Fix usr/sys/ctx/majf/minf for -USR1 usage
[fio.git] / backend.c
index 52791040ff21384d283b2a09de3ff9718e19d866..fcb74dcbcfc3d39b7916c3c1f2f8fcb43127e889 100644 (file)
--- a/backend.c
+++ b/backend.c
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <sys/ipc.h>
-#include <sys/shm.h>
 #include <sys/mman.h>
 
 #include "fio.h"
+#ifndef FIO_NO_HAVE_SHM_H
+#include <sys/shm.h>
+#endif
 #include "hash.h"
 #include "smalloc.h"
 #include "verify.h"
 #include "lib/rand.h"
 #include "memalign.h"
 #include "server.h"
+#include "lib/getrusage.h"
+#include "idletime.h"
 
 static pthread_t disk_util_thread;
+static struct fio_mutex *disk_thread_mutex;
 static struct fio_mutex *startup_mutex;
 static struct fio_mutex *writeout_mutex;
 static struct flist_head *cgroup_list;
@@ -57,15 +62,17 @@ static char *cgroup_mnt;
 static int exit_value;
 static volatile int fio_abort;
 
-struct io_log *agg_io_log[2];
+struct io_log *agg_io_log[DDIR_RWDIR_CNT];
 
 int groupid = 0;
 unsigned int thread_number = 0;
+unsigned int stat_number = 0;
 unsigned int nr_process = 0;
 unsigned int nr_thread = 0;
 int shm_id = 0;
 int temp_stall_ts;
 unsigned long done_secs = 0;
+volatile int disk_util_exit = 0;
 
 #define PAGE_ALIGN(buf)        \
        (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
@@ -87,6 +94,11 @@ static void sig_int(int sig)
        }
 }
 
+static void sig_show_status(int sig)
+{
+       show_running_run_stats();
+}
+
 static void set_sig_handlers(void)
 {
        struct sigaction act;
@@ -101,6 +113,19 @@ static void set_sig_handlers(void)
        act.sa_flags = SA_RESTART;
        sigaction(SIGTERM, &act, NULL);
 
+/* Windows uses SIGBREAK as a quit signal from other applications */
+#ifdef WIN32
+       memset(&act, 0, sizeof(act));
+       act.sa_handler = sig_int;
+       act.sa_flags = SA_RESTART;
+       sigaction(SIGBREAK, &act, NULL);
+#endif
+
+       memset(&act, 0, sizeof(act));
+       act.sa_handler = sig_show_status;
+       act.sa_flags = SA_RESTART;
+       sigaction(SIGUSR1, &act, NULL);
+
        if (is_backend) {
                memset(&act, 0, sizeof(act));
                act.sa_handler = sig_int;
@@ -193,14 +218,16 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
 }
 
 static int check_min_rate(struct thread_data *td, struct timeval *now,
-                         unsigned long *bytes_done)
+                         uint64_t *bytes_done)
 {
        int ret = 0;
 
-       if (bytes_done[0])
-               ret |= __check_min_rate(td, now, 0);
-       if (bytes_done[1])
-               ret |= __check_min_rate(td, now, 1);
+       if (bytes_done[DDIR_READ])
+               ret |= __check_min_rate(td, now, DDIR_READ);
+       if (bytes_done[DDIR_WRITE])
+               ret |= __check_min_rate(td, now, DDIR_WRITE);
+       if (bytes_done[DDIR_TRIM])
+               ret |= __check_min_rate(td, now, DDIR_TRIM);
 
        return ret;
 }
@@ -295,6 +322,21 @@ requeue:
        return 0;
 }
 
+static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
+{
+       int ret;
+
+       if (fio_file_open(f))
+               return fio_io_sync(td, f);
+
+       if (td_io_open_file(td, f))
+               return 1;
+
+       ret = fio_io_sync(td, f);
+       td_io_close_file(td, f);
+       return ret;
+}
+
 static inline void __update_tv_cache(struct thread_data *td)
 {
        fio_gettime(&td->tv_cache, NULL);
@@ -324,17 +366,17 @@ static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
        int ret = *retptr;
 
        if (ret < 0 || td->error) {
-               int err;
+               int err = td->error;
+               enum error_type_bit eb;
 
                if (ret < 0)
                        err = -ret;
-               else
-                       err = td->error;
 
