datapath/brcompat.c

   1 /*
   2  * Copyright (c) 2009 Nicira Networks.
   3  * Distributed under the terms of the GNU GPL version 2.
   4  *
   5  * Significant portions of this file may be copied from parts of the Linux
   6  * kernel, by Linus Torvalds and others.
   7  */
   8
   9 #include <linux/kernel.h>
  10 #include <asm/uaccess.h>
  11 #include <linux/completion.h>
  12 #include <linux/etherdevice.h>
  13 #include <linux/if_bridge.h>
  14 #include <linux/netdevice.h>
  15 #include <linux/rtnetlink.h>
  16 #include <net/genetlink.h>
  17
  18 #include "compat.h"
  19 #include "openvswitch/brcompat-netlink.h"
  20 #include "brc_procfs.h"
  21 #include "datapath.h"
  22
  23 static struct genl_family brc_genl_family;
  24 static struct genl_multicast_group brc_mc_group;
  25
  26 /* Time to wait for ovs-vswitchd to respond to a datapath action, in
  27  * jiffies. */
  28 #define BRC_TIMEOUT (HZ * 5)
  29
  30 /* Mutex to serialize ovs-brcompatd callbacks.  (Some callbacks naturally hold
  31  * br_ioctl_mutex, others hold rtnl_lock, but we can't take the former
  32  * ourselves and we don't want to hold the latter over a potentially long
  33  * period of time.) */
  34 static DEFINE_MUTEX(brc_serial);
  35
  36 /* Userspace communication. */
  37 static DEFINE_SPINLOCK(brc_lock);    /* Ensure atomic access to these vars. */
  38 static DECLARE_COMPLETION(brc_done); /* Userspace signaled operation done? */
  39 static struct sk_buff *brc_reply;    /* Reply from userspace. */
  40 static u32 brc_seq;                  /* Sequence number for current op. */
  41
  42 static struct sk_buff *brc_send_command(struct sk_buff *, struct nlattr **attrs);
  43 static int brc_send_simple_command(struct sk_buff *);
  44
  45 static struct sk_buff *
  46 brc_make_request(int op, const char *bridge, const char *port)
  47 {
  48         struct sk_buff *skb = genlmsg_new(NLMSG_GOODSIZE, GFP_KERNEL);
  49         if (!skb)
  50                 goto error;
  51
  52         genlmsg_put(skb, 0, 0, &brc_genl_family, 0, op);
  53         if (bridge)
  54                 NLA_PUT_STRING(skb, BRC_GENL_A_DP_NAME, bridge);
  55         if (port)
  56                 NLA_PUT_STRING(skb, BRC_GENL_A_PORT_NAME, port);
  57         return skb;
  58
  59 nla_put_failure:
  60         kfree_skb(skb);
  61 error:
  62         return NULL;
  63 }
  64
  65 static int brc_send_simple_command(struct sk_buff *request)
  66 {
  67         struct nlattr *attrs[BRC_GENL_A_MAX + 1];
  68         struct sk_buff *reply;
  69         int error;
  70
  71         reply = brc_send_command(request, attrs);
  72         if (IS_ERR(reply))
  73                 return PTR_ERR(reply);
  74
  75         error = nla_get_u32(attrs[BRC_GENL_A_ERR_CODE]);
  76         kfree_skb(reply);
  77         return -error;
  78 }
  79
  80 static int brc_add_del_bridge(char __user *uname, int add)
  81 {
  82         struct sk_buff *request;
  83         char name[IFNAMSIZ];
  84
  85         if (copy_from_user(name, uname, IFNAMSIZ))
  86                 return -EFAULT;
  87
  88         name[IFNAMSIZ - 1] = 0;
  89         request = brc_make_request(add ? BRC_GENL_C_DP_ADD : BRC_GENL_C_DP_DEL,
  90                                    name, NULL);
  91         if (!request)
  92                 return -ENOMEM;
  93
  94         return brc_send_simple_command(request);
  95 }
  96
  97 static int brc_get_indices(int op, const char *br_name,
  98                            int __user *uindices, int n)
  99 {
 100         struct nlattr *attrs[BRC_GENL_A_MAX + 1];
 101         struct sk_buff *request, *reply;
 102         int *indices;
 103         int ret;
 104         int len;
 105
 106         if (n < 0)
 107                 return -EINVAL;
 108         if (n >= 2048)
 109                 return -ENOMEM;
 110
 111         request = brc_make_request(op, br_name, NULL);
