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 lock_sock(sk); 523 con->orig_data_ready = sk->sk_data_ready; 524 con->orig_state_change = sk->sk_state_change; 525 con->orig_write_space = sk->sk_write_space; 526 con->orig_error_report = sk->sk_error_report; 527 release_sock(sk); 528 } 529 530 static void restore_callbacks(struct connection *con, struct sock *sk) 531 { 532 write_lock_bh(&sk->sk_callback_lock); 533 lock_sock(sk); 534 sk->sk_user_data = NULL; 535 sk->sk_data_ready = con->orig_data_ready; 536 sk->sk_state_change = con->orig_state_change; 537 sk->sk_write_space = con->orig_write_space; 538 sk->sk_error_report = con->orig_error_report; 539 release_sock(sk); 540 write_unlock_bh(&sk->sk_callback_lock); 541 } 542 543 /* Make a socket active */ 544 static void add_sock(struct socket *sock, struct connection *con, bool save_cb) 545 { 546 struct sock *sk = sock->sk; 547 548 write_lock_bh(&sk->sk_callback_lock); 549 con->sock = sock; 550 551 sk->sk_user_data = con; 552 if (save_cb) 553 save_callbacks(con, sk); 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 clear_bit(CF_CONNECT_PENDING, &con->flags); 587 clear_bit(CF_WRITE_PENDING, &con->flags); 588 if (tx && cancel_work_sync(&con->swork)) 589 log_print("canceled swork for node %d", con->nodeid); 590 if (rx && 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 if (!test_bit(CF_IS_OTHERCON, &con->flags)) 596 restore_callbacks(con, con->sock->sk); 597 sock_release(con->sock); 598 con->sock = NULL; 599 } 600 if (con->othercon && and_other) { 601 /* Will only re-enter once. */ 602 close_connection(con->othercon, false, true, true); 603 } 604 if (con->rx_page) { 605 __free_page(con->rx_page); 606 con->rx_page = NULL; 607 } 608 609 con->retries = 0; 610 mutex_unlock(&con->sock_mutex); 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, false, 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_kern(&init_net, 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); 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 if (!othercon->sock) { 806 newcon->othercon = othercon; 807 othercon->sock = newsock; 808 newsock->sk->sk_user_data = othercon; 809 add_sock(newsock, othercon, false); 810 addcon = othercon; 811 } 812 else { 813 printk("Extra connection from node %d attempted\n", nodeid); 814 result = -EAGAIN; 815 mutex_unlock(&newcon->sock_mutex); 816 goto accept_err; 817 } 818 } 819 else { 820 newsock->sk->sk_user_data = newcon; 821 newcon->rx_action = receive_from_sock; 822 /* accept copies the sk after we've saved the callbacks, so we 823 don't want to save them a second time or comm errors will 824 result in calling sk_error_report recursively. */ 825 add_sock(newsock, newcon, false); 826 addcon = newcon; 827 } 828 829 mutex_unlock(&newcon->sock_mutex); 830 831 /* 832 * Add it to the active queue in case we got data 833 * between processing the accept adding the socket 834 * to the read_sockets list 835 */ 836 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags)) 837 queue_work(recv_workqueue, &addcon->rwork); 838 mutex_unlock(&con->sock_mutex); 839 840 return 0; 841 842 accept_err: 843 mutex_unlock(&con->sock_mutex); 844 sock_release(newsock); 845 846 if (result != -EAGAIN) 847 log_print("error accepting connection from node: %d", result); 848 return result; 849 } 850 851 static int sctp_accept_from_sock(struct connection *con) 852 { 853 /* Check that the new node is in the lockspace */ 854 struct sctp_prim prim; 855 int nodeid; 856 int prim_len, ret; 857 int addr_len; 858 struct connection *newcon; 859 struct connection *addcon; 860 struct socket *newsock; 861 862 mutex_lock(&connections_lock); 863 if (!dlm_allow_conn) { 864 mutex_unlock(&connections_lock); 865 return -1; 866 } 867 mutex_unlock(&connections_lock); 868 869 mutex_lock_nested(&con->sock_mutex, 0); 870 871 ret = kernel_accept(con->sock, &newsock, O_NONBLOCK); 872 if (ret < 0) 873 goto accept_err; 874 875 memset(&prim, 0, sizeof(struct sctp_prim)); 876 prim_len = sizeof(struct sctp_prim); 877 878 ret = kernel_getsockopt(newsock, IPPROTO_SCTP, SCTP_PRIMARY_ADDR, 