mmc: sunxi: Add support for DMA transfers

Allwinner MMC controller supports DMA via internal DMA controller,
use it.

Signed-off-by: Ondrej Jirman <megous@megous.com>

arch/arm/include/asm/arch-sunxi/mmc.h

@@ -112,6 +112,10 @@ struct sunxi_mmc {
 	 SUNXI_MMC_RINT_COMMAND_DONE |		\
 	 SUNXI_MMC_RINT_VOLTAGE_CHANGE_DONE)
 
+#define SUNXI_MMC_FTRGLEVEL_BURST_SIZE(v)	(((v) & 0x7) << 28)
+#define SUNXI_MMC_FTRGLEVEL_RX_TL(v)		(((v) & 0xfff) << 16)
+#define SUNXI_MMC_FTRGLEVEL_TX_TL(v)		(((v) & 0xffff) << 0)
+
 #define SUNXI_MMC_STATUS_RXWL_FLAG		(0x1 << 0)
 #define SUNXI_MMC_STATUS_TXWL_FLAG		(0x1 << 1)
 #define SUNXI_MMC_STATUS_FIFO_EMPTY		(0x1 << 2)
@@ -129,6 +133,9 @@ struct sunxi_mmc {
 #define SUNXI_MMC_IDIE_TXIRQ		(0x1 << 0)
 #define SUNXI_MMC_IDIE_RXIRQ		(0x1 << 1)
 
+#define SUNXI_MMC_IDST_TXIRQ		(0x1 << 0)
+#define SUNXI_MMC_IDST_RXIRQ		(0x1 << 1)
+
 #define SUNXI_MMC_COMMON_CLK_GATE		(1 << 16)
 #define SUNXI_MMC_COMMON_RESET			(1 << 18)
 

drivers/mmc/sunxi_mmc.c

@@ -9,6 +9,7 @@
 
 #include <common.h>
 #include <dm.h>
+#include <cpu_func.h>
 #include <errno.h>
 #include <log.h>
 #include <malloc.h>
@@ -23,6 +24,27 @@
 #include <asm-generic/gpio.h>
 #include <linux/delay.h>
 
+#define DMA_CONFIG_DIC BIT(1)  // flag: disable interrupt after this descriptor's buffer is processed
+#define DMA_CONFIG_LAST BIT(2) // flag: last descriptor
+#define DMA_CONFIG_FIRST BIT(3) // flag: first descriptor
+#define DMA_CONFIG_CHAIN BIT(4) // flag: buf_addr_ptr2 points to next descriptor
+#define DMA_CONFIG_ERROR BIT(30) // flag: out: error happened
+#define DMA_CONFIG_HOLD BIT(31) // flag: desc owned by IDMAC (set to 1)
+
+#if defined(CONFIG_MACH_SUN50I)
+// mmc2 on A64 only allows for 8k
+#define DMA_BUF_MAX_SIZE (1 << 13)
+#else
+#define DMA_BUF_MAX_SIZE (1 << 16)
+#endif
+
+struct sunxi_idma_desc {
+        u32 config;
+        u32 buf_size;
+        u32 buf_addr_ptr1;
+        u32 buf_addr_ptr2;
+};
+
 #ifdef CONFIG_DM_MMC
 struct sunxi_mmc_variant {
 	u16 mclk_offset;
@@ -42,6 +64,8 @@ struct sunxi_mmc_priv {
 	int cd_inverted;		/* Inverted Card Detect */
 	struct sunxi_mmc *reg;
 	struct mmc_config cfg;
+	unsigned n_dma_descs;
+	struct sunxi_idma_desc* dma_descs;
 #ifdef CONFIG_DM_MMC
 	const struct sunxi_mmc_variant *variant;
 #endif
@@ -319,7 +343,7 @@ static int mmc_trans_data_by_cpu(struct sunxi_mmc_priv *priv, struct mmc *mmc,
 	if (timeout_msecs < 2000)
 		timeout_msecs = 2000;
 
