On Tue, Jun 07, 2022 at 06:30:03PM +0800, Qu Wenruo wrote:
>
>
> On 2022/6/7 17:40, Filipe Manana wrote:
> > On Tue, Jun 07, 2022 at 01:09:13PM +0800, Qu Wenruo wrote:
> > > [BUG]
> > > After creating a dirty log tree, although
> > > btrfs_super_block::log_root and log_root_level is correctly populated,
> > > its generation is still left 0:
> > >
> > > log_root 30474240
> > > log_root_transid 0 <<<
> > > log_root_level 0
> > >
> > > [CAUSE]
> > > We just forgot to update btrfs_super_block::log_root_transid completely.
> > >
> > > Thus it's always the original value (0) from the initial super block.
> > >
> > > Thankfully this old behavior won't break log replay, as in
> > > btrfs_read_tree(), parent generation 0 means we just skip the generation
> >
> > btrfs_read_tree() does not exists, it's btrfs_read_tree_root().
> >
> > This is actually irrelevant, because we don't read the root log tree with
> > btrfs_read_tree_root(). We use read_tree_block() for that (at btrfs_replay_log()),
>
> Oh, right, I forgot to check that code, and just assumed every root read
> from superblock would has its generation checked, but it's not the case
> for log tree root.
>
> > and we use a generation matching the committed transaction + 1 (as it can never
> > be anything else).
> >
> > For every other log tree, we use btrfs_read_tree_root(), but the generation is
> > stored in the root's root item stored in the root log tree.
> >
> > The log_root_transid field was added to the super block by:
> >
> > commit c3027eb5523d6983f12628f3fe13d8a7576db701
> > Author: Chris Mason <chris.mason@xxxxxxxxxx>
> > Date: Mon Dec 8 16:40:21 2008 -0500
> >
> > Btrfs: Add inode sequence number for NFS and reserved space in a few structs
> >
> > But it was never used.
> >
> > So this change is not needed.
>
> Then, considering we have never really set log_root_transid anywhere,
> and in theory we're always using btrfs_super_block::generation + 1, why
It's not in theory only, it's in practice too.
We use committed generation + 1 since the first day the log tree was added:
commit e02119d5a7b4396c5a872582fddc8bd6d305a70a
Author: Chris Mason <chris.mason@xxxxxxxxxx>
Date: Fri Sep 5 16:13:11 2008 -0400
Btrfs: Add a write ahead tree log to optimize synchronous operations
(back then we read the root log tree directly at open_ctree()).
> not just deprecate that member?
>
> In fact, every time (thankfully not that many times for me) I checked
> log_root_transid, I can not help but to wonder why it's always 0.
You'd have to ask Chris why he added the field if it was never used and,
as far as I can see, was always useless since the life of a log tree
has never spanned more than 1 transaction, its generation must necessarily
be "committed transaction generation + 1".
Maybe just add a comment on top of the field saying it's unused and should
always be 0.
It's technically part of the disk format, so it's probably better not being
renamed to '__le64 unused;'.
Thanks.
>
> Thanks,
> Qu
>
> >
> > Thanks.
> >
> >
> > > check.
> > >
> > > And per-root log tree is still done properly using the root generation,
> > > so here we really only missed the generation check for log tree root,
> > > and even we fixed it, it should not cause any compatible problem.
> > >
> > > [FIX]
> > > Just update btrfs_super_block::log_root_transid properly.
> >
> > We don't need this.
> >
> > The log_root_transid field was added to the super block by:
> >
> > commit c3027eb5523d6983f12628f3fe13d8a7576db701
> > Author: Chris Mason <chris.mason@xxxxxxxxxx>
> > Date: Mon Dec 8 16:40:21 2008 -0500
> >
> > Btrfs: Add inode sequence number for NFS and reserved space in a few structs
> >
> > But it was never used.
> > For btrfs_read_tree_root(), what we use is the
> >
> >
> >
> > >
> > > Cc: stable@xxxxxxxxxxxxxxx #4.9+
> > > Signed-off-by: Qu Wenruo <wqu@xxxxxxxx>
> > > ---
> > > fs/btrfs/tree-log.c | 5 ++++-
> > > 1 file changed, 4 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/fs/btrfs/tree-log.c b/fs/btrfs/tree-log.c
> > > index 370388fadf96..27a76d6fef8c 100644
> > > --- a/fs/btrfs/tree-log.c
> > > +++ b/fs/btrfs/tree-log.c
> > > @@ -3083,7 +3083,8 @@ int btrfs_sync_log(struct btrfs_trans_handle *trans,
> > > struct btrfs_log_ctx root_log_ctx;
> > > struct blk_plug plug;
> > > u64 log_root_start;
> > > - u64 log_root_level;
> > > + u64 log_root_transid;
> > > + u8 log_root_level;
> > >
> > > mutex_lock(&root->log_mutex);
> > > log_transid = ctx->log_transid;
> > > @@ -3297,6 +3298,7 @@ int btrfs_sync_log(struct btrfs_trans_handle *trans,
> > >
> > > log_root_start = log_root_tree->node->start;
> > > log_root_level = btrfs_header_level(log_root_tree->node);
> > > + log_root_transid = btrfs_header_generation(log_root_tree->node);
> > > log_root_tree->log_transid++;
> > > mutex_unlock(&log_root_tree->log_mutex);
> > >
> > > @@ -3334,6 +3336,7 @@ int btrfs_sync_log(struct btrfs_trans_handle *trans,
> > >
> > > btrfs_set_super_log_root(fs_info->super_for_commit, log_root_start);
> > > btrfs_set_super_log_root_level(fs_info->super_for_commit, log_root_level);
> > > + btrfs_set_super_log_root_transid(fs_info->super_for_commit, log_root_transid);
> > > ret = write_all_supers(fs_info, 1);
> > > mutex_unlock(&fs_info->tree_log_mutex);
> > > if (ret) {
> > > --
> > > 2.36.1
> > >
