lnet/klnds/socklnd/socklnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/klnds/socklnd/socklnd.c
  33  *
  34  * Author: Zach Brown <zab@zabbo.net>
  35  * Author: Peter J. Braam <braam@clusterfs.com>
  36  * Author: Phil Schwan <phil@clusterfs.com>
  37  * Author: Eric Barton <eric@bartonsoftware.com>
  38  */
  39
  40 #include <linux/inetdevice.h>
  41 #include "socklnd.h"
  42
  43 static const struct lnet_lnd the_ksocklnd;
  44 struct ksock_nal_data ksocknal_data;
  45
  46 static struct ksock_interface *
  47 ksocknal_ip2iface(struct lnet_ni *ni, __u32 ip)
  48 {
  49         struct ksock_net *net = ni->ni_data;
  50         int i;
  51         struct ksock_interface *iface;
  52
  53         for (i = 0; i < net->ksnn_ninterfaces; i++) {
  54                 LASSERT(i < LNET_INTERFACES_NUM);
  55                 iface = &net->ksnn_interfaces[i];
  56
  57                 if (iface->ksni_ipaddr == ip)
  58                         return iface;
  59         }
  60
  61         return NULL;
  62 }
  63
  64 static struct ksock_interface *
  65 ksocknal_index2iface(struct lnet_ni *ni, int index)
  66 {
  67         struct ksock_net *net = ni->ni_data;
  68         int i;
  69         struct ksock_interface *iface;
  70
  71         for (i = 0; i < net->ksnn_ninterfaces; i++) {
  72                 LASSERT(i < LNET_INTERFACES_NUM);
  73                 iface = &net->ksnn_interfaces[i];
  74
  75                 if (iface->ksni_index == index)
  76                         return iface;
  77         }
  78
  79         return NULL;
  80 }
  81
  82 static int ksocknal_ip2index(__u32 ipaddress, struct lnet_ni *ni)
  83 {
  84         struct net_device *dev;
  85         int ret = -1;
  86         DECLARE_CONST_IN_IFADDR(ifa);
  87
  88         rcu_read_lock();
  89         for_each_netdev(ni->ni_net_ns, dev) {
  90                 int flags = dev_get_flags(dev);
  91                 struct in_device *in_dev;
  92
  93                 if (flags & IFF_LOOPBACK) /* skip the loopback IF */
  94                         continue;
  95
  96                 if (!(flags & IFF_UP))
  97                         continue;
  98
  99                 in_dev = __in_dev_get_rcu(dev);
 100                 if (!in_dev)
 101                         continue;
 102
 103                 in_dev_for_each_ifa_rcu(ifa, in_dev) {
 104                         if (ntohl(ifa->ifa_local) == ipaddress)
 105                                 ret = dev->ifindex;
 106                 }
 107                 endfor_ifa(in_dev);
 108                 if (ret >= 0)
 109                         break;
 110         }
 111         rcu_read_unlock();
 112
 113         return ret;
 114 }
 115
 116 static struct ksock_route *
 117 ksocknal_create_route(__u32 ipaddr, int port)
 118 {
 119         struct ksock_route *route;
 120
 121         LIBCFS_ALLOC (route, sizeof (*route));
 122         if (route == NULL)
 123                 return (NULL);
 124
 125         atomic_set (&route->ksnr_refcount, 1);
 126         route->ksnr_peer = NULL;
 127         route->ksnr_retry_interval = 0;         /* OK to connect at any time */
 128         route->ksnr_ipaddr = ipaddr;
 129         route->ksnr_myiface = -1;
 130         route->ksnr_port = port;
 131         route->ksnr_scheduled = 0;
 132         route->ksnr_connecting = 0;
 133         route->ksnr_connected = 0;
 134         route->ksnr_deleted = 0;
 135         route->ksnr_conn_count = 0;
 136         route->ksnr_share_count = 0;
 137
 138         return route;
 139 }
 140
 141 void
 142 ksocknal_destroy_route(struct ksock_route *route)
 143 {
 144         LASSERT (atomic_read(&route->ksnr_refcount) == 0);
 145
 146         if (route->ksnr_peer != NULL)
 147                 ksocknal_peer_decref(route->ksnr_peer);
 148
 149         LIBCFS_FREE (route, sizeof (*route));
 150 }
 151
 152 static struct ksock_peer_ni *
 153 ksocknal_create_peer(struct lnet_ni *ni, struct lnet_process_id id)
 154 {
 155         int cpt = lnet_cpt_of_nid(id.nid, ni);
 156         struct ksock_net *net = ni->ni_data;
 157         struct ksock_peer_ni *peer_ni;
 158
 159         LASSERT(id.nid != LNET_NID_ANY);
 160         LASSERT(id.pid != LNET_PID_ANY);
 161         LASSERT(!in_interrupt());
 162
 163         if (!atomic_inc_unless_negative(&net->ksnn_npeers)) {
 164                 CERROR("Can't create peer_ni: network shutdown\n");
 165                 return ERR_PTR(-ESHUTDOWN);
 166         }
 167
 168         LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
 169         if (!peer_ni) {
 170                 atomic_dec(&net->ksnn_npeers);
 171                 return ERR_PTR(-ENOMEM);
 172         }
 173
 174         peer_ni->ksnp_ni = ni;
 175         peer_ni->ksnp_id = id;
 176         atomic_set(&peer_ni->ksnp_refcount, 1); /* 1 ref for caller */
 177         peer_ni->ksnp_closing = 0;
 178         peer_ni->ksnp_accepting = 0;
 179         peer_ni->ksnp_proto = NULL;
 180         peer_ni->ksnp_last_alive = 0;
 181         peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 182
 183         INIT_LIST_HEAD(&peer_ni->ksnp_conns);
 184         INIT_LIST_HEAD(&peer_ni->ksnp_routes);
 185         INIT_LIST_HEAD(&peer_ni->ksnp_tx_queue);
 186         INIT_LIST_HEAD(&peer_ni->ksnp_zc_req_list);
 187         spin_lock_init(&peer_ni->ksnp_lock);
 188
 189         return peer_ni;
 190 }
 191
 192 void
 193 ksocknal_destroy_peer(struct ksock_peer_ni *peer_ni)
 194 {
 195         struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
 196
 197         CDEBUG (D_NET, "peer_ni %s %p deleted\n",
 198                 libcfs_id2str(peer_ni->ksnp_id), peer_ni);
 199
 200         LASSERT(atomic_read(&peer_ni->ksnp_refcount) == 0);
 201         LASSERT(peer_ni->ksnp_accepting == 0);
 202         LASSERT(list_empty(&peer_ni->ksnp_conns));
 203         LASSERT(list_empty(&peer_ni->ksnp_routes));
 204         LASSERT(list_empty(&peer_ni->ksnp_tx_queue));
 205         LASSERT(list_empty(&peer_ni->ksnp_zc_req_list));
 206
 207         LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
 208
 209         /* NB a peer_ni's connections and routes keep a reference on their
 210          * peer_ni until they are destroyed, so we can be assured that _all_
 211          * state to do with this peer_ni has been cleaned up when its refcount
 212          * drops to zero.
 213          */
 214         if (atomic_dec_and_test(&net->ksnn_npeers))
 215                 wake_up_var(&net->ksnn_npeers);
 216 }
 217
 218 struct ksock_peer_ni *
 219 ksocknal_find_peer_locked(struct lnet_ni *ni, struct lnet_process_id id)
 220 {
 221         struct ksock_peer_ni *peer_ni;
 222
 223         hash_for_each_possible(ksocknal_data.ksnd_peers, peer_ni,
 224                                ksnp_list, id.nid) {
 225                 LASSERT(!peer_ni->ksnp_closing);
 226
 227                 if (peer_ni->ksnp_ni != ni)
 228                         continue;
 229
 230                 if (peer_ni->ksnp_id.nid != id.nid ||
 231                     peer_ni->ksnp_id.pid != id.pid)
 232                         continue;
 233
 234                 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d)\n",
 235                        peer_ni, libcfs_id2str(id),
 236                        atomic_read(&peer_ni->ksnp_refcount));
 237                 return peer_ni;
 238         }
 239         return NULL;
 240 }
 241
 242 struct ksock_peer_ni *
 243 ksocknal_find_peer(struct lnet_ni *ni, struct lnet_process_id id)
 244 {
 245         struct ksock_peer_ni *peer_ni;
 246
 247         read_lock(&ksocknal_data.ksnd_global_lock);
 248         peer_ni = ksocknal_find_peer_locked(ni, id);
 249         if (peer_ni != NULL)                    /* +1 ref for caller? */
 250                 ksocknal_peer_addref(peer_ni);
 251         read_unlock(&ksocknal_data.ksnd_global_lock);
 252
 253         return (peer_ni);
 254 }
 255
 256 static void
 257 ksocknal_unlink_peer_locked(struct ksock_peer_ni *peer_ni)
 258 {
 259         int i;
 260         __u32 ip;
 261         struct ksock_interface *iface;
 262
 263         for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++) {
 264                 LASSERT(i < LNET_INTERFACES_NUM);
 265                 ip = peer_ni->ksnp_passive_ips[i];
 266
 267                 iface = ksocknal_ip2iface(peer_ni->ksnp_ni, ip);
 268                 /*
 269                  * All IPs in peer_ni->ksnp_passive_ips[] come from the
 270                  * interface list, therefore the call must succeed.