-               if (!(td->o.continue_on_error & td_error_type(ddir, err)))
+               eb = td_error_type(ddir, err);
+               if (!(td->o.continue_on_error & (1 << eb)))
                        return 1;
 
-               if (td_non_fatal_error(err)) {
+               if (td_non_fatal_error(td, eb, err)) {
                        /*
                         * Continue with the I/Os in case of
                         * a non fatal error.
@@ -364,12 +406,22 @@ static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
        return 0;
 }
 
+static void check_update_rusage(struct thread_data *td)
+{
+       if (td->update_rusage) {
+               td->update_rusage = 0;
+               update_rusage_stat(td);
+               fio_mutex_up(td->rusage_sem);
+       }
+}
+
 /*
  * 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)
+static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 {
+       uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
        struct fio_file *f;
        struct io_u *io_u;
        int ret, min_events;
@@ -390,6 +442,8 @@ static void do_verify(struct thread_data *td)
                        break;
        }
 
+       check_update_rusage(td);
+
        if (td->error)
                return;
 
@@ -397,9 +451,11 @@ static void do_verify(struct thread_data *td)
 
        io_u = NULL;
        while (!td->terminate) {
+               enum fio_ddir ddir;
                int ret2, full;
 
                update_tv_cache(td);
+               check_update_rusage(td);
 
                if (runtime_exceeded(td, &td->tv_cache)) {
                        __update_tv_cache(td);
@@ -412,18 +468,53 @@ static void do_verify(struct thread_data *td)
                if (flow_threshold_exceeded(td))
                        continue;
 
-               io_u = __get_io_u(td);
-               if (!io_u)
-                       break;
+               if (!td->o.experimental_verify) {
+                       io_u = __get_io_u(td);
+                       if (!io_u)
+                               break;
 
-               if (get_next_verify(td, io_u)) {
-                       put_io_u(td, io_u);
-                       break;
-               }
+                       if (get_next_verify(td, io_u)) {
+                               put_io_u(td, io_u);
+                               break;
+                       }
 
-               if (td_io_prep(td, io_u)) {
-                       put_io_u(td, io_u);
-                       break;
+                       if (td_io_prep(td, io_u)) {
+                               put_io_u(td, io_u);
+                               break;
+                       }
+               } else {
+                       if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes)
+                               break;
+
+                       while ((io_u = get_io_u(td)) != NULL) {
+                               /*
+                                * We are only interested in the places where
+                                * we wrote or trimmed IOs. Turn those into
+                                * reads for verification purposes.
+                                */
+                               if (io_u->ddir == DDIR_READ) {
+                                       /*
+                                        * Pretend we issued it for rwmix
+                                        * accounting
+                                        */
+                                       td->io_issues[DDIR_READ]++;
+                                       put_io_u(td, io_u);
+                                       continue;
+                               } else if (io_u->ddir == DDIR_TRIM) {
+                                       io_u->ddir = DDIR_READ;
+                                       io_u->flags |= IO_U_F_TRIMMED;
+                                       break;
+                               } else if (io_u->ddir == DDIR_WRITE) {
+                                       io_u->ddir = DDIR_READ;
+                                       break;
+                               } else {
+                                       put_io_u(td, io_u);
+                                       continue;
+                               }
+                       }
+
+                       if (!io_u)
+                               break;
                }
 
                if (td->o.verify_async)
@@ -431,6 +522,8 @@ static void do_verify(struct thread_data *td)
                else
                        io_u->end_io = verify_io_u;
 
