Fix problem with too many opened files
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 37a425e85ee6fde71ec7aa537b0f6947687cc128..6464b6d917ff39c7ecf5fc5efcd87fc4f93c079b 100644 (file)
--- a/fio.c
+++ b/fio.c
@@ -37,6 +37,8 @@
 #include "fio.h"
 #include "hash.h"
 #include "smalloc.h"
+#include "verify.h"
+#include "diskutil.h"
 
 unsigned long page_mask;
 unsigned long page_size;
@@ -52,21 +54,24 @@ int temp_stall_ts;
 unsigned long done_secs = 0;
 
 static struct fio_mutex *startup_mutex;
+static struct fio_mutex *writeout_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;
 
-       dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", td->pid, td->runstate,
-                                                               runstate);
+       dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid,
+                                               td->runstate, runstate);
        td->runstate = runstate;
 }
 
@@ -80,7 +85,7 @@ static void terminate_threads(int group_id)
        for_each_td(td, i) {
                if (group_id == TERMINATE_ALL || groupid == td->groupid) {
                        dprint(FD_PROCESS, "setting terminate on %s/%d\n",
-                                                       td->o.name, td->pid);
+                                               td->o.name, (int) td->pid);
                        td->terminate = 1;
                        td->o.start_delay = 0;
 
@@ -99,40 +104,87 @@ static void terminate_threads(int group_id)
        }
 }
 
-static void sig_handler(int sig)
+static void status_timer_arm(void)
 {
-       if (!threads)
-               return;
+       itimer.it_value.tv_sec = 0;
+       itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
+       setitimer(ITIMER_REAL, &itimer, NULL);
+}
 
-       switch (sig) {
-       case SIGALRM:
+static void sig_alrm(int fio_unused sig)
+{
+       if (threads) {
                update_io_ticks();
-               disk_util_timer_arm();
                print_thread_status();
-               break;
-       default:
+               status_timer_arm();
+       }
+}
+
+/*
+ * 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) {
                printf("\nfio: terminating on signal %d\n", sig);
                fflush(stdout);
                terminate_threads(TERMINATE_ALL);
-               break;
        }
 }
 
+static void sig_ill(int fio_unused sig)
+{
+       if (!threads)
+               return;
+
+       log_err("fio: illegal instruction. your cpu does not support "
+               "the sse4.2 instruction for crc32c\n");
+       terminate_threads(TERMINATE_ALL);
+       exit(4);
+}
+
+static void set_sig_handlers(void)
+{
+       struct sigaction act;
+
+       memset(&act, 0, sizeof(act));
+       act.sa_handler = sig_alrm;
+       act.sa_flags = SA_RESTART;
+       sigaction(SIGALRM, &act, NULL);
+
+       memset(&act, 0, sizeof(act));
+       act.sa_handler = sig_int;
+       act.sa_flags = SA_RESTART;
+       sigaction(SIGINT, &act, NULL);
+
+       memset(&act, 0, sizeof(act));
+       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);
+}
+
 /*
  * Check if we are above the minimum rate given.
  */
-static int check_min_rate(struct thread_data *td, struct timeval *now)
+static int __check_min_rate(struct thread_data *td, struct timeval *now,
+                           enum td_ddir ddir)
 {
        unsigned long long bytes = 0;
        unsigned long iops = 0;
        unsigned long spent;
        unsigned long rate;
-
-       /*
-        * No minimum rate set, always ok
-        */
-       if (!td->o.ratemin && !td->o.rate_iops_min)
-               return 0;
+       unsigned int ratemin = 0;
+       unsigned int rate_iops = 0;
+       unsigned int rate_iops_min = 0;
 
        /*
         * allow a 2 second settle period in the beginning
@@ -140,38 +192,35 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
        if (mtime_since(&td->start, now) < 2000)
                return 0;
 
-       if (td_read(td)) {
-               iops += td->io_blocks[DDIR_READ];
-               bytes += td->this_io_bytes[DDIR_READ];
-       }
-       if (td_write(td)) {
-               iops += td->io_blocks[DDIR_WRITE];
-               bytes += td->this_io_bytes[DDIR_WRITE];
-       }
+       iops += td->io_blocks[ddir];
+       bytes += td->this_io_bytes[ddir];
+       ratemin += td->o.ratemin[ddir];
+       rate_iops += td->o.rate_iops[ddir];
+       rate_iops_min += td->o.rate_iops_min[ddir];
 
