Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* -*- linux-c -*- |
2 | * sysctl_net_core.c: sysctl interface to net core subsystem. | |
3 | * | |
4 | * Begun April 1, 1996, Mike Shaver. | |
5 | * Added /proc/sys/net/core directory entry (empty =) ). [MS] | |
6 | */ | |
7 | ||
8 | #include <linux/mm.h> | |
9 | #include <linux/sysctl.h> | |
1da177e4 | 10 | #include <linux/module.h> |
20380731 | 11 | #include <linux/socket.h> |
a37ae408 | 12 | #include <linux/netdevice.h> |
3fff4c42 | 13 | #include <linux/ratelimit.h> |
fec5e652 | 14 | #include <linux/vmalloc.h> |
33eb9cfc | 15 | #include <linux/init.h> |
5a0e3ad6 | 16 | #include <linux/slab.h> |
7426a564 | 17 | #include <linux/kmemleak.h> |
3fff4c42 | 18 | |
63d819ca | 19 | #include <net/ip.h> |
20380731 | 20 | #include <net/sock.h> |
c5c177b4 | 21 | #include <net/net_ratelimit.h> |
076bb0c8 | 22 | #include <net/busy_poll.h> |
1da177e4 | 23 | |
5f671d6b | 24 | static int zero = 0; |
cdda8891 | 25 | static int one = 1; |
5f671d6b | 26 | static int ushort_max = USHRT_MAX; |
cdda8891 | 27 | |
fec5e652 | 28 | #ifdef CONFIG_RPS |
fe2c6338 | 29 | static int rps_sock_flow_sysctl(struct ctl_table *table, int write, |
fec5e652 TH |
30 | void __user *buffer, size_t *lenp, loff_t *ppos) |
31 | { | |
32 | unsigned int orig_size, size; | |
33 | int ret, i; | |
fe2c6338 | 34 | struct ctl_table tmp = { |
fec5e652 TH |
35 | .data = &size, |
36 | .maxlen = sizeof(size), | |
37 | .mode = table->mode | |
38 | }; | |
39 | struct rps_sock_flow_table *orig_sock_table, *sock_table; | |
40 | static DEFINE_MUTEX(sock_flow_mutex); | |
41 | ||
42 | mutex_lock(&sock_flow_mutex); | |
43 | ||
6e3f7faf ED |
44 | orig_sock_table = rcu_dereference_protected(rps_sock_flow_table, |
45 | lockdep_is_held(&sock_flow_mutex)); | |
fec5e652 TH |
46 | size = orig_size = orig_sock_table ? orig_sock_table->mask + 1 : 0; |
47 | ||
48 | ret = proc_dointvec(&tmp, write, buffer, lenp, ppos); | |
49 | ||
50 | if (write) { | |
51 | if (size) { | |
52 | if (size > 1<<30) { | |
53 | /* Enforce limit to prevent overflow */ | |
54 | mutex_unlock(&sock_flow_mutex); | |
55 | return -EINVAL; | |
56 | } | |
57 | size = roundup_pow_of_two(size); | |
58 | if (size != orig_size) { | |
59 | sock_table = | |
60 | vmalloc(RPS_SOCK_FLOW_TABLE_SIZE(size)); | |
61 | if (!sock_table) { | |
62 | mutex_unlock(&sock_flow_mutex); | |
63 | return -ENOMEM; | |
64 | } | |
65 | ||
66 | sock_table->mask = size - 1; | |
67 | } else | |
68 | sock_table = orig_sock_table; | |
69 | ||
70 | for (i = 0; i < size; i++) | |
71 | sock_table->ents[i] = RPS_NO_CPU; | |
72 | } else | |
73 | sock_table = NULL; | |
74 | ||
75 | if (sock_table != orig_sock_table) { | |
76 | rcu_assign_pointer(rps_sock_flow_table, sock_table); | |
adc9300e | 77 | if (sock_table) |
c5905afb | 78 | static_key_slow_inc(&rps_needed); |
adc9300e | 79 | if (orig_sock_table) { |
c5905afb | 80 | static_key_slow_dec(&rps_needed); |
adc9300e ED |
81 | synchronize_rcu(); |
82 | vfree(orig_sock_table); | |
83 | } | |
fec5e652 TH |
84 | } |
85 | } | |
86 | ||
87 | mutex_unlock(&sock_flow_mutex); | |
88 | ||
89 | return ret; | |
90 | } | |
91 | #endif /* CONFIG_RPS */ | |
92 | ||
99bbc707 WB |
93 | #ifdef CONFIG_NET_FLOW_LIMIT |
94 | static DEFINE_MUTEX(flow_limit_update_mutex); | |
95 | ||
fe2c6338 | 96 | static int flow_limit_cpu_sysctl(struct ctl_table *table, int write, |
