static void disk_util_timer_arm(void);
static void print_thread_status(void);
+extern unsigned long long mlock_size;
+
/*
* thread life cycle
*/
enum {
TD_NOT_CREATED = 0,
TD_CREATED,
+ TD_INITIALIZED,
TD_RUNNING,
TD_VERIFYING,
TD_EXITED,
static sem_t startup_sem;
#define TERMINATE_ALL (-1)
+#define JOB_START_TIMEOUT (5 * 1000)
static void terminate_threads(int group_id)
{
static void mark_random_map(struct thread_data *td, struct io_u *io_u)
{
- unsigned long block = io_u->offset / td->min_bs;
+ unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
unsigned int blocks = 0;
while (blocks < (io_u->buflen / td->min_bs)) {
long r;
if (!td->sequential) {
- unsigned long max_blocks = td->io_size / td->min_bs;
+ unsigned long long max_blocks = td->io_size / td->min_bs;
int loops = 50;
do {
lrand48_r(&td->random_state, &r);
- b = ((max_blocks - 1) * r / (RAND_MAX+1.0));
+ b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
rb = b + (td->file_offset / td->min_bs);
loops--;
} while (!random_map_free(td, rb) && loops);
memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
}
-unsigned int hweight32(unsigned int w)
-{
- unsigned int res = w - ((w >> 1) & 0x55555555);
-
- res = (res & 0x33333333) + ((res >> 2) & 0x33333333);
- res = (res + (res >> 4)) & 0x0F0F0F0F;
- res = res + (res >> 8);
-
- return (res + (res >> 16)) & 0x000000FF;
-}
-
-unsigned long hweight64(unsigned long long w)
+static int get_rw_ddir(struct thread_data *td)
{
-#if __WORDSIZE == 32
- return hweight32((unsigned int)(w >> 32)) + hweight32((unsigned int)w);
-#elif __WORDSIZE == 64
- unsigned long long v = w - ((w >> 1) & 0x5555555555555555ul);
+ if (td_rw(td)) {
+ struct timeval now;
+ unsigned long elapsed;
- v = (v & 0x3333333333333333ul) + ((v >> 2) & 0x3333333333333333ul);
- v = (v + (v >> 4)) & 0x0F0F0F0F0F0F0F0Ful;
- v = v + (v >> 8);
- v = v + (v >> 16);
+ gettimeofday(&now, NULL);
+ elapsed = mtime_since_now(&td->rwmix_switch);
- return (v + (v >> 32)) & 0x00000000000000FFul;
-#else
-#error __WORDSIZE not defined
-#endif
-}
+ /*
+ * Check if it's time to seed a new data direction.
+ */
+ if (elapsed >= td->rwmixcycle) {
+ unsigned long v;
+ long r;
-static int get_rw_ddir(struct thread_data *td)
-{
- /*
- * perhaps cheasy, but use the hamming weight of the position
- * as a randomizer for data direction.
- */
- if (td_rw(td))
- return hweight64(td->last_pos) & 1;
- else if (td_read(td))
+ lrand48_r(&td->random_state, &r);
+ v = 100UL * r / (unsigned long) (RAND_MAX + 1.0);
+ if (v < td->rwmixread)
+ td->rwmix_ddir = DDIR_READ;
+ else
+ td->rwmix_ddir = DDIR_WRITE;
+ memcpy(&td->rwmix_switch, &now, sizeof(now));
+ }
+ return td->rwmix_ddir;
+ } else if (td_read(td))
return DDIR_READ;
else
return DDIR_WRITE;
memcpy(io_u->buf, &hdr, sizeof(hdr));
}
-static int td_io_prep(struct thread_data *td, struct io_u *io_u, int read)
+static int td_io_prep(struct thread_data *td, struct io_u *io_u)
{
- if (read)
- io_u->ddir = DDIR_READ;
- else
- io_u->ddir = DDIR_WRITE;
-
if (td->io_prep && td->io_prep(td, io_u))
return 1;
td->cur_depth--;
}
+static int fill_io_u(struct thread_data *td, struct io_u *io_u)
+{
+ /*
+ * If using an iolog, grab next piece if any available.
+ */
+ if (td->iolog) {
+ struct io_piece *ipo;
+
+ if (list_empty(&td->io_log_list))
+ return 1;
+
+ ipo = list_entry(td->io_log_list.next, struct io_piece, list);
+ list_del(&ipo->list);
+ io_u->offset = ipo->offset;
+ io_u->buflen = ipo->len;
+ io_u->ddir = ipo->ddir;
+ free(ipo);
+ return 0;
+ }
+
+ /*
+ * No log, let the seq/rand engine retrieve the next position.
