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.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2015, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lnet/klnds/socklnd/socklnd.c
  37  *
  38  * Author: Zach Brown <zab@zabbo.net>
  39  * Author: Peter J. Braam <braam@clusterfs.com>
  40  * Author: Phil Schwan <phil@clusterfs.com>
  41  * Author: Eric Barton <eric@bartonsoftware.com>
  42  */
  43
  44 #include "socklnd.h"
  45
  46 static lnd_t                   the_ksocklnd;
  47 ksock_nal_data_t        ksocknal_data;
  48
  49 static ksock_interface_t *
  50 ksocknal_ip2iface(lnet_ni_t *ni, __u32 ip)
  51 {
  52         ksock_net_t       *net = ni->ni_data;
  53         int                i;
  54         ksock_interface_t *iface;
  55
  56         for (i = 0; i < net->ksnn_ninterfaces; i++) {
  57                 LASSERT(i < LNET_MAX_INTERFACES);
  58                 iface = &net->ksnn_interfaces[i];
  59
  60                 if (iface->ksni_ipaddr == ip)
  61                         return (iface);
  62         }
  63
  64         return (NULL);
  65 }
  66
  67 static ksock_route_t *
  68 ksocknal_create_route (__u32 ipaddr, int port)
  69 {
  70         ksock_route_t *route;
  71
  72         LIBCFS_ALLOC (route, sizeof (*route));
  73         if (route == NULL)
  74                 return (NULL);
  75
  76         atomic_set (&route->ksnr_refcount, 1);
  77         route->ksnr_peer = NULL;
  78         route->ksnr_retry_interval = 0;         /* OK to connect at any time */
  79         route->ksnr_ipaddr = ipaddr;
  80         route->ksnr_port = port;
  81         route->ksnr_scheduled = 0;
  82         route->ksnr_connecting = 0;
  83         route->ksnr_connected = 0;
  84         route->ksnr_deleted = 0;
  85         route->ksnr_conn_count = 0;
  86         route->ksnr_share_count = 0;
  87
  88         return (route);
  89 }
  90
  91 void
  92 ksocknal_destroy_route (ksock_route_t *route)
  93 {
  94         LASSERT (atomic_read(&route->ksnr_refcount) == 0);
  95
  96         if (route->ksnr_peer != NULL)
  97                 ksocknal_peer_decref(route->ksnr_peer);
  98
  99         LIBCFS_FREE (route, sizeof (*route));
 100 }
 101
 102 static int
 103 ksocknal_create_peer(ksock_peer_t **peerp, lnet_ni_t *ni, lnet_process_id_t id)
 104 {
 105         int             cpt = lnet_cpt_of_nid(id.nid);
 106         ksock_net_t     *net = ni->ni_data;
 107         ksock_peer_t    *peer;
 108
 109         LASSERT(id.nid != LNET_NID_ANY);
 110         LASSERT(id.pid != LNET_PID_ANY);
 111         LASSERT(!in_interrupt());
 112
 113         LIBCFS_CPT_ALLOC(peer, lnet_cpt_table(), cpt, sizeof(*peer));
 114         if (peer == NULL)
 115                 return -ENOMEM;
 116
 117         peer->ksnp_ni = ni;
 118         peer->ksnp_id = id;
 119         atomic_set(&peer->ksnp_refcount, 1);    /* 1 ref for caller */
 120         peer->ksnp_closing = 0;
 121         peer->ksnp_accepting = 0;
 122         peer->ksnp_proto = NULL;
 123         peer->ksnp_last_alive = 0;
 124         peer->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 125
 126         INIT_LIST_HEAD(&peer->ksnp_conns);
 127         INIT_LIST_HEAD(&peer->ksnp_routes);
 128         INIT_LIST_HEAD(&peer->ksnp_tx_queue);
 129         INIT_LIST_HEAD(&peer->ksnp_zc_req_list);
 130         spin_lock_init(&peer->ksnp_lock);
 131
 132         spin_lock_bh(&net->ksnn_lock);
 133
 134         if (net->ksnn_shutdown) {
 135                 spin_unlock_bh(&net->ksnn_lock);
 136
 137                 LIBCFS_FREE(peer, sizeof(*peer));
 138                 CERROR("Can't create peer: network shutdown\n");
 139                 return -ESHUTDOWN;
 140         }
 141
 142         net->ksnn_npeers++;
 143
 144         spin_unlock_bh(&net->ksnn_lock);
 145
 146         *peerp = peer;
 147         return 0;
 148 }
 149
 150 void
 151 ksocknal_destroy_peer (ksock_peer_t *peer)
 152 {
 153         ksock_net_t    *net = peer->ksnp_ni->ni_data;
 154
 155         CDEBUG (D_NET, "peer %s %p deleted\n",
 156                 libcfs_id2str(peer->ksnp_id), peer);
 157
 158         LASSERT(atomic_read(&peer->ksnp_refcount) == 0);
 159         LASSERT(peer->ksnp_accepting == 0);
 160         LASSERT(list_empty(&peer->ksnp_conns));
 161         LASSERT(list_empty(&peer->ksnp_routes));
 162         LASSERT(list_empty(&peer->ksnp_tx_queue));
 163         LASSERT(list_empty(&peer->ksnp_zc_req_list));
 164
 165         LIBCFS_FREE(peer, sizeof(*peer));
 166
 167         /* NB a peer's connections and routes keep a reference on their peer
 168          * until they are destroyed, so we can be assured that _all_ state to
 169          * do with this peer has been cleaned up when its refcount drops to
 170          * zero. */
 171         spin_lock_bh(&net->ksnn_lock);
 172         net->ksnn_npeers--;
 173         spin_unlock_bh(&net->ksnn_lock);
 174 }
 175
 176 ksock_peer_t *
 177 ksocknal_find_peer_locked (lnet_ni_t *ni, lnet_process_id_t id)
 178 {
 179         struct list_head *peer_list = ksocknal_nid2peerlist(id.nid);
 180         struct list_head *tmp;
 181         ksock_peer_t     *peer;
 182
 183         list_for_each(tmp, peer_list) {
 184
 185                 peer = list_entry(tmp, ksock_peer_t, ksnp_list);
 186
 187                 LASSERT(!peer->ksnp_closing);
 188
 189                 if (peer->ksnp_ni != ni)
 190                         continue;
 191
 192                 if (peer->ksnp_id.nid != id.nid ||
 193                     peer->ksnp_id.pid != id.pid)
 194                         continue;
 195
 196                 CDEBUG(D_NET, "got peer [%p] -> %s (%d)\n",
 197                        peer, libcfs_id2str(id),
 198                        atomic_read(&peer->ksnp_refcount));
 199                 return peer;
 200         }
 201         return NULL;
 202 }
 203
 204 ksock_peer_t *
 205 ksocknal_find_peer (lnet_ni_t *ni, lnet_process_id_t id)
 206 {
 207         ksock_peer_t     *peer;
 208
 209         read_lock(&ksocknal_data.ksnd_global_lock);
 210         peer = ksocknal_find_peer_locked(ni, id);
 211         if (peer != NULL)                       /* +1 ref for caller? */
 212                 ksocknal_peer_addref(peer);
 213         read_unlock(&ksocknal_data.ksnd_global_lock);
 214
 215         return (peer);
 216 }
 217
 218 static void
 219 ksocknal_unlink_peer_locked (ksock_peer_t *peer)
 220 {
 221         int                i;
 222         __u32              ip;
 223         ksock_interface_t *iface;
 224
 225         for (i = 0; i < peer->ksnp_n_passive_ips; i++) {
 226                 LASSERT (i < LNET_MAX_INTERFACES);
 227                 ip = peer->ksnp_passive_ips[i];
 228
 229                 iface = ksocknal_ip2iface(peer->ksnp_ni, ip);
 230                 /* All IPs in peer->ksnp_passive_ips[] come from the
 231                  * interface list, therefore the call must succeed. */
 232                 LASSERT (iface != NULL);
 233
 234                 CDEBUG(D_NET, "peer=%p iface=%p ksni_nroutes=%d\n",
 235                        peer, iface, iface->ksni_nroutes);
 236                 iface->ksni_npeers--;
 237         }
 238
 239         LASSERT(list_empty(&peer->ksnp_conns));
 240         LASSERT(list_empty(&peer->ksnp_routes));
 241         LASSERT(!peer->ksnp_closing);
 242         peer->ksnp_closing = 1;
 243         list_del(&peer->ksnp_list);
 244         /* lose peerlist's ref */
 245         ksocknal_peer_decref(peer);
 246 }
 247
 248 static int
 249 ksocknal_get_peer_info (lnet_ni_t *ni, int index,
 250                         lnet_process_id_t *id, __u32 *myip, __u32 *peer_ip,
 251                         int *port, int *conn_count, int *share_count)
 252 {
 253         ksock_peer_t      *peer;
 254         struct list_head  *ptmp;
 255         ksock_route_t     *route;
 256         struct list_head  *rtmp;
 257         int                i;
 258         int                j;
 259         int                rc = -ENOENT;
 260
 261         read_lock(&ksocknal_data.ksnd_global_lock);
 262
 263         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
 264                 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
 265                         peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
 266
 267                         if (peer->ksnp_ni != ni)
 268                                 continue;
 269
 270                         if (peer->ksnp_n_passive_ips == 0 &&
 271                             list_empty(&peer->ksnp_routes)) {
 272                                 if (index-- > 0)
 273                                         continue;
 274
 275                                 *id = peer->ksnp_id;
 276                                 *myip = 0;
 277                                 *peer_ip = 0;
 278                                 *port = 0;
 279                                 *conn_count = 0;
 280                                 *share_count = 0;
 281                                 rc = 0;
 282                                 goto out;
 283                         }
 284
 285                         for (j = 0; j < peer->ksnp_n_passive_ips; j++) {
 286                                 if (index-- > 0)
 287                                         continue;
 288
 289                                 *id = peer->ksnp_id;
 290                                 *myip = peer->ksnp_passive_ips[j];
 291                                 *peer_ip = 0;
 292                                 *port = 0;
 293                                 *conn_count = 0;
 294                                 *share_count = 0;
 295                                 rc = 0;
 296                                 goto out;
 297                         }
 298
 299                         list_for_each(rtmp, &peer->ksnp_routes) {
 300                                 if (index-- > 0)
 301                                         continue;
 302
 303                                 route = list_entry(rtmp, ksock_route_t,
 304                                                    ksnr_list);
 305
 306                                 *id = peer->ksnp_id;
 307                                 *myip = route->ksnr_myipaddr;
 308                                 *peer_ip = route->ksnr_ipaddr;
 309                                 *port = route->ksnr_port;
 310                                 *conn_count = route->ksnr_conn_count;
 311                                 *share_count = route->ksnr_share_count;
 312                                 rc = 0;
 313                                 goto out;
 314                         }
 315                 }
 316         }
 317 out:
 318         read_unlock(&ksocknal_data.ksnd_global_lock);
 319         return rc;
 320 }
 321
 322 static void
 323 ksocknal_associate_route_conn_locked(ksock_route_t *route, ksock_conn_t *conn)
 324 {
 325         ksock_peer_t      *peer = route->ksnr_peer;
 326         int                type = conn->ksnc_type;
 327         ksock_interface_t *iface;
 328
 329         conn->ksnc_route = route;
 330         ksocknal_route_addref(route);
 331
 332         if (route->ksnr_myipaddr != conn->ksnc_myipaddr) {
 333                 if (route->ksnr_myipaddr == 0) {
 334                         /* route wasn't bound locally yet (the initial route) */
 335                         CDEBUG(D_NET, "Binding %s %pI4h to %pI4h\n",
 336                                libcfs_id2str(peer->ksnp_id),
 337                                &route->ksnr_ipaddr,
 338                                &conn->ksnc_myipaddr);
 339                 } else {
 340                         CDEBUG(D_NET, "Rebinding %s %pI4h from %pI4h "
 341                                "to %pI4h\n", libcfs_id2str(peer->ksnp_id),
 342                                &route->ksnr_ipaddr,
 343                                &route->ksnr_myipaddr,
 344                                &conn->ksnc_myipaddr);
 345
 346                         iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
 347                                                   route->ksnr_myipaddr);
 348                         if (iface != NULL)
 349                                 iface->ksni_nroutes--;
 350                 }
 351                 route->ksnr_myipaddr = conn->ksnc_myipaddr;
 352                 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
 353                                           route->ksnr_myipaddr);
 354                 if (iface != NULL)
 355                         iface->ksni_nroutes++;
 356         }
 357
 358         route->ksnr_connected |= (1<<type);
 359         route->ksnr_conn_count++;
 360
 361         /* Successful connection => further attempts can
 362          * proceed immediately */
 363         route->ksnr_retry_interval = 0;
 364 }
 365
 366 static void
 367 ksocknal_add_route_locked (ksock_peer_t *peer, ksock_route_t *route)
 368 {
 369         struct list_head *tmp;
 370         ksock_conn_t     *conn;
 371         ksock_route_t    *route2;
 372
 373         LASSERT(!peer->ksnp_closing);
 374         LASSERT(route->ksnr_peer == NULL);
 375         LASSERT(!route->ksnr_scheduled);
 376         LASSERT(!route->ksnr_connecting);
 377         LASSERT(route->ksnr_connected == 0);
 378
 379         /* LASSERT(unique) */
 380         list_for_each(tmp, &peer->ksnp_routes) {
 381                 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
 382
 383                 if (route2->ksnr_ipaddr == route->ksnr_ipaddr) {
 384                         CERROR("Duplicate route %s %pI4h\n",
 385                                libcfs_id2str(peer->ksnp_id),
 386                                &route->ksnr_ipaddr);
 387                         LBUG();
 388                 }
 389         }
 390
 391         route->ksnr_peer = peer;
 392         ksocknal_peer_addref(peer);
 393         /* peer's routelist takes over my ref on 'route' */
 394         list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
 395
 396         list_for_each(tmp, &peer->ksnp_conns) {
 397                 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
 398
 399                 if (conn->ksnc_ipaddr != route->ksnr_ipaddr)
 400                         continue;
 401
 402                 ksocknal_associate_route_conn_locked(route, conn);
 403                 /* keep going (typed routes) */
 404         }
 405 }
 406
 407 static void
