Style fixup
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 3ffe1e17a601eebc54debbb1f51580a969b56bc1..b095527e4d028d18ab30d185a7b25c4a93eba7ab 100644 (file)
--- a/fio.c
+++ b/fio.c
 #include <string.h>
 #include <signal.h>
 #include <time.h>
+#include <locale.h>
 #include <assert.h>
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <sys/ipc.h>
 #include <sys/shm.h>
-#include <sys/ioctl.h>
 #include <sys/mman.h>
 
 #include "fio.h"
 #include "os.h"
 
-#define MASK   (4095)
-
-#define ALIGN(buf)     (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
+static unsigned long page_mask;
+#define ALIGN(buf)     \
+       (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
 
 int groupid = 0;
 int thread_number = 0;
@@ -47,11 +47,20 @@ int shm_id = 0;
 int temp_stall_ts;
 
 static volatile int startup_sem;
+static volatile int fio_abort;
+static int exit_value;
+
+struct io_log *agg_io_log[2];
 
 #define TERMINATE_ALL          (-1)
 #define JOB_START_TIMEOUT      (5 * 1000)
 
-static void terminate_threads(int group_id)
+static inline void td_set_runstate(struct thread_data *td, int runstate)
+{
+       td->runstate = runstate;
+}
+
+static void terminate_threads(int group_id, int forced_kill)
 {
        struct thread_data *td;
        int i;
@@ -60,6 +69,8 @@ static void terminate_threads(int group_id)
                if (group_id == TERMINATE_ALL || groupid == td->groupid) {
                        td->terminate = 1;
                        td->start_delay = 0;
+                       if (forced_kill)
+                               td_set_runstate(td, TD_EXITED);
                }
        }
 }
@@ -73,9 +84,9 @@ static void sig_handler(int sig)
                        print_thread_status();
                        break;
                default:
-                       printf("\nfio: terminating on signal\n");
+                       printf("\nfio: terminating on signal %d\n", sig);
                        fflush(stdout);
-                       terminate_threads(TERMINATE_ALL);
+                       terminate_threads(TERMINATE_ALL, 0);
                        break;
        }
 }
@@ -85,9 +96,9 @@ 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 spent;
        unsigned long rate;
-       int ddir = td->ddir;
 
        /*
         * allow a 2 second settle period in the beginning
@@ -95,6 +106,11 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
        if (mtime_since(&td->start, now) < 2000)
                return 0;
 
+       if (td_read(td))
+               bytes += td->this_io_bytes[DDIR_READ];
+       if (td_write(td))
+               bytes += td->this_io_bytes[DDIR_WRITE];
+
        /*
         * if rate blocks is set, sample is running
         */
@@ -103,14 +119,19 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
                if (spent < td->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);
+               if (bytes < td->rate_bytes) {
+                       fprintf(f_out, "%s: min rate %u not met\n", td->name, td->ratemin);
                        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);
+                               return 1;
+                       }
                }
        }
 
-       td->rate_bytes = td->this_io_bytes[ddir];
+       td->rate_bytes = bytes;
        memcpy(&td->lastrate, now, sizeof(*now));
        return 0;
 }
@@ -125,52 +146,22 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
        return 0;
 }
 
-static inline void td_set_runstate(struct thread_data *td, int runstate)
-{
-       td->runstate = runstate;
-}
-
-static struct fio_file *get_next_file(struct thread_data *td)
-{
-       unsigned int old_next_file = td->next_file;
-       struct fio_file *f;
-
-       do {
-               f = &td->files[td->next_file];
-
-               td->next_file++;
-               if (td->next_file >= td->nr_files)
-                       td->next_file = 0;
-
-               if (f->fd != -1)
-                       break;
-
-               f = NULL;
-       } while (td->next_file != old_next_file);
-
-       return f;
-}
-
 /*
  * When job exits, we can cancel the in-flight IO if we are using async
  * io. Attempt to do so.
  */
 static void cleanup_pending_aio(struct thread_data *td)
 {
-       struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
        struct list_head *entry, *n;
-       struct io_completion_data icd;
        struct io_u *io_u;
        int r;
 
        /*
         * get immediately available events, if any
         */
-       r = td_io_getevents(td, 0, td->cur_depth, &ts);
-       if (r > 0) {
-               icd.nr = r;
-               ios_completed(td, &icd);
-       }
+       r = io_u_queued_complete(td, 0);
+       if (r < 0)
+               return;
 
