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 #define RDS_TCP_DEFAULT_BUFSIZE (128 * 1024)
  56
  57 static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
  58                                  void __user *buffer, size_t *lenp,
  59                                  loff_t *fpos);
  60
  61 int rds_tcp_min_sndbuf = SOCK_MIN_SNDBUF;
  62 int rds_tcp_min_rcvbuf = SOCK_MIN_RCVBUF;
  63
  64 static struct ctl_table rds_tcp_sysctl_table[] = {
  65 #define RDS_TCP_SNDBUF  0
  66         {
  67                 .procname       = "rds_tcp_sndbuf",
  68                 /* data is per-net pointer */
  69                 .maxlen         = sizeof(int),
  70                 .mode           = 0644,
  71                 .proc_handler   = rds_tcp_skbuf_handler,
  72                 .extra1         = &rds_tcp_min_sndbuf,
  73         },
  74 #define RDS_TCP_RCVBUF  1
  75         {
  76                 .procname       = "rds_tcp_rcvbuf",
  77                 /* data is per-net pointer */
  78                 .maxlen         = sizeof(int),
  79                 .mode           = 0644,
  80                 .proc_handler   = rds_tcp_skbuf_handler,
  81                 .extra1         = &rds_tcp_min_rcvbuf,
  82         },
  83         { }
  84 };
  85
  86 /* doing it this way avoids calling tcp_sk() */
  87 void rds_tcp_nonagle(struct socket *sock)
  88 {
  89         mm_segment_t oldfs = get_fs();
  90         int val = 1;
  91
  92         set_fs(KERNEL_DS);
  93         sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY, (char __user *)&val,
  94                               sizeof(val));
  95         set_fs(oldfs);
  96 }
  97
  98 u32 rds_tcp_snd_nxt(struct rds_tcp_connection *tc)
  99 {
 100         return tcp_sk(tc->t_sock->sk)->snd_nxt;
 101 }
 102
 103 u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
 104 {
 105         return tcp_sk(tc->t_sock->sk)->snd_una;
 106 }
 107
 108 void rds_tcp_restore_callbacks(struct socket *sock,
 109                                struct rds_tcp_connection *tc)
 110 {
 111         rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
 112         write_lock_bh(&sock->sk->sk_callback_lock);
 113
 114         /* done under the callback_lock to serialize with write_space */
 115         spin_lock(&rds_tcp_tc_list_lock);
 116         list_del_init(&tc->t_list_item);
 117         rds_tcp_tc_count--;
 118         spin_unlock(&rds_tcp_tc_list_lock);
 119
 120         tc->t_sock = NULL;
 121
 122         sock->sk->sk_write_space = tc->t_orig_write_space;
 123         sock->sk->sk_data_ready = tc->t_orig_data_ready;
 124         sock->sk->sk_state_change = tc->t_orig_state_change;
 125         sock->sk->sk_user_data = NULL;
 126
 127         write_unlock_bh(&sock->sk->sk_callback_lock);
 128 }
 129
 130 /*
 131  * This is the only path that sets tc->t_sock.  Send and receive trust that
 132  * it is set.  The RDS_CONN_CONNECTED bit protects those paths from being
 133  * called while it isn't set.
