2 * This file contains job initialization and setup functions.
14 #include <sys/types.h>
20 static char fio_version_string[] = "fio 1.17.3";
22 #define FIO_RANDSEED (0xb1899bedUL)
24 static char **ini_file;
25 static int max_jobs = MAX_JOBS;
26 static int dump_cmdline;
28 struct thread_data def_thread;
29 struct thread_data *threads = NULL;
31 int exitall_on_terminate = 0;
34 unsigned long long mlock_size = 0;
41 static int def_timeout = 0;
42 static int write_lat_log = 0;
44 static int prev_group_jobs;
46 unsigned long fio_debug = 0;
49 * Command line options. These will contain the above, plus a few
50 * extra that only pertain to fio itself and not jobs.
52 static struct option long_options[FIO_NR_OPTIONS] = {
55 .has_arg = required_argument,
60 .has_arg = required_argument,
64 .name = "latency-log",
65 .has_arg = required_argument,
69 .name = "bandwidth-log",
70 .has_arg = required_argument,
75 .has_arg = optional_argument,
80 .has_arg = no_argument,
85 .has_arg = no_argument,
90 .has_arg = optional_argument,
95 .has_arg = no_argument,
100 .has_arg = no_argument,
105 .has_arg = required_argument,
110 .has_arg = required_argument,
129 * Return a free job structure.
131 static struct thread_data *get_new_job(int global, struct thread_data *parent)
133 struct thread_data *td;
137 if (thread_number >= max_jobs)
140 td = &threads[thread_number++];
143 dup_files(td, parent);
144 options_mem_dupe(td);
146 td->thread_number = thread_number;
150 static void put_job(struct thread_data *td)
152 if (td == &def_thread)
156 log_info("fio: %s\n", td->verror);
158 memset(&threads[td->thread_number - 1], 0, sizeof(*td));
162 static int setup_rate(struct thread_data *td)
164 unsigned long nr_reads_per_msec;
165 unsigned long long rate;
168 if (!td->o.rate && !td->o.rate_iops)
172 bs = td->o.rw_min_bs;
173 else if (td_read(td))
174 bs = td->o.min_bs[DDIR_READ];
176 bs = td->o.min_bs[DDIR_WRITE];
180 nr_reads_per_msec = (rate * 1024 * 1000LL) / bs;
182 nr_reads_per_msec = td->o.rate_iops * 1000UL;
184 if (!nr_reads_per_msec) {
185 log_err("rate lower than supported\n");
189 td->rate_usec_cycle = 1000000000ULL / nr_reads_per_msec;
190 td->rate_pending_usleep = 0;
195 * Lazy way of fixing up options that depend on each other. We could also
196 * define option callback handlers, but this is easier.
198 static int fixup_options(struct thread_data *td)
200 struct thread_options *o = &td->o;
202 if (read_only && td_write(td)) {
203 log_err("fio: job <%s> has write bit set, but fio is in read-only mode\n", td->o.name);
207 if (o->rwmix[DDIR_READ] + o->rwmix[DDIR_WRITE] > 100)
208 o->rwmix[DDIR_WRITE] = 100 - o->rwmix[DDIR_READ];
210 if (o->write_iolog_file && o->read_iolog_file) {
211 log_err("fio: read iolog overrides write_iolog\n");
212 free(o->write_iolog_file);
213 o->write_iolog_file = NULL;
217 * only really works for sequential io for now, and with 1 file
219 if (o->zone_size && td_random(td) && o->open_files == 1)
223 * Reads can do overwrites, we always need to pre-create the file
225 if (td_read(td) || td_rw(td))
228 if (!o->min_bs[DDIR_READ])
229 o->min_bs[DDIR_READ]= o->bs[DDIR_READ];
230 if (!o->max_bs[DDIR_READ])
231 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
232 if (!o->min_bs[DDIR_WRITE])
233 o->min_bs[DDIR_WRITE]= o->bs[DDIR_WRITE];
234 if (!o->max_bs[DDIR_WRITE])
235 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
237 o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
239 if (!o->file_size_high)
