Add support for reserved keywords
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 0e1eadb35ce2d11555ee1a1ad3213083d136c5c5..dfd15e0f5db1e943a9aec595454cf1f601e97fea 100644 (file)
--- a/fio.c
+++ b/fio.c
 #include <sys/mman.h>
 
 #include "fio.h"
-#include "os.h"
+#include "hash.h"
+#include "smalloc.h"
+#include "verify.h"
+#include "diskutil.h"
 
-static unsigned long page_mask;
-#define ALIGN(buf)     \
+unsigned long page_mask;
+unsigned long page_size;
+
+#define PAGE_ALIGN(buf)        \
        (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
 
 int groupid = 0;
 int thread_number = 0;
+int nr_process = 0;
+int nr_thread = 0;
 int shm_id = 0;
 int temp_stall_ts;
+unsigned long done_secs = 0;
 
-static volatile int startup_sem;
+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", (int) td->pid,
+                                               td->runstate, runstate);
        td->runstate = runstate;
 }
 
-static void terminate_threads(int group_id, int forced_kill)
+static void terminate_threads(int group_id)
 {
        struct thread_data *td;
        int i;
 
+       dprint(FD_PROCESS, "terminate group_id=%d\n", 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, (int) td->pid);
                        td->terminate = 1;
-                       td->start_delay = 0;
-                       if (forced_kill)
-                               td_set_runstate(td, TD_EXITED);
+                       td->o.start_delay = 0;
+
+                       /*
+                        * if the thread is running, just let it exit
+                        */
+                       if (td->runstate < TD_RUNNING)
+                               kill(td->pid, SIGQUIT);
+                       else {
+                               struct ioengine_ops *ops = td->io_ops;
+
+                               if (ops && (ops->flags & FIO_SIGQUIT))
+                                       kill(td->pid, SIGQUIT);
+                       }
                }
        }
 }
 
-static void sig_handler(int sig)
+static void status_timer_arm(void)
 {
-       switch (sig) {
-               case SIGALRM:
-                       update_io_ticks();
-                       disk_util_timer_arm();
-                       print_thread_status();
-                       break;
-               default:
-                       printf("\nfio: terminating on signal %d\n", sig);
-                       fflush(stdout);
-                       terminate_threads(TERMINATE_ALL, 0);
-                       break;
+       itimer.it_value.tv_sec = 0;
+       itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
+       setitimer(ITIMER_REAL, &itimer, NULL);
+}
+
+static void sig_alrm(int fio_unused sig)
+{
+       if (threads) {
+               update_io_ticks();
+               print_thread_status();
+               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);
+       }
+}
+
+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;
-       int ddir = td->ddir;
+       unsigned int ratemin = 0;
+       unsigned int rate_iops = 0;
+       unsigned int rate_iops_min = 0;
+
+       if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
+               return 0;
 
        /*
         * allow a 2 second settle period in the beginning
@@ -106,31 +196,82 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
        if (mtime_since(&td->start, now) < 2000)
                return 0;
 
+       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) {
-               spent = mtime_since(&td->lastrate, now);
-               if (spent < td->ratecycle)
+       if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
+               spent = mtime_since(&td->lastrate[ddir], now);
+               if (spent < td->o.ratecycle)
                        return 0;
 
-               rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
-               if (rate < td->ratemin) {
-                       fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
-                       return 1;
+               if (td->o.rate[ddir]) {
+                       /*
+                        * check bandwidth specified rate
+                        */
+                       if (bytes < td->rate_bytes[ddir]) {
+                               log_err("%s: min rate %u not met\n", td->o.name,
+                                                               ratemin);
+                               return 1;
+                       } else {
+                               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"
+                                               " %luKB/sec\n", td->o.name,
+                                                       ratemin, rate);
+                                       return 1;
+                               }
+                       }
+               } else {
+                       /*
+                        * checks iops specified rate
+                        */
+                       if (iops < rate_iops) {
+                               log_err("%s: min iops rate %u not met\n",
+                                               td->o.name, rate_iops);
+                               return 1;
+                       } else {
+                               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,
+                                                       rate_iops_min, rate);
+                               }
+                       }
                }
        }
 
-       td->rate_bytes = td->this_io_bytes[ddir];
-       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->timeout)
+       if (!td->o.timeout)
                return 0;
-       if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
+       if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
                return 1;
 
        return 0;
@@ -142,7 +283,7 @@ 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;
 
