Re: [PATCH 2/2] mm: support multi-size THP numa balancing

From: Baolin Wang
Date: Thu Mar 28 2024 - 07:34:57 EST




On 2024/3/28 17:25, David Hildenbrand wrote:
On 26.03.24 12:51, Baolin Wang wrote:
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.

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);
+    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;
+

Should you be using folio_pte_batch() in the caller to collect all applicable PTEs and then only have function that batch-changes a given nr of PTEs?

(just like we are now batching other stuff)

Seems folio_pte_batch() is not suitable for numa balancing, since we did not care about other PTE bits, only care about the protnone bits. And after more thinking, I think I can drop the vm_normal_folio() validation, since all PTEs are ensured to be within the range of the folio size.