 112         if (!request)
 113                 return -ENOMEM;
 114
 115         reply = brc_send_command(request, attrs);
 116         ret = PTR_ERR(reply);
 117         if (IS_ERR(reply))
 118                 goto exit;
 119
 120         ret = -nla_get_u32(attrs[BRC_GENL_A_ERR_CODE]);
 121         if (ret < 0)
 122                 goto exit_free_skb;
 123
 124         ret = -EINVAL;
 125         if (!attrs[BRC_GENL_A_IFINDEXES])
 126                 goto exit_free_skb;
 127
 128         len = nla_len(attrs[BRC_GENL_A_IFINDEXES]);
 129         indices = nla_data(attrs[BRC_GENL_A_IFINDEXES]);
 130         if (len % sizeof(int))
 131                 goto exit_free_skb;
 132
 133         n = min_t(int, n, len / sizeof(int));
 134         ret = copy_to_user(uindices, indices, n * sizeof(int)) ? -EFAULT : n;
 135
 136 exit_free_skb:
 137         kfree_skb(reply);
 138 exit:
 139         return ret;
 140 }
 141
 142 /* Called with br_ioctl_mutex. */
 143 static int brc_get_bridges(int __user *uindices, int n)
 144 {
 145         return brc_get_indices(BRC_GENL_C_GET_BRIDGES, NULL, uindices, n);
 146 }
 147
 148 /* Legacy deviceless bridge ioctl's.  Called with br_ioctl_mutex. */
 149 static int
 150 old_deviceless(void __user *uarg)
 151 {
 152         unsigned long args[3];
 153
 154         if (copy_from_user(args, uarg, sizeof(args)))
 155                 return -EFAULT;
 156
 157         switch (args[0]) {
 158         case BRCTL_GET_BRIDGES:
 159                 return brc_get_bridges((int __user *)args[1], args[2]);
 160
 161         case BRCTL_ADD_BRIDGE:
 162                 return brc_add_del_bridge((void __user *)args[1], 1);
 163         case BRCTL_DEL_BRIDGE:
 164                 return brc_add_del_bridge((void __user *)args[1], 0);
 165         }
 166
 167         return -EOPNOTSUPP;
 168 }
 169
 170 /* Called with the br_ioctl_mutex. */
 171 static int
 172 #if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,23)
 173 brc_ioctl_deviceless_stub(unsigned int cmd, void __user *uarg)
 174 #else
 175 brc_ioctl_deviceless_stub(struct net *net, unsigned int cmd, void __user *uarg)
 176 #endif
 177 {
 178         switch (cmd) {
 179         case SIOCGIFBR:
 180         case SIOCSIFBR:
 181                 return old_deviceless(uarg);
 182
 183         case SIOCBRADDBR:
 184                 return brc_add_del_bridge(uarg, 1);
 185         case SIOCBRDELBR:
 186                 return brc_add_del_bridge(uarg, 0);
 187         }
 188
 189         return -EOPNOTSUPP;
 190 }
 191
 192 static int
 193 brc_add_del_port(struct net_device *dev, int port_ifindex, int add)
 194 {
 195         struct sk_buff *request;
 196         struct net_device *port;
 197         int err;
 198
 199         port = __dev_get_by_index(&init_net, port_ifindex);
 200         if (!port)
 201                 return -EINVAL;
 202
 203         /* Save name of dev and port because there's a race between the
 204          * rtnl_unlock() and the brc_send_simple_command(). */
 205         request = brc_make_request(add ? BRC_GENL_C_PORT_ADD : BRC_GENL_C_PORT_DEL,
 206                                    dev->name, port->name);
 207         if (!request)
 208                 return -ENOMEM;
 209
 210         rtnl_unlock();
 211         err = brc_send_simple_command(request);
 212         rtnl_lock();
 213
 214         return err;
 215 }
 216
 217 static int
 218 brc_get_bridge_info(struct net_device *dev, struct __bridge_info __user *ub)
 219 {
 220         struct __bridge_info b;
 221         u64 id = 0;
 222         int i;
 223
 224         memset(&b, 0, sizeof(struct __bridge_info));
 225
 226         for (i=0; i<ETH_ALEN; i++)
 227                 id |= (u64)dev->dev_addr[i] << (8*(ETH_ALEN-1 - i));
 228         b.bridge_id = cpu_to_be64(id);
 229         b.stp_enabled = 0;
 230
 231         if (copy_to_user(ub, &b, sizeof(struct __bridge_info)))
