From: Aaro Koskinen <[email protected]>
Commit 0e80bdc9a72d ("stmmac: first frame prep at the end of xmit
routine") overlooked jumbo frames when re-ordering the code, and as a
result the own bit was not getting set anymore for the first jumbo frame
descriptor. Commit 487e2e22ab79 ("net: stmmac: Set OWN bit for jumbo
frames") tried to fix this, but now the bit is getting set too early and
the DMA may start while we are still setting up the remaing descriptors.
And with the chain mode the own bit remains still unset.
Fix by setting the own bit at the end of xmit also with jumbo frames.
Fixes: 0e80bdc9a72d ("stmmac: first frame prep at the end of xmit routine")
Fixes: 487e2e22ab79 ("net: stmmac: Set OWN bit for jumbo frames")
Signed-off-by: Aaro Koskinen <[email protected]>
---
drivers/net/ethernet/stmicro/stmmac/ring_mode.c | 4 ++--
drivers/net/ethernet/stmicro/stmmac/stmmac_main.c | 14 ++++++++------
2 files changed, 10 insertions(+), 8 deletions(-)
diff --git a/drivers/net/ethernet/stmicro/stmmac/ring_mode.c b/drivers/net/ethernet/stmicro/stmmac/ring_mode.c
index d8c5bc412219..bc83ced94e1b 100644
--- a/drivers/net/ethernet/stmicro/stmmac/ring_mode.c
+++ b/drivers/net/ethernet/stmicro/stmmac/ring_mode.c
@@ -59,7 +59,7 @@ static int jumbo_frm(void *p, struct sk_buff *skb, int csum)
desc->des3 = cpu_to_le32(des2 + BUF_SIZE_4KiB);
stmmac_prepare_tx_desc(priv, desc, 1, bmax, csum,
- STMMAC_RING_MODE, 1, false, skb->len);
+ STMMAC_RING_MODE, 0, false, skb->len);
tx_q->tx_skbuff[entry] = NULL;
entry = STMMAC_GET_ENTRY(entry, DMA_TX_SIZE);
@@ -91,7 +91,7 @@ static int jumbo_frm(void *p, struct sk_buff *skb, int csum)
tx_q->tx_skbuff_dma[entry].is_jumbo = true;
desc->des3 = cpu_to_le32(des2 + BUF_SIZE_4KiB);
stmmac_prepare_tx_desc(priv, desc, 1, nopaged_len, csum,
- STMMAC_RING_MODE, 1, true, skb->len);
+ STMMAC_RING_MODE, 0, true, skb->len);
}
tx_q->cur_tx = entry;
diff --git a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
index 97c5e1aad88f..6a2e1031a62a 100644
--- a/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
+++ b/drivers/net/ethernet/stmicro/stmmac/stmmac_main.c
@@ -3216,14 +3216,16 @@ static netdev_tx_t stmmac_xmit(struct sk_buff *skb, struct net_device *dev)
stmmac_prepare_tx_desc(priv, first, 1, nopaged_len,
csum_insertion, priv->mode, 1, last_segment,
skb->len);
-
- /* The own bit must be the latest setting done when prepare the
- * descriptor and then barrier is needed to make sure that
- * all is coherent before granting the DMA engine.
- */
- wmb();
+ } else {
+ stmmac_set_tx_owner(priv, first);
}
+ /* The own bit must be the latest setting done when prepare the
+ * descriptor and then barrier is needed to make sure that
+ * all is coherent before granting the DMA engine.
+ */
+ wmb();
+
netdev_tx_sent_queue(netdev_get_tx_queue(dev, queue), skb->len);
stmmac_enable_dma_transmission(priv, priv->ioaddr);
--
2.17.0
From: Aaro Koskinen <[email protected]>
When sending non-linear skbs with jumbo frames, we set up the non-paged
data and mark that as a last segment, although the paged fragments are
also prepared. This will stall the TX queue and trigger a watchdog warning
(a simple reproducer is to run an iperf client mode TCP test with a large
MTU - networking fails instantly).
Fix by checking if the skb is non-linear.
Signed-off-by: Aaro Koskinen <[email protected]>
---
drivers/net/ethernet/stmicro/stmmac/ring_mode.c | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/drivers/net/ethernet/stmicro/stmmac/ring_mode.c b/drivers/net/ethernet/stmicro/stmmac/ring_mode.c
index bc83ced94e1b..f936166d8910 100644
--- a/drivers/net/ethernet/stmicro/stmmac/ring_mode.c
+++ b/drivers/net/ethernet/stmicro/stmmac/ring_mode.c
@@ -79,7 +79,8 @@ static int jumbo_frm(void *p, struct sk_buff *skb, int csum)
desc->des3 = cpu_to_le32(des2 + BUF_SIZE_4KiB);
stmmac_prepare_tx_desc(priv, desc, 0, len, csum,
- STMMAC_RING_MODE, 1, true, skb->len);
+ STMMAC_RING_MODE, 1, !skb_is_nonlinear(skb),
+ skb->len);
} else {
des2 = dma_map_single(priv->device, skb->data,
nopaged_len, DMA_TO_DEVICE);
@@ -91,7 +92,8 @@ static int jumbo_frm(void *p, struct sk_buff *skb, int csum)
tx_q->tx_skbuff_dma[entry].is_jumbo = true;
desc->des3 = cpu_to_le32(des2 + BUF_SIZE_4KiB);
stmmac_prepare_tx_desc(priv, desc, 1, nopaged_len, csum,
- STMMAC_RING_MODE, 0, true, skb->len);
+ STMMAC_RING_MODE, 0, !skb_is_nonlinear(skb),
+ skb->len);
}
tx_q->cur_tx = entry;
--
2.17.0
Hello!
