From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ingo Molnar Subject: Re: [PATCH v8 4/6] MCS Lock: Move mcs_lock/unlock function into its own Date: Tue, 21 Jan 2014 20:06:58 +0100 Message-ID: <20140121190658.GA5862@gmail.com> References: <1390267471.3138.38.camel@schen9-DESK> <20140121101915.GS31570@twins.programming.kicks-ass.net> <20140121104140.GA4092@gmail.com> <1390330623.3138.56.camel@schen9-DESK> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Content-Disposition: inline In-Reply-To: <1390330623.3138.56.camel@schen9-DESK> Sender: owner-linux-mm@kvack.org To: Tim Chen Cc: Peter Zijlstra , Ingo Molnar , Andrew Morton , Thomas Gleixner , "Paul E.McKenney" , Will Deacon , linux-kernel@vger.kernel.org, linux-mm , linux-arch@vger.kernel.org, Linus Torvalds , Waiman Long , Andrea Arcangeli , Alex Shi , Andi Kleen , Michel Lespinasse , Davidlohr Bueso , Matthew R Wilcox , Dave Hansen , Rik van Riel , Peter Hurley , Raghavendra K T , George Spelvin , "H. Peter Anvin" , Arnd List-Id: linux-arch.vger.kernel.org * Tim Chen wrote: > On Tue, 2014-01-21 at 11:41 +0100, Ingo Molnar wrote: > > * Peter Zijlstra wrote: > > > > > On Mon, Jan 20, 2014 at 05:24:31PM -0800, Tim Chen wrote: > > > > +EXPORT_SYMBOL_GPL(mcs_spin_lock); > > > > +EXPORT_SYMBOL_GPL(mcs_spin_unlock); > > > > > > Do we really need the EXPORTs? The only user so far is mutex and that's > > > core code. The other planned users are rwsems and rwlocks, for both it > > > would be in the slow path, which is also core code. > > > > > > We should generally only add EXPORTs once theres a need. > > > > In fact I'd argue the hot path needs to be inlined. > > > > We only don't inline regular locking primitives because it would blow > > up the kernel's size in too many critical places. > > > > But inlining an _internal_ locking implementation used in just a > > handful of places is a no-brainer... > > The original mspin_lock primitive from which mcs_spin_lock was > derived has an explicit noinline annotation. The comment says that > it is so that perf can properly account for time spent in the lock > function. So it wasn't inlined in previous kernels when we started. Not sure what comment that was, but it's not a valid argument: profiling and measurement is in almost all cases secondary to any performance considerations! If we keep it out of line then we want to do it only if it's faster that way. > For the time being, I'll just remove the EXPORT. If people feel > that inline is the right way to go, then we'll leave the function in > mcs_spin_lock.h and not create mcs_spin_lock.c. Well, 'people' could be you, the person touching the code? This is really something that is discoverable: look at the critical path in the inlined and the out of line case, and compare the number of instructions. This can be done based on disassembly of the affected code. Thanks, Ingo -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@kvack.org. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: email@kvack.org