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                 ksocknal_route_decref(route);   /* drop conn's ref on route */
1443         }
1444
1445         if (list_empty(&peer->ksnp_conns)) {
1446                 /* No more connections to this peer */
1447
1448                 if (!list_empty(&peer->ksnp_tx_queue)) {
1449                                 ksock_tx_t *tx;
1450
1451                         LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1452
1453                         /* throw them to the last connection...,
1454                          * these TXs will be send to /dev/null by scheduler */
1455                         list_for_each_entry(tx, &peer->ksnp_tx_queue,
1456                                             tx_list)
1457                                 ksocknal_tx_prep(conn, tx);
1458
1459                         spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1460                         list_splice_init(&peer->ksnp_tx_queue,
1461                                          &conn->ksnc_tx_queue);
1462                         spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1463                 }
1464
1465                 /* renegotiate protocol version */
1466                 peer->ksnp_proto = NULL;
1467                 /* stash last conn close reason */
1468                 peer->ksnp_error = error;
1469
1470                 if (list_empty(&peer->ksnp_routes)) {
1471                         /* I've just closed last conn belonging to a
1472                          * peer with no routes to it */
1473                         ksocknal_unlink_peer_locked(peer);
1474                 }
1475         }
1476
1477         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1478
1479         list_add_tail(&conn->ksnc_list,
1480                       &ksocknal_data.ksnd_deathrow_conns);
1481         wake_up(&ksocknal_data.ksnd_reaper_waitq);
1482
1483         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1484 }
1485
1486 void
1487 ksocknal_peer_failed (ksock_peer_t *peer)
1488 {
1489         int        notify = 0;
1490         cfs_time_t last_alive = 0;
1491
1492         /* There has been a connection failure or comms error; but I'll only
1493          * tell LNET I think the peer is dead if it's to another kernel and
1494          * there are no connections or connection attempts in existence. */
1495
1496         read_lock(&ksocknal_data.ksnd_global_lock);
1497
1498         if ((peer->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1499              list_empty(&peer->ksnp_conns) &&
1500              peer->ksnp_accepting == 0 &&
1501              ksocknal_find_connecting_route_locked(peer) == NULL) {
1502                 notify = 1;
1503                 last_alive = peer->ksnp_last_alive;
1504         }
1505
1506         read_unlock(&ksocknal_data.ksnd_global_lock);
1507
1508         if (notify)
1509                 lnet_notify(peer->ksnp_ni, peer->ksnp_id.nid, 0,
1510                             last_alive);
1511 }
1512
1513 void
1514 ksocknal_finalize_zcreq(ksock_conn_t *conn)
1515 {
1516         ksock_peer_t     *peer = conn->ksnc_peer;
1517         ksock_tx_t       *tx;
1518         ksock_tx_t       *tmp;
1519         struct list_head  zlist = LIST_HEAD_INIT(zlist);
1520
1521         /* NB safe to finalize TXs because closing of socket will
1522          * abort all buffered data */
1523         LASSERT(conn->ksnc_sock == NULL);
1524
1525         spin_lock(&peer->ksnp_lock);
1526
1527         list_for_each_entry_safe(tx, tmp, &peer->ksnp_zc_req_list, tx_zc_list) {
1528                 if (tx->tx_conn != conn)
1529                         continue;
1530
1531                 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1532
1533                 tx->tx_msg.ksm_zc_cookies[0] = 0;
1534                 tx->tx_zc_aborted = 1;  /* mark it as not-acked */
1535                 list_del(&tx->tx_zc_list);
1536                 list_add(&tx->tx_zc_list, &zlist);
1537         }
1538
1539         spin_unlock(&peer->ksnp_lock);
1540
1541         while (!list_empty(&zlist)) {
1542                 tx = list_entry(zlist.next, ksock_tx_t, tx_zc_list);
1543
1544                 list_del(&tx->tx_zc_list);
1545                 ksocknal_tx_decref(tx);
1546         }
1547 }
1548
1549 void
1550 ksocknal_terminate_conn(ksock_conn_t *conn)
1551 {
1552         /* This gets called by the reaper (guaranteed thread context) to
1553          * disengage the socket from its callbacks and close it.
1554          * ksnc_refcount will eventually hit zero, and then the reaper will
1555          * destroy it. */
1556         ksock_peer_t     *peer = conn->ksnc_peer;
1557         ksock_sched_t    *sched = conn->ksnc_scheduler;
1558         int               failed = 0;
1559
1560         LASSERT(conn->ksnc_closing);
1561
1562         /* wake up the scheduler to "send" all remaining packets to /dev/null */
1563         spin_lock_bh(&sched->kss_lock);
1564
1565         /* a closing conn is always ready to tx */
1566         conn->ksnc_tx_ready = 1;
1567
1568         if (!conn->ksnc_tx_scheduled &&
1569             !list_empty(&conn->ksnc_tx_queue)) {
1570                 list_add_tail(&conn->ksnc_tx_list,
1571                                &sched->kss_tx_conns);
1572                 conn->ksnc_tx_scheduled = 1;
1573                 /* extra ref for scheduler */
1574                 ksocknal_conn_addref(conn);
1575
1576                 wake_up (&sched->kss_waitq);
1577         }
1578
1579         spin_unlock_bh(&sched->kss_lock);
1580
1581         /* serialise with callbacks */
1582         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1583
1584         ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1585
1586         /* OK, so this conn may not be completely disengaged from its
1587          * scheduler yet, but it _has_ committed to terminate... */
1588         conn->ksnc_scheduler->kss_nconns--;
1589
1590         if (peer->ksnp_error != 0) {
1591                 /* peer's last conn closed in error */
1592                 LASSERT(list_empty(&peer->ksnp_conns));
1593                 failed = 1;
1594                 peer->ksnp_error = 0;     /* avoid multiple notifications */
1595         }
1596
1597         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1598
1599         if (failed)
1600                 ksocknal_peer_failed(peer);
1601
1602         /* The socket is closed on the final put; either here, or in
1603          * ksocknal_{send,recv}msg().  Since we set up the linger2 option
1604          * when the connection was established, this will close the socket
1605          * immediately, aborting anything buffered in it. Any hung
1606          * zero-copy transmits will therefore complete in finite time. */
1607         ksocknal_connsock_decref(conn);
1608 }
1609
1610 void
1611 ksocknal_queue_zombie_conn (ksock_conn_t *conn)
1612 {
1613         /* Queue the conn for the reaper to destroy */
1614
1615         LASSERT(atomic_read(&conn->ksnc_conn_refcount) == 0);
1616         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1617
1618         list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1619         wake_up(&ksocknal_data.ksnd_reaper_waitq);
1620
1621         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1622 }
1623
1624 void
1625 ksocknal_destroy_conn (ksock_conn_t *conn)
1626 {
1627         cfs_time_t      last_rcv;
1628
1629         /* Final coup-de-grace of the reaper */
1630         CDEBUG (D_NET, "connection %p\n", conn);
1631
1632         LASSERT (atomic_read (&conn->ksnc_conn_refcount) == 0);
1633         LASSERT (atomic_read (&conn->ksnc_sock_refcount) == 0);
1634         LASSERT (conn->ksnc_sock == NULL);
1635         LASSERT (conn->ksnc_route == NULL);
1636         LASSERT (!conn->ksnc_tx_scheduled);
1637         LASSERT (!conn->ksnc_rx_scheduled);
1638         LASSERT(list_empty(&conn->ksnc_tx_queue));
1639
1640         /* complete current receive if any */
1641         switch (conn->ksnc_rx_state) {
1642         case SOCKNAL_RX_LNET_PAYLOAD:
1643                 last_rcv = conn->ksnc_rx_deadline -
1644                            cfs_time_seconds(*ksocknal_tunables.ksnd_timeout);
1645                 CERROR("Completing partial receive from %s[%d], "
1646                        "ip %pI4h:%d, with error, wanted: %d, left: %d, "
1647                        "last alive is %ld secs ago\n",
1648                        libcfs_id2str(conn->ksnc_peer->ksnp_id), conn->ksnc_type,
1649                        &conn->ksnc_ipaddr, conn->ksnc_port,
1650                        conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1651                        cfs_duration_sec(cfs_time_sub(cfs_time_current(),
1652                                         last_rcv)));
1653                 lnet_finalize (conn->ksnc_peer->ksnp_ni,
1654                                conn->ksnc_cookie, -EIO);
1655                 break;
1656         case SOCKNAL_RX_LNET_HEADER:
1657                 if (conn->ksnc_rx_started)
1658                         CERROR("Incomplete receive of lnet header from %s, "
1659                                "ip %pI4h:%d, with error, protocol: %d.x.\n",
1660                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1661                                &conn->ksnc_ipaddr, conn->ksnc_port,
1662                                conn->ksnc_proto->pro_version);
1663                 break;
1664         case SOCKNAL_RX_KSM_HEADER:
1665                 if (conn->ksnc_rx_started)
1666                         CERROR("Incomplete receive of ksock message from %s, "
1667                                "ip %pI4h:%d, with error, protocol: %d.x.\n",
1668                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1669                                &conn->ksnc_ipaddr, conn->ksnc_port,
1670                                conn->ksnc_proto->pro_version);
1671                 break;
1672         case SOCKNAL_RX_SLOP:
1673                 if (conn->ksnc_rx_started)
1674                         CERROR("Incomplete receive of slops from %s, "
1675                                "ip %pI4h:%d, with error\n",
1676                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1677                                &conn->ksnc_ipaddr, conn->ksnc_port);
1678                break;
1679         default:
1680                 LBUG ();
1681                 break;
1682         }
1683
1684         ksocknal_peer_decref(conn->ksnc_peer);
1685
1686         LIBCFS_FREE (conn, sizeof (*conn));
1687 }
1688
1689 int
1690 ksocknal_close_peer_conns_locked (ksock_peer_t *peer, __u32 ipaddr, int why)
1691 {
1692         ksock_conn_t       *conn;
1693         struct list_head         *ctmp;
1694         struct list_head         *cnxt;
1695         int                 count = 0;
1696
1697         list_for_each_safe(ctmp, cnxt, &peer->ksnp_conns) {
1698                 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
1699
1700                 if (ipaddr == 0 ||
1701                     conn->ksnc_ipaddr == ipaddr) {
1702                         count++;
1703                         ksocknal_close_conn_locked (conn, why);
1704                 }
1705         }
1706
1707         return (count);
1708 }
1709
1710 int
1711 ksocknal_close_conn_and_siblings (ksock_conn_t *conn, int why)
1712 {
1713         ksock_peer_t     *peer = conn->ksnc_peer;
1714         __u32             ipaddr = conn->ksnc_ipaddr;
1715         int               count;
1716
1717         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1718
1719         count = ksocknal_close_peer_conns_locked (peer, ipaddr, why);
1720
1721         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1722
1723         return (count);
1724 }
1725
1726 int
1727 ksocknal_close_matching_conns (lnet_process_id_t id, __u32 ipaddr)
1728 {
1729         ksock_peer_t       *peer;
1730         struct list_head         *ptmp;
1731         struct list_head         *pnxt;
1732         int                 lo;
1733         int                 hi;
1734         int                 i;
1735         int                 count = 0;
1736
1737         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1738
1739         if (id.nid != LNET_NID_ANY)
1740                 lo = hi = (int)(ksocknal_nid2peerlist(id.nid) - ksocknal_data.ksnd_peers);
1741         else {
1742                 lo = 0;
1743                 hi = ksocknal_data.ksnd_peer_hash_size - 1;
1744         }
1745
1746         for (i = lo; i <= hi; i++) {
1747                 list_for_each_safe(ptmp, pnxt, &ksocknal_data.ksnd_peers[i]) {
1748
1749                         peer = list_entry(ptmp, ksock_peer_t, ksnp_list);
1750
1751                         if (!((id.nid == LNET_NID_ANY || id.nid == peer->ksnp_id.nid) &&
1752                               (id.pid == LNET_PID_ANY || id.pid == peer->ksnp_id.pid)))
1753                                 continue;
1754
1755                         count += ksocknal_close_peer_conns_locked (peer, ipaddr, 0);
1756                 }
1757         }
1758
1759         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1760
1761         /* wildcards always succeed */
1762         if (id.nid == LNET_NID_ANY || id.pid == LNET_PID_ANY || ipaddr == 0)
1763                 return (0);
1764
1765         return (count == 0 ? -ENOENT : 0);
1766 }
1767
1768 void
1769 ksocknal_notify (lnet_ni_t *ni, lnet_nid_t gw_nid, int alive)
1770 {
1771         /* The router is telling me she's been notified of a change in
1772          * gateway state.... */
1773         lnet_process_id_t  id = {0};
1774
1775         id.nid = gw_nid;
1776         id.pid = LNET_PID_ANY;
1777
1778         CDEBUG (D_NET, "gw %s %s\n", libcfs_nid2str(gw_nid),
1779                 alive ? "up" : "down");
1780
1781         if (!alive) {
1782                 /* If the gateway crashed, close all open connections... */
1783                 ksocknal_close_matching_conns (id, 0);
1784                 return;
1785         }
1786
1787         /* ...otherwise do nothing.  We can only establish new connections
1788          * if we have autroutes, and these connect on demand. */
1789 }
1790
1791 void
1792 ksocknal_query (lnet_ni_t *ni, lnet_nid_t nid, cfs_time_t *when)
1793 {
1794         int                connect = 1;
1795         cfs_time_t         last_alive = 0;
1796         cfs_time_t         now = cfs_time_current();
1797         ksock_peer_t      *peer = NULL;
1798         rwlock_t                *glock = &ksocknal_data.ksnd_global_lock;
1799         lnet_process_id_t  id = {
1800                 .nid = nid,
1801                 .pid = LNET_PID_LUSTRE,
1802         };
1803
1804         read_lock(glock);
1805
1806         peer = ksocknal_find_peer_locked(ni, id);
1807         if (peer != NULL) {
1808                 struct list_head       *tmp;
1809                 ksock_conn_t     *conn;