        /*
         * if rate blocks is set, sample is running
         */
-       if (td->rate_bytes || td->rate_blocks) {
-               spent = mtime_since(&td->lastrate, now);
+       if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
+               spent = mtime_since(&td->lastrate[ddir], now);
                if (spent < td->o.ratecycle)
                        return 0;
 
-               if (td->o.rate) {
+               if (td->o.rate[ddir]) {
                        /*
                         * check bandwidth specified rate
                         */
-                       if (bytes < td->rate_bytes) {
+                       if (bytes < td->rate_bytes[ddir]) {
                                log_err("%s: min rate %u not met\n", td->o.name,
-                                                               td->o.ratemin);
+                                                               ratemin);
                                return 1;
                        } else {
-                               rate = (bytes - td->rate_bytes) / spent;
-                               if (rate < td->o.ratemin ||
-                                   bytes < td->rate_bytes) {
+                               rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
+                               if (rate < ratemin ||
+                                   bytes < td->rate_bytes[ddir]) {
                                        log_err("%s: min rate %u not met, got"
                                                " %luKiB/sec\n", td->o.name,
-                                                       td->o.ratemin, rate);
+                                                       ratemin, rate);
                                        return 1;
                                }
                        }
@@ -179,29 +228,41 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
                        /*
                         * checks iops specified rate
                         */
-                       if (iops < td->o.rate_iops) {
+                       if (iops < rate_iops) {
                                log_err("%s: min iops rate %u not met\n",
-                                               td->o.name, td->o.rate_iops);
+                                               td->o.name, rate_iops);
                                return 1;
                        } else {
-                               rate = (iops - td->rate_blocks) / spent;
-                               if (rate < td->o.rate_iops_min ||
-                                   iops < td->rate_blocks) {
+                               rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
+                               if (rate < rate_iops_min ||
+                                   iops < td->rate_blocks[ddir]) {
                                        log_err("%s: min iops rate %u not met,"
                                                " got %lu\n", td->o.name,
-                                                       td->o.rate_iops_min,
-                                                       rate);
+                                                       rate_iops_min, rate);
                                }
                        }
                }
        }
 
-       td->rate_bytes = bytes;
-       td->rate_blocks = iops;
-       memcpy(&td->lastrate, now, sizeof(*now));
+       td->rate_bytes[ddir] = bytes;
+       td->rate_blocks[ddir] = iops;
+       memcpy(&td->lastrate[ddir], now, sizeof(*now));
        return 0;
 }
 
+static int check_min_rate(struct thread_data *td, struct timeval *now,
+                         unsigned long *bytes_done)
+{
+       int ret = 0;
+
+       if (bytes_done[0])
+               ret |= __check_min_rate(td, now, 0);
+       if (bytes_done[1])
+               ret |= __check_min_rate(td, now, 1);
+
+       return ret;
+}
+
 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 {
        if (!td->o.timeout)
@@ -218,14 +279,14 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
  */
 static void cleanup_pending_aio(struct thread_data *td)
 {
-       struct list_head *entry, *n;
+       struct flist_head *entry, *n;
        struct io_u *io_u;
        int r;
 
        /*
         * get immediately available events, if any
         */
-       r = io_u_queued_complete(td, 0);
+       r = io_u_queued_complete(td, 0, NULL);
        if (r < 0)
                return;
 
@@ -233,8 +294,8 @@ static void cleanup_pending_aio(struct thread_data *td)
         * now cancel remaining active events
         */
        if (td->io_ops->cancel) {
-               list_for_each_safe(entry, n, &td->io_u_busylist) {
-                       io_u = list_entry(entry, struct io_u, list);
+               flist_for_each_safe(entry, n, &td->io_u_busylist) {
+                       io_u = flist_entry(entry, struct io_u, list);
 
                        /*
                         * if the io_u isn't in flight, then that generally
@@ -253,7 +314,7 @@ static void cleanup_pending_aio(struct thread_data *td)
        }
 
        if (td->cur_depth)
-               r = io_u_queued_complete(td, td->cur_depth);
+               r = io_u_queued_complete(td, td->cur_depth, NULL);
 }
 
