target/linux/bcm27xx/patches-5.10/950-0615-media-rpivid-Map-cmd-buffer-directly.patch

   1 From d4f8e47d60c180cf57eba4093a343230a824ecbf Mon Sep 17 00:00:00 2001
   2 From: John Cox <jc@kynesim.co.uk>
   3 Date: Mon, 29 Mar 2021 17:42:16 +0100
   4 Subject: [PATCH] media: rpivid: Map cmd buffer directly
   5
   6 It is unnecessary to have a separate dmabuf to hold the cmd buffer.
   7 Map it directly from the kmalloc.
   8
   9 Signed-off-by: John Cox <jc@kynesim.co.uk>
  10 ---
  11  drivers/staging/media/rpivid/rpivid.h      |  3 +-
  12  drivers/staging/media/rpivid/rpivid_h265.c | 48 ++++++++++------------
  13  drivers/staging/media/rpivid/rpivid_hw.c   |  2 +
  14  3 files changed, 25 insertions(+), 28 deletions(-)
  15
  16 --- a/drivers/staging/media/rpivid/rpivid.h
  17 +++ b/drivers/staging/media/rpivid/rpivid.h
  18 @@ -114,7 +114,6 @@ struct rpivid_ctx {
  19         unsigned int p1idx;
  20         atomic_t p1out;
  21         struct rpivid_gptr bitbufs[RPIVID_P1BUF_COUNT];
  22 -       struct rpivid_gptr cmdbufs[RPIVID_P1BUF_COUNT];
  23
  24         /* *** Should be in dev *** */
  25         unsigned int p2idx;
  26 @@ -183,6 +182,8 @@ struct rpivid_dev {
  27         struct clk              *clock;
  28         struct clk_request      *hevc_req;
  29
  30 +       int                     cache_align;
  31 +
  32         struct rpivid_hw_irq_ctrl ic_active1;
  33         struct rpivid_hw_irq_ctrl ic_active2;
  34  };
  35 --- a/drivers/staging/media/rpivid/rpivid_h265.c
  36 +++ b/drivers/staging/media/rpivid/rpivid_h265.c
  37 @@ -227,6 +227,9 @@ struct rpivid_dec_env {
  38         struct rpivid_q_aux *frame_aux;
  39         struct rpivid_q_aux *col_aux;
  40
  41 +       dma_addr_t cmd_addr;
  42 +       size_t cmd_size;
  43 +
  44         dma_addr_t pu_base_vc;
  45         dma_addr_t coeff_base_vc;
  46         u32 pu_stride;
  47 @@ -234,7 +237,6 @@ struct rpivid_dec_env {
  48
  49         struct rpivid_gptr *bit_copy_gptr;
  50         size_t bit_copy_len;
  51 -       struct rpivid_gptr *cmd_copy_gptr;
  52
  53  #define SLICE_MSGS_MAX (2 * HEVC_MAX_REFS * 8 + 3)
  54         u16 slice_msgs[SLICE_MSGS_MAX];
  55 @@ -1499,22 +1501,17 @@ static int write_cmd_buffer(struct rpivi
  56                             struct rpivid_dec_env *const de,
  57                             const struct rpivid_dec_state *const s)
  58  {
  59 -       // Copy commands out to dma buf
  60 -       const size_t cmd_size = de->cmd_len * sizeof(de->cmd_fifo[0]);
  61 -
  62 -       if (!de->cmd_copy_gptr->ptr || cmd_size > de->cmd_copy_gptr->size) {
  63 -               size_t cmd_alloc = round_up_size(cmd_size);
  64 +       const size_t cmd_size = ALIGN(de->cmd_len * sizeof(de->cmd_fifo[0]),
  65 +                                     dev->cache_align);
  66
  67 -               if (gptr_realloc_new(dev, de->cmd_copy_gptr, cmd_alloc)) {
  68 -                       v4l2_err(&dev->v4l2_dev,
  69 -                                "Alloc cmd buffer (%zu): FAILED\n", cmd_alloc);
  70 -                       return -ENOMEM;
  71 -               }
  72 -               v4l2_info(&dev->v4l2_dev, "Alloc cmd buffer (%zu): OK\n",
  73 -                         cmd_alloc);
  74 +       de->cmd_addr = dma_map_single(dev->dev, de->cmd_fifo,
  75 +                                     cmd_size, DMA_TO_DEVICE);
  76 +       if (dma_mapping_error(dev->dev, de->cmd_addr)) {
  77 +               v4l2_err(&dev->v4l2_dev,
  78 +                        "Map cmd buffer (%zu): FAILED\n", cmd_size);
