tcp: Introduce optional per-netns ehash.
Subsystems (10)
v3 cover tcp: Introduce optional per-netns ehash.
The more sockets we have in the hash table, the longer we spend looking up the socket. While running a number of small workloads on the same host, they penalise each other and cause performance degradation.
The root cause might be a single workload that consumes much more resources than the others. It often happens on a cloud service where different workloads share the same computing resource.
On EC2 c5.24xlarge instance (196 GiB memory and 524288 (1Mi / 2) ehash entries), after running iperf3 in different netns, creating 24Mi sockets without data transfer in the root netns causes about 10% performance regression for the iperf3's connection.
thash_entries sockets length Gbps
524288 1 1 50.7 24Mi 48 45.1
It is basically related to the length of the list of each hash bucket. For testing purposes to see how performance drops along the length, I set 131072 (1Mi / 8) to thash_entries, and here's the result.
thash_entries sockets length Gbps
131072 1 1 50.7 1Mi 8 49.9 2Mi 16 48.9 4Mi 32 47.3 8Mi 64 44.6 16Mi 128 40.6 24Mi 192 36.3 32Mi 256 32.5 40Mi 320 27.0 48Mi 384 25.0
To resolve the socket lookup degradation, we introduce an optional per-netns hash table for TCP, but it's just ehash, and we still share the global bhash, bhash2 and lhash2.
With a smaller ehash, we can look up non-listener sockets faster and isolate such noisy neighbours. Also, we can reduce lock contention.
For details, please see the last patch.
patch 1 - 3: prep for per-netns ehash
patch 4: small optimisation for netns dismantle without TIME_WAIT sockets
patch 5: add per-netns ehash
Changes:
v3:
* Patch 3
* Drop mellanox and netronome driver changes (Eric Dumazet)
* Patch 4
* Add test results in the changelog
* Patch 5
* Move roundup_pow_of_two() into tcp_set_hashinfo() (Eric Dumazet)
* Remove proc_tcp_child_ehash_entries() and use proc_douintvec_minmax()
v2: https://lore.kernel.org/netdev/[email protected]/
* Drop flock() and UDP stuff
* Patch 2
* Rename inet_get_hashinfo() to tcp_or_dccp_get_hashinfo() (Eric Dumazet)
* Patch 4
* Remove unnecessary inet_twsk_purge() calls for unshare()
* Factorise inet_twsk_purge() calls (Eric Dumazet)
* Patch 5
* Change max buckets size as 16Mi
* Use unsigned int for ehash size (Eric Dumazet)
* Use GFP_KERNEL_ACCOUNT for the per-netns ehash allocation (Eric Dumazet)
* Use current->nsproxy->net_ns for parent netns (Eric Dumazet)
v1: https://lore.kernel.org/netdev/[email protected]/
Kuniyuki Iwashima (5):
tcp: Clean up some functions.
tcp: Set NULL to sk->sk_prot->h.hashinfo.
tcp: Access &tcp_hashinfo via net.
tcp: Save unnecessary inet_twsk_purge() calls.
tcp: Introduce optional per-netns ehash.
Documentation/networking/ip-sysctl.rst | 23 ++++ .../chelsio/inline_crypto/chtls/chtls_cm.c | 5 +- include/net/inet_hashtables.h | 16 +++ include/net/netns/ipv4.h | 1 + include/net/tcp.h | 1 + net/core/filter.c | 5 +- net/dccp/proto.c | 2 + net/ipv4/af_inet.c | 2 +- net/ipv4/esp4.c | 3 +- net/ipv4/inet_connection_sock.c | 22 ++-- net/ipv4/inet_hashtables.c | 102 ++++++++++++---- net/ipv4/inet_timewait_sock.c | 4 +- net/ipv4/netfilter/nf_socket_ipv4.c | 2 +- net/ipv4/netfilter/nf_tproxy_ipv4.c | 17 ++- net/ipv4/sysctl_net_ipv4.c | 39 ++++++ net/ipv4/tcp.c | 1 + net/ipv4/tcp_diag.c | 18 ++- net/ipv4/tcp_ipv4.c | 114 ++++++++++++------ net/ipv4/tcp_minisocks.c | 31 ++++- net/ipv6/esp6.c | 3 +- net/ipv6/inet6_hashtables.c | 4 +- net/ipv6/netfilter/nf_socket_ipv6.c | 2 +- net/ipv6/netfilter/nf_tproxy_ipv6.c | 5 +- net/ipv6/tcp_ipv6.c | 20 +-- net/mptcp/mptcp_diag.c | 7 +- 25 files changed, 339 insertions(+), 110 deletions(-)
Signature · 1 line
2.30.2
PATCH v3 1/5 tcp: Clean up some functions.
This patch adds no functional change and cleans up some functions that the following patches touch around so that we make them tidy and easy to review/revert. The changes are
- Keep reverse christmas tree order
- Remove unnecessary init of port in inet_csk_find_open_port()
- Use req_to_sk() once in reqsk_queue_unlink()
--- net/ipv4/inet_connection_sock.c | 21 ++++++++++----------- net/ipv4/inet_hashtables.c | 29 +++++++++++++++-------------- net/ipv4/tcp_ipv4.c | 4 ++-- 3 files changed, 27 insertions(+), 27 deletions(-)
net/ipv4/inet_connection_sock.c +10 -11
@@ -286,15 +286,13 @@ inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret, struct inet_bind_hashbucket **head2_ret, int *port_ret) { struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo; -int port =0 ; +int i, low, high, attempt_half, port, l3mdev;struct inet_bind_hashbucket *head, *head2;struct net *net =sock_net (sk); -bool relax = false; -int i, low, high, attempt_half;struct inet_bind2_bucket *tb2;struct inet_bind_bucket *tb;u32 remaining, offset; -int l3mdev; +bool relax = false; l3mdev =inet_sk_bound_l3mdev (sk); ports_exhausted:
@@ -471,15 +469,14 @@ int inet_csk_get_port(struct sock *sk, unsigned short snum) {bool reuse = sk->sk_reuse && sk->sk_state !=TCP_LISTEN ;struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo; -int ret =1 , port = snum; -struct net *net =sock_net (sk);bool found_port = false, check_bind_conflict = true;bool bhash_created = false, bhash2_created = false;struct inet_bind_hashbucket *head, *head2;struct inet_bind2_bucket *tb2 =NULL ;struct inet_bind_bucket *tb =NULL ;bool head2_lock_acquired = false; -int l3mdev; +int ret =1 , port = snum, l3mdev; +struct net *net =sock_net (sk); l3mdev =inet_sk_bound_l3mdev (sk);
@@ -909,14 +906,16 @@ static void reqsk_migrate_reset(struct request_sock *req)/* return true if req was found in the ehash table */ static bool reqsk_queue_unlink (struct request_sock *req) { -struct inet_hashinfo *hashinfo =req_to_sk (req)->sk_prot->h.hashinfo; +struct sock *sk =req_to_sk (req);bool found = false; -if (sk_hashed (req_to_sk (req))) { -spinlock_t *lock =inet_ehash_lockp (hashinfo, req->rsk_hash); +if (sk_hashed (sk)) { +struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; +spinlock_t *lock; + lock =inet_ehash_lockp (hashinfo, req->rsk_hash);spin_lock (lock); - found =__sk_nulls_del_node_init_rcu (req_to_sk (req)); + found =__sk_nulls_del_node_init_rcu (sk);spin_unlock (lock); }if (timer_pending (&req->rsk_timer) &&del_timer_sync (&req->rsk_timer))
net/ipv4/inet_hashtables.c +15 -14
@@ -169,13 +169,14 @@ void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,static void __inet_put_port (struct sock *sk) {struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; -const int bhash =inet_bhashfn (sock_net (sk),inet_sk (sk)->inet_num, - hashinfo->bhash_size); -struct inet_bind_hashbucket *head = &hashinfo->bhash[bhash]; -struct inet_bind_hashbucket *head2 = -inet_bhashfn_portaddr (hashinfo, sk,sock_net (sk), -inet_sk (sk)->inet_num); +struct inet_bind_hashbucket *head, *head2; +struct net *net =sock_net (sk);struct inet_bind_bucket *tb; +int bhash; + + bhash =inet_bhashfn (net,inet_sk (sk)->inet_num, hashinfo->bhash_size); + head = &hashinfo->bhash[bhash]; + head2 =inet_bhashfn_portaddr (hashinfo, sk, net,inet_sk (sk)->inet_num);spin_lock (&head->lock); tb =inet_csk (sk)->icsk_bind_hash;
@@ -209,17 +210,17 @@ int __inet_inherit_port(const struct sock *sk, struct sock *child) {struct inet_hashinfo *table = sk->sk_prot->h.hashinfo;unsigned short port =inet_sk (child)->inet_num; -const int bhash =inet_bhashfn (sock_net (sk), port, - table->bhash_size); -struct inet_bind_hashbucket *head = &table->bhash[bhash]; -struct inet_bind_hashbucket *head2 = -inet_bhashfn_portaddr (table, child,sock_net (sk), port); +struct inet_bind_hashbucket *head, *head2;bool created_inet_bind_bucket = false; -bool update_fastreuse = false;struct net *net =sock_net (sk); +bool update_fastreuse = false;struct inet_bind2_bucket *tb2;struct inet_bind_bucket *tb; -int l3mdev; +int bhash, l3mdev; + + bhash =inet_bhashfn (net, port, table->bhash_size); + head = &table->bhash[bhash]; + head2 =inet_bhashfn_portaddr (table, child, net, port);spin_lock (&head->lock);spin_lock (&head2->lock);
@@ -629,8 +630,8 @@ static bool inet_ehash_lookup_by_sk(struct sock *sk,bool inet_ehash_insert (struct sock *sk,struct sock *osk,bool *found_dup_sk) {struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; -struct hlist_nulls_head *list;struct inet_ehash_bucket *head; +struct hlist_nulls_head *list;spinlock_t *lock;bool ret = true;
net/ipv4/tcp_ipv4.c +2 -2
@@ -2406,9 +2406,9 @@ static void *established_get_first(struct seq_file *seq)static void *established_get_next (struct seq_file *seq,void *cur) { -struct sock *sk = cur; -struct hlist_nulls_node *node;struct tcp_iter_state *st = seq->private; +struct hlist_nulls_node *node; +struct sock *sk = cur; ++st->num; ++st->offset;
Signature · 1 line
2.30.2
PATCH v3 2/5 tcp: Set NULL to sk->sk_prot->h.hashinfo.
In reply to Kuniyuki IwashimaWe will soon introduce an optional per-netns ehash.
This means we cannot use the global sk->sk_prot->h.hashinfo to fetch a TCP hashinfo.
Instead, set NULL to sk->sk_prot->h.hashinfo for TCP and get a proper hashinfo from net->ipv4.tcp_death_row->hashinfo.
Note that we need not use sk->sk_prot->h.hashinfo if DCCP is disabled.
