pages pinned for BO lifetime and security

[Daniel Vetter]

On Wed, Jul 22, 2020 at 9:19 AM Christian König
<christian.koenig at amd.com> wrote:
>
> Am 22.07.20 um 02:22 schrieb Gurchetan Singh:
>
> +Christian who added DMABUF_MOVE_NOTIFY which added the relevant blurb:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/dma-buf/Kconfig#n46
>
> Currently, the user seems to amdgpu for P2P dma-buf and it seems to plumb ttm (*move_notify) callback to dma-buf.  We're not sure if it's a security issue occurring across DRM drivers, or one more specific to the new amdgpu use case.
>
> On Tue, Jul 21, 2020 at 1:03 PM Chia-I Wu <olvaffe at gmail.com> wrote:
>>
>> Hi list,
>>
>> virtio-gpu is moving in the direction where BO pages are pinned for
>> the lifetime for simplicity.  I am wondering if that is considered a
>> security issue in general, especially after running into the
>> description of the new DMABUF_MOVE_NOTIFY config option.
>
>
> Yes, that is generally considered a deny of service possibility and so far Dave and Daniel have rejected all tries to upstream stuff like this as far as I know.

Uh we have merged pretty much all arm-soc drivers without real
shrinkers. Whether that was a good idea or not is maybe different
question - now that we do have pretty good helpers maybe we should
poke this a bit more. But then SoCs Suck (tm).

But for real gpus they do indeed all have shrinkers, and not just "pin
everything forever" model. Real gpus = stuff you might run on servers
or multi-app and all that stuff, not with a simple "we just kill all
background jobs if memory gets low" model like on android and other
such things.

> DMA-buf an pinning for scanout are the only exceptions since the implementation wouldn't have been possible otherwise.
>
>>
>> Most drivers do not have a shrinker, or whether a BO is purgeable is
>> entirely controlled by the userspace (madvice).  They can be
>> categorized as "a security problem where userspace is able to pin
>> unrestricted amounts of memory".  But those drivers are normally found
>> on systems without swap.  I don't think the issue applies.
>
>
> This is completely independent of the availability of swap or not.
>
> Pinning of pages in large quantities can result in all kind of problems and needs to be prevented even without swap.

Yeah you don't just kill swap, you kill a ton of other kernel services
with mass pinning. I think even the pinning of scanout buffers for
i915 from system memory is somewhat questionable (but I guess small
enough to not matter in practice).

> Otherwise you can ran into problems even with simple I/O operations for example.
>
>>
>> Of the desktop GPU drivers, i915's shrinker certainly supports purging
>> to swap.  TTM is a bit hard to follow.  I can't really tell if amdgpu
>> or nouveau supports that.  virtio-gpu is more commonly found on
>> systems with swaps so I think it should follow the desktop practices?
>
>
> What we do at least in the amdgpu, radeon, i915 and nouveau is to only allow it for scanout and that in turn is limited by the physical number of CRTCs on the board.
>
>>
>> Truth is, the emulated virtio-gpu device always supports page moves
>> with VIRTIO_GPU_CMD_RESOURCE_{ATTACH,DETACH}_BACKING.  It is just that
>> the driver does not make use of them.  That makes this less of an
>> issue because the driver can be fixed anytime (finger crossed that the
>> emulator won't have bugs in these untested paths).  This issue becomes
>> more urgent because we are considering adding a new HW command[1]
>> where page moves will be disallowed.  We definitely don't want a HW
>> command that is inherently insecure, if BO pages pinned for the
>> lifetime is considered a security issue on desktops.
>
>
> Yeah, that's probably not such a good idea :)

Well if the pinning is just for the duration of the hw command, it's
fine, just like batch buffers. But if it's long term pinning then that
doesn't sound like a good idea. RDMA has this as their inherit hw
programming model (except if your hw is really fancy and has hw page
fault handling on the rdma nic), and they hard limit such pins to what
you can mlock (or something similar within rdma).
-Daniel

>
> Regards,
> Christian.
>
>>
>> [1] VIRTIO_GPU_CMD_RESOURCE_CREATE_BLOB
>> https://gitlab.freedesktop.org/virgl/drm-misc-next/-/blob/virtio-gpu-next/include/uapi/linux/virtio_gpu.h#L396
>
>
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel



-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch