[PATCH 0/2] dmaengine: bcm2835: Cyclic DMA fixes

In downstream Raspberry Pi kernel we noticed that audio didn't work
as expected, we got stuttering and overruns/underruns. Here's the
link to the original discussion on GitHub:
https://github.com/raspberrypi/linux/issues/1517

This issue is caused by a small bug in the period-splitting-code
and fixed by the first patch.

The second patch, avoiding very small chunks, is mainly a precaution.
While small chunks are not known to have caused any problems so far
they have the potentical to cause very hard to track down issues.
So better avoid such situations in the first place.

Matthias Reichl (2):
  dmaengine: bcm2835: Fix cyclic DMA period splitting
  dmaengine: bcm2835: Avoid splitting periods into very small chunks

 drivers/dma/bcm2835-dma.c | 19 ++++++++++++++++++-
 1 file changed, 18 insertions(+), 1 deletion(-)

-- 
2.1.4

The current cyclic DMA period splitting implementation can generate
very small chunks at the end of each period. For example a 65536 byte
period will be split into a 65532 byte chunk and a 4 byte chunk on
the "lite" DMA channels.

This increases pressure on the RAM controller as the DMA controller
needs to fetch two control blocks from RAM in quick succession and
could potentially cause latency issues if the RAM is tied up by other
devices.

We can easily avoid these situations by distributing the remaining
length evenly between the last-but-one and the last chunk, making
sure that split chunks will be at least half the maximum length the
DMA controller can handle.

This patch checks if the last chunk would be less than half of
the maximum DMA length and if yes distributes the max len+4...max_len*1.5
bytes evenly between the last 2 chunks. This results in chunk sizes
between max_len/2 and max_len*0.75 bytes.

Signed-off-by: Matthias Reichl <redacted>
Signed-off-by: Martin Sperl <redacted>
Tested-by: Clive Messer <redacted>
---
 drivers/dma/bcm2835-dma.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/drivers/dma/bcm2835-dma.c b/drivers/dma/bcm2835-dma.c
index 344bcf92..36b998d 100644
--- a/drivers/dma/bcm2835-dma.c
+++ b/drivers/dma/bcm2835-dma.c

@@ -252,6 +252,20 @@ static void bcm2835_dma_create_cb_set_length(
 
 	/* have we filled in period_length yet? */
 	if (*total_len + control_block->length < period_len) {
+		/*
+		 * If the next control block is the last in the period
+		 * and it's length would be less than half of max_len
+		 * change it so that both control blocks are (almost)
+		 * equally long. This avoids generating very short
+		 * control blocks (worst case would be 4 bytes) which
+		 * might be problematic. We also have to make sure the
+		 * new length is a multiple of 4 bytes.
+		 */
+		if (*total_len + control_block->length + max_len / 2 >
+		    period_len) {
+			control_block->length =
+				DIV_ROUND_UP(period_len - *total_len, 8) * 4;
+		}
 		/* update number of bytes in this period so far */
 		*total_len += control_block->length;
 		return;

-- 
2.1.4

The code responsible for splitting periods into chunks that
can be handled by the DMA controller missed to update total_len,
the number of bytes processed in the current period, when there
are more chunks to follow.

Therefore total_len was stuck at 0 and the code didn't work at all.
This resulted in a wrong control block layout and audio issues because
the cyclic DMA callback wasn't executing on period boundaries.

Fix this by adding the missing total_len update.

Signed-off-by: Matthias Reichl <redacted>
Signed-off-by: Martin Sperl <redacted>
Tested-by: Clive Messer <redacted>
---
 drivers/dma/bcm2835-dma.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/dma/bcm2835-dma.c b/drivers/dma/bcm2835-dma.c
index 6149b27..344bcf92 100644
--- a/drivers/dma/bcm2835-dma.c
+++ b/drivers/dma/bcm2835-dma.c

@@ -251,8 +251,11 @@ static void bcm2835_dma_create_cb_set_length(
 	 */
 
 	/* have we filled in period_length yet? */
-	if (*total_len + control_block->length < period_len)
+	if (*total_len + control_block->length < period_len) {
+		/* update number of bytes in this period so far */
+		*total_len += control_block->length;
 		return;
+	}
 
 	/* calculate the length that remains to reach period_length */
 	control_block->length = period_len - *total_len;

-- 
2.1.4

Matthias Reichl [off-list ref] writes:

The code responsible for splitting periods into chunks that
can be handled by the DMA controller missed to update total_len,
the number of bytes processed in the current period, when there
are more chunks to follow.

