Re: [PATCH v2] sparc64: Add support for Application Data Integrity (ADI)

[David Miller]

From: Khalid Aziz <khalid.aziz@xxxxxxxxxx>
Date: Mon, 7 Mar 2016 14:33:56 -0700

> On 03/07/2016 12:16 PM, David Miller wrote:
>> From: Khalid Aziz <khalid.aziz@xxxxxxxxxx>
>> Date: Mon, 7 Mar 2016 11:24:54 -0700
>>
>>> Tags can be cleared by user by setting tag to 0. Tags are
>>> automatically cleared by the hardware when the mapping for a virtual
>>> address is removed from TSB (which is why swappable pages are a
>>> problem), so kernel does not have to do it as part of clean up.
>>
>> You might be able to crib some bits for the Tag in the swp_entry_t,
>> it's
>> 64-bit and you can therefore steal bits from the offset field.
>>
>> That way you'll have the ADI tag in the page tables, ready to
>> re-install
>> at swapin time.
>>
> 
> That is a possibility but limited in scope. An address range covered
> by a single TTE can have large number of tags. Version tags are set on
> cacheline. In extreme case, one could set a tag for each set of
> 64-bytes in a page. Also tags are set completely in userspace and no
> transition occurs to kernel space, so kernel has no idea of what tags
> have been set. I have not found a way to query the MMU on tags.
> 
> I will think some more about it.

That would mean that ADI is impossible to use for swappable memory.

...

If that's true I'm extremely disappointed that they devoted so much
silicon and engineering to this feature yet didn't take that one
critical step to make it generally useful. :(

We could have a way to do this via the kernel, wherein the user has a
contract with us.  Basically we have a call to pass the Tags (what
granularity to use for this is a design point, pages, cache lines,
etc.)  into the kernel and the user agrees not to change them behind
the kernel's back.

In return the kernel agrees to restore the tags upon swapin.

So we could support something for swappable pages, it would just be
more work.