nanosecond: update completion latency recording and normal, json output to use nanose...
[fio.git] / io_u.c
diff --git a/io_u.c b/io_u.c
index d815ad9..375413f 100644 (file)
--- a/io_u.c
+++ b/io_u.c
 #include "lib/rand.h"
 #include "lib/axmap.h"
 #include "err.h"
+#include "lib/pow2.h"
+#include "minmax.h"
 
 struct io_completion_data {
        int nr;                         /* input */
 
        int error;                      /* output */
        uint64_t bytes_done[DDIR_RWDIR_CNT];    /* output */
-       struct timeval time;            /* output */
+       struct timespec time;           /* output */
 };
 
 /*
  * The ->io_axmap contains a map of blocks we have or have not done io
  * to yet. Used to make sure we cover the entire range in a fair fashion.
  */
-static int random_map_free(struct fio_file *f, const uint64_t block)
+static bool random_map_free(struct fio_file *f, const uint64_t block)
 {
        return !axmap_isset(f->io_axmap, block);
 }
@@ -60,6 +62,7 @@ static uint64_t last_block(struct thread_data *td, struct fio_file *f,
 
        /*
         * Hmm, should we make sure that ->io_size <= ->real_file_size?
+        * -> not for now since there is code assuming it could go either.
         */
        max_size = f->io_size;
        if (max_size > f->real_file_size)
@@ -68,6 +71,9 @@ static uint64_t last_block(struct thread_data *td, struct fio_file *f,
        if (td->o.zone_range)
                max_size = td->o.zone_range;
 
+       if (td->o.min_bs[ddir] > td->o.ba[ddir])
+               max_size -= td->o.min_bs[ddir] - td->o.ba[ddir];
+
        max_blocks = max_size / (uint64_t) td->o.ba[ddir];
        if (!max_blocks)
                return 0;
@@ -81,24 +87,25 @@ struct rand_off {
 };
 
 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
-                                 enum fio_ddir ddir, uint64_t *b)
+                                 enum fio_ddir ddir, uint64_t *b,
+                                 uint64_t lastb)
 {
-       uint64_t r, lastb;
+       uint64_t r;
 
-       lastb = last_block(td, f, ddir);
-       if (!lastb)
-               return 1;
+       if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
+           td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64) {
 
-       if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
                r = __rand(&td->random_state);
 
                dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
 
-               *b = lastb * (r / ((uint64_t) FRAND_MAX + 1.0));
+               *b = lastb * (r / (rand_max(&td->random_state) + 1.0));
        } else {
                uint64_t off = 0;
 
-               if (lfsr_next(&f->lfsr, &off, lastb))
+               assert(fio_file_lfsr(f));
+
+               if (lfsr_next(&f->lfsr, &off))
                        return 1;
 
                *b = off;
@@ -142,6 +149,79 @@ static int __get_next_rand_offset_pareto(struct thread_data *td,
        return 0;
 }
 
+static int __get_next_rand_offset_gauss(struct thread_data *td,
+                                       struct fio_file *f, enum fio_ddir ddir,
+                                       uint64_t *b)
+{
+       *b = gauss_next(&f->gauss);
+       return 0;
+}
+
+static int __get_next_rand_offset_zoned(struct thread_data *td,
+                                       struct fio_file *f, enum fio_ddir ddir,
+                                       uint64_t *b)
+{
+       unsigned int v, send, stotal;
+       uint64_t offset, lastb;
+       static int warned;
+       struct zone_split_index *zsi;
+
+       lastb = last_block(td, f, ddir);
+       if (!lastb)
+               return 1;
+
+       if (!td->o.zone_split_nr[ddir]) {
+bail:
+               return __get_next_rand_offset(td, f, ddir, b, lastb);
+       }
+
+       /*
+        * Generate a value, v, between 1 and 100, both inclusive
+        */
+       v = rand32_between(&td->zone_state, 1, 100);
+
+       zsi = &td->zone_state_index[ddir][v - 1];
+       stotal = zsi->size_perc_prev;
+       send = zsi->size_perc;
+
+       /*
+        * Should never happen
+        */
+       if (send == -1U) {
+               if (!warned) {
+                       log_err("fio: bug in zoned generation\n");
+                       warned = 1;
+               }
+               goto bail;
+       }
+
+       /*
+        * 'send' is some percentage below or equal to 100 that
+        * marks the end of the current IO range. 'stotal' marks
+        * the start, in percent.
+        */
+       if (stotal)
+               offset = stotal * lastb / 100ULL;
+       else
+               offset = 0;
+
+       lastb = lastb * (send - stotal) / 100ULL;
+
+       /*
+        * Generate index from 0..send-of-lastb
+        */
+       if (__get_next_rand_offset(td, f, ddir, b, lastb) == 1)
+               return 1;
+
+       /*
+        * Add our start offset, if any
+        */
+       if (offset)
+               *b += offset;
+
+       return 0;
+}
+
 static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
 {
        struct rand_off *r1 = flist_entry(a, struct rand_off, list);
@@ -153,12 +233,22 @@ static int flist_cmp(void *data, struct flist_head *a, struct flist_head *b)
 static int get_off_from_method(struct thread_data *td, struct fio_file *f,
                               enum fio_ddir ddir, uint64_t *b)
 {
-       if (td->o.random_distribution == FIO_RAND_DIST_RANDOM)
-               return __get_next_rand_offset(td, f, ddir, b);
-       else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
+       if (td->o.random_distribution == FIO_RAND_DIST_RANDOM) {
+               uint64_t lastb;
+
+               lastb = last_block(td, f, ddir);
+               if (!lastb)
+                       return 1;
+
+               return __get_next_rand_offset(td, f, ddir, b, lastb);
+       } else if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
                return __get_next_rand_offset_zipf(td, f, ddir, b);
        else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
                return __get_next_rand_offset_pareto(td, f, ddir, b);
+       else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
+               return __get_next_rand_offset_gauss(td, f, ddir, b);
+       else if (td->o.random_distribution == FIO_RAND_DIST_ZONED)
+               return __get_next_rand_offset_zoned(td, f, ddir, b);
 
        log_err("fio: unknown random distribution: %d\n", td->o.random_distribution);
        return 1;
@@ -168,30 +258,29 @@ static int get_off_from_method(struct thread_data *td, struct fio_file *f,
  * Sort the reads for a verify phase in batches of verifysort_nr, if
  * specified.
  */
-static inline int should_sort_io(struct thread_data *td)
+static inline bool should_sort_io(struct thread_data *td)
 {
        if (!td->o.verifysort_nr || !td->o.do_verify)
-               return 0;
+               return false;
        if (!td_random(td))
-               return 0;
+               return false;
        if (td->runstate != TD_VERIFYING)
-               return 0;
-       if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE)
-               return 0;
+               return false;
+       if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE ||
+           td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE64)
+               return false;
 
