staging: mt7621-spi: revised half-duplex message handling

The mt7621 SPI engine has a 32 byte buffer and the driver
currently only allows 32-byte read requests and 36 bytes writes
(there is a 4byte op/addr buffer).

This is an unnecessary limitation.  As the SPI clock is controlled
by the host it is quite acceptable to send a larger message in
multiple smaller transactions.  As long as Chip Select is kept asserted
the whole time, the SPI engine can be run multiple times for
a single SPI message.

This patch factors out the transaction logic and calls for each
transfer in the message.  A write transfer might leave bytes in the
buffer to be combined with a following read transfer, as this is
a common pattern.

With this in place, we can remove the current max_transfer_size limit.

In testing, this increases the read throughput for a NOR flash chip
from 1.4MB/s to 2.3MB/s, a 50% improvement.
Signed-off-by: NeilBrown <neil@brown.name>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

drivers/staging/mt7621-spi/spi-mt7621.c

@@ -65,6 +65,7 @@ struct mt7621_spi {
 	unsigned int		sys_freq;
 	unsigned int		speed;
 	struct clk		*clk;
+	int			pending_write;
 
 	struct mt7621_spi_ops	*ops;
 };
@@ -96,6 +97,7 @@ static void mt7621_spi_reset(struct mt7621_spi *rs, int duplex)
 		master &= ~(1 << 10);
 
 	mt7621_spi_write(rs, MT7621_SPI_MASTER, master);
+	rs->pending_write = 0;
 }
 
 static void mt7621_spi_set_cs(struct spi_device *spi, int enable)
@@ -173,90 +175,124 @@ static inline int mt7621_spi_wait_till_ready(struct mt7621_spi *rs)
 	return -ETIMEDOUT;
 }
 
-static int mt7621_spi_transfer_half_duplex(struct spi_master *master,
-					   struct spi_message *m)
+static void mt7621_spi_read_half_duplex(struct mt7621_spi *rs,
+					int rx_len, u8 *buf)
 {
-	struct mt7621_spi *rs = spi_master_get_devdata(master);
-	struct spi_device *spi = m->spi;
-	unsigned int speed = spi->max_speed_hz;
-	struct spi_transfer *t = NULL;
-	int status = 0;
-	int i, len = 0;
-	int rx_len = 0;
-	u32 data[9] = { 0 };
-	u32 val;
+	/* Combine with any pending write, and perform one or
+	 * more half-duplex transactions reading 'len' bytes.
+	 * Data to be written is already in MT7621_SPI_DATA*
+	 */
+	int tx_len = rs->pending_write;
 
-	mt7621_spi_wait_till_ready(rs);
+	rs->pending_write = 0;
 
-	list_for_each_entry(t, &m->transfers, transfer_list) {
-		const u8 *buf = t->tx_buf;
+	while (rx_len || tx_len) {
+		int i;
+		u32 val = (min(tx_len, 4) * 8) << 24;
+		int rx = min(rx_len, 32);
 
-		if (t->rx_buf)
-			rx_len += t->len;
+		if (tx_len > 4)
+			val |= (tx_len - 4) * 8;
+		val |= (rx * 8) << 12;
+		mt7621_spi_write(rs, MT7621_SPI_MOREBUF, val);
 
-		if (!buf)
-			continue;
+		tx_len = 0;
 
-		if (t->speed_hz < speed)
-			speed = t->speed_hz;
+		val = mt7621_spi_read(rs, MT7621_SPI_TRANS);
+		val |= SPI_CTL_START;
+		mt7621_spi_write(rs, MT7621_SPI_TRANS, val);
 
-		if (WARN_ON(len + t->len > 36)) {
-			status = -EIO;
-			goto msg_done;
-		}
+		mt7621_spi_wait_till_ready(rs);
 
-		for (i = 0; i < t->len; i++, len++)
-			data[len / 4] |= buf[i] << (8 * (len & 3));
+		for (i = 0; i < rx; i++) {
+			if ((i % 4) == 0)
+				val = mt7621_spi_read(rs, MT7621_SPI_DATA0 + i);
+			*buf++ = val & 0xff;
+			val >>= 8;
+		}
+		rx_len -= i;
 	}
+}
 
-	if (WARN_ON(rx_len > 32)) {
-		status = -EIO;
-		goto msg_done;
-	}
+static inline void mt7621_spi_flush(struct mt7621_spi *rs)
+{
+	mt7621_spi_read_half_duplex(rs, 0, NULL);
+}
 
-	if (mt7621_spi_prepare(spi, speed)) {
-		status = -EIO;
-		goto msg_done;
+static void mt7621_spi_write_half_duplex(struct mt7621_spi *rs,
+					 int tx_len, const u8 *buf)
+{
+	int val = 0;
+	int len = rs->pending_write;
+
+	if (len & 3) {
+		val = mt7621_spi_read(rs, MT7621_SPI_OPCODE + (len & ~3));
+		if (len < 4) {
+			val <<= (4 - len) * 8;
+			val = swab32(val);
+		}
 	}
-	data[0] = swab32(data[0]);
-	if (len < 4)
-		data[0] >>= (4 - len) * 8;
 
