Re: [PATCH] mm/ksm: fix userfaultfd protection during merge and unmerge

[Lance Yang <lance.yang@linux.dev>]

On 2026/10/26 22:27, Lorenzo Stoakes (ARM) wrote:
> On Mon, Oct 26, 2026 at 09:00:54PM +0800, Lance Yang wrote:
>> KSM loses userfaultfd protection when replace_page() builds a new PTE
>> from vma->vm_page_prot. Dropping its uffd bit lets a subsequent write go
>> through ordinary COW instead of userfaultfd write protection handling. For
>> RWP, a merged page also becomes readable without faulting.
> 
> I feel we probably shouldn't have allowed an extension to UFFD while it's still
> fundamentally a broken feature :))
> 
> Which is evidenced by the SLEW of AI bug reports, and human-reported bug
> reports, and endless history of bug reports and it infecting the rest of mm,
> etc. etc. /rant

Heh, UFFD isn't winning you over today :D

>>
>> Keeping the uffd bit and restoring PAGE_NONE fixes merging, but we also
>> have to handle unmerging. An existing KSM page can be protected with RWP
>> before MADV_UNMERGEABLE, so unmerging already needs fixing independently
>> of preserving RWP during merging.
>>
>> Unsharing replaces a shared page with a private one and should leave
>> userfaultfd protection in place, as do_wp_page() already does for write
>> protection. Currently, an unshare fault reaches do_uffd_rwp() before it
>> can proceed to do_wp_page().
>>
>> - With synchronous RWP, handle_userfault() returns VM_FAULT_SIGBUS.
>>    break_ksm() then reports success with a KSM page still merged.
>>
>> - With asynchronous RWP, protection is cleared before unsharing.
>>
>> Preserve the uffd bit when merging, and restore PAGE_NONE for RWP. Let
>> FAULT_FLAG_UNSHARE skip RWP fault handling and keep PAGE_NONE on each
>> private copy as well. Unsharing can then complete with RWP still armed
>> for a subsequent access.
>>
>> Fixes: 63b2d4174c4a ("userfaultfd: wp: add the writeprotect API to userfaultfd ioctl")
>> Fixes: 2d427bb01639 ("mm/userfaultfd: add RWP fault delivery and expose UFFDIO_REGISTER_MODE_RWP")
>> Link: https://lore.kernel.org/all/20261019010934.697895-1-samarth.colleges@gmail.com/
>> Link: https://lore.kernel.org/all/20261020055518.36087-1-lance.yang@linux.dev/
>> Cc: stable@vger.kernel.org
>> Co-developed-by: Samarth Narang <samarth.colleges@gmail.com>
>> Signed-off-by: Samarth Narang <samarth.colleges@gmail.com>
>> Signed-off-by: Lance Yang <lance.yang@linux.dev>
> 
> Looks sensible to me so:
> 
> Acked-by: Lorenzo Stoakes (ARM) <ljs@kernel.org>

Really appreciated!

>> ---
>>   mm/ksm.c    |  8 +++++++-
>>   mm/memory.c | 14 ++++++++++----
>>   2 files changed, 17 insertions(+), 5 deletions(-)
>>
>> diff --git a/mm/ksm.c b/mm/ksm.c
>> index 910a1307ac9f..d6f736a1aa2f 100644
>> --- a/mm/ksm.c
>> +++ b/mm/ksm.c
>> @@ -34,6 +34,7 @@
>>   #include <linux/memory.h>
>>   #include <linux/mmu_notifier.h>
>>   #include <linux/swap.h>
>> +#include <linux/userfaultfd_k.h>
>>   #include <linux/ksm.h>
>>   #include <linux/hashtable.h>
>>   #include <linux/freezer.h>
>> @@ -1463,9 +1464,14 @@ static int replace_page(struct vm_area_struct *vma, struct page *page,
>>   		dec_mm_counter(mm, MM_ANONPAGES);
>>   	}
>>
>> -	/* Preserve soft-dirty when replacing the page. */
>> +	/* Preserve soft-dirty and uffd protection when replacing the page. */
>>   	if (pte_soft_dirty(orig_pte))
>>   		newpte = pte_mksoft_dirty(newpte);
>> +	if (pte_uffd(orig_pte)) {
>> +		newpte = pte_mkuffd(newpte);
>> +		if (userfaultfd_rwp(vma))
>> +			newpte = pte_modify(newpte, PAGE_NONE);
> 
> OK, so by definition, RWP pages are protnone, and this resets it correctly.

Right, we need PAGE_NONE back as well as the uffd bit :)

>> +	}
>>
>>   	flush_cache_page(vma, addr, pte_pfn(ptep_get(ptep)));
>>   	/*
>> diff --git a/mm/memory.c b/mm/memory.c
>> index 6a93b7fa48b2..99439a7afa63 100644
>> --- a/mm/memory.c
>> +++ b/mm/memory.c
>> @@ -4122,8 +4122,11 @@ static vm_fault_t wp_page_copy(struct vm_fault *vmf)
>>   		if (unlikely(unshare)) {
>>   			if (pte_soft_dirty(vmf->orig_pte))
>>   				entry = pte_mksoft_dirty(entry);
>> -			if (pte_uffd(vmf->orig_pte))
>> +			if (pte_uffd(vmf->orig_pte)) {
>>   				entry = pte_mkuffd(entry);
>> +				if (userfaultfd_rwp(vma))
>> +					entry = pte_modify(entry, PAGE_NONE);
>> +			}
> 
> Could this be abstracted? You're doing the same pattern here, which also exists
> in copy_present_page() and presumably many other places.
> 
> But maybe that's one for a follow up?

Yeah, a helper would be nice here. I'd keep this fix small and do that
as a follow-up :D

Cheers, Lance

>>   		} else {
>>   			entry = maybe_mkwrite(pte_mkdirty(entry), vma);
>>   		}
>> @@ -6659,9 +6662,12 @@ static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
>>   		 * VM_UFFD_RWP VMA, a protnone PTE without the uffd bit is
>>   		 * NUMA hinting and must still fall through to do_numa_page().
>>   		 */
>> -		if (userfaultfd_pte_rwp(vmf->vma, vmf->orig_pte))
>> -			return do_uffd_rwp(vmf);
>> -		return do_numa_page(vmf);
>> +		if (userfaultfd_pte_rwp(vmf->vma, vmf->orig_pte)) {
>> +			if (!(vmf->flags & FAULT_FLAG_UNSHARE))
>> +				return do_uffd_rwp(vmf);
>> +		} else {
>> +			return do_numa_page(vmf);
>> +		}
>>   	}
>>
>>   	spin_lock(vmf->ptl);
>> --
>> 2.49.0
>>
> 
> --
> Cheers, Lorenzo