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 /* 3 minutes wait to sync ending of dlm */
 149 #define DLM_SHUTDOWN_TIMEOUT    msecs_to_jiffies(3 * 60 * 1000)
 150 #define DLM_VERSION_NOT_SET     0
 151
 152 struct midcomms_node {
 153         int nodeid;
 154         uint32_t version;
 155         uint32_t seq_send;
 156         uint32_t seq_next;
 157         /* These queues are unbound because we cannot drop any message in dlm.
 158          * We could send a fence signal for a specific node to the cluster
 159          * manager if queues hits some maximum value, however this handling
 160          * not supported yet.
 161          */
 162         struct list_head send_queue;
 163         spinlock_t send_queue_lock;
 164         atomic_t send_queue_cnt;
 165 #define DLM_NODE_FLAG_CLOSE     1
 166 #define DLM_NODE_FLAG_STOP_TX   2
 167 #define DLM_NODE_FLAG_STOP_RX   3
 168 #define DLM_NODE_ULP_DELIVERED  4
 169         unsigned long flags;
 170         wait_queue_head_t shutdown_wait;
 171
 172         /* dlm tcp termination state */
 173 #define DLM_CLOSED      1
 174 #define DLM_ESTABLISHED 2
 175 #define DLM_FIN_WAIT1   3
 176 #define DLM_FIN_WAIT2   4
 177 #define DLM_CLOSE_WAIT  5
 178 #define DLM_LAST_ACK    6
 179 #define DLM_CLOSING     7
 180         int state;
 181         spinlock_t state_lock;
 182
 183         /* counts how many lockspaces are using this node
 184          * this refcount is necessary to determine if the
 185          * node wants to disconnect.
 186          */
 187         int users;
 188
 189         /* not protected by srcu, node_hash lifetime */
 190         void *debugfs;
 191
 192         struct hlist_node hlist;
 193         struct rcu_head rcu;
 194 };
 195
 196 struct dlm_mhandle {
 197         const union dlm_packet *inner_p;
 198         struct midcomms_node *node;
 199         struct dlm_opts *opts;
 200         struct dlm_msg *msg;
 201         bool committed;
 202         uint32_t seq;
 203
 204         void (*ack_rcv)(struct midcomms_node *node);
 205
 206         /* get_mhandle/commit srcu idx exchange */
 207         int idx;
 208
 209         struct list_head list;
 210         struct rcu_head rcu;
 211 };
 212
 213 static struct hlist_head node_hash[CONN_HASH_SIZE];
 214 static DEFINE_SPINLOCK(nodes_lock);
 215 DEFINE_STATIC_SRCU(nodes_srcu);
 216
 217 /* This mutex prevents that midcomms_close() is running while
 218  * stop() or remove(). As I experienced invalid memory access
 219  * behaviours when DLM_DEBUG_FENCE_TERMINATION is enabled and
 220  * resetting machines. I will end in some double deletion in nodes
 221  * datastructure.
 222  */
 223 static DEFINE_MUTEX(close_lock);
 224
 225 struct kmem_cache *dlm_midcomms_cache_create(void)
 226 {
 227         return kmem_cache_create("dlm_mhandle", sizeof(struct dlm_mhandle),
 228                                  0, 0, NULL);
 229 }
 230
 231 static inline const char *dlm_state_str(int state)
 232 {
 233         switch (state) {
 234         case DLM_CLOSED:
 235                 return "CLOSED";
 236         case DLM_ESTABLISHED:
 237                 return "ESTABLISHED";
 238         case DLM_FIN_WAIT1:
 239                 return "FIN_WAIT1";
 240         case DLM_FIN_WAIT2:
 241                 return "FIN_WAIT2";
 242         case DLM_CLOSE_WAIT:
 243                 return "CLOSE_WAIT";
 244         case DLM_LAST_ACK:
 245                 return "LAST_ACK";
 246         case DLM_CLOSING:
 247                 return "CLOSING";
 248         default:
 249                 return "UNKNOWN";
 250         }
 251 }
 252
 253 const char *dlm_midcomms_state(struct midcomms_node *node)
 254 {
 255         return dlm_state_str(node->state);
 256 }
 257
 258 unsigned long dlm_midcomms_flags(struct midcomms_node *node)
 259 {
 260         return node->flags;
 261 }
 262
 263 int dlm_midcomms_send_queue_cnt(struct midcomms_node *node)
 264 {
 265         return atomic_read(&node->send_queue_cnt);
 266 }
 267
 268 uint32_t dlm_midcomms_version(struct midcomms_node *node)
 269 {
 270         return node->version;
 271 }
 272
 273 static struct midcomms_node *__find_node(int nodeid, int r)
 274 {
 275         struct midcomms_node *node;
 276
 277         hlist_for_each_entry_rcu(node, &node_hash[r], hlist) {
 278                 if (node->nodeid == nodeid)
 279                         return node;
 280         }
 281
 282         return NULL;
 283 }
 284
 285 static void dlm_mhandle_release(struct rcu_head *rcu)
 286 {
 287         struct dlm_mhandle *mh = container_of(rcu, struct dlm_mhandle, rcu);
 288
 289         dlm_lowcomms_put_msg(mh->msg);
 290         dlm_free_mhandle(mh);
 291 }
 292
 293 static void dlm_mhandle_delete(struct midcomms_node *node,
 294                                struct dlm_mhandle *mh)
 295 {
 296         list_del_rcu(&mh->list);
 297         atomic_dec(&node->send_queue_cnt);
 298         call_rcu(&mh->rcu, dlm_mhandle_release);
 299 }
 300
 301 static void dlm_send_queue_flush(struct midcomms_node *node)
 302 {
 303         struct dlm_mhandle *mh;
 304
 305         pr_debug("flush midcomms send queue of node %d\n", node->nodeid);
 306
 307         rcu_read_lock();
 308         spin_lock(&node->send_queue_lock);
 309         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 310                 dlm_mhandle_delete(node, mh);
 311         }
 312         spin_unlock(&node->send_queue_lock);
 313         rcu_read_unlock();
 314 }
 315
 316 static void midcomms_node_reset(struct midcomms_node *node)
 317 {
 318         pr_debug("reset node %d\n", node->nodeid);
 319
 320         node->seq_next = DLM_SEQ_INIT;
 321         node->seq_send = DLM_SEQ_INIT;
 322         node->version = DLM_VERSION_NOT_SET;
 323         node->flags = 0;
 324
 325         dlm_send_queue_flush(node);
 326         node->state = DLM_CLOSED;
 327         wake_up(&node->shutdown_wait);
 328 }
 329
 330 static struct midcomms_node *nodeid2node(int nodeid, gfp_t alloc)
 331 {
 332         struct midcomms_node *node, *tmp;
 333         int r = nodeid_hash(nodeid);
 334
 335         node = __find_node(nodeid, r);
 336         if (node || !alloc)
 337                 return node;
 338
 339         node = kmalloc(sizeof(*node), alloc);
 340         if (!node)
 341                 return NULL;
 342
 343         node->nodeid = nodeid;
 344         spin_lock_init(&node->state_lock);
