iolog: remove 'td' from trim_io_piece()
[fio.git] / backend.c
index 60cea3c682a8dc8c5c95a40f2a438ff06a054fbd..151210a439ba303a18c995d11fa08baac1c3bec7 100644 (file)
--- a/backend.c
+++ b/backend.c
  *
  *  You should have received a copy of the GNU General Public License
  *  along with this program; if not, write to the Free Software
- *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
+ *  Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
  *
  */
 #include <unistd.h>
-#include <fcntl.h>
 #include <string.h>
-#include <limits.h>
 #include <signal.h>
-#include <time.h>
-#include <locale.h>
 #include <assert.h>
-#include <time.h>
 #include <inttypes.h>
 #include <sys/stat.h>
 #include <sys/wait.h>
-#include <sys/ipc.h>
-#include <sys/mman.h>
 #include <math.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 "trim.h"
 #include "diskutil.h"
 #include "cgroup.h"
 #include "profile.h"
@@ -58,8 +46,9 @@
 #include "lib/mountcheck.h"
 #include "rate-submit.h"
 #include "helper_thread.h"
+#include "pshared.h"
 
-static struct fio_mutex *startup_mutex;
+static struct fio_sem *startup_sem;
 static struct flist_head *cgroup_list;
 static char *cgroup_mnt;
 static int exit_value;
@@ -76,9 +65,6 @@ int shm_id = 0;
 int temp_stall_ts;
 unsigned long done_secs = 0;
 
-#define PAGE_ALIGN(buf)        \
-       (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
-
 #define JOB_START_TIMEOUT      (5 * 1000)
 
 static void sig_int(int sig)
@@ -139,7 +125,7 @@ static void set_sig_handlers(void)
 /*
  * Check if we are above the minimum rate given.
  */
-static bool __check_min_rate(struct thread_data *td, struct timeval *now,
+static bool __check_min_rate(struct thread_data *td, struct timespec *now,
                             enum fio_ddir ddir)
 {
        unsigned long long bytes = 0;
@@ -180,8 +166,8 @@ static bool __check_min_rate(struct thread_data *td, struct timeval *now,
                         * check bandwidth specified rate
                         */
                        if (bytes < td->rate_bytes[ddir]) {
-                               log_err("%s: min rate %u not met\n", td->o.name,
-                                                               ratemin);
+                               log_err("%s: rate_min=%uB/s not met, only transferred %lluB\n",
+                                       td->o.name, ratemin, bytes);
                                return true;
                        } else {
                                if (spent)
@@ -191,9 +177,8 @@ static bool __check_min_rate(struct thread_data *td, struct timeval *now,
 
                                if (rate < ratemin ||
                                    bytes < td->rate_bytes[ddir]) {
-                                       log_err("%s: min rate %u not met, got"
-                                               " %luKB/sec\n", td->o.name,
-                                                       ratemin, rate);
+                                       log_err("%s: rate_min=%uB/s not met, got %luB/s\n",
+                                               td->o.name, ratemin, rate);
                                        return true;
                                }
                        }
@@ -202,8 +187,8 @@ static bool __check_min_rate(struct thread_data *td, struct timeval *now,
                         * checks iops specified rate
                         */
                        if (iops < rate_iops) {
-                               log_err("%s: min iops rate %u not met\n",
-                                               td->o.name, rate_iops);
+                               log_err("%s: rate_iops_min=%u not met, only performed %lu IOs\n",
+                                               td->o.name, rate_iops, iops);
                                return true;
                        } else {
                                if (spent)
@@ -213,9 +198,8 @@ static bool __check_min_rate(struct thread_data *td, struct timeval *now,
 
                                if (rate < rate_iops_min ||
                                    iops < td->rate_blocks[ddir]) {
-                                       log_err("%s: min iops rate %u not met,"
-                                               " got %lu\n", td->o.name,
-                                                       rate_iops_min, rate);
+                                       log_err("%s: rate_iops_min=%u not met, got %lu IOPS\n",
+                                               td->o.name, rate_iops_min, rate);
                                        return true;
                                }
                        }
@@ -228,7 +212,7 @@ static bool __check_min_rate(struct thread_data *td, struct timeval *now,
        return false;
 }
 
-static bool check_min_rate(struct thread_data *td, struct timeval *now)
+static bool check_min_rate(struct thread_data *td, struct timespec *now)
 {
        bool ret = false;
 
