lnet/klnds/socklnd/socklnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2015, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/klnds/socklnd/socklnd.c
  33  *
  34  * Author: Zach Brown <zab@zabbo.net>
  35  * Author: Peter J. Braam <braam@clusterfs.com>
  36  * Author: Phil Schwan <phil@clusterfs.com>
  37  * Author: Eric Barton <eric@bartonsoftware.com>
  38  */
  39
  40 #include <linux/pci.h>
  41 #include "socklnd.h"
  42
  43 static struct lnet_lnd the_ksocklnd;
  44 ksock_nal_data_t        ksocknal_data;
  45
  46 static ksock_interface_t *
  47 ksocknal_ip2iface(struct lnet_ni *ni, __u32 ip)
  48 {
  49         ksock_net_t *net = ni->ni_data;
  50         int i;
  51         ksock_interface_t *iface;
  52
  53         for (i = 0; i < net->ksnn_ninterfaces; i++) {
  54                 LASSERT(i < LNET_NUM_INTERFACES);
  55                 iface = &net->ksnn_interfaces[i];
  56
  57                 if (iface->ksni_ipaddr == ip)
  58                         return iface;
  59         }
  60
  61         return NULL;
  62 }
  63
  64 static ksock_route_t *
  65 ksocknal_create_route (__u32 ipaddr, int port)
  66 {
  67         ksock_route_t *route;
  68
  69         LIBCFS_ALLOC (route, sizeof (*route));
  70         if (route == NULL)
  71                 return (NULL);
  72
  73         atomic_set (&route->ksnr_refcount, 1);
  74         route->ksnr_peer = NULL;
  75         route->ksnr_retry_interval = 0;         /* OK to connect at any time */
  76         route->ksnr_ipaddr = ipaddr;
  77         route->ksnr_port = port;
  78         route->ksnr_scheduled = 0;
  79         route->ksnr_connecting = 0;
  80         route->ksnr_connected = 0;
  81         route->ksnr_deleted = 0;
  82         route->ksnr_conn_count = 0;
  83         route->ksnr_share_count = 0;
  84
  85         return (route);
  86 }
  87
  88 void
  89 ksocknal_destroy_route (ksock_route_t *route)
  90 {
  91         LASSERT (atomic_read(&route->ksnr_refcount) == 0);
  92
  93         if (route->ksnr_peer != NULL)
  94                 ksocknal_peer_decref(route->ksnr_peer);
  95
  96         LIBCFS_FREE (route, sizeof (*route));
  97 }
  98
  99 static int
 100 ksocknal_create_peer(ksock_peer_ni_t **peerp, struct lnet_ni *ni,
 101                      struct lnet_process_id id)
 102 {
 103         int             cpt = lnet_cpt_of_nid(id.nid, ni);
 104         ksock_net_t     *net = ni->ni_data;
 105         ksock_peer_ni_t *peer_ni;
 106
 107         LASSERT(id.nid != LNET_NID_ANY);
 108         LASSERT(id.pid != LNET_PID_ANY);
 109         LASSERT(!in_interrupt());
 110
 111         LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
 112         if (peer_ni == NULL)
 113                 return -ENOMEM;
 114
 115         peer_ni->ksnp_ni = ni;
 116         peer_ni->ksnp_id = id;
 117         atomic_set(&peer_ni->ksnp_refcount, 1); /* 1 ref for caller */
 118         peer_ni->ksnp_closing = 0;
 119         peer_ni->ksnp_accepting = 0;
 120         peer_ni->ksnp_proto = NULL;
 121         peer_ni->ksnp_last_alive = 0;
 122         peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 123
 124         INIT_LIST_HEAD(&peer_ni->ksnp_conns);
 125         INIT_LIST_HEAD(&peer_ni->ksnp_routes);
 126         INIT_LIST_HEAD(&peer_ni->ksnp_tx_queue);
 127         INIT_LIST_HEAD(&peer_ni->ksnp_zc_req_list);
 128         spin_lock_init(&peer_ni->ksnp_lock);
 129
 130         spin_lock_bh(&net->ksnn_lock);
 131
 132         if (net->ksnn_shutdown) {
 133                 spin_unlock_bh(&net->ksnn_lock);
 134
 135                 LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
 136                 CERROR("Can't create peer_ni: network shutdown\n");
 137                 return -ESHUTDOWN;
 138         }
 139
 140         net->ksnn_npeers++;
 141
 142         spin_unlock_bh(&net->ksnn_lock);
 143
 144         *peerp = peer_ni;
 145         return 0;
 146 }
 147
 148 void
 149 ksocknal_destroy_peer (ksock_peer_ni_t *peer_ni)
 150 {
 151         ksock_net_t    *net = peer_ni->ksnp_ni->ni_data;
 152
 153         CDEBUG (D_NET, "peer_ni %s %p deleted\n",
 154                 libcfs_id2str(peer_ni->ksnp_id), peer_ni);
 155
 156         LASSERT(atomic_read(&peer_ni->ksnp_refcount) == 0);
 157         LASSERT(peer_ni->ksnp_accepting == 0);
 158         LASSERT(list_empty(&peer_ni->ksnp_conns));
 159         LASSERT(list_empty(&peer_ni->ksnp_routes));
 160         LASSERT(list_empty(&peer_ni->ksnp_tx_queue));
 161         LASSERT(list_empty(&peer_ni->ksnp_zc_req_list));
 162
 163         LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
 164
 165         /* NB a peer_ni's connections and routes keep a reference on their peer_ni
 166          * until they are destroyed, so we can be assured that _all_ state to
 167          * do with this peer_ni has been cleaned up when its refcount drops to
 168          * zero. */
 169         spin_lock_bh(&net->ksnn_lock);
 170         net->ksnn_npeers--;
 171         spin_unlock_bh(&net->ksnn_lock);
 172 }
 173
 174 ksock_peer_ni_t *
 175 ksocknal_find_peer_locked(struct lnet_ni *ni, struct lnet_process_id id)
 176 {
 177         struct list_head *peer_list = ksocknal_nid2peerlist(id.nid);
 178         struct list_head *tmp;
 179         ksock_peer_ni_t  *peer_ni;
 180
 181         list_for_each(tmp, peer_list) {
 182
 183                 peer_ni = list_entry(tmp, ksock_peer_ni_t, ksnp_list);
 184
 185                 LASSERT(!peer_ni->ksnp_closing);
 186
 187                 if (peer_ni->ksnp_ni != ni)
 188                         continue;
 189
 190                 if (peer_ni->ksnp_id.nid != id.nid ||
 191                     peer_ni->ksnp_id.pid != id.pid)
 192                         continue;
 193
 194                 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d)\n",
 195                        peer_ni, libcfs_id2str(id),
 196                        atomic_read(&peer_ni->ksnp_refcount));
 197                 return peer_ni;
 198         }
 199         return NULL;
 200 }
 201
 202 ksock_peer_ni_t *
 203 ksocknal_find_peer(struct lnet_ni *ni, struct lnet_process_id id)
 204 {
 205         ksock_peer_ni_t     *peer_ni;
 206
 207         read_lock(&ksocknal_data.ksnd_global_lock);
 208         peer_ni = ksocknal_find_peer_locked(ni, id);
 209         if (peer_ni != NULL)                    /* +1 ref for caller? */
 210                 ksocknal_peer_addref(peer_ni);
 211         read_unlock(&ksocknal_data.ksnd_global_lock);
 212
 213         return (peer_ni);
 214 }
 215
 216 static void
 217 ksocknal_unlink_peer_locked(ksock_peer_ni_t *peer_ni)
 218 {
 219         int i;
 220         __u32 ip;
 221         ksock_interface_t *iface;
 222
 223         for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++) {
 224                 LASSERT(i < LNET_NUM_INTERFACES);
 225                 ip = peer_ni->ksnp_passive_ips[i];
 226
 227                 iface = ksocknal_ip2iface(peer_ni->ksnp_ni, ip);
 228                 /*
 229                  * All IPs in peer_ni->ksnp_passive_ips[] come from the
 230                  * interface list, therefore the call must succeed.
 231                  */
 232                 LASSERT(iface != NULL);
 233
 234                 CDEBUG(D_NET, "peer_ni=%p iface=%p ksni_nroutes=%d\n",
 235                        peer_ni, iface, iface->ksni_nroutes);
 236                 iface->ksni_npeers--;
 237         }
 238
 239         LASSERT(list_empty(&peer_ni->ksnp_conns));
 240         LASSERT(list_empty(&peer_ni->ksnp_routes));
 241         LASSERT(!peer_ni->ksnp_closing);
 242         peer_ni->ksnp_closing = 1;
 243         list_del(&peer_ni->ksnp_list);
 244         /* lose peerlist's ref */
 245         ksocknal_peer_decref(peer_ni);
 246 }
 247
 248 static int
 249 ksocknal_get_peer_info(struct lnet_ni *ni, int index,
 250                        struct lnet_process_id *id, __u32 *myip, __u32 *peer_ip,
 251                        int *port, int *conn_count, int *share_count)
 252 {
 253         ksock_peer_ni_t   *peer_ni;
 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_ni = list_entry(ptmp, ksock_peer_ni_t, ksnp_list);
 266
 267                         if (peer_ni->ksnp_ni != ni)
 268                                 continue;
 269
 270                         if (peer_ni->ksnp_n_passive_ips == 0 &&
 271                             list_empty(&peer_ni->ksnp_routes)) {
 272                                 if (index-- > 0)
 273                                         continue;
 274
 275                                 *id = peer_ni->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_ni->ksnp_n_passive_ips; j++) {
 286                                 if (index-- > 0)
 287                                         continue;
 288
 289                                 *id = peer_ni->ksnp_id;
 290                                 *myip = peer_ni->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_ni->ksnp_routes) {
 300                                 if (index-- > 0)
 301                                         continue;
 302
 303                                 route = list_entry(rtmp, ksock_route_t,
 304                                                    ksnr_list);
 305
 306                                 *id = peer_ni->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_ni_t   *peer_ni = 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_ni->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_ni->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_ni_t *peer_ni, ksock_route_t *route)
 368 {
 369         struct list_head *tmp;
 370         ksock_conn_t     *conn;
 371         ksock_route_t    *route2;
 372
 373         LASSERT(!peer_ni->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_ni->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_ni->ksnp_id),
 386                                &route->ksnr_ipaddr);
 387                         LBUG();
 388                 }
 389         }
 390
 391         route->ksnr_peer = peer_ni;
 392         ksocknal_peer_addref(peer_ni);
 393         /* peer_ni's routelist takes over my ref on 'route' */
 394         list_add_tail(&route->ksnr_list, &peer_ni->ksnp_routes);
 395
 396         list_for_each(tmp, &peer_ni->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_ni_t   *peer_ni = 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_ni->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_ni's ref */
 438
 439         if (list_empty(&peer_ni->ksnp_routes) &&
 440             list_empty(&peer_ni->ksnp_conns)) {
 441                 /* I've just removed the last route to a peer_ni with no active
 442                  * connections */
 443                 ksocknal_unlink_peer_locked(peer_ni);
 444         }
 445 }
 446
 447 int
 448 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ipaddr,
 449                   int port)
 450 {
 451         struct list_head *tmp;
 452         ksock_peer_ni_t  *peer_ni;
 453         ksock_peer_ni_t  *peer2;
 454         ksock_route_t    *route;
 455         ksock_route_t    *route2;
 456         int               rc;
 457
 458         if (id.nid == LNET_NID_ANY ||
 459             id.pid == LNET_PID_ANY)
 460                 return (-EINVAL);
 461
 462         /* Have a brand new peer_ni ready... */
 463         rc = ksocknal_create_peer(&peer_ni, ni, id);
 464         if (rc != 0)
 465                 return rc;
 466
 467         route = ksocknal_create_route (ipaddr, port);
 468         if (route == NULL) {
 469                 ksocknal_peer_decref(peer_ni);
 470                 return (-ENOMEM);
 471         }
 472
 473         write_lock_bh(&ksocknal_data.ksnd_global_lock);
 474
 475         /* always called with a ref on ni, so shutdown can't have started */
 476         LASSERT (((ksock_net_t *) ni->ni_data)->ksnn_shutdown == 0);
 477
 478         peer2 = ksocknal_find_peer_locked(ni, id);
 479         if (peer2 != NULL) {
