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_CONNECT_PENDING 3 111 #define CF_INIT_PENDING 4 112 #define CF_IS_OTHERCON 5 113 #define CF_CLOSE 6 114 #define CF_APP_LIMITED 7 115 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 void (*orig_error_report)(struct sock *); 128 void (*orig_data_ready)(struct sock *); 129 void (*orig_state_change)(struct sock *); 130 void (*orig_write_space)(struct sock *); 131 }; 132 #define sock2con(x) ((struct connection *)(x)->sk_user_data) 133 134 /* An entry waiting to be sent */ 135 struct writequeue_entry { 136 struct list_head list; 137 struct page *page; 138 int offset; 139 int len; 140 int end; 141 int users; 142 struct connection *con; 143 }; 144 145 struct dlm_node_addr { 146 struct list_head list; 147 int nodeid; 148 int addr_count; 149 int curr_addr_index; 150 struct sockaddr_storage *addr[DLM_MAX_ADDR_COUNT]; 151 }; 152 153 static LIST_HEAD(dlm_node_addrs); 154 static DEFINE_SPINLOCK(dlm_node_addrs_spin); 155 156 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT]; 157 static int dlm_local_count; 158 static int dlm_allow_conn; 159 160 /* Work queues */ 161 static struct workqueue_struct *recv_workqueue; 162 static struct workqueue_struct *send_workqueue; 163 164 static struct hlist_head connection_hash[CONN_HASH_SIZE]; 165 static DEFINE_MUTEX(connections_lock); 166 static struct kmem_cache *con_cache; 167 168 static void process_recv_sockets(struct work_struct *work); 169 static void process_send_sockets(struct work_struct *work); 170 171 172 /* This is deliberately very simple because most clusters have simple 173 sequential nodeids, so we should be able to go straight to a connection 174 struct in the array */ 175 static inline int nodeid_hash(int nodeid) 176 { 177 return nodeid & (CONN_HASH_SIZE-1); 178 } 179 180 static struct connection *__find_con(int nodeid) 181 { 182 int r; 183 struct connection *con; 184 185 r = nodeid_hash(nodeid); 186 187 hlist_for_each_entry(con, &connection_hash[r], list) { 188 if (con->nodeid == nodeid) 189 return con; 190 } 191 return NULL; 192 } 193 194 /* 195 * If 'allocation' is zero then we don't attempt to create a new 196 * connection structure for this node. 197 */ 198 static struct connection *__nodeid2con(int nodeid, gfp_t alloc) 199 { 200 struct connection *con = NULL; 201 int r; 202 203 con = __find_con(nodeid); 204 if (con || !alloc) 205 return con; 206 207 con = kmem_cache_zalloc(con_cache, alloc); 208 if (!con) 209 return NULL; 210 211 r = nodeid_hash(nodeid); 212 hlist_add_head(&con->list, &connection_hash[r]); 213 214 con->nodeid = nodeid; 215 mutex_init(&con->sock_mutex); 216 INIT_LIST_HEAD(&con->writequeue); 217 spin_lock_init(&con->writequeue_lock); 218 INIT_WORK(&con->swork, process_send_sockets); 219 INIT_WORK(&con->rwork, process_recv_sockets); 220 221 /* Setup action pointers for child sockets */ 222 if (con->nodeid) { 223 struct connection *zerocon = __find_con(0); 224 225 con->connect_action = zerocon->connect_action; 226 if (!con->rx_action) 227 con->rx_action = zerocon->rx_action; 228 } 229 230 return con; 231 } 232 233 /* Loop round all connections */ 234 static void foreach_conn(void (*conn_func)(struct connection *c)) 235 { 236 int i; 237 struct hlist_node *n; 238 struct connection *con; 239 240 for (i = 0; i < CONN_HASH_SIZE; i++) { 241 hlist_for_each_entry_safe(con, n, &connection_hash[i], list) 242 conn_func(con); 243 } 244 } 245 246 static struct connection *nodeid2con(int nodeid, gfp_t allocation) 247 { 248 struct connection *con; 249 250 mutex_lock(&connections_lock); 251 con = __nodeid2con(nodeid, allocation); 252 mutex_unlock(&connections_lock); 253 254 return con; 255 } 256 257 static struct dlm_node_addr *find_node_addr(int nodeid) 258 { 259 struct dlm_node_addr *na; 260 261 list_for_each_entry(na, &dlm_node_addrs, list) { 262 if (na->nodeid == nodeid) 263 return na; 264 } 265 return NULL; 266 } 267 268 static int addr_compare(struct sockaddr_storage *x, struct sockaddr_storage *y) 269 { 270 switch (x->ss_family) { 271 case AF_INET: { 272 struct sockaddr_in *sinx = (struct sockaddr_in *)x; 273 struct sockaddr_in *siny = (struct sockaddr_in *)y; 274 if (sinx->sin_addr.s_addr != siny->sin_addr.s_addr) 275 return 0; 276 if (sinx->sin_port != siny->sin_port) 277 return 0; 278 break; 279 } 280 case AF_INET6: { 281 struct sockaddr_in6 *sinx = (struct sockaddr_in6 *)x; 282 struct sockaddr_in6 *siny = (struct sockaddr_in6 *)y; 283 if (!ipv6_addr_equal(&sinx->sin6_addr, &siny->sin6_addr)) 284 return 0; 285 if (sinx->sin6_port != siny->sin6_port) 286 return 0; 287 break; 288 } 289 default: 290 return 0; 291 } 292 return 1; 293 } 294 295 static int nodeid_to_addr(int nodeid, struct sockaddr_storage *sas_out, 296 struct sockaddr *sa_out, bool try_new_addr) 297 { 298 struct sockaddr_storage sas; 299 struct dlm_node_addr *na; 300 301 if (!dlm_local_count) 302 return -1; 303 304 spin_lock(&dlm_node_addrs_spin); 305 na = find_node_addr(nodeid); 306 if (na && na->addr_count) { 307 memcpy(&sas, na->addr[na->curr_addr_index], 308 sizeof(struct sockaddr_storage)); 309 310 if (try_new_addr) { 311 na->curr_addr_index++; 312 if (na->curr_addr_index == na->addr_count) 313 na->curr_addr_index = 0; 314 } 315 } 316 spin_unlock(&dlm_node_addrs_spin); 317 318 if (!na) 319 return -EEXIST; 320 321 if (!na->addr_count) 322 return -ENOENT; 323 324 if (sas_out) 325 memcpy(sas_out, &sas, sizeof(struct sockaddr_storage)); 326 327 if (!sa_out) 328 return 0; 329 330 if (dlm_local_addr[0]->ss_family == AF_INET) { 331 struct sockaddr_in *in4 = (struct sockaddr_in *) &sas; 332 struct sockaddr_in *ret4 = (struct sockaddr_in *) sa_out; 333 ret4->sin_addr.s_addr = in4->sin_addr.s_addr; 334 } else { 335 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) &sas; 336 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) sa_out; 337 ret6->sin6_addr = in6->sin6_addr; 338 } 339 340 return 0; 341 } 342 343 static int addr_to_nodeid(struct sockaddr_storage *addr, int *nodeid) 344 { 345 struct dlm_node_addr *na; 346 int rv = -EEXIST; 347 int addr_i; 348 349 spin_lock(&dlm_node_addrs_spin); 350 list_for_each_entry(na, &dlm_node_addrs, list) { 351 if (!na->addr_count) 352 continue; 353 354 for (addr_i = 0; addr_i < na->addr_count; addr_i++) { 355 if (addr_compare(na->addr[addr_i], addr)) { 356 *nodeid = na->nodeid; 357 rv = 0; 358 goto unlock; 359 } 360 } 361 } 362 unlock: 363 spin_unlock(&dlm_node_addrs_spin); 364 return rv; 365 } 366 367 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len) 368 { 369 struct sockaddr_storage *new_addr; 370 struct dlm_node_addr *new_node, *na; 371 372 new_node = kzalloc(sizeof(struct dlm_node_addr), GFP_NOFS); 373 if (!new_node) 374 return -ENOMEM; 375 376 new_addr = kzalloc(sizeof(struct sockaddr_storage), GFP_NOFS); 377 if (!new_addr) { 378 kfree(new_node); 379 return -ENOMEM; 380 } 381 382 memcpy(new_addr, addr, len); 383 384 spin_lock(&dlm_node_addrs_spin); 385 na = find_node_addr(nodeid); 386 if (!na) { 387 new_node->nodeid = nodeid; 388 new_node->addr[0] = new_addr; 389 new_node->addr_count = 1; 390 list_add(&new_node->list, &dlm_node_addrs); 391 spin_unlock(&dlm_node_addrs_spin); 392 return 0; 393 } 394 395 if (na->addr_count >= DLM_MAX_ADDR_COUNT) { 396 spin_unlock(&dlm_node_addrs_spin); 397 kfree(new_addr); 398 kfree(new_node); 399 return -ENOSPC; 400 } 401 402 na->addr[na->addr_count++] = new_addr; 403 spin_unlock(&dlm_node_addrs_spin); 404 kfree(new_node); 405 return 0; 406 } 407 408 /* Data available on socket or listen socket received a connect */ 409 static void lowcomms_data_ready(struct sock *sk) 410 { 411 struct connection *con = sock2con(sk); 412 if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags)) 413 queue_work(recv_workqueue, &con->rwork); 414 } 415 416 static void lowcomms_write_space(struct sock *sk) 417 { 418 struct connection *con = sock2con(sk); 419 420 if (!con) 421 return; 422 423 clear_bit(SOCK_NOSPACE, &con->sock->flags); 424 425 if (test_and_clear_bit(CF_APP_LIMITED, &con->flags)) { 426 con->sock->sk->sk_write_pending--; 427 clear_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags); 428 } 429 430 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) 431 queue_work(send_workqueue, &con->swork); 432 } 433 434 static inline void lowcomms_connect_sock(struct connection *con) 435 { 436 if (test_bit(CF_CLOSE, &con->flags)) 437 return; 438 if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags)) 439 queue_work(send_workqueue, &con->swork); 440 } 441 442 static void lowcomms_state_change(struct sock *sk) 443 { 444 /* SCTP layer is not calling sk_data_ready when the connection 445 * is done, so we catch the signal through here. Also, it 446 * doesn't switch socket state when entering shutdown, so we 447 * skip the write in that case. 448 */ 449 if (sk->sk_shutdown) { 450 if (sk->sk_shutdown == RCV_SHUTDOWN) 451 lowcomms_data_ready(sk); 452 } else if (sk->sk_state == TCP_ESTABLISHED) { 453 lowcomms_write_space(sk); 454 } 455 } 456 457 int dlm_lowcomms_connect_node(int nodeid) 458 { 459 struct connection *con; 460 461 if (nodeid == dlm_our_nodeid()) 462 return 0; 463 464 con = nodeid2con(nodeid, GFP_NOFS); 465 if (!con) 466 return -ENOMEM; 467 lowcomms_connect_sock(con); 468 return 0; 469 } 470 471 static void lowcomms_error_report(struct sock *sk) 472 { 473 struct connection *con; 474 struct sockaddr_storage saddr; 475 int buflen; 476 void (*orig_report)(struct sock *) = NULL; 477 478 read_lock_bh(&sk->sk_callback_lock); 479 con = sock2con(sk); 480 if (con == NULL) 481 goto out; 482 483 orig_report = con->orig_error_report; 484 if (con->sock == NULL || 485 kernel_getpeername(con->sock, (struct sockaddr *)&saddr, &buflen)) { 486 printk_ratelimited(KERN_ERR "dlm: node %d: socket error " 487 "sending to node %d, port %d, " 488 "sk_err=%d/%d\n", dlm_our_nodeid(), 489 con->nodeid, dlm_config.ci_tcp_port, 490 sk->sk_err, sk->sk_err_soft); 491 } else if (saddr.ss_family == AF_INET) { 492 struct sockaddr_in *sin4 = (struct sockaddr_in *)&saddr; 493 494 printk_ratelimited(KERN_ERR "dlm: node %d: socket error " 495 "sending to node %d at %pI4, port %d, " 496 "sk_err=%d/%d\n", dlm_our_nodeid(), 497 con->nodeid, &sin4->sin_addr.s_addr, 498 dlm_config.ci_tcp_port, sk->sk_err, 499 sk->sk_err_soft); 500 } else { 501 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)&saddr; 502 503 printk_ratelimited(KERN_ERR "dlm: node %d: socket error " 504 "sending to node %d at %u.%u.%u.