On 6.02.2023 12:28, Rafał Miłecki wrote:
On 3.02.2023 21:26, Florian Fainelli wrote:
Have you managed to get a `perf record` or `perf top` capture to see whether we
might be going a little too far into iptables-raw code rather than bailing out
early if we are not utilizing it?
Just did.
It doesn't hele me much but maybe you can figure out sth out of it. It
seems that with iptable_raw loaded v7_dma_inv_range() starts getting
partially called from swapper. Without iptable_raw it's called from
ksoftirqd only.
I used FlameGraph to visualize perf results. The best summary is
generated diff one:
http://files.zajec.net/openwrt/bcm53xx-iptable_raw-nat-slow-down/diff.svg
It shows CPU spends more time executing nf_hook_slow which in turn calls
ip_tables.
We also also see CPU spending LESS time executing v7_dma_* which is BAD
for my case as that is what actually handles packets.
Full FlameGraphs:
http://files.zajec.net/openwrt/bcm53xx-iptable_raw-nat-slow-down/fast-without-iptable_raw.svg
http://files.zajec.net/openwrt/bcm53xx-iptable_raw-nat-slow-down/slow-with-iptable_raw.svg
_______________________________________________
openwrt-devel mailing list
[email protected]
https://lists.openwrt.org/mailman/listinfo/openwrt-devel
