net/psample/psample.c

   1 /*
   2  * net/psample/psample.c - Netlink channel for packet sampling
   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/skbuff.h>
  13 #include <linux/module.h>
  14 #include <net/net_namespace.h>
  15 #include <net/sock.h>
  16 #include <net/netlink.h>
  17 #include <net/genetlink.h>
  18 #include <net/psample.h>
  19 #include <linux/spinlock.h>
  20
  21 #define PSAMPLE_MAX_PACKET_SIZE 0xffff
  22
  23 static LIST_HEAD(psample_groups_list);
  24 static DEFINE_SPINLOCK(psample_groups_lock);
  25
  26 /* multicast groups */
  27 enum psample_nl_multicast_groups {
  28         PSAMPLE_NL_MCGRP_CONFIG,
  29         PSAMPLE_NL_MCGRP_SAMPLE,
  30 };
  31
  32 static const struct genl_multicast_group psample_nl_mcgrps[] = {
  33         [PSAMPLE_NL_MCGRP_CONFIG] = { .name = PSAMPLE_NL_MCGRP_CONFIG_NAME },
  34         [PSAMPLE_NL_MCGRP_SAMPLE] = { .name = PSAMPLE_NL_MCGRP_SAMPLE_NAME },
  35 };
  36
  37 static struct genl_family psample_nl_family __ro_after_init;
  38
  39 static int psample_group_nl_fill(struct sk_buff *msg,
  40                                  struct psample_group *group,
  41                                  enum psample_command cmd, u32 portid, u32 seq,
  42                                  int flags)
  43 {
  44         void *hdr;
  45         int ret;
  46
  47         hdr = genlmsg_put(msg, portid, seq, &psample_nl_family, flags, cmd);
  48         if (!hdr)
  49                 return -EMSGSIZE;
  50
  51         ret = nla_put_u32(msg, PSAMPLE_ATTR_SAMPLE_GROUP, group->group_num);
  52         if (ret < 0)
  53                 goto error;
  54
  55         ret = nla_put_u32(msg, PSAMPLE_ATTR_GROUP_REFCOUNT, group->refcount);
  56         if (ret < 0)
  57                 goto error;
  58
  59         ret = nla_put_u32(msg, PSAMPLE_ATTR_GROUP_SEQ, group->seq);
  60         if (ret < 0)
  61                 goto error;
  62
  63         genlmsg_end(msg, hdr);
  64         return 0;
  65
  66 error:
  67         genlmsg_cancel(msg, hdr);
  68         return -EMSGSIZE;
  69 }
  70
  71 static int psample_nl_cmd_get_group_dumpit(struct sk_buff *msg,
  72                                            struct netlink_callback *cb)
  73 {
  74         struct psample_group *group;
  75         int start = cb->args[0];
  76         int idx = 0;
  77         int err;
  78
  79         spin_lock(&psample_groups_lock);
  80         list_for_each_entry(group, &psample_groups_list, list) {
  81                 if (!net_eq(group->net, sock_net(msg->sk)))
  82                         continue;
  83                 if (idx < start) {
  84                         idx++;
  85                         continue;
  86                 }
  87                 err = psample_group_nl_fill(msg, group, PSAMPLE_CMD_NEW_GROUP,
  88                                             NETLINK_CB(cb->skb).portid,
  89                                             cb->nlh->nlmsg_seq, NLM_F_MULTI);
  90                 if (err)
  91                         break;
  92                 idx++;
  93         }
  94
  95         spin_unlock(&psample_groups_lock);
  96         cb->args[0] = idx;
  97         return msg->len;
  98 }
  99
 100 static const struct genl_ops psample_nl_ops[] = {
 101         {
 102                 .cmd = PSAMPLE_CMD_GET_GROUP,
 103                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 104                 .dumpit = psample_nl_cmd_get_group_dumpit,
 105                 /* can be retrieved by unprivileged users */
 106         }
