[PATCH v9 4/4] arm: ARMv7 dirty page logging 2nd stage page fault handling support

Mario Smarduch m.smarduch at samsung.com
Wed Aug 13 18:20:19 PDT 2014


On 08/13/2014 12:30 AM, Christoffer Dall wrote:
> On Tue, Aug 12, 2014 at 06:27:11PM -0700, Mario Smarduch wrote:
>> On 08/12/2014 02:50 AM, Christoffer Dall wrote:
>>> On Mon, Aug 11, 2014 at 06:25:05PM -0700, Mario Smarduch wrote:
>>>> On 08/11/2014 12:13 PM, Christoffer Dall wrote:
>>>>> On Thu, Jul 24, 2014 at 05:56:08PM -0700, Mario Smarduch wrote:
> 
> [...]
> 
>>>>>> @@ -1151,7 +1170,7 @@ static void kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data)
>>>>>>  {
>>>>>>  	pte_t *pte = (pte_t *)data;
>>>>>>  
>>>>>> -	stage2_set_pte(kvm, NULL, gpa, pte, false);
>>>>>> +	stage2_set_pte(kvm, NULL, gpa, pte, false, false);
>>>>>
>>>>> why is logging never active if we are called from MMU notifiers?
>>>>
>>>> mmu notifiers update sptes, but I don't see how these updates
>>>> can result in guest dirty pages. Also guest pages are marked dirty
>>>> from 2nd stage page fault handlers (searching through the code).
>>>>
>>> Ok, then add:
>>>
>>> /*
>>>  * We can always call stage2_set_pte with logging_active == false,
>>>  * because MMU notifiers will have unmapped a huge PMD before calling
>>>  * ->change_pte() (which in turn calls kvm_set_spte_hva()) and therefore
>>>  * stage2_set_pte() never needs to clear out a huge PMD through this
>>>  * calling path.
>>>  */
>>
>> So here on permission change to primary pte's kernel first invalidates
>> related s2ptes followed by ->change_pte calls to synchronize s2ptes. As
>> consequence of invalidation we unmap huge PMDs, if a page falls in that
>> range.
>>
>> Is the comment to point out use of logging flag under various scenarios?
> 
> The comment is because when you look at this function it is not obvious
> why we pass logging_active=false, despite logging may actually be
> active.  This could suggest that the parameter to stage2_set_pte()
> should be named differently (break_huge_pmds) or something like that,
> but we can also be satisfied with the comment.

Ok I see, I was thinking you thought it was breaking something.
Yeah I'll add the comment, in reality this is another use case
where a PMD may need to be converted to page table so it makes sense
to contrast use cases.

> 
>>
>> Should I add comments on flag use in other places as well?
>>
> 
> It's always a judgement call.  I didn't find it necessarry to put a
> comment elsewhere because I think it's pretty obivous that we would
> never care about logging writes to device regions.
> 
> However, this made me think, are we making sure that we are not marking
> device mappings as read-only in the wp_range functions?  I think it's

KVM_SET_USER_MEMORY_REGION ioctl doesn't check type of region being
installed/operated on (KVM_MEM_LOG_DIRTY_PAGES), in case of QEMU
these regions wind up in KVMState->KVMSlot[], when
memory_region_add_subregion() is called KVM listener installs it.
For migration and dirty page logging QEMU walks the KVMSlot[] array.

For QEMU VFIO (PCI) mmap()ing BAR of type IORESOURCE_MEM,
causes the memory region to be added to KVMState->KVMSlot[].
In that case it's possible to walk KVMState->KVMSlot[] issue
the ioctl and  come across  a device mapping with normal memory and
WP it's s2ptes (VFIO sets unmigrateble state though).

But I'm not sure what's there to stop someone calling the ioctl and
install a region with device memory type. Most likely though if you
installed that kind of region migration would be disabled.

But just for logging use not checking memory type could be an issue.

> quite bad if we mark the VCPU interface as read-only for example.
> 
> -Christoffer
> 




More information about the linux-arm-kernel mailing list