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 "lockspace.h"
 140 #include "lowcomms.h"
 141 #include "config.h"
 142 #include "memory.h"
 143 #include "lock.h"
 144 #include "util.h"
 145 #include "midcomms.h"
 146
 147 /* init value for sequence numbers for testing purpose only e.g. overflows */
 148 #define DLM_SEQ_INIT            0
 149 /* 5 seconds wait to sync ending of dlm */
 150 #define DLM_SHUTDOWN_TIMEOUT    msecs_to_jiffies(5000)
 151 #define DLM_VERSION_NOT_SET     0
 152
 153 struct midcomms_node {
 154         int nodeid;
 155         uint32_t version;
 156         uint32_t seq_send;
 157         uint32_t seq_next;
 158         /* These queues are unbound because we cannot drop any message in dlm.
 159          * We could send a fence signal for a specific node to the cluster
 160          * manager if queues hits some maximum value, however this handling
 161          * not supported yet.
 162          */
 163         struct list_head send_queue;
 164         spinlock_t send_queue_lock;
 165         atomic_t send_queue_cnt;
 166 #define DLM_NODE_FLAG_CLOSE     1
 167 #define DLM_NODE_FLAG_STOP_TX   2
 168 #define DLM_NODE_FLAG_STOP_RX   3
 169 #define DLM_NODE_ULP_DELIVERED  4
 170         unsigned long flags;
 171         wait_queue_head_t shutdown_wait;
 172
 173         /* dlm tcp termination state */
 174 #define DLM_CLOSED      1
 175 #define DLM_ESTABLISHED 2
 176 #define DLM_FIN_WAIT1   3
 177 #define DLM_FIN_WAIT2   4
 178 #define DLM_CLOSE_WAIT  5
 179 #define DLM_LAST_ACK    6
 180 #define DLM_CLOSING     7
 181         int state;
 182         spinlock_t state_lock;
 183
 184         /* counts how many lockspaces are using this node
 185          * this refcount is necessary to determine if the
 186          * node wants to disconnect.
 187          */
 188         int users;
 189
 190         /* not protected by srcu, node_hash lifetime */
 191         void *debugfs;
 192
 193         struct hlist_node hlist;
 194         struct rcu_head rcu;
 195 };
 196
 197 struct dlm_mhandle {
 198         const union dlm_packet *inner_p;
 199         struct midcomms_node *node;
 200         struct dlm_opts *opts;
 201         struct dlm_msg *msg;
 202         bool committed;
 203         uint32_t seq;
 204
 205         void (*ack_rcv)(struct midcomms_node *node);
 206
 207         /* get_mhandle/commit srcu idx exchange */
 208         int idx;
 209
 210         struct list_head list;
 211         struct rcu_head rcu;
 212 };
 213
 214 static struct hlist_head node_hash[CONN_HASH_SIZE];
 215 static DEFINE_SPINLOCK(nodes_lock);
 216 DEFINE_STATIC_SRCU(nodes_srcu);
 217
 218 /* This mutex prevents that midcomms_close() is running while
 219  * stop() or remove(). As I experienced invalid memory access
 220  * behaviours when DLM_DEBUG_FENCE_TERMINATION is enabled and
 221  * resetting machines. I will end in some double deletion in nodes
 222  * datastructure.
 223  */
 224 static DEFINE_MUTEX(close_lock);
 225
 226 struct kmem_cache *dlm_midcomms_cache_create(void)
 227 {
 228         return kmem_cache_create("dlm_mhandle", sizeof(struct dlm_mhandle),
 229                                  0, 0, NULL);
 230 }
 231
 232 static inline const char *dlm_state_str(int state)
 233 {
 234         switch (state) {
 235         case DLM_CLOSED:
 236                 return "CLOSED";
 237         case DLM_ESTABLISHED:
 238                 return "ESTABLISHED";
 239         case DLM_FIN_WAIT1:
 240                 return "FIN_WAIT1";
 241         case DLM_FIN_WAIT2:
 242                 return "FIN_WAIT2";
 243         case DLM_CLOSE_WAIT:
 244                 return "CLOSE_WAIT";
 245         case DLM_LAST_ACK:
 246                 return "LAST_ACK";
 247         case DLM_CLOSING:
 248                 return "CLOSING";
 249         default:
 250                 return "UNKNOWN";
 251         }
 252 }
 253
 254 const char *dlm_midcomms_state(struct midcomms_node *node)
 255 {
 256         return dlm_state_str(node->state);
 257 }
 258
 259 unsigned long dlm_midcomms_flags(struct midcomms_node *node)
 260 {
 261         return node->flags;
 262 }
 263
 264 int dlm_midcomms_send_queue_cnt(struct midcomms_node *node)
 265 {
 266         return atomic_read(&node->send_queue_cnt);
 267 }
 268
 269 uint32_t dlm_midcomms_version(struct midcomms_node *node)
 270 {
 271         return node->version;
 272 }
 273
 274 static struct midcomms_node *__find_node(int nodeid, int r)
 275 {
 276         struct midcomms_node *node;
 277
 278         hlist_for_each_entry_rcu(node, &node_hash[r], hlist) {
 279                 if (node->nodeid == nodeid)
 280                         return node;
 281         }
 282
 283         return NULL;
 284 }
 285
 286 static void dlm_mhandle_release(struct rcu_head *rcu)
 287 {
 288         struct dlm_mhandle *mh = container_of(rcu, struct dlm_mhandle, rcu);
 289
 290         dlm_lowcomms_put_msg(mh->msg);
 291         dlm_free_mhandle(mh);
 292 }
 293
 294 static void dlm_mhandle_delete(struct midcomms_node *node,
 295                                struct dlm_mhandle *mh)
 296 {
 297         list_del_rcu(&mh->list);
 298         atomic_dec(&node->send_queue_cnt);
 299         call_rcu(&mh->rcu, dlm_mhandle_release);
 300 }
 301
 302 static void dlm_send_queue_flush(struct midcomms_node *node)
 303 {
 304         struct dlm_mhandle *mh;
 305
 306         pr_debug("flush midcomms send queue of node %d\n", node->nodeid);
 307
 308         rcu_read_lock();
 309         spin_lock_bh(&node->send_queue_lock);
 310         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 311                 dlm_mhandle_delete(node, mh);
 312         }
 313         spin_unlock_bh(&node->send_queue_lock);
 314         rcu_read_unlock();
 315 }
 316
 317 static void midcomms_node_reset(struct midcomms_node *node)
 318 {
 319         pr_debug("reset node %d\n", node->nodeid);
 320
 321         node->seq_next = DLM_SEQ_INIT;
 322         node->seq_send = DLM_SEQ_INIT;
 323         node->version = DLM_VERSION_NOT_SET;
 324         node->flags = 0;
 325
 326         dlm_send_queue_flush(node);
 327         node->state = DLM_CLOSED;
 328         wake_up(&node->shutdown_wait);
 329 }
 330
 331 static struct midcomms_node *nodeid2node(int nodeid, gfp_t alloc)
 332 {
 333         struct midcomms_node *node, *tmp;
 334         int r = nodeid_hash(nodeid);
 335
 336         node = __find_node(nodeid, r);
 337         if (node || !alloc)
