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

[Ben Dooks]

On 18/03/2021 15:18, Dmitry Vyukov wrote:
> On Mon, Mar 15, 2021 at 3:41 PM Ben Dooks <ben.dooks at codethink.co.uk> wrote:
>>
>> On 15/03/2021 11:52, Dmitry Vyukov wrote:
>>> On Mon, Mar 15, 2021 at 12:30 PM Ben Dooks <ben.dooks at codethink.co.uk> wrote:
>>>>
>>>> On 14/03/2021 11:03, Dmitry Vyukov wrote:
>>>>> On Sun, Mar 14, 2021 at 11:01 AM Dmitry Vyukov <dvyukov at google.com> wrote:
>>>>>>> On Wed, Mar 10, 2021 at 7:28 PM syzbot
>>>>>>> <syzbot+c23c5421600e9b454849 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=122c343ad00000
>>>>>>>> kernel config:  https://syzkaller.appspot.com/x/.config?x=e3c595255fb2d136
>>>>>>>> dashboard link: https://syzkaller.appspot.com/bug?extid=c23c5421600e9b454849
>>>>>>>> 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+c23c5421600e9b454849 at syzkaller.appspotmail.com
>>>>>>>
>>>>>>> +riscv maintainers
>>>>>>>
>>>>>>> Another case of put_user crashing.
>>>>>>
>>>>>> There are 58 crashes in sock_ioctl already. Somehow there is a very
>>>>>> significant skew towards crashing with this "user memory without
>>>>>> uaccess routines" in schedule_tail and sock_ioctl of all places in the
>>>>>> kernel that use put_user... This looks very strange... Any ideas
>>>>>> what's special about these 2 locations?
>>>>>
>>>>> I could imagine if such a crash happens after a previous stack
>>>>> overflow and now task data structures are corrupted. But f_getown does
>>>>> not look like a function that consumes way more than other kernel
>>>>> syscalls...
>>>>
>>>> The last crash I looked at suggested somehow put_user got re-entered
>>>> with the user protection turned back on. Either there is a path through
>>>> one of the kernel handlers where this happens or there's something
>>>> weird going on with qemu.
>>>
>>> Is there any kind of tracking/reporting that would help to localize
>>> it? I could re-reproduce with that code.
>>
>> I'm not sure. I will have a go at debugging on qemu today just to make
>> sure I can reproduce here before I have to go into the office and fix
>> my Icicle board for real hardware tests.
>>
>> I think my first plan post reproduction is to stuff some trace points
>> into the fault handlers to see if we can get a idea of faults being
>> processed, etc.
>>
>> Maybe also add a check in the fault handler to see if the fault was
>> in a fixable region and post an error if that happens / maybe retry
>> the instruction with the relevant SR_SUM flag set.
>>
>> Hopefully tomorrow I can get a run on real hardware to confirm.
>> Would have been better if the Unmatched board I ordered last year
>> would turn up.
> 
> In retrospect it's obvious what's common between these 2 locations:
> they both call a function inside of put_user.
> 
> #syz dup:
> BUG: unable to handle kernel access to user memory in schedule_tail

I think so. I've posted a patch that you can test, which should force
the flags to be saved over switch_to(). I think the sanitisers are just
making it easier to see.

There is a seperate issue of passing complicated things to put_user()
as for security, the function may be executed with the user-space
protections turned off. I plan to raise this on the kernel list later
once I've done some more testing.

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

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