-       return 1;
+       return true;
 }
 
-static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
+static bool should_do_random(struct thread_data *td, enum fio_ddir ddir)
 {
        unsigned int v;
-       unsigned long r;
 
        if (td->o.perc_rand[ddir] == 100)
-               return 1;
+               return true;
 
-       r = __rand(&td->seq_rand_state[ddir]);
-       v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
+       v = rand32_between(&td->seq_rand_state[ddir], 1, 100);
 
        return v <= td->o.perc_rand[ddir];
 }
@@ -206,7 +295,6 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
                return get_off_from_method(td, f, ddir, b);
 
        if (!flist_empty(&td->next_rand_list)) {
-               struct rand_off *r;
 fetch:
                r = flist_first_entry(&td->next_rand_list, struct rand_off, list);
                flist_del(&r->list);
@@ -241,14 +329,15 @@ static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
        if (!get_next_rand_offset(td, f, ddir, b))
                return 0;
 
-       if (td->o.time_based) {
+       if (td->o.time_based ||
+           (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)) {
                fio_file_reset(td, f);
                if (!get_next_rand_offset(td, f, ddir, b))
                        return 0;
        }
 
        dprint(FD_IO, "%s: rand offset failed, last=%llu, size=%llu\n",
-                       f->file_name, (unsigned long long) f->last_pos,
+                       f->file_name, (unsigned long long) f->last_pos[ddir],
                        (unsigned long long) f->real_file_size);
        return 1;
 }
@@ -260,27 +349,49 @@ static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
 
        assert(ddir_rw(ddir));
 
-       if (f->last_pos >= f->io_size + get_start_offset(td, f) &&
-           o->time_based)
-               f->last_pos = f->last_pos - f->io_size;
+       if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
+           o->time_based) {
+               struct thread_options *o = &td->o;
+               uint64_t io_size = f->io_size + (f->io_size % o->min_bs[ddir]);
 
-       if (f->last_pos < f->real_file_size) {
+               if (io_size > f->last_pos[ddir])
+                       f->last_pos[ddir] = 0;
+               else
+                       f->last_pos[ddir] = f->last_pos[ddir] - io_size;
+       }
+
+       if (f->last_pos[ddir] < f->real_file_size) {
                uint64_t pos;
 
-               if (f->last_pos == f->file_offset && o->ddir_seq_add < 0)
-                       f->last_pos = f->real_file_size;
+               if (f->last_pos[ddir] == f->file_offset && o->ddir_seq_add < 0) {
+                       if (f->real_file_size > f->io_size)
+                               f->last_pos[ddir] = f->io_size;
+                       else
+                               f->last_pos[ddir] = f->real_file_size;
+               }
 
-               pos = f->last_pos - f->file_offset;
+               pos = f->last_pos[ddir] - f->file_offset;
                if (pos && o->ddir_seq_add) {
                        pos += o->ddir_seq_add;
 
                        /*
                         * If we reach beyond the end of the file
                         * with holed IO, wrap around to the
-                        * beginning again.
+                        * beginning again. If we're doing backwards IO,
+                        * wrap to the end.
                         */
-                       if (pos >= f->real_file_size)
-                               pos = f->file_offset;
+                       if (pos >= f->real_file_size) {
+                               if (o->ddir_seq_add > 0)
+                                       pos = f->file_offset;
+                               else {
+                                       if (f->real_file_size > f->io_size)
+                                               pos = f->io_size;
+                                       else
+                                               pos = f->real_file_size;
+
+                                       pos += o->ddir_seq_add;
+                               }
+                       }
                }
 
                *offset = pos;
@@ -309,7 +420,7 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
                                *is_random = 1;
                        } else {
                                *is_random = 0;
-                               io_u->flags |= IO_U_F_BUSY_OK;
+                               io_u_set(td, io_u, IO_U_F_BUSY_OK);
                                ret = get_next_seq_offset(td, f, ddir, &offset);
                                if (ret)
                                        ret = get_next_rand_block(td, f, ddir, &b);
@@ -319,7 +430,7 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
                        ret = get_next_seq_offset(td, f, ddir, &offset);
                }
        } else {
-               io_u->flags |= IO_U_F_BUSY_OK;
+               io_u_set(td, io_u, IO_U_F_BUSY_OK);
                *is_random = 0;
 
                if (td->o.rw_seq == RW_SEQ_SEQ) {
@@ -329,8 +440,8 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
                                *is_random = 0;
                        }
                } else if (td->o.rw_seq == RW_SEQ_IDENT) {
-                       if (f->last_start != -1ULL)
-                               offset = f->last_start - f->file_offset;
+                       if (f->last_start[ddir] != -1ULL)
+                               offset = f->last_start[ddir] - f->file_offset;
                        else
                                offset = 0;
                        ret = 0;
@@ -407,8 +518,8 @@ static int get_next_offset(struct thread_data *td, struct io_u *io_u,
        return __get_next_offset(td, io_u, is_random);
 }
 
-static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
-                           unsigned int buflen)
+static inline bool io_u_fits(struct thread_data *td, struct io_u *io_u,
+                            unsigned int buflen)
 {
        struct fio_file *f = io_u->file;
 
@@ -421,7 +532,8 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
        int ddir = io_u->ddir;
        unsigned int buflen = 0;
        unsigned int minbs, maxbs;
-       unsigned long r;
+       uint64_t frand_max, r;
+       bool power_2;
 
        assert(ddir_rw(ddir));
 
@@ -440,16 +552,17 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
        if (!io_u_fits(td, io_u, minbs))
                return 0;
 
+       frand_max = rand_max(&td->bsrange_state);
        do {
                r = __rand(&td->bsrange_state);
 
                if (!td->o.bssplit_nr[ddir]) {
                        buflen = 1 + (unsigned int) ((double) maxbs *
-                                       (r / (FRAND_MAX + 1.0)));
+                                       (r / (frand_max + 1.0)));
                        if (buflen < minbs)
                                buflen = minbs;
                } else {
-                       long perc = 0;
+                       long long perc = 0;
                        unsigned int i;
 
