diff mbox series

[net,1/3] net: hns3: fix for TX queue not restarted problem

Message ID 1575342535-2981-2-git-send-email-tanhuazhong@huawei.com
State Changes Requested
Delegated to: David Miller
Headers show
Series net: hns3: fixes for -net | expand

Commit Message

tanhuazhong Dec. 3, 2019, 3:08 a.m. UTC 
From: Yunsheng Lin <linyunsheng@huawei.com>

There is timing window between ring_space checking and
netif_stop_subqueue when transmiting a SKB, and the TX BD
cleaning may be executed during the time window, which may
caused TX queue not restarted problem.

This patch fixes it by rechecking the ring_space after
netif_stop_subqueue to make sure TX queue is restarted.

Also, the ring->next_to_clean is updated even when pkts is
zero, because all the TX BD cleaned may be non-SKB, so it
needs to check if TX queue need to be restarted.

Fixes: 76ad4f0ee747 ("net: hns3: Add support of HNS3 Ethernet Driver for hip08 SoC")
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
---
 drivers/net/ethernet/hisilicon/hns3/hns3_enet.c | 36 ++++++++++++++++---------
 1 file changed, 23 insertions(+), 13 deletions(-)

Comments

David Miller Dec. 3, 2019, 3:25 a.m. UTC | #1 
From: Huazhong Tan <tanhuazhong@huawei.com>
Date: Tue, 3 Dec 2019 11:08:53 +0800

> diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
> index ba05368..b2bb8e2 100644
> --- a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
> +++ b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
> @@ -1286,13 +1286,16 @@ static bool hns3_skb_need_linearized(struct sk_buff *skb, unsigned int *bd_size,
>  	return false;
>  }
>  
> -static int hns3_nic_maybe_stop_tx(struct hns3_enet_ring *ring,
> +static int hns3_nic_maybe_stop_tx(struct net_device *netdev,
>  				  struct sk_buff **out_skb)
>  {
> +	struct hns3_nic_priv *priv = netdev_priv(netdev);
>  	unsigned int bd_size[HNS3_MAX_TSO_BD_NUM + 1U];
>  	struct sk_buff *skb = *out_skb;
> +	struct hns3_enet_ring *ring;
>  	unsigned int bd_num;
>  
> +	ring = &priv->ring[skb->queue_mapping];

Please just pass the ring pointer into hns3_nic_maybe_stop_tx() instead of
needlessly recalculating it.

Thank you.
Yunsheng Lin Dec. 3, 2019, 4:28 a.m. UTC | #2 
On 2019/12/3 11:25, David Miller wrote:
> From: Huazhong Tan <tanhuazhong@huawei.com>
> Date: Tue, 3 Dec 2019 11:08:53 +0800
> 
>> diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
>> index ba05368..b2bb8e2 100644
>> --- a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
>> +++ b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
>> @@ -1286,13 +1286,16 @@ static bool hns3_skb_need_linearized(struct sk_buff *skb, unsigned int *bd_size,
>>  	return false;
>>  }
>>  
>> -static int hns3_nic_maybe_stop_tx(struct hns3_enet_ring *ring,
>> +static int hns3_nic_maybe_stop_tx(struct net_device *netdev,
>>  				  struct sk_buff **out_skb)
>>  {
>> +	struct hns3_nic_priv *priv = netdev_priv(netdev);
>>  	unsigned int bd_size[HNS3_MAX_TSO_BD_NUM + 1U];
>>  	struct sk_buff *skb = *out_skb;
>> +	struct hns3_enet_ring *ring;
>>  	unsigned int bd_num;
>>  
>> +	ring = &priv->ring[skb->queue_mapping];
> 
> Please just pass the ring pointer into hns3_nic_maybe_stop_tx() instead of
> needlessly recalculating it.

The reason that I am passing the netdev instead of ring pointer is
that the netif_start_subqueue() need a netdev parameter, and the
netdev can not be derived from the ring pointer.

