[PATCH] ARM: add PrimeCell generic DMA to MMCI/PL180

* [PATCH] ARM: add PrimeCell generic DMA to MMCI/PL180
@ 2010-10-06  9:42 Linus Walleij
  2010-12-19 16:32 ` Russell King - ARM Linux
  0 siblings, 1 reply; 18+ messages in thread
From: Linus Walleij @ 2010-10-06  9:42 UTC (permalink / raw)
  To: linux-arm-kernel

This extends the MMCI/PL180 driver with generic DMA engine support
using the PrimeCell DMA engine interface.

Signed-off-by: Linus Walleij <linus.walleij@stericsson.com>
---
Changes to previous version:

- Added burst size compensation for the Ux500 variants: these
  require that you send single words (=32 bits, 4 bytes) of bursts
  when the burst is < 16 bytes.
---
 drivers/mmc/host/mmci.c   |  270 ++++++++++++++++++++++++++++++++++++++++++---
 drivers/mmc/host/mmci.h   |   13 ++
 include/linux/amba/mmci.h |   16 +++
 3 files changed, 285 insertions(+), 14 deletions(-)

diff --git a/drivers/mmc/host/mmci.c b/drivers/mmc/host/mmci.c
index 09bcce7..951ac37 100644
--- a/drivers/mmc/host/mmci.c
+++ b/drivers/mmc/host/mmci.c
@@ -2,7 +2,7 @@
  *  linux/drivers/mmc/host/mmci.c - ARM PrimeCell MMCI PL180/1 driver
  *
  *  Copyright (C) 2003 Deep Blue Solutions, Ltd, All Rights Reserved.
- *  Copyright (C) 2010 ST-Ericsson AB.
+ *  Copyright (C) 2010 ST-Ericsson SA
  *
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License version 2 as
@@ -23,8 +23,10 @@
 #include <linux/clk.h>
 #include <linux/scatterlist.h>
 #include <linux/gpio.h>
-#include <linux/amba/mmci.h>
 #include <linux/regulator/consumer.h>
+#include <linux/dmaengine.h>
+#include <linux/dma-mapping.h>
+#include <linux/amba/mmci.h>
 
 #include <asm/div64.h>
 #include <asm/io.h>
@@ -40,20 +42,26 @@ static unsigned int fmax = 515633;
  * struct variant_data - MMCI variant-specific quirks
  * @clkreg: default value for MCICLOCK register
  * @clkreg_enable: enable value for MMCICLOCK register
+ * @dmareg_enable: enable value for MMCIDATACTRL register
  * @datalength_bits: number of bits in the MMCIDATALENGTH register
  * @fifosize: number of bytes that can be written when MMCI_TXFIFOEMPTY
  *	      is asserted (likewise for RX)
  * @fifohalfsize: number of bytes that can be written when MCI_TXFIFOHALFEMPTY
  *		  is asserted (likewise for RX)
  * @singleirq: true if only one IRQ line is available
+ * @txsize_threshold: Sets burst size to minimal if transfer size is less or
+ *		equal to this threshold. This shall be specified in
+ *		number of bytes. Set 0 for no burst compensation.
  */
 struct variant_data {
 	unsigned int		clkreg;
 	unsigned int		clkreg_enable;
+	unsigned int		dmareg_enable;
 	unsigned int		datalength_bits;
 	unsigned int		fifosize;
 	unsigned int		fifohalfsize;
 	bool			singleirq;
+	unsigned int		txsize_threshold;
 };
 
 static struct variant_data variant_arm = {
@@ -74,8 +82,10 @@ static struct variant_data variant_ux500 = {
 	.fifohalfsize		= 8 * 4,
 	.clkreg			= MCI_CLK_ENABLE,
 	.clkreg_enable		= 1 << 14, /* HWFCEN */
+	.dmareg_enable		= 1 << 12, /* DMAREQCTRL */
 	.datalength_bits	= 24,
 	.singleirq		= true,
+	.txsize_threshold	= 16,
 };
 
 /*
@@ -169,6 +179,209 @@ static void mmci_init_sg(struct mmci_host *host, struct mmc_data *data)
 	sg_miter_start(&host->sg_miter, data->sg, data->sg_len, flags);
 }
 