1810                 int               bufnob;
1811
1812                 list_for_each(tmp, &peer->ksnp_conns) {
1813                         conn = list_entry(tmp, ksock_conn_t, ksnc_list);
1814                         bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
1815
1816                         if (bufnob < conn->ksnc_tx_bufnob) {
1817                                 /* something got ACKed */
1818                                 conn->ksnc_tx_deadline =
1819                                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1820                                 peer->ksnp_last_alive = now;
1821                                 conn->ksnc_tx_bufnob = bufnob;
1822                         }
1823                 }
1824
1825                 last_alive = peer->ksnp_last_alive;
1826                 if (ksocknal_find_connectable_route_locked(peer) == NULL)
1827                         connect = 0;
1828         }
1829
1830         read_unlock(glock);
1831
1832         if (last_alive != 0)
1833                 *when = last_alive;
1834
1835         CDEBUG(D_NET, "Peer %s %p, alive %ld secs ago, connect %d\n",
1836                libcfs_nid2str(nid), peer,
1837                last_alive ? cfs_duration_sec(now - last_alive) : -1,
1838                connect);
1839
1840         if (!connect)
1841                 return;
1842
1843         ksocknal_add_peer(ni, id, LNET_NIDADDR(nid), lnet_acceptor_port());
1844
1845         write_lock_bh(glock);
1846
1847         peer = ksocknal_find_peer_locked(ni, id);
1848         if (peer != NULL)
1849                 ksocknal_launch_all_connections_locked(peer);
1850
1851         write_unlock_bh(glock);
1852         return;
1853 }
1854
1855 static void
1856 ksocknal_push_peer (ksock_peer_t *peer)
1857 {
1858         int               index;
1859         int               i;
1860         struct list_head       *tmp;
1861         ksock_conn_t     *conn;
1862
1863         for (index = 0; ; index++) {
1864                 read_lock(&ksocknal_data.ksnd_global_lock);
1865
1866                 i = 0;
1867                 conn = NULL;
1868
1869                 list_for_each(tmp, &peer->ksnp_conns) {
1870                         if (i++ == index) {
1871                                 conn = list_entry(tmp, ksock_conn_t,
1872                                                        ksnc_list);
1873                                 ksocknal_conn_addref(conn);
1874                                 break;
1875                         }
1876                 }
1877
1878                 read_unlock(&ksocknal_data.ksnd_global_lock);
1879
1880                 if (conn == NULL)
1881                         break;
1882
1883                 ksocknal_lib_push_conn (conn);
1884                 ksocknal_conn_decref(conn);
1885         }
1886 }
1887
1888 static int
1889 ksocknal_push (lnet_ni_t *ni, lnet_process_id_t id)
1890 {
1891         struct list_head *start;
1892         struct list_head *end;
1893         struct list_head *tmp;
1894         int               rc = -ENOENT;
1895         unsigned int      hsize = ksocknal_data.ksnd_peer_hash_size;
1896
1897         if (id.nid == LNET_NID_ANY) {
1898                 start = &ksocknal_data.ksnd_peers[0];
1899                 end = &ksocknal_data.ksnd_peers[hsize - 1];
1900         } else {
1901                 start = end = ksocknal_nid2peerlist(id.nid);
1902         }
1903
1904         for (tmp = start; tmp <= end; tmp++) {
1905                 int     peer_off; /* searching offset in peer hash table */
1906
1907                 for (peer_off = 0; ; peer_off++) {
1908                         ksock_peer_t *peer;
1909                         int           i = 0;
1910
1911                         read_lock(&ksocknal_data.ksnd_global_lock);
1912                         list_for_each_entry(peer, tmp, ksnp_list) {
1913                                 if (!((id.nid == LNET_NID_ANY ||
1914                                        id.nid == peer->ksnp_id.nid) &&
1915                                       (id.pid == LNET_PID_ANY ||
1916                                        id.pid == peer->ksnp_id.pid)))
1917                                         continue;
1918
1919                                 if (i++ == peer_off) {
1920                                         ksocknal_peer_addref(peer);
1921                                         break;
1922                                 }
1923                         }
1924                         read_unlock(&ksocknal_data.ksnd_global_lock);
1925
1926                         if (i == 0) /* no match */
1927                                 break;
1928
1929                         rc = 0;
1930                         ksocknal_push_peer(peer);
1931                         ksocknal_peer_decref(peer);
1932                 }
1933         }
1934         return rc;
1935 }
1936
1937 static int
1938 ksocknal_add_interface(lnet_ni_t *ni, __u32 ipaddress, __u32 netmask)
1939 {
1940         ksock_net_t       *net = ni->ni_data;
1941         ksock_interface_t *iface;
1942         int                rc;
1943         int                i;
1944         int                j;
1945         struct list_head        *ptmp;
1946         ksock_peer_t      *peer;
1947         struct list_head        *rtmp;
1948         ksock_route_t     *route;
1949
1950         if (ipaddress == 0 ||
1951             netmask == 0)
1952                 return (-EINVAL);
1953
1954         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1955
1956         iface = ksocknal_ip2iface(ni, ipaddress);
1957         if (iface != NULL) {
1958                 /* silently ignore dups */
1959                 rc = 0;
1960         } else if (net->ksnn_ninterfaces == LNET_MAX_INTERFACES) {
1961                 rc = -ENOSPC;
1962         } else {
1963                 iface = &net->ksnn_interfaces[net->ksnn_ninterfaces++];
1964
1965                 iface->ksni_ipaddr = ipaddress;
1966                 iface->ksni_netmask = netmask;
1967                 iface->ksni_nroutes = 0;
1968                 iface->ksni_npeers = 0;
1969
1970                 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1971                         list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
1972                                 peer = list_entry(ptmp, ksock_peer_t,
1973                                                       ksnp_list);
1974
1975                                 for (j = 0; j < peer->ksnp_n_passive_ips; j++)
1976                                         if (peer->ksnp_passive_ips[j] == ipaddress)
1977                                                 iface->ksni_npeers++;
1978
1979                                 list_for_each(rtmp, &peer->ksnp_routes) {
1980                                         route = list_entry(rtmp,
1981                                                                ksock_route_t,
1982                                                                ksnr_list);
1983
1984                                         if (route->ksnr_myipaddr == ipaddress)
1985                                                 iface->ksni_nroutes++;
1986                                 }
1987                         }
1988                 }
1989
1990                 rc = 0;
1991                 /* NB only new connections will pay attention to the new interface! */
1992         }
1993
1994         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1995
