Merge branch 'aux-bus-v11' of https://github.com/ajitkhaparde1/linux
[linux-block.git] / net / sched / act_sample.c
CommitLineData
d2912cb1 1// SPDX-License-Identifier: GPL-2.0-only
5c5670fa
YG
2/*
3 * net/sched/act_sample.c - Packet sampling tc action
4 * Copyright (c) 2017 Yotam Gigi <yotamg@mellanox.com>
5c5670fa
YG
5 */
6
7#include <linux/types.h>
8#include <linux/kernel.h>
9#include <linux/string.h>
10#include <linux/errno.h>
11#include <linux/skbuff.h>
12#include <linux/rtnetlink.h>
13#include <linux/module.h>
14#include <linux/init.h>
15#include <linux/gfp.h>
16#include <net/net_namespace.h>
17#include <net/netlink.h>
18#include <net/pkt_sched.h>
19#include <linux/tc_act/tc_sample.h>
20#include <net/tc_act/tc_sample.h>
21#include <net/psample.h>
e8c87c64 22#include <net/pkt_cls.h>
871cf386 23#include <net/tc_wrapper.h>
5c5670fa
YG
24
25#include <linux/if_arp.h>
26
5c5670fa
YG
27static struct tc_action_ops act_sample_ops;
28
29static const struct nla_policy sample_policy[TCA_SAMPLE_MAX + 1] = {
30 [TCA_SAMPLE_PARMS] = { .len = sizeof(struct tc_sample) },
31 [TCA_SAMPLE_RATE] = { .type = NLA_U32 },
32 [TCA_SAMPLE_TRUNC_SIZE] = { .type = NLA_U32 },
33 [TCA_SAMPLE_PSAMPLE_GROUP] = { .type = NLA_U32 },
34};
35
36static int tcf_sample_init(struct net *net, struct nlattr *nla,
695176bf
CW
37 struct nlattr *est, struct tc_action **a,
38 struct tcf_proto *tp,
abbb0d33 39 u32 flags, struct netlink_ext_ack *extack)
5c5670fa 40{
acd0a7ab 41 struct tc_action_net *tn = net_generic(net, act_sample_ops.net_id);
695176bf 42 bool bind = flags & TCA_ACT_FLAGS_BIND;
5c5670fa
YG
43 struct nlattr *tb[TCA_SAMPLE_MAX + 1];
44 struct psample_group *psample_group;
7be8ef2c 45 u32 psample_group_num, rate, index;
e8c87c64 46 struct tcf_chain *goto_ch = NULL;
5c5670fa
YG
47 struct tc_sample *parm;
48 struct tcf_sample *s;
49 bool exists = false;
0190c1d4 50 int ret, err;
5c5670fa
YG
51
52 if (!nla)
53 return -EINVAL;
8cb08174
JB
54 ret = nla_parse_nested_deprecated(tb, TCA_SAMPLE_MAX, nla,
55 sample_policy, NULL);
5c5670fa
YG
56 if (ret < 0)
57 return ret;
58 if (!tb[TCA_SAMPLE_PARMS] || !tb[TCA_SAMPLE_RATE] ||
59 !tb[TCA_SAMPLE_PSAMPLE_GROUP])
60 return -EINVAL;
61
62 parm = nla_data(tb[TCA_SAMPLE_PARMS]);
7be8ef2c
DL
63 index = parm->index;
64 err = tcf_idr_check_alloc(tn, &index, a, bind);
0190c1d4
VB
65 if (err < 0)
66 return err;
67 exists = err;
5c5670fa
YG
68 if (exists && bind)
69 return 0;
70
71 if (!exists) {
7be8ef2c 72 ret = tcf_idr_create(tn, index, est, a,
40bd094d 73 &act_sample_ops, bind, true, flags);
0190c1d4 74 if (ret) {
7be8ef2c 75 tcf_idr_cleanup(tn, index);
5c5670fa 76 return ret;
0190c1d4 77 }
5c5670fa 78 ret = ACT_P_CREATED;
695176bf 79 } else if (!(flags & TCA_ACT_FLAGS_REPLACE)) {
65a206c0 80 tcf_idr_release(*a, bind);
4e8ddd7f 81 return -EEXIST;
5c5670fa 82 }