 345         spin_lock_init(&node->send_queue_lock);
 346         atomic_set(&node->send_queue_cnt, 0);
 347         INIT_LIST_HEAD(&node->send_queue);
 348         init_waitqueue_head(&node->shutdown_wait);
 349         node->users = 0;
 350         midcomms_node_reset(node);
 351
 352         spin_lock(&nodes_lock);
 353         /* check again if there was somebody else
 354          * earlier here to add the node
 355          */
 356         tmp = __find_node(nodeid, r);
 357         if (tmp) {
 358                 spin_unlock(&nodes_lock);
 359                 kfree(node);
 360                 return tmp;
 361         }
 362
 363         hlist_add_head_rcu(&node->hlist, &node_hash[r]);
 364         spin_unlock(&nodes_lock);
 365
 366         node->debugfs = dlm_create_debug_comms_file(nodeid, node);
 367         return node;
 368 }
 369
 370 static int dlm_send_ack(int nodeid, uint32_t seq)
 371 {
 372         int mb_len = sizeof(struct dlm_header);
 373         struct dlm_header *m_header;
 374         struct dlm_msg *msg;
 375         char *ppc;
 376
 377         msg = dlm_lowcomms_new_msg(nodeid, mb_len, GFP_ATOMIC, &ppc,
 378                                    NULL, NULL);
 379         if (!msg)
 380                 return -ENOMEM;
 381
 382         m_header = (struct dlm_header *)ppc;
 383
 384         m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
 385         m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
 386         m_header->h_length = cpu_to_le16(mb_len);
 387         m_header->h_cmd = DLM_ACK;
 388         m_header->u.h_seq = cpu_to_le32(seq);
 389
 390         dlm_lowcomms_commit_msg(msg);
 391         dlm_lowcomms_put_msg(msg);
 392
 393         return 0;
 394 }
 395
 396 static int dlm_send_fin(struct midcomms_node *node,
 397                         void (*ack_rcv)(struct midcomms_node *node))
 398 {
 399         int mb_len = sizeof(struct dlm_header);
 400         struct dlm_header *m_header;
 401         struct dlm_mhandle *mh;
 402         char *ppc;
 403
 404         mh = dlm_midcomms_get_mhandle(node->nodeid, mb_len, GFP_ATOMIC, &ppc);
 405         if (!mh)
 406                 return -ENOMEM;
 407
 408         set_bit(DLM_NODE_FLAG_STOP_TX, &node->flags);
 409         mh->ack_rcv = ack_rcv;
 410
 411         m_header = (struct dlm_header *)ppc;
 412
 413         m_header->h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
 414         m_header->h_nodeid = cpu_to_le32(dlm_our_nodeid());
 415         m_header->h_length = cpu_to_le16(mb_len);
 416         m_header->h_cmd = DLM_FIN;
 417
 418         pr_debug("sending fin msg to node %d\n", node->nodeid);
 419         dlm_midcomms_commit_mhandle(mh, NULL, 0);
 420
 421         return 0;
 422 }
 423
 424 static void dlm_receive_ack(struct midcomms_node *node, uint32_t seq)
 425 {
 426         struct dlm_mhandle *mh;
 427
 428         rcu_read_lock();
 429         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 430                 if (before(mh->seq, seq)) {
 431                         if (mh->ack_rcv)
 432                                 mh->ack_rcv(node);
 433                 } else {
 434                         /* send queue should be ordered */
 435                         break;
 436                 }
 437         }
 438
 439         spin_lock(&node->send_queue_lock);
 440         list_for_each_entry_rcu(mh, &node->send_queue, list) {
 441                 if (before(mh->seq, seq)) {
 442                         dlm_mhandle_delete(node, mh);
 443                 } else {
 444                         /* send queue should be ordered */
 445                         break;
 446                 }
 447         }
 448         spin_unlock(&node->send_queue_lock);
 449         rcu_read_unlock();
 450 }
 451
 452 static void dlm_pas_fin_ack_rcv(struct midcomms_node *node)
 453 {
 454         spin_lock(&node->state_lock);
 455         pr_debug("receive passive fin ack from node %d with state %s\n",
 456                  node->nodeid, dlm_state_str(node->state));
 457
 458         switch (node->state) {
 459         case DLM_LAST_ACK:
 460                 /* DLM_CLOSED */
 461                 midcomms_node_reset(node);
 462                 break;
 463         case DLM_CLOSED:
 464                 /* not valid but somehow we got what we want */
 465                 wake_up(&node->shutdown_wait);
 466                 break;
 467         default:
 468                 spin_unlock(&node->state_lock);
 469                 log_print("%s: unexpected state: %d\n",
 470                           __func__, node->state);
 471                 WARN_ON_ONCE(1);
 472                 return;
 473         }
 474         spin_unlock(&node->state_lock);
 475 }
 476
 477 static void dlm_receive_buffer_3_2_trace(uint32_t seq, union dlm_packet *p)
 478 {
 479         switch (p->header.h_cmd) {
 480         case DLM_MSG:
 481                 trace_dlm_recv_message(seq, &p->message);
 482                 break;
 483         case DLM_RCOM:
 484                 trace_dlm_recv_rcom(seq, &p->rcom);
 485                 break;
 486         default:
 487                 break;
 488         }
 489 }
 490
 491 static void dlm_midcomms_receive_buffer(union dlm_packet *p,
 492                                         struct midcomms_node *node,
 493                                         uint32_t seq)
 494 {
 495         if (seq == node->seq_next) {
 496                 node->seq_next++;
 497
 498                 switch (p->header.h_cmd) {
 499                 case DLM_FIN:
 500                         spin_lock(&node->state_lock);
 501                         pr_debug("receive fin msg from node %d with state %s\n",
 502                                  node->nodeid, dlm_state_str(node->state));
 503
 504                         switch (node->state) {
 505                         case DLM_ESTABLISHED:
 506                                 dlm_send_ack(node->nodeid, node->seq_next);
 507
 508                                 node->state = DLM_CLOSE_WAIT;
 509                                 pr_debug("switch node %d to state %s\n",
 510                                          node->nodeid, dlm_state_str(node->state));
 511                                 /* passive shutdown DLM_LAST_ACK case 1
 512                                  * additional we check if the node is used by
 513                                  * cluster manager events at all.
