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 int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF;
  60 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  * This is the only path that sets tc->t_sock.  Send and receive trust that
 130  * it is set.  The RDS_CONN_UP bit protects those paths from being
 131  * called while it isn't set.
 132  */
 133 void rds_tcp_set_callbacks(struct socket *sock, struct rds_connection *conn)
 134 {
 135         struct rds_tcp_connection *tc = conn->c_transport_data;
 136
 137         rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
 138         write_lock_bh(&sock->sk->sk_callback_lock);
 139
 140         /* done under the callback_lock to serialize with write_space */
 141         spin_lock(&rds_tcp_tc_list_lock);
 142         list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
 143         rds_tcp_tc_count++;
 144         spin_unlock(&rds_tcp_tc_list_lock);
 145
 146         /* accepted sockets need our listen data ready undone */
 147         if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
 148                 sock->sk->sk_data_ready = sock->sk->sk_user_data;
 149
 150         tc->t_sock = sock;
 151         tc->conn = conn;
 152         tc->t_orig_data_ready = sock->sk->sk_data_ready;
 153         tc->t_orig_write_space = sock->sk->sk_write_space;
 154         tc->t_orig_state_change = sock->sk->sk_state_change;
 155
 156         sock->sk->sk_user_data = conn;
 157         sock->sk->sk_data_ready = rds_tcp_data_ready;
 158         sock->sk->sk_write_space = rds_tcp_write_space;
 159         sock->sk->sk_state_change = rds_tcp_state_change;
 160
 161         write_unlock_bh(&sock->sk->sk_callback_lock);
 162 }
 163
 164 static void rds_tcp_tc_info(struct socket *sock, unsigned int len,
 165                             struct rds_info_iterator *iter,
 166                             struct rds_info_lengths *lens)
 167 {
 168         struct rds_info_tcp_socket tsinfo;
 169         struct rds_tcp_connection *tc;
 170         unsigned long flags;
 171         struct sockaddr_in sin;
 172         int sinlen;
 173
 174         spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
 175
 176         if (len / sizeof(tsinfo) < rds_tcp_tc_count)
 177                 goto out;
 178
 179         list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
 180
 181                 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 0);
 182                 tsinfo.local_addr = sin.sin_addr.s_addr;
 183                 tsinfo.local_port = sin.sin_port;
 184                 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 1);
 185                 tsinfo.peer_addr = sin.sin_addr.s_addr;
 186                 tsinfo.peer_port = sin.sin_port;
 187
 188                 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
 189                 tsinfo.data_rem = tc->t_tinc_data_rem;
 190                 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
 191                 tsinfo.last_expected_una = tc->t_last_expected_una;
 192                 tsinfo.last_seen_una = tc->t_last_seen_una;
 193
 194                 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
 195         }
 196
 197 out:
 198         lens->nr = rds_tcp_tc_count;
 199         lens->each = sizeof(tsinfo);
 200
 201         spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
 202 }
 203
 204 static int rds_tcp_laddr_check(struct net *net, __be32 addr)
 205 {
 206         if (inet_addr_type(net, addr) == RTN_LOCAL)
 207                 return 0;
 208         return -EADDRNOTAVAIL;
 209 }
 210
 211 static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 212 {
 213         struct rds_tcp_connection *tc;
 214
 215         tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
 216         if (!tc)
 217                 return -ENOMEM;
 218
 219         mutex_init(&tc->t_conn_lock);
 220         tc->t_sock = NULL;
 221         tc->t_tinc = NULL;
 222         tc->t_tinc_hdr_rem = sizeof(struct rds_header);
 223         tc->t_tinc_data_rem = 0;
 224
 225         conn->c_transport_data = tc;
 226
 227         spin_lock_irq(&rds_tcp_conn_lock);
 228         list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
 229         spin_unlock_irq(&rds_tcp_conn_lock);
 230
 231         rdsdebug("alloced tc %p\n", conn->c_transport_data);
 232         return 0;
 233 }
 234
