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, bool rtnl_held,
  41                            struct netlink_ext_ack *extack)
  42 {
  43         struct tc_action_net *tn = net_generic(net, sample_net_id);
  44         struct nlattr *tb[TCA_SAMPLE_MAX + 1];
  45         struct psample_group *psample_group;
  46         u32 psample_group_num, rate, index;
  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(tb, TCA_SAMPLE_MAX, nla, sample_policy, NULL);
  55         if (ret < 0)
  56                 return ret;
  57         if (!tb[TCA_SAMPLE_PARMS] || !tb[TCA_SAMPLE_RATE] ||
  58             !tb[TCA_SAMPLE_PSAMPLE_GROUP])
  59                 return -EINVAL;
  60
  61         parm = nla_data(tb[TCA_SAMPLE_PARMS]);
  62         index = parm->index;
  63         err = tcf_idr_check_alloc(tn, &index, a, bind);
  64         if (err < 0)
  65                 return err;
  66         exists = err;
  67         if (exists && bind)
  68                 return 0;
  69
  70         if (!exists) {
  71                 ret = tcf_idr_create(tn, index, est, a,
  72                                      &act_sample_ops, bind, true);
  73                 if (ret) {
  74                         tcf_idr_cleanup(tn, index);
  75                         return ret;
  76                 }
  77                 ret = ACT_P_CREATED;
  78         } else if (!ovr) {
  79                 tcf_idr_release(*a, bind);
  80                 return -EEXIST;
  81         }
  82
  83         rate = nla_get_u32(tb[TCA_SAMPLE_RATE]);
  84         if (!rate) {
  85                 NL_SET_ERR_MSG(extack, "invalid sample rate");
  86                 tcf_idr_release(*a, bind);
  87                 return -EINVAL;
  88         }
  89         psample_group_num = nla_get_u32(tb[TCA_SAMPLE_PSAMPLE_GROUP]);
  90         psample_group = psample_group_get(net, psample_group_num);
  91         if (!psample_group) {
  92                 tcf_idr_release(*a, bind);
  93                 return -ENOMEM;
  94         }
  95
  96         s = to_sample(*a);
  97
  98         spin_lock_bh(&s->tcf_lock);
  99         s->tcf_action = parm->action;
 100         s->rate = rate;
 101         s->psample_group_num = psample_group_num;
 102         rcu_swap_protected(s->psample_group, psample_group,
 103                            lockdep_is_held(&s->tcf_lock));
 104
 105         if (tb[TCA_SAMPLE_TRUNC_SIZE]) {
 106                 s->truncate = true;
 107                 s->trunc_size = nla_get_u32(tb[TCA_SAMPLE_TRUNC_SIZE]);
 108         }
 109         spin_unlock_bh(&s->tcf_lock);
 110
 111         if (psample_group)
 112                 psample_group_put(psample_group);
 113         if (ret == ACT_P_CREATED)
 114                 tcf_idr_insert(tn, *a);
 115         return ret;
 116 }
 117
 118 static void tcf_sample_cleanup(struct tc_action *a)
 119 {
 120         struct tcf_sample *s = to_sample(a);
 121         struct psample_group *psample_group;
 122
 123         /* last reference to action, no need to lock */
 124         psample_group = rcu_dereference_protected(s->psample_group, 1);
 125         RCU_INIT_POINTER(s->psample_group, NULL);
 126         if (psample_group)
 127                 psample_group_put(psample_group);
 128 }
 129
 130 static bool tcf_sample_dev_ok_push(struct net_device *dev)
 131 {
 132         switch (dev->type) {
 133         case ARPHRD_TUNNEL:
 134         case ARPHRD_TUNNEL6:
 135         case ARPHRD_SIT:
 136         case ARPHRD_IPGRE:
 137         case ARPHRD_IP6GRE:
 138         case ARPHRD_VOID:
 139         case ARPHRD_NONE:
 140                 return false;
 141         default:
 142                 return true;
 143         }
 144 }
 145
 146 static int tcf_sample_act(struct sk_buff *skb, const struct tc_action *a,
 147                           struct tcf_result *res)
 148 {
 149         struct tcf_sample *s = to_sample(a);
 150         struct psample_group *psample_group;
 151         int retval;
 152         int size;
 153         int iif;
 154         int oif;
 155
 156         tcf_lastuse_update(&s->tcf_tm);
 157         bstats_cpu_update(this_cpu_ptr(s->common.cpu_bstats), skb);
 158         retval = READ_ONCE(s->tcf_action);
 159
 160         psample_group = rcu_dereference_bh(s->psample_group);
 161
 162         /* randomly sample packets according to rate */
 163         if (psample_group && (prandom_u32() % s->rate == 0)) {
 164                 if (!skb_at_tc_ingress(skb)) {
