hugetlb: guarantee that COW faults for a process that called mmap(MAP_PRIVATE) on hugetlbfs will succeed

After patch 2 in this series, a process that successfully calls mmap() for
a MAP_PRIVATE mapping will be guaranteed to successfully fault until a
process calls fork().  At that point, the next write fault from the parent
could fail due to COW if the child still has a reference.

We only reserve pages for the parent but a copy must be made to avoid
leaking data from the parent to the child after fork().  Reserves could be
taken for both parent and child at fork time to guarantee faults but if
the mapping is large it is highly likely we will not have sufficient pages
for the reservation, and it is common to fork only to exec() immediatly
after.  A failure here would be very undesirable.

Note that the current behaviour of mainline with MAP_PRIVATE pages is
pretty bad.  The following situation is allowed to occur today.

1. Process calls mmap(MAP_PRIVATE)
2. Process calls mlock() to fault all pages and makes sure it succeeds
3. Process forks()
4. Process writes to MAP_PRIVATE mapping while child still exists
5. If the COW fails at this point, the process gets SIGKILLed even though it
   had taken care to ensure the pages existed

This patch improves the situation by guaranteeing the reliability of the
process that successfully calls mmap().  When the parent performs COW, it
will try to satisfy the allocation without using reserves.  If that fails
the parent will steal the page leaving any children without a page.
Faults from the child after that point will result in failure.  If the
child COW happens first, an attempt will be made to allocate the page
without reserves and the child will get SIGKILLed on failure.

To summarise the new behaviour:

1. If the original mapper performs COW on a private mapping with multiple
   references, it will attempt to allocate a hugepage from the pool or
   the buddy allocator without using the existing reserves. On fail, VMAs
   mapping the same area are traversed and the page being COW'd is unmapped
   where found. It will then steal the original page as the last mapper in
   the normal way.

2. The VMAs the pages were unmapped from are flagged to note that pages
   with data no longer exist. Future no-page faults on those VMAs will
   terminate the process as otherwise it would appear that data was corrupted.
   A warning is printed to the console that this situation occured.

2. If the child performs COW first, it will attempt to satisfy the COW
   from the pool if there are enough pages or via the buddy allocator if
   overcommit is allowed and the buddy allocator can satisfy the request. If
   it fails, the child will be killed.

If the pool is large enough, existing applications will not notice that
the reserves were a factor.  Existing applications depending on the
no-reserves been set are unlikely to exist as for much of the history of
hugetlbfs, pages were prefaulted at mmap(), allocating the pages at that
point or failing the mmap().

[npiggin@suse.de: fix CONFIG_HUGETLB=n build]
Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Acked-by: Adam Litke <agl@us.ibm.com>
Cc: Andy Whitcroft <apw@shadowen.org>
Cc: William Lee Irwin III <wli@holomorphy.com>
Cc: Hugh Dickins <hugh@veritas.com>
Cc: Nick Piggin <npiggin@suse.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

fs/hugetlbfs/inode.c

@@ -441,7 +441,7 @@ hugetlb_vmtruncate_list(struct prio_tree_root *root, pgoff_t pgoff)
 			v_offset = 0;
 
 		__unmap_hugepage_range(vma,
-				vma->vm_start + v_offset, vma->vm_end);
+				vma->vm_start + v_offset, vma->vm_end, NULL);
 	}
 }
 

include/linux/hugetlb.h

@@ -23,8 +23,10 @@ int hugetlb_overcommit_handler(struct ctl_table *, int, struct file *, void __us
 int hugetlb_treat_movable_handler(struct ctl_table *, int, struct file *, void __user *, size_t *, loff_t *);
 int copy_hugetlb_page_range(struct mm_struct *, struct mm_struct *, struct vm_area_struct *);
 int follow_hugetlb_page(struct mm_struct *, struct vm_area_struct *, struct page **, struct vm_area_struct **, unsigned long *, int *, int, int);
-void unmap_hugepage_range(struct vm_area_struct *, unsigned long, unsigned long);
+void unmap_hugepage_range(struct vm_area_struct *,
-void __unmap_hugepage_range(struct vm_area_struct *, unsigned long, unsigned long);
+			unsigned long, unsigned long, struct page *);
+void __unmap_hugepage_range(struct vm_area_struct *,
+			unsigned long, unsigned long, struct page *);
 int hugetlb_prefault(struct address_space *, struct vm_area_struct *);
 int hugetlb_report_meminfo(char *);
 int hugetlb_report_node_meminfo(int, char *);
@@ -74,7 +76,7 @@ static inline unsigned long hugetlb_total_pages(void)
 #define follow_huge_addr(mm, addr, write)	ERR_PTR(-EINVAL)
 #define copy_hugetlb_page_range(src, dst, vma)	({ BUG(); 0; })
 #define hugetlb_prefault(mapping, vma)		({ BUG(); 0; })
-#define unmap_hugepage_range(vma, start, end)	BUG()
+#define unmap_hugepage_range(vma, start, end, page)	BUG()
 #define hugetlb_report_meminfo(buf)		0
 #define hugetlb_report_node_meminfo(n, buf)	0
 #define follow_huge_pmd(mm, addr, pmd, write)	NULL

mm/hugetlb.c

@@ -40,6 +40,9 @@ static int hugetlb_next_nid;
  */
 static DEFINE_SPINLOCK(hugetlb_lock);
 
