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 <trace/events/dlm.h>
 136 #include <net/tcp.h>
 137
 138 #include "dlm_internal.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 /* 5 seconds wait to sync ending of dlm */
 149 #define DLM_SHUTDOWN_TIMEOUT    msecs_to_jiffies(5000)
 150 #define DLM_VERSION_NOT_SET     0
 151 #define DLM_SEND_ACK_BACK_MSG_THRESHOLD 32
 152 #define DLM_RECV_ACK_BACK_MSG_THRESHOLD (DLM_SEND_ACK_BACK_MSG_THRESHOLD * 8)
 153
 154 struct midcomms_node {
 155         int nodeid;
 156         uint32_t version;
 157         atomic_t seq_send;
 158         atomic_t seq_next;
 159         /* These queues are unbound because we cannot drop any message in dlm.
 160          * We could send a fence signal for a specific node to the cluster
 161          * manager if queues hits some maximum value, however this handling
 162          * not supported yet.
 163          */
 164         struct list_head send_queue;
 165         spinlock_t send_queue_lock;
 166         atomic_t send_queue_cnt;
 167 #define DLM_NODE_FLAG_CLOSE     1
 168 #define DLM_NODE_FLAG_STOP_TX   2
 169 #define DLM_NODE_FLAG_STOP_RX   3
 170         atomic_t ulp_delivered;
 171         unsigned long flags;
 172         wait_queue_head_t shutdown_wait;
 173
 174         /* dlm tcp termination state */
 175 #define DLM_CLOSED      1
 176 #define DLM_ESTABLISHED 2
 177 #define DLM_FIN_WAIT1   3
 178 #define DLM_FIN_WAIT2   4
 179 #define DLM_CLOSE_WAIT  5
 180 #define DLM_LAST_ACK    6
 181 #define DLM_CLOSING     7
 182         int state;
 183         spinlock_t state_lock;
 184
 185         /* counts how many lockspaces are using this node
 186          * this refcount is necessary to determine if the
 187          * node wants to disconnect.
 188          */
 189         int users;
 190
 191         /* not protected by srcu, node_hash lifetime */
 192         void *debugfs;
 193
 194         struct hlist_node hlist;
 195         struct rcu_head rcu;
 196 };
 197
 198 struct dlm_mhandle {
 199         const union dlm_packet *inner_p;
 200         struct midcomms_node *node;
 201         struct dlm_opts *opts;
 202         struct dlm_msg *msg;
 203         bool committed;
 204         uint32_t seq;
 205
 206         void (*ack_rcv)(struct midcomms_node *node);
 207
 208         /* get_mhandle/commit srcu idx exchange */
 209         int idx;
 210
 211         struct list_head list;
 212         struct rcu_head rcu;
 213 };
 214
 215 static struct hlist_head node_hash[CONN_HASH_SIZE];
 216 static DEFINE_SPINLOCK(nodes_lock);
 217 DEFINE_STATIC_SRCU(nodes_srcu);
 218
 219 /* This mutex prevents that midcomms_close() is running while
 220  * stop() or remove(). As I experienced invalid memory access
 221  * behaviours when DLM_DEBUG_FENCE_TERMINATION is enabled and
 222  * resetting machines. I will end in some double deletion in nodes
 223  * datastructure.
 224  */
 225 static DEFINE_MUTEX(close_lock);
 226
 227 struct kmem_cache *dlm_midcomms_cache_create(void)
 228 {
 229         return kmem_cache_create("dlm_mhandle", sizeof(struct dlm_mhandle),
 230                                  0, 0, NULL);
 231 }
 232
 233 static inline const char *dlm_state_str(int state)
 234 {
 235         switch (state) {
 236         case DLM_CLOSED:
 237                 return "CLOSED";
 238         case DLM_ESTABLISHED:
 239                 return "ESTABLISHED";
 240         case DLM_FIN_WAIT1:
 241                 return "FIN_WAIT1";
 242         case DLM_FIN_WAIT2:
 243                 return "FIN_WAIT2";
 244         case DLM_CLOSE_WAIT:
 245                 return "CLOSE_WAIT";
 246         case DLM_LAST_ACK:
 247                 return "LAST_ACK";
 248         case DLM_CLOSING:
 249                 return "CLOSING";
 250         default:
 251                 return "UNKNOWN";
 252         }
 253 }
 254
 255 const char *dlm_midcomms_state(struct midcomms_node *node)
 256 {
 257         return dlm_state_str(node->state);
 258 }
 259
 260 unsigned long dlm_midcomms_flags(struct midcomms_node *node)
 261 {
 262         return node->flags;
 263 }
 264
 265 int dlm_midcomms_send_queue_cnt(struct midcomms_node *node)
 266 {
 267         return atomic_read(&node->send_queue_cnt);
 268 }
 269
 270 uint32_t dlm_midcomms_version(struct midcomms_node *node)
 271 {
 272         return node->version;
 273 }
 274
 275 static struct midcomms_node *__find_node(int nodeid, int r)
 276 {
 277         struct midcomms_node *node;
 278
 279         hlist_for_each_entry_rcu(node, &node_hash[r], hlist) {
 280                 if (node->nodeid == nodeid)
 281                         return node;
 282         }
 283
 284         return NULL;
 285 }
 286
 287 static void dlm_mhandle_release(struct rcu_head *rcu)
 288 {
 289         struct dlm_mhandle *mh = container_of(rcu, struct dlm_mhandle, rcu);
 290
 291         dlm_lowcomms_put_msg(mh->msg);
 292         dlm_free_mhandle(mh);
 293 }
 294
 295 static void dlm_mhandle_delete(struct midcomms_node *node,
 296                                struct dlm_mhandle *mh)
 297 {
 298         list_del_rcu(&mh->list);
 299         atomic_dec(&node->send_queue_cnt);
 300         call_rcu(&mh->rcu, dlm_mhandle_release);
 301 }
 302
 303 static void dlm_send_queue_flush(struct midcomms_node *node)
 304 {
 305         struct dlm_mhandle *mh;
 306
 307         pr_debug("flush midcomms send queue of node %d\n", node->nodeid);
 308
 309         rcu_read_lock();
 310         spin_lock_bh(&node->send_queue_lock);
 311         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 312                 dlm_mhandle_delete(node, mh);
 313         }
 314         spin_unlock_bh(&node->send_queue_lock);
 315         rcu_read_unlock();
 316 }
 317
 318 static void midcomms_node_reset(struct midcomms_node *node)
 319 {
 320         pr_debug("reset node %d\n", node->nodeid);
 321
 322         atomic_set(&node->seq_next, DLM_SEQ_INIT);
 323         atomic_set(&node->seq_send, DLM_SEQ_INIT);
