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, gfp_t alloc)
 334 {
 335         struct midcomms_node *node, *tmp;
 336         int r = nodeid_hash(nodeid);
 337
 338         node = __find_node(nodeid, r);
 339         if (node || !alloc)
 340                 return node;
 341
 342         node = kmalloc(sizeof(*node), alloc);
 343         if (!node)
 344                 return NULL;
 345
 346         node->nodeid = nodeid;
 347         spin_lock_init(&node->state_lock);
 348         spin_lock_init(&node->send_queue_lock);
 349         atomic_set(&node->send_queue_cnt, 0);
 350         INIT_LIST_HEAD(&node->send_queue);
 351         init_waitqueue_head(&node->shutdown_wait);
 352         node->users = 0;
 353         midcomms_node_reset(node);
 354
 355         spin_lock(&nodes_lock);
 356         /* check again if there was somebody else
 357          * earlier here to add the node
 358          */
 359         tmp = __find_node(nodeid, r);
 360         if (tmp) {
 361                 spin_unlock(&nodes_lock);
 362                 kfree(node);
 363                 return tmp;
 364         }
 365
 366         hlist_add_head_rcu(&node->hlist, &node_hash[r]);
 367         spin_unlock(&nodes_lock);
 368
 369         node->debugfs = dlm_create_debug_comms_file(nodeid, node);
 370         return node;
 371 }
 372
 373 static int dlm_send_ack(int nodeid, uint32_t seq)
 374 {
 375         int mb_len = sizeof(struct dlm_header);
 376         struct dlm_header *m_header;
 377         struct dlm_msg *msg;
 378         char *ppc;
 379
 380         msg = dlm_lowcomms_new_msg(nodeid, mb_len, GFP_ATOMIC, &ppc,
 381                                    NULL, NULL);
 382         if (!msg)
 383                 return -ENOMEM;
 384
 385         m_header = (struct dlm_header *)ppc;
 386
 387         m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
 388         m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
 389         m_header->h_length = cpu_to_le16(mb_len);
 390         m_header->h_cmd = DLM_ACK;
 391         m_header->u.h_seq = cpu_to_le32(seq);
 392
 393         dlm_lowcomms_commit_msg(msg);
 394         dlm_lowcomms_put_msg(msg);
 395
 396         return 0;
 397 }
 398
 399 static void dlm_send_ack_threshold(struct midcomms_node *node,
 400                                    uint32_t threshold)
 401 {
 402         uint32_t oval, nval;
 403         bool send_ack;
 404
 405         /* let only send one user trigger threshold to send ack back */
 406         do {
 407                 oval = atomic_read(&node->ulp_delivered);
 408                 send_ack = (oval > threshold);
 409                 /* abort if threshold is not reached */
 410                 if (!send_ack)
 411                         break;
 412
 413                 nval = 0;
 414                 /* try to reset ulp_delivered counter */
 415         } while (atomic_cmpxchg(&node->ulp_delivered, oval, nval) != oval);
 416
 417         if (send_ack)
 418                 dlm_send_ack(node->nodeid, atomic_read(&node->seq_next));
 419 }
 420
 421 static int dlm_send_fin(struct midcomms_node *node,
 422                         void (*ack_rcv)(struct midcomms_node *node))
 423 {
 424         int mb_len = sizeof(struct dlm_header);
 425         struct dlm_header *m_header;
 426         struct dlm_mhandle *mh;
 427         char *ppc;
 428
 429         mh = dlm_midcomms_get_mhandle(node->nodeid, mb_len, GFP_ATOMIC, &ppc);
 430         if (!mh)
 431                 return -ENOMEM;
 432
 433         set_bit(DLM_NODE_FLAG_STOP_TX, &node->flags);
 434         mh->ack_rcv = ack_rcv;
 435
 436         m_header = (struct dlm_header *)ppc;
 437
 438         m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
 439         m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
 440         m_header->h_length = cpu_to_le16(mb_len);
 441         m_header->h_cmd = DLM_FIN;
 442
 443         pr_debug("sending fin msg to node %d\n", node->nodeid);
 444         dlm_midcomms_commit_mhandle(mh, NULL, 0);
 445
 446         return 0;
 447 }
 448
 449 static void dlm_receive_ack(struct midcomms_node *node, uint32_t seq)
 450 {
 451         struct dlm_mhandle *mh;
 452
 453         rcu_read_lock();
 454         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 455                 if (before(mh->seq, seq)) {
 456                         if (mh->ack_rcv)
 457                                 mh->ack_rcv(node);
 458                 } else {
 459                         /* send queue should be ordered */
 460                         break;
 461                 }
 462         }
 463
 464         spin_lock_bh(&node->send_queue_lock);
 465         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 466                 if (before(mh->seq, seq)) {
 467                         dlm_mhandle_delete(node, mh);
 468                 } else {
 469                         /* send queue should be ordered */
 470                         break;
 471                 }
 472         }
 473         spin_unlock_bh(&node->send_queue_lock);
 474         rcu_read_unlock();
 475 }
 476
 477 static void dlm_pas_fin_ack_rcv(struct midcomms_node *node)
 478 {
 479         spin_lock(&node->state_lock);
 480         pr_debug("receive passive fin ack from node %d with state %s\n",
 481                  node->nodeid, dlm_state_str(node->state));
 482
 483         switch (node->state) {
 484         case DLM_LAST_ACK:
 485                 /* DLM_CLOSED */
 486                 midcomms_node_reset(node);
 487                 break;
