Update fio io engine version
[fio.git] / io_u.c
diff --git a/io_u.c b/io_u.c
index 5331f8efee4d052e0aa555f80839fd9f6e66c191..31d3b14be50d0e124a8de277245cc043f4d1c624 100644 (file)
--- a/io_u.c
+++ b/io_u.c
@@ -52,7 +52,11 @@ static void mark_random_map(struct thread_data *td, struct io_u *io_u)
        while (blocks < nr_blocks) {
                unsigned int idx, bit;
 
-               if (!random_map_free(td, f, block))
+               /*
+                * If we have a mixed random workload, we may
+                * encounter blocks we already did IO to.
+                */
+               if (!td->o.ddir_nr && !random_map_free(td, f, block))
                        break;
 
                idx = RAND_MAP_IDX(td, f, block);
@@ -93,6 +97,35 @@ static int get_next_free_block(struct thread_data *td, struct fio_file *f,
        return 1;
 }
 
+static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
+                               int ddir, unsigned long long *b)
+{
+       unsigned long long max_blocks = f->io_size / td->o.min_bs[ddir];
+       unsigned long long r, rb;
+       int loops = 5;
+
+       do {
+               r = os_random_long(&td->random_state);
+               if (!max_blocks)
+                       *b = 0;
+               else
+                       *b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
+               if (td->o.norandommap)
+                       break;
+               rb = *b + (f->file_offset / td->o.min_bs[ddir]);
+               loops--;
+       } while (!random_map_free(td, f, rb) && loops);
+
+       /*
+        * if we failed to retrieve a truly random offset within
+        * the loops assigned, see if there are free ones left at all
+        */
+       if (!loops && get_next_free_block(td, f, b))
+               return 1;
+
+       return 0;
+}
+
 /*
  * For random io, generate a random new block and see if it's used. Repeat
  * until we find a free one. For sequential io, just return the end of
@@ -102,30 +135,12 @@ static int get_next_offset(struct thread_data *td, struct io_u *io_u)
 {
        struct fio_file *f = io_u->file;
        const int ddir = io_u->ddir;
-       unsigned long long b, rb;
-       long r;
+       unsigned long long b;
 
-       if (td_random(td)) {
-               unsigned long long max_blocks = f->file_size / td->o.min_bs[ddir];
-               int loops = 5;
-
-               do {
-                       r = os_random_long(&td->random_state);
-                       if (!max_blocks)
-                               b = 0;
-                       else
-                               b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
-                       if (td->o.norandommap)
-                               break;
-                       rb = b + (f->file_offset / td->o.min_bs[ddir]);
-                       loops--;
-               } while (!random_map_free(td, f, rb) && loops);
+       if (td_random(td) && (td->o.ddir_nr && !--td->ddir_nr)) {
+               td->ddir_nr = td->o.ddir_nr;
 
-               /*
-                * if we failed to retrieve a truly random offset within
-                * the loops assigned, see if there are free ones left at all
-                */
-               if (!loops && get_next_free_block(td, f, &b))
+               if (get_next_rand_offset(td, f, ddir, &b))
                        return 1;
        } else
                b = f->last_pos / td->o.min_bs[ddir];
@@ -169,6 +184,35 @@ static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
        return buflen;
 }
 
+static void set_rwmix_bytes(struct thread_data *td)
+{
+       unsigned long long rbytes;
+       unsigned int diff;
+
+       /*
+        * we do time or byte based switch. this is needed because
+        * buffered writes may issue a lot quicker than they complete,
+        * whereas reads do not.
+        */
+       rbytes = td->io_bytes[td->rwmix_ddir] - td->rwmix_bytes;
+       diff = td->o.rwmix[td->rwmix_ddir ^ 1];
+
+       td->rwmix_bytes = td->io_bytes[td->rwmix_ddir] + (rbytes * ((100 - diff)) / diff);
+}
+
+static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
+{
+       unsigned int v;
+       long r;
+
+       r = os_random_long(&td->rwmix_state);
+       v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
+       if (v < td->o.rwmix[DDIR_READ])
+               return DDIR_READ;
+
+       return DDIR_WRITE;
+}
+
 /*
  * Return the data direction for the next io_u. If the job is a
  * mixed read/write workload, check the rwmix cycle and switch if
@@ -179,23 +223,45 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
        if (td_rw(td)) {
                struct timeval now;
                unsigned long elapsed;
+               unsigned int cycle;
 
                fio_gettime(&now, NULL);
                elapsed = mtime_since_now(&td->rwmix_switch);
 