%u, " 505 "port %d, sk_err=%d/%d\n", dlm_our_nodeid(), 506 con->nodeid, sin6->sin6_addr.s6_addr32[0], 507 sin6->sin6_addr.s6_addr32[1], 508 sin6->sin6_addr.s6_addr32[2], 509 sin6->sin6_addr.s6_addr32[3], 510 dlm_config.ci_tcp_port, sk->sk_err, 511 sk->sk_err_soft); 512 } 513 out: 514 read_unlock_bh(&sk->sk_callback_lock); 515 if (orig_report) 516 orig_report(sk); 517 } 518 519 /* Note: sk_callback_lock must be locked before calling this function. */ 520 static void save_callbacks(struct connection *con, struct sock *sk) 521 { 522 con->orig_data_ready = sk->sk_data_ready; 523 con->orig_state_change = sk->sk_state_change; 524 con->orig_write_space = sk->sk_write_space; 525 con->orig_error_report = sk->sk_error_report; 526 } 527 528 static void restore_callbacks(struct connection *con, struct sock *sk) 529 { 530 write_lock_bh(&sk->sk_callback_lock); 531 sk->sk_user_data = NULL; 532 sk->sk_data_ready = con->orig_data_ready; 533 sk->sk_state_change = con->orig_state_change; 534 sk->sk_write_space = con->orig_write_space; 535 sk->sk_error_report = con->orig_error_report; 536 write_unlock_bh(&sk->sk_callback_lock); 537 } 538 539 /* Make a socket active */ 540 static void add_sock(struct socket *sock, struct connection *con, bool save_cb) 541 { 542 struct sock *sk = sock->sk; 543 544 write_lock_bh(&sk->sk_callback_lock); 545 con->sock = sock; 546 547 sk->sk_user_data = con; 548 if (save_cb) 549 save_callbacks(con, sk); 550 /* Install a data_ready callback */ 551 sk->sk_data_ready = lowcomms_data_ready; 552 sk->sk_write_space = lowcomms_write_space; 553 sk->sk_state_change = lowcomms_state_change; 554 sk->sk_allocation = GFP_NOFS; 555 sk->sk_error_report = lowcomms_error_report; 556 write_unlock_bh(&sk->sk_callback_lock); 557 } 558 559 /* Add the port number to an IPv6 or 4 sockaddr and return the address 560 length */ 561 static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port, 562 int *addr_len) 563 { 564 saddr->ss_family = dlm_local_addr[0]->ss_family; 565 if (saddr->ss_family == AF_INET) { 566 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr; 567 in4_addr->sin_port = cpu_to_be16(port); 568 *addr_len = sizeof(struct sockaddr_in); 569 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero)); 570 } else { 571 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr; 572 in6_addr->sin6_port = cpu_to_be16(port); 573 *addr_len = sizeof(struct sockaddr_in6); 574 } 575 memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len); 576 } 577 578 /* Close a remote connection and tidy up */ 579 static void close_connection(struct connection *con, bool and_other, 580 bool tx, bool rx) 581 { 582 clear_bit(CF_CONNECT_PENDING, &con->flags); 583 clear_bit(CF_WRITE_PENDING, &con->flags); 584 if (tx && cancel_work_sync(&con->swork)) 585 log_print("canceled swork for node %d", con->nodeid); 586 if (rx && cancel_work_sync(&con->rwork)) 587 log_print("canceled rwork for node %d", con->nodeid); 588 589 mutex_lock(&con->sock_mutex); 590 if (con->sock) { 591 if (!test_bit(CF_IS_OTHERCON, &con->flags)) 592 restore_callbacks(con, con->sock->sk); 593 sock_release(con->sock); 594 con->sock = NULL; 595 } 596 if (con->othercon && and_other) { 597 /* Will only re-enter once. */ 598 close_connection(con->othercon, false, true, true); 599 } 600 if (con->rx_page) { 601 __free_page(con->rx_page); 602 con->rx_page = NULL; 603 } 604 605 con->retries = 0; 606 mutex_unlock(&con->sock_mutex); 607 } 608 609 /* Data received from remote end */ 610 static int receive_from_sock(struct connection *con) 611 { 612 int ret = 0; 613 struct msghdr msg = {}; 614 struct kvec iov[2]; 615 unsigned len; 616 int r; 617 int call_again_soon = 0; 618 int nvec; 619 620 mutex_lock(&con->sock_mutex); 621 622 if (con->sock == NULL) { 623 ret = -EAGAIN; 624 goto out_close; 625 } 626 if (con->nodeid == 0) { 627 ret = -EINVAL; 628 goto out_close; 629 } 630 631 if (con->rx_page == NULL) { 632 /* 633 * This doesn't need to be atomic, but I think it should 634 * improve performance if it is. 635 */ 636 con->rx_page = alloc_page(GFP_ATOMIC); 637 if (con->rx_page == NULL) 638 goto out_resched; 639 cbuf_init(&con->cb, PAGE_SIZE); 640 } 641 642 /* 643 * iov[0] is the bit of the circular buffer between the current end 644 * point (cb.base + cb.len) and the end of the buffer. 645 */ 646 iov[0].iov_len = con->cb.base - cbuf_data(&con->cb); 647 iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb); 648 iov[1].iov_len = 0; 649 nvec = 1; 650 651 /* 652 * iov[1] is the bit of the circular buffer between the start of the 653 * buffer and the start of the currently used section (cb.base) 654 */ 655 if (cbuf_data(&con->cb) >= con->cb.base) { 656 iov[0].iov_len = PAGE_SIZE - cbuf_data(&con->cb); 657 iov[1].iov_len = con->cb.base; 658 iov[1].iov_base = page_address(con->rx_page); 659 nvec = 2; 660 } 661 len = iov[0].iov_len + iov[1].iov_len; 662 663 r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len, 664 MSG_DONTWAIT | MSG_NOSIGNAL); 665 if (ret <= 0) 666 goto out_close; 667 else if (ret == len) 668 call_again_soon = 1; 669 670 cbuf_add(&con->cb, ret); 671 ret = dlm_process_incoming_buffer(con->nodeid, 672 page_address(con->rx_page), 673 con->cb.base, con->cb.len, 674 PAGE_SIZE); 675 if (ret == -EBADMSG) { 676 log_print("lowcomms: addr=%p, base=%u, len=%u, read=%d", 677 page_address(con->rx_page), con->cb.base, 678 con->cb.len, r); 679 } 680 if (ret < 0) 681 goto out_close; 682 cbuf_eat(&con->cb, ret); 683 684 if (cbuf_empty(&con->cb) && !call_again_soon) { 685 __free_page(con->rx_page); 686 con->rx_page = NULL; 687 } 688 689 if (call_again_soon) 690 goto out_resched; 691 mutex_unlock(&con->sock_mutex); 692 return 0; 693 694 out_resched: 695 if (!test_and_set_bit(CF_READ_PENDING, &con->flags)) 696 queue_work(recv_workqueue, &con->rwork); 697 mutex_unlock(&con->sock_mutex); 698 return -EAGAIN; 699 700 out_close: 701 mutex_unlock(&con->sock_mutex); 702 if (ret != -EAGAIN) { 703 close_connection(con, false, true, false); 704 /* Reconnect when there is something to send */ 705 } 706 /* Don't return success