On Thu, Jul 1, 2021 at 11:39 AM David Ahern <dsahern@xxxxxxxxx> wrote: > > [ adding Paolo, author of 18f25dc39990 ] > > On 7/1/21 4:47 AM, Matthias Treydte wrote: > > Hello, > > > > we recently upgraded the Linux kernel from 5.11.21 to 5.12.12 in our > > video stream receiver appliance and noticed compression artifacts on > > video streams that were previously looking fine. We are receiving UDP > > multicast MPEG TS streams through an FFMpeg / libav layer which does the > > socket and lower level protocol handling. For affected kernels it spills > > the log with messages like > > > >> [mpegts @ 0x7fa130000900] Packet corrupt (stream = 0, dts = 6870802195). > >> [mpegts @ 0x7fa11c000900] Packet corrupt (stream = 0, dts = 6870821068). > > > > Bisecting identified commit 18f25dc399901426dff61e676ba603ff52c666f7 as > > the one introducing the problem in the mainline kernel. It was > > backported to the 5.12 series in > > 450687386cd16d081b58cd7a342acff370a96078. Some random observations which > > may help to understand what's going on: > > > > * the problem exists in Linux 5.13 > > * reverting that commit on top of 5.13 makes the problem go away > > * Linux 5.10.45 is fine > > * no relevant output in dmesg > > * can be reproduced on different hardware (Intel, AMD, different > > NICs, ...) > > * we do use the bonding driver on the systems (but I did not yet > > verify that this is related) > > * we do not use vxlan (mentioned in the commit message) > > * the relevant code in FFMpeg identifying packet corruption is here: > > > > https://github.com/FFmpeg/FFmpeg/blob/master/libavformat/mpegts.c#L2758 > > > > And the bonding configuration: > > > > # cat /proc/net/bonding/bond0 > > Ethernet Channel Bonding Driver: v5.10.45 > > > > Bonding Mode: fault-tolerance (active-backup) > > Primary Slave: None > > Currently Active Slave: enp2s0 > > MII Status: up > > MII Polling Interval (ms): 100 > > Up Delay (ms): 0 > > Down Delay (ms): 0 > > Peer Notification Delay (ms): 0 > > > > Slave Interface: enp2s0 > > MII Status: up > > Speed: 1000 Mbps > > Duplex: full > > Link Failure Count: 0 > > Permanent HW addr: 80:ee:73:XX:XX:XX > > Slave queue ID: 0 > > > > Slave Interface: enp3s0 > > MII Status: down > > Speed: Unknown > > Duplex: Unknown > > Link Failure Count: 0 > > Permanent HW addr: 80:ee:73:XX:XX:XX > > Slave queue ID: 0 > > > > > > If there is anything else I can do to help tracking this down please let > > me know. That library does not enable UDP_GRO. You do not have any UDP based tunnel devices (besides vxlan) configured, either, right? Then no socket lookup should take place, so sk is NULL. It is also unlikely that the device has either of NETIF_F_GRO_FRAGLIST or NETIF_F_GRO_UDP_FWD configured. This can be checked with `ethtool -K $DEV`, shown as "rx-gro-list" and "rx-udp-gro-forwarding", respectively. Then udp_gro_receive_segment is not called. So this should just return the packet without applying any GRO. I'm referring to this block of code in udp_gro_receive: if (!sk || !udp_sk(sk)->gro_receive) { if (skb->dev->features & NETIF_F_GRO_FRAGLIST) NAPI_GRO_CB(skb)->is_flist = sk ? !udp_sk(sk)->gro_enabled : 1; if ((!sk && (skb->dev->features & NETIF_F_GRO_UDP_FWD)) || (sk && udp_sk(sk)->gro_enabled) || NAPI_GRO_CB(skb)->is_flist) pp = call_gro_receive(udp_gro_receive_segment, head, skb); return pp; } I don't see what could be up. One possible short-term workaround is to disable GRO. If this commit is implicated, that should fix it. At some obvious possible cycle cost.