+               /*
+                * if this is the first cycle, make it shorter
+                */
+               cycle = td->o.rwmixcycle;
+               if (!td->rwmix_bytes)
+                       cycle /= 10;
+
                /*
                 * Check if it's time to seed a new data direction.
                 */
-               if (elapsed >= td->o.rwmixcycle) {
-                       unsigned int v;
-                       long r;
-
-                       r = os_random_long(&td->rwmix_state);
-                       v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
-                       if (v < td->o.rwmixread)
-                               td->rwmix_ddir = DDIR_READ;
-                       else
-                               td->rwmix_ddir = DDIR_WRITE;
+               if (elapsed >= cycle ||
+                   td->io_bytes[td->rwmix_ddir] >= td->rwmix_bytes) {
+                       unsigned long long max_bytes;
+                       enum fio_ddir ddir;                     
+
+                       /*
+                        * Put a top limit on how many bytes we do for
+                        * one data direction, to avoid overflowing the
+                        * ranges too much
+                        */
+                       ddir = get_rand_ddir(td);
+                       max_bytes = td->this_io_bytes[ddir];
+                       if (max_bytes >= (td->o.size * td->o.rwmix[ddir] / 100)) {
+                               if (!td->rw_end_set[ddir]) {
+                                       td->rw_end_set[ddir] = 1;
+                                       memcpy(&td->rw_end[ddir], &now, sizeof(now));
+                               }
+                               ddir ^= 1;
+                       }
+
+                       if (ddir != td->rwmix_ddir)
+                               set_rwmix_bytes(td);
+
+                       td->rwmix_ddir = ddir;
                        memcpy(&td->rwmix_switch, &now, sizeof(now));
                }
                return td->rwmix_ddir;
@@ -244,7 +310,7 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
           !(td->io_issues[DDIR_WRITE] % td->o.fsync_blocks) &&
             td->io_issues[DDIR_WRITE] && should_fsync(td)) {
                io_u->ddir = DDIR_SYNC;
-               return 0;
+               goto out;
        }
 
        io_u->ddir = get_rw_ddir(td);
@@ -263,12 +329,13 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
        /*
         * mark entry before potentially trimming io_u
         */
-       if (!td->o.read_iolog && td_random(td) && !td->o.norandommap)
+       if (td_random(td) && !td->o.norandommap)
                mark_random_map(td, io_u);
 
        /*
         * If using a write iolog, store this entry.
         */
+out:
        if (td->o.write_iolog_file)
                write_iolog_put(td, io_u);
 
@@ -594,13 +661,14 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
                add_bw_sample(td, idx, &icd->time);
                io_u_mark_latency(td, msec);
 
-               if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
+               if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE &&
+                   td->o.verify != VERIFY_NONE)
                        log_io_piece(td, io_u);
 
                icd->bytes_done[idx] += bytes;
 
                if (io_u->end_io) {
-                       ret = io_u->end_io(io_u);
+                       ret = io_u->end_io(td, io_u);
                        if (ret && !icd->error)
                                icd->error = ret;
                }
@@ -660,12 +728,10 @@ long io_u_queued_complete(struct thread_data *td, int min_events)
        struct io_completion_data icd;
        struct timespec *tvp = NULL;
        int ret;
+       struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
 
-       if (!min_events) {
-               struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
-
+       if (!min_events)
                tvp = &ts;
-       }
 
        ret = td_io_getevents(td, min_events, td->cur_depth, tvp);
        if (ret < 0) {