lnet/klnds/socklnd/socklnd.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  * Lustre is a trademark of Sun Microsystems, Inc.
  31  *
  32  * lnet/klnds/socklnd/socklnd.c
  33  *
  34  * Author: Zach Brown <zab@zabbo.net>
  35  * Author: Peter J. Braam <braam@clusterfs.com>
  36  * Author: Phil Schwan <phil@clusterfs.com>
  37  * Author: Eric Barton <eric@bartonsoftware.com>
  38  */
  39
  40 #include "socklnd.h"
  41 #include <linux/inetdevice.h>
  42
  43 static const struct lnet_lnd the_ksocklnd;
  44 struct ksock_nal_data ksocknal_data;
  45
  46 static struct ksock_interface *
  47 ksocknal_ip2iface(struct lnet_ni *ni, __u32 ip)
  48 {
  49         struct ksock_net *net = ni->ni_data;
  50         int i;
  51         struct ksock_interface *iface;
  52
  53         for (i = 0; i < net->ksnn_ninterfaces; i++) {
  54                 LASSERT(i < LNET_INTERFACES_NUM);
  55                 iface = &net->ksnn_interfaces[i];
  56
  57                 if (iface->ksni_ipaddr == ip)
  58                         return iface;
  59         }
  60
  61         return NULL;
  62 }
  63
  64 static struct ksock_route *
  65 ksocknal_create_route(__u32 ipaddr, int port)
  66 {
  67         struct ksock_route *route;
  68
  69         LIBCFS_ALLOC (route, sizeof (*route));
  70         if (route == NULL)
  71                 return (NULL);
  72
  73         atomic_set (&route->ksnr_refcount, 1);
  74         route->ksnr_peer = NULL;
  75         route->ksnr_retry_interval = 0;         /* OK to connect at any time */
  76         route->ksnr_ipaddr = ipaddr;
  77         route->ksnr_port = port;
  78         route->ksnr_scheduled = 0;
  79         route->ksnr_connecting = 0;
  80         route->ksnr_connected = 0;
  81         route->ksnr_deleted = 0;
  82         route->ksnr_conn_count = 0;
  83         route->ksnr_share_count = 0;
  84
  85         return (route);
  86 }
  87
  88 void
  89 ksocknal_destroy_route(struct ksock_route *route)
  90 {
  91         LASSERT (atomic_read(&route->ksnr_refcount) == 0);
  92
  93         if (route->ksnr_peer != NULL)
  94                 ksocknal_peer_decref(route->ksnr_peer);
  95
  96         LIBCFS_FREE (route, sizeof (*route));
  97 }
  98
  99 static struct ksock_peer_ni *
 100 ksocknal_create_peer(struct lnet_ni *ni, struct lnet_process_id id)
 101 {
 102         int cpt = lnet_cpt_of_nid(id.nid, ni);
 103         struct ksock_net *net = ni->ni_data;
 104         struct ksock_peer_ni *peer_ni;
 105
 106         LASSERT(id.nid != LNET_NID_ANY);
 107         LASSERT(id.pid != LNET_PID_ANY);
 108         LASSERT(!in_interrupt());
 109
 110         if (!atomic_inc_unless_negative(&net->ksnn_npeers)) {
 111                 CERROR("Can't create peer_ni: network shutdown\n");
 112                 return ERR_PTR(-ESHUTDOWN);
 113         }
 114
 115         LIBCFS_CPT_ALLOC(peer_ni, lnet_cpt_table(), cpt, sizeof(*peer_ni));
 116         if (!peer_ni) {
 117                 atomic_dec(&net->ksnn_npeers);
 118                 return ERR_PTR(-ENOMEM);
 119         }
 120
 121         peer_ni->ksnp_ni = ni;
 122         peer_ni->ksnp_id = id;
 123         atomic_set(&peer_ni->ksnp_refcount, 1); /* 1 ref for caller */
 124         peer_ni->ksnp_closing = 0;
 125         peer_ni->ksnp_accepting = 0;
 126         peer_ni->ksnp_proto = NULL;
 127         peer_ni->ksnp_last_alive = 0;
 128         peer_ni->ksnp_zc_next_cookie = SOCKNAL_KEEPALIVE_PING + 1;
 129
 130         INIT_LIST_HEAD(&peer_ni->ksnp_conns);
 131         INIT_LIST_HEAD(&peer_ni->ksnp_routes);
 132         INIT_LIST_HEAD(&peer_ni->ksnp_tx_queue);
 133         INIT_LIST_HEAD(&peer_ni->ksnp_zc_req_list);
 134         spin_lock_init(&peer_ni->ksnp_lock);
 135
 136         return peer_ni;
 137 }
 138
 139 void
 140 ksocknal_destroy_peer(struct ksock_peer_ni *peer_ni)
 141 {
 142         struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
 143
 144         CDEBUG (D_NET, "peer_ni %s %p deleted\n",
 145                 libcfs_id2str(peer_ni->ksnp_id), peer_ni);
 146
 147         LASSERT(atomic_read(&peer_ni->ksnp_refcount) == 0);
 148         LASSERT(peer_ni->ksnp_accepting == 0);
 149         LASSERT(list_empty(&peer_ni->ksnp_conns));
 150         LASSERT(list_empty(&peer_ni->ksnp_routes));
 151         LASSERT(list_empty(&peer_ni->ksnp_tx_queue));
 152         LASSERT(list_empty(&peer_ni->ksnp_zc_req_list));
 153
 154         LIBCFS_FREE(peer_ni, sizeof(*peer_ni));
 155
 156         /* NB a peer_ni's connections and routes keep a reference on their
 157          * peer_ni until they are destroyed, so we can be assured that _all_
 158          * state to do with this peer_ni has been cleaned up when its refcount
 159          * drops to zero.
 160          */
 161         atomic_dec(&net->ksnn_npeers);
 162 }
 163
 164 struct ksock_peer_ni *
 165 ksocknal_find_peer_locked(struct lnet_ni *ni, struct lnet_process_id id)
 166 {
 167         struct ksock_peer_ni *peer_ni;
 168
 169         hash_for_each_possible(ksocknal_data.ksnd_peers, peer_ni,
 170                                ksnp_list, id.nid) {
 171                 LASSERT(!peer_ni->ksnp_closing);
 172
 173                 if (peer_ni->ksnp_ni != ni)
 174                         continue;
 175
 176                 if (peer_ni->ksnp_id.nid != id.nid ||
 177                     peer_ni->ksnp_id.pid != id.pid)
 178                         continue;
 179
 180                 CDEBUG(D_NET, "got peer_ni [%p] -> %s (%d)\n",
 181                        peer_ni, libcfs_id2str(id),
 182                        atomic_read(&peer_ni->ksnp_refcount));
 183                 return peer_ni;
 184         }
 185         return NULL;
 186 }
 187
 188 struct ksock_peer_ni *
 189 ksocknal_find_peer(struct lnet_ni *ni, struct lnet_process_id id)
 190 {
 191         struct ksock_peer_ni *peer_ni;
 192
 193         read_lock(&ksocknal_data.ksnd_global_lock);
 194         peer_ni = ksocknal_find_peer_locked(ni, id);
 195         if (peer_ni != NULL)                    /* +1 ref for caller? */
 196                 ksocknal_peer_addref(peer_ni);
 197         read_unlock(&ksocknal_data.ksnd_global_lock);
 198
 199         return (peer_ni);
 200 }
 201
 202 static void
 203 ksocknal_unlink_peer_locked(struct ksock_peer_ni *peer_ni)
 204 {
 205         int i;
 206         __u32 ip;
 207         struct ksock_interface *iface;
 208
 209         for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++) {
 210                 LASSERT(i < LNET_INTERFACES_NUM);
 211                 ip = peer_ni->ksnp_passive_ips[i];
 212
 213                 iface = ksocknal_ip2iface(peer_ni->ksnp_ni, ip);
 214                 /*
 215                  * All IPs in peer_ni->ksnp_passive_ips[] come from the
 216                  * interface list, therefore the call must succeed.
 217                  */
 218                 LASSERT(iface != NULL);
 219
 220                 CDEBUG(D_NET, "peer_ni=%p iface=%p ksni_nroutes=%d\n",
 221                        peer_ni, iface, iface->ksni_nroutes);
 222                 iface->ksni_npeers--;
 223         }
 224
 225         LASSERT(list_empty(&peer_ni->ksnp_conns));
 226         LASSERT(list_empty(&peer_ni->ksnp_routes));
 227         LASSERT(!peer_ni->ksnp_closing);
 228         peer_ni->ksnp_closing = 1;
 229         hlist_del(&peer_ni->ksnp_list);
 230         /* lose peerlist's ref */
 231         ksocknal_peer_decref(peer_ni);
 232 }
 233
 234 static int
 235 ksocknal_get_peer_info(struct lnet_ni *ni, int index,
 236                        struct lnet_process_id *id, __u32 *myip, __u32 *peer_ip,
 237                        int *port, int *conn_count, int *share_count)
 238 {
 239         struct ksock_peer_ni *peer_ni;
 240         struct ksock_route *route;
 241         struct list_head *rtmp;
 242         int i;
 243         int j;
 244         int rc = -ENOENT;
 245
 246         read_lock(&ksocknal_data.ksnd_global_lock);
 247
 248         hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
 249
 250                 if (peer_ni->ksnp_ni != ni)
 251                         continue;
 252
 253                 if (peer_ni->ksnp_n_passive_ips == 0 &&
 254                     list_empty(&peer_ni->ksnp_routes)) {
 255                         if (index-- > 0)
 256                                 continue;
 257
 258                         *id = peer_ni->ksnp_id;
 259                         *myip = 0;
 260                         *peer_ip = 0;
 261                         *port = 0;
 262                         *conn_count = 0;
 263                         *share_count = 0;
 264                         rc = 0;
 265                         goto out;
 266                 }
 267
 268                 for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++) {
 269                         if (index-- > 0)
 270                                 continue;
 271
 272                         *id = peer_ni->ksnp_id;
 273                         *myip = peer_ni->ksnp_passive_ips[j];
 274                         *peer_ip = 0;
 275                         *port = 0;
 276                         *conn_count = 0;
 277                         *share_count = 0;
 278                         rc = 0;
 279                         goto out;
 280                 }
 281
 282                 list_for_each(rtmp, &peer_ni->ksnp_routes) {
 283                         if (index-- > 0)
 284                                 continue;
 285
 286                         route = list_entry(rtmp, struct ksock_route,
 287                                            ksnr_list);
 288
 289                         *id = peer_ni->ksnp_id;
 290                         *myip = route->ksnr_myipaddr;
 291                         *peer_ip = route->ksnr_ipaddr;
 292                         *port = route->ksnr_port;
 293                         *conn_count = route->ksnr_conn_count;
 294                         *share_count = route->ksnr_share_count;
 295                         rc = 0;
 296                         goto out;
 297                 }
 298         }
 299 out:
 300         read_unlock(&ksocknal_data.ksnd_global_lock);
 301         return rc;
 302 }
 303
 304 static void
 305 ksocknal_associate_route_conn_locked(struct ksock_route *route, struct ksock_conn *conn)
 306 {
 307         struct ksock_peer_ni *peer_ni = route->ksnr_peer;
 308         int type = conn->ksnc_type;
 309         struct ksock_interface *iface;
 310
 311         conn->ksnc_route = route;
 312         ksocknal_route_addref(route);
 313
 314         if (route->ksnr_myipaddr != conn->ksnc_myipaddr) {
 315                 if (route->ksnr_myipaddr == 0) {
 316                         /* route wasn't bound locally yet (the initial route) */
 317                         CDEBUG(D_NET, "Binding %s %pI4h to %pI4h\n",
 318                                libcfs_id2str(peer_ni->ksnp_id),
 319                                &route->ksnr_ipaddr,
 320                                &conn->ksnc_myipaddr);
 321                 } else {
 322                         CDEBUG(D_NET, "Rebinding %s %pI4h from %pI4h "
 323                                "to %pI4h\n", libcfs_id2str(peer_ni->ksnp_id),
 324                                &route->ksnr_ipaddr,
 325                                &route->ksnr_myipaddr,
 326                                &conn->ksnc_myipaddr);
 327
 328                         iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
 329                                                   route->ksnr_myipaddr);
 330                         if (iface != NULL)
 331                                 iface->ksni_nroutes--;
 332                 }
 333                 route->ksnr_myipaddr = conn->ksnc_myipaddr;
 334                 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
 335                                           route->ksnr_myipaddr);
 336                 if (iface != NULL)
 337                         iface->ksni_nroutes++;
 338         }
 339
 340         route->ksnr_connected |= (1<<type);
 341         route->ksnr_conn_count++;
 342
 343         /* Successful connection => further attempts can
 344          * proceed immediately */
 345         route->ksnr_retry_interval = 0;
 346 }
 347
 348 static void
 349 ksocknal_add_route_locked(struct ksock_peer_ni *peer_ni, struct ksock_route *route)
 350 {
 351         struct list_head *tmp;
 352         struct ksock_conn *conn;
 353         struct ksock_route *route2;
 354
 355         LASSERT(!peer_ni->ksnp_closing);
 356         LASSERT(route->ksnr_peer == NULL);
 357         LASSERT(!route->ksnr_scheduled);
 358         LASSERT(!route->ksnr_connecting);
 359         LASSERT(route->ksnr_connected == 0);
 360
 361         /* LASSERT(unique) */
 362         list_for_each(tmp, &peer_ni->ksnp_routes) {
 363                 route2 = list_entry(tmp, struct ksock_route, ksnr_list);
 364
 365                 if (route2->ksnr_ipaddr == route->ksnr_ipaddr) {
 366                         CERROR("Duplicate route %s %pI4h\n",
 367                                libcfs_id2str(peer_ni->ksnp_id),
 368                                &route->ksnr_ipaddr);
 369                         LBUG();
 370                 }
 371         }
 372
 373         route->ksnr_peer = peer_ni;
 374         ksocknal_peer_addref(peer_ni);
 375         /* peer_ni's routelist takes over my ref on 'route' */
 376         list_add_tail(&route->ksnr_list, &peer_ni->ksnp_routes);
 377
 378         list_for_each(tmp, &peer_ni->ksnp_conns) {
 379                 conn = list_entry(tmp, struct ksock_conn, ksnc_list);
 380
 381                 if (conn->ksnc_ipaddr != route->ksnr_ipaddr)
 382                         continue;
 383
 384                 ksocknal_associate_route_conn_locked(route, conn);
 385                 /* keep going (typed routes) */
