Re: [PATCH RFC v2 03/29] mm: asi: Introduce ASI core API

[Junaid Shahid <junaids@google.com>]

On 3/18/25 6:03 AM, Brendan Jackman wrote:
> On Tue Mar 18, 2025 at 12:50 AM UTC, Junaid Shahid wrote:
>> On 3/17/25 4:40 AM, Brendan Jackman wrote:
>>>
>>> static inline void asi_start(void)
>>> {
>>> 	/*
>>> 	 * Cannot currently context switch in the restricted adddress
>>> 	 * space.
>>> 	 */
>>> 	lockdep_assert_preemption_disabled();
>>
>> I assume that this limitation is just for the initial version in this RFC,
>> right?
> 
> Well I think we also wanna get ASI in-tree with this limitation,
> otherwise the initial series will be too big and complex. 

Agreed. That is what I meant as well.

> 
>> But even in that case, I think this should be in asi_start_critical()
>> below, not asi_start(), since IIRC the KVM run loop does contain preemptible
>> code as well. And we would need an explicit asi_exit() in the context switch
>> code like we had in an earlier RFC.
> 
> Oh. Yeah. In my proposal below I had totally forgotten we had
> asi_exit() in the context_switch() path (it is there in this patch).
> 
> So we only need the asi_exit() in the KVM code in order to avoid
> actually hitting e.g. exit_to_user_mode() in the restricted address
> space.
> 
> But... we can just put an asi_exit() there explicitly instead of
> dumping all this weirdness into the "core API" and the KVM codebase.
> 
> So... I think all we really need is asi_start_critical() and
> asi_end_critical()? And make everything else happen as part of the
> normal functioning of the entry and context-switching logic. Am I
> forgetting something else?

Yes, I think this should work.


Thanks,
Junaid