Skip to contentSkip to conversation

mm/hugetlb: Fix calculation of adjust_range_if_pmd_sharing_possible

PATCH Applied Peter Xu 2 participants 4 messages Jul 30, 2020 – Jul 31, 2020

Subsystems (2)
Memory management HugeTLB
lore
Share & tools
lore

Peter Xu

author
#0 lore raw

PATCH v2 1/1 mm/hugetlb: Fix calculation of adjust_range_if_pmd_sharing_possible

This is found by code observation only.

Firstly, the worst case scenario should assume the whole range was covered by pmd sharing. The old algorithm might not work as expected for ranges like (1g-2m, 1g+2m), where the adjusted range should be (0, 1g+2m) but the expected range should be (0, 2g).

Since at it, remove the loop since it should not be required. With that, the new code should be faster too when the invalidating range is huge.

CC Andrea Arcangeli <[email protected]>
CC Mike Kravetz <[email protected]>
CC Andrew Morton <[email protected]>
CC Matthew Wilcox <[email protected]>
CC [email protected]
CC [email protected]
Signed-off-by Peter Xu <[email protected]>

---
v2:
- use min/max instead of custom MIN/MAX [Matthew]

---
 mm/hugetlb.c | 24 ++++++++++--------------
 1 file changed, 10 insertions(+), 14 deletions(-)
mm/hugetlb.c +10 -14
index 4645f1441d32..7332f3c4b8ec 100644
@@ -5321,25 +5321,21 @@ static bool vma_shareable(struct vm_area_struct *vma, unsigned long addr)
 void adjust_range_if_pmd_sharing_possible(struct vm_area_struct *vma,
 				unsigned long *start, unsigned long *end)
 {
-	unsigned long check_addr;
+	unsigned long a_start, a_end;
 
 	if (!(vma->vm_flags & VM_MAYSHARE))
 		return;
 
-	for (check_addr = *start; check_addr < *end; check_addr += PUD_SIZE) {
-		unsigned long a_start = check_addr & PUD_MASK;
-		unsigned long a_end = a_start + PUD_SIZE;
+	/* Extend the range to be PUD aligned for a worst case scenario */
+	a_start = ALIGN_DOWN(*start, PUD_SIZE);
+	a_end = ALIGN(*end, PUD_SIZE);
 
-		/*
-		 * If sharing is possible, adjust start/end if necessary.
-		 */
-		if (range_in_vma(vma, a_start, a_end)) {
-			if (a_start < *start)
-				*start = a_start;
-			if (a_end > *end)
-				*end = a_end;
-		}
-	}
+	/*
+	 * Intersect the range with the vma range, since pmd sharing won't be
+	 * across vma after all
+	 */
+	*start = max(vma->vm_start, a_start);
+	*end = min(vma->vm_end, a_end);
 }
 
 /*
Signature · 1 line
2.26.2
#1 lore raw
Quoted 9 lines · Peter Xu

This is found by code observation only.

Firstly, the worst case scenario should assume the whole range was covered by pmd sharing. The old algorithm might not work as expected for ranges like (1g-2m, 1g+2m), where the adjusted range should be (0, 1g+2m) but the expected range should be (0, 2g).

Since at it, remove the loop since it should not be required. With that, the new code should be faster too when the invalidating range is huge.

Thanks Peter!

That is certainly much simpler than the loop in current code. You say there are instances where old code 'might not work' for ranges like (1g-2m, 1g+2m). Not sure I understand what you mean by adjusted and expected ranges in the message. Both are possible 'adjusted' ranges depending on vma size.

Just trying to figure out if there is an actual problem in the existing code that needs to be fixed in stable. I think the existing code is correct, just inefficient.

Signature · 1 line
Mike Kravetz
author
#2 lore raw

Hi, Mike,

Quoted 13 lines · Mike Kravetz
Quoted 9 lines · Peter Xu

This is found by code observation only.

Firstly, the worst case scenario should assume the whole range was covered by pmd sharing. The old algorithm might not work as expected for ranges like (1g-2m, 1g+2m), where the adjusted range should be (0, 1g+2m) but the expected range should be (0, 2g).

Since at it, remove the loop since it should not be required. With that, the new code should be faster too when the invalidating range is huge.

Thanks Peter!

Quoted 8 lines

That is certainly much simpler than the loop in current code. You say there are instances where old code 'might not work' for ranges like (1g-2m, 1g+2m). Not sure I understand what you mean by adjusted and expected ranges in the message. Both are possible 'adjusted' ranges depending on vma size.

Just trying to figure out if there is an actual problem in the existing code that needs to be fixed in stable. I think the existing code is correct, just inefficient.

Thanks for the quick review!

I'm not sure whether that will cause a real problem, but iiuc in my previous example of (1g-2m, 1g+2m) in the commit message, the old code will extend the range to (0, 1g+2m). In this case, if unluckily the (1g, 2g) range is a pud with shared pmd, then imho we face the risk of partial tlb flushing with the old code, because it will only flush tlb for range (0, 1g+2m) but not (0, 2g). If that's the case, maybe it worths cc stable.

Anyway, I'd like to double confirm with you in case I missed something.

Thanks,

Signature · 1 line
Peter Xu
Reviewed-by
#3 lore raw
Quoted 27 lines · Peter Xu

Hi, Mike,

Quoted 13 lines · Mike Kravetz
Quoted 9 lines · Peter Xu

This is found by code observation only.

Firstly, the worst case scenario should assume the whole range was covered by pmd sharing. The old algorithm might not work as expected for ranges like (1g-2m, 1g+2m), where the adjusted range should be (0, 1g+2m) but the expected range should be (0, 2g).

Since at it, remove the loop since it should not be required. With that, the new code should be faster too when the invalidating range is huge.

Thanks Peter!

Quoted 8 lines

That is certainly much simpler than the loop in current code. You say there are instances where old code 'might not work' for ranges like (1g-2m, 1g+2m). Not sure I understand what you mean by adjusted and expected ranges in the message. Both are possible 'adjusted' ranges depending on vma size.

Just trying to figure out if there is an actual problem in the existing code that needs to be fixed in stable. I think the existing code is correct, just inefficient.

Thanks for the quick review!

Quoted 8 lines

I'm not sure whether that will cause a real problem, but iiuc in my previous example of (1g-2m, 1g+2m) in the commit message, the old code will extend the range to (0, 1g+2m). In this case, if unluckily the (1g, 2g) range is a pud with shared pmd, then imho we face the risk of partial tlb flushing with the old code, because it will only flush tlb for range (0, 1g+2m) but not (0, 2g). If that's the case, maybe it worths cc stable.

Anyway, I'd like to double confirm with you in case I missed something.

You are correct. With range (1g-2m, 1g+2m) within a vma (0, 2g) the existing code will only adjust to (0, 1g+2m) which is incorrect.

We should cc stable. The original reason for adjusting the range was to prevent data corruption (getting wrong page). Since the range is not always adjusted correctly, the potential for corruption still exists.

However, I am fairly confident that adjust_range_if_pmd_sharing_possible is only gong to be called in two cases:
1) for a single page
2) for range == entire vma
In those cases, the current code should produce the correct results.

To be safe, let's just cc stable.

Also,

Fixes 017b1660df89 ("mm: migration: fix migration of huge PMD shared pages")
Reviewed-by Mike Kravetz <[email protected]>
Signature · 1 line
Mike Kravetz