2 * This file contains job initialization and setup functions.
13 #include <sys/types.h>
23 #define DEF_TIMEOUT (0)
24 #define DEF_RATE_CYCLE (1000)
25 #define DEF_ODIRECT (1)
26 #define DEF_IO_ENGINE (FIO_SYNCIO)
27 #define DEF_IO_ENGINE_NAME "sync"
28 #define DEF_SEQUENTIAL (1)
29 #define DEF_RAND_REPEAT (1)
30 #define DEF_OVERWRITE (1)
31 #define DEF_INVALIDATE (1)
32 #define DEF_SYNCIO (0)
33 #define DEF_RANDSEED (0xb1899bedUL)
34 #define DEF_BWAVGTIME (500)
35 #define DEF_CREATE_SER (1)
36 #define DEF_CREATE_FSYNC (1)
38 #define DEF_VERIFY (0)
39 #define DEF_STONEWALL (0)
40 #define DEF_NUMJOBS (1)
41 #define DEF_USE_THREAD (0)
42 #define DEF_FILE_SIZE (1024 * 1024 * 1024UL)
43 #define DEF_ZONE_SIZE (0)
44 #define DEF_ZONE_SKIP (0)
45 #define DEF_RWMIX_CYCLE (500)
46 #define DEF_RWMIX_READ (50)
48 #define DEF_NR_FILES (1)
49 #define DEF_UNLINK (0)
50 #define DEF_WRITE_BW_LOG (0)
51 #define DEF_WRITE_LAT_LOG (0)
53 #define td_var_offset(var) ((size_t) &((struct thread_data *)0)->var)
55 static int str_rw_cb(void *, char *);
56 static int str_ioengine_cb(void *, char *);
57 static int str_mem_cb(void *, char *);
58 static int str_verify_cb(void *, char *);
59 static int str_lockmem_cb(void *, unsigned long *);
60 static int str_prio_cb(void *, unsigned int *);
61 static int str_prioclass_cb(void *, unsigned int *);
62 static int str_exitall_cb(void);
63 static int str_cpumask_cb(void *, unsigned int *);
66 * Map of job/command line options
68 static struct fio_option options[] = {
71 .type = FIO_OPT_STR_STORE,
72 .off1 = td_var_offset(name),
76 .type = FIO_OPT_STR_STORE,
77 .off1 = td_var_offset(directory),
81 .type = FIO_OPT_STR_STORE,
82 .off1 = td_var_offset(filename),
92 .cb = str_ioengine_cb,
105 .name = "write_iolog",
106 .type = FIO_OPT_STR_STORE,
107 .off1 = td_var_offset(write_iolog),
111 .type = FIO_OPT_STR_STORE,
112 .off1 = td_var_offset(iolog),
115 .name = "exec_prerun",
116 .type = FIO_OPT_STR_STORE,
117 .off1 = td_var_offset(exec_prerun),
120 .name = "exec_postrun",
121 .type = FIO_OPT_STR_STORE,
122 .off1 = td_var_offset(exec_postrun),
124 #ifdef FIO_HAVE_IOSCHED_SWITCH
126 .name = "ioscheduler",
127 .type = FIO_OPT_STR_STORE,
128 .off1 = td_var_offset(ioscheduler),
133 .type = FIO_OPT_STR_VAL,
134 .off1 = td_var_offset(total_file_size),
138 .type = FIO_OPT_STR_VAL,
139 .off1 = td_var_offset(bs),
143 .type = FIO_OPT_STR_VAL,
144 .off1 = td_var_offset(start_offset),
