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 /* doing it this way avoids calling tcp_sk() */
  58 void rds_tcp_nonagle(struct socket *sock)
  59 {
  60         mm_segment_t oldfs = get_fs();
  61         int val = 1;
  62
  63         set_fs(KERNEL_DS);
  64         sock->ops->setsockopt(sock, SOL_TCP, TCP_NODELAY, (char __user *)&val,
  65                               sizeof(val));
  66         set_fs(oldfs);
  67 }
  68
  69 /* All module specific customizations to the RDS-TCP socket should be done in
  70  * rds_tcp_tune() and applied after socket creation. In general these
  71  * customizations should be tunable via module_param()
  72  */
  73 void rds_tcp_tune(struct socket *sock)
  74 {
  75         rds_tcp_nonagle(sock);
  76 }
  77
  78 u32 rds_tcp_snd_nxt(struct rds_tcp_connection *tc)
  79 {
  80         return tcp_sk(tc->t_sock->sk)->snd_nxt;
  81 }
  82
  83 u32 rds_tcp_snd_una(struct rds_tcp_connection *tc)
  84 {
  85         return tcp_sk(tc->t_sock->sk)->snd_una;
  86 }
  87
  88 void rds_tcp_restore_callbacks(struct socket *sock,
  89                                struct rds_tcp_connection *tc)
  90 {
  91         rdsdebug("restoring sock %p callbacks from tc %p\n", sock, tc);
  92         write_lock_bh(&sock->sk->sk_callback_lock);
  93
  94         /* done under the callback_lock to serialize with write_space */
  95         spin_lock(&rds_tcp_tc_list_lock);
  96         list_del_init(&tc->t_list_item);
  97         rds_tcp_tc_count--;
  98         spin_unlock(&rds_tcp_tc_list_lock);
  99
 100         tc->t_sock = NULL;
 101
 102         sock->sk->sk_write_space = tc->t_orig_write_space;
 103         sock->sk->sk_data_ready = tc->t_orig_data_ready;
 104         sock->sk->sk_state_change = tc->t_orig_state_change;
 105         sock->sk->sk_user_data = NULL;
 106
 107         write_unlock_bh(&sock->sk->sk_callback_lock);
 108 }
 109
 110 /*
 111  * This is the only path that sets tc->t_sock.  Send and receive trust that
 112  * it is set.  The RDS_CONN_CONNECTED bit protects those paths from being
 113  * called while it isn't set.
 114  */
 115 void rds_tcp_set_callbacks(struct socket *sock, struct rds_connection *conn)
 116 {
 117         struct rds_tcp_connection *tc = conn->c_transport_data;
 118
 119         rdsdebug("setting sock %p callbacks to tc %p\n", sock, tc);
 120         write_lock_bh(&sock->sk->sk_callback_lock);
 121
 122         /* done under the callback_lock to serialize with write_space */
 123         spin_lock(&rds_tcp_tc_list_lock);
 124         list_add_tail(&tc->t_list_item, &rds_tcp_tc_list);
 125         rds_tcp_tc_count++;
 126         spin_unlock(&rds_tcp_tc_list_lock);
 127
 128         /* accepted sockets need our listen data ready undone */
 129         if (sock->sk->sk_data_ready == rds_tcp_listen_data_ready)
 130                 sock->sk->sk_data_ready = sock->sk->sk_user_data;
 131
 132         tc->t_sock = sock;
 133         tc->conn = conn;
 134         tc->t_orig_data_ready = sock->sk->sk_data_ready;
 135         tc->t_orig_write_space = sock->sk->sk_write_space;
 136         tc->t_orig_state_change = sock->sk->sk_state_change;
 137
 138         sock->sk->sk_user_data = conn;
 139         sock->sk->sk_data_ready = rds_tcp_data_ready;
 140         sock->sk->sk_write_space = rds_tcp_write_space;
 141         sock->sk->sk_state_change = rds_tcp_state_change;
 142
