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 <net/sctp/sctp.h> 56 #include <net/ipv6.h> 57 58 #include "dlm_internal.h" 59 #include "lowcomms.h" 60 #include "midcomms.h" 61 #include "config.h" 62 63 #define NEEDED_RMEM (4*1024*1024) 64 #define CONN_HASH_SIZE 32 65 66 /* Number of messages to send before rescheduling */ 67 #define MAX_SEND_MSG_COUNT 25 68 69 struct cbuf { 70 unsigned int base; 71 unsigned int len; 72 unsigned int mask; 73 }; 74 75 static void cbuf_add(struct cbuf *cb, int n) 76 { 77 cb->len += n; 78 } 79 80 static int cbuf_data(struct cbuf *cb) 81 { 82 return ((cb->base + cb->len) & cb->mask); 83 } 84 85 static void cbuf_init(struct cbuf *cb, int size) 86 { 87 cb->base = cb->len = 0; 88 cb->mask = size-1; 89 } 90 91 static void cbuf_eat(struct cbuf *cb, int n) 92 { 93 cb->len -= n; 94 cb->base += n; 95 cb->base &= cb->mask; 96 } 97 98 static bool cbuf_empty(struct cbuf *cb) 99 { 100 return cb->len == 0; 101 } 102 103 struct connection { 104 struct socket *sock; /* NULL if not connected */ 105 uint32_t nodeid; /* So we know who we are in the list */ 106 struct mutex sock_mutex; 107 unsigned long flags; 108 #define CF_READ_PENDING 1 109 #define CF_WRITE_PENDING 2 110 #define CF_INIT_PENDING 4 111 #define CF_IS_OTHERCON 5 112 #define CF_CLOSE 6 113 #define CF_APP_LIMITED 7 114 #define CF_CLOSING 8 115 struct list_head writequeue; /* List of outgoing writequeue_entries */ 116 spinlock_t writequeue_lock; 117 int (*rx_action) (struct connection *); /* What to do when active */ 118 void (*connect_action) (struct connection *); /* What to do to connect */ 119 struct page *rx_page; 120 struct cbuf cb; 121 int retries; 122 #define MAX_CONNECT_RETRIES 3 123 struct hlist_node list; 124 struct connection *othercon; 125 struct work_struct rwork; /* Receive workqueue */ 126 struct work_struct swork; /* Send workqueue */ 127 }; 128 #define sock2con(x) ((struct connection *)(x)->sk_user_data) 129 130 /* An entry waiting to be sent */ 131 struct writequeue_entry { 132 struct list_head list; 133 struct page *page; 134 int offset; 135 int len; 136 int end; 137 int users; 138 struct connection *con; 139 }; 140 141 struct dlm_node_addr { 142 struct list_head list; 143 int nodeid; 144 int addr_count; 145 int curr_addr_index; 146 struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT]; 147 }; 148 149 static struct listen_sock_callbacks { 150 void (*sk_error_report)(struct sock *); 151 void (*sk_data_ready)(struct sock *); 152 void (*sk_state_change)(struct sock *); 153 void (*sk_write_space)(struct sock *); 154 } listen_sock; 155 156 static LIST_HEAD(dlm_node_addrs); 157 static DEFINE_SPINLOCK(dlm_node_addrs_spin); 158 159 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT]; 160 static int dlm_local_count; 161 static int dlm_allow_conn; 162 163 /* Work queues */ 164 static struct workqueue_struct *recv_workqueue; 165 static struct workqueue_struct *send_workqueue; 166 167 static struct hlist_head connection_hash[CONN_HASH_SIZE]; 168 static DEFINE_MUTEX(connections_lock); 169 static struct kmem_cache *con_cache; 170 171 static void process_recv_sockets(struct work_struct *work); 172 static void process_send_sockets(struct work_struct *work); 173 174 175 /* This is deliberately very simple because most clusters have simple 176 sequential nodeids, so we should be able to go straight to a connection 177 struct in the array */ 178 static inline int nodeid_hash(int nodeid) 179 { 180 return nodeid & (CONN_HASH_SIZE-1); 181 } 182 183 static struct connection *__find_con(int nodeid) 184 { 185 int r; 186 struct connection *con; 187 188 r = nodeid_hash(nodeid); 189 190 hlist_for_each_entry(con, &connection_hash[r], list) { 191 if (con->nodeid == nodeid) 192 return con; 193 } 194 return NULL; 195 } 196 197 /* 198 * If 'allocation' is zero then we don't attempt to create a new 199 * connection structure for this node. 200 */ 201 static struct connection *__nodeid2con(int nodeid, gfp_t alloc) 202 { 203 struct connection *con = NULL; 204 int r; 205 206 con = __find_con(nodeid); 207 if (con || !alloc) 208 return con; 209 210 con = kmem_cache_zalloc(con_cache, alloc); 211 if (!con) 212 return NULL; 213 214 r = nodeid_hash(nodeid); 215 hlist_add_head(&con->list, &connection_hash[r]); 216 217 con->nodeid = nodeid; 218 mutex_init(&con->sock_mutex); 219 INIT_LIST_HEAD(&con->writequeue); 220 spin_lock_init(&con->writequeue_lock); 221 INIT_WORK(&con->swork, process_send_sockets); 222 INIT_WORK(&con->rwork, process_recv_sockets); 223 224 /* Setup action pointers for child sockets */ 225 if (con->nodeid) { 226 struct connection *zerocon = __find_con(0); 227 228 con->connect_action = zerocon->connect_action; 229 if (!con->rx_action) 230 con->rx_action = zerocon->rx_action; 231 } 232 233 return con; 234 } 235 236 /* Loop round all connections */ 237 static void foreach_conn(void (*conn_func)(struct connection *c)) 238 { 239 int i; 240 struct hlist_node *n; 241 struct connection *con; 242 243 for (i = 0; i < CONN_HASH_SIZE; i++) { 244 hlist_for_each_entry_safe(con, n, &connection_hash[i], list) 245 conn_func(con); 246 } 247 } 248 249 static struct connection *nodeid2con(int nodeid, gfp_t allocation) 250 { 251 struct connection *con; 252 253 mutex_lock(&connections_lock); 254 con = __nodeid2con(nodeid, allocation); 255 mutex_unlock(&connections_lock); 256 257 return con; 258 } 259 260 static struct dlm_node_addr *find_node_addr(int nodeid) 261 { 262 struct dlm_node_addr *na; 263 264 list_for_each_entry(na, &dlm_node_addrs, list) { 265 if (na->nodeid == nodeid) 266 return na; 267 } 268 return NULL; 269 } 270 271 static int addr_compare(struct sockaddr_storage *x, struct sockaddr_storage *y) 272 { 273 switch (x->ss_family) { 274 case AF_INET: { 275 struct sockaddr_in *sinx = (struct sockaddr_in *)x; 276 struct sockaddr_in *siny = (struct sockaddr_in *)y; 277 