fs/dlm/midcomms.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /******************************************************************************
   3 *******************************************************************************
   4 **
   5 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
   6 **  Copyright (C) 2004-2021 Red Hat, Inc.  All rights reserved.
   7 **
   8 **
   9 *******************************************************************************
  10 ******************************************************************************/
  11
  12 /*
  13  * midcomms.c
  14  *
  15  * This is the appallingly named "mid-level" comms layer. It takes care about
  16  * deliver an on application layer "reliable" communication above the used
  17  * lowcomms transport layer.
  18  *
  19  * How it works:
  20  *
  21  * Each nodes keeps track of all send DLM messages in send_queue with a sequence
  22  * number. The receive will send an DLM_ACK message back for every DLM message
  23  * received at the other side. If a reconnect happens in lowcomms we will send
  24  * all unacknowledged dlm messages again. The receiving side might drop any already
  25  * received message by comparing sequence numbers.
  26  *
  27  * How version detection works:
  28  *
  29  * Due the fact that dlm has pre-configured node addresses on every side
  30  * it is in it's nature that every side connects at starts to transmit
  31  * dlm messages which ends in a race. However DLM_RCOM_NAMES, DLM_RCOM_STATUS
  32  * and their replies are the first messages which are exchanges. Due backwards
  33  * compatibility these messages are not covered by the midcomms re-transmission
  34  * layer. These messages have their own re-transmission handling in the dlm
  35  * application layer. The version field of every node will be set on these RCOM
  36  * messages as soon as they arrived and the node isn't yet part of the nodes
  37  * hash. There exists also logic to detect version mismatched if something weird
  38  * going on or the first messages isn't an expected one.
  39  *
  40  * Termination:
  41  *
  42  * The midcomms layer does a 4 way handshake for termination on DLM protocol
  43  * like TCP supports it with half-closed socket support. SCTP doesn't support
  44  * half-closed socket, so we do it on DLM layer. Also socket shutdown() can be
  45  * interrupted by .e.g. tcp reset itself. Additional there exists the othercon
  46  * paradigm in lowcomms which cannot be easily without breaking backwards
  47  * compatibility. A node cannot send anything to another node when a DLM_FIN
  48  * message was send. There exists additional logic to print a warning if
  49  * DLM wants to do it. There exists a state handling like RFC 793 but reduced
  50  * to termination only. The event "member removal event" describes the cluster
  51  * manager removed the node from internal lists, at this point DLM does not
  52  * send any message to the other node. There exists two cases:
  53  *
  54  * 1. The cluster member was removed and we received a FIN
  55  * OR
  56  * 2. We received a FIN but the member was not removed yet
  57  *
  58  * One of these cases will do the CLOSE_WAIT to LAST_ACK change.
  59  *
  60  *
  61  *                              +---------+
  62  *                              | CLOSED  |
  63  *                              +---------+
  64  *                                   | add member/receive RCOM version
  65  *                                   |            detection msg
  66  *                                   V
  67  *                              +---------+
  68  *                              |  ESTAB  |
  69  *                              +---------+
  70  *                       CLOSE    |     |    rcv FIN
  71  *                      -------   |     |    -------
  72  * +---------+          snd FIN  /       \   snd ACK          +---------+
  73  * |  FIN    |<-----------------           ------------------>|  CLOSE  |
  74  * | WAIT-1  |------------------                              |   WAIT  |
  75  * +---------+          rcv FIN  \                            +---------+
  76  * | rcv ACK of FIN   -------   |                            CLOSE  | member
  77  * | --------------   snd ACK   |                           ------- | removal
  78  * V        x                   V                           snd FIN V event
  79  * +---------+                  +---------+                   +---------+
  80  * |FINWAIT-2|                  | CLOSING |                   | LAST-ACK|
  81  * +---------+                  +---------+                   +---------+
  82  * |                rcv ACK of FIN |                 rcv ACK of FIN |
  83  * |  rcv FIN       -------------- |                 -------------- |
  84  * |  -------              x       V                        x       V
  85  *  \ snd ACK                 +---------+                   +---------+
  86  *   ------------------------>| CLOSED  |                   | CLOSED  |
  87  *                            +---------+                   +---------+
  88  *
  89  * NOTE: any state can interrupted by midcomms_close() and state will be
  90  * switched to CLOSED in case of fencing. There exists also some timeout
  91  * handling when we receive the version detection RCOM messages which is
  92  * made by observation.
  93  *
  94  * Future improvements:
  95  *
  96  * There exists some known issues/improvements of the dlm handling. Some
  97  * of them should be done in a next major dlm version bump which makes
  98  * it incompatible with previous versions.
  99  *
 100  * Unaligned memory access:
 101  *
 102  * There exists cases when the dlm message buffer length is not aligned
 103  * to 8 byte. However seems nobody detected any problem with it. This
 104  * can be fixed in the next major version bump of dlm.
 105  *
 106  * Version detection:
 107  *
 108  * The version detection and how it's done is related to backwards
 109  * compatibility. There exists better ways to make a better handling.
 110  * However this should be changed in the next major version bump of dlm.
 111  *
 112  * Tail Size checking:
 113  *
 114  * There exists a message tail payload in e.g. DLM_MSG however we don't
 115  * check it against the message length yet regarding to the receive buffer
 116  * length. That need to be validated.
 117  *
 118  * Fencing bad nodes:
 119  *
 120  * At timeout places or weird sequence number behaviours we should send
 121  * a fencing request to the cluster manager.
 122  */
 123
 124 /* Debug switch to enable a 5 seconds sleep waiting of a termination.
 125  * This can be useful to test fencing while termination is running.
 126  * This requires a setup with only gfs2 as dlm user, so that the
 127  * last umount will terminate the connection.
 128  *
 129  * However it became useful to test, while the 5 seconds block in umount
 130  * just press the reset button. In a lot of dropping the termination
 131  * process can could take several seconds.
 132  */
 133 #define DLM_DEBUG_FENCE_TERMINATION     0
 134
 135 #include <net/tcp.h>
 136
 137 #include "dlm_internal.h"
 138 #include "lockspace.h"
 139 #include "lowcomms.h"
 140 #include "config.h"
 141 #include "memory.h"
 142 #include "lock.h"
 143 #include "util.h"
 144 #include "midcomms.h"
 145
 146 /* init value for sequence numbers for testing purpose only e.g. overflows */
 147 #define DLM_SEQ_INIT            0
 148 /* 3 minutes wait to sync ending of dlm */
 149 #define DLM_SHUTDOWN_TIMEOUT    msecs_to_jiffies(3 * 60 * 1000)
 150 #define DLM_VERSION_NOT_SET     0
 151
 152 struct midcomms_node {
 153         int nodeid;
 154         uint32_t version;
 155         uint32_t seq_send;
 156         uint32_t seq_next;
 157         /* These queues are unbound because we cannot drop any message in dlm.