+#define HPAGE_RESV_OWNER    (1UL << (BITS_PER_LONG - 1))
+#define HPAGE_RESV_UNMAPPED (1UL << (BITS_PER_LONG - 2))
+#define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
 /*
  * These helpers are used to track how many pages are reserved for
  * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
@@ -54,17 +57,32 @@ static unsigned long vma_resv_huge_pages(struct vm_area_struct *vma)
 {
 	VM_BUG_ON(!is_vm_hugetlb_page(vma));
 	if (!(vma->vm_flags & VM_SHARED))
-		return (unsigned long)vma->vm_private_data;
+		return (unsigned long)vma->vm_private_data & ~HPAGE_RESV_MASK;
 	return 0;
 }
 
 static void set_vma_resv_huge_pages(struct vm_area_struct *vma,
 							unsigned long reserve)
 {
+	unsigned long flags;
 	VM_BUG_ON(!is_vm_hugetlb_page(vma));
 	VM_BUG_ON(vma->vm_flags & VM_SHARED);
 
-	vma->vm_private_data = (void *)reserve;
+	flags = (unsigned long)vma->vm_private_data & HPAGE_RESV_MASK;
+	vma->vm_private_data = (void *)(reserve | flags);
+}
+
+static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
+{
+	unsigned long reserveflags = (unsigned long)vma->vm_private_data;
+	VM_BUG_ON(!is_vm_hugetlb_page(vma));
+	vma->vm_private_data = (void *)(reserveflags | flags);
+}
+
+static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
+{
+	VM_BUG_ON(!is_vm_hugetlb_page(vma));
+	return ((unsigned long)vma->vm_private_data & flag) != 0;
 }
 
 /* Decrement the reserved pages in the hugepage pool by one */
@@ -78,14 +96,18 @@ static void decrement_hugepage_resv_vma(struct vm_area_struct *vma)
 		 * Only the process that called mmap() has reserves for
 		 * private mappings.
 		 */
-		if (vma_resv_huge_pages(vma)) {
+		if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
+			unsigned long flags, reserve;
 			resv_huge_pages--;
+			flags = (unsigned long)vma->vm_private_data &
+							HPAGE_RESV_MASK;
 			reserve = (unsigned long)vma->vm_private_data - 1;
-			vma->vm_private_data = (void *)reserve;
+			vma->vm_private_data = (void *)(reserve | flags);
 		}
 	}
 }
 
+/* Reset counters to 0 and clear all HPAGE_RESV_* flags */
 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
 {
 	VM_BUG_ON(!is_vm_hugetlb_page(vma));
@@ -153,7 +175,7 @@ static struct page *dequeue_huge_page(void)
 }
 
 static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
-				unsigned long address)
+				unsigned long address, int avoid_reserve)
 {
 	int nid;
 	struct page *page = NULL;
@@ -173,6 +195,10 @@ static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
 			free_huge_pages - resv_huge_pages == 0)
 		return NULL;
 
+	/* If reserves cannot be used, ensure enough pages are in the pool */
+	if (avoid_reserve && free_huge_pages - resv_huge_pages == 0)
+		return NULL;
+
 	for_each_zone_zonelist_nodemask(zone, z, zonelist,
 						MAX_NR_ZONES - 1, nodemask) {
 		nid = zone_to_nid(zone);
@@ -183,6 +209,8 @@ static struct page *dequeue_huge_page_vma(struct vm_area_struct *vma,
 			list_del(&page->lru);
 			free_huge_pages--;
 			free_huge_pages_node[nid]--;
+
+			if (!avoid_reserve)
 				decrement_hugepage_resv_vma(vma);
 
 			break;
@@ -534,7 +562,7 @@ static void return_unused_surplus_pages(unsigned long unused_resv_pages)
 }
 