 408 ksocknal_del_route_locked (ksock_route_t *route)
 409 {
 410         ksock_peer_t      *peer = route->ksnr_peer;
 411         ksock_interface_t *iface;
 412         ksock_conn_t      *conn;
 413         struct list_head  *ctmp;
 414         struct list_head  *cnxt;
 415
 416         LASSERT(!route->ksnr_deleted);
 417
 418         /* Close associated conns */
 419         list_for_each_safe(ctmp, cnxt, &peer->ksnp_conns) {
 420                 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
 421
 422                 if (conn->ksnc_route != route)
 423                         continue;
 424
 425                 ksocknal_close_conn_locked(conn, 0);
 426         }
 427
 428         if (route->ksnr_myipaddr != 0) {
 429                 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
 430                                           route->ksnr_myipaddr);
 431                 if (iface != NULL)
 432                         iface->ksni_nroutes--;
 433         }
 434
 435         route->ksnr_deleted = 1;
 436         list_del(&route->ksnr_list);
 437         ksocknal_route_decref(route);           /* drop peer's ref */
 438
 439         if (list_empty(&peer->ksnp_routes) &&
 440             list_empty(&peer->ksnp_conns)) {
 441                 /* I've just removed the last route to a peer with no active
 442                  * connections */
 443                 ksocknal_unlink_peer_locked(peer);
 444         }
 445 }
 446
 447 int
 448 ksocknal_add_peer(lnet_ni_t *ni, lnet_process_id_t id, __u32 ipaddr, int port)
 449 {
 450         struct list_head *tmp;
 451         ksock_peer_t     *peer;
 452         ksock_peer_t     *peer2;
 453         ksock_route_t    *route;
 454         ksock_route_t    *route2;
 455         int               rc;
 456
 457         if (id.nid == LNET_NID_ANY ||
 458             id.pid == LNET_PID_ANY)
 459                 return (-EINVAL);
 460
 461         /* Have a brand new peer ready... */
 462         rc = ksocknal_create_peer(&peer, ni, id);
 463         if (rc != 0)
 464                 return rc;
 465
 466         route = ksocknal_create_route (ipaddr, port);
 467         if (route == NULL) {
 468                 ksocknal_peer_decref(peer);
 469                 return (-ENOMEM);
 470         }
 471
 472         write_lock_bh(&ksocknal_data.ksnd_global_lock);
 473
 474         /* always called with a ref on ni, so shutdown can't have started */
 475         LASSERT (((ksock_net_t *) ni->ni_data)->ksnn_shutdown == 0);
 476
 477         peer2 = ksocknal_find_peer_locked(ni, id);
 478         if (peer2 != NULL) {
 479                 ksocknal_peer_decref(peer);
 480                 peer = peer2;
 481         } else {
 482                 /* peer table takes my ref on peer */
 483                 list_add_tail(&peer->ksnp_list,
 484                               ksocknal_nid2peerlist(id.nid));
 485         }
 486
 487         route2 = NULL;
 488         list_for_each(tmp, &peer->ksnp_routes) {
 489                 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
 490
 491                 if (route2->ksnr_ipaddr == ipaddr)
 492                         break;
 493
 494                 route2 = NULL;
 495         }
 496         if (route2 == NULL) {
 497                 ksocknal_add_route_locked(peer, route);
 498                 route->ksnr_share_count++;
 499         } else {
 500                 ksocknal_route_decref(route);
 501                 route2->ksnr_share_count++;
 502         }
 503
 504         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 505
 506         return 0;
 507 }
 508
 509 static void
 510 ksocknal_del_peer_locked (ksock_peer_t *peer, __u32 ip)
 511 {
 512         ksock_conn_t     *conn;
 513         ksock_route_t    *route;
 514         struct list_head *tmp;
 515         struct list_head *nxt;
 516         int               nshared;
 517
 518         LASSERT(!peer->ksnp_closing);
 519
 520         /* Extra ref prevents peer disappearing until I'm done with it */
 521         ksocknal_peer_addref(peer);
 522
 523         list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
 524                 route = list_entry(tmp, ksock_route_t, ksnr_list);
 525
 526                 /* no match */
 527                 if (!(ip == 0 || route->ksnr_ipaddr == ip))
 528                         continue;
 529
 530                 route->ksnr_share_count = 0;
 531                 /* This deletes associated conns too */
 532                 ksocknal_del_route_locked(route);
 533         }
 534
 535         nshared = 0;
 536         list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
 537                 route = list_entry(tmp, ksock_route_t, ksnr_list);
 538                 nshared += route->ksnr_share_count;
 539         }
 540
 541         if (nshared == 0) {
 542                 /* remove everything else if there are no explicit entries
 543                  * left */
 544
 545                 list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
 546                         route = list_entry(tmp, ksock_route_t, ksnr_list);
 547
 548                         /* we should only be removing auto-entries */
 549                         LASSERT(route->ksnr_share_count == 0);
 550                         ksocknal_del_route_locked(route);
 551                 }
 552
 553                 list_for_each_safe(tmp, nxt, &peer->ksnp_conns) {
 554                         conn = list_entry(tmp, ksock_conn_t, ksnc_list);
 555
 556                         ksocknal_close_conn_locked(conn, 0);
 557                 }
 558         }
 559
 560         ksocknal_peer_decref(peer);
 561                 /* NB peer unlinks itself when last conn/route is removed */
 562 }
 563
 564 static int
 565 ksocknal_del_peer (lnet_ni_t *ni, lnet_process_id_t id, __u32 ip)
 566 {
 567         struct list_head  zombies = LIST_HEAD_INIT(zombies);
 568         struct list_head *ptmp;
 569         struct list_head *pnxt;
 570         ksock_peer_t     *peer;
 571         int               lo;
 572         int               hi;
 573         int               i;
 574         int               rc = -ENOENT;
 575
 576         write_lock_bh(&ksocknal_data.ksnd_global_lock);
 577
 578         if (id.nid != LNET_NID_ANY) {
 579                 hi = (int)(ksocknal_nid2peerlist(id.nid) -
 580                            ksocknal_data.ksnd_peers);
 581                 lo = hi;
 582         } else {
 583                 lo = 0;
 584                 hi = ksocknal_data.ksnd_peer_hash_size - 1;
 585         }
 586
 587         for (i = lo; i <= hi; i++) {
 588                 list_for_each_safe(ptmp, pnxt,
 589                                    &ksocknal_data.ksnd_peers[i]) {
 590                         peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
 591
 592                         if (peer->ksnp_ni != ni)
 593                                 continue;
 594
 595                         if (!((id.nid == LNET_NID_ANY ||
 596                                peer->ksnp_id.nid == id.nid) &&
 597                               (id.pid == LNET_PID_ANY ||
 598                                peer->ksnp_id.pid == id.pid)))
 599                                 continue;
 600
 601                         ksocknal_peer_addref(peer);     /* a ref for me... */
 602
 603                         ksocknal_del_peer_locked(peer, ip);
 604
 605                         if (peer->ksnp_closing &&
 606                             !list_empty(&peer->ksnp_tx_queue)) {
 607                                 LASSERT(list_empty(&peer->ksnp_conns));
 608                                 LASSERT(list_empty(&peer->ksnp_routes));
 609
 610                                 list_splice_init(&peer->ksnp_tx_queue,
 611                                                  &zombies);
 612                         }
 613
 614                         ksocknal_peer_decref(peer);     /* ...till here */
 615
 616                         rc = 0;                         /* matched! */
 617                 }
 618         }
 619
 620         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 621
 622         ksocknal_txlist_done(ni, &zombies, 1);
 623
 624         return rc;
 625 }
 626
 627 static ksock_conn_t *
 628 ksocknal_get_conn_by_idx (lnet_ni_t *ni, int index)
 629 {
 630         ksock_peer_t     *peer;
 631         struct list_head *ptmp;
 632         ksock_conn_t     *conn;
 633         struct list_head *ctmp;
 634         int               i;
 635
 636         read_lock(&ksocknal_data.ksnd_global_lock);
 637
 638         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
 639                 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
 640                         peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
 641
 642                         LASSERT(!peer->ksnp_closing);
 643
 644                         if (peer->ksnp_ni != ni)
 645                                 continue;
 646
 647                         list_for_each(ctmp, &peer->ksnp_conns) {
 648                                 if (index-- > 0)
 649                                         continue;
 650
 651                                 conn = list_entry(ctmp, ksock_conn_t,
 652                                                   ksnc_list);
 653                                 ksocknal_conn_addref(conn);
 654                                 read_unlock(&ksocknal_data. \
 655                                             ksnd_global_lock);
 656                                 return conn;
 657                         }
 658                 }
 659         }
 660
 661         read_unlock(&ksocknal_data.ksnd_global_lock);
 662         return NULL;
 663 }
 664
 665 static ksock_sched_t *
 666 ksocknal_choose_scheduler_locked(unsigned int cpt)
 667 {
 668         struct ksock_sched_info *info = ksocknal_data.ksnd_sched_info[cpt];
 669         ksock_sched_t           *sched;
 670         int                     i;
 671
 672         LASSERT(info->ksi_nthreads > 0);
 673
 674         sched = &info->ksi_scheds[0];
 675         /*
 676          * NB: it's safe so far, but info->ksi_nthreads could be changed
 677          * at runtime when we have dynamic LNet configuration, then we
 678          * need to take care of this.
 679          */
 680         for (i = 1; i < info->ksi_nthreads; i++) {
 681                 if (sched->kss_nconns > info->ksi_scheds[i].kss_nconns)
 682                         sched = &info->ksi_scheds[i];
 683         }
 684
 685         return sched;
 686 }
 687
 688 static int
 689 ksocknal_local_ipvec (lnet_ni_t *ni, __u32 *ipaddrs)
 690 {
 691         ksock_net_t       *net = ni->ni_data;
 692         int                i;
 693         int                nip;
 694
 695         read_lock(&ksocknal_data.ksnd_global_lock);
 696
 697         nip = net->ksnn_ninterfaces;
 698         LASSERT (nip <= LNET_MAX_INTERFACES);
 699
 700         /* Only offer interfaces for additional connections if I have
 701          * more than one. */
 702         if (nip < 2) {
 703                 read_unlock(&ksocknal_data.ksnd_global_lock);
 704                 return 0;
 705         }
 706
 707         for (i = 0; i < nip; i++) {
 708                 ipaddrs[i] = net->ksnn_interfaces[i].ksni_ipaddr;
 709                 LASSERT (ipaddrs[i] != 0);
 710         }
 711
 712         read_unlock(&ksocknal_data.ksnd_global_lock);
 713         return (nip);
 714 }
 715
 716 static int
 717 ksocknal_match_peerip (ksock_interface_t *iface, __u32 *ips, int nips)
 718 {
 719         int   best_netmatch = 0;
 720         int   best_xor      = 0;
 721         int   best          = -1;
 722         int   this_xor;
 723         int   this_netmatch;
 724         int   i;
 725
 726         for (i = 0; i < nips; i++) {
 727                 if (ips[i] == 0)
 728                         continue;
 729
 730                 this_xor = (ips[i] ^ iface->ksni_ipaddr);
 731                 this_netmatch = ((this_xor & iface->ksni_netmask) == 0) ? 1 : 0;
 732
 733                 if (!(best < 0 ||
 734                       best_netmatch < this_netmatch ||
 735                       (best_netmatch == this_netmatch &&
 736                        best_xor > this_xor)))
 737                         continue;
 738
 739                 best = i;
 740                 best_netmatch = this_netmatch;
 741                 best_xor = this_xor;
 742         }
 743
 744         LASSERT (best >= 0);
 745         return (best);
 746 }
 747
 748 static int
 749 ksocknal_select_ips(ksock_peer_t *peer, __u32 *peerips, int n_peerips)
 750 {
 751         rwlock_t                *global_lock = &ksocknal_data.ksnd_global_lock;
 752         ksock_net_t        *net = peer->ksnp_ni->ni_data;
 753         ksock_interface_t  *iface;
 754         ksock_interface_t  *best_iface;
 755         int                 n_ips;
 756         int                 i;
 757         int                 j;
 758         int                 k;
 759         __u32               ip;
 760         __u32               xor;
 761         int                 this_netmatch;
 762         int                 best_netmatch;
 763         int                 best_npeers;
 764
 765         /* CAVEAT EMPTOR: We do all our interface matching with an
 766          * exclusive hold of global lock at IRQ priority.  We're only
 767          * expecting to be dealing with small numbers of interfaces, so the
 768          * O(n**3)-ness shouldn't matter */
 769
 770         /* Also note that I'm not going to return more than n_peerips
 771          * interfaces, even if I have more myself */
 772
 773         write_lock_bh(global_lock);
 774
 775         LASSERT (n_peerips <= LNET_MAX_INTERFACES);
 776         LASSERT (net->ksnn_ninterfaces <= LNET_MAX_INTERFACES);
 777
 778         /* Only match interfaces for additional connections
 779          * if I have > 1 interface */
 780         n_ips = (net->ksnn_ninterfaces < 2) ? 0 :
 781                 MIN(n_peerips, net->ksnn_ninterfaces);
 782
 783         for (i = 0; peer->ksnp_n_passive_ips < n_ips; i++) {
 784                 /*              ^ yes really... */
 785
 786                 /* If we have any new interfaces, first tick off all the
 787                  * peer IPs that match old interfaces, then choose new
 788                  * interfaces to match the remaining peer IPS.