 514                                  */
 515                                 if (node->users == 0) {
 516                                         node->state = DLM_LAST_ACK;
 517                                         pr_debug("switch node %d to state %s case 1\n",
 518                                                  node->nodeid, dlm_state_str(node->state));
 519                                         set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
 520                                         dlm_send_fin(node, dlm_pas_fin_ack_rcv);
 521                                 }
 522                                 break;
 523                         case DLM_FIN_WAIT1:
 524                                 dlm_send_ack(node->nodeid, node->seq_next);
 525                                 node->state = DLM_CLOSING;
 526                                 set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
 527                                 pr_debug("switch node %d to state %s\n",
 528                                          node->nodeid, dlm_state_str(node->state));
 529                                 break;
 530                         case DLM_FIN_WAIT2:
 531                                 dlm_send_ack(node->nodeid, node->seq_next);
 532                                 midcomms_node_reset(node);
 533                                 pr_debug("switch node %d to state %s\n",
 534                                          node->nodeid, dlm_state_str(node->state));
 535                                 wake_up(&node->shutdown_wait);
 536                                 break;
 537                         case DLM_LAST_ACK:
 538                                 /* probably remove_member caught it, do nothing */
 539                                 break;
 540                         default:
 541                                 spin_unlock(&node->state_lock);
 542                                 log_print("%s: unexpected state: %d\n",
 543                                           __func__, node->state);
 544                                 WARN_ON_ONCE(1);
 545                                 return;
 546                         }
 547                         spin_unlock(&node->state_lock);
 548                         break;
 549                 default:
 550                         WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
 551                         dlm_receive_buffer_3_2_trace(seq, p);
 552                         dlm_receive_buffer(p, node->nodeid);
 553                         set_bit(DLM_NODE_ULP_DELIVERED, &node->flags);
 554                         break;
 555                 }
 556         } else {
 557                 /* retry to ack message which we already have by sending back
 558                  * current node->seq_next number as ack.
 559                  */
 560                 if (seq < node->seq_next)
 561                         dlm_send_ack(node->nodeid, node->seq_next);
 562
 563                 log_print_ratelimited("ignore dlm msg because seq mismatch, seq: %u, expected: %u, nodeid: %d",
 564                                       seq, node->seq_next, node->nodeid);
 565         }
 566 }
 567
 568 static struct midcomms_node *
 569 dlm_midcomms_recv_node_lookup(int nodeid, const union dlm_packet *p,
 570                               uint16_t msglen, int (*cb)(struct midcomms_node *node))
 571 {
 572         struct midcomms_node *node = NULL;
 573         gfp_t allocation = 0;
 574         int ret;
 575
 576         switch (p->header.h_cmd) {
 577         case DLM_RCOM:
 578                 if (msglen < sizeof(struct dlm_rcom)) {
 579                         log_print("rcom msg too small: %u, will skip this message from node %d",
 580                                   msglen, nodeid);
 581                         return NULL;
 582                 }
 583
 584                 switch (p->rcom.rc_type) {
 585                 case cpu_to_le32(DLM_RCOM_NAMES):
 586                         fallthrough;
 587                 case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
 588                         fallthrough;
 589                 case cpu_to_le32(DLM_RCOM_STATUS):
 590                         fallthrough;
 591                 case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
 592                         node = nodeid2node(nodeid, 0);
 593                         if (node) {
 594                                 spin_lock(&node->state_lock);
 595                                 if (node->state != DLM_ESTABLISHED)
 596                                         pr_debug("receive begin RCOM msg from node %d with state %s\n",
 597                                                  node->nodeid, dlm_state_str(node->state));
 598
 599                                 switch (node->state) {
 600                                 case DLM_CLOSED:
 601                                         node->state = DLM_ESTABLISHED;
 602                                         pr_debug("switch node %d to state %s\n",
 603                                                  node->nodeid, dlm_state_str(node->state));
 604                                         break;
 605                                 case DLM_ESTABLISHED:
 606                                         break;
 607                                 default:
 608                                         /* some invalid state passive shutdown
 609                                          * was failed, we try to reset and
 610                                          * hope it will go on.
 611                                          */
 612                                         log_print("reset node %d because shutdown stuck",
 613                                                   node->nodeid);
 614
 615                                         midcomms_node_reset(node);
 616                                         node->state = DLM_ESTABLISHED;
 617                                         break;
 618                                 }
 619                                 spin_unlock(&node->state_lock);
 620                         }
 621
 622                         allocation = GFP_NOFS;
 623                         break;
 624                 default:
 625                         break;
 626                 }
 627
 628                 break;
 629         default:
 630                 break;
 631         }
 632
 633         node = nodeid2node(nodeid, allocation);
 634         if (!node) {
 635                 switch (p->header.h_cmd) {
 636                 case DLM_OPTS:
 637                         if (msglen < sizeof(struct dlm_opts)) {
 638                                 log_print("opts msg too small: %u, will skip this message from node %d",
 639                                           msglen, nodeid);
 640                                 return NULL;
 641                         }
 642
 643                         log_print_ratelimited("received dlm opts message nextcmd %d from node %d in an invalid sequence",
 644                                               p->opts.o_nextcmd, nodeid);
 645                         break;
 646                 default:
 647                         log_print_ratelimited("received dlm message cmd %d from node %d in an invalid sequence",
 648                                               p->header.h_cmd, nodeid);
 649                         break;
 650                 }
 651
 652                 return NULL;
 653         }
 654
 655         ret = cb(node);
 656         if (ret < 0)
 657                 return NULL;
 658
 659         return node;
 660 }
 661
 662 static int dlm_midcomms_version_check_3_2(struct midcomms_node *node)
 663 {
 664         switch (node->version) {
 665         case DLM_VERSION_NOT_SET:
 666                 node->version = DLM_VERSION_3_2;
 667                 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_2,
 668                           node->nodeid);
 669                 break;
 670         case DLM_VERSION_3_2:
 671                 break;
 672         default:
 673                 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
 674                                       DLM_VERSION_3_2, node->nodeid, node->version);
 675                 return -1;
 676         }
 677
 678         return 0;
 679 }
 680
 681 static int dlm_opts_check_msglen(union dlm_packet *p, uint16_t msglen, int nodeid)
 682 {
 683         int len = msglen;
 684
 685         /* we only trust outer header msglen because
 686          * it's checked against receive buffer length.