 338                 return node;
 339
 340         node = kmalloc(sizeof(*node), alloc);
 341         if (!node)
 342                 return NULL;
 343
 344         node->nodeid = nodeid;
 345         spin_lock_init(&node->state_lock);
 346         spin_lock_init(&node->send_queue_lock);
 347         atomic_set(&node->send_queue_cnt, 0);
 348         INIT_LIST_HEAD(&node->send_queue);
 349         init_waitqueue_head(&node->shutdown_wait);
 350         node->users = 0;
 351         midcomms_node_reset(node);
 352
 353         spin_lock(&nodes_lock);
 354         /* check again if there was somebody else
 355          * earlier here to add the node
 356          */
 357         tmp = __find_node(nodeid, r);
 358         if (tmp) {
 359                 spin_unlock(&nodes_lock);
 360                 kfree(node);
 361                 return tmp;
 362         }
 363
 364         hlist_add_head_rcu(&node->hlist, &node_hash[r]);
 365         spin_unlock(&nodes_lock);
 366
 367         node->debugfs = dlm_create_debug_comms_file(nodeid, node);
 368         return node;
 369 }
 370
 371 static int dlm_send_ack(int nodeid, uint32_t seq)
 372 {
 373         int mb_len = sizeof(struct dlm_header);
 374         struct dlm_header *m_header;
 375         struct dlm_msg *msg;
 376         char *ppc;
 377
 378         msg = dlm_lowcomms_new_msg(nodeid, mb_len, GFP_ATOMIC, &ppc,
 379                                    NULL, NULL);
 380         if (!msg)
 381                 return -ENOMEM;
 382
 383         m_header = (struct dlm_header *)ppc;
 384
 385         m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
 386         m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
 387         m_header->h_length = cpu_to_le16(mb_len);
 388         m_header->h_cmd = DLM_ACK;
 389         m_header->u.h_seq = cpu_to_le32(seq);
 390
 391         dlm_lowcomms_commit_msg(msg);
 392         dlm_lowcomms_put_msg(msg);
 393
 394         return 0;
 395 }
 396
 397 static int dlm_send_fin(struct midcomms_node *node,
 398                         void (*ack_rcv)(struct midcomms_node *node))
 399 {
 400         int mb_len = sizeof(struct dlm_header);
 401         struct dlm_header *m_header;
 402         struct dlm_mhandle *mh;
 403         char *ppc;
 404
 405         mh = dlm_midcomms_get_mhandle(node->nodeid, mb_len, GFP_ATOMIC, &ppc);
 406         if (!mh)
 407                 return -ENOMEM;
 408
 409         set_bit(DLM_NODE_FLAG_STOP_TX, &node->flags);
 410         mh->ack_rcv = ack_rcv;
 411
 412         m_header = (struct dlm_header *)ppc;
 413
 414         m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
 415         m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
 416         m_header->h_length = cpu_to_le16(mb_len);
 417         m_header->h_cmd = DLM_FIN;
 418
 419         pr_debug("sending fin msg to node %d\n", node->nodeid);
 420         dlm_midcomms_commit_mhandle(mh, NULL, 0);
 421
 422         return 0;
 423 }
 424
 425 static void dlm_receive_ack(struct midcomms_node *node, uint32_t seq)
 426 {
 427         struct dlm_mhandle *mh;
 428
 429         rcu_read_lock();
 430         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 431                 if (before(mh->seq, seq)) {
 432                         if (mh->ack_rcv)
 433                                 mh->ack_rcv(node);
 434                 } else {
 435                         /* send queue should be ordered */
 436                         break;
 437                 }
 438         }
 439
 440         spin_lock_bh(&node->send_queue_lock);
 441         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 442                 if (before(mh->seq, seq)) {
 443                         dlm_mhandle_delete(node, mh);
 444                 } else {
 445                         /* send queue should be ordered */
 446                         break;
 447                 }
 448         }
 449         spin_unlock_bh(&node->send_queue_lock);
 450         rcu_read_unlock();
 451 }
 452
 453 static void dlm_pas_fin_ack_rcv(struct midcomms_node *node)
 454 {
 455         spin_lock(&node->state_lock);
 456         pr_debug("receive passive fin ack from node %d with state %s\n",
 457                  node->nodeid, dlm_state_str(node->state));
 458
 459         switch (node->state) {
 460         case DLM_LAST_ACK:
 461                 /* DLM_CLOSED */
 462                 midcomms_node_reset(node);
 463                 break;
 464         case DLM_CLOSED:
 465                 /* not valid but somehow we got what we want */
 466                 wake_up(&node->shutdown_wait);
 467                 break;
 468         default:
 469                 spin_unlock(&node->state_lock);
 470                 log_print("%s: unexpected state: %d",
 471                           __func__, node->state);
 472                 WARN_ON_ONCE(1);
 473                 return;
 474         }
 475         spin_unlock(&node->state_lock);
 476 }
 477
 478 static void dlm_receive_buffer_3_2_trace(uint32_t seq, union dlm_packet *p)
 479 {
 480         switch (p->header.h_cmd) {
 481         case DLM_MSG:
 482                 trace_dlm_recv_message(dlm_our_nodeid(), seq, &p->message);
 483                 break;
 484         case DLM_RCOM:
 485                 trace_dlm_recv_rcom(dlm_our_nodeid(), seq, &p->rcom);
 486                 break;
 487         default:
 488                 break;
 489         }
 490 }
 491
 492 static void dlm_midcomms_receive_buffer(union dlm_packet *p,
 493                                         struct midcomms_node *node,
 494                                         uint32_t seq)
 495 {
 496         if (seq == node->seq_next) {
 497                 node->seq_next++;
 498
 499                 switch (p->header.h_cmd) {
 500                 case DLM_FIN:
 501                         spin_lock(&node->state_lock);
 502                         pr_debug("receive fin msg from node %d with state %s\n",
 503                                  node->nodeid, dlm_state_str(node->state));
 504
 505                         switch (node->state) {
 506                         case DLM_ESTABLISHED:
 507                                 dlm_send_ack(node->nodeid, node->seq_next);
 508
 509                                 /* passive shutdown DLM_LAST_ACK case 1
 510                                  * additional we check if the node is used by
 511                                  * cluster manager events at all.