 324         atomic_set(&node->ulp_delivered, 0);
 325         node->version = DLM_VERSION_NOT_SET;
 326         node->flags = 0;
 327
 328         dlm_send_queue_flush(node);
 329         node->state = DLM_CLOSED;
 330         wake_up(&node->shutdown_wait);
 331 }
 332
 333 static struct midcomms_node *nodeid2node(int nodeid)
 334 {
 335         return __find_node(nodeid, nodeid_hash(nodeid));
 336 }
 337
 338 int dlm_midcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
 339 {
 340         int ret, r = nodeid_hash(nodeid);
 341         struct midcomms_node *node;
 342
 343         ret = dlm_lowcomms_addr(nodeid, addr, len);
 344         if (ret)
 345                 return ret;
 346
 347         node = kmalloc(sizeof(*node), GFP_NOFS);
 348         if (!node)
 349                 return -ENOMEM;
 350
 351         node->nodeid = nodeid;
 352         spin_lock_init(&node->state_lock);
 353         spin_lock_init(&node->send_queue_lock);
 354         atomic_set(&node->send_queue_cnt, 0);
 355         INIT_LIST_HEAD(&node->send_queue);
 356         init_waitqueue_head(&node->shutdown_wait);
 357         node->users = 0;
 358         midcomms_node_reset(node);
 359
 360         spin_lock(&nodes_lock);
 361         hlist_add_head_rcu(&node->hlist, &node_hash[r]);
 362         spin_unlock(&nodes_lock);
 363
 364         node->debugfs = dlm_create_debug_comms_file(nodeid, node);
 365         return 0;
 366 }
 367
 368 static int dlm_send_ack(int nodeid, uint32_t seq)
 369 {
 370         int mb_len = sizeof(struct dlm_header);
 371         struct dlm_header *m_header;
 372         struct dlm_msg *msg;
 373         char *ppc;
 374
 375         msg = dlm_lowcomms_new_msg(nodeid, mb_len, GFP_ATOMIC, &ppc,
 376                                    NULL, NULL);
 377         if (!msg)
 378                 return -ENOMEM;
 379
 380         m_header = (struct dlm_header *)ppc;
 381
 382         m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
 383         m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
 384         m_header->h_length = cpu_to_le16(mb_len);
 385         m_header->h_cmd = DLM_ACK;
 386         m_header->u.h_seq = cpu_to_le32(seq);
 387
 388         dlm_lowcomms_commit_msg(msg);
 389         dlm_lowcomms_put_msg(msg);
 390
 391         return 0;
 392 }
 393
 394 static void dlm_send_ack_threshold(struct midcomms_node *node,
 395                                    uint32_t threshold)
 396 {
 397         uint32_t oval, nval;
 398         bool send_ack;
 399
 400         /* let only send one user trigger threshold to send ack back */
 401         do {
 402                 oval = atomic_read(&node->ulp_delivered);
 403                 send_ack = (oval > threshold);
 404                 /* abort if threshold is not reached */
 405                 if (!send_ack)
 406                         break;
 407
 408                 nval = 0;
 409                 /* try to reset ulp_delivered counter */
 410         } while (atomic_cmpxchg(&node->ulp_delivered, oval, nval) != oval);
 411
 412         if (send_ack)
 413                 dlm_send_ack(node->nodeid, atomic_read(&node->seq_next));
 414 }
 415
 416 static int dlm_send_fin(struct midcomms_node *node,
 417                         void (*ack_rcv)(struct midcomms_node *node))
 418 {
 419         int mb_len = sizeof(struct dlm_header);
 420         struct dlm_header *m_header;
 421         struct dlm_mhandle *mh;
 422         char *ppc;
 423
 424         mh = dlm_midcomms_get_mhandle(node->nodeid, mb_len, GFP_ATOMIC, &ppc);
 425         if (!mh)
 426                 return -ENOMEM;
 427
 428         set_bit(DLM_NODE_FLAG_STOP_TX, &node->flags);
 429         mh->ack_rcv = ack_rcv;
 430
 431         m_header = (struct dlm_header *)ppc;
 432
 433         m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
 434         m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
 435         m_header->h_length = cpu_to_le16(mb_len);
 436         m_header->h_cmd = DLM_FIN;
 437
 438         pr_debug("sending fin msg to node %d\n", node->nodeid);
 439         dlm_midcomms_commit_mhandle(mh, NULL, 0);
 440
 441         return 0;
 442 }
 443
 444 static void dlm_receive_ack(struct midcomms_node *node, uint32_t seq)
 445 {
 446         struct dlm_mhandle *mh;
 447
 448         rcu_read_lock();
 449         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 450                 if (before(mh->seq, seq)) {
 451                         if (mh->ack_rcv)
 452                                 mh->ack_rcv(node);
 453                 } else {
 454                         /* send queue should be ordered */
 455                         break;
 456                 }
 457         }
 458
 459         spin_lock_bh(&node->send_queue_lock);
 460         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 461                 if (before(mh->seq, seq)) {
 462                         dlm_mhandle_delete(node, mh);
 463                 } else {
 464                         /* send queue should be ordered */
 465                         break;
 466                 }
 467         }
 468         spin_unlock_bh(&node->send_queue_lock);
 469         rcu_read_unlock();
 470 }
 471
 472 static void dlm_pas_fin_ack_rcv(struct midcomms_node *node)
 473 {
 474         spin_lock(&node->state_lock);
 475         pr_debug("receive passive fin ack from node %d with state %s\n",
 476                  node->nodeid, dlm_state_str(node->state));
 477
 478         switch (node->state) {
 479         case DLM_LAST_ACK:
 480                 /* DLM_CLOSED */
 481                 midcomms_node_reset(node);
 482                 break;
 483         case DLM_CLOSED:
 484                 /* not valid but somehow we got what we want */
 485                 wake_up(&node->shutdown_wait);
 486                 break;
 487         default:
 488                 spin_unlock(&node->state_lock);
 489                 log_print("%s: unexpected state: %d",
