2 * This file contains job initialization and setup functions.
14 #include <sys/types.h>
23 static char fio_version_string[] = "fio 1.28";
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, enum fio_ddir ddir)
180 unsigned int bs = td->o.min_bs[ddir];
181 unsigned long long rate;
182 unsigned long ios_per_msec;
184 if (td->o.rate[ddir]) {
185 rate = td->o.rate[ddir];
186 ios_per_msec = (rate * 1000LL) / bs;
188 ios_per_msec = td->o.rate_iops[ddir] * 1000UL;
191 log_err("rate lower than supported\n");
195 td->rate_usec_cycle[ddir] = 1000000000ULL / ios_per_msec;
196 td->rate_pending_usleep[ddir] = 0;
200 static int setup_rate(struct thread_data *td)
204 if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
205 ret = __setup_rate(td, DDIR_READ);
206 if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
207 ret |= __setup_rate(td, DDIR_WRITE);
212 static int fixed_block_size(struct thread_options *o)
214 return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
215 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
216 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
220 * Lazy way of fixing up options that depend on each other. We could also
221 * define option callback handlers, but this is easier.
223 static int fixup_options(struct thread_data *td)
225 struct thread_options *o = &td->o;
227 #ifndef FIO_HAVE_PSHARED_MUTEX
228 if (!td->o.use_thread) {
229 log_info("fio: this platform does not support process shared"
230 " mutexes, forcing use of threads. Use the 'thread'"
231 " option to get rid of this warning.\n");
232 td->o.use_thread = 1;
236 #ifndef FIO_HAVE_CPU_AFFINITY
237 if (td->o.gtod_cpu) {
238 log_err("fio: platform must support CPU affinity for"
239 "gettimeofday() offloading\n");
244 if (read_only && td_write(td)) {
245 log_err("fio: job <%s> has write bit set, but fio is in"
246 " read-only mode\n", td->o.name);
250 if (o->write_iolog_file && o->read_iolog_file) {
251 log_err("fio: read iolog overrides write_iolog\n");
252 free(o->write_iolog_file);
253 o->write_iolog_file = NULL;
257 * only really works for sequential io for now, and with 1 file
259 if (o->zone_size && td_random(td) && o->open_files == 1)
263 * Reads can do overwrites, we always need to pre-create the file
265 if (td_read(td) || td_rw(td))
268 if (!o->min_bs[DDIR_READ])
269 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
270 if (!o->max_bs[DDIR_READ])
271 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
272 if (!o->min_bs[DDIR_WRITE])
273 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
274 if (!o->max_bs[DDIR_WRITE])
275 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
277 o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
280 * For random IO, allow blockalign offset other than min_bs.
282 if (!o->ba[DDIR_READ] || !td_random(td))
283 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
284 if (!o->ba[DDIR_WRITE] || !td_random(td))
285 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
287 if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
288 o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
289 !td->o.norandommap) {
290 log_err("fio: Any use of blockalign= turns off randommap\n");
291 td->o.norandommap = 1;
294 if (!o->file_size_high)
295 o->file_size_high = o->file_size_low;
297 if (o->norandommap && o->verify != VERIFY_NONE
298 && !fixed_block_size(o)) {
299 log_err("fio: norandommap given for variable block sizes, "
300 "verify disabled\n");
301 o->verify = VERIFY_NONE;
303 if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
304 log_err("fio: bs_unaligned may not work with raw io\n");
307 * thinktime_spin must be less than thinktime
309 if (o->thinktime_spin > o->thinktime)
310 o->thinktime_spin = o->thinktime;
313 * The low water mark cannot be bigger than the iodepth
315 if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
317 * syslet work around - if the workload is sequential,
318 * we want to let the queue drain all the way down to
319 * avoid seeking between async threads
321 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
324 o->iodepth_low = o->iodepth;
328 * If batch number isn't set, default to the same as iodepth
330 if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
331 o->iodepth_batch = o->iodepth;
333 if (o->nr_files > td->files_index)
334 o->nr_files = td->files_index;
336 if (o->open_files > o->nr_files || !o->open_files)
337 o->open_files = o->nr_files;
339 if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
340 ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
341 o->rate_iops_min[1]))) {
342 log_err("fio: rate and rate_iops are mutually exclusive\n");
345 if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
346 (o->rate_iops[0] < o->rate_iops_min[0]) ||
347 (o->rate_iops[1] < o->rate_iops_min[1])) {
348 log_err("fio: minimum rate exceeds rate\n");
352 if (!o->timeout && o->time_based) {
353 log_err("fio: time_based requires a runtime/timeout setting\n");
357 if (o->fill_device && !o->size)
