Re: [PATCH 5/8] x86/mmu: Add mm-based PASID refcounting

[Andy Lutomirski <luto@kernel.org>]

On 9/29/21 05:28, Thomas Gleixner wrote:
> On Wed, Sep 29 2021 at 11:54, Peter Zijlstra wrote:
>> On Fri, Sep 24, 2021 at 04:03:53PM -0700, Andy Lutomirski wrote:
>>> I think the perfect and the good are a bit confused here. If we go for
>>> "good", then we have an mm owning a PASID for its entire lifetime.  If
>>> we want "perfect", then we should actually do it right: teach the
>>> kernel to update an entire mm's PASID setting all at once.  This isn't
>>> *that* hard -- it involves two things:
>>>
>>> 1. The context switch code needs to resync PASID.  Unfortunately, this
>>> adds some overhead to every context switch, although a static_branch
>>> could minimize it for non-PASID users.
>>
>>> 2. A change to an mm's PASID needs to sent an IPI, but that IPI can't
>>> touch FPU state.  So instead the IPI should use task_work_add() to
>>> make sure PASID gets resynced.
>>
>> What do we need 1 for? Any PASID change can be achieved using 2 no?
>>
>> Basically, call task_work_add() on all relevant tasks [1], then IPI
>> spray the current running of those and presto.
>>
>> [1] it is nigh on impossible to find all tasks sharing an mm in any sane
>> way due to CLONE_MM && !CLONE_THREAD.
> 
> Why would we want any of that at all?
> 
>    Process starts, no PASID assigned.
> 
>    bind to device -> PASID is allocated and assigned to the mm
> 
>    some task of the process issues ENQCMD -> #GP -> write PASID MSR
> 
>    After that the PASID is saved and restored as part of the XSTATE and
>    there is no extra overhead in context switch or return to user space.
> 
>    All tasks of the process which did never use ENQCMD don't care and their
>    PASID xstate is in init state.
> 
> There is absolutely no point in enforcing that all tasks of the process
> have the PASID activated immediately when it is assigned. If they need
> it they get it via the #GP fixup and everything just works.
> 
> Looking at that patch again, none of this muck in fpu__pasid_write() is
> required at all. The whole exception fixup is:
> 
>      if (!user_mode(regs))
>               return false;
> 
>      if (!current->mm->pasid)
>               return false;
> 
>      if (current->pasid_activated)
>      	     return false;

<-- preemption or BH here: kaboom.

> 
>      wrmsrl(MSR_IA32_PASID, current->mm->pasid);

This needs the actual sane fpstate writing helper -- see other email.