On 8/8/19 10:08 PM, Heiner Kallweit wrote: (..snip..)
I was about to ask exactly that, whether you have TSO enabled. I don't know what can trigger the HW issue, it was just confirmed by Realtek that this chip version has a problem with TSO. So the logical conclusion is: test w/o TSO, ideally the linux-next version.So disabling TSO alone didn't work - it leads to reduced throughout (~70 MB/s in iperf). Instead I decided to backport 93681cd7d94f ("r8169: enable HW csum and TSO"), which wasn't easy due to cleanups/renamings of dependencies, but I managed to backport it and .. got the same problem of reduced throughout. wat?! After lots of trial & error I started disabling all offloads and finally found that sg (Scatter-Gather) enabled alone - without TSO - will lead to the throughput drop. So the culprit seems 93681cd7d94f, which disabled TSO on my NIC, but left sg on by default. This weas repeatable - switch on sg, throughput drop; turn it off - smooth sailing, now with reduced buffers. I modified the relevant bits to disable tso & sg like this: /* RTL8168e-vl has a HW issue with TSO */ if (tp->mac_version == RTL_GIGA_MAC_VER_34) { + dev->vlan_features &= ~(NETIF_F_ALL_TSO|NETIF_F_SG); + dev->hw_features &= ~(NETIF_F_ALL_TSO|NETIF_F_SG); + dev->features &= ~(NETIF_F_ALL_TSO|NETIF_F_SG); } This seems to work since it restores performance without sg/tso by default and without any additional offloads, yet with xmit_more in the mix. We'll see whether that is stable over the next few days, but I strongly suspect it will be good and that the hiccups were due to xmit_more/TSO interaction.
So that didn't take long - got another timeout this morning during some random light usage, despite sg/tso being disabled this time. Again the only common element is the xmit_more patch. :( Not sure whether you want to revert this right away or wait for 5.4-rc1 feedback. Maybe this too is chipset-specific?
Thanks a lot for the analysis and testing. Then I'll submit the disabling of SG on RTL8168evl (on your behalf), independent of whether it fixes the timeout issue.
Got it, thanks! Holger