@@ -284,7 +268,7 @@ static void cleanup_pending_aio(struct thread_data *td)
 static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 {
        struct io_u *io_u = __get_io_u(td);
-       int ret;
+       enum fio_q_status ret;
 
        if (!io_u)
                return true;
@@ -299,16 +283,13 @@ static bool fio_io_sync(struct thread_data *td, struct fio_file *f)
 
 requeue:
        ret = td_io_queue(td, io_u);
-       if (ret < 0) {
-               td_verror(td, io_u->error, "td_io_queue");
-               put_io_u(td, io_u);
-               return true;
-       } else if (ret == FIO_Q_QUEUED) {
-               if (td_io_commit(td))
-                       return true;
+       switch (ret) {
+       case FIO_Q_QUEUED:
+               td_io_commit(td);
                if (io_u_queued_complete(td, 1) < 0)
                        return true;
-       } else if (ret == FIO_Q_COMPLETED) {
+               break;
+       case FIO_Q_COMPLETED:
                if (io_u->error) {
                        td_verror(td, io_u->error, "td_io_queue");
                        return true;
@@ -316,9 +297,9 @@ requeue:
 
                if (io_u_sync_complete(td, io_u) < 0)
                        return true;
-       } else if (ret == FIO_Q_BUSY) {
-               if (td_io_commit(td))
-                       return true;
+               break;
+       case FIO_Q_BUSY:
+               td_io_commit(td);
                goto requeue;
        }
 
@@ -340,18 +321,18 @@ static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
        return ret;
 }
 
-static inline void __update_tv_cache(struct thread_data *td)
+static inline void __update_ts_cache(struct thread_data *td)
 {
-       fio_gettime(&td->tv_cache, NULL);
+       fio_gettime(&td->ts_cache, NULL);
 }
 
-static inline void update_tv_cache(struct thread_data *td)
+static inline void update_ts_cache(struct thread_data *td)
 {
-       if ((++td->tv_cache_nr & td->tv_cache_mask) == td->tv_cache_mask)
-               __update_tv_cache(td);
+       if ((++td->ts_cache_nr & td->ts_cache_mask) == td->ts_cache_mask)
+               __update_ts_cache(td);
 }
 
-static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
+static inline bool runtime_exceeded(struct thread_data *td, struct timespec *t)
 {
        if (in_ramp_time(td))
                return false;
@@ -431,21 +412,18 @@ 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);
+               fio_sem_up(td->rusage_sem);
        }
 }
 
-static int wait_for_completions(struct thread_data *td, struct timeval *time)
+static int wait_for_completions(struct thread_data *td, struct timespec *time)
 {
        const int full = queue_full(td);
        int min_evts = 0;
        int ret;
 
-       if (td->flags & TD_F_REGROW_LOGS) {
-               ret = io_u_quiesce(td);
-               regrow_logs(td);
-               return ret;
-       }
+       if (td->flags & TD_F_REGROW_LOGS)
+               return io_u_quiesce(td);
 
        /*
         * if the queue is full, we MUST reap at least 1 event
@@ -454,9 +432,7 @@ static int wait_for_completions(struct thread_data *td, struct timeval *time)
        if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
                min_evts = 1;
 
-       if (time && (__should_check_rate(td, DDIR_READ) ||
-           __should_check_rate(td, DDIR_WRITE) ||
-           __should_check_rate(td, DDIR_TRIM)))
+       if (time && __should_check_rate(td))
                fio_gettime(time, NULL);
 
        do {
@@ -470,10 +446,8 @@ static int wait_for_completions(struct thread_data *td, struct timeval *time)
 
 int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
                   enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
-                  struct timeval *comp_time)
+                  struct timespec *comp_time)
 {
-       int ret2;
-
        switch (*ret) {
        case FIO_Q_COMPLETED:
                if (io_u->error) {
@@ -487,7 +461,7 @@ int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
                                *bytes_issued += bytes;
 
                        if (!from_verify)
-                               trim_io_piece(td, io_u);
+                               trim_io_piece(io_u);
 
                        /*
                         * zero read, fail
@@ -507,16 +481,13 @@ int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
                        if (ddir_rw(io_u->ddir))
                                td->ts.short_io_u[io_u->ddir]++;
 
-                       f = io_u->file;
                        if (io_u->offset == f->real_file_size)
                                goto sync_done;
 
                        requeue_io_u(td, &io_u);
                } else {
 sync_done:
-                       if (comp_time && (__should_check_rate(td, DDIR_READ) ||
-                           __should_check_rate(td, DDIR_WRITE) ||
-                           __should_check_rate(td, DDIR_TRIM)))
+                       if (comp_time && __should_check_rate(td))
                                fio_gettime(comp_time, NULL);
 
