Fix bad latency reporting for rated IO jobs
[fio.git] / io_u.c
diff --git a/io_u.c b/io_u.c
index e9ce37ef6804670371d86e4833181c6cbc6abc64..7709df5fe55bf7fb6237d6b93476b2f09a2d3d3d 100644 (file)
--- a/io_u.c
+++ b/io_u.c
@@ -210,7 +210,10 @@ static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
                if (!get_next_free_block(td, f, ddir, b))
                        goto ret;
 
-               r = os_random_long(&td->random_state);
+               if (td->o.use_os_rand)
+                       r = os_random_long(&td->random_state);
+               else
+                       r = __rand(&td->__random_state);
        } while (--loops);
 
        /*
@@ -246,7 +249,12 @@ static int get_next_seq_block(struct thread_data *td, struct fio_file *f,
        assert(ddir_rw(ddir));
 
        if (f->last_pos < f->real_file_size) {
-               *b = (f->last_pos - f->file_offset) / td->o.min_bs[ddir];
+               unsigned long long pos = f->last_pos - f->file_offset;
+
+               if (pos)
+                       pos += td->o.ddir_seq_add;
+
+               *b = pos / td->o.min_bs[ddir];
                return 0;
        }
 
@@ -338,12 +346,20 @@ static int get_next_offset(struct thread_data *td, struct io_u *io_u)
        return __get_next_offset(td, io_u);
 }
 
+static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
+                           unsigned int buflen)
+{
+       struct fio_file *f = io_u->file;
+
+       return io_u->offset + buflen <= f->io_size + td->o.start_offset;
+}
+
 static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
 {
        const int ddir = io_u->ddir;
        unsigned int uninitialized_var(buflen);
        unsigned int minbs, maxbs;
-       long r;
+       unsigned long r, rand_max;
 
        assert(ddir_rw(ddir));
 
@@ -351,12 +367,22 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
        maxbs = td->o.max_bs[ddir];
 
        if (minbs == maxbs)
-               buflen = minbs;
-       else {
-               r = os_random_long(&td->bsrange_state);
+               return minbs;
+
+       if (td->o.use_os_rand)
+               rand_max = OS_RAND_MAX;
+       else
+               rand_max = FRAND_MAX;
+
+       do {
+               if (td->o.use_os_rand)
+                       r = os_random_long(&td->bsrange_state);
+               else
+                       r = __rand(&td->__bsrange_state);
+
                if (!td->o.bssplit_nr[ddir]) {
                        buflen = 1 + (unsigned int) ((double) maxbs *
-                                       (r / (OS_RAND_MAX + 1.0)));
+                                       (r / (rand_max + 1.0)));
                        if (buflen < minbs)
                                buflen = minbs;
                } else {
@@ -368,19 +394,16 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
 
                                buflen = bsp->bs;
                                perc += bsp->perc;
-                               if (r <= ((OS_RAND_MAX / 100L) * perc))
+                               if ((r <= ((rand_max / 100L) * perc)) &&
+                                   io_u_fits(td, io_u, buflen))
                                        break;
                        }
                }
+
                if (!td->o.bs_unaligned && is_power_of_2(minbs))
                        buflen = (buflen + minbs - 1) & ~(minbs - 1);
-       }
 
-       if (io_u->offset + buflen > io_u->file->real_file_size) {
-               dprint(FD_IO, "lower buflen %u -> %u (ddir=%d)\n", buflen,
-                                               minbs, ddir);
-               buflen = minbs;
-       }
+       } while (!io_u_fits(td, io_u, buflen));
 
        return buflen;
 }
@@ -411,10 +434,16 @@ static void set_rwmix_bytes(struct thread_data *td)
 static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
 {
        unsigned int v;
-       long r;
+       unsigned long r;
+
+       if (td->o.use_os_rand) {
+               r = os_random_long(&td->rwmix_state);
+               v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
+       } else {
+               r = __rand(&td->__rwmix_state);
+               v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
+       }
 
-       r = os_random_long(&td->rwmix_state);
-       v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
        if (v <= td->o.rwmix[DDIR_READ])
                return DDIR_READ;
 
@@ -457,6 +486,16 @@ static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
        } else
                usec = td->rate_pending_usleep[ddir];
 
+       /*
+        * We are going to sleep, ensure that we flush anything pending as
+        * not to skew our latency numbers
+        */
+       if (td->cur_depth) {
+               int fio_unused ret;
+
+               ret = io_u_queued_complete(td, td->cur_depth, NULL);
+       }
+
        fio_gettime(&t, NULL);
        usec_sleep(td, usec);
        usec = utime_since_now(&t);
@@ -830,11 +869,19 @@ static struct fio_file *get_next_file_rand(struct thread_data *td,
        int fno;
 
        do {
-               long r = os_random_long(&td->next_file_state);
                int opened = 0;
+               unsigned long r;
+
+               if (td->o.use_os_rand) {
+                       r = os_random_long(&td->next_file_state);
+                       fno = (unsigned int) ((double) td->o.nr_files
+                               * (r / (OS_RAND_MAX + 1.0)));
+               } else {
+                       r = __rand(&td->__next_file_state);
+                       fno = (unsigned int) ((double) td->o.nr_files
+                               * (r / (FRAND_MAX + 1.0)));
+               }
 
-               fno = (unsigned int) ((double) td->o.nr_files
-                       * (r / (OS_RAND_MAX + 1.0)));
                f = td->files[fno];
                if (fio_file_done(f))
                        continue;