148 .type = FIO_OPT_STR_VAL,
149 .off1 = td_var_offset(zone_size),
153 .type = FIO_OPT_STR_VAL,
154 .off1 = td_var_offset(zone_skip),
158 .type = FIO_OPT_STR_VAL,
159 .cb = str_lockmem_cb,
163 .type = FIO_OPT_RANGE,
164 .off1 = td_var_offset(min_bs),
165 .off2 = td_var_offset(max_bs),
170 .off1 = td_var_offset(nr_files),
175 .off1 = td_var_offset(iodepth),
180 .off1 = td_var_offset(fsync_blocks),
183 .name = "rwmixcycle",
185 .off1 = td_var_offset(rwmixcycle),
190 .off1 = td_var_offset(rwmixread),
194 .name = "rwmixwrite",
196 .off1 = td_var_offset(rwmixwrite),
202 .off1 = td_var_offset(nice),
204 #ifdef FIO_HAVE_IOPRIO
213 .cb = str_prioclass_cb,
219 .off1 = td_var_offset(thinktime)
224 .off1 = td_var_offset(rate)
229 .off1 = td_var_offset(ratemin)
234 .off1 = td_var_offset(ratecycle)
237 .name = "startdelay",
239 .off1 = td_var_offset(start_delay)
243 .type = FIO_OPT_STR_VAL_TIME,
244 .off1 = td_var_offset(timeout)
247 .name = "invalidate",
249 .off1 = td_var_offset(invalidate_cache)
254 .off1 = td_var_offset(sync_io)
259 .off1 = td_var_offset(bw_avg_time)
262 .name = "create_serialize",
264 .off1 = td_var_offset(create_serialize)
267 .name = "create_fsync",
269 .off1 = td_var_offset(create_fsync)
274 .off1 = td_var_offset(loops)
279 .off1 = td_var_offset(numjobs)
284 .off1 = td_var_offset(cpuload)
289 .off1 = td_var_offset(cpucycle)
294 .off1 = td_var_offset(odirect)
299 .off1 = td_var_offset(overwrite)
301 #ifdef FIO_HAVE_CPU_AFFINITY
305 .cb = str_cpumask_cb,
311 .off1 = td_var_offset(end_fsync)
315 .type = FIO_OPT_STR_SET,
316 .off1 = td_var_offset(unlink),
320 .type = FIO_OPT_STR_SET,
321 .cb = str_exitall_cb,
325 .type = FIO_OPT_STR_SET,
326 .off1 = td_var_offset(stonewall),
330 .type = FIO_OPT_STR_SET,
331 .off1 = td_var_offset(thread),
334 .name = "write_bw_log",
335 .type = FIO_OPT_STR_SET,
336 .off1 = td_var_offset(write_bw_log),
339 .name = "write_lat_log",
340 .type = FIO_OPT_STR_SET,
341 .off1 = td_var_offset(write_lat_log),
348 static int def_timeout = DEF_TIMEOUT;
350 static char fio_version_string[] = "fio 1.5";
352 static char **ini_file;
353 static int max_jobs = MAX_JOBS;
355 struct thread_data def_thread;
356 struct thread_data *threads = NULL;
359 int exitall_on_terminate = 0;
360 int terse_output = 0;
361 unsigned long long mlock_size = 0;
365 static int write_lat_log = DEF_WRITE_LAT_LOG;
366 static int write_bw_log = DEF_WRITE_BW_LOG;
369 * Return a free job structure.
