zebra/rule_netlink.c

   1 /*
   2  * Zebra Policy Based Routing (PBR) interaction with the kernel using
   3  * netlink.
   4  * Copyright (C) 2018  Cumulus Networks, Inc.
   5  *
   6  * This file is part of FRR.
   7  *
   8  * FRR is free software; you can redistribute it and/or modify it
   9  * under the terms of the GNU General Public License as published by the
  10  * Free Software Foundation; either version 2, or (at your option) any
  11  * later version.
  12  *
  13  * FRR is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * along with FRR; see the file COPYING.  If not, write to the Free
  20  * Software Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA
  21  * 02111-1307, USA.
  22  */
  23
  24 #include <zebra.h>
  25
  26 #ifdef HAVE_NETLINK
  27
  28 #include "if.h"
  29 #include "prefix.h"
  30 #include "vrf.h"
  31
  32 #include <linux/fib_rules.h>
  33 #include "zebra/zserv.h"
  34 #include "zebra/zebra_ns.h"
  35 #include "zebra/zebra_vrf.h"
  36 #include "zebra/rt.h"
  37 #include "zebra/interface.h"
  38 #include "zebra/debug.h"
  39 #include "zebra/rtadv.h"
  40 #include "zebra/kernel_netlink.h"
  41 #include "zebra/rule_netlink.h"
  42 #include "zebra/zebra_pbr.h"
  43 #include "zebra/zebra_errors.h"
  44 #include "zebra/zebra_dplane.h"
  45
  46 /* definitions */
  47
  48 /* static function declarations */
  49
  50 /* Private functions */
  51
  52
  53 /*
  54  * netlink_rule_msg_encode
  55  *
  56  * Encodes netlink RTM_ADDRULE/RTM_DELRULE message to buffer buf of size buflen.
  57  *
  58  * Returns -1 on failure, 0 when the msg doesn't fit entirely in the buffer
  59  * or the number of bytes written to buf.
  60  */
  61 static ssize_t
  62 netlink_rule_msg_encode(int cmd, const struct zebra_dplane_ctx *ctx,
  63                         uint32_t filter_bm, uint32_t priority, uint32_t table,
  64                         const struct prefix *src_ip,
  65                         const struct prefix *dst_ip, uint32_t fwmark,
  66                         uint8_t dsfield, void *buf, size_t buflen)
  67 {
  68         uint8_t protocol = RTPROT_ZEBRA;
  69         int family;
  70         int bytelen;
  71         struct {
  72                 struct nlmsghdr n;
  73                 struct fib_rule_hdr frh;
  74                 char buf[];
  75         } *req = buf;
  76
  77         const char *ifname = dplane_ctx_rule_get_ifname(ctx);
  78
  79         if (buflen < sizeof(*req))
  80                 return 0;
  81         memset(req, 0, sizeof(*req));
  82
  83         /* Assume ipv4 if no src/dst set, we only support ipv4/ipv6 */
  84         if (PREFIX_FAMILY(src_ip))
  85                 family = PREFIX_FAMILY(src_ip);
  86         else if (PREFIX_FAMILY(dst_ip))
  87                 family = PREFIX_FAMILY(dst_ip);
  88         else
  89                 family = AF_INET;
  90
  91         bytelen = (family == AF_INET ? 4 : 16);
  92
  93         req->n.nlmsg_type = cmd;
  94         req->n.nlmsg_len = NLMSG_LENGTH(sizeof(struct rtmsg));
  95         req->n.nlmsg_flags = NLM_F_REQUEST;
  96
  97         req->frh.family = family;
  98         req->frh.action = FR_ACT_TO_TBL;
  99
 100         if (!nl_attr_put(&req->n, buflen, FRA_PROTOCOL, &protocol,
 101                          sizeof(protocol)))
 102                 return 0;
 103
 104         /* rule's pref # */
 105         if (!nl_attr_put32(&req->n, buflen, FRA_PRIORITY, priority))
 106                 return 0;
 107
 108         /* interface on which applied */
 109         if (!nl_attr_put(&req->n, buflen, FRA_IFNAME, ifname,
 110                          strlen(ifname) + 1))
 111                 return 0;
 112
 113         /* source IP, if specified */
 114         if (filter_bm & PBR_FILTER_SRC_IP) {
 115                 req->frh.src_len = src_ip->prefixlen;
 116                 if (!nl_attr_put(&req->n, buflen, FRA_SRC, &src_ip->u.prefix,
