zbd: Support zone capacity smaller than zone size
[fio.git] / zbd.c
diff --git a/zbd.c b/zbd.c
index cf2cded93cff4371263106d3e7c0c9c02a6640c0..c738a58b206b4ee6a29f2837108f9a2c4ae06b21 100644 (file)
--- a/zbd.c
+++ b/zbd.c
@@ -140,6 +140,24 @@ static inline bool zbd_zone_swr(struct fio_zone_info *z)
        return z->type == ZBD_ZONE_TYPE_SWR;
 }
 
+/**
+ * zbd_zone_end - Return zone end location
+ * @z: zone info pointer.
+ */
+static inline uint64_t zbd_zone_end(const struct fio_zone_info *z)
+{
+       return (z+1)->start;
+}
+
+/**
+ * zbd_zone_capacity_end - Return zone capacity limit end location
+ * @z: zone info pointer.
+ */
+static inline uint64_t zbd_zone_capacity_end(const struct fio_zone_info *z)
+{
+       return z->start + z->capacity;
+}
+
 /**
  * zbd_zone_full - verify whether a minimum number of bytes remain in a zone
  * @f: file pointer.
@@ -154,7 +172,7 @@ static bool zbd_zone_full(const struct fio_file *f, struct fio_zone_info *z,
        assert((required & 511) == 0);
 
        return zbd_zone_swr(z) &&
-               z->wp + required > z->start + f->zbd_info->zone_size;
+               z->wp + required > zbd_zone_capacity_end(z);
 }
 
 static void zone_lock(struct thread_data *td, struct fio_file *f, struct fio_zone_info *z)
@@ -271,7 +289,7 @@ static bool zbd_verify_sizes(void)
                        z = &f->zbd_info->zone_info[zone_idx];
                        if ((f->file_offset != z->start) &&
                            (td->o.td_ddir != TD_DDIR_READ)) {
-                               new_offset = (z+1)->start;
+                               new_offset = zbd_zone_end(z);
                                if (new_offset >= f->file_offset + f->io_size) {
                                        log_info("%s: io_size must be at least one zone\n",
                                                 f->file_name);
@@ -384,6 +402,7 @@ static int init_zone_info(struct thread_data *td, struct fio_file *f)
                p->wp = p->start;
                p->type = ZBD_ZONE_TYPE_SWR;
                p->cond = ZBD_ZONE_COND_EMPTY;
+               p->capacity = zone_size;
        }
        /* a sentinel */
        p->start = nr_zones * zone_size;
@@ -456,10 +475,11 @@ static int parse_zone_info(struct thread_data *td, struct fio_file *f)
                        mutex_init_pshared_with_type(&p->mutex,
                                                     PTHREAD_MUTEX_RECURSIVE);
                        p->start = z->start;
+                       p->capacity = z->capacity;
                        switch (z->cond) {
                        case ZBD_ZONE_COND_NOT_WP:
                        case ZBD_ZONE_COND_FULL:
-                               p->wp = p->start + zone_size;
+                               p->wp = p->start + p->capacity;
                                break;
                        default:
                                assert(z->start <= z->wp);
@@ -707,7 +727,7 @@ static int zbd_reset_zone(struct thread_data *td, struct fio_file *f,
        dprint(FD_ZBD, "%s: resetting wp of zone %u.\n", f->file_name,
                zbd_zone_nr(f->zbd_info, z));
 
-       return zbd_reset_range(td, f, z->start, (z+1)->start - z->start);
+       return zbd_reset_range(td, f, z->start, zbd_zone_end(z) - z->start);
 }
 
 /* The caller must hold f->zbd_info->mutex */
@@ -1068,7 +1088,7 @@ found_candidate_zone:
        /* Both z->mutex and f->zbd_info->mutex are held. */
 
 examine_zone:
-       if (z->wp + min_bs <= (z+1)->start) {
+       if (z->wp + min_bs <= zbd_zone_capacity_end(z)) {
                pthread_mutex_unlock(&f->zbd_info->mutex);
                goto out;
        }
@@ -1112,7 +1132,7 @@ examine_zone:
                z = &f->zbd_info->zone_info[zone_idx];
 
                zone_lock(td, f, z);
-               if (z->wp + min_bs <= (z+1)->start)
+               if (z->wp + min_bs <= zbd_zone_capacity_end(z))
                        goto out;
                pthread_mutex_lock(&f->zbd_info->mutex);
        }
@@ -1143,9 +1163,9 @@ static struct fio_zone_info *zbd_replay_write_order(struct thread_data *td,
                assert(z);
        }
 