 789                  * We don't forget interfaces we've stopped using; we might
 790                  * start using them again... */
 791
 792                 if (i < peer->ksnp_n_passive_ips) {
 793                         /* Old interface. */
 794                         ip = peer->ksnp_passive_ips[i];
 795                         best_iface = ksocknal_ip2iface(peer->ksnp_ni, ip);
 796
 797                         /* peer passive ips are kept up to date */
 798                         LASSERT(best_iface != NULL);
 799                 } else {
 800                         /* choose a new interface */
 801                         LASSERT (i == peer->ksnp_n_passive_ips);
 802
 803                         best_iface = NULL;
 804                         best_netmatch = 0;
 805                         best_npeers = 0;
 806
 807                         for (j = 0; j < net->ksnn_ninterfaces; j++) {
 808                                 iface = &net->ksnn_interfaces[j];
 809                                 ip = iface->ksni_ipaddr;
 810
 811                                 for (k = 0; k < peer->ksnp_n_passive_ips; k++)
 812                                         if (peer->ksnp_passive_ips[k] == ip)
 813                                                 break;
 814
 815                                 if (k < peer->ksnp_n_passive_ips) /* using it already */
 816                                         continue;
 817
 818                                 k = ksocknal_match_peerip(iface, peerips, n_peerips);
 819                                 xor = (ip ^ peerips[k]);
 820                                 this_netmatch = ((xor & iface->ksni_netmask) == 0) ? 1 : 0;
 821
 822                                 if (!(best_iface == NULL ||
 823                                       best_netmatch < this_netmatch ||
 824                                       (best_netmatch == this_netmatch &&
 825                                        best_npeers > iface->ksni_npeers)))
 826                                         continue;
 827
 828                                 best_iface = iface;
 829                                 best_netmatch = this_netmatch;
 830                                 best_npeers = iface->ksni_npeers;
 831                         }
 832
 833                         LASSERT(best_iface != NULL);
 834
 835                         best_iface->ksni_npeers++;
 836                         ip = best_iface->ksni_ipaddr;
 837                         peer->ksnp_passive_ips[i] = ip;
 838                         peer->ksnp_n_passive_ips = i+1;
 839                 }
 840
 841                 /* mark the best matching peer IP used */
 842                 j = ksocknal_match_peerip(best_iface, peerips, n_peerips);
 843                 peerips[j] = 0;
 844         }
 845
 846         /* Overwrite input peer IP addresses */
 847         memcpy(peerips, peer->ksnp_passive_ips, n_ips * sizeof(*peerips));
 848
 849         write_unlock_bh(global_lock);
 850
 851         return (n_ips);
 852 }
 853
 854 static void
 855 ksocknal_create_routes(ksock_peer_t *peer, int port,
 856                        __u32 *peer_ipaddrs, int npeer_ipaddrs)
 857 {
 858         ksock_route_t           *newroute = NULL;
 859         rwlock_t                *global_lock = &ksocknal_data.ksnd_global_lock;
 860         lnet_ni_t               *ni = peer->ksnp_ni;
 861         ksock_net_t             *net = ni->ni_data;
 862         struct list_head        *rtmp;
 863         ksock_route_t           *route;
 864         ksock_interface_t       *iface;
 865         ksock_interface_t       *best_iface;
 866         int                     best_netmatch;
 867         int                     this_netmatch;
 868         int                     best_nroutes;
 869         int                     i;
 870         int                     j;
 871
 872         /* CAVEAT EMPTOR: We do all our interface matching with an
 873          * exclusive hold of global lock at IRQ priority.  We're only
 874          * expecting to be dealing with small numbers of interfaces, so the
 875          * O(n**3)-ness here shouldn't matter */
 876
 877         write_lock_bh(global_lock);
 878
 879         if (net->ksnn_ninterfaces < 2) {
 880                 /* Only create additional connections
 881                  * if I have > 1 interface */
 882                 write_unlock_bh(global_lock);
 883                 return;
 884         }
 885
 886         LASSERT (npeer_ipaddrs <= LNET_MAX_INTERFACES);
 887
 888         for (i = 0; i < npeer_ipaddrs; i++) {
 889                 if (newroute != NULL) {
 890                         newroute->ksnr_ipaddr = peer_ipaddrs[i];
 891                 } else {
 892                         write_unlock_bh(global_lock);
 893
 894                         newroute = ksocknal_create_route(peer_ipaddrs[i], port);
 895                         if (newroute == NULL)
 896                                 return;
 897
 898                         write_lock_bh(global_lock);
 899                 }
 900
 901                 if (peer->ksnp_closing) {
 902                         /* peer got closed under me */
 903                         break;
 904                 }
 905
 906                 /* Already got a route? */
 907                 route = NULL;
 908                 list_for_each(rtmp, &peer->ksnp_routes) {
 909                         route = list_entry(rtmp, ksock_route_t, ksnr_list);
 910
 911                         if (route->ksnr_ipaddr == newroute->ksnr_ipaddr)
 912                                 break;
 913
 914                         route = NULL;
 915                 }
 916                 if (route != NULL)
 917                         continue;
 918
 919                 best_iface = NULL;
 920                 best_nroutes = 0;
 921                 best_netmatch = 0;
 922
 923                 LASSERT(net->ksnn_ninterfaces <= LNET_MAX_INTERFACES);
 924
 925                 /* Select interface to connect from */
 926                 for (j = 0; j < net->ksnn_ninterfaces; j++) {
 927                         iface = &net->ksnn_interfaces[j];
 928
 929                         /* Using this interface already? */
 930                         list_for_each(rtmp, &peer->ksnp_routes) {
 931                                 route = list_entry(rtmp, ksock_route_t,
 932                                                    ksnr_list);
 933
 934                                 if (route->ksnr_myipaddr == iface->ksni_ipaddr)
 935                                         break;
 936
 937                                 route = NULL;
 938                         }
 939                         if (route != NULL)
 940                                 continue;
 941
 942                         this_netmatch = (((iface->ksni_ipaddr ^
 943                                            newroute->ksnr_ipaddr) &
 944                                            iface->ksni_netmask) == 0) ? 1 : 0;
 945
 946                         if (!(best_iface == NULL ||
 947                               best_netmatch < this_netmatch ||
 948                               (best_netmatch == this_netmatch &&
 949                                best_nroutes > iface->ksni_nroutes)))
 950                                 continue;
 951
 952                         best_iface = iface;
 953                         best_netmatch = this_netmatch;
 954                         best_nroutes = iface->ksni_nroutes;
 955                 }
 956
 957                 if (best_iface == NULL)
 958                         continue;
 959
 960                 newroute->ksnr_myipaddr = best_iface->ksni_ipaddr;
 961                 best_iface->ksni_nroutes++;
 962
 963                 ksocknal_add_route_locked(peer, newroute);
 964                 newroute = NULL;
 965         }
 966
 967         write_unlock_bh(global_lock);
 968         if (newroute != NULL)
 969                 ksocknal_route_decref(newroute);
 970 }
 971
 972 int
 973 ksocknal_accept(lnet_ni_t *ni, struct socket *sock)
 974 {
 975         ksock_connreq_t *cr;
 976         int              rc;
 977         __u32            peer_ip;
 978         int              peer_port;
 979
 980         rc = lnet_sock_getaddr(sock, true, &peer_ip, &peer_port);
 981         LASSERT(rc == 0);               /* we succeeded before */
 982
 983         LIBCFS_ALLOC(cr, sizeof(*cr));
 984         if (cr == NULL) {
 985                 LCONSOLE_ERROR_MSG(0x12f, "Dropping connection request from "
 986                                    "%pI4h: memory exhausted\n", &peer_ip);
 987                 return -ENOMEM;
 988         }
 989
 990         lnet_ni_addref(ni);
 991         cr->ksncr_ni   = ni;
 992         cr->ksncr_sock = sock;
 993
 994         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
 995
 996         list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
 997         wake_up(&ksocknal_data.ksnd_connd_waitq);
 998
 999         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
1000         return 0;
1001 }
1002
1003 static int
1004 ksocknal_connecting (ksock_peer_t *peer, __u32 ipaddr)
1005 {
1006         ksock_route_t *route;
1007
1008         list_for_each_entry(route, &peer->ksnp_routes, ksnr_list) {
1009                 if (route->ksnr_ipaddr == ipaddr)
1010                         return route->ksnr_connecting;
1011         }
1012         return 0;
1013 }
1014
1015 int
1016 ksocknal_create_conn(lnet_ni_t *ni, ksock_route_t *route,
1017                      struct socket *sock, int type)
1018 {
1019         rwlock_t                *global_lock = &ksocknal_data.ksnd_global_lock;
1020         struct list_head        zombies = LIST_HEAD_INIT(zombies);
1021         lnet_process_id_t       peerid;
1022         struct list_head        *tmp;
1023         __u64              incarnation;
1024         ksock_conn_t      *conn;
1025         ksock_conn_t      *conn2;
1026         ksock_peer_t      *peer = NULL;
1027         ksock_peer_t      *peer2;
1028         ksock_sched_t     *sched;
1029         ksock_hello_msg_t *hello;
1030         int                cpt;
1031         ksock_tx_t        *tx;
1032         ksock_tx_t        *txtmp;
1033         int                rc;
1034         int                active;
1035         char              *warn = NULL;
1036
1037         active = (route != NULL);
1038
1039         LASSERT (active == (type != SOCKLND_CONN_NONE));
1040
1041         LIBCFS_ALLOC(conn, sizeof(*conn));
1042         if (conn == NULL) {
1043                 rc = -ENOMEM;
1044                 goto failed_0;
1045         }
1046
1047         conn->ksnc_peer = NULL;
1048         conn->ksnc_route = NULL;
1049         conn->ksnc_sock = sock;
1050         /* 2 ref, 1 for conn, another extra ref prevents socket
1051          * being closed before establishment of connection */
1052         atomic_set (&conn->ksnc_sock_refcount, 2);
1053         conn->ksnc_type = type;
1054         ksocknal_lib_save_callback(sock, conn);
1055         atomic_set (&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
1056
1057         conn->ksnc_rx_ready = 0;
1058         conn->ksnc_rx_scheduled = 0;
1059
1060         INIT_LIST_HEAD(&conn->ksnc_tx_queue);
1061         conn->ksnc_tx_ready = 0;
1062         conn->ksnc_tx_scheduled = 0;
1063         conn->ksnc_tx_carrier = NULL;
1064         atomic_set (&conn->ksnc_tx_nob, 0);
1065
1066         LIBCFS_ALLOC(hello, offsetof(ksock_hello_msg_t,
1067                                      kshm_ips[LNET_MAX_INTERFACES]));
1068         if (hello == NULL) {
1069                 rc = -ENOMEM;
1070                 goto failed_1;
1071         }
1072
1073         /* stash conn's local and remote addrs */
1074         rc = ksocknal_lib_get_conn_addrs (conn);
1075         if (rc != 0)
1076                 goto failed_1;
1077
1078         /* Find out/confirm peer's NID and connection type and get the
1079          * vector of interfaces she's willing to let me connect to.
