@@ -540,6 +540,134 @@ u64 add_new_free_space(struct btrfs_block_group *block_group, u64 start, u64 end
return total_added;
}
+/*
+ * Get an arbitrary extent item index / max_index through the block group
+ *
+ * @block_group: the block group to sample from
+ * @index: the integral step through the block group to grab from
+ * @max_index: the granularity of the sampling
+ * @key: return value parameter for the item we find
+ *
+ * pre-conditions on indices:
+ * 0 <= index <= max_index
+ * 0 < max_index
+ *
+ * Returns: 0 on success, 1 if the search didn't yield a useful item, negative
+ * error code on error.
+ */
+static int sample_block_group_extent_item(struct btrfs_block_group *block_group,
+ int index, int max_index,
+ struct btrfs_key *key)
+{
+ struct btrfs_fs_info *fs_info = block_group->fs_info;
+ struct btrfs_root *extent_root;
+ int ret = 0;
+ u64 search_offset;
+ struct btrfs_path *path;
+
+ ASSERT(index >= 0);
+ ASSERT(index <= max_index);
+ ASSERT(max_index > 0);
+
+ path = btrfs_alloc_path();
+ if (!path)
+ return -ENOMEM;
+
+ down_read(&fs_info->commit_root_sem);
+ extent_root = btrfs_extent_root(fs_info, max_t(u64, block_group->start,
+ BTRFS_SUPER_INFO_OFFSET));
+
+ path->skip_locking = 1;
+ path->search_commit_root = 1;
+ path->reada = READA_FORWARD;
+
+ search_offset = index * (block_group->length / max_index);
+ key->objectid = block_group->start + search_offset;
+ key->offset = 0;
+ key->type = BTRFS_EXTENT_ITEM_KEY;
+
+ ret = btrfs_search_slot(NULL, extent_root, key, path, 0, 0);
+ if (ret != 0)
+ goto out;
+ if (key->objectid < block_group->start ||
+ key->objectid > block_group->start + block_group->length) {
+ ret = 1;
+ goto out;
+ }
+ if (key->type != BTRFS_EXTENT_ITEM_KEY) {
+ ret = 1;
+ goto out;
+ }
+out:
+ btrfs_free_path(path);
+ up_read(&fs_info->commit_root_sem);
+ return ret;
+}
+
+/*
+ * Best effort attempt to compute a block group's size class while caching it.
+ *
+ * @block_group: the block group we are caching
+ *
+ * We cannot infer the size class while adding free space extents, because that
+ * logic doesn't care about contiguous file extents (it doesn't differentiate
+ * between a 100M extent and 100 contiguous 1M extents). So we need to read the
+ * file extent items. Reading all of them is quite wasteful, because usually
+ * only a handful are enough to give a good answer. Therefore, we just grab 5 of
+ * them at even steps through the block group and pick the smallest size class
+ * we see. Since size class is best effort, and not guaranteed in general,
+ * inaccuracy is acceptable.
+ *
+ * To be more explicit about why this algorithm makes sense:
+ *
+ * If we are caching in a block group from disk, then there are three major cases
+ * to consider:
+ * 1. the block group is well behaved and all extents in it are the same size
+ * class.
+ * 2. the block group is mostly one size class with rare exceptions for last
+ * ditch allocations
+ * 3. the block group was populated before size classes and can have a totally
+ * arbitrary mix of size classes.
+ *
+ * In case 1, looking at any extent in the block group will yield the correct
+ * result. For the mixed cases, taking the minimum size class seems like a good
+ * approximation, since gaps from frees will be usable to the size class. For
+ * 2., a small handful of file extents is likely to yield the right answer. For
+ * 3, we can either read every file extent, or admit that this is best effort
+ * anyway and try to stay fast.
+ *
+ * Returns: 0 on success, negative error code on error.
+ */
+static int load_block_group_size_class(struct btrfs_block_group *block_group)
+{
+ struct btrfs_key key;
+ int i;
+ u64 min_size = block_group->length;
+ enum btrfs_block_group_size_class size_class = BTRFS_BG_SZ_NONE;
+ int ret;
+
+ if (!btrfs_is_block_group_data_only(block_group))
+ return 0;
+
+ for (i = 0; i < 5; ++i) {
+ ret = sample_block_group_extent_item(block_group, i, 5, &key);
+ if (ret < 0)
+ goto out;
+ if (ret > 0)
+ continue;
+ min_size = min_t(u64, min_size, key.offset);
+ size_class = btrfs_calc_block_group_size_class(min_size);
+ }
+ if (size_class != BTRFS_BG_SZ_NONE) {
+ spin_lock(&block_group->lock);
+ block_group->size_class = size_class;
+ spin_unlock(&block_group->lock);
+ }
+
+out:
+ return ret;
+}
+
static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
{
struct btrfs_block_group *block_group = caching_ctl->block_group;
@@ -739,6 +867,8 @@ static noinline void caching_thread(struct btrfs_work *work)
wake_up(&caching_ctl->wait);
+ load_block_group_size_class(block_group);
+
btrfs_put_caching_control(caching_ctl);
btrfs_put_block_group(block_group);
}
Since the size class is an artifact of an arbitrary anti fragmentation strategy, it doesn't really make sense to persist it. Furthermore, most of the size class logic assumes fresh block groups. That is of course not a reasonable assumption -- we will be upgrading kernels with existing filesystems whose block groups are not classified. To work around those issues, implement logic to compute the size class of the block groups as we cache them in. To perfectly assess the state of a block group, we would have to read the entire extent tree (since the free space cache mashes together contiguous extent items) which would be prohibitively expensive for larger file systems with more extents. We can do it relatively cheaply by implementing a simple heuristic of sampling a handful of extents and picking the smallest one we see. In the happy case where the block group was classified, we will only see extents of the correct size. In the unhappy case, we will hopefully find one of the smaller extents, but there is no perfect answer anyway. Autorelocation will eventually churn up the block group if there is significant free-ing anyway. The work is done in the caching thread but after marking the block group cached, as we tradeoff classification accuracy vs. slowing down allocations. There was no regression in mount performance at end state of the fsperf test suite. Signed-off-by: Boris Burkov <boris@bur.io> --- fs/btrfs/block-group.c | 130 +++++++++++++++++++++++++++++++++++++++++ 1 file changed, 130 insertions(+)