Re: [PATCH v4 05/11] iommu/sva: Assign a PASID to mm on PASID allocation and free it on mm exit
From: Jean-Philippe Brucker
Date: Mon Apr 25 2022 - 12:13:43 EST
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.
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 .