net/rds/tcp.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/kernel.h>
  34 #include <linux/slab.h>
  35 #include <linux/in.h>
  36 #include <linux/module.h>
  37 #include <net/tcp.h>
  38 #include <net/net_namespace.h>
  39 #include <net/netns/generic.h>
  40
  41 #include "rds.h"
  42 #include "tcp.h"
  43
  44 /* only for info exporting */
  45 static DEFINE_SPINLOCK(rds_tcp_tc_list_lock);
  46 static LIST_HEAD(rds_tcp_tc_list);
  47 static unsigned int rds_tcp_tc_count;
  48
  49 /* Track rds_tcp_connection structs so they can be cleaned up */
  50 static DEFINE_SPINLOCK(rds_tcp_conn_lock);
  51 static LIST_HEAD(rds_tcp_conn_list);
  52
  53 static struct kmem_cache *rds_tcp_conn_slab;
  54
  55 static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
  56                                  void __user *buffer, size_t *lenp,
  57                                  loff_t *fpos);
  58
  59 static int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF;
  60 static int rds_tcp_min_rcvbuf = SOCK_MIN_RCVBUF;
  61
  62 static struct ctl_table rds_tcp_sysctl_table[] = {
  63 #define RDS_TCP_SNDBUF  0
  64         {
  65                 .procname       = "rds_tcp_sndbuf",
  66                 /* data is per-net pointer */
  67                 .maxlen         = sizeof(int),
  68                 .mode           = 0644,
  69                 .proc_handler   = rds_tcp_skbuf_handler,
  70                 .extra1         = &rds_tcp_min_sndbuf,
  71         },
  72 #define RDS_TCP_RCVBUF  1
  73         {
  74                 .procname       = "rds_tcp_rcvbuf",
  75                 /* data is per-net pointer */
  76                 .maxlen         = sizeof(int),
  77                 .mode           = 0644,
  78                 .proc_handler   = rds_tcp_skbuf_handler,
  79                 .extra1         = &rds_tcp_min_rcvbuf,
  80         },
  81         { }
  82 };
  83
  84 /* doing it this way avoids calling tcp_sk() */
  85 void rds_tcp_nonagle(struct socket *sock)
  86 {
  87         mm_segment_t oldfs = get_fs();
  88         int val = 1;
  89
  90         set_fs(KERNEL_DS);
  91         sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY, (char __user *)&val,
  92                               sizeof(val));
  93         set_fs(oldfs);
  94 }
  95
  96 u32 rds_tcp_snd_nxt(struct rds_tcp_connection *tc)
  97 {
  98         return tcp_sk(tc->t_sock->sk)->snd_nxt;
  99 }
 100
 101 u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
 102 {
 103         return tcp_sk(tc->t_sock->sk)->snd_una;
 104 }
 105
 106 void rds_tcp_restore_callbacks(struct socket *sock,
 107                                struct rds_tcp_connection *tc)
 108 {
 109         rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
 110         write_lock_bh(&sock->sk->sk_callback_lock);
 111
 112         /* done under the callback_lock to serialize with write_space */
 113         spin_lock(&rds_tcp_tc_list_lock);
 114         list_del_init(&tc->t_list_item);
 115         rds_tcp_tc_count--;
 116         spin_unlock(&rds_tcp_tc_list_lock);
 117
 118         tc->t_sock = NULL;
 119
 120         sock->sk->sk_write_space = tc->t_orig_write_space;
 121         sock->sk->sk_data_ready = tc->t_orig_data_ready;
 122         sock->sk->sk_state_change = tc->t_orig_state_change;
 123         sock->sk->sk_user_data = NULL;
 124
 125         write_unlock_bh(&sock->sk->sk_callback_lock);
 126 }
 127
 128 /*
 129  * rds_tcp_reset_callbacks() switches the to the new sock and
 130  * returns the existing tc->t_sock.
 131  *
 132  * The only functions that set tc->t_sock are rds_tcp_set_callbacks
 133  * and rds_tcp_reset_callbacks.  Send and receive trust that
 134  * it is set.  The absence of RDS_CONN_UP bit protects those paths
 135  * from being called while it isn't set.