Do you think it is better to keep it as this patch, or add a new
netdevice parameter? like below:

static int hns3_nic_maybe_stop_tx(struct net_device *netdev,
				  struct hns3_enet_ring *ring,
				  struct sk_buff **out_skb)



> 
> Thank you.
> 
> .
>
David Miller Dec. 3, 2019, 8:48 a.m. UTC | #3 
From: Yunsheng Lin <linyunsheng@huawei.com>
Date: Tue, 3 Dec 2019 12:28:22 +0800

> On 2019/12/3 11:25, David Miller wrote:
>> From: Huazhong Tan <tanhuazhong@huawei.com>
>> Date: Tue, 3 Dec 2019 11:08:53 +0800
>> 
>>> diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
>>> index ba05368..b2bb8e2 100644
>>> --- a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
>>> +++ b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
>>> @@ -1286,13 +1286,16 @@ static bool hns3_skb_need_linearized(struct sk_buff *skb, unsigned int *bd_size,
>>>  	return false;
>>>  }
>>>  
>>> -static int hns3_nic_maybe_stop_tx(struct hns3_enet_ring *ring,
>>> +static int hns3_nic_maybe_stop_tx(struct net_device *netdev,
>>>  				  struct sk_buff **out_skb)
>>>  {
>>> +	struct hns3_nic_priv *priv = netdev_priv(netdev);
>>>  	unsigned int bd_size[HNS3_MAX_TSO_BD_NUM + 1U];
>>>  	struct sk_buff *skb = *out_skb;
>>> +	struct hns3_enet_ring *ring;
>>>  	unsigned int bd_num;
>>>  
>>> +	ring = &priv->ring[skb->queue_mapping];
>> 
>> Please just pass the ring pointer into hns3_nic_maybe_stop_tx() instead of
>> needlessly recalculating it.
> 
> The reason that I am passing the netdev instead of ring pointer is
> that the netif_start_subqueue() need a netdev parameter, and the
> netdev can not be derived from the ring pointer.
> 
> Do you think it is better to keep it as this patch, or add a new
> netdevice parameter? like below:

Just add the netdev parameter, in addition to the ring parameter.

All arguments fit in the register argument passing conventions of
various cpus so the cost of adding the parameter is zero.
Yunsheng Lin Dec. 4, 2019, 1:32 a.m. UTC | #4 
On 2019/12/3 16:48, David Miller wrote:
> From: Yunsheng Lin <linyunsheng@huawei.com>
> Date: Tue, 3 Dec 2019 12:28:22 +0800
> 
>> On 2019/12/3 11:25, David Miller wrote:
>>> From: Huazhong Tan <tanhuazhong@huawei.com>
>>> Date: Tue, 3 Dec 2019 11:08:53 +0800
>>>
>>>> diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
>>>> index ba05368..b2bb8e2 100644
>>>> --- a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
>>>> +++ b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
>>>> @@ -1286,13 +1286,16 @@ static bool hns3_skb_need_linearized(struct sk_buff *skb, unsigned int *bd_size,
>>>>  	return false;
>>>>  }
>>>>  
>>>> -static int hns3_nic_maybe_stop_tx(struct hns3_enet_ring *ring,
>>>> +static int hns3_nic_maybe_stop_tx(struct net_device *netdev,
>>>>  				  struct sk_buff **out_skb)
>>>>  {
>>>> +	struct hns3_nic_priv *priv = netdev_priv(netdev);
>>>>  	unsigned int bd_size[HNS3_MAX_TSO_BD_NUM + 1U];
>>>>  	struct sk_buff *skb = *out_skb;
>>>> +	struct hns3_enet_ring *ring;
>>>>  	unsigned int bd_num;
>>>>  
>>>> +	ring = &priv->ring[skb->queue_mapping];
>>>
>>> Please just pass the ring pointer into hns3_nic_maybe_stop_tx() instead of
>>> needlessly recalculating it.
>>
>> The reason that I am passing the netdev instead of ring pointer is
>> that the netif_start_subqueue() need a netdev parameter, and the
>> netdev can not be derived from the ring pointer.
>>
>> Do you think it is better to keep it as this patch, or add a new
>> netdevice parameter? like below:
> 
> Just add the netdev parameter, in addition to the ring parameter.
> 
> All arguments fit in the register argument passing conventions of
> various cpus so the cost of adding the parameter is zero.

Ok, thanks.

