mm, thp: fix collapsing of hugepages on madvise

If an anonymous mapping is not allowed to fault thp memory and then madvise(MADV_HUGEPAGE) is used after fault, khugepaged will never collapse this memory into thp memory. This occurs because the madvise(2) handler for thp, hugepage_madvise(), clears VM_NOHUGEPAGE on the stack and it isn't stored in vma->vm_flags until the final action of madvise_behavior(). This causes the khugepaged_enter_vma_merge() to be a no-op in hugepage_madvise() when the vma had previously had VM_NOHUGEPAGE set. Fix this by passing the correct vma flags to the khugepaged mm slot handler. There's no chance khugepaged can run on this vma until after madvise_behavior() returns since we hold mm->mmap_sem. It would be possible to clear VM_NOHUGEPAGE directly from vma->vm_flags in hugepage_advise(), but I didn't want to introduce special case behavior into madvise_behavior(). I think it's best to just let it always set vma->vm_flags itself.

David Rientjes committed 12 years ago

Commit 6d50e601 parent 47f29df3 files · +20 −16Browse files

Unified · Split

3 files

include/linux/khugepaged.h +10 −7
mm/huge_memory.c +6 −5
mm/mmap.c +4 −4

include/linux/khugepaged.h+10−7

@@ -6,7 +6,8 @@ · 5 unmodified lines
6	6	#ifdef CONFIG_TRANSPARENT_HUGEPAGE
7	7	extern int __khugepaged_enter(struct mm_struct *mm);
8	8	extern void __khugepaged_exit(struct mm_struct *mm);
9		−extern int khugepaged_enter_vma_merge(struct vm_area_struct *vma);
	9	+extern int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
	10	+ unsigned long vm_flags);
10	11
11	12	#define khugepaged_enabled() \
12	13	(transparent_hugepage_flags & \
@@ -35,13 +36,13 @@ static inline void khugepaged_exit(struct mm_struct *mm) · 22 unmodified lines
35	36	__khugepaged_exit(mm);
36	37	}
37	38
38		−static inline int khugepaged_enter(struct vm_area_struct *vma)
	39	+static inline int khugepaged_enter(struct vm_area_struct *vma,
	40	+ unsigned long vm_flags)
39	41	{
40	42	if (!test_bit(MMF_VM_HUGEPAGE, &vma->vm_mm->flags))
41	43	if ((khugepaged_always() \|\|
42		− (khugepaged_req_madv() &&
43		− vma->vm_flags & VM_HUGEPAGE)) &&
44		− !(vma->vm_flags & VM_NOHUGEPAGE))
	44	+ (khugepaged_req_madv() && (vm_flags & VM_HUGEPAGE))) &&
	45	+ !(vm_flags & VM_NOHUGEPAGE))
45	46	if (__khugepaged_enter(vma->vm_mm))
46	47	return -ENOMEM;
47	48	return 0;
@@ -54,11 +55,13 @@ static inline int khugepaged_fork(struct mm_struct mm, struct mm_struct oldmm) · 6 unmodified lines
54	55	static inline void khugepaged_exit(struct mm_struct *mm)
55	56	{
56	57	}
57		−static inline int khugepaged_enter(struct vm_area_struct *vma)
	58	+static inline int khugepaged_enter(struct vm_area_struct *vma,
	59	+ unsigned long vm_flags)
58	60	{
59	61	return 0;
60	62	}
61		−static inline int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
	63	+static inline int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
	64	+ unsigned long vm_flags)
62	65	{
63	66	return 0;
64	67	}

