Re: [PATCH v2 3/5] spi: spi-mtk-nor: use dma_alloc_coherent() for bounce buffer

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi!

On Fri, Sep 18, 2020 at 4:35 PM Ikjoon Jang <ikjn@xxxxxxxxxxxx> wrote:
>
> Use dma_alloc_coherent() for bounce buffer instead of kmalloc.

The commit message should explain why such a change is
needed. (i.e. why using dma_alloc_coherent here is better
than kmalloc.) And if there's no benefit for this change I'd prefer
leaving it untouched.
I remembered reading somewhere that stream DMA api is
prefered over dma_alloc_coherent for this kind of single-direction
DMA operation.

>
> Signed-off-by: Ikjoon Jang <ikjn@xxxxxxxxxxxx>
>
> ---
>
> (no changes since v1)
>
>  drivers/spi/spi-mtk-nor.c | 60 +++++++++++++++++++++++----------------
>  1 file changed, 35 insertions(+), 25 deletions(-)
>
> diff --git a/drivers/spi/spi-mtk-nor.c b/drivers/spi/spi-mtk-nor.c
> index 54b2c0fde95b..e14798a6e7d0 100644
> --- a/drivers/spi/spi-mtk-nor.c
> +++ b/drivers/spi/spi-mtk-nor.c
> @@ -96,6 +96,7 @@ struct mtk_nor {
>         struct device *dev;
>         void __iomem *base;
>         u8 *buffer;
> +       dma_addr_t buffer_dma;
>         struct clk *spi_clk;
>         struct clk *ctlr_clk;
>         unsigned int spi_freq;
> @@ -275,19 +276,16 @@ static void mtk_nor_setup_bus(struct mtk_nor *sp, const struct spi_mem_op *op)
>         mtk_nor_rmw(sp, MTK_NOR_REG_BUSCFG, reg, MTK_NOR_BUS_MODE_MASK);
>  }
>
> -static int mtk_nor_read_dma(struct mtk_nor *sp, u32 from, unsigned int length,
> -                           u8 *buffer)
> +static int read_dma(struct mtk_nor *sp, u32 from, unsigned int length,

This name is a bit confusing considering there's a mtk_nor_read_dma
below.
As this function now only executes dma readings and wait it to finish,
what about mtk_nor_dma_exec instead?

> +                   dma_addr_t dma_addr)
>  {
>         int ret = 0;
>         ulong delay;
>         u32 reg;
> -       dma_addr_t dma_addr;
>
> -       dma_addr = dma_map_single(sp->dev, buffer, length, DMA_FROM_DEVICE);
> -       if (dma_mapping_error(sp->dev, dma_addr)) {
> -               dev_err(sp->dev, "failed to map dma buffer.\n");
> +       if (WARN_ON((length & MTK_NOR_DMA_ALIGN_MASK) ||
> +                   (dma_addr & MTK_NOR_DMA_ALIGN_MASK)))

These alignment is guaranteed by callers of this function if all
my comments below are addressed. This check isn't needed.

>                 return -EINVAL;
> -       }
>
>         writel(from, sp->base + MTK_NOR_REG_DMA_FADR);
>         writel(dma_addr, sp->base + MTK_NOR_REG_DMA_DADR);
> @@ -312,30 +310,39 @@ static int mtk_nor_read_dma(struct mtk_nor *sp, u32 from, unsigned int length,
>                                          (delay + 1) * 100);
>         }
>
> -       dma_unmap_single(sp->dev, dma_addr, length, DMA_FROM_DEVICE);
>         if (ret < 0)
>                 dev_err(sp->dev, "dma read timeout.\n");
>
>         return ret;
>  }
>
> -static int mtk_nor_read_bounce(struct mtk_nor *sp, u32 from,
> -                              unsigned int length, u8 *buffer)
> +static int mtk_nor_read_dma(struct mtk_nor *sp, u32 from,
> +                           unsigned int length, u8 *buffer)
>  {
> -       unsigned int rdlen;
>         int ret;
> +       dma_addr_t dma_addr;
> +       bool bounce = need_bounce(buffer, length);
>
> -       if (length & MTK_NOR_DMA_ALIGN_MASK)
> -               rdlen = (length + MTK_NOR_DMA_ALIGN) & ~MTK_NOR_DMA_ALIGN_MASK;

The intention of this rdlen alignment is explained in 2/5.
Please make sure this rdlen alignment logic is present
only for PIO reading.