 512                                  */
 513                                 if (node->users == 0) {
 514                                         node->state = DLM_LAST_ACK;
 515                                         pr_debug("switch node %d to state %s case 1\n",
 516                                                  node->nodeid, dlm_state_str(node->state));
 517                                         set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
 518                                         dlm_send_fin(node, dlm_pas_fin_ack_rcv);
 519                                 } else {
 520                                         node->state = DLM_CLOSE_WAIT;
 521                                         pr_debug("switch node %d to state %s\n",
 522                                                  node->nodeid, dlm_state_str(node->state));
 523                                 }
 524                                 break;
 525                         case DLM_FIN_WAIT1:
 526                                 dlm_send_ack(node->nodeid, node->seq_next);
 527                                 node->state = DLM_CLOSING;
 528                                 set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
 529                                 pr_debug("switch node %d to state %s\n",
 530                                          node->nodeid, dlm_state_str(node->state));
 531                                 break;
 532                         case DLM_FIN_WAIT2:
 533                                 dlm_send_ack(node->nodeid, node->seq_next);
 534                                 midcomms_node_reset(node);
 535                                 pr_debug("switch node %d to state %s\n",
 536                                          node->nodeid, dlm_state_str(node->state));
 537                                 break;
 538                         case DLM_LAST_ACK:
 539                                 /* probably remove_member caught it, do nothing */
 540                                 break;
 541                         default:
 542                                 spin_unlock(&node->state_lock);
 543                                 log_print("%s: unexpected state: %d",
 544                                           __func__, node->state);
 545                                 WARN_ON_ONCE(1);
 546                                 return;
 547                         }
 548                         spin_unlock(&node->state_lock);
 549                         break;
 550                 default:
 551                         WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
 552                         dlm_receive_buffer_3_2_trace(seq, p);
 553                         dlm_receive_buffer(p, node->nodeid);
 554                         set_bit(DLM_NODE_ULP_DELIVERED, &node->flags);
 555                         break;
 556                 }
 557         } else {
 558                 /* retry to ack message which we already have by sending back
 559                  * current node->seq_next number as ack.
 560                  */
 561                 if (seq < node->seq_next)
 562                         dlm_send_ack(node->nodeid, node->seq_next);
 563
 564                 log_print_ratelimited("ignore dlm msg because seq mismatch, seq: %u, expected: %u, nodeid: %d",
 565                                       seq, node->seq_next, node->nodeid);
 566         }
 567 }
 568
 569 static struct midcomms_node *
 570 dlm_midcomms_recv_node_lookup(int nodeid, const union dlm_packet *p,
 571                               uint16_t msglen, int (*cb)(struct midcomms_node *node))
 572 {
 573         struct midcomms_node *node = NULL;
 574         gfp_t allocation = 0;
 575         int ret;
 576
 577         switch (p->header.h_cmd) {
 578         case DLM_RCOM:
 579                 if (msglen < sizeof(struct dlm_rcom)) {
 580                         log_print("rcom msg too small: %u, will skip this message from node %d",
 581                                   msglen, nodeid);
 582                         return NULL;
 583                 }
 584
 585                 switch (p->rcom.rc_type) {
 586                 case cpu_to_le32(DLM_RCOM_NAMES):
 587                         fallthrough;
 588                 case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
 589                         fallthrough;
 590                 case cpu_to_le32(DLM_RCOM_STATUS):
 591                         fallthrough;
 592                 case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
 593                         node = nodeid2node(nodeid, 0);
 594                         if (node) {
 595                                 spin_lock(&node->state_lock);
 596                                 if (node->state != DLM_ESTABLISHED)
 597                                         pr_debug("receive begin RCOM msg from node %d with state %s\n",
 598                                                  node->nodeid, dlm_state_str(node->state));
 599
 600                                 switch (node->state) {
 601                                 case DLM_CLOSED:
 602                                         node->state = DLM_ESTABLISHED;
 603                                         pr_debug("switch node %d to state %s\n",
 604                                                  node->nodeid, dlm_state_str(node->state));
 605                                         break;
 606                                 case DLM_ESTABLISHED:
 607                                         break;
 608                                 default:
 609                                         spin_unlock(&node->state_lock);
 610                                         return NULL;
 611                                 }
 612                                 spin_unlock(&node->state_lock);
 613                         }
 614
 615                         allocation = GFP_NOFS;
 616                         break;
 617                 default:
 618                         break;
 619                 }
 620
 621                 break;
 622         default:
 623                 break;
 624         }
 625
 626         node = nodeid2node(nodeid, allocation);
 627         if (!node) {
 628                 switch (p->header.h_cmd) {
 629                 case DLM_OPTS:
 630                         if (msglen < sizeof(struct dlm_opts)) {
 631                                 log_print("opts msg too small: %u, will skip this message from node %d",
 632                                           msglen, nodeid);
 633                                 return NULL;
 634                         }
 635
 636                         log_print_ratelimited("received dlm opts message nextcmd %d from node %d in an invalid sequence",
 637                                               p->opts.o_nextcmd, nodeid);
 638                         break;
 639                 default:
 640                         log_print_ratelimited("received dlm message cmd %d from node %d in an invalid sequence",
 641                                               p->header.h_cmd, nodeid);
 642                         break;
 643                 }
 644
 645                 return NULL;
 646         }
 647
 648         ret = cb(node);
 649         if (ret < 0)
 650                 return NULL;
 651
 652         return node;
 653 }
 654
 655 static int dlm_midcomms_version_check_3_2(struct midcomms_node *node)
 656 {
 657         switch (node->version) {
 658         case DLM_VERSION_NOT_SET:
 659                 node->version = DLM_VERSION_3_2;
 660                 wake_up(&node->shutdown_wait);
 661                 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_2,
 662                           node->nodeid);
 663                 break;
 664         case DLM_VERSION_3_2:
 665                 break;
 666         default:
 667                 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
 668                                       DLM_VERSION_3_2, node->nodeid, node->version);
 669                 return -1;
 670         }
 671
 672         return 0;
 673 }
 674
 675 static int dlm_opts_check_msglen(union dlm_packet *p, uint16_t msglen, int nodeid)
 676 {
 677         int len = msglen;
 678
 679         /* we only trust outer header msglen because
 680          * it's checked against receive buffer length.