 386         }
 387 }
 388
 389 static void
 390 ksocknal_del_route_locked(struct ksock_route *route)
 391 {
 392         struct ksock_peer_ni *peer_ni = route->ksnr_peer;
 393         struct ksock_interface *iface;
 394         struct ksock_conn *conn;
 395         struct list_head *ctmp;
 396         struct list_head *cnxt;
 397
 398         LASSERT(!route->ksnr_deleted);
 399
 400         /* Close associated conns */
 401         list_for_each_safe(ctmp, cnxt, &peer_ni->ksnp_conns) {
 402                 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
 403
 404                 if (conn->ksnc_route != route)
 405                         continue;
 406
 407                 ksocknal_close_conn_locked(conn, 0);
 408         }
 409
 410         if (route->ksnr_myipaddr != 0) {
 411                 iface = ksocknal_ip2iface(route->ksnr_peer->ksnp_ni,
 412                                           route->ksnr_myipaddr);
 413                 if (iface != NULL)
 414                         iface->ksni_nroutes--;
 415         }
 416
 417         route->ksnr_deleted = 1;
 418         list_del(&route->ksnr_list);
 419         ksocknal_route_decref(route);           /* drop peer_ni's ref */
 420
 421         if (list_empty(&peer_ni->ksnp_routes) &&
 422             list_empty(&peer_ni->ksnp_conns)) {
 423                 /* I've just removed the last route to a peer_ni with no active
 424                  * connections */
 425                 ksocknal_unlink_peer_locked(peer_ni);
 426         }
 427 }
 428
 429 int
 430 ksocknal_add_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ipaddr,
 431                   int port)
 432 {
 433         struct list_head *tmp;
 434         struct ksock_peer_ni *peer_ni;
 435         struct ksock_peer_ni *peer2;
 436         struct ksock_route *route;
 437         struct ksock_route *route2;
 438
 439         if (id.nid == LNET_NID_ANY ||
 440             id.pid == LNET_PID_ANY)
 441                 return (-EINVAL);
 442
 443         /* Have a brand new peer_ni ready... */
 444         peer_ni = ksocknal_create_peer(ni, id);
 445         if (IS_ERR(peer_ni))
 446                 return PTR_ERR(peer_ni);
 447
 448         route = ksocknal_create_route (ipaddr, port);
 449         if (route == NULL) {
 450                 ksocknal_peer_decref(peer_ni);
 451                 return (-ENOMEM);
 452         }
 453
 454         write_lock_bh(&ksocknal_data.ksnd_global_lock);
 455
 456         /* always called with a ref on ni, so shutdown can't have started */
 457         LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers)
 458                 >= 0);
 459
 460         peer2 = ksocknal_find_peer_locked(ni, id);
 461         if (peer2 != NULL) {
 462                 ksocknal_peer_decref(peer_ni);
 463                 peer_ni = peer2;
 464         } else {
 465                 /* peer_ni table takes my ref on peer_ni */
 466                 hash_add(ksocknal_data.ksnd_peers, &peer_ni->ksnp_list, id.nid);
 467         }
 468
 469         route2 = NULL;
 470         list_for_each(tmp, &peer_ni->ksnp_routes) {
 471                 route2 = list_entry(tmp, struct ksock_route, ksnr_list);
 472
 473                 if (route2->ksnr_ipaddr == ipaddr)
 474                         break;
 475
 476                 route2 = NULL;
 477         }
 478         if (route2 == NULL) {
 479                 ksocknal_add_route_locked(peer_ni, route);
 480                 route->ksnr_share_count++;
 481         } else {
 482                 ksocknal_route_decref(route);
 483                 route2->ksnr_share_count++;
 484         }
 485
 486         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 487
 488         return 0;
 489 }
 490
 491 static void
 492 ksocknal_del_peer_locked(struct ksock_peer_ni *peer_ni, __u32 ip)
 493 {
 494         struct ksock_conn *conn;
 495         struct ksock_route *route;
 496         struct list_head *tmp;
 497         struct list_head *nxt;
 498         int nshared;
 499
 500         LASSERT(!peer_ni->ksnp_closing);
 501
 502         /* Extra ref prevents peer_ni disappearing until I'm done with it */
 503         ksocknal_peer_addref(peer_ni);
 504
 505         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
 506                 route = list_entry(tmp, struct ksock_route, ksnr_list);
 507
 508                 /* no match */
 509                 if (!(ip == 0 || route->ksnr_ipaddr == ip))
 510                         continue;
 511
 512                 route->ksnr_share_count = 0;
 513                 /* This deletes associated conns too */
 514                 ksocknal_del_route_locked(route);
 515         }
 516
 517         nshared = 0;
 518         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
 519                 route = list_entry(tmp, struct ksock_route, ksnr_list);
 520                 nshared += route->ksnr_share_count;
 521         }
 522
 523         if (nshared == 0) {
 524                 /* remove everything else if there are no explicit entries
 525                  * left */
 526
 527                 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
 528                         route = list_entry(tmp, struct ksock_route, ksnr_list);
 529
 530                         /* we should only be removing auto-entries */
 531                         LASSERT(route->ksnr_share_count == 0);
 532                         ksocknal_del_route_locked(route);
 533                 }
 534
 535                 list_for_each_safe(tmp, nxt, &peer_ni->ksnp_conns) {
 536                         conn = list_entry(tmp, struct ksock_conn, ksnc_list);
 537
 538                         ksocknal_close_conn_locked(conn, 0);
 539                 }
 540         }
 541
 542         ksocknal_peer_decref(peer_ni);
 543         /* NB peer_ni unlinks itself when last conn/route is removed */
 544 }
 545
 546 static int
 547 ksocknal_del_peer(struct lnet_ni *ni, struct lnet_process_id id, __u32 ip)
 548 {
 549         LIST_HEAD(zombies);
 550         struct hlist_node *pnxt;
 551         struct ksock_peer_ni *peer_ni;
 552         int lo;
 553         int hi;
 554         int i;
 555         int rc = -ENOENT;
 556
 557         write_lock_bh(&ksocknal_data.ksnd_global_lock);
 558
 559         if (id.nid != LNET_NID_ANY) {
 560                 lo = hash_min(id.nid, HASH_BITS(ksocknal_data.ksnd_peers));
 561                 hi = lo;
 562         } else {
 563                 lo = 0;
 564                 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
 565         }
 566
 567         for (i = lo; i <= hi; i++) {
 568                 hlist_for_each_entry_safe(peer_ni, pnxt,
 569                                           &ksocknal_data.ksnd_peers[i],
 570                                           ksnp_list) {
 571                         if (peer_ni->ksnp_ni != ni)
 572                                 continue;
 573
 574                         if (!((id.nid == LNET_NID_ANY ||
 575                                peer_ni->ksnp_id.nid == id.nid) &&
 576                               (id.pid == LNET_PID_ANY ||
 577                                peer_ni->ksnp_id.pid == id.pid)))
 578                                 continue;
 579
 580                         ksocknal_peer_addref(peer_ni);  /* a ref for me... */
 581
 582                         ksocknal_del_peer_locked(peer_ni, ip);
 583
 584                         if (peer_ni->ksnp_closing &&
 585                             !list_empty(&peer_ni->ksnp_tx_queue)) {
 586                                 LASSERT(list_empty(&peer_ni->ksnp_conns));
 587                                 LASSERT(list_empty(&peer_ni->ksnp_routes));
 588
 589                                 list_splice_init(&peer_ni->ksnp_tx_queue,
 590                                                  &zombies);
 591                         }
 592
 593                         ksocknal_peer_decref(peer_ni);  /* ...till here */
 594
 595                         rc = 0;                         /* matched! */
 596                 }
 597         }
 598
 599         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
 600
 601         ksocknal_txlist_done(ni, &zombies, -ENETDOWN);
 602
 603         return rc;
 604 }
 605
 606 static struct ksock_conn *
 607 ksocknal_get_conn_by_idx(struct lnet_ni *ni, int index)
 608 {
 609         struct ksock_peer_ni *peer_ni;
 610         struct ksock_conn *conn;
 611         struct list_head *ctmp;
 612         int i;
 613
 614         read_lock(&ksocknal_data.ksnd_global_lock);
 615
 616         hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
 617                 LASSERT(!peer_ni->ksnp_closing);
 618
 619                 if (peer_ni->ksnp_ni != ni)
 620                         continue;
 621
 622                 list_for_each(ctmp, &peer_ni->ksnp_conns) {
 623                         if (index-- > 0)
 624                                 continue;
 625
 626                         conn = list_entry(ctmp, struct ksock_conn,
 627                                           ksnc_list);
 628                         ksocknal_conn_addref(conn);
 629                         read_unlock(&ksocknal_data.ksnd_global_lock);
 630                         return conn;
 631                 }
 632         }
 633
 634         read_unlock(&ksocknal_data.ksnd_global_lock);
 635         return NULL;
 636 }
 637
 638 static struct ksock_sched *
 639 ksocknal_choose_scheduler_locked(unsigned int cpt)
 640 {
 641         struct ksock_sched *sched = ksocknal_data.ksnd_schedulers[cpt];
 642         int i;
 643
 644         if (sched->kss_nthreads == 0) {
 645                 cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
 646                         if (sched->kss_nthreads > 0) {
 647                                 CDEBUG(D_NET, "scheduler[%d] has no threads. selected scheduler[%d]\n",
 648                                        cpt, sched->kss_cpt);
 649                                 return sched;
 650                         }
 651                 }
 652                 return NULL;
 653         }
 654
 655         return sched;
 656 }
 657
 658 static int
 659 ksocknal_local_ipvec(struct lnet_ni *ni, __u32 *ipaddrs)
 660 {
 661         struct ksock_net *net = ni->ni_data;
 662         int i;
 663         int nip;
 664
 665         read_lock(&ksocknal_data.ksnd_global_lock);
 666
 667         nip = net->ksnn_ninterfaces;
 668         LASSERT(nip <= LNET_INTERFACES_NUM);
 669
 670         /*
 671          * Only offer interfaces for additional connections if I have
 672          * more than one.
 673          */
 674         if (nip < 2) {
 675                 read_unlock(&ksocknal_data.ksnd_global_lock);
 676                 return 0;
 677         }
 678
 679         for (i = 0; i < nip; i++) {
 680                 ipaddrs[i] = net->ksnn_interfaces[i].ksni_ipaddr;
 681                 LASSERT(ipaddrs[i] != 0);
 682         }
 683
 684         read_unlock(&ksocknal_data.ksnd_global_lock);
 685         return nip;
 686 }
 687
 688 static int
 689 ksocknal_match_peerip(struct ksock_interface *iface, __u32 *ips, int nips)
 690 {
 691         int best_netmatch = 0;
 692         int best_xor = 0;
 693         int best = -1;
 694         int this_xor;
 695         int this_netmatch;
 696         int i;
 697
 698         for (i = 0; i < nips; i++) {
 699                 if (ips[i] == 0)
 700                         continue;
 701
 702                 this_xor = (ips[i] ^ iface->ksni_ipaddr);
 703                 this_netmatch = ((this_xor & iface->ksni_netmask) == 0) ? 1 : 0;
 704
 705                 if (!(best < 0 ||
 706                       best_netmatch < this_netmatch ||
 707                       (best_netmatch == this_netmatch &&
 708                        best_xor > this_xor)))
 709                         continue;
 710
 711                 best = i;
 712                 best_netmatch = this_netmatch;
 713                 best_xor = this_xor;
 714         }
 715
 716         LASSERT (best >= 0);
 717         return (best);
 718 }
 719
 720 static int
 721 ksocknal_select_ips(struct ksock_peer_ni *peer_ni, __u32 *peerips, int n_peerips)
 722 {
 723         rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
 724         struct ksock_net *net = peer_ni->ksnp_ni->ni_data;
 725         struct ksock_interface *iface;
 726         struct ksock_interface *best_iface;
 727         int n_ips;
 728         int i;
 729         int j;
 730         int k;
 731         u32 ip;
 732         u32 xor;
 733         int this_netmatch;
 734         int best_netmatch;
 735         int best_npeers;
 736
 737         /* CAVEAT EMPTOR: We do all our interface matching with an
 738          * exclusive hold of global lock at IRQ priority.  We're only
 739          * expecting to be dealing with small numbers of interfaces, so the
 740          * O(n**3)-ness shouldn't matter */
 741
 742         /* Also note that I'm not going to return more than n_peerips
 743          * interfaces, even if I have more myself */
 744
 745         write_lock_bh(global_lock);
 746
 747         LASSERT(n_peerips <= LNET_INTERFACES_NUM);
 748         LASSERT(net->ksnn_ninterfaces <= LNET_INTERFACES_NUM);
 749
 750         /* Only match interfaces for additional connections
 751          * if I have > 1 interface
 752          */
 753         n_ips = (net->ksnn_ninterfaces < 2) ? 0 :
 754                 min(n_peerips, net->ksnn_ninterfaces);
 755
 756         for (i = 0; peer_ni->ksnp_n_passive_ips < n_ips; i++) {
 757                 /*              ^ yes really... */
 758
 759                 /* If we have any new interfaces, first tick off all the
 760                  * peer_ni IPs that match old interfaces, then choose new
 761                  * interfaces to match the remaining peer_ni IPS.
