tcp/dccp: lockless listener
From: | Eric Dumazet <edumazet@google.com> | |
To: | "David S . Miller" <davem@davemloft.net> | |
Subject: | [PATCH net-next 00/17] tcp/dccp: lockless listener | |
Date: | Fri, 2 Oct 2015 11:43:22 -0700 | |
Message-ID: | <1443811419-4798-1-git-send-email-edumazet@google.com> | |
Cc: | netdev <netdev@vger.kernel.org>, Eric Dumazet <edumazet@google.com>, Ying Cai <ycai@google.com>, Willem de Bruijn <willemb@google.com>, Eric Dumazet <eric.dumazet@gmail.com> | |
Archive‑link: | Article |
TCP listener refactoring : this is becoming interesting ! This patch series takes the steps to use normal TCP/DCCP ehash table to store SYN_RECV requests, instead of the private per-listener hash table we had until now. SYNACK skb are now attached to their syn_recv request socket, so that we no longer heavily modify listener sk_wmem_alloc. listener lock is no longer held in fast path, including SYNCOOKIE mode. During my tests, my server was able to process 3,500,000 SYN packets per second on one listener and still had available cpu cycles. That is about 2 to 3 order of magnitude what we had with older kernels. This effort started two years ago and I am pleased to reach expectations. We'll probably extend SO_REUSEPORT to add proper cpu/numa affinities, so that heavy duty TCP servers can get proper siloing thanks to multi-queues NIC. Eric Dumazet (17): tcp: add a spinlock to protect struct request_sock_queue tcp: move qlen/young out of struct listen_sock tcp: move synflood_warned into struct request_sock_queue tcp: call sk_mark_napi_id() on the child, not the listener tcp/dccp: init sk_prot and call sk_node_init() in reqsk_alloc() tcp: cleanup tcp_v[46]_inbound_md5_hash() tcp: remove BUG_ON() in tcp_check_req() tcp: get_openreq[46]() changes tcp/dccp: remove inet_csk_reqsk_queue_added() timeout argument tcp/dccp: install syn_recv requests into ehash table tcp/dccp: shrink struct listen_sock ipv6: remove obsolete inet6 functions tcp: attach SYNACK messages to request sockets instead of listener tcp/dccp: remove struct listen_sock tcp: remove max_qlen_log tcp/dccp: add a reschedule point in inet_csk_listen_stop() tcp: do not lock listener to process SYN packets include/net/inet6_connection_sock.h | 9 -- include/net/inet_connection_sock.h | 9 +- include/net/inet_hashtables.h | 1 + include/net/request_sock.h | 132 ++++++++++-------------------- include/net/tcp.h | 10 +-- net/core/request_sock.c | 84 +------------------ net/dccp/ipv4.c | 64 +++++---------- net/dccp/ipv6.c | 72 +++++----------- net/ipv4/inet_connection_sock.c | 145 ++++++++------------------------ net/ipv4/inet_diag.c | 96 ++-------------------- net/ipv4/inet_hashtables.c | 14 +++- net/ipv4/syncookies.c | 4 + net/ipv4/tcp_fastopen.c | 4 +- net/ipv4/tcp_input.c | 30 +++---- net/ipv4/tcp_ipv4.c | 159 +++++++++++++----------------------- net/ipv4/tcp_minisocks.c | 2 - net/ipv4/tcp_output.c | 22 +++-- net/ipv6/inet6_connection_sock.c | 67 --------------- net/ipv6/tcp_ipv6.c | 117 +++++++++++++------------- net/sched/sch_fq.c | 12 +-- 20 files changed, 307 insertions(+), 746 deletions(-) -- 2.6.0.rc2.230.g3dd15c0