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 static atomic_t rds_tcp_unloading = ATOMIC_INIT(0);
  53
  54 static struct kmem_cache *rds_tcp_conn_slab;
  55
  56 static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
  57                                  void __user *buffer, size_t *lenp,
  58                                  loff_t *fpos);
  59
  60 static int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF;
  61 static int rds_tcp_min_rcvbuf = SOCK_MIN_RCVBUF;
  62
  63 static struct ctl_table rds_tcp_sysctl_table[] = {
  64 #define RDS_TCP_SNDBUF  0
  65         {
  66                 .procname       = "rds_tcp_sndbuf",
  67                 /* data is per-net pointer */
  68                 .maxlen         = sizeof(int),
  69                 .mode           = 0644,
  70                 .proc_handler   = rds_tcp_skbuf_handler,
  71                 .extra1         = &rds_tcp_min_sndbuf,
  72         },
  73 #define RDS_TCP_RCVBUF  1
  74         {
  75                 .procname       = "rds_tcp_rcvbuf",
  76                 /* data is per-net pointer */
  77                 .maxlen         = sizeof(int),
  78                 .mode           = 0644,
  79                 .proc_handler   = rds_tcp_skbuf_handler,
  80                 .extra1         = &rds_tcp_min_rcvbuf,
  81         },
  82         { }
  83 };
  84
  85 /* doing it this way avoids calling tcp_sk() */
  86 void rds_tcp_nonagle(struct socket *sock)
  87 {
  88         int val = 1;
  89
  90         kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (void *)&val,
  91                               sizeof(val));
  92 }
  93
  94 u32 rds_tcp_write_seq(struct rds_tcp_connection *tc)
  95 {
  96         /* seq# of the last byte of data in tcp send buffer */
  97         return tcp_sk(tc->t_sock->sk)->write_seq;
  98 }
  99
 100 u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
 101 {
 102         return tcp_sk(tc->t_sock->sk)->snd_una;
 103 }
 104
 105 void rds_tcp_restore_callbacks(struct socket *sock,
 106                                struct rds_tcp_connection *tc)
 107 {
 108         rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
 109         write_lock_bh(&sock->sk->sk_callback_lock);
 110
 111         /* done under the callback_lock to serialize with write_space */
 112         spin_lock(&rds_tcp_tc_list_lock);
 113         list_del_init(&tc->t_list_item);
 114         rds_tcp_tc_count--;
 115         spin_unlock(&rds_tcp_tc_list_lock);
 116
 117         tc->t_sock = NULL;
 118
 119         sock->sk->sk_write_space = tc->t_orig_write_space;
 120         sock->sk->sk_data_ready = tc->t_orig_data_ready;
 121         sock->sk->sk_state_change = tc->t_orig_state_change;
 122         sock->sk->sk_user_data = NULL;
 123
 124         write_unlock_bh(&sock->sk->sk_callback_lock);
 125 }
 126
 127 /*
 128  * rds_tcp_reset_callbacks() switches the to the new sock and
 129  * returns the existing tc->t_sock.
 130  *
 131  * The only functions that set tc->t_sock are rds_tcp_set_callbacks
 132  * and rds_tcp_reset_callbacks.  Send and receive trust that
 133  * it is set.  The absence of RDS_CONN_UP bit protects those paths
 134  * from being called while it isn't set.
 135  */
 136 void rds_tcp_reset_callbacks(struct socket *sock,
 137                              struct rds_conn_path *cp)
 138 {
 139         struct rds_tcp_connection *tc = cp->cp_transport_data;
 140         struct socket *osock = tc->t_sock;
 141
 142         if (!osock)
 143                 goto newsock;
 144
 145         /* Need to resolve a duelling SYN between peers.
 146          * We have an outstanding SYN to this peer, which may
 147          * potentially have transitioned to the RDS_CONN_UP state,
 148          * so we must quiesce any send threads before resetting
 149          * cp_transport_data. We quiesce these threads by setting
 150          * cp_state to something other than RDS_CONN_UP, and then
 151          * waiting for any existing threads in rds_send_xmit to
 152          * complete release_in_xmit(). (Subsequent threads entering
 153          * rds_send_xmit() will bail on !rds_conn_up().