 107 };
 108
 109 static struct genl_family psample_nl_family __ro_after_init = {
 110         .name           = PSAMPLE_GENL_NAME,
 111         .version        = PSAMPLE_GENL_VERSION,
 112         .maxattr        = PSAMPLE_ATTR_MAX,
 113         .netnsok        = true,
 114         .module         = THIS_MODULE,
 115         .mcgrps         = psample_nl_mcgrps,
 116         .ops            = psample_nl_ops,
 117         .n_ops          = ARRAY_SIZE(psample_nl_ops),
 118         .n_mcgrps       = ARRAY_SIZE(psample_nl_mcgrps),
 119 };
 120
 121 static void psample_group_notify(struct psample_group *group,
 122                                  enum psample_command cmd)
 123 {
 124         struct sk_buff *msg;
 125         int err;
 126
 127         msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
 128         if (!msg)
 129                 return;
 130
 131         err = psample_group_nl_fill(msg, group, cmd, 0, 0, NLM_F_MULTI);
 132         if (!err)
 133                 genlmsg_multicast_netns(&psample_nl_family, group->net, msg, 0,
 134                                         PSAMPLE_NL_MCGRP_CONFIG, GFP_ATOMIC);
 135         else
 136                 nlmsg_free(msg);
 137 }
 138
 139 static struct psample_group *psample_group_create(struct net *net,
 140                                                   u32 group_num)
 141 {
 142         struct psample_group *group;
 143
 144         group = kzalloc(sizeof(*group), GFP_ATOMIC);
 145         if (!group)
 146                 return NULL;
 147
 148         group->net = net;
 149         group->group_num = group_num;
 150         list_add_tail(&group->list, &psample_groups_list);
 151
 152         psample_group_notify(group, PSAMPLE_CMD_NEW_GROUP);
 153         return group;
 154 }
 155
 156 static void psample_group_destroy(struct psample_group *group)
 157 {
 158         psample_group_notify(group, PSAMPLE_CMD_DEL_GROUP);
 159         list_del(&group->list);
 160         kfree(group);
 161 }
 162
 163 static struct psample_group *
 164 psample_group_lookup(struct net *net, u32 group_num)
 165 {
 166         struct psample_group *group;
 167
 168         list_for_each_entry(group, &psample_groups_list, list)
 169                 if ((group->group_num == group_num) && (group->net == net))
 170                         return group;
 171         return NULL;
 172 }
 173
 174 struct psample_group *psample_group_get(struct net *net, u32 group_num)
 175 {
 176         struct psample_group *group;
 177
 178         spin_lock(&psample_groups_lock);
 179
 180         group = psample_group_lookup(net, group_num);
 181         if (!group) {
 182                 group = psample_group_create(net, group_num);
 183                 if (!group)
 184                         goto out;
 185         }
 186         group->refcount++;
 187
 188 out:
 189         spin_unlock(&psample_groups_lock);
 190         return group;
 191 }
 192 EXPORT_SYMBOL_GPL(psample_group_get);
 193
 194 void psample_group_put(struct psample_group *group)
 195 {
 196         spin_lock(&psample_groups_lock);
 197
 198         if (--group->refcount == 0)
 199                 psample_group_destroy(group);
 200
 201         spin_unlock(&psample_groups_lock);
 202 }
 203 EXPORT_SYMBOL_GPL(psample_group_put);
 204
 205 void psample_sample_packet(struct psample_group *group, struct sk_buff *skb,
 206                            u32 trunc_size, int in_ifindex, int out_ifindex,
 207                            u32 sample_rate)
 208 {
 209         struct sk_buff *nl_skb;
