1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * net/sched/act_sample.c - Packet sampling tc action
4 * Copyright (c) 2017 Yotam Gigi <yotamg@mellanox.com>
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>
22#include <net/pkt_cls.h>
23#include <net/tc_wrapper.h>
24
25#include <linux/if_arp.h>
26
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,
37			   struct nlattr *est, struct tc_action **a,
38			   struct tcf_proto *tp,
39			   u32 flags, struct netlink_ext_ack *extack)
40{
41	struct tc_action_net *tn = net_generic(net, act_sample_ops.net_id);
42	bool bind = flags & TCA_ACT_FLAGS_BIND;
43	struct nlattr *tb[TCA_SAMPLE_MAX + 1];
44	struct psample_group *psample_group;
45	u32 psample_group_num, rate, index;
46	struct tcf_chain *goto_ch = NULL;
47	struct tc_sample *parm;
48	struct tcf_sample *s;
49	bool exists = false;
50	int ret, err;
51
52	if (!nla)
53		return -EINVAL;
54	ret = nla_parse_nested_deprecated(tb, TCA_SAMPLE_MAX, nla,
55					  sample_policy, NULL);
56	if (ret < 0)
57		return ret;
58
59	if (!tb[TCA_SAMPLE_PARMS])
60		return -EINVAL;
61
62	parm = nla_data(tb[TCA_SAMPLE_PARMS]);
63	index = parm->index;
64	err = tcf_idr_check_alloc(tn, &index, a, bind);
65	if (err < 0)
66		return err;
67	exists = err;
68	if (exists && bind)
69		return ACT_P_BOUND;
70
71	if (!exists) {
72		ret = tcf_idr_create(tn, index, est, a,
73				     &act_sample_ops, bind, true, flags);
74		if (ret) {
75			tcf_idr_cleanup(tn, index);
76			return ret;
77		}
78		ret = ACT_P_CREATED;
79	} else if (!(flags & TCA_ACT_FLAGS_REPLACE)) {
80		tcf_idr_release(*a, bind);
81		return -EEXIST;
82	}
83
84	if (!tb[TCA_SAMPLE_RATE] || !tb[TCA_SAMPLE_PSAMPLE_GROUP]) {
85		NL_SET_ERR_MSG(extack, "sample rate and group are required");
86		err = -EINVAL;
87		goto release_idr;
88	}
89
90	err = tcf_action_check_ctrlact(parm->action, tp, &goto_ch, extack);
91	if (err < 0)
92		goto release_idr;
93
94	rate = nla_get_u32(tb[TCA_SAMPLE_RATE]);
95	if (!rate) {
96		NL_SET_ERR_MSG(extack, "invalid sample rate");
97		err = -EINVAL;
98		goto put_chain;
99	}
100	psample_group_num = nla_get_u32(tb[TCA_SAMPLE_PSAMPLE_GROUP]);
101	psample_group = psample_group_get(net, psample_group_num);
102	if (!psample_group) {
103		err = -ENOMEM;
104		goto put_chain;
105	}
106
107	s = to_sample(*a);
108
109	spin_lock_bh(&s->tcf_lock);
110	goto_ch = tcf_action_set_ctrlact(*a, parm->action, goto_ch);
111	s->rate = rate;
112	s->psample_group_num = psample_group_num;
113	psample_group = rcu_replace_pointer(s->psample_group, psample_group,
114					    lockdep_is_held(&s->tcf_lock));
115
116	if (tb[TCA_SAMPLE_TRUNC_SIZE]) {
117		s->truncate = true;
118		s->trunc_size = nla_get_u32(tb[TCA_SAMPLE_TRUNC_SIZE]);
119	}
120	spin_unlock_bh(&s->tcf_lock);
121
122	if (psample_group)
123		psample_group_put(psample_group);
124	if (goto_ch)
125		tcf_chain_put_by_act(goto_ch);
126
127	return ret;
128put_chain:
129	if (goto_ch)
130		tcf_chain_put_by_act(goto_ch);
131release_idr:
132	tcf_idr_release(*a, bind);
133	return err;
134}
135
136static void tcf_sample_cleanup(struct tc_action *a)
137{
138	struct tcf_sample *s = to_sample(a);
139	struct psample_group *psample_group;
