Martin KaFai Lau wrote:

On 2/28/24 7:53 AM, Willem de Bruijn wrote:

quoted

Sidenote: with sk_clockid, FQ could detect when skb->tstamp is not
set in monotonic (i.e., set by SO_TXTIME) and drop the packet or
ignore the embedded timestamp, warn, etc.

Thanks for cc-ing me. Sorry for the late reply. I just catch up to this thread 
and the v1.

I think it is needed to detect if skb->tstamp is monotonic or not in fq. The 
container (with the veth setup) may use sch_etf while the host usually uses fq 
at the physical NIC and expects monotonic skb->tstamp.

During forward (e.g. by bpf_redirect / ip[6]_forward from a veth to a physical 
NIC), skb_clear_tstamp() only forwards the monotonic skb->tstamp now. While 
sch_etf does check sk_clockid first before using skb->tstamp, fq does not check 
that now.
or fq_packet_beyond_horizon() is enough to catch this clock discrepancy?

Before your patch, I believe FQ had no such guard rails. An skb with
any clockid from SO_TXTIME can arrive at FQ.

With the new clockid field, we could add guard rails in fq_enqueue.
If the bit is set, look up sk_clockid.