371 static struct thread_data *get_new_job(int global, struct thread_data *parent)
373 struct thread_data *td;
377 if (thread_number >= max_jobs)
380 td = &threads[thread_number++];
384 td->thread_number = thread_number;
388 static void put_job(struct thread_data *td)
390 memset(&threads[td->thread_number - 1], 0, sizeof(*td));
394 static void fixup_options(struct thread_data *td)
401 if (!td->rwmixread && td->rwmixwrite)
402 td->rwmixread = 100 - td->rwmixwrite;
406 * Adds a job to the list of things todo. Sanitizes the various options
407 * to make sure we don't have conflicts, and initializes various
410 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
412 char *ddir_str[] = { "read", "write", "randread", "randwrite",
413 "rw", NULL, "randrw" };
415 int numjobs, ddir, i;
418 #ifndef FIO_HAVE_LIBAIO
419 if (td->io_engine == FIO_LIBAIO) {
420 log_err("Linux libaio not available\n");
424 #ifndef FIO_HAVE_POSIXAIO
425 if (td->io_engine == FIO_POSIXAIO) {
426 log_err("posix aio not available\n");
434 * the def_thread is just for options, it's not a real job
436 if (td == &def_thread)
440 * Set default io engine, if none set
443 td->io_ops = load_ioengine(td, DEF_IO_ENGINE_NAME);
445 log_err("default engine %s not there?\n", DEF_IO_ENGINE_NAME);
450 if (td->io_ops->flags & FIO_SYNCIO)
454 td->iodepth = td->nr_files;
458 * only really works for sequential io for now, and with 1 file
460 if (td->zone_size && !td->sequential && td->nr_files == 1)
464 * Reads can do overwrites, we always need to pre-create the file
466 if (td_read(td) || td_rw(td))
469 td->filetype = FIO_TYPE_FILE;
470 if (!stat(jobname, &sb)) {
471 if (S_ISBLK(sb.st_mode))
472 td->filetype = FIO_TYPE_BD;
473 else if (S_ISCHR(sb.st_mode))
474 td->filetype = FIO_TYPE_CHAR;
478 td->io_ops->flags |= FIO_RAWIO;
481 td->nr_uniq_files = 1;
483 td->nr_uniq_files = td->nr_files;
485 if (td->filetype == FIO_TYPE_FILE || td->filename) {
490 if (td->directory && td->directory[0] != '\0')
491 sprintf(tmp, "%s/", td->directory);
493 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
495 for_each_file(td, f, i) {
496 memset(f, 0, sizeof(*f));
500 sprintf(tmp + len, "%s", td->filename);
502 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
503 f->file_name = strdup(tmp);
507 td->files = malloc(sizeof(struct fio_file));
510 memset(f, 0, sizeof(*f));
512 f->file_name = strdup(jobname);
515 for_each_file(td, f, i) {
516 f->file_size = td->total_file_size / td->nr_files;
517 f->file_offset = td->start_offset;
520 fio_sem_init(&td->mutex, 0);
522 td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
523 td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
524 td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
526 if (td->min_bs == -1U)
528 if (td->max_bs == -1U)
530 if (td_read(td) && !td_rw(td))
533 if (td->stonewall && td->thread_number > 1)
536 td->groupid = groupid;
541 if (td->write_lat_log) {
542 setup_log(&td->slat_log);
543 setup_log(&td->clat_log);
545 if (td->write_bw_log)
546 setup_log(&td->bw_log);
548 if (td->name[0] == '\0')
549 snprintf(td->name, sizeof(td->name)-1, "client%d", td->thread_number);
551 ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
555 if (td->io_ops->flags & FIO_CPUIO)
556 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
558 fprintf(f_out, "%s: (g=%d): rw=%s, odir=%d, bs=%d-%d, rate=%d, ioengine=%s, iodepth=%d\n", td->name, td->groupid, ddir_str[ddir], td->odirect, td->min_bs, td->max_bs, td->rate, td->io_ops->name, td->iodepth);
559 } else if (job_add_num == 1)
560 fprintf(f_out, "...\n");
564 * recurse add identical jobs, clear numjobs and stonewall options
565 * as they don't apply to sub-jobs
567 numjobs = td->numjobs;
569 struct thread_data *td_new = get_new_job(0, td);
575 td_new->stonewall = 0;
576 job_add_num = numjobs - 1;
578 if (add_job(td_new, jobname, job_add_num))
588 * Initialize the various random states we need (random io, block size ranges,
589 * read/write mix, etc).