 681          */
 682         if (len < sizeof(struct dlm_opts))
 683                 return -1;
 684         len -= sizeof(struct dlm_opts);
 685
 686         if (len < le16_to_cpu(p->opts.o_optlen))
 687                 return -1;
 688         len -= le16_to_cpu(p->opts.o_optlen);
 689
 690         switch (p->opts.o_nextcmd) {
 691         case DLM_FIN:
 692                 if (len < sizeof(struct dlm_header)) {
 693                         log_print("fin too small: %d, will skip this message from node %d",
 694                                   len, nodeid);
 695                         return -1;
 696                 }
 697
 698                 break;
 699         case DLM_MSG:
 700                 if (len < sizeof(struct dlm_message)) {
 701                         log_print("msg too small: %d, will skip this message from node %d",
 702                                   msglen, nodeid);
 703                         return -1;
 704                 }
 705
 706                 break;
 707         case DLM_RCOM:
 708                 if (len < sizeof(struct dlm_rcom)) {
 709                         log_print("rcom msg too small: %d, will skip this message from node %d",
 710                                   len, nodeid);
 711                         return -1;
 712                 }
 713
 714                 break;
 715         default:
 716                 log_print("unsupported o_nextcmd received: %u, will skip this message from node %d",
 717                           p->opts.o_nextcmd, nodeid);
 718                 return -1;
 719         }
 720
 721         return 0;
 722 }
 723
 724 static void dlm_midcomms_receive_buffer_3_2(union dlm_packet *p, int nodeid)
 725 {
 726         uint16_t msglen = le16_to_cpu(p->header.h_length);
 727         struct midcomms_node *node;
 728         uint32_t seq;
 729         int ret, idx;
 730
 731         idx = srcu_read_lock(&nodes_srcu);
 732         node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
 733                                              dlm_midcomms_version_check_3_2);
 734         if (!node)
 735                 goto out;
 736
 737         switch (p->header.h_cmd) {
 738         case DLM_RCOM:
 739                 /* these rcom message we use to determine version.
 740                  * they have their own retransmission handling and
 741                  * are the first messages of dlm.
 742                  *
 743                  * length already checked.
 744                  */
 745                 switch (p->rcom.rc_type) {
 746                 case cpu_to_le32(DLM_RCOM_NAMES):
 747                         fallthrough;
 748                 case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
 749                         fallthrough;
 750                 case cpu_to_le32(DLM_RCOM_STATUS):
 751                         fallthrough;
 752                 case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
 753                         break;
 754                 default:
 755                         log_print("unsupported rcom type received: %u, will skip this message from node %d",
 756                                   le32_to_cpu(p->rcom.rc_type), nodeid);
 757                         goto out;
 758                 }
 759
 760                 WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
 761                 dlm_receive_buffer(p, nodeid);
 762                 break;
 763         case DLM_OPTS:
 764                 seq = le32_to_cpu(p->header.u.h_seq);
 765
 766                 ret = dlm_opts_check_msglen(p, msglen, nodeid);
 767                 if (ret < 0) {
 768                         log_print("opts msg too small: %u, will skip this message from node %d",
 769                                   msglen, nodeid);
 770                         goto out;
 771                 }
 772
 773                 p = (union dlm_packet *)((unsigned char *)p->opts.o_opts +
 774                                          le16_to_cpu(p->opts.o_optlen));
 775
 776                 /* recheck inner msglen just if it's not garbage */
 777                 msglen = le16_to_cpu(p->header.h_length);
 778                 switch (p->header.h_cmd) {
 779                 case DLM_RCOM:
 780                         if (msglen < sizeof(struct dlm_rcom)) {
 781                                 log_print("inner rcom msg too small: %u, will skip this message from node %d",
 782                                           msglen, nodeid);
 783                                 goto out;
 784                         }
 785
 786                         break;
 787                 case DLM_MSG:
 788                         if (msglen < sizeof(struct dlm_message)) {
 789                                 log_print("inner msg too small: %u, will skip this message from node %d",
 790                                           msglen, nodeid);
 791                                 goto out;
 792                         }
 793
 794                         break;
 795                 case DLM_FIN:
 796                         if (msglen < sizeof(struct dlm_header)) {
 797                                 log_print("inner fin too small: %u, will skip this message from node %d",
 798                                           msglen, nodeid);
 799                                 goto out;
 800                         }
 801
 802                         break;
 803                 default:
 804                         log_print("unsupported inner h_cmd received: %u, will skip this message from node %d",
 805                                   msglen, nodeid);
 806                         goto out;
 807                 }
 808
 809                 dlm_midcomms_receive_buffer(p, node, seq);
 810                 break;
 811         case DLM_ACK:
 812                 seq = le32_to_cpu(p->header.u.h_seq);
 813                 dlm_receive_ack(node, seq);
 814                 break;
 815         default:
 816                 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
 817                           p->header.h_cmd, nodeid);
 818                 break;
 819         }
 820
 821 out:
 822         srcu_read_unlock(&nodes_srcu, idx);
 823 }
 824
 825 static int dlm_midcomms_version_check_3_1(struct midcomms_node *node)
 826 {
 827         switch (node->version) {
 828         case DLM_VERSION_NOT_SET:
 829                 node->version = DLM_VERSION_3_1;
 830                 wake_up(&node->shutdown_wait);
 831                 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_1,
 832                           node->nodeid);
 833                 break;
 834         case DLM_VERSION_3_1:
 835                 break;
 836         default:
 837                 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
 838                                       DLM_VERSION_3_1, node->nodeid, node->version);
 839                 return -1;
 840         }
 841
 842         return 0;
 843 }
 844
 845 static void dlm_midcomms_receive_buffer_3_1(union dlm_packet *p, int nodeid)
 846 {
 847         uint16_t msglen = le16_to_cpu(p->header.h_length);
 848         struct midcomms_node *node;
 849         int idx;
 850
 851         idx = srcu_read_lock(&nodes_srcu);
 852         node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
 853                                              dlm_midcomms_version_check_3_1);
 854         if (!node) {
 855                 srcu_read_unlock(&nodes_srcu, idx);
 856                 return;
 857         }
 858         srcu_read_unlock(&nodes_srcu, idx);
 859
 860         switch (p->header.h_cmd) {
 861         case DLM_RCOM:
 862                 /* length already checked */
 863                 break;
 864         case DLM_MSG:
 865                 if (msglen < sizeof(struct dlm_message)) {
 866                         log_print("msg too small: %u, will skip this message from node %d",
 867                                   msglen, nodeid);
 868                         return;
 869                 }
 870
 871                 break;
 872         default:
 873                 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
 874                           p->header.h_cmd, nodeid);
 875                 return;
 876         }
 877
 878         dlm_receive_buffer(p, nodeid);
 879 }
 880
 881 int dlm_validate_incoming_buffer(int nodeid, unsigned char *buf, int len)
 882 {
 883         const unsigned char *ptr = buf;
 884         const struct dlm_header *hd;
 885         uint16_t msglen;
 886         int ret = 0;
 887
 888         while (len >= sizeof(struct dlm_header)) {
 889                 hd = (struct dlm_header *)ptr;
 890
 891                 /* no message should be more than DLM_MAX_SOCKET_BUFSIZE or
 892                  * less than dlm_header size.