1996         return (rc);
1997 }
1998
1999 static void
2000 ksocknal_peer_del_interface_locked(ksock_peer_t *peer, __u32 ipaddr)
2001 {
2002         struct list_head         *tmp;
2003         struct list_head         *nxt;
2004         ksock_route_t      *route;
2005         ksock_conn_t       *conn;
2006         int                 i;
2007         int                 j;
2008
2009         for (i = 0; i < peer->ksnp_n_passive_ips; i++)
2010                 if (peer->ksnp_passive_ips[i] == ipaddr) {
2011                         for (j = i+1; j < peer->ksnp_n_passive_ips; j++)
2012                                 peer->ksnp_passive_ips[j-1] =
2013                                         peer->ksnp_passive_ips[j];
2014                         peer->ksnp_n_passive_ips--;
2015                         break;
2016                 }
2017
2018         list_for_each_safe(tmp, nxt, &peer->ksnp_routes) {
2019                 route = list_entry(tmp, ksock_route_t, ksnr_list);
2020
2021                 if (route->ksnr_myipaddr != ipaddr)
2022                         continue;
2023
2024                 if (route->ksnr_share_count != 0) {
2025                         /* Manually created; keep, but unbind */
2026                         route->ksnr_myipaddr = 0;
2027                 } else {
2028                         ksocknal_del_route_locked(route);
2029                 }
2030         }
2031
2032         list_for_each_safe(tmp, nxt, &peer->ksnp_conns) {
2033                 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2034
2035                 if (conn->ksnc_myipaddr == ipaddr)
2036                         ksocknal_close_conn_locked (conn, 0);
2037         }
2038 }
2039
2040 static int
2041 ksocknal_del_interface(lnet_ni_t *ni, __u32 ipaddress)
2042 {
2043         ksock_net_t       *net = ni->ni_data;
2044         int                rc = -ENOENT;
2045         struct list_head        *tmp;
2046         struct list_head        *nxt;
2047         ksock_peer_t      *peer;
2048         __u32              this_ip;
2049         int                i;
2050         int                j;
2051
2052         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2053
2054         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2055                 this_ip = net->ksnn_interfaces[i].ksni_ipaddr;
2056
2057                 if (!(ipaddress == 0 ||
2058                       ipaddress == this_ip))
2059                         continue;
2060
2061                 rc = 0;
2062
2063                 for (j = i+1; j < net->ksnn_ninterfaces; j++)
2064                         net->ksnn_interfaces[j-1] =
2065                                 net->ksnn_interfaces[j];
2066
2067                 net->ksnn_ninterfaces--;
2068
2069                 for (j = 0; j < ksocknal_data.ksnd_peer_hash_size; j++) {
2070                         list_for_each_safe(tmp, nxt,
2071                                                &ksocknal_data.ksnd_peers[j]) {
2072                                 peer = list_entry(tmp, ksock_peer_t,
2073                                                       ksnp_list);
2074
2075                                 if (peer->ksnp_ni != ni)
2076                                         continue;
2077
2078                                 ksocknal_peer_del_interface_locked(peer, this_ip);
2079                         }
2080                 }
2081         }
2082
2083         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2084
2085         return (rc);
2086 }
2087
2088 int
2089 ksocknal_ctl(lnet_ni_t *ni, unsigned int cmd, void *arg)
2090 {
2091         lnet_process_id_t id = {0};
2092         struct libcfs_ioctl_data *data = arg;
2093         int rc;
2094
2095         switch(cmd) {
2096         case IOC_LIBCFS_GET_INTERFACE: {
2097                 ksock_net_t       *net = ni->ni_data;
2098                 ksock_interface_t *iface;
2099
2100                 read_lock(&ksocknal_data.ksnd_global_lock);
2101
2102                 if (data->ioc_count >= (__u32)net->ksnn_ninterfaces) {
2103                         rc = -ENOENT;
2104                 } else {
2105                         rc = 0;
2106                         iface = &net->ksnn_interfaces[data->ioc_count];
2107
2108                         data->ioc_u32[0] = iface->ksni_ipaddr;
2109                         data->ioc_u32[1] = iface->ksni_netmask;
2110                         data->ioc_u32[2] = iface->ksni_npeers;
2111                         data->ioc_u32[3] = iface->ksni_nroutes;
2112                 }
2113
2114                 read_unlock(&ksocknal_data.ksnd_global_lock);
2115                 return rc;
2116         }
2117
2118         case IOC_LIBCFS_ADD_INTERFACE:
2119                 return ksocknal_add_interface(ni,
2120                                               data->ioc_u32[0], /* IP address */
2121                                               data->ioc_u32[1]); /* net mask */
2122
2123         case IOC_LIBCFS_DEL_INTERFACE:
2124                 return ksocknal_del_interface(ni,
2125                                               data->ioc_u32[0]); /* IP address */
2126
2127         case IOC_LIBCFS_GET_PEER: {
2128                 __u32            myip = 0;
2129                 __u32            ip = 0;
2130                 int              port = 0;
2131                 int              conn_count = 0;
2132                 int              share_count = 0;
2133
2134                 rc = ksocknal_get_peer_info(ni, data->ioc_count,
2135                                             &id, &myip, &ip, &port,
2136                                             &conn_count,  &share_count);
2137                 if (rc != 0)
2138                         return rc;
2139
2140                 data->ioc_nid    = id.nid;
2141                 data->ioc_count  = share_count;
2142                 data->ioc_u32[0] = ip;
2143                 data->ioc_u32[1] = port;
2144                 data->ioc_u32[2] = myip;
2145                 data->ioc_u32[3] = conn_count;
2146                 data->ioc_u32[4] = id.pid;
2147                 return 0;
2148         }
2149
2150         case IOC_LIBCFS_ADD_PEER:
2151                 id.nid = data->ioc_nid;
2152                 id.pid = LNET_PID_LUSTRE;
2153                 return ksocknal_add_peer (ni, id,
2154                                           data->ioc_u32[0], /* IP */
2155                                           data->ioc_u32[1]); /* port */
2156
2157         case IOC_LIBCFS_DEL_PEER:
2158                 id.nid = data->ioc_nid;
2159                 id.pid = LNET_PID_ANY;
2160                 return ksocknal_del_peer (ni, id,
2161                                           data->ioc_u32[0]); /* IP */
2162
2163         case IOC_LIBCFS_GET_CONN: {
2164                 int           txmem;
2165                 int           rxmem;
2166                 int           nagle;
2167                 ksock_conn_t *conn = ksocknal_get_conn_by_idx (ni, data->ioc_count);
2168
2169                 if (conn == NULL)
2170                         return -ENOENT;
2171
2172                 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
2173
2174                 data->ioc_count  = txmem;
2175                 data->ioc_nid    = conn->ksnc_peer->ksnp_id.nid;
2176                 data->ioc_flags  = nagle;
2177                 data->ioc_u32[0] = conn->ksnc_ipaddr;
2178                 data->ioc_u32[1] = conn->ksnc_port;
