On Mon, 2018-05-21 at 10:54 -0700, Darrick J. Wong wrote:
> On Sat, May 19, 2018 at 08:36:19AM -0700, Matthew Wilcox wrote:
> > On Sat, May 19, 2018 at 08:27:00AM -0700, Darrick J. Wong wrote:
> > > From: Darrick J. Wong <darrick.wong@xxxxxxxxxx>
> > >
> > > In inode_init_always(), we clear the inode mapping flags, which clears
> > > any retained error (AS_EIO, AS_ENOSC) bits. Unfortunately, we do not
> >
> > typo of ENOSPC in case you do a new version
>
> Fixed, thanks.
>
> > > also clear wb_err, which means that old mapping errors can leak through
> > > to new inodes.
> > >
> > > This is crucial for the XFS inode allocation path because we recycle old
> > > in-core inodes and we do not want error state from an old file to leak
> > > into the new file. This bug was discovered by running generic/036 and
> > > generic/047 in a loop and noticing that the EIOs generated by the
> > > collision of direct and buffered writes in generic/036 would survive the
> > > remount between 036 and 047, and get reported to the fsyncs (on
> > > different files on a reformatted fs!) in generic/047.
> > >
> > > Since we're changing the semantics of inode_init_always, we must also
> > > change xfs_reinit_inode to retain the writeback error state when we go
> > > to recover an inode that has been torn down in the vfs but not yet
> > > disposed of by XFS.
> >
> > Don't you also need to preserve inode->i_mapping->flags across the
> > reinitialisation for the users which aren't yet using ->wb_err?
>
> At first I thought (mistakenly) that xfs had been completely converted,
> but digging further we still use the old filemap_check_errors. It seems
> reasonable to me that if we're going to resurrect an incore inode then
> we should try to hang on to AS_EIO/AS_ENOSPC for as long as we can.
>
>
Yes, I think the patch you sent makes sense.
Most of the XFS callers are using filemap_write_and_wait{_range}, and
most of those seem to be called from ioctls (->setattr op being the
notable exception).
We could (in principle) pass a pointer to file->f_wb_err to most of
those callers, and use that to check for errors instead of looking for
AS_EIO/AS_ENOSPC. We wouldn't want to advance the error cursor for those
(as that should only be done in the context of an fsync), but it might
be more reliable than using the flags here.
> > > +++ b/fs/xfs/xfs_icache.c
> > > @@ -298,6 +298,7 @@ xfs_reinit_inode(
> > > uint64_t version = inode_peek_iversion(inode);
> > > umode_t mode = inode->i_mode;
> > > dev_t dev = inode->i_rdev;
> > > + errseq_t old_err = inode->i_mapping->wb_err;
> > >
> > > error = inode_init_always(mp->m_super, inode);
> > >
> > > @@ -306,6 +307,7 @@ xfs_reinit_inode(
> > > inode_set_iversion_queried(inode, version);
> > > inode->i_mode = mode;
> > > inode->i_rdev = dev;
> > > + inode->i_mapping->wb_err = old_err;
> > > return error;
> > > }
> > >
> >
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> > the body of a message to majordomo@xxxxxxxxxxxxxxx
> > More majordomo info at http://vger.kernel.org/majordomo-info.html
--
Jeff Layton <jlayton@xxxxxxxxxx>
