fs/dlm/lowcomms.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) 2004-2009 Red Hat, Inc.  All rights reserved.
   7 **
   8 **
   9 *******************************************************************************
  10 ******************************************************************************/
  11
  12 /*
  13  * lowcomms.c
  14  *
  15  * This is the "low-level" comms layer.
  16  *
  17  * It is responsible for sending/receiving messages
  18  * from other nodes in the cluster.
  19  *
  20  * Cluster nodes are referred to by their nodeids. nodeids are
  21  * simply 32 bit numbers to the locking module - if they need to
  22  * be expanded for the cluster infrastructure then that is its
  23  * responsibility. It is this layer's
  24  * responsibility to resolve these into IP address or
  25  * whatever it needs for inter-node communication.
  26  *
  27  * The comms level is two kernel threads that deal mainly with
  28  * the receiving of messages from other nodes and passing them
  29  * up to the mid-level comms layer (which understands the
  30  * message format) for execution by the locking core, and
  31  * a send thread which does all the setting up of connections
  32  * to remote nodes and the sending of data. Threads are not allowed
  33  * to send their own data because it may cause them to wait in times
  34  * of high load. Also, this way, the sending thread can collect together
  35  * messages bound for one node and send them in one block.
  36  *
  37  * lowcomms will choose to use either TCP or SCTP as its transport layer
  38  * depending on the configuration variable 'protocol'. This should be set
  39  * to 0 (default) for TCP or 1 for SCTP. It should be configured using a
  40  * cluster-wide mechanism as it must be the same on all nodes of the cluster
  41  * for the DLM to function.
  42  *
  43  */
  44
  45 #include <asm/ioctls.h>
  46 #include <net/sock.h>
  47 #include <net/tcp.h>
  48 #include <linux/pagemap.h>
  49 #include <linux/file.h>
  50 #include <linux/mutex.h>
  51 #include <linux/sctp.h>
  52 #include <linux/slab.h>
  53 #include <net/sctp/sctp.h>
  54 #include <net/ipv6.h>
  55
  56 #include <trace/events/dlm.h>
  57
  58 #include "dlm_internal.h"
  59 #include "lowcomms.h"
  60 #include "midcomms.h"
  61 #include "memory.h"
  62 #include "config.h"
  63
  64 #define NEEDED_RMEM (4*1024*1024)
  65
  66 /* Number of messages to send before rescheduling */
  67 #define MAX_SEND_MSG_COUNT 25
  68 #define DLM_SHUTDOWN_WAIT_TIMEOUT msecs_to_jiffies(10000)
  69
  70 struct connection {
  71         struct socket *sock;    /* NULL if not connected */
  72         uint32_t nodeid;        /* So we know who we are in the list */
  73         struct mutex sock_mutex;
  74         unsigned long flags;
  75 #define CF_READ_PENDING 1
  76 #define CF_WRITE_PENDING 2
  77 #define CF_INIT_PENDING 4
  78 #define CF_IS_OTHERCON 5
  79 #define CF_CLOSE 6
  80 #define CF_APP_LIMITED 7
  81 #define CF_CLOSING 8
  82 #define CF_SHUTDOWN 9
  83 #define CF_CONNECTED 10
  84 #define CF_RECONNECT 11
  85 #define CF_DELAY_CONNECT 12
  86 #define CF_EOF 13
  87         struct list_head writequeue;  /* List of outgoing writequeue_entries */
  88         spinlock_t writequeue_lock;
  89         atomic_t writequeue_cnt;
  90         int retries;
  91 #define MAX_CONNECT_RETRIES 3
  92         struct hlist_node list;
  93         struct connection *othercon;
  94         struct connection *sendcon;
  95         struct work_struct rwork; /* Receive workqueue */
  96         struct work_struct swork; /* Send workqueue */
  97         wait_queue_head_t shutdown_wait; /* wait for graceful shutdown */
  98         unsigned char *rx_buf;
  99         int rx_buflen;
 100         int rx_leftover;
 101         struct rcu_head rcu;
 102 };
 103 #define sock2con(x) ((struct connection *)(x)->sk_user_data)
 104
 105 struct listen_connection {
 106         struct socket *sock;
 107         struct work_struct rwork;
 108 };
 109
 110 #define DLM_WQ_REMAIN_BYTES(e) (PAGE_SIZE - e->end)
 111 #define DLM_WQ_LENGTH_BYTES(e) (e->end - e->offset)
 112
 113 /* An entry waiting to be sent */
 114 struct writequeue_entry {
 115         struct list_head list;
 116         struct page *page;
 117         int offset;
 118         int len;
 119         int end;
 120         int users;
 121         bool dirty;
 122         struct connection *con;
 123         struct list_head msgs;
 124         struct kref ref;
 125 };
 126
 127 struct dlm_msg {
 128         struct writequeue_entry *entry;
 129         struct dlm_msg *orig_msg;
 130         bool retransmit;
 131         void *ppc;
 132         int len;
 133         int idx; /* new()/commit() idx exchange */
 134
 135         struct list_head list;
 136         struct kref ref;
 137 };
 138
 139 struct dlm_node_addr {
 140         struct list_head list;
 141         int nodeid;
 142         int mark;
 143         int addr_count;
 144         int curr_addr_index;
 145         struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT];
 146 };
 147
 148 struct dlm_proto_ops {
 149         bool try_new_addr;
 150         const char *name;
 151         int proto;
 152
 153         int (*connect)(struct connection *con, struct socket *sock,
 154                        struct sockaddr *addr, int addr_len);
 155         void (*sockopts)(struct socket *sock);
 156         int (*bind)(struct socket *sock);
 157         int (*listen_validate)(void);
 158         void (*listen_sockopts)(struct socket *sock);
 159         int (*listen_bind)(struct socket *sock);
 160         /* What to do to shutdown */
 161         void (*shutdown_action)(struct connection *con);
 162         /* What to do to eof check */
 163         bool (*eof_condition)(struct connection *con);
 164 };
 165
 166 static struct listen_sock_callbacks {
 167         void (*sk_error_report)(struct sock *);
 168         void (*sk_data_ready)(struct sock *);
 169         void (*sk_state_change)(struct sock *);
 170         void (*sk_write_space)(struct sock *);
 171 } listen_sock;
 172
 173 static LIST_HEAD(dlm_node_addrs);
 174 static DEFINE_SPINLOCK(dlm_node_addrs_spin);
 175
 176 static struct listen_connection listen_con;
 177 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 178 static int dlm_local_count;
 179 int dlm_allow_conn;
 180
 181 /* Work queues */
 182 static struct workqueue_struct *recv_workqueue;
 183 static struct workqueue_struct *send_workqueue;
 184
 185 static struct hlist_head connection_hash[CONN_HASH_SIZE];
 186 static DEFINE_SPINLOCK(connections_lock);
 187 DEFINE_STATIC_SRCU(connections_srcu);
 188
 189 static const struct dlm_proto_ops *dlm_proto_ops;
 190
 191 static void process_recv_sockets(struct work_struct *work);
 192 static void process_send_sockets(struct work_struct *work);
 193
 194 static void writequeue_entry_ctor(void *data)
 195 {
 196         struct writequeue_entry *entry = data;
 197
 198         INIT_LIST_HEAD(&entry->msgs);
 199 }
 200
 201 struct kmem_cache *dlm_lowcomms_writequeue_cache_create(void)
 202 {
 203         return kmem_cache_create("dlm_writequeue", sizeof(struct writequeue_entry),
 204                                  0, 0, writequeue_entry_ctor);
 205 }
 206
 207 struct kmem_cache *dlm_lowcomms_msg_cache_create(void)
 208 {
 209         return kmem_cache_create("dlm_msg", sizeof(struct dlm_msg), 0, 0, NULL);
 210 }
 211
 212 /* need to held writequeue_lock */
 213 static struct writequeue_entry *con_next_wq(struct connection *con)
 214 {
 215         struct writequeue_entry *e;
 216
 217         if (list_empty(&con->writequeue))
 218                 return NULL;
 219
 220         e = list_first_entry(&con->writequeue, struct writequeue_entry,
 221                              list);
 222         /* if len is zero nothing is to send, if there are users filling
 223          * buffers we wait until the users are done so we can send more.