 /*
@@ -283,7 +344,7 @@ requeue:
                put_io_u(td, io_u);
                return 1;
        } else if (ret == FIO_Q_QUEUED) {
-               if (io_u_queued_complete(td, 1) < 0)
+               if (io_u_queued_complete(td, 1, NULL) < 0)
                        return 1;
        } else if (ret == FIO_Q_COMPLETED) {
                if (io_u->error) {
@@ -291,7 +352,7 @@ requeue:
                        return 1;
                }
 
-               if (io_u_sync_complete(td, io_u) < 0)
+               if (io_u_sync_complete(td, io_u, NULL) < 0)
                        return 1;
        } else if (ret == FIO_Q_BUSY) {
                if (td_io_commit(td))
@@ -302,6 +363,12 @@ 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.
@@ -318,7 +385,7 @@ static void do_verify(struct thread_data *td)
         * read from disk.
         */
        for_each_file(td, f, i) {
-               if (!(f->flags & FIO_FILE_OPEN))
+               if (!fio_file_open(f))
                        continue;
                if (fio_io_sync(td, f))
                        break;
@@ -333,13 +400,15 @@ static void do_verify(struct thread_data *td)
 
        io_u = NULL;
        while (!td->terminate) {
-               int ret2;
+               int ret2, full;
 
                io_u = __get_io_u(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;
@@ -370,7 +439,7 @@ static void do_verify(struct thread_data *td)
                                 * zero read, fail
                                 */
                                if (!bytes) {
-                                       td_verror(td, ENODATA, "full resid");
+                                       td_verror(td, EIO, "full resid");
                                        put_io_u(td, io_u);
                                        break;
                                }
@@ -387,7 +456,7 @@ static void do_verify(struct thread_data *td)
                                requeue_io_u(td, &io_u);
                        } else {
 sync_done:
-                               ret = io_u_sync_complete(td, io_u);
+                               ret = io_u_sync_complete(td, io_u, NULL);
                                if (ret < 0)
                                        break;
                        }
@@ -413,19 +482,25 @@ sync_done:
                 * if we can queue more, do so. but check if there are
                 * completed io_u's first.
                 */
-               min_events = 0;
-               if (queue_full(td) || ret == FIO_Q_BUSY) {
-                       if (td->cur_depth >= td->o.iodepth_low)
-                               min_events = td->cur_depth - td->o.iodepth_low;
-                       if (!min_events)
+               full = queue_full(td) || ret == FIO_Q_BUSY;
+               if (full || !td->o.iodepth_batch_complete) {
+                       min_events = td->o.iodepth_batch_complete;
+                       if (full && !min_events)
                                min_events = 1;
-               }
 
-               /*
-                * Reap required number of io units, if any, and do the
-                * verification on them through the callback handler
-                */
-               if (io_u_queued_complete(td, min_events) < 0)
+                       do {
+                               /*
+                                * Reap required number of io units, if any,
+                                * and do the verification on them through
+                                * the callback handler
+                                */
+                               if (io_u_queued_complete(td, min_events, NULL) < 0) {
+                                       ret = -1;
+                                       break;
+                               }
+                       } while (full && (td->cur_depth > td->o.iodepth_low));
+               }
+               if (ret < 0)
                        break;
        }
 
@@ -433,7 +508,7 @@ sync_done:
                min_events = td->cur_depth;
 
                if (min_events)
-                       ret = io_u_queued_complete(td, min_events);
+                       ret = io_u_queued_complete(td, min_events, NULL);
        } else
                cleanup_pending_aio(td);
 
@@ -446,19 +521,20 @@ 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;
-               long bytes_done = 0;
+               unsigned long bytes_done[2] = { 0, 0 };
                int min_evts = 0;
                struct io_u *io_u;
-               int ret2;
+               int ret2, full;
 
