Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1755604Ab3I3OlF (ORCPT ); Mon, 30 Sep 2013 10:41:05 -0400 Received: from mx1.redhat.com ([209.132.183.28]:51595 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1755472Ab3I3OlD (ORCPT ); Mon, 30 Sep 2013 10:41:03 -0400 Date: Mon, 30 Sep 2013 10:40:17 -0400 From: Jerome Glisse To: Andrea Arcangeli Cc: Ingo Molnar , Linus Torvalds , Waiman Long , Ingo Molnar , Andrew Morton , Linux Kernel Mailing List , Rik van Riel , Peter Hurley , Davidlohr Bueso , Alex Shi , Tim Chen , Peter Zijlstra , Matthew R Wilcox , Dave Hansen , Michel Lespinasse , Andi Kleen , "Chandramouleeswaran, Aswin" , "Norton, Scott J" , Haggai Eran , Sagi Grimberg , Or Gerlitz Subject: Re: [PATCH] anon_vmas: Convert the rwsem to an rwlock_t Message-ID: <20130930144016.GA2880@redhat.com> References: <1380308424-31011-1-git-send-email-Waiman.Long@hp.com> <20130928074144.GA17773@gmail.com> <20130928192123.GA8228@gmail.com> <20130928193739.GA8642@gmail.com> <20130930085243.GA25685@redhat.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20130930085243.GA25685@redhat.com> 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 Content-Length: 4619 Lines: 98 On Mon, Sep 30, 2013 at 10:52:43AM +0200, Andrea Arcangeli wrote: > Hi everyone, > > On Sat, Sep 28, 2013 at 09:37:39PM +0200, Ingo Molnar wrote: > > > > * Ingo Molnar wrote: > > > > > If we do that then I suspect the next step will be queued rwlocks :-/ > > > The current rwlock_t implementation is rather primitive by modern > > > standards. (We'd probably have killed rwlock_t long ago if not for the > > > tasklist_lock.) > > > > > > But yeah, it would work and conceptually a hard spinlock fits something > > > as lowlevel as the anon-vma lock. > > > > > > I did a quick review pass and it appears nothing obvious is scheduling > > > with the anon-vma lock held. If it did in a non-obvious way it's likely > > > a bug anyway. The hugepage code grew a lot of logic running under the > > > anon-vma lock, but it all seems atomic. > > > > > > So a conversion to rwlock_t could be attempted. (It should be relatively > > > easy patch as well, because the locking operation is now nicely > > > abstracted out.) > > > > Here's a totally untested patch to convert the anon vma lock to an > > rwlock_t. > > Sorry having to break the party but the sleepable locks for anon_vma > and i_mmap_mutex are now requirement for the "pageable RDMA" effort > recently achieved upstream by mellanox with the MMU notifier. > > And as far as I can tell that's the only single good reason for why > those locks shouldn't be spinlocks (otherwise I would have also > complianed at the time of that conversion, the original regression was > known, ask Andi). After the lock conversion it took a while to fix all > other minor bits to make mmu notifier methods fully sleepable. > > The problem with the spinlocks is that in the rmap code (like > try_to_unmap) we need to call mmu_notifier_invalidate_page with an > "mm" as parameter, and the callee assumes the "mm" won't go away under > it. The other second requirement is that the page cannot be freed > until we call the mmu_notifier_invalidate_page (secondary MMU is ok to > still access the page after the linux pte has been dropped and the TLB > flushed). > > In the rmap code the only things that keep things afloat is either the > anon_vma lock or the i_mmap_mutex so it is quite tricky to drop that > lock while keeping "mm" and "page" both afloat for the invalidate > post-anon-vma-unlock. > > Maybe there are ways to makes that safe? (reference counting, > trylocking the mmap_sem). But there isn't a very strightforward way to > do that. > > It isn't just mellanox drivers: originally SGI XPMEM driver also > needed to schedule in those methods (then they figured how to get away > in only scheduling in the mmu_notifier range calls but I suppose they > prefer to be able to schedule in all invalidate methods including > mmu_notifier_invalidate_page). > > Nvidia is now also going to use mmu notifier to allow the GPU (acting > as a secondary MMU with pagetables) to access main memory without > requiring RAM pinning and without disabling all VM features on the > graphics memory (and without GART physical). Probably they don't need > to schedule but I CC'ed Jerome just in case they need (as that would > add one more relevant user of the feature). Thanks for the cc if it was on mm mailing list i would have seen it but i am way behind on my lkml folder. Yes right now we have to sleep in mmu_notifier_invalidate_page but it's somewhat solvable. The mkclean path is solvable more or less easily but the try_to_unmap do have to sleep and that's bad. Especialy the shrinker path. My plan is to provide informations down the mmu notifier API and fails quickly on shrinker path. For shrinking i am not sure if i should just rely on shrinker API to force device or not. It's not something i want to tackle in first patchset anyway. I am hoping to be able to send patchset in next few weeks. But the sleeping inside invalidate page is definitly on my WHATTHEHECKLIST. Cheers, Jerome > > As far as KVM is concerned, there is no benefit in scheduling in the > methods. The KVM mmu notifier invalidate consists in zeroing out one > or more sptes and sending an IPI to flush the guest TLBs if others > vcpus are running in other cpus. It's a mechanism pretty much > identical to the one used by the primary MMU and it only requires irqs > enabled to avoid deadlocks in case of cross-IPIs. -- 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/