Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
1da177e4 LT |
2 | /* -*- linux-c -*- |
3 | * sysctl_net_core.c: sysctl interface to net core subsystem. | |
4 | * | |
5 | * Begun April 1, 1996, Mike Shaver. | |
6 | * Added /proc/sys/net/core directory entry (empty =) ). [MS] | |
7 | */ | |
8 | ||
9 | #include <linux/mm.h> | |
10 | #include <linux/sysctl.h> | |
1da177e4 | 11 | #include <linux/module.h> |
20380731 | 12 | #include <linux/socket.h> |
a37ae408 | 13 | #include <linux/netdevice.h> |
3fff4c42 | 14 | #include <linux/ratelimit.h> |
fec5e652 | 15 | #include <linux/vmalloc.h> |
33eb9cfc | 16 | #include <linux/init.h> |
5a0e3ad6 | 17 | #include <linux/slab.h> |
7426a564 | 18 | #include <linux/kmemleak.h> |
3fff4c42 | 19 | |
63d819ca | 20 | #include <net/ip.h> |
20380731 | 21 | #include <net/sock.h> |
c5c177b4 | 22 | #include <net/net_ratelimit.h> |
076bb0c8 | 23 | #include <net/busy_poll.h> |
6da7c8fc | 24 | #include <net/pkt_sched.h> |
1da177e4 | 25 | |
5f671d6b | 26 | static int zero = 0; |
cdda8891 | 27 | static int one = 1; |
b1cb59cf AK |
28 | static int min_sndbuf = SOCK_MIN_SNDBUF; |
29 | static int min_rcvbuf = SOCK_MIN_RCVBUF; | |
5f74f82e | 30 | static int max_skb_frags = MAX_SKB_FRAGS; |
cdda8891 | 31 | |
ba7a46f1 JP |
32 | static int net_msg_warn; /* Unused, but still a sysctl */ |
33 | ||
fec5e652 | 34 | #ifdef CONFIG_RPS |
fe2c6338 | 35 | static int rps_sock_flow_sysctl(struct ctl_table *table, int write, |
fec5e652 TH |
36 | void __user *buffer, size_t *lenp, loff_t *ppos) |
37 | { | |
38 | unsigned int orig_size, size; | |
39 | int ret, i; | |
fe2c6338 | 40 | struct ctl_table tmp = { |
fec5e652 TH |
41 | .data = &size, |
42 | .maxlen = sizeof(size), | |
43 | .mode = table->mode | |
44 | }; | |
45 | struct rps_sock_flow_table *orig_sock_table, *sock_table; | |
46 | static DEFINE_MUTEX(sock_flow_mutex); | |
47 | ||
48 | mutex_lock(&sock_flow_mutex); | |
49 | ||
6e3f7faf ED |
50 | orig_sock_table = rcu_dereference_protected(rps_sock_flow_table, |
51 | lockdep_is_held(&sock_flow_mutex)); | |
fec5e652 TH |
52 | size = orig_size = orig_sock_table ? orig_sock_table->mask + 1 : 0; |
53 | ||
54 | ret = proc_dointvec(&tmp, write, buffer, lenp, ppos); | |
55 | ||
56 | if (write) { | |
57 | if (size) { | |
93c1af6c | 58 | if (size > 1<<29) { |
fec5e652 TH |
59 | /* Enforce limit to prevent overflow */ |
60 | mutex_unlock(&sock_flow_mutex); | |
61 | return -EINVAL; | |
62 | } | |
63 | size = roundup_pow_of_two(size); | |
64 | if (size != orig_size) { | |
65 | sock_table = | |
66 | vmalloc(RPS_SOCK_FLOW_TABLE_SIZE(size)); | |
67 | if (!sock_table) { | |
68 | mutex_unlock(&sock_flow_mutex); | |
69 | return -ENOMEM; | |
70 | } | |
567e4b79 | 71 | rps_cpu_mask = roundup_pow_of_two(nr_cpu_ids) - 1; |
fec5e652 TH |
72 | sock_table->mask = size - 1; |
73 | } else | |
74 | sock_table = orig_sock_table; | |
75 | ||
76 | for (i = 0; i < size; i++) | |
77 | sock_table->ents[i] = RPS_NO_CPU; | |
78 | } else | |
79 | sock_table = NULL; | |
