]>
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> |
6da7c8fc | 23 | #include <net/pkt_sched.h> |
1da177e4 | 24 | |
5f671d6b | 25 | static int zero = 0; |
cdda8891 | 26 | static int one = 1; |
b1cb59cf AK |
27 | static int min_sndbuf = SOCK_MIN_SNDBUF; |
28 | static int min_rcvbuf = SOCK_MIN_RCVBUF; | |
5f74f82e | 29 | static int max_skb_frags = MAX_SKB_FRAGS; |
cdda8891 | 30 | |
ba7a46f1 JP |
31 | static int net_msg_warn; /* Unused, but still a sysctl */ |
32 | ||
fec5e652 | 33 | #ifdef CONFIG_RPS |
fe2c6338 | 34 | static int rps_sock_flow_sysctl(struct ctl_table *table, int write, |
fec5e652 TH |
35 | void __user *buffer, size_t *lenp, loff_t *ppos) |
36 | { | |
37 | unsigned int orig_size, size; | |
38 | int ret, i; | |
fe2c6338 | 39 | struct ctl_table tmp = { |
fec5e652 TH |
40 | .data = &size, |
41 | .maxlen = sizeof(size), | |
42 | .mode = table->mode | |
43 | }; | |
44 | struct rps_sock_flow_table *orig_sock_table, *sock_table; | |
45 | static DEFINE_MUTEX(sock_flow_mutex); | |
46 | ||
47 | mutex_lock(&sock_flow_mutex); | |
48 | ||
6e3f7faf ED |
49 | orig_sock_table = rcu_dereference_protected(rps_sock_flow_table, |
50 | lockdep_is_held(&sock_flow_mutex)); | |
fec5e652 TH |
51 | size = orig_size = orig_sock_table ? orig_sock_table->mask + 1 : 0; |
52 | ||
53 | ret = proc_dointvec(&tmp, write, buffer, lenp, ppos); | |
54 | ||
55 | if (write) { | |
56 | if (size) { | |
93c1af6c | 57 | if (size > 1<<29) { |
fec5e652 TH |
58 | /* Enforce limit to prevent overflow */ |
59 | mutex_unlock(&sock_flow_mutex); | |
60 | return -EINVAL; | |
61 | } | |
62 | size = roundup_pow_of_two(size); | |
63 | if (size != orig_size) { | |
64 | sock_table = | |
65 | vmalloc(RPS_SOCK_FLOW_TABLE_SIZE(size)); | |
66 | if (!sock_table) { | |
67 | mutex_unlock(&sock_flow_mutex); | |
68 | return -ENOMEM; | |
69 | } | |
567e4b79 | 70 | rps_cpu_mask = roundup_pow_of_two(nr_cpu_ids) - 1; |
fec5e652 TH |
71 | sock_table->mask = size - 1; |
72 | } else | |
73 | sock_table = orig_sock_table; | |
74 | ||
75 | for (i = 0; i < size; i++) | |
76 | sock_table->ents[i] = RPS_NO_CPU; | |
77 | } else | |
78 | sock_table = NULL; | |
79 | ||
80 | if (sock_table != orig_sock_table) { | |
81 | rcu_assign_pointer(rps_sock_flow_table, sock_table); | |
adc9300e | 82 | if (sock_table) |
c5905afb | 83 | static_key_slow_inc(&rps_needed); |
adc9300e | 84 | if (orig_sock_table) { |
c5905afb | 85 | static_key_slow_dec(&rps_needed); |
adc9300e ED |
86 | synchronize_rcu(); |
87 | vfree(orig_sock_table); | |
88 | } | |
fec5e652 TH |
89 | } |
90 | } | |
91 | ||
92 | mutex_unlock(&sock_flow_mutex); | |
93 | ||
94 | return ret; | |
95 | } | |
96 | #endif /* CONFIG_RPS */ | |
97 | ||
99bbc707 WB |
98 | #ifdef CONFIG_NET_FLOW_LIMIT |
99 | static DEFINE_MUTEX(flow_limit_update_mutex); | |
100 | ||
fe2c6338 | 101 | static int flow_limit_cpu_sysctl(struct ctl_table *table, int write, |
99bbc707 WB |
102 | void __user *buffer, size_t *lenp, |
103 | loff_t *ppos) | |
104 | { | |
105 | struct sd_flow_limit *cur; | |
106 | struct softnet_data *sd; | |
107 | cpumask_var_t mask; | |
108 | int i, len, ret = 0; | |
109 | ||
110 | if (!alloc_cpumask_var(&mask, GFP_KERNEL)) | |
111 | return -ENOMEM; | |
112 | ||
113 | if (write) { | |
114 | ret = cpumask_parse_user(buffer, *lenp, mask); | |
115 | if (ret) | |
116 | goto done; | |
117 | ||
118 | mutex_lock(&flow_limit_update_mutex); | |
119 | len = sizeof(*cur) + netdev_flow_limit_table_len; | |
120 | for_each_possible_cpu(i) { | |
121 | sd = &per_cpu(softnet_data, i); | |
122 | cur = rcu_dereference_protected(sd->flow_limit, | |
123 | lockdep_is_held(&flow_limit_update_mutex)); | |
124 | if (cur && !cpumask_test_cpu(i, mask)) { | |
125 | RCU_INIT_POINTER(sd->flow_limit, NULL); | |
126 | synchronize_rcu(); | |
127 | kfree(cur); | |
128 | } else if (!cur && cpumask_test_cpu(i, mask)) { | |
5b59d467 ED |
129 | cur = kzalloc_node(len, GFP_KERNEL, |
130 | cpu_to_node(i)); | |
99bbc707 WB |
131 | if (!cur) { |
132 | /* not unwinding previous changes */ | |
133 | ret = -ENOMEM; | |
134 | goto write_unlock; | |
135 | } | |
136 | cur->num_buckets = netdev_flow_limit_table_len; | |
137 | rcu_assign_pointer(sd->flow_limit, cur); | |
138 | } | |
139 | } | |
140 | write_unlock: | |
141 | mutex_unlock(&flow_limit_update_mutex); | |
142 | } else { | |
5f121b9a WB |
143 | char kbuf[128]; |
144 | ||
99bbc707 WB |
145 | if (*ppos || !*lenp) { |
146 | *lenp = 0; | |
147 | goto done; | |
148 | } | |
149 | ||
150 | cpumask_clear(mask); | |
151 | rcu_read_lock(); | |
152 | for_each_possible_cpu(i) { | |
153 | sd = &per_cpu(softnet_data, i); | |
154 | if (rcu_dereference(sd->flow_limit)) | |
155 | cpumask_set_cpu(i, mask); | |
156 | } | |
157 | rcu_read_unlock(); | |
158 | ||
5f121b9a | 159 | len = min(sizeof(kbuf) - 1, *lenp); |
f0906827 | 160 | len = scnprintf(kbuf, len, "%*pb", cpumask_pr_args(mask)); |
5f121b9a WB |
161 | if (!len) { |
162 | *lenp = 0; | |
163 | goto done; | |
164 | } | |
165 | if (len < *lenp) | |
166 | kbuf[len++] = '\n'; | |
167 | if (copy_to_user(buffer, kbuf, len)) { | |
168 | ret = -EFAULT; | |
169 | goto done; | |
170 | } | |
171 | *lenp = len; | |
172 | *ppos += len; | |
99bbc707 WB |
173 | } |
174 | ||
175 | done: | |
176 | free_cpumask_var(mask); | |
177 | return ret; | |
178 | } | |
179 | ||
fe2c6338 | 180 | static int flow_limit_table_len_sysctl(struct ctl_table *table, int write, |
99bbc707 WB |
181 | void __user *buffer, size_t *lenp, |
182 | loff_t *ppos) | |
183 | { | |
184 | unsigned int old, *ptr; | |
185 | int ret; | |
186 | ||
187 | mutex_lock(&flow_limit_update_mutex); | |
188 | ||
189 | ptr = table->data; | |
190 | old = *ptr; | |
191 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
192 | if (!ret && write && !is_power_of_2(*ptr)) { | |
193 | *ptr = old; | |
194 | ret = -EINVAL; | |
195 | } | |
196 | ||
197 | mutex_unlock(&flow_limit_update_mutex); | |
198 | return ret; | |
199 | } | |
200 | #endif /* CONFIG_NET_FLOW_LIMIT */ | |
201 | ||
6da7c8fc | 202 | #ifdef CONFIG_NET_SCHED |
203 | static int set_default_qdisc(struct ctl_table *table, int write, | |
