Start and stop status interval (or file) thread separately
[fio.git] / backend.c
index 6b6da67645870b74a59eb6c619d82ed870aee60b..4437d196c0d5cdf2b5a9ce34c1982d353e7ea480 100644 (file)
--- a/backend.c
+++ b/backend.c
 #include "server.h"
 #include "lib/getrusage.h"
 #include "idletime.h"
+#include "err.h"
+#include "lib/tp.h"
 
 static pthread_t disk_util_thread;
-static struct fio_mutex *disk_thread_mutex;
+static pthread_cond_t du_cond;
+static pthread_mutex_t du_lock;
+
 static struct fio_mutex *startup_mutex;
-static struct fio_mutex *writeout_mutex;
 static struct flist_head *cgroup_list;
 static char *cgroup_mnt;
 static int exit_value;
@@ -86,7 +89,7 @@ static void sig_int(int sig)
                        fio_server_got_signal(sig);
                else {
                        log_info("\nfio: terminating on signal %d\n", sig);
-                       fflush(stdout);
+                       log_info_flush();
                        exit_value = 128;
                }
 
@@ -182,7 +185,11 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
                                                                ratemin);
                                return 1;
                        } else {
-                               rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
+                               if (spent)
+                                       rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
+                               else
+                                       rate = 0;
+
                                if (rate < ratemin ||
                                    bytes < td->rate_bytes[ddir]) {
                                        log_err("%s: min rate %u not met, got"
@@ -200,7 +207,11 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
                                                td->o.name, rate_iops);
                                return 1;
                        } else {
-                               rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
+                               if (spent)
+                                       rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
+                               else
+                                       rate = 0;
+
                                if (rate < rate_iops_min ||
                                    iops < td->rate_blocks[ddir]) {
                                        log_err("%s: min iops rate %u not met,"
@@ -238,8 +249,6 @@ static int check_min_rate(struct thread_data *td, struct timeval *now,
  */
 static void cleanup_pending_aio(struct thread_data *td)
 {
-       struct flist_head *entry, *n;
-       struct io_u *io_u;
        int r;
 
        /*
@@ -253,18 +262,11 @@ static void cleanup_pending_aio(struct thread_data *td)
         * now cancel remaining active events
         */
        if (td->io_ops->cancel) {
-               flist_for_each_safe(entry, n, &td->io_u_busylist) {
-                       io_u = flist_entry(entry, struct io_u, list);
+               struct io_u *io_u;
+               int i;
 
-                       /*
-                        * 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 {
+               io_u_qiter(&td->io_u_all, io_u, i) {
+                       if (io_u->flags & IO_U_F_FLIGHT) {
                                r = td->io_ops->cancel(td, io_u);
                                if (!r)
                                        put_io_u(td, io_u);
@@ -354,7 +356,7 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
                return 0;
        if (!td->o.timeout)
                return 0;
-       if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
+       if (utime_since(&td->epoch, t) >= td->o.timeout)
                return 1;
 
        return 0;
@@ -391,7 +393,7 @@ static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
                         * fill_device option is set.
                         */
                        td_clear_error(td);
-                       td->terminate = 1;
+                       fio_mark_td_terminate(td);
                        return 1;
                } else {
                        /*
@@ -460,7 +462,7 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
                if (runtime_exceeded(td, &td->tv_cache)) {
                        __update_tv_cache(td);
                        if (runtime_exceeded(td, &td->tv_cache)) {
-                               td->terminate = 1;
+                               fio_mark_td_terminate(td);
                                break;
                        }
                }
@@ -487,6 +489,12 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
                                break;
 
                        while ((io_u = get_io_u(td)) != NULL) {
+                               if (IS_ERR(io_u)) {
+                                       io_u = NULL;
+                                       ret = FIO_Q_BUSY;
+                                       goto reap;
+                               }
+
                                /*
                                 * We are only interested in the places where
                                 * we wrote or trimmed IOs. Turn those into
@@ -583,6 +591,7 @@ sync_done:
                 * completed io_u's first. Note that we can get BUSY even
                 * without IO queued, if the system is resource starved.
                 */
+reap:
                full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
                if (full || !td->o.iodepth_batch_complete) {
                        min_events = min(td->o.iodepth_batch_complete,
@@ -624,9 +633,22 @@ sync_done:
        dprint(FD_VERIFY, "exiting loop\n");
 }
 