 271                  */
 272                 LASSERT(iface != NULL);
 273
 274                 CDEBUG(D_NET, "peer_ni=%p iface=%p ksni_nroutes=%d\n",
 275                        peer_ni, iface, iface->ksni_nroutes);
 276                 iface->ksni_npeers--;
 277         }
 278
 279         LASSERT(list_empty(&peer_ni->ksnp_conns));
 280         LASSERT(list_empty(&peer_ni->ksnp_routes));
 281         LASSERT(!peer_ni->ksnp_closing);
 282         peer_ni->ksnp_closing = 1;
 283         hlist_del(&peer_ni->ksnp_list);
 284         /* lose peerlist's ref */
 285         ksocknal_peer_decref(peer_ni);
 286 }
 287
 288 static int
 289 ksocknal_get_peer_info(struct lnet_ni *ni, int index,
 290                        struct lnet_process_id *id, __u32 *myip, __u32 *peer_ip,
 291                        int *port, int *conn_count, int *share_count)
 292 {
 293         struct ksock_peer_ni *peer_ni;
 294         struct ksock_route *route;
 295         struct list_head *rtmp;
 296         int i;
 297         int j;
 298         int rc = -ENOENT;
 299
 300         read_lock(&ksocknal_data.ksnd_global_lock);
 301
 302         hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
 303
 304                 if (peer_ni->ksnp_ni != ni)
 305                         continue;
 306
 307                 if (peer_ni->ksnp_n_passive_ips == 0 &&
 308                     list_empty(&peer_ni->ksnp_routes)) {
 309                         if (index-- > 0)
 310                                 continue;
 311
 312                         *id = peer_ni->ksnp_id;
 313                         *myip = 0;
 314                         *peer_ip = 0;
 315                         *port = 0;
 316                         *conn_count = 0;
 317                         *share_count = 0;
 318                         rc = 0;
 319                         goto out;
 320                 }
 321
 322                 for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++) {
 323                         if (index-- > 0)
 324                                 continue;
 325
 326                         *id = peer_ni->ksnp_id;
 327                         *myip = peer_ni->ksnp_passive_ips[j];
 328                         *peer_ip = 0;
 329                         *port = 0;
 330                         *conn_count = 0;
 331                         *share_count = 0;
 332                         rc = 0;
 333                         goto out;
 334                 }
 335
 336                 list_for_each(rtmp, &peer_ni->ksnp_routes) {
 337                         if (index-- > 0)
 338                                 continue;
 339
 340                         route = list_entry(rtmp, struct ksock_route,
 341                                            ksnr_list);
 342
 343                         *id = peer_ni->ksnp_id;
 344                         rc = choose_ipv4_src(myip, route->ksnr_myiface,
 345                                              route->ksnr_ipaddr,
 346                                              ni->ni_net_ns);
 347                         *peer_ip = route->ksnr_ipaddr;
 348                         *port = route->ksnr_port;
 349                         *conn_count = route->ksnr_conn_count;
 350                         *share_count = route->ksnr_share_count;
 351                         goto out;
 352                 }
 353         }
 354 out:
 355         read_unlock(&ksocknal_data.ksnd_global_lock);
 356         return rc;
 357 }
 358
 359 static void
 360 ksocknal_associate_route_conn_locked(struct ksock_route *route,
 361                                      struct ksock_conn *conn)
 362 {
 363         struct ksock_peer_ni *peer_ni = route->ksnr_peer;
 364         int type = conn->ksnc_type;
 365         struct ksock_interface *iface;
 366         int conn_iface = ksocknal_ip2index(conn->ksnc_myipaddr,
 367                                            route->ksnr_peer->ksnp_ni);
 368
 369         conn->ksnc_route = route;
 370         ksocknal_route_addref(route);
 371
 372         if (route->ksnr_myiface != conn_iface) {
 373                 if (route->ksnr_myiface < 0) {
 374                         /* route wasn't bound locally yet (the initial route) */
 375                         CDEBUG(D_NET, "Binding %s %pI4h to interface %d\n",
 376                                libcfs_id2str(peer_ni->ksnp_id),
 377                                &route->ksnr_ipaddr,
 378                                conn_iface);
 379                 } else {
 380                         CDEBUG(D_NET,
 381                                "Rebinding %s %pI4h from interface %d to %d\n",
 382                                libcfs_id2str(peer_ni->ksnp_id),
 383                                &route->ksnr_ipaddr,
 384                                route->ksnr_myiface,
 385                                conn_iface);
 386
 387                         iface = ksocknal_index2iface(route->ksnr_peer->ksnp_ni,
 388                                                      route->ksnr_myiface);
 389                         if (iface)
 390                                 iface->ksni_nroutes--;
 391                 }
 392                 route->ksnr_myiface = conn_iface;
 393                 iface = ksocknal_index2iface(route->ksnr_peer->ksnp_ni,
 394                                              route->ksnr_myiface);
 395                 if (iface)
 396                         iface->ksni_nroutes++;
 397         }
 398
 399         route->ksnr_connected |= (1<<type);
 400         route->ksnr_conn_count++;
 401
 402         /* Successful connection => further attempts can
 403          * proceed immediately
 404          */
 405         route->ksnr_retry_interval = 0;
 406 }
 407
 408 static void
 409 ksocknal_add_route_locked(struct ksock_peer_ni *peer_ni, struct ksock_route *route)
 410 {
 411         struct list_head *tmp;
 412         struct ksock_conn *conn;
 413         struct ksock_route *route2;
 414         struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
 415
 416         LASSERT(!peer_ni->ksnp_closing);
 417         LASSERT(route->ksnr_peer == NULL);
 418         LASSERT(!route->ksnr_scheduled);
 419         LASSERT(!route->ksnr_connecting);
 420         LASSERT(route->ksnr_connected == 0);
 421         LASSERT(net->ksnn_ninterfaces > 0);
 422
 423         /* LASSERT(unique) */
 424         list_for_each(tmp, &peer_ni->ksnp_routes) {
 425                 route2 = list_entry(tmp, struct ksock_route, ksnr_list);
 426
 427                 if (route2->ksnr_ipaddr == route->ksnr_ipaddr) {
 428                         CERROR("Duplicate route %s %pI4h\n",
 429                                libcfs_id2str(peer_ni->ksnp_id),
 430                                &route->ksnr_ipaddr);
 431                         LBUG();
 432                 }
 433         }
 434
 435         route->ksnr_peer = peer_ni;
 436         ksocknal_peer_addref(peer_ni);
 437
 438         /* set the route's interface to the current net's interface */
 439         route->ksnr_myiface = net->ksnn_interfaces[0].ksni_index;
 440         net->ksnn_interfaces[0].ksni_nroutes++;
 441
 442         /* peer_ni's routelist takes over my ref on 'route' */
 443         list_add_tail(&route->ksnr_list, &peer_ni->ksnp_routes);
 444
 445         list_for_each(tmp, &peer_ni->ksnp_conns) {
 446                 conn = list_entry(tmp, struct ksock_conn, ksnc_list);
 447
 448                 if (conn->ksnc_ipaddr != route->ksnr_ipaddr)
 449                         continue;
 450
 451                 ksocknal_associate_route_conn_locked(route, conn);
 452                 /* keep going (typed routes) */
 453         }
 454 }
 455
 456 static void
 457 ksocknal_del_route_locked(struct ksock_route *route)
 458 {
 459         struct ksock_peer_ni *peer_ni = route->ksnr_peer;
 460         struct ksock_interface *iface;
 461         struct ksock_conn *conn;
 462         struct list_head *ctmp;
 463         struct list_head *cnxt;
 464
 465         LASSERT(!route->ksnr_deleted);
 466
 467         /* Close associated conns */
 468         list_for_each_safe(ctmp, cnxt, &peer_ni->ksnp_conns) {
 469                 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
 470
 471                 if (conn->ksnc_route != route)
 472                         continue;
 473
 474                 ksocknal_close_conn_locked(conn, 0);
 475         }
 476
 477         if (route->ksnr_myiface >= 0) {
 478                 iface = ksocknal_index2iface(route->ksnr_peer->ksnp_ni,
 479                                              route->ksnr_myiface);
 480                 if (iface)
 481                         iface->ksni_nroutes--;
 482         }
 483
 484         route->ksnr_deleted = 1;
 485         list_del(&route->ksnr_list);
 486         ksocknal_route_decref(route);           /* drop peer_ni's ref */
 487
 488         if (list_empty(&peer_ni->ksnp_routes) &&
 489             list_empty(&peer_ni->ksnp_conns)) {
 490                 /* I've just removed the last route to a peer_ni with no active
 491                  * connections */
 492                 ksocknal_unlink_peer_locked(peer_ni);
 493         }
 494 }
 495
 496 int
 497 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ipaddr,
 498                   int port)
 499 {
 500         struct list_head *tmp;
 501         struct ksock_peer_ni *peer_ni;
 502         struct ksock_peer_ni *peer2;
 503         struct ksock_route *route;
 504         struct ksock_route *route2;
 505
 506         if (id.nid == LNET_NID_ANY ||
 507             id.pid == LNET_PID_ANY)
 508                 return (-EINVAL);
 509
 510         /* Have a brand new peer_ni ready... */
 511         peer_ni = ksocknal_create_peer(ni, id);
 512         if (IS_ERR(peer_ni))
 513                 return PTR_ERR(peer_ni);
 514
 515         route = ksocknal_create_route (ipaddr, port);
 516         if (route == NULL) {
 517                 ksocknal_peer_decref(peer_ni);
 518                 return (-ENOMEM);
 519         }
 520
 521         write_lock_bh(&ksocknal_data.ksnd_global_lock);
 522
 523         /* always called with a ref on ni, so shutdown can't have started */
 524         LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers)
 525                 >= 0);
 526
 527         peer2 = ksocknal_find_peer_locked(ni, id);
 528         if (peer2 != NULL) {
 529                 ksocknal_peer_decref(peer_ni);
 530                 peer_ni = peer2;
 531         } else {
 532                 /* peer_ni table takes my ref on peer_ni */
 533                 hash_add(ksocknal_data.ksnd_peers, &peer_ni->ksnp_list, id.nid);
 534         }
 535
 536         route2 = NULL;
 537         list_for_each(tmp, &peer_ni->ksnp_routes) {
 538                 route2 = list_entry(tmp, struct ksock_route, ksnr_list);
 539
 540                 if (route2->ksnr_ipaddr == ipaddr)
 541                         break;
 542
 543                 route2 = NULL;
 544         }
 545         if (route2 == NULL) {
 546                 ksocknal_add_route_locked(peer_ni, route);
 547                 route->ksnr_share_count++;
 548         } else {
 549                 ksocknal_route_decref(route);
 550                 route2->ksnr_share_count++;
 551         }
 552
 553         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 554
 555         return 0;
 556 }
 557
 558 static void
 559 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni, __u32 ip)
 560 {
 561         struct ksock_conn *conn;
 562         struct ksock_route *route;
 563         struct list_head *tmp;
 564         struct list_head *nxt;
 565         int nshared;
 566
 567         LASSERT(!peer_ni->ksnp_closing);
 568
 569         /* Extra ref prevents peer_ni disappearing until I'm done with it */
 570         ksocknal_peer_addref(peer_ni);
 571
 572         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
 573                 route = list_entry(tmp, struct ksock_route, ksnr_list);
 574
 575                 /* no match */
 576                 if (!(ip == 0 || route->ksnr_ipaddr == ip))
 577                         continue;
 578
 579                 route->ksnr_share_count = 0;
 580                 /* This deletes associated conns too */
 581                 ksocknal_del_route_locked(route);
 582         }
 583
 584         nshared = 0;
 585         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
 586                 route = list_entry(tmp, struct ksock_route, ksnr_list);
 587                 nshared += route->ksnr_share_count;
 588         }
 589
 590         if (nshared == 0) {
 591                 /* remove everything else if there are no explicit entries
 592                  * left */
 593
 594                 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
 595                         route = list_entry(tmp, struct ksock_route, ksnr_list);
 596
 597                         /* we should only be removing auto-entries */
 598                         LASSERT(route->ksnr_share_count == 0);
 599                         ksocknal_del_route_locked(route);
 600                 }
 601
 602                 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_conns) {
 603                         conn = list_entry(tmp, struct ksock_conn, ksnc_list);
 604
 605                         ksocknal_close_conn_locked(conn, 0);
 606                 }
 607         }
 608
 609         ksocknal_peer_decref(peer_ni);
 610         /* NB peer_ni unlinks itself when last conn/route is removed */
 611 }
 612
 613 static int
 614 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ip)
 615 {
 616         LIST_HEAD(zombies);
 617         struct hlist_node *pnxt;
 618         struct ksock_peer_ni *peer_ni;
 619         int lo;
 620         int hi;
 621         int i;
 622         int rc = -ENOENT;
 623
 624         write_lock_bh(&ksocknal_data.ksnd_global_lock);
 625
 626         if (id.nid != LNET_NID_ANY) {
 627                 lo = hash_min(id.nid, HASH_BITS(ksocknal_data.ksnd_peers));
 628                 hi = lo;
 629         } else {
 630                 lo = 0;
 631                 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
 632         }
 633
 634         for (i = lo; i <= hi; i++) {
 635                 hlist_for_each_entry_safe(peer_ni, pnxt,
 636                                           &ksocknal_data.ksnd_peers[i],
 637                                           ksnp_list) {
 638                         if (peer_ni->ksnp_ni != ni)
 639                                 continue;
 640
 641                         if (!((id.nid == LNET_NID_ANY ||
 642                                peer_ni->ksnp_id.nid == id.nid) &&
 643                               (id.pid == LNET_PID_ANY ||
 644                                peer_ni->ksnp_id.pid == id.pid)))
 645                                 continue;
 646
 647                         ksocknal_peer_addref(peer_ni);  /* a ref for me... */
 648
 649                         ksocknal_del_peer_locked(peer_ni, ip);
 650
 651                         if (peer_ni->ksnp_closing &&
 652                             !list_empty(&peer_ni->ksnp_tx_queue)) {
 653                                 LASSERT(list_empty(&peer_ni->ksnp_conns));
 654                                 LASSERT(list_empty(&peer_ni->ksnp_routes));
 655
 656                                 list_splice_init(&peer_ni->ksnp_tx_queue,
 657                                                  &zombies);
 658                         }
 659
 660                         ksocknal_peer_decref(peer_ni);  /* ...till here */
 661
 662                         rc = 0;                         /* matched! */
 663                 }
 664         }
 665
 666         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 667
 668         ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
 669
 670         return rc;
 671 }
 672
 673 static struct ksock_conn *
 674 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
 675 {
 676         struct ksock_peer_ni *peer_ni;
 677         struct ksock_conn *conn;
 678         struct list_head *ctmp;
 679         int i;
 680
 681         read_lock(&ksocknal_data.ksnd_global_lock);
 682
 683         hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
 684                 LASSERT(!peer_ni->ksnp_closing);
 685
 686                 if (peer_ni->ksnp_ni != ni)
 687                         continue;
 688
 689                 list_for_each(ctmp, &peer_ni->ksnp_conns) {
 690                         if (index-- > 0)
 691                                 continue;
 692
 693                         conn = list_entry(ctmp, struct ksock_conn,
 694                                           ksnc_list);
 695                         ksocknal_conn_addref(conn);
 696                         read_unlock(&ksocknal_data.ksnd_global_lock);
 697                         return conn;
 698                 }
 699         }
 700
 701         read_unlock(&ksocknal_data.ksnd_global_lock);
 702         return NULL;
 703 }
 704
 705 static struct ksock_sched *
 706 ksocknal_choose_scheduler_locked(unsigned int cpt)
 707 {
 708         struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
 709         int i;
 710
 711         if (sched->kss_nthreads == 0) {
 712                 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
 713                         if (sched->kss_nthreads > 0) {
 714                                 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 715                                        cpt, sched->kss_cpt);
 716                                 return sched;
 717                         }
 718                 }
 719                 return NULL;
 720         }
 721
 722         return sched;
 723 }
 724
 725 static int
 726 ksocknal_local_ipvec(struct lnet_ni *ni, __u32 *ipaddrs)
 727 {
 728         struct ksock_net *net = ni->ni_data;
 729         int i;
 730         int nip;
 731
 732         read_lock(&ksocknal_data.ksnd_global_lock);
 733
 734         nip = net->ksnn_ninterfaces;
 735         LASSERT(nip <= LNET_INTERFACES_NUM);
 736
 737         /*
 738          * Only offer interfaces for additional connections if I have
 739          * more than one.
