[v3,net-next,1/2] net: dsa: sja1105: send multiple spi_messages instead of using cs_change

Commit Message

Vladimir Oltean May 20, 2021, 9:16 p.m. UTC

From: Vladimir Oltean <[e-mail]>

The sja1105 driver has been described by Mark Brown as "not using the
[ SPI ] API at all idiomatically" due to the use of cs_change:
https://patchwork.kernel.org/project/netdevbpf/patch/[e-mail]/

According to include/linux/spi/spi.h, the chip select is supposed to be
asserted for the entire length of a SPI message, as long as cs_change is
false for all member transfers. The cs_change flag changes the following:

(i) When a non-final SPI transfer has cs_change = true, the chip select
    should temporarily deassert and then reassert starting with the next
    transfer.
(ii) When a final SPI transfer has cs_change = true, the chip select
     should remain asserted until the following SPI message.

The sja1105 driver only uses cs_change for its first property, to form a
single SPI message whose layout can be seen below:

                                             this is an entire, single spi_message
           _______________________________________________________________________________________________
          /                                                                                               \
          +-------------+---------------+-------------+---------------+ ... +-------------+---------------+
          | hdr_xfer[0] | chunk_xfer[0] | hdr_xfer[1] | chunk_xfer[1] |     | hdr_xfer[n] | chunk_xfer[n] |
          +-------------+---------------+-------------+---------------+ ... +-------------+---------------+
cs_change      false          true           false           true                false          false

           ____________________________  _____________________________       _____________________________
CS line __/                            \/                             \ ... /                             \__

The fact of the matter is that spi_max_message_size() has an ambiguous
meaning if any non-final transfer has cs_change = true.

If the SPI master has a limitation in that it cannot keep the chip
select asserted for more than, say, 200 bytes (like the spi-sc18is602),
the normal thing for it to do is to implement .max_transfer_size and
.max_message_size, and limit both to 200: in the "worst case" where
cs_change is always false, then the controller can, indeed, not send
messages larger than 200 bytes.

But the fact that the SPI controller's max_message_size does not
necessarily mean that we cannot send messages larger than that.
Notably, if the SPI master special-cases the transfers with cs_change
and treats every chip select toggling as an entirely new transaction,
then a SPI message can easily exceed that limit. So there is a
temptation to ignore the controller's reported max_message_size when
using cs_change = true in non-final transfers.

But that can lead to false conclusions. As Mark points out, the SPI
controller might have a different kind of limitation with the max
message size, that has nothing at all to do with how long it can keep
the chip select asserted.
For example, that might be the case if the device is able to offload the
chip select changes to the hardware as part of the data stream, and it
packs the entire stream of commands+data (corresponding to a SPI
message) into a single DMA transfer that is itself limited in size.

So the only thing we can do is avoid ambiguity by not using cs_change at
all. Instead of sending a single spi_message, we now send multiple SPI
messages as follows:

                  spi_message 0                 spi_message 1                       spi_message n
           ____________________________   ___________________________        _____________________________
          /                            \ /                           \      /                             \
          +-------------+---------------+-------------+---------------+ ... +-------------+---------------+
          | hdr_xfer[0] | chunk_xfer[0] | hdr_xfer[1] | chunk_xfer[1] |     | hdr_xfer[n] | chunk_xfer[n] |
          +-------------+---------------+-------------+---------------+ ... +-------------+---------------+
cs_change      false          true           false           true                false          false

           ____________________________  _____________________________       _____________________________
CS line __/                            \/                             \ ... /                             \__

which is clearer because the max_message_size limit is now easier to
enforce. What is transmitted on the wire stays, of course, the same.

Additionally, because we send no more than 2 transfers at a time, we now
avoid dynamic memory allocation too, which might be seen as an
improvement by some.

Signed-off-by: Vladimir Oltean <[e-mail]>
---
 drivers/net/dsa/sja1105/sja1105_spi.c | 52 +++++++--------------------
 1 file changed, 12 insertions(+), 40 deletions(-)

Comments

Mark Brown May 24, 2021, 8:35 a.m. UTC | #1

On Fri, May 21, 2021 at 12:16:56AM +0300, Vladimir Oltean wrote:

> The fact of the matter is that spi_max_message_size() has an ambiguous
> meaning if any non-final transfer has cs_change = true.

This is not the case, spi_message_max_size() is a limit on the size of a
spi_message.

Vladimir Oltean May 24, 2021, 1:02 p.m. UTC | #2

On Mon, May 24, 2021 at 09:35:29AM +0100, Mark Brown wrote:
> On Fri, May 21, 2021 at 12:16:56AM +0300, Vladimir Oltean wrote:
> 
> > The fact of the matter is that spi_max_message_size() has an ambiguous
> > meaning if any non-final transfer has cs_change = true.
> 
> This is not the case, spi_message_max_size() is a limit on the size of a
> spi_message.

That is true, although it doesn't mean much, since in the presence of
cs_change, a spi_message has no correspondent in the physical world
(i.e. you can't look at a logic analyzer dump and say "this spi_message
was from this to this point"), and that is the problem really.
Describing the controller's inability to send more than N SPI words with
continuous chip select using spi_message_max_size() is what seems flawed
to me, but it's what we have, and what I've adapted to.

