1 /****************************************************************************** 2 ******************************************************************************* 3 ** 4 ** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved. 5 ** Copyright (C) 2004-2009 Red Hat, Inc. All rights reserved. 6 ** 7 ** This copyrighted material is made available to anyone wishing to use, 8 ** modify, copy, or redistribute it subject to the terms and conditions 9 ** of the GNU General Public License v.2. 10 ** 11 ******************************************************************************* 12 ******************************************************************************/ 13 14 /* 15 * lowcomms.c 16 * 17 * This is the "low-level" comms layer. 18 * 19 * It is responsible for sending/receiving messages 20 * from other nodes in the cluster. 21 * 22 * Cluster nodes are referred to by their nodeids. nodeids are 23 * simply 32 bit numbers to the locking module - if they need to 24 * be expanded for the cluster infrastructure then that is its 25 * responsibility. It is this layer's 26 * responsibility to resolve these into IP address or 27 * whatever it needs for inter-node communication. 28 * 29 * The comms level is two kernel threads that deal mainly with 30 * the receiving of messages from other nodes and passing them 31 * up to the mid-level comms layer (which understands the 32 * message format) for execution by the locking core, and 33 * a send thread which does all the setting up of connections 34 * to remote nodes and the sending of data. Threads are not allowed 35 * to send their own data because it may cause them to wait in times 36 * of high load. Also, this way, the sending thread can collect together 37 * messages bound for one node and send them in one block. 38 * 39 * lowcomms will choose to use either TCP or SCTP as its transport layer 40 * depending on the configuration variable 'protocol'. This should be set 41 * to 0 (default) for TCP or 1 for SCTP. It should be configured using a 42 * cluster-wide mechanism as it must be the same on all nodes of the cluster 43 * for the DLM to function. 44 * 45 */ 46 47 #include <asm/ioctls.h> 48 #include <net/sock.h> 49 #include <net/tcp.h> 50 #include <linux/pagemap.h> 51 #include <linux/file.h> 52 #include <linux/mutex.h> 53 #include <linux/sctp.h> 54 #include <linux/slab.h> 55 #include <linux/sctp.h> 56 #include <net/sctp/sctp.h> 57 #include <net/ipv6.h> 58 59 #include "dlm_internal.h" 60 #include "lowcomms.h" 61 #include "midcomms.h" 62 #include "config.h" 63 64 #define NEEDED_RMEM (4*1024*1024) 65 #define CONN_HASH_SIZE 32 66 67 /* Number of messages to send before rescheduling */ 68 #define MAX_SEND_MSG_COUNT 25 69 70 struct cbuf { 71 unsigned int base; 72 unsigned int len; 73 unsigned int mask; 74 }; 75 76 static void cbuf_add(struct cbuf *cb, int n) 77 { 78 cb->len += n; 79 } 80 81 static int cbuf_data(struct cbuf *cb) 82 { 83 return ((cb->base + cb->len) & cb->mask); 84 } 85 86 static void cbuf_init(struct cbuf *cb, int size) 87 { 88 cb->base = cb->len = 0; 89 cb->mask = size-1; 90 } 91 92 static void cbuf_eat(struct cbuf *cb, int n) 93 { 94 cb->len -= n; 95 cb->base += n; 96 cb->base &= cb->mask; 97 } 98 99 static bool cbuf_empty(struct cbuf *cb) 100 { 101 return cb->len == 0; 102 } 103 104 struct connection { 105 struct socket *sock; /* NULL if not connected */ 106 uint32_t nodeid; /* So we know who we are in the list */ 107 struct mutex sock_mutex; 108 unsigned long flags; 109 #define CF_READ_PENDING 1 110 #define CF_WRITE_PENDING 2 111 #define CF_CONNECT_PENDING 3 112 #define CF_INIT_PENDING 4 113 #define CF_IS_OTHERCON 5 114 #define CF_CLOSE 6 115 #define CF_APP_LIMITED 7 116 struct list_head writequeue; /* List of outgoing writequeue_entries */ 117 spinlock_t writequeue_lock; 118 int (*rx_action) (struct connection *); /* What to do when active */ 119 void (*connect_action) (struct connection *); /* What to do to connect */ 120 struct page *rx_page; 121 struct cbuf cb; 122 int retries; 123 #define MAX_CONNECT_RETRIES 3 124 int sctp_assoc; 125 struct hlist_node list; 126 struct connection *othercon; 127 struct work_struct rwork; /* Receive workqueue */ 128 struct work_struct swork; /* Send workqueue */ 129 bool try_new_addr; 130 }; 131 #define sock2con(x) ((struct connection *)(x)->sk_user_data) 132 133 /* An entry waiting to be sent */ 134 struct writequeue_entry { 135 struct list_head list; 136 struct page *page; 137 int offset; 138 int len; 139 int end; 140 int users; 141 struct connection *con; 142 }; 143 144 struct dlm_node_addr { 145 struct list_head list; 146 int nodeid; 147 int addr_count; 148 int curr_addr_index; 149 struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT]; 150 }; 151 152 static LIST_HEAD(dlm_node_addrs); 153 static DEFINE_SPINLOCK(dlm_node_addrs_spin); 154 155 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT]; 156 static int dlm_local_count; 157 static int dlm_allow_conn; 158 159 /* Work queues */ 160 static struct workqueue_struct *recv_workqueue; 161 static struct workqueue_struct *send_workqueue; 162 163 static struct hlist_head connection_hash[CONN_HASH_SIZE]; 164 static DEFINE_MUTEX(connections_lock); 165 static struct kmem_cache *con_cache; 166 167 static void process_recv_sockets(struct work_struct *work); 168 static void process_send_sockets(struct work_struct *work); 169 170 171 /* This is deliberately very simple because most clusters have simple 172 sequential nodeids, so we should be able to go straight to a connection 173 struct in the array */ 174 static inline int nodeid_hash(int nodeid) 175 { 176 return nodeid & (CONN_HASH_SIZE-1); 177 } 178 179 static struct connection *__find_con(int nodeid) 180 { 181 int r; 182 struct connection *con; 183 184 r = nodeid_hash(nodeid); 185 186 hlist_for_each_entry(con, &connection_hash[r], list) { 187 if (con->nodeid == nodeid) 188 return con; 189 } 190 return NULL; 191 } 192 193 /* 194 * If 'allocation' is zero then we don't attempt to create a new 195 * connection structure for this node. 196 */ 197 static struct connection *__nodeid2con(int nodeid, gfp_t alloc) 198 { 199 struct connection *con = NULL; 200 int r; 201 202 con = __find_con(nodeid); 203 if (con || !alloc) 204 return con; 205 206 con = kmem_cache_zalloc(con_cache, alloc); 207 if (!con) 