                        for (i = 0; i < td->o.bssplit_nr[ddir]; i++) {
@@ -457,19 +570,19 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
 
                                buflen = bsp->bs;
                                perc += bsp->perc;
-                               if ((r <= ((FRAND_MAX / 100L) * perc)) &&
+                               if (!perc)
+                                       break;
+                               if ((r / perc <= frand_max / 100ULL) &&
                                    io_u_fits(td, io_u, buflen))
                                        break;
                        }
                }
 
-               if (td->o.do_verify && td->o.verify != VERIFY_NONE)
-                       buflen = (buflen + td->o.verify_interval - 1) &
-                               ~(td->o.verify_interval - 1);
-
-               if (!td->o.bs_unaligned && is_power_of_2(minbs))
-                       buflen = (buflen + minbs - 1) & ~(minbs - 1);
-
+               power_2 = is_power_of_2(minbs);
+               if (!td->o.bs_unaligned && power_2)
+                       buflen &= ~(minbs - 1);
+               else if (!td->o.bs_unaligned && !power_2) 
+                       buflen -= buflen % minbs; 
        } while (!io_u_fits(td, io_u, buflen));
 
        return buflen;
@@ -504,10 +617,8 @@ static void set_rwmix_bytes(struct thread_data *td)
 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
 {
        unsigned int v;
-       unsigned long r;
 
-       r = __rand(&td->rwmix_state);
-       v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
+       v = rand32_between(&td->rwmix_state, 1, 100);
 
        if (v <= td->o.rwmix[DDIR_READ])
                return DDIR_READ;
@@ -515,8 +626,10 @@ static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
        return DDIR_WRITE;
 }
 
-void io_u_quiesce(struct thread_data *td)
+int io_u_quiesce(struct thread_data *td)
 {
+       int completed = 0;
+
        /*
         * We are going to sleep, ensure that we flush anything pending as
         * not to skew our latency numbers.
@@ -526,63 +639,70 @@ void io_u_quiesce(struct thread_data *td)
         * io's that have been actually submitted to an async engine,
         * and cur_depth is meaningless for sync engines.
         */
-       while (td->io_u_in_flight) {
+       if (td->io_u_queued || td->cur_depth) {
                int fio_unused ret;
 
-               ret = io_u_queued_complete(td, 1, NULL);
+               ret = td_io_commit(td);
        }
+
+       while (td->io_u_in_flight) {
+               int ret;
+
+               ret = io_u_queued_complete(td, 1);
+               if (ret > 0)
+                       completed += ret;
+       }
+
+       if (td->flags & TD_F_REGROW_LOGS)
+               regrow_logs(td);
+
+       return completed;
 }
 
 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
 {
        enum fio_ddir odir = ddir ^ 1;
-       struct timeval t;
        long usec;
+       uint64_t now;
 
        assert(ddir_rw(ddir));
+       now = utime_since_now(&td->start);
 
-       if (td->rate_pending_usleep[ddir] <= 0)
+       /*
+        * if rate_next_io_time is in the past, need to catch up to rate
+        */
+       if (td->rate_next_io_time[ddir] <= now)
                return ddir;
 
        /*
-        * We have too much pending sleep in this direction. See if we
+        * We are ahead of rate in this direction. See if we
         * should switch.
         */
        if (td_rw(td) && td->o.rwmix[odir]) {
                /*
-                * Other direction does not have too much pending, switch
+                * Other direction is behind rate, switch
                 */
-               if (td->rate_pending_usleep[odir] < 100000)
+               if (td->rate_next_io_time[odir] <= now)
                        return odir;
 
                /*
-                * Both directions have pending sleep. Sleep the minimum time
-                * and deduct from both.
+                * Both directions are ahead of rate. sleep the min
+                * switch if necissary
                 */
-               if (td->rate_pending_usleep[ddir] <=
-                       td->rate_pending_usleep[odir]) {
-                       usec = td->rate_pending_usleep[ddir];
+               if (td->rate_next_io_time[ddir] <=
+                       td->rate_next_io_time[odir]) {
+                       usec = td->rate_next_io_time[ddir] - now;
                } else {
-                       usec = td->rate_pending_usleep[odir];
+                       usec = td->rate_next_io_time[odir] - now;
                        ddir = odir;
                }
        } else
-               usec = td->rate_pending_usleep[ddir];
+               usec = td->rate_next_io_time[ddir] - now;
 
-       io_u_quiesce(td);
-
-       fio_gettime(&t, NULL);
-       usec_sleep(td, usec);
-       usec = utime_since_now(&t);
+       if (td->o.io_submit_mode == IO_MODE_INLINE)
+               io_u_quiesce(td);
 
-       td->rate_pending_usleep[ddir] -= usec;
-
-       odir = ddir ^ 1;
-       if (td_rw(td) && __should_check_rate(td, odir))
-               td->rate_pending_usleep[odir] -= usec;
-
-       if (ddir_trim(ddir))
-               return ddir;
+       usec = usec_sleep(td, usec);
 
        return ddir;
 }
@@ -597,28 +717,22 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
        enum fio_ddir ddir;
 
        /*
-        * see if it's time to fsync
-        */
-       if (td->o.fsync_blocks &&
-          !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
-            td->io_issues[DDIR_WRITE] && should_fsync(td))
-               return DDIR_SYNC;
-
-       /*
-        * see if it's time to fdatasync
-        */
-       if (td->o.fdatasync_blocks &&
-          !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks) &&
-            td->io_issues[DDIR_WRITE] && should_fsync(td))
-               return DDIR_DATASYNC;
-
-       /*
-        * see if it's time to sync_file_range
+        * See if it's time to fsync/fdatasync/sync_file_range first,
+        * and if not then move on to check regular I/Os.
         */
-       if (td->sync_file_range_nr &&
-          !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr) &&
-            td->io_issues[DDIR_WRITE] && should_fsync(td))
-               return DDIR_SYNC_FILE_RANGE;
+       if (should_fsync(td)) {
+               if (td->o.fsync_blocks && td->io_issues[DDIR_WRITE] &&
+                   !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks))
+                       return DDIR_SYNC;
+
+               if (td->o.fdatasync_blocks && td->io_issues[DDIR_WRITE] &&
+                   !(td->io_issues[DDIR_WRITE] % td->o.fdatasync_blocks))
+                       return DDIR_DATASYNC;
+
+               if (td->sync_file_range_nr && td->io_issues[DDIR_WRITE] &&
+                   !(td->io_issues[DDIR_WRITE] % td->sync_file_range_nr))
+                       return DDIR_SYNC_FILE_RANGE;
+       }
 
