lnet/klnds/socklnd/socklnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/klnds/socklnd/socklnd.c
  33  *
  34  * Author: Zach Brown <zab@zabbo.net>
  35  * Author: Peter J. Braam <braam@clusterfs.com>
  36  * Author: Phil Schwan <phil@clusterfs.com>
  37  * Author: Eric Barton <eric@bartonsoftware.com>
  38  */
  39
  40 #include "socklnd.h"
  41 #include <linux/inetdevice.h>
  42
  43 static struct lnet_lnd the_ksocklnd;
  44 struct ksock_nal_data ksocknal_data;
  45
  46 static struct ksock_interface *
  47 ksocknal_ip2iface(struct lnet_ni *ni, __u32 ip)
  48 {
  49         struct ksock_net *net = ni->ni_data;
  50         int i;
  51         struct ksock_interface *iface;
  52
  53         for (i = 0; i < net->ksnn_ninterfaces; i++) {
  54                 LASSERT(i < LNET_INTERFACES_NUM);
  55                 iface = &net->ksnn_interfaces[i];
  56
  57                 if (iface->ksni_ipaddr == ip)
  58                         return iface;
  59         }
  60
  61         return NULL;
  62 }
  63
  64 static struct ksock_route *
  65 ksocknal_create_route(__u32 ipaddr, int port)
  66 {
  67         struct ksock_route *route;
  68
  69         LIBCFS_ALLOC (route, sizeof (*route));
  70         if (route == NULL)
  71                 return (NULL);
  72
  73         atomic_set (&route->ksnr_refcount, 1);
  74         route->ksnr_peer = NULL;
  75         route->ksnr_retry_interval = 0;         /* OK to connect at any time */
  76         route->ksnr_ipaddr = ipaddr;
  77         route->ksnr_port = port;
  78         route->ksnr_scheduled = 0;
  79         route->ksnr_connecting = 0;
  80         route->ksnr_connected = 0;
  81         route->ksnr_deleted = 0;
  82         route->ksnr_conn_count = 0;
  83         route->ksnr_share_count = 0;
  84
  85         return (route);
  86 }
  87
  88 void
  89 ksocknal_destroy_route(struct ksock_route *route)
  90 {
  91         LASSERT (atomic_read(&route->ksnr_refcount) == 0);
  92
  93         if (route->ksnr_peer != NULL)
  94                 ksocknal_peer_decref(route->ksnr_peer);
  95
  96         LIBCFS_FREE (route, sizeof (*route));
  97 }
  98
  99 static int
 100 ksocknal_create_peer(struct ksock_peer_ni **peerp, struct lnet_ni *ni,
 101                      struct lnet_process_id id)
 102 {
 103         int cpt = lnet_cpt_of_nid(id.nid, ni);
 104         struct ksock_net *net = ni->ni_data;
 105         struct ksock_peer_ni *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(struct ksock_peer_ni *peer_ni)
 150 {
 151         struct ksock_net *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 struct ksock_peer_ni *
 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         struct ksock_peer_ni *peer_ni;
 180
 181         list_for_each(tmp, peer_list) {
 182                 peer_ni = list_entry(tmp, struct ksock_peer_ni, ksnp_list);
 183
 184                 LASSERT(!peer_ni->ksnp_closing);
 185
 186                 if (peer_ni->ksnp_ni != ni)
 187                         continue;
 188
 189                 if (peer_ni->ksnp_id.nid != id.nid ||
 190                     peer_ni->ksnp_id.pid != id.pid)
 191                         continue;
 192
 193                 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d)\n",
 194                        peer_ni, libcfs_id2str(id),
 195                        atomic_read(&peer_ni->ksnp_refcount));
 196                 return peer_ni;
 197         }
 198         return NULL;
 199 }
 200
 201 struct ksock_peer_ni *
 202 ksocknal_find_peer(struct lnet_ni *ni, struct lnet_process_id id)
 203 {
 204         struct ksock_peer_ni *peer_ni;
 205
 206         read_lock(&ksocknal_data.ksnd_global_lock);
 207         peer_ni = ksocknal_find_peer_locked(ni, id);
 208         if (peer_ni != NULL)                    /* +1 ref for caller? */
 209                 ksocknal_peer_addref(peer_ni);
 210         read_unlock(&ksocknal_data.ksnd_global_lock);
 211
 212         return (peer_ni);
 213 }
 214
 215 static void
 216 ksocknal_unlink_peer_locked(struct ksock_peer_ni *peer_ni)
 217 {
 218         int i;
 219         __u32 ip;
 220         struct ksock_interface *iface;
 221
 222         for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++) {
 223                 LASSERT(i < LNET_INTERFACES_NUM);
 224                 ip = peer_ni->ksnp_passive_ips[i];
 225
 226                 iface = ksocknal_ip2iface(peer_ni->ksnp_ni, ip);
 227                 /*
 228                  * All IPs in peer_ni->ksnp_passive_ips[] come from the
 229                  * interface list, therefore the call must succeed.
 230                  */
 231                 LASSERT(iface != NULL);
 232
 233                 CDEBUG(D_NET, "peer_ni=%p iface=%p ksni_nroutes=%d\n",
 234                        peer_ni, iface, iface->ksni_nroutes);
 235                 iface->ksni_npeers--;
 236         }
 237
 238         LASSERT(list_empty(&peer_ni->ksnp_conns));
 239         LASSERT(list_empty(&peer_ni->ksnp_routes));
 240         LASSERT(!peer_ni->ksnp_closing);
 241         peer_ni->ksnp_closing = 1;
 242         list_del(&peer_ni->ksnp_list);
 243         /* lose peerlist's ref */
 244         ksocknal_peer_decref(peer_ni);
 245 }
 246
 247 static int
 248 ksocknal_get_peer_info(struct lnet_ni *ni, int index,
 249                        struct lnet_process_id *id, __u32 *myip, __u32 *peer_ip,
 250                        int *port, int *conn_count, int *share_count)
 251 {
 252         struct ksock_peer_ni *peer_ni;
 253         struct list_head *ptmp;
 254         struct ksock_route *route;
 255         struct list_head *rtmp;
 256         int i;
 257         int j;
 258         int rc = -ENOENT;
 259
 260         read_lock(&ksocknal_data.ksnd_global_lock);
 261
 262         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
 263                 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
 264                         peer_ni = list_entry(ptmp, struct ksock_peer_ni, ksnp_list);
 265
 266                         if (peer_ni->ksnp_ni != ni)
 267                                 continue;
 268
 269                         if (peer_ni->ksnp_n_passive_ips == 0 &&
 270                             list_empty(&peer_ni->ksnp_routes)) {
 271                                 if (index-- > 0)
 272                                         continue;
 273
 274                                 *id = peer_ni->ksnp_id;
 275                                 *myip = 0;
 276                                 *peer_ip = 0;
 277                                 *port = 0;
 278                                 *conn_count = 0;
 279                                 *share_count = 0;
 280                                 rc = 0;
 281                                 goto out;
 282                         }
 283
 284                         for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++) {
 285                                 if (index-- > 0)
 286                                         continue;
 287
 288                                 *id = peer_ni->ksnp_id;
 289                                 *myip = peer_ni->ksnp_passive_ips[j];
 290                                 *peer_ip = 0;
 291                                 *port = 0;
 292                                 *conn_count = 0;
 293                                 *share_count = 0;
 294                                 rc = 0;
 295                                 goto out;
 296                         }
 297
 298                         list_for_each(rtmp, &peer_ni->ksnp_routes) {
 299                                 if (index-- > 0)
 300                                         continue;
 301
 302                                 route = list_entry(rtmp, struct ksock_route,
 303                                                    ksnr_list);
 304
 305                                 *id = peer_ni->ksnp_id;
 306                                 *myip = route->ksnr_myipaddr;
 307                                 *peer_ip = route->ksnr_ipaddr;
 308                                 *port = route->ksnr_port;
 309                                 *conn_count = route->ksnr_conn_count;
 310                                 *share_count = route->ksnr_share_count;
 311                                 rc = 0;
 312                                 goto out;
 313                         }
 314                 }
 315         }
 316 out:
 317         read_unlock(&ksocknal_data.ksnd_global_lock);
 318         return rc;
 319 }
 320
 321 static void
 322 ksocknal_associate_route_conn_locked(struct ksock_route *route, struct ksock_conn *conn)
 323 {
 324         struct ksock_peer_ni *peer_ni = route->ksnr_peer;
 325         int type = conn->ksnc_type;
 326         struct ksock_interface *iface;
 327
 328         conn->ksnc_route = route;
 329         ksocknal_route_addref(route);
 330
 331         if (route->ksnr_myipaddr != conn->ksnc_myipaddr) {
 332                 if (route->ksnr_myipaddr == 0) {
 333                         /* route wasn't bound locally yet (the initial route) */
 334                         CDEBUG(D_NET, "Binding %s %pI4h to %pI4h\n",
 335                                libcfs_id2str(peer_ni->ksnp_id),
 336                                &route->ksnr_ipaddr,
 337                                &conn->ksnc_myipaddr);
 338                 } else {
 339                         CDEBUG(D_NET, "Rebinding %s %pI4h from %pI4h "
 340                                "to %pI4h\n", libcfs_id2str(peer_ni->ksnp_id),
 341                                &route->ksnr_ipaddr,
 342                                &route->ksnr_myipaddr,
 343                                &conn->ksnc_myipaddr);
 344
 345                         iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
 346                                                   route->ksnr_myipaddr);
 347                         if (iface != NULL)
 348                                 iface->ksni_nroutes--;
 349                 }
 350                 route->ksnr_myipaddr = conn->ksnc_myipaddr;
 351                 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
 352                                           route->ksnr_myipaddr);
 353                 if (iface != NULL)
 354                         iface->ksni_nroutes++;
 355         }
 356
 357         route->ksnr_connected |= (1<<type);
 358         route->ksnr_conn_count++;
 359
 360         /* Successful connection => further attempts can
 361          * proceed immediately */
 362         route->ksnr_retry_interval = 0;
 363 }
 364
 365 static void
 366 ksocknal_add_route_locked(struct ksock_peer_ni *peer_ni, struct ksock_route *route)
 367 {
 368         struct list_head *tmp;
 369         struct ksock_conn *conn;
 370         struct ksock_route *route2;
 371
 372         LASSERT(!peer_ni->ksnp_closing);
 373         LASSERT(route->ksnr_peer == NULL);
 374         LASSERT(!route->ksnr_scheduled);
 375         LASSERT(!route->ksnr_connecting);
 376         LASSERT(route->ksnr_connected == 0);
 377
 378         /* LASSERT(unique) */
 379         list_for_each(tmp, &peer_ni->ksnp_routes) {
 380                 route2 = list_entry(tmp, struct ksock_route, ksnr_list);
 381
 382                 if (route2->ksnr_ipaddr == route->ksnr_ipaddr) {
 383                         CERROR("Duplicate route %s %pI4h\n",
 384                                libcfs_id2str(peer_ni->ksnp_id),
 385                                &route->ksnr_ipaddr);
 386                         LBUG();
 387                 }
 388         }
 389
 390         route->ksnr_peer = peer_ni;
 391         ksocknal_peer_addref(peer_ni);
 392         /* peer_ni's routelist takes over my ref on 'route' */
 393         list_add_tail(&route->ksnr_list, &peer_ni->ksnp_routes);
 394
 395         list_for_each(tmp, &peer_ni->ksnp_conns) {
 396                 conn = list_entry(tmp, struct ksock_conn, ksnc_list);
 397
 398                 if (conn->ksnc_ipaddr != route->ksnr_ipaddr)
 399                         continue;
 400
 401                 ksocknal_associate_route_conn_locked(route, conn);
 402                 /* keep going (typed routes) */
 403         }
 404 }
 405
 406 static void
 407 ksocknal_del_route_locked(struct ksock_route *route)
 408 {
 409         struct ksock_peer_ni *peer_ni = route->ksnr_peer;
 410         struct ksock_interface *iface;
 411         struct ksock_conn *conn;
 412         struct list_head *ctmp;
 413         struct list_head *cnxt;
 414
 415         LASSERT(!route->ksnr_deleted);
 416
 417         /* Close associated conns */
 418         list_for_each_safe(ctmp, cnxt, &peer_ni->ksnp_conns) {
 419                 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
 420
 421                 if (conn->ksnc_route != route)
 422                         continue;
 423
 424                 ksocknal_close_conn_locked(conn, 0);
 425         }
 426
 427         if (route->ksnr_myipaddr != 0) {
 428                 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
 429                                           route->ksnr_myipaddr);
 430                 if (iface != NULL)
 431                         iface->ksni_nroutes--;
 432         }
 433
 434         route->ksnr_deleted = 1;