 488         case DLM_CLOSED:
 489                 /* not valid but somehow we got what we want */
 490                 wake_up(&node->shutdown_wait);
 491                 break;
 492         default:
 493                 spin_unlock(&node->state_lock);
 494                 log_print("%s: unexpected state: %d",
 495                           __func__, node->state);
 496                 WARN_ON_ONCE(1);
 497                 return;
 498         }
 499         spin_unlock(&node->state_lock);
 500 }
 501
 502 static void dlm_receive_buffer_3_2_trace(uint32_t seq, union dlm_packet *p)
 503 {
 504         switch (p->header.h_cmd) {
 505         case DLM_MSG:
 506                 trace_dlm_recv_message(dlm_our_nodeid(), seq, &p->message);
 507                 break;
 508         case DLM_RCOM:
 509                 trace_dlm_recv_rcom(dlm_our_nodeid(), seq, &p->rcom);
 510                 break;
 511         default:
 512                 break;
 513         }
 514 }
 515
 516 static void dlm_midcomms_receive_buffer(union dlm_packet *p,
 517                                         struct midcomms_node *node,
 518                                         uint32_t seq)
 519 {
 520         bool is_expected_seq;
 521         uint32_t oval, nval;
 522
 523         do {
 524                 oval = atomic_read(&node->seq_next);
 525                 is_expected_seq = (oval == seq);
 526                 if (!is_expected_seq)
 527                         break;
 528
 529                 nval = oval + 1;
 530         } while (atomic_cmpxchg(&node->seq_next, oval, nval) != oval);
 531
 532         if (is_expected_seq) {
 533                 switch (p->header.h_cmd) {
 534                 case DLM_FIN:
 535                         spin_lock(&node->state_lock);
 536                         pr_debug("receive fin msg from node %d with state %s\n",
 537                                  node->nodeid, dlm_state_str(node->state));
 538
 539                         switch (node->state) {
 540                         case DLM_ESTABLISHED:
 541                                 dlm_send_ack(node->nodeid, nval);
 542
 543                                 /* passive shutdown DLM_LAST_ACK case 1
 544                                  * additional we check if the node is used by
 545                                  * cluster manager events at all.
 546                                  */
 547                                 if (node->users == 0) {
 548                                         node->state = DLM_LAST_ACK;
 549                                         pr_debug("switch node %d to state %s case 1\n",
 550                                                  node->nodeid, dlm_state_str(node->state));
 551                                         set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
 552                                         dlm_send_fin(node, dlm_pas_fin_ack_rcv);
 553                                 } else {
 554                                         node->state = DLM_CLOSE_WAIT;
 555                                         pr_debug("switch node %d to state %s\n",
 556                                                  node->nodeid, dlm_state_str(node->state));
 557                                 }
 558                                 break;
 559                         case DLM_FIN_WAIT1:
 560                                 dlm_send_ack(node->nodeid, nval);
 561                                 node->state = DLM_CLOSING;
 562                                 set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
 563                                 pr_debug("switch node %d to state %s\n",
 564                                          node->nodeid, dlm_state_str(node->state));
 565                                 break;
 566                         case DLM_FIN_WAIT2:
 567                                 dlm_send_ack(node->nodeid, nval);
 568                                 midcomms_node_reset(node);
 569                                 pr_debug("switch node %d to state %s\n",
 570                                          node->nodeid, dlm_state_str(node->state));
 571                                 break;
 572                         case DLM_LAST_ACK:
 573                                 /* probably remove_member caught it, do nothing */
 574                                 break;
 575                         default:
 576                                 spin_unlock(&node->state_lock);
 577                                 log_print("%s: unexpected state: %d",
 578                                           __func__, node->state);
 579                                 WARN_ON_ONCE(1);
 580                                 return;
 581                         }
 582                         spin_unlock(&node->state_lock);
 583                         break;
 584                 default:
 585                         WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
 586                         dlm_receive_buffer_3_2_trace(seq, p);
 587                         dlm_receive_buffer(p, node->nodeid);
 588                         atomic_inc(&node->ulp_delivered);
 589                         /* unlikely case to send ack back when we don't transmit */
 590                         dlm_send_ack_threshold(node, DLM_RECV_ACK_BACK_MSG_THRESHOLD);
 591                         break;
 592                 }
 593         } else {
 594                 /* retry to ack message which we already have by sending back
 595                  * current node->seq_next number as ack.