1080          * Passive connections use the listener timeout since the peer sends
1081          * eagerly */
1082
1083         if (active) {
1084                 peer = route->ksnr_peer;
1085                 LASSERT(ni == peer->ksnp_ni);
1086
1087                 /* Active connection sends HELLO eagerly */
1088                 hello->kshm_nips = ksocknal_local_ipvec(ni, hello->kshm_ips);
1089                 peerid = peer->ksnp_id;
1090
1091                 write_lock_bh(global_lock);
1092                 conn->ksnc_proto = peer->ksnp_proto;
1093                 write_unlock_bh(global_lock);
1094
1095                 if (conn->ksnc_proto == NULL) {
1096                          conn->ksnc_proto = &ksocknal_protocol_v3x;
1097 #if SOCKNAL_VERSION_DEBUG
1098                          if (*ksocknal_tunables.ksnd_protocol == 2)
1099                                  conn->ksnc_proto = &ksocknal_protocol_v2x;
1100                          else if (*ksocknal_tunables.ksnd_protocol == 1)
1101                                  conn->ksnc_proto = &ksocknal_protocol_v1x;
1102 #endif
1103                 }
1104
1105                 rc = ksocknal_send_hello (ni, conn, peerid.nid, hello);
1106                 if (rc != 0)
1107                         goto failed_1;
1108         } else {
1109                 peerid.nid = LNET_NID_ANY;
1110                 peerid.pid = LNET_PID_ANY;
1111
1112                 /* Passive, get protocol from peer */
1113                 conn->ksnc_proto = NULL;
1114         }
1115
1116         rc = ksocknal_recv_hello (ni, conn, hello, &peerid, &incarnation);
1117         if (rc < 0)
1118                 goto failed_1;
1119
1120         LASSERT (rc == 0 || active);
1121         LASSERT (conn->ksnc_proto != NULL);
1122         LASSERT (peerid.nid != LNET_NID_ANY);
1123
1124         cpt = lnet_cpt_of_nid(peerid.nid);
1125
1126         if (active) {
1127                 ksocknal_peer_addref(peer);
1128                 write_lock_bh(global_lock);
1129         } else {
1130                 rc = ksocknal_create_peer(&peer, ni, peerid);
1131                 if (rc != 0)
1132                         goto failed_1;
1133
1134                 write_lock_bh(global_lock);
1135
1136                 /* called with a ref on ni, so shutdown can't have started */
1137                 LASSERT (((ksock_net_t *) ni->ni_data)->ksnn_shutdown == 0);
1138
1139                 peer2 = ksocknal_find_peer_locked(ni, peerid);
1140                 if (peer2 == NULL) {
1141                         /* NB this puts an "empty" peer in the peer
1142                          * table (which takes my ref) */
1143                         list_add_tail(&peer->ksnp_list,
1144                                       ksocknal_nid2peerlist(peerid.nid));
1145                 } else {
1146                         ksocknal_peer_decref(peer);
1147                         peer = peer2;
1148                 }
1149
1150                 /* +1 ref for me */
1151                 ksocknal_peer_addref(peer);
1152                 peer->ksnp_accepting++;
1153
1154                 /* Am I already connecting to this guy?  Resolve in
1155                  * favour of higher NID... */
1156                 if (peerid.nid < ni->ni_nid &&
1157                     ksocknal_connecting(peer, conn->ksnc_ipaddr)) {
1158                         rc = EALREADY;
1159                         warn = "connection race resolution";
1160                         goto failed_2;
1161                 }
1162         }
1163
1164         if (peer->ksnp_closing ||
1165             (active && route->ksnr_deleted)) {
1166                 /* peer/route got closed under me */
1167                 rc = -ESTALE;
1168                 warn = "peer/route removed";
1169                 goto failed_2;
1170         }
1171
1172         if (peer->ksnp_proto == NULL) {
1173                 /* Never connected before.
1174                  * NB recv_hello may have returned EPROTO to signal my peer
1175                  * wants a different protocol than the one I asked for.
1176                  */
1177                 LASSERT(list_empty(&peer->ksnp_conns));
1178
1179                 peer->ksnp_proto = conn->ksnc_proto;
1180                 peer->ksnp_incarnation = incarnation;
1181         }
1182
1183         if (peer->ksnp_proto != conn->ksnc_proto ||
1184             peer->ksnp_incarnation != incarnation) {
1185                 /* Peer rebooted or I've got the wrong protocol version */
1186                 ksocknal_close_peer_conns_locked(peer, 0, 0);
1187
1188                 peer->ksnp_proto = NULL;
1189                 rc = ESTALE;
1190                 warn = peer->ksnp_incarnation != incarnation ?
1191                        "peer rebooted" :
1192                        "wrong proto version";
1193                 goto failed_2;
1194         }
1195
1196         switch (rc) {
1197         default:
1198                 LBUG();
1199         case 0:
1200                 break;
1201         case EALREADY:
1202                 warn = "lost conn race";
1203                 goto failed_2;
1204         case EPROTO:
1205                 warn = "retry with different protocol version";
1206                 goto failed_2;
1207         }
1208
1209         /* Refuse to duplicate an existing connection, unless this is a
1210          * loopback connection */
1211         if (conn->ksnc_ipaddr != conn->ksnc_myipaddr) {
1212                 list_for_each(tmp, &peer->ksnp_conns) {
1213                         conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1214
1215                         if (conn2->ksnc_ipaddr != conn->ksnc_ipaddr ||
1216                             conn2->ksnc_myipaddr != conn->ksnc_myipaddr ||
1217                             conn2->ksnc_type != conn->ksnc_type)
1218                                 continue;
1219
1220                         /* Reply on a passive connection attempt so the peer
1221                          * realises we're connected. */
1222                         LASSERT (rc == 0);
1223                         if (!active)
1224                                 rc = EALREADY;
1225
1226                         warn = "duplicate";
1227                         goto failed_2;
1228                 }
1229         }
1230
1231         /* If the connection created by this route didn't bind to the IP
1232          * address the route connected to, the connection/route matching
1233          * code below probably isn't going to work. */
1234         if (active &&
1235             route->ksnr_ipaddr != conn->ksnc_ipaddr) {
1236                 CERROR("Route %s %pI4h connected to %pI4h\n",
1237                        libcfs_id2str(peer->ksnp_id),
1238                        &route->ksnr_ipaddr,
1239                        &conn->ksnc_ipaddr);
1240         }
1241
1242         /* Search for a route corresponding to the new connection and
1243          * create an association.  This allows incoming connections created
1244          * by routes in my peer to match my own route entries so I don't
1245          * continually create duplicate routes. */
1246         list_for_each(tmp, &peer->ksnp_routes) {
1247                 route = list_entry(tmp, ksock_route_t, ksnr_list);
1248
1249                 if (route->ksnr_ipaddr != conn->ksnc_ipaddr)
1250                         continue;
1251
1252                 ksocknal_associate_route_conn_locked(route, conn);
1253                 break;
1254         }
1255
1256         conn->ksnc_peer = peer;                 /* conn takes my ref on peer */
1257         peer->ksnp_last_alive = cfs_time_current();
1258         peer->ksnp_send_keepalive = 0;
1259         peer->ksnp_error = 0;
1260
1261         sched = ksocknal_choose_scheduler_locked(cpt);
1262         sched->kss_nconns++;
1263         conn->ksnc_scheduler = sched;
1264
1265         conn->ksnc_tx_last_post = cfs_time_current();
1266         /* Set the deadline for the outgoing HELLO to drain */
1267         conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1268         conn->ksnc_tx_deadline = cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1269         smp_mb();   /* order with adding to peer's conn list */
1270
1271         list_add(&conn->ksnc_list, &peer->ksnp_conns);
1272         ksocknal_conn_addref(conn);
1273
1274         ksocknal_new_packet(conn, 0);
1275
1276         conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1277
1278         /* Take packets blocking for this connection. */
1279         list_for_each_entry_safe(tx, txtmp, &peer->ksnp_tx_queue, tx_list) {
1280                 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1281                     SOCKNAL_MATCH_NO)
1282                         continue;
1283
1284                 list_del(&tx->tx_list);
1285                 ksocknal_queue_tx_locked(tx, conn);
1286         }
1287
1288         write_unlock_bh(global_lock);
1289
1290         /* We've now got a new connection.  Any errors from here on are just
1291          * like "normal" comms errors and we close the connection normally.
1292          * NB (a) we still have to send the reply HELLO for passive
1293          *        connections,
1294          *    (b) normal I/O on the conn is blocked until I setup and call the
1295          *        socket callbacks.
1296          */
1297
1298         CDEBUG(D_NET, "New conn %s p %d.x %pI4h -> %pI4h/%d"
1299                " incarnation:%lld sched[%d:%d]\n",
1300                libcfs_id2str(peerid), conn->ksnc_proto->pro_version,
1301                &conn->ksnc_myipaddr, &conn->ksnc_ipaddr,
1302                conn->ksnc_port, incarnation, cpt,
1303                (int)(sched - &sched->kss_info->ksi_scheds[0]));
1304
1305         if (active) {
1306                 /* additional routes after interface exchange? */
1307                 ksocknal_create_routes(peer, conn->ksnc_port,
1308                                        hello->kshm_ips, hello->kshm_nips);
1309         } else {
1310                 hello->kshm_nips = ksocknal_select_ips(peer, hello->kshm_ips,
1311                                                        hello->kshm_nips);
1312                 rc = ksocknal_send_hello(ni, conn, peerid.nid, hello);
1313         }
1314
1315         LIBCFS_FREE(hello, offsetof(ksock_hello_msg_t,
1316                                     kshm_ips[LNET_MAX_INTERFACES]));
1317
1318         /* setup the socket AFTER I've received hello (it disables
1319          * SO_LINGER).  I might call back to the acceptor who may want
1320          * to send a protocol version response and then close the
1321          * socket; this ensures the socket only tears down after the
1322          * response has been sent. */
1323         if (rc == 0)
1324                 rc = ksocknal_lib_setup_sock(sock);
1325
1326         write_lock_bh(global_lock);
1327
1328         /* NB my callbacks block while I hold ksnd_global_lock */
1329         ksocknal_lib_set_callback(sock, conn);
1330
1331         if (!active)
1332                 peer->ksnp_accepting--;
1333
1334         write_unlock_bh(global_lock);
1335
1336         if (rc != 0) {
1337                 write_lock_bh(global_lock);
1338                 if (!conn->ksnc_closing) {
1339                         /* could be closed by another thread */
1340                         ksocknal_close_conn_locked(conn, rc);
1341                 }
1342                 write_unlock_bh(global_lock);
1343         } else if (ksocknal_connsock_addref(conn) == 0) {
1344                 /* Allow I/O to proceed. */
1345                 ksocknal_read_callback(conn);
1346                 ksocknal_write_callback(conn);
1347                 ksocknal_connsock_decref(conn);
1348         }
1349
1350         ksocknal_connsock_decref(conn);
1351         ksocknal_conn_decref(conn);
1352         return rc;
1353
1354 failed_2:
1355         if (!peer->ksnp_closing &&
1356             list_empty(&peer->ksnp_conns) &&
1357             list_empty(&peer->ksnp_routes)) {
1358                 list_add(&zombies, &peer->ksnp_tx_queue);
1359                 list_del_init(&peer->ksnp_tx_queue);
1360                 ksocknal_unlink_peer_locked(peer);
1361         }
1362
1363         write_unlock_bh(global_lock);
1364
1365         if (warn != NULL) {
1366                 if (rc < 0)
1367                         CERROR("Not creating conn %s type %d: %s\n",
1368                                libcfs_id2str(peerid), conn->ksnc_type, warn);
1369                 else
1370                         CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1371                               libcfs_id2str(peerid), conn->ksnc_type, warn);
1372         }
1373
1374         if (!active) {
1375                 if (rc > 0) {
1376                         /* Request retry by replying with CONN_NONE
1377                          * ksnc_proto has been set already */
1378                         conn->ksnc_type = SOCKLND_CONN_NONE;
1379                         hello->kshm_nips = 0;
1380                         ksocknal_send_hello(ni, conn, peerid.nid, hello);
1381                 }
1382
1383                 write_lock_bh(global_lock);
1384                 peer->ksnp_accepting--;
1385                 write_unlock_bh(global_lock);
1386         }
1387
1388         ksocknal_txlist_done(ni, &zombies, 1);
1389         ksocknal_peer_decref(peer);
1390
1391  failed_1:
1392         if (hello != NULL)
1393                 LIBCFS_FREE(hello, offsetof(ksock_hello_msg_t,
1394                                             kshm_ips[LNET_MAX_INTERFACES]));
1395
1396         LIBCFS_FREE(conn, sizeof(*conn));
1397
1398 failed_0:
1399         sock_release(sock);
1400         return rc;
1401 }
1402
1403 void
1404 ksocknal_close_conn_locked (ksock_conn_t *conn, int error)
1405 {
1406         /* This just does the immmediate housekeeping, and queues the
1407          * connection for the reaper to terminate.
