spi/ep93xx: add DMA support
authorMika Westerberg <mika.westerberg@iki.fi>
Sun, 29 May 2011 10:10:06 +0000 (13:10 +0300)
committerGrant Likely <grant.likely@secretlab.ca>
Wed, 8 Jun 2011 21:46:12 +0000 (15:46 -0600)
This patch adds DMA support for the EP93xx SPI driver. By default the DMA is
not enabled but it can be enabled by setting ep93xx_spi_info.use_dma to true
in board configuration file.

Note that the SPI driver still uses PIO for small transfers (<= 8 bytes) for
performance reasons.

Signed-off-by: Mika Westerberg <mika.westerberg@iki.fi>
Acked-by: H Hartley Sweeten <hsweeten@visionengravers.com>
Acked-by: Vinod Koul <vinod.koul@intel.com>
Signed-off-by: Grant Likely <grant.likely@secretlab.ca>
Documentation/spi/ep93xx_spi
arch/arm/mach-ep93xx/core.c
arch/arm/mach-ep93xx/include/mach/ep93xx_spi.h
drivers/spi/ep93xx_spi.c

index 6325f5b486357e7949d2adc90f62165b1330bee2..d8eb01c15db1cc9cf801138a87e21a7464ad0045 100644 (file)
@@ -88,6 +88,16 @@ static void __init ts72xx_init_machine(void)
                            ARRAY_SIZE(ts72xx_spi_devices));
 }
 
+The driver can use DMA for the transfers also. In this case ts72xx_spi_info
+becomes:
+
+static struct ep93xx_spi_info ts72xx_spi_info = {
+       .num_chipselect = ARRAY_SIZE(ts72xx_spi_devices),
+       .use_dma        = true;
+};
+
+Note that CONFIG_EP93XX_DMA should be enabled as well.
+
 Thanks to
 =========
 Martin Guy, H. Hartley Sweeten and others who helped me during development of
index 82079545adc4e8b788e98509cc8e77c7a5e95508..cc9f1d4b104dafded9fcd0d3e22dced6f3f873f1 100644 (file)
@@ -488,11 +488,15 @@ static struct resource ep93xx_spi_resources[] = {
        },
 };
 
+static u64 ep93xx_spi_dma_mask = DMA_BIT_MASK(32);
+
 static struct platform_device ep93xx_spi_device = {
        .name           = "ep93xx-spi",
        .id             = 0,
        .dev            = {
-               .platform_data = &ep93xx_spi_master_data,
+               .platform_data          = &ep93xx_spi_master_data,
+               .coherent_dma_mask      = DMA_BIT_MASK(32),
+               .dma_mask               = &ep93xx_spi_dma_mask,
        },
        .num_resources  = ARRAY_SIZE(ep93xx_spi_resources),
        .resource       = ep93xx_spi_resources,
index 0a37961b3453747d48aea9ee86b240ae3a7f266c..9bb63ac13f04574d5cf84a87067f4917dc79a5f8 100644 (file)
@@ -7,9 +7,11 @@ struct spi_device;
  * struct ep93xx_spi_info - EP93xx specific SPI descriptor
  * @num_chipselect: number of chip selects on this board, must be
  *                  at least one
+ * @use_dma: use DMA for the transfers
  */
 struct ep93xx_spi_info {
        int     num_chipselect;
+       bool    use_dma;
 };
 
 /**
index d3570071e98f47e8b6ecd3d4834eab7278972898..1cf645479bfec6b1cc7a2d2dc76d4e0e295fb8bc 100644 (file)
@@ -1,7 +1,7 @@
 /*
  * Driver for Cirrus Logic EP93xx SPI controller.
  *
- * Copyright (c) 2010 Mika Westerberg
+ * Copyright (C) 2010-2011 Mika Westerberg
  *
  * Explicit FIFO handling code was inspired by amba-pl022 driver.
  *
 #include <linux/err.h>
 #include <linux/delay.h>
 #include <linux/device.h>
+#include <linux/dmaengine.h>
 #include <linux/bitops.h>
 #include <linux/interrupt.h>
 #include <linux/platform_device.h>
 #include <linux/workqueue.h>
 #include <linux/sched.h>
+#include <linux/scatterlist.h>
 #include <linux/spi/spi.h>
 
