Re: [PATCH] mm/thp: Correctly differentiate between mapped THP and PMD migration entry

[Anshuman Khandual]

On 11/06/2018 06:05 AM, Will Deacon wrote:
> On Fri, Nov 02, 2018 at 11:45:00AM +0530, Anshuman Khandual wrote:
>> On 10/17/2018 07:39 AM, Andrea Arcangeli wrote:
>>> What we need to do during split is an invalidate of the huge TLB.
>>> There's no pmd_trans_splitting anymore, so we only clear the present
>>> bit in the PTE despite pmd_present still returns true (just like
>>> PROT_NONE, nothing new in this respect). pmd_present never meant the
>>
>> On arm64, the problem is that pmd_present() is tied with pte_present() which
>> checks for PTE_VALID (also PTE_PROT_NONE) but which gets cleared during PTE
>> invalidation. pmd_present() returns false just after the first step of PMD
>> splitting. So pmd_present() needs to be decoupled from PTE_VALID which is
>> same as PMD_SECT_VALID and instead should depend upon a pte bit which sticks
>> around like PAGE_PSE as in case of x86. I am working towards a solution.
> 
> Could we not just go via a PROT_NONE mapping during the split, instead of
> having to allocate a new software bit to treat these invalid ptes as
> present?

The problem might occur during page fault (i.e __handle_mm_fault). As discussed
previously on this thread any potential PTE sticky bit would be used for both
pmd_trans_huge() and pmd_present() wrappers to maintain existing semantics. At
present, PMD state analysis during page fault has conditional block like this.

                if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
                        if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
                                return do_huge_pmd_numa_page(&vmf, orig_pmd);

Using PROT_NONE for pmd_trans_huge() might force PMD page fault to go through
NUMA fault handling all the time as both pmd_trans_huge() and pmd_protnone()
will return true in that situation.