591 int init_random_state(struct thread_data *td)
593 unsigned long seeds[4];
594 int fd, num_maps, blocks, i;
597 fd = open("/dev/urandom", O_RDONLY);
599 td_verror(td, errno);
603 if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
611 os_random_seed(seeds[0], &td->bsrange_state);
612 os_random_seed(seeds[1], &td->verify_state);
613 os_random_seed(seeds[2], &td->rwmix_state);
618 if (td->rand_repeatable)
619 seeds[3] = DEF_RANDSEED;
621 for_each_file(td, f, i) {
622 blocks = (f->file_size + td->min_bs - 1) / td->min_bs;
623 num_maps = blocks / BLOCKS_PER_MAP;
624 f->file_map = malloc(num_maps * sizeof(long));
625 f->num_maps = num_maps;
626 memset(f->file_map, 0, num_maps * sizeof(long));
629 os_random_seed(seeds[3], &td->random_state);
633 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
635 #ifdef FIO_HAVE_CPU_AFFINITY
640 for (i = 0; i < sizeof(int) * 8; i++) {
642 CPU_SET(i, &cpumask);
647 static int is_empty_or_comment(char *line)
651 for (i = 0; i < strlen(line); i++) {
654 if (!isspace(line[i]) && !iscntrl(line[i]))
661 static int str_rw_cb(void *data, char *mem)
663 struct thread_data *td = data;
665 if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
666 td->ddir = DDIR_READ;
669 } else if (!strncmp(mem, "randread", 8)) {
670 td->ddir = DDIR_READ;
673 } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
674 td->ddir = DDIR_WRITE;
677 } else if (!strncmp(mem, "randwrite", 9)) {
678 td->ddir = DDIR_WRITE;
681 } else if (!strncmp(mem, "rw", 2)) {
686 } else if (!strncmp(mem, "randrw", 6)) {
693 log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
697 static int str_verify_cb(void *data, char *mem)
699 struct thread_data *td = data;
701 if (!strncmp(mem, "0", 1)) {
702 td->verify = VERIFY_NONE;
704 } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
705 td->verify = VERIFY_MD5;
707 } else if (!strncmp(mem, "crc32", 5)) {
708 td->verify = VERIFY_CRC32;
712 log_err("fio: verify types: md5, crc32\n");
716 static int str_mem_cb(void *data, char *mem)
718 struct thread_data *td = data;
720 if (!strncmp(mem, "malloc", 6)) {
721 td->mem_type = MEM_MALLOC;
723 } else if (!strncmp(mem, "shm", 3)) {
724 td->mem_type = MEM_SHM;
726 } else if (!strncmp(mem, "mmap", 4)) {
727 td->mem_type = MEM_MMAP;
731 log_err("fio: mem type: malloc, shm, mmap\n");
735 static int str_ioengine_cb(void *data, char *str)
737 struct thread_data *td = data;
739 td->io_ops = load_ioengine(td, str);
743 log_err("fio: ioengine: { linuxaio, aio, libaio }, posixaio, sync, mmap, sgio, splice, cpu\n");
747 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
753 static int str_prioclass_cb(void *data, unsigned int *val)
755 struct thread_data *td = data;
757 td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
761 static int str_prio_cb(void *data, unsigned int *val)
763 struct thread_data *td = data;
769 static int str_exitall_cb(void)
771 exitall_on_terminate = 1;
775 static int str_cpumask_cb(void *data, unsigned int *val)
777 struct thread_data *td = data;
779 fill_cpu_mask(td->cpumask, *val);
784 * This is our [ini] type file parser.
786 int parse_jobs_ini(char *file, int stonewall_flag)
789 struct thread_data *td;
790 char *string, *name, *tmpbuf;
794 int ret = 0, stonewall;
796 f = fopen(file, "r");
798 perror("fopen job file");
802 string = malloc(4096);
804 tmpbuf = malloc(4096);
806 stonewall = stonewall_flag;
807 while ((p = fgets(string, 4096, f)) != NULL) {
810 if (is_empty_or_comment(p))
812 if (sscanf(p, "[%s]", name) != 1)
815 global = !strncmp(name, "global", 6);
817 name[strlen(name) - 1] = '\0';
819 td = get_new_job(global, &def_thread);
826 * Seperate multiple job files by a stonewall
828 if (!global && stonewall) {
829 td->stonewall = stonewall;
834 while ((p = fgets(string, 4096, f)) != NULL) {
835 if (is_empty_or_comment(p))
840 strip_blank_front(&p);
846 * Don't break here, continue parsing options so we
847 * dump all the bad ones. Makes trial/error fixups
848 * easier on the user.
