From bd456f283b66704920fae8e655ebc769cb743420 Mon Sep 17 00:00:00 2001 From: Mubashir Adnan Qureshi Date: Wed, 26 Oct 2022 13:51:11 +0000 Subject: tcp: add sysctls for TCP PLB parameters PLB (Protective Load Balancing) is a host based mechanism for load balancing across switch links. It leverages congestion signals(e.g. ECN) from transport layer to randomly change the path of the connection experiencing congestion. PLB changes the path of the connection by changing the outgoing IPv6 flow label for IPv6 connections (implemented in Linux by calling sk_rethink_txhash()). Because of this implementation mechanism, PLB can currently only work for IPv6 traffic. For more information, see the SIGCOMM 2022 paper: https://doi.org/10.1145/3544216.3544226 This commit adds new sysctl knobs and sets their default values for TCP PLB. Signed-off-by: Mubashir Adnan Qureshi Signed-off-by: Yuchung Cheng Signed-off-by: Neal Cardwell Reviewed-by: Eric Dumazet Signed-off-by: David S. Miller --- include/net/netns/ipv4.h | 5 +++++ 1 file changed, 5 insertions(+) (limited to 'include') diff --git a/include/net/netns/ipv4.h b/include/net/netns/ipv4.h index 1b8004679445..25f90bba4889 100644 --- a/include/net/netns/ipv4.h +++ b/include/net/netns/ipv4.h @@ -183,6 +183,11 @@ struct netns_ipv4 { unsigned long tfo_active_disable_stamp; u32 tcp_challenge_timestamp; u32 tcp_challenge_count; + u8 sysctl_tcp_plb_enabled; + u8 sysctl_tcp_plb_idle_rehash_rounds; + u8 sysctl_tcp_plb_rehash_rounds; + u8 sysctl_tcp_plb_suspend_rto_sec; + int sysctl_tcp_plb_cong_thresh; int sysctl_udp_wmem_min; int sysctl_udp_rmem_min; -- cgit v1.2.3 From 1a91bb7c3ebf95e908ec33220defbcda1ecc072f Mon Sep 17 00:00:00 2001 From: Mubashir Adnan Qureshi Date: Wed, 26 Oct 2022 13:51:12 +0000 Subject: tcp: add PLB functionality for TCP Congestion control algorithms track PLB state and cause the connection to trigger a path change when either of the 2 conditions is satisfied: - No packets are in flight and (# consecutive congested rounds >= sysctl_tcp_plb_idle_rehash_rounds) - (# consecutive congested rounds >= sysctl_tcp_plb_rehash_rounds) A round (RTT) is marked as congested when congestion signal (ECN ce_ratio) over an RTT is greater than sysctl_tcp_plb_cong_thresh. In the event of RTO, PLB (via tcp_write_timeout()) triggers a path change and disables congestion-triggered path changes for random time between (sysctl_tcp_plb_suspend_rto_sec, 2*sysctl_tcp_plb_suspend_rto_sec) to avoid hopping onto the "connectivity blackhole". RTO-triggered path changes can still happen during this cool-off period. Signed-off-by: Mubashir Adnan Qureshi Signed-off-by: Yuchung Cheng Signed-off-by: Neal Cardwell Reviewed-by: Eric Dumazet Signed-off-by: David S. Miller --- include/net/tcp.h | 28 ++++++++++++++ net/ipv4/Makefile | 2 +- net/ipv4/tcp_ipv4.c | 2 +- net/ipv4/tcp_plb.c | 107 ++++++++++++++++++++++++++++++++++++++++++++++++++++ 4 files changed, 137 insertions(+), 2 deletions(-) create mode 100644 net/ipv4/tcp_plb.c (limited to 'include') diff --git a/include/net/tcp.h b/include/net/tcp.h index 14d45661a84d..6b814e788f00 