Date:   Thu, 9 Jan 2020 11:05:13 +0100
From:   Dmitry Vyukov <dvyukov@...gle.com>
To:     Casey Schaufler <casey@...aufler-ca.com>,
        Daniel Axtens <dja@...ens.net>,
        Alexander Potapenko <glider@...gle.com>,
        clang-built-linux <clang-built-linux@...glegroups.com>
Cc:     Tetsuo Handa <penguin-kernel@...ove.sakura.ne.jp>,
        syzbot <syzbot+de8d933e7d153aa0c1bb@...kaller.appspotmail.com>,
        kasan-dev <kasan-dev@...glegroups.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        LKML <linux-kernel@...r.kernel.org>,
        syzkaller-bugs <syzkaller-bugs@...glegroups.com>
Subject: Re: INFO: rcu detected stall in sys_kill

On Thu, Jan 9, 2020 at 10:29 AM Dmitry Vyukov <dvyukov@...gle.com> wrote:
> > > >
> > > > On 1/8/2020 2:25 AM, Tetsuo Handa wrote:
> > > > > On 2020/01/08 15:20, Dmitry Vyukov wrote:
> > > > >> I temporarily re-enabled smack instance and it produced another 50
> > > > >> stalls all over the kernel, and now keeps spewing a dozen every hour.
> > > >
> > > > Do I have to be using clang to test this? I'm setting up to work on this,
> > > > and don't want to waste time using my current tool chain if the problem
> > > > is clang specific.
> > >
> > > Humm, interesting. Initially I was going to say that most likely it's
> > > not clang-related. Bug smack instance is actually the only one that
> > > uses clang as well (except for KMSAN of course). So maybe it's indeed
> > > clang-related rather than smack-related. Let me try to build a kernel
> > > with clang.
> >
> > +clang-built-linux, glider
> >
> > [clang-built linux is severe broken since early Dec]
> >
> > Building kernel with clang I can immediately reproduce this locally:
> >
> > $ syz-manager
> > 2020/01/09 09:27:15 loading corpus...
> > 2020/01/09 09:27:17 serving http on http://0.0.0.0:50001
> > 2020/01/09 09:27:17 serving rpc on tcp://[::]:45851
> > 2020/01/09 09:27:17 booting test machines...
> > 2020/01/09 09:27:17 wait for the connection from test machine...
> > 2020/01/09 09:29:23 machine check:
> > 2020/01/09 09:29:23 syscalls                : 2961/3195
> > 2020/01/09 09:29:23 code coverage           : enabled
> > 2020/01/09 09:29:23 comparison tracing      : enabled
> > 2020/01/09 09:29:23 extra coverage          : enabled
> > 2020/01/09 09:29:23 setuid sandbox          : enabled
> > 2020/01/09 09:29:23 namespace sandbox       : enabled
> > 2020/01/09 09:29:23 Android sandbox         : /sys/fs/selinux/policy
> > does not exist
> > 2020/01/09 09:29:23 fault injection         : enabled
> > 2020/01/09 09:29:23 leak checking           : CONFIG_DEBUG_KMEMLEAK is
> > not enabled
> > 2020/01/09 09:29:23 net packet injection    : enabled
> > 2020/01/09 09:29:23 net device setup        : enabled
> > 2020/01/09 09:29:23 concurrency sanitizer   : /sys/kernel/debug/kcsan
> > does not exist
> > 2020/01/09 09:29:23 devlink PCI setup       : PCI device 0000:00:10.0
> > is not available
> > 2020/01/09 09:29:27 corpus                  : 50226 (0 deleted)
> > 2020/01/09 09:29:27 VMs 20, executed 0, cover 0, crashes 0, repro 0
> > 2020/01/09 09:29:37 VMs 20, executed 45, cover 0, crashes 0, repro 0
> > 2020/01/09 09:29:47 VMs 20, executed 74, cover 0, crashes 0, repro 0
> > 2020/01/09 09:29:57 VMs 20, executed 80, cover 0, crashes 0, repro 0
> > 2020/01/09 09:30:07 VMs 20, executed 80, cover 0, crashes 0, repro 0
> > 2020/01/09 09:30:17 VMs 20, executed 80, cover 0, crashes 0, repro 0
> > 2020/01/09 09:30:27 VMs 20, executed 80, cover 0, crashes 0, repro 0
> > 2020/01/09 09:30:37 VMs 20, executed 80, cover 0, crashes 0, repro 0
> > 2020/01/09 09:30:47 VMs 20, executed 80, cover 0, crashes 0, repro 0
> > 2020/01/09 09:30:57 VMs 20, executed 80, cover 0, crashes 0, repro 0
> > 2020/01/09 09:31:07 VMs 20, executed 80, cover 0, crashes 0, repro 0
> > 2020/01/09 09:31:17 VMs 20, executed 80, cover 0, crashes 0, repro 0
> > 2020/01/09 09:31:26 vm-10: crash: INFO: rcu detected stall in do_idle
> > 2020/01/09 09:31:27 VMs 13, executed 80, cover 0, crashes 0, repro 0
> > 2020/01/09 09:31:28 vm-1: crash: INFO: rcu detected stall in sys_futex
> > 2020/01/09 09:31:29 vm-4: crash: INFO: rcu detected stall in sys_futex
