On 2024/2/23 22:19, Oscar Salvador wrote:
On Fri, Feb 23, 2024 at 10:56:48AM +0800, Baolin Wang wrote:
I previously considered passing the MR_reason argument to the
htlb_modify_alloc_mask(), which is only used by hugetlb migration.
But in alloc_hugetlb_folio_nodemask(), if there are available hugetlb on
other nodes, we should allow migrating, that will not break the per-node
hugetlb pool.
That's why I just change the gfp_mask for allocating a new hguetlb when
migration, that can break the pool.
Code-wise I think this is good, but I'm having some feelings
about where filter out the mask.
Ok, I'm trying to get my head around this.
It's been a while since I looked into hugetlb code, so here we go.
You mentioned that the only reason not to fiddle with gfp_mask before calling
in alloc_hugetlb_folio_nodemask(), was that we might be able to find a hugetlb
page in another node, and that that's ok because since all nodes remain with
the same number of hugetlb pages, per-node pool doesn't get broken.
Now, I see that dequeue_hugetlb_folio_nodemask() first tries to get the zonelist
of the preferred node, and AFAICS, if it has !GFP_THISNODE, it should also
get the zonelists of all other nodes, so we might fallback.
Right.
In the hope of finding a way to be able to filter out in htlb_modify_alloc_mask(),
I was trying to see whether we could skip GFP_THISNODE in
dequeue_hugetlb_folio_nodemask() but no because we might end up dequeueing
a hugetlb which sits in another node, while we really specified __GFP_THISNODE.
The only way might be to somehow decouple dequeue_hugetlb_folio_nodemask()
from alloc_hugetlb_folio_nodemask() and do some kind of gfp modification
between the two calls.
IMO, I'm not sure whether it's appropriate to decouple
dequeue_hugetlb_folio_nodemask() from alloc_hugetlb_folio_nodemask()
into two separate functions for the users to call, because these details
should be hidden within the hugetlb core implementation.
Instead, I can move the gfp_mask fiddling into a new helper, and move
the helper into alloc_migrate_hugetlb_folio(). Temporary hugetlb
allocation has its own gfp strategy seems reasonable to me.
Another thing I dislike is the "-1" in alloc_hugetlb_folio_vma().
I think at least it deserves a comment like "Passing -1 will make us stick
to GFP_THISNODE".
Sure, will add some comments.
Although that is another thing, we will pass "-1" which forces GFP_THISNODE
when allocating a newly fresh hugetlb page, but in dequeue_hugetlb_folio_nodemask()
we might get a page from a different node.
That doesn't break per-node pool, but it is somehow odd?
Only hugetlb_mfill_atomic_pte() will use -1, which is used to allocate a
temporary hugetlb to hold the copied content that will be immediately
released if uffd copy completes (see commmit 8cc5fcbb5be8). Therefore,
here it is allowed to fallback to available hugetlb on other nodes, but
it is not allowed to break the per-node pool.