 480                 ksocknal_peer_decref(peer_ni);
 481                 peer_ni = peer2;
 482         } else {
 483                 /* peer_ni table takes my ref on peer_ni */
 484                 list_add_tail(&peer_ni->ksnp_list,
 485                               ksocknal_nid2peerlist(id.nid));
 486         }
 487
 488         route2 = NULL;
 489         list_for_each(tmp, &peer_ni->ksnp_routes) {
 490                 route2 = list_entry(tmp, ksock_route_t, ksnr_list);
 491
 492                 if (route2->ksnr_ipaddr == ipaddr)
 493                         break;
 494
 495                 route2 = NULL;
 496         }
 497         if (route2 == NULL) {
 498                 ksocknal_add_route_locked(peer_ni, route);
 499                 route->ksnr_share_count++;
 500         } else {
 501                 ksocknal_route_decref(route);
 502                 route2->ksnr_share_count++;
 503         }
 504
 505         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 506
 507         return 0;
 508 }
 509
 510 static void
 511 ksocknal_del_peer_locked (ksock_peer_ni_t *peer_ni, __u32 ip)
 512 {
 513         ksock_conn_t     *conn;
 514         ksock_route_t    *route;
 515         struct list_head *tmp;
 516         struct list_head *nxt;
 517         int               nshared;
 518
 519         LASSERT(!peer_ni->ksnp_closing);
 520
 521         /* Extra ref prevents peer_ni disappearing until I'm done with it */
 522         ksocknal_peer_addref(peer_ni);
 523
 524         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
 525                 route = list_entry(tmp, ksock_route_t, ksnr_list);
 526
 527                 /* no match */
 528                 if (!(ip == 0 || route->ksnr_ipaddr == ip))
 529                         continue;
 530
 531                 route->ksnr_share_count = 0;
 532                 /* This deletes associated conns too */
 533                 ksocknal_del_route_locked(route);
 534         }
 535
 536         nshared = 0;
 537         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
 538                 route = list_entry(tmp, ksock_route_t, ksnr_list);
 539                 nshared += route->ksnr_share_count;
 540         }
 541
 542         if (nshared == 0) {
 543                 /* remove everything else if there are no explicit entries
 544                  * left */
 545
 546                 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
 547                         route = list_entry(tmp, ksock_route_t, ksnr_list);
 548
 549                         /* we should only be removing auto-entries */
 550                         LASSERT(route->ksnr_share_count == 0);
 551                         ksocknal_del_route_locked(route);
 552                 }
 553
 554                 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_conns) {
 555                         conn = list_entry(tmp, ksock_conn_t, ksnc_list);
 556
 557                         ksocknal_close_conn_locked(conn, 0);
 558                 }
 559         }
 560
 561         ksocknal_peer_decref(peer_ni);
 562                 /* NB peer_ni unlinks itself when last conn/route is removed */
 563 }
 564
 565 static int
 566 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ip)
 567 {
 568         struct list_head  zombies = LIST_HEAD_INIT(zombies);
 569         struct list_head *ptmp;
 570         struct list_head *pnxt;
 571         ksock_peer_ni_t     *peer_ni;
 572         int               lo;
 573         int               hi;
 574         int               i;
 575         int               rc = -ENOENT;
 576
 577         write_lock_bh(&ksocknal_data.ksnd_global_lock);
 578
 579         if (id.nid != LNET_NID_ANY) {
 580                 hi = (int)(ksocknal_nid2peerlist(id.nid) -
 581                            ksocknal_data.ksnd_peers);
 582                 lo = hi;
 583         } else {
 584                 lo = 0;
 585                 hi = ksocknal_data.ksnd_peer_hash_size - 1;
 586         }
 587
 588         for (i = lo; i <= hi; i++) {
 589                 list_for_each_safe(ptmp, pnxt,
 590                                    &ksocknal_data.ksnd_peers[i]) {
 591                         peer_ni = list_entry(ptmp, ksock_peer_ni_t, ksnp_list);
 592
 593                         if (peer_ni->ksnp_ni != ni)
 594                                 continue;
 595
 596                         if (!((id.nid == LNET_NID_ANY ||
 597                                peer_ni->ksnp_id.nid == id.nid) &&
 598                               (id.pid == LNET_PID_ANY ||
 599                                peer_ni->ksnp_id.pid == id.pid)))
 600                                 continue;
 601
 602                         ksocknal_peer_addref(peer_ni);  /* a ref for me... */
 603
 604                         ksocknal_del_peer_locked(peer_ni, ip);
 605
 606                         if (peer_ni->ksnp_closing &&
 607                             !list_empty(&peer_ni->ksnp_tx_queue)) {
 608                                 LASSERT(list_empty(&peer_ni->ksnp_conns));
 609                                 LASSERT(list_empty(&peer_ni->ksnp_routes));
 610
 611                                 list_splice_init(&peer_ni->ksnp_tx_queue,
 612                                                  &zombies);
 613                         }
 614
 615                         ksocknal_peer_decref(peer_ni);  /* ...till here */
 616
 617                         rc = 0;                         /* matched! */
 618                 }
 619         }
 620
 621         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 622
 623         ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
 624
 625         return rc;
 626 }
 627
 628 static ksock_conn_t *
 629 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
 630 {
 631         ksock_peer_ni_t  *peer_ni;
 632         struct list_head *ptmp;
 633         ksock_conn_t     *conn;
 634         struct list_head *ctmp;
 635         int               i;
 636
 637         read_lock(&ksocknal_data.ksnd_global_lock);
 638
 639         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
 640                 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
 641                         peer_ni = list_entry(ptmp, ksock_peer_ni_t, ksnp_list);
 642
 643                         LASSERT(!peer_ni->ksnp_closing);
 644
 645                         if (peer_ni->ksnp_ni != ni)
 646                                 continue;
 647
 648                         list_for_each(ctmp, &peer_ni->ksnp_conns) {
 649                                 if (index-- > 0)
 650                                         continue;
 651
 652                                 conn = list_entry(ctmp, ksock_conn_t,
 653                                                   ksnc_list);
 654                                 ksocknal_conn_addref(conn);
 655                                 read_unlock(&ksocknal_data. \
 656                                             ksnd_global_lock);
 657                                 return conn;
 658                         }
 659                 }
 660         }
 661
 662         read_unlock(&ksocknal_data.ksnd_global_lock);
 663         return NULL;
 664 }
 665
 666 static ksock_sched_t *
 667 ksocknal_choose_scheduler_locked(unsigned int cpt)
 668 {
 669         struct ksock_sched_info *info = ksocknal_data.ksnd_sched_info[cpt];
 670         ksock_sched_t           *sched;
 671         int                     i;
 672
 673         LASSERT(info->ksi_nthreads > 0);
 674
 675         sched = &info->ksi_scheds[0];
 676         /*
 677          * NB: it's safe so far, but info->ksi_nthreads could be changed
 678          * at runtime when we have dynamic LNet configuration, then we
 679          * need to take care of this.
 680          */
 681         for (i = 1; i < info->ksi_nthreads; i++) {
 682                 if (sched->kss_nconns > info->ksi_scheds[i].kss_nconns)
 683                         sched = &info->ksi_scheds[i];
 684         }
 685
 686         return sched;
 687 }
 688
 689 static int
 690 ksocknal_local_ipvec(struct lnet_ni *ni, __u32 *ipaddrs)
 691 {
 692         ksock_net_t *net = ni->ni_data;
 693         int i;
 694         int nip;
 695
 696         read_lock(&ksocknal_data.ksnd_global_lock);
 697
 698         nip = net->ksnn_ninterfaces;
 699         LASSERT(nip <= LNET_NUM_INTERFACES);
 700
 701         /*
 702          * Only offer interfaces for additional connections if I have
 703          * more than one.
 704          */
 705         if (nip < 2) {
 706                 read_unlock(&ksocknal_data.ksnd_global_lock);
 707                 return 0;
 708         }
 709
 710         for (i = 0; i < nip; i++) {
 711                 ipaddrs[i] = net->ksnn_interfaces[i].ksni_ipaddr;
 712                 LASSERT(ipaddrs[i] != 0);
 713         }
 714
 715         read_unlock(&ksocknal_data.ksnd_global_lock);
 716         return nip;
 717 }
 718
 719 static int
 720 ksocknal_match_peerip (ksock_interface_t *iface, __u32 *ips, int nips)
 721 {
 722         int   best_netmatch = 0;
 723         int   best_xor      = 0;
 724         int   best          = -1;
 725         int   this_xor;
 726         int   this_netmatch;
 727         int   i;
 728
 729         for (i = 0; i < nips; i++) {
 730                 if (ips[i] == 0)
 731                         continue;
 732
 733                 this_xor = (ips[i] ^ iface->ksni_ipaddr);
 734                 this_netmatch = ((this_xor & iface->ksni_netmask) == 0) ? 1 : 0;
 735
 736                 if (!(best < 0 ||
 737                       best_netmatch < this_netmatch ||
 738                       (best_netmatch == this_netmatch &&
 739                        best_xor > this_xor)))
 740                         continue;
 741
 742                 best = i;
 743                 best_netmatch = this_netmatch;
 744                 best_xor = this_xor;
 745         }
 746
 747         LASSERT (best >= 0);
 748         return (best);
 749 }
 750
 751 static int
 752 ksocknal_select_ips(ksock_peer_ni_t *peer_ni, __u32 *peerips, int n_peerips)
 753 {
 754         rwlock_t                *global_lock = &ksocknal_data.ksnd_global_lock;
 755         ksock_net_t        *net = peer_ni->ksnp_ni->ni_data;
 756         ksock_interface_t  *iface;
 757         ksock_interface_t  *best_iface;
 758         int                 n_ips;
 759         int                 i;
 760         int                 j;
 761         int                 k;
 762         __u32               ip;
 763         __u32               xor;
 764         int                 this_netmatch;
 765         int                 best_netmatch;
 766         int                 best_npeers;
 767
 768         /* CAVEAT EMPTOR: We do all our interface matching with an
 769          * exclusive hold of global lock at IRQ priority.  We're only
 770          * expecting to be dealing with small numbers of interfaces, so the
 771          * O(n**3)-ness shouldn't matter */
 772
 773         /* Also note that I'm not going to return more than n_peerips
 774          * interfaces, even if I have more myself */
 775
 776         write_lock_bh(global_lock);
 777
 778         LASSERT(n_peerips <= LNET_NUM_INTERFACES);
 779         LASSERT(net->ksnn_ninterfaces <= LNET_NUM_INTERFACES);
 780
 781         /* Only match interfaces for additional connections
 782          * if I have > 1 interface */
 783         n_ips = (net->ksnn_ninterfaces < 2) ? 0 :
 784                 MIN(n_peerips, net->ksnn_ninterfaces);
 785
 786         for (i = 0; peer_ni->ksnp_n_passive_ips < n_ips; i++) {
 787                 /*              ^ yes really... */
 788
 789                 /* If we have any new interfaces, first tick off all the
 790                  * peer_ni IPs that match old interfaces, then choose new
 791                  * interfaces to match the remaining peer_ni IPS.