> -       else
> -               rdlen = length;
> +       if (!bounce) {
> +               dma_addr = dma_map_single(sp->dev, buffer, length,
> +                                         DMA_FROM_DEVICE);
> +               if (dma_mapping_error(sp->dev, dma_addr)) {
> +                       dev_err(sp->dev, "failed to map dma buffer.\n");
> +                       return -EINVAL;
> +               }
> +       } else {
> +               dma_addr = sp->buffer_dma;
> +       }
>
> -       ret = mtk_nor_read_dma(sp, from, rdlen, sp->buffer);
> -       if (ret)
> -               return ret;
> +       ret = read_dma(sp, from, length, dma_addr);
>
> -       memcpy(buffer, sp->buffer, length);
> -       return 0;
> +       if (!bounce)
> +               dma_unmap_single(sp->dev, dma_addr, length,
> +                                DMA_FROM_DEVICE);
> +       else
> +               memcpy(buffer, sp->buffer, length);
> +
> +       return ret;
>  }

I think a separated read_dma and read_bounce function will be
cleaner than this if-else implementation:
read_dma:
1. call dma_map_single to get physical address
2. call read_dma to execute operation
3. call dma_unmap_single

read_bounce:
1. align reading length
2. call read_dma
3. call memcpy

>
>  static int mtk_nor_read_pio(struct mtk_nor *sp, const struct spi_mem_op *op)
> @@ -439,11 +446,6 @@ static int mtk_nor_exec_op(struct spi_mem *mem, const struct spi_mem_op *op)
>                 if (op->data.nbytes == 1) {
>                         mtk_nor_set_addr(sp, op);
>                         return mtk_nor_read_pio(sp, op);
> -               } else if (((ulong)(op->data.buf.in) &
> -                           MTK_NOR_DMA_ALIGN_MASK)) {
> -                       return mtk_nor_read_bounce(sp, op->addr.val,
> -                                                  op->data.nbytes,
> -                                                  op->data.buf.in);
>                 } else {
>                         return mtk_nor_read_dma(sp, op->addr.val,
>                                                 op->data.nbytes,
> @@ -654,6 +656,10 @@ static int mtk_nor_probe(struct platform_device *pdev)
>         sp->dev = &pdev->dev;
>         sp->spi_clk = spi_clk;
>         sp->ctlr_clk = ctlr_clk;

There is extra memory allocation code for sp->buffer in mtk_nor_probe.
If you intend to replace this with dma_alloc_coherent you should
drop those devm_kmalloc code as well.

> +       sp->buffer = dma_alloc_coherent(&pdev->dev, MTK_NOR_BOUNCE_BUF_SIZE,
> +                                       &sp->buffer_dma, GFP_KERNEL);

There's a devm variant: dmam_alloc_coherent(dev, size, dma_handle, gfp)

> +       if (!sp->buffer)
> +               return -ENOMEM;

This spi-nor controller requires all addresses to be 16-byte aligned.
Although it should be guaranteed by a usually way larger page
alignment address from dma_alloc_coherent I'd prefer an explicit
check for address alignment here rather than letting it probe
successfully and fail for every dma_read with bounce buffer.


>
>         irq = platform_get_irq_optional(pdev, 0);
>         if (irq < 0) {
> @@ -674,6 +680,8 @@ static int mtk_nor_probe(struct platform_device *pdev)
>         ret = mtk_nor_init(sp);
>         if (ret < 0) {
>                 kfree(ctlr);
> +               dma_free_coherent(&pdev->dev, MTK_NOR_BOUNCE_BUF_SIZE,
> +                                 sp->buffer, sp->buffer_dma);
>                 return ret;
>         }
>
> @@ -692,6 +700,8 @@ static int mtk_nor_remove(struct platform_device *pdev)
>
>         mtk_nor_disable_clk(sp);
>
> +       dma_free_coherent(&pdev->dev, MTK_NOR_BOUNCE_BUF_SIZE,
> +                         sp->buffer, sp->buffer_dma);
>         return 0;
>  }
>
> --
> 2.28.0.681.g6f77f65b4e-goog
>


--
Regards,
Chuanhong Guo



[Index of Archives]     [Device Tree Compilter]     [Device Tree Spec]     [Linux Driver Backports]     [Video for Linux]     [Linux USB Devel]     [Linux PCI Devel]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Yosemite Backpacking]


  Powered by Linux