-       if (z->verify_block * min_bs >= f->zbd_info->zone_size)
+       if (z->verify_block * min_bs >= z->capacity)
                log_err("%s: %d * %d >= %llu\n", f->file_name, z->verify_block,
-                       min_bs, (unsigned long long) f->zbd_info->zone_size);
+                       min_bs, (unsigned long long)z->capacity);
        io_u->offset = z->start + z->verify_block++ * min_bs;
        return z;
 }
@@ -1231,7 +1251,7 @@ static void zbd_queue_io(struct io_u *io_u, int q, bool success)
        switch (io_u->ddir) {
        case DDIR_WRITE:
                zone_end = min((uint64_t)(io_u->offset + io_u->buflen),
-                              (z + 1)->start);
+                              zbd_zone_capacity_end(z));
                pthread_mutex_lock(&zbd_info->mutex);
                /*
                 * z->wp > zone_end means that one or more I/O errors
@@ -1327,6 +1347,26 @@ void setup_zbd_zone_mode(struct thread_data *td, struct io_u *io_u)
        assert(td->o.zone_mode == ZONE_MODE_ZBD);
        assert(td->o.zone_size);
 
+       zone_idx = zbd_zone_idx(f, f->last_pos[ddir]);
+       z = &f->zbd_info->zone_info[zone_idx];
+
+       /*
+        * When the zone capacity is smaller than the zone size and the I/O is
+        * sequential write, skip to zone end if the latest position is at the
+        * zone capacity limit.
+        */
+       if (z->capacity < f->zbd_info->zone_size && !td_random(td) &&
+           ddir == DDIR_WRITE &&
+           f->last_pos[ddir] >= zbd_zone_capacity_end(z)) {
+               dprint(FD_ZBD,
+                      "%s: Jump from zone capacity limit to zone end:"
+                      " (%lu -> %lu) for zone %u (%ld)\n",
+                      f->file_name, f->last_pos[ddir], zbd_zone_end(z),
+                      zbd_zone_nr(f->zbd_info, z), z->capacity);
+               td->io_skip_bytes += zbd_zone_end(z) - f->last_pos[ddir];
+               f->last_pos[ddir] = zbd_zone_end(z);
+       }
+
        /*
         * zone_skip is valid only for sequential workloads.
         */
@@ -1340,11 +1380,8 @@ void setup_zbd_zone_mode(struct thread_data *td, struct io_u *io_u)
         * - For reads with td->o.read_beyond_wp == false, the last position
         *   reached the zone write pointer.
         */
-       zone_idx = zbd_zone_idx(f, f->last_pos[ddir]);
-       z = &f->zbd_info->zone_info[zone_idx];
-
        if (td->zone_bytes >= td->o.zone_size ||
-           f->last_pos[ddir] >= (z+1)->start ||
+           f->last_pos[ddir] >= zbd_zone_end(z) ||
            (ddir == DDIR_READ &&
             (!td->o.read_beyond_wp) && f->last_pos[ddir] >= z->wp)) {
                /*
@@ -1530,6 +1567,13 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
                        zb->reset_zone = 0;
                        if (zbd_reset_zone(td, f, zb) < 0)
                                goto eof;
+
+                       if (zb->capacity < min_bs) {
+                               log_err("zone capacity %llu smaller than minimum block size %d\n",
+                                       (unsigned long long)zb->capacity,
+                                       min_bs);
+                               goto eof;
+                       }
                }
                /* Make writes occur at the write pointer */
                assert(!zbd_zone_full(f, zb, min_bs));
@@ -1545,7 +1589,7 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
                 * small.
                 */
                new_len = min((unsigned long long)io_u->buflen,
-                             (zb + 1)->start - io_u->offset);
+                             zbd_zone_capacity_end(zb) - io_u->offset);
                new_len = new_len / min_bs * min_bs;
                if (new_len == io_u->buflen)
                        goto accept;
@@ -1556,7 +1600,7 @@ enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u)
                        goto accept;
                }
                log_err("Zone remainder %lld smaller than minimum block size %d\n",
-                       ((zb + 1)->start - io_u->offset),
+                       (zbd_zone_capacity_end(zb) - io_u->offset),
                        min_bs);
                goto eof;
        case DDIR_TRIM: