Reseed random generator for loop/time based runs
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 2a09fbe2d71b68fe40eaa0f7fb4f77056219d81f..ad2282d7516fe7fc0033b3bc721a12b16d370fe1 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -55,13 +55,14 @@ static struct fio_mutex *startup_mutex;
 static volatile int fio_abort;
 static int exit_value;
 static struct itimerval itimer;
+static pthread_t gtod_thread;
 
 struct io_log *agg_io_log[2];
 
 #define TERMINATE_ALL          (-1)
 #define JOB_START_TIMEOUT      (5 * 1000)
 
-static inline void td_set_runstate(struct thread_data *td, int runstate)
+void td_set_runstate(struct thread_data *td, int runstate)
 {
        if (td->runstate == runstate)
                return;
@@ -107,7 +108,7 @@ static void status_timer_arm(void)
        setitimer(ITIMER_REAL, &itimer, NULL);
 }
 
-static void sig_alrm(int sig)
+static void sig_alrm(int fio_unused sig)
 {
        if (threads) {
                update_io_ticks();
@@ -116,6 +117,13 @@ static void sig_alrm(int sig)
        }
 }
 
+/*
+ * Happens on thread runs with ctrl-c, ignore our own SIGQUIT
+ */
+static void sig_quit(int sig)
+{
+}
+
 static void sig_int(int sig)
 {
        if (threads) {
@@ -125,7 +133,7 @@ static void sig_int(int sig)
        }
 }
 
-static void sig_ill(int sig)
+static void sig_ill(int fio_unused sig)
 {
        if (!threads)
                return;
@@ -154,6 +162,24 @@ static void set_sig_handlers(void)
        act.sa_handler = sig_ill;
        act.sa_flags = SA_RESTART;
        sigaction(SIGILL, &act, NULL);
+
+       memset(&act, 0, sizeof(act));
+       act.sa_handler = sig_quit;
+       act.sa_flags = SA_RESTART;
+       sigaction(SIGQUIT, &act, NULL);
+}
+
+static inline int should_check_rate(struct thread_data *td)
+{
+       struct thread_options *o = &td->o;
+
+       /*
+        * If some rate setting was given, we need to check it
+        */
+       if (o->rate || o->ratemin || o->rate_iops || o->rate_iops_min)
+               return 1;
+
+       return 0;
 }
 
 /*
@@ -166,12 +192,6 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
        unsigned long spent;
        unsigned long rate;
 
-       /*
-        * No minimum rate set, always ok
-        */
-       if (!td->o.ratemin && !td->o.rate_iops_min)
-               return 0;
-
        /*
         * allow a 2 second settle period in the beginning
         */
@@ -340,6 +360,12 @@ requeue:
        return 0;
 }
 
+static inline void update_tv_cache(struct thread_data *td)
+{
+       if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
+               fio_gettime(&td->tv_cache, NULL);
+}
+
 /*
  * The main verify engine. Runs over the writes we previously submitted,
  * reads the blocks back in, and checks the crc/md5 of the data.
@@ -377,7 +403,9 @@ static void do_verify(struct thread_data *td)
                if (!io_u)
                        break;
 
-               if (runtime_exceeded(td, &io_u->start_time)) {
+               update_tv_cache(td);
+
+               if (runtime_exceeded(td, &td->tv_cache)) {
                        put_io_u(td, io_u);
                        td->terminate = 1;
                        break;
@@ -490,12 +518,14 @@ sync_done:
  */
 static void do_io(struct thread_data *td)
 {
-       struct timeval s;
        unsigned long usec;
        unsigned int i;
        int ret = 0;
 
-       td_set_runstate(td, TD_RUNNING);
+       if (in_ramp_time(td))
+               td_set_runstate(td, TD_RAMP);
+       else
+               td_set_runstate(td, TD_RUNNING);
 
        while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
                struct timeval comp_time;
@@ -511,9 +541,9 @@ static void do_io(struct thread_data *td)
                if (!io_u)
                        break;
 
