Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * net/psample/psample.c - Netlink channel for packet sampling
4 * Copyright (c) 2017 Yotam Gigi <yotamg@mellanox.com>
5 */
6
7#include <linux/types.h>
8#include <linux/kernel.h>
9#include <linux/skbuff.h>
10#include <linux/module.h>
11#include <net/net_namespace.h>
12#include <net/sock.h>
13#include <net/netlink.h>
14#include <net/genetlink.h>
15#include <net/psample.h>
16#include <linux/spinlock.h>
17#include <net/ip_tunnels.h>
18#include <net/dst_metadata.h>
19
20#define PSAMPLE_MAX_PACKET_SIZE 0xffff
21
22static LIST_HEAD(psample_groups_list);
23static DEFINE_SPINLOCK(psample_groups_lock);
24
25/* multicast groups */
26enum psample_nl_multicast_groups {
27 PSAMPLE_NL_MCGRP_CONFIG,
28 PSAMPLE_NL_MCGRP_SAMPLE,
29};
30
31static const struct genl_multicast_group psample_nl_mcgrps[] = {
32 [PSAMPLE_NL_MCGRP_CONFIG] = { .name = PSAMPLE_NL_MCGRP_CONFIG_NAME },
33 [PSAMPLE_NL_MCGRP_SAMPLE] = { .name = PSAMPLE_NL_MCGRP_SAMPLE_NAME },
34};
35
36static struct genl_family psample_nl_family __ro_after_init;
37
38static int psample_group_nl_fill(struct sk_buff *msg,
39 struct psample_group *group,
40 enum psample_command cmd, u32 portid, u32 seq,
41 int flags)
42{
43 void *hdr;
44 int ret;
45
46 hdr = genlmsg_put(msg, portid, seq, &psample_nl_family, flags, cmd);
47 if (!hdr)
48 return -EMSGSIZE;
49
50 ret = nla_put_u32(msg, PSAMPLE_ATTR_SAMPLE_GROUP, group->group_num);
51 if (ret < 0)
52 goto error;
53
54 ret = nla_put_u32(msg, PSAMPLE_ATTR_GROUP_REFCOUNT, group->refcount);
55 if (ret < 0)
56 goto error;
57
58 ret = nla_put_u32(msg, PSAMPLE_ATTR_GROUP_SEQ, group->seq);
59 if (ret < 0)
60 goto error;
61
62 genlmsg_end(msg, hdr);
63 return 0;
64
65error:
66 genlmsg_cancel(msg, hdr);
67 return -EMSGSIZE;
68}
69
70static int psample_nl_cmd_get_group_dumpit(struct sk_buff *msg,
71 struct netlink_callback *cb)
72{
73 struct psample_group *group;
74 int start = cb->args[0];
75 int idx = 0;
76 int err;
77
78 spin_lock_bh(&psample_groups_lock);
79 list_for_each_entry(group, &psample_groups_list, list) {
80 if (!net_eq(group->net, sock_net(msg->sk)))
81 continue;
82 if (idx < start) {
83 idx++;
84 continue;
85 }
86 err = psample_group_nl_fill(msg, group, PSAMPLE_CMD_NEW_GROUP,
87 NETLINK_CB(cb->skb).portid,
88 cb->nlh->nlmsg_seq, NLM_F_MULTI);
89 if (err)
90 break;
91 idx++;
92 }
93
94 spin_unlock_bh(&psample_groups_lock);
95 cb->args[0] = idx;
96 return msg->len;
97}
98
99static const struct genl_ops psample_nl_ops[] = {
100 {
101 .cmd = PSAMPLE_CMD_GET_GROUP,
102 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
103 .dumpit = psample_nl_cmd_get_group_dumpit,
104 /* can be retrieved by unprivileged users */
105 }
106};
107
108static struct genl_family psample_nl_family __ro_after_init = {
109 .name = PSAMPLE_GENL_NAME,
110 .version = PSAMPLE_GENL_VERSION,
111 .maxattr = PSAMPLE_ATTR_MAX,
112 .netnsok = true,
113 .module = THIS_MODULE,
114 .mcgrps = psample_nl_mcgrps,
115 .ops = psample_nl_ops,
116 .n_ops = ARRAY_SIZE(psample_nl_ops),
117 .n_mcgrps = ARRAY_SIZE(psample_nl_mcgrps),
118};
119
120static void psample_group_notify(struct psample_group *group,
121 enum psample_command cmd)
122{
123 struct sk_buff *msg;
124 int err;
125
126 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
127 if (!msg)
128 return;
129
130 err = psample_group_nl_fill(msg, group, cmd, 0, 0, NLM_F_MULTI);
131 if (!err)
132 genlmsg_multicast_netns(&psample_nl_family, group->net, msg, 0,
