From 6ddb811a579f87b8506344020002d396f814f7c8 Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Mon, 13 Oct 2025 15:22:31 +0000 Subject: [PATCH 1/4] net: add SK_WMEM_ALLOC_BIAS constant sk->sk_wmem_alloc is initialized to 1, and sk_wmem_alloc_get() takes care of this initial value. Add SK_WMEM_ALLOC_BIAS define to not spread this magic value. Signed-off-by: Eric Dumazet Reviewed-by: Neal Cardwell Reviewed-by: Kuniyuki Iwashima Link: https://patch.msgid.link/20251013152234.842065-2-edumazet@google.com Signed-off-by: Jakub Kicinski --- include/net/sock.h | 3 ++- net/atm/common.c | 2 +- net/core/sock.c | 5 ++--- 3 files changed, 5 insertions(+), 5 deletions(-) diff --git a/include/net/sock.h b/include/net/sock.h index 60bcb13f045c..2794bc5c5654 100644 --- a/include/net/sock.h +++ b/include/net/sock.h @@ -2303,6 +2303,7 @@ static inline int skb_copy_to_page_nocache(struct sock *sk, struct iov_iter *fro return 0; } +#define SK_WMEM_ALLOC_BIAS 1 /** * sk_wmem_alloc_get - returns write allocations * @sk: socket @@ -2311,7 +2312,7 @@ static inline int skb_copy_to_page_nocache(struct sock *sk, struct iov_iter *fro */ static inline int sk_wmem_alloc_get(const struct sock *sk) { - return refcount_read(&sk->sk_wmem_alloc) - 1; + return refcount_read(&sk->sk_wmem_alloc) - SK_WMEM_ALLOC_BIAS; } /** diff --git a/net/atm/common.c b/net/atm/common.c index 881c7f259dbd..cecc71a8bee1 100644 --- a/net/atm/common.c +++ b/net/atm/common.c @@ -157,7 +157,7 @@ int vcc_create(struct net *net, struct socket *sock, int protocol, int family, i memset(&vcc->local, 0, sizeof(struct sockaddr_atmsvc)); memset(&vcc->remote, 0, sizeof(struct sockaddr_atmsvc)); vcc->qos.txtp.max_sdu = 1 << 16; /* for meta VCs */ - refcount_set(&sk->sk_wmem_alloc, 1); + refcount_set(&sk->sk_wmem_alloc, SK_WMEM_ALLOC_BIAS); atomic_set(&sk->sk_rmem_alloc, 0); vcc->push = NULL; vcc->pop = NULL; diff --git a/net/core/sock.c b/net/core/sock.c index dc03d4b5909a..542cfa16ee12 100644 --- a/net/core/sock.c +++ b/net/core/sock.c @@ -2313,7 +2313,7 @@ struct sock *sk_alloc(struct net *net, int family, gfp_t priority, } sock_net_set(sk, net); - refcount_set(&sk->sk_wmem_alloc, 1); + refcount_set(&sk->sk_wmem_alloc, SK_WMEM_ALLOC_BIAS); mem_cgroup_sk_alloc(sk); cgroup_sk_alloc(&sk->sk_cgrp_data); @@ -2494,8 +2494,7 @@ struct sock *sk_clone_lock(const struct sock *sk, const gfp_t priority) atomic_set(&newsk->sk_rmem_alloc, 0); - /* sk_wmem_alloc set to one (see sk_free() and sock_wfree()) */ - refcount_set(&newsk->sk_wmem_alloc, 1); + refcount_set(&newsk->sk_wmem_alloc, SK_WMEM_ALLOC_BIAS); atomic_set(&newsk->sk_omem_alloc, 0); sk_init_common(newsk); From d365c9bca35cdeb534aac279c81d1fc9730bb100 Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Mon, 13 Oct 2025 15:22:32 +0000 Subject: [PATCH 2/4] net: control skb->ooo_okay from skb_set_owner_w() 15 years after Tom Herbert added skb->ooo_okay, only TCP transport benefits from it. We can support other transports directly from skb_set_owner_w(). If no other TX packet for this socket is in a host queue (qdisc, NIC queue) there is no risk of self-inflicted reordering, we can set skb->ooo_okay. This allows netdev_pick_tx() to