--- include/net/inet_hashtables.h | 10 ++++++++++ net/ipv4/af_inet.c | 2 +- net/ipv4/inet_connection_sock.c | 9 ++++----- net/ipv4/inet_hashtables.c | 14 +++++++------- net/ipv4/tcp_ipv4.c | 2 +- net/ipv6/tcp_ipv6.c | 2 +- 6 files changed, 24 insertions(+), 15 deletions(-)
include/net/inet_hashtables.h +10 -0
@@ -170,6 +170,16 @@ struct inet_hashinfo {struct inet_listen_hashbucket *lhash2; }; +static inline struct inet_hashinfo *tcp_or_dccp_get_hashinfo (const struct sock *sk) +{ +#if IS_ENABLED (CONFIG_IP_DCCP ) +return sk->sk_prot->h.hashinfo ? : +sock_net (sk)->ipv4.tcp_death_row->hashinfo; +#else +return sock_net (sk)->ipv4.tcp_death_row->hashinfo; +#endif +} +static inline struct inet_listen_hashbucket *inet_lhash2_bucket (struct inet_hashinfo *h,u32 hash) {
net/ipv4/af_inet.c +1 -1
@@ -1250,7 +1250,7 @@ static int inet_sk_reselect_saddr(struct sock *sk) } prev_addr_hashbucket = -inet_bhashfn_portaddr (sk->sk_prot->h.hashinfo, sk, +inet_bhashfn_portaddr (tcp_or_dccp_get_hashinfo (sk), sk,sock_net (sk), inet->inet_num); inet->inet_saddr = inet->inet_rcv_saddr = new_saddr;
net/ipv4/inet_connection_sock.c +4 -5
@@ -285,7 +285,7 @@ inet_csk_find_open_port(const struct sock *sk, struct inet_bind_bucket **tb_ret,struct inet_bind2_bucket **tb2_ret, struct inet_bind_hashbucket **head2_ret, int *port_ret) { - struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo; + struct inet_hashinfo *hinfo = tcp_or_dccp_get_hashinfo(sk);int i, low, high, attempt_half, port, l3mdev;struct inet_bind_hashbucket *head, *head2;struct net *net =sock_net (sk);
@@ -467,8 +467,8 @@ void inet_csk_update_fastreuse(struct inet_bind_bucket *tb,*/ int inet_csk_get_port (struct sock *sk,unsigned short snum) { +struct inet_hashinfo *hinfo =tcp_or_dccp_get_hashinfo (sk);bool reuse = sk->sk_reuse && sk->sk_state !=TCP_LISTEN ; -struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo;bool found_port = false, check_bind_conflict = true;bool bhash_created = false, bhash2_created = false;struct inet_bind_hashbucket *head, *head2;
@@ -910,10 +910,9 @@ static bool reqsk_queue_unlink(struct request_sock *req)bool found = false;if (sk_hashed (sk)) { -struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; -spinlock_t *lock; +struct inet_hashinfo *hashinfo =tcp_or_dccp_get_hashinfo (sk); +spinlock_t *lock =inet_ehash_lockp (hashinfo, req->rsk_hash); - lock =inet_ehash_lockp (hashinfo, req->rsk_hash);spin_lock (lock); found =__sk_nulls_del_node_init_rcu (sk);spin_unlock (lock);
net/ipv4/inet_hashtables.c +7 -7
@@ -168,7 +168,7 @@ void inet_bind_hash(struct sock *sk, struct inet_bind_bucket *tb,*/ static void __inet_put_port (struct sock *sk) { -struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; +struct inet_hashinfo *hashinfo =tcp_or_dccp_get_hashinfo (sk);struct inet_bind_hashbucket *head, *head2;struct net *net =sock_net (sk);struct inet_bind_bucket *tb;
@@ -208,7 +208,7 @@ EXPORT_SYMBOL(inet_put_port);int __inet_inherit_port (const struct sock *sk,struct sock *child) { -struct inet_hashinfo *table = sk->sk_prot->h.hashinfo; +struct inet_hashinfo *table =tcp_or_dccp_get_hashinfo (sk);unsigned short port =inet_sk (child)->inet_num;struct inet_bind_hashbucket *head, *head2;bool created_inet_bind_bucket = false;
@@ -629,7 +629,7 @@ static bool inet_ehash_lookup_by_sk(struct sock *sk,*/ bool inet_ehash_insert (struct sock *sk,struct sock *osk,bool *found_dup_sk) { -struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; +struct inet_hashinfo *hashinfo =tcp_or_dccp_get_hashinfo (sk);struct inet_ehash_bucket *head;struct hlist_nulls_head *list;spinlock_t *lock;
@@ -701,7 +701,7 @@ static int inet_reuseport_add_sock(struct sock *sk,int __inet_hash (struct sock *sk,struct sock *osk) { -struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; +struct inet_hashinfo *hashinfo =tcp_or_dccp_get_hashinfo (sk);struct inet_listen_hashbucket *ilb2;int err =0 ;
@@ -747,7 +747,7 @@ EXPORT_SYMBOL_GPL(inet_hash);void inet_unhash (struct sock *sk) { -struct inet_hashinfo *hashinfo = sk->sk_prot->h.hashinfo; +struct inet_hashinfo *hashinfo =tcp_or_dccp_get_hashinfo (sk);if (sk_unhashed (sk))return ;
@@ -834,7 +834,7 @@ inet_bind2_bucket_find(const struct inet_bind_hashbucket *head, const struct netstruct inet_bind_hashbucket *inet_bhash2_addr_any_hashbucket (const struct sock *sk,const struct net *net,int port) { -struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo; +struct inet_hashinfo *hinfo =tcp_or_dccp_get_hashinfo (sk);u32 hash;#if IS_ENABLED (CONFIG_IPV6 )struct in6_addr addr_any = {};
@@ -850,7 +850,7 @@ inet_bhash2_addr_any_hashbucket(const struct sock *sk, const struct net *net, inint inet_bhash2_update_saddr (struct inet_bind_hashbucket *prev_saddr,struct sock *sk) { -struct inet_hashinfo *hinfo = sk->sk_prot->h.hashinfo; +struct inet_hashinfo *hinfo =tcp_or_dccp_get_hashinfo (sk);struct inet_bind2_bucket *tb2, *new_tb2;int l3mdev =inet_sk_bound_l3mdev (sk);struct inet_bind_hashbucket *head2;
net/ipv4/tcp_ipv4.c +1 -1
@@ -3083,7 +3083,7 @@ struct proto tcp_prot = { .slab_flags = SLAB_TYPESAFE_BY_RCU, .twsk_prot = &tcp_timewait_sock_ops, .rsk_prot = &tcp_request_sock_ops, - .h.hashinfo = &tcp_hashinfo, + .h.hashinfo = NULL, .no_autobind = true, .diag_destroy = tcp_abort, };
net/ipv6/tcp_ipv6.c +1 -1
@@ -2193,7 +2193,7 @@ struct proto tcpv6_prot = { .slab_flags = SLAB_TYPESAFE_BY_RCU, .twsk_prot = &tcp6_timewait_sock_ops, .rsk_prot = &tcp6_request_sock_ops, - .h.hashinfo = &tcp_hashinfo, + .h.hashinfo = NULL, .no_autobind = true, .diag_destroy = tcp_abort, };
Signature · 1 line
2.30.2
PATCH v3 3/5 tcp: Access &tcp_hashinfo via net.
In reply to Kuniyuki IwashimaWe will soon introduce an optional per-netns ehash.
This means we cannot use tcp_hashinfo directly in most places.
Instead, access it via net->ipv4.tcp_death_row->hashinfo.
The access will be valid only while initialising tcp_hashinfo itself and creating/destroying each netns.
--- .../chelsio/inline_crypto/chtls/chtls_cm.c | 5 +- net/core/filter.c | 5 +- net/ipv4/esp4.c | 3 +- net/ipv4/inet_hashtables.c | 2 +- net/ipv4/netfilter/nf_socket_ipv4.c | 2 +- net/ipv4/netfilter/nf_tproxy_ipv4.c | 17 +++-- net/ipv4/tcp_diag.c | 18 +++++- net/ipv4/tcp_ipv4.c | 63 +++++++++++-------- net/ipv4/tcp_minisocks.c | 2 +- net/ipv6/esp6.c | 3 +- net/ipv6/inet6_hashtables.c | 4 +- net/ipv6/netfilter/nf_socket_ipv6.c | 2 +- net/ipv6/netfilter/nf_tproxy_ipv6.c | 5 +- net/ipv6/tcp_ipv6.c | 16 ++--- net/mptcp/mptcp_diag.c | 7 ++- 15 files changed, 92 insertions(+), 62 deletions(-)
drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c +2 -3
@@ -1069,8 +1069,7 @@ static void chtls_pass_accept_rpl(struct sk_buff *skb,cxgb4_l2t_send (csk->egress_dev, skb, csk->l2t_entry); } -static void inet_inherit_port (struct inet_hashinfo *hash_info, -struct sock *lsk,struct sock *newsk) +static void inet_inherit_port (struct sock *lsk,struct sock *newsk) {local_bh_disable ();__inet_inherit_port (lsk, newsk);
@@ -1240,7 +1239,7 @@ static struct sock *chtls_recv_sock(struct sock *lsk, ipv4.sysctl_tcp_window_scaling), tp->window_clamp);neigh_release (n); -inet_inherit_port (&tcp_hashinfo, lsk, newsk); +inet_inherit_port (lsk, newsk);csk_set_flag (csk,CSK_CONN_INLINE );bh_unlock_sock (newsk);/* tcp_create_openreq_child ->sk_clone_lock */
net/core/filter.c +3 -2
@@ -6469,6 +6469,7 @@ static const struct bpf_func_proto bpf_lwt_seg6_adjust_srh_proto = {static struct sock *sk_lookup (struct net *net,struct bpf_sock_tuple *tuple,int dif,int sdif,u8 family,u8 proto) { +struct inet_hashinfo *hinfo = net->ipv4.tcp_death_row->hashinfo;bool refcounted = false;struct sock *sk =NULL ;
@@ -6477,7 +6478,7 @@ static struct sock *sk_lookup(struct net *net, struct bpf_sock_tuple *tuple,__be32 dst4 = tuple->ipv4.daddr;if (proto ==IPPROTO_TCP ) - sk =__inet_lookup (net, &tcp_hashinfo,NULL ,0 , + sk =__inet_lookup (net, hinfo,NULL ,0 , src4, tuple->ipv4.sport, dst4, tuple->ipv4.dport, dif, sdif, &refcounted);
@@ -6491,7 +6492,7 @@ static struct sock *sk_lookup(struct net *net, struct bpf_sock_tuple *tuple,struct in6_addr *dst6 = (struct in6_addr *)&tuple->ipv6.daddr;if (proto ==IPPROTO_TCP ) - sk =__inet6_lookup (net, &tcp_hashinfo,NULL ,0 , + sk =__inet6_lookup (net, hinfo,NULL ,0 , src6, tuple->ipv6.sport, dst6,ntohs (tuple->ipv6.dport), dif, sdif, &refcounted);
net/ipv4/esp4.c +2 -1
@@ -134,6 +134,7 @@ static void esp_free_tcp_sk(struct rcu_head *head)static struct sock *esp_find_tcp_sk (struct xfrm_state *x) {struct xfrm_encap_tmpl *encap = x->encap; +struct net *net =xs_net (x);struct esp_tcp_sk *esk;__be16 sport, dport;struct sock *nsk;
@@ -160,7 +161,7 @@ static struct sock *esp_find_tcp_sk(struct xfrm_state *x) }spin_unlock_bh (&x->lock); - sk =inet_lookup_established (xs_net (x), &tcp_hashinfo, x->id.daddr.a4, + sk =inet_lookup_established (net, net->ipv4.tcp_death_row->hashinfo, x->id.daddr.a4, dport, x->props.saddr.a4, sport,0 );if (!sk)return ERR_PTR (-ENOENT );
net/ipv4/inet_hashtables.c +1 -1
@@ -386,7 +386,7 @@ static inline struct sock *inet_lookup_run_bpf(struct net *net,struct sock *sk, *reuse_sk;bool no_reuseport; -if (hashinfo != &tcp_hashinfo) +if (hashinfo != net->ipv4.tcp_death_row->hashinfo)return NULL ;/* only TCP is supported */ no_reuseport = bpf_sk_lookup_run_v4(net, IPPROTO_TCP, saddr, sport,
net/ipv4/netfilter/nf_socket_ipv4.c +1 -1
@@ -71,7 +71,7 @@ nf_socket_get_sock_v4(struct net *net, struct sk_buff *skb, const int doff, {switch (protocol) {case IPPROTO_TCP : -return inet_lookup (net, &tcp_hashinfo, skb, doff, +return inet_lookup (net, net->ipv4.tcp_death_row->hashinfo, skb, doff, saddr, sport, daddr, dport, in->ifindex);case IPPROTO_UDP :
net/ipv4/netfilter/nf_tproxy_ipv4.c +8 -9
@@ -79,6 +79,7 @@ nf_tproxy_get_sock_v4(struct net *net, struct sk_buff *skb,const struct net_device *in, const enum nf_tproxy_lookup_t lookup_type) { + struct inet_hashinfo *hinfo = net->ipv4.tcp_death_row->hashinfo;struct sock *sk;switch (protocol) {
@@ -92,12 +93,10 @@ nf_tproxy_get_sock_v4(struct net *net, struct sk_buff *skb,switch (lookup_type) {case NF_TPROXY_LOOKUP_LISTENER : - sk =inet_lookup_listener (net, &tcp_hashinfo, skb, -ip_hdrlen (skb) + -__tcp_hdrlen (hp), - saddr, sport, - daddr, dport, - in->ifindex,0 ); + sk =inet_lookup_listener (net, hinfo, skb, +ip_hdrlen (skb) +__tcp_hdrlen (hp), + saddr, sport, daddr, dport, + in->ifindex,0 );if (sk && !refcount_inc_not_zero (&sk->sk_refcnt)) sk =NULL ;
@@ -108,9 +107,9 @@ nf_tproxy_get_sock_v4(struct net *net, struct sk_buff *skb,*/ break ;case NF_TPROXY_LOOKUP_ESTABLISHED : - sk =inet_lookup_established (net, &tcp_hashinfo, - saddr, sport, daddr, dport, - in->ifindex); + sk =inet_lookup_established (net, hinfo, + saddr, sport, daddr, dport, + in->ifindex);break ;default :BUG ();
net/ipv4/tcp_diag.c +15 -3