e8c87c64
DC
83 err = tcf_action_check_ctrlact(parm->action, tp, &goto_ch, extack);
84 if (err < 0)
85 goto release_idr;
5c5670fa 86
fae27081
DC
87 rate = nla_get_u32(tb[TCA_SAMPLE_RATE]);
88 if (!rate) {
89 NL_SET_ERR_MSG(extack, "invalid sample rate");
90 err = -EINVAL;
91 goto put_chain;
92 }
653cd284
VB
93 psample_group_num = nla_get_u32(tb[TCA_SAMPLE_PSAMPLE_GROUP]);
94 psample_group = psample_group_get(net, psample_group_num);
cadb9c9f 95 if (!psample_group) {
e8c87c64
DC
96 err = -ENOMEM;
97 goto put_chain;
cadb9c9f 98 }
653cd284
VB
99
100 s = to_sample(*a);
101
102 spin_lock_bh(&s->tcf_lock);
e8c87c64 103 goto_ch = tcf_action_set_ctrlact(*a, parm->action, goto_ch);
fae27081 104 s->rate = rate;
653cd284 105 s->psample_group_num = psample_group_num;
445d3749
PM
106 psample_group = rcu_replace_pointer(s->psample_group, psample_group,
107 lockdep_is_held(&s->tcf_lock));
5c5670fa
YG
108
109 if (tb[TCA_SAMPLE_TRUNC_SIZE]) {
110 s->truncate = true;
111 s->trunc_size = nla_get_u32(tb[TCA_SAMPLE_TRUNC_SIZE]);
112 }
653cd284 113 spin_unlock_bh(&s->tcf_lock);
dbf47a2a
VB
114
115 if (psample_group)
116 psample_group_put(psample_group);
e8c87c64
DC
117 if (goto_ch)
118 tcf_chain_put_by_act(goto_ch);
5c5670fa 119
5c5670fa 120 return ret;
e8c87c64
DC
121put_chain:
122 if (goto_ch)
123 tcf_chain_put_by_act(goto_ch);
124release_idr:
125 tcf_idr_release(*a, bind);
126 return err;
5c5670fa
YG
127}
128
9a63b255 129static void tcf_sample_cleanup(struct tc_action *a)
5c5670fa 130{
90a6ec85 131 struct tcf_sample *s = to_sample(a);
5c5670fa
YG
132 struct psample_group *psample_group;
133
d7728495
VB
134 /* last reference to action, no need to lock */
135 psample_group = rcu_dereference_protected(s->psample_group, 1);
5c5670fa 136 RCU_INIT_POINTER(s->psample_group, NULL);
1f110e7c
DC
137 if (psample_group)
138 psample_group_put(psample_group);
5c5670fa
YG
139}
140
5c5670fa
YG
141static bool tcf_sample_dev_ok_push(struct net_device *dev)
142{
143 switch (dev->type) {
144 case ARPHRD_TUNNEL:
145 case ARPHRD_TUNNEL6:
146 case ARPHRD_SIT:
147 case ARPHRD_IPGRE:
92974a1d 148 case ARPHRD_IP6GRE:
5c5670fa
YG
149 case ARPHRD_VOID:
150 case ARPHRD_NONE:
151 return false;
152 default:
153 return true;
154 }
155}
156
871cf386
PT
157TC_INDIRECT_SCOPE int tcf_sample_act(struct sk_buff *skb,
158 const struct tc_action *a,
159 struct tcf_result *res)
5c5670fa
YG
160{
161 struct tcf_sample *s = to_sample(a);
162 struct psample_group *psample_group;
a03e99d3 163 struct psample_metadata md = {};
5c5670fa 164 int retval;
5c5670fa
YG
165
166 tcf_lastuse_update(&s->tcf_tm);
50dc9a85 167 bstats_update(this_cpu_ptr(s->common.cpu_bstats), skb);
5c5670fa
YG
168 retval = READ_ONCE(s->tcf_action);
169
7fd4b288 170 psample_group = rcu_dereference_bh(s->psample_group);
5c5670fa
YG
171
