mmc: tmio_mmc: implement a bounce buffer for unaligned DMA
authorGuennadi Liakhovetski <g.liakhovetski@gmx.de>
Wed, 22 Dec 2010 11:02:15 +0000 (12:02 +0100)
committerChris Ball <cjb@laptop.org>
Sun, 9 Jan 2011 04:52:29 +0000 (23:52 -0500)
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>
Acked-by: Samuel Ortiz <sameo@linux.intel.com>
Signed-off-by: Chris Ball <cjb@laptop.org>
drivers/mmc/host/tmio_mmc.c
include/linux/mfd/tmio.h

index e04c032abb1c42d377d59854f5581829b2e8308c..595b7b3f160d3b5cbf905cacddff27d1e9acac46 100644 (file)
                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;
 
@@ -139,9 +142,13 @@ 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
 };
 
+static void tmio_check_bounce_buffer(struct tmio_mmc_host *host);
+
 static u16 sd_ctrl_read16(struct tmio_mmc_host *host, int addr)
 {
        return readw(host->ctl + (addr << host->bus_shift));
@@ -180,6 +187,7 @@ 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;
 }
 
@@ -438,6 +446,8 @@ static void tmio_mmc_do_data_irq(struct tmio_mmc_host *host)
        if (data->flags & MMC_DATA_READ) {
                if (!host->chan_rx)
                        disable_mmc_irqs(host, TMIO_MASK_READOP);
+               else
+                       tmio_check_bounce_buffer(host);
                dev_dbg(&host->pdev->dev, "Complete Rx request %p\n",
                        host->mrq);
        } else {
@@ -529,8 +539,7 @@ 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);
@@ -612,6 +621,16 @@ out:
 }
 
 #ifdef CONFIG_TMIO_MMC_DMA
+static void tmio_check_bounce_buffer(struct tmio_mmc_host *host)
+{
+       if (host->sg_ptr == &host->bounce_sg) {
+               unsigned long flags;
+               void *sg_vaddr = tmio_mmc_kmap_atomic(host->sg_orig, &flags);
+               memcpy(sg_vaddr, host->bounce_buf, host->bounce_sg.length);
+               tmio_mmc_kunmap_atomic(sg_vaddr, &flags);
+       }
+}
+
 static void tmio_mmc_enable_dma(struct tmio_mmc_host *host, bool enable)
 {
 #if defined(CONFIG_SUPERH) || defined(CONFIG_ARCH_SHMOBILE)
@@ -634,11 +653,35 @@ 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) {
+               sg_init_one(&host->bounce_sg, host->bounce_buf, sg->length);
+               host->sg_ptr = &host->bounce_sg;
+               sg = host->sg_ptr;
+       }
 
        ret = dma_map_sg(&host->pdev->dev, sg, host->sg_len, DMA_FROM_DEVICE);
        if (ret > 0) {
@@ -661,6 +704,7 @@ 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 +728,39 @@ 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);
+               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;
+       }
 
        ret = dma_map_sg(&host->pdev->dev, sg, host->sg_len, DMA_TO_DEVICE);
        if (ret > 0) {
@@ -709,6 +781,7 @@ 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)
@@ -822,6 +895,10 @@ static void tmio_mmc_release_dma(struct tmio_mmc_host *host)
        }
 }
 #else
+static void tmio_check_bounce_buffer(struct tmio_mmc_host *host)
+{
+}
+
 static void tmio_mmc_start_dma(struct tmio_mmc_host *host,
                               struct mmc_data *data)
 {
index 085f041197dce74a49a691fe89a550826f7b4c05..dbfc053b1f95c98733a032126fddb9bfe1cdeb3c 100644 (file)
@@ -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;
 };
 
 /*