        if (td_rw(td)) {
                /*
@@ -642,8 +756,10 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
                ddir = DDIR_READ;
        else if (td_write(td))
                ddir = DDIR_WRITE;
-       else
+       else if (td_trim(td))
                ddir = DDIR_TRIM;
+       else
+               ddir = DDIR_INVAL;
 
        td->rwmix_ddir = rate_ddir(td, ddir);
        return td->rwmix_ddir;
@@ -651,13 +767,23 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 
 static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
 {
-       io_u->ddir = io_u->acct_ddir = get_rw_ddir(td);
+       enum fio_ddir ddir = get_rw_ddir(td);
 
-       if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
+       if (td_trimwrite(td)) {
+               struct fio_file *f = io_u->file;
+               if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
+                       ddir = DDIR_TRIM;
+               else
+                       ddir = DDIR_WRITE;
+       }
+
+       io_u->ddir = io_u->acct_ddir = ddir;
+
+       if (io_u->ddir == DDIR_WRITE && td_ioengine_flagged(td, FIO_BARRIER) &&
            td->o.barrier_blocks &&
           !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
             td->io_issues[DDIR_WRITE])
-               io_u->flags |= IO_U_F_BARRIER;
+               io_u_set(td, io_u, IO_U_F_BARRIER);
 }
 
 void put_file_log(struct thread_data *td, struct fio_file *f)
@@ -670,16 +796,21 @@ void put_file_log(struct thread_data *td, struct fio_file *f)
 
 void put_io_u(struct thread_data *td, struct io_u *io_u)
 {
+       if (td->parent)
+               td = td->parent;
+
        td_io_u_lock(td);
 
        if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
                put_file_log(td, io_u->file);
 
        io_u->file = NULL;
-       io_u->flags |= IO_U_F_FREE;
+       io_u_set(td, io_u, IO_U_F_FREE);
 
-       if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
+       if (io_u->flags & IO_U_F_IN_CUR_DEPTH) {
                td->cur_depth--;
+               assert(!(td->flags & TD_F_CHILD));
+       }
        io_u_qpush(&td->io_u_freelist, io_u);
        td_io_u_unlock(td);
        td_io_u_free_notify(td);
@@ -687,7 +818,7 @@ void put_io_u(struct thread_data *td, struct io_u *io_u)
 
 void clear_io_u(struct thread_data *td, struct io_u *io_u)
 {
-       io_u->flags &= ~IO_U_F_FLIGHT;
+       io_u_clear(td, io_u, IO_U_F_FLIGHT);
        put_io_u(td, io_u);
 }
 
@@ -698,18 +829,24 @@ void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 
        dprint(FD_IO, "requeue %p\n", __io_u);
 
+       if (td->parent)
+               td = td->parent;
+
        td_io_u_lock(td);
 
-       __io_u->flags |= IO_U_F_FREE;
+       io_u_set(td, __io_u, IO_U_F_FREE);
        if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
                td->io_issues[ddir]--;
 
-       __io_u->flags &= ~IO_U_F_FLIGHT;
-       if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
+       io_u_clear(td, __io_u, IO_U_F_FLIGHT);
+       if (__io_u->flags & IO_U_F_IN_CUR_DEPTH) {
                td->cur_depth--;
+               assert(!(td->flags & TD_F_CHILD));
+       }
 
        io_u_rpush(&td->io_u_requeues, __io_u);
        td_io_u_unlock(td);
+       td_io_u_free_notify(td);
        *io_u = NULL;
 }
 
@@ -717,7 +854,7 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 {
        unsigned int is_random;
 
-       if (td->io_ops->flags & FIO_NOIO)
+       if (td_ioengine_flagged(td, FIO_NOIO))
                goto out;
 
        set_rw_ddir(td, io_u);
@@ -742,7 +879,7 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
                 */
                if (f->file_offset >= f->real_file_size)
                        f->file_offset = f->real_file_size - f->file_offset;
-               f->last_pos = f->file_offset;
+               f->last_pos[io_u->ddir] = f->file_offset;
                td->io_skip_bytes += td->o.zone_skip;
        }
 
@@ -762,8 +899,9 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
        }
 
        if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
-               dprint(FD_IO, "io_u %p, offset too large\n", io_u);
-               dprint(FD_IO, "  off=%llu/%lu > %llu\n",
+               dprint(FD_IO, "io_u %p, offset + buflen exceeds file size\n",
+                       io_u);
+               dprint(FD_IO, "  offset=%llu/buflen=%lu > %llu\n",
                        (unsigned long long) io_u->offset, io_u->buflen,
                        (unsigned long long) io_u->file->real_file_size);
                return 1;
@@ -851,11 +989,52 @@ void io_u_mark_depth(struct thread_data *td, unsigned int nr)
        td->ts.io_u_map[idx] += nr;
 }
 
-static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
+static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
+{
+       int idx = 0;
+
+       assert(nsec < 1000);
+
+       switch (nsec) {
+       case 750 ... 999:
+               idx = 9;
+               break;
+       case 500 ... 749:
+               idx = 8;
+               break;
+       case 250 ... 499:
+               idx = 7;
+               break;
+       case 100 ... 249:
+               idx = 6;
+               break;
+       case 50 ... 99:
+               idx = 5;
+               break;
+       case 20 ... 49:
+               idx = 4;
+               break;
+       case 10 ... 19:
+               idx = 3;
+               break;
+       case 4 ... 9:
+               idx = 2;
+               break;
+       case 2 ... 3:
+               idx = 1;
+       case 0 ... 1:
+               break;
+       }
+
+       assert(idx < FIO_IO_U_LAT_N_NR);
+       td->ts.io_u_lat_n[idx]++;
+}
+
+static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
 {
        int idx = 0;
 
-       assert(usec < 1000);
+       assert(usec < 1000 && usec >= 1);
 
        switch (usec) {
        case 750 ... 999:
@@ -892,10 +1071,12 @@ static void io_u_mark_lat_usec(struct thread_data *td, unsigned long usec)
        td->ts.io_u_lat_u[idx]++;
 }
 
