Re: [PATCH net-next 0/3] net: bcmgenet: collapse TX priority queues

Hi Jakub

On 13.6.2026 23:57, Jakub Kicinski wrote:

On Fri, 12 Jun 2026 22:59:12 +0200 Nicolai Buchwitz wrote:

quoted

Tested on Raspberry Pi CM4 (BCM2711):
  - Ovidiu's reproducer (iperf3 -u -b0 -P16 -t60) no longer trips
    NETDEV_WATCHDOG.
  - UDP sustains 956 Mbit/s line rate over 60 s with 0 datagrams
    lost (0/4952890).
  - Single-stream TCP throughput unchanged at 943 Mbit/s.

Of course it has no impact on a single TCP stream test, since TCP
stream can only use one queue. If anything it should help.
The testing here is not very convincing. At least install a realistic
qdisc (fq/fq_codel/cake) and run multi-stream test with multiple cores?
What's the CPU idle delta in such a test?

Fair. Tests I ran with fq_codel:

   # TCP
   iperf3 -c <remote> -P 16
   # UDP
   iperf3 -c <remote> -u -b 1000M -P 16
   # RR
   iperf3 -c <remote> -t 60 &
   netperf -H <remote> -t TCP_RR -l 30 -- -r 1,1

With the following results (all based on net-next with PP):
                               PP       PP+WRR    PP+series
   TCP -P16 Mbit/s               938       941        941
   TCP retransmits             56228     57679      48606
   UDP -b1000M -P16 Mbit/s       956       956        956
   TCP_RR under TCP load       451.7     453.8      596.8
   CPU idle                    91.55%   90.53%     90.97%
   CPU0 softirq                32.6%    33.4%      33.5%
   CPU1-3 idle (avg)           97.6%    97.7%      97.7%

So WRR fixes the watchdog issue, otherwise is within noise. This series
adds ~31% TCP_RR under load and reduces retransmits by ~14%. Note, that
bcmgenet on BCM2711 has only two IRQs and no per-queue affinity, so all
HW interrupts are handled by one core regardless of queue count.

The reason for this change is not coming thru from the submission.
Ovidiu's patch makes much more intuitive sense. I'll apply that,
please rebase.

Will rebase and resend with a hopefully better reasoning :)

Thanks
Nicolai