Skip to content

Commit 5628f3f

Browse files
edumazetPaolo Abeni
edumazet
authored and
Paolo Abeni
committed
net: add NUMA awareness to skb_attempt_defer_free()
Instead of sharing sd->defer_list & sd->defer_count with many cpus, add one pair for each NUMA node. Signed-off-by: Eric Dumazet <edumazet@google.com> Reviewed-by: Jason Xing <kerneljasonxing@gmail.com> Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com> Link: https://patch.msgid.link/20250928084934.3266948-4-edumazet@google.com Signed-off-by: Paolo Abeni <pabeni@redhat.com>
1 parent 844c9db commit 5628f3f

File tree

5 files changed

+37
-22
lines changed

5 files changed

+37
-22
lines changed

‎include/linux/netdevice.h‎

Lines changed: 0 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -3536,10 +3536,6 @@ struct softnet_data {
35363536

35373537
struct numa_drop_counters drop_counters;
35383538

3539-
/* Another possibly contended cache line */
3540-
struct llist_head defer_list ____cacheline_aligned_in_smp;
3541-
atomic_long_t defer_count;
3542-
35433539
int defer_ipi_scheduled ____cacheline_aligned_in_smp;
35443540
call_single_data_t defer_csd;
35453541
};

‎include/net/hotdata.h‎

Lines changed: 7 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -2,10 +2,16 @@
22
#ifndef _NET_HOTDATA_H
33
#define _NET_HOTDATA_H
44

5+
#include <linux/llist.h>
56
#include <linux/types.h>
67
#include <linux/netdevice.h>
78
#include <net/protocol.h>
89

10+
struct skb_defer_node {
11+
struct llist_head defer_list;
12+
atomic_long_t defer_count;
13+
} ____cacheline_aligned_in_smp;
14+
915
/* Read mostly data used in network fast paths. */
1016
struct net_hotdata {
1117
#if IS_ENABLED(CONFIG_INET)
@@ -30,6 +36,7 @@ struct net_hotdata {
3036
struct rps_sock_flow_table __rcu *rps_sock_flow_table;
3137
u32 rps_cpu_mask;
3238
#endif
39+
struct skb_defer_node __percpu *skb_defer_nodes;
3340
int gro_normal_batch;
3441
int netdev_budget;
3542
int netdev_budget_usecs;

‎net/core/dev.c‎

Lines changed: 23 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -5180,8 +5180,9 @@ static void napi_schedule_rps(struct softnet_data *sd)
51805180
__napi_schedule_irqoff(&mysd->backlog);
51815181
}
51825182

5183-
void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu)
5183+
void kick_defer_list_purge(unsigned int cpu)
51845184
{
5185+
struct softnet_data *sd = &per_cpu(softnet_data, cpu);
51855186
unsigned long flags;
51865187

51875188
if (use_backlog_threads()) {
@@ -6715,18 +6716,24 @@ bool napi_complete_done(struct napi_struct *n, int work_done)
67156716
}
67166717
EXPORT_SYMBOL(napi_complete_done);
67176718

6718-
static void skb_defer_free_flush(struct softnet_data *sd)
6719+
static void skb_defer_free_flush(void)
67196720
{
67206721
struct llist_node *free_list;
67216722
struct sk_buff *skb, *next;
6723+
struct skb_defer_node *sdn;
6724+
int node;
67226725

6723-
if (llist_empty(&sd->defer_list))
6724-
return;
6725-
atomic_long_set(&sd->defer_count, 0);
6726-
free_list = llist_del_all(&sd->defer_list);
6726+
for_each_node(node) {
6727+
sdn = this_cpu_ptr(net_hotdata.skb_defer_nodes) + node;
6728+
6729+
if (llist_empty(&sdn->defer_list))
6730+
continue;
6731+
atomic_long_set(&sdn->defer_count, 0);
6732+
free_list = llist_del_all(&sdn->defer_list);
67276733

6728-
llist_for_each_entry_safe(skb, next, free_list, ll_node) {
6729-
napi_consume_skb(skb, 1);
6734+
llist_for_each_entry_safe(skb, next, free_list, ll_node) {
6735+
napi_consume_skb(skb, 1);
6736+
}
67306737
}
67316738
}
67326739