                if (td->terminate)
                        break;
@@ -467,9 +543,9 @@ 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;
@@ -482,7 +558,9 @@ 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);
@@ -498,7 +576,7 @@ static void do_io(struct thread_data *td)
                                 * zero read, fail
                                 */
                                if (!bytes) {
-                                       td_verror(td, ENODATA, "full resid");
+                                       td_verror(td, EIO, "full resid");
                                        put_io_u(td, io_u);
                                        break;
                                }
@@ -515,10 +593,13 @@ static void do_io(struct thread_data *td)
                                requeue_io_u(td, &io_u);
                        } else {
 sync_done:
-                               fio_gettime(&comp_time, NULL);
-                               bytes_done = io_u_sync_complete(td, io_u);
-                               if (bytes_done < 0)
-                                       ret = bytes_done;
+                               if (__should_check_rate(td, 0) ||
+                                   __should_check_rate(td, 1))
+                                       fio_gettime(&comp_time, NULL);
+
+                               ret = io_u_sync_complete(td, io_u, bytes_done);
+                               if (ret < 0)
+                                       break;
                        }
                        break;
                case FIO_Q_QUEUED:
@@ -548,35 +629,36 @@ sync_done:
                /*
                 * See if we need to complete some commands
                 */
-               if (queue_full(td) || ret == FIO_Q_BUSY) {
-                       min_evts = 0;
-                       if (td->cur_depth >= td->o.iodepth_low)
-                               min_evts = td->cur_depth - td->o.iodepth_low;
-                       if (!min_evts)
+               full = queue_full(td) || ret == FIO_Q_BUSY;
+               if (full || !td->o.iodepth_batch_complete) {
+                       min_evts = td->o.iodepth_batch_complete;
+                       if (full && !min_evts)
                                min_evts = 1;
-                       fio_gettime(&comp_time, NULL);
-                       bytes_done = io_u_queued_complete(td, min_evts);
-                       if (bytes_done < 0)
-                               break;
-               }
 
-               if (!bytes_done)
-                       continue;
+                       if (__should_check_rate(td, 0) ||
+                           __should_check_rate(td, 1))
+                               fio_gettime(&comp_time, NULL);
 
-               /*
-                * the rate is batched for now, it should work for batches
-                * of completions except the very first one which may look
-                * a little bursty
-                */
-               usec = utime_since(&s, &comp_time);
+                       do {
+                               ret = io_u_queued_complete(td, min_evts, bytes_done);
+                               if (ret < 0)
+                                       break;
 
-               rate_throttle(td, usec, bytes_done);
+                       } while (full && (td->cur_depth > td->o.iodepth_low));
+               }
 
-               if (check_min_rate(td, &comp_time)) {
-                       if (exitall_on_terminate)
-                               terminate_threads(td->groupid);
-                       td_verror(td, ENODATA, "check_min_rate");
+               if (ret < 0)
                        break;
+               if (!(bytes_done[0] + bytes_done[1]))
+                       continue;
+
+               if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
+                       if (check_min_rate(td, &comp_time, bytes_done)) {
+                               if (exitall_on_terminate)
+                                       terminate_threads(td->groupid);
+                               td_verror(td, EIO, "check_min_rate");
+                               break;
+                       }
                }
 
                if (td->o.thinktime) {
@@ -587,7 +669,7 @@ 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)
@@ -605,13 +687,13 @@ sync_done:
 
                i = td->cur_depth;
                if (i)
-                       ret = io_u_queued_complete(td, i);
+                       ret = io_u_queued_complete(td, i, NULL);
 
                if (should_fsync(td) && td->o.end_fsync) {
                        td_set_runstate(td, TD_FSYNCING);
 
                        for_each_file(td, f, i) {
-                               if (!(f->flags & FIO_FILE_OPEN))
+                               if (!fio_file_open(f))
                                        continue;
                                fio_io_sync(td, f);
                        }
@@ -628,13 +710,13 @@ sync_done:
 
 static void cleanup_io_u(struct thread_data *td)
 {
-       struct list_head *entry, *n;
+       struct flist_head *entry, *n;
        struct io_u *io_u;
 
-       list_for_each_safe(entry, n, &td->io_u_freelist) {
-               io_u = list_entry(entry, struct io_u, list);
+       flist_for_each_safe(entry, n, &td->io_u_freelist) {
+               io_u = flist_entry(entry, struct io_u, list);
 
-               list_del(&io_u->list);
+               flist_del(&io_u->list);
                free(io_u);
        }
 
@@ -645,7 +727,7 @@ static int init_io_u(struct thread_data *td)
 {
        struct io_u *io_u;
        unsigned int max_bs;
-       int i, max_units;
+       int cl_align, i, max_units;
        char *p;
 
        max_units = td->o.iodepth;
@@ -673,12 +755,22 @@ static int init_io_u(struct thread_data *td)
        else
                p = td->orig_buffer;
 
