On 21.12.24 13:04, Ge Yang wrote:
在 2024/12/21 0:30, David Hildenbrand 写道:
On 20.12.24 09:56, Ge Yang wrote:
在 2024/12/20 0:40, David Hildenbrand 写道:
On 18.12.24 07:33, yangge1116@xxxxxxx wrote:
From: yangge <yangge1116@xxxxxxx>
CCing Oscar, who worked on migrating these pages during memory offlining
and alloc_contig_range().
My machine has 4 NUMA nodes, each equipped with 32GB of memory. I
have configured each NUMA node with 16GB of CMA and 16GB of in-use
hugetlb pages. The allocation of contiguous memory via the
cma_alloc() function can fail probabilistically.
The cma_alloc() function may fail if it sees an in-use hugetlb page
within the allocation range, even if that page has already been
migrated. When in-use hugetlb pages are migrated, they may simply
be released back into the free hugepage pool instead of being
returned to the buddy system. This can cause the
test_pages_isolated() function check to fail, ultimately leading
to the failure of the cma_alloc() function:
cma_alloc()
__alloc_contig_migrate_range() // migrate in-use hugepage
test_pages_isolated()
__test_page_isolated_in_pageblock()
PageBuddy(page) // check if the page is in buddy
I thought this would be working as expected, at least we tested it with
alloc_contig_range / virtio-mem a while ago.
On the memory_offlining path, we migrate hugetlb folios, but also
dissolve any remaining free folios even if it means that we will going
below the requested number of hugetlb pages in our pool.
During alloc_contig_range(), we only migrate them, to then free them up
after migration.
Under which circumstances doe sit apply that "they may simply be
released back into the free hugepage pool instead of being returned to
the buddy system"?
After migration, in-use hugetlb pages are only released back to the
hugetlb pool and are not returned to the buddy system.
We had
commit ae37c7ff79f1f030e28ec76c46ee032f8fd07607
Author: Oscar Salvador <osalvador@xxxxxxx>
Date: Tue May 4 18:35:29 2021 -0700
mm: make alloc_contig_range handle in-use hugetlb pages
alloc_contig_range() will fail if it finds a HugeTLB page within the
range, without a chance to handle them. Since HugeTLB pages can be
migrated as any LRU or Movable page, it does not make sense to bail
out
without trying. Enable the interface to recognize in-use HugeTLB
pages so
we can migrate them, and have much better chances to succeed the call.
And I am trying to figure out if it never worked correctly, or if
something changed that broke it.
In start_isolate_page_range()->isolate_migratepages_block(), we do the
ret = isolate_or_dissolve_huge_page(page, &cc->migratepages);
to add these folios to the cc->migratepages list.
In __alloc_contig_migrate_range(), we migrate the pages using
migrate_pages().
After that, the src hugetlb folios should still be isolated?
Yes.
But I'm
getting
confused when these pages get un-silated and putback to hugetlb/freed.
If the migration is successful, call folio_putback_active_hugetlb to
release the src hugetlb folios back to the free hugetlb pool.
trace:
unmap_and_move_huge_page
folio_putback_active_hugetlb
folio_put
free_huge_folio
alloc_contig_range_noprof
__alloc_contig_migrate_range
if (test_pages_isolated()) //to determine if hugetlb pages in buddy
isolate_freepages_range //grab isolated pages from freelists.
else
undo_isolate_page_range //undo isolate
Ah, now I remember, thanks.
So when we free an ordinary page, we put it onto the buddy isolate list,
from where we can grab it later and nobody can allocate it in the meantime.
In case of hugetlb, we simply free it back to hugetlb, from where it can
likely even get allocated immediately again.
I think that can actually happen in your proposal: the now-free page
will get reallocated, for example for migrating the next folio. Or some
concurrent system activity can simply allocate the now-free folio. Or am
I missing something that prevents these now-free hugetlb folios from
getting re-allocated after migration succeeded?
Conceptually, I think we would want migration code in the case of
alloc_contig_range() to allocate a new folio from the buddy, and to free
the old one back to the buddy immediately, without ever allowing
re-allocation of it.
What needs to be handled is detecting that
(a) we want to allocate a fresh hugetlb folio as migration target
(b) if migration succeeds, we have to free the hugetlb folio back to the
buddy
(c) if migation fails, we have to free the allocated hugetlb foliio back
to the buddy
We could provide a custom alloc_migration_target that we pass to
migrate_page to allocate a fresh hugetlb folio to handle (a). Using the
put_new_folio callback we could handle (c). (b) would need some thought.
Maybe we can also just mark the source folio as we isolate it, and
enlighten migration+freeing code to handle it automatically?
Hoping to get some feedback from hugetlb maintainers.
--
Cheers,
David / dhildenb