+               ddir = io_u->ddir;
+
                ret = td_io_queue(td, io_u);
                switch (ret) {
                case FIO_Q_COMPLETED:
@@ -463,7 +556,7 @@ static void do_verify(struct thread_data *td)
                                requeue_io_u(td, &io_u);
                        } else {
 sync_done:
-                               ret = io_u_sync_complete(td, io_u, NULL);
+                               ret = io_u_sync_complete(td, io_u, bytes_done);
                                if (ret < 0)
                                        break;
                        }
@@ -482,7 +575,7 @@ sync_done:
                        break;
                }
 
-               if (break_on_this_error(td, io_u->ddir, &ret))
+               if (break_on_this_error(td, ddir, &ret))
                        break;
 
                /*
@@ -494,7 +587,10 @@ sync_done:
                if (full || !td->o.iodepth_batch_complete) {
                        min_events = min(td->o.iodepth_batch_complete,
                                         td->cur_depth);
-                       if (full && !min_events && td->o.iodepth_batch_complete != 0)
+                       /*
+                        * if the queue is full, we MUST reap at least 1 event
+                        */
+                       if (full && !min_events)
                                min_events = 1;
 
                        do {
@@ -503,7 +599,7 @@ sync_done:
                                 * and do the verification on them through
                                 * the callback handler
                                 */
-                               if (io_u_queued_complete(td, min_events, NULL) < 0) {
+                               if (io_u_queued_complete(td, min_events, bytes_done) < 0) {
                                        ret = -1;
                                        break;
                                }
@@ -513,6 +609,8 @@ sync_done:
                        break;
        }
 
+       check_update_rusage(td);
+
        if (!td->error) {
                min_events = td->cur_depth;
 
@@ -526,12 +624,31 @@ sync_done:
        dprint(FD_VERIFY, "exiting loop\n");
 }
 
+static int io_bytes_exceeded(struct thread_data *td)
+{
+       unsigned long long bytes;
+
+       if (td_rw(td))
+               bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE];
+       else if (td_write(td))
+               bytes = td->this_io_bytes[DDIR_WRITE];
+       else if (td_read(td))
+               bytes = td->this_io_bytes[DDIR_READ];
+       else
+               bytes = td->this_io_bytes[DDIR_TRIM];
+
+       return bytes >= td->o.size;
+}
+
 /*
  * Main IO worker function. It retrieves io_u's to process and queues
  * and reaps them, checking for rate and errors along the way.
+ *
+ * Returns number of bytes written and trimmed.
  */
-static void do_io(struct thread_data *td)
+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;
 
@@ -540,17 +657,18 @@ static void do_io(struct thread_data *td)
        else
                td_set_runstate(td, TD_RUNNING);
 
-       while ( (td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
-               (!flist_empty(&td->trim_list)) ||
-               ((td->this_io_bytes[0] + td->this_io_bytes[1]) < 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) {
                struct timeval comp_time;
-               unsigned long bytes_done[2] = { 0, 0 };
                int min_evts = 0;
                struct io_u *io_u;
                int ret2, full;
                enum fio_ddir ddir;
 
-               if (td->terminate)
+               check_update_rusage(td);
+
+               if (td->terminate || td->done)
                        break;
 
                update_tv_cache(td);
@@ -573,11 +691,12 @@ static void do_io(struct thread_data *td)
                ddir = io_u->ddir;
 
                /*
-                * Add verification end_io handler, if asked to verify
-                * a previously written file.
+                * Add verification end_io handler if:
+                *      - Asked to verify (!td_rw(td))
+                *      - Or the io_u is from our verify list (mixed write/ver)
                 */
                if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ &&
-                   !td_rw(td)) {
+                   ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) {
                        if (td->o.verify_async)
                                io_u->end_io = verify_io_u_async;
                        else
@@ -620,8 +739,9 @@ static void do_io(struct thread_data *td)
                                requeue_io_u(td, &io_u);
                        } else {
 sync_done:
-                               if (__should_check_rate(td, 0) ||
-                                   __should_check_rate(td, 1))
+                               if (__should_check_rate(td, DDIR_READ) ||
+                                   __should_check_rate(td, DDIR_WRITE) ||
+                                   __should_check_rate(td, DDIR_TRIM))
                                        fio_gettime(&comp_time, NULL);
 