133 PSAMPLE_NL_MCGRP_CONFIG, GFP_ATOMIC);
134 else
135 nlmsg_free(msg);
136}
137
138static struct psample_group *psample_group_create(struct net *net,
139 u32 group_num)
140{
141 struct psample_group *group;
142
143 group = kzalloc(sizeof(*group), GFP_ATOMIC);
144 if (!group)
145 return NULL;
146
147 group->net = net;
148 group->group_num = group_num;
149 list_add_tail(&group->list, &psample_groups_list);
150
151 psample_group_notify(group, PSAMPLE_CMD_NEW_GROUP);
152 return group;
153}
154
155static void psample_group_destroy(struct psample_group *group)
156{
157 psample_group_notify(group, PSAMPLE_CMD_DEL_GROUP);
158 list_del(&group->list);
159 kfree_rcu(group, rcu);
160}
161
162static struct psample_group *
163psample_group_lookup(struct net *net, u32 group_num)
164{
165 struct psample_group *group;
166
167 list_for_each_entry(group, &psample_groups_list, list)
168 if ((group->group_num == group_num) && (group->net == net))
169 return group;
170 return NULL;
171}
172
173struct psample_group *psample_group_get(struct net *net, u32 group_num)
174{
175 struct psample_group *group;
176
177 spin_lock_bh(&psample_groups_lock);
178
179 group = psample_group_lookup(net, group_num);
180 if (!group) {
181 group = psample_group_create(net, group_num);
182 if (!group)
183 goto out;
184 }
185 group->refcount++;
186
187out:
188 spin_unlock_bh(&psample_groups_lock);
189 return group;
190}
191EXPORT_SYMBOL_GPL(psample_group_get);
192
193void psample_group_take(struct psample_group *group)
194{
195 spin_lock_bh(&psample_groups_lock);
196 group->refcount++;
197 spin_unlock_bh(&psample_groups_lock);
198}
199EXPORT_SYMBOL_GPL(psample_group_take);
200
201void psample_group_put(struct psample_group *group)
202{
203 spin_lock_bh(&psample_groups_lock);
204
205 if (--group->refcount == 0)
206 psample_group_destroy(group);
207
208 spin_unlock_bh(&psample_groups_lock);
209}
210EXPORT_SYMBOL_GPL(psample_group_put);
211
212#ifdef CONFIG_INET
213static int __psample_ip_tun_to_nlattr(struct sk_buff *skb,
214 struct ip_tunnel_info *tun_info)
215{
216 unsigned short tun_proto = ip_tunnel_info_af(tun_info);
217 const void *tun_opts = ip_tunnel_info_opts(tun_info);
218 const struct ip_tunnel_key *tun_key = &tun_info->key;
219 int tun_opts_len = tun_info->options_len;
220
221 if (tun_key->tun_flags & TUNNEL_KEY &&
222 nla_put_be64(skb, PSAMPLE_TUNNEL_KEY_ATTR_ID, tun_key->tun_id,
223 PSAMPLE_TUNNEL_KEY_ATTR_PAD))
224 return -EMSGSIZE;
225
226 if (tun_info->mode & IP_TUNNEL_INFO_BRIDGE &&
227 nla_put_flag(skb, PSAMPLE_TUNNEL_KEY_ATTR_IPV4_INFO_BRIDGE))
228 return -EMSGSIZE;
229
230 switch (tun_proto) {
231 case AF_INET:
232 if (tun_key->u.ipv4.src &&
233 nla_put_in_addr(skb, PSAMPLE_TUNNEL_KEY_ATTR_IPV4_SRC,
234 tun_key->u.ipv4.src))
235 return -EMSGSIZE;
236 if (tun_key->u.ipv4.dst &&
237 nla_put_in_addr(skb, PSAMPLE_TUNNEL_KEY_ATTR_IPV4_DST,
238 tun_key->u.ipv4.dst))
239 return -EMSGSIZE;
240 break;
241 case AF_INET6:
242 if (!ipv6_addr_any(&tun_key->u.ipv6.src) &&
243 nla_put_in6_addr(skb, PSAMPLE_TUNNEL_KEY_ATTR_IPV6_SRC,
244 &tun_key->u.ipv6.src))
245 return -EMSGSIZE;
246 if (!ipv6_addr_any(&tun_key->u.ipv6.dst) &&
247 nla_put_in6_addr(skb, PSAMPLE_TUNNEL_KEY_ATTR_IPV6_DST,
248 &tun_key->u.ipv6.dst))
249 return -EMSGSIZE;
250 break;
251 }
252 if (tun_key->tos &&
253 nla_put_u8(skb, PSAMPLE_TUNNEL_KEY_ATTR_TOS, tun_key->tos))
254 return -EMSGSIZE;
255 if (nla_put_u8(skb, PSAMPLE_TUNNEL_KEY_ATTR_TTL, tun_key->ttl))