 435         list_del(&route->ksnr_list);
 436         ksocknal_route_decref(route);           /* drop peer_ni's ref */
 437
 438         if (list_empty(&peer_ni->ksnp_routes) &&
 439             list_empty(&peer_ni->ksnp_conns)) {
 440                 /* I've just removed the last route to a peer_ni with no active
 441                  * connections */
 442                 ksocknal_unlink_peer_locked(peer_ni);
 443         }
 444 }
 445
 446 int
 447 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ipaddr,
 448                   int port)
 449 {
 450         struct list_head *tmp;
 451         struct ksock_peer_ni *peer_ni;
 452         struct ksock_peer_ni *peer2;
 453         struct ksock_route *route;
 454         struct ksock_route *route2;
 455         int rc;
 456
 457         if (id.nid == LNET_NID_ANY ||
 458             id.pid == LNET_PID_ANY)
 459                 return (-EINVAL);
 460
 461         /* Have a brand new peer_ni ready... */
 462         rc = ksocknal_create_peer(&peer_ni, ni, id);
 463         if (rc != 0)
 464                 return rc;
 465
 466         route = ksocknal_create_route (ipaddr, port);
 467         if (route == NULL) {
 468                 ksocknal_peer_decref(peer_ni);
 469                 return (-ENOMEM);
 470         }
 471
 472         write_lock_bh(&ksocknal_data.ksnd_global_lock);
 473
 474         /* always called with a ref on ni, so shutdown can't have started */
 475         LASSERT(((struct ksock_net *) ni->ni_data)->ksnn_shutdown == 0);
 476
 477         peer2 = ksocknal_find_peer_locked(ni, id);
 478         if (peer2 != NULL) {
 479                 ksocknal_peer_decref(peer_ni);
 480                 peer_ni = peer2;
 481         } else {
 482                 /* peer_ni table takes my ref on peer_ni */
 483                 list_add_tail(&peer_ni->ksnp_list,
 484                               ksocknal_nid2peerlist(id.nid));
 485         }
 486
 487         route2 = NULL;
 488         list_for_each(tmp, &peer_ni->ksnp_routes) {
 489                 route2 = list_entry(tmp, struct ksock_route, ksnr_list);
 490
 491                 if (route2->ksnr_ipaddr == ipaddr)
 492                         break;
 493
 494                 route2 = NULL;
 495         }
 496         if (route2 == NULL) {
 497                 ksocknal_add_route_locked(peer_ni, route);
 498                 route->ksnr_share_count++;
 499         } else {
 500                 ksocknal_route_decref(route);
 501                 route2->ksnr_share_count++;
 502         }
 503
 504         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 505
 506         return 0;
 507 }
 508
 509 static void
 510 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni, __u32 ip)
 511 {
 512         struct ksock_conn *conn;
 513         struct ksock_route *route;
 514         struct list_head *tmp;
 515         struct list_head *nxt;
 516         int nshared;
 517
 518         LASSERT(!peer_ni->ksnp_closing);
 519
 520         /* Extra ref prevents peer_ni disappearing until I'm done with it */
 521         ksocknal_peer_addref(peer_ni);
 522
 523         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
 524                 route = list_entry(tmp, struct ksock_route, ksnr_list);
 525
 526                 /* no match */
 527                 if (!(ip == 0 || route->ksnr_ipaddr == ip))
 528                         continue;
 529
 530                 route->ksnr_share_count = 0;
 531                 /* This deletes associated conns too */
 532                 ksocknal_del_route_locked(route);
 533         }
 534
 535         nshared = 0;
 536         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
 537                 route = list_entry(tmp, struct ksock_route, ksnr_list);
 538                 nshared += route->ksnr_share_count;
 539         }
 540
 541         if (nshared == 0) {
 542                 /* remove everything else if there are no explicit entries
 543                  * left */
 544
 545                 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
 546                         route = list_entry(tmp, struct ksock_route, ksnr_list);
 547
 548                         /* we should only be removing auto-entries */
 549                         LASSERT(route->ksnr_share_count == 0);
 550                         ksocknal_del_route_locked(route);
 551                 }
 552
 553                 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_conns) {
 554                         conn = list_entry(tmp, struct ksock_conn, ksnc_list);
 555
 556                         ksocknal_close_conn_locked(conn, 0);
 557                 }
 558         }
 559
 560         ksocknal_peer_decref(peer_ni);
 561         /* NB peer_ni unlinks itself when last conn/route is removed */
 562 }
 563
 564 static int
 565 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ip)
 566 {
 567         struct list_head zombies = LIST_HEAD_INIT(zombies);
 568         struct list_head *ptmp;
 569         struct list_head *pnxt;
 570         struct ksock_peer_ni *peer_ni;
 571         int lo;
 572         int hi;
 573         int i;
 574         int rc = -ENOENT;
 575
 576         write_lock_bh(&ksocknal_data.ksnd_global_lock);
 577
 578         if (id.nid != LNET_NID_ANY) {
 579                 hi = (int)(ksocknal_nid2peerlist(id.nid) -
 580                            ksocknal_data.ksnd_peers);
 581                 lo = hi;
 582         } else {
 583                 lo = 0;
 584                 hi = ksocknal_data.ksnd_peer_hash_size - 1;
 585         }
 586
 587         for (i = lo; i <= hi; i++) {
 588                 list_for_each_safe(ptmp, pnxt,
 589                                    &ksocknal_data.ksnd_peers[i]) {
 590                         peer_ni = list_entry(ptmp, struct ksock_peer_ni, ksnp_list);
 591
 592                         if (peer_ni->ksnp_ni != ni)
 593                                 continue;
 594
 595                         if (!((id.nid == LNET_NID_ANY ||
 596                                peer_ni->ksnp_id.nid == id.nid) &&
 597                               (id.pid == LNET_PID_ANY ||
 598                                peer_ni->ksnp_id.pid == id.pid)))
 599                                 continue;
 600
 601                         ksocknal_peer_addref(peer_ni);  /* a ref for me... */
 602
 603                         ksocknal_del_peer_locked(peer_ni, ip);
 604
 605                         if (peer_ni->ksnp_closing &&
 606                             !list_empty(&peer_ni->ksnp_tx_queue)) {
 607                                 LASSERT(list_empty(&peer_ni->ksnp_conns));
 608                                 LASSERT(list_empty(&peer_ni->ksnp_routes));
 609
 610                                 list_splice_init(&peer_ni->ksnp_tx_queue,
 611                                                  &zombies);
 612                         }
 613
 614                         ksocknal_peer_decref(peer_ni);  /* ...till here */
 615
 616                         rc = 0;                         /* matched! */
 617                 }
 618         }
 619
 620         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 621
 622         ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
 623
 624         return rc;
 625 }
 626
 627 static struct ksock_conn *
 628 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
 629 {
 630         struct ksock_peer_ni *peer_ni;
 631         struct list_head *ptmp;
 632         struct ksock_conn *conn;
 633         struct list_head *ctmp;
 634         int i;
 635
 636         read_lock(&ksocknal_data.ksnd_global_lock);
 637
 638         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
 639                 list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
 640                         peer_ni = list_entry(ptmp, struct ksock_peer_ni, ksnp_list);
 641
 642                         LASSERT(!peer_ni->ksnp_closing);
 643
 644                         if (peer_ni->ksnp_ni != ni)
 645                                 continue;
 646
 647                         list_for_each(ctmp, &peer_ni->ksnp_conns) {
 648                                 if (index-- > 0)
 649                                         continue;
 650
 651                                 conn = list_entry(ctmp, struct ksock_conn,
 652                                                   ksnc_list);
 653                                 ksocknal_conn_addref(conn);
 654                                 read_unlock(&ksocknal_data. \
 655                                             ksnd_global_lock);
 656                                 return conn;
 657                         }
 658                 }
 659         }
 660
 661         read_unlock(&ksocknal_data.ksnd_global_lock);
 662         return NULL;
 663 }
 664
 665 static struct ksock_sched *
 666 ksocknal_choose_scheduler_locked(unsigned int cpt)
 667 {
 668         struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
 669         int i;
 670
 671         if (sched->kss_nthreads == 0) {
 672                 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
 673                         if (sched->kss_nthreads > 0) {
 674                                 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 675                                        cpt, sched->kss_cpt);
 676                                 return sched;
 677                         }
 678                 }
 679                 return NULL;
 680         }
 681
 682         return sched;
 683 }
 684
 685 static int
 686 ksocknal_local_ipvec(struct lnet_ni *ni, __u32 *ipaddrs)
 687 {
 688         struct ksock_net *net = ni->ni_data;
 689         int i;
 690         int nip;
 691
 692         read_lock(&ksocknal_data.ksnd_global_lock);
 693
 694         nip = net->ksnn_ninterfaces;
 695         LASSERT(nip <= LNET_INTERFACES_NUM);
 696
 697         /*
 698          * Only offer interfaces for additional connections if I have
 699          * more than one.
 700          */
 701         if (nip < 2) {
 702                 read_unlock(&ksocknal_data.ksnd_global_lock);
 703                 return 0;
 704         }
 705
 706         for (i = 0; i < nip; i++) {
 707                 ipaddrs[i] = net->ksnn_interfaces[i].ksni_ipaddr;
 708                 LASSERT(ipaddrs[i] != 0);
 709         }
 710
 711         read_unlock(&ksocknal_data.ksnd_global_lock);
 712         return nip;
 713 }
 714
 715 static int
 716 ksocknal_match_peerip(struct ksock_interface *iface, __u32 *ips, int nips)
 717 {
 718         int best_netmatch = 0;
 719         int best_xor = 0;
 720         int best = -1;
 721         int this_xor;
 722         int this_netmatch;
 723         int i;
 724
 725         for (i = 0; i < nips; i++) {
 726                 if (ips[i] == 0)
 727                         continue;
 728
 729                 this_xor = (ips[i] ^ iface->ksni_ipaddr);
 730                 this_netmatch = ((this_xor & iface->ksni_netmask) == 0) ? 1 : 0;
 731
 732                 if (!(best < 0 ||
 733                       best_netmatch < this_netmatch ||
 734                       (best_netmatch == this_netmatch &&
 735                        best_xor > this_xor)))
 736                         continue;
 737
 738                 best = i;
 739                 best_netmatch = this_netmatch;
 740                 best_xor = this_xor;
 741         }
 742
 743         LASSERT (best >= 0);
 744         return (best);
 745 }
 746
 747 static int
 748 ksocknal_select_ips(struct ksock_peer_ni *peer_ni, __u32 *peerips, int n_peerips)
 749 {
 750         rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
 751         struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
 752         struct ksock_interface *iface;
 753         struct ksock_interface *best_iface;
 754         int n_ips;
 755         int i;
 756         int j;
 757         int k;
 758         u32 ip;
 759         u32 xor;
 760         int this_netmatch;
 761         int best_netmatch;
 762         int best_npeers;
 763
 764         /* CAVEAT EMPTOR: We do all our interface matching with an
 765          * exclusive hold of global lock at IRQ priority.  We're only
 766          * expecting to be dealing with small numbers of interfaces, so the
 767          * O(n**3)-ness shouldn't matter */
 768
 769         /* Also note that I'm not going to return more than n_peerips
 770          * interfaces, even if I have more myself */
 771
 772         write_lock_bh(global_lock);
 773
 774         LASSERT(n_peerips <= LNET_INTERFACES_NUM);
 775         LASSERT(net->ksnn_ninterfaces <= LNET_INTERFACES_NUM);
 776
 777         /* Only match interfaces for additional connections
 778          * if I have > 1 interface */
 779         n_ips = (net->ksnn_ninterfaces < 2) ? 0 :
 780                 MIN(n_peerips, net->ksnn_ninterfaces);
 781
 782         for (i = 0; peer_ni->ksnp_n_passive_ips < n_ips; i++) {
 783                 /*              ^ yes really... */
 784
 785                 /* If we have any new interfaces, first tick off all the
 786                  * peer_ni IPs that match old interfaces, then choose new
 787                  * interfaces to match the remaining peer_ni IPS.