 136  */
 137 void rds_tcp_reset_callbacks(struct socket *sock,
 138                              struct rds_conn_path *cp)
 139 {
 140         struct rds_tcp_connection *tc = cp->cp_transport_data;
 141         struct socket *osock = tc->t_sock;
 142
 143         if (!osock)
 144                 goto newsock;
 145
 146         /* Need to resolve a duelling SYN between peers.
 147          * We have an outstanding SYN to this peer, which may
 148          * potentially have transitioned to the RDS_CONN_UP state,
 149          * so we must quiesce any send threads before resetting
 150          * cp_transport_data. We quiesce these threads by setting
 151          * cp_state to something other than RDS_CONN_UP, and then
 152          * waiting for any existing threads in rds_send_xmit to
 153          * complete release_in_xmit(). (Subsequent threads entering
 154          * rds_send_xmit() will bail on !rds_conn_up().
 155          *
 156          * However an incoming syn-ack at this point would end up
 157          * marking the conn as RDS_CONN_UP, and would again permit
 158          * rds_send_xmi() threads through, so ideally we would
 159          * synchronize on RDS_CONN_UP after lock_sock(), but cannot
 160          * do that: waiting on !RDS_IN_XMIT after lock_sock() may
 161          * end up deadlocking with tcp_sendmsg(), and the RDS_IN_XMIT
 162          * would not get set. As a result, we set c_state to
 163          * RDS_CONN_RESETTTING, to ensure that rds_tcp_state_change
 164          * cannot mark rds_conn_path_up() in the window before lock_sock()
 165          */
 166         atomic_set(&cp->cp_state, RDS_CONN_RESETTING);
 167         wait_event(cp->cp_waitq, !test_bit(RDS_IN_XMIT, &cp->cp_flags));
 168         lock_sock(osock->sk);
 169         /* reset receive side state for rds_tcp_data_recv() for osock  */
 170         cancel_delayed_work_sync(&cp->cp_send_w);
 171         cancel_delayed_work_sync(&cp->cp_recv_w);
 172         if (tc->t_tinc) {
 173                 rds_inc_put(&tc->t_tinc->ti_inc);
 174                 tc->t_tinc = NULL;
 175         }
 176         tc->t_tinc_hdr_rem = sizeof(struct rds_header);
 177         tc->t_tinc_data_rem = 0;
 178         rds_tcp_restore_callbacks(osock, tc);
 179         release_sock(osock->sk);
 180         sock_release(osock);
 181 newsock:
 182         rds_send_path_reset(cp);
 183         lock_sock(sock->sk);
 184         rds_tcp_set_callbacks(sock, cp);
 185         release_sock(sock->sk);
 186 }
 187
 188 /* Add tc to rds_tcp_tc_list and set tc->t_sock. See comments
 189  * above rds_tcp_reset_callbacks for notes about synchronization
 190  * with data path
 191  */
 192 void rds_tcp_set_callbacks(struct socket *sock, struct rds_conn_path *cp)
 193 {
 194         struct rds_tcp_connection *tc = cp->cp_transport_data;
 195
 196         rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
 197         write_lock_bh(&sock->sk->sk_callback_lock);
 198
 199         /* done under the callback_lock to serialize with write_space */
 200         spin_lock(&rds_tcp_tc_list_lock);
 201         list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
 202         rds_tcp_tc_count++;
 203         spin_unlock(&rds_tcp_tc_list_lock);
 204
 205         /* accepted sockets need our listen data ready undone */
 206         if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
 207                 sock->sk->sk_data_ready = sock->sk->sk_user_data;
 208
 209         tc->t_sock = sock;
 210         tc->t_cpath = cp;
 211         tc->t_orig_data_ready = sock->sk->sk_data_ready;
 212         tc->t_orig_write_space = sock->sk->sk_write_space;
 213         tc->t_orig_state_change = sock->sk->sk_state_change;
 214
 215         sock->sk->sk_user_data = cp;
 216         sock->sk->sk_data_ready = rds_tcp_data_ready;
 217         sock->sk->sk_write_space = rds_tcp_write_space;
 218         sock->sk->sk_state_change = rds_tcp_state_change;
