[PATCH] Improve rate functionality
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 76710e19fba08166902c52cdb8218cda45235928..225aaa4d4550c2ea385f8eaaee70b9b8d1bc321d 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -45,14 +45,19 @@ int groupid = 0;
 int thread_number = 0;
 int shm_id = 0;
 int temp_stall_ts;
-char *fio_inst_prefix = _INST_PREFIX;
 
 static volatile int startup_sem;
+static volatile int fio_abort;
 
 #define TERMINATE_ALL          (-1)
 #define JOB_START_TIMEOUT      (5 * 1000)
 
-static void terminate_threads(int group_id)
+static inline void td_set_runstate(struct thread_data *td, int runstate)
+{
+       td->runstate = runstate;
+}
+
+static void terminate_threads(int group_id, int forced_kill)
 {
        struct thread_data *td;
        int i;
@@ -61,6 +66,8 @@ static void terminate_threads(int group_id)
                if (group_id == TERMINATE_ALL || groupid == td->groupid) {
                        td->terminate = 1;
                        td->start_delay = 0;
+                       if (forced_kill)
+                               td_set_runstate(td, TD_EXITED);
                }
        }
 }
@@ -73,10 +80,15 @@ static void sig_handler(int sig)
                        disk_util_timer_arm();
                        print_thread_status();
                        break;
+               case SIGSEGV:
+                       fprintf(stderr, "fio: got segfault, aborting\n");
+                       terminate_threads(TERMINATE_ALL, 1);
+                       fio_abort = 1;
+                       exit(0);
                default:
-                       printf("\nfio: terminating on signal\n");
+                       printf("\nfio: terminating on signal %d\n", sig);
                        fflush(stdout);
-                       terminate_threads(TERMINATE_ALL);
+                       terminate_threads(TERMINATE_ALL, 0);
                        break;
        }
 }
@@ -106,7 +118,7 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
 
                rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
                if (rate < td->ratemin) {
-                       fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate);
+                       fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
                        return 1;
                }
        }
@@ -126,11 +138,6 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
        return 0;
 }
 
-static inline void td_set_runstate(struct thread_data *td, int runstate)
-{
-       td->runstate = runstate;
-}
-
 static struct fio_file *get_next_file(struct thread_data *td)
 {
        unsigned int old_next_file = td->next_file;
@@ -243,9 +250,8 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
  * The main verify engine. Runs over the writes we previusly submitted,
  * reads the blocks back in, and checks the crc/md5 of the data.
  */
-void do_verify(struct thread_data *td)
+static void do_verify(struct thread_data *td)
 {
-       struct timeval t;
        struct io_u *io_u, *v_io_u = NULL;
        struct io_completion_data icd;
        struct fio_file *f;
@@ -266,14 +272,15 @@ void do_verify(struct thread_data *td)
                if (td->terminate)
                        break;
 
-               gettimeofday(&t, NULL);
-               if (runtime_exceeded(td, &t))
-                       break;
-
                io_u = __get_io_u(td);
                if (!io_u)
                        break;
 
+               if (runtime_exceeded(td, &io_u->start_time)) {
+                       put_io_u(td, io_u);
+                       break;
+               }
+
                if (get_next_verify(td, io_u)) {
                        put_io_u(td, io_u);
                        break;
@@ -314,6 +321,7 @@ void do_verify(struct thread_data *td)
                v_io_u = td->io_ops->event(td, 0);
                icd.nr = 1;
                icd.error = 0;
+               fio_gettime(&icd.time, NULL);
                io_completed(td, v_io_u, &icd);
 
                if (icd.error) {
@@ -350,7 +358,7 @@ static void do_cpuio(struct thread_data *td)
        int i = 0;
 
        while (!td->terminate) {
-               gettimeofday(&e, NULL);
+               fio_gettime(&e, NULL);
 
                if (runtime_exceeded(td, &e))
                        break;
@@ -371,7 +379,7 @@ static void do_cpuio(struct thread_data *td)
 static void do_io(struct thread_data *td)
 {
        struct io_completion_data icd;
-       struct timeval s, e;
+       struct timeval s;
        unsigned long usec;
        struct fio_file *f;
        int i, ret = 0;
@@ -379,7 +387,6 @@ static void do_io(struct thread_data *td)
        td_set_runstate(td, TD_RUNNING);
 