 740          */
 741         if (nip < 2) {
 742                 read_unlock(&ksocknal_data.ksnd_global_lock);
 743                 return 0;
 744         }
 745
 746         for (i = 0; i < nip; i++) {
 747                 ipaddrs[i] = net->ksnn_interfaces[i].ksni_ipaddr;
 748                 LASSERT(ipaddrs[i] != 0);
 749         }
 750
 751         read_unlock(&ksocknal_data.ksnd_global_lock);
 752         return nip;
 753 }
 754
 755 static int
 756 ksocknal_match_peerip(struct ksock_interface *iface, __u32 *ips, int nips)
 757 {
 758         int best_netmatch = 0;
 759         int best_xor = 0;
 760         int best = -1;
 761         int this_xor;
 762         int this_netmatch;
 763         int i;
 764
 765         for (i = 0; i < nips; i++) {
 766                 if (ips[i] == 0)
 767                         continue;
 768
 769                 this_xor = (ips[i] ^ iface->ksni_ipaddr);
 770                 this_netmatch = ((this_xor & iface->ksni_netmask) == 0) ? 1 : 0;
 771
 772                 if (!(best < 0 ||
 773                       best_netmatch < this_netmatch ||
 774                       (best_netmatch == this_netmatch &&
 775                        best_xor > this_xor)))
 776                         continue;
 777
 778                 best = i;
 779                 best_netmatch = this_netmatch;
 780                 best_xor = this_xor;
 781         }
 782
 783         LASSERT (best >= 0);
 784         return (best);
 785 }
 786
 787 static int
 788 ksocknal_select_ips(struct ksock_peer_ni *peer_ni, __u32 *peerips, int n_peerips)
 789 {
 790         rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
 791         struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
 792         struct ksock_interface *iface;
 793         struct ksock_interface *best_iface;
 794         int n_ips;
 795         int i;
 796         int j;
 797         int k;
 798         u32 ip;
 799         u32 xor;
 800         int this_netmatch;
 801         int best_netmatch;
 802         int best_npeers;
 803
 804         /* CAVEAT EMPTOR: We do all our interface matching with an
 805          * exclusive hold of global lock at IRQ priority.  We're only
 806          * expecting to be dealing with small numbers of interfaces, so the
 807          * O(n**3)-ness shouldn't matter */
 808
 809         /* Also note that I'm not going to return more than n_peerips
 810          * interfaces, even if I have more myself */
 811
 812         write_lock_bh(global_lock);
 813
 814         LASSERT(n_peerips <= LNET_INTERFACES_NUM);
 815         LASSERT(net->ksnn_ninterfaces <= LNET_INTERFACES_NUM);
 816
 817         /* Only match interfaces for additional connections
 818          * if I have > 1 interface
 819          */
 820         n_ips = (net->ksnn_ninterfaces < 2) ? 0 :
 821                 min(n_peerips, net->ksnn_ninterfaces);
 822
 823         for (i = 0; peer_ni->ksnp_n_passive_ips < n_ips; i++) {
 824                 /*              ^ yes really... */
 825
 826                 /* If we have any new interfaces, first tick off all the
 827                  * peer_ni IPs that match old interfaces, then choose new
 828                  * interfaces to match the remaining peer_ni IPS.
 829                  * We don't forget interfaces we've stopped using; we might
 830                  * start using them again... */
 831
 832                 if (i < peer_ni->ksnp_n_passive_ips) {
 833                         /* Old interface. */
 834                         ip = peer_ni->ksnp_passive_ips[i];
 835                         best_iface = ksocknal_ip2iface(peer_ni->ksnp_ni, ip);
 836
 837                         /* peer_ni passive ips are kept up to date */
 838                         LASSERT(best_iface != NULL);
 839                 } else {
 840                         /* choose a new interface */
 841                         LASSERT (i == peer_ni->ksnp_n_passive_ips);
 842
 843                         best_iface = NULL;
 844                         best_netmatch = 0;
 845                         best_npeers = 0;
 846
 847                         for (j = 0; j < net->ksnn_ninterfaces; j++) {
 848                                 iface = &net->ksnn_interfaces[j];
 849                                 ip = iface->ksni_ipaddr;
 850
 851                                 for (k = 0; k < peer_ni->ksnp_n_passive_ips; k++)
 852                                         if (peer_ni->ksnp_passive_ips[k] == ip)
 853                                                 break;
 854
 855                                 if (k < peer_ni->ksnp_n_passive_ips) /* using it already */
 856                                         continue;
 857
 858                                 k = ksocknal_match_peerip(iface, peerips, n_peerips);
 859                                 xor = (ip ^ peerips[k]);
 860                                 this_netmatch = ((xor & iface->ksni_netmask) == 0) ? 1 : 0;
 861
 862                                 if (!(best_iface == NULL ||
 863                                       best_netmatch < this_netmatch ||
 864                                       (best_netmatch == this_netmatch &&
 865                                        best_npeers > iface->ksni_npeers)))
 866                                         continue;
 867
 868                                 best_iface = iface;
 869                                 best_netmatch = this_netmatch;
 870                                 best_npeers = iface->ksni_npeers;
 871                         }
 872
 873                         LASSERT(best_iface != NULL);
 874
 875                         best_iface->ksni_npeers++;
 876                         ip = best_iface->ksni_ipaddr;
 877                         peer_ni->ksnp_passive_ips[i] = ip;
 878                         peer_ni->ksnp_n_passive_ips = i+1;
 879                 }
 880
 881                 /* mark the best matching peer_ni IP used */
 882                 j = ksocknal_match_peerip(best_iface, peerips, n_peerips);
 883                 peerips[j] = 0;
 884         }
 885
 886         /* Overwrite input peer_ni IP addresses */
 887         memcpy(peerips, peer_ni->ksnp_passive_ips, n_ips * sizeof(*peerips));
 888
 889         write_unlock_bh(global_lock);
 890
 891         return (n_ips);
 892 }
 893
 894 static void
 895 ksocknal_create_routes(struct ksock_peer_ni *peer_ni, int port,
 896                        __u32 *peer_ipaddrs, int npeer_ipaddrs)
 897 {
 898         struct ksock_route              *newroute = NULL;
 899         rwlock_t                *global_lock = &ksocknal_data.ksnd_global_lock;
 900         struct lnet_ni *ni = peer_ni->ksnp_ni;
 901         struct ksock_net                *net = ni->ni_data;
 902         struct list_head        *rtmp;
 903         struct ksock_route              *route;
 904         struct ksock_interface  *iface;
 905         struct ksock_interface  *best_iface;
 906         int                     best_netmatch;
 907         int                     this_netmatch;
 908         int                     best_nroutes;
 909         int                     i;
 910         int                     j;
 911
 912         /* CAVEAT EMPTOR: We do all our interface matching with an
 913          * exclusive hold of global lock at IRQ priority.  We're only
 914          * expecting to be dealing with small numbers of interfaces, so the
 915          * O(n**3)-ness here shouldn't matter */
 916
 917         write_lock_bh(global_lock);
 918
 919         if (net->ksnn_ninterfaces < 2) {
 920                 /* Only create additional connections
 921                  * if I have > 1 interface */
 922                 write_unlock_bh(global_lock);
 923                 return;
 924         }
 925
 926         LASSERT(npeer_ipaddrs <= LNET_INTERFACES_NUM);
 927
 928         for (i = 0; i < npeer_ipaddrs; i++) {
 929                 if (newroute != NULL) {
 930                         newroute->ksnr_ipaddr = peer_ipaddrs[i];
 931                 } else {
 932                         write_unlock_bh(global_lock);
 933
 934                         newroute = ksocknal_create_route(peer_ipaddrs[i], port);
 935                         if (newroute == NULL)
 936                                 return;
 937
 938                         write_lock_bh(global_lock);
 939                 }
 940
 941                 if (peer_ni->ksnp_closing) {
 942                         /* peer_ni got closed under me */
 943                         break;
 944                 }
 945
 946                 /* Already got a route? */
 947                 route = NULL;
 948                 list_for_each(rtmp, &peer_ni->ksnp_routes) {
 949                         route = list_entry(rtmp, struct ksock_route, ksnr_list);
 950
 951                         if (route->ksnr_ipaddr == newroute->ksnr_ipaddr)
 952                                 break;
 953
 954                         route = NULL;
 955                 }
 956                 if (route != NULL)
 957                         continue;
 958
 959                 best_iface = NULL;
 960                 best_nroutes = 0;
 961                 best_netmatch = 0;
 962
 963                 LASSERT(net->ksnn_ninterfaces <= LNET_INTERFACES_NUM);
 964
 965                 /* Select interface to connect from */
 966                 for (j = 0; j < net->ksnn_ninterfaces; j++) {
 967                         iface = &net->ksnn_interfaces[j];
 968
 969                         /* Using this interface already? */
 970                         list_for_each(rtmp, &peer_ni->ksnp_routes) {
 971                                 route = list_entry(rtmp, struct ksock_route,
 972                                                    ksnr_list);
 973
 974                                 if (route->ksnr_myiface == iface->ksni_index)
 975                                         break;
 976
 977                                 route = NULL;
 978                         }
 979                         if (route != NULL)
 980                                 continue;
 981
 982                         this_netmatch = (((iface->ksni_ipaddr ^
 983                                            newroute->ksnr_ipaddr) &
 984                                           iface->ksni_netmask) == 0) ? 1 : 0;
 985
 986                         if (!(best_iface == NULL ||
 987                               best_netmatch < this_netmatch ||
 988                               (best_netmatch == this_netmatch &&
 989                                best_nroutes > iface->ksni_nroutes)))
 990                                 continue;
 991
 992                         best_iface = iface;
 993                         best_netmatch = this_netmatch;
 994                         best_nroutes = iface->ksni_nroutes;
 995                 }
 996
 997                 if (best_iface == NULL)
 998                         continue;
 999
1000                 newroute->ksnr_myiface = best_iface->ksni_index;
1001                 best_iface->ksni_nroutes++;
1002
1003                 ksocknal_add_route_locked(peer_ni, newroute);
1004                 newroute = NULL;
1005         }
1006
1007         write_unlock_bh(global_lock);
1008         if (newroute != NULL)
1009                 ksocknal_route_decref(newroute);
1010 }
1011
1012 int
1013 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
1014 {
1015         struct ksock_connreq *cr;
1016         int rc;
1017         u32 peer_ip;
1018         int peer_port;
1019
1020         rc = lnet_sock_getaddr(sock, true, &peer_ip, &peer_port);
1021         LASSERT(rc == 0);               /* we succeeded before */
1022
1023         LIBCFS_ALLOC(cr, sizeof(*cr));
1024         if (cr == NULL) {
1025                 LCONSOLE_ERROR_MSG(0x12f, "Dropping connection request from "
1026                                    "%pI4h: memory exhausted\n", &peer_ip);
1027                 return -ENOMEM;
1028         }
1029
1030         lnet_ni_addref(ni);
1031         cr->ksncr_ni   = ni;
1032         cr->ksncr_sock = sock;
1033
1034         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
1035
1036         list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
1037         wake_up(&ksocknal_data.ksnd_connd_waitq);
1038
1039         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
1040         return 0;
1041 }
1042
1043 static int
1044 ksocknal_connecting(struct ksock_peer_ni *peer_ni, __u32 ipaddr)
1045 {
1046         struct ksock_route *route;
1047
1048         list_for_each_entry(route, &peer_ni->ksnp_routes, ksnr_list) {
1049                 if (route->ksnr_ipaddr == ipaddr)
1050                         return route->ksnr_connecting;
1051         }
1052         return 0;
1053 }
1054
1055 int
1056 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_route *route,
1057                      struct socket *sock, int type)
1058 {
1059         rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
1060         LIST_HEAD(zombies);
1061         struct lnet_process_id peerid;
1062         struct list_head *tmp;
1063         u64 incarnation;
1064         struct ksock_conn *conn;
1065         struct ksock_conn *conn2;
1066         struct ksock_peer_ni *peer_ni = NULL;
1067         struct ksock_peer_ni *peer2;
1068         struct ksock_sched *sched;
1069         struct ksock_hello_msg *hello;
1070         int cpt;
1071         struct ksock_tx *tx;
1072         struct ksock_tx *txtmp;
1073         int rc;
1074         int rc2;
1075         int active;
1076         char *warn = NULL;
1077
1078         active = (route != NULL);
1079
1080         LASSERT (active == (type != SOCKLND_CONN_NONE));
1081
1082         LIBCFS_ALLOC(conn, sizeof(*conn));
1083         if (conn == NULL) {
1084                 rc = -ENOMEM;
1085                 goto failed_0;
1086         }
1087
1088         conn->ksnc_peer = NULL;
1089         conn->ksnc_route = NULL;
1090         conn->ksnc_sock = sock;
1091         /* 2 ref, 1 for conn, another extra ref prevents socket
1092          * being closed before establishment of connection */
1093         atomic_set (&conn->ksnc_sock_refcount, 2);
1094         conn->ksnc_type = type;
1095         ksocknal_lib_save_callback(sock, conn);
1096         atomic_set (&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
1097
1098         conn->ksnc_rx_ready = 0;
1099         conn->ksnc_rx_scheduled = 0;
1100
1101         INIT_LIST_HEAD(&conn->ksnc_tx_queue);
1102         conn->ksnc_tx_ready = 0;
1103         conn->ksnc_tx_scheduled = 0;
1104         conn->ksnc_tx_carrier = NULL;
1105         atomic_set (&conn->ksnc_tx_nob, 0);
1106
1107         LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
1108                                      kshm_ips[LNET_INTERFACES_NUM]));
1109         if (hello == NULL) {
1110                 rc = -ENOMEM;
1111                 goto failed_1;
1112         }
1113
1114         /* stash conn's local and remote addrs */
1115         rc = ksocknal_lib_get_conn_addrs (conn);
1116         if (rc != 0)
1117                 goto failed_1;
1118
1119         /* Find out/confirm peer_ni's NID and connection type and get the
1120          * vector of interfaces she's willing to let me connect to.
