CC netfilter
On Tue, 15 Mar 2022 23:15:08 +0000 McLean, Patrick wrote:
> When we upgraded from the 5.10 (5.10.61) series to the 5.15 (5.15.16) series, we encountered an intermittent performance regression that appears to be related to iptables / ipset. This regression was noticed on Kubernetes hosts that run kube-router and experience a high amount of churn to both iptables and ipsets. Specifically, when we run the nftables (iptables-1.8.7 / nftables-1.0.0) iptables wrapper xtables-nft-multi on the 5.15 series kernel, we end up getting extremely laggy response times when iptables attempts to lookup information on the ipsets that are used in the iptables definition. This issue isn’t reproducible on all hosts. However, our experience has been that across a fleet of ~50 hosts we experienced this issue on ~40% of the hosts. When the problem evidences, the time that it takes to run unrestricted iptables list commands like iptables -L or iptables-save gradually increases over the course of about 1 - 2 hours. Growing from less than a second to run, to taking sometimes over 2 minutes to run. After that 2 hour mark it seems to plateau and not grow any longer. Flushing tables or ipsets doesn’t seem to have any affect on the issue. However, rebooting the host does reset the issue. Occasionally, a machine that was evidencing the problem may no longer evidence it after being rebooted.
>
> We did try to debug this to find a root cause, but ultimately ran short on time. We were not able to perform a set of bisects to hopefully narrow down the issue as the problem isn’t consistently reproducible. We were able to get some straces where it appears that most of the time is spent on getsockopt() operations. It appears that during iptables operations, it attempts to do some work to resolve the ipsets that are linked to the iptables definitions (perhaps getting the names of the ipsets themselves?). Slowly that getsockopt request takes more and more time on affected hosts. Here is an example strace of the operation in question:
>
> 0.000074 newfstatat(AT_FDCWD, "/etc/nsswitch.conf", {st_mode=S_IFREG|0644, st_size=539, ...}, 0) = 0 <0.000017>
> 0.000064 openat(AT_FDCWD, "/var/db/protocols.db", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory) <0.000017>
> 0.000057 openat(AT_FDCWD, "/etc/protocols", O_RDONLY|O_CLOEXEC) = 4 <0.000013>
> 0.000034 newfstatat(4, "", {st_mode=S_IFREG|0644, st_size=6108, ...}, AT_EMPTY_PATH) = 0 <0.000009>
> 0.000032 lseek(4, 0, SEEK_SET) = 0 <0.000008>
> 0.000025 read(4, "# /etc/protocols\n#\n# Internet (I"..., 4096) = 4096 <0.000010>
> 0.000036 close(4) = 0 <0.000008>
> 0.000028 write(1, "ANGME7BF25 - [0:0]\n:KUBE-POD-FW-"..., 4096) = 4096 <0.000028>
> 0.000049 socket(AF_INET, SOCK_RAW, IPPROTO_RAW) = 4 <0.000015>
> 0.000032 fcntl(4, F_SETFD, FD_CLOEXEC) = 0 <0.000008>
> 0.000024 getsockopt(4, SOL_IP, 0x53 /* IP_??? */, "\0\1\0\0\7\0\0\0", [8]) = 0 <0.000024>
> 0.000046 getsockopt(4, SOL_IP, 0x53 /* IP_??? */, "\7\0\0\0\7\0\0\0KUBE-DST-VBH27M7NWLDOZIE"..., [40]) = 0 <0.109384>
> 0.109456 close(4) = 0 <0.000022>
>
> On a host that is not evidencing the performance regression we normally see that operation take ~ 0.00001 as opposed to 0.109384.Additionally, hosts that were evidencing the problem we also saw high lock times with `klockstat` (unfortunately at the time we did not know about or run echo "0" > /proc/sys/kernel/kptr_restrict to get the callers of the below commands).
>
> klockstat -i 5 -n 10 (on a host experiencing the problem)
> Caller Avg Hold Count Max hold Total hold
> b'[unknown]' 118490772 83 179899470 9834734132
> b'[unknown]' 118416941 83 179850047 9828606138
> # or somewhere later while iptables -vnL was running:
> Caller Avg Hold Count Max hold Total hold
> b'[unknown]' 496466236 46 17919955720 22837446860
> b'[unknown]' 496391064 46 17919893843 22833988950
>
> klockstat -i 5 -n 10 (on a host not experiencing the problem)
> Caller Avg Hold Count Max hold Total hold
> b'[unknown]' 120316 1510 85537797 181677885
> b'[unknown]' 7119070 24 85527251 170857690
