[PATCH v12 4/8] arm64: kvm: Introduce MTE VM feature

Steven Price steven.price at arm.com
Fri May 21 02:28:52 PDT 2021


On 20/05/2021 18:50, Catalin Marinas wrote:
> On Thu, May 20, 2021 at 04:05:46PM +0100, Steven Price wrote:
>> On 20/05/2021 12:54, Catalin Marinas wrote:
>>> On Mon, May 17, 2021 at 01:32:35PM +0100, Steven Price wrote:
>>>> diff --git a/arch/arm64/kvm/mmu.c b/arch/arm64/kvm/mmu.c
>>>> index c5d1f3c87dbd..8660f6a03f51 100644
>>>> --- a/arch/arm64/kvm/mmu.c
>>>> +++ b/arch/arm64/kvm/mmu.c
>>>> @@ -822,6 +822,31 @@ transparent_hugepage_adjust(struct kvm_memory_slot *memslot,
>>>>  	return PAGE_SIZE;
>>>>  }
>>>>  
>>>> +static int sanitise_mte_tags(struct kvm *kvm, unsigned long size,
>>>> +			     kvm_pfn_t pfn)
>>>> +{
>>>> +	if (kvm_has_mte(kvm)) {
>>>> +		/*
>>>> +		 * The page will be mapped in stage 2 as Normal Cacheable, so
>>>> +		 * the VM will be able to see the page's tags and therefore
>>>> +		 * they must be initialised first. If PG_mte_tagged is set,
>>>> +		 * tags have already been initialised.
>>>> +		 */
>>>> +		unsigned long i, nr_pages = size >> PAGE_SHIFT;
>>>> +		struct page *page = pfn_to_online_page(pfn);
>>>> +
>>>> +		if (!page)
>>>> +			return -EFAULT;
>>>
>>> IIRC we ended up with pfn_to_online_page() to reject ZONE_DEVICE pages
>>> that may be mapped into a guest and we have no idea whether they support
>>> MTE. It may be worth adding a comment, otherwise, as Marc said, the page
>>> wouldn't disappear.
>>
>> I'll add a comment.
>>
>>>> +
>>>> +		for (i = 0; i < nr_pages; i++, page++) {
>>>> +			if (!test_and_set_bit(PG_mte_tagged, &page->flags))
>>>> +				mte_clear_page_tags(page_address(page));
>>>
>>> We started the page->flags thread and ended up fixing it for the host
>>> set_pte_at() as per the first patch:
>>>
>>> https://lore.kernel.org/r/c3293d47-a5f2-ea4a-6730-f5cae26d8a7e@arm.com
>>>
>>> Now, can we have a race between the stage 2 kvm_set_spte_gfn() and a
>>> stage 1 set_pte_at()? Only the latter takes a lock. Or between two
>>> kvm_set_spte_gfn() in different VMs? I think in the above thread we
>>> concluded that there's only a problem if the page is shared between
>>> multiple VMMs (MAP_SHARED). How realistic is this and what's the
>>> workaround?
>>>
>>> Either way, I think it's worth adding a comment here on the race on
>>> page->flags as it looks strange that here it's just a test_and_set_bit()
>>> while set_pte_at() uses a spinlock.
>>>
>>
>> Very good point! I should have thought about that. I think splitting the
>> test_and_set_bit() in two (as with the cache flush) is sufficient. While
>> there technically still is a race which could lead to user space tags
>> being clobbered:
>>
>> a) It's very odd for a VMM to be doing an mprotect() after the fact to
>> add PROT_MTE, or to be sharing the memory with another process which
>> sets PROT_MTE.
>>
>> b) The window for the race is incredibly small and the VMM (generally)
>> needs to be robust against the guest changing tags anyway.
>>
>> But I'll add a comment here as well:
>>
>> 	/*
>> 	 * There is a potential race between sanitising the
>> 	 * flags here and user space using mprotect() to add
>> 	 * PROT_MTE to access the tags, however by splitting
>> 	 * the test/set the only risk is user space tags
>> 	 * being overwritten by the mte_clear_page_tags() call.
>> 	 */
> 
> I think (well, I haven't re-checked), an mprotect() in the VMM ends up
> calling set_pte_at_notify() which would call kvm_set_spte_gfn() and that
> will map the page in the guest. So the problem only appears between
> different VMMs sharing the same page. In principle they can be
> MAP_PRIVATE but they'd be CoW so the race wouldn't matter. So it's left
> with MAP_SHARED between multiple VMMs.

mprotect.c only has a call to set_pte_at() not set_pte_at_notify(). And
AFAICT the MMU notifiers are called to invalidate only in
change_pmd_range(). So the stage 2 mappings would be invalidated rather
than populated. However I believe this should cause synchronisation
because of the KVM mmu_lock. So from my reading you are right an
mprotect() can't race.

MAP_SHARED between multiple VMs is then the only potential problem.

> I think we should just state that this is unsafe and they can delete
> each-others tags. If we are really worried, we can export that lock you
> added in mte.c.
> 

I'll just update the comment for now.

Thanks,

Steve



More information about the linux-arm-kernel mailing list