 687          */
 688         if (len < sizeof(struct dlm_opts))
 689                 return -1;
 690         len -= sizeof(struct dlm_opts);
 691
 692         if (len < le16_to_cpu(p->opts.o_optlen))
 693                 return -1;
 694         len -= le16_to_cpu(p->opts.o_optlen);
 695
 696         switch (p->opts.o_nextcmd) {
 697         case DLM_FIN:
 698                 if (len < sizeof(struct dlm_header)) {
 699                         log_print("fin too small: %d, will skip this message from node %d",
 700                                   len, nodeid);
 701                         return -1;
 702                 }
 703
 704                 break;
 705         case DLM_MSG:
 706                 if (len < sizeof(struct dlm_message)) {
 707                         log_print("msg too small: %d, will skip this message from node %d",
 708                                   msglen, nodeid);
 709                         return -1;
 710                 }
 711
 712                 break;
 713         case DLM_RCOM:
 714                 if (len < sizeof(struct dlm_rcom)) {
 715                         log_print("rcom msg too small: %d, will skip this message from node %d",
 716                                   len, nodeid);
 717                         return -1;
 718                 }
 719
 720                 break;
 721         default:
 722                 log_print("unsupported o_nextcmd received: %u, will skip this message from node %d",
 723                           p->opts.o_nextcmd, nodeid);
 724                 return -1;
 725         }
 726
 727         return 0;
 728 }
 729
 730 static void dlm_midcomms_receive_buffer_3_2(union dlm_packet *p, int nodeid)
 731 {
 732         uint16_t msglen = le16_to_cpu(p->header.h_length);
 733         struct midcomms_node *node;
 734         uint32_t seq;
 735         int ret, idx;
 736
 737         idx = srcu_read_lock(&nodes_srcu);
 738         node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
 739                                              dlm_midcomms_version_check_3_2);
 740         if (!node)
 741                 goto out;
 742
 743         switch (p->header.h_cmd) {
 744         case DLM_RCOM:
 745                 /* these rcom message we use to determine version.
 746                  * they have their own retransmission handling and
 747                  * are the first messages of dlm.
 748                  *
 749                  * length already checked.
 750                  */
 751                 switch (p->rcom.rc_type) {
 752                 case cpu_to_le32(DLM_RCOM_NAMES):
 753                         fallthrough;
 754                 case cpu_to_le32(DLM_RCOM_NAMES_REPLY):
 755                         fallthrough;
 756                 case cpu_to_le32(DLM_RCOM_STATUS):
 757                         fallthrough;
 758                 case cpu_to_le32(DLM_RCOM_STATUS_REPLY):
 759                         break;
 760                 default:
 761                         log_print("unsupported rcom type received: %u, will skip this message from node %d",
 762                                   le32_to_cpu(p->rcom.rc_type), nodeid);
 763                         goto out;
 764                 }
 765
 766                 WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags));
 767                 dlm_receive_buffer(p, nodeid);
 768                 break;
 769         case DLM_OPTS:
 770                 seq = le32_to_cpu(p->header.u.h_seq);
 771
 772                 ret = dlm_opts_check_msglen(p, msglen, nodeid);
 773                 if (ret < 0) {
 774                         log_print("opts msg too small: %u, will skip this message from node %d",
 775                                   msglen, nodeid);
 776                         goto out;
 777                 }
 778
 779                 p = (union dlm_packet *)((unsigned char *)p->opts.o_opts +
 780                                          le16_to_cpu(p->opts.o_optlen));
 781
 782                 /* recheck inner msglen just if it's not garbage */
 783                 msglen = le16_to_cpu(p->header.h_length);
 784                 switch (p->header.h_cmd) {
 785                 case DLM_RCOM:
 786                         if (msglen < sizeof(struct dlm_rcom)) {
 787                                 log_print("inner rcom msg too small: %u, will skip this message from node %d",
 788                                           msglen, nodeid);
 789                                 goto out;
 790                         }
 791
 792                         break;
 793                 case DLM_MSG:
 794                         if (msglen < sizeof(struct dlm_message)) {
 795                                 log_print("inner msg too small: %u, will skip this message from node %d",
 796                                           msglen, nodeid);
 797                                 goto out;
 798                         }
 799
 800                         break;
 801                 case DLM_FIN:
 802                         if (msglen < sizeof(struct dlm_header)) {
 803                                 log_print("inner fin too small: %u, will skip this message from node %d",
 804                                           msglen, nodeid);
 805                                 goto out;
 806                         }
 807
 808                         break;
 809                 default:
 810                         log_print("unsupported inner h_cmd received: %u, will skip this message from node %d",
 811                                   msglen, nodeid);
 812                         goto out;
 813                 }
 814
 815                 dlm_midcomms_receive_buffer(p, node, seq);
 816                 break;
 817         case DLM_ACK:
 818                 seq = le32_to_cpu(p->header.u.h_seq);
 819                 dlm_receive_ack(node, seq);
 820                 break;
 821         default:
 822                 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
 823                           p->header.h_cmd, nodeid);
 824                 break;
 825         }
 826
 827 out:
 828         srcu_read_unlock(&nodes_srcu, idx);
 829 }
 830
 831 static int dlm_midcomms_version_check_3_1(struct midcomms_node *node)
 832 {
 833         switch (node->version) {
 834         case DLM_VERSION_NOT_SET:
 835                 node->version = DLM_VERSION_3_1;
 836                 log_print("version 0x%08x for node %d detected", DLM_VERSION_3_1,
 837                           node->nodeid);
 838                 break;
 839         case DLM_VERSION_3_1:
 840                 break;
 841         default:
 842                 log_print_ratelimited("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
 843                                       DLM_VERSION_3_1, node->nodeid, node->version);
 844                 return -1;
 845         }
 846
 847         return 0;
 848 }
 849
 850 static void dlm_midcomms_receive_buffer_3_1(union dlm_packet *p, int nodeid)
 851 {
 852         uint16_t msglen = le16_to_cpu(p->header.h_length);
 853         struct midcomms_node *node;
 854         int idx;
 855
 856         idx = srcu_read_lock(&nodes_srcu);
 857         node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
 858                                              dlm_midcomms_version_check_3_1);
 859         if (!node) {
 860                 srcu_read_unlock(&nodes_srcu, idx);
 861                 return;
 862         }
 863         srcu_read_unlock(&nodes_srcu, idx);
 864
 865         switch (p->header.h_cmd) {
 866         case DLM_RCOM:
 867                 /* length already checked */
 868                 break;
 869         case DLM_MSG:
 870                 if (msglen < sizeof(struct dlm_message)) {
 871                         log_print("msg too small: %u, will skip this message from node %d",
 872                                   msglen, nodeid);
 873                         return;
 874                 }
 875
 876                 break;
 877         default:
 878                 log_print("unsupported h_cmd received: %u, will skip this message from node %d",
 879                           p->header.h_cmd, nodeid);
 880                 return;
 881         }
 882
 883         dlm_receive_buffer(p, nodeid);
 884 }
 885
 886 /*
 887  * Called from the low-level comms layer to process a buffer of
 888  * commands.
