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_CONNECTED 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         tc->t_sock = NULL;
 220         tc->t_tinc = NULL;
 221         tc->t_tinc_hdr_rem = sizeof(struct rds_header);
 222         tc->t_tinc_data_rem = 0;
 223
 224         conn->c_transport_data = tc;
 225
 226         spin_lock_irq(&rds_tcp_conn_lock);
 227         list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
 228         spin_unlock_irq(&rds_tcp_conn_lock);
 229
 230         rdsdebug("alloced tc %p\n", conn->c_transport_data);
 231         return 0;
 232 }
 233
 234 static void rds_tcp_conn_free(void *arg)
 235 {
 236         struct rds_tcp_connection *tc = arg;
 237         unsigned long flags;
 238         rdsdebug("freeing tc %p\n", tc);
 239
 240         spin_lock_irqsave(&rds_tcp_conn_lock, flags);
 241         list_del(&tc->t_tcp_node);
 242         spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
 243
 244         kmem_cache_free(rds_tcp_conn_slab, tc);
 245 }
 246
 247 static void rds_tcp_destroy_conns(void)
 248 {
 249         struct rds_tcp_connection *tc, *_tc;
 250         LIST_HEAD(tmp_list);
 251
 252         /* avoid calling conn_destroy with irqs off */
 253         spin_lock_irq(&rds_tcp_conn_lock);
 254         list_splice(&rds_tcp_conn_list, &tmp_list);
 255         INIT_LIST_HEAD(&rds_tcp_conn_list);
 256         spin_unlock_irq(&rds_tcp_conn_lock);
 257
 258         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
 259                 if (tc->conn->c_passive)
 260                         rds_conn_destroy(tc->conn->c_passive);
 261                 rds_conn_destroy(tc->conn);
 262         }
 263 }
 264
 265 static void rds_tcp_exit(void);
 266
 267 struct rds_transport rds_tcp_transport = {
 268         .laddr_check            = rds_tcp_laddr_check,
 269         .xmit_prepare           = rds_tcp_xmit_prepare,
 270         .xmit_complete          = rds_tcp_xmit_complete,
 271         .xmit                   = rds_tcp_xmit,
 272         .recv                   = rds_tcp_recv,
 273         .conn_alloc             = rds_tcp_conn_alloc,
 274         .conn_free              = rds_tcp_conn_free,
 275         .conn_connect           = rds_tcp_conn_connect,
 276         .conn_shutdown          = rds_tcp_conn_shutdown,
 277         .inc_copy_to_user       = rds_tcp_inc_copy_to_user,
 278         .inc_free               = rds_tcp_inc_free,
 279         .stats_info_copy        = rds_tcp_stats_info_copy,
 280         .exit                   = rds_tcp_exit,
 281         .t_owner                = THIS_MODULE,
 282         .t_name                 = "tcp",
 283         .t_type                 = RDS_TRANS_TCP,
 284         .t_prefer_loopback      = 1,
 285 };
 286
 287 static int rds_tcp_netid;
 288
 289 /* per-network namespace private data for this module */
 290 struct rds_tcp_net {
 291         struct socket *rds_tcp_listen_sock;
 292         struct work_struct rds_tcp_accept_w;
 293         struct ctl_table_header *rds_tcp_sysctl;
 294         struct ctl_table *ctl_table;
 295         int sndbuf_size;
 296         int rcvbuf_size;
 297 };
 298
 299 /* All module specific customizations to the RDS-TCP socket should be done in
 300  * rds_tcp_tune() and applied after socket creation.