 762                  * We don't forget interfaces we've stopped using; we might
 763                  * start using them again... */
 764
 765                 if (i < peer_ni->ksnp_n_passive_ips) {
 766                         /* Old interface. */
 767                         ip = peer_ni->ksnp_passive_ips[i];
 768                         best_iface = ksocknal_ip2iface(peer_ni->ksnp_ni, ip);
 769
 770                         /* peer_ni passive ips are kept up to date */
 771                         LASSERT(best_iface != NULL);
 772                 } else {
 773                         /* choose a new interface */
 774                         LASSERT (i == peer_ni->ksnp_n_passive_ips);
 775
 776                         best_iface = NULL;
 777                         best_netmatch = 0;
 778                         best_npeers = 0;
 779
 780                         for (j = 0; j < net->ksnn_ninterfaces; j++) {
 781                                 iface = &net->ksnn_interfaces[j];
 782                                 ip = iface->ksni_ipaddr;
 783
 784                                 for (k = 0; k < peer_ni->ksnp_n_passive_ips; k++)
 785                                         if (peer_ni->ksnp_passive_ips[k] == ip)
 786                                                 break;
 787
 788                                 if (k < peer_ni->ksnp_n_passive_ips) /* using it already */
 789                                         continue;
 790
 791                                 k = ksocknal_match_peerip(iface, peerips, n_peerips);
 792                                 xor = (ip ^ peerips[k]);
 793                                 this_netmatch = ((xor & iface->ksni_netmask) == 0) ? 1 : 0;
 794
 795                                 if (!(best_iface == NULL ||
 796                                       best_netmatch < this_netmatch ||
 797                                       (best_netmatch == this_netmatch &&
 798                                        best_npeers > iface->ksni_npeers)))
 799                                         continue;
 800
 801                                 best_iface = iface;
 802                                 best_netmatch = this_netmatch;
 803                                 best_npeers = iface->ksni_npeers;
 804                         }
 805
 806                         LASSERT(best_iface != NULL);
 807
 808                         best_iface->ksni_npeers++;
 809                         ip = best_iface->ksni_ipaddr;
 810                         peer_ni->ksnp_passive_ips[i] = ip;
 811                         peer_ni->ksnp_n_passive_ips = i+1;
 812                 }
 813
 814                 /* mark the best matching peer_ni IP used */
 815                 j = ksocknal_match_peerip(best_iface, peerips, n_peerips);
 816                 peerips[j] = 0;
 817         }
 818
 819         /* Overwrite input peer_ni IP addresses */
 820         memcpy(peerips, peer_ni->ksnp_passive_ips, n_ips * sizeof(*peerips));
 821
 822         write_unlock_bh(global_lock);
 823
 824         return (n_ips);
 825 }
 826
 827 static void
 828 ksocknal_create_routes(struct ksock_peer_ni *peer_ni, int port,
 829                        __u32 *peer_ipaddrs, int npeer_ipaddrs)
 830 {
 831         struct ksock_route              *newroute = NULL;
 832         rwlock_t                *global_lock = &ksocknal_data.ksnd_global_lock;
 833         struct lnet_ni *ni = peer_ni->ksnp_ni;
 834         struct ksock_net                *net = ni->ni_data;
 835         struct list_head        *rtmp;
 836         struct ksock_route              *route;
 837         struct ksock_interface  *iface;
 838         struct ksock_interface  *best_iface;
 839         int                     best_netmatch;
 840         int                     this_netmatch;
 841         int                     best_nroutes;
 842         int                     i;
 843         int                     j;
 844
 845         /* CAVEAT EMPTOR: We do all our interface matching with an
 846          * exclusive hold of global lock at IRQ priority.  We're only
 847          * expecting to be dealing with small numbers of interfaces, so the
 848          * O(n**3)-ness here shouldn't matter */
 849
 850         write_lock_bh(global_lock);
 851
 852         if (net->ksnn_ninterfaces < 2) {
 853                 /* Only create additional connections
 854                  * if I have > 1 interface */
 855                 write_unlock_bh(global_lock);
 856                 return;
 857         }
 858
 859         LASSERT(npeer_ipaddrs <= LNET_INTERFACES_NUM);
 860
 861         for (i = 0; i < npeer_ipaddrs; i++) {
 862                 if (newroute != NULL) {
 863                         newroute->ksnr_ipaddr = peer_ipaddrs[i];
 864                 } else {
 865                         write_unlock_bh(global_lock);
 866
 867                         newroute = ksocknal_create_route(peer_ipaddrs[i], port);
 868                         if (newroute == NULL)
 869                                 return;
 870
 871                         write_lock_bh(global_lock);
 872                 }
 873
 874                 if (peer_ni->ksnp_closing) {
 875                         /* peer_ni got closed under me */
 876                         break;
 877                 }
 878
 879                 /* Already got a route? */
 880                 route = NULL;
 881                 list_for_each(rtmp, &peer_ni->ksnp_routes) {
 882                         route = list_entry(rtmp, struct ksock_route, ksnr_list);
 883
 884                         if (route->ksnr_ipaddr == newroute->ksnr_ipaddr)
 885                                 break;
 886
 887                         route = NULL;
 888                 }
 889                 if (route != NULL)
 890                         continue;
 891
 892                 best_iface = NULL;
 893                 best_nroutes = 0;
 894                 best_netmatch = 0;
 895
 896                 LASSERT(net->ksnn_ninterfaces <= LNET_INTERFACES_NUM);
 897
 898                 /* Select interface to connect from */
 899                 for (j = 0; j < net->ksnn_ninterfaces; j++) {
 900                         iface = &net->ksnn_interfaces[j];
 901
 902                         /* Using this interface already? */
 903                         list_for_each(rtmp, &peer_ni->ksnp_routes) {
 904                                 route = list_entry(rtmp, struct ksock_route,
 905                                                    ksnr_list);
 906
 907                                 if (route->ksnr_myipaddr == iface->ksni_ipaddr)
 908                                         break;
 909
 910                                 route = NULL;
 911                         }
 912                         if (route != NULL)
 913                                 continue;
 914
 915                         this_netmatch = (((iface->ksni_ipaddr ^
 916                                            newroute->ksnr_ipaddr) &
 917                                            iface->ksni_netmask) == 0) ? 1 : 0;
 918
 919                         if (!(best_iface == NULL ||
 920                               best_netmatch < this_netmatch ||
 921                               (best_netmatch == this_netmatch &&
 922                                best_nroutes > iface->ksni_nroutes)))
 923                                 continue;
 924
 925                         best_iface = iface;
 926                         best_netmatch = this_netmatch;
 927                         best_nroutes = iface->ksni_nroutes;
 928                 }
 929
 930                 if (best_iface == NULL)
 931                         continue;
 932
 933                 newroute->ksnr_myipaddr = best_iface->ksni_ipaddr;
 934                 best_iface->ksni_nroutes++;
 935
 936                 ksocknal_add_route_locked(peer_ni, newroute);
 937                 newroute = NULL;
 938         }
 939
 940         write_unlock_bh(global_lock);
 941         if (newroute != NULL)
 942                 ksocknal_route_decref(newroute);
 943 }
 944
 945 int
 946 ksocknal_accept(struct lnet_ni *ni, struct socket *sock)
 947 {
 948         struct ksock_connreq *cr;
 949         int rc;
 950         u32 peer_ip;
 951         int peer_port;
 952
 953         rc = lnet_sock_getaddr(sock, true, &peer_ip, &peer_port);
 954         LASSERT(rc == 0);               /* we succeeded before */
 955
 956         LIBCFS_ALLOC(cr, sizeof(*cr));
 957         if (cr == NULL) {
 958                 LCONSOLE_ERROR_MSG(0x12f, "Dropping connection request from "
 959                                    "%pI4h: memory exhausted\n", &peer_ip);
 960                 return -ENOMEM;
 961         }
 962
 963         lnet_ni_addref(ni);
 964         cr->ksncr_ni   = ni;
 965         cr->ksncr_sock = sock;
 966
 967         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
 968
 969         list_add_tail(&cr->ksncr_list, &ksocknal_data.ksnd_connd_connreqs);
 970         wake_up(&ksocknal_data.ksnd_connd_waitq);
 971
 972         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
 973         return 0;
 974 }
 975
 976 static int
 977 ksocknal_connecting(struct ksock_peer_ni *peer_ni, __u32 ipaddr)
 978 {
 979         struct ksock_route *route;
 980
 981         list_for_each_entry(route, &peer_ni->ksnp_routes, ksnr_list) {
 982                 if (route->ksnr_ipaddr == ipaddr)
 983                         return route->ksnr_connecting;
 984         }
 985         return 0;
 986 }
 987
 988 int
 989 ksocknal_create_conn(struct lnet_ni *ni, struct ksock_route *route,
 990                      struct socket *sock, int type)
 991 {
 992         rwlock_t *global_lock = &ksocknal_data.ksnd_global_lock;
 993         LIST_HEAD(zombies);
 994         struct lnet_process_id peerid;
 995         struct list_head *tmp;
 996         u64 incarnation;
 997         struct ksock_conn *conn;
 998         struct ksock_conn *conn2;
 999         struct ksock_peer_ni *peer_ni = NULL;
1000         struct ksock_peer_ni *peer2;
1001         struct ksock_sched *sched;
1002         struct ksock_hello_msg *hello;
1003         int cpt;
1004         struct ksock_tx *tx;
1005         struct ksock_tx *txtmp;
1006         int rc;
1007         int rc2;
1008         int active;
1009         char *warn = NULL;
1010
1011         active = (route != NULL);
1012
1013         LASSERT (active == (type != SOCKLND_CONN_NONE));
1014
1015         LIBCFS_ALLOC(conn, sizeof(*conn));
1016         if (conn == NULL) {
1017                 rc = -ENOMEM;
1018                 goto failed_0;
1019         }
1020
1021         conn->ksnc_peer = NULL;
1022         conn->ksnc_route = NULL;
1023         conn->ksnc_sock = sock;
1024         /* 2 ref, 1 for conn, another extra ref prevents socket
1025          * being closed before establishment of connection */
1026         atomic_set (&conn->ksnc_sock_refcount, 2);
1027         conn->ksnc_type = type;
1028         ksocknal_lib_save_callback(sock, conn);
1029         atomic_set (&conn->ksnc_conn_refcount, 1); /* 1 ref for me */
1030
1031         conn->ksnc_rx_ready = 0;
1032         conn->ksnc_rx_scheduled = 0;
1033
1034         INIT_LIST_HEAD(&conn->ksnc_tx_queue);
1035         conn->ksnc_tx_ready = 0;
1036         conn->ksnc_tx_scheduled = 0;
1037         conn->ksnc_tx_carrier = NULL;
1038         atomic_set (&conn->ksnc_tx_nob, 0);
1039
1040         LIBCFS_ALLOC(hello, offsetof(struct ksock_hello_msg,
1041                                      kshm_ips[LNET_INTERFACES_NUM]));
1042         if (hello == NULL) {
1043                 rc = -ENOMEM;
1044                 goto failed_1;
1045         }
1046
1047         /* stash conn's local and remote addrs */
1048         rc = ksocknal_lib_get_conn_addrs (conn);
1049         if (rc != 0)
1050                 goto failed_1;
1051
1052         /* Find out/confirm peer_ni's NID and connection type and get the
1053          * vector of interfaces she's willing to let me connect to.