+static unsigned int exceeds_number_ios(struct thread_data *td)
+{
+       unsigned long long number_ios;
+
+       if (!td->o.number_ios)
+               return 0;
+
+       number_ios = ddir_rw_sum(td->this_io_blocks);
+       number_ios += td->io_u_queued + td->io_u_in_flight;
+
+       return number_ios >= td->o.number_ios;
+}
+
 static int io_bytes_exceeded(struct thread_data *td)
 {
-       unsigned long long bytes;
+       unsigned long long bytes, limit;
 
        if (td_rw(td))
                bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
@@ -637,7 +659,12 @@ static int io_bytes_exceeded(struct thread_data *td)
        else
                bytes = td->this_io_bytes[DDIR_TRIM];
 
-       return bytes >= td->o.size;
+       if (td->o.io_limit)
+               limit = td->o.io_limit;
+       else
+               limit = td->o.size;
+
+       return bytes >= limit || exceeds_number_ios(td);
 }
 
 /*
@@ -651,12 +678,25 @@ static uint64_t do_io(struct thread_data *td)
        uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
        unsigned int i;
        int ret = 0;
+       uint64_t total_bytes, bytes_issued = 0;
 
        if (in_ramp_time(td))
                td_set_runstate(td, TD_RAMP);
        else
                td_set_runstate(td, TD_RUNNING);
 
+       lat_target_init(td);
+
+       /*
+        * If verify_backlog is enabled, we'll run the verify in this
+        * handler as well. For that case, we may need up to twice the
+        * amount of bytes.
+        */
+       total_bytes = td->o.size;
+       if (td->o.verify != VERIFY_NONE &&
+          (td_write(td) && td->o.verify_backlog))
+               total_bytes += td->o.size;
+
        while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
                (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
                td->o.time_based) {
@@ -676,7 +716,7 @@ static uint64_t do_io(struct thread_data *td)
                if (runtime_exceeded(td, &td->tv_cache)) {
                        __update_tv_cache(td);
                        if (runtime_exceeded(td, &td->tv_cache)) {
-                               td->terminate = 1;
+                               fio_mark_td_terminate(td);
                                break;
                        }
                }
@@ -684,9 +724,22 @@ static uint64_t do_io(struct thread_data *td)
                if (flow_threshold_exceeded(td))
                        continue;
 
+               if (bytes_issued >= total_bytes)
+                       break;
+
                io_u = get_io_u(td);
-               if (!io_u)
+               if (IS_ERR_OR_NULL(io_u)) {
+                       int err = PTR_ERR(io_u);
+
+                       io_u = NULL;
+                       if (err == -EBUSY) {
+                               ret = FIO_Q_BUSY;
+                               goto reap;
+                       }
+                       if (td->o.latency_target)
+                               goto reap;
                        break;
+               }
 
                ddir = io_u->ddir;
 
@@ -697,6 +750,13 @@ static uint64_t do_io(struct thread_data *td)
                 */
                if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
                    ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
+
+                       if (!td->o.verify_pattern_bytes) {
+                               io_u->rand_seed = __rand(&td->__verify_state);
+                               if (sizeof(int) != sizeof(long *))
+                                       io_u->rand_seed *= __rand(&td->__verify_state);
+                       }
+
                        if (td->o.verify_async)
                                io_u->end_io = verify_io_u_async;
                        else
@@ -707,20 +767,37 @@ static uint64_t do_io(struct thread_data *td)
                else
                        td_set_runstate(td, TD_RUNNING);
 
+               /*
+                * Always log IO before it's issued, so we know the specific
+                * order of it. The logged unit will track when the IO has
+                * completed.
+                */
+               if (td_write(td) && io_u->ddir == DDIR_WRITE &&
+                   td->o.do_verify &&
+                   td->o.verify != VERIFY_NONE &&
+                   !td->o.experimental_verify)
+                       log_io_piece(td, io_u);
+
                ret = td_io_queue(td, io_u);
                switch (ret) {
                case FIO_Q_COMPLETED:
                        if (io_u->error) {
                                ret = -io_u->error;
+                               unlog_io_piece(td, io_u);
                                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;
 
