Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Date: Fri, 29 Sep 2017 00:17:25 +0300
From: Alexander Popov <alex.popov@...ux.com>
To: Tycho Andersen <tycho@...ker.com>, Kees Cook <keescook@...omium.org>
Cc: "kernel-hardening@...ts.openwall.com"
 <kernel-hardening@...ts.openwall.com>, PaX Team <pageexec@...email.hu>,
 Brad Spengler <spender@...ecurity.net>, Laura Abbott <labbott@...hat.com>,
 Mark Rutland <mark.rutland@....com>,
 Ard Biesheuvel <ard.biesheuvel@...aro.org>, "x86@...nel.org"
 <x86@...nel.org>, Andy Lutomirski <luto@...capital.net>
Subject: Re: [PATCH RFC v3 1/1] gcc-plugins: Add stackleak feature erasing the
 kernel stack at the end of syscalls

On 21.09.2017 16:26, Alexander Popov wrote:
> On 20.09.2017 17:13, Tycho Andersen wrote:
>> On Wed, Sep 20, 2017 at 02:27:05PM +0300, Alexander Popov wrote:
>>>> +/*
>>>> + * Note that the way this test fails is kind of ugly; it hits the BUG() in
>>>> + * track_stack, but then the BUG() handler blows the stack and hits the stack
>>>> + * guard page.
>>>> + */
>>>
>>> Yes, actually, the reason is deeper.
>>>
>>> When there are less than (THREAD_SIZE / 16) bytes left in the kernel stack, the
>>> BUG() in track_stack() is hit. But do_error_trap(), which handles the invalid
>>> opcode, has a big stack frame. So it is instrumented by the STACKLEAK gcc plugin
>>> and itself calls track_stack() at the beginning. Hence we have a recursive
>>> BUG(), which eventually hits the guard page.
>>>
>>> I banned the instrumentation of do_error_trap() in the plugin, but it didn't
>>> really help, since there are several other instrumented functions called during
>>> BUG() handling.
>>>
>>> So it seems to me that this BUG() in track_stack() is really useless and can be
>>> dropped. Moreover:
>>>  - it is not a part of the PaX patch;
>>>  - it never worked in Grsecurity kernel because of the error spotted by Tycho.
>>>
>>> What do you think about it?
>>
>> We'll only have a stack guard page in the case of vmap stack, so maybe
> 
> Thanks, that's an important aspect.
> 
>> we can do:
>>
>> diff --git a/fs/exec.c b/fs/exec.c
>> index 8333c4dce59b..8351369cd1e4 100644
>> --- a/fs/exec.c
>> +++ b/fs/exec.c
>> @@ -1960,7 +1960,8 @@ void __used track_stack(void)
>>  		current->thread.lowest_stack = sp;
>>  	}
>>  
>> -	if (unlikely((sp & (THREAD_SIZE - 1)) < (THREAD_SIZE / 16)))
>> +	if (!IS_ENABLED(CONFIG_VMAP_STACK) &&
>> +	    unlikely((sp & (THREAD_SIZE - 1)) < (THREAD_SIZE / 16)))
>>  		BUG();
>>  }
>>  EXPORT_SYMBOL(track_stack);
> 
> In that case the recursive BUG() in track_stack() will happen anyway. You know,
> I would better make CONFIG_GCC_PLUGIN_STACKLEAK depend on CONFIG_VMAP_STACK.

That turned out to be a bad idea. Unfortunately, VMAP_STACK is not available on
x86_32, but STACKLEAK works on that platform. So I'll put the check behind
#ifdef. Maybe having it is better than having a silent stack overflow.

Best regards,
Alexander

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.