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         int val = 1;
  88
  89         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (void *)&val,
  90                               sizeof(val));
  91 }
  92
  93 u32 rds_tcp_snd_nxt(struct rds_tcp_connection *tc)
  94 {
  95         return tcp_sk(tc->t_sock->sk)->snd_nxt;
  96 }
  97
  98 u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
  99 {
 100         return tcp_sk(tc->t_sock->sk)->snd_una;
 101 }
 102
 103 void rds_tcp_restore_callbacks(struct socket *sock,
 104                                struct rds_tcp_connection *tc)
 105 {
 106         rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
 107         write_lock_bh(&sock->sk->sk_callback_lock);
 108
 109         /* done under the callback_lock to serialize with write_space */
 110         spin_lock(&rds_tcp_tc_list_lock);
 111         list_del_init(&tc->t_list_item);
 112         rds_tcp_tc_count--;
 113         spin_unlock(&rds_tcp_tc_list_lock);
 114
 115         tc->t_sock = NULL;
 116
 117         sock->sk->sk_write_space = tc->t_orig_write_space;
 118         sock->sk->sk_data_ready = tc->t_orig_data_ready;
 119         sock->sk->sk_state_change = tc->t_orig_state_change;
 120         sock->sk->sk_user_data = NULL;
 121
 122         write_unlock_bh(&sock->sk->sk_callback_lock);
 123 }
 124
 125 /*
 126  * rds_tcp_reset_callbacks() switches the to the new sock and
 127  * returns the existing tc->t_sock.
 128  *
 129  * The only functions that set tc->t_sock are rds_tcp_set_callbacks
 130  * and rds_tcp_reset_callbacks.  Send and receive trust that
 131  * it is set.  The absence of RDS_CONN_UP bit protects those paths
 132  * from being called while it isn't set.
 133  */
 134 void rds_tcp_reset_callbacks(struct socket *sock,
 135                              struct rds_conn_path *cp)
 136 {
 137         struct rds_tcp_connection *tc = cp->cp_transport_data;
 138         struct socket *osock = tc->t_sock;
 139
 140         if (!osock)
 141                 goto newsock;
 142
 143         /* Need to resolve a duelling SYN between peers.
 144          * We have an outstanding SYN to this peer, which may
 145          * potentially have transitioned to the RDS_CONN_UP state,
 146          * so we must quiesce any send threads before resetting
 147          * cp_transport_data. We quiesce these threads by setting
 148          * cp_state to something other than RDS_CONN_UP, and then
 149          * waiting for any existing threads in rds_send_xmit to
 150          * complete release_in_xmit(). (Subsequent threads entering
 151          * rds_send_xmit() will bail on !rds_conn_up().
 152          *
 153          * However an incoming syn-ack at this point would end up
 154          * marking the conn as RDS_CONN_UP, and would again permit
 155          * rds_send_xmi() threads through, so ideally we would
 156          * synchronize on RDS_CONN_UP after lock_sock(), but cannot
 157          * do that: waiting on !RDS_IN_XMIT after lock_sock() may
 158          * end up deadlocking with tcp_sendmsg(), and the RDS_IN_XMIT
 159          * would not get set. As a result, we set c_state to
 160          * RDS_CONN_RESETTTING, to ensure that rds_tcp_state_change
 161          * cannot mark rds_conn_path_up() in the window before lock_sock()
 162          */
 163         atomic_set(&cp->cp_state, RDS_CONN_RESETTING);
 164         wait_event(cp->cp_waitq, !test_bit(RDS_IN_XMIT, &cp->cp_flags));
 165         lock_sock(osock->sk);
 166         /* reset receive side state for rds_tcp_data_recv() for osock  */
 167         cancel_delayed_work_sync(&cp->cp_send_w);
 168         cancel_delayed_work_sync(&cp->cp_recv_w);
 169         if (tc->t_tinc) {
 170                 rds_inc_put(&tc->t_tinc->ti_inc);
 171                 tc->t_tinc = NULL;
 172         }
 173         tc->t_tinc_hdr_rem = sizeof(struct rds_header);
 174         tc->t_tinc_data_rem = 0;
 175         rds_tcp_restore_callbacks(osock, tc);
 176         release_sock(osock->sk);
 177         sock_release(osock);
 178 newsock:
 179         rds_send_path_reset(cp);
 180         lock_sock(sock->sk);
 181         rds_tcp_set_callbacks(sock, cp);
 182         release_sock(sock->sk);
 183 }
 184
 185 /* Add tc to rds_tcp_tc_list and set tc->t_sock. See comments
 186  * above rds_tcp_reset_callbacks for notes about synchronization
 187  * with data path
 188  */
 189 void rds_tcp_set_callbacks(struct socket *sock, struct rds_conn_path *cp)
 190 {
 191         struct rds_tcp_connection *tc = cp->cp_transport_data;
 192
 193         rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
