Re: [Patch 2/4] KVM:SVM: Introduce set_spte_notify support

[Eric van Tassell <evantass@amd.com>]

On 8/20/20 6:59 PM, Sean Christopherson wrote:
> On Thu, Aug 20, 2020 at 12:05:00PM -0500, Eric van Tassell wrote:
>>
>>
>> On 8/19/20 11:05 AM, Sean Christopherson wrote:
>>> On Wed, Aug 19, 2020 at 11:03:48AM -0500, Eric van Tassell wrote:
>>>>
>>>>
>>>> On 8/3/20 11:27 AM, Sean Christopherson wrote:
>>>>> On Sun, Aug 02, 2020 at 03:53:54PM -0500, Eric van Tassell wrote:
>>>>>>
>>>>>> On 7/31/20 3:25 PM, Sean Christopherson wrote:
>>>>>>> On Fri, Jul 24, 2020 at 06:54:46PM -0500, eric van tassell wrote:
>>>>>>>> Improve SEV guest startup time from O(n) to a constant by deferring
>>>>>>>> guest page pinning until the pages are used to satisfy nested page faults.
>>>>>>>>
>>>>>>>> Implement the code to do the pinning (sev_get_page) and the notifier
>>>>>>>> sev_set_spte_notify().
>>>>>>>>
>>>>>>>> Track the pinned pages with xarray so they can be released during guest
>>>>>>>> termination.
>>>>>>>
>>>>>>> I like that SEV is trying to be a better citizen, but this is trading one
>>>>>>> hack for another.
>>>>>>>
>>>>>>>     - KVM goes through a lot of effort to ensure page faults don't need to
>>>>>>>       allocate memory, and this throws all that effort out the window.
>>>>>>>
>>>>>> can you elaborate on that?
>>>>>
>>>>> mmu_topup_memory_caches() is called from the page fault handlers before
>>>>> acquiring mmu_lock to pre-allocate shadow pages, PTE list descriptors, GFN
>>>>> arrays, etc... that may be needed to handle the page fault.  This allows
>>>>> using standard GFP flags for the allocation and obviates the need for error
>>>>> handling in the consumers.
>>>>>
>>>>
>>>> I see what you meant. The issue that causes us to use this approach is that
>>>> we need to be able to unpin the pages when the VM exits.
>>>
>>> Yes, but using a software available flag in the SPTE to track pinned pages
>>> should be very doable.
>>>
>>
>> The issue, as I understand it, is that when spte(s) get zapped/unzapped, the
>> flags are lost so we'd have to have some mechanism to, before zapping, cache
>> the pfn <-> spte mapping
> 
> The issue is that code doesn't exist :-)

let me look into that and discuss it in our team meeting

> 
> The idea is to leave the pfn in the spte itself when a pinned spte is zapped,
> and use software available bits in the spte to indicate the page is pinned
> and has zap.  When the VM is destroyed, remove all sptes and drop the page
> reference for pinned pages.
>