@@ -181,13 +181,21 @@ static size_t tcp_diag_get_aux_size(struct sock *sk, bool net_admin)static void tcp_diag_dump (struct sk_buff *skb,struct netlink_callback *cb,const struct inet_diag_req_v2 *r) { -inet_diag_dump_icsk (&tcp_hashinfo, skb, cb, r); +struct inet_hashinfo *hinfo; + + hinfo =sock_net (cb->skb->sk)->ipv4.tcp_death_row->hashinfo; + +inet_diag_dump_icsk (hinfo, skb, cb, r); }static int tcp_diag_dump_one (struct netlink_callback *cb,const struct inet_diag_req_v2 *req) { -return inet_diag_dump_one_icsk (&tcp_hashinfo, cb, req); +struct inet_hashinfo *hinfo; + + hinfo =sock_net (cb->skb->sk)->ipv4.tcp_death_row->hashinfo; + +return inet_diag_dump_one_icsk (hinfo, cb, req); }#ifdef CONFIG_INET_DIAG_DESTROY
@@ -195,9 +203,13 @@ static int tcp_diag_destroy(struct sk_buff *in_skb, const struct inet_diag_req_v2 *req) {struct net *net =sock_net (in_skb->sk); -struct sock *sk =inet_diag_find_one_icsk (net, &tcp_hashinfo, req); +struct inet_hashinfo *hinfo; +struct sock *sk;int err; + hinfo = net->ipv4.tcp_death_row->hashinfo; + sk =inet_diag_find_one_icsk (net, hinfo, req); +if (IS_ERR (sk))return PTR_ERR (sk);
net/ipv4/tcp_ipv4.c +38 -25
233 lines not on this page · raw message
net/ipv4/tcp_minisocks.c +1 -1
@@ -319,7 +319,7 @@ void tcp_time_wait(struct sock *sk, int state, int timeo)/* Linkage updates. * Note that access to tw after this point is illegal. */ -inet_twsk_hashdance (tw, sk, &tcp_hashinfo); +inet_twsk_hashdance (tw, sk, tcp_death_row->hashinfo);local_bh_enable (); }else {/* Sorry, if we're out of memory, just CLOSE this
net/ipv6/esp6.c +2 -1
@@ -151,6 +151,7 @@ static void esp_free_tcp_sk(struct rcu_head *head)static struct sock *esp6_find_tcp_sk (struct xfrm_state *x) {struct xfrm_encap_tmpl *encap = x->encap; +struct net *net =xs_net (x);struct esp_tcp_sk *esk;__be16 sport, dport;struct sock *nsk;
@@ -177,7 +178,7 @@ static struct sock *esp6_find_tcp_sk(struct xfrm_state *x) }spin_unlock_bh (&x->lock); - sk =__inet6_lookup_established (xs_net (x), &tcp_hashinfo, &x->id.daddr.in6, + sk =__inet6_lookup_established (net, net->ipv4.tcp_death_row->hashinfo, &x->id.daddr.in6, dport, &x->props.saddr.in6,ntohs (sport),0 ,0 );if (!sk)return ERR_PTR (-ENOENT );
net/ipv6/inet6_hashtables.c +1 -3
@@ -21,8 +21,6 @@#include <net/ip.h> #include <net/sock_reuseport.h> -extern struct inet_hashinfo tcp_hashinfo; -u32 inet6_ehashfn (const struct net *net,const struct in6_addr *laddr,const u16 lport,const struct in6_addr *faddr,const __be16 fport)
@@ -169,7 +167,7 @@ static inline struct sock *inet6_lookup_run_bpf(struct net *net,struct sock *sk, *reuse_sk;bool no_reuseport; -if (hashinfo != &tcp_hashinfo) +if (hashinfo != net->ipv4.tcp_death_row->hashinfo)return NULL ;/* only TCP is supported */ no_reuseport = bpf_sk_lookup_run_v6(net, IPPROTO_TCP, saddr, sport,
net/ipv6/netfilter/nf_socket_ipv6.c +1 -1
@@ -83,7 +83,7 @@ nf_socket_get_sock_v6(struct net *net, struct sk_buff *skb, int doff, {switch (protocol) {case IPPROTO_TCP : -return inet6_lookup (net, &tcp_hashinfo, skb, doff, +return inet6_lookup (net, net->ipv4.tcp_death_row->hashinfo, skb, doff, saddr, sport, daddr, dport, in->ifindex);case IPPROTO_UDP :
net/ipv6/netfilter/nf_tproxy_ipv6.c +3 -2
@@ -80,6 +80,7 @@ nf_tproxy_get_sock_v6(struct net *net, struct sk_buff *skb, int thoff,const struct net_device *in, const enum nf_tproxy_lookup_t lookup_type) { + struct inet_hashinfo *hinfo = net->ipv4.tcp_death_row->hashinfo;struct sock *sk;switch (protocol) {
@@ -93,7 +94,7 @@ nf_tproxy_get_sock_v6(struct net *net, struct sk_buff *skb, int thoff,switch (lookup_type) {case NF_TPROXY_LOOKUP_LISTENER : - sk = inet6_lookup_listener(net, &tcp_hashinfo, skb, + sk = inet6_lookup_listener(net, hinfo, skb, thoff + __tcp_hdrlen(hp), saddr, sport, daddr, ntohs(dport),
@@ -108,7 +109,7 @@ nf_tproxy_get_sock_v6(struct net *net, struct sk_buff *skb, int thoff,*/ break ;case NF_TPROXY_LOOKUP_ESTABLISHED : - sk =__inet6_lookup_established (net, &tcp_hashinfo, + sk =__inet6_lookup_established (net, hinfo, saddr, sport, daddr,ntohs (dport), in->ifindex,0 );break ;
net/ipv6/tcp_ipv6.c +9 -7
@@ -286,12 +286,14 @@ static int tcp_v6_connect(struct sock *sk, struct sockaddr *uaddr, goto failure; } + tcp_death_row =sock_net (sk)->ipv4.tcp_death_row; +if (!saddr) {struct inet_bind_hashbucket *prev_addr_hashbucket =NULL ;struct in6_addr prev_v6_rcv_saddr;if (icsk->icsk_bind2_hash) { - prev_addr_hashbucket =inet_bhashfn_portaddr (&tcp_hashinfo, + prev_addr_hashbucket =inet_bhashfn_portaddr (tcp_death_row->hashinfo, sk,sock_net (sk), inet->inet_num); prev_v6_rcv_saddr = sk->sk_v6_rcv_saddr;
@@ -325,7 +327,6 @@ static int tcp_v6_connect(struct sock *sk, struct sockaddr *uaddr, inet->inet_dport = usin->sin6_port;tcp_set_state (sk,TCP_SYN_SENT ); - tcp_death_row =sock_net (sk)->ipv4.tcp_death_row; err =inet6_hash_connect (tcp_death_row, sk);if (err) goto late_failure;
@@ -403,7 +404,7 @@ static int tcp_v6_err(struct sk_buff *skb, struct inet6_skb_parm *opt,bool fatal;int err; - sk =__inet6_lookup_established (net, &tcp_hashinfo, + sk =__inet6_lookup_established (net, net->ipv4.tcp_death_row->hashinfo, &hdr->daddr, th->dest, &hdr->saddr,ntohs (th->source), skb->dev->ifindex,inet6_sdif (skb));
@@ -1037,7 +1038,7 @@ static void tcp_v6_send_reset(const struct sock *sk, struct sk_buff *skb)* no RST generated if md5 hash doesn't match. */ sk1 =inet6_lookup_listener (net, - &tcp_hashinfo,NULL ,0 , + net->ipv4.tcp_death_row->hashinfo,NULL ,0 , &ipv6h->saddr, th->source, &ipv6h->daddr,ntohs (th->source), dif, sdif);
@@ -1636,7 +1637,7 @@ INDIRECT_CALLABLE_SCOPE int tcp_v6_rcv(struct sk_buff *skb) hdr =ipv6_hdr (skb); lookup: - sk =__inet6_lookup_skb (&tcp_hashinfo, skb,__tcp_hdrlen (th), + sk =__inet6_lookup_skb (net->ipv4.tcp_death_row->hashinfo, skb,__tcp_hdrlen (th), th->source, th->dest,inet6_iif (skb), sdif, &refcounted);if (!sk)
@@ -1811,7 +1812,7 @@ INDIRECT_CALLABLE_SCOPE int tcp_v6_rcv(struct sk_buff *skb) {struct sock *sk2; - sk2 = inet6_lookup_listener(dev_net(skb->dev), &tcp_hashinfo, + sk2 = inet6_lookup_listener(net, net->ipv4.tcp_death_row->hashinfo, skb, __tcp_hdrlen(th), &ipv6_hdr(skb)->saddr, th->source, &ipv6_hdr(skb)->daddr,
@@ -1844,6 +1845,7 @@ INDIRECT_CALLABLE_SCOPE int tcp_v6_rcv(struct sk_buff *skb)void tcp_v6_early_demux (struct sk_buff *skb) { +struct net *net =dev_net (skb->dev);const struct ipv6hdr *hdr;const struct tcphdr *th;struct sock *sk;
@@ -1861,7 +1863,7 @@ void tcp_v6_early_demux(struct sk_buff *skb)return ;/* Note : We use inet6_iif() here, not tcp_v6_iif() */ - sk =__inet6_lookup_established (dev_net (skb->dev), &tcp_hashinfo, + sk =__inet6_lookup_established (net, net->ipv4.tcp_death_row->hashinfo, &hdr->saddr, th->source, &hdr->daddr,ntohs (th->dest),inet6_iif (skb),inet6_sdif (skb));
net/mptcp/mptcp_diag.c +5 -2
@@ -81,15 +81,18 @@ static void mptcp_diag_dump_listeners(struct sk_buff *skb, struct netlink_callbastruct mptcp_diag_ctx *diag_ctx = (void *)cb->ctx;struct nlattr *bc = cb_data->inet_diag_nla_bc;struct net *net =sock_net (skb->sk); +struct inet_hashinfo *hinfo;int i; -for (i = diag_ctx->l_slot; i <= tcp_hashinfo.lhash2_mask; i++) { + hinfo = net->ipv4.tcp_death_row->hashinfo; + +for (i = diag_ctx->l_slot; i <= hinfo->lhash2_mask; i++) {struct inet_listen_hashbucket *ilb;struct hlist_nulls_node *node;struct sock *sk;int num =0 ; - ilb = &tcp_hashinfo.lhash2[i]; + ilb = &hinfo->lhash2[i];rcu_read_lock ();spin_lock (&ilb->lock);
Signature · 1 line
2.30.2
Paolo Abeni
maintainerRe: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
Quoted 43 lines · Kuniyuki Iwashima
We will soon introduce an optional per-netns ehash.
This means we cannot use
tcp_hashinfodirectly in most places.Instead, access it via
net->ipv4.tcp_death_row->hashinfo.The access will be valid only while initialising
tcp_hashinfoitself and creating/destroying each netns.Signed-off-by Kuniyuki Iwashima <[email protected]>--- .../chelsio/inline_crypto/chtls/chtls_cm.c | 5 +- net/core/filter.c | 5 +- net/ipv4/esp4.c | 3 +- net/ipv4/inet_hashtables.c | 2 +- net/ipv4/netfilter/nf_socket_ipv4.c | 2 +- net/ipv4/netfilter/nf_tproxy_ipv4.c | 17 +++-- net/ipv4/tcp_diag.c | 18 +++++- net/ipv4/tcp_ipv4.c | 63 +++++++++++-------- net/ipv4/tcp_minisocks.c | 2 +- net/ipv6/esp6.c | 3 +- net/ipv6/inet6_hashtables.c | 4 +- net/ipv6/netfilter/nf_socket_ipv6.c | 2 +- net/ipv6/netfilter/nf_tproxy_ipv6.c | 5 +- net/ipv6/tcp_ipv6.c | 16 ++--- net/mptcp/mptcp_diag.c | 7 ++- 15 files changed, 92 insertions(+), 62 deletions(-)drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c +1
-2@@ -1069,8 +1069,7 @@ static void chtls_pass_accept_rpl(struct sk_buff *skb,cxgb4_l2t_send (csk->egress_dev, skb, csk->l2t_entry); } -static void inet_inherit_port (struct inet_hashinfo *hash_info, -struct sock *lsk,struct sock *newsk) +static void inet_inherit_port (struct sock *lsk,struct sock *newsk) {local_bh_disable ();__inet_inherit_port (lsk, newsk);
drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c static struct sock *chtls_recv_sock(struct sock *lsk, · line 1244
Quoted 7 lines
drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c +1
-1@@ -1240,7 +1239,7 @@ static struct sock *chtls_recv_sock(struct sock *lsk, ipv4.sysctl_tcp_window_scaling), tp->window_clamp);neigh_release (n); -inet_inherit_port (&tcp_hashinfo, lsk, newsk); +inet_inherit_port (lsk, newsk);csk_set_flag (csk,CSK_CONN_INLINE );bh_unlock_sock (newsk);/* tcp_create_openreq_child ->sk_clone_lock */
I looks to me that the above chunks are functionally a no-op and I think that omitting the 2 drivers from the v2:
https://lore.kernel.org/netdev/[email protected]/
should break mlx5/nfp inside a netns. I don't understand why including the above and skipping the latters?!? I guess is a question mostly for Eric :)
Quoted 9 lines
drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c +1
-0@@ -1728,6 +1728,7 @@ EXPORT_SYMBOL(tcp_v4_do_rcv);int tcp_v4_early_demux (struct sk_buff *skb) { +struct net *net =dev_net (skb->dev);const struct iphdr *iph;const struct tcphdr *th;struct sock *sk;
drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c int tcp_v4_early_demux(struct sk_buff *skb) · line 1751
Quoted 8 lines
drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c +1
-1@@ -1744,7 +1745,7 @@ int tcp_v4_early_demux(struct sk_buff *skb)if (th->doff <sizeof (struct tcphdr ) /4 )return 0 ; - sk =__inet_lookup_established (dev_net (skb->dev), &tcp_hashinfo, + sk =__inet_lookup_established (net, net->ipv4.tcp_death_row->hashinfo, iph->saddr, th->source, iph->daddr,ntohs (th->dest), skb->skb_iif,inet_sdif (skb));
/Me is thinking aloud...
I'm wondering if the above has some measurable negative effect for large deployments using only the main netns?
Specifically, are net->ipv4.tcp_death_row and net->ipv4.tcp_death_row-
Quoted 1 line
hashinfo already into the working set data for established socket?
Would the above increase the WSS by 2 cache-lines?
Thanks!
Paolo
Re: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
From: Paolo Abeni <[email protected]> Date: Thu, 01 Sep 2022 12:57:33 +0200
Quoted 55 lines · Kuniyuki Iwashima
Quoted 43 lines
We will soon introduce an optional per-netns ehash.