1408          * Caller holds ksnd_global_lock exclusively in irq context */
1409         ksock_peer_t      *peer = conn->ksnc_peer;
1410         ksock_route_t     *route;
1411         ksock_conn_t      *conn2;
1412         struct list_head  *tmp;
1413
1414         LASSERT(peer->ksnp_error == 0);
1415         LASSERT(!conn->ksnc_closing);
1416         conn->ksnc_closing = 1;
1417
1418         /* ksnd_deathrow_conns takes over peer's ref */
1419         list_del(&conn->ksnc_list);
1420
1421         route = conn->ksnc_route;
1422         if (route != NULL) {
1423                 /* dissociate conn from route... */
1424                 LASSERT(!route->ksnr_deleted);
1425                 LASSERT((route->ksnr_connected & (1 << conn->ksnc_type)) != 0);
1426
1427                 conn2 = NULL;
1428                 list_for_each(tmp, &peer->ksnp_conns) {
1429                         conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1430
1431                         if (conn2->ksnc_route == route &&
1432                             conn2->ksnc_type == conn->ksnc_type)
1433                                 break;
1434
1435                         conn2 = NULL;
1436                 }
1437                 if (conn2 == NULL)
1438                         route->ksnr_connected &= ~(1 << conn->ksnc_type);
1439
1440                 conn->ksnc_route = NULL;
1441
1442 #if 0           /* irrelevent with only eager routes */
1443                 /* make route least favourite */
1444                 list_del(&route->ksnr_list);
1445                 list_add_tail(&route->ksnr_list, &peer->ksnp_routes);
1446 #endif
1447                 ksocknal_route_decref(route);   /* drop conn's ref on route */
1448         }
1449
1450         if (list_empty(&peer->ksnp_conns)) {
1451                 /* No more connections to this peer */
1452
1453                 if (!list_empty(&peer->ksnp_tx_queue)) {
1454                                 ksock_tx_t *tx;
1455
1456                         LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1457
1458                         /* throw them to the last connection...,
1459                          * these TXs will be send to /dev/null by scheduler */
1460                         list_for_each_entry(tx, &peer->ksnp_tx_queue,
1461                                             tx_list)
1462                                 ksocknal_tx_prep(conn, tx);
1463
1464                         spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1465                         list_splice_init(&peer->ksnp_tx_queue,
1466                                          &conn->ksnc_tx_queue);
1467                         spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1468                 }
1469
1470                 /* renegotiate protocol version */
1471                 peer->ksnp_proto = NULL;
1472                 /* stash last conn close reason */
1473                 peer->ksnp_error = error;
1474
1475                 if (list_empty(&peer->ksnp_routes)) {
1476                         /* I've just closed last conn belonging to a
1477                          * peer with no routes to it */
1478                         ksocknal_unlink_peer_locked(peer);
1479                 }
1480         }
1481
1482         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1483
1484         list_add_tail(&conn->ksnc_list,
1485                       &ksocknal_data.ksnd_deathrow_conns);
1486         wake_up(&ksocknal_data.ksnd_reaper_waitq);
1487
1488         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1489 }
1490
1491 void
1492 ksocknal_peer_failed (ksock_peer_t *peer)
1493 {
1494         int        notify = 0;
1495         cfs_time_t last_alive = 0;
1496
1497         /* There has been a connection failure or comms error; but I'll only
1498          * tell LNET I think the peer is dead if it's to another kernel and
1499          * there are no connections or connection attempts in existence. */
1500
1501         read_lock(&ksocknal_data.ksnd_global_lock);
1502
1503         if ((peer->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1504              list_empty(&peer->ksnp_conns) &&
1505              peer->ksnp_accepting == 0 &&
1506              ksocknal_find_connecting_route_locked(peer) == NULL) {
1507                 notify = 1;
1508                 last_alive = peer->ksnp_last_alive;
1509         }
1510
1511         read_unlock(&ksocknal_data.ksnd_global_lock);
1512
1513         if (notify)
1514                 lnet_notify(peer->ksnp_ni, peer->ksnp_id.nid, 0,
1515                             last_alive);
1516 }
1517
1518 void
1519 ksocknal_finalize_zcreq(ksock_conn_t *conn)
1520 {
1521         ksock_peer_t     *peer = conn->ksnc_peer;
1522         ksock_tx_t       *tx;
1523         ksock_tx_t       *tmp;
1524         struct list_head  zlist = LIST_HEAD_INIT(zlist);
1525
1526         /* NB safe to finalize TXs because closing of socket will
1527          * abort all buffered data */
1528         LASSERT(conn->ksnc_sock == NULL);
1529
1530         spin_lock(&peer->ksnp_lock);
1531
1532         list_for_each_entry_safe(tx, tmp, &peer->ksnp_zc_req_list, tx_zc_list) {
1533                 if (tx->tx_conn != conn)
1534                         continue;
1535
1536                 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1537
1538                 tx->tx_msg.ksm_zc_cookies[0] = 0;
1539                 tx->tx_zc_aborted = 1;  /* mark it as not-acked */
1540                 list_del(&tx->tx_zc_list);
1541                 list_add(&tx->tx_zc_list, &zlist);
1542         }
1543
1544         spin_unlock(&peer->ksnp_lock);
1545
1546         while (!list_empty(&zlist)) {
1547                 tx = list_entry(zlist.next, ksock_tx_t, tx_zc_list);
1548
1549                 list_del(&tx->tx_zc_list);
1550                 ksocknal_tx_decref(tx);
1551         }
1552 }
1553
1554 void
1555 ksocknal_terminate_conn(ksock_conn_t *conn)
1556 {
1557         /* This gets called by the reaper (guaranteed thread context) to
1558          * disengage the socket from its callbacks and close it.
1559          * ksnc_refcount will eventually hit zero, and then the reaper will
1560          * destroy it. */
1561         ksock_peer_t     *peer = conn->ksnc_peer;
1562         ksock_sched_t    *sched = conn->ksnc_scheduler;
1563         int               failed = 0;
1564
1565         LASSERT(conn->ksnc_closing);
1566
1567         /* wake up the scheduler to "send" all remaining packets to /dev/null */
1568         spin_lock_bh(&sched->kss_lock);
1569
1570         /* a closing conn is always ready to tx */
1571         conn->ksnc_tx_ready = 1;
1572
1573         if (!conn->ksnc_tx_scheduled &&
1574             !list_empty(&conn->ksnc_tx_queue)) {
1575                 list_add_tail(&conn->ksnc_tx_list,
1576                                &sched->kss_tx_conns);
1577                 conn->ksnc_tx_scheduled = 1;
1578                 /* extra ref for scheduler */
1579                 ksocknal_conn_addref(conn);
1580
1581                 wake_up (&sched->kss_waitq);
1582         }
1583
1584         spin_unlock_bh(&sched->kss_lock);
1585
1586         /* serialise with callbacks */
1587         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1588
1589         ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1590
1591         /* OK, so this conn may not be completely disengaged from its
1592          * scheduler yet, but it _has_ committed to terminate... */
1593         conn->ksnc_scheduler->kss_nconns--;
1594
1595         if (peer->ksnp_error != 0) {
1596                 /* peer's last conn closed in error */
1597                 LASSERT(list_empty(&peer->ksnp_conns));
1598                 failed = 1;
1599                 peer->ksnp_error = 0;     /* avoid multiple notifications */
1600         }
1601
1602         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1603
1604         if (failed)
1605                 ksocknal_peer_failed(peer);
1606
1607         /* The socket is closed on the final put; either here, or in
1608          * ksocknal_{send,recv}msg().  Since we set up the linger2 option
1609          * when the connection was established, this will close the socket
1610          * immediately, aborting anything buffered in it. Any hung
1611          * zero-copy transmits will therefore complete in finite time. */
1612         ksocknal_connsock_decref(conn);
1613 }
1614
1615 void
1616 ksocknal_queue_zombie_conn (ksock_conn_t *conn)
1617 {
1618         /* Queue the conn for the reaper to destroy */
1619
1620         LASSERT(atomic_read(&conn->ksnc_conn_refcount) == 0);
1621         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1622
1623         list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1624         wake_up(&ksocknal_data.ksnd_reaper_waitq);
1625
1626         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1627 }
1628
1629 void
1630 ksocknal_destroy_conn (ksock_conn_t *conn)
1631 {
1632         cfs_time_t      last_rcv;
1633
1634         /* Final coup-de-grace of the reaper */
1635         CDEBUG (D_NET, "connection %p\n", conn);
1636
1637         LASSERT (atomic_read (&conn->ksnc_conn_refcount) == 0);
1638         LASSERT (atomic_read (&conn->ksnc_sock_refcount) == 0);
1639         LASSERT (conn->ksnc_sock == NULL);
1640         LASSERT (conn->ksnc_route == NULL);
1641         LASSERT (!conn->ksnc_tx_scheduled);
1642         LASSERT (!conn->ksnc_rx_scheduled);
1643         LASSERT(list_empty(&conn->ksnc_tx_queue));
1644
1645         /* complete current receive if any */
1646         switch (conn->ksnc_rx_state) {
1647         case SOCKNAL_RX_LNET_PAYLOAD:
1648                 last_rcv = conn->ksnc_rx_deadline -
1649                            cfs_time_seconds(*ksocknal_tunables.ksnd_timeout);
1650                 CERROR("Completing partial receive from %s[%d], "
1651                        "ip %pI4h:%d, with error, wanted: %d, left: %d, "
1652                        "last alive is %ld secs ago\n",
1653                        libcfs_id2str(conn->ksnc_peer->ksnp_id), conn->ksnc_type,
1654                        &conn->ksnc_ipaddr, conn->ksnc_port,
1655                        conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1656                        cfs_duration_sec(cfs_time_sub(cfs_time_current(),
1657                                         last_rcv)));
1658                 lnet_finalize (conn->ksnc_peer->ksnp_ni,
1659                                conn->ksnc_cookie, -EIO);
1660                 break;
1661         case SOCKNAL_RX_LNET_HEADER:
1662                 if (conn->ksnc_rx_started)
1663                         CERROR("Incomplete receive of lnet header from %s, "
1664                                "ip %pI4h:%d, with error, protocol: %d.x.\n",
1665                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1666                                &conn->ksnc_ipaddr, conn->ksnc_port,
1667                                conn->ksnc_proto->pro_version);
1668                 break;
1669         case SOCKNAL_RX_KSM_HEADER:
1670                 if (conn->ksnc_rx_started)
1671                         CERROR("Incomplete receive of ksock message from %s, "
1672                                "ip %pI4h:%d, with error, protocol: %d.x.\n",
1673                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1674                                &conn->ksnc_ipaddr, conn->ksnc_port,
1675                                conn->ksnc_proto->pro_version);
1676                 break;
1677         case SOCKNAL_RX_SLOP:
1678                 if (conn->ksnc_rx_started)
1679                         CERROR("Incomplete receive of slops from %s, "
1680                                "ip %pI4h:%d, with error\n",
1681                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1682                                &conn->ksnc_ipaddr, conn->ksnc_port);
1683                break;
1684         default:
1685                 LBUG ();
1686                 break;
1687         }
1688
1689         ksocknal_peer_decref(conn->ksnc_peer);
1690
1691         LIBCFS_FREE (conn, sizeof (*conn));
1692 }
1693
1694 int
1695 ksocknal_close_peer_conns_locked (ksock_peer_t *peer, __u32 ipaddr, int why)
1696 {
1697         ksock_conn_t       *conn;
1698         struct list_head         *ctmp;
1699         struct list_head         *cnxt;
1700         int                 count = 0;
1701
1702         list_for_each_safe(ctmp, cnxt, &peer->ksnp_conns) {
1703                 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
1704
1705                 if (ipaddr == 0 ||
1706                     conn->ksnc_ipaddr == ipaddr) {
1707                         count++;
1708                         ksocknal_close_conn_locked (conn, why);
1709                 }
1710         }
1711
1712         return (count);
1713 }
1714
1715 int
1716 ksocknal_close_conn_and_siblings (ksock_conn_t *conn, int why)
1717 {
1718         ksock_peer_t     *peer = conn->ksnc_peer;
1719         __u32             ipaddr = conn->ksnc_ipaddr;
1720         int               count;
1721
1722         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1723
1724         count = ksocknal_close_peer_conns_locked (peer, ipaddr, why);
1725
1726         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1727
1728         return (count);
1729 }
1730
1731 int
1732 ksocknal_close_matching_conns (lnet_process_id_t id, __u32 ipaddr)
1733 {
1734         ksock_peer_t       *peer;
1735         struct list_head         *ptmp;
1736         struct list_head         *pnxt;
1737         int                 lo;
1738         int                 hi;
1739         int                 i;
1740         int                 count = 0;
1741
1742         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1743
1744         if (id.nid != LNET_NID_ANY)
1745                 lo = hi = (int)(ksocknal_nid2peerlist(id.nid) - ksocknal_data.ksnd_peers);
1746         else {
1747                 lo = 0;
1748                 hi = ksocknal_data.ksnd_peer_hash_size - 1;
1749         }
1750
1751         for (i = lo; i <= hi; i++) {
1752                 list_for_each_safe(ptmp, pnxt, &ksocknal_data.ksnd_peers[i]) {
1753
1754                         peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
1755
1756                         if (!((id.nid == LNET_NID_ANY || id.nid == peer->ksnp_id.nid) &&
1757                               (id.pid == LNET_PID_ANY || id.pid == peer->ksnp_id.pid)))
1758                                 continue;
1759
1760                         count += ksocknal_close_peer_conns_locked (peer, ipaddr, 0);
1761                 }
1762         }
1763
1764         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1765
1766         /* wildcards always succeed */
1767         if (id.nid == LNET_NID_ANY || id.pid == LNET_PID_ANY || ipaddr == 0)
1768                 return (0);
1769
1770         return (count == 0 ? -ENOENT : 0);
1771 }
1772
1773 void
1774 ksocknal_notify (lnet_ni_t *ni, lnet_nid_t gw_nid, int alive)
1775 {
1776         /* The router is telling me she's been notified of a change in
1777          * gateway state.... */
1778         lnet_process_id_t  id = {0};
1779
1780         id.nid = gw_nid;
1781         id.pid = LNET_PID_ANY;
1782
