From mboxrd@z Thu Jan 1 00:00:00 1970 From: Avi Kivity Subject: Re: [PATCH 2/4] KVM: Avoid checking huge page mappings in get_dirty_log() Date: Tue, 13 Mar 2012 12:12:33 +0200 Message-ID: <4F5F1D91.6030706@redhat.com> References: <20120301193007.04b2db8e.yoshikawa.takuya@oss.ntt.co.jp> <20120301193216.b14538bb.yoshikawa.takuya@oss.ntt.co.jp> <4F5E3AAB.7020603@redhat.com> <20120313182045.97f42399.yoshikawa.takuya@oss.ntt.co.jp> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 7bit Cc: mtosatti@redhat.com, kvm@vger.kernel.org, linux-kernel@vger.kernel.org To: Takuya Yoshikawa Return-path: In-Reply-To: <20120313182045.97f42399.yoshikawa.takuya@oss.ntt.co.jp> Sender: linux-kernel-owner@vger.kernel.org List-Id: kvm.vger.kernel.org On 03/13/2012 11:20 AM, Takuya Yoshikawa wrote: > Avi Kivity wrote: > > > It occurs to me that we should write-protect huge page tables, since it > > makes write protection much faster (we make up for this later at write > > fault time, but that might not occur, and even if it does we reduce > > guest jitter). In fact I once proposed a more involved variant, called > > Do you mean protecting huge page tables when we start logging and dropping > them, one by one, at the time of write fault? Yes. Even more generally, protecting PML4Es, then at fault time un-protecting the faulting PML4E and write-protecting all underlying PDPEs except the one for the faulting address, and similarly for PDEs and PTEs. > If so, we do not need to change get_dirty_log() implementation. > Changing kvm_mmu_remove_write_access() and fault handler seems to be enough. > > > O(1) write protection, in which we write-protect the topmost page table > > only and only un-write-protect the paths that fault. > > > That can be done later however and shouldn't affect this patchset. > > I have some additional patches to optimize rmap handling which seems to > improve get_dirty_log() about 10% in the worst case. Great, looking forward. > After that, I want to take some time to prepare for further developments > because my current test tools are not enough now. > -- error compiling committee.c: too many arguments to function