On Fri, Mar 18, 2022 at 9:36 PM Daniel Borkmann <daniel@xxxxxxxxxxxxx> wrote:
>
> On 3/17/22 9:20 AM, Paolo Abeni wrote:
> > On Tue, 2022-03-15 at 20:48 +0800, Tonghao Zhang wrote:
> >> On Tue, Mar 15, 2022 at 5:59 AM Daniel Borkmann <daniel@xxxxxxxxxxxxx> wrote:
> >>> On 3/14/22 3:15 PM, xiangxia.m.yue@xxxxxxxxx wrote:
> >>> [...]
> >>>> include/linux/netdevice.h | 3 +++
> >>>> include/linux/rtnetlink.h | 1 +
> >>>> net/core/dev.c | 31 +++++++++++++++++++++++++++++--
> >>>> net/sched/act_skbedit.c | 6 +++++-
> >>>> 4 files changed, 38 insertions(+), 3 deletions(-)
> >>>>
> >>>> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
> >>>> index 0d994710b335..f33fb2d6712a 100644
> >>>> --- a/include/linux/netdevice.h
> >>>> +++ b/include/linux/netdevice.h
> >>>> @@ -3065,6 +3065,9 @@ struct softnet_data {
> >>>> struct {
> >>>> u16 recursion;
> >>>> u8 more;
> >>>> +#ifdef CONFIG_NET_EGRESS
> >>>> + u8 skip_txqueue;
> >>>> +#endif
> >>>> } xmit;
> >>>> #ifdef CONFIG_RPS
> >>>> /* input_queue_head should be written by cpu owning this struct,
> >>>> diff --git a/include/linux/rtnetlink.h b/include/linux/rtnetlink.h
> >>>> index 7f970b16da3a..ae2c6a3cec5d 100644
> >>>> --- a/include/linux/rtnetlink.h
> >>>> +++ b/include/linux/rtnetlink.h
> >>>> @@ -100,6 +100,7 @@ void net_dec_ingress_queue(void);
> >>>> #ifdef CONFIG_NET_EGRESS
> >>>> void net_inc_egress_queue(void);
> >>>> void net_dec_egress_queue(void);
> >>>> +void netdev_xmit_skip_txqueue(bool skip);
> >>>> #endif
> >>>>
> >>>> void rtnetlink_init(void);
> >>>> diff --git a/net/core/dev.c b/net/core/dev.c
> >>>> index 75bab5b0dbae..8e83b7099977 100644
> >>>> --- a/net/core/dev.c
> >>>> +++ b/net/core/dev.c
> >>>> @@ -3908,6 +3908,25 @@ sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
> >>>>
> >>>> return skb;
> >>>> }
> >>>> +
> >>>> +static struct netdev_queue *
> >>>> +netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb)
> >>>> +{
> >>>> + int qm = skb_get_queue_mapping(skb);
> >>>> +
> >>>> + return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm));
> >>>> +}
> >>>> +
> >>>> +static bool netdev_xmit_txqueue_skipped(void)
> >>>> +{
> >>>> + return __this_cpu_read(softnet_data.xmit.skip_txqueue);
> >>>> +}
> >>>> +
> >>>> +void netdev_xmit_skip_txqueue(bool skip)
> >>>> +{
> >>>> + __this_cpu_write(softnet_data.xmit.skip_txqueue, skip);
> >>>> +}
> >>>> +EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue);
> >>>> #endif /* CONFIG_NET_EGRESS */
> >>>>
> >>>> #ifdef CONFIG_XPS
> >>>> @@ -4078,7 +4097,7 @@ struct netdev_queue *netdev_core_pick_tx(struct net_device *dev,
> >>>> static int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
> >>>> {
> >>>> struct net_device *dev = skb->dev;
> >>>> - struct netdev_queue *txq;
> >>>> + struct netdev_queue *txq = NULL;
> >>>> struct Qdisc *q;
> >>>> int rc = -ENOMEM;
> >>>> bool again = false;
> >>>> @@ -4106,11 +4125,17 @@ static int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
> >>>> if (!skb)
> >>>> goto out;
> >>>> }
> >>>> +
> >>>> + netdev_xmit_skip_txqueue(false);
> >>>> +
> >>>> nf_skip_egress(skb, true);
> >>>> skb = sch_handle_egress(skb, &rc, dev);
> >>>> if (!skb)
> >>>> goto out;
> >>>> nf_skip_egress(skb, false);
> >>>> +
> >>>> + if (netdev_xmit_txqueue_skipped())
> >>>> + txq = netdev_tx_queue_mapping(dev, skb);
> >>>> }
> >>>> #endif
> >>>> /* If device/qdisc don't need skb->dst, release it right now while
> >>>> @@ -4121,7 +4146,9 @@ static int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
> >>>> else
> >>>> skb_dst_force(skb);
> >>>>
> >>>> - txq = netdev_core_pick_tx(dev, skb, sb_dev);
> >>>> + if (likely(!txq))
> >>>
> >>> nit: Drop likely(). If the feature is used from sch_handle_egress(), then this would always be the case.
> >> Hi Daniel
> >> I think in most case, we don't use skbedit queue_mapping in the
> >> sch_handle_egress() , so I add likely in fast path.
>
> Yeah, but then let branch predictor do its work ? We can still change and drop the
> likely() once we add support for BPF though..
Hi
if you are ok that introducing the bpf helper shown below, I will drop
likely() in next patch.
>
> >>>> + txq = netdev_core_pick_tx(dev, skb, sb_dev);
> >>>> +
> >>>> q = rcu_dereference_bh(txq->qdisc);
> >>>
> >>> How will the `netdev_xmit_skip_txqueue(true)` be usable from BPF side (see bpf_convert_ctx_access() ->
> >>> queue_mapping)?
> >> Good questions, In other patch, I introduce the
> >> bpf_netdev_skip_txqueue, so we can use netdev_xmit_skip_txqueue in bpf
> >> side
>
> Yeah, that bpf_netdev_skip_txqueue() won't fly. It's basically a helper doing quirk for
> an implementation detail (aka calling netdev_xmit_skip_txqueue()). Was hoping you have
> something better we could use along with the context rewrite of __sk_buff's queue_mapping,
Hi Daniel
I review the bpf codes, we introduce a lot helper to change the skb field:
skb_change_proto
skb_change_type
skb_change_tail
skb_pull_data
skb_change_head
skb_ecn_set_ce
skb_cgroup_classid
skb_vlan_push
skb_set_tunnel_key
did you mean that, we introduce bpf_skb_set_queue_mapping is better
than bpf_netdev_skip_txqueue.
for example:
BPF_CALL_2(bpf_skb_set_queue_mapping, struct sk_buff *, skb, u32, txq)
{
skb->queue_mapping = txq;
netdev_xmit_skip_txqueue(true);
return 0;
};
> but worst case we need to rework a bit for BPF. :/
> Thanks,
> Daniel
--
Best regards, Tonghao
