10 #include <sys/types.h>
16 #define DEF_TIMEOUT (0)
17 #define DEF_RATE_CYCLE (1000)
18 #define DEF_ODIRECT (1)
19 #define DEF_IO_ENGINE (FIO_SYNCIO)
20 #define DEF_IO_ENGINE_NAME "sync"
21 #define DEF_SEQUENTIAL (1)
22 #define DEF_RAND_REPEAT (1)
23 #define DEF_OVERWRITE (1)
24 #define DEF_CREATE (1)
25 #define DEF_INVALIDATE (1)
26 #define DEF_SYNCIO (0)
27 #define DEF_RANDSEED (0xb1899bedUL)
28 #define DEF_BWAVGTIME (500)
29 #define DEF_CREATE_SER (1)
30 #define DEF_CREATE_FSYNC (1)
32 #define DEF_VERIFY (0)
33 #define DEF_STONEWALL (0)
34 #define DEF_NUMJOBS (1)
35 #define DEF_USE_THREAD (0)
36 #define DEF_FILE_SIZE (1024 * 1024 * 1024UL)
37 #define DEF_ZONE_SIZE (0)
38 #define DEF_ZONE_SKIP (0)
40 static char fio_version_string[] = "fio 1.3";
42 static int repeatable = DEF_RAND_REPEAT;
43 static char *ini_file;
44 static int max_jobs = MAX_JOBS;
46 struct thread_data def_thread;
47 struct thread_data *threads = NULL;
50 int write_lat_log = 0;
52 int exitall_on_terminate = 0;
54 static int setup_rate(struct thread_data *td)
61 if (td->rate < td->ratemin) {
62 fprintf(stderr, "min rate larger than nominal rate\n");
66 nr_reads_per_sec = (td->rate * 1024) / td->min_bs;
67 td->rate_usec_cycle = 1000000 / nr_reads_per_sec;
68 td->rate_pending_usleep = 0;
72 static void setup_log(struct io_log **log)
74 struct io_log *l = malloc(sizeof(*l));
77 l->max_samples = 1024;
78 l->log = malloc(l->max_samples * sizeof(struct io_sample));
82 void finish_log(struct thread_data *td, struct io_log *log, const char *name)
88 snprintf(file_name, 200, "client%d_%s.log", td->thread_number, name);
89 f = fopen(file_name, "w");
95 for (i = 0; i < log->nr_samples; i++)
96 fprintf(f, "%lu, %lu, %u\n", log->log[i].time, log->log[i].val, log->log[i].ddir);
103 static struct thread_data *get_new_job(int global, struct thread_data *parent)
105 struct thread_data *td;
109 if (thread_number >= max_jobs)
112 td = &threads[thread_number++];
116 memset(td, 0, sizeof(*td));
119 td->thread_number = thread_number;
121 td->ddir = parent->ddir;
122 td->ioprio = parent->ioprio;
123 td->sequential = parent->sequential;
125 td->min_bs = parent->min_bs;
126 td->max_bs = parent->max_bs;
127 td->odirect = parent->odirect;
128 td->thinktime = parent->thinktime;
129 td->fsync_blocks = parent->fsync_blocks;
130 td->start_delay = parent->start_delay;
131 td->timeout = parent->timeout;
132 td->io_engine = parent->io_engine;
133 td->create_file = parent->create_file;
134 td->overwrite = parent->overwrite;
135 td->invalidate_cache = parent->invalidate_cache;
136 td->file_size = parent->file_size;
137 td->file_offset = parent->file_offset;
138 td->zone_size = parent->zone_size;
139 td->zone_skip = parent->zone_skip;
140 td->rate = parent->rate;
141 td->ratemin = parent->ratemin;
142 td->ratecycle = parent->ratecycle;
143 td->iodepth = parent->iodepth;
144 td->sync_io = parent->sync_io;
145 td->mem_type = parent->mem_type;
146 td->bw_avg_time = parent->bw_avg_time;
147 td->create_serialize = parent->create_serialize;
148 td->create_fsync = parent->create_fsync;
149 td->loops = parent->loops;
150 td->verify = parent->verify;
151 td->stonewall = parent->stonewall;
152 td->numjobs = parent->numjobs;
153 td->use_thread = parent->use_thread;