-	/* Always read / write data through the CPU */
+	/* Read / write data through the CPU */
 	setbits_le32(&priv->reg->gctrl, SUNXI_MMC_GCTRL_ACCESS_BY_AHB);
 
 	start = get_timer(0);
@@ -327,7 +351,7 @@ static int mmc_trans_data_by_cpu(struct sunxi_mmc_priv *priv, struct mmc *mmc,
 	for (i = 0; i < (byte_cnt >> 2); i++) {
 		while (readl(&priv->reg->status) & status_bit) {
 			if (get_timer(start) > timeout_msecs)
-				return -1;
+				return -ETIMEDOUT;
 		}
 
 		if (reading)
@@ -339,21 +363,142 @@ static int mmc_trans_data_by_cpu(struct sunxi_mmc_priv *priv, struct mmc *mmc,
 	return 0;
 }
 
+static void flush_cache_auto_align(void* buf, size_t len)
+{
+	uintptr_t mask = ~((uintptr_t)CONFIG_SYS_CACHELINE_SIZE - 1);
+	uintptr_t start = (uintptr_t)buf & mask;
+
+	len = (len + 2 * CONFIG_SYS_CACHELINE_SIZE) & mask;
+
+	flush_cache(start, len);
+}
+
+static int mmc_trans_data_by_dma(struct sunxi_mmc_priv *priv, struct mmc *mmc,
+				 struct mmc_data *data)
+{
+	const int reading = !!(data->flags & MMC_DATA_READ);
+	uint8_t *buff = (uint8_t*)(reading ? data->dest : data->src);
+	unsigned byte_cnt = data->blocksize * data->blocks;
+	unsigned i, n_desc, last_block_size;
+	u32 rval;
+
+	/* data pointer and transfer size needs to be aligned to 4 bytes */
+
+	/* Read / write data through IDMAC */
+	clrbits_le32(&priv->reg->gctrl, SUNXI_MMC_GCTRL_ACCESS_BY_AHB);
+
+	n_desc = byte_cnt / DMA_BUF_MAX_SIZE;
+	last_block_size = byte_cnt % DMA_BUF_MAX_SIZE;
+	if (last_block_size)
+		n_desc++;
+
+	if (n_desc > priv->n_dma_descs)
+		return -ENOMEM;
+
+	memset(priv->dma_descs, 0, sizeof(struct sunxi_idma_desc) * n_desc);
+
+	for (i = 0; i < n_desc; i++) {
+		struct sunxi_idma_desc* desc = &priv->dma_descs[i];
+		bool is_last = i == n_desc - 1;
+		bool is_first = i == 0;
+
+		desc->config = DMA_CONFIG_CHAIN | DMA_CONFIG_HOLD
+			| (is_last ? DMA_CONFIG_LAST : DMA_CONFIG_DIC)
+			| (is_first ? DMA_CONFIG_FIRST : 0);
+
+		if (is_last && last_block_size)
+			desc->buf_size = last_block_size;
+		else
+			desc->buf_size = DMA_BUF_MAX_SIZE;
+
+		desc->buf_addr_ptr1 = (uintptr_t)buff + i * DMA_BUF_MAX_SIZE;
+		if (!is_last)
+			desc->buf_addr_ptr2 = (uintptr_t)(desc + 1);
+	}
+
+	/*
+	 * Make sure everyhting needed for a transfer is in DRAM.
+	 */
+
+	flush_cache_auto_align(buff, byte_cnt);
+	flush_cache_auto_align(priv->dma_descs,
+			       sizeof(struct sunxi_idma_desc) * n_desc);
+
+	dsb();
+	isb();
+
+	/* dma enable */
+	setbits_le32(&priv->reg->gctrl, SUNXI_MMC_GCTRL_DMA_RESET
+		     | SUNXI_MMC_GCTRL_DMA_ENABLE);
+
+	/* idma reset */
+	writel(SUNXI_MMC_IDMAC_RESET, &priv->reg->dmac);
+
+	/* wait idma reset done */
+	while (readl(&priv->reg->dmac) & SUNXI_MMC_IDMAC_RESET);
+
+	/* idma on */
+	writel(SUNXI_MMC_IDMAC_ENABLE | SUNXI_MMC_IDMAC_FIXBURST,
+	       &priv->reg->dmac);
+
+	/* enable interrupt flags */
+	rval = readl(&priv->reg->idie)
+		& ~(SUNXI_MMC_IDIE_RXIRQ | SUNXI_MMC_IDIE_TXIRQ);
+	rval |= reading ? SUNXI_MMC_IDIE_RXIRQ : SUNXI_MMC_IDIE_TXIRQ;
+	writel(rval, &priv->reg->idie);
+
+	/* set address of the first descriptor */
+	writel((uintptr_t)priv->dma_descs, &priv->reg->dlba);
+
+	/* set fifo fill tresholds for issuing dma */
+
+#if defined(CONFIG_MACH_SUN50I) || defined(CONFIG_MACH_SUN50I_H6)
+	if (priv->mmc_no == 2) {
+		// for mmc 2 we need to set this differently
+		writel(SUNXI_MMC_FTRGLEVEL_BURST_SIZE(3) // burst-16
+		       | SUNXI_MMC_FTRGLEVEL_RX_TL(15)
+		       | SUNXI_MMC_FTRGLEVEL_TX_TL(240),
+		       &priv->reg->ftrglevel);
+	} else {
+		writel(SUNXI_MMC_FTRGLEVEL_BURST_SIZE(2) // burst-8
+		       | SUNXI_MMC_FTRGLEVEL_RX_TL(7)
+		       | SUNXI_MMC_FTRGLEVEL_TX_TL(248),
+		       &priv->reg->ftrglevel);
+	}
+#else
+	writel(SUNXI_MMC_FTRGLEVEL_BURST_SIZE(2) // burst-8
+	       | SUNXI_MMC_FTRGLEVEL_RX_TL(7)
+	       | SUNXI_MMC_FTRGLEVEL_TX_TL(8),
+	       &priv->reg->ftrglevel);
+#endif
+
+	writel(0xffffffff, &priv->reg->idst);
+
+	return 0;
+}
+
 static int mmc_rint_wait(struct sunxi_mmc_priv *priv, struct mmc *mmc,
-			 uint timeout_msecs, uint done_bit, const char *what)
+			 uint timeout_msecs, uint done_bit, bool wait_dma,
+			 const char *what)
 {
 	unsigned int status;
 	unsigned long start = get_timer(0);
+	bool dma_done = true;
 