 117                                  bytelen))
 118                         return 0;
 119         }
 120
 121         /* destination IP, if specified */
 122         if (filter_bm & PBR_FILTER_DST_IP) {
 123                 req->frh.dst_len = dst_ip->prefixlen;
 124                 if (!nl_attr_put(&req->n, buflen, FRA_DST, &dst_ip->u.prefix,
 125                                  bytelen))
 126                         return 0;
 127         }
 128
 129         /* fwmark, if specified */
 130         if (filter_bm & PBR_FILTER_FWMARK) {
 131                 if (!nl_attr_put32(&req->n, buflen, FRA_FWMARK, fwmark))
 132                         return 0;
 133         }
 134
 135         /* dsfield, if specified */
 136         if (filter_bm & PBR_FILTER_DSFIELD)
 137                 req->frh.tos = dsfield;
 138
 139         /* Route table to use to forward, if filter criteria matches. */
 140         if (table < 256)
 141                 req->frh.table = table;
 142         else {
 143                 req->frh.table = RT_TABLE_UNSPEC;
 144                 if (!nl_attr_put32(&req->n, buflen, FRA_TABLE, table))
 145                         return 0;
 146         }
 147
 148         if (IS_ZEBRA_DEBUG_KERNEL)
 149                 zlog_debug(
 150                         "Tx %s family %s IF %s Pref %u Fwmark %u Src %pFX Dst %pFX Table %u",
 151                         nl_msg_type_to_str(cmd), nl_family_to_str(family),
 152                         ifname, priority, fwmark, src_ip, dst_ip, table);
 153
 154         return NLMSG_ALIGN(req->n.nlmsg_len);
 155 }
 156
 157 static ssize_t netlink_rule_msg_encoder(struct zebra_dplane_ctx *ctx, void *buf,
 158                                         size_t buflen)
 159 {
 160         int cmd = RTM_NEWRULE;
 161
 162         if (dplane_ctx_get_op(ctx) == DPLANE_OP_RULE_DELETE)
 163                 cmd = RTM_DELRULE;
 164
 165         return netlink_rule_msg_encode(
 166                 cmd, ctx, dplane_ctx_rule_get_filter_bm(ctx),
 167                 dplane_ctx_rule_get_priority(ctx),
 168                 dplane_ctx_rule_get_table(ctx), dplane_ctx_rule_get_src_ip(ctx),
 169                 dplane_ctx_rule_get_dst_ip(ctx),
 170                 dplane_ctx_rule_get_fwmark(ctx),
 171                 dplane_ctx_rule_get_dsfield(ctx), buf, buflen);
 172 }
 173
 174 static ssize_t netlink_oldrule_msg_encoder(struct zebra_dplane_ctx *ctx,
 175                                            void *buf, size_t buflen)
 176 {
 177         return netlink_rule_msg_encode(
 178                 RTM_DELRULE, ctx, dplane_ctx_rule_get_old_filter_bm(ctx),
 179                 dplane_ctx_rule_get_old_priority(ctx),
 180                 dplane_ctx_rule_get_old_table(ctx),
 181                 dplane_ctx_rule_get_old_src_ip(ctx),
 182                 dplane_ctx_rule_get_old_dst_ip(ctx),
 183                 dplane_ctx_rule_get_old_fwmark(ctx),
 184                 dplane_ctx_rule_get_old_dsfield(ctx), buf, buflen);
 185 }
 186
 187 /* Public functions */
 188
 189 enum netlink_msg_status
 190 netlink_put_rule_update_msg(struct nl_batch *bth, struct zebra_dplane_ctx *ctx)
 191 {
 192         enum dplane_op_e op;
 193         enum netlink_msg_status ret;
 194
 195         op = dplane_ctx_get_op(ctx);
 196         if (!(op == DPLANE_OP_RULE_ADD || op == DPLANE_OP_RULE_UPDATE
 197               || op == DPLANE_OP_RULE_DELETE)) {
 198                 flog_err(
 199                         EC_ZEBRA_PBR_RULE_UPDATE,
 200                         "Context received for kernel rule update with incorrect OP code (%u)",
 201                         op);
 202                 return FRR_NETLINK_ERROR;
 203         }
 204
 205         ret = netlink_batch_add_msg(bth, ctx, netlink_rule_msg_encoder, false);
 206
 207         /**
 208          * Delete the old one.
 209          *
 210          * Don't care about this result right?
