1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
|
From: Felix Fietkau <nbd@nbd.name>
Date: Thu, 16 Feb 2023 18:39:04 +0100
Subject: [PATCH] net/core: add optional threading for backlog processing
When dealing with few flows or an imbalance on CPU utilization, static RPS
CPU assignment can be too inflexible. Add support for enabling threaded NAPI
for backlog processing in order to allow the scheduler to better balance
processing. This helps better spread the load across idle CPUs.
Signed-off-by: Felix Fietkau <nbd@nbd.name>
---
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -558,6 +558,7 @@ static inline bool napi_complete(struct
}
int dev_set_threaded(struct net_device *dev, bool threaded);
+int backlog_set_threaded(bool threaded);
/**
* napi_disable - prevent NAPI from scheduling
@@ -3238,6 +3239,7 @@ struct softnet_data {
/* stats */
unsigned int processed;
unsigned int time_squeeze;
+ unsigned int process_queue_empty;
#ifdef CONFIG_RPS
struct softnet_data *rps_ipi_list;
#endif
--- a/net/core/dev.c
+++ b/net/core/dev.c
@@ -4729,7 +4729,7 @@ static void napi_schedule_rps(struct sof
struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
#ifdef CONFIG_RPS
- if (sd != mysd) {
+ if (sd != mysd && !test_bit(NAPI_STATE_THREADED, &sd->backlog.state)) {
sd->rps_ipi_next = mysd->rps_ipi_list;
mysd->rps_ipi_list = sd;
@@ -5848,6 +5848,8 @@ static DEFINE_PER_CPU(struct work_struct
/* Network device is going away, flush any packets still pending */
static void flush_backlog(struct work_struct *work)
{
+ unsigned int process_queue_empty;
+ bool threaded, flush_processq;
struct sk_buff *skb, *tmp;
struct softnet_data *sd;
@@ -5862,8 +5864,17 @@ static void flush_backlog(struct work_st
input_queue_head_incr(sd);
}
}
+
+ threaded = test_bit(NAPI_STATE_THREADED, &sd->backlog.state);
+ flush_processq = threaded &&
+ !skb_queue_empty_lockless(&sd->process_queue);
+ if (flush_processq)
+ process_queue_empty = sd->process_queue_empty;
rps_unlock_irq_enable(sd);
+ if (threaded)
+ goto out;
+
skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
if (skb->dev->reg_state == NETREG_UNREGISTERING) {
__skb_unlink(skb, &sd->process_queue);
@@ -5871,7 +5882,16 @@ static void flush_backlog(struct work_st
input_queue_head_incr(sd);
}
}
+
+out:
local_bh_enable();
+
+ while (flush_processq) {
+ msleep(1);
+ rps_lock_irq_disable(sd);
+ flush_processq = process_queue_empty == sd->process_queue_empty;
+ rps_unlock_irq_enable(sd);
+ }
}
static bool flush_required(int cpu)
@@ -6003,6 +6023,7 @@ static int process_backlog(struct napi_s
}
rps_lock_irq_disable(sd);
+ sd->process_queue_empty++;
if (skb_queue_empty(&sd->input_pkt_queue)) {
/*
* Inline a custom version of __napi_complete().
@@ -6012,7 +6033,8 @@ static int process_backlog(struct napi_s
* We can use a plain write instead of clear_bit(),
* and we dont need an smp_mb() memory barrier.
*/
- napi->state = 0;
+ napi->state &= ~(NAPIF_STATE_SCHED |
+ NAPIF_STATE_SCHED_THREADED);
again = false;
} else {
skb_queue_splice_tail_init(&sd->input_pkt_queue,
@@ -6426,6 +6448,55 @@ int dev_set_threaded(struct net_device *
}
EXPORT_SYMBOL(dev_set_threaded);
+int backlog_set_threaded(bool threaded)
+{
+ static bool backlog_threaded;
+ int err = 0;
+ int i;
+
+ if (backlog_threaded == threaded)
+ return 0;
+
+ for_each_possible_cpu(i) {
+ struct softnet_data *sd = &per_cpu(softnet_data, i);
+ struct napi_struct *n = &sd->backlog;
+
+ if (n->thread)
+ continue;
+ n->thread = kthread_run(napi_threaded_poll, n, "napi/backlog-%d", i);
+ if (IS_ERR(n->thread)) {
+ err = PTR_ERR(n->thread);
+ pr_err("kthread_run failed with err %d\n", err);
+ n->thread = NULL;
+ threaded = false;
+ break;
+ }
+
+ }
+
+ backlog_threaded = threaded;
+
+ /* Make sure kthread is created before THREADED bit
+ * is set.
+ */
+ smp_mb__before_atomic();
+
+ for_each_possible_cpu(i) {
+ struct softnet_data *sd = &per_cpu(softnet_data, i);
+ struct napi_struct *n = &sd->backlog;
+ unsigned long flags;
+
+ rps_lock_irqsave(sd, &flags);
+ if (threaded)
+ n->state |= NAPIF_STATE_THREADED;
+ else
+ n->state &= ~NAPIF_STATE_THREADED;
+ rps_unlock_irq_restore(sd, &flags);
+ }
+
+ return err;
+}
+
void netif_napi_add_weight(struct net_device *dev, struct napi_struct *napi,
int (*poll)(struct napi_struct *, int), int weight)
{
@@ -11351,6 +11422,9 @@ static int dev_cpu_dead(unsigned int old
raise_softirq_irqoff(NET_TX_SOFTIRQ);
local_irq_enable();
+ if (test_bit(NAPI_STATE_THREADED, &oldsd->backlog.state))
+ return 0;
+
#ifdef CONFIG_RPS
remsd = oldsd->rps_ipi_list;
oldsd->rps_ipi_list = NULL;
@@ -11666,6 +11740,7 @@ static int __init net_dev_init(void)
INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd);
spin_lock_init(&sd->defer_lock);
+ INIT_LIST_HEAD(&sd->backlog.poll_list);
init_gro_hash(&sd->backlog);
sd->backlog.poll = process_backlog;
sd->backlog.weight = weight_p;
--- a/net/core/sysctl_net_core.c
+++ b/net/core/sysctl_net_core.c
@@ -30,6 +30,7 @@ static int int_3600 = 3600;
static int min_sndbuf = SOCK_MIN_SNDBUF;
static int min_rcvbuf = SOCK_MIN_RCVBUF;
static int max_skb_frags = MAX_SKB_FRAGS;
+static int backlog_threaded;
static int net_msg_warn; /* Unused, but still a sysctl */
@@ -188,6 +189,23 @@ static int rps_sock_flow_sysctl(struct c
}
#endif /* CONFIG_RPS */
+static int backlog_threaded_sysctl(struct ctl_table *table, int write,
+ void *buffer, size_t *lenp, loff_t *ppos)
+{
+ static DEFINE_MUTEX(backlog_threaded_mutex);
+ int ret;
+
+ mutex_lock(&backlog_threaded_mutex);
+
+ ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
+ if (write && !ret)
+ ret = backlog_set_threaded(backlog_threaded);
+
+ mutex_unlock(&backlog_threaded_mutex);
+
+ return ret;
+}
+
#ifdef CONFIG_NET_FLOW_LIMIT
static DEFINE_MUTEX(flow_limit_update_mutex);
@@ -532,6 +550,15 @@ static struct ctl_table net_core_table[]
.proc_handler = rps_sock_flow_sysctl
},
#endif
+ {
+ .procname = "backlog_threaded",
+ .data = &backlog_threaded,
+ .maxlen = sizeof(unsigned int),
+ .mode = 0644,
+ .proc_handler = backlog_threaded_sysctl,
+ .extra1 = SYSCTL_ZERO,
+ .extra2 = SYSCTL_ONE
+ },
#ifdef CONFIG_NET_FLOW_LIMIT
{
.procname = "flow_limit_cpu_bitmap",
|