 	do {
 		status = readl(&priv->reg->rint);
+
 		if ((get_timer(start) > timeout_msecs) ||
 		    (status & SUNXI_MMC_RINT_INTERRUPT_ERROR_BIT)) {
 			debug("%s timeout %x\n", what,
 			      status & SUNXI_MMC_RINT_INTERRUPT_ERROR_BIT);
 			return -ETIMEDOUT;
 		}
-	} while (!(status & done_bit));
+
+		if (wait_dma)
+			dma_done = readl(&priv->reg->idst)
+				& (SUNXI_MMC_IDST_TXIRQ | SUNXI_MMC_IDST_RXIRQ);
+	} while (!(status & done_bit) || !dma_done);
 
 	return 0;
 }
@@ -367,6 +512,7 @@ static int sunxi_mmc_send_cmd_common(struct sunxi_mmc_priv *priv,
 	int error = 0;
 	unsigned int status = 0;
 	unsigned int bytecnt = 0;
+	bool usedma = false;
 
 	if (priv->fatal_err)
 		return -1;
@@ -403,42 +549,45 @@ static int sunxi_mmc_send_cmd_common(struct sunxi_mmc_priv *priv,
 	      cmd->cmdidx, cmdval | cmd->cmdidx, cmd->cmdarg);
 	writel(cmd->cmdarg, &priv->reg->arg);
 