256 return -EMSGSIZE;
257 if ((tun_key->tun_flags & TUNNEL_DONT_FRAGMENT) &&
258 nla_put_flag(skb, PSAMPLE_TUNNEL_KEY_ATTR_DONT_FRAGMENT))
259 return -EMSGSIZE;
260 if ((tun_key->tun_flags & TUNNEL_CSUM) &&
261 nla_put_flag(skb, PSAMPLE_TUNNEL_KEY_ATTR_CSUM))
262 return -EMSGSIZE;
263 if (tun_key->tp_src &&
264 nla_put_be16(skb, PSAMPLE_TUNNEL_KEY_ATTR_TP_SRC, tun_key->tp_src))
265 return -EMSGSIZE;
266 if (tun_key->tp_dst &&
267 nla_put_be16(skb, PSAMPLE_TUNNEL_KEY_ATTR_TP_DST, tun_key->tp_dst))
268 return -EMSGSIZE;
269 if ((tun_key->tun_flags & TUNNEL_OAM) &&
270 nla_put_flag(skb, PSAMPLE_TUNNEL_KEY_ATTR_OAM))
271 return -EMSGSIZE;
272 if (tun_opts_len) {
273 if (tun_key->tun_flags & TUNNEL_GENEVE_OPT &&
274 nla_put(skb, PSAMPLE_TUNNEL_KEY_ATTR_GENEVE_OPTS,
275 tun_opts_len, tun_opts))
276 return -EMSGSIZE;
277 else if (tun_key->tun_flags & TUNNEL_ERSPAN_OPT &&
278 nla_put(skb, PSAMPLE_TUNNEL_KEY_ATTR_ERSPAN_OPTS,
279 tun_opts_len, tun_opts))
280 return -EMSGSIZE;
281 }
282
283 return 0;
284}
285
286static int psample_ip_tun_to_nlattr(struct sk_buff *skb,
287 struct ip_tunnel_info *tun_info)
288{
289 struct nlattr *nla;
290 int err;
291
292 nla = nla_nest_start_noflag(skb, PSAMPLE_ATTR_TUNNEL);
293 if (!nla)
294 return -EMSGSIZE;
295
296 err = __psample_ip_tun_to_nlattr(skb, tun_info);
297 if (err) {
298 nla_nest_cancel(skb, nla);
299 return err;
300 }
301
302 nla_nest_end(skb, nla);
303
304 return 0;
305}
306
307static int psample_tunnel_meta_len(struct ip_tunnel_info *tun_info)
308{
309 unsigned short tun_proto = ip_tunnel_info_af(tun_info);
310 const struct ip_tunnel_key *tun_key = &tun_info->key;
311 int tun_opts_len = tun_info->options_len;
312 int sum = 0;
313
314 if (tun_key->tun_flags & TUNNEL_KEY)
315 sum += nla_total_size(sizeof(u64));
316
317 if (tun_info->mode & IP_TUNNEL_INFO_BRIDGE)
318 sum += nla_total_size(0);
319
320 switch (tun_proto) {
321 case AF_INET:
322 if (tun_key->u.ipv4.src)
323 sum += nla_total_size(sizeof(u32));
324 if (tun_key->u.ipv4.dst)
325 sum += nla_total_size(sizeof(u32));
326 break;
327 case AF_INET6:
328 if (!ipv6_addr_any(&tun_key->u.ipv6.src))
329 sum += nla_total_size(sizeof(struct in6_addr));
330 if (!ipv6_addr_any(&tun_key->u.ipv6.dst))
331 sum += nla_total_size(sizeof(struct in6_addr));
332 break;
333 }
334 if (tun_key->tos)
335 sum += nla_total_size(sizeof(u8));
336 sum += nla_total_size(sizeof(u8)); /* TTL */
337 if (tun_key->tun_flags & TUNNEL_DONT_FRAGMENT)
338 sum += nla_total_size(0);
339 if (tun_key->tun_flags & TUNNEL_CSUM)
340 sum += nla_total_size(0);
341 if (tun_key->tp_src)
342 sum += nla_total_size(sizeof(u16));
343 if (tun_key->tp_dst)
344 sum += nla_total_size(sizeof(u16));
345 if (tun_key->tun_flags & TUNNEL_OAM)
346 sum += nla_total_size(0);
347 if (tun_opts_len) {
348 if (tun_key->tun_flags & TUNNEL_GENEVE_OPT)
349 sum += nla_total_size(tun_opts_len);
350 else if (tun_key->tun_flags & TUNNEL_ERSPAN_OPT)
351 sum += nla_total_size(tun_opts_len);
352 }
353
354 return sum;
355}
356#endif
357
358void psample_sample_packet(struct psample_group *group, struct sk_buff *skb,
359 u32 trunc_size, int in_ifindex, int out_ifindex,
360 u32 sample_rate)
361{
362#ifdef CONFIG_INET
363 struct ip_tunnel_info *tun_info;
364#endif
365 struct sk_buff *nl_skb;
366 int data_len;
367 int meta_len;
368 void *data;
369 int ret;
370
371 meta_len = (in_ifindex ? nla_total_size(sizeof(u16)) : 0) +
372 (out_ifindex ? nla_total_size(sizeof(u16)) : 0) +