 158          * We could send a fence signal for a specific node to the cluster
 159          * manager if queues hits some maximum value, however this handling
 160          * not supported yet.
 161          */
 162         struct list_head send_queue;
 163         spinlock_t send_queue_lock;
 164         atomic_t send_queue_cnt;
 165 #define DLM_NODE_FLAG_CLOSE     1
 166 #define DLM_NODE_FLAG_STOP_TX   2
 167 #define DLM_NODE_FLAG_STOP_RX   3
 168 #define DLM_NODE_ULP_DELIVERED  4
 169         unsigned long flags;
 170         wait_queue_head_t shutdown_wait;
 171
 172         /* dlm tcp termination state */
 173 #define DLM_CLOSED      1
 174 #define DLM_ESTABLISHED 2
 175 #define DLM_FIN_WAIT1   3
 176 #define DLM_FIN_WAIT2   4
 177 #define DLM_CLOSE_WAIT  5
 178 #define DLM_LAST_ACK    6
 179 #define DLM_CLOSING     7
 180         int state;
 181         spinlock_t state_lock;
 182
 183         /* counts how many lockspaces are using this node
 184          * this refcount is necessary to determine if the
 185          * node wants to disconnect.
 186          */
 187         int users;
 188
 189         /* not protected by srcu, node_hash lifetime */
 190         void *debugfs;
 191
 192         struct hlist_node hlist;
 193         struct rcu_head rcu;
 194 };
 195
 196 struct dlm_mhandle {
 197         const struct dlm_header *inner_hd;
 198         struct midcomms_node *node;
 199         struct dlm_opts *opts;
 200         struct dlm_msg *msg;
 201         bool committed;
 202         uint32_t seq;
 203
 204         void (*ack_rcv)(struct midcomms_node *node);
 205
 206         /* get_mhandle/commit srcu idx exchange */
 207         int idx;
 208
 209         struct list_head list;
 210         struct rcu_head rcu;
 211 };
 212
 213 static struct hlist_head node_hash[CONN_HASH_SIZE];
 214 static DEFINE_SPINLOCK(nodes_lock);
 215 DEFINE_STATIC_SRCU(nodes_srcu);
 216
 217 /* This mutex prevents that midcomms_close() is running while
 218  * stop() or remove(). As I experienced invalid memory access
 219  * behaviours when DLM_DEBUG_FENCE_TERMINATION is enabled and
 220  * resetting machines. I will end in some double deletion in nodes
 221  * datastructure.
 222  */
 223 static DEFINE_MUTEX(close_lock);
 224
 225 struct kmem_cache *dlm_midcomms_cache_create(void)
 226 {
 227         return kmem_cache_create("dlm_mhandle", sizeof(struct dlm_mhandle),
 228                                  0, 0, NULL);
 229 }
 230
 231 static inline const char *dlm_state_str(int state)
 232 {
 233         switch (state) {
 234         case DLM_CLOSED:
 235                 return "CLOSED";
 236         case DLM_ESTABLISHED:
 237                 return "ESTABLISHED";
 238         case DLM_FIN_WAIT1:
 239                 return "FIN_WAIT1";
 240         case DLM_FIN_WAIT2:
 241                 return "FIN_WAIT2";
 242         case DLM_CLOSE_WAIT:
 243                 return "CLOSE_WAIT";
 244         case DLM_LAST_ACK:
 245                 return "LAST_ACK";
 246         case DLM_CLOSING:
 247                 return "CLOSING";
 248         default:
 249                 return "UNKNOWN";
 250         }
 251 }
 252
 253 const char *dlm_midcomms_state(struct midcomms_node *node)
 254 {
 255         return dlm_state_str(node->state);
 256 }
 257
 258 unsigned long dlm_midcomms_flags(struct midcomms_node *node)
 259 {
 260         return node->flags;
 261 }
 262
 263 int dlm_midcomms_send_queue_cnt(struct midcomms_node *node)
 264 {
 265         return atomic_read(&node->send_queue_cnt);
 266 }
 267
 268 uint32_t dlm_midcomms_version(struct midcomms_node *node)
 269 {
 270         return node->version;
 271 }
 272
 273 static struct midcomms_node *__find_node(int nodeid, int r)
 274 {
 275         struct midcomms_node *node;
 276
 277         hlist_for_each_entry_rcu(node, &node_hash[r], hlist) {
 278                 if (node->nodeid == nodeid)
 279                         return node;
 280         }
 281
 282         return NULL;
 283 }
 284
 285 static void dlm_mhandle_release(struct rcu_head *rcu)
 286 {
 287         struct dlm_mhandle *mh = container_of(rcu, struct dlm_mhandle, rcu);
 288
 289         dlm_lowcomms_put_msg(mh->msg);
 290         dlm_free_mhandle(mh);
 291 }
 292
 293 static void dlm_mhandle_delete(struct midcomms_node *node,
 294                                struct dlm_mhandle *mh)
 295 {
 296         list_del_rcu(&mh->list);
 297         atomic_dec(&node->send_queue_cnt);
 298         call_rcu(&mh->rcu, dlm_mhandle_release);
 299 }
 300
 301 static void dlm_send_queue_flush(struct midcomms_node *node)
 302 {
 303         struct dlm_mhandle *mh;
 304
 305         pr_debug("flush midcomms send queue of node %d\n", node->nodeid);
 306
 307         rcu_read_lock();
 308         spin_lock(&node->send_queue_lock);
 309         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 310                 dlm_mhandle_delete(node, mh);
 311         }
 312         spin_unlock(&node->send_queue_lock);
 313         rcu_read_unlock();
 314 }
 315
 316 static void midcomms_node_reset(struct midcomms_node *node)
 317 {
 318         pr_debug("reset node %d\n", node->nodeid);
 319
 320         node->seq_next = DLM_SEQ_INIT;
 321         node->seq_send = DLM_SEQ_INIT;
 322         node->version = DLM_VERSION_NOT_SET;
 323         node->flags = 0;
 324
 325         dlm_send_queue_flush(node);
 326         node->state = DLM_CLOSED;
 327         wake_up(&node->shutdown_wait);
 328 }
 329
 330 static struct midcomms_node *nodeid2node(int nodeid, gfp_t alloc)
 331 {
 332         struct midcomms_node *node, *tmp;
 333         int r = nodeid_hash(nodeid);
 334
 335         node = __find_node(nodeid, r);
 336         if (node || !alloc)
 337                 return node;
 338
 339         node = kmalloc(sizeof(*node), alloc);
 340         if (!node)
 341                 return NULL;
 342
 343         node->nodeid = nodeid;
 344         spin_lock_init(&node->state_lock);
 345         spin_lock_init(&node->send_queue_lock);
 346         atomic_set(&node->send_queue_cnt, 0);
 347         INIT_LIST_HEAD(&node->send_queue);
 348         init_waitqueue_head(&node->shutdown_wait);
 349         node->users = 0;
 350         midcomms_node_reset(node);
 351
 352         spin_lock(&nodes_lock);
 353         /* check again if there was somebody else
 354          * earlier here to add the node
 355          */
 356         tmp = __find_node(nodeid, r);
 357         if (tmp) {
 358                 spin_unlock(&nodes_lock);
 359                 kfree(node);
 360                 return tmp;
 361         }
 362
 363         hlist_add_head_rcu(&node->hlist, &node_hash[r]);
 364         spin_unlock(&nodes_lock);
 365
 366         node->debugfs = dlm_create_debug_comms_file(nodeid, node);
 367         return node;
 368 }
 369
 370 static int dlm_send_ack(int nodeid, uint32_t seq)
 371 {
 372         int mb_len = sizeof(struct dlm_header);
 373         struct dlm_header *m_header;
 374         struct dlm_msg *msg;
 375         char *ppc;
 376
 377         msg = dlm_lowcomms_new_msg(nodeid, mb_len, GFP_NOFS, &ppc,
 378                                    NULL, NULL);
 379         if (!msg)
 380                 return -ENOMEM;
 381
 382         m_header = (struct dlm_header *)ppc;