-static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
+static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
 {
        int idx = 0;
 
+       assert(msec >= 1);
+
        switch (msec) {
        default:
                idx = 11;
@@ -937,12 +1118,42 @@ static void io_u_mark_lat_msec(struct thread_data *td, unsigned long msec)
        td->ts.io_u_lat_m[idx]++;
 }
 
-static void io_u_mark_latency(struct thread_data *td, unsigned long usec)
+static void io_u_mark_latency(struct thread_data *td, unsigned long long nsec)
 {
-       if (usec < 1000)
-               io_u_mark_lat_usec(td, usec);
+       if (nsec < 1000)
+               io_u_mark_lat_nsec(td, nsec);
+       else if (nsec < 1000000)
+               io_u_mark_lat_usec(td, nsec / 1000);
        else
-               io_u_mark_lat_msec(td, usec / 1000);
+               io_u_mark_lat_msec(td, nsec / 1000000);
+}
+
+static unsigned int __get_next_fileno_rand(struct thread_data *td)
+{
+       unsigned long fileno;
+
+       if (td->o.file_service_type == FIO_FSERVICE_RANDOM) {
+               uint64_t frand_max = rand_max(&td->next_file_state);
+               unsigned long r;
+
+               r = __rand(&td->next_file_state);
+               return (unsigned int) ((double) td->o.nr_files
+                               * (r / (frand_max + 1.0)));
+       }
+
+       if (td->o.file_service_type == FIO_FSERVICE_ZIPF)
+               fileno = zipf_next(&td->next_file_zipf);
+       else if (td->o.file_service_type == FIO_FSERVICE_PARETO)
+               fileno = pareto_next(&td->next_file_zipf);
+       else if (td->o.file_service_type == FIO_FSERVICE_GAUSS)
+               fileno = gauss_next(&td->next_file_gauss);
+       else {
+               log_err("fio: bad file service type: %d\n", td->o.file_service_type);
+               assert(0);
+               return 0;
+       }
+
+       return fileno >> FIO_FSERVICE_SHIFT;
 }
 
 /*
@@ -957,11 +1168,8 @@ static struct fio_file *get_next_file_rand(struct thread_data *td,
 
        do {
                int opened = 0;
-               unsigned long r;
 
-               r = __rand(&td->next_file_state);
-               fno = (unsigned int) ((double) td->o.nr_files
-                               * (r / (FRAND_MAX + 1.0)));
+               fno = __get_next_fileno_rand(td);
 
                f = td->files[fno];
                if (fio_file_done(f))
@@ -1114,10 +1322,14 @@ static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
                put_file_log(td, f);
                td_io_close_file(td, f);
                io_u->file = NULL;
-               fio_file_set_done(f);
-               td->nr_done_files++;
-               dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
+               if (td->o.file_service_type & __FIO_FSERVICE_NONUNIFORM)
+                       fio_file_reset(td, f);
+               else {
+                       fio_file_set_done(f);
+                       td->nr_done_files++;
+                       dprint(FD_FILE, "%s: is done (%d of %d)\n", f->file_name,
                                        td->nr_done_files, td->o.nr_files);
+               }
        } while (1);
 
        return 0;
@@ -1143,10 +1355,10 @@ static void lat_new_cycle(struct thread_data *td)
  * We had an IO outside the latency target. Reduce the queue depth. If we
  * are at QD=1, then it's time to give up.
  */
-static int __lat_target_failed(struct thread_data *td)
+static bool __lat_target_failed(struct thread_data *td)
 {
        if (td->latency_qd == 1)
-               return 1;
+               return true;
 
        td->latency_qd_high = td->latency_qd;
 
@@ -1163,16 +1375,16 @@ static int __lat_target_failed(struct thread_data *td)
         */
        io_u_quiesce(td);
        lat_new_cycle(td);
-       return 0;
+       return false;
 }
 
-static int lat_target_failed(struct thread_data *td)
+static bool lat_target_failed(struct thread_data *td)
 {
        if (td->o.latency_percentile.u.f == 100.0)
                return __lat_target_failed(td);
 
        td->latency_failed++;
-       return 0;
+       return false;
 }
 
 void lat_target_init(struct thread_data *td)
@@ -1267,14 +1479,14 @@ void lat_target_check(struct thread_data *td)
  * If latency target is enabled, we might be ramping up or down and not
  * using the full queue depth available.
  */
-int queue_full(const struct thread_data *td)
+bool queue_full(const struct thread_data *td)
 {
        const int qempty = io_u_qempty(&td->io_u_freelist);
 
        if (qempty)
-               return 1;
+               return true;
        if (!td->o.latency_target)
-               return 0;
+               return false;
 
        return td->cur_depth >= td->latency_qd;
 }
@@ -1283,6 +1495,9 @@ struct io_u *__get_io_u(struct thread_data *td)
 {
        struct io_u *io_u = NULL;
 
+       if (td->stop_io)
+               return NULL;
+
        td_io_u_lock(td);
 
 again:
@@ -1299,21 +1514,23 @@ again:
 
        if (io_u) {
                assert(io_u->flags & IO_U_F_FREE);
-               io_u->flags &= ~(IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
+               io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
                                 IO_U_F_TRIMMED | IO_U_F_BARRIER |
                                 IO_U_F_VER_LIST);
 
                io_u->error = 0;
                io_u->acct_ddir = -1;
                td->cur_depth++;
-               io_u->flags |= IO_U_F_IN_CUR_DEPTH;
+               assert(!(td->flags & TD_F_CHILD));
+               io_u_set(td, io_u, IO_U_F_IN_CUR_DEPTH);
                io_u->ipo = NULL;
-       } else if (td->o.verify_async) {
+       } else if (td_async_processing(td)) {
                /*
                 * We ran out, wait for async verify threads to finish and
                 * return one
                 */
-               pthread_cond_wait(&td->free_cond, &td->io_u_lock);
+               assert(!(td->flags & TD_F_CHILD));
+               assert(!pthread_cond_wait(&td->free_cond, &td->io_u_lock));
                goto again;
        }
 
@@ -1321,10 +1538,10 @@ again:
        return io_u;
 }
 
-static int check_get_trim(struct thread_data *td, struct io_u *io_u)
+static bool check_get_trim(struct thread_data *td, struct io_u *io_u)
 {
        if (!(td->flags & TD_F_TRIM_BACKLOG))
-               return 0;
+               return false;
 
        if (td->trim_entries) {
                int get_trim = 0;
@@ -1340,17 +1557,17 @@ static int check_get_trim(struct thread_data *td, struct io_u *io_u)
                        get_trim = 1;
                }
 