373 nla_total_size(sizeof(u32)) + /* sample_rate */
374 nla_total_size(sizeof(u32)) + /* orig_size */
375 nla_total_size(sizeof(u32)) + /* group_num */
376 nla_total_size(sizeof(u32)); /* seq */
377
378#ifdef CONFIG_INET
379 tun_info = skb_tunnel_info(skb);
380 if (tun_info)
381 meta_len += psample_tunnel_meta_len(tun_info);
382#endif
383
384 data_len = min(skb->len, trunc_size);
385 if (meta_len + nla_total_size(data_len) > PSAMPLE_MAX_PACKET_SIZE)
386 data_len = PSAMPLE_MAX_PACKET_SIZE - meta_len - NLA_HDRLEN
387 - NLA_ALIGNTO;
388
389 nl_skb = genlmsg_new(meta_len + nla_total_size(data_len), GFP_ATOMIC);
390 if (unlikely(!nl_skb))
391 return;
392
393 data = genlmsg_put(nl_skb, 0, 0, &psample_nl_family, 0,
394 PSAMPLE_CMD_SAMPLE);
395 if (unlikely(!data))
396 goto error;
397
398 if (in_ifindex) {
399 ret = nla_put_u16(nl_skb, PSAMPLE_ATTR_IIFINDEX, in_ifindex);
400 if (unlikely(ret < 0))
401 goto error;
402 }
403
404 if (out_ifindex) {
405 ret = nla_put_u16(nl_skb, PSAMPLE_ATTR_OIFINDEX, out_ifindex);
406 if (unlikely(ret < 0))
407 goto error;
408 }
409
410 ret = nla_put_u32(nl_skb, PSAMPLE_ATTR_SAMPLE_RATE, sample_rate);
411 if (unlikely(ret < 0))
412 goto error;
413
414 ret = nla_put_u32(nl_skb, PSAMPLE_ATTR_ORIGSIZE, skb->len);
415 if (unlikely(ret < 0))
416 goto error;
417
418 ret = nla_put_u32(nl_skb, PSAMPLE_ATTR_SAMPLE_GROUP, group->group_num);
419 if (unlikely(ret < 0))
420 goto error;
421
422 ret = nla_put_u32(nl_skb, PSAMPLE_ATTR_GROUP_SEQ, group->seq++);
423 if (unlikely(ret < 0))
424 goto error;
425
426 if (data_len) {
427 int nla_len = nla_total_size(data_len);
428 struct nlattr *nla;
429
430 nla = skb_put(nl_skb, nla_len);
431 nla->nla_type = PSAMPLE_ATTR_DATA;
432 nla->nla_len = nla_attr_size(data_len);
433
434 if (skb_copy_bits(skb, 0, nla_data(nla), data_len))
435 goto error;
436 }
437
438#ifdef CONFIG_INET
439 if (tun_info) {
440 ret = psample_ip_tun_to_nlattr(nl_skb, tun_info);
441 if (unlikely(ret < 0))
442 goto error;
443 }
444#endif
445
446 genlmsg_end(nl_skb, data);
447 genlmsg_multicast_netns(&psample_nl_family, group->net, nl_skb, 0,
448 PSAMPLE_NL_MCGRP_SAMPLE, GFP_ATOMIC);
449
450 return;
451error:
452 pr_err_ratelimited("Could not create psample log message\n");
453 nlmsg_free(nl_skb);
454}
455EXPORT_SYMBOL_GPL(psample_sample_packet);
456
457static int __init psample_module_init(void)
458{
459 return genl_register_family(&psample_nl_family);
460}
461
462static void __exit psample_module_exit(void)
463{
464 genl_unregister_family(&psample_nl_family);
465}
466
467module_init(psample_module_init);
468module_exit(psample_module_exit);
469
470MODULE_AUTHOR("Yotam Gigi <yotam.gi@gmail.com>");
471MODULE_DESCRIPTION("netlink channel for packet sampling");
472MODULE_LICENSE("GPL v2");
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * net/psample/psample.c - Netlink channel for packet sampling
4 * Copyright (c) 2017 Yotam Gigi <yotamg@mellanox.com>
5 */
6
7#include <linux/types.h>
8#include <linux/kernel.h>
9#include <linux/skbuff.h>
10#include <linux/module.h>
11#include <linux/timekeeping.h>
12#include <net/net_namespace.h>
13#include <net/sock.h>
14#include <net/netlink.h>
15#include <net/genetlink.h>
16#include <net/psample.h>
17#include <linux/spinlock.h>
18#include <net/ip_tunnels.h>
19#include <net/dst_metadata.h>
20
21#define PSAMPLE_MAX_PACKET_SIZE 0xffff
22
23static LIST_HEAD(psample_groups_list);
24static DEFINE_SPINLOCK(psample_groups_lock);
25
26/* multicast groups */
