Re: [PATCH] KVM: guest_memfd: fix NUMA interleave index double-counting

["Garg, Shivank" <shivankg@amd.com>]

On 6/5/2026 8:25 PM, Michael S. Tsirkin wrote:
> On Fri, Jun 05, 2026 at 06:31:51PM +0530, Garg, Shivank wrote:
>>
>>
>> On 6/5/2026 5:16 AM, Michael S. Tsirkin wrote:
>>> [You don't often get email from mst@redhat.com. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
>>>
>>> On Thu, Jun 04, 2026 at 12:21:15AM +0530, Garg, Shivank wrote:
>>>>
>>>>
>>>> On 6/3/2026 9:27 PM, Michael S. Tsirkin wrote:
>>>>> kvm_gmem_get_policy() sets *ilx to the full page offset
>>>>> (vm_pgoff + vma offset).  But get_vma_policy() adds the page
>>>>> offset on top of *ilx, so the offset is counted twice.  This
>>>>> causes NUMA interleaving to skip nodes: for order-0 pages the
>>>>> effective index jumps by 2 for each consecutive page.
>>>>>
>>>>> The get_policy vm_op should return only a per-file bias in *ilx
>>>>> (like shmem_get_policy does with inode->i_ino), letting
>>>>> get_vma_policy() add the page-offset component.
>>>>>
>>>>> Fix by setting *ilx to inode->i_ino instead of the full page
>>>>> offset.  The page offset is computed by get_vma_policy() in
>>>>> mm/mempolicy.c. The full offset is still computed
>>>>> in kvm_gmem_get_policy() for mpol_shared_policy_lookup().
>>>>> shmem_get_policy() follows the same pattern.
>>>>>
>>>>> Found by Sashiko (sashiko.dev) AI code review.
>>>>>
>>>>> Fixes: ed1ffa810bd6 ("KVM: guest_memfd: Enforce NUMA mempolicy using shared policy")
>>>>> Cc: Sean Christopherson <seanjc@google.com>
>>>>> Cc: Paolo Bonzini <pbonzini@redhat.com>
>>>>> Assisted-by: Claude:claude-opus-4-6
>>>>> Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
>>>>> ---
>>>>>  virt/kvm/guest_memfd.c | 7 ++++---
>>>>>  1 file changed, 4 insertions(+), 3 deletions(-)
>>>>>
>>>>> diff --git a/virt/kvm/guest_memfd.c b/virt/kvm/guest_memfd.c
>>>>> index 69c9d6d546b2..0bcf6fc08e2d 100644
>>>>> --- a/virt/kvm/guest_memfd.c
>>>>> +++ b/virt/kvm/guest_memfd.c
>>>>> @@ -438,11 +438,12 @@ static int kvm_gmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpo
>>>>>  }
>>>>>
>>>>>  static struct mempolicy *kvm_gmem_get_policy(struct vm_area_struct *vma,
>>>>> -                                            unsigned long addr, pgoff_t *pgoff)
>>>>> +                                            unsigned long addr, pgoff_t *ilx)
>>>>>  {
>>>>>         struct inode *inode = file_inode(vma->vm_file);
>>>>> +       pgoff_t pgoff = vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT);
>>>>>
>>>>> -       *pgoff = vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT);
>>>>> +       *ilx = inode->i_ino;
>>>>>
>>>>>         /*
>>>>>          * Return the memory policy for this index, or NULL if none is set.
>>>>> @@ -453,7 +454,7 @@ static struct mempolicy *kvm_gmem_get_policy(struct vm_area_struct *vma,
>>>>>          * can then replace NULL with the default memory policy instead of the
>>>>>          * current task's memory policy.
>>>>>          */
>>>>> -       return mpol_shared_policy_lookup(&GMEM_I(inode)->policy, *pgoff);
>>>>> +       return mpol_shared_policy_lookup(&GMEM_I(inode)->policy, pgoff);
>>>>>  }
>>>>>  #endif /* CONFIG_NUMA */
>>>>>
>>>>> --
>>>>> MST
>>>>>
>>>>
>>>> Thanks for fixing this. LGTM!
>>>>
>>>> Reviewed-by: Shivank Garg <shivankg@amd.com>
>>>
>>>
>>> Can u actually test it though pls?
>>> Because I think another patch I sent in response so Sashiko
>>> is also needed.
>>
>> Hi Michael,
>>
>> Yes, I tested this.
>>
>> I used kretprobes to read *ilx on each kvm_gmem_get_policy(), while calling
>> get_mempolicy(MPOL_F_ADDR) on consecutive offsets(0..7) of guest_memfd mapping:
>>
>> BEFORE:
>> page offset:  0   1   2   3   4   5   6   7
>> *ilx:         0   1   2   3   4   5   6   7
>>   
>> get_vma_policy() again add the page offset on top. so, it will increase by stride 2.
>>
>> AFTER Fix:
>> page offset:  0       1       2       3      ...  7
>> *ilx:         128376  128376  128376  128376 ...  128376
>>
>> It store i_no, so after get_vma_policy(), it will increase by just 1.
>>
>> It's hard to show any wrong allocation with the bug because this index value is not
>> used by allocation path, which uses NO_INTERLEAVE_INDEX.
>>
>> Tested-by: Shivank Garg <shivankg@amd.com>
>>
>> Thanks,
>> Shivank
>>
> 
> 
> So for this to be useful at all
> we do need the patch I sent in response to sashiko, right?
> Mind trying out that one?
> 

I could not find the other patch from you.
Are you talking about this response
https://lore.kernel.org/all/20260604034539-mutt-send-email-mst@kernel.org?

If you send any separate patch to test elsewhere, please point me.

Sashiko comment:
> Doesn't this defeat the purpose of the shared policy, causing interleaving
> to be randomized by the chronological order of vCPU page faults rather than
> deterministically spread based on the guest physical address?

It doesn't defeat the shared policy's purpose.
This design is intentional, The shared policy exists so the policy is a
property of the inode and selected by file offset, and that works.

I don't think we need determinstic interleave now, it can be revisited later.

Thanks,
Shivank