-               if (get_trim && !get_next_trim(td, io_u))
-                       return 1;
+               if (get_trim && get_next_trim(td, io_u))
+                       return true;
        }
 
-       return 0;
+       return false;
 }
 
-static int check_get_verify(struct thread_data *td, struct io_u *io_u)
+static bool check_get_verify(struct thread_data *td, struct io_u *io_u)
 {
        if (!(td->flags & TD_F_VER_BACKLOG))
-               return 0;
+               return false;
 
        if (td->io_hist_len) {
                int get_verify = 0;
@@ -1367,11 +1584,11 @@ static int check_get_verify(struct thread_data *td, struct io_u *io_u)
 
                if (get_verify && !get_next_verify(td, io_u)) {
                        td->verify_batch--;
-                       return 1;
+                       return true;
                }
        }
 
-       return 0;
+       return false;
 }
 
 /*
@@ -1400,7 +1617,7 @@ static void small_content_scramble(struct io_u *io_u)
                 * the buffer, given by the product of the usec time
                 * and the actual offset.
                 */
-               offset = (io_u->start_time.tv_usec ^ boffset) & 511;
+               offset = ((io_u->start_time.tv_nsec/1000) ^ boffset) & 511;
                offset &= ~(sizeof(uint64_t) - 1);
                if (offset >= 512 - sizeof(uint64_t))
                        offset -= sizeof(uint64_t);
@@ -1462,19 +1679,19 @@ struct io_u *get_io_u(struct thread_data *td)
        assert(fio_file_open(f));
 
        if (ddir_rw(io_u->ddir)) {
-               if (!io_u->buflen && !(td->io_ops->flags & FIO_NOIO)) {
+               if (!io_u->buflen && !td_ioengine_flagged(td, FIO_NOIO)) {
                        dprint(FD_IO, "get_io_u: zero buflen on %p\n", io_u);
                        goto err_put;
                }
 
-               f->last_start = io_u->offset;
-               f->last_pos = io_u->offset + io_u->buflen;
+               f->last_start[io_u->ddir] = io_u->offset;
+               f->last_pos[io_u->ddir] = io_u->offset + io_u->buflen;
 
                if (io_u->ddir == DDIR_WRITE) {
                        if (td->flags & TD_F_REFILL_BUFFERS) {
                                io_u_fill_buffer(td, io_u,
                                        td->o.min_bs[DDIR_WRITE],
-                                       io_u->xfer_buflen);
+                                       io_u->buflen);
                        } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
                                   !(td->flags & TD_F_COMPRESS))
                                do_scramble = 1;
@@ -1500,10 +1717,12 @@ struct io_u *get_io_u(struct thread_data *td)
 out:
        assert(io_u->file);
        if (!td_io_prep(td, io_u)) {
-               if (!td->o.disable_slat)
+               if (!td->o.disable_lat)
                        fio_gettime(&io_u->start_time, NULL);
+
                if (do_scramble)
                        small_content_scramble(io_u);
+
                return io_u;
        }
 err_put:
@@ -1512,7 +1731,7 @@ err_put:
        return ERR_PTR(ret);
 }
 
-void io_u_log_error(struct thread_data *td, struct io_u *io_u)
+static void __io_u_log_error(struct thread_data *td, struct io_u *io_u)
 {
        enum error_type_bit eb = td_error_type(io_u->ddir, io_u->error);
 
@@ -1526,67 +1745,114 @@ void io_u_log_error(struct thread_data *td, struct io_u *io_u)
                io_ddir_name(io_u->ddir),
                io_u->offset, io_u->xfer_buflen);
 
+       if (td->io_ops->errdetails) {
+               char *err = td->io_ops->errdetails(io_u);
+
+               log_err("fio: %s\n", err);
+               free(err);
+       }
+
        if (!td->error)
                td_verror(td, io_u->error, "io_u error");
 }
 
-static inline int gtod_reduce(struct thread_data *td)
+void io_u_log_error(struct thread_data *td, struct io_u *io_u)
+{
+       __io_u_log_error(td, io_u);
+       if (td->parent)
+               __io_u_log_error(td->parent, io_u);
+}
+
+static inline bool gtod_reduce(struct thread_data *td)
 {
-       return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
-               && td->o.disable_bw;
+       return (td->o.disable_clat && td->o.disable_slat && td->o.disable_bw)
+                       || td->o.gtod_reduce;
 }
 
 static void account_io_completion(struct thread_data *td, struct io_u *io_u,
                                  struct io_completion_data *icd,
                                  const enum fio_ddir idx, unsigned int bytes)
 {
-       unsigned long lusec = 0;
+       const int no_reduce = !gtod_reduce(td);
+       unsigned long long llnsec = 0;
 
-       if (!gtod_reduce(td))
-               lusec = utime_since(&io_u->issue_time, &icd->time);
+       if (td->parent)
+               td = td->parent;
+
+       if (!td->o.stats)
+               return;
+
+       if (no_reduce)
+               llnsec = ntime_since(&io_u->issue_time, &icd->time);
 
        if (!td->o.disable_lat) {
-               unsigned long tusec;
+               unsigned long long tnsec;
 
-               tusec = utime_since(&io_u->start_time, &icd->time);
-               add_lat_sample(td, idx, tusec, bytes, io_u->offset);
+               tnsec = ntime_since(&io_u->start_time, &icd->time);
+               add_lat_sample(td, idx, tnsec, bytes, io_u->offset);
 
                if (td->flags & TD_F_PROFILE_OPS) {
                        struct prof_io_ops *ops = &td->prof_io_ops;
 
                        if (ops->io_u_lat)
-                               icd->error = ops->io_u_lat(td, tusec);
+                               icd->error = ops->io_u_lat(td, tnsec/1000);
                }
 
-               if (td->o.max_latency && tusec > td->o.max_latency)
-                       lat_fatal(td, icd, tusec, td->o.max_latency);
-               if (td->o.latency_target && tusec > td->o.latency_target) {
+               if (td->o.max_latency && tnsec/1000 > td->o.max_latency)
+                       lat_fatal(td, icd, tnsec/1000, td->o.max_latency);
+               if (td->o.latency_target && tnsec/1000 > td->o.latency_target) {
                        if (lat_target_failed(td))
-                               lat_fatal(td, icd, tusec, td->o.latency_target);
+                               lat_fatal(td, icd, tnsec/1000, td->o.latency_target);
                }
        }
 