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -2140,6 +2140,34 @@ extern void tcp_rack_advance(struct tcp_sock *tp, u8 sacked, u32 end_seq, extern void tcp_rack_reo_timeout(struct sock *sk); extern void tcp_rack_update_reo_wnd(struct sock *sk, struct rate_sample *rs); +/* tcp_plb.c */ + +/* + * Scaling factor for fractions in PLB. For example, tcp_plb_update_state + * expects cong_ratio which represents fraction of traffic that experienced + * congestion over a single RTT. In order to avoid floating point operations, + * this fraction should be mapped to (1 << TCP_PLB_SCALE) and passed in. + */ +#define TCP_PLB_SCALE 8 + +/* State for PLB (Protective Load Balancing) for a single TCP connection. */ +struct tcp_plb_state { + u8 consec_cong_rounds:5, /* consecutive congested rounds */ + unused:3; + u32 pause_until; /* jiffies32 when PLB can resume rerouting */ +}; + +static inline void tcp_plb_init(const struct sock *sk, + struct tcp_plb_state *plb) +{ + plb->consec_cong_rounds = 0; + plb->pause_until = 0; +} +void tcp_plb_update_state(const struct sock *sk, struct tcp_plb_state *plb, + const int cong_ratio); +void tcp_plb_check_rehash(struct sock *sk, struct tcp_plb_state *plb); +void tcp_plb_update_state_upon_rto(struct sock *sk, struct tcp_plb_state *plb); + /* At how many usecs into the future should the RTO fire? */ static inline s64 tcp_rto_delta_us(const struct sock *sk) { diff --git a/net/ipv4/Makefile b/net/ipv4/Makefile index bbdd9c44f14e..af7d2cf490fb 100644 --- a/net/ipv4/Makefile +++ b/net/ipv4/Makefile @@ -10,7 +10,7 @@ obj-y := route.o inetpeer.o protocol.o \ tcp.o tcp_input.o tcp_output.o tcp_timer.o tcp_ipv4.o \ tcp_minisocks.o tcp_cong.o tcp_metrics.o tcp_fastopen.o \ tcp_rate.o tcp_recovery.o tcp_ulp.o \ - tcp_offload.o datagram.o raw.o udp.o udplite.o \ + tcp_offload.o tcp_plb.o datagram.o raw.o udp.o udplite.o \ udp_offload.o arp.o icmp.o devinet.o af_inet.o igmp.o \ fib_frontend.o fib_semantics.o fib_trie.o fib_notifier.o \ inet_fragment.o ping.o ip_tunnel_core.o gre_offload.o \ diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c index 58b838b56c7f..ebab9e8b184c 100644 --- a/net/ipv4/tcp_ipv4.c +++ b/net/ipv4/tcp_ipv4.c @@ -3224,7 +3224,7 @@ static int __net_init tcp_sk_init(struct net *net) net->ipv4.sysctl_tcp_plb_rehash_rounds = 12; net->ipv4.sysctl_tcp_plb_suspend_rto_sec = 60; /* Default congestion threshold for PLB to mark a round is 50% */ - net->ipv4.sysctl_tcp_plb_cong_thresh = 128; + net->ipv4.sysctl_tcp_plb_cong_thresh = (1 << TCP_PLB_SCALE) / 2; /* Reno is always built in */ if (!net_eq(net, &init_net) && diff --git a/net/ipv4/tcp_plb.c b/net/ipv4/tcp_plb.c new file mode 100644 index 000000000000..f4ced370acad --- /dev/null +++ b/net/ipv4/tcp_plb.c @@ -0,0 +1,107 @@ +/* Protective Load Balancing (PLB) + * + * PLB was designed to reduce link load imbalance across datacenter + * switches. PLB is a host-based optimization; it leverages congestion + * signals from the transport layer to randomly change the path of the + * connection experiencing sustained congestion. PLB prefers to repath + * after idle periods to minimize