+                               bytes_issued += bytes;
+
+                               trim_io_piece(td, io_u);
+
                                /*
                                 * zero read, fail
                                 */
                                if (!bytes) {
+                                       unlog_io_piece(td, io_u);
                                        td_verror(td, EIO, "full resid");
                                        put_io_u(td, io_u);
                                        break;
@@ -747,6 +824,7 @@ sync_done:
                                ret = io_u_sync_complete(td, io_u, bytes_done);
                                if (ret < 0)
                                        break;
+                               bytes_issued += io_u->xfer_buflen;
                        }
                        break;
                case FIO_Q_QUEUED:
@@ -757,8 +835,10 @@ sync_done:
                         */
                        if (td->io_ops->commit == NULL)
                                io_u_queued(td, io_u);
+                       bytes_issued += io_u->xfer_buflen;
                        break;
                case FIO_Q_BUSY:
+                       unlog_io_piece(td, io_u);
                        requeue_io_u(td, &io_u);
                        ret2 = td_io_commit(td);
                        if (ret2 < 0)
@@ -778,6 +858,7 @@ sync_done:
                 * can get BUSY even without IO queued, if the system is
                 * resource starved.
                 */
+reap:
                full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
                if (full || !td->o.iodepth_batch_complete) {
                        min_evts = min(td->o.iodepth_batch_complete,
@@ -814,6 +895,8 @@ sync_done:
                                break;
                        }
                }
+               if (!in_ramp_time(td) && td->o.latency_target)
+                       lat_target_check(td);
 
                if (td->o.thinktime) {
                        unsigned long long b;
@@ -841,7 +924,7 @@ sync_done:
 
        if (td->o.fill_device && td->error == ENOSPC) {
                td->error = 0;
-               td->terminate = 1;
+               fio_mark_td_terminate(td);
        }
        if (!td->error) {
                struct fio_file *f;
@@ -878,13 +961,9 @@ sync_done:
 
 static void cleanup_io_u(struct thread_data *td)
 {
-       struct flist_head *entry, *n;
        struct io_u *io_u;
 
-       flist_for_each_safe(entry, n, &td->io_u_freelist) {
-               io_u = flist_entry(entry, struct io_u, list);
-
-               flist_del(&io_u->list);
+       while ((io_u = io_u_qpop(&td->io_u_freelist)) != NULL) {
 
                if (td->io_ops->io_u_free)
                        td->io_ops->io_u_free(td, io_u);
@@ -893,6 +972,10 @@ static void cleanup_io_u(struct thread_data *td)
        }
 
        free_io_mem(td);
+
+       io_u_rexit(&td->io_u_requeues);
+       io_u_qexit(&td->io_u_freelist);
+       io_u_qexit(&td->io_u_all);
 }
 
 static int init_io_u(struct thread_data *td)
@@ -900,7 +983,7 @@ static int init_io_u(struct thread_data *td)
        struct io_u *io_u;
        unsigned int max_bs, min_write;
        int cl_align, i, max_units;
-       int data_xfer = 1;
+       int data_xfer = 1, err;
        char *p;
 
        max_units = td->o.iodepth;
@@ -912,13 +995,24 @@ static int init_io_u(struct thread_data *td)
        if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
                data_xfer = 0;
 
+       err = 0;
+       err += io_u_rinit(&td->io_u_requeues, td->o.iodepth);
+       err += io_u_qinit(&td->io_u_freelist, td->o.iodepth);
+       err += io_u_qinit(&td->io_u_all, td->o.iodepth);
+
+       if (err) {
+               log_err("fio: failed setting up IO queues\n");
+               return 1;
+       }
+
        /*
         * if we may later need to do address alignment, then add any
         * possible adjustment here so that we don't cause a buffer
         * overflow later. this adjustment may be too much if we get
         * lucky and the allocator gives us an aligned address.
         */
-       if (td->o.odirect || td->o.mem_align || (td->io_ops->flags & FIO_RAWIO))
+       if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
+           (td->io_ops->flags & FIO_RAWIO))
                td->orig_buffer_size += page_mask + td->o.mem_align;
 
        if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
@@ -936,7 +1030,7 @@ static int init_io_u(struct thread_data *td)
        if (data_xfer && allocate_io_mem(td))
                return 1;
 