Therefore total_len was stuck at 0 and the code didn't work at all.
This resulted in a wrong control block layout and audio issues because
the cyclic DMA callback wasn't executing on period boundaries.

Fix this by adding the missing total_len update.

It looks like this issue has been around for a long time, and this fix
is pretty dependent on the recent refactors.

Reviewed-by: Eric Anholt <redacted>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 818 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20160613/57a4493c/attachment.sig>

Matthias Reichl [off-list ref] writes:

It seems to me like this would all be a lot simpler if we always split
the last 2 control blocks evenly (other than 4-byte rounding):

u32 period_remaining = period_len - *total_len;

/* Early exit if we aren't finishing this period */
if (period_remaining >= max_len) {
	/*
	 * Split the length between the last 2 CBs, to help hide the
	 * latency of fetching the CBs.
	 */
	if (period_remaining < max_len * 2) {
		control_block->length =
                	DIV_ROUND_UP(period_remaining, 8) * 4;
        }
	/* update number of bytes in this period so far */
	*total_len += control_block->length;
}

I'm about to go semi-AFK for a couple weeks.  If there's a good reason
to only do this when the last block is very short, I'm fine with:

Acked-by: Eric Anholt <redacted>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 818 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20160613/51301430/attachment.sig>

On Mon, Jun 13, 2016 at 10:06:49PM -0700, Eric Anholt wrote:

Matthias Reichl [off-list ref] writes:

quoted

The current cyclic DMA period splitting implementation can generate
very small chunks at the end of each period. For example a 65536 byte
period will be split into a 65532 byte chunk and a 4 byte chunk on
the "lite" DMA channels.

This increases pressure on the RAM controller as the DMA controller
needs to fetch two control blocks from RAM in quick succession and
could potentially cause latency issues if the RAM is tied up by other
devices.

We can easily avoid these situations by distributing the remaining
length evenly between the last-but-one and the last chunk, making
sure that split chunks will be at least half the maximum length the
DMA controller can handle.

This patch checks if the last chunk would be less than half of
the maximum DMA length and if yes distributes the max len+4...max_len*1.5
bytes evenly between the last 2 chunks. This results in chunk sizes
between max_len/2 and max_len*0.75 bytes.

Signed-off-by: Matthias Reichl <redacted>
Signed-off-by: Martin Sperl <redacted>
Tested-by: Clive Messer <redacted>
---
 drivers/dma/bcm2835-dma.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/drivers/dma/bcm2835-dma.c b/drivers/dma/bcm2835-dma.c
index 344bcf92..36b998d 100644
--- a/drivers/dma/bcm2835-dma.c
+++ b/drivers/dma/bcm2835-dma.c

@@ -252,6 +252,20 @@ static void bcm2835_dma_create_cb_set_length(
 
 	/* have we filled in period_length yet? */
 	if (*total_len + control_block->length < period_len) {
+		/*
+		 * If the next control block is the last in the period
+		 * and it's length would be less than half of max_len
+		 * change it so that both control blocks are (almost)
+		 * equally long. This avoids generating very short
+		 * control blocks (worst case would be 4 bytes) which
+		 * might be problematic. We also have to make sure the
+		 * new length is a multiple of 4 bytes.
+		 */
+		if (*total_len + control_block->length + max_len / 2 >
+		    period_len) {
+			control_block->length =
+				DIV_ROUND_UP(period_len - *total_len, 8) * 4;
+		}
 		/* update number of bytes in this period so far */
 		*total_len += control_block->length;
 		return;

It seems to me like this would all be a lot simpler if we always split
the last 2 control blocks evenly (other than 4-byte rounding):

Agreed and thanks a lot for the feedback!

I'll do it that way and then send out a v2.

u32 period_remaining = period_len - *total_len;

/* Early exit if we aren't finishing this period */
if (period_remaining >= max_len) {

This has to be > max_len, but the rest seems fine. We want to split
if we have more than max_len but less than max_len*2 bytes.

	/*
	 * Split the length between the last 2 CBs, to help hide the
	 * latency of fetching the CBs.
	 */
	if (period_remaining < max_len * 2) {
		control_block->length =
                	DIV_ROUND_UP(period_remaining, 8) * 4;
        }
	/* update number of bytes in this period so far */
	*total_len += control_block->length;
}

I'm about to go semi-AFK for a couple weeks.  If there's a good reason
to only do this when the last block is very short, I'm fine with:

Acked-by: Eric Anholt <redacted>