                        *ret = io_u_sync_complete(td, io_u);
@@ -550,9 +521,7 @@ sync_done:
                if (!from_verify)
                        unlog_io_piece(td, io_u);
                requeue_io_u(td, &io_u);
-               ret2 = td_io_commit(td);
-               if (ret2 < 0)
-                       *ret = ret2;
+               td_io_commit(td);
                break;
        default:
                assert(*ret < 0);
@@ -594,6 +563,50 @@ static int unlink_all_files(struct thread_data *td)
        return ret;
 }
 
+/*
+ * Check if io_u will overlap an in-flight IO in the queue
+ */
+static bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u)
+{
+       bool overlap;
+       struct io_u *check_io_u;
+       unsigned long long x1, x2, y1, y2;
+       int i;
+
+       x1 = io_u->offset;
+       x2 = io_u->offset + io_u->buflen;
+       overlap = false;
+       io_u_qiter(q, check_io_u, i) {
+               if (check_io_u->flags & IO_U_F_FLIGHT) {
+                       y1 = check_io_u->offset;
+                       y2 = check_io_u->offset + check_io_u->buflen;
+
+                       if (x1 < y2 && y1 < x2) {
+                               overlap = true;
+                               dprint(FD_IO, "in-flight overlap: %llu/%lu, %llu/%lu\n",
+                                               x1, io_u->buflen,
+                                               y1, check_io_u->buflen);
+                               break;
+                       }
+               }
+       }
+
+       return overlap;
+}
+
+static enum fio_q_status io_u_submit(struct thread_data *td, struct io_u *io_u)
+{
+       /*
+        * Check for overlap if the user asked us to, and we have
+        * at least one IO in flight besides this one.
+        */
+       if (td->o.serialize_overlap && td->cur_depth > 1 &&
+           in_flight_overlap(&td->io_u_all, io_u))
+               return FIO_Q_BUSY;
+
+       return td_io_queue(td, io_u);
+}
+
 /*
  * The main verify engine. Runs over the writes we previously submitted,
  * reads the blocks back in, and checks the crc/md5 of the data.
@@ -641,12 +654,12 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
                enum fio_ddir ddir;
                int full;
 
-               update_tv_cache(td);
+               update_ts_cache(td);
                check_update_rusage(td);
 
-               if (runtime_exceeded(td, &td->tv_cache)) {
-                       __update_tv_cache(td);
-                       if (runtime_exceeded(td, &td->tv_cache)) {
+               if (runtime_exceeded(td, &td->ts_cache)) {
+                       __update_ts_cache(td);
+                       if (runtime_exceeded(td, &td->ts_cache)) {
                                fio_mark_td_terminate(td);
                                break;
                        }
@@ -699,6 +712,7 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
                                        break;
                                } else if (io_u->ddir == DDIR_WRITE) {
                                        io_u->ddir = DDIR_READ;
+                                       populate_verify_io_u(td, io_u);
                                        break;
                                } else {
                                        put_io_u(td, io_u);
@@ -724,7 +738,7 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
                if (!td->o.disable_slat)
                        fio_gettime(&io_u->start_time, NULL);
 
-               ret = td_io_queue(td, io_u);
+               ret = io_u_submit(td, io_u);
 
                if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL))
                        break;
@@ -784,8 +798,8 @@ static bool io_bytes_exceeded(struct thread_data *td, uint64_t *this_bytes)
        else
                bytes = this_bytes[DDIR_TRIM];
 
-       if (td->o.io_limit)
-               limit = td->o.io_limit;
+       if (td->o.io_size)
+               limit = td->o.io_size;
        else
                limit = td->o.size;
 
@@ -809,32 +823,76 @@ static bool io_complete_bytes_exceeded(struct thread_data *td)
  */
 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
 {
-       uint64_t secs, remainder, bps, bytes, iops;
+       uint64_t bps = td->rate_bps[ddir];
 
        assert(!(td->flags & TD_F_CHILD));
-       bytes = td->rate_io_issue_bytes[ddir];
-       bps = td->rate_bps[ddir];
 