27enum psample_nl_multicast_groups {
28 PSAMPLE_NL_MCGRP_CONFIG,
29 PSAMPLE_NL_MCGRP_SAMPLE,
30};
31
32static const struct genl_multicast_group psample_nl_mcgrps[] = {
33 [PSAMPLE_NL_MCGRP_CONFIG] = { .name = PSAMPLE_NL_MCGRP_CONFIG_NAME },
34 [PSAMPLE_NL_MCGRP_SAMPLE] = { .name = PSAMPLE_NL_MCGRP_SAMPLE_NAME,
35 .flags = GENL_MCAST_CAP_NET_ADMIN, },
36};
37
38static struct genl_family psample_nl_family __ro_after_init;
39
40static int psample_group_nl_fill(struct sk_buff *msg,
41 struct psample_group *group,
42 enum psample_command cmd, u32 portid, u32 seq,
43 int flags)
44{
45 void *hdr;
46 int ret;
47
48 hdr = genlmsg_put(msg, portid, seq, &psample_nl_family, flags, cmd);
49 if (!hdr)
50 return -EMSGSIZE;
51
52 ret = nla_put_u32(msg, PSAMPLE_ATTR_SAMPLE_GROUP, group->group_num);
53 if (ret < 0)
54 goto error;
55
56 ret = nla_put_u32(msg, PSAMPLE_ATTR_GROUP_REFCOUNT, group->refcount);
57 if (ret < 0)
58 goto error;
59
60 ret = nla_put_u32(msg, PSAMPLE_ATTR_GROUP_SEQ, group->seq);
61 if (ret < 0)
62 goto error;
63
64 genlmsg_end(msg, hdr);
65 return 0;
66
67error:
68 genlmsg_cancel(msg, hdr);
69 return -EMSGSIZE;
70}
71
72static int psample_nl_cmd_get_group_dumpit(struct sk_buff *msg,
73 struct netlink_callback *cb)
74{
75 struct psample_group *group;
76 int start = cb->args[0];
77 int idx = 0;
78 int err;
79
80 spin_lock_bh(&psample_groups_lock);
81 list_for_each_entry(group, &psample_groups_list, list) {
82 if (!net_eq(group->net, sock_net(msg->sk)))
83 continue;
84 if (idx < start) {
85 idx++;
86 continue;
87 }
88 err = psample_group_nl_fill(msg, group, PSAMPLE_CMD_NEW_GROUP,
89 NETLINK_CB(cb->skb).portid,
90 cb->nlh->nlmsg_seq, NLM_F_MULTI);
91 if (err)
92 break;
93 idx++;
94 }
95
96 spin_unlock_bh(&psample_groups_lock);
97 cb->args[0] = idx;
98 return msg->len;
99}
100
101static const struct genl_small_ops psample_nl_ops[] = {
102 {
103 .cmd = PSAMPLE_CMD_GET_GROUP,
104 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
105 .dumpit = psample_nl_cmd_get_group_dumpit,
106 /* can be retrieved by unprivileged users */
107 }
108};
109
110static struct genl_family psample_nl_family __ro_after_init = {
111 .name = PSAMPLE_GENL_NAME,
112 .version = PSAMPLE_GENL_VERSION,
113 .maxattr = PSAMPLE_ATTR_MAX,
114 .netnsok = true,
115 .module = THIS_MODULE,
116 .mcgrps = psample_nl_mcgrps,
117 .small_ops = psample_nl_ops,
118 .n_small_ops = ARRAY_SIZE(psample_nl_ops),
119 .resv_start_op = PSAMPLE_CMD_GET_GROUP + 1,
120 .n_mcgrps = ARRAY_SIZE(psample_nl_mcgrps),
121};
122
123static void psample_group_notify(struct psample_group *group,
124 enum psample_command cmd)
125{
126 struct sk_buff *msg;
127 int err;
128
129 msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
130 if (!msg)
131 return;
132
133 err = psample_group_nl_fill(msg, group, cmd, 0, 0, NLM_F_MULTI);
134 if (!err)
135 genlmsg_multicast_netns(&psample_nl_family, group->net, msg, 0,
136 PSAMPLE_NL_MCGRP_CONFIG, GFP_ATOMIC);
137 else
138 nlmsg_free(msg);
139}
140
141static struct psample_group *psample_group_create(struct net *net,
142 u32 group_num)
143{
144 struct psample_group *group;
145
146 group = kzalloc(sizeof(*group), GFP_ATOMIC);
147 if (!group)
148 return NULL;
149
150 group->net = net;
151 group->group_num = group_num;
152 list_add_tail(&group->list, &psample_groups_list);
153
154 psample_group_notify(group, PSAMPLE_CMD_NEW_GROUP);
155 return group;
156}
157
158static void psample_group_destroy(struct psample_group *group)
