[PATCH v2 2/8] media: videobuf2: Make bufs array dynamic allocated

[Tomasz Figa]

On Fri, Mar 24, 2023 at 09:56:34AM +0100, Benjamin Gaignard wrote:
> 
> Le 24/03/2023 à 09:52, Hans Verkuil a écrit :
> > On 24/03/2023 09:48, Laurent Pinchart wrote:
> > > On Fri, Mar 24, 2023 at 09:31:35AM +0100, Hans Verkuil wrote:
> > > > On 24/03/2023 09:11, Benjamin Gaignard wrote:
> > > > > Le 24/03/2023 à 06:01, Dan Carpenter a écrit :
> > > > > > Hi Benjamin,
> > > > > > 
> > > > > > https://git-scm.com/docs/git-format-patch#_base_tree_information]
> > > > > > 
> > > > > > url:    https://github.com/intel-lab-lkp/linux/commits/Benjamin-Gaignard/media-videobuf2-Access-vb2_queue-bufs-array-through-helper-functions/20230321-183154
> > > > > > base:   git://linuxtv.org/media_tree.git master
> > > > > > patch link:    https://lore.kernel.org/r/20230321102855.346732-3-benjamin.gaignard%40collabora.com
> > > > > > patch subject: [PATCH v2 2/8] media: videobuf2: Make bufs array dynamic allocated
> > > > > > config: arm64-randconfig-m041-20230319 (https://download.01.org/0day-ci/archive/20230324/202303240148.lKRnUqW9-lkp@intel.com/config)
> > > > > > compiler: aarch64-linux-gcc (GCC) 12.1.0
> > > > > > 
> > > > > > If you fix the issue, kindly add following tag where applicable
> > > > > > | Reported-by: kernel test robot <lkp at intel.com>
> > > > > > | Reported-by: Dan Carpenter <error27 at gmail.com>
> > > > > > | Link: https://lore.kernel.org/r/202303240148.lKRnUqW9-lkp@intel.com/
> > > > > > 
> > > > > > smatch warnings:
> > > > > > include/media/videobuf2-core.h:1272 vb2_queue_add_buffer() warn: sleeping in atomic context
> > > > > > drivers/media/common/videobuf2/videobuf2-core.c:2456 vb2_core_queue_init() warn: Please consider using kcalloc instead of kmalloc_array
> > > > > > 
> > > > > > vim +1272 include/media/videobuf2-core.h
> > > > > > 
> > > > > > 625d46c1c1fe8e Benjamin Gaignard 2023-03-21  1263  static inline bool vb2_queue_add_buffer(struct vb2_queue *q, struct vb2_buffer *vb)
> > > > > > 625d46c1c1fe8e Benjamin Gaignard 2023-03-21  1264  {
> > > > > > 487d3f14d12ecf Benjamin Gaignard 2023-03-21  1265      bool ret = false;
> > > > > > 487d3f14d12ecf Benjamin Gaignard 2023-03-21  1266
> > > > > > 487d3f14d12ecf Benjamin Gaignard 2023-03-21  1267      spin_lock(&q->bufs_lock);
> > > > > >                                                           ^^^^^^^^^^^^^^^^^^^^^^^
> > > > > > Holding a spin lock.
> > > > > > 
> > > > > > 487d3f14d12ecf Benjamin Gaignard 2023-03-21  1268
> > > > > > 487d3f14d12ecf Benjamin Gaignard 2023-03-21  1269      if (vb->index >= q->max_num_bufs) {
> > > > > > 487d3f14d12ecf Benjamin Gaignard 2023-03-21  1270          struct vb2_buffer **tmp;
> > > > > > 487d3f14d12ecf Benjamin Gaignard 2023-03-21  1271
> > > > > > 487d3f14d12ecf Benjamin Gaignard 2023-03-21 @1272          tmp = krealloc_array(q->bufs, q->max_num_bufs * 2, sizeof(*q->bufs), GFP_KERNEL);
> > > > > >                                                                                                                                        ^^^^^^^^^^
> > > > > > Sleeping allocation.  GFP_ATOMIC?  Or is there a way to move the
> > > > > > allocation outside the lock?
> > > > > I will add GFP_ATOMIC flag in next version.
> > > > No need. Instead, don't use realloc here, just allocate a new array, copy over all
> > > > the data from the old, and then switch q->bufs with the spinlock held. Then you
> > > > can free the old one.
> > > > 
> > > > It's only when you update q->bufs that you need the lock.
> > > The copy also needs to be protected by the lock.
> > I suspect that that is not needed, since you shouldn't be able to add buffers here
> > since a mutex should be held at this time.
> > 
> > That said, it's something that Benjamin needs to analyze.

I spent some time looking through the call sites of vb2_get_buffer() and
how those can be called and it turned out to be a massive list of code
paths, including a lot of calls originating from codec drivers calling
vb2_find_buffer() in random contexts (possibly even interrupt). So a
spinlock protecting the array pointer makes sense indeed.

> 
> Does using GFP_ATOMIC is problematic ?
> 

Yes, because the ability to reclaim memory is drastically limited and
the allocation is more likely to fail (as in: it's actually possible).
(And generally the time with interrupts disabled should be minimized to
keep system latency reasonable.)

Best regards,
Tomasz