-	if (!data)
-		writel(cmdval | cmd->cmdidx, &priv->reg->cmd);
-
 	/*
 	 * transfer data and check status
 	 * STATREG[2] : FIFO empty
 	 * STATREG[3] : FIFO full
 	 */
 	if (data) {
-		int ret = 0;
-
 		bytecnt = data->blocksize * data->blocks;
 		debug("trans data %d bytes\n", bytecnt);
-		writel(cmdval | cmd->cmdidx, &priv->reg->cmd);
-		ret = mmc_trans_data_by_cpu(priv, mmc, data);
-		if (ret) {
-			error = readl(&priv->reg->rint) &
-				SUNXI_MMC_RINT_INTERRUPT_ERROR_BIT;
-			error = -ETIMEDOUT;
-			goto out;
+
+		if (bytecnt > 64 && !IS_ENABLED(SPL_BUILD)) {
+			debug("  using dma %d\n", bytecnt);
+			error = mmc_trans_data_by_dma(priv, mmc, data);
+			writel(cmdval | cmd->cmdidx, &priv->reg->cmd);
+			usedma = true;
+		} else {
+			debug("  using pio\n");
+			writel(cmdval | cmd->cmdidx, &priv->reg->cmd);
+			error = mmc_trans_data_by_cpu(priv, mmc, data);
 		}
+
+		if (error)
+			goto out;
+	} else {
+		writel(cmdval | cmd->cmdidx, &priv->reg->cmd);
 	}
 
 	error = mmc_rint_wait(priv, mmc, 1000, SUNXI_MMC_RINT_COMMAND_DONE,
-			      "cmd");
+			      false, "cmd");
 	if (error)
 		goto out;
 
 	if (data) {
-		timeout_msecs = 120;
+		timeout_msecs = 10000;
 		debug("cacl timeout %x msec\n", timeout_msecs);
 		error = mmc_rint_wait(priv, mmc, timeout_msecs,
 				      data->blocks > 1 ?
 				      SUNXI_MMC_RINT_AUTO_COMMAND_DONE :
 				      SUNXI_MMC_RINT_DATA_OVER,
-				      "data");
+				      usedma, "data");
 		if (error)
 			goto out;
 	}
@@ -470,6 +619,14 @@ static int sunxi_mmc_send_cmd_common(struct sunxi_mmc_priv *priv,
 		debug("mmc resp 0x%08x\n", cmd->response[0]);
 	}
 out:
+	if (data && usedma) {
+		//status = readl(&reg->idst);
+		writel(0, &priv->reg->idie);
+		writel(0xffffffff, &priv->reg->idst);
+		writel(0, &priv->reg->dmac);
+		clrbits_le32(&priv->reg->gctrl, SUNXI_MMC_GCTRL_DMA_ENABLE);
+	}
+
 	if (error < 0) {
 		writel(SUNXI_MMC_GCTRL_RESET, &priv->reg->gctrl);
 		mmc_update_clk(priv);
@@ -636,6 +793,15 @@ static int sunxi_mmc_probe(struct udevice *dev)
 	priv->variant =
 		(const struct sunxi_mmc_variant *)dev_get_driver_data(dev);
 
+	// make sure we have enough space for descritors for BLK_SIZE * b_max
+	priv->n_dma_descs = 512 * 65536 / DMA_BUF_MAX_SIZE;
+	priv->dma_descs = malloc(sizeof(struct sunxi_idma_desc)
+				 * priv->n_dma_descs);
+	if (priv->dma_descs == NULL) {
+		debug("init mmc alloc failed\n");
+		return -ENOMEM;
+	}
+
 	/* We don't have a sunxi clock driver so find the clock address here */
 	ret = dev_read_phandle_with_args(dev, "clocks", "#clock-cells", 0,
 					  1, &args);