 490                           __func__, node->state);
 491                 WARN_ON_ONCE(1);
 492                 return;
 493         }
 494         spin_unlock(&node->state_lock);
 495 }
 496
 497 static void dlm_receive_buffer_3_2_trace(uint32_t seq,
 498                                          const union dlm_packet *p)
 499 {
 500         switch (p->header.h_cmd) {
 501         case DLM_MSG:
 502                 trace_dlm_recv_message(dlm_our_nodeid(), seq, &p->message);
 503                 break;
 504         case DLM_RCOM:
 505                 trace_dlm_recv_rcom(dlm_our_nodeid(), seq, &p->rcom);
 506                 break;
 507         default:
 508                 break;
 509         }
 510 }
 511
 512 static void dlm_midcomms_receive_buffer(const union dlm_packet *p,
 513                                         struct midcomms_node *node,
 514                                         uint32_t seq)
 515 {
 516         bool is_expected_seq;
 517         uint32_t oval, nval;
 518
 519         do {
 520                 oval = atomic_read(&node->seq_next);
 521                 is_expected_seq = (oval == seq);
 522                 if (!is_expected_seq)
 523                         break;
 524
 525                 nval = oval + 1;
 526         } while (atomic_cmpxchg(&node->seq_next, oval, nval) != oval);
 527
 528         if (is_expected_seq) {
 529                 switch (p->header.h_cmd) {
 530                 case DLM_FIN:
 531                         spin_lock(&node->state_lock);
 532                         pr_debug("receive fin msg from node %d with state %s\n",
 533                                  node->nodeid, dlm_state_str(node->state));
 534
 535                         switch (node->state) {
 536                         case DLM_ESTABLISHED:
 537                                 dlm_send_ack(node->nodeid, nval);
 538
 539                                 /* passive shutdown DLM_LAST_ACK case 1
 540                                  * additional we check if the node is used by
 541                                  * cluster manager events at all.
 542                                  */
 543                                 if (node->users == 0) {
 544                                         node->state = DLM_LAST_ACK;
 545                                         pr_debug("switch node %d to state %s case 1\n",
 546                                                  node->nodeid, dlm_state_str(node->state));
 547                                         set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
 548                                         dlm_send_fin(node, dlm_pas_fin_ack_rcv);
 549                                 } else {
 550                                         node->state = DLM_CLOSE_WAIT;
 551                                         pr_debug("switch node %d to state %s\n",
 552                                                  node->nodeid, dlm_state_str(node->state));
 553                                 }
 554                                 break;
 555                         case DLM_FIN_WAIT1:
 556                                 dlm_send_ack(node->nodeid, nval);
 557                                 node->state = DLM_CLOSING;
 558                                 set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
 559                                 pr_debug("switch node %d to state %s\n",
 560                                          node->nodeid, dlm_state_str(node->state));
 561                                 break;
 562                         case DLM_FIN_WAIT2:
 563                                 dlm_send_ack(node->nodeid, nval);
 564                                 midcomms_node_reset(node);
 565                                 pr_debug("switch node %d to state %s\n",
 566                                          node->nodeid, dlm_state_str(node->state));
 567                                 break;
 568                         case DLM_LAST_ACK:
 569                                 /* probably remove_member caught it, do nothing */
 570                                 break;
 571                         default:
 572                                 spin_unlock(&node->state_lock);
 573                                 log_print("%s: unexpected state: %d",
 574                                           __func__, node->state);
 575                                 WARN_ON_ONCE(1);
 576                                 return;
 577                         }
 578                         spin_unlock(&node->state_lock);
 579                         break;
 580                 default:
 581                         WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
 582                         dlm_receive_buffer_3_2_trace(seq, p);
 583                         dlm_receive_buffer(p, node->nodeid);
 584                         atomic_inc(&node->ulp_delivered);
 585                         /* unlikely case to send ack back when we don't transmit */
 586                         dlm_send_ack_threshold(node, DLM_RECV_ACK_BACK_MSG_THRESHOLD);
 587                         break;
 588                 }
 589         } else {
 590                 /* retry to ack message which we already have by sending back
 591                  * current node->seq_next number as ack.
 592                  */
 593                 if (seq < oval)
 594                         dlm_send_ack(node->nodeid, oval);
 595
 596                 log_print_ratelimited("ignore dlm msg because seq mismatch, seq: %u, expected: %u, nodeid: %d",
 597                                       seq, oval, node->nodeid);
 598         }
 599 }
 600
 601 static int dlm_opts_check_msglen(const union dlm_packet *p, uint16_t msglen,
 602                                  int nodeid)
 603 {
 604         int len = msglen;
 605
 606         /* we only trust outer header msglen because
 607          * it's checked against receive buffer length.
