Re: [RFC PATCH] mm/hmm: Add userfaultfd support to fault handling

["David Hildenbrand (Arm)" <david@kernel.org>]

On 4/1/26 00:24, Stanislav Kinsburskii wrote:
> Add support for userfaultfd-enabled VMAs to the HMM framework.
> 
> Extract fault handling logic into hmm_handle_mm_fault() to handle both
> regular and userfaultfd-backed mappings. The implementation follows
> fixup_user_fault() for handling VM_FAULT_RETRY and VM_FAULT_COMPLETED, with
> a key difference: instead of retrying or moving forward respectively,
> return -EBUSY after reacquiring mmap_read_lock. Since the lock was
> released, the VMA could have changed, so defer retry logic to the caller.
> 
> This approach is inefficient for userfaultfd-backed VMAs, as HMM can only
> populate one page at a time, but keeps the framework simple by avoiding
> complex retry logic within HMM itself.
> 
> Signed-off-by: Stanislav Kinsburskii <skinsburskii@linux.microsoft.com>
> ---
>  mm/hmm.c |   40 ++++++++++++++++++++++++++++++++++++----
>  1 file changed, 36 insertions(+), 4 deletions(-)
> 
> diff --git a/mm/hmm.c b/mm/hmm.c
> index f6c4ddff4bd6..d04d68e21473 100644
> --- a/mm/hmm.c
> +++ b/mm/hmm.c
> @@ -59,6 +59,35 @@ static int hmm_pfns_fill(unsigned long addr, unsigned long end,
>  	return 0;
>  }
>  
> +static int hmm_handle_mm_fault(struct vm_area_struct *vma,
> +			       unsigned long addr,
> +			       unsigned int fault_flags)
> +{
> +	int ret;
> +
> +	if (userfaultfd_missing(vma)) {
> +		struct mm_struct *mm = vma->vm_mm;
> +
> +		fault_flags |= FAULT_FLAG_ALLOW_RETRY |
> +			       FAULT_FLAG_USER;
> +
> +		ret = handle_mm_fault(vma, addr, fault_flags, NULL);
> +
> +		if (ret & (VM_FAULT_COMPLETED | VM_FAULT_RETRY)) {
> +			mmap_read_lock(mm);
> +			return -EBUSY;
> +		}
> +
> +		if (ret & VM_FAULT_ERROR)
> +			return vm_fault_to_errno(ret, 0);
> +	} else {
> +		ret = handle_mm_fault(vma, addr, fault_flags, NULL);
> +		if (ret & VM_FAULT_ERROR)
> +			return vm_fault_to_errno(ret, 0);
> +	}

I'm surprised that there is userfaultfd_missing() logic required here at
all.

What prevents us from always calling handle_mm_fault() in a way +
handling return values, such that we will just do the right thing
independent of userfaultfd_missing()?

-- 
Cheers,

David