2 * This file contains job initialization and setup functions.
14 #include <sys/types.h>
23 static char fio_version_string[] = "fio 1.26.4";
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,
145 * Return a free job structure.
147 static struct thread_data *get_new_job(int global, struct thread_data *parent)
149 struct thread_data *td;
153 if (thread_number >= max_jobs)
156 td = &threads[thread_number++];
159 dup_files(td, parent);
160 options_mem_dupe(td);
162 td->thread_number = thread_number;
166 static void put_job(struct thread_data *td)
168 if (td == &def_thread)
172 log_info("fio: %s\n", td->verror);
174 memset(&threads[td->thread_number - 1], 0, sizeof(*td));
178 static int setup_rate(struct thread_data *td)
180 unsigned long nr_reads_per_msec;
181 unsigned long long rate;
184 if (!td->o.rate && !td->o.rate_iops)
188 bs = td->o.rw_min_bs;
189 else if (td_read(td))
190 bs = td->o.min_bs[DDIR_READ];
192 bs = td->o.min_bs[DDIR_WRITE];
196 nr_reads_per_msec = (rate * 1024 * 1000LL) / bs;
198 nr_reads_per_msec = td->o.rate_iops * 1000UL;
200 if (!nr_reads_per_msec) {
201 log_err("rate lower than supported\n");
205 td->rate_usec_cycle = 1000000000ULL / nr_reads_per_msec;
206 td->rate_pending_usleep = 0;
210 static int fixed_block_size(struct thread_options *o)
212 return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
213 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
214 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
218 * Lazy way of fixing up options that depend on each other. We could also
219 * define option callback handlers, but this is easier.
221 static int fixup_options(struct thread_data *td)
223 struct thread_options *o = &td->o;
225 #ifndef FIO_HAVE_PSHARED_MUTEX
226 if (!td->o.use_thread) {
227 log_info("fio: this platform does not support process shared"
228 " mutexes, forcing use of threads. Use the 'thread'"
229 " option to get rid of this warning.\n");
230 td->o.use_thread = 1;
234 #ifndef FIO_HAVE_CPU_AFFINITY
235 if (td->o.gtod_cpu) {
236 log_err("fio: platform must support CPU affinity for"
237 "gettimeofday() offloading\n");
242 if (read_only && td_write(td)) {
243 log_err("fio: job <%s> has write bit set, but fio is in"
244 " read-only mode\n", td->o.name);
248 if (o->write_iolog_file && o->read_iolog_file) {
249 log_err("fio: read iolog overrides write_iolog\n");
250 free(o->write_iolog_file);
251 o->write_iolog_file = NULL;
255 * only really works for sequential io for now, and with 1 file
257 if (o->zone_size && td_random(td) && o->open_files == 1)
261 * Reads can do overwrites, we always need to pre-create the file
263 if (td_read(td) || td_rw(td))
266 if (!o->min_bs[DDIR_READ])
267 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
268 if (!o->max_bs[DDIR_READ])
269 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
270 if (!o->min_bs[DDIR_WRITE])
271 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
272 if (!o->max_bs[DDIR_WRITE])
273 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
275 o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
278 * For random IO, allow blockalign offset other than min_bs.
280 if (!o->ba[DDIR_READ] || !td_random(td))
281 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
282 if (!o->ba[DDIR_WRITE] || !td_random(td))
283 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
285 if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
286 o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
287 !td->o.norandommap) {
288 log_err("fio: Any use of blockalign= turns off randommap\n");
289 td->o.norandommap = 1;
292 if (!o->file_size_high)
293 o->file_size_high = o->file_size_low;
295 if (o->norandommap && o->verify != VERIFY_NONE
296 && !fixed_block_size(o)) {
297 log_err("fio: norandommap given for variable block sizes, "
298 "verify disabled\n");
299 o->verify = VERIFY_NONE;
301 if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
302 log_err("fio: bs_unaligned may not work with raw io\n");
305 * thinktime_spin must be less than thinktime
307 if (o->thinktime_spin > o->thinktime)
308 o->thinktime_spin = o->thinktime;
311 * The low water mark cannot be bigger than the iodepth
313 if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
315 * syslet work around - if the workload is sequential,
316 * we want to let the queue drain all the way down to
317 * avoid seeking between async threads
319 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
322 o->iodepth_low = o->iodepth;
326 * If batch number isn't set, default to the same as iodepth
328 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
329 o->iodepth_batch = o->iodepth;
331 if (o->nr_files > td->files_index)
332 o->nr_files = td->files_index;
334 if (o->open_files > o->nr_files || !o->open_files)
335 o->open_files = o->nr_files;
337 if ((o->rate && o->rate_iops) || (o->ratemin && o->rate_iops_min)) {
338 log_err("fio: rate and rate_iops are mutually exclusive\n");
341 if ((o->rate < o->ratemin) || (o->rate_iops < o->rate_iops_min)) {
