* [PATCH] mm: skip temporary folio allocation for anonymous UFFD missing faults
@ 2026-11-19 18:00 Avi Weiss
2026-11-19 18:22 ` Lorenzo Stoakes (ARM)
0 siblings, 1 reply; 2+ messages in thread
From: Avi Weiss @ 2026-11-19 18:00 UTC (permalink / raw)
To: david; +Cc: akpm, ljs, rppt, peterx, linux-mm, linux-kernel, Avi Weiss
An anonymous write fault on a VMA registered for missing userfaultfd
faults currently allocates, zeroes and memcg-charges a folio before
checking the PTE and handing the fault to userspace. The folio is then
immediately discarded without ever being mapped.
Check the PTE and address-space stability under the PTE lock before
allocating a folio when missing faults are registered. Deliver the fault
to userfaultfd without allocating and discarding a temporary folio.
In a four-boot native B/P/B/P comparison, the mean UFFD fault median
fell from 13,997.75 ns to 13,093.55 ns (-904.2 ns, -6.46%). A guest
allocation trace saw 256 fewer allocations in the faulting task for
256 registered missing faults.
Note: this changes OOM ordering. In a fixed 18 MiB memcg test, the
baseline OOM-killed only the faulting client before delivering a UFFD
event. With this change, the event reached an unprotected shared handler,
and its UFFDIO_COPY allocation could trigger an additional OOM kill of the
handler. A separate surviving client consequently lost its handler and
failed a page-content check after the UFFD registration was released.
The same handler-death mechanism was reproducible on the unmodified
(mainline) kernel when memory pressure was imposed after event delivery
but before UFFDIO_COPY, although the fixed-limit outcomes differed.
Signed-off-by: Avi Weiss <thnkslprpt@gmail.com>
---
mm/memory.c | 22 ++++++++++++++++------
1 file changed, 16 insertions(+), 6 deletions(-)
diff --git a/mm/memory.c b/mm/memory.c
index 0fb676d..63b39fa 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -5547,6 +5547,22 @@ static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
ret = vmf_anon_prepare(vmf);
if (ret)
return ret;
+
+ if (unlikely(userfaultfd_missing(vma))) {
+ vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
+ vmf->address, &vmf->ptl);
+ if (!vmf->pte)
+ goto unlock;
+ if (vmf_pte_changed(vmf)) {
+ update_mmu_tlb(vma, vmf->address, vmf->pte);
+ goto unlock;
+ }
+ ret = check_stable_address_space(vma->vm_mm);
+ if (ret)
+ goto unlock;
+ pte_unmap_unlock(vmf->pte, vmf->ptl);
+ return handle_userfault(vmf, VM_UFFD_MISSING);
+ }
/* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */
folio = alloc_anon_folio(vmf);
if (IS_ERR(folio))
@@ -5579,12 +5595,6 @@ static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
if (ret)
goto release;
- /* Deliver the page fault to userland, check inside PT lock */
- if (userfaultfd_missing(vma)) {
- pte_unmap_unlock(vmf->pte, vmf->ptl);
- folio_put(folio);
- return handle_userfault(vmf, VM_UFFD_MISSING);
- }
map_anon_folio_pte_pf(folio, vmf->pte, vma, addr,
vmf_orig_pte_uffd_wp(vmf));
unlock:
--
2.43.0
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] mm: skip temporary folio allocation for anonymous UFFD missing faults
2026-11-19 18:00 [PATCH] mm: skip temporary folio allocation for anonymous UFFD missing faults Avi Weiss
@ 2026-11-19 18:22 ` Lorenzo Stoakes (ARM)
0 siblings, 0 replies; 2+ messages in thread
From: Lorenzo Stoakes (ARM) @ 2026-11-19 18:22 UTC (permalink / raw)
To: Avi Weiss
Cc: david, akpm, rppt, peterx, linux-mm, linux-kernel, liam, mhocko,
surenb, vbabka
+cc missing.