99bbc707 WB |
97 | void __user *buffer, size_t *lenp, |
98 | loff_t *ppos) | |
99 | { | |
100 | struct sd_flow_limit *cur; | |
101 | struct softnet_data *sd; | |
102 | cpumask_var_t mask; | |
103 | int i, len, ret = 0; | |
104 | ||
105 | if (!alloc_cpumask_var(&mask, GFP_KERNEL)) | |
106 | return -ENOMEM; | |
107 | ||
108 | if (write) { | |
109 | ret = cpumask_parse_user(buffer, *lenp, mask); | |
110 | if (ret) | |
111 | goto done; | |
112 | ||
113 | mutex_lock(&flow_limit_update_mutex); | |
114 | len = sizeof(*cur) + netdev_flow_limit_table_len; | |
115 | for_each_possible_cpu(i) { | |
116 | sd = &per_cpu(softnet_data, i); | |
117 | cur = rcu_dereference_protected(sd->flow_limit, | |
118 | lockdep_is_held(&flow_limit_update_mutex)); | |
119 | if (cur && !cpumask_test_cpu(i, mask)) { | |
120 | RCU_INIT_POINTER(sd->flow_limit, NULL); | |
121 | synchronize_rcu(); | |
122 | kfree(cur); | |
123 | } else if (!cur && cpumask_test_cpu(i, mask)) { | |
124 | cur = kzalloc(len, GFP_KERNEL); | |
125 | if (!cur) { | |
126 | /* not unwinding previous changes */ | |
127 | ret = -ENOMEM; | |
128 | goto write_unlock; | |
129 | } | |
130 | cur->num_buckets = netdev_flow_limit_table_len; | |
131 | rcu_assign_pointer(sd->flow_limit, cur); | |
132 | } | |
133 | } | |
134 | write_unlock: | |
135 | mutex_unlock(&flow_limit_update_mutex); | |
136 | } else { | |
5f121b9a WB |
137 | char kbuf[128]; |
138 | ||
99bbc707 WB |
139 | if (*ppos || !*lenp) { |
140 | *lenp = 0; | |
141 | goto done; | |
142 | } | |
143 | ||
144 | cpumask_clear(mask); | |
145 | rcu_read_lock(); | |
146 | for_each_possible_cpu(i) { | |
147 | sd = &per_cpu(softnet_data, i); | |
148 | if (rcu_dereference(sd->flow_limit)) | |
149 | cpumask_set_cpu(i, mask); | |
150 | } | |
151 | rcu_read_unlock(); | |
152 | ||
5f121b9a WB |
153 | len = min(sizeof(kbuf) - 1, *lenp); |
154 | len = cpumask_scnprintf(kbuf, len, mask); | |
155 | if (!len) { | |
156 | *lenp = 0; | |
157 | goto done; | |
158 | } | |
159 | if (len < *lenp) | |
160 | kbuf[len++] = '\n'; | |
161 | if (copy_to_user(buffer, kbuf, len)) { | |
162 | ret = -EFAULT; | |
163 | goto done; | |
164 | } | |
165 | *lenp = len; | |
166 | *ppos += len; | |
99bbc707 WB |
167 | } |
168 | ||
169 | done: | |
170 | free_cpumask_var(mask); | |
171 | return ret; | |
172 | } | |
173 | ||
fe2c6338 | 174 | static int flow_limit_table_len_sysctl(struct ctl_table *table, int write, |
99bbc707 WB |
175 | void __user *buffer, size_t *lenp, |
176 | loff_t *ppos) | |
177 | { | |
178 | unsigned int old, *ptr; | |
179 | int ret; | |
180 | ||
181 | mutex_lock(&flow_limit_update_mutex); | |
182 | ||
183 | ptr = table->data; | |
184 | old = *ptr; | |
185 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
186 | if (!ret && write && !is_power_of_2(*ptr)) { | |
187 | *ptr = old; | |
188 | ret = -EINVAL; | |
189 | } | |
190 | ||
191 | mutex_unlock(&flow_limit_update_mutex); | |
192 | return ret; | |
193 | } | |
194 | #endif /* CONFIG_NET_FLOW_LIMIT */ | |
195 | ||
33eb9cfc | 196 | static struct ctl_table net_core_table[] = { |
1da177e4 LT |
197 | #ifdef CONFIG_NET |
198 | { | |
1da177e4 LT |
199 | .procname = "wmem_max", |
200 | .data = &sysctl_wmem_max, | |
201 | .maxlen = sizeof(int), | |
202 | .mode = 0644, | |
cdda8891 | 203 | .proc_handler = proc_dointvec_minmax, |
204 | .extra1 = &one, | |
1da177e4 LT |
205 | }, |
206 | { | |
1da177e4 LT |