 219
 220         write_unlock_bh(&sock->sk->sk_callback_lock);
 221 }
 222
 223 static void rds_tcp_tc_info(struct socket *sock, unsigned int len,
 224                             struct rds_info_iterator *iter,
 225                             struct rds_info_lengths *lens)
 226 {
 227         struct rds_info_tcp_socket tsinfo;
 228         struct rds_tcp_connection *tc;
 229         unsigned long flags;
 230         struct sockaddr_in sin;
 231         int sinlen;
 232
 233         spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
 234
 235         if (len / sizeof(tsinfo) < rds_tcp_tc_count)
 236                 goto out;
 237
 238         list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
 239
 240                 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 0);
 241                 tsinfo.local_addr = sin.sin_addr.s_addr;
 242                 tsinfo.local_port = sin.sin_port;
 243                 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 1);
 244                 tsinfo.peer_addr = sin.sin_addr.s_addr;
 245                 tsinfo.peer_port = sin.sin_port;
 246
 247                 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
 248                 tsinfo.data_rem = tc->t_tinc_data_rem;
 249                 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
 250                 tsinfo.last_expected_una = tc->t_last_expected_una;
 251                 tsinfo.last_seen_una = tc->t_last_seen_una;
 252
 253                 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
 254         }
 255
 256 out:
 257         lens->nr = rds_tcp_tc_count;
 258         lens->each = sizeof(tsinfo);
 259
 260         spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
 261 }
 262
 263 static int rds_tcp_laddr_check(struct net *net, __be32 addr)
 264 {
 265         if (inet_addr_type(net, addr) == RTN_LOCAL)
 266                 return 0;
 267         return -EADDRNOTAVAIL;
 268 }
 269
 270 static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 271 {
 272         struct rds_tcp_connection *tc;
 273         int i;
 274
 275         for (i = 0; i < RDS_MPATH_WORKERS; i++) {
 276                 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
 277                 if (!tc)
 278                         return -ENOMEM;
 279
 280                 mutex_init(&tc->t_conn_path_lock);
 281                 tc->t_sock = NULL;
 282                 tc->t_tinc = NULL;
 283                 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
 284                 tc->t_tinc_data_rem = 0;
 285
 286                 conn->c_path[i].cp_transport_data = tc;
 287                 tc->t_cpath = &conn->c_path[i];
 288
 289                 spin_lock_irq(&rds_tcp_conn_lock);
 290                 list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
 291                 spin_unlock_irq(&rds_tcp_conn_lock);
 292                 rdsdebug("rds_conn_path [%d] tc %p\n", i,
 293                          conn->c_path[i].cp_transport_data);
 294         }
 295
 296         return 0;
 297 }
 298
 299 static void rds_tcp_conn_free(void *arg)
 300 {
 301         struct rds_tcp_connection *tc = arg;
 302         unsigned long flags;
 303         rdsdebug("freeing tc %p\n", tc);
 304
 305         spin_lock_irqsave(&rds_tcp_conn_lock, flags);
 306         list_del(&tc->t_tcp_node);
 307         spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
 308
 309         kmem_cache_free(rds_tcp_conn_slab, tc);
 310 }
 311
 312 static bool list_has_conn(struct list_head *list, struct rds_connection *conn)
 313 {
 314         struct rds_tcp_connection *tc, *_tc;
 315
 316         list_for_each_entry_safe(tc, _tc, list, t_tcp_node) {
 317                 if (tc->t_cpath->cp_conn == conn)
 318                         return true;
 319         }
 320         return false;
 321 }
 322
 323 static void rds_tcp_destroy_conns(void)
 324 {
 325         struct rds_tcp_connection *tc, *_tc;