 224          */
 225         if (e->users || e->len == 0)
 226                 return NULL;
 227
 228         return e;
 229 }
 230
 231 static struct connection *__find_con(int nodeid, int r)
 232 {
 233         struct connection *con;
 234
 235         hlist_for_each_entry_rcu(con, &connection_hash[r], list) {
 236                 if (con->nodeid == nodeid)
 237                         return con;
 238         }
 239
 240         return NULL;
 241 }
 242
 243 static bool tcp_eof_condition(struct connection *con)
 244 {
 245         return atomic_read(&con->writequeue_cnt);
 246 }
 247
 248 static int dlm_con_init(struct connection *con, int nodeid)
 249 {
 250         con->rx_buflen = dlm_config.ci_buffer_size;
 251         con->rx_buf = kmalloc(con->rx_buflen, GFP_NOFS);
 252         if (!con->rx_buf)
 253                 return -ENOMEM;
 254
 255         con->nodeid = nodeid;
 256         mutex_init(&con->sock_mutex);
 257         INIT_LIST_HEAD(&con->writequeue);
 258         spin_lock_init(&con->writequeue_lock);
 259         atomic_set(&con->writequeue_cnt, 0);
 260         INIT_WORK(&con->swork, process_send_sockets);
 261         INIT_WORK(&con->rwork, process_recv_sockets);
 262         init_waitqueue_head(&con->shutdown_wait);
 263
 264         return 0;
 265 }
 266
 267 /*
 268  * If 'allocation' is zero then we don't attempt to create a new
 269  * connection structure for this node.
 270  */
 271 static struct connection *nodeid2con(int nodeid, gfp_t alloc)
 272 {
 273         struct connection *con, *tmp;
 274         int r, ret;
 275
 276         r = nodeid_hash(nodeid);
 277         con = __find_con(nodeid, r);
 278         if (con || !alloc)
 279                 return con;
 280
 281         con = kzalloc(sizeof(*con), alloc);
 282         if (!con)
 283                 return NULL;
 284
 285         ret = dlm_con_init(con, nodeid);
 286         if (ret) {
 287                 kfree(con);
 288                 return NULL;
 289         }
 290
 291         spin_lock(&connections_lock);
 292         /* Because multiple workqueues/threads calls this function it can
 293          * race on multiple cpu's. Instead of locking hot path __find_con()
 294          * we just check in rare cases of recently added nodes again
 295          * under protection of connections_lock. If this is the case we
 296          * abort our connection creation and return the existing connection.
 297          */
 298         tmp = __find_con(nodeid, r);
 299         if (tmp) {
 300                 spin_unlock(&connections_lock);
 301                 kfree(con->rx_buf);
 302                 kfree(con);
 303                 return tmp;
 304         }
 305
 306         hlist_add_head_rcu(&con->list, &connection_hash[r]);
 307         spin_unlock(&connections_lock);
 308
 309         return con;
 310 }
 311
 312 /* Loop round all connections */
 313 static void foreach_conn(void (*conn_func)(struct connection *c))
 314 {
 315         int i;
 316         struct connection *con;
 317
 318         for (i = 0; i < CONN_HASH_SIZE; i++) {
 319                 hlist_for_each_entry_rcu(con, &connection_hash[i], list)
 320                         conn_func(con);
 321         }
 322 }
 323
 324 static struct dlm_node_addr *find_node_addr(int nodeid)
 325 {
 326         struct dlm_node_addr *na;
 327
 328         list_for_each_entry(na, &dlm_node_addrs, list) {
 329                 if (na->nodeid == nodeid)
 330                         return na;
 331         }
 332         return NULL;
 333 }
 334
 335 static int addr_compare(const struct sockaddr_storage *x,
 336                         const struct sockaddr_storage *y)
 337 {
 338         switch (x->ss_family) {
 339         case AF_INET: {
 340                 struct sockaddr_in *sinx = (struct sockaddr_in *)x;
 341                 struct sockaddr_in *siny = (struct sockaddr_in *)y;
 342                 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr)
 343                         return 0;
 344                 if (sinx->sin_port != siny->sin_port)
 345                         return 0;
 346                 break;
 347         }
 348         case AF_INET6: {
 349                 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x;
 350                 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y;
 351                 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr))
 352                         return 0;
 353                 if (sinx->sin6_port != siny->sin6_port)
 354                         return 0;
 355                 break;
 356         }
 357         default:
 358                 return 0;
 359         }
 360         return 1;
 361 }
 362
 363 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out,
 364                           struct sockaddr *sa_out, bool try_new_addr,
 365                           unsigned int *mark)
 366 {
 367         struct sockaddr_storage sas;
 368         struct dlm_node_addr *na;
 369
 370         if (!dlm_local_count)
 371                 return -1;
 372
 373         spin_lock(&dlm_node_addrs_spin);
 374         na = find_node_addr(nodeid);
 375         if (na && na->addr_count) {
 376                 memcpy(&sas, na->addr[na->curr_addr_index],
 377                        sizeof(struct sockaddr_storage));
 378
 379                 if (try_new_addr) {
 380                         na->curr_addr_index++;
 381                         if (na->curr_addr_index == na->addr_count)
 382                                 na->curr_addr_index = 0;
 383                 }
 384         }
 385         spin_unlock(&dlm_node_addrs_spin);
 386
 387         if (!na)
 388                 return -EEXIST;
 389
 390         if (!na->addr_count)
 391                 return -ENOENT;
 392
 393         *mark = na->mark;
 394
 395         if (sas_out)
 396                 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage));
 397
 398         if (!sa_out)
 399                 return 0;
 400
 401         if (dlm_local_addr[0]->ss_family == AF_INET) {
 402                 struct sockaddr_in *in4  = (struct sockaddr_in *) &sas;
 403                 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out;
 404                 ret4->sin_addr.s_addr = in4->sin_addr.s_addr;
 405         } else {
 406                 struct sockaddr_in6 *in6  = (struct sockaddr_in6 *) &sas;
 407                 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out;
 408                 ret6->sin6_addr = in6->sin6_addr;
 409         }
 410
 411         return 0;
 412 }
 413
 414 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid,
 415                           unsigned int *mark)
 416 {
 417         struct dlm_node_addr *na;
 418         int rv = -EEXIST;
 419         int addr_i;
 420
 421         spin_lock(&dlm_node_addrs_spin);
 422         list_for_each_entry(na, &dlm_node_addrs, list) {
 423                 if (!na->addr_count)
 424                         continue;
 425
 426                 for (addr_i = 0; addr_i < na->addr_count; addr_i++) {
 427                         if (addr_compare(na->addr[addr_i], addr)) {
 428                                 *nodeid = na->nodeid;
 429                                 *mark = na->mark;
 430                                 rv = 0;
 431                                 goto unlock;
 432                         }
 433                 }
 434         }
 435 unlock:
 436         spin_unlock(&dlm_node_addrs_spin);
 437         return rv;
 438 }
 439
 440 /* caller need to held dlm_node_addrs_spin lock */
 441 static bool dlm_lowcomms_na_has_addr(const struct dlm_node_addr *na,
 442                                      const struct sockaddr_storage *addr)
 443 {
 444         int i;
 445
 446         for (i = 0; i < na->addr_count; i++) {
 447                 if (addr_compare(na->addr[i], addr))
 448                         return true;
 449         }
 450
 451         return false;
 452 }
 453
 454 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len)
 455 {
 456         struct sockaddr_storage *new_addr;
 457         struct dlm_node_addr *new_node, *na;
 458         bool ret;
 459
 460         new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS);
 461         if (!new_node)
 462                 return -ENOMEM;
 463
 464         new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS);
 465         if (!new_addr) {
 466                 kfree(new_node);
 467                 return -ENOMEM;
 468         }
 469
 470         memcpy(new_addr, addr, len);
 471
 472         spin_lock(&dlm_node_addrs_spin);
 473         na = find_node_addr(nodeid);
 474         if (!na) {
 475                 new_node->nodeid = nodeid;
 476                 new_node->addr[0] = new_addr;
 477                 new_node->addr_count = 1;
 478                 new_node->mark = dlm_config.ci_mark;
 479                 list_add(&new_node->list, &dlm_node_addrs);
 480                 spin_unlock(&dlm_node_addrs_spin);
 481                 return 0;
 482         }
 483
 484         ret = dlm_lowcomms_na_has_addr(na, addr);
 485         if (ret) {
 486                 spin_unlock(&dlm_node_addrs_spin);
 487                 kfree(new_addr);
 488                 kfree(new_node);
 489                 return -EEXIST;
 490         }
 491
 492         if (na->addr_count >= DLM_MAX_ADDR_COUNT) {
 493                 spin_unlock(&dlm_node_addrs_spin);
 494                 kfree(new_addr);
 495                 kfree(new_node);
 496                 return -ENOSPC;
 497         }
 498
 499         na->addr[na->addr_count++] = new_addr;
 500         spin_unlock(&dlm_node_addrs_spin);
 501         kfree(new_node);
 502         return 0;
 503 }
 504
 505 /* Data available on socket or listen socket received a connect */
 506 static void lowcomms_data_ready(struct sock *sk)
 507 {
 508         struct connection *con;
 509
 510         con = sock2con(sk);
 511         if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags))
 512                 queue_work(recv_workqueue, &con->rwork);
 513 }
 514
 515 static void lowcomms_listen_data_ready(struct sock *sk)
 516 {
 517         if (!dlm_allow_conn)
 518                 return;
 519
 520         queue_work(recv_workqueue, &listen_con.rwork);
 521 }
 522
 523 static void lowcomms_write_space(struct sock *sk)
 524 {
 525         struct connection *con;
 526
 527         con = sock2con(sk);
 528         if (!con)
 529                 return;
 530
 531         if (!test_and_set_bit(CF_CONNECTED, &con->flags)) {
 532                 log_print("successful connected to node %d", con->nodeid);
 533                 queue_work(send_workqueue, &con->swork);
 534                 return;
 535         }
 536
 537         clear_bit(SOCK_NOSPACE, &con->sock->flags);
 538
 539         if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) {
 540                 con->sock->sk->sk_write_pending--;
 541                 clear_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags);
 542         }
 543
 544         queue_work(send_workqueue, &con->swork);
 545 }
 546
 547 static inline void lowcomms_connect_sock(struct connection *con)
 548 {
 549         if (test_bit(CF_CLOSE, &con->flags))
 550                 return;
 551         queue_work(send_workqueue, &con->swork);
 552         cond_resched();
 553 }
 554
 555 static void lowcomms_state_change(struct sock *sk)
 556 {
 557         /* SCTP layer is not calling sk_data_ready when the connection
 558          * is done, so we catch the signal through here. Also, it
 559          * doesn't switch socket state when entering shutdown, so we
 560          * skip the write in that case.
