media: rpi: hevc_dec: allow up to 64 capture buffers

The capture queue does not set max_num_buffers, so videobuf2 limits it to
VB2_MAX_FRAME (32) buffers, and the per-buffer auxiliary (colocated
motion vector) table is sized to match. Every reference frame must be a
capture buffer, so a userspace that keeps decoded frames queued for
display or further processing can exhaust that limit, and
VIDIOC_CREATE_BUFS then fails with -ENOBUFS.

Kodi's Dolby Vision profile 7 FEL playback holds up to 16 base layer
pictures while it waits for the matching enhancement layer pictures, on
top of the HEVC DPB and its render queue, which exceeds 32 buffers for
the 4K base layer. Size the auxiliary table for 64 entries and allow 64
capture buffers. Buffers and their auxiliary data are still allocated on
demand, and the decoder is behind an IOMMU, so they do not use CMA.

Assisted-by: Claude:claude-opus-5-5
Signed-off-by: Christian Hewitt <christianshewitt@gmail.com>

Author: chewitt

drivers/media/platform/raspberrypi/hevc_dec/hevc_d.h

@@ -94,7 +94,7 @@ struct hevc_d_dev;
 typedef void (*hevc_d_irq_callback)(struct hevc_d_dev *dev, void *ctx);
 
 struct hevc_d_q_aux;
-#define HEVC_D_AUX_ENT_COUNT VB2_MAX_FRAME
+#define HEVC_D_AUX_ENT_COUNT 64
 
 struct hevc_d_ctx {
 	struct v4l2_fh			fh;

drivers/media/platform/raspberrypi/hevc_dec/hevc_d_video.c

@@ -642,6 +642,7 @@ int hevc_d_queue_init(void *priv, struct vb2_queue *src_vq,
 	dst_vq->drv_priv = ctx;
 	dst_vq->buf_struct_size = sizeof(struct hevc_d_buffer);
 	dst_vq->min_queued_buffers = 1;
+	dst_vq->max_num_buffers = HEVC_D_AUX_ENT_COUNT;
 	dst_vq->ops = &hevc_d_qops;
 	dst_vq->mem_ops = &vb2_dma_contig_memops;
 	dst_vq->timestamp_flags = V4L2_BUF_FLAG_TIMESTAMP_COPY;