2179                 data->ioc_u32[2] = conn->ksnc_myipaddr;
2180                 data->ioc_u32[3] = conn->ksnc_type;
2181                 data->ioc_u32[4] = conn->ksnc_scheduler->kss_info->ksi_cpt;
2182                 data->ioc_u32[5] = rxmem;
2183                 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
2184                 ksocknal_conn_decref(conn);
2185                 return 0;
2186         }
2187
2188         case IOC_LIBCFS_CLOSE_CONNECTION:
2189                 id.nid = data->ioc_nid;
2190                 id.pid = LNET_PID_ANY;
2191                 return ksocknal_close_matching_conns (id,
2192                                                       data->ioc_u32[0]);
2193
2194         case IOC_LIBCFS_REGISTER_MYNID:
2195                 /* Ignore if this is a noop */
2196                 if (data->ioc_nid == ni->ni_nid)
2197                         return 0;
2198
2199                 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
2200                        libcfs_nid2str(data->ioc_nid),
2201                        libcfs_nid2str(ni->ni_nid));
2202                 return -EINVAL;
2203
2204         case IOC_LIBCFS_PUSH_CONNECTION:
2205                 id.nid = data->ioc_nid;
2206                 id.pid = LNET_PID_ANY;
2207                 return ksocknal_push(ni, id);
2208
2209         default:
2210                 return -EINVAL;
2211         }
2212         /* not reached */
2213 }
2214
2215 static void
2216 ksocknal_free_buffers (void)
2217 {
2218         LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
2219
2220         if (ksocknal_data.ksnd_sched_info != NULL) {
2221                 struct ksock_sched_info *info;
2222                 int                     i;
2223
2224                 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2225                         if (info->ksi_scheds != NULL) {
2226                                 LIBCFS_FREE(info->ksi_scheds,
2227                                             info->ksi_nthreads_max *
2228                                             sizeof(info->ksi_scheds[0]));
2229                         }
2230                 }
2231                 cfs_percpt_free(ksocknal_data.ksnd_sched_info);
2232         }
2233
2234         LIBCFS_FREE (ksocknal_data.ksnd_peers,
2235                      sizeof(struct list_head) *
2236                      ksocknal_data.ksnd_peer_hash_size);
2237
2238         spin_lock(&ksocknal_data.ksnd_tx_lock);
2239
2240         if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
2241                 struct list_head        zlist;
2242                 ksock_tx_t      *tx;
2243
2244                 list_add(&zlist, &ksocknal_data.ksnd_idle_noop_txs);
2245                 list_del_init(&ksocknal_data.ksnd_idle_noop_txs);
2246                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2247
2248                 while (!list_empty(&zlist)) {
2249                         tx = list_entry(zlist.next, ksock_tx_t, tx_list);
2250                         list_del(&tx->tx_list);
2251                         LIBCFS_FREE(tx, tx->tx_desc_size);
2252                 }
2253         } else {
2254                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2255         }
2256 }
2257
2258 static void
2259 ksocknal_base_shutdown(void)
2260 {
2261         struct ksock_sched_info *info;
2262         ksock_sched_t           *sched;
2263         int                     i;
2264         int                     j;
2265
2266         CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2267                atomic_read (&libcfs_kmemory));
2268         LASSERT (ksocknal_data.ksnd_nnets == 0);
2269
2270         switch (ksocknal_data.ksnd_init) {
2271         default:
2272                 LASSERT (0);
2273
2274         case SOCKNAL_INIT_ALL:
2275         case SOCKNAL_INIT_DATA:
2276                 LASSERT (ksocknal_data.ksnd_peers != NULL);
2277                 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2278                         LASSERT(list_empty(&ksocknal_data.ksnd_peers[i]));
2279                 }
2280
2281                 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2282                 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2283                 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2284                 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2285                 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2286
2287                 if (ksocknal_data.ksnd_sched_info != NULL) {
2288                         cfs_percpt_for_each(info, i,
2289                                             ksocknal_data.ksnd_sched_info) {
2290                                 if (info->ksi_scheds == NULL)
2291                                         continue;
2292
2293                                 for (j = 0; j < info->ksi_nthreads_max; j++) {
2294
2295                                         sched = &info->ksi_scheds[j];
2296                                         LASSERT(list_empty(&sched->\
2297                                                                kss_tx_conns));
2298                                         LASSERT(list_empty(&sched->\
2299                                                                kss_rx_conns));
2300                                         LASSERT(list_empty(&sched-> \
2301                                                   kss_zombie_noop_txs));
2302                                         LASSERT(sched->kss_nconns == 0);
2303                                 }
2304                         }
2305                 }
2306
2307                 /* flag threads to terminate; wake and wait for them to die */
2308                 ksocknal_data.ksnd_shuttingdown = 1;
2309                 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2310                 wake_up_all(&ksocknal_data.ksnd_reaper_waitq);
2311
2312                 if (ksocknal_data.ksnd_sched_info != NULL) {
2313                         cfs_percpt_for_each(info, i,
2314                                             ksocknal_data.ksnd_sched_info) {
2315                                 if (info->ksi_scheds == NULL)
2316                                         continue;
2317
2318                                 for (j = 0; j < info->ksi_nthreads_max; j++) {
2319                                         sched = &info->ksi_scheds[j];
2320                                         wake_up_all(&sched->kss_waitq);
2321                                 }
2322                         }
2323                 }
2324
2325                 i = 4;
2326                 read_lock(&ksocknal_data.ksnd_global_lock);
2327                 while (ksocknal_data.ksnd_nthreads != 0) {
2328                         i++;
2329                         /* power of 2? */
2330                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2331                                 "waiting for %d threads to terminate\n",
2332                                 ksocknal_data.ksnd_nthreads);
2333                         read_unlock(&ksocknal_data.ksnd_global_lock);
2334                         set_current_state(TASK_UNINTERRUPTIBLE);
2335                         schedule_timeout(cfs_time_seconds(1));
2336                         read_lock(&ksocknal_data.ksnd_global_lock);
2337                 }
2338                 read_unlock(&ksocknal_data.ksnd_global_lock);
2339
2340                 ksocknal_free_buffers();
2341