 893                  *
 894                  * Some messages does not have a 8 byte length boundary yet
 895                  * which can occur in a unaligned memory access of some dlm
 896                  * messages. However this problem need to be fixed at the
 897                  * sending side, for now it seems nobody run into architecture
 898                  * related issues yet but it slows down some processing.
 899                  * Fixing this issue should be scheduled in future by doing
 900                  * the next major version bump.
 901                  */
 902                 msglen = le16_to_cpu(hd->h_length);
 903                 if (msglen > DLM_MAX_SOCKET_BUFSIZE ||
 904                     msglen < sizeof(struct dlm_header)) {
 905                         log_print("received invalid length header: %u from node %d, will abort message parsing",
 906                                   msglen, nodeid);
 907                         return -EBADMSG;
 908                 }
 909
 910                 /* caller will take care that leftover
 911                  * will be parsed next call with more data
 912                  */
 913                 if (msglen > len)
 914                         break;
 915
 916                 ret += msglen;
 917                 len -= msglen;
 918                 ptr += msglen;
 919         }
 920
 921         return ret;
 922 }
 923
 924 /*
 925  * Called from the low-level comms layer to process a buffer of
 926  * commands.
 927  */
 928 int dlm_process_incoming_buffer(int nodeid, unsigned char *buf, int len)
 929 {
 930         const unsigned char *ptr = buf;
 931         const struct dlm_header *hd;
 932         uint16_t msglen;
 933         int ret = 0;
 934
 935         while (len >= sizeof(struct dlm_header)) {
 936                 hd = (struct dlm_header *)ptr;
 937
 938                 msglen = le16_to_cpu(hd->h_length);
 939                 if (msglen > len)
 940                         break;
 941
 942                 switch (hd->h_version) {
 943                 case cpu_to_le32(DLM_VERSION_3_1):
 944                         dlm_midcomms_receive_buffer_3_1((union dlm_packet *)ptr, nodeid);
 945                         break;
 946                 case cpu_to_le32(DLM_VERSION_3_2):
 947                         dlm_midcomms_receive_buffer_3_2((union dlm_packet *)ptr, nodeid);
 948                         break;
 949                 default:
 950                         log_print("received invalid version header: %u from node %d, will skip this message",
 951                                   le32_to_cpu(hd->h_version), nodeid);
 952                         break;
 953                 }
 954
 955                 ret += msglen;
 956                 len -= msglen;
 957                 ptr += msglen;
 958         }
 959
 960         return ret;
 961 }
 962
 963 void dlm_midcomms_receive_done(int nodeid)
 964 {
 965         struct midcomms_node *node;
 966         int idx;
 967
 968         idx = srcu_read_lock(&nodes_srcu);
 969         node = nodeid2node(nodeid, 0);
 970         if (!node) {
 971                 srcu_read_unlock(&nodes_srcu, idx);
 972                 return;
 973         }
 974
 975         /* old protocol, we do nothing */
 976         switch (node->version) {
 977         case DLM_VERSION_3_2:
 978                 break;
 979         default:
 980                 srcu_read_unlock(&nodes_srcu, idx);
 981                 return;
 982         }
 983
 984         /* do nothing if we didn't delivered stateful to ulp */
 985         if (!test_and_clear_bit(DLM_NODE_ULP_DELIVERED,
 986                                 &node->flags)) {
 987                 srcu_read_unlock(&nodes_srcu, idx);
 988                 return;
 989         }
 990
 991         spin_lock(&node->state_lock);
 992         /* we only ack if state is ESTABLISHED */
 993         switch (node->state) {
 994         case DLM_ESTABLISHED:
 995                 spin_unlock(&node->state_lock);
 996                 dlm_send_ack(node->nodeid, node->seq_next);
 997                 break;
 998         default:
 999                 spin_unlock(&node->state_lock);
1000                 /* do nothing FIN has it's own ack send */
1001                 break;
1002         }
1003         srcu_read_unlock(&nodes_srcu, idx);
1004 }
1005
1006 void dlm_midcomms_unack_msg_resend(int nodeid)
1007 {
1008         struct midcomms_node *node;
1009         struct dlm_mhandle *mh;
1010         int idx, ret;
1011
1012         idx = srcu_read_lock(&nodes_srcu);
1013         node = nodeid2node(nodeid, 0);
1014         if (!node) {
1015                 srcu_read_unlock(&nodes_srcu, idx);
1016                 return;
1017         }
1018
1019         /* old protocol, we don't support to retransmit on failure */
1020         switch (node->version) {
1021         case DLM_VERSION_3_2:
1022                 break;
1023         default:
1024                 srcu_read_unlock(&nodes_srcu, idx);
1025                 return;
1026         }
1027
1028         rcu_read_lock();
1029         list_for_each_entry_rcu(mh, &node->send_queue, list) {
1030                 if (!mh->committed)
1031                         continue;
1032
1033                 ret = dlm_lowcomms_resend_msg(mh->msg);
1034                 if (!ret)
1035                         log_print_ratelimited("retransmit dlm msg, seq %u, nodeid %d",
1036                                               mh->seq, node->nodeid);
1037         }
1038         rcu_read_unlock();
1039         srcu_read_unlock(&nodes_srcu, idx);
1040 }
1041
1042 static void dlm_fill_opts_header(struct dlm_opts *opts, uint16_t inner_len,
1043                                  uint32_t seq)
1044 {
1045         opts->o_header.h_cmd = DLM_OPTS;
1046         opts->o_header.h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
1047         opts->o_header.h_nodeid = cpu_to_le32(dlm_our_nodeid());
