On 2024/3/27 10:04, Huang, Ying wrote:
Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx> writes:
Now the anonymous page allocation already supports multi-size THP (mTHP),
but the numa balancing still prohibits mTHP migration even though it is an
exclusive mapping, which is unreasonable.
Allow scanning mTHP:
Commit 859d4adc3415 ("mm: numa: do not trap faults on shared data section
pages") skips shared CoW pages' NUMA page migration to avoid shared data
segment migration. In addition, commit 80d47f5de5e3 ("mm: don't try to
NUMA-migrate COW pages that have other uses") change to use page_count()
to avoid GUP pages migration, that will also skip the mTHP numa scaning.
Theoretically, we can use folio_maybe_dma_pinned() to detect the GUP
issue, although there is still a GUP race, the issue seems to have been
resolved by commit 80d47f5de5e3. Meanwhile, use the folio_likely_mapped_shared()
to skip shared CoW pages though this is not a precise sharers count. To
check if the folio is shared, ideally we want to make sure every page is
mapped to the same process, but doing that seems expensive and using
the estimated mapcount seems can work when running autonuma benchmark.
Because now we can deal with shared mTHP, it appears even possible to
remove folio_likely_mapped_shared() check?
IMO, the issue solved by commit 859d4adc3415 is about shared CoW
mapping, and I prefer to measure it in another patch:)
Allow migrating mTHP:
As mentioned in the previous thread[1], large folios (including THP) are
more susceptible to false sharing issues among threads than 4K base page,
leading to pages ping-pong back and forth during numa balancing, which is
currently not easy to resolve. Therefore, as a start to support mTHP numa
balancing, we can follow the PMD mapped THP's strategy, that means we can
reuse the 2-stage filter in should_numa_migrate_memory() to check if the
mTHP is being heavily contended among threads (through checking the CPU id
and pid of the last access) to avoid false sharing at some degree. Thus,
we can restore all PTE maps upon the first hint page fault of a large folio
to follow the PMD mapped THP's strategy. In the future, we can continue to
optimize the NUMA balancing algorithm to avoid the false sharing issue with
large folios as much as possible.
Performance data:
Machine environment: 2 nodes, 128 cores Intel(R) Xeon(R) Platinum
Base: 2024-03-25 mm-unstable branch
Enable mTHP to run autonuma-benchmark
mTHP:16K
Base Patched
numa01 numa01
224.70 137.23
numa01_THREAD_ALLOC numa01_THREAD_ALLOC
118.05 50.57
numa02 numa02
13.45 9.30
numa02_SMT numa02_SMT
14.80 7.43
mTHP:64K
Base Patched
numa01 numa01
216.15 135.20
numa01_THREAD_ALLOC numa01_THREAD_ALLOC
115.35 46.93
numa02 numa02
13.24 9.24
numa02_SMT numa02_SMT
14.67 7.31
mTHP:128K
Base Patched
numa01 numa01
205.13 140.41
numa01_THREAD_ALLOC numa01_THREAD_ALLOC
112.93 44.78
numa02 numa02
13.16 9.19
numa02_SMT numa02_SMT
14.81 7.39
[1] https://lore.kernel.org/all/20231117100745.fnpijbk4xgmals3k@xxxxxxxxxxxxxxxxxxx/
Signed-off-by: Baolin Wang <baolin.wang@xxxxxxxxxxxxxxxxx>
---
mm/memory.c | 56 +++++++++++++++++++++++++++++++++++++++++++--------
mm/mprotect.c | 3 ++-
2 files changed, 50 insertions(+), 9 deletions(-)
diff --git a/mm/memory.c b/mm/memory.c
index c30fb4b95e15..36191a9c799c 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -5068,16 +5068,55 @@ static void numa_rebuild_single_mapping(struct vm_fault *vmf, struct vm_area_str
update_mmu_cache_range(vmf, vma, vmf->address, vmf->pte, 1);
}
+static void numa_rebuild_large_mapping(struct vm_fault *vmf, struct vm_area_struct *vma,
+ struct folio *folio, pte_t fault_pte, bool ignore_writable)
+{
+ int nr = pte_pfn(fault_pte) - folio_pfn(folio);
+ unsigned long start = max(vmf->address - nr * PAGE_SIZE, vma->vm_start);
+ unsigned long end = min(start + folio_nr_pages(folio) * PAGE_SIZE, vma->vm_end);
If start is in the middle of folio, it's possible for end to go beyond
the end of folio. So, should be something like below?
Yes, good catch, even though below iteration can skip over the parts
that exceed the size of that folio.
unsigned long end = min(vmf->address + (folio_nr_pages(folio) - nr) * PAGE_SIZE, vma->vm_end);
Yes, this looks good to me. Will do in next version. Thanks.
+ pte_t *start_ptep = vmf->pte - (vmf->address - start) / PAGE_SIZE;
+ bool pte_write_upgrade = vma_wants_manual_pte_write_upgrade(vma);
+ unsigned long addr;
+
+ /* Restore all PTEs' mapping of the large folio */
+ for (addr = start; addr != end; start_ptep++, addr += PAGE_SIZE) {
+ pte_t pte, old_pte;
+ pte_t ptent = ptep_get(start_ptep);
+ bool writable = false;
+
+ if (!pte_present(ptent) || !pte_protnone(ptent))
+ continue;
+
+ if (vm_normal_folio(vma, addr, ptent) != folio)
+ continue;
+
+ if (!ignore_writable) {
+ writable = pte_write(pte);
+ if (!writable && pte_write_upgrade &&
+ can_change_pte_writable(vma, addr, pte))
+ writable = true;
+ }
+
+ old_pte = ptep_modify_prot_start(vma, addr, start_ptep);
+ pte = pte_modify(old_pte, vma->vm_page_prot);
+ pte = pte_mkyoung(pte);
+ if (writable)
+ pte = pte_mkwrite(pte, vma);
+ ptep_modify_prot_commit(vma, addr, start_ptep, old_pte, pte);
+ update_mmu_cache_range(vmf, vma, addr, start_ptep, 1);
Can this be batched for the whole folio?
I thought about it, but things are a little tricky. The folio may not
contain continuous protnone PTEs, should skip non-present or
non-protnone PTEs.
Moreover, it is necessary to define architecture-specified
ptep_modify_prot_start*_nr and ptep_modify_prot_commit*_nr that can
handle multiple PTEs, which is in my TODO list including batch numa
scanning in change_pte_range().
