> On Jul 11, 2018, at 2:51 PM, Jann Horn <jannh@xxxxxxxxxx> wrote:
>
> On Wed, Jul 11, 2018 at 2:34 PM Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>>> On Jul 11, 2018, at 2:10 PM, Jann Horn <jannh@xxxxxxxxxx> wrote:
>>>
>>>> On Tue, Jul 10, 2018 at 3:31 PM Yu-cheng Yu <yu-cheng.yu@xxxxxxxxx> wrote:
>>>>
>>>> This patch adds basic shadow stack enabling/disabling routines.
>>>> A task's shadow stack is allocated from memory with VM_SHSTK
>>>> flag set and read-only protection. The shadow stack is
>>>> allocated to a fixed size.
>>>>
>>>> Signed-off-by: Yu-cheng Yu <yu-cheng.yu@xxxxxxxxx>
>>> [...]
>>>> diff --git a/arch/x86/kernel/cet.c b/arch/x86/kernel/cet.c
>>>> new file mode 100644
>>>> index 000000000000..96bf69db7da7
>>>> --- /dev/null
>>>> +++ b/arch/x86/kernel/cet.c
>>> [...]
>>>> +static unsigned long shstk_mmap(unsigned long addr, unsigned long len)
>>>> +{
>>>> + struct mm_struct *mm = current->mm;
>>>> + unsigned long populate;
>>>> +
>>>> + down_write(&mm->mmap_sem);
>>>> + addr = do_mmap(NULL, addr, len, PROT_READ,
>>>> + MAP_ANONYMOUS | MAP_PRIVATE, VM_SHSTK,
>>>> + 0, &populate, NULL);
>>>> + up_write(&mm->mmap_sem);
>>>> +
>>>> + if (populate)
>>>> + mm_populate(addr, populate);
>>>> +
>>>> + return addr;
>>>> +}
> [...]
>>> Should the kernel enforce that two shadow stacks must have a guard
>>> page between them so that they can not be directly adjacent, so that
>>> if you have too much recursion, you can't end up corrupting an
>>> adjacent shadow stack?
>>
>> I think the answer is a qualified “no”. I would like to instead enforce a general guard page on all mmaps that don’t use MAP_FORCE. We *might* need to exempt any mmap with an address hint for compatibility.
>
> I like this idea a lot.
>
>> My commercial software has been manually adding guard pages on every single mmap done by tcmalloc for years, and it has caught a couple bugs and costs essentially nothing.
>>
>> Hmm. Linux should maybe add something like Windows’ “reserved” virtual memory. It’s basically a way to ask for a VA range that explicitly contains nothing and can be subsequently be turned into something useful with the equivalent of MAP_FORCE.
>
> What's the benefit over creating an anonymous PROT_NONE region? That
> the kernel won't have to scan through the corresponding PTEs when
> tearing down the mapping?
Make it more obvious what’s happening and avoid accounting issues? What I’ve actually used is MAP_NORESERVE | PROT_NONE, but I think this still counts against the VA rlimit. But maybe that’s actually the desired behavior.