1783         CDEBUG (D_NET, "gw %s %s\n", libcfs_nid2str(gw_nid),
1784                 alive ? "up" : "down");
1785
1786         if (!alive) {
1787                 /* If the gateway crashed, close all open connections... */
1788                 ksocknal_close_matching_conns (id, 0);
1789                 return;
1790         }
1791
1792         /* ...otherwise do nothing.  We can only establish new connections
1793          * if we have autroutes, and these connect on demand. */
1794 }
1795
1796 void
1797 ksocknal_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1798 {
1799         int                connect = 1;
1800         cfs_time_t         last_alive = 0;
1801         cfs_time_t         now = cfs_time_current();
1802         ksock_peer_t      *peer = NULL;
1803         rwlock_t                *glock = &ksocknal_data.ksnd_global_lock;
1804         lnet_process_id_t  id = {
1805                 .nid = nid,
1806                 .pid = LNET_PID_LUSTRE,
1807         };
1808
1809         read_lock(glock);
1810
1811         peer = ksocknal_find_peer_locked(ni, id);
1812         if (peer != NULL) {
1813                 struct list_head       *tmp;
1814                 ksock_conn_t     *conn;
1815                 int               bufnob;
1816
1817                 list_for_each(tmp, &peer->ksnp_conns) {
1818                         conn = list_entry(tmp, ksock_conn_t, ksnc_list);
1819                         bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
1820
1821                         if (bufnob < conn->ksnc_tx_bufnob) {
1822                                 /* something got ACKed */
1823                                 conn->ksnc_tx_deadline =
1824                                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1825                                 peer->ksnp_last_alive = now;
1826                                 conn->ksnc_tx_bufnob = bufnob;
1827                         }
1828                 }
1829
1830                 last_alive = peer->ksnp_last_alive;
1831                 if (ksocknal_find_connectable_route_locked(peer) == NULL)
1832                         connect = 0;
1833         }
1834
1835         read_unlock(glock);
1836
1837         if (last_alive != 0)
1838                 *when = last_alive;
1839
1840         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago, connect %d\n",
1841                libcfs_nid2str(nid), peer,
1842                last_alive ? cfs_duration_sec(now - last_alive) : -1,
1843                connect);
1844
1845         if (!connect)
1846                 return;
1847
1848         ksocknal_add_peer(ni, id, LNET_NIDADDR(nid), lnet_acceptor_port());
1849
1850         write_lock_bh(glock);
1851
1852         peer = ksocknal_find_peer_locked(ni, id);
1853         if (peer != NULL)
1854                 ksocknal_launch_all_connections_locked(peer);
1855
1856         write_unlock_bh(glock);
1857         return;
1858 }
1859
1860 static void
1861 ksocknal_push_peer (ksock_peer_t *peer)
1862 {
1863         int               index;
1864         int               i;
1865         struct list_head       *tmp;
1866         ksock_conn_t     *conn;
1867
1868         for (index = 0; ; index++) {
1869                 read_lock(&ksocknal_data.ksnd_global_lock);
1870
1871                 i = 0;
1872                 conn = NULL;
1873
1874                 list_for_each(tmp, &peer->ksnp_conns) {
1875                         if (i++ == index) {
1876                                 conn = list_entry(tmp, ksock_conn_t,
1877                                                        ksnc_list);
1878                                 ksocknal_conn_addref(conn);
1879                                 break;
1880                         }
1881                 }
1882
1883                 read_unlock(&ksocknal_data.ksnd_global_lock);
1884
1885                 if (conn == NULL)
1886                         break;
1887
1888                 ksocknal_lib_push_conn (conn);
1889                 ksocknal_conn_decref(conn);
1890         }
1891 }
1892
1893 static int
1894 ksocknal_push (lnet_ni_t *ni, lnet_process_id_t id)
1895 {
1896         struct list_head *start;
1897         struct list_head *end;
1898         struct list_head *tmp;
1899         int               rc = -ENOENT;
1900         unsigned int      hsize = ksocknal_data.ksnd_peer_hash_size;
1901
1902         if (id.nid == LNET_NID_ANY) {
1903                 start = &ksocknal_data.ksnd_peers[0];
1904                 end = &ksocknal_data.ksnd_peers[hsize - 1];
1905         } else {
1906                 start = end = ksocknal_nid2peerlist(id.nid);
1907         }
1908
1909         for (tmp = start; tmp <= end; tmp++) {
1910                 int     peer_off; /* searching offset in peer hash table */
1911
1912                 for (peer_off = 0; ; peer_off++) {
1913                         ksock_peer_t *peer;
1914                         int           i = 0;
1915
1916                         read_lock(&ksocknal_data.ksnd_global_lock);
1917                         list_for_each_entry(peer, tmp, ksnp_list) {
1918                                 if (!((id.nid == LNET_NID_ANY ||
1919                                        id.nid == peer->ksnp_id.nid) &&
1920                                       (id.pid == LNET_PID_ANY ||
1921                                        id.pid == peer->ksnp_id.pid)))
1922                                         continue;
1923
1924                                 if (i++ == peer_off) {
1925                                         ksocknal_peer_addref(peer);
1926                                         break;
1927                                 }
1928                         }
1929                         read_unlock(&ksocknal_data.ksnd_global_lock);
1930
1931                         if (i == 0) /* no match */
1932                                 break;
1933
1934                         rc = 0;
1935                         ksocknal_push_peer(peer);
1936                         ksocknal_peer_decref(peer);
1937                 }
1938         }
1939         return rc;
1940 }
1941
1942 static int
1943 ksocknal_add_interface(lnet_ni_t *ni, __u32 ipaddress, __u32 netmask)
1944 {
1945         ksock_net_t       *net = ni->ni_data;
1946         ksock_interface_t *iface;
1947         int                rc;
1948         int                i;
1949         int                j;
1950         struct list_head        *ptmp;
1951         ksock_peer_t      *peer;
1952         struct list_head        *rtmp;
1953         ksock_route_t     *route;
1954
1955         if (ipaddress == 0 ||
1956             netmask == 0)
1957                 return (-EINVAL);
1958
1959         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1960
1961         iface = ksocknal_ip2iface(ni, ipaddress);
1962         if (iface != NULL) {
1963                 /* silently ignore dups */
1964                 rc = 0;
1965         } else if (net->ksnn_ninterfaces == LNET_MAX_INTERFACES) {
1966                 rc = -ENOSPC;
1967         } else {
1968                 iface = &net->ksnn_interfaces[net->ksnn_ninterfaces++];
1969
1970                 iface->ksni_ipaddr = ipaddress;
1971                 iface->ksni_netmask = netmask;
1972                 iface->ksni_nroutes = 0;
1973                 iface->ksni_npeers = 0;
1974
1975                 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1976                         list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
1977                                 peer = list_entry(ptmp, ksock_peer_t,
1978                                                       ksnp_list);
1979
1980                                 for (j = 0; j < peer->ksnp_n_passive_ips; j++)
1981                                         if (peer->ksnp_passive_ips[j] == ipaddress)
1982                                                 iface->ksni_npeers++;
1983
1984                                 list_for_each(rtmp, &peer->ksnp_routes) {
1985                                         route = list_entry(rtmp,
1986                                                                ksock_route_t,
1987                                                                ksnr_list);
1988
1989                                         if (route->ksnr_myipaddr == ipaddress)
1990                                                 iface->ksni_nroutes++;
1991                                 }
1992                         }
1993                 }
1994
1995                 rc = 0;
1996                 /* NB only new connections will pay attention to the new interface! */
1997         }
1998
1999         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2000
2001         return (rc);
2002 }
2003
2004 static void
2005 ksocknal_peer_del_interface_locked(ksock_peer_t *peer, __u32 ipaddr)
2006 {
2007         struct list_head         *tmp;
2008         struct list_head         *nxt;
2009         ksock_route_t      *route;
2010         ksock_conn_t       *conn;
2011         int                 i;
2012         int                 j;
2013
2014         for (i = 0; i < peer->ksnp_n_passive_ips; i++)
2015                 if (peer->ksnp_passive_ips[i] == ipaddr) {
2016                         for (j = i+1; j < peer->ksnp_n_passive_ips; j++)
2017                                 peer->ksnp_passive_ips[j-1] =
2018                                         peer->ksnp_passive_ips[j];
2019                         peer->ksnp_n_passive_ips--;
2020                         break;
2021                 }
2022
2023         list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
2024                 route = list_entry(tmp, ksock_route_t, ksnr_list);
2025
2026                 if (route->ksnr_myipaddr != ipaddr)
2027                         continue;
2028
2029                 if (route->ksnr_share_count != 0) {
2030                         /* Manually created; keep, but unbind */
2031                         route->ksnr_myipaddr = 0;
2032                 } else {
2033                         ksocknal_del_route_locked(route);
2034                 }
2035         }
2036
2037         list_for_each_safe(tmp, nxt, &peer->ksnp_conns) {
2038                 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2039
2040                 if (conn->ksnc_myipaddr == ipaddr)
2041                         ksocknal_close_conn_locked (conn, 0);
2042         }
2043 }
2044
2045 static int
2046 ksocknal_del_interface(lnet_ni_t *ni, __u32 ipaddress)
2047 {
2048         ksock_net_t       *net = ni->ni_data;
2049         int                rc = -ENOENT;
2050         struct list_head        *tmp;
2051         struct list_head        *nxt;
2052         ksock_peer_t      *peer;
2053         __u32              this_ip;
2054         int                i;
2055         int                j;
2056
2057         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2058
2059         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2060                 this_ip = net->ksnn_interfaces[i].ksni_ipaddr;
2061
2062                 if (!(ipaddress == 0 ||
2063                       ipaddress == this_ip))
2064                         continue;
2065
2066                 rc = 0;
2067
2068                 for (j = i+1; j < net->ksnn_ninterfaces; j++)
2069                         net->ksnn_interfaces[j-1] =
2070                                 net->ksnn_interfaces[j];
2071
2072                 net->ksnn_ninterfaces--;
2073
2074                 for (j = 0; j < ksocknal_data.ksnd_peer_hash_size; j++) {
2075                         list_for_each_safe(tmp, nxt,
2076                                                &ksocknal_data.ksnd_peers[j]) {
2077                                 peer = list_entry(tmp, ksock_peer_t,
2078                                                       ksnp_list);
2079
2080                                 if (peer->ksnp_ni != ni)
2081                                         continue;
2082
2083                                 ksocknal_peer_del_interface_locked(peer, this_ip);
2084                         }
2085                 }
2086         }
2087
2088         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2089
2090         return (rc);
2091 }
2092
2093 int
2094 ksocknal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
2095 {
2096         lnet_process_id_t id = {0};
2097         struct libcfs_ioctl_data *data = arg;
2098         int rc;
2099
2100         switch(cmd) {
2101         case IOC_LIBCFS_GET_INTERFACE: {
2102                 ksock_net_t       *net = ni->ni_data;
2103                 ksock_interface_t *iface;
2104
2105                 read_lock(&ksocknal_data.ksnd_global_lock);
2106
2107                 if (data->ioc_count >= (__u32)net->ksnn_ninterfaces) {
2108                         rc = -ENOENT;
2109                 } else {
2110                         rc = 0;
2111                         iface = &net->ksnn_interfaces[data->ioc_count];
2112
2113                         data->ioc_u32[0] = iface->ksni_ipaddr;
2114                         data->ioc_u32[1] = iface->ksni_netmask;
2115                         data->ioc_u32[2] = iface->ksni_npeers;
2116                         data->ioc_u32[3] = iface->ksni_nroutes;
2117                 }
2118
2119                 read_unlock(&ksocknal_data.ksnd_global_lock);
2120                 return rc;
2121         }
2122
2123         case IOC_LIBCFS_ADD_INTERFACE:
2124                 return ksocknal_add_interface(ni,
2125                                               data->ioc_u32[0], /* IP address */
2126                                               data->ioc_u32[1]); /* net mask */
2127
2128         case IOC_LIBCFS_DEL_INTERFACE:
2129                 return ksocknal_del_interface(ni,
2130                                               data->ioc_u32[0]); /* IP address */
2131
2132         case IOC_LIBCFS_GET_PEER: {
2133                 __u32            myip = 0;
2134                 __u32            ip = 0;
2135                 int              port = 0;
2136                 int              conn_count = 0;
2137                 int              share_count = 0;
2138
2139                 rc = ksocknal_get_peer_info(ni, data->ioc_count,
2140                                             &id, &myip, &ip, &port,
2141                                             &conn_count,  &share_count);
2142                 if (rc != 0)
2143                         return rc;
2144
2145                 data->ioc_nid    = id.nid;
2146                 data->ioc_count  = share_count;
2147                 data->ioc_u32[0] = ip;
2148                 data->ioc_u32[1] = port;
2149                 data->ioc_u32[2] = myip;
2150                 data->ioc_u32[3] = conn_count;
2151                 data->ioc_u32[4] = id.pid;
2152                 return 0;
2153         }
2154
2155         case IOC_LIBCFS_ADD_PEER:
2156                 id.nid = data->ioc_nid;
2157                 id.pid = LNET_PID_LUSTRE;
2158                 return ksocknal_add_peer (ni, id,
2159                                           data->ioc_u32[0], /* IP */
2160                                           data->ioc_u32[1]); /* port */
2161
2162         case IOC_LIBCFS_DEL_PEER:
2163                 id.nid = data->ioc_nid;
2164                 id.pid = LNET_PID_ANY;
2165                 return ksocknal_del_peer (ni, id,
