Re: [PATCH v2 04/11] mm: convert VM_PFNMAP tracking to pfnmap_track() + pfnmap_untrack()

[David Hildenbrand <david@redhat.com>]

On 12.05.25 18:42, Lorenzo Stoakes wrote:
> On Mon, May 12, 2025 at 02:34:17PM +0200, David Hildenbrand wrote:
>> Let's use our new interface. In remap_pfn_range(), we'll now decide
>> whether we have to track (full VMA covered) or only lookup the
>> cachemode (partial VMA covered).
>>
>> Remember what we have to untrack by linking it from the VMA. When
>> duplicating VMAs (e.g., splitting, mremap, fork), we'll handle it similar
>> to anon VMA names, and use a kref to share the tracking.
>>
>> Once the last VMA un-refs our tracking data, we'll do the untracking,
>> which simplifies things a lot and should sort our various issues we saw
>> recently, for example, when partially unmapping/zapping a tracked VMA.
>>
>> This change implies that we'll keep tracking the original PFN range even
>> after splitting + partially unmapping it: not too bad, because it was
>> not working reliably before. The only thing that kind-of worked before
>> was shrinking such a mapping using mremap(): we managed to adjust the
>> reservation in a hacky way, now we won't adjust the reservation but
>> leave it around until all involved VMAs are gone.
>>
>> If that ever turns out to be an issue, we could hook into VM splitting
>> code and split the tracking; however, that adds complexity that might
>> not be required, so we'll keep it simple for now.
>>
>> Acked-by: Ingo Molnar <mingo@kernel.org> # x86 bits
>> Signed-off-by: David Hildenbrand <david@redhat.com>
> 
> Other than small nit below,
> 
> Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
> 
>> ---
>>   include/linux/mm_inline.h |  2 +
>>   include/linux/mm_types.h  | 11 ++++++
>>   mm/memory.c               | 82 +++++++++++++++++++++++++++++++--------
>>   mm/mmap.c                 |  5 ---
>>   mm/mremap.c               |  4 --
>>   mm/vma_init.c             | 50 ++++++++++++++++++++++++
>>   6 files changed, 129 insertions(+), 25 deletions(-)
>>
>> diff --git a/include/linux/mm_inline.h b/include/linux/mm_inline.h
>> index f9157a0c42a5c..89b518ff097e6 100644
>> --- a/include/linux/mm_inline.h
>> +++ b/include/linux/mm_inline.h
>> @@ -447,6 +447,8 @@ static inline bool anon_vma_name_eq(struct anon_vma_name *anon_name1,
>>
>>   #endif  /* CONFIG_ANON_VMA_NAME */
>>
>> +void pfnmap_track_ctx_release(struct kref *ref);
>> +
>>   static inline void init_tlb_flush_pending(struct mm_struct *mm)
>>   {
>>   	atomic_set(&mm->tlb_flush_pending, 0);
>> diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
>> index 15808cad2bc1a..3e934dc6057c4 100644
>> --- a/include/linux/mm_types.h
>> +++ b/include/linux/mm_types.h
>> @@ -763,6 +763,14 @@ struct vma_numab_state {
>>   	int prev_scan_seq;
>>   };
>>
>> +#ifdef __HAVE_PFNMAP_TRACKING
>> +struct pfnmap_track_ctx {
>> +	struct kref kref;
>> +	unsigned long pfn;
>> +	unsigned long size;	/* in bytes */
>> +};
>> +#endif
>> +
>>   /*
>>    * Describes a VMA that is about to be mmap()'ed. Drivers may choose to
>>    * manipulate mutable fields which will cause those fields to be updated in the
>> @@ -900,6 +908,9 @@ struct vm_area_struct {
>>   	struct anon_vma_name *anon_name;
>>   #endif
>>   	struct vm_userfaultfd_ctx vm_userfaultfd_ctx;
>> +#ifdef __HAVE_PFNMAP_TRACKING
>> +	struct pfnmap_track_ctx *pfnmap_track_ctx;
>> +#endif
>>   } __randomize_layout;
>>
>>   #ifdef CONFIG_NUMA
