[PATCH net-next 0/3] Introduce and use netif_xmit_timeout_ms() helper

[PATCH net-next 0/3] Introduce and use netif_xmit_timeout_ms() helper

[Tariq Toukan]

Hi,

This series by Shahar introduces a new helper function
netif_xmit_timeout_ms() to check if a TX queue has timed out and report
the timeout duration.

Replace open-coded timeout checks in dev_watchdog() and the hns3 driver
with the new helper.

For mlx5e, refine the TX timeout recovery flow to act only on SQs whose
transmit timestamp indicates an actual timeout, as determined by the
helper. This prevents unnecessary channel reopen events caused by
attempting recovery on queues that are merely stopped but not truly
timed out.

Regards,
Tariq

Shahar Shitrit (3):
  net: Introduce netif_xmit_time_out_duration() helper
  net: hns3: Use netif_xmit_timeout_ms() helper
  net/mlx5e: Refine TX timeout handling to skip non-timed-out SQ

 drivers/net/ethernet/hisilicon/hns3/hns3_enet.c   |  5 +----
 drivers/net/ethernet/mellanox/mlx5/core/en_main.c |  2 +-
 include/linux/netdevice.h                         | 15 +++++++++++++++
 net/sched/sch_generic.c                           |  7 +++----
 4 files changed, 20 insertions(+), 9 deletions(-)


base-commit: e05021a829b834fecbd42b173e55382416571b2c
-- 
2.31.1

[PATCH net-next 1/3] net: Introduce netif_xmit_time_out_duration() helper

[Tariq Toukan]

From: Shahar Shitrit <shshitrit@nvidia.com>

Introduce a new helper function netif_xmit_time_out_duration() to
check if a TX queue has timed out and report the timeout duration.
This helper consolidates the logic that is duplicated in several
locations and also encapsulates the check for whether the TX queue
is stopped.

As the first user, convert dev_watchdog() to use this helper.

Signed-off-by: Shahar Shitrit <shshitrit@nvidia.com>
Reviewed-by: Yael Chemla <ychemla@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
---
 include/linux/netdevice.h | 15 +++++++++++++++
 net/sched/sch_generic.c   |  7 +++----
 2 files changed, 18 insertions(+), 4 deletions(-)

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index e808071dbb7d..3cd73769fcfa 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -3680,6 +3680,21 @@ static inline bool netif_xmit_stopped(const struct netdev_queue *dev_queue)
 	return dev_queue->state & QUEUE_STATE_ANY_XOFF;
 }
 
+static inline unsigned int
+netif_xmit_timeout_ms(struct netdev_queue *txq, unsigned long *trans_start)
+{
+	unsigned long txq_trans_start = READ_ONCE(txq->trans_start);
+
+	if (trans_start)
+		*trans_start = txq_trans_start;
+
+	if (netif_xmit_stopped(txq) &&
+	    time_after(jiffies, txq_trans_start + txq->dev->watchdog_timeo))
+		return jiffies_to_msecs(jiffies - txq_trans_start);
+
+	return 0;
+}
+
 static inline bool
 netif_xmit_frozen_or_stopped(const struct netdev_queue *dev_queue)
 {
diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c
index 852e603c1755..aa6192781a24 100644
--- a/net/sched/sch_generic.c
+++ b/net/sched/sch_generic.c
@@ -523,10 +523,9 @@ static void dev_watchdog(struct timer_list *t)
 				 * netdev_tx_sent_queue() and netif_tx_stop_queue().
 				 */
 				smp_mb();
-				trans_start = READ_ONCE(txq->trans_start);
-
-				if (time_after(jiffies, trans_start + dev->watchdog_timeo)) {
-					timedout_ms = jiffies_to_msecs(jiffies - trans_start);
+				timedout_ms = netif_xmit_timeout_ms(txq,
+								    &trans_start);
+				if (timedout_ms) {
 					atomic_long_inc(&txq->trans_timeout);
 					break;
 				}
-- 
2.31.1

Re: [PATCH net-next 1/3] net: Introduce netif_xmit_time_out_duration() helper

[Paolo Abeni]