360 if (td_rw(td) && td->o.verify != VERIFY_NONE)
361 log_info("fio: mixed read/write workload with verify. May not "
362 "work as expected, unless you pre-populated the file\n");
364 if (td->o.verify != VERIFY_NONE)
365 td->o.refill_buffers = 1;
368 td->o.invalidate_cache = 0;
370 if (td->o.mem_align) {
371 if (td->o.odirect && !is_power_of_2(td->o.mem_align)) {
372 log_err("fio: given IO mem alignment conflicts with"
373 " direct=1. Resetting.\n");
374 td->o.mem_align = page_mask;
382 * This function leaks the buffer
384 static char *to_kmg(unsigned int val)
386 char *buf = malloc(32);
387 char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
398 snprintf(buf, 31, "%u%c", val, *p);
402 /* External engines are specified by "external:name.o") */
403 static const char *get_engine_name(const char *str)
405 char *p = strstr(str, ":");
411 strip_blank_front(&p);
416 static int exists_and_not_file(const char *filename)
420 if (lstat(filename, &sb) == -1)
423 if (S_ISREG(sb.st_mode))
429 void td_fill_rand_seeds(struct thread_data *td)
431 os_random_seed(td->rand_seeds[0], &td->bsrange_state);
432 os_random_seed(td->rand_seeds[1], &td->verify_state);
433 os_random_seed(td->rand_seeds[2], &td->rwmix_state);
435 if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
436 os_random_seed(td->rand_seeds[3], &td->next_file_state);
438 os_random_seed(td->rand_seeds[5], &td->file_size_state);
443 if (td->o.rand_repeatable)
444 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
446 os_random_seed(td->rand_seeds[4], &td->random_state);
450 * Initialize the various random states we need (random io, block size ranges,
451 * read/write mix, etc).
453 static int init_random_state(struct thread_data *td)
457 fd = open("/dev/urandom", O_RDONLY);
459 td_verror(td, errno, "open");
463 if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
464 (int) sizeof(td->rand_seeds)) {
465 td_verror(td, EIO, "read");
471 td_fill_rand_seeds(td);
476 * Adds a job to the list of things todo. Sanitizes the various options
477 * to make sure we don't have conflicts, and initializes various
480 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
482 const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
483 "randread", "randwrite", "randrw" };
486 char fname[PATH_MAX];
487 int numjobs, file_alloced;
490 * the def_thread is just for options, it's not a real job
492 if (td == &def_thread)
496 * if we are just dumping the output command line, don't add the job
503 engine = get_engine_name(td->o.ioengine);
504 td->io_ops = load_ioengine(td, engine);
506 log_err("fio: failed to load engine %s\n", engine);
510 if (td->o.use_thread)
516 td->io_ops->flags |= FIO_RAWIO;
519 if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
522 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
523 add_file(td, jobname);
525 for (i = 0; i < td->o.nr_files; i++) {
526 sprintf(fname, "%s.%d.%d", jobname,
527 td->thread_number, i);
533 if (fixup_options(td))
536 if (td->io_ops->flags & FIO_DISKLESSIO) {
539 for_each_file(td, f, i)
540 f->real_file_size = -1ULL;
543 td->mutex = fio_mutex_init(0);
545 td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
546 td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
547 td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
548 td->ddir_nr = td->o.ddir_nr;
550 if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
551 && prev_group_jobs) {
556 td->groupid = groupid;
559 if (init_random_state(td))
565 if (td->o.write_lat_log) {
566 setup_log(&td->ts.slat_log);
567 setup_log(&td->ts.clat_log);
569 if (td->o.write_bw_log)
570 setup_log(&td->ts.bw_log);
573 td->o.name = strdup(jobname);
577 if (!strcmp(td->io_ops->name, "cpuio")) {
578 log_info("%s: ioengine=cpu, cpuload=%u,"
579 " cpucycle=%u\n", td->o.name,
583 char *c1, *c2, *c3, *c4;
585 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
586 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
587 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
588 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
590 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
591 " ioengine=%s, iodepth=%u\n",
592 td->o.name, td->groupid,
593 ddir_str[td->o.td_ddir],
603 } else if (job_add_num == 1)
608 * recurse add identical jobs, clear numjobs and stonewall options
609 * as they don't apply to sub-jobs
611 numjobs = td->o.numjobs;
613 struct thread_data *td_new = get_new_job(0, td);
618 td_new->o.numjobs = 1;
619 td_new->o.stonewall = 0;
620 td_new->o.new_group = 0;
623 td_new->o.filename = NULL;
624 td_new->files_index = 0;
625 td_new->files_size = 0;
626 td_new->files = NULL;
629 job_add_num = numjobs - 1;
631 if (add_job(td_new, jobname, job_add_num))
641 static int skip_this_section(const char *name)
645 if (!strncmp(name, "global", 6))
648 return strcmp(job_section, name);
651 static int is_empty_or_comment(char *line)
655 for (i = 0; i < strlen(line); i++) {
660 if (!isspace(line[i]) && !iscntrl(line[i]))
668 * This is our [ini] type file parser.