        if (td->o.rate_process == RATE_PROCESS_POISSON) {
-               uint64_t val;
+               uint64_t val, iops;
+
                iops = bps / td->o.bs[ddir];
                val = (int64_t) (1000000 / iops) *
-                               -logf(__rand_0_1(&td->poisson_state));
+                               -logf(__rand_0_1(&td->poisson_state[ddir]));
                if (val) {
-                       dprint(FD_RATE, "poisson rate iops=%llu\n",
-                                       (unsigned long long) 1000000 / val);
+                       dprint(FD_RATE, "poisson rate iops=%llu, ddir=%d\n",
+                                       (unsigned long long) 1000000 / val,
+                                       ddir);
                }
-               td->last_usec += val;
-               return td->last_usec;
+               td->last_usec[ddir] += val;
+               return td->last_usec[ddir];
        } else if (bps) {
-               secs = bytes / bps;
-               remainder = bytes % bps;
+               uint64_t bytes = td->rate_io_issue_bytes[ddir];
+               uint64_t secs = bytes / bps;
+               uint64_t remainder = bytes % bps;
+
                return remainder * 1000000 / bps + secs * 1000000;
        }
 
        return 0;
 }
 
+static void handle_thinktime(struct thread_data *td, enum fio_ddir ddir)
+{
+       unsigned long long b;
+       uint64_t total;
+       int left;
+
+       b = ddir_rw_sum(td->io_blocks);
+       if (b % td->o.thinktime_blocks)
+               return;
+
+       io_u_quiesce(td);
+
+       total = 0;
+       if (td->o.thinktime_spin)
+               total = usec_spin(td->o.thinktime_spin);
+
+       left = td->o.thinktime - total;
+       if (left)
+               total += usec_sleep(td, left);
+
+       /*
+        * If we're ignoring thinktime for the rate, add the number of bytes
+        * we would have done while sleeping, minus one block to ensure we
+        * start issuing immediately after the sleep.
+        */
+       if (total && td->rate_bps[ddir] && td->o.rate_ign_think) {
+               uint64_t missed = (td->rate_bps[ddir] * total) / 1000000ULL;
+               uint64_t bs = td->o.min_bs[ddir];
+               uint64_t usperop = bs * 1000000ULL / td->rate_bps[ddir];
+               uint64_t over;
+
+               if (usperop <= total)
+                       over = bs;
+               else
+                       over = (usperop - total) / usperop * -bs;
+
+               td->rate_io_issue_bytes[ddir] += (missed - over);
+               /* adjust for rate_process=poisson */
+               td->last_usec[ddir] += total;
+       }
+}
+
 /*
  * Main IO worker function. It retrieves io_u's to process and queues
  * and reaps them, checking for rate and errors along the way.
@@ -859,11 +917,11 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
 
        total_bytes = td->o.size;
        /*
-       * Allow random overwrite workloads to write up to io_limit
+       * Allow random overwrite workloads to write up to io_size
        * before starting verification phase as 'size' doesn't apply.
        */
        if (td_write(td) && td_random(td) && td->o.norandommap)
-               total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
+               total_bytes = max(total_bytes, (uint64_t) td->o.io_size);
        /*
         * 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
@@ -881,7 +939,7 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
        while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
                (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
                td->o.time_based) {
-               struct timeval comp_time;
+               struct timespec comp_time;
                struct io_u *io_u;
                int full;
                enum fio_ddir ddir;
@@ -891,11 +949,11 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
                if (td->terminate || td->done)
                        break;
 
-               update_tv_cache(td);
+               update_ts_cache(td);
 
-               if (runtime_exceeded(td, &td->tv_cache)) {
-                       __update_tv_cache(td);
-                       if (runtime_exceeded(td, &td->tv_cache)) {
+               if (runtime_exceeded(td, &td->ts_cache)) {
+                       __update_ts_cache(td);
+                       if (runtime_exceeded(td, &td->ts_cache)) {
                                fio_mark_td_terminate(td);
                                break;
                        }
@@ -919,6 +977,7 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
                        int err = PTR_ERR(io_u);
 
                        io_u = NULL;
+                       ddir = DDIR_INVAL;
                        if (err == -EBUSY) {
                                ret = FIO_Q_BUSY;
                                goto reap;
@@ -928,6 +987,9 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
                        break;
                }
 
+               if (io_u->ddir == DDIR_WRITE && td->flags & TD_F_DO_VERIFY)
+                       populate_verify_io_u(td, io_u);
+
                ddir = io_u->ddir;
 
                /*
@@ -990,7 +1052,7 @@ static void do_io(struct thread_data *td, uint64_t *bytes_done)
                                td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
 
                } else {
-                       ret = td_io_queue(td, io_u);
+                       ret = io_u_submit(td, io_u);
 
                        if (should_check_rate(td))
                                td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
@@ -1026,23 +1088,8 @@ reap:
                if (!in_ramp_time(td) && td->o.latency_target)
                        lat_target_check(td);
 
