From mboxrd@z Thu Jan 1 00:00:00 1970 From: Xiao Guangrong Subject: Re: [PATCH v3 1/7] KVM: MMU: fix release noslot pfn Date: Mon, 24 Sep 2012 20:32:17 +0800 Message-ID: <506052D1.5000006@linux.vnet.ibm.com> References: <505C0FA8.5070007@linux.vnet.ibm.com> <505C0FCF.2070308@linux.vnet.ibm.com> <20120923091335.GA20907@redhat.com> <505FE8B4.8030309@linux.vnet.ibm.com> <20120924112416.GA23096@redhat.com> <506048D1.70403@linux.vnet.ibm.com> <20120924120424.GB23096@redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 7bit Cc: Avi Kivity , Marcelo Tosatti , LKML , KVM To: Gleb Natapov Return-path: In-Reply-To: <20120924120424.GB23096@redhat.com> Sender: linux-kernel-owner@vger.kernel.org List-Id: kvm.vger.kernel.org On 09/24/2012 08:04 PM, Gleb Natapov wrote: > On Mon, Sep 24, 2012 at 07:49:37PM +0800, Xiao Guangrong wrote: >> On 09/24/2012 07:24 PM, Gleb Natapov wrote: >>> On Mon, Sep 24, 2012 at 12:59:32PM +0800, Xiao Guangrong wrote: >>>> On 09/23/2012 05:13 PM, Gleb Natapov wrote: >>>>> On Fri, Sep 21, 2012 at 02:57:19PM +0800, Xiao Guangrong wrote: >>>>>> We can not directly call kvm_release_pfn_clean to release the pfn >>>>>> since we can meet noslot pfn which is used to cache mmio info into >>>>>> spte >>>>>> >>>>> Wouldn't it be better to move the check into kvm_release_pfn_clean()? >>>> >>>> I think there is no reason for us to prefer to adding this branch in >>>> the common code. :) >>> >>> Is the function performance critical? Is function called without the check >>> on a hot path? The function already contains much heavier kvm_is_mmio_pfn() >>> check. If most/all function invocation require check before call it's >>> better to move it inside. >> >> It is not most/all functions need do this check - it is only needed on x86 mmu >> page-fault/prefetch path. > At least on x86 there 7 calls to kvm_release_pfn_clean(), 5 of them are > guarded by is_noslot_pfn() (after this patch) 3 places after the whole patchset (There are some cleanups after this patch). > and one by even stronger is_error_pfn(). This one is: | if (!is_error_pfn(pfn)) { | kvm_release_pfn_clean(pfn); | return true; | } | | return false; We can change it to: | if (is_error_pfn(pfn)) | return false; | | kvm_release_pfn_clean(pfn); | return true; > I guess when/if other architectures will add MMIO MMU > caching they will need to guard kvm_release_pfn_clean() by is_noslot_pfn() > too in most cases. I am not insisting, but as this patch shows it is > easy to miss the check before calling the function. Sounds reasonable. I will consider it if Avi/Marcelo have no object on it.