+/*
+ * All the DMA operation mode stuff goes inside this ifdef.
+ * This assumes that you have a generic DMA device interface,
+ * no custom DMA interfaces are supported.
+ */
+#ifdef CONFIG_DMA_ENGINE
+static void __devinit mmci_setup_dma(struct mmci_host *host)
+{
+	struct mmci_platform_data *plat = host->plat;
+	dma_cap_mask_t mask;
+
+	if (!plat || !plat->dma_filter) {
+		dev_err(mmc_dev(host->mmc), "no DMA platform data!\n");
+		return;
+	}
+
+	/* Try to acquire a generic DMA engine slave channel */
+	dma_cap_zero(mask);
+	dma_cap_set(DMA_SLAVE, mask);
+	/*
+	 * If only an RX channel is specified, the driver will
+	 * attempt to use it bidirectionally, however if it is
+	 * is specified but cannot be located, DMA will be disabled.
+	 */
+	host->dma_rx_channel = dma_request_channel(mask,
+						plat->dma_filter,
+						plat->dma_rx_param);
+	/* E.g if no DMA hardware is present */
+	if (!host->dma_rx_channel) {
+		dev_err(mmc_dev(host->mmc), "no RX DMA channel!\n");
+		return;
+	}
+	if (plat->dma_tx_param) {
+		host->dma_tx_channel = dma_request_channel(mask,
+							   plat->dma_filter,
+							   plat->dma_tx_param);
+		if (!host->dma_tx_channel) {
+			dma_release_channel(host->dma_rx_channel);
+			host->dma_rx_channel = NULL;
+			return;
+		}
+	} else {
+		host->dma_tx_channel = host->dma_rx_channel;
+	}
+	host->dma_enable = true;
+	dev_info(mmc_dev(host->mmc), "use DMA channels DMA RX %s, DMA TX %s\n",
+		 dma_chan_name(host->dma_rx_channel),
+		 dma_chan_name(host->dma_tx_channel));
+}
+
+/*
+ * This is used in __devinit or __devexit so inline it
+ * so it can be discarded.
+ */
+static inline void mmci_disable_dma(struct mmci_host *host)
+{
+	if (host->dma_rx_channel)
+		dma_release_channel(host->dma_rx_channel);
+	if (host->dma_tx_channel)
+		dma_release_channel(host->dma_tx_channel);
+	host->dma_enable = false;
+}
+
+static void mmci_dma_data_end(struct mmci_host *host)
+{
+	struct mmc_data *data = host->data;
+
+	dma_unmap_sg(mmc_dev(host->mmc), data->sg, data->sg_len,
+		     (data->flags & MMC_DATA_WRITE)
+		     ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
+}
+
+static void mmci_dma_data_error(struct mmci_host *host)
+{
+	struct mmc_data *data = host->data;
+	struct dma_chan *chan;
+
+	dev_err(mmc_dev(host->mmc), "error during DMA transfer!\n");
+	if (data->flags & MMC_DATA_READ)
+		chan = host->dma_rx_channel;
+	else
+		chan = host->dma_tx_channel;
+	dma_unmap_sg(mmc_dev(host->mmc), data->sg, data->sg_len,
+		     (data->flags & MMC_DATA_WRITE)
+		     ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
+	chan->device->device_control(chan, DMA_TERMINATE_ALL, 0);
+}
+
+static int mmci_dma_start_data(struct mmci_host *host, unsigned int datactrl)
+{
+	struct variant_data *variant = host->variant;
+	struct dma_slave_config rx_conf = {
+		.src_addr = host->phybase + MMCIFIFO,
+		.src_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES,
+		.direction = DMA_FROM_DEVICE,
+		.src_maxburst = variant->fifohalfsize >> 2, /* # of words */
+	};
+	struct dma_slave_config tx_conf = {
+		.dst_addr = host->phybase + MMCIFIFO,
+		.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES,
+		.direction = DMA_TO_DEVICE,