+ */
+ if (!get_next_offset(td, &io_u->offset)) {
+ io_u->buflen = get_next_buflen(td);
+
+ if (io_u->buflen) {
+ io_u->ddir = get_rw_ddir(td);
+ return 0;
+ }
+ }
+
+ return 1;
+}
+
#define queue_full(td) (list_empty(&(td)->io_u_freelist))
struct io_u *__get_io_u(struct thread_data *td)
td->last_pos += td->zone_skip;
}
- if (get_next_offset(td, &io_u->offset)) {
- put_io_u(td, io_u);
- return NULL;
- }
-
- io_u->buflen = get_next_buflen(td);
- if (!io_u->buflen) {
+ if (fill_io_u(td, io_u)) {
put_io_u(td, io_u);
return NULL;
}
return NULL;
}
- if (!td->sequential)
+ if (!td->iolog && !td->sequential)
mark_random_map(td, io_u);
td->last_pos += io_u->buflen;
if (td->verify != VERIFY_NONE)
populate_io_u(td, io_u);
- if (td_io_prep(td, io_u, get_rw_ddir(td))) {
+ if (td_io_prep(td, io_u)) {
put_io_u(td, io_u);
return NULL;
}
td->runstate = runstate;
}
-static int get_next_verify(struct thread_data *td,
- unsigned long long *offset, unsigned int *len)
+static int get_next_verify(struct thread_data *td, struct io_u *io_u)
{
struct io_piece *ipo;
ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
list_del(&ipo->list);
- *offset = ipo->offset;
- *len = ipo->len;
+ io_u->offset = ipo->offset;
+ io_u->buflen = ipo->len;
+ io_u->ddir = DDIR_READ;
free(ipo);
return 0;
}
list_add(&ipo->list, entry);
}
+static int init_iolog(struct thread_data *td)
+{
+ unsigned long long offset;
+ unsigned int bytes;
+ char *str, *p;
+ FILE *f;
+ int rw, i, reads, writes;
+
+ if (!td->iolog)
+ return 0;
+
+ f = fopen(td->iolog_file, "r");
+ if (!f) {
+ perror("fopen iolog");
+ return 1;
+ }
+
+ str = malloc(4096);
+ reads = writes = i = 0;
+ while ((p = fgets(str, 4096, f)) != NULL) {
+ struct io_piece *ipo;
+
+ if (sscanf(p, "%d,%llu,%u", &rw, &offset, &bytes) != 3) {
+ fprintf(stderr, "bad iolog: %s\n", p);
+ continue;
+ }
+ if (rw == DDIR_READ)
+ reads++;
+ else if (rw == DDIR_WRITE)
+ writes++;
+ else {
+ fprintf(stderr, "bad ddir: %d\n", rw);
+ continue;
+ }
+
+ ipo = malloc(sizeof(*ipo));
+ INIT_LIST_HEAD(&ipo->list);
+ ipo->offset = offset;
+ ipo->len = bytes;
+ if (bytes > td->max_bs)
+ td->max_bs = bytes;
+ ipo->ddir = rw;
+ list_add_tail(&ipo->list, &td->io_log_list);
+ i++;
+ }
+
+ free(str);
+ fclose(f);
+
+ if (!i)
+ return 1;
+
+ if (reads && !writes)
+ td->ddir = DDIR_READ;
+ else if (!reads && writes)
+ td->ddir = DDIR_READ;
+ else
+ td->iomix = 1;
+
+ return 0;
+}
+
static int sync_td(struct thread_data *td)
{
if (td->io_sync)
if (!io_u)
break;
- if (get_next_verify(td, &io_u->offset, &io_u->buflen)) {
+ if (get_next_verify(td, io_u)) {
put_io_u(td, io_u);
break;
}
- if (td_io_prep(td, io_u, 1)) {
+ if (td_io_prep(td, io_u)) {
put_io_u(td, io_u);
break;
}
if (td->cur_depth)
cleanup_pending_aio(td);
- if (should_fsync(td))
+ if (should_fsync(td) && td->end_fsync)
sync_td(td);
}
return fio_posixaio_init(td);
else if (td->io_engine == FIO_SGIO)
return fio_sgio_init(td);
+ else if (td->io_engine == FIO_SPLICEIO)
+ return fio_spliceio_init(td);
else {
fprintf(stderr, "bad io_engine %d\n", td->io_engine);
return 1;
}
}
- INIT_LIST_HEAD(&td->io_u_freelist);
- INIT_LIST_HEAD(&td->io_u_busylist);
- INIT_LIST_HEAD(&td->io_hist_list);
-
p = ALIGN(td->orig_buffer);
for (i = 0; i < max_units; i++) {
io_u = malloc(sizeof(*io_u));
return;
/*
- * for md/dm, there's no queue dir. we already have the right place
+ * If there's a ../queue/ directory there, we are inside a partition.