 232                 return -EFAULT;
 233
 234         return 0;
 235 }
 236
 237 static int
 238 brc_get_port_list(struct net_device *dev, int __user *uindices, int num)
 239 {
 240         int retval;
 241
 242         rtnl_unlock();
 243         retval = brc_get_indices(BRC_GENL_C_GET_PORTS, dev->name,
 244                                  uindices, num);
 245         rtnl_lock();
 246
 247         return retval;
 248 }
 249
 250 /*
 251  * Format up to a page worth of forwarding table entries
 252  * userbuf -- where to copy result
 253  * maxnum  -- maximum number of entries desired
 254  *            (limited to a page for sanity)
 255  * offset  -- number of records to skip
 256  */
 257 static int brc_get_fdb_entries(struct net_device *dev, void __user *userbuf,
 258                                unsigned long maxnum, unsigned long offset)
 259 {
 260         struct nlattr *attrs[BRC_GENL_A_MAX + 1];
 261         struct sk_buff *request, *reply;
 262         int retval;
 263         int len;
 264
 265         /* Clamp size to PAGE_SIZE, test maxnum to avoid overflow */
 266         if (maxnum > PAGE_SIZE/sizeof(struct __fdb_entry))
 267                 maxnum = PAGE_SIZE/sizeof(struct __fdb_entry);
 268
 269         request = brc_make_request(BRC_GENL_C_FDB_QUERY, dev->name, NULL);
 270         if (!request)
 271                 return -ENOMEM;
 272         NLA_PUT_U64(request, BRC_GENL_A_FDB_COUNT, maxnum);
 273         NLA_PUT_U64(request, BRC_GENL_A_FDB_SKIP, offset);
 274
 275         rtnl_unlock();
 276         reply = brc_send_command(request, attrs);
 277         retval = PTR_ERR(reply);
 278         if (IS_ERR(reply))
 279                 goto exit;
 280
 281         retval = -nla_get_u32(attrs[BRC_GENL_A_ERR_CODE]);
 282         if (retval < 0)
 283                 goto exit_free_skb;
 284
 285         retval = -EINVAL;
 286         if (!attrs[BRC_GENL_A_FDB_DATA])
 287                 goto exit_free_skb;
 288         len = nla_len(attrs[BRC_GENL_A_FDB_DATA]);
 289         if (len % sizeof(struct __fdb_entry) ||
 290             len / sizeof(struct __fdb_entry) > maxnum)
 291                 goto exit_free_skb;
 292
 293         retval = len / sizeof(struct __fdb_entry);
 294         if (copy_to_user(userbuf, nla_data(attrs[BRC_GENL_A_FDB_DATA]), len))
 295                 retval = -EFAULT;
 296
 297 exit_free_skb:
 298         kfree_skb(reply);
 299 exit:
 300         rtnl_lock();
 301         return retval;
 302
 303 nla_put_failure:
 304         kfree_skb(request);
 305         return -ENOMEM;
 306 }
 307
 308 /* Legacy ioctl's through SIOCDEVPRIVATE.  Called with rtnl_lock. */
 309 static int
 310 old_dev_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
 311 {
 312         unsigned long args[4];
 313
 314         if (copy_from_user(args, rq->ifr_data, sizeof(args)))
 315                 return -EFAULT;
 316
 317         switch (args[0]) {
 318         case BRCTL_ADD_IF:
 319                 return brc_add_del_port(dev, args[1], 1);
 320         case BRCTL_DEL_IF:
 321                 return brc_add_del_port(dev, args[1], 0);
 322
 323         case BRCTL_GET_BRIDGE_INFO:
 324                 return brc_get_bridge_info(dev, (struct __bridge_info __user *)args[1]);
 325
 326         case BRCTL_GET_PORT_LIST:
 327                 return brc_get_port_list(dev, (int __user *)args[1], args[2]);
 328
 329         case BRCTL_GET_FDB_ENTRIES:
 330                 return brc_get_fdb_entries(dev, (void __user *)args[1],
 331                                            args[2], args[3]);
 332         }
 333
 334         return -EOPNOTSUPP;
 335 }
 336
 337 /* Called with the rtnl_lock. */
 338 static int
 339 brc_dev_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
 340 {
 341         int err;
 342
 343         switch (cmd) {
 344                 case SIOCDEVPRIVATE:
 345                         err = old_dev_ioctl(dev, rq, cmd);
 346                         break;
 347
 348                 case SIOCBRADDIF:
 349                         return brc_add_del_port(dev, rq->ifr_ifindex, 1);
 350                 case SIOCBRDELIF:
 351                         return brc_add_del_port(dev, rq->ifr_ifindex, 0);
 352
 353                 default:
 354                         err = -EOPNOTSUPP;
 355                         break;
 356         }
 357
 358         return err;
 359 }
 360
 361
 362 static struct genl_family brc_genl_family = {
 363         .id = GENL_ID_GENERATE,
 364         .hdrsize = 0,
 365         .name = BRC_GENL_FAMILY_NAME,
 366         .version = 1,
 367         .maxattr = BRC_GENL_A_MAX,
 368 };
 369
 370 static int brc_genl_query(struct sk_buff *skb, struct genl_info *info)
 371 {
 372         int err = -EINVAL;
 373         struct sk_buff *ans_skb;
 374         void *data;
 375
 376         ans_skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
 377         if (!ans_skb)
 378                 return -ENOMEM;
 379
 380         data = genlmsg_put_reply(ans_skb, info, &brc_genl_family,
 381                                  0, BRC_GENL_C_QUERY_MC);
 382         if (data == NULL) {
 383                 err = -ENOMEM;
 384                 goto err;
 385         }
 386         NLA_PUT_U32(ans_skb, BRC_GENL_A_MC_GROUP, brc_mc_group.id);
 387
 388         genlmsg_end(ans_skb, data);
 389         return genlmsg_reply(ans_skb, info);
 390
 391 err:
 392 nla_put_failure:
 393         kfree_skb(ans_skb);
 394         return err;
 395 }
 396
 397 static struct genl_ops brc_genl_ops_query_dp = {
 398         .cmd = BRC_GENL_C_QUERY_MC,
 399         .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privelege. */
 400         .policy = NULL,
 401         .doit = brc_genl_query,
 402         .dumpit = NULL
 403 };
 404
 405 /* Attribute policy: what each attribute may contain.  */
 406 static struct nla_policy brc_genl_policy[BRC_GENL_A_MAX + 1] = {
 407         [BRC_GENL_A_ERR_CODE] = { .type = NLA_U32 },
 408
 409         [BRC_GENL_A_PROC_DIR] = { .type = NLA_NUL_STRING },
 410         [BRC_GENL_A_PROC_NAME] = { .type = NLA_NUL_STRING },
 411         [BRC_GENL_A_PROC_DATA] = { .type = NLA_NUL_STRING },
 412
 413         [BRC_GENL_A_FDB_DATA] = { .type = NLA_UNSPEC },
 414 };
 415
 416 static int
 417 brc_genl_dp_result(struct sk_buff *skb, struct genl_info *info)
 418 {
 419         unsigned long int flags;
 420         int err;
 421
 422         if (!info->attrs[BRC_GENL_A_ERR_CODE])
 423                 return -EINVAL;
 424
 425         skb = skb_clone(skb, GFP_KERNEL);
 426         if (!skb)
 427                 return -ENOMEM;
 428
 429         spin_lock_irqsave(&brc_lock, flags);
 430         if (brc_seq == info->snd_seq) {
 431                 brc_seq++;
 432
 433                 if (brc_reply)
 434                         kfree_skb(brc_reply);
 435                 brc_reply = skb;
 436
 437                 complete(&brc_done);
 438                 err = 0;
 439         } else {
 440                 kfree_skb(skb);
 441                 err = -ESTALE;
 442         }
 443         spin_unlock_irqrestore(&brc_lock, flags);
 444
 445         return err;
 446 }
 447
 448 static struct genl_ops brc_genl_ops_dp_result = {
 449         .cmd = BRC_GENL_C_DP_RESULT,
 450         .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privelege. */
 451         .policy = brc_genl_policy,
 452         .doit = brc_genl_dp_result,
 453         .dumpit = NULL
 454 };
 455
 456 static struct genl_ops brc_genl_ops_set_proc = {
 457         .cmd = BRC_GENL_C_SET_PROC,
 458         .flags = GENL_ADMIN_PERM, /* Requires CAP_NET_ADMIN privelege. */
 459         .policy = brc_genl_policy,
 460         .doit = brc_genl_set_proc,
 461         .dumpit = NULL
 462 };