 301  */
 302 void rds_tcp_tune(struct socket *sock)
 303 {
 304         struct sock *sk = sock->sk;
 305         struct net *net = sock_net(sk);
 306         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 307
 308         rds_tcp_nonagle(sock);
 309         lock_sock(sk);
 310         if (rtn->sndbuf_size > 0) {
 311                 sk->sk_sndbuf = rtn->sndbuf_size;
 312                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
 313         }
 314         if (rtn->rcvbuf_size > 0) {
 315                 sk->sk_sndbuf = rtn->rcvbuf_size;
 316                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
 317         }
 318         release_sock(sk);
 319 }
 320
 321 static void rds_tcp_accept_worker(struct work_struct *work)
 322 {
 323         struct rds_tcp_net *rtn = container_of(work,
 324                                                struct rds_tcp_net,
 325                                                rds_tcp_accept_w);
 326
 327         while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
 328                 cond_resched();
 329 }
 330
 331 void rds_tcp_accept_work(struct sock *sk)
 332 {
 333         struct net *net = sock_net(sk);
 334         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 335
 336         queue_work(rds_wq, &rtn->rds_tcp_accept_w);
 337 }
 338
 339 static __net_init int rds_tcp_init_net(struct net *net)
 340 {
 341         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 342         struct ctl_table *tbl;
 343         int err = 0;
 344
 345         memset(rtn, 0, sizeof(*rtn));
 346
 347         /* {snd, rcv}buf_size default to 0, which implies we let the
 348          * stack pick the value, and permit auto-tuning of buffer size.
 349          */
 350         if (net == &init_net) {
 351                 tbl = rds_tcp_sysctl_table;
 352         } else {
 353                 tbl = kmemdup(rds_tcp_sysctl_table,
 354                               sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
 355                 if (!tbl) {
 356                         pr_warn("could not set allocate syctl table\n");
 357                         return -ENOMEM;
 358                 }
 359                 rtn->ctl_table = tbl;
 360         }
 361         tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
 362         tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
 363         rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
 364         if (!rtn->rds_tcp_sysctl) {
 365                 pr_warn("could not register sysctl\n");
 366                 err = -ENOMEM;
 367                 goto fail;
 368         }
 369         rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
 370         if (!rtn->rds_tcp_listen_sock) {
 371                 pr_warn("could not set up listen sock\n");
 372                 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
 373                 rtn->rds_tcp_sysctl = NULL;
 374                 err = -EAFNOSUPPORT;
 375                 goto fail;
 376         }
 377         INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
 378         return 0;
 379
 380 fail:
 381         if (net != &init_net)
 382                 kfree(tbl);
 383         return err;
 384 }
 385
 386 static void __net_exit rds_tcp_exit_net(struct net *net)
 387 {
 388         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 389
 390         if (rtn->rds_tcp_sysctl)
 391                 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
 392
 393         if (net != &init_net && rtn->ctl_table)
 394                 kfree(rtn->ctl_table);
 395
 396         /* If rds_tcp_exit_net() is called as a result of netns deletion,
 397          * the rds_tcp_kill_sock() device notifier would already have cleaned
 398          * up the listen socket, thus there is no work to do in this function.
 399          *
 400          * If rds_tcp_exit_net() is called as a result of module unload,
 401          * i.e., due to rds_tcp_exit() -> unregister_pernet_subsys(), then
 402          * we do need to clean up the listen socket here.
 403          */
 404         if (rtn->rds_tcp_listen_sock) {
 405                 rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 406                 rtn->rds_tcp_listen_sock = NULL;
 407                 flush_work(&rtn->rds_tcp_accept_w);
 408         }
 409 }
 410
 411 static struct pernet_operations rds_tcp_net_ops = {
 412         .init = rds_tcp_init_net,
 413         .exit = rds_tcp_exit_net,
 414         .id = &rds_tcp_netid,
 415         .size = sizeof(struct rds_tcp_net),
 416 };
 417
 418 static void rds_tcp_kill_sock(struct net *net)
 419 {
 420         struct rds_tcp_connection *tc, *_tc;
 421         struct sock *sk;
 422         LIST_HEAD(tmp_list);
 423         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 424
 425         rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 426         rtn->rds_tcp_listen_sock = NULL;
 427         flush_work(&rtn->rds_tcp_accept_w);
 428         spin_lock_irq(&rds_tcp_conn_lock);
 429         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 430                 struct net *c_net = read_pnet(&tc->conn->c_net);
 431
 432                 if (net != c_net || !tc->t_sock)
 433                         continue;
 434                 list_move_tail(&tc->t_tcp_node, &tmp_list);
 435         }
 436         spin_unlock_irq(&rds_tcp_conn_lock);
 437         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
 438                 sk = tc->t_sock->sk;
 439                 sk->sk_prot->disconnect(sk, 0);
 440                 tcp_done(sk);
 441                 if (tc->conn->c_passive)
 442                         rds_conn_destroy(tc->conn->c_passive);
 443                 rds_conn_destroy(tc->conn);
 444         }
 445 }
 446
 447 static int rds_tcp_dev_event(struct notifier_block *this,
 448                              unsigned long event, void *ptr)
 449 {
 450         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 451
 452         /* rds-tcp registers as a pernet subys, so the ->exit will only
 453          * get invoked after network acitivity has quiesced. We need to
 454          * clean up all sockets  to quiesce network activity, and use
 455          * the unregistration of the per-net loopback device as a trigger
 456          * to start that cleanup.