2342                 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2343                 break;
2344         }
2345
2346         CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2347                atomic_read (&libcfs_kmemory));
2348
2349         module_put(THIS_MODULE);
2350 }
2351
2352 static __u64 ksocknal_new_incarnation(void)
2353 {
2354         struct timeval tv;
2355
2356         /* The incarnation number is the time this module loaded and it
2357          * identifies this particular instance of the socknal.  Hopefully
2358          * we won't be able to reboot more frequently than 1MHz for the
2359          * forseeable future :) */
2360
2361         do_gettimeofday(&tv);
2362
2363         return (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
2364 }
2365
2366 static int
2367 ksocknal_base_startup(void)
2368 {
2369         struct ksock_sched_info *info;
2370         int                     rc;
2371         int                     i;
2372
2373         LASSERT (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2374         LASSERT (ksocknal_data.ksnd_nnets == 0);
2375
2376         memset (&ksocknal_data, 0, sizeof (ksocknal_data)); /* zero pointers */
2377
2378         ksocknal_data.ksnd_peer_hash_size = SOCKNAL_PEER_HASH_SIZE;
2379         LIBCFS_ALLOC(ksocknal_data.ksnd_peers,
2380                      sizeof(struct list_head) *
2381                      ksocknal_data.ksnd_peer_hash_size);
2382         if (ksocknal_data.ksnd_peers == NULL)
2383                 return -ENOMEM;
2384
2385         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++)
2386                 INIT_LIST_HEAD(&ksocknal_data.ksnd_peers[i]);
2387
2388         rwlock_init(&ksocknal_data.ksnd_global_lock);
2389         INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2390
2391         spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2392         INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2393         INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2394         INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2395         init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2396
2397         spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2398         INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2399         INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2400         init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2401
2402         spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2403         INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2404
2405         /* NB memset above zeros whole of ksocknal_data */
2406
2407         /* flag lists/ptrs/locks initialised */
2408         ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2409         try_module_get(THIS_MODULE);
2410
2411         ksocknal_data.ksnd_sched_info = cfs_percpt_alloc(lnet_cpt_table(),
2412                                                          sizeof(*info));
2413         if (ksocknal_data.ksnd_sched_info == NULL)
2414                 goto failed;
2415
2416         cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2417                 ksock_sched_t   *sched;
2418                 int             nthrs;
2419
2420                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2421                 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2422                         nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2423                 } else {
2424                         /* max to half of CPUs, assume another half should be
2425                          * reserved for upper layer modules */
2426                         nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2427                 }
2428
2429                 info->ksi_nthreads_max = nthrs;
2430                 info->ksi_cpt = i;
2431
2432                 LIBCFS_CPT_ALLOC(info->ksi_scheds, lnet_cpt_table(), i,
2433                                  info->ksi_nthreads_max * sizeof(*sched));
2434                 if (info->ksi_scheds == NULL)
2435                         goto failed;
2436
2437                 for (; nthrs > 0; nthrs--) {
2438                         sched = &info->ksi_scheds[nthrs - 1];
2439
2440                         sched->kss_info = info;
2441                         spin_lock_init(&sched->kss_lock);
2442                         INIT_LIST_HEAD(&sched->kss_rx_conns);
2443                         INIT_LIST_HEAD(&sched->kss_tx_conns);
2444                         INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2445                         init_waitqueue_head(&sched->kss_waitq);
2446                 }
2447         }
2448
2449         ksocknal_data.ksnd_connd_starting         = 0;
2450         ksocknal_data.ksnd_connd_failed_stamp     = 0;
2451         ksocknal_data.ksnd_connd_starting_stamp   = cfs_time_current_sec();
2452         /* must have at least 2 connds to remain responsive to accepts while
2453          * connecting */
2454         if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2455                 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2456
2457         if (*ksocknal_tunables.ksnd_nconnds_max <
2458             *ksocknal_tunables.ksnd_nconnds) {
2459                 ksocknal_tunables.ksnd_nconnds_max =
2460                         ksocknal_tunables.ksnd_nconnds;
2461         }
2462
2463         for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2464                 char name[16];
2465                 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2466                 ksocknal_data.ksnd_connd_starting++;
2467                 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2468
2469
2470                 snprintf(name, sizeof(name), "socknal_cd%02d", i);
2471                 rc = ksocknal_thread_start(ksocknal_connd,
2472                                            (void *)((uintptr_t)i), name);
2473                 if (rc != 0) {
2474                         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2475                         ksocknal_data.ksnd_connd_starting--;
2476                         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2477                         CERROR("Can't spawn socknal connd: %d\n", rc);
2478                         goto failed;
2479                 }
2480         }
2481
2482         rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2483         if (rc != 0) {
2484                 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2485                 goto failed;
2486         }
2487
2488         /* flag everything initialised */
2489         ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2490
2491         return 0;
2492
2493  failed:
2494         ksocknal_base_shutdown();
2495         return -ENETDOWN;
2496 }
2497
2498 static void
2499 ksocknal_debug_peerhash (lnet_ni_t *ni)
2500 {
2501         ksock_peer_t    *peer = NULL;
2502         struct list_head        *tmp;
2503         int             i;
2504
2505         read_lock(&ksocknal_data.ksnd_global_lock);
2506
2507         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2508                 list_for_each(tmp, &ksocknal_data.ksnd_peers[i]) {
2509                         peer = list_entry(tmp, ksock_peer_t, ksnp_list);
2510
2511                         if (peer->ksnp_ni == ni) break;
2512
2513                         peer = NULL;
2514                 }
2515         }
2516
2517         if (peer != NULL) {