  79 +               return -ENOMEM;
  80         }
  81 -
  82 -       memcpy(de->cmd_copy_gptr->ptr, de->cmd_fifo, cmd_size);
  83 +       de->cmd_size = cmd_size;
  84         return 0;
  85  }
  86
  87 @@ -1551,6 +1548,12 @@ static void dec_env_delete(struct rpivid
  88         struct rpivid_ctx * const ctx = de->ctx;
  89         unsigned long lock_flags;
  90
  91 +       if (de->cmd_size) {
  92 +               dma_unmap_single(ctx->dev->dev, de->cmd_addr, de->cmd_size,
  93 +                                DMA_TO_DEVICE);
  94 +               de->cmd_size = 0;
  95 +       }
  96 +
  97         aux_q_release(ctx, &de->frame_aux);
  98         aux_q_release(ctx, &de->col_aux);
  99
 100 @@ -1603,7 +1606,8 @@ static int dec_env_init(struct rpivid_ct
 101
 102                 de->ctx = ctx;
 103                 de->decode_order = i;
 104 -               de->cmd_max = 1024;
 105 +//             de->cmd_max = 1024;
 106 +               de->cmd_max = 8096;
 107                 de->cmd_fifo = kmalloc_array(de->cmd_max,
 108                                              sizeof(struct rpi_cmd),
 109                                              GFP_KERNEL);
 110 @@ -1748,7 +1752,6 @@ static void rpivid_h265_setup(struct rpi
 111
 112                 de->bit_copy_gptr = ctx->bitbufs + ctx->p1idx;
 113                 de->bit_copy_len = 0;
 114 -               de->cmd_copy_gptr = ctx->cmdbufs + ctx->p1idx;
 115
 116                 de->frame_c_offset = ctx->dst_fmt.height * 128;
 117                 de->frame_stride = ctx->dst_fmt.plane_fmt[0].bytesperline * 128;
 118 @@ -2356,7 +2359,7 @@ static void phase1_claimed(struct rpivid
 119         rpivid_hw_irq_active1_irq(dev, &de->irq_ent, phase1_cb, de);
 120
 121         // And start the h/w
 122 -       apb_write_vc_addr_final(dev, RPI_CFBASE, de->cmd_copy_gptr->addr);
 123 +       apb_write_vc_addr_final(dev, RPI_CFBASE, de->cmd_addr);
 124
 125         xtrace_ok(dev, de);
 126         return;
 127 @@ -2400,8 +2403,6 @@ static void rpivid_h265_stop(struct rpiv
 128
 129         for (i = 0; i != ARRAY_SIZE(ctx->bitbufs); ++i)
 130                 gptr_free(dev, ctx->bitbufs + i);
 131 -       for (i = 0; i != ARRAY_SIZE(ctx->cmdbufs); ++i)
 132 -               gptr_free(dev, ctx->cmdbufs + i);
 133         for (i = 0; i != ARRAY_SIZE(ctx->pu_bufs); ++i)
 134                 gptr_free(dev, ctx->pu_bufs + i);
 135         for (i = 0; i != ARRAY_SIZE(ctx->coeff_bufs); ++i)
 136 @@ -2451,13 +2452,6 @@ static int rpivid_h265_start(struct rpiv
 137                 goto fail;
 138         }
 139
 140 -       // 16k is plenty for most purposes but we will realloc if needed
 141 -       for (i = 0; i != ARRAY_SIZE(ctx->cmdbufs); ++i) {
 142 -               if (gptr_alloc(dev, ctx->cmdbufs + i, 0x4000,
 143 -                              DMA_ATTR_FORCE_CONTIGUOUS))
 144 -                       goto fail;
 145 -       }
 146 -
 147         // Finger in the air PU & Coeff alloc
 148         // Will be realloced if too small
 149         coeff_alloc = round_up_size(wxh);
 150 --- a/drivers/staging/media/rpivid/rpivid_hw.c
 151 +++ b/drivers/staging/media/rpivid/rpivid_hw.c
 152 @@ -331,6 +331,8 @@ int rpivid_hw_probe(struct rpivid_dev *d
 153         if (IS_ERR(dev->clock))
 154                 return PTR_ERR(dev->clock);
 155
 156 +       dev->cache_align = dma_get_cache_alignment();
 157 +
 158         // Disable IRQs & reset anything pending
 159         irq_write(dev, 0,
 160                   ARG_IC_ICTRL_ACTIVE1_EN_SET | ARG_IC_ICTRL_ACTIVE2_EN_SET);