 235 static void rds_tcp_conn_free(void *arg)
 236 {
 237         struct rds_tcp_connection *tc = arg;
 238         unsigned long flags;
 239         rdsdebug("freeing tc %p\n", tc);
 240
 241         spin_lock_irqsave(&rds_tcp_conn_lock, flags);
 242         list_del(&tc->t_tcp_node);
 243         spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
 244
 245         kmem_cache_free(rds_tcp_conn_slab, tc);
 246 }
 247
 248 static void rds_tcp_destroy_conns(void)
 249 {
 250         struct rds_tcp_connection *tc, *_tc;
 251         LIST_HEAD(tmp_list);
 252
 253         /* avoid calling conn_destroy with irqs off */
 254         spin_lock_irq(&rds_tcp_conn_lock);
 255         list_splice(&rds_tcp_conn_list, &tmp_list);
 256         INIT_LIST_HEAD(&rds_tcp_conn_list);
 257         spin_unlock_irq(&rds_tcp_conn_lock);
 258
 259         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
 260                 if (tc->conn->c_passive)
 261                         rds_conn_destroy(tc->conn->c_passive);
 262                 rds_conn_destroy(tc->conn);
 263         }
 264 }
 265
 266 static void rds_tcp_exit(void);
 267
 268 struct rds_transport rds_tcp_transport = {
 269         .laddr_check            = rds_tcp_laddr_check,
 270         .xmit_prepare           = rds_tcp_xmit_prepare,
 271         .xmit_complete          = rds_tcp_xmit_complete,
 272         .xmit                   = rds_tcp_xmit,
 273         .recv                   = rds_tcp_recv,
 274         .conn_alloc             = rds_tcp_conn_alloc,
 275         .conn_free              = rds_tcp_conn_free,
 276         .conn_connect           = rds_tcp_conn_connect,
 277         .conn_shutdown          = rds_tcp_conn_shutdown,
 278         .inc_copy_to_user       = rds_tcp_inc_copy_to_user,
 279         .inc_free               = rds_tcp_inc_free,
 280         .stats_info_copy        = rds_tcp_stats_info_copy,
 281         .exit                   = rds_tcp_exit,
 282         .t_owner                = THIS_MODULE,
 283         .t_name                 = "tcp",
 284         .t_type                 = RDS_TRANS_TCP,
 285         .t_prefer_loopback      = 1,
 286 };
 287
 288 static int rds_tcp_netid;
 289
 290 /* per-network namespace private data for this module */
 291 struct rds_tcp_net {
 292         struct socket *rds_tcp_listen_sock;
 293         struct work_struct rds_tcp_accept_w;
 294         struct ctl_table_header *rds_tcp_sysctl;
 295         struct ctl_table *ctl_table;
 296         int sndbuf_size;
 297         int rcvbuf_size;
 298 };
 299
 300 /* All module specific customizations to the RDS-TCP socket should be done in
 301  * rds_tcp_tune() and applied after socket creation.
 302  */
 303 void rds_tcp_tune(struct socket *sock)
 304 {
 305         struct sock *sk = sock->sk;
 306         struct net *net = sock_net(sk);
 307         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 308
 309         rds_tcp_nonagle(sock);
 310         lock_sock(sk);
 311         if (rtn->sndbuf_size > 0) {
 312                 sk->sk_sndbuf = rtn->sndbuf_size;
 313                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
 314         }
 315         if (rtn->rcvbuf_size > 0) {
 316                 sk->sk_sndbuf = rtn->rcvbuf_size;
 317                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
 318         }
 319         release_sock(sk);
 320 }
 321
 322 static void rds_tcp_accept_worker(struct work_struct *work)
 323 {
 324         struct rds_tcp_net *rtn = container_of(work,
 325                                                struct rds_tcp_net,
 326                                                rds_tcp_accept_w);
 327
 328         while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
 329                 cond_resched();
 330 }
 331
 332 void rds_tcp_accept_work(struct sock *sk)
 333 {
 334         struct net *net = sock_net(sk);
 335         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 336
 337         queue_work(rds_wq, &rtn->rds_tcp_accept_w);
 338 }
 339
 340 static __net_init int rds_tcp_init_net(struct net *net)
 341 {
 342         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 343         struct ctl_table *tbl;
 344         int err = 0;
 345
 346         memset(rtn, 0, sizeof(*rtn));
 347
 348         /* {snd, rcv}buf_size default to 0, which implies we let the
 349          * stack pick the value, and permit auto-tuning of buffer size.