2166                                           data->ioc_u32[0]); /* IP */
2167
2168         case IOC_LIBCFS_GET_CONN: {
2169                 int           txmem;
2170                 int           rxmem;
2171                 int           nagle;
2172                 ksock_conn_t *conn = ksocknal_get_conn_by_idx (ni, data->ioc_count);
2173
2174                 if (conn == NULL)
2175                         return -ENOENT;
2176
2177                 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
2178
2179                 data->ioc_count  = txmem;
2180                 data->ioc_nid    = conn->ksnc_peer->ksnp_id.nid;
2181                 data->ioc_flags  = nagle;
2182                 data->ioc_u32[0] = conn->ksnc_ipaddr;
2183                 data->ioc_u32[1] = conn->ksnc_port;
2184                 data->ioc_u32[2] = conn->ksnc_myipaddr;
2185                 data->ioc_u32[3] = conn->ksnc_type;
2186                 data->ioc_u32[4] = conn->ksnc_scheduler->kss_info->ksi_cpt;
2187                 data->ioc_u32[5] = rxmem;
2188                 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
2189                 ksocknal_conn_decref(conn);
2190                 return 0;
2191         }
2192
2193         case IOC_LIBCFS_CLOSE_CONNECTION:
2194                 id.nid = data->ioc_nid;
2195                 id.pid = LNET_PID_ANY;
2196                 return ksocknal_close_matching_conns (id,
2197                                                       data->ioc_u32[0]);
2198
2199         case IOC_LIBCFS_REGISTER_MYNID:
2200                 /* Ignore if this is a noop */
2201                 if (data->ioc_nid == ni->ni_nid)
2202                         return 0;
2203
2204                 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
2205                        libcfs_nid2str(data->ioc_nid),
2206                        libcfs_nid2str(ni->ni_nid));
2207                 return -EINVAL;
2208
2209         case IOC_LIBCFS_PUSH_CONNECTION:
2210                 id.nid = data->ioc_nid;
2211                 id.pid = LNET_PID_ANY;
2212                 return ksocknal_push(ni, id);
2213
2214         default:
2215                 return -EINVAL;
2216         }
2217         /* not reached */
2218 }
2219
2220 static void
2221 ksocknal_free_buffers (void)
2222 {
2223         LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
2224
2225         if (ksocknal_data.ksnd_sched_info != NULL) {
2226                 struct ksock_sched_info *info;
2227                 int                     i;
2228
2229                 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2230                         if (info->ksi_scheds != NULL) {
2231                                 LIBCFS_FREE(info->ksi_scheds,
2232                                             info->ksi_nthreads_max *
2233                                             sizeof(info->ksi_scheds[0]));
2234                         }
2235                 }
2236                 cfs_percpt_free(ksocknal_data.ksnd_sched_info);
2237         }
2238
2239         LIBCFS_FREE (ksocknal_data.ksnd_peers,
2240                      sizeof(struct list_head) *
2241                      ksocknal_data.ksnd_peer_hash_size);
2242
2243         spin_lock(&ksocknal_data.ksnd_tx_lock);
2244
2245         if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
2246                 struct list_head        zlist;
2247                 ksock_tx_t      *tx;
2248
2249                 list_add(&zlist, &ksocknal_data.ksnd_idle_noop_txs);
2250                 list_del_init(&ksocknal_data.ksnd_idle_noop_txs);
2251                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2252
2253                 while (!list_empty(&zlist)) {
2254                         tx = list_entry(zlist.next, ksock_tx_t, tx_list);
2255                         list_del(&tx->tx_list);
2256                         LIBCFS_FREE(tx, tx->tx_desc_size);
2257                 }
2258         } else {
2259                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2260         }
2261 }
2262
2263 static void
2264 ksocknal_base_shutdown(void)
2265 {
2266         struct ksock_sched_info *info;
2267         ksock_sched_t           *sched;
2268         int                     i;
2269         int                     j;
2270
2271         CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2272                atomic_read (&libcfs_kmemory));
2273         LASSERT (ksocknal_data.ksnd_nnets == 0);
2274
2275         switch (ksocknal_data.ksnd_init) {
2276         default:
2277                 LASSERT (0);
2278
2279         case SOCKNAL_INIT_ALL:
2280         case SOCKNAL_INIT_DATA:
2281                 LASSERT (ksocknal_data.ksnd_peers != NULL);
2282                 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2283                         LASSERT(list_empty(&ksocknal_data.ksnd_peers[i]));
2284                 }
2285
2286                 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2287                 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2288                 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2289                 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2290                 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2291
2292                 if (ksocknal_data.ksnd_sched_info != NULL) {
2293                         cfs_percpt_for_each(info, i,
2294                                             ksocknal_data.ksnd_sched_info) {
2295                                 if (info->ksi_scheds == NULL)
2296                                         continue;
2297
2298                                 for (j = 0; j < info->ksi_nthreads_max; j++) {
2299
2300                                         sched = &info->ksi_scheds[j];
2301                                         LASSERT(list_empty(&sched->\
2302                                                                kss_tx_conns));
2303                                         LASSERT(list_empty(&sched->\
2304                                                                kss_rx_conns));
2305                                         LASSERT(list_empty(&sched-> \
2306                                                   kss_zombie_noop_txs));
2307                                         LASSERT(sched->kss_nconns == 0);
2308                                 }
2309                         }
2310                 }
2311
2312                 /* flag threads to terminate; wake and wait for them to die */
2313                 ksocknal_data.ksnd_shuttingdown = 1;
2314                 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2315                 wake_up_all(&ksocknal_data.ksnd_reaper_waitq);
2316
2317                 if (ksocknal_data.ksnd_sched_info != NULL) {
2318                         cfs_percpt_for_each(info, i,
2319                                             ksocknal_data.ksnd_sched_info) {
2320                                 if (info->ksi_scheds == NULL)
2321                                         continue;
2322
2323                                 for (j = 0; j < info->ksi_nthreads_max; j++) {
2324                                         sched = &info->ksi_scheds[j];
2325                                         wake_up_all(&sched->kss_waitq);
2326                                 }
2327                         }
2328                 }
2329
2330                 i = 4;
2331                 read_lock(&ksocknal_data.ksnd_global_lock);
2332                 while (ksocknal_data.ksnd_nthreads != 0) {
2333                         i++;
2334                         /* power of 2? */
2335                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2336                                 "waiting for %d threads to terminate\n",
2337                                 ksocknal_data.ksnd_nthreads);
2338                         read_unlock(&ksocknal_data.ksnd_global_lock);
2339                         set_current_state(TASK_UNINTERRUPTIBLE);
2340                         schedule_timeout(cfs_time_seconds(1));
2341                         read_lock(&ksocknal_data.ksnd_global_lock);
2342                 }
2343                 read_unlock(&ksocknal_data.ksnd_global_lock);
2344
2345                 ksocknal_free_buffers();
2346
2347                 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2348                 break;
2349         }
2350
2351         CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2352                atomic_read (&libcfs_kmemory));
2353
2354         module_put(THIS_MODULE);
2355 }
2356
2357 static __u64 ksocknal_new_incarnation(void)
2358 {
2359         struct timeval tv;
2360
2361         /* The incarnation number is the time this module loaded and it
2362          * identifies this particular instance of the socknal.  Hopefully
2363          * we won't be able to reboot more frequently than 1MHz for the
2364          * forseeable future :) */
2365
2366         do_gettimeofday(&tv);
2367
2368         return (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2369 }
2370
2371 static int
2372 ksocknal_base_startup(void)
2373 {
2374         struct ksock_sched_info *info;
2375         int                     rc;
2376         int                     i;
2377
2378         LASSERT (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2379         LASSERT (ksocknal_data.ksnd_nnets == 0);
2380
2381         memset (&ksocknal_data, 0, sizeof (ksocknal_data)); /* zero pointers */
2382
2383         ksocknal_data.ksnd_peer_hash_size = SOCKNAL_PEER_HASH_SIZE;
2384         LIBCFS_ALLOC(ksocknal_data.ksnd_peers,
2385                      sizeof(struct list_head) *
2386                      ksocknal_data.ksnd_peer_hash_size);
2387         if (ksocknal_data.ksnd_peers == NULL)
2388                 return -ENOMEM;
2389
2390         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++)
2391                 INIT_LIST_HEAD(&ksocknal_data.ksnd_peers[i]);
2392
2393         rwlock_init(&ksocknal_data.ksnd_global_lock);
2394         INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2395
2396         spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2397         INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2398         INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2399         INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2400         init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2401
2402         spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2403         INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2404         INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2405         init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2406
2407         spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2408         INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2409
2410         /* NB memset above zeros whole of ksocknal_data */
2411
2412         /* flag lists/ptrs/locks initialised */
2413         ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2414         try_module_get(THIS_MODULE);
2415
2416         ksocknal_data.ksnd_sched_info = cfs_percpt_alloc(lnet_cpt_table(),
2417                                                          sizeof(*info));
2418         if (ksocknal_data.ksnd_sched_info == NULL)
2419                 goto failed;
2420
2421         cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2422                 ksock_sched_t   *sched;
2423                 int             nthrs;
2424
2425                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2426                 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2427                         nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2428                 } else {
2429                         /* max to half of CPUs, assume another half should be
2430                          * reserved for upper layer modules */
2431                         nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2432                 }
2433
2434                 info->ksi_nthreads_max = nthrs;
2435                 info->ksi_cpt = i;
2436
2437                 LIBCFS_CPT_ALLOC(info->ksi_scheds, lnet_cpt_table(), i,
2438                                  info->ksi_nthreads_max * sizeof(*sched));
2439                 if (info->ksi_scheds == NULL)
2440                         goto failed;
2441
2442                 for (; nthrs > 0; nthrs--) {
2443                         sched = &info->ksi_scheds[nthrs - 1];
2444
2445                         sched->kss_info = info;
2446                         spin_lock_init(&sched->kss_lock);
2447                         INIT_LIST_HEAD(&sched->kss_rx_conns);
2448                         INIT_LIST_HEAD(&sched->kss_tx_conns);
2449                         INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2450                         init_waitqueue_head(&sched->kss_waitq);
2451                 }
2452         }
2453
2454         ksocknal_data.ksnd_connd_starting         = 0;
2455         ksocknal_data.ksnd_connd_failed_stamp     = 0;
2456         ksocknal_data.ksnd_connd_starting_stamp   = cfs_time_current_sec();
2457         /* must have at least 2 connds to remain responsive to accepts while
2458          * connecting */
2459         if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2460                 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2461
2462         if (*ksocknal_tunables.ksnd_nconnds_max <
2463             *ksocknal_tunables.ksnd_nconnds) {
2464                 ksocknal_tunables.ksnd_nconnds_max =
2465                         ksocknal_tunables.ksnd_nconnds;
2466         }
2467
2468         for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2469                 char name[16];
2470                 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2471                 ksocknal_data.ksnd_connd_starting++;
2472                 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2473
2474
2475                 snprintf(name, sizeof(name), "socknal_cd%02d", i);
2476                 rc = ksocknal_thread_start(ksocknal_connd,
2477                                            (void *)((ulong_ptr_t)i), name);
2478                 if (rc != 0) {
2479                         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2480                         ksocknal_data.ksnd_connd_starting--;
2481                         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2482                         CERROR("Can't spawn socknal connd: %d\n", rc);
2483                         goto failed;
2484                 }
2485         }
2486
2487         rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2488         if (rc != 0) {
2489                 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2490                 goto failed;
2491         }
2492
2493         /* flag everything initialised */
2494         ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2495
2496         return 0;
2497
2498  failed:
2499         ksocknal_base_shutdown();
2500         return -ENETDOWN;
2501 }
2502
2503 static void
2504 ksocknal_debug_peerhash (lnet_ni_t *ni)
2505 {
2506         ksock_peer_t    *peer = NULL;
2507         struct list_head        *tmp;
2508         int             i;
2509
2510         read_lock(&ksocknal_data.ksnd_global_lock);
2511