@@ -157,8 +298,8 @@ 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
@@ -203,7 +344,7 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 requeue:
        ret = td_io_queue(td, io_u);
        if (ret < 0) {
-               td_verror(td, io_u->error);
+               td_verror(td, io_u->error, "td_io_queue");
                put_io_u(td, io_u);
                return 1;
        } else if (ret == FIO_Q_QUEUED) {
@@ -211,7 +352,7 @@ requeue:
                        return 1;
        } else if (ret == FIO_Q_COMPLETED) {
                if (io_u->error) {
-                       td_verror(td, io_u->error);
+                       td_verror(td, io_u->error, "td_io_queue");
                        return 1;
                }
 
@@ -226,21 +367,75 @@ 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);
+}
+
+static int break_on_this_error(struct thread_data *td, int *retptr)
+{
+       int ret = *retptr;
+
+       if (ret < 0 || td->error) {
+               int err;
+
+               if (!td->o.continue_on_error)
+                       return 1;
+
+               if (ret < 0)
+                       err = -ret;
+               else
+                       err = td->error;
+
+               if (td_non_fatal_error(err)) {
+                       /*
+                        * Continue with the I/Os in case of
+                        * a non fatal error.
+                        */
+                       update_error_count(td, err);
+                       td_clear_error(td);
+                       *retptr = 0;
+                       return 0;
+               } else if (td->o.fill_device && err == ENOSPC) {
+                       /*
+                        * We expect to hit this error if
+                        * fill_device option is set.
+                        */
+                       td_clear_error(td);
+                       td->terminate = 1;
+                       return 1;
+               } else {
+                       /*
+                        * Stop the I/O in case of a fatal
+                        * error.
+                        */
+                       update_error_count(td, err);
+                       return 1;
+               }
+       }
+
+       return 0;
+}
+
 /*
- * The main verify engine. Runs over the writes we previusly submitted,
+ * The main verify engine. Runs over the writes we previously submitted,
  * reads the blocks back in, and checks the crc/md5 of the data.
  */
 static void do_verify(struct thread_data *td)
 {
        struct fio_file *f;
        struct io_u *io_u;
-       int ret, i, min_events;
+       int ret, min_events;
+       unsigned int i;
 
        /*
         * sync io first and invalidate cache, to make sure we really
         * read from disk.
         */
        for_each_file(td, f, i) {
+               if (!fio_file_open(f))
+                       continue;
                if (fio_io_sync(td, f))
                        break;
                if (file_invalidate_cache(td, f))
@@ -254,15 +449,19 @@ static void do_verify(struct thread_data *td)
 
        io_u = NULL;
        while (!td->terminate) {
-               io_u = __get_io_u(td);
-               if (!io_u)
-                       break;
+               int ret2, full;
 
-               if (runtime_exceeded(td, &io_u->start_time)) {
-                       put_io_u(td, io_u);
+               update_tv_cache(td);
+
+               if (runtime_exceeded(td, &td->tv_cache)) {
+                       td->terminate = 1;
                        break;
                }
 
+               io_u = __get_io_u(td);
+               if (!io_u)
+                       break;
+
                if (get_next_verify(td, io_u)) {
                        put_io_u(td, io_u);
                        break;
@@ -272,141 +471,191 @@ static void do_verify(struct thread_data *td)
                        put_io_u(td, io_u);
                        break;
                }
-requeue:
-               ret = td_io_queue(td, io_u);
 
+               if (td->o.verify_async)
+                       io_u->end_io = verify_io_u_async;
+               else
+                       io_u->end_io = verify_io_u;
+
+               ret = td_io_queue(td, io_u);
                switch (ret) {
                case FIO_Q_COMPLETED:
-                       if (io_u->error)
+                       if (io_u->error) {
                                ret = -io_u->error;
-                       if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
+                               clear_io_u(td, io_u);
+                       } else if (io_u->resid) {
                                int bytes = io_u->xfer_buflen - io_u->resid;
+                               struct fio_file *f = io_u->file;
+
+                               /*
+                                * zero read, fail
+                                */
+                               if (!bytes) {
+                                       td_verror(td, EIO, "full resid");
+                                       put_io_u(td, io_u);
+                                       break;
+                               }
 
                                io_u->xfer_buflen = io_u->resid;
                                io_u->xfer_buf += bytes;
-                               goto requeue;
+                               io_u->offset += bytes;
+
+                               td->ts.short_io_u[io_u->ddir]++;
+
+                               if (io_u->offset == f->real_file_size)
+                                       goto sync_done;
+
+                               requeue_io_u(td, &io_u);
+                       } else {
+sync_done:
+                               ret = io_u_sync_complete(td, io_u, NULL);
+                               if (ret < 0)
+                                       break;
                        }
-                       ret = io_u_sync_complete(td, io_u, verify_io_u);
-                       if (ret < 0)
-                               break;
                        continue;
                case FIO_Q_QUEUED:
                        break;
                case FIO_Q_BUSY:
                        requeue_io_u(td, &io_u);
-                       ret = td_io_commit(td);
+                       ret2 = td_io_commit(td);
+                       if (ret2 < 0)
+                               ret = ret2;
                        break;
                default:
                        assert(ret < 0);
-                       td_verror(td, -ret);
+                       td_verror(td, -ret, "td_io_queue");
                        break;
                }
 