-               if (td->o.thinktime) {
-                       unsigned long long b;
-
-                       b = ddir_rw_sum(td->io_blocks);
-                       if (!(b % td->o.thinktime_blocks)) {
-                               int left;
-
-                               io_u_quiesce(td);
-
-                               if (td->o.thinktime_spin)
-                                       usec_spin(td->o.thinktime_spin);
-
-                               left = td->o.thinktime - td->o.thinktime_spin;
-                               if (left)
-                                       usec_sleep(td, left);
-                       }
-               }
+               if (ddir_rw(ddir) && td->o.thinktime)
+                       handle_thinktime(td, ddir);
        }
 
        check_update_rusage(td);
@@ -1167,9 +1214,9 @@ static int init_io_u(struct thread_data *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);
+       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");
@@ -1203,7 +1250,7 @@ static int init_io_u(struct thread_data *td)
 
        if (td->o.odirect || td->o.mem_align || td->o.oatomic ||
            td_ioengine_flagged(td, FIO_RAWIO))
-               p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
+               p = PTR_ALIGN(td->orig_buffer, page_mask) + td->o.mem_align;
        else
                p = td->orig_buffer;
 
@@ -1269,17 +1316,22 @@ static int init_io_u(struct thread_data *td)
        return 0;
 }
 
+/*
+ * This function is Linux specific.
+ * FIO_HAVE_IOSCHED_SWITCH enabled currently means it's Linux.
+ */
 static int switch_ioscheduler(struct thread_data *td)
 {
 #ifdef FIO_HAVE_IOSCHED_SWITCH
-       char tmp[256], tmp2[128];
+       char tmp[256], tmp2[128], *p;
        FILE *f;
        int ret;
 
        if (td_ioengine_flagged(td, FIO_DISKLESSIO))
                return 0;
 
-       sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
+       assert(td->files && td->files[0]);
+       sprintf(tmp, "%s/queue/scheduler", td->files[0]->du->sysfs_root);
 
        f = fopen(tmp, "r+");
        if (!f) {
@@ -1307,17 +1359,19 @@ static int switch_ioscheduler(struct thread_data *td)
        /*
         * Read back and check that the selected scheduler is now the default.
         */
-       memset(tmp, 0, sizeof(tmp));
-       ret = fread(tmp, sizeof(tmp), 1, f);
+       ret = fread(tmp, 1, sizeof(tmp) - 1, f);
        if (ferror(f) || ret < 0) {
                td_verror(td, errno, "fread");
                fclose(f);
                return 1;
        }
+       tmp[ret] = '\0';
        /*
-        * either a list of io schedulers or "none\n" is expected.
+        * either a list of io schedulers or "none\n" is expected. Strip the
+        * trailing newline.
         */
-       tmp[strlen(tmp) - 1] = '\0';
+       p = tmp;
+       strsep(&p, "\n");
 
        /*
         * Write to "none" entry doesn't fail, so check the result here.
@@ -1349,6 +1403,8 @@ static bool keep_running(struct thread_data *td)
 
        if (td->done)
                return false;
+       if (td->terminate)
+               return false;
        if (td->o.time_based)
                return true;
        if (td->o.loops) {
@@ -1358,8 +1414,8 @@ static bool keep_running(struct thread_data *td)
        if (exceeds_number_ios(td))
                return false;
 
-       if (td->o.io_limit)
-               limit = td->o.io_limit;
+       if (td->o.io_size)
+               limit = td->o.io_size;
        else
                limit = td->o.size;
 
@@ -1367,14 +1423,14 @@ static bool keep_running(struct thread_data *td)
                uint64_t diff;
 
                /*
-                * If the difference is less than the minimum IO size, we
+                * If the difference is less than the maximum IO size, we
                 * are done.
                 */
                diff = limit - ddir_rw_sum(td->io_bytes);
                if (diff < td_max_bs(td))
                        return false;
 
-               if (fio_files_done(td) && !td->o.io_limit)
+               if (fio_files_done(td) && !td->o.io_size)
                        return false;
 
                return true;
@@ -1459,8 +1515,9 @@ static void *thread_main(void *data)
        struct thread_data *td = fd->td;
        struct thread_options *o = &td->o;
        struct sk_out *sk_out = fd->sk_out;
+       uint64_t bytes_done[DDIR_RWDIR_CNT];
        int deadlock_loop_cnt;
-       int clear_state;
+       bool clear_state, did_some_io;
        int ret;
 
        sk_out_assign(sk_out);
@@ -1472,7 +1529,7 @@ static void *thread_main(void *data)
        } else
                td->pid = gettid();
 