 788                  * We don't forget interfaces we've stopped using; we might
 789                  * start using them again... */
 790
 791                 if (i < peer_ni->ksnp_n_passive_ips) {
 792                         /* Old interface. */
 793                         ip = peer_ni->ksnp_passive_ips[i];
 794                         best_iface = ksocknal_ip2iface(peer_ni->ksnp_ni, ip);
 795
 796                         /* peer_ni passive ips are kept up to date */
 797                         LASSERT(best_iface != NULL);
 798                 } else {
 799                         /* choose a new interface */
 800                         LASSERT (i == peer_ni->ksnp_n_passive_ips);
 801
 802                         best_iface = NULL;
 803                         best_netmatch = 0;
 804                         best_npeers = 0;
 805
 806                         for (j = 0; j < net->ksnn_ninterfaces; j++) {
 807                                 iface = &net->ksnn_interfaces[j];
 808                                 ip = iface->ksni_ipaddr;
 809
 810                                 for (k = 0; k < peer_ni->ksnp_n_passive_ips; k++)
 811                                         if (peer_ni->ksnp_passive_ips[k] == ip)
 812                                                 break;
 813
 814                                 if (k < peer_ni->ksnp_n_passive_ips) /* using it already */
 815                                         continue;
 816
 817                                 k = ksocknal_match_peerip(iface, peerips, n_peerips);
 818                                 xor = (ip ^ peerips[k]);
 819                                 this_netmatch = ((xor & iface->ksni_netmask) == 0) ? 1 : 0;
 820
 821                                 if (!(best_iface == NULL ||
 822                                       best_netmatch < this_netmatch ||
 823                                       (best_netmatch == this_netmatch &&
 824                                        best_npeers > iface->ksni_npeers)))
 825                                         continue;
 826
 827                                 best_iface = iface;
 828                                 best_netmatch = this_netmatch;
 829                                 best_npeers = iface->ksni_npeers;
 830                         }
 831
 832                         LASSERT(best_iface != NULL);
 833
 834                         best_iface->ksni_npeers++;
 835                         ip = best_iface->ksni_ipaddr;
 836                         peer_ni->ksnp_passive_ips[i] = ip;
 837                         peer_ni->ksnp_n_passive_ips = i+1;
 838                 }
 839
 840                 /* mark the best matching peer_ni IP used */
 841                 j = ksocknal_match_peerip(best_iface, peerips, n_peerips);
 842                 peerips[j] = 0;
 843         }
 844
 845         /* Overwrite input peer_ni IP addresses */
 846         memcpy(peerips, peer_ni->ksnp_passive_ips, n_ips * sizeof(*peerips));
 847
 848         write_unlock_bh(global_lock);
 849
 850         return (n_ips);
 851 }
 852
 853 static void
 854 ksocknal_create_routes(struct ksock_peer_ni *peer_ni, int port,
 855                        __u32 *peer_ipaddrs, int npeer_ipaddrs)
 856 {
 857         struct ksock_route              *newroute = NULL;
 858         rwlock_t                *global_lock = &ksocknal_data.ksnd_global_lock;
 859         struct lnet_ni *ni = peer_ni->ksnp_ni;
 860         struct ksock_net                *net = ni->ni_data;
 861         struct list_head        *rtmp;
 862         struct ksock_route              *route;
 863         struct ksock_interface  *iface;
 864         struct ksock_interface  *best_iface;
 865         int                     best_netmatch;
 866         int                     this_netmatch;
 867         int                     best_nroutes;
 868         int                     i;
 869         int                     j;
 870
 871         /* CAVEAT EMPTOR: We do all our interface matching with an
 872          * exclusive hold of global lock at IRQ priority.  We're only
 873          * expecting to be dealing with small numbers of interfaces, so the
 874          * O(n**3)-ness here shouldn't matter */
 875
 876         write_lock_bh(global_lock);
 877
 878         if (net->ksnn_ninterfaces < 2) {
 879                 /* Only create additional connections
 880                  * if I have > 1 interface */
 881                 write_unlock_bh(global_lock);
 882                 return;
 883         }
 884
 885         LASSERT(npeer_ipaddrs <= LNET_INTERFACES_NUM);
 886
 887         for (i = 0; i < npeer_ipaddrs; i++) {
 888                 if (newroute != NULL) {
 889                         newroute->ksnr_ipaddr = peer_ipaddrs[i];
 890                 } else {
 891                         write_unlock_bh(global_lock);
 892
 893                         newroute = ksocknal_create_route(peer_ipaddrs[i], port);
 894                         if (newroute == NULL)
 895                                 return;
 896
 897                         write_lock_bh(global_lock);
 898                 }
 899
 900                 if (peer_ni->ksnp_closing) {
 901                         /* peer_ni got closed under me */
 902                         break;
 903                 }
 904
 905                 /* Already got a route? */
 906                 route = NULL;
 907                 list_for_each(rtmp, &peer_ni->ksnp_routes) {
 908                         route = list_entry(rtmp, struct ksock_route, ksnr_list);
 909
 910                         if (route->ksnr_ipaddr == newroute->ksnr_ipaddr)
 911                                 break;
 912
 913                         route = NULL;
 914                 }
 915                 if (route != NULL)
 916                         continue;
 917
 918                 best_iface = NULL;
 919                 best_nroutes = 0;
 920                 best_netmatch = 0;
 921
 922                 LASSERT(net->ksnn_ninterfaces <= LNET_INTERFACES_NUM);
 923
 924                 /* Select interface to connect from */
 925                 for (j = 0; j < net->ksnn_ninterfaces; j++) {
 926                         iface = &net->ksnn_interfaces[j];
 927
 928                         /* Using this interface already? */
 929                         list_for_each(rtmp, &peer_ni->ksnp_routes) {
 930                                 route = list_entry(rtmp, struct ksock_route,
 931                                                    ksnr_list);
 932
 933                                 if (route->ksnr_myipaddr == iface->ksni_ipaddr)
 934                                         break;
 935
 936                                 route = NULL;
 937                         }
 938                         if (route != NULL)
 939                                 continue;
 940
 941                         this_netmatch = (((iface->ksni_ipaddr ^
 942                                            newroute->ksnr_ipaddr) &
 943                                            iface->ksni_netmask) == 0) ? 1 : 0;
 944
 945                         if (!(best_iface == NULL ||
 946                               best_netmatch < this_netmatch ||
 947                               (best_netmatch == this_netmatch &&
 948                                best_nroutes > iface->ksni_nroutes)))
 949                                 continue;
 950
 951                         best_iface = iface;
 952                         best_netmatch = this_netmatch;
 953                         best_nroutes = iface->ksni_nroutes;
 954                 }
 955
 956                 if (best_iface == NULL)
 957                         continue;
 958
 959                 newroute->ksnr_myipaddr = best_iface->ksni_ipaddr;
 960                 best_iface->ksni_nroutes++;
 961
 962                 ksocknal_add_route_locked(peer_ni, newroute);
 963                 newroute = NULL;
 964         }
 965
 966         write_unlock_bh(global_lock);
 967         if (newroute != NULL)
 968                 ksocknal_route_decref(newroute);
 969 }
 970
 971 int
 972 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
 973 {
 974         struct ksock_connreq *cr;
 975         int rc;
 976         u32 peer_ip;
 977         int peer_port;
 978
 979         rc = lnet_sock_getaddr(sock, true, &peer_ip, &peer_port);
 980         LASSERT(rc == 0);               /* we succeeded before */
 981
 982         LIBCFS_ALLOC(cr, sizeof(*cr));
 983         if (cr == NULL) {
 984                 LCONSOLE_ERROR_MSG(0x12f, "Dropping connection request from "
 985                                    "%pI4h: memory exhausted\n", &peer_ip);
 986                 return -ENOMEM;
 987         }
 988
 989         lnet_ni_addref(ni);
 990         cr->ksncr_ni   = ni;
 991         cr->ksncr_sock = sock;
 992
 993         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
 994
 995         list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
 996         wake_up(&ksocknal_data.ksnd_connd_waitq);
 997
 998         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
 999         return 0;
1000 }
1001
1002 static int
1003 ksocknal_connecting(struct ksock_peer_ni *peer_ni, __u32 ipaddr)
1004 {
1005         struct ksock_route *route;
1006
1007         list_for_each_entry(route, &peer_ni->ksnp_routes, ksnr_list) {
1008                 if (route->ksnr_ipaddr == ipaddr)
1009                         return route->ksnr_connecting;
1010         }
1011         return 0;
1012 }
1013
1014 int
1015 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_route *route,
1016                      struct socket *sock, int type)
1017 {
1018         rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
1019         struct list_head zombies = LIST_HEAD_INIT(zombies);
1020         struct lnet_process_id peerid;
1021         struct list_head *tmp;
1022         u64 incarnation;
1023         struct ksock_conn *conn;
1024         struct ksock_conn *conn2;
1025         struct ksock_peer_ni *peer_ni = NULL;
1026         struct ksock_peer_ni *peer2;
1027         struct ksock_sched *sched;
1028         struct ksock_hello_msg *hello;
1029         int cpt;
1030         struct ksock_tx *tx;
1031         struct ksock_tx *txtmp;
1032         int rc;
1033         int rc2;
1034         int active;
1035         char *warn = NULL;
1036
1037         active = (route != NULL);
1038
1039         LASSERT (active == (type != SOCKLND_CONN_NONE));
1040
1041         LIBCFS_ALLOC(conn, sizeof(*conn));
1042         if (conn == NULL) {
1043                 rc = -ENOMEM;
1044                 goto failed_0;
1045         }
1046
1047         conn->ksnc_peer = NULL;
1048         conn->ksnc_route = NULL;
1049         conn->ksnc_sock = sock;
1050         /* 2 ref, 1 for conn, another extra ref prevents socket
1051          * being closed before establishment of connection */
1052         atomic_set (&conn->ksnc_sock_refcount, 2);
1053         conn->ksnc_type = type;
1054         ksocknal_lib_save_callback(sock, conn);
1055         atomic_set (&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
1056
1057         conn->ksnc_rx_ready = 0;
1058         conn->ksnc_rx_scheduled = 0;
1059
1060         INIT_LIST_HEAD(&conn->ksnc_tx_queue);
1061         conn->ksnc_tx_ready = 0;
1062         conn->ksnc_tx_scheduled = 0;
1063         conn->ksnc_tx_carrier = NULL;
1064         atomic_set (&conn->ksnc_tx_nob, 0);
1065
1066         LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
1067                                      kshm_ips[LNET_INTERFACES_NUM]));
1068         if (hello == NULL) {
1069                 rc = -ENOMEM;
1070                 goto failed_1;
1071         }
1072
1073         /* stash conn's local and remote addrs */
1074         rc = ksocknal_lib_get_conn_addrs (conn);
1075         if (rc != 0)
1076                 goto failed_1;
1077
1078         /* Find out/confirm peer_ni's NID and connection type and get the
1079          * vector of interfaces she's willing to let me connect to.
1080          * Passive connections use the listener timeout since the peer_ni sends
1081          * eagerly */
1082
1083         if (active) {
1084                 peer_ni = route->ksnr_peer;
1085                 LASSERT(ni == peer_ni->ksnp_ni);
1086
1087                 /* Active connection sends HELLO eagerly */
1088                 hello->kshm_nips = ksocknal_local_ipvec(ni, hello->kshm_ips);
1089                 peerid = peer_ni->ksnp_id;
1090
1091                 write_lock_bh(global_lock);
1092                 conn->ksnc_proto = peer_ni->ksnp_proto;
1093                 write_unlock_bh(global_lock);
1094
1095                 if (conn->ksnc_proto == NULL) {
1096                          conn->ksnc_proto = &ksocknal_protocol_v3x;
1097 #if SOCKNAL_VERSION_DEBUG
1098                          if (*ksocknal_tunables.ksnd_protocol == 2)
1099                                  conn->ksnc_proto = &ksocknal_protocol_v2x;
1100                          else if (*ksocknal_tunables.ksnd_protocol == 1)
1101                                  conn->ksnc_proto = &ksocknal_protocol_v1x;
1102 #endif
1103                 }
1104
1105                 rc = ksocknal_send_hello (ni, conn, peerid.nid, hello);
1106                 if (rc != 0)
1107                         goto failed_1;
1108         } else {
1109                 peerid.nid = LNET_NID_ANY;
1110                 peerid.pid = LNET_PID_ANY;
1111
1112                 /* Passive, get protocol from peer_ni */
1113                 conn->ksnc_proto = NULL;
1114         }
1115
1116         rc = ksocknal_recv_hello (ni, conn, hello, &peerid, &incarnation);
1117         if (rc < 0)
1118                 goto failed_1;
1119
1120         LASSERT (rc == 0 || active);
1121         LASSERT (conn->ksnc_proto != NULL);
1122         LASSERT (peerid.nid != LNET_NID_ANY);
1123
1124         cpt = lnet_cpt_of_nid(peerid.nid, ni);
1125
1126         if (active) {
1127                 ksocknal_peer_addref(peer_ni);
1128                 write_lock_bh(global_lock);
1129         } else {
1130                 rc = ksocknal_create_peer(&peer_ni, ni, peerid);
1131                 if (rc != 0)
1132                         goto failed_1;
1133
1134                 write_lock_bh(global_lock);
1135
1136                 /* called with a ref on ni, so shutdown can't have started */
1137                 LASSERT(((struct ksock_net *) ni->ni_data)->ksnn_shutdown == 0);
1138
1139                 peer2 = ksocknal_find_peer_locked(ni, peerid);
1140                 if (peer2 == NULL) {
1141                         /* NB this puts an "empty" peer_ni in the peer_ni
1142                          * table (which takes my ref) */
1143                         list_add_tail(&peer_ni->ksnp_list,
1144                                       ksocknal_nid2peerlist(peerid.nid));
1145                 } else {
1146                         ksocknal_peer_decref(peer_ni);
1147                         peer_ni = peer2;
1148                 }
1149
1150                 /* +1 ref for me */
1151                 ksocknal_peer_addref(peer_ni);
1152                 peer_ni->ksnp_accepting++;
1153
1154                 /* Am I already connecting to this guy?  Resolve in
1155                  * favour of higher NID... */
1156                 if (peerid.nid < ni->ni_nid &&
1157                     ksocknal_connecting(peer_ni, conn->ksnc_ipaddr)) {
1158                         rc = EALREADY;
1159                         warn = "connection race resolution";
1160                         goto failed_2;
1161                 }
1162         }
1163
1164         if (peer_ni->ksnp_closing ||
1165             (active && route->ksnr_deleted)) {
1166                 /* peer_ni/route got closed under me */
1167                 rc = -ESTALE;
1168                 warn = "peer_ni/route removed";
1169                 goto failed_2;
1170         }
1171
1172         if (peer_ni->ksnp_proto == NULL) {
1173                 /* Never connected before.