if we really got EOF */ 707 if (ret == 0) 708 ret = -EAGAIN; 709 710 return ret; 711 } 712 713 /* Listening socket is busy, accept a connection */ 714 static int tcp_accept_from_sock(struct connection *con) 715 { 716 int result; 717 struct sockaddr_storage peeraddr; 718 struct socket *newsock; 719 int len; 720 int nodeid; 721 struct connection *newcon; 722 struct connection *addcon; 723 724 mutex_lock(&connections_lock); 725 if (!dlm_allow_conn) { 726 mutex_unlock(&connections_lock); 727 return -1; 728 } 729 mutex_unlock(&connections_lock); 730 731 memset(&peeraddr, 0, sizeof(peeraddr)); 732 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 733 SOCK_STREAM, IPPROTO_TCP, &newsock); 734 if (result < 0) 735 return -ENOMEM; 736 737 mutex_lock_nested(&con->sock_mutex, 0); 738 739 result = -ENOTCONN; 740 if (con->sock == NULL) 741 goto accept_err; 742 743 newsock->type = con->sock->type; 744 newsock->ops = con->sock->ops; 745 746 result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK); 747 if (result < 0) 748 goto accept_err; 749 750 /* Get the connected socket's peer */ 751 memset(&peeraddr, 0, sizeof(peeraddr)); 752 if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr, 753 &len, 2)) { 754 result = -ECONNABORTED; 755 goto accept_err; 756 } 757 758 /* Get the new node's NODEID */ 759 make_sockaddr(&peeraddr, 0, &len); 760 if (addr_to_nodeid(&peeraddr, &nodeid)) { 761 unsigned char *b=(unsigned char *)&peeraddr; 762 log_print("connect from non cluster node"); 763 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE, 764 b, sizeof(struct sockaddr_storage)); 765 sock_release(newsock); 766 mutex_unlock(&con->sock_mutex); 767 return -1; 768 } 769 770 log_print("got connection from %d", nodeid); 771 772 /* Check to see if we already have a connection to this node. This 773 * could happen if the two nodes initiate a connection at roughly 774 * the same time and the connections cross on the wire. 775 * In this case we store the incoming one in "othercon" 776 */ 777 newcon = nodeid2con(nodeid, GFP_NOFS); 778 if (!newcon) { 779 result = -ENOMEM; 780 goto accept_err; 781 } 782 mutex_lock_nested(&newcon->sock_mutex, 1); 783 if (newcon->sock) { 784 struct connection *othercon = newcon->othercon; 785 786 if (!othercon) { 787 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS); 788 if (!othercon) { 789 log_print("failed to allocate incoming socket"); 790 mutex_unlock(&newcon->sock_mutex); 791 result = -ENOMEM; 792 goto accept_err; 793 } 794 othercon->nodeid = nodeid; 795 othercon->rx_action = receive_from_sock; 796 mutex_init(&othercon->sock_mutex); 797 INIT_WORK(&othercon->swork, process_send_sockets); 798 INIT_WORK(&othercon->rwork, process_recv_sockets); 799 set_bit(CF_IS_OTHERCON, &othercon->flags); 800 } 801 if (!othercon->sock) { 802 newcon->othercon = othercon; 803 othercon->sock = newsock; 804 newsock->sk->sk_user_data = othercon; 805 add_sock(newsock, othercon, false); 806 addcon = othercon; 807 } 808 else { 809 printk("Extra connection from node %d attempted\n", nodeid); 810 result = -EAGAIN; 811 mutex_unlock(&newcon->sock_mutex); 812 goto accept_err; 813 } 814 } 815 else { 816 newsock->sk->sk_user_data = newcon; 817 newcon->rx_action = receive_from_sock; 818 /* accept copies the sk after we've saved the callbacks, so we 819 don't want to save them a second time or comm errors will 820 result in calling sk_error_report recursively. */ 821 add_sock(newsock, newcon, false); 822 addcon = newcon; 823 } 824 825 mutex_unlock(&newcon->sock_mutex); 826 827 /* 828 * Add it to the active queue in case we got data 829 * between processing the accept adding the socket 830 * to the read_sockets list 831 */ 832 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags)) 833 queue_work(recv_workqueue, &addcon->rwork); 834 mutex_unlock(&con->sock_mutex); 835 836 return 0; 837 838 accept_err: 839 mutex_unlock(&con->sock_mutex); 840 sock_release(newsock); 841 842 if (result != -EAGAIN) 843 log_print("error accepting connection from node: %d", result); 844 return result; 845 } 846 847 static int sctp_accept_from_sock(struct connection *con) 848 { 849 /* Check that the new node is in the lockspace */ 850 struct sctp_prim prim; 851 int nodeid; 852 int prim_len, ret; 853 int addr_len; 854 struct connection *newcon; 855 struct connection *addcon; 856 struct socket *newsock; 857 858 mutex_lock(&connections_lock); 859 if (!dlm_allow_conn) { 860 mutex_unlock(&connections_lock); 861 return -1; 862 } 863 mutex_unlock(&connections_lock); 864 865 mutex_lock_nested(&con->sock_mutex, 0); 866 867 ret = kernel_accept(con->sock, &newsock, O_NONBLOCK); 868 if (ret < 0) 869 goto accept_err; 870 871 memset(&prim, 0, sizeof(struct sctp_prim)); 872 prim_len = sizeof(struct sctp_prim); 873 874 ret = kernel_getsockopt(newsock, IPPROTO_SCTP, SCTP_PRIMARY_ADDR, 875 (char *)&prim, &prim_len); 876 if (ret < 0) { 877 log_print("getsockopt/sctp_primary_addr failed: %d", ret); 878 goto accept_err; 879 } 880 881 make_sockaddr(&prim.ssp_addr, 0, &addr_len); 882 if (addr_to_nodeid(&prim.ssp_addr, &nodeid)) { 883 unsigned char *b = (unsigned char *)&prim.ssp_addr; 884 885 log_print("reject connect from unknown addr"); 886 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE, 887 b, sizeof(struct sockaddr_storage)); 888 goto accept_err; 889 } 890 891 newcon = nodeid2con(nodeid, GFP_NOFS); 892 if (!newcon) { 893 ret = -ENOMEM; 894 goto accept_err; 895 } 896 897 mutex_lock_nested(&newcon->sock_mutex, 1); 898 899 if (newcon->sock) { 