 350          */
 351         if (net == &init_net) {
 352                 tbl = rds_tcp_sysctl_table;
 353         } else {
 354                 tbl = kmemdup(rds_tcp_sysctl_table,
 355                               sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
 356                 if (!tbl) {
 357                         pr_warn("could not set allocate syctl table\n");
 358                         return -ENOMEM;
 359                 }
 360                 rtn->ctl_table = tbl;
 361         }
 362         tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
 363         tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
 364         rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
 365         if (!rtn->rds_tcp_sysctl) {
 366                 pr_warn("could not register sysctl\n");
 367                 err = -ENOMEM;
 368                 goto fail;
 369         }
 370         rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
 371         if (!rtn->rds_tcp_listen_sock) {
 372                 pr_warn("could not set up listen sock\n");
 373                 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
 374                 rtn->rds_tcp_sysctl = NULL;
 375                 err = -EAFNOSUPPORT;
 376                 goto fail;
 377         }
 378         INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
 379         return 0;
 380
 381 fail:
 382         if (net != &init_net)
 383                 kfree(tbl);
 384         return err;
 385 }
 386
 387 static void __net_exit rds_tcp_exit_net(struct net *net)
 388 {
 389         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 390
 391         if (rtn->rds_tcp_sysctl)
 392                 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
 393
 394         if (net != &init_net && rtn->ctl_table)
 395                 kfree(rtn->ctl_table);
 396
 397         /* If rds_tcp_exit_net() is called as a result of netns deletion,
 398          * the rds_tcp_kill_sock() device notifier would already have cleaned
 399          * up the listen socket, thus there is no work to do in this function.
 400          *
 401          * If rds_tcp_exit_net() is called as a result of module unload,
 402          * i.e., due to rds_tcp_exit() -> unregister_pernet_subsys(), then
 403          * we do need to clean up the listen socket here.
 404          */
 405         if (rtn->rds_tcp_listen_sock) {
 406                 rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 407                 rtn->rds_tcp_listen_sock = NULL;
 408                 flush_work(&rtn->rds_tcp_accept_w);
 409         }
 410 }
 411
 412 static struct pernet_operations rds_tcp_net_ops = {
 413         .init = rds_tcp_init_net,
 414         .exit = rds_tcp_exit_net,
 415         .id = &rds_tcp_netid,
 416         .size = sizeof(struct rds_tcp_net),
 417 };
 418
 419 static void rds_tcp_kill_sock(struct net *net)
 420 {
 421         struct rds_tcp_connection *tc, *_tc;
 422         struct sock *sk;
 423         LIST_HEAD(tmp_list);
 424         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 425
 426         rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 427         rtn->rds_tcp_listen_sock = NULL;
 428         flush_work(&rtn->rds_tcp_accept_w);
 429         spin_lock_irq(&rds_tcp_conn_lock);
 430         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 431                 struct net *c_net = read_pnet(&tc->conn->c_net);
 432
 433                 if (net != c_net || !tc->t_sock)
 434                         continue;
 435                 list_move_tail(&tc->t_tcp_node, &tmp_list);
 436         }
 437         spin_unlock_irq(&rds_tcp_conn_lock);
 438         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
 439                 sk = tc->t_sock->sk;
 440                 sk->sk_prot->disconnect(sk, 0);
 441                 tcp_done(sk);
 442                 if (tc->conn->c_passive)
 443                         rds_conn_destroy(tc->conn->c_passive);
 444                 rds_conn_destroy(tc->conn);
 445         }
 446 }
 447
 448 static int rds_tcp_dev_event(struct notifier_block *this,
 449                              unsigned long event, void *ptr)
 450 {
 451         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 452
 453         /* rds-tcp registers as a pernet subys, so the ->exit will only
 454          * get invoked after network acitivity has quiesced. We need to
 455          * clean up all sockets  to quiesce network activity, and use
 456          * the unregistration of the per-net loopback device as a trigger
 457          * to start that cleanup.
