Skip to contentSkip to conversation

mm: khugepaged: don't pass swap entry value to trace_mm_khugepaged_scan_file()

PATCH In review Vernon Yang 3 participants 5 messages Aug 11 – Aug 12

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 mm: khugepaged: don't pass swap entry value to trace_mm_khugepaged_scan_file()

From: Vernon Yang <[email protected]>

When the swap entries found exceed max_ptes_swap, the loop is left via break with folio still holding the xarray value that encodes the swap entry, not valid folio pointer.

That value is passed to trace_mm_khugepaged_scan_file(), which feeds it to folio_pfn(). On FLATMEM and SPARSEMEM_VMEMMAP, the page_to_pfn() is plain pointer arithmetic, so the trace event merely prints bogus scan_pfn. On classic SPARSEMEM, the page_to_pfn() reads page->flags, dereferencing the tiny encoded integer and oopsing khugepaged whenever the trace event is enabled.

So set folio to NULL before breaking out, the tracepoint maps NULL to scan_pfn of -1, just like exhausted scan naturally.

Fixes d41fd2016ed0 ("mm/khugepaged: add tracepoint to hpage_collapse_scan_file()")
---
 mm/khugepaged.c | 1 +
 1 file changed, 1 insertion(+)
mm/khugepaged.c +1 -0
index 617bca76db49..bc0d04c9162d 100644
@@ -2696,6 +2696,7 @@ static enum scan_result collapse_scan_file(struct mm_struct *mm,
 		if (xa_is_value(folio)) {
 			swap += 1 << xas_get_order(&xas);
 			if (swap > max_ptes_swap) {
+				folio = NULL;
 				result = SCAN_EXCEED_SWAP_PTE;
 				count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
 				break;

base-commit: 075b74841bd0065a3bda3440873c747938e69b68

Signature · 1 line
2.53.0
#1 lore raw
Quoted 28 lines · Vernon Yang

From: Vernon Yang <[email protected]>

When the swap entries found exceed max_ptes_swap, the loop is left via break with folio still holding the xarray value that encodes the swap entry, not valid folio pointer.

That value is passed to trace_mm_khugepaged_scan_file(), which feeds it to folio_pfn(). On FLATMEM and SPARSEMEM_VMEMMAP, the page_to_pfn() is plain pointer arithmetic, so the trace event merely prints bogus scan_pfn. On classic SPARSEMEM, the page_to_pfn() reads page->flags, dereferencing the tiny encoded integer and oopsing khugepaged whenever the trace event is enabled.

So set folio to NULL before breaking out, the tracepoint maps NULL to scan_pfn of -1, just like exhausted scan naturally.

Fixes d41fd2016ed0 ("mm/khugepaged: add tracepoint to hpage_collapse_scan_file()")
---
 mm/khugepaged.c | 1 +
 1 file changed, 1 insertion(+)
mm/khugepaged.c +0 -0
index 617bca76db49..bc0d04c9162d 100644

mm/khugepaged.c static enum scan_result collapse_scan_file(struct mm_struct *mm, · line 2702

Quoted 7 lines
mm/khugepaged.c +1 -0
@@ -2696,6 +2696,7 @@ static enum scan_result collapse_scan_file(struct mm_struct *mm,
 		if (xa_is_value(folio)) {
 			swap += 1 << xas_get_order(&xas);
 			if (swap > max_ptes_swap) {
+				folio = NULL;
 				result = SCAN_EXCEED_SWAP_PTE;
 				count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
 				break;

Yes, we'll do a folio_pfn(), and used to do a page_to_pfn().

Using the folio after dropping the reference is rather nasty.

Instead of passing the folio, should we just pass the pfn directly?

Signature · 3 lines
Cheers,

David
#2 lore raw
Quoted 39 lines · David Hildenbrand (Arm)
Quoted 28 lines · Vernon Yang

From: Vernon Yang <[email protected]>

When the swap entries found exceed max_ptes_swap, the loop is left via break with folio still holding the xarray value that encodes the swap entry, not valid folio pointer.

That value is passed to trace_mm_khugepaged_scan_file(), which feeds it to folio_pfn(). On FLATMEM and SPARSEMEM_VMEMMAP, the page_to_pfn() is plain pointer arithmetic, so the trace event merely prints bogus scan_pfn. On classic SPARSEMEM, the page_to_pfn() reads page->flags, dereferencing the tiny encoded integer and oopsing khugepaged whenever the trace event is enabled.

So set folio to NULL before breaking out, the tracepoint maps NULL to scan_pfn of -1, just like exhausted scan naturally.

Fixes d41fd2016ed0 ("mm/khugepaged: add tracepoint to hpage_collapse_scan_file()")
---
 mm/khugepaged.c | 1 +
 1 file changed, 1 insertion(+)
mm/khugepaged.c +0 -0
index 617bca76db49..bc0d04c9162d 100644

mm/khugepaged.c static enum scan_result collapse_scan_file(struct mm_struct *mm, · line 2702

Quoted 7 lines
mm/khugepaged.c +1 -0
@@ -2696,6 +2696,7 @@ static enum scan_result collapse_scan_file(struct mm_struct *mm,
 		if (xa_is_value(folio)) {
 			swap += 1 << xas_get_order(&xas);
 			if (swap > max_ptes_swap) {
+				folio = NULL;
 				result = SCAN_EXCEED_SWAP_PTE;
 				count_vm_event(THP_SCAN_EXCEED_SWAP_PTE);
 				break;

Yes, we'll do a folio_pfn(), and used to do a page_to_pfn().

Quoted 3 lines

Using the folio after dropping the reference is rather nasty.

Instead of passing the folio, should we just pass the pfn directly?

Yes, LGTM.