> > 2020/01/09 09:31:31 vm-0: crash: INFO: rcu detected stall in sys_getsockopt
> > 2020/01/09 09:31:33 vm-18: crash: INFO: rcu detected stall in sys_clone3
> > 2020/01/09 09:31:35 vm-3: crash: INFO: rcu detected stall in sys_futex
> > 2020/01/09 09:31:36 vm-8: crash: INFO: rcu detected stall in do_idle
> > 2020/01/09 09:31:37 VMs 7, executed 80, cover 0, crashes 6, repro 0
> > 2020/01/09 09:31:38 vm-19: crash: INFO: rcu detected stall in schedule_tail
> > 2020/01/09 09:31:40 vm-6: crash: INFO: rcu detected stall in schedule_tail
> > 2020/01/09 09:31:42 vm-2: crash: INFO: rcu detected stall in schedule_tail
> > 2020/01/09 09:31:44 vm-12: crash: INFO: rcu detected stall in sys_futex
> > 2020/01/09 09:31:46 vm-15: crash: INFO: rcu detected stall in sys_nanosleep
> > 2020/01/09 09:31:47 VMs 1, executed 80, cover 0, crashes 11, repro 0
> > 2020/01/09 09:31:48 vm-16: crash: INFO: rcu detected stall in sys_futex
> > 2020/01/09 09:31:50 vm-9: crash: INFO: rcu detected stall in schedule
> > 2020/01/09 09:31:52 vm-13: crash: INFO: rcu detected stall in schedule_tail
> > 2020/01/09 09:31:54 vm-11: crash: INFO: rcu detected stall in schedule_tail
> > 2020/01/09 09:31:56 vm-17: crash: INFO: rcu detected stall in sys_futex
> > 2020/01/09 09:31:57 VMs 0, executed 80, cover 0, crashes 16, repro 0
> > 2020/01/09 09:31:58 vm-7: crash: INFO: rcu detected stall in sys_futex
> > 2020/01/09 09:32:00 vm-5: crash: INFO: rcu detected stall in dput
> > 2020/01/09 09:32:02 vm-14: crash: INFO: rcu detected stall in sys_nanosleep
> >
> >
> > Then I switched LSM to selinux and I _still_ can reproduce this. So,
> > Casey, you may relax, this is not smack-specific :)
> >
> > Then I disabled CONFIG_KASAN_VMALLOC and CONFIG_VMAP_STACK and it
> > started working normally.
> >
> > So this is somehow related to both clang and KASAN/VMAP_STACK.
> >
> > The clang I used is:
> > https://storage.googleapis.com/syzkaller/clang-kmsan-362913.tar.gz
> > (the one we use on syzbot).
>
>
> Clustering hangs, they all happen within very limited section of the code:
>
>       1  free_thread_stack+0x124/0x590 kernel/fork.c:284
>       5  free_thread_stack+0x12e/0x590 kernel/fork.c:280
>      39  free_thread_stack+0x12e/0x590 kernel/fork.c:284
>       6  free_thread_stack+0x133/0x590 kernel/fork.c:280
>       5  free_thread_stack+0x13d/0x590 kernel/fork.c:280
>       2  free_thread_stack+0x141/0x590 kernel/fork.c:280
>       6  free_thread_stack+0x14c/0x590 kernel/fork.c:280
>       9  free_thread_stack+0x151/0x590 kernel/fork.c:280
>       3  free_thread_stack+0x15b/0x590 kernel/fork.c:280
>      67  free_thread_stack+0x168/0x590 kernel/fork.c:280
>       6  free_thread_stack+0x16d/0x590 kernel/fork.c:284
>       2  free_thread_stack+0x177/0x590 kernel/fork.c:284
>       1  free_thread_stack+0x182/0x590 kernel/fork.c:284
>       1  free_thread_stack+0x186/0x590 kernel/fork.c:284
>      16  free_thread_stack+0x18b/0x590 kernel/fork.c:284
>       4  free_thread_stack+0x195/0x590 kernel/fork.c:284
>
> Here is disass of the function:
> https://gist.githubusercontent.com/dvyukov/a283d1aaf2ef7874001d56525279ccbd/raw/ac2478bff6472bc473f57f91a75f827cd72bb6bf/gistfile1.txt
>
> But if I am not mistaken, the function only ever jumps down. So how
> can it loop?...


This is a miscompilation related to static branches.

objdump shows:

ffffffff814878f8: 0f 1f 44 00 00        nopl   0x0(%rax,%rax,1)
 ./arch/x86/include/asm/jump_label.h:25
asm_volatile_goto("1:"

However, the actual instruction in memory at the time is:

   0xffffffff814878f8 <+408>: jmpq   0xffffffff8148787f <free_thread_stack+287>

Which jumps to a wrong location in free_thread_stack and makes it loop.

The static branch is this:

static inline bool memcg_kmem_enabled(void)
{
  return static_branch_unlikely(&memcg_kmem_enabled_key);
}

static inline void memcg_kmem_uncharge(struct page *page, int order)
{
  if (memcg_kmem_enabled())
    __memcg_kmem_uncharge(page, order);
}

I suspect it may have something to do with loop unrolling. It may jump
to the right location, but in the wrong unrolled iteration.

Powered by blists - more mailing lists