The residue calculation in pl330_tx_status doesn't handle transitional
states that occur at the time one descriptor (A) is completed and the
next (B) is started. Specifically, both A and B can simultaneously be in
the BUSY state and at this time the thread's 'req_running' may (or may
not) be -1.
To cope with this situation we change the code to ensure A is treated as
complete and B as having not yet started. Prior to the change, the code
would calculate a transferred byte count as if both A and B had
completed.
Fixes: aee4d1fac887 ("dmaengine: pl330: improve pl330_tx_status() function")
Signed-off-by: Jon Medhurst <tixy@xxxxxxxxxx>
---
I discovered this issue when trying to work out why audio stopped
working on ARM's Juno platform and bisected it to commit aee4d1fac887.
Whilst this patch seems to fix the problems I was seeing, I can't help
but think there are more race conditions with this code. E.g. if the
running descriptor changes under us, pl330_get_current_xferred_count
can end up reading values from hardware that relate to a different
descriptor. And if we're really unlucky, the reading of the 'val' and
'addr' values in pl330_get_current_xferred_count can come from different
descriptors. I don't know if there is any locks we can use to prevent
such races or if we need to try and detect when things have changed and
redo/abort the residue calculation...
drivers/dma/pl330.c | 24 ++++++++++++++++++++----
1 file changed, 20 insertions(+), 4 deletions(-)
diff --git a/drivers/dma/pl330.c b/drivers/dma/pl330.c
index 17ee758..55e3c5f 100644
--- a/drivers/dma/pl330.c
+++ b/drivers/dma/pl330.c
@@ -2240,6 +2240,7 @@ pl330_tx_status(struct dma_chan *chan, dma_cookie_t cookie,
struct dma_pl330_desc *desc, *running = NULL;
struct dma_pl330_chan *pch = to_pchan(chan);
unsigned int transferred, residual = 0;
+ bool first_busy;
ret = dma_cookie_status(chan, cookie, txstate);
@@ -2253,16 +2254,31 @@ pl330_tx_status(struct dma_chan *chan, dma_cookie_t cookie,
if (pch->thread->req_running != -1)
running = pch->thread->req[pch->thread->req_running].desc;
+ first_busy = true;
/* Check in pending list */
list_for_each_entry(desc, &pch->work_list, node) {
if (desc->status == DONE)
transferred = desc->bytes_requested;
- else if (running && desc == running)
- transferred =
- pl330_get_current_xferred_count(pch, desc);
- else
+ else if (desc->status == BUSY && first_busy) {
+ first_busy = false;
+ if (running && desc == running) {
+ transferred =
+ pl330_get_current_xferred_count(pch, desc);
+ } else {
+ /* BUSY but not running means it's just completed */
+ transferred = desc->bytes_requested;
+ }
+ } else {
+ /*
+ * Descriptor is either in PREP state queued for future
+ * transfer or it is the second BUSY descriptor we have
+ * seen. The latter case means it has just, or is about
+ * to be, started, so treat it as having not yet
+ * transferred any bytes, the same as PREP.
+ */
transferred = 0;
+ }
residual += desc->bytes_requested - transferred;
if (desc->txd.cookie == cookie) {
switch (desc->status) {
--
2.1.4
--
To unsubscribe from this list: send the line "unsubscribe dmaengine" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