 792                  * We don't forget interfaces we've stopped using; we might
 793                  * start using them again... */
 794
 795                 if (i < peer_ni->ksnp_n_passive_ips) {
 796                         /* Old interface. */
 797                         ip = peer_ni->ksnp_passive_ips[i];
 798                         best_iface = ksocknal_ip2iface(peer_ni->ksnp_ni, ip);
 799
 800                         /* peer_ni passive ips are kept up to date */
 801                         LASSERT(best_iface != NULL);
 802                 } else {
 803                         /* choose a new interface */
 804                         LASSERT (i == peer_ni->ksnp_n_passive_ips);
 805
 806                         best_iface = NULL;
 807                         best_netmatch = 0;
 808                         best_npeers = 0;
 809
 810                         for (j = 0; j < net->ksnn_ninterfaces; j++) {
 811                                 iface = &net->ksnn_interfaces[j];
 812                                 ip = iface->ksni_ipaddr;
 813
 814                                 for (k = 0; k < peer_ni->ksnp_n_passive_ips; k++)
 815                                         if (peer_ni->ksnp_passive_ips[k] == ip)
 816                                                 break;
 817
 818                                 if (k < peer_ni->ksnp_n_passive_ips) /* using it already */
 819                                         continue;
 820
 821                                 k = ksocknal_match_peerip(iface, peerips, n_peerips);
 822                                 xor = (ip ^ peerips[k]);
 823                                 this_netmatch = ((xor & iface->ksni_netmask) == 0) ? 1 : 0;
 824
 825                                 if (!(best_iface == NULL ||
 826                                       best_netmatch < this_netmatch ||
 827                                       (best_netmatch == this_netmatch &&
 828                                        best_npeers > iface->ksni_npeers)))
 829                                         continue;
 830
 831                                 best_iface = iface;
 832                                 best_netmatch = this_netmatch;
 833                                 best_npeers = iface->ksni_npeers;
 834                         }
 835
 836                         LASSERT(best_iface != NULL);
 837
 838                         best_iface->ksni_npeers++;
 839                         ip = best_iface->ksni_ipaddr;
 840                         peer_ni->ksnp_passive_ips[i] = ip;
 841                         peer_ni->ksnp_n_passive_ips = i+1;
 842                 }
 843
 844                 /* mark the best matching peer_ni IP used */
 845                 j = ksocknal_match_peerip(best_iface, peerips, n_peerips);
 846                 peerips[j] = 0;
 847         }
 848
 849         /* Overwrite input peer_ni IP addresses */
 850         memcpy(peerips, peer_ni->ksnp_passive_ips, n_ips * sizeof(*peerips));
 851
 852         write_unlock_bh(global_lock);
 853
 854         return (n_ips);
 855 }
 856
 857 static void
 858 ksocknal_create_routes(ksock_peer_ni_t *peer_ni, int port,
 859                        __u32 *peer_ipaddrs, int npeer_ipaddrs)
 860 {
 861         ksock_route_t           *newroute = NULL;
 862         rwlock_t                *global_lock = &ksocknal_data.ksnd_global_lock;
 863         struct lnet_ni *ni = peer_ni->ksnp_ni;
 864         ksock_net_t             *net = ni->ni_data;
 865         struct list_head        *rtmp;
 866         ksock_route_t           *route;
 867         ksock_interface_t       *iface;
 868         ksock_interface_t       *best_iface;
 869         int                     best_netmatch;
 870         int                     this_netmatch;
 871         int                     best_nroutes;
 872         int                     i;
 873         int                     j;
 874
 875         /* CAVEAT EMPTOR: We do all our interface matching with an
 876          * exclusive hold of global lock at IRQ priority.  We're only
 877          * expecting to be dealing with small numbers of interfaces, so the
 878          * O(n**3)-ness here shouldn't matter */
 879
 880         write_lock_bh(global_lock);
 881
 882         if (net->ksnn_ninterfaces < 2) {
 883                 /* Only create additional connections
 884                  * if I have > 1 interface */
 885                 write_unlock_bh(global_lock);
 886                 return;
 887         }
 888
 889         LASSERT(npeer_ipaddrs <= LNET_NUM_INTERFACES);
 890
 891         for (i = 0; i < npeer_ipaddrs; i++) {
 892                 if (newroute != NULL) {
 893                         newroute->ksnr_ipaddr = peer_ipaddrs[i];
 894                 } else {
 895                         write_unlock_bh(global_lock);
 896
 897                         newroute = ksocknal_create_route(peer_ipaddrs[i], port);
 898                         if (newroute == NULL)
 899                                 return;
 900
 901                         write_lock_bh(global_lock);
 902                 }
 903
 904                 if (peer_ni->ksnp_closing) {
 905                         /* peer_ni got closed under me */
 906                         break;
 907                 }
 908
 909                 /* Already got a route? */
 910                 route = NULL;
 911                 list_for_each(rtmp, &peer_ni->ksnp_routes) {
 912                         route = list_entry(rtmp, ksock_route_t, ksnr_list);
 913
 914                         if (route->ksnr_ipaddr == newroute->ksnr_ipaddr)
 915                                 break;
 916
 917                         route = NULL;
 918                 }
 919                 if (route != NULL)
 920                         continue;
 921
 922                 best_iface = NULL;
 923                 best_nroutes = 0;
 924                 best_netmatch = 0;
 925
 926                 LASSERT(net->ksnn_ninterfaces <= LNET_NUM_INTERFACES);
 927
 928                 /* Select interface to connect from */
 929                 for (j = 0; j < net->ksnn_ninterfaces; j++) {
 930                         iface = &net->ksnn_interfaces[j];
 931
 932                         /* Using this interface already? */
 933                         list_for_each(rtmp, &peer_ni->ksnp_routes) {
 934                                 route = list_entry(rtmp, ksock_route_t,
 935                                                    ksnr_list);
 936
 937                                 if (route->ksnr_myipaddr == iface->ksni_ipaddr)
 938                                         break;
 939
 940                                 route = NULL;
 941                         }
 942                         if (route != NULL)
 943                                 continue;
 944
 945                         this_netmatch = (((iface->ksni_ipaddr ^
 946                                            newroute->ksnr_ipaddr) &
 947                                            iface->ksni_netmask) == 0) ? 1 : 0;
 948
 949                         if (!(best_iface == NULL ||
 950                               best_netmatch < this_netmatch ||
 951                               (best_netmatch == this_netmatch &&
 952                                best_nroutes > iface->ksni_nroutes)))
 953                                 continue;
 954
 955                         best_iface = iface;
 956                         best_netmatch = this_netmatch;
 957                         best_nroutes = iface->ksni_nroutes;
 958                 }
 959
 960                 if (best_iface == NULL)
 961                         continue;
 962
 963                 newroute->ksnr_myipaddr = best_iface->ksni_ipaddr;
 964                 best_iface->ksni_nroutes++;
 965
 966                 ksocknal_add_route_locked(peer_ni, newroute);
 967                 newroute = NULL;
 968         }
 969
 970         write_unlock_bh(global_lock);
 971         if (newroute != NULL)
 972                 ksocknal_route_decref(newroute);
 973 }
 974
 975 int
 976 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
 977 {
 978         ksock_connreq_t *cr;
 979         int              rc;
 980         __u32            peer_ip;
 981         int              peer_port;
 982
 983         rc = lnet_sock_getaddr(sock, true, &peer_ip, &peer_port);
 984         LASSERT(rc == 0);               /* we succeeded before */
 985
 986         LIBCFS_ALLOC(cr, sizeof(*cr));
 987         if (cr == NULL) {
 988                 LCONSOLE_ERROR_MSG(0x12f, "Dropping connection request from "
 989                                    "%pI4h: memory exhausted\n", &peer_ip);
 990                 return -ENOMEM;
 991         }
 992
 993         lnet_ni_addref(ni);
 994         cr->ksncr_ni   = ni;
 995         cr->ksncr_sock = sock;
 996
 997         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
 998
 999         list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
1000         wake_up(&ksocknal_data.ksnd_connd_waitq);
1001
1002         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
1003         return 0;
1004 }
1005
1006 static int
1007 ksocknal_connecting (ksock_peer_ni_t *peer_ni, __u32 ipaddr)
1008 {
1009         ksock_route_t *route;
1010
1011         list_for_each_entry(route, &peer_ni->ksnp_routes, ksnr_list) {
1012                 if (route->ksnr_ipaddr == ipaddr)
1013                         return route->ksnr_connecting;
1014         }
1015         return 0;
1016 }
1017
1018 int
1019 ksocknal_create_conn(struct lnet_ni *ni, ksock_route_t *route,
1020                      struct socket *sock, int type)
1021 {
1022         rwlock_t                *global_lock = &ksocknal_data.ksnd_global_lock;
1023         struct list_head        zombies = LIST_HEAD_INIT(zombies);
1024         struct lnet_process_id peerid;
1025         struct list_head        *tmp;
1026         __u64              incarnation;
1027         ksock_conn_t      *conn;
1028         ksock_conn_t      *conn2;
1029         ksock_peer_ni_t      *peer_ni = NULL;
1030         ksock_peer_ni_t      *peer2;
1031         ksock_sched_t     *sched;
1032         struct ksock_hello_msg *hello;
1033         int                cpt;
1034         ksock_tx_t        *tx;
1035         ksock_tx_t        *txtmp;
1036         int                rc;
1037         int                rc2;
1038         int                active;
1039         char              *warn = NULL;
1040
1041         active = (route != NULL);
1042
1043         LASSERT (active == (type != SOCKLND_CONN_NONE));
1044
1045         LIBCFS_ALLOC(conn, sizeof(*conn));
1046         if (conn == NULL) {
1047                 rc = -ENOMEM;
1048                 goto failed_0;
1049         }
1050
1051         conn->ksnc_peer = NULL;
1052         conn->ksnc_route = NULL;
1053         conn->ksnc_sock = sock;
1054         /* 2 ref, 1 for conn, another extra ref prevents socket
1055          * being closed before establishment of connection */
1056         atomic_set (&conn->ksnc_sock_refcount, 2);
1057         conn->ksnc_type = type;
1058         ksocknal_lib_save_callback(sock, conn);
1059         atomic_set (&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
1060
1061         conn->ksnc_rx_ready = 0;
1062         conn->ksnc_rx_scheduled = 0;
1063
1064         INIT_LIST_HEAD(&conn->ksnc_tx_queue);
1065         conn->ksnc_tx_ready = 0;
1066         conn->ksnc_tx_scheduled = 0;
1067         conn->ksnc_tx_carrier = NULL;
1068         atomic_set (&conn->ksnc_tx_nob, 0);
1069
1070         LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
1071                                      kshm_ips[LNET_NUM_INTERFACES]));
1072         if (hello == NULL) {
1073                 rc = -ENOMEM;
1074                 goto failed_1;
1075         }
1076
1077         /* stash conn's local and remote addrs */
1078         rc = ksocknal_lib_get_conn_addrs (conn);
1079         if (rc != 0)
1080                 goto failed_1;
1081
1082         /* Find out/confirm peer_ni's NID and connection type and get the
1083          * vector of interfaces she's willing to let me connect to.
1084          * Passive connections use the listener timeout since the peer_ni sends
1085          * eagerly */
1086
1087         if (active) {
1088                 peer_ni = route->ksnr_peer;
1089                 LASSERT(ni == peer_ni->ksnp_ni);
1090
1091                 /* Active connection sends HELLO eagerly */
1092                 hello->kshm_nips = ksocknal_local_ipvec(ni, hello->kshm_ips);
1093                 peerid = peer_ni->ksnp_id;
1094
1095                 write_lock_bh(global_lock);
1096                 conn->ksnc_proto = peer_ni->ksnp_proto;
1097                 write_unlock_bh(global_lock);
1098
1099                 if (conn->ksnc_proto == NULL) {
1100                          conn->ksnc_proto = &ksocknal_protocol_v3x;
1101 #if SOCKNAL_VERSION_DEBUG
1102                          if (*ksocknal_tunables.ksnd_protocol == 2)
1103                                  conn->ksnc_proto = &ksocknal_protocol_v2x;
1104                          else if (*ksocknal_tunables.ksnd_protocol == 1)
1105                                  conn->ksnc_proto = &ksocknal_protocol_v1x;
1106 #endif
1107                 }
1108
1109                 rc = ksocknal_send_hello (ni, conn, peerid.nid, hello);
1110                 if (rc != 0)
1111                         goto failed_1;
1112         } else {
1113                 peerid.nid = LNET_NID_ANY;
1114                 peerid.pid = LNET_PID_ANY;
1115
1116                 /* Passive, get protocol from peer_ni */
1117                 conn->ksnc_proto = NULL;
1118         }
1119
1120         rc = ksocknal_recv_hello (ni, conn, hello, &peerid, &incarnation);
1121         if (rc < 0)
1122                 goto failed_1;
1123
1124         LASSERT (rc == 0 || active);
1125         LASSERT (conn->ksnc_proto != NULL);
1126         LASSERT (peerid.nid != LNET_NID_ANY);
1127
1128         cpt = lnet_cpt_of_nid(peerid.nid, ni);
1129
1130         if (active) {
1131                 ksocknal_peer_addref(peer_ni);
1132                 write_lock_bh(global_lock);
1133         } else {
1134                 rc = ksocknal_create_peer(&peer_ni, ni, peerid);
1135                 if (rc != 0)
1136                         goto failed_1;
1137
1138                 write_lock_bh(global_lock);
1139
1140                 /* called with a ref on ni, so shutdown can't have started */
1141                 LASSERT (((ksock_net_t *) ni->ni_data)->ksnn_shutdown == 0);
1142
1143                 peer2 = ksocknal_find_peer_locked(ni, peerid);
1144                 if (peer2 == NULL) {
1145                         /* NB this puts an "empty" peer_ni in the peer_ni
1146                          * table (which takes my ref) */
1147                         list_add_tail(&peer_ni->ksnp_list,
1148                                       ksocknal_nid2peerlist(peerid.nid));
1149                 } else {
1150                         ksocknal_peer_decref(peer_ni);
1151                         peer_ni = peer2;
1152                 }
1153
1154                 /* +1 ref for me */
1155                 ksocknal_peer_addref(peer_ni);
1156                 peer_ni->ksnp_accepting++;
1157
1158                 /* Am I already connecting to this guy?  Resolve in
1159                  * favour of higher NID... */
1160                 if (peerid.nid < ni->ni_nid &&
1161                     ksocknal_connecting(peer_ni, conn->ksnc_ipaddr)) {
1162                         rc = EALREADY;
1163                         warn = "connection race resolution";
1164                         goto failed_2;
1165                 }
1166         }
1167
1168         if (peer_ni->ksnp_closing ||
1169             (active && route->ksnr_deleted)) {
1170                 /* peer_ni/route got closed under me */
1171                 rc = -ESTALE;
1172                 warn = "peer_ni/route removed";
1173                 goto failed_2;
1174         }
1175
1176         if (peer_ni->ksnp_proto == NULL) {
1177                 /* Never connected before.
