net/sched/act_sample.c

   1 /*
   2  * net/sched/act_sample.c - Packet sampling tc action
   3  * Copyright (c) 2017 Yotam Gigi <yotamg@mellanox.com>
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License version 2 as
   7  * published by the Free Software Foundation.
   8  */
   9
  10 #include <linux/types.h>
  11 #include <linux/kernel.h>
  12 #include <linux/string.h>
  13 #include <linux/errno.h>
  14 #include <linux/skbuff.h>
  15 #include <linux/rtnetlink.h>
  16 #include <linux/module.h>
  17 #include <linux/init.h>
  18 #include <linux/gfp.h>
  19 #include <net/net_namespace.h>
  20 #include <net/netlink.h>
  21 #include <net/pkt_sched.h>
  22 #include <linux/tc_act/tc_sample.h>
  23 #include <net/tc_act/tc_sample.h>
  24 #include <net/psample.h>
  25
  26 #include <linux/if_arp.h>
  27
  28 static unsigned int sample_net_id;
  29 static struct tc_action_ops act_sample_ops;
  30
  31 static const struct nla_policy sample_policy[TCA_SAMPLE_MAX + 1] = {
  32         [TCA_SAMPLE_PARMS]              = { .len = sizeof(struct tc_sample) },
  33         [TCA_SAMPLE_RATE]               = { .type = NLA_U32 },
  34         [TCA_SAMPLE_TRUNC_SIZE]         = { .type = NLA_U32 },
  35         [TCA_SAMPLE_PSAMPLE_GROUP]      = { .type = NLA_U32 },
  36 };
  37
  38 static int tcf_sample_init(struct net *net, struct nlattr *nla,
  39                            struct nlattr *est, struct tc_action **a, int ovr,
  40                            int bind)
  41 {
  42         struct tc_action_net *tn = net_generic(net, sample_net_id);
  43         struct nlattr *tb[TCA_SAMPLE_MAX + 1];
  44         struct psample_group *psample_group;
  45         struct tc_sample *parm;
  46         struct tcf_sample *s;
  47         bool exists = false;
  48         u32 rate;
  49         int ret;
  50
  51         if (!nla)
  52                 return -EINVAL;
  53         ret = nla_parse_nested(tb, TCA_SAMPLE_MAX, nla, sample_policy, NULL);
  54         if (ret < 0)
  55                 return ret;
  56         if (!tb[TCA_SAMPLE_PARMS] || !tb[TCA_SAMPLE_RATE] ||
  57             !tb[TCA_SAMPLE_PSAMPLE_GROUP])
  58                 return -EINVAL;
  59
  60         parm = nla_data(tb[TCA_SAMPLE_PARMS]);
  61
  62         exists = tcf_idr_check(tn, parm->index, a, bind);
  63         if (exists && bind)
  64                 return 0;
  65
  66         if (!exists) {
  67                 ret = tcf_idr_create(tn, parm->index, est, a,
  68                                      &act_sample_ops, bind, true);
  69                 if (ret)
  70                         return ret;
  71                 ret = ACT_P_CREATED;
  72         } else {
  73                 tcf_idr_release(*a, bind);
  74                 if (!ovr)
  75                         return -EEXIST;
  76         }
  77
  78         rate = nla_get_u32(tb[TCA_SAMPLE_RATE]);
  79         if (!rate) {
  80                 tcf_idr_release(*a, bind);
  81                 return -EINVAL;
  82         }
  83
  84         s = to_sample(*a);
  85         s->tcf_action = parm->action;
  86         s->rate = nla_get_u32(tb[TCA_SAMPLE_RATE]);
  87         s->rate = rate;
  88         s->psample_group_num = nla_get_u32(tb[TCA_SAMPLE_PSAMPLE_GROUP]);
  89         psample_group = psample_group_get(net, s->psample_group_num);
  90         if (!psample_group) {
  91                 if (ret == ACT_P_CREATED)
  92                         tcf_idr_release(*a, bind);
  93                 return -ENOMEM;
