Poor __inet_check_established() implementation

[krizhanovsky]

Linux kernel 4.19 (standard Debian 9.12 kernel). Couple of runs of wrk like

wrk -H 'Connection: close' -c 32768 -d 30 -t 8

can easily lead to significantly degraded performance on the client machine, sometime up to 20-30 times, with following perf profile.

  36.28%  [kernel]            [k] __inet_check_established
  20.68%  [kernel]            [k] _raw_spin_lock_bh
  14.76%  [kernel]            [k] _raw_spin_lock
  11.17%  [kernel]            [k] native_queued_spin_lock_slowpath
   9.23%  [kernel]            [k] __inet_hash_connect
   3.14%  [kernel]            [k] inet_ehashfn
   1.32%  [kernel]            [k] tcp_twsk_unique
   1.15%  [kernel]            [k] __local_bh_enable_ip
   0.64%  [kernel]            [k] _cond_resched
   0.43%  [kernel]            [k] _raw_spin_unlock_bh
   0.42%  [kernel]            [k] __indirect_thunk_start
   0.19%  [kernel]            [k] rcu_all_qs

Need to check the current kernel implementation. Last time I checked it, it was a dummy hash table with lists. Probably the recent our research in high performance concurrent hash tables for MariaDB can be employed here. Also see how VPP TCP manages per-CPU/per-thread connections hash tables.

[krizhanovsky]

While the hash table tcp_hashinfo is really naive (most of time the function spends in scanning a linked list), __inet_check_established() is called on connect() path only and the issue was observed on the system with multiple TIME-WAIT sockets. The problem can be mitigates with setting tcp_tw_reuse or reducing tcp_max_tw_buckets sysctls.

Meantime, the hash table is used also scanned in the main receiving routine tcp_v4_rcv(), so it still can cause a contention on a multi-core system servicing thousands TCP connections.

The problem was also reported by CloudFlare.

BTW, it seems FreeBSD/F-stack use very similar approach with inpcbinfo (see netinet/in_pcb.h) - see for an example tcp_input() -> in_pcblookup_mbuf(&V_tcbinfo, ...).

[defanator]

TWIMC, https://lore.kernel.org/netdev/20210220110356.84399-1-redsky110@gmail.com/#t

[krizhanovsky]

The same issue is relevant for server-side connections. We saw 4ms delays in __inet_lookup_established().