+       cl_align = os_cache_line_size();
+
        for (i = 0; i < max_units; i++) {
+               void *ptr;
+
                if (td->terminate)
                        return 1;
-               io_u = malloc(sizeof(*io_u));
+
+               if (posix_memalign(&ptr, cl_align, sizeof(*io_u))) {
+                       log_err("fio: posix_memalign=%s\n", strerror(errno));
+                       break;
+               }
+
+               io_u = ptr;
                memset(io_u, 0, sizeof(*io_u));
-               INIT_LIST_HEAD(&io_u->list);
+               INIT_FLIST_HEAD(&io_u->list);
 
                if (!(td->io_ops->flags & FIO_NOIO)) {
                        io_u->buf = p + max_bs * i;
@@ -689,11 +781,9 @@ static int init_io_u(struct thread_data *td)
 
                io_u->index = i;
                io_u->flags = IO_U_F_FREE;
-               list_add(&io_u->list, &td->io_u_freelist);
+               flist_add(&io_u->list, &td->io_u_freelist);
        }
 
-       io_u_init_timeout();
-
        return 0;
 }
 
@@ -774,18 +864,13 @@ 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;
-       td->rate_bytes = 0;
-       td->rate_blocks = 0;
-       td->rw_end_set[0] = td->rw_end_set[1] = 0;
+       td->rate_bytes[0] = td->rate_bytes[1] = 0;
+       td->rate_blocks[0] = td->rate_blocks[1] = 0;
 
        td->last_was_sync = 0;
 
@@ -795,16 +880,56 @@ 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));
+}
+
+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)
+               fio_file_clear_done(f);
+}
 
+static int exec_string(const char *string)
+{
+       int ret, newlen = strlen(string) + 1 + 8;
+       char *str;
+
+       str = malloc(newlen);
+       sprintf(str, "sh -c %s", string);
+
+       ret = system(str);
+       if (ret == -1)
+               log_err("fio: exec of cmd <%s> failed\n", str);
+
+       free(str);
        return ret;
 }
 
@@ -823,18 +948,21 @@ static void *thread_main(void *data)
 
        td->pid = getpid();
 
-       dprint(FD_PROCESS, "jobs pid=%d started\n", td->pid);
+       dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 
-       INIT_LIST_HEAD(&td->io_u_freelist);
-       INIT_LIST_HEAD(&td->io_u_busylist);
-       INIT_LIST_HEAD(&td->io_u_requeues);
-       INIT_LIST_HEAD(&td->io_log_list);
-       INIT_LIST_HEAD(&td->io_hist_list);
+       INIT_FLIST_HEAD(&td->io_u_freelist);
+       INIT_FLIST_HEAD(&td->io_u_busylist);
+       INIT_FLIST_HEAD(&td->io_u_requeues);
+       INIT_FLIST_HEAD(&td->io_log_list);
+       INIT_FLIST_HEAD(&td->io_hist_list);
        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
@@ -857,6 +985,18 @@ 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");
@@ -878,19 +1018,20 @@ 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;
 
        if (td->o.exec_prerun) {
-               if (system(td->o.exec_prerun) < 0)
+               if (exec_string(td->o.exec_prerun))
+                       goto err;
+       }
+
+       if (td->o.pre_read) {
+               if (pre_read_files(td) < 0)
                        goto err;
        }
 
        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;
@@ -898,13 +1039,14 @@ 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)
+               if (td->o.ratemin[0] || td->o.ratemin[1])
                        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);
 
@@ -913,21 +1055,11 @@ static void *thread_main(void *data)
                clear_state = 1;
 
                if (td_read(td) && td->io_bytes[DDIR_READ]) {
-                       if (td->rw_end_set[DDIR_READ])
-                               elapsed = utime_since(&td->start,
-                                                     &td->rw_end[DDIR_READ]);
-                       else
-                               elapsed = utime_since_now(&td->start);
-
+                       elapsed = utime_since_now(&td->start);
                        runtime[DDIR_READ] += elapsed;
                }
                if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
-                       if (td->rw_end_set[DDIR_WRITE])
-                               elapsed = utime_since(&td->start,
-                                                     &td->rw_end[DDIR_WRITE]);
-                       else
-                               elapsed = utime_since_now(&td->start);
-
+                       elapsed = utime_since_now(&td->start);
                        runtime[DDIR_WRITE] += elapsed;
                }
 