 326         LIST_HEAD(tmp_list);
 327
 328         /* avoid calling conn_destroy with irqs off */
 329         spin_lock_irq(&rds_tcp_conn_lock);
 330         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 331                 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn))
 332                         list_move_tail(&tc->t_tcp_node, &tmp_list);
 333         }
 334         spin_unlock_irq(&rds_tcp_conn_lock);
 335
 336         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
 337                 rds_conn_destroy(tc->t_cpath->cp_conn);
 338 }
 339
 340 static void rds_tcp_exit(void);
 341
 342 struct rds_transport rds_tcp_transport = {
 343         .laddr_check            = rds_tcp_laddr_check,
 344         .xmit_path_prepare      = rds_tcp_xmit_path_prepare,
 345         .xmit_path_complete     = rds_tcp_xmit_path_complete,
 346         .xmit                   = rds_tcp_xmit,
 347         .recv_path              = rds_tcp_recv_path,
 348         .conn_alloc             = rds_tcp_conn_alloc,
 349         .conn_free              = rds_tcp_conn_free,
 350         .conn_path_connect      = rds_tcp_conn_path_connect,
 351         .conn_path_shutdown     = rds_tcp_conn_path_shutdown,
 352         .inc_copy_to_user       = rds_tcp_inc_copy_to_user,
 353         .inc_free               = rds_tcp_inc_free,
 354         .stats_info_copy        = rds_tcp_stats_info_copy,
 355         .exit                   = rds_tcp_exit,
 356         .t_owner                = THIS_MODULE,
 357         .t_name                 = "tcp",
 358         .t_type                 = RDS_TRANS_TCP,
 359         .t_prefer_loopback      = 1,
 360         .t_mp_capable           = 1,
 361 };
 362
 363 static int rds_tcp_netid;
 364
 365 /* per-network namespace private data for this module */
 366 struct rds_tcp_net {
 367         struct socket *rds_tcp_listen_sock;
 368         struct work_struct rds_tcp_accept_w;
 369         struct ctl_table_header *rds_tcp_sysctl;
 370         struct ctl_table *ctl_table;
 371         int sndbuf_size;
 372         int rcvbuf_size;
 373 };
 374
 375 /* All module specific customizations to the RDS-TCP socket should be done in
 376  * rds_tcp_tune() and applied after socket creation.
 377  */
 378 void rds_tcp_tune(struct socket *sock)
 379 {
 380         struct sock *sk = sock->sk;
 381         struct net *net = sock_net(sk);
 382         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 383
 384         rds_tcp_nonagle(sock);
 385         lock_sock(sk);
 386         if (rtn->sndbuf_size > 0) {
 387                 sk->sk_sndbuf = rtn->sndbuf_size;
 388                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
 389         }
 390         if (rtn->rcvbuf_size > 0) {
 391                 sk->sk_sndbuf = rtn->rcvbuf_size;
 392                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
 393         }
 394         release_sock(sk);
 395 }
 396
 397 static void rds_tcp_accept_worker(struct work_struct *work)
 398 {
 399         struct rds_tcp_net *rtn = container_of(work,
 400                                                struct rds_tcp_net,
 401                                                rds_tcp_accept_w);
 402
 403         while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
 404                 cond_resched();
 405 }
 406
 407 void rds_tcp_accept_work(struct sock *sk)
 408 {
 409         struct net *net = sock_net(sk);
 410         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 411
 412         queue_work(rds_wq, &rtn->rds_tcp_accept_w);
 413 }
 414
 415 static __net_init int rds_tcp_init_net(struct net *net)
 416 {
 417         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 418         struct ctl_table *tbl;
 419         int err = 0;
 420
 421         memset(rtn, 0, sizeof(*rtn));
 422
 423         /* {snd, rcv}buf_size default to 0, which implies we let the
 424          * stack pick the value, and permit auto-tuning of buffer size.