 608          */
 609         if (len < sizeof(struct dlm_opts))
 610                 return -1;
 611         len -= sizeof(struct dlm_opts);
 612
 613         if (len < le16_to_cpu(p->opts.o_optlen))
 614                 return -1;
 615         len -= le16_to_cpu(p->opts.o_optlen);
 616
 617         switch (p->opts.o_nextcmd) {
 618         case DLM_FIN:
 619                 if (len < sizeof(struct dlm_header)) {
 620                         log_print("fin too small: %d, will skip this message from node %d",
 621                                   len, nodeid);
 622                         return -1;
 623                 }
 624
 625                 break;
 626         case DLM_MSG:
 627                 if (len < sizeof(struct dlm_message)) {
 628                         log_print("msg too small: %d, will skip this message from node %d",
 629                                   msglen, nodeid);
 630                         return -1;
 631                 }
 632
 633                 break;
 634         case DLM_RCOM:
 635                 if (len < sizeof(struct dlm_rcom)) {
 636                         log_print("rcom msg too small: %d, will skip this message from node %d",
 637                                   len, nodeid);
 638                         return -1;
 639                 }
 640
 641                 break;
 642         default:
 643                 log_print("unsupported o_nextcmd received: %u, will skip this message from node %d",
 644                           p->opts.o_nextcmd, nodeid);
 645                 return -1;
 646         }
 647
 648         return 0;
 649 }
 650
 651 static void dlm_midcomms_receive_buffer_3_2(const union dlm_packet *p, int nodeid)
 652 {
 653         uint16_t msglen = le16_to_cpu(p->header.h_length);
 654         struct midcomms_node *node;
 655         uint32_t seq;
 656         int ret, idx;
 657
 658         idx = srcu_read_lock(&nodes_srcu);
 659         node = nodeid2node(nodeid);
 660         if (WARN_ON_ONCE(!node))
 661                 goto out;
 662
 663         switch (node->version) {
 664         case DLM_VERSION_NOT_SET:
 665                 node->version = DLM_VERSION_3_2;
 666                 wake_up(&node->shutdown_wait);
 667                 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_2,
 668                           node->nodeid);
 669
 670                 spin_lock(&node->state_lock);
 671                 switch (node->state) {
 672                 case DLM_CLOSED:
 673                         node->state = DLM_ESTABLISHED;
 674                         pr_debug("switch node %d to state %s\n",
 675                                  node->nodeid, dlm_state_str(node->state));
 676                         break;
 677                 default:
 678                         break;
 679                 }
 680                 spin_unlock(&node->state_lock);
 681
 682                 break;
 683         case DLM_VERSION_3_2:
 684                 break;
 685         default:
 686                 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
 687                                       DLM_VERSION_3_2, node->nodeid, node->version);
 688                 goto out;
 689         }
 690
 691         switch (p->header.h_cmd) {
 692         case DLM_RCOM:
 693                 /* these rcom message we use to determine version.
 694                  * they have their own retransmission handling and
 695                  * are the first messages of dlm.
 696                  *
 697                  * length already checked.
 698                  */
 699                 switch (p->rcom.rc_type) {
 700                 case cpu_to_le32(DLM_RCOM_NAMES):
 701                         fallthrough;
 702                 case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
 703                         fallthrough;
 704                 case cpu_to_le32(DLM_RCOM_STATUS):
 705                         fallthrough;
 706                 case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
 707                         break;
 708                 default:
 709                         log_print("unsupported rcom type received: %u, will skip this message from node %d",
 710                                   le32_to_cpu(p->rcom.rc_type), nodeid);
 711                         goto out;
 712                 }
 713
 714                 WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
 715                 dlm_receive_buffer(p, nodeid);
 716                 break;
 717         case DLM_OPTS:
 718                 seq = le32_to_cpu(p->header.u.h_seq);
 719
 720                 ret = dlm_opts_check_msglen(p, msglen, nodeid);
 721                 if (ret < 0) {
 722                         log_print("opts msg too small: %u, will skip this message from node %d",
 723                                   msglen, nodeid);
 724                         goto out;
 725                 }
 726
 727                 p = (union dlm_packet *)((unsigned char *)p->opts.o_opts +
 728                                          le16_to_cpu(p->opts.o_optlen));
 729
 730                 /* recheck inner msglen just if it's not garbage */
 731                 msglen = le16_to_cpu(p->header.h_length);
 732                 switch (p->header.h_cmd) {
 733                 case DLM_RCOM:
 734                         if (msglen < sizeof(struct dlm_rcom)) {
 735                                 log_print("inner rcom msg too small: %u, will skip this message from node %d",
 736                                           msglen, nodeid);
 737                                 goto out;
 738                         }
 739
 740                         break;
 741                 case DLM_MSG:
 742                         if (msglen < sizeof(struct dlm_message)) {
 743                                 log_print("inner msg too small: %u, will skip this message from node %d",
 744                                           msglen, nodeid);
 745                                 goto out;
 746                         }
 747
 748                         break;
 749                 case DLM_FIN:
 750                         if (msglen < sizeof(struct dlm_header)) {
 751                                 log_print("inner fin too small: %u, will skip this message from node %d",
 752                                           msglen, nodeid);
 753                                 goto out;
 754                         }
 755
 756                         break;
 757                 default:
 758                         log_print("unsupported inner h_cmd received: %u, will skip this message from node %d",
 759                                   msglen, nodeid);
 760                         goto out;
 761                 }
 762
 763                 dlm_midcomms_receive_buffer(p, node, seq);
 764                 break;
 765         case DLM_ACK:
 766                 seq = le32_to_cpu(p->header.u.h_seq);
 767                 dlm_receive_ack(node, seq);
 768                 break;
 769         default:
 770                 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
 771                           p->header.h_cmd, nodeid);
 772                 break;
 773         }
 774
 775 out:
 776         srcu_read_unlock(&nodes_srcu, idx);
 777 }
 778
 779 static void dlm_midcomms_receive_buffer_3_1(const union dlm_packet *p, int nodeid)
 780 {
 781         uint16_t msglen = le16_to_cpu(p->header.h_length);
 782         struct midcomms_node *node;
 783         int idx;
 784
 785         idx = srcu_read_lock(&nodes_srcu);
 786         node = nodeid2node(nodeid);
 787         if (WARN_ON_ONCE(!node)) {
 788                 srcu_read_unlock(&nodes_srcu, idx);
 789                 return;
 790         }
 791
 792         switch (node->version) {
 793         case DLM_VERSION_NOT_SET:
 794                 node->version = DLM_VERSION_3_1;
 795                 wake_up(&node->shutdown_wait);
 796                 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_1,
 797                           node->nodeid);
 798                 break;
 799         case DLM_VERSION_3_1:
 800                 break;
 801         default:
 802                 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
 803                                       DLM_VERSION_3_1, node->nodeid, node->version);
 804                 srcu_read_unlock(&nodes_srcu, idx);
 805                 return;
 806         }
 807         srcu_read_unlock(&nodes_srcu, idx);
 808
 809         switch (p->header.h_cmd) {
 810         case DLM_RCOM:
 811                 /* length already checked */
 812                 break;
 813         case DLM_MSG:
 814                 if (msglen < sizeof(struct dlm_message)) {
 815                         log_print("msg too small: %u, will skip this message from node %d",
 816                                   msglen, nodeid);
 817                         return;
 818                 }
 819
 820                 break;
 821         default:
 822                 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
 823                           p->header.h_cmd, nodeid);
 824                 return;
 825         }
 826
 827         dlm_receive_buffer(p, nodeid);
 828 }
 829
 830 int dlm_validate_incoming_buffer(int nodeid, unsigned char *buf, int len)
 831 {
 832         const unsigned char *ptr = buf;
 833         const struct dlm_header *hd;
 834         uint16_t msglen;
 835         int ret = 0;
 836
 837         while (len >= sizeof(struct dlm_header)) {
 838                 hd = (struct dlm_header *)ptr;
 839
 840                 /* no message should be more than DLM_MAX_SOCKET_BUFSIZE or
 841                  * less than dlm_header size.