-               if (ret < 0 || td->error)
+               if (break_on_this_error(td, &ret))
                        break;
 
                /*
                 * 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) {
-                       min_events = 1;
-
-                       if (td->cur_depth > td->iodepth_low)
-                               min_events = td->cur_depth - td->iodepth_low;
+               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;
+
+                       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));
                }
-
-               /*
-                * 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, verify_io_u) < 0)
+               if (ret < 0)
                        break;
        }
 
-       if (td->cur_depth)
+       if (!td->error) {
+               min_events = td->cur_depth;
+
+               if (min_events)
+                       ret = io_u_queued_complete(td, min_events, NULL);
+       } else
                cleanup_pending_aio(td);
 
        td_set_runstate(td, TD_RUNNING);
 }
 
-/*
- * Not really an io thread, all it does is burn CPU cycles in the specified
- * manner.
- */
-static void do_cpuio(struct thread_data *td)
-{
-       struct timeval e;
-       int split = 100 / td->cpuload;
-       int i = 0;
-
-       while (!td->terminate) {
-               fio_gettime(&e, NULL);
-
-               if (runtime_exceeded(td, &e))
-                       break;
-
-               if (!(i % split))
-                       __usec_sleep(10000);
-               else
-                       usec_sleep(td, 10000);
-
-               i++;
-       }
-}
-
 /*
  * Main IO worker function. It retrieves io_u's to process and queues
  * and reaps them, checking for rate and errors along the way.
  */
 static void do_io(struct thread_data *td)
 {
-       struct timeval s;
-       unsigned long usec;
-       int i, ret = 0;
+       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->io_size) {
+       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, full;
 
                if (td->terminate)
                        break;
 
+               update_tv_cache(td);
+
+               if (runtime_exceeded(td, &td->tv_cache)) {
+                       td->terminate = 1;
+                       break;
+               }
+
                io_u = get_io_u(td);
                if (!io_u)
                        break;
 
-               memcpy(&s, &io_u->start_time, sizeof(s));
+               /*
+                * Add verification end_io handler, if asked to verify
+                * a previously written file.
+                */
+               if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
+                       if (td->o.verify_async)
+                               io_u->end_io = verify_io_u_async;
+                       else
+                               io_u->end_io = verify_io_u;
+                       td_set_runstate(td, TD_VERIFYING);
+               } else if (in_ramp_time(td))
+                       td_set_runstate(td, TD_RAMP);
+               else
+                       td_set_runstate(td, TD_RUNNING);
 
-               if (runtime_exceeded(td, &s)) {
-                       put_io_u(td, io_u);
-                       break;
-               }
-requeue:
                ret = td_io_queue(td, io_u);
-
                switch (ret) {
                case FIO_Q_COMPLETED:
                        if (io_u->error) {
-                               ret = io_u->error;
-                               break;
-                       }
-                       if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
+                               ret = -io_u->error;
+                               clear_io_u(td, io_u);
+                       } else if (io_u->resid) {
                                int bytes = io_u->xfer_buflen - io_u->resid;
+                               struct fio_file *f = io_u->file;
+
+                               /*
+                                * zero read, fail
+                                */
+                               if (!bytes) {
+                                       td_verror(td, EIO, "full resid");
+                                       put_io_u(td, io_u);
+                                       break;
+                               }
 
                                io_u->xfer_buflen = io_u->resid;
                                io_u->xfer_buf += bytes;
-                               goto requeue;
+                               io_u->offset += bytes;
+
+                               td->ts.short_io_u[io_u->ddir]++;
+
+                               if (io_u->offset == f->real_file_size)
+                                       goto sync_done;
+
+                               requeue_io_u(td, &io_u);
+                       } else {
+sync_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;
                        }
-                       fio_gettime(&comp_time, NULL);
-                       bytes_done = io_u_sync_complete(td, io_u, NULL);
-                       if (bytes_done < 0)
-                               ret = bytes_done;
                        break;
                case FIO_Q_QUEUED:
                        /*
@@ -419,7 +668,9 @@ requeue:
                        break;
                case FIO_Q_BUSY:
                        requeue_io_u(td, &io_u);
-                       ret = td_io_commit(td);
+                       ret2 = td_io_commit(td);
+                       if (ret2 < 0)
+                               ret = ret2;
                        break;
                default:
                        assert(ret < 0);
@@ -427,148 +678,169 @@ requeue:
                        break;
                }
 