 154          *
 155          * However an incoming syn-ack at this point would end up
 156          * marking the conn as RDS_CONN_UP, and would again permit
 157          * rds_send_xmi() threads through, so ideally we would
 158          * synchronize on RDS_CONN_UP after lock_sock(), but cannot
 159          * do that: waiting on !RDS_IN_XMIT after lock_sock() may
 160          * end up deadlocking with tcp_sendmsg(), and the RDS_IN_XMIT
 161          * would not get set. As a result, we set c_state to
 162          * RDS_CONN_RESETTTING, to ensure that rds_tcp_state_change
 163          * cannot mark rds_conn_path_up() in the window before lock_sock()
 164          */
 165         atomic_set(&cp->cp_state, RDS_CONN_RESETTING);
 166         wait_event(cp->cp_waitq, !test_bit(RDS_IN_XMIT, &cp->cp_flags));
 167         lock_sock(osock->sk);
 168         /* reset receive side state for rds_tcp_data_recv() for osock  */
 169         cancel_delayed_work_sync(&cp->cp_send_w);
 170         cancel_delayed_work_sync(&cp->cp_recv_w);
 171         if (tc->t_tinc) {
 172                 rds_inc_put(&tc->t_tinc->ti_inc);
 173                 tc->t_tinc = NULL;
 174         }
 175         tc->t_tinc_hdr_rem = sizeof(struct rds_header);
 176         tc->t_tinc_data_rem = 0;
 177         rds_tcp_restore_callbacks(osock, tc);
 178         release_sock(osock->sk);
 179         sock_release(osock);
 180 newsock:
 181         rds_send_path_reset(cp);
 182         lock_sock(sock->sk);
 183         rds_tcp_set_callbacks(sock, cp);
 184         release_sock(sock->sk);
 185 }
 186
 187 /* Add tc to rds_tcp_tc_list and set tc->t_sock. See comments
 188  * above rds_tcp_reset_callbacks for notes about synchronization
 189  * with data path
 190  */
 191 void rds_tcp_set_callbacks(struct socket *sock, struct rds_conn_path *cp)
 192 {
 193         struct rds_tcp_connection *tc = cp->cp_transport_data;
 194
 195         rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
 196         write_lock_bh(&sock->sk->sk_callback_lock);
 197
 198         /* done under the callback_lock to serialize with write_space */
 199         spin_lock(&rds_tcp_tc_list_lock);
 200         list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
 201         rds_tcp_tc_count++;
 202         spin_unlock(&rds_tcp_tc_list_lock);
 203
 204         /* accepted sockets need our listen data ready undone */
 205         if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
 206                 sock->sk->sk_data_ready = sock->sk->sk_user_data;
 207
 208         tc->t_sock = sock;
 209         tc->t_cpath = cp;
 210         tc->t_orig_data_ready = sock->sk->sk_data_ready;
 211         tc->t_orig_write_space = sock->sk->sk_write_space;
 212         tc->t_orig_state_change = sock->sk->sk_state_change;
 213
 214         sock->sk->sk_user_data = cp;
 215         sock->sk->sk_data_ready = rds_tcp_data_ready;
 216         sock->sk->sk_write_space = rds_tcp_write_space;
 217         sock->sk->sk_state_change = rds_tcp_state_change;
 218
 219         write_unlock_bh(&sock->sk->sk_callback_lock);
 220 }
 221
 222 static void rds_tcp_tc_info(struct socket *rds_sock, unsigned int len,
 223                             struct rds_info_iterator *iter,
 224                             struct rds_info_lengths *lens)
 225 {
 226         struct rds_info_tcp_socket tsinfo;
 227         struct rds_tcp_connection *tc;
 228         unsigned long flags;
 229         struct sockaddr_in sin;
 230         struct socket *sock;
 231
 232         spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
 233
 234         if (len / sizeof(tsinfo) < rds_tcp_tc_count)
 235                 goto out;
 236
 237         list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
 238
 239                 sock = tc->t_sock;
 240                 if (sock) {
 241                         sock->ops->getname(sock, (struct sockaddr *)&sin, 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, 1);
