On 8/11/21 11:18 PM, Mike Rapoport wrote:
On Wed, Aug 11, 2021 at 06:37:19PM -0700, Vineet Gupta wrote:
On 8/11/21 5:31 AM, Mike Rapoport wrote:
+/*
+ * For ARC, pgtable_t is not struct page *, but pte_t * (to avoid
+ * extraneous page_address() calculations) hence can't use
+ * use asm-generic/pgalloc.h which assumes it being struct page *
+ */
Another reason to leave ARC without asm-generic/pgalloc.h was
__get_order_pte() that other arches don't have.
So this and pgtable_t aliased to pte_t * seemed to me too much to bother
then, but probably it's worth reconsidering with addition of 3rd and 4th
levels.
I agree that savings of not havign page_address() might not be huge. However
asm-generic/pgalloc.h only has pte allocation routines and all other
allocation levels come from arch file
asm-generic/pgalloc.h has allocation routines up to PUD.
There is also pgtable_pmd_page_ctor() and pgtable_pmd_page_dtor() called in
the generic versions of PMD allocation, it seems they are not called in ARC
implementation.
:-(
So using asm-generic/pgalloc.h would probably save you some THP debugging ;-)
We may even probably accommodate multi-page PTEs in asm-generic/pgalloc.h
with something like
#ifndef __HAVE_ARCH_PTE_GET_ORDER
static inline int __pte_get_order(void)
{
return 0;
}
#endif
Not needed - those cases are unreal, esoteric at best. I'm working on
switching back to canonical struct page based pgtable_t, will do that in v2.
Also for ARCv2, given the arbitrary address split and ensuing paging levels
(as well as support for different page sizes) we will need to make sure that
one page is enough to hold any level's paging using say BUILD_BUG_ON. In
fact that should also be done for 3rd and 4th levels for sanity.
Right, these sanity checks would be useful, but they may live in one of .c
files in arch/arc/mm.
Sure !
Thx,
-Vineet