+		.dst_maxburst = variant->fifohalfsize >> 2, /* # of words */
+	};
+	struct mmc_data *data = host->data;
+	enum dma_data_direction direction;
+	struct dma_chan *chan;
+	struct dma_async_tx_descriptor *desc;
+	struct scatterlist *sg;
+	dma_cookie_t cookie;
+	int i;
+
+	datactrl |= MCI_DPSM_DMAENABLE;
+	datactrl |= variant->dmareg_enable;
+
+	if (data->flags & MMC_DATA_READ) {
+		if (host->size <= variant->txsize_threshold)
+			rx_conf.src_maxburst = 1;
+
+		direction = DMA_FROM_DEVICE;
+		chan = host->dma_rx_channel;
+		chan->device->device_control(chan, DMA_SLAVE_CONFIG,
+					     (unsigned long) &rx_conf);
+	} else {
+		if (host->size <= variant->txsize_threshold)
+			tx_conf.dst_maxburst = 1;
+
+		direction = DMA_TO_DEVICE;
+		chan = host->dma_tx_channel;
+		chan->device->device_control(chan, DMA_SLAVE_CONFIG,
+					     (unsigned long) &tx_conf);
+	}
+
+	/* Check for weird stuff in the sg list */
+	for_each_sg(data->sg, sg, data->sg_len, i) {
+		dev_vdbg(mmc_dev(host->mmc), "MMCI SGlist %d dir %d: length: %08x\n",
+			 i, direction, sg->length);
+		if (sg->offset & 3 || sg->length & 3)
+			return -EINVAL;
+	}
+
+	dma_map_sg(mmc_dev(host->mmc), data->sg,
+		   data->sg_len, direction);
+
+	desc = chan->device->device_prep_slave_sg(chan,
+					data->sg, data->sg_len, direction,
+					DMA_CTRL_ACK);
+	if (!desc)
+		goto unmap_exit;
+
+	host->dma_desc = desc;
+	dev_vdbg(mmc_dev(host->mmc), "Submit MMCI DMA job, sglen %d "
+		 "blksz %04x blks %04x flags %08x\n",
+		 data->sg_len, data->blksz, data->blocks, data->flags);
+	cookie = desc->tx_submit(desc);
+
+	/* Here overloaded DMA controllers may fail */
+	if (dma_submit_error(cookie))
+		goto unmap_exit;
+
+	host->dma_on_current_xfer = true;
+	chan->device->device_issue_pending(chan);
+
+	/* Trigger the DMA transfer */
+	writel(datactrl, host->base + MMCIDATACTRL);
+	/*
+	 * Let the MMCI say when the data is ended and it's time
+	 * to fire next DMA request. When that happens, MMCI will
+	 * call mmci_data_end()
+	 */
+	writel(readl(host->base + MMCIMASK0) | MCI_DATAENDMASK,
+	       host->base + MMCIMASK0);
+	return 0;
+
+unmap_exit:
+	chan->device->device_control(chan, DMA_TERMINATE_ALL, 0);
+	dma_unmap_sg(mmc_dev(host->mmc), data->sg, data->sg_len, direction);
+	host->dma_on_current_xfer = false;
+	return -ENOMEM;
+}
+#else
+/* Blank functions if the DMA engine is not available */
+static inline void mmci_setup_dma(struct mmci_host *host)
+{
+}
+
+static inline void mmci_disable_dma(struct mmci_host *host)
+{
+}
+
+static inline void mmci_dma_data_end(struct mmci_host *host)
+{
+}
+
+static inline void mmci_dma_data_error(struct mmci_host *host)
+{
+}
+
+static inline int mmci_dma_start_data(struct mmci_host *host, unsigned int datactrl)
+{
+	return -ENOSYS;
+}
+#endif
+
 static void mmci_start_data(struct mmci_host *host, struct mmc_data *data)
 {
 	struct variant_data *variant = host->variant;
@@ -183,8 +396,8 @@ static void mmci_start_data(struct mmci_host *host, struct mmc_data *data)
 	host->data = data;
 	host->size = data->blksz * data->blocks;
 	host->data_xfered = 0;
-
-	mmci_init_sg(host, data);
+	host->blockend = false;
+	host->dataend = false;
 
