Re: [PATCH 2/3] mmc: tmio: implement a bounce buffer for unaligned DMA

[Chris Ball <cjb@laptop.org>]

Hi Guennadi, could you resend this one with some style changes:

On Tue, Nov 23, 2010 at 05:24:15PM +0100, Guennadi Liakhovetski wrote:
> For example, with SDIO WLAN cards, some transfers happen with buffers at odd
> addresses, whereas the SH-Mobile DMA engine requires even addresses for SDHI.
> This patch extends the tmio driver with a bounce buffer, that is used for
> single entry scatter-gather lists both for sending and receiving. If we ever
> encounter unaligned transfers with multi-element sg lists, this patch will have
> to be extended. For now it just falls back to PIO in this and other unsupported
> cases.
> 
> Signed-off-by: Guennadi Liakhovetski <g.liakhovetski@gmx.de>

Please wrap commit messages at 72 cols, not 80.

> ---
>  drivers/mmc/host/tmio_mmc.c |   81 +++++++++++++++++++++++++++++++++++++++----
>  include/linux/mfd/tmio.h    |    1 +
>  2 files changed, 75 insertions(+), 7 deletions(-)
> 
> diff --git a/drivers/mmc/host/tmio_mmc.c b/drivers/mmc/host/tmio_mmc.c
> index 118ad86..57ece9d 100644
> --- a/drivers/mmc/host/tmio_mmc.c
> +++ b/drivers/mmc/host/tmio_mmc.c
> @@ -111,6 +111,8 @@
>  		sd_ctrl_write32((host), CTL_STATUS, ~(i)); \
>  	} while (0)
>  
> +/* This is arbitrary, just noone needed any higher alignment yet */
> +#define MAX_ALIGN 4
>  
>  struct tmio_mmc_host {
>  	void __iomem *ctl;
> @@ -127,6 +129,7 @@ struct tmio_mmc_host {
>  
>  	/* pio related stuff */
>  	struct scatterlist      *sg_ptr;
> +	struct scatterlist      *sg_orig;
>  	unsigned int            sg_len;
>  	unsigned int            sg_off;
>  
> @@ -139,6 +142,8 @@ struct tmio_mmc_host {
>  	struct tasklet_struct	dma_issue;
>  #ifdef CONFIG_TMIO_MMC_DMA
>  	unsigned int            dma_sglen;
> +	u8			bounce_buf[PAGE_CACHE_SIZE] __attribute__((aligned(MAX_ALIGN)));
> +	struct scatterlist	bounce_sg;
>  #endif
>  };
>  
> @@ -180,6 +185,7 @@ static void tmio_mmc_init_sg(struct tmio_mmc_host *host, struct mmc_data *data)
>  {
>  	host->sg_len = data->sg_len;
>  	host->sg_ptr = data->sg;
> +	host->sg_orig = data->sg;
>  	host->sg_off = 0;
>  }
>  
> @@ -436,8 +442,14 @@ static void tmio_mmc_do_data_irq(struct tmio_mmc_host *host)
>  	 */
>  
>  	if (data->flags & MMC_DATA_READ) {
> -		if (!host->chan_rx)
> +		if (!host->chan_rx) {
>  			disable_mmc_irqs(host, TMIO_MASK_READOP);
> +		} else if (host->sg_ptr == &host->bounce_sg) {
> +			unsigned long flags;
> +			void *sg_vaddr = tmio_mmc_kmap_atomic(host->sg_orig, &flags);

Would be nice to stay within 80-chars where possible.

> +			memcpy(sg_vaddr, host->bounce_buf, host->bounce_sg.length);
> +			tmio_mmc_kunmap_atomic(sg_vaddr, &flags);
> +		}
>  		dev_dbg(&host->pdev->dev, "Complete Rx request %p\n",
>  			host->mrq);
>  	} else {
> @@ -529,8 +541,7 @@ static void tmio_mmc_cmd_irq(struct tmio_mmc_host *host,
>  			if (!host->chan_rx)
>  				enable_mmc_irqs(host, TMIO_MASK_READOP);
>  		} else {
> -			struct dma_chan *chan = host->chan_tx;
> -			if (!chan)
> +			if (!host->chan_tx)
>  				enable_mmc_irqs(host, TMIO_MASK_WRITEOP);
>  			else
>  				tasklet_schedule(&host->dma_issue);
> @@ -634,11 +645,36 @@ static void tmio_dma_complete(void *arg)
>  
>  static void tmio_mmc_start_dma_rx(struct tmio_mmc_host *host)
>  {
> -	struct scatterlist *sg = host->sg_ptr;
> +	struct scatterlist *sg = host->sg_ptr, *sg_tmp;
>  	struct dma_async_tx_descriptor *desc = NULL;
>  	struct dma_chan *chan = host->chan_rx;
> +	struct mfd_cell	*cell = host->pdev->dev.platform_data;
> +	struct tmio_mmc_data *pdata = cell->driver_data;
>  	dma_cookie_t cookie;
> -	int ret;
> +	int ret, i;
> +	bool aligned = true, multiple = true;
> +	unsigned int align = (1 << pdata->dma->alignment_shift) - 1;
> +
> +	for_each_sg(sg, sg_tmp, host->sg_len, i) {
> +		if (sg_tmp->offset & align)
> +			aligned = false;
> +		if (sg_tmp->length & align) {
> +			multiple = false;
> +			break;
> +		}
> +	}
> +
> +	if ((!aligned && (host->sg_len > 1 || sg->length > PAGE_CACHE_SIZE ||
> +			  align >= MAX_ALIGN)) || !multiple)
> +		goto pio;
> +
> +	/* The only sg element can be unaligned, use our bounce buffer then */
> +	if (!aligned) {
> +		/* The first sg element unaligned, use our bounce-buffer */
> +		sg_init_one(&host->bounce_sg, host->bounce_buf, sg->length);
> +		host->sg_ptr = &host->bounce_sg;
> +		sg = host->sg_ptr;
> +	}