 static struct page *alloc_huge_page(struct vm_area_struct *vma,
-				    unsigned long addr)
+				    unsigned long addr, int avoid_reserve)
 {
 	struct page *page;
 	struct address_space *mapping = vma->vm_file->f_mapping;
@@ -546,14 +574,15 @@ static struct page *alloc_huge_page(struct vm_area_struct *vma,
 	 * will not have accounted against quota. Check that the quota can be
 	 * made before satisfying the allocation
 	 */
-	if (!vma_has_private_reserves(vma)) {
+	if (!(vma->vm_flags & VM_SHARED) &&
+			!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
 		chg = 1;
 		if (hugetlb_get_quota(inode->i_mapping, chg))
 			return ERR_PTR(-ENOSPC);
 	}
 
 	spin_lock(&hugetlb_lock);
-	page = dequeue_huge_page_vma(vma, addr);
+	page = dequeue_huge_page_vma(vma, addr, avoid_reserve);
 	spin_unlock(&hugetlb_lock);
 
 	if (!page) {
@@ -909,7 +938,7 @@ nomem:
 }
 
 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
-			    unsigned long end)
+			    unsigned long end, struct page *ref_page)
 {
 	struct mm_struct *mm = vma->vm_mm;
 	unsigned long address;
@@ -937,6 +966,27 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 		if (huge_pmd_unshare(mm, &address, ptep))
 			continue;
 
+		/*
+		 * If a reference page is supplied, it is because a specific
+		 * page is being unmapped, not a range. Ensure the page we
+		 * are about to unmap is the actual page of interest.
+		 */
+		if (ref_page) {
+			pte = huge_ptep_get(ptep);
+			if (huge_pte_none(pte))
+				continue;
+			page = pte_page(pte);
+			if (page != ref_page)
+				continue;
+
+			/*
+			 * Mark the VMA as having unmapped its page so that
+			 * future faults in this VMA will fail rather than
+			 * looking like data was lost
+			 */
+			set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
+		}
+
 		pte = huge_ptep_get_and_clear(mm, address, ptep);
 		if (huge_pte_none(pte))
 			continue;
@@ -955,7 +1005,7 @@ void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 }
 
 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
-			  unsigned long end)
+			  unsigned long end, struct page *ref_page)
 {
 	/*
 	 * It is undesirable to test vma->vm_file as it should be non-null
@@ -967,19 +1017,68 @@ void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
 	 */
 	if (vma->vm_file) {
 		spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
-		__unmap_hugepage_range(vma, start, end);
+		__unmap_hugepage_range(vma, start, end, ref_page);
 		spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
 	}
 }
 
+/*
+ * This is called when the original mapper is failing to COW a MAP_PRIVATE
+ * mappping it owns the reserve page for. The intention is to unmap the page
+ * from other VMAs and let the children be SIGKILLed if they are faulting the
+ * same region.
+ */
+int unmap_ref_private(struct mm_struct *mm,
+					struct vm_area_struct *vma,
+					struct page *page,
+					unsigned long address)
+{
+	struct vm_area_struct *iter_vma;
+	struct address_space *mapping;
+	struct prio_tree_iter iter;
+	pgoff_t pgoff;
+
+	/*
+	 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
+	 * from page cache lookup which is in HPAGE_SIZE units.
+	 */
+	address = address & huge_page_mask(hstate_vma(vma));
+	pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
+		+ (vma->vm_pgoff >> PAGE_SHIFT);
+	mapping = (struct address_space *)page_private(page);
+
+	vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
+		/* Do not unmap the current VMA */
+		if (iter_vma == vma)
+			continue;
+
+		/*
+		 * Unmap the page from other VMAs without their own reserves.
+		 * They get marked to be SIGKILLed if they fault in these
+		 * areas. This is because a future no-page fault on this VMA
+		 * could insert a zeroed page instead of the data existing
+		 * from the time of fork. This would look like data corruption
+		 */
+		if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
+			unmap_hugepage_range(iter_vma,
+				address, address + HPAGE_SIZE,
+				page);
+	}
+
+	return 1;
+}
+
 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
-			unsigned long address, pte_t *ptep, pte_t pte)
+			unsigned long address, pte_t *ptep, pte_t pte,
+			struct page *pagecache_page)
 {
 	struct page *old_page, *new_page;
 	int avoidcopy;
+	int outside_reserve = 0;
 
 	old_page = pte_page(pte);
 
+retry_avoidcopy:
 	/* If no-one else is actually using this page, avoid the copy
 	 * and just make the page writable */
 	avoidcopy = (page_count(old_page) == 1);
@@ -988,11 +1087,43 @@ static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 		return 0;
 	}
 
