mm: thp: move pmd check inside ptl for freeze_page()

I found a race condition triggering VM_BUG_ON() in freeze_page(), when
running a testcase with 3 processes:
  - process 1: keep writing thp,
  - process 2: keep clearing soft-dirty bits from virtual address of process 1
  - process 3: call migratepages for process 1,

The kernel message is like this:

  kernel BUG at /src/linux-dev/mm/huge_memory.c:3096!
  invalid opcode: 0000 [#1] SMP
  Modules linked in: cfg80211 rfkill crc32c_intel ppdev serio_raw pcspkr virtio_balloon virtio_console parport_pc parport pvpanic acpi_cpufreq tpm_tis tpm i2c_piix4 virtio_blk virtio_net ata_generic pata_acpi floppy virtio_pci virtio_ring virtio
  CPU: 0 PID: 28863 Comm: migratepages Not tainted 4.6.0-v4.6-160602-0827-+ #2
  Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
  task: ffff880037320000 ti: ffff88007cdd0000 task.ti: ffff88007cdd0000
  RIP: 0010:[<ffffffff811f8e06>]  [<ffffffff811f8e06>] split_huge_page_to_list+0x496/0x590
  RSP: 0018:ffff88007cdd3b70  EFLAGS: 00010202
  RAX: 0000000000000001 RBX: ffff88007c7b88c0 RCX: 0000000000000000
  RDX: 0000000000000000 RSI: 0000000700000200 RDI: ffffea0003188000
  RBP: ffff88007cdd3bb8 R08: 0000000000000001 R09: 00003ffffffff000
  R10: ffff880000000000 R11: ffffc000001fffff R12: ffffea0003188000
  R13: ffffea0003188000 R14: 0000000000000000 R15: 0400000000000080
  FS:  00007f8ec241d740(0000) GS:ffff88007dc00000(0000) knlGS:0000000000000000             CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
  CR2: 00007f8ec1f3ed20 CR3: 000000003707b000 CR4: 00000000000006f0
  Call Trace:
    ? list_del+0xd/0x30
    queue_pages_pte_range+0x4d1/0x590
    __walk_page_range+0x204/0x4e0
    walk_page_range+0x71/0xf0
    queue_pages_range+0x75/0x90
    ? queue_pages_hugetlb+0x190/0x190
    ? new_node_page+0xc0/0xc0
    ? change_prot_numa+0x40/0x40
    migrate_to_node+0x71/0xd0
    do_migrate_pages+0x1c3/0x210
    SyS_migrate_pages+0x261/0x290
    entry_SYSCALL_64_fastpath+0x1a/0xa4
  Code: e8 b0 87 fb ff 0f 0b 48 c7 c6 30 32 9f 81 e8 a2 87 fb ff 0f 0b 48 c7 c6 b8 46 9f 81 e8 94 87 fb ff 0f 0b 85 c0 0f 84 3e fd ff ff <0f> 0b 85 c0 0f 85 a6 00 00 00 48 8b 75 c0 4c 89 f7 41 be f0 ff
  RIP   split_huge_page_to_list+0x496/0x590

I'm not sure of the full scenario of the reproduction, but my debug
showed that split_huge_pmd_address(freeze=true) returned without running
main code of pmd splitting because pmd_present(*pmd) in precheck somehow
returned 0.  If this happens, the subsequent try_to_unmap() fails and
returns non-zero (because page_mapcount() still > 0), and finally
VM_BUG_ON() fires.  This patch tries to fix it by prechecking pmd state
inside ptl.

Link: http://lkml.kernel.org/r/1466990929-7452-1-git-send-email-n-horiguchi@ah.jp.nec.com
Signed-off-by: Naoya Horiguchi <n-horiguchi@ah.jp.nec.com>
Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

1 files changed, 12 insertions, 19 deletions

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 55940d1..343a2b7 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -2984,7 +2984,7 @@ static void __split_huge_pmd_locked(struct vm_area_struct *vma, pmd_t *pmd,
 }
 
 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
-		unsigned long address, bool freeze)
+		unsigned long address, bool freeze, struct page *page)
 {
 	spinlock_t *ptl;
 	struct mm_struct *mm = vma->vm_mm;
@@ -2992,8 +2992,17 @@ void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 
 	mmu_notifier_invalidate_range_start(mm, haddr, haddr + HPAGE_PMD_SIZE);
 	ptl = pmd_lock(mm, pmd);
+
+	/*
+	 * If caller asks to setup a migration entries, we need a page to check
+	 * pmd against. Otherwise we can end up replacing wrong page.
+	 */
+	VM_BUG_ON(freeze && !page);
+	if (page && page != pmd_page(*pmd))
+	        goto out;
+
 	if (pmd_trans_huge(*pmd)) {
-		struct page *page = pmd_page(*pmd);
+		page = pmd_page(*pmd);
 		if (PageMlocked(page))
 			clear_page_mlock(page);
 	} else if (!pmd_devmap(*pmd))
@@ -3020,24 +3029,8 @@ void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 		return;
 
 	pmd = pmd_offset(pud, address);
-	if (!pmd_present(*pmd) || (!pmd_trans_huge(*pmd) && !pmd_devmap(*pmd)))
-		return;
 
-	/*
-	 * If caller asks to setup a migration entries, we need a page to check
-	 * pmd against. Otherwise we can end up replacing wrong page.
-	 */
-	VM_BUG_ON(freeze && !page);
-	if (page && page != pmd_page(*pmd))
-		return;
-
-	/*
-	 * Caller holds the mmap_sem write mode or the anon_vma lock,
-	 * so a huge pmd cannot materialize from under us (khugepaged
-	 * holds both the mmap_sem write mode and the anon_vma lock
-	 * write mode).
-	 */
-	__split_huge_pmd(vma, pmd, address, freeze);
+	__split_huge_pmd(vma, pmd, address, freeze, page);
 }
 
 void vma_adjust_trans_huge(struct vm_area_struct *vma,