 457          */
 458         if (event == NETDEV_UNREGISTER_FINAL &&
 459             dev->ifindex == LOOPBACK_IFINDEX)
 460                 rds_tcp_kill_sock(dev_net(dev));
 461
 462         return NOTIFY_DONE;
 463 }
 464
 465 static struct notifier_block rds_tcp_dev_notifier = {
 466         .notifier_call        = rds_tcp_dev_event,
 467         .priority = -10, /* must be called after other network notifiers */
 468 };
 469
 470 /* when sysctl is used to modify some kernel socket parameters,this
 471  * function  resets the RDS connections in that netns  so that we can
 472  * restart with new parameters.  The assumption is that such reset
 473  * events are few and far-between.
 474  */
 475 static void rds_tcp_sysctl_reset(struct net *net)
 476 {
 477         struct rds_tcp_connection *tc, *_tc;
 478
 479         spin_lock_irq(&rds_tcp_conn_lock);
 480         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 481                 struct net *c_net = read_pnet(&tc->conn->c_net);
 482
 483                 if (net != c_net || !tc->t_sock)
 484                         continue;
 485
 486                 rds_conn_drop(tc->conn); /* reconnect with new parameters */
 487         }
 488         spin_unlock_irq(&rds_tcp_conn_lock);
 489 }
 490
 491 static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
 492                                  void __user *buffer, size_t *lenp,
 493                                  loff_t *fpos)
 494 {
 495         struct net *net = current->nsproxy->net_ns;
 496         int err;
 497
 498         err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
 499         if (err < 0) {
 500                 pr_warn("Invalid input. Must be >= %d\n",
 501                         *(int *)(ctl->extra1));
 502                 return err;
 503         }
 504         if (write)
 505                 rds_tcp_sysctl_reset(net);
 506         return 0;
 507 }
 508
 509 static void rds_tcp_exit(void)
 510 {
 511         rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 512         unregister_pernet_subsys(&rds_tcp_net_ops);
 513         if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
 514                 pr_warn("could not unregister rds_tcp_dev_notifier\n");
 515         rds_tcp_destroy_conns();
 516         rds_trans_unregister(&rds_tcp_transport);
 517         rds_tcp_recv_exit();
 518         kmem_cache_destroy(rds_tcp_conn_slab);
 519 }
 520 module_exit(rds_tcp_exit);
 521
 522 static int rds_tcp_init(void)
 523 {
 524         int ret;
 525
 526         rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
 527                                               sizeof(struct rds_tcp_connection),
 528                                               0, 0, NULL);
 529         if (!rds_tcp_conn_slab) {
 530                 ret = -ENOMEM;
 531                 goto out;
 532         }
 533
 534         ret = register_netdevice_notifier(&rds_tcp_dev_notifier);
 535         if (ret) {
 536                 pr_warn("could not register rds_tcp_dev_notifier\n");
 537                 goto out;
 538         }
 539
 540         ret = register_pernet_subsys(&rds_tcp_net_ops);
 541         if (ret)
 542                 goto out_slab;
 543
 544         ret = rds_tcp_recv_init();
 545         if (ret)
 546                 goto out_slab;
 547
 548         ret = rds_trans_register(&rds_tcp_transport);
 549         if (ret)
 550                 goto out_recv;
 551
 552         rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 553
 554         goto out;
 555
 556 out_recv:
 557         rds_tcp_recv_exit();
 558 out_slab:
 559         unregister_pernet_subsys(&rds_tcp_net_ops);
 560         kmem_cache_destroy(rds_tcp_conn_slab);
 561 out:
 562         return ret;
 563 }
 564 module_init(rds_tcp_init);
 565
 566 MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
 567 MODULE_DESCRIPTION("RDS: TCP transport");
 568 MODULE_LICENSE("Dual BSD/GPL");
 569