        /*
         * now cancel remaining active events
@@ -179,19 +170,24 @@ static void cleanup_pending_aio(struct thread_data *td)
                list_for_each_safe(entry, n, &td->io_u_busylist) {
                        io_u = list_entry(entry, struct io_u, list);
 
-                       r = td->io_ops->cancel(td, io_u);
-                       if (!r)
+                       /*
+                        * if the io_u isn't in flight, then that generally
+                        * means someone leaked an io_u. complain but fix
+                        * it up, so we don't stall here.
+                        */
+                       if ((io_u->flags & IO_U_F_FLIGHT) == 0) {
+                               log_err("fio: non-busy IO on busy list\n");
                                put_io_u(td, io_u);
+                       } else {
+                               r = td->io_ops->cancel(td, io_u);
+                               if (!r)
+                                       put_io_u(td, io_u);
+                       }
                }
        }
 
-       if (td->cur_depth) {
-               r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
-               if (r > 0) {
-                       icd.nr = r;
-                       ios_completed(td, &icd);
-               }
-       }
+       if (td->cur_depth)
+               r = io_u_queued_complete(td, td->cur_depth);
 }
 
 /*
@@ -201,7 +197,6 @@ static void cleanup_pending_aio(struct thread_data *td)
 static int fio_io_sync(struct thread_data *td, struct fio_file *f)
 {
        struct io_u *io_u = __get_io_u(td);
-       struct io_completion_data icd;
        int ret;
 
        if (!io_u)
@@ -215,24 +210,27 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
                return 1;
        }
 
+requeue:
        ret = td_io_queue(td, io_u);
-       if (ret) {
-               td_verror(td, io_u->error);
-               put_io_u(td, io_u);
-               return 1;
-       }
-
-       ret = td_io_getevents(td, 1, td->cur_depth, NULL);
        if (ret < 0) {
-               td_verror(td, ret);
+               td_verror(td, io_u->error, "td_io_queue");
+               put_io_u(td, io_u);
                return 1;
-       }
+       } else if (ret == FIO_Q_QUEUED) {
+               if (io_u_queued_complete(td, 1) < 0)
+                       return 1;
+       } else if (ret == FIO_Q_COMPLETED) {
+               if (io_u->error) {
+                       td_verror(td, io_u->error, "td_io_queue");
+                       return 1;
+               }
 
-       icd.nr = ret;
-       ios_completed(td, &icd);
-       if (icd.error) {
-               td_verror(td, icd.error);
-               return 1;
+               if (io_u_sync_complete(td, io_u) < 0)
+                       return 1;
+       } else if (ret == FIO_Q_BUSY) {
+               if (td_io_commit(td))
+                       return 1;
+               goto requeue;
        }
 
        return 0;
@@ -244,26 +242,28 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f)
  */
 static void do_verify(struct thread_data *td)
 {
-       struct io_u *io_u, *v_io_u = NULL;
-       struct io_completion_data icd;
        struct fio_file *f;
-       int ret, i;
+       struct io_u *io_u;
+       int ret, i, min_events;
 
        /*
         * sync io first and invalidate cache, to make sure we really
         * read from disk.
         */
        for_each_file(td, f, i) {
-               fio_io_sync(td, f);
-               file_invalidate_cache(td, f);
+               if (fio_io_sync(td, f))
+                       break;
+               if (file_invalidate_cache(td, f))
+                       break;
        }
 
-       td_set_runstate(td, TD_VERIFYING);
+       if (td->error)
+               return;
 
-       do {
-               if (td->terminate)
-                       break;
+       td_set_runstate(td, TD_VERIFYING);
 
+       io_u = NULL;
+       while (!td->terminate) {
                io_u = __get_io_u(td);
                if (!io_u)
                        break;
@@ -278,62 +278,71 @@ static void do_verify(struct thread_data *td)
                        break;
                }
 
-               f = get_next_file(td);
-               if (!f)
-                       break;
-
-               io_u->file = f;
-
                if (td_io_prep(td, io_u)) {
                        put_io_u(td, io_u);
                        break;
                }
 
+               io_u->end_io = verify_io_u;
+requeue:
                ret = td_io_queue(td, io_u);
-               if (ret) {
-                       td_verror(td, io_u->error);
-                       put_io_u(td, io_u);
-                       break;
-               }
 