 383
 384         m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
 385         m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
 386         m_header->h_length = cpu_to_le16(mb_len);
 387         m_header->h_cmd = DLM_ACK;
 388         m_header->u.h_seq = cpu_to_le32(seq);
 389
 390         dlm_lowcomms_commit_msg(msg);
 391         dlm_lowcomms_put_msg(msg);
 392
 393         return 0;
 394 }
 395
 396 static int dlm_send_fin(struct midcomms_node *node,
 397                         void (*ack_rcv)(struct midcomms_node *node))
 398 {
 399         int mb_len = sizeof(struct dlm_header);
 400         struct dlm_header *m_header;
 401         struct dlm_mhandle *mh;
 402         char *ppc;
 403
 404         mh = dlm_midcomms_get_mhandle(node->nodeid, mb_len, GFP_NOFS, &ppc);
 405         if (!mh)
 406                 return -ENOMEM;
 407
 408         mh->ack_rcv = ack_rcv;
 409
 410         m_header = (struct dlm_header *)ppc;
 411
 412         m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
 413         m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
 414         m_header->h_length = cpu_to_le16(mb_len);
 415         m_header->h_cmd = DLM_FIN;
 416
 417         pr_debug("sending fin msg to node %d\n", node->nodeid);
 418         dlm_midcomms_commit_mhandle(mh);
 419         set_bit(DLM_NODE_FLAG_STOP_TX, &node->flags);
 420
 421         return 0;
 422 }
 423
 424 static void dlm_receive_ack(struct midcomms_node *node, uint32_t seq)
 425 {
 426         struct dlm_mhandle *mh;
 427
 428         rcu_read_lock();
 429         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 430                 if (before(mh->seq, seq)) {
 431                         if (mh->ack_rcv)
 432                                 mh->ack_rcv(node);
 433                 } else {
 434                         /* send queue should be ordered */
 435                         break;
 436                 }
 437         }
 438
 439         spin_lock(&node->send_queue_lock);
 440         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 441                 if (before(mh->seq, seq)) {
 442                         dlm_mhandle_delete(node, mh);
 443                 } else {
 444                         /* send queue should be ordered */
 445                         break;
 446                 }
 447         }
 448         spin_unlock(&node->send_queue_lock);
 449         rcu_read_unlock();
 450 }
 451
 452 static void dlm_pas_fin_ack_rcv(struct midcomms_node *node)
 453 {
 454         spin_lock(&node->state_lock);
 455         pr_debug("receive passive fin ack from node %d with state %s\n",
 456                  node->nodeid, dlm_state_str(node->state));
 457
 458         switch (node->state) {
 459         case DLM_LAST_ACK:
 460                 /* DLM_CLOSED */
 461                 midcomms_node_reset(node);
 462                 break;
 463         case DLM_CLOSED:
 464                 /* not valid but somehow we got what we want */
 465                 wake_up(&node->shutdown_wait);
 466                 break;
 467         default:
 468                 spin_unlock(&node->state_lock);
 469                 log_print("%s: unexpected state: %d\n",
 470                           __func__, node->state);
 471                 WARN_ON(1);
 472                 return;
 473         }
 474         spin_unlock(&node->state_lock);
 475 }
 476
 477 static void dlm_midcomms_receive_buffer(union dlm_packet *p,
 478                                         struct midcomms_node *node,
 479                                         uint32_t seq)
 480 {
 481         if (seq == node->seq_next) {
 482                 node->seq_next++;
 483
 484                 switch (p->header.h_cmd) {
 485                 case DLM_FIN:
 486                         /* send ack before fin */
 487                         dlm_send_ack(node->nodeid, node->seq_next);
 488
 489                         spin_lock(&node->state_lock);
 490                         pr_debug("receive fin msg from node %d with state %s\n",
 491                                  node->nodeid, dlm_state_str(node->state));
 492
 493                         switch (node->state) {
 494                         case DLM_ESTABLISHED:
 495                                 node->state = DLM_CLOSE_WAIT;
 496                                 pr_debug("switch node %d to state %s\n",
 497                                          node->nodeid, dlm_state_str(node->state));
 498                                 /* passive shutdown DLM_LAST_ACK case 1
 499                                  * additional we check if the node is used by
 500                                  * cluster manager events at all.
 501                                  */
 502                                 if (node->users == 0) {
 503                                         node->state = DLM_LAST_ACK;
 504                                         pr_debug("switch node %d to state %s case 1\n",
 505                                                  node->nodeid, dlm_state_str(node->state));
 506                                         spin_unlock(&node->state_lock);
 507                                         goto send_fin;
 508                                 }
 509                                 break;
 510                         case DLM_FIN_WAIT1:
 511                                 node->state = DLM_CLOSING;
 512                                 pr_debug("switch node %d to state %s\n",
 513                                          node->nodeid, dlm_state_str(node->state));
 514                                 break;
 515                         case DLM_FIN_WAIT2:
 516                                 midcomms_node_reset(node);
 517                                 pr_debug("switch node %d to state %s\n",
 518                                          node->nodeid, dlm_state_str(node->state));
 519                                 wake_up(&node->shutdown_wait);
 520                                 break;
 521                         case DLM_LAST_ACK:
 522                                 /* probably remove_member caught it, do nothing */
 523                                 break;
 524                         default:
 525                                 spin_unlock(&node->state_lock);
 526                                 log_print("%s: unexpected state: %d\n",
 527                                           __func__, node->state);
 528                                 WARN_ON(1);
 529                                 return;
 530                         }
 531                         spin_unlock(&node->state_lock);
 532
 533                         set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
 534                         break;
 535                 default:
 536                         WARN_ON(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
 537                         dlm_receive_buffer(p, node->nodeid);
 538                         set_bit(DLM_NODE_ULP_DELIVERED, &node->flags);
 539                         break;
 540                 }
 541         } else {
 542                 /* retry to ack message which we already have by sending back
 543                  * current node->seq_next number as ack.