159{
160 psample_group_notify(group, PSAMPLE_CMD_DEL_GROUP);
161 list_del(&group->list);
162 kfree_rcu(group, rcu);
163}
164
165static struct psample_group *
166psample_group_lookup(struct net *net, u32 group_num)
167{
168 struct psample_group *group;
169
170 list_for_each_entry(group, &psample_groups_list, list)
171 if ((group->group_num == group_num) && (group->net == net))
172 return group;
173 return NULL;
174}
175
176struct psample_group *psample_group_get(struct net *net, u32 group_num)
177{
178 struct psample_group *group;
179
180 spin_lock_bh(&psample_groups_lock);
181
182 group = psample_group_lookup(net, group_num);
183 if (!group) {
184 group = psample_group_create(net, group_num);
185 if (!group)
186 goto out;
187 }
188 group->refcount++;
189
190out:
191 spin_unlock_bh(&psample_groups_lock);
192 return group;
193}
194EXPORT_SYMBOL_GPL(psample_group_get);
195
196void psample_group_take(struct psample_group *group)
197{
198 spin_lock_bh(&psample_groups_lock);
199 group->refcount++;
200 spin_unlock_bh(&psample_groups_lock);
201}
202EXPORT_SYMBOL_GPL(psample_group_take);
203
204void psample_group_put(struct psample_group *group)
205{
206 spin_lock_bh(&psample_groups_lock);
207
208 if (--group->refcount == 0)
209 psample_group_destroy(group);
210
211 spin_unlock_bh(&psample_groups_lock);
212}
213EXPORT_SYMBOL_GPL(psample_group_put);
214
215#ifdef CONFIG_INET
216static int __psample_ip_tun_to_nlattr(struct sk_buff *skb,
217 struct ip_tunnel_info *tun_info)
218{
219 unsigned short tun_proto = ip_tunnel_info_af(tun_info);
220 const void *tun_opts = ip_tunnel_info_opts(tun_info);
221 const struct ip_tunnel_key *tun_key = &tun_info->key;
222 int tun_opts_len = tun_info->options_len;
223
224 if (tun_key->tun_flags & TUNNEL_KEY &&
225 nla_put_be64(skb, PSAMPLE_TUNNEL_KEY_ATTR_ID, tun_key->tun_id,
226 PSAMPLE_TUNNEL_KEY_ATTR_PAD))
227 return -EMSGSIZE;
228
229 if (tun_info->mode & IP_TUNNEL_INFO_BRIDGE &&
230 nla_put_flag(skb, PSAMPLE_TUNNEL_KEY_ATTR_IPV4_INFO_BRIDGE))
231 return -EMSGSIZE;
232
233 switch (tun_proto) {
234 case AF_INET:
235 if (tun_key->u.ipv4.src &&
236 nla_put_in_addr(skb, PSAMPLE_TUNNEL_KEY_ATTR_IPV4_SRC,
237 tun_key->u.ipv4.src))
238 return -EMSGSIZE;
239 if (tun_key->u.ipv4.dst &&
240 nla_put_in_addr(skb, PSAMPLE_TUNNEL_KEY_ATTR_IPV4_DST,
241 tun_key->u.ipv4.dst))
242 return -EMSGSIZE;
243 break;
244 case AF_INET6:
245 if (!ipv6_addr_any(&tun_key->u.ipv6.src) &&
246 nla_put_in6_addr(skb, PSAMPLE_TUNNEL_KEY_ATTR_IPV6_SRC,
247 &tun_key->u.ipv6.src))
248 return -EMSGSIZE;
249 if (!ipv6_addr_any(&tun_key->u.ipv6.dst) &&
250 nla_put_in6_addr(skb, PSAMPLE_TUNNEL_KEY_ATTR_IPV6_DST,
251 &tun_key->u.ipv6.dst))
252 return -EMSGSIZE;
253 break;
254 }
255 if (tun_key->tos &&
256 nla_put_u8(skb, PSAMPLE_TUNNEL_KEY_ATTR_TOS, tun_key->tos))
257 return -EMSGSIZE;
258 if (nla_put_u8(skb, PSAMPLE_TUNNEL_KEY_ATTR_TTL, tun_key->ttl))
259 return -EMSGSIZE;
260 if ((tun_key->tun_flags & TUNNEL_DONT_FRAGMENT) &&
261 nla_put_flag(skb, PSAMPLE_TUNNEL_KEY_ATTR_DONT_FRAGMENT))
262 return -EMSGSIZE;
263 if ((tun_key->tun_flags & TUNNEL_CSUM) &&
264 nla_put_flag(skb, PSAMPLE_TUNNEL_KEY_ATTR_CSUM))
265 return -EMSGSIZE;
266 if (tun_key->tp_src &&
267 nla_put_be16(skb, PSAMPLE_TUNNEL_KEY_ATTR_TP_SRC, tun_key->tp_src))
268 return -EMSGSIZE;
269 if (tun_key->tp_dst &&
270 nla_put_be16(skb, PSAMPLE_TUNNEL_KEY_ATTR_TP_DST, tun_key->tp_dst))
271 return -EMSGSIZE;
272 if ((tun_key->tun_flags & TUNNEL_OAM) &&