154 td->do_disk_util = parent->do_disk_util;
155 memcpy(&td->cpumask, &parent->cpumask, sizeof(td->cpumask));
156 strcpy(td->io_engine_name, parent->io_engine_name);
161 static void put_job(struct thread_data *td)
163 memset(&threads[td->thread_number - 1], 0, sizeof(*td));
167 static int add_job(struct thread_data *td, const char *jobname)
169 char *ddir_str[] = { "read", "write", "randread", "randwrite",
170 "rw", NULL, "randrw" };
174 #ifndef FIO_HAVE_LIBAIO
175 if (td->io_engine == FIO_LIBAIO) {
176 fprintf(stderr, "Linux libaio not available\n");
180 #ifndef FIO_HAVE_POSIXAIO
181 if (td->io_engine == FIO_POSIXAIO) {
182 fprintf(stderr, "posix aio not available\n");
188 * the def_thread is just for options, it's not a real job
190 if (td == &def_thread)
193 if (td->io_engine & FIO_SYNCIO)
201 * only really works for sequential io for now
203 if (td->zone_size && !td->sequential)
206 td->filetype = FIO_TYPE_FILE;
207 if (!stat(jobname, &sb)) {
208 if (S_ISBLK(sb.st_mode))
209 td->filetype = FIO_TYPE_BD;
210 else if (S_ISCHR(sb.st_mode))
211 td->filetype = FIO_TYPE_CHAR;
214 if (td->filetype == FIO_TYPE_FILE) {
215 if (td->directory[0] != '\0')
216 sprintf(td->file_name, "%s/%s.%d", td->directory, jobname, td->jobnum);
218 sprintf(td->file_name, "%s.%d", jobname, td->jobnum);
220 strncpy(td->file_name, jobname, sizeof(td->file_name) - 1);
222 sem_init(&td->mutex, 0, 0);
224 td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
225 td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
226 td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
228 if (td->min_bs == -1U)
230 if (td->max_bs == -1U)
232 if (td_read(td) && !td_rw(td))
235 if (td->stonewall && td->thread_number > 1)
238 td->groupid = groupid;
244 setup_log(&td->slat_log);
245 setup_log(&td->clat_log);
248 setup_log(&td->bw_log);
250 ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
251 printf("Client%d (g=%d): rw=%s, odir=%d, bs=%d-%d, rate=%d, ioengine=%s, iodepth=%d\n", td->thread_number, td->groupid, ddir_str[ddir], td->odirect, td->min_bs, td->max_bs, td->rate, td->io_engine_name, td->iodepth);
254 * recurse add identical jobs, clear numjobs and stonewall options
255 * as they don't apply to sub-jobs
257 numjobs = td->numjobs;
259 struct thread_data *td_new = get_new_job(0, td);
265 td_new->stonewall = 0;
266 td_new->jobnum = numjobs;
268 if (add_job(td_new, jobname))
277 int init_random_state(struct thread_data *td)
280 int fd, num_maps, blocks;
282 fd = open("/dev/random", O_RDONLY);
284 td_verror(td, errno);
288 if (read(fd, &seed, sizeof(seed)) < (int) sizeof(seed)) {
296 srand48_r(seed, &td->bsrange_state);
297 srand48_r(seed, &td->verify_state);
305 blocks = (td->io_size + td->min_bs - 1) / td->min_bs;
306 num_maps = blocks / BLOCKS_PER_MAP;
307 td->file_map = malloc(num_maps * sizeof(long));
308 td->num_maps = num_maps;
309 memset(td->file_map, 0, num_maps * sizeof(long));
311 srand48_r(seed, &td->random_state);
315 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
317 #ifdef FIO_HAVE_CPU_AFFINITY
322 for (i = 0; i < sizeof(int) * 8; i++) {
324 CPU_SET(i, &cpumask);
329 static unsigned long get_mult(char c)
340 return 1024 * 1024 * 1024;
347 * convert string after '=' into decimal value, noting any size suffix