On 11/25/25 8:12 AM, Tariq Toukan wrote:
> From: Shahar Shitrit <shshitrit@nvidia.com>
> 
> Introduce a new helper function netif_xmit_time_out_duration() to
> check if a TX queue has timed out and report the timeout duration.
> This helper consolidates the logic that is duplicated in several
> locations and also encapsulates the check for whether the TX queue
> is stopped.
> 
> As the first user, convert dev_watchdog() to use this helper.
> 
> Signed-off-by: Shahar Shitrit <shshitrit@nvidia.com>
> Reviewed-by: Yael Chemla <ychemla@nvidia.com>
> Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
> ---
>  include/linux/netdevice.h | 15 +++++++++++++++
>  net/sched/sch_generic.c   |  7 +++----
>  2 files changed, 18 insertions(+), 4 deletions(-)
> 
> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
> index e808071dbb7d..3cd73769fcfa 100644
> --- a/include/linux/netdevice.h
> +++ b/include/linux/netdevice.h
> @@ -3680,6 +3680,21 @@ static inline bool netif_xmit_stopped(const struct netdev_queue *dev_queue)
>  	return dev_queue->state & QUEUE_STATE_ANY_XOFF;
>  }
>  
> +static inline unsigned int
> +netif_xmit_timeout_ms(struct netdev_queue *txq, unsigned long *trans_start)
> +{
> +	unsigned long txq_trans_start = READ_ONCE(txq->trans_start);
> +
> +	if (trans_start)
> +		*trans_start = txq_trans_start;

What about making this argument mandatory?

> +
> +	if (netif_xmit_stopped(txq) &&

Why restricting to the <queue stopped> case? AFAICS the watchdog is
intended to additionally catch the scenarios where the rx ring is not
full but the H/W is stuck for whatever reasons, and this change will not
catch them anymore.

/P

Re: [PATCH net-next 1/3] net: Introduce netif_xmit_time_out_duration() helper

[Shahar Shitrit]

On 27/11/2025 12:53, Paolo Abeni wrote:
> On 11/25/25 8:12 AM, Tariq Toukan wrote:
>> From: Shahar Shitrit <shshitrit@nvidia.com>
>>
>> Introduce a new helper function netif_xmit_time_out_duration() to
>> check if a TX queue has timed out and report the timeout duration.
>> This helper consolidates the logic that is duplicated in several
>> locations and also encapsulates the check for whether the TX queue
>> is stopped.
>>
>> As the first user, convert dev_watchdog() to use this helper.
>>
>> Signed-off-by: Shahar Shitrit <shshitrit@nvidia.com>
>> Reviewed-by: Yael Chemla <ychemla@nvidia.com>
>> Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
>> ---
>>  include/linux/netdevice.h | 15 +++++++++++++++
>>  net/sched/sch_generic.c   |  7 +++----
>>  2 files changed, 18 insertions(+), 4 deletions(-)
>>
>> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
>> index e808071dbb7d..3cd73769fcfa 100644
>> --- a/include/linux/netdevice.h
>> +++ b/include/linux/netdevice.h
>> @@ -3680,6 +3680,21 @@ static inline bool netif_xmit_stopped(const struct netdev_queue *dev_queue)
>>  	return dev_queue->state & QUEUE_STATE_ANY_XOFF;
>>  }
>>  
>> +static inline unsigned int
>> +netif_xmit_timeout_ms(struct netdev_queue *txq, unsigned long *trans_start)
>> +{
>> +	unsigned long txq_trans_start = READ_ONCE(txq->trans_start);
>> +
>> +	if (trans_start)
>> +		*trans_start = txq_trans_start;
> 
> What about making this argument mandatory?

Since not all callers are interested in this return value, as in the
case of mlx5, it would be nice to allow them pass NULL.>
>> +
>> +	if (netif_xmit_stopped(txq) &&
> 
> Why restricting to the <queue stopped> case? AFAICS the watchdog is
> intended to additionally catch the scenarios where the rx ring is not
> full but the H/W is stuck for whatever reasons, and this change will not
> catch them anymore.
> 
> /P
>

dev_watchdog catches only the cases where tx queue is both full (queue
stopped) and timed out (last transmit timestamp was longer ago than the
watchdog timeout period). We wanted to preserve the same conditions.

By the way, I notice now that the new helper name in the title doesn't
match the one in the code. We will fix.

Re: [PATCH net-next 1/3] net: Introduce netif_xmit_time_out_duration() helper

[Jakub Kicinski]

On Tue, 25 Nov 2025 09:12:54 +0200 Tariq Toukan wrote:
> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
> index e808071dbb7d..3cd73769fcfa 100644
> --- a/include/linux/netdevice.h
> +++ b/include/linux/netdevice.h

include/net/net_queue.h seems like a better place for new code

> @@ -3680,6 +3680,21 @@ static inline bool netif_xmit_stopped(const struct netdev_queue *dev_queue)
>  	return dev_queue->state & QUEUE_STATE_ANY_XOFF;
>  }
>  
> +static inline unsigned int
> +netif_xmit_timeout_ms(struct netdev_queue *txq, unsigned long *trans_start)
> +{
> +	unsigned long txq_trans_start = READ_ONCE(txq->trans_start);
> +
> +	if (trans_start)
> +		*trans_start = txq_trans_start;

The drivers don't really care about this, AFAICT hns3 uses this
to calculate the stall length (return value of this func.