1121          * Passive connections use the listener timeout since the peer_ni sends
1122          * eagerly */
1123
1124         if (active) {
1125                 peer_ni = route->ksnr_peer;
1126                 LASSERT(ni == peer_ni->ksnp_ni);
1127
1128                 /* Active connection sends HELLO eagerly */
1129                 hello->kshm_nips = ksocknal_local_ipvec(ni, hello->kshm_ips);
1130                 peerid = peer_ni->ksnp_id;
1131
1132                 write_lock_bh(global_lock);
1133                 conn->ksnc_proto = peer_ni->ksnp_proto;
1134                 write_unlock_bh(global_lock);
1135
1136                 if (conn->ksnc_proto == NULL) {
1137                          conn->ksnc_proto = &ksocknal_protocol_v3x;
1138 #if SOCKNAL_VERSION_DEBUG
1139                          if (*ksocknal_tunables.ksnd_protocol == 2)
1140                                  conn->ksnc_proto = &ksocknal_protocol_v2x;
1141                          else if (*ksocknal_tunables.ksnd_protocol == 1)
1142                                  conn->ksnc_proto = &ksocknal_protocol_v1x;
1143 #endif
1144                 }
1145
1146                 rc = ksocknal_send_hello (ni, conn, peerid.nid, hello);
1147                 if (rc != 0)
1148                         goto failed_1;
1149         } else {
1150                 peerid.nid = LNET_NID_ANY;
1151                 peerid.pid = LNET_PID_ANY;
1152
1153                 /* Passive, get protocol from peer_ni */
1154                 conn->ksnc_proto = NULL;
1155         }
1156
1157         rc = ksocknal_recv_hello (ni, conn, hello, &peerid, &incarnation);
1158         if (rc < 0)
1159                 goto failed_1;
1160
1161         LASSERT (rc == 0 || active);
1162         LASSERT (conn->ksnc_proto != NULL);
1163         LASSERT (peerid.nid != LNET_NID_ANY);
1164
1165         cpt = lnet_cpt_of_nid(peerid.nid, ni);
1166
1167         if (active) {
1168                 ksocknal_peer_addref(peer_ni);
1169                 write_lock_bh(global_lock);
1170         } else {
1171                 peer_ni = ksocknal_create_peer(ni, peerid);
1172                 if (IS_ERR(peer_ni)) {
1173                         rc = PTR_ERR(peer_ni);
1174                         goto failed_1;
1175                 }
1176
1177                 write_lock_bh(global_lock);
1178
1179                 /* called with a ref on ni, so shutdown can't have started */
1180                 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers) >= 0);
1181
1182                 peer2 = ksocknal_find_peer_locked(ni, peerid);
1183                 if (peer2 == NULL) {
1184                         /* NB this puts an "empty" peer_ni in the peer_ni
1185                          * table (which takes my ref) */
1186                         hash_add(ksocknal_data.ksnd_peers,
1187                                  &peer_ni->ksnp_list, peerid.nid);
1188                 } else {
1189                         ksocknal_peer_decref(peer_ni);
1190                         peer_ni = peer2;
1191                 }
1192
1193                 /* +1 ref for me */
1194                 ksocknal_peer_addref(peer_ni);
1195                 peer_ni->ksnp_accepting++;
1196
1197                 /* Am I already connecting to this guy?  Resolve in
1198                  * favour of higher NID... */
1199                 if (peerid.nid < ni->ni_nid &&
1200                     ksocknal_connecting(peer_ni, conn->ksnc_ipaddr)) {
1201                         rc = EALREADY;
1202                         warn = "connection race resolution";
1203                         goto failed_2;
1204                 }
1205         }
1206
1207         if (peer_ni->ksnp_closing ||
1208             (active && route->ksnr_deleted)) {
1209                 /* peer_ni/route got closed under me */
1210                 rc = -ESTALE;
1211                 warn = "peer_ni/route removed";
1212                 goto failed_2;
1213         }
1214
1215         if (peer_ni->ksnp_proto == NULL) {
1216                 /* Never connected before.
1217                  * NB recv_hello may have returned EPROTO to signal my peer_ni
1218                  * wants a different protocol than the one I asked for.
1219                  */
1220                 LASSERT(list_empty(&peer_ni->ksnp_conns));
1221
1222                 peer_ni->ksnp_proto = conn->ksnc_proto;
1223                 peer_ni->ksnp_incarnation = incarnation;
1224         }
1225
1226         if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1227             peer_ni->ksnp_incarnation != incarnation) {
1228                 /* peer_ni rebooted or I've got the wrong protocol version */
1229                 ksocknal_close_peer_conns_locked(peer_ni, 0, 0);
1230
1231                 peer_ni->ksnp_proto = NULL;
1232                 rc = ESTALE;
1233                 warn = peer_ni->ksnp_incarnation != incarnation ?
1234                        "peer_ni rebooted" :
1235                        "wrong proto version";
1236                 goto failed_2;
1237         }
1238
1239         switch (rc) {
1240         default:
1241                 LBUG();
1242         case 0:
1243                 break;
1244         case EALREADY:
1245                 warn = "lost conn race";
1246                 goto failed_2;
1247         case EPROTO:
1248                 warn = "retry with different protocol version";
1249                 goto failed_2;
1250         }
1251
1252         /* Refuse to duplicate an existing connection, unless this is a
1253          * loopback connection */
1254         if (conn->ksnc_ipaddr != conn->ksnc_myipaddr) {
1255                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1256                         conn2 = list_entry(tmp, struct ksock_conn, ksnc_list);
1257
1258                         if (conn2->ksnc_ipaddr != conn->ksnc_ipaddr ||
1259                             conn2->ksnc_myipaddr != conn->ksnc_myipaddr ||
1260                             conn2->ksnc_type != conn->ksnc_type)
1261                                 continue;
1262
1263                         /* Reply on a passive connection attempt so the peer_ni
1264                          * realises we're connected. */
1265                         LASSERT (rc == 0);
1266                         if (!active)
1267                                 rc = EALREADY;
1268
1269                         warn = "duplicate";
1270                         goto failed_2;
1271                 }
1272         }
1273
1274         /* If the connection created by this route didn't bind to the IP
1275          * address the route connected to, the connection/route matching
1276          * code below probably isn't going to work. */
1277         if (active &&
1278             route->ksnr_ipaddr != conn->ksnc_ipaddr) {
1279                 CERROR("Route %s %pI4h connected to %pI4h\n",
1280                        libcfs_id2str(peer_ni->ksnp_id),
1281                        &route->ksnr_ipaddr,
1282                        &conn->ksnc_ipaddr);
1283         }
1284
1285         /* Search for a route corresponding to the new connection and
1286          * create an association.  This allows incoming connections created
1287          * by routes in my peer_ni to match my own route entries so I don't
1288          * continually create duplicate routes. */
1289         list_for_each(tmp, &peer_ni->ksnp_routes) {
1290                 route = list_entry(tmp, struct ksock_route, ksnr_list);
1291
1292                 if (route->ksnr_ipaddr != conn->ksnc_ipaddr)
1293                         continue;
1294
1295                 ksocknal_associate_route_conn_locked(route, conn);
1296                 break;
1297         }
1298
1299         conn->ksnc_peer = peer_ni;                 /* conn takes my ref on peer_ni */
1300         peer_ni->ksnp_last_alive = ktime_get_seconds();
1301         peer_ni->ksnp_send_keepalive = 0;
1302         peer_ni->ksnp_error = 0;
1303
1304         sched = ksocknal_choose_scheduler_locked(cpt);
1305         if (!sched) {
1306                 CERROR("no schedulers available. node is unhealthy\n");
1307                 goto failed_2;
1308         }
1309         /*
1310          * The cpt might have changed if we ended up selecting a non cpt
1311          * native scheduler. So use the scheduler's cpt instead.
1312          */
1313         cpt = sched->kss_cpt;
1314         sched->kss_nconns++;
1315         conn->ksnc_scheduler = sched;
1316
1317         conn->ksnc_tx_last_post = ktime_get_seconds();
1318         /* Set the deadline for the outgoing HELLO to drain */
1319         conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1320         conn->ksnc_tx_deadline = ktime_get_seconds() +
1321                                  ksocknal_timeout();
1322         smp_mb();   /* order with adding to peer_ni's conn list */
1323
1324         list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1325         ksocknal_conn_addref(conn);
1326
1327         ksocknal_new_packet(conn, 0);
1328
1329         conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1330
1331         /* Take packets blocking for this connection. */
1332         list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1333                 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1334                     SOCKNAL_MATCH_NO)
1335                         continue;
1336
1337                 list_del(&tx->tx_list);
1338                 ksocknal_queue_tx_locked(tx, conn);
1339         }
1340
1341         write_unlock_bh(global_lock);
1342
1343         /* We've now got a new connection.  Any errors from here on are just
1344          * like "normal" comms errors and we close the connection normally.
1345          * NB (a) we still have to send the reply HELLO for passive
1346          *        connections,
1347          *    (b) normal I/O on the conn is blocked until I setup and call the
1348          *        socket callbacks.