@@ -939,8 +1071,7 @@ 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);
 
@@ -959,28 +1090,49 @@ 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->o.exec_postrun) {
-               if (system(td->o.exec_postrun) < 0)
-                       log_err("fio: postrun %s failed\n", td->o.exec_postrun);
+       fio_mutex_down(writeout_mutex);
+       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, "slat");
+               } 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");
+       }
+       fio_mutex_up(writeout_mutex);
+       if (td->o.exec_postrun)
+               exec_string(td->o.exec_postrun);
 
        if (exitall_on_terminate)
                terminate_threads(td->groupid);
 
 err:
        if (td->error)
-               printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error,
+               printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error,
                                                        td->verror);
        close_and_free_files(td);
        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
         */
@@ -1063,8 +1215,8 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                ret = waitpid(td->pid, &status, flags);
                if (ret < 0) {
                        if (errno == ECHILD) {
-                               log_err("fio: pid=%d disappeared %d\n", td->pid,
-                                                               td->runstate);
+                               log_err("fio: pid=%d disappeared %d\n",
+                                               (int) td->pid, td->runstate);
                                td_set_runstate(td, TD_REAPED);
                                goto reaped;
                        }
@@ -1075,7 +1227,7 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 
                                if (sig != SIGQUIT)
                                        log_err("fio: pid=%d, got signal=%d\n",
-                                                               td->pid, sig);
+                                                       (int) td->pid, sig);
                                td_set_runstate(td, TD_REAPED);
                                goto reaped;
                        }
@@ -1095,8 +1247,8 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                continue;
 reaped:
                (*nr_running)--;
-               (*m_rate) -= td->o.ratemin;
-               (*t_rate) -= td->o.rate;
+               (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
+               (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
                if (!td->pid)
                        pending--;
 
@@ -1110,6 +1262,46 @@ 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.
  */
@@ -1122,6 +1314,9 @@ 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)
@@ -1137,8 +1332,7 @@ static void run_threads(void)
                fflush(stdout);
        }
 
-       signal(SIGINT, sig_handler);
-       signal(SIGALRM, sig_handler);
+       set_sig_handlers();
 
        todo = thread_number;
        nr_running = 0;
@@ -1161,8 +1355,8 @@ static void run_threads(void)
                if (setup_files(td)) {
                        exit_value++;
                        if (td->error)
-                               log_err("fio: pid=%d, err=%d/%s\n", td->pid,
-                                                       td->error, td->verror);
+                               log_err("fio: pid=%d, err=%d/%s\n",
+                                       (int) td->pid, td->error, td->verror);
                        td_set_runstate(td, TD_REAPED);
                        todo--;
                } else {
@@ -1174,8 +1368,12 @@ static void run_threads(void)
                         * its own files. so close them, if we opened them
                         * for creation
                         */
-                       for_each_file(td, f, i)
-                               td_io_close_file(td, f);
+                       for_each_file(td, f, i) {
+                               if (fio_file_open(f))
+                                       td_io_close_file(td, f);
+                               else
+                                       assert(f->fd == -1);
+                       }
                }
 
                init_disk_util(td);
@@ -1226,15 +1424,21 @@ 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");
@@ -1246,7 +1450,9 @@ 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");
                }
 
                /*
@@ -1295,11 +1501,14 @@ 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;
-                       t_rate += td->o.rate;
+                       m_rate += td->o.ratemin[0] + td->o.ratemin[1];
+                       t_rate += td->o.rate[0] + td->o.rate[1];
                        todo--;
                        fio_mutex_up(td->mutex);
                }
@@ -1353,10 +1562,11 @@ int main(int argc, char *argv[])
        }
 
        startup_mutex = fio_mutex_init(0);
+       writeout_mutex = fio_mutex_init(1);
 
        set_genesis_time();
 
-       disk_util_timer_arm();
+       status_timer_arm();
 
        run_threads();
 
@@ -1370,5 +1580,6 @@ int main(int argc, char *argv[])
        }
 
        fio_mutex_remove(startup_mutex);
+       fio_mutex_remove(writeout_mutex);
        return exit_value;
 }