Re: spurious sillyrename after O_DIRECT writes get ENOSPC

["J. Bruce Fields" <bfields@fieldses.org>]

On Fri, Dec 08, 2017 at 05:16:26PM -0500, J. Bruce Fields wrote:
> Last year Christoph noticed a bug that could result in a file being
> unnecessarily sillyrenamed after O_DIRECT writes get ENOSPC:
> 
> 	http://lkml.kernel.org/r/20160616150146.GA14015@infradead.org
> 
> It's reproduceable on upstream, over v3 or v4.
> 
> I looked into it some more, and it seems to reproduce whenever a write
> system call results in multiple WRITE calls, only some of which receive
> ENOSPC.  I think that's resulting in a leak of the wb_kref on some
> nfs_pages (possibly the ones corresponding to the ENOSPC failures?).
> Those nfs_pages in turn hold references on nfs_{lock,open}_contexts.  So
> a "rm" on the client (even after the file is closed) results in a
> sillyrename.
> 
> I'll keep looking at this, but the relevant code is pretty opaque to me
> so far.  Any ideas welcomed.

Actually it looks like a leak of dreq->io_count?  That prevents commits
from being sent (which I'm also seeing in network traces--the succesfull
WRITEs are unstable but never get committed), which means
nfs_direct_commit_complete() is never called, and the reference taken on
wb_kref in the request_commit case of nfs_direct_write_completion is
never put.

--b.