1349          */
1350
1351         CDEBUG(D_NET, "New conn %s p %d.x %pI4h -> %pI4h/%d"
1352                " incarnation:%lld sched[%d]\n",
1353                libcfs_id2str(peerid), conn->ksnc_proto->pro_version,
1354                &conn->ksnc_myipaddr, &conn->ksnc_ipaddr,
1355                conn->ksnc_port, incarnation, cpt);
1356
1357         if (active) {
1358                 /* additional routes after interface exchange? */
1359                 ksocknal_create_routes(peer_ni, conn->ksnc_port,
1360                                        hello->kshm_ips, hello->kshm_nips);
1361         } else {
1362                 hello->kshm_nips = ksocknal_select_ips(peer_ni, hello->kshm_ips,
1363                                                        hello->kshm_nips);
1364                 rc = ksocknal_send_hello(ni, conn, peerid.nid, hello);
1365         }
1366
1367         LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1368                                     kshm_ips[LNET_INTERFACES_NUM]));
1369
1370         /* setup the socket AFTER I've received hello (it disables
1371          * SO_LINGER).  I might call back to the acceptor who may want
1372          * to send a protocol version response and then close the
1373          * socket; this ensures the socket only tears down after the
1374          * response has been sent. */
1375         if (rc == 0)
1376                 rc = ksocknal_lib_setup_sock(sock);
1377
1378         write_lock_bh(global_lock);
1379
1380         /* NB my callbacks block while I hold ksnd_global_lock */
1381         ksocknal_lib_set_callback(sock, conn);
1382
1383         if (!active)
1384                 peer_ni->ksnp_accepting--;
1385
1386         write_unlock_bh(global_lock);
1387
1388         if (rc != 0) {
1389                 write_lock_bh(global_lock);
1390                 if (!conn->ksnc_closing) {
1391                         /* could be closed by another thread */
1392                         ksocknal_close_conn_locked(conn, rc);
1393                 }
1394                 write_unlock_bh(global_lock);
1395         } else if (ksocknal_connsock_addref(conn) == 0) {
1396                 /* Allow I/O to proceed. */
1397                 ksocknal_read_callback(conn);
1398                 ksocknal_write_callback(conn);
1399                 ksocknal_connsock_decref(conn);
1400         }
1401
1402         ksocknal_connsock_decref(conn);
1403         ksocknal_conn_decref(conn);
1404         return rc;
1405
1406 failed_2:
1407         if (!peer_ni->ksnp_closing &&
1408             list_empty(&peer_ni->ksnp_conns) &&
1409             list_empty(&peer_ni->ksnp_routes)) {
1410                 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
1411                 ksocknal_unlink_peer_locked(peer_ni);
1412         }
1413
1414         write_unlock_bh(global_lock);
1415
1416         if (warn != NULL) {
1417                 if (rc < 0)
1418                         CERROR("Not creating conn %s type %d: %s\n",
1419                                libcfs_id2str(peerid), conn->ksnc_type, warn);
1420                 else
1421                         CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1422                               libcfs_id2str(peerid), conn->ksnc_type, warn);
1423         }
1424
1425         if (!active) {
1426                 if (rc > 0) {
1427                         /* Request retry by replying with CONN_NONE
1428                          * ksnc_proto has been set already */
1429                         conn->ksnc_type = SOCKLND_CONN_NONE;
1430                         hello->kshm_nips = 0;
1431                         ksocknal_send_hello(ni, conn, peerid.nid, hello);
1432                 }
1433
1434                 write_lock_bh(global_lock);
1435                 peer_ni->ksnp_accepting--;
1436                 write_unlock_bh(global_lock);
1437         }
1438
1439         /*
1440          * If we get here without an error code, just use -EALREADY.
1441          * Depending on how we got here, the error may be positive
1442          * or negative. Normalize the value for ksocknal_txlist_done().
1443          */
1444         rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1445         ksocknal_txlist_done(ni, &zombies, rc2);
1446         ksocknal_peer_decref(peer_ni);
1447
1448 failed_1:
1449         if (hello != NULL)
1450                 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1451                                             kshm_ips[LNET_INTERFACES_NUM]));
1452
1453         LIBCFS_FREE(conn, sizeof(*conn));
1454
1455 failed_0:
1456         sock_release(sock);
1457         return rc;
1458 }
1459
1460 void
1461 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1462 {
1463         /* This just does the immmediate housekeeping, and queues the
1464          * connection for the reaper to terminate.
1465          * Caller holds ksnd_global_lock exclusively in irq context */
1466         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1467         struct ksock_route *route;
1468         struct ksock_conn *conn2;
1469         struct list_head *tmp;
1470
1471         LASSERT(peer_ni->ksnp_error == 0);
1472         LASSERT(!conn->ksnc_closing);
1473         conn->ksnc_closing = 1;
1474
1475         /* ksnd_deathrow_conns takes over peer_ni's ref */
1476         list_del(&conn->ksnc_list);
1477
1478         route = conn->ksnc_route;
1479         if (route != NULL) {
1480                 /* dissociate conn from route... */
1481                 LASSERT(!route->ksnr_deleted);
1482                 LASSERT((route->ksnr_connected & BIT(conn->ksnc_type)) != 0);
1483
1484                 conn2 = NULL;
1485                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1486                         conn2 = list_entry(tmp, struct ksock_conn, ksnc_list);
1487
1488                         if (conn2->ksnc_route == route &&
1489                             conn2->ksnc_type == conn->ksnc_type)
1490                                 break;
1491
1492                         conn2 = NULL;
1493                 }
1494                 if (conn2 == NULL)
1495                         route->ksnr_connected &= ~BIT(conn->ksnc_type);
1496
1497                 conn->ksnc_route = NULL;
1498
1499                 ksocknal_route_decref(route);   /* drop conn's ref on route */
1500         }
1501
1502         if (list_empty(&peer_ni->ksnp_conns)) {
1503                 /* No more connections to this peer_ni */
1504
1505                 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1506                         struct ksock_tx *tx;
1507
1508                         LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1509
1510                         /* throw them to the last connection...,
1511                          * these TXs will be send to /dev/null by scheduler */
1512                         list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1513                                             tx_list)
1514                                 ksocknal_tx_prep(conn, tx);
1515
1516                         spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1517                         list_splice_init(&peer_ni->ksnp_tx_queue,
1518                                          &conn->ksnc_tx_queue);
1519                         spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1520                 }
1521
1522                 /* renegotiate protocol version */
1523                 peer_ni->ksnp_proto = NULL;
1524                 /* stash last conn close reason */
1525                 peer_ni->ksnp_error = error;
1526
1527                 if (list_empty(&peer_ni->ksnp_routes)) {
1528                         /* I've just closed last conn belonging to a
1529                          * peer_ni with no routes to it */
1530                         ksocknal_unlink_peer_locked(peer_ni);
1531                 }
1532         }
1533
1534         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1535
1536         list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1537         wake_up(&ksocknal_data.ksnd_reaper_waitq);
1538
1539         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1540 }
1541
1542 void
1543 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1544 {
1545         int notify = 0;
1546         time64_t last_alive = 0;
1547
1548         /* There has been a connection failure or comms error; but I'll only
1549          * tell LNET I think the peer_ni is dead if it's to another kernel and
1550          * there are no connections or connection attempts in existence. */
1551
1552         read_lock(&ksocknal_data.ksnd_global_lock);
1553
1554         if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1555              list_empty(&peer_ni->ksnp_conns) &&
1556              peer_ni->ksnp_accepting == 0 &&
1557              ksocknal_find_connecting_route_locked(peer_ni) == NULL) {
1558                 notify = 1;
1559                 last_alive = peer_ni->ksnp_last_alive;
1560         }
1561
1562         read_unlock(&ksocknal_data.ksnd_global_lock);
1563
1564         if (notify)
1565                 lnet_notify(peer_ni->ksnp_ni, peer_ni->ksnp_id.nid,
1566                             false, false, last_alive);
1567 }
1568
1569 void
1570 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1571 {
1572         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1573         struct ksock_tx *tx;
1574         struct ksock_tx *tmp;
1575         LIST_HEAD(zlist);
1576
1577         /* NB safe to finalize TXs because closing of socket will
1578          * abort all buffered data */
1579         LASSERT(conn->ksnc_sock == NULL);
1580
1581         spin_lock(&peer_ni->ksnp_lock);
1582
1583         list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list, tx_zc_list) {
1584                 if (tx->tx_conn != conn)
1585                         continue;
1586
1587                 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1588
1589                 tx->tx_msg.ksm_zc_cookies[0] = 0;
1590                 tx->tx_zc_aborted = 1;  /* mark it as not-acked */
1591                 list_move(&tx->tx_zc_list, &zlist);
1592         }
1593
1594         spin_unlock(&peer_ni->ksnp_lock);
1595
1596         while (!list_empty(&zlist)) {
1597                 tx = list_entry(zlist.next, struct ksock_tx, tx_zc_list);
1598
1599                 list_del(&tx->tx_zc_list);
1600                 ksocknal_tx_decref(tx);
1601         }
1602 }
1603
1604 void
1605 ksocknal_terminate_conn(struct ksock_conn *conn)
1606 {
1607         /* This gets called by the reaper (guaranteed thread context) to
1608          * disengage the socket from its callbacks and close it.
1609          * ksnc_refcount will eventually hit zero, and then the reaper will
1610          * destroy it. */
1611         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1612         struct ksock_sched *sched = conn->ksnc_scheduler;
1613         int failed = 0;
1614
1615         LASSERT(conn->ksnc_closing);
1616
1617         /* wake up the scheduler to "send" all remaining packets to /dev/null */
1618         spin_lock_bh(&sched->kss_lock);
1619
1620         /* a closing conn is always ready to tx */
1621         conn->ksnc_tx_ready = 1;
1622
1623         if (!conn->ksnc_tx_scheduled &&
1624             !list_empty(&conn->ksnc_tx_queue)) {
1625                 list_add_tail(&conn->ksnc_tx_list,
1626                                &sched->kss_tx_conns);
1627                 conn->ksnc_tx_scheduled = 1;
1628                 /* extra ref for scheduler */
1629                 ksocknal_conn_addref(conn);
1630
1631                 wake_up (&sched->kss_waitq);
1632         }
1633
1634         spin_unlock_bh(&sched->kss_lock);
1635
1636         /* serialise with callbacks */
1637         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1638
1639         ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1640
1641         /* OK, so this conn may not be completely disengaged from its
1642          * scheduler yet, but it _has_ committed to terminate... */
1643         conn->ksnc_scheduler->kss_nconns--;
1644
1645         if (peer_ni->ksnp_error != 0) {
1646                 /* peer_ni's last conn closed in error */
1647                 LASSERT(list_empty(&peer_ni->ksnp_conns));
1648                 failed = 1;
1649                 peer_ni->ksnp_error = 0;     /* avoid multiple notifications */
1650         }
1651
1652         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1653
1654         if (failed)
1655                 ksocknal_peer_failed(peer_ni);
1656
1657         /* The socket is closed on the final put; either here, or in
1658          * ksocknal_{send,recv}msg().  Since we set up the linger2 option
1659          * when the connection was established, this will close the socket
1660          * immediately, aborting anything buffered in it. Any hung
