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_PSHARED_MUTEX
218 if (!td->o.use_thread) {
219 log_info("fio: this platform does not support process shared"
220 " mutexes, forcing use of threads. Use the 'thread'"
221 " option to get rid of this warning.\n");
222 td->o.use_thread = 1;
226 #ifndef FIO_HAVE_CPU_AFFINITY
227 if (td->o.gtod_cpu) {
228 log_err("fio: platform must support CPU affinity for"
229 "gettimeofday() offloading\n");
234 if (read_only && td_write(td)) {
235 log_err("fio: job <%s> has write bit set, but fio is in"
236 " read-only mode\n", td->o.name);
240 if (o->write_iolog_file && o->read_iolog_file) {
241 log_err("fio: read iolog overrides write_iolog\n");
242 free(o->write_iolog_file);
243 o->write_iolog_file = NULL;
247 * only really works for sequential io for now, and with 1 file
249 if (o->zone_size && td_random(td) && o->open_files == 1)
253 * Reads can do overwrites, we always need to pre-create the file
255 if (td_read(td) || td_rw(td))
258 if (!o->min_bs[DDIR_READ])
259 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
260 if (!o->max_bs[DDIR_READ])
261 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
262 if (!o->min_bs[DDIR_WRITE])
263 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
264 if (!o->max_bs[DDIR_WRITE])
265 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
267 o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
269 if (!o->file_size_high)
270 o->file_size_high = o->file_size_low;
272 if (o->norandommap && o->verify != VERIFY_NONE) {
273 log_err("fio: norandommap given, verify disabled\n");
274 o->verify = VERIFY_NONE;
276 if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
277 log_err("fio: bs_unaligned may not work with raw io\n");
280 * thinktime_spin must be less than thinktime
282 if (o->thinktime_spin > o->thinktime)
283 o->thinktime_spin = o->thinktime;
286 * The low water mark cannot be bigger than the iodepth
288 if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
290 * syslet work around - if the workload is sequential,
291 * we want to let the queue drain all the way down to
292 * avoid seeking between async threads
294 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
297 o->iodepth_low = o->iodepth;
301 * If batch number isn't set, default to the same as iodepth
303 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
304 o->iodepth_batch = o->iodepth;
306 if (o->nr_files > td->files_index)
307 o->nr_files = td->files_index;
309 if (o->open_files > o->nr_files || !o->open_files)
310 o->open_files = o->nr_files;
312 if ((o->rate && o->rate_iops) || (o->ratemin && o->rate_iops_min)) {
313 log_err("fio: rate and rate_iops are mutually exclusive\n");
316 if ((o->rate < o->ratemin) || (o->rate_iops < o->rate_iops_min)) {
317 log_err("fio: minimum rate exceeds rate\n");
321 if (!o->timeout && o->time_based) {
322 log_err("fio: time_based requires a runtime/timeout setting\n");
326 if (o->fill_device && !o->size)
329 if (td_rw(td) && td->o.verify != VERIFY_NONE)
330 log_info("fio: mixed read/write workload with verify. May not "
331 "work as expected, unless you pre-populated the file\n");
333 if (td->o.verify != VERIFY_NONE)
334 td->o.refill_buffers = 1;
340 * This function leaks the buffer
342 static char *to_kmg(unsigned int val)
344 char *buf = malloc(32);
345 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
356 snprintf(buf, 31, "%u%c", val, *p);
360 /* External engines are specified by "external:name.o") */
361 static const char *get_engine_name(const char *str)
363 char *p = strstr(str, ":");
369 strip_blank_front(&p);
374 static int exists_and_not_file(const char *filename)
378 if (lstat(filename, &sb) == -1)
381 if (S_ISREG(sb.st_mode))
388 * Initialize the various random states we need (random io, block size ranges,
389 * read/write mix, etc).
