mxs/spi: Fix issues when doing long continuous transfer

When doing long continuous transfer, eg. from SPI flash via /dev/mtd,
the driver dies. This is caused by a bug in the DMA chaining. Rework
the DMA transfer code so that this issue does not happen any longer.

This involves proper allocation of correct amount of sg-list members.
Also, this means proper creation of DMA descriptors. There is actually an
important catch to this, the data transfer descriptors must be interleaved
with PIO register write descriptor, otherwise the transfer stalls. This
can be done in one descriptor, but due to the limitation of the DMA API,
it's not possible.

It turns out that in order for the SPI DMA to properly support
continuous transfers longer than 65280 bytes, there are some very
important parts that were left out from the documentation about about
the PIO transfer that is used.

Firstly, the XFER_SIZE register is not written with the whole length
of a transfer, but is written by each and every chained descriptor
with the length of the descriptors data buffer.

Next, unlike the demo code supplied by FSL, which only writes one PIO
word per descriptor, this does not apply if the descriptors are chained,
since the XFER_SIZE register must be written. Therefore, it is essential
to use four PIO words, CTRL0, CMD0, CMD1, XFER_SIZE. CMD0 and CMD1 are
written with zero, since they don't apply. The DMA programs the PIO words
in an incrementing order, so four PIO words.

Finally, unlike the demo code supplied by FSL, the SSP_CTRL0_IGNORE_CRC
must not be set during the whole transfer, but it must be set only on the
last descriptor in the chain.

Lastly, this code lends code from drivers/mtd/nand/omap2.c, which solves
trouble when the buffer supplied to the DMA transfer was vmalloc()'d. So
with this patch, it's safe to use /dev/mtdblockX interface again.

Signed-off-by: Marek Vasut <marex@denx.de>
Signed-off-by: Mark Brown <broonie@opensource.wolfsonmicro.com>

1 files changed, 88 insertions, 53 deletions

diff --git a/drivers/spi/spi-mxs.c b/drivers/spi/spi-mxs.c
index 10d34ebe9ca..bcba098e97c 100644
--- a/drivers/spi/spi-mxs.c
+++ b/drivers/spi/spi-mxs.c
@@ -53,9 +53,9 @@
 
 #define DRIVER_NAME		"mxs-spi"
 
-#define SSP_TIMEOUT		1000	/* 1000 ms */
+/* Use 10S timeout for very long transfers, it should suffice. */
+#define SSP_TIMEOUT		10000
 
-#define SG_NUM			4
 #define SG_MAXLEN		0xff00
 
 struct mxs_spi {
@@ -219,61 +219,94 @@ static int mxs_spi_txrx_dma(struct mxs_spi *spi, int cs,
 			    int *first, int *last, int write)
 {
 	struct mxs_ssp *ssp = &spi->ssp;
-	struct dma_async_tx_descriptor *desc;
-	struct scatterlist sg[SG_NUM];
+	struct dma_async_tx_descriptor *desc = NULL;
+	const bool vmalloced_buf = is_vmalloc_addr(buf);
+	const int desc_len = vmalloced_buf ? PAGE_SIZE : SG_MAXLEN;
+	const int sgs = DIV_ROUND_UP(len, desc_len);
 	int sg_count;
-	uint32_t pio = BM_SSP_CTRL0_DATA_XFER | mxs_spi_cs_to_reg(cs);
-	int ret;
-
-	if (len > SG_NUM * SG_MAXLEN) {
-		dev_err(ssp->dev, "Data chunk too big for DMA\n");
+	int min, ret;
+	uint32_t ctrl0;
+	struct page *vm_page;
+	void *sg_buf;
+	struct {
+		uint32_t		pio[4];
+		struct scatterlist	sg;
+	} *dma_xfer;
+
+	if (!len)
 		return -EINVAL;
-	}
+
+	dma_xfer = kzalloc(sizeof(*dma_xfer) * sgs, GFP_KERNEL);
+	if (!dma_xfer)
+		return -ENOMEM;
 
 	INIT_COMPLETION(spi->c);
 