 245                         tsinfo.peer_addr = sin.sin_addr.s_addr;
 246                         tsinfo.peer_port = sin.sin_port;
 247                 }
 248
 249                 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
 250                 tsinfo.data_rem = tc->t_tinc_data_rem;
 251                 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
 252                 tsinfo.last_expected_una = tc->t_last_expected_una;
 253                 tsinfo.last_seen_una = tc->t_last_seen_una;
 254
 255                 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
 256         }
 257
 258 out:
 259         lens->nr = rds_tcp_tc_count;
 260         lens->each = sizeof(tsinfo);
 261
 262         spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
 263 }
 264
 265 static int rds_tcp_laddr_check(struct net *net, __be32 addr)
 266 {
 267         if (inet_addr_type(net, addr) == RTN_LOCAL)
 268                 return 0;
 269         return -EADDRNOTAVAIL;
 270 }
 271
 272 static void rds_tcp_conn_free(void *arg)
 273 {
 274         struct rds_tcp_connection *tc = arg;
 275
 276         rdsdebug("freeing tc %p\n", tc);
 277
 278         spin_lock_bh(&rds_tcp_conn_lock);
 279         if (!tc->t_tcp_node_detached)
 280                 list_del(&tc->t_tcp_node);
 281         spin_unlock_bh(&rds_tcp_conn_lock);
 282
 283         kmem_cache_free(rds_tcp_conn_slab, tc);
 284 }
 285
 286 static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 287 {
 288         struct rds_tcp_connection *tc;
 289         int i, j;
 290         int ret = 0;
 291
 292         for (i = 0; i < RDS_MPATH_WORKERS; i++) {
 293                 tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
 294                 if (!tc) {
 295                         ret = -ENOMEM;
 296                         goto fail;
 297                 }
 298                 mutex_init(&tc->t_conn_path_lock);
 299                 tc->t_sock = NULL;
 300                 tc->t_tinc = NULL;
 301                 tc->t_tinc_hdr_rem = sizeof(struct rds_header);
 302                 tc->t_tinc_data_rem = 0;
 303
 304                 conn->c_path[i].cp_transport_data = tc;
 305                 tc->t_cpath = &conn->c_path[i];
 306                 tc->t_tcp_node_detached = true;
 307
 308                 rdsdebug("rds_conn_path [%d] tc %p\n", i,
 309                          conn->c_path[i].cp_transport_data);
 310         }
 311         spin_lock_bh(&rds_tcp_conn_lock);
 312         for (i = 0; i < RDS_MPATH_WORKERS; i++) {
 313                 tc = conn->c_path[i].cp_transport_data;
 314                 tc->t_tcp_node_detached = false;
 315                 list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
 316         }
 317         spin_unlock_bh(&rds_tcp_conn_lock);
 318 fail:
 319         if (ret) {
 320                 for (j = 0; j < i; j++)
 321                         rds_tcp_conn_free(conn->c_path[j].cp_transport_data);
 322         }
 323         return ret;
 324 }
 325
 326 static bool list_has_conn(struct list_head *list, struct rds_connection *conn)
 327 {
 328         struct rds_tcp_connection *tc, *_tc;
 329
 330         list_for_each_entry_safe(tc, _tc, list, t_tcp_node) {
 331                 if (tc->t_cpath->cp_conn == conn)
 332                         return true;
 333         }
 334         return false;
 335 }
 336
 337 static void rds_tcp_set_unloading(void)
 338 {
 339         atomic_set(&rds_tcp_unloading, 1);
 340 }
 341
 342 static bool rds_tcp_is_unloading(struct rds_connection *conn)
 343 {
 344         return atomic_read(&rds_tcp_unloading) != 0;
 345 }
 346
 347 static void rds_tcp_destroy_conns(void)
 348 {
 349         struct rds_tcp_connection *tc, *_tc;
 350         LIST_HEAD(tmp_list);
 351
 352         /* avoid calling conn_destroy with irqs off */
 353         spin_lock_irq(&rds_tcp_conn_lock);