208 return NULL; 209 210 r = nodeid_hash(nodeid); 211 hlist_add_head(&con->list, &connection_hash[r]); 212 213 con->nodeid = nodeid; 214 mutex_init(&con->sock_mutex); 215 INIT_LIST_HEAD(&con->writequeue); 216 spin_lock_init(&con->writequeue_lock); 217 INIT_WORK(&con->swork, process_send_sockets); 218 INIT_WORK(&con->rwork, process_recv_sockets); 219 220 /* Setup action pointers for child sockets */ 221 if (con->nodeid) { 222 struct connection *zerocon = __find_con(0); 223 224 con->connect_action = zerocon->connect_action; 225 if (!con->rx_action) 226 con->rx_action = zerocon->rx_action; 227 } 228 229 return con; 230 } 231 232 /* Loop round all connections */ 233 static void foreach_conn(void (*conn_func)(struct connection *c)) 234 { 235 int i; 236 struct hlist_node *n; 237 struct connection *con; 238 239 for (i = 0; i < CONN_HASH_SIZE; i++) { 240 hlist_for_each_entry_safe(con, n, &connection_hash[i], list) 241 conn_func(con); 242 } 243 } 244 245 static struct connection *nodeid2con(int nodeid, gfp_t allocation) 246 { 247 struct connection *con; 248 249 mutex_lock(&connections_lock); 250 con = __nodeid2con(nodeid, allocation); 251 mutex_unlock(&connections_lock); 252 253 return con; 254 } 255 256 /* This is a bit drastic, but only called when things go wrong */ 257 static struct connection *assoc2con(int assoc_id) 258 { 259 int i; 260 struct connection *con; 261 262 mutex_lock(&connections_lock); 263 264 for (i = 0 ; i < CONN_HASH_SIZE; i++) { 265 hlist_for_each_entry(con, &connection_hash[i], list) { 266 if (con->sctp_assoc == assoc_id) { 267 mutex_unlock(&connections_lock); 268 return con; 269 } 270 } 271 } 272 mutex_unlock(&connections_lock); 273 return NULL; 274 } 275 276 static struct dlm_node_addr *find_node_addr(int nodeid) 277 { 278 struct dlm_node_addr *na; 279 280 list_for_each_entry(na, &dlm_node_addrs, list) { 281 if (na->nodeid == nodeid) 282 return na; 283 } 284 return NULL; 285 } 286 287 static int addr_compare(struct sockaddr_storage *x, struct sockaddr_storage *y) 288 { 289 switch (x->ss_family) { 290 case AF_INET: { 291 struct sockaddr_in *sinx = (struct sockaddr_in *)x; 292 struct sockaddr_in *siny = (struct sockaddr_in *)y; 293 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr) 294 return 0; 295 if (sinx->sin_port != siny->sin_port) 296 return 0; 297 break; 298 } 299 case AF_INET6: { 300 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x; 301 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y; 302 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr)) 303 return 0; 304 if (sinx->sin6_port != siny->sin6_port) 305 return 0; 306 break; 307 } 308 default: 309 return 0; 310 } 311 return 1; 312 } 313 314 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out, 315 struct sockaddr *sa_out, bool try_new_addr) 316 { 317 struct sockaddr_storage sas; 318 struct dlm_node_addr *na; 319 320 if (!dlm_local_count) 321 return -1; 322 323 spin_lock(&dlm_node_addrs_spin); 324 na = find_node_addr(nodeid); 325 if (na && na->addr_count) { 326 if (try_new_addr) { 327 na->curr_addr_index++; 328 if (na->curr_addr_index == na->addr_count) 329 na->curr_addr_index = 0; 330 } 331 332 memcpy(&sas, na->addr[na->curr_addr_index ], 333 sizeof(struct sockaddr_storage)); 334 } 335 spin_unlock(&dlm_node_addrs_spin); 336 337 if (!na) 338 return -EEXIST; 339 340 if (!na->addr_count) 341 return -ENOENT; 342 343 if (sas_out) 344 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage)); 345 346 if (!sa_out) 347 return 0; 348 349 if (dlm_local_addr[0]->ss_family == AF_INET) { 350 struct sockaddr_in *in4 = (struct sockaddr_in *) &sas; 351 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out; 352 ret4->sin_addr.s_addr = in4->sin_addr.s_addr; 353 } else { 354 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) &sas; 355 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out; 356 ret6->sin6_addr = in6->sin6_addr; 357 } 358 359 return 0; 360 } 361 362 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid) 363 { 364 struct dlm_node_addr *na; 365 int rv = -EEXIST; 366 int addr_i; 367 368 spin_lock(&dlm_node_addrs_spin); 369 list_for_each_entry(na, &dlm_node_addrs, list) { 370 if (!na->addr_count) 371 continue; 372 373 for (addr_i = 0; addr_i < na->addr_count; addr_i++) { 374 if (addr_compare(na->addr[addr_i], addr)) { 375 *nodeid = na->nodeid; 376 rv = 0; 377 goto unlock; 378 } 379 } 380 } 381 unlock: 382 spin_unlock(&dlm_node_addrs_spin); 383 return rv; 384 } 385 386 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len) 387 { 388 struct sockaddr_storage *new_addr; 389 struct dlm_node_addr *new_node, *na; 390 391 new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS); 392 if (!new_node) 393 return -ENOMEM; 394 395 new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS); 396 if (!new_addr) { 397 kfree(new_node); 398 return -ENOMEM; 399 } 400 401 memcpy(new_addr, addr, len); 402 403 spin_lock(&dlm_node_addrs_spin); 404 na = find_node_addr(nodeid); 405 if (!na) { 406 new_node->nodeid = nodeid; 407 new_node->addr[0] = new_addr; 408 new_node->addr_count = 1; 409 list_add(&new_node->list, &dlm_node_addrs); 410 spin_unlock(&dlm_node_addrs_spin); 411 return 0; 412 } 413 414 if (na->addr_count >= DLM_MAX_ADDR_COUNT) { 415 spin_unlock(&dlm_node_addrs_spin); 416 kfree(new_addr); 417 kfree(new_node); 418 return -ENOSPC; 419 } 420 421 na->addr[na->addr_count++] = new_addr; 422 spin_unlock(&dlm_node_addrs_spin); 423 kfree(new_node); 424 return 0; 425 } 426 427 /* Data available on socket or listen socket received a connect */ 428 static void lowcomms_data_ready(struct sock *sk, int count_unused) 429 { 430 struct connection *con = sock2con(sk); 431 if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags)) 432 queue_work(recv_workqueue, &con->rwork); 433 } 434 435 static void lowcomms_write_space(struct sock *sk) 436 { 437 struct connection *con = sock2con(sk); 438 439 if (!con) 440 return; 441 442 clear_bit(SOCK_NOSPACE, &con->sock->flags); 443 444 if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) { 445 con->sock->sk->sk_write_pending--; 446 clear_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags); 447 } 448 449 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) 450 queue_work(send_workqueue, &con->swork); 451 } 452 453 static inline void lowcomms_connect_sock(struct