 561          */
 562         if (sk->sk_shutdown) {
 563                 if (sk->sk_shutdown == RCV_SHUTDOWN)
 564                         lowcomms_data_ready(sk);
 565         } else if (sk->sk_state == TCP_ESTABLISHED) {
 566                 lowcomms_write_space(sk);
 567         }
 568 }
 569
 570 int dlm_lowcomms_connect_node(int nodeid)
 571 {
 572         struct connection *con;
 573         int idx;
 574
 575         if (nodeid == dlm_our_nodeid())
 576                 return 0;
 577
 578         idx = srcu_read_lock(&connections_srcu);
 579         con = nodeid2con(nodeid, GFP_NOFS);
 580         if (!con) {
 581                 srcu_read_unlock(&connections_srcu, idx);
 582                 return -ENOMEM;
 583         }
 584
 585         lowcomms_connect_sock(con);
 586         srcu_read_unlock(&connections_srcu, idx);
 587
 588         return 0;
 589 }
 590
 591 int dlm_lowcomms_nodes_set_mark(int nodeid, unsigned int mark)
 592 {
 593         struct dlm_node_addr *na;
 594
 595         spin_lock(&dlm_node_addrs_spin);
 596         na = find_node_addr(nodeid);
 597         if (!na) {
 598                 spin_unlock(&dlm_node_addrs_spin);
 599                 return -ENOENT;
 600         }
 601
 602         na->mark = mark;
 603         spin_unlock(&dlm_node_addrs_spin);
 604
 605         return 0;
 606 }
 607
 608 static void lowcomms_error_report(struct sock *sk)
 609 {
 610         struct connection *con;
 611         void (*orig_report)(struct sock *) = NULL;
 612         struct inet_sock *inet;
 613
 614         con = sock2con(sk);
 615         if (con == NULL)
 616                 goto out;
 617
 618         orig_report = listen_sock.sk_error_report;
 619
 620         inet = inet_sk(sk);
 621         switch (sk->sk_family) {
 622         case AF_INET:
 623                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 624                                    "sending to node %d at %pI4, dport %d, "
 625                                    "sk_err=%d/%d\n", dlm_our_nodeid(),
 626                                    con->nodeid, &inet->inet_daddr,
 627                                    ntohs(inet->inet_dport), sk->sk_err,
 628                                    sk->sk_err_soft);
 629                 break;
 630 #if IS_ENABLED(CONFIG_IPV6)
 631         case AF_INET6:
 632                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 633                                    "sending to node %d at %pI6c, "
 634                                    "dport %d, sk_err=%d/%d\n", dlm_our_nodeid(),
 635                                    con->nodeid, &sk->sk_v6_daddr,
 636                                    ntohs(inet->inet_dport), sk->sk_err,
 637                                    sk->sk_err_soft);
 638                 break;
 639 #endif
 640         default:
 641                 printk_ratelimited(KERN_ERR "dlm: node %d: socket error "
 642                                    "invalid socket family %d set, "
 643                                    "sk_err=%d/%d\n", dlm_our_nodeid(),
 644                                    sk->sk_family, sk->sk_err, sk->sk_err_soft);
 645                 goto out;
 646         }
 647
 648         /* below sendcon only handling */
 649         if (test_bit(CF_IS_OTHERCON, &con->flags))
 650                 con = con->sendcon;
 651
 652         switch (sk->sk_err) {
 653         case ECONNREFUSED:
 654                 set_bit(CF_DELAY_CONNECT, &con->flags);
 655                 break;
 656         default:
 657                 break;
 658         }
 659
 660         if (!test_and_set_bit(CF_RECONNECT, &con->flags))
 661                 queue_work(send_workqueue, &con->swork);
 662
 663 out:
 664         if (orig_report)
 665                 orig_report(sk);
 666 }
 667
 668 /* Note: sk_callback_lock must be locked before calling this function. */
 669 static void save_listen_callbacks(struct socket *sock)
 670 {
 671         struct sock *sk = sock->sk;
 672
 673         listen_sock.sk_data_ready = sk->sk_data_ready;
 674         listen_sock.sk_state_change = sk->sk_state_change;
 675         listen_sock.sk_write_space = sk->sk_write_space;
 676         listen_sock.sk_error_report = sk->sk_error_report;
 677 }
 678
 679 static void restore_callbacks(struct socket *sock)
 680 {
 681         struct sock *sk = sock->sk;
 682
 683         lock_sock(sk);
 684         sk->sk_user_data = NULL;
 685         sk->sk_data_ready = listen_sock.sk_data_ready;
 686         sk->sk_state_change = listen_sock.sk_state_change;
 687         sk->sk_write_space = listen_sock.sk_write_space;
 688         sk->sk_error_report = listen_sock.sk_error_report;
 689         release_sock(sk);
 690 }
 691
 692 static void add_listen_sock(struct socket *sock, struct listen_connection *con)
 693 {
 694         struct sock *sk = sock->sk;
 695
 696         lock_sock(sk);
 697         save_listen_callbacks(sock);
 698         con->sock = sock;
 699
 700         sk->sk_user_data = con;
 701         sk->sk_allocation = GFP_NOFS;
 702         /* Install a data_ready callback */
 703         sk->sk_data_ready = lowcomms_listen_data_ready;
 704         release_sock(sk);
 705 }
 706
 707 /* Make a socket active */
 708 static void add_sock(struct socket *sock, struct connection *con)
 709 {
 710         struct sock *sk = sock->sk;
 711
 712         lock_sock(sk);
 713         con->sock = sock;
 714
 715         sk->sk_user_data = con;
 716         /* Install a data_ready callback */
 717         sk->sk_data_ready = lowcomms_data_ready;
 718         sk->sk_write_space = lowcomms_write_space;
 719         sk->sk_state_change = lowcomms_state_change;
 720         sk->sk_allocation = GFP_NOFS;
 721         sk->sk_error_report = lowcomms_error_report;
 722         release_sock(sk);
 723 }
 724
 725 /* Add the port number to an IPv6 or 4 sockaddr and return the address
 726    length */
 727 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 728                           int *addr_len)
 729 {
 730         saddr->ss_family =  dlm_local_addr[0]->ss_family;
 731         if (saddr->ss_family == AF_INET) {
 732                 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr;
 733                 in4_addr->sin_port = cpu_to_be16(port);
 734                 *addr_len = sizeof(struct sockaddr_in);
 735                 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero));
 736         } else {
 737                 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr;
 738                 in6_addr->sin6_port = cpu_to_be16(port);
 739                 *addr_len = sizeof(struct sockaddr_in6);
 740         }
 741         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 742 }
 743
 744 static void dlm_page_release(struct kref *kref)
 745 {
 746         struct writequeue_entry *e = container_of(kref, struct writequeue_entry,
 747                                                   ref);
 748
 749         __free_page(e->page);
 750         dlm_free_writequeue(e);
 751 }
 752
 753 static void dlm_msg_release(struct kref *kref)
 754 {
 755         struct dlm_msg *msg = container_of(kref, struct dlm_msg, ref);
 756
 757         kref_put(&msg->entry->ref, dlm_page_release);
 758         dlm_free_msg(msg);
 759 }
 760
 761 static void free_entry(struct writequeue_entry *e)
 762 {
 763         struct dlm_msg *msg, *tmp;
 764
 765         list_for_each_entry_safe(msg, tmp, &e->msgs, list) {
 766                 if (msg->orig_msg) {
 767                         msg->orig_msg->retransmit = false;
 768                         kref_put(&msg->orig_msg->ref, dlm_msg_release);
 769                 }
 770
 771                 list_del(&msg->list);
 772                 kref_put(&msg->ref, dlm_msg_release);
 773         }
 774
 775         list_del(&e->list);
 776         atomic_dec(&e->con->writequeue_cnt);
 777         kref_put(&e->ref, dlm_page_release);
 778 }
 779
 780 static void dlm_close_sock(struct socket **sock)
 781 {
 782         if (*sock) {
 783                 restore_callbacks(*sock);
 784                 sock_release(*sock);
 785                 *sock = NULL;
 786         }
 787 }
 788
 789 /* Close a remote connection and tidy up */
 790 static void close_connection(struct connection *con, bool and_other,
 791                              bool tx, bool rx)
 792 {
 793         bool closing = test_and_set_bit(CF_CLOSING, &con->flags);
 794         struct writequeue_entry *e;
 795
 796         if (tx && !closing && cancel_work_sync(&con->swork)) {
 797                 log_print("canceled swork for node %d", con->nodeid);
 798                 clear_bit(CF_WRITE_PENDING, &con->flags);
 799         }
 800         if (rx && !closing && cancel_work_sync(&con->rwork)) {
 801                 log_print("canceled rwork for node %d", con->nodeid);
 802                 clear_bit(CF_READ_PENDING, &con->flags);
 803         }
 804
 805         mutex_lock(&con->sock_mutex);
 806         dlm_close_sock(&con->sock);
 807
 808         if (con->othercon && and_other) {
 809                 /* Will only re-enter once. */
 810                 close_connection(con->othercon, false, tx, rx);
 811         }
 812
 813         /* if we send a writequeue entry only a half way, we drop the
 814          * whole entry because reconnection and that we not start of the
 815          * middle of a msg which will confuse the other end.
