Re: [kernel-hardening] Re: [RFC v2][PATCH 04/11] x86: Implement __arch_rare_write_begin/unmap()

[Andy Lutomirski]

On Sun, Apr 9, 2017 at 5:47 AM, PaX Team <pageexec@xxxxxxxxxxx> wrote:
> On 7 Apr 2017 at 21:58, Andy Lutomirski wrote:
>
>> On Fri, Apr 7, 2017 at 12:58 PM, PaX Team <pageexec@xxxxxxxxxxx> wrote:
>> > On 7 Apr 2017 at 9:14, Andy Lutomirski wrote:
>> >> Then someone who cares about performance can benchmark the CR0.WP
>> >> approach against it and try to argue that it's a good idea.  This
>> >> benchmark should wait until I'm done with my PCID work, because PCID
>> >> is going to make use_mm() a whole heck of a lot faster.
>> >
>> > in my measurements switching PCID is hovers around 230 cycles for snb-ivb
>> > and 200-220 for hsw-skl whereas cr0 writes are around 230-240 cycles. there's
>> > of course a whole lot more impact for switching address spaces so it'll never
>> > be fast enough to beat cr0.wp.
>> >
>>
>> If I'm reading this right, you're saying that a non-flushing CR3 write
>> is about the same cost as a CR0.WP write.  If so, then why should CR0
>> be preferred over the (arch-neutral) CR3 approach?
>
> cr3 (page table switching) isn't arch neutral at all ;). you probably meant
> the higher level primitives except they're not enough to implement the scheme
> as discussed before since the enter/exit paths are very much arch dependent.

Yes.

>
> on x86 the cost of the pax_open/close_kernel primitives comes from the cr0
> writes and nothing else, use_mm suffers not only from the cr3 writes but
> also locking/atomic ops and cr4 writes on its path and the inevitable TLB
> entry costs. and if cpu vendors cared enough, they could make toggling cr0.wp
> a fast path in the microcode and reduce its overhead by an order of magnitude.
>

If the CR4 writes happen in for this use case, that's a bug.

>>  And why would switching address spaces obviously be much slower?
>> There'll be a very small number of TLB fills needed for the actual
>> protected access.
>
> you'll be duplicating TLB entries in the alternative PCID for both code
> and data, where they will accumulate (=take room away from the normal PCID
> and expose unwanted memory for access) unless you also flush them when
> switching back (which then will cost even more cycles). also i'm not sure
> that processors implement all the 12 PCID bits so depending on how many PCIDs
> you plan to use, you could be causing even more unnecessary TLB replacements.
>

Unless the CPU is rather dumber than I expect, the only duplicated
entries should be for the writable aliases of pages that are written.
The rest of the pages are global and should be shared for all PCIDs.

--Andy