1174                  * NB recv_hello may have returned EPROTO to signal my peer_ni
1175                  * wants a different protocol than the one I asked for.
1176                  */
1177                 LASSERT(list_empty(&peer_ni->ksnp_conns));
1178
1179                 peer_ni->ksnp_proto = conn->ksnc_proto;
1180                 peer_ni->ksnp_incarnation = incarnation;
1181         }
1182
1183         if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1184             peer_ni->ksnp_incarnation != incarnation) {
1185                 /* peer_ni rebooted or I've got the wrong protocol version */
1186                 ksocknal_close_peer_conns_locked(peer_ni, 0, 0);
1187
1188                 peer_ni->ksnp_proto = NULL;
1189                 rc = ESTALE;
1190                 warn = peer_ni->ksnp_incarnation != incarnation ?
1191                        "peer_ni rebooted" :
1192                        "wrong proto version";
1193                 goto failed_2;
1194         }
1195
1196         switch (rc) {
1197         default:
1198                 LBUG();
1199         case 0:
1200                 break;
1201         case EALREADY:
1202                 warn = "lost conn race";
1203                 goto failed_2;
1204         case EPROTO:
1205                 warn = "retry with different protocol version";
1206                 goto failed_2;
1207         }
1208
1209         /* Refuse to duplicate an existing connection, unless this is a
1210          * loopback connection */
1211         if (conn->ksnc_ipaddr != conn->ksnc_myipaddr) {
1212                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1213                         conn2 = list_entry(tmp, struct ksock_conn, ksnc_list);
1214
1215                         if (conn2->ksnc_ipaddr != conn->ksnc_ipaddr ||
1216                             conn2->ksnc_myipaddr != conn->ksnc_myipaddr ||
1217                             conn2->ksnc_type != conn->ksnc_type)
1218                                 continue;
1219
1220                         /* Reply on a passive connection attempt so the peer_ni
1221                          * realises we're connected. */
1222                         LASSERT (rc == 0);
1223                         if (!active)
1224                                 rc = EALREADY;
1225
1226                         warn = "duplicate";
1227                         goto failed_2;
1228                 }
1229         }
1230
1231         /* If the connection created by this route didn't bind to the IP
1232          * address the route connected to, the connection/route matching
1233          * code below probably isn't going to work. */
1234         if (active &&
1235             route->ksnr_ipaddr != conn->ksnc_ipaddr) {
1236                 CERROR("Route %s %pI4h connected to %pI4h\n",
1237                        libcfs_id2str(peer_ni->ksnp_id),
1238                        &route->ksnr_ipaddr,
1239                        &conn->ksnc_ipaddr);
1240         }
1241
1242         /* Search for a route corresponding to the new connection and
1243          * create an association.  This allows incoming connections created
1244          * by routes in my peer_ni to match my own route entries so I don't
1245          * continually create duplicate routes. */
1246         list_for_each(tmp, &peer_ni->ksnp_routes) {
1247                 route = list_entry(tmp, struct ksock_route, ksnr_list);
1248
1249                 if (route->ksnr_ipaddr != conn->ksnc_ipaddr)
1250                         continue;
1251
1252                 ksocknal_associate_route_conn_locked(route, conn);
1253                 break;
1254         }
1255
1256         conn->ksnc_peer = peer_ni;                 /* conn takes my ref on peer_ni */
1257         peer_ni->ksnp_last_alive = ktime_get_seconds();
1258         peer_ni->ksnp_send_keepalive = 0;
1259         peer_ni->ksnp_error = 0;
1260
1261         sched = ksocknal_choose_scheduler_locked(cpt);
1262         if (!sched) {
1263                 CERROR("no schedulers available. node is unhealthy\n");
1264                 goto failed_2;
1265         }
1266         /*
1267          * The cpt might have changed if we ended up selecting a non cpt
1268          * native scheduler. So use the scheduler's cpt instead.
1269          */
1270         cpt = sched->kss_cpt;
1271         sched->kss_nconns++;
1272         conn->ksnc_scheduler = sched;
1273
1274         conn->ksnc_tx_last_post = ktime_get_seconds();
1275         /* Set the deadline for the outgoing HELLO to drain */
1276         conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1277         conn->ksnc_tx_deadline = ktime_get_seconds() +
1278                                  lnet_get_lnd_timeout();
1279         smp_mb();   /* order with adding to peer_ni's conn list */
1280
1281         list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1282         ksocknal_conn_addref(conn);
1283
1284         ksocknal_new_packet(conn, 0);
1285
1286         conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1287
1288         /* Take packets blocking for this connection. */
1289         list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1290                 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1291                     SOCKNAL_MATCH_NO)
1292                         continue;
1293
1294                 list_del(&tx->tx_list);
1295                 ksocknal_queue_tx_locked(tx, conn);
1296         }
1297
1298         write_unlock_bh(global_lock);
1299
1300         /* We've now got a new connection.  Any errors from here on are just
1301          * like "normal" comms errors and we close the connection normally.
1302          * NB (a) we still have to send the reply HELLO for passive
1303          *        connections,
1304          *    (b) normal I/O on the conn is blocked until I setup and call the
1305          *        socket callbacks.
1306          */
1307
1308         CDEBUG(D_NET, "New conn %s p %d.x %pI4h -> %pI4h/%d"
1309                " incarnation:%lld sched[%d]\n",
1310                libcfs_id2str(peerid), conn->ksnc_proto->pro_version,
1311                &conn->ksnc_myipaddr, &conn->ksnc_ipaddr,
1312                conn->ksnc_port, incarnation, cpt);
1313
1314         if (active) {
1315                 /* additional routes after interface exchange? */
1316                 ksocknal_create_routes(peer_ni, conn->ksnc_port,
1317                                        hello->kshm_ips, hello->kshm_nips);
1318         } else {
1319                 hello->kshm_nips = ksocknal_select_ips(peer_ni, hello->kshm_ips,
1320                                                        hello->kshm_nips);
1321                 rc = ksocknal_send_hello(ni, conn, peerid.nid, hello);
1322         }
1323
1324         LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1325                                     kshm_ips[LNET_INTERFACES_NUM]));
1326
1327         /* setup the socket AFTER I've received hello (it disables
1328          * SO_LINGER).  I might call back to the acceptor who may want
1329          * to send a protocol version response and then close the
1330          * socket; this ensures the socket only tears down after the
1331          * response has been sent. */
1332         if (rc == 0)
1333                 rc = ksocknal_lib_setup_sock(sock);
1334
1335         write_lock_bh(global_lock);
1336
1337         /* NB my callbacks block while I hold ksnd_global_lock */
1338         ksocknal_lib_set_callback(sock, conn);
1339
1340         if (!active)
1341                 peer_ni->ksnp_accepting--;
1342
1343         write_unlock_bh(global_lock);
1344
1345         if (rc != 0) {
1346                 write_lock_bh(global_lock);
1347                 if (!conn->ksnc_closing) {
1348                         /* could be closed by another thread */
1349                         ksocknal_close_conn_locked(conn, rc);
1350                 }
1351                 write_unlock_bh(global_lock);
1352         } else if (ksocknal_connsock_addref(conn) == 0) {
1353                 /* Allow I/O to proceed. */
1354                 ksocknal_read_callback(conn);
1355                 ksocknal_write_callback(conn);
1356                 ksocknal_connsock_decref(conn);
1357         }
1358
1359         ksocknal_connsock_decref(conn);
1360         ksocknal_conn_decref(conn);
1361         return rc;
1362
1363 failed_2:
1364         if (!peer_ni->ksnp_closing &&
1365             list_empty(&peer_ni->ksnp_conns) &&
1366             list_empty(&peer_ni->ksnp_routes)) {
1367                 list_add(&zombies, &peer_ni->ksnp_tx_queue);
1368                 list_del_init(&peer_ni->ksnp_tx_queue);
1369                 ksocknal_unlink_peer_locked(peer_ni);
1370         }
1371
1372         write_unlock_bh(global_lock);
1373
1374         if (warn != NULL) {
1375                 if (rc < 0)
1376                         CERROR("Not creating conn %s type %d: %s\n",
1377                                libcfs_id2str(peerid), conn->ksnc_type, warn);
1378                 else
1379                         CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1380                               libcfs_id2str(peerid), conn->ksnc_type, warn);
1381         }
1382
1383         if (!active) {
1384                 if (rc > 0) {
1385                         /* Request retry by replying with CONN_NONE
1386                          * ksnc_proto has been set already */
1387                         conn->ksnc_type = SOCKLND_CONN_NONE;
1388                         hello->kshm_nips = 0;
1389                         ksocknal_send_hello(ni, conn, peerid.nid, hello);
1390                 }
1391
1392                 write_lock_bh(global_lock);
1393                 peer_ni->ksnp_accepting--;
1394                 write_unlock_bh(global_lock);
1395         }
1396
1397         /*
1398          * If we get here without an error code, just use -EALREADY.
1399          * Depending on how we got here, the error may be positive
1400          * or negative. Normalize the value for ksocknal_txlist_done().
1401          */
1402         rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1403         ksocknal_txlist_done(ni, &zombies, rc2);
1404         ksocknal_peer_decref(peer_ni);
1405
1406 failed_1:
1407         if (hello != NULL)
1408                 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1409                                             kshm_ips[LNET_INTERFACES_NUM]));
1410
1411         LIBCFS_FREE(conn, sizeof(*conn));
1412
1413 failed_0:
1414         sock_release(sock);
1415         return rc;
1416 }
1417
1418 void
1419 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1420 {
1421         /* This just does the immmediate housekeeping, and queues the
1422          * connection for the reaper to terminate.