1054          * Passive connections use the listener timeout since the peer_ni sends
1055          * eagerly */
1056
1057         if (active) {
1058                 peer_ni = route->ksnr_peer;
1059                 LASSERT(ni == peer_ni->ksnp_ni);
1060
1061                 /* Active connection sends HELLO eagerly */
1062                 hello->kshm_nips = ksocknal_local_ipvec(ni, hello->kshm_ips);
1063                 peerid = peer_ni->ksnp_id;
1064
1065                 write_lock_bh(global_lock);
1066                 conn->ksnc_proto = peer_ni->ksnp_proto;
1067                 write_unlock_bh(global_lock);
1068
1069                 if (conn->ksnc_proto == NULL) {
1070                          conn->ksnc_proto = &ksocknal_protocol_v3x;
1071 #if SOCKNAL_VERSION_DEBUG
1072                          if (*ksocknal_tunables.ksnd_protocol == 2)
1073                                  conn->ksnc_proto = &ksocknal_protocol_v2x;
1074                          else if (*ksocknal_tunables.ksnd_protocol == 1)
1075                                  conn->ksnc_proto = &ksocknal_protocol_v1x;
1076 #endif
1077                 }
1078
1079                 rc = ksocknal_send_hello (ni, conn, peerid.nid, hello);
1080                 if (rc != 0)
1081                         goto failed_1;
1082         } else {
1083                 peerid.nid = LNET_NID_ANY;
1084                 peerid.pid = LNET_PID_ANY;
1085
1086                 /* Passive, get protocol from peer_ni */
1087                 conn->ksnc_proto = NULL;
1088         }
1089
1090         rc = ksocknal_recv_hello (ni, conn, hello, &peerid, &incarnation);
1091         if (rc < 0)
1092                 goto failed_1;
1093
1094         LASSERT (rc == 0 || active);
1095         LASSERT (conn->ksnc_proto != NULL);
1096         LASSERT (peerid.nid != LNET_NID_ANY);
1097
1098         cpt = lnet_cpt_of_nid(peerid.nid, ni);
1099
1100         if (active) {
1101                 ksocknal_peer_addref(peer_ni);
1102                 write_lock_bh(global_lock);
1103         } else {
1104                 peer_ni = ksocknal_create_peer(ni, peerid);
1105                 if (IS_ERR(peer_ni)) {
1106                         rc = PTR_ERR(peer_ni);
1107                         goto failed_1;
1108                 }
1109
1110                 write_lock_bh(global_lock);
1111
1112                 /* called with a ref on ni, so shutdown can't have started */
1113                 LASSERT(atomic_read(&((struct ksock_net *)ni->ni_data)->ksnn_npeers) >= 0);
1114
1115                 peer2 = ksocknal_find_peer_locked(ni, peerid);
1116                 if (peer2 == NULL) {
1117                         /* NB this puts an "empty" peer_ni in the peer_ni
1118                          * table (which takes my ref) */
1119                         hash_add(ksocknal_data.ksnd_peers,
1120                                  &peer_ni->ksnp_list, peerid.nid);
1121                 } else {
1122                         ksocknal_peer_decref(peer_ni);
1123                         peer_ni = peer2;
1124                 }
1125
1126                 /* +1 ref for me */
1127                 ksocknal_peer_addref(peer_ni);
1128                 peer_ni->ksnp_accepting++;
1129
1130                 /* Am I already connecting to this guy?  Resolve in
1131                  * favour of higher NID... */
1132                 if (peerid.nid < ni->ni_nid &&
1133                     ksocknal_connecting(peer_ni, conn->ksnc_ipaddr)) {
1134                         rc = EALREADY;
1135                         warn = "connection race resolution";
1136                         goto failed_2;
1137                 }
1138         }
1139
1140         if (peer_ni->ksnp_closing ||
1141             (active && route->ksnr_deleted)) {
1142                 /* peer_ni/route got closed under me */
1143                 rc = -ESTALE;
1144                 warn = "peer_ni/route removed";
1145                 goto failed_2;
1146         }
1147
1148         if (peer_ni->ksnp_proto == NULL) {
1149                 /* Never connected before.
1150                  * NB recv_hello may have returned EPROTO to signal my peer_ni
1151                  * wants a different protocol than the one I asked for.
1152                  */
1153                 LASSERT(list_empty(&peer_ni->ksnp_conns));
1154
1155                 peer_ni->ksnp_proto = conn->ksnc_proto;
1156                 peer_ni->ksnp_incarnation = incarnation;
1157         }
1158
1159         if (peer_ni->ksnp_proto != conn->ksnc_proto ||
1160             peer_ni->ksnp_incarnation != incarnation) {
1161                 /* peer_ni rebooted or I've got the wrong protocol version */
1162                 ksocknal_close_peer_conns_locked(peer_ni, 0, 0);
1163
1164                 peer_ni->ksnp_proto = NULL;
1165                 rc = ESTALE;
1166                 warn = peer_ni->ksnp_incarnation != incarnation ?
1167                        "peer_ni rebooted" :
1168                        "wrong proto version";
1169                 goto failed_2;
1170         }
1171
1172         switch (rc) {
1173         default:
1174                 LBUG();
1175         case 0:
1176                 break;
1177         case EALREADY:
1178                 warn = "lost conn race";
1179                 goto failed_2;
1180         case EPROTO:
1181                 warn = "retry with different protocol version";
1182                 goto failed_2;
1183         }
1184
1185         /* Refuse to duplicate an existing connection, unless this is a
1186          * loopback connection */
1187         if (conn->ksnc_ipaddr != conn->ksnc_myipaddr) {
1188                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1189                         conn2 = list_entry(tmp, struct ksock_conn, ksnc_list);
1190
1191                         if (conn2->ksnc_ipaddr != conn->ksnc_ipaddr ||
1192                             conn2->ksnc_myipaddr != conn->ksnc_myipaddr ||
1193                             conn2->ksnc_type != conn->ksnc_type)
1194                                 continue;
1195
1196                         /* Reply on a passive connection attempt so the peer_ni
1197                          * realises we're connected. */
1198                         LASSERT (rc == 0);
1199                         if (!active)
1200                                 rc = EALREADY;
1201
1202                         warn = "duplicate";
1203                         goto failed_2;
1204                 }
1205         }
1206
1207         /* If the connection created by this route didn't bind to the IP
1208          * address the route connected to, the connection/route matching
1209          * code below probably isn't going to work. */
1210         if (active &&
1211             route->ksnr_ipaddr != conn->ksnc_ipaddr) {
1212                 CERROR("Route %s %pI4h connected to %pI4h\n",
1213                        libcfs_id2str(peer_ni->ksnp_id),
1214                        &route->ksnr_ipaddr,
1215                        &conn->ksnc_ipaddr);
1216         }
1217
1218         /* Search for a route corresponding to the new connection and
1219          * create an association.  This allows incoming connections created
1220          * by routes in my peer_ni to match my own route entries so I don't
1221          * continually create duplicate routes. */
1222         list_for_each(tmp, &peer_ni->ksnp_routes) {
1223                 route = list_entry(tmp, struct ksock_route, ksnr_list);
1224
1225                 if (route->ksnr_ipaddr != conn->ksnc_ipaddr)
1226                         continue;
1227
1228                 ksocknal_associate_route_conn_locked(route, conn);
1229                 break;
1230         }
1231
1232         conn->ksnc_peer = peer_ni;                 /* conn takes my ref on peer_ni */
1233         peer_ni->ksnp_last_alive = ktime_get_seconds();
1234         peer_ni->ksnp_send_keepalive = 0;
1235         peer_ni->ksnp_error = 0;
1236
1237         sched = ksocknal_choose_scheduler_locked(cpt);
1238         if (!sched) {
1239                 CERROR("no schedulers available. node is unhealthy\n");
1240                 goto failed_2;
1241         }
1242         /*
1243          * The cpt might have changed if we ended up selecting a non cpt
1244          * native scheduler. So use the scheduler's cpt instead.
1245          */
1246         cpt = sched->kss_cpt;
1247         sched->kss_nconns++;
1248         conn->ksnc_scheduler = sched;
1249
1250         conn->ksnc_tx_last_post = ktime_get_seconds();
1251         /* Set the deadline for the outgoing HELLO to drain */
1252         conn->ksnc_tx_bufnob = sock->sk->sk_wmem_queued;
1253         conn->ksnc_tx_deadline = ktime_get_seconds() +
1254                                  lnet_get_lnd_timeout();
1255         smp_mb();   /* order with adding to peer_ni's conn list */
1256
1257         list_add(&conn->ksnc_list, &peer_ni->ksnp_conns);
1258         ksocknal_conn_addref(conn);
1259
1260         ksocknal_new_packet(conn, 0);
1261
1262         conn->ksnc_zc_capable = ksocknal_lib_zc_capable(conn);
1263
1264         /* Take packets blocking for this connection. */
1265         list_for_each_entry_safe(tx, txtmp, &peer_ni->ksnp_tx_queue, tx_list) {
1266                 if (conn->ksnc_proto->pro_match_tx(conn, tx, tx->tx_nonblk) ==
1267                     SOCKNAL_MATCH_NO)
1268                         continue;
1269
1270                 list_del(&tx->tx_list);
1271                 ksocknal_queue_tx_locked(tx, conn);
1272         }
1273
1274         write_unlock_bh(global_lock);
1275
1276         /* We've now got a new connection.  Any errors from here on are just
1277          * like "normal" comms errors and we close the connection normally.
1278          * NB (a) we still have to send the reply HELLO for passive
1279          *        connections,
1280          *    (b) normal I/O on the conn is blocked until I setup and call the
1281          *        socket callbacks.
1282          */
1283
1284         CDEBUG(D_NET, "New conn %s p %d.x %pI4h -> %pI4h/%d"
1285                " incarnation:%lld sched[%d]\n",
1286                libcfs_id2str(peerid), conn->ksnc_proto->pro_version,
1287                &conn->ksnc_myipaddr, &conn->ksnc_ipaddr,
1288                conn->ksnc_port, incarnation, cpt);
1289
1290         if (active) {
1291                 /* additional routes after interface exchange? */
1292                 ksocknal_create_routes(peer_ni, conn->ksnc_port,
1293                                        hello->kshm_ips, hello->kshm_nips);
1294         } else {
1295                 hello->kshm_nips = ksocknal_select_ips(peer_ni, hello->kshm_ips,
1296                                                        hello->kshm_nips);
1297                 rc = ksocknal_send_hello(ni, conn, peerid.nid, hello);
1298         }
1299
1300         LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1301                                     kshm_ips[LNET_INTERFACES_NUM]));
1302
1303         /* setup the socket AFTER I've received hello (it disables
1304          * SO_LINGER).  I might call back to the acceptor who may want
1305          * to send a protocol version response and then close the
1306          * socket; this ensures the socket only tears down after the
1307          * response has been sent. */
1308         if (rc == 0)
1309                 rc = ksocknal_lib_setup_sock(sock);
1310
1311         write_lock_bh(global_lock);
1312
1313         /* NB my callbacks block while I hold ksnd_global_lock */
1314         ksocknal_lib_set_callback(sock, conn);
1315
1316         if (!active)
1317                 peer_ni->ksnp_accepting--;
1318
1319         write_unlock_bh(global_lock);
1320
1321         if (rc != 0) {
1322                 write_lock_bh(global_lock);
1323                 if (!conn->ksnc_closing) {
1324                         /* could be closed by another thread */
1325                         ksocknal_close_conn_locked(conn, rc);
1326                 }
1327                 write_unlock_bh(global_lock);
1328         } else if (ksocknal_connsock_addref(conn) == 0) {
1329                 /* Allow I/O to proceed. */
1330                 ksocknal_read_callback(conn);
1331                 ksocknal_write_callback(conn);
1332                 ksocknal_connsock_decref(conn);
1333         }
1334
1335         ksocknal_connsock_decref(conn);
1336         ksocknal_conn_decref(conn);
1337         return rc;
1338
1339 failed_2:
1340         if (!peer_ni->ksnp_closing &&
1341             list_empty(&peer_ni->ksnp_conns) &&
1342             list_empty(&peer_ni->ksnp_routes)) {
1343                 list_splice_init(&peer_ni->ksnp_tx_queue, &zombies);
1344                 ksocknal_unlink_peer_locked(peer_ni);
1345         }
1346
1347         write_unlock_bh(global_lock);
1348
1349         if (warn != NULL) {
1350                 if (rc < 0)
1351                         CERROR("Not creating conn %s type %d: %s\n",
1352                                libcfs_id2str(peerid), conn->ksnc_type, warn);
1353                 else
1354                         CDEBUG(D_NET, "Not creating conn %s type %d: %s\n",
1355                               libcfs_id2str(peerid), conn->ksnc_type, warn);
1356         }
1357
1358         if (!active) {
1359                 if (rc > 0) {
1360                         /* Request retry by replying with CONN_NONE
1361                          * ksnc_proto has been set already */
1362                         conn->ksnc_type = SOCKLND_CONN_NONE;
1363                         hello->kshm_nips = 0;
1364                         ksocknal_send_hello(ni, conn, peerid.nid, hello);
1365                 }
1366
1367                 write_lock_bh(global_lock);
1368                 peer_ni->ksnp_accepting--;
1369                 write_unlock_bh(global_lock);
1370         }
1371
1372         /*
1373          * If we get here without an error code, just use -EALREADY.
1374          * Depending on how we got here, the error may be positive
1375          * or negative. Normalize the value for ksocknal_txlist_done().
1376          */
1377         rc2 = (rc == 0 ? -EALREADY : (rc > 0 ? -rc : rc));
1378         ksocknal_txlist_done(ni, &zombies, rc2);
1379         ksocknal_peer_decref(peer_ni);
1380
1381 failed_1:
1382         if (hello != NULL)
1383                 LIBCFS_FREE(hello, offsetof(struct ksock_hello_msg,
1384                                             kshm_ips[LNET_INTERFACES_NUM]));
1385
1386         LIBCFS_FREE(conn, sizeof(*conn));
1387
1388 failed_0:
1389         sock_release(sock);
1390         return rc;
1391 }
1392
1393 void
1394 ksocknal_close_conn_locked(struct ksock_conn *conn, int error)
1395 {
1396         /* This just does the immmediate housekeeping, and queues the
1397          * connection for the reaper to terminate.