349 static int str_cnv(char *p, unsigned long long *val)
354 str = strchr(p, '=');
361 *val = strtoul(str, NULL, 10);
362 if (*val == ULONG_MAX && errno == ERANGE)
365 *val *= get_mult(str[len - 2]);
369 static int check_strcnv(char *p, char *name, unsigned long long *val)
371 if (strncmp(p, name, strlen(name) - 1))
374 return str_cnv(p, val);
377 static void strip_blank_front(char **p)
385 static void strip_blank_end(char *p)
387 char *s = p + strlen(p) - 1;
389 while (isspace(*s) || iscntrl(*s))
395 typedef int (str_cb_fn)(struct thread_data *, char *);
397 static int check_str(char *p, char *name, str_cb_fn *cb, struct thread_data *td)
399 char *s = strstr(p, name);
409 strip_blank_front(&s);
413 static int check_strstore(char *p, char *name, char *dest)
415 char *s = strstr(p, name);
425 strip_blank_front(&s);
431 static int __check_range(char *str, unsigned long *val)
435 if (sscanf(str, "%lu%c", val, &suffix) == 2) {
436 *val *= get_mult(suffix);
440 if (sscanf(str, "%lu", val) == 1)
446 static int check_range(char *p, char *name, unsigned long *s, unsigned long *e)
454 str = strstr(p, name);
460 str = strchr(p, '=');
465 * 'p' now holds whatever is after the '=' sign
470 * terminate p1 at the '-' sign
479 if (!__check_range(p1, s) && !__check_range(p2, e))
485 static int check_int(char *p, char *name, unsigned int *val)
489 str = strstr(p, name);
493 str = strchr(p, '=');
499 if (sscanf(str, "%u", val) == 1)
505 static int check_strset(char *p, char *name)
507 return strncmp(p, name, strlen(name));
510 static int is_empty_or_comment(char *line)
514 for (i = 0; i < strlen(line); i++) {
517 if (!isspace(line[i]) && !iscntrl(line[i]))
524 static int str_rw_cb(struct thread_data *td, char *mem)
526 if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
527 td->ddir = DDIR_READ;
530 } else if (!strncmp(mem, "randread", 8)) {
531 td->ddir = DDIR_READ;
534 } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
535 td->ddir = DDIR_WRITE;
538 } else if (!strncmp(mem, "randwrite", 9)) {
539 td->ddir = DDIR_WRITE;
542 } else if (!strncmp(mem, "rw", 2)) {
547 } else if (!strncmp(mem, "randrw", 6)) {
554 fprintf(stderr, "bad data direction: %s\n", mem);
558 static int str_verify_cb(struct thread_data *td, char *mem)
560 if (!strncmp(mem, "0", 1)) {
561 td->verify = VERIFY_NONE;
563 } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
564 td->verify = VERIFY_MD5;
566 } else if (!strncmp(mem, "crc32", 5)) {
567 td->verify = VERIFY_CRC32;
571 fprintf(stderr, "bad verify type: %s\n", mem);
575 static int str_mem_cb(struct thread_data *td, char *mem)
577 if (!strncmp(mem, "malloc", 6)) {
578 td->mem_type = MEM_MALLOC;
580 } else if (!strncmp(mem, "shm", 3)) {
581 td->mem_type = MEM_SHM;
583 } else if (!strncmp(mem, "mmap", 4)) {
584 td->mem_type = MEM_MMAP;
588 fprintf(stderr, "bad mem type: %s\n", mem);
592 static int str_ioengine_cb(struct thread_data *td, char *str)
594 if (!strncmp(str, "linuxaio", 8) || !strncmp(str, "aio", 3) ||
595 !strncmp(str, "libaio", 6)) {
596 strcpy(td->io_engine_name, "libaio");
597 td->io_engine = FIO_LIBAIO;
599 } else if (!strncmp(str, "posixaio", 8)) {
600 strcpy(td->io_engine_name, "posixaio");
601 td->io_engine = FIO_POSIXAIO;
603 } else if (!strncmp(str, "sync", 4)) {