1423          * Caller holds ksnd_global_lock exclusively in irq context */
1424         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1425         struct ksock_route *route;
1426         struct ksock_conn *conn2;
1427         struct list_head *tmp;
1428
1429         LASSERT(peer_ni->ksnp_error == 0);
1430         LASSERT(!conn->ksnc_closing);
1431         conn->ksnc_closing = 1;
1432
1433         /* ksnd_deathrow_conns takes over peer_ni's ref */
1434         list_del(&conn->ksnc_list);
1435
1436         route = conn->ksnc_route;
1437         if (route != NULL) {
1438                 /* dissociate conn from route... */
1439                 LASSERT(!route->ksnr_deleted);
1440                 LASSERT((route->ksnr_connected & (1 << conn->ksnc_type)) != 0);
1441
1442                 conn2 = NULL;
1443                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1444                         conn2 = list_entry(tmp, struct ksock_conn, ksnc_list);
1445
1446                         if (conn2->ksnc_route == route &&
1447                             conn2->ksnc_type == conn->ksnc_type)
1448                                 break;
1449
1450                         conn2 = NULL;
1451                 }
1452                 if (conn2 == NULL)
1453                         route->ksnr_connected &= ~(1 << conn->ksnc_type);
1454
1455                 conn->ksnc_route = NULL;
1456
1457                 ksocknal_route_decref(route);   /* drop conn's ref on route */
1458         }
1459
1460         if (list_empty(&peer_ni->ksnp_conns)) {
1461                 /* No more connections to this peer_ni */
1462
1463                 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1464                                 struct ksock_tx *tx;
1465
1466                         LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1467
1468                         /* throw them to the last connection...,
1469                          * these TXs will be send to /dev/null by scheduler */
1470                         list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1471                                             tx_list)
1472                                 ksocknal_tx_prep(conn, tx);
1473
1474                         spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1475                         list_splice_init(&peer_ni->ksnp_tx_queue,
1476                                          &conn->ksnc_tx_queue);
1477                         spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1478                 }
1479
1480                 /* renegotiate protocol version */
1481                 peer_ni->ksnp_proto = NULL;
1482                 /* stash last conn close reason */
1483                 peer_ni->ksnp_error = error;
1484
1485                 if (list_empty(&peer_ni->ksnp_routes)) {
1486                         /* I've just closed last conn belonging to a
1487                          * peer_ni with no routes to it */
1488                         ksocknal_unlink_peer_locked(peer_ni);
1489                 }
1490         }
1491
1492         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1493
1494         list_add_tail(&conn->ksnc_list,
1495                       &ksocknal_data.ksnd_deathrow_conns);
1496         wake_up(&ksocknal_data.ksnd_reaper_waitq);
1497
1498         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1499 }
1500
1501 void
1502 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1503 {
1504         int notify = 0;
1505         time64_t last_alive = 0;
1506
1507         /* There has been a connection failure or comms error; but I'll only
1508          * tell LNET I think the peer_ni is dead if it's to another kernel and
1509          * there are no connections or connection attempts in existence. */
1510
1511         read_lock(&ksocknal_data.ksnd_global_lock);
1512
1513         if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1514              list_empty(&peer_ni->ksnp_conns) &&
1515              peer_ni->ksnp_accepting == 0 &&
1516              ksocknal_find_connecting_route_locked(peer_ni) == NULL) {
1517                 notify = 1;
1518                 last_alive = peer_ni->ksnp_last_alive;
1519         }
1520
1521         read_unlock(&ksocknal_data.ksnd_global_lock);
1522
1523         if (notify)
1524                 lnet_notify(peer_ni->ksnp_ni, peer_ni->ksnp_id.nid,
1525                             false, false, last_alive);
1526 }
1527
1528 void
1529 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1530 {
1531         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1532         struct ksock_tx *tx;
1533         struct ksock_tx *tmp;
1534         struct list_head zlist = LIST_HEAD_INIT(zlist);
1535
1536         /* NB safe to finalize TXs because closing of socket will
1537          * abort all buffered data */
1538         LASSERT(conn->ksnc_sock == NULL);
1539
1540         spin_lock(&peer_ni->ksnp_lock);
1541
1542         list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list, tx_zc_list) {
1543                 if (tx->tx_conn != conn)
1544                         continue;
1545
1546                 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1547
1548                 tx->tx_msg.ksm_zc_cookies[0] = 0;
1549                 tx->tx_zc_aborted = 1;  /* mark it as not-acked */
1550                 list_del(&tx->tx_zc_list);
1551                 list_add(&tx->tx_zc_list, &zlist);
1552         }
1553
1554         spin_unlock(&peer_ni->ksnp_lock);
1555
1556         while (!list_empty(&zlist)) {
1557                 tx = list_entry(zlist.next, struct ksock_tx, tx_zc_list);
1558
1559                 list_del(&tx->tx_zc_list);
1560                 ksocknal_tx_decref(tx);
1561         }
1562 }
1563
1564 void
1565 ksocknal_terminate_conn(struct ksock_conn *conn)
1566 {
1567         /* This gets called by the reaper (guaranteed thread context) to
1568          * disengage the socket from its callbacks and close it.
1569          * ksnc_refcount will eventually hit zero, and then the reaper will
1570          * destroy it. */
1571         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1572         struct ksock_sched *sched = conn->ksnc_scheduler;
1573         int failed = 0;
1574
1575         LASSERT(conn->ksnc_closing);
1576
1577         /* wake up the scheduler to "send" all remaining packets to /dev/null */
1578         spin_lock_bh(&sched->kss_lock);
1579
1580         /* a closing conn is always ready to tx */
1581         conn->ksnc_tx_ready = 1;
1582
1583         if (!conn->ksnc_tx_scheduled &&
1584             !list_empty(&conn->ksnc_tx_queue)) {
1585                 list_add_tail(&conn->ksnc_tx_list,
1586                                &sched->kss_tx_conns);
1587                 conn->ksnc_tx_scheduled = 1;
1588                 /* extra ref for scheduler */
1589                 ksocknal_conn_addref(conn);
1590
1591                 wake_up (&sched->kss_waitq);
1592         }
1593
1594         spin_unlock_bh(&sched->kss_lock);
1595
1596         /* serialise with callbacks */
1597         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1598
1599         ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1600
1601         /* OK, so this conn may not be completely disengaged from its
1602          * scheduler yet, but it _has_ committed to terminate... */
1603         conn->ksnc_scheduler->kss_nconns--;
1604
1605         if (peer_ni->ksnp_error != 0) {
1606                 /* peer_ni's last conn closed in error */
1607                 LASSERT(list_empty(&peer_ni->ksnp_conns));
1608                 failed = 1;
1609                 peer_ni->ksnp_error = 0;     /* avoid multiple notifications */
1610         }
1611
1612         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1613
1614         if (failed)
1615                 ksocknal_peer_failed(peer_ni);
1616
1617         /* The socket is closed on the final put; either here, or in
1618          * ksocknal_{send,recv}msg().  Since we set up the linger2 option
1619          * when the connection was established, this will close the socket
1620          * immediately, aborting anything buffered in it. Any hung
1621          * zero-copy transmits will therefore complete in finite time. */
1622         ksocknal_connsock_decref(conn);
1623 }
1624
1625 void
1626 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1627 {
1628         /* Queue the conn for the reaper to destroy */
1629         LASSERT(atomic_read(&conn->ksnc_conn_refcount) == 0);
1630         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1631
1632         list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1633         wake_up(&ksocknal_data.ksnd_reaper_waitq);
1634
1635         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1636 }
1637
1638 void
1639 ksocknal_destroy_conn(struct ksock_conn *conn)
1640 {
1641         time64_t last_rcv;
1642
1643         /* Final coup-de-grace of the reaper */
1644         CDEBUG (D_NET, "connection %p\n", conn);
1645
1646         LASSERT (atomic_read (&conn->ksnc_conn_refcount) == 0);
1647         LASSERT (atomic_read (&conn->ksnc_sock_refcount) == 0);
1648         LASSERT (conn->ksnc_sock == NULL);
1649         LASSERT (conn->ksnc_route == NULL);
1650         LASSERT (!conn->ksnc_tx_scheduled);
1651         LASSERT (!conn->ksnc_rx_scheduled);
1652         LASSERT(list_empty(&conn->ksnc_tx_queue));
1653
1654         /* complete current receive if any */
1655         switch (conn->ksnc_rx_state) {
1656         case SOCKNAL_RX_LNET_PAYLOAD:
1657                 last_rcv = conn->ksnc_rx_deadline -
1658                            lnet_get_lnd_timeout();
1659                 CERROR("Completing partial receive from %s[%d], "
1660                        "ip %pI4h:%d, with error, wanted: %d, left: %d, "
1661                        "last alive is %lld secs ago\n",
1662                        libcfs_id2str(conn->ksnc_peer->ksnp_id), conn->ksnc_type,
1663                        &conn->ksnc_ipaddr, conn->ksnc_port,
1664                        conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1665                        ktime_get_seconds() - last_rcv);
1666                 if (conn->ksnc_lnet_msg)
1667                         conn->ksnc_lnet_msg->msg_health_status =
1668                                 LNET_MSG_STATUS_REMOTE_ERROR;
1669                 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1670                 break;
1671         case SOCKNAL_RX_LNET_HEADER:
1672                 if (conn->ksnc_rx_started)
1673                         CERROR("Incomplete receive of lnet header from %s, "
1674                                "ip %pI4h:%d, with error, protocol: %d.x.\n",
1675                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1676                                &conn->ksnc_ipaddr, conn->ksnc_port,
1677                                conn->ksnc_proto->pro_version);
1678                 break;
1679         case SOCKNAL_RX_KSM_HEADER:
1680                 if (conn->ksnc_rx_started)
1681                         CERROR("Incomplete receive of ksock message from %s, "
1682                                "ip %pI4h:%d, with error, protocol: %d.x.\n",
1683                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1684                                &conn->ksnc_ipaddr, conn->ksnc_port,
1685                                conn->ksnc_proto->pro_version);
1686                 break;
1687         case SOCKNAL_RX_SLOP:
1688                 if (conn->ksnc_rx_started)
1689                         CERROR("Incomplete receive of slops from %s, "
1690                                "ip %pI4h:%d, with error\n",
1691                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1692                                &conn->ksnc_ipaddr, conn->ksnc_port);
1693                break;
1694         default:
1695                 LBUG ();
1696                 break;
1697         }
1698
1699         ksocknal_peer_decref(conn->ksnc_peer);
1700
1701         LIBCFS_FREE (conn, sizeof (*conn));
1702 }
1703
1704 int
1705 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni, __u32 ipaddr, int why)
1706 {
1707         struct ksock_conn *conn;
1708         struct list_head *ctmp;
1709         struct list_head *cnxt;
1710         int count = 0;
1711
1712         list_for_each_safe(ctmp, cnxt, &peer_ni->ksnp_conns) {
1713                 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
1714
1715                 if (ipaddr == 0 ||
1716                     conn->ksnc_ipaddr == ipaddr) {
1717                         count++;
1718                         ksocknal_close_conn_locked (conn, why);
1719                 }
1720         }
1721
1722         return (count);
1723 }
1724
1725 int
1726 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1727 {
1728         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1729         u32 ipaddr = conn->ksnc_ipaddr;
1730         int count;
1731
1732         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1733
1734         count = ksocknal_close_peer_conns_locked (peer_ni, ipaddr, why);
1735
1736         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1737
1738         return (count);
1739 }
1740
1741 int
1742 ksocknal_close_matching_conns(struct lnet_process_id id, __u32 ipaddr)
1743 {
1744         struct ksock_peer_ni *peer_ni;
1745         struct list_head *ptmp;
1746         struct list_head *pnxt;
1747         int lo;
1748         int hi;
1749         int i;
1750         int count = 0;
1751
1752         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1753
1754         if (id.nid != LNET_NID_ANY)
1755                 lo = hi = (int)(ksocknal_nid2peerlist(id.nid) - ksocknal_data.ksnd_peers);
1756         else {
1757                 lo = 0;
1758                 hi = ksocknal_data.ksnd_peer_hash_size - 1;
1759         }
1760
1761         for (i = lo; i <= hi; i++) {
1762                 list_for_each_safe(ptmp, pnxt, &ksocknal_data.ksnd_peers[i]) {
1763
1764                         peer_ni = list_entry(ptmp, struct ksock_peer_ni, ksnp_list);
1765
1766                         if (!((id.nid == LNET_NID_ANY || id.nid == peer_ni->ksnp_id.nid) &&
1767                               (id.pid == LNET_PID_ANY || id.pid == peer_ni->ksnp_id.pid)))
1768                                 continue;
1769
1770                         count += ksocknal_close_peer_conns_locked (peer_ni, ipaddr, 0);
1771                 }
1772         }
1773
1774         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1775
1776         /* wildcards always succeed */
1777         if (id.nid == LNET_NID_ANY || id.pid == LNET_PID_ANY || ipaddr == 0)
1778                 return (0);
1779
1780         return (count == 0 ? -ENOENT : 0);
1781 }
1782
1783 void
1784 ksocknal_notify_gw_down(lnet_nid_t gw_nid)
1785 {
1786         /* The router is telling me she's been notified of a change in
1787          * gateway state....