 425          */
 426         if (net == &init_net) {
 427                 tbl = rds_tcp_sysctl_table;
 428         } else {
 429                 tbl = kmemdup(rds_tcp_sysctl_table,
 430                               sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
 431                 if (!tbl) {
 432                         pr_warn("could not set allocate syctl table\n");
 433                         return -ENOMEM;
 434                 }
 435                 rtn->ctl_table = tbl;
 436         }
 437         tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
 438         tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
 439         rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
 440         if (!rtn->rds_tcp_sysctl) {
 441                 pr_warn("could not register sysctl\n");
 442                 err = -ENOMEM;
 443                 goto fail;
 444         }
 445         rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
 446         if (!rtn->rds_tcp_listen_sock) {
 447                 pr_warn("could not set up listen sock\n");
 448                 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
 449                 rtn->rds_tcp_sysctl = NULL;
 450                 err = -EAFNOSUPPORT;
 451                 goto fail;
 452         }
 453         INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
 454         return 0;
 455
 456 fail:
 457         if (net != &init_net)
 458                 kfree(tbl);
 459         return err;
 460 }
 461
 462 static void __net_exit rds_tcp_exit_net(struct net *net)
 463 {
 464         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 465
 466         if (rtn->rds_tcp_sysctl)
 467                 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
 468
 469         if (net != &init_net && rtn->ctl_table)
 470                 kfree(rtn->ctl_table);
 471
 472         /* If rds_tcp_exit_net() is called as a result of netns deletion,
 473          * the rds_tcp_kill_sock() device notifier would already have cleaned
 474          * up the listen socket, thus there is no work to do in this function.
 475          *
 476          * If rds_tcp_exit_net() is called as a result of module unload,
 477          * i.e., due to rds_tcp_exit() -> unregister_pernet_subsys(), then
 478          * we do need to clean up the listen socket here.
 479          */
 480         if (rtn->rds_tcp_listen_sock) {
 481                 rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 482                 rtn->rds_tcp_listen_sock = NULL;
 483                 flush_work(&rtn->rds_tcp_accept_w);
 484         }
 485 }
 486
 487 static struct pernet_operations rds_tcp_net_ops = {
 488         .init = rds_tcp_init_net,
 489         .exit = rds_tcp_exit_net,
 490         .id = &rds_tcp_netid,
 491         .size = sizeof(struct rds_tcp_net),
 492 };
 493
 494 /* explicitly send a RST on each socket, thereby releasing any socket refcnts
 495  * that may otherwise hold up netns deletion.
 496  */
 497 static void rds_tcp_conn_paths_destroy(struct rds_connection *conn)
 498 {
 499         struct rds_conn_path *cp;
 500         struct rds_tcp_connection *tc;
 501         int i;
 502         struct sock *sk;
 503
 504         for (i = 0; i < RDS_MPATH_WORKERS; i++) {
 505                 cp = &conn->c_path[i];
 506                 tc = cp->cp_transport_data;
 507                 if (!tc->t_sock)
 508                         continue;
 509                 sk = tc->t_sock->sk;
 510                 sk->sk_prot->disconnect(sk, 0);
 511                 tcp_done(sk);
 512         }
 513 }
 514
 515 static void rds_tcp_kill_sock(struct net *net)
 516 {
 517         struct rds_tcp_connection *tc, *_tc;
 518         LIST_HEAD(tmp_list);
 519         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 520
 521         rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 522         rtn->rds_tcp_listen_sock = NULL;
 523         flush_work(&rtn->rds_tcp_accept_w);
 524         spin_lock_irq(&rds_tcp_conn_lock);
 525         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 526                 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
 527
 528                 if (net != c_net || !tc->t_sock)
 529                         continue;
 530                 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn))
 531                         list_move_tail(&tc->t_tcp_node, &tmp_list);
 532         }
 533         spin_unlock_irq(&rds_tcp_conn_lock);
 534         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
 535                 rds_tcp_conn_paths_destroy(tc->t_cpath->cp_conn);
 536                 rds_conn_destroy(tc->t_cpath->cp_conn);
 537         }
 538 }
 539
 540 void *rds_tcp_listen_sock_def_readable(struct net *net)
 541 {
 542         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 543
 544         return rtn->rds_tcp_listen_sock->sk->sk_user_data;
 545 }
 546
 547 static int rds_tcp_dev_event(struct notifier_block *this,
 548                              unsigned long event, void *ptr)
 549 {
 550         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 551
 552         /* rds-tcp registers as a pernet subys, so the ->exit will only
 553          * get invoked after network acitivity has quiesced. We need to
 554          * clean up all sockets  to quiesce network activity, and use
 555          * the unregistration of the per-net loopback device as a trigger
 556          * to start that cleanup.