 816          *
 817          * we can always drop messages because retransmits, but what we
 818          * cannot allow is to transmit half messages which may be processed
 819          * at the other side.
 820          *
 821          * our policy is to start on a clean state when disconnects, we don't
 822          * know what's send/received on transport layer in this case.
 823          */
 824         spin_lock(&con->writequeue_lock);
 825         if (!list_empty(&con->writequeue)) {
 826                 e = list_first_entry(&con->writequeue, struct writequeue_entry,
 827                                      list);
 828                 if (e->dirty)
 829                         free_entry(e);
 830         }
 831         spin_unlock(&con->writequeue_lock);
 832
 833         con->rx_leftover = 0;
 834         con->retries = 0;
 835         clear_bit(CF_APP_LIMITED, &con->flags);
 836         clear_bit(CF_CONNECTED, &con->flags);
 837         clear_bit(CF_DELAY_CONNECT, &con->flags);
 838         clear_bit(CF_RECONNECT, &con->flags);
 839         clear_bit(CF_EOF, &con->flags);
 840         mutex_unlock(&con->sock_mutex);
 841         clear_bit(CF_CLOSING, &con->flags);
 842 }
 843
 844 static void shutdown_connection(struct connection *con)
 845 {
 846         int ret;
 847
 848         flush_work(&con->swork);
 849
 850         mutex_lock(&con->sock_mutex);
 851         /* nothing to shutdown */
 852         if (!con->sock) {
 853                 mutex_unlock(&con->sock_mutex);
 854                 return;
 855         }
 856
 857         set_bit(CF_SHUTDOWN, &con->flags);
 858         ret = kernel_sock_shutdown(con->sock, SHUT_WR);
 859         mutex_unlock(&con->sock_mutex);
 860         if (ret) {
 861                 log_print("Connection %p failed to shutdown: %d will force close",
 862                           con, ret);
 863                 goto force_close;
 864         } else {
 865                 ret = wait_event_timeout(con->shutdown_wait,
 866                                          !test_bit(CF_SHUTDOWN, &con->flags),
 867                                          DLM_SHUTDOWN_WAIT_TIMEOUT);
 868                 if (ret == 0) {
 869                         log_print("Connection %p shutdown timed out, will force close",
 870                                   con);
 871                         goto force_close;
 872                 }
 873         }
 874
 875         return;
 876
 877 force_close:
 878         clear_bit(CF_SHUTDOWN, &con->flags);
 879         close_connection(con, false, true, true);
 880 }
 881
 882 static void dlm_tcp_shutdown(struct connection *con)
 883 {
 884         if (con->othercon)
 885                 shutdown_connection(con->othercon);
 886         shutdown_connection(con);
 887 }
 888
 889 static int con_realloc_receive_buf(struct connection *con, int newlen)
 890 {
 891         unsigned char *newbuf;
 892
 893         newbuf = kmalloc(newlen, GFP_NOFS);
 894         if (!newbuf)
 895                 return -ENOMEM;
 896
 897         /* copy any leftover from last receive */
 898         if (con->rx_leftover)
 899                 memmove(newbuf, con->rx_buf, con->rx_leftover);
 900
 901         /* swap to new buffer space */
 902         kfree(con->rx_buf);
 903         con->rx_buflen = newlen;
 904         con->rx_buf = newbuf;
 905
 906         return 0;
 907 }
 908
 909 /* Data received from remote end */
 910 static int receive_from_sock(struct connection *con)
 911 {
 912         struct msghdr msg;
 913         struct kvec iov;
 914         int ret, buflen;
 915
 916         mutex_lock(&con->sock_mutex);
 917
 918         if (con->sock == NULL) {
 919                 ret = -EAGAIN;
 920                 goto out_close;
 921         }
 922
 923         /* realloc if we get new buffer size to read out */
 924         buflen = dlm_config.ci_buffer_size;
 925         if (con->rx_buflen != buflen && con->rx_leftover <= buflen) {
 926                 ret = con_realloc_receive_buf(con, buflen);
 927                 if (ret < 0)
 928                         goto out_resched;
 929         }
 930
 931         for (;;) {
 932                 /* calculate new buffer parameter regarding last receive and
 933                  * possible leftover bytes
 934                  */
 935                 iov.iov_base = con->rx_buf + con->rx_leftover;
 936                 iov.iov_len = con->rx_buflen - con->rx_leftover;
 937
 938                 memset(&msg, 0, sizeof(msg));
 939                 msg.msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
 940                 ret = kernel_recvmsg(con->sock, &msg, &iov, 1, iov.iov_len,
 941                                      msg.msg_flags);
 942                 trace_dlm_recv(con->nodeid, ret);
 943                 if (ret == -EAGAIN)
 944                         break;
 945                 else if (ret <= 0)
 946                         goto out_close;
 947
 948                 /* new buflen according readed bytes and leftover from last receive */
 949                 buflen = ret + con->rx_leftover;
 950                 ret = dlm_process_incoming_buffer(con->nodeid, con->rx_buf, buflen);
 951                 if (ret < 0)
 952                         goto out_close;
 953
 954                 /* calculate leftover bytes from process and put it into begin of
 955                  * the receive buffer, so next receive we have the full message
 956                  * at the start address of the receive buffer.