1788          */
1789         struct lnet_process_id id = {
1790                 .nid    = gw_nid,
1791                 .pid    = LNET_PID_ANY,
1792         };
1793
1794         CDEBUG(D_NET, "gw %s down\n", libcfs_nid2str(gw_nid));
1795
1796         /* If the gateway crashed, close all open connections... */
1797         ksocknal_close_matching_conns(id, 0);
1798         return;
1799
1800         /* We can only establish new connections
1801          * if we have autroutes, and these connect on demand. */
1802 }
1803
1804 void
1805 ksocknal_query(struct lnet_ni *ni, lnet_nid_t nid, time64_t *when)
1806 {
1807         int connect = 1;
1808         time64_t last_alive = 0;
1809         time64_t now = ktime_get_seconds();
1810         struct ksock_peer_ni *peer_ni = NULL;
1811         rwlock_t *glock = &ksocknal_data.ksnd_global_lock;
1812         struct lnet_process_id id = {
1813                 .nid = nid,
1814                 .pid = LNET_PID_LUSTRE,
1815         };
1816
1817         read_lock(glock);
1818
1819         peer_ni = ksocknal_find_peer_locked(ni, id);
1820         if (peer_ni != NULL) {
1821                 struct list_head *tmp;
1822                 struct ksock_conn *conn;
1823                 int bufnob;
1824
1825                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1826                         conn = list_entry(tmp, struct ksock_conn, ksnc_list);
1827                         bufnob = conn->ksnc_sock->sk->sk_wmem_queued;
1828
1829                         if (bufnob < conn->ksnc_tx_bufnob) {
1830                                 /* something got ACKed */
1831                                 conn->ksnc_tx_deadline = ktime_get_seconds() +
1832                                                          lnet_get_lnd_timeout();
1833                                 peer_ni->ksnp_last_alive = now;
1834                                 conn->ksnc_tx_bufnob = bufnob;
1835                         }
1836                 }
1837
1838                 last_alive = peer_ni->ksnp_last_alive;
1839                 if (ksocknal_find_connectable_route_locked(peer_ni) == NULL)
1840                         connect = 0;
1841         }
1842
1843         read_unlock(glock);
1844
1845         if (last_alive != 0)
1846                 *when = last_alive;
1847
1848         CDEBUG(D_NET, "peer_ni %s %p, alive %lld secs ago, connect %d\n",
1849                libcfs_nid2str(nid), peer_ni,
1850                last_alive ? now - last_alive : -1,
1851                connect);
1852
1853         if (!connect)
1854                 return;
1855
1856         ksocknal_add_peer(ni, id, LNET_NIDADDR(nid), lnet_acceptor_port());
1857
1858         write_lock_bh(glock);
1859
1860         peer_ni = ksocknal_find_peer_locked(ni, id);
1861         if (peer_ni != NULL)
1862                 ksocknal_launch_all_connections_locked(peer_ni);
1863
1864         write_unlock_bh(glock);
1865         return;
1866 }
1867
1868 static void
1869 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1870 {
1871         int index;
1872         int i;
1873         struct list_head *tmp;
1874         struct ksock_conn *conn;
1875
1876         for (index = 0; ; index++) {
1877                 read_lock(&ksocknal_data.ksnd_global_lock);
1878
1879                 i = 0;
1880                 conn = NULL;
1881
1882                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1883                         if (i++ == index) {
1884                                 conn = list_entry(tmp, struct ksock_conn,
1885                                                   ksnc_list);
1886                                 ksocknal_conn_addref(conn);
1887                                 break;
1888                         }
1889                 }
1890
1891                 read_unlock(&ksocknal_data.ksnd_global_lock);
1892
1893                 if (conn == NULL)
1894                         break;
1895
1896                 ksocknal_lib_push_conn (conn);
1897                 ksocknal_conn_decref(conn);
1898         }
1899 }
1900
1901 static int
1902 ksocknal_push(struct lnet_ni *ni, struct lnet_process_id id)
1903 {
1904         struct list_head *start;
1905         struct list_head *end;
1906         struct list_head *tmp;
1907         int               rc = -ENOENT;
1908         unsigned int      hsize = ksocknal_data.ksnd_peer_hash_size;
1909
1910         if (id.nid == LNET_NID_ANY) {
1911                 start = &ksocknal_data.ksnd_peers[0];
1912                 end = &ksocknal_data.ksnd_peers[hsize - 1];
1913         } else {
1914                 start = end = ksocknal_nid2peerlist(id.nid);
1915         }
1916
1917         for (tmp = start; tmp <= end; tmp++) {
1918                 int     peer_off; /* searching offset in peer_ni hash table */
1919
1920                 for (peer_off = 0; ; peer_off++) {
1921                         struct ksock_peer_ni *peer_ni;
1922                         int           i = 0;
1923
1924                         read_lock(&ksocknal_data.ksnd_global_lock);
1925                         list_for_each_entry(peer_ni, tmp, ksnp_list) {
1926                                 if (!((id.nid == LNET_NID_ANY ||
1927                                        id.nid == peer_ni->ksnp_id.nid) &&
1928                                       (id.pid == LNET_PID_ANY ||
1929                                        id.pid == peer_ni->ksnp_id.pid)))
1930                                         continue;
1931
1932                                 if (i++ == peer_off) {
1933                                         ksocknal_peer_addref(peer_ni);
1934                                         break;
1935                                 }
1936                         }
1937                         read_unlock(&ksocknal_data.ksnd_global_lock);
1938
1939                         if (i == 0) /* no match */
1940                                 break;
1941
1942                         rc = 0;
1943                         ksocknal_push_peer(peer_ni);
1944                         ksocknal_peer_decref(peer_ni);
1945                 }
1946         }
1947         return rc;
1948 }
1949
1950 static int
1951 ksocknal_add_interface(struct lnet_ni *ni, __u32 ipaddress, __u32 netmask)
1952 {
1953         struct ksock_net *net = ni->ni_data;
1954         struct ksock_interface *iface;
1955         int rc;
1956         int i;
1957         int j;
1958         struct list_head *ptmp;
1959         struct ksock_peer_ni *peer_ni;
1960         struct list_head *rtmp;
1961         struct ksock_route *route;
1962
1963         if (ipaddress == 0 ||
1964             netmask == 0)
1965                 return -EINVAL;
1966
1967         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1968
1969         iface = ksocknal_ip2iface(ni, ipaddress);
1970         if (iface != NULL) {
1971                 /* silently ignore dups */
1972                 rc = 0;
1973         } else if (net->ksnn_ninterfaces == LNET_INTERFACES_NUM) {
1974                 rc = -ENOSPC;
1975         } else {
1976                 iface = &net->ksnn_interfaces[net->ksnn_ninterfaces++];
1977
1978                 iface->ksni_ipaddr = ipaddress;
1979                 iface->ksni_netmask = netmask;
1980                 iface->ksni_nroutes = 0;
1981                 iface->ksni_npeers = 0;
1982
1983                 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
1984                         list_for_each(ptmp, &ksocknal_data.ksnd_peers[i]) {
1985                                 peer_ni = list_entry(ptmp, struct ksock_peer_ni,
1986                                                      ksnp_list);
1987
1988                                 for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++)
1989                                         if (peer_ni->ksnp_passive_ips[j] == ipaddress)
1990                                                 iface->ksni_npeers++;
1991
1992                                 list_for_each(rtmp, &peer_ni->ksnp_routes) {
1993                                         route = list_entry(rtmp,
1994                                                            struct ksock_route,
1995                                                            ksnr_list);
1996
1997                                         if (route->ksnr_myipaddr == ipaddress)
1998                                                 iface->ksni_nroutes++;
1999                                 }
2000                         }
2001                 }
2002
2003                 rc = 0;
2004                 /* NB only new connections will pay attention to the new interface! */
2005         }
2006
2007         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2008
2009         return rc;
2010 }
2011
2012 static void
2013 ksocknal_peer_del_interface_locked(struct ksock_peer_ni *peer_ni, __u32 ipaddr)
2014 {
2015         struct list_head *tmp;
2016         struct list_head *nxt;
2017         struct ksock_route *route;
2018         struct ksock_conn *conn;
2019         int i;
2020         int j;
2021
2022         for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++)
2023                 if (peer_ni->ksnp_passive_ips[i] == ipaddr) {
2024                         for (j = i+1; j < peer_ni->ksnp_n_passive_ips; j++)
2025                                 peer_ni->ksnp_passive_ips[j-1] =
2026                                         peer_ni->ksnp_passive_ips[j];
2027                         peer_ni->ksnp_n_passive_ips--;
2028                         break;
2029                 }
2030
2031         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
2032                 route = list_entry(tmp, struct ksock_route, ksnr_list);
2033
2034                 if (route->ksnr_myipaddr != ipaddr)
2035                         continue;
2036
2037                 if (route->ksnr_share_count != 0) {
2038                         /* Manually created; keep, but unbind */
2039                         route->ksnr_myipaddr = 0;
2040                 } else {
2041                         ksocknal_del_route_locked(route);
2042                 }
2043         }
2044
2045         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_conns) {
2046                 conn = list_entry(tmp, struct ksock_conn, ksnc_list);
2047
2048                 if (conn->ksnc_myipaddr == ipaddr)
2049                         ksocknal_close_conn_locked (conn, 0);
2050         }
2051 }
2052
2053 static int
2054 ksocknal_del_interface(struct lnet_ni *ni, __u32 ipaddress)
2055 {
2056         struct ksock_net *net = ni->ni_data;
2057         int rc = -ENOENT;
2058         struct list_head *tmp;
2059         struct list_head *nxt;
2060         struct ksock_peer_ni *peer_ni;
2061         u32 this_ip;
2062         int i;
2063         int j;
2064
2065         write_lock_bh(&ksocknal_data.ksnd_global_lock);
2066
2067         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2068                 this_ip = net->ksnn_interfaces[i].ksni_ipaddr;
2069
2070                 if (!(ipaddress == 0 ||
2071                       ipaddress == this_ip))
2072                         continue;
2073
2074                 rc = 0;
2075
2076                 for (j = i+1; j < net->ksnn_ninterfaces; j++)
2077                         net->ksnn_interfaces[j-1] =
2078                                 net->ksnn_interfaces[j];
2079
2080                 net->ksnn_ninterfaces--;
2081
2082                 for (j = 0; j < ksocknal_data.ksnd_peer_hash_size; j++) {
2083                         list_for_each_safe(tmp, nxt,
2084                                            &ksocknal_data.ksnd_peers[j]) {
2085                                 peer_ni = list_entry(tmp, struct ksock_peer_ni,
2086                                                      ksnp_list);
2087
2088                                 if (peer_ni->ksnp_ni != ni)
2089                                         continue;
2090
2091                                 ksocknal_peer_del_interface_locked(peer_ni, this_ip);
2092                         }
2093                 }
2094         }
2095
2096         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2097
2098         return (rc);
2099 }
2100
2101 int
2102 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
2103 {
2104         struct lnet_process_id id = {0};
2105         struct libcfs_ioctl_data *data = arg;
2106         int rc;
2107
2108         switch(cmd) {
2109         case IOC_LIBCFS_GET_INTERFACE: {
2110                 struct ksock_net *net = ni->ni_data;
2111                 struct ksock_interface *iface;
2112
2113                 read_lock(&ksocknal_data.ksnd_global_lock);
2114
2115                 if (data->ioc_count >= (__u32)net->ksnn_ninterfaces) {
2116                         rc = -ENOENT;
2117                 } else {
2118                         rc = 0;
2119                         iface = &net->ksnn_interfaces[data->ioc_count];
2120
2121                         data->ioc_u32[0] = iface->ksni_ipaddr;
2122                         data->ioc_u32[1] = iface->ksni_netmask;
2123                         data->ioc_u32[2] = iface->ksni_npeers;
2124                         data->ioc_u32[3] = iface->ksni_nroutes;
2125                 }
2126
2127                 read_unlock(&ksocknal_data.ksnd_global_lock);
2128                 return rc;
2129         }
2130
2131         case IOC_LIBCFS_ADD_INTERFACE:
2132                 return ksocknal_add_interface(ni,
2133                                               data->ioc_u32[0], /* IP address */
2134                                               data->ioc_u32[1]); /* net mask */
2135
2136         case IOC_LIBCFS_DEL_INTERFACE:
2137                 return ksocknal_del_interface(ni,
2138                                               data->ioc_u32[0]); /* IP address */
2139
2140         case IOC_LIBCFS_GET_PEER: {
2141                 __u32            myip = 0;
2142                 __u32            ip = 0;
2143                 int              port = 0;
2144                 int              conn_count = 0;
2145                 int              share_count = 0;
2146
2147                 rc = ksocknal_get_peer_info(ni, data->ioc_count,
2148                                             &id, &myip, &ip, &port,
2149                                             &conn_count,  &share_count);
2150                 if (rc != 0)
2151                         return rc;
2152
2153                 data->ioc_nid    = id.nid;
2154                 data->ioc_count  = share_count;
2155                 data->ioc_u32[0] = ip;
2156                 data->ioc_u32[1] = port;
2157                 data->ioc_u32[2] = myip;
2158                 data->ioc_u32[3] = conn_count;
2159                 data->ioc_u32[4] = id.pid;
2160                 return 0;
2161         }
2162
2163         case IOC_LIBCFS_ADD_PEER:
2164                 id.nid = data->ioc_nid;
2165                 id.pid = LNET_PID_LUSTRE;
2166                 return ksocknal_add_peer (ni, id,
2167                                           data->ioc_u32[0], /* IP */
2168                                           data->ioc_u32[1]); /* port */
2169
2170         case IOC_LIBCFS_DEL_PEER:
2171                 id.nid = data->ioc_nid;
2172                 id.pid = LNET_PID_ANY;
2173                 return ksocknal_del_peer (ni, id,
2174                                           data->ioc_u32[0]); /* IP */
2175
2176         case IOC_LIBCFS_GET_CONN: {
2177                 int           txmem;
2178                 int           rxmem;
2179                 int           nagle;
2180                 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
2181
2182                 if (conn == NULL)
2183                         return -ENOENT;
2184
2185                 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
2186
2187                 data->ioc_count  = txmem;
2188                 data->ioc_nid    = conn->ksnc_peer->ksnp_id.nid;
2189                 data->ioc_flags  = nagle;
2190                 data->ioc_u32[0] = conn->ksnc_ipaddr;
2191                 data->ioc_u32[1] = conn->ksnc_port;
2192                 data->ioc_u32[2] = conn->ksnc_myipaddr;
2193                 data->ioc_u32[3] = conn->ksnc_type;
2194                 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
2195                 data->ioc_u32[5] = rxmem;
2196                 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