 557          */
 558         if (event == NETDEV_UNREGISTER_FINAL &&
 559             dev->ifindex == LOOPBACK_IFINDEX)
 560                 rds_tcp_kill_sock(dev_net(dev));
 561
 562         return NOTIFY_DONE;
 563 }
 564
 565 static struct notifier_block rds_tcp_dev_notifier = {
 566         .notifier_call        = rds_tcp_dev_event,
 567         .priority = -10, /* must be called after other network notifiers */
 568 };
 569
 570 /* when sysctl is used to modify some kernel socket parameters,this
 571  * function  resets the RDS connections in that netns  so that we can
 572  * restart with new parameters.  The assumption is that such reset
 573  * events are few and far-between.
 574  */
 575 static void rds_tcp_sysctl_reset(struct net *net)
 576 {
 577         struct rds_tcp_connection *tc, *_tc;
 578
 579         spin_lock_irq(&rds_tcp_conn_lock);
 580         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 581                 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
 582
 583                 if (net != c_net || !tc->t_sock)
 584                         continue;
 585
 586                 /* reconnect with new parameters */
 587                 rds_conn_path_drop(tc->t_cpath);
 588         }
 589         spin_unlock_irq(&rds_tcp_conn_lock);
 590 }
 591
 592 static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
 593                                  void __user *buffer, size_t *lenp,
 594                                  loff_t *fpos)
 595 {
 596         struct net *net = current->nsproxy->net_ns;
 597         int err;
 598
 599         err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
 600         if (err < 0) {
 601                 pr_warn("Invalid input. Must be >= %d\n",
 602                         *(int *)(ctl->extra1));
 603                 return err;
 604         }
 605         if (write)
 606                 rds_tcp_sysctl_reset(net);
 607         return 0;
 608 }
 609
 610 static void rds_tcp_exit(void)
 611 {
 612         rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 613         unregister_pernet_subsys(&rds_tcp_net_ops);
 614         if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
 615                 pr_warn("could not unregister rds_tcp_dev_notifier\n");
 616         rds_tcp_destroy_conns();
 617         rds_trans_unregister(&rds_tcp_transport);
 618         rds_tcp_recv_exit();
 619         kmem_cache_destroy(rds_tcp_conn_slab);
 620 }
 621 module_exit(rds_tcp_exit);
 622
 623 static int rds_tcp_init(void)
 624 {
 625         int ret;
 626
 627         rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
 628                                               sizeof(struct rds_tcp_connection),
 629                                               0, 0, NULL);
 630         if (!rds_tcp_conn_slab) {
 631                 ret = -ENOMEM;
 632                 goto out;
 633         }
 634
 635         ret = register_netdevice_notifier(&rds_tcp_dev_notifier);
 636         if (ret) {
 637                 pr_warn("could not register rds_tcp_dev_notifier\n");
 638                 goto out;
 639         }
 640
 641         ret = register_pernet_subsys(&rds_tcp_net_ops);
 642         if (ret)
 643                 goto out_slab;
 644
 645         ret = rds_tcp_recv_init();
 646         if (ret)
 647                 goto out_pernet;
 648
 649         ret = rds_trans_register(&rds_tcp_transport);
 650         if (ret)
 651                 goto out_recv;
 652
 653         rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 654
 655         goto out;
 656
 657 out_recv:
 658         rds_tcp_recv_exit();
 659 out_pernet:
 660         unregister_pernet_subsys(&rds_tcp_net_ops);
 661 out_slab:
 662         kmem_cache_destroy(rds_tcp_conn_slab);
 663 out:
 664         return ret;
 665 }
 666 module_init(rds_tcp_init);
 667
 668 MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
 669 MODULE_DESCRIPTION("RDS: TCP transport");
 670 MODULE_LICENSE("Dual BSD/GPL");
 671