mm/huge_memory.c+6−5

@@ -803,7 +803,7 @@ int do_huge_pmd_anonymous_page(struct mm_struct mm, struct vm_area_struct vma, · 802 unmodified lines
803	803	return VM_FAULT_FALLBACK;
804	804	if (unlikely(anon_vma_prepare(vma)))
805	805	return VM_FAULT_OOM;
806		− if (unlikely(khugepaged_enter(vma)))
	806	+ if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
807	807	return VM_FAULT_OOM;
808	808	if (!(flags & FAULT_FLAG_WRITE) &&
809	809	transparent_hugepage_use_zero_page()) {
@@ -1970,7 +1970,7 @@ int hugepage_madvise(struct vm_area_struct *vma, · 1160 unmodified lines
1970	1970	* register it here without waiting a page fault that
1971	1971	* may not happen any time soon.
1972	1972	*/
1973		− if (unlikely(khugepaged_enter_vma_merge(vma)))
	1973	+ if (unlikely(khugepaged_enter_vma_merge(vma, *vm_flags)))
1974	1974	return -ENOMEM;
1975	1975	break;
1976	1976	case MADV_NOHUGEPAGE:
@@ -2071,7 +2071,8 @@ int __khugepaged_enter(struct mm_struct *mm) · 94 unmodified lines
2071	2071	return 0;
2072	2072	}
2073	2073
2074		−int khugepaged_enter_vma_merge(struct vm_area_struct *vma)
	2074	+int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
	2075	+ unsigned long vm_flags)
2075	2076	{
2076	2077	unsigned long hstart, hend;
2077	2078	if (!vma->anon_vma)
@@ -2083,11 +2084,11 @@ int khugepaged_enter_vma_merge(struct vm_area_struct *vma) · 5 unmodified lines
2083	2084	if (vma->vm_ops)
2084	2085	/* khugepaged not yet working on file or special mappings */
2085	2086	return 0;
2086		− VM_BUG_ON_VMA(vma->vm_flags & VM_NO_THP, vma);
	2087	+ VM_BUG_ON_VMA(vm_flags & VM_NO_THP, vma);
2087	2088	hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
2088	2089	hend = vma->vm_end & HPAGE_PMD_MASK;
2089	2090	if (hstart < hend)
2090		− return khugepaged_enter(vma);
	2091	+ return khugepaged_enter(vma, vm_flags);
2091	2092	return 0;
2092	2093	}
2093	2094

mm/mmap.c+4−4

@@ -1080,7 +1080,7 @@ struct vm_area_struct vma_merge(struct mm_struct mm, · 1079 unmodified lines
1080	1080	end, prev->vm_pgoff, NULL);
1081	1081	if (err)
1082	1082	return NULL;
1083		− khugepaged_enter_vma_merge(prev);
	1083	+ khugepaged_enter_vma_merge(prev, vm_flags);
1084	1084	return prev;
1085	1085	}
1086	1086
@@ -1099,7 +1099,7 @@ struct vm_area_struct vma_merge(struct mm_struct mm, · 12 unmodified lines
1099	1099	next->vm_pgoff - pglen, NULL);
1100	1100	if (err)
1101	1101	return NULL;
1102		− khugepaged_enter_vma_merge(area);
	1102	+ khugepaged_enter_vma_merge(area, vm_flags);
1103	1103	return area;
1104	1104	}
1105	1105
@@ -2208,7 +2208,7 @@ int expand_upwards(struct vm_area_struct *vma, unsigned long address) · 1102 unmodified lines
2208	2208	}
2209	2209	}
2210	2210	vma_unlock_anon_vma(vma);
2211		− khugepaged_enter_vma_merge(vma);
	2211	+ khugepaged_enter_vma_merge(vma, vma->vm_flags);
2212	2212	validate_mm(vma->vm_mm);
2213	2213	return error;
2214	2214	}
@@ -2277,7 +2277,7 @@ int expand_downwards(struct vm_area_struct *vma, · 62 unmodified lines
2277	2277	}
2278	2278	}
2279	2279	vma_unlock_anon_vma(vma);
2280		− khugepaged_enter_vma_merge(vma);
	2280	+ khugepaged_enter_vma_merge(vma, vma->vm_flags);
2281	2281	validate_mm(vma->vm_mm);
2282	2282	return error;
2283	2283	}