1178                  * NB recv_hello may have returned EPROTO to signal my peer_ni
1179                  * wants a different protocol than the one I asked for.
1180                  */
1181                 LASSERT(list_empty(&peer_ni->ksnp_conns));
1182
1183                 peer_ni->ksnp_proto = conn->ksnc_proto;
1184                 peer_ni->ksnp_incarnation = incarnation;
1185         }
1186
1187         if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1188             peer_ni->ksnp_incarnation != incarnation) {
1189                 /* peer_ni rebooted or I've got the wrong protocol version */
1190                 ksocknal_close_peer_conns_locked(peer_ni, 0, 0);
1191
1192                 peer_ni->ksnp_proto = NULL;
1193                 rc = ESTALE;
1194                 warn = peer_ni->ksnp_incarnation != incarnation ?
1195                        "peer_ni rebooted" :
1196                        "wrong proto version";
1197                 goto failed_2;
1198         }
1199
1200         switch (rc) {
1201         default:
1202                 LBUG();
1203         case 0:
1204                 break;
1205         case EALREADY:
1206                 warn = "lost conn race";
1207                 goto failed_2;
1208         case EPROTO:
1209                 warn = "retry with different protocol version";
1210                 goto failed_2;
1211         }
1212
1213         /* Refuse to duplicate an existing connection, unless this is a
1214          * loopback connection */
1215         if (conn->ksnc_ipaddr != conn->ksnc_myipaddr) {
1216                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1217                         conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1218
1219                         if (conn2->ksnc_ipaddr != conn->ksnc_ipaddr ||
1220                             conn2->ksnc_myipaddr != conn->ksnc_myipaddr ||
1221                             conn2->ksnc_type != conn->ksnc_type)
1222                                 continue;
1223
1224                         /* Reply on a passive connection attempt so the peer_ni
1225                          * realises we're connected. */
1226                         LASSERT (rc == 0);
1227                         if (!active)
1228                                 rc = EALREADY;
1229
1230                         warn = "duplicate";
1231                         goto failed_2;
1232                 }
1233         }
1234
1235         /* If the connection created by this route didn't bind to the IP
1236          * address the route connected to, the connection/route matching
1237          * code below probably isn't going to work. */
1238         if (active &&
1239             route->ksnr_ipaddr != conn->ksnc_ipaddr) {
1240                 CERROR("Route %s %pI4h connected to %pI4h\n",
1241                        libcfs_id2str(peer_ni->ksnp_id),
1242                        &route->ksnr_ipaddr,
1243                        &conn->ksnc_ipaddr);
1244         }
1245
1246         /* Search for a route corresponding to the new connection and
1247          * create an association.  This allows incoming connections created
1248          * by routes in my peer_ni to match my own route entries so I don't
1249          * continually create duplicate routes. */
1250         list_for_each(tmp, &peer_ni->ksnp_routes) {
1251                 route = list_entry(tmp, ksock_route_t, ksnr_list);
1252
1253                 if (route->ksnr_ipaddr != conn->ksnc_ipaddr)
1254                         continue;
1255
1256                 ksocknal_associate_route_conn_locked(route, conn);
1257                 break;
1258         }
1259
1260         conn->ksnc_peer = peer_ni;                 /* conn takes my ref on peer_ni */
1261         peer_ni->ksnp_last_alive = ktime_get_real_seconds();
1262         peer_ni->ksnp_send_keepalive = 0;
1263         peer_ni->ksnp_error = 0;
1264
1265         sched = ksocknal_choose_scheduler_locked(cpt);
1266         sched->kss_nconns++;
1267         conn->ksnc_scheduler = sched;
1268
1269         conn->ksnc_tx_last_post = ktime_get_real_seconds();
1270         /* Set the deadline for the outgoing HELLO to drain */
1271         conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1272         conn->ksnc_tx_deadline = cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1273         smp_mb();   /* order with adding to peer_ni's conn list */
1274
1275         list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1276         ksocknal_conn_addref(conn);
1277
1278         ksocknal_new_packet(conn, 0);
1279
1280         conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1281
1282         /* Take packets blocking for this connection. */
1283         list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1284                 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1285                     SOCKNAL_MATCH_NO)
1286                         continue;
1287
1288                 list_del(&tx->tx_list);
1289                 ksocknal_queue_tx_locked(tx, conn);
1290         }
1291
1292         write_unlock_bh(global_lock);
1293
1294         /* We've now got a new connection.  Any errors from here on are just
1295          * like "normal" comms errors and we close the connection normally.
1296          * NB (a) we still have to send the reply HELLO for passive
1297          *        connections,
1298          *    (b) normal I/O on the conn is blocked until I setup and call the
1299          *        socket callbacks.
1300          */
1301
1302         CDEBUG(D_NET, "New conn %s p %d.x %pI4h -> %pI4h/%d"
1303                " incarnation:%lld sched[%d:%d]\n",
1304                libcfs_id2str(peerid), conn->ksnc_proto->pro_version,
1305                &conn->ksnc_myipaddr, &conn->ksnc_ipaddr,
1306                conn->ksnc_port, incarnation, cpt,
1307                (int)(sched - &sched->kss_info->ksi_scheds[0]));
1308
1309         if (active) {
1310                 /* additional routes after interface exchange? */
1311                 ksocknal_create_routes(peer_ni, conn->ksnc_port,
1312                                        hello->kshm_ips, hello->kshm_nips);
1313         } else {
1314                 hello->kshm_nips = ksocknal_select_ips(peer_ni, hello->kshm_ips,
1315                                                        hello->kshm_nips);
1316                 rc = ksocknal_send_hello(ni, conn, peerid.nid, hello);
1317         }
1318
1319         LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1320                                     kshm_ips[LNET_NUM_INTERFACES]));
1321
1322         /* setup the socket AFTER I've received hello (it disables
1323          * SO_LINGER).  I might call back to the acceptor who may want
1324          * to send a protocol version response and then close the
1325          * socket; this ensures the socket only tears down after the
1326          * response has been sent. */
1327         if (rc == 0)
1328                 rc = ksocknal_lib_setup_sock(sock);
1329
1330         write_lock_bh(global_lock);
1331
1332         /* NB my callbacks block while I hold ksnd_global_lock */
1333         ksocknal_lib_set_callback(sock, conn);
1334
1335         if (!active)
1336                 peer_ni->ksnp_accepting--;
1337
1338         write_unlock_bh(global_lock);
1339
1340         if (rc != 0) {
1341                 write_lock_bh(global_lock);
1342                 if (!conn->ksnc_closing) {
1343                         /* could be closed by another thread */
1344                         ksocknal_close_conn_locked(conn, rc);
1345                 }
1346                 write_unlock_bh(global_lock);
1347         } else if (ksocknal_connsock_addref(conn) == 0) {
1348                 /* Allow I/O to proceed. */
1349                 ksocknal_read_callback(conn);
1350                 ksocknal_write_callback(conn);
1351                 ksocknal_connsock_decref(conn);
1352         }
1353
1354         ksocknal_connsock_decref(conn);
1355         ksocknal_conn_decref(conn);
1356         return rc;
1357
1358 failed_2:
1359         if (!peer_ni->ksnp_closing &&
1360             list_empty(&peer_ni->ksnp_conns) &&
1361             list_empty(&peer_ni->ksnp_routes)) {
1362                 list_add(&zombies, &peer_ni->ksnp_tx_queue);
1363                 list_del_init(&peer_ni->ksnp_tx_queue);
1364                 ksocknal_unlink_peer_locked(peer_ni);
1365         }
1366
1367         write_unlock_bh(global_lock);
1368
1369         if (warn != NULL) {
1370                 if (rc < 0)
1371                         CERROR("Not creating conn %s type %d: %s\n",
1372                                libcfs_id2str(peerid), conn->ksnc_type, warn);
1373                 else
1374                         CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1375                               libcfs_id2str(peerid), conn->ksnc_type, warn);
1376         }
1377
1378         if (!active) {
1379                 if (rc > 0) {
1380                         /* Request retry by replying with CONN_NONE
1381                          * ksnc_proto has been set already */
1382                         conn->ksnc_type = SOCKLND_CONN_NONE;
1383                         hello->kshm_nips = 0;
1384                         ksocknal_send_hello(ni, conn, peerid.nid, hello);
1385                 }
1386
1387                 write_lock_bh(global_lock);
1388                 peer_ni->ksnp_accepting--;
1389                 write_unlock_bh(global_lock);
1390         }
1391
1392         /*
1393          * If we get here without an error code, just use -EALREADY.
1394          * Depending on how we got here, the error may be positive
1395          * or negative. Normalize the value for ksocknal_txlist_done().
1396          */
1397         rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1398         ksocknal_txlist_done(ni, &zombies, rc2);
1399         ksocknal_peer_decref(peer_ni);
1400
1401 failed_1:
1402         if (hello != NULL)
1403                 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1404                                             kshm_ips[LNET_NUM_INTERFACES]));
1405
1406         LIBCFS_FREE(conn, sizeof(*conn));
1407
1408 failed_0:
1409         sock_release(sock);
1410         return rc;
1411 }
1412
1413 void
1414 ksocknal_close_conn_locked (ksock_conn_t *conn, int error)
1415 {
1416         /* This just does the immmediate housekeeping, and queues the
1417          * connection for the reaper to terminate.
1418          * Caller holds ksnd_global_lock exclusively in irq context */
1419         ksock_peer_ni_t      *peer_ni = conn->ksnc_peer;
1420         ksock_route_t     *route;
1421         ksock_conn_t      *conn2;
1422         struct list_head  *tmp;
1423
1424         LASSERT(peer_ni->ksnp_error == 0);
1425         LASSERT(!conn->ksnc_closing);
1426         conn->ksnc_closing = 1;
1427
1428         /* ksnd_deathrow_conns takes over peer_ni's ref */
1429         list_del(&conn->ksnc_list);
1430
1431         route = conn->ksnc_route;
1432         if (route != NULL) {
1433                 /* dissociate conn from route... */
1434                 LASSERT(!route->ksnr_deleted);
1435                 LASSERT((route->ksnr_connected & (1 << conn->ksnc_type)) != 0);
1436
1437                 conn2 = NULL;
1438                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1439                         conn2 = list_entry(tmp, ksock_conn_t, ksnc_list);
1440
1441                         if (conn2->ksnc_route == route &&
1442                             conn2->ksnc_type == conn->ksnc_type)
1443                                 break;
1444
1445                         conn2 = NULL;
1446                 }
1447                 if (conn2 == NULL)
1448                         route->ksnr_connected &= ~(1 << conn->ksnc_type);
1449
1450                 conn->ksnc_route = NULL;
1451
1452                 ksocknal_route_decref(route);   /* drop conn's ref on route */
1453         }
1454
1455         if (list_empty(&peer_ni->ksnp_conns)) {
1456                 /* No more connections to this peer_ni */
1457
1458                 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1459                                 ksock_tx_t *tx;
1460
1461                         LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1462
1463                         /* throw them to the last connection...,
1464                          * these TXs will be send to /dev/null by scheduler */
1465                         list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1466                                             tx_list)
1467                                 ksocknal_tx_prep(conn, tx);
1468
1469                         spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1470                         list_splice_init(&peer_ni->ksnp_tx_queue,
1471                                          &conn->ksnc_tx_queue);
1472                         spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1473                 }
1474
1475                 /* renegotiate protocol version */
1476                 peer_ni->ksnp_proto = NULL;
1477                 /* stash last conn close reason */
1478                 peer_ni->ksnp_error = error;
1479
1480                 if (list_empty(&peer_ni->ksnp_routes)) {
1481                         /* I've just closed last conn belonging to a
1482                          * peer_ni with no routes to it */
1483                         ksocknal_unlink_peer_locked(peer_ni);
1484                 }
1485         }
1486
1487         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1488
1489         list_add_tail(&conn->ksnc_list,
1490                       &ksocknal_data.ksnd_deathrow_conns);
1491         wake_up(&ksocknal_data.ksnd_reaper_waitq);
1492
1493         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1494 }
1495
1496 void
1497 ksocknal_peer_failed (ksock_peer_ni_t *peer_ni)
1498 {
1499         int        notify = 0;
1500         cfs_time_t last_alive = 0;
1501
1502         /* There has been a connection failure or comms error; but I'll only
1503          * tell LNET I think the peer_ni is dead if it's to another kernel and
1504          * there are no connections or connection attempts in existence. */
1505
1506         read_lock(&ksocknal_data.ksnd_global_lock);
1507
1508         if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1509              list_empty(&peer_ni->ksnp_conns) &&
1510              peer_ni->ksnp_accepting == 0 &&
1511              ksocknal_find_connecting_route_locked(peer_ni) == NULL) {
1512                 notify = 1;
1513                 last_alive = peer_ni->ksnp_last_alive;
1514         }
1515
1516         read_unlock(&ksocknal_data.ksnd_global_lock);
1517
1518         if (notify)
1519                 lnet_notify(peer_ni->ksnp_ni, peer_ni->ksnp_id.nid, 0,
1520                             last_alive);
1521 }
1522
1523 void
1524 ksocknal_finalize_zcreq(ksock_conn_t *conn)
1525 {
1526         ksock_peer_ni_t  *peer_ni = conn->ksnc_peer;
1527         ksock_tx_t       *tx;
1528         ksock_tx_t       *tmp;
1529         struct list_head  zlist = LIST_HEAD_INIT(zlist);
1530
1531         /* NB safe to finalize TXs because closing of socket will
1532          * abort all buffered data */
1533         LASSERT(conn->ksnc_sock == NULL);
1534
1535         spin_lock(&peer_ni->ksnp_lock);
1536
1537         list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list, tx_zc_list) {
1538                 if (tx->tx_conn != conn)
1539                         continue;
1540
1541                 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1542
1543                 tx->tx_msg.ksm_zc_cookies[0] = 0;
1544                 tx->tx_zc_aborted = 1;  /* mark it as not-acked */
1545                 list_del(&tx->tx_zc_list);
1546                 list_add(&tx->tx_zc_list, &zlist);
1547         }
1548
1549         spin_unlock(&peer_ni->ksnp_lock);
1550
1551         while (!list_empty(&zlist)) {
1552                 tx = list_entry(zlist.next, ksock_tx_t, tx_zc_list);
1553
1554                 list_del(&tx->tx_zc_list);
1555                 ksocknal_tx_decref(tx);
1556         }
1557 }
1558
1559 void
1560 ksocknal_terminate_conn(ksock_conn_t *conn)
1561 {
1562         /* This gets called by the reaper (guaranteed thread context) to
1563          * disengage the socket from its callbacks and close it.