-       if (!td->o.disable_clat) {
-               add_clat_sample(td, idx, lusec, bytes, io_u->offset);
-               io_u_mark_latency(td, lusec);
-       }
+       if (ddir_rw(idx)) {
+               if (!td->o.disable_clat) {
+                       add_clat_sample(td, idx, llnsec, bytes, io_u->offset);
+                       io_u_mark_latency(td, llnsec);
+               }
 
-       if (!td->o.disable_bw)
-               add_bw_sample(td, idx, bytes, &icd->time);
+               if (!td->o.disable_bw && per_unit_log(td->bw_log))
+                       add_bw_sample(td, io_u, bytes, llnsec);
 
-       if (!gtod_reduce(td))
-               add_iops_sample(td, idx, bytes, &icd->time);
+               if (no_reduce && per_unit_log(td->iops_log))
+                       add_iops_sample(td, io_u, bytes);
+       }
+
+       if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
+               uint32_t *info = io_u_block_info(td, io_u);
+               if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
+                       if (io_u->ddir == DDIR_TRIM) {
+                               *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
+                                               BLOCK_INFO_TRIMS(*info) + 1);
+                       } else if (io_u->ddir == DDIR_WRITE) {
+                               *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
+                                                               *info);
+                       }
+               }
+       }
 }
 
-static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
+static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
+                               uint64_t offset, unsigned int bytes)
 {
-       uint64_t secs, remainder, bps, bytes;
+       int idx;
+
+       if (!f)
+               return;
 
-       bytes = td->this_io_bytes[ddir];
-       bps = td->rate_bps[ddir];
-       secs = bytes / bps;
-       remainder = bytes % bps;
-       return remainder * 1000000 / bps + secs * 1000000;
+       if (f->first_write == -1ULL || offset < f->first_write)
+               f->first_write = offset;
+       if (f->last_write == -1ULL || ((offset + bytes) > f->last_write))
+               f->last_write = offset + bytes;
+
+       if (!f->last_write_comp)
+               return;
+
+       idx = f->last_write_idx++;
+       f->last_write_comp[idx] = offset;
+       if (f->last_write_idx == td->o.iodepth)
+               f->last_write_idx = 0;
 }
 
 static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
@@ -1598,9 +1864,8 @@ static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
 
        dprint_io_u(io_u, "io complete");
 
-       td_io_u_lock(td);
        assert(io_u->flags & IO_U_F_FLIGHT);
-       io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
+       io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
 
        /*
         * Mark IO ok to verify
@@ -1617,8 +1882,6 @@ static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
                }
        }
 
-       td_io_u_unlock(td);
-
        if (ddir_sync(ddir)) {
                td->last_was_sync = 1;
                if (f) {
@@ -1633,7 +1896,6 @@ static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
 
        if (!io_u->error && ddir_rw(ddir)) {
                unsigned int bytes = io_u->buflen - io_u->resid;
-               const enum fio_ddir oddir = ddir ^ 1;
                int ret;
 
                td->io_blocks[ddir]++;
@@ -1643,32 +1905,13 @@ static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
                if (!(io_u->flags & IO_U_F_VER_LIST))
                        td->this_io_bytes[ddir] += bytes;
 
-               if (ddir == DDIR_WRITE && f) {
-                       if (f->first_write == -1ULL ||
-                           io_u->offset < f->first_write)
-                               f->first_write = io_u->offset;
-                       if (f->last_write == -1ULL ||
-                           ((io_u->offset + bytes) > f->last_write))
-                               f->last_write = io_u->offset + bytes;
-               }
+               if (ddir == DDIR_WRITE)
+                       file_log_write_comp(td, f, io_u->offset, bytes);
 
                if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
-                                          td->runstate == TD_VERIFYING)) {
+                                          td->runstate == TD_VERIFYING))
                        account_io_completion(td, io_u, icd, ddir, bytes);
 
-                       if (__should_check_rate(td, ddir)) {
-                               td->rate_pending_usleep[ddir] =
-                                       (usec_for_io(td, ddir) -
-                                        utime_since_now(&td->start));
-                       }
-                       if (ddir != DDIR_TRIM &&
-                           __should_check_rate(td, oddir)) {
-                               td->rate_pending_usleep[oddir] =
-                                       (usec_for_io(td, oddir) -
-                                        utime_since_now(&td->start));
-                       }
-               }
-
                icd->bytes_done[ddir] += bytes;
 
                if (io_u->end_io) {
@@ -1710,7 +1953,7 @@ static void init_icd(struct thread_data *td, struct io_completion_data *icd,
        icd->nr = nr;
 
        icd->error = 0;
-       for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
+       for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
                icd->bytes_done[ddir] = 0;
 }
 
@@ -1733,10 +1976,10 @@ static void ios_completed(struct thread_data *td,
 /*
  * Complete a single io_u for the sync engines.
  */
-int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
-                      uint64_t *bytes)
+int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
 {
        struct io_completion_data icd;
+       int ddir;
 
        init_icd(td, &icd, 1);
        io_completed(td, &io_u, &icd);
@@ -1749,12 +1992,8 @@ int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
                return -1;
        }
 
-       if (bytes) {
-               int ddir;
-
-               for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
-                       bytes[ddir] += icd.bytes_done[ddir];
-       }
+       for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
+               td->bytes_done[ddir] += icd.bytes_done[ddir];
 
        return 0;
 }
@@ -1762,22 +2001,23 @@ int io_u_sync_complete(struct thread_data *td, struct io_u *io_u,
 /*
  * Called to complete min_events number of io for the async engines.
  */
-int io_u_queued_complete(struct thread_data *td, int min_evts,
-                        uint64_t *bytes)
+int io_u_queued_complete(struct thread_data *td, int min_evts)
 {
        struct io_completion_data icd;
        struct timespec *tvp = NULL;
-       int ret;
+       int ret, ddir;
        struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
 
-       dprint(FD_IO, "io_u_queued_completed: min=%d\n", min_evts);
+       dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
 
        if (!min_evts)
                tvp = &ts;
        else if (min_evts > td->cur_depth)
                min_evts = td->cur_depth;
 