240 o->file_size_high = o->file_size_low;
242 if (o->norandommap && o->verify != VERIFY_NONE) {
243 log_err("fio: norandommap given, verify disabled\n");
244 o->verify = VERIFY_NONE;
246 if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
247 log_err("fio: bs_unaligned may not work with raw io\n");
250 * thinktime_spin must be less than thinktime
252 if (o->thinktime_spin > o->thinktime)
253 o->thinktime_spin = o->thinktime;
256 * The low water mark cannot be bigger than the iodepth
258 if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
260 * syslet work around - if the workload is sequential,
261 * we want to let the queue drain all the way down to
262 * avoid seeking between async threads
264 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
267 o->iodepth_low = o->iodepth;
271 * If batch number isn't set, default to the same as iodepth
273 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
274 o->iodepth_batch = o->iodepth;
276 if (o->nr_files > td->files_index)
277 o->nr_files = td->files_index;
279 if (o->open_files > o->nr_files || !o->open_files)
280 o->open_files = o->nr_files;
282 if ((o->rate && o->rate_iops) || (o->ratemin && o->rate_iops_min)) {
283 log_err("fio: rate and rate_iops are mutually exclusive\n");
286 if ((o->rate < o->ratemin) || (o->rate_iops < o->rate_iops_min)) {
287 log_err("fio: minimum rate exceeds rate\n");
291 if (!o->timeout && o->time_based) {
292 log_err("fio: time_based requires a runtime/timeout setting\n");
296 if (o->fill_device && !o->size)
297 o->size = ULONG_LONG_MAX;
303 * This function leaks the buffer
305 static char *to_kmg(unsigned int val)
307 char *buf = malloc(32);
308 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
319 snprintf(buf, 31, "%u%c", val, *p);
323 /* External engines are specified by "external:name.o") */
324 static const char *get_engine_name(const char *str)
326 char *p = strstr(str, ":");
332 strip_blank_front(&p);
337 static int exists_and_not_file(const char *filename)
341 if (lstat(filename, &sb) == -1)
344 if (S_ISREG(sb.st_mode))
351 * Initialize the various random states we need (random io, block size ranges,
352 * read/write mix, etc).
354 static int init_random_state(struct thread_data *td)
356 unsigned long seeds[6];
359 fd = open("/dev/urandom", O_RDONLY);
361 td_verror(td, errno, "open");
365 if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
366 td_verror(td, EIO, "read");
373 os_random_seed(seeds[0], &td->bsrange_state);
374 os_random_seed(seeds[1], &td->verify_state);
375 os_random_seed(seeds[2], &td->rwmix_state);
377 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
378 os_random_seed(seeds[3], &td->next_file_state);
380 os_random_seed(seeds[5], &td->file_size_state);
385 if (td->o.rand_repeatable)
386 seeds[4] = FIO_RANDSEED * td->thread_number;
388 os_random_seed(seeds[4], &td->random_state);
393 * Adds a job to the list of things todo. Sanitizes the various options
394 * to make sure we don't have conflicts, and initializes various
397 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
399 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
400 "randread", "randwrite", "randrw" };
403 char fname[PATH_MAX];
404 int numjobs, file_alloced;
407 * the def_thread is just for options, it's not a real job
409 if (td == &def_thread)
413 * if we are just dumping the output command line, don't add the job
420 engine = get_engine_name(td->o.ioengine);
421 td->io_ops = load_ioengine(td, engine);
