2020년 6월 29일 (월) 오후 4:55, Michal Hocko <mhocko@xxxxxxxxxx>님이 작성:
>
> On Mon 29-06-20 15:27:25, Joonsoo Kim wrote:
> [...]
> > Solution that Introduces a new
> > argument doesn't cause this problem while avoiding CMA regions.
>
> My primary argument is that there is no real reason to treat hugetlb
> dequeing somehow differently. So if we simply exclude __GFP_MOVABLE for
> _any_ other allocation then this certainly has some drawbacks on the
> usable memory for the migration target and it can lead to allocation
> failures (especially on movable_node setups where the amount of movable
> memory might be really high) and therefore longterm gup failures. And
> yes those failures might be premature. But my point is that the behavior
> would be _consistent_. So a user wouldn't see random failures for some
> types of pages while a success for others.
Hmm... I don't agree with your argument. Excluding __GFP_MOVABLE is
a *work-around* way to exclude CMA regions. Implementation for dequeuing
in this patch is a right way to exclude CMA regions. Why do we use a work-around
for this case? To be consistent is important but it's only meaningful
if it is correct.
It should not disrupt to make a better code. And, dequeing is already a special
process that is only available for hugetlb. I think that using
different (correct)
implementations there doesn't break any consistency.
> Let's have a look at this patch. It is simply working that around the
> restriction for a very limited types of pages - only hugetlb pages
> which have reserves in non-cma movable pools. I would claim that many
> setups will simply not have many (if any) spare hugetlb pages in the
> pool except for temporary time periods when a workload is (re)starting
> because this would be effectively a wasted memory.
This can not be a stopper to make the correct code.
> The patch is adding a special case flag to claim what the code already
> does by memalloc_nocma_{save,restore} API so the information is already
> there. Sorry I didn't bring this up earlier but I have completely forgot
> about its existence. With that one in place I do agree that dequeing
> needs a fixup but that should be something like the following instead.
Thanks for letting me know. I don't know it until now. It looks like it's
better to use this API rather than introducing a new argument.
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index 57ece74e3aae..c1595b1d36f3 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -1092,10 +1092,14 @@ static struct page *dequeue_huge_page_nodemask(struct hstate *h, gfp_t gfp_mask,
> /* Movability of hugepages depends on migration support. */
> static inline gfp_t htlb_alloc_mask(struct hstate *h)
> {
> + gfp_t gfp;
> +
> if (hugepage_movable_supported(h))
> - return GFP_HIGHUSER_MOVABLE;
> + gfp = GFP_HIGHUSER_MOVABLE;
> else
> - return GFP_HIGHUSER;
> + gfp = GFP_HIGHUSER;
> +
> + return current_gfp_context(gfp);
> }
>
> static struct page *dequeue_huge_page_vma(struct hstate *h,
>
> If we even fix this general issue for other allocations and allow a
> better CMA exclusion then it would be implemented consistently for
> everybody.
Yes, I have reviewed the memalloc_nocma_{} APIs and found the better way
for CMA exclusion. I will do it after this patch is finished.
> Does this make more sense to you are we still not on the same page wrt
> to the actual problem?
Yes, but we have different opinions about it. As said above, I will make
a patch for better CMA exclusion after this patchset. It will make
code consistent.
I'd really appreciate it if you wait until then.
Thanks.