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr) 278 return 0; 279 if (sinx->sin_port != siny->sin_port) 280 return 0; 281 break; 282 } 283 case AF_INET6: { 284 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x; 285 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y; 286 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr)) 287 return 0; 288 if (sinx->sin6_port != siny->sin6_port) 289 return 0; 290 break; 291 } 292 default: 293 return 0; 294 } 295 return 1; 296 } 297 298 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out, 299 struct sockaddr *sa_out, bool try_new_addr) 300 { 301 struct sockaddr_storage sas; 302 struct dlm_node_addr *na; 303 304 if (!dlm_local_count) 305 return -1; 306 307 spin_lock(&dlm_node_addrs_spin); 308 na = find_node_addr(nodeid); 309 if (na && na->addr_count) { 310 memcpy(&sas, na->addr[na->curr_addr_index], 311 sizeof(struct sockaddr_storage)); 312 313 if (try_new_addr) { 314 na->curr_addr_index++; 315 if (na->curr_addr_index == na->addr_count) 316 na->curr_addr_index = 0; 317 } 318 } 319 spin_unlock(&dlm_node_addrs_spin); 320 321 if (!na) 322 return -EEXIST; 323 324 if (!na->addr_count) 325 return -ENOENT; 326 327 if (sas_out) 328 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage)); 329 330 if (!sa_out) 331 return 0; 332 333 if (dlm_local_addr[0]->ss_family == AF_INET) { 334 struct sockaddr_in *in4 = (struct sockaddr_in *) &sas; 335 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out; 336 ret4->sin_addr.s_addr = in4->sin_addr.s_addr; 337 } else { 338 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) &sas; 339 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out; 340 ret6->sin6_addr = in6->sin6_addr; 341 } 342 343 return 0; 344 } 345 346 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid) 347 { 348 struct dlm_node_addr *na; 349 int rv = -EEXIST; 350 int addr_i; 351 352 spin_lock(&dlm_node_addrs_spin); 353 list_for_each_entry(na, &dlm_node_addrs, list) { 354 if (!na->addr_count) 355 continue; 356 357 for (addr_i = 0; addr_i < na->addr_count; addr_i++) { 358 if (addr_compare(na->addr[addr_i], addr)) { 359 *nodeid = na->nodeid; 360 rv = 0; 361 goto unlock; 362 } 363 } 364 } 365 unlock: 366 spin_unlock(&dlm_node_addrs_spin); 367 return rv; 368 } 369 370 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len) 371 { 372 struct sockaddr_storage *new_addr; 373 struct dlm_node_addr *new_node, *na; 374 375 new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS); 376 if (!new_node) 377 return -ENOMEM; 378 379 new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS); 380 if (!new_addr) { 381 kfree(new_node); 382 return -ENOMEM; 383 } 384 385 memcpy(new_addr, addr, len); 386 387 spin_lock(&dlm_node_addrs_spin); 388 na = find_node_addr(nodeid); 389 if (!na) { 390 new_node->nodeid = nodeid; 391 new_node->addr[0] = new_addr; 392 new_node->addr_count = 1; 393 list_add(&new_node->list, &dlm_node_addrs); 394 spin_unlock(&dlm_node_addrs_spin); 395 return 0; 396 } 397 398 if (na->addr_count >= DLM_MAX_ADDR_COUNT) { 399 spin_unlock(&dlm_node_addrs_spin); 400 kfree(new_addr); 401 kfree(new_node); 402 return -ENOSPC; 403 } 404 405 na->addr[na->addr_count++] = new_addr; 406 spin_unlock(&dlm_node_addrs_spin); 407 kfree(new_node); 408 return 0; 409 } 410 411 /* Data available on socket or listen socket received a connect */ 412 static void lowcomms_data_ready(struct sock *sk) 413 { 414 struct connection *con = sock2con(sk); 415 if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags)) 416 queue_work(recv_workqueue, &con->rwork); 417 } 418 419 static void lowcomms_write_space(struct sock *sk) 420 { 421 struct connection *con = sock2con(sk); 422 423 if (!con) 424 return; 425 426 clear_bit(SOCK_NOSPACE, &con->sock->flags); 427 428 if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) { 429 con->sock->sk->sk_write_pending--; 430 clear_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags); 431 } 432 433 queue_work(send_workqueue, &con->swork); 434 } 435 436 static inline void lowcomms_connect_sock(struct connection *con) 437 { 438 if (test_bit(CF_CLOSE, &con->flags)) 439 return; 440 queue_work(send_workqueue, &con->swork); 441 cond_resched(); 442 } 443 444 static void lowcomms_state_change(struct sock *sk) 445 { 446 /* SCTP layer is not calling sk_data_ready when the connection 447 * is done, so we catch the signal through here. Also, it 448 * doesn't switch socket state when entering shutdown, so we 449 * skip the write in that case. 450 */ 451 if (sk->sk_shutdown) { 452 if (sk->sk_shutdown == RCV_SHUTDOWN) 453 lowcomms_data_ready(sk); 454 } else if (sk->sk_state == TCP_ESTABLISHED) { 455 lowcomms_write_space(sk); 456 } 457 } 458 459 int dlm_lowcomms_connect_node(int nodeid) 460 { 461 struct connection *con; 462 463 if (nodeid == dlm_our_nodeid()) 464 return 0; 465 466 con = nodeid2con(nodeid, GFP_NOFS); 467 if (!con) 468 return -ENOMEM; 469 lowcomms_connect_sock(con); 470 return 0; 471 } 472 473 static void lowcomms_error_report(struct sock *sk) 474 { 475 struct connection *con; 476 struct sockaddr_storage saddr; 477 int buflen; 478 void (*orig_report)(struct sock *) = NULL; 479 480 read_lock_bh(&sk->sk_callback_lock); 481 con = sock2con(sk); 482 if (con == NULL) 483 goto out; 484 485 orig_report = listen_sock.sk_error_report; 486 if (con->sock == NULL || 487 kernel_getpeername(con->sock, (struct sockaddr *)&saddr, &buflen)) { 488 printk_ratelimited(KERN_ERR "dlm: node %d: socket error " 489 "sending to node %d, port %d, " 490 "sk_err=%d/%d\n", dlm_our_nodeid(), 491 con->nodeid, dlm_config.ci_tcp_port, 492 sk->sk_err, sk->sk_err_soft); 493 } else if (saddr.ss_family == AF_INET) { 494 struct sockaddr_in *sin4 = (struct sockaddr_in *)&saddr; 495 496 printk_ratelimited(KERN_ERR "dlm: node %d: socket error " 497 "sending to node %d at %pI4, port %d, " 498 "sk_err=%d/%d\n", dlm_our_nodeid(), 499 con->nodeid, &sin4->sin_addr.s_addr, 500 dlm_config.ci_tcp_port, sk->sk_err, 501 sk->sk_err_soft); 502 } else { 503 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&saddr; 504 505 printk_ratelimited(KERN_ERR "dlm: node %d: socket error " 506 "sending to node %d at %u.