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