> On Jun 7, 2019, at 9:45 AM, Yu-cheng Yu <yu-cheng.yu@xxxxxxxxx> wrote:
>
> On Fri, 2019-06-07 at 09:35 -0700, Andy Lutomirski wrote:
>>> On Jun 7, 2019, at 9:23 AM, Yu-cheng Yu <yu-cheng.yu@xxxxxxxxx> wrote:
>>>
>>>>> On Fri, 2019-06-07 at 10:08 +0200, Peter Zijlstra wrote:
>>>>> On Thu, Jun 06, 2019 at 01:09:15PM -0700, Yu-cheng Yu wrote:
>>>>> Indirect Branch Tracking (IBT) provides an optional legacy code bitmap
>>>>> that allows execution of legacy, non-IBT compatible library by an
>>>>> IBT-enabled application. When set, each bit in the bitmap indicates
>>>>> one page of legacy code.
>>>>>
>>>>> The bitmap is allocated and setup from the application.
>>>>> +int cet_setup_ibt_bitmap(unsigned long bitmap, unsigned long size)
>>>>> +{
>>>>> + u64 r;
>>>>> +
>>>>> + if (!current->thread.cet.ibt_enabled)
>>>>> + return -EINVAL;
>>>>> +
>>>>> + if (!PAGE_ALIGNED(bitmap) || (size > TASK_SIZE_MAX))
>>>>> + return -EINVAL;
>>>>> +
>>>>> + current->thread.cet.ibt_bitmap_addr = bitmap;
>>>>> + current->thread.cet.ibt_bitmap_size = size;
>>>>> +
>>>>> + /*
>>>>> + * Turn on IBT legacy bitmap.
>>>>> + */
>>>>> + modify_fpu_regs_begin();
>>>>> + rdmsrl(MSR_IA32_U_CET, r);
>>>>> + r |= (MSR_IA32_CET_LEG_IW_EN | bitmap);
>>>>> + wrmsrl(MSR_IA32_U_CET, r);
>>>>> + modify_fpu_regs_end();
>>>>> +
>>>>> + return 0;
>>>>> +}
>>>>
>>>> So you just program a random user supplied address into the hardware.
>>>> What happens if there's not actually anything at that address or the
>>>> user munmap()s the data after doing this?
>>>
>>> This function checks the bitmap's alignment and size, and anything else is
>>> the
>>> app's responsibility. What else do you think the kernel should check?
>>>
>>
>> One might reasonably wonder why this state is privileged in the first place
>> and, given that, why we’re allowing it to be written like this.
>>
>> Arguably we should have another prctl to lock these values (until exec) as a
>> gardening measure.
>
> We can prevent the bitmap from being set more than once. I will test it.
>
I think it would be better to make locking an explicit opt-in.