 354         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 355                 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn))
 356                         list_move_tail(&tc->t_tcp_node, &tmp_list);
 357         }
 358         spin_unlock_irq(&rds_tcp_conn_lock);
 359
 360         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
 361                 rds_conn_destroy(tc->t_cpath->cp_conn);
 362 }
 363
 364 static void rds_tcp_exit(void);
 365
 366 struct rds_transport rds_tcp_transport = {
 367         .laddr_check            = rds_tcp_laddr_check,
 368         .xmit_path_prepare      = rds_tcp_xmit_path_prepare,
 369         .xmit_path_complete     = rds_tcp_xmit_path_complete,
 370         .xmit                   = rds_tcp_xmit,
 371         .recv_path              = rds_tcp_recv_path,
 372         .conn_alloc             = rds_tcp_conn_alloc,
 373         .conn_free              = rds_tcp_conn_free,
 374         .conn_path_connect      = rds_tcp_conn_path_connect,
 375         .conn_path_shutdown     = rds_tcp_conn_path_shutdown,
 376         .inc_copy_to_user       = rds_tcp_inc_copy_to_user,
 377         .inc_free               = rds_tcp_inc_free,
 378         .stats_info_copy        = rds_tcp_stats_info_copy,
 379         .exit                   = rds_tcp_exit,
 380         .t_owner                = THIS_MODULE,
 381         .t_name                 = "tcp",
 382         .t_type                 = RDS_TRANS_TCP,
 383         .t_prefer_loopback      = 1,
 384         .t_mp_capable           = 1,
 385         .t_unloading            = rds_tcp_is_unloading,
 386 };
 387
 388 static unsigned int rds_tcp_netid;
 389
 390 /* per-network namespace private data for this module */
 391 struct rds_tcp_net {
 392         struct socket *rds_tcp_listen_sock;
 393         struct work_struct rds_tcp_accept_w;
 394         struct ctl_table_header *rds_tcp_sysctl;
 395         struct ctl_table *ctl_table;
 396         int sndbuf_size;
 397         int rcvbuf_size;
 398 };
 399
 400 /* All module specific customizations to the RDS-TCP socket should be done in
 401  * rds_tcp_tune() and applied after socket creation.
 402  */
 403 void rds_tcp_tune(struct socket *sock)
 404 {
 405         struct sock *sk = sock->sk;
 406         struct net *net = sock_net(sk);
 407         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 408
 409         rds_tcp_nonagle(sock);
 410         lock_sock(sk);
 411         if (rtn->sndbuf_size > 0) {
 412                 sk->sk_sndbuf = rtn->sndbuf_size;
 413                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
 414         }
 415         if (rtn->rcvbuf_size > 0) {
 416                 sk->sk_sndbuf = rtn->rcvbuf_size;
 417                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
 418         }
 419         release_sock(sk);
 420 }
 421
 422 static void rds_tcp_accept_worker(struct work_struct *work)
 423 {
 424         struct rds_tcp_net *rtn = container_of(work,
 425                                                struct rds_tcp_net,
 426                                                rds_tcp_accept_w);
 427
 428         while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
 429                 cond_resched();
 430 }
 431
 432 void rds_tcp_accept_work(struct sock *sk)
 433 {
 434         struct net *net = sock_net(sk);
 435         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 436
 437         queue_work(rds_wq, &rtn->rds_tcp_accept_w);
 438 }
 439
 440 static __net_init int rds_tcp_init_net(struct net *net)
 441 {
 442         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 443         struct ctl_table *tbl;
 444         int err = 0;
 445
 446         memset(rtn, 0, sizeof(*rtn));
 447
 448         /* {snd, rcv}buf_size default to 0, which implies we let the
 449          * stack pick the value, and permit auto-tuning of buffer size.
