Re: [PATCH 1/2] nfs: ensure correct writeback errors are returned on close()

[Scott Mayhew <smayhew@redhat.com>]

On Fri, 31 Jul 2020, Trond Myklebust wrote:

> On Fri, 2020-07-31 at 13:46 -0400, Scott Mayhew wrote:
> > nfs_wb_all() calls filemap_write_and_wait(), which uses
> > filemap_check_errors() to determine the error to return.
> > filemap_check_errors() only looks at the mapping->flags and will
> > therefore only return either -ENOSPC or -EIO.  To ensure that the
> > correct error is returned on close(), nfs{,4}_file_flush() should
> > call
> > file_check_and_advance_wb_err() which looks at the errseq value in
> > mapping->wb_err.
> > 
> > Fixes: 6fbda89b257f ("NFS: Replace custom error reporting mechanism
> > with
> > generic one")
> > Signed-off-by: Scott Mayhew <smayhew@redhat.com>
> > ---
> >  fs/nfs/file.c     | 3 ++-
> >  fs/nfs/nfs4file.c | 3 ++-
> >  2 files changed, 4 insertions(+), 2 deletions(-)
> > 
> > diff --git a/fs/nfs/file.c b/fs/nfs/file.c
> > index f96367a2463e..eeef6580052f 100644
> > --- a/fs/nfs/file.c
> > +++ b/fs/nfs/file.c
> > @@ -148,7 +148,8 @@ nfs_file_flush(struct file *file, fl_owner_t id)
> >  		return 0;
> >  
> >  	/* Flush writes to the server and return any errors */
> > -	return nfs_wb_all(inode);
> > +	nfs_wb_all(inode);
> > +	return file_check_and_advance_wb_err(file);
> >  }
> >  
> >  ssize_t
> > diff --git a/fs/nfs/nfs4file.c b/fs/nfs/nfs4file.c
> > index 8e5d6223ddd3..77bf9c12734c 100644
> > --- a/fs/nfs/nfs4file.c
> > +++ b/fs/nfs/nfs4file.c
> > @@ -125,7 +125,8 @@ nfs4_file_flush(struct file *file, fl_owner_t id)
> >  		return filemap_fdatawrite(file->f_mapping);
> >  
> >  	/* Flush writes to the server and return any errors */
> > -	return nfs_wb_all(inode);
> > +	nfs_wb_all(inode);
> > +	return file_check_and_advance_wb_err(file);
> >  }
> >  
> >  #ifdef CONFIG_NFS_V4_2
> 
> I don't think this one is correct. The contract with POSIX is that we
> always deliver the error on fsync(). If we call
> file_check_and_advance_wb_err() here in nfs_file_flush(), then that
> means we eat the error before it can get delivered to fsync().

I was looking at callers of the flush f_op and the only one I saw was
filp_close(), so I assumed that there wouldn't be any other calls to
fsync() for that struct file... I guess that's not the case if the file
descriptor was duplicated though.

Would a solution using filemap_sample_wb_err() & filemap_check_wb_err()
be acceptable (like in the 2nd patch)?

-Scott
> 
> -- 
> Trond Myklebust
> Linux NFS client maintainer, Hammerspace
> trond.myklebust@hammerspace.com
> 
>