604 strcpy(td->io_engine_name, "sync");
605 td->io_engine = FIO_SYNCIO;
607 } else if (!strncmp(str, "mmap", 4)) {
608 strcpy(td->io_engine_name, "mmap");
609 td->io_engine = FIO_MMAPIO;
611 } else if (!strncmp(str, "sgio", 4)) {
612 strcpy(td->io_engine_name, "sgio");
613 td->io_engine = FIO_SGIO;
615 } else if (!strncmp(str, "splice", 6)) {
616 strcpy(td->io_engine_name, "splice");
617 td->io_engine = FIO_SPLICEIO;
621 fprintf(stderr, "bad ioengine type: %s\n", str);
625 static int str_iolog_cb(struct thread_data *td, char *file)
627 strncpy(td->iolog_file, file, sizeof(td->iolog_file) - 1);
632 int parse_jobs_ini(char *file)
634 unsigned int prioclass, prio, cpu, global;
635 unsigned long long ull;
636 unsigned long ul1, ul2;
637 struct thread_data *td;
643 f = fopen(file, "r");
645 perror("fopen job file");
649 string = malloc(4096);
652 while ((p = fgets(string, 4096, f)) != NULL) {
653 if (is_empty_or_comment(p))
655 if (sscanf(p, "[%s]", name) != 1)
658 global = !strncmp(name, "global", 6);
660 name[strlen(name) - 1] = '\0';
662 td = get_new_job(global, &def_thread);
667 while ((p = fgets(string, 4096, f)) != NULL) {
668 if (is_empty_or_comment(p))
674 if (!check_int(p, "prio", &prio)) {
675 #ifndef FIO_HAVE_IOPRIO
676 fprintf(stderr, "io priorities not available\n");
683 if (!check_int(p, "prioclass", &prioclass)) {
684 #ifndef FIO_HAVE_IOPRIO
685 fprintf(stderr, "io priorities not available\n");
688 td->ioprio |= prioclass << IOPRIO_CLASS_SHIFT;
692 if (!check_int(p, "direct", &td->odirect)) {
696 if (!check_int(p, "rate", &td->rate)) {
700 if (!check_int(p, "ratemin", &td->ratemin)) {
704 if (!check_int(p, "ratecycle", &td->ratecycle)) {
708 if (!check_int(p, "thinktime", &td->thinktime)) {
712 if (!check_int(p, "cpumask", &cpu)) {
713 #ifndef FIO_HAVE_CPU_AFFINITY
714 fprintf(stderr, "cpu affinity not available\n");
717 fill_cpu_mask(td->cpumask, cpu);
721 if (!check_int(p, "fsync", &td->fsync_blocks)) {
726 if (!check_int(p, "startdelay", &td->start_delay)) {
730 if (!check_int(p, "timeout", &td->timeout)) {
734 if (!check_int(p, "invalidate",&td->invalidate_cache)) {
738 if (!check_int(p, "iodepth", &td->iodepth)) {
742 if (!check_int(p, "sync", &td->sync_io)) {
746 if (!check_int(p, "bwavgtime", &td->bw_avg_time)) {
750 if (!check_int(p, "create_serialize", &td->create_serialize)) {
754 if (!check_int(p, "create_fsync", &td->create_fsync)) {
758 if (!check_int(p, "end_fsync", &td->end_fsync)) {
762 if (!check_int(p, "loops", &td->loops)) {
766 if (!check_int(p, "numjobs", &td->numjobs)) {
770 if (!check_int(p, "overwrite", &td->overwrite)) {
774 if (!check_range(p, "bsrange", &ul1, &ul2)) {
785 if (!check_strcnv(p, "bs", &ull)) {
790 if (!check_strcnv(p, "size", &td->file_size)) {
794 if (!check_strcnv(p, "offset", &td->file_offset)) {
798 if (!check_strcnv(p, "zonesize", &td->zone_size)) {
802 if (!check_strcnv(p, "zoneskip", &td->zone_skip)) {
806 if (!check_strstore(p, "directory", td->directory)) {
810 if (!check_str(p, "mem", str_mem_cb, td)) {
814 if (!check_str(p, "verify", str_verify_cb, td)) {
818 if (!check_str(p, "rw", str_rw_cb, td)) {
822 if (!check_str(p, "ioengine", str_ioengine_cb, td)) {
826 if (!check_strset(p, "create")) {
831 if (!check_strset(p, "exitall")) {