900 struct connection *othercon = newcon->othercon; 901 902 if (!othercon) { 903 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS); 904 if (!othercon) { 905 log_print("failed to allocate incoming socket"); 906 mutex_unlock(&newcon->sock_mutex); 907 ret = -ENOMEM; 908 goto accept_err; 909 } 910 othercon->nodeid = nodeid; 911 othercon->rx_action = receive_from_sock; 912 mutex_init(&othercon->sock_mutex); 913 INIT_WORK(&othercon->swork, process_send_sockets); 914 INIT_WORK(&othercon->rwork, process_recv_sockets); 915 set_bit(CF_IS_OTHERCON, &othercon->flags); 916 } 917 if (!othercon->sock) { 918 newcon->othercon = othercon; 919 othercon->sock = newsock; 920 newsock->sk->sk_user_data = othercon; 921 add_sock(newsock, othercon, false); 922 addcon = othercon; 923 } else { 924 printk("Extra connection from node %d attempted\n", nodeid); 925 ret = -EAGAIN; 926 mutex_unlock(&newcon->sock_mutex); 927 goto accept_err; 928 } 929 } else { 930 newsock->sk->sk_user_data = newcon; 931 newcon->rx_action = receive_from_sock; 932 add_sock(newsock, newcon, false); 933 addcon = newcon; 934 } 935 936 log_print("connected to %d", nodeid); 937 938 mutex_unlock(&newcon->sock_mutex); 939 940 /* 941 * Add it to the active queue in case we got data 942 * between processing the accept adding the socket 943 * to the read_sockets list 944 */ 945 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags)) 946 queue_work(recv_workqueue, &addcon->rwork); 947 mutex_unlock(&con->sock_mutex); 948 949 return 0; 950 951 accept_err: 952 mutex_unlock(&con->sock_mutex); 953 if (newsock) 954 sock_release(newsock); 955 if (ret != -EAGAIN) 956 log_print("error accepting connection from node: %d", ret); 957 958 return ret; 959 } 960 961 static void free_entry(struct writequeue_entry *e) 962 { 963 __free_page(e->page); 964 kfree(e); 965 } 966 967 /* 968 * writequeue_entry_complete - try to delete and free write queue entry 969 * @e: write queue entry to try to delete 970 * @completed: bytes completed 971 * 972 * writequeue_lock must be held. 973 */ 974 static void writequeue_entry_complete(struct writequeue_entry *e, int completed) 975 { 976 e->offset += completed; 977 e->len -= completed; 978 979 if (e->len == 0 && e->users == 0) { 980 list_del(&e->list); 981 free_entry(e); 982 } 983 } 984 985 /* 986 * sctp_bind_addrs - bind a SCTP socket to all our addresses 987 */ 988 static int sctp_bind_addrs(struct connection *con, uint16_t port) 989 { 990 struct sockaddr_storage localaddr; 991 int i, addr_len, result = 0; 992 993 for (i = 0; i < dlm_local_count; i++) { 994 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr)); 995 make_sockaddr(&localaddr, port, &addr_len); 996 997 if (!i) 998 result = kernel_bind(con->sock, 999 (struct sockaddr *)&localaddr, 1000 addr_len); 1001 else 1002 result = kernel_setsockopt(con->sock, SOL_SCTP, 1003 SCTP_SOCKOPT_BINDX_ADD, 1004 (char *)&localaddr, addr_len); 1005 1006 if (result < 0) { 1007 log_print("Can't bind to %d addr number %d, %d.\n", 1008 port, i + 1, result); 1009 break; 1010 } 1011 } 1012 return result; 1013 } 1014 1015 /* Initiate an SCTP association. 1016 This is a special case of send_to_sock() in that we don't yet have a 1017 peeled-off socket for this association, so we use the listening socket 1018 and add the primary IP address of the remote node. 1019 */ 1020 static void sctp_connect_to_sock(struct connection *con) 1021 { 1022 struct sockaddr_storage daddr; 1023 int one = 1; 1024 int result; 1025 int addr_len; 1026 struct socket *sock; 1027 1028 if (con->nodeid == 0) { 1029 log_print("attempt to connect sock 0 foiled"); 1030 return; 1031 } 1032 1033 mutex_lock(&con->sock_mutex); 1034 1035 /* Some odd races can cause double-connects, ignore them */ 1036 if (con->retries++ > MAX_CONNECT_RETRIES) 1037 goto out; 1038 1039 if (con->sock) { 1040 log_print("node %d already connected.", con->nodeid); 1041 goto out; 1042 } 1043 1044 memset(&daddr, 0, sizeof(daddr)); 1045 result = nodeid_to_addr(con->nodeid, &daddr, NULL, true); 1046 if (result < 0) { 1047 log_print("no address for nodeid %d", con->nodeid); 1048 goto out; 1049 } 1050 1051 /* Create a socket to communicate with */ 1052 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 1053 SOCK_STREAM, IPPROTO_SCTP, &sock); 1054 if (result < 0) 1055 goto socket_err; 1056 1057 sock->sk->sk_user_data = con; 1058 con->rx_action = receive_from_sock; 1059 con->connect_action = sctp_connect_to_sock; 1060 add_sock(sock, con, true); 1061 1062 /* Bind to all addresses. */ 1063 if (sctp_bind_addrs(con, 0)) 1064 goto bind_err; 1065 1066 make_sockaddr(&daddr, dlm_config.ci_tcp_port, &addr_len); 1067 1068 log_print("connecting to %d", con->nodeid); 1069 1070 /* Turn off Nagle's algorithm */ 1071 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one, 1072 sizeof(one)); 1073 1074 result = sock->ops->connect(sock, (struct sockaddr *)&daddr, addr_len, 1075 O_NONBLOCK); 1076 if (result == -EINPROGRESS) 1077 result = 0; 1078 if (result == 0) 1079 goto out; 1080 1081 1082 bind_err: 1083 con->sock = NULL; 1084 sock_release(sock); 1085 1086 socket_err: 1087 /* 1088 * Some errors are fatal and this list might need adjusting. For other 1089 * errors we try again until the max number of retries is reached. 