 889  */
 890
 891 int dlm_process_incoming_buffer(int nodeid, unsigned char *buf, int len)
 892 {
 893         const unsigned char *ptr = buf;
 894         const struct dlm_header *hd;
 895         uint16_t msglen;
 896         int ret = 0;
 897
 898         while (len >= sizeof(struct dlm_header)) {
 899                 hd = (struct dlm_header *)ptr;
 900
 901                 /* no message should be more than DLM_MAX_SOCKET_BUFSIZE or
 902                  * less than dlm_header size.
 903                  *
 904                  * Some messages does not have a 8 byte length boundary yet
 905                  * which can occur in a unaligned memory access of some dlm
 906                  * messages. However this problem need to be fixed at the
 907                  * sending side, for now it seems nobody run into architecture
 908                  * related issues yet but it slows down some processing.
 909                  * Fixing this issue should be scheduled in future by doing
 910                  * the next major version bump.
 911                  */
 912                 msglen = le16_to_cpu(hd->h_length);
 913                 if (msglen > DLM_MAX_SOCKET_BUFSIZE ||
 914                     msglen < sizeof(struct dlm_header)) {
 915                         log_print("received invalid length header: %u from node %d, will abort message parsing",
 916                                   msglen, nodeid);
 917                         return -EBADMSG;
 918                 }
 919
 920                 /* caller will take care that leftover
 921                  * will be parsed next call with more data
 922                  */
 923                 if (msglen > len)
 924                         break;
 925
 926                 switch (hd->h_version) {
 927                 case cpu_to_le32(DLM_VERSION_3_1):
 928                         dlm_midcomms_receive_buffer_3_1((union dlm_packet *)ptr, nodeid);
 929                         break;
 930                 case cpu_to_le32(DLM_VERSION_3_2):
 931                         dlm_midcomms_receive_buffer_3_2((union dlm_packet *)ptr, nodeid);
 932                         break;
 933                 default:
 934                         log_print("received invalid version header: %u from node %d, will skip this message",
 935                                   le32_to_cpu(hd->h_version), nodeid);
 936                         break;
 937                 }
 938
 939                 ret += msglen;
 940                 len -= msglen;
 941                 ptr += msglen;
 942         }
 943
 944         return ret;
 945 }
 946
 947 void dlm_midcomms_receive_done(int nodeid)
 948 {
 949         struct midcomms_node *node;
 950         int idx;
 951
 952         idx = srcu_read_lock(&nodes_srcu);
 953         node = nodeid2node(nodeid, 0);
 954         if (!node) {
 955                 srcu_read_unlock(&nodes_srcu, idx);
 956                 return;
 957         }
 958
 959         /* old protocol, we do nothing */
 960         switch (node->version) {
 961         case DLM_VERSION_3_2:
 962                 break;
 963         default:
 964                 srcu_read_unlock(&nodes_srcu, idx);
 965                 return;
 966         }
 967
 968         /* do nothing if we didn't delivered stateful to ulp */
 969         if (!test_and_clear_bit(DLM_NODE_ULP_DELIVERED,
 970                                 &node->flags)) {
 971                 srcu_read_unlock(&nodes_srcu, idx);
 972                 return;
 973         }
 974
 975         spin_lock(&node->state_lock);
 976         /* we only ack if state is ESTABLISHED */
 977         switch (node->state) {
 978         case DLM_ESTABLISHED:
 979                 spin_unlock(&node->state_lock);
 980                 dlm_send_ack(node->nodeid, node->seq_next);
 981                 break;
 982         default:
 983                 spin_unlock(&node->state_lock);
 984                 /* do nothing FIN has it's own ack send */
 985                 break;
 986         }
 987         srcu_read_unlock(&nodes_srcu, idx);
 988 }
 989
 990 void dlm_midcomms_unack_msg_resend(int nodeid)
 991 {
 992         struct midcomms_node *node;
 993         struct dlm_mhandle *mh;
 994         int idx, ret;
 995
 996         idx = srcu_read_lock(&nodes_srcu);
 997         node = nodeid2node(nodeid, 0);
 998         if (!node) {
 999                 srcu_read_unlock(&nodes_srcu, idx);
1000                 return;
1001         }
1002
1003         /* old protocol, we don't support to retransmit on failure */
1004         switch (node->version) {
1005         case DLM_VERSION_3_2:
1006                 break;
1007         default:
1008                 srcu_read_unlock(&nodes_srcu, idx);
1009                 return;
1010         }
1011
1012         rcu_read_lock();
1013         list_for_each_entry_rcu(mh, &node->send_queue, list) {
1014                 if (!mh->committed)
1015                         continue;
1016
1017                 ret = dlm_lowcomms_resend_msg(mh->msg);
1018                 if (!ret)
1019                         log_print_ratelimited("retransmit dlm msg, seq %u, nodeid %d",
1020                                               mh->seq, node->nodeid);
1021         }
1022         rcu_read_unlock();
1023         srcu_read_unlock(&nodes_srcu, idx);
1024 }
1025
1026 static void dlm_fill_opts_header(struct dlm_opts *opts, uint16_t inner_len,
1027                                  uint32_t seq)
1028 {
1029         opts->o_header.h_cmd = DLM_OPTS;
1030         opts->o_header.h_version = cpu_to_le32(DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
1031         opts->o_header.h_nodeid = cpu_to_le32(dlm_our_nodeid());