connection *con) 454 { 455 if (test_bit(CF_CLOSE, &con->flags)) 456 return; 457 if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags)) 458 queue_work(send_workqueue, &con->swork); 459 } 460 461 static void lowcomms_state_change(struct sock *sk) 462 { 463 if (sk->sk_state == TCP_ESTABLISHED) 464 lowcomms_write_space(sk); 465 } 466 467 int dlm_lowcomms_connect_node(int nodeid) 468 { 469 struct connection *con; 470 471 /* with sctp there's no connecting without sending */ 472 if (dlm_config.ci_protocol != 0) 473 return 0; 474 475 if (nodeid == dlm_our_nodeid()) 476 return 0; 477 478 con = nodeid2con(nodeid, GFP_NOFS); 479 if (!con) 480 return -ENOMEM; 481 lowcomms_connect_sock(con); 482 return 0; 483 } 484 485 /* Make a socket active */ 486 static void add_sock(struct socket *sock, struct connection *con) 487 { 488 con->sock = sock; 489 490 /* Install a data_ready callback */ 491 con->sock->sk->sk_data_ready = lowcomms_data_ready; 492 con->sock->sk->sk_write_space = lowcomms_write_space; 493 con->sock->sk->sk_state_change = lowcomms_state_change; 494 con->sock->sk->sk_user_data = con; 495 con->sock->sk->sk_allocation = GFP_NOFS; 496 } 497 498 /* Add the port number to an IPv6 or 4 sockaddr and return the address 499 length */ 500 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port, 501 int *addr_len) 502 { 503 saddr->ss_family = dlm_local_addr[0]->ss_family; 504 if (saddr->ss_family == AF_INET) { 505 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr; 506 in4_addr->sin_port = cpu_to_be16(port); 507 *addr_len = sizeof(struct sockaddr_in); 508 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero)); 509 } else { 510 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr; 511 in6_addr->sin6_port = cpu_to_be16(port); 512 *addr_len = sizeof(struct sockaddr_in6); 513 } 514 memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len); 515 } 516 517 /* Close a remote connection and tidy up */ 518 static void close_connection(struct connection *con, bool and_other) 519 { 520 mutex_lock(&con->sock_mutex); 521 522 if (con->sock) { 523 sock_release(con->sock); 524 con->sock = NULL; 525 } 526 if (con->othercon && and_other) { 527 /* Will only re-enter once. */ 528 close_connection(con->othercon, false); 529 } 530 if (con->rx_page) { 531 __free_page(con->rx_page); 532 con->rx_page = NULL; 533 } 534 535 con->retries = 0; 536 mutex_unlock(&con->sock_mutex); 537 } 538 539 /* We only send shutdown messages to nodes that are not part of the cluster */ 540 static void sctp_send_shutdown(sctp_assoc_t associd) 541 { 542 static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))]; 543 struct msghdr outmessage; 544 struct cmsghdr *cmsg; 545 struct sctp_sndrcvinfo *sinfo; 546 int ret; 547 struct connection *con; 548 549 con = nodeid2con(0,0); 550 BUG_ON(con == NULL); 551 552 outmessage.msg_name = NULL; 553 outmessage.msg_namelen = 0; 554 outmessage.msg_control = outcmsg; 555 outmessage.msg_controllen = sizeof(outcmsg); 556 outmessage.msg_flags = MSG_EOR; 557 558 cmsg = CMSG_FIRSTHDR(&outmessage); 559 cmsg->cmsg_level = IPPROTO_SCTP; 560 cmsg->cmsg_type = SCTP_SNDRCV; 561 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo)); 562 outmessage.msg_controllen = cmsg->cmsg_len; 563 sinfo = CMSG_DATA(cmsg); 564 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo)); 565 566 sinfo->sinfo_flags |= MSG_EOF; 567 sinfo->sinfo_assoc_id = associd; 568 569 ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0); 570 571 if (ret != 0) 572 log_print("send EOF to node failed: %d", ret); 573 } 574 575 static void sctp_init_failed_foreach(struct connection *con) 576 { 577 578 /* 579 * Don't try to recover base con and handle race where the 580 * other node's assoc init creates a assoc and we get that 581 * notification, then we get a notification that our attempt 582 * failed due. This happens when we are still trying the primary 583 * address, but the other node has already tried secondary addrs 584 * and found one that worked. 585 */ 586 if (!con->nodeid || con->sctp_assoc) 587 return; 588 589 log_print("Retrying SCTP association init for node %d\n", con->nodeid); 590 591 con->try_new_addr = true; 592 con->sctp_assoc = 0; 593 if (test_and_clear_bit(CF_INIT_PENDING, &con->flags)) { 594 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) 595 queue_work(send_workqueue, &con->swork); 596 } 597 } 598 599 /* INIT failed but we don't know which node... 600 restart INIT on all pending nodes */ 601 static void sctp_init_failed(void) 602 { 603 mutex_lock(&connections_lock); 604 605 foreach_conn(sctp_init_failed_foreach); 606 607 mutex_unlock(&connections_lock); 608 } 609 610 static void retry_failed_sctp_send(struct connection *recv_con, 611 struct sctp_send_failed *sn_send_failed, 612 char *buf) 613 { 614 int len = sn_send_failed->ssf_length - sizeof(struct sctp_send_failed); 615 struct dlm_mhandle *mh; 616 struct connection *con; 617 char *retry_buf; 618 int nodeid = sn_send_failed->ssf_info.sinfo_ppid; 619 620 log_print("Retry sending %d bytes to node id %d", len, nodeid); 621 622 con = nodeid2con(nodeid, 0); 623 if (!con) { 624 log_print("Could not look up con for nodeid %d\n", 625 nodeid); 626 return; 627 } 628 629 mh = dlm_lowcomms_get_buffer(nodeid, len, GFP_NOFS, &retry_buf); 630 if (!mh) { 631 log_print("Could not allocate buf for retry."); 632 return; 633 } 634 memcpy(retry_buf, buf + sizeof(struct sctp_send_failed), len); 635 dlm_lowcomms_commit_buffer(mh); 636 637 /* 638 * If we got a assoc changed event before the send failed event then 639 * we only need to retry the send. 640 */ 641 if (con->sctp_assoc) { 642 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) 643 queue_work(send_workqueue, &con->swork); 644 } else 645 sctp_init_failed_foreach(con); 646 } 647 648 /* Something happened to an association */ 649 static void process_sctp_notification(struct connection *con, 650 struct msghdr *msg, char *buf) 651 { 652 union sctp_notification *sn = (union sctp_notification *)buf; 653 654 switch (sn->sn_header.sn_type) { 655 case SCTP_SEND_FAILED: 656 retry_failed_sctp_send(con, &sn->sn_send_failed, buf); 657 break; 658 case SCTP_ASSOC_CHANGE: 659 switch (sn->sn_assoc_change.sac_state) { 660 case SCTP_COMM_UP: 661 case SCTP_RESTART: 662 { 663 /* Check that the new node is in the lockspace */ 664 struct sctp_prim prim; 665 int nodeid; 666 int prim_len, ret; 667 int addr_len; 668 struct connection *new_con; 669 670 /* 671 * We get this before any data for an association. 