1398          * Caller holds ksnd_global_lock exclusively in irq context */
1399         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1400         struct ksock_route *route;
1401         struct ksock_conn *conn2;
1402         struct list_head *tmp;
1403
1404         LASSERT(peer_ni->ksnp_error == 0);
1405         LASSERT(!conn->ksnc_closing);
1406         conn->ksnc_closing = 1;
1407
1408         /* ksnd_deathrow_conns takes over peer_ni's ref */
1409         list_del(&conn->ksnc_list);
1410
1411         route = conn->ksnc_route;
1412         if (route != NULL) {
1413                 /* dissociate conn from route... */
1414                 LASSERT(!route->ksnr_deleted);
1415                 LASSERT((route->ksnr_connected & (1 << conn->ksnc_type)) != 0);
1416
1417                 conn2 = NULL;
1418                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1419                         conn2 = list_entry(tmp, struct ksock_conn, ksnc_list);
1420
1421                         if (conn2->ksnc_route == route &&
1422                             conn2->ksnc_type == conn->ksnc_type)
1423                                 break;
1424
1425                         conn2 = NULL;
1426                 }
1427                 if (conn2 == NULL)
1428                         route->ksnr_connected &= ~(1 << conn->ksnc_type);
1429
1430                 conn->ksnc_route = NULL;
1431
1432                 ksocknal_route_decref(route);   /* drop conn's ref on route */
1433         }
1434
1435         if (list_empty(&peer_ni->ksnp_conns)) {
1436                 /* No more connections to this peer_ni */
1437
1438                 if (!list_empty(&peer_ni->ksnp_tx_queue)) {
1439                         struct ksock_tx *tx;
1440
1441                         LASSERT(conn->ksnc_proto == &ksocknal_protocol_v3x);
1442
1443                         /* throw them to the last connection...,
1444                          * these TXs will be send to /dev/null by scheduler */
1445                         list_for_each_entry(tx, &peer_ni->ksnp_tx_queue,
1446                                             tx_list)
1447                                 ksocknal_tx_prep(conn, tx);
1448
1449                         spin_lock_bh(&conn->ksnc_scheduler->kss_lock);
1450                         list_splice_init(&peer_ni->ksnp_tx_queue,
1451                                          &conn->ksnc_tx_queue);
1452                         spin_unlock_bh(&conn->ksnc_scheduler->kss_lock);
1453                 }
1454
1455                 /* renegotiate protocol version */
1456                 peer_ni->ksnp_proto = NULL;
1457                 /* stash last conn close reason */
1458                 peer_ni->ksnp_error = error;
1459
1460                 if (list_empty(&peer_ni->ksnp_routes)) {
1461                         /* I've just closed last conn belonging to a
1462                          * peer_ni with no routes to it */
1463                         ksocknal_unlink_peer_locked(peer_ni);
1464                 }
1465         }
1466
1467         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1468
1469         list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_deathrow_conns);
1470         wake_up(&ksocknal_data.ksnd_reaper_waitq);
1471
1472         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1473 }
1474
1475 void
1476 ksocknal_peer_failed(struct ksock_peer_ni *peer_ni)
1477 {
1478         int notify = 0;
1479         time64_t last_alive = 0;
1480
1481         /* There has been a connection failure or comms error; but I'll only
1482          * tell LNET I think the peer_ni is dead if it's to another kernel and
1483          * there are no connections or connection attempts in existence. */
1484
1485         read_lock(&ksocknal_data.ksnd_global_lock);
1486
1487         if ((peer_ni->ksnp_id.pid & LNET_PID_USERFLAG) == 0 &&
1488              list_empty(&peer_ni->ksnp_conns) &&
1489              peer_ni->ksnp_accepting == 0 &&
1490              ksocknal_find_connecting_route_locked(peer_ni) == NULL) {
1491                 notify = 1;
1492                 last_alive = peer_ni->ksnp_last_alive;
1493         }
1494
1495         read_unlock(&ksocknal_data.ksnd_global_lock);
1496
1497         if (notify)
1498                 lnet_notify(peer_ni->ksnp_ni, peer_ni->ksnp_id.nid,
1499                             false, false, last_alive);
1500 }
1501
1502 void
1503 ksocknal_finalize_zcreq(struct ksock_conn *conn)
1504 {
1505         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1506         struct ksock_tx *tx;
1507         struct ksock_tx *tmp;
1508         LIST_HEAD(zlist);
1509
1510         /* NB safe to finalize TXs because closing of socket will
1511          * abort all buffered data */
1512         LASSERT(conn->ksnc_sock == NULL);
1513
1514         spin_lock(&peer_ni->ksnp_lock);
1515
1516         list_for_each_entry_safe(tx, tmp, &peer_ni->ksnp_zc_req_list, tx_zc_list) {
1517                 if (tx->tx_conn != conn)
1518                         continue;
1519
1520                 LASSERT(tx->tx_msg.ksm_zc_cookies[0] != 0);
1521
1522                 tx->tx_msg.ksm_zc_cookies[0] = 0;
1523                 tx->tx_zc_aborted = 1;  /* mark it as not-acked */
1524                 list_move(&tx->tx_zc_list, &zlist);
1525         }
1526
1527         spin_unlock(&peer_ni->ksnp_lock);
1528
1529         while (!list_empty(&zlist)) {
1530                 tx = list_entry(zlist.next, struct ksock_tx, tx_zc_list);
1531
1532                 list_del(&tx->tx_zc_list);
1533                 ksocknal_tx_decref(tx);
1534         }
1535 }
1536
1537 void
1538 ksocknal_terminate_conn(struct ksock_conn *conn)
1539 {
1540         /* This gets called by the reaper (guaranteed thread context) to
1541          * disengage the socket from its callbacks and close it.
1542          * ksnc_refcount will eventually hit zero, and then the reaper will
1543          * destroy it. */
1544         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1545         struct ksock_sched *sched = conn->ksnc_scheduler;
1546         int failed = 0;
1547
1548         LASSERT(conn->ksnc_closing);
1549
1550         /* wake up the scheduler to "send" all remaining packets to /dev/null */
1551         spin_lock_bh(&sched->kss_lock);
1552
1553         /* a closing conn is always ready to tx */
1554         conn->ksnc_tx_ready = 1;
1555
1556         if (!conn->ksnc_tx_scheduled &&
1557             !list_empty(&conn->ksnc_tx_queue)) {
1558                 list_add_tail(&conn->ksnc_tx_list,
1559                                &sched->kss_tx_conns);
1560                 conn->ksnc_tx_scheduled = 1;
1561                 /* extra ref for scheduler */
1562                 ksocknal_conn_addref(conn);
1563
1564                 wake_up (&sched->kss_waitq);
1565         }
1566
1567         spin_unlock_bh(&sched->kss_lock);
1568
1569         /* serialise with callbacks */
1570         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1571
1572         ksocknal_lib_reset_callback(conn->ksnc_sock, conn);
1573
1574         /* OK, so this conn may not be completely disengaged from its
1575          * scheduler yet, but it _has_ committed to terminate... */
1576         conn->ksnc_scheduler->kss_nconns--;
1577
1578         if (peer_ni->ksnp_error != 0) {
1579                 /* peer_ni's last conn closed in error */
1580                 LASSERT(list_empty(&peer_ni->ksnp_conns));
1581                 failed = 1;
1582                 peer_ni->ksnp_error = 0;     /* avoid multiple notifications */
1583         }
1584
1585         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1586
1587         if (failed)
1588                 ksocknal_peer_failed(peer_ni);
1589
1590         /* The socket is closed on the final put; either here, or in
1591          * ksocknal_{send,recv}msg().  Since we set up the linger2 option
1592          * when the connection was established, this will close the socket
1593          * immediately, aborting anything buffered in it. Any hung
1594          * zero-copy transmits will therefore complete in finite time. */
1595         ksocknal_connsock_decref(conn);
1596 }
1597
1598 void
1599 ksocknal_queue_zombie_conn(struct ksock_conn *conn)
1600 {
1601         /* Queue the conn for the reaper to destroy */
1602         LASSERT(atomic_read(&conn->ksnc_conn_refcount) == 0);
1603         spin_lock_bh(&ksocknal_data.ksnd_reaper_lock);
1604
1605         list_add_tail(&conn->ksnc_list, &ksocknal_data.ksnd_zombie_conns);
1606         wake_up(&ksocknal_data.ksnd_reaper_waitq);
1607
1608         spin_unlock_bh(&ksocknal_data.ksnd_reaper_lock);
1609 }
1610
1611 void
1612 ksocknal_destroy_conn(struct ksock_conn *conn)
1613 {
1614         time64_t last_rcv;
1615
1616         /* Final coup-de-grace of the reaper */
1617         CDEBUG (D_NET, "connection %p\n", conn);
1618
1619         LASSERT (atomic_read (&conn->ksnc_conn_refcount) == 0);
1620         LASSERT (atomic_read (&conn->ksnc_sock_refcount) == 0);
1621         LASSERT (conn->ksnc_sock == NULL);
1622         LASSERT (conn->ksnc_route == NULL);
1623         LASSERT (!conn->ksnc_tx_scheduled);
1624         LASSERT (!conn->ksnc_rx_scheduled);
1625         LASSERT(list_empty(&conn->ksnc_tx_queue));
1626
1627         /* complete current receive if any */
1628         switch (conn->ksnc_rx_state) {
1629         case SOCKNAL_RX_LNET_PAYLOAD:
1630                 last_rcv = conn->ksnc_rx_deadline -
1631                            lnet_get_lnd_timeout();
1632                 CERROR("Completing partial receive from %s[%d], "
1633                        "ip %pI4h:%d, with error, wanted: %d, left: %d, "
1634                        "last alive is %lld secs ago\n",
1635                        libcfs_id2str(conn->ksnc_peer->ksnp_id), conn->ksnc_type,
1636                        &conn->ksnc_ipaddr, conn->ksnc_port,
1637                        conn->ksnc_rx_nob_wanted, conn->ksnc_rx_nob_left,
1638                        ktime_get_seconds() - last_rcv);
1639                 if (conn->ksnc_lnet_msg)
1640                         conn->ksnc_lnet_msg->msg_health_status =
1641                                 LNET_MSG_STATUS_REMOTE_ERROR;
1642                 lnet_finalize(conn->ksnc_lnet_msg, -EIO);
1643                 break;
1644         case SOCKNAL_RX_LNET_HEADER:
1645                 if (conn->ksnc_rx_started)
1646                         CERROR("Incomplete receive of lnet header from %s, "
1647                                "ip %pI4h:%d, with error, protocol: %d.x.\n",
1648                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1649                                &conn->ksnc_ipaddr, conn->ksnc_port,
1650                                conn->ksnc_proto->pro_version);
1651                 break;
1652         case SOCKNAL_RX_KSM_HEADER:
1653                 if (conn->ksnc_rx_started)
1654                         CERROR("Incomplete receive of ksock message from %s, "
1655                                "ip %pI4h:%d, with error, protocol: %d.x.\n",
1656                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1657                                &conn->ksnc_ipaddr, conn->ksnc_port,
1658                                conn->ksnc_proto->pro_version);
1659                 break;
1660         case SOCKNAL_RX_SLOP:
1661                 if (conn->ksnc_rx_started)
1662                         CERROR("Incomplete receive of slops from %s, "
1663                                "ip %pI4h:%d, with error\n",
1664                                libcfs_id2str(conn->ksnc_peer->ksnp_id),
1665                                &conn->ksnc_ipaddr, conn->ksnc_port);
1666                break;
1667         default:
1668                 LBUG ();
1669                 break;
1670         }
1671
1672         ksocknal_peer_decref(conn->ksnc_peer);
1673
1674         LIBCFS_FREE (conn, sizeof (*conn));
1675 }
1676
1677 int
1678 ksocknal_close_peer_conns_locked(struct ksock_peer_ni *peer_ni, __u32 ipaddr, int why)
1679 {
1680         struct ksock_conn *conn;
1681         struct list_head *ctmp;
1682         struct list_head *cnxt;
1683         int count = 0;
1684
1685         list_for_each_safe(ctmp, cnxt, &peer_ni->ksnp_conns) {
1686                 conn = list_entry(ctmp, struct ksock_conn, ksnc_list);
1687
1688                 if (ipaddr == 0 ||
1689                     conn->ksnc_ipaddr == ipaddr) {
1690                         count++;
1691                         ksocknal_close_conn_locked (conn, why);
1692                 }
1693         }
1694
1695         return (count);
1696 }
1697
1698 int
1699 ksocknal_close_conn_and_siblings(struct ksock_conn *conn, int why)
1700 {
1701         struct ksock_peer_ni *peer_ni = conn->ksnc_peer;
1702         u32 ipaddr = conn->ksnc_ipaddr;
1703         int count;
1704
1705         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1706
1707         count = ksocknal_close_peer_conns_locked (peer_ni, ipaddr, why);
1708
1709         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1710
1711         return (count);
1712 }
1713
1714 int
1715 ksocknal_close_matching_conns(struct lnet_process_id id, __u32 ipaddr)
1716 {
1717         struct ksock_peer_ni *peer_ni;
1718         struct hlist_node *pnxt;
1719         int lo;
1720         int hi;
1721         int i;
1722         int count = 0;
1723
1724         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1725
1726         if (id.nid != LNET_NID_ANY) {
1727                 lo = hash_min(id.nid, HASH_BITS(ksocknal_data.ksnd_peers));
1728                 hi = lo;
1729         } else {
1730                 lo = 0;
1731                 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1732         }
1733
1734         for (i = lo; i <= hi; i++) {
1735                 hlist_for_each_entry_safe(peer_ni, pnxt,
1736                                           &ksocknal_data.ksnd_peers[i],
1737                                           ksnp_list) {
1738
1739                         if (!((id.nid == LNET_NID_ANY ||
1740                                id.nid == peer_ni->ksnp_id.nid) &&
1741                               (id.pid == LNET_PID_ANY ||
1742                                id.pid == peer_ni->ksnp_id.pid)))
1743                                 continue;
1744
1745                         count += ksocknal_close_peer_conns_locked(peer_ni,
1746                                                                   ipaddr, 0);
1747                 }
1748         }
1749
1750         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1751
1752         /* wildcards always succeed */
1753         if (id.nid == LNET_NID_ANY || id.pid == LNET_PID_ANY || ipaddr == 0)
1754                 return 0;
1755
1756         return (count == 0 ? -ENOENT : 0);
1757 }
1758
1759 void
1760 ksocknal_notify_gw_down(lnet_nid_t gw_nid)
1761 {
1762         /* The router is telling me she's been notified of a change in
1763          * gateway state....
