net-sched: When TX over-limit happens on netdev, what does kernel do?

net-sched: When TX over-limit happens on netdev, what does kernel do?

[Liu ping fan]

Hi,

Could anybody tell me when TX over-limit happens on netdev, what does kernel do?
Is skb buffered in Qdisc->q and the sending process keeps running?
Or we block the sending process and make it scheduled?

Found some code in net/sched/sch_cbq.c
cbq_dequeue(struct Qdisc *sch) --> qdisc_watchdog_schedule(),  so I
guess the skb is buffered in Qdisc, but what about sending process?


Thanks and regards,
pingfan

Re: net-sched: When TX over-limit happens on netdev, what does kernel do?

[Eric Dumazet]

On Sun, 2012-04-01 at 16:27 +0800, Liu ping fan wrote:
> Hi,
> 
> Could anybody tell me when TX over-limit happens on netdev, what does kernel do?
> Is skb buffered in Qdisc->q and the sending process keeps running?

Are you referring to the NETDEV_TX_BUSY return code processing ?

> Or we block the sending process and make it scheduled?
> 

We never block the sender process in qdisc layer.

Only upper protocol layers can eventually block if too many packets are
in flight on the socket.

> Found some code in net/sched/sch_cbq.c
> cbq_dequeue(struct Qdisc *sch) --> qdisc_watchdog_schedule(),  so I
> guess the skb is buffered in Qdisc, but what about sending process?

sender queued its packet and doesnt care if the packet is sent or not.

Only problem is when qdisc limit (number of packets in queue) is hit,
then we can return an error to sender (packet dropped or congestion
notification)

Re: net-sched: When TX over-limit happens on netdev, what does kernel do?

[Liu ping fan]

On Sun, Apr 1, 2012 at 4:44 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
> On Sun, 2012-04-01 at 16:27 +0800, Liu ping fan wrote:
>> Hi,
>>
>> Could anybody tell me when TX over-limit happens on netdev, what does kernel do?
>> Is skb buffered in Qdisc->q and the sending process keeps running?
>
> Are you referring to the NETDEV_TX_BUSY return code processing ?
>
Yes.
>> Or we block the sending process and make it scheduled?
>>
>
> We never block the sender process in qdisc layer.
>
For qdisc layer in softirq?

> Only upper protocol layers can eventually block if too many packets are
> in flight on the socket.
>
>> Found some code in net/sched/sch_cbq.c
>> cbq_dequeue(struct Qdisc *sch) --> qdisc_watchdog_schedule(),  so I
>> guess the skb is buffered in Qdisc, but what about sending process?
>
> sender queued its packet and doesnt care if the packet is sent or not.
>
> Only problem is when qdisc limit (number of packets in queue) is hit,
> then we can return an error to sender (packet dropped or congestion
> notification)
>
Got it, thanks.

Regards,
pingfan
>
>