Re: Hugepages demand paging V1 [2/4]: set_huge_pte() arch updates

[William Lee Irwin III]

On Fri, 22 Oct 2004, William Lee Irwin III wrote:
>> What's described above is not what the patch implements. The patch is
>> calling update_mmu_cache() in a loop on all the virtual base pages of a
>> virtual hugepage, which won't help at all, as it doesn't understand how
>> to find the hugepages regardless of virtual address. AFAICT code to
>> actually do the equivalent of update_mmu_cache() on hugepages most
>> likely involves privileged instructions and perhaps digging around some
>> cpu-specific data structures (e.g. the natively architected pagetables
>> bearing no resemblance to Linux') for almost every non-x86 architecture.

On Fri, Oct 22, 2004 at 08:32:34AM -0700, Christoph Lameter wrote:
> The looping is architecture specific. But you are right for the
> architectures where I simply did the loop that is wrong. The address given
> needs to be correctly calculated which it is not.
> Again this is arch specific stuff and can be done as needed for any
> architecture. There is no intend to generalize this.

The "model", as it were, for pagetable updats, is a 3-stage model:
(1) prepare
(2) update
(3) commit

update_mmu_cache() is stage (3). Linux' software pagetable
modifications are step (2). Generally, architectures are trying to fold
stages (2) and (3) together in set_huge_pte(), so update_mmu_cache() is
not so much of an issue for them (and it's actually incorrect to do
this multiple times or without the architectures' awareness of hugepages
in update_mmu_cache() and so on). To completely regularize the hugetlb
case, merely move the commitment to cpu-native structures or other TLB
insertion done within set_huge_pte() to a new case in update_mmu_cache()
for large pages.

What is in fact far more pressing is flush_dcache_page(), without a
correct implementation of which for hugetlb, user-visible data
corruption follows.


-- wli
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/