[syzbot] BUG: unable to handle kernel access to user memory in schedule_tail

[Ben Dooks]

On 12/03/2021 15:12, Dmitry Vyukov wrote:
> On Fri, Mar 12, 2021 at 2:50 PM Ben Dooks <ben.dooks at codethink.co.uk> wrote:
>>
>> On 10/03/2021 17:16, Dmitry Vyukov wrote:
>>> On Wed, Mar 10, 2021 at 5:46 PM syzbot
>>> <syzbot+e74b94fe601ab9552d69 at syzkaller.appspotmail.com> wrote:
>>>>
>>>> Hello,
>>>>
>>>> syzbot found the following issue on:
>>>>
>>>> HEAD commit:    0d7588ab riscv: process: Fix no prototype for arch_dup_tas..
>>>> git tree:       git://git.kernel.org/pub/scm/linux/kernel/git/riscv/linux.git fixes
>>>> console output: https://syzkaller.appspot.com/x/log.txt?x=1212c6e6d00000
>>>> kernel config:  https://syzkaller.appspot.com/x/.config?x=e3c595255fb2d136
>>>> dashboard link: https://syzkaller.appspot.com/bug?extid=e74b94fe601ab9552d69
>>>> userspace arch: riscv64
>>>>
>>>> Unfortunately, I don't have any reproducer for this issue yet.
>>>>
>>>> IMPORTANT: if you fix the issue, please add the following tag to the commit:
>>>> Reported-by: syzbot+e74b94fe601ab9552d69 at syzkaller.appspotmail.com
>>>
>>> +riscv maintainers
>>>
>>> This is riscv64-specific.
>>> I've seen similar crashes in put_user in other places. It looks like
>>> put_user crashes in the user address is not mapped/protected (?).
>>
>> I've been having a look, and this seems to be down to access of the
>> tsk->set_child_tid variable. I assume the fuzzing here is to pass a
>> bad address to clone?
>>
>>   From looking at the code, the put_user() code should have set the
>> relevant SR_SUM bit (the value for this, which is 1<<18 is in the
>> s2 register in the crash report) and from looking at the compiler
>> output from my gcc-10, the code looks to be dong the relevant csrs
>> and then csrc around the put_user
>>
>> So currently I do not understand how the above could have happened
>> over than something re-tried the code seqeunce and ended up retrying
>> the faulting instruction without the SR_SUM bit set.
> 
> I would maybe blame qemu for randomly resetting SR_SUM, but it's
> strange that 99% of these crashes are in schedule_tail. If it would be
> qemu, then they would be more evenly distributed...
> 
> Another observation: looking at a dozen of crash logs, in none of
> these cases fuzzer was actually trying to fuzz clone with some insane
> arguments. So it looks like completely normal clone's (e..g coming
> from pthread_create) result in this crash.
> 
> I also wonder why there is ret_from_exception, is it normal? I see
> handle_exception disables SR_SUM:
> https://elixir.bootlin.com/linux/v5.12-rc2/source/arch/riscv/kernel/entry.S#L73

So I think if SR_SUM is set, then it faults the access to user memory
which the _user() routines clear to allow them access.

I'm thinking there is at least one issue here:

- the test in fault is the wrong way around for die kernel
- the handler only catches this if the page has yet to be mapped.

So I think the test should be:

         if (!user_mode(regs) && addr < TASK_SIZE &&
                         unlikely(regs->status & SR_SUM)

This then should continue on and allow the rest of the handler to
complete mapping the page if it is not there.

I have been trying to create a very simple clone test, but so far it
has yet to actually trigger anything.

-- 
Ben Dooks				http://www.codethink.co.uk/
Senior Engineer				Codethink - Providing Genius

https://www.codethink.co.uk/privacy.html