Re: [PATCH v2 3/8] media: videobuf2: Add a module param to limit vb2 queue buffer storage

[Laurent Pinchart <laurent.pinchart@ideasonboard.com>]

Hi Tomasz,

On Thu, Jun 08, 2023 at 07:24:29PM +0900, Tomasz Figa wrote:
> On Wed, May 31, 2023 at 9:39 PM Laurent Pinchart wrote:
> > On Wed, May 31, 2023 at 10:30:36AM +0200, Hans Verkuil wrote:
> > > On 5/31/23 10:03, Laurent Pinchart wrote:
> > > > On Wed, May 31, 2023 at 08:36:59AM +0200, Hans Verkuil wrote:
> > > >> On 21/03/2023 11:28, Benjamin Gaignard wrote:
> > > >>> Add module parameter "max_vb_buffer_per_queue" to be able to limit
> > > >>> the number of vb2 buffers store in queue.
> > > >>>
> > > >>> Signed-off-by: Benjamin Gaignard <benjamin.gaignard@collabora.com>
> > > >>> ---
> > > >>>  drivers/media/common/videobuf2/videobuf2-core.c | 15 +++------------
> > > >>>  include/media/videobuf2-core.h                  | 11 +++++++++--
> > > >>>  2 files changed, 12 insertions(+), 14 deletions(-)
> > > >>>
> > > >>> diff --git a/drivers/media/common/videobuf2/videobuf2-core.c b/drivers/media/common/videobuf2/videobuf2-core.c
> > > >>> index ae9d72f4d181..f4da917ccf3f 100644
> > > >>> --- a/drivers/media/common/videobuf2/videobuf2-core.c
> > > >>> +++ b/drivers/media/common/videobuf2/videobuf2-core.c
> > > >>> @@ -34,6 +34,8 @@
> > > >>>  static int debug;
> > > >>>  module_param(debug, int, 0644);
> > > >>>
> > > >>> +module_param(max_vb_buffer_per_queue, ulong, 0644);
> > > >>
> > > >> There is no MODULE_PARM_DESC here? Please add. I see it is not there for
> > > >> the debug param either, it should be added for that as well.
> > > >
> > > > Would this be the right time to consider resource accounting in V4L2 for
> > > > buffers ? Having a module parameter doesn't sound very useful, an
> > > > application could easily allocate more buffers by using buffer orphaning
> > > > (allocating buffers, exporting them as dmabuf objects, and freeing them,
> > > > which leaves the memory allocated). Repeating allocation cycles up to
> > > > max_vb_buffer_per_queue will allow allocating an unbounded number of
> > > > buffers, using all the available system memory. I'd rather not add a
> > > > module argument that only gives the impression of some kind of safety
> > > > without actually providing any value.
> 
> Good point. It's even simpler, just keep opening new vim2m instances
> and requesting max buffers :).
> 
> > >
> > > Does dmabuf itself provide some accounting mechanism? Just wondering.
> > >
> > > More specific to V4L2: I'm not so sure about this module parameter either.
> > > It makes sense to have a check somewhere against ridiculous values (i.e.
> > > allocating MAXINT buffers), but that can be a define as well. But otherwise
> > > I am fine with allowing applications to allocate buffers until the memory
> > > is full.
> > >
> > > The question is really: what is this parameter supposed to do? The only
> > > thing it does is to sanitize unlikely inputs (e.g. allocating MAXINT buffers).
> > >
> > > I prefer that as a define, to be honest.
> > >
> > > I think it is perfectly fine for users to try to request more buffers than
> > > memory allows. It will just fail in that case, not a problem.
> > >
> > > And if an application is doing silly things like buffer orphaning, then so
> > > what? Is that any different than allocating memory and not freeing it?
> > > Eventually it will run out of memory and crash, which is normal.
> >
> > Linux provides APIs to account for and limit usage of resources,
> > including memory. A system administrator can prevent rogue processes
> > from starving system resources. The memory consumed by vb2 buffer isn't
> > taken into account, making V4L2 essentially unsafe for untrusted
> > processes.
> 
> I agree that proper accounting would be useful, although I wouldn't
> really make this patch series depend on it, since it's not introducing
> the loophole in the first place.

No disagreement here, my concern was about introducing a workaround for
the lack of proper memory accounting. I'd like to avoid the workaround,
but it doesn't mean memory accounting has to be implement now.

> We had some discussion about this in ChromeOS long ago and we thought
> it would be really useful for killing browser tabs with big videos,
> but otherwise using very little regular memory (e.g. via javascript).
> 
> One challenge with accounting V4L2 allocations is how to count shared
> DMA-bufs. If one process allocates a V4L2 buffer, exports it to
> DMA-buf and then sends it to another process that keeps it alive, but
> frees the V4L2 buffer (and even closes the DMA-buf fd), should that
> memory be still accounted to it even though it doesn't hold a
> reference to it anymore?

I've thought about that too. It's an annoying problem, it should
probably be discussed with memory management developers.

> > Now, to be fair, there are many reasons why allowing access to v4L2
> > devices to untrusted applications is a bad idea, and memory consumption
> > is likely not even the worst one. Still, is this something we want to
> > fix, or do we want to consider V4L2 to be priviledged API only ? Right
> > now we can't do so, but with many Linux systems moving towards pipewire,
> > we could possibly have a system daemon isolating untrusted applications
> > from the rest of the system. We may thus not need to fix this in the
> > V4L2 API.

-- 
Regards,

Laurent Pinchart