choose a TX queue based on XPS settings, instead of reusing the queue chosen at the time the first packet was sent for connected sockets. Tested: 500 concurrent UDP_RR connected UDP flows, host with 32 TX queues, 512 cpus, XPS setup. super_netperf 500 -t UDP_RR -H -l 1000 -- -r 100,100 -Nn & This patch saves between 10% and 20% of cycles, depending on how process scheduler migrates threads among cpus. Using following bpftrace script, we can see the effect on Qdisc/NIC tx queues being better used (less cache line misses). bpftrace -e ' k:__dev_queue_xmit { @start[cpu] = nsecs; } kr:__dev_queue_xmit { if (@start[cpu]) { $delay = nsecs - @start[cpu]; delete(@start[cpu]); @__dev_queue_xmit_ns = hist($delay); } } END { clear(@start); }' Before: @__dev_queue_xmit_ns: [128, 256) 6 | | [256, 512) 116283 | | [512, 1K) 1888205 |@@@@@@@@@@@ | [1K, 2K) 8106167 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ | [2K, 4K) 8699293 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@| [4K, 8K) 2600676 |@@@@@@@@@@@@@@@ | [8K, 16K) 721688 |@@@@ | [16K, 32K) 122995 | | [32K, 64K) 10639 | | [64K, 128K) 119 | | [128K, 256K) 1 | | After: @__dev_queue_xmit_ns: [128, 256) 3 | | [256, 512) 651112 |@@ | [512, 1K) 8109938 |@@@@@@@@@@@@@@@@@@@@@@@@@@ | [1K, 2K) 16081031 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@| [2K, 4K) 2411692 |@@@@@@@ | [4K, 8K) 98994 | | [8K, 16K) 1536 | | [16K, 32K) 587 | | [32K, 64K) 2 | | Signed-off-by: Eric Dumazet Reviewed-by: Neal Cardwell Reviewed-by: Kuniyuki Iwashima Link: https://patch.msgid.link/20251013152234.842065-3-edumazet@google.com Signed-off-by: Jakub Kicinski --- net/core/sock.c | 12 +++++++++++- 1 file changed, 11 insertions(+), 1 deletion(-) diff --git a/net/core/sock.c b/net/core/sock.c index 542cfa16ee12..08ae20069b6d 100644 --- a/net/core/sock.c +++ b/net/core/sock.c @@ -2694,6 +2694,8 @@ void __sock_wfree(struct sk_buff *skb) void skb_set_owner_w(struct sk_buff *skb, struct sock *sk) { + int old_wmem; + skb_orphan(skb); #ifdef CONFIG_INET if (unlikely(!sk_fullsock(sk))) @@ -2707,7 +2709,15 @@ void skb_set_owner_w(struct sk_buff *skb, struct sock *sk) * is enough to guarantee sk_free() won't free this sock until * all in-flight packets are completed */ - refcount_add(skb->truesize, &sk->sk_wmem_alloc); + __refcount_add(skb->truesize, &sk->sk_wmem_alloc, &old_wmem); + + /* (old_wmem == SK_WMEM_ALLOC_BIAS) if no other TX packet for this socket + * is in a host queue (qdisc, NIC queue). + * Set skb->ooo_okay so that netdev_pick_tx() can choose a TX queue + * based on XPS for better performance. + * Otherwise clear ooo_okay to not risk Out Of Order delivery. + */ + skb->ooo_okay = (old_wmem == SK_WMEM_ALLOC_BIAS); } EXPORT_SYMBOL(skb_set_owner_w); From 2ddef3462b3a5d62e5485e22ce128a5c02276438 Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Mon, 13 Oct 2025 15:22:33 +0000 Subject: [PATCH 3/4] net: add /proc/sys/net/core/txq_reselection_ms control Add a new sysctl to control how often a queue reselection can happen even if a flow has a persistent queue of skbs in a Qdisc or NIC queue. A value of zero means the feature is disabled. Default is 1000 (1 second). This sysctl is used in