> 
> .
>
diff mbox series

Patch

diff --git a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
index ba05368..b2bb8e2 100644
--- a/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
+++ b/drivers/net/ethernet/hisilicon/hns3/hns3_enet.c
@@ -1286,13 +1286,16 @@  static bool hns3_skb_need_linearized(struct sk_buff *skb, unsigned int *bd_size,
 	return false;
 }
 
-static int hns3_nic_maybe_stop_tx(struct hns3_enet_ring *ring,
+static int hns3_nic_maybe_stop_tx(struct net_device *netdev,
 				  struct sk_buff **out_skb)
 {
+	struct hns3_nic_priv *priv = netdev_priv(netdev);
 	unsigned int bd_size[HNS3_MAX_TSO_BD_NUM + 1U];
 	struct sk_buff *skb = *out_skb;
+	struct hns3_enet_ring *ring;
 	unsigned int bd_num;
 
+	ring = &priv->ring[skb->queue_mapping];
 	bd_num = hns3_tx_bd_num(skb, bd_size);
 	if (unlikely(bd_num > HNS3_MAX_NON_TSO_BD_NUM)) {
 		struct sk_buff *new_skb;
@@ -1320,10 +1323,23 @@  static int hns3_nic_maybe_stop_tx(struct hns3_enet_ring *ring,
 	}
 
 out:
-	if (unlikely(ring_space(ring) < bd_num))
-		return -EBUSY;
+	if (likely(ring_space(ring) >= bd_num))
+		return bd_num;
 
-	return bd_num;
+	netif_stop_subqueue(netdev, ring->queue_index);
+	smp_mb(); /* Memory barrier before checking ring_space */
+
+	/* Start queue in case hns3_clean_tx_ring has just made room
+	 * available and has not seen the queue stopped state performed
+	 * by netif_stop_subqueue above.
+	 */
+	if (ring_space(ring) >= bd_num && netif_carrier_ok(netdev) &&
+	    !test_bit(HNS3_NIC_STATE_DOWN, &priv->state)) {
+		netif_start_subqueue(netdev, ring->queue_index);
+		return bd_num;
+	}
+
+	return -EBUSY;
 }
 
 static void hns3_clear_desc(struct hns3_enet_ring *ring, int next_to_use_orig)
@@ -1400,13 +1416,13 @@  netdev_tx_t hns3_nic_net_xmit(struct sk_buff *skb, struct net_device *netdev)
 	/* Prefetch the data used later */
 	prefetch(skb->data);
 
-	ret = hns3_nic_maybe_stop_tx(ring, &skb);
+	ret = hns3_nic_maybe_stop_tx(netdev, &skb);
 	if (unlikely(ret <= 0)) {
 		if (ret == -EBUSY) {
 			u64_stats_update_begin(&ring->syncp);
 			ring->stats.tx_busy++;
 			u64_stats_update_end(&ring->syncp);
-			goto out_net_tx_busy;
+			return NETDEV_TX_BUSY;
 		} else if (ret == -ENOMEM) {
 			u64_stats_update_begin(&ring->syncp);
 			ring->stats.sw_err_cnt++;
@@ -1457,12 +1473,6 @@  netdev_tx_t hns3_nic_net_xmit(struct sk_buff *skb, struct net_device *netdev)
 out_err_tx_ok:
 	dev_kfree_skb_any(skb);
 	return NETDEV_TX_OK;
-
-out_net_tx_busy:
-	netif_stop_subqueue(netdev, ring->queue_index);
-	smp_mb(); /* Commit all data before submit */
-
-	return NETDEV_TX_BUSY;
 }
 
 static int hns3_nic_net_set_mac_address(struct net_device *netdev, void *p)
@@ -2519,7 +2529,7 @@  void hns3_clean_tx_ring(struct hns3_enet_ring *ring)
 	dev_queue = netdev_get_tx_queue(netdev, ring->tqp->tqp_index);
 	netdev_tx_completed_queue(dev_queue, pkts, bytes);
 
-	if (unlikely(pkts && netif_carrier_ok(netdev) &&
+	if (unlikely(netif_carrier_ok(netdev) &&
 		     ring_space(ring) > HNS3_MAX_TSO_BD_NUM)) {
 		/* Make sure that anybody stopping the queue after this
 		 * sees the new next_to_clean.