Fix potential crash in terminate_threads()
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 7460492..d345064 100644 (file)
--- a/fio.c
+++ b/fio.c
 #include <sys/mman.h>
 
 #include "fio.h"
-#include "os.h"
+#include "hash.h"
 
-static unsigned long page_mask;
+unsigned long page_mask;
+unsigned long page_size;
 #define 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;
 
-static volatile int startup_sem;
+static struct fio_mutex *startup_mutex;
 static volatile int fio_abort;
 static int exit_value;
 
@@ -57,20 +60,34 @@ struct io_log *agg_io_log[2];
 
 static inline void td_set_runstate(struct thread_data *td, int runstate)
 {
+       dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", 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;
 
        for_each_td(td, i) {
                if (group_id == TERMINATE_ALL || groupid == td->groupid) {
+                       dprint(FD_PROCESS, "setting terminate on %d\n",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);
+                       }
                }
        }
 }
@@ -86,7 +103,7 @@ static void sig_handler(int sig)
                default:
                        printf("\nfio: terminating on signal %d\n", sig);
                        fflush(stdout);
-                       terminate_threads(TERMINATE_ALL, 0);
+                       terminate_threads(TERMINATE_ALL);
                        break;
        }
 }
@@ -97,50 +114,80 @@ static void sig_handler(int sig)
 static int check_min_rate(struct thread_data *td, struct timeval *now)
 {
        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;
+
        /*
         * allow a 2 second settle period in the beginning
         */
        if (mtime_since(&td->start, now) < 2000)
                return 0;
 
-       if (td_read(td))
+       if (td_read(td)) {
+               iops += td->io_blocks[DDIR_READ];
                bytes += td->this_io_bytes[DDIR_READ];
-       if (td_write(td))
+       }
+       if (td_write(td)) {
+               iops += td->io_blocks[DDIR_WRITE];
                bytes += td->this_io_bytes[DDIR_WRITE];
+       }
 
        /*
         * if rate blocks is set, sample is running
         */
-       if (td->rate_bytes) {
+       if (td->rate_bytes || td->rate_blocks) {
                spent = mtime_since(&td->lastrate, now);
-               if (spent < td->ratecycle)
+               if (spent < td->o.ratecycle)
                        return 0;
 
-               if (bytes < td->rate_bytes) {
-                       fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
-                       return 1;
+               if (td->o.rate) {
+                       /*
+                        * check bandwidth specified rate
+                        */
+                       if (bytes < td->rate_bytes) {
+                               log_err("%s: min rate %u not met\n", td->o.name, td->o.ratemin);
+                               return 1;
+                       } else {
+                               rate = (bytes - td->rate_bytes) / spent;
+                               if (rate < td->o.ratemin || bytes < td->rate_bytes) {
+                                       log_err("%s: min rate %u not met, got %luKiB/sec\n", td->o.name, td->o.ratemin, rate);
+                                       return 1;
+                               }
+                       }
                } else {
-                       rate = (bytes - td->rate_bytes) / spent;
-                       if (rate < td->ratemin || bytes < td->rate_bytes) {
-                               fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
+                       /*
+                        * checks iops specified rate
+                        */
+                       if (iops < td->o.rate_iops) {
+                               log_err("%s: min iops rate %u not met\n", td->o.name, td->o.rate_iops);
                                return 1;
+                       } else {
+                               rate = (iops - td->rate_blocks) / spent;
+                               if (rate < td->o.rate_iops_min || iops < td->rate_blocks) {
+                                       log_err("%s: min iops rate %u not met, got %lu\n", td->o.name, td->o.rate_iops_min, rate);
+                               }
                        }
                }
        }
 
        td->rate_bytes = bytes;
+       td->rate_blocks = iops;
        memcpy(&td->lastrate, now, sizeof(*now));
        return 0;
 }
 
 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;
@@ -237,20 +284,23 @@ requeue:
 }
 
 /*
- * 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 (!(f->flags & FIO_FILE_OPEN))
+                       continue;
                if (fio_io_sync(td, f))
                        break;
                if (file_invalidate_cache(td, f))
@@ -264,12 +314,15 @@ static void do_verify(struct thread_data *td)
 
        io_u = NULL;
        while (!td->terminate) {
+               int ret2;
+
                io_u = __get_io_u(td);
                if (!io_u)
                        break;
 
                if (runtime_exceeded(td, &io_u->start_time)) {
                        put_io_u(td, io_u);
+                       td->terminate = 1;
                        break;
                }
 