 143         write_unlock_bh(&sock->sk->sk_callback_lock);
 144 }
 145
 146 static void rds_tcp_tc_info(struct socket *sock, unsigned int len,
 147                             struct rds_info_iterator *iter,
 148                             struct rds_info_lengths *lens)
 149 {
 150         struct rds_info_tcp_socket tsinfo;
 151         struct rds_tcp_connection *tc;
 152         unsigned long flags;
 153         struct sockaddr_in sin;
 154         int sinlen;
 155
 156         spin_lock_irqsave(&rds_tcp_tc_list_lock, flags);
 157
 158         if (len / sizeof(tsinfo) < rds_tcp_tc_count)
 159                 goto out;
 160
 161         list_for_each_entry(tc, &rds_tcp_tc_list, t_list_item) {
 162
 163                 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 0);
 164                 tsinfo.local_addr = sin.sin_addr.s_addr;
 165                 tsinfo.local_port = sin.sin_port;
 166                 sock->ops->getname(sock, (struct sockaddr *)&sin, &sinlen, 1);
 167                 tsinfo.peer_addr = sin.sin_addr.s_addr;
 168                 tsinfo.peer_port = sin.sin_port;
 169
 170                 tsinfo.hdr_rem = tc->t_tinc_hdr_rem;
 171                 tsinfo.data_rem = tc->t_tinc_data_rem;
 172                 tsinfo.last_sent_nxt = tc->t_last_sent_nxt;
 173                 tsinfo.last_expected_una = tc->t_last_expected_una;
 174                 tsinfo.last_seen_una = tc->t_last_seen_una;
 175
 176                 rds_info_copy(iter, &tsinfo, sizeof(tsinfo));
 177         }
 178
 179 out:
 180         lens->nr = rds_tcp_tc_count;
 181         lens->each = sizeof(tsinfo);
 182
 183         spin_unlock_irqrestore(&rds_tcp_tc_list_lock, flags);
 184 }
 185
 186 static int rds_tcp_laddr_check(struct net *net, __be32 addr)
 187 {
 188         if (inet_addr_type(net, addr) == RTN_LOCAL)
 189                 return 0;
 190         return -EADDRNOTAVAIL;
 191 }
 192
 193 static int rds_tcp_conn_alloc(struct rds_connection *conn, gfp_t gfp)
 194 {
 195         struct rds_tcp_connection *tc;
 196
 197         tc = kmem_cache_alloc(rds_tcp_conn_slab, gfp);
 198         if (!tc)
 199                 return -ENOMEM;
 200
 201         tc->t_sock = NULL;
 202         tc->t_tinc = NULL;
 203         tc->t_tinc_hdr_rem = sizeof(struct rds_header);
 204         tc->t_tinc_data_rem = 0;
 205
 206         conn->c_transport_data = tc;
 207
 208         spin_lock_irq(&rds_tcp_conn_lock);
 209         list_add_tail(&tc->t_tcp_node, &rds_tcp_conn_list);
 210         spin_unlock_irq(&rds_tcp_conn_lock);
 211
 212         rdsdebug("alloced tc %p\n", conn->c_transport_data);
 213         return 0;
 214 }
 215
 216 static void rds_tcp_conn_free(void *arg)
 217 {
 218         struct rds_tcp_connection *tc = arg;
 219         unsigned long flags;
 220         rdsdebug("freeing tc %p\n", tc);
 221
 222         spin_lock_irqsave(&rds_tcp_conn_lock, flags);
 223         list_del(&tc->t_tcp_node);
 224         spin_unlock_irqrestore(&rds_tcp_conn_lock, flags);
 225
 226         kmem_cache_free(rds_tcp_conn_slab, tc);
 227 }
 228
 229 static void rds_tcp_destroy_conns(void)
 230 {
 231         struct rds_tcp_connection *tc, *_tc;
 232         LIST_HEAD(tmp_list);
 233
 234         /* avoid calling conn_destroy with irqs off */
 235         spin_lock_irq(&rds_tcp_conn_lock);
 236         list_splice(&rds_tcp_conn_list, &tmp_list);