 957                  */
 958                 con->rx_leftover = buflen - ret;
 959                 if (con->rx_leftover) {
 960                         memmove(con->rx_buf, con->rx_buf + ret,
 961                                 con->rx_leftover);
 962                 }
 963         }
 964
 965         dlm_midcomms_receive_done(con->nodeid);
 966         mutex_unlock(&con->sock_mutex);
 967         return 0;
 968
 969 out_resched:
 970         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 971                 queue_work(recv_workqueue, &con->rwork);
 972         mutex_unlock(&con->sock_mutex);
 973         return -EAGAIN;
 974
 975 out_close:
 976         if (ret == 0) {
 977                 log_print("connection %p got EOF from %d",
 978                           con, con->nodeid);
 979
 980                 if (dlm_proto_ops->eof_condition &&
 981                     dlm_proto_ops->eof_condition(con)) {
 982                         set_bit(CF_EOF, &con->flags);
 983                         mutex_unlock(&con->sock_mutex);
 984                 } else {
 985                         mutex_unlock(&con->sock_mutex);
 986                         close_connection(con, false, true, false);
 987
 988                         /* handling for tcp shutdown */
 989                         clear_bit(CF_SHUTDOWN, &con->flags);
 990                         wake_up(&con->shutdown_wait);
 991                 }
 992
 993                 /* signal to breaking receive worker */
 994                 ret = -1;
 995         } else {
 996                 mutex_unlock(&con->sock_mutex);
 997         }
 998         return ret;
 999 }
1000
1001 /* Listening socket is busy, accept a connection */
1002 static int accept_from_sock(struct listen_connection *con)
1003 {
1004         int result;
1005         struct sockaddr_storage peeraddr;
1006         struct socket *newsock;
1007         int len, idx;
1008         int nodeid;
1009         struct connection *newcon;
1010         struct connection *addcon;
1011         unsigned int mark;
1012
1013         if (!con->sock)
1014                 return -ENOTCONN;
1015
1016         result = kernel_accept(con->sock, &newsock, O_NONBLOCK);
1017         if (result < 0)
1018                 goto accept_err;
1019
1020         /* Get the connected socket's peer */
1021         memset(&peeraddr, 0, sizeof(peeraddr));
1022         len = newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr, 2);
1023         if (len < 0) {
1024                 result = -ECONNABORTED;
1025                 goto accept_err;
1026         }
1027
1028         /* Get the new node's NODEID */
1029         make_sockaddr(&peeraddr, 0, &len);
1030         if (addr_to_nodeid(&peeraddr, &nodeid, &mark)) {
1031                 switch (peeraddr.ss_family) {
1032                 case AF_INET: {
1033                         struct sockaddr_in *sin = (struct sockaddr_in *)&peeraddr;
1034
1035                         log_print("connect from non cluster IPv4 node %pI4",
1036                                   &sin->sin_addr);
1037                         break;
1038                 }
1039 #if IS_ENABLED(CONFIG_IPV6)
1040                 case AF_INET6: {
1041                         struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&peeraddr;
1042
1043                         log_print("connect from non cluster IPv6 node %pI6c",
1044                                   &sin6->sin6_addr);
1045                         break;
1046                 }
1047 #endif
1048                 default:
1049                         log_print("invalid family from non cluster node");
1050                         break;
1051                 }
1052
1053                 sock_release(newsock);
1054                 return -1;
1055         }
1056
1057         log_print("got connection from %d", nodeid);
1058
1059         /*  Check to see if we already have a connection to this node. This
1060          *  could happen if the two nodes initiate a connection at roughly
1061          *  the same time and the connections cross on the wire.
1062          *  In this case we store the incoming one in "othercon"
1063          */
1064         idx = srcu_read_lock(&connections_srcu);
1065         newcon = nodeid2con(nodeid, GFP_NOFS);
1066         if (!newcon) {
1067                 srcu_read_unlock(&connections_srcu, idx);
1068                 result = -ENOMEM;
1069                 goto accept_err;
1070         }
1071
1072         sock_set_mark(newsock->sk, mark);
1073
1074         mutex_lock(&newcon->sock_mutex);
1075         if (newcon->sock) {
1076                 struct connection *othercon = newcon->othercon;
1077
1078                 if (!othercon) {
1079                         othercon = kzalloc(sizeof(*othercon), GFP_NOFS);
1080                         if (!othercon) {
1081                                 log_print("failed to allocate incoming socket");
1082                                 mutex_unlock(&newcon->sock_mutex);
1083                                 srcu_read_unlock(&connections_srcu, idx);
1084                                 result = -ENOMEM;
1085                                 goto accept_err;
1086                         }
1087
1088                         result = dlm_con_init(othercon, nodeid);
1089                         if (result < 0) {
1090                                 kfree(othercon);
1091                                 mutex_unlock(&newcon->sock_mutex);
1092                                 srcu_read_unlock(&connections_srcu, idx);
1093                                 goto accept_err;
1094                         }
1095
1096                         lockdep_set_subclass(&othercon->sock_mutex, 1);
1097                         set_bit(CF_IS_OTHERCON, &othercon->flags);
1098                         newcon->othercon = othercon;
1099                         othercon->sendcon = newcon;
1100                 } else {
1101                         /* close other sock con if we have something new */
1102                         close_connection(othercon, false, true, false);
1103                 }
1104
1105                 mutex_lock(&othercon->sock_mutex);
1106                 add_sock(newsock, othercon);
1107                 addcon = othercon;
1108                 mutex_unlock(&othercon->sock_mutex);
1109         }
1110         else {
1111                 /* accept copies the sk after we've saved the callbacks, so we
1112                    don't want to save them a second time or comm errors will
1113                    result in calling sk_error_report recursively. */
1114                 add_sock(newsock, newcon);
1115                 addcon = newcon;
1116         }
1117
1118         set_bit(CF_CONNECTED, &addcon->flags);
1119         mutex_unlock(&newcon->sock_mutex);
1120
1121         /*
1122          * Add it to the active queue in case we got data
1123          * between processing the accept adding the socket
1124          * to the read_sockets list
1125          */
1126         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
1127                 queue_work(recv_workqueue, &addcon->rwork);
1128
1129         srcu_read_unlock(&connections_srcu, idx);
1130
1131         return 0;
1132
1133 accept_err:
1134         if (newsock)
1135                 sock_release(newsock);
1136
1137         if (result != -EAGAIN)
1138                 log_print("error accepting connection from node: %d", result);
1139         return result;
1140 }
1141
1142 /*
1143  * writequeue_entry_complete - try to delete and free write queue entry
1144  * @e: write queue entry to try to delete
1145  * @completed: bytes completed
1146  *
1147  * writequeue_lock must be held.
1148  */
1149 static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
1150 {
1151         e->offset += completed;
1152         e->len -= completed;
1153         /* signal that page was half way transmitted */
1154         e->dirty = true;
1155
1156         if (e->len == 0 && e->users == 0)
1157                 free_entry(e);
1158 }
1159
1160 /*
1161  * sctp_bind_addrs - bind a SCTP socket to all our addresses
1162  */
1163 static int sctp_bind_addrs(struct socket *sock, uint16_t port)
1164 {
1165         struct sockaddr_storage localaddr;
1166         struct sockaddr *addr = (struct sockaddr *)&localaddr;
1167         int i, addr_len, result = 0;
1168
1169         for (i = 0; i < dlm_local_count; i++) {
1170                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1171                 make_sockaddr(&localaddr, port, &addr_len);
1172
1173                 if (!i)
1174                         result = kernel_bind(sock, addr, addr_len);
1175                 else
1176                         result = sock_bind_add(sock->sk, addr, addr_len);
1177
1178                 if (result < 0) {
1179                         log_print("Can't bind to %d addr number %d, %d.\n",
1180                                   port, i + 1, result);
1181                         break;
1182                 }
1183         }
1184         return result;
1185 }
1186
1187 /* Get local addresses */
1188 static void init_local(void)
1189 {
1190         struct sockaddr_storage sas, *addr;
1191         int i;
1192
1193         dlm_local_count = 0;
1194         for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) {
1195                 if (dlm_our_addr(&sas, i))
1196                         break;
1197
1198                 addr = kmemdup(&sas, sizeof(*addr), GFP_NOFS);
1199                 if (!addr)
1200                         break;
1201                 dlm_local_addr[dlm_local_count++] = addr;
1202         }
1203 }
1204
1205 static void deinit_local(void)
1206 {
1207         int i;
1208
1209         for (i = 0; i < dlm_local_count; i++)
1210                 kfree(dlm_local_addr[i]);
1211 }
1212
1213 static struct writequeue_entry *new_writequeue_entry(struct connection *con)
1214 {
1215         struct writequeue_entry *entry;
1216
1217         entry = dlm_allocate_writequeue();
1218         if (!entry)
1219                 return NULL;
1220
1221         entry->page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
1222         if (!entry->page) {
1223                 dlm_free_writequeue(entry);
1224                 return NULL;
1225         }
1226
1227         entry->offset = 0;
1228         entry->len = 0;
1229         entry->end = 0;
1230         entry->dirty = false;
1231         entry->con = con;
1232         entry->users = 1;
1233         kref_init(&entry->ref);
1234         return entry;
1235 }
1236
1237 static struct writequeue_entry *new_wq_entry(struct connection *con, int len,
1238                                              char **ppc, void (*cb)(void *data),
1239                                              void *data)
1240 {
1241         struct writequeue_entry *e;
1242
1243         spin_lock(&con->writequeue_lock);
1244         if (!list_empty(&con->writequeue)) {
1245                 e = list_last_entry(&con->writequeue, struct writequeue_entry, list);
1246                 if (DLM_WQ_REMAIN_BYTES(e) >= len) {
1247                         kref_get(&e->ref);
1248
1249                         *ppc = page_address(e->page) + e->end;
1250                         if (cb)
1251                                 cb(data);
1252
1253                         e->end += len;
1254                         e->users++;
1255                         goto out;