670 static int parse_jobs_ini(char *file, int stonewall_flag)
673 struct thread_data *td;
677 int ret = 0, stonewall;
682 int i, alloc_opts, num_opts;
684 if (!strcmp(file, "-"))
687 f = fopen(file, "r");
690 perror("fopen job file");
694 string = malloc(4096);
697 * it's really 256 + small bit, 280 should suffice
700 memset(name, 0, 280);
703 opts = malloc(sizeof(char *) * alloc_opts);
706 stonewall = stonewall_flag;
709 * if skip_fgets is set, we already have loaded a line we
713 p = fgets(string, 4095, f);
719 strip_blank_front(&p);
722 if (is_empty_or_comment(p))
724 if (sscanf(p, "[%255s]", name) != 1) {
727 log_err("fio: option <%s> outside of [] job section\n",
732 if (skip_this_section(name)) {
738 global = !strncmp(name, "global", 6);
740 name[strlen(name) - 1] = '\0';
746 log_info("--name=%s ", name);
750 td = get_new_job(global, &def_thread);
757 * Seperate multiple job files by a stonewall
759 if (!global && stonewall) {
760 td->o.stonewall = stonewall;
765 memset(opts, 0, alloc_opts * sizeof(char *));
767 while ((p = fgets(string, 4096, f)) != NULL) {
768 if (is_empty_or_comment(p))
771 strip_blank_front(&p);
774 * new section, break out and make sure we don't
775 * fgets() a new line at the top.
784 if (num_opts == alloc_opts) {
787 alloc_opts * sizeof(char *));
790 opts[num_opts] = strdup(p);
794 ret = fio_options_parse(td, opts, num_opts);
797 for (i = 0; i < num_opts; i++)
798 log_info("--%s ", opts[i]);
800 ret = add_job(td, name, 0);
802 log_err("fio: job %s dropped\n", name);
806 for (i = 0; i < num_opts; i++)
814 for (i = 0; i < num_opts; i++)
825 static int fill_def_thread(void)
827 memset(&def_thread, 0, sizeof(def_thread));
829 fio_getaffinity(getpid(), &def_thread.o.cpumask);
832 * fill default options
834 fio_fill_default_options(&def_thread);
836 def_thread.o.timeout = def_timeout;
837 def_thread.o.write_bw_log = write_bw_log;
838 def_thread.o.write_lat_log = write_lat_log;
843 static void free_shm(void)
845 struct shmid_ds sbuf;
852 fio_debug_jobp = NULL;
854 shmctl(shm_id, IPC_RMID, &sbuf);
861 * The thread area is shared between the main process and the job
862 * threads/processes. So setup a shared memory segment that will hold
863 * all the job info. We use the end of the region for keeping track of
864 * open files across jobs, for file sharing.