1661          * zero-copy transmits will therefore complete in finite time. */
1662         ksocknal_connsock_decref(conn);
1663 }
1664
1665 void
1666 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1667 {
1668         /* Queue the conn for the reaper to destroy */
1669         LASSERT(atomic_read(&conn->ksnc_conn_refcount) == 0);
1670         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1671
1672         list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1673         wake_up(&ksocknal_data.ksnd_reaper_waitq);
1674
1675         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1676 }
1677
1678 void
1679 ksocknal_destroy_conn(struct ksock_conn *conn)
1680 {
1681         time64_t last_rcv;
1682
1683         /* Final coup-de-grace of the reaper */
1684         CDEBUG (D_NET, "connection %p\n", conn);
1685
1686         LASSERT (atomic_read (&conn->ksnc_conn_refcount) == 0);
1687         LASSERT (atomic_read (&conn->ksnc_sock_refcount) == 0);
1688         LASSERT (conn->ksnc_sock == NULL);
1689         LASSERT (conn->ksnc_route == NULL);
1690         LASSERT (!conn->ksnc_tx_scheduled);
1691         LASSERT (!conn->ksnc_rx_scheduled);
1692         LASSERT(list_empty(&conn->ksnc_tx_queue));
1693
1694         /* complete current receive if any */
1695         switch (conn->ksnc_rx_state) {
1696         case SOCKNAL_RX_LNET_PAYLOAD:
1697                 last_rcv = conn->ksnc_rx_deadline -
1698                            ksocknal_timeout();
1699                 CERROR("Completing partial receive from %s[%d], "
1700                        "ip %pI4h:%d, with error, wanted: %d, left: %d, "
1701                        "last alive is %lld secs ago\n",
1702                        libcfs_id2str(conn->ksnc_peer->ksnp_id), conn->ksnc_type,
1703                        &conn->ksnc_ipaddr, conn->ksnc_port,
1704                        conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1705                        ktime_get_seconds() - last_rcv);
1706                 if (conn->ksnc_lnet_msg)
1707                         conn->ksnc_lnet_msg->msg_health_status =
1708                                 LNET_MSG_STATUS_REMOTE_ERROR;
1709                 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1710                 break;
1711         case SOCKNAL_RX_LNET_HEADER:
1712                 if (conn->ksnc_rx_started)
1713                         CERROR("Incomplete receive of lnet header from %s, "
1714                                "ip %pI4h:%d, with error, protocol: %d.x.\n",
1715                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1716                                &conn->ksnc_ipaddr, conn->ksnc_port,
1717                                conn->ksnc_proto->pro_version);
1718                 break;
1719         case SOCKNAL_RX_KSM_HEADER:
1720                 if (conn->ksnc_rx_started)
1721                         CERROR("Incomplete receive of ksock message from %s, "
1722                                "ip %pI4h:%d, with error, protocol: %d.x.\n",
1723                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1724                                &conn->ksnc_ipaddr, conn->ksnc_port,
1725                                conn->ksnc_proto->pro_version);
1726                 break;
1727         case SOCKNAL_RX_SLOP:
1728                 if (conn->ksnc_rx_started)
1729                         CERROR("Incomplete receive of slops from %s, "
1730                                "ip %pI4h:%d, with error\n",
1731                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1732                                &conn->ksnc_ipaddr, conn->ksnc_port);
1733                break;
1734         default:
1735                 LBUG ();
1736                 break;
1737         }
1738
1739         ksocknal_peer_decref(conn->ksnc_peer);
1740
1741         LIBCFS_FREE (conn, sizeof (*conn));
1742 }
1743
1744 int
1745 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni, __u32 ipaddr, int why)
1746 {
1747         struct ksock_conn *conn;
1748         struct list_head *ctmp;
1749         struct list_head *cnxt;
1750         int count = 0;
1751
1752         list_for_each_safe(ctmp, cnxt, &peer_ni->ksnp_conns) {
1753                 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
1754
1755                 if (ipaddr == 0 ||
1756                     conn->ksnc_ipaddr == ipaddr) {
1757                         count++;
1758                         ksocknal_close_conn_locked (conn, why);
1759                 }
1760         }
1761
1762         return (count);
1763 }
1764
1765 int
1766 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1767 {
1768         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1769         u32 ipaddr = conn->ksnc_ipaddr;
1770         int count;
1771
1772         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1773
1774         count = ksocknal_close_peer_conns_locked (peer_ni, ipaddr, why);
1775
1776         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1777
1778         return (count);
1779 }
1780
1781 int
1782 ksocknal_close_matching_conns(struct lnet_process_id id, __u32 ipaddr)
1783 {
1784         struct ksock_peer_ni *peer_ni;
1785         struct hlist_node *pnxt;
1786         int lo;
1787         int hi;
1788         int i;
1789         int count = 0;
1790
1791         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1792
1793         if (id.nid != LNET_NID_ANY) {
1794                 lo = hash_min(id.nid, HASH_BITS(ksocknal_data.ksnd_peers));
1795                 hi = lo;
1796         } else {
1797                 lo = 0;
1798                 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1799         }
1800
1801         for (i = lo; i <= hi; i++) {
1802                 hlist_for_each_entry_safe(peer_ni, pnxt,
1803                                           &ksocknal_data.ksnd_peers[i],
1804                                           ksnp_list) {
1805
1806                         if (!((id.nid == LNET_NID_ANY ||
1807                                id.nid == peer_ni->ksnp_id.nid) &&
1808                               (id.pid == LNET_PID_ANY ||
1809                                id.pid == peer_ni->ksnp_id.pid)))
1810                                 continue;
1811
1812                         count += ksocknal_close_peer_conns_locked(peer_ni,
1813                                                                   ipaddr, 0);
1814                 }
1815         }
1816
1817         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1818
1819         /* wildcards always succeed */
1820         if (id.nid == LNET_NID_ANY || id.pid == LNET_PID_ANY || ipaddr == 0)
1821                 return 0;
1822
1823         return (count == 0 ? -ENOENT : 0);
1824 }
1825
1826 void
1827 ksocknal_notify_gw_down(lnet_nid_t gw_nid)
1828 {
1829         /* The router is telling me she's been notified of a change in
1830          * gateway state....
1831          */
1832         struct lnet_process_id id = {
1833                 .nid    = gw_nid,
1834                 .pid    = LNET_PID_ANY,
1835         };
1836
1837         CDEBUG(D_NET, "gw %s down\n", libcfs_nid2str(gw_nid));
1838
1839         /* If the gateway crashed, close all open connections... */
1840         ksocknal_close_matching_conns(id, 0);
1841         return;
1842
1843         /* We can only establish new connections
1844          * if we have autroutes, and these connect on demand. */
1845 }
1846
1847 static void
1848 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1849 {
1850         int index;
1851         int i;
1852         struct list_head *tmp;
1853         struct ksock_conn *conn;
1854
1855         for (index = 0; ; index++) {
1856                 read_lock(&ksocknal_data.ksnd_global_lock);
1857
1858                 i = 0;
1859                 conn = NULL;
1860
1861                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1862                         if (i++ == index) {
1863                                 conn = list_entry(tmp, struct ksock_conn,
1864                                                   ksnc_list);
1865                                 ksocknal_conn_addref(conn);
1866                                 break;
1867                         }
1868                 }
1869
1870                 read_unlock(&ksocknal_data.ksnd_global_lock);
1871
1872                 if (conn == NULL)
1873                         break;
1874
1875                 ksocknal_lib_push_conn (conn);
1876                 ksocknal_conn_decref(conn);
1877         }
1878 }
1879
1880 static int
1881 ksocknal_push(struct lnet_ni *ni, struct lnet_process_id id)
1882 {
1883         int lo;
1884         int hi;
1885         int bkt;
1886         int rc = -ENOENT;
1887
1888         if (id.nid != LNET_NID_ANY) {
1889                 lo = hash_min(id.nid, HASH_BITS(ksocknal_data.ksnd_peers));
1890                 hi = lo;
1891         } else {
1892                 lo = 0;
1893                 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1894         }
1895
1896         for (bkt = lo; bkt <= hi; bkt++) {
1897                 int peer_off; /* searching offset in peer_ni hash table */
1898
1899                 for (peer_off = 0; ; peer_off++) {
1900                         struct ksock_peer_ni *peer_ni;
1901                         int           i = 0;
1902
1903                         read_lock(&ksocknal_data.ksnd_global_lock);
1904                         hlist_for_each_entry(peer_ni,
1905                                              &ksocknal_data.ksnd_peers[bkt],
1906                                              ksnp_list) {
1907                                 if (!((id.nid == LNET_NID_ANY ||
1908                                        id.nid == peer_ni->ksnp_id.nid) &&
1909                                       (id.pid == LNET_PID_ANY ||
1910                                        id.pid == peer_ni->ksnp_id.pid)))
1911                                         continue;
1912
1913                                 if (i++ == peer_off) {
1914                                         ksocknal_peer_addref(peer_ni);
1915                                         break;
1916                                 }
1917                         }
1918                         read_unlock(&ksocknal_data.ksnd_global_lock);
1919
1920                         if (i <= peer_off) /* no match */
1921                                 break;
1922
1923                         rc = 0;
1924                         ksocknal_push_peer(peer_ni);
1925                         ksocknal_peer_decref(peer_ni);
1926                 }
1927         }
1928         return rc;
1929 }
1930
1931 static int
1932 ksocknal_add_interface(struct lnet_ni *ni, __u32 ipaddress, __u32 netmask)
1933 {
1934         struct ksock_net *net = ni->ni_data;
1935         struct ksock_interface *iface;
1936         int rc;
1937         int i;
1938         int j;
1939         struct ksock_peer_ni *peer_ni;
1940         struct list_head *rtmp;
1941         struct ksock_route *route;
1942
1943         if (ipaddress == 0 ||
1944             netmask == 0)
1945                 return -EINVAL;
1946
1947         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1948
1949         iface = ksocknal_ip2iface(ni, ipaddress);
1950         if (iface != NULL) {
1951                 /* silently ignore dups */
1952                 rc = 0;
1953         } else if (net->ksnn_ninterfaces == LNET_INTERFACES_NUM) {
1954                 rc = -ENOSPC;
1955         } else {
1956                 iface = &net->ksnn_interfaces[net->ksnn_ninterfaces++];
1957
1958                 iface->ksni_index = ksocknal_ip2index(ipaddress, ni);
1959                 iface->ksni_ipaddr = ipaddress;
1960                 iface->ksni_netmask = netmask;
1961                 iface->ksni_nroutes = 0;
1962                 iface->ksni_npeers = 0;
1963
1964                 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
1965                         for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++)
1966                                 if (peer_ni->ksnp_passive_ips[j] == ipaddress)
1967                                         iface->ksni_npeers++;
1968
1969                         list_for_each(rtmp, &peer_ni->ksnp_routes) {
1970                                 route = list_entry(rtmp,
1971                                                    struct ksock_route,
1972                                                    ksnr_list);
1973
1974                                 if (route->ksnr_myiface ==
1975                                             iface->ksni_index)
1976                                         iface->ksni_nroutes++;
1977                         }
1978                 }
1979
1980                 rc = 0;
1981                 /* NB only new connections will pay attention to the new
1982                  * interface!