850 ret = parse_option(p, options, td);
855 ret = add_job(td, name, 0);
868 static int fill_def_thread(void)
870 memset(&def_thread, 0, sizeof(def_thread));
872 if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
873 perror("sched_getaffinity");
880 def_thread.ddir = DDIR_READ;
881 def_thread.iomix = 0;
882 def_thread.bs = DEF_BS;
883 def_thread.min_bs = -1;
884 def_thread.max_bs = -1;
885 def_thread.odirect = DEF_ODIRECT;
886 def_thread.ratecycle = DEF_RATE_CYCLE;
887 def_thread.sequential = DEF_SEQUENTIAL;
888 def_thread.timeout = def_timeout;
889 def_thread.overwrite = DEF_OVERWRITE;
890 def_thread.invalidate_cache = DEF_INVALIDATE;
891 def_thread.sync_io = DEF_SYNCIO;
892 def_thread.mem_type = MEM_MALLOC;
893 def_thread.bw_avg_time = DEF_BWAVGTIME;
894 def_thread.create_serialize = DEF_CREATE_SER;
895 def_thread.create_fsync = DEF_CREATE_FSYNC;
896 def_thread.loops = DEF_LOOPS;
897 def_thread.verify = DEF_VERIFY;
898 def_thread.stonewall = DEF_STONEWALL;
899 def_thread.numjobs = DEF_NUMJOBS;
900 def_thread.use_thread = DEF_USE_THREAD;
901 def_thread.rwmixcycle = DEF_RWMIX_CYCLE;
902 def_thread.rwmixread = DEF_RWMIX_READ;
903 def_thread.nice = DEF_NICE;
904 def_thread.rand_repeatable = DEF_RAND_REPEAT;
905 def_thread.nr_files = DEF_NR_FILES;
906 def_thread.unlink = DEF_UNLINK;
907 def_thread.write_bw_log = write_bw_log;
908 def_thread.write_lat_log = write_lat_log;
909 #ifdef FIO_HAVE_DISK_UTIL
910 def_thread.do_disk_util = 1;
916 static void usage(void)
918 printf("%s\n", fio_version_string);
919 printf("\t-o Write output to file\n");
920 printf("\t-t Runtime in seconds\n");
921 printf("\t-l Generate per-job latency logs\n");
922 printf("\t-w Generate per-job bandwidth logs\n");
923 printf("\t-m Minimal (terse) output\n");
924 printf("\t-v Print version info and exit\n");
927 static int parse_cmd_line(int argc, char *argv[])
929 int c, idx = 1, ini_idx = 0;
931 while ((c = getopt(argc, argv, "t:o:lwvhm")) != EOF) {
934 def_timeout = atoi(optarg);
946 f_out = fopen(optarg, "w+");
948 perror("fopen output");
962 printf("%s\n", fio_version_string);
969 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
970 ini_file[ini_idx - 1] = strdup(argv[idx]);
982 static void free_shm(void)
984 struct shmid_ds sbuf;
987 shmdt((void *) threads);
989 shmctl(shm_id, IPC_RMID, &sbuf);
994 * The thread area is shared between the main process and the job
995 * threads/processes. So setup a shared memory segment that will hold
998 static int setup_thread_area(void)
1001 * 1024 is too much on some machines, scale max_jobs if
1002 * we get a failure that looks like too large a shm segment
1005 size_t size = max_jobs * sizeof(struct thread_data);
1007 shm_id = shmget(0, size, IPC_CREAT | 0600);
1010 if (errno != EINVAL) {
1021 threads = shmat(shm_id, NULL, 0);
1022 if (threads == (void *) -1) {
1031 int parse_options(int argc, char *argv[])
1035 if (setup_thread_area())
1037 if (fill_def_thread())
1040 job_files = parse_cmd_line(argc, argv);
1042 log_err("Need job file(s)\n");
1047 for (i = 0; i < job_files; i++) {
1048 if (fill_def_thread())
1050 if (parse_jobs_ini(ini_file[i], i))