 544                  */
 545                 if (seq < node->seq_next)
 546                         dlm_send_ack(node->nodeid, node->seq_next);
 547
 548                 log_print_ratelimited("ignore dlm msg because seq mismatch, seq: %u, expected: %u, nodeid: %d",
 549                                       seq, node->seq_next, node->nodeid);
 550         }
 551
 552         return;
 553
 554 send_fin:
 555         set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
 556         dlm_send_fin(node, dlm_pas_fin_ack_rcv);
 557 }
 558
 559 static struct midcomms_node *
 560 dlm_midcomms_recv_node_lookup(int nodeid, const union dlm_packet *p,
 561                               uint16_t msglen, int (*cb)(struct midcomms_node *node))
 562 {
 563         struct midcomms_node *node = NULL;
 564         gfp_t allocation = 0;
 565         int ret;
 566
 567         switch (p->header.h_cmd) {
 568         case DLM_RCOM:
 569                 if (msglen < sizeof(struct dlm_rcom)) {
 570                         log_print("rcom msg too small: %u, will skip this message from node %d",
 571                                   msglen, nodeid);
 572                         return NULL;
 573                 }
 574
 575                 switch (p->rcom.rc_type) {
 576                 case cpu_to_le32(DLM_RCOM_NAMES):
 577                         fallthrough;
 578                 case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
 579                         fallthrough;
 580                 case cpu_to_le32(DLM_RCOM_STATUS):
 581                         fallthrough;
 582                 case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
 583                         node = nodeid2node(nodeid, 0);
 584                         if (node) {
 585                                 spin_lock(&node->state_lock);
 586                                 if (node->state != DLM_ESTABLISHED)
 587                                         pr_debug("receive begin RCOM msg from node %d with state %s\n",
 588                                                  node->nodeid, dlm_state_str(node->state));
 589
 590                                 switch (node->state) {
 591                                 case DLM_CLOSED:
 592                                         node->state = DLM_ESTABLISHED;
 593                                         pr_debug("switch node %d to state %s\n",
 594                                                  node->nodeid, dlm_state_str(node->state));
 595                                         break;
 596                                 case DLM_ESTABLISHED:
 597                                         break;
 598                                 default:
 599                                         /* some invalid state passive shutdown
 600                                          * was failed, we try to reset and
 601                                          * hope it will go on.
 602                                          */
 603                                         log_print("reset node %d because shutdown stuck",
 604                                                   node->nodeid);
 605
 606                                         midcomms_node_reset(node);
 607                                         node->state = DLM_ESTABLISHED;
 608                                         break;
 609                                 }
 610                                 spin_unlock(&node->state_lock);
 611                         }
 612
 613                         allocation = GFP_NOFS;
 614                         break;
 615                 default:
 616                         break;
 617                 }
 618
 619                 break;
 620         default:
 621                 break;
 622         }
 623
 624         node = nodeid2node(nodeid, allocation);
 625         if (!node) {
 626                 switch (p->header.h_cmd) {
 627                 case DLM_OPTS:
 628                         if (msglen < sizeof(struct dlm_opts)) {
 629                                 log_print("opts msg too small: %u, will skip this message from node %d",
 630                                           msglen, nodeid);
 631                                 return NULL;
 632                         }
 633
 634                         log_print_ratelimited("received dlm opts message nextcmd %d from node %d in an invalid sequence",
 635                                               p->opts.o_nextcmd, nodeid);
 636                         break;
 637                 default:
 638                         log_print_ratelimited("received dlm message cmd %d from node %d in an invalid sequence",
 639                                               p->header.h_cmd, nodeid);
 640                         break;
 641                 }
 642
 643                 return NULL;
 644         }
 645
 646         ret = cb(node);
 647         if (ret < 0)
 648                 return NULL;
 649
 650         return node;
 651 }
 652
 653 static int dlm_midcomms_version_check_3_2(struct midcomms_node *node)
 654 {
 655         switch (node->version) {
 656         case DLM_VERSION_NOT_SET:
 657                 node->version = DLM_VERSION_3_2;
 658                 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_2,
 659                           node->nodeid);
 660                 break;
 661         case DLM_VERSION_3_2:
 662                 break;
 663         default:
 664                 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
 665                                       DLM_VERSION_3_2, node->nodeid, node->version);
 666                 return -1;
 667         }
 668
 669         return 0;
 670 }
 671
 672 static int dlm_opts_check_msglen(union dlm_packet *p, uint16_t msglen, int nodeid)
 673 {
 674         int len = msglen;
 675
 676         /* we only trust outer header msglen because
 677          * it's checked against receive buffer length.
 678          */
 679         if (len < sizeof(struct dlm_opts))
 680                 return -1;
 681         len -= sizeof(struct dlm_opts);
 682
 683         if (len < le16_to_cpu(p->opts.o_optlen))
 684                 return -1;
 685         len -= le16_to_cpu(p->opts.o_optlen);
 686
 687         switch (p->opts.o_nextcmd) {
 688         case DLM_FIN:
 689                 if (len < sizeof(struct dlm_header)) {
 690                         log_print("fin too small: %d, will skip this message from node %d",
 691                                   len, nodeid);
 692                         return -1;
 693                 }
 694
 695                 break;
 696         case DLM_MSG:
 697                 if (len < sizeof(struct dlm_message)) {
 698                         log_print("msg too small: %d, will skip this message from node %d",
 699                                   msglen, nodeid);
 700                         return -1;
 701                 }
 702
 703                 break;
 704         case DLM_RCOM:
 705                 if (len < sizeof(struct dlm_rcom)) {
 706                         log_print("rcom msg too small: %d, will skip this message from node %d",
 707                                   len, nodeid);
 708                         return -1;
 709                 }
 710
 711                 break;
 712         default:
 713                 log_print("unsupported o_nextcmd received: %u, will skip this message from node %d",
 714                           p->opts.o_nextcmd, nodeid);
 715                 return -1;
 716         }
 717
 718         return 0;
 719 }
 720
 721 static void dlm_midcomms_receive_buffer_3_2(union dlm_packet *p, int nodeid)
 722 {
 723         uint16_t msglen = le16_to_cpu(p->header.h_length);
 724         struct midcomms_node *node;
 725         uint32_t seq;
 726         int ret, idx;
 727
 728         idx = srcu_read_lock(&nodes_srcu);
 729         node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
 730                                              dlm_midcomms_version_check_3_2);
 731         if (!node)
 732                 goto out;
 733
 734         switch (p->header.h_cmd) {
 735         case DLM_RCOM:
 736                 /* these rcom message we use to determine version.