1764          */
1765         struct lnet_process_id id = {
1766                 .nid    = gw_nid,
1767                 .pid    = LNET_PID_ANY,
1768         };
1769
1770         CDEBUG(D_NET, "gw %s down\n", libcfs_nid2str(gw_nid));
1771
1772         /* If the gateway crashed, close all open connections... */
1773         ksocknal_close_matching_conns(id, 0);
1774         return;
1775
1776         /* We can only establish new connections
1777          * if we have autroutes, and these connect on demand. */
1778 }
1779
1780 static void
1781 ksocknal_push_peer(struct ksock_peer_ni *peer_ni)
1782 {
1783         int index;
1784         int i;
1785         struct list_head *tmp;
1786         struct ksock_conn *conn;
1787
1788         for (index = 0; ; index++) {
1789                 read_lock(&ksocknal_data.ksnd_global_lock);
1790
1791                 i = 0;
1792                 conn = NULL;
1793
1794                 list_for_each(tmp, &peer_ni->ksnp_conns) {
1795                         if (i++ == index) {
1796                                 conn = list_entry(tmp, struct ksock_conn,
1797                                                   ksnc_list);
1798                                 ksocknal_conn_addref(conn);
1799                                 break;
1800                         }
1801                 }
1802
1803                 read_unlock(&ksocknal_data.ksnd_global_lock);
1804
1805                 if (conn == NULL)
1806                         break;
1807
1808                 ksocknal_lib_push_conn (conn);
1809                 ksocknal_conn_decref(conn);
1810         }
1811 }
1812
1813 static int
1814 ksocknal_push(struct lnet_ni *ni, struct lnet_process_id id)
1815 {
1816         int lo;
1817         int hi;
1818         int bkt;
1819         int rc = -ENOENT;
1820
1821         if (id.nid != LNET_NID_ANY) {
1822                 lo = hash_min(id.nid, HASH_BITS(ksocknal_data.ksnd_peers));
1823                 hi = lo;
1824         } else {
1825                 lo = 0;
1826                 hi = HASH_SIZE(ksocknal_data.ksnd_peers) - 1;
1827         }
1828
1829         for (bkt = lo; bkt <= hi; bkt++) {
1830                 int peer_off; /* searching offset in peer_ni hash table */
1831
1832                 for (peer_off = 0; ; peer_off++) {
1833                         struct ksock_peer_ni *peer_ni;
1834                         int           i = 0;
1835
1836                         read_lock(&ksocknal_data.ksnd_global_lock);
1837                         hlist_for_each_entry(peer_ni,
1838                                              &ksocknal_data.ksnd_peers[bkt],
1839                                              ksnp_list) {
1840                                 if (!((id.nid == LNET_NID_ANY ||
1841                                        id.nid == peer_ni->ksnp_id.nid) &&
1842                                       (id.pid == LNET_PID_ANY ||
1843                                        id.pid == peer_ni->ksnp_id.pid)))
1844                                         continue;
1845
1846                                 if (i++ == peer_off) {
1847                                         ksocknal_peer_addref(peer_ni);
1848                                         break;
1849                                 }
1850                         }
1851                         read_unlock(&ksocknal_data.ksnd_global_lock);
1852
1853                         if (i <= peer_off) /* no match */
1854                                 break;
1855
1856                         rc = 0;
1857                         ksocknal_push_peer(peer_ni);
1858                         ksocknal_peer_decref(peer_ni);
1859                 }
1860         }
1861         return rc;
1862 }
1863
1864 static int
1865 ksocknal_add_interface(struct lnet_ni *ni, __u32 ipaddress, __u32 netmask)
1866 {
1867         struct ksock_net *net = ni->ni_data;
1868         struct ksock_interface *iface;
1869         int rc;
1870         int i;
1871         int j;
1872         struct ksock_peer_ni *peer_ni;
1873         struct list_head *rtmp;
1874         struct ksock_route *route;
1875
1876         if (ipaddress == 0 ||
1877             netmask == 0)
1878                 return -EINVAL;
1879
1880         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1881
1882         iface = ksocknal_ip2iface(ni, ipaddress);
1883         if (iface != NULL) {
1884                 /* silently ignore dups */
1885                 rc = 0;
1886         } else if (net->ksnn_ninterfaces == LNET_INTERFACES_NUM) {
1887                 rc = -ENOSPC;
1888         } else {
1889                 iface = &net->ksnn_interfaces[net->ksnn_ninterfaces++];
1890
1891                 iface->ksni_ipaddr = ipaddress;
1892                 iface->ksni_netmask = netmask;
1893                 iface->ksni_nroutes = 0;
1894                 iface->ksni_npeers = 0;
1895
1896                 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
1897                         for (j = 0; j < peer_ni->ksnp_n_passive_ips; j++)
1898                                 if (peer_ni->ksnp_passive_ips[j] == ipaddress)
1899                                         iface->ksni_npeers++;
1900
1901                         list_for_each(rtmp, &peer_ni->ksnp_routes) {
1902                                 route = list_entry(rtmp,
1903                                                    struct ksock_route,
1904                                                    ksnr_list);
1905
1906                                 if (route->ksnr_myipaddr == ipaddress)
1907                                         iface->ksni_nroutes++;
1908                         }
1909                 }
1910
1911                 rc = 0;
1912                 /* NB only new connections will pay attention to the new
1913                  * interface!
1914                  */
1915         }
1916
1917         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
1918
1919         return rc;
1920 }
1921
1922 static void
1923 ksocknal_peer_del_interface_locked(struct ksock_peer_ni *peer_ni, __u32 ipaddr)
1924 {
1925         struct list_head *tmp;
1926         struct list_head *nxt;
1927         struct ksock_route *route;
1928         struct ksock_conn *conn;
1929         int i;
1930         int j;
1931
1932         for (i = 0; i < peer_ni->ksnp_n_passive_ips; i++)
1933                 if (peer_ni->ksnp_passive_ips[i] == ipaddr) {
1934                         for (j = i+1; j < peer_ni->ksnp_n_passive_ips; j++)
1935                                 peer_ni->ksnp_passive_ips[j-1] =
1936                                         peer_ni->ksnp_passive_ips[j];
1937                         peer_ni->ksnp_n_passive_ips--;
1938                         break;
1939                 }
1940
1941         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_routes) {
1942                 route = list_entry(tmp, struct ksock_route, ksnr_list);
1943
1944                 if (route->ksnr_myipaddr != ipaddr)
1945                         continue;
1946
1947                 if (route->ksnr_share_count != 0) {
1948                         /* Manually created; keep, but unbind */
1949                         route->ksnr_myipaddr = 0;
1950                 } else {
1951                         ksocknal_del_route_locked(route);
1952                 }
1953         }
1954
1955         list_for_each_safe(tmp, nxt, &peer_ni->ksnp_conns) {
1956                 conn = list_entry(tmp, struct ksock_conn, ksnc_list);
1957
1958                 if (conn->ksnc_myipaddr == ipaddr)
1959                         ksocknal_close_conn_locked (conn, 0);
1960         }
1961 }
1962
1963 static int
1964 ksocknal_del_interface(struct lnet_ni *ni, __u32 ipaddress)
1965 {
1966         struct ksock_net *net = ni->ni_data;
1967         int rc = -ENOENT;
1968         struct hlist_node *nxt;
1969         struct ksock_peer_ni *peer_ni;
1970         u32 this_ip;
1971         int i;
1972         int j;
1973
1974         write_lock_bh(&ksocknal_data.ksnd_global_lock);
1975
1976         for (i = 0; i < net->ksnn_ninterfaces; i++) {
1977                 this_ip = net->ksnn_interfaces[i].ksni_ipaddr;
1978
1979                 if (!(ipaddress == 0 ||
1980                       ipaddress == this_ip))
1981                         continue;
1982
1983                 rc = 0;
1984
1985                 for (j = i+1; j < net->ksnn_ninterfaces; j++)
1986                         net->ksnn_interfaces[j-1] =
1987                                 net->ksnn_interfaces[j];
1988
1989                 net->ksnn_ninterfaces--;
1990
1991                 hash_for_each_safe(ksocknal_data.ksnd_peers, j,
1992                                    nxt, peer_ni, ksnp_list) {
1993                         if (peer_ni->ksnp_ni != ni)
1994                                 continue;
1995
1996                         ksocknal_peer_del_interface_locked(peer_ni, this_ip);
1997                 }
1998         }
1999
2000         write_unlock_bh(&ksocknal_data.ksnd_global_lock);
2001
2002         return rc;
2003 }
2004
2005 int
2006 ksocknal_ctl(struct lnet_ni *ni, unsigned int cmd, void *arg)
2007 {
2008         struct lnet_process_id id = {0};
2009         struct libcfs_ioctl_data *data = arg;
2010         int rc;
2011
2012         switch(cmd) {
2013         case IOC_LIBCFS_GET_INTERFACE: {
2014                 struct ksock_net *net = ni->ni_data;
2015                 struct ksock_interface *iface;
2016
2017                 read_lock(&ksocknal_data.ksnd_global_lock);
2018
2019                 if (data->ioc_count >= (__u32)net->ksnn_ninterfaces) {
2020                         rc = -ENOENT;
2021                 } else {
2022                         rc = 0;
2023                         iface = &net->ksnn_interfaces[data->ioc_count];
2024
2025                         data->ioc_u32[0] = iface->ksni_ipaddr;
2026                         data->ioc_u32[1] = iface->ksni_netmask;
2027                         data->ioc_u32[2] = iface->ksni_npeers;
2028                         data->ioc_u32[3] = iface->ksni_nroutes;
2029                 }
2030
2031                 read_unlock(&ksocknal_data.ksnd_global_lock);
2032                 return rc;
2033         }
2034
2035         case IOC_LIBCFS_ADD_INTERFACE:
2036                 return ksocknal_add_interface(ni,
2037                                               data->ioc_u32[0], /* IP address */
2038                                               data->ioc_u32[1]); /* net mask */
2039
2040         case IOC_LIBCFS_DEL_INTERFACE:
2041                 return ksocknal_del_interface(ni,
2042                                               data->ioc_u32[0]); /* IP address */
2043
2044         case IOC_LIBCFS_GET_PEER: {
2045                 __u32            myip = 0;
2046                 __u32            ip = 0;
2047                 int              port = 0;
2048                 int              conn_count = 0;
2049                 int              share_count = 0;
2050
2051                 rc = ksocknal_get_peer_info(ni, data->ioc_count,
2052                                             &id, &myip, &ip, &port,
2053                                             &conn_count,  &share_count);
2054                 if (rc != 0)
2055                         return rc;
2056
2057                 data->ioc_nid    = id.nid;
2058                 data->ioc_count  = share_count;
2059                 data->ioc_u32[0] = ip;
2060                 data->ioc_u32[1] = port;
2061                 data->ioc_u32[2] = myip;
2062                 data->ioc_u32[3] = conn_count;
2063                 data->ioc_u32[4] = id.pid;
2064                 return 0;
2065         }
2066
2067         case IOC_LIBCFS_ADD_PEER:
2068                 id.nid = data->ioc_nid;
2069                 id.pid = LNET_PID_LUSTRE;
2070                 return ksocknal_add_peer (ni, id,
2071                                           data->ioc_u32[0], /* IP */
2072                                           data->ioc_u32[1]); /* port */
2073
2074         case IOC_LIBCFS_DEL_PEER:
2075                 id.nid = data->ioc_nid;
2076                 id.pid = LNET_PID_ANY;
2077                 return ksocknal_del_peer (ni, id,
2078                                           data->ioc_u32[0]); /* IP */
2079
2080         case IOC_LIBCFS_GET_CONN: {
2081                 int           txmem;
2082                 int           rxmem;
2083                 int           nagle;
2084                 struct ksock_conn *conn = ksocknal_get_conn_by_idx(ni, data->ioc_count);
2085
2086                 if (conn == NULL)
2087                         return -ENOENT;
2088
2089                 ksocknal_lib_get_conn_tunables(conn, &txmem, &rxmem, &nagle);
2090
2091                 data->ioc_count  = txmem;
2092                 data->ioc_nid    = conn->ksnc_peer->ksnp_id.nid;
2093                 data->ioc_flags  = nagle;
2094                 data->ioc_u32[0] = conn->ksnc_ipaddr;
2095                 data->ioc_u32[1] = conn->ksnc_port;
2096                 data->ioc_u32[2] = conn->ksnc_myipaddr;
2097                 data->ioc_u32[3] = conn->ksnc_type;
2098                 data->ioc_u32[4] = conn->ksnc_scheduler->kss_cpt;