207 | .procname = "rmem_max", |
208 | .data = &sysctl_rmem_max, | |
209 | .maxlen = sizeof(int), | |
210 | .mode = 0644, | |
cdda8891 | 211 | .proc_handler = proc_dointvec_minmax, |
212 | .extra1 = &one, | |
1da177e4 LT |
213 | }, |
214 | { | |
1da177e4 LT |
215 | .procname = "wmem_default", |
216 | .data = &sysctl_wmem_default, | |
217 | .maxlen = sizeof(int), | |
218 | .mode = 0644, | |
cdda8891 | 219 | .proc_handler = proc_dointvec_minmax, |
220 | .extra1 = &one, | |
1da177e4 LT |
221 | }, |
222 | { | |
1da177e4 LT |
223 | .procname = "rmem_default", |
224 | .data = &sysctl_rmem_default, | |
225 | .maxlen = sizeof(int), | |
226 | .mode = 0644, | |
cdda8891 | 227 | .proc_handler = proc_dointvec_minmax, |
228 | .extra1 = &one, | |
1da177e4 LT |
229 | }, |
230 | { | |
1da177e4 LT |
231 | .procname = "dev_weight", |
232 | .data = &weight_p, | |
233 | .maxlen = sizeof(int), | |
234 | .mode = 0644, | |
6d9f239a | 235 | .proc_handler = proc_dointvec |
1da177e4 LT |
236 | }, |
237 | { | |
1da177e4 LT |
238 | .procname = "netdev_max_backlog", |
239 | .data = &netdev_max_backlog, | |
240 | .maxlen = sizeof(int), | |
241 | .mode = 0644, | |
6d9f239a | 242 | .proc_handler = proc_dointvec |
1da177e4 | 243 | }, |
0a14842f ED |
244 | #ifdef CONFIG_BPF_JIT |
245 | { | |
246 | .procname = "bpf_jit_enable", | |
247 | .data = &bpf_jit_enable, | |
248 | .maxlen = sizeof(int), | |
249 | .mode = 0644, | |
250 | .proc_handler = proc_dointvec | |
251 | }, | |
252 | #endif | |
3b098e2d ED |
253 | { |
254 | .procname = "netdev_tstamp_prequeue", | |
255 | .data = &netdev_tstamp_prequeue, | |
256 | .maxlen = sizeof(int), | |
257 | .mode = 0644, | |
258 | .proc_handler = proc_dointvec | |
259 | }, | |
1da177e4 | 260 | { |
1da177e4 | 261 | .procname = "message_cost", |
717115e1 | 262 | .data = &net_ratelimit_state.interval, |
1da177e4 LT |
263 | .maxlen = sizeof(int), |
264 | .mode = 0644, | |
6d9f239a | 265 | .proc_handler = proc_dointvec_jiffies, |
1da177e4 LT |
266 | }, |
267 | { | |
1da177e4 | 268 | .procname = "message_burst", |
717115e1 | 269 | .data = &net_ratelimit_state.burst, |
1da177e4 LT |
270 | .maxlen = sizeof(int), |
271 | .mode = 0644, | |
6d9f239a | 272 | .proc_handler = proc_dointvec, |
1da177e4 LT |
273 | }, |
274 | { | |
1da177e4 LT |
275 | .procname = "optmem_max", |
276 | .data = &sysctl_optmem_max, | |
277 | .maxlen = sizeof(int), | |
278 | .mode = 0644, | |
6d9f239a | 279 | .proc_handler = proc_dointvec |
1da177e4 | 280 | }, |
fec5e652 TH |
281 | #ifdef CONFIG_RPS |
282 | { | |
283 | .procname = "rps_sock_flow_entries", | |
284 | .maxlen = sizeof(int), | |
285 | .mode = 0644, | |
286 | .proc_handler = rps_sock_flow_sysctl | |
287 | }, | |
288 | #endif | |
99bbc707 WB |
289 | #ifdef CONFIG_NET_FLOW_LIMIT |
290 | { | |
291 | .procname = "flow_limit_cpu_bitmap", | |
292 | .mode = 0644, | |
293 | .proc_handler = flow_limit_cpu_sysctl | |
294 | }, | |
295 | { | |
296 | .procname = "flow_limit_table_len", | |
297 | .data = &netdev_flow_limit_table_len, | |
298 | .maxlen = sizeof(int), | |
299 | .mode = 0644, | |
300 | .proc_handler = flow_limit_table_len_sysctl | |
301 | }, | |
302 | #endif /* CONFIG_NET_FLOW_LIMIT */ | |
e0d1095a | 303 | #ifdef CONFIG_NET_RX_BUSY_POLL |
06021292 | 304 | { |
64b0dc51 ET |
305 | .procname = "busy_poll", |
306 | .data = &sysctl_net_busy_poll, | |
eb6db622 | 307 | .maxlen = sizeof(unsigned int), |
