[PATCH v2] sparse: Track the boundaries of memory sections for accurate checks

[KarimAllah Ahmed]

When sparse memory model is used an array of memory sections is created to
track each block of contiguous physical pages. Each element of this array
contains PAGES_PER_SECTION pages. During the creation of this array the actual
boundaries of the memory block is lost, so the whole block is either considered
as present or not.

pfn_valid() in the sparse memory configuration checks which memory sections the
pfn belongs to then checks whether it's present or not. This yields sub-optimal
results when the available memory doesn't cover the whole memory section,
because pfn_valid will return 'true' even for the unavailable pfns at the
boundaries of the memory section.

If pfn_valid() returns 'true' this means that this is a valid RAM page and
that it is controlled by the kernel (there's a 'struct page' backing it) which
is not the case if this pfn happens to be unavailable and at the boundaries of
the memory section and given the pattern of using pfn_valid just before
accessing the 'struct page' (through pfn_to_page) which can lead to a lot of
surprises.

For example this hunk of code in '__ioremap_check_ram':

	if (pfn_valid(start_pfn + i) &&
		!PageReserved(pfn_to_page(start_pfn + i)))
		return 1;

which can return '1' even for a pfn that's not valid!

or this other hunk (which is almost the same pattern) in 'kvm_is_reserved_pfn':

	if (pfn_valid(pfn))
		return PageReserved(pfn_to_page(pfn));

which can return false for the same reason (which will trigger a BUG_ON at the
call-site).

Using 'mem=' kernel parameter will have the same effect on pfn_valid() because
even though the memory at the memory section boundary can be RAM, it's not
valid because there's no 'struct page' for it.

Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
Cc: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
Cc: Vlastimil Babka <vbabka@xxxxxxx>
Cc: Michal Hocko <mhocko@xxxxxxxx>
Cc: Johannes Weiner <hannes@xxxxxxxxxxx>
Cc: Yaowei Bai <baiyaowei@xxxxxxxxxxxxxxxxxxxx>
Cc: Dan Williams <dan.j.williams@xxxxxxxxx>
Cc: Joe Perches <joe@xxxxxxxxxxx>
Cc: Tejun Heo <tj@xxxxxxxxxx>
Cc: Anthony Liguori <aliguori@xxxxxxxxxx>
Cc: linux-mm@xxxxxxxxx
Cc: linux-kernel@xxxxxxxxxxxxxxx
Signed-off-by: KarimAllah Ahmed <karahmed@xxxxxxxxx>
Signed-off-by: Jan H. SchÃnherr <jschoenh@xxxxxxxxx>

---
v2: A little bit more verbose commit message to explain why 'sub-optimal'
    results can actually cause problems.
---
 include/linux/mmzone.h | 22 ++++++++++++++++------
 mm/sparse.c            | 37 ++++++++++++++++++++++++++++++++++++-
 2 files changed, 52 insertions(+), 7 deletions(-)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 02069c2..f76a0e1 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -1067,8 +1067,12 @@ struct mem_section {
 	 * section. (see page_ext.h about this.)
 	 */
 	struct page_ext *page_ext;
-	unsigned long pad;
+	unsigned long pad[3];
 #endif
+
+	unsigned long first_pfn;
+	unsigned long last_pfn;
+
 	/*
 	 * WARNING: mem_section must be a power-of-2 in size for the
 	 * calculation and use of SECTION_ROOT_MASK to make sense.
@@ -1140,23 +1144,29 @@ static inline int valid_section_nr(unsigned long nr)
 
 static inline struct mem_section *__pfn_to_section(unsigned long pfn)
 {
+	if (pfn_to_section_nr(pfn) >= NR_MEM_SECTIONS)
+		return NULL;
+
 	return __nr_to_section(pfn_to_section_nr(pfn));
 }
 
 #ifndef CONFIG_HAVE_ARCH_PFN_VALID
 static inline int pfn_valid(unsigned long pfn)
 {
-	if (pfn_to_section_nr(pfn) >= NR_MEM_SECTIONS)
+	struct mem_section *ms;
+
+	ms = __pfn_to_section(pfn);
+
+	if (ms && !(ms->first_pfn <= pfn && ms->last_pfn >= pfn))
 		return 0;
-	return valid_section(__nr_to_section(pfn_to_section_nr(pfn)));
+
+	return valid_section(ms);
 }
 #endif
 
 static inline int pfn_present(unsigned long pfn)
 {
-	if (pfn_to_section_nr(pfn) >= NR_MEM_SECTIONS)
-		return 0;
-	return present_section(__nr_to_section(pfn_to_section_nr(pfn)));
+	return present_section(__pfn_to_section(pfn));
 }
 
 /*
diff --git a/mm/sparse.c b/mm/sparse.c
index 5d0cf45..3c91837 100644
--- a/mm/sparse.c
+++ b/mm/sparse.c
@@ -166,24 +166,59 @@ void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 	}
 }
 
+static int __init
+overlaps(u64 start1, u64 end1, u64 start2, u64 end2)
+{
+	u64 start, end;
+
+	start = max(start1, start2);
+	end = min(end1, end2);
+	return start <= end;
+}
+
 /* Record a memory area against a node. */
 void __init memory_present(int nid, unsigned long start, unsigned long end)
 {
+	unsigned long first_pfn = start;
 	unsigned long pfn;
 
 	start &= PAGE_SECTION_MASK;
 	mminit_validate_memmodel_limits(&start, &end);
 	for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 		unsigned long section = pfn_to_section_nr(pfn);
+		unsigned long last_pfn = min(pfn + PAGES_PER_SECTION, end) - 1;
 		struct mem_section *ms;
 
 		sparse_index_init(section, nid);
 		set_section_nid(section, nid);
 
 		ms = __nr_to_section(section);
-		if (!ms->section_mem_map)
+		if (!ms->section_mem_map) {
 			ms->section_mem_map = sparse_encode_early_nid(nid) |
 							SECTION_MARKED_PRESENT;
+		} else {
+			/* Merge the two regions */
+			WARN_ON(sparse_early_nid(ms) != nid);
+
+			/*
+			 * If they don't overlap there will be a hole in
+			 * between where meta-data says it's valid even though
+			 * it's not.
+			 */
+			if (!overlaps(first_pfn, last_pfn + 1,
+				      ms->first_pfn, ms->last_pfn + 1))	{
+				pr_info("Merging non-contiguous pfn ranges 0x%lx-0x%lx and 0x%lx-0x%lx\n",
+					ms->first_pfn, ms->last_pfn,
+					first_pfn, last_pfn);
+			}
+			first_pfn = min(first_pfn, ms->first_pfn);
+			last_pfn = max(last_pfn, ms->last_pfn);
+		}
+
+		ms->first_pfn = first_pfn;
+		ms->last_pfn = last_pfn;
+
+		first_pfn = pfn + PAGES_PER_SECTION;
 	}
 }
 
-- 
2.8.2