 596                  */
 597                 if (seq < oval)
 598                         dlm_send_ack(node->nodeid, oval);
 599
 600                 log_print_ratelimited("ignore dlm msg because seq mismatch, seq: %u, expected: %u, nodeid: %d",
 601                                       seq, oval, node->nodeid);
 602         }
 603 }
 604
 605 static struct midcomms_node *
 606 dlm_midcomms_recv_node_lookup(int nodeid, const union dlm_packet *p,
 607                               uint16_t msglen, int (*cb)(struct midcomms_node *node))
 608 {
 609         struct midcomms_node *node = NULL;
 610         gfp_t allocation = 0;
 611         int ret;
 612
 613         switch (p->header.h_cmd) {
 614         case DLM_RCOM:
 615                 if (msglen < sizeof(struct dlm_rcom)) {
 616                         log_print("rcom msg too small: %u, will skip this message from node %d",
 617                                   msglen, nodeid);
 618                         return NULL;
 619                 }
 620
 621                 switch (p->rcom.rc_type) {
 622                 case cpu_to_le32(DLM_RCOM_NAMES):
 623                         fallthrough;
 624                 case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
 625                         fallthrough;
 626                 case cpu_to_le32(DLM_RCOM_STATUS):
 627                         fallthrough;
 628                 case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
 629                         node = nodeid2node(nodeid, 0);
 630                         if (node) {
 631                                 spin_lock(&node->state_lock);
 632                                 if (node->state != DLM_ESTABLISHED)
 633                                         pr_debug("receive begin RCOM msg from node %d with state %s\n",
 634                                                  node->nodeid, dlm_state_str(node->state));
 635
 636                                 switch (node->state) {
 637                                 case DLM_CLOSED:
 638                                         node->state = DLM_ESTABLISHED;
 639                                         pr_debug("switch node %d to state %s\n",
 640                                                  node->nodeid, dlm_state_str(node->state));
 641                                         break;
 642                                 case DLM_ESTABLISHED:
 643                                         break;
 644                                 default:
 645                                         spin_unlock(&node->state_lock);
 646                                         return NULL;
 647                                 }
 648                                 spin_unlock(&node->state_lock);
 649                         }
 650
 651                         allocation = GFP_NOFS;
 652                         break;
 653                 default:
 654                         break;
 655                 }
 656
 657                 break;
 658         default:
 659                 break;
 660         }
 661
 662         node = nodeid2node(nodeid, allocation);
 663         if (!node) {
 664                 switch (p->header.h_cmd) {
 665                 case DLM_OPTS:
 666                         if (msglen < sizeof(struct dlm_opts)) {
 667                                 log_print("opts msg too small: %u, will skip this message from node %d",
 668                                           msglen, nodeid);
 669                                 return NULL;
 670                         }
 671
 672                         log_print_ratelimited("received dlm opts message nextcmd %d from node %d in an invalid sequence",
 673                                               p->opts.o_nextcmd, nodeid);
 674                         break;
 675                 default:
 676                         log_print_ratelimited("received dlm message cmd %d from node %d in an invalid sequence",
 677                                               p->header.h_cmd, nodeid);
 678                         break;
 679                 }
 680
 681                 return NULL;
 682         }
 683
 684         ret = cb(node);
 685         if (ret < 0)
 686                 return NULL;
 687
 688         return node;
 689 }
 690
 691 static int dlm_midcomms_version_check_3_2(struct midcomms_node *node)
 692 {
 693         switch (node->version) {
 694         case DLM_VERSION_NOT_SET:
 695                 node->version = DLM_VERSION_3_2;
 696                 wake_up(&node->shutdown_wait);
 697                 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_2,
 698                           node->nodeid);
 699                 break;
 700         case DLM_VERSION_3_2:
 701                 break;
 702         default:
 703                 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
 704                                       DLM_VERSION_3_2, node->nodeid, node->version);
 705                 return -1;
 706         }
 707
 708         return 0;
 709 }
 710
 711 static int dlm_opts_check_msglen(union dlm_packet *p, uint16_t msglen, int nodeid)
 712 {
 713         int len = msglen;
 714
 715         /* we only trust outer header msglen because
 716          * it's checked against receive buffer length.
 717          */
 718         if (len < sizeof(struct dlm_opts))
 719                 return -1;
 720         len -= sizeof(struct dlm_opts);
 721
 722         if (len < le16_to_cpu(p->opts.o_optlen))
 723                 return -1;
 724         len -= le16_to_cpu(p->opts.o_optlen);
 725
 726         switch (p->opts.o_nextcmd) {
 727         case DLM_FIN:
 728                 if (len < sizeof(struct dlm_header)) {
 729                         log_print("fin too small: %d, will skip this message from node %d",
 730                                   len, nodeid);
 731                         return -1;
 732                 }
 733
 734                 break;
 735         case DLM_MSG:
 736                 if (len < sizeof(struct dlm_message)) {
 737                         log_print("msg too small: %d, will skip this message from node %d",
 738                                   msglen, nodeid);
 739                         return -1;
 740                 }
 741
 742                 break;
 743         case DLM_RCOM:
 744                 if (len < sizeof(struct dlm_rcom)) {
 745                         log_print("rcom msg too small: %d, will skip this message from node %d",
 746                                   len, nodeid);
 747                         return -1;
 748                 }
 749
 750                 break;
 751         default:
 752                 log_print("unsupported o_nextcmd received: %u, will skip this message from node %d",
 753                           p->opts.o_nextcmd, nodeid);
 754                 return -1;
 755         }
 756
 757         return 0;
 758 }
 759
 760 static void dlm_midcomms_receive_buffer_3_2(union dlm_packet *p, int nodeid)
 761 {
 762         uint16_t msglen = le16_to_cpu(p->header.h_length);
 763         struct midcomms_node *node;
 764         uint32_t seq;
 765         int ret, idx;
 766
 767         idx = srcu_read_lock(&nodes_srcu);
 768         node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
 769                                              dlm_midcomms_version_check_3_2);
 770         if (!node)
 771                 goto out;
 772
 773         switch (p->header.h_cmd) {
 774         case DLM_RCOM:
 775                 /* these rcom message we use to determine version.
 776                  * they have their own retransmission handling and
 777                  * are the first messages of dlm.
 778                  *
 779                  * length already checked.