 450          */
 451         if (net == &init_net) {
 452                 tbl = rds_tcp_sysctl_table;
 453         } else {
 454                 tbl = kmemdup(rds_tcp_sysctl_table,
 455                               sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
 456                 if (!tbl) {
 457                         pr_warn("could not set allocate syctl table\n");
 458                         return -ENOMEM;
 459                 }
 460                 rtn->ctl_table = tbl;
 461         }
 462         tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
 463         tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
 464         rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
 465         if (!rtn->rds_tcp_sysctl) {
 466                 pr_warn("could not register sysctl\n");
 467                 err = -ENOMEM;
 468                 goto fail;
 469         }
 470         rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
 471         if (!rtn->rds_tcp_listen_sock) {
 472                 pr_warn("could not set up listen sock\n");
 473                 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
 474                 rtn->rds_tcp_sysctl = NULL;
 475                 err = -EAFNOSUPPORT;
 476                 goto fail;
 477         }
 478         INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
 479         return 0;
 480
 481 fail:
 482         if (net != &init_net)
 483                 kfree(tbl);
 484         return err;
 485 }
 486
 487 static void __net_exit rds_tcp_exit_net(struct net *net)
 488 {
 489         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 490
 491         if (rtn->rds_tcp_sysctl)
 492                 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
 493
 494         if (net != &init_net && rtn->ctl_table)
 495                 kfree(rtn->ctl_table);
 496
 497         /* If rds_tcp_exit_net() is called as a result of netns deletion,
 498          * the rds_tcp_kill_sock() device notifier would already have cleaned
 499          * up the listen socket, thus there is no work to do in this function.
 500          *
 501          * If rds_tcp_exit_net() is called as a result of module unload,
 502          * i.e., due to rds_tcp_exit() -> unregister_pernet_subsys(), then
 503          * we do need to clean up the listen socket here.
 504          */
 505         if (rtn->rds_tcp_listen_sock) {
 506                 struct socket *lsock = rtn->rds_tcp_listen_sock;
 507
 508                 rtn->rds_tcp_listen_sock = NULL;
 509                 rds_tcp_listen_stop(lsock, &rtn->rds_tcp_accept_w);
 510         }
 511 }
 512
 513 static struct pernet_operations rds_tcp_net_ops = {
 514         .init = rds_tcp_init_net,
 515         .exit = rds_tcp_exit_net,
 516         .id = &rds_tcp_netid,
 517         .size = sizeof(struct rds_tcp_net),
 518 };
 519
 520 static void rds_tcp_kill_sock(struct net *net)
 521 {
 522         struct rds_tcp_connection *tc, *_tc;
 523         LIST_HEAD(tmp_list);
 524         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 525         struct socket *lsock = rtn->rds_tcp_listen_sock;
 526
 527         rtn->rds_tcp_listen_sock = NULL;
 528         rds_tcp_listen_stop(lsock, &rtn->rds_tcp_accept_w);
 529         spin_lock_bh(&rds_tcp_conn_lock);
 530         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 531                 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
 532
 533                 if (net != c_net || !tc->t_sock)
 534                         continue;
 535                 if (!list_has_conn(&tmp_list, tc->t_cpath->cp_conn)) {
 536                         list_move_tail(&tc->t_tcp_node, &tmp_list);
 537                 } else {
 538                         list_del(&tc->t_tcp_node);
 539                         tc->t_tcp_node_detached = true;
 540                 }
 541         }
 542         spin_unlock_bh(&rds_tcp_conn_lock);
 543         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node)
 544                 rds_conn_destroy(tc->t_cpath->cp_conn);
 545 }
 546
 547 void *rds_tcp_listen_sock_def_readable(struct net *net)
 548 {
 549         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 550         struct socket *lsock = rtn->rds_tcp_listen_sock;
 551
 552         if (!lsock)
 553                 return NULL;
 554
 555         return lsock->sk->sk_user_data;
 556 }
 557
 558 static int rds_tcp_dev_event(struct notifier_block *this,
 559                              unsigned long event, void *ptr)
 560 {
 561         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 562
 563         /* rds-tcp registers as a pernet subys, so the ->exit will only
 564          * get invoked after network acitivity has quiesced. We need to
 565          * clean up all sockets  to quiesce network activity, and use
 566          * the unregistration of the per-net loopback device as a trigger
 567          * to start that cleanup.