-               if (ret < 0 || td->error)
+               if (break_on_this_error(td, &ret))
                        break;
 
                /*
                 * See if we need to complete some commands
                 */
-               if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
-                       min_evts = 0;
-                       if (queue_full(td) || ret == FIO_Q_BUSY) {
+               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;
 
-                               if (td->cur_depth > td->iodepth_low)
-                                       min_evts = td->cur_depth - td->iodepth_low;
-                       }
+                       if (__should_check_rate(td, 0) ||
+                           __should_check_rate(td, 1))
+                               fio_gettime(&comp_time, NULL);
 
-                       fio_gettime(&comp_time, NULL);
-                       bytes_done = io_u_queued_complete(td, min_evts, NULL);
-                       if (bytes_done < 0)
-                               break;
+                       do {
+                               ret = io_u_queued_complete(td, min_evts, bytes_done);
+                               if (ret < 0)
+                                       break;
+
+                       } while (full && (td->cur_depth > td->o.iodepth_low));
                }
 
-               if (!bytes_done)
+               if (ret < 0)
+                       break;
+               if (!(bytes_done[0] + bytes_done[1]))
                        continue;
 
-               /*
-                * 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);
-
-               rate_throttle(td, usec, bytes_done, td->ddir);
-
-               if (check_min_rate(td, &comp_time)) {
-                       if (exitall_on_terminate)
-                               terminate_threads(td->groupid, 0);
-                       td_verror(td, ENODATA);
-                       break;
+               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->thinktime) {
+               if (td->o.thinktime) {
                        unsigned long long b;
 
                        b = td->io_blocks[0] + td->io_blocks[1];
-                       if (!(b % td->thinktime_blocks)) {
+                       if (!(b % td->o.thinktime_blocks)) {
                                int left;
 
-                               if (td->thinktime_spin)
-                                       __usec_sleep(td->thinktime_spin);
+                               if (td->o.thinktime_spin)
+                                       usec_spin(td->o.thinktime_spin);
 
-                               left = td->thinktime - td->thinktime_spin;
+                               left = td->o.thinktime - td->o.thinktime_spin;
                                if (left)
                                        usec_sleep(td, left);
                        }
                }
        }
 
+       if (td->o.fill_device && td->error == ENOSPC) {
+               td->error = 0;
+               td->terminate = 1;
+       }
        if (!td->error) {
                struct fio_file *f;
 
-               if (td->cur_depth)
-                       cleanup_pending_aio(td);
+               i = td->cur_depth;
+               if (i)
+                       ret = io_u_queued_complete(td, i, NULL);
 
-               if (should_fsync(td) && td->end_fsync) {
+               if (should_fsync(td) && td->o.end_fsync) {
                        td_set_runstate(td, TD_FSYNCING);
-                       for_each_file(td, f, i)
+
+                       for_each_file(td, f, i) {
+                               if (!fio_file_open(f))
+                                       continue;
                                fio_io_sync(td, f);
+                       }
                }
-       }
+       } else
+               cleanup_pending_aio(td);
+
+       /*
+        * stop job if we failed doing any IO
+        */
+       if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
+               td->done = 1;
 }
 
 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);
        }
 
        free_io_mem(td);
 }
 
-/*
- * "randomly" fill the buffer contents
- */
-static void fill_rand_buf(struct io_u *io_u, int max_bs)
-{
-       int *ptr = io_u->buf;
-
-       while ((void *) ptr - io_u->buf < max_bs) {
-               *ptr = rand() * 0x9e370001;
-               ptr++;
-       }
-}
-
 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;
 
-       if (td->io_ops->flags & FIO_CPUIO)
-               return 0;
+       max_units = td->o.iodepth;
+       max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
+       td->orig_buffer_size = (unsigned long long) max_bs
+                                       * (unsigned long long) max_units;
 
-       if (td->io_ops->flags & FIO_SYNCIO)
-               max_units = 1;
-       else
-               max_units = td->iodepth;
+       if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
+               unsigned long bs;
 
-       max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
-       td->orig_buffer_size = max_bs * max_units;
+               bs = td->orig_buffer_size + td->o.hugepage_size - 1;
+               td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1);
+       }
 
-       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 += page_mask;
+       if (td->orig_buffer_size != (size_t) td->orig_buffer_size) {
+               log_err("fio: IO memory too large. Reduce max_bs or iodepth\n");
+               return 1;
+       }
 
        if (allocate_io_mem(td))
                return 1;
 