 237         INIT_LIST_HEAD(&rds_tcp_conn_list);
 238         spin_unlock_irq(&rds_tcp_conn_lock);
 239
 240         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
 241                 if (tc->conn->c_passive)
 242                         rds_conn_destroy(tc->conn->c_passive);
 243                 rds_conn_destroy(tc->conn);
 244         }
 245 }
 246
 247 static void rds_tcp_exit(void);
 248
 249 struct rds_transport rds_tcp_transport = {
 250         .laddr_check            = rds_tcp_laddr_check,
 251         .xmit_prepare           = rds_tcp_xmit_prepare,
 252         .xmit_complete          = rds_tcp_xmit_complete,
 253         .xmit                   = rds_tcp_xmit,
 254         .recv                   = rds_tcp_recv,
 255         .conn_alloc             = rds_tcp_conn_alloc,
 256         .conn_free              = rds_tcp_conn_free,
 257         .conn_connect           = rds_tcp_conn_connect,
 258         .conn_shutdown          = rds_tcp_conn_shutdown,
 259         .inc_copy_to_user       = rds_tcp_inc_copy_to_user,
 260         .inc_free               = rds_tcp_inc_free,
 261         .stats_info_copy        = rds_tcp_stats_info_copy,
 262         .exit                   = rds_tcp_exit,
 263         .t_owner                = THIS_MODULE,
 264         .t_name                 = "tcp",
 265         .t_type                 = RDS_TRANS_TCP,
 266         .t_prefer_loopback      = 1,
 267 };
 268
 269 static int rds_tcp_netid;
 270
 271 /* per-network namespace private data for this module */
 272 struct rds_tcp_net {
 273         struct socket *rds_tcp_listen_sock;
 274         struct work_struct rds_tcp_accept_w;
 275 };
 276
 277 static void rds_tcp_accept_worker(struct work_struct *work)
 278 {
 279         struct rds_tcp_net *rtn = container_of(work,
 280                                                struct rds_tcp_net,
 281                                                rds_tcp_accept_w);
 282
 283         while (rds_tcp_accept_one(rtn->rds_tcp_listen_sock) == 0)
 284                 cond_resched();
 285 }
 286
 287 void rds_tcp_accept_work(struct sock *sk)
 288 {
 289         struct net *net = sock_net(sk);
 290         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 291
 292         queue_work(rds_wq, &rtn->rds_tcp_accept_w);
 293 }
 294
 295 static __net_init int rds_tcp_init_net(struct net *net)
 296 {
 297         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 298
 299         rtn->rds_tcp_listen_sock = rds_tcp_listen_init(net);
 300         if (!rtn->rds_tcp_listen_sock) {
 301                 pr_warn("could not set up listen sock\n");
 302                 return -EAFNOSUPPORT;
 303         }
 304         INIT_WORK(&rtn->rds_tcp_accept_w, rds_tcp_accept_worker);
 305         return 0;
 306 }
 307
 308 static void __net_exit rds_tcp_exit_net(struct net *net)
 309 {
 310         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 311
 312         /* If rds_tcp_exit_net() is called as a result of netns deletion,
 313          * the rds_tcp_kill_sock() device notifier would already have cleaned
 314          * up the listen socket, thus there is no work to do in this function.
 315          *
 316          * If rds_tcp_exit_net() is called as a result of module unload,
 317          * i.e., due to rds_tcp_exit() -> unregister_pernet_subsys(), then
 318          * we do need to clean up the listen socket here.