832 exitall_on_terminate = 1;
836 if (!check_strset(p, "stonewall")) {
841 if (!check_strset(p, "thread")) {
846 if (!check_str(p, "iolog", str_iolog_cb, td)) {
852 printf("Client%d: bad option %s\n",td->thread_number,p);
857 if (add_job(td, name))
867 static int fill_def_thread(void)
869 memset(&def_thread, 0, sizeof(def_thread));
871 if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
872 perror("sched_getaffinity");
879 def_thread.ddir = DDIR_READ;
880 def_thread.iomix = 0;
881 def_thread.bs = DEF_BS;
882 def_thread.min_bs = -1;
883 def_thread.max_bs = -1;
884 def_thread.io_engine = DEF_IO_ENGINE;
885 strcpy(def_thread.io_engine_name, DEF_IO_ENGINE_NAME);
886 def_thread.odirect = DEF_ODIRECT;
887 def_thread.ratecycle = DEF_RATE_CYCLE;
888 def_thread.sequential = DEF_SEQUENTIAL;
889 def_thread.timeout = DEF_TIMEOUT;
890 def_thread.create_file = DEF_CREATE;
891 def_thread.overwrite = DEF_OVERWRITE;
892 def_thread.invalidate_cache = DEF_INVALIDATE;
893 def_thread.sync_io = DEF_SYNCIO;
894 def_thread.mem_type = MEM_MALLOC;
895 def_thread.bw_avg_time = DEF_BWAVGTIME;
896 def_thread.create_serialize = DEF_CREATE_SER;
897 def_thread.create_fsync = DEF_CREATE_FSYNC;
898 def_thread.loops = DEF_LOOPS;
899 def_thread.verify = DEF_VERIFY;
900 def_thread.stonewall = DEF_STONEWALL;
901 def_thread.numjobs = DEF_NUMJOBS;
902 def_thread.use_thread = DEF_USE_THREAD;
903 #ifdef FIO_HAVE_DISK_UTIL
904 def_thread.do_disk_util = 1;
910 static void usage(char *name)
912 printf("%s\n", fio_version_string);
913 printf("\t-s IO is sequential\n");
914 printf("\t-b Block size in KiB for each IO\n");
915 printf("\t-t Runtime in seconds\n");
916 printf("\t-R Exit all threads on failure to meet rate goal\n");
917 printf("\t-o Use O_DIRECT\n");
918 printf("\t-l Generate per-job latency logs\n");
919 printf("\t-w Generate per-job bandwidth logs\n");
920 printf("\t-f Job file (Required)\n");
921 printf("\t-v Print version info and exit\n");
924 static void parse_cmd_line(int argc, char *argv[])
928 while ((c = getopt(argc, argv, "s:b:t:r:R:o:f:lwvh")) != EOF) {
931 def_thread.sequential = !!atoi(optarg);
934 def_thread.bs = atoi(optarg);
935 def_thread.bs <<= 10;
936 if (!def_thread.bs) {
937 printf("bad block size\n");
938 def_thread.bs = DEF_BS;
942 def_thread.timeout = atoi(optarg);
945 repeatable = !!atoi(optarg);
948 rate_quit = !!atoi(optarg);
951 def_thread.odirect = !!atoi(optarg);
954 ini_file = strdup(optarg);
966 printf("%s\n", fio_version_string);
971 if (!ini_file && argc > 1 && argv[argc - 1][0] != '-')
972 ini_file = strdup(argv[argc - 1]);
975 static void free_shm(void)
977 struct shmid_ds sbuf;
982 shmctl(shm_id, IPC_RMID, &sbuf);
986 static int setup_thread_area(void)
989 * 1024 is too much on some machines, scale max_jobs if
990 * we get a failure that looks like too large a shm segment
993 int s = max_jobs * sizeof(struct thread_data);
995 shm_id = shmget(0, s, IPC_CREAT | 0600);
998 if (errno != EINVAL) {
1009 threads = shmat(shm_id, NULL, 0);
1010 if (threads == (void *) -1) {
1019 int parse_options(int argc, char *argv[])
1021 if (setup_thread_area())
1023 if (fill_def_thread())
1026 parse_cmd_line(argc, argv);
1029 printf("Need job file\n");
1034 if (parse_jobs_ini(ini_file)) {