Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752307AbaANV1C (ORCPT ); Tue, 14 Jan 2014 16:27:02 -0500 Received: from fieldses.org ([174.143.236.118]:39976 "EHLO fieldses.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752288AbaANV06 (ORCPT ); Tue, 14 Jan 2014 16:26:58 -0500 Date: Tue, 14 Jan 2014 16:26:52 -0500 From: "J. Bruce Fields" To: Andy Lutomirski Cc: Frank Filz , Jeff Layton , Linux FS Devel , nfs-ganesha-devel@lists.sourceforge.net, samba-technical@lists.samba.org, "linux-kernel@vger.kernel.org" , Richard Hipp Subject: Re: [PATCH v5 13/14] locks: skip deadlock detection on FL_FILE_PVT locks Message-ID: <20140114212652.GC23999@fieldses.org> References: <1389277187-18211-1-git-send-email-jlayton@redhat.com> <1389277187-18211-14-git-send-email-jlayton@redhat.com> <52CF05B5.5080700@amacapital.net> <20140109194930.1692fbbe@tlielax.poochiereds.net> <20140114192713.GA22262@fieldses.org> <20140114211009.GA23999@fieldses.org> <039101cf116e$56c2de80$04489b80$@mindspring.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: User-Agent: Mutt/1.5.21 (2010-09-15) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Tue, Jan 14, 2014 at 01:24:23PM -0800, Andy Lutomirski wrote: > On Tue, Jan 14, 2014 at 1:19 PM, Frank Filz wrote: > >> On Tue, Jan 14, 2014 at 12:29:17PM -0800, Andy Lutomirski wrote: > >> > [cc: drh, who I suspect is responsible for the most widespread > >> > userspace software that uses this stuff] > >> > > >> > On Tue, Jan 14, 2014 at 11:27 AM, J. Bruce Fields > >> wrote: > >> > > On Thu, Jan 09, 2014 at 04:58:59PM -0800, Andy Lutomirski wrote: > >> > >> On Thu, Jan 9, 2014 at 4:49 PM, Jeff Layton > >> wrote: > >> > >> > On Thu, 09 Jan 2014 12:25:25 -0800 Andy Lutomirski > >> > >> > wrote: > >> > >> >> When I think of deadlocks caused by r/w locks (which these are), > >> > >> >> I think of two kinds. First is what the current code tries to > >> > >> >> detect: two processes that are each waiting for each other. I > >> > >> >> don't know whether POSIX enshrines the idea of detecting that, > >> > >> >> but I wouldn't be surprised, considering how awful the old POSIX > >> locks are. > >> > > ... > >> > >> >> The sensible kind of detectable deadlock involves just one lock, > >> > >> >> and it happens when two processes both hold read locks and try > >> > >> >> to upgrade to write locks. This should be efficiently > >> > >> >> detectable and makes upgrading locks safe(r). > >> > > > >> > > This also involves two processes waiting on each other, and the > >> > > current code should detect either case equally well. > >> > > > >> > > ... > >> > >> For this kind of deadlock detection, nothing global is needed -- > >> > >> I'm only talking about detecting deadlocks due to two tasks > >> > >> upgrading locks on the same file (with overlapping ranges) at the > > same > >> time. > >> > >> > >> > >> This is actually useful for SQL-like things. Imagine this scenario: > >> > >> > >> > >> Program 1: > >> > >> > >> > >> Open a file > >> > >> BEGIN; > >> > >> SELECT whatever; -- acquires a read lock > >> > >> > >> > >> Program 2: > >> > >> > >> > >> Open the same file > >> > >> BEGIN; > >> > >> SELECT whatever; -- acquires a read lock > >> > >> > >> > >> Program 1: > >> > >> UPDATE something; -- upgrades to write > >> > >> > >> > >> Now program 1 is waiting for program 2 to release its lock. But if > >> > >> program 2 tries to UPDATE, then it deadlocks. A friendly MySQL > >> > >> implementation (which, sadly, does not include sqlite) will fail > >> > >> the abort the transaction instead. > >> > > > >> > > And then I suppose you'd need to get an exclusive lock when you > >> > > retry, to guarantee forward progress in the face of multiple > >> > > processes retrying at once. > >> > > >> > I don't think so -- as long as deadlock detection is 100% reliable and > >> > if you have writer priority, > >> > >> We don't have writer priority. Depending on how it worked I'm not > >> convinced it would help. E.g. consider the above but with 3 processes: > >> > >> processes 1, 2, and 3 each get a whole-file read lock. > >> > >> process 1 requests a write lock, blocks because it conflicts > >> with read locks held by 2 and 3. > >> > >> process 2 requests a write lock, gets -EDEADLK, unlocks and > >> requests a new read lock. That request succeeds because there > >> is no conflicting lock. (Note the lock manager had no > >> opportunity to upgrade 1's lock here thanks to the conflict with > >> 3's lock.) > > > > As I understand write lock priority, process 2 requesting a new read lock > > would block, once there is a write lock waiter, no further read locks would > > be granted that would conflict with that waiting write lock. > > ...which reminds me -- if anyone implements writer priority, please > make it optional (either w/ a writer-priority-ignoring read lock or a > non-priority-granting write lock). I have an application for which > writer priority would be really annoying. Is it something you could describe briefly? --b. > > Even better: Have read-lock-and-wait-for-pending-writers -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/