> +	if (netif_xmit_stopped(txq) &&
> +	    time_after(jiffies, txq_trans_start + txq->dev->watchdog_timeo))
> +		return jiffies_to_msecs(jiffies - txq_trans_start);
> +
> +	return 0;
> +}
> +
>  static inline bool
>  netif_xmit_frozen_or_stopped(const struct netdev_queue *dev_queue)
>  {
> diff --git a/net/sched/sch_generic.c b/net/sched/sch_generic.c
> index 852e603c1755..aa6192781a24 100644
> --- a/net/sched/sch_generic.c
> +++ b/net/sched/sch_generic.c
> @@ -523,10 +523,9 @@ static void dev_watchdog(struct timer_list *t)
>  				 * netdev_tx_sent_queue() and netif_tx_stop_queue().
>  				 */
>  				smp_mb();
> -				trans_start = READ_ONCE(txq->trans_start);
> -
> -				if (time_after(jiffies, trans_start + dev->watchdog_timeo)) {
> -					timedout_ms = jiffies_to_msecs(jiffies - trans_start);
> +				timedout_ms = netif_xmit_timeout_ms(txq,
> +								    &trans_start);
> +				if (timedout_ms) {

The use of the new helper in the core feels a bit forced, I'd leave 
the core as is. Otherwise you need the awkward output param, and
core now duplicates the netif_xmit_stopped(txq) check

>  					atomic_long_inc(&txq->trans_timeout);
>  					break;
>  				}

[PATCH net-next 2/3] net: hns3: Use netif_xmit_timeout_ms() helper

[Tariq Toukan]

From: Shahar Shitrit <shshitrit@nvidia.com>

Replace the open-coded TX queue timeout check in
hns3_get_timeout_queue() with a call to netif_xmit_timeout_ms() helper.

Signed-off-by: Shahar Shitrit <shshitrit@nvidia.com>
Reviewed-by: Yael Chemla <ychemla@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
---
 drivers/net/ethernet/hisilicon/hns3/hns3_enet.c | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)

diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
index 7a0654e2d3dd..3e8fe3b5d32b 100644
--- a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
+++ b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
@@ -2811,10 +2811,7 @@ static int hns3_get_timeout_queue(struct net_device *ndev)
 		unsigned long trans_start;
 
 		q = netdev_get_tx_queue(ndev, i);
-		trans_start = READ_ONCE(q->trans_start);
-		if (netif_xmit_stopped(q) &&
-		    time_after(jiffies,
-			       (trans_start + ndev->watchdog_timeo))) {
+		if (netif_xmit_timeout_ms(q, &trans_start)) {
 #ifdef CONFIG_BQL
 			struct dql *dql = &q->dql;
 
-- 
2.31.1

[PATCH net-next 3/3] net/mlx5e: Refine TX timeout handling to skip non-timed-out SQ

[Tariq Toukan]

From: Shahar Shitrit <shshitrit@nvidia.com>

mlx5e_tx_timeout_work() is invoked when the dev_watchdog reports a
timed-out TX queue. Currently, the recovery flow is triggered for all
stopped SQs, which is not always correct — some SQs may be temporarily
stopped without actually timing out. Attempting to recover such SQs
results in no EQE being polled (since no real timeout occurred), which
the driver misinterprets as a recovery failure, unnecessarily causing
channel reopening.

Improve the logic to initiate recovery only for SQs that are both
stopped and timed out. Utilize the helper introduced in the previous
patch to determine whether the netdevice watchdog timeout period has
elapsed since the SQ’s last transmit timestamp.

Signed-off-by: Shahar Shitrit <shshitrit@nvidia.com>
Reviewed-by: Yael Chemla <ychemla@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
---
 drivers/net/ethernet/mellanox/mlx5/core/en_main.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
index e537df670758..cd146df29ada 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
@@ -5139,7 +5139,7 @@ static void mlx5e_tx_timeout_work(struct work_struct *work)
 			netdev_get_tx_queue(netdev, i);
 		struct mlx5e_txqsq *sq = priv->txq2sq[i];
 
-		if (!netif_xmit_stopped(dev_queue))
+		if (!netif_xmit_timeout_ms(dev_queue, NULL))
 			continue;
 
 		if (mlx5e_reporter_tx_timeout(sq))
-- 
2.31.1