 211          */
 212         if (op == DPLANE_OP_RULE_UPDATE)
 213                 netlink_batch_add_msg(bth, ctx, netlink_oldrule_msg_encoder,
 214                                       true);
 215
 216         return ret;
 217 }
 218
 219 /*
 220  * Handle netlink notification informing a rule add or delete.
 221  * Handling of an ADD is TBD.
 222  * DELs are notified up, if other attributes indicate it may be a
 223  * notification of interest. The expectation is that if this corresponds
 224  * to a PBR rule added by FRR, it will be readded.
 225  *
 226  * If startup and we see a rule we created, delete it as its leftover
 227  * from a previous instance and should have been removed on shutdown.
 228  *
 229  */
 230 int netlink_rule_change(struct nlmsghdr *h, ns_id_t ns_id, int startup)
 231 {
 232         struct zebra_ns *zns;
 233         struct fib_rule_hdr *frh;
 234         struct rtattr *tb[FRA_MAX + 1];
 235         int len;
 236         char *ifname;
 237         struct zebra_pbr_rule rule = {};
 238         uint8_t proto = 0;
 239
 240         /* Basic validation followed by extracting attributes. */
 241         if (h->nlmsg_type != RTM_NEWRULE && h->nlmsg_type != RTM_DELRULE)
 242                 return 0;
 243
 244         len = h->nlmsg_len - NLMSG_LENGTH(sizeof(struct fib_rule_hdr));
 245         if (len < 0) {
 246                 zlog_err(
 247                         "%s: Message received from netlink is of a broken size: %d %zu",
 248                         __func__, h->nlmsg_len,
 249                         (size_t)NLMSG_LENGTH(sizeof(struct fib_rule_hdr)));
 250                 return -1;
 251         }
 252
 253         frh = NLMSG_DATA(h);
 254
 255         if (frh->family != AF_INET && frh->family != AF_INET6) {
 256                 if (frh->family == RTNL_FAMILY_IPMR
 257                     || frh->family == RTNL_FAMILY_IP6MR) {
 258                         if (IS_ZEBRA_DEBUG_KERNEL)
 259                                 zlog_debug(
 260                                         "Received rule netlink that we are ignoring for family %u, rule change: %u",
 261                                         frh->family, h->nlmsg_type);
 262                         return 0;
 263                 }
 264                 flog_warn(
 265                         EC_ZEBRA_NETLINK_INVALID_AF,
 266                         "Invalid address family: %u received from kernel rule change: %u",
 267                         frh->family, h->nlmsg_type);
 268                 return 0;
 269         }
 270         if (frh->action != FR_ACT_TO_TBL)
 271                 return 0;
 272
 273         memset(tb, 0, sizeof(tb));
 274         netlink_parse_rtattr(tb, FRA_MAX, RTM_RTA(frh), len);
 275
 276         if (tb[FRA_PRIORITY])
 277                 rule.rule.priority = *(uint32_t *)RTA_DATA(tb[FRA_PRIORITY]);
 278
 279         if (tb[FRA_SRC]) {
 280                 if (frh->family == AF_INET)
 281                         memcpy(&rule.rule.filter.src_ip.u.prefix4,
 282                                RTA_DATA(tb[FRA_SRC]), 4);
 283                 else
 284                         memcpy(&rule.rule.filter.src_ip.u.prefix6,
 285                                RTA_DATA(tb[FRA_SRC]), 16);
 286                 rule.rule.filter.src_ip.prefixlen = frh->src_len;
 287                 rule.rule.filter.src_ip.family = frh->family;
 288                 rule.rule.filter.filter_bm |= PBR_FILTER_SRC_IP;
 289         }
 290
 291         if (tb[FRA_DST]) {
 292                 if (frh->family == AF_INET)
 293                         memcpy(&rule.rule.filter.dst_ip.u.prefix4,
 294                                RTA_DATA(tb[FRA_DST]), 4);
 295                 else
 296                         memcpy(&rule.rule.filter.dst_ip.u.prefix6,
 297                                RTA_DATA(tb[FRA_DST]), 16);
 298                 rule.rule.filter.dst_ip.prefixlen = frh->dst_len;
 299                 rule.rule.filter.dst_ip.family = frh->family;
 300                 rule.rule.filter.filter_bm |= PBR_FILTER_DST_IP;
 301         }
 302
 303         if (tb[FRA_TABLE])
 304                 rule.rule.action.table = *(uint32_t *)RTA_DATA(tb[FRA_TABLE]);
 305         else
 306                 rule.rule.action.table = frh->table;
 307
 308         /* TBD: We don't care about rules not specifying an IIF. */
 309         if (tb[FRA_IFNAME] == NULL)
 310                 return 0;
 311
 312         if (tb[FRA_PROTOCOL])
 313                 proto = *(uint8_t *)RTA_DATA(tb[FRA_PROTOCOL]);
 314
 315         ifname = (char *)RTA_DATA(tb[FRA_IFNAME]);
 316         strlcpy(rule.ifname, ifname, sizeof(rule.ifname));
 317
 318         if (h->nlmsg_type == RTM_NEWRULE) {
 319                 /*
 320                  * If we see a rule at startup we created, delete it now.