 842                  *
 843                  * Some messages does not have a 8 byte length boundary yet
 844                  * which can occur in a unaligned memory access of some dlm
 845                  * messages. However this problem need to be fixed at the
 846                  * sending side, for now it seems nobody run into architecture
 847                  * related issues yet but it slows down some processing.
 848                  * Fixing this issue should be scheduled in future by doing
 849                  * the next major version bump.
 850                  */
 851                 msglen = le16_to_cpu(hd->h_length);
 852                 if (msglen > DLM_MAX_SOCKET_BUFSIZE ||
 853                     msglen < sizeof(struct dlm_header)) {
 854                         log_print("received invalid length header: %u from node %d, will abort message parsing",
 855                                   msglen, nodeid);
 856                         return -EBADMSG;
 857                 }
 858
 859                 /* caller will take care that leftover
 860                  * will be parsed next call with more data
 861                  */
 862                 if (msglen > len)
 863                         break;
 864
 865                 ret += msglen;
 866                 len -= msglen;
 867                 ptr += msglen;
 868         }
 869
 870         return ret;
 871 }
 872
 873 /*
 874  * Called from the low-level comms layer to process a buffer of
 875  * commands.
 876  */
 877 int dlm_process_incoming_buffer(int nodeid, unsigned char *buf, int len)
 878 {
 879         const unsigned char *ptr = buf;
 880         const struct dlm_header *hd;
 881         uint16_t msglen;
 882         int ret = 0;
 883
 884         while (len >= sizeof(struct dlm_header)) {
 885                 hd = (struct dlm_header *)ptr;
 886
 887                 msglen = le16_to_cpu(hd->h_length);
 888                 if (msglen > len)
 889                         break;
 890
 891                 switch (hd->h_version) {
 892                 case cpu_to_le32(DLM_VERSION_3_1):
 893                         dlm_midcomms_receive_buffer_3_1((const union dlm_packet *)ptr, nodeid);
 894                         break;
 895                 case cpu_to_le32(DLM_VERSION_3_2):
 896                         dlm_midcomms_receive_buffer_3_2((const union dlm_packet *)ptr, nodeid);
 897                         break;
 898                 default:
 899                         log_print("received invalid version header: %u from node %d, will skip this message",
 900                                   le32_to_cpu(hd->h_version), nodeid);
 901                         break;
 902                 }
 903
 904                 ret += msglen;
 905                 len -= msglen;
 906                 ptr += msglen;
 907         }
 908
 909         return ret;
 910 }
 911
 912 void dlm_midcomms_unack_msg_resend(int nodeid)
 913 {
 914         struct midcomms_node *node;
 915         struct dlm_mhandle *mh;
 916         int idx, ret;
 917
 918         idx = srcu_read_lock(&nodes_srcu);
 919         node = nodeid2node(nodeid);
 920         if (WARN_ON_ONCE(!node)) {
 921                 srcu_read_unlock(&nodes_srcu, idx);
 922                 return;
 923         }
 924
 925         /* old protocol, we don't support to retransmit on failure */
 926         switch (node->version) {
 927         case DLM_VERSION_3_2:
 928                 break;
 929         default:
 930                 srcu_read_unlock(&nodes_srcu, idx);
 931                 return;
 932         }
 933
 934         rcu_read_lock();
 935         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 936                 if (!mh->committed)
 937                         continue;
 938
 939                 ret = dlm_lowcomms_resend_msg(mh->msg);
 940                 if (!ret)
 941                         log_print_ratelimited("retransmit dlm msg, seq %u, nodeid %d",
 942                                               mh->seq, node->nodeid);
 943         }
 944         rcu_read_unlock();
 945         srcu_read_unlock(&nodes_srcu, idx);
 946 }
 947
 948 static void dlm_fill_opts_header(struct dlm_opts *opts, uint16_t inner_len,
 949                                  uint32_t seq)
 950 {
 951         opts->o_header.h_cmd = DLM_OPTS;
 952         opts->o_header.h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
 953         opts->o_header.h_nodeid = cpu_to_le32(dlm_our_nodeid());
 954         opts->o_header.h_length = cpu_to_le16(DLM_MIDCOMMS_OPT_LEN + inner_len);
 955         opts->o_header.u.h_seq = cpu_to_le32(seq);
 956 }
 957
 958 static void midcomms_new_msg_cb(void *data)
 959 {
 960         struct dlm_mhandle *mh = data;
 961
 962         atomic_inc(&mh->node->send_queue_cnt);
 963
 964         spin_lock_bh(&mh->node->send_queue_lock);
 965         list_add_tail_rcu(&mh->list, &mh->node->send_queue);
 966         spin_unlock_bh(&mh->node->send_queue_lock);
 967
 968         mh->seq = atomic_fetch_inc(&mh->node->seq_send);
 969 }
 970
 971 static struct dlm_msg *dlm_midcomms_get_msg_3_2(struct dlm_mhandle *mh, int nodeid,
 972                                                 int len, gfp_t allocation, char **ppc)
 973 {