-       if (td->o.odirect || td->o.mem_align ||
+       if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
            (td->io_ops->flags & FIO_RAWIO))
                p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
        else
@@ -958,7 +1052,7 @@ static int init_io_u(struct thread_data *td)
 
                io_u = ptr;
                memset(io_u, 0, sizeof(*io_u));
-               INIT_FLIST_HEAD(&io_u->list);
+               INIT_FLIST_HEAD(&io_u->verify_list);
                dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 
                if (data_xfer) {
@@ -972,13 +1066,19 @@ static int init_io_u(struct thread_data *td)
                                 * Fill the buffer with the pattern if we are
                                 * going to be doing writes.
                                 */
-                               fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
+                               fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0);
                        }
                }
 
                io_u->index = i;
                io_u->flags = IO_U_F_FREE;
-               flist_add(&io_u->list, &td->io_u_freelist);
+               io_u_qpush(&td->io_u_freelist, io_u);
+
+               /*
+                * io_u never leaves this stack, used for iteration of all
+                * io_u buffers.
+                */
+               io_u_qpush(&td->io_u_all, io_u);
 
                if (td->io_ops->io_u_init) {
                        int ret = td->io_ops->io_u_init(td, io_u);
@@ -1032,12 +1132,14 @@ static int switch_ioscheduler(struct thread_data *td)
        /*
         * Read back and check that the selected scheduler is now the default.
         */
-       ret = fread(tmp, 1, sizeof(tmp), f);
+       ret = fread(tmp, sizeof(tmp), 1, f);
        if (ferror(f) || ret < 0) {
                td_verror(td, errno, "fread");
                fclose(f);
                return 1;
        }
+       tmp[sizeof(tmp) - 1] = '\0';
+
 
        sprintf(tmp2, "[%s]", td->o.ioscheduler);
        if (!strstr(tmp, tmp2)) {
@@ -1053,6 +1155,8 @@ static int switch_ioscheduler(struct thread_data *td)
 
 static int keep_running(struct thread_data *td)
 {
+       unsigned long long limit;
+
        if (td->done)
                return 0;
        if (td->o.time_based)
@@ -1061,32 +1165,43 @@ static int keep_running(struct thread_data *td)
                td->o.loops--;
                return 1;
        }
+       if (exceeds_number_ios(td))
+               return 0;
+
+       if (td->o.io_limit)
+               limit = td->o.io_limit;
+       else
+               limit = td->o.size;
 
-       if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
+       if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) {
                uint64_t diff;
 
                /*
                 * If the difference is less than the minimum IO size, we
                 * are done.
                 */
-               diff = td->o.size - ddir_rw_sum(td->io_bytes);
+               diff = limit - ddir_rw_sum(td->io_bytes);
                if (diff < td_max_bs(td))
                        return 0;
 
+               if (fio_files_done(td))
+                       return 0;
+
                return 1;
        }
 
        return 0;
 }
 
-static int exec_string(const char *string)
+static int exec_string(struct thread_options *o, const char *string, const char *mode)
 {
-       int ret, newlen = strlen(string) + 1 + 8;
+       int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
        char *str;
 
        str = malloc(newlen);
-       sprintf(str, "sh -c %s", string);
+       sprintf(str, "%s &> %s.%s.txt", string, o->name, mode);
 
+       log_info("%s : Saving output of %s in %s.%s.txt\n",o->name, mode, o->name, mode);
        ret = system(str);
        if (ret == -1)
                log_err("fio: exec of cmd <%s> failed\n", str);
@@ -1095,6 +1210,50 @@ static int exec_string(const char *string)
        return ret;
 }
 