-	for (i = 0; i < len; i += 4)
-		mt7621_spi_write(rs, MT7621_SPI_OPCODE + i, data[i / 4]);
-
-	val = (min_t(int, len, 4) * 8) << 24;
-	if (len > 4)
-		val |= (len - 4) * 8;
-	val |= (rx_len * 8) << 12;
-	mt7621_spi_write(rs, MT7621_SPI_MOREBUF, val);
+	while (tx_len > 0) {
+		if (len >= 36) {
+			rs->pending_write = len;
+			mt7621_spi_flush(rs);
+			len = 0;
+		}
 
-	mt7621_spi_set_cs(spi, 1);
+		val |= *buf++ << (8 * (len & 3));
+		len++;
+		if ((len & 3) == 0) {
+			if (len == 4)
+				/* The byte-order of the opcode is weird! */
+				val = swab32(val);
+			mt7621_spi_write(rs, MT7621_SPI_OPCODE + len - 4, val);
+			val = 0;
+		}
+		tx_len -= 1;
+	}
+	if (len & 3) {
+		if (len < 4) {
+			val = swab32(val);
+			val >>= (4 - len) * 8;
+		}
+		mt7621_spi_write(rs, MT7621_SPI_OPCODE + (len & ~3), val);
+	}
+	rs->pending_write = len;
+}
 
-	val = mt7621_spi_read(rs, MT7621_SPI_TRANS);
-	val |= SPI_CTL_START;
-	mt7621_spi_write(rs, MT7621_SPI_TRANS, val);
+static int mt7621_spi_transfer_half_duplex(struct spi_master *master,
+					   struct spi_message *m)
+{
+	struct mt7621_spi *rs = spi_master_get_devdata(master);
+	struct spi_device *spi = m->spi;
+	unsigned int speed = spi->max_speed_hz;
+	struct spi_transfer *t = NULL;
+	int status = 0;
 
 	mt7621_spi_wait_till_ready(rs);
 
-	mt7621_spi_set_cs(spi, 0);
-
-	for (i = 0; i < rx_len; i += 4)
-		data[i / 4] = mt7621_spi_read(rs, MT7621_SPI_DATA0 + i);
+	list_for_each_entry(t, &m->transfers, transfer_list)
+		if (t->speed_hz < speed)
+			speed = t->speed_hz;
 
-	m->actual_length = len + rx_len;
+	if (mt7621_spi_prepare(spi, speed)) {
+		status = -EIO;
+		goto msg_done;
+	}
 
-	len = 0;
+	mt7621_spi_set_cs(spi, 1);
+	m->actual_length = 0;
 	list_for_each_entry(t, &m->transfers, transfer_list) {
-		u8 *buf = t->rx_buf;
-
-		if (!buf)
-			continue;
-
-		for (i = 0; i < t->len; i++, len++)
-			buf[i] = data[len / 4] >> (8 * (len & 3));
+		if (t->rx_buf)
+			mt7621_spi_read_half_duplex(rs, t->len, t->rx_buf);
+		else if (t->tx_buf)
+			mt7621_spi_write_half_duplex(rs, t->len, t->tx_buf);
+		m->actual_length += t->len;
 	}
+	mt7621_spi_flush(rs);
 
+	mt7621_spi_set_cs(spi, 0);
 msg_done:
 	m->status = status;
 	spi_finalize_current_message(master);
@@ -383,11 +419,6 @@ static const struct of_device_id mt7621_spi_match[] = {
 };
 MODULE_DEVICE_TABLE(of, mt7621_spi_match);
 
-static size_t max_transfer_size(struct spi_device *spi)
-{
-	return 32;
-}
-
 static int mt7621_spi_probe(struct platform_device *pdev)
 {
 	const struct of_device_id *match;
@@ -433,7 +464,6 @@ static int mt7621_spi_probe(struct platform_device *pdev)
 	master->bits_per_word_mask = SPI_BPW_MASK(8);
 	master->dev.of_node = pdev->dev.of_node;
 	master->num_chipselect = 2;
-	master->max_transfer_size = max_transfer_size;
 
 	dev_set_drvdata(&pdev->dev, master);
 
@@ -443,6 +473,7 @@ static int mt7621_spi_probe(struct platform_device *pdev)
 	rs->master = master;
 	rs->sys_freq = clk_get_rate(rs->clk);
 	rs->ops = ops;
+	rs->pending_write = 0;
 	dev_info(&pdev->dev, "sys_freq: %u\n", rs->sys_freq);
 
 	device_reset(&pdev->dev);