 568          */
 569         if (event == NETDEV_UNREGISTER_FINAL &&
 570             dev->ifindex == LOOPBACK_IFINDEX)
 571                 rds_tcp_kill_sock(dev_net(dev));
 572
 573         return NOTIFY_DONE;
 574 }
 575
 576 static struct notifier_block rds_tcp_dev_notifier = {
 577         .notifier_call        = rds_tcp_dev_event,
 578         .priority = -10, /* must be called after other network notifiers */
 579 };
 580
 581 /* when sysctl is used to modify some kernel socket parameters,this
 582  * function  resets the RDS connections in that netns  so that we can
 583  * restart with new parameters.  The assumption is that such reset
 584  * events are few and far-between.
 585  */
 586 static void rds_tcp_sysctl_reset(struct net *net)
 587 {
 588         struct rds_tcp_connection *tc, *_tc;
 589
 590         spin_lock_bh(&rds_tcp_conn_lock);
 591         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 592                 struct net *c_net = read_pnet(&tc->t_cpath->cp_conn->c_net);
 593
 594                 if (net != c_net || !tc->t_sock)
 595                         continue;
 596
 597                 /* reconnect with new parameters */
 598                 rds_conn_path_drop(tc->t_cpath, false);
 599         }
 600         spin_unlock_bh(&rds_tcp_conn_lock);
 601 }
 602
 603 static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
 604                                  void __user *buffer, size_t *lenp,
 605                                  loff_t *fpos)
 606 {
 607         struct net *net = current->nsproxy->net_ns;
 608         int err;
 609
 610         err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
 611         if (err < 0) {
 612                 pr_warn("Invalid input. Must be >= %d\n",
 613                         *(int *)(ctl->extra1));
 614                 return err;
 615         }
 616         if (write)
 617                 rds_tcp_sysctl_reset(net);
 618         return 0;
 619 }
 620
 621 static void rds_tcp_exit(void)
 622 {
 623         rds_tcp_set_unloading();
 624         synchronize_rcu();
 625         rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 626         unregister_pernet_subsys(&rds_tcp_net_ops);
 627         if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
 628                 pr_warn("could not unregister rds_tcp_dev_notifier\n");
 629         rds_tcp_destroy_conns();
 630         rds_trans_unregister(&rds_tcp_transport);
 631         rds_tcp_recv_exit();
 632         kmem_cache_destroy(rds_tcp_conn_slab);
 633 }
 634 module_exit(rds_tcp_exit);
 635
 636 static int rds_tcp_init(void)
 637 {
 638         int ret;
 639
 640         rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
 641                                               sizeof(struct rds_tcp_connection),
 642                                               0, 0, NULL);
 643         if (!rds_tcp_conn_slab) {
 644                 ret = -ENOMEM;
 645                 goto out;
 646         }
 647
 648         ret = rds_tcp_recv_init();
 649         if (ret)
 650                 goto out_slab;
 651
 652         ret = register_pernet_subsys(&rds_tcp_net_ops);
 653         if (ret)
 654                 goto out_recv;
 655
 656         ret = register_netdevice_notifier(&rds_tcp_dev_notifier);
 657         if (ret) {
 658                 pr_warn("could not register rds_tcp_dev_notifier\n");
 659                 goto out_pernet;
 660         }
 661
 662         rds_trans_register(&rds_tcp_transport);
 663
 664         rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 665
 666         goto out;
 667
 668 out_pernet:
 669         unregister_pernet_subsys(&rds_tcp_net_ops);
 670 out_recv:
 671         rds_tcp_recv_exit();
 672 out_slab:
 673         kmem_cache_destroy(rds_tcp_conn_slab);
 674 out:
 675         return ret;
 676 }
 677 module_init(rds_tcp_init);
 678
 679 MODULE_AUTHOR("Oracle Corporation <[email protected]>");
 680 MODULE_DESCRIPTION("RDS: TCP transport");
 681 MODULE_LICENSE("Dual BSD/GPL");
 682