273 nla_put_flag(skb, PSAMPLE_TUNNEL_KEY_ATTR_OAM))
274 return -EMSGSIZE;
275 if (tun_opts_len) {
276 if (tun_key->tun_flags & TUNNEL_GENEVE_OPT &&
277 nla_put(skb, PSAMPLE_TUNNEL_KEY_ATTR_GENEVE_OPTS,
278 tun_opts_len, tun_opts))
279 return -EMSGSIZE;
280 else if (tun_key->tun_flags & TUNNEL_ERSPAN_OPT &&
281 nla_put(skb, PSAMPLE_TUNNEL_KEY_ATTR_ERSPAN_OPTS,
282 tun_opts_len, tun_opts))
283 return -EMSGSIZE;
284 }
285
286 return 0;
287}
288
289static int psample_ip_tun_to_nlattr(struct sk_buff *skb,
290 struct ip_tunnel_info *tun_info)
291{
292 struct nlattr *nla;
293 int err;
294
295 nla = nla_nest_start_noflag(skb, PSAMPLE_ATTR_TUNNEL);
296 if (!nla)
297 return -EMSGSIZE;
298
299 err = __psample_ip_tun_to_nlattr(skb, tun_info);
300 if (err) {
301 nla_nest_cancel(skb, nla);
302 return err;
303 }
304
305 nla_nest_end(skb, nla);
306
307 return 0;
308}
309
310static int psample_tunnel_meta_len(struct ip_tunnel_info *tun_info)
311{
312 unsigned short tun_proto = ip_tunnel_info_af(tun_info);
313 const struct ip_tunnel_key *tun_key = &tun_info->key;
314 int tun_opts_len = tun_info->options_len;
315 int sum = nla_total_size(0); /* PSAMPLE_ATTR_TUNNEL */
316
317 if (tun_key->tun_flags & TUNNEL_KEY)
318 sum += nla_total_size_64bit(sizeof(u64));
319
320 if (tun_info->mode & IP_TUNNEL_INFO_BRIDGE)
321 sum += nla_total_size(0);
322
323 switch (tun_proto) {
324 case AF_INET:
325 if (tun_key->u.ipv4.src)
326 sum += nla_total_size(sizeof(u32));
327 if (tun_key->u.ipv4.dst)
328 sum += nla_total_size(sizeof(u32));
329 break;
330 case AF_INET6:
331 if (!ipv6_addr_any(&tun_key->u.ipv6.src))
332 sum += nla_total_size(sizeof(struct in6_addr));
333 if (!ipv6_addr_any(&tun_key->u.ipv6.dst))
334 sum += nla_total_size(sizeof(struct in6_addr));
335 break;
336 }
337 if (tun_key->tos)
338 sum += nla_total_size(sizeof(u8));
339 sum += nla_total_size(sizeof(u8)); /* TTL */
340 if (tun_key->tun_flags & TUNNEL_DONT_FRAGMENT)
341 sum += nla_total_size(0);
342 if (tun_key->tun_flags & TUNNEL_CSUM)
343 sum += nla_total_size(0);
344 if (tun_key->tp_src)
345 sum += nla_total_size(sizeof(u16));
346 if (tun_key->tp_dst)
347 sum += nla_total_size(sizeof(u16));
348 if (tun_key->tun_flags & TUNNEL_OAM)
349 sum += nla_total_size(0);
350 if (tun_opts_len) {
351 if (tun_key->tun_flags & TUNNEL_GENEVE_OPT)
352 sum += nla_total_size(tun_opts_len);
353 else if (tun_key->tun_flags & TUNNEL_ERSPAN_OPT)
354 sum += nla_total_size(tun_opts_len);
355 }
356
357 return sum;
358}
359#endif
360
361void psample_sample_packet(struct psample_group *group, struct sk_buff *skb,
362 u32 sample_rate, const struct psample_metadata *md)
363{
364 ktime_t tstamp = ktime_get_real();
365 int out_ifindex = md->out_ifindex;
366 int in_ifindex = md->in_ifindex;
367 u32 trunc_size = md->trunc_size;
368#ifdef CONFIG_INET
369 struct ip_tunnel_info *tun_info;
370#endif
371 struct sk_buff *nl_skb;
372 int data_len;
373 int meta_len;
374 void *data;
375 int ret;
376
377 meta_len = (in_ifindex ? nla_total_size(sizeof(u16)) : 0) +
378 (out_ifindex ? nla_total_size(sizeof(u16)) : 0) +
379 (md->out_tc_valid ? nla_total_size(sizeof(u16)) : 0) +
380 (md->out_tc_occ_valid ? nla_total_size_64bit(sizeof(u64)) : 0) +
381 (md->latency_valid ? nla_total_size_64bit(sizeof(u64)) : 0) +
382 nla_total_size(sizeof(u32)) + /* sample_rate */
383 nla_total_size(sizeof(u32)) + /* orig_size */
384 nla_total_size(sizeof(u32)) + /* group_num */
385 nla_total_size(sizeof(u32)) + /* seq */