1564          * ksnc_refcount will eventually hit zero, and then the reaper will
1565          * destroy it. */
1566         ksock_peer_ni_t     *peer_ni = conn->ksnc_peer;
1567         ksock_sched_t    *sched = conn->ksnc_scheduler;
1568         int               failed = 0;
1569
1570         LASSERT(conn->ksnc_closing);
1571
1572         /* wake up the scheduler to "send" all remaining packets to /dev/null */
1573         spin_lock_bh(&sched->kss_lock);
1574
1575         /* a closing conn is always ready to tx */
1576         conn->ksnc_tx_ready = 1;
1577
1578         if (!conn->ksnc_tx_scheduled &&
1579             !list_empty(&conn->ksnc_tx_queue)) {
1580                 list_add_tail(&conn->ksnc_tx_list,
1581                                &sched->kss_tx_conns);
1582                 conn->ksnc_tx_scheduled = 1;
1583                 /* extra ref for scheduler */
1584                 ksocknal_conn_addref(conn);
1585
1586                 wake_up (&sched->kss_waitq);
1587         }
1588
1589         spin_unlock_bh(&sched->kss_lock);
1590
1591         /* serialise with callbacks */
1592         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1593
1594         ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1595
1596         /* OK, so this conn may not be completely disengaged from its
1597          * scheduler yet, but it _has_ committed to terminate... */
1598         conn->ksnc_scheduler->kss_nconns--;
1599
1600         if (peer_ni->ksnp_error != 0) {
1601                 /* peer_ni's last conn closed in error */
1602                 LASSERT(list_empty(&peer_ni->ksnp_conns));
1603                 failed = 1;
1604                 peer_ni->ksnp_error = 0;     /* avoid multiple notifications */
1605         }
1606
1607         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1608
1609         if (failed)
1610                 ksocknal_peer_failed(peer_ni);
1611
1612         /* The socket is closed on the final put; either here, or in
1613          * ksocknal_{send,recv}msg().  Since we set up the linger2 option
1614          * when the connection was established, this will close the socket
1615          * immediately, aborting anything buffered in it. Any hung
1616          * zero-copy transmits will therefore complete in finite time. */
1617         ksocknal_connsock_decref(conn);
1618 }
1619
1620 void
1621 ksocknal_queue_zombie_conn (ksock_conn_t *conn)
1622 {
1623         /* Queue the conn for the reaper to destroy */
1624
1625         LASSERT(atomic_read(&conn->ksnc_conn_refcount) == 0);
1626         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1627
1628         list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1629         wake_up(&ksocknal_data.ksnd_reaper_waitq);
1630
1631         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1632 }
1633
1634 void
1635 ksocknal_destroy_conn (ksock_conn_t *conn)
1636 {
1637         cfs_time_t      last_rcv;
1638
1639         /* Final coup-de-grace of the reaper */
1640         CDEBUG (D_NET, "connection %p\n", conn);
1641
1642         LASSERT (atomic_read (&conn->ksnc_conn_refcount) == 0);
1643         LASSERT (atomic_read (&conn->ksnc_sock_refcount) == 0);
1644         LASSERT (conn->ksnc_sock == NULL);
1645         LASSERT (conn->ksnc_route == NULL);
1646         LASSERT (!conn->ksnc_tx_scheduled);
1647         LASSERT (!conn->ksnc_rx_scheduled);
1648         LASSERT(list_empty(&conn->ksnc_tx_queue));
1649
1650         /* complete current receive if any */
1651         switch (conn->ksnc_rx_state) {
1652         case SOCKNAL_RX_LNET_PAYLOAD:
1653                 last_rcv = conn->ksnc_rx_deadline -
1654                            cfs_time_seconds(*ksocknal_tunables.ksnd_timeout);
1655                 CERROR("Completing partial receive from %s[%d], "
1656                        "ip %pI4h:%d, with error, wanted: %d, left: %d, "
1657                        "last alive is %ld secs ago\n",
1658                        libcfs_id2str(conn->ksnc_peer->ksnp_id), conn->ksnc_type,
1659                        &conn->ksnc_ipaddr, conn->ksnc_port,
1660                        conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1661                        cfs_duration_sec(cfs_time_sub(ktime_get_real_seconds(),
1662                                         last_rcv)));
1663                 lnet_finalize(conn->ksnc_cookie, -EIO);
1664                 break;
1665         case SOCKNAL_RX_LNET_HEADER:
1666                 if (conn->ksnc_rx_started)
1667                         CERROR("Incomplete receive of lnet header from %s, "
1668                                "ip %pI4h:%d, with error, protocol: %d.x.\n",
1669                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1670                                &conn->ksnc_ipaddr, conn->ksnc_port,
1671                                conn->ksnc_proto->pro_version);
1672                 break;
1673         case SOCKNAL_RX_KSM_HEADER:
1674                 if (conn->ksnc_rx_started)
1675                         CERROR("Incomplete receive of ksock message from %s, "
1676                                "ip %pI4h:%d, with error, protocol: %d.x.\n",
1677                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1678                                &conn->ksnc_ipaddr, conn->ksnc_port,
1679                                conn->ksnc_proto->pro_version);
1680                 break;
1681         case SOCKNAL_RX_SLOP:
1682                 if (conn->ksnc_rx_started)
1683                         CERROR("Incomplete receive of slops from %s, "
1684                                "ip %pI4h:%d, with error\n",
1685                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1686                                &conn->ksnc_ipaddr, conn->ksnc_port);
1687                break;
1688         default:
1689                 LBUG ();
1690                 break;
1691         }
1692
1693         ksocknal_peer_decref(conn->ksnc_peer);
1694
1695         LIBCFS_FREE (conn, sizeof (*conn));
1696 }
1697
1698 int
1699 ksocknal_close_peer_conns_locked (ksock_peer_ni_t *peer_ni, __u32 ipaddr, int why)
1700 {
1701         ksock_conn_t       *conn;
1702         struct list_head         *ctmp;
1703         struct list_head         *cnxt;
1704         int                 count = 0;
1705
1706         list_for_each_safe(ctmp, cnxt, &peer_ni->ksnp_conns) {
1707                 conn = list_entry(ctmp, ksock_conn_t, ksnc_list);
1708
1709                 if (ipaddr == 0 ||
1710                     conn->ksnc_ipaddr == ipaddr) {
1711                         count++;
1712                         ksocknal_close_conn_locked (conn, why);
1713                 }
1714         }
1715
1716         return (count);
1717 }
1718
1719 int
1720 ksocknal_close_conn_and_siblings (ksock_conn_t *conn, int why)
1721 {
1722         ksock_peer_ni_t     *peer_ni = conn->ksnc_peer;
1723         __u32             ipaddr = conn->ksnc_ipaddr;
1724         int               count;
1725
1726         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1727
1728         count = ksocknal_close_peer_conns_locked (peer_ni, ipaddr, why);
1729
1730         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1731
1732         return (count);
1733 }
1734
1735 int
1736 ksocknal_close_matching_conns(struct lnet_process_id id, __u32 ipaddr)
1737 {
1738         ksock_peer_ni_t       *peer_ni;
1739         struct list_head         *ptmp;
1740         struct list_head         *pnxt;
1741         int                 lo;
1742         int                 hi;
1743         int                 i;
1744         int                 count = 0;
1745
1746         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1747
1748         if (id.nid != LNET_NID_ANY)
1749                 lo = hi = (int)(ksocknal_nid2peerlist(id.nid) - ksocknal_data.ksnd_peers);
1750         else {
1751                 lo = 0;
1752                 hi = ksocknal_data.ksnd_peer_hash_size - 1;
1753         }
1754
1755         for (i = lo; i <= hi; i++) {
1756                 list_for_each_safe(ptmp, pnxt, &ksocknal_data.ksnd_peers[i]) {
1757
1758                         peer_ni = list_entry(ptmp, ksock_peer_ni_t, ksnp_list);
1759
1760                         if (!((id.nid == LNET_NID_ANY || id.nid == peer_ni->ksnp_id.nid) &&
1761                               (id.pid == LNET_PID_ANY || id.pid == peer_ni->ksnp_id.pid)))
1762                                 continue;
1763
1764                         count += ksocknal_close_peer_conns_locked (peer_ni, ipaddr, 0);
1765                 }
1766         }
1767
1768         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1769
1770         /* wildcards always succeed */
1771         if (id.nid == LNET_NID_ANY || id.pid == LNET_PID_ANY || ipaddr == 0)
1772                 return (0);
1773
1774         return (count == 0 ? -ENOENT : 0);
1775 }
1776
1777 void
1778 ksocknal_notify(struct lnet_ni *ni, lnet_nid_t gw_nid, int alive)
1779 {
1780         /* The router is telling me she's been notified of a change in
1781          * gateway state....