80 | ||
81 | if (sock_table != orig_sock_table) { | |
82 | rcu_assign_pointer(rps_sock_flow_table, sock_table); | |
13bfff25 | 83 | if (sock_table) { |
c5905afb | 84 | static_key_slow_inc(&rps_needed); |
13bfff25 ED |
85 | static_key_slow_inc(&rfs_needed); |
86 | } | |
adc9300e | 87 | if (orig_sock_table) { |
c5905afb | 88 | static_key_slow_dec(&rps_needed); |
13bfff25 | 89 | static_key_slow_dec(&rfs_needed); |
adc9300e ED |
90 | synchronize_rcu(); |
91 | vfree(orig_sock_table); | |
92 | } | |
fec5e652 TH |
93 | } |
94 | } | |
95 | ||
96 | mutex_unlock(&sock_flow_mutex); | |
97 | ||
98 | return ret; | |
99 | } | |
100 | #endif /* CONFIG_RPS */ | |
101 | ||
99bbc707 WB |
102 | #ifdef CONFIG_NET_FLOW_LIMIT |
103 | static DEFINE_MUTEX(flow_limit_update_mutex); | |
104 | ||
fe2c6338 | 105 | static int flow_limit_cpu_sysctl(struct ctl_table *table, int write, |
99bbc707 WB |
106 | void __user *buffer, size_t *lenp, |
107 | loff_t *ppos) | |
108 | { | |
109 | struct sd_flow_limit *cur; | |
110 | struct softnet_data *sd; | |
111 | cpumask_var_t mask; | |
112 | int i, len, ret = 0; | |
113 | ||
114 | if (!alloc_cpumask_var(&mask, GFP_KERNEL)) | |
115 | return -ENOMEM; | |
116 | ||
117 | if (write) { | |
118 | ret = cpumask_parse_user(buffer, *lenp, mask); | |
119 | if (ret) | |
120 | goto done; | |
121 | ||
122 | mutex_lock(&flow_limit_update_mutex); | |
123 | len = sizeof(*cur) + netdev_flow_limit_table_len; | |
124 | for_each_possible_cpu(i) { | |
125 | sd = &per_cpu(softnet_data, i); | |
126 | cur = rcu_dereference_protected(sd->flow_limit, | |
127 | lockdep_is_held(&flow_limit_update_mutex)); | |
128 | if (cur && !cpumask_test_cpu(i, mask)) { | |
129 | RCU_INIT_POINTER(sd->flow_limit, NULL); | |
130 | synchronize_rcu(); | |
131 | kfree(cur); | |
132 | } else if (!cur && cpumask_test_cpu(i, mask)) { | |
5b59d467 ED |
133 | cur = kzalloc_node(len, GFP_KERNEL, |
134 | cpu_to_node(i)); | |
99bbc707 WB |
135 | if (!cur) { |
136 | /* not unwinding previous changes */ | |
137 | ret = -ENOMEM; | |
138 | goto write_unlock; | |
139 | } | |
140 | cur->num_buckets = netdev_flow_limit_table_len; | |
141 | rcu_assign_pointer(sd->flow_limit, cur); | |
142 | } | |
143 | } | |
144 | write_unlock: | |
145 | mutex_unlock(&flow_limit_update_mutex); | |
146 | } else { | |
5f121b9a WB |
147 | char kbuf[128]; |
148 | ||
99bbc707 WB |
149 | if (*ppos || !*lenp) { |
150 | *lenp = 0; | |
151 | goto done; | |
152 | } | |
153 | ||
154 | cpumask_clear(mask); | |
155 | rcu_read_lock(); | |
156 | for_each_possible_cpu(i) { | |
157 | sd = &per_cpu(softnet_data, i); | |
158 | if (rcu_dereference(sd->flow_limit)) | |
159 | cpumask_set_cpu(i, mask); | |
160 | } | |
161 | rcu_read_unlock(); | |
162 | ||
5f121b9a | 163 | len = min(sizeof(kbuf) - 1, *lenp); |
f0906827 | 164 | len = scnprintf(kbuf, len, "%*pb", cpumask_pr_args(mask)); |
5f121b9a WB |
165 | if (!len) { |
166 | *lenp = 0; | |
167 | goto done; | |
168 | } | |
169 | if (len < *lenp) | |