 194         write_lock_bh(&sock->sk->sk_callback_lock);
 195
 196         /* done under the callback_lock to serialize with write_space */
 197         spin_lock(&rds_tcp_tc_list_lock);
 198         list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
 199         rds_tcp_tc_count++;
 200         spin_unlock(&rds_tcp_tc_list_lock);
 201
 202         /* accepted sockets need our listen data ready undone */
 203         if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
 204                 sock->sk->sk_data_ready = sock->sk->sk_user_data;
 205
 206         tc->t_sock = sock;
 207         tc->t_cpath = cp;
 208         tc->t_orig_data_ready = sock->sk->sk_data_ready;
 209         tc->t_orig_write_space = sock->sk->sk_write_space;
 210         tc->t_orig_state_change = sock->sk->sk_state_change;
 211
 212         sock->sk->sk_user_data = cp;
 213         sock->sk->sk_data_ready = rds_tcp_data_ready;
 214         sock->sk->sk_write_space = rds_tcp_write_space;
 215         sock->sk->sk_state_change = rds_tcp_state_change;
 216
 217         write_unlock_bh(&sock->sk->sk_callback_lock);
 218 }
 219
 220 static void rds_tcp_tc_info(struct socket *rds_sock, unsigned int len,
 221                             struct rds_info_iterator *iter,
 222                             struct rds_info_lengths *lens)
 223 {
 224         struct rds_info_tcp_socket tsinfo;
 225         struct rds_tcp_connection *tc;
 226         unsigned long flags;
 227         struct sockaddr_in sin;
 228         int sinlen;
 229         struct socket *sock;
 230
 231         spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
 232
 233         if (len / sizeof(tsinfo) < rds_tcp_tc_count)
 234                 goto out;
 235
 236         list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
 237
 238                 sock = tc->t_sock;
 239                 if (sock) {
 240                         sock->ops->getname(sock, (struct sockaddr *)&sin,
 241                                            &sinlen, 0);
 242                         tsinfo.local_addr = sin.sin_addr.s_addr;
 243                         tsinfo.local_port = sin.sin_port;
 244                         sock->ops->getname(sock, (struct sockaddr *)&sin,
 245                                            &sinlen, 1);
 246                         tsinfo.peer_addr = sin.sin_addr.s_addr;
 247                         tsinfo.peer_port = sin.sin_port;
 248                 }
 249
 250                 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
 251                 tsinfo.data_rem = tc->t_tinc_data_rem;
 252                 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
 253                 tsinfo.last_expected_una = tc->t_last_expected_una;
 254                 tsinfo.last_seen_una = tc->t_last_seen_una;
 255
 256                 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
 257         }
 258
 259 out:
 260         lens->nr = rds_tcp_tc_count;
 261         lens->each = sizeof(tsinfo);
 262
 263         spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
 264 }
 265
 266 static int rds_tcp_laddr_check(struct net *net, __be32 addr)
 267 {
 268         if (inet_addr_type(net, addr) == RTN_LOCAL)
 269                 return 0;
 270         return -EADDRNOTAVAIL;
 271 }
 272
 273 static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 274 {
 275         struct rds_tcp_connection *tc;
 276         int i;
 277
 278         for (i = 0; i < RDS_MPATH_WORKERS; i++) {
 279                 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
 280                 if (!tc)
 281                         return -ENOMEM;
 282
 283                 mutex_init(&tc->t_conn_path_lock);
 284                 tc->t_sock = NULL;
 285                 tc->t_tinc = NULL;
 286                 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
 287                 tc->t_tinc_data_rem = 0;
 288
 289                 conn->c_path[i].cp_transport_data = tc;
 290                 tc->t_cpath = &conn->c_path[i];
 291
 292                 spin_lock_irq(&rds_tcp_conn_lock);
 293                 list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
 294                 spin_unlock_irq(&rds_tcp_conn_lock);
 295                 rdsdebug("rds_conn_path [%d] tc %p\n", i,
 296                          conn->c_path[i].cp_transport_data);
 297         }
 298
 299         return 0;
 300 }
 301
 302 static void rds_tcp_conn_free(void *arg)
 303 {
 304         struct rds_tcp_connection *tc = arg;
 305         unsigned long flags;
 306         rdsdebug("freeing tc %p\n", tc);
 307
 308         spin_lock_irqsave(&rds_tcp_conn_lock, flags);