 780                  */
 781                 switch (p->rcom.rc_type) {
 782                 case cpu_to_le32(DLM_RCOM_NAMES):
 783                         fallthrough;
 784                 case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
 785                         fallthrough;
 786                 case cpu_to_le32(DLM_RCOM_STATUS):
 787                         fallthrough;
 788                 case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
 789                         break;
 790                 default:
 791                         log_print("unsupported rcom type received: %u, will skip this message from node %d",
 792                                   le32_to_cpu(p->rcom.rc_type), nodeid);
 793                         goto out;
 794                 }
 795
 796                 WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
 797                 dlm_receive_buffer(p, nodeid);
 798                 break;
 799         case DLM_OPTS:
 800                 seq = le32_to_cpu(p->header.u.h_seq);
 801
 802                 ret = dlm_opts_check_msglen(p, msglen, nodeid);
 803                 if (ret < 0) {
 804                         log_print("opts msg too small: %u, will skip this message from node %d",
 805                                   msglen, nodeid);
 806                         goto out;
 807                 }
 808
 809                 p = (union dlm_packet *)((unsigned char *)p->opts.o_opts +
 810                                          le16_to_cpu(p->opts.o_optlen));
 811
 812                 /* recheck inner msglen just if it's not garbage */
 813                 msglen = le16_to_cpu(p->header.h_length);
 814                 switch (p->header.h_cmd) {
 815                 case DLM_RCOM:
 816                         if (msglen < sizeof(struct dlm_rcom)) {
 817                                 log_print("inner rcom msg too small: %u, will skip this message from node %d",
 818                                           msglen, nodeid);
 819                                 goto out;
 820                         }
 821
 822                         break;
 823                 case DLM_MSG:
 824                         if (msglen < sizeof(struct dlm_message)) {
 825                                 log_print("inner msg too small: %u, will skip this message from node %d",
 826                                           msglen, nodeid);
 827                                 goto out;
 828                         }
 829
 830                         break;
 831                 case DLM_FIN:
 832                         if (msglen < sizeof(struct dlm_header)) {
 833                                 log_print("inner fin too small: %u, will skip this message from node %d",
 834                                           msglen, nodeid);
 835                                 goto out;
 836                         }
 837
 838                         break;
 839                 default:
 840                         log_print("unsupported inner h_cmd received: %u, will skip this message from node %d",
 841                                   msglen, nodeid);
 842                         goto out;
 843                 }
 844
 845                 dlm_midcomms_receive_buffer(p, node, seq);
 846                 break;
 847         case DLM_ACK:
 848                 seq = le32_to_cpu(p->header.u.h_seq);
 849                 dlm_receive_ack(node, seq);
 850                 break;
 851         default:
 852                 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
 853                           p->header.h_cmd, nodeid);
 854                 break;
 855         }
 856
 857 out:
 858         srcu_read_unlock(&nodes_srcu, idx);
 859 }
 860
 861 static int dlm_midcomms_version_check_3_1(struct midcomms_node *node)
 862 {
 863         switch (node->version) {
 864         case DLM_VERSION_NOT_SET:
 865                 node->version = DLM_VERSION_3_1;
 866                 wake_up(&node->shutdown_wait);
 867                 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_1,
 868                           node->nodeid);
 869                 break;
 870         case DLM_VERSION_3_1:
 871                 break;
 872         default:
 873                 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
 874                                       DLM_VERSION_3_1, node->nodeid, node->version);
 875                 return -1;
 876         }
 877
 878         return 0;
 879 }
 880
 881 static void dlm_midcomms_receive_buffer_3_1(union dlm_packet *p, int nodeid)
 882 {
 883         uint16_t msglen = le16_to_cpu(p->header.h_length);
 884         struct midcomms_node *node;
 885         int idx;
 886
 887         idx = srcu_read_lock(&nodes_srcu);
 888         node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
 889                                              dlm_midcomms_version_check_3_1);
 890         if (!node) {
 891                 srcu_read_unlock(&nodes_srcu, idx);
 892                 return;
 893         }
 894         srcu_read_unlock(&nodes_srcu, idx);
 895
 896         switch (p->header.h_cmd) {
 897         case DLM_RCOM:
 898                 /* length already checked */
 899                 break;
 900         case DLM_MSG:
 901                 if (msglen < sizeof(struct dlm_message)) {
 902                         log_print("msg too small: %u, will skip this message from node %d",
 903                                   msglen, nodeid);
 904                         return;
 905                 }
 906
 907                 break;
 908         default:
 909                 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
 910                           p->header.h_cmd, nodeid);
 911                 return;
 912         }
 913
 914         dlm_receive_buffer(p, nodeid);
 915 }
 916
 917 int dlm_validate_incoming_buffer(int nodeid, unsigned char *buf, int len)
 918 {
 919         const unsigned char *ptr = buf;
 920         const struct dlm_header *hd;
 921         uint16_t msglen;
 922         int ret = 0;
 923
 924         while (len >= sizeof(struct dlm_header)) {
 925                 hd = (struct dlm_header *)ptr;
 926
 927                 /* no message should be more than DLM_MAX_SOCKET_BUFSIZE or
 928                  * less than dlm_header size.
 929                  *
 930                  * Some messages does not have a 8 byte length boundary yet
 931                  * which can occur in a unaligned memory access of some dlm
 932                  * messages. However this problem need to be fixed at the
 933                  * sending side, for now it seems nobody run into architecture
 934                  * related issues yet but it slows down some processing.
 935                  * Fixing this issue should be scheduled in future by doing
 936                  * the next major version bump.
 937                  */
 938                 msglen = le16_to_cpu(hd->h_length);
 939                 if (msglen > DLM_MAX_SOCKET_BUFSIZE ||
 940                     msglen < sizeof(struct dlm_header)) {
 941                         log_print("received invalid length header: %u from node %d, will abort message parsing",
 942                                   msglen, nodeid);
 943                         return -EBADMSG;
 944                 }
 945
 946                 /* caller will take care that leftover
 947                  * will be parsed next call with more data
 948                  */
 949                 if (msglen > len)
 950                         break;
 951
 952                 ret += msglen;
 953                 len -= msglen;
 954                 ptr += msglen;
 955         }
 956
 957         return ret;
 958 }
 959
 960 /*
 961  * Called from the low-level comms layer to process a buffer of
 962  * commands.