This means we cannot use
tcp_hashinfodirectly in most places.Instead, access it via
net->ipv4.tcp_death_row->hashinfo.The access will be valid only while initialising
tcp_hashinfoitself and creating/destroying each netns.Signed-off-by Kuniyuki Iwashima <[email protected]>--- .../chelsio/inline_crypto/chtls/chtls_cm.c | 5 +- net/core/filter.c | 5 +- net/ipv4/esp4.c | 3 +- net/ipv4/inet_hashtables.c | 2 +- net/ipv4/netfilter/nf_socket_ipv4.c | 2 +- net/ipv4/netfilter/nf_tproxy_ipv4.c | 17 +++-- net/ipv4/tcp_diag.c | 18 +++++- net/ipv4/tcp_ipv4.c | 63 +++++++++++-------- net/ipv4/tcp_minisocks.c | 2 +- net/ipv6/esp6.c | 3 +- net/ipv6/inet6_hashtables.c | 4 +- net/ipv6/netfilter/nf_socket_ipv6.c | 2 +- net/ipv6/netfilter/nf_tproxy_ipv6.c | 5 +- net/ipv6/tcp_ipv6.c | 16 ++--- net/mptcp/mptcp_diag.c | 7 ++- 15 files changed, 92 insertions(+), 62 deletions(-)drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c +1
-2@@ -1069,8 +1069,7 @@ static void chtls_pass_accept_rpl(struct sk_buff *skb,cxgb4_l2t_send (csk->egress_dev, skb, csk->l2t_entry); } -static void inet_inherit_port (struct inet_hashinfo *hash_info, -struct sock *lsk,struct sock *newsk) +static void inet_inherit_port (struct sock *lsk,struct sock *newsk) {local_bh_disable ();__inet_inherit_port (lsk, newsk);drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c static struct sock *chtls_recv_sock(struct sock *lsk, · line 1244
Quoted 7 lines
drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c +1
-1@@ -1240,7 +1239,7 @@ static struct sock *chtls_recv_sock(struct sock *lsk, ipv4.sysctl_tcp_window_scaling), tp->window_clamp);neigh_release (n); -inet_inherit_port (&tcp_hashinfo, lsk, newsk); +inet_inherit_port (lsk, newsk);csk_set_flag (csk,CSK_CONN_INLINE );bh_unlock_sock (newsk);/* tcp_create_openreq_child ->sk_clone_lock */ I looks to me that the above chunks are functionally a no-op and I think that omitting the 2 drivers from the v2:
Quoted 5 lines
https://lore.kernel.org/netdev/[email protected]/
should break mlx5/nfp inside a netns. I don't understand why including the above and skipping the latters?!? I guess is a question mostly for Eric :)
My best guess is that it's ok unless it does not touch TCP stack deeply and if it does, the driver developer must catch up with the core changes not to burden maintainers...?
If so, I understand that take. OTOH, I also don't want to break anything when we know the change would do.
So, I'm fine to either stay as is or add the change in v4 again.
Quoted 21 lines
Quoted 10 lines
drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c +1
-0@@ -1728,6 +1728,7 @@ EXPORT_SYMBOL(tcp_v4_do_rcv);int tcp_v4_early_demux (struct sk_buff *skb) { +struct net *net =dev_net (skb->dev);const struct iphdr *iph;const struct tcphdr *th;struct sock *sk;drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c int tcp_v4_early_demux(struct sk_buff *skb) · line 1751
Quoted 8 lines
drivers/net/ethernet/chelsio/inline_crypto/chtls/chtls_cm.c +1
-1@@ -1744,7 +1745,7 @@ int tcp_v4_early_demux(struct sk_buff *skb)if (th->doff <sizeof (struct tcphdr ) /4 )return 0 ; - sk =__inet_lookup_established (dev_net (skb->dev), &tcp_hashinfo, + sk =__inet_lookup_established (net, net->ipv4.tcp_death_row->hashinfo, iph->saddr, th->source, iph->daddr,ntohs (th->dest), skb->skb_iif,inet_sdif (skb));/Me is thinking aloud...
Quoted 6 lines
I'm wondering if the above has some measurable negative effect for large deployments using only the main netns?
Specifically, are
net->ipv4.tcp_death_rowand net->ipv4.tcp_death_row-Quoted 1 line
hashinfo already into the working set data for established socket?
Would the above increase the WSS by 2 cache-lines?
Currently, the death_row and hashinfo are touched around tw sockets or connect(). If connections on the deployment are short-lived or frequently initiated by itself, that would be host and included in WSS.
If the workload is server and there's no active-close() socket or connections are long-lived, then it might not be included in WSS. But I think it's not likely than the former if the deployment is large enough.
If this change had large impact, then we could revert fbb8295248e1 which converted net->ipv4.tcp_death_row into pointer for 0dad4087a86a that tried to fire a TW timer after netns is freed, but 0dad4087a86a has already reverted.
Quoted 4 lines
Thanks!
Paolo
Re: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
Quoted 21 lines · Kuniyuki Iwashima
From: Paolo Abeni <[email protected]>
Quoted 8 lines
/Me is thinking aloud...
I'm wondering if the above has some measurable negative effect for large deployments using only the main netns?
Specifically, are
net->ipv4.tcp_death_rowand net->ipv4.tcp_death_row-Quoted 1 line
hashinfo already into the working set data for established socket?
Would the above increase the WSS by 2 cache-lines?
Currently, the
death_rowand hashinfo are touched around tw sockets orconnect(). If connections on the deployment are short-lived or frequently initiated by itself, that would be host and included in WSS.If the workload is server and there's no active-close() socket or connections are long-lived, then it might not be included in WSS. But I think it's not likely than the former if the deployment is large enough.
Quoted 4 lines
If this change had large impact, then we could revert
fbb8295248e1which convertednet->ipv4.tcp_death_rowinto pointer for0dad4087a86athat tried to fire a TW timer after netns is freed, but0dad4087a86ahas already reverted.
Concern was fast path.
Each incoming packet does a socket lookup.
Fetching hashinfo (instead of &tcp_hashinfo) with a dereference of a field in 'struct net' might inccurr a new cache line miss.
Previously, first cache line of tcp_info was enough to bring a lot of fields in cpu cache.
Re: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
From: Eric Dumazet <[email protected]> Date: Thu, 1 Sep 2022 14:30:43 -0700
Quoted 30 lines · Kuniyuki Iwashima
Quoted 21 lines
From: Paolo Abeni <[email protected]>
Quoted 8 lines
/Me is thinking aloud...
I'm wondering if the above has some measurable negative effect for large deployments using only the main netns?
Specifically, are
net->ipv4.tcp_death_rowand net->ipv4.tcp_death_row-Quoted 1 line
hashinfo already into the working set data for established socket?
Would the above increase the WSS by 2 cache-lines?
Currently, the
death_rowand hashinfo are touched around tw sockets orconnect(). If connections on the deployment are short-lived or frequently initiated by itself, that would be host and included in WSS.If the workload is server and there's no active-close() socket or connections are long-lived, then it might not be included in WSS. But I think it's not likely than the former if the deployment is large enough.
Quoted 4 lines
If this change had large impact, then we could revert
fbb8295248e1which convertednet->ipv4.tcp_death_rowinto pointer for0dad4087a86athat tried to fire a TW timer after netns is freed, but0dad4087a86ahas already reverted.Concern was fast path.
Quoted 7 lines
Each incoming packet does a socket lookup.
Fetching hashinfo (instead of
&tcp_hashinfo) with a dereference of a field in 'struct net' might inccurr a new cache line miss.Previously, first cache line of
tcp_infowas enough to bring a lot of fields in cpu cache.
Ok, let me test on that if there could be regressions.
Re: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
From: Kuniyuki Iwashima <[email protected]> Date: Thu, 1 Sep 2022 15:12:16 -0700
Quoted 34 lines
From: Eric Dumazet <[email protected]> Date: Thu, 1 Sep 2022 14:30:43 -0700Quoted 32 lines · Kuniyuki Iwashima
Quoted 21 lines
From: Paolo Abeni <[email protected]>
Quoted 8 lines
/Me is thinking aloud...
I'm wondering if the above has some measurable negative effect for large deployments using only the main netns?
Specifically, are
net->ipv4.tcp_death_rowand net->ipv4.tcp_death_row-Quoted 1 line
hashinfo already into the working set data for established socket?
Would the above increase the WSS by 2 cache-lines?
Currently, the
death_rowand hashinfo are touched around tw sockets orconnect(). If connections on the deployment are short-lived or frequently initiated by itself, that would be host and included in WSS.If the workload is server and there's no active-close() socket or connections are long-lived, then it might not be included in WSS. But I think it's not likely than the former if the deployment is large enough.
Quoted 4 lines
If this change had large impact, then we could revert
fbb8295248e1which convertednet->ipv4.tcp_death_rowinto pointer for0dad4087a86athat tried to fire a TW timer after netns is freed, but0dad4087a86ahas already reverted.Concern was fast path.
Each incoming packet does a socket lookup.
Quoted 7 lines
Quoted 5 lines
Fetching hashinfo (instead of
&tcp_hashinfo) with a dereference of a field in 'struct net' might inccurr a new cache line miss.Previously, first cache line of
tcp_infowas enough to bring a lot of fields in cpu cache.Ok, let me test on that if there could be regressions.
I tested tcp_hashinfo vs tcp_death_row->hashinfo with super_netperf and collected HW cache-related metrics with perf.
After the patch the number of L1 miss seems to increase, but the instructions per cycle also increases, and cache miss rate did not change. Also, there was not performance regression for netperf.
Tested:
# cat perf_super_netperf echo 0 > /proc/sys/kernel/nmi_watchdog echo 3 > /proc/sys/vm/drop_caches
perf stat -a \
-e cycles,instructions,cache-references,cache-misses,bus-cycles \
-e L1-dcache-loads,L1-dcache-load-misses,L1-dcache-stores \
-e dTLB-loads,dTLB-load-misses \
-e LLC-loads,LLC-load-misses,LLC-stores \
./super_netperf $(($(nproc) * 2)) -H 10.0.0.142 -l 60 -fMecho 1 > /proc/sys/kernel/nmi_watchdog
Before:
# ./perf_super_netperf 2929.81
Performance counter stats for 'system wide':
494,002,600,338 cycles (23.07%)
241,230,662,890 instructions # 0.49 insn per cycle (30.76%)
6,303,603,008 cache-references (38.45%)
1,421,440,332 cache-misses # 22.550 % of all cache refs (46.15%)
4,861,179,308 bus-cycles (46.15%)
65,410,735,599 L1-dcache-loads (46.15%)
12,647,247,339 L1-dcache-load-misses # 19.34% of all L1-dcache accesses (30.77%)
32,912,656,369 L1-dcache-stores (30.77%)
66,015,779,361 dTLB-loads (30.77%)
81,293,994 dTLB-load-misses # 0.12% of all dTLB cache accesses (30.77%)
2,946,386,949 LLC-loads (30.77%)
257,223,942 LLC-load-misses # 8.73% of all LL-cache accesses (30.77%)
1,183,820,461 LLC-stores (15.38%)62.132250590 seconds time elapsed
After:
# ./perf_super_netperf 2930.17
Performance counter stats for 'system wide':
479,595,776,631 cycles (23.07%)
243,318,957,230 instructions # 0.51 insn per cycle (30.76%)
6,169,892,840 cache-references (38.46%)
1,381,992,694 cache-misses # 22.399 % of all cache refs (46.15%)
4,534,304,190 bus-cycles (46.16%)
66,059,178,377 L1-dcache-loads (46.17%)
12,759,529,139 L1-dcache-load-misses # 19.32% of all L1-dcache accesses (30.78%)
33,292,513,002 L1-dcache-stores (30.78%)
66,482,176,008 dTLB-loads (30.77%)
72,877,970 dTLB-load-misses # 0.11% of all dTLB cache accesses (30.76%)
2,984,881,101 LLC-loads (30.76%)
234,747,930 LLC-load-misses # 7.86% of all LL-cache accesses (30.76%)
1,165,606,022 LLC-stores (15.38%)62.110708964 seconds time elapsed
Re: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
Quoted 90 lines · Kuniyuki Iwashima
From: Kuniyuki Iwashima <[email protected]> Date: Thu, 1 Sep 2022 15:12:16 -0700Quoted 34 lines
From: Eric Dumazet <[email protected]> Date: Thu, 1 Sep 2022 14:30:43 -0700Quoted 32 lines · Kuniyuki Iwashima
Quoted 21 lines
From: Paolo Abeni <[email protected]>
Quoted 8 lines
/Me is thinking aloud...
I'm wondering if the above has some measurable negative effect for large deployments using only the main netns?
Specifically, are
net->ipv4.tcp_death_rowand net->ipv4.tcp_death_row-Quoted 1 line
hashinfo already into the working set data for established socket?
Would the above increase the WSS by 2 cache-lines?
Currently, the
death_rowand hashinfo are touched around tw sockets orconnect(). If connections on the deployment are short-lived or frequently initiated by itself, that would be host and included in WSS.If the workload is server and there's no active-close() socket or connections are long-lived, then it might not be included in WSS. But I think it's not likely than the former if the deployment is large enough.
Quoted 4 lines
If this change had large impact, then we could revert
fbb8295248e1which convertednet->ipv4.tcp_death_rowinto pointer for0dad4087a86athat tried to fire a TW timer after netns is freed, but0dad4087a86ahas already reverted.Concern was fast path.
Each incoming packet does a socket lookup.