+ * Check if that is the case and jump back. For loop/md/dm etc we
+ * are already in the right spot.
*/
- sprintf(tmp, "%s/stat", foo);
- if (stat(tmp, &st)) {
- /*
- * if this is inside a partition dir, jump back to parent
- */
- sprintf(tmp, "%s/queue", foo);
+ sprintf(tmp, "%s/../queue", foo);
+ if (!stat(tmp, &st)) {
+ p = dirname(foo);
+ sprintf(tmp, "%s/queue", p);
if (stat(tmp, &st)) {
- p = dirname(foo);
- sprintf(tmp, "%s/queue", p);
- if (stat(tmp, &st)) {
- fprintf(stderr, "unknown sysfs layout\n");
- return;
- }
- sprintf(foo, "%s", p);
+ fprintf(stderr, "unknown sysfs layout\n");
+ return;
}
+ sprintf(foo, "%s", p);
}
disk_util_add(dev, foo);
static void *thread_main(void *data)
{
struct thread_data *td = data;
- int ret = 1;
if (!td->use_thread)
setsid();
td->pid = getpid();
+ INIT_LIST_HEAD(&td->io_u_freelist);
+ INIT_LIST_HEAD(&td->io_u_busylist);
+ INIT_LIST_HEAD(&td->io_hist_list);
+ INIT_LIST_HEAD(&td->io_log_list);
+
if (init_io_u(td))
goto err;
if (init_io(td))
goto err;
+ if (init_iolog(td))
+ goto err;
+
if (td->ioprio) {
if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
td_verror(td, errno);
}
}
+ if (nice(td->nice) < 0) {
+ td_verror(td, errno);
+ goto err;
+ }
+
+ if (init_random_state(td))
+ goto err;
+
+ td_set_runstate(td, TD_INITIALIZED);
sem_post(&startup_sem);
sem_wait(&td->mutex);
if (!td->create_serialize && setup_file(td))
goto err;
- if (init_random_state(td))
- goto err;
-
gettimeofday(&td->epoch, NULL);
while (td->loops--) {
do_io(td);
td->runtime[td->ddir] += mtime_since_now(&td->start);
- if (td_rw(td))
+ if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
update_rusage_stat(td);
break;
}
- ret = 0;
-
if (td->bw_log)
finish_log(td, td->bw_log, "bw");
if (td->slat_log)
munmap(td->mmap, td->file_size);
cleanup_io(td);
cleanup_io_u(td);
- if (ret) {
- sem_post(&startup_sem);
- sem_wait(&td->mutex);
- }
td_set_runstate(td, TD_EXITED);
return NULL;
int ddir)
{
char *ddir_str[] = { "read ", "write" };
- unsigned long min, max, bw;
+ unsigned long min, max;
+ unsigned long long bw;
double mean, dev;
if (!td->runtime[ddir])
return;
bw = td->io_bytes[ddir] / td->runtime[ddir];
- printf(" %s: io=%6lluMiB, bw=%6luKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
+ printf(" %s: io=%6lluMiB, bw=%6lluKiB/s, runt=%6lumsec\n", ddir_str[ddir], td->io_bytes[ddir] >> 20, bw, td->runtime[ddir]);
if (calc_lat(&td->slat_stat[ddir], &min, &max, &mean, &dev))
printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
printf("Client%d (groupid=%d): err=%2d:\n", td->thread_number, td->groupid, td->error);
show_ddir_status(td, rs, td->ddir);
- show_ddir_status(td, rs, td->ddir ^ 1);
+ if (td->io_bytes[td->ddir ^ 1])
+ show_ddir_status(td, rs, td->ddir ^ 1);
if (td->runtime[0] + td->runtime[1]) {
double runt = td->runtime[0] + td->runtime[1];
case TD_CREATED:
c = 'C';
break;
+ case TD_INITIALIZED:
+ c = 'I';
+ break;
case TD_NOT_CREATED:
c = 'P';
break;
if (td->timeout && eta_sec > (td->timeout - elapsed))
eta_sec = td->timeout - elapsed;
- } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED) {
+ } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED
+ || td->runstate == TD_INITIALIZED) {
int t_eta = 0, r_eta = 0;
/*
}
}
+static void fio_unpin_memory(void *pinned)
+{
+ if (pinned) {
+ if (munlock(pinned, mlock_size) < 0)
