Return-Path: linux-nfs-owner@vger.kernel.org Received: from fieldses.org ([174.143.236.118]:34008 "EHLO fieldses.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754402AbaGUVSG (ORCPT ); Mon, 21 Jul 2014 17:18:06 -0400 Date: Mon, 21 Jul 2014 17:17:57 -0400 From: "J. Bruce Fields" To: NeilBrown Cc: Jeff Layton , hch@infradead.org, linux-nfs@vger.kernel.org Subject: Re: [PATCH v4 10/10] nfsd: give block_delegation and delegation_blocked its own spinlock Message-ID: <20140721211757.GL8438@fieldses.org> References: <1405696416-32585-1-git-send-email-jlayton@primarydata.com> <1405696416-32585-11-git-send-email-jlayton@primarydata.com> <20140721170254.0289ab9f@notabene.brown> <20140721074412.4d9be086@tlielax.poochiereds.net> <20140722064049.6478e74d@notabene.brown> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii In-Reply-To: <20140722064049.6478e74d@notabene.brown> Sender: linux-nfs-owner@vger.kernel.org List-ID: On Tue, Jul 22, 2014 at 06:40:49AM +1000, NeilBrown wrote: > On Mon, 21 Jul 2014 07:44:12 -0400 Jeff Layton > wrote: > > > On Mon, 21 Jul 2014 17:02:54 +1000 > > NeilBrown wrote: > > > > > hash = arch_fast_hash(&fh->fh_base, fh->fh_size, 0); > > > > > > > > __set_bit(hash&255, bd->set[bd->new]); > > > > __set_bit((hash>>8)&255, bd->set[bd->new]); > > > > __set_bit((hash>>16)&255, bd->set[bd->new]); > > > > + spin_lock(&blocked_delegations_lock); > > > > > > __set_bit isn't atomic. The spin_lock should be taken *before* these > > > __set_bit() calls. > > > > > > Otherwise, looks fine. > > > > > > Thanks, > > > NeilBrown > > > > > > > > > > Ok. I guess the worry is that we could end up setting bits in the > > middle of swapping the two fields? Makes sense -- fixed in my repo. > > It is more subtle than that. > __set_bit() will: > read a value from memory to a register > set a bit in the register > write the register back out to memory > > If two threads both run __set_bit on the same word of memory at the same > time, one of the updates can get lost. > set_bit() (no underscore) performs an atomic RMW to avoid this, but is more > expensive. > spin_lock() obviously ensures the required exclusion and as we are going to > take the lock anyway we may as well take it before setting bits so we can use > the non-atomic (cheaper) __set_bit function. > > > I'll send out the updated set later today (it also includes a few nits > > that HCH pointed out last week). > > > > As a side note...I wonder how much we'll get in the way of false > > positives with this scheme? > > If a future version of NFSv4 could allow delegations to be granted while a > file is open (oh, it seems you are the only client using this file at the > moment, you can treat this "open" as a delegation if you like) a few false > positives would be a complete non-issue. For what it's worth, I think 4.1 provides what you're asking for here; see http://tools.ietf.org/html/rfc5661#section-20.7 and the discussion of the various WANT_ flags in http://tools.ietf.org/html/rfc5661#section-18.16.3 As far as I know none of that is implemented yet. --b.