2099                 data->ioc_u32[5] = rxmem;
2100                 data->ioc_u32[6] = conn->ksnc_peer->ksnp_id.pid;
2101                 ksocknal_conn_decref(conn);
2102                 return 0;
2103         }
2104
2105         case IOC_LIBCFS_CLOSE_CONNECTION:
2106                 id.nid = data->ioc_nid;
2107                 id.pid = LNET_PID_ANY;
2108                 return ksocknal_close_matching_conns (id,
2109                                                       data->ioc_u32[0]);
2110
2111         case IOC_LIBCFS_REGISTER_MYNID:
2112                 /* Ignore if this is a noop */
2113                 if (data->ioc_nid == ni->ni_nid)
2114                         return 0;
2115
2116                 CERROR("obsolete IOC_LIBCFS_REGISTER_MYNID: %s(%s)\n",
2117                        libcfs_nid2str(data->ioc_nid),
2118                        libcfs_nid2str(ni->ni_nid));
2119                 return -EINVAL;
2120
2121         case IOC_LIBCFS_PUSH_CONNECTION:
2122                 id.nid = data->ioc_nid;
2123                 id.pid = LNET_PID_ANY;
2124                 return ksocknal_push(ni, id);
2125
2126         default:
2127                 return -EINVAL;
2128         }
2129         /* not reached */
2130 }
2131
2132 static void
2133 ksocknal_free_buffers (void)
2134 {
2135         LASSERT (atomic_read(&ksocknal_data.ksnd_nactive_txs) == 0);
2136
2137         if (ksocknal_data.ksnd_schedulers != NULL)
2138                 cfs_percpt_free(ksocknal_data.ksnd_schedulers);
2139
2140         spin_lock(&ksocknal_data.ksnd_tx_lock);
2141
2142         if (!list_empty(&ksocknal_data.ksnd_idle_noop_txs)) {
2143                 LIST_HEAD(zlist);
2144                 struct ksock_tx *tx;
2145
2146                 list_splice_init(&ksocknal_data.ksnd_idle_noop_txs, &zlist);
2147                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2148
2149                 while (!list_empty(&zlist)) {
2150                         tx = list_entry(zlist.next, struct ksock_tx, tx_list);
2151                         list_del(&tx->tx_list);
2152                         LIBCFS_FREE(tx, tx->tx_desc_size);
2153                 }
2154         } else {
2155                 spin_unlock(&ksocknal_data.ksnd_tx_lock);
2156         }
2157 }
2158
2159 static void
2160 ksocknal_base_shutdown(void)
2161 {
2162         struct ksock_sched *sched;
2163         struct ksock_peer_ni *peer_ni;
2164         int i;
2165
2166         CDEBUG(D_MALLOC, "before NAL cleanup: kmem %d\n",
2167                atomic_read (&libcfs_kmemory));
2168         LASSERT (ksocknal_data.ksnd_nnets == 0);
2169
2170         switch (ksocknal_data.ksnd_init) {
2171         default:
2172                 LASSERT(0);
2173                 /* fallthrough */
2174
2175         case SOCKNAL_INIT_ALL:
2176         case SOCKNAL_INIT_DATA:
2177                 hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list)
2178                         LASSERT(0);
2179
2180                 LASSERT(list_empty(&ksocknal_data.ksnd_nets));
2181                 LASSERT(list_empty(&ksocknal_data.ksnd_enomem_conns));
2182                 LASSERT(list_empty(&ksocknal_data.ksnd_zombie_conns));
2183                 LASSERT(list_empty(&ksocknal_data.ksnd_connd_connreqs));
2184                 LASSERT(list_empty(&ksocknal_data.ksnd_connd_routes));
2185
2186                 if (ksocknal_data.ksnd_schedulers != NULL) {
2187                         cfs_percpt_for_each(sched, i,
2188                                             ksocknal_data.ksnd_schedulers) {
2189
2190                                 LASSERT(list_empty(&sched->kss_tx_conns));
2191                                 LASSERT(list_empty(&sched->kss_rx_conns));
2192                                 LASSERT(list_empty(&sched->kss_zombie_noop_txs));
2193                                 LASSERT(sched->kss_nconns == 0);
2194                         }
2195                 }
2196
2197                 /* flag threads to terminate; wake and wait for them to die */
2198                 ksocknal_data.ksnd_shuttingdown = 1;
2199                 wake_up_all(&ksocknal_data.ksnd_connd_waitq);
2200                 wake_up_all(&ksocknal_data.ksnd_reaper_waitq);
2201
2202                 if (ksocknal_data.ksnd_schedulers != NULL) {
2203                         cfs_percpt_for_each(sched, i,
2204                                             ksocknal_data.ksnd_schedulers)
2205                                         wake_up_all(&sched->kss_waitq);
2206                 }
2207
2208                 i = 4;
2209                 read_lock(&ksocknal_data.ksnd_global_lock);
2210                 while (ksocknal_data.ksnd_nthreads != 0) {
2211                         i++;
2212                         /* power of 2? */
2213                         CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET,
2214                                 "waiting for %d threads to terminate\n",
2215                                 ksocknal_data.ksnd_nthreads);
2216                         read_unlock(&ksocknal_data.ksnd_global_lock);
2217                         schedule_timeout_uninterruptible(cfs_time_seconds(1));
2218                         read_lock(&ksocknal_data.ksnd_global_lock);
2219                 }
2220                 read_unlock(&ksocknal_data.ksnd_global_lock);
2221
2222                 ksocknal_free_buffers();
2223
2224                 ksocknal_data.ksnd_init = SOCKNAL_INIT_NOTHING;
2225                 break;
2226         }
2227
2228         CDEBUG(D_MALLOC, "after NAL cleanup: kmem %d\n",
2229                atomic_read (&libcfs_kmemory));
2230
2231         module_put(THIS_MODULE);
2232 }
2233
2234 static int
2235 ksocknal_base_startup(void)
2236 {
2237         struct ksock_sched *sched;
2238         int rc;
2239         int i;
2240
2241         LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING);
2242         LASSERT(ksocknal_data.ksnd_nnets == 0);
2243
2244         memset(&ksocknal_data, 0, sizeof(ksocknal_data)); /* zero pointers */
2245
2246         hash_init(ksocknal_data.ksnd_peers);
2247
2248         rwlock_init(&ksocknal_data.ksnd_global_lock);
2249         INIT_LIST_HEAD(&ksocknal_data.ksnd_nets);
2250
2251         spin_lock_init(&ksocknal_data.ksnd_reaper_lock);
2252         INIT_LIST_HEAD(&ksocknal_data.ksnd_enomem_conns);
2253         INIT_LIST_HEAD(&ksocknal_data.ksnd_zombie_conns);
2254         INIT_LIST_HEAD(&ksocknal_data.ksnd_deathrow_conns);
2255         init_waitqueue_head(&ksocknal_data.ksnd_reaper_waitq);
2256
2257         spin_lock_init(&ksocknal_data.ksnd_connd_lock);
2258         INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_connreqs);
2259         INIT_LIST_HEAD(&ksocknal_data.ksnd_connd_routes);
2260         init_waitqueue_head(&ksocknal_data.ksnd_connd_waitq);
2261
2262         spin_lock_init(&ksocknal_data.ksnd_tx_lock);
2263         INIT_LIST_HEAD(&ksocknal_data.ksnd_idle_noop_txs);
2264
2265         /* NB memset above zeros whole of ksocknal_data */
2266
2267         /* flag lists/ptrs/locks initialised */
2268         ksocknal_data.ksnd_init = SOCKNAL_INIT_DATA;
2269         if (!try_module_get(THIS_MODULE))
2270                 goto failed;
2271
2272         /* Create a scheduler block per available CPT */
2273         ksocknal_data.ksnd_schedulers = cfs_percpt_alloc(lnet_cpt_table(),
2274                                                          sizeof(*sched));
2275         if (ksocknal_data.ksnd_schedulers == NULL)
2276                 goto failed;
2277
2278         cfs_percpt_for_each(sched, i, ksocknal_data.ksnd_schedulers) {
2279                 int nthrs;
2280
2281                 /*
2282                  * make sure not to allocate more threads than there are
2283                  * cores/CPUs in teh CPT
2284                  */
2285                 nthrs = cfs_cpt_weight(lnet_cpt_table(), i);
2286                 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2287                         nthrs = min(nthrs, *ksocknal_tunables.ksnd_nscheds);
2288                 } else {
2289                         /*
2290                          * max to half of CPUs, assume another half should be
2291                          * reserved for upper layer modules
2292                          */
2293                         nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2294                 }
2295
2296                 sched->kss_nthreads_max = nthrs;
2297                 sched->kss_cpt = i;
2298
2299                 spin_lock_init(&sched->kss_lock);
2300                 INIT_LIST_HEAD(&sched->kss_rx_conns);
2301                 INIT_LIST_HEAD(&sched->kss_tx_conns);
2302                 INIT_LIST_HEAD(&sched->kss_zombie_noop_txs);
2303                 init_waitqueue_head(&sched->kss_waitq);
2304         }
2305
2306         ksocknal_data.ksnd_connd_starting         = 0;
2307         ksocknal_data.ksnd_connd_failed_stamp     = 0;
2308         ksocknal_data.ksnd_connd_starting_stamp   = ktime_get_real_seconds();
2309         /* must have at least 2 connds to remain responsive to accepts while
2310          * connecting */
2311         if (*ksocknal_tunables.ksnd_nconnds < SOCKNAL_CONND_RESV + 1)
2312                 *ksocknal_tunables.ksnd_nconnds = SOCKNAL_CONND_RESV + 1;
2313
2314         if (*ksocknal_tunables.ksnd_nconnds_max <
2315             *ksocknal_tunables.ksnd_nconnds) {
2316                 ksocknal_tunables.ksnd_nconnds_max =
2317                         ksocknal_tunables.ksnd_nconnds;
2318         }
2319
2320         for (i = 0; i < *ksocknal_tunables.ksnd_nconnds; i++) {
2321                 char name[16];
2322                 spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2323                 ksocknal_data.ksnd_connd_starting++;
2324                 spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2325
2326
2327                 snprintf(name, sizeof(name), "socknal_cd%02d", i);
2328                 rc = ksocknal_thread_start(ksocknal_connd,
2329                                            (void *)((uintptr_t)i), name);
2330                 if (rc != 0) {
2331                         spin_lock_bh(&ksocknal_data.ksnd_connd_lock);
2332                         ksocknal_data.ksnd_connd_starting--;
2333                         spin_unlock_bh(&ksocknal_data.ksnd_connd_lock);
2334                         CERROR("Can't spawn socknal connd: %d\n", rc);
2335                         goto failed;
2336                 }
2337         }
2338
2339         rc = ksocknal_thread_start(ksocknal_reaper, NULL, "socknal_reaper");
2340         if (rc != 0) {
2341                 CERROR ("Can't spawn socknal reaper: %d\n", rc);
2342                 goto failed;
2343         }
2344
2345         /* flag everything initialised */
2346         ksocknal_data.ksnd_init = SOCKNAL_INIT_ALL;
2347
2348         return 0;
2349
2350  failed:
2351         ksocknal_base_shutdown();
2352         return -ENETDOWN;
2353 }
2354
2355 static void
2356 ksocknal_debug_peerhash(struct lnet_ni *ni)
2357 {
2358         struct ksock_peer_ni *peer_ni;
2359         int i;
2360
2361         read_lock(&ksocknal_data.ksnd_global_lock);
2362
2363         hash_for_each(ksocknal_data.ksnd_peers, i, peer_ni, ksnp_list) {
2364                 struct ksock_route *route;
2365                 struct ksock_conn *conn;
2366
2367                 if (peer_ni->ksnp_ni != ni)
2368                         continue;
2369
2370                 CWARN("Active peer_ni on shutdown: %s, ref %d, "
2371                       "closing %d, accepting %d, err %d, zcookie %llu, "
2372                       "txq %d, zc_req %d\n", libcfs_id2str(peer_ni->ksnp_id),
2373                       atomic_read(&peer_ni->ksnp_refcount),
2374                       peer_ni->ksnp_closing,
2375                       peer_ni->ksnp_accepting, peer_ni->ksnp_error,
2376                       peer_ni->ksnp_zc_next_cookie,
2377                       !list_empty(&peer_ni->ksnp_tx_queue),
2378                       !list_empty(&peer_ni->ksnp_zc_req_list));
2379
2380                 list_for_each_entry(route, &peer_ni->ksnp_routes, ksnr_list) {
2381                         CWARN("Route: ref %d, schd %d, conn %d, cnted %d, "
2382                               "del %d\n", atomic_read(&route->ksnr_refcount),
2383                               route->ksnr_scheduled, route->ksnr_connecting,
2384                               route->ksnr_connected, route->ksnr_deleted);
2385                 }
2386
2387                 list_for_each_entry(conn, &peer_ni->ksnp_conns, ksnc_list) {
2388                         CWARN("Conn: ref %d, sref %d, t %d, c %d\n",
2389                               atomic_read(&conn->ksnc_conn_refcount),
2390                               atomic_read(&conn->ksnc_sock_refcount),
2391                               conn->ksnc_type, conn->ksnc_closing);
2392                 }
2393                 break;
2394         }
2395
2396         read_unlock(&ksocknal_data.ksnd_global_lock);
2397 }
2398
2399 void
2400 ksocknal_shutdown(struct lnet_ni *ni)
2401 {
2402         struct ksock_net *net = ni->ni_data;
2403         struct lnet_process_id anyid = {
