On Mon, Jul 22, 2013 at 12:00:07PM -0400, Dwight Engen wrote:
> On Fri, 19 Jul 2013 16:02:21 +1000
> Dave Chinner <david@xxxxxxxxxxxxx> wrote:
>
> [...]
> > > diff --git a/fs/xfs/xfs_icache.c b/fs/xfs/xfs_icache.c
> > > index d873ab9e..728283a 100644
> > > --- a/fs/xfs/xfs_icache.c
> > > +++ b/fs/xfs/xfs_icache.c
> > > @@ -1247,6 +1247,10 @@ xfs_inode_free_eofblocks(
> > > if (!xfs_inode_match_id(ip, eofb))
> > > return 0;
> > >
> > > + if (eofb->eof_flags & XFS_EOF_FLAGS_PERM_CHECK &&
> > > + inode_permission(VFS_I(ip), MAY_WRITE))
> > > + return 0;
> >
> > This assumes we are walking fully instantiated VFS inodes. That's
> > not necessarily true - we may be walking inodes that have already
> > been dropped from the VFS and are waiting for background reclaim to
>
> Hi Dave, in looking at this a bit I don't see how they can be dropped
> from the VFS since they are igrab()ed in the flow:
>
> xfs_icache_free_eofblocks
> xfs_inode_ag_iterator_tag
> xfs_inode_ag_walk
> xfs_inode_ag_walk_grab
> igrab
Ah, right, I forgot that we have two different methods of walking
that tree, and the one that xfs_icache_free_eofblocks() avoids
inodes in reclaim.
> and I don't see a way for xfs_inode_free_eofblocks() to be called other
> than the ag_walk flow.
>
> If there is a way to get into xfs_inode_free_eofblocks where we can't
> use VFS_I(ip) then it will be a problem for the new code in
> xfs_inode_match_id() as well.
We can always use VFS_I(ip) because the struct inode is embedded in
the struct xfs_inode, so that's never an issue.
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
|