objtool clac/stac handling change..

[Christophe Leroy]

Le 03/07/2020 à 05:17, Michael Ellerman a écrit :
> Christophe Leroy <christophe.leroy at csgroup.eu> writes:
>> Le 02/07/2020 à 15:34, Michael Ellerman a écrit :
>>> Linus Torvalds <torvalds at linux-foundation.org> writes:
>>>> On Wed, Jul 1, 2020 at 12:59 PM Al Viro <viro at zeniv.linux.org.uk> wrote:
>>>>> On Wed, Jul 01, 2020 at 12:04:36PM -0700, Linus Torvalds wrote:
>>>>>>
>>>>>> That's actually for the access granting. Shutting the access down ends
>>>>>> up always doing the same thing anyway..
>>>>>
>>>>> #define user_read_access_end            prevent_current_read_from_user
>>>>> #define user_write_access_end           prevent_current_write_to_user
>>>>> static inline void prevent_current_read_from_user(void)
>>>>> {
>>>>>           prevent_user_access(NULL, NULL, ~0UL, KUAP_CURRENT_READ);
>>>>> }
>>>>>
>>>>> static inline void prevent_current_write_to_user(void)
>>>>> {
>>>>>           prevent_user_access(NULL, NULL, ~0UL, KUAP_CURRENT_WRITE);
>>>>> }
>>>>>
>>>>> and prevent_user_access() has instances that do care about the direction...
>>>>
>>>> Go and look closer.
>>>>
>>>> There are three cases:
>>>>
>>>>    (a) the 32-bit book3s case. It looks like it cares, but when you look
>>>> closer, it ends up not caring about the read side, and saving the
>>>> "which address to I allow user writes to" in current->thread.kuap
>>>>
>>>>    (b) the nohash 32-bit case - doesn't care
>>>>
>>>>    (c) the 64-bit books case - doesn't care
>>>>
>>>> So yes, in the (a) case it does make a difference between reads and
>>>> writes, but at least as far as I can tell, it ignores the read case,
>>>> and has code to avoid the unnecessary "disable user writes" case when
>>>> there was only a read enable done.
>>>
>>> Yeah that's my understanding too.
>>>
>>> Christophe is the expert on that code so I'll defer to him if I'm wrong.
>>>
>>>> Now, it's possible that I'm wrong, but the upshot of that is that even
>>>> on powerpc, I think that if we just made the rule be that "taking a
>>>> user exception should automatically do the 'user_access_end()' for us"
>>>> is trivial.
>>>
>>> I think we can do something to make it work.
>>>
>>> We don't have an equivalent of x86's ex_handler_uaccess(), so it's not
>>> quite as easy as whacking a user_access_end() in there.
>>
>> Isn't it something easy to do in bad_page_fault() ?
> 
> We'd need to do it there at least.
> 
> But I'm not convinced that's the only place we'd need to do it. We could
> theoretically take a machine check on a user access, and those are
> handled differently on each sub-(sub-sub)-platform, and I think all or
> most of them don't call bad_page_fault().

Indeed, it needs to be done everywhere we do

	regs->nip = extable_fixup(entry)

There are half a dozen of places that do that, in additional of 
bad_page_fault() that's mainly machine checks, also kprobe.

I think we can create a fixup_exception() function which takes regs and 
entry as parameters and does the nip fixup and kuap closuse.

> 
>> Not exactly a call to user_access_end() but altering regs->kuap so that
>> user access is not restored on exception exit.
> 
> Yes.
> 
>>> Probably the simplest option for us is to just handle it in our
>>> unsafe_op_wrap(). I'll try and come up with something tomorrow.
>>
>> unsafe_op_wrap() is not used anymore for unsafe_put_user() as we are now
>> using asm goto.
> 
> Sure, but we could change it back to use unsafe_op_wrap().

But the whole purpose of using goto in unsafe_???_user() is to allow the 
use of asm goto. See explanations in commit 
https://github.com/linuxppc/linux/commit/1bd4403d86a1c06cb6cc9ac87664a0c9d3413d51#diff-eba084de047bb8a9087dac10c06f44bc


> 
> I did a quick hack to do that and see no difference in the generated
> code, but your commit adding put_user_goto() did show better code
> generation, so possibly it depends on compiler version, or my example
> wasn't complicated enough (filldir()).

Yes as explained above it should remove the error checking in the caller 
so your exemple was most likely too trivial.

Christophe