+	/*
+	 * If the process that created a MAP_PRIVATE mapping is about to
+	 * perform a COW due to a shared page count, attempt to satisfy
+	 * the allocation without using the existing reserves. The pagecache
+	 * page is used to determine if the reserve at this address was
+	 * consumed or not. If reserves were used, a partial faulted mapping
+	 * at the time of fork() could consume its reserves on COW instead
+	 * of the full address range.
+	 */
+	if (!(vma->vm_flags & VM_SHARED) &&
+			is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
+			old_page != pagecache_page)
+		outside_reserve = 1;
+
 	page_cache_get(old_page);
-	new_page = alloc_huge_page(vma, address);
+	new_page = alloc_huge_page(vma, address, outside_reserve);
 
 	if (IS_ERR(new_page)) {
 		page_cache_release(old_page);
+
+		/*
+		 * If a process owning a MAP_PRIVATE mapping fails to COW,
+		 * it is due to references held by a child and an insufficient
+		 * huge page pool. To guarantee the original mappers
+		 * reliability, unmap the page from child processes. The child
+		 * may get SIGKILLed if it later faults.
+		 */
+		if (outside_reserve) {
+			BUG_ON(huge_pte_none(pte));
+			if (unmap_ref_private(mm, vma, old_page, address)) {
+				BUG_ON(page_count(old_page) != 1);
+				BUG_ON(huge_pte_none(pte));
+				goto retry_avoidcopy;
+			}
+			WARN_ON_ONCE(1);
+		}
+
 		return -PTR_ERR(new_page);
 	}
 
@@ -1015,6 +1146,20 @@ static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
 	return 0;
 }
 
+/* Return the pagecache page at a given address within a VMA */
+static struct page *hugetlbfs_pagecache_page(struct vm_area_struct *vma,
+			unsigned long address)
+{
+	struct address_space *mapping;
+	unsigned long idx;
+
+	mapping = vma->vm_file->f_mapping;
+	idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
+		+ (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
+
+	return find_lock_page(mapping, idx);
+}
+
 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 			unsigned long address, pte_t *ptep, int write_access)
 {
@@ -1025,6 +1170,18 @@ static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
 	struct address_space *mapping;
 	pte_t new_pte;
 
+	/*
+	 * Currently, we are forced to kill the process in the event the
+	 * original mapper has unmapped pages from the child due to a failed
+	 * COW. Warn that such a situation has occured as it may not be obvious
+	 */
+	if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
+		printk(KERN_WARNING
+			"PID %d killed due to inadequate hugepage pool\n",
+			current->pid);
+		return ret;
+	}
+
 	mapping = vma->vm_file->f_mapping;
 	idx = ((address - vma->vm_start) >> HPAGE_SHIFT)
 		+ (vma->vm_pgoff >> (HPAGE_SHIFT - PAGE_SHIFT));
@@ -1039,7 +1196,7 @@ retry:
 		size = i_size_read(mapping->host) >> HPAGE_SHIFT;
 		if (idx >= size)
 			goto out;
-		page = alloc_huge_page(vma, address);
+		page = alloc_huge_page(vma, address, 0);
 		if (IS_ERR(page)) {
 			ret = -PTR_ERR(page);
 			goto out;
@@ -1081,7 +1238,7 @@ retry:
 
 	if (write_access && !(vma->vm_flags & VM_SHARED)) {
 		/* Optimization, do the COW without a second fault */
-		ret = hugetlb_cow(mm, vma, address, ptep, new_pte);
+		ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
 	}
 
 	spin_unlock(&mm->page_table_lock);
@@ -1126,8 +1283,15 @@ int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
 	spin_lock(&mm->page_table_lock);
 	/* Check for a racing update before calling hugetlb_cow */
 	if (likely(pte_same(entry, huge_ptep_get(ptep))))
-		if (write_access && !pte_write(entry))
+		if (write_access && !pte_write(entry)) {
-			ret = hugetlb_cow(mm, vma, address, ptep, entry);
+			struct page *page;
+			page = hugetlbfs_pagecache_page(vma, address);
+			ret = hugetlb_cow(mm, vma, address, ptep, entry, page);
+			if (page) {
+				unlock_page(page);
+				put_page(page);
+			}
+		}
 	spin_unlock(&mm->page_table_lock);
 	mutex_unlock(&hugetlb_instantiation_mutex);
 
@@ -1371,6 +1535,7 @@ int hugetlb_reserve_pages(struct inode *inode,
 	else {
 		chg = to - from;
 		set_vma_resv_huge_pages(vma, chg);
+		set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
 	}
 
 	if (chg < 0)

mm/memory.c

@@ -901,7 +901,7 @@ unsigned long unmap_vmas(struct mmu_gather **tlbp,
 			}
 
 			if (unlikely(is_vm_hugetlb_page(vma))) {
-				unmap_hugepage_range(vma, start, end);
+				unmap_hugepage_range(vma, start, end, NULL);
 				zap_work -= (end - start) /
 						(HPAGE_SIZE / PAGE_SIZE);
 				start = end;