+#include <mach/dma.h>
 #include <mach/ep93xx_spi.h>
 
 #define SSPCR0                 0x0000
@@ -71,6 +74,7 @@
  * @pdev: pointer to platform device
  * @clk: clock for the controller
  * @regs_base: pointer to ioremap()'d registers
+ * @sspdr_phys: physical address of the SSPDR register
  * @irq: IRQ number used by the driver
  * @min_rate: minimum clock rate (in Hz) supported by the controller
  * @max_rate: maximum clock rate (in Hz) supported by the controller
  * @rx: current byte in transfer to receive
  * @fifo_level: how full is FIFO (%0..%SPI_FIFO_SIZE - %1). Receiving one
  *              frame decreases this level and sending one frame increases it.
+ * @dma_rx: RX DMA channel
+ * @dma_tx: TX DMA channel
+ * @dma_rx_data: RX parameters passed to the DMA engine
+ * @dma_tx_data: TX parameters passed to the DMA engine
+ * @rx_sgt: sg table for RX transfers
+ * @tx_sgt: sg table for TX transfers
+ * @zeropage: dummy page used as RX buffer when only TX buffer is passed in by
+ *            the client
  *
  * This structure holds EP93xx SPI controller specific information. When
  * @running is %true, driver accepts transfer requests from protocol drivers.
@@ -100,6 +112,7 @@ struct ep93xx_spi {
        const struct platform_device    *pdev;
        struct clk                      *clk;
        void __iomem                    *regs_base;
+       unsigned long                   sspdr_phys;
        int                             irq;
        unsigned long                   min_rate;
        unsigned long                   max_rate;
@@ -112,6 +125,13 @@ struct ep93xx_spi {
        size_t                          tx;
        size_t                          rx;
        size_t                          fifo_level;
+       struct dma_chan                 *dma_rx;
+       struct dma_chan                 *dma_tx;
+       struct ep93xx_dma_data          dma_rx_data;
+       struct ep93xx_dma_data          dma_tx_data;
+       struct sg_table                 rx_sgt;
+       struct sg_table                 tx_sgt;
+       void                            *zeropage;
 };
 
 /**
@@ -496,14 +516,195 @@ static int ep93xx_spi_read_write(struct ep93xx_spi *espi)
                espi->fifo_level++;
        }
 
-       if (espi->rx == t->len) {
-               msg->actual_length += t->len;
+       if (espi->rx == t->len)
                return 0;
-       }
 