-               /*
-                * we have one pending to verify, do that while
-                * we are doing io on the next one
-                */
-               if (do_io_u_verify(td, &v_io_u))
-                       break;
+               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) {
+                               int bytes = io_u->xfer_buflen - io_u->resid;
 
-               ret = td_io_getevents(td, 1, 1, NULL);
-               if (ret != 1) {
+                               io_u->xfer_buflen = io_u->resid;
+                               io_u->xfer_buf += bytes;
+                               goto requeue;
+                       }
+                       ret = io_u_sync_complete(td, io_u);
                        if (ret < 0)
-                               td_verror(td, ret);
+                               break;
+                       continue;
+               case FIO_Q_QUEUED:
+                       break;
+               case FIO_Q_BUSY:
+                       requeue_io_u(td, &io_u);
+                       ret = td_io_commit(td);
+                       break;
+               default:
+                       assert(ret < 0);
+                       td_verror(td, -ret, "td_io_queue");
                        break;
                }
 
-               v_io_u = td->io_ops->event(td, 0);
-               icd.nr = 1;
-               icd.error = 0;
-               fio_gettime(&icd.time, NULL);
-               io_completed(td, v_io_u, &icd);
-
-               if (icd.error) {
-                       td_verror(td, icd.error);
-                       put_io_u(td, v_io_u);
-                       v_io_u = NULL;
+               if (ret < 0 || td->error)
                        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;
                }
 
                /*
-                * if we can't submit more io, we need to verify now
+                * Reap required number of io units, if any, and do the
+                * verification on them through the callback handler
                 */
-               if (queue_full(td) && do_io_u_verify(td, &v_io_u))
+               if (io_u_queued_complete(td, min_events) < 0)
                        break;
+       }
 
-       } while (1);
-
-       do_io_u_verify(td, &v_io_u);
+       if (!td->error) {
+               min_events = td->cur_depth;
 
-       if (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);
@@ -370,97 +379,142 @@ static void do_cpuio(struct thread_data *td)
  */
 static void do_io(struct thread_data *td)
 {
-       struct io_completion_data icd;
        struct timeval s;
        unsigned long usec;
-       struct fio_file *f;
        int i, ret = 0;
 
        td_set_runstate(td, TD_RUNNING);
 
-       while (td->this_io_bytes[td->ddir] < td->io_size) {
-               struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
-               struct timespec *timeout;
+       while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) {
+               struct timeval comp_time;
+               long bytes_done = 0;
                int min_evts = 0;
                struct io_u *io_u;
 
                if (td->terminate)
                        break;
 
-               f = get_next_file(td);
-               if (!f)
-                       break;
-
-               io_u = get_io_u(td, f);
+               io_u = get_io_u(td);
                if (!io_u)
                        break;
 
                memcpy(&s, &io_u->start_time, sizeof(s));
 
-               ret = td_io_queue(td, io_u);
-               if (ret) {
-                       td_verror(td, io_u->error);
+               if (runtime_exceeded(td, &s)) {
                        put_io_u(td, io_u);
                        break;
                }
+requeue:
+               ret = td_io_queue(td, io_u);
 
-               add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
+               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) {
+                               int bytes = io_u->xfer_buflen - io_u->resid;
 
-               if (td->cur_depth < td->iodepth) {
-                       timeout = &ts;
-                       min_evts = 0;
-               } else {
-                       timeout = NULL;
-                       min_evts = 1;
+                               io_u->xfer_buflen = io_u->resid;
+                               io_u->xfer_buf += bytes;
+                               goto requeue;
+                       }
+                       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:
+                       /*
+                        * if the engine doesn't have a commit hook,
+                        * the io_u is really queued. if it does have such
+                        * a hook, it has to call io_u_queued() itself.
+                        */
+                       if (td->io_ops->commit == NULL)
+                               io_u_queued(td, io_u);
+                       break;
+               case FIO_Q_BUSY:
+                       requeue_io_u(td, &io_u);
+                       ret = td_io_commit(td);
+                       break;
+               default:
+                       assert(ret < 0);
+                       put_io_u(td, io_u);
+                       break;
                }
 
-               ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
-               if (ret < 0) {
-                       td_verror(td, ret);
+               if (ret < 0 || td->error)
                        break;
-               } else if (!ret)
-                       continue;
 