1090 */ 1091 if (result != -EHOSTUNREACH && 1092 result != -ENETUNREACH && 1093 result != -ENETDOWN && 1094 result != -EINVAL && 1095 result != -EPROTONOSUPPORT) { 1096 log_print("connect %d try %d error %d", con->nodeid, 1097 con->retries, result); 1098 mutex_unlock(&con->sock_mutex); 1099 msleep(1000); 1100 clear_bit(CF_CONNECT_PENDING, &con->flags); 1101 lowcomms_connect_sock(con); 1102 return; 1103 } 1104 1105 out: 1106 mutex_unlock(&con->sock_mutex); 1107 set_bit(CF_WRITE_PENDING, &con->flags); 1108 } 1109 1110 /* Connect a new socket to its peer */ 1111 static void tcp_connect_to_sock(struct connection *con) 1112 { 1113 struct sockaddr_storage saddr, src_addr; 1114 int addr_len; 1115 struct socket *sock = NULL; 1116 int one = 1; 1117 int result; 1118 1119 if (con->nodeid == 0) { 1120 log_print("attempt to connect sock 0 foiled"); 1121 return; 1122 } 1123 1124 mutex_lock(&con->sock_mutex); 1125 if (con->retries++ > MAX_CONNECT_RETRIES) 1126 goto out; 1127 1128 /* Some odd races can cause double-connects, ignore them */ 1129 if (con->sock) 1130 goto out; 1131 1132 /* Create a socket to communicate with */ 1133 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 1134 SOCK_STREAM, IPPROTO_TCP, &sock); 1135 if (result < 0) 1136 goto out_err; 1137 1138 memset(&saddr, 0, sizeof(saddr)); 1139 result = nodeid_to_addr(con->nodeid, &saddr, NULL, false); 1140 if (result < 0) { 1141 log_print("no address for nodeid %d", con->nodeid); 1142 goto out_err; 1143 } 1144 1145 sock->sk->sk_user_data = con; 1146 con->rx_action = receive_from_sock; 1147 con->connect_action = tcp_connect_to_sock; 1148 add_sock(sock, con, true); 1149 1150 /* Bind to our cluster-known address connecting to avoid 1151 routing problems */ 1152 memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr)); 1153 make_sockaddr(&src_addr, 0, &addr_len); 1154 result = sock->ops->bind(sock, (struct sockaddr *) &src_addr, 1155 addr_len); 1156 if (result < 0) { 1157 log_print("could not bind for connect: %d", result); 1158 /* This *may* not indicate a critical error */ 1159 } 1160 1161 make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len); 1162 1163 log_print("connecting to %d", con->nodeid); 1164 1165 /* Turn off Nagle's algorithm */ 1166 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one, 1167 sizeof(one)); 1168 1169 result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len, 1170 O_NONBLOCK); 1171 if (result == -EINPROGRESS) 1172 result = 0; 1173 if (result == 0) 1174 goto out; 1175 1176 out_err: 1177 if (con->sock) { 1178 sock_release(con->sock); 1179 con->sock = NULL; 1180 } else if (sock) { 1181 sock_release(sock); 1182 } 1183 /* 1184 * Some errors are fatal and this list might need adjusting. For other 1185 * errors we try again until the max number of retries is reached. 1186 */ 1187 if (result != -EHOSTUNREACH && 1188 result != -ENETUNREACH && 1189 result != -ENETDOWN && 1190 result != -EINVAL && 1191 result != -EPROTONOSUPPORT) { 1192 log_print("connect %d try %d error %d", con->nodeid, 1193 con->retries, result); 1194 mutex_unlock(&con->sock_mutex); 1195 msleep(1000); 1196 clear_bit(CF_CONNECT_PENDING, &con->flags); 1197 lowcomms_connect_sock(con); 1198 return; 1199 } 1200 out: 1201 mutex_unlock(&con->sock_mutex); 1202 set_bit(CF_WRITE_PENDING, &con->flags); 1203 return; 1204 } 1205 1206 static struct socket *tcp_create_listen_sock(struct connection *con, 1207 struct sockaddr_storage *saddr) 1208 { 1209 struct socket *sock = NULL; 1210 int result = 0; 1211 int one = 1; 1212 int addr_len; 1213 1214 if (dlm_local_addr[0]->ss_family == AF_INET) 1215 addr_len = sizeof(struct sockaddr_in); 1216 else 1217 addr_len = sizeof(struct sockaddr_in6); 1218 1219 /* Create a socket to communicate with */ 1220 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 1221 SOCK_STREAM, IPPROTO_TCP, &sock); 1222 if (result < 0) { 1223 log_print("Can't create listening comms socket"); 1224 goto create_out; 1225 } 1226 1227 /* Turn off Nagle's algorithm */ 1228 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one, 1229 sizeof(one)); 1230 1231 result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, 1232 (char *)&one, sizeof(one)); 1233 1234 if (result < 0) { 1235 log_print("Failed to set SO_REUSEADDR on socket: %d", result); 1236 } 1237 sock->sk->sk_user_data = con; 1238 1239 con->rx_action = tcp_accept_from_sock; 1240 con->connect_action = tcp_connect_to_sock; 1241 1242 /* Bind to our port */ 1243 make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len); 1244 result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len); 1245 if (result < 0) { 1246 log_print("Can't bind to port %d", dlm_config.ci_tcp_port); 1247 sock_release(sock); 1248 sock = NULL; 1249 con->sock = NULL; 1250 goto create_out; 1251 } 1252 result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE, 1253 (char *)&one, sizeof(one)); 1254 if (result < 0) { 1255 log_print("Set keepalive failed: %d", result); 1256 } 1257 1258 result = sock->ops->listen(sock, 5); 1259 if (result < 0) { 1260 log_print("Can't listen on port %d", dlm_config.ci_tcp_port); 1261 sock_release(sock); 1262 sock = NULL; 1263 goto create_out; 1264 } 1265 1266 create_out: 1267 return sock; 1268 } 1269 1270 /* Get local addresses */ 1271 static void init_local(void) 1272 { 1273 struct sockaddr_storage sas, *addr; 1274 int i; 1275 1276 dlm_local_count = 0; 1277 for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) { 1278 if (dlm_our_addr(&sas, i)) 1279 break; 1280 1281 addr = kmemdup(&sas, sizeof(*addr), GFP_NOFS); 1282 if (!addr) 1283 break; 1284 dlm_local_addr[dlm_local_count++] = addr; 1285 } 1286 } 1287 1288 /* Initialise SCTP socket and bind to all interfaces */ 1289 static int sctp_listen_for_all(void) 1290 { 