 737                  * they have their own retransmission handling and
 738                  * are the first messages of dlm.
 739                  *
 740                  * length already checked.
 741                  */
 742                 switch (p->rcom.rc_type) {
 743                 case cpu_to_le32(DLM_RCOM_NAMES):
 744                         fallthrough;
 745                 case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
 746                         fallthrough;
 747                 case cpu_to_le32(DLM_RCOM_STATUS):
 748                         fallthrough;
 749                 case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
 750                         break;
 751                 default:
 752                         log_print("unsupported rcom type received: %u, will skip this message from node %d",
 753                                   le32_to_cpu(p->rcom.rc_type), nodeid);
 754                         goto out;
 755                 }
 756
 757                 WARN_ON(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
 758                 dlm_receive_buffer(p, nodeid);
 759                 break;
 760         case DLM_OPTS:
 761                 seq = le32_to_cpu(p->header.u.h_seq);
 762
 763                 ret = dlm_opts_check_msglen(p, msglen, nodeid);
 764                 if (ret < 0) {
 765                         log_print("opts msg too small: %u, will skip this message from node %d",
 766                                   msglen, nodeid);
 767                         goto out;
 768                 }
 769
 770                 p = (union dlm_packet *)((unsigned char *)p->opts.o_opts +
 771                                          le16_to_cpu(p->opts.o_optlen));
 772
 773                 /* recheck inner msglen just if it's not garbage */
 774                 msglen = le16_to_cpu(p->header.h_length);
 775                 switch (p->header.h_cmd) {
 776                 case DLM_RCOM:
 777                         if (msglen < sizeof(struct dlm_rcom)) {
 778                                 log_print("inner rcom msg too small: %u, will skip this message from node %d",
 779                                           msglen, nodeid);
 780                                 goto out;
 781                         }
 782
 783                         break;
 784                 case DLM_MSG:
 785                         if (msglen < sizeof(struct dlm_message)) {
 786                                 log_print("inner msg too small: %u, will skip this message from node %d",
 787                                           msglen, nodeid);
 788                                 goto out;
 789                         }
 790
 791                         break;
 792                 case DLM_FIN:
 793                         if (msglen < sizeof(struct dlm_header)) {
 794                                 log_print("inner fin too small: %u, will skip this message from node %d",
 795                                           msglen, nodeid);
 796                                 goto out;
 797                         }
 798
 799                         break;
 800                 default:
 801                         log_print("unsupported inner h_cmd received: %u, will skip this message from node %d",
 802                                   msglen, nodeid);
 803                         goto out;
 804                 }
 805
 806                 dlm_midcomms_receive_buffer(p, node, seq);
 807                 break;
 808         case DLM_ACK:
 809                 seq = le32_to_cpu(p->header.u.h_seq);
 810                 dlm_receive_ack(node, seq);
 811                 break;
 812         default:
 813                 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
 814                           p->header.h_cmd, nodeid);
 815                 break;
 816         }
 817
 818 out:
 819         srcu_read_unlock(&nodes_srcu, idx);
 820 }
 821
 822 static int dlm_midcomms_version_check_3_1(struct midcomms_node *node)
 823 {
 824         switch (node->version) {
 825         case DLM_VERSION_NOT_SET:
 826                 node->version = DLM_VERSION_3_1;
 827                 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_1,
 828                           node->nodeid);
 829                 break;
 830         case DLM_VERSION_3_1:
 831                 break;
 832         default:
 833                 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
 834                                       DLM_VERSION_3_1, node->nodeid, node->version);
 835                 return -1;
 836         }
 837
 838         return 0;
 839 }
 840
 841 static void dlm_midcomms_receive_buffer_3_1(union dlm_packet *p, int nodeid)
 842 {
 843         uint16_t msglen = le16_to_cpu(p->header.h_length);
 844         struct midcomms_node *node;
 845         int idx;
 846
 847         idx = srcu_read_lock(&nodes_srcu);
 848         node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
 849                                              dlm_midcomms_version_check_3_1);
 850         if (!node) {
 851                 srcu_read_unlock(&nodes_srcu, idx);
 852                 return;
 853         }
 854         srcu_read_unlock(&nodes_srcu, idx);
 855
 856         switch (p->header.h_cmd) {
 857         case DLM_RCOM:
 858                 /* length already checked */
 859                 break;
 860         case DLM_MSG:
 861                 if (msglen < sizeof(struct dlm_message)) {
 862                         log_print("msg too small: %u, will skip this message from node %d",
 863                                   msglen, nodeid);
 864                         return;
 865                 }
 866
 867                 break;
 868         default:
 869                 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
 870                           p->header.h_cmd, nodeid);
 871                 return;
 872         }
 873
 874         dlm_receive_buffer(p, nodeid);
 875 }
 876
 877 /*
 878  * Called from the low-level comms layer to process a buffer of
 879  * commands.
 880  */
 881
 882 int dlm_process_incoming_buffer(int nodeid, unsigned char *buf, int len)
 883 {
 884         const unsigned char *ptr = buf;
 885         const struct dlm_header *hd;
 886         uint16_t msglen;
 887         int ret = 0;
 888
 889         while (len >= sizeof(struct dlm_header)) {
 890                 hd = (struct dlm_header *)ptr;
 891
 892                 /* no message should be more than DLM_MAX_SOCKET_BUFSIZE or
 893                  * less than dlm_header size.
 894                  *
 895                  * Some messages does not have a 8 byte length boundary yet
 896                  * which can occur in a unaligned memory access of some dlm
 897                  * messages. However this problem need to be fixed at the
 898                  * sending side, for now it seems nobody run into architecture
 899                  * related issues yet but it slows down some processing.
 900                  * Fixing this issue should be scheduled in future by doing
 901                  * the next major version bump.
