Re: [PATCH] mm: mmu_notifier: fix inconsistent memory between secondary MMU and host

[Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>]

On 08/23/2012 12:37 AM, Andrea Arcangeli wrote:
> On Wed, Aug 22, 2012 at 11:51:17AM +0800, Xiao Guangrong wrote:
>> Hmm, in KSM code, i found this code in replace_page:
>>
>> set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot));
>>
>> It is possible to establish a writable pte, no?
> 
> Hugh already answered this thanks. Further details on the vm_page_prot
> are in top of mmap.c, and KSM never scans MAP_SHARED vmas.
> 

Yes, i see that, thank you, Andrea!

>> Unfortunately, all these bugs are triggered by test cases.
> 
> Sure, I've seen the very Oops for the other one, and this one also can
> trigger if unlucky.
> 
> This one can trigger with KVM but only if KSM is enabled or with live
> migration or with device hotplug or some other event that triggers a
> fork in qemu.
> 
> My curiosity about the other one in the exit/unregister/release paths
> is if it really ever triggered with KVM. Because I can't easily see
> how it could trigger. By the time kvm_destroy_vm or exit_mmap() runs,
> no vcpu can be in guest mode anymore, so it cannot matter whatever the
> status of any leftover spte at that time.
> 

vcpu is not in guest mode, but the memory can be still hold in KVM MMU.

Consider this case:

   CPU 0                      CPU 1

   create kvm
   create vcpu thread

   [ Guest is happily running ]

                          send kill signal to the process

   call do_exit
      mmput mm
      exit_mmap, then
         delete mmu_notify

                         reclaim the memory of these threads
                            !!!!!!
                         Now, the page has been reclaimed but
                         it is still hold in KVM MMU

        mmu_notify->release
             !!!!!!
           delete spte, and call
           mark_page_accessed/mark_page_dirty for
           the page which has already been freed on CPU 1


     exit_files
        release kvm/vcpu file, then
        kvm and vcpu are destroyed.