What was stated is that things broke with the 5.15 kernel push when it was a WIP, and that the issue has never been addressed; apparently jow has been working on changes that address the issue.
Can you test on snapshot, youll need to use hardware offload to enable soft offload for all interfaces not only bridges likely.
Above test was on master version:
OpenWrt SNAPSHOT, r27119-c609c3bd6e
-----------------------------------------------------
root@bsaedgy:~# uname -a
Linux bsaedgy 6.6.45 #0 SMP Sun Aug 11 16:56:27 2024 armv7l GNU/Linux
HW offload not really applicable here.
hw offload even with soft fallback will add all leaf devices to offload pool, the "improved" soft offload will stay with only original bridges and interfaces.
I will later tonight
Do i need to report to GitHub or you can do it.
You need real devvice tags'
Sorry didn´t follow you
See the github issue template. Nobody will take seriously that I add x86 version tags and tell wrt1200ac is broken....
I think maybe this WIP(?) is a throughput killer:
root@bsaedgy:/usr/share/ucode# ls -las
0 drwxr-xr-x 1 root root 368 Sep 14 17:31 .
0 drwxr-xr-x 1 root root 224 Sep 4 04:39 ..
0 lrwxrwxrwx 1 root root 9 Sep 14 17:31 fw4.uc -> fw4ORG.uc
80 -rw-r--r-- 1 root root 79432 Sep 4 13:58 fw4ORG.uc
80 -rw-r--r-- 1 root root 79820 Sep 4 13:57 fw4TEST.uc
Necessary evil. If you have better idea on how to keep roaming with offload --ideas are welcome.
Rather non sequitur post. I am simply noting an empirical observation with the intent of informing and perhaps garnering others to proffer their own. To call this a necessary evil is just absurd. If I'm not mistaken, the basic idea behind software offload is to increase throughput, not to decrease things to 1/6 of normal. And
well...
Working around limitations. You can click hw offload if MAC+netdev does not change (and netdev does not flop as in this case)
https://www.kernel.org/doc/html/latest/networking/nf_flowtable.html#limitations
Can you help with measurements - either waveform test or iperf across forward or routing - without offload, with offload old way with physical devices prefered, or new way with bridges. 1043v2 - 150Mbps ... 400Mbps ... 400Mbps
Since hardware and software heuristics are split once can improve either of those past current one-line difference between them.
24.10.0-rc4 and snapshot since a week ago both include the workaround / fix. Shout out loud if similar effect ever returns.
2 Likes
This topic was automatically closed 10 days after the last reply. New replies are no longer allowed.