204 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
205 | { | |
206 | char id[IFNAMSIZ]; | |
207 | struct ctl_table tbl = { | |
208 | .data = id, | |
209 | .maxlen = IFNAMSIZ, | |
210 | }; | |
211 | int ret; | |
212 | ||
213 | qdisc_get_default(id, IFNAMSIZ); | |
214 | ||
215 | ret = proc_dostring(&tbl, write, buffer, lenp, ppos); | |
216 | if (write && ret == 0) | |
217 | ret = qdisc_set_default(id); | |
218 | return ret; | |
219 | } | |
220 | #endif | |
221 | ||
960fb622 ED |
222 | static int proc_do_rss_key(struct ctl_table *table, int write, |
223 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
224 | { | |
225 | struct ctl_table fake_table; | |
226 | char buf[NETDEV_RSS_KEY_LEN * 3]; | |
227 | ||
228 | snprintf(buf, sizeof(buf), "%*phC", NETDEV_RSS_KEY_LEN, netdev_rss_key); | |
229 | fake_table.data = buf; | |
230 | fake_table.maxlen = sizeof(buf); | |
231 | return proc_dostring(&fake_table, write, buffer, lenp, ppos); | |
232 | } | |
233 | ||
33eb9cfc | 234 | static struct ctl_table net_core_table[] = { |
1da177e4 LT |
235 | #ifdef CONFIG_NET |
236 | { | |
1da177e4 LT |
237 | .procname = "wmem_max", |
238 | .data = &sysctl_wmem_max, | |
239 | .maxlen = sizeof(int), | |
240 | .mode = 0644, | |
cdda8891 | 241 | .proc_handler = proc_dointvec_minmax, |
b1cb59cf | 242 | .extra1 = &min_sndbuf, |
1da177e4 LT |
243 | }, |
244 | { | |
1da177e4 LT |
245 | .procname = "rmem_max", |
246 | .data = &sysctl_rmem_max, | |
247 | .maxlen = sizeof(int), | |
248 | .mode = 0644, | |
cdda8891 | 249 | .proc_handler = proc_dointvec_minmax, |
b1cb59cf | 250 | .extra1 = &min_rcvbuf, |
1da177e4 LT |
251 | }, |
252 | { | |
1da177e4 LT |
253 | .procname = "wmem_default", |
254 | .data = &sysctl_wmem_default, | |
255 | .maxlen = sizeof(int), | |
256 | .mode = 0644, | |
cdda8891 | 257 | .proc_handler = proc_dointvec_minmax, |
b1cb59cf | 258 | .extra1 = &min_sndbuf, |
1da177e4 LT |
259 | }, |
260 | { | |
1da177e4 LT |
261 | .procname = "rmem_default", |
262 | .data = &sysctl_rmem_default, | |
263 | .maxlen = sizeof(int), | |
264 | .mode = 0644, | |
cdda8891 | 265 | .proc_handler = proc_dointvec_minmax, |
b1cb59cf | 266 | .extra1 = &min_rcvbuf, |
1da177e4 LT |
267 | }, |
268 | { | |
1da177e4 LT |
269 | .procname = "dev_weight", |
270 | .data = &weight_p, | |
271 | .maxlen = sizeof(int), | |
272 | .mode = 0644, | |
6d9f239a | 273 | .proc_handler = proc_dointvec |
1da177e4 LT |
274 | }, |
275 | { | |
1da177e4 LT |
276 | .procname = "netdev_max_backlog", |
277 | .data = &netdev_max_backlog, | |
278 | .maxlen = sizeof(int), | |
279 | .mode = 0644, | |
6d9f239a | 280 | .proc_handler = proc_dointvec |
1da177e4 | 281 | }, |
960fb622 ED |
282 | { |
283 | .procname = "netdev_rss_key", | |
284 | .data = &netdev_rss_key, | |
285 | .maxlen = sizeof(int), | |
286 | .mode = 0444, | |
287 | .proc_handler = proc_do_rss_key, | |
288 | }, | |
0a14842f ED |
289 | #ifdef CONFIG_BPF_JIT |
290 | { | |
291 | .procname = "bpf_jit_enable", | |
292 | .data = &bpf_jit_enable, | |
293 | .maxlen = sizeof(int), | |
294 | .mode = 0644, | |