-       ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete, tvp);
+       /* No worries, td_io_getevents fixes min and max if they are
+        * set incorrectly */
+       ret = td_io_getevents(td, min_evts, td->o.iodepth_batch_complete_max, tvp);
        if (ret < 0) {
                td_verror(td, -ret, "td_io_getevents");
                return ret;
@@ -1791,14 +2031,10 @@ int io_u_queued_complete(struct thread_data *td, int min_evts,
                return -1;
        }
 
-       if (bytes) {
-               int ddir;
+       for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
+               td->bytes_done[ddir] += icd.bytes_done[ddir];
 
-               for (ddir = DDIR_READ; ddir < DDIR_RWDIR_CNT; ddir++)
-                       bytes[ddir] += icd.bytes_done[ddir];
-       }
-
-       return 0;
+       return ret;
 }
 
 /*
@@ -1806,10 +2042,14 @@ int io_u_queued_complete(struct thread_data *td, int min_evts,
  */
 void io_u_queued(struct thread_data *td, struct io_u *io_u)
 {
-       if (!td->o.disable_slat) {
+       if (!td->o.disable_slat && ramp_time_over(td) && td->o.stats) {
                unsigned long slat_time;
 
-               slat_time = utime_since(&io_u->start_time, &io_u->issue_time);
+               slat_time = ntime_since(&io_u->start_time, &io_u->issue_time);
+
+               if (td->parent)
+                       td = td->parent;
+
                add_slat_sample(td, io_u->ddir, slat_time, io_u->xfer_buflen,
                                io_u->offset);
        }
@@ -1821,15 +2061,15 @@ void io_u_queued(struct thread_data *td, struct io_u *io_u)
 static struct frand_state *get_buf_state(struct thread_data *td)
 {
        unsigned int v;
-       unsigned long r;
 
        if (!td->o.dedupe_percentage)
                return &td->buf_state;
-       else if (td->o.dedupe_percentage == 100)
-               return &td->buf_state_prev;
+       else if (td->o.dedupe_percentage == 100) {
+               frand_copy(&td->buf_state_prev, &td->buf_state);
+               return &td->buf_state;
+       }
 
-       r = __rand(&td->dedupe_state);
-       v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
+       v = rand32_between(&td->dedupe_state, 1, 100);
 
        if (v <= td->o.dedupe_percentage)
                return &td->buf_state_prev;
@@ -1839,19 +2079,25 @@ static struct frand_state *get_buf_state(struct thread_data *td)
 
 static void save_buf_state(struct thread_data *td, struct frand_state *rs)
 {
-       if (rs == &td->buf_state)
+       if (td->o.dedupe_percentage == 100)
+               frand_copy(rs, &td->buf_state_prev);
+       else if (rs == &td->buf_state)
                frand_copy(&td->buf_state_prev, rs);
 }
 
 void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
                    unsigned int max_bs)
 {
-       if (td->o.buffer_pattern_bytes)
-               fill_buffer_pattern(td, buf, max_bs);
-       else if (!td->o.zero_buffers) {
+       struct thread_options *o = &td->o;
+
+       if (o->mem_type == MEM_CUDA_MALLOC)
+               return;
+
+       if (o->compress_percentage || o->dedupe_percentage) {
                unsigned int perc = td->o.compress_percentage;
                struct frand_state *rs;
                unsigned int left = max_bs;
+               unsigned int this_write;
 
                do {
                        rs = get_buf_state(td);
@@ -1859,23 +2105,28 @@ void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
                        min_write = min(min_write, left);
 
                        if (perc) {
-                               unsigned int seg = min_write;
-
-                               seg = min(min_write, td->o.compress_chunk);
-                               if (!seg)
-                                       seg = min_write;
+                               this_write = min_not_zero(min_write,
+                                                       td->o.compress_chunk);
 
-                               fill_random_buf_percentage(rs, buf, perc, seg,
-                                                               min_write);
-                       } else
+                               fill_random_buf_percentage(rs, buf, perc,
+                                       this_write, this_write,
+                                       o->buffer_pattern,
+                                       o->buffer_pattern_bytes);
+                       } else {
                                fill_random_buf(rs, buf, min_write);
+                               this_write = min_write;
+                       }
 
-                       buf += min_write;
-                       left -= min_write;
+                       buf += this_write;
+                       left -= this_write;
                        save_buf_state(td, rs);
                } while (left);
-       } else
+       } else if (o->buffer_pattern_bytes)
+               fill_buffer_pattern(td, buf, max_bs);
+       else if (o->zero_buffers)
                memset(buf, 0, max_bs);
+       else
+               fill_random_buf(get_buf_state(td), buf, max_bs);
 }
 
 /*
@@ -1887,3 +2138,61 @@ void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
        io_u->buf_filled_len = 0;
        fill_io_buffer(td, io_u->buf, min_write, max_bs);
 }
+
+static int do_sync_file_range(const struct thread_data *td,
+                             struct fio_file *f)
+{
+       off64_t offset, nbytes;
+
+       offset = f->first_write;
+       nbytes = f->last_write - f->first_write;
+
+       if (!nbytes)
+               return 0;
+
+       return sync_file_range(f->fd, offset, nbytes, td->o.sync_file_range);
+}
+
+int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
+{
+       int ret;
+
+       if (io_u->ddir == DDIR_SYNC) {
+               ret = fsync(io_u->file->fd);
+       } else if (io_u->ddir == DDIR_DATASYNC) {
+#ifdef CONFIG_FDATASYNC
+               ret = fdatasync(io_u->file->fd);
+#else
+               ret = io_u->xfer_buflen;
+               io_u->error = EINVAL;
+#endif
+       } else if (io_u->ddir == DDIR_SYNC_FILE_RANGE)
+               ret = do_sync_file_range(td, io_u->file);
+       else {
+               ret = io_u->xfer_buflen;
+               io_u->error = EINVAL;
+       }
+
+       if (ret < 0)
+               io_u->error = errno;
+
+       return ret;
+}
+
+int do_io_u_trim(const struct thread_data *td, struct io_u *io_u)
+{
+#ifndef FIO_HAVE_TRIM
+       io_u->error = EINVAL;
+       return 0;
+#else
+       struct fio_file *f = io_u->file;
+       int ret;
+
+       ret = os_trim(f->fd, io_u->offset, io_u->xfer_buflen);
+       if (!ret)
+               return io_u->xfer_buflen;
+
+       io_u->error = ret;
+       return 0;
+#endif
+}