@@ -284,29 +337,50 @@ static void do_verify(struct thread_data *td)
                }
 
                io_u->end_io = verify_io_u;
-requeue:
-               ret = td_io_queue(td, io_u);
 
+               ret = td_io_queue(td, io_u);
                switch (ret) {
                case FIO_Q_COMPLETED:
                        if (io_u->error)
                                ret = -io_u->error;
-                       if (io_u->xfer_buflen != io_u->resid && io_u->resid) {
+                       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, ENODATA, "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;
+                               f->last_completed_pos = io_u->offset;
+
+                               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);
+                               if (ret < 0)
+                                       break;
                        }
-                       ret = io_u_sync_complete(td, 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);
@@ -325,8 +399,8 @@ requeue:
                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;
+                       if (td->cur_depth > td->o.iodepth_low)
+                               min_events = td->cur_depth - td->o.iodepth_low;
                }
 
                /*
@@ -337,37 +411,17 @@ requeue:
                        break;
        }
 
-       if (td->cur_depth)
+       if (!td->error) {
+               min_events = td->cur_depth;
+
+               if (min_events)
+                       ret = io_u_queued_complete(td, min_events);
+       } 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.
@@ -376,15 +430,17 @@ 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);
 
-       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;
                int min_evts = 0;
                struct io_u *io_u;
+               int ret2;
 
                if (td->terminate)
                        break;
@@ -397,28 +453,53 @@ static void do_io(struct thread_data *td)
 
                if (runtime_exceeded(td, &s)) {
                        put_io_u(td, io_u);
+                       td->terminate = 1;
                        break;
                }
-requeue:
-               ret = td_io_queue(td, io_u);
 
+               /*
+                * Add verification end_io handler, if asked to verify
+                * a previously written file.
+                */
+               if (td->o.verify != VERIFY_NONE)
+                       io_u->end_io = verify_io_u;
+
+               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) {
+                       if (io_u->error)
+                               ret = -io_u->error;
+                       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, ENODATA, "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;
+                               f->last_completed_pos = io_u->offset;
+
+                               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:
+                               fio_gettime(&comp_time, NULL);
+                               bytes_done = io_u_sync_complete(td, io_u);
+                               if (bytes_done < 0)
+                                       ret = bytes_done;
                        }
-                       fio_gettime(&comp_time, NULL);
-                       bytes_done = io_u_sync_complete(td, io_u);
-                       if (bytes_done < 0)
-                               ret = bytes_done;
                        break;
                case FIO_Q_QUEUED:
                        /*
@@ -431,7 +512,9 @@ 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);
@@ -450,8 +533,8 @@ requeue:
                        if (queue_full(td) || ret == FIO_Q_BUSY) {
                                min_evts = 1;
 
-                               if (td->cur_depth > td->iodepth_low)
-                                       min_evts = td->cur_depth - td->iodepth_low;
+                               if (td->cur_depth > td->o.iodepth_low)
+                                       min_evts = td->cur_depth - td->o.iodepth_low;
                        }
 
                        fio_gettime(&comp_time, NULL);
@@ -474,40 +557,56 @@ requeue:
 
                if (check_min_rate(td, &comp_time)) {
                        if (exitall_on_terminate)
-                               terminate_threads(td->groupid, 0);
+                               terminate_threads(td->groupid);
                        td_verror(td, ENODATA, "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_sleep(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);
 
-               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 (!(f->flags & FIO_FILE_OPEN))
+                                       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)
@@ -528,14 +627,17 @@ static void cleanup_io_u(struct thread_data *td)
 /*
  * "randomly" fill the buffer contents
  */
-static void fill_rand_buf(struct io_u *io_u, int max_bs)
+static void fill_io_buf(struct thread_data *td, struct io_u *io_u, int max_bs)
 {
-       int *ptr = io_u->buf;
+       long *ptr = io_u->buf;
 
-       while ((void *) ptr - io_u->buf < max_bs) {
-               *ptr = rand() * 0x9e370001;
-               ptr++;
-       }
+       if (!td->o.zero_buffers) {
+               while ((void *) ptr - io_u->buf < max_bs) {
+                       *ptr = rand() * GOLDEN_RATIO_PRIME;
+                       ptr++;
+               }
+       } else
+               memset(ptr, 0, max_bs);
 }
 
 static int init_io_u(struct thread_data *td)
@@ -545,34 +647,39 @@ static int init_io_u(struct thread_data *td)
        int i, max_units;
        char *p;
 