06021292 | 308 | .mode = 0644, |
eb6db622 | 309 | .proc_handler = proc_dointvec |
06021292 | 310 | }, |
2d48d67f | 311 | { |
64b0dc51 ET |
312 | .procname = "busy_read", |
313 | .data = &sysctl_net_busy_read, | |
2d48d67f ET |
314 | .maxlen = sizeof(unsigned int), |
315 | .mode = 0644, | |
316 | .proc_handler = proc_dointvec | |
317 | }, | |
318 | # | |
06021292 | 319 | #endif |
1da177e4 | 320 | #endif /* CONFIG_NET */ |
51b0bded | 321 | { |
51b0bded SH |
322 | .procname = "netdev_budget", |
323 | .data = &netdev_budget, | |
324 | .maxlen = sizeof(int), | |
325 | .mode = 0644, | |
6d9f239a | 326 | .proc_handler = proc_dointvec |
51b0bded | 327 | }, |
a2a316fd | 328 | { |
a2a316fd SH |
329 | .procname = "warnings", |
330 | .data = &net_msg_warn, | |
331 | .maxlen = sizeof(int), | |
332 | .mode = 0644, | |
6d9f239a | 333 | .proc_handler = proc_dointvec |
a2a316fd | 334 | }, |
f8572d8f | 335 | { } |
1da177e4 | 336 | }; |
33eb9cfc | 337 | |
d5a4502e PE |
338 | static struct ctl_table netns_core_table[] = { |
339 | { | |
d5a4502e PE |
340 | .procname = "somaxconn", |
341 | .data = &init_net.core.sysctl_somaxconn, | |
342 | .maxlen = sizeof(int), | |
343 | .mode = 0644, | |
5f671d6b RG |
344 | .extra1 = &zero, |
345 | .extra2 = &ushort_max, | |
346 | .proc_handler = proc_dointvec_minmax | |
d5a4502e | 347 | }, |
f8572d8f | 348 | { } |
d5a4502e PE |
349 | }; |
350 | ||
024626e3 | 351 | static __net_init int sysctl_core_net_init(struct net *net) |
33eb9cfc | 352 | { |
d5a4502e | 353 | struct ctl_table *tbl; |
024626e3 | 354 | |
8efa6e93 | 355 | net->core.sysctl_somaxconn = SOMAXCONN; |
b8e1f9b5 | 356 | |
d5a4502e | 357 | tbl = netns_core_table; |
09ad9bc7 | 358 | if (!net_eq(net, &init_net)) { |
d5a4502e | 359 | tbl = kmemdup(tbl, sizeof(netns_core_table), GFP_KERNEL); |
024626e3 PE |
360 | if (tbl == NULL) |
361 | goto err_dup; | |
362 | ||
d5a4502e | 363 | tbl[0].data = &net->core.sysctl_somaxconn; |
464dc801 EB |
364 | |
365 | /* Don't export any sysctls to unprivileged users */ | |
366 | if (net->user_ns != &init_user_ns) { | |
367 | tbl[0].procname = NULL; | |
368 | } | |
024626e3 PE |
369 | } |
370 | ||
ec8f23ce | 371 | net->core.sysctl_hdr = register_net_sysctl(net, "net/core", tbl); |
8efa6e93 | 372 | if (net->core.sysctl_hdr == NULL) |
024626e3 | 373 | goto err_reg; |
33eb9cfc | 374 | |
024626e3 PE |
375 | return 0; |
376 | ||
377 | err_reg: | |
d5a4502e | 378 | if (tbl != netns_core_table) |
024626e3 PE |
379 | kfree(tbl); |
380 | err_dup: | |
381 | return -ENOMEM; | |
382 | } | |
383 | ||
384 | static __net_exit void sysctl_core_net_exit(struct net *net) | |
385 | { | |
386 | struct ctl_table *tbl; | |
387 | ||
8efa6e93 PE |
388 | tbl = net->core.sysctl_hdr->ctl_table_arg; |
389 | unregister_net_sysctl_table(net->core.sysctl_hdr); | |
d5a4502e | 390 | BUG_ON(tbl == netns_core_table); |
024626e3 PE |
391 | kfree(tbl); |
392 | } | |
393 | ||
394 | static __net_initdata struct pernet_operations sysctl_core_ops = { | |
395 | .init = sysctl_core_net_init, | |
396 | .exit = sysctl_core_net_exit, | |
397 | }; | |
398 | ||
399 | static __init int sysctl_core_init(void) | |
400 | { | |
43444757 | 401 | register_net_sysctl(&init_net, "net/core", net_core_table); |
024626e3 | 402 | return register_pernet_subsys(&sysctl_core_ops); |
33eb9cfc PE |
403 | } |
404 | ||
b27aeadb | 405 | fs_initcall(sysctl_core_init); |