170 | kbuf[len++] = '\n'; | |
171 | if (copy_to_user(buffer, kbuf, len)) { | |
172 | ret = -EFAULT; | |
173 | goto done; | |
174 | } | |
175 | *lenp = len; | |
176 | *ppos += len; | |
99bbc707 WB |
177 | } |
178 | ||
179 | done: | |
180 | free_cpumask_var(mask); | |
181 | return ret; | |
182 | } | |
183 | ||
fe2c6338 | 184 | static int flow_limit_table_len_sysctl(struct ctl_table *table, int write, |
99bbc707 WB |
185 | void __user *buffer, size_t *lenp, |
186 | loff_t *ppos) | |
187 | { | |
188 | unsigned int old, *ptr; | |
189 | int ret; | |
190 | ||
191 | mutex_lock(&flow_limit_update_mutex); | |
192 | ||
193 | ptr = table->data; | |
194 | old = *ptr; | |
195 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
196 | if (!ret && write && !is_power_of_2(*ptr)) { | |
197 | *ptr = old; | |
198 | ret = -EINVAL; | |
199 | } | |
200 | ||
201 | mutex_unlock(&flow_limit_update_mutex); | |
202 | return ret; | |
203 | } | |
204 | #endif /* CONFIG_NET_FLOW_LIMIT */ | |
205 | ||
6da7c8fc | 206 | #ifdef CONFIG_NET_SCHED |
207 | static int set_default_qdisc(struct ctl_table *table, int write, | |
208 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
209 | { | |
210 | char id[IFNAMSIZ]; | |
211 | struct ctl_table tbl = { | |
212 | .data = id, | |
213 | .maxlen = IFNAMSIZ, | |
214 | }; | |
215 | int ret; | |
216 | ||
217 | qdisc_get_default(id, IFNAMSIZ); | |
218 | ||
219 | ret = proc_dostring(&tbl, write, buffer, lenp, ppos); | |
220 | if (write && ret == 0) | |
221 | ret = qdisc_set_default(id); | |
222 | return ret; | |
223 | } | |
224 | #endif | |
225 | ||
3d48b53f MT |
226 | static int proc_do_dev_weight(struct ctl_table *table, int write, |
227 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
228 | { | |
229 | int ret; | |
230 | ||
231 | ret = proc_dointvec(table, write, buffer, lenp, ppos); | |
232 | if (ret != 0) | |
233 | return ret; | |
234 | ||
235 | dev_rx_weight = weight_p * dev_weight_rx_bias; | |
236 | dev_tx_weight = weight_p * dev_weight_tx_bias; | |
237 | ||
238 | return ret; | |
239 | } | |
240 | ||
960fb622 ED |
241 | static int proc_do_rss_key(struct ctl_table *table, int write, |
242 | void __user *buffer, size_t *lenp, loff_t *ppos) | |
243 | { | |
244 | struct ctl_table fake_table; | |
245 | char buf[NETDEV_RSS_KEY_LEN * 3]; | |
246 | ||
247 | snprintf(buf, sizeof(buf), "%*phC", NETDEV_RSS_KEY_LEN, netdev_rss_key); | |
248 | fake_table.data = buf; | |
249 | fake_table.maxlen = sizeof(buf); | |
250 | return proc_dostring(&fake_table, write, buffer, lenp, ppos); | |
251 | } | |
252 | ||
33eb9cfc | 253 | static struct ctl_table net_core_table[] = { |
1da177e4 LT |
254 | #ifdef CONFIG_NET |
255 | { | |
1da177e4 LT |
256 | .procname = "wmem_max", |
257 | .data = &sysctl_wmem_max, | |
258 | .maxlen = sizeof(int), | |
259 | .mode = 0644, | |
cdda8891 | 260 | .proc_handler = proc_dointvec_minmax, |
b1cb59cf | 261 | .extra1 = &min_sndbuf, |
1da177e4 LT |
262 | }, |
263 | { | |
1da177e4 LT |
264 | .procname = "rmem_max", |