  94         }
  95         rcu_swap_protected(s->psample_group, psample_group,
  96                            lockdep_is_held(&s->tcf_lock));
  97
  98         if (tb[TCA_SAMPLE_TRUNC_SIZE]) {
  99                 s->truncate = true;
 100                 s->trunc_size = nla_get_u32(tb[TCA_SAMPLE_TRUNC_SIZE]);
 101         }
 102
 103         if (psample_group)
 104                 psample_group_put(psample_group);
 105         if (ret == ACT_P_CREATED)
 106                 tcf_idr_insert(tn, *a);
 107         return ret;
 108 }
 109
 110 static void tcf_sample_cleanup_rcu(struct rcu_head *rcu)
 111 {
 112         struct tcf_sample *s = container_of(rcu, struct tcf_sample, rcu);
 113         struct psample_group *psample_group;
 114
 115         psample_group = rcu_dereference_protected(s->psample_group, 1);
 116         RCU_INIT_POINTER(s->psample_group, NULL);
 117         if (psample_group)
 118                 psample_group_put(psample_group);
 119 }
 120
 121 static void tcf_sample_cleanup(struct tc_action *a, int bind)
 122 {
 123         struct tcf_sample *s = to_sample(a);
 124
 125         call_rcu(&s->rcu, tcf_sample_cleanup_rcu);
 126 }
 127
 128 static bool tcf_sample_dev_ok_push(struct net_device *dev)
 129 {
 130         switch (dev->type) {
 131         case ARPHRD_TUNNEL:
 132         case ARPHRD_TUNNEL6:
 133         case ARPHRD_SIT:
 134         case ARPHRD_IPGRE:
 135         case ARPHRD_IP6GRE:
 136         case ARPHRD_VOID:
 137         case ARPHRD_NONE:
 138                 return false;
 139         default:
 140                 return true;
 141         }
 142 }
 143
 144 static int tcf_sample_act(struct sk_buff *skb, const struct tc_action *a,
 145                           struct tcf_result *res)
 146 {
 147         struct tcf_sample *s = to_sample(a);
 148         struct psample_group *psample_group;
 149         int retval;
 150         int size;
 151         int iif;
 152         int oif;
 153
 154         tcf_lastuse_update(&s->tcf_tm);
 155         bstats_cpu_update(this_cpu_ptr(s->common.cpu_bstats), skb);
 156         retval = READ_ONCE(s->tcf_action);
 157
 158         rcu_read_lock();
 159         psample_group = rcu_dereference(s->psample_group);
 160
 161         /* randomly sample packets according to rate */
 162         if (psample_group && (prandom_u32() % s->rate == 0)) {
 163                 if (!skb_at_tc_ingress(skb)) {
 164                         iif = skb->skb_iif;
 165                         oif = skb->dev->ifindex;
 166                 } else {
 167                         iif = skb->dev->ifindex;
 168                         oif = 0;
 169                 }
 170
 171                 /* on ingress, the mac header gets popped, so push it back */
 172                 if (skb_at_tc_ingress(skb) && tcf_sample_dev_ok_push(skb->dev))
 173                         skb_push(skb, skb->mac_len);
 174
 175                 size = s->truncate ? s->trunc_size : skb->len;
 176                 psample_sample_packet(psample_group, skb, size, iif, oif,
 177                                       s->rate);
 178
 179                 if (skb_at_tc_ingress(skb) && tcf_sample_dev_ok_push(skb->dev))
 180                         skb_pull(skb, skb->mac_len);
 181         }
 182
 183         rcu_read_unlock();
 184         return retval;
 185 }
 186
 187 static int tcf_sample_dump(struct sk_buff *skb, struct tc_action *a,
 188                            int bind, int ref)
 189 {