342 log_err("fio: minimum rate exceeds rate\n");
346 if (!o->timeout && o->time_based) {
347 log_err("fio: time_based requires a runtime/timeout setting\n");
351 if (o->fill_device && !o->size)
354 if (td_rw(td) && td->o.verify != VERIFY_NONE)
355 log_info("fio: mixed read/write workload with verify. May not "
356 "work as expected, unless you pre-populated the file\n");
358 if (td->o.verify != VERIFY_NONE)
359 td->o.refill_buffers = 1;
362 td->o.invalidate_cache = 0;
368 * This function leaks the buffer
370 static char *to_kmg(unsigned int val)
372 char *buf = malloc(32);
373 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
384 snprintf(buf, 31, "%u%c", val, *p);
388 /* External engines are specified by "external:name.o") */
389 static const char *get_engine_name(const char *str)
391 char *p = strstr(str, ":");
397 strip_blank_front(&p);
402 static int exists_and_not_file(const char *filename)
406 if (lstat(filename, &sb) == -1)
409 if (S_ISREG(sb.st_mode))
415 void td_fill_rand_seeds(struct thread_data *td)
417 os_random_seed(td->rand_seeds[0], &td->bsrange_state);
418 os_random_seed(td->rand_seeds[1], &td->verify_state);
419 os_random_seed(td->rand_seeds[2], &td->rwmix_state);
421 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
422 os_random_seed(td->rand_seeds[3], &td->next_file_state);
424 os_random_seed(td->rand_seeds[5], &td->file_size_state);
429 if (td->o.rand_repeatable)
430 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
432 os_random_seed(td->rand_seeds[4], &td->random_state);
436 * Initialize the various random states we need (random io, block size ranges,
437 * read/write mix, etc).
439 static int init_random_state(struct thread_data *td)
443 fd = open("/dev/urandom", O_RDONLY);
445 td_verror(td, errno, "open");
449 if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
450 (int) sizeof(td->rand_seeds)) {
451 td_verror(td, EIO, "read");
457 td_fill_rand_seeds(td);
462 * Adds a job to the list of things todo. Sanitizes the various options
463 * to make sure we don't have conflicts, and initializes various
466 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
468 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
469 "randread", "randwrite", "randrw" };
472 char fname[PATH_MAX];
473 int numjobs, file_alloced;
476 * the def_thread is just for options, it's not a real job
478 if (td == &def_thread)
482 * if we are just dumping the output command line, don't add the job
489 engine = get_engine_name(td->o.ioengine);
490 td->io_ops = load_ioengine(td, engine);
492 log_err("fio: failed to load engine %s\n", engine);
496 if (td->o.use_thread)
502 td->io_ops->flags |= FIO_RAWIO;
505 if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
508 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
509 add_file(td, jobname);
511 for (i = 0; i < td->o.nr_files; i++) {
512 sprintf(fname, "%s.%d.%d", jobname,
513 td->thread_number, i);
519 if (fixup_options(td))
522 if (td->io_ops->flags & FIO_DISKLESSIO) {
525 for_each_file(td, f, i)
526 f->real_file_size = -1ULL;
529 td->mutex = fio_mutex_init(0);
531 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
532 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
533 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
534 td->ddir_nr = td->o.ddir_nr;
536 if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
537 && prev_group_jobs) {
542 td->groupid = groupid;
545 if (init_random_state(td))
551 if (td->o.write_lat_log) {
552 setup_log(&td->ts.slat_log);
553 setup_log(&td->ts.clat_log);
555 if (td->o.write_bw_log)
556 setup_log(&td->ts.bw_log);
559 td->o.name = strdup(jobname);
563 if (!strcmp(td->io_ops->name, "cpuio")) {
564 log_info("%s: ioengine=cpu, cpuload=%u,"
565 " cpucycle=%u\n", td->o.name,
569 char *c1, *c2, *c3, *c4;
571 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
572 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
573 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
574 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
576 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
577 " ioengine=%s, iodepth=%u\n",
578 td->o.name, td->groupid,
579 ddir_str[td->o.td_ddir],
589 } else if (job_add_num == 1)
594 * recurse add identical jobs, clear numjobs and stonewall options
595 * as they don't apply to sub-jobs
597 numjobs = td->o.numjobs;
599 struct thread_data *td_new = get_new_job(0, td);
604 td_new->o.numjobs = 1;
605 td_new->o.stonewall = 0;
606 td_new->o.new_group = 0;
609 td_new->o.filename = NULL;
610 td_new->files_index = 0;
611 td_new->files_size = 0;
612 td_new->files = NULL;
615 job_add_num = numjobs - 1;
617 if (add_job(td_new, jobname, job_add_num))
627 static int skip_this_section(const char *name)
631 if (!strncmp(name, "global", 6))
634 return strcmp(job_section, name);
637 static int is_empty_or_comment(char *line)
641 for (i = 0; i < strlen(line); i++) {
646 if (!isspace(line[i]) && !iscntrl(line[i]))
654 * This is our [ini] type file parser.