-       fio_local_clock_init(o->use_thread);
+       fio_local_clock_init();
 
        dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 
@@ -1483,7 +1540,6 @@ 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);
        td->io_hist_tree = RB_ROOT;
 
        ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond);
@@ -1498,11 +1554,11 @@ static void *thread_main(void *data)
        }
 
        td_set_runstate(td, TD_INITIALIZED);
-       dprint(FD_MUTEX, "up startup_mutex\n");
-       fio_mutex_up(startup_mutex);
-       dprint(FD_MUTEX, "wait on td->mutex\n");
-       fio_mutex_down(td->mutex);
-       dprint(FD_MUTEX, "done waiting on td->mutex\n");
+       dprint(FD_MUTEX, "up startup_sem\n");
+       fio_sem_up(startup_sem);
+       dprint(FD_MUTEX, "wait on td->sem\n");
+       fio_sem_down(td->sem);
+       dprint(FD_MUTEX, "done waiting on td->sem\n");
 
        /*
         * A new gid requires privilege, so we need to do this before setting
@@ -1612,7 +1668,7 @@ static void *thread_main(void *data)
         * May alter parameters that init_io_u() will use, so we need to
         * do this first.
         */
-       if (init_iolog(td))
+       if (!init_iolog(td))
                goto err;
 
        if (init_io_u(td))
@@ -1648,16 +1704,14 @@ static void *thread_main(void *data)
        if (td_io_init(td))
                goto err;
 
-       if (init_random_map(td))
+       if (!init_random_map(td))
                goto err;
 
        if (o->exec_prerun && exec_string(o, o->exec_prerun, (const char *)"prerun"))
                goto err;
 
-       if (o->pre_read) {
-               if (pre_read_files(td) < 0)
-                       goto err;
-       }
+       if (o->pre_read && !pre_read_files(td))
+               goto err;
 
        fio_verify_init(td);
 
@@ -1668,6 +1722,7 @@ static void *thread_main(void *data)
        fio_getrusage(&td->ru_start);
        memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
        memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
+       memcpy(&td->ss.prev_time, &td->epoch, sizeof(td->epoch));
 
        if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
                        o->ratemin[DDIR_TRIM]) {
@@ -1679,12 +1734,15 @@ static void *thread_main(void *data)
                                        sizeof(td->bw_sample_time));
        }
 
-       clear_state = 0;
+       memset(bytes_done, 0, sizeof(bytes_done));
+       clear_state = false;
+       did_some_io = false;
+
        while (keep_running(td)) {
                uint64_t verify_bytes;
 
                fio_gettime(&td->start, NULL);
-               memcpy(&td->tv_cache, &td->start, sizeof(td->start));
+               memcpy(&td->ts_cache, &td->start, sizeof(td->start));
 
                if (clear_state) {
                        clear_io_state(td, 0);
@@ -1695,11 +1753,9 @@ static void *thread_main(void *data)
 
                prune_io_piece_log(td);
 
-               if (td->o.verify_only && (td_write(td) || td_rw(td)))
+               if (td->o.verify_only && td_write(td))
                        verify_bytes = do_dry_run(td);
                else {
-                       uint64_t bytes_done[DDIR_RWDIR_CNT];
-
                        do_io(td, bytes_done);
 
                        if (!ddir_rw_sum(bytes_done)) {
@@ -1719,7 +1775,7 @@ static void *thread_main(void *data)
                if (td->runstate >= TD_EXITED)
                        break;
 
-               clear_state = 1;
+               clear_state = true;
 
                /*
                 * Make sure we've successfully updated the rusage stats
@@ -1731,11 +1787,11 @@ static void *thread_main(void *data)
                deadlock_loop_cnt = 0;
                do {
                        check_update_rusage(td);
-                       if (!fio_mutex_down_trylock(stat_mutex))
+                       if (!fio_sem_down_trylock(stat_sem))
                                break;
                        usleep(1000);
                        if (deadlock_loop_cnt++ > 5000) {
-                               log_err("fio seems to be stuck grabbing stat_mutex, forcibly exiting\n");
+                               log_err("fio seems to be stuck grabbing stat_sem, forcibly exiting\n");
                                td->error = EDEADLK;
                                goto err;
                        }
@@ -1748,7 +1804,7 @@ static void *thread_main(void *data)
                if (td_trim(td) && td->io_bytes[DDIR_TRIM])
                        update_runtime(td, elapsed_us, DDIR_TRIM);
                fio_gettime(&td->start, NULL);
-               fio_mutex_up(stat_mutex);
+               fio_sem_up(stat_sem);
 
                if (td->error || td->terminate)
                        break;
@@ -1758,6 +1814,9 @@ static void *thread_main(void *data)
                    td_ioengine_flagged(td, FIO_UNIDIR))
                        continue;
 