1782          */
1783         struct lnet_process_id id = {
1784                 .nid    = gw_nid,
1785                 .pid    = LNET_PID_ANY,
1786         };
1787
1788         CDEBUG (D_NET, "gw %s %s\n", libcfs_nid2str(gw_nid),
1789                 alive ? "up" : "down");
1790
1791         if (!alive) {
1792                 /* If the gateway crashed, close all open connections... */
1793                 ksocknal_close_matching_conns (id, 0);
1794                 return;
1795         }
1796
1797         /* ...otherwise do nothing.  We can only establish new connections
1798          * if we have autroutes, and these connect on demand. */
1799 }
1800
1801 void
1802 ksocknal_query(struct lnet_ni *ni, lnet_nid_t nid, cfs_time_t *when)
1803 {
1804         int connect = 1;
1805         time64_t last_alive = 0;
1806         time64_t now = ktime_get_real_seconds();
1807         ksock_peer_ni_t *peer_ni = NULL;
1808         rwlock_t *glock = &ksocknal_data.ksnd_global_lock;
1809         struct lnet_process_id id = {
1810                 .nid = nid,
1811                 .pid = LNET_PID_LUSTRE,
1812         };
1813
1814         read_lock(glock);
1815
1816         peer_ni = ksocknal_find_peer_locked(ni, id);
1817         if (peer_ni != NULL) {
1818                 struct list_head       *tmp;
1819                 ksock_conn_t     *conn;
1820                 int               bufnob;
1821
1822                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1823                         conn = list_entry(tmp, ksock_conn_t, ksnc_list);
1824                         bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
1825
1826                         if (bufnob < conn->ksnc_tx_bufnob) {
1827                                 /* something got ACKed */
1828                                 conn->ksnc_tx_deadline =
1829                                         cfs_time_shift(*ksocknal_tunables.ksnd_timeout);
1830                                 peer_ni->ksnp_last_alive = now;
1831                                 conn->ksnc_tx_bufnob = bufnob;
1832                         }
1833                 }
1834
1835                 last_alive = peer_ni->ksnp_last_alive;
1836                 if (ksocknal_find_connectable_route_locked(peer_ni) == NULL)
1837                         connect = 0;
1838         }
1839
1840         read_unlock(glock);
1841
1842         if (last_alive != 0)
1843                 *when = last_alive;
1844
1845         CDEBUG(D_NET, "peer_ni %s %p, alive %ld secs ago, connect %d\n",
1846                libcfs_nid2str(nid), peer_ni,
1847                last_alive ? cfs_duration_sec(now - last_alive) : -1,
1848                connect);
1849
1850         if (!connect)
1851                 return;
1852
1853         ksocknal_add_peer(ni, id, LNET_NIDADDR(nid), lnet_acceptor_port());
1854
1855         write_lock_bh(glock);
1856
1857         peer_ni = ksocknal_find_peer_locked(ni, id);
1858         if (peer_ni != NULL)
1859                 ksocknal_launch_all_connections_locked(peer_ni);
1860
1861         write_unlock_bh(glock);
1862         return;
1863 }
1864
1865 static void
1866 ksocknal_push_peer (ksock_peer_ni_t *peer_ni)
1867 {
1868         int               index;
1869         int               i;
1870         struct list_head       *tmp;
1871         ksock_conn_t     *conn;
1872
1873         for (index = 0; ; index++) {
1874                 read_lock(&ksocknal_data.ksnd_global_lock);
1875
1876                 i = 0;
1877                 conn = NULL;
1878
1879                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1880                         if (i++ == index) {
1881                                 conn = list_entry(tmp, ksock_conn_t,
1882                                                        ksnc_list);
1883                                 ksocknal_conn_addref(conn);
1884                                 break;
1885                         }
1886                 }
1887
1888                 read_unlock(&ksocknal_data.ksnd_global_lock);
1889
1890                 if (conn == NULL)
1891                         break;
1892
1893                 ksocknal_lib_push_conn (conn);
1894                 ksocknal_conn_decref(conn);
1895         }
1896 }
1897
1898 static int
1899 ksocknal_push(struct lnet_ni *ni, struct lnet_process_id id)
1900 {
1901         struct list_head *start;
1902         struct list_head *end;
1903         struct list_head *tmp;
1904         int               rc = -ENOENT;
1905         unsigned int      hsize = ksocknal_data.ksnd_peer_hash_size;
1906
1907         if (id.nid == LNET_NID_ANY) {
1908                 start = &ksocknal_data.ksnd_peers[0];
1909                 end = &ksocknal_data.ksnd_peers[hsize - 1];
1910         } else {
1911                 start = end = ksocknal_nid2peerlist(id.nid);
1912         }
1913
1914         for (tmp = start; tmp <= end; tmp++) {
1915                 int     peer_off; /* searching offset in peer_ni hash table */
1916
1917                 for (peer_off = 0; ; peer_off++) {
1918                         ksock_peer_ni_t *peer_ni;
1919                         int           i = 0;
1920
1921                         read_lock(&ksocknal_data.ksnd_global_lock);
1922                         list_for_each_entry(peer_ni, tmp, ksnp_list) {
1923                                 if (!((id.nid == LNET_NID_ANY ||
1924                                        id.nid == peer_ni->ksnp_id.nid) &&
1925                                       (id.pid == LNET_PID_ANY ||
1926                                        id.pid == peer_ni->ksnp_id.pid)))
1927                                         continue;
1928
1929                                 if (i++ == peer_off) {
1930                                         ksocknal_peer_addref(peer_ni);
1931                                         break;
1932                                 }
1933                         }
1934                         read_unlock(&ksocknal_data.ksnd_global_lock);
1935
1936                         if (i == 0) /* no match */
1937                                 break;
1938
1939                         rc = 0;
1940                         ksocknal_push_peer(peer_ni);
1941                         ksocknal_peer_decref(peer_ni);
1942                 }
1943         }
1944         return rc;
1945 }
1946
1947 static int
1948 ksocknal_add_interface(struct lnet_ni *ni, __u32 ipaddress, __u32 netmask)
1949 {
1950         ksock_net_t *net = ni->ni_data;
1951         ksock_interface_t *iface;
1952         int rc;
1953         int i;
1954         int j;
1955         struct list_head *ptmp;
1956         ksock_peer_ni_t *peer_ni;
1957         struct list_head *rtmp;
1958         ksock_route_t *route;
1959
1960         if (ipaddress == 0 ||
1961             netmask == 0)
1962                 return -EINVAL;
1963
1964         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1965
1966         iface = ksocknal_ip2iface(ni, ipaddress);
1967         if (iface != NULL) {
1968                 /* silently ignore dups */
1969                 rc = 0;
1970         } else if (net->ksnn_ninterfaces == LNET_NUM_INTERFACES) {
1971                 rc = -ENOSPC;
1972         } else {
1973                 iface = &net->ksnn_interfaces[net->ksnn_ninterfaces++];
1974
1975                 iface->ksni_ipaddr = ipaddress;
1976                 iface->ksni_netmask = netmask;
1977                 iface->ksni_nroutes = 0;
1978                 iface->ksni_npeers = 0;
1979
1980                 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1981                         list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
1982                                 peer_ni = list_entry(ptmp, ksock_peer_ni_t,
1983                                                      ksnp_list);
1984
1985                                 for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++)
1986                                         if (peer_ni->ksnp_passive_ips[j] == ipaddress)
1987                                                 iface->ksni_npeers++;
1988
1989                                 list_for_each(rtmp, &peer_ni->ksnp_routes) {
1990                                         route = list_entry(rtmp,
1991                                                            ksock_route_t,
1992                                                            ksnr_list);
1993
1994                                         if (route->ksnr_myipaddr == ipaddress)
1995                                                 iface->ksni_nroutes++;
1996                                 }
1997                         }
1998                 }
1999
2000                 rc = 0;
2001                 /* NB only new connections will pay attention to the new interface! */
2002         }
2003
2004         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2005
2006         return rc;
2007 }
2008
2009 static void
2010 ksocknal_peer_del_interface_locked(ksock_peer_ni_t *peer_ni, __u32 ipaddr)
2011 {
2012         struct list_head         *tmp;
2013         struct list_head         *nxt;
2014         ksock_route_t      *route;
2015         ksock_conn_t       *conn;
2016         int                 i;
2017         int                 j;
2018
2019         for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++)
2020                 if (peer_ni->ksnp_passive_ips[i] == ipaddr) {
2021                         for (j = i+1; j < peer_ni->ksnp_n_passive_ips; j++)
2022                                 peer_ni->ksnp_passive_ips[j-1] =
2023                                         peer_ni->ksnp_passive_ips[j];
2024                         peer_ni->ksnp_n_passive_ips--;
2025                         break;
2026                 }
2027
2028         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
2029                 route = list_entry(tmp, ksock_route_t, ksnr_list);
2030
2031                 if (route->ksnr_myipaddr != ipaddr)
2032                         continue;
2033
2034                 if (route->ksnr_share_count != 0) {
2035                         /* Manually created; keep, but unbind */
2036                         route->ksnr_myipaddr = 0;
2037                 } else {
2038                         ksocknal_del_route_locked(route);
2039                 }
2040         }
2041
2042         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_conns) {
2043                 conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2044
2045                 if (conn->ksnc_myipaddr == ipaddr)
2046                         ksocknal_close_conn_locked (conn, 0);
2047         }
2048 }
2049
2050 static int
2051 ksocknal_del_interface(struct lnet_ni *ni, __u32 ipaddress)
2052 {
2053         ksock_net_t       *net = ni->ni_data;
2054         int                rc = -ENOENT;
2055         struct list_head        *tmp;
2056         struct list_head        *nxt;
2057         ksock_peer_ni_t      *peer_ni;
2058         __u32              this_ip;
2059         int                i;
2060         int                j;
2061
2062         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2063
2064         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2065                 this_ip = net->ksnn_interfaces[i].ksni_ipaddr;
2066
2067                 if (!(ipaddress == 0 ||
2068                       ipaddress == this_ip))
2069                         continue;
2070
2071                 rc = 0;
2072
2073                 for (j = i+1; j < net->ksnn_ninterfaces; j++)
2074                         net->ksnn_interfaces[j-1] =
2075                                 net->ksnn_interfaces[j];
2076
2077                 net->ksnn_ninterfaces--;
2078
2079                 for (j = 0; j < ksocknal_data.ksnd_peer_hash_size; j++) {
2080                         list_for_each_safe(tmp, nxt,
2081                                                &ksocknal_data.ksnd_peers[j]) {
2082                                 peer_ni = list_entry(tmp, ksock_peer_ni_t,
2083                                                       ksnp_list);
2084
2085                                 if (peer_ni->ksnp_ni != ni)
2086                                         continue;
2087
2088                                 ksocknal_peer_del_interface_locked(peer_ni, this_ip);
2089                         }
2090                 }
2091         }
2092
2093         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2094
2095         return (rc);
2096 }
2097
2098 int
2099 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
2100 {
2101         struct lnet_process_id id = {0};
2102         struct libcfs_ioctl_data *data = arg;
2103         int rc;
2104
2105         switch(cmd) {
2106         case IOC_LIBCFS_GET_INTERFACE: {
2107                 ksock_net_t       *net = ni->ni_data;
2108                 ksock_interface_t *iface;
2109
2110                 read_lock(&ksocknal_data.ksnd_global_lock);
2111
2112                 if (data->ioc_count >= (__u32)net->ksnn_ninterfaces) {
2113                         rc = -ENOENT;
2114                 } else {
2115                         rc = 0;
2116                         iface = &net->ksnn_interfaces[data->ioc_count];
2117
2118                         data->ioc_u32[0] = iface->ksni_ipaddr;
2119                         data->ioc_u32[1] = iface->ksni_netmask;
2120                         data->ioc_u32[2] = iface->ksni_npeers;
2121                         data->ioc_u32[3] = iface->ksni_nroutes;
2122                 }
2123
2124                 read_unlock(&ksocknal_data.ksnd_global_lock);
2125                 return rc;
2126         }
2127
2128         case IOC_LIBCFS_ADD_INTERFACE:
2129                 return ksocknal_add_interface(ni,
2130                                               data->ioc_u32[0], /* IP address */
2131                                               data->ioc_u32[1]); /* net mask */
2132
2133         case IOC_LIBCFS_DEL_INTERFACE:
2134                 return ksocknal_del_interface(ni,
2135                                               data->ioc_u32[0]); /* IP address */
2136
2137         case IOC_LIBCFS_GET_PEER: {
2138                 __u32            myip = 0;
2139                 __u32            ip = 0;
2140                 int              port = 0;
2141                 int              conn_count = 0;
2142                 int              share_count = 0;
2143
2144                 rc = ksocknal_get_peer_info(ni, data->ioc_count,
2145                                             &id, &myip, &ip, &port,
2146                                             &conn_count,  &share_count);
2147                 if (rc != 0)
2148                         return rc;
2149
2150                 data->ioc_nid    = id.nid;
2151                 data->ioc_count  = share_count;
2152                 data->ioc_u32[0] = ip;
2153                 data->ioc_u32[1] = port;
2154                 data->ioc_u32[2] = myip;
2155                 data->ioc_u32[3] = conn_count;
2156                 data->ioc_u32[4] = id.pid;
2157                 return 0;
2158         }
2159
2160         case IOC_LIBCFS_ADD_PEER:
2161                 id.nid = data->ioc_nid;
2162                 id.pid = LNET_PID_LUSTRE;
2163                 return ksocknal_add_peer (ni, id,
2164                                           data->ioc_u32[0], /* IP */
2165                                           data->ioc_u32[1]); /* port */
2166
2167         case IOC_LIBCFS_DEL_PEER:
2168                 id.nid = data->ioc_nid;
2169                 id.pid = LNET_PID_ANY;
2170                 return ksocknal_del_peer (ni, id,