1291 struct socket *sock = NULL; 1292 int result = -EINVAL; 1293 struct connection *con = nodeid2con(0, GFP_NOFS); 1294 int bufsize = NEEDED_RMEM; 1295 int one = 1; 1296 1297 if (!con) 1298 return -ENOMEM; 1299 1300 log_print("Using SCTP for communications"); 1301 1302 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 1303 SOCK_STREAM, IPPROTO_SCTP, &sock); 1304 if (result < 0) { 1305 log_print("Can't create comms socket, check SCTP is loaded"); 1306 goto out; 1307 } 1308 1309 result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE, 1310 (char *)&bufsize, sizeof(bufsize)); 1311 if (result) 1312 log_print("Error increasing buffer space on socket %d", result); 1313 1314 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one, 1315 sizeof(one)); 1316 if (result < 0) 1317 log_print("Could not set SCTP NODELAY error %d\n", result); 1318 1319 write_lock_bh(&sock->sk->sk_callback_lock); 1320 /* Init con struct */ 1321 sock->sk->sk_user_data = con; 1322 con->sock = sock; 1323 con->sock->sk->sk_data_ready = lowcomms_data_ready; 1324 con->rx_action = sctp_accept_from_sock; 1325 con->connect_action = sctp_connect_to_sock; 1326 1327 write_unlock_bh(&sock->sk->sk_callback_lock); 1328 1329 /* Bind to all addresses. */ 1330 if (sctp_bind_addrs(con, dlm_config.ci_tcp_port)) 1331 goto create_delsock; 1332 1333 result = sock->ops->listen(sock, 5); 1334 if (result < 0) { 1335 log_print("Can't set socket listening"); 1336 goto create_delsock; 1337 } 1338 1339 return 0; 1340 1341 create_delsock: 1342 sock_release(sock); 1343 con->sock = NULL; 1344 out: 1345 return result; 1346 } 1347 1348 static int tcp_listen_for_all(void) 1349 { 1350 struct socket *sock = NULL; 1351 struct connection *con = nodeid2con(0, GFP_NOFS); 1352 int result = -EINVAL; 1353 1354 if (!con) 1355 return -ENOMEM; 1356 1357 /* We don't support multi-homed hosts */ 1358 if (dlm_local_addr[1] != NULL) { 1359 log_print("TCP protocol can't handle multi-homed hosts, " 1360 "try SCTP"); 1361 return -EINVAL; 1362 } 1363 1364 log_print("Using TCP for communications"); 1365 1366 sock = tcp_create_listen_sock(con, dlm_local_addr[0]); 1367 if (sock) { 1368 add_sock(sock, con, true); 1369 result = 0; 1370 } 1371 else { 1372 result = -EADDRINUSE; 1373 } 1374 1375 return result; 1376 } 1377 1378 1379 1380 static struct writequeue_entry *new_writequeue_entry(struct connection *con, 1381 gfp_t allocation) 1382 { 1383 struct writequeue_entry *entry; 1384 1385 entry = kmalloc(sizeof(struct writequeue_entry), allocation); 1386 if (!entry) 1387 return NULL; 1388 1389 entry->page = alloc_page(allocation); 1390 if (!entry->page) { 1391 kfree(entry); 1392 return NULL; 1393 } 1394 1395 entry->offset = 0; 1396 entry->len = 0; 1397 entry->end = 0; 1398 entry->users = 0; 1399 entry->con = con; 1400 1401 return entry; 1402 } 1403 1404 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc) 1405 { 1406 struct connection *con; 1407 struct writequeue_entry *e; 1408 int offset = 0; 1409 1410 con = nodeid2con(nodeid, allocation); 1411 if (!con) 1412 return NULL; 1413 1414 spin_lock(&con->writequeue_lock); 1415 e = list_entry(con->writequeue.prev, struct writequeue_entry, list); 1416 if ((&e->list == &con->writequeue) || 1417 (PAGE_SIZE - e->end < len)) { 1418 e = NULL; 1419 } else { 1420 offset = e->end; 1421 e->end += len; 1422 e->users++; 1423 } 1424 spin_unlock(&con->writequeue_lock); 1425 1426 if (e) { 1427 got_one: 1428 *ppc = page_address(e->page) + offset; 1429 return e; 1430 } 1431 1432 e = new_writequeue_entry(con, allocation); 1433 if (e) { 1434 spin_lock(&con->writequeue_lock); 1435 offset = e->end; 1436 e->end += len; 1437 e->users++; 1438 list_add_tail(&e->list, &con->writequeue); 1439 spin_unlock(&con->writequeue_lock); 1440 goto got_one; 1441 } 1442 return NULL; 1443 } 1444 1445 void dlm_lowcomms_commit_buffer(void *mh) 1446 { 1447 struct writequeue_entry *e = (struct writequeue_entry *)mh; 1448 struct connection *con = e->con; 1449 int users; 1450 1451 spin_lock(&con->writequeue_lock); 1452 users = --e->users; 1453 if (users) 1454 goto out; 1455 e->len = e->end - e->offset; 1456 spin_unlock(&con->writequeue_lock); 1457 1458 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) { 1459 queue_work(send_workqueue, &con->swork); 1460 } 1461 return; 1462 1463 out: 1464 spin_unlock(&con->writequeue_lock); 1465 return; 1466 } 1467 1468 /* Send a message */ 1469 static void send_to_sock(struct connection *con) 1470 { 1471 int ret = 0; 1472 const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL; 1473 struct writequeue_entry *e; 1474 int len, offset; 1475 int count = 0; 1476 1477 mutex_lock(&con->sock_mutex); 1478 if (con->sock == NULL) 1479 goto out_connect; 1480 1481 spin_lock(&con->writequeue_lock); 1482 for (;;) { 1483 e = list_entry(con->writequeue.next, struct writequeue_entry, 1484 list); 1485 if ((struct list_head *) e == &con->writequeue) 1486 break; 1487 1488 len = e->len; 1489 offset = e->offset; 1490 BUG_ON(len == 0 && e->users == 0); 1491 spin_unlock(&con->writequeue_lock); 1492 1493 ret = 0; 1494 if (len) { 1495 ret = kernel_sendpage(con->sock, e->page, offset, len, 1496 msg_flags); 1497 if (ret == -EAGAIN || ret == 0) { 1498 if (ret == -EAGAIN && 1499 test_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags) && 1500 !test_and_set_bit(CF_APP_LIMITED, &con->flags)) { 1501 /* Notify TCP that we're limited by the 1502 * application window size. 1503 */ 1504 set_bit(SOCK_NOSPACE, &con->sock->flags); 1505 con->sock->sk->sk_write_pending++; 