the following patch. Signed-off-by: Eric Dumazet Reviewed-by: Neal Cardwell Reviewed-by: Kuniyuki Iwashima Link: https://patch.msgid.link/20251013152234.842065-4-edumazet@google.com Signed-off-by: Jakub Kicinski --- Documentation/admin-guide/sysctl/net.rst | 17 +++++++++++++++++ include/net/netns/core.h | 1 + net/core/net_namespace.c | 1 + net/core/sysctl_net_core.c | 7 +++++++ 4 files changed, 26 insertions(+) diff --git a/Documentation/admin-guide/sysctl/net.rst b/Documentation/admin-guide/sysctl/net.rst index 2ef50828aff1..40749b3cd356 100644 --- a/Documentation/admin-guide/sysctl/net.rst +++ b/Documentation/admin-guide/sysctl/net.rst @@ -406,6 +406,23 @@ to SOCK_TXREHASH_DEFAULT (i. e. not overridden by setsockopt). If set to 1 (default), hash rethink is performed on listening socket. If set to 0, hash rethink is not performed. +txq_reselection_ms +------------------ + +Controls how often (in ms) a busy connected flow can select another tx queue. + +A resection is desirable when/if user thread has migrated and XPS +would select a different queue. Same can occur without XPS +if the flow hash has changed. + +But switching txq can introduce reorders, especially if the +old queue is under high pressure. Modern TCP stacks deal +well with reorders if they happen not too often. + +To disable this feature, set the value to 0. + +Default : 1000 + gro_normal_batch ---------------- diff --git a/include/net/netns/core.h b/include/net/netns/core.h index 9b36f0ff0c20..cb9c3e4cd738 100644 --- a/include/net/netns/core.h +++ b/include/net/netns/core.h @@ -13,6 +13,7 @@ struct netns_core { struct ctl_table_header *sysctl_hdr; int sysctl_somaxconn; + int sysctl_txq_reselection; int sysctl_optmem_max; u8 sysctl_txrehash; u8 sysctl_tstamp_allow_data; diff --git a/net/core/net_namespace.c b/net/core/net_namespace.c index b0e0f22d7b21..adcfef55a66f 100644 --- a/net/core/net_namespace.c +++ b/net/core/net_namespace.c @@ -395,6 +395,7 @@ static __net_init void preinit_net_sysctl(struct net *net) net->core.sysctl_optmem_max = 128 * 1024; net->core.sysctl_txrehash = SOCK_TXREHASH_ENABLED; net->core.sysctl_tstamp_allow_data = 1; + net->core.sysctl_txq_reselection = msecs_to_jiffies(1000); } /* init code that must occur even if setup_net() is not called. */ diff --git a/net/core/sysctl_net_core.c b/net/core/sysctl_net_core.c index 8cf04b57ade1..f79137826d7f 100644 --- a/net/core/sysctl_net_core.c +++ b/net/core/sysctl_net_core.c @@ -667,6 +667,13 @@ static struct ctl_table netns_core_table[] = { .extra2 = SYSCTL_ONE, .proc_handler = proc_dou8vec_minmax, }, + { + .procname = "txq_reselection_ms", + .data = &init_net.core.sysctl_txq_reselection, + .maxlen = sizeof(int), + .mode = 0644, + .proc_handler = proc_dointvec_ms_jiffies, + }, { .procname = "tstamp_allow_data", .data = &init_net.core.sysctl_tstamp_allow_data, From 4a7708443dec13b074bc43855f494358fedbd3c0 Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Mon, 13 Oct 2025 15:22:34 +0000 Subject: [PATCH 4/4] net: allow busy connected flows to switch tx queues This is a followup of commit 726e9e8b94b9 ("tcp: refine skb->ooo_okay setting") and of prior commit in this series ("net: control