From 5db127a534e1aa381c3ba60f51455863f37ea96b Mon Sep 17 00:00:00 2001 From: Jernej Skrabec Date: Sun, 10 Oct 2021 11:45:14 +0100 Subject: [PATCH] media: cedrus: Don't kernel map most buffers Except for VP8 probability coefficients buffer, all other buffers are never accessed by CPU. That allows us to mark them with DMA_ATTR_NO_KERNEL_MAPPING flag. This helps with decoding big (like 4k) videos on 32-bit ARM platforms where default vmalloc size is relatively small - 240 MiB. Since auxiliary buffer are not yet efficiently allocated, this can be easily exceeded. Even if allocation is optimized, 4k videos will still often exceed this limit. Signed-off-by: Jernej Skrabec Signed-off-by: Hans Verkuil Signed-off-by: Mauro Carvalho Chehab --- drivers/staging/media/sunxi/cedrus/cedrus_h264.c | 113 ++++++++++++---------- drivers/staging/media/sunxi/cedrus/cedrus_h265.c | 30 +++--- drivers/staging/media/sunxi/cedrus/cedrus_video.c | 1 + 3 files changed, 82 insertions(+), 62 deletions(-) diff --git a/drivers/staging/media/sunxi/cedrus/cedrus_h264.c b/drivers/staging/media/sunxi/cedrus/cedrus_h264.c index de7442d..b4173a8 100644 --- a/drivers/staging/media/sunxi/cedrus/cedrus_h264.c +++ b/drivers/staging/media/sunxi/cedrus/cedrus_h264.c @@ -520,6 +520,11 @@ static int cedrus_h264_start(struct cedrus_ctx *ctx) unsigned int mv_col_size; int ret; + /* + * NOTE: All buffers allocated here are only used by HW, so we + * can add DMA_ATTR_NO_KERNEL_MAPPING flag when allocating them. + */ + /* Formula for picture buffer size is taken from CedarX source. */ if (ctx->src_fmt.width > 2048) @@ -538,23 +543,23 @@ static int cedrus_h264_start(struct cedrus_ctx *ctx) ctx->codec.h264.pic_info_buf_size = pic_info_size; ctx->codec.h264.pic_info_buf = - dma_alloc_coherent(dev->dev, ctx->codec.h264.pic_info_buf_size, - &ctx->codec.h264.pic_info_buf_dma, - GFP_KERNEL); + dma_alloc_attrs(dev->dev, ctx->codec.h264.pic_info_buf_size, + &ctx->codec.h264.pic_info_buf_dma, + GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING); if (!ctx->codec.h264.pic_info_buf) return -ENOMEM; /* * That buffer is supposed to be 16kiB in size, and be aligned - * on 16kiB as well. However, dma_alloc_coherent provides the - * guarantee that we'll have a CPU and DMA address aligned on - * the smallest page order that is greater to the requested - * size, so we don't have to overallocate. + * on 16kiB as well. However, dma_alloc_attrs provides the + * guarantee that we'll have a DMA address aligned on the + * smallest page order that is greater to the requested size, + * so we don't have to overallocate. */ ctx->codec.h264.neighbor_info_buf = - dma_alloc_coherent(dev->dev, CEDRUS_NEIGHBOR_INFO_BUF_SIZE, - &ctx->codec.h264.neighbor_info_buf_dma, - GFP_KERNEL); + dma_alloc_attrs(dev->dev, CEDRUS_NEIGHBOR_INFO_BUF_SIZE, + &ctx->codec.h264.neighbor_info_buf_dma, + GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING); if (!ctx->codec.h264.neighbor_info_buf) { ret = -ENOMEM; goto err_pic_buf; @@ -582,10 +587,11 @@ static int cedrus_h264_start(struct cedrus_ctx *ctx) mv_col_size = field_size * 2 * CEDRUS_H264_FRAME_NUM; ctx->codec.h264.mv_col_buf_size = mv_col_size; - ctx->codec.h264.mv_col_buf = dma_alloc_coherent(dev->dev, - ctx->codec.h264.mv_col_buf_size, - &ctx->codec.h264.mv_col_buf_dma, - GFP_KERNEL); + ctx->codec.h264.mv_col_buf = + dma_alloc_attrs(dev->dev, + ctx->codec.h264.mv_col_buf_size, + &ctx->codec.h264.mv_col_buf_dma, + GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING); if (!ctx->codec.h264.mv_col_buf) { ret = -ENOMEM; goto err_neighbor_buf; @@ -600,10 +606,10 @@ static int cedrus_h264_start(struct cedrus_ctx *ctx) ctx->codec.h264.deblk_buf_size = ALIGN(ctx->src_fmt.width, 32) * 12; ctx->codec.h264.deblk_buf = - dma_alloc_coherent(dev->dev, - ctx->codec.h264.deblk_buf_size, - &ctx->codec.h264.deblk_buf_dma, - GFP_KERNEL); + dma_alloc_attrs(dev->dev, + ctx->codec.h264.deblk_buf_size, + &ctx->codec.h264.deblk_buf_dma, + GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING); if (!ctx->codec.h264.deblk_buf) { ret = -ENOMEM; goto err_mv_col_buf; @@ -616,10 +622,10 @@ static int cedrus_h264_start(struct cedrus_ctx *ctx) ctx->codec.h264.intra_pred_buf_size = ALIGN(ctx->src_fmt.width, 64) * 5 * 2; ctx->codec.h264.intra_pred_buf = - dma_alloc_coherent(dev->dev, - ctx->codec.h264.intra_pred_buf_size, - &ctx->codec.h264.intra_pred_buf_dma, - GFP_KERNEL); + dma_alloc_attrs(dev->dev, + ctx->codec.h264.intra_pred_buf_size, + &ctx->codec.h264.intra_pred_buf_dma, + GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING); if (!ctx->codec.h264.intra_pred_buf) { ret = -ENOMEM; goto err_deblk_buf; @@ -629,24 +635,28 @@ static int cedrus_h264_start(struct cedrus_ctx *ctx) return 0; err_deblk_buf: - dma_free_coherent(dev->dev, ctx->codec.h264.deblk_buf_size, - ctx->codec.h264.deblk_buf, - ctx->codec.h264.deblk_buf_dma); + dma_free_attrs(dev->dev, ctx->codec.h264.deblk_buf_size, + ctx->codec.h264.deblk_buf, + ctx->codec.h264.deblk_buf_dma, + DMA_ATTR_NO_KERNEL_MAPPING); err_mv_col_buf: - dma_free_coherent(dev->dev, ctx->codec.h264.mv_col_buf_size, - ctx->codec.h264.mv_col_buf, - ctx->codec.h264.mv_col_buf_dma); + dma_free_attrs(dev->dev, ctx->codec.h264.mv_col_buf_size, + ctx->codec.h264.mv_col_buf, + ctx->codec.h264.mv_col_buf_dma, + DMA_ATTR_NO_KERNEL_MAPPING); err_neighbor_buf: - dma_free_coherent(dev->dev, CEDRUS_NEIGHBOR_INFO_BUF_SIZE, - ctx->codec.h264.neighbor_info_buf, - ctx->codec.h264.neighbor_info_buf_dma); + dma_free_attrs(dev->dev, CEDRUS_NEIGHBOR_INFO_BUF_SIZE, + ctx->codec.h264.neighbor_info_buf, + ctx->codec.h264.neighbor_info_buf_dma, + DMA_ATTR_NO_KERNEL_MAPPING); err_pic_buf: - dma_free_coherent(dev->dev, ctx->codec.h264.pic_info_buf_size, - ctx->codec.h264.pic_info_buf, - ctx->codec.h264.pic_info_buf_dma); + dma_free_attrs(dev->dev, ctx->codec.h264.pic_info_buf_size, + ctx->codec.h264.pic_info_buf, + ctx->codec.h264.pic_info_buf_dma, + DMA_ATTR_NO_KERNEL_MAPPING); return ret; } @@ -654,23 +664,28 @@ static void cedrus_h264_stop(struct cedrus_ctx *ctx) { struct cedrus_dev *dev = ctx->dev; - dma_free_coherent(dev->dev, ctx->codec.h264.mv_col_buf_size, - ctx->codec.h264.mv_col_buf, - ctx->codec.h264.mv_col_buf_dma); - dma_free_coherent(dev->dev, CEDRUS_NEIGHBOR_INFO_BUF_SIZE, - ctx->codec.h264.neighbor_info_buf, - ctx->codec.h264.neighbor_info_buf_dma); - dma_free_coherent(dev->dev, ctx->codec.h264.pic_info_buf_size, - ctx->codec.h264.pic_info_buf, - ctx->codec.h264.pic_info_buf_dma); + dma_free_attrs(dev->dev, ctx->codec.h264.mv_col_buf_size, + ctx->codec.h264.mv_col_buf, + ctx->codec.h264.mv_col_buf_dma, + DMA_ATTR_NO_KERNEL_MAPPING); + dma_free_attrs(dev->dev, CEDRUS_NEIGHBOR_INFO_BUF_SIZE, + ctx->codec.h264.neighbor_info_buf, + ctx->codec.h264.neighbor_info_buf_dma, + DMA_ATTR_NO_KERNEL_MAPPING); + dma_free_attrs(dev->dev, ctx->codec.h264.pic_info_buf_size, + ctx->codec.h264.pic_info_buf, + ctx->codec.h264.pic_info_buf_dma, + DMA_ATTR_NO_KERNEL_MAPPING); if (ctx->codec.h264.deblk_buf_size) - dma_free_coherent(dev->dev, ctx->codec.h264.deblk_buf_size, - ctx->codec.h264.deblk_buf, - ctx->codec.h264.deblk_buf_dma); + dma_free_attrs(dev->dev, ctx->codec.h264.deblk_buf_size, + ctx->codec.h264.deblk_buf, + ctx->codec.h264.deblk_buf_dma, + DMA_ATTR_NO_KERNEL_MAPPING); if (ctx->codec.h264.intra_pred_buf_size) - dma_free_coherent(dev->dev, ctx->codec.h264.intra_pred_buf_size, - ctx->codec.h264.intra_pred_buf, - ctx->codec.h264.intra_pred_buf_dma); + dma_free_attrs(dev->dev, ctx->codec.h264.intra_pred_buf_size, + ctx->codec.h264.intra_pred_buf, + ctx->codec.h264.intra_pred_buf_dma, + DMA_ATTR_NO_KERNEL_MAPPING); } static void cedrus_h264_trigger(struct cedrus_ctx *ctx) diff --git a/drivers/staging/media/sunxi/cedrus/cedrus_h265.c b/drivers/staging/media/sunxi/cedrus/cedrus_h265.c index 3d9561d..8829a7b 100644 --- a/drivers/staging/media/sunxi/cedrus/cedrus_h265.c +++ b/drivers/staging/media/sunxi/cedrus/cedrus_h265.c @@ -350,11 +350,12 @@ static void cedrus_h265_setup(struct cedrus_ctx *ctx, ctx->codec.h265.mv_col_buf_size = num_buffers * ctx->codec.h265.mv_col_buf_unit_size; + /* Buffer is never accessed by CPU, so we can skip kernel mapping. */ ctx->codec.h265.mv_col_buf = - dma_alloc_coherent(dev->dev, - ctx->codec.h265.mv_col_buf_size, - &ctx->codec.h265.mv_col_buf_addr, - GFP_KERNEL); + dma_alloc_attrs(dev->dev, + ctx->codec.h265.mv_col_buf_size, + &ctx->codec.h265.mv_col_buf_addr, + GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING); if (!ctx->codec.h265.mv_col_buf) { ctx->codec.h265.mv_col_buf_size = 0; // TODO: Abort the process here. @@ -667,10 +668,11 @@ static int cedrus_h265_start(struct cedrus_ctx *ctx) /* The buffer size is calculated at setup time. */ ctx->codec.h265.mv_col_buf_size = 0; + /* Buffer is never accessed by CPU, so we can skip kernel mapping. */ ctx->codec.h265.neighbor_info_buf = - dma_alloc_coherent(dev->dev, CEDRUS_H265_NEIGHBOR_INFO_BUF_SIZE, - &ctx->codec.h265.neighbor_info_buf_addr, - GFP_KERNEL); + dma_alloc_attrs(dev->dev, CEDRUS_H265_NEIGHBOR_INFO_BUF_SIZE, + &ctx->codec.h265.neighbor_info_buf_addr, + GFP_KERNEL, DMA_ATTR_NO_KERNEL_MAPPING); if (!ctx->codec.h265.neighbor_info_buf) return -ENOMEM; @@ -682,16 +684,18 @@ static void cedrus_h265_stop(struct cedrus_ctx *ctx) struct cedrus_dev *dev = ctx->dev; if (ctx->codec.h265.mv_col_buf_size > 0) { - dma_free_coherent(dev->dev, ctx->codec.h265.mv_col_buf_size, - ctx->codec.h265.mv_col_buf, - ctx->codec.h265.mv_col_buf_addr); + dma_free_attrs(dev->dev, ctx->codec.h265.mv_col_buf_size, + ctx->codec.h265.mv_col_buf, + ctx->codec.h265.mv_col_buf_addr, + DMA_ATTR_NO_KERNEL_MAPPING); ctx->codec.h265.mv_col_buf_size = 0; } - dma_free_coherent(dev->dev, CEDRUS_H265_NEIGHBOR_INFO_BUF_SIZE, - ctx->codec.h265.neighbor_info_buf, - ctx->codec.h265.neighbor_info_buf_addr); + dma_free_attrs(dev->dev, CEDRUS_H265_NEIGHBOR_INFO_BUF_SIZE, + ctx->codec.h265.neighbor_info_buf, + ctx->codec.h265.neighbor_info_buf_addr, + DMA_ATTR_NO_KERNEL_MAPPING); } static void cedrus_h265_trigger(struct cedrus_ctx *ctx) diff --git a/drivers/staging/media/sunxi/cedrus/cedrus_video.c b/drivers/staging/media/sunxi/cedrus/cedrus_video.c index 6671460..3372617 100644 --- a/drivers/staging/media/sunxi/cedrus/cedrus_video.c +++ b/drivers/staging/media/sunxi/cedrus/cedrus_video.c @@ -568,6 +568,7 @@ int cedrus_queue_init(void *priv, struct vb2_queue *src_vq, src_vq->type = V4L2_BUF_TYPE_VIDEO_OUTPUT; src_vq->io_modes = VB2_MMAP | VB2_DMABUF; + src_vq->dma_attrs = DMA_ATTR_NO_KERNEL_MAPPING; src_vq->drv_priv = ctx; src_vq->buf_struct_size = sizeof(struct cedrus_buffer); src_vq->ops = &cedrus_qops; -- 2.7.4