 902                  */
 903                 msglen = le16_to_cpu(hd->h_length);
 904                 if (msglen > DLM_MAX_SOCKET_BUFSIZE ||
 905                     msglen < sizeof(struct dlm_header)) {
 906                         log_print("received invalid length header: %u from node %d, will abort message parsing",
 907                                   msglen, nodeid);
 908                         return -EBADMSG;
 909                 }
 910
 911                 /* caller will take care that leftover
 912                  * will be parsed next call with more data
 913                  */
 914                 if (msglen > len)
 915                         break;
 916
 917                 switch (hd->h_version) {
 918                 case cpu_to_le32(DLM_VERSION_3_1):
 919                         dlm_midcomms_receive_buffer_3_1((union dlm_packet *)ptr, nodeid);
 920                         break;
 921                 case cpu_to_le32(DLM_VERSION_3_2):
 922                         dlm_midcomms_receive_buffer_3_2((union dlm_packet *)ptr, nodeid);
 923                         break;
 924                 default:
 925                         log_print("received invalid version header: %u from node %d, will skip this message",
 926                                   le32_to_cpu(hd->h_version), nodeid);
 927                         break;
 928                 }
 929
 930                 ret += msglen;
 931                 len -= msglen;
 932                 ptr += msglen;
 933         }
 934
 935         return ret;
 936 }
 937
 938 void dlm_midcomms_receive_done(int nodeid)
 939 {
 940         struct midcomms_node *node;
 941         int idx;
 942
 943         idx = srcu_read_lock(&nodes_srcu);
 944         node = nodeid2node(nodeid, 0);
 945         if (!node) {
 946                 srcu_read_unlock(&nodes_srcu, idx);
 947                 return;
 948         }
 949
 950         /* old protocol, we do nothing */
 951         switch (node->version) {
 952         case DLM_VERSION_3_2:
 953                 break;
 954         default:
 955                 srcu_read_unlock(&nodes_srcu, idx);
 956                 return;
 957         }
 958
 959         /* do nothing if we didn't delivered stateful to ulp */
 960         if (!test_and_clear_bit(DLM_NODE_ULP_DELIVERED,
 961                                 &node->flags)) {
 962                 srcu_read_unlock(&nodes_srcu, idx);
 963                 return;
 964         }
 965
 966         spin_lock(&node->state_lock);
 967         /* we only ack if state is ESTABLISHED */
 968         switch (node->state) {
 969         case DLM_ESTABLISHED:
 970                 spin_unlock(&node->state_lock);
 971                 dlm_send_ack(node->nodeid, node->seq_next);
 972                 break;
 973         default:
 974                 spin_unlock(&node->state_lock);
 975                 /* do nothing FIN has it's own ack send */
 976                 break;
 977         }
 978         srcu_read_unlock(&nodes_srcu, idx);
 979 }
 980
 981 void dlm_midcomms_unack_msg_resend(int nodeid)
 982 {
 983         struct midcomms_node *node;
 984         struct dlm_mhandle *mh;
 985         int idx, ret;
 986
 987         idx = srcu_read_lock(&nodes_srcu);
 988         node = nodeid2node(nodeid, 0);
 989         if (!node) {
 990                 srcu_read_unlock(&nodes_srcu, idx);
 991                 return;
 992         }
 993
 994         /* old protocol, we don't support to retransmit on failure */
 995         switch (node->version) {
 996         case DLM_VERSION_3_2:
 997                 break;
 998         default:
 999                 srcu_read_unlock(&nodes_srcu, idx);
1000                 return;
1001         }
1002
1003         rcu_read_lock();
1004         list_for_each_entry_rcu(mh, &node->send_queue, list) {
1005                 if (!mh->committed)
1006                         continue;
1007
1008                 ret = dlm_lowcomms_resend_msg(mh->msg);
1009                 if (!ret)
1010                         log_print_ratelimited("retransmit dlm msg, seq %u, nodeid %d",
1011                                               mh->seq, node->nodeid);
1012         }
1013         rcu_read_unlock();
1014         srcu_read_unlock(&nodes_srcu, idx);
1015 }
1016
1017 static void dlm_fill_opts_header(struct dlm_opts *opts, uint16_t inner_len,
1018                                  uint32_t seq)
1019 {
1020         opts->o_header.h_cmd = DLM_OPTS;
1021         opts->o_header.h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
1022         opts->o_header.h_nodeid = cpu_to_le32(dlm_our_nodeid());
1023         opts->o_header.h_length = cpu_to_le16(DLM_MIDCOMMS_OPT_LEN + inner_len);
1024         opts->o_header.u.h_seq = cpu_to_le32(seq);
1025 }
1026
1027 static void midcomms_new_msg_cb(void *data)
1028 {
1029         struct dlm_mhandle *mh = data;
1030
1031         atomic_inc(&mh->node->send_queue_cnt);
1032
1033         spin_lock(&mh->node->send_queue_lock);
1034         list_add_tail_rcu(&mh->list, &mh->node->send_queue);
1035         spin_unlock(&mh->node->send_queue_lock);
1036
1037         mh->seq = mh->node->seq_send++;
1038 }
1039
1040 static struct dlm_msg *dlm_midcomms_get_msg_3_2(struct dlm_mhandle *mh, int nodeid,
1041                                                 int len, gfp_t allocation, char **ppc)
1042 {
1043         struct dlm_opts *opts;
1044         struct dlm_msg *msg;
1045
1046         msg = dlm_lowcomms_new_msg(nodeid, len + DLM_MIDCOMMS_OPT_LEN,
1047                                    allocation, ppc, midcomms_new_msg_cb, mh);
1048         if (!msg)
1049                 return NULL;
1050
1051         opts = (struct dlm_opts *)*ppc;
1052         mh->opts = opts;
1053
1054         /* add possible options here */
1055         dlm_fill_opts_header(opts, len, mh->seq);
1056
1057         *ppc += sizeof(*opts);
1058         mh->inner_hd = (const struct dlm_header *)*ppc;
1059         return msg;
1060 }
1061
1062 /* avoid false positive for nodes_srcu, unlock happens in
1063  * dlm_midcomms_commit_mhandle which is a must call if success
1064  */
1065 #ifndef __CHECKER__
1066 struct dlm_mhandle *dlm_midcomms_get_mhandle(int nodeid, int len,
1067                                              gfp_t allocation, char **ppc)
1068 {
1069         struct midcomms_node *node;
1070         struct dlm_mhandle *mh;
1071         struct dlm_msg *msg;
1072         int idx;
1073
1074         idx = srcu_read_lock(&nodes_srcu);
1075         node = nodeid2node(nodeid, 0);
1076         if (!node) {
1077                 WARN_ON_ONCE(1);
1078                 goto err;
1079         }
1080
1081         /* this is a bug, however we going on and hope it will be resolved */
1082         WARN_ON(test_bit(DLM_NODE_FLAG_STOP_TX, &node->flags));
1083
1084         mh = dlm_allocate_mhandle();
1085         if (!mh)
1086                 goto err;
1087
1088         mh->committed = false;
1089         mh->ack_rcv = NULL;
1090         mh->idx = idx;
1091         mh->node = node;
1092
1093         switch (node->version) {
1094         case DLM_VERSION_3_1:
1095                 msg = dlm_lowcomms_new_msg(nodeid, len, allocation, ppc,
1096                                            NULL, NULL);
1097                 if (!msg) {
1098                         dlm_free_mhandle(mh);
1099                         goto err;
1100                 }
1101
1102                 break;
1103         case DLM_VERSION_3_2:
1104                 msg = dlm_midcomms_get_msg_3_2(mh, nodeid, len, allocation,
1105                                                ppc);
1106                 if (!msg) {
1107                         dlm_free_mhandle(mh);
1108                         goto err;
1109                 }
1110
1111                 break;
1112         default:
1113                 dlm_free_mhandle(mh);
1114                 WARN_ON(1);
1115                 goto err;
1116         }
1117
1118         mh->msg = msg;
1119
1120         /* keep in mind that is a must to call
1121          * dlm_midcomms_commit_msg() which releases
1122          * nodes_srcu using mh->idx which is assumed
1123          * here that the application will call it.