>> diff --git a/mm/memory.c b/mm/memory.c
>> index 064fc55d8eab9..4cf4adb0de266 100644
>> --- a/mm/memory.c
>> +++ b/mm/memory.c
>> @@ -1371,7 +1371,7 @@ copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
>>   	struct mm_struct *dst_mm = dst_vma->vm_mm;
>>   	struct mm_struct *src_mm = src_vma->vm_mm;
>>   	struct mmu_notifier_range range;
>> -	unsigned long next, pfn = 0;
>> +	unsigned long next;
>>   	bool is_cow;
>>   	int ret;
>>
>> @@ -1381,12 +1381,6 @@ copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
>>   	if (is_vm_hugetlb_page(src_vma))
>>   		return copy_hugetlb_page_range(dst_mm, src_mm, dst_vma, src_vma);
>>
>> -	if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
>> -		ret = track_pfn_copy(dst_vma, src_vma, &pfn);
>> -		if (ret)
>> -			return ret;
>> -	}
>> -
>>   	/*
>>   	 * We need to invalidate the secondary MMU mappings only when
>>   	 * there could be a permission downgrade on the ptes of the
>> @@ -1428,8 +1422,6 @@ copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
>>   		raw_write_seqcount_end(&src_mm->write_protect_seq);
>>   		mmu_notifier_invalidate_range_end(&range);
>>   	}
>> -	if (ret && unlikely(src_vma->vm_flags & VM_PFNMAP))
>> -		untrack_pfn_copy(dst_vma, pfn);
>>   	return ret;
>>   }
>>
>> @@ -1924,9 +1916,6 @@ static void unmap_single_vma(struct mmu_gather *tlb,
>>   	if (vma->vm_file)
>>   		uprobe_munmap(vma, start, end);
>>
>> -	if (unlikely(vma->vm_flags & VM_PFNMAP))
>> -		untrack_pfn(vma, 0, 0, mm_wr_locked);
>> -
>>   	if (start != end) {
>>   		if (unlikely(is_vm_hugetlb_page(vma))) {
>>   			/*
>> @@ -2872,6 +2861,36 @@ int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
>>   	return error;
>>   }
>>
>> +#ifdef __HAVE_PFNMAP_TRACKING
>> +static inline struct pfnmap_track_ctx *pfnmap_track_ctx_alloc(unsigned long pfn,
>> +		unsigned long size, pgprot_t *prot)
>> +{
>> +	struct pfnmap_track_ctx *ctx;
>> +
>> +	if (pfnmap_track(pfn, size, prot))
>> +		return ERR_PTR(-EINVAL);
>> +
>> +	ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
>> +	if (unlikely(!ctx)) {
>> +		pfnmap_untrack(pfn, size);
>> +		return ERR_PTR(-ENOMEM);
>> +	}
>> +
>> +	ctx->pfn = pfn;
>> +	ctx->size = size;
>> +	kref_init(&ctx->kref);
>> +	return ctx;
>> +}
>> +
>> +void pfnmap_track_ctx_release(struct kref *ref)
>> +{
>> +	struct pfnmap_track_ctx *ctx = container_of(ref, struct pfnmap_track_ctx, kref);
>> +
>> +	pfnmap_untrack(ctx->pfn, ctx->size);
>> +	kfree(ctx);
>> +}
>> +#endif /* __HAVE_PFNMAP_TRACKING */
>> +
>>   /**
>>    * remap_pfn_range - remap kernel memory to userspace
>>    * @vma: user vma to map to
>> @@ -2884,20 +2903,51 @@ int remap_pfn_range_notrack(struct vm_area_struct *vma, unsigned long addr,
>>    *
>>    * Return: %0 on success, negative error code otherwise.
>>    */
>> +#ifdef __HAVE_PFNMAP_TRACKING
>>   int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
>>   		    unsigned long pfn, unsigned long size, pgprot_t prot)
>>   {
>> +	struct pfnmap_track_ctx *ctx = NULL;
>>   	int err;
>>
>> -	err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
>> -	if (err)
>> +	size = PAGE_ALIGN(size);
>> +
>> +	/*
>> +	 * If we cover the full VMA, we'll perform actual tracking, and
>> +	 * remember to untrack when the last reference to our tracking
>> +	 * context from a VMA goes away. We'll keep tracking the whole pfn
>> +	 * range even during VMA splits and partial unmapping.
>> +	 *
>> +	 * If we only cover parts of the VMA, we'll only setup the cachemode