172 /* randomly sample packets according to rate */
8032bf12 173 if (psample_group && (get_random_u32_below(s->rate) == 0)) {
5c5670fa 174 if (!skb_at_tc_ingress(skb)) {
a03e99d3
IS
175 md.in_ifindex = skb->skb_iif;
176 md.out_ifindex = skb->dev->ifindex;
5c5670fa 177 } else {
a03e99d3 178 md.in_ifindex = skb->dev->ifindex;
5c5670fa
YG
179 }
180
181 /* on ingress, the mac header gets popped, so push it back */
182 if (skb_at_tc_ingress(skb) && tcf_sample_dev_ok_push(skb->dev))
183 skb_push(skb, skb->mac_len);
184
a03e99d3
IS
185 md.trunc_size = s->truncate ? s->trunc_size : skb->len;
186 psample_sample_packet(psample_group, skb, s->rate, &md);
5c5670fa
YG
187
188 if (skb_at_tc_ingress(skb) && tcf_sample_dev_ok_push(skb->dev))
189 skb_pull(skb, skb->mac_len);
190 }
191
5c5670fa
YG
192 return retval;
193}
194
58c04397
IS
195static void tcf_sample_stats_update(struct tc_action *a, u64 bytes, u64 packets,
196 u64 drops, u64 lastuse, bool hw)
197{
198 struct tcf_sample *s = to_sample(a);
199 struct tcf_t *tm = &s->tcf_tm;
200
201 tcf_action_update_stats(a, bytes, packets, drops, hw);
202 tm->lastuse = max_t(u64, tm->lastuse, lastuse);
203}
204
5c5670fa
YG
205static int tcf_sample_dump(struct sk_buff *skb, struct tc_action *a,
206 int bind, int ref)
207{
208 unsigned char *b = skb_tail_pointer(skb);
209 struct tcf_sample *s = to_sample(a);
210 struct tc_sample opt = {
211 .index = s->tcf_index,
036bb443
VB
212 .refcnt = refcount_read(&s->tcf_refcnt) - ref,
213 .bindcnt = atomic_read(&s->tcf_bindcnt) - bind,
5c5670fa
YG
214 };
215 struct tcf_t t;
216
653cd284 217 spin_lock_bh(&s->tcf_lock);
d7728495 218 opt.action = s->tcf_action;
5c5670fa
YG
219 if (nla_put(skb, TCA_SAMPLE_PARMS, sizeof(opt), &opt))
220 goto nla_put_failure;
221
222 tcf_tm_dump(&t, &s->tcf_tm);
223 if (nla_put_64bit(skb, TCA_SAMPLE_TM, sizeof(t), &t, TCA_SAMPLE_PAD))
224 goto nla_put_failure;
225
226 if (nla_put_u32(skb, TCA_SAMPLE_RATE, s->rate))
227 goto nla_put_failure;
228
229 if (s->truncate)
230 if (nla_put_u32(skb, TCA_SAMPLE_TRUNC_SIZE, s->trunc_size))
231 goto nla_put_failure;
232
233 if (nla_put_u32(skb, TCA_SAMPLE_PSAMPLE_GROUP, s->psample_group_num))
234 goto nla_put_failure;
653cd284 235 spin_unlock_bh(&s->tcf_lock);
d7728495 236
5c5670fa
YG
237 return skb->len;
238
239nla_put_failure:
653cd284 240 spin_unlock_bh(&s->tcf_lock);
5c5670fa
YG
241 nlmsg_trim(skb, b);
242 return -1;
243}
244
4a5da47d
VB
245static void tcf_psample_group_put(void *priv)
246{
247 struct psample_group *group = priv;
248
249 psample_group_put(group);
250}
251
252static struct psample_group *
253tcf_sample_get_group(const struct tc_action *a,
254 tc_action_priv_destructor *destructor)
255{
256 struct tcf_sample *s = to_sample(a);
257 struct psample_group *group;
258
4a5da47d
VB
259 group = rcu_dereference_protected(s->psample_group,