Quoted 7 lines
Quoted 5 lines
Fetching hashinfo (instead of
&tcp_hashinfo) with a dereference of a field in 'struct net' might inccurr a new cache line miss.Previously, first cache line of
tcp_infowas enough to bring a lot of fields in cpu cache.Ok, let me test on that if there could be regressions.
I tested
tcp_hashinfovstcp_death_row->hashinfowithsuper_netperfand collected HW cache-related metrics with perf.After the patch the number of L1 miss seems to increase, but the instructions per cycle also increases, and cache miss rate did not change. Also, there was not performance regression for netperf.
Tested:
# cat perf_super_netperf echo 0 > /proc/sys/kernel/nmi_watchdog echo 3 > /proc/sys/vm/drop_cachesperf stat -a \ -e cycles,instructions,cache-references,cache-misses,bus-cycles \ -e L1-dcache-loads,L1-dcache-load-misses,L1-dcache-stores \ -e dTLB-loads,dTLB-load-misses \ -e LLC-loads,LLC-load-misses,LLC-stores \ ./super_netperf $(($(nproc) * 2)) -H 10.0.0.142 -l 60 -fMecho 1 >
/proc/sys/kernel/nmi_watchdogBefore:
# ./perf_super_netperf 2929.81Performance counter stats for 'system wide':
494,002,600,338 cycles (23.07%) 241,230,662,890 instructions # 0.49 insn per cycle (30.76%) 6,303,603,008 cache-references (38.45%) 1,421,440,332 cache-misses # 22.550 % of all cache refs (46.15%) 4,861,179,308 bus-cycles (46.15%) 65,410,735,599 L1-dcache-loads (46.15%) 12,647,247,339 L1-dcache-load-misses # 19.34% of all L1-dcache accesses (30.77%) 32,912,656,369 L1-dcache-stores (30.77%) 66,015,779,361 dTLB-loads (30.77%) 81,293,994 dTLB-load-misses # 0.12% of all dTLB cache accesses (30.77%) 2,946,386,949 LLC-loads (30.77%) 257,223,942 LLC-load-misses # 8.73% of all LL-cache accesses (30.77%) 1,183,820,461 LLC-stores (15.38%)
Quoted 2 lines
62.132250590 seconds time elapsed
This test will not be able to see a difference really...
What is needed is to measure the latency when nothing at all is in the caches.
Vast majority of real world TCP traffic is light or moderate. Packets are received and cpu has to bring X cache lines into L1 in order to process one packet.
We slowly are increasing X over time :/
pahole is your friend, more than a stress-test.
Re: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
From: Eric Dumazet <[email protected]> Date: Fri, 2 Sep 2022 17:53:18 -0700
Quoted 98 lines · Kuniyuki Iwashima
Quoted 90 lines
From: Kuniyuki Iwashima <[email protected]> Date: Thu, 1 Sep 2022 15:12:16 -0700Quoted 34 lines
From: Eric Dumazet <[email protected]> Date: Thu, 1 Sep 2022 14:30:43 -0700Quoted 32 lines · Kuniyuki Iwashima
Quoted 21 lines
From: Paolo Abeni <[email protected]>
Quoted 8 lines
/Me is thinking aloud...
I'm wondering if the above has some measurable negative effect for large deployments using only the main netns?
Specifically, are
net->ipv4.tcp_death_rowand net->ipv4.tcp_death_row-Quoted 1 line
hashinfo already into the working set data for established socket?
Would the above increase the WSS by 2 cache-lines?
Currently, the
death_rowand hashinfo are touched around tw sockets orconnect(). If connections on the deployment are short-lived or frequently initiated by itself, that would be host and included in WSS.If the workload is server and there's no active-close() socket or connections are long-lived, then it might not be included in WSS. But I think it's not likely than the former if the deployment is large enough.
Quoted 4 lines
If this change had large impact, then we could revert
fbb8295248e1which convertednet->ipv4.tcp_death_rowinto pointer for0dad4087a86athat tried to fire a TW timer after netns is freed, but0dad4087a86ahas already reverted.Concern was fast path.
Each incoming packet does a socket lookup.
Quoted 7 lines
Quoted 5 lines
Fetching hashinfo (instead of
&tcp_hashinfo) with a dereference of a field in 'struct net' might inccurr a new cache line miss.Previously, first cache line of
tcp_infowas enough to bring a lot of fields in cpu cache.Ok, let me test on that if there could be regressions.
I tested
tcp_hashinfovstcp_death_row->hashinfowithsuper_netperfand collected HW cache-related metrics with perf.After the patch the number of L1 miss seems to increase, but the instructions per cycle also increases, and cache miss rate did not change. Also, there was not performance regression for netperf.
Tested:
# cat perf_super_netperf echo 0 > /proc/sys/kernel/nmi_watchdog echo 3 > /proc/sys/vm/drop_cachesperf stat -a \ -e cycles,instructions,cache-references,cache-misses,bus-cycles \ -e L1-dcache-loads,L1-dcache-load-misses,L1-dcache-stores \ -e dTLB-loads,dTLB-load-misses \ -e LLC-loads,LLC-load-misses,LLC-stores \ ./super_netperf $(($(nproc) * 2)) -H 10.0.0.142 -l 60 -fMecho 1 >
/proc/sys/kernel/nmi_watchdogBefore:
# ./perf_super_netperf 2929.81Performance counter stats for 'system wide':
494,002,600,338 cycles (23.07%) 241,230,662,890 instructions # 0.49 insn per cycle (30.76%) 6,303,603,008 cache-references (38.45%) 1,421,440,332 cache-misses # 22.550 % of all cache refs (46.15%) 4,861,179,308 bus-cycles (46.15%) 65,410,735,599 L1-dcache-loads (46.15%) 12,647,247,339 L1-dcache-load-misses # 19.34% of all L1-dcache accesses (30.77%) 32,912,656,369 L1-dcache-stores (30.77%) 66,015,779,361 dTLB-loads (30.77%) 81,293,994 dTLB-load-misses # 0.12% of all dTLB cache accesses (30.77%) 2,946,386,949 LLC-loads (30.77%) 257,223,942 LLC-load-misses # 8.73% of all LL-cache accesses (30.77%) 1,183,820,461 LLC-stores (15.38%)Quoted 2 lines
62.132250590 seconds time elapsedThis test will not be able to see a difference really...
What is needed is to measure the latency when nothing at all is in the caches.
Quoted 7 lines
Vast majority of real world TCP traffic is light or moderate. Packets are received and cpu has to bring X cache lines into L1 in order to process one packet.
We slowly are increasing X over time :/
pahole is your friend, more than a stress-test.
Here's pahole result on my local build. As Paolo said, we need 2 cachelines for tcp_death_row and the hashinfo?
How about moving hashinfo as the first member of struct inet_timewait_death_row and convert it to just struct instead of pointer so that we need 1 cache line to read hashinfo?
$ pahole -C netns_ipv4,inet_timewait_death_row vmlinux
struct netns_ipv4 {
struct inet_timewait_death_row * tcp_death_row; /* 0 8 */
struct ctl_table_header * forw_hdr; /* 8 8 */
struct ctl_table_header * frags_hdr; /* 16 8 */
struct ctl_table_header * ipv4_hdr; /* 24 8 */
struct ctl_table_header * route_hdr; /* 32 8 */
struct ctl_table_header * xfrm4_hdr; /* 40 8 */
struct ipv4_devconf * devconf_all; /* 48 8 */
struct ipv4_devconf * devconf_dflt; /* 56 8 */
/* --- cacheline 1 boundary (64 bytes) --- */
...
};
struct inet_timewait_death_row {
refcount_t tw_refcount; /* 0 4 *//* XXX 60 bytes hole, try to pack */ /* --- cacheline 1 boundary (64 bytes) --- */ struct inet_hashinfo * hashinfo __attribute__((__aligned__(64))); /* 64 8 */ int sysctl_max_tw_buckets; /* 72 4 */ /* size: 128, cachelines: 2, members: 3 */ /* sum members: 16, holes: 1, sum holes: 60 */ /* padding: 52 */ /* forced alignments: 1, forced holes: 1, sum forced holes: 60 */ } __attribute__((__aligned__(64)));
Re: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
From: Kuniyuki Iwashima <[email protected]> Date: Fri, 2 Sep 2022 18:12:43 -0700
Quoted 108 lines
From: Eric Dumazet <[email protected]> Date: Fri, 2 Sep 2022 17:53:18 -0700Quoted 98 lines · Kuniyuki Iwashima
Quoted 90 lines
From: Kuniyuki Iwashima <[email protected]> Date: Thu, 1 Sep 2022 15:12:16 -0700Quoted 34 lines
From: Eric Dumazet <[email protected]> Date: Thu, 1 Sep 2022 14:30:43 -0700Quoted 32 lines · Kuniyuki Iwashima
Quoted 21 lines
From: Paolo Abeni <[email protected]>
Quoted 8 lines
/Me is thinking aloud...
I'm wondering if the above has some measurable negative effect for large deployments using only the main netns?
Specifically, are
net->ipv4.tcp_death_rowand net->ipv4.tcp_death_row-Quoted 1 line
hashinfo already into the working set data for established socket?
Would the above increase the WSS by 2 cache-lines?
Currently, the
death_rowand hashinfo are touched around tw sockets orconnect(). If connections on the deployment are short-lived or frequently initiated by itself, that would be host and included in WSS.If the workload is server and there's no active-close() socket or connections are long-lived, then it might not be included in WSS. But I think it's not likely than the former if the deployment is large enough.
Quoted 4 lines
If this change had large impact, then we could revert
fbb8295248e1which convertednet->ipv4.tcp_death_rowinto pointer for0dad4087a86athat tried to fire a TW timer after netns is freed, but0dad4087a86ahas already reverted.Concern was fast path.
Each incoming packet does a socket lookup.
Quoted 7 lines
Quoted 5 lines
Fetching hashinfo (instead of
&tcp_hashinfo) with a dereference of a field in 'struct net' might inccurr a new cache line miss.Previously, first cache line of
tcp_infowas enough to bring a lot of fields in cpu cache.Ok, let me test on that if there could be regressions.
I tested
tcp_hashinfovstcp_death_row->hashinfowithsuper_netperfand collected HW cache-related metrics with perf.After the patch the number of L1 miss seems to increase, but the instructions per cycle also increases, and cache miss rate did not change. Also, there was not performance regression for netperf.
Tested:
# cat perf_super_netperf echo 0 > /proc/sys/kernel/nmi_watchdog echo 3 > /proc/sys/vm/drop_cachesperf stat -a \ -e cycles,instructions,cache-references,cache-misses,bus-cycles \ -e L1-dcache-loads,L1-dcache-load-misses,L1-dcache-stores \ -e dTLB-loads,dTLB-load-misses \ -e LLC-loads,LLC-load-misses,LLC-stores \ ./super_netperf $(($(nproc) * 2)) -H 10.0.0.142 -l 60 -fMecho 1 >
/proc/sys/kernel/nmi_watchdogBefore:
# ./perf_super_netperf 2929.81Performance counter stats for 'system wide':
494,002,600,338 cycles (23.07%) 241,230,662,890 instructions # 0.49 insn per cycle (30.76%) 6,303,603,008 cache-references (38.45%) 1,421,440,332 cache-misses # 22.550 % of all cache refs (46.15%) 4,861,179,308 bus-cycles (46.15%) 65,410,735,599 L1-dcache-loads (46.15%) 12,647,247,339 L1-dcache-load-misses # 19.34% of all L1-dcache accesses (30.77%) 32,912,656,369 L1-dcache-stores (30.77%) 66,015,779,361 dTLB-loads (30.77%) 81,293,994 dTLB-load-misses # 0.12% of all dTLB cache accesses (30.77%) 2,946,386,949 LLC-loads (30.77%) 257,223,942 LLC-load-misses # 8.73% of all LL-cache accesses (30.77%) 1,183,820,461 LLC-stores (15.38%)Quoted 2 lines
62.132250590 seconds time elapsedThis test will not be able to see a difference really...
What is needed is to measure the latency when nothing at all is in the caches.
Quoted 7 lines
Vast majority of real world TCP traffic is light or moderate. Packets are received and cpu has to bring X cache lines into L1 in order to process one packet.
We slowly are increasing X over time :/
pahole is your friend, more than a stress-test.
Quoted 7 lines
Here's pahole result on my local build. As Paolo said, we need 2 cachelines for
tcp_death_rowand the hashinfo?How about moving hashinfo as the first member of struct
inet_timewait_death_rowand convert it to just struct instead of pointer so that we need 1 cache line to read hashinfo?
Like this.
$ pahole -EC netns_ipv4 vmlinux
struct netns_ipv4 {
struct inet_timewait_death_row {
struct inet_hashinfo * hashinfo __attribute__((__aligned__(64))); /* 0 8 */
/* typedef refcount_t */ struct refcount_struct {
/* typedef atomic_t */ struct {
int counter; /* 8 4 */
} refs; /* 8 4 */
} tw_refcount; /* 8 4 */
int sysctl_max_tw_buckets; /* 12 4 */
} tcp_death_row __attribute__((__aligned__(64))) __attribute__((__aligned__(64))); /* 0 64 *//* XXX last struct has 48 bytes of padding */ /* --- cacheline 1 boundary (64 bytes) --- */
...