2518                 ksock_route_t *route;
2519                 ksock_conn_t  *conn;
2520
2521                 CWARN ("Active peer on shutdown: %s, ref %d, scnt %d, "
2522                        "closing %d, accepting %d, err %d, zcookie %llu, "
2523                        "txq %d, zc_req %d\n", libcfs_id2str(peer->ksnp_id),
2524                        atomic_read(&peer->ksnp_refcount),
2525                        peer->ksnp_sharecount, peer->ksnp_closing,
2526                        peer->ksnp_accepting, peer->ksnp_error,
2527                        peer->ksnp_zc_next_cookie,
2528                        !list_empty(&peer->ksnp_tx_queue),
2529                        !list_empty(&peer->ksnp_zc_req_list));
2530
2531                 list_for_each(tmp, &peer->ksnp_routes) {
2532                         route = list_entry(tmp, ksock_route_t, ksnr_list);
2533                         CWARN ("Route: ref %d, schd %d, conn %d, cnted %d, "
2534                                "del %d\n", atomic_read(&route->ksnr_refcount),
2535                                route->ksnr_scheduled, route->ksnr_connecting,
2536                                route->ksnr_connected, route->ksnr_deleted);
2537                 }
2538
2539                 list_for_each(tmp, &peer->ksnp_conns) {
2540                         conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2541                         CWARN ("Conn: ref %d, sref %d, t %d, c %d\n",
2542                                atomic_read(&conn->ksnc_conn_refcount),
2543                                atomic_read(&conn->ksnc_sock_refcount),
2544                                conn->ksnc_type, conn->ksnc_closing);
2545                 }
2546         }
2547
2548         read_unlock(&ksocknal_data.ksnd_global_lock);
2549         return;
2550 }
2551
2552 void
2553 ksocknal_shutdown (lnet_ni_t *ni)
2554 {
2555         ksock_net_t      *net = ni->ni_data;
2556         int               i;
2557         lnet_process_id_t anyid = {0};
2558
2559         anyid.nid =  LNET_NID_ANY;
2560         anyid.pid =  LNET_PID_ANY;
2561
2562         LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2563         LASSERT(ksocknal_data.ksnd_nnets > 0);
2564
2565         spin_lock_bh(&net->ksnn_lock);
2566         net->ksnn_shutdown = 1;                 /* prevent new peers */
2567         spin_unlock_bh(&net->ksnn_lock);
2568
2569         /* Delete all peers */
2570         ksocknal_del_peer(ni, anyid, 0);
2571
2572         /* Wait for all peer state to clean up */
2573         i = 2;
2574         spin_lock_bh(&net->ksnn_lock);
2575         while (net->ksnn_npeers != 0) {
2576                 spin_unlock_bh(&net->ksnn_lock);
2577
2578                 i++;
2579                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2580                        "waiting for %d peers to disconnect\n",
2581                        net->ksnn_npeers);
2582                 set_current_state(TASK_UNINTERRUPTIBLE);
2583                 schedule_timeout(cfs_time_seconds(1));
2584
2585                 ksocknal_debug_peerhash(ni);
2586
2587                 spin_lock_bh(&net->ksnn_lock);
2588         }
2589         spin_unlock_bh(&net->ksnn_lock);
2590
2591         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2592                 LASSERT (net->ksnn_interfaces[i].ksni_npeers == 0);
2593                 LASSERT (net->ksnn_interfaces[i].ksni_nroutes == 0);
2594         }
2595
2596         list_del(&net->ksnn_list);
2597         LIBCFS_FREE(net, sizeof(*net));
2598
2599         ksocknal_data.ksnd_nnets--;
2600         if (ksocknal_data.ksnd_nnets == 0)
2601                 ksocknal_base_shutdown();
2602 }
2603
2604 static int
2605 ksocknal_enumerate_interfaces(ksock_net_t *net)
2606 {
2607         char      **names;
2608         int         i;
2609         int         j;
2610         int         rc;
2611         int         n;
2612
2613         n = lnet_ipif_enumerate(&names);
2614         if (n <= 0) {
2615                 CERROR("Can't enumerate interfaces: %d\n", n);
2616                 return n;
2617         }
2618
2619         for (i = j = 0; i < n; i++) {
2620                 int        up;
2621                 __u32      ip;
2622                 __u32      mask;
2623
2624                 if (!strcmp(names[i], "lo")) /* skip the loopback IF */
2625                         continue;
2626
2627                 rc = lnet_ipif_query(names[i], &up, &ip, &mask);
2628                 if (rc != 0) {
2629                         CWARN("Can't get interface %s info: %d\n",
2630                               names[i], rc);
2631                         continue;
2632                 }
2633
2634                 if (!up) {
2635                         CWARN("Ignoring interface %s (down)\n",
2636                               names[i]);
2637                         continue;
2638                 }
2639
2640                 if (j == LNET_MAX_INTERFACES) {
2641                         CWARN("Ignoring interface %s (too many interfaces)\n",
2642                               names[i]);
2643                         continue;
2644                 }
2645
2646                 net->ksnn_interfaces[j].ksni_ipaddr = ip;
2647                 net->ksnn_interfaces[j].ksni_netmask = mask;
2648                 strlcpy(net->ksnn_interfaces[j].ksni_name,
2649                         names[i], sizeof(net->ksnn_interfaces[j].ksni_name));
2650                 j++;
2651         }
2652
2653         lnet_ipif_free_enumeration(names, n);
2654
2655         if (j == 0)
2656                 CERROR("Can't find any usable interfaces\n");
2657
2658         return j;
2659 }
2660
2661 static int
2662 ksocknal_search_new_ipif(ksock_net_t *net)
2663 {
2664         int     new_ipif = 0;
2665         int     i;
2666
2667         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2668                 char            *ifnam = &net->ksnn_interfaces[i].ksni_name[0];
2669                 char            *colon = strchr(ifnam, ':');
2670                 int             found  = 0;
2671                 ksock_net_t     *tmp;
2672                 int             j;
2673
2674                 if (colon != NULL) /* ignore alias device */
2675                         *colon = 0;
2676
2677                 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets,
2678                                         ksnn_list) {
2679                         for (j = 0; !found && j < tmp->ksnn_ninterfaces; j++) {
2680                                 char *ifnam2 = &tmp->ksnn_interfaces[j].\
2681                                              ksni_name[0];
2682                                 char *colon2 = strchr(ifnam2, ':');
2683
2684                                 if (colon2 != NULL)
2685                                         *colon2 = 0;
2686
2687                                 found = strcmp(ifnam, ifnam2) == 0;
2688                                 if (colon2 != NULL)
2689                                         *colon2 = ':';
2690                         }
2691                         if (found)
2692                                 break;
2693                 }
2694
2695                 new_ipif += !found;
2696                 if (colon != NULL)
2697                         *colon = ':';
2698         }
2699
2700         return new_ipif;
2701 }
2702
2703 static int
2704 ksocknal_start_schedulers(struct ksock_sched_info *info)
2705 {
2706         int     nthrs;