                                ret = io_u_sync_complete(td, io_u, bytes_done);
@@ -662,11 +782,15 @@ sync_done:
                if (full || !td->o.iodepth_batch_complete) {
                        min_evts = min(td->o.iodepth_batch_complete,
                                        td->cur_depth);
-                       if (full && !min_evts && td->o.iodepth_batch_complete != 0)
+                       /*
+                        * if the queue is full, we MUST reap at least 1 event
+                        */
+                       if (full && !min_evts)
                                min_evts = 1;
 
-                       if (__should_check_rate(td, 0) ||
-                           __should_check_rate(td, 1))
+                       if (__should_check_rate(td, DDIR_READ) ||
+                           __should_check_rate(td, DDIR_WRITE) ||
+                           __should_check_rate(td, DDIR_TRIM))
                                fio_gettime(&comp_time, NULL);
 
                        do {
@@ -679,7 +803,7 @@ sync_done:
 
                if (ret < 0)
                        break;
-               if (!(bytes_done[0] + bytes_done[1]))
+               if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
                        continue;
 
                if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
@@ -694,7 +818,7 @@ sync_done:
                if (td->o.thinktime) {
                        unsigned long long b;
 
-                       b = td->io_blocks[0] + td->io_blocks[1];
+                       b = ddir_rw_sum(td->io_blocks);
                        if (!(b % td->o.thinktime_blocks)) {
                                int left;
 
@@ -708,6 +832,8 @@ sync_done:
                }
        }
 
+       check_update_rusage(td);
+
        if (td->trim_entries)
                log_err("fio: %d trim entries leaked?\n", td->trim_entries);
 
@@ -720,7 +846,7 @@ sync_done:
 
                i = td->cur_depth;
                if (i) {
-                       ret = io_u_queued_complete(td, i, NULL);
+                       ret = io_u_queued_complete(td, i, bytes_done);
                        if (td->o.fill_device && td->error == ENOSPC)
                                td->error = 0;
                }
@@ -729,9 +855,11 @@ sync_done:
                        td_set_runstate(td, TD_FSYNCING);
 
                        for_each_file(td, f, i) {
-                               if (!fio_file_open(f))
+                               if (!fio_file_fsync(td, f))
                                        continue;
-                               fio_io_sync(td, f);
+
+                               log_err("fio: end_fsync failed for file %s\n",
+                                                               f->file_name);
                        }
                }
        } else
@@ -740,8 +868,10 @@ sync_done:
        /*
         * stop job if we failed doing any IO
         */
-       if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0)
+       if (!ddir_rw_sum(td->this_io_bytes))
                td->done = 1;
+
+       return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
 }
 
 static void cleanup_io_u(struct thread_data *td)
@@ -753,6 +883,10 @@ static void cleanup_io_u(struct thread_data *td)
                io_u = flist_entry(entry, struct io_u, list);
 
                flist_del(&io_u->list);
+
+               if (td->io_ops->io_u_free)
+                       td->io_ops->io_u_free(td, io_u);
+
                fio_memfree(io_u, sizeof(*io_u));
        }
 
@@ -762,15 +896,20 @@ static void cleanup_io_u(struct thread_data *td)
 static int init_io_u(struct thread_data *td)
 {
        struct io_u *io_u;
-       unsigned int max_bs;
+       unsigned int max_bs, min_write;
        int cl_align, i, max_units;
+       int data_xfer = 1;
        char *p;
 
        max_units = td->o.iodepth;
-       max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
+       max_bs = td_max_bs(td);
+       min_write = td->o.min_bs[DDIR_WRITE];
        td->orig_buffer_size = (unsigned long long) max_bs
                                        * (unsigned long long) max_units;
 
+       if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
+               data_xfer = 0;
+
        if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
                unsigned long bs;
 
@@ -783,7 +922,7 @@ static int init_io_u(struct thread_data *td)
                return 1;
        }
 