 321                  * It should have been flushed on a previous shutdown.
 322                  */
 323                 if (startup && proto == RTPROT_ZEBRA) {
 324                         enum zebra_dplane_result ret;
 325
 326                         ret = dplane_pbr_rule_delete(&rule);
 327
 328                         zlog_debug(
 329                                 "%s: %s leftover rule: family %s IF %s Pref %u Src %pFX Dst %pFX Table %u",
 330                                 __func__,
 331                                 ((ret == ZEBRA_DPLANE_REQUEST_FAILURE)
 332                                          ? "Failed to remove"
 333                                          : "Removed"),
 334                                 nl_family_to_str(frh->family), rule.ifname,
 335                                 rule.rule.priority, &rule.rule.filter.src_ip,
 336                                 &rule.rule.filter.dst_ip,
 337                                 rule.rule.action.table);
 338                 }
 339
 340                 /* TBD */
 341                 return 0;
 342         }
 343
 344         zns = zebra_ns_lookup(ns_id);
 345
 346         /* If we don't know the interface, we don't care. */
 347         if (!if_lookup_by_name_per_ns(zns, ifname))
 348                 return 0;
 349
 350         if (IS_ZEBRA_DEBUG_KERNEL)
 351                 zlog_debug(
 352                         "Rx %s family %s IF %s Pref %u Src %pFX Dst %pFX Table %u",
 353                         nl_msg_type_to_str(h->nlmsg_type),
 354                         nl_family_to_str(frh->family), rule.ifname,
 355                         rule.rule.priority, &rule.rule.filter.src_ip,
 356                         &rule.rule.filter.dst_ip, rule.rule.action.table);
 357
 358         return kernel_pbr_rule_del(&rule);
 359 }
 360
 361 /*
 362  * Request rules from the kernel
 363  */
 364 static int netlink_request_rules(struct zebra_ns *zns, int family, int type)
 365 {
 366         struct {
 367                 struct nlmsghdr n;
 368                 struct fib_rule_hdr frh;
 369                 char buf[NL_PKT_BUF_SIZE];
 370         } req;
 371
 372         memset(&req, 0, sizeof(req));
 373         req.n.nlmsg_type = type;
 374         req.n.nlmsg_flags = NLM_F_ROOT | NLM_F_MATCH | NLM_F_REQUEST;
 375         req.n.nlmsg_len = NLMSG_LENGTH(sizeof(struct fib_rule_hdr));
 376         req.frh.family = family;
 377
 378         return netlink_request(&zns->netlink_cmd, &req);
 379 }
 380
 381 /*
 382  * Get to know existing PBR rules in the kernel - typically called at startup.
 383  */
 384 int netlink_rules_read(struct zebra_ns *zns)
 385 {
 386         int ret;
 387         struct zebra_dplane_info dp_info;
 388
 389         zebra_dplane_info_from_zns(&dp_info, zns, true);
 390
 391         ret = netlink_request_rules(zns, AF_INET, RTM_GETRULE);
 392         if (ret < 0)
 393                 return ret;
 394
 395         ret = netlink_parse_info(netlink_rule_change, &zns->netlink_cmd,
 396                                  &dp_info, 0, 1);
 397         if (ret < 0)
 398                 return ret;
 399
 400         ret = netlink_request_rules(zns, AF_INET6, RTM_GETRULE);
 401         if (ret < 0)
 402                 return ret;
 403
 404         ret = netlink_parse_info(netlink_rule_change, &zns->netlink_cmd,
 405                                  &dp_info, 0, 1);
 406         return ret;
 407 }
 408
 409 #endif /* HAVE_NETLINK */