 210         int data_len;
 211         int meta_len;
 212         void *data;
 213         int ret;
 214
 215         meta_len = (in_ifindex ? nla_total_size(sizeof(u16)) : 0) +
 216                    (out_ifindex ? nla_total_size(sizeof(u16)) : 0) +
 217                    nla_total_size(sizeof(u32)) +        /* sample_rate */
 218                    nla_total_size(sizeof(u32)) +        /* orig_size */
 219                    nla_total_size(sizeof(u32)) +        /* group_num */
 220                    nla_total_size(sizeof(u32));         /* seq */
 221
 222         data_len = min(skb->len, trunc_size);
 223         if (meta_len + nla_total_size(data_len) > PSAMPLE_MAX_PACKET_SIZE)
 224                 data_len = PSAMPLE_MAX_PACKET_SIZE - meta_len - NLA_HDRLEN
 225                             - NLA_ALIGNTO;
 226
 227         nl_skb = genlmsg_new(meta_len + data_len, GFP_ATOMIC);
 228         if (unlikely(!nl_skb))
 229                 return;
 230
 231         data = genlmsg_put(nl_skb, 0, 0, &psample_nl_family, 0,
 232                            PSAMPLE_CMD_SAMPLE);
 233         if (unlikely(!data))
 234                 goto error;
 235
 236         if (in_ifindex) {
 237                 ret = nla_put_u16(nl_skb, PSAMPLE_ATTR_IIFINDEX, in_ifindex);
 238                 if (unlikely(ret < 0))
 239                         goto error;
 240         }
 241
 242         if (out_ifindex) {
 243                 ret = nla_put_u16(nl_skb, PSAMPLE_ATTR_OIFINDEX, out_ifindex);
 244                 if (unlikely(ret < 0))
 245                         goto error;
 246         }
 247
 248         ret = nla_put_u32(nl_skb, PSAMPLE_ATTR_SAMPLE_RATE, sample_rate);
 249         if (unlikely(ret < 0))
 250                 goto error;
 251
 252         ret = nla_put_u32(nl_skb, PSAMPLE_ATTR_ORIGSIZE, skb->len);
 253         if (unlikely(ret < 0))
 254                 goto error;
 255
 256         ret = nla_put_u32(nl_skb, PSAMPLE_ATTR_SAMPLE_GROUP, group->group_num);
 257         if (unlikely(ret < 0))
 258                 goto error;
 259
 260         ret = nla_put_u32(nl_skb, PSAMPLE_ATTR_GROUP_SEQ, group->seq++);
 261         if (unlikely(ret < 0))
 262                 goto error;
 263
 264         if (data_len) {
 265                 int nla_len = nla_total_size(data_len);
 266                 struct nlattr *nla;
 267
 268                 nla = skb_put(nl_skb, nla_len);
 269                 nla->nla_type = PSAMPLE_ATTR_DATA;
 270                 nla->nla_len = nla_attr_size(data_len);
 271
 272                 if (skb_copy_bits(skb, 0, nla_data(nla), data_len))
 273                         goto error;
 274         }
 275
 276         genlmsg_end(nl_skb, data);
 277         genlmsg_multicast_netns(&psample_nl_family, group->net, nl_skb, 0,
 278                                 PSAMPLE_NL_MCGRP_SAMPLE, GFP_ATOMIC);
 279
 280         return;
 281 error:
 282         pr_err_ratelimited("Could not create psample log message\n");
 283         nlmsg_free(nl_skb);
 284 }
 285 EXPORT_SYMBOL_GPL(psample_sample_packet);
 286
 287 static int __init psample_module_init(void)
 288 {
 289         return genl_register_family(&psample_nl_family);
 290 }
 291
 292 static void __exit psample_module_exit(void)
 293 {
 294         genl_unregister_family(&psample_nl_family);
 295 }
 296
 297 module_init(psample_module_init);
 298 module_exit(psample_module_exit);
 299
 300 MODULE_AUTHOR("Yotam Gigi <yotam.gi@gmail.com>");
 301 MODULE_DESCRIPTION("netlink channel for packet sampling");
 302 MODULE_LICENSE("GPL v2");