From mboxrd@z Thu Jan 1 00:00:00 1970 From: Avi Kivity Subject: Re: [PATCH v5 2/9] KVM: MMU: fix race between 'walk_addr' and 'fetch' Date: Sun, 11 Jul 2010 15:52:32 +0300 Message-ID: <4C39BE90.1030802@redhat.com> References: <4C330918.6040709@cn.fujitsu.com> <4C330948.1070305@cn.fujitsu.com> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8; format=flowed Content-Transfer-Encoding: 7bit Cc: Marcelo Tosatti , LKML , KVM list To: Xiao Guangrong Return-path: In-Reply-To: <4C330948.1070305@cn.fujitsu.com> Sender: linux-kernel-owner@vger.kernel.org List-Id: kvm.vger.kernel.org On 07/06/2010 01:45 PM, Xiao Guangrong wrote: > 'walk_addr' is out of mmu_lock's protection, so while we handle 'fetch', > then guest's mapping has modifited by other vcpu's write path, such as > invlpg, pte_write and other fetch path > > Fixed by checking all level's mapping > > > @@ -319,22 +319,23 @@ static u64 *FNAME(fetch)(struct kvm_vcpu *vcpu, gva_t addr, > direct_access&= ~ACC_WRITE_MASK; > > for_each_shadow_entry(vcpu, addr, iterator) { > + bool nonpresent = false, last_mapping = false; > + > I don't like these two new variables, but no suggestion at the moment. I'll try to simplify this loop later. One idea may be: while (level > walker.level) { handle indirect pages } while (level > hlevel) { handle direct pages } handle last spte I'm worried that this change is too big for backporting, but no suggestions on how to make it smaller, so we'll have to accept it. > - spte = __pa(sp->spt) > - | PT_PRESENT_MASK | PT_ACCESSED_MASK > - | PT_WRITABLE_MASK | PT_USER_MASK; > - *sptep = spte; > + if (nonpresent) { > + spte = __pa(sp->spt) > + | PT_PRESENT_MASK | PT_ACCESSED_MASK > + | PT_WRITABLE_MASK | PT_USER_MASK; > + *sptep = spte; > + continue; > + } > Should be __set_spte(), but this already exists, so can be in a later patch. -- error compiling committee.c: too many arguments to function