-       if (td->io_ops->flags & FIO_CPUIO)
-               return 0;
-
-       if (td->io_ops->flags & FIO_SYNCIO)
-               max_units = 1;
-       else
-               max_units = td->iodepth;
+       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;
 
-       max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
-       td->orig_buffer_size = max_bs * max_units;
+       if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE)
+               td->orig_buffer_size = (td->orig_buffer_size + td->o.hugepage_size - 1) & ~(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)
+               p = ALIGN(td->orig_buffer);
+       else
+               p = td->orig_buffer;
+
        for (i = 0; i < max_units; i++) {
+               if (td->terminate)
+                       return 1;
                io_u = malloc(sizeof(*io_u));
                memset(io_u, 0, sizeof(*io_u));
                INIT_LIST_HEAD(&io_u->list);
 
-               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;
+
+                       if (td_write(td))
+                               fill_io_buf(td, io_u, max_bs);
+               }
 
                io_u->index = i;
                io_u->flags = IO_U_F_FREE;
@@ -590,21 +697,25 @@ 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, "fopen");
+               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, "fwrite");
                fclose(f);
@@ -623,9 +734,9 @@ static int switch_ioscheduler(struct thread_data *td)
                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);
+               log_err("fio: io scheduler %s not found\n", td->o.ioscheduler);
                td_verror(td, EINVAL, "iosched_switch");
                fclose(f);
                return 1;
@@ -635,27 +746,59 @@ 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)
+{
+       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 int clear_io_state(struct thread_data *td)
 {
        struct fio_file *f;
-       int i;
+       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->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);
+       for_each_file(td, f, i)
+               td_io_close_file(td, f);
 
-               if (f->file_map)
-                       memset(f->file_map, 0, f->num_maps * sizeof(long));
+       ret = 0;
+       for_each_file(td, f, i) {
+               f->flags &= ~FIO_FILE_DONE;
+               ret = td_io_open_file(td, f);
+               if (ret)
+                       break;
        }
+
+       return ret;
 }
 
 /*
@@ -664,67 +807,78 @@ 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;
+       int clear_state;
 
-       if (!td->use_thread)
+       if (!td->o.use_thread)
                setsid();
 
        td->pid = getpid();
 
+       dprint(FD_PROCESS, "jobs pid=%d started\n", 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_hist_list);
        INIT_LIST_HEAD(&td->io_log_list);
+       INIT_LIST_HEAD(&td->io_hist_list);
+       td->io_hist_tree = RB_ROOT;
+
+       td_set_runstate(td, TD_INITIALIZED);
+       fio_mutex_up(startup_mutex);
+       fio_mutex_down(td->mutex);
+
+       /*
+        * 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) {
+       if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
                td_verror(td, errno, "cpu_set_affinity");
                goto err;
        }
 
-       if (init_iolog(td))
-               goto err;
-
-       if (td->ioprio) {
+       if (td->ioprio_set) {
                if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
                        td_verror(td, errno, "ioprio_set");
                        goto err;
                }
        }
 
-       if (nice(td->nice) == -1) {
+       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))
+       if (td_io_init(td))
                goto err;
+
        if (open_files(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 (system(td->o.exec_prerun) < 0)
                        goto err;
        }
 
@@ -733,33 +887,51 @@ static void *thread_main(void *data)
        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));
 
-               if (td->ratemin)
+               if (td->o.ratemin)
                        memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 
-               clear_io_state(td);
+               if (clear_state && clear_io_state(td))
+                       break;
+
                prune_io_piece_log(td);
 
-               if (td->io_ops->flags & FIO_CPUIO)
-                       do_cpuio(td);
-               else
-                       do_io(td);
+               do_io(td);
+
+               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);
 