-       p = ALIGN(td->orig_buffer);
+       if (td->o.odirect || td->o.mem_align)
+               p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
+       else
+               p = td->orig_buffer;
+
+       cl_align = os_cache_line_size();
+
        for (i = 0; i < max_units; i++) {
-               io_u = malloc(sizeof(*io_u));
+               void *ptr;
+
+               if (td->terminate)
+                       return 1;
+
+               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);
+               dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 
-               io_u->buf = p + max_bs * i;
-               if (td_write(td) || td_rw(td))
-                       fill_rand_buf(io_u, max_bs);
+               if (!(td->io_ops->flags & FIO_NOIO)) {
+                       io_u->buf = p + max_bs * i;
+                       dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
+
+                       if (td_write(td) && !td->o.refill_buffers)
+                               io_u_fill_buffer(td, io_u, max_bs);
+               }
 
                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;
 }
 
@@ -578,23 +850,28 @@ static int switch_ioscheduler(struct thread_data *td)
        FILE *f;
        int ret;
 
-       if (td->io_ops->flags & FIO_CPUIO)
+       if (td->io_ops->flags & FIO_DISKLESSIO)
                return 0;
 
        sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
 
        f = fopen(tmp, "r+");
        if (!f) {
-               td_verror(td, errno);
+               if (errno == ENOENT) {
+                       log_err("fio: os or kernel doesn't support IO scheduler"
+                               " switching\n");
+                       return 0;
+               }
+               td_verror(td, errno, "fopen iosched");
                return 1;
        }
 
        /*
         * Set io scheduler.
         */
-       ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
+       ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f);
        if (ferror(f) || ret != 1) {
-               td_verror(td, errno);
+               td_verror(td, errno, "fwrite");
                fclose(f);
                return 1;
        }
@@ -606,15 +883,15 @@ static int switch_ioscheduler(struct thread_data *td)
         */
        ret = fread(tmp, 1, sizeof(tmp), f);
        if (ferror(f) || ret < 0) {
-               td_verror(td, errno);
+               td_verror(td, errno, "fread");
                fclose(f);
                return 1;
        }
 
-       sprintf(tmp2, "[%s]", td->ioscheduler);
+       sprintf(tmp2, "[%s]", td->o.ioscheduler);
        if (!strstr(tmp, tmp2)) {
-               log_err("fio: io scheduler %s not found\n", td->ioscheduler);
-               td_verror(td, EINVAL);
+               log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
+               td_verror(td, EINVAL, "iosched_switch");
                fclose(f);
                return 1;
        }
@@ -623,27 +900,94 @@ static int switch_ioscheduler(struct thread_data *td)
        return 0;
 }
 
-static void clear_io_state(struct thread_data *td)
+static int keep_running(struct thread_data *td)
 {
-       struct fio_file *f;
-       int i;
+       unsigned long long io_done;
 
+       if (td->done)
+               return 0;
+       if (td->o.time_based)
+               return 1;
+       if (td->o.loops) {
+               td->o.loops--;
+               return 1;
+       }
+
+       io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
+                       + td->io_skip_bytes;
+       if (io_done < td->o.size)
+               return 1;
+
+       return 0;
+}
+
+static void reset_io_counters(struct thread_data *td)
+{
        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_bytes[1] = 0;
+       td->rate_blocks[0] = td->rate_blocks[1] = 0;
 
        td->last_was_sync = 0;
 
-       for_each_file(td, f, i) {
-               f->last_completed_pos = 0;
+       /*
+        * reset file done count if we are to start over
+        */
+       if (td->o.time_based || td->o.loops)
+               td->nr_done_files = 0;
 
-               f->last_pos = 0;
-               if (td->io_ops->flags & FIO_SYNCIO)
-                       lseek(f->fd, SEEK_SET, 0);
+       /*
+        * Set the same seed to get repeatable runs
+        */
+       td_fill_rand_seeds(td);
+}
 
-               if (f->file_map)
-                       memset(f->file_map, 0, f->num_maps * sizeof(long));
+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;
 }
 
 /*
@@ -652,101 +996,153 @@ static void clear_io_state(struct thread_data *td)
  */
 static void *thread_main(void *data)
 {
-       unsigned long long runtime[2];
+       unsigned long long runtime[2], elapsed;
        struct thread_data *td = data;
+       pthread_condattr_t attr;
+       int clear_state;
 