656 static int parse_jobs_ini(char *file, int stonewall_flag)
659 struct thread_data *td;
663 int ret = 0, stonewall;
668 int i, alloc_opts, num_opts;
670 if (!strcmp(file, "-"))
673 f = fopen(file, "r");
676 perror("fopen job file");
680 string = malloc(4096);
683 * it's really 256 + small bit, 280 should suffice
686 memset(name, 0, 280);
689 opts = malloc(sizeof(char *) * alloc_opts);
692 stonewall = stonewall_flag;
695 * if skip_fgets is set, we already have loaded a line we
699 p = fgets(string, 4095, f);
705 strip_blank_front(&p);
708 if (is_empty_or_comment(p))
710 if (sscanf(p, "[%255s]", name) != 1) {
713 log_err("fio: option <%s> outside of [] job section\n",
718 if (skip_this_section(name)) {
724 global = !strncmp(name, "global", 6);
726 name[strlen(name) - 1] = '\0';
732 log_info("--name=%s ", name);
736 td = get_new_job(global, &def_thread);
743 * Seperate multiple job files by a stonewall
745 if (!global && stonewall) {
746 td->o.stonewall = stonewall;
751 memset(opts, 0, alloc_opts * sizeof(char *));
753 while ((p = fgets(string, 4096, f)) != NULL) {
754 if (is_empty_or_comment(p))
757 strip_blank_front(&p);
760 * new section, break out and make sure we don't
761 * fgets() a new line at the top.
770 if (num_opts == alloc_opts) {
773 alloc_opts * sizeof(char *));
776 opts[num_opts] = strdup(p);
780 ret = fio_options_parse(td, opts, num_opts);
783 for (i = 0; i < num_opts; i++)
784 log_info("--%s ", opts[i]);
786 ret = add_job(td, name, 0);
788 log_err("fio: job %s dropped\n", name);
792 for (i = 0; i < num_opts; i++)
800 for (i = 0; i < num_opts; i++)
811 static int fill_def_thread(void)
813 memset(&def_thread, 0, sizeof(def_thread));
815 fio_getaffinity(getpid(), &def_thread.o.cpumask);
818 * fill default options
820 fio_fill_default_options(&def_thread);
822 def_thread.o.timeout = def_timeout;
823 def_thread.o.write_bw_log = write_bw_log;
824 def_thread.o.write_lat_log = write_lat_log;
829 static void free_shm(void)
831 struct shmid_ds sbuf;
838 fio_debug_jobp = NULL;
840 shmctl(shm_id, IPC_RMID, &sbuf);
847 * The thread area is shared between the main process and the job
848 * threads/processes. So setup a shared memory segment that will hold
849 * all the job info. We use the end of the region for keeping track of
850 * open files across jobs, for file sharing.