265 | .data = &sysctl_rmem_max, | |
266 | .maxlen = sizeof(int), | |
267 | .mode = 0644, | |
cdda8891 | 268 | .proc_handler = proc_dointvec_minmax, |
b1cb59cf | 269 | .extra1 = &min_rcvbuf, |
1da177e4 LT |
270 | }, |
271 | { | |
1da177e4 LT |
272 | .procname = "wmem_default", |
273 | .data = &sysctl_wmem_default, | |
274 | .maxlen = sizeof(int), | |
275 | .mode = 0644, | |
cdda8891 | 276 | .proc_handler = proc_dointvec_minmax, |
b1cb59cf | 277 | .extra1 = &min_sndbuf, |
1da177e4 LT |
278 | }, |
279 | { | |
1da177e4 LT |
280 | .procname = "rmem_default", |
281 | .data = &sysctl_rmem_default, | |
282 | .maxlen = sizeof(int), | |
283 | .mode = 0644, | |
cdda8891 | 284 | .proc_handler = proc_dointvec_minmax, |
b1cb59cf | 285 | .extra1 = &min_rcvbuf, |
1da177e4 LT |
286 | }, |
287 | { | |
1da177e4 LT |
288 | .procname = "dev_weight", |
289 | .data = &weight_p, | |
290 | .maxlen = sizeof(int), | |
291 | .mode = 0644, | |
3d48b53f MT |
292 | .proc_handler = proc_do_dev_weight, |
293 | }, | |
294 | { | |
295 | .procname = "dev_weight_rx_bias", | |
296 | .data = &dev_weight_rx_bias, | |
297 | .maxlen = sizeof(int), | |
298 | .mode = 0644, | |
299 | .proc_handler = proc_do_dev_weight, | |
300 | }, | |
301 | { | |
302 | .procname = "dev_weight_tx_bias", | |
303 | .data = &dev_weight_tx_bias, | |
304 | .maxlen = sizeof(int), | |
305 | .mode = 0644, | |
306 | .proc_handler = proc_do_dev_weight, | |
1da177e4 LT |
307 | }, |
308 | { | |
1da177e4 LT |
309 | .procname = "netdev_max_backlog", |
310 | .data = &netdev_max_backlog, | |
311 | .maxlen = sizeof(int), | |
312 | .mode = 0644, | |
6d9f239a | 313 | .proc_handler = proc_dointvec |
1da177e4 | 314 | }, |
960fb622 ED |
315 | { |
316 | .procname = "netdev_rss_key", | |
317 | .data = &netdev_rss_key, | |
318 | .maxlen = sizeof(int), | |
319 | .mode = 0444, | |
320 | .proc_handler = proc_do_rss_key, | |
321 | }, | |
0a14842f ED |
322 | #ifdef CONFIG_BPF_JIT |
323 | { | |
324 | .procname = "bpf_jit_enable", | |
325 | .data = &bpf_jit_enable, | |
326 | .maxlen = sizeof(int), | |
327 | .mode = 0644, | |
328 | .proc_handler = proc_dointvec | |
329 | }, | |
4f3446bb DB |
330 | # ifdef CONFIG_HAVE_EBPF_JIT |
331 | { | |
332 | .procname = "bpf_jit_harden", | |
333 | .data = &bpf_jit_harden, | |
334 | .maxlen = sizeof(int), | |
335 | .mode = 0600, | |
336 | .proc_handler = proc_dointvec, | |
337 | }, | |
74451e66 DB |
338 | { |
339 | .procname = "bpf_jit_kallsyms", | |
340 | .data = &bpf_jit_kallsyms, | |
341 | .maxlen = sizeof(int), | |
342 | .mode = 0600, | |
343 | .proc_handler = proc_dointvec, | |
344 | }, | |
4f3446bb | 345 | # endif |
0a14842f | 346 | #endif |
3b098e2d ED |
347 | { |
348 | .procname = "netdev_tstamp_prequeue", | |
349 | .data = &netdev_tstamp_prequeue, | |
350 | .maxlen = sizeof(int), | |
351 | .mode = 0644, | |
352 | .proc_handler = proc_dointvec | |
353 | }, | |
1da177e4 | 354 | { |
1da177e4 | 355 | .procname = "message_cost", |
717115e1 | 356 | .data = &net_ratelimit_state.interval, |
