Re: [PATCH v4 03/10] fs: Introduce ->corrupted_range() for superblock

[Dan Williams <dan.j.williams@intel.com>]

On Wed, Jun 16, 2021 at 11:51 PM ruansy.fnst@fujitsu.com
<ruansy.fnst@fujitsu.com> wrote:
>
> > -----Original Message-----
> > From: Dan Williams <dan.j.williams@intel.com>
> > Subject: Re: [PATCH v4 03/10] fs: Introduce ->corrupted_range() for superblock
> >
> > [ drop old linux-nvdimm@lists.01.org, add nvdimm@lists.linux.dev ]
> >
> > On Thu, Jun 3, 2021 at 6:19 PM Shiyang Ruan <ruansy.fnst@fujitsu.com> wrote:
> > >
> > > Memory failure occurs in fsdax mode will finally be handled in
> > > filesystem.  We introduce this interface to find out files or metadata
> > > affected by the corrupted range, and try to recover the corrupted data
> > > if possiable.
> > >
> > > Signed-off-by: Shiyang Ruan <ruansy.fnst@fujitsu.com>
> > > ---
> > >  include/linux/fs.h | 2 ++
> > >  1 file changed, 2 insertions(+)
> > >
> > > diff --git a/include/linux/fs.h b/include/linux/fs.h index
> > > c3c88fdb9b2a..92af36c4225f 100644
> > > --- a/include/linux/fs.h
> > > +++ b/include/linux/fs.h
> > > @@ -2176,6 +2176,8 @@ struct super_operations {
> > >                                   struct shrink_control *);
> > >         long (*free_cached_objects)(struct super_block *,
> > >                                     struct shrink_control *);
> > > +       int (*corrupted_range)(struct super_block *sb, struct block_device
> > *bdev,
> > > +                              loff_t offset, size_t len, void *data);
> >
> > Why does the superblock need a new operation? Wouldn't whatever function is
> > specified here just be specified to the dax_dev as the
> > ->notify_failure() holder callback?
>
> Because we need to find out which file is effected by the given poison page so that memory-failure code can do collect_procs() and kill_procs() jobs.  And it needs filesystem to use its rmap feature to search the file from a given offset.  So, we need this implemented by the specified filesystem and called by dax_device's holder.
>
> This is the call trace I described in cover letter:
> memory_failure()
>  * fsdax case
>  pgmap->ops->memory_failure()      => pmem_pgmap_memory_failure()
>   dax_device->holder_ops->corrupted_range() =>
>                                       - fs_dax_corrupted_range()
>                                       - md_dax_corrupted_range()
>    sb->s_ops->currupted_range()    => xfs_fs_corrupted_range()  <== **HERE**
>     xfs_rmap_query_range()
>      xfs_currupt_helper()
>       * corrupted on metadata
>           try to recover data, call xfs_force_shutdown()
>       * corrupted on file data
>           try to recover data, call mf_dax_kill_procs()
>  * normal case
>  mf_generic_kill_procs()
>
> As you can see, this new added operation is an important for the whole progress.

I don't think you need either fs_dax_corrupted_range() nor
sb->s_ops->corrupted_range(). In fact that fs_dax_corrupted_range()
looks broken because the filesystem may not even be mounted on the
device associated with the error. The holder_data and holder_op should
be sufficient from communicating the stack of notifications:

pgmap->notify_memory_failure() => pmem_pgmap_notify_failure()
pmem_dax_dev->holder_ops->notify_failure(pmem_dax_dev) =>
md_dax_notify_failure()
md_dax_dev->holder_ops->notify_failure() => xfs_notify_failure()

I.e. the entire chain just walks dax_dev holder ops.