2512         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2513                 list_for_each(tmp, &ksocknal_data.ksnd_peers[i]) {
2514                         peer = list_entry(tmp, ksock_peer_t, ksnp_list);
2515
2516                         if (peer->ksnp_ni == ni) break;
2517
2518                         peer = NULL;
2519                 }
2520         }
2521
2522         if (peer != NULL) {
2523                 ksock_route_t *route;
2524                 ksock_conn_t  *conn;
2525
2526                 CWARN ("Active peer on shutdown: %s, ref %d, scnt %d, "
2527                        "closing %d, accepting %d, err %d, zcookie "LPU64", "
2528                        "txq %d, zc_req %d\n", libcfs_id2str(peer->ksnp_id),
2529                        atomic_read(&peer->ksnp_refcount),
2530                        peer->ksnp_sharecount, peer->ksnp_closing,
2531                        peer->ksnp_accepting, peer->ksnp_error,
2532                        peer->ksnp_zc_next_cookie,
2533                        !list_empty(&peer->ksnp_tx_queue),
2534                        !list_empty(&peer->ksnp_zc_req_list));
2535
2536                 list_for_each(tmp, &peer->ksnp_routes) {
2537                         route = list_entry(tmp, ksock_route_t, ksnr_list);
2538                         CWARN ("Route: ref %d, schd %d, conn %d, cnted %d, "
2539                                "del %d\n", atomic_read(&route->ksnr_refcount),
2540                                route->ksnr_scheduled, route->ksnr_connecting,
2541                                route->ksnr_connected, route->ksnr_deleted);
2542                 }
2543
2544                 list_for_each(tmp, &peer->ksnp_conns) {
2545                         conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2546                         CWARN ("Conn: ref %d, sref %d, t %d, c %d\n",
2547                                atomic_read(&conn->ksnc_conn_refcount),
2548                                atomic_read(&conn->ksnc_sock_refcount),
2549                                conn->ksnc_type, conn->ksnc_closing);
2550                 }
2551         }
2552
2553         read_unlock(&ksocknal_data.ksnd_global_lock);
2554         return;
2555 }
2556
2557 void
2558 ksocknal_shutdown (lnet_ni_t *ni)
2559 {
2560         ksock_net_t      *net = ni->ni_data;
2561         int               i;
2562         lnet_process_id_t anyid = {0};
2563
2564         anyid.nid =  LNET_NID_ANY;
2565         anyid.pid =  LNET_PID_ANY;
2566
2567         LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2568         LASSERT(ksocknal_data.ksnd_nnets > 0);
2569
2570         spin_lock_bh(&net->ksnn_lock);
2571         net->ksnn_shutdown = 1;                 /* prevent new peers */
2572         spin_unlock_bh(&net->ksnn_lock);
2573
2574         /* Delete all peers */
2575         ksocknal_del_peer(ni, anyid, 0);
2576
2577         /* Wait for all peer state to clean up */
2578         i = 2;
2579         spin_lock_bh(&net->ksnn_lock);
2580         while (net->ksnn_npeers != 0) {
2581                 spin_unlock_bh(&net->ksnn_lock);
2582
2583                 i++;
2584                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2585                        "waiting for %d peers to disconnect\n",
2586                        net->ksnn_npeers);
2587                 set_current_state(TASK_UNINTERRUPTIBLE);
2588                 schedule_timeout(cfs_time_seconds(1));
2589
2590                 ksocknal_debug_peerhash(ni);
2591
2592                 spin_lock_bh(&net->ksnn_lock);
2593         }
2594         spin_unlock_bh(&net->ksnn_lock);
2595
2596         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2597                 LASSERT (net->ksnn_interfaces[i].ksni_npeers == 0);
2598                 LASSERT (net->ksnn_interfaces[i].ksni_nroutes == 0);
2599         }
2600
2601         list_del(&net->ksnn_list);
2602         LIBCFS_FREE(net, sizeof(*net));
2603
2604         ksocknal_data.ksnd_nnets--;
2605         if (ksocknal_data.ksnd_nnets == 0)
2606                 ksocknal_base_shutdown();
2607 }
2608
2609 static int
2610 ksocknal_enumerate_interfaces(ksock_net_t *net)
2611 {
2612         char      **names;
2613         int         i;
2614         int         j;
2615         int         rc;
2616         int         n;
2617
2618         n = lnet_ipif_enumerate(&names);
2619         if (n <= 0) {
2620                 CERROR("Can't enumerate interfaces: %d\n", n);
2621                 return n;
2622         }
2623
2624         for (i = j = 0; i < n; i++) {
2625                 int        up;
2626                 __u32      ip;
2627                 __u32      mask;
2628
2629                 if (!strcmp(names[i], "lo")) /* skip the loopback IF */
2630                         continue;
2631
2632                 rc = lnet_ipif_query(names[i], &up, &ip, &mask);
2633                 if (rc != 0) {
2634                         CWARN("Can't get interface %s info: %d\n",
2635                               names[i], rc);
2636                         continue;
2637                 }
2638
2639                 if (!up) {
2640                         CWARN("Ignoring interface %s (down)\n",
2641                               names[i]);
2642                         continue;
2643                 }
2644
2645                 if (j == LNET_MAX_INTERFACES) {
2646                         CWARN("Ignoring interface %s (too many interfaces)\n",
2647                               names[i]);
2648                         continue;
2649                 }
2650
2651                 net->ksnn_interfaces[j].ksni_ipaddr = ip;
2652                 net->ksnn_interfaces[j].ksni_netmask = mask;
2653                 strlcpy(net->ksnn_interfaces[j].ksni_name,
2654                         names[i], sizeof(net->ksnn_interfaces[j].ksni_name));
2655                 j++;
2656         }
2657
2658         lnet_ipif_free_enumeration(names, n);
2659
2660         if (j == 0)
2661                 CERROR("Can't find any usable interfaces\n");
2662
2663         return j;
2664 }
2665
2666 static int
2667 ksocknal_search_new_ipif(ksock_net_t *net)
2668 {
2669         int     new_ipif = 0;
2670         int     i;
2671
2672         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2673                 char            *ifnam = &net->ksnn_interfaces[i].ksni_name[0];
2674                 char            *colon = strchr(ifnam, ':');
2675                 int             found  = 0;
2676                 ksock_net_t     *tmp;
2677                 int             j;
2678
2679                 if (colon != NULL) /* ignore alias device */
2680                         *colon = 0;
2681
2682                 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets,
2683                                         ksnn_list) {
2684                         for (j = 0; !found && j < tmp->ksnn_ninterfaces; j++) {
2685                                 char *ifnam2 = &tmp->ksnn_interfaces[j].\
2686                                              ksni_name[0];
2687                                 char *colon2 = strchr(ifnam2, ':');
2688
2689                                 if (colon2 != NULL)
2690                                         *colon2 = 0;
2691
2692                                 found = strcmp(ifnam, ifnam2) == 0;
2693                                 if (colon2 != NULL)
2694                                         *colon2 = ':';
2695                         }
2696                         if (found)
2697                                 break;
2698                 }
2699
2700                 new_ipif += !found;
2701                 if (colon != NULL)
2702                         *colon = ':';
2703         }
2704
2705         return new_ipif;
2706 }
2707
2708 static int
2709 ksocknal_start_schedulers(struct ksock_sched_info *info)
2710 {
2711         int     nthrs;
2712         int     rc = 0;
2713         int     i;
2714
2715         if (info->ksi_nthreads == 0) {
2716                 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2717                         nthrs = info->ksi_nthreads_max;
2718                 } else {
2719                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2720                                                info->ksi_cpt);
2721                         nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2722                         nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2723                 }
2724                 nthrs = min(nthrs, info->ksi_nthreads_max);
2725         } else {
2726                 LASSERT(info->ksi_nthreads <= info->ksi_nthreads_max);
2727                 /* increase two threads if there is new interface */
2728                 nthrs = min(2, info->ksi_nthreads_max - info->ksi_nthreads);
2729         }
2730
2731         for (i = 0; i < nthrs; i++) {
2732                 long            id;
2733                 char            name[20];
2734                 ksock_sched_t   *sched;
2735                 id = KSOCK_THREAD_ID(info->ksi_cpt, info->ksi_nthreads + i);
2736                 sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
2737                 snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
2738                          info->ksi_cpt, (int)(sched - &info->ksi_scheds[0]));
2739
2740                 rc = ksocknal_thread_start(ksocknal_scheduler,
2741                                            (void *)id, name);
2742                 if (rc == 0)
2743                         continue;
2744
2745                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2746                        info->ksi_cpt, info->ksi_nthreads + i, rc);
2747                 break;
2748         }
2749
2750         info->ksi_nthreads += i;
2751         return rc;
2752 }
2753
2754 static int
2755 ksocknal_net_start_threads(ksock_net_t *net, __u32 *cpts, int ncpts)
2756 {
2757         int     newif = ksocknal_search_new_ipif(net);
2758         int     rc;
2759         int     i;
2760
2761         LASSERT(ncpts > 0 && ncpts <= cfs_cpt_number(lnet_cpt_table()));
2762
2763         for (i = 0; i < ncpts; i++) {
2764                 struct ksock_sched_info *info;
2765                 int cpt = (cpts == NULL) ? i : cpts[i];
2766
2767                 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2768                 info = ksocknal_data.ksnd_sched_info[cpt];
2769
2770                 if (!newif && info->ksi_nthreads > 0)
2771                         continue;
2772
2773                 rc = ksocknal_start_schedulers(info);
2774                 if (rc != 0)
2775                         return rc;
2776         }
2777         return 0;
2778 }
2779
2780 int
2781 ksocknal_startup (lnet_ni_t *ni)
2782 {
2783         ksock_net_t  *net;
2784         int           rc;
2785         int           i;
2786
2787         LASSERT (ni->ni_lnd == &the_ksocklnd);
2788
2789         if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2790                 rc = ksocknal_base_startup();
2791                 if (rc != 0)
2792                         return rc;
2793         }
2794
2795         LIBCFS_ALLOC(net, sizeof(*net));
2796         if (net == NULL)
2797                 goto fail_0;
2798
2799         spin_lock_init(&net->ksnn_lock);
2800         net->ksnn_incarnation = ksocknal_new_incarnation();
2801         ni->ni_data = net;
2802         ni->ni_peertimeout    = *ksocknal_tunables.ksnd_peertimeout;
2803         ni->ni_maxtxcredits   = *ksocknal_tunables.ksnd_credits;
2804         ni->ni_peertxcredits  = *ksocknal_tunables.ksnd_peertxcredits;
2805         ni->ni_peerrtrcredits = *ksocknal_tunables.ksnd_peerrtrcredits;
2806
2807         if (ni->ni_interfaces[0] == NULL) {
2808                 rc = ksocknal_enumerate_interfaces(net);
2809                 if (rc <= 0)
2810                         goto fail_1;
2811
2812                 net->ksnn_ninterfaces = 1;
2813         } else {
2814                 for (i = 0; i < LNET_MAX_INTERFACES; i++) {
2815                         int    up;
2816
2817                         if (ni->ni_interfaces[i] == NULL)
2818                                 break;
2819
2820                         rc = lnet_ipif_query(ni->ni_interfaces[i], &up,
2821                                 &net->ksnn_interfaces[i].ksni_ipaddr,
2822                                 &net->ksnn_interfaces[i].ksni_netmask);
2823
2824                         if (rc != 0) {
2825                                 CERROR("Can't get interface %s info: %d\n",
2826                                        ni->ni_interfaces[i], rc);
2827                                 goto fail_1;
2828                         }
2829
2830                         if (!up) {
2831                                 CERROR("Interface %s is down\n",
2832                                        ni->ni_interfaces[i]);
2833                                 goto fail_1;
2834                         }
2835
2836                         strlcpy(net->ksnn_interfaces[i].ksni_name,
2837                                 ni->ni_interfaces[i],
2838                                 sizeof(net->ksnn_interfaces[i].ksni_name));
2839                 }
2840                 net->ksnn_ninterfaces = i;
2841         }
2842
2843         /* call it before add it to ksocknal_data.ksnd_nets */
2844         rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2845         if (rc != 0)
2846                 goto fail_1;
2847
2848         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid),
2849                                 net->ksnn_interfaces[0].ksni_ipaddr);
2850         list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2851
2852         ksocknal_data.ksnd_nnets++;
2853
2854         return 0;
2855
2856  fail_1:
2857         LIBCFS_FREE(net, sizeof(*net));
2858  fail_0:
2859         if (ksocknal_data.ksnd_nnets == 0)
2860                 ksocknal_base_shutdown();
2861
2862         return -ENETDOWN;
2863 }
2864
2865
2866 static void __exit ksocklnd_exit(void)
2867 {
2868         lnet_unregister_lnd(&the_ksocklnd);
2869         ksocknal_tunables_fini();
2870 }
2871
2872 static int __init ksocklnd_init(void)
2873 {
2874         int rc;
2875
2876         /* check ksnr_connected/connecting field large enough */
2877         CLASSERT(SOCKLND_CONN_NTYPES <= 4);
2878         CLASSERT(SOCKLND_CONN_ACK == SOCKLND_CONN_BULK_IN);
2879
2880         /* initialize the_ksocklnd */
2881         the_ksocklnd.lnd_type     = SOCKLND;
2882         the_ksocklnd.lnd_startup  = ksocknal_startup;
2883         the_ksocklnd.lnd_shutdown = ksocknal_shutdown;
2884         the_ksocklnd.lnd_ctl      = ksocknal_ctl;
2885         the_ksocklnd.lnd_send     = ksocknal_send;
2886         the_ksocklnd.lnd_recv     = ksocknal_recv;
2887         the_ksocklnd.lnd_notify   = ksocknal_notify;
2888         the_ksocklnd.lnd_query    = ksocknal_query;
2889         the_ksocklnd.lnd_accept   = ksocknal_accept;
2890
2891         rc = ksocknal_tunables_init();
2892         if (rc != 0)
2893                 return rc;
2894
2895         lnet_register_lnd(&the_ksocklnd);
2896
2897         return 0;
2898 }
2899
2900 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2901 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2902 MODULE_VERSION("2.8.0");
2903 MODULE_LICENSE("GPL");
2904
2905 module_init(ksocklnd_init);
2906 module_exit(ksocklnd_exit);