-       if (!td->use_thread)
+       if (!td->o.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_u_requeues);
-       INIT_LIST_HEAD(&td->io_hist_list);
-       INIT_LIST_HEAD(&td->io_log_list);
+       dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
+
+       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);
+       INIT_FLIST_HEAD(&td->verify_list);
+       pthread_mutex_init(&td->io_u_lock, NULL);
+       td->io_hist_tree = RB_ROOT;
+
+       pthread_condattr_init(&attr);
+       pthread_cond_init(&td->verify_cond, &attr);
+       pthread_cond_init(&td->free_cond, &attr);
+
+       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
+        * eating a file descriptor
+        */
+       fio_mutex_remove(td->mutex);
+
+       /*
+        * May alter parameters that init_io_u() will use, so we need to
+        * do this first.
+        */
+       if (init_iolog(td))
+               goto err;
 
        if (init_io_u(td))
                goto err;
 
-       if (fio_setaffinity(td) == -1) {
-               td_verror(td, errno);
+       if (td->o.verify_async && verify_async_init(td))
                goto err;
-       }
 
-       if (init_iolog(td))
+       if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
+               td_verror(td, errno, "cpu_set_affinity");
                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->pid, td->o.cpumask) == -1) {
+                       td_verror(td, errno, "cpu_set_affinity");
+                       goto err;
+               }
+       }
 
-       if (td->ioprio) {
+       if (td->ioprio_set) {
                if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
-                       td_verror(td, errno);
+                       td_verror(td, errno, "ioprio_set");
                        goto err;
                }
        }
 
-       if (nice(td->nice) == -1) {
-               td_verror(td, errno);
+       if (nice(td->o.nice) == -1) {
+               td_verror(td, errno, "nice");
                goto err;
        }
 
-       if (init_random_state(td))
+       if (td->o.ioscheduler && switch_ioscheduler(td))
                goto err;
 
-       if (td->ioscheduler && switch_ioscheduler(td))
+       if (!td->o.create_serialize && setup_files(td))
                goto err;
 
-       td_set_runstate(td, TD_INITIALIZED);
-       fio_sem_up(&startup_sem);
-       fio_sem_down(&td->mutex);
-
-       if (!td->create_serialize && setup_files(td))
-               goto err;
-       if (open_files(td))
+       if (td_io_init(td))
                goto err;
 
-       /*
-        * Do this late, as some IO engines would like to have the
-        * files setup prior to initializing structures.
-        */
-       if (td_io_init(td))
+       if (init_random_map(td))
                goto err;
 
-       if (td->exec_prerun) {
-               if (system(td->exec_prerun) < 0)
+       if (td->o.exec_prerun) {
+               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;
-       while (td->loops--) {
+       clear_state = 0;
+       while (keep_running(td)) {
                fio_gettime(&td->start, NULL);
-               memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
+               memcpy(&td->ts.stat_sample_time[0], &td->start,
+                               sizeof(td->start));
+               memcpy(&td->ts.stat_sample_time[1], &td->start,
+                               sizeof(td->start));
+               memcpy(&td->tv_cache, &td->start, sizeof(td->start));
 
-               if (td->ratemin)
-                       memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
+               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);
 
-               clear_io_state(td);
                prune_io_piece_log(td);
 
-               if (td->io_ops->flags & FIO_CPUIO)
-                       do_cpuio(td);
-               else
-                       do_io(td);
+               do_io(td);
 
-               runtime[td->ddir] += utime_since_now(&td->start);
-               if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
-                       runtime[td->ddir ^ 1] = runtime[td->ddir];
+               clear_state = 1;
+
+               if (td_read(td) && td->io_bytes[DDIR_READ]) {
+                       elapsed = utime_since_now(&td->start);
+                       runtime[DDIR_READ] += elapsed;
+               }
+               if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
+                       elapsed = utime_since_now(&td->start);
+                       runtime[DDIR_WRITE] += elapsed;
+               }
 
                if (td->error || td->terminate)
                        break;
 
-               if (td->verify == VERIFY_NONE)
+               if (!td->o.do_verify ||
+                   td->o.verify == VERIFY_NONE ||
+                   (td->io_ops->flags & FIO_UNIDIR))
                        continue;
 
                clear_io_state(td);
+
                fio_gettime(&td->start, NULL);
 
                do_verify(td);
@@ -758,32 +1154,65 @@ static void *thread_main(void *data)
        }
 
        update_rusage_stat(td);
-       fio_gettime(&td->end_time, NULL);
-       td->runtime[0] = runtime[0] / 1000;
-       td->runtime[1] = runtime[1] / 1000;
-
-       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->write_iolog_file)
-               write_iolog_close(td);
-       if (td->exec_postrun) {
-               if (system(td->exec_postrun) < 0)
-                       log_err("fio: postrun %s failed\n", td->exec_postrun);
+       td->ts.runtime[0] = (runtime[0] + 999) / 1000;
+       td->ts.runtime[1] = (runtime[1] + 999) / 1000;
+       td->ts.total_run_time = mtime_since_now(&td->epoch);
+       td->ts.io_bytes[0] = td->io_bytes[0];
+       td->ts.io_bytes[1] = td->io_bytes[1];
+
+       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, 0);
+               terminate_threads(td->groupid);
 