 463
 464 static struct sk_buff *brc_send_command(struct sk_buff *request, struct nlattr **attrs)
 465 {
 466         unsigned long int flags;
 467         struct sk_buff *reply;
 468         int error;
 469
 470         mutex_lock(&brc_serial);
 471
 472         /* Increment sequence number first, so that we ignore any replies
 473          * to stale requests. */
 474         spin_lock_irqsave(&brc_lock, flags);
 475         nlmsg_hdr(request)->nlmsg_seq = ++brc_seq;
 476         INIT_COMPLETION(brc_done);
 477         spin_unlock_irqrestore(&brc_lock, flags);
 478
 479         nlmsg_end(request, nlmsg_hdr(request));
 480
 481         /* Send message. */
 482         error = genlmsg_multicast(request, 0, brc_mc_group.id, GFP_KERNEL);
 483         if (error < 0)
 484                 goto error;
 485
 486         /* Wait for reply. */
 487         error = -ETIMEDOUT;
 488         if (!wait_for_completion_timeout(&brc_done, BRC_TIMEOUT))
 489                 goto error;
 490
 491         /* Grab reply. */
 492         spin_lock_irqsave(&brc_lock, flags);
 493         reply = brc_reply;
 494         brc_reply = NULL;
 495         spin_unlock_irqrestore(&brc_lock, flags);
 496
 497         mutex_unlock(&brc_serial);
 498
 499         /* Re-parse message.  Can't fail, since it parsed correctly once
 500          * already. */
 501         error = nlmsg_parse(nlmsg_hdr(reply), GENL_HDRLEN,
 502                             attrs, BRC_GENL_A_MAX, brc_genl_policy);
 503         WARN_ON(error);
 504
 505         return reply;
 506
 507 error:
 508         mutex_unlock(&brc_serial);
 509         return ERR_PTR(error);
 510 }
 511
 512 static int
 513 __init brc_init(void)
 514 {
 515         int err;
 516
 517         printk("Open vSwitch Bridge Compatibility, built "__DATE__" "__TIME__"\n");
 518
 519         /* Set the bridge ioctl handler */
 520         brioctl_set(brc_ioctl_deviceless_stub);
 521
 522         /* Set the openvswitch_mod device ioctl handler */
 523         dp_ioctl_hook = brc_dev_ioctl;
 524
 525         /* Randomize the initial sequence number.  This is not a security
 526          * feature; it only helps avoid crossed wires between userspace and
 527          * the kernel when the module is unloaded and reloaded. */
 528         brc_seq = net_random();
 529
 530         /* Register generic netlink family to communicate changes to
 531          * userspace. */
 532         err = genl_register_family(&brc_genl_family);
 533         if (err)
 534                 goto error;
 535
 536         err = genl_register_ops(&brc_genl_family, &brc_genl_ops_query_dp);
 537         if (err != 0)
 538                 goto err_unregister;
 539
 540         err = genl_register_ops(&brc_genl_family, &brc_genl_ops_dp_result);
 541         if (err != 0)
 542                 goto err_unregister;
 543
 544         err = genl_register_ops(&brc_genl_family, &brc_genl_ops_set_proc);
 545         if (err != 0)
 546                 goto err_unregister;
 547
 548         strcpy(brc_mc_group.name, "brcompat");
 549         err = genl_register_mc_group(&brc_genl_family, &brc_mc_group);
 550         if (err < 0)
 551                 goto err_unregister;
 552
 553         return 0;
 554
 555 err_unregister:
 556         genl_unregister_family(&brc_genl_family);
 557 error:
 558         printk(KERN_EMERG "brcompat: failed to install!");
 559         return err;
 560 }
 561
 562 static void
 563 brc_cleanup(void)
 564 {
 565         /* Unregister ioctl hooks */
 566         dp_ioctl_hook = NULL;
 567         brioctl_set(NULL);
 568
 569         genl_unregister_family(&brc_genl_family);
 570         brc_procfs_exit();
 571 }
 572
 573 module_init(brc_init);
 574 module_exit(brc_cleanup);
 575
 576 MODULE_DESCRIPTION("Open vSwitch bridge compatibility");
 577 MODULE_AUTHOR("Nicira Networks");
 578 MODULE_LICENSE("GPL");