packet reordering. It repaths by + * changing the IPv6 Flow Label on the packets of a connection, which + * datacenter switches include as part of ECMP/WCMP hashing. + * + * PLB is described in detail in: + * + * Mubashir Adnan Qureshi, Yuchung Cheng, Qianwen Yin, Qiaobin Fu, + * Gautam Kumar, Masoud Moshref, Junhua Yan, Van Jacobson, + * David Wetherall,Abdul Kabbani: + * "PLB: Congestion Signals are Simple and Effective for + * Network Load Balancing" + * In ACM SIGCOMM 2022, Amsterdam Netherlands. + * + */ + +#include + +/* Called once per round-trip to update PLB state for a connection. */ +void tcp_plb_update_state(const struct sock *sk, struct tcp_plb_state *plb, + const int cong_ratio) +{ + struct net *net = sock_net(sk); + + if (!READ_ONCE(net->ipv4.sysctl_tcp_plb_enabled)) + return; + + if (cong_ratio >= 0) { + if (cong_ratio < READ_ONCE(net->ipv4.sysctl_tcp_plb_cong_thresh)) + plb->consec_cong_rounds = 0; + else if (plb->consec_cong_rounds < + READ_ONCE(net->ipv4.sysctl_tcp_plb_rehash_rounds)) + plb->consec_cong_rounds++; + } +} +EXPORT_SYMBOL_GPL(tcp_plb_update_state); + +/* Check whether recent congestion has been persistent enough to warrant + * a load balancing decision that switches the connection to another path. + */ +void tcp_plb_check_rehash(struct sock *sk, struct tcp_plb_state *plb) +{ + struct net *net = sock_net(sk); + u32 max_suspend; + bool forced_rehash = false, idle_rehash = false; + + if (!READ_ONCE(net->ipv4.sysctl_tcp_plb_enabled)) + return; + + forced_rehash = plb->consec_cong_rounds >= + READ_ONCE(net->ipv4.sysctl_tcp_plb_rehash_rounds); + /* If sender goes idle then we check whether to rehash. */ + idle_rehash = READ_ONCE(net->ipv4.sysctl_tcp_plb_idle_rehash_rounds) && + !tcp_sk(sk)->packets_out && + plb->consec_cong_rounds >= + READ_ONCE(net->ipv4.sysctl_tcp_plb_idle_rehash_rounds); + + if (!forced_rehash && !idle_rehash) + return; + + /* Note that tcp_jiffies32 can wrap; we detect wraps by checking for + * cases where the max suspension end is before the actual suspension + * end. We clear pause_until to 0 to indicate there is no recent + * RTO event that constrains PLB rehashing. + */ + max_suspend = 2 * READ_ONCE(net->ipv4.sysctl_tcp_plb_suspend_rto_sec) * HZ; + if (plb->pause_until && + (!before(tcp_jiffies32, plb->pause_until) || + before(tcp_jiffies32 + max_suspend, plb->pause_until))) + plb->pause_until = 0; + + if (plb->pause_until) + return; + + sk_rethink_txhash(sk); + plb->consec_cong_rounds = 0; +} +EXPORT_SYMBOL_GPL(tcp_plb_check_rehash); + +/* Upon RTO, disallow load balancing for a while, to avoid having load + * balancing decisions switch traffic to a black-holed path that was + * previously avoided with a sk_rethink_txhash() call at RTO time. + */ +void tcp_plb_update_state_upon_rto(struct sock *sk, struct tcp_plb_state *plb) +{ + struct net *net = sock_net(sk); + u32 pause; + + if (!READ_ONCE(net->ipv4.sysctl_tcp_plb_enabled)) + return; + + pause = READ_ONCE(net->ipv4.sysctl_tcp_plb_suspend_rto_sec) * HZ; + pause += prandom_u32_max(pause); + plb->pause_until = tcp_jiffies32 + pause; + + /* Reset