-               memcpy(&s, &io_u->start_time, sizeof(s));
+               update_tv_cache(td);
 
-               if (runtime_exceeded(td, &s)) {
+               if (runtime_exceeded(td, &td->tv_cache)) {
                        put_io_u(td, io_u);
                        td->terminate = 1;
                        break;
@@ -526,7 +556,9 @@ static void do_io(struct thread_data *td)
                if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
                        io_u->end_io = verify_io_u;
                        td_set_runstate(td, TD_VERIFYING);
-               } else
+               } else if (in_ramp_time(td))
+                       td_set_runstate(td, TD_RAMP);
+               else
                        td_set_runstate(td, TD_RUNNING);
 
                ret = td_io_queue(td, io_u);
@@ -559,7 +591,9 @@ static void do_io(struct thread_data *td)
                                requeue_io_u(td, &io_u);
                        } else {
 sync_done:
-                               fio_gettime(&comp_time, NULL);
+                               if (should_check_rate(td))
+                                       fio_gettime(&comp_time, NULL);
+
                                bytes_done = io_u_sync_complete(td, io_u);
                                if (bytes_done < 0)
                                        ret = bytes_done;
@@ -598,7 +632,8 @@ sync_done:
                        if (full && !min_evts)
                                min_evts = 1;
 
-                       fio_gettime(&comp_time, NULL);
+                       if (should_check_rate(td))
+                               fio_gettime(&comp_time, NULL);
 
                        do {
                                ret = io_u_queued_complete(td, min_evts);
@@ -619,8 +654,8 @@ sync_done:
                 * of completions except the very first one which may look
                 * a little bursty
                 */
-               if (ramp_time_over(td)) {
-                       usec = utime_since(&s, &comp_time);
+               if (!in_ramp_time(td) && should_check_rate(td)) {
+                       usec = utime_since(&td->tv_cache, &comp_time);
 
                        rate_throttle(td, usec, bytes_done);
 
@@ -640,7 +675,7 @@ sync_done:
                                int left;
 
                                if (td->o.thinktime_spin)
-                                       __usec_sleep(td->o.thinktime_spin);
+                                       usec_spin(td->o.thinktime_spin);
 
                                left = td->o.thinktime - td->o.thinktime_spin;
                                if (left)
@@ -745,8 +780,6 @@ static int init_io_u(struct thread_data *td)
                flist_add(&io_u->list, &td->io_u_freelist);
        }
 
-       io_u_init_timeout();
-
        return 0;
 }
 
@@ -827,12 +860,8 @@ static int keep_running(struct thread_data *td)
        return 0;
 }
 
-static int clear_io_state(struct thread_data *td)
+static void reset_io_counters(struct thread_data *td)
 {
-       struct fio_file *f;
-       unsigned int i;
-       int ret;
-
        td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
        td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
        td->zone_bytes = 0;
@@ -848,17 +877,41 @@ static int clear_io_state(struct thread_data *td)
        if (td->o.time_based || td->o.loops)
                td->nr_done_files = 0;
 
-       close_files(td);
+       /*
+        * Set the same seed to get repeatable runs
+        */
+       td_fill_rand_seeds(td);
+}
 
-       ret = 0;
-       for_each_file(td, f, i) {
-               f->flags &= ~FIO_FILE_DONE;
-               ret = td_io_open_file(td, f);
-               if (ret)
-                       break;
+void reset_all_stats(struct thread_data *td)
+{
+       struct timeval tv;
+       int i;
+
+       reset_io_counters(td);
+
+       for (i = 0; i < 2; i++) {
+               td->io_bytes[i] = 0;
+               td->io_blocks[i] = 0;
+               td->io_issues[i] = 0;
+               td->ts.total_io_u[i] = 0;
        }
+       
+       fio_gettime(&tv, NULL);
+       memcpy(&td->epoch, &tv, sizeof(tv));
+       memcpy(&td->start, &tv, sizeof(tv));
+}
 