On 14.03.2019 22:43, Aaro Koskinen wrote:
> From: Aaro Koskinen <[email protected]>
>
> Commit 0e80bdc9a72d ("stmmac: first frame prep at the end of xmit
> routine") overlooked jumbo frames when re-ordering the code, and as a
> result the own bit was not getting set anymore for the first jumbo frame
> descriptor. Commit 487e2e22ab79 ("net: stmmac: Set OWN bit for jumbo
> frames") tried to fix this, but now the bit is getting set too early and
> the DMA may start while we are still setting up the remaing descriptors.
Remaining?
> And with the chain mode the own bit remains still unset.
>
> Fix by setting the own bit at the end of xmit also with jumbo frames.
>
> Fixes: 0e80bdc9a72d ("stmmac: first frame prep at the end of xmit routine")
> Fixes: 487e2e22ab79 ("net: stmmac: Set OWN bit for jumbo frames")
> Signed-off-by: Aaro Koskinen <[email protected]>
[...]
MBR, Sergei
On 3/14/2019 7:43 PM, Aaro Koskinen wrote:
> From: Aaro Koskinen <[email protected]>
>
> Commit 0e80bdc9a72d ("stmmac: first frame prep at the end of xmit
> routine") overlooked jumbo frames when re-ordering the code, and as a
> result the own bit was not getting set anymore for the first jumbo frame
> descriptor. Commit 487e2e22ab79 ("net: stmmac: Set OWN bit for jumbo
> frames") tried to fix this, but now the bit is getting set too early and
> the DMA may start while we are still setting up the remaing descriptors.
> And with the chain mode the own bit remains still unset.
>
> Fix by setting the own bit at the end of xmit also with jumbo frames.
>
> Fixes: 0e80bdc9a72d ("stmmac: first frame prep at the end of xmit routine")
> Fixes: 487e2e22ab79 ("net: stmmac: Set OWN bit for jumbo frames")
> Signed-off-by: Aaro Koskinen <[email protected]>
With Sergei comment fix:
Acked-by: Jose Abreu <[email protected]>
Thanks,
Jose Miguel Abreu
On 3/14/2019 7:43 PM, Aaro Koskinen wrote:
> From: Aaro Koskinen <[email protected]>
>
> When sending non-linear skbs with jumbo frames, we set up the non-paged
> data and mark that as a last segment, although the paged fragments are
> also prepared. This will stall the TX queue and trigger a watchdog warning
> (a simple reproducer is to run an iperf client mode TCP test with a large
> MTU - networking fails instantly).
>
> Fix by checking if the skb is non-linear.
>
> Signed-off-by: Aaro Koskinen <[email protected]>
Acked-by: Jose Abreu <[email protected]>
Thanks,
Jose Miguel Abreu
From: Aaro Koskinen <[email protected]>
Date: Thu, 14 Mar 2019 21:43:19 +0200
> From: Aaro Koskinen <[email protected]>
>
> Commit 0e80bdc9a72d ("stmmac: first frame prep at the end of xmit
> routine") overlooked jumbo frames when re-ordering the code, and as a
> result the own bit was not getting set anymore for the first jumbo frame
> descriptor. Commit 487e2e22ab79 ("net: stmmac: Set OWN bit for jumbo
> frames") tried to fix this, but now the bit is getting set too early and
> the DMA may start while we are still setting up the remaing descriptors.
> And with the chain mode the own bit remains still unset.
>
> Fix by setting the own bit at the end of xmit also with jumbo frames.
>
> Fixes: 0e80bdc9a72d ("stmmac: first frame prep at the end of xmit routine")
> Fixes: 487e2e22ab79 ("net: stmmac: Set OWN bit for jumbo frames")
> Signed-off-by: Aaro Koskinen <[email protected]>
Applied with the "remaining" typo fixed.
From: Aaro Koskinen <[email protected]>
Date: Thu, 14 Mar 2019 21:43:20 +0200
> From: Aaro Koskinen <[email protected]>
>
> When sending non-linear skbs with jumbo frames, we set up the non-paged
> data and mark that as a last segment, although the paged fragments are
> also prepared. This will stall the TX queue and trigger a watchdog warning
> (a simple reproducer is to run an iperf client mode TCP test with a large
> MTU - networking fails instantly).
>
> Fix by checking if the skb is non-linear.
>
> Signed-off-by: Aaro Koskinen <[email protected]>
Applied.