1124          */
1125         return mh;
1126
1127 err:
1128         srcu_read_unlock(&nodes_srcu, idx);
1129         return NULL;
1130 }
1131 #endif
1132
1133 static void dlm_midcomms_commit_msg_3_2(struct dlm_mhandle *mh)
1134 {
1135         /* nexthdr chain for fast lookup */
1136         mh->opts->o_nextcmd = mh->inner_hd->h_cmd;
1137         mh->committed = true;
1138         dlm_lowcomms_commit_msg(mh->msg);
1139 }
1140
1141 /* avoid false positive for nodes_srcu, lock was happen in
1142  * dlm_midcomms_get_mhandle
1143  */
1144 #ifndef __CHECKER__
1145 void dlm_midcomms_commit_mhandle(struct dlm_mhandle *mh)
1146 {
1147         switch (mh->node->version) {
1148         case DLM_VERSION_3_1:
1149                 srcu_read_unlock(&nodes_srcu, mh->idx);
1150
1151                 dlm_lowcomms_commit_msg(mh->msg);
1152                 dlm_lowcomms_put_msg(mh->msg);
1153                 /* mh is not part of rcu list in this case */
1154                 dlm_free_mhandle(mh);
1155                 break;
1156         case DLM_VERSION_3_2:
1157                 dlm_midcomms_commit_msg_3_2(mh);
1158                 srcu_read_unlock(&nodes_srcu, mh->idx);
1159                 break;
1160         default:
1161                 srcu_read_unlock(&nodes_srcu, mh->idx);
1162                 WARN_ON(1);
1163                 break;
1164         }
1165 }
1166 #endif
1167
1168 int dlm_midcomms_start(void)
1169 {
1170         int i;
1171
1172         for (i = 0; i < CONN_HASH_SIZE; i++)
1173                 INIT_HLIST_HEAD(&node_hash[i]);
1174
1175         return dlm_lowcomms_start();
1176 }
1177
1178 static void dlm_act_fin_ack_rcv(struct midcomms_node *node)
1179 {
1180         spin_lock(&node->state_lock);
1181         pr_debug("receive active fin ack from node %d with state %s\n",
1182                  node->nodeid, dlm_state_str(node->state));
1183
1184         switch (node->state) {
1185         case DLM_FIN_WAIT1:
1186                 node->state = DLM_FIN_WAIT2;
1187                 pr_debug("switch node %d to state %s\n",
1188                          node->nodeid, dlm_state_str(node->state));
1189                 break;
1190         case DLM_CLOSING:
1191                 midcomms_node_reset(node);
1192                 pr_debug("switch node %d to state %s\n",
1193                          node->nodeid, dlm_state_str(node->state));
1194                 wake_up(&node->shutdown_wait);
1195                 break;
1196         case DLM_CLOSED:
1197                 /* not valid but somehow we got what we want */
1198                 wake_up(&node->shutdown_wait);
1199                 break;
1200         default:
1201                 spin_unlock(&node->state_lock);
1202                 log_print("%s: unexpected state: %d\n",
1203                           __func__, node->state);
1204                 WARN_ON(1);
1205                 return;
1206         }
1207         spin_unlock(&node->state_lock);
1208 }
1209
1210 void dlm_midcomms_add_member(int nodeid)
1211 {
1212         struct midcomms_node *node;
1213         int idx;
1214
1215         if (nodeid == dlm_our_nodeid())
1216                 return;
1217
1218         idx = srcu_read_lock(&nodes_srcu);
1219         node = nodeid2node(nodeid, GFP_NOFS);
1220         if (!node) {
1221                 srcu_read_unlock(&nodes_srcu, idx);
1222                 return;
1223         }
1224
1225         spin_lock(&node->state_lock);
1226         if (!node->users) {
1227                 pr_debug("receive add member from node %d with state %s\n",
1228                          node->nodeid, dlm_state_str(node->state));
1229                 switch (node->state) {
1230                 case DLM_ESTABLISHED:
1231                         break;
1232                 case DLM_CLOSED:
1233                         node->state = DLM_ESTABLISHED;
1234                         pr_debug("switch node %d to state %s\n",
1235                                  node->nodeid, dlm_state_str(node->state));
1236                         break;
1237                 default:
1238                         /* some invalid state passive shutdown
1239                          * was failed, we try to reset and
1240                          * hope it will go on.
1241                          */
1242                         log_print("reset node %d because shutdown stuck",
1243                                   node->nodeid);
1244
1245                         midcomms_node_reset(node);
1246                         node->state = DLM_ESTABLISHED;
1247                         break;
1248                 }
1249         }
1250
1251         node->users++;
1252         pr_debug("node %d users inc count %d\n", nodeid, node->users);
1253         spin_unlock(&node->state_lock);
1254
1255         srcu_read_unlock(&nodes_srcu, idx);
1256 }
1257
1258 void dlm_midcomms_remove_member(int nodeid)
1259 {
1260         struct midcomms_node *node;
1261         int idx;
1262
1263         if (nodeid == dlm_our_nodeid())
1264                 return;
1265
1266         idx = srcu_read_lock(&nodes_srcu);
1267         node = nodeid2node(nodeid, 0);
1268         if (!node) {
1269                 srcu_read_unlock(&nodes_srcu, idx);
1270                 return;
1271         }
1272
1273         spin_lock(&node->state_lock);
1274         node->users--;
1275         pr_debug("node %d users dec count %d\n", nodeid, node->users);
1276
1277         /* hitting users count to zero means the
1278          * other side is running dlm_midcomms_stop()
1279          * we meet us to have a clean disconnect.
