Re: INFO: rcu detected stall in sys_kill

From: Nick Desaulniers
Date: Thu Jan 09 2020 - 12:17:10 EST

On Thu, Jan 9, 2020 at 8:23 AM 'Alexander Potapenko' via Clang Built
Linux <clang-built-linux@xxxxxxxxxxxxxxxx> wrote:
>
> On Thu, Jan 9, 2020 at 11:39 AM Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> >
> > On Thu, Jan 9, 2020 at 11:05 AM Dmitry Vyukov <dvyukov@xxxxxxxxxx> 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]

Is there automated reporting? Consider adding our mailing list for
Clang specific failures.
clang-built-linux <clang-built-linux@xxxxxxxxxxxxxxxx>
Our CI looks green, but there's a very long tail of combinations of
configs that we don't have coverage of, so bug reports are
appreciated:
https://github.com/ClangBuiltLinux/linux/issues

> > > > >
> > > > > 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.

I disabled loop unrolling and loop unswitching in LLVM when the loop
contained asm goto in:
https://github.com/llvm/llvm-project/commit/c4f245b40aad7e8627b37a8bf1bdcdbcd541e665
I have a fix for loop unrolling in:
https://reviews.llvm.org/D64101
that I should dust off. I haven't looked into loop unswitching yet.

> >
> >
> > Kernel built with clang version 10.0.0
> > (https://github.com/llvm/llvm-project.git
> > c2443155a0fb245c8f17f2c1c72b6ea391e86e81) works fine.
> >
> > Alex, please update clang on syzbot machines.
>
> Done ~3 hours ago, guess we'll see the results within a day.

Please let me know if you otherwise encounter any miscompiles with
Clang, particularly `asm goto` I treat as P0.
--
Thanks,
~Nick Desaulniers