2197                 ksocknal_conn_decref(conn);
2198                 return 0;
2199         }
2200
2201         case IOC_LIBCFS_CLOSE_CONNECTION:
2202                 id.nid = data->ioc_nid;
2203                 id.pid = LNET_PID_ANY;
2204                 return ksocknal_close_matching_conns (id,
2205                                                       data->ioc_u32[0]);
2206
2207         case IOC_LIBCFS_REGISTER_MYNID:
2208                 /* Ignore if this is a noop */
2209                 if (data->ioc_nid == ni->ni_nid)
2210                         return 0;
2211
2212                 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
2213                        libcfs_nid2str(data->ioc_nid),
2214                        libcfs_nid2str(ni->ni_nid));
2215                 return -EINVAL;
2216
2217         case IOC_LIBCFS_PUSH_CONNECTION:
2218                 id.nid = data->ioc_nid;
2219                 id.pid = LNET_PID_ANY;
2220                 return ksocknal_push(ni, id);
2221
2222         default:
2223                 return -EINVAL;
2224         }
2225         /* not reached */
2226 }
2227
2228 static void
2229 ksocknal_free_buffers (void)
2230 {
2231         LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
2232
2233         if (ksocknal_data.ksnd_schedulers != NULL)
2234                 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
2235
2236         LIBCFS_FREE (ksocknal_data.ksnd_peers,
2237                      sizeof(struct list_head) *
2238                      ksocknal_data.ksnd_peer_hash_size);
2239
2240         spin_lock(&ksocknal_data.ksnd_tx_lock);
2241
2242         if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
2243                 struct list_head zlist;
2244                 struct ksock_tx *tx;
2245
2246                 list_add(&zlist, &ksocknal_data.ksnd_idle_noop_txs);
2247                 list_del_init(&ksocknal_data.ksnd_idle_noop_txs);
2248                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2249
2250                 while (!list_empty(&zlist)) {
2251                         tx = list_entry(zlist.next, struct ksock_tx, tx_list);
2252                         list_del(&tx->tx_list);
2253                         LIBCFS_FREE(tx, tx->tx_desc_size);
2254                 }
2255         } else {
2256                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2257         }
2258 }
2259
2260 static void
2261 ksocknal_base_shutdown(void)
2262 {
2263         struct ksock_sched *sched;
2264         int i;
2265
2266         CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2267                atomic_read (&libcfs_kmemory));
2268         LASSERT (ksocknal_data.ksnd_nnets == 0);
2269
2270         switch (ksocknal_data.ksnd_init) {
2271         default:
2272                 LASSERT (0);
2273
2274         case SOCKNAL_INIT_ALL:
2275         case SOCKNAL_INIT_DATA:
2276                 LASSERT (ksocknal_data.ksnd_peers != NULL);
2277                 for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2278                         LASSERT(list_empty(&ksocknal_data.ksnd_peers[i]));
2279                 }
2280
2281                 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2282                 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2283                 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2284                 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2285                 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2286
2287                 if (ksocknal_data.ksnd_schedulers != NULL) {
2288                         cfs_percpt_for_each(sched, i,
2289                                             ksocknal_data.ksnd_schedulers) {
2290
2291                                 LASSERT(list_empty(&sched->kss_tx_conns));
2292                                 LASSERT(list_empty(&sched->kss_rx_conns));
2293                                 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
2294                                 LASSERT(sched->kss_nconns == 0);
2295                         }
2296                 }
2297
2298                 /* flag threads to terminate; wake and wait for them to die */
2299                 ksocknal_data.ksnd_shuttingdown = 1;
2300                 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2301                 wake_up_all(&ksocknal_data.ksnd_reaper_waitq);
2302
2303                 if (ksocknal_data.ksnd_schedulers != NULL) {
2304                         cfs_percpt_for_each(sched, i,
2305                                             ksocknal_data.ksnd_schedulers)
2306                                         wake_up_all(&sched->kss_waitq);
2307                 }
2308
2309                 i = 4;
2310                 read_lock(&ksocknal_data.ksnd_global_lock);
2311                 while (ksocknal_data.ksnd_nthreads != 0) {
2312                         i++;
2313                         /* power of 2? */
2314                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2315                                 "waiting for %d threads to terminate\n",
2316                                 ksocknal_data.ksnd_nthreads);
2317                         read_unlock(&ksocknal_data.ksnd_global_lock);
2318                         set_current_state(TASK_UNINTERRUPTIBLE);
2319                         schedule_timeout(cfs_time_seconds(1));
2320                         read_lock(&ksocknal_data.ksnd_global_lock);
2321                 }
2322                 read_unlock(&ksocknal_data.ksnd_global_lock);
2323
2324                 ksocknal_free_buffers();
2325
2326                 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2327                 break;
2328         }
2329
2330         CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2331                atomic_read (&libcfs_kmemory));
2332
2333         module_put(THIS_MODULE);
2334 }
2335
2336 static int
2337 ksocknal_base_startup(void)
2338 {
2339         struct ksock_sched *sched;
2340         int rc;
2341         int i;
2342
2343         LASSERT (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2344         LASSERT (ksocknal_data.ksnd_nnets == 0);
2345
2346         memset (&ksocknal_data, 0, sizeof (ksocknal_data)); /* zero pointers */
2347
2348         ksocknal_data.ksnd_peer_hash_size = SOCKNAL_PEER_HASH_SIZE;
2349         LIBCFS_ALLOC(ksocknal_data.ksnd_peers,
2350                      sizeof(struct list_head) *
2351                      ksocknal_data.ksnd_peer_hash_size);
2352         if (ksocknal_data.ksnd_peers == NULL)
2353                 return -ENOMEM;
2354
2355         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++)
2356                 INIT_LIST_HEAD(&ksocknal_data.ksnd_peers[i]);
2357
2358         rwlock_init(&ksocknal_data.ksnd_global_lock);
2359         INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2360
2361         spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2362         INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2363         INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2364         INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2365         init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2366
2367         spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2368         INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2369         INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2370         init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2371
2372         spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2373         INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2374
2375         /* NB memset above zeros whole of ksocknal_data */
2376
2377         /* flag lists/ptrs/locks initialised */
2378         ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2379         try_module_get(THIS_MODULE);
2380
2381         /* Create a scheduler block per available CPT */
2382         ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2383                                                          sizeof(*sched));
2384         if (ksocknal_data.ksnd_schedulers == NULL)
2385                 goto failed;
2386
2387         cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2388                 int nthrs;
2389
2390                 /*
2391                  * make sure not to allocate more threads than there are
2392                  * cores/CPUs in teh CPT
2393                  */
2394                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2395                 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2396                         nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2397                 } else {
2398                         /*
2399                          * max to half of CPUs, assume another half should be
2400                          * reserved for upper layer modules
2401                          */
2402                         nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2403                 }
2404
2405                 sched->kss_nthreads_max = nthrs;
2406                 sched->kss_cpt = i;
2407
2408                 spin_lock_init(&sched->kss_lock);
2409                 INIT_LIST_HEAD(&sched->kss_rx_conns);
2410                 INIT_LIST_HEAD(&sched->kss_tx_conns);
2411                 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2412                 init_waitqueue_head(&sched->kss_waitq);
2413         }
2414
2415         ksocknal_data.ksnd_connd_starting         = 0;
2416         ksocknal_data.ksnd_connd_failed_stamp     = 0;
2417         ksocknal_data.ksnd_connd_starting_stamp   = ktime_get_real_seconds();
2418         /* must have at least 2 connds to remain responsive to accepts while
2419          * connecting */
2420         if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2421                 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2422
2423         if (*ksocknal_tunables.ksnd_nconnds_max <
2424             *ksocknal_tunables.ksnd_nconnds) {
2425                 ksocknal_tunables.ksnd_nconnds_max =
2426                         ksocknal_tunables.ksnd_nconnds;
2427         }
2428
2429         for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2430                 char name[16];
2431                 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2432                 ksocknal_data.ksnd_connd_starting++;
2433                 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2434
2435
2436                 snprintf(name, sizeof(name), "socknal_cd%02d", i);
2437                 rc = ksocknal_thread_start(ksocknal_connd,
2438                                            (void *)((uintptr_t)i), name);
2439                 if (rc != 0) {
2440                         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2441                         ksocknal_data.ksnd_connd_starting--;
2442                         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2443                         CERROR("Can't spawn socknal connd: %d\n", rc);
2444                         goto failed;
2445                 }
2446         }
2447
2448         rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2449         if (rc != 0) {
2450                 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2451                 goto failed;
2452         }
2453
2454         /* flag everything initialised */
2455         ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2456
2457         return 0;
2458
2459  failed:
2460         ksocknal_base_shutdown();
2461         return -ENETDOWN;
2462 }
2463
2464 static void
2465 ksocknal_debug_peerhash(struct lnet_ni *ni)
2466 {
2467         struct ksock_peer_ni *peer_ni = NULL;
2468         struct list_head *tmp;
2469         int i;
2470
2471         read_lock(&ksocknal_data.ksnd_global_lock);
2472
2473         for (i = 0; i < ksocknal_data.ksnd_peer_hash_size; i++) {
2474                 list_for_each(tmp, &ksocknal_data.ksnd_peers[i]) {
2475                         peer_ni = list_entry(tmp, struct ksock_peer_ni, ksnp_list);
2476
2477                         if (peer_ni->ksnp_ni == ni) break;
2478
2479                         peer_ni = NULL;
2480                 }
2481         }
2482
2483         if (peer_ni != NULL) {
2484                 struct ksock_route *route;
2485                 struct ksock_conn  *conn;
2486
2487                 CWARN ("Active peer_ni on shutdown: %s, ref %d, scnt %d, "
2488                        "closing %d, accepting %d, err %d, zcookie %llu, "
2489                        "txq %d, zc_req %d\n", libcfs_id2str(peer_ni->ksnp_id),
2490                        atomic_read(&peer_ni->ksnp_refcount),
2491                        peer_ni->ksnp_sharecount, peer_ni->ksnp_closing,
2492                        peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2493                        peer_ni->ksnp_zc_next_cookie,
2494                        !list_empty(&peer_ni->ksnp_tx_queue),
2495                        !list_empty(&peer_ni->ksnp_zc_req_list));
2496
2497                 list_for_each(tmp, &peer_ni->ksnp_routes) {
2498                         route = list_entry(tmp, struct ksock_route, ksnr_list);
2499                         CWARN ("Route: ref %d, schd %d, conn %d, cnted %d, "
2500                                "del %d\n", atomic_read(&route->ksnr_refcount),
2501                                route->ksnr_scheduled, route->ksnr_connecting,
2502                                route->ksnr_connected, route->ksnr_deleted);
2503                 }
2504
2505                 list_for_each(tmp, &peer_ni->ksnp_conns) {
2506                         conn = list_entry(tmp, struct ksock_conn, ksnc_list);
2507                         CWARN ("Conn: ref %d, sref %d, t %d, c %d\n",
2508                                atomic_read(&conn->ksnc_conn_refcount),
2509                                atomic_read(&conn->ksnc_sock_refcount),
2510                                conn->ksnc_type, conn->ksnc_closing);
2511                 }
2512         }
2513
2514         read_unlock(&ksocknal_data.ksnd_global_lock);
2515         return;
2516 }
2517
2518 void
2519 ksocknal_shutdown(struct lnet_ni *ni)
2520 {
2521         struct ksock_net *net = ni->ni_data;
2522         struct lnet_process_id anyid = {
2523                 .nid = LNET_NID_ANY,
2524                 .pid = LNET_PID_ANY,
2525         };
2526         int i;
2527
2528         LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2529         LASSERT(ksocknal_data.ksnd_nnets > 0);
2530
2531         spin_lock_bh(&net->ksnn_lock);
2532         net->ksnn_shutdown = 1;                 /* prevent new peers */
2533         spin_unlock_bh(&net->ksnn_lock);
2534
2535         /* Delete all peers */
2536         ksocknal_del_peer(ni, anyid, 0);
2537
2538         /* Wait for all peer_ni state to clean up */
2539         i = 2;
2540         spin_lock_bh(&net->ksnn_lock);
2541         while (net->ksnn_npeers != 0) {
2542                 spin_unlock_bh(&net->ksnn_lock);
2543
2544                 i++;
2545                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2546                        "waiting for %d peers to disconnect\n",
2547                        net->ksnn_npeers);
2548                 set_current_state(TASK_UNINTERRUPTIBLE);
2549                 schedule_timeout(cfs_time_seconds(1));
2550
2551                 ksocknal_debug_peerhash(ni);
2552
2553                 spin_lock_bh(&net->ksnn_lock);
2554         }
2555         spin_unlock_bh(&net->ksnn_lock);
2556
2557         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2558                 LASSERT (net->ksnn_interfaces[i].ksni_npeers == 0);
2559                 LASSERT (net->ksnn_interfaces[i].ksni_nroutes == 0);
2560         }
2561
2562         list_del(&net->ksnn_list);
2563         LIBCFS_FREE(net, sizeof(*net));
2564
2565         ksocknal_data.ksnd_nnets--;
2566         if (ksocknal_data.ksnd_nnets == 0)
2567                 ksocknal_base_shutdown();
2568 }
2569
2570 static int
2571 ksocknal_enumerate_interfaces(struct ksock_net *net, char *iname)
2572 {
2573         struct net_device *dev;
2574
2575         rtnl_lock();
2576         for_each_netdev(&init_net, dev) {
2577                 /* The iname specified by an user land configuration can
2578                  * map to an ifa_label so always treat iname as an ifa_label.