 309         list_del(&tc->t_tcp_node);
 310         spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
 311
 312         kmem_cache_free(rds_tcp_conn_slab, tc);
 313 }
 314
 315 static bool list_has_conn(struct list_head *list, struct rds_connection *conn)
 316 {
 317         struct rds_tcp_connection *tc, *_tc;
 318
 319         list_for_each_entry_safe(tc, _tc, list, t_tcp_node) {
 320                 if (tc->t_cpath->cp_conn == conn)
 321                         return true;
 322         }
 323         return false;
 324 }
 325
 326 static void rds_tcp_destroy_conns(void)
 327 {
 328         struct rds_tcp_connection *tc, *_tc;
 329         LIST_HEAD(tmp_list);
 330
 331         /* avoid calling conn_destroy with irqs off */
 332         spin_lock_irq(&rds_tcp_conn_lock);
 333         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 334                 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn))
 335                         list_move_tail(&tc->t_tcp_node, &tmp_list);
 336         }
 337         spin_unlock_irq(&rds_tcp_conn_lock);
 338
 339         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
 340                 rds_conn_destroy(tc->t_cpath->cp_conn);
 341 }
 342
 343 static void rds_tcp_exit(void);
 344
 345 struct rds_transport rds_tcp_transport = {
 346         .laddr_check            = rds_tcp_laddr_check,
 347         .xmit_path_prepare      = rds_tcp_xmit_path_prepare,
 348         .xmit_path_complete     = rds_tcp_xmit_path_complete,
 349         .xmit                   = rds_tcp_xmit,
 350         .recv_path              = rds_tcp_recv_path,
 351         .conn_alloc             = rds_tcp_conn_alloc,
 352         .conn_free              = rds_tcp_conn_free,
 353         .conn_path_connect      = rds_tcp_conn_path_connect,
 354         .conn_path_shutdown     = rds_tcp_conn_path_shutdown,
 355         .inc_copy_to_user       = rds_tcp_inc_copy_to_user,
 356         .inc_free               = rds_tcp_inc_free,
 357         .stats_info_copy        = rds_tcp_stats_info_copy,
 358         .exit                   = rds_tcp_exit,
 359         .t_owner                = THIS_MODULE,
 360         .t_name                 = "tcp",
 361         .t_type                 = RDS_TRANS_TCP,
 362         .t_prefer_loopback      = 1,
 363         .t_mp_capable           = 1,
 364 };
 365
 366 static unsigned int rds_tcp_netid;
 367
 368 /* per-network namespace private data for this module */
 369 struct rds_tcp_net {
 370         struct socket *rds_tcp_listen_sock;
 371         struct work_struct rds_tcp_accept_w;
 372         struct ctl_table_header *rds_tcp_sysctl;
 373         struct ctl_table *ctl_table;
 374         int sndbuf_size;
 375         int rcvbuf_size;
 376 };
 377
 378 /* All module specific customizations to the RDS-TCP socket should be done in
 379  * rds_tcp_tune() and applied after socket creation.
 380  */
 381 void rds_tcp_tune(struct socket *sock)
 382 {
 383         struct sock *sk = sock->sk;
 384         struct net *net = sock_net(sk);
 385         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 386
 387         rds_tcp_nonagle(sock);
 388         lock_sock(sk);
 389         if (rtn->sndbuf_size > 0) {
 390                 sk->sk_sndbuf = rtn->sndbuf_size;
 391                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
 392         }
 393         if (rtn->rcvbuf_size > 0) {
 394                 sk->sk_sndbuf = rtn->rcvbuf_size;
 395                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
 396         }
 397         release_sock(sk);
 398 }
 399
 400 static void rds_tcp_accept_worker(struct work_struct *work)
 401 {
 402         struct rds_tcp_net *rtn = container_of(work,
 403                                                struct rds_tcp_net,
 404                                                rds_tcp_accept_w);
 405
 406         while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
 407                 cond_resched();
 408 }
 409
 410 void rds_tcp_accept_work(struct sock *sk)
 411 {
 412         struct net *net = sock_net(sk);
 413         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 414
 415         queue_work(rds_wq, &rtn->rds_tcp_accept_w);
 416 }
 417
 418 static __net_init int rds_tcp_init_net(struct net *net)
 419 {
 420         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 421         struct ctl_table *tbl;
 422         int err = 0;
 423
 424         memset(rtn, 0, sizeof(*rtn));
 425
 426         /* {snd, rcv}buf_size default to 0, which implies we let the
 427          * stack pick the value, and permit auto-tuning of buffer size.