 134  */
 135 void rds_tcp_set_callbacks(struct socket *sock, struct rds_connection *conn)
 136 {
 137         struct rds_tcp_connection *tc = conn->c_transport_data;
 138
 139         rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
 140         write_lock_bh(&sock->sk->sk_callback_lock);
 141
 142         /* done under the callback_lock to serialize with write_space */
 143         spin_lock(&rds_tcp_tc_list_lock);
 144         list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
 145         rds_tcp_tc_count++;
 146         spin_unlock(&rds_tcp_tc_list_lock);
 147
 148         /* accepted sockets need our listen data ready undone */
 149         if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
 150                 sock->sk->sk_data_ready = sock->sk->sk_user_data;
 151
 152         tc->t_sock = sock;
 153         tc->conn = conn;
 154         tc->t_orig_data_ready = sock->sk->sk_data_ready;
 155         tc->t_orig_write_space = sock->sk->sk_write_space;
 156         tc->t_orig_state_change = sock->sk->sk_state_change;
 157
 158         sock->sk->sk_user_data = conn;
 159         sock->sk->sk_data_ready = rds_tcp_data_ready;
 160         sock->sk->sk_write_space = rds_tcp_write_space;
 161         sock->sk->sk_state_change = rds_tcp_state_change;
 162
 163         write_unlock_bh(&sock->sk->sk_callback_lock);
 164 }
 165
 166 static void rds_tcp_tc_info(struct socket *sock, unsigned int len,
 167                             struct rds_info_iterator *iter,
 168                             struct rds_info_lengths *lens)
 169 {
 170         struct rds_info_tcp_socket tsinfo;
 171         struct rds_tcp_connection *tc;
 172         unsigned long flags;
 173         struct sockaddr_in sin;
 174         int sinlen;
 175
 176         spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
 177
 178         if (len / sizeof(tsinfo) < rds_tcp_tc_count)
 179                 goto out;
 180
 181         list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
 182
 183                 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 0);
 184                 tsinfo.local_addr = sin.sin_addr.s_addr;
 185                 tsinfo.local_port = sin.sin_port;
 186                 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 1);
 187                 tsinfo.peer_addr = sin.sin_addr.s_addr;
 188                 tsinfo.peer_port = sin.sin_port;
 189
 190                 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
 191                 tsinfo.data_rem = tc->t_tinc_data_rem;
 192                 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
 193                 tsinfo.last_expected_una = tc->t_last_expected_una;
 194                 tsinfo.last_seen_una = tc->t_last_seen_una;
 195
 196                 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
 197         }
 198
 199 out:
 200         lens->nr = rds_tcp_tc_count;
 201         lens->each = sizeof(tsinfo);
 202
 203         spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
 204 }
 205
 206 static int rds_tcp_laddr_check(struct net *net, __be32 addr)
 207 {
 208         if (inet_addr_type(net, addr) == RTN_LOCAL)
 209                 return 0;
 210         return -EADDRNOTAVAIL;
 211 }
 212
 213 static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 214 {
 215         struct rds_tcp_connection *tc;
 216
 217         tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
 218         if (!tc)
 219                 return -ENOMEM;
 220
 221         tc->t_sock = NULL;
 222         tc->t_tinc = NULL;
 223         tc->t_tinc_hdr_rem = sizeof(struct rds_header);
 224         tc->t_tinc_data_rem = 0;
 225
 226         conn->c_transport_data = tc;
 227
 228         spin_lock_irq(&rds_tcp_conn_lock);
 229         list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
 230         spin_unlock_irq(&rds_tcp_conn_lock);
 231
 232         rdsdebug("alloced tc %p\n", conn->c_transport_data);
 233         return 0;
 234 }
 235
 236 static void rds_tcp_conn_free(void *arg)
 237 {
 238         struct rds_tcp_connection *tc = arg;
 239         unsigned long flags;
 240         rdsdebug("freeing tc %p\n", tc);
 241
 242         spin_lock_irqsave(&rds_tcp_conn_lock, flags);
