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 #include <net/pkt_cls.h>
  26
  27 #include <linux/if_arp.h>
  28
  29 static unsigned int sample_net_id;
  30 static struct tc_action_ops act_sample_ops;
  31
  32 static const struct nla_policy sample_policy[TCA_SAMPLE_MAX + 1] = {
  33         [TCA_SAMPLE_PARMS]              = { .len = sizeof(struct tc_sample) },
  34         [TCA_SAMPLE_RATE]               = { .type = NLA_U32 },
  35         [TCA_SAMPLE_TRUNC_SIZE]         = { .type = NLA_U32 },
  36         [TCA_SAMPLE_PSAMPLE_GROUP]      = { .type = NLA_U32 },
  37 };
  38
  39 static int tcf_sample_init(struct net *net, struct nlattr *nla,
  40                            struct nlattr *est, struct tc_action **a, int ovr,
  41                            int bind, bool rtnl_held, struct tcf_proto *tp,
  42                            struct netlink_ext_ack *extack)
  43 {
  44         struct tc_action_net *tn = net_generic(net, sample_net_id);
  45         struct nlattr *tb[TCA_SAMPLE_MAX + 1];
  46         struct psample_group *psample_group;
  47         struct tcf_chain *goto_ch = NULL;
  48         u32 psample_group_num, rate;
  49         struct tc_sample *parm;
  50         struct tcf_sample *s;
  51         bool exists = false;
  52         int ret, err;
  53
  54         if (!nla)
  55                 return -EINVAL;
  56         ret = nla_parse_nested(tb, TCA_SAMPLE_MAX, nla, sample_policy, NULL);
  57         if (ret < 0)
  58                 return ret;
  59         if (!tb[TCA_SAMPLE_PARMS] || !tb[TCA_SAMPLE_RATE] ||
  60             !tb[TCA_SAMPLE_PSAMPLE_GROUP])
  61                 return -EINVAL;
  62
  63         parm = nla_data(tb[TCA_SAMPLE_PARMS]);
  64
  65         err = tcf_idr_check_alloc(tn, &parm->index, a, bind);
  66         if (err < 0)
  67                 return err;
  68         exists = err;
  69         if (exists && bind)
  70                 return 0;
  71
  72         if (!exists) {
  73                 ret = tcf_idr_create(tn, parm->index, est, a,
  74                                      &act_sample_ops, bind, true);
  75                 if (ret) {
  76                         tcf_idr_cleanup(tn, parm->index);
  77                         return ret;
  78                 }
  79                 ret = ACT_P_CREATED;
  80         } else if (!ovr) {
  81                 tcf_idr_release(*a, bind);
  82                 return -EEXIST;
  83         }
  84         err = tcf_action_check_ctrlact(parm->action, tp, &goto_ch, extack);
  85         if (err < 0)
  86                 goto release_idr;
  87
  88         rate = nla_get_u32(tb[TCA_SAMPLE_RATE]);
  89         if (!rate) {
  90                 NL_SET_ERR_MSG(extack, "invalid sample rate");
  91                 err = -EINVAL;
  92                 goto put_chain;
  93         }
  94         psample_group_num = nla_get_u32(tb[TCA_SAMPLE_PSAMPLE_GROUP]);
  95         psample_group = psample_group_get(net, psample_group_num);
  96         if (!psample_group) {
  97                 err = -ENOMEM;
  98                 goto put_chain;
  99         }
 100
 101         s = to_sample(*a);
 102
 103         spin_lock_bh(&s->tcf_lock);
 104         goto_ch = tcf_action_set_ctrlact(*a, parm->action, goto_ch);
 105         s->rate = rate;
 106         s->psample_group_num = psample_group_num;
 107         RCU_INIT_POINTER(s->psample_group, psample_group);
 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         }
 113         spin_unlock_bh(&s->tcf_lock);
 114         if (goto_ch)
 115                 tcf_chain_put_by_act(goto_ch);
 116
 117         if (ret == ACT_P_CREATED)
 118                 tcf_idr_insert(tn, *a);
 119         return ret;
 120 put_chain:
 121         if (goto_ch)
 122                 tcf_chain_put_by_act(goto_ch);
 123 release_idr:
 124         tcf_idr_release(*a, bind);
 125         return err;
 126 }
 127
 128 static void tcf_sample_cleanup(struct tc_action *a)
 129 {
 130         struct tcf_sample *s = to_sample(a);
 131         struct psample_group *psample_group;
 132
 133         /* last reference to action, no need to lock */
 134         psample_group = rcu_dereference_protected(s->psample_group, 1);
 135         RCU_INIT_POINTER(s->psample_group, NULL);
 136         if (psample_group)
 137                 psample_group_put(psample_group);
 138 }
 139
 140 static bool tcf_sample_dev_ok_push(struct net_device *dev)
 141 {
 142         switch (dev->type) {
 143         case ARPHRD_TUNNEL:
 144         case ARPHRD_TUNNEL6:
 145         case ARPHRD_SIT:
 146         case ARPHRD_IPGRE:
 147         case ARPHRD_VOID:
 148         case ARPHRD_NONE:
 149                 return false;
 150         default:
 151                 return true;
 152         }
 153 }
 154
 155 static int tcf_sample_act(struct sk_buff *skb, const struct tc_action *a,
 156                           struct tcf_result *res)
 157 {
 158         struct tcf_sample *s = to_sample(a);
 159         struct psample_group *psample_group;