Would similar modifications like the following match the effect you want? If so, I'll make these changes in the next version.

mm/khugepaged.c +9 -5
index 617bca76db49..e7830761d3a2 100644
@@ -2683,6 +2683,7 @@ static enum scan_result collapse_scan_file(struct mm_struct *mm,
 	int present, swap;
 	int node = NUMA_NO_NODE;
 	enum scan_result result = SCAN_SUCCEED;
+	unsigned long pfn;
 
 	present = 0;
 	swap = 0;
@@ -2720,27 +2721,23 @@ static enum scan_result collapse_scan_file(struct mm_struct *mm,
 			 * PMD-sized THP implies that we can only try
 			 * retracting the PTE table.
 			 */
-			folio_put(folio);
 			break;
 		}
 
 		node = folio_nid(folio);
 		if (collapse_scan_abort(node, cc)) {
 			result = SCAN_SCAN_ABORT;
-			folio_put(folio);
 			break;
 		}
 		cc->node_load[node]++;
 
 		if (!folio_test_lru(folio)) {
 			result = SCAN_PAGE_LRU;
-			folio_put(folio);
 			break;
 		}
 
 		if (folio_expected_ref_count(folio) + 1 != folio_ref_count(folio)) {
 			result = SCAN_PAGE_COUNT;
-			folio_put(folio);
 			break;
 		}
 
@@ -2759,7 +2756,14 @@ static enum scan_result collapse_scan_file(struct mm_struct *mm,
 			cond_resched_rcu();
 		}
 	}
+	if (!folio || xa_is_value(folio)) {
+		pfn = -1;
+	} else {
+		pfn = folio_pfn(folio);
+		folio_put(folio);
+	}
 	rcu_read_unlock();
+
 	if (result == SCAN_PTE_MAPPED_HUGEPAGE)
 		cc->progress++;
 	else
@@ -2774,7 +2778,7 @@ static enum scan_result collapse_scan_file(struct mm_struct *mm,
 		}
 	}
 
-	trace_mm_khugepaged_scan_file(mm, folio, file, present, swap, result);
+	trace_mm_khugepaged_scan_file(mm, pfn, file, present, swap, result);
 	return result;
 }
 

--
Cheers,
Vernon

maintainer
#3 lore raw
In reply to Vernon Yang
Quoted 11 lines · On Tue, 11 Aug 2026 21:36:55 +0800 Vernon Yang <[email protected]> wrote:

When the swap entries found exceed max_ptes_swap, the loop is left via break with folio still holding the xarray value that encodes the swap entry, not valid folio pointer.

That value is passed to trace_mm_khugepaged_scan_file(), which feeds it to folio_pfn(). On FLATMEM and SPARSEMEM_VMEMMAP, the page_to_pfn() is plain pointer arithmetic, so the trace event merely prints bogus scan_pfn. On classic SPARSEMEM, the page_to_pfn() reads page->flags, dereferencing the tiny encoded integer and oopsing khugepaged whenever the trace event is enabled.

Quoted 4 lines

So set folio to NULL before breaking out, the tracepoint maps NULL to scan_pfn of -1, just like exhausted scan naturally.

Fixes d41fd2016ed0 ("mm/khugepaged: add tracepoint to hpage_collapse_scan_file()")

Added in 2022. Why so long - do people not use tracing?

Sashiko might have a found a couple of other tracing bugs in this code, which I suggest are on-topic for your patch:

	https://sashiko.dev/#/patchset/[email protected]

Also a possible bug mapping large folios which straddle i_size, which is a separate thing.

#4 lore raw
Quoted 13 lines · Andrew Morton
Quoted 11 lines · On Tue, 11 Aug 2026 21:36:55 +0800 Vernon Yang <[email protected]> wrote:

When the swap entries found exceed max_ptes_swap, the loop is left via break with folio still holding the xarray value that encodes the swap entry, not valid folio pointer.

That value is passed to trace_mm_khugepaged_scan_file(), which feeds it to folio_pfn(). On FLATMEM and SPARSEMEM_VMEMMAP, the page_to_pfn() is plain pointer arithmetic, so the trace event merely prints bogus scan_pfn. On classic SPARSEMEM, the page_to_pfn() reads page->flags, dereferencing the tiny encoded integer and oopsing khugepaged whenever the trace event is enabled.

Quoted 6 lines
Quoted 4 lines

So set folio to NULL before breaking out, the tracepoint maps NULL to scan_pfn of -1, just like exhausted scan naturally.

Fixes d41fd2016ed0 ("mm/khugepaged: add tracepoint to hpage_collapse_scan_file()")

Added in 2022. Why so long - do people not use tracing?

For most architectures, SPARSEMEM_VMEMMAP is the default, where page_to_pfn() is plain pointer arithmetic, merely printing bogus scan_pfn. Only on classic SPARSEMEM, the swap entry value cause khugepaged to oops.

Quoted 5 lines

Sashiko might have a found a couple of other tracing bugs in this code, which I suggest are on-topic for your patch:

	https://sashiko.dev/#/patchset/[email protected]

David also mentioned "Using the folio after dropping the reference", and the same issue exists in trace_mm_khugepaged_collapse_file()/trace_mm_khugepaged_scan_pmd(), which merely prints a overdue scan_pfn without oopsing khugepaged. However, this approach does have potential issues, just that they haven't been triggered yet. I can fix them together.

I will add PATCH#2 and PATCH#3 in the next version of the patchset to fix the other two trace_xxx() functions, rather than mixing them together into one large patch.

Quoted 2 lines

Also a possible bug mapping large folios which straddle i_size, which is a separate thing.

Yes, I will submit a separate fix patch later to resolve this issue.

--
Cheers,
Vernon