%u.%u.%u, " 507 "port %d, sk_err=%d/%d\n", dlm_our_nodeid(), 508 con->nodeid, sin6->sin6_addr.s6_addr32[0], 509 sin6->sin6_addr.s6_addr32[1], 510 sin6->sin6_addr.s6_addr32[2], 511 sin6->sin6_addr.s6_addr32[3], 512 dlm_config.ci_tcp_port, sk->sk_err, 513 sk->sk_err_soft); 514 } 515 out: 516 read_unlock_bh(&sk->sk_callback_lock); 517 if (orig_report) 518 orig_report(sk); 519 } 520 521 /* Note: sk_callback_lock must be locked before calling this function. */ 522 static void save_listen_callbacks(struct socket *sock) 523 { 524 struct sock *sk = sock->sk; 525 526 listen_sock.sk_data_ready = sk->sk_data_ready; 527 listen_sock.sk_state_change = sk->sk_state_change; 528 listen_sock.sk_write_space = sk->sk_write_space; 529 listen_sock.sk_error_report = sk->sk_error_report; 530 } 531 532 static void restore_callbacks(struct socket *sock) 533 { 534 struct sock *sk = sock->sk; 535 536 write_lock_bh(&sk->sk_callback_lock); 537 sk->sk_user_data = NULL; 538 sk->sk_data_ready = listen_sock.sk_data_ready; 539 sk->sk_state_change = listen_sock.sk_state_change; 540 sk->sk_write_space = listen_sock.sk_write_space; 541 sk->sk_error_report = listen_sock.sk_error_report; 542 write_unlock_bh(&sk->sk_callback_lock); 543 } 544 545 /* Make a socket active */ 546 static void add_sock(struct socket *sock, struct connection *con) 547 { 548 struct sock *sk = sock->sk; 549 550 write_lock_bh(&sk->sk_callback_lock); 551 con->sock = sock; 552 553 sk->sk_user_data = con; 554 /* Install a data_ready callback */ 555 sk->sk_data_ready = lowcomms_data_ready; 556 sk->sk_write_space = lowcomms_write_space; 557 sk->sk_state_change = lowcomms_state_change; 558 sk->sk_allocation = GFP_NOFS; 559 sk->sk_error_report = lowcomms_error_report; 560 write_unlock_bh(&sk->sk_callback_lock); 561 } 562 563 /* Add the port number to an IPv6 or 4 sockaddr and return the address 564 length */ 565 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port, 566 int *addr_len) 567 { 568 saddr->ss_family = dlm_local_addr[0]->ss_family; 569 if (saddr->ss_family == AF_INET) { 570 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr; 571 in4_addr->sin_port = cpu_to_be16(port); 572 *addr_len = sizeof(struct sockaddr_in); 573 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero)); 574 } else { 575 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr; 576 in6_addr->sin6_port = cpu_to_be16(port); 577 *addr_len = sizeof(struct sockaddr_in6); 578 } 579 memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len); 580 } 581 582 /* Close a remote connection and tidy up */ 583 static void close_connection(struct connection *con, bool and_other, 584 bool tx, bool rx) 585 { 586 bool closing = test_and_set_bit(CF_CLOSING, &con->flags); 587 588 if (tx && !closing && cancel_work_sync(&con->swork)) 589 log_print("canceled swork for node %d", con->nodeid); 590 if (rx && !closing && cancel_work_sync(&con->rwork)) 591 log_print("canceled rwork for node %d", con->nodeid); 592 593 mutex_lock(&con->sock_mutex); 594 if (con->sock) { 595 restore_callbacks(con->sock); 596 sock_release(con->sock); 597 con->sock = NULL; 598 } 599 if (con->othercon && and_other) { 600 /* Will only re-enter once. */ 601 close_connection(con->othercon, false, true, true); 602 } 603 if (con->rx_page) { 604 __free_page(con->rx_page); 605 con->rx_page = NULL; 606 } 607 608 con->retries = 0; 609 mutex_unlock(&con->sock_mutex); 610 clear_bit(CF_CLOSING, &con->flags); 611 } 612 613 /* Data received from remote end */ 614 static int receive_from_sock(struct connection *con) 615 { 616 int ret = 0; 617 struct msghdr msg = {}; 618 struct kvec iov[2]; 619 unsigned len; 620 int r; 621 int call_again_soon = 0; 622 int nvec; 623 624 mutex_lock(&con->sock_mutex); 625 626 if (con->sock == NULL) { 627 ret = -EAGAIN; 628 goto out_close; 629 } 630 if (con->nodeid == 0) { 631 ret = -EINVAL; 632 goto out_close; 633 } 634 635 if (con->rx_page == NULL) { 636 /* 637 * This doesn't need to be atomic, but I think it should 638 * improve performance if it is. 639 */ 640 con->rx_page = alloc_page(GFP_ATOMIC); 641 if (con->rx_page == NULL) 642 goto out_resched; 643 cbuf_init(&con->cb, PAGE_SIZE); 644 } 645 646 /* 647 * iov[0] is the bit of the circular buffer between the current end 648 * point (cb.base + cb.len) and the end of the buffer. 649 */ 650 iov[0].iov_len = con->cb.base - cbuf_data(&con->cb); 651 iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb); 652 iov[1].iov_len = 0; 653 nvec = 1; 654 655 /* 656 * iov[1] is the bit of the circular buffer between the start of the 657 * buffer and the start of the currently used section (cb.base) 658 */ 659 if (cbuf_data(&con->cb) >= con->cb.base) { 660 iov[0].iov_len = PAGE_SIZE - cbuf_data(&con->cb); 661 iov[1].iov_len = con->cb.base; 662 iov[1].iov_base = page_address(con->rx_page); 663 nvec = 2; 664 } 665 len = iov[0].iov_len + iov[1].iov_len; 666 667 r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len, 668 MSG_DONTWAIT | MSG_NOSIGNAL); 669 if (ret <= 0) 670 goto out_close; 671 else if (ret == len) 672 call_again_soon = 1; 673 674 cbuf_add(&con->cb, ret); 675 ret = dlm_process_incoming_buffer(con->nodeid, 676 page_address(con->rx_page), 677 con->cb.base, con->cb.len, 678 PAGE_SIZE); 679 if (ret == -EBADMSG) { 680 log_print("lowcomms: addr=%p, base=%u, len=%u, read=%d", 681 page_address(con->rx_page), con->cb.base, 682 con->cb.len, r); 683 } 684 if (ret < 0) 685 goto out_close; 686 cbuf_eat(&con->cb, ret); 687 688 if (cbuf_empty(&con->cb) && !call_again_soon) { 689 __free_page(con->rx_page); 690 con->rx_page = NULL; 691 } 692 693 if (call_again_soon) 694 goto out_resched; 695 mutex_unlock(&con->sock_mutex); 696 return 0; 697 698 out_resched: 699 if (!test_and_set_bit(CF_READ_PENDING, &con->flags)) 700 queue_work(recv_workqueue, &con->rwork); 701 mutex_unlock(&con->sock_mutex); 702 return -EAGAIN; 703 704 out_close: 705 mutex_unlock(&con->sock_mutex); 706 if (ret != -EAGAIN) { 707 close_connection(con, true, true, false); 708 /* Reconnect when there is something to send */ 709 } 710 /* Don't return success if we really got EOF */ 711 if (ret == 0) 712 ret = -EAGAIN; 713 714 return ret; 715 } 716 717 /* Listening socket is busy, accept a connection */ 718 static int tcp_accept_from_sock(struct connection *con) 719 { 720 int result; 721 struct sockaddr_storage peeraddr; 722 struct socket *newsock; 723 int len; 724 int nodeid; 725 struct connection *newcon; 726 struct connection *addcon; 727 728 mutex_lock(&connections_lock); 729 if (!dlm_allow_conn) { 730 mutex_unlock(&connections_lock); 731 return -1; 732 } 733 mutex_unlock(&connections_lock); 734 735 memset(&peeraddr, 0, sizeof(peeraddr)); 736 result = sock_create_lite(dlm_local_addr[0]->ss_family, 737 SOCK_STREAM, IPPROTO_TCP, &newsock); 738 if (result < 0) 739 return -ENOMEM; 740 741 mutex_lock_nested(&con->sock_mutex, 0); 742 743 result = -ENOTCONN; 744 if (con->sock == NULL) 745 goto accept_err; 746 747 newsock->type = con->sock->type; 748 newsock->ops = con->sock->ops; 749 750 result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK, true); 751 if (result < 0) 752 goto accept_err; 753 754 /* Get the connected socket's peer */ 755 memset(&peeraddr, 0, sizeof(peeraddr)); 756 if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr, 757 &len, 2)) { 758 result = -ECONNABORTED; 759 goto accept_err; 760 } 761 762 /* Get the new node's NODEID */ 763 make_sockaddr(&peeraddr, 0, &len); 764 if (addr_to_nodeid(&peeraddr, &nodeid)) { 765 unsigned char *b=(unsigned char *)&peeraddr; 766 log_print("connect from non cluster node"); 767 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE, 768 b, sizeof(struct sockaddr_storage)); 769 sock_release(newsock); 770 mutex_unlock(&con->sock_mutex); 771 return -1; 772 } 773 774 log_print("got connection from %d", nodeid); 775 776 /* Check to see if we already have a connection to this node. This 777 * could happen if the two nodes initiate a connection at roughly 778 * the same time and the connections cross on the wire. 779 * In this case we store the incoming one in "othercon" 780 */ 781 newcon = nodeid2con(nodeid, GFP_NOFS); 782 if (!newcon) { 783 result = -ENOMEM; 784 goto accept_err; 785 } 786 mutex_lock_nested(&newcon->sock_mutex, 1); 787 if (newcon->sock) { 788 struct connection *othercon = newcon->othercon; 789 790 if (!othercon) { 791 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS); 792 if (!othercon) { 793 log_print("failed to allocate incoming socket"); 794 mutex_unlock(&newcon->sock_mutex); 795 result = -ENOMEM; 796 goto accept_err; 797 } 798 othercon->nodeid = nodeid; 799 othercon->rx_action = receive_from_sock; 800 mutex_init(&othercon->sock_mutex); 801 INIT_WORK(&othercon->swork, process_send_sockets); 802 INIT_WORK(&othercon->rwork, process_recv_sockets); 803 set_bit(CF_IS_OTHERCON, &othercon->flags); 804 } 805 mutex_lock_nested(&othercon->sock_mutex, 2); 806 if (!othercon->sock) { 807 newcon->othercon = othercon; 808 othercon->sock = newsock; 809 newsock->sk->sk_user_data = othercon; 810 add_sock(newsock, othercon); 811 addcon = othercon; 812 mutex_unlock(&othercon->sock_mutex); 813 } 814 else { 815 printk("Extra connection from node %d attempted\n", nodeid); 816 result = -EAGAIN; 817 mutex_unlock(&othercon->sock_mutex); 818 mutex_unlock(&newcon->sock_mutex); 819 goto accept_err; 820 } 821 } 822 else { 823 newsock->sk->sk_user_data = newcon; 824 newcon->rx_action = receive_from_sock; 825 /* accept copies the sk after we've saved the callbacks, so we 826 don't want to save them a second time or comm errors will 827 result in calling sk_error_report recursively. */ 828 add_sock(newsock, newcon); 829 addcon = newcon; 830 } 831 832 mutex_unlock(&newcon->sock_mutex); 833 834 /* 835 * Add it to the active queue in case we got data 836 * between processing the accept adding the socket 837 * to the read_sockets list 838 */ 839 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags)) 840 queue_work(recv_workqueue, &addcon->rwork); 841 mutex_unlock(&con->sock_mutex); 842 843 return 0; 844 845 accept_err: 846 mutex_unlock(&con->sock_mutex); 847 sock_release(newsock); 848 849 if (result != -EAGAIN) 850 log_print("error accepting connection from node: %d", result); 851 return result; 852 } 853 854 static int sctp_accept_from_sock(struct connection *con) 855 { 856 /* Check that the new node is in the lockspace */ 857 struct sctp_prim prim; 858 int nodeid; 859 int prim_len, ret; 860 int addr_len; 861 struct connection *newcon; 862 struct connection *addcon; 863 struct socket *newsock; 864 865 mutex_lock(&connections_lock); 866 if (!dlm_allow_conn) { 867 mutex_unlock(&connections_lock); 868 return -1; 869 } 870 mutex_unlock(&connections_lock); 871 872 mutex_lock_nested(&con->sock_mutex, 0); 873 874 ret = kernel_accept(con->sock, &newsock, O_NONBLOCK); 875 if (ret < 0) 876 goto accept_err; 877 878 memset(&prim, 0, sizeof(struct sctp_prim)); 879 prim_len = sizeof(struct sctp_prim); 880 881 ret = kernel_getsockopt(newsock, IPPROTO_SCTP, SCTP_PRIMARY_ADDR, 882 (char *)&prim, &prim_len); 883 if (ret < 0) { 884 log_print("getsockopt/sctp_primary_addr failed: %d", ret); 885 goto accept_err; 886 } 887 888 make_sockaddr(&prim.ssp_addr, 0, &addr_len); 889 ret = addr_to_nodeid(&prim.ssp_addr, &nodeid); 890 if (ret) { 891 unsigned char *b = (unsigned char *)&prim.ssp_addr; 892 893 log_print("reject connect from unknown addr"); 894 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE, 895 