852 static int setup_thread_area(void)
857 * 1024 is too much on some machines, scale max_jobs if
858 * we get a failure that looks like too large a shm segment
861 size_t size = max_jobs * sizeof(struct thread_data);
863 size += file_hash_size;
864 size += sizeof(unsigned int);
866 shm_id = shmget(0, size, IPC_CREAT | 0600);
869 if (errno != EINVAL) {
880 threads = shmat(shm_id, NULL, 0);
881 if (threads == (void *) -1) {
886 memset(threads, 0, max_jobs * sizeof(struct thread_data));
887 hash = (void *) threads + max_jobs * sizeof(struct thread_data);
888 fio_debug_jobp = (void *) hash + file_hash_size;
889 *fio_debug_jobp = -1;
890 file_hash_init(hash);
895 static void usage(const char *name)
897 printf("%s\n", fio_version_string);
898 printf("%s [options] [job options] <job file(s)>\n", name);
899 printf("\t--debug=options\tEnable debug logging\n");
900 printf("\t--output\tWrite output to file\n");
901 printf("\t--timeout\tRuntime in seconds\n");
902 printf("\t--latency-log\tGenerate per-job latency logs\n");
903 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
904 printf("\t--minimal\tMinimal (terse) output\n");
905 printf("\t--version\tPrint version info and exit\n");
906 printf("\t--help\t\tPrint this page\n");
907 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
909 printf("\t--showcmd\tTurn a job file into command line options\n");
910 printf("\t--eta=when\tWhen ETA estimate should be printed\n");
911 printf("\t \tMay be \"always\", \"never\" or \"auto\"\n");
912 printf("\t--readonly\tTurn on safety read-only checks, preventing"
914 printf("\t--section=name\tOnly run specified section in job file\n");
915 printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
920 struct debug_level debug_levels[] = {
921 { .name = "process", .shift = FD_PROCESS, },
922 { .name = "file", .shift = FD_FILE, },
923 { .name = "io", .shift = FD_IO, },
924 { .name = "mem", .shift = FD_MEM, },
925 { .name = "blktrace", .shift = FD_BLKTRACE },
926 { .name = "verify", .shift = FD_VERIFY },
927 { .name = "random", .shift = FD_RANDOM },
928 { .name = "parse", .shift = FD_PARSE },
929 { .name = "diskutil", .shift = FD_DISKUTIL },
930 { .name = "job", .shift = FD_JOB },
931 { .name = "mutex", .shift = FD_MUTEX },
935 static int set_debug(const char *string)
937 struct debug_level *dl;
938 char *p = (char *) string;
942 if (!strcmp(string, "?") || !strcmp(string, "help")) {
945 log_info("fio: dumping debug options:");
946 for (i = 0; debug_levels[i].name; i++) {
947 dl = &debug_levels[i];
948 log_info("%s,", dl->name);
954 while ((opt = strsep(&p, ",")) != NULL) {
957 if (!strncmp(opt, "all", 3)) {
958 log_info("fio: set all debug options\n");
963 for (i = 0; debug_levels[i].name; i++) {
964 dl = &debug_levels[i];
965 found = !strncmp(opt, dl->name, strlen(dl->name));
969 if (dl->shift == FD_JOB) {
970 opt = strchr(opt, ':');
972 log_err("fio: missing job number\n");
976 fio_debug_jobno = atoi(opt);
977 log_info("fio: set debug jobno %d\n",
980 log_info("fio: set debug option %s\n", opt);
981 fio_debug |= (1UL << dl->shift);
987 log_err("fio: debug mask %s not found\n", opt);
992 static int set_debug(const char *string)
994 log_err("fio: debug tracing not included in build\n");
999 static int parse_cmd_line(int argc, char *argv[])
1001 struct thread_data *td = NULL;
1002 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1004 while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1007 smalloc_pool_size = atoi(optarg);
1010 def_timeout = atoi(optarg);
1019 f_out = fopen(optarg, "w+");
1021 perror("fopen output");
1033 exit(fio_show_option_help(optarg));
1041 printf("%s\n", fio_version_string);
1044 if (!strcmp("always", optarg))
1045 eta_print = FIO_ETA_ALWAYS;
1046 else if (!strcmp("never", optarg))
1047 eta_print = FIO_ETA_NEVER;
1050 if (set_debug(optarg))
1054 if (!strcmp(optarg, "global")) {
1055 log_err("fio: can't use global as only "
1063 job_section = strdup(optarg);
1065 case FIO_GETOPT_JOB: {
1066 const char *opt = l_opts[lidx].name;
1069 if (!strncmp(opt, "name", 4) && td) {
1070 ret = add_job(td, td->o.name ?: "fio", 0);
1078 int is_section = !strncmp(opt, "name", 4);
1081 if (!is_section || !strncmp(val, "global", 6))
1084 if (is_section && skip_this_section(val))
1087 td = get_new_job(global, &def_thread);
1092 ret = fio_cmd_option_parse(td, opt, val);
1107 ret = add_job(td, td->o.name ?: "fio", 0);
1112 while (optind < argc) {
1114 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1115 ini_file[ini_idx - 1] = strdup(argv[optind]);
1122 int parse_options(int argc, char *argv[])
1129 fio_options_dup_and_init(l_opts);
1131 if (setup_thread_area())
1133 if (fill_def_thread())
1136 job_files = parse_cmd_line(argc, argv);
1138 for (i = 0; i < job_files; i++) {
1139 if (fill_def_thread())
1141 if (parse_jobs_ini(ini_file[i], i))
1147 options_mem_free(&def_thread);
1149 if (!thread_number) {
1153 log_err("No jobs defined(s)\n");
1158 if (def_thread.o.gtod_offload) {
1160 fio_gtod_offload = 1;
1161 fio_gtod_cpu = def_thread.o.gtod_cpu;