+               if (ddir_rw_sum(bytes_done))
+                       did_some_io = true;
+
                clear_io_state(td, 0);
 
                fio_gettime(&td->start, NULL);
@@ -1769,15 +1828,28 @@ static void *thread_main(void *data)
                 */
                check_update_rusage(td);
 
-               fio_mutex_down(stat_mutex);
+               fio_sem_down(stat_sem);
                update_runtime(td, elapsed_us, DDIR_READ);
                fio_gettime(&td->start, NULL);
-               fio_mutex_up(stat_mutex);
+               fio_sem_up(stat_sem);
 
                if (td->error || td->terminate)
                        break;
        }
 
+       /*
+        * If td ended up with no I/O when it should have had,
+        * then something went wrong unless FIO_NOIO or FIO_DISKLESSIO.
+        * (Are we not missing other flags that can be ignored ?)
+        */
+       if ((td->o.size || td->o.io_size) && !ddir_rw_sum(bytes_done) &&
+           !did_some_io && !td->o.create_only &&
+           !(td_ioengine_flagged(td, FIO_NOIO) ||
+             td_ioengine_flagged(td, FIO_DISKLESSIO)))
+               log_err("%s: No I/O performed by %s, "
+                        "perhaps try --debug=io option for details?\n",
+                        td->o.name, td->io_ops->name);
+
        td_set_runstate(td, TD_FINISHING);
 
        update_rusage_stat(td);
@@ -1838,9 +1910,6 @@ err:
        if (o->write_iolog_file)
                write_iolog_close(td);
 
-       fio_mutex_remove(td->mutex);
-       td->mutex = NULL;
-
        td_set_runstate(td, TD_EXITED);
 
        /*
@@ -1853,19 +1922,11 @@ err:
        return (void *) (uintptr_t) td->error;
 }
 
-static void dump_td_info(struct thread_data *td)
-{
-       log_err("fio: job '%s' (state=%d) hasn't exited in %lu seconds, it "
-               "appears to be stuck. Doing forceful exit of this job.\n",
-                       td->o.name, td->runstate,
-                       (unsigned long) time_since_now(&td->terminate_time));
-}
-
 /*
  * Run over the job map and reap the threads that have exited, if any.
  */
-static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
-                        unsigned int *m_rate)
+static void reap_threads(unsigned int *nr_running, uint64_t *t_rate,
+                        uint64_t *m_rate)
 {
        struct thread_data *td;
        unsigned int cputhreads, realthreads, pending;
@@ -1878,11 +1939,7 @@ static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
        for_each_td(td, i) {
                int flags = 0;
 
-               /*
-                * ->io_ops is NULL for a thread that has closed its
-                * io engine
-                */
-               if (td->io_ops && !strcmp(td->io_ops->name, "cpuio"))
+                if (!strcmp(td->o.ioengine, "cpuio"))
                        cputhreads++;
                else
                        realthreads++;
@@ -1945,7 +2002,11 @@ static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
                if (td->terminate &&
                    td->runstate < TD_FSYNCING &&
                    time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) {
-                       dump_td_info(td);
+                       log_err("fio: job '%s' (state=%d) hasn't exited in "
+                               "%lu seconds, it appears to be stuck. Doing "
+                               "forceful exit of this job.\n",
+                               td->o.name, td->runstate,
+                               (unsigned long) time_since_now(&td->terminate_time));
                        td_set_runstate(td, TD_REAPED);
                        goto reaped;
                }
@@ -1993,7 +2054,10 @@ static bool __check_trigger_file(void)
 static bool trigger_timedout(void)
 {
        if (trigger_timeout)
-               return time_since_genesis() >= trigger_timeout;
+               if (time_since_genesis() >= trigger_timeout) {
+                       trigger_timeout = 0;
+                       return true;
+               }
 
        return false;
 }
@@ -2002,7 +2066,7 @@ void exec_trigger(const char *cmd)
 {
        int ret;
 
-       if (!cmd)
+       if (!cmd || cmd[0] == '\0')
                return;
 
        ret = system(cmd);
@@ -2058,8 +2122,16 @@ static bool check_mount_writes(struct thread_data *td)
        if (!td_write(td) || td->o.allow_mounted_write)
                return false;
 