b, sizeof(struct sockaddr_storage)); 896 goto accept_err; 897 } 898 899 newcon = nodeid2con(nodeid, GFP_NOFS); 900 if (!newcon) { 901 ret = -ENOMEM; 902 goto accept_err; 903 } 904 905 mutex_lock_nested(&newcon->sock_mutex, 1); 906 907 if (newcon->sock) { 908 struct connection *othercon = newcon->othercon; 909 910 if (!othercon) { 911 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS); 912 if (!othercon) { 913 log_print("failed to allocate incoming socket"); 914 mutex_unlock(&newcon->sock_mutex); 915 ret = -ENOMEM; 916 goto accept_err; 917 } 918 othercon->nodeid = nodeid; 919 othercon->rx_action = receive_from_sock; 920 mutex_init(&othercon->sock_mutex); 921 INIT_WORK(&othercon->swork, process_send_sockets); 922 INIT_WORK(&othercon->rwork, process_recv_sockets); 923 set_bit(CF_IS_OTHERCON, &othercon->flags); 924 } 925 mutex_lock_nested(&othercon->sock_mutex, 2); 926 if (!othercon->sock) { 927 newcon->othercon = othercon; 928 othercon->sock = newsock; 929 newsock->sk->sk_user_data = othercon; 930 add_sock(newsock, othercon); 931 addcon = othercon; 932 mutex_unlock(&othercon->sock_mutex); 933 } else { 934 printk("Extra connection from node %d attempted\n", nodeid); 935 ret = -EAGAIN; 936 mutex_unlock(&othercon->sock_mutex); 937 mutex_unlock(&newcon->sock_mutex); 938 goto accept_err; 939 } 940 } else { 941 newsock->sk->sk_user_data = newcon; 942 newcon->rx_action = receive_from_sock; 943 add_sock(newsock, newcon); 944 addcon = newcon; 945 } 946 947 log_print("connected to %d", nodeid); 948 949 mutex_unlock(&newcon->sock_mutex); 950 951 /* 952 * Add it to the active queue in case we got data 953 * between processing the accept adding the socket 954 * to the read_sockets list 955 */ 956 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags)) 957 queue_work(recv_workqueue, &addcon->rwork); 958 mutex_unlock(&con->sock_mutex); 959 960 return 0; 961 962 accept_err: 963 mutex_unlock(&con->sock_mutex); 964 if (newsock) 965 sock_release(newsock); 966 if (ret != -EAGAIN) 967 log_print("error accepting connection from node: %d", ret); 968 969 return ret; 970 } 971 972 static void free_entry(struct writequeue_entry *e) 973 { 974 __free_page(e->page); 975 kfree(e); 976 } 977 978 /* 979 * writequeue_entry_complete - try to delete and free write queue entry 980 * @e: write queue entry to try to delete 981 * @completed: bytes completed 982 * 983 * writequeue_lock must be held. 984 */ 985 static void writequeue_entry_complete(struct writequeue_entry *e, int completed) 986 { 987 e->offset += completed; 988 e->len -= completed; 989 990 if (e->len == 0 && e->users == 0) { 991 list_del(&e->list); 992 free_entry(e); 993 } 994 } 995 996 /* 997 * sctp_bind_addrs - bind a SCTP socket to all our addresses 998 */ 999 static int sctp_bind_addrs(struct connection *con, uint16_t port) 1000 { 1001 struct sockaddr_storage localaddr; 1002 int i, addr_len, result = 0; 1003 1004 for (i = 0; i < dlm_local_count; i++) { 1005 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr)); 1006 make_sockaddr(&localaddr, port, &addr_len); 1007 1008 if (!i) 1009 result = kernel_bind(con->sock, 1010 (struct sockaddr *)&localaddr, 1011 addr_len); 1012 else 1013 result = kernel_setsockopt(con->sock, SOL_SCTP, 1014 SCTP_SOCKOPT_BINDX_ADD, 1015 (char *)&localaddr, addr_len); 1016 1017 if (result < 0) { 1018 log_print("Can't bind to %d addr number %d, %d.\n", 1019 port, i + 1, result); 1020 break; 1021 } 1022 } 1023 return result; 1024 } 1025 1026 /* Initiate an SCTP association. 1027 This is a special case of send_to_sock() in that we don't yet have a 1028 peeled-off socket for this association, so we use the listening socket 1029 and add the primary IP address of the remote node. 1030 */ 1031 static void sctp_connect_to_sock(struct connection *con) 1032 { 1033 struct sockaddr_storage daddr; 1034 int one = 1; 1035 int result; 1036 int addr_len; 1037 struct socket *sock; 1038 1039 if (con->nodeid == 0) { 1040 log_print("attempt to connect sock 0 foiled"); 1041 return; 1042 } 1043 1044 mutex_lock(&con->sock_mutex); 1045 1046 /* Some odd races can cause double-connects, ignore them */ 1047 if (con->retries++ > MAX_CONNECT_RETRIES) 1048 goto out; 1049 1050 if (con->sock) { 1051 log_print("node %d already connected.", con->nodeid); 1052 goto out; 1053 } 1054 1055 memset(&daddr, 0, sizeof(daddr)); 1056 result = nodeid_to_addr(con->nodeid, &daddr, NULL, true); 1057 if (result < 0) { 1058 log_print("no address for nodeid %d", con->nodeid); 1059 goto out; 1060 } 1061 1062 /* Create a socket to communicate with */ 1063 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 1064 SOCK_STREAM, IPPROTO_SCTP, &sock); 1065 if (result < 0) 1066 goto socket_err; 1067 1068 sock->sk->sk_user_data = con; 1069 con->rx_action = receive_from_sock; 1070 con->connect_action = sctp_connect_to_sock; 1071 add_sock(sock, con); 1072 1073 /* Bind to all addresses. */ 1074 if (sctp_bind_addrs(con, 0)) 1075 goto bind_err; 1076 1077 make_sockaddr(&daddr, dlm_config.ci_tcp_port, &addr_len); 1078 1079 log_print("connecting to %d", con->nodeid); 1080 1081 /* Turn off Nagle's algorithm */ 1082 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one, 1083 sizeof(one)); 1084 1085 result = sock->ops->connect(sock, (struct sockaddr *)&daddr, addr_len, 1086 O_NONBLOCK); 1087 if (result == -EINPROGRESS) 1088 result = 0; 1089 if (result == 0) 1090 goto out; 1091 1092 bind_err: 1093 con->sock = NULL; 1094 sock_release(sock); 1095 1096 socket_err: 1097 /* 1098 * Some errors are fatal and this list might need adjusting. For other 1099 * errors we try again until the max number of retries is reached. 