Re: [edk2-devel] [PATCH] UefiCpuPkg: Fix issue that IsModified is wrongly set in PageTableMap

["Laszlo Ersek" <lersek@redhat.com>]

On 1/15/24 10:43, Pedro Falcato wrote:
> On Thu, Jan 11, 2024 at 8:56 AM Laszlo Ersek <lersek@redhat.com> wrote:
>>
>> On 1/11/24 03:03, Ni, Ray wrote:
>>>> This function is incredibly complicated, so reviewing this patch is
>>>> hard, even after reading the bugzilla ticket.
>>>>
>>>> The commit message is useless. It should contain a brief description of
>>>> the problem, and how the fix resolves the problem.
>>>>
>>>> The documentation of the PageTableLibMapInLevel() function is wrong,
>>>> even before this patch. It documents the "IsModified" output-only
>>>> parameter as follows:
>>>>
>>>> "TRUE means page table is modified. FALSE means page table is not
>>>> modified."
>>>>
>>>> This states that "IsModified" is always set on output, to either FALSE
>>>> or TRUE. Which is an incorrect statement; in reality the caller is
>>>> expected to pre-set (*IsModified) to FALSE, and PageTableLibMapInLevel()
>>>> will (conditionally!) perform a FALSE->TRUE transition only.
>>>>
>>>> Now, this patch may fix a bug, but it makes the above-described
>>>> documentation issue worse, by further restricting the condition for said
>>>> FALSE->TRUE transition.
>>>
>>> Laszlo, thanks for the comments!
>>> Though the fixing looks simple, Zhiguang and I did have several rounds of offline discussions
>>> regarding how to fix it.
>>>
>>> When the lib accesses the page table content, CPU would set the "Access" bit in the page entry
>>> that points to the page table memory being accessed by the lib.
>>>
>>> So, even when the "Modify" is FALSE (indicating caller doesn't want the lib to modify the page table),
>>> lib code should not modify the page table but CPU still sets the "Access" bit in some of the entries due to
>>> the reasons above.
>>
>> Huh, tricky!
>>
>> Should the comparison explicitly mask out the Accessed bit from each of
>> the "before" page table entry and the "after" one, perhaps?
> 
> FWIW, clearing the A and D bits off of PTEs requires a TLB flush and,
> as such, that change would break them.

I didn't mean to clear the A/D bits inside the actual live PTEs, only in
those temporary / helper variables (or even just expressions) that we
use for comparing the before/after states.

> 
> In general:
>  - You need a TLB flush when unmapping a page
>  - You need a TLB flush when changing an already-mapped PTE (unless
> you tolerate a stale TLB and want to eat a spurious page fault, which
> is a valid technique)
>  - You don't need a TLB flush when freshly mapping a page (unmapped ->
> mapped) as x86 doesn't cache non-present PTEs
> 
> so you shouldn't need to inspect the PTE before and after;

this seems to invite further discussion wrt. what the function is
supposed to do at all...

> in fact,
> that's erroneous as Intel CPUs can speculatively set the A and D bits
> (they're slightly more careful since CET rolled around, but as far as
> I've heard older Intel used to wildly set those bits speculatively)
> and AMD ones can too (although they cannot speculatively set D).

What is the error behavior here? Assuming we consider a speculative A/D
setting by the hardware, the worst that can happen is that we spuriously
flush the TLB. Is that right? Doesn't seem extremely harmful.

> 
> I'd love to give out more feedback on this patch, but I *really* don't
> understand what any of that function is doing :/
> 

Yup.

(In general I'm just acknowledging that right now this is quite out of
my league...)

Thanks,
Laszlo



-=-=-=-=-=-=-=-=-=-=-=-
Groups.io Links: You receive all messages sent to this group.
View/Reply Online (#113839): https://edk2.groups.io/g/devel/message/113839
Mute This Topic: https://groups.io/mt/103636407/7686176
Group Owner: devel+owner@edk2.groups.io
Unsubscribe: https://edk2.groups.io/g/devel/leave/12367111/7686176/1913456212/xyzzy [rebecca@openfw.io]
-=-=-=-=-=-=-=-=-=-=-=-