 243         list_del(&tc->t_tcp_node);
 244         spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
 245
 246         kmem_cache_free(rds_tcp_conn_slab, tc);
 247 }
 248
 249 static void rds_tcp_destroy_conns(void)
 250 {
 251         struct rds_tcp_connection *tc, *_tc;
 252         LIST_HEAD(tmp_list);
 253
 254         /* avoid calling conn_destroy with irqs off */
 255         spin_lock_irq(&rds_tcp_conn_lock);
 256         list_splice(&rds_tcp_conn_list, &tmp_list);
 257         INIT_LIST_HEAD(&rds_tcp_conn_list);
 258         spin_unlock_irq(&rds_tcp_conn_lock);
 259
 260         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
 261                 if (tc->conn->c_passive)
 262                         rds_conn_destroy(tc->conn->c_passive);
 263                 rds_conn_destroy(tc->conn);
 264         }
 265 }
 266
 267 static void rds_tcp_exit(void);
 268
 269 struct rds_transport rds_tcp_transport = {
 270         .laddr_check            = rds_tcp_laddr_check,
 271         .xmit_prepare           = rds_tcp_xmit_prepare,
 272         .xmit_complete          = rds_tcp_xmit_complete,
 273         .xmit                   = rds_tcp_xmit,
 274         .recv                   = rds_tcp_recv,
 275         .conn_alloc             = rds_tcp_conn_alloc,
 276         .conn_free              = rds_tcp_conn_free,
 277         .conn_connect           = rds_tcp_conn_connect,
 278         .conn_shutdown          = rds_tcp_conn_shutdown,
 279         .inc_copy_to_user       = rds_tcp_inc_copy_to_user,
 280         .inc_free               = rds_tcp_inc_free,
 281         .stats_info_copy        = rds_tcp_stats_info_copy,
 282         .exit                   = rds_tcp_exit,
 283         .t_owner                = THIS_MODULE,
 284         .t_name                 = "tcp",
 285         .t_type                 = RDS_TRANS_TCP,
 286         .t_prefer_loopback      = 1,
 287 };
 288
 289 static int rds_tcp_netid;
 290
 291 /* per-network namespace private data for this module */
 292 struct rds_tcp_net {
 293         struct socket *rds_tcp_listen_sock;
 294         struct work_struct rds_tcp_accept_w;
 295         struct ctl_table_header *rds_tcp_sysctl;
 296         struct ctl_table *ctl_table;
 297         int sndbuf_size;
 298         int rcvbuf_size;
 299 };
 300
 301 /* All module specific customizations to the RDS-TCP socket should be done in
 302  * rds_tcp_tune() and applied after socket creation.
 303  */
 304 void rds_tcp_tune(struct socket *sock)
 305 {
 306         struct sock *sk = sock->sk;
 307         struct net *net = sock_net(sk);
 308         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 309
 310         rds_tcp_nonagle(sock);
 311         lock_sock(sk);
 312         if (rtn->sndbuf_size > 0) {
 313                 sk->sk_sndbuf = rtn->sndbuf_size;
 314                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
 315         }
 316         if (rtn->rcvbuf_size > 0) {
 317                 sk->sk_sndbuf = rtn->rcvbuf_size;
 318                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
 319         }
 320         release_sock(sk);
 321 }
 322
 323 static void rds_tcp_accept_worker(struct work_struct *work)
 324 {
 325         struct rds_tcp_net *rtn = container_of(work,
 326                                                struct rds_tcp_net,
 327                                                rds_tcp_accept_w);
 328
 329         while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
 330                 cond_resched();
 331 }
 332
 333 void rds_tcp_accept_work(struct sock *sk)
 334 {
 335         struct net *net = sock_net(sk);
 336         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 337
 338         queue_work(rds_wq, &rtn->rds_tcp_accept_w);
 339 }
 340
 341 static __net_init int rds_tcp_init_net(struct net *net)
 342 {
 343         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 344         struct ctl_table *tbl;
 345         int err = 0;
 346
 347         memset(rtn, 0, sizeof(*rtn));
 348
 349         /* {snd, rcv}buf_size default to 0, which implies we let the
 350          * stack pick the value, and permit auto-tuning of buffer size.