The second comment is missing a verb, and looks redundant with the first.

>
>  	ret = dma_map_sg(&host->pdev->dev, sg, host->sg_len, DMA_FROM_DEVICE);
>  	if (ret > 0) {
> @@ -661,6 +697,7 @@ static void tmio_mmc_start_dma_rx(struct tmio_mmc_host *host)
>  	dev_dbg(&host->pdev->dev, "%s(): mapped %d -> %d, cookie %d, rq %p\n",
>  		__func__, host->sg_len, ret, cookie, host->mrq);
>  
> +pio:
>  	if (!desc) {
>  		/* DMA failed, fall back to PIO */
>  		if (ret >= 0)
> @@ -684,11 +721,40 @@ static void tmio_mmc_start_dma_rx(struct tmio_mmc_host *host)
>  
>  static void tmio_mmc_start_dma_tx(struct tmio_mmc_host *host)
>  {
> -	struct scatterlist *sg = host->sg_ptr;
> +	struct scatterlist *sg = host->sg_ptr, *sg_tmp;
>  	struct dma_async_tx_descriptor *desc = NULL;
>  	struct dma_chan *chan = host->chan_tx;
> +	struct mfd_cell	*cell = host->pdev->dev.platform_data;
> +	struct tmio_mmc_data *pdata = cell->driver_data;
>  	dma_cookie_t cookie;
> -	int ret;
> +	int ret, i;
> +	bool aligned = true, multiple = true;
> +	unsigned int align = (1 << pdata->dma->alignment_shift) - 1;
> +
> +	for_each_sg(sg, sg_tmp, host->sg_len, i) {
> +		if (sg_tmp->offset & align)
> +			aligned = false;
> +		if (sg_tmp->length & align) {
> +			multiple = false;
> +			break;
> +		}
> +	}
> +
> +	if ((!aligned && (host->sg_len > 1 || sg->length > PAGE_CACHE_SIZE ||
> +			  align >= MAX_ALIGN)) || !multiple)
> +		goto pio;
> +
> +	/* The only sg element can be unaligned, use our bounce buffer then */
> +	if (!aligned) {
> +		unsigned long flags;
> +		void *sg_vaddr = tmio_mmc_kmap_atomic(sg, &flags);
> +		/* The first sg element unaligned, use our bounce-buffer */
> +		sg_init_one(&host->bounce_sg, host->bounce_buf, sg->length);
> +		memcpy(host->bounce_buf, sg_vaddr, host->bounce_sg.length);
> +		tmio_mmc_kunmap_atomic(sg_vaddr, &flags);
> +		host->sg_ptr = &host->bounce_sg;
> +		sg = host->sg_ptr;
> +	}

Same here.

>  
>  	ret = dma_map_sg(&host->pdev->dev, sg, host->sg_len, DMA_TO_DEVICE);
>  	if (ret > 0) {
> @@ -709,6 +775,7 @@ static void tmio_mmc_start_dma_tx(struct tmio_mmc_host *host)
>  	dev_dbg(&host->pdev->dev, "%s(): mapped %d -> %d, cookie %d, rq %p\n",
>  		__func__, host->sg_len, ret, cookie, host->mrq);
>  
> +pio:
>  	if (!desc) {
>  		/* DMA failed, fall back to PIO */
>  		if (ret >= 0)
> diff --git a/include/linux/mfd/tmio.h b/include/linux/mfd/tmio.h
> index 085f041..dbfc053 100644
> --- a/include/linux/mfd/tmio.h
> +++ b/include/linux/mfd/tmio.h
> @@ -66,6 +66,7 @@ void tmio_core_mmc_clk_div(void __iomem *cnf, int shift, int state);
>  struct tmio_mmc_dma {
>  	void *chan_priv_tx;
>  	void *chan_priv_rx;
> +	int alignment_shift;
>  };
>  
>  /*

Thanks,

-- 
Chris Ball   <cjb@laptop.org>   <http://printf.net/>
One Laptop Per Child