140
141	/* last reference to action, no need to lock */
142	psample_group = rcu_dereference_protected(s->psample_group, 1);
143	RCU_INIT_POINTER(s->psample_group, NULL);
144	if (psample_group)
145		psample_group_put(psample_group);
146}
147
148static bool tcf_sample_dev_ok_push(struct net_device *dev)
149{
150	switch (dev->type) {
151	case ARPHRD_TUNNEL:
152	case ARPHRD_TUNNEL6:
153	case ARPHRD_SIT:
154	case ARPHRD_IPGRE:
155	case ARPHRD_IP6GRE:
156	case ARPHRD_VOID:
157	case ARPHRD_NONE:
158		return false;
159	default:
160		return true;
161	}
162}
163
164TC_INDIRECT_SCOPE int tcf_sample_act(struct sk_buff *skb,
165				     const struct tc_action *a,
166				     struct tcf_result *res)
167{
168	struct tcf_sample *s = to_sample(a);
169	struct psample_group *psample_group;
170	struct psample_metadata md = {};
171	int retval;
172
173	tcf_lastuse_update(&s->tcf_tm);
174	bstats_update(this_cpu_ptr(s->common.cpu_bstats), skb);
175	retval = READ_ONCE(s->tcf_action);
176
177	psample_group = rcu_dereference_bh(s->psample_group);
178
179	/* randomly sample packets according to rate */
180	if (psample_group && (get_random_u32_below(s->rate) == 0)) {
181		if (!skb_at_tc_ingress(skb)) {
182			md.in_ifindex = skb->skb_iif;
183			md.out_ifindex = skb->dev->ifindex;
184		} else {
185			md.in_ifindex = skb->dev->ifindex;
186		}
187
188		/* on ingress, the mac header gets popped, so push it back */
189		if (skb_at_tc_ingress(skb) && tcf_sample_dev_ok_push(skb->dev))
190			skb_push(skb, skb->mac_len);
191
192		md.trunc_size = s->truncate ? s->trunc_size : skb->len;
193		psample_sample_packet(psample_group, skb, s->rate, &md);
194
195		if (skb_at_tc_ingress(skb) && tcf_sample_dev_ok_push(skb->dev))
196			skb_pull(skb, skb->mac_len);
197	}
198
199	return retval;
200}
201
202static void tcf_sample_stats_update(struct tc_action *a, u64 bytes, u64 packets,
203				    u64 drops, u64 lastuse, bool hw)
204{
205	struct tcf_sample *s = to_sample(a);
206	struct tcf_t *tm = &s->tcf_tm;
207
208	tcf_action_update_stats(a, bytes, packets, drops, hw);
209	tm->lastuse = max_t(u64, tm->lastuse, lastuse);
210}
211
212static int tcf_sample_dump(struct sk_buff *skb, struct tc_action *a,
213			   int bind, int ref)
214{
215	unsigned char *b = skb_tail_pointer(skb);
216	struct tcf_sample *s = to_sample(a);
217	struct tc_sample opt = {
218		.index      = s->tcf_index,
219		.refcnt     = refcount_read(&s->tcf_refcnt) - ref,
220		.bindcnt    = atomic_read(&s->tcf_bindcnt) - bind,
221	};
222	struct tcf_t t;
223
224	spin_lock_bh(&s->tcf_lock);
225	opt.action = s->tcf_action;
226	if (nla_put(skb, TCA_SAMPLE_PARMS, sizeof(opt), &opt))
227		goto nla_put_failure;
228
229	tcf_tm_dump(&t, &s->tcf_tm);
230	if (nla_put_64bit(skb, TCA_SAMPLE_TM, sizeof(t), &t, TCA_SAMPLE_PAD))
231		goto nla_put_failure;
232
233	if (nla_put_u32(skb, TCA_SAMPLE_RATE, s->rate))
234		goto nla_put_failure;
235
236	if (s->truncate)
237		if (nla_put_u32(skb, TCA_SAMPLE_TRUNC_SIZE, s->trunc_size))
238			goto nla_put_failure;
239
240	if (nla_put_u32(skb, TCA_SAMPLE_PSAMPLE_GROUP, s->psample_group_num))
241		goto nla_put_failure;
242	spin_unlock_bh(&s->tcf_lock);