-       return ret;
+static void clear_io_state(struct thread_data *td)
+{
+       struct fio_file *f;
+       unsigned int i;
+
+       reset_io_counters(td);
+
+       close_files(td);
+       for_each_file(td, f, i)
+               f->flags &= ~FIO_FILE_DONE;
 }
 
 /*
@@ -886,8 +939,11 @@ static void *thread_main(void *data)
        td->io_hist_tree = RB_ROOT;
 
        td_set_runstate(td, TD_INITIALIZED);
+       dprint(FD_MUTEX, "up startup_mutex\n");
        fio_mutex_up(startup_mutex);
+       dprint(FD_MUTEX, "wait on td->mutex\n");
        fio_mutex_down(td->mutex);
+       dprint(FD_MUTEX, "done waiting on td->mutex\n");
 
        /*
         * the ->mutex mutex is now no longer used, close it to avoid
@@ -910,6 +966,18 @@ static void *thread_main(void *data)
                goto err;
        }
 
+       /*
+        * If we have a gettimeofday() thread, make sure we exclude that
+        * thread from this job
+        */
+       if (td->o.gtod_cpu) {
+               fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
+               if (fio_setaffinity(td) == -1) {
+                       td_verror(td, errno, "cpu_set_affinity");
+                       goto err;
+               }
+       }
+
        if (td->ioprio_set) {
                if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
                        td_verror(td, errno, "ioprio_set");
@@ -931,9 +999,6 @@ static void *thread_main(void *data)
        if (td_io_init(td))
                goto err;
 
-       if (open_files(td))
-               goto err;
-
        if (init_random_map(td))
                goto err;
 
@@ -943,7 +1008,6 @@ static void *thread_main(void *data)
        }
 
        fio_gettime(&td->epoch, NULL);
-       memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
        getrusage(RUSAGE_SELF, &td->ts.ru_start);
 
        runtime[0] = runtime[1] = 0;
@@ -951,13 +1015,14 @@ static void *thread_main(void *data)
        while (keep_running(td)) {
                fio_gettime(&td->start, NULL);
                memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
+               memcpy(&td->tv_cache, &td->start, sizeof(td->start));
 
                if (td->o.ratemin)
                        memcpy(&td->lastrate, &td->ts.stat_sample_time,
                                                        sizeof(td->lastrate));
 
-               if (clear_state && clear_io_state(td))
-                       break;
+               if (clear_state)
+                       clear_io_state(td);
 
                prune_io_piece_log(td);
 
@@ -992,8 +1057,7 @@ static void *thread_main(void *data)
                    (td->io_ops->flags & FIO_UNIDIR))
                        continue;
 
-               if (clear_io_state(td))
-                       break;
+               clear_io_state(td);
 
                fio_gettime(&td->start, NULL);
 
@@ -1012,12 +1076,27 @@ static void *thread_main(void *data)
        td->ts.io_bytes[0] = td->io_bytes[0];
        td->ts.io_bytes[1] = td->io_bytes[1];
 
-       if (td->ts.bw_log)
-               finish_log(td, td->ts.bw_log, "bw");
-       if (td->ts.slat_log)
-               finish_log(td, td->ts.slat_log, "slat");
-       if (td->ts.clat_log)
-               finish_log(td, td->ts.clat_log, "clat");
+       if (td->ts.bw_log) {
+               if (td->o.bw_log_file) {
+                       finish_log_named(td, td->ts.bw_log,
+                                               td->o.bw_log_file, "bw");
+               } else
+                       finish_log(td, td->ts.bw_log, "bw");
+       }
+       if (td->ts.slat_log) {
+               if (td->o.lat_log_file) {
+                       finish_log_named(td, td->ts.slat_log,
+                                               td->o.lat_log_file, "clat");
+               } else
+                       finish_log(td, td->ts.slat_log, "slat");
+       }
+       if (td->ts.clat_log) {
+               if (td->o.lat_log_file) {
+                       finish_log_named(td, td->ts.clat_log,
+                                               td->o.lat_log_file, "clat");
+               } else
+                       finish_log(td, td->ts.clat_log, "clat");
+       }
        if (td->o.exec_postrun) {
                if (system(td->o.exec_postrun) < 0)
                        log_err("fio: postrun %s failed\n", td->o.exec_postrun);
@@ -1034,6 +1113,12 @@ err:
        close_ioengine(td);
        cleanup_io_u(td);
 