PLB state upon RTO, since an RTO causes a sk_rethink_txhash() call + * that may switch this connection to a path with completely different + * congestion characteristics. + */ + plb->consec_cong_rounds = 0; +} +EXPORT_SYMBOL_GPL(tcp_plb_update_state_upon_rto); -- cgit v1.2.3 From 29c1c44646aec5d5134f2365259a84becc1ee7d3 Mon Sep 17 00:00:00 2001 From: Mubashir Adnan Qureshi Date: Wed, 26 Oct 2022 13:51:14 +0000 Subject: tcp: add u32 counter in tcp_sock and an SNMP counter for PLB A u32 counter is added to tcp_sock for counting the number of PLB triggered rehashes for a TCP connection. An SNMP counter is also added to count overall PLB triggered rehash events for a host. These counters are hooked up to PLB implementation for DCTCP. TCP_NLA_REHASH is added to SCM_TIMESTAMPING_OPT_STATS that reports the rehash attempts triggered due to PLB or timeouts. This gives a historical view of sustained congestion or timeouts experienced by the TCP connection. Signed-off-by: Mubashir Adnan Qureshi Signed-off-by: Yuchung Cheng Signed-off-by: Neal Cardwell Reviewed-by: Eric Dumazet Signed-off-by: David S. Miller --- include/linux/tcp.h | 1 + include/uapi/linux/snmp.h | 1 + include/uapi/linux/tcp.h | 1 + net/ipv4/proc.c | 1 + net/ipv4/tcp.c | 3 +++ net/ipv4/tcp_plb.c | 2 ++ 6 files changed, 9 insertions(+) (limited to 'include') diff --git a/include/linux/tcp.h b/include/linux/tcp.h index 41b1da621a45..ca7f05a130d2 100644 --- a/include/linux/tcp.h +++ b/include/linux/tcp.h @@ -423,6 +423,7 @@ struct tcp_sock { u32 probe_seq_start; u32 probe_seq_end; } mtu_probe; + u32 plb_rehash; /* PLB-triggered rehash attempts */ u32 mtu_info; /* We received an ICMP_FRAG_NEEDED / ICMPV6_PKT_TOOBIG * while socket was owned by user. */ diff --git a/include/uapi/linux/snmp.h b/include/uapi/linux/snmp.h index 4d7470036a8b..6600cb0164c2 100644 --- a/include/uapi/linux/snmp.h +++ b/include/uapi/linux/snmp.h @@ -292,6 +292,7 @@ enum LINUX_MIB_TCPDSACKIGNOREDDUBIOUS, /* TCPDSACKIgnoredDubious */ LINUX_MIB_TCPMIGRATEREQSUCCESS, /* TCPMigrateReqSuccess */ LINUX_MIB_TCPMIGRATEREQFAILURE, /* TCPMigrateReqFailure */ + LINUX_MIB_TCPPLBREHASH, /* TCPPLBRehash */ __LINUX_MIB_MAX }; diff --git a/include/uapi/linux/tcp.h b/include/uapi/linux/tcp.h index 8fc09e8638b3..c9abe86eda5f 100644 --- a/include/uapi/linux/tcp.h +++ b/include/uapi/linux/tcp.h @@ -315,6 +315,7 @@ enum { TCP_NLA_BYTES_NOTSENT, /* Bytes in write queue not yet sent */ TCP_NLA_EDT, /* Earliest departure time (CLOCK_MONOTONIC) */ TCP_NLA_TTL, /* TTL or hop limit of a packet received */ + TCP_NLA_REHASH, /* PLB and timeout triggered rehash attempts */ }; /* for TCP_MD5SIG socket option */ diff --git a/net/ipv4/proc.c b/net/ipv4/proc.c index 5386f460bd20..f88daace9de3 100644 --- a/net/ipv4/proc.c +++ b/net/ipv4/proc.c @@ -297,6 +297,7 @@ static const struct snmp_mib snmp4_net_list[] = { SNMP_MIB_ITEM("TCPDSACKIgnoredDubious", LINUX_MIB_TCPDSACKIGNOREDDUBIOUS), SNMP_MIB_ITEM("TCPMigrateReqSuccess", LINUX_MIB_TCPMIGRATEREQSUCCESS), SNMP_MIB_ITEM("TCPMigrateReqFailure", LINUX_MIB_TCPMIGRATEREQFAILURE), + SNMP_MIB_ITEM("TCPPLBRehash", LINUX_MIB_TCPPLBREHASH), SNMP_MIB_SENTINEL }; diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c index ef14efa1fb70..1da7c53b6cb5 100644 --- a/net/ipv4/tcp.c +++ b/net/ipv4/tcp.c @@ -3176,6 +3176,7 @@ int tcp_disconnect(struct sock *sk, int flags) tp->sacked_out = 0; tp->tlp_high_seq = 0; tp->last_oow_ack_time = 0; + tp->plb_rehash = 0; /* There's a bubble in the pipe until at least the first ACK. */ tp->app_limited = ~0U; tp->rack.mstamp = 0; @@ -3973,6 +3974,7 @@ static size_t tcp_opt_stats_get_size(void) nla_total_size(sizeof(u32)) + /* TCP_NLA_BYTES_NOTSENT */ nla_total_size_64bit(sizeof(u64)) + /* TCP_NLA_EDT */ nla_total_size(sizeof(u8)) + /* TCP_NLA_TTL */ + nla_total_size(sizeof(u32)) + /* TCP_NLA_REHASH */ 0; } @@ -4049,6 +4051,7 @@ struct sk_buff *tcp_get_timestamping_opt_stats(const struct sock *sk, nla_put_u8(stats, TCP_NLA_TTL, tcp_skb_ttl_or_hop_limit(ack_skb)); + nla_put_u32(stats, TCP_NLA_REHASH, tp->plb_rehash + tp->timeout_rehash); return stats; } diff --git a/net/ipv4/tcp_plb.c b/net/ipv4/tcp_plb.c index f4ced370acad..bb1a08fda113 100644 --- a/net/ipv4/tcp_plb.c +++ b/net/ipv4/tcp_plb.c @@ -79,6 +79,8 @@ void tcp_plb_check_rehash(struct sock *sk, struct tcp_plb_state *plb) sk_rethink_txhash(sk); plb->consec_cong_rounds = 0; + tcp_sk(sk)->plb_rehash++; + NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPPLBREHASH); } EXPORT_SYMBOL_GPL(tcp_plb_check_rehash); -- cgit v1.2.3 From 71fc704768f601ed3fa36310822a5e03f310f781 Mon Sep 17 00:00:00 2001 From: Mubashir Adnan Qureshi Date: Wed, 26 Oct 2022 13:51:15 +0000 Subject: tcp: add rcv_wnd and plb_rehash to TCP_INFO rcv_wnd can be useful to diagnose TCP performance where receiver window becomes the bottleneck. rehash reports the PLB and timeout triggered rehash attempts by the TCP connection. Signed-off-by: Mubashir Adnan Qureshi Signed-off-by: Yuchung Cheng Signed-off-by: Neal Cardwell Reviewed-by: Eric Dumazet Signed-off-by: David S. Miller --- include/uapi/linux/tcp.h | 5 +++++ net/ipv4/tcp.c | 2 ++ 2 files changed, 7 insertions(+) (limited to 'include') diff --git a/include/uapi/linux/tcp.h b/include/uapi/linux/tcp.h index c9abe86eda5f..879eeb0a084b 100644 --- a/include/uapi/linux/tcp.h +++ b/include/uapi/linux/tcp.h @@ -284,6 +284,11 @@ struct tcp_info { __u32 tcpi_snd_wnd; /* peer's advertised receive window after * scaling (bytes) */ + __u32 tcpi_rcv_wnd; /* local advertised receive window after + * scaling (bytes) + */ + + __u32 tcpi_rehash; /* PLB or timeout triggered rehash attempts */ }; /* netlink attributes types for SCM_TIMESTAMPING_OPT_STATS */ diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c index 1da7c53b6cb5..de8f0cd7cb32 100644 --- a/net/ipv4/tcp.c +++ b/net/ipv4/tcp.c @@ -3940,6 +3940,8 @@ void tcp_get_info(struct sock *sk, struct tcp_info *info) info->tcpi_reord_seen = tp->reord_seen; info->tcpi_rcv_ooopack = tp->rcv_ooopack; info->tcpi_snd_wnd = tp->snd_wnd; + info->tcpi_rcv_wnd = tp->rcv_wnd; + info->tcpi_rehash = tp->plb_rehash + tp->timeout_rehash; info->tcpi_fastopen_client_fail = tp->fastopen_client_fail; unlock_sock_fast(sk, slow); } -- cgit v1.2.3