-       if (allocate_io_mem(td))
+       if (data_xfer && allocate_io_mem(td))
                return 1;
 
        if (td->o.odirect || td->o.mem_align ||
@@ -811,12 +950,12 @@ static int init_io_u(struct thread_data *td)
                INIT_FLIST_HEAD(&io_u->list);
                dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 
-               if (!(td->io_ops->flags & FIO_NOIO)) {
+               if (data_xfer) {
                        io_u->buf = p;
                        dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 
                        if (td_write(td))
-                               io_u_fill_buffer(td, io_u, max_bs);
+                               io_u_fill_buffer(td, io_u, min_write, max_bs);
                        if (td_write(td) && td->o.verify_pattern_bytes) {
                                /*
                                 * Fill the buffer with the pattern if we are
@@ -829,6 +968,16 @@ static int init_io_u(struct thread_data *td)
                io_u->index = i;
                io_u->flags = IO_U_F_FREE;
                flist_add(&io_u->list, &td->io_u_freelist);
+
+               if (td->io_ops->io_u_init) {
+                       int ret = td->io_ops->io_u_init(td, io_u);
+
+                       if (ret) {
+                               log_err("fio: failed to init engine data: %d\n", ret);
+                               return 1;
+                       }
+               }
+
                p += max_bs;
        }
 
@@ -893,8 +1042,6 @@ static int switch_ioscheduler(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)
@@ -904,10 +1051,19 @@ static int keep_running(struct thread_data *td)
                return 1;
        }
 
-       io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]
-                       + td->io_skip_bytes;
-       if (io_done < td->o.size)
+       if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
+               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);
+               if (diff < td_max_bs(td))
+                       return 0;
+
                return 1;
+       }
 
        return 0;
 }
@@ -945,6 +1101,8 @@ static void *thread_main(void *data)
        } else
                td->pid = gettid();
 
+       fio_local_clock_init(td->o.use_thread);
+
        dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 
        INIT_FLIST_HEAD(&td->io_u_freelist);
@@ -954,6 +1112,7 @@ static void *thread_main(void *data)
        INIT_FLIST_HEAD(&td->io_hist_list);
        INIT_FLIST_HEAD(&td->verify_list);
        INIT_FLIST_HEAD(&td->trim_list);
+       INIT_FLIST_HEAD(&td->next_rand_list);
        pthread_mutex_init(&td->io_u_lock, NULL);
        td->io_hist_tree = RB_ROOT;
 
@@ -1003,6 +1162,49 @@ static void *thread_main(void *data)
                goto err;
        }
 
+#ifdef CONFIG_LIBNUMA
+       /* numa node setup */
+       if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
+               int ret;
+
+               if (numa_available() < 0) {
+                       td_verror(td, errno, "Does not support NUMA API\n");
+                       goto err;
+               }
+
+               if (td->o.numa_cpumask_set) {
+                       ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
+                       if (ret == -1) {
+                               td_verror(td, errno, \
+                                       "numa_run_on_node_mask failed\n");
+                               goto err;
+                       }
+               }
+
+               if (td->o.numa_memmask_set) {
+
+                       switch (td->o.numa_mem_mode) {
+                       case MPOL_INTERLEAVE:
+                               numa_set_interleave_mask(td->o.numa_memnodesmask);
+                               break;
+                       case MPOL_BIND:
+                               numa_set_membind(td->o.numa_memnodesmask);
+                               break;
+                       case MPOL_LOCAL:
+                               numa_set_localalloc();
+                               break;
+                       case MPOL_PREFERRED:
+                               numa_set_preferred(td->o.numa_mem_prefer_node);
+                               break;
+                       case MPOL_DEFAULT:
+                       default:
+                               break;
+                       }
+
+               }
+       }
+#endif
+
        /*
         * May alter parameters that init_io_u() will use, so we need to
         * do this first.
@@ -1023,7 +1225,7 @@ static void *thread_main(void *data)
                }
        }
 
