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

[Luck, Tony]

On Thu, Sep 23, 2021 at 04:36:50PM +0200, Thomas Gleixner wrote:
> On Mon, Sep 20 2021 at 19:23, Fenghua Yu wrote:
> >  
> > +#ifdef CONFIG_INTEL_IOMMU_SVM
> > +void pasid_put(struct task_struct *tsk, struct mm_struct *mm);
> > +#else
> > +static inline void pasid_put(struct task_struct *tsk, struct mm_struct *mm) { }
> > +#endif
> 
> This code is again defining that PASID is entirely restricted to
> INTEL. It's true, that no other vendor supports this, but PASID is
> a non-vendor specific concept.
> 
> Sticking this into INTEL code means that any other PASID implementation
> has to rip it out again from INTEL code and make it a run time property.
> 
> The refcounting issue should be the same for all PASID mechanisms which
> attach PASID to a mm. What's INTEL specific about that?
> 
> So can we pretty please do that correct right away?

It's a bit messy (surprise).

There are two reasons to hold a refcount on a PASID

1) The process has done a bind on a device that uses PASIDs

	This one isn't dependent on Intel.

2) A task within a process is using ENQCMD (and thus holds
   a reference on the PASID because IA32_PASID MSR for this
   task has the PASID value loaded with the enable bit set).

	This is (currently) Intel specific (until others
	implement an ENQCMD-like feature to allow apps to
	access PASID enabled devices without going through
	the OS).

Perhaps some better function naming might help?  E.g. have
a task_pasid_put() function that handles the process exit
case separatley from the device unbind case.

void task_pasid_put(void)
{
	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
		return;

	if (current->has_valid_pasid) {
		mutex_lock(&pasid_mutex);
		iommu_sva_free_pasid(mm);
		mutex_unlock(&pasid_mutex);
	}
}

-Tony