Re: [RFC 1/2] arm64/mm: Change THP helpers to comply with generic MM semantics

[Anshuman Khandual]

On 07/03/2019 11:22 PM, Catalin Marinas wrote:
> On Tue, Jul 02, 2019 at 09:07:28AM +0530, Anshuman Khandual wrote:
>> On 06/28/2019 03:50 PM, Catalin Marinas wrote:
>>> On Thu, Jun 27, 2019 at 06:18:15PM +0530, Anshuman Khandual wrote:
>>>> pmd_present() and pmd_trans_huge() are expected to behave in the following
>>>> manner during various phases of a given PMD. It is derived from a previous
>>>> detailed discussion on this topic [1] and present THP documentation [2].
>>>>
>>>> pmd_present(pmd):
>>>>
>>>> - Returns true if pmd refers to system RAM with a valid pmd_page(pmd)
>>>> - Returns false if pmd does not refer to system RAM - Invalid pmd_page(pmd)
>>>>
>>>> pmd_trans_huge(pmd):
>>>>
>>>> - Returns true if pmd refers to system RAM and is a trans huge mapping
> [...]
>>> Before we actually start fixing this, I would strongly suggest that you
>>> add a boot selftest (see lib/Kconfig.debug for other similar cases)
>>> which checks the consistency of the page table macros w.r.t. the
>>> expected mm semantics. Once the mm maintainers agreed with the
>>> semantics, it will really help architecture maintainers in implementing
>>> them correctly.
>>
>> Sure and it will help all architectures to be in sync wrt semantics.
>>
>>> You wouldn't need actual page tables, just things like assertions on
>>> pmd_trans_huge(pmd_mkhuge(pmd)) == true. You could go further and have
>>> checks on pmdp_invalidate(&dummy_vma, dummy_addr, &dummy_pmd) with the
>>> dummy_* variables on the stack.
>>
>> Hmm. I guess macros which operate directly on a page table entry will be
>> okay but the ones which check on specific states for VMA or MM might be
>> bit tricky. Try to emulate VMA/MM states while on stack ?. But sure, will
>> explore adding such a test.
> 
> You can pretend that the page table is on the stack. See the _pmd
> variable in do_huge_pmd_wp_page_fallback() and
> __split_huge_zero_page_pmd(). Similarly, the vma and even the mm can be
> faked on the stack (see the arm64 tlb_flush()).

Sure will explore them and other similar examples. I am already working on a module
which will test various architecture page table accessors semantics as expected from
generic MM. This should help us making sure that all architectures are on same page.

> 
>>>> The problem:
>>>>
>>>> PMD is first invalidated with pmdp_invalidate() before it's splitting. This
>>>> invalidation clears PMD_SECT_VALID as below.
>>>>
>>>> PMD Split -> pmdp_invalidate() -> pmd_mknotpresent -> Clears PMD_SECT_VALID
>>>>
>>>> Once PMD_SECT_VALID gets cleared, it results in pmd_present() return false
>>>> on the PMD entry.
>>>
>>> I think that's an inconsistency in the expected semantics here. Do you
>>> mean that pmd_present(pmd_mknotpresent(pmd)) should be true? If not, do
> [...]
>> pmd_present() and pmd_mknotpresent() are not exact inverse.
> 
> I find this very confusing (not your fault, just the semantics expected
> by the core code). I can see that x86 is using _PAGE_PSE to make
> pmd_present(pmd_mknotpresent()) == true. However, for pud that's not the
> case (because it's not used for transhuge).
> 
> I'd rather have this renamed to pmd_mknotvalid().

Right, it makes sense to do the renaming even without considering this proposal.

> 
>> In absence of a positive section mapping bit on arm64, PTE_SPECIAL is being set
>> temporarily to remember that it was a mapped PMD which got invalidated recently
>> but which still points to memory. Hence pmd_present() must evaluate true.
> 
> I wonder if we can encode this safely for arm64 in the bottom two bits
> of a pmd :
> 
> 0b00 - not valid, not present
> 0b10 - not valid, present, huge
> 0b01 - valid, present, huge
> 0b11 - valid, table (not huge)
> 
> Do we ever call pmdp_invalidate() on a table entry? I don't think we do.
> 
> So a pte_mknotvalid would set bit 1 and I think swp_entry_to_pmd() would
> have to clear it so that pmd_present() actually returns false for a swp
> pmd entry.

All these makes it riskier for collision with other core MM paths as compared to
using a an isolated SW bit like PTE_SPECIAL exclusively for this purpose. This
is in line with using PTE_PROTNONE. PTE_SPECIAL seems to be well away from core
PMD path. Is there any particular concern about using PTE_SPECIAL ? Nonetheless
I will evaluate above proposal of using (0b10) to represent invalid but present
huge PMD entry during splitting.

> 
>>> we need to implement our own pmdp_invalidate() or change the generic one
>>> to set a "special" bit instead of just a pmd_mknotpresent?
>>
>> Though arm64 can subscribe __HAVE_ARCH_PMDP_INVALIDATE and implement it's own
>> pmdp_invalidate() in order to not call pmd_mknotpresent() and instead operate
>> on the invalid and special bits directly. But its not going to alter relevant
>> semantics here. AFAICS it might be bit better as it saves pmd_mknotpresent()
>> from putting in that special bit in there which it is not supposed do.
>>
>> IFAICS there is no compelling reason for generic pmdp_invalidate() to change
>> either. It calls pmd_mknotpresent() which invalidates the entry through valid
>> or present bit and platforms which have dedicated huge page bit can still test
>> positive for pmd_present() after it's invalidation. It works for such platforms.
>> Platform specific override is required when invalidation via pmd_mknotpresent()
>> is not enough.
> 
> I'd really like the mknotpresent to be renamed to mknotvalid and then we
> can keep pmdp_invalidate unchanged (well, calling mknotvalid instead).
> 

Though this change really makes sense just from fixing generic pmdp_invalidate()
perspective as all it asks is to invalidate the PMD entry not mark them non-present
and currently calling pmd_mknotpresent() in that sense is bit misleading.

But for arm64 I believe implementing arch specific pmdp_invalidate() via subscribing
__HAVE_ARCH_PMDP_INVALIDATE is bit better. Because the implementation needs more than
just a PMD entry invalidation even with above proposed 0b10 method or with PTE_SPECIAL.
pmd_mknotvalid() should not do that additional stuff but instead a platform specific 
pmdp_invalidate() can incorporate that after doing the real invalidation i.e clearing
the bit 0 in pmd_mknotvalid().