On 2021/11/10 03:24, Josef Bacik wrote:

On Tue, Nov 09, 2021 at 09:14:06AM +0800, Qu Wenruo wrote:

quoted

On 2021/11/9 03:36, Josef Bacik wrote:

quoted

On Sat, Nov 06, 2021 at 09:11:44AM +0800, Qu Wenruo wrote:

quoted

On 2021/11/6 04:49, Josef Bacik wrote:

quoted

This code adds the on disk structures for the block group root, which
will hold the block group items for extent tree v2.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
---
    fs/btrfs/ctree.h                | 26 ++++++++++++++++-
    fs/btrfs/disk-io.c              | 49 ++++++++++++++++++++++++++++-----
    fs/btrfs/disk-io.h              |  2 ++
    fs/btrfs/print-tree.c           |  1 +
    include/trace/events/btrfs.h    |  1 +
    include/uapi/linux/btrfs_tree.h |  3 ++
    6 files changed, 74 insertions(+), 8 deletions(-)

diff --git a/fs/btrfs/ctree.h b/fs/btrfs/ctree.h
index 8ec2f068a1c2..b57367141b95 100644
--- a/fs/btrfs/ctree.h
+++ b/fs/btrfs/ctree.h

@@ -271,8 +271,13 @@ struct btrfs_super_block {
    	/* the UUID written into btree blocks */
    	u8 metadata_uuid[BTRFS_FSID_SIZE];

+	__le64 block_group_root;
+	__le64 block_group_root_generation;
+	u8 block_group_root_level;
+

Is there any special reason that, block group root can't be put into
root tree?

Yes, I'm so glad you asked!

One of the planned changes with extent-tree-v2 is how we do relocation.  With no
longer being able to track metadata in the extent tree, relocation becomes much
more of a pain in the ass.

I'm even surprised that relocation can even be done without proper metadata
tracking in the new extent tree(s).

quoted

In addition, relocation currently has a pretty big problem, it can generate
unlimited delayed refs because it absolutely has to update all paths that point
to a relocated block in a single transaction.

Yep, that's also the biggest problem I attacked for the qgroup balance
optimization.

quoted

I'm fixing both of these problems with a new relocation thing, which will walk
through a block group, copy those extents to a new block group, and then update
a tree that maps the old logical address to the new logical address.

That sounds like the proposal from Johannes for zoned support of RAID56.
An FTL-like layer.

But I'm still not sure how we could even get all the tree blocks in one
block group in the first place, as there is no longer backref in the extent
tree(s).

By iterating all tree blocks? That doesn't sound sane to me...

No, iterating the free areas in the free space tree.  We no longer care about
the metadata itself, just the space that is utilized in the block group.  We
will mark the block group as read only, search through the free space tree for
that block group to find extents, copy them to new locations, insert a mapping
object for that block group to say "X range is now at Y".

OK, this makes sense now.

As extent's are free'd their new respective ranges are freed.  Once a relocated
block groups ->used hits 0 its mapping items are deleted.

quoted

quoted

Because of this we could end up with blocks in the tree root that need to be
remapped from a relocated block group into a new block group.  Thus we need to
be able to know what that mapping is before we go read the tree root.  This
means we have to store the block group root (and the new mapping root I'll
introduce later) in the super block.

Wouldn't the new mapping root becoming a new bottleneck then?

If we relocate the full fs, then the mapping root (block group root) would
be no different than an old extent tree?

Especially the mapping is done in extent level, not chunk level, thus it can
cause tons of mapping entries, really not that better than old extent tree
then.

Except the problem with the old extent tree is we are constantly modifying it.
The mapping's are never modified once they're created, unless we're remapping
and already remapped range.  Once the remapped extent is free'd it's new
location will be normal, and won't update anything in the mapping tree.

Oh, so the block group tree would be an colder version of extent tree, 
that would be really much nicer.

But that also means, to determine if a tree block/data extent is really 
belonging to a chunk/bg, we need to search for the new tree to be sure.

Would there be something to do reverse search? Or it may be a problem 
for balance again.

quoted

quoted

These two roots will behave like the chunk root, they'll have to be read first
in order to know where to find any remapped metadata blocks.  Because of that we
have to keep pointers to them in the super block instead of the tree root.

Got the reason now, but I'm not yet convinced the block group root mapping
is the proper way to go....

And still not sure how can we find out all tree blocks in one block group
without backref for each tree blocks...

We won't, we'll find allocated ranges.  It's certainly less precise than the
backref tree, but waaaaaaaay faster, because we only care about the range that
is allocated and moving that range.

Also it gives us another neat ability, we can relocate parts of extents instead
of being required to move full extents.  Before we had to move the whole extent
because we have to modify the file extent items to point at exactly the same
range.

Here the translation happens at the logical level, so we can easily split up
large extents and simply split up any bio's across the new logical locations and
stitch it back together at the end.  Thanks,

That really sounds better, and is what I'm going to do as a preparation 
for the iomap work.

Move all the bio split part into chunk layer, not in the current layer.

Thanks,
Qu

Josef