1100 */ 1101 if (result != -EHOSTUNREACH && 1102 result != -ENETUNREACH && 1103 result != -ENETDOWN && 1104 result != -EINVAL && 1105 result != -EPROTONOSUPPORT) { 1106 log_print("connect %d try %d error %d", con->nodeid, 1107 con->retries, result); 1108 mutex_unlock(&con->sock_mutex); 1109 msleep(1000); 1110 lowcomms_connect_sock(con); 1111 return; 1112 } 1113 1114 out: 1115 mutex_unlock(&con->sock_mutex); 1116 } 1117 1118 /* Connect a new socket to its peer */ 1119 static void tcp_connect_to_sock(struct connection *con) 1120 { 1121 struct sockaddr_storage saddr, src_addr; 1122 int addr_len; 1123 struct socket *sock = NULL; 1124 int one = 1; 1125 int result; 1126 1127 if (con->nodeid == 0) { 1128 log_print("attempt to connect sock 0 foiled"); 1129 return; 1130 } 1131 1132 mutex_lock(&con->sock_mutex); 1133 if (con->retries++ > MAX_CONNECT_RETRIES) 1134 goto out; 1135 1136 /* Some odd races can cause double-connects, ignore them */ 1137 if (con->sock) 1138 goto out; 1139 1140 /* Create a socket to communicate with */ 1141 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 1142 SOCK_STREAM, IPPROTO_TCP, &sock); 1143 if (result < 0) 1144 goto out_err; 1145 1146 memset(&saddr, 0, sizeof(saddr)); 1147 result = nodeid_to_addr(con->nodeid, &saddr, NULL, false); 1148 if (result < 0) { 1149 log_print("no address for nodeid %d", con->nodeid); 1150 goto out_err; 1151 } 1152 1153 sock->sk->sk_user_data = con; 1154 con->rx_action = receive_from_sock; 1155 con->connect_action = tcp_connect_to_sock; 1156 add_sock(sock, con); 1157 1158 /* Bind to our cluster-known address connecting to avoid 1159 routing problems */ 1160 memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr)); 1161 make_sockaddr(&src_addr, 0, &addr_len); 1162 result = sock->ops->bind(sock, (struct sockaddr *) &src_addr, 1163 addr_len); 1164 if (result < 0) { 1165 log_print("could not bind for connect: %d", result); 1166 /* This *may* not indicate a critical error */ 1167 } 1168 1169 make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len); 1170 1171 log_print("connecting to %d", con->nodeid); 1172 1173 /* Turn off Nagle's algorithm */ 1174 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one, 1175 sizeof(one)); 1176 1177 result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len, 1178 O_NONBLOCK); 1179 if (result == -EINPROGRESS) 1180 result = 0; 1181 if (result == 0) 1182 goto out; 1183 1184 out_err: 1185 if (con->sock) { 1186 sock_release(con->sock); 1187 con->sock = NULL; 1188 } else if (sock) { 1189 sock_release(sock); 1190 } 1191 /* 1192 * Some errors are fatal and this list might need adjusting. For other 1193 * errors we try again until the max number of retries is reached. 1194 */ 1195 if (result != -EHOSTUNREACH && 1196 result != -ENETUNREACH && 1197 result != -ENETDOWN && 1198 result != -EINVAL && 1199 result != -EPROTONOSUPPORT) { 1200 log_print("connect %d try %d error %d", con->nodeid, 1201 con->retries, result); 1202 mutex_unlock(&con->sock_mutex); 1203 msleep(1000); 1204 lowcomms_connect_sock(con); 1205 return; 1206 } 1207 out: 1208 mutex_unlock(&con->sock_mutex); 1209 return; 1210 } 1211 1212 static struct socket *tcp_create_listen_sock(struct connection *con, 1213 struct sockaddr_storage *saddr) 1214 { 1215 struct socket *sock = NULL; 1216 int result = 0; 1217 int one = 1; 1218 int addr_len; 1219 1220 if (dlm_local_addr[0]->ss_family == AF_INET) 1221 addr_len = sizeof(struct sockaddr_in); 1222 else 1223 addr_len = sizeof(struct sockaddr_in6); 1224 1225 /* Create a socket to communicate with */ 1226 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 1227 SOCK_STREAM, IPPROTO_TCP, &sock); 1228 if (result < 0) { 1229 log_print("Can't create listening comms socket"); 1230 goto create_out; 1231 } 1232 1233 /* Turn off Nagle's algorithm */ 1234 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one, 1235 sizeof(one)); 1236 1237 result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, 1238 (char *)&one, sizeof(one)); 1239 1240 if (result < 0) { 1241 log_print("Failed to set SO_REUSEADDR on socket: %d", result); 1242 } 1243 sock->sk->sk_user_data = con; 1244 save_listen_callbacks(sock); 1245 con->rx_action = tcp_accept_from_sock; 1246 con->connect_action = tcp_connect_to_sock; 1247 1248 /* Bind to our port */ 1249 make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len); 1250 result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len); 1251 if (result < 0) { 1252 log_print("Can't bind to port %d", dlm_config.ci_tcp_port); 1253 sock_release(sock); 1254 sock = NULL; 1255 con->sock = NULL; 1256 goto create_out; 1257 } 1258 result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE, 1259 (char *)&one, sizeof(one)); 1260 if (result < 0) { 1261 log_print("Set keepalive failed: %d", result); 1262 } 1263 1264 result = sock->ops->listen(sock, 5); 1265 if (result < 0) { 1266 log_print("Can't listen on port %d", dlm_config.ci_tcp_port); 1267 sock_release(sock); 1268 sock = NULL; 1269 goto create_out; 1270 } 1271 1272 create_out: 1273 return sock; 1274 } 1275 1276 /* Get local addresses */ 1277 static void init_local(void) 1278 { 1279 struct sockaddr_storage sas, *addr; 1280 int i; 1281 1282 dlm_local_count = 0; 1283 for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) { 1284 if (dlm_our_addr(&sas, i)) 1285 break; 1286 1287 addr = kmemdup(&sas, sizeof(*addr), GFP_NOFS); 1288 if (!addr) 1289 break; 1290 dlm_local_addr[dlm_local_count++] = addr; 1291 } 1292 } 1293 1294 /* Initialise SCTP socket and bind to all interfaces */ 1295 static int sctp_listen_for_all(void) 1296 { 1297 struct socket *sock = NULL; 1298 int result = -EINVAL; 1299 struct connection *con = nodeid2con(0, GFP_NOFS); 1300 int bufsize = NEEDED_RMEM; 1301 int one = 1; 1302 1303 if (!con) 1304 return -ENOMEM; 1305 1306 log_print("Using SCTP for communications"); 1307 1308 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 1309 SOCK_STREAM, IPPROTO_SCTP, &sock); 1310 if (result < 0) { 1311 log_print("Can't create comms socket, check SCTP is loaded"); 1312 goto out; 1313 } 1314 1315 result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE, 