Re: mm: BUG in page_move_anon_rmap

[Kirill A. Shutemov]

On Mon, Jul 04, 2016 at 04:10:53PM -0700, Hugh Dickins wrote:
> On Fri, 1 Jul 2016, Dmitry Vyukov wrote:
> > Hello,
> > 
> > I am getting the following crashes while running syzkaller fuzzer on
> > 00bf377d19ad3d80cbc7a036521279a86e397bfb (Jun 29). So far I did not
> > manage to reproduce it outside of fuzzer, but fuzzer hits it once per
> > hour or so.
> > 
> > flags: 0xfffe0000044079(locked|uptodate|dirty|lru|active|head|swapbacked)
> > page dumped because: VM_BUG_ON_PAGE(page->index !=
> > linear_page_index(vma, address))
> > page->mem_cgroup:ffff88003e829be0
> > ------------[ cut here ]------------
> > kernel BUG at mm/rmap.c:1103!
> > invalid opcode: 0000 [#2] SMP DEBUG_PAGEALLOC KASAN
> > Modules linked in:
> > CPU: 0 PID: 7043 Comm: syz-fuzzer Tainted: G      D         4.7.0-rc5+ #22
> > Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
> > task: ffff8800342f46c0 ti: ffff880034008000 task.ti: ffff880034008000
> > RIP: 0010:[<ffffffff817693d8>] [<ffffffff817693d8>]
> > page_move_anon_rmap+0x278/0x310 mm/rmap.c:1103
> > RSP: 0000:ffff88003400fad0  EFLAGS: 00010286
> > RAX: ffff8800342f46c0 RBX: ffffea0000928000 RCX: 0000000000000000
> > RDX: 0000000000000000 RSI: ffff88003ec16de8 RDI: ffffed0006801f41
> > RBP: ffff88003400fb00 R08: 0000000000000001 R09: 0000000000000000
> > R10: 0000000000000000 R11: ffffed000fffea01 R12: ffff88006776b8e8
> > R13: 001000000c829e00 R14: ffff88006247c3e8 R15: 000000000c829e00
> > FS:  00007f7627bc5700(0000) GS:ffff88003ec00000(0000) knlGS:0000000000000000
> > CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > CR2: 000000c829fd8000 CR3: 0000000034b23000 CR4: 00000000000006f0
> > Stack:
> >  ffffea0000928000 ffffea000092f600 ffff88006776b8e8 ffffea0000928000
> >  ffffea0000928001 000000c829fd8000 ffff88003400fc38 ffffffff8173a25f
> >  0000000000000086 ffff88003400fbd0 ffffea0000928001 ffff880036cd3ec0
> > Call Trace:
> >  [<ffffffff8173a25f>] do_wp_page+0x7df/0x1c90 mm/memory.c:2402
> >  [<ffffffff817404f5>] handle_pte_fault+0x1e85/0x4960 mm/memory.c:3381
> >  [<     inline     >] __handle_mm_fault mm/memory.c:3489
> >  [<ffffffff8174443b>] handle_mm_fault+0xeab/0x11a0 mm/memory.c:3518
> >  [<ffffffff81290f77>] __do_page_fault+0x457/0xbb0 arch/x86/mm/fault.c:1356
> >  [<ffffffff8129181f>] trace_do_page_fault+0xdf/0x5b0 arch/x86/mm/fault.c:1449
> >  [<ffffffff81281c24>] do_async_page_fault+0x14/0xd0 arch/x86/kernel/kvm.c:265
> >  [<ffffffff86a9d538>] async_page_fault+0x28/0x30 arch/x86/entry/entry_64.S:923
> > Code: 0b e8 dd d5 e2 ff 48 c7 c6 40 f7 d0 86 48 89 df e8 2e 4a fc ff
> > 0f 0b e8 c7 d5 e2 ff 48 c7 c6 c0 f7 d0 86 48 89 df e8 18 4a fc ff <0f>
> > 0b e8 b1 d5 e2 ff 4c 89 ee 4c 89 e7 e8 96 80 02 00 49 89 c5
> > RIP  [<ffffffff817693d8>] page_move_anon_rmap+0x278/0x310 mm/rmap.c:1103
> >  RSP <ffff88003400fad0>
> > ---[ end trace b6c02a1136e2a9ec ]---
> > BUG: sleeping function called from invalid context at include/linux/sched.h:2955
> > in_atomic(): 1, irqs_disabled(): 0, pid: 7043, name: syz-fuzzer
> > lockdep is turned off.
> > CPU: 0 PID: 7043 Comm: syz-fuzzer Tainted: G      D         4.7.0-rc5+ #22
> > Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