 	clks = (unsigned long long)data->timeout_ns * host->cclk;
 	do_div(clks, 1000000000UL);
@@ -199,13 +412,32 @@ static void mmci_start_data(struct mmci_host *host, struct mmc_data *data)
 	BUG_ON(1 << blksz_bits != data->blksz);
 
 	datactrl = MCI_DPSM_ENABLE | blksz_bits << 4;
-	if (data->flags & MMC_DATA_READ) {
+
+	if (data->flags & MMC_DATA_READ)
 		datactrl |= MCI_DPSM_DIRECTION;
+
+	if (host->dma_enable) {
+		int ret;
+
+		/*
+		 * Attempt to use DMA operation mode, if this
+		 * should fail, fall back to PIO mode
+		 */
+		ret = mmci_dma_start_data(host, datactrl);
+		if (!ret)
+			return;
+	}
+
+	/* IRQ mode, map the SG list for CPU reading/writing */
+	mmci_init_sg(host, data);
+
+	if (data->flags & MMC_DATA_READ) {
 		irqmask = MCI_RXFIFOHALFFULLMASK;
 
 		/*
-		 * If we have less than a FIFOSIZE of bytes to transfer,
-		 * trigger a PIO interrupt as soon as any data is available.
+		 * If we have less than a FIFOSIZE of bytes to
+		 * transfer, trigger a PIO interrupt as soon as any
+		 * data is available.
 		 */
 		if (host->size < variant->fifosize)
 			irqmask |= MCI_RXDATAAVLBLMASK;
@@ -270,9 +502,12 @@ mmci_data_irq(struct mmci_host *host, struct mmc_data *data,
 
 		/*
 		 * We hit an error condition.  Ensure that any data
-		 * partially written to a page is properly coherent.
+		 * partially written to a page is properly coherent,
+		 * unless we're using DMA.
 		 */
-		if (data->flags & MMC_DATA_READ) {
+		if (host->dma_on_current_xfer)
+			mmci_dma_data_error(host);
+		else if (data->flags & MMC_DATA_READ) {
 			struct sg_mapping_iter *sg_miter = &host->sg_miter;
 			unsigned long flags;
 
@@ -308,7 +543,7 @@ mmci_data_irq(struct mmci_host *host, struct mmc_data *data,
 		 * use this progressive update in DMA or single-IRQ
 		 * mode.
 		 */
-		if (!host->variant->singleirq)
+		if (!host->variant->singleirq && !host->dma_on_current_xfer)
 			host->data_xfered += data->blksz;
 		host->blockend = true;
 	}
@@ -321,6 +556,7 @@ mmci_data_irq(struct mmci_host *host, struct mmc_data *data,
 	 * with the blockend signal since they can appear out-of-order.
 	 */
 	if (host->dataend && (host->blockend || host->variant->singleirq)) {
+		mmci_dma_data_end(host);
 		mmci_stop_data(host);
 
 		/* Reset these flags */
@@ -332,7 +568,7 @@ mmci_data_irq(struct mmci_host *host, struct mmc_data *data,
 		 * and in DMA mode this signal need to be synchronized
 		 * with MCI_DATEND
 		 */
-		if (host->variant->singleirq && !data->error)
+		if ((host->variant->singleirq || host->dma_on_current_xfer) && !data->error)
 			host->data_xfered += data->blksz * data->blocks;
 
 		if (!data->stop) {
@@ -761,6 +997,7 @@ static int __devinit mmci_probe(struct amba_device *dev, struct amba_id *id)
 		dev_dbg(mmc_dev(mmc), "eventual mclk rate: %u Hz\n",
 			host->mclk);
 	}
+	host->phybase = dev->res.start;
 	host->base = ioremap(dev->res.start, resource_size(&dev->res));
 	if (!host->base) {
 		ret = -ENOMEM;
@@ -871,6 +1108,8 @@ static int __devinit mmci_probe(struct amba_device *dev, struct amba_id *id)
 	    && host->gpio_cd_irq < 0)
 		mmc->caps |= MMC_CAP_NEEDS_POLL;
 
+	mmci_setup_dma(host);
+
 	ret = request_irq(dev->irq[0], mmci_irq, IRQF_SHARED, DRIVER_NAME " (cmd)", host);
 	if (ret)
 		goto unmap;
@@ -897,15 +1136,17 @@ static int __devinit mmci_probe(struct amba_device *dev, struct amba_id *id)
 
 	mmc_add_host(mmc);
 