391 static int init_random_state(struct thread_data *td)
393 unsigned long seeds[6];
396 fd = open("/dev/urandom", O_RDONLY);
398 td_verror(td, errno, "open");
402 if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
403 td_verror(td, EIO, "read");
410 os_random_seed(seeds[0], &td->bsrange_state);
411 os_random_seed(seeds[1], &td->verify_state);
412 os_random_seed(seeds[2], &td->rwmix_state);
414 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
415 os_random_seed(seeds[3], &td->next_file_state);
417 os_random_seed(seeds[5], &td->file_size_state);
422 if (td->o.rand_repeatable)
423 seeds[4] = FIO_RANDSEED * td->thread_number;
425 os_random_seed(seeds[4], &td->random_state);
430 * Adds a job to the list of things todo. Sanitizes the various options
431 * to make sure we don't have conflicts, and initializes various
434 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
436 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
437 "randread", "randwrite", "randrw" };
440 char fname[PATH_MAX];
441 int numjobs, file_alloced;
444 * the def_thread is just for options, it's not a real job
446 if (td == &def_thread)
450 * if we are just dumping the output command line, don't add the job
457 engine = get_engine_name(td->o.ioengine);
458 td->io_ops = load_ioengine(td, engine);
460 log_err("fio: failed to load engine %s\n", engine);
464 if (td->o.use_thread)
470 td->io_ops->flags |= FIO_RAWIO;
473 if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
476 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
477 add_file(td, jobname);
479 for (i = 0; i < td->o.nr_files; i++) {
480 sprintf(fname, "%s.%d.%d", jobname,
481 td->thread_number, i);
487 if (fixup_options(td))
490 if (td->io_ops->flags & FIO_DISKLESSIO) {
493 for_each_file(td, f, i)
494 f->real_file_size = -1ULL;
497 td->mutex = fio_mutex_init(0);
499 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
500 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
501 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
502 td->ddir_nr = td->o.ddir_nr;
504 if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
505 && prev_group_jobs) {
510 td->groupid = groupid;
513 if (init_random_state(td))
519 if (td->o.write_lat_log) {
520 setup_log(&td->ts.slat_log);
521 setup_log(&td->ts.clat_log);
523 if (td->o.write_bw_log)
524 setup_log(&td->ts.bw_log);
527 td->o.name = strdup(jobname);
531 if (!strcmp(td->io_ops->name, "cpuio")) {
532 log_info("%s: ioengine=cpu, cpuload=%u,"
533 " cpucycle=%u\n", td->o.name,
537 char *c1, *c2, *c3, *c4;
539 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
540 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
541 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
542 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
544 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
545 " ioengine=%s, iodepth=%u\n",
546 td->o.name, td->groupid,
547 ddir_str[td->o.td_ddir],
557 } else if (job_add_num == 1)
562 * recurse add identical jobs, clear numjobs and stonewall options
563 * as they don't apply to sub-jobs
565 numjobs = td->o.numjobs;
567 struct thread_data *td_new = get_new_job(0, td);
572 td_new->o.numjobs = 1;
573 td_new->o.stonewall = 0;
574 td_new->o.new_group = 0;
577 td_new->o.filename = NULL;
578 td_new->files_index = 0;
579 td_new->files = NULL;
582 job_add_num = numjobs - 1;
584 if (add_job(td_new, jobname, job_add_num))
594 static int skip_this_section(const char *name)
598 if (!strncmp(name, "global", 6))
601 return strncmp(job_section, name, strlen(job_section));
604 static int is_empty_or_comment(char *line)
608 for (i = 0; i < strlen(line); i++) {
613 if (!isspace(line[i]) && !iscntrl(line[i]))
621 * This is our [ini] type file parser.