1032         opts->o_header.h_length = cpu_to_le16(DLM_MIDCOMMS_OPT_LEN + inner_len);
1033         opts->o_header.u.h_seq = cpu_to_le32(seq);
1034 }
1035
1036 static void midcomms_new_msg_cb(void *data)
1037 {
1038         struct dlm_mhandle *mh = data;
1039
1040         atomic_inc(&mh->node->send_queue_cnt);
1041
1042         spin_lock(&mh->node->send_queue_lock);
1043         list_add_tail_rcu(&mh->list, &mh->node->send_queue);
1044         spin_unlock(&mh->node->send_queue_lock);
1045
1046         mh->seq = mh->node->seq_send++;
1047 }
1048
1049 static struct dlm_msg *dlm_midcomms_get_msg_3_2(struct dlm_mhandle *mh, int nodeid,
1050                                                 int len, gfp_t allocation, char **ppc)
1051 {
1052         struct dlm_opts *opts;
1053         struct dlm_msg *msg;
1054
1055         msg = dlm_lowcomms_new_msg(nodeid, len + DLM_MIDCOMMS_OPT_LEN,
1056                                    allocation, ppc, midcomms_new_msg_cb, mh);
1057         if (!msg)
1058                 return NULL;
1059
1060         opts = (struct dlm_opts *)*ppc;
1061         mh->opts = opts;
1062
1063         /* add possible options here */
1064         dlm_fill_opts_header(opts, len, mh->seq);
1065
1066         *ppc += sizeof(*opts);
1067         mh->inner_p = (const union dlm_packet *)*ppc;
1068         return msg;
1069 }
1070
1071 /* avoid false positive for nodes_srcu, unlock happens in
1072  * dlm_midcomms_commit_mhandle which is a must call if success
1073  */
1074 #ifndef __CHECKER__
1075 struct dlm_mhandle *dlm_midcomms_get_mhandle(int nodeid, int len,
1076                                              gfp_t allocation, char **ppc)
1077 {
1078         struct midcomms_node *node;
1079         struct dlm_mhandle *mh;
1080         struct dlm_msg *msg;
1081         int idx;
1082
1083         idx = srcu_read_lock(&nodes_srcu);
1084         node = nodeid2node(nodeid, 0);
1085         if (!node) {
1086                 WARN_ON_ONCE(1);
1087                 goto err;
1088         }
1089
1090         /* this is a bug, however we going on and hope it will be resolved */
1091         WARN_ON_ONCE(test_bit(DLM_NODE_FLAG_STOP_TX, &node->flags));
1092
1093         mh = dlm_allocate_mhandle();
1094         if (!mh)
1095                 goto err;
1096
1097         mh->committed = false;
1098         mh->ack_rcv = NULL;
1099         mh->idx = idx;
1100         mh->node = node;
1101
1102         switch (node->version) {
1103         case DLM_VERSION_3_1:
1104                 msg = dlm_lowcomms_new_msg(nodeid, len, allocation, ppc,
1105                                            NULL, NULL);
1106                 if (!msg) {
1107                         dlm_free_mhandle(mh);
1108                         goto err;
1109                 }
1110
1111                 break;
1112         case DLM_VERSION_3_2:
1113                 msg = dlm_midcomms_get_msg_3_2(mh, nodeid, len, allocation,
1114                                                ppc);
1115                 if (!msg) {
1116                         dlm_free_mhandle(mh);
1117                         goto err;
1118                 }
1119
1120                 break;
1121         default:
1122                 dlm_free_mhandle(mh);
1123                 WARN_ON_ONCE(1);
1124                 goto err;
1125         }
1126
1127         mh->msg = msg;
1128
1129         /* keep in mind that is a must to call
1130          * dlm_midcomms_commit_msg() which releases
1131          * nodes_srcu using mh->idx which is assumed
1132          * here that the application will call it.
1133          */
1134         return mh;
1135
1136 err:
1137         srcu_read_unlock(&nodes_srcu, idx);
1138         return NULL;
1139 }
1140 #endif
1141
1142 static void dlm_midcomms_commit_msg_3_2_trace(const struct dlm_mhandle *mh,
1143                                               const void *name, int namelen)
1144 {
1145         switch (mh->inner_p->header.h_cmd) {
1146         case DLM_MSG:
1147                 trace_dlm_send_message(mh->seq, &mh->inner_p->message,
1148                                        name, namelen);
1149                 break;
1150         case DLM_RCOM:
1151                 trace_dlm_send_rcom(mh->seq, &mh->inner_p->rcom);
1152                 break;
1153         default:
1154                 /* nothing to trace */
1155                 break;
1156         }
1157 }
1158
1159 static void dlm_midcomms_commit_msg_3_2(struct dlm_mhandle *mh,
1160                                         const void *name, int namelen)
1161 {
1162         /* nexthdr chain for fast lookup */
1163         mh->opts->o_nextcmd = mh->inner_p->header.h_cmd;
1164         mh->committed = true;
1165         dlm_midcomms_commit_msg_3_2_trace(mh, name, namelen);
1166         dlm_lowcomms_commit_msg(mh->msg);
1167 }
1168
1169 /* avoid false positive for nodes_srcu, lock was happen in
1170  * dlm_midcomms_get_mhandle
1171  */
1172 #ifndef __CHECKER__
1173 void dlm_midcomms_commit_mhandle(struct dlm_mhandle *mh,
1174                                  const void *name, int namelen)
1175 {
1176
1177         switch (mh->node->version) {
1178         case DLM_VERSION_3_1:
1179                 srcu_read_unlock(&nodes_srcu, mh->idx);
1180
1181                 dlm_lowcomms_commit_msg(mh->msg);
1182                 dlm_lowcomms_put_msg(mh->msg);
1183                 /* mh is not part of rcu list in this case */
1184                 dlm_free_mhandle(mh);
1185                 break;
1186         case DLM_VERSION_3_2:
1187                 /* held rcu read lock here, because we sending the
1188                  * dlm message out, when we do that we could receive
1189                  * an ack back which releases the mhandle and we
1190                  * get a use after free.