skb->ooo_okay from skb_set_owner_w()") skb->ooo_okay might never be set for bulk flows that always have at least one skb in a qdisc queue of NIC queue, especially if TX completion is delayed because of a stressed cpu. The so-called "strange attractors" has caused many performance issues (see for instance 9b462d02d6dd ("tcp: TCP Small Queues and strange attractors")), we need to do better. We have tried very hard to avoid reorders because TCP was not dealing with them nicely a decade ago. Use the new net.core.txq_reselection_ms sysctl to let flows follow XPS and select a more efficient queue. After this patch, we no longer have to make sure threads are pinned to cpus, they now can be migrated without adding too much spinlock/qdisc/TX completion pressure anymore. TX completion part was problematic, because it added false sharing on various socket fields, but also added false sharing and spinlock contention in mm layers. Calling skb_orphan() from ndo_start_xmit() is not an option unfortunately. Note for later: 1) move sk->sk_tx_queue_mapping closer to sk_tx_queue_mapping_jiffies for better cache locality. 2) Study if 9b462d02d6dd ("tcp: TCP Small Queues and strange attractors") could be revised. Tested: Used a host with 32 TX queues, shared by groups of 8 cores. XPS setup : echo ff >/sys/class/net/eth1/queue/tx-0/xps_cpus echo ff00 >/sys/class/net/eth1/queue/tx-1/xps_cpus echo ff0000 >/sys/class/net/eth1/queue/tx-2/xps_cpus echo ff000000 >/sys/class/net/eth1/queue/tx-3/xps_cpus echo ff,00000000 >/sys/class/net/eth1/queue/tx-4/xps_cpus echo ff00,00000000 >/sys/class/net/eth1/queue/tx-5/xps_cpus echo ff0000,00000000 >/sys/class/net/eth1/queue/tx-6/xps_cpus echo ff000000,00000000 >/sys/class/net/eth1/queue/tx-7/xps_cpus ... Launched a tcp_stream with 15 threads and 1000 flows, initially affined to core 0-15 taskset -c 0-15 tcp_stream -T15 -F1000 -l1000 -c -H target_host Checked that only queues 0 and 1 are used as instructed by XPS : tc -s qdisc show dev eth1|grep backlog|grep -v "backlog 0b 0p" backlog 123489410b 1890p backlog 69809026b 1064p backlog 52401054b 805p Then force each thread to run on cpu 1,9,17,25,33,41,49,57,65,73,81,89,97,105,113,121 C=1;PID=`pidof tcp_stream`;for P in `ls /proc/$PID/task`; do taskset -pc $C $P; C=$(($C + 8));done Set txq_reselection_ms to 1000 echo 1000 > /proc/sys/net/core/txq_reselection_ms Check that the flows have migrated nicely: tc -s qdisc show dev eth1|grep backlog|grep -v "backlog 0b 0p" backlog 130508314b 1916p backlog 8584380b 126p backlog 8584380b 126p backlog 8379990b 123p backlog 8584380b 126p backlog 8487484b 125p backlog 8584380b 126p backlog 8448120b 124p backlog 8584380b 126p backlog 8720640b 128p backlog 8856900b 130p backlog 8584380b 126p backlog 8652510b 127p backlog 8448120b 124p backlog 8516250b 125p backlog 7834950b 115p Signed-off-by: Eric Dumazet Reviewed-by: Kuniyuki Iwashima Link: https://patch.msgid.link/20251013152234.842065-5-edumazet@google.com Signed-off-by: Jakub Kicinski --- include/net/sock.h | 26 ++++++++++++-------------- net/core/dev.c | 29 +++++++++++++++++++++++++++-- 2 files changed, 39 insertions(+), 16 deletions(-) diff --git a/include/net/sock.h