On 27.03.2019 16:20, Mike Crowe wrote: > On Wednesday 27 March 2019 at 15:05:46 +0000, Mike Crowe wrote: >> After upgrading the kernel to v4.19 we started seeing increased packet >> latency on two machines with ASUS H87M-E motherboards with on-board Realtek >> Gigabit Ethernet. This didn't happen with v4.18. (More details, including >> kernel log and lspci output at >> https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=925496 ) >> >> Bisecting revealed that commit >> >> a99790bf5c7f3d68d8b01e015d3212a98ee7bd57 r8169: Reinstate ASPM Support >> >> provoked this change in behaviour. >> >> If I patch v4.9.31's drivers/net/ethernet/realtek/r8169.c >> rtl_hw_aspm_clkreq_enable function to never enable the clocks (which, >> AFAICT, is equivalent to reverting the aforementioned commit) then packet >> latency is normal. >> >> The increased packet latency can also be resolved by: >> >> 1. Running 'ethtool -C eth0 rx-usecs 0 rx-frames 1' >> >> 2. Disabling "EPU Power Saving Mode" in the BIOS (on the Ai Tweaker page.) >> This option is enabled when the "ASUS Optimal" setting is selected on >> the BIOS front page. >> >> but neither of these solutions are ideal because they require the user to >> realise that there's a problem and discover these solutions. >> >> Is there a way to fix the code to work correctly with this motherboard even >> with "EPU Power Saving Mode" enabled? > > Unfortunately, my searching prior to sending that email did not reveal that > this problem has been reported before[1]. Apologies for the noise. > > Mike. > > [1] > https://lore.kernel.org/netdev/camlo_r7q6qcr7kx5goovj_bzdlu_tnw3ukmkzwdml4c2bhu...@mail.gmail.com/ > Thanks for the report. Actually the ASPM-related issue seems to occur in different flavors, but it affects only a certain number of network chip / board combinations (even though it's not known which combinations). And it's not clear who's to blame, board/BIOS or network chip. Most likely both. Unfortunately disabling ASPM would result in notebook users complaining, because ASPM provides a significant power saving.
But interesting is that the mentioned coalesce setting fixes the issue for you. I'll check with other affected users whether it's the same for them. Heiner