 351          */
 352         if (net == &init_net) {
 353                 tbl = rds_tcp_sysctl_table;
 354         } else {
 355                 tbl = kmemdup(rds_tcp_sysctl_table,
 356                               sizeof(rds_tcp_sysctl_table), GFP_KERNEL);
 357                 if (!tbl) {
 358                         pr_warn("could not set allocate syctl table\n");
 359                         return -ENOMEM;
 360                 }
 361                 rtn->ctl_table = tbl;
 362         }
 363         tbl[RDS_TCP_SNDBUF].data = &rtn->sndbuf_size;
 364         tbl[RDS_TCP_RCVBUF].data = &rtn->rcvbuf_size;
 365         rtn->rds_tcp_sysctl = register_net_sysctl(net, "net/rds/tcp", tbl);
 366         if (!rtn->rds_tcp_sysctl) {
 367                 pr_warn("could not register sysctl\n");
 368                 err = -ENOMEM;
 369                 goto fail;
 370         }
 371         rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
 372         if (!rtn->rds_tcp_listen_sock) {
 373                 pr_warn("could not set up listen sock\n");
 374                 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
 375                 rtn->rds_tcp_sysctl = NULL;
 376                 err = -EAFNOSUPPORT;
 377                 goto fail;
 378         }
 379         INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
 380         return 0;
 381
 382 fail:
 383         if (net != &init_net)
 384                 kfree(tbl);
 385         return err;
 386 }
 387
 388 static void __net_exit rds_tcp_exit_net(struct net *net)
 389 {
 390         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 391
 392         if (rtn->rds_tcp_sysctl)
 393                 unregister_net_sysctl_table(rtn->rds_tcp_sysctl);
 394
 395         if (net != &init_net && rtn->ctl_table)
 396                 kfree(rtn->ctl_table);
 397
 398         /* If rds_tcp_exit_net() is called as a result of netns deletion,
 399          * the rds_tcp_kill_sock() device notifier would already have cleaned
 400          * up the listen socket, thus there is no work to do in this function.
 401          *
 402          * If rds_tcp_exit_net() is called as a result of module unload,
 403          * i.e., due to rds_tcp_exit() -> unregister_pernet_subsys(), then
 404          * we do need to clean up the listen socket here.
 405          */
 406         if (rtn->rds_tcp_listen_sock) {
 407                 rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 408                 rtn->rds_tcp_listen_sock = NULL;
 409                 flush_work(&rtn->rds_tcp_accept_w);
 410         }
 411 }
 412
 413 static struct pernet_operations rds_tcp_net_ops = {
 414         .init = rds_tcp_init_net,
 415         .exit = rds_tcp_exit_net,
 416         .id = &rds_tcp_netid,
 417         .size = sizeof(struct rds_tcp_net),
 418 };
 419
 420 static void rds_tcp_kill_sock(struct net *net)
 421 {
 422         struct rds_tcp_connection *tc, *_tc;
 423         struct sock *sk;
 424         LIST_HEAD(tmp_list);
 425         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 426
 427         rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 428         rtn->rds_tcp_listen_sock = NULL;
 429         flush_work(&rtn->rds_tcp_accept_w);
 430         spin_lock_irq(&rds_tcp_conn_lock);
 431         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 432                 struct net *c_net = read_pnet(&tc->conn->c_net);
 433
 434                 if (net != c_net || !tc->t_sock)
 435                         continue;
 436                 list_move_tail(&tc->t_tcp_node, &tmp_list);
 437         }
 438         spin_unlock_irq(&rds_tcp_conn_lock);
 439         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
 440                 sk = tc->t_sock->sk;
 441                 sk->sk_prot->disconnect(sk, 0);
 442                 tcp_done(sk);
 443                 if (tc->conn->c_passive)
 444                         rds_conn_destroy(tc->conn->c_passive);
 445                 rds_conn_destroy(tc->conn);
 446         }
 447 }
 448
 449 static int rds_tcp_dev_event(struct notifier_block *this,
 450                              unsigned long event, void *ptr)
 451 {
 452         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 453
 454         /* rds-tcp registers as a pernet subys, so the ->exit will only
 455          * get invoked after network acitivity has quiesced. We need to
 456          * clean up all sockets  to quiesce network activity, and use
 457          * the unregistration of the per-net loopback device as a trigger
 458          * to start that cleanup.