        while (td->this_io_bytes[td->ddir] < td->io_size) {
-               struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
                struct timespec *timeout;
                int min_evts = 0;
                struct io_u *io_u;
@@ -407,6 +414,8 @@ static void do_io(struct thread_data *td)
                add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
 
                if (td->cur_depth < td->iodepth) {
+                       struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
+
                        timeout = &ts;
                        min_evts = 0;
                } else {
@@ -433,26 +442,30 @@ static void do_io(struct thread_data *td)
                 * of completions except the very first one which may look
                 * a little bursty
                 */
-               gettimeofday(&e, NULL);
-               usec = utime_since(&s, &e);
+               usec = utime_since(&s, &icd.time);
 
-               rate_throttle(td, usec, icd.bytes_done[td->ddir]);
+               rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
 
-               if (check_min_rate(td, &e)) {
-                       if (rate_quit)
-                               terminate_threads(td->groupid);
-                       td_verror(td, ENOMEM);
+               if (check_min_rate(td, &icd.time)) {
+                       if (exitall_on_terminate)
+                               terminate_threads(td->groupid, 0);
+                       td_verror(td, ENODATA);
                        break;
                }
 
-               if (runtime_exceeded(td, &e))
+               if (runtime_exceeded(td, &icd.time))
                        break;
 
-               if (td->thinktime)
-                       usec_sleep(td, td->thinktime);
+               if (td->thinktime) {
+                       unsigned long long b;
+
+                       b = td->io_blocks[0] + td->io_blocks[1];
+                       if (!(b % td->thinktime_blocks))
+                               usec_sleep(td, td->thinktime);
+               }
        }
 
-       if (!ret) {
+       if (!td->error) {
                if (td->cur_depth)
                        cleanup_pending_aio(td);
 
@@ -495,6 +508,7 @@ static void fill_rand_buf(struct io_u *io_u, int max_bs)
 static int init_io_u(struct thread_data *td)
 {
        struct io_u *io_u;
+       unsigned int max_bs;
        int i, max_units;
        char *p;
 
@@ -506,7 +520,13 @@ static int init_io_u(struct thread_data *td)
        else
                max_units = td->iodepth;
 
-       td->orig_buffer_size = td->max_bs * max_units + MASK;
+       max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
+       td->orig_buffer_size = max_bs * max_units;
+
+       if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
+               td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
+       else
+               td->orig_buffer_size += MASK;
 
        if (allocate_io_mem(td))
                return 1;
@@ -517,9 +537,9 @@ static int init_io_u(struct thread_data *td)
                memset(io_u, 0, sizeof(*io_u));
                INIT_LIST_HEAD(&io_u->list);
 
-               io_u->buf = p + td->max_bs * i;
+               io_u->buf = p + max_bs * i;
                if (td_write(td) || td_rw(td))
-                       fill_rand_buf(io_u, td->max_bs);
+                       fill_rand_buf(io_u, max_bs);
 
                io_u->index = i;
                list_add(&io_u->list, &td->io_u_freelist);
@@ -604,6 +624,7 @@ static void clear_io_state(struct thread_data *td)
  */
 static void *thread_main(void *data)
 {
+       unsigned long long runtime[2];
        struct thread_data *td = data;
 
        if (!td->use_thread)
@@ -654,15 +675,18 @@ static void *thread_main(void *data)
 
        if (!td->create_serialize && setup_files(td))
                goto err;
-
-       gettimeofday(&td->epoch, NULL);
+       if (open_files(td))
+               goto err;
 
        if (td->exec_prerun)
                system(td->exec_prerun);
 
+       fio_gettime(&td->epoch, NULL);
+       getrusage(RUSAGE_SELF, &td->ru_start);
+
+       runtime[0] = runtime[1] = 0;
        while (td->loops--) {
-               getrusage(RUSAGE_SELF, &td->ru_start);
-               gettimeofday(&td->start, NULL);
+               fio_gettime(&td->start, NULL);
                memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
 
                if (td->ratemin)
@@ -676,11 +700,9 @@ static void *thread_main(void *data)
                else
                        do_io(td);
 
