Re: [PATCH v1 1/2] arm64/mm: Move PTE_PROT_NONE and PMD_PRESENT_INVALID

[Ryan Roberts]

On 29/04/2024 13:38, Catalin Marinas wrote:
> On Mon, Apr 29, 2024 at 11:04:53AM +0100, Ryan Roberts wrote:
>> On 26/04/2024 15:48, Catalin Marinas wrote:
>>> On Thu, Apr 25, 2024 at 11:37:42AM +0100, Ryan Roberts wrote:
>>>> Also, IMHO we shouldn't really need to reserve PMD_PRESENT_INVALID for swap
>>>> ptes; it would be cleaner to have one bit that defines "present" when valid is
>>>> clear (similar to PTE_PROT_NONE today) then another bit which is only defined
>>>> when "present && !valid" which tells us if this is PTE_PROT_NONE or
>>>> PMD_PRESENT_INVALID (I don't think you can ever have both at the same time?).
>>>
>>> I think this make sense, maybe rename the above to PTE_PRESENT_INVALID
>>> and use it for both ptes and pmds.
>>
>> Yep, sounds good. I've already got a patch to do this, but it's exposed a bug in
>> core-mm so will now fix that before I can validate my change. see
>> https://lore.kernel.org/linux-arm-kernel/ZiuyGXt0XWwRgFh9@x1n/
>>
>> With this in place, I'm proposing to remove PTE_PROT_NONE entirely and instead
>> represent PROT_NONE as a present but invalid pte (PTE_VALID=0, PTE_INVALID=1)
>> with both PTE_WRITE=0 and PTE_RDONLY=0.
>>
>> While the HW would interpret PTE_WRITE=0/PTE_RDONLY=0 as "RW without dirty bit
>> modification", this is not a problem as the pte is invalid, so the HW doesn't
>> interpret it. And SW always uses the PTE_WRITE bit to interpret the writability
>> of the pte. So PTE_WRITE=0/PTE_RDONLY=0 was previously an unused combination
>> that we now repurpose for PROT_NONE.
> 
> Why not just keep the bits currently in PAGE_NONE (PTE_RDONLY would be
> set) and check PTE_USER|PTE_UXN == 0b01 which is a unique combination
> for PAGE_NONE (bar the kernel mappings).

Yes I guess that works. I personally prefer my proposal because it is more
intuitive; you have an R bit and a W bit, and you encode RO, WR, and NONE. But
if you think reusing the kernel mapping check (PTE_USER|PTE_UXN == 0b01) is
preferable, then I'll go with that.

> 
> For ptes, it doesn't matter, we can assume that PTE_PRESENT_INVALID
> means pte_protnone(). For pmds, however, we can end up with
> pmd_protnone(pmd_mkinvalid(pmd)) == true for any of the PAGE_*
> permissions encoded into a valid pmd. That's where a dedicated
> PTE_PROT_NONE bit helped.

Yes agreed.

> 
> Let's say a CPU starts splitting a pmd and does a pmdp_invalidate*()
> first to set PTE_PRESENT_INVALID. A different CPU gets a fault and since
> the pmd is present, it goes and checks pmd_protnone() which returns
> true, ending up on do_huge_pmd_numa_page() path. Maybe some locks help
> but it looks fragile to rely on them.
> 
> So I think for protnone we need to check some other bits (like USER and
> UXN) in addition to PTE_PRESENT_INVALID.

Yes 100% agree. But using PTE_WRITE|PTE_RDONLY==0b00 is just as valid for that
purpose, I think?

> 
>> This will subtly change behaviour in an edge case though. Imagine:
>>
>> pte_t pte;
>>
>> pte = pte_modify(pte, PAGE_NONE);
>> pte = pte_mkwrite_novma(pte);
>> WARN_ON(pte_protnone(pte));
>>
>> Should that warning fire or not? Previously, because we had a dedicated bit for
>> PTE_PROT_NONE it would fire. With my proposed change it will not fire. To me
>> it's more intuitive if it doesn't fire. Regardless there is no core code that
>> ever does this. Once you have a protnone pte, its terminal - nothing ever
>> modifies it with these helpers AFAICS.
> 
> I don't think any core code should try to make page a PAGE_NONE pte
> writeable.

I looked at some other arches; some (at least alpha and hexagon) will not fire
this warning because they have R and W bits and 0b00 means NONE. Others (x86)
will fire it because they have an explicit NONE bit and don't remove it on
permission change. So I conclude its UB and fine to do either.

> 
>> Personally I think this is a nice tidy up that saves a SW bit in both present
>> and swap ptes. What do you think? (I'll just post the series if its easier to
>> provide feedback in that context).
> 
> It would be nice to tidy this up and get rid of PTE_PROT_NONE as long as
> it doesn't affect the pmd case I mentioned above.
> 
>>>> But there is a problem with this: __split_huge_pmd_locked() calls
>>>> pmdp_invalidate() for a pmd before it determines that it is pmd_present(). So
>>>> the PMD_PRESENT_INVALID can be set in a swap pte today. That feels wrong to me,
>>>> but was trying to avoid the whole thing unravelling so didn't persue.
>>>
>>> Maybe what's wrong is the arm64 implementation setting this bit on a
>>> swap/migration pmd (though we could handle this in the core code as
>>> well, it depends what the other architectures do). The only check for
>>> the PMD_PRESENT_INVALID bit is in the arm64 code and it can be absorbed
>>> into the pmd_present() check. I think it is currently broken as
>>> pmd_present() can return true for a swap pmd after pmd_mkinvalid().
>>
>> I've posted a fix here:
>> https://lore.kernel.org/linux-mm/20240425170704.3379492-1-ryan.roberts@xxxxxxx/
>>
>> My position is that you shouldn't be calling pmd_mkinvalid() on a non-present pmd.
> 
> I agree, thanks.
>