Please just stop. You sent 7 series across 6 different subsystems over 11
hours. You(r AI) even failed to cc the right people (added).
On Thu, Nov 19, 2026 at 08:00:15PM +0200, Avi Weiss wrote:
> An anonymous write fault on a VMA registered for missing userfaultfd
> faults currently allocates, zeroes and memcg-charges a folio before
> checking the PTE and handing the fault to userspace. The folio is then
> immediately discarded without ever being mapped.
>
> Check the PTE and address-space stability under the PTE lock before
> allocating a folio when missing faults are registered. Deliver the fault
> to userfaultfd without allocating and discarding a temporary folio.
>
> In a four-boot native B/P/B/P comparison, the mean UFFD fault median
> fell from 13,997.75 ns to 13,093.55 ns (-904.2 ns, -6.46%). A guest
> allocation trace saw 256 fewer allocations in the faulting task for
> 256 registered missing faults.
A .9us delta on a _uffd_ fault (nobody in their right mind puts it on a hot
path) with no standard deviation? Seriously?
>
> Note: this changes OOM ordering. In a fixed 18 MiB memcg test, the
> baseline OOM-killed only the faulting client before delivering a UFFD
> event. With this change, the event reached an unprotected shared handler,
> and its UFFDIO_COPY allocation could trigger an additional OOM kill of the
> handler. A separate surviving client consequently lost its handler and
> failed a page-content check after the UFFD registration was released.
> The same handler-death mechanism was reproducible on the unmodified
> (mainline) kernel when memory pressure was imposed after event delivery
> but before UFFDIO_COPY, although the fixed-limit outcomes differed.
This is both unreadable word salad and alludes to you causing possibly
breaking changes for a barely-distinguishable-from-noise fault time
improvement.
>
> Signed-off-by: Avi Weiss <thnkslprpt@gmail.com>
Please stop submitting patches, thanks.
"You are expected to understand and to be able to defend everything you
submit. If you are unable to do so, then do not submit the resulting
changes."
https://docs.kernel.org/process/generated-content.html
> ---
> mm/memory.c | 22 ++++++++++++++++------
> 1 file changed, 16 insertions(+), 6 deletions(-)
>
> diff --git a/mm/memory.c b/mm/memory.c
> index 0fb676d..63b39fa 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -5547,6 +5547,22 @@ static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
> ret = vmf_anon_prepare(vmf);
> if (ret)
> return ret;
> +
> + if (unlikely(userfaultfd_missing(vma))) {
> + vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
> + vmf->address, &vmf->ptl);
> + if (!vmf->pte)
> + goto unlock;
> + if (vmf_pte_changed(vmf)) {
> + update_mmu_tlb(vma, vmf->address, vmf->pte);
> + goto unlock;
> + }
> + ret = check_stable_address_space(vma->vm_mm);
> + if (ret)
> + goto unlock;
> + pte_unmap_unlock(vmf->pte, vmf->ptl);
> + return handle_userfault(vmf, VM_UFFD_MISSING);
> + }
> /* Returns NULL on OOM or ERR_PTR(-EAGAIN) if we must retry the fault */
> folio = alloc_anon_folio(vmf);
> if (IS_ERR(folio))
> @@ -5579,12 +5595,6 @@ static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
> if (ret)
> goto release;
>
> - /* Deliver the page fault to userland, check inside PT lock */
> - if (userfaultfd_missing(vma)) {
> - pte_unmap_unlock(vmf->pte, vmf->ptl);
> - folio_put(folio);
> - return handle_userfault(vmf, VM_UFFD_MISSING);
> - }
> map_anon_folio_pte_pf(folio, vmf->pte, vma, addr,
> vmf_orig_pte_uffd_wp(vmf));
> unlock:
> --
> 2.43.0
>
--
Cheers, Lorenzo
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2026-11-19 18:22 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-19 18:00 [PATCH] mm: skip temporary folio allocation for anonymous UFFD missing faults Avi Weiss
2026-11-19 18:22 ` Lorenzo Stoakes (ARM)
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®