Scramble IO buffers unless explicitly told not
[fio.git] / io_u.c
diff --git a/io_u.c b/io_u.c
index af3b415160609d9dbb5317ef27c4d67d004be23b..a45dd4021c4e161f945367a3aec59dde0ebd4c57 100644 (file)
--- a/io_u.c
+++ b/io_u.c
@@ -68,6 +68,9 @@ 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;
@@ -83,32 +86,26 @@ struct rand_off {
 static int __get_next_rand_offset(struct thread_data *td, struct fio_file *f,
                                  enum fio_ddir ddir, uint64_t *b)
 {
-       uint64_t r, lastb;
-
-       lastb = last_block(td, f, ddir);
-       if (!lastb)
-               return 1;
+       uint64_t r;
 
        if (td->o.random_generator == FIO_RAND_GEN_TAUSWORTHE) {
-               uint64_t rmax;
+               uint64_t lastb;
 
-               rmax = td->o.use_os_rand ? OS_RAND_MAX : FRAND_MAX;
+               lastb = last_block(td, f, ddir);
+               if (!lastb)
+                       return 1;
 
-               if (td->o.use_os_rand) {
-                       rmax = OS_RAND_MAX;
-                       r = os_random_long(&td->random_state);
-               } else {
-                       rmax = FRAND_MAX;
-                       r = __rand(&td->__random_state);
-               }
+               r = __rand(&td->random_state);
 
                dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
 
-               *b = lastb * (r / ((uint64_t) rmax + 1.0));
+               *b = lastb * (r / ((uint64_t) FRAND_MAX + 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;
@@ -200,13 +197,8 @@ static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
        if (td->o.perc_rand[ddir] == 100)
                return 1;
 
-       if (td->o.use_os_rand) {
-               r = os_random_long(&td->seq_rand_state[ddir]);
-               v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
-       } else {
-               r = __rand(&td->__seq_rand_state[ddir]);
-               v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
-       }
+       r = __rand(&td->seq_rand_state[ddir]);
+       v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
 
        return v <= td->o.perc_rand[ddir];
 }
@@ -221,7 +213,6 @@ 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);
@@ -263,7 +254,7 @@ static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
        }
 
        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;
 }
@@ -271,20 +262,32 @@ static int get_next_rand_block(struct thread_data *td, struct fio_file *f,
 static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
                               enum fio_ddir ddir, uint64_t *offset)
 {
+       struct thread_options *o = &td->o;
+
        assert(ddir_rw(ddir));
 
-       if (f->last_pos >= f->io_size + get_start_offset(td, f) && td->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)
+               f->last_pos[ddir] = f->last_pos[ddir] - f->io_size;
 
-       if (f->last_pos < f->real_file_size) {
+       if (f->last_pos[ddir] < f->real_file_size) {
                uint64_t pos;
 
-               if (f->last_pos == f->file_offset && td->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)
+                       f->last_pos[ddir] = f->real_file_size;
+
+               pos = f->last_pos[ddir] - f->file_offset;
+               if (pos && o->ddir_seq_add) {
+                       pos += o->ddir_seq_add;
 
-               pos = f->last_pos - f->file_offset;
-               if (pos)
-                       pos += td->o.ddir_seq_add;
+                       /*
+                        * If we reach beyond the end of the file
+                        * with holed IO, wrap around to the
+                        * beginning again.
+                        */
+                       if (pos >= f->real_file_size)
+                               pos = f->file_offset;
+               }
 
                *offset = pos;
                return 0;
@@ -332,8 +335,8 @@ 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;
@@ -424,7 +427,7 @@ 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, rand_max;
+       unsigned long r;
 
        assert(ddir_rw(ddir));
 
@@ -443,20 +446,12 @@ 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;
 
-       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);
+               r = __rand(&td->bsrange_state);
 
                if (!td->o.bssplit_nr[ddir]) {
                        buflen = 1 + (unsigned int) ((double) maxbs *
-                                       (r / (rand_max + 1.0)));
+                                       (r / (FRAND_MAX + 1.0)));
                        if (buflen < minbs)
                                buflen = minbs;
                } else {
@@ -468,7 +463,7 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
 
                                buflen = bsp->bs;
                                perc += bsp->perc;
-                               if ((r <= ((rand_max / 100L) * perc)) &&
+                               if ((r <= ((FRAND_MAX / 100L) * perc)) &&
                                    io_u_fits(td, io_u, buflen))
                                        break;
                        }
@@ -517,13 +512,8 @@ static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
        unsigned int v;
        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 = __rand(&td->rwmix_state);
+       v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
 
        if (v <= td->o.rwmix[DDIR_READ])
                return DDIR_READ;
@@ -542,6 +532,12 @@ 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.
         */
+       if (td->io_u_queued || td->cur_depth) {
+               int fio_unused ret;
+
+               ret = td_io_commit(td);
+       }
+
        while (td->io_u_in_flight) {
                int fio_unused ret;
 