 428          */
 429         if (net == &init_net) {
 430                 tbl = rds_tcp_sysctl_table;
 431         } else {
 432                 tbl = kmemdup(rds_tcp_sysctl_table,
 433                               sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
 434                 if (!tbl) {
 435                         pr_warn("could not set allocate syctl table\n");
 436                         return -ENOMEM;
 437                 }
 438                 rtn->ctl_table = tbl;
 439         }
 440         tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
 441         tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
 442         rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
 443         if (!rtn->rds_tcp_sysctl) {
 444                 pr_warn("could not register sysctl\n");
 445                 err = -ENOMEM;
 446                 goto fail;
 447         }
 448         rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
 449         if (!rtn->rds_tcp_listen_sock) {
 450                 pr_warn("could not set up listen sock\n");
 451                 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
 452                 rtn->rds_tcp_sysctl = NULL;
 453                 err = -EAFNOSUPPORT;
 454                 goto fail;
 455         }
 456         INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
 457         return 0;
 458
 459 fail:
 460         if (net != &init_net)
 461                 kfree(tbl);
 462         return err;
 463 }
 464
 465 static void __net_exit rds_tcp_exit_net(struct net *net)
 466 {
 467         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 468
 469         if (rtn->rds_tcp_sysctl)
 470                 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
 471
 472         if (net != &init_net && rtn->ctl_table)
 473                 kfree(rtn->ctl_table);
 474
 475         /* If rds_tcp_exit_net() is called as a result of netns deletion,
 476          * the rds_tcp_kill_sock() device notifier would already have cleaned
 477          * up the listen socket, thus there is no work to do in this function.
 478          *
 479          * If rds_tcp_exit_net() is called as a result of module unload,
 480          * i.e., due to rds_tcp_exit() -> unregister_pernet_subsys(), then
 481          * we do need to clean up the listen socket here.
 482          */
 483         if (rtn->rds_tcp_listen_sock) {
 484                 rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 485                 rtn->rds_tcp_listen_sock = NULL;
 486                 flush_work(&rtn->rds_tcp_accept_w);
 487         }
 488 }
 489
 490 static struct pernet_operations rds_tcp_net_ops = {
 491         .init = rds_tcp_init_net,
 492         .exit = rds_tcp_exit_net,
 493         .id = &rds_tcp_netid,
 494         .size = sizeof(struct rds_tcp_net),
 495 };
 496
 497 /* explicitly send a RST on each socket, thereby releasing any socket refcnts
 498  * that may otherwise hold up netns deletion.
 499  */
 500 static void rds_tcp_conn_paths_destroy(struct rds_connection *conn)
 501 {
 502         struct rds_conn_path *cp;
 503         struct rds_tcp_connection *tc;
 504         int i;
 505         struct sock *sk;
 506
 507         for (i = 0; i < RDS_MPATH_WORKERS; i++) {
 508                 cp = &conn->c_path[i];
 509                 tc = cp->cp_transport_data;
 510                 if (!tc->t_sock)
 511                         continue;
 512                 sk = tc->t_sock->sk;
 513                 sk->sk_prot->disconnect(sk, 0);
 514                 tcp_done(sk);
 515         }
 516 }
 517
 518 static void rds_tcp_kill_sock(struct net *net)
 519 {
 520         struct rds_tcp_connection *tc, *_tc;
 521         LIST_HEAD(tmp_list);
 522         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 523
 524         rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 525         rtn->rds_tcp_listen_sock = NULL;
 526         flush_work(&rtn->rds_tcp_accept_w);
 527         spin_lock_irq(&rds_tcp_conn_lock);
 528         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 529                 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
 530
 531                 if (net != c_net || !tc->t_sock)
 532                         continue;
 533                 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn))
 534                         list_move_tail(&tc->t_tcp_node, &tmp_list);
 535         }
 536         spin_unlock_irq(&rds_tcp_conn_lock);
 537         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
 538                 rds_tcp_conn_paths_destroy(tc->t_cpath->cp_conn);
 539                 rds_conn_destroy(tc->t_cpath->cp_conn);
 540         }
 541 }
 542
 543 void *rds_tcp_listen_sock_def_readable(struct net *net)
 544 {
 545         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 546
 547         return rtn->rds_tcp_listen_sock->sk->sk_user_data;
 548 }
 549
 550 static int rds_tcp_dev_event(struct notifier_block *this,
 551                              unsigned long event, void *ptr)
 552 {
 553         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 554
 555         /* rds-tcp registers as a pernet subys, so the ->exit will only
 556          * get invoked after network acitivity has quiesced. We need to
 557          * clean up all sockets  to quiesce network activity, and use
 558          * the unregistration of the per-net loopback device as a trigger
 559          * to start that cleanup.