+/*
+ * Dry run to compute correct state of numberio for verification.
+ */
+static uint64_t do_dry_run(struct thread_data *td)
+{
+       uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
+
+       td_set_runstate(td, TD_RUNNING);
+
+       while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
+               (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
+               struct io_u *io_u;
+               int ret;
+
+               if (td->terminate || td->done)
+                       break;
+
+               io_u = get_io_u(td);
+               if (!io_u)
+                       break;
+
+               io_u->flags |= IO_U_F_FLIGHT;
+               io_u->error = 0;
+               io_u->resid = 0;
+               if (ddir_rw(acct_ddir(io_u)))
+                       td->io_issues[acct_ddir(io_u)]++;
+               if (ddir_rw(io_u->ddir)) {
+                       io_u_mark_depth(td, 1);
+                       td->ts.total_io_u[io_u->ddir]++;
+               }
+
+               if (td_write(td) && io_u->ddir == DDIR_WRITE &&
+                   td->o.do_verify &&
+                   td->o.verify != VERIFY_NONE &&
+                   !td->o.experimental_verify)
+                       log_io_piece(td, io_u);
+
+               ret = io_u_sync_complete(td, io_u, bytes_done);
+               (void) ret;
+       }
+
+       return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
+}
+
 /*
  * Entry point for the thread based jobs. The process based jobs end up
  * here as well, after a little setup.
@@ -1121,9 +1280,6 @@ static void *thread_main(void *data)
        if (is_backend)
                fio_server_send_start(td);
 
-       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);
@@ -1143,13 +1299,6 @@ static void *thread_main(void *data)
        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);
-       td->mutex = NULL;
-
        /*
         * A new gid requires privilege, so we need to do this before setting
         * the uid.
@@ -1175,6 +1324,15 @@ static void *thread_main(void *data)
         * allocations.
         */
        if (o->cpumask_set) {
+               if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
+                       ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
+                       if (!ret) {
+                               log_err("fio: no CPUs set\n");
+                               log_err("fio: Try increasing number of available CPUs\n");
+                               td_verror(td, EINVAL, "cpus_split");
+                               goto err;
+                       }
+               }
                ret = fio_setaffinity(td->pid, o->cpumask);
                if (ret == -1) {
                        td_verror(td, errno, "cpu_set_affinity");
@@ -1185,6 +1343,7 @@ static void *thread_main(void *data)
 #ifdef CONFIG_LIBNUMA
        /* numa node setup */
        if (o->numa_cpumask_set || o->numa_memmask_set) {
+               struct bitmask *mask;
                int ret;
 
                if (numa_available() < 0) {
@@ -1193,7 +1352,9 @@ static void *thread_main(void *data)
                }
 
                if (o->numa_cpumask_set) {
-                       ret = numa_run_on_node_mask(o->numa_cpunodesmask);
+                       mask = numa_parse_nodestring(o->numa_cpunodes);
+                       ret = numa_run_on_node_mask(mask);
+                       numa_free_nodemask(mask);
                        if (ret == -1) {
                                td_verror(td, errno, \
                                        "numa_run_on_node_mask failed\n");
@@ -1203,12 +1364,16 @@ static void *thread_main(void *data)
 
                if (o->numa_memmask_set) {
 
+                       mask = NULL;
+                       if (o->numa_memnodes)
+                               mask = numa_parse_nodestring(o->numa_memnodes);
+
                        switch (o->numa_mem_mode) {
                        case MPOL_INTERLEAVE:
-                               numa_set_interleave_mask(o->numa_memnodesmask);
+                               numa_set_interleave_mask(mask);
                                break;
                        case MPOL_BIND:
-                               numa_set_membind(o->numa_memnodesmask);
+                               numa_set_membind(mask);
                                break;
                        case MPOL_LOCAL:
                                numa_set_localalloc();
@@ -1221,6 +1386,9 @@ static void *thread_main(void *data)
                                break;
                        }
 
+                       if (mask)
+                               numa_free_nodemask(mask);
+
                }
        }
 #endif
@@ -1270,7 +1438,7 @@ static void *thread_main(void *data)
        if (init_random_map(td))
                goto err;
 
-       if (o->exec_prerun && exec_string(o->exec_prerun))
+       if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
                goto err;
 
        if (o->pre_read) {
@@ -1278,6 +1446,9 @@ static void *thread_main(void *data)
                        goto err;
        }
 
+       if (td->flags & TD_F_COMPRESS_LOG)
+               tp_init(&td->tp_data);
+
        fio_verify_init(td);
 
        fio_gettime(&td->epoch, NULL);
@@ -1306,7 +1477,10 @@ static void *thread_main(void *data)
 
                prune_io_piece_log(td);
 
-               verify_bytes = do_io(td);
+               if (td->o.verify_only && (td_write(td) || td_rw(td)))
+                       verify_bytes = do_dry_run(td);
+               else
+                       verify_bytes = do_io(td);
 
                clear_state = 1;
 