 160         int retval;
 161         int size;
 162         int iif;
 163         int oif;
 164
 165         tcf_lastuse_update(&s->tcf_tm);
 166         bstats_cpu_update(this_cpu_ptr(s->common.cpu_bstats), skb);
 167         retval = READ_ONCE(s->tcf_action);
 168
 169         psample_group = rcu_dereference_bh(s->psample_group);
 170
 171         /* randomly sample packets according to rate */
 172         if (psample_group && (prandom_u32() % s->rate == 0)) {
 173                 if (!skb_at_tc_ingress(skb)) {
 174                         iif = skb->skb_iif;
 175                         oif = skb->dev->ifindex;
 176                 } else {
 177                         iif = skb->dev->ifindex;
 178                         oif = 0;
 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
 185                 size = s->truncate ? s->trunc_size : skb->len;
 186                 psample_sample_packet(psample_group, skb, size, iif, oif,
 187                                       s->rate);
 188
 189                 if (skb_at_tc_ingress(skb) && tcf_sample_dev_ok_push(skb->dev))
 190                         skb_pull(skb, skb->mac_len);
 191         }
 192
 193         return retval;
 194 }
 195
 196 static int tcf_sample_dump(struct sk_buff *skb, struct tc_action *a,
 197                            int bind, int ref)
 198 {
 199         unsigned char *b = skb_tail_pointer(skb);
 200         struct tcf_sample *s = to_sample(a);
 201         struct tc_sample opt = {
 202                 .index      = s->tcf_index,
 203                 .refcnt     = refcount_read(&s->tcf_refcnt) - ref,
 204                 .bindcnt    = atomic_read(&s->tcf_bindcnt) - bind,
 205         };
 206         struct tcf_t t;
 207
 208         spin_lock_bh(&s->tcf_lock);
 209         opt.action = s->tcf_action;
 210         if (nla_put(skb, TCA_SAMPLE_PARMS, sizeof(opt), &opt))
 211                 goto nla_put_failure;
 212
 213         tcf_tm_dump(&t, &s->tcf_tm);
 214         if (nla_put_64bit(skb, TCA_SAMPLE_TM, sizeof(t), &t, TCA_SAMPLE_PAD))
 215                 goto nla_put_failure;
 216
 217         if (nla_put_u32(skb, TCA_SAMPLE_RATE, s->rate))
 218                 goto nla_put_failure;
 219
 220         if (s->truncate)
 221                 if (nla_put_u32(skb, TCA_SAMPLE_TRUNC_SIZE, s->trunc_size))
 222                         goto nla_put_failure;
 223
 224         if (nla_put_u32(skb, TCA_SAMPLE_PSAMPLE_GROUP, s->psample_group_num))
 225                 goto nla_put_failure;
 226         spin_unlock_bh(&s->tcf_lock);
 227
 228         return skb->len;
 229
 230 nla_put_failure:
 231         spin_unlock_bh(&s->tcf_lock);
 232         nlmsg_trim(skb, b);
 233         return -1;
 234 }
 235
 236 static int tcf_sample_walker(struct net *net, struct sk_buff *skb,
 237                              struct netlink_callback *cb, int type,
 238                              const struct tc_action_ops *ops,
 239                              struct netlink_ext_ack *extack)
 240 {
 241         struct tc_action_net *tn = net_generic(net, sample_net_id);
 242
 243         return tcf_generic_walker(tn, skb, cb, type, ops, extack);
 244 }
 245
 246 static int tcf_sample_search(struct net *net, struct tc_action **a, u32 index)
 247 {
 248         struct tc_action_net *tn = net_generic(net, sample_net_id);
 249
 250         return tcf_idr_search(tn, a, index);
 251 }
 252
 253 static struct tc_action_ops act_sample_ops = {
 254         .kind     = "sample",
 255         .id       = TCA_ID_SAMPLE,
 256         .owner    = THIS_MODULE,
 257         .act      = tcf_sample_act,
 258         .dump     = tcf_sample_dump,
 259         .init     = tcf_sample_init,
 260         .cleanup  = tcf_sample_cleanup,
 261         .walk     = tcf_sample_walker,
 262         .lookup   = tcf_sample_search,
 263         .size     = sizeof(struct tcf_sample),
 264 };
 265
 266 static __net_init int sample_init_net(struct net *net)
 267 {
 268         struct tc_action_net *tn = net_generic(net, sample_net_id);
 269
 270         return tc_action_net_init(tn, &act_sample_ops);
 271 }
 272
 273 static void __net_exit sample_exit_net(struct list_head *net_list)
 274 {
 275         tc_action_net_exit(net_list, sample_net_id);
 276 }
 277
 278 static struct pernet_operations sample_net_ops = {
 279         .init = sample_init_net,
 280         .exit_batch = sample_exit_net,
 281         .id   = &sample_net_id,
 282         .size = sizeof(struct tc_action_net),
 283 };
 284
 285 static int __init sample_init_module(void)
 286 {
 287         return tcf_register_action(&act_sample_ops, &sample_net_ops);
 288 }
 289
 290 static void __exit sample_cleanup_module(void)
 291 {
 292         tcf_unregister_action(&act_sample_ops, &sample_net_ops);
 293 }
 294
 295 module_init(sample_init_module);
 296 module_exit(sample_cleanup_module);
 297
 298 MODULE_AUTHOR("Yotam Gigi <yotam.gi@gmail.com>");
 299 MODULE_DESCRIPTION("Packet sampling action");
 300 MODULE_LICENSE("GPL v2");