1048         opts->o_header.h_length = cpu_to_le16(DLM_MIDCOMMS_OPT_LEN + inner_len);
1049         opts->o_header.u.h_seq = cpu_to_le32(seq);
1050 }
1051
1052 static void midcomms_new_msg_cb(void *data)
1053 {
1054         struct dlm_mhandle *mh = data;
1055
1056         atomic_inc(&mh->node->send_queue_cnt);
1057
1058         spin_lock_bh(&mh->node->send_queue_lock);
1059         list_add_tail_rcu(&mh->list, &mh->node->send_queue);
1060         spin_unlock_bh(&mh->node->send_queue_lock);
1061
1062         mh->seq = mh->node->seq_send++;
1063 }
1064
1065 static struct dlm_msg *dlm_midcomms_get_msg_3_2(struct dlm_mhandle *mh, int nodeid,
1066                                                 int len, gfp_t allocation, char **ppc)
1067 {
1068         struct dlm_opts *opts;
1069         struct dlm_msg *msg;
1070
1071         msg = dlm_lowcomms_new_msg(nodeid, len + DLM_MIDCOMMS_OPT_LEN,
1072                                    allocation, ppc, midcomms_new_msg_cb, mh);
1073         if (!msg)
1074                 return NULL;
1075
1076         opts = (struct dlm_opts *)*ppc;
1077         mh->opts = opts;
1078
1079         /* add possible options here */
1080         dlm_fill_opts_header(opts, len, mh->seq);
1081
1082         *ppc += sizeof(*opts);
1083         mh->inner_p = (const union dlm_packet *)*ppc;
1084         return msg;
1085 }
1086
1087 /* avoid false positive for nodes_srcu, unlock happens in
1088  * dlm_midcomms_commit_mhandle which is a must call if success
1089  */
1090 #ifndef __CHECKER__
1091 struct dlm_mhandle *dlm_midcomms_get_mhandle(int nodeid, int len,
1092                                              gfp_t allocation, char **ppc)
1093 {
1094         struct midcomms_node *node;
1095         struct dlm_mhandle *mh;
1096         struct dlm_msg *msg;
1097         int idx;
1098
1099         idx = srcu_read_lock(&nodes_srcu);
1100         node = nodeid2node(nodeid, 0);
1101         if (!node) {
1102                 WARN_ON_ONCE(1);
1103                 goto err;
1104         }
1105
1106         /* this is a bug, however we going on and hope it will be resolved */
1107         WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_TX, &node->flags));
1108
1109         mh = dlm_allocate_mhandle(allocation);
1110         if (!mh)
1111                 goto err;
1112
1113         mh->committed = false;
1114         mh->ack_rcv = NULL;
1115         mh->idx = idx;
1116         mh->node = node;
1117
1118         switch (node->version) {
1119         case DLM_VERSION_3_1:
1120                 msg = dlm_lowcomms_new_msg(nodeid, len, allocation, ppc,
1121                                            NULL, NULL);
1122                 if (!msg) {
1123                         dlm_free_mhandle(mh);
1124                         goto err;
1125                 }
1126
1127                 break;
1128         case DLM_VERSION_3_2:
1129                 msg = dlm_midcomms_get_msg_3_2(mh, nodeid, len, allocation,
1130                                                ppc);
1131                 if (!msg) {
1132                         dlm_free_mhandle(mh);
1133                         goto err;
1134                 }
1135
1136                 break;
1137         default:
1138                 dlm_free_mhandle(mh);
1139                 WARN_ON_ONCE(1);
1140                 goto err;
1141         }
1142
1143         mh->msg = msg;
1144
1145         /* keep in mind that is a must to call
1146          * dlm_midcomms_commit_msg() which releases
1147          * nodes_srcu using mh->idx which is assumed
1148          * here that the application will call it.
1149          */
1150         return mh;
1151
1152 err:
1153         srcu_read_unlock(&nodes_srcu, idx);
1154         return NULL;
1155 }
1156 #endif
1157
1158 static void dlm_midcomms_commit_msg_3_2_trace(const struct dlm_mhandle *mh,
1159                                               const void *name, int namelen)
1160 {
1161         switch (mh->inner_p->header.h_cmd) {
1162         case DLM_MSG:
1163                 trace_dlm_send_message(mh->node->nodeid, mh->seq,
1164                                        &mh->inner_p->message,
1165                                        name, namelen);
1166                 break;
1167         case DLM_RCOM:
1168                 trace_dlm_send_rcom(mh->node->nodeid, mh->seq,
1169                                     &mh->inner_p->rcom);
1170                 break;
1171         default:
1172                 /* nothing to trace */
1173                 break;
1174         }
1175 }
1176
1177 static void dlm_midcomms_commit_msg_3_2(struct dlm_mhandle *mh,
1178                                         const void *name, int namelen)
1179 {
1180         /* nexthdr chain for fast lookup */
1181         mh->opts->o_nextcmd = mh->inner_p->header.h_cmd;
1182         mh->committed = true;
1183         dlm_midcomms_commit_msg_3_2_trace(mh, name, namelen);
1184         dlm_lowcomms_commit_msg(mh->msg);
1185 }
1186
1187 /* avoid false positive for nodes_srcu, lock was happen in
1188  * dlm_midcomms_get_mhandle
1189  */
1190 #ifndef __CHECKER__
1191 void dlm_midcomms_commit_mhandle(struct dlm_mhandle *mh,
1192                                  const void *name, int namelen)
1193 {
1194
1195         switch (mh->node->version) {
1196         case DLM_VERSION_3_1:
1197                 srcu_read_unlock(&nodes_srcu, mh->idx);
1198
1199                 dlm_lowcomms_commit_msg(mh->msg);
1200                 dlm_lowcomms_put_msg(mh->msg);
1201                 /* mh is not part of rcu list in this case */
1202                 dlm_free_mhandle(mh);
1203                 break;
1204         case DLM_VERSION_3_2:
1205                 /* held rcu read lock here, because we sending the
1206                  * dlm message out, when we do that we could receive
1207                  * an ack back which releases the mhandle and we
1208                  * get a use after free.