@@ -1354,46 +1528,13 @@ static void *thread_main(void *data)
 
        fio_unpin_memory(td);
 
-       fio_mutex_down(writeout_mutex);
-       if (td->bw_log) {
-               if (o->bw_log_file) {
-                       finish_log_named(td, td->bw_log,
-                                               o->bw_log_file, "bw");
-               } else
-                       finish_log(td, td->bw_log, "bw");
-       }
-       if (td->lat_log) {
-               if (o->lat_log_file) {
-                       finish_log_named(td, td->lat_log,
-                                               o->lat_log_file, "lat");
-               } else
-                       finish_log(td, td->lat_log, "lat");
-       }
-       if (td->slat_log) {
-               if (o->lat_log_file) {
-                       finish_log_named(td, td->slat_log,
-                                               o->lat_log_file, "slat");
-               } else
-                       finish_log(td, td->slat_log, "slat");
-       }
-       if (td->clat_log) {
-               if (o->lat_log_file) {
-                       finish_log_named(td, td->clat_log,
-                                               o->lat_log_file, "clat");
-               } else
-                       finish_log(td, td->clat_log, "clat");
-       }
-       if (td->iops_log) {
-               if (o->iops_log_file) {
-                       finish_log_named(td, td->iops_log,
-                                               o->iops_log_file, "iops");
-               } else
-                       finish_log(td, td->iops_log, "iops");
-       }
+       fio_writeout_logs(td);
+
+       if (td->flags & TD_F_COMPRESS_LOG)
+               tp_exit(&td->tp_data);
 
-       fio_mutex_up(writeout_mutex);
        if (o->exec_postrun)
-               exec_string(o->exec_postrun);
+               exec_string(o, o->exec_postrun, (const char *)"postrun");
 
        if (exitall_on_terminate)
                fio_terminate_threads(td->groupid);
@@ -1426,6 +1567,9 @@ err:
        fio_mutex_remove(td->rusage_sem);
        td->rusage_sem = NULL;
 
+       fio_mutex_remove(td->mutex);
+       td->mutex = NULL;
+
        td_set_runstate(td, TD_EXITED);
        return (void *) (uintptr_t) td->error;
 }
@@ -1440,7 +1584,7 @@ static int fork_main(int shmid, int offset)
        struct thread_data *td;
        void *data, *ret;
 
-#ifndef __hpux
+#if !defined(__hpux) && !defined(CONFIG_NO_SHM)
        data = shmat(shmid, NULL, 0);
        if (data == (void *) -1) {
                int __err = errno;
@@ -1461,6 +1605,13 @@ static int fork_main(int shmid, int offset)
        return (int) (uintptr_t) ret;
 }
 
+static void dump_td_info(struct thread_data *td)
+{
+       log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to "
+               "be stuck. Doing forceful exit of this job.\n", td->o.name,
+                       (unsigned long) time_since_now(&td->terminate_time));
+}
+
 /*
  * Run over the job map and reap the threads that have exited, if any.
  */
@@ -1538,6 +1689,17 @@ static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
                        }
                }
 
+               /*
+                * If the job is stuck, do a forceful timeout of it and
+                * move on.
+                */
+               if (td->terminate &&
+                   time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
+                       dump_td_info(td);
+                       td_set_runstate(td, TD_REAPED);
+                       goto reaped;
+               }
+
                /*
                 * thread is not dead, continue
                 */
@@ -1554,6 +1716,7 @@ reaped:
                        exit_value++;
 
                done_secs += mtime_since_now(&td->epoch) / 1000;
+               profile_td_exit(td);
        }
 
        if (*nr_running == cputhreads && !pending && realthreads)
@@ -1572,12 +1735,12 @@ static void do_usleep(unsigned int usecs)
 static void run_threads(void)
 {
        struct thread_data *td;
-       unsigned long spent;
        unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
+       uint64_t spent;
 
        if (fio_gtod_offload && fio_start_gtod_thread())
                return;
-       
+
        fio_idle_prof_init();
 
        set_sig_handlers();
@@ -1602,7 +1765,7 @@ static void run_threads(void)
                                                nr_process > 1 ? "es" : "");
                }
                log_info("\n");
-               fflush(stdout);
+               log_info_flush();
        }
 
        todo = thread_number;
@@ -1671,9 +1834,9 @@ static void run_threads(void)
                        }
 
