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

[Shan Hai <haishan.bai@gmail.com>]

On 07/15/2011 11:24 AM, Peter Zijlstra wrote:
> On Fri, 2011-07-15 at 11:18 -0400, Shan Hai wrote:
>
>>>> +	vma = find_vma(mm, address);
>>> Uhm, find_vma() needs mmap_sem, and futex_atomic_cmpxchg_inatomic() is
>>> most certainly not called with that lock held.
>>>
>> My fault, that will be fixed in the V2 patch.
> But you cannot, the function isn't called _atomic_ just for kicks, its
> used while holding spinlocks.
>

Yes we can do that, _atomic_ here is just atomic for cmpxchg
implemented by the combination of 'lwarx' and 'stwcx.' instructions
as done in the spin lock implementation, so even we hold the
mmap_sem that has no impact on the _atomic_ feature of the
futex_atomic_cmpxchg_inatomic().

>>>> +	if (likely(vma)) {
>>>> +		/* only fixup present page */
>>>> +		if (follow_page(vma, address, FOLL_WRITE)) {
>>>> +			handle_mm_fault(mm, vma, address, FAULT_FLAG_WRITE);
>>> So how can this toggle your sw dirty/young tracking, that's pretty much
>>> what gup(.write=1) does too!
>>>
>> because of the kernel read only permission of the page is transparent
>> to the follow_page(),  the handle_mm_fault() is not to be activated
>> in the __get_use_pages(), so the gup(.write=1) could not help to fixup
>> the write permission.
> So why do you need the vma? Is it like I wrote earlier that you don't
> have spare PTE bits and need the vma flags to see if it may become
> writable?
>

Need vma for the reason to call handle_mm_fault(), that's all.

> gup(.write=1) not triggering this is a serious problem though, not
> something you can just paper over. I wouldn't be at all surprised to
> find there's more things broken because of that.

In my opinion another solution might be check the read only for kernel
feature of a page in the follow_page() on gup(.write=1) to avoid this
problem on all architectures.

Thanks
Shan Hai