423 log_err("fio: failed to load engine %s\n", engine);
427 if (td->o.use_thread)
433 td->io_ops->flags |= FIO_RAWIO;
436 if (!td->o.filename && !td->files_index) {
439 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
440 add_file(td, jobname);
442 for (i = 0; i < td->o.nr_files; i++) {
443 sprintf(fname, "%s.%d.%d", jobname, td->thread_number, i);
449 if (fixup_options(td))
452 if (td->io_ops->flags & FIO_DISKLESSIO) {
455 for_each_file(td, f, i)
456 f->real_file_size = -1ULL;
459 td->mutex = fio_sem_init(0);
461 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
462 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
463 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
464 td->ddir_nr = td->o.ddir_nr;
466 if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
467 && prev_group_jobs) {
472 td->groupid = groupid;
475 if (init_random_state(td))
481 if (td->o.write_lat_log) {
482 setup_log(&td->ts.slat_log);
483 setup_log(&td->ts.clat_log);
485 if (td->o.write_bw_log)
486 setup_log(&td->ts.bw_log);
489 td->o.name = strdup(jobname);
493 if (!strcmp(td->io_ops->name, "cpuio"))
494 log_info("%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->o.name, td->o.cpuload, td->o.cpucycle);
496 char *c1, *c2, *c3, *c4;
498 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
499 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
500 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
501 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
503 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->o.name, td->groupid, ddir_str[td->o.td_ddir], c1, c2, c3, c4, td->io_ops->name, td->o.iodepth);
510 } else if (job_add_num == 1)
515 * recurse add identical jobs, clear numjobs and stonewall options
516 * as they don't apply to sub-jobs
518 numjobs = td->o.numjobs;
520 struct thread_data *td_new = get_new_job(0, td);
525 td_new->o.numjobs = 1;
526 td_new->o.stonewall = 0;
527 td_new->o.new_group = 0;
530 td_new->o.filename = NULL;
531 td_new->files_index = 0;
532 td_new->files = NULL;
535 job_add_num = numjobs - 1;
537 if (add_job(td_new, jobname, job_add_num))
547 static int is_empty_or_comment(char *line)
551 for (i = 0; i < strlen(line); i++) {
556 if (!isspace(line[i]) && !iscntrl(line[i]))
564 * This is our [ini] type file parser.
566 static int parse_jobs_ini(char *file, int stonewall_flag)
569 struct thread_data *td;
573 int ret = 0, stonewall;
577 if (!strcmp(file, "-"))
580 f = fopen(file, "r");
583 perror("fopen job file");
587 string = malloc(4096);
590 * it's really 256 + small bit, 280 should suffice
593 memset(name, 0, 280);
595 stonewall = stonewall_flag;
598 * if skip_fgets is set, we already have loaded a line we
602 p = fgets(string, 4095, f);
608 strip_blank_front(&p);
611 if (is_empty_or_comment(p))
613 if (sscanf(p, "[%255s]", name) != 1) {
614 log_err("fio: option <%s> outside of [] job section\n", p);
618 global = !strncmp(name, "global", 6);
620 name[strlen(name) - 1] = '\0';
626 log_info("--name=%s ", name);
630 td = get_new_job(global, &def_thread);
637 * Seperate multiple job files by a stonewall
639 if (!global && stonewall) {
640 td->o.stonewall = stonewall;
644 while ((p = fgets(string, 4096, f)) != NULL) {
645 if (is_empty_or_comment(p))
648 strip_blank_front(&p);
651 * new section, break out and make sure we don't
652 * fgets() a new line at the top.
662 * Don't break here, continue parsing options so we
663 * dump all the bad ones. Makes trial/error fixups
664 * easier on the user.