 165                         iif = skb->skb_iif;
 166                         oif = skb->dev->ifindex;
 167                 } else {
 168                         iif = skb->dev->ifindex;
 169                         oif = 0;
 170                 }
 171
 172                 /* on ingress, the mac header gets popped, so push it back */
 173                 if (skb_at_tc_ingress(skb) && tcf_sample_dev_ok_push(skb->dev))
 174                         skb_push(skb, skb->mac_len);
 175
 176                 size = s->truncate ? s->trunc_size : skb->len;
 177                 psample_sample_packet(psample_group, skb, size, iif, oif,
 178                                       s->rate);
 179
 180                 if (skb_at_tc_ingress(skb) && tcf_sample_dev_ok_push(skb->dev))
 181                         skb_pull(skb, skb->mac_len);
 182         }
 183
 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                 .refcnt     = refcount_read(&s->tcf_refcnt) - ref,
 195                 .bindcnt    = atomic_read(&s->tcf_bindcnt) - bind,
 196         };
 197         struct tcf_t t;
 198
 199         spin_lock_bh(&s->tcf_lock);
 200         opt.action = s->tcf_action;
 201         if (nla_put(skb, TCA_SAMPLE_PARMS, sizeof(opt), &opt))
 202                 goto nla_put_failure;
 203
 204         tcf_tm_dump(&t, &s->tcf_tm);
 205         if (nla_put_64bit(skb, TCA_SAMPLE_TM, sizeof(t), &t, TCA_SAMPLE_PAD))
 206                 goto nla_put_failure;
 207
 208         if (nla_put_u32(skb, TCA_SAMPLE_RATE, s->rate))
 209                 goto nla_put_failure;
 210
 211         if (s->truncate)
 212                 if (nla_put_u32(skb, TCA_SAMPLE_TRUNC_SIZE, s->trunc_size))
 213                         goto nla_put_failure;
 214
 215         if (nla_put_u32(skb, TCA_SAMPLE_PSAMPLE_GROUP, s->psample_group_num))
 216                 goto nla_put_failure;
 217         spin_unlock_bh(&s->tcf_lock);
 218
 219         return skb->len;
 220
 221 nla_put_failure:
 222         spin_unlock_bh(&s->tcf_lock);
 223         nlmsg_trim(skb, b);
 224         return -1;
 225 }
 226
 227 static int tcf_sample_walker(struct net *net, struct sk_buff *skb,
 228                              struct netlink_callback *cb, int type,
 229                              const struct tc_action_ops *ops,
 230                              struct netlink_ext_ack *extack)
 231 {
 232         struct tc_action_net *tn = net_generic(net, sample_net_id);
 233
 234         return tcf_generic_walker(tn, skb, cb, type, ops, extack);
 235 }
 236
 237 static int tcf_sample_search(struct net *net, struct tc_action **a, u32 index,
 238                              struct netlink_ext_ack *extack)
 239 {
 240         struct tc_action_net *tn = net_generic(net, sample_net_id);
 241
 242         return tcf_idr_search(tn, a, index);
 243 }
 244
 245 static struct tc_action_ops act_sample_ops = {
 246         .kind     = "sample",
 247         .type     = TCA_ACT_SAMPLE,
 248         .owner    = THIS_MODULE,
 249         .act      = tcf_sample_act,
 250         .dump     = tcf_sample_dump,
 251         .init     = tcf_sample_init,
 252         .cleanup  = tcf_sample_cleanup,
 253         .walk     = tcf_sample_walker,
 254         .lookup   = tcf_sample_search,
 255         .size     = sizeof(struct tcf_sample),
 256 };
 257
 258 static __net_init int sample_init_net(struct net *net)
 259 {
 260         struct tc_action_net *tn = net_generic(net, sample_net_id);
 261
 262         return tc_action_net_init(net, tn, &act_sample_ops);
 263 }
 264
 265 static void __net_exit sample_exit_net(struct list_head *net_list)
 266 {
 267         tc_action_net_exit(net_list, sample_net_id);
 268 }
 269
 270 static struct pernet_operations sample_net_ops = {
 271         .init = sample_init_net,
 272         .exit_batch = sample_exit_net,
 273         .id   = &sample_net_id,
 274         .size = sizeof(struct tc_action_net),
 275 };
 276
 277 static int __init sample_init_module(void)
 278 {
 279         return tcf_register_action(&act_sample_ops, &sample_net_ops);
 280 }
 281
 282 static void __exit sample_cleanup_module(void)
 283 {
 284         tcf_unregister_action(&act_sample_ops, &sample_net_ops);
 285 }
 286
 287 module_init(sample_init_module);
 288 module_exit(sample_cleanup_module);
 289
 290 MODULE_AUTHOR("Yotam Gigi <yotam.gi@gmail.com>");
 291 MODULE_DESCRIPTION("Packet sampling action");
 292 MODULE_LICENSE("GPL v2");