-               td->runtime[td->ddir] += mtime_since_now(&td->start);
+               runtime[td->ddir] += utime_since_now(&td->start);
                if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
-                       td->runtime[td->ddir ^ 1] = td->runtime[td->ddir];
-
-               update_rusage_stat(td);
+                       runtime[td->ddir ^ 1] = runtime[td->ddir];
 
                if (td->error || td->terminate)
                        break;
@@ -689,16 +711,21 @@ static void *thread_main(void *data)
                        continue;
 
                clear_io_state(td);
-               gettimeofday(&td->start, NULL);
+               fio_gettime(&td->start, NULL);
 
                do_verify(td);
 
-               td->runtime[DDIR_READ] += mtime_since_now(&td->start);
+               runtime[DDIR_READ] += utime_since_now(&td->start);
 
                if (td->error || td->terminate)
                        break;
        }
 
+       update_rusage_stat(td);
+       fio_gettime(&td->end_time, NULL);
+       td->runtime[0] = runtime[0] / 1000;
+       td->runtime[1] = runtime[1] / 1000;
+
        if (td->bw_log)
                finish_log(td, td->bw_log, "bw");
        if (td->slat_log)
@@ -711,7 +738,7 @@ static void *thread_main(void *data)
                system(td->exec_postrun);
 
        if (exitall_on_terminate)
-               terminate_threads(td->groupid);
+               terminate_threads(td->groupid, 0);
 
 err:
        close_files(td);
@@ -749,12 +776,12 @@ static void *fork_main(int shmid, int offset)
 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 {
        struct thread_data *td;
-       int i, cputhreads;
+       int i, cputhreads, pending;
 
        /*
         * reap exited threads (TD_EXITED -> TD_REAPED)
         */
-       cputhreads = 0;
+       pending = cputhreads = 0;
        for_each_td(td, i) {
                /*
                 * ->io_ops is NULL for a thread that has closed its
@@ -763,8 +790,12 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
                        cputhreads++;
 
-               if (td->runstate != TD_EXITED)
+               if (td->runstate != TD_EXITED) {
+                       if (td->runstate < TD_RUNNING)
+                               pending++;
+
                        continue;
+               }
 
                td_set_runstate(td, TD_REAPED);
 
@@ -781,8 +812,8 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                (*t_rate) -= td->rate;
        }
 
-       if (*nr_running == cputhreads)
-               terminate_threads(TERMINATE_ALL);
+       if (*nr_running == cputhreads && !pending)
+               terminate_threads(TERMINATE_ALL, 0);
 }
 
 /*
@@ -804,6 +835,7 @@ static void run_threads(void)
 
        signal(SIGINT, sig_handler);
        signal(SIGALRM, sig_handler);
+       signal(SIGSEGV, sig_handler);
 
        todo = thread_number;
        nr_running = 0;
@@ -813,10 +845,10 @@ static void run_threads(void)
        for_each_td(td, i) {
                print_status_init(td->thread_number - 1);
 
-               init_disk_util(td);
-
-               if (!td->create_serialize)
+               if (!td->create_serialize) {
+                       init_disk_util(td);
                        continue;
+               }
 
                /*
                 * do file setup here so it happens sequentially,
@@ -827,9 +859,9 @@ static void run_threads(void)
                        td_set_runstate(td, TD_REAPED);
                        todo--;
                }
-       }
 
-       time_init();
+               init_disk_util(td);
+       }
 
        while (todo) {
                struct thread_data *map[MAX_JOBS];
@@ -890,9 +922,9 @@ static void run_threads(void)
                 * Wait for the started threads to transition to
                 * TD_INITIALIZED.
                 */
-               gettimeofday(&this_start, NULL);
+               fio_gettime(&this_start, NULL);
                left = this_jobs;
-               while (left) {
+               while (left && !fio_abort) {
                        if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
                                break;
 
@@ -969,7 +1001,9 @@ int main(int argc, char *argv[])
        disk_util_timer_arm();
 
        run_threads();
-       show_run_stats();
+
+       if (!fio_abort)
+               show_run_stats();
 
        return 0;
 }