 319          */
 320         if (rtn->rds_tcp_listen_sock) {
 321                 rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 322                 rtn->rds_tcp_listen_sock = NULL;
 323                 flush_work(&rtn->rds_tcp_accept_w);
 324         }
 325 }
 326
 327 static struct pernet_operations rds_tcp_net_ops = {
 328         .init = rds_tcp_init_net,
 329         .exit = rds_tcp_exit_net,
 330         .id = &rds_tcp_netid,
 331         .size = sizeof(struct rds_tcp_net),
 332 };
 333
 334 static void rds_tcp_kill_sock(struct net *net)
 335 {
 336         struct rds_tcp_connection *tc, *_tc;
 337         struct sock *sk;
 338         LIST_HEAD(tmp_list);
 339         struct rds_tcp_net *rtn = net_generic(net, rds_tcp_netid);
 340
 341         rds_tcp_listen_stop(rtn->rds_tcp_listen_sock);
 342         rtn->rds_tcp_listen_sock = NULL;
 343         flush_work(&rtn->rds_tcp_accept_w);
 344         spin_lock_irq(&rds_tcp_conn_lock);
 345         list_for_each_entry_safe(tc, _tc, &rds_tcp_conn_list, t_tcp_node) {
 346                 struct net *c_net = read_pnet(&tc->conn->c_net);
 347
 348                 if (net != c_net || !tc->t_sock)
 349                         continue;
 350                 list_move_tail(&tc->t_tcp_node, &tmp_list);
 351         }
 352         spin_unlock_irq(&rds_tcp_conn_lock);
 353         list_for_each_entry_safe(tc, _tc, &tmp_list, t_tcp_node) {
 354                 sk = tc->t_sock->sk;
 355                 sk->sk_prot->disconnect(sk, 0);
 356                 tcp_done(sk);
 357                 if (tc->conn->c_passive)
 358                         rds_conn_destroy(tc->conn->c_passive);
 359                 rds_conn_destroy(tc->conn);
 360         }
 361 }
 362
 363 static int rds_tcp_dev_event(struct notifier_block *this,
 364                              unsigned long event, void *ptr)
 365 {
 366         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 367
 368         /* rds-tcp registers as a pernet subys, so the ->exit will only
 369          * get invoked after network acitivity has quiesced. We need to
 370          * clean up all sockets  to quiesce network activity, and use
 371          * the unregistration of the per-net loopback device as a trigger
 372          * to start that cleanup.
 373          */
 374         if (event == NETDEV_UNREGISTER_FINAL &&
 375             dev->ifindex == LOOPBACK_IFINDEX)
 376                 rds_tcp_kill_sock(dev_net(dev));
 377
 378         return NOTIFY_DONE;
 379 }
 380
 381 static struct notifier_block rds_tcp_dev_notifier = {
 382         .notifier_call        = rds_tcp_dev_event,
 383         .priority = -10, /* must be called after other network notifiers */
 384 };
 385
 386 static void rds_tcp_exit(void)
 387 {
 388         rds_info_deregister_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 389         unregister_pernet_subsys(&rds_tcp_net_ops);
 390         if (unregister_netdevice_notifier(&rds_tcp_dev_notifier))
 391                 pr_warn("could not unregister rds_tcp_dev_notifier\n");
 392         rds_tcp_destroy_conns();
 393         rds_trans_unregister(&rds_tcp_transport);
 394         rds_tcp_recv_exit();
 395         kmem_cache_destroy(rds_tcp_conn_slab);
 396 }
 397 module_exit(rds_tcp_exit);
 398
 399 static int rds_tcp_init(void)
 400 {
 401         int ret;
 402
 403         rds_tcp_conn_slab = kmem_cache_create("rds_tcp_connection",
 404                                               sizeof(struct rds_tcp_connection),
 405                                               0, 0, NULL);
 406         if (!rds_tcp_conn_slab) {
 407                 ret = -ENOMEM;
 408                 goto out;
 409         }
 410
 411         ret = register_netdevice_notifier(&rds_tcp_dev_notifier);
 412         if (ret) {
 413                 pr_warn("could not register rds_tcp_dev_notifier\n");
 414                 goto out;
 415         }
 416
 417         ret = register_pernet_subsys(&rds_tcp_net_ops);
 418         if (ret)
 419                 goto out_slab;
 420
 421         ret = rds_tcp_recv_init();
 422         if (ret)
 423                 goto out_slab;
 424
 425         ret = rds_trans_register(&rds_tcp_transport);
 426         if (ret)
 427                 goto out_recv;
 428
 429         rds_info_register_func(RDS_INFO_TCP_SOCKETS, rds_tcp_tc_info);
 430
 431         goto out;
 432
 433 out_recv:
 434         rds_tcp_recv_exit();
 435 out_slab:
 436         unregister_pernet_subsys(&rds_tcp_net_ops);
 437         kmem_cache_destroy(rds_tcp_conn_slab);
 438 out:
 439         return ret;
 440 }
 441 module_init(rds_tcp_init);
 442
 443 MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
 444 MODULE_DESCRIPTION("RDS: TCP transport");
 445 MODULE_LICENSE("Dual BSD/GPL");
 446