386 nla_total_size_64bit(sizeof(u64)) + /* timestamp */
387 nla_total_size(sizeof(u16)); /* protocol */
388
389#ifdef CONFIG_INET
390 tun_info = skb_tunnel_info(skb);
391 if (tun_info)
392 meta_len += psample_tunnel_meta_len(tun_info);
393#endif
394
395 data_len = min(skb->len, trunc_size);
396 if (meta_len + nla_total_size(data_len) > PSAMPLE_MAX_PACKET_SIZE)
397 data_len = PSAMPLE_MAX_PACKET_SIZE - meta_len - NLA_HDRLEN
398 - NLA_ALIGNTO;
399
400 nl_skb = genlmsg_new(meta_len + nla_total_size(data_len), GFP_ATOMIC);
401 if (unlikely(!nl_skb))
402 return;
403
404 data = genlmsg_put(nl_skb, 0, 0, &psample_nl_family, 0,
405 PSAMPLE_CMD_SAMPLE);
406 if (unlikely(!data))
407 goto error;
408
409 if (in_ifindex) {
410 ret = nla_put_u16(nl_skb, PSAMPLE_ATTR_IIFINDEX, in_ifindex);
411 if (unlikely(ret < 0))
412 goto error;
413 }
414
415 if (out_ifindex) {
416 ret = nla_put_u16(nl_skb, PSAMPLE_ATTR_OIFINDEX, out_ifindex);
417 if (unlikely(ret < 0))
418 goto error;
419 }
420
421 ret = nla_put_u32(nl_skb, PSAMPLE_ATTR_SAMPLE_RATE, sample_rate);
422 if (unlikely(ret < 0))
423 goto error;
424
425 ret = nla_put_u32(nl_skb, PSAMPLE_ATTR_ORIGSIZE, skb->len);
426 if (unlikely(ret < 0))
427 goto error;
428
429 ret = nla_put_u32(nl_skb, PSAMPLE_ATTR_SAMPLE_GROUP, group->group_num);
430 if (unlikely(ret < 0))
431 goto error;
432
433 ret = nla_put_u32(nl_skb, PSAMPLE_ATTR_GROUP_SEQ, group->seq++);
434 if (unlikely(ret < 0))
435 goto error;
436
437 if (md->out_tc_valid) {
438 ret = nla_put_u16(nl_skb, PSAMPLE_ATTR_OUT_TC, md->out_tc);
439 if (unlikely(ret < 0))
440 goto error;
441 }
442
443 if (md->out_tc_occ_valid) {
444 ret = nla_put_u64_64bit(nl_skb, PSAMPLE_ATTR_OUT_TC_OCC,
445 md->out_tc_occ, PSAMPLE_ATTR_PAD);
446 if (unlikely(ret < 0))
447 goto error;
448 }
449
450 if (md->latency_valid) {
451 ret = nla_put_u64_64bit(nl_skb, PSAMPLE_ATTR_LATENCY,
452 md->latency, PSAMPLE_ATTR_PAD);
453 if (unlikely(ret < 0))
454 goto error;
455 }
456
457 ret = nla_put_u64_64bit(nl_skb, PSAMPLE_ATTR_TIMESTAMP,
458 ktime_to_ns(tstamp), PSAMPLE_ATTR_PAD);
459 if (unlikely(ret < 0))
460 goto error;
461
462 ret = nla_put_u16(nl_skb, PSAMPLE_ATTR_PROTO,
463 be16_to_cpu(skb->protocol));
464 if (unlikely(ret < 0))
465 goto error;
466
467 if (data_len) {
468 int nla_len = nla_total_size(data_len);
469 struct nlattr *nla;
470
471 nla = skb_put(nl_skb, nla_len);
472 nla->nla_type = PSAMPLE_ATTR_DATA;
473 nla->nla_len = nla_attr_size(data_len);
474
475 if (skb_copy_bits(skb, 0, nla_data(nla), data_len))
476 goto error;
477 }
478
479#ifdef CONFIG_INET
480 if (tun_info) {
481 ret = psample_ip_tun_to_nlattr(nl_skb, tun_info);
482 if (unlikely(ret < 0))
483 goto error;
484 }
485#endif
486
487 genlmsg_end(nl_skb, data);
488 genlmsg_multicast_netns(&psample_nl_family, group->net, nl_skb, 0,
489 PSAMPLE_NL_MCGRP_SAMPLE, GFP_ATOMIC);
490
491 return;
492error:
493 pr_err_ratelimited("Could not create psample log message\n");
494 nlmsg_free(nl_skb);
495}
496EXPORT_SYMBOL_GPL(psample_sample_packet);
497
498static int __init psample_module_init(void)
499{
500 return genl_register_family(&psample_nl_family);
501}
502
503static void __exit psample_module_exit(void)
504{
505 genl_unregister_family(&psample_nl_family);
506}
507
508module_init(psample_module_init);
509module_exit(psample_module_exit);
510
511MODULE_AUTHOR("Yotam Gigi <yotam.gi@gmail.com>");
512MODULE_DESCRIPTION("netlink channel for packet sampling");
513MODULE_LICENSE("GPL v2");