Re: [PATCH v2 6/9] mm: set mapping error when launder_pages fails

[Trond Myklebust <trondmy@primarydata.com>]

On Wed, 2017-03-08 at 13:38 -0500, Jeff Layton wrote:
> On Wed, 2017-03-08 at 18:01 +0000, Trond Myklebust wrote:
> > On Wed, 2017-03-08 at 11:29 -0500, Jeff Layton wrote:
> > > If launder_page fails, then we hit a problem writing back some
> > > inode
> > > data. Ensure that we communicate that fact in a subsequent fsync
> > > since
> > > another task could still have it open for write.
> > > 
> > > Signed-off-by: Jeff Layton <jlayton@redhat.com>
> > > ---
> > >  mm/truncate.c | 6 +++++-
> > >  1 file changed, 5 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/mm/truncate.c b/mm/truncate.c
> > > index 6263affdef88..29ae420a5bf9 100644
> > > --- a/mm/truncate.c
> > > +++ b/mm/truncate.c
> > > @@ -594,11 +594,15 @@ invalidate_complete_page2(struct
> > > address_space
> > > *mapping, struct page *page)
> > >  
> > >  static int do_launder_page(struct address_space *mapping, struct
> > > page *page)
> > >  {
> > > +	int ret;
> > > +
> > >  	if (!PageDirty(page))
> > >  		return 0;
> > >  	if (page->mapping != mapping || mapping->a_ops-
> > > >launder_page 
> > > == NULL)
> > >  		return 0;
> > > -	return mapping->a_ops->launder_page(page);
> > > +	ret = mapping->a_ops->launder_page(page);
> > > +	mapping_set_error(mapping, ret);
> > > +	return ret;
> > >  }
> > >  
> > >  /**
> > 
> > No. At that layer, you don't know that this is a page error. In the
> > NFS
> > case, it could, for instance, just as well be a fatal signal.
> > 
> 
> Ok...don't we have the same problem with writepage then? Most of the
> writepage callers will set an error in the mapping if writepage
> returns
> any sort of error? A fatal signal in that codepath could cause the
> same
> problem, it seems. We don't dip into direct reclaim so much anymore,
> so
> maybe signals aren't an issue there?

If writepage() fails due to a signal, then it has the option of marking
the page as dirty and returning AOP_WRITEPAGE_ACTIVATE. That's not
possible for launder_page().

> The alternative here would be to push this down into the callers. I
> worry a bit though about getting this right across filesystems
> though.
> It'd be preferable it if we could keep the mapping_set_error call in
> generic VFS code instead, but if not then I'll just plan to do that.
> 



-- 
Trond Myklebust
Linux NFS client maintainer, PrimaryData
trond.myklebust@primarydata.com