Re: [PATCH v4 05/11] iommu/sva: Assign a PASID to mm on PASID allocation and free it on mm exit

[Jacob Pan]

Hi Jean-Philippe,

On Mon, 25 Apr 2022 17:13:02 +0100, Jean-Philippe Brucker
<jean-philippe@xxxxxxxxxx> wrote:

> Hi Jacob,
> 
> On Mon, Apr 25, 2022 at 08:34:44AM -0700, Jacob Pan wrote:
> > Hi Jean-Philippe,
> > 
> > On Mon, 25 Apr 2022 15:26:40 +0100, Jean-Philippe Brucker
> > <jean-philippe@xxxxxxxxxx> wrote:
> >   
> > > On Mon, Apr 25, 2022 at 07:18:36AM -0700, Dave Hansen wrote:  
> > > > On 4/25/22 06:53, Jean-Philippe Brucker wrote:    
> > > > > On Sat, Apr 23, 2022 at 07:13:39PM +0800, zhangfei.gao@xxxxxxxxxxx
> > > > > wrote:    
> > > > >>>> On 5.17
> > > > >>>> fops_release is called automatically, as well as
> > > > >>>> iommu_sva_unbind_device. On 5.18-rc1.
> > > > >>>> fops_release is not called, have to manually call close(fd)    
> > > > >>> Right that's weird    
> > > > >> Looks it is caused by the fix patch, via mmget, which may add
> > > > >> refcount of fd.    
> > > > > Yes indirectly I think: when the process mmaps the queue,
> > > > > mmap_region() takes a reference to the uacce fd. That reference is
> > > > > released either by explicit close() or munmap(), or by exit_mmap()
> > > > > (which is triggered by mmput()). Since there is an mm->fd
> > > > > dependency, we cannot add a fd->mm dependency, so no
> > > > > mmget()/mmput() in bind()/unbind().
> > > > > 
> > > > > I guess we should go back to refcounted PASIDs instead, to avoid
> > > > > freeing them until unbind().    
> > > > 
> > > > Yeah, this is a bit gnarly for -rc4.  Let's just make sure there's
> > > > nothing else simple we can do.
> > > > 
> > > > How does the IOMMU hardware know that all activity to a given PASID
> > > > is finished?  That activity should, today, be independent of an mm
> > > > or a fd's lifetime.    
> > > 
> > > In the case of uacce, it's tied to the fd lifetime: opening an
> > > accelerator queue calls iommu_sva_bind_device(), which sets up the
> > > PASID context in the IOMMU. Closing the queue calls
> > > iommu_sva_unbind_device() which destroys the PASID context (after the
> > > device driver stopped all DMA for this PASID).
> > >   
> > For VT-d, it is essentially the same flow except managed by the
> > individual drivers such as DSA.
> > If free() happens before unbind(), we deactivate the PASIDs and suppress
> > faults from the device. When the unbind finally comes, we finalize the
> > PASID teardown. It seems we have a need for an intermediate state where
> > PASID is "pending free"?  
> 
> Yes we do have that state, though I'm not sure we need to make it explicit
> in the ioasid allocator.
> 
IMHO, making it explicit would fail ioasid_get() on a "pending free" PASID.
Making free a one-way trip and prevent further complications.

> Could we move mm_pasid_drop() to __mmdrop() instead of __mmput()?  For Arm
> we do need to hold the mm_count until unbind(), and mmgrab()/mmdrop() is
> also part of Lu's rework [1].
> 
Yes, I would agree. IIRC, Fenghua's early patch was doing pasid drop
in mmdrop. Maybe I missed something.

> Thanks,
> Jean
> 
> [1]
> https://lore.kernel.org/linux-iommu/20220421052121.3464100-9-baolu.lu@xxxxxxxxxxxxxxx/


Thanks,

Jacob