-               if (td_read(td) && td->io_bytes[DDIR_READ])
-                       runtime[DDIR_READ] += utime_since_now(&td->start);
-               if (td_write(td) && td->io_bytes[DDIR_WRITE])
-                       runtime[DDIR_WRITE] += 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);
+
+                       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);
+               if (clear_io_state(td))
+                       break;
+
                fio_gettime(&td->start, NULL);
 
                do_verify(td);
@@ -771,9 +943,11 @@ 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;
+       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];
 
        if (td->ts.bw_log)
                finish_log(td, td->ts.bw_log, "bw");
@@ -781,15 +955,13 @@ static void *thread_main(void *data)
                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);
+       if (td->o.exec_postrun) {
+               if (system(td->o.exec_postrun) < 0)
+                       log_err("fio: postrun %s failed\n", td->o.exec_postrun);
        }
 
        if (exitall_on_terminate)
-               terminate_threads(td->groupid, 0);
+               terminate_threads(td->groupid);
 
 err:
        if (td->error)
@@ -797,6 +969,14 @@ err:
        close_files(td);
        close_ioengine(td);
        cleanup_io_u(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;
 }
@@ -830,12 +1010,12 @@ 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;
 
@@ -843,12 +1023,18 @@ 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;
@@ -875,7 +1061,8 @@ 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", td->pid, sig);
                                td_set_runstate(td, TD_REAPED);
                                goto reaped;
                        }
@@ -891,25 +1078,30 @@ 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) {
+               if (td->o.use_thread) {
                        long ret;
 
-                       if (pthread_join(td->thread, (void *) &ret))
+                       dprint(FD_PROCESS, "joining tread %d\n", td->pid);
+                       if (pthread_join(td->thread, (void *) &ret)) {
+                               dprint(FD_PROCESS, "join failed %ld\n", ret);
                                perror("pthread_join");
+                       }
                }
 
                (*nr_running)--;
-               (*m_rate) -= td->ratemin;
-               (*t_rate) -= td->rate;
+               (*m_rate) -= td->o.ratemin;
+               (*t_rate) -= td->o.rate;
+               pending--;
 
                if (td->error)
                        exit_value++;
        }
 
-       if (*nr_running == cputhreads && !pending)
-               terminate_threads(TERMINATE_ALL, 0);
+       if (*nr_running == cputhreads && !pending && realthreads)
+               terminate_threads(TERMINATE_ALL);
 }
 
 /*
@@ -925,7 +1117,15 @@ static void run_threads(void)
                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);
        }
 
@@ -940,7 +1140,7 @@ 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;
                }
@@ -984,14 +1184,14 @@ 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))
                                break;
 
                        /*
@@ -1000,23 +1200,24 @@ 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 (td->o.use_thread) {
+                               dprint(FD_PROCESS, "will pthread_create\n");
                                if (pthread_create(&td->thread, NULL, thread_main, td)) {
                                        perror("thread_create");
                                        nr_started--;
+                                       break;
                                }
                        } else {
-                               if (fork())
-                                       fio_sem_down(&startup_sem);
-                               else {
+                               dprint(FD_PROCESS, "will fork\n");
+                               if (!fork()) {
                                        int ret = fork_main(shm_id, i);
 
                                        exit(ret);
                                }
                        }
+                       fio_mutex_down(startup_mutex);
                }
 
                /*
@@ -1068,10 +1269,10 @@ static void run_threads(void)
                        td_set_runstate(td, TD_RUNNING);
                        nr_running++;
                        nr_started--;
-                       m_rate += td->ratemin;
-                       t_rate += td->rate;
+                       m_rate += td->o.ratemin;
+                       t_rate += td->o.rate;
                        todo--;
-                       fio_sem_up(&td->mutex);
+                       fio_mutex_up(td->mutex);
                }
 
                reap_threads(&nr_running, &t_rate, &m_rate);
@@ -1103,10 +1304,8 @@ int main(int argc, char *argv[])
        if (parse_options(argc, argv))
                return 1;
 
-       if (!thread_number) {
-               log_err("Nothing to do\n");
-               return 1;
-       }
+       if (!thread_number)
+               return 0;
 
        ps = sysconf(_SC_PAGESIZE);
        if (ps < 0) {
@@ -1114,6 +1313,7 @@ int main(int argc, char *argv[])
                return 1;
        }
 
+       page_size = ps;
        page_mask = ps - 1;
 
        if (write_bw_log) {
@@ -1121,6 +1321,8 @@ int main(int argc, char *argv[])
                setup_log(&agg_io_log[DDIR_WRITE]);
        }
 
+       startup_mutex = fio_mutex_init(0);
+
        set_genesis_time();
 
        disk_util_timer_arm();
@@ -1135,5 +1337,6 @@ int main(int argc, char *argv[])
                }
        }
 
+       fio_mutex_remove(startup_mutex);
        return exit_value;
 }