1983                  */
1984         }
1985
1986         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1987
1988         return rc;
1989 }
1990
1991 static void
1992 ksocknal_peer_del_interface_locked(struct ksock_peer_ni *peer_ni,
1993                                    __u32 ipaddr, int index)
1994 {
1995         struct list_head *tmp;
1996         struct list_head *nxt;
1997         struct ksock_route *route;
1998         struct ksock_conn *conn;
1999         int i;
2000         int j;
2001
2002         for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++)
2003                 if (peer_ni->ksnp_passive_ips[i] == ipaddr) {
2004                         for (j = i+1; j < peer_ni->ksnp_n_passive_ips; j++)
2005                                 peer_ni->ksnp_passive_ips[j-1] =
2006                                         peer_ni->ksnp_passive_ips[j];
2007                         peer_ni->ksnp_n_passive_ips--;
2008                         break;
2009                 }
2010
2011         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
2012                 route = list_entry(tmp, struct ksock_route, ksnr_list);
2013
2014                 if (route->ksnr_myiface != index)
2015                         continue;
2016
2017                 if (route->ksnr_share_count != 0) {
2018                         /* Manually created; keep, but unbind */
2019                         route->ksnr_myiface = -1;
2020                 } else {
2021                         ksocknal_del_route_locked(route);
2022                 }
2023         }
2024
2025         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_conns) {
2026                 conn = list_entry(tmp, struct ksock_conn, ksnc_list);
2027
2028                 if (conn->ksnc_myipaddr == ipaddr)
2029                         ksocknal_close_conn_locked (conn, 0);
2030         }
2031 }
2032
2033 static int
2034 ksocknal_del_interface(struct lnet_ni *ni, __u32 ipaddress)
2035 {
2036         struct ksock_net *net = ni->ni_data;
2037         int rc = -ENOENT;
2038         struct hlist_node *nxt;
2039         struct ksock_peer_ni *peer_ni;
2040         u32 this_ip;
2041         int index;
2042         int i;
2043         int j;
2044
2045         index = ksocknal_ip2index(ipaddress, ni);
2046
2047         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2048
2049         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2050                 this_ip = net->ksnn_interfaces[i].ksni_ipaddr;
2051
2052                 if (!(ipaddress == 0 ||
2053                       ipaddress == this_ip))
2054                         continue;
2055
2056                 rc = 0;
2057
2058                 for (j = i+1; j < net->ksnn_ninterfaces; j++)
2059                         net->ksnn_interfaces[j-1] =
2060                                 net->ksnn_interfaces[j];
2061
2062                 net->ksnn_ninterfaces--;
2063
2064                 hash_for_each_safe(ksocknal_data.ksnd_peers, j,
2065                                    nxt, peer_ni, ksnp_list) {
2066                         if (peer_ni->ksnp_ni != ni)
2067                                 continue;
2068
2069                         ksocknal_peer_del_interface_locked(peer_ni,
2070                                                            this_ip, index);
2071                 }
2072         }
2073
2074         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2075
2076         return rc;
2077 }
2078
2079 int
2080 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
2081 {
2082         struct lnet_process_id id = {0};
2083         struct libcfs_ioctl_data *data = arg;
2084         int rc;
2085
2086         switch(cmd) {
2087         case IOC_LIBCFS_GET_INTERFACE: {
2088                 struct ksock_net *net = ni->ni_data;
2089                 struct ksock_interface *iface;
2090
2091                 read_lock(&ksocknal_data.ksnd_global_lock);
2092
2093                 if (data->ioc_count >= (__u32)net->ksnn_ninterfaces) {
2094                         rc = -ENOENT;
2095                 } else {
2096                         rc = 0;
2097                         iface = &net->ksnn_interfaces[data->ioc_count];
2098
2099                         data->ioc_u32[0] = iface->ksni_ipaddr;
2100                         data->ioc_u32[1] = iface->ksni_netmask;
2101                         data->ioc_u32[2] = iface->ksni_npeers;
2102                         data->ioc_u32[3] = iface->ksni_nroutes;
2103                 }
2104
2105                 read_unlock(&ksocknal_data.ksnd_global_lock);
2106                 return rc;
2107         }
2108
2109         case IOC_LIBCFS_ADD_INTERFACE:
2110                 return ksocknal_add_interface(ni,
2111                                               data->ioc_u32[0], /* IP address */
2112                                               data->ioc_u32[1]); /* net mask */
2113
2114         case IOC_LIBCFS_DEL_INTERFACE:
2115                 return ksocknal_del_interface(ni,
2116                                               data->ioc_u32[0]); /* IP address */
2117
2118         case IOC_LIBCFS_GET_PEER: {
2119                 __u32            myip = 0;
2120                 __u32            ip = 0;
2121                 int              port = 0;
2122                 int              conn_count = 0;
2123                 int              share_count = 0;
2124
2125                 rc = ksocknal_get_peer_info(ni, data->ioc_count,
2126                                             &id, &myip, &ip, &port,
2127                                             &conn_count,  &share_count);
2128                 if (rc != 0)
2129                         return rc;
2130
2131                 data->ioc_nid    = id.nid;
2132                 data->ioc_count  = share_count;
2133                 data->ioc_u32[0] = ip;
2134                 data->ioc_u32[1] = port;
2135                 data->ioc_u32[2] = myip;
2136                 data->ioc_u32[3] = conn_count;
2137                 data->ioc_u32[4] = id.pid;
2138                 return 0;
2139         }
2140
2141         case IOC_LIBCFS_ADD_PEER:
2142                 id.nid = data->ioc_nid;
2143                 id.pid = LNET_PID_LUSTRE;
2144                 return ksocknal_add_peer (ni, id,
2145                                           data->ioc_u32[0], /* IP */
2146                                           data->ioc_u32[1]); /* port */
2147
2148         case IOC_LIBCFS_DEL_PEER:
2149                 id.nid = data->ioc_nid;
2150                 id.pid = LNET_PID_ANY;
2151                 return ksocknal_del_peer (ni, id,
2152                                           data->ioc_u32[0]); /* IP */
2153
2154         case IOC_LIBCFS_GET_CONN: {
2155                 int           txmem;
2156                 int           rxmem;
2157                 int           nagle;
2158                 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
2159
2160                 if (conn == NULL)
2161                         return -ENOENT;
2162
2163                 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
2164
2165                 data->ioc_count  = txmem;
2166                 data->ioc_nid    = conn->ksnc_peer->ksnp_id.nid;
2167                 data->ioc_flags  = nagle;
2168                 data->ioc_u32[0] = conn->ksnc_ipaddr;
2169                 data->ioc_u32[1] = conn->ksnc_port;
2170                 data->ioc_u32[2] = conn->ksnc_myipaddr;
2171                 data->ioc_u32[3] = conn->ksnc_type;
2172                 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
2173                 data->ioc_u32[5] = rxmem;
2174                 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
2175                 ksocknal_conn_decref(conn);
2176                 return 0;
2177         }
2178
2179         case IOC_LIBCFS_CLOSE_CONNECTION:
2180                 id.nid = data->ioc_nid;
2181                 id.pid = LNET_PID_ANY;
2182                 return ksocknal_close_matching_conns (id,
2183                                                       data->ioc_u32[0]);
2184
2185         case IOC_LIBCFS_REGISTER_MYNID:
2186                 /* Ignore if this is a noop */
2187                 if (data->ioc_nid == ni->ni_nid)
2188                         return 0;
2189
2190                 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
2191                        libcfs_nid2str(data->ioc_nid),
2192                        libcfs_nid2str(ni->ni_nid));
2193                 return -EINVAL;
2194
2195         case IOC_LIBCFS_PUSH_CONNECTION:
2196                 id.nid = data->ioc_nid;
2197                 id.pid = LNET_PID_ANY;
2198                 return ksocknal_push(ni, id);
2199
2200         default:
2201                 return -EINVAL;
2202         }
2203         /* not reached */
2204 }
2205
2206 static void
2207 ksocknal_free_buffers (void)
2208 {
2209         LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
2210
2211         if (ksocknal_data.ksnd_schedulers != NULL)
2212                 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
2213
2214         spin_lock(&ksocknal_data.ksnd_tx_lock);
2215
2216         if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
2217                 LIST_HEAD(zlist);
2218                 struct ksock_tx *tx;
2219
2220                 list_splice_init(&ksocknal_data.ksnd_idle_noop_txs, &zlist);
2221                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2222
2223                 while (!list_empty(&zlist)) {
2224                         tx = list_entry(zlist.next, struct ksock_tx, tx_list);
2225                         list_del(&tx->tx_list);
2226                         LIBCFS_FREE(tx, tx->tx_desc_size);
2227                 }
2228         } else {
2229                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2230         }
2231 }
2232
2233 static void
2234 ksocknal_base_shutdown(void)
2235 {
2236         struct ksock_sched *sched;
2237         struct ksock_peer_ni *peer_ni;
2238         int i;
2239
2240         CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2241                atomic_read (&libcfs_kmemory));
2242         LASSERT (ksocknal_data.ksnd_nnets == 0);
2243
2244         switch (ksocknal_data.ksnd_init) {
2245         default:
2246                 LASSERT(0);
2247                 /* fallthrough */
2248
2249         case SOCKNAL_INIT_ALL:
2250         case SOCKNAL_INIT_DATA:
2251                 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list)
2252                         LASSERT(0);
2253
2254                 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2255                 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2256                 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2257                 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2258                 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2259
2260                 if (ksocknal_data.ksnd_schedulers != NULL) {
2261                         cfs_percpt_for_each(sched, i,
2262                                             ksocknal_data.ksnd_schedulers) {
2263
2264                                 LASSERT(list_empty(&sched->kss_tx_conns));
2265                                 LASSERT(list_empty(&sched->kss_rx_conns));
2266                                 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
2267                                 LASSERT(sched->kss_nconns == 0);
2268                         }
2269                 }
2270
2271                 /* flag threads to terminate; wake and wait for them to die */
2272                 ksocknal_data.ksnd_shuttingdown = 1;
2273                 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2274                 wake_up_all(&ksocknal_data.ksnd_reaper_waitq);
2275
2276                 if (ksocknal_data.ksnd_schedulers != NULL) {
2277                         cfs_percpt_for_each(sched, i,
2278                                             ksocknal_data.ksnd_schedulers)
2279                                         wake_up_all(&sched->kss_waitq);
2280                 }
2281
2282                 wait_var_event_warning(&ksocknal_data.ksnd_nthreads,
2283                                        ksocknal_data.ksnd_nthreads == 0,
2284                                        "waiting for %d threads to terminate\n",
2285                                        ksocknal_data.ksnd_nthreads);
2286
2287                 ksocknal_free_buffers();
2288
2289                 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2290                 break;
2291         }
2292
2293         CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2294                atomic_read (&libcfs_kmemory));
2295
2296         module_put(THIS_MODULE);
2297 }
2298
2299 static int
2300 ksocknal_base_startup(void)
2301 {
2302         struct ksock_sched *sched;
2303         int rc;
2304         int i;
2305
2306         LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2307         LASSERT(ksocknal_data.ksnd_nnets == 0);
2308
2309         memset(&ksocknal_data, 0, sizeof(ksocknal_data)); /* zero pointers */
2310
2311         hash_init(ksocknal_data.ksnd_peers);
2312
2313         rwlock_init(&ksocknal_data.ksnd_global_lock);
2314         INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2315
2316         spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2317         INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2318         INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2319         INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2320         init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2321
2322         spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2323         INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2324         INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2325         init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2326
2327         spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2328         INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2329
2330         /* NB memset above zeros whole of ksocknal_data */
2331
2332         /* flag lists/ptrs/locks initialised */
2333         ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2334         if (!try_module_get(THIS_MODULE))
2335                 goto failed;
2336
2337         /* Create a scheduler block per available CPT */
2338         ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2339                                                          sizeof(*sched));
2340         if (ksocknal_data.ksnd_schedulers == NULL)
2341                 goto failed;
2342
2343         cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2344                 int nthrs;
2345
2346                 /*
2347                  * make sure not to allocate more threads than there are
2348                  * cores/CPUs in teh CPT
2349                  */
2350                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2351                 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2352                         nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2353                 } else {
2354                         /*
2355                          * max to half of CPUs, assume another half should be
2356                          * reserved for upper layer modules
2357                          */
2358                         nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2359                 }
2360
2361                 sched->kss_nthreads_max = nthrs;
2362                 sched->kss_cpt = i;
2363
2364                 spin_lock_init(&sched->kss_lock);
2365                 INIT_LIST_HEAD(&sched->kss_rx_conns);
2366                 INIT_LIST_HEAD(&sched->kss_tx_conns);
2367                 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2368                 init_waitqueue_head(&sched->kss_waitq);
2369         }
2370
2371         ksocknal_data.ksnd_connd_starting         = 0;
2372         ksocknal_data.ksnd_connd_failed_stamp     = 0;
2373         ksocknal_data.ksnd_connd_starting_stamp   = ktime_get_real_seconds();
2374         /* must have at least 2 connds to remain responsive to accepts while
2375          * connecting */
2376         if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2377                 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2378
2379         if (*ksocknal_tunables.ksnd_nconnds_max <
2380             *ksocknal_tunables.ksnd_nconnds) {
2381                 ksocknal_tunables.ksnd_nconnds_max =
2382                         ksocknal_tunables.ksnd_nconnds;
2383         }
2384
2385         for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2386                 char name[16];
