On 27 Jun 2019, at 8:48, 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
>
> -------------------------------------------------------------------------
> |	PMD states	|	pmd_present	|	pmd_trans_huge	|
> -------------------------------------------------------------------------
> |	Mapped		|	Yes		|	Yes		|
> -------------------------------------------------------------------------
> |	Splitting	|	Yes		|	Yes		|
> -------------------------------------------------------------------------
> |	Migration/Swap	|	No		|	No		|
> -------------------------------------------------------------------------
>
> 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. It will need another bit apart from PMD_SECT_VALID to re-
> affirm pmd_present() as true during the THP split process. To comply with
> above mentioned semantics, pmd_trans_huge() should also check pmd_present()
> first before testing presence of an actual transparent huge mapping.
>
> The solution:
>
> Ideally PMD_TYPE_SECT should have been used here instead. But it shares the
> bit position with PMD_SECT_VALID which is used for THP invalidation. Hence
> it will not be there for pmd_present() check after pmdp_invalidate().
>
> PTE_SPECIAL never gets used for PMD mapping i.e there is no pmd_special().
> Hence this bit can be set on the PMD entry during invalidation which can
> help in making pmd_present() return true and in recognizing the fact that
> it still points to memory.
>
> This bit is transient. During the split is process it will be overridden
> by a page table page representing the normal pages in place of erstwhile
> huge page. Other pmdp_invalidate() callers always write a fresh PMD value
> on the entry overriding this transient PTE_SPECIAL making it safe. In the
> past former pmd_[mk]splitting() functions used PTE_SPECIAL.
>
> [1]: https://lkml.org/lkml/2018/10/17/231

Just want to point out that lkml.org link might not be stable.
This one would be better: https://lore.kernel.org/linux-mm/20181017020930.GN30832@redhat.com/


—
Best Regards,
Yan Zi