Re: [PATCH mm-new 1/1] mm/khugepaged: abort collapse scan on non-swap entries

From: Lance Yang

Date: Mon Sep 29 2025 - 06:39:34 EST




On 2025/9/29 18:29, David Hildenbrand wrote:
On 24.09.25 13:47, Lance Yang wrote:


On 2025/9/24 18:10, David Hildenbrand wrote:
On 24.09.25 12:02, Lance Yang wrote:
From: Lance Yang <lance.yang@xxxxxxxxx>

The existing check in hpage_collapse_scan_pmd() is specific to uffd-wp
markers. Other special markers (e.g., GUARD, POISONED) would not be
caught
early, leading to failures deeper in the swap-in logic.

hpage_collapse_scan_pmd()
   `- collapse_huge_page()
       `- __collapse_huge_page_swapin() -> fails!

As David suggested[1], this patch skips any such non-swap entries early.
If a special marker is found, the scan is aborted immediately with the
SCAN_PTE_NON_PRESENT result, as Lorenzo suggested[2], avoiding wasted
work.

Note that I suggested to skip all non-present entries except swap
entries, which includes migration entries, hwpoisoned entries etc.

Oops, I completely misunderstood your suggestion :(

It should be to handle all special non-present entries (migration,
hwpoison, markers), not just a specific type of marker ...

How about this version, which handles all non-swap entries as you
suggested?

diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index 7ab2d1a42df3..27f432e7f07c 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -1284,7 +1284,23 @@ static int hpage_collapse_scan_pmd(struct
mm_struct *mm,
          for (addr = start_addr, _pte = pte; _pte < pte + HPAGE_PMD_NR;
               _pte++, addr += PAGE_SIZE) {
                  pte_t pteval = ptep_get(_pte);
-               if (is_swap_pte(pteval)) {
+               if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
+                       ++none_or_zero;
+                       if (!userfaultfd_armed(vma) &&
+                           (!cc->is_khugepaged ||
+                            none_or_zero <= khugepaged_max_ptes_none)) {
+                               continue;
+                       } else {
+                               result = SCAN_EXCEED_NONE_PTE;
+                               count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
+                               goto out_unmap;
+                       }
+               } else if (!pte_present(pteval)) {
+                       if (non_swap_entry(pte_to_swp_entry(pteval))) {
+                               result = SCAN_PTE_NON_PRESENT;
+                               goto out_unmap;
+                       }
+
                          ++unmapped;
                          if (!cc->is_khugepaged ||
                              unmapped <= khugepaged_max_ptes_swap) {
@@ -1293,7 +1309,7 @@ static int hpage_collapse_scan_pmd(struct
mm_struct *mm,
                                   * enabled swap entries.  Please see
                                   * comment below for pte_uffd_wp().
                                   */
-                               if (pte_swp_uffd_wp_any(pteval)) {
+                               if (pte_swp_uffd_wp(pteval)) {
                                          result = SCAN_PTE_UFFD_WP;
                                          goto out_unmap;
                                  }
@@ -1304,18 +1320,6 @@ static int hpage_collapse_scan_pmd(struct
mm_struct *mm,
                                  goto out_unmap;
                          }
                  }
-               if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
-                       ++none_or_zero;
-                       if (!userfaultfd_armed(vma) &&
-                           (!cc->is_khugepaged ||
-                            none_or_zero <= khugepaged_max_ptes_none)) {
-                               continue;
-                       } else {
-                               result = SCAN_EXCEED_NONE_PTE;
-                               count_vm_event(THP_SCAN_EXCEED_NONE_PTE);
-                               goto out_unmap;
-                       }
-               }
                  if (pte_uffd_wp(pteval)) {

From a quick glimpse, this should work. And as raised, we might be able to unify later the scanning with the almost-duplicated code when we do the second scan.

Sounds good! Let's get this one merged first, and I'll send a follow-up
patch to unify the duplicated code as you suggested ;)