1da177e4 LT |
357 | .maxlen = sizeof(int), |
358 | .mode = 0644, | |
6d9f239a | 359 | .proc_handler = proc_dointvec_jiffies, |
1da177e4 LT |
360 | }, |
361 | { | |
1da177e4 | 362 | .procname = "message_burst", |
717115e1 | 363 | .data = &net_ratelimit_state.burst, |
1da177e4 LT |
364 | .maxlen = sizeof(int), |
365 | .mode = 0644, | |
6d9f239a | 366 | .proc_handler = proc_dointvec, |
1da177e4 LT |
367 | }, |
368 | { | |
1da177e4 LT |
369 | .procname = "optmem_max", |
370 | .data = &sysctl_optmem_max, | |
371 | .maxlen = sizeof(int), | |
372 | .mode = 0644, | |
6d9f239a | 373 | .proc_handler = proc_dointvec |
1da177e4 | 374 | }, |
b245be1f WB |
375 | { |
376 | .procname = "tstamp_allow_data", | |
377 | .data = &sysctl_tstamp_allow_data, | |
378 | .maxlen = sizeof(int), | |
379 | .mode = 0644, | |
380 | .proc_handler = proc_dointvec_minmax, | |
381 | .extra1 = &zero, | |
382 | .extra2 = &one | |
383 | }, | |
fec5e652 TH |
384 | #ifdef CONFIG_RPS |
385 | { | |
386 | .procname = "rps_sock_flow_entries", | |
387 | .maxlen = sizeof(int), | |
388 | .mode = 0644, | |
389 | .proc_handler = rps_sock_flow_sysctl | |
390 | }, | |
391 | #endif | |
99bbc707 WB |
392 | #ifdef CONFIG_NET_FLOW_LIMIT |
393 | { | |
394 | .procname = "flow_limit_cpu_bitmap", | |
395 | .mode = 0644, | |
396 | .proc_handler = flow_limit_cpu_sysctl | |
397 | }, | |
398 | { | |
399 | .procname = "flow_limit_table_len", | |
400 | .data = &netdev_flow_limit_table_len, | |
401 | .maxlen = sizeof(int), | |
402 | .mode = 0644, | |
403 | .proc_handler = flow_limit_table_len_sysctl | |
404 | }, | |
405 | #endif /* CONFIG_NET_FLOW_LIMIT */ | |
e0d1095a | 406 | #ifdef CONFIG_NET_RX_BUSY_POLL |
06021292 | 407 | { |
64b0dc51 ET |
408 | .procname = "busy_poll", |
409 | .data = &sysctl_net_busy_poll, | |
eb6db622 | 410 | .maxlen = sizeof(unsigned int), |
06021292 | 411 | .mode = 0644, |
95f25521 AD |
412 | .proc_handler = proc_dointvec_minmax, |
413 | .extra1 = &zero, | |
06021292 | 414 | }, |
2d48d67f | 415 | { |
64b0dc51 ET |
416 | .procname = "busy_read", |
417 | .data = &sysctl_net_busy_read, | |
2d48d67f ET |
418 | .maxlen = sizeof(unsigned int), |
419 | .mode = 0644, | |
95f25521 AD |
420 | .proc_handler = proc_dointvec_minmax, |
421 | .extra1 = &zero, | |
2d48d67f | 422 | }, |
6da7c8fc | 423 | #endif |
424 | #ifdef CONFIG_NET_SCHED | |
425 | { | |
426 | .procname = "default_qdisc", | |
427 | .mode = 0644, | |
428 | .maxlen = IFNAMSIZ, | |
429 | .proc_handler = set_default_qdisc | |
430 | }, | |
06021292 | 431 | #endif |
1da177e4 | 432 | #endif /* CONFIG_NET */ |
51b0bded | 433 | { |
51b0bded SH |
434 | .procname = "netdev_budget", |
435 | .data = &netdev_budget, | |
436 | .maxlen = sizeof(int), | |
437 | .mode = 0644, | |
6d9f239a | 438 | .proc_handler = proc_dointvec |
51b0bded | 439 | }, |
a2a316fd | 440 | { |
a2a316fd SH |
441 | .procname = "warnings", |
442 | .data = &net_msg_warn, | |
443 | .maxlen = sizeof(int), | |
444 | .mode = 0644, | |