1316 (char *)&bufsize, sizeof(bufsize)); 1317 if (result) 1318 log_print("Error increasing buffer space on socket %d", result); 1319 1320 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one, 1321 sizeof(one)); 1322 if (result < 0) 1323 log_print("Could not set SCTP NODELAY error %d\n", result); 1324 1325 write_lock_bh(&sock->sk->sk_callback_lock); 1326 /* Init con struct */ 1327 sock->sk->sk_user_data = con; 1328 save_listen_callbacks(sock); 1329 con->sock = sock; 1330 con->sock->sk->sk_data_ready = lowcomms_data_ready; 1331 con->rx_action = sctp_accept_from_sock; 1332 con->connect_action = sctp_connect_to_sock; 1333 1334 write_unlock_bh(&sock->sk->sk_callback_lock); 1335 1336 /* Bind to all addresses. */ 1337 if (sctp_bind_addrs(con, dlm_config.ci_tcp_port)) 1338 goto create_delsock; 1339 1340 result = sock->ops->listen(sock, 5); 1341 if (result < 0) { 1342 log_print("Can't set socket listening"); 1343 goto create_delsock; 1344 } 1345 1346 return 0; 1347 1348 create_delsock: 1349 sock_release(sock); 1350 con->sock = NULL; 1351 out: 1352 return result; 1353 } 1354 1355 static int tcp_listen_for_all(void) 1356 { 1357 struct socket *sock = NULL; 1358 struct connection *con = nodeid2con(0, GFP_NOFS); 1359 int result = -EINVAL; 1360 1361 if (!con) 1362 return -ENOMEM; 1363 1364 /* We don't support multi-homed hosts */ 1365 if (dlm_local_addr[1] != NULL) { 1366 log_print("TCP protocol can't handle multi-homed hosts, " 1367 "try SCTP"); 1368 return -EINVAL; 1369 } 1370 1371 log_print("Using TCP for communications"); 1372 1373 sock = tcp_create_listen_sock(con, dlm_local_addr[0]); 1374 if (sock) { 1375 add_sock(sock, con); 1376 result = 0; 1377 } 1378 else { 1379 result = -EADDRINUSE; 1380 } 1381 1382 return result; 1383 } 1384 1385 1386 1387 static struct writequeue_entry *new_writequeue_entry(struct connection *con, 1388 gfp_t allocation) 1389 { 1390 struct writequeue_entry *entry; 1391 1392 entry = kmalloc(sizeof(struct writequeue_entry), allocation); 1393 if (!entry) 1394 return NULL; 1395 1396 entry->page = alloc_page(allocation); 1397 if (!entry->page) { 1398 kfree(entry); 1399 return NULL; 1400 } 1401 1402 entry->offset = 0; 1403 entry->len = 0; 1404 entry->end = 0; 1405 entry->users = 0; 1406 entry->con = con; 1407 1408 return entry; 1409 } 1410 1411 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc) 1412 { 1413 struct connection *con; 1414 struct writequeue_entry *e; 1415 int offset = 0; 1416 1417 con = nodeid2con(nodeid, allocation); 1418 if (!con) 1419 return NULL; 1420 1421 spin_lock(&con->writequeue_lock); 1422 e = list_entry(con->writequeue.prev, struct writequeue_entry, list); 1423 if ((&e->list == &con->writequeue) || 1424 (PAGE_SIZE - e->end < len)) { 1425 e = NULL; 1426 } else { 1427 offset = e->end; 1428 e->end += len; 1429 e->users++; 1430 } 1431 spin_unlock(&con->writequeue_lock); 1432 1433 if (e) { 1434 got_one: 1435 *ppc = page_address(e->page) + offset; 1436 return e; 1437 } 1438 1439 e = new_writequeue_entry(con, allocation); 1440 if (e) { 1441 spin_lock(&con->writequeue_lock); 1442 offset = e->end; 1443 e->end += len; 1444 e->users++; 1445 list_add_tail(&e->list, &con->writequeue); 1446 spin_unlock(&con->writequeue_lock); 1447 goto got_one; 1448 } 1449 return NULL; 1450 } 1451 1452 void dlm_lowcomms_commit_buffer(void *mh) 1453 { 1454 struct writequeue_entry *e = (struct writequeue_entry *)mh; 1455 struct connection *con = e->con; 1456 int users; 1457 1458 spin_lock(&con->writequeue_lock); 1459 users = --e->users; 1460 if (users) 1461 goto out; 1462 e->len = e->end - e->offset; 1463 spin_unlock(&con->writequeue_lock); 1464 1465 queue_work(send_workqueue, &con->swork); 1466 return; 1467 1468 out: 1469 spin_unlock(&con->writequeue_lock); 1470 return; 1471 } 1472 1473 /* Send a message */ 1474 static void send_to_sock(struct connection *con) 1475 { 1476 int ret = 0; 1477 const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL; 1478 struct writequeue_entry *e; 1479 int len, offset; 1480 int count = 0; 1481 1482 mutex_lock(&con->sock_mutex); 1483 if (con->sock == NULL) 1484 goto out_connect; 1485 1486 spin_lock(&con->writequeue_lock); 1487 for (;;) { 1488 e = list_entry(con->writequeue.next, struct writequeue_entry, 1489 list); 1490 if ((struct list_head *) e == &con->writequeue) 1491 break; 1492 1493 len = e->len; 1494 offset = e->offset; 1495 BUG_ON(len == 0 && e->users == 0); 1496 spin_unlock(&con->writequeue_lock); 1497 1498 ret = 0; 1499 if (len) { 1500 ret = kernel_sendpage(con->sock, e->page, offset, len, 1501 msg_flags); 1502 if (ret == -EAGAIN || ret == 0) { 1503 if (ret == -EAGAIN && 1504 test_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags) && 1505 !test_and_set_bit(CF_APP_LIMITED, &con->flags)) { 1506 /* Notify TCP that we're limited by the 1507 * application window size. 1508 */ 1509 set_bit(SOCK_NOSPACE, &con->sock->flags); 1510 con->sock->sk->sk_write_pending++; 1511 } 1512 cond_resched(); 1513 goto out; 1514 } else if (ret < 0) 1515 goto send_error; 1516 } 1517 1518 /* Don't starve people filling buffers */ 1519 if (++count >= MAX_SEND_MSG_COUNT) { 1520 cond_resched(); 1521 count = 0; 1522 } 1523 1524 spin_lock(&con->writequeue_lock); 1525 writequeue_entry_complete(e, ret); 1526 } 1527 spin_unlock(&con->writequeue_lock); 1528 out: 1529 mutex_unlock(&con->sock_mutex); 1530 return; 1531 1532 send_error: 1533 mutex_unlock(&con->sock_mutex); 1534 close_connection(con, true, false, true); 1535 /* Requeue the send work. When the work daemon runs again, it will try 1536 a new connection, then call this function again. */ 1537 queue_work(send_workqueue, &con->swork); 1538 return; 1539 1540 out_connect: 1541 mutex_unlock(&con->sock_mutex); 1542 queue_work(send_workqueue, &con->swork); 1543 cond_resched(); 1544 } 1545 1546 static void clean_one_writequeue(struct connection *con) 