                        if (td->o.start_delay) {
-                               spent = mtime_since_genesis();
+                               spent = utime_since_genesis();
 
-                               if (td->o.start_delay * 1000 > spent)
+                               if (td->o.start_delay > spent)
                                        continue;
                        }
 
@@ -1809,16 +1972,20 @@ static void run_threads(void)
        update_io_ticks();
 }
 
-void wait_for_disk_thread_exit(void)
+static void wait_for_disk_thread_exit(void)
 {
-       fio_mutex_down(disk_thread_mutex);
+       void *ret;
+
+       disk_util_start_exit();
+       pthread_cond_signal(&du_cond);
+       pthread_join(disk_util_thread, &ret);
 }
 
 static void free_disk_util(void)
 {
-       disk_util_start_exit();
-       wait_for_disk_thread_exit();
        disk_util_prune_entries();
+
+       pthread_cond_destroy(&du_cond);
 }
 
 static void *disk_thread_main(void *data)
@@ -1827,17 +1994,30 @@ static void *disk_thread_main(void *data)
 
        fio_mutex_up(startup_mutex);
 
-       while (threads && !ret) {
-               usleep(DISK_UTIL_MSEC * 1000);
-               if (!threads)
+       while (!ret) {
+               uint64_t sec = DISK_UTIL_MSEC / 1000;
+               uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000;
+               struct timespec ts;
+               struct timeval tv;
+
+               gettimeofday(&tv, NULL);
+               ts.tv_sec = tv.tv_sec + sec;
+               ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
+               if (ts.tv_nsec > 1000000000ULL) {
+                       ts.tv_nsec -= 1000000000ULL;
+                       ts.tv_sec++;
+               }
+
+               ret = pthread_cond_timedwait(&du_cond, &du_lock, &ts);
+               if (ret != ETIMEDOUT)
                        break;
+
                ret = update_io_ticks();
 
                if (!is_backend)
                        print_thread_status();
        }
 
-       fio_mutex_up(disk_thread_mutex);
        return NULL;
 }
 
@@ -1847,22 +2027,15 @@ static int create_disk_util_thread(void)
 
        setup_disk_util();
 
-       disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
+       pthread_cond_init(&du_cond, NULL);
+       pthread_mutex_init(&du_lock, NULL);
 
        ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL);
        if (ret) {
-               fio_mutex_remove(disk_thread_mutex);
                log_err("Can't create disk util thread: %s\n", strerror(ret));
                return 1;
        }
 
-       ret = pthread_detach(disk_util_thread);
-       if (ret) {
-               fio_mutex_remove(disk_thread_mutex);
-               log_err("Can't detatch disk util 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");
@@ -1884,35 +2057,43 @@ int fio_backend(void)
                return 0;
 
        if (write_bw_log) {
-               setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW);
-               setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW);
-               setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW);
+               struct log_params p = {
+                       .log_type = IO_LOG_TYPE_BW,
+               };
+
+               setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log");
+               setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log");
+               setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
        }
 
        startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
        if (startup_mutex == NULL)
                return 1;
-       writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
-       if (writeout_mutex == NULL)
-               return 1;
 
        set_genesis_time();
        stat_init();
        create_disk_util_thread();
+       create_status_interval_thread();
 
        cgroup_list = smalloc(sizeof(*cgroup_list));
        INIT_FLIST_HEAD(cgroup_list);
 
        run_threads();
 
+       wait_for_disk_thread_exit();
+       wait_for_status_interval_thread_exit();
+
        if (!fio_abort) {
-               show_run_stats();
+               __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_TRIM],
-                                       "agg-write_bw.log");
+                       int i;
+
+                       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
+                               struct io_log *log = agg_io_log[i];
+
+                               flush_log(log);
+                               free_log(log);
+                       }
                }
        }
 
@@ -1925,8 +2106,6 @@ int fio_backend(void)
        sfree(cgroup_mnt);
 
        fio_mutex_remove(startup_mutex);
-       fio_mutex_remove(writeout_mutex);
-       fio_mutex_remove(disk_thread_mutex);
        stat_exit();
        return exit_value;
 }