1209                  */
1210                 rcu_read_lock();
1211                 dlm_midcomms_commit_msg_3_2(mh, name, namelen);
1212                 srcu_read_unlock(&nodes_srcu, mh->idx);
1213                 rcu_read_unlock();
1214                 break;
1215         default:
1216                 srcu_read_unlock(&nodes_srcu, mh->idx);
1217                 WARN_ON_ONCE(1);
1218                 break;
1219         }
1220 }
1221 #endif
1222
1223 int dlm_midcomms_start(void)
1224 {
1225         return dlm_lowcomms_start();
1226 }
1227
1228 void dlm_midcomms_stop(void)
1229 {
1230         dlm_lowcomms_stop();
1231 }
1232
1233 void dlm_midcomms_init(void)
1234 {
1235         int i;
1236
1237         for (i = 0; i < CONN_HASH_SIZE; i++)
1238                 INIT_HLIST_HEAD(&node_hash[i]);
1239
1240         dlm_lowcomms_init();
1241 }
1242
1243 void dlm_midcomms_exit(void)
1244 {
1245         dlm_lowcomms_exit();
1246 }
1247
1248 static void dlm_act_fin_ack_rcv(struct midcomms_node *node)
1249 {
1250         spin_lock(&node->state_lock);
1251         pr_debug("receive active fin ack from node %d with state %s\n",
1252                  node->nodeid, dlm_state_str(node->state));
1253
1254         switch (node->state) {
1255         case DLM_FIN_WAIT1:
1256                 node->state = DLM_FIN_WAIT2;
1257                 pr_debug("switch node %d to state %s\n",
1258                          node->nodeid, dlm_state_str(node->state));
1259                 break;
1260         case DLM_CLOSING:
1261                 midcomms_node_reset(node);
1262                 pr_debug("switch node %d to state %s\n",
1263                          node->nodeid, dlm_state_str(node->state));
1264                 break;
1265         case DLM_CLOSED:
1266                 /* not valid but somehow we got what we want */
1267                 wake_up(&node->shutdown_wait);
1268                 break;
1269         default:
1270                 spin_unlock(&node->state_lock);
1271                 log_print("%s: unexpected state: %d",
1272                           __func__, node->state);
1273                 WARN_ON_ONCE(1);
1274                 return;
1275         }
1276         spin_unlock(&node->state_lock);
1277 }
1278
1279 void dlm_midcomms_add_member(int nodeid)
1280 {
1281         struct midcomms_node *node;
1282         int idx;
1283
1284         if (nodeid == dlm_our_nodeid())
1285                 return;
1286
1287         idx = srcu_read_lock(&nodes_srcu);
1288         node = nodeid2node(nodeid, GFP_NOFS);
1289         if (!node) {
1290                 srcu_read_unlock(&nodes_srcu, idx);
1291                 return;
1292         }
1293
1294         spin_lock(&node->state_lock);
1295         if (!node->users) {
1296                 pr_debug("receive add member from node %d with state %s\n",
1297                          node->nodeid, dlm_state_str(node->state));
1298                 switch (node->state) {
1299                 case DLM_ESTABLISHED:
1300                         break;
1301                 case DLM_CLOSED:
1302                         node->state = DLM_ESTABLISHED;
1303                         pr_debug("switch node %d to state %s\n",
1304                                  node->nodeid, dlm_state_str(node->state));
1305                         break;
1306                 default:
1307                         /* some invalid state passive shutdown
1308                          * was failed, we try to reset and
1309                          * hope it will go on.
1310                          */
1311                         log_print("reset node %d because shutdown stuck",
1312                                   node->nodeid);
1313
1314                         midcomms_node_reset(node);
1315                         node->state = DLM_ESTABLISHED;
1316                         break;
1317                 }
1318         }
1319
1320         node->users++;
1321         pr_debug("node %d users inc count %d\n", nodeid, node->users);
1322         spin_unlock(&node->state_lock);
1323
1324         srcu_read_unlock(&nodes_srcu, idx);
1325 }
1326
1327 void dlm_midcomms_remove_member(int nodeid)
1328 {
1329         struct midcomms_node *node;
1330         int idx;
1331
1332         if (nodeid == dlm_our_nodeid())
1333                 return;
1334
1335         idx = srcu_read_lock(&nodes_srcu);
1336         node = nodeid2node(nodeid, 0);
1337         if (!node) {
1338                 srcu_read_unlock(&nodes_srcu, idx);
1339                 return;
1340         }
1341
1342         spin_lock(&node->state_lock);
1343         node->users--;
1344         pr_debug("node %d users dec count %d\n", nodeid, node->users);
1345
1346         /* hitting users count to zero means the
1347          * other side is running dlm_midcomms_stop()
1348          * we meet us to have a clean disconnect.