Mark Brown June 7, 2021, 5:56 p.m. UTC | #3

On Mon, May 24, 2021 at 04:02:12PM +0300, Vladimir Oltean wrote:
> On Mon, May 24, 2021 at 09:35:29AM +0100, Mark Brown wrote:

> > This is not the case, spi_message_max_size() is a limit on the size of a
> > spi_message.

> That is true, although it doesn't mean much, since in the presence of
> cs_change, a spi_message has no correspondent in the physical world
> (i.e. you can't look at a logic analyzer dump and say "this spi_message
> was from this to this point"), and that is the problem really.

It may affect how things are implemented by the driver, for example if
the driver can send a command stream to the hardware the limit might be
due to that command stream.  There is no need or expectation for drivers
to pattern match what the're being asked to do and parse out something
that should be a string of messages from the spi_message they get, it is
expected that client drivers should split things up naturally.

> Describing the controller's inability to send more than N SPI words with
> continuous chip select using spi_message_max_size() is what seems flawed
> to me, but it's what we have, and what I've adapted to.

I can't entirely parse that but the limit here isn't to do with how long
chip select is asserted for.

Patch

diff --git a/drivers/net/dsa/sja1105/sja1105_spi.c b/drivers/net/dsa/sja1105/sja1105_spi.c
index f7a1514f81e8..8746e3f158a0 100644
--- a/drivers/net/dsa/sja1105/sja1105_spi.c
+++ b/drivers/net/dsa/sja1105/sja1105_spi.c
@@ -29,13 +29,6 @@  sja1105_spi_message_pack(void *buf, const struct sja1105_spi_message *msg)
 	sja1105_pack(buf, &msg->address,    24,  4, size);
 }
 
-#define sja1105_hdr_xfer(xfers, chunk) \
-	((xfers) + 2 * (chunk))
-#define sja1105_chunk_xfer(xfers, chunk) \
-	((xfers) + 2 * (chunk) + 1)
-#define sja1105_hdr_buf(hdr_bufs, chunk) \
-	((hdr_bufs) + (chunk) * SJA1105_SIZE_SPI_MSG_HEADER)
-
 /* If @rw is:
  * - SPI_WRITE: creates and sends an SPI write message at absolute
  *		address reg_addr, taking @len bytes from *buf
@@ -46,41 +39,25 @@  static int sja1105_xfer(const struct sja1105_private *priv,
 			sja1105_spi_rw_mode_t rw, u64 reg_addr, u8 *buf,
 			size_t len, struct ptp_system_timestamp *ptp_sts)
 {
+	u8 hdr_buf[SJA1105_SIZE_SPI_MSG_HEADER] = {0};
 	struct sja1105_chunk chunk = {
 		.len = min_t(size_t, len, SJA1105_SIZE_SPI_MSG_MAXLEN),
 		.reg_addr = reg_addr,
 		.buf = buf,
 	};
 	struct spi_device *spi = priv->spidev;
-	struct spi_transfer *xfers;
+	struct spi_transfer xfers[2] = {0};
+	struct spi_transfer *chunk_xfer;
+	struct spi_transfer *hdr_xfer;
 	int num_chunks;
 	int rc, i = 0;
-	u8 *hdr_bufs;
 
 	num_chunks = DIV_ROUND_UP(len, SJA1105_SIZE_SPI_MSG_MAXLEN);
 
-	/* One transfer for each message header, one for each message
-	 * payload (chunk).
-	 */
-	xfers = kcalloc(2 * num_chunks, sizeof(struct spi_transfer),
-			GFP_KERNEL);
-	if (!xfers)
-		return -ENOMEM;
-
-	/* Packed buffers for the num_chunks SPI message headers,
-	 * stored as a contiguous array
-	 */
-	hdr_bufs = kcalloc(num_chunks, SJA1105_SIZE_SPI_MSG_HEADER,
-			   GFP_KERNEL);
-	if (!hdr_bufs) {
-		kfree(xfers);
-		return -ENOMEM;
-	}
+	hdr_xfer = &xfers[0];
+	chunk_xfer = &xfers[1];
 
 	for (i = 0; i < num_chunks; i++) {
-		struct spi_transfer *chunk_xfer = sja1105_chunk_xfer(xfers, i);
-		struct spi_transfer *hdr_xfer = sja1105_hdr_xfer(xfers, i);
-		u8 *hdr_buf = sja1105_hdr_buf(hdr_bufs, i);
 		struct spi_transfer *ptp_sts_xfer;
 		struct sja1105_spi_message msg;
 
@@ -129,19 +106,14 @@  static int sja1105_xfer(const struct sja1105_private *priv,
 		chunk.len = min_t(size_t, (ptrdiff_t)(buf + len - chunk.buf),
 				  SJA1105_SIZE_SPI_MSG_MAXLEN);
 
-		/* De-assert the chip select after each chunk. */
-		if (chunk.len)
-			chunk_xfer->cs_change = 1;
+		rc = spi_sync_transfer(spi, xfers, 2);
+		if (rc < 0) {
+			dev_err(&spi->dev, "SPI transfer failed: %d\n", rc);
+			return rc;
+		}
 	}
 
-	rc = spi_sync_transfer(spi, xfers, 2 * num_chunks);
-	if (rc < 0)
-		dev_err(&spi->dev, "SPI transfer failed: %d\n", rc);
-
-	kfree(hdr_bufs);
-	kfree(xfers);
-
-	return rc;
+	return 0;
 }
 
 int sja1105_xfer_buf(const struct sja1105_private *priv,