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

[Andy Lutomirski]

On Thu, Sep 23, 2021, at 4:22 PM, Luck, Tony wrote:
> On Thu, Sep 23, 2021 at 04:09:18PM -0700, Andy Lutomirski wrote:
>> On Mon, Sep 20, 2021, at 12:23 PM, Fenghua Yu wrote:
>
>> I think this is unnecessarily complicated because it's buying in to the
>> existing ISA misconception that PASID has anything to do with a task.
>> A PASID belongs to an mm, full stop.  Now the ISA is nasty and we have
>> tasks that have *noticed* that their mm has a PASID and tasks that have
>> not noticed this fact, but that should be irrelevant to essentially
>> everything except the fault handler.
>> 
>> So just refcount the thing the obvious way: take a reference when you
>> stick the PASID in the mm_struct and drop the reference in __mmdrop().
>> Problem solved.  You could probably drop it more aggressively in
>> __mmput(), and the comment explaining why is left as an exercise to the
>> reader -- if a kernel thread starts doing ENQCMD, we have worse things
>> to worry about :)
>
> That doesn't match well with the non-x86 usage of PASIDs. The code there
> bumps the reference count on each device bind, and decrements on each
> device unbind.

Can you elaborate on how that works?  Is there an architecture where there is a bona fide per task PASID?

>
> If we don't keep a reference count for each task that has IA32_PASID
> set up we could have this sequence
>
> 1) Process binds to a PASID capable device

Okay, so the mm has that PASID set up and a reference is taken.

> 2) Task uses ENQCMD, so PASID MSR is set up.

Yep.

> 3) Process unbinds the device, reference count on PASID
>    goes to zero. PASID is freed. PASID is reallocated to
>    another task.

It had better not.  We had an entire phone call in which we agreed that the entire lazy-MSR-setup approach only makes any sense if everyone pinky swears that an mm will *never* change its PASID once it has a PASID.

> 4) Task from step #2 uses ENQCMD to submit a descriptor
>    and device now processes virtual addresses based on mappings
>    in the new task.
>
> Now you might say that at step 3 we need to hunt down all the
> tasks that have PASID enabled and disabled ... but that's the
> same painful code that we avoided when we said that we would
> not make Linux hand out a PASID to all existing tasks in a
> process on the first bind operation.
>

Exactly.  Which means that the mm ought to pin that PASID for as long as it exists.  What am I missing?

Sure, one can invent a situation in which you start two threads, and one of those threads binds a device, does ENQCMD, unbinds the device, and exits.  Then the other thread *in the same mm* binds another device and gets a new PASID.  And it all works.  But I really don't think this special case is worth optimizing for.

> -Tony