672 * We verify that the node is in the cluster and 673 * then peel off a socket for it. 674 */ 675 if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) { 676 log_print("COMM_UP for invalid assoc ID %d", 677 (int)sn->sn_assoc_change.sac_assoc_id); 678 sctp_init_failed(); 679 return; 680 } 681 memset(&prim, 0, sizeof(struct sctp_prim)); 682 prim_len = sizeof(struct sctp_prim); 683 prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id; 684 685 ret = kernel_getsockopt(con->sock, 686 IPPROTO_SCTP, 687 SCTP_PRIMARY_ADDR, 688 (char*)&prim, 689 &prim_len); 690 if (ret < 0) { 691 log_print("getsockopt/sctp_primary_addr on " 692 "new assoc %d failed : %d", 693 (int)sn->sn_assoc_change.sac_assoc_id, 694 ret); 695 696 /* Retry INIT later */ 697 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id); 698 if (new_con) 699 clear_bit(CF_CONNECT_PENDING, &con->flags); 700 return; 701 } 702 make_sockaddr(&prim.ssp_addr, 0, &addr_len); 703 if (addr_to_nodeid(&prim.ssp_addr, &nodeid)) { 704 unsigned char *b=(unsigned char *)&prim.ssp_addr; 705 log_print("reject connect from unknown addr"); 706 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE, 707 b, sizeof(struct sockaddr_storage)); 708 sctp_send_shutdown(prim.ssp_assoc_id); 709 return; 710 } 711 712 new_con = nodeid2con(nodeid, GFP_NOFS); 713 if (!new_con) 714 return; 715 716 /* Peel off a new sock */ 717 sctp_lock_sock(con->sock->sk); 718 ret = sctp_do_peeloff(con->sock->sk, 719 sn->sn_assoc_change.sac_assoc_id, 720 &new_con->sock); 721 sctp_release_sock(con->sock->sk); 722 if (ret < 0) { 723 log_print("Can't peel off a socket for " 724 "connection %d to node %d: err=%d", 725 (int)sn->sn_assoc_change.sac_assoc_id, 726 nodeid, ret); 727 return; 728 } 729 add_sock(new_con->sock, new_con); 730 731 log_print("connecting to %d sctp association %d", 732 nodeid, (int)sn->sn_assoc_change.sac_assoc_id); 733 734 new_con->sctp_assoc = sn->sn_assoc_change.sac_assoc_id; 735 new_con->try_new_addr = false; 736 /* Send any pending writes */ 737 clear_bit(CF_CONNECT_PENDING, &new_con->flags); 738 clear_bit(CF_INIT_PENDING, &new_con->flags); 739 if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) { 740 queue_work(send_workqueue, &new_con->swork); 741 } 742 if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags)) 743 queue_work(recv_workqueue, &new_con->rwork); 744 } 745 break; 746 747 case SCTP_COMM_LOST: 748 case SCTP_SHUTDOWN_COMP: 749 { 750 con = assoc2con(sn->sn_assoc_change.sac_assoc_id); 751 if (con) { 752 con->sctp_assoc = 0; 753 } 754 } 755 break; 756 757 case SCTP_CANT_STR_ASSOC: 758 { 759 /* Will retry init when we get the send failed notification */ 760 log_print("Can't start SCTP association - retrying"); 761 } 762 break; 763 764 default: 765 log_print("unexpected SCTP assoc change id=%d state=%d", 766 (int)sn->sn_assoc_change.sac_assoc_id, 767 sn->sn_assoc_change.sac_state); 768 } 769 default: 770 ; /* fall through */ 771 } 772 } 773 774 /* Data received from remote end */ 775 static int receive_from_sock(struct connection *con) 776 { 777 int ret = 0; 778 struct msghdr msg = {}; 779 struct kvec iov[2]; 780 unsigned len; 781 int r; 782 int call_again_soon = 0; 783 int nvec; 784 char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))]; 785 786 mutex_lock(&con->sock_mutex); 787 788 if (con->sock == NULL) { 789 ret = -EAGAIN; 790 goto out_close; 791 } 792 793 if (con->rx_page == NULL) { 794 /* 795 * This doesn't need to be atomic, but I think it should 796 * improve performance if it is. 797 */ 798 con->rx_page = alloc_page(GFP_ATOMIC); 799 if (con->rx_page == NULL) 800 goto out_resched; 801 cbuf_init(&con->cb, PAGE_CACHE_SIZE); 802 } 803 804 /* Only SCTP needs these really */ 805 memset(&incmsg, 0, sizeof(incmsg)); 806 msg.msg_control = incmsg; 807 msg.msg_controllen = sizeof(incmsg); 808 809 /* 810 * iov[0] is the bit of the circular buffer between the current end 811 * point (cb.base + cb.len) and the end of the buffer. 812 */ 813 iov[0].iov_len = con->cb.base - cbuf_data(&con->cb); 814 iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb); 815 iov[1].iov_len = 0; 816 nvec = 1; 817 818 /* 819 * iov[1] is the bit of the circular buffer between the start of the 820 * buffer and the start of the currently used section (cb.base) 821 */ 822 if (cbuf_data(&con->cb) >= con->cb.base) { 823 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb); 824 iov[1].iov_len = con->cb.base; 825 iov[1].iov_base = page_address(con->rx_page); 826 nvec = 2; 827 } 828 len = iov[0].iov_len + iov[1].iov_len; 829 830 r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len, 831 MSG_DONTWAIT | MSG_NOSIGNAL); 832 if (ret <= 0) 833 goto out_close; 834 835 /* Process SCTP notifications */ 836 if (msg.msg_flags & MSG_NOTIFICATION) { 837 msg.msg_control = incmsg; 838 msg.msg_controllen = sizeof(incmsg); 839 840 process_sctp_notification(con, &msg, 841 page_address(con->rx_page) + con->cb.base); 842 mutex_unlock(&con->sock_mutex); 843 return 0; 844 } 845 BUG_ON(con->nodeid == 0); 846 847 if (ret == len) 848 call_again_soon = 1; 849 cbuf_add(&con->cb, ret); 850 ret = dlm_process_incoming_buffer(con->nodeid, 851 page_address(con->rx_page), 852 con->cb.base, con->cb.len, 853 PAGE_CACHE_SIZE); 854 if (ret == -EBADMSG) { 855 log_print("lowcomms: addr=%p, base=%u, len=%u, " 856 "iov_len=%u, iov_base[0]=%p, read=%d", 857 page_address(con->rx_page), con->cb.base, con->cb.len, 858 len, iov[0].iov_base, r); 859 } 860 if (ret < 0) 861 goto out_close; 862 cbuf_eat(&con->cb, ret); 863 864 if (cbuf_empty(&con->cb) && !call_again_soon) { 865 __free_page(con->rx_page); 866 con->rx_page = NULL; 867 } 868 869 if (call_again_soon) 870 goto out_resched; 871 mutex_unlock(&con->sock_mutex); 872 return 0; 873 874 out_resched: 875 if (!test_and_set_bit(CF_READ_PENDING, &con->flags)) 876 queue_work(recv_workqueue, &con->rwork); 877 mutex_unlock(&con->sock_mutex); 878 return -EAGAIN; 879 880 out_close: 881 mutex_unlock(&con->sock_mutex); 882 if (ret != -EAGAIN) { 883 close_connection(con, false); 884 /* Reconnect when there is something to send */ 885 } 886 /* Don't return success if we really got EOF */ 887 if (ret == 0) 888 ret = -EAGAIN; 889 890 return ret; 891 } 892 893 /* Listening socket is busy, accept a connection */ 894 static int tcp_accept_from_sock(struct connection *con) 895 { 896 int result; 897 struct sockaddr_storage peeraddr; 898 struct socket *newsock; 899 int len; 900 int nodeid; 901 struct connection *newcon; 902 struct connection *addcon; 903 904 mutex_lock(&connections_lock); 905 if (!dlm_allow_conn) { 906 mutex_unlock(&connections_lock); 907 return -1; 908 } 909 mutex_unlock(&connections_lock); 910 911 memset(&peeraddr, 0, sizeof(peeraddr)); 912 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM, 913 IPPROTO_TCP, &newsock); 914 if (result < 0) 915 return -ENOMEM; 916 917 mutex_lock_nested(&con->sock_mutex, 0); 918 919 result = -ENOTCONN; 920 if (con->sock == NULL) 921 goto accept_err; 922 923 newsock->type = con->sock->type; 924 newsock->ops = con->sock->ops; 925 926 result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK); 927 if (result < 0) 928 goto accept_err; 929 930 /* Get the connected socket's peer */ 931 memset(&peeraddr, 0, sizeof(peeraddr)); 932 if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr, 933 &len, 2)) { 934 result = -ECONNABORTED; 935 goto accept_err; 936 } 937 938 /* Get the new node's NODEID */ 939 make_sockaddr(&peeraddr, 0, &len); 940 if (addr_to_nodeid(&peeraddr, &nodeid)) { 941 unsigned char *b=(unsigned char *)&peeraddr; 942 log_print("connect from non cluster node"); 943 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE, 944 b, sizeof(struct sockaddr_storage)); 945 sock_release(newsock); 946 mutex_unlock(&con->sock_mutex); 947 return -1; 948 } 949 950 log_print("got connection from %d", nodeid); 951 952 /* Check to see if we already have a connection to this node. This 953 * could happen if the two nodes initiate a connection at roughly 954 * the same time and the connections cross on the wire. 955 * In this case we store the incoming one in "othercon" 956 */ 957 newcon = nodeid2con(nodeid, GFP_NOFS); 958 if (!newcon) { 959 result = -ENOMEM; 960 goto accept_err; 961 } 962 mutex_lock_nested(&newcon->sock_mutex, 1); 963 if (newcon->sock) { 964 struct connection *othercon = newcon->othercon; 965 966 if (!othercon) { 967 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS); 968 if (!othercon) { 969 log_print("failed to allocate incoming socket"); 970 mutex_unlock(&newcon->sock_mutex); 971 result = -ENOMEM; 972 goto accept_err; 973 } 974 othercon->nodeid = nodeid; 975 othercon->rx_action = receive_from_sock; 976 mutex_init(&othercon->sock_mutex); 977 INIT_WORK(&othercon->swork, process_send_sockets); 978 INIT_WORK(&othercon->rwork, process_recv_sockets); 979 set_bit(CF_IS_OTHERCON, &othercon->flags); 980 } 981 if (!othercon->sock) { 982 newcon->othercon = othercon; 983 othercon->sock = newsock; 984 newsock->sk->sk_user_data = othercon; 985 add_sock(newsock, othercon); 986 addcon = othercon; 987 } 988 else { 989 printk("Extra connection from node %d attempted\n", nodeid); 990 result = -EAGAIN; 991 mutex_unlock(&newcon->sock_mutex); 992 goto accept_err; 993 } 994 } 995 else { 996 newsock->sk->sk_user_data = newcon; 997 newcon->rx_action = receive_from_sock; 998 add_sock(newsock, newcon); 999 addcon = newcon; 1000 } 1001 1002 mutex_unlock(&newcon->sock_mutex); 1003 1004 /* 1005 * Add it to the active queue in case we got data 1006 * between processing the accept adding the socket 1007 * to the read_sockets list 1008 */ 1009 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags)) 1010 queue_work(recv_workqueue, &addcon->rwork); 1011 mutex_unlock(&con->sock_mutex); 1012 1013 return 0; 1014 1015 accept_err: 1016 mutex_unlock(&con->sock_mutex); 1017 sock_release(newsock); 1018 1019 if (result != -EAGAIN) 1020 log_print("error accepting connection from node: %d", result); 1021 return result; 1022 } 1023 1024 static void free_entry(struct writequeue_entry *e) 1025 { 1026 __free_page(e->page); 1027 kfree(e); 1028 } 1029 1030 /* 1031 * writequeue_entry_complete - try to delete and free write queue entry 1032 * @e: write queue entry to try to delete 1033 * @completed: bytes completed 1034 * 1035 * writequeue_lock must be held. 1036 */ 1037 static void writequeue_entry_complete(struct writequeue_entry *e, int completed) 1038 { 1039 e->offset += completed; 1040 e->len -= completed; 1041 1042 if (e->len == 0 && e->users == 0) { 1043 list_del(&e->list); 1044 free_entry(e); 1045 } 1046 } 1047 1048 /* Initiate an SCTP association. 1049 This is a special case of send_to_sock() in that we don't yet have a 1050 peeled-off socket for this association, so we use the listening socket 1051 and add the primary IP address of the remote node. 1052 */ 1053 static void sctp_init_assoc(struct connection *con) 1054 { 1055 struct sockaddr_storage rem_addr; 1056 char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))]; 1057 struct msghdr outmessage; 1058 struct cmsghdr *cmsg; 1059 struct sctp_sndrcvinfo *sinfo; 1060 struct connection *base_con; 1061 struct writequeue_entry *e; 1062 int len, offset; 1063 int ret; 1064 int addrlen; 1065 struct kvec iov[1]; 1066 1067 mutex_lock(&con->sock_mutex); 1068 if (test_and_set_bit(CF_INIT_PENDING, &con->flags)) 1069 goto unlock; 1070 1071 if (nodeid_to_addr(con->nodeid, NULL, (struct sockaddr *)&rem_addr, 1072 con->try_new_addr)) { 1073 log_print("no address for nodeid %d", con->nodeid); 1074 goto unlock; 1075 } 1076 base_con = nodeid2con(0, 0); 1077 BUG_ON(base_con == NULL); 1078 1079 make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen); 1080 1081 outmessage.msg_name = &rem_addr; 1082 outmessage.msg_namelen = addrlen; 1083 outmessage.msg_control = outcmsg; 1084 outmessage.msg_controllen = sizeof(outcmsg); 1085 outmessage.msg_flags = MSG_EOR; 1086 1087 spin_lock(&con->writequeue_lock); 1088 1089 if (list_empty(&con->writequeue)) { 1090 spin_unlock(&con->writequeue_lock); 1091 log_print("writequeue empty for nodeid %d", con->nodeid); 1092 goto unlock; 1093 } 1094 1095 e = list_first_entry(&con->writequeue, struct writequeue_entry, list); 1096 len = e->len; 1097 offset = e->offset; 1098 1099 /* Send the first block off the write queue */ 1100 iov[0].iov_base = page_address(e->page)+offset; 1101 iov[0].iov_len = len; 1102 spin_unlock(&con->writequeue_lock); 1103 1104 if (rem_addr.ss_family == AF_INET) { 1105 struct sockaddr_in *sin = (struct sockaddr_in *)&rem_addr; 1106 log_print("Trying to connect to %pI4", &sin->sin_addr.s_addr); 1107 } else { 1108 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&rem_addr; 1109 log_print("Trying to connect to %pI6", &sin6->sin6_addr); 1110 } 1111 1112 cmsg = CMSG_FIRSTHDR(&outmessage); 1113 cmsg->cmsg_level = IPPROTO_SCTP; 1114 cmsg->cmsg_type = SCTP_SNDRCV; 1115 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo)); 1116 sinfo = CMSG_DATA(cmsg); 1117 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo)); 1118 sinfo->sinfo_ppid = cpu_to_le32(con->nodeid); 1119 outmessage.msg_controllen = cmsg->cmsg_len; 1120 sinfo->sinfo_flags |= SCTP_ADDR_OVER; 1121 1122 ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len); 1123 if (ret < 0) { 1124 log_print("Send first packet to node %d failed: %d", 1125 con->nodeid, ret); 1126 1127 /* Try again later */ 1128 clear_bit(CF_CONNECT_PENDING, &con->flags); 1129 clear_bit(CF_INIT_PENDING, &con->flags); 1130 } 1131 else { 1132 spin_lock(&con->writequeue_lock); 1133 writequeue_entry_complete(e, ret); 1134 spin_unlock(&con->writequeue_lock); 1135 } 1136 1137 unlock: 1138 mutex_unlock(&con->sock_mutex); 1139 } 1140 1141 /* Connect a new socket to its peer */ 1142 static void tcp_connect_to_sock(struct connection *con) 1143 { 1144 struct sockaddr_storage saddr, src_addr; 1145 int addr_len; 1146 struct socket *sock = NULL; 1147 int one = 1; 1148 int result; 1149 1150 if (con->nodeid == 0) { 1151 log_print("attempt to connect sock 0 foiled"); 1152 return; 1153 } 1154 1155 mutex_lock(&con->sock_mutex); 1156 if (con->retries++ > MAX_CONNECT_RETRIES) 1157 goto out; 1158 1159 /* Some odd races can cause double-connects, ignore them */ 1160 if (con->sock) 1161 goto out; 1162 1163 /* Create a socket to communicate with */ 1164 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM, 1165 IPPROTO_TCP, &sock); 1166 if (result < 0) 1167 goto out_err; 1168 1169 memset(&saddr, 0, sizeof(saddr)); 1170 result = nodeid_to_addr(con->nodeid, &saddr, NULL, false); 1171 if (result < 0) { 1172 log_print("no address for nodeid %d", con->nodeid); 1173 goto out_err; 1174 } 1175 1176 sock->sk->sk_user_data = con; 1177 con->rx_action = receive_from_sock; 1178 con->connect_action = tcp_connect_to_sock; 1179 add_sock(sock, con); 1180 1181 /* Bind to our cluster-known address connecting to avoid 1182 routing problems */ 1183 memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr)); 1184 make_sockaddr(&src_addr, 0, &addr_len); 1185 result = sock->ops->bind(sock, (struct sockaddr *) &src_addr, 1186 addr_len); 1187 if (result < 0) { 1188 log_print("could not bind for connect: %d", result); 1189 /* This *may* not indicate a critical error */ 1190 } 1191 1192 make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len); 1193 1194 log_print("connecting to %d", con->nodeid); 1195 1196 /* Turn off Nagle's algorithm */ 1197 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one, 1198 sizeof(one)); 1199 1200 result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len, 1201 O_NONBLOCK); 1202 if (result == -EINPROGRESS) 1203 result = 0; 1204 if (result == 0) 1205 goto out; 1206 1207 out_err: 1208 if (con->sock) { 1209 sock_release(con->sock); 1210 con->sock = NULL; 1211 } else if (sock) { 1212 sock_release(sock); 1213 } 1214 /* 1215 * Some errors are fatal and this list might need adjusting. For other 1216 * errors we try again until the max number of retries is reached. 1217 */ 1218 if (result != -EHOSTUNREACH && 1219 result != -ENETUNREACH && 1220 result != -ENETDOWN && 1221 result != -EINVAL && 1222 result != -EPROTONOSUPPORT) { 1223 log_print("connect %d try %d error %d", con->nodeid, 1224 con->retries, result); 1225 mutex_unlock(&con->sock_mutex); 1226 msleep(1000); 1227 lowcomms_connect_sock(con); 1228 return; 1229 } 1230 out: 1231 mutex_unlock(&con->sock_mutex); 1232 return; 1233 } 1234 1235 static struct socket *tcp_create_listen_sock(struct connection *con, 1236 struct sockaddr_storage *saddr) 1237 { 1238 struct socket *sock = NULL; 1239 int result = 0; 1240 int one = 1; 1241 int addr_len; 1242 1243 if (dlm_local_addr[0]->ss_family == AF_INET) 1244 addr_len = sizeof(struct sockaddr_in); 1245 else 1246 addr_len = sizeof(struct sockaddr_in6); 1247 1248 /* Create a socket to communicate with */ 1249 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM, 1250 IPPROTO_TCP, &sock); 1251 if (result < 0) { 1252 log_print("Can't create listening comms socket"); 1253 goto create_out; 1254 } 1255 1256 /* Turn off Nagle's algorithm */ 1257 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one, 1258 sizeof(one)); 1259 1260 result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, 1261 (char *)&one, sizeof(one)); 1262 1263 if (result < 0) { 1264 log_print("Failed to set SO_REUSEADDR on socket: %d", result); 1265 } 1266 con->rx_action = tcp_accept_from_sock; 1267 con->connect_action = tcp_connect_to_sock; 1268 1269 /* Bind to our port */ 1270 make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len); 1271 result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len); 1272 if (result < 0) { 1273 log_print("Can't bind to port %d", dlm_config.ci_tcp_port); 1274 sock_release(sock); 1275 sock = NULL; 1276 con->sock = NULL; 1277 goto create_out; 1278 } 1279 result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE, 1280 (char *)&one, sizeof(one)); 1281 if (result < 0) { 1282 log_print("Set keepalive failed: %d", result); 1283 } 1284 1285 result = sock->ops->listen(sock, 5); 1286 if (result < 0) { 1287 log_print("Can't listen on port %d", dlm_config.ci_tcp_port); 1288 sock_release(sock); 1289 sock = NULL; 1290 goto create_out; 1291 } 1292 1293 create_out: 1294 return sock; 1295 } 1296 1297 /* Get local addresses */ 1298 static void init_local(void) 1299 { 1300 struct sockaddr_storage sas, *addr; 1301 int i; 1302 1303 dlm_local_count = 0; 1304 for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) { 1305 if (dlm_our_addr(&sas, i)) 1306 break; 1307 1308 addr = kmalloc(sizeof(*addr), GFP_NOFS); 1309 if (!addr) 1310 break; 1311 memcpy(addr, &sas, sizeof(*addr)); 1312 dlm_local_addr[dlm_local_count++] = addr; 1313 } 1314 } 1315 1316 /* Bind to an IP address. SCTP allows multiple address so it can do 1317 multi-homing */ 1318 