2707         int     rc = 0;
2708         int     i;
2709
2710         if (info->ksi_nthreads == 0) {
2711                 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2712                         nthrs = info->ksi_nthreads_max;
2713                 } else {
2714                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2715                                                info->ksi_cpt);
2716                         nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2717                         nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2718                 }
2719                 nthrs = min(nthrs, info->ksi_nthreads_max);
2720         } else {
2721                 LASSERT(info->ksi_nthreads <= info->ksi_nthreads_max);
2722                 /* increase two threads if there is new interface */
2723                 nthrs = min(2, info->ksi_nthreads_max - info->ksi_nthreads);
2724         }
2725
2726         for (i = 0; i < nthrs; i++) {
2727                 long            id;
2728                 char            name[20];
2729                 ksock_sched_t   *sched;
2730                 id = KSOCK_THREAD_ID(info->ksi_cpt, info->ksi_nthreads + i);
2731                 sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
2732                 snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
2733                          info->ksi_cpt, (int)(sched - &info->ksi_scheds[0]));
2734
2735                 rc = ksocknal_thread_start(ksocknal_scheduler,
2736                                            (void *)id, name);
2737                 if (rc == 0)
2738                         continue;
2739
2740                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2741                        info->ksi_cpt, info->ksi_nthreads + i, rc);
2742                 break;
2743         }
2744
2745         info->ksi_nthreads += i;
2746         return rc;
2747 }
2748
2749 static int
2750 ksocknal_net_start_threads(ksock_net_t *net, __u32 *cpts, int ncpts)
2751 {
2752         int     newif = ksocknal_search_new_ipif(net);
2753         int     rc;
2754         int     i;
2755
2756         LASSERT(ncpts > 0 && ncpts <= cfs_cpt_number(lnet_cpt_table()));
2757
2758         for (i = 0; i < ncpts; i++) {
2759                 struct ksock_sched_info *info;
2760                 int cpt = (cpts == NULL) ? i : cpts[i];
2761
2762                 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2763                 info = ksocknal_data.ksnd_sched_info[cpt];
2764
2765                 if (!newif && info->ksi_nthreads > 0)
2766                         continue;
2767
2768                 rc = ksocknal_start_schedulers(info);
2769                 if (rc != 0)
2770                         return rc;
2771         }
2772         return 0;
2773 }
2774
2775 int
2776 ksocknal_startup (lnet_ni_t *ni)
2777 {
2778         ksock_net_t  *net;
2779         int           rc;
2780         int           i;
2781
2782         LASSERT (ni->ni_lnd == &the_ksocklnd);
2783
2784         if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2785                 rc = ksocknal_base_startup();
2786                 if (rc != 0)
2787                         return rc;
2788         }
2789
2790         LIBCFS_ALLOC(net, sizeof(*net));
2791         if (net == NULL)
2792                 goto fail_0;
2793
2794         spin_lock_init(&net->ksnn_lock);
2795         net->ksnn_incarnation = ksocknal_new_incarnation();
2796         ni->ni_data = net;
2797         ni->ni_peertimeout    = *ksocknal_tunables.ksnd_peertimeout;
2798         ni->ni_maxtxcredits   = *ksocknal_tunables.ksnd_credits;
2799         ni->ni_peertxcredits  = *ksocknal_tunables.ksnd_peertxcredits;
2800         ni->ni_peerrtrcredits = *ksocknal_tunables.ksnd_peerrtrcredits;
2801
2802         if (ni->ni_interfaces[0] == NULL) {
2803                 rc = ksocknal_enumerate_interfaces(net);
2804                 if (rc <= 0)
2805                         goto fail_1;
2806
2807                 net->ksnn_ninterfaces = 1;
2808         } else {
2809                 for (i = 0; i < LNET_MAX_INTERFACES; i++) {
2810                         int    up;
2811
2812                         if (ni->ni_interfaces[i] == NULL)
2813                                 break;
2814
2815                         rc = lnet_ipif_query(ni->ni_interfaces[i], &up,
2816                                 &net->ksnn_interfaces[i].ksni_ipaddr,
2817                                 &net->ksnn_interfaces[i].ksni_netmask);
2818
2819                         if (rc != 0) {
2820                                 CERROR("Can't get interface %s info: %d\n",
2821                                        ni->ni_interfaces[i], rc);
2822                                 goto fail_1;
2823                         }
2824
2825                         if (!up) {
2826                                 CERROR("Interface %s is down\n",
2827                                        ni->ni_interfaces[i]);
2828                                 goto fail_1;
2829                         }
2830
2831                         strlcpy(net->ksnn_interfaces[i].ksni_name,
2832                                 ni->ni_interfaces[i],
2833                                 sizeof(net->ksnn_interfaces[i].ksni_name));
2834                 }
2835                 net->ksnn_ninterfaces = i;
2836         }
2837
2838         /* call it before add it to ksocknal_data.ksnd_nets */
2839         rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2840         if (rc != 0)
2841                 goto fail_1;
2842
2843         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid),
2844                                 net->ksnn_interfaces[0].ksni_ipaddr);
2845         list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2846
2847         ksocknal_data.ksnd_nnets++;
2848
2849         return 0;
2850
2851  fail_1:
2852         LIBCFS_FREE(net, sizeof(*net));
2853  fail_0:
2854         if (ksocknal_data.ksnd_nnets == 0)
2855                 ksocknal_base_shutdown();
2856
2857         return -ENETDOWN;
2858 }
2859
2860
2861 static void __exit ksocklnd_exit(void)
2862 {
2863         lnet_unregister_lnd(&the_ksocklnd);
2864 }
2865
2866 static int __init ksocklnd_init(void)
2867 {
2868         int rc;
2869
2870         /* check ksnr_connected/connecting field large enough */
2871         CLASSERT(SOCKLND_CONN_NTYPES <= 4);
2872         CLASSERT(SOCKLND_CONN_ACK == SOCKLND_CONN_BULK_IN);
2873
2874         /* initialize the_ksocklnd */
2875         the_ksocklnd.lnd_type     = SOCKLND;
2876         the_ksocklnd.lnd_startup  = ksocknal_startup;
2877         the_ksocklnd.lnd_shutdown = ksocknal_shutdown;
2878         the_ksocklnd.lnd_ctl      = ksocknal_ctl;
2879         the_ksocklnd.lnd_send     = ksocknal_send;
2880         the_ksocklnd.lnd_recv     = ksocknal_recv;
2881         the_ksocklnd.lnd_notify   = ksocknal_notify;
2882         the_ksocklnd.lnd_query    = ksocknal_query;
2883         the_ksocklnd.lnd_accept   = ksocknal_accept;
2884
2885         rc = ksocknal_tunables_init();
2886         if (rc != 0)
2887                 return rc;
2888
2889         lnet_register_lnd(&the_ksocklnd);
2890
2891         return 0;
2892 }
2893
2894 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2895 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2896 MODULE_VERSION("2.8.0");
2897 MODULE_LICENSE("GPL");
2898
2899 module_init(ksocklnd_init);
2900 module_exit(ksocklnd_exit);