866 static int setup_thread_area(void)
871 * 1024 is too much on some machines, scale max_jobs if
872 * we get a failure that looks like too large a shm segment
875 size_t size = max_jobs * sizeof(struct thread_data);
877 size += file_hash_size;
878 size += sizeof(unsigned int);
880 shm_id = shmget(0, size, IPC_CREAT | 0600);
883 if (errno != EINVAL) {
894 threads = shmat(shm_id, NULL, 0);
895 if (threads == (void *) -1) {
900 memset(threads, 0, max_jobs * sizeof(struct thread_data));
901 hash = (void *) threads + max_jobs * sizeof(struct thread_data);
902 fio_debug_jobp = (void *) hash + file_hash_size;
903 *fio_debug_jobp = -1;
904 file_hash_init(hash);
909 static void usage(const char *name)
911 printf("%s\n", fio_version_string);
912 printf("%s [options] [job options] <job file(s)>\n", name);
913 printf("\t--debug=options\tEnable debug logging\n");
914 printf("\t--output\tWrite output to file\n");
915 printf("\t--timeout\tRuntime in seconds\n");
916 printf("\t--latency-log\tGenerate per-job latency logs\n");
917 printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
918 printf("\t--minimal\tMinimal (terse) output\n");
919 printf("\t--version\tPrint version info and exit\n");
920 printf("\t--help\t\tPrint this page\n");
921 printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
923 printf("\t--showcmd\tTurn a job file into command line options\n");
924 printf("\t--eta=when\tWhen ETA estimate should be printed\n");
925 printf("\t \tMay be \"always\", \"never\" or \"auto\"\n");
926 printf("\t--readonly\tTurn on safety read-only checks, preventing"
928 printf("\t--section=name\tOnly run specified section in job file\n");
929 printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
931 printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>\n");
935 struct debug_level debug_levels[] = {
936 { .name = "process", .shift = FD_PROCESS, },
937 { .name = "file", .shift = FD_FILE, },
938 { .name = "io", .shift = FD_IO, },
939 { .name = "mem", .shift = FD_MEM, },
940 { .name = "blktrace", .shift = FD_BLKTRACE },
941 { .name = "verify", .shift = FD_VERIFY },
942 { .name = "random", .shift = FD_RANDOM },
943 { .name = "parse", .shift = FD_PARSE },
944 { .name = "diskutil", .shift = FD_DISKUTIL },
945 { .name = "job", .shift = FD_JOB },
946 { .name = "mutex", .shift = FD_MUTEX },
950 static int set_debug(const char *string)
952 struct debug_level *dl;
953 char *p = (char *) string;
957 if (!strcmp(string, "?") || !strcmp(string, "help")) {
960 log_info("fio: dumping debug options:");
961 for (i = 0; debug_levels[i].name; i++) {
962 dl = &debug_levels[i];
963 log_info("%s,", dl->name);
969 while ((opt = strsep(&p, ",")) != NULL) {
972 if (!strncmp(opt, "all", 3)) {
973 log_info("fio: set all debug options\n");
978 for (i = 0; debug_levels[i].name; i++) {
979 dl = &debug_levels[i];
980 found = !strncmp(opt, dl->name, strlen(dl->name));
984 if (dl->shift == FD_JOB) {
985 opt = strchr(opt, ':');
987 log_err("fio: missing job number\n");
991 fio_debug_jobno = atoi(opt);
992 log_info("fio: set debug jobno %d\n",
995 log_info("fio: set debug option %s\n", opt);
996 fio_debug |= (1UL << dl->shift);
1002 log_err("fio: debug mask %s not found\n", opt);
1007 static int set_debug(const char *string)
1009 log_err("fio: debug tracing not included in build\n");
1014 static int parse_cmd_line(int argc, char *argv[])
1016 struct thread_data *td = NULL;
1017 int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1019 while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1022 smalloc_pool_size = atoi(optarg);
1025 def_timeout = atoi(optarg);
1034 f_out = fopen(optarg, "w+");
1036 perror("fopen output");
1048 exit(fio_show_option_help(optarg));
1056 printf("%s\n", fio_version_string);
1059 if (!strcmp("always", optarg))
1060 eta_print = FIO_ETA_ALWAYS;
1061 else if (!strcmp("never", optarg))
1062 eta_print = FIO_ETA_NEVER;
1065 if (set_debug(optarg))
1069 if (!strcmp(optarg, "global")) {
1070 log_err("fio: can't use global as only "
1078 job_section = strdup(optarg);
1080 case FIO_GETOPT_JOB: {
1081 const char *opt = l_opts[lidx].name;
1084 if (!strncmp(opt, "name", 4) && td) {
1085 ret = add_job(td, td->o.name ?: "fio", 0);
1093 int is_section = !strncmp(opt, "name", 4);
1096 if (!is_section || !strncmp(val, "global", 6))
1099 if (is_section && skip_this_section(val))
1102 td = get_new_job(global, &def_thread);
1107 ret = fio_cmd_option_parse(td, opt, val);
1122 ret = add_job(td, td->o.name ?: "fio", 0);
1127 while (optind < argc) {
1129 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1130 ini_file[ini_idx - 1] = strdup(argv[optind]);
1137 int parse_options(int argc, char *argv[])
1144 fio_options_dup_and_init(l_opts);
1146 if (setup_thread_area())
1148 if (fill_def_thread())
1151 job_files = parse_cmd_line(argc, argv);
1153 for (i = 0; i < job_files; i++) {
1154 if (fill_def_thread())
1156 if (parse_jobs_ini(ini_file[i], i))
1162 options_mem_free(&def_thread);
1164 if (!thread_number) {
1168 log_err("No jobs defined(s)\n\n");
1173 if (def_thread.o.gtod_offload) {
1175 fio_gtod_offload = 1;
1176 fio_gtod_cpu = def_thread.o.gtod_cpu;