>> +	 * in the pgprot for the pfn range.
>> +	 */
>> +	if (addr == vma->vm_start && addr + size == vma->vm_end) {
>> +		if (vma->pfnmap_track_ctx)
>> +			return -EINVAL;
>> +		ctx = pfnmap_track_ctx_alloc(pfn, size, &prot);
>> +		if (IS_ERR(ctx))
>> +			return PTR_ERR(ctx);
>> +	} else if (pfnmap_setup_cachemode(pfn, size, &prot)) {
>>   		return -EINVAL;
>> +	}
>>
>>   	err = remap_pfn_range_notrack(vma, addr, pfn, size, prot);
>> -	if (err)
>> -		untrack_pfn(vma, pfn, PAGE_ALIGN(size), true);
>> +	if (ctx) {
>> +		if (err)
>> +			kref_put(&ctx->kref, pfnmap_track_ctx_release);
>> +		else
>> +			vma->pfnmap_track_ctx = ctx;
>> +	}
>>   	return err;
>>   }
>> +
>> +#else
>> +int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
>> +		    unsigned long pfn, unsigned long size, pgprot_t prot)
>> +{
>> +	return remap_pfn_range_notrack(vma, addr, pfn, size, prot);
>> +}
>> +#endif
>>   EXPORT_SYMBOL(remap_pfn_range);
>>
>>   /**
>> diff --git a/mm/mmap.c b/mm/mmap.c
>> index 50f902c08341a..09c563c951123 100644
>> --- a/mm/mmap.c
>> +++ b/mm/mmap.c
>> @@ -1784,11 +1784,6 @@ __latent_entropy int dup_mmap(struct mm_struct *mm, struct mm_struct *oldmm)
>>   		tmp = vm_area_dup(mpnt);
>>   		if (!tmp)
>>   			goto fail_nomem;
>> -
>> -		/* track_pfn_copy() will later take care of copying internal state. */
>> -		if (unlikely(tmp->vm_flags & VM_PFNMAP))
>> -			untrack_pfn_clear(tmp);
>> -
>>   		retval = vma_dup_policy(mpnt, tmp);
>>   		if (retval)
>>   			goto fail_nomem_policy;
>> diff --git a/mm/mremap.c b/mm/mremap.c
>> index 7db9da609c84f..6e78e02f74bd3 100644
>> --- a/mm/mremap.c
>> +++ b/mm/mremap.c
>> @@ -1191,10 +1191,6 @@ static int copy_vma_and_data(struct vma_remap_struct *vrm,
>>   	if (is_vm_hugetlb_page(vma))
>>   		clear_vma_resv_huge_pages(vma);
>>
>> -	/* Tell pfnmap has moved from this vma */
>> -	if (unlikely(vma->vm_flags & VM_PFNMAP))
>> -		untrack_pfn_clear(vma);
>> -
>>   	*new_vma_ptr = new_vma;
>>   	return err;
>>   }
>> diff --git a/mm/vma_init.c b/mm/vma_init.c
>> index 967ca85179864..8e53c7943561e 100644
>> --- a/mm/vma_init.c
>> +++ b/mm/vma_init.c
>> @@ -71,7 +71,51 @@ static void vm_area_init_from(const struct vm_area_struct *src,
>>   #ifdef CONFIG_NUMA
>>   	dest->vm_policy = src->vm_policy;
>>   #endif
>> +#ifdef __HAVE_PFNMAP_TRACKING
>> +	dest->pfnmap_track_ctx = NULL;
>> +#endif
>> +}
>> +
>> +#ifdef __HAVE_PFNMAP_TRACKING
>> +static inline int vma_pfnmap_track_ctx_dup(struct vm_area_struct *orig,
>> +		struct vm_area_struct *new)
>> +{
>> +	struct pfnmap_track_ctx *ctx = orig->pfnmap_track_ctx;
>> +
>> +	if (likely(!ctx))
>> +		return 0;
>> +
>> +	/*
>> +	 * We don't expect to ever hit this. If ever required, we would have
>> +	 * to duplicate the tracking.
>> +	 */
>> +	if (unlikely(kref_read(&ctx->kref) >= REFCOUNT_MAX))
> 
> How not expected is this? :) maybe use WARN_ON_ONCE() if it really should
> never happen?
I guess if we mmap a large PFNMAP and then split it into individual 
PTE-sized chunks, we could get many VMAs per-process referencing that 
tracing.

Combine that with fork() and I assume one could hit this -- when really 
trying hard to achieve it. (probably as privileged user to get a big 
VM_PFNMAP, though, but not sure)

In that case, a WARN_ON_ONCE() would be bad -- because it could be 
triggered by the user.

We could do a pr_warn_once() instead, stating that this is not supported 
right now?

-- 
Cheers,

David / dhildenb