Re: dcache: remove trylock loops (was Re: [BUG] lock_parent() breakage when used from shrink_dentry_list())

From: ebiederm@xmission.com (Eric W. Biederman)
To: Al Viro <viro@ZenIV.linux.org.uk>
Cc: John Ogness <john.ogness@linutronix.de>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	linux-fsdevel <linux-fsdevel@vger.kernel.org>,
	Christoph Hellwig <hch@lst.de>,
	Thomas Gleixner <tglx@linutronix.de>,
	Peter Zijlstra <peterz@infradead.org>,
	Sebastian Andrzej Siewior <bigeasy@linutronix.de>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
Subject: Re: dcache: remove trylock loops (was Re: [BUG] lock_parent() breakage when used from shrink_dentry_list())
Date: Mon, 12 Mar 2018 18:28:30 -0500	[thread overview]
Message-ID: <87woygan6p.fsf@xmission.com> (raw)
In-Reply-To: <20180312203916.GQ30522@ZenIV.linux.org.uk> (Al Viro's message of "Mon, 12 Mar 2018 20:39:17 +0000")

Al Viro <viro@ZenIV.linux.org.uk> writes:

> On Mon, Mar 12, 2018 at 03:23:44PM -0500, Eric W. Biederman wrote:
>
>> Of the two code paths you are concert about:
>> 
>> For path path_connected looking at s_root is a heuristic to avoid
>> calling is_subdir every time we need to do that check.  If the heuristic
>> fails we still have is_subdir which should remain accurate.  If
>> is_subdir fails the path is genuinely not connected at that moment
>> and failing is the correct thing to do.
>  
> Umm...  That might be not good enough - the logics is "everything's
> reachable from ->s_root anyway, so we might as well not bother checking".
> For NFS it's simply not true.

If I am parsing the code correctly path_connected is broken for nfsv2
and nfsv3 when NFS_MOUNT_UNSHARED is not set.   nfsv4 appears to make
a kernel mount of the real root of the filesystem properly setting
s_root and then finds the child it is mounting.

> We can mount server:/foo/bar/baz on /tmp/a, then server:/foo on /tmp/b
> and we'll have ->s_root pointing to a subtree of what's reachable at
> /tmp/b.  Play with renames under /tmp/b and you just might end up with
> a problem.  And mount on /tmp/a will be (mistakenly) considered to
> be safe, since it satisfies the heuristics in path_connected().

Agreed.

Which means that if you mount server:/foo/bar/baz first and then
mount server:/foo with an appropriate rename you might be able
to see all of server:/foo or possibly server:/

Hmm..

Given that nfs_kill_super calls generic_shutdown_super and
generic_shutdown_super calls shrink_dcache_for_umount
I would argue that nfsv2 and nfsv3 are buggy in the same
case, as shrink_dcache_for_umount is called on something
that is not the root of the filesystem's dentry tree.

Eric