1349          */
1350         if (node->users == 0) {
1351                 pr_debug("receive remove member from node %d with state %s\n",
1352                          node->nodeid, dlm_state_str(node->state));
1353                 switch (node->state) {
1354                 case DLM_ESTABLISHED:
1355                         break;
1356                 case DLM_CLOSE_WAIT:
1357                         /* passive shutdown DLM_LAST_ACK case 2 */
1358                         node->state = DLM_LAST_ACK;
1359                         pr_debug("switch node %d to state %s case 2\n",
1360                                  node->nodeid, dlm_state_str(node->state));
1361                         set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
1362                         dlm_send_fin(node, dlm_pas_fin_ack_rcv);
1363                         break;
1364                 case DLM_LAST_ACK:
1365                         /* probably receive fin caught it, do nothing */
1366                         break;
1367                 case DLM_CLOSED:
1368                         /* already gone, do nothing */
1369                         break;
1370                 default:
1371                         log_print("%s: unexpected state: %d",
1372                                   __func__, node->state);
1373                         break;
1374                 }
1375         }
1376         spin_unlock(&node->state_lock);
1377
1378         srcu_read_unlock(&nodes_srcu, idx);
1379 }
1380
1381 static void midcomms_node_release(struct rcu_head *rcu)
1382 {
1383         struct midcomms_node *node = container_of(rcu, struct midcomms_node, rcu);
1384
1385         WARN_ON_ONCE(atomic_read(&node->send_queue_cnt));
1386         dlm_send_queue_flush(node);
1387         kfree(node);
1388 }
1389
1390 void dlm_midcomms_version_wait(void)
1391 {
1392         struct midcomms_node *node;
1393         int i, idx, ret;
1394
1395         idx = srcu_read_lock(&nodes_srcu);
1396         for (i = 0; i < CONN_HASH_SIZE; i++) {
1397                 hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1398                         ret = wait_event_timeout(node->shutdown_wait,
1399                                                  node->version != DLM_VERSION_NOT_SET ||
1400                                                  node->state == DLM_CLOSED ||
1401                                                  test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1402                                                  DLM_SHUTDOWN_TIMEOUT);
1403                         if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
1404                                 pr_debug("version wait timed out for node %d with state %s\n",
1405                                          node->nodeid, dlm_state_str(node->state));
1406                 }
1407         }
1408         srcu_read_unlock(&nodes_srcu, idx);
1409 }
1410
1411 static void midcomms_shutdown(struct midcomms_node *node)
1412 {
1413         int ret;
1414
1415         /* old protocol, we don't wait for pending operations */
1416         switch (node->version) {
1417         case DLM_VERSION_3_2:
1418                 break;
1419         default:
1420                 return;
1421         }
1422
1423         spin_lock(&node->state_lock);
1424         pr_debug("receive active shutdown for node %d with state %s\n",
1425                  node->nodeid, dlm_state_str(node->state));
1426         switch (node->state) {
1427         case DLM_ESTABLISHED:
1428                 node->state = DLM_FIN_WAIT1;
1429                 pr_debug("switch node %d to state %s case 2\n",
1430                          node->nodeid, dlm_state_str(node->state));
1431                 dlm_send_fin(node, dlm_act_fin_ack_rcv);
1432                 break;
1433         case DLM_CLOSED:
1434                 /* we have what we want */
1435                 break;
1436         default:
1437                 /* busy to enter DLM_FIN_WAIT1, wait until passive
1438                  * done in shutdown_wait to enter DLM_CLOSED.
1439                  */
1440                 break;
1441         }
1442         spin_unlock(&node->state_lock);
1443
1444         if (DLM_DEBUG_FENCE_TERMINATION)
1445                 msleep(5000);
1446
1447         /* wait for other side dlm + fin */
1448         ret = wait_event_timeout(node->shutdown_wait,
1449                                  node->state == DLM_CLOSED ||
1450                                  test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1451                                  DLM_SHUTDOWN_TIMEOUT);
1452         if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
1453                 pr_debug("active shutdown timed out for node %d with state %s\n",
1454                          node->nodeid, dlm_state_str(node->state));
1455         else
1456                 pr_debug("active shutdown done for node %d with state %s\n",
1457                          node->nodeid, dlm_state_str(node->state));
1458 }
1459
1460 void dlm_midcomms_shutdown(void)
1461 {
1462         struct midcomms_node *node;
1463         int i, idx;
1464
1465         mutex_lock(&close_lock);
1466         idx = srcu_read_lock(&nodes_srcu);
1467         for (i = 0; i < CONN_HASH_SIZE; i++) {
1468                 hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1469                         midcomms_shutdown(node);
1470
1471                         dlm_delete_debug_comms_file(node->debugfs);
1472
1473                         spin_lock(&nodes_lock);
1474                         hlist_del_rcu(&node->hlist);
1475                         spin_unlock(&nodes_lock);
1476
1477                         call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
1478                 }
1479         }
1480         srcu_read_unlock(&nodes_srcu, idx);
1481         mutex_unlock(&close_lock);
1482
1483         dlm_lowcomms_shutdown();
1484 }
1485
1486 int dlm_midcomms_close(int nodeid)
1487 {
1488         struct midcomms_node *node;
1489         int idx, ret;
1490
1491         if (nodeid == dlm_our_nodeid())
1492                 return 0;
1493
1494         dlm_stop_lockspaces_check();
1495
1496         idx = srcu_read_lock(&nodes_srcu);
1497         /* Abort pending close/remove operation */
1498         node = nodeid2node(nodeid, 0);
1499         if (node) {
1500                 /* let shutdown waiters leave */
1501                 set_bit(DLM_NODE_FLAG_CLOSE, &node->flags);
1502                 wake_up(&node->shutdown_wait);
1503         }
1504         srcu_read_unlock(&nodes_srcu, idx);
1505
1506         synchronize_srcu(&nodes_srcu);
1507
1508         idx = srcu_read_lock(&nodes_srcu);
1509         mutex_lock(&close_lock);
1510         node = nodeid2node(nodeid, 0);
1511         if (!node) {
1512                 mutex_unlock(&close_lock);
1513                 srcu_read_unlock(&nodes_srcu, idx);
1514                 return dlm_lowcomms_close(nodeid);
1515         }
1516
1517         ret = dlm_lowcomms_close(nodeid);
1518         spin_lock(&node->state_lock);
1519         midcomms_node_reset(node);
1520         spin_unlock(&node->state_lock);
1521         srcu_read_unlock(&nodes_srcu, idx);
1522         mutex_unlock(&close_lock);
1523
1524         return ret;
1525 }
1526
1527 /* debug functionality to send raw dlm msg from user space */
1528 struct dlm_rawmsg_data {
1529         struct midcomms_node *node;
1530         void *buf;
1531 };
1532
1533 static void midcomms_new_rawmsg_cb(void *data)
1534 {
1535         struct dlm_rawmsg_data *rd = data;
1536         struct dlm_header *h = rd->buf;
1537
1538         switch (h->h_version) {
1539         case cpu_to_le32(DLM_VERSION_3_1):
1540                 break;
1541         default:
1542                 switch (h->h_cmd) {
1543                 case DLM_OPTS:
1544                         if (!h->u.h_seq)
1545                                 h->u.h_seq = cpu_to_le32(rd->node->seq_send++);
1546                         break;
1547                 default:
1548                         break;
1549                 }
1550                 break;
1551         }
1552 }
1553
1554 int dlm_midcomms_rawmsg_send(struct midcomms_node *node, void *buf,
1555                              int buflen)
1556 {
1557         struct dlm_rawmsg_data rd;
1558         struct dlm_msg *msg;
1559         char *msgbuf;
1560
1561         rd.node = node;
1562         rd.buf = buf;
1563
1564         msg = dlm_lowcomms_new_msg(node->nodeid, buflen, GFP_NOFS,
1565                                    &msgbuf, midcomms_new_rawmsg_cb, &rd);
1566         if (!msg)
1567                 return -ENOMEM;
1568
1569         memcpy(msgbuf, buf, buflen);
1570         dlm_lowcomms_commit_msg(msg);
1571         return 0;
1572 }
1573