Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1757868AbXFOVEy (ORCPT ); Fri, 15 Jun 2007 17:04:54 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1754856AbXFOVEq (ORCPT ); Fri, 15 Jun 2007 17:04:46 -0400 Received: from e4.ny.us.ibm.com ([32.97.182.144]:32790 "EHLO e4.ny.us.ibm.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753033AbXFOVEp (ORCPT ); Fri, 15 Jun 2007 17:04:45 -0400 Date: Fri, 15 Jun 2007 14:04:43 -0700 From: "Paul E. McKenney" To: Dmitry Torokhov Cc: Dipankar Sarma , Peter Zijlstra , LKML Subject: Re: Using RCU with rcu_read_lock()? Message-ID: <20070615210443.GF9301@linux.vnet.ibm.com> Reply-To: paulmck@linux.vnet.ibm.com References: <1181934259.11113.6.camel@lappy> <20070615192940.GA2996@in.ibm.com> <20070615201238.GE9301@linux.vnet.ibm.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: User-Agent: Mutt/1.5.13 (2006-08-11) Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 3609 Lines: 80 On Fri, Jun 15, 2007 at 04:25:02PM -0400, Dmitry Torokhov wrote: > On 6/15/07, Paul E. McKenney wrote: > >On Sat, Jun 16, 2007 at 12:59:40AM +0530, Dipankar Sarma wrote: > >> On Fri, Jun 15, 2007 at 09:04:19PM +0200, Peter Zijlstra wrote: > >> > On Fri, 2007-06-15 at 15:00 -0400, Dmitry Torokhov wrote: > >> > > Hi, > >> > > > >> > > I have a piece of code that is always called under a spinlock with > >> > > interrups disabled. Within that piece of code I iterate through a > >> > > list. I have another piece of code that wants to modify that list. I > >> > > have 2 options: > >> > > > >> > > I don't want to do 1) because the otheir piece of code does not > >really > >> > > care about object owning the spinlock and so acquiring the spinlock > >is > >> > > "not nice". However it is guaranteed that the piece of code that > >> > > accesses lock runs atomically with interrupts disabled. So > >> > > rcu_read_lock() would be superfluos there. > >> > > > >> > > Is it possible to still use list_for_each_rcu() and friends to access > >> > > that list without rcu_read_lock()? Or it is betteruse complete RCU > >> > > interface and eat cost of couple of extra instrctions? > >> > > >> > Yes, preemptible rcu requires that you use the full interface, also, it > >> > more clearly documents the code. Trying to find code that breaks these > >> > assumptions is very tedious work after the fact. > >> > > >> > Please do use the RCU interface in full. > >> > >> As Peter said, you should use the strict RCU APIs and not rely > >> on the current implementation of RCU to optimize. Things change. > >> Plus static/dynamic checking becomes easier that way. > > > >What they said!!! > > > >There are a couple of other options, however: > > > >1. Use preempt_disable() and preempt_enable() on the read side, > > and synchronize_sched() on the update side. > > > >2. Use local_irq_save() and local_irq_restore() on the read side, > > and synchronize_sched() on the update side. Usually not > > competitive -- unless interrupts needed to be disabled for some > > other reason anyway. Which you in fact say that you do. > > Right. The callsite that iterates through the list is essentially > protected by spin_lock_irqsave()/spin_unlock_irqrestore() - needed for > other reasons (such as updating internal state of a device - and that > can happen from different contexts). That will work! > >I believe that #2 might do what you want. But please, PLEASE carefully > >comment this usage!!! > > Would there be a reson not to use #2 but rather full RCU with > rcu_read_lock()/synchronize_rcu()? Probably not, but here are a couple of situations where the full RCU might be preferred: 1. If you were relying on interrupts being disabled within an interrupt handler (which they are -not- in -rt), then you would either need to add some form of local_irq_save() or, as you say, go to the rcu_read_lock() and synchronize_rcu() interfaces. 2. If updates needed to use callbacks rather than synchronous waits for grace periods, in other words, if you needed call_rcu() instead of synchronize_rcu(). Of course, a callback API for _sched (call_rcu_sched() or some such) could be added if needed, though it would be better to avoid the API proliferation unless really badly needed. Thanx, Paul - 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/