On Mon, Oct 07, 2019 at 04:37:28PM -0400, Josef Bacik wrote:
> On Mon, Oct 07, 2019 at 04:17:34PM -0400, Dennis Zhou wrote:
> > Async discard will use the free space cache as backing knowledge for
> > which extents to discard. This patch plumbs knowledge about which
> > extents need to be discarded into the free space cache from
> > unpin_extent_range().
> >
> > An untrimmed extent can merge with everything as this is a new region.
> > Absorbing trimmed extents is a tradeoff to for greater coalescing which
> > makes life better for find_free_extent(). Additionally, it seems the
> > size of a trim isn't as problematic as the trim io itself.
> >
> > When reading in the free space cache from disk, if sync is set, mark all
> > extents as trimmed. The current code ensures at transaction commit that
> > all free space is trimmed when sync is set, so this reflects that.
> >
> > Signed-off-by: Dennis Zhou <dennis@xxxxxxxxxx>
> > ---
> > fs/btrfs/extent-tree.c | 15 ++++++++++-----
> > fs/btrfs/free-space-cache.c | 38 ++++++++++++++++++++++++++++++-------
> > fs/btrfs/free-space-cache.h | 10 +++++++++-
> > fs/btrfs/inode-map.c | 13 +++++++------
> > 4 files changed, 57 insertions(+), 19 deletions(-)
> >
> > diff --git a/fs/btrfs/extent-tree.c b/fs/btrfs/extent-tree.c
> > index 77a5904756c5..b9e3bedad878 100644
> > --- a/fs/btrfs/extent-tree.c
> > +++ b/fs/btrfs/extent-tree.c
> > @@ -2782,7 +2782,7 @@ fetch_cluster_info(struct btrfs_fs_info *fs_info,
> > }
> >
> > static int unpin_extent_range(struct btrfs_fs_info *fs_info,
> > - u64 start, u64 end,
> > + u64 start, u64 end, u32 fsc_flags,
> > const bool return_free_space)
> > {
> > struct btrfs_block_group_cache *cache = NULL;
> > @@ -2816,7 +2816,9 @@ static int unpin_extent_range(struct btrfs_fs_info *fs_info,
> > if (start < cache->last_byte_to_unpin) {
> > len = min(len, cache->last_byte_to_unpin - start);
> > if (return_free_space)
> > - btrfs_add_free_space(cache, start, len);
> > + __btrfs_add_free_space(fs_info,
> > + cache->free_space_ctl,
> > + start, len, fsc_flags);
> > }
> >
> > start += len;
> > @@ -2894,6 +2896,7 @@ int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
> >
> > while (!trans->aborted) {
> > struct extent_state *cached_state = NULL;
> > + u32 fsc_flags = 0;
> >
> > mutex_lock(&fs_info->unused_bg_unpin_mutex);
> > ret = find_first_extent_bit(unpin, 0, &start, &end,
> > @@ -2903,12 +2906,14 @@ int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans)
> > break;
> > }
> >
> > - if (btrfs_test_opt(fs_info, DISCARD_SYNC))
> > + if (btrfs_test_opt(fs_info, DISCARD_SYNC)) {
> > ret = btrfs_discard_extent(fs_info, start,
> > end + 1 - start, NULL);
> > + fsc_flags |= BTRFS_FSC_TRIMMED;
> > + }
> >
> > clear_extent_dirty(unpin, start, end, &cached_state);
> > - unpin_extent_range(fs_info, start, end, true);
> > + unpin_extent_range(fs_info, start, end, fsc_flags, true);
> > mutex_unlock(&fs_info->unused_bg_unpin_mutex);
> > free_extent_state(cached_state);
> > cond_resched();
> > @@ -5512,7 +5517,7 @@ u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
> > int btrfs_error_unpin_extent_range(struct btrfs_fs_info *fs_info,
> > u64 start, u64 end)
> > {
> > - return unpin_extent_range(fs_info, start, end, false);
> > + return unpin_extent_range(fs_info, start, end, 0, false);
> > }
> >
> > /*
> > diff --git a/fs/btrfs/free-space-cache.c b/fs/btrfs/free-space-cache.c
> > index d54dcd0ab230..f119895292b8 100644
> > --- a/fs/btrfs/free-space-cache.c
> > +++ b/fs/btrfs/free-space-cache.c
> > @@ -747,6 +747,14 @@ static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
> > goto free_cache;
> > }
> >
> > + /*
> > + * Sync discard ensures that the free space cache is always
> > + * trimmed. So when reading this in, the state should reflect
> > + * that.
> > + */
> > + if (btrfs_test_opt(fs_info, DISCARD_SYNC))
> > + e->flags |= BTRFS_FSC_TRIMMED;
> > +
> > if (!e->bytes) {
> > kmem_cache_free(btrfs_free_space_cachep, e);
> > goto free_cache;
> > @@ -2165,6 +2173,7 @@ static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
> > bool merged = false;
> > u64 offset = info->offset;
> > u64 bytes = info->bytes;
> > + bool is_trimmed = btrfs_free_space_trimmed(info);
> >
> > /*
> > * first we want to see if there is free space adjacent to the range we
> > @@ -2178,7 +2187,8 @@ static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
> > else
> > left_info = tree_search_offset(ctl, offset - 1, 0, 0);
> >
> > - if (right_info && !right_info->bitmap) {
> > + if (right_info && !right_info->bitmap &&
> > + (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
> > if (update_stat)
> > unlink_free_space(ctl, right_info);
> > else
> > @@ -2189,7 +2199,8 @@ static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
> > }
> >
> > if (left_info && !left_info->bitmap &&
> > - left_info->offset + left_info->bytes == offset) {
> > + left_info->offset + left_info->bytes == offset &&
> > + (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
>
> So we allow merging if we haven't trimmed this entry, or if the adjacent entry
> is already trimmed? This means we'll merge if we trimmed the new entry
> regardless of the adjacent entries status, or if the new entry is drity. Why is
> that? Thanks,
>
This is the tradeoff I called out above here:
> > Absorbing trimmed extents is a tradeoff to for greater coalescing which
> > makes life better for find_free_extent(). Additionally, it seems the
> > size of a trim isn't as problematic as the trim io itself.
A problematic example case:
|----trimmed----|/////X/////|-----trimmed-----|
If region X gets freed and returned to the free space cache, we end up
with the following:
|----trimmed----|-untrimmed-|-----trimmed-----|
This isn't great because now we need to teach find_free_extent() to span
multiple btrfs_free_space entries, something I didn't want to do. So the
other option is to overtrim trading for a simpler find_free_extent().
Then the above becomes:
|-------------------trimmed-------------------|
It makes the assumption that if we're inserting, it's generally is free
space being returned rather than we needed to slice out from the middle
of a block. It does still have degenerative cases, but it's better than
the above. The merging also allows for stuff to come out of bitmaps more
proactively too.
Also from what it seems, the cost of a discard operation is quite costly
relative to the amount your discarding (1 larger discard is better than
several smaller discards) as it will clog up the device too.
Thanks,
Dennis