-               icd.nr = ret;
-               ios_completed(td, &icd);
-               if (icd.error) {
-                       td_verror(td, icd.error);
-                       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) {
+                               min_evts = 1;
+
+                               if (td->cur_depth > td->iodepth_low)
+                                       min_evts = td->cur_depth - td->iodepth_low;
+                       }
+
+                       fio_gettime(&comp_time, NULL);
+                       bytes_done = io_u_queued_complete(td, min_evts);
+                       if (bytes_done < 0)
+                               break;
                }
 
+               if (!bytes_done)
+                       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, &icd.time);
+               usec = utime_since(&s, &comp_time);
 
-               rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
+               rate_throttle(td, usec, bytes_done);
 
-               if (check_min_rate(td, &icd.time)) {
+               if (check_min_rate(td, &comp_time)) {
                        if (exitall_on_terminate)
-                               terminate_threads(td->groupid);
-                       td_verror(td, ENOMEM);
+                               terminate_threads(td->groupid, 0);
+                       td_verror(td, ENODATA, "check_min_rate");
                        break;
                }
 
-               if (runtime_exceeded(td, &icd.time))
-                       break;
+               if (td->thinktime) {
+                       unsigned long long b;
 
-               if (td->thinktime)
-                       usec_sleep(td, td->thinktime);
+                       b = td->io_blocks[0] + td->io_blocks[1];
+                       if (!(b % td->thinktime_blocks)) {
+                               int left;
+
+                               if (td->thinktime_spin)
+                                       __usec_sleep(td->thinktime_spin);
+
+                               left = td->thinktime - td->thinktime_spin;
+                               if (left)
+                                       usec_sleep(td, left);
+                       }
+               }
        }
 
        if (!td->error) {
-               if (td->cur_depth)
-                       cleanup_pending_aio(td);
+               struct fio_file *f;
+
+               i = td->cur_depth;
+               if (i)
+                       ret = io_u_queued_complete(td, i);
 
                if (should_fsync(td) && td->end_fsync) {
                        td_set_runstate(td, TD_FSYNCING);
                        for_each_file(td, f, i)
                                fio_io_sync(td, f);
                }
-       }
+       } else
+               cleanup_pending_aio(td);
 }
 
 static void cleanup_io_u(struct thread_data *td)
@@ -507,7 +561,12 @@ static int init_io_u(struct thread_data *td)
                max_units = td->iodepth;
 
        max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
-       td->orig_buffer_size = max_bs * max_units + MASK;
+       td->orig_buffer_size = max_bs * max_units;
+
+       if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
+               td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
+       else
+               td->orig_buffer_size += page_mask;
 
        if (allocate_io_mem(td))
                return 1;
@@ -523,9 +582,12 @@ static int init_io_u(struct thread_data *td)
                        fill_rand_buf(io_u, max_bs);
 
                io_u->index = i;
+               io_u->flags = IO_U_F_FREE;
                list_add(&io_u->list, &td->io_u_freelist);
        }
 
+       io_u_init_timeout();
+
        return 0;
 }
 
@@ -542,7 +604,7 @@ static int switch_ioscheduler(struct thread_data *td)
 
        f = fopen(tmp, "r+");
        if (!f) {
-               td_verror(td, errno);
+               td_verror(td, errno, "fopen");
                return 1;
        }
 
@@ -551,7 +613,7 @@ static int switch_ioscheduler(struct thread_data *td)
         */
        ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
        if (ferror(f) || ret != 1) {
-               td_verror(td, errno);
+               td_verror(td, errno, "fwrite");
                fclose(f);
                return 1;
        }
@@ -563,7 +625,7 @@ 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;
        }
@@ -571,7 +633,7 @@ static int switch_ioscheduler(struct thread_data *td)
        sprintf(tmp2, "[%s]", td->ioscheduler);
        if (!strstr(tmp, tmp2)) {
                log_err("fio: io scheduler %s not found\n", td->ioscheduler);
-               td_verror(td, EINVAL);
+               td_verror(td, EINVAL, "iosched_switch");
                fclose(f);
                return 1;
        }
@@ -585,11 +647,15 @@ static void clear_io_state(struct thread_data *td)
        struct fio_file *f;
        int i;
 
-       td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
+       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->last_was_sync = 0;
+
        for_each_file(td, f, i) {
+               f->last_completed_pos = 0;
+
                f->last_pos = 0;
                if (td->io_ops->flags & FIO_SYNCIO)
                        lseek(f->fd, SEEK_SET, 0);
@@ -615,6 +681,7 @@ static void *thread_main(void *data)
 