623 static int parse_jobs_ini(char *file, int stonewall_flag)
626 struct thread_data *td;
630 int ret = 0, stonewall;
635 int i, alloc_opts, num_opts;
637 if (!strcmp(file, "-"))
640 f = fopen(file, "r");
643 perror("fopen job file");
647 string = malloc(4096);
650 * it's really 256 + small bit, 280 should suffice
653 memset(name, 0, 280);
656 opts = malloc(sizeof(char *) * alloc_opts);
659 stonewall = stonewall_flag;
662 * if skip_fgets is set, we already have loaded a line we
666 p = fgets(string, 4095, f);
672 strip_blank_front(&p);
675 if (is_empty_or_comment(p))
677 if (sscanf(p, "[%255s]", name) != 1) {
680 log_err("fio: option <%s> outside of [] job section\n",
685 if (skip_this_section(name)) {
691 global = !strncmp(name, "global", 6);
693 name[strlen(name) - 1] = '\0';
699 log_info("--name=%s ", name);
703 td = get_new_job(global, &def_thread);
710 * Seperate multiple job files by a stonewall
712 if (!global && stonewall) {
713 td->o.stonewall = stonewall;
718 memset(opts, 0, alloc_opts * sizeof(char *));
720 while ((p = fgets(string, 4096, f)) != NULL) {
721 if (is_empty_or_comment(p))
724 strip_blank_front(&p);
727 * new section, break out and make sure we don't
728 * fgets() a new line at the top.
737 if (num_opts == alloc_opts) {
740 alloc_opts * sizeof(char *));
743 opts[num_opts] = strdup(p);
747 ret = fio_options_parse(td, opts, num_opts);
750 for (i = 0; i < num_opts; i++)
751 log_info("--%s ", opts[i]);
753 ret = add_job(td, name, 0);
755 log_err("fio: job %s dropped\n", name);
759 for (i = 0; i < num_opts; i++)
767 for (i = 0; i < num_opts; i++)
778 static int fill_def_thread(void)
780 memset(&def_thread, 0, sizeof(def_thread));
782 fio_getaffinity(getpid(), &def_thread.o.cpumask);
785 * fill default options
787 fio_fill_default_options(&def_thread);
789 def_thread.o.timeout = def_timeout;
790 def_thread.o.write_bw_log = write_bw_log;
791 def_thread.o.write_lat_log = write_lat_log;
796 static void free_shm(void)
798 struct shmid_ds sbuf;
805 fio_debug_jobp = NULL;
807 shmctl(shm_id, IPC_RMID, &sbuf);
814 * The thread area is shared between the main process and the job
815 * threads/processes. So setup a shared memory segment that will hold
816 * all the job info. We use the end of the region for keeping track of
817 * open files across jobs, for file sharing.
819 static int setup_thread_area(void)
824 * 1024 is too much on some machines, scale max_jobs if
825 * we get a failure that looks like too large a shm segment
828 size_t size = max_jobs * sizeof(struct thread_data);
830 size += file_hash_size;
831 size += sizeof(unsigned int);
833 shm_id = shmget(0, size, IPC_CREAT | 0600);
836 if (errno != EINVAL) {
847 threads = shmat(shm_id, NULL, 0);
848 if (threads == (void *) -1) {
853 memset(threads, 0, max_jobs * sizeof(struct thread_data));
854 hash = (void *) threads + max_jobs * sizeof(struct thread_data);
855 fio_debug_jobp = (void *) hash + file_hash_size;
856 *fio_debug_jobp = -1;
857 file_hash_init(hash);
862 static void usage(const char *name)
864 printf("%s\n", fio_version_string);
865 printf("%s [options] [job options] <job file(s)>\n", name);
866 printf("\t--debug=options\tEnable debug logging\n");
867 printf("\t--output\tWrite output to file\n");
868 printf("\t--timeout\tRuntime in seconds\n");
869 printf("\t--latency-log\tGenerate per-job latency logs\n");
870 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
871 printf("\t--minimal\tMinimal (terse) output\n");
872 printf("\t--version\tPrint version info and exit\n");
873 printf("\t--help\t\tPrint this page\n");
874 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
876 printf("\t--showcmd\tTurn a job file into command line options\n");