 459          */
 460         if (event == NETDEV_UNREGISTER_FINAL &&
 461             dev->ifindex == LOOPBACK_IFINDEX)
 462                 rds_tcp_kill_sock(dev_net(dev));
 463
 464         return NOTIFY_DONE;
 465 }
 466
 467 static struct notifier_block rds_tcp_dev_notifier = {
 468         .notifier_call        = rds_tcp_dev_event,
 469         .priority = -10, /* must be called after other network notifiers */
 470 };
 471
 472 /* when sysctl is used to modify some kernel socket parameters,this
 473  * function  resets the RDS connections in that netns  so that we can
 474  * restart with new parameters.  The assumption is that such reset
 475  * events are few and far-between.
 476  */
 477 static void rds_tcp_sysctl_reset(struct net *net)
 478 {
 479         struct rds_tcp_connection *tc, *_tc;
 480
 481         spin_lock_irq(&rds_tcp_conn_lock);
 482         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 483                 struct net *c_net = read_pnet(&tc->conn->c_net);
 484
 485                 if (net != c_net || !tc->t_sock)
 486                         continue;
 487
 488                 rds_conn_drop(tc->conn); /* reconnect with new parameters */
 489         }
 490         spin_unlock_irq(&rds_tcp_conn_lock);
 491 }
 492
 493 static int rds_tcp_skbuf_handler(struct ctl_table *ctl, int write,
 494                                  void __user *buffer, size_t *lenp,
 495                                  loff_t *fpos)
 496 {
 497         struct net *net = current->nsproxy->net_ns;
 498         int err;
 499
 500         err = proc_dointvec_minmax(ctl, write, buffer, lenp, fpos);
 501         if (err < 0) {
 502                 pr_warn("Invalid input. Must be >= %d\n",
 503                         *(int *)(ctl->extra1));
 504                 return err;
 505         }
 506         if (write)
 507                 rds_tcp_sysctl_reset(net);
 508         return 0;
 509 }
 510
 511 static void rds_tcp_exit(void)
 512 {
 513         rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 514         unregister_pernet_subsys(&rds_tcp_net_ops);
 515         if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
 516                 pr_warn("could not unregister rds_tcp_dev_notifier\n");
 517         rds_tcp_destroy_conns();
 518         rds_trans_unregister(&rds_tcp_transport);
 519         rds_tcp_recv_exit();
 520         kmem_cache_destroy(rds_tcp_conn_slab);
 521 }
 522 module_exit(rds_tcp_exit);
 523
 524 static int rds_tcp_init(void)
 525 {
 526         int ret;
 527
 528         rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
 529                                               sizeof(struct rds_tcp_connection),
 530                                               0, 0, NULL);
 531         if (!rds_tcp_conn_slab) {
 532                 ret = -ENOMEM;
 533                 goto out;
 534         }
 535
 536         ret = register_netdevice_notifier(&rds_tcp_dev_notifier);
 537         if (ret) {
 538                 pr_warn("could not register rds_tcp_dev_notifier\n");
 539                 goto out;
 540         }
 541
 542         ret = register_pernet_subsys(&rds_tcp_net_ops);
 543         if (ret)
 544                 goto out_slab;
 545
 546         ret = rds_tcp_recv_init();
 547         if (ret)
 548                 goto out_slab;
 549
 550         ret = rds_trans_register(&rds_tcp_transport);
 551         if (ret)
 552                 goto out_recv;
 553
 554         rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 555
 556         goto out;
 557
 558 out_recv:
 559         rds_tcp_recv_exit();
 560 out_slab:
 561         unregister_pernet_subsys(&rds_tcp_net_ops);
 562         kmem_cache_destroy(rds_tcp_conn_slab);
 563 out:
 564         return ret;
 565 }
 566 module_init(rds_tcp_init);
 567
 568 MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
 569 MODULE_DESCRIPTION("RDS: TCP transport");
 570 MODULE_LICENSE("Dual BSD/GPL");
 571