Re: [RFC PATCH v1 2/2] randomize_kstack: Unify random source across arches

[Ryan Roberts <ryan.roberts@arm.com>]

On 02/12/2025 09:53, Mark Rutland wrote:
> On Tue, Dec 02, 2025 at 09:39:40AM +0000, Ryan Roberts wrote:
>> On 02/12/2025 09:35, Mark Rutland wrote:
>>> On Tue, Dec 02, 2025 at 10:15:22AM +0100, Ard Biesheuvel wrote:
>>>> On Mon, 1 Dec 2025 at 19:20, Ryan Roberts <ryan.roberts@arm.com> wrote:
>>>>> On 28/11/2025 11:01, Ard Biesheuvel wrote:
>>>>>> On Thu, 27 Nov 2025 at 12:00, Ryan Roberts <ryan.roberts@arm.com> wrote:
>>>>>>> diff --git a/include/linux/randomize_kstack.h b/include/linux/randomize_kstack.h
>>>>>>> index 089b1432f7e6..83c7e6710f6d 100644
>>>>>>> --- a/include/linux/randomize_kstack.h
>>>>>>> +++ b/include/linux/randomize_kstack.h
>>>>>>> @@ -6,6 +6,7 @@
>>>>>>>  #include <linux/kernel.h>
>>>>>>>  #include <linux/jump_label.h>
>>>>>>>  #include <linux/percpu-defs.h>
>>>>>>> +#include <linux/prandom.h>
>>>>>>>
>>>>>>>  DECLARE_STATIC_KEY_MAYBE(CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT,
>>>>>>>                          randomize_kstack_offset);
>>>>>>> @@ -45,9 +46,13 @@ DECLARE_STATIC_KEY_MAYBE(CONFIG_RANDOMIZE_KSTACK_OFFSET_DEFAULT,
>>>>>>>  #define KSTACK_OFFSET_MAX(x)   ((x) & 0b1111111100)
>>>>>>>  #endif
>>>>>>>
>>>>>>> +static __always_inline u32 get_update_kstack_offset(void)
>>>>>>> +{
>>>>>>> +       return prandom_u32_state(&current->kstack_rnd_state);
>>>>>
>>>>> I've got bot warnings because this is being called from noinstr code. I guess
>>>>> the best option is to just move add_random_kstack_offset() to after
>>>>> instrumentation is enabled for the affected arches.
>>>>
>>>> Just put instrumentation_begin()/instrumentation_end() around the call
>>>> to prandom_u32_state() - that seems to be the common approach for
>>>> punching holes into the 'noinstr' validation.
>>>
>>> That silences the warning, but isn't necessarily safe, so please DO NOT
>>> do that blindly. The instrumentation_{begin,end}() annotations are only
>>> supposed to be used when we know by construction that instrumentation is
>>> safe.
>>>
>>> Generally, if you can move this to after instrumentation is already
>>> enabled, that should be safe, and so that'd be the better approach.
>>>
>>> Ryan, can you share those warnings (e.g. link to those reports)?
>>
>> https://lore.kernel.org/oe-kbuild-all/202511302137.u4iIA2kf-lkp@intel.com/
>> https://lore.kernel.org/oe-kbuild-all/202511302317.BGFIlAYX-lkp@intel.com/
> 
> Cool, so those are __do_fast_syscall_32() and do_syscall_64(), which I'd
> expect (and those are legitimate warnings).
> 
> From a quick scan of v6.18, it looks like arm64, loongarch, powerpc, and
> riscv perform all the necessary entry work before calling
> add_random_kstack_offset(), and we'll need to change:
> 
> * s390's __do_syscall()
> * x86's do_syscall_64()
> * x86's do_int80_syscall_32()
> * x86's __do_fast_syscall_32()

Yup. I'll have a play with inline vs outline. If I go with outline, I'll fix
these sites up.

> 
> Mark.