 974         struct dlm_opts *opts;
 975         struct dlm_msg *msg;
 976
 977         msg = dlm_lowcomms_new_msg(nodeid, len + DLM_MIDCOMMS_OPT_LEN,
 978                                    allocation, ppc, midcomms_new_msg_cb, mh);
 979         if (!msg)
 980                 return NULL;
 981
 982         opts = (struct dlm_opts *)*ppc;
 983         mh->opts = opts;
 984
 985         /* add possible options here */
 986         dlm_fill_opts_header(opts, len, mh->seq);
 987
 988         *ppc += sizeof(*opts);
 989         mh->inner_p = (const union dlm_packet *)*ppc;
 990         return msg;
 991 }
 992
 993 /* avoid false positive for nodes_srcu, unlock happens in
 994  * dlm_midcomms_commit_mhandle which is a must call if success
 995  */
 996 #ifndef __CHECKER__
 997 struct dlm_mhandle *dlm_midcomms_get_mhandle(int nodeid, int len,
 998                                              gfp_t allocation, char **ppc)
 999 {
1000         struct midcomms_node *node;
1001         struct dlm_mhandle *mh;
1002         struct dlm_msg *msg;
1003         int idx;
1004
1005         idx = srcu_read_lock(&nodes_srcu);
1006         node = nodeid2node(nodeid);
1007         if (WARN_ON_ONCE(!node))
1008                 goto err;
1009
1010         /* this is a bug, however we going on and hope it will be resolved */
1011         WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_TX, &node->flags));
1012
1013         mh = dlm_allocate_mhandle(allocation);
1014         if (!mh)
1015                 goto err;
1016
1017         mh->committed = false;
1018         mh->ack_rcv = NULL;
1019         mh->idx = idx;
1020         mh->node = node;
1021
1022         switch (node->version) {
1023         case DLM_VERSION_3_1:
1024                 msg = dlm_lowcomms_new_msg(nodeid, len, allocation, ppc,
1025                                            NULL, NULL);
1026                 if (!msg) {
1027                         dlm_free_mhandle(mh);
1028                         goto err;
1029                 }
1030
1031                 break;
1032         case DLM_VERSION_3_2:
1033                 msg = dlm_midcomms_get_msg_3_2(mh, nodeid, len, allocation,
1034                                                ppc);
1035                 if (!msg) {
1036                         dlm_free_mhandle(mh);
1037                         goto err;
1038                 }
1039
1040                 /* send ack back if necessary */
1041                 dlm_send_ack_threshold(node, DLM_SEND_ACK_BACK_MSG_THRESHOLD);
1042                 break;
1043         default:
1044                 dlm_free_mhandle(mh);
1045                 WARN_ON_ONCE(1);
1046                 goto err;
1047         }
1048
1049         mh->msg = msg;
1050
1051         /* keep in mind that is a must to call
1052          * dlm_midcomms_commit_msg() which releases
1053          * nodes_srcu using mh->idx which is assumed
1054          * here that the application will call it.
1055          */
1056         return mh;
1057
1058 err:
1059         srcu_read_unlock(&nodes_srcu, idx);
1060         return NULL;
1061 }
1062 #endif
1063
1064 static void dlm_midcomms_commit_msg_3_2_trace(const struct dlm_mhandle *mh,
1065                                               const void *name, int namelen)
1066 {
1067         switch (mh->inner_p->header.h_cmd) {
1068         case DLM_MSG:
1069                 trace_dlm_send_message(mh->node->nodeid, mh->seq,
1070                                        &mh->inner_p->message,
1071                                        name, namelen);
1072                 break;
1073         case DLM_RCOM:
1074                 trace_dlm_send_rcom(mh->node->nodeid, mh->seq,
1075                                     &mh->inner_p->rcom);
1076                 break;
1077         default:
1078                 /* nothing to trace */
1079                 break;
1080         }
1081 }
1082
1083 static void dlm_midcomms_commit_msg_3_2(struct dlm_mhandle *mh,
1084                                         const void *name, int namelen)
1085 {
1086         /* nexthdr chain for fast lookup */
1087         mh->opts->o_nextcmd = mh->inner_p->header.h_cmd;
1088         mh->committed = true;
1089         dlm_midcomms_commit_msg_3_2_trace(mh, name, namelen);
1090         dlm_lowcomms_commit_msg(mh->msg);
1091 }
1092
1093 /* avoid false positive for nodes_srcu, lock was happen in
1094  * dlm_midcomms_get_mhandle
1095  */
1096 #ifndef __CHECKER__
1097 void dlm_midcomms_commit_mhandle(struct dlm_mhandle *mh,
1098                                  const void *name, int namelen)
1099 {
1100
1101         switch (mh->node->version) {
1102         case DLM_VERSION_3_1:
1103                 srcu_read_unlock(&nodes_srcu, mh->idx);
1104
1105                 dlm_lowcomms_commit_msg(mh->msg);
1106                 dlm_lowcomms_put_msg(mh->msg);
1107                 /* mh is not part of rcu list in this case */
1108                 dlm_free_mhandle(mh);
1109                 break;
1110         case DLM_VERSION_3_2:
1111                 /* held rcu read lock here, because we sending the
1112                  * dlm message out, when we do that we could receive
1113                  * an ack back which releases the mhandle and we
1114                  * get a use after free.