243
244	return skb->len;
245
246nla_put_failure:
247	spin_unlock_bh(&s->tcf_lock);
248	nlmsg_trim(skb, b);
249	return -1;
250}
251
252static void tcf_psample_group_put(void *priv)
253{
254	struct psample_group *group = priv;
255
256	psample_group_put(group);
257}
258
259static struct psample_group *
260tcf_sample_get_group(const struct tc_action *a,
261		     tc_action_priv_destructor *destructor)
262{
263	struct tcf_sample *s = to_sample(a);
264	struct psample_group *group;
265
266	group = rcu_dereference_protected(s->psample_group,
267					  lockdep_is_held(&s->tcf_lock));
268	if (group) {
269		psample_group_take(group);
270		*destructor = tcf_psample_group_put;
271	}
272
273	return group;
274}
275
276static void tcf_offload_sample_get_group(struct flow_action_entry *entry,
277					 const struct tc_action *act)
278{
279	entry->sample.psample_group =
280		act->ops->get_psample_group(act, &entry->destructor);
281	entry->destructor_priv = entry->sample.psample_group;
282}
283
284static int tcf_sample_offload_act_setup(struct tc_action *act, void *entry_data,
285					u32 *index_inc, bool bind,
286					struct netlink_ext_ack *extack)
287{
288	if (bind) {
289		struct flow_action_entry *entry = entry_data;
290
291		entry->id = FLOW_ACTION_SAMPLE;
292		entry->sample.trunc_size = tcf_sample_trunc_size(act);
293		entry->sample.truncate = tcf_sample_truncate(act);
294		entry->sample.rate = tcf_sample_rate(act);
295		tcf_offload_sample_get_group(entry, act);
296		*index_inc = 1;
297	} else {
298		struct flow_offload_action *fl_action = entry_data;
299
300		fl_action->id = FLOW_ACTION_SAMPLE;
301	}
302
303	return 0;
304}
305
306static struct tc_action_ops act_sample_ops = {
307	.kind	  = "sample",
308	.id	  = TCA_ID_SAMPLE,
309	.owner	  = THIS_MODULE,
310	.act	  = tcf_sample_act,
311	.stats_update = tcf_sample_stats_update,
312	.dump	  = tcf_sample_dump,
313	.init	  = tcf_sample_init,
314	.cleanup  = tcf_sample_cleanup,
315	.get_psample_group = tcf_sample_get_group,
316	.offload_act_setup    = tcf_sample_offload_act_setup,
317	.size	  = sizeof(struct tcf_sample),
318};
319MODULE_ALIAS_NET_ACT("sample");
320
321static __net_init int sample_init_net(struct net *net)
322{
323	struct tc_action_net *tn = net_generic(net, act_sample_ops.net_id);
324
325	return tc_action_net_init(net, tn, &act_sample_ops);
326}
327
328static void __net_exit sample_exit_net(struct list_head *net_list)
329{
330	tc_action_net_exit(net_list, act_sample_ops.net_id);
331}
332
333static struct pernet_operations sample_net_ops = {
334	.init = sample_init_net,
335	.exit_batch = sample_exit_net,
336	.id   = &act_sample_ops.net_id,
337	.size = sizeof(struct tc_action_net),
338};
339
340static int __init sample_init_module(void)
341{
342	return tcf_register_action(&act_sample_ops, &sample_net_ops);
343}
344
345static void __exit sample_cleanup_module(void)
346{
347	tcf_unregister_action(&act_sample_ops, &sample_net_ops);
348}
349
350module_init(sample_init_module);
351module_exit(sample_cleanup_module);
352
353MODULE_AUTHOR("Yotam Gigi <yotam.gi@gmail.com>");
354MODULE_DESCRIPTION("Packet sampling action");
355MODULE_LICENSE("GPL v2");
356