6d9f239a | 445 | .proc_handler = proc_dointvec |
a2a316fd | 446 | }, |
5f74f82e HWR |
447 | { |
448 | .procname = "max_skb_frags", | |
449 | .data = &sysctl_max_skb_frags, | |
450 | .maxlen = sizeof(int), | |
451 | .mode = 0644, | |
452 | .proc_handler = proc_dointvec_minmax, | |
453 | .extra1 = &one, | |
454 | .extra2 = &max_skb_frags, | |
455 | }, | |
7acf8a1e MW |
456 | { |
457 | .procname = "netdev_budget_usecs", | |
458 | .data = &netdev_budget_usecs, | |
459 | .maxlen = sizeof(unsigned int), | |
460 | .mode = 0644, | |
461 | .proc_handler = proc_dointvec_minmax, | |
462 | .extra1 = &zero, | |
463 | }, | |
f8572d8f | 464 | { } |
1da177e4 | 465 | }; |
33eb9cfc | 466 | |
d5a4502e PE |
467 | static struct ctl_table netns_core_table[] = { |
468 | { | |
d5a4502e PE |
469 | .procname = "somaxconn", |
470 | .data = &init_net.core.sysctl_somaxconn, | |
471 | .maxlen = sizeof(int), | |
472 | .mode = 0644, | |
5f671d6b | 473 | .extra1 = &zero, |
5f671d6b | 474 | .proc_handler = proc_dointvec_minmax |
d5a4502e | 475 | }, |
f8572d8f | 476 | { } |
d5a4502e PE |
477 | }; |
478 | ||
024626e3 | 479 | static __net_init int sysctl_core_net_init(struct net *net) |
33eb9cfc | 480 | { |
d5a4502e | 481 | struct ctl_table *tbl; |
024626e3 | 482 | |
d5a4502e | 483 | tbl = netns_core_table; |
09ad9bc7 | 484 | if (!net_eq(net, &init_net)) { |
d5a4502e | 485 | tbl = kmemdup(tbl, sizeof(netns_core_table), GFP_KERNEL); |
024626e3 PE |
486 | if (tbl == NULL) |
487 | goto err_dup; | |
488 | ||
d5a4502e | 489 | tbl[0].data = &net->core.sysctl_somaxconn; |
464dc801 EB |
490 | |
491 | /* Don't export any sysctls to unprivileged users */ | |
492 | if (net->user_ns != &init_user_ns) { | |
493 | tbl[0].procname = NULL; | |
494 | } | |
024626e3 PE |
495 | } |
496 | ||
ec8f23ce | 497 | net->core.sysctl_hdr = register_net_sysctl(net, "net/core", tbl); |
8efa6e93 | 498 | if (net->core.sysctl_hdr == NULL) |
024626e3 | 499 | goto err_reg; |
33eb9cfc | 500 | |
024626e3 PE |
501 | return 0; |
502 | ||
503 | err_reg: | |
d5a4502e | 504 | if (tbl != netns_core_table) |
024626e3 PE |
505 | kfree(tbl); |
506 | err_dup: | |
507 | return -ENOMEM; | |
508 | } | |
509 | ||
510 | static __net_exit void sysctl_core_net_exit(struct net *net) | |
511 | { | |
512 | struct ctl_table *tbl; | |
513 | ||
8efa6e93 PE |
514 | tbl = net->core.sysctl_hdr->ctl_table_arg; |
515 | unregister_net_sysctl_table(net->core.sysctl_hdr); | |
d5a4502e | 516 | BUG_ON(tbl == netns_core_table); |
024626e3 PE |
517 | kfree(tbl); |
518 | } | |
519 | ||
520 | static __net_initdata struct pernet_operations sysctl_core_ops = { | |
521 | .init = sysctl_core_net_init, | |
522 | .exit = sysctl_core_net_exit, | |
523 | }; | |
524 | ||
525 | static __init int sysctl_core_init(void) | |
526 | { | |
43444757 | 527 | register_net_sysctl(&init_net, "net/core", net_core_table); |
024626e3 | 528 | return register_pernet_subsys(&sysctl_core_ops); |
33eb9cfc PE |
529 | } |
530 | ||
b27aeadb | 531 | fs_initcall(sysctl_core_init); |