2404                 .nid = LNET_NID_ANY,
2405                 .pid = LNET_PID_ANY,
2406         };
2407         int i;
2408
2409         LASSERT(ksocknal_data.ksnd_init == SOCKNAL_INIT_ALL);
2410         LASSERT(ksocknal_data.ksnd_nnets > 0);
2411
2412         /* prevent new peers */
2413         atomic_add(SOCKNAL_SHUTDOWN_BIAS, &net->ksnn_npeers);
2414
2415         /* Delete all peers */
2416         ksocknal_del_peer(ni, anyid, 0);
2417
2418         /* Wait for all peer_ni state to clean up */
2419         i = 2;
2420         while (atomic_read(&net->ksnn_npeers) > SOCKNAL_SHUTDOWN_BIAS) {
2421                 i++;
2422                 CDEBUG(((i & (-i)) == i) ? D_WARNING : D_NET, /* power of 2? */
2423                        "waiting for %d peers to disconnect\n",
2424                        atomic_read(&net->ksnn_npeers) - SOCKNAL_SHUTDOWN_BIAS);
2425                 schedule_timeout_uninterruptible(cfs_time_seconds(1));
2426
2427                 ksocknal_debug_peerhash(ni);
2428         }
2429
2430         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2431                 LASSERT(net->ksnn_interfaces[i].ksni_npeers == 0);
2432                 LASSERT(net->ksnn_interfaces[i].ksni_nroutes == 0);
2433         }
2434
2435         list_del(&net->ksnn_list);
2436         LIBCFS_FREE(net, sizeof(*net));
2437
2438         ksocknal_data.ksnd_nnets--;
2439         if (ksocknal_data.ksnd_nnets == 0)
2440                 ksocknal_base_shutdown();
2441 }
2442
2443 static int
2444 ksocknal_search_new_ipif(struct ksock_net *net)
2445 {
2446         int new_ipif = 0;
2447         int i;
2448
2449         for (i = 0; i < net->ksnn_ninterfaces; i++) {
2450                 char *ifnam = &net->ksnn_interfaces[i].ksni_name[0];
2451                 char *colon = strchr(ifnam, ':');
2452                 int found  = 0;
2453                 struct ksock_net *tmp;
2454                 int j;
2455
2456                 if (colon != NULL) /* ignore alias device */
2457                         *colon = 0;
2458
2459                 list_for_each_entry(tmp, &ksocknal_data.ksnd_nets,
2460                                         ksnn_list) {
2461                         for (j = 0; !found && j < tmp->ksnn_ninterfaces; j++) {
2462                                 char *ifnam2 = &tmp->ksnn_interfaces[j].\
2463                                              ksni_name[0];
2464                                 char *colon2 = strchr(ifnam2, ':');
2465
2466                                 if (colon2 != NULL)
2467                                         *colon2 = 0;
2468
2469                                 found = strcmp(ifnam, ifnam2) == 0;
2470                                 if (colon2 != NULL)
2471                                         *colon2 = ':';
2472                         }
2473                         if (found)
2474                                 break;
2475                 }
2476
2477                 new_ipif += !found;
2478                 if (colon != NULL)
2479                         *colon = ':';
2480         }
2481
2482         return new_ipif;
2483 }
2484
2485 static int
2486 ksocknal_start_schedulers(struct ksock_sched *sched)
2487 {
2488         int     nthrs;
2489         int     rc = 0;
2490         int     i;
2491
2492         if (sched->kss_nthreads == 0) {
2493                 if (*ksocknal_tunables.ksnd_nscheds > 0) {
2494                         nthrs = sched->kss_nthreads_max;
2495                 } else {
2496                         nthrs = cfs_cpt_weight(lnet_cpt_table(),
2497                                                sched->kss_cpt);
2498                         nthrs = min(max(SOCKNAL_NSCHEDS, nthrs >> 1), nthrs);
2499                         nthrs = min(SOCKNAL_NSCHEDS_HIGH, nthrs);
2500                 }
2501                 nthrs = min(nthrs, sched->kss_nthreads_max);
2502         } else {
2503                 LASSERT(sched->kss_nthreads <= sched->kss_nthreads_max);
2504                 /* increase two threads if there is new interface */
2505                 nthrs = min(2, sched->kss_nthreads_max - sched->kss_nthreads);
2506         }
2507
2508         for (i = 0; i < nthrs; i++) {
2509                 long id;
2510                 char name[20];
2511
2512                 id = KSOCK_THREAD_ID(sched->kss_cpt, sched->kss_nthreads + i);
2513                 snprintf(name, sizeof(name), "socknal_sd%02d_%02d",
2514                          sched->kss_cpt, (int)KSOCK_THREAD_SID(id));
2515
2516                 rc = ksocknal_thread_start(ksocknal_scheduler,
2517                                            (void *)id, name);
2518                 if (rc == 0)
2519                         continue;
2520
2521                 CERROR("Can't spawn thread %d for scheduler[%d]: %d\n",
2522                        sched->kss_cpt, (int) KSOCK_THREAD_SID(id), rc);
2523                 break;
2524         }
2525
2526         sched->kss_nthreads += i;
2527         return rc;
2528 }
2529
2530 static int
2531 ksocknal_net_start_threads(struct ksock_net *net, __u32 *cpts, int ncpts)
2532 {
2533         int newif = ksocknal_search_new_ipif(net);
2534         int rc;
2535         int i;
2536
2537         if (ncpts > 0 && ncpts > cfs_cpt_number(lnet_cpt_table()))
2538                 return -EINVAL;
2539
2540         for (i = 0; i < ncpts; i++) {
2541                 struct ksock_sched *sched;
2542                 int cpt = (cpts == NULL) ? i : cpts[i];
2543
2544                 LASSERT(cpt < cfs_cpt_number(lnet_cpt_table()));
2545                 sched = ksocknal_data.ksnd_schedulers[cpt];
2546
2547                 if (!newif && sched->kss_nthreads > 0)
2548                         continue;
2549
2550                 rc = ksocknal_start_schedulers(sched);
2551                 if (rc != 0)
2552                         return rc;
2553         }
2554         return 0;
2555 }
2556
2557 int
2558 ksocknal_startup(struct lnet_ni *ni)
2559 {
2560         struct ksock_net *net;
2561         struct lnet_ioctl_config_lnd_cmn_tunables *net_tunables;
2562         struct ksock_interface *ksi = NULL;
2563         struct lnet_inetdev *ifaces = NULL;
2564         int i = 0;
2565         int rc;
2566
2567         LASSERT (ni->ni_net->net_lnd == &the_ksocklnd);
2568
2569         if (ksocknal_data.ksnd_init == SOCKNAL_INIT_NOTHING) {
2570                 rc = ksocknal_base_startup();
2571                 if (rc != 0)
2572                         return rc;
2573         }
2574
2575         LIBCFS_ALLOC(net, sizeof(*net));
2576         if (net == NULL)
2577                 goto fail_0;
2578
2579         net->ksnn_incarnation = ktime_get_real_ns();
2580         ni->ni_data = net;
2581         net_tunables = &ni->ni_net->net_tunables;
2582
2583         if (net_tunables->lct_peer_timeout == -1)
2584                 net_tunables->lct_peer_timeout =
2585                         *ksocknal_tunables.ksnd_peertimeout;
2586
2587         if (net_tunables->lct_max_tx_credits == -1)
2588                 net_tunables->lct_max_tx_credits =
2589                         *ksocknal_tunables.ksnd_credits;
2590
2591         if (net_tunables->lct_peer_tx_credits == -1)
2592                 net_tunables->lct_peer_tx_credits =
2593                         *ksocknal_tunables.ksnd_peertxcredits;
2594
2595         if (net_tunables->lct_peer_tx_credits >
2596             net_tunables->lct_max_tx_credits)
2597                 net_tunables->lct_peer_tx_credits =
2598                         net_tunables->lct_max_tx_credits;
2599
2600         if (net_tunables->lct_peer_rtr_credits == -1)
2601                 net_tunables->lct_peer_rtr_credits =
2602                         *ksocknal_tunables.ksnd_peerrtrcredits;
2603
2604         rc = lnet_inet_enumerate(&ifaces, ni->ni_net_ns);
2605         if (rc < 0)
2606                 goto fail_1;
2607
2608         if (!ni->ni_interfaces[0]) {
2609                 ksi = &net->ksnn_interfaces[0];
2610
2611                 /* Use the first discovered interface */
2612                 net->ksnn_ninterfaces = 1;
2613                 ni->ni_dev_cpt = ifaces[0].li_cpt;
2614                 ksi->ksni_ipaddr = ifaces[0].li_ipaddr;
2615                 ksi->ksni_netmask = ifaces[0].li_netmask;
2616                 strlcpy(ksi->ksni_name, ifaces[0].li_name,
2617                         sizeof(ksi->ksni_name));
2618         } else {
2619                 /* Before Multi-Rail ksocklnd would manage
2620                  * multiple interfaces with its own tcp bonding.
2621                  * If we encounter an old configuration using
2622                  * this tcp bonding approach then we need to
2623                  * handle more than one ni_interfaces.
2624                  *
2625                  * In Multi-Rail configuration only ONE ni_interface
2626                  * should exist. Each IP alias should be mapped to
2627                  * each 'struct net_ni'.
2628                  */
2629                 for (i = 0; i < LNET_INTERFACES_NUM; i++) {
2630                         int j;
2631
2632                         if (!ni->ni_interfaces[i])
2633                                 break;
2634
2635                         for (j = 0; j < LNET_INTERFACES_NUM;  j++) {
2636                                 if (i != j && ni->ni_interfaces[j] &&
2637                                     strcmp(ni->ni_interfaces[i],
2638                                            ni->ni_interfaces[j]) == 0) {
2639                                         rc = -EEXIST;
2640                                         CERROR("ksocklnd: found duplicate %s at %d and %d, rc = %d\n",
2641                                                ni->ni_interfaces[i], i, j, rc);
2642                                         goto fail_1;
2643                                 }
2644                         }
2645
2646                         for (j = 0; j < rc; j++) {
2647                                 if (strcmp(ifaces[j].li_name,
2648                                            ni->ni_interfaces[i]) != 0)
2649                                         continue;
2650
2651                                 ksi = &net->ksnn_interfaces[j];
2652                                 ni->ni_dev_cpt = ifaces[j].li_cpt;
2653                                 ksi->ksni_ipaddr = ifaces[j].li_ipaddr;
2654                                 ksi->ksni_netmask = ifaces[j].li_netmask;
2655                                 strlcpy(ksi->ksni_name, ifaces[j].li_name,
2656                                         sizeof(ksi->ksni_name));
2657                                 net->ksnn_ninterfaces++;
2658                                 break;
2659                         }
2660                 }
2661                 /* ni_interfaces don't map to all network interfaces */
2662                 if (!ksi || net->ksnn_ninterfaces != i) {
2663                         CERROR("ksocklnd: requested %d but only %d interfaces found\n",
2664                                i, net->ksnn_ninterfaces);
2665                         goto fail_1;
2666                 }
2667         }
2668
2669         /* call it before add it to ksocknal_data.ksnd_nets */
2670         rc = ksocknal_net_start_threads(net, ni->ni_cpts, ni->ni_ncpts);
2671         if (rc != 0)
2672                 goto fail_1;
2673
2674         LASSERT(ksi);
2675         ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ksi->ksni_ipaddr);
2676         list_add(&net->ksnn_list, &ksocknal_data.ksnd_nets);
2677
2678         ksocknal_data.ksnd_nnets++;
2679
2680         return 0;
2681
2682  fail_1:
2683         LIBCFS_FREE(net, sizeof(*net));
2684  fail_0:
2685         if (ksocknal_data.ksnd_nnets == 0)
2686                 ksocknal_base_shutdown();
2687
2688         return -ENETDOWN;
2689 }
2690
2691
2692 static void __exit ksocklnd_exit(void)
2693 {
2694         lnet_unregister_lnd(&the_ksocklnd);
2695 }
2696
2697 static const struct lnet_lnd the_ksocklnd = {
2698         .lnd_type               = SOCKLND,
2699         .lnd_startup            = ksocknal_startup,
2700         .lnd_shutdown           = ksocknal_shutdown,
2701         .lnd_ctl                = ksocknal_ctl,
2702         .lnd_send               = ksocknal_send,
2703         .lnd_recv               = ksocknal_recv,
2704         .lnd_notify_peer_down   = ksocknal_notify_gw_down,
2705         .lnd_accept             = ksocknal_accept,
2706 };
2707
2708 static int __init ksocklnd_init(void)
2709 {
2710         int rc;
2711
2712         /* check ksnr_connected/connecting field large enough */
2713         BUILD_BUG_ON(SOCKLND_CONN_NTYPES > 4);
2714         BUILD_BUG_ON(SOCKLND_CONN_ACK != SOCKLND_CONN_BULK_IN);
2715
2716         rc = ksocknal_tunables_init();
2717         if (rc != 0)
2718                 return rc;
2719
2720         lnet_register_lnd(&the_ksocklnd);
2721
2722         return 0;
2723 }
2724
2725 MODULE_AUTHOR("OpenSFS, Inc. <http://www.lustre.org/>");
2726 MODULE_DESCRIPTION("TCP Socket LNet Network Driver");
2727 MODULE_VERSION("2.8.0");
2728 MODULE_LICENSE("GPL");
2729
2730 module_init(ksocklnd_init);
2731 module_exit(ksocklnd_exit);