 963  */
 964 int dlm_process_incoming_buffer(int nodeid, unsigned char *buf, int len)
 965 {
 966         const unsigned char *ptr = buf;
 967         const struct dlm_header *hd;
 968         uint16_t msglen;
 969         int ret = 0;
 970
 971         while (len >= sizeof(struct dlm_header)) {
 972                 hd = (struct dlm_header *)ptr;
 973
 974                 msglen = le16_to_cpu(hd->h_length);
 975                 if (msglen > len)
 976                         break;
 977
 978                 switch (hd->h_version) {
 979                 case cpu_to_le32(DLM_VERSION_3_1):
 980                         dlm_midcomms_receive_buffer_3_1((union dlm_packet *)ptr, nodeid);
 981                         break;
 982                 case cpu_to_le32(DLM_VERSION_3_2):
 983                         dlm_midcomms_receive_buffer_3_2((union dlm_packet *)ptr, nodeid);
 984                         break;
 985                 default:
 986                         log_print("received invalid version header: %u from node %d, will skip this message",
 987                                   le32_to_cpu(hd->h_version), nodeid);
 988                         break;
 989                 }
 990
 991                 ret += msglen;
 992                 len -= msglen;
 993                 ptr += msglen;
 994         }
 995
 996         return ret;
 997 }
 998
 999 void dlm_midcomms_unack_msg_resend(int nodeid)
1000 {
1001         struct midcomms_node *node;
1002         struct dlm_mhandle *mh;
1003         int idx, ret;
1004
1005         idx = srcu_read_lock(&nodes_srcu);
1006         node = nodeid2node(nodeid, 0);
1007         if (!node) {
1008                 srcu_read_unlock(&nodes_srcu, idx);
1009                 return;
1010         }
1011
1012         /* old protocol, we don't support to retransmit on failure */
1013         switch (node->version) {
1014         case DLM_VERSION_3_2:
1015                 break;
1016         default:
1017                 srcu_read_unlock(&nodes_srcu, idx);
1018                 return;
1019         }
1020
1021         rcu_read_lock();
1022         list_for_each_entry_rcu(mh, &node->send_queue, list) {
1023                 if (!mh->committed)
1024                         continue;
1025
1026                 ret = dlm_lowcomms_resend_msg(mh->msg);
1027                 if (!ret)
1028                         log_print_ratelimited("retransmit dlm msg, seq %u, nodeid %d",
1029                                               mh->seq, node->nodeid);
1030         }
1031         rcu_read_unlock();
1032         srcu_read_unlock(&nodes_srcu, idx);
1033 }
1034
1035 static void dlm_fill_opts_header(struct dlm_opts *opts, uint16_t inner_len,
1036                                  uint32_t seq)
1037 {
1038         opts->o_header.h_cmd = DLM_OPTS;
1039         opts->o_header.h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
1040         opts->o_header.h_nodeid = cpu_to_le32(dlm_our_nodeid());
1041         opts->o_header.h_length = cpu_to_le16(DLM_MIDCOMMS_OPT_LEN + inner_len);
1042         opts->o_header.u.h_seq = cpu_to_le32(seq);
1043 }
1044
1045 static void midcomms_new_msg_cb(void *data)
1046 {
1047         struct dlm_mhandle *mh = data;
1048
1049         atomic_inc(&mh->node->send_queue_cnt);
1050
1051         spin_lock_bh(&mh->node->send_queue_lock);
1052         list_add_tail_rcu(&mh->list, &mh->node->send_queue);
1053         spin_unlock_bh(&mh->node->send_queue_lock);
1054
1055         mh->seq = atomic_fetch_inc(&mh->node->seq_send);
1056 }
1057
1058 static struct dlm_msg *dlm_midcomms_get_msg_3_2(struct dlm_mhandle *mh, int nodeid,
1059                                                 int len, gfp_t allocation, char **ppc)
1060 {
1061         struct dlm_opts *opts;
1062         struct dlm_msg *msg;
1063
1064         msg = dlm_lowcomms_new_msg(nodeid, len + DLM_MIDCOMMS_OPT_LEN,
1065                                    allocation, ppc, midcomms_new_msg_cb, mh);
1066         if (!msg)
1067                 return NULL;
1068
1069         opts = (struct dlm_opts *)*ppc;
1070         mh->opts = opts;
1071
1072         /* add possible options here */
1073         dlm_fill_opts_header(opts, len, mh->seq);
1074
1075         *ppc += sizeof(*opts);
1076         mh->inner_p = (const union dlm_packet *)*ppc;
1077         return msg;
1078 }
1079
1080 /* avoid false positive for nodes_srcu, unlock happens in
1081  * dlm_midcomms_commit_mhandle which is a must call if success
1082  */
1083 #ifndef __CHECKER__
1084 struct dlm_mhandle *dlm_midcomms_get_mhandle(int nodeid, int len,
1085                                              gfp_t allocation, char **ppc)
1086 {
1087         struct midcomms_node *node;
1088         struct dlm_mhandle *mh;
1089         struct dlm_msg *msg;
1090         int idx;
1091
1092         idx = srcu_read_lock(&nodes_srcu);
1093         node = nodeid2node(nodeid, 0);
1094         if (!node) {
1095                 WARN_ON_ONCE(1);
1096                 goto err;
1097         }
1098
1099         /* this is a bug, however we going on and hope it will be resolved */
1100         WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_TX, &node->flags));
1101
1102         mh = dlm_allocate_mhandle(allocation);
1103         if (!mh)
1104                 goto err;
1105
1106         mh->committed = false;
1107         mh->ack_rcv = NULL;
1108         mh->idx = idx;
1109         mh->node = node;
1110
1111         switch (node->version) {
1112         case DLM_VERSION_3_1:
1113                 msg = dlm_lowcomms_new_msg(nodeid, len, allocation, ppc,
1114                                            NULL, NULL);
1115                 if (!msg) {
1116                         dlm_free_mhandle(mh);
1117                         goto err;
1118                 }
1119
1120                 break;
1121         case DLM_VERSION_3_2:
1122                 msg = dlm_midcomms_get_msg_3_2(mh, nodeid, len, allocation,
1123                                                ppc);
1124                 if (!msg) {
1125                         dlm_free_mhandle(mh);
1126                         goto err;
1127                 }
1128
1129                 /* send ack back if necessary */
1130                 dlm_send_ack_threshold(node, DLM_SEND_ACK_BACK_MSG_THRESHOLD);
1131                 break;
1132         default:
1133                 dlm_free_mhandle(mh);
1134                 WARN_ON_ONCE(1);
1135                 goto err;
1136         }
1137
1138         mh->msg = msg;
1139
1140         /* keep in mind that is a must to call
1141          * dlm_midcomms_commit_msg() which releases
1142          * nodes_srcu using mh->idx which is assumed
1143          * here that the application will call it.