1547 { 1548 struct writequeue_entry *e, *safe; 1549 1550 spin_lock(&con->writequeue_lock); 1551 list_for_each_entry_safe(e, safe, &con->writequeue, list) { 1552 list_del(&e->list); 1553 free_entry(e); 1554 } 1555 spin_unlock(&con->writequeue_lock); 1556 } 1557 1558 /* Called from recovery when it knows that a node has 1559 left the cluster */ 1560 int dlm_lowcomms_close(int nodeid) 1561 { 1562 struct connection *con; 1563 struct dlm_node_addr *na; 1564 1565 log_print("closing connection to node %d", nodeid); 1566 con = nodeid2con(nodeid, 0); 1567 if (con) { 1568 set_bit(CF_CLOSE, &con->flags); 1569 close_connection(con, true, true, true); 1570 clean_one_writequeue(con); 1571 } 1572 1573 spin_lock(&dlm_node_addrs_spin); 1574 na = find_node_addr(nodeid); 1575 if (na) { 1576 list_del(&na->list); 1577 while (na->addr_count--) 1578 kfree(na->addr[na->addr_count]); 1579 kfree(na); 1580 } 1581 spin_unlock(&dlm_node_addrs_spin); 1582 1583 return 0; 1584 } 1585 1586 /* Receive workqueue function */ 1587 static void process_recv_sockets(struct work_struct *work) 1588 { 1589 struct connection *con = container_of(work, struct connection, rwork); 1590 int err; 1591 1592 clear_bit(CF_READ_PENDING, &con->flags); 1593 do { 1594 err = con->rx_action(con); 1595 } while (!err); 1596 } 1597 1598 /* Send workqueue function */ 1599 static void process_send_sockets(struct work_struct *work) 1600 { 1601 struct connection *con = container_of(work, struct connection, swork); 1602 1603 clear_bit(CF_WRITE_PENDING, &con->flags); 1604 if (con->sock == NULL) /* not mutex protected so check it inside too */ 1605 con->connect_action(con); 1606 if (!list_empty(&con->writequeue)) 1607 send_to_sock(con); 1608 } 1609 1610 1611 /* Discard all entries on the write queues */ 1612 static void clean_writequeues(void) 1613 { 1614 foreach_conn(clean_one_writequeue); 1615 } 1616 1617 static void work_stop(void) 1618 { 1619 destroy_workqueue(recv_workqueue); 1620 destroy_workqueue(send_workqueue); 1621 } 1622 1623 static int work_start(void) 1624 { 1625 recv_workqueue = alloc_workqueue("dlm_recv", 1626 WQ_UNBOUND | WQ_MEM_RECLAIM, 1); 1627 if (!recv_workqueue) { 1628 log_print("can't start dlm_recv"); 1629 return -ENOMEM; 1630 } 1631 1632 send_workqueue = alloc_workqueue("dlm_send", 1633 WQ_UNBOUND | WQ_MEM_RECLAIM, 1); 1634 if (!send_workqueue) { 1635 log_print("can't start dlm_send"); 1636 destroy_workqueue(recv_workqueue); 1637 return -ENOMEM; 1638 } 1639 1640 return 0; 1641 } 1642 1643 static void _stop_conn(struct connection *con, bool and_other) 1644 { 1645 mutex_lock(&con->sock_mutex); 1646 set_bit(CF_READ_PENDING, &con->flags); 1647 set_bit(CF_WRITE_PENDING, &con->flags); 1648 if (con->sock && con->sock->sk) 1649 con->sock->sk->sk_user_data = NULL; 1650 if (con->othercon && and_other) 1651 _stop_conn(con->othercon, false); 1652 mutex_unlock(&con->sock_mutex); 1653 } 1654 1655 static void stop_conn(struct connection *con) 1656 { 1657 _stop_conn(con, true); 1658 } 1659 1660 static void free_conn(struct connection *con) 1661 { 1662 close_connection(con, true, true, true); 1663 if (con->othercon) 1664 kmem_cache_free(con_cache, con->othercon); 1665 hlist_del(&con->list); 1666 kmem_cache_free(con_cache, con); 1667 } 1668 1669 static void work_flush(void) 1670 { 1671 int ok; 1672 int i; 1673 struct hlist_node *n; 1674 struct connection *con; 1675 1676 flush_workqueue(recv_workqueue); 1677 flush_workqueue(send_workqueue); 1678 do { 1679 ok = 1; 1680 foreach_conn(stop_conn); 1681 flush_workqueue(recv_workqueue); 1682 flush_workqueue(send_workqueue); 1683 for (i = 0; i < CONN_HASH_SIZE && ok; i++) { 1684 hlist_for_each_entry_safe(con, n, 1685 &connection_hash[i], list) { 1686 ok &= test_bit(CF_READ_PENDING, &con->flags); 1687 ok &= test_bit(CF_WRITE_PENDING, &con->flags); 1688 if (con->othercon) { 1689 ok &= test_bit(CF_READ_PENDING, 1690 &con->othercon->flags); 1691 ok &= test_bit(CF_WRITE_PENDING, 1692 &con->othercon->flags); 1693 } 1694 } 1695 } 1696 } while (!ok); 1697 } 1698 1699 void dlm_lowcomms_stop(void) 1700 { 1701 /* Set all the flags to prevent any 1702 socket activity. 1703 */ 1704 mutex_lock(&connections_lock); 1705 dlm_allow_conn = 0; 1706 mutex_unlock(&connections_lock); 1707 work_flush(); 1708 clean_writequeues(); 1709 foreach_conn(free_conn); 1710 work_stop(); 1711 1712 kmem_cache_destroy(con_cache); 1713 } 1714 1715 int dlm_lowcomms_start(void) 1716 { 1717 int error = -EINVAL; 1718 struct connection *con; 1719 int i; 1720 1721 for (i = 0; i < CONN_HASH_SIZE; i++) 1722 INIT_HLIST_HEAD(&connection_hash[i]); 1723 1724 init_local(); 1725 if (!dlm_local_count) { 1726 error = -ENOTCONN; 1727 log_print("no local IP address has been set"); 1728 goto fail; 1729 } 1730 1731 error = -ENOMEM; 1732 con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection), 1733 __alignof__(struct connection), 0, 1734 NULL); 1735 if (!con_cache) 1736 goto fail; 1737 1738 error = work_start(); 1739 if (error) 1740 goto fail_destroy; 1741 1742 dlm_allow_conn = 1; 1743 1744 /* Start listening */ 1745 if (dlm_config.ci_protocol == 0) 1746 error = tcp_listen_for_all(); 1747 else 1748 error = sctp_listen_for_all(); 1749 if (error) 1750 goto fail_unlisten; 1751 1752 return 0; 1753 1754 fail_unlisten: 1755 dlm_allow_conn = 0; 1756 con = nodeid2con(0,0); 1757 if (con) { 1758 close_connection(con, false, true, true); 1759 kmem_cache_free(con_cache, con); 1760 } 1761 fail_destroy: 1762 kmem_cache_destroy(con_cache); 1763 fail: 1764 return error; 1765 } 1766 1767 void dlm_lowcomms_exit(void) 1768 { 1769 struct dlm_node_addr *na, *safe; 1770 1771 spin_lock(&dlm_node_addrs_spin); 1772 list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) { 1773 list_del(&na->list); 1774 while (na->addr_count--) 1775 kfree(na->addr[na->addr_count]); 1776 kfree(na); 1777 } 1778 spin_unlock(&dlm_node_addrs_spin); 1779 } 1780