 err:
        if (td->error)
-               printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
-       close_files(td);
+               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");
+       }
+
+       if (td->o.verify_async)
+               verify_async_exit(td);
+
+       /*
+        * do this very late, it will log file closing as well
+        */
+       if (td->o.write_iolog_file)
+               write_iolog_close(td);
+
+       options_mem_free(td);
        td_set_runstate(td, TD_EXITED);
        return (void *) (unsigned long) td->error;
 }
@@ -817,12 +1246,12 @@ static int 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, pending, status, ret;
+       int i, cputhreads, realthreads, pending, status, ret;
 
        /*
         * reap exited threads (TD_EXITED -> TD_REAPED)
         */
-       pending = cputhreads = 0;
+       realthreads = pending = cputhreads = 0;
        for_each_td(td, i) {
                int flags = 0;
 
@@ -830,12 +1259,18 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                 * ->io_ops is NULL for a thread that has closed its
                 * io engine
                 */
-               if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
+               if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
                        cputhreads++;
+               else
+                       realthreads++;
 
-               if (!td->pid || td->runstate == TD_REAPED)
+               if (!td->pid) {
+                       pending++;
                        continue;
-               if (td->use_thread) {
+               }
+               if (td->runstate == TD_REAPED)
+                       continue;
+               if (td->o.use_thread) {
                        if (td->runstate == TD_EXITED) {
                                td_set_runstate(td, TD_REAPED);
                                goto reaped;
@@ -853,7 +1288,8 @@ 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;
                        }
@@ -862,7 +1298,9 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                        if (WIFSIGNALED(status)) {
                                int sig = WTERMSIG(status);
 
-                               log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
+                               if (sig != SIGQUIT)
+                                       log_err("fio: pid=%d, got signal=%d\n",
+                                                       (int) td->pid, sig);
                                td_set_runstate(td, TD_REAPED);
                                goto reaped;
                        }
@@ -878,25 +1316,63 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                /*
                 * thread is not dead, continue
                 */
+               pending++;
                continue;
 reaped:
-               if (td->use_thread) {
-                       long ret;
-
-                       if (pthread_join(td->thread, (void *) &ret))
-                               perror("pthread_join");
-               }
-
                (*nr_running)--;
-               (*m_rate) -= td->ratemin;
-               (*t_rate) -= td->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--;
 
                if (td->error)
                        exit_value++;
+
+               done_secs += mtime_since_now(&td->epoch) / 1000;
        }
 
-       if (*nr_running == cputhreads && !pending)
-               terminate_threads(TERMINATE_ALL, 0);
+       if (*nr_running == cputhreads && !pending && realthreads)
+               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;
 }
 
 /*
@@ -911,13 +1387,25 @@ static void run_threads(void)
        if (fio_pin_memory())
                return;
 
+       if (fio_gtod_offload && fio_start_gtod_thread())
+               return;
+
        if (!terse_output) {
-               printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
+               printf("Starting ");
+               if (nr_thread)
+                       printf("%d thread%s", nr_thread,
+                                               nr_thread > 1 ? "s" : "");
+               if (nr_process) {
+                       if (nr_thread)
+                               printf(" and ");
+                       printf("%d process%s", nr_process,
+                                               nr_process > 1 ? "es" : "");
+               }
+               printf("\n");
                fflush(stdout);
        }
 
-       signal(SIGINT, sig_handler);
-       signal(SIGALRM, sig_handler);
+       set_sig_handlers();
 
        todo = thread_number;
        nr_running = 0;
@@ -927,7 +1415,7 @@ static void run_threads(void)
        for_each_td(td, i) {
                print_status_init(td->thread_number - 1);
 
-               if (!td->create_serialize) {
+               if (!td->o.create_serialize) {
                        init_disk_util(td);
                        continue;
                }
@@ -940,9 +1428,23 @@ 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 {
+                       struct fio_file *f;
+                       unsigned int i;
+
+                       /*
+                        * for sharing to work, each job must always open
+                        * its own files. so close them, if we opened them
+                        * for creation
+                        */
+                       for_each_file(td, f, i) {
+                               if (fio_file_open(f))
+                                       td_io_close_file(td, f);
+                       }
                }
 
                init_disk_util(td);
@@ -971,15 +1473,18 @@ static void run_threads(void)
                                continue;
                        }
 