2171                                           data->ioc_u32[0]); /* IP */
2172
2173         case IOC_LIBCFS_GET_CONN: {
2174                 int           txmem;
2175                 int           rxmem;
2176                 int           nagle;
2177                 ksock_conn_t *conn = ksocknal_get_conn_by_idx (ni, data->ioc_count);
2178
2179                 if (conn == NULL)
2180                         return -ENOENT;
2181
2182                 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
2183
2184                 data->ioc_count  = txmem;
2185                 data->ioc_nid    = conn->ksnc_peer->ksnp_id.nid;
2186                 data->ioc_flags  = nagle;
2187                 data->ioc_u32[0] = conn->ksnc_ipaddr;
2188                 data->ioc_u32[1] = conn->ksnc_port;
2189                 data->ioc_u32[2] = conn->ksnc_myipaddr;
2190                 data->ioc_u32[3] = conn->ksnc_type;
2191                 data->ioc_u32[4] = conn->ksnc_scheduler->kss_info->ksi_cpt;
2192                 data->ioc_u32[5] = rxmem;
2193                 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
2194                 ksocknal_conn_decref(conn);
2195                 return 0;
2196         }
2197
2198         case IOC_LIBCFS_CLOSE_CONNECTION:
2199                 id.nid = data->ioc_nid;
2200                 id.pid = LNET_PID_ANY;
2201                 return ksocknal_close_matching_conns (id,
2202                                                       data->ioc_u32[0]);
2203
2204         case IOC_LIBCFS_REGISTER_MYNID:
2205                 /* Ignore if this is a noop */
2206                 if (data->ioc_nid == ni->ni_nid)
2207                         return 0;
2208
2209                 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
2210                        libcfs_nid2str(data->ioc_nid),
2211                        libcfs_nid2str(ni->ni_nid));
2212                 return -EINVAL;
2213
2214         case IOC_LIBCFS_PUSH_CONNECTION:
2215                 id.nid = data->ioc_nid;
2216                 id.pid = LNET_PID_ANY;
2217                 return ksocknal_push(ni, id);
2218
2219         default:
2220                 return -EINVAL;
2221         }
2222         /* not reached */
2223 }
2224
2225 static void
2226 ksocknal_free_buffers (void)
2227 {
2228         LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
2229
2230         if (ksocknal_data.ksnd_sched_info != NULL) {
2231                 struct ksock_sched_info *info;
2232                 int                     i;
2233
2234                 cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2235                         if (info->ksi_scheds != NULL) {
2236                                 LIBCFS_FREE(info->ksi_scheds,
2237                                             info->ksi_nthreads_max *
2238                                             sizeof(info->ksi_scheds[0]));
2239                         }
2240                 }
2241                 cfs_percpt_free(ksocknal_data.ksnd_sched_info);
2242         }
2243
2244         LIBCFS_FREE (ksocknal_data.ksnd_peers,
2245                      sizeof(struct list_head) *
2246                      ksocknal_data.ksnd_peer_hash_size);
2247
2248         spin_lock(&ksocknal_data.ksnd_tx_lock);
2249
2250         if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
2251                 struct list_head        zlist;
2252                 ksock_tx_t      *tx;
2253
2254                 list_add(&zlist, &ksocknal_data.ksnd_idle_noop_txs);
2255                 list_del_init(&ksocknal_data.ksnd_idle_noop_txs);
2256                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2257
2258                 while (!list_empty(&zlist)) {
2259                         tx = list_entry(zlist.next, ksock_tx_t, tx_list);
2260                         list_del(&tx->tx_list);
2261                         LIBCFS_FREE(tx, tx->tx_desc_size);
2262                 }
2263         } else {
2264                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2265         }
2266 }
2267
2268 static void
2269 ksocknal_base_shutdown(void)
2270 {
2271         struct ksock_sched_info *info;
2272         ksock_sched_t           *sched;
2273         int                     i;
2274         int                     j;
2275
2276         CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2277                atomic_read (&libcfs_kmemory));
2278         LASSERT (ksocknal_data.ksnd_nnets == 0);
2279
2280         switch (ksocknal_data.ksnd_init) {
2281         default:
2282                 LASSERT (0);
2283
2284         case SOCKNAL_INIT_ALL:
2285         case SOCKNAL_INIT_DATA:
2286                 LASSERT (ksocknal_data.ksnd_peers != NULL);
2287                 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2288                         LASSERT(list_empty(&ksocknal_data.ksnd_peers[i]));
2289                 }
2290
2291                 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2292                 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2293                 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2294                 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2295                 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2296
2297                 if (ksocknal_data.ksnd_sched_info != NULL) {
2298                         cfs_percpt_for_each(info, i,
2299                                             ksocknal_data.ksnd_sched_info) {
2300                                 if (info->ksi_scheds == NULL)
2301                                         continue;
2302
2303                                 for (j = 0; j < info->ksi_nthreads_max; j++) {
2304
2305                                         sched = &info->ksi_scheds[j];
2306                                         LASSERT(list_empty(&sched->\
2307                                                                kss_tx_conns));
2308                                         LASSERT(list_empty(&sched->\
2309                                                                kss_rx_conns));
2310                                         LASSERT(list_empty(&sched-> \
2311                                                   kss_zombie_noop_txs));
2312                                         LASSERT(sched->kss_nconns == 0);
2313                                 }
2314                         }
2315                 }
2316
2317                 /* flag threads to terminate; wake and wait for them to die */
2318                 ksocknal_data.ksnd_shuttingdown = 1;
2319                 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2320                 wake_up_all(&ksocknal_data.ksnd_reaper_waitq);
2321
2322                 if (ksocknal_data.ksnd_sched_info != NULL) {
2323                         cfs_percpt_for_each(info, i,
2324                                             ksocknal_data.ksnd_sched_info) {
2325                                 if (info->ksi_scheds == NULL)
2326                                         continue;
2327
2328                                 for (j = 0; j < info->ksi_nthreads_max; j++) {
2329                                         sched = &info->ksi_scheds[j];
2330                                         wake_up_all(&sched->kss_waitq);
2331                                 }
2332                         }
2333                 }
2334
2335                 i = 4;
2336                 read_lock(&ksocknal_data.ksnd_global_lock);
2337                 while (ksocknal_data.ksnd_nthreads != 0) {
2338                         i++;
2339                         /* power of 2? */
2340                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2341                                 "waiting for %d threads to terminate\n",
2342                                 ksocknal_data.ksnd_nthreads);
2343                         read_unlock(&ksocknal_data.ksnd_global_lock);
2344                         set_current_state(TASK_UNINTERRUPTIBLE);
2345                         schedule_timeout(cfs_time_seconds(1));
2346                         read_lock(&ksocknal_data.ksnd_global_lock);
2347                 }
2348                 read_unlock(&ksocknal_data.ksnd_global_lock);
2349
2350                 ksocknal_free_buffers();
2351
2352                 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2353                 break;
2354         }
2355
2356         CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2357                atomic_read (&libcfs_kmemory));
2358
2359         module_put(THIS_MODULE);
2360 }
2361
2362 static int
2363 ksocknal_base_startup(void)
2364 {
2365         struct ksock_sched_info *info;
2366         int                     rc;
2367         int                     i;
2368
2369         LASSERT (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2370         LASSERT (ksocknal_data.ksnd_nnets == 0);
2371
2372         memset (&ksocknal_data, 0, sizeof (ksocknal_data)); /* zero pointers */
2373
2374         ksocknal_data.ksnd_peer_hash_size = SOCKNAL_PEER_HASH_SIZE;
2375         LIBCFS_ALLOC(ksocknal_data.ksnd_peers,
2376                      sizeof(struct list_head) *
2377                      ksocknal_data.ksnd_peer_hash_size);
2378         if (ksocknal_data.ksnd_peers == NULL)
2379                 return -ENOMEM;
2380
2381         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++)
2382                 INIT_LIST_HEAD(&ksocknal_data.ksnd_peers[i]);
2383
2384         rwlock_init(&ksocknal_data.ksnd_global_lock);
2385         INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2386
2387         spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2388         INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2389         INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2390         INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2391         init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2392
2393         spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2394         INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2395         INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2396         init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2397
2398         spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2399         INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2400
2401         /* NB memset above zeros whole of ksocknal_data */
2402
2403         /* flag lists/ptrs/locks initialised */
2404         ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2405         try_module_get(THIS_MODULE);
2406
2407         ksocknal_data.ksnd_sched_info = cfs_percpt_alloc(lnet_cpt_table(),
2408                                                          sizeof(*info));
2409         if (ksocknal_data.ksnd_sched_info == NULL)
2410                 goto failed;
2411
2412         cfs_percpt_for_each(info, i, ksocknal_data.ksnd_sched_info) {
2413                 ksock_sched_t   *sched;
2414                 int             nthrs;
2415
2416                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2417                 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2418                         nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2419                 } else {
2420                         /* max to half of CPUs, assume another half should be
2421                          * reserved for upper layer modules */
2422                         nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2423                 }
2424
2425                 info->ksi_nthreads_max = nthrs;
2426                 info->ksi_cpt = i;
2427
2428                 LIBCFS_CPT_ALLOC(info->ksi_scheds, lnet_cpt_table(), i,
2429                                  info->ksi_nthreads_max * sizeof(*sched));
2430                 if (info->ksi_scheds == NULL)
2431                         goto failed;
2432
2433                 for (; nthrs > 0; nthrs--) {
2434                         sched = &info->ksi_scheds[nthrs - 1];
2435
2436                         sched->kss_info = info;
2437                         spin_lock_init(&sched->kss_lock);
2438                         INIT_LIST_HEAD(&sched->kss_rx_conns);
2439                         INIT_LIST_HEAD(&sched->kss_tx_conns);
2440                         INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2441                         init_waitqueue_head(&sched->kss_waitq);
2442                 }
2443         }
2444
2445         ksocknal_data.ksnd_connd_starting         = 0;
2446         ksocknal_data.ksnd_connd_failed_stamp     = 0;
2447         ksocknal_data.ksnd_connd_starting_stamp   = ktime_get_real_seconds();
2448         /* must have at least 2 connds to remain responsive to accepts while
2449          * connecting */
2450         if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2451                 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2452
2453         if (*ksocknal_tunables.ksnd_nconnds_max <
2454             *ksocknal_tunables.ksnd_nconnds) {
2455                 ksocknal_tunables.ksnd_nconnds_max =
2456                         ksocknal_tunables.ksnd_nconnds;
2457         }
2458
2459         for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2460                 char name[16];
2461                 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2462                 ksocknal_data.ksnd_connd_starting++;
2463                 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2464
2465
2466                 snprintf(name, sizeof(name), "socknal_cd%02d", i);
2467                 rc = ksocknal_thread_start(ksocknal_connd,
2468                                            (void *)((uintptr_t)i), name);
2469                 if (rc != 0) {
2470                         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2471                         ksocknal_data.ksnd_connd_starting--;
2472                         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2473                         CERROR("Can't spawn socknal connd: %d\n", rc);
2474                         goto failed;
2475                 }
2476         }
2477
2478         rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2479         if (rc != 0) {
2480                 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2481                 goto failed;
2482         }
2483
2484         /* flag everything initialised */
2485         ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2486
2487         return 0;
2488
2489  failed:
2490         ksocknal_base_shutdown();
2491         return -ENETDOWN;
2492 }
2493
2494 static void
2495 ksocknal_debug_peerhash(struct lnet_ni *ni)
2496 {
2497         ksock_peer_ni_t *peer_ni = NULL;
2498         struct list_head        *tmp;
2499         int             i;
2500
2501         read_lock(&ksocknal_data.ksnd_global_lock);
2502
2503         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2504                 list_for_each(tmp, &ksocknal_data.ksnd_peers[i]) {
2505                         peer_ni = list_entry(tmp, ksock_peer_ni_t, ksnp_list);
2506
2507                         if (peer_ni->ksnp_ni == ni) break;
2508
2509                         peer_ni = NULL;
2510                 }
2511         }
2512
2513         if (peer_ni != NULL) {
2514                 ksock_route_t *route;
2515                 ksock_conn_t  *conn;
2516
2517                 CWARN ("Active peer_ni on shutdown: %s, ref %d, scnt %d, "
2518                        "closing %d, accepting %d, err %d, zcookie %llu, "