1191                  */
1192                 rcu_read_lock();
1193                 dlm_midcomms_commit_msg_3_2(mh, name, namelen);
1194                 srcu_read_unlock(&nodes_srcu, mh->idx);
1195                 rcu_read_unlock();
1196                 break;
1197         default:
1198                 srcu_read_unlock(&nodes_srcu, mh->idx);
1199                 WARN_ON_ONCE(1);
1200                 break;
1201         }
1202 }
1203 #endif
1204
1205 int dlm_midcomms_start(void)
1206 {
1207         return dlm_lowcomms_start();
1208 }
1209
1210 void dlm_midcomms_stop(void)
1211 {
1212         dlm_lowcomms_stop();
1213 }
1214
1215 void dlm_midcomms_init(void)
1216 {
1217         int i;
1218
1219         for (i = 0; i < CONN_HASH_SIZE; i++)
1220                 INIT_HLIST_HEAD(&node_hash[i]);
1221
1222         dlm_lowcomms_init();
1223 }
1224
1225 void dlm_midcomms_exit(void)
1226 {
1227         dlm_lowcomms_exit();
1228 }
1229
1230 static void dlm_act_fin_ack_rcv(struct midcomms_node *node)
1231 {
1232         spin_lock(&node->state_lock);
1233         pr_debug("receive active fin ack from node %d with state %s\n",
1234                  node->nodeid, dlm_state_str(node->state));
1235
1236         switch (node->state) {
1237         case DLM_FIN_WAIT1:
1238                 node->state = DLM_FIN_WAIT2;
1239                 pr_debug("switch node %d to state %s\n",
1240                          node->nodeid, dlm_state_str(node->state));
1241                 break;
1242         case DLM_CLOSING:
1243                 midcomms_node_reset(node);
1244                 pr_debug("switch node %d to state %s\n",
1245                          node->nodeid, dlm_state_str(node->state));
1246                 wake_up(&node->shutdown_wait);
1247                 break;
1248         case DLM_CLOSED:
1249                 /* not valid but somehow we got what we want */
1250                 wake_up(&node->shutdown_wait);
1251                 break;
1252         default:
1253                 spin_unlock(&node->state_lock);
1254                 log_print("%s: unexpected state: %d\n",
1255                           __func__, node->state);
1256                 WARN_ON_ONCE(1);
1257                 return;
1258         }
1259         spin_unlock(&node->state_lock);
1260 }
1261
1262 void dlm_midcomms_add_member(int nodeid)
1263 {
1264         struct midcomms_node *node;
1265         int idx;
1266
1267         if (nodeid == dlm_our_nodeid())
1268                 return;
1269
1270         idx = srcu_read_lock(&nodes_srcu);
1271         node = nodeid2node(nodeid, GFP_NOFS);
1272         if (!node) {
1273                 srcu_read_unlock(&nodes_srcu, idx);
1274                 return;
1275         }
1276
1277         spin_lock(&node->state_lock);
1278         if (!node->users) {
1279                 pr_debug("receive add member from node %d with state %s\n",
1280                          node->nodeid, dlm_state_str(node->state));
1281                 switch (node->state) {
1282                 case DLM_ESTABLISHED:
1283                         break;
1284                 case DLM_CLOSED:
1285                         node->state = DLM_ESTABLISHED;
1286                         pr_debug("switch node %d to state %s\n",
1287                                  node->nodeid, dlm_state_str(node->state));
1288                         break;
1289                 default:
1290                         /* some invalid state passive shutdown
1291                          * was failed, we try to reset and
1292                          * hope it will go on.
1293                          */
1294                         log_print("reset node %d because shutdown stuck",
1295                                   node->nodeid);
1296
1297                         midcomms_node_reset(node);
1298                         node->state = DLM_ESTABLISHED;
1299                         break;
1300                 }
1301         }
1302
1303         node->users++;
1304         pr_debug("node %d users inc count %d\n", nodeid, node->users);
1305         spin_unlock(&node->state_lock);
1306
1307         srcu_read_unlock(&nodes_srcu, idx);
1308 }
1309
1310 void dlm_midcomms_remove_member(int nodeid)
1311 {
1312         struct midcomms_node *node;
1313         int idx;
1314
1315         if (nodeid == dlm_our_nodeid())
1316                 return;
1317
1318         idx = srcu_read_lock(&nodes_srcu);
1319         node = nodeid2node(nodeid, 0);
1320         if (!node) {
1321                 srcu_read_unlock(&nodes_srcu, idx);
1322                 return;
1323         }
1324
1325         spin_lock(&node->state_lock);
1326         node->users--;
1327         pr_debug("node %d users dec count %d\n", nodeid, node->users);
1328
1329         /* hitting users count to zero means the
1330          * other side is running dlm_midcomms_stop()
1331          * we meet us to have a clean disconnect.