1506 } 1507 cond_resched(); 1508 goto out; 1509 } else if (ret < 0) 1510 goto send_error; 1511 } 1512 1513 /* Don't starve people filling buffers */ 1514 if (++count >= MAX_SEND_MSG_COUNT) { 1515 cond_resched(); 1516 count = 0; 1517 } 1518 1519 spin_lock(&con->writequeue_lock); 1520 writequeue_entry_complete(e, ret); 1521 } 1522 spin_unlock(&con->writequeue_lock); 1523 out: 1524 mutex_unlock(&con->sock_mutex); 1525 return; 1526 1527 send_error: 1528 mutex_unlock(&con->sock_mutex); 1529 close_connection(con, false, false, true); 1530 lowcomms_connect_sock(con); 1531 return; 1532 1533 out_connect: 1534 mutex_unlock(&con->sock_mutex); 1535 lowcomms_connect_sock(con); 1536 } 1537 1538 static void clean_one_writequeue(struct connection *con) 1539 { 1540 struct writequeue_entry *e, *safe; 1541 1542 spin_lock(&con->writequeue_lock); 1543 list_for_each_entry_safe(e, safe, &con->writequeue, list) { 1544 list_del(&e->list); 1545 free_entry(e); 1546 } 1547 spin_unlock(&con->writequeue_lock); 1548 } 1549 1550 /* Called from recovery when it knows that a node has 1551 left the cluster */ 1552 int dlm_lowcomms_close(int nodeid) 1553 { 1554 struct connection *con; 1555 struct dlm_node_addr *na; 1556 1557 log_print("closing connection to node %d", nodeid); 1558 con = nodeid2con(nodeid, 0); 1559 if (con) { 1560 set_bit(CF_CLOSE, &con->flags); 1561 close_connection(con, true, true, true); 1562 clean_one_writequeue(con); 1563 } 1564 1565 spin_lock(&dlm_node_addrs_spin); 1566 na = find_node_addr(nodeid); 1567 if (na) { 1568 list_del(&na->list); 1569 while (na->addr_count--) 1570 kfree(na->addr[na->addr_count]); 1571 kfree(na); 1572 } 1573 spin_unlock(&dlm_node_addrs_spin); 1574 1575 return 0; 1576 } 1577 1578 /* Receive workqueue function */ 1579 static void process_recv_sockets(struct work_struct *work) 1580 { 1581 struct connection *con = container_of(work, struct connection, rwork); 1582 int err; 1583 1584 clear_bit(CF_READ_PENDING, &con->flags); 1585 do { 1586 err = con->rx_action(con); 1587 } while (!err); 1588 } 1589 1590 /* Send workqueue function */ 1591 static void process_send_sockets(struct work_struct *work) 1592 { 1593 struct connection *con = container_of(work, struct connection, swork); 1594 1595 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) 1596 con->connect_action(con); 1597 if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags)) 1598 send_to_sock(con); 1599 } 1600 1601 1602 /* Discard all entries on the write queues */ 1603 static void clean_writequeues(void) 1604 { 1605 foreach_conn(clean_one_writequeue); 1606 } 1607 1608 static void work_stop(void) 1609 { 1610 destroy_workqueue(recv_workqueue); 1611 destroy_workqueue(send_workqueue); 1612 } 1613 1614 static int work_start(void) 1615 { 1616 recv_workqueue = alloc_workqueue("dlm_recv", 1617 WQ_UNBOUND | WQ_MEM_RECLAIM, 1); 1618 if (!recv_workqueue) { 1619 log_print("can't start dlm_recv"); 1620 return -ENOMEM; 1621 } 1622 1623 send_workqueue = alloc_workqueue("dlm_send", 1624 WQ_UNBOUND | WQ_MEM_RECLAIM, 1); 1625 if (!send_workqueue) { 1626 log_print("can't start dlm_send"); 1627 destroy_workqueue(recv_workqueue); 1628 return -ENOMEM; 1629 } 1630 1631 return 0; 1632 } 1633 1634 static void stop_conn(struct connection *con) 1635 { 1636 con->flags |= 0x0F; 1637 if (con->sock && con->sock->sk) 1638 con->sock->sk->sk_user_data = NULL; 1639 } 1640 1641 static void free_conn(struct connection *con) 1642 { 1643 close_connection(con, true, true, true); 1644 if (con->othercon) 1645 kmem_cache_free(con_cache, con->othercon); 1646 hlist_del(&con->list); 1647 kmem_cache_free(con_cache, con); 1648 } 1649 1650 void dlm_lowcomms_stop(void) 1651 { 1652 /* Set all the flags to prevent any 1653 socket activity. 1654 */ 1655 mutex_lock(&connections_lock); 1656 dlm_allow_conn = 0; 1657 foreach_conn(stop_conn); 1658 clean_writequeues(); 1659 foreach_conn(free_conn); 1660 mutex_unlock(&connections_lock); 1661 1662 work_stop(); 1663 1664 kmem_cache_destroy(con_cache); 1665 } 1666 1667 int dlm_lowcomms_start(void) 1668 { 1669 int error = -EINVAL; 1670 struct connection *con; 1671 int i; 1672 1673 for (i = 0; i < CONN_HASH_SIZE; i++) 1674 INIT_HLIST_HEAD(&connection_hash[i]); 1675 1676 init_local(); 1677 if (!dlm_local_count) { 1678 error = -ENOTCONN; 1679 log_print("no local IP address has been set"); 1680 goto fail; 1681 } 1682 1683 error = -ENOMEM; 1684 con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection), 1685 __alignof__(struct connection), 0, 1686 NULL); 1687 if (!con_cache) 1688 goto fail; 1689 1690 error = work_start(); 1691 if (error) 1692 goto fail_destroy; 1693 1694 dlm_allow_conn = 1; 1695 1696 /* Start listening */ 1697 if (dlm_config.ci_protocol == 0) 1698 error = tcp_listen_for_all(); 1699 else 1700 error = sctp_listen_for_all(); 1701 if (error) 1702 goto fail_unlisten; 1703 1704 return 0; 1705 1706 fail_unlisten: 1707 dlm_allow_conn = 0; 1708 con = nodeid2con(0,0); 1709 if (con) { 1710 close_connection(con, false, true, true); 1711 kmem_cache_free(con_cache, con); 1712 } 1713 fail_destroy: 1714 kmem_cache_destroy(con_cache); 1715 fail: 1716 return error; 1717 } 1718 1719 void dlm_lowcomms_exit(void) 1720 { 1721 struct dlm_node_addr *na, *safe; 1722 1723 spin_lock(&dlm_node_addrs_spin); 1724 list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) { 1725 list_del(&na->list); 1726 while (na->addr_count--) 1727 kfree(na->addr[na->addr_count]); 1728 kfree(na); 1729 } 1730 spin_unlock(&dlm_node_addrs_spin); 1731 } 1732