1144          */
1145         return mh;
1146
1147 err:
1148         srcu_read_unlock(&nodes_srcu, idx);
1149         return NULL;
1150 }
1151 #endif
1152
1153 static void dlm_midcomms_commit_msg_3_2_trace(const struct dlm_mhandle *mh,
1154                                               const void *name, int namelen)
1155 {
1156         switch (mh->inner_p->header.h_cmd) {
1157         case DLM_MSG:
1158                 trace_dlm_send_message(mh->node->nodeid, mh->seq,
1159                                        &mh->inner_p->message,
1160                                        name, namelen);
1161                 break;
1162         case DLM_RCOM:
1163                 trace_dlm_send_rcom(mh->node->nodeid, mh->seq,
1164                                     &mh->inner_p->rcom);
1165                 break;
1166         default:
1167                 /* nothing to trace */
1168                 break;
1169         }
1170 }
1171
1172 static void dlm_midcomms_commit_msg_3_2(struct dlm_mhandle *mh,
1173                                         const void *name, int namelen)
1174 {
1175         /* nexthdr chain for fast lookup */
1176         mh->opts->o_nextcmd = mh->inner_p->header.h_cmd;
1177         mh->committed = true;
1178         dlm_midcomms_commit_msg_3_2_trace(mh, name, namelen);
1179         dlm_lowcomms_commit_msg(mh->msg);
1180 }
1181
1182 /* avoid false positive for nodes_srcu, lock was happen in
1183  * dlm_midcomms_get_mhandle
1184  */
1185 #ifndef __CHECKER__
1186 void dlm_midcomms_commit_mhandle(struct dlm_mhandle *mh,
1187                                  const void *name, int namelen)
1188 {
1189
1190         switch (mh->node->version) {
1191         case DLM_VERSION_3_1:
1192                 srcu_read_unlock(&nodes_srcu, mh->idx);
1193
1194                 dlm_lowcomms_commit_msg(mh->msg);
1195                 dlm_lowcomms_put_msg(mh->msg);
1196                 /* mh is not part of rcu list in this case */
1197                 dlm_free_mhandle(mh);
1198                 break;
1199         case DLM_VERSION_3_2:
1200                 /* held rcu read lock here, because we sending the
1201                  * dlm message out, when we do that we could receive
1202                  * an ack back which releases the mhandle and we
1203                  * get a use after free.