 458          */
 459         if (event == NETDEV_UNREGISTER_FINAL &&
 460             dev->ifindex == LOOPBACK_IFINDEX)
 461                 rds_tcp_kill_sock(dev_net(dev));
 462
 463         return NOTIFY_DONE;
 464 }
 465
 466 static struct notifier_block rds_tcp_dev_notifier = {
 467         .notifier_call        = rds_tcp_dev_event,
 468         .priority = -10, /* must be called after other network notifiers */
 469 };
 470
 471 /* when sysctl is used to modify some kernel socket parameters,this
 472  * function  resets the RDS connections in that netns  so that we can
 473  * restart with new parameters.  The assumption is that such reset
 474  * events are few and far-between.
 475  */
 476 static void rds_tcp_sysctl_reset(struct net *net)
 477 {
 478         struct rds_tcp_connection *tc, *_tc;
 479
 480         spin_lock_irq(&rds_tcp_conn_lock);
 481         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 482                 struct net *c_net = read_pnet(&tc->conn->c_net);
 483
 484                 if (net != c_net || !tc->t_sock)
 485                         continue;
 486
 487                 rds_conn_drop(tc->conn); /* reconnect with new parameters */
 488         }
 489         spin_unlock_irq(&rds_tcp_conn_lock);
 490 }
 491
 492 static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
 493                                  void __user *buffer, size_t *lenp,
 494                                  loff_t *fpos)
 495 {
 496         struct net *net = current->nsproxy->net_ns;
 497         int err;
 498
 499         err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
 500         if (err < 0) {
 501                 pr_warn("Invalid input. Must be >= %d\n",
 502                         *(int *)(ctl->extra1));
 503                 return err;
 504         }
 505         if (write)
 506                 rds_tcp_sysctl_reset(net);
 507         return 0;
 508 }
 509
 510 static void rds_tcp_exit(void)
 511 {
 512         rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 513         unregister_pernet_subsys(&rds_tcp_net_ops);
 514         if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
 515                 pr_warn("could not unregister rds_tcp_dev_notifier\n");
 516         rds_tcp_destroy_conns();
 517         rds_trans_unregister(&rds_tcp_transport);
 518         rds_tcp_recv_exit();
 519         kmem_cache_destroy(rds_tcp_conn_slab);
 520 }
 521 module_exit(rds_tcp_exit);
 522
 523 static int rds_tcp_init(void)
 524 {
 525         int ret;
 526
 527         rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
 528                                               sizeof(struct rds_tcp_connection),
 529                                               0, 0, NULL);
 530         if (!rds_tcp_conn_slab) {
 531                 ret = -ENOMEM;
 532                 goto out;
 533         }
 534
 535         ret = register_netdevice_notifier(&rds_tcp_dev_notifier);
 536         if (ret) {
 537                 pr_warn("could not register rds_tcp_dev_notifier\n");
 538                 goto out;
 539         }
 540
 541         ret = register_pernet_subsys(&rds_tcp_net_ops);
 542         if (ret)
 543                 goto out_slab;
 544
 545         ret = rds_tcp_recv_init();
 546         if (ret)
 547                 goto out_slab;
 548
 549         ret = rds_trans_register(&rds_tcp_transport);
 550         if (ret)
 551                 goto out_recv;
 552
 553         rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 554
 555         goto out;
 556
 557 out_recv:
 558         rds_tcp_recv_exit();
 559 out_slab:
 560         unregister_pernet_subsys(&rds_tcp_net_ops);
 561         kmem_cache_destroy(rds_tcp_conn_slab);
 562 out:
 563         return ret;
 564 }
 565 module_init(rds_tcp_init);
 566
 567 MODULE_AUTHOR("Oracle Corporation <[email protected]>");
 568 MODULE_DESCRIPTION("RDS: TCP transport");
 569 MODULE_LICENSE("Dual BSD/GPL");
 570