        return -EINPROGRESS;
 }
 
+static void ep93xx_spi_pio_transfer(struct ep93xx_spi *espi)
+{
+       /*
+        * Now everything is set up for the current transfer. We prime the TX
+        * FIFO, enable interrupts, and wait for the transfer to complete.
+        */
+       if (ep93xx_spi_read_write(espi)) {
+               ep93xx_spi_enable_interrupts(espi);
+               wait_for_completion(&espi->wait);
+       }
+}
+
+/**
+ * ep93xx_spi_dma_prepare() - prepares a DMA transfer
+ * @espi: ep93xx SPI controller struct
+ * @dir: DMA transfer direction
+ *
+ * Function configures the DMA, maps the buffer and prepares the DMA
+ * descriptor. Returns a valid DMA descriptor in case of success and ERR_PTR
+ * in case of failure.
+ */
+static struct dma_async_tx_descriptor *
+ep93xx_spi_dma_prepare(struct ep93xx_spi *espi, enum dma_data_direction dir)
+{
+       struct spi_transfer *t = espi->current_msg->state;
+       struct dma_async_tx_descriptor *txd;
+       enum dma_slave_buswidth buswidth;
+       struct dma_slave_config conf;
+       struct scatterlist *sg;
+       struct sg_table *sgt;
+       struct dma_chan *chan;
+       const void *buf, *pbuf;
+       size_t len = t->len;
+       int i, ret, nents;
+
+       if (bits_per_word(espi) > 8)
+               buswidth = DMA_SLAVE_BUSWIDTH_2_BYTES;
+       else
+               buswidth = DMA_SLAVE_BUSWIDTH_1_BYTE;
+
+       memset(&conf, 0, sizeof(conf));
+       conf.direction = dir;
+
+       if (dir == DMA_FROM_DEVICE) {
+               chan = espi->dma_rx;
+               buf = t->rx_buf;
+               sgt = &espi->rx_sgt;
+
+               conf.src_addr = espi->sspdr_phys;
+               conf.src_addr_width = buswidth;
+       } else {
+               chan = espi->dma_tx;
+               buf = t->tx_buf;
+               sgt = &espi->tx_sgt;
+
+               conf.dst_addr = espi->sspdr_phys;
+               conf.dst_addr_width = buswidth;
+       }
+
+       ret = dmaengine_slave_config(chan, &conf);
+       if (ret)
+               return ERR_PTR(ret);
+
+       /*
+        * We need to split the transfer into PAGE_SIZE'd chunks. This is
+        * because we are using @espi->zeropage to provide a zero RX buffer
+        * for the TX transfers and we have only allocated one page for that.
+        *
+        * For performance reasons we allocate a new sg_table only when
+        * needed. Otherwise we will re-use the current one. Eventually the
+        * last sg_table is released in ep93xx_spi_release_dma().
+        */
+
+       nents = DIV_ROUND_UP(len, PAGE_SIZE);
+       if (nents != sgt->nents) {
+               sg_free_table(sgt);
+
+               ret = sg_alloc_table(sgt, nents, GFP_KERNEL);
+               if (ret)
+                       return ERR_PTR(ret);
+       }
+
+       pbuf = buf;
+       for_each_sg(sgt->sgl, sg, sgt->nents, i) {
+               size_t bytes = min_t(size_t, len, PAGE_SIZE);
+
+               if (buf) {
+                       sg_set_page(sg, virt_to_page(pbuf), bytes,
+                                   offset_in_page(pbuf));
+               } else {
