summaryrefslogtreecommitdiffstats
path: root/mm
diff options
context:
space:
mode:
authorMel Gorman <mgorman@suse.de>2012-02-03 15:37:18 -0800
committerLinus Torvalds <torvalds@linux-foundation.org>2012-02-03 16:16:41 -0800
commit0bf380bc70ecba68cb4d74dc656cc2fa8c4d801a (patch)
tree6d5869a979b62c3b21491f007519185af30442a0 /mm
parent3deaa7190a8da38453c4fabd9dec7f66d17fff67 (diff)
downloadlinux-0bf380bc70ecba68cb4d74dc656cc2fa8c4d801a.tar.bz2
mm: compaction: check pfn_valid when entering a new MAX_ORDER_NR_PAGES block during isolation for migration
When isolating for migration, migration starts at the start of a zone which is not necessarily pageblock aligned. Further, it stops isolating when COMPACT_CLUSTER_MAX pages are isolated so migrate_pfn is generally not aligned. This allows isolate_migratepages() to call pfn_to_page() on an invalid PFN which can result in a crash. This was originally reported against a 3.0-based kernel with the following trace in a crash dump. PID: 9902 TASK: d47aecd0 CPU: 0 COMMAND: "memcg_process_s" #0 [d72d3ad0] crash_kexec at c028cfdb #1 [d72d3b24] oops_end at c05c5322 #2 [d72d3b38] __bad_area_nosemaphore at c0227e60 #3 [d72d3bec] bad_area at c0227fb6 #4 [d72d3c00] do_page_fault at c05c72ec #5 [d72d3c80] error_code (via page_fault) at c05c47a4 EAX: 00000000 EBX: 000c0000 ECX: 00000001 EDX: 00000807 EBP: 000c0000 DS: 007b ESI: 00000001 ES: 007b EDI: f3000a80 GS: 6f50 CS: 0060 EIP: c030b15a ERR: ffffffff EFLAGS: 00010002 #6 [d72d3cb4] isolate_migratepages at c030b15a #7 [d72d3d14] zone_watermark_ok at c02d26cb #8 [d72d3d2c] compact_zone at c030b8de #9 [d72d3d68] compact_zone_order at c030bba1 #10 [d72d3db4] try_to_compact_pages at c030bc84 #11 [d72d3ddc] __alloc_pages_direct_compact at c02d61e7 #12 [d72d3e08] __alloc_pages_slowpath at c02d66c7 #13 [d72d3e78] __alloc_pages_nodemask at c02d6a97 #14 [d72d3eb8] alloc_pages_vma at c030a845 #15 [d72d3ed4] do_huge_pmd_anonymous_page at c03178eb #16 [d72d3f00] handle_mm_fault at c02f36c6 #17 [d72d3f30] do_page_fault at c05c70ed #18 [d72d3fb0] error_code (via page_fault) at c05c47a4 EAX: b71ff000 EBX: 00000001 ECX: 00001600 EDX: 00000431 DS: 007b ESI: 08048950 ES: 007b EDI: bfaa3788 SS: 007b ESP: bfaa36e0 EBP: bfaa3828 GS: 6f50 CS: 0073 EIP: 080487c8 ERR: ffffffff EFLAGS: 00010202 It was also reported by Herbert van den Bergh against 3.1-based kernel with the following snippet from the console log. BUG: unable to handle kernel paging request at 01c00008 IP: [<c0522399>] isolate_migratepages+0x119/0x390 *pdpt = 000000002f7ce001 *pde = 0000000000000000 It is expected that it also affects 3.2.x and current mainline. The problem is that pfn_valid is only called on the first PFN being checked and that PFN is not necessarily aligned. Lets say we have a case like this H = MAX_ORDER_NR_PAGES boundary | = pageblock boundary m = cc->migrate_pfn f = cc->free_pfn o = memory hole H------|------H------|----m-Hoooooo|ooooooH-f----|------H The migrate_pfn is just below a memory hole and the free scanner is beyond the hole. When isolate_migratepages started, it scans from migrate_pfn to migrate_pfn+pageblock_nr_pages which is now in a memory hole. It checks pfn_valid() on the first PFN but then scans into the hole where there are not necessarily valid struct pages. This patch ensures that isolate_migratepages calls pfn_valid when necessary. Reported-by: Herbert van den Bergh <herbert.van.den.bergh@oracle.com> Tested-by: Herbert van den Bergh <herbert.van.den.bergh@oracle.com> Signed-off-by: Mel Gorman <mgorman@suse.de> Acked-by: Michal Nazarewicz <mina86@mina86.com> Cc: <stable@vger.kernel.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'mm')
-rw-r--r--mm/compaction.c13
1 files changed, 13 insertions, 0 deletions
diff --git a/mm/compaction.c b/mm/compaction.c
index 71a58f67f481..bd939a574b84 100644
--- a/mm/compaction.c
+++ b/mm/compaction.c
@@ -313,6 +313,19 @@ static isolate_migrate_t isolate_migratepages(struct zone *zone,
} else if (!locked)
spin_lock_irq(&zone->lru_lock);
+ /*
+ * migrate_pfn does not necessarily start aligned to a
+ * pageblock. Ensure that pfn_valid is called when moving
+ * into a new MAX_ORDER_NR_PAGES range in case of large
+ * memory holes within the zone
+ */
+ if ((low_pfn & (MAX_ORDER_NR_PAGES - 1)) == 0) {
+ if (!pfn_valid(low_pfn)) {
+ low_pfn += MAX_ORDER_NR_PAGES - 1;
+ continue;
+ }
+ }
+
if (!pfn_valid_within(low_pfn))
continue;
nr_scanned++;