-	dev_info(&dev->dev, "%s: MMCI rev %x cfg %02x at 0x%016llx irq %d,%d\n",
-		mmc_hostname(mmc), amba_rev(dev), amba_config(dev),
-		(unsigned long long)dev->res.start, dev->irq[0], dev->irq[1]);
+	dev_info(&dev->dev, "%s: MMCI/PL180 manf %x rev %x cfg %02x at 0x%016llx\n",
+		 mmc_hostname(mmc), amba_manf(dev), amba_rev(dev), amba_config(dev),
+		 (unsigned long long)dev->res.start);
+	dev_info(&dev->dev, "IRQ %d, %d (pio)\n", dev->irq[0], dev->irq[1]);
 
 	return 0;
 
  irq0_free:
 	free_irq(dev->irq[0], host);
  unmap:
+	mmci_disable_dma(host);
 	if (host->gpio_wp != -ENOSYS)
 		gpio_free(host->gpio_wp);
  err_gpio_wp:
@@ -945,6 +1186,7 @@ static int __devexit mmci_remove(struct amba_device *dev)
 		writel(0, host->base + MMCICOMMAND);
 		writel(0, host->base + MMCIDATACTRL);
 
+		mmci_disable_dma(host);
 		free_irq(dev->irq[0], host);
 		if (!variant->singleirq)
 			free_irq(dev->irq[1], host);
diff --git a/drivers/mmc/host/mmci.h b/drivers/mmc/host/mmci.h
index acac081..9e2a4c6 100644
--- a/drivers/mmc/host/mmci.h
+++ b/drivers/mmc/host/mmci.h
@@ -142,8 +142,11 @@
 
 struct clk;
 struct variant_data;
+struct dma_chan;
+struct dma_async_tx_descriptor;
 
 struct mmci_host {
+	phys_addr_t		phybase;
 	void __iomem		*base;
 	struct mmc_request	*mrq;
 	struct mmc_command	*cmd;
@@ -176,6 +179,16 @@ struct mmci_host {
 	/* pio stuff */
 	struct sg_mapping_iter	sg_miter;
 	unsigned int		size;
+
 	struct regulator	*vcc;
+
+	/* DMA stuff */
+	bool			dma_enable;
+	bool			dma_on_current_xfer;
+#ifdef CONFIG_DMA_ENGINE
+	struct dma_chan		*dma_rx_channel;
+	struct dma_chan		*dma_tx_channel;
+	struct dma_async_tx_descriptor *dma_desc;
+#endif
 };
 
diff --git a/include/linux/amba/mmci.h b/include/linux/amba/mmci.h
index f4ee9ac..dbeba68 100644
--- a/include/linux/amba/mmci.h
+++ b/include/linux/amba/mmci.h
@@ -6,6 +6,8 @@
 
 #include <linux/mmc/host.h>
 
+/* Just some dummy forwarding */
+struct dma_chan;
 /**
  * struct mmci_platform_data - platform configuration for the MMCI
  * (also known as PL180) block.
@@ -27,6 +29,17 @@
  * @cd_invert: true if the gpio_cd pin value is active low
  * @capabilities: the capabilities of the block as implemented in
  * this platform, signify anything MMC_CAP_* from mmc/host.h
+ * @dma_filter: function used to select an apropriate RX and TX
+ * DMA channel to be used for DMA, if and only if you're deploying the
+ * generic DMA engine
+ * @dma_rx_param: parameter passed to the DMA allocation
+ * filter in order to select an apropriate RX channel. If
+ * there is a bidirectional RX+TX channel, then just specify
+ * this and leave dma_tx_param set to NULL
+ * @dma_tx_param: parameter passed to the DMA allocation
+ * filter in order to select an apropriate TX channel. If this
+ * is NULL the driver will attempt to use the RX channel as a
+ * bidirectional channel
  */
 struct mmci_platform_data {
 	unsigned int f_max;
@@ -38,6 +51,9 @@ struct mmci_platform_data {
 	int	gpio_cd;
 	bool	cd_invert;
 	unsigned long capabilities;
+	bool (*dma_filter)(struct dma_chan *chan, void *filter_param);
+	void *dma_rx_param;
+	void *dma_tx_param;
 };
 
 #endif
-- 
1.6.3.3

^ permalink raw reply related	[flat|nested] 18+ messages in thread