+	ctrl0 = readl(ssp->base + HW_SSP_CTRL0);
+	ctrl0 |= BM_SSP_CTRL0_DATA_XFER | mxs_spi_cs_to_reg(cs);
+
 	if (*first)
-		pio |= BM_SSP_CTRL0_LOCK_CS;
-	if (*last)
-		pio |= BM_SSP_CTRL0_IGNORE_CRC;
+		ctrl0 |= BM_SSP_CTRL0_LOCK_CS;
 	if (!write)
-		pio |= BM_SSP_CTRL0_READ;
-
-	if (ssp->devid == IMX23_SSP)
-		pio |= len;
-	else
-		writel(len, ssp->base + HW_SSP_XFER_SIZE);
-
-	/* Queue the PIO register write transfer. */
-	desc = dmaengine_prep_slave_sg(ssp->dmach,
-			(struct scatterlist *)&pio,
-			1, DMA_TRANS_NONE, 0);
-	if (!desc) {
-		dev_err(ssp->dev,
-			"Failed to get PIO reg. write descriptor.\n");
-		return -EINVAL;
-	}
+		ctrl0 |= BM_SSP_CTRL0_READ;
 
 	/* Queue the DMA data transfer. */
-	sg_init_table(sg, (len / SG_MAXLEN) + 1);
-	sg_count = 0;
-	while (len) {
-		sg_set_buf(&sg[sg_count++], buf, min(len, SG_MAXLEN));
-		len -= min(len, SG_MAXLEN);
-		buf += min(len, SG_MAXLEN);
-	}
-	dma_map_sg(ssp->dev, sg, sg_count,
-		write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
-
-	desc = dmaengine_prep_slave_sg(ssp->dmach, sg, sg_count,
-			write ? DMA_MEM_TO_DEV : DMA_DEV_TO_MEM,
-			DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
-
-	if (!desc) {
-		dev_err(ssp->dev,
-			"Failed to get DMA data write descriptor.\n");
-		ret = -EINVAL;
-		goto err;
+	for (sg_count = 0; sg_count < sgs; sg_count++) {
+		min = min(len, desc_len);
+
+		/* Prepare the transfer descriptor. */
+		if ((sg_count + 1 == sgs) && *last)
+			ctrl0 |= BM_SSP_CTRL0_IGNORE_CRC;
+
+		if (ssp->devid == IMX23_SSP)
+			ctrl0 |= min;
+
+		dma_xfer[sg_count].pio[0] = ctrl0;
+		dma_xfer[sg_count].pio[3] = min;
+
+		if (vmalloced_buf) {
+			vm_page = vmalloc_to_page(buf);
+			if (!vm_page) {
+				ret = -ENOMEM;
+				goto err_vmalloc;
+			}
+			sg_buf = page_address(vm_page) +
+				((size_t)buf & ~PAGE_MASK);
+		} else {
+			sg_buf = buf;
+		}
+
+		sg_init_one(&dma_xfer[sg_count].sg, sg_buf, min);
+		ret = dma_map_sg(ssp->dev, &dma_xfer[sg_count].sg, 1,
+			write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
+
+		len -= min;
+		buf += min;
+
+		/* Queue the PIO register write transfer. */
+		desc = dmaengine_prep_slave_sg(ssp->dmach,
+				(struct scatterlist *)dma_xfer[sg_count].pio,
+				(ssp->devid == IMX23_SSP) ? 1 : 4,
+				DMA_TRANS_NONE,
+				sg_count ? DMA_PREP_INTERRUPT : 0);
+		if (!desc) {
+			dev_err(ssp->dev,
+				"Failed to get PIO reg. write descriptor.\n");
+			ret = -EINVAL;
+			goto err_mapped;
+		}
+
+		desc = dmaengine_prep_slave_sg(ssp->dmach,
+				&dma_xfer[sg_count].sg, 1,
+				write ? DMA_MEM_TO_DEV : DMA_DEV_TO_MEM,
+				DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
+
+		if (!desc) {
+			dev_err(ssp->dev,
+				"Failed to get DMA data write descriptor.\n");
+			ret = -EINVAL;
+			goto err_mapped;
+		}
 	}
 
 	/*
@@ -289,21 +322,23 @@ static int mxs_spi_txrx_dma(struct mxs_spi *spi, int cs,
 
 	ret = wait_for_completion_timeout(&spi->c,
 				msecs_to_jiffies(SSP_TIMEOUT));
-
 	if (!ret) {
 		dev_err(ssp->dev, "DMA transfer timeout\n");
 		ret = -ETIMEDOUT;
-		goto err;
+		goto err_vmalloc;
 	}
 
 	ret = 0;
 
-err:
-	for (--sg_count; sg_count >= 0; sg_count--) {
-		dma_unmap_sg(ssp->dev, &sg[sg_count], 1,
+err_vmalloc:
+	while (--sg_count >= 0) {
+err_mapped:
+		dma_unmap_sg(ssp->dev, &dma_xfer[sg_count].sg, 1,
 			write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 	}
 
+	kfree(dma_xfer);
+
 	return ret;
 }