1332          */
1333         if (node->users == 0) {
1334                 pr_debug("receive remove member from node %d with state %s\n",
1335                          node->nodeid, dlm_state_str(node->state));
1336                 switch (node->state) {
1337                 case DLM_ESTABLISHED:
1338                         break;
1339                 case DLM_CLOSE_WAIT:
1340                         /* passive shutdown DLM_LAST_ACK case 2 */
1341                         node->state = DLM_LAST_ACK;
1342                         pr_debug("switch node %d to state %s case 2\n",
1343                                  node->nodeid, dlm_state_str(node->state));
1344                         set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
1345                         dlm_send_fin(node, dlm_pas_fin_ack_rcv);
1346                         break;
1347                 case DLM_LAST_ACK:
1348                         /* probably receive fin caught it, do nothing */
1349                         break;
1350                 case DLM_CLOSED:
1351                         /* already gone, do nothing */
1352                         break;
1353                 default:
1354                         log_print("%s: unexpected state: %d\n",
1355                                   __func__, node->state);
1356                         break;
1357                 }
1358         }
1359         spin_unlock(&node->state_lock);
1360
1361         srcu_read_unlock(&nodes_srcu, idx);
1362 }
1363
1364 static void midcomms_node_release(struct rcu_head *rcu)
1365 {
1366         struct midcomms_node *node = container_of(rcu, struct midcomms_node, rcu);
1367
1368         WARN_ON_ONCE(atomic_read(&node->send_queue_cnt));
1369         dlm_send_queue_flush(node);
1370         kfree(node);
1371 }
1372
1373 static void midcomms_shutdown(struct midcomms_node *node)
1374 {
1375         int ret;
1376
1377         /* old protocol, we don't wait for pending operations */
1378         switch (node->version) {
1379         case DLM_VERSION_3_2:
1380                 break;
1381         default:
1382                 return;
1383         }
1384
1385         spin_lock(&node->state_lock);
1386         pr_debug("receive active shutdown for node %d with state %s\n",
1387                  node->nodeid, dlm_state_str(node->state));
1388         switch (node->state) {
1389         case DLM_ESTABLISHED:
1390                 node->state = DLM_FIN_WAIT1;
1391                 pr_debug("switch node %d to state %s case 2\n",
1392                          node->nodeid, dlm_state_str(node->state));
1393                 dlm_send_fin(node, dlm_act_fin_ack_rcv);
1394                 break;
1395         case DLM_CLOSED:
1396                 /* we have what we want */
1397                 spin_unlock(&node->state_lock);
1398                 return;
1399         default:
1400                 /* busy to enter DLM_FIN_WAIT1, wait until passive
1401                  * done in shutdown_wait to enter DLM_CLOSED.
1402                  */
1403                 break;
1404         }
1405         spin_unlock(&node->state_lock);
1406
1407         if (DLM_DEBUG_FENCE_TERMINATION)
1408                 msleep(5000);
1409
1410         /* wait for other side dlm + fin */
1411         ret = wait_event_timeout(node->shutdown_wait,
1412                                  node->state == DLM_CLOSED ||
1413                                  test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
1414                                  DLM_SHUTDOWN_TIMEOUT);
1415         if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags)) {
1416                 pr_debug("active shutdown timed out for node %d with state %s\n",
1417                          node->nodeid, dlm_state_str(node->state));
1418                 midcomms_node_reset(node);
1419                 return;
1420         }
1421
1422         pr_debug("active shutdown done for node %d with state %s\n",
1423                  node->nodeid, dlm_state_str(node->state));
1424 }
1425
1426 void dlm_midcomms_shutdown(void)
1427 {
1428         struct midcomms_node *node;
1429         int i, idx;
1430
1431         mutex_lock(&close_lock);
1432         idx = srcu_read_lock(&nodes_srcu);
1433         for (i = 0; i < CONN_HASH_SIZE; i++) {
1434                 hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
1435                         midcomms_shutdown(node);
1436
1437                         dlm_delete_debug_comms_file(node->debugfs);
1438
1439                         spin_lock(&nodes_lock);
1440                         hlist_del_rcu(&node->hlist);
1441                         spin_unlock(&nodes_lock);
1442
1443                         call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
1444                 }
1445         }
1446         srcu_read_unlock(&nodes_srcu, idx);
1447         mutex_unlock(&close_lock);
1448
1449         dlm_lowcomms_shutdown();
1450 }
1451
1452 int dlm_midcomms_close(int nodeid)
1453 {
1454         struct midcomms_node *node;
1455         int idx, ret;
1456
1457         if (nodeid == dlm_our_nodeid())
1458                 return 0;
1459
1460         idx = srcu_read_lock(&nodes_srcu);
1461         /* Abort pending close/remove operation */
1462         node = nodeid2node(nodeid, 0);
1463         if (node) {
1464                 /* let shutdown waiters leave */
1465                 set_bit(DLM_NODE_FLAG_CLOSE, &node->flags);
1466                 wake_up(&node->shutdown_wait);
1467         }
1468         srcu_read_unlock(&nodes_srcu, idx);
1469
1470         synchronize_srcu(&nodes_srcu);
1471
1472         idx = srcu_read_lock(&nodes_srcu);
1473         mutex_lock(&close_lock);
1474         node = nodeid2node(nodeid, 0);
1475         if (!node) {
1476                 mutex_unlock(&close_lock);
1477                 srcu_read_unlock(&nodes_srcu, idx);
1478                 return dlm_lowcomms_close(nodeid);
1479         }
1480
1481         ret = dlm_lowcomms_close(nodeid);
1482         spin_lock(&node->state_lock);
1483         midcomms_node_reset(node);
1484         spin_unlock(&node->state_lock);
1485         srcu_read_unlock(&nodes_srcu, idx);
1486         mutex_unlock(&close_lock);
1487
1488         return ret;
1489 }
1490
1491 /* debug functionality to send raw dlm msg from user space */
1492 struct dlm_rawmsg_data {
1493         struct midcomms_node *node;
1494         void *buf;
1495 };
1496
1497 static void midcomms_new_rawmsg_cb(void *data)
1498 {
1499         struct dlm_rawmsg_data *rd = data;
1500         struct dlm_header *h = rd->buf;
1501
1502         switch (h->h_version) {
1503         case cpu_to_le32(DLM_VERSION_3_1):
1504                 break;
1505         default:
1506                 switch (h->h_cmd) {
1507                 case DLM_OPTS:
1508                         if (!h->u.h_seq)
1509                                 h->u.h_seq = cpu_to_le32(rd->node->seq_send++);
1510                         break;
1511                 default:
1512                         break;
1513                 }
1514                 break;
1515         }
1516 }
1517
1518 int dlm_midcomms_rawmsg_send(struct midcomms_node *node, void *buf,
1519                              int buflen)
1520 {
1521         struct dlm_rawmsg_data rd;
1522         struct dlm_msg *msg;
1523         char *msgbuf;
1524
1525         rd.node = node;
1526         rd.buf = buf;
1527
1528         msg = dlm_lowcomms_new_msg(node->nodeid, buflen, GFP_NOFS,
1529                                    &msgbuf, midcomms_new_rawmsg_cb, &rd);
1530         if (!msg)
1531                 return -ENOMEM;
1532
1533         memcpy(msgbuf, buf, buflen);
1534         dlm_lowcomms_commit_msg(msg);
1535         return 0;
1536 }
1537