Re: [PATCH] ext4: cache es->s_journal_inum in ext4_sb_info

[Ojaswin Mujoo]

On Wed, Mar 26, 2025 at 12:01:45PM +0800, Zhang Yi wrote:
> On 2025/3/26 10:16, Baokun Li wrote:
> > On 2025/3/26 1:57, Ojaswin Mujoo wrote:
> >> On Tue, Mar 18, 2025 at 10:31:29PM -0400, Theodore Ts'o wrote:
> >>> On Tue, Mar 18, 2025 at 01:42:31PM +0530, Ojaswin Mujoo wrote:
> >>>>> So this is something we need to do if the journal is actived, and if
> >>>>> it's active, then sbi->s_journal will be non-NULL, and so we can just
> >>>>> check to see if inode == sbi->s_journal instead.  This will simplify
> >>>> I believe you mean inode == sbi->s_journal->j_inode here right?
> >>> Yes, that's what I meant; sorry for the not catching this before I
> >>> sent my reply.
> >>>
> >>> Cheers,
> >>>
> >>>                     - Ted
> >> Hi Ted, Baokun,
> >>
> >> I got some time to revisit this. Seems like checking against
> >> s_journal->j_inode is not enough. This is because both
> >> ext4_check_blockref() and check_block_validity() can be called even
> >> before journal->j_inode is set:
> >>
> >> ext4_open_inode_journal
> >>    ext4_get_journal_inode
> >>       __ext4_iget
> >>           ext4_ind_check_inode
> >>               ext4_check_blockref  /* j_inode not set */
> >>
> >>    journal = jbd2_journal_init_inode
> >>       bmap
> >>           ext4_bmap
> >>              iomap_bmap
> >>                ext4_iomap_begin
> >>                    ext4_map_blocks
> >>                        check_block_validity
> >>
> >>    journal->j_inode = inode
> >>
> >>
> >> Now, I think in this case the best solution might be to use the extra
> >> field like we do in this patch but set  EXT4_SB(sb)->s_journal_ino
> >> sufficiently early.
> >>
> >> Thoughts?
> > 
> > Because system zone setup happens after the journal are loaded, I think we
> > can skip the check if the journal haven't been loaded yet, like this:
> > 
> > diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c
> > index d04d8a7f12e7..38dc72ff7e78 100644
> > --- a/fs/ext4/inode.c
> > +++ b/fs/ext4/inode.c
> > @@ -383,9 +383,10 @@ static int __check_block_validity(struct inode *inode, const char *func,
> >                                 unsigned int line,
> >                                 struct ext4_map_blocks *map)
> >  {
> > +       journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
> > +
> >         if (ext4_has_feature_journal(inode->i_sb) &&
> > -           (inode->i_ino ==
> > - le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_journal_inum)))
> > +           (!journal || inode == journal->j_inode))
> >                 return 0;
> >         if (!ext4_inode_block_valid(inode, map->m_pblk, map->m_len)) {
> >                 ext4_error_inode(inode, func, line, map->m_pblk,
> > 
> > If any part of the journal area overlaps with the system zone, we'll catch
> > it when we add the journal area to the system zone later.
> > 
> > 
> 
> Since the creation of the system zone relies on the journal being
> loaded, I think there is no risk in proceeding to call
> ext4_inode_block_valid() to perform a basic block range check for
> the journal inode, or even better.
> 
> Thanks,
> Yi.

Got it Yi, makes sense to me. So I believe you are suggesting something
like:

@@ -384,9 +384,10 @@ static int __check_block_validity(struct inode *inode, const char *func,
                                unsigned int line,
                                struct ext4_map_blocks *map)
 {
+       journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
+
        if (ext4_has_feature_journal(inode->i_sb) &&
-           (inode->i_ino ==
-            le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_journal_inum)))
+           (journal && journal->j_inode == inode))
                return 0;
        if (!ext4_inode_block_valid(inode, map->m_pblk, map->m_len)) {
                ext4_error_inode(inode, func, line, map->m_pblk,

> 

So that even if it is a journal inode we can go ahead and perform some basic checks
as the system zone rbtree will anyways be NULL at this point. From a cursory look,
it seems that __ext4_iget(..., journal_inode) -> ext4_ext_check_inode() already relies
on the fact that system zone is NULL, so we should be okay here as well.

If this looks good, I'll send a v2 with the suggested changes.

Thanks,
ojaswin