} __attribute__((__aligned__(64)));
---8<---
include/net/netns/ipv4.h +2 -3
@@ -32,16 +32,15 @@ struct ping_group_range {struct inet_hashinfo ;struct inet_timewait_death_row { -refcount_t tw_refcount; -struct inet_hashinfo *hashinfo____cacheline_aligned_in_smp ; +refcount_t tw_refcount;int sysctl_max_tw_buckets; };struct tcp_fastopen_context ;struct netns_ipv4 { -struct inet_timewait_death_row *tcp_death_row; +struct inet_timewait_death_row tcp_death_row;#ifdef CONFIG_SYSCTL struct ctl_table_header *forw_hdr;
---8<---
Quoted 28 lines
$ pahole -C netns_ipv4,inet_timewait_death_row vmlinux struct netns_ipv4 { struct inet_timewait_death_row * tcp_death_row; /* 0 8 */ struct ctl_table_header * forw_hdr; /* 8 8 */ struct ctl_table_header * frags_hdr; /* 16 8 */ struct ctl_table_header * ipv4_hdr; /* 24 8 */ struct ctl_table_header * route_hdr; /* 32 8 */ struct ctl_table_header * xfrm4_hdr; /* 40 8 */ struct ipv4_devconf * devconf_all; /* 48 8 */ struct ipv4_devconf * devconf_dflt; /* 56 8 */ /* --- cacheline 1 boundary (64 bytes) --- */ ... }; struct inet_timewait_death_row { refcount_t tw_refcount; /* 0 4 *//* XXX 60 bytes hole, try to pack */ /* --- cacheline 1 boundary (64 bytes) --- */ struct inet_hashinfo * hashinfo __attribute__((__aligned__(64))); /* 64 8 */ int sysctl_max_tw_buckets; /* 72 4 */ /* size: 128, cachelines: 2, members: 3 */ /* sum members: 16, holes: 1, sum holes: 60 */ /* padding: 52 */ /* forced alignments: 1, forced holes: 1, sum forced holes: 60 */ } __attribute__((__aligned__(64)));
Re: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
Quoted 146 lines · Kuniyuki Iwashima
From: Kuniyuki Iwashima <[email protected]> Date: Fri, 2 Sep 2022 18:12:43 -0700Quoted 108 lines
From: Eric Dumazet <[email protected]> Date: Fri, 2 Sep 2022 17:53:18 -0700Quoted 98 lines · Kuniyuki Iwashima
Quoted 90 lines
From: Kuniyuki Iwashima <[email protected]> Date: Thu, 1 Sep 2022 15:12:16 -0700Quoted 34 lines
From: Eric Dumazet <[email protected]> Date: Thu, 1 Sep 2022 14:30:43 -0700Quoted 32 lines · Kuniyuki Iwashima
Quoted 21 lines
From: Paolo Abeni <[email protected]>
Quoted 8 lines
/Me is thinking aloud...
I'm wondering if the above has some measurable negative effect for large deployments using only the main netns?
Specifically, are
net->ipv4.tcp_death_rowand net->ipv4.tcp_death_row-Quoted 1 line
hashinfo already into the working set data for established socket?
Would the above increase the WSS by 2 cache-lines?
Currently, the
death_rowand hashinfo are touched around tw sockets orconnect(). If connections on the deployment are short-lived or frequently initiated by itself, that would be host and included in WSS.If the workload is server and there's no active-close() socket or connections are long-lived, then it might not be included in WSS. But I think it's not likely than the former if the deployment is large enough.
Quoted 4 lines
If this change had large impact, then we could revert
fbb8295248e1which convertednet->ipv4.tcp_death_rowinto pointer for0dad4087a86athat tried to fire a TW timer after netns is freed, but0dad4087a86ahas already reverted.Concern was fast path.
Each incoming packet does a socket lookup.
Quoted 7 lines
Quoted 5 lines
Fetching hashinfo (instead of
&tcp_hashinfo) with a dereference of a field in 'struct net' might inccurr a new cache line miss.Previously, first cache line of
tcp_infowas enough to bring a lot of fields in cpu cache.Ok, let me test on that if there could be regressions.
I tested
tcp_hashinfovstcp_death_row->hashinfowithsuper_netperfand collected HW cache-related metrics with perf.After the patch the number of L1 miss seems to increase, but the instructions per cycle also increases, and cache miss rate did not change. Also, there was not performance regression for netperf.
Tested:
# cat perf_super_netperf echo 0 > /proc/sys/kernel/nmi_watchdog echo 3 > /proc/sys/vm/drop_cachesperf stat -a \ -e cycles,instructions,cache-references,cache-misses,bus-cycles \ -e L1-dcache-loads,L1-dcache-load-misses,L1-dcache-stores \ -e dTLB-loads,dTLB-load-misses \ -e LLC-loads,LLC-load-misses,LLC-stores \ ./super_netperf $(($(nproc) * 2)) -H 10.0.0.142 -l 60 -fMecho 1 >
/proc/sys/kernel/nmi_watchdogBefore:
# ./perf_super_netperf 2929.81Performance counter stats for 'system wide':
494,002,600,338 cycles (23.07%) 241,230,662,890 instructions # 0.49 insn per cycle (30.76%) 6,303,603,008 cache-references (38.45%) 1,421,440,332 cache-misses # 22.550 % of all cache refs (46.15%) 4,861,179,308 bus-cycles (46.15%) 65,410,735,599 L1-dcache-loads (46.15%) 12,647,247,339 L1-dcache-load-misses # 19.34% of all L1-dcache accesses (30.77%) 32,912,656,369 L1-dcache-stores (30.77%) 66,015,779,361 dTLB-loads (30.77%) 81,293,994 dTLB-load-misses # 0.12% of all dTLB cache accesses (30.77%) 2,946,386,949 LLC-loads (30.77%) 257,223,942 LLC-load-misses # 8.73% of all LL-cache accesses (30.77%) 1,183,820,461 LLC-stores (15.38%)Quoted 2 lines
62.132250590 seconds time elapsedThis test will not be able to see a difference really...
What is needed is to measure the latency when nothing at all is in the caches.
Quoted 7 lines
Vast majority of real world TCP traffic is light or moderate. Packets are received and cpu has to bring X cache lines into L1 in order to process one packet.
We slowly are increasing X over time :/
pahole is your friend, more than a stress-test.
Quoted 7 lines
Here's pahole result on my local build. As Paolo said, we need 2 cachelines for
tcp_death_rowand the hashinfo?How about moving hashinfo as the first member of struct
inet_timewait_death_rowand convert it to just struct instead of pointer so that we need 1 cache line to read hashinfo?Like this.
$ pahole -EC netns_ipv4 vmlinux struct netns_ipv4 { struct inet_timewait_death_row { struct inet_hashinfo * hashinfo __attribute__((__aligned__(64))); /* 0 8 */ /* typedef refcount_t */ struct refcount_struct { /* typedef atomic_t */ struct { int counter; /* 8 4 */ } refs; /* 8 4 */ } tw_refcount; /* 8 4 */ int sysctl_max_tw_buckets; /* 12 4 */ } tcp_death_row __attribute__((__aligned__(64))) __attribute__((__aligned__(64))); /* 0 64 *//* XXX last struct has 48 bytes of padding */ /* --- cacheline 1 boundary (64 bytes) --- */...
} __attribute__((__aligned__(64)));---8<---
include/net/netns/ipv4.h +0
-0
include/net/netns/ipv4.h struct ping_group_range { · line 36
Quoted 7 lines
include/net/netns/ipv4.h +1
-2@@ -32,16 +32,15 @@ struct ping_group_range {struct inet_hashinfo ;struct inet_timewait_death_row { -refcount_t tw_refcount; -struct inet_hashinfo *hashinfo____cacheline_aligned_in_smp ; +refcount_t tw_refcount;
This would be very bad. tw_refcount would share a cache line with hashinfo.
false sharing is more problematic than a cache line miss in read mode.
Quoted 13 lines
include/net/netns/ipv4.h +1
-2int sysctl_max_tw_buckets; };struct tcp_fastopen_context ;struct netns_ipv4 { -struct inet_timewait_death_row *tcp_death_row; +struct inet_timewait_death_row tcp_death_row;#ifdef CONFIG_SYSCTL struct ctl_table_header *forw_hdr; ---8<---
Re: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
From: Eric Dumazet <[email protected]> Date: Fri, 2 Sep 2022 19:30:31 -0700
Quoted 157 lines · Kuniyuki Iwashima
Quoted 146 lines
From: Kuniyuki Iwashima <[email protected]> Date: Fri, 2 Sep 2022 18:12:43 -0700Quoted 108 lines
From: Eric Dumazet <[email protected]> Date: Fri, 2 Sep 2022 17:53:18 -0700Quoted 98 lines · Kuniyuki Iwashima
Quoted 90 lines
From: Kuniyuki Iwashima <[email protected]> Date: Thu, 1 Sep 2022 15:12:16 -0700Quoted 34 lines
From: Eric Dumazet <[email protected]> Date: Thu, 1 Sep 2022 14:30:43 -0700Quoted 32 lines · Kuniyuki Iwashima
Quoted 21 lines
From: Paolo Abeni <[email protected]>
Quoted 8 lines
/Me is thinking aloud...
I'm wondering if the above has some measurable negative effect for large deployments using only the main netns?
Specifically, are
net->ipv4.tcp_death_rowand net->ipv4.tcp_death_row-Quoted 1 line
hashinfo already into the working set data for established socket?
Would the above increase the WSS by 2 cache-lines?
Currently, the
death_rowand hashinfo are touched around tw sockets orconnect(). If connections on the deployment are short-lived or frequently initiated by itself, that would be host and included in WSS.If the workload is server and there's no active-close() socket or connections are long-lived, then it might not be included in WSS. But I think it's not likely than the former if the deployment is large enough.
Quoted 4 lines
If this change had large impact, then we could revert
fbb8295248e1which convertednet->ipv4.tcp_death_rowinto pointer for0dad4087a86athat tried to fire a TW timer after netns is freed, but0dad4087a86ahas already reverted.Concern was fast path.
Each incoming packet does a socket lookup.
Quoted 7 lines
Quoted 5 lines
Fetching hashinfo (instead of
&tcp_hashinfo) with a dereference of a field in 'struct net' might inccurr a new cache line miss.Previously, first cache line of
tcp_infowas enough to bring a lot of fields in cpu cache.Ok, let me test on that if there could be regressions.
I tested
tcp_hashinfovstcp_death_row->hashinfowithsuper_netperfand collected HW cache-related metrics with perf.After the patch the number of L1 miss seems to increase, but the instructions per cycle also increases, and cache miss rate did not change. Also, there was not performance regression for netperf.
Tested:
# cat perf_super_netperf echo 0 > /proc/sys/kernel/nmi_watchdog echo 3 > /proc/sys/vm/drop_cachesperf stat -a \ -e cycles,instructions,cache-references,cache-misses,bus-cycles \ -e L1-dcache-loads,L1-dcache-load-misses,L1-dcache-stores \ -e dTLB-loads,dTLB-load-misses \ -e LLC-loads,LLC-load-misses,LLC-stores \ ./super_netperf $(($(nproc) * 2)) -H 10.0.0.142 -l 60 -fMecho 1 >
/proc/sys/kernel/nmi_watchdogBefore:
# ./perf_super_netperf 2929.81Performance counter stats for 'system wide':
494,002,600,338 cycles (23.07%) 241,230,662,890 instructions # 0.49 insn per cycle (30.76%) 6,303,603,008 cache-references (38.45%) 1,421,440,332 cache-misses # 22.550 % of all cache refs (46.15%) 4,861,179,308 bus-cycles (46.15%) 65,410,735,599 L1-dcache-loads (46.15%) 12,647,247,339 L1-dcache-load-misses # 19.34% of all L1-dcache accesses (30.77%) 32,912,656,369 L1-dcache-stores (30.77%) 66,015,779,361 dTLB-loads (30.77%) 81,293,994 dTLB-load-misses # 0.12% of all dTLB cache accesses (30.77%) 2,946,386,949 LLC-loads (30.77%) 257,223,942 LLC-load-misses # 8.73% of all LL-cache accesses (30.77%) 1,183,820,461 LLC-stores (15.38%)Quoted 2 lines
62.132250590 seconds time elapsedThis test will not be able to see a difference really...
What is needed is to measure the latency when nothing at all is in the caches.
Quoted 7 lines
Vast majority of real world TCP traffic is light or moderate. Packets are received and cpu has to bring X cache lines into L1 in order to process one packet.
We slowly are increasing X over time :/
pahole is your friend, more than a stress-test.
Quoted 7 lines
Here's pahole result on my local build. As Paolo said, we need 2 cachelines for
tcp_death_rowand the hashinfo?How about moving hashinfo as the first member of struct
inet_timewait_death_rowand convert it to just struct instead of pointer so that we need 1 cache line to read hashinfo?Like this.
$ pahole -EC netns_ipv4 vmlinux struct netns_ipv4 { struct inet_timewait_death_row { struct inet_hashinfo * hashinfo __attribute__((__aligned__(64))); /* 0 8 */ /* typedef refcount_t */ struct refcount_struct { /* typedef atomic_t */ struct { int counter; /* 8 4 */ } refs; /* 8 4 */ } tw_refcount; /* 8 4 */ int sysctl_max_tw_buckets; /* 12 4 */ } tcp_death_row __attribute__((__aligned__(64))) __attribute__((__aligned__(64))); /* 0 64 *//* XXX last struct has 48 bytes of padding */ /* --- cacheline 1 boundary (64 bytes) --- */...
