Re: Strange hang with gcc 8 of kprobe multiple_kprobes test

[Masami Hiramatsu]

Hi Steve,

On Mon, 3 Dec 2018 21:18:07 -0500
Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:

> Hi Masami,
> 
> I started testing some of my new code and the system got into a
> strange state. Debugging further, I found the cause came from the
> kprobe tests. It became stranger to me that I could reproduce it with
> older kernels. I went back as far as 4.16 and it triggered. I thought
> this very strange because I ran this test on all those kernels in the
> past.
> 
> After a bit of hair pulling, I figured out what changed. I upgraded to
> gcc 8.1 (and I reproduce it with 8.2 as well). I convert back to gcc 7
> and the tests pass without issue.

OK, let me see.

> The issue that I notice when the system gets into this strange state is
> that I can't log into the box. Nor can I reboot. Basically it's
> anything to do with systemd just doesn't work (insert your jokes here
> now, and then let's move on).
> 
> I was able to narrow down what the exact function was that caused the
> issues and it is: update_vsyscall()
> 
> gcc 7 looks like this:
> 
> ffffffff81004bf0 <update_vsyscall>:
> ffffffff81004bf0:       e8 0b cc 9f 00          callq  ffffffff81a01800 <__fentry__>
>                         ffffffff81004bf1: R_X86_64_PC32 __fentry__-0x4
> ffffffff81004bf5:       48 8b 07                mov    (%rdi),%rax
> ffffffff81004bf8:       8b 15 96 5f 34 01       mov    0x1345f96(%rip),%edx        # ffffffff8234ab94 <vclocks_used>
>                         ffffffff81004bfa: R_X86_64_PC32 vclocks_used-0x4
> ffffffff81004bfe:       83 05 7b 84 6f 01 01    addl   $0x1,0x16f847b(%rip)        # ffffffff826fd080 <vsyscall_gtod_data>
>                         ffffffff81004c00: R_X86_64_PC32 vsyscall_gtod_data-0x5
> ffffffff81004c05:       8b 48 24                mov    0x24(%rax),%ecx
> ffffffff81004c08:       b8 01 00 00 00          mov    $0x1,%eax
> ffffffff81004c0d:       d3 e0                   shl    %cl,%eax
> 
> And gcc 8 looks like this:
> 
> ffffffff81004c90 <update_vsyscall>:
> ffffffff81004c90:       e8 6b cb 9f 00          callq  ffffffff81a01800 <__fentry__>
>                         ffffffff81004c91: R_X86_64_PC32 __fentry__-0x4
> ffffffff81004c95:       48 8b 07                mov    (%rdi),%rax
> ffffffff81004c98:       83 05 e1 93 6f 01 01    addl   $0x1,0x16f93e1(%rip)        # ffffffff826fe080 <vsyscall_gtod_data>

Hm this is a RIP relative instruction, it should be modified by kprobes.

>                         ffffffff81004c9a: R_X86_64_PC32 vsyscall_gtod_data-0x5
> ffffffff81004c9f:       8b 50 24                mov    0x24(%rax),%edx
> ffffffff81004ca2:       8b 05 ec 5e 34 01       mov    0x1345eec(%rip),%eax        # ffffffff8234ab94 <vclocks_used>
>                         ffffffff81004ca4: R_X86_64_PC32 vclocks_used-0x4
> 
> The test adds a kprobe (optimized) at udpate_vsyscall+5. And will
> insert a jump on the two instructions after fentry. The difference
> between v7 and v8 is that v7 is touching vclocks_used and v8 is
> touching vsyscall_gtod_data.
> 
> Is there some black magic going on with the vsyscall area with
> vsyscall_gtod_data that is causing havoc when a kprobe is added there?

I think it might miss something when preprocessing RIP relative instruction.
Could you disable jump optimization as below and test what happen on
update_vsyscall+5 AND update_vsyscall+8? (RIP relative preprocess must
happen even if the jump optimization is disabled)

# echo 0 > /proc/sys/debug/kprobes-optimization


> I can dig a little more into this, but I'm currently at my HQ office
> with a lot of other objectives that I must get done, and I can't work
> on this much more this week.

OK, let me try to reproduce it in my environment.

> 
> I included my config (for my virt machine, which I was also able to
> trigger it with).

Thanks, but I think it should not depend on the kconfig.

> 
> The test that triggers this bug is:
> 
>  tools/testing/selftests/ftrace/test.d/kprobe/multiple_kprobes.tc
> 
> It runs the test fine, but other things just start to act up after I
> run it.

Yeah, thank you for digging it down. It is now much easier to me.

> 
> I notice that when I get into the state, journald and the dbus_daemon
> are constantly running. Perhaps the userspace time keeping went bad?

Yeah, I think so. Maybe addl instruction becomes broken.

Thank you,

-- 
Masami Hiramatsu <mhiramat@xxxxxxxxxx>