1115                  */
1116                 rcu_read_lock();
1117                 dlm_midcomms_commit_msg_3_2(mh, name, namelen);
1118                 srcu_read_unlock(&nodes_srcu, mh->idx);
1119                 rcu_read_unlock();
1120                 break;
1121         default:
1122                 srcu_read_unlock(&nodes_srcu, mh->idx);
1123                 WARN_ON_ONCE(1);
1124                 break;
1125         }
1126 }
1127 #endif
1128
1129 int dlm_midcomms_start(void)
1130 {
1131         return dlm_lowcomms_start();
1132 }
1133
1134 void dlm_midcomms_stop(void)
1135 {
1136         dlm_lowcomms_stop();
1137 }
1138
1139 void dlm_midcomms_init(void)
1140 {
1141         int i;
1142
1143         for (i = 0; i < CONN_HASH_SIZE; i++)
1144                 INIT_HLIST_HEAD(&node_hash[i]);
1145
1146         dlm_lowcomms_init();
1147 }
1148
1149 static void midcomms_node_release(struct rcu_head *rcu)
1150 {
1151         struct midcomms_node *node = container_of(rcu, struct midcomms_node, rcu);
1152
1153         WARN_ON_ONCE(atomic_read(&node->send_queue_cnt));
1154         dlm_send_queue_flush(node);
1155         kfree(node);
1156 }
1157
1158 void dlm_midcomms_exit(void)
1159 {
1160         struct midcomms_node *node;
1161         int i, idx;
1162
1163         idx = srcu_read_lock(&nodes_srcu);
1164         for (i = 0; i < CONN_HASH_SIZE; i++) {
1165                 hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1166                         dlm_delete_debug_comms_file(node->debugfs);
1167
1168                         spin_lock(&nodes_lock);
1169                         hlist_del_rcu(&node->hlist);
1170                         spin_unlock(&nodes_lock);
1171
1172                         call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
1173                 }
1174         }
1175         srcu_read_unlock(&nodes_srcu, idx);
1176
1177         dlm_lowcomms_exit();
1178 }
1179
1180 static void dlm_act_fin_ack_rcv(struct midcomms_node *node)
1181 {
1182         spin_lock(&node->state_lock);
1183         pr_debug("receive active fin ack from node %d with state %s\n",
1184                  node->nodeid, dlm_state_str(node->state));
1185
1186         switch (node->state) {
1187         case DLM_FIN_WAIT1:
1188                 node->state = DLM_FIN_WAIT2;
1189                 pr_debug("switch node %d to state %s\n",
1190                          node->nodeid, dlm_state_str(node->state));
1191                 break;
1192         case DLM_CLOSING:
1193                 midcomms_node_reset(node);
1194                 pr_debug("switch node %d to state %s\n",
1195                          node->nodeid, dlm_state_str(node->state));
1196                 break;
1197         case DLM_CLOSED:
1198                 /* not valid but somehow we got what we want */
1199                 wake_up(&node->shutdown_wait);
1200                 break;
1201         default:
1202                 spin_unlock(&node->state_lock);
1203                 log_print("%s: unexpected state: %d",
1204                           __func__, node->state);
1205                 WARN_ON_ONCE(1);
1206                 return;
1207         }
1208         spin_unlock(&node->state_lock);
1209 }
1210
1211 void dlm_midcomms_add_member(int nodeid)
1212 {
1213         struct midcomms_node *node;
1214         int idx;
1215
1216         idx = srcu_read_lock(&nodes_srcu);
1217         node = nodeid2node(nodeid);
1218         if (WARN_ON_ONCE(!node)) {
1219                 srcu_read_unlock(&nodes_srcu, idx);
1220                 return;
1221         }
1222
1223         spin_lock(&node->state_lock);
1224         if (!node->users) {
1225                 pr_debug("receive add member from node %d with state %s\n",
1226                          node->nodeid, dlm_state_str(node->state));
1227                 switch (node->state) {
1228                 case DLM_ESTABLISHED:
1229                         break;
1230                 case DLM_CLOSED:
1231                         node->state = DLM_ESTABLISHED;
1232                         pr_debug("switch node %d to state %s\n",
1233                                  node->nodeid, dlm_state_str(node->state));
1234                         break;
1235                 default:
1236                         /* some invalid state passive shutdown
1237                          * was failed, we try to reset and
1238                          * hope it will go on.
1239                          */
1240                         log_print("reset node %d because shutdown stuck",
1241                                   node->nodeid);
1242
1243                         midcomms_node_reset(node);
1244                         node->state = DLM_ESTABLISHED;
1245                         break;
1246                 }
1247         }
1248
1249         node->users++;
1250         pr_debug("node %d users inc count %d\n", nodeid, node->users);
1251         spin_unlock(&node->state_lock);
1252
1253         srcu_read_unlock(&nodes_srcu, idx);
1254 }
1255
1256 void dlm_midcomms_remove_member(int nodeid)
1257 {
1258         struct midcomms_node *node;
1259         int idx;
1260
1261         idx = srcu_read_lock(&nodes_srcu);
1262         node = nodeid2node(nodeid);
1263         if (WARN_ON_ONCE(!node)) {
1264                 srcu_read_unlock(&nodes_srcu, idx);
1265                 return;
1266         }
1267
1268         spin_lock(&node->state_lock);
1269         node->users--;
1270         pr_debug("node %d users dec count %d\n", nodeid, node->users);
1271
1272         /* hitting users count to zero means the
1273          * other side is running dlm_midcomms_stop()
1274          * we meet us to have a clean disconnect.
1275          */
1276         if (node->users == 0) {
1277                 pr_debug("receive remove member from node %d with state %s\n",
1278                          node->nodeid, dlm_state_str(node->state));
1279                 switch (node->state) {
1280                 case DLM_ESTABLISHED:
1281                         break;
1282                 case DLM_CLOSE_WAIT:
1283                         /* passive shutdown DLM_LAST_ACK case 2 */
1284                         node->state = DLM_LAST_ACK;
1285                         pr_debug("switch node %d to state %s case 2\n",
1286                                  node->nodeid, dlm_state_str(node->state));
1287                         set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
1288                         dlm_send_fin(node, dlm_pas_fin_ack_rcv);
1289                         break;
1290                 case DLM_LAST_ACK:
1291                         /* probably receive fin caught it, do nothing */
1292                         break;
1293                 case DLM_CLOSED:
1294                         /* already gone, do nothing */
1295                         break;
1296                 default:
1297                         log_print("%s: unexpected state: %d",
1298                                   __func__, node->state);
1299                         break;
1300                 }
1301         }
1302         spin_unlock(&node->state_lock);
1303
1304         srcu_read_unlock(&nodes_srcu, idx);
1305 }
1306
1307 void dlm_midcomms_version_wait(void)
1308 {
1309         struct midcomms_node *node;
1310         int i, idx, ret;
1311
1312         idx = srcu_read_lock(&nodes_srcu);
1313         for (i = 0; i < CONN_HASH_SIZE; i++) {
1314                 hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1315                         ret = wait_event_timeout(node->shutdown_wait,
1316                                                  node->version != DLM_VERSION_NOT_SET ||
1317                                                  node->state == DLM_CLOSED ||
1318                                                  test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1319                                                  DLM_SHUTDOWN_TIMEOUT);
1320                         if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
1321                                 pr_debug("version wait timed out for node %d with state %s\n",
1322                                          node->nodeid, dlm_state_str(node->state));
1323                 }
1324         }
1325         srcu_read_unlock(&nodes_srcu, idx);
1326 }
1327
1328 static void midcomms_shutdown(struct midcomms_node *node)
1329 {
1330         int ret;
1331
1332         /* old protocol, we don't wait for pending operations */
1333         switch (node->version) {
1334         case DLM_VERSION_3_2:
1335                 break;
1336         default:
1337                 return;
1338         }
1339
1340         spin_lock(&node->state_lock);
1341         pr_debug("receive active shutdown for node %d with state %s\n",
1342                  node->nodeid, dlm_state_str(node->state));
1343         switch (node->state) {
1344         case DLM_ESTABLISHED:
1345                 node->state = DLM_FIN_WAIT1;
1346                 pr_debug("switch node %d to state %s case 2\n",
1347                          node->nodeid, dlm_state_str(node->state));
1348                 dlm_send_fin(node, dlm_act_fin_ack_rcv);
1349                 break;
1350         case DLM_CLOSED:
1351                 /* we have what we want */
1352                 break;
1353         default:
1354                 /* busy to enter DLM_FIN_WAIT1, wait until passive
1355                  * done in shutdown_wait to enter DLM_CLOSED.