260 lockdep_is_held(&s->tcf_lock));
261 if (group) {
262 psample_group_take(group);
263 *destructor = tcf_psample_group_put;
264 }
4a5da47d
VB
265
266 return group;
267}
268
c54e1d92
BZ
269static void tcf_offload_sample_get_group(struct flow_action_entry *entry,
270 const struct tc_action *act)
271{
272 entry->sample.psample_group =
273 act->ops->get_psample_group(act, &entry->destructor);
274 entry->destructor_priv = entry->sample.psample_group;
275}
276
277static int tcf_sample_offload_act_setup(struct tc_action *act, void *entry_data,
c2ccf84e
IS
278 u32 *index_inc, bool bind,
279 struct netlink_ext_ack *extack)
c54e1d92
BZ
280{
281 if (bind) {
282 struct flow_action_entry *entry = entry_data;
283
284 entry->id = FLOW_ACTION_SAMPLE;
285 entry->sample.trunc_size = tcf_sample_trunc_size(act);
286 entry->sample.truncate = tcf_sample_truncate(act);
287 entry->sample.rate = tcf_sample_rate(act);
288 tcf_offload_sample_get_group(entry, act);
289 *index_inc = 1;
290 } else {
8cbfe939
BZ
291 struct flow_offload_action *fl_action = entry_data;
292
293 fl_action->id = FLOW_ACTION_SAMPLE;
c54e1d92
BZ
294 }
295
296 return 0;
297}
298
5c5670fa
YG
299static struct tc_action_ops act_sample_ops = {
300 .kind = "sample",
eddd2cf1 301 .id = TCA_ID_SAMPLE,
5c5670fa
YG
302 .owner = THIS_MODULE,
303 .act = tcf_sample_act,
58c04397 304 .stats_update = tcf_sample_stats_update,
5c5670fa
YG
305 .dump = tcf_sample_dump,
306 .init = tcf_sample_init,
307 .cleanup = tcf_sample_cleanup,
4a5da47d 308 .get_psample_group = tcf_sample_get_group,
c54e1d92 309 .offload_act_setup = tcf_sample_offload_act_setup,
5c5670fa
YG
310 .size = sizeof(struct tcf_sample),
311};
312
313static __net_init int sample_init_net(struct net *net)
314{
acd0a7ab 315 struct tc_action_net *tn = net_generic(net, act_sample_ops.net_id);
5c5670fa 316
981471bd 317 return tc_action_net_init(net, tn, &act_sample_ops);
5c5670fa
YG
318}
319
039af9c6 320static void __net_exit sample_exit_net(struct list_head *net_list)
5c5670fa 321{
acd0a7ab 322 tc_action_net_exit(net_list, act_sample_ops.net_id);
5c5670fa
YG
323}
324
325static struct pernet_operations sample_net_ops = {
326 .init = sample_init_net,
039af9c6 327 .exit_batch = sample_exit_net,
acd0a7ab 328 .id = &act_sample_ops.net_id,
5c5670fa
YG
329 .size = sizeof(struct tc_action_net),
330};
331
332static int __init sample_init_module(void)
333{
334 return tcf_register_action(&act_sample_ops, &sample_net_ops);
335}
336
337static void __exit sample_cleanup_module(void)
338{
339 tcf_unregister_action(&act_sample_ops, &sample_net_ops);
340}
341
342module_init(sample_init_module);
343module_exit(sample_cleanup_module);
344
f1fd20c3 345MODULE_AUTHOR("Yotam Gigi <yotam.gi@gmail.com>");
5c5670fa
YG
346MODULE_DESCRIPTION("Packet sampling action");
347MODULE_LICENSE("GPL v2");