877 printf("\t--eta=when\tWhen ETA estimate should be printed\n");
878 printf("\t \tMay be \"always\", \"never\" or \"auto\"\n");
879 printf("\t--readonly\tTurn on safety read-only checks, preventing"
881 printf("\t--section=name\tOnly run specified section in job file\n");
882 printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
887 struct debug_level debug_levels[] = {
888 { .name = "process", .shift = FD_PROCESS, },
889 { .name = "file", .shift = FD_FILE, },
890 { .name = "io", .shift = FD_IO, },
891 { .name = "mem", .shift = FD_MEM, },
892 { .name = "blktrace", .shift = FD_BLKTRACE },
893 { .name = "verify", .shift = FD_VERIFY },
894 { .name = "random", .shift = FD_RANDOM },
895 { .name = "parse", .shift = FD_PARSE },
896 { .name = "diskutil", .shift = FD_DISKUTIL },
897 { .name = "job", .shift = FD_JOB },
898 { .name = "mutex", .shift = FD_MUTEX },
902 static int set_debug(const char *string)
904 struct debug_level *dl;
905 char *p = (char *) string;
909 if (!strcmp(string, "?") || !strcmp(string, "help")) {
912 log_info("fio: dumping debug options:");
913 for (i = 0; debug_levels[i].name; i++) {
914 dl = &debug_levels[i];
915 log_info("%s,", dl->name);
921 while ((opt = strsep(&p, ",")) != NULL) {
924 if (!strncmp(opt, "all", 3)) {
925 log_info("fio: set all debug options\n");
930 for (i = 0; debug_levels[i].name; i++) {
931 dl = &debug_levels[i];
932 found = !strncmp(opt, dl->name, strlen(dl->name));
936 if (dl->shift == FD_JOB) {
937 opt = strchr(opt, ':');
939 log_err("fio: missing job number\n");
943 fio_debug_jobno = atoi(opt);
944 log_info("fio: set debug jobno %d\n",
947 log_info("fio: set debug option %s\n", opt);
948 fio_debug |= (1UL << dl->shift);
954 log_err("fio: debug mask %s not found\n", opt);
959 static int set_debug(const char *string)
961 log_err("fio: debug tracing not included in build\n");
966 static int parse_cmd_line(int argc, char *argv[])
968 struct thread_data *td = NULL;
969 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
971 while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
974 smalloc_pool_size = atoi(optarg);
977 def_timeout = atoi(optarg);
986 f_out = fopen(optarg, "w+");
988 perror("fopen output");
1000 exit(fio_show_option_help(optarg));
1008 printf("%s\n", fio_version_string);
1011 if (!strcmp("always", optarg))
1012 eta_print = FIO_ETA_ALWAYS;
1013 else if (!strcmp("never", optarg))
1014 eta_print = FIO_ETA_NEVER;
1017 if (set_debug(optarg))
1021 if (!strcmp(optarg, "global")) {
1022 log_err("fio: can't use global as only "
1030 job_section = strdup(optarg);
1032 case FIO_GETOPT_JOB: {
1033 const char *opt = l_opts[lidx].name;
1036 if (!strncmp(opt, "name", 4) && td) {
1037 ret = add_job(td, td->o.name ?: "fio", 0);
1045 int is_section = !strncmp(opt, "name", 4);
1048 if (!is_section || !strncmp(val, "global", 6))
1051 if (is_section && skip_this_section(val))
1054 td = get_new_job(global, &def_thread);
1059 ret = fio_cmd_option_parse(td, opt, val);
1074 ret = add_job(td, td->o.name ?: "fio", 0);
1079 while (optind < argc) {
1081 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1082 ini_file[ini_idx - 1] = strdup(argv[optind]);
1089 int parse_options(int argc, char *argv[])
1096 fio_options_dup_and_init(l_opts);
1098 if (setup_thread_area())
1100 if (fill_def_thread())
1103 job_files = parse_cmd_line(argc, argv);
1105 for (i = 0; i < job_files; i++) {
1106 if (fill_def_thread())
1108 if (parse_jobs_ini(ini_file[i], i))
1114 options_mem_free(&def_thread);
1116 if (!thread_number) {
1120 log_err("No jobs defined(s)\n");
1125 if (def_thread.o.gtod_offload) {
1127 fio_gtod_offload = 1;
1128 fio_gtod_cpu = def_thread.o.gtod_cpu;