1204                  */
1205                 rcu_read_lock();
1206                 dlm_midcomms_commit_msg_3_2(mh, name, namelen);
1207                 srcu_read_unlock(&nodes_srcu, mh->idx);
1208                 rcu_read_unlock();
1209                 break;
1210         default:
1211                 srcu_read_unlock(&nodes_srcu, mh->idx);
1212                 WARN_ON_ONCE(1);
1213                 break;
1214         }
1215 }
1216 #endif
1217
1218 int dlm_midcomms_start(void)
1219 {
1220         return dlm_lowcomms_start();
1221 }
1222
1223 void dlm_midcomms_stop(void)
1224 {
1225         dlm_lowcomms_stop();
1226 }
1227
1228 void dlm_midcomms_init(void)
1229 {
1230         int i;
1231
1232         for (i = 0; i < CONN_HASH_SIZE; i++)
1233                 INIT_HLIST_HEAD(&node_hash[i]);
1234
1235         dlm_lowcomms_init();
1236 }
1237
1238 void dlm_midcomms_exit(void)
1239 {
1240         dlm_lowcomms_exit();
1241 }
1242
1243 static void dlm_act_fin_ack_rcv(struct midcomms_node *node)
1244 {
1245         spin_lock(&node->state_lock);
1246         pr_debug("receive active fin ack from node %d with state %s\n",
1247                  node->nodeid, dlm_state_str(node->state));
1248
1249         switch (node->state) {
1250         case DLM_FIN_WAIT1:
1251                 node->state = DLM_FIN_WAIT2;
1252                 pr_debug("switch node %d to state %s\n",
1253                          node->nodeid, dlm_state_str(node->state));
1254                 break;
1255         case DLM_CLOSING:
1256                 midcomms_node_reset(node);
1257                 pr_debug("switch node %d to state %s\n",
1258                          node->nodeid, dlm_state_str(node->state));
1259                 break;
1260         case DLM_CLOSED:
1261                 /* not valid but somehow we got what we want */
1262                 wake_up(&node->shutdown_wait);
1263                 break;
1264         default:
1265                 spin_unlock(&node->state_lock);
1266                 log_print("%s: unexpected state: %d",
1267                           __func__, node->state);
1268                 WARN_ON_ONCE(1);
1269                 return;
1270         }
1271         spin_unlock(&node->state_lock);
1272 }
1273
1274 void dlm_midcomms_add_member(int nodeid)
1275 {
1276         struct midcomms_node *node;
1277         int idx;
1278
1279         idx = srcu_read_lock(&nodes_srcu);
1280         node = nodeid2node(nodeid, GFP_NOFS);
1281         if (!node) {
1282                 srcu_read_unlock(&nodes_srcu, idx);
1283                 return;
1284         }
1285
1286         spin_lock(&node->state_lock);
1287         if (!node->users) {
1288                 pr_debug("receive add member from node %d with state %s\n",
1289                          node->nodeid, dlm_state_str(node->state));
1290                 switch (node->state) {
1291                 case DLM_ESTABLISHED:
1292                         break;
1293                 case DLM_CLOSED:
1294                         node->state = DLM_ESTABLISHED;
1295                         pr_debug("switch node %d to state %s\n",
1296                                  node->nodeid, dlm_state_str(node->state));
1297                         break;
1298                 default:
1299                         /* some invalid state passive shutdown
1300                          * was failed, we try to reset and
1301                          * hope it will go on.
1302                          */
1303                         log_print("reset node %d because shutdown stuck",
1304                                   node->nodeid);
1305
1306                         midcomms_node_reset(node);
1307                         node->state = DLM_ESTABLISHED;
1308                         break;
1309                 }
1310         }
1311
1312         node->users++;
1313         pr_debug("node %d users inc count %d\n", nodeid, node->users);
1314         spin_unlock(&node->state_lock);
1315
1316         srcu_read_unlock(&nodes_srcu, idx);
1317 }
1318
1319 void dlm_midcomms_remove_member(int nodeid)
1320 {
1321         struct midcomms_node *node;
1322         int idx;
1323
1324         idx = srcu_read_lock(&nodes_srcu);
1325         node = nodeid2node(nodeid, 0);
1326         if (!node) {
1327                 srcu_read_unlock(&nodes_srcu, idx);
1328                 return;
1329         }
1330
1331         spin_lock(&node->state_lock);
1332         node->users--;
1333         pr_debug("node %d users dec count %d\n", nodeid, node->users);
1334
1335         /* hitting users count to zero means the
1336          * other side is running dlm_midcomms_stop()
1337          * we meet us to have a clean disconnect.
1338          */
1339         if (node->users == 0) {
1340                 pr_debug("receive remove member from node %d with state %s\n",
1341                          node->nodeid, dlm_state_str(node->state));
1342                 switch (node->state) {
1343                 case DLM_ESTABLISHED:
1344                         break;
1345                 case DLM_CLOSE_WAIT:
1346                         /* passive shutdown DLM_LAST_ACK case 2 */
1347                         node->state = DLM_LAST_ACK;
1348                         pr_debug("switch node %d to state %s case 2\n",
1349                                  node->nodeid, dlm_state_str(node->state));
1350                         set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
1351                         dlm_send_fin(node, dlm_pas_fin_ack_rcv);
1352                         break;
1353                 case DLM_LAST_ACK:
1354                         /* probably receive fin caught it, do nothing */
1355                         break;
1356                 case DLM_CLOSED:
1357                         /* already gone, do nothing */
1358                         break;
1359                 default:
1360                         log_print("%s: unexpected state: %d",
1361                                   __func__, node->state);
1362                         break;
1363                 }
1364         }
1365         spin_unlock(&node->state_lock);
1366
1367         srcu_read_unlock(&nodes_srcu, idx);
1368 }
1369
1370 static void midcomms_node_release(struct rcu_head *rcu)
1371 {
1372         struct midcomms_node *node = container_of(rcu, struct midcomms_node, rcu);
1373
1374         WARN_ON_ONCE(atomic_read(&node->send_queue_cnt));
1375         dlm_send_queue_flush(node);
1376         kfree(node);
1377 }
1378
1379 void dlm_midcomms_version_wait(void)
1380 {
1381         struct midcomms_node *node;
1382         int i, idx, ret;
1383
1384         idx = srcu_read_lock(&nodes_srcu);
1385         for (i = 0; i < CONN_HASH_SIZE; i++) {
1386                 hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1387                         ret = wait_event_timeout(node->shutdown_wait,
1388                                                  node->version != DLM_VERSION_NOT_SET ||
1389                                                  node->state == DLM_CLOSED ||
1390                                                  test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1391                                                  DLM_SHUTDOWN_TIMEOUT);
1392                         if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
1393                                 pr_debug("version wait timed out for node %d with state %s\n",
1394                                          node->nodeid, dlm_state_str(node->state));
1395                 }
1396         }
1397         srcu_read_unlock(&nodes_srcu, idx);
1398 }
1399
1400 static void midcomms_shutdown(struct midcomms_node *node)
1401 {
1402         int ret;
1403
1404         /* old protocol, we don't wait for pending operations */
1405         switch (node->version) {
1406         case DLM_VERSION_3_2:
1407                 break;
1408         default:
1409                 return;
1410         }
1411
1412         spin_lock(&node->state_lock);
1413         pr_debug("receive active shutdown for node %d with state %s\n",
1414                  node->nodeid, dlm_state_str(node->state));
1415         switch (node->state) {
1416         case DLM_ESTABLISHED:
1417                 node->state = DLM_FIN_WAIT1;
1418                 pr_debug("switch node %d to state %s case 2\n",
1419                          node->nodeid, dlm_state_str(node->state));
1420                 dlm_send_fin(node, dlm_act_fin_ack_rcv);
1421                 break;
1422         case DLM_CLOSED:
1423                 /* we have what we want */
1424                 break;
1425         default:
1426                 /* busy to enter DLM_FIN_WAIT1, wait until passive
1427                  * done in shutdown_wait to enter DLM_CLOSED.
