From mboxrd@z Thu Jan 1 00:00:00 1970 From: Xiao Guangrong Subject: Re: [PATCH v3 11/11] KVM: MMU: improve write flooding detected Date: Wed, 14 Sep 2011 03:19:25 +0800 Message-ID: <4E6FACBD.1020801@cn.fujitsu.com> References: <4E5C4C20.3000403@cn.fujitsu.com> <4E5C4D17.2080700@cn.fujitsu.com> <4E6F395A.5030609@redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 7bit Cc: Marcelo Tosatti , LKML , KVM To: Avi Kivity Return-path: In-Reply-To: <4E6F395A.5030609@redhat.com> Sender: linux-kernel-owner@vger.kernel.org List-Id: kvm.vger.kernel.org On 09/13/2011 07:07 PM, Avi Kivity wrote: > On 08/30/2011 05:38 AM, Xiao Guangrong wrote: >> Detecting write-flooding does not work well, when we handle page written, if >> the last speculative spte is not accessed, we treat the page is >> write-flooding, however, we can speculative spte on many path, such as pte >> prefetch, page synced, that means the last speculative spte may be not point >> to the written page and the written page can be accessed via other sptes, so >> depends on the Accessed bit of the last speculative spte is not enough >> >> Instead of detected page accessed, we can detect whether the spte is accessed >> after it is written, if the spte is not accessed but it is written frequently, >> we treat is not a page table or it not used for a long time >> >> > > The spte may not be accessed, but other sptes in the same page can be accessed. An example is the fixmap area for kmap_atomic(), there will be a lot of pte writes but other sptes will be accessed without going through soft-mmu at all. I think this kind of shadow pae is mostly the last page table(level=1), maybe we can skip the write-flooding for the last shadow page, because the last shadow page can become unsync and it can not let page table write-protected. > I think you have to read the parent_ptes->spte.accessed bits to be sure. > I guess the overload of this way is little high: - it needs to walk parent ptes for every shadow pages - we need to clear the parent_ptes->spte.accessed bit when the page is written, and the tlb flush is needed. no?