1256                 }
1257         }
1258
1259         e = new_writequeue_entry(con);
1260         if (!e)
1261                 goto out;
1262
1263         kref_get(&e->ref);
1264         *ppc = page_address(e->page);
1265         e->end += len;
1266         atomic_inc(&con->writequeue_cnt);
1267         if (cb)
1268                 cb(data);
1269
1270         list_add_tail(&e->list, &con->writequeue);
1271
1272 out:
1273         spin_unlock(&con->writequeue_lock);
1274         return e;
1275 };
1276
1277 static struct dlm_msg *dlm_lowcomms_new_msg_con(struct connection *con, int len,
1278                                                 gfp_t allocation, char **ppc,
1279                                                 void (*cb)(void *data),
1280                                                 void *data)
1281 {
1282         struct writequeue_entry *e;
1283         struct dlm_msg *msg;
1284
1285         msg = dlm_allocate_msg(allocation);
1286         if (!msg)
1287                 return NULL;
1288
1289         kref_init(&msg->ref);
1290
1291         e = new_wq_entry(con, len, ppc, cb, data);
1292         if (!e) {
1293                 dlm_free_msg(msg);
1294                 return NULL;
1295         }
1296
1297         msg->retransmit = false;
1298         msg->orig_msg = NULL;
1299         msg->ppc = *ppc;
1300         msg->len = len;
1301         msg->entry = e;
1302
1303         return msg;
1304 }
1305
1306 struct dlm_msg *dlm_lowcomms_new_msg(int nodeid, int len, gfp_t allocation,
1307                                      char **ppc, void (*cb)(void *data),
1308                                      void *data)
1309 {
1310         struct connection *con;
1311         struct dlm_msg *msg;
1312         int idx;
1313
1314         if (len > DLM_MAX_SOCKET_BUFSIZE ||
1315             len < sizeof(struct dlm_header)) {
1316                 BUILD_BUG_ON(PAGE_SIZE < DLM_MAX_SOCKET_BUFSIZE);
1317                 log_print("failed to allocate a buffer of size %d", len);
1318                 WARN_ON(1);
1319                 return NULL;
1320         }
1321
1322         idx = srcu_read_lock(&connections_srcu);
1323         con = nodeid2con(nodeid, allocation);
1324         if (!con) {
1325                 srcu_read_unlock(&connections_srcu, idx);
1326                 return NULL;
1327         }
1328
1329         msg = dlm_lowcomms_new_msg_con(con, len, allocation, ppc, cb, data);
1330         if (!msg) {
1331                 srcu_read_unlock(&connections_srcu, idx);
1332                 return NULL;
1333         }
1334
1335         /* we assume if successful commit must called */
1336         msg->idx = idx;
1337         return msg;
1338 }
1339
1340 static void _dlm_lowcomms_commit_msg(struct dlm_msg *msg)
1341 {
1342         struct writequeue_entry *e = msg->entry;
1343         struct connection *con = e->con;
1344         int users;
1345
1346         spin_lock(&con->writequeue_lock);
1347         kref_get(&msg->ref);
1348         list_add(&msg->list, &e->msgs);
1349
1350         users = --e->users;
1351         if (users)
1352                 goto out;
1353
1354         e->len = DLM_WQ_LENGTH_BYTES(e);
1355         spin_unlock(&con->writequeue_lock);
1356
1357         queue_work(send_workqueue, &con->swork);
1358         return;
1359
1360 out:
1361         spin_unlock(&con->writequeue_lock);
1362         return;
1363 }
1364
1365 void dlm_lowcomms_commit_msg(struct dlm_msg *msg)
1366 {
1367         _dlm_lowcomms_commit_msg(msg);
1368         srcu_read_unlock(&connections_srcu, msg->idx);
1369 }
1370
1371 void dlm_lowcomms_put_msg(struct dlm_msg *msg)
1372 {
1373         kref_put(&msg->ref, dlm_msg_release);
1374 }
1375
1376 /* does not held connections_srcu, usage workqueue only */
1377 int dlm_lowcomms_resend_msg(struct dlm_msg *msg)
1378 {
1379         struct dlm_msg *msg_resend;
1380         char *ppc;
1381
1382         if (msg->retransmit)
1383                 return 1;
1384
1385         msg_resend = dlm_lowcomms_new_msg_con(msg->entry->con, msg->len,
1386                                               GFP_ATOMIC, &ppc, NULL, NULL);
1387         if (!msg_resend)
1388                 return -ENOMEM;
1389
1390         msg->retransmit = true;
1391         kref_get(&msg->ref);
1392         msg_resend->orig_msg = msg;
1393
1394         memcpy(ppc, msg->ppc, msg->len);
1395         _dlm_lowcomms_commit_msg(msg_resend);
1396         dlm_lowcomms_put_msg(msg_resend);
1397
1398         return 0;
1399 }
1400
1401 /* Send a message */
1402 static void send_to_sock(struct connection *con)
1403 {
1404         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1405         struct writequeue_entry *e;
1406         int len, offset, ret;
1407         int count = 0;
1408
1409         mutex_lock(&con->sock_mutex);
1410         if (con->sock == NULL)
1411                 goto out_connect;
1412
1413         spin_lock(&con->writequeue_lock);
1414         for (;;) {
1415                 e = con_next_wq(con);
1416                 if (!e)
1417                         break;
1418
1419                 len = e->len;
1420                 offset = e->offset;
1421                 BUG_ON(len == 0 && e->users == 0);
1422                 spin_unlock(&con->writequeue_lock);
1423
1424                 ret = kernel_sendpage(con->sock, e->page, offset, len,
1425                                       msg_flags);
1426                 trace_dlm_send(con->nodeid, ret);
1427                 if (ret == -EAGAIN || ret == 0) {
1428                         if (ret == -EAGAIN &&
1429                             test_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags) &&
1430                             !test_and_set_bit(CF_APP_LIMITED, &con->flags)) {
1431                                 /* Notify TCP that we're limited by the
1432                                  * application window size.
1433                                  */
1434                                 set_bit(SOCK_NOSPACE, &con->sock->flags);
1435                                 con->sock->sk->sk_write_pending++;
1436                         }
1437                         cond_resched();
1438                         goto out;
1439                 } else if (ret < 0)
1440                         goto out;
1441
1442                 /* Don't starve people filling buffers */
1443                 if (++count >= MAX_SEND_MSG_COUNT) {
1444                         cond_resched();
1445                         count = 0;
1446                 }
1447
1448                 spin_lock(&con->writequeue_lock);
1449                 writequeue_entry_complete(e, ret);
1450         }
1451         spin_unlock(&con->writequeue_lock);
1452
1453         /* close if we got EOF */
1454         if (test_and_clear_bit(CF_EOF, &con->flags)) {
1455                 mutex_unlock(&con->sock_mutex);
1456                 close_connection(con, false, false, true);
1457
1458                 /* handling for tcp shutdown */
1459                 clear_bit(CF_SHUTDOWN, &con->flags);
1460                 wake_up(&con->shutdown_wait);
1461         } else {
1462                 mutex_unlock(&con->sock_mutex);
1463         }
1464
1465         return;
1466
1467 out:
1468         mutex_unlock(&con->sock_mutex);
1469         return;
1470
1471 out_connect:
1472         mutex_unlock(&con->sock_mutex);
1473         queue_work(send_workqueue, &con->swork);
1474         cond_resched();
1475 }
1476
1477 static void clean_one_writequeue(struct connection *con)
1478 {
1479         struct writequeue_entry *e, *safe;
1480
1481         spin_lock(&con->writequeue_lock);
1482         list_for_each_entry_safe(e, safe, &con->writequeue, list) {
1483                 free_entry(e);
1484         }
1485         spin_unlock(&con->writequeue_lock);
1486 }
1487
1488 /* Called from recovery when it knows that a node has
1489    left the cluster */
1490 int dlm_lowcomms_close(int nodeid)
1491 {
1492         struct connection *con;
1493         struct dlm_node_addr *na;
1494         int idx;
1495
1496         log_print("closing connection to node %d", nodeid);
1497         idx = srcu_read_lock(&connections_srcu);
1498         con = nodeid2con(nodeid, 0);
1499         if (con) {
1500                 set_bit(CF_CLOSE, &con->flags);
1501                 close_connection(con, true, true, true);
1502                 clean_one_writequeue(con);
1503                 if (con->othercon)
1504                         clean_one_writequeue(con->othercon);
1505         }
1506         srcu_read_unlock(&connections_srcu, idx);
1507
1508         spin_lock(&dlm_node_addrs_spin);
1509         na = find_node_addr(nodeid);
1510         if (na) {
1511                 list_del(&na->list);
1512                 while (na->addr_count--)
1513                         kfree(na->addr[na->addr_count]);
1514                 kfree(na);
1515         }
1516         spin_unlock(&dlm_node_addrs_spin);
1517
1518         return 0;
1519 }
1520
1521 /* Receive workqueue function */
1522 static void process_recv_sockets(struct work_struct *work)
1523 {
1524         struct connection *con = container_of(work, struct connection, rwork);
1525
1526         clear_bit(CF_READ_PENDING, &con->flags);
1527         receive_from_sock(con);
1528 }
1529
1530 static void process_listen_recv_socket(struct work_struct *work)
1531 {
1532         accept_from_sock(&listen_con);
1533 }
1534
1535 static void dlm_connect(struct connection *con)
1536 {
1537         struct sockaddr_storage addr;
1538         int result, addr_len;
1539         struct socket *sock;
1540         unsigned int mark;
1541
1542         /* Some odd races can cause double-connects, ignore them */
1543         if (con->retries++ > MAX_CONNECT_RETRIES)
1544                 return;
1545
1546         if (con->sock) {
1547                 log_print("node %d already connected.", con->nodeid);
1548                 return;
1549         }
1550
1551         memset(&addr, 0, sizeof(addr));
1552         result = nodeid_to_addr(con->nodeid, &addr, NULL,
1553                                 dlm_proto_ops->try_new_addr, &mark);
1554         if (result < 0) {
1555                 log_print("no address for nodeid %d", con->nodeid);
1556                 return;
1557         }
1558
1559         /* Create a socket to communicate with */
1560         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1561                                   SOCK_STREAM, dlm_proto_ops->proto, &sock);
1562         if (result < 0)
1563                 goto socket_err;
1564
1565         sock_set_mark(sock->sk, mark);
1566         dlm_proto_ops->sockopts(sock);
1567
1568         add_sock(sock, con);
1569
1570         result = dlm_proto_ops->bind(sock);
1571         if (result < 0)
1572                 goto add_sock_err;
1573
1574         log_print_ratelimited("connecting to %d", con->nodeid);
1575         make_sockaddr(&addr, dlm_config.ci_tcp_port, &addr_len);
1576         result = dlm_proto_ops->connect(con, sock, (struct sockaddr *)&addr,
1577                                         addr_len);
1578         if (result < 0)
1579                 goto add_sock_err;
1580
1581         return;
1582
1583 add_sock_err:
1584         dlm_close_sock(&con->sock);
1585
1586 socket_err:
1587         /*
1588          * Some errors are fatal and this list might need adjusting. For other
1589          * errors we try again until the max number of retries is reached.