b/include/net/sock.h index 2794bc5c5654..f0d00928db9e 100644 --- a/include/net/sock.h +++ b/include/net/sock.h @@ -313,6 +313,7 @@ struct sk_filter; * @sk_bind_phc: SO_TIMESTAMPING bind PHC index of PTP virtual clock * for timestamping * @sk_tskey: counter to disambiguate concurrent tstamp requests + * @sk_tx_queue_mapping_jiffies: time in jiffies of last @sk_tx_queue_mapping refresh. * @sk_zckey: counter to order MSG_ZEROCOPY notifications * @sk_socket: Identd and reporting IO signals * @sk_user_data: RPC layer private data. Write-protected by @sk_callback_lock. @@ -485,6 +486,7 @@ struct sock { unsigned long sk_pacing_rate; /* bytes per second */ atomic_t sk_zckey; atomic_t sk_tskey; + unsigned long sk_tx_queue_mapping_jiffies; __cacheline_group_end(sock_write_tx); __cacheline_group_begin(sock_read_tx); @@ -1992,7 +1994,15 @@ static inline void sk_tx_queue_set(struct sock *sk, int tx_queue) /* Paired with READ_ONCE() in sk_tx_queue_get() and * other WRITE_ONCE() because socket lock might be not held. */ - WRITE_ONCE(sk->sk_tx_queue_mapping, tx_queue); + if (READ_ONCE(sk->sk_tx_queue_mapping) != tx_queue) { + WRITE_ONCE(sk->sk_tx_queue_mapping, tx_queue); + WRITE_ONCE(sk->sk_tx_queue_mapping_jiffies, jiffies); + return; + } + + /* Refresh sk_tx_queue_mapping_jiffies if too old. */ + if (time_is_before_jiffies(READ_ONCE(sk->sk_tx_queue_mapping_jiffies) + HZ)) + WRITE_ONCE(sk->sk_tx_queue_mapping_jiffies, jiffies); } #define NO_QUEUE_MAPPING USHRT_MAX @@ -2005,19 +2015,7 @@ static inline void sk_tx_queue_clear(struct sock *sk) WRITE_ONCE(sk->sk_tx_queue_mapping, NO_QUEUE_MAPPING); } -static inline int sk_tx_queue_get(const struct sock *sk) -{ - if (sk) { - /* Paired with WRITE_ONCE() in sk_tx_queue_clear() - * and sk_tx_queue_set(). - */ - int val = READ_ONCE(sk->sk_tx_queue_mapping); - - if (val != NO_QUEUE_MAPPING) - return val; - } - return -1; -} +int sk_tx_queue_get(const struct sock *sk); static inline void __sk_rx_queue_set(struct sock *sk, const struct sk_buff *skb, diff --git a/net/core/dev.c b/net/core/dev.c index a64cef2c537e..33e6101dbc45 100644 --- a/net/core/dev.c +++ b/net/core/dev.c @@ -4591,6 +4591,32 @@ u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, } EXPORT_SYMBOL(dev_pick_tx_zero); +int sk_tx_queue_get(const struct sock *sk) +{ + int resel, val; + + if (!sk) + return -1; + /* Paired with WRITE_ONCE() in sk_tx_queue_clear() + * and sk_tx_queue_set(). + */ + val = READ_ONCE(sk->sk_tx_queue_mapping); + + if (val == NO_QUEUE_MAPPING) + return -1; + + if (!sk_fullsock(sk)) + return val; + + resel = READ_ONCE(sock_net(sk)->core.sysctl_txq_reselection); + if (resel && time_is_before_jiffies( + READ_ONCE(sk->sk_tx_queue_mapping_jiffies) + resel)) + return -1; + + return val; +} +EXPORT_SYMBOL(sk_tx_queue_get); + u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, struct net_device *sb_dev) { @@ -4606,8 +4632,7 @@ u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, if (new_index < 0) new_index = skb_tx_hash(dev, sb_dev, skb); - if (queue_index != new_index && sk && - sk_fullsock(sk) && + if (sk && sk_fullsock(sk) && rcu_access_pointer(sk->sk_dst_cache)) sk_tx_queue_set(sk, new_index);