Re: [PATCH 1/2] mm/numa: no task_numa_fault() call if page table is changed

On 8 Aug 2024, at 4:22, David Hildenbrand wrote:

On 08.08.24 05:19, Baolin Wang wrote:

...

On 2024/8/8 02:47, Zi Yan wrote:

...

When handling a numa page fault, task_numa_fault() should be called by a process that restores the page table of the faulted folio to avoid duplicated stats counting. Commit b99a342d4f11 ("NUMA balancing: reduce TLB flush via delaying mapping on hint page fault") restructured do_numa_page() and do_huge_pmd_numa_page() and did not avoid task_numa_fault() call in the second page table check after a numa migration failure. Fix it by making all !pte_same()/!pmd_same() return immediately.

This issue can cause task_numa_fault() being called more than necessary and lead to unexpected numa balancing results (It is hard to tell whether the issue will cause positive or negative performance impact due to duplicated numa fault counting).

Reported-by: "Huang, Ying" ying.huang@intel.com Closes: https://lore.kernel.org/linux-mm/87zfqfw0yw.fsf@yhuang6-desk2.ccr.corp.intel... Fixes: b99a342d4f11 ("NUMA balancing: reduce TLB flush via delaying mapping on hint page fault") Cc: stable@vger.kernel.org Signed-off-by: Zi Yan ziy@nvidia.com

The fix looks reasonable to me. Feel free to add: Reviewed-by: Baolin Wang baolin.wang@linux.alibaba.com

(Nit: These goto labels are a bit confusing and might need some cleanup in the future.)

Agreed, maybe we should simply handle that right away and replace the "goto out;" users by "return 0;".

Then, just copy the 3 LOC.

For mm/memory.c that would be:

diff --git a/mm/memory.c b/mm/memory.c index 67496dc5064f..410ba50ca746 100644 --- a/mm/memory.c +++ b/mm/memory.c @@ -5461,7 +5461,7 @@ static vm_fault_t do_numa_page(struct vm_fault *vmf) if (unlikely(!pte_same(old_pte, vmf->orig_pte))) { pte_unmap_unlock(vmf->pte, vmf->ptl);

•           goto out;
  

•           return 0;
    }
     pte = pte_modify(old_pte, vma->vm_page_prot);
  

@@ -5528,15 +5528,14 @@ static vm_fault_t do_numa_page(struct vm_fault *vmf) vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address, &vmf->ptl); if (unlikely(!vmf->pte))

•                   goto out;
  

•                   return 0;
            if (unlikely(!pte_same(ptep_get(vmf->pte), vmf->orig_pte))) {
                    pte_unmap_unlock(vmf->pte, vmf->ptl);
  

•                   goto out;
  

•                   return 0;
            }
            goto out_map;
    }
  

-out: if (nid != NUMA_NO_NODE) task_numa_fault(last_cpupid, nid, nr_pages, flags); return 0; @@ -5552,7 +5551,9 @@ static vm_fault_t do_numa_page(struct vm_fault *vmf) numa_rebuild_single_mapping(vmf, vma, vmf->address, vmf->pte, writable); pte_unmap_unlock(vmf->pte, vmf->ptl);

•   goto out;
  

•   if (nid != NUMA_NO_NODE)
  

•           task_numa_fault(last_cpupid, nid, nr_pages, flags);
  

•   return 0;
  

}

Looks good to me. Thanks.

Hi Andrew,

Should I resend this for an easy back porting? Or you want to fold David’s changes in directly?

Best Regards, Yan, Zi