Re: [PATCH] x86/asm/entry/64: Clean up entry_64.S

[Sasha Levin <sasha.levin@oracle.com>]

On 07/08/2015 08:59 PM, Andy Lutomirski wrote:
> Having failed to bisect, let's look at the trace:
> 
> On Mon, Jul 6, 2015 at 8:00 AM, Sasha Levin <sasha.levin@oracle.com> wrote:
>> [3157054.661763] ------------[ cut here ]------------
>> [3157054.662552] kernel BUG at arch/x86/kernel/nmi.c:533!
>> [3157054.663277] invalid opcode: 0000 [#1] PREEMPT SMP KASAN
>> [3157054.664164] Dumping ftrace buffer:
>> [3157054.664740]    (ftrace buffer empty)
>> [3157054.665274] Modules linked in:
>> [3157054.665768] CPU: 16 PID: 11446 Comm: trinity-main Not tainted 4.1.0-next-20150703-sasha-00040-gd868f14-dirty #2292
>> [3157054.667203] task: ffff880408813000 ti: ffff8803d29c8000 task.ti: ffff8803d29c8000
>> [3157054.668256] RIP: do_nmi (arch/x86/kernel/nmi.c:533 (discriminator 1))
>> [3157054.669378] RSP: 0018:ffff88077800bed8  EFLAGS: 00010006
>> [3157054.670141] ==================================================================
>> [3157054.671268] BUG: KASan: out of bounds on stack in __show_regs+0x7f6/0x940 at addr ffff88077800be50
> 
> I bet that__show_regs interacts poorly with KASan for some reason.
> But that's not the underlying bug.  In fact, the bad read is quite
> close the RSP, so this is almost certainly a bug in KASan or
> __show_regs.
> 
>> [3157054.674604] Read of size 8 by task trinity-main/11446
>> [3157054.676521] page:ffffea001de002c0 count:1 mapcount:0 mapping:          (null) index:0x0
>> [3157054.679451] flags: 0x42fffff80000400(reserved)
>> [3157054.681237] page dumped because: kasan: bad access detected
>> [3157054.683326] CPU: 16 PID: 11446 Comm: trinity-main Not tainted 4.1.0-next-20150703-sasha-00040-gd868f14-dirty #2292
>> [3157054.687097]  ffff88077800be50 000000009c65e33f ffff88077800b9f8 ffffffffa0ac8938
>> [3157054.690303]  1ffffd4003bc0058 ffff88077800ba88 ffff88077800ba78 ffffffff9759796e
>> [3157054.693365]  ffff88077800bab8 ffffffffa0abe0b3 0000000000000082 ffffffffa2fe39e4
>> [3157054.696209] Call Trace:
>> [3157054.697180] <NMI> dump_stack (lib/dump_stack.c:52)
>> [3157054.699390] kasan_report_error (mm/kasan/report.c:132 mm/kasan/report.c:193)
>> [3157054.701663] ? printk (kernel/printk/printk.c:1896)
>> [3157054.703531] ? bitmap_weight (include/linux/bitmap.h:303)
>> [3157054.705553] __asan_report_load8_noabort (mm/kasan/report.c:230 mm/kasan/report.c:251)
>> [3157054.708014] ? __show_regs (arch/x86/kernel/process_64.c:68)
>> [3157054.710046] __show_regs (arch/x86/kernel/process_64.c:68)
>> [3157054.712066] ? printk (kernel/printk/printk.c:1896)
>> [3157054.713878] ? bitmap_weight (include/linux/bitmap.h:303)
>> [3157054.715875] ? start_thread_common.constprop.0 (arch/x86/kernel/process_64.c:58)
>> [3157054.718336] ? dump_stack_print_info (kernel/printk/printk.c:3121)
>> [3157054.720619] show_regs (arch/x86/kernel/dumpstack_64.c:313)
>> [3157054.722530] __die (arch/x86/kernel/dumpstack.c:294)
>> [3157054.724290] die (arch/x86/kernel/dumpstack.c:316)
>> [3157054.725962] do_trap (arch/x86/kernel/traps.c:214 arch/x86/kernel/traps.c:260)
>> [3157054.727805] do_error_trap (arch/x86/kernel/traps.c:298 include/linux/jump_label.h:125 include/linux/context_tracking_state.h:29 include/linux/context_tracking.h:46 arch/x86/kernel/traps.c:302)
>> [3157054.729843] ? do_device_not_available (arch/x86/kernel/traps.c:291)
>> [3157054.732211] ? do_nmi (arch/x86/kernel/nmi.c:533 (discriminator 1))
>> [3157054.734101] ? kvm_clock_read (./arch/x86/include/asm/preempt.h:87 arch/x86/kernel/kvmclock.c:86)
>> [3157054.736165] ? sched_clock (arch/x86/kernel/tsc.c:305)
>> [3157054.738126] ? nmi_handle (arch/x86/kernel/nmi.c:134)
>> [3157054.740133] ? trace_hardirqs_off_thunk (arch/x86/entry/thunk_64.S:40)
>> [3157054.742997] do_invalid_op (arch/x86/kernel/traps.c:313)
>> [3157054.744991] invalid_op (arch/x86/entry/entry_64.S:925)
> 
> So we got #UD somewhere...
> 
>> [3157054.746873] ? do_nmi (arch/x86/kernel/nmi.c:533 (discriminator 1))
>> [3157054.748769] ? do_nmi (arch/x86/kernel/nmi.c:515 arch/x86/kernel/nmi.c:531)
>> [3157054.750658] end_repeat_nmi (arch/x86/entry/entry_64.S:1435)
> 
> ...here, perhaps?
> 
> Do you know what line 1435 was in the version you tested?  There
> shouldn't be funny instructions in end_repeat_nmi, though.  Did we end
> up off an instruction boundary?

Yes, arch/x86/entry/entry_64.S:1435 is 'call    do_nmi', do_nmi does nmi_enter(),
which is actually:

	? do_nmi (arch/x86/kernel/nmi.c:533

And that has a BUG_ON(in_nmi()); in it.

> 
> Here's my wild guess.  The repeat_nmi thing is really rare.  What if
> there's a CPU or emulator that can't do mov %cr2, %r12 or vice versa?
> mov from cr has a somewhat unusual encoding.  What platform is this?
> Does KASan play games that would cause KVM to emulate a mov to or from
> cr2?

It's just a regular KVM guest, I'm not sure about KASan (Andrey Cc'ed).


Thanks,
Sasha