1428                  */
1429                 break;
1430         }
1431         spin_unlock(&node->state_lock);
1432
1433         if (DLM_DEBUG_FENCE_TERMINATION)
1434                 msleep(5000);
1435
1436         /* wait for other side dlm + fin */
1437         ret = wait_event_timeout(node->shutdown_wait,
1438                                  node->state == DLM_CLOSED ||
1439                                  test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1440                                  DLM_SHUTDOWN_TIMEOUT);
1441         if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
1442                 pr_debug("active shutdown timed out for node %d with state %s\n",
1443                          node->nodeid, dlm_state_str(node->state));
1444         else
1445                 pr_debug("active shutdown done for node %d with state %s\n",
1446                          node->nodeid, dlm_state_str(node->state));
1447 }
1448
1449 void dlm_midcomms_shutdown(void)
1450 {
1451         struct midcomms_node *node;
1452         int i, idx;
1453
1454         mutex_lock(&close_lock);
1455         idx = srcu_read_lock(&nodes_srcu);
1456         for (i = 0; i < CONN_HASH_SIZE; i++) {
1457                 hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1458                         midcomms_shutdown(node);
1459
1460                         dlm_delete_debug_comms_file(node->debugfs);
1461
1462                         spin_lock(&nodes_lock);
1463                         hlist_del_rcu(&node->hlist);
1464                         spin_unlock(&nodes_lock);
1465
1466                         call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
1467                 }
1468         }
1469         srcu_read_unlock(&nodes_srcu, idx);
1470         mutex_unlock(&close_lock);
1471
1472         dlm_lowcomms_shutdown();
1473 }
1474
1475 int dlm_midcomms_close(int nodeid)
1476 {
1477         struct midcomms_node *node;
1478         int idx, ret;
1479
1480         idx = srcu_read_lock(&nodes_srcu);
1481         /* Abort pending close/remove operation */
1482         node = nodeid2node(nodeid, 0);
1483         if (node) {
1484                 /* let shutdown waiters leave */
1485                 set_bit(DLM_NODE_FLAG_CLOSE, &node->flags);
1486                 wake_up(&node->shutdown_wait);
1487         }
1488         srcu_read_unlock(&nodes_srcu, idx);
1489
1490         synchronize_srcu(&nodes_srcu);
1491
1492         idx = srcu_read_lock(&nodes_srcu);
1493         mutex_lock(&close_lock);
1494         node = nodeid2node(nodeid, 0);
1495         if (!node) {
1496                 mutex_unlock(&close_lock);
1497                 srcu_read_unlock(&nodes_srcu, idx);
1498                 return dlm_lowcomms_close(nodeid);
1499         }
1500
1501         ret = dlm_lowcomms_close(nodeid);
1502         spin_lock(&node->state_lock);
1503         midcomms_node_reset(node);
1504         spin_unlock(&node->state_lock);
1505         srcu_read_unlock(&nodes_srcu, idx);
1506         mutex_unlock(&close_lock);
1507
1508         return ret;
1509 }
1510
1511 /* debug functionality to send raw dlm msg from user space */
1512 struct dlm_rawmsg_data {
1513         struct midcomms_node *node;
1514         void *buf;
1515 };
1516
1517 static void midcomms_new_rawmsg_cb(void *data)
1518 {
1519         struct dlm_rawmsg_data *rd = data;
1520         struct dlm_header *h = rd->buf;
1521
1522         switch (h->h_version) {
1523         case cpu_to_le32(DLM_VERSION_3_1):
1524                 break;
1525         default:
1526                 switch (h->h_cmd) {
1527                 case DLM_OPTS:
1528                         if (!h->u.h_seq)
1529                                 h->u.h_seq = cpu_to_le32(atomic_fetch_inc(&rd->node->seq_send));
1530                         break;
1531                 default:
1532                         break;
1533                 }
1534                 break;
1535         }
1536 }
1537
1538 int dlm_midcomms_rawmsg_send(struct midcomms_node *node, void *buf,
1539                              int buflen)
1540 {
1541         struct dlm_rawmsg_data rd;
1542         struct dlm_msg *msg;
1543         char *msgbuf;
1544
1545         rd.node = node;
1546         rd.buf = buf;
1547
1548         msg = dlm_lowcomms_new_msg(node->nodeid, buflen, GFP_NOFS,
1549                                    &msgbuf, midcomms_new_rawmsg_cb, &rd);
1550         if (!msg)
1551                 return -ENOMEM;
1552
1553         memcpy(msgbuf, buf, buflen);
1554         dlm_lowcomms_commit_msg(msg);
1555         return 0;
1556 }
1557