static int add_sctp_bind_addr(struct connection *sctp_con, 1319 struct sockaddr_storage *addr, 1320 int addr_len, int num) 1321 { 1322 int result = 0; 1323 1324 if (num == 1) 1325 result = kernel_bind(sctp_con->sock, 1326 (struct sockaddr *) addr, 1327 addr_len); 1328 else 1329 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP, 1330 SCTP_SOCKOPT_BINDX_ADD, 1331 (char *)addr, addr_len); 1332 1333 if (result < 0) 1334 log_print("Can't bind to port %d addr number %d", 1335 dlm_config.ci_tcp_port, num); 1336 1337 return result; 1338 } 1339 1340 /* Initialise SCTP socket and bind to all interfaces */ 1341 static int sctp_listen_for_all(void) 1342 { 1343 struct socket *sock = NULL; 1344 struct sockaddr_storage localaddr; 1345 struct sctp_event_subscribe subscribe; 1346 int result = -EINVAL, num = 1, i, addr_len; 1347 struct connection *con = nodeid2con(0, GFP_NOFS); 1348 int bufsize = NEEDED_RMEM; 1349 1350 if (!con) 1351 return -ENOMEM; 1352 1353 log_print("Using SCTP for communications"); 1354 1355 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET, 1356 IPPROTO_SCTP, &sock); 1357 if (result < 0) { 1358 log_print("Can't create comms socket, check SCTP is loaded"); 1359 goto out; 1360 } 1361 1362 /* Listen for events */ 1363 memset(&subscribe, 0, sizeof(subscribe)); 1364 subscribe.sctp_data_io_event = 1; 1365 subscribe.sctp_association_event = 1; 1366 subscribe.sctp_send_failure_event = 1; 1367 subscribe.sctp_shutdown_event = 1; 1368 subscribe.sctp_partial_delivery_event = 1; 1369 1370 result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE, 1371 (char *)&bufsize, sizeof(bufsize)); 1372 if (result) 1373 log_print("Error increasing buffer space on socket %d", result); 1374 1375 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS, 1376 (char *)&subscribe, sizeof(subscribe)); 1377 if (result < 0) { 1378 log_print("Failed to set SCTP_EVENTS on socket: result=%d", 1379 result); 1380 goto create_delsock; 1381 } 1382 1383 /* Init con struct */ 1384 sock->sk->sk_user_data = con; 1385 con->sock = sock; 1386 con->sock->sk->sk_data_ready = lowcomms_data_ready; 1387 con->rx_action = receive_from_sock; 1388 con->connect_action = sctp_init_assoc; 1389 1390 /* Bind to all interfaces. */ 1391 for (i = 0; i < dlm_local_count; i++) { 1392 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr)); 1393 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len); 1394 1395 result = add_sctp_bind_addr(con, &localaddr, addr_len, num); 1396 if (result) 1397 goto create_delsock; 1398 ++num; 1399 } 1400 1401 result = sock->ops->listen(sock, 5); 1402 if (result < 0) { 1403 log_print("Can't set socket listening"); 1404 goto create_delsock; 1405 } 1406 1407 return 0; 1408 1409 create_delsock: 1410 sock_release(sock); 1411 con->sock = NULL; 1412 out: 1413 return result; 1414 } 1415 1416 static int tcp_listen_for_all(void) 1417 { 1418 struct socket *sock = NULL; 1419 struct connection *con = nodeid2con(0, GFP_NOFS); 1420 int result = -EINVAL; 1421 1422 if (!con) 1423 return -ENOMEM; 1424 1425 /* We don't support multi-homed hosts */ 1426 if (dlm_local_addr[1] != NULL) { 1427 log_print("TCP protocol can't handle multi-homed hosts, " 1428 "try SCTP"); 1429 return -EINVAL; 1430 } 1431 1432 log_print("Using TCP for communications"); 1433 1434 sock = tcp_create_listen_sock(con, dlm_local_addr[0]); 1435 if (sock) { 1436 add_sock(sock, con); 1437 result = 0; 1438 } 1439 else { 1440 result = -EADDRINUSE; 1441 } 1442 1443 return result; 1444 } 1445 1446 1447 1448 static struct writequeue_entry *new_writequeue_entry(struct connection *con, 1449 gfp_t allocation) 1450 { 1451 struct writequeue_entry *entry; 1452 1453 entry = kmalloc(sizeof(struct writequeue_entry), allocation); 1454 if (!entry) 1455 return NULL; 1456 1457 entry->page = alloc_page(allocation); 1458 if (!entry->page) { 1459 kfree(entry); 1460 return NULL; 1461 } 1462 1463 entry->offset = 0; 1464 entry->len = 0; 1465 entry->end = 0; 1466 entry->users = 0; 1467 entry->con = con; 1468 1469 return entry; 1470 } 1471 1472 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc) 1473 { 1474 struct connection *con; 1475 struct writequeue_entry *e; 1476 int offset = 0; 1477 1478 con = nodeid2con(nodeid, allocation); 1479 if (!con) 1480 return NULL; 1481 1482 spin_lock(&con->writequeue_lock); 1483 e = list_entry(con->writequeue.prev, struct writequeue_entry, list); 1484 if ((&e->list == &con->writequeue) || 1485 (PAGE_CACHE_SIZE - e->end < len)) { 1486 e = NULL; 1487 } else { 1488 offset = e->end; 1489 e->end += len; 1490 e->users++; 1491 } 1492 spin_unlock(&con->writequeue_lock); 1493 1494 if (e) { 1495 got_one: 1496 *ppc = page_address(e->page) + offset; 1497 return e; 1498 } 1499 1500 e = new_writequeue_entry(con, allocation); 1501 if (e) { 1502 spin_lock(&con->writequeue_lock); 1503 offset = e->end; 1504 e->end += len; 1505 e->users++; 1506 list_add_tail(&e->list, &con->writequeue); 1507 spin_unlock(&con->writequeue_lock); 1508 goto got_one; 1509 } 1510 return NULL; 1511 } 1512 1513 void dlm_lowcomms_commit_buffer(void *mh) 1514 { 1515 struct writequeue_entry *e = (struct writequeue_entry *)mh; 1516 struct connection *con = e->con; 1517 int users; 1518 1519 spin_lock(&con->writequeue_lock); 1520 users = --e->users; 1521 if (users) 1522 goto out; 1523 e->len = e->end - e->offset; 1524 spin_unlock(&con->writequeue_lock); 1525 1526 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) { 1527 queue_work(send_workqueue, &con->swork); 1528 } 1529 return; 1530 1531 out: 1532 spin_unlock(&con->writequeue_lock); 1533 return; 1534 } 1535 1536 /* Send a message */ 1537 static void send_to_sock(struct connection *con) 1538 { 1539 int ret = 0; 1540 const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL; 1541 struct writequeue_entry *e; 1542 int len, offset; 1543 int count = 0; 1544 1545 mutex_lock(&con->sock_mutex); 1546 if (con->sock == NULL) 1547 goto out_connect; 1548 1549 spin_lock(&con->writequeue_lock); 1550 for (;;) { 1551 e = list_entry(con->writequeue.next, struct writequeue_entry, 1552 list); 1553 if ((struct list_head *) e == &con->writequeue) 1554 break; 1555 1556 len = e->len; 1557 offset = e->offset; 1558 BUG_ON(len == 0 && e->users == 0); 1559 spin_unlock(&con->writequeue_lock); 1560 1561 ret = 0; 1562 if (len) { 1563 ret = kernel_sendpage(con->sock, e->page, offset, len, 1564 msg_flags); 1565 if (ret == -EAGAIN || ret == 0) { 1566 if (ret == -EAGAIN && 1567 test_bit(SOCK_ASYNC_NOSPACE, &con->sock->flags) && 1568 !test_and_set_bit(CF_APP_LIMITED, &con->flags)) { 1569 /* Notify TCP that we're limited by the 1570 * application window size. 1571 */ 1572 set_bit(SOCK_NOSPACE, &con->sock->flags); 1573 con->sock->sk->sk_write_pending++; 1574 } 1575 cond_resched(); 1576 goto out; 1577 } else if (ret < 0) 1578 goto send_error; 1579 } 1580 1581 /* Don't starve people filling buffers */ 1582 if (++count >= MAX_SEND_MSG_COUNT) { 1583 cond_resched(); 1584 count = 0; 1585 } 1586 1587 spin_lock(&con->writequeue_lock); 1588 writequeue_entry_complete(e, ret); 1589 } 1590 spin_unlock(&con->writequeue_lock); 1591 out: 1592 mutex_unlock(&con->sock_mutex); 1593 return; 1594 1595 send_error: 1596 mutex_unlock(&con->sock_mutex); 1597 close_connection(con, false); 1598 lowcomms_connect_sock(con); 1599 return; 1600 1601 out_connect: 1602 mutex_unlock(&con->sock_mutex); 1603 if (!test_bit(CF_INIT_PENDING, &con->flags)) 1604 lowcomms_connect_sock(con); 1605 } 1606 1607 static void clean_one_writequeue(struct connection *con) 1608 { 1609 struct writequeue_entry *e, *safe; 1610 1611 spin_lock(&con->writequeue_lock); 1612 list_for_each_entry_safe(e, safe, &con->writequeue, list) { 1613 list_del(&e->list); 1614 free_entry(e); 1615 } 1616 spin_unlock(&con->writequeue_lock); 1617 } 1618 1619 /* Called from recovery when it knows that a node has 1620 left the cluster */ 1621 int dlm_lowcomms_close(int nodeid) 1622 { 1623 struct connection *con; 1624 struct dlm_node_addr *na; 1625 1626 log_print("closing connection to node %d", nodeid); 1627 con = nodeid2con(nodeid, 0); 1628 if (con) { 1629 clear_bit(CF_CONNECT_PENDING, &con->flags); 1630 clear_bit(CF_WRITE_PENDING, &con->flags); 1631 set_bit(CF_CLOSE, &con->flags); 1632 if (cancel_work_sync(&con->swork)) 1633 log_print("canceled swork for node %d", nodeid); 1634 if (cancel_work_sync(&con->rwork)) 1635 log_print("canceled rwork for node %d", nodeid); 1636 clean_one_writequeue(con); 1637 close_connection(con, true); 1638 } 1639 1640 spin_lock(&dlm_node_addrs_spin); 1641 na = find_node_addr(nodeid); 1642 if (na) { 1643 list_del(&na->list); 1644 while (na->addr_count--) 1645 kfree(na->addr[na->addr_count]); 1646 kfree(na); 1647 } 1648 spin_unlock(&dlm_node_addrs_spin); 1649 1650 return 0; 1651 } 1652 1653 /* Receive workqueue function */ 1654 static void process_recv_sockets(struct work_struct *work) 1655 { 1656 struct connection *con = container_of(work, struct connection, rwork); 1657 int err; 1658 1659 clear_bit(CF_READ_PENDING, &con->flags); 1660 do { 1661 err = con->rx_action(con); 1662 } while (!err); 1663 } 1664 1665 /* Send workqueue function */ 1666 static void process_send_sockets(struct work_struct *work) 1667 { 1668 struct connection *con = container_of(work, struct connection, swork); 1669 1670 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) { 1671 con->connect_action(con); 1672 set_bit(CF_WRITE_PENDING, &con->flags); 1673 } 1674 if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags)) 1675 send_to_sock(con); 1676 } 1677 1678 1679 /* Discard all entries on the write queues */ 1680 static void clean_writequeues(void) 1681 { 1682 foreach_conn(clean_one_writequeue); 1683 } 1684 1685 static void work_stop(void) 1686 { 1687 destroy_workqueue(recv_workqueue); 1688 destroy_workqueue(send_workqueue); 1689 } 1690 1691 static int work_start(void) 1692 { 1693 recv_workqueue = alloc_workqueue("dlm_recv", 1694 WQ_UNBOUND | WQ_MEM_RECLAIM, 1); 1695 if (!recv_workqueue) { 1696 log_print("can't start dlm_recv"); 1697 return -ENOMEM; 1698 } 1699 1700 send_workqueue = alloc_workqueue("dlm_send", 1701 WQ_UNBOUND | WQ_MEM_RECLAIM, 1); 1702 if (!send_workqueue) { 1703 log_print("can't start dlm_send"); 1704 destroy_workqueue(recv_workqueue); 1705 return -ENOMEM; 1706 } 1707 1708 return 0; 1709 } 1710 1711 static void stop_conn(struct connection *con) 1712 { 1713 con->flags |= 0x0F; 1714 if (con->sock && con->sock->sk) 1715 con->sock->sk->sk_user_data = NULL; 1716 } 1717 1718 static void free_conn(struct connection *con) 1719 { 1720 close_connection(con, true); 1721 if (con->othercon) 1722 kmem_cache_free(con_cache, con->othercon); 1723 hlist_del(&con->list); 1724 kmem_cache_free(con_cache, con); 1725 } 1726 1727 void dlm_lowcomms_stop(void) 1728 { 1729 /* Set all the flags to prevent any 1730 socket activity. 1731 */ 1732 mutex_lock(&connections_lock); 1733 dlm_allow_conn = 0; 1734 foreach_conn(stop_conn); 1735 mutex_unlock(&connections_lock); 1736 1737 work_stop(); 1738 1739 mutex_lock(&connections_lock); 1740 clean_writequeues(); 1741 1742 foreach_conn(free_conn); 1743 1744 mutex_unlock(&connections_lock); 1745 kmem_cache_destroy(con_cache); 1746 } 1747 1748 int dlm_lowcomms_start(void) 1749 { 1750 int error = -EINVAL; 1751 struct connection *con; 1752 int i; 1753 1754 for (i = 0; i < CONN_HASH_SIZE; i++) 1755 INIT_HLIST_HEAD(&connection_hash[i]); 1756 1757 init_local(); 1758 if (!dlm_local_count) { 1759 error = -ENOTCONN; 1760 log_print("no local IP address has been set"); 1761 goto fail; 1762 } 1763 1764 error = -ENOMEM; 1765 con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection), 1766 __alignof__(struct connection), 0, 1767 NULL); 1768 if (!con_cache) 1769 goto fail; 1770 1771 error = work_start(); 1772 if (error) 1773 goto fail_destroy; 1774 1775 dlm_allow_conn = 1; 1776 1777 /* Start listening */ 1778 if (dlm_config.ci_protocol == 0) 1779 error = tcp_listen_for_all(); 1780 else 1781 error = sctp_listen_for_all(); 1782 if (error) 1783 goto fail_unlisten; 1784 1785 return 0; 1786 1787 fail_unlisten: 1788 dlm_allow_conn = 0; 1789 con = nodeid2con(0,0); 1790 if (con) { 1791 close_connection(con, false); 1792 kmem_cache_free(con_cache, con); 1793 } 1794 fail_destroy: 1795 kmem_cache_destroy(con_cache); 1796 fail: 1797 return error; 1798 } 1799 1800 void dlm_lowcomms_exit(void) 1801 { 1802 struct dlm_node_addr *na, *safe; 1803 1804 spin_lock(&dlm_node_addrs_spin); 1805 list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) { 1806 list_del(&na->list); 1807 while (na->addr_count--) 1808 kfree(na->addr[na->addr_count]); 1809 kfree(na); 1810 } 1811 spin_unlock(&dlm_node_addrs_spin); 1812 } 1813