295 | .proc_handler = proc_dointvec | |
296 | }, | |
297 | #endif | |
3b098e2d ED |
298 | { |
299 | .procname = "netdev_tstamp_prequeue", | |
300 | .data = &netdev_tstamp_prequeue, | |
301 | .maxlen = sizeof(int), | |
302 | .mode = 0644, | |
303 | .proc_handler = proc_dointvec | |
304 | }, | |
1da177e4 | 305 | { |
1da177e4 | 306 | .procname = "message_cost", |
717115e1 | 307 | .data = &net_ratelimit_state.interval, |
1da177e4 LT |
308 | .maxlen = sizeof(int), |
309 | .mode = 0644, | |
6d9f239a | 310 | .proc_handler = proc_dointvec_jiffies, |
1da177e4 LT |
311 | }, |
312 | { | |
1da177e4 | 313 | .procname = "message_burst", |
717115e1 | 314 | .data = &net_ratelimit_state.burst, |
1da177e4 LT |
315 | .maxlen = sizeof(int), |
316 | .mode = 0644, | |
6d9f239a | 317 | .proc_handler = proc_dointvec, |
1da177e4 LT |
318 | }, |
319 | { | |
1da177e4 LT |
320 | .procname = "optmem_max", |
321 | .data = &sysctl_optmem_max, | |
322 | .maxlen = sizeof(int), | |
323 | .mode = 0644, | |
6d9f239a | 324 | .proc_handler = proc_dointvec |
1da177e4 | 325 | }, |
b245be1f WB |
326 | { |
327 | .procname = "tstamp_allow_data", | |
328 | .data = &sysctl_tstamp_allow_data, | |
329 | .maxlen = sizeof(int), | |
330 | .mode = 0644, | |
331 | .proc_handler = proc_dointvec_minmax, | |
332 | .extra1 = &zero, | |
333 | .extra2 = &one | |
334 | }, | |
fec5e652 TH |
335 | #ifdef CONFIG_RPS |
336 | { | |
337 | .procname = "rps_sock_flow_entries", | |
338 | .maxlen = sizeof(int), | |
339 | .mode = 0644, | |
340 | .proc_handler = rps_sock_flow_sysctl | |
341 | }, | |
342 | #endif | |
99bbc707 WB |
343 | #ifdef CONFIG_NET_FLOW_LIMIT |
344 | { | |
345 | .procname = "flow_limit_cpu_bitmap", | |
346 | .mode = 0644, | |
347 | .proc_handler = flow_limit_cpu_sysctl | |
348 | }, | |
349 | { | |
350 | .procname = "flow_limit_table_len", | |
351 | .data = &netdev_flow_limit_table_len, | |
352 | .maxlen = sizeof(int), | |
353 | .mode = 0644, | |
354 | .proc_handler = flow_limit_table_len_sysctl | |
355 | }, | |
356 | #endif /* CONFIG_NET_FLOW_LIMIT */ | |
e0d1095a | 357 | #ifdef CONFIG_NET_RX_BUSY_POLL |
06021292 | 358 | { |
64b0dc51 ET |
359 | .procname = "busy_poll", |
360 | .data = &sysctl_net_busy_poll, | |
eb6db622 | 361 | .maxlen = sizeof(unsigned int), |
06021292 | 362 | .mode = 0644, |
eb6db622 | 363 | .proc_handler = proc_dointvec |
06021292 | 364 | }, |
2d48d67f | 365 | { |
64b0dc51 ET |
366 | .procname = "busy_read", |
367 | .data = &sysctl_net_busy_read, | |
2d48d67f ET |
368 | .maxlen = sizeof(unsigned int), |
369 | .mode = 0644, | |
370 | .proc_handler = proc_dointvec | |
371 | }, | |
6da7c8fc | 372 | #endif |
373 | #ifdef CONFIG_NET_SCHED | |
374 | { | |
375 | .procname = "default_qdisc", | |
376 | .mode = 0644, | |
377 | .maxlen = IFNAMSIZ, | |
378 | .proc_handler = set_default_qdisc | |
379 | }, | |
06021292 | 380 | #endif |
1da177e4 | 381 | #endif /* CONFIG_NET */ |
51b0bded | 382 | { |
51b0bded SH |
383 | .procname = "netdev_budget", |
384 | .data = &netdev_budget, | |
385 | .maxlen = sizeof(int), | |
386 | .mode = 0644, | |