 190         unsigned char *b = skb_tail_pointer(skb);
 191         struct tcf_sample *s = to_sample(a);
 192         struct tc_sample opt = {
 193                 .index      = s->tcf_index,
 194                 .action     = s->tcf_action,
 195                 .refcnt     = s->tcf_refcnt - ref,
 196                 .bindcnt    = s->tcf_bindcnt - bind,
 197         };
 198         struct tcf_t t;
 199
 200         if (nla_put(skb, TCA_SAMPLE_PARMS, sizeof(opt), &opt))
 201                 goto nla_put_failure;
 202
 203         tcf_tm_dump(&t, &s->tcf_tm);
 204         if (nla_put_64bit(skb, TCA_SAMPLE_TM, sizeof(t), &t, TCA_SAMPLE_PAD))
 205                 goto nla_put_failure;
 206
 207         if (nla_put_u32(skb, TCA_SAMPLE_RATE, s->rate))
 208                 goto nla_put_failure;
 209
 210         if (s->truncate)
 211                 if (nla_put_u32(skb, TCA_SAMPLE_TRUNC_SIZE, s->trunc_size))
 212                         goto nla_put_failure;
 213
 214         if (nla_put_u32(skb, TCA_SAMPLE_PSAMPLE_GROUP, s->psample_group_num))
 215                 goto nla_put_failure;
 216         return skb->len;
 217
 218 nla_put_failure:
 219         nlmsg_trim(skb, b);
 220         return -1;
 221 }
 222
 223 static int tcf_sample_walker(struct net *net, struct sk_buff *skb,
 224                              struct netlink_callback *cb, int type,
 225                              const struct tc_action_ops *ops)
 226 {
 227         struct tc_action_net *tn = net_generic(net, sample_net_id);
 228
 229         return tcf_generic_walker(tn, skb, cb, type, ops);
 230 }
 231
 232 static int tcf_sample_search(struct net *net, struct tc_action **a, u32 index)
 233 {
 234         struct tc_action_net *tn = net_generic(net, sample_net_id);
 235
 236         return tcf_idr_search(tn, a, index);
 237 }
 238
 239 static struct tc_action_ops act_sample_ops = {
 240         .kind     = "sample",
 241         .type     = TCA_ACT_SAMPLE,
 242         .owner    = THIS_MODULE,
 243         .act      = tcf_sample_act,
 244         .dump     = tcf_sample_dump,
 245         .init     = tcf_sample_init,
 246         .cleanup  = tcf_sample_cleanup,
 247         .walk     = tcf_sample_walker,
 248         .lookup   = tcf_sample_search,
 249         .size     = sizeof(struct tcf_sample),
 250 };
 251
 252 static __net_init int sample_init_net(struct net *net)
 253 {
 254         struct tc_action_net *tn = net_generic(net, sample_net_id);
 255
 256         return tc_action_net_init(net, tn, &act_sample_ops);
 257 }
 258
 259 static void __net_exit sample_exit_net(struct net *net)
 260 {
 261         struct tc_action_net *tn = net_generic(net, sample_net_id);
 262
 263         tc_action_net_exit(tn);
 264 }
 265
 266 static struct pernet_operations sample_net_ops = {
 267         .init = sample_init_net,
 268         .exit = sample_exit_net,
 269         .id   = &sample_net_id,
 270         .size = sizeof(struct tc_action_net),
 271 };
 272
 273 static int __init sample_init_module(void)
 274 {
 275         return tcf_register_action(&act_sample_ops, &sample_net_ops);
 276 }
 277
 278 static void __exit sample_cleanup_module(void)
 279 {
 280         rcu_barrier();
 281         tcf_unregister_action(&act_sample_ops, &sample_net_ops);
 282 }
 283
 284 module_init(sample_init_module);
 285 module_exit(sample_cleanup_module);
 286
 287 MODULE_AUTHOR("Yotam Gigi <yotam.gi@gmail.com>");
 288 MODULE_DESCRIPTION("Packet sampling action");
 289 MODULE_LICENSE("GPL v2");