[PATCH 0/1] Fixup write permission of TLB on powerpc e500 core

[Shan Hai]

On 07/15/2011 08:20 PM, Benjamin Herrenschmidt wrote:
> On Fri, 2011-07-15 at 11:32 -0400, Shan Hai wrote:
>> I agree with you, the problem could be triggered by accessing
>> any user space page which has kernel read only permission
>> in the page fault disabled context, the problem also affects
>> architectures which depend on SW dirty/young tracking as
>> stated by Benjamin in this thread.
>>
>> In the e500 case, the commit 6cfd8990e27d3a491c1c605d6cbc18a46ae51fef
>> removed the write permission fixup from TLB miss handlers and left it to
>> generic code, so it might be right time to fixup the write permission here
>> in the generic code.
> But we can't. The must not modify the PTE from an interrupt context and
> the "atomic" variants of user accesses can be called in such contexts.
>
> I think the problem is that we try to actually do things other than just
> "peek" at user memory (for backtraces etc...) but actually useful things
> in page fault disabled contexts. That's bad and various archs mm were
> designed with the assumption that this never happens.
>

Yes I understood, the *here* above means 'generic code' like futex code,
I am sorry for my ambiguous description.

> If the futex case is seldom here, we could probably find a way to work
> around in that specific case.
>

That's what my patch wants to do.

> However, I -still- don't understand why gup didn't fixup the write
> permission. gup doesn't set dirty ?
>

Yep, gup doesn't set dirty, because when the page fault
occurs on the kernel accessing a user page which is
read only to the kernel the following conditions hold,
- the page is present, because its a shared page
- the page is writable, because demand paging
     sets up the pte for the current  process to so

The follow_page() called in the __get_user_page()
returns non NULL to its caller on the above mentioned
present and writable page, so the gup(.write=1) has no
chance to set pte dirty by calling handle_mm_fault

Thanks
Shan Hai
s
> Cheers,
> Ben.
>
>