-                       if (td->start_delay) {
+                       if (td->o.start_delay) {
                                spent = mtime_since_genesis();
 
-                               if (td->start_delay * 1000 > spent)
+                               if (td->o.start_delay * 1000 > spent)
                                        continue;
                        }
 
-                       if (td->stonewall && (nr_started || nr_running))
+                       if (td->o.stonewall && (nr_started || nr_running)) {
+                               dprint(FD_PROCESS, "%s: stonewall wait\n",
+                                                       td->o.name);
                                break;
+                       }
 
                        /*
                         * Set state to created. Thread will transition
@@ -987,23 +1492,44 @@ static void run_threads(void)
                         */
                        td_set_runstate(td, TD_CREATED);
                        map[this_jobs++] = td;
-                       fio_sem_init(&startup_sem, 1);
                        nr_started++;
 
-                       if (td->use_thread) {
-                               if (pthread_create(&td->thread, NULL, thread_main, td)) {
-                                       perror("thread_create");
+                       if (td->o.use_thread) {
+                               int ret;
+
+                               dprint(FD_PROCESS, "will pthread_create\n");
+                               ret = pthread_create(&td->thread, NULL,
+                                                       thread_main, td);
+                               if (ret) {
+                                       log_err("pthread_create: %s\n",
+                                                       strerror(ret));
                                        nr_started--;
+                                       break;
                                }
+                               ret = pthread_detach(td->thread);
+                               if (ret)
+                                       log_err("pthread_detach: %s",
+                                                       strerror(ret));
                        } else {
-                               if (fork())
-                                       fio_sem_down(&startup_sem);
-                               else {
+                               pid_t pid;
+                               dprint(FD_PROCESS, "will fork\n");
+                               pid = fork();
+                               if (!pid) {
                                        int ret = fork_main(shm_id, i);
 
-                                       exit(ret);
-                               }
+                                       _exit(ret);
+                               } else if (i == fio_debug_jobno)
+                                       *fio_debug_jobp = pid;
+                       }
+                       dprint(FD_MUTEX, "wait on startup_mutex\n");
+                       if (fio_mutex_down_timeout(startup_mutex, 10)) {
+                               log_err("fio: job startup hung? exiting.\n");
+                               terminate_threads(TERMINATE_ALL);
+                               fio_abort = 1;
+                               nr_started--;
+                               break;
                        }
+                       dprint(FD_MUTEX, "done waiting on startup_mutex\n");
                }
 
                /*
@@ -1052,13 +1578,16 @@ 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->ratemin;
-                       t_rate += td->rate;
+                       m_rate += td->o.ratemin[0] + td->o.ratemin[1];
+                       t_rate += td->o.rate[0] + td->o.rate[1];
                        todo--;
-                       fio_sem_up(&td->mutex);
+                       fio_mutex_up(td->mutex);
                }
 
                reap_threads(&nr_running, &t_rate, &m_rate);
@@ -1080,6 +1609,8 @@ int main(int argc, char *argv[])
 {
        long ps;
 
+       sinit();
+
        /*
         * We need locale for number printing, if it isn't set then just
         * go with the US format.
@@ -1087,40 +1618,47 @@ int main(int argc, char *argv[])
        if (!getenv("LC_NUMERIC"))
                setlocale(LC_NUMERIC, "en_US");
 
-       if (parse_options(argc, argv))
-               return 1;
-
-       if (!thread_number) {
-               log_err("Nothing to do\n");
-               return 1;
-       }
-
        ps = sysconf(_SC_PAGESIZE);
        if (ps < 0) {
                log_err("Failed to get page size\n");
                return 1;
        }
 
+       page_size = ps;
        page_mask = ps - 1;
 
+       fio_keywords_init();
+
+       if (parse_options(argc, argv))
+               return 1;
+
+       if (!thread_number)
+               return 0;
+
        if (write_bw_log) {
                setup_log(&agg_io_log[DDIR_READ]);
                setup_log(&agg_io_log[DDIR_WRITE]);
        }
 
+       startup_mutex = fio_mutex_init(0);
+       writeout_mutex = fio_mutex_init(1);
+
        set_genesis_time();
 
-       disk_util_timer_arm();
+       status_timer_arm();
 
        run_threads();
 
        if (!fio_abort) {
                show_run_stats();
                if (write_bw_log) {
-                       __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
-                       __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
+                       __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log");
+                       __finish_log(agg_io_log[DDIR_WRITE],
+                                       "agg-write_bw.log");
                }
        }
 
+       fio_mutex_remove(startup_mutex);
+       fio_mutex_remove(writeout_mutex);
        return exit_value;
 }