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