2579                  * If iname is NULL then fall back to the net device name.
2580                  */
2581                 const char *name = iname ? iname : dev->name;
2582                 struct in_device *in_dev;
2583
2584                 if (strcmp(dev->name, "lo") == 0) /* skip the loopback IF */
2585                         continue;
2586
2587                 if (!(dev_get_flags(dev) & IFF_UP)) {
2588                         CWARN("Ignoring interface %s (down)\n", dev->name);
2589                         continue;
2590                 }
2591
2592                 in_dev = __in_dev_get_rtnl(dev);
2593                 if (!in_dev) {
2594                         CWARN("Interface %s has no IPv4 status.\n", dev->name);
2595                         continue;
2596                 }
2597
2598                 for_ifa(in_dev)
2599                         if (strcmp(name, ifa->ifa_label) == 0) {
2600                                 int idx = net->ksnn_ninterfaces;
2601                                 struct ksock_interface *ksi;
2602
2603                                 if (idx >= ARRAY_SIZE(net->ksnn_interfaces)) {
2604                                         rtnl_unlock();
2605                                         return -E2BIG;
2606                                 }
2607
2608                                 ksi = &net->ksnn_interfaces[idx];
2609                                 ksi->ksni_ipaddr = ntohl(ifa->ifa_local);
2610                                 ksi->ksni_netmask = ifa->ifa_mask;
2611                                 strlcpy(ksi->ksni_name,
2612                                         name, sizeof(ksi->ksni_name));
2613                                 net->ksnn_ninterfaces++;
2614                                 break;
2615                         }
2616                 endfor_ifa(in_dev);
2617         }
2618         rtnl_unlock();
2619
2620         if (net->ksnn_ninterfaces == 0)
2621                 CERROR("Can't find any usable interfaces\n");
2622
2623         return net->ksnn_ninterfaces > 0 ? 0 : -ENOENT;
2624 }
2625
2626 static int
2627 ksocknal_search_new_ipif(struct ksock_net *net)
2628 {
2629         int new_ipif = 0;
2630         int i;
2631
2632         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2633                 char *ifnam = &net->ksnn_interfaces[i].ksni_name[0];
2634                 char *colon = strchr(ifnam, ':');
2635                 int found  = 0;
2636                 struct ksock_net *tmp;
2637                 int j;
2638
2639                 if (colon != NULL) /* ignore alias device */
2640                         *colon = 0;
2641
2642                 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets,
2643                                         ksnn_list) {
2644                         for (j = 0; !found && j < tmp->ksnn_ninterfaces; j++) {
2645                                 char *ifnam2 = &tmp->ksnn_interfaces[j].\
2646                                              ksni_name[0];
2647                                 char *colon2 = strchr(ifnam2, ':');
2648
2649                                 if (colon2 != NULL)
2650                                         *colon2 = 0;
2651
2652                                 found = strcmp(ifnam, ifnam2) == 0;
2653                                 if (colon2 != NULL)
2654                                         *colon2 = ':';
2655                         }
2656                         if (found)
2657                                 break;
2658                 }
2659
2660                 new_ipif += !found;
2661                 if (colon != NULL)
2662                         *colon = ':';
2663         }
2664
2665         return new_ipif;
2666 }
2667
2668 static int
2669 ksocknal_start_schedulers(struct ksock_sched *sched)
2670 {
2671         int     nthrs;
2672         int     rc = 0;
2673         int     i;
2674
2675         if (sched->kss_nthreads == 0) {
2676                 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2677                         nthrs = sched->kss_nthreads_max;
2678                 } else {
2679                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2680                                                sched->kss_cpt);
2681                         nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2682                         nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2683                 }
2684                 nthrs = min(nthrs, sched->kss_nthreads_max);
2685         } else {
2686                 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2687                 /* increase two threads if there is new interface */
2688                 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2689         }
2690
2691         for (i = 0; i < nthrs; i++) {
2692                 long id;
2693                 char name[20];
2694
2695                 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2696                 snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
2697                          sched->kss_cpt, (int)KSOCK_THREAD_SID(id));
2698
2699                 rc = ksocknal_thread_start(ksocknal_scheduler,
2700                                            (void *)id, name);
2701                 if (rc == 0)
2702                         continue;
2703
2704                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2705                        sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2706                 break;
2707         }
2708
2709         sched->kss_nthreads += i;
2710         return rc;
2711 }
2712
2713 static int
2714 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2715 {
2716         int newif = ksocknal_search_new_ipif(net);
2717         int rc;
2718         int i;
2719
2720         if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2721                 return -EINVAL;
2722
2723         for (i = 0; i < ncpts; i++) {
2724                 struct ksock_sched *sched;
2725                 int cpt = (cpts == NULL) ? i : cpts[i];
2726
2727                 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2728                 sched = ksocknal_data.ksnd_schedulers[cpt];
2729
2730                 if (!newif && sched->kss_nthreads > 0)
2731                         continue;
2732
2733                 rc = ksocknal_start_schedulers(sched);
2734                 if (rc != 0)
2735                         return rc;
2736         }
2737         return 0;
2738 }
2739
2740 int
2741 ksocknal_startup(struct lnet_ni *ni)
2742 {
2743         struct ksock_net *net;
2744         struct lnet_ioctl_config_lnd_cmn_tunables *net_tunables;
2745         int rc;
2746         int i;
2747         struct net_device *net_dev;
2748         int node_id;
2749
2750         LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2751
2752         if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2753                 rc = ksocknal_base_startup();
2754                 if (rc != 0)
2755                         return rc;
2756         }
2757
2758         LIBCFS_ALLOC(net, sizeof(*net));
2759         if (net == NULL)
2760                 goto fail_0;
2761
2762         spin_lock_init(&net->ksnn_lock);
2763         net->ksnn_incarnation = ktime_get_real_ns();
2764         ni->ni_data = net;
2765         net_tunables = &ni->ni_net->net_tunables;
2766
2767         if (net_tunables->lct_peer_timeout == -1)
2768                 net_tunables->lct_peer_timeout =
2769                         *ksocknal_tunables.ksnd_peertimeout;
2770
2771         if (net_tunables->lct_max_tx_credits == -1)
2772                 net_tunables->lct_max_tx_credits =
2773                         *ksocknal_tunables.ksnd_credits;
2774
2775         if (net_tunables->lct_peer_tx_credits == -1)
2776                 net_tunables->lct_peer_tx_credits =
2777                         *ksocknal_tunables.ksnd_peertxcredits;
2778
2779         if (net_tunables->lct_peer_tx_credits >
2780             net_tunables->lct_max_tx_credits)
2781                 net_tunables->lct_peer_tx_credits =
2782                         net_tunables->lct_max_tx_credits;
2783
2784         if (net_tunables->lct_peer_rtr_credits == -1)
2785                 net_tunables->lct_peer_rtr_credits =
2786                         *ksocknal_tunables.ksnd_peerrtrcredits;
2787
2788         if (!ni->ni_interfaces[0]) {
2789                 rc = ksocknal_enumerate_interfaces(net, NULL);
2790                 if (rc < 0)
2791                         goto fail_1;
2792         } else {
2793                 /* Before Multi-Rail ksocklnd would manage
2794                  * multiple interfaces with its own tcp bonding.
2795                  * If we encounter an old configuration using
2796                  * this tcp bonding approach then we need to
2797                  * handle more than one ni_interfaces.
2798                  *
2799                  * In Multi-Rail configuration only ONE ni_interface
2800                  * should exist. Each IP alias should be mapped to
2801                  * each 'struct net_ni'.
2802                  */
2803                 for (i = 0; i < LNET_INTERFACES_NUM; i++) {
2804                         int j;
2805
2806                         if (!ni->ni_interfaces[i])
2807                                 break;
2808
2809                         for (j = 0; j < net->ksnn_ninterfaces;  j++) {
2810                                 struct ksock_interface *ksi;
2811
2812                                 ksi = &net->ksnn_interfaces[j];
2813
2814                                 if (strcmp(ni->ni_interfaces[i],
2815                                            ksi->ksni_name) == 0) {
2816                                         CERROR("found duplicate %s\n",
2817                                                ksi->ksni_name);
2818                                         rc = -EEXIST;
2819                                         goto fail_1;
2820                                 }
2821                         }
2822
2823                         rc = ksocknal_enumerate_interfaces(net, ni->ni_interfaces[i]);
2824                         if (rc < 0)
2825                                 goto fail_1;
2826                 }
2827         }
2828
2829         net_dev = dev_get_by_name(&init_net,
2830                                   net->ksnn_interfaces[0].ksni_name);
2831         if (net_dev != NULL) {
2832                 node_id = dev_to_node(&net_dev->dev);
2833                 ni->ni_dev_cpt = cfs_cpt_of_node(lnet_cpt_table(), node_id);
2834                 dev_put(net_dev);
2835         } else {
2836                 ni->ni_dev_cpt = CFS_CPT_ANY;
2837         }
2838
2839         /* call it before add it to ksocknal_data.ksnd_nets */
2840         rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2841         if (rc != 0)
2842                 goto fail_1;
2843
2844         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid),
2845                                 net->ksnn_interfaces[0].ksni_ipaddr);
2846         list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2847
2848         ksocknal_data.ksnd_nnets++;
2849
2850         return 0;
2851
2852  fail_1:
2853         LIBCFS_FREE(net, sizeof(*net));
2854  fail_0:
2855         if (ksocknal_data.ksnd_nnets == 0)
2856                 ksocknal_base_shutdown();
2857
2858         return -ENETDOWN;
2859 }
2860
2861
2862 static void __exit ksocklnd_exit(void)
2863 {
2864         lnet_unregister_lnd(&the_ksocklnd);
2865 }
2866
2867 static int __init ksocklnd_init(void)
2868 {
2869         int rc;
2870
2871         /* check ksnr_connected/connecting field large enough */
2872         CLASSERT(SOCKLND_CONN_NTYPES <= 4);
2873         CLASSERT(SOCKLND_CONN_ACK == SOCKLND_CONN_BULK_IN);
2874
2875         /* initialize the_ksocklnd */
2876         the_ksocklnd.lnd_type     = SOCKLND;
2877         the_ksocklnd.lnd_startup  = ksocknal_startup;
2878         the_ksocklnd.lnd_shutdown = ksocknal_shutdown;
2879         the_ksocklnd.lnd_ctl      = ksocknal_ctl;
2880         the_ksocklnd.lnd_send     = ksocknal_send;
2881         the_ksocklnd.lnd_recv     = ksocknal_recv;
2882         the_ksocklnd.lnd_notify_peer_down   = ksocknal_notify_gw_down;
2883         the_ksocklnd.lnd_query    = ksocknal_query;
2884         the_ksocklnd.lnd_accept   = ksocknal_accept;
2885
2886         rc = ksocknal_tunables_init();
2887         if (rc != 0)
2888                 return rc;
2889
2890         lnet_register_lnd(&the_ksocklnd);
2891
2892         return 0;
2893 }
2894
2895 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2896 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2897 MODULE_VERSION("2.8.0");
2898 MODULE_LICENSE("GPL");
2899
2900 module_init(ksocklnd_init);
2901 module_exit(ksocklnd_exit);