        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);
 
@@ -622,25 +689,22 @@ static void *thread_main(void *data)
                goto err;
 
        if (fio_setaffinity(td) == -1) {
-               td_verror(td, errno);
+               td_verror(td, errno, "cpu_set_affinity");
                goto err;
        }
 
-       if (td_io_init(td))
-               goto err;
-
        if (init_iolog(td))
                goto err;
 
        if (td->ioprio) {
                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);
+               td_verror(td, errno, "nice");
                goto err;
        }
 
@@ -659,19 +723,29 @@ static void *thread_main(void *data)
        if (open_files(td))
                goto err;
 
-       if (td->exec_prerun)
-               system(td->exec_prerun);
+       /*
+        * Do this late, as some IO engines would like to have the
+        * files setup prior to initializing structures.
+        */
+       if (td_io_init(td))
+               goto err;
+
+       if (td->exec_prerun) {
+               if (system(td->exec_prerun) < 0)
+                       goto err;
+       }
 
        fio_gettime(&td->epoch, NULL);
-       getrusage(RUSAGE_SELF, &td->ru_start);
+       memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
+       getrusage(RUSAGE_SELF, &td->ts.ru_start);
 
        runtime[0] = runtime[1] = 0;
        while (td->loops--) {
                fio_gettime(&td->start, NULL);
-               memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
+               memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
 
                if (td->ratemin)
-                       memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
+                       memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
 
                clear_io_state(td);
                prune_io_piece_log(td);
@@ -681,10 +755,11 @@ static void *thread_main(void *data)
                else
                        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];
-
+               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);
+               
                if (td->error || td->terminate)
                        break;
 
@@ -707,48 +782,53 @@ static void *thread_main(void *data)
        td->runtime[0] = runtime[0] / 1000;
        td->runtime[1] = runtime[1] / 1000;
 
-       if (td->bw_log)
-               finish_log(td, td->bw_log, "bw");
-       if (td->slat_log)
-               finish_log(td, td->slat_log, "slat");
-       if (td->clat_log)
-               finish_log(td, td->clat_log, "clat");
+       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)
-               system(td->exec_postrun);
+       if (td->exec_postrun) {
+               if (system(td->exec_postrun) < 0)
+                       log_err("fio: postrun %s failed\n", td->exec_postrun);
+       }
 
        if (exitall_on_terminate)
-               terminate_threads(td->groupid);
+               terminate_threads(td->groupid, 0);
 
 err:
+       if (td->error)
+               printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
        close_files(td);
        close_ioengine(td);
        cleanup_io_u(td);
        td_set_runstate(td, TD_EXITED);
-       return NULL;
-
+       return (void *) (unsigned long) td->error;
 }
 
 /*
  * We cannot pass the td data into a forked process, so attach the td and
  * pass it to the thread worker.
  */
-static void *fork_main(int shmid, int offset)
+static int fork_main(int shmid, int offset)
 {
        struct thread_data *td;
-       void *data;
+       void *data, *ret;
 
        data = shmat(shmid, NULL, 0);
        if (data == (void *) -1) {
+               int __err = errno;
+
                perror("shmat");
-               return NULL;
+               return __err;
        }
 
        td = data + offset * sizeof(struct thread_data);
-       thread_main(td);
+       ret = thread_main(td);
        shmdt(data);
-       return NULL;
+       return (int) (unsigned long) ret;
 }
 
 /*
@@ -757,13 +837,15 @@ static void *fork_main(int shmid, int offset)
 static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
 {
        struct thread_data *td;
-       int i, cputhreads, pending;
+       int i, cputhreads, pending, status, ret;
 
        /*
         * reap exited threads (TD_EXITED -> TD_REAPED)
         */
        pending = cputhreads = 0;
        for_each_td(td, i) {
+               int flags = 0;
+
                /*
                 * ->io_ops is NULL for a thread that has closed its
                 * io engine
@@ -771,30 +853,70 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
                        cputhreads++;
 
-               if (td->runstate != TD_EXITED) {
-                       if (td->runstate < TD_RUNNING)
-                               pending++;
-
+               if (!td->pid || td->runstate == TD_REAPED)
+                       continue;
+               if (td->use_thread) {
+                       if (td->runstate == TD_EXITED) {
+                               td_set_runstate(td, TD_REAPED);
+                               goto reaped;
+                       }
                        continue;
                }
 
