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