Re: [PATCH v11 6/9] x86/cet: Add PTRACE interface for CET

[Andy Lutomirski]

> On Sep 2, 2020, at 7:53 PM, Yu, Yu-cheng <yu-cheng.yu@xxxxxxxxx> wrote:
> 
> On 9/2/2020 4:50 PM, Andy Lutomirski wrote:
>>>> On Sep 2, 2020, at 3:13 PM, Yu, Yu-cheng <yu-cheng.yu@xxxxxxxxx> wrote:
>>> 
>>> On 9/2/2020 1:03 PM, Jann Horn wrote:
>>>>> On Tue, Aug 25, 2020 at 2:30 AM Yu-cheng Yu <yu-cheng.yu@xxxxxxxxx> wrote:
>>>>> Add REGSET_CET64/REGSET_CET32 to get/set CET MSRs:
>>>>> 
>>>>>     IA32_U_CET (user-mode CET settings) and
>>>>>     IA32_PL3_SSP (user-mode Shadow Stack)
>>>> [...]
>>>>> diff --git a/arch/x86/kernel/fpu/regset.c b/arch/x86/kernel/fpu/regset.c
>>>> [...]
>>>>> +int cetregs_get(struct task_struct *target, const struct user_regset *regset,
>>>>> +               struct membuf to)
>>>>> +{
>>>>> +       struct fpu *fpu = &target->thread.fpu;
>>>>> +       struct cet_user_state *cetregs;
>>>>> +
>>>>> +       if (!boot_cpu_has(X86_FEATURE_SHSTK))
>>>>> +               return -ENODEV;
>>>>> +
>>>>> +       fpu__prepare_read(fpu);
>>>>> +       cetregs = get_xsave_addr(&fpu->state.xsave, XFEATURE_CET_USER);
>>>>> +       if (!cetregs)
>>>>> +               return -EFAULT;
>>>> Can this branch ever be hit without a kernel bug? If yes, I think
>>>> -EFAULT is probably a weird error code to choose here. If no, this
>>>> should probably use WARN_ON(). Same thing in cetregs_set().
>>> 
>>> When a thread is not CET-enabled, its CET state does not exist.  I looked at EFAULT, and it means "Bad address".  Maybe this can be ENODEV, which means "No such device"?

Having read the code, I’m unconvinced. It looks like a get_xsave_addr() failure means “state not saved; task sees INIT state”.  So *maybe* it’s reasonable -ENODEV this, but I’m not really convinced. I tend to think we should return the actual INIT state and that we should permit writes and handle them correctly.

Dave, what do you think?

>>> 
>>> [...]
>>> 
>>>>> @@ -1284,6 +1293,13 @@ static struct user_regset x86_32_regsets[] __ro_after_init = {
>>>> [...]
>>>>> +       [REGSET_CET32] = {
>>>>> +               .core_note_type = NT_X86_CET,
>>>>> +               .n = sizeof(struct cet_user_state) / sizeof(u64),
>>>>> +               .size = sizeof(u64), .align = sizeof(u64),
>>>>> +               .active = cetregs_active, .regset_get = cetregs_get,
>>>>> +               .set = cetregs_set
>>>>> +       },
>>>>>  };
>>>> Why are there different identifiers for 32-bit CET and 64-bit CET when
>>>> they operate on the same structs and have the same handlers? If
>>>> there's a good reason for that, the commit message should probably
>>>> point that out.
>>> 
>>> Yes, the reason for two regsets is that fill_note_info() does not expect any holes in a regsets.  I will put this in the commit log.
>>> 
>>> 
>> Perhaps we could fix that instead?
> 
> As long as we understand the root cause, leaving it as-is may be OK.

The regset mechanism’s interactions with compat are awful. Let’s please not make it worse.  One CET regret is good; two is not good.

> 
> I had a patch in the past, but did not follow up on it.
> 
> https://lore.kernel.org/lkml/20180717162502.32274-1-yu-cheng.yu@xxxxxxxxx/
> 
> Yu-cheng