@@ -552,7 +548,6 @@ void io_u_quiesce(struct thread_data *td)
 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;
 
        assert(ddir_rw(ddir));
@@ -587,9 +582,7 @@ static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
 
        io_u_quiesce(td);
 
-       fio_gettime(&t, NULL);
-       usec_sleep(td, usec);
-       usec = utime_since_now(&t);
+       usec = usec_sleep(td, usec);
 
        td->rate_pending_usleep[ddir] -= usec;
 
@@ -597,9 +590,6 @@ static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
        if (td_rw(td) && __should_check_rate(td, odir))
                td->rate_pending_usleep[odir] -= usec;
 
-       if (ddir_trim(ddir))
-               return ddir;
-
        return ddir;
 }
 
@@ -748,9 +738,17 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
         * See if it's time to switch to a new zone
         */
        if (td->zone_bytes >= td->o.zone_size && td->o.zone_skip) {
+               struct fio_file *f = io_u->file;
+
                td->zone_bytes = 0;
-               io_u->file->file_offset += td->o.zone_range + td->o.zone_skip;
-               io_u->file->last_pos = io_u->file->file_offset;
+               f->file_offset += td->o.zone_range + td->o.zone_skip;
+
+               /*
+                * Wrap from the beginning, if we exceed the file size
+                */
+               if (f->file_offset >= f->real_file_size)
+                       f->file_offset = f->real_file_size - f->file_offset;
+               f->last_pos[io_u->ddir] = f->file_offset;
                td->io_skip_bytes += td->o.zone_skip;
        }
 
@@ -967,15 +965,9 @@ static struct fio_file *get_next_file_rand(struct thread_data *td,
                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 = __rand(&td->next_file_state);
+               fno = (unsigned int) ((double) td->o.nr_files
                                * (r / (FRAND_MAX + 1.0)));
-               }
 
                f = td->files[fno];
                if (fio_file_done(f))
@@ -1281,7 +1273,7 @@ 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(struct thread_data *td)
+int queue_full(const struct thread_data *td)
 {
        const int qempty = io_u_qempty(&td->io_u_freelist);
 
@@ -1297,6 +1289,9 @@ 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:
@@ -1481,14 +1476,16 @@ struct io_u *get_io_u(struct thread_data *td)
                        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,
-                                       io_u->xfer_buflen, io_u->xfer_buflen);
-                       } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
+                                       td->o.min_bs[DDIR_WRITE],
+                                       io_u->xfer_buflen);
+                       }
+                       if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
                                   !(td->flags & TD_F_COMPRESS))
                                do_scramble = 1;
                        if (td->flags & TD_F_VER_NONE) {
@@ -1656,13 +1653,22 @@ 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) {
+                       if (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 (td->last_write_comp) {
+                               int idx = td->last_write_idx++;
+
+                               td->last_write_comp[idx] = io_u->offset;
+                               if (td->last_write_idx == td->o.iodepth)
+                                       td->last_write_idx = 0;
+                       }
                }
 
                if (ramp_time_over(td) && (td->runstate == TD_RUNNING ||
@@ -1838,6 +1844,8 @@ static struct frand_state *get_buf_state(struct thread_data *td)
 
        if (!td->o.dedupe_percentage)
                return &td->buf_state;
+       else if (td->o.dedupe_percentage == 100)
+               return &td->buf_state_prev;
 
        r = __rand(&td->dedupe_state);
        v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
@@ -1857,29 +1865,41 @@ static void save_buf_state(struct thread_data *td, struct frand_state *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->compress_percentage || o->dedupe_percentage) {
                unsigned int perc = td->o.compress_percentage;
                struct frand_state *rs;
+               unsigned int left = max_bs;
 
-               rs = get_buf_state(td);
+               do {
+                       rs = get_buf_state(td);
 
-               if (perc) {
-                       unsigned int seg = min_write;
+                       min_write = min(min_write, left);
 
-                       seg = min(min_write, td->o.compress_chunk);
-                       if (!seg)
-                               seg = min_write;
+                       if (perc) {
+                               unsigned int seg = min_write;
 
-                       fill_random_buf_percentage(rs, buf, perc, seg,max_bs);
-                       save_buf_state(td, rs);
-               } else {
-                       fill_random_buf(rs, buf, max_bs);
+                               seg = min(min_write, td->o.compress_chunk);
+                               if (!seg)
+                                       seg = min_write;
+
+                               fill_random_buf_percentage(rs, buf, perc, seg,
+                                       min_write, o->buffer_pattern,
+                                                  o->buffer_pattern_bytes);
+                       } else
+                               fill_random_buf(rs, buf, min_write);
+
+                       buf += min_write;
+                       left -= min_write;
                        save_buf_state(td, rs);
-               }
-       } else
+               } while (left);
+       } 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);
 }
 
 /*