> >  ffffffff880b58e0 ffff88003400f5c0 ffffffff82cc924f ffffffff342f46c0
> >  fffffbfff1016b1c ffff8800342f46c0 0000000000001b83 0000000000000000
> >  0000000000000000 dffffc0000000000 ffff88003400f5e8 ffffffff813efbfb
> > Call Trace:
> >  [<     inline     >] __dump_stack lib/dump_stack.c:15
> >  [<ffffffff82cc924f>] dump_stack+0x12e/0x18f lib/dump_stack.c:51
> >  [<ffffffff813efbfb>] ___might_sleep+0x27b/0x3a0 kernel/sched/core.c:7573
> >  [<ffffffff813efdb0>] __might_sleep+0x90/0x1a0 kernel/sched/core.c:7535
> >  [<     inline     >] threadgroup_change_begin include/linux/sched.h:2955
> >  [<ffffffff813a175f>] exit_signals+0x7f/0x430 kernel/signal.c:2392
> >  [<ffffffff8137a6a4>] do_exit+0x234/0x2c80 kernel/exit.c:701
> >  [<ffffffff81204331>] oops_end+0xa1/0xd0 arch/x86/kernel/dumpstack.c:250
> >  [<ffffffff812045c6>] die+0x46/0x60 arch/x86/kernel/dumpstack.c:308
> >  [<     inline     >] do_trap_no_signal arch/x86/kernel/traps.c:192
> >  [<ffffffff811fd9f2>] do_trap+0x192/0x380 arch/x86/kernel/traps.c:238
> >  [<ffffffff811fde4e>] do_error_trap+0x11e/0x280 arch/x86/kernel/traps.c:275
> >  [<ffffffff811ff18b>] do_invalid_op+0x1b/0x20 arch/x86/kernel/traps.c:288
> >  [<ffffffff86a9cf0e>] invalid_op+0x1e/0x30 arch/x86/entry/entry_64.S:761
> >  [<ffffffff8173a25f>] do_wp_page+0x7df/0x1c90 mm/memory.c:2402
> >  [<ffffffff817404f5>] handle_pte_fault+0x1e85/0x4960 mm/memory.c:3381
> >  [<     inline     >] __handle_mm_fault mm/memory.c:3489
> >  [<ffffffff8174443b>] handle_mm_fault+0xeab/0x11a0 mm/memory.c:3518
> >  [<ffffffff81290f77>] __do_page_fault+0x457/0xbb0 arch/x86/mm/fault.c:1356
> >  [<ffffffff8129181f>] trace_do_page_fault+0xdf/0x5b0 arch/x86/mm/fault.c:1449
> >  [<ffffffff81281c24>] do_async_page_fault+0x14/0xd0 arch/x86/kernel/kvm.c:265
> >  [<ffffffff86a9d538>] async_page_fault+0x28/0x30 arch/x86/entry/entry_64.S:923
> > note: syz-fuzzer[7043] exited with preempt_count 1
> 
> I think 0798d3c022dc ("mm: thp: avoid false positive VM_BUG_ON_PAGE in
> page_move_anon_rmap()") is flawed.  It would certainly be interesting
> to hear whether this patch fixes your crashes:
> 
> --- 4.7-rc6/include/linux/pagemap.h	2016-05-29 15:47:38.303064058 -0700
> +++ linux/include/linux/pagemap.h	2016-07-04 15:44:46.635147739 -0700
> @@ -408,7 +408,7 @@ static inline pgoff_t linear_page_index(
>  	pgoff_t pgoff;
>  	if (unlikely(is_vm_hugetlb_page(vma)))
>  		return linear_hugepage_index(vma, address);
> -	pgoff = (address - vma->vm_start) >> PAGE_SHIFT;
> +	pgoff = (long)(address - vma->vm_start) >> PAGE_SHIFT;
>  	pgoff += vma->vm_pgoff;
>  	return pgoff;
>  }
> 
> But if it does work, I'm not sure that we really want to extend
> linear_page_index() to go in a backward direction, just for this case.
> 
> I think I'd prefer to skip page_move_anon_rmap()'s linear_page_index()
> check in the PageTransHuge case; or, indeed, remove that check (and
> the address arg) completely - address plays no further part there.
> But let's wait to see what Kirill prefers before committing.

Yeah, I would rather kill the VM_BUG_ON_PAGE() altogether and simplify
page_move_anon_rmap() interface. It doesn't make sense to add even more
glue to get the assert work.

-- 
 Kirill A. Shutemov