+       /*
+        * If FIO_HAVE_CHARDEV_SIZE is defined, it's likely that chrdevs
+        * are mkfs'd and mounted.
+        */
        for_each_file(td, f, i) {
-               if (f->filetype != FIO_TYPE_BD)
+#ifdef FIO_HAVE_CHARDEV_SIZE
+               if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR)
+#else
+               if (f->filetype != FIO_TYPE_BLOCK)
+#endif
                        continue;
                if (device_is_mounted(f->file_name))
                        goto mounted;
@@ -2067,7 +2139,7 @@ static bool check_mount_writes(struct thread_data *td)
 
        return false;
 mounted:
-       log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
+       log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.\n", f->file_name);
        return true;
 }
 
@@ -2103,7 +2175,8 @@ static bool waitee_running(struct thread_data *me)
 static void run_threads(struct sk_out *sk_out)
 {
        struct thread_data *td;
-       unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
+       unsigned int i, todo, nr_running, nr_started;
+       uint64_t m_rate, t_rate;
        uint64_t spent;
 
        if (fio_gtod_offload && fio_start_gtod_thread())
@@ -2188,7 +2261,7 @@ reap:
 
        while (todo) {
                struct thread_data *map[REAL_MAX_JOBS];
-               struct timeval this_start;
+               struct timespec this_start;
                int this_jobs = 0, left;
                struct fork_data *fd;
 
@@ -2229,7 +2302,7 @@ reap:
 
                        init_disk_util(td);
 
-                       td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED);
+                       td->rusage_sem = fio_sem_init(FIO_SEM_LOCKED);
                        td->update_rusage = 0;
 
                        /*
@@ -2257,6 +2330,7 @@ reap:
                                        nr_started--;
                                        break;
                                }
+                               fd = NULL;
                                ret = pthread_detach(td->thread);
                                if (ret)
                                        log_err("pthread_detach: %s",
@@ -2273,15 +2347,16 @@ reap:
                                } else if (i == fio_debug_jobno)
                                        *fio_debug_jobp = pid;
                        }
-                       dprint(FD_MUTEX, "wait on startup_mutex\n");
-                       if (fio_mutex_down_timeout(startup_mutex, 10000)) {
+                       dprint(FD_MUTEX, "wait on startup_sem\n");
+                       if (fio_sem_down_timeout(startup_sem, 10000)) {
                                log_err("fio: job startup hung? exiting.\n");
                                fio_terminate_threads(TERMINATE_ALL);
                                fio_abort = 1;
                                nr_started--;
+                               free(fd);
                                break;
                        }
-                       dprint(FD_MUTEX, "done waiting on startup_mutex\n");
+                       dprint(FD_MUTEX, "done waiting on startup_sem\n");
                }
 
                /*
@@ -2340,7 +2415,7 @@ reap:
                        m_rate += ddir_rw_sum(td->o.ratemin);
                        t_rate += ddir_rw_sum(td->o.rate);
                        todo--;
-                       fio_mutex_up(td->mutex);
+                       fio_sem_up(td->sem);
                }
 
                reap_threads(&nr_running, &t_rate, &m_rate);
@@ -2389,16 +2464,17 @@ int fio_backend(struct sk_out *sk_out)
                setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log");
        }
 
-       startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
-       if (startup_mutex == NULL)
+       startup_sem = fio_sem_init(FIO_SEM_LOCKED);
+       if (startup_sem == NULL)
                return 1;
 
        set_genesis_time();
        stat_init();
-       helper_thread_create(startup_mutex, sk_out);
+       helper_thread_create(startup_sem, sk_out);
 
        cgroup_list = smalloc(sizeof(*cgroup_list));
-       INIT_FLIST_HEAD(cgroup_list);
+       if (cgroup_list)
+               INIT_FLIST_HEAD(cgroup_list);
 
        run_threads(sk_out);
 
@@ -2417,19 +2493,24 @@ int fio_backend(struct sk_out *sk_out)
        }
 
        for_each_td(td, i) {
+               steadystate_free(td);
                fio_options_free(td);
                if (td->rusage_sem) {
-                       fio_mutex_remove(td->rusage_sem);
+                       fio_sem_remove(td->rusage_sem);
                        td->rusage_sem = NULL;
                }
+               fio_sem_remove(td->sem);
+               td->sem = NULL;
        }
 
        free_disk_util();
-       cgroup_kill(cgroup_list);
-       sfree(cgroup_list);
+       if (cgroup_list) {
+               cgroup_kill(cgroup_list);
+               sfree(cgroup_list);
+       }
        sfree(cgroup_mnt);
 
-       fio_mutex_remove(startup_mutex);
+       fio_sem_remove(startup_sem);
        stat_exit();
        return exit_value;
 }