2387                 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2388                 ksocknal_data.ksnd_connd_starting++;
2389                 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2390
2391
2392                 snprintf(name, sizeof(name), "socknal_cd%02d", i);
2393                 rc = ksocknal_thread_start(ksocknal_connd,
2394                                            (void *)((uintptr_t)i), name);
2395                 if (rc != 0) {
2396                         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2397                         ksocknal_data.ksnd_connd_starting--;
2398                         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2399                         CERROR("Can't spawn socknal connd: %d\n", rc);
2400                         goto failed;
2401                 }
2402         }
2403
2404         rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2405         if (rc != 0) {
2406                 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2407                 goto failed;
2408         }
2409
2410         /* flag everything initialised */
2411         ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2412
2413         return 0;
2414
2415  failed:
2416         ksocknal_base_shutdown();
2417         return -ENETDOWN;
2418 }
2419
2420 static int
2421 ksocknal_debug_peerhash(struct lnet_ni *ni)
2422 {
2423         struct ksock_peer_ni *peer_ni;
2424         int i;
2425
2426         read_lock(&ksocknal_data.ksnd_global_lock);
2427
2428         hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
2429                 struct ksock_route *route;
2430                 struct ksock_conn *conn;
2431
2432                 if (peer_ni->ksnp_ni != ni)
2433                         continue;
2434
2435                 CWARN("Active peer_ni on shutdown: %s, ref %d, "
2436                       "closing %d, accepting %d, err %d, zcookie %llu, "
2437                       "txq %d, zc_req %d\n", libcfs_id2str(peer_ni->ksnp_id),
2438                       atomic_read(&peer_ni->ksnp_refcount),
2439                       peer_ni->ksnp_closing,
2440                       peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2441                       peer_ni->ksnp_zc_next_cookie,
2442                       !list_empty(&peer_ni->ksnp_tx_queue),
2443                       !list_empty(&peer_ni->ksnp_zc_req_list));
2444
2445                 list_for_each_entry(route, &peer_ni->ksnp_routes, ksnr_list) {
2446                         CWARN("Route: ref %d, schd %d, conn %d, cnted %d, "
2447                               "del %d\n", atomic_read(&route->ksnr_refcount),
2448                               route->ksnr_scheduled, route->ksnr_connecting,
2449                               route->ksnr_connected, route->ksnr_deleted);
2450                 }
2451
2452                 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2453                         CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2454                               atomic_read(&conn->ksnc_conn_refcount),
2455                               atomic_read(&conn->ksnc_sock_refcount),
2456                               conn->ksnc_type, conn->ksnc_closing);
2457                 }
2458                 break;
2459         }
2460
2461         read_unlock(&ksocknal_data.ksnd_global_lock);
2462         return 0;
2463 }
2464
2465 void
2466 ksocknal_shutdown(struct lnet_ni *ni)
2467 {
2468         struct ksock_net *net = ni->ni_data;
2469         struct lnet_process_id anyid = {
2470                 .nid = LNET_NID_ANY,
2471                 .pid = LNET_PID_ANY,
2472         };
2473         int i;
2474
2475         LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2476         LASSERT(ksocknal_data.ksnd_nnets > 0);
2477
2478         /* prevent new peers */
2479         atomic_add(SOCKNAL_SHUTDOWN_BIAS, &net->ksnn_npeers);
2480
2481         /* Delete all peers */
2482         ksocknal_del_peer(ni, anyid, 0);
2483
2484         /* Wait for all peer_ni state to clean up */
2485         wait_var_event_warning(&net->ksnn_npeers,
2486                                atomic_read(&net->ksnn_npeers) ==
2487                                SOCKNAL_SHUTDOWN_BIAS,
2488                                "waiting for %d peers to disconnect\n",
2489                                ksocknal_debug_peerhash(ni) +
2490                                atomic_read(&net->ksnn_npeers) -
2491                                SOCKNAL_SHUTDOWN_BIAS);
2492
2493         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2494                 LASSERT(net->ksnn_interfaces[i].ksni_npeers == 0);
2495                 LASSERT(net->ksnn_interfaces[i].ksni_nroutes == 0);
2496         }
2497
2498         list_del(&net->ksnn_list);
2499         LIBCFS_FREE(net, sizeof(*net));
2500
2501         ksocknal_data.ksnd_nnets--;
2502         if (ksocknal_data.ksnd_nnets == 0)
2503                 ksocknal_base_shutdown();
2504 }
2505
2506 static int
2507 ksocknal_search_new_ipif(struct ksock_net *net)
2508 {
2509         int new_ipif = 0;
2510         int i;
2511
2512         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2513                 char *ifnam = &net->ksnn_interfaces[i].ksni_name[0];
2514                 char *colon = strchr(ifnam, ':');
2515                 int found  = 0;
2516                 struct ksock_net *tmp;
2517                 int j;
2518
2519                 if (colon != NULL) /* ignore alias device */
2520                         *colon = 0;
2521
2522                 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets,
2523                                         ksnn_list) {
2524                         for (j = 0; !found && j < tmp->ksnn_ninterfaces; j++) {
2525                                 char *ifnam2 = &tmp->ksnn_interfaces[j].\
2526                                              ksni_name[0];
2527                                 char *colon2 = strchr(ifnam2, ':');
2528
2529                                 if (colon2 != NULL)
2530                                         *colon2 = 0;
2531
2532                                 found = strcmp(ifnam, ifnam2) == 0;
2533                                 if (colon2 != NULL)
2534                                         *colon2 = ':';
2535                         }
2536                         if (found)
2537                                 break;
2538                 }
2539
2540                 new_ipif += !found;
2541                 if (colon != NULL)
2542                         *colon = ':';
2543         }
2544
2545         return new_ipif;
2546 }
2547
2548 static int
2549 ksocknal_start_schedulers(struct ksock_sched *sched)
2550 {
2551         int     nthrs;
2552         int     rc = 0;
2553         int     i;
2554
2555         if (sched->kss_nthreads == 0) {
2556                 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2557                         nthrs = sched->kss_nthreads_max;
2558                 } else {
2559                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2560                                                sched->kss_cpt);
2561                         nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2562                         nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2563                 }
2564                 nthrs = min(nthrs, sched->kss_nthreads_max);
2565         } else {
2566                 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2567                 /* increase two threads if there is new interface */
2568                 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2569         }
2570
2571         for (i = 0; i < nthrs; i++) {
2572                 long id;
2573                 char name[20];
2574
2575                 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2576                 snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
2577                          sched->kss_cpt, (int)KSOCK_THREAD_SID(id));
2578
2579                 rc = ksocknal_thread_start(ksocknal_scheduler,
2580                                            (void *)id, name);
2581                 if (rc == 0)
2582                         continue;
2583
2584                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2585                        sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2586                 break;
2587         }
2588
2589         sched->kss_nthreads += i;
2590         return rc;
2591 }
2592
2593 static int
2594 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2595 {
2596         int newif = ksocknal_search_new_ipif(net);
2597         int rc;
2598         int i;
2599
2600         if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2601                 return -EINVAL;
2602
2603         for (i = 0; i < ncpts; i++) {
2604                 struct ksock_sched *sched;
2605                 int cpt = (cpts == NULL) ? i : cpts[i];
2606
2607                 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2608                 sched = ksocknal_data.ksnd_schedulers[cpt];
2609
2610                 if (!newif && sched->kss_nthreads > 0)
2611                         continue;
2612
2613                 rc = ksocknal_start_schedulers(sched);
2614                 if (rc != 0)
2615                         return rc;
2616         }
2617         return 0;
2618 }
2619
2620 int
2621 ksocknal_startup(struct lnet_ni *ni)
2622 {
2623         struct ksock_net *net;
2624         struct lnet_ioctl_config_lnd_cmn_tunables *net_tunables;
2625         struct ksock_interface *ksi = NULL;
2626         struct lnet_inetdev *ifaces = NULL;
2627         int i = 0;
2628         int rc;
2629
2630         LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2631
2632         if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2633                 rc = ksocknal_base_startup();
2634                 if (rc != 0)
2635                         return rc;
2636         }
2637
2638         LIBCFS_ALLOC(net, sizeof(*net));
2639         if (net == NULL)
2640                 goto fail_0;
2641
2642         net->ksnn_incarnation = ktime_get_real_ns();
2643         ni->ni_data = net;
2644         net_tunables = &ni->ni_net->net_tunables;
2645
2646         if (net_tunables->lct_peer_timeout == -1)
2647                 net_tunables->lct_peer_timeout =
2648                         *ksocknal_tunables.ksnd_peertimeout;
2649
2650         if (net_tunables->lct_max_tx_credits == -1)
2651                 net_tunables->lct_max_tx_credits =
2652                         *ksocknal_tunables.ksnd_credits;
2653
2654         if (net_tunables->lct_peer_tx_credits == -1)
2655                 net_tunables->lct_peer_tx_credits =
2656                         *ksocknal_tunables.ksnd_peertxcredits;
2657
2658         if (net_tunables->lct_peer_tx_credits >
2659             net_tunables->lct_max_tx_credits)
2660                 net_tunables->lct_peer_tx_credits =
2661                         net_tunables->lct_max_tx_credits;
2662
2663         if (net_tunables->lct_peer_rtr_credits == -1)
2664                 net_tunables->lct_peer_rtr_credits =
2665                         *ksocknal_tunables.ksnd_peerrtrcredits;
2666
2667         rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
2668         if (rc < 0)
2669                 goto fail_1;
2670
2671         if (!ni->ni_interfaces[0]) {
2672                 ksi = &net->ksnn_interfaces[0];
2673
2674                 /* Use the first discovered interface */
2675                 net->ksnn_ninterfaces = 1;
2676                 ni->ni_dev_cpt = ifaces[0].li_cpt;
2677                 ksi->ksni_ipaddr = ifaces[0].li_ipaddr;
2678                 ksi->ksni_index = ksocknal_ip2index(ksi->ksni_ipaddr, ni);
2679                 ksi->ksni_netmask = ifaces[0].li_netmask;
2680                 strlcpy(ksi->ksni_name, ifaces[0].li_name,
2681                         sizeof(ksi->ksni_name));
2682         } else {
2683                 /* Before Multi-Rail ksocklnd would manage
2684                  * multiple interfaces with its own tcp bonding.
2685                  * If we encounter an old configuration using
2686                  * this tcp bonding approach then we need to
2687                  * handle more than one ni_interfaces.
2688                  *
2689                  * In Multi-Rail configuration only ONE ni_interface
2690                  * should exist. Each IP alias should be mapped to
2691                  * each 'struct net_ni'.
2692                  */
2693                 for (i = 0; i < LNET_INTERFACES_NUM; i++) {
2694                         int j;
2695
2696                         if (!ni->ni_interfaces[i])
2697                                 break;
2698
2699                         for (j = 0; j < LNET_INTERFACES_NUM;  j++) {
2700                                 if (i != j && ni->ni_interfaces[j] &&
2701                                     strcmp(ni->ni_interfaces[i],
2702                                            ni->ni_interfaces[j]) == 0) {
2703                                         rc = -EEXIST;
2704                                         CERROR("ksocklnd: found duplicate %s at %d and %d, rc = %d\n",
2705                                                ni->ni_interfaces[i], i, j, rc);
2706                                         goto fail_1;
2707                                 }
2708                         }
2709
2710                         for (j = 0; j < rc; j++) {
2711                                 if (strcmp(ifaces[j].li_name,
2712                                            ni->ni_interfaces[i]) != 0)
2713                                         continue;
2714
2715                                 ksi = &net->ksnn_interfaces[j];
2716                                 ni->ni_dev_cpt = ifaces[j].li_cpt;
2717                                 ksi->ksni_ipaddr = ifaces[j].li_ipaddr;
2718                                 ksi->ksni_index =
2719                                   ksocknal_ip2index(ksi->ksni_ipaddr, ni);
2720                                 ksi->ksni_netmask = ifaces[j].li_netmask;
2721                                 strlcpy(ksi->ksni_name, ifaces[j].li_name,
2722                                         sizeof(ksi->ksni_name));
2723                                 net->ksnn_ninterfaces++;
2724                                 break;
2725                         }
2726                 }
2727                 /* ni_interfaces don't map to all network interfaces */
2728                 if (!ksi || net->ksnn_ninterfaces != i) {
2729                         CERROR("ksocklnd: requested %d but only %d interfaces found\n",
2730                                i, net->ksnn_ninterfaces);
2731                         goto fail_1;
2732                 }
2733         }
2734
2735         /* call it before add it to ksocknal_data.ksnd_nets */
2736         rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2737         if (rc != 0)
2738                 goto fail_1;
2739
2740         LASSERT(ksi);
2741         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ksi->ksni_ipaddr);
2742         list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2743
2744         ksocknal_data.ksnd_nnets++;
2745
2746         return 0;
2747
2748  fail_1:
2749         LIBCFS_FREE(net, sizeof(*net));
2750  fail_0:
2751         if (ksocknal_data.ksnd_nnets == 0)
2752                 ksocknal_base_shutdown();
2753
2754         return -ENETDOWN;
2755 }
2756
2757
2758 static void __exit ksocklnd_exit(void)
2759 {
2760         lnet_unregister_lnd(&the_ksocklnd);
2761 }
2762
2763 static const struct lnet_lnd the_ksocklnd = {
2764         .lnd_type               = SOCKLND,
2765         .lnd_startup            = ksocknal_startup,
2766         .lnd_shutdown           = ksocknal_shutdown,
2767         .lnd_ctl                = ksocknal_ctl,
2768         .lnd_send               = ksocknal_send,
2769         .lnd_recv               = ksocknal_recv,
2770         .lnd_notify_peer_down   = ksocknal_notify_gw_down,
2771         .lnd_accept             = ksocknal_accept,
2772 };
2773
2774 static int __init ksocklnd_init(void)
2775 {
2776         int rc;
2777
2778         /* check ksnr_connected/connecting field large enough */
2779         BUILD_BUG_ON(SOCKLND_CONN_NTYPES > 4);
2780         BUILD_BUG_ON(SOCKLND_CONN_ACK != SOCKLND_CONN_BULK_IN);
2781
2782         rc = ksocknal_tunables_init();
2783         if (rc != 0)
2784                 return rc;
2785
2786         lnet_register_lnd(&the_ksocklnd);
2787
2788         return 0;
2789 }
2790
2791 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2792 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2793 MODULE_VERSION("2.8.0");
2794 MODULE_LICENSE("GPL");
2795
2796 module_init(ksocklnd_init);
2797 module_exit(ksocklnd_exit);