Re: [GIT PULL] Please pull NFS client bugfixes....

From: Linus Torvalds <torvalds@linux-foundation.org>
To: Trond Myklebust <Trond.Myklebust@netapp.com>
Cc: Andi Kleen <andi@firstfloor.org>, linux-kernel@vger.kernel.org
Subject: Re: [GIT PULL] Please pull NFS client bugfixes....
Date: Thu, 7 Jan 2010 17:12:29 -0800 (PST)	[thread overview]
Message-ID: <alpine.LFD.2.00.1001071707430.7821@localhost.localdomain> (raw)
In-Reply-To: <1262912601.2659.89.camel@localhost>

On Thu, 7 Jan 2010, Trond Myklebust wrote:
> > 
> > Because it means that you can trivially take page faults before the thing 
> > is validated (think threads).
> 
> Which would mean that another process/thread already has part of the
> file mmapped on the same client. I'm not arguing that have to revalidate
> in _that_ case.

No, I'm talking about the new mapping. Nothing else.

If the mmap'ing thread releases mmap_sem, and then does the revalidate, 
then you can have

	thread1				thread2
	-------				-------

	mmap
	map it in
	release mmap_sem
					page-fault the mapping before it got validated
	->post_mmap()
	revalidate outside mmap_sem

See? No "already part of the file mmapped" case at all. The exact mmap 
that you just set up - without the revalidation having happened.

In fact, because of this kind of _fundamental_ race, I don't see why I 
would ever accept any patches that add multiple mmap() down-calls at 
different phases to the filesystem at the VFS layer.

A filesystem that depends on the different phases would be a fundamentally 
buggy filesystem. Right now mmap is "atomic", and you can pre-populate (or 
pre-verify, like NFS does) the mapping in the _knowledge_ that there are 
no page faults that will populate it concurrently. Exactly because we hold 
the mmap_sem for writing.

				Linus