+                       sg_set_page(sg, virt_to_page(espi->zeropage),
+                                   bytes, 0);
+               }
+
+               pbuf += bytes;
+               len -= bytes;
+       }
+
+       if (WARN_ON(len)) {
+               dev_warn(&espi->pdev->dev, "len = %d expected 0!", len);
+               return ERR_PTR(-EINVAL);
+       }
+
+       nents = dma_map_sg(chan->device->dev, sgt->sgl, sgt->nents, dir);
+       if (!nents)
+               return ERR_PTR(-ENOMEM);
+
+       txd = chan->device->device_prep_slave_sg(chan, sgt->sgl, nents,
+                                                dir, DMA_CTRL_ACK);
+       if (!txd) {
+               dma_unmap_sg(chan->device->dev, sgt->sgl, sgt->nents, dir);
+               return ERR_PTR(-ENOMEM);
+       }
+       return txd;
+}
+
+/**
+ * ep93xx_spi_dma_finish() - finishes with a DMA transfer
+ * @espi: ep93xx SPI controller struct
+ * @dir: DMA transfer direction
+ *
+ * Function finishes with the DMA transfer. After this, the DMA buffer is
+ * unmapped.
+ */
+static void ep93xx_spi_dma_finish(struct ep93xx_spi *espi,
+                                 enum dma_data_direction dir)
+{
+       struct dma_chan *chan;
+       struct sg_table *sgt;
+
+       if (dir == DMA_FROM_DEVICE) {
+               chan = espi->dma_rx;
+               sgt = &espi->rx_sgt;
+       } else {
+               chan = espi->dma_tx;
+               sgt = &espi->tx_sgt;
+       }
+
+       dma_unmap_sg(chan->device->dev, sgt->sgl, sgt->nents, dir);
+}
+
+static void ep93xx_spi_dma_callback(void *callback_param)
+{
+       complete(callback_param);
+}
+
+static void ep93xx_spi_dma_transfer(struct ep93xx_spi *espi)
+{
+       struct spi_message *msg = espi->current_msg;
+       struct dma_async_tx_descriptor *rxd, *txd;
+
+       rxd = ep93xx_spi_dma_prepare(espi, DMA_FROM_DEVICE);
+       if (IS_ERR(rxd)) {
+               dev_err(&espi->pdev->dev, "DMA RX failed: %ld\n", PTR_ERR(rxd));
+               msg->status = PTR_ERR(rxd);
+               return;
+       }
+
+       txd = ep93xx_spi_dma_prepare(espi, DMA_TO_DEVICE);
+       if (IS_ERR(txd)) {
+               ep93xx_spi_dma_finish(espi, DMA_FROM_DEVICE);
+               dev_err(&espi->pdev->dev, "DMA TX failed: %ld\n", PTR_ERR(rxd));
+               msg->status = PTR_ERR(txd);
+               return;
+       }
+
+       /* We are ready when RX is done */
+       rxd->callback = ep93xx_spi_dma_callback;
+       rxd->callback_param = &espi->wait;
+
+       /* Now submit both descriptors and wait while they finish */
+       dmaengine_submit(rxd);
+       dmaengine_submit(txd);
+
+       dma_async_issue_pending(espi->dma_rx);
+       dma_async_issue_pending(espi->dma_tx);
+
+       wait_for_completion(&espi->wait);
+
+       ep93xx_spi_dma_finish(espi, DMA_TO_DEVICE);
+       ep93xx_spi_dma_finish(espi, DMA_FROM_DEVICE);
+}
+
 /**
  * ep93xx_spi_process_transfer() - processes one SPI transfer
  * @espi: ep93xx SPI controller struct
@@ -556,13 +757,14 @@ static void ep93xx_spi_process_transfer(struct ep93xx_spi *espi,
        espi->tx = 0;
 