} __attribute__((__aligned__(64)));---8<---
include/net/netns/ipv4.h +0
-0include/net/netns/ipv4.h struct ping_group_range { · line 36
Quoted 7 lines
include/net/netns/ipv4.h +1
-2@@ -32,16 +32,15 @@ struct ping_group_range {struct inet_hashinfo ;struct inet_timewait_death_row { -refcount_t tw_refcount; -struct inet_hashinfo *hashinfo____cacheline_aligned_in_smp ; +refcount_t tw_refcount;This would be very bad.
tw_refcountwould share a cache line with hashinfo.
Quoted 1 line
false sharing is more problematic than a cache line miss in read mode.
Ah, exactly.
Then I will add ____cacheline_aligned_in_smp to tw_refcount, instead of keeping the original order, to avoid invalidation by sysctl_max_tw_buckets change, which wouldn't be so frequently done though.
struct inet_timewait_death_row {
- refcount_t tw_refcount;
-
- struct inet_hashinfo *hashinfo ____cacheline_aligned_in_smp;
+ struct inet_hashinfo *hashinfo;
+ refcount_t tw_refcount ____cacheline_aligned_in_smp;
int sysctl_max_tw_buckets;};
Quoted 14 lines
Quoted 14 lines
include/net/netns/ipv4.h +1
-2int sysctl_max_tw_buckets; };struct tcp_fastopen_context ;struct netns_ipv4 { -struct inet_timewait_death_row *tcp_death_row; +struct inet_timewait_death_row tcp_death_row;#ifdef CONFIG_SYSCTL struct ctl_table_header *forw_hdr; ---8<---
Re: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
include/net/netns/ipv4.h struct ping_group_range { · line 37
Quoted 12 lines · Kuniyuki Iwashima
Ah, exactly.
Then I will add____cacheline_aligned_in_smptotw_refcount, instead of keeping the original order, to avoid invalidation bysysctl_max_tw_bucketschange, which wouldn't be so frequently done though.struct
inet_timewait_death_row{include/net/netns/ipv4.h +2
-3-refcount_t tw_refcount; - -struct inet_hashinfo *hashinfo____cacheline_aligned_in_smp ; +struct inet_hashinfo *hashinfo; +refcount_t tw_refcount____cacheline_aligned_in_smp ;int sysctl_max_tw_buckets;
This would move sysctl_max_tw_buckets in a separate/dedicated cache line :/
-->
{
refcount_t tw_refcount;
/* Padding to avoid false sharing, tw_refcount can be often written */
struct inet_hashinfo *hashinfo ____cacheline_aligned_in_smp;
int sysctl_max_tw_buckets;
.. other read only fields could fit here.
Quoted 1 line
};
Explicit alignment of the structure or first field is not needed, they will already be cache line aligned.
Re: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
From: Eric Dumazet <[email protected]> Date: Fri, 2 Sep 2022 20:16:11 -0700
Quoted 25 lines · Kuniyuki Iwashima
include/net/netns/ipv4.h struct ping_group_range { · line 37
Quoted 13 lines
Ah, exactly.
Then I will add____cacheline_aligned_in_smptotw_refcount, instead of keeping the original order, to avoid invalidation bysysctl_max_tw_bucketschange, which wouldn't be so frequently done though.struct
inet_timewait_death_row{include/net/netns/ipv4.h +2
-3-refcount_t tw_refcount; - -struct inet_hashinfo *hashinfo____cacheline_aligned_in_smp ; +struct inet_hashinfo *hashinfo; +refcount_t tw_refcount____cacheline_aligned_in_smp ;int sysctl_max_tw_buckets;This would move
sysctl_max_tw_bucketsin a separate/dedicated cache line :/-->
{ refcount_t tw_refcount; /* Padding to avoid false sharing, tw_refcount can be often written */ struct inet_hashinfo *hashinfo ____cacheline_aligned_in_smp; int sysctl_max_tw_buckets;
Quoted 6 lines
.. other read only fields could fit here.
Quoted 1 line
};
Explicit alignment of the structure or first field is not needed, they will already be cache line aligned.
I got it. I'll add that change.
Thank you so much!
Jakub Kicinski
maintainerRe: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
In reply to Kuniyuki IwashimaQuoted 9 lines · On Thu, 1 Sep 2022 14:25:20 -0700 Kuniyuki Iwashima wrote:
Quoted 8 lines
I looks to me that the above chunks are functionally a no-op and I think that omitting the 2 drivers from the v2:
https://lore.kernel.org/netdev/[email protected]/
should break mlx5/nfp inside a netns. I don't understand why including the above and skipping the latters?!? I guess is a question mostly for Eric :)
Quoted 8 lines
My best guess is that it's ok unless it does not touch TCP stack deeply and if it does, the driver developer must catch up with the core changes not to burden maintainers...?
If so, I understand that take. OTOH, I also don't want to break anything when we know the change would do.
So, I'm fine to either stay as is or add the change in v4 again.
FWIW I share Paolo's concern. If we don't want the drivers to be twiddling with the hash tables we should factor out that code to a common helper in net/tls/
Re: [PATCH v3 net-next 3/5] tcp: Access &tcp_hashinfo via net.
From: Jakub Kicinski <[email protected]> Date: Thu, 1 Sep 2022 14:49:36 -0700
Quoted 14 lines · On Thu, 1 Sep 2022 14:25:20 -0700 Kuniyuki Iwashima wrote:
Quoted 13 lines
Quoted 8 lines
I looks to me that the above chunks are functionally a no-op and I think that omitting the 2 drivers from the v2:
https://lore.kernel.org/netdev/[email protected]/
should break mlx5/nfp inside a netns. I don't understand why including the above and skipping the latters?!? I guess is a question mostly for Eric :)
My best guess is that it's ok unless it does not touch TCP stack deeply and if it does, the driver developer must catch up with the core changes not to burden maintainers...?
Quoted 8 lines
Quoted 4 lines
If so, I understand that take. OTOH, I also don't want to break anything when we know the change would do.
So, I'm fine to either stay as is or add the change in v4 again.
FWIW I share Paolo's concern. If we don't want the drivers to be twiddling with the hash tables we should factor out that code to a common helper in
net/tls/
That makes sense.
For the moment, I'll add the changes back in v4.
PATCH v3 4/5 tcp: Save unnecessary inet_twsk_purge() calls.
In reply to Kuniyuki IwashimaWhile destroying netns, we call inet_twsk_purge() in tcp_sk_exit_batch() and tcpv6_net_exit_batch() for AF_INET and AF_INET6. These commands trigger the kernel to walk through the potentially big ehash twice even though the netns has no TIME_WAIT sockets.
# ip netns add test # ip netns del test
or
# unshare -n /bin/true >/dev/null
AF_INET6 uses module_init() to be loaded after AF_INET which uses fs_initcall(), so tcpv6_net_ops is always registered after tcp_sk_ops. Also, we clean up netns in the reverse order, so tcpv6_net_exit_batch() is always called before tcp_sk_exit_batch().
The characteristic enables us to save such unneeded iterations when all netns in net_exit_list have no tw sockets. This change eliminates the tax by the additional unshare() described in the next patch to guarantee the per-netns ehash size.
Tested:
# echo cleanup_net > /sys/kernel/debug/tracing/set_ftrace_filter
# echo inet_twsk_purge >> /sys/kernel/debug/tracing/set_ftrace_filter
# echo function > /sys/kernel/debug/tracing/current_tracer
# cat ./add_del_unshare.sh
for i in `seq 1 40`
do
(for j in `seq 1 100` ; do unshare -n /bin/true >/dev/null ; done) &
done
wait;
# ./add_del_unshare.shBefore the patch:
# cat /sys/kernel/debug/tracing/trace_pipe
kworker/u128:0-8 [031] ...1. 174.162765: cleanup_net <-process_one_work
kworker/u128:0-8 [031] ...1. 174.240796: inet_twsk_purge <-cleanup_net
kworker/u128:0-8 [032] ...1. 174.244759: inet_twsk_purge <-tcp_sk_exit_batch
kworker/u128:0-8 [034] ...1. 174.290861: cleanup_net <-process_one_work
kworker/u128:0-8 [039] ...1. 175.245027: inet_twsk_purge <-cleanup_net
kworker/u128:0-8 [046] ...1. 175.290541: inet_twsk_purge <-tcp_sk_exit_batch
kworker/u128:0-8 [037] ...1. 175.321046: cleanup_net <-process_one_work
kworker/u128:0-8 [024] ...1. 175.941633: inet_twsk_purge <-cleanup_net
kworker/u128:0-8 [025] ...1. 176.242539: inet_twsk_purge <-tcp_sk_exit_batchAfter:
# cat /sys/kernel/debug/tracing/trace_pipe
kworker/u128:0-8 [038] ...1. 428.116174: cleanup_net <-process_one_work
kworker/u128:0-8 [038] ...1. 428.262532: cleanup_net <-process_one_work
kworker/u128:0-8 [030] ...1. 429.292645: cleanup_net <-process_one_work--- include/net/tcp.h | 1 + net/ipv4/tcp_ipv4.c | 6 ++++-- net/ipv4/tcp_minisocks.c | 24 ++++++++++++++++++++++++ net/ipv6/tcp_ipv6.c | 2 +- 4 files changed, 30 insertions(+), 3 deletions(-)
include/net/tcp.h +1 -0
@@ -346,6 +346,7 @@ void tcp_rcv_established(struct sock *sk, struct sk_buff *skb);void tcp_rcv_space_adjust (struct sock *sk);int tcp_twsk_unique (struct sock *sk,struct sock *sktw,void *twp);void tcp_twsk_destructor (struct sock *sk); +void tcp_twsk_purge (struct list_head *net_exit_list,int family);ssize_t tcp_splice_read (struct socket *sk,loff_t *ppos,struct pipe_inode_info *pipe,size_t len,unsigned int flags);
net/ipv4/tcp_ipv4.c +4 -2
@@ -3109,8 +3109,10 @@ static void __net_exit tcp_sk_exit(struct net *net)if (net->ipv4.tcp_congestion_control)bpf_module_put (net->ipv4.tcp_congestion_control, net->ipv4.tcp_congestion_control->owner); -if (refcount_dec_and_test (&tcp_death_row->tw_refcount)) +if (refcount_dec_and_test (&tcp_death_row->tw_refcount)) {kfree (tcp_death_row); + net->ipv4.tcp_death_row =NULL ; + } }static int __net_init tcp_sk_init (struct net *net)
@@ -3210,7 +3212,7 @@ static void __net_exit tcp_sk_exit_batch(struct list_head *net_exit_list) {struct net *net; -inet_twsk_purge (&tcp_hashinfo,AF_INET ); +tcp_twsk_purge (net_exit_list,AF_INET );list_for_each_entry (net, net_exit_list, exit_list)tcp_fastopen_ctx_destroy (net);
net/ipv4/tcp_minisocks.c +24 -0
@@ -347,6 +347,30 @@ void tcp_twsk_destructor(struct sock *sk) }EXPORT_SYMBOL_GPL (tcp_twsk_destructor); +void tcp_twsk_purge (struct list_head *net_exit_list,int family) +{ +struct net *net; + +list_for_each_entry (net, net_exit_list, exit_list) { +if (!net->ipv4.tcp_death_row) +continue ; + +/* AF_INET6 using module_init() is always registered after +* AF_INET using fs_initcall() and cleaned up in the reverse +* order. +* +* The last refcount is decremented later in tcp_sk_exit(). +*/ +if (IS_ENABLED (CONFIG_IPV6 ) && family ==AF_INET6 && +refcount_read (&net->ipv4.tcp_death_row->tw_refcount) ==1 ) +continue ; + +inet_twsk_purge (&tcp_hashinfo, family); +break ; + } +} +EXPORT_SYMBOL_GPL (tcp_twsk_purge); +/* Warning : This function is called without sk_listener being locked. * Be sure to read socket fields once, as their value could change under us. */
net/ipv6/tcp_ipv6.c +1 -1
@@ -2229,7 +2229,7 @@ static void __net_exit tcpv6_net_exit(struct net *net) static void __net_exit tcpv6_net_exit_batch(struct list_head *net_exit_list) { -inet_twsk_purge (&tcp_hashinfo,AF_INET6 ); +tcp_twsk_purge (net_exit_list,AF_INET6 ); }static struct pernet_operations tcpv6_net_ops = {
Signature · 1 line
2.30.2
PATCH v3 5/5 tcp: Introduce optional per-netns ehash.
In reply to Kuniyuki IwashimaThe more sockets we have in the hash table, the longer we spend looking up the socket. While running a number of small workloads on the same host, they penalise each other and cause performance degradation.
The root cause might be a single workload that consumes much more resources than the others. It often happens on a cloud service where different workloads share the same computing resource.
On EC2 c5.24xlarge instance (196 GiB memory and 524288 (1Mi / 2) ehash entries), after running iperf3 in different netns, creating 24Mi sockets without data transfer in the root netns causes about 10% performance regression for the iperf3's connection.
thash_entries sockets length Gbps
524288 1 1 50.7 24Mi 48 45.1
It is basically related to the length of the list of each hash bucket. For testing purposes to see how performance drops along the length, I set 131072 (1Mi / 8) to thash_entries, and here's the result.
thash_entries sockets length Gbps
131072 1 1 50.7 1Mi 8 49.9 2Mi 16 48.9 4Mi 32 47.3 8Mi 64 44.6 16Mi 128 40.6 24Mi 192 36.3 32Mi 256 32.5 40Mi 320 27.0 48Mi 384 25.0
To resolve the socket lookup degradation, we introduce an optional per-netns hash table for TCP, but it's just ehash, and we still share the global bhash, bhash2 and lhash2.