@@ -6854,7 +6861,7 @@ static void __napi_busy_loop(unsigned int napi_id,
68546861
if (work > 0)
68556862
__NET_ADD_STATS(dev_net(napi->dev),
68566863
LINUX_MIB_BUSYPOLLRXPACKETS, work);
6857-
skb_defer_free_flush(this_cpu_ptr(&softnet_data));
6864+
skb_defer_free_flush();
68586865
bpf_net_ctx_clear(bpf_net_ctx);
68596866
local_bh_enable();
68606867

@@ -7713,7 +7720,7 @@ static void napi_threaded_poll_loop(struct napi_struct *napi)
77137720
local_irq_disable();
77147721
net_rps_action_and_irq_enable(sd);
77157722
}
7716-
skb_defer_free_flush(sd);
7723+
skb_defer_free_flush();
77177724
bpf_net_ctx_clear(bpf_net_ctx);
77187725
local_bh_enable();
77197726

@@ -7755,7 +7762,7 @@ static __latent_entropy void net_rx_action(void)
77557762
for (;;) {
77567763
struct napi_struct *n;
77577764

7758-
skb_defer_free_flush(sd);
7765+
skb_defer_free_flush();
77597766

77607767
if (list_empty(&list)) {
77617768
if (list_empty(&repoll)) {
@@ -12989,7 +12996,6 @@ static int __init net_dev_init(void)
1298912996
sd->cpu = i;
1299012997
#endif
1299112998
INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd);
12992-
init_llist_head(&sd->defer_list);
1299312999

1299413000
gro_init(&sd->backlog.gro);
1299513001
sd->backlog.poll = process_backlog;
@@ -12999,6 +13005,11 @@ static int __init net_dev_init(void)
1299913005
if (net_page_pool_create(i))
1300013006
goto out;
1300113007
}
13008+
net_hotdata.skb_defer_nodes =
13009+
__alloc_percpu(sizeof(struct skb_defer_node) * nr_node_ids,
13010+
__alignof__(struct skb_defer_node));
13011+
if (!net_hotdata.skb_defer_nodes)
13012+
goto out;
1300213013
if (use_backlog_threads())
1300313014
smpboot_register_percpu_thread(&backlog_threads);
1300413015

‎net/core/dev.h‎

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -357,7 +357,7 @@ static inline void napi_assert_will_not_race(const struct napi_struct *napi)
357357
WARN_ON(READ_ONCE(napi->list_owner) != -1);
358358
}
359359

360-
void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu);
360+
void kick_defer_list_purge(unsigned int cpu);
361361

362362
#define XMIT_RECURSION_LIMIT 8
363363

‎net/core/skbuff.c‎

Lines changed: 6 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -7185,9 +7185,9 @@ static void kfree_skb_napi_cache(struct sk_buff *skb)
71857185
*/
71867186
void skb_attempt_defer_free(struct sk_buff *skb)
71877187
{
7188+
struct skb_defer_node *sdn;
71887189
unsigned long defer_count;
71897190
int cpu = skb->alloc_cpu;
7190-
struct softnet_data *sd;
71917191
unsigned int defer_max;
71927192
bool kick;
71937193

@@ -7201,14 +7201,15 @@ nodefer: kfree_skb_napi_cache(skb);
72017201
DEBUG_NET_WARN_ON_ONCE(skb_dst(skb));
72027202
DEBUG_NET_WARN_ON_ONCE(skb->destructor);
72037203

7204-
sd = &per_cpu(softnet_data, cpu);
7204+
sdn = per_cpu_ptr(net_hotdata.skb_defer_nodes, cpu) + numa_node_id();
7205+
72057206
defer_max = READ_ONCE(net_hotdata.sysctl_skb_defer_max);
7206-
defer_count = atomic_long_inc_return(&sd->defer_count);
7207+
defer_count = atomic_long_inc_return(&sdn->defer_count);
72077208

72087209
if (defer_count >= defer_max)
72097210
goto nodefer;
72107211

7211-
llist_add(&skb->ll_node, &sd->defer_list);
7212+
llist_add(&skb->ll_node, &sdn->defer_list);
72127213

72137214
/* Send an IPI every time queue reaches half capacity. */
72147215
kick = (defer_count - 1) == (defer_max >> 1);
@@ -7217,7 +7218,7 @@ nodefer: kfree_skb_napi_cache(skb);
72177218
* if we are unlucky enough (this seems very unlikely).
72187219
*/
72197220
if (unlikely(kick))
7220-
kick_defer_list_purge(sd, cpu);
7221+
kick_defer_list_purge(cpu);
72217222
}
72227223

72237224
static void skb_splice_csum_page(struct sk_buff *skb, struct page *page,

0 commit comments

Comments
 (0)