jobs if
878 * we get a failure that looks like too large a shm segment
881 size_t size = max_jobs * sizeof(struct thread_data);
883 size += file_hash_size;
884 size += sizeof(unsigned int);
886 shm_id = shmget(0, size, IPC_CREAT | 0600);
889 if (errno != EINVAL) {
900 threads = shmat(shm_id, NULL, 0);
901 if (threads == (void *) -1) {
906 memset(threads, 0, max_jobs * sizeof(struct thread_data));
907 hash = (void *) threads + max_jobs * sizeof(struct thread_data);
908 fio_debug_jobp = (void *) hash + file_hash_size;
909 *fio_debug_jobp = -1;
910 file_hash_init(hash);
915 static void usage(const char *name)
917 printf("%s\n", fio_version_string);
918 printf("%s [options] [job options] <job file(s)>\n", name);
919 printf("\t--debug=options\tEnable debug logging\n");
920 printf("\t--output\tWrite output to file\n");
921 printf("\t--timeout\tRuntime in seconds\n");
922 printf("\t--latency-log\tGenerate per-job latency logs\n");
923 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
924 printf("\t--minimal\tMinimal (terse) output\n");
925 printf("\t--version\tPrint version info and exit\n");
926 printf("\t--help\t\tPrint this page\n");
927 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
929 printf("\t--showcmd\tTurn a job file into command line options\n");
930 printf("\t--eta=when\tWhen ETA estimate should be printed\n");
931 printf("\t \tMay be \"always\", \"never\" or \"auto\"\n");
932 printf("\t--readonly\tTurn on safety read-only checks, preventing"
934 printf("\t--section=name\tOnly run specified section in job file\n");
935 printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
937 printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>\n");
941 struct debug_level debug_levels[] = {
942 { .name = "process", .shift = FD_PROCESS, },
943 { .name = "file", .shift = FD_FILE, },
944 { .name = "io", .shift = FD_IO, },
945 { .name = "mem", .shift = FD_MEM, },
946 { .name = "blktrace", .shift = FD_BLKTRACE },
947 { .name = "verify", .shift = FD_VERIFY },
948 { .name = "random", .shift = FD_RANDOM },
949 { .name = "parse", .shift = FD_PARSE },
950 { .name = "diskutil", .shift = FD_DISKUTIL },
951 { .name = "job", .shift = FD_JOB },
952 { .name = "mutex", .shift = FD_MUTEX },
956 static int set_debug(const char *string)
958 struct debug_level *dl;
959 char *p = (char *) string;
963 if (!strcmp(string, "?") || !strcmp(string, "help")) {
966 log_info("fio: dumping debug options:");
967 for (i = 0; debug_levels[i].name; i++) {
968 dl = &debug_levels[i];
969 log_info("%s,", dl->name);
975 while ((opt = strsep(&p, ",")) != NULL) {
978 if (!strncmp(opt, "all", 3)) {
979 log_info("fio: set all debug options\n");
984 for (i = 0; debug_levels[i].name; i++) {
985 dl = &debug_levels[i];
986 found = !strncmp(opt, dl->name, strlen(dl->name));
990 if (dl->shift == FD_JOB) {
991 opt = strchr(opt, ':');
993 log_err("fio: missing job number\n");
997 fio_debug_jobno = atoi(opt);
998 log_info("fio: set debug jobno %d\n",
1001 log_info("fio: set debug option %s\n", opt);
1002 fio_debug |= (1UL << dl->shift);
1008 log_err("fio: debug mask %s not found\n", opt);
1013 static int set_debug(const char *string)
1015 log_err("fio: debug tracing not included in build\n");
1020 static int load_profile(const char *profile)
1022 struct thread_data *td, *td_parent;
1024 int i, in_global = 1;
1027 dprint(FD_PARSE, "loading profile %s\n", profile);
1029 for (i = 0; i < PROFILE_END; i++) {
1030 if (!strcmp(profile, profiles[i]))
1034 if (i == PROFILE_END) {
1035 log_err("fio: unknown profile %s\n", profile);
1039 o = fio_prof_strings[i];
1044 td_parent = td = NULL;
1046 if (!strncmp(o[i], "name", 4)) {
1049 add_job(td, jobname, 0);
1051 sprintf(jobname, "%s", o[i] + 5);
1053 if (in_global && !td_parent)
1054 td_parent = get_new_job(1, &def_thread);
1055 else if (!in_global && !td) {
1057 td_parent = &def_thread;
1058 td = get_new_job(0, td_parent);
1061 fio_options_parse(td_parent, (char **) &o[i], 1);
1063 fio_options_parse(td, (char **) &o[i], 1);
1067 add_job(td, jobname, 0);
1071 static int parse_cmd_line(int argc, char *argv[])
1073 struct thread_data *td = NULL;
1074 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1076 while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1079 smalloc_pool_size = atoi(optarg);
1082 def_timeout = atoi(optarg);
1091 f_out = fopen(optarg, "w+");
1093 perror("fopen output");
1105 exit(fio_show_option_help(optarg));
1113 printf("%s\n", fio_version_string);
1116 if (!strcmp("always", optarg))
1117 eta_print = FIO_ETA_ALWAYS;
1118 else if (!strcmp("never", optarg))
1119 eta_print = FIO_ETA_NEVER;
1122 if (set_debug(optarg))
1126 if (!strcmp(optarg, "global")) {
1127 log_err("fio: can't use global as only "
1135 job_section = strdup(optarg);
1138 if (load_profile(optarg))
1141 case FIO_GETOPT_JOB: {
1142 const char *opt = l_opts[lidx].name;
1145 if (!strncmp(opt, "name", 4) && td) {
1146 ret = add_job(td, td->o.name ?: "fio", 0);
1154 int is_section = !strncmp(opt, "name", 4);
1157 if (!is_section || !strncmp(val, "global", 6))
1160 if (is_section && skip_this_section(val))
1163 td = get_new_job(global, &def_thread);
1168 ret = fio_cmd_option_parse(td, opt, val);
1183 ret = add_job(td, td->o.name ?: "fio", 0);
1188 while (optind < argc) {
1190 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1191 ini_file[ini_idx - 1] = strdup(argv[optind]);
1198 int parse_options(int argc, char *argv[])
1205 fio_options_dup_and_init(l_opts);
1207 if (setup_thread_area())
1209 if (fill_def_thread())
1212 job_files = parse_cmd_line(argc, argv);
1214 for (i = 0; i < job_files; i++) {
1215 if (fill_def_thread())
1217 if (parse_jobs_ini(ini_file[i], i))
1223 options_mem_free(&def_thread);
1225 if (!thread_number) {
1229 log_err("No jobs defined(s)\n\n");
1234 if (def_thread.o.gtod_offload) {
1236 fio_gtod_offload = 1;
1237 fio_gtod_cpu = def_thread.o.gtod_cpu;