1590          */
1591         if (result != -EHOSTUNREACH &&
1592             result != -ENETUNREACH &&
1593             result != -ENETDOWN &&
1594             result != -EINVAL &&
1595             result != -EPROTONOSUPPORT) {
1596                 log_print("connect %d try %d error %d", con->nodeid,
1597                           con->retries, result);
1598                 msleep(1000);
1599                 lowcomms_connect_sock(con);
1600         }
1601 }
1602
1603 /* Send workqueue function */
1604 static void process_send_sockets(struct work_struct *work)
1605 {
1606         struct connection *con = container_of(work, struct connection, swork);
1607
1608         WARN_ON(test_bit(CF_IS_OTHERCON, &con->flags));
1609
1610         clear_bit(CF_WRITE_PENDING, &con->flags);
1611
1612         if (test_and_clear_bit(CF_RECONNECT, &con->flags)) {
1613                 close_connection(con, false, false, true);
1614                 dlm_midcomms_unack_msg_resend(con->nodeid);
1615         }
1616
1617         if (con->sock == NULL) {
1618                 if (test_and_clear_bit(CF_DELAY_CONNECT, &con->flags))
1619                         msleep(1000);
1620
1621                 mutex_lock(&con->sock_mutex);
1622                 dlm_connect(con);
1623                 mutex_unlock(&con->sock_mutex);
1624         }
1625
1626         if (!list_empty(&con->writequeue))
1627                 send_to_sock(con);
1628 }
1629
1630 static void work_stop(void)
1631 {
1632         if (recv_workqueue) {
1633                 destroy_workqueue(recv_workqueue);
1634                 recv_workqueue = NULL;
1635         }
1636
1637         if (send_workqueue) {
1638                 destroy_workqueue(send_workqueue);
1639                 send_workqueue = NULL;
1640         }
1641 }
1642
1643 static int work_start(void)
1644 {
1645         recv_workqueue = alloc_ordered_workqueue("dlm_recv", WQ_MEM_RECLAIM);
1646         if (!recv_workqueue) {
1647                 log_print("can't start dlm_recv");
1648                 return -ENOMEM;
1649         }
1650
1651         send_workqueue = alloc_ordered_workqueue("dlm_send", WQ_MEM_RECLAIM);
1652         if (!send_workqueue) {
1653                 log_print("can't start dlm_send");
1654                 destroy_workqueue(recv_workqueue);
1655                 recv_workqueue = NULL;
1656                 return -ENOMEM;
1657         }
1658
1659         return 0;
1660 }
1661
1662 static void shutdown_conn(struct connection *con)
1663 {
1664         if (dlm_proto_ops->shutdown_action)
1665                 dlm_proto_ops->shutdown_action(con);
1666 }
1667
1668 void dlm_lowcomms_shutdown(void)
1669 {
1670         int idx;
1671
1672         /* Set all the flags to prevent any
1673          * socket activity.
1674          */
1675         dlm_allow_conn = 0;
1676
1677         if (recv_workqueue)
1678                 flush_workqueue(recv_workqueue);
1679         if (send_workqueue)
1680                 flush_workqueue(send_workqueue);
1681
1682         dlm_close_sock(&listen_con.sock);
1683
1684         idx = srcu_read_lock(&connections_srcu);
1685         foreach_conn(shutdown_conn);
1686         srcu_read_unlock(&connections_srcu, idx);
1687 }
1688
1689 static void _stop_conn(struct connection *con, bool and_other)
1690 {
1691         mutex_lock(&con->sock_mutex);
1692         set_bit(CF_CLOSE, &con->flags);
1693         set_bit(CF_READ_PENDING, &con->flags);
1694         set_bit(CF_WRITE_PENDING, &con->flags);
1695         if (con->sock && con->sock->sk) {
1696                 lock_sock(con->sock->sk);
1697                 con->sock->sk->sk_user_data = NULL;
1698                 release_sock(con->sock->sk);
1699         }
1700         if (con->othercon && and_other)
1701                 _stop_conn(con->othercon, false);
1702         mutex_unlock(&con->sock_mutex);
1703 }
1704
1705 static void stop_conn(struct connection *con)
1706 {
1707         _stop_conn(con, true);
1708 }
1709
1710 static void connection_release(struct rcu_head *rcu)
1711 {
1712         struct connection *con = container_of(rcu, struct connection, rcu);
1713
1714         kfree(con->rx_buf);
1715         kfree(con);
1716 }
1717
1718 static void free_conn(struct connection *con)
1719 {
1720         close_connection(con, true, true, true);
1721         spin_lock(&connections_lock);
1722         hlist_del_rcu(&con->list);
1723         spin_unlock(&connections_lock);
1724         if (con->othercon) {
1725                 clean_one_writequeue(con->othercon);
1726                 call_srcu(&connections_srcu, &con->othercon->rcu,
1727                           connection_release);
1728         }
1729         clean_one_writequeue(con);
1730         call_srcu(&connections_srcu, &con->rcu, connection_release);
1731 }
1732
1733 static void work_flush(void)
1734 {
1735         int ok;
1736         int i;
1737         struct connection *con;
1738
1739         do {
1740                 ok = 1;
1741                 foreach_conn(stop_conn);
1742                 if (recv_workqueue)
1743                         flush_workqueue(recv_workqueue);
1744                 if (send_workqueue)
1745                         flush_workqueue(send_workqueue);
1746                 for (i = 0; i < CONN_HASH_SIZE && ok; i++) {
1747                         hlist_for_each_entry_rcu(con, &connection_hash[i],
1748                                                  list) {
1749                                 ok &= test_bit(CF_READ_PENDING, &con->flags);
1750                                 ok &= test_bit(CF_WRITE_PENDING, &con->flags);
1751                                 if (con->othercon) {
1752                                         ok &= test_bit(CF_READ_PENDING,
1753                                                        &con->othercon->flags);
1754                                         ok &= test_bit(CF_WRITE_PENDING,
1755                                                        &con->othercon->flags);
1756                                 }
1757                         }
1758                 }
1759         } while (!ok);
1760 }
1761
1762 void dlm_lowcomms_stop(void)
1763 {
1764         int idx;
1765
1766         idx = srcu_read_lock(&connections_srcu);
1767         work_flush();
1768         foreach_conn(free_conn);
1769         srcu_read_unlock(&connections_srcu, idx);
1770         work_stop();
1771         deinit_local();
1772
1773         dlm_proto_ops = NULL;
1774 }
1775
1776 static int dlm_listen_for_all(void)
1777 {
1778         struct socket *sock;
1779         int result;
1780
1781         log_print("Using %s for communications",
1782                   dlm_proto_ops->name);
1783
1784         result = dlm_proto_ops->listen_validate();
1785         if (result < 0)
1786                 return result;
1787
1788         result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family,
1789                                   SOCK_STREAM, dlm_proto_ops->proto, &sock);
1790         if (result < 0) {
1791                 log_print("Can't create comms socket: %d", result);
1792                 goto out;
1793         }
1794
1795         sock_set_mark(sock->sk, dlm_config.ci_mark);
1796         dlm_proto_ops->listen_sockopts(sock);
1797
1798         result = dlm_proto_ops->listen_bind(sock);
1799         if (result < 0)
1800                 goto out;
1801
1802         save_listen_callbacks(sock);
1803         add_listen_sock(sock, &listen_con);
1804
1805         INIT_WORK(&listen_con.rwork, process_listen_recv_socket);
1806         result = sock->ops->listen(sock, 5);
1807         if (result < 0) {
1808                 dlm_close_sock(&listen_con.sock);
1809                 goto out;
1810         }
1811
1812         return 0;
1813
1814 out:
1815         sock_release(sock);
1816         return result;
1817 }
1818
1819 static int dlm_tcp_bind(struct socket *sock)
1820 {
1821         struct sockaddr_storage src_addr;
1822         int result, addr_len;
1823
1824         /* Bind to our cluster-known address connecting to avoid
1825          * routing problems.