-       if (td->o.cgroup_weight && cgroup_setup(td, cgroup_list, &cgroup_mnt))
+       if (td->o.cgroup && cgroup_setup(td, cgroup_list, &cgroup_mnt))
                goto err;
 
        errno = 0;
@@ -1055,19 +1257,23 @@ static void *thread_main(void *data)
        }
 
        fio_gettime(&td->epoch, NULL);
-       getrusage(RUSAGE_SELF, &td->ru_start);
-
+       fio_getrusage(&td->ru_start);
        clear_state = 0;
        while (keep_running(td)) {
+               uint64_t verify_bytes;
+
                fio_gettime(&td->start, NULL);
                memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
                memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
                memcpy(&td->tv_cache, &td->start, sizeof(td->start));
 
-               if (td->o.ratemin[0] || td->o.ratemin[1]) {
-                       memcpy(&td->lastrate[0], &td->bw_sample_time,
+               if (td->o.ratemin[DDIR_READ] || td->o.ratemin[DDIR_WRITE] ||
+                               td->o.ratemin[DDIR_TRIM]) {
+                       memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
+                                               sizeof(td->bw_sample_time));
+                       memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
                                                sizeof(td->bw_sample_time));
-                       memcpy(&td->lastrate[1], &td->bw_sample_time,
+                       memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
                                                sizeof(td->bw_sample_time));
                }
 
@@ -1076,7 +1282,7 @@ static void *thread_main(void *data)
 
                prune_io_piece_log(td);
 
-               do_io(td);
+               verify_bytes = do_io(td);
 
                clear_state = 1;
 
@@ -1088,6 +1294,10 @@ static void *thread_main(void *data)
                        elapsed = utime_since_now(&td->start);
                        td->ts.runtime[DDIR_WRITE] += elapsed;
                }
+               if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
+                       elapsed = utime_since_now(&td->start);
+                       td->ts.runtime[DDIR_TRIM] += elapsed;
+               }
 
                if (td->error || td->terminate)
                        break;
@@ -1101,7 +1311,7 @@ static void *thread_main(void *data)
 
                fio_gettime(&td->start, NULL);
 
-               do_verify(td);
+               do_verify(td, verify_bytes);
 
                td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
 
@@ -1110,11 +1320,13 @@ static void *thread_main(void *data)
        }
 
        update_rusage_stat(td);
-       td->ts.runtime[0] = (td->ts.runtime[0] + 999) / 1000;
-       td->ts.runtime[1] = (td->ts.runtime[1] + 999) / 1000;
+       td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000;
+       td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000;
+       td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 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];
+       td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ];
+       td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE];
+       td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM];
 
        fio_mutex_down(writeout_mutex);
        if (td->bw_log) {
@@ -1169,8 +1381,8 @@ err:
                verify_async_exit(td);
 
        close_and_free_files(td);
-       close_ioengine(td);
        cleanup_io_u(td);
+       close_ioengine(td);
        cgroup_shutdown(td, &cgroup_mnt);
 
        if (td->o.cpumask_set) {
@@ -1185,6 +1397,9 @@ err:
        if (td->o.write_iolog_file)
                write_iolog_close(td);
 
+       fio_mutex_remove(td->rusage_sem);
+       td->rusage_sem = NULL;
+
        td_set_runstate(td, TD_EXITED);
        return (void *) (uintptr_t) td->error;
 }
@@ -1272,6 +1487,7 @@ static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
                        if (errno == ECHILD) {
                                log_err("fio: pid=%d disappeared %d\n",
                                                (int) td->pid, td->runstate);
+                               td->sig = ECHILD;
                                td_set_runstate(td, TD_REAPED);
                                goto reaped;
                        }
@@ -1280,9 +1496,10 @@ static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
                        if (WIFSIGNALED(status)) {
                                int sig = WTERMSIG(status);
 
-                               if (sig != SIGTERM)
+                               if (sig != SIGTERM && sig != SIGUSR2)
                                        log_err("fio: pid=%d, got signal=%d\n",
                                                        (int) td->pid, sig);
+                               td->sig = sig;
                                td_set_runstate(td, TD_REAPED);
                                goto reaped;
                        }
@@ -1302,8 +1519,8 @@ static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
                continue;
 reaped:
                (*nr_running)--;
-               (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
-               (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
+               (*m_rate) -= ddir_rw_sum(td->o.ratemin);
+               (*t_rate) -= ddir_rw_sum(td->o.rate);
                if (!td->pid)
                        pending--;
 