 560          */
 561         if (event == NETDEV_UNREGISTER_FINAL &&
 562             dev->ifindex == LOOPBACK_IFINDEX)
 563                 rds_tcp_kill_sock(dev_net(dev));
 564
 565         return NOTIFY_DONE;
 566 }
 567
 568 static struct notifier_block rds_tcp_dev_notifier = {
 569         .notifier_call        = rds_tcp_dev_event,
 570         .priority = -10, /* must be called after other network notifiers */
 571 };
 572
 573 /* when sysctl is used to modify some kernel socket parameters,this
 574  * function  resets the RDS connections in that netns  so that we can
 575  * restart with new parameters.  The assumption is that such reset
 576  * events are few and far-between.
 577  */
 578 static void rds_tcp_sysctl_reset(struct net *net)
 579 {
 580         struct rds_tcp_connection *tc, *_tc;
 581
 582         spin_lock_irq(&rds_tcp_conn_lock);
 583         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 584                 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
 585
 586                 if (net != c_net || !tc->t_sock)
 587                         continue;
 588
 589                 /* reconnect with new parameters */
 590                 rds_conn_path_drop(tc->t_cpath);
 591         }
 592         spin_unlock_irq(&rds_tcp_conn_lock);
 593 }
 594
 595 static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
 596                                  void __user *buffer, size_t *lenp,
 597                                  loff_t *fpos)
 598 {
 599         struct net *net = current->nsproxy->net_ns;
 600         int err;
 601
 602         err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
 603         if (err < 0) {
 604                 pr_warn("Invalid input. Must be >= %d\n",
 605                         *(int *)(ctl->extra1));
 606                 return err;
 607         }
 608         if (write)
 609                 rds_tcp_sysctl_reset(net);
 610         return 0;
 611 }
 612
 613 static void rds_tcp_exit(void)
 614 {
 615         rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 616         unregister_pernet_subsys(&rds_tcp_net_ops);
 617         if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
 618                 pr_warn("could not unregister rds_tcp_dev_notifier\n");
 619         rds_tcp_destroy_conns();
 620         rds_trans_unregister(&rds_tcp_transport);
 621         rds_tcp_recv_exit();
 622         kmem_cache_destroy(rds_tcp_conn_slab);
 623 }
 624 module_exit(rds_tcp_exit);
 625
 626 static int rds_tcp_init(void)
 627 {
 628         int ret;
 629
 630         rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
 631                                               sizeof(struct rds_tcp_connection),
 632                                               0, 0, NULL);
 633         if (!rds_tcp_conn_slab) {
 634                 ret = -ENOMEM;
 635                 goto out;
 636         }
 637
 638         ret = register_netdevice_notifier(&rds_tcp_dev_notifier);
 639         if (ret) {
 640                 pr_warn("could not register rds_tcp_dev_notifier\n");
 641                 goto out_slab;
 642         }
 643
 644         ret = register_pernet_subsys(&rds_tcp_net_ops);
 645         if (ret)
 646                 goto out_notifier;
 647
 648         ret = rds_tcp_recv_init();
 649         if (ret)
 650                 goto out_pernet;
 651
 652         rds_trans_register(&rds_tcp_transport);
 653
 654         rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 655
 656         goto out;
 657
 658 out_pernet:
 659         unregister_pernet_subsys(&rds_tcp_net_ops);
 660 out_notifier:
 661         if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
 662                 pr_warn("could not unregister rds_tcp_dev_notifier\n");
 663 out_slab:
 664         kmem_cache_destroy(rds_tcp_conn_slab);
 665 out:
 666         return ret;
 667 }
 668 module_init(rds_tcp_init);
 669
 670 MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
 671 MODULE_DESCRIPTION("RDS: TCP transport");
 672 MODULE_LICENSE("Dual BSD/GPL");
 673