With a smaller ehash, we can look up non-listener sockets faster and isolate such noisy neighbours. Also, we can reduce lock contention.
We can control the ehash size by a new sysctl knob. However, depending on workloads, it will require very sensitive tuning, so we disable the feature by default (net.ipv4.tcp_child_ehash_entries == 0). Moreover, we can fall back to using the global ehash in case we fail to allocate enough memory for a new ehash. The maximum size is 16Mi, which is large enough that even if we have 48Mi sockets, the average list length is 3, and regression would be less than 1%.
We can check the current ehash size by another read-only sysctl knob, net.ipv4.tcp_ehash_entries. A negative value means the netns shares the global ehash (per-netns ehash is disabled or failed to allocate memory).
# dmesg | cut -d ' ' -f 5- | grep "established hash" TCP established hash table entries: 524288 (order: 10, 4194304 bytes, vmalloc hugepage)
# sysctl net.ipv4.tcp_ehash_entries net.ipv4.tcp_ehash_entries = 524288 # can be changed by thash_entries
# sysctl net.ipv4.tcp_child_ehash_entries net.ipv4.tcp_child_ehash_entries = 0 # disabled by default
# ip netns add test1 # ip netns exec test1 sysctl net.ipv4.tcp_ehash_entries net.ipv4.tcp_ehash_entries = -524288 # share the global ehash
# sysctl -w net.ipv4.tcp_child_ehash_entries=100 net.ipv4.tcp_child_ehash_entries = 100
# ip netns add test2 # ip netns exec test2 sysctl net.ipv4.tcp_ehash_entries net.ipv4.tcp_ehash_entries = 128 # own a per-netns ehash with 2^n buckets
When more than two processes in the same netns create per-netns ehash concurrently with different sizes, we need to guarantee the size in one of the following ways:
1) Share the global ehash and create per-netns ehash
First, unshare() with tcp_child_ehash_entries==0. It creates dedicated netns sysctl knobs where we can safely change tcp_child_ehash_entries and clone()/unshare() to create a per-netns ehash.
2) Control write on sysctl by BPF
We can use BPF_PROG_TYPE_CGROUP_SYSCTL to allow/deny read/write on sysctl knobs.
Note the default values of two sysctl knobs depend on the ehash size and should be tuned carefully:
tcp_max_tw_buckets : tcp_child_ehash_entries / 2 tcp_max_syn_backlog : max(128, tcp_child_ehash_entries / 128)
Also, we could optimise ehash lookup/iteration further by removing netns comparison for the per-netns ehash in the future.
As a bonus, we can dismantle netns faster. Currently, while destroying netns, we call inet_twsk_purge(), which walks through the global ehash. It can be potentially big because it can have many sockets other than TIME_WAIT in all netns. Splitting ehash changes that situation, where it's only necessary for inet_twsk_purge() to clean up TIME_WAIT sockets in each netns.
--- Documentation/networking/ip-sysctl.rst | 23 +++++++++++ include/net/inet_hashtables.h | 6 +++ include/net/netns/ipv4.h | 1 + net/dccp/proto.c | 2 + net/ipv4/inet_hashtables.c | 57 ++++++++++++++++++++++++++ net/ipv4/inet_timewait_sock.c | 4 +- net/ipv4/sysctl_net_ipv4.c | 39 ++++++++++++++++++ net/ipv4/tcp.c | 1 + net/ipv4/tcp_ipv4.c | 39 +++++++++++++++--- net/ipv4/tcp_minisocks.c | 9 +++- 10 files changed, 172 insertions(+), 9 deletions(-)
Documentation/networking/ip-sysctl.rst +23 -0
@@ -1037,6 +1037,29 @@ tcp_challenge_ack_limit - INTEGER in RFC 5961 (Improving TCP's Robustness to Blind In-Window Attacks) Default: 1000 +tcp_ehash_entries - INTEGER + Read-only number of hash buckets for TCP sockets in the current + networking namespace. + + A negative value means the networking namespace does not own its + hash buckets and shares the initial networking namespace's one. + +tcp_child_ehash_entries - INTEGER + Control the number of hash buckets for TCP sockets in the child + networking namespace, which must be set before clone() or unshare(). + + If the value is not 0, the kernel uses a value rounded up to 2^n + as the actual hash bucket size. 0 is a special value, meaning + the child networking namespace will share the initial networking + namespace's hash buckets. + + Note that the child will use the global one in case the kernel + fails to allocate enough memory. + + Possible values: 0, 2^n (n: 0 - 24 (16Mi)) + + Default: 0 + UDP variables =============
include/net/inet_hashtables.h +6 -0
@@ -168,6 +168,8 @@ struct inet_hashinfo {/* The 2nd listener table hashed by local port and address */ unsigned int lhash2_mask;struct inet_listen_hashbucket *lhash2; + +bool pernet; };static inline struct inet_hashinfo *tcp_or_dccp_get_hashinfo (const struct sock *sk)
@@ -214,6 +216,10 @@ static inline void inet_ehash_locks_free(struct inet_hashinfo *hashinfo) hashinfo->ehash_locks =NULL ; } +struct inet_hashinfo *inet_pernet_hashinfo_alloc (struct inet_hashinfo *hashinfo, +unsigned int ehash_entries); +void inet_pernet_hashinfo_free (struct inet_hashinfo *hashinfo); +struct inet_bind_bucket * inet_bind_bucket_create(struct kmem_cache *cachep, struct net *net, struct inet_bind_hashbucket *head,
include/net/netns/ipv4.h +1 -0
@@ -170,6 +170,7 @@ struct netns_ipv4 {int sysctl_tcp_pacing_ca_ratio;int sysctl_tcp_wmem[3 ];int sysctl_tcp_rmem[3 ]; +unsigned int sysctl_tcp_child_ehash_entries;unsigned long sysctl_tcp_comp_sack_delay_ns;unsigned long sysctl_tcp_comp_sack_slack_ns;int sysctl_max_syn_backlog;
net/dccp/proto.c +2 -0
@@ -1197,6 +1197,8 @@ static int __init dccp_init(void)INIT_HLIST_HEAD (&dccp_hashinfo.bhash2[i].chain); } + dccp_hashinfo.pernet = false; + rc =dccp_mib_init ();if (rc) goto out_free_dccp_bhash2;
net/ipv4/inet_hashtables.c +57 -0
60 lines not on this page · raw message
net/ipv4/inet_timewait_sock.c +3 -1
@@ -59,8 +59,10 @@ static void inet_twsk_kill(struct inet_timewait_sock *tw)inet_twsk_bind_unhash (tw, hashinfo);spin_unlock (&bhead->lock); -if (refcount_dec_and_test (&tw->tw_dr->tw_refcount)) +if (refcount_dec_and_test (&tw->tw_dr->tw_refcount)) { +inet_pernet_hashinfo_free (hashinfo);kfree (tw->tw_dr); + }inet_twsk_put (tw); }
net/ipv4/sysctl_net_ipv4.c +39 -0
@@ -39,6 +39,7 @@ static u32 u32_max_div_HZ = UINT_MAX / HZ;static int one_day_secs =24 *3600 ;static u32 fib_multipath_hash_fields_all_mask__maybe_unused =FIB_MULTIPATH_HASH_FIELD_ALL_MASK ; +static unsigned int tcp_child_ehash_entries_max =16 *1024 *1024 ;/* obsolete */ static int sysctl_tcp_low_latency__read_mostly ;
@@ -382,6 +383,29 @@ static int proc_tcp_available_ulp(struct ctl_table *ctl,return ret; } +static int proc_tcp_ehash_entries (struct ctl_table *table,int write, +void *buffer,size_t *lenp,loff_t *ppos) +{ + struct net *net = container_of(table->data, struct net, + ipv4.sysctl_tcp_child_ehash_entries); +struct inet_hashinfo *hinfo = net->ipv4.tcp_death_row->hashinfo; +int tcp_ehash_entries; +struct ctl_table tbl; + + tcp_ehash_entries = hinfo->ehash_mask +1 ; + +/* A negative number indicates that the child netns +* shares the global ehash. +*/ +if (!net_eq (net, &init_net) && !hinfo->pernet) + tcp_ehash_entries *=-1 ; + + tbl.data = &tcp_ehash_entries; + tbl.maxlen =sizeof (int ); + +return proc_dointvec (&tbl, write, buffer, lenp, ppos); +} +#ifdef CONFIG_IP_ROUTE_MULTIPATH static int proc_fib_multipath_hash_policy(struct ctl_table *table, int write, void *buffer, size_t *lenp,
@@ -1321,6 +1345,21 @@ static struct ctl_table ipv4_net_table[] = { .extra1 = SYSCTL_ZERO, .extra2 = SYSCTL_ONE, }, + { + .procname = "tcp_ehash_entries", + .data = &init_net.ipv4.sysctl_tcp_child_ehash_entries, + .mode = 0444, + .proc_handler = proc_tcp_ehash_entries, + }, + { + .procname = "tcp_child_ehash_entries", + .data = &init_net.ipv4.sysctl_tcp_child_ehash_entries, + .maxlen = sizeof(unsigned int), + .mode = 0644, + .proc_handler = proc_douintvec_minmax, + .extra1 = SYSCTL_ZERO, + .extra2 = &tcp_child_ehash_entries_max, + }, { .procname = "udp_rmem_min", .data = &init_net.ipv4.sysctl_udp_rmem_min,
net/ipv4/tcp.c +1 -0
@@ -4788,6 +4788,7 @@ void __init tcp_init(void)INIT_HLIST_HEAD (&tcp_hashinfo.bhash2[i].chain); } + tcp_hashinfo.pernet = false; cnt = tcp_hashinfo.ehash_mask +1 ; sysctl_tcp_max_orphans = cnt /2 ;
net/ipv4/tcp_ipv4.c +33 -6
@@ -3110,15 +3110,44 @@ static void __net_exit tcp_sk_exit(struct net *net)bpf_module_put (net->ipv4.tcp_congestion_control, net->ipv4.tcp_congestion_control->owner);if (refcount_dec_and_test (&tcp_death_row->tw_refcount)) { +inet_pernet_hashinfo_free (tcp_death_row->hashinfo);kfree (tcp_death_row); net->ipv4.tcp_death_row =NULL ; } } -static int __net_init tcp_sk_init (struct net *net) +static void __net_init tcp_set_hashinfo (struct net *net) { -int cnt; +struct inet_hashinfo *hinfo; +unsigned int ehash_entries; +struct net *old_net; + +if (net_eq (net, &init_net)) + goto fallback; + + old_net = current->nsproxy->net_ns; + ehash_entries =READ_ONCE (old_net->ipv4.sysctl_tcp_child_ehash_entries); +if (!ehash_entries) + goto fallback; + + ehash_entries =roundup_pow_of_two (ehash_entries); + hinfo =inet_pernet_hashinfo_alloc (&tcp_hashinfo, ehash_entries); +if (!hinfo) { +pr_warn ("Failed to allocate TCP ehash (entries: %u) " +"for a netns, fallback to the global one\n" , + ehash_entries); +fallback: + hinfo = &tcp_hashinfo; + ehash_entries = tcp_hashinfo.ehash_mask +1 ; + } + net->ipv4.tcp_death_row->hashinfo = hinfo; + net->ipv4.tcp_death_row->sysctl_max_tw_buckets = ehash_entries /2 ; + net->ipv4.sysctl_max_syn_backlog =max (128U , ehash_entries /128 ); +} + +static int __net_init tcp_sk_init (struct net *net) +{ net->ipv4.sysctl_tcp_ecn =2 ; net->ipv4.sysctl_tcp_ecn_fallback =1 ;
@@ -3147,12 +3176,10 @@ static int __net_init tcp_sk_init(struct net *net) net->ipv4.tcp_death_row =kzalloc (sizeof (struct inet_timewait_death_row ),GFP_KERNEL );if (!net->ipv4.tcp_death_row)return -ENOMEM ; +refcount_set (&net->ipv4.tcp_death_row->tw_refcount,1 ); - cnt = tcp_hashinfo.ehash_mask +1 ; - net->ipv4.tcp_death_row->sysctl_max_tw_buckets = cnt /2 ; - net->ipv4.tcp_death_row->hashinfo = &tcp_hashinfo; +tcp_set_hashinfo (net); - net->ipv4.sysctl_max_syn_backlog =max (128 , cnt /128 ); net->ipv4.sysctl_tcp_sack =1 ; net->ipv4.sysctl_tcp_window_scaling =1 ; net->ipv4.sysctl_tcp_timestamps =1 ;
net/ipv4/tcp_minisocks.c +7 -2
@@ -349,6 +349,7 @@ EXPORT_SYMBOL_GPL(tcp_twsk_destructor);void tcp_twsk_purge (struct list_head *net_exit_list,int family) { +bool purged_once = false;struct net *net;list_for_each_entry (net, net_exit_list, exit_list) {
@@ -365,8 +366,12 @@ void tcp_twsk_purge(struct list_head *net_exit_list, int family) refcount_read(&net->ipv4.tcp_death_row->tw_refcount) == 1) continue; -inet_twsk_purge (&tcp_hashinfo, family); -break ; +if (net->ipv4.tcp_death_row->hashinfo->pernet) { +inet_twsk_purge (net->ipv4.tcp_death_row->hashinfo, family); + }else if (!purged_once) { +inet_twsk_purge (&tcp_hashinfo, family); + purged_once = true; + } } }EXPORT_SYMBOL_GPL (tcp_twsk_purge);
Signature · 1 line
2.30.2