@@ -1331,10 +1548,12 @@ static void run_threads(void)
 
        if (fio_gtod_offload && fio_start_gtod_thread())
                return;
+       
+       fio_idle_prof_init();
 
        set_sig_handlers();
 
-       if (!terse_output) {
+       if (output_format == FIO_OUTPUT_NORMAL) {
                log_info("Starting ");
                if (nr_thread)
                        log_info("%d thread%s", nr_thread,
@@ -1388,6 +1607,9 @@ static void run_threads(void)
                }
        }
 
+       /* start idle threads before io threads start to run */
+       fio_idle_prof_start();
+
        set_genesis_time();
 
        while (todo) {
@@ -1426,6 +1648,9 @@ static void run_threads(void)
 
                        init_disk_util(td);
 
+                       td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
+                       td->update_rusage = 0;
+
                        /*
                         * Set state to created. Thread will transition
                         * to TD_INITIALIZED when it's done setting up.
@@ -1525,8 +1750,8 @@ static void run_threads(void)
                                td_set_runstate(td, TD_RUNNING);
                        nr_running++;
                        nr_started--;
-                       m_rate += td->o.ratemin[0] + td->o.ratemin[1];
-                       t_rate += td->o.rate[0] + td->o.rate[1];
+                       m_rate += ddir_rw_sum(td->o.ratemin);
+                       t_rate += ddir_rw_sum(td->o.rate);
                        todo--;
                        fio_mutex_up(td->mutex);
                }
@@ -1550,24 +1775,41 @@ static void run_threads(void)
                        usleep(10000);
        }
 
+       fio_idle_prof_stop();
+
        update_io_ticks();
        fio_unpin_memory();
 }
 
+void wait_for_disk_thread_exit(void)
+{
+       fio_mutex_down(disk_thread_mutex);
+}
+
+static void free_disk_util(void)
+{
+       disk_util_start_exit();
+       wait_for_disk_thread_exit();
+       disk_util_prune_entries();
+}
+
 static void *disk_thread_main(void *data)
 {
+       int ret = 0;
+
        fio_mutex_up(startup_mutex);
 
-       while (threads) {
+       while (threads && !ret) {
                usleep(DISK_UTIL_MSEC * 1000);
                if (!threads)
                        break;
-               update_io_ticks();
+               ret = update_io_ticks();
 
                if (!is_backend)
                        print_thread_status();
        }
 
+       fio_mutex_up(disk_thread_mutex);
        return NULL;
 }
 
@@ -1575,14 +1817,20 @@ static int create_disk_util_thread(void)
 {
        int ret;
 
+       setup_disk_util();
+
+       disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
+
        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;
        }
@@ -1610,12 +1858,13 @@ int fio_backend(void)
        if (write_bw_log) {
                setup_log(&agg_io_log[DDIR_READ], 0);
                setup_log(&agg_io_log[DDIR_WRITE], 0);
+               setup_log(&agg_io_log[DDIR_TRIM], 0);
        }
 
-       startup_mutex = fio_mutex_init(0);
+       startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
        if (startup_mutex == NULL)
                return 1;
-       writeout_mutex = fio_mutex_init(1);
+       writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
        if (writeout_mutex == NULL)
                return 1;
 
@@ -1633,17 +1882,21 @@ int fio_backend(void)
                        __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");
                }
        }
 
        for_each_td(td, i)
                fio_options_free(td);
 
+       free_disk_util();
        cgroup_kill(cgroup_list);
        sfree(cgroup_list);
        sfree(cgroup_mnt);
 
        fio_mutex_remove(startup_mutex);
        fio_mutex_remove(writeout_mutex);
+       fio_mutex_remove(disk_thread_mutex);
        return exit_value;
 }