The mvneta controller provides a 8-bit register to update the pending
Tx descriptor counter. Then, a maximum of 255 Tx descriptors can be
added at once. In the current code the mvneta_txq_pend_desc_add function
assumes the caller takes care of this limit. But it is not the case. In
some situations (xmit_more flag), more than 255 descriptors are added.
When this happens, the Tx descriptor counter register is updated with a
wrong value, which breaks the whole Tx queue management.
This patch fixes the issue by allowing the mvneta_txq_pend_desc_add
function to process more than 255 Tx descriptors.
Fixes: 2a90f7e1d5d0 ("net: mvneta: add xmit_more support")
Cc: stable@xxxxxxxxxxxxxxx # 4.11+
Signed-off-by: Simon Guinot <simon.guinot@xxxxxxxxxxxx>
---
drivers/net/ethernet/marvell/mvneta.c | 13 ++++++++-----
1 file changed, 8 insertions(+), 5 deletions(-)
Changes for v2:
- Use do {} while instead of while {}.
- Keep "txq->pending + frags > MVNETA_TXQ_DEC_SENT_MASK" as a flushing
condition for the Tx queue.
diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c
index 64a04975bcf8..bc93b69cfd1e 100644
--- a/drivers/net/ethernet/marvell/mvneta.c
+++ b/drivers/net/ethernet/marvell/mvneta.c
@@ -816,11 +816,14 @@ static void mvneta_txq_pend_desc_add(struct mvneta_port *pp,
{
u32 val;
- /* Only 255 descriptors can be added at once ; Assume caller
- * process TX desriptors in quanta less than 256
- */
- val = pend_desc + txq->pending;
- mvreg_write(pp, MVNETA_TXQ_UPDATE_REG(txq->id), val);
+ pend_desc += txq->pending;
+
+ /* Only 255 Tx descriptors can be added at once */
+ do {
+ val = min(pend_desc, 255);
+ mvreg_write(pp, MVNETA_TXQ_UPDATE_REG(txq->id), val);
+ pend_desc -= val;
+ } while (pend_desc > 0);
txq->pending = 0;
}
--
2.15.0