Sasha reports unmap_page_range tears down pmd range which is race with thp
splitting during page reclaim. Transparent huge page will be splitting
during page reclaim. However, split pmd lock which held by __split_trans_huge_lock
can't prevent __split_huge_page_refcount running in parallel. This patch fix
it by hold compound lock to check if __split_huge_page_refcount is running
underneath, in that case zap huge pmd range should be fallback.
[ 265.474585] kernel BUG at mm/huge_memory.c:1440!
[ 265.475129] invalid opcode: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
[ 265.476684] Dumping ftrace buffer:
[ 265.477144] (ftrace buffer empty)
[ 265.478398] Modules linked in:
[ 265.478807] CPU: 8 PID: 11344 Comm: trinity-c206 Tainted: G W 3.13.0-rc5-next-20131223-sasha-00015-gec22156-dirty #8
[ 265.480172] task: ffff8801cb573000 ti: ffff8801cbd3a000 task.ti: ffff8801cbd3a000
[ 265.480172] RIP: 0010:[<ffffffff812c7f70>] [<ffffffff812c7f70>] zap_huge_pmd+0x170/0x1f0
[ 265.480172] RSP: 0000:ffff8801cbd3bc78 EFLAGS: 00010246
[ 265.480172] RAX: 015fffff80090018 RBX: ffff8801cbd3bde8 RCX: ffffffffffffff9c
[ 265.480172] RDX: ffffffffffffffff RSI: 0000000000000008 RDI: ffff8800bffd2000
[ 265.480172] RBP: ffff8801cbd3bcb8 R08: 0000000000000000 R09: 0000000000000000
[ 265.480172] R10: 0000000000000001 R11: 0000000000000000 R12: ffffea0002856740
[ 265.480172] R13: ffffea0002d50000 R14: 00007ff915000000 R15: 00007ff930e48fff
[ 265.480172] FS: 00007ff934899700(0000) GS:ffff88014d400000(0000) knlGS:0000000000000000
[ 265.480172] CS: 0010 DS: 0000 ES: 0000 CR0: 000000008005003b
[ 265.480172] CR2: 00007ff93428a000 CR3: 000000010babe000 CR4: 00000000000006e0
[ 265.480172] Stack:
[ 265.480172] 00000000000004dd ffff8801ccbfbb60 ffff8801cbd3bcb8 ffff8801cbb15540
[ 265.480172] 00007ff915000000 00007ff930e49000 ffff8801cbd3bde8 00007ff930e48fff
[ 265.480172] ffff8801cbd3bd48 ffffffff812885b6 ffff88005f5d20c0 00007ff915200000
[ 265.480172] Call Trace:
[ 265.480172] [<ffffffff812885b6>] unmap_page_range+0x2c6/0x410
[ 265.480172] [<ffffffff81288801>] unmap_single_vma+0x101/0x120
[ 265.480172] [<ffffffff81288881>] unmap_vmas+0x61/0xa0
[ 265.480172] [<ffffffff8128f730>] exit_mmap+0xd0/0x170
[ 265.480172] [<ffffffff81138860>] mmput+0x70/0xe0
[ 265.480172] [<ffffffff8113c89d>] exit_mm+0x18d/0x1a0
[ 265.480172] [<ffffffff811ea355>] ? acct_collect+0x175/0x1b0
[ 265.480172] [<ffffffff8113ed0f>] do_exit+0x26f/0x520
[ 265.480172] [<ffffffff8113f069>] do_group_exit+0xa9/0xe0
[ 265.480172] [<ffffffff8113f0b7>] SyS_exit_group+0x17/0x20
[ 265.480172] [<ffffffff845f10d0>] tracesys+0xdd/0xe2
[ 265.480172] Code: 0f 0b 66 0f 1f 84 00 00 00 00 00 eb fe 66 0f 1f
44 00 00 48 8b 03 f0 48 81 80 50 03 00 00 00 fe ff ff 49 8b 45 00 f6
c4 40 75 10 <0f> 0b 66 0f 1f 44 00 00 eb fe 66 0f 1f 44 00 00 48 8b 03
f0 48
[ 265.480172] RIP [<ffffffff812c7f70>] zap_huge_pmd+0x170/0x1f0
[ 265.480172] RSP <ffff8801cbd3bc78>
Reported-by: Sasha Levin <sasha.levin@xxxxxxxxxx>
Signed-off-by: Wanpeng Li <liwanp@xxxxxxxxxxxxxxxxxx>
---
mm/huge_memory.c | 13 ++++++++++++-
1 file changed, 12 insertions(+), 1 deletion(-)
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 7de1bf8..d1e0c80 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -1414,6 +1414,7 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
{
spinlock_t *ptl;
int ret = 0;
+ unsigned long flags;
if (__pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
struct page *page;
@@ -1426,6 +1427,15 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
* operations.
*/
orig_pmd = pmdp_get_and_clear(tlb->mm, addr, pmd);
+ page = pmd_page(orig_pmd);
+ flags = compound_lock_irqsave(page);
+ if (unlikely(!PageHead(page))) {
+ /*
+ * __split_huge_page_refcount run before us
+ */
+ compound_unlock_irqrestore(page, flags);
+ return ret;
+ }
tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
pgtable = pgtable_trans_huge_withdraw(tlb->mm, pmd);
if (is_huge_zero_pmd(orig_pmd)) {
@@ -1433,7 +1443,6 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
spin_unlock(ptl);
put_huge_zero_page();
} else {
- page = pmd_page(orig_pmd);
page_remove_rmap(page);
VM_BUG_ON(page_mapcount(page) < 0);
add_mm_counter(tlb->mm, MM_ANONPAGES, -HPAGE_PMD_NR);
@@ -1442,6 +1451,8 @@ int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
spin_unlock(ptl);
tlb_remove_page(tlb, page);
}
+ compound_unlock_irqrestore(page, flags);
+
pte_free(tlb->mm, pgtable);
ret = 1;
}
--
1.8.3.2
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>