-               td_set_runstate(td, TD_REAPED);
+               flags = WNOHANG;
+               if (td->runstate == TD_EXITED)
+                       flags = 0;
+
+               /*
+                * check if someone quit or got killed in an unusual way
+                */
+               ret = waitpid(td->pid, &status, flags);
+               if (ret < 0) {
+                       if (errno == ECHILD) {
+                               log_err("fio: pid=%d disappeared %d\n", td->pid, td->runstate);
+                               td_set_runstate(td, TD_REAPED);
+                               goto reaped;
+                       }
+                       perror("waitpid");
+               } else if (ret == td->pid) {
+                       if (WIFSIGNALED(status)) {
+                               int sig = WTERMSIG(status);
+
+                               log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
+                               td_set_runstate(td, TD_REAPED);
+                               goto reaped;
+                       }
+                       if (WIFEXITED(status)) {
+                               if (WEXITSTATUS(status) && !td->error)
+                                       td->error = WEXITSTATUS(status);
+
+                               td_set_runstate(td, TD_REAPED);
+                               goto reaped;
+                       }
+               }
 
+               /*
+                * thread is not dead, continue
+                */
+               continue;
+reaped:
                if (td->use_thread) {
                        long ret;
 
                        if (pthread_join(td->thread, (void *) &ret))
-                               perror("thread_join");
-               } else
-                       waitpid(td->pid, NULL, 0);
+                               perror("pthread_join");
+               }
 
                (*nr_running)--;
                (*m_rate) -= td->ratemin;
                (*t_rate) -= td->rate;
+
+               if (td->error)
+                       exit_value++;
        }
 
        if (*nr_running == cputhreads && !pending)
-               terminate_threads(TERMINATE_ALL);
+               terminate_threads(TERMINATE_ALL, 0);
 }
 
 /*
@@ -825,10 +947,10 @@ static void run_threads(void)
        for_each_td(td, i) {
                print_status_init(td->thread_number - 1);
 
-               init_disk_util(td);
-
-               if (!td->create_serialize)
+               if (!td->create_serialize) {
+                       init_disk_util(td);
                        continue;
+               }
 
                /*
                 * do file setup here so it happens sequentially,
@@ -836,11 +958,18 @@ static void run_threads(void)
                 * client data interspersed on disk
                 */
                if (setup_files(td)) {
+                       exit_value++;
+                       if (td->error)
+                               log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror);
                        td_set_runstate(td, TD_REAPED);
                        todo--;
                }
+
+               init_disk_util(td);
        }
 
+       set_genesis_time();
+
        while (todo) {
                struct thread_data *map[MAX_JOBS];
                struct timeval this_start;
@@ -890,8 +1019,9 @@ static void run_threads(void)
                                if (fork())
                                        fio_sem_down(&startup_sem);
                                else {
-                                       fork_main(shm_id, i);
-                                       exit(0);
+                                       int ret = fork_main(shm_id, i);
+
+                                       exit(ret);
                                }
                        }
                }
@@ -902,7 +1032,7 @@ static void run_threads(void)
                 */
                fio_gettime(&this_start, NULL);
                left = this_jobs;
-               while (left) {
+               while (left && !fio_abort) {
                        if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
                                break;
 
@@ -968,6 +1098,15 @@ static void run_threads(void)
 
 int main(int argc, char *argv[])
 {
+       long ps;
+
+       /*
+        * We need locale for number printing, if it isn't set then just
+        * go with the US format.
+        */
+       if (!getenv("LC_NUMERIC"))
+               setlocale(LC_NUMERIC, "en_US");
+
        if (parse_options(argc, argv))
                return 1;
 
@@ -976,10 +1115,32 @@ int main(int argc, char *argv[])
                return 1;
        }
 
+       ps = sysconf(_SC_PAGESIZE);
+       if (ps < 0) {
+               log_err("Failed to get page size\n");
+               return 1;
+       }
+
+       page_mask = ps - 1;
+
+       if (write_bw_log) {
+               setup_log(&agg_io_log[DDIR_READ]);
+               setup_log(&agg_io_log[DDIR_WRITE]);
+       }
+
+       set_genesis_time();
+
        disk_util_timer_arm();
 
        run_threads();
-       show_run_stats();
 
-       return 0;
+       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");
+               }
+       }
+
+       return exit_value;
 }