666 ret |= fio_option_parse(td, p);
667 if (!ret && dump_cmdline)
668 log_info("--%s ", p);
672 ret = add_job(td, name, 0);
674 log_err("fio: job %s dropped\n", name);
689 static int fill_def_thread(void)
691 memset(&def_thread, 0, sizeof(def_thread));
693 fio_getaffinity(getpid(), &def_thread.o.cpumask);
696 * fill default options
698 fio_fill_default_options(&def_thread);
700 def_thread.o.timeout = def_timeout;
701 def_thread.o.write_bw_log = write_bw_log;
702 def_thread.o.write_lat_log = write_lat_log;
707 static void free_shm(void)
709 struct shmid_ds sbuf;
712 shmdt((void *) threads);
714 shmctl(shm_id, IPC_RMID, &sbuf);
719 * The thread area is shared between the main process and the job
720 * threads/processes. So setup a shared memory segment that will hold
723 static int setup_thread_area(void)
726 * 1024 is too much on some machines, scale max_jobs if
727 * we get a failure that looks like too large a shm segment
730 size_t size = max_jobs * sizeof(struct thread_data);
732 shm_id = shmget(0, size, IPC_CREAT | 0600);
735 if (errno != EINVAL) {
746 threads = shmat(shm_id, NULL, 0);
747 if (threads == (void *) -1) {
752 memset(threads, 0, max_jobs * sizeof(struct thread_data));
757 static void usage(const char *name)
759 printf("%s\n", fio_version_string);
760 printf("%s [options] [job options] <job file(s)>\n", name);
761 printf("\t--debug=options\tEnable debug logging\n");
762 printf("\t--output\tWrite output to file\n");
763 printf("\t--timeout\tRuntime in seconds\n");
764 printf("\t--latency-log\tGenerate per-job latency logs\n");
765 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
766 printf("\t--minimal\tMinimal (terse) output\n");
767 printf("\t--version\tPrint version info and exit\n");
768 printf("\t--help\t\tPrint this page\n");
769 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
770 printf("\t--showcmd\tTurn a job file into command line options\n");
771 printf("\t--eta=when\tWhen ETA estimate should be printed\n");
772 printf("\t \tMay be \"always\", \"never\" or \"auto\"\n");
776 struct debug_level debug_levels[] = {
777 { .name = "process", .shift = FD_PROCESS, },
778 { .name = "file", .shift = FD_FILE, },
779 { .name = "io", .shift = FD_IO, },
780 { .name = "mem", .shift = FD_MEM, },
781 { .name = "blktrace", .shift = FD_BLKTRACE },
782 { .name = "verify", .shift = FD_VERIFY },
786 static void set_debug(const char *string)
788 struct debug_level *dl;
789 char *p = (char *) string;
793 if (!strcmp(string, "?") || !strcmp(string, "help")) {
796 log_info("fio: dumping debug options:");
797 for (i = 0; debug_levels[i].name; i++) {
798 dl = &debug_levels[i];
799 log_info("%s,", dl->name);
803 } else if (!strcmp(string, "all")) {
808 while ((opt = strsep(&p, ",")) != NULL) {
811 for (i = 0; debug_levels[i].name; i++) {
812 dl = &debug_levels[i];
813 if (!strncmp(opt, dl->name, strlen(opt))) {
814 log_info("fio: set debug option %s\n", opt);
816 fio_debug |= (1UL << dl->shift);
822 log_err("fio: debug mask %s not found\n", opt);
826 static void set_debug(const char *string)
828 log_err("fio: debug tracing not included in build\n");
832 static int parse_cmd_line(int argc, char *argv[])
834 struct thread_data *td = NULL;
835 int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
837 while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
840 def_timeout = atoi(optarg);
849 f_out = fopen(optarg, "w+");
851 perror("fopen output");
863 exit(fio_show_option_help(optarg));
871 printf("%s\n", fio_version_string);
874 if (!strcmp("always", optarg))
875 eta_print = FIO_ETA_ALWAYS;
876 else if (!strcmp("never", optarg))
877 eta_print = FIO_ETA_NEVER;
882 case FIO_GETOPT_JOB: {
883 const char *opt = long_options[lidx].name;
886 if (!strncmp(opt, "name", 4) && td) {
887 ret = add_job(td, td->o.name ?: "fio", 0);
897 if (strncmp(opt, "name", 4) ||
898 !strncmp(val, "global", 6))
901 td = get_new_job(global, &def_thread);
906 ret = fio_cmd_option_parse(td, opt, val);
920 ret = add_job(td, td->o.name ?: "fio", 0);
926 while (optind < argc) {
928 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
929 ini_file[ini_idx - 1] = strdup(argv[optind]);
936 int parse_options(int argc, char *argv[])
943 fio_options_dup_and_init(long_options);
945 if (setup_thread_area())
947 if (fill_def_thread())
950 job_files = parse_cmd_line(argc, argv);
952 for (i = 0; i < job_files; i++) {
953 if (fill_def_thread())
955 if (parse_jobs_ini(ini_file[i], i))
961 options_mem_free(&def_thread);
963 if (!thread_number) {
967 log_err("No jobs defined(s)\n");