1280          */
1281         if (node->users == 0) {
1282                 pr_debug("receive remove member from node %d with state %s\n",
1283                          node->nodeid, dlm_state_str(node->state));
1284                 switch (node->state) {
1285                 case DLM_ESTABLISHED:
1286                         break;
1287                 case DLM_CLOSE_WAIT:
1288                         /* passive shutdown DLM_LAST_ACK case 2 */
1289                         node->state = DLM_LAST_ACK;
1290                         spin_unlock(&node->state_lock);
1291
1292                         pr_debug("switch node %d to state %s case 2\n",
1293                                  node->nodeid, dlm_state_str(node->state));
1294                         goto send_fin;
1295                 case DLM_LAST_ACK:
1296                         /* probably receive fin caught it, do nothing */
1297                         break;
1298                 case DLM_CLOSED:
1299                         /* already gone, do nothing */
1300                         break;
1301                 default:
1302                         log_print("%s: unexpected state: %d\n",
1303                                   __func__, node->state);
1304                         break;
1305                 }
1306         }
1307         spin_unlock(&node->state_lock);
1308
1309         srcu_read_unlock(&nodes_srcu, idx);
1310         return;
1311
1312 send_fin:
1313         set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
1314         dlm_send_fin(node, dlm_pas_fin_ack_rcv);
1315         srcu_read_unlock(&nodes_srcu, idx);
1316 }
1317
1318 static void midcomms_node_release(struct rcu_head *rcu)
1319 {
1320         struct midcomms_node *node = container_of(rcu, struct midcomms_node, rcu);
1321
1322         WARN_ON(atomic_read(&node->send_queue_cnt));
1323         kfree(node);
1324 }
1325
1326 static void midcomms_shutdown(struct midcomms_node *node)
1327 {
1328         int ret;
1329
1330         /* old protocol, we don't wait for pending operations */
1331         switch (node->version) {
1332         case DLM_VERSION_3_2:
1333                 break;
1334         default:
1335                 return;
1336         }
1337
1338         spin_lock(&node->state_lock);
1339         pr_debug("receive active shutdown for node %d with state %s\n",
1340                  node->nodeid, dlm_state_str(node->state));
1341         switch (node->state) {
1342         case DLM_ESTABLISHED:
1343                 node->state = DLM_FIN_WAIT1;
1344                 pr_debug("switch node %d to state %s case 2\n",
1345                          node->nodeid, dlm_state_str(node->state));
1346                 break;
1347         case DLM_CLOSED:
1348                 /* we have what we want */
1349                 spin_unlock(&node->state_lock);
1350                 return;
1351         default:
1352                 /* busy to enter DLM_FIN_WAIT1, wait until passive
1353                  * done in shutdown_wait to enter DLM_CLOSED.
1354                  */
1355                 break;
1356         }
1357         spin_unlock(&node->state_lock);
1358
1359         if (node->state == DLM_FIN_WAIT1) {
1360                 dlm_send_fin(node, dlm_act_fin_ack_rcv);
1361
1362                 if (DLM_DEBUG_FENCE_TERMINATION)
1363                         msleep(5000);
1364         }
1365
1366         /* wait for other side dlm + fin */
1367         ret = wait_event_timeout(node->shutdown_wait,
1368                                  node->state == DLM_CLOSED ||
1369                                  test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1370                                  DLM_SHUTDOWN_TIMEOUT);
1371         if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags)) {
1372                 pr_debug("active shutdown timed out for node %d with state %s\n",
1373                          node->nodeid, dlm_state_str(node->state));
1374                 midcomms_node_reset(node);
1375                 return;
1376         }
1377
1378         pr_debug("active shutdown done for node %d with state %s\n",
1379                  node->nodeid, dlm_state_str(node->state));
1380 }
1381
1382 void dlm_midcomms_shutdown(void)
1383 {
1384         struct midcomms_node *node;
1385         int i, idx;
1386
1387         mutex_lock(&close_lock);
1388         idx = srcu_read_lock(&nodes_srcu);
1389         for (i = 0; i < CONN_HASH_SIZE; i++) {
1390                 hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1391                         midcomms_shutdown(node);
1392
1393                         dlm_delete_debug_comms_file(node->debugfs);
1394
1395                         spin_lock(&nodes_lock);
1396                         hlist_del_rcu(&node->hlist);
1397                         spin_unlock(&nodes_lock);
1398
1399                         call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
1400                 }
1401         }
1402         srcu_read_unlock(&nodes_srcu, idx);
1403         mutex_unlock(&close_lock);
1404
1405         dlm_lowcomms_shutdown();
1406 }
1407
1408 int dlm_midcomms_close(int nodeid)
1409 {
1410         struct midcomms_node *node;
1411         int idx, ret;
1412
1413         if (nodeid == dlm_our_nodeid())
1414                 return 0;
1415
1416         dlm_stop_lockspaces_check();
1417
1418         idx = srcu_read_lock(&nodes_srcu);
1419         /* Abort pending close/remove operation */
1420         node = nodeid2node(nodeid, 0);
1421         if (node) {
1422                 /* let shutdown waiters leave */
1423                 set_bit(DLM_NODE_FLAG_CLOSE, &node->flags);
1424                 wake_up(&node->shutdown_wait);
1425         }
1426         srcu_read_unlock(&nodes_srcu, idx);
1427
1428         synchronize_srcu(&nodes_srcu);
1429
1430         idx = srcu_read_lock(&nodes_srcu);
1431         mutex_lock(&close_lock);
1432         node = nodeid2node(nodeid, 0);
1433         if (!node) {
1434                 mutex_unlock(&close_lock);
1435                 srcu_read_unlock(&nodes_srcu, idx);
1436                 return dlm_lowcomms_close(nodeid);
1437         }
1438
1439         ret = dlm_lowcomms_close(nodeid);
1440         spin_lock(&node->state_lock);
1441         midcomms_node_reset(node);
1442         spin_unlock(&node->state_lock);
1443         srcu_read_unlock(&nodes_srcu, idx);
1444         mutex_unlock(&close_lock);
1445
1446         return ret;
1447 }
1448
1449 /* debug functionality to send raw dlm msg from user space */
1450 struct dlm_rawmsg_data {
1451         struct midcomms_node *node;
1452         void *buf;
1453 };
1454
1455 static void midcomms_new_rawmsg_cb(void *data)
1456 {
1457         struct dlm_rawmsg_data *rd = data;
1458         struct dlm_header *h = rd->buf;
1459
1460         switch (h->h_version) {
1461         case cpu_to_le32(DLM_VERSION_3_1):
1462                 break;
1463         default:
1464                 switch (h->h_cmd) {
1465                 case DLM_OPTS:
1466                         if (!h->u.h_seq)
1467                                 h->u.h_seq = cpu_to_le32(rd->node->seq_send++);
1468                         break;
1469                 default:
1470                         break;
1471                 }
1472                 break;
1473         }
1474 }
1475
1476 int dlm_midcomms_rawmsg_send(struct midcomms_node *node, void *buf,
1477                              int buflen)
1478 {
1479         struct dlm_rawmsg_data rd;
1480         struct dlm_msg *msg;
1481         char *msgbuf;
1482
1483         rd.node = node;
1484         rd.buf = buf;
1485
1486         msg = dlm_lowcomms_new_msg(node->nodeid, buflen, GFP_NOFS,
1487                                    &msgbuf, midcomms_new_rawmsg_cb, &rd);
1488         if (!msg)
1489                 return -ENOMEM;
1490
1491         memcpy(msgbuf, buf, buflen);
1492         dlm_lowcomms_commit_msg(msg);
1493         return 0;
1494 }
1495