+ perror("munlock");
+ munmap(pinned, mlock_size);
+ }
+}
+
+static void *fio_pin_memory(void)
+{
+ long pagesize, pages;
+ void *ptr;
+
+ if (!mlock_size)
+ return NULL;
+
+ /*
+ * Don't allow mlock of more than real_mem-128MB
+ */
+ pagesize = sysconf(_SC_PAGESIZE);
+ pages = sysconf(_SC_PHYS_PAGES);
+ if (pages != -1 && pagesize != -1) {
+ unsigned long long real_mem = pages * pagesize;
+
+ if ((mlock_size + 128 * 1024 * 1024) > real_mem) {
+ mlock_size = real_mem - 128 * 1024 * 1024;
+ printf("fio: limiting mlocked memory to %lluMiB\n",
+ mlock_size >> 20);
+ }
+ }
+
+ ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
+ if (!ptr) {
+ perror("malloc locked mem");
+ return NULL;
+ }
+ if (mlock(ptr, mlock_size) < 0) {
+ munmap(ptr, mlock_size);
+ perror("mlock");
+ return NULL;
+ }
+
+ return ptr;
+}
+
static void run_threads(void)
{
struct thread_data *td;
unsigned long spent;
int i, todo, nr_running, m_rate, t_rate, nr_started;
+ void *mlocked_mem;
+
+ mlocked_mem = fio_pin_memory();
printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
fflush(stdout);
gettimeofday(&genesis, NULL);
while (todo) {
+ struct thread_data *map[MAX_JOBS];
+ struct timeval this_start;
+ int this_jobs = 0, left;
+
/*
* create threads (TD_NOT_CREATED -> TD_CREATED)
*/
if (td->stonewall && (nr_started || nr_running))
break;
+ /*
+ * Set state to created. Thread will transition
+ * to TD_INITIALIZED when it's done setting up.
+ */
td_set_runstate(td, TD_CREATED);
+ map[this_jobs++] = td;
sem_init(&startup_sem, 0, 1);
- todo--;
nr_started++;
if (td->use_thread) {
}
/*
- * start created threads (TD_CREATED -> TD_RUNNING)
+ * Wait for the started threads to transition to
+ * TD_INITIALIZED.
+ */
+ printf("fio: Waiting for threads to initialize...\n");
+ gettimeofday(&this_start, NULL);
+ left = this_jobs;
+ while (left) {
+ if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
+ break;
+
+ usleep(100000);
+
+ for (i = 0; i < this_jobs; i++) {
+ td = map[i];
+ if (!td)
+ continue;
+ if (td->runstate == TD_INITIALIZED) {
+ map[i] = NULL;
+ left--;
+ } else if (td->runstate >= TD_EXITED) {
+ map[i] = NULL;
+ left--;
+ todo--;
+ nr_running++; /* work-around... */
+ }
+ }
+ }
+
+ if (left) {
+ fprintf(stderr, "fio: %d jobs failed to start\n", left);
+ for (i = 0; i < this_jobs; i++) {
+ td = map[i];
+ if (!td)
+ continue;
+ kill(td->pid, SIGTERM);
+ }
+ break;
+ }
+
+ /*
+ * start created threads (TD_INITIALIZED -> TD_RUNNING).
*/
+ printf("fio: Go for launch\n");
for (i = 0; i < thread_number; i++) {
td = &threads[i];
- if (td->runstate != TD_CREATED)
+ if (td->runstate != TD_INITIALIZED)
continue;
td_set_runstate(td, TD_RUNNING);
nr_started--;
m_rate += td->ratemin;
t_rate += td->rate;
+ todo--;
sem_post(&td->mutex);
}
}
update_io_ticks();
+ fio_unpin_memory(mlocked_mem);
}
static void show_group_stats(struct group_run_stats *rs, int id)
}
for (i = 0; i < thread_number; i++) {
- unsigned long rbw, wbw;
+ unsigned long long rbw, wbw;
td = &threads[i];
rbw = wbw = 0;
if (td->runtime[0])
- rbw = td->io_bytes[0] / td->runtime[0];
+ rbw = td->io_bytes[0] / (unsigned long long) td->runtime[0];
if (td->runtime[1])
- wbw = td->io_bytes[1] / td->runtime[1];
+ wbw = td->io_bytes[1] / (unsigned long long) td->runtime[1];
if (rbw < rs->min_bw[0])
rs->min_bw[0] = rbw;