2519                        "txq %d, zc_req %d\n", libcfs_id2str(peer_ni->ksnp_id),
2520                        atomic_read(&peer_ni->ksnp_refcount),
2521                        peer_ni->ksnp_sharecount, peer_ni->ksnp_closing,
2522                        peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2523                        peer_ni->ksnp_zc_next_cookie,
2524                        !list_empty(&peer_ni->ksnp_tx_queue),
2525                        !list_empty(&peer_ni->ksnp_zc_req_list));
2526
2527                 list_for_each(tmp, &peer_ni->ksnp_routes) {
2528                         route = list_entry(tmp, ksock_route_t, ksnr_list);
2529                         CWARN ("Route: ref %d, schd %d, conn %d, cnted %d, "
2530                                "del %d\n", atomic_read(&route->ksnr_refcount),
2531                                route->ksnr_scheduled, route->ksnr_connecting,
2532                                route->ksnr_connected, route->ksnr_deleted);
2533                 }
2534
2535                 list_for_each(tmp, &peer_ni->ksnp_conns) {
2536                         conn = list_entry(tmp, ksock_conn_t, ksnc_list);
2537                         CWARN ("Conn: ref %d, sref %d, t %d, c %d\n",
2538                                atomic_read(&conn->ksnc_conn_refcount),
2539                                atomic_read(&conn->ksnc_sock_refcount),
2540                                conn->ksnc_type, conn->ksnc_closing);
2541                 }
2542         }
2543
2544         read_unlock(&ksocknal_data.ksnd_global_lock);
2545         return;
2546 }
2547
2548 void
2549 ksocknal_shutdown(struct lnet_ni *ni)
2550 {
2551         ksock_net_t *net = ni->ni_data;
2552         struct lnet_process_id anyid = {
2553                 .nid = LNET_NID_ANY,
2554                 .pid = LNET_PID_ANY,
2555         };
2556         int i;
2557
2558         LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2559         LASSERT(ksocknal_data.ksnd_nnets > 0);
2560
2561         spin_lock_bh(&net->ksnn_lock);
2562         net->ksnn_shutdown = 1;                 /* prevent new peers */
2563         spin_unlock_bh(&net->ksnn_lock);
2564
2565         /* Delete all peers */
2566         ksocknal_del_peer(ni, anyid, 0);
2567
2568         /* Wait for all peer_ni state to clean up */
2569         i = 2;
2570         spin_lock_bh(&net->ksnn_lock);
2571         while (net->ksnn_npeers != 0) {
2572                 spin_unlock_bh(&net->ksnn_lock);
2573
2574                 i++;
2575                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2576                        "waiting for %d peers to disconnect\n",
2577                        net->ksnn_npeers);
2578                 set_current_state(TASK_UNINTERRUPTIBLE);
2579                 schedule_timeout(cfs_time_seconds(1));
2580
2581                 ksocknal_debug_peerhash(ni);
2582
2583                 spin_lock_bh(&net->ksnn_lock);
2584         }
2585         spin_unlock_bh(&net->ksnn_lock);
2586
2587         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2588                 LASSERT (net->ksnn_interfaces[i].ksni_npeers == 0);
2589                 LASSERT (net->ksnn_interfaces[i].ksni_nroutes == 0);
2590         }
2591
2592         list_del(&net->ksnn_list);
2593         LIBCFS_FREE(net, sizeof(*net));
2594
2595         ksocknal_data.ksnd_nnets--;
2596         if (ksocknal_data.ksnd_nnets == 0)
2597                 ksocknal_base_shutdown();
2598 }
2599
2600 static int
2601 ksocknal_enumerate_interfaces(ksock_net_t *net)
2602 {
2603         char      **names;
2604         int         i;
2605         int         j;
2606         int         rc;
2607         int         n;
2608
2609         n = lnet_ipif_enumerate(&names);
2610         if (n <= 0) {
2611                 CERROR("Can't enumerate interfaces: %d\n", n);
2612                 return n;
2613         }
2614
2615         for (i = j = 0; i < n; i++) {
2616                 int        up;
2617                 __u32      ip;
2618                 __u32      mask;
2619
2620                 if (!strcmp(names[i], "lo")) /* skip the loopback IF */
2621                         continue;
2622
2623                 rc = lnet_ipif_query(names[i], &up, &ip, &mask);
2624                 if (rc != 0) {
2625                         CWARN("Can't get interface %s info: %d\n",
2626                               names[i], rc);
2627                         continue;
2628                 }
2629
2630                 if (!up) {
2631                         CWARN("Ignoring interface %s (down)\n",
2632                               names[i]);
2633                         continue;
2634                 }
2635
2636                 if (j == LNET_NUM_INTERFACES) {
2637                         CWARN("Ignoring interface %s (too many interfaces)\n",
2638                               names[i]);
2639                         continue;
2640                 }
2641
2642                 net->ksnn_interfaces[j].ksni_ipaddr = ip;
2643                 net->ksnn_interfaces[j].ksni_netmask = mask;
2644                 strlcpy(net->ksnn_interfaces[j].ksni_name,
2645                         names[i], sizeof(net->ksnn_interfaces[j].ksni_name));
2646                 j++;
2647         }
2648
2649         lnet_ipif_free_enumeration(names, n);
2650
2651         if (j == 0)
2652                 CERROR("Can't find any usable interfaces\n");
2653
2654         return j;
2655 }
2656
2657 static int
2658 ksocknal_search_new_ipif(ksock_net_t *net)
2659 {
2660         int     new_ipif = 0;
2661         int     i;
2662
2663         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2664                 char            *ifnam = &net->ksnn_interfaces[i].ksni_name[0];
2665                 char            *colon = strchr(ifnam, ':');
2666                 int             found  = 0;
2667                 ksock_net_t     *tmp;
2668                 int             j;
2669
2670                 if (colon != NULL) /* ignore alias device */
2671                         *colon = 0;
2672
2673                 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets,
2674                                         ksnn_list) {
2675                         for (j = 0; !found && j < tmp->ksnn_ninterfaces; j++) {
2676                                 char *ifnam2 = &tmp->ksnn_interfaces[j].\
2677                                              ksni_name[0];
2678                                 char *colon2 = strchr(ifnam2, ':');
2679
2680                                 if (colon2 != NULL)
2681                                         *colon2 = 0;
2682
2683                                 found = strcmp(ifnam, ifnam2) == 0;
2684                                 if (colon2 != NULL)
2685                                         *colon2 = ':';
2686                         }
2687                         if (found)
2688                                 break;
2689                 }
2690
2691                 new_ipif += !found;
2692                 if (colon != NULL)
2693                         *colon = ':';
2694         }
2695
2696         return new_ipif;
2697 }
2698
2699 static int
2700 ksocknal_start_schedulers(struct ksock_sched_info *info)
2701 {
2702         int     nthrs;
2703         int     rc = 0;
2704         int     i;
2705
2706         if (info->ksi_nthreads == 0) {
2707                 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2708                         nthrs = info->ksi_nthreads_max;
2709                 } else {
2710                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2711                                                info->ksi_cpt);
2712                         nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2713                         nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2714                 }
2715                 nthrs = min(nthrs, info->ksi_nthreads_max);
2716         } else {
2717                 LASSERT(info->ksi_nthreads <= info->ksi_nthreads_max);
2718                 /* increase two threads if there is new interface */
2719                 nthrs = min(2, info->ksi_nthreads_max - info->ksi_nthreads);
2720         }
2721
2722         for (i = 0; i < nthrs; i++) {
2723                 long            id;
2724                 char            name[20];
2725                 ksock_sched_t   *sched;
2726                 id = KSOCK_THREAD_ID(info->ksi_cpt, info->ksi_nthreads + i);
2727                 sched = &info->ksi_scheds[KSOCK_THREAD_SID(id)];
2728                 snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
2729                          info->ksi_cpt, (int)(sched - &info->ksi_scheds[0]));
2730
2731                 rc = ksocknal_thread_start(ksocknal_scheduler,
2732                                            (void *)id, name);
2733                 if (rc == 0)
2734                         continue;
2735
2736                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2737                        info->ksi_cpt, info->ksi_nthreads + i, rc);
2738                 break;
2739         }
2740
2741         info->ksi_nthreads += i;
2742         return rc;
2743 }
2744
2745 static int
2746 ksocknal_net_start_threads(ksock_net_t *net, __u32 *cpts, int ncpts)
2747 {
2748         int     newif = ksocknal_search_new_ipif(net);
2749         int     rc;
2750         int     i;
2751
2752         if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2753                 return -EINVAL;
2754
2755         for (i = 0; i < ncpts; i++) {
2756                 struct ksock_sched_info *info;
2757                 int cpt = (cpts == NULL) ? i : cpts[i];
2758
2759                 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2760                 info = ksocknal_data.ksnd_sched_info[cpt];
2761
2762                 if (!newif && info->ksi_nthreads > 0)
2763                         continue;
2764
2765                 rc = ksocknal_start_schedulers(info);
2766                 if (rc != 0)
2767                         return rc;
2768         }
2769         return 0;
2770 }
2771
2772 int
2773 ksocknal_startup(struct lnet_ni *ni)
2774 {
2775         ksock_net_t  *net;
2776         int           rc;
2777         int           i;
2778         struct net_device *net_dev;
2779         int node_id;
2780
2781         LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2782
2783         if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2784                 rc = ksocknal_base_startup();
2785                 if (rc != 0)
2786                         return rc;
2787         }
2788
2789         LIBCFS_ALLOC(net, sizeof(*net));
2790         if (net == NULL)
2791                 goto fail_0;
2792
2793         spin_lock_init(&net->ksnn_lock);
2794         net->ksnn_incarnation = ktime_get_real_ns();
2795         ni->ni_data = net;
2796         if (!ni->ni_net->net_tunables_set) {
2797                 ni->ni_net->net_tunables.lct_peer_timeout =
2798                         *ksocknal_tunables.ksnd_peertimeout;
2799                 ni->ni_net->net_tunables.lct_max_tx_credits =
2800                         *ksocknal_tunables.ksnd_credits;
2801                 ni->ni_net->net_tunables.lct_peer_tx_credits =
2802                         *ksocknal_tunables.ksnd_peertxcredits;
2803                 ni->ni_net->net_tunables.lct_peer_rtr_credits =
2804                         *ksocknal_tunables.ksnd_peerrtrcredits;
2805                 ni->ni_net->net_tunables_set = true;
2806         }
2807
2808
2809         if (ni->ni_interfaces[0] == NULL) {
2810                 rc = ksocknal_enumerate_interfaces(net);
2811                 if (rc <= 0)
2812                         goto fail_1;
2813
2814                 net->ksnn_ninterfaces = 1;
2815         } else {
2816                 for (i = 0; i < LNET_NUM_INTERFACES; i++) {
2817                         int up;
2818
2819                         if (ni->ni_interfaces[i] == NULL)
2820                                 break;
2821
2822                         rc = lnet_ipif_query(ni->ni_interfaces[i], &up,
2823                                 &net->ksnn_interfaces[i].ksni_ipaddr,
2824                                 &net->ksnn_interfaces[i].ksni_netmask);
2825
2826                         if (rc != 0) {
2827                                 CERROR("Can't get interface %s info: %d\n",
2828                                        ni->ni_interfaces[i], rc);
2829                                 goto fail_1;
2830                         }
2831
2832                         if (!up) {
2833                                 CERROR("Interface %s is down\n",
2834                                        ni->ni_interfaces[i]);
2835                                 goto fail_1;
2836                         }
2837
2838                         strlcpy(net->ksnn_interfaces[i].ksni_name,
2839                                 ni->ni_interfaces[i],
2840                                 sizeof(net->ksnn_interfaces[i].ksni_name));
2841
2842                 }
2843                 net->ksnn_ninterfaces = i;
2844         }
2845
2846         net_dev = dev_get_by_name(&init_net,
2847                                   net->ksnn_interfaces[0].ksni_name);
2848         if (net_dev != NULL) {
2849                 node_id = dev_to_node(&net_dev->dev);
2850                 ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
2851                 dev_put(net_dev);
2852         } else {
2853                 ni->ni_dev_cpt = CFS_CPT_ANY;
2854         }
2855
2856         /* call it before add it to ksocknal_data.ksnd_nets */
2857         rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2858         if (rc != 0)
2859                 goto fail_1;
2860
2861         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid),
2862                                 net->ksnn_interfaces[0].ksni_ipaddr);
2863         list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2864
2865         ksocknal_data.ksnd_nnets++;
2866
2867         return 0;
2868
2869  fail_1:
2870         LIBCFS_FREE(net, sizeof(*net));
2871  fail_0:
2872         if (ksocknal_data.ksnd_nnets == 0)
2873                 ksocknal_base_shutdown();
2874
2875         return -ENETDOWN;
2876 }
2877
2878
2879 static void __exit ksocklnd_exit(void)
2880 {
2881         lnet_unregister_lnd(&the_ksocklnd);
2882 }
2883
2884 static int __init ksocklnd_init(void)
2885 {
2886         int rc;
2887
2888         /* check ksnr_connected/connecting field large enough */
2889         CLASSERT(SOCKLND_CONN_NTYPES <= 4);
2890         CLASSERT(SOCKLND_CONN_ACK == SOCKLND_CONN_BULK_IN);
2891
2892         /* initialize the_ksocklnd */
2893         the_ksocklnd.lnd_type     = SOCKLND;
2894         the_ksocklnd.lnd_startup  = ksocknal_startup;
2895         the_ksocklnd.lnd_shutdown = ksocknal_shutdown;
2896         the_ksocklnd.lnd_ctl      = ksocknal_ctl;
2897         the_ksocklnd.lnd_send     = ksocknal_send;
2898         the_ksocklnd.lnd_recv     = ksocknal_recv;
2899         the_ksocklnd.lnd_notify   = ksocknal_notify;
2900         the_ksocklnd.lnd_query    = ksocknal_query;
2901         the_ksocklnd.lnd_accept   = ksocknal_accept;
2902
2903         rc = ksocknal_tunables_init();
2904         if (rc != 0)
2905                 return rc;
2906
2907         lnet_register_lnd(&the_ksocklnd);
2908
2909         return 0;
2910 }
2911
2912 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2913 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2914 MODULE_VERSION("2.8.0");
2915 MODULE_LICENSE("GPL");
2916
2917 module_init(ksocklnd_init);
2918 module_exit(ksocklnd_exit);