6d9f239a | 387 | .proc_handler = proc_dointvec |
51b0bded | 388 | }, |
a2a316fd | 389 | { |
a2a316fd SH |
390 | .procname = "warnings", |
391 | .data = &net_msg_warn, | |
392 | .maxlen = sizeof(int), | |
393 | .mode = 0644, | |
6d9f239a | 394 | .proc_handler = proc_dointvec |
a2a316fd | 395 | }, |
5f74f82e HWR |
396 | { |
397 | .procname = "max_skb_frags", | |
398 | .data = &sysctl_max_skb_frags, | |
399 | .maxlen = sizeof(int), | |
400 | .mode = 0644, | |
401 | .proc_handler = proc_dointvec_minmax, | |
402 | .extra1 = &one, | |
403 | .extra2 = &max_skb_frags, | |
404 | }, | |
f8572d8f | 405 | { } |
1da177e4 | 406 | }; |
33eb9cfc | 407 | |
d5a4502e PE |
408 | static struct ctl_table netns_core_table[] = { |
409 | { | |
d5a4502e PE |
410 | .procname = "somaxconn", |
411 | .data = &init_net.core.sysctl_somaxconn, | |
412 | .maxlen = sizeof(int), | |
413 | .mode = 0644, | |
5f671d6b | 414 | .extra1 = &zero, |
5f671d6b | 415 | .proc_handler = proc_dointvec_minmax |
d5a4502e | 416 | }, |
f8572d8f | 417 | { } |
d5a4502e PE |
418 | }; |
419 | ||
024626e3 | 420 | static __net_init int sysctl_core_net_init(struct net *net) |
33eb9cfc | 421 | { |
d5a4502e | 422 | struct ctl_table *tbl; |
024626e3 | 423 | |
8efa6e93 | 424 | net->core.sysctl_somaxconn = SOMAXCONN; |
b8e1f9b5 | 425 | |
d5a4502e | 426 | tbl = netns_core_table; |
09ad9bc7 | 427 | if (!net_eq(net, &init_net)) { |
d5a4502e | 428 | tbl = kmemdup(tbl, sizeof(netns_core_table), GFP_KERNEL); |
024626e3 PE |
429 | if (tbl == NULL) |
430 | goto err_dup; | |
431 | ||
d5a4502e | 432 | tbl[0].data = &net->core.sysctl_somaxconn; |
464dc801 EB |
433 | |
434 | /* Don't export any sysctls to unprivileged users */ | |
435 | if (net->user_ns != &init_user_ns) { | |
436 | tbl[0].procname = NULL; | |
437 | } | |
024626e3 PE |
438 | } |
439 | ||
ec8f23ce | 440 | net->core.sysctl_hdr = register_net_sysctl(net, "net/core", tbl); |
8efa6e93 | 441 | if (net->core.sysctl_hdr == NULL) |
024626e3 | 442 | goto err_reg; |
33eb9cfc | 443 | |
024626e3 PE |
444 | return 0; |
445 | ||
446 | err_reg: | |
d5a4502e | 447 | if (tbl != netns_core_table) |
024626e3 PE |
448 | kfree(tbl); |
449 | err_dup: | |
450 | return -ENOMEM; | |
451 | } | |
452 | ||
453 | static __net_exit void sysctl_core_net_exit(struct net *net) | |
454 | { | |
455 | struct ctl_table *tbl; | |
456 | ||
8efa6e93 PE |
457 | tbl = net->core.sysctl_hdr->ctl_table_arg; |
458 | unregister_net_sysctl_table(net->core.sysctl_hdr); | |
d5a4502e | 459 | BUG_ON(tbl == netns_core_table); |
024626e3 PE |
460 | kfree(tbl); |
461 | } | |
462 | ||
463 | static __net_initdata struct pernet_operations sysctl_core_ops = { | |
464 | .init = sysctl_core_net_init, | |
465 | .exit = sysctl_core_net_exit, | |
466 | }; | |
467 | ||
468 | static __init int sysctl_core_init(void) | |
469 | { | |
43444757 | 470 | register_net_sysctl(&init_net, "net/core", net_core_table); |
024626e3 | 471 | return register_pernet_subsys(&sysctl_core_ops); |
33eb9cfc PE |
472 | } |
473 | ||
b27aeadb | 474 | fs_initcall(sysctl_core_init); |