[Intel-gfx] [PATCH] drm/i915: avoid leaking DMA mappings

[Imre Deak]

On ma, 2015-07-06 at 17:04 +0100, Chris Wilson wrote:
> On Mon, Jul 06, 2015 at 06:56:00PM +0300, Imre Deak wrote:
> > On ma, 2015-07-06 at 16:33 +0100, Chris Wilson wrote:
> > > On Mon, Jul 06, 2015 at 05:29:39PM +0200, Daniel Vetter wrote:
> > > > On Mon, Jul 06, 2015 at 03:57:44PM +0100, Chris Wilson wrote:
> > > > > On Mon, Jul 06, 2015 at 05:50:37PM +0300, Imre Deak wrote:
> > > > > > We have 3 types of DMA mappings for GEM objects:
> > > > > > 1. physically contiguous for stolen and for objects needing contiguous
> > > > > >    memory
> > > > > > 2. DMA-buf mappings imported via a DMA-buf attach operation
> > > > > > 3. SG DMA mappings for shmem backed and userptr objects
> > > > > > 
> > > > > > For 1. and 2. the lifetime of the DMA mapping matches the lifetime of the
> > > > > > corresponding backing pages and so in practice we create/release the
> > > > > > mapping in the object's get_pages/put_pages callback.
> > > > > > 
> > > > > > For 3. the lifetime of the mapping matches that of any existing GPU binding
> > > > > > of the object, so we'll create the mapping when the object is bound to
> > > > > > the first vma and release the mapping when the object is unbound from its
> > > > > > last vma.
> > > > > > 
> > > > > > Since the object can be bound to multiple vmas, we can end up creating a
> > > > > > new DMA mapping in the 3. case even if the object already had one. This
> > > > > > is not allowed by the DMA API and can lead to leaked mapping data and
> > > > > > IOMMU memory space starvation in certain cases. For example HW IOMMU
> > > > > > drivers (intel_iommu) allocate a new range from their memory space
> > > > > > whenever a mapping is created, silently overriding a pre-existing
> > > > > > mapping.
> > > > 
> > > > How does this happen? Essentially list_empty(obj->vmas) ==
> > > > !dma_mapping_exists should hold for objects of the 3rd type. I don't
> > > > understand how this is broken in the current code. There was definitely
> > > > versions of the ppgtt code where this wasn't working properly, but I
> > > > thought we've fixed that up again.
> > > 
> > > Every g/ppgtt binding remapped the obj->pages through the iommu. Even
> > > with the DMAR disabled, we still pay the cpu cost of sw iommu (which is
> > > itself an annoying kernel bug that you can't disable).
> > >  
> > > > > > Fix this by adding new callbacks to create/release the DMA mapping. This
> > > > > > way we can use the has_dma_mapping flag for objects of the 3. case also
> > > > > > (so far the flag was only used for the 1. and 2. case) and skip creating
> > > > > > a new mapping if one exists already.
> > > > > > 
> > > > > > Note that I also thought about simply creating/releasing the mapping
> > > > > > when get_pages/put_pages is called. However since creating a DMA mapping
> > > > > > may have associated resources (at least in case of HW IOMMU) it does
> > > > > > make sense to release these resources as early as possible. We can
> > > > > > release the DMA mapping as soon as the object is unbound from the last
> > > > > > vma, before we drop the backing pages, hence it's worth keeping the two
> > > > > > operations separate.
> > > > > > 
> > > > > > I noticed this issue by enabling DMA debugging, which got disabled after
> > > > > > a while due to its internal mapping tables getting full. It also reported
> > > > > > errors in connection to random other drivers that did a DMA mapping for
> > > > > > an address that was previously mapped by i915 but was never released.
> > > > > > Besides these diagnostic messages and the memory space starvation
> > > > > > problem for IOMMUs, I'm not aware of this causing a real issue.
> > > > > 
> > > > > Nope, it is much much simpler. Since we only do the dma prepare/finish
> > > > > from inside get_pages/put_pages, we can put the calls there. The only
> > > > > caveat there is userptr worker, but that can be easily fixed up.
> > > > 
> > > > I do kinda like the distinction between just grabbing the backing storage
> > > > and making it accessible to the hw. Small one, but I think it does help if
> > > > we keep these two maps separate. Now the function names otoh are
> > > > super-confusing, that I agree with.
> > > 
> > > But that is the raison-d'etre of get_pages(). We call it preciselly when
> > > we want the backing storage available to the hw. We relaxed that for
> > > set-domain to avoid one type of bug, and stolen/dma-buf have their own
> > > notion of dma mapping. userptr is the odd one out due to its worker
> > > asynchronously grabbing the pages.
> > 
> > Isn't the DMA mapping operation more tied to binding the object to a
> > VMA? As far as I can see we call put_pages only when destroying the
> > object (or attaching a physically contiguous mapping to it) and that's
> > because at that point we also give up on the content of the buffer.
> > Otherwise we just do unbinding when reclaiming memory. At this point it
> > make sense to release the DMA mapping independently of releasing the
> > buffer contents.
> 
> No. As proved above, it is not about each VMA, it about preparing the
> object for access by the hw - i.e. a natural fit for the
> get_pages/put_pages() greedy scheme, and if you look at the workloads
> where we benefit from the current scheme, we also massively benefit from
> avoiding the remapping. A dma shrinker would also simply call
> i915_gem_shrink(), and we can do that today cf get_pages_gtt() and do
> our own shrinking first.

Right, misunderstood this. Adding new callbacks doesn't have a benefit
then.

--Imre