Re: [PATCH v2] mm/huge_memory: bypass THP tuneables for huge pfnmap mappings

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

On 9/7/26 19:33, Lorenzo Stoakes (ARM) wrote:
> The sysfs THP tuneables at /sys/kernel/mm/transparent_huge_pages/ rather
> confusingly only control the behaviour of THP in some instances.
> 
> They are not applicable to MADV_COLLAPSE operations, nor to DAX mappings.
> 
> Long-term, THP is predicated upon compaction being able to obtain large
> folios to populate THP ranges.
> 
> However, vm_normal_folio() returns NULL for PFN map mappings, thus their
> reference count is maintained by the driver, not core mm.
> 
> As a consequence, the folios are not subject to reclaim nor compaction, so
> are not truly part of the THP mechanism at all.
> 
> However, since commit 5dd40721f147 ("mm: allow THP orders for PFNMAPs")
> introduced the ability to establish huge PFN maps, they have been subject
> to THP tuneables.
> 
> This is incorrect - if a huge PFN map is available (defined by
> vma->vm_ops->huge_fault being non-NULL for a VMA_PFNMAP_BIT VMA), then it
> should be mapped huge upon fault-in.
> 
> Correct this by explicitly checking for this while ensuring that smaps
> continues to accurately report THPeligible statistics.
> 
> While here, abstract the entire file-backed THP check in
> vma_can_map_huge_file(), with sensible separation of logic into helper
> functions.
> 
> Note that drm_gem_shmem_mmap() and panthor_gem_mmap() establish huge PFN
> maps of shmem folios, however they are marked unevictable in
> drm_gem_get_pages(), and in any case would fail the reference check in
> __remove_mapping() even if they weren't.
> 
> Failing to map huge PFN maps has resulted in significant real-world
> performance degradation, see links for details.
> 
> Reported-by: Cedric Le Goater <clg@redhat.com>
> Closes: https://lore.kernel.org/linux-mm/20260805055544.1568534-1-clg@redhat.com/
> Reported-by: Saravanan D <saravanand@crusoe.ai>
> Closes: https://lore.kernel.org/linux-mm/20260821070520.25759-1-saravanand@crusoe.ai/
> Fixes: 5dd40721f147 ("mm: allow THP orders for PFNMAPs")
> Cc: stable@vger.kernel.org
> Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
> Reviewed-by: SJ Park <sj@kernel.org>
> Tested-by: Saravanan D <saravanand@crusoe.ai>
> Tested-by: Lance Yang <lance.yang@linux.dev>
> Reviewed-by: Zi Yan <ziy@nvidia.com>
> Signed-off-by: Lorenzo Stoakes (ARM) <ljs@kernel.org>
> ---
> v2:
> - Added tags thanks everybody!
> - Updated to use vma_file_*() naming as per Zi.
> - Also updated vma_can_map_huge_file() to vma_file_can_map_huge() as per
>   David.
> 
> v1:
> https://lore.kernel.org/r/20260827-hugepfn-allowable-orders-v1-1-94819c8807c8@kernel.org
> ---
>  mm/huge_memory.c | 86 +++++++++++++++++++++++++++++++++++++++++---------------
>  1 file changed, 64 insertions(+), 22 deletions(-)
> 
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index afbb5974bd22..75c204af40cb 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -92,7 +92,7 @@ unsigned long huge_anon_orders_madvise __read_mostly;
>  unsigned long huge_anon_orders_inherit __read_mostly;
>  static bool anon_orders_configured __initdata;
>  
> -static inline bool file_thp_enabled(struct vm_area_struct *vma)
> +static inline bool file_thp_enabled(const struct vm_area_struct *vma)
>  {
>  	struct inode *inode;
>  
> @@ -118,6 +118,67 @@ static bool vma_is_special_huge(const struct vm_area_struct *vma)
>  	return vma_test_any(vma, VMA_PFNMAP_BIT, VMA_MIXEDMAP_BIT);
>  }
>  
> +static bool vma_file_bypass_thp_tuneables(const struct vm_area_struct *vma,
> +		enum tva_type type)
> +{
> +	const bool has_huge_fault = vma->vm_ops->huge_fault;
> +
> +	/* MADV_COLLAPSE ignores tuneables. */
> +	if (type == TVA_FORCED_COLLAPSE)
> +		return true;
> +	/* Huge PFN mappings are uncompactable so the policy doesn't apply. */
> +	if (vma_test(vma, VMA_PFNMAP_BIT) && has_huge_fault)
> +		return true;
> +	return false;

/* Huge PFN mappings are uncompactable so the policy doesn't apply. */
return vma_test(vma, VMA_PFNMAP_BIT) && has_huge_fault;

> +}
> +
> +static bool vma_file_allow_thp_tuneables(vm_flags_t vm_flags)
> +{
> +	/* THP=always? */
> +	if (hugepage_global_always())
> +		return true;
> +	/* THP=madvise and marked MADV_HUGEPAGE? */
> +	if (hugepage_global_enabled() && (vm_flags & VM_HUGEPAGE))
> +		return true;
> +	return false;

/* THP=madvise and marked MADV_HUGEPAGE? */
return hugepage_global_enabled() && (vm_flags & VM_HUGEPAGE);

> +}
> +
> +static bool vma_file_check_thp_tuneables(const struct vm_area_struct *vma,
> +		vm_flags_t vm_flags, enum tva_type type)
> +{
> +	return vma_file_bypass_thp_tuneables(vma, type) ||
> +		vma_file_allow_thp_tuneables(vm_flags);
> +}
> +
> +static bool vma_file_can_map_huge(const struct vm_area_struct *vma,
> +		vm_flags_t vm_flags, enum tva_type type)
> +{
> +	const bool has_huge_fault = vma->vm_ops->huge_fault;
> +
> +	/*
> +	 * Enforce THP collapse requirements as necessary. Anonymous vmas
> +	 * were already handled in thp_vma_allowable_orders().
> +	 */

The reference to an anon vmas is not really suitable in the vma_file_* helper
(would be better kept in the caller). I'd just drop the reference to anonymous
vmas entirely.

> +	if (!vma_file_check_thp_tuneables(vma, vm_flags, type))
> +		return false;
> +
> +	switch (type) {
> +	case TVA_PAGEFAULT:
> +		/*
> +		 * Trust that ->huge_fault() handlers know what they are doing
> +		 * in fault path.
> +		 */
> +		return has_huge_fault;
> +	case TVA_SMAPS:
> +		if (has_huge_fault)
> +			return true;
> +		fallthrough;
> +	default:
> +		/* Only regular file is valid in collapse path. */
> +		return file_thp_enabled(vma);
> +	}
> +}

Much nicer to read, thanks

Acked-by: David Hildenbrand (Arm) <david@kernel.org>

-- 
Cheers,

David