        /*
-        * Now everything is set up for the current transfer. We prime the TX
-        * FIFO, enable interrupts, and wait for the transfer to complete.
+        * There is no point of setting up DMA for the transfers which will
+        * fit into the FIFO and can be transferred with a single interrupt.
+        * So in these cases we will be using PIO and don't bother for DMA.
         */
-       if (ep93xx_spi_read_write(espi)) {
-               ep93xx_spi_enable_interrupts(espi);
-               wait_for_completion(&espi->wait);
-       }
+       if (espi->dma_rx && t->len > SPI_FIFO_SIZE)
+               ep93xx_spi_dma_transfer(espi);
+       else
+               ep93xx_spi_pio_transfer(espi);
 
        /*
         * In case of error during transmit, we bail out from processing
@@ -571,6 +773,8 @@ static void ep93xx_spi_process_transfer(struct ep93xx_spi *espi,
        if (msg->status)
                return;
 
+       msg->actual_length += t->len;
+
        /*
         * After this transfer is finished, perform any possible
         * post-transfer actions requested by the protocol driver.
@@ -752,6 +956,75 @@ static irqreturn_t ep93xx_spi_interrupt(int irq, void *dev_id)
        return IRQ_HANDLED;
 }
 
+static bool ep93xx_spi_dma_filter(struct dma_chan *chan, void *filter_param)
+{
+       if (ep93xx_dma_chan_is_m2p(chan))
+               return false;
+
+       chan->private = filter_param;
+       return true;
+}
+
+static int ep93xx_spi_setup_dma(struct ep93xx_spi *espi)
+{
+       dma_cap_mask_t mask;
+       int ret;
+
+       espi->zeropage = (void *)get_zeroed_page(GFP_KERNEL);
+       if (!espi->zeropage)
+               return -ENOMEM;
+
+       dma_cap_zero(mask);
+       dma_cap_set(DMA_SLAVE, mask);
+
+       espi->dma_rx_data.port = EP93XX_DMA_SSP;
+       espi->dma_rx_data.direction = DMA_FROM_DEVICE;
+       espi->dma_rx_data.name = "ep93xx-spi-rx";
+
+       espi->dma_rx = dma_request_channel(mask, ep93xx_spi_dma_filter,
+                                          &espi->dma_rx_data);
+       if (!espi->dma_rx) {
+               ret = -ENODEV;
+               goto fail_free_page;
+       }
+
+       espi->dma_tx_data.port = EP93XX_DMA_SSP;
+       espi->dma_tx_data.direction = DMA_TO_DEVICE;
+       espi->dma_tx_data.name = "ep93xx-spi-tx";
+
+       espi->dma_tx = dma_request_channel(mask, ep93xx_spi_dma_filter,
+                                          &espi->dma_tx_data);
+       if (!espi->dma_tx) {
+               ret = -ENODEV;
+               goto fail_release_rx;
+       }
+
+       return 0;
+
+fail_release_rx:
+       dma_release_channel(espi->dma_rx);
+       espi->dma_rx = NULL;
+fail_free_page:
+       free_page((unsigned long)espi->zeropage);
+
+       return ret;
+}
+
+static void ep93xx_spi_release_dma(struct ep93xx_spi *espi)
+{
+       if (espi->dma_rx) {
+               dma_release_channel(espi->dma_rx);
+               sg_free_table(&espi->rx_sgt);
+       }
+       if (espi->dma_tx) {
+               dma_release_channel(espi->dma_tx);
+               sg_free_table(&espi->tx_sgt);
+       }
+
+       if (espi->zeropage)
+               free_page((unsigned long)espi->zeropage);
+}
+
 static int __init ep93xx_spi_probe(struct platform_device *pdev)
 {
        struct spi_master *master;
@@ -818,6 +1091,7 @@ static int __init ep93xx_spi_probe(struct platform_device *pdev)
                goto fail_put_clock;
        }
 
+       espi->sspdr_phys = res->start + SSPDR;
        espi->regs_base = ioremap(res->start, resource_size(res));
        if (!espi->regs_base) {
                dev_err(&pdev->dev, "failed to map resources\n");
@@ -832,10 +1106,13 @@ static int __init ep93xx_spi_probe(struct platform_device *pdev)
                goto fail_unmap_regs;
        }
 
+       if (info->use_dma && ep93xx_spi_setup_dma(espi))
+               dev_warn(&pdev->dev, "DMA setup failed. Falling back to PIO\n");
+
        espi->wq = create_singlethread_workqueue("ep93xx_spid");
        if (!espi->wq) {
                dev_err(&pdev->dev, "unable to create workqueue\n");
-               goto fail_free_irq;
+               goto fail_free_dma;
        }
        INIT_WORK(&espi->msg_work, ep93xx_spi_work);
        INIT_LIST_HEAD(&espi->msg_queue);
@@ -857,7 +1134,8 @@ static int __init ep93xx_spi_probe(struct platform_device *pdev)
 
 fail_free_queue:
        destroy_workqueue(espi->wq);
-fail_free_irq:
+fail_free_dma:
+       ep93xx_spi_release_dma(espi);
        free_irq(espi->irq, espi);
 fail_unmap_regs:
        iounmap(espi->regs_base);
@@ -901,6 +1179,7 @@ static int __exit ep93xx_spi_remove(struct platform_device *pdev)
        }
        spin_unlock_irq(&espi->lock);
 
+       ep93xx_spi_release_dma(espi);
        free_irq(espi->irq, espi);
        iounmap(espi->regs_base);
        res = platform_get_resource(pdev, IORESOURCE_MEM, 0);