1356                  */
1357                 break;
1358         }
1359         spin_unlock(&node->state_lock);
1360
1361         if (DLM_DEBUG_FENCE_TERMINATION)
1362                 msleep(5000);
1363
1364         /* wait for other side dlm + fin */
1365         ret = wait_event_timeout(node->shutdown_wait,
1366                                  node->state == DLM_CLOSED ||
1367                                  test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1368                                  DLM_SHUTDOWN_TIMEOUT);
1369         if (!ret)
1370                 pr_debug("active shutdown timed out for node %d with state %s\n",
1371                          node->nodeid, dlm_state_str(node->state));
1372         else
1373                 pr_debug("active shutdown done for node %d with state %s\n",
1374                          node->nodeid, dlm_state_str(node->state));
1375 }
1376
1377 void dlm_midcomms_shutdown(void)
1378 {
1379         struct midcomms_node *node;
1380         int i, idx;
1381
1382         mutex_lock(&close_lock);
1383         idx = srcu_read_lock(&nodes_srcu);
1384         for (i = 0; i < CONN_HASH_SIZE; i++) {
1385                 hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1386                         midcomms_shutdown(node);
1387                 }
1388         }
1389         srcu_read_unlock(&nodes_srcu, idx);
1390         mutex_unlock(&close_lock);
1391
1392         dlm_lowcomms_shutdown();
1393 }
1394
1395 int dlm_midcomms_close(int nodeid)
1396 {
1397         struct midcomms_node *node;
1398         int idx, ret;
1399
1400         idx = srcu_read_lock(&nodes_srcu);
1401         /* Abort pending close/remove operation */
1402         node = nodeid2node(nodeid);
1403         if (node) {
1404                 /* let shutdown waiters leave */
1405                 set_bit(DLM_NODE_FLAG_CLOSE, &node->flags);
1406                 wake_up(&node->shutdown_wait);
1407         }
1408         srcu_read_unlock(&nodes_srcu, idx);
1409
1410         synchronize_srcu(&nodes_srcu);
1411
1412         mutex_lock(&close_lock);
1413         idx = srcu_read_lock(&nodes_srcu);
1414         node = nodeid2node(nodeid);
1415         if (!node) {
1416                 srcu_read_unlock(&nodes_srcu, idx);
1417                 mutex_unlock(&close_lock);
1418                 return dlm_lowcomms_close(nodeid);
1419         }
1420
1421         ret = dlm_lowcomms_close(nodeid);
1422         dlm_delete_debug_comms_file(node->debugfs);
1423
1424         spin_lock(&nodes_lock);
1425         hlist_del_rcu(&node->hlist);
1426         spin_unlock(&nodes_lock);
1427         srcu_read_unlock(&nodes_srcu, idx);
1428
1429         /* wait that all readers left until flush send queue */
1430         synchronize_srcu(&nodes_srcu);
1431
1432         /* drop all pending dlm messages, this is fine as
1433          * this function get called when the node is fenced
1434          */
1435         dlm_send_queue_flush(node);
1436
1437         call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
1438         mutex_unlock(&close_lock);
1439
1440         return ret;
1441 }
1442
1443 /* debug functionality to send raw dlm msg from user space */
1444 struct dlm_rawmsg_data {
1445         struct midcomms_node *node;
1446         void *buf;
1447 };
1448
1449 static void midcomms_new_rawmsg_cb(void *data)
1450 {
1451         struct dlm_rawmsg_data *rd = data;
1452         struct dlm_header *h = rd->buf;
1453
1454         switch (h->h_version) {
1455         case cpu_to_le32(DLM_VERSION_3_1):
1456                 break;
1457         default:
1458                 switch (h->h_cmd) {
1459                 case DLM_OPTS:
1460                         if (!h->u.h_seq)
1461                                 h->u.h_seq = cpu_to_le32(atomic_fetch_inc(&rd->node->seq_send));
1462                         break;
1463                 default:
1464                         break;
1465                 }
1466                 break;
1467         }
1468 }
1469
1470 int dlm_midcomms_rawmsg_send(struct midcomms_node *node, void *buf,
1471                              int buflen)
1472 {
1473         struct dlm_rawmsg_data rd;
1474         struct dlm_msg *msg;
1475         char *msgbuf;
1476
1477         rd.node = node;
1478         rd.buf = buf;
1479
1480         msg = dlm_lowcomms_new_msg(node->nodeid, buflen, GFP_NOFS,
1481                                    &msgbuf, midcomms_new_rawmsg_cb, &rd);
1482         if (!msg)
1483                 return -ENOMEM;
1484
1485         memcpy(msgbuf, buf, buflen);
1486         dlm_lowcomms_commit_msg(msg);
1487         return 0;
1488 }
1489