+       if (td->o.cpumask_set) {
+               int ret = fio_cpuset_exit(&td->o.cpumask);
+
+               td_verror(td, ret, "fio_cpuset_exit");
+       }
+
        /*
         * do this very late, it will log file closing as well
         */
@@ -1163,6 +1248,46 @@ reaped:
                terminate_threads(TERMINATE_ALL);
 }
 
+static void *gtod_thread_main(void *data)
+{
+       fio_mutex_up(startup_mutex);
+
+       /*
+        * As long as we have jobs around, update the clock. It would be nice
+        * to have some way of NOT hammering that CPU with gettimeofday(),
+        * but I'm not sure what to use outside of a simple CPU nop to relax
+        * it - we don't want to lose precision.
+        */
+       while (threads) {
+               fio_gtod_update();
+               nop;
+       }
+
+       return NULL;
+}
+
+static int fio_start_gtod_thread(void)
+{
+       int ret;
+
+       ret = pthread_create(&gtod_thread, NULL, gtod_thread_main, NULL);
+       if (ret) {
+               log_err("Can't create gtod thread: %s\n", strerror(ret));
+               return 1;
+       }
+
+       ret = pthread_detach(gtod_thread);
+       if (ret) {
+               log_err("Can't detatch gtod thread: %s\n", strerror(ret));
+               return 1;
+       }
+
+       dprint(FD_MUTEX, "wait on startup_mutex\n");
+       fio_mutex_down(startup_mutex);
+       dprint(FD_MUTEX, "done waiting on startup_mutex\n");
+       return 0;
+}
+
 /*
  * Main function for kicking off and reaping jobs, as needed.
  */
@@ -1175,6 +1300,9 @@ static void run_threads(void)
        if (fio_pin_memory())
                return;
 
+       if (fio_gtod_offload && fio_start_gtod_thread())
+               return;
+
        if (!terse_output) {
                printf("Starting ");
                if (nr_thread)
@@ -1278,15 +1406,21 @@ static void run_threads(void)
                        nr_started++;
 
                        if (td->o.use_thread) {
+                               int ret;
+
                                dprint(FD_PROCESS, "will pthread_create\n");
-                               if (pthread_create(&td->thread, NULL,
-                                                  thread_main, td)) {
-                                       perror("pthread_create");
+                               ret = pthread_create(&td->thread, NULL,
+                                                       thread_main, td);
+                               if (ret) {
+                                       log_err("pthread_create: %s\n",
+                                                       strerror(ret));
                                        nr_started--;
                                        break;
                                }
-                               if (pthread_detach(td->thread) < 0)
-                                       perror("pthread_detach");
+                               ret = pthread_detach(td->thread);
+                               if (ret)
+                                       log_err("pthread_detach: %s",
+                                                       strerror(ret));
                        } else {
                                pid_t pid;
                                dprint(FD_PROCESS, "will fork\n");
@@ -1298,7 +1432,9 @@ static void run_threads(void)
                                } else if (i == fio_debug_jobno)
                                        *fio_debug_jobp = pid;
                        }
+                       dprint(FD_MUTEX, "wait on startup_mutex\n");
                        fio_mutex_down(startup_mutex);
+                       dprint(FD_MUTEX, "done waiting on startup_mutex\n");
                }
 
                /*
@@ -1347,7 +1483,10 @@ static void run_threads(void)
                        if (td->runstate != TD_INITIALIZED)
                                continue;
 
-                       td_set_runstate(td, TD_RUNNING);
+                       if (in_ramp_time(td))
+                               td_set_runstate(td, TD_RAMP);
+                       else
+                               td_set_runstate(td, TD_RUNNING);
                        nr_running++;
                        nr_started--;
                        m_rate += td->o.ratemin;