1826          */
1827         memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr));
1828         make_sockaddr(&src_addr, 0, &addr_len);
1829
1830         result = sock->ops->bind(sock, (struct sockaddr *)&src_addr,
1831                                  addr_len);
1832         if (result < 0) {
1833                 /* This *may* not indicate a critical error */
1834                 log_print("could not bind for connect: %d", result);
1835         }
1836
1837         return 0;
1838 }
1839
1840 static int dlm_tcp_connect(struct connection *con, struct socket *sock,
1841                            struct sockaddr *addr, int addr_len)
1842 {
1843         int ret;
1844
1845         ret = sock->ops->connect(sock, addr, addr_len, O_NONBLOCK);
1846         switch (ret) {
1847         case -EINPROGRESS:
1848                 fallthrough;
1849         case 0:
1850                 return 0;
1851         }
1852
1853         return ret;
1854 }
1855
1856 static int dlm_tcp_listen_validate(void)
1857 {
1858         /* We don't support multi-homed hosts */
1859         if (dlm_local_count > 1) {
1860                 log_print("TCP protocol can't handle multi-homed hosts, try SCTP");
1861                 return -EINVAL;
1862         }
1863
1864         return 0;
1865 }
1866
1867 static void dlm_tcp_sockopts(struct socket *sock)
1868 {
1869         /* Turn off Nagle's algorithm */
1870         tcp_sock_set_nodelay(sock->sk);
1871 }
1872
1873 static void dlm_tcp_listen_sockopts(struct socket *sock)
1874 {
1875         dlm_tcp_sockopts(sock);
1876         sock_set_reuseaddr(sock->sk);
1877 }
1878
1879 static int dlm_tcp_listen_bind(struct socket *sock)
1880 {
1881         int addr_len;
1882
1883         /* Bind to our port */
1884         make_sockaddr(dlm_local_addr[0], dlm_config.ci_tcp_port, &addr_len);
1885         return sock->ops->bind(sock, (struct sockaddr *)dlm_local_addr[0],
1886                                addr_len);
1887 }
1888
1889 static const struct dlm_proto_ops dlm_tcp_ops = {
1890         .name = "TCP",
1891         .proto = IPPROTO_TCP,
1892         .connect = dlm_tcp_connect,
1893         .sockopts = dlm_tcp_sockopts,
1894         .bind = dlm_tcp_bind,
1895         .listen_validate = dlm_tcp_listen_validate,
1896         .listen_sockopts = dlm_tcp_listen_sockopts,
1897         .listen_bind = dlm_tcp_listen_bind,
1898         .shutdown_action = dlm_tcp_shutdown,
1899         .eof_condition = tcp_eof_condition,
1900 };
1901
1902 static int dlm_sctp_bind(struct socket *sock)
1903 {
1904         return sctp_bind_addrs(sock, 0);
1905 }
1906
1907 static int dlm_sctp_connect(struct connection *con, struct socket *sock,
1908                             struct sockaddr *addr, int addr_len)
1909 {
1910         int ret;
1911
1912         /*
1913          * Make sock->ops->connect() function return in specified time,
1914          * since O_NONBLOCK argument in connect() function does not work here,
1915          * then, we should restore the default value of this attribute.
1916          */
1917         sock_set_sndtimeo(sock->sk, 5);
1918         ret = sock->ops->connect(sock, addr, addr_len, 0);
1919         sock_set_sndtimeo(sock->sk, 0);
1920         if (ret < 0)
1921                 return ret;
1922
1923         if (!test_and_set_bit(CF_CONNECTED, &con->flags))
1924                 log_print("successful connected to node %d", con->nodeid);
1925
1926         return 0;
1927 }
1928
1929 static int dlm_sctp_listen_validate(void)
1930 {
1931         if (!IS_ENABLED(CONFIG_IP_SCTP)) {
1932                 log_print("SCTP is not enabled by this kernel");
1933                 return -EOPNOTSUPP;
1934         }
1935
1936         request_module("sctp");
1937         return 0;
1938 }
1939
1940 static int dlm_sctp_bind_listen(struct socket *sock)
1941 {
1942         return sctp_bind_addrs(sock, dlm_config.ci_tcp_port);
1943 }
1944
1945 static void dlm_sctp_sockopts(struct socket *sock)
1946 {
1947         /* Turn off Nagle's algorithm */
1948         sctp_sock_set_nodelay(sock->sk);
1949         sock_set_rcvbuf(sock->sk, NEEDED_RMEM);
1950 }
1951
1952 static const struct dlm_proto_ops dlm_sctp_ops = {
1953         .name = "SCTP",
1954         .proto = IPPROTO_SCTP,
1955         .try_new_addr = true,
1956         .connect = dlm_sctp_connect,
1957         .sockopts = dlm_sctp_sockopts,
1958         .bind = dlm_sctp_bind,
1959         .listen_validate = dlm_sctp_listen_validate,
1960         .listen_sockopts = dlm_sctp_sockopts,
1961         .listen_bind = dlm_sctp_bind_listen,
1962 };
1963
1964 int dlm_lowcomms_start(void)
1965 {
1966         int error = -EINVAL;
1967         int i;
1968
1969         for (i = 0; i < CONN_HASH_SIZE; i++)
1970                 INIT_HLIST_HEAD(&connection_hash[i]);
1971
1972         init_local();
1973         if (!dlm_local_count) {
1974                 error = -ENOTCONN;
1975                 log_print("no local IP address has been set");
1976                 goto fail;
1977         }
1978
1979         INIT_WORK(&listen_con.rwork, process_listen_recv_socket);
1980
1981         error = work_start();
1982         if (error)
1983                 goto fail_local;
1984
1985         dlm_allow_conn = 1;
1986
1987         /* Start listening */
1988         switch (dlm_config.ci_protocol) {
1989         case DLM_PROTO_TCP:
1990                 dlm_proto_ops = &dlm_tcp_ops;
1991                 break;
1992         case DLM_PROTO_SCTP:
1993                 dlm_proto_ops = &dlm_sctp_ops;
1994                 break;
1995         default:
1996                 log_print("Invalid protocol identifier %d set",
1997                           dlm_config.ci_protocol);
1998                 error = -EINVAL;
1999                 goto fail_proto_ops;
2000         }
2001
2002         error = dlm_listen_for_all();
2003         if (error)
2004                 goto fail_listen;
2005
2006         return 0;
2007
2008 fail_listen:
2009         dlm_proto_ops = NULL;
2010 fail_proto_ops:
2011         dlm_allow_conn = 0;
2012         dlm_close_sock(&listen_con.sock);
2013         work_stop();
2014 fail_local:
2015         deinit_local();
2016 fail:
2017         return error;
2018 }
2019
2020 void dlm_lowcomms_exit(void)
2021 {
2022         struct dlm_node_addr *na, *safe;
2023
2024         spin_lock(&dlm_node_addrs_spin);
2025         list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) {
2026                 list_del(&na->list);
2027                 while (na->addr_count--)
2028                         kfree(na->addr[na->addr_count]);
2029                 kfree(na);
2030         }
2031         spin_unlock(&dlm_node_addrs_spin);
2032 }