879 (char *)&prim, &prim_len); 880 if (ret < 0) { 881 log_print("getsockopt/sctp_primary_addr failed: %d", ret); 882 goto accept_err; 883 } 884 885 make_sockaddr(&prim.ssp_addr, 0, &addr_len); 886 if (addr_to_nodeid(&prim.ssp_addr, &nodeid)) { 887 unsigned char *b = (unsigned char *)&prim.ssp_addr; 888 889 log_print("reject connect from unknown addr"); 890 print_hex_dump_bytes("ss: ", DUMP_PREFIX_NONE, 891 b, sizeof(struct sockaddr_storage)); 892 goto accept_err; 893 } 894 895 newcon = nodeid2con(nodeid, GFP_NOFS); 896 if (!newcon) { 897 ret = -ENOMEM; 898 goto accept_err; 899 } 900 901 mutex_lock_nested(&newcon->sock_mutex, 1); 902 903 if (newcon->sock) { 904 struct connection *othercon = newcon->othercon; 905 906 if (!othercon) { 907 othercon = kmem_cache_zalloc(con_cache, GFP_NOFS); 908 if (!othercon) { 909 log_print("failed to allocate incoming socket"); 910 mutex_unlock(&newcon->sock_mutex); 911 ret = -ENOMEM; 912 goto accept_err; 913 } 914 othercon->nodeid = nodeid; 915 othercon->rx_action = receive_from_sock; 916 mutex_init(&othercon->sock_mutex); 917 INIT_WORK(&othercon->swork, process_send_sockets); 918 INIT_WORK(&othercon->rwork, process_recv_sockets); 919 set_bit(CF_IS_OTHERCON, &othercon->flags); 920 } 921 if (!othercon->sock) { 922 newcon->othercon = othercon; 923 othercon->sock = newsock; 924 newsock->sk->sk_user_data = othercon; 925 add_sock(newsock, othercon, false); 926 addcon = othercon; 927 } else { 928 printk("Extra connection from node %d attempted\n", nodeid); 929 ret = -EAGAIN; 930 mutex_unlock(&newcon->sock_mutex); 931 goto accept_err; 932 } 933 } else { 934 newsock->sk->sk_user_data = newcon; 935 newcon->rx_action = receive_from_sock; 936 add_sock(newsock, newcon, false); 937 addcon = newcon; 938 } 939 940 log_print("connected to %d", nodeid); 941 942 mutex_unlock(&newcon->sock_mutex); 943 944 /* 945 * Add it to the active queue in case we got data 946 * between processing the accept adding the socket 947 * to the read_sockets list 948 */ 949 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags)) 950 queue_work(recv_workqueue, &addcon->rwork); 951 mutex_unlock(&con->sock_mutex); 952 953 return 0; 954 955 accept_err: 956 mutex_unlock(&con->sock_mutex); 957 if (newsock) 958 sock_release(newsock); 959 if (ret != -EAGAIN) 960 log_print("error accepting connection from node: %d", ret); 961 962 return ret; 963 } 964 965 static void free_entry(struct writequeue_entry *e) 966 { 967 __free_page(e->page); 968 kfree(e); 969 } 970 971 /* 972 * writequeue_entry_complete - try to delete and free write queue entry 973 * @e: write queue entry to try to delete 974 * @completed: bytes completed 975 * 976 * writequeue_lock must be held. 977 */ 978 static void writequeue_entry_complete(struct writequeue_entry *e, int completed) 979 { 980 e->offset += completed; 981 e->len -= completed; 982 983 if (e->len == 0 && e->users == 0) { 984 list_del(&e->list); 985 free_entry(e); 986 } 987 } 988 989 /* 990 * sctp_bind_addrs - bind a SCTP socket to all our addresses 991 */ 992 static int sctp_bind_addrs(struct connection *con, uint16_t port) 993 { 994 struct sockaddr_storage localaddr; 995 int i, addr_len, result = 0; 996 997 for (i = 0; i < dlm_local_count; i++) { 998 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr)); 999 make_sockaddr(&localaddr, port, &addr_len); 1000 1001 if (!i) 1002 result = kernel_bind(con->sock, 1003 (struct sockaddr *)&localaddr, 1004 addr_len); 1005 else 1006 result = kernel_setsockopt(con->sock, SOL_SCTP, 1007 SCTP_SOCKOPT_BINDX_ADD, 1008 (char *)&localaddr, addr_len); 1009 1010 if (result < 0) { 1011 log_print("Can't bind to %d addr number %d, %d.\n", 1012 port, i + 1, result); 1013 break; 1014 } 1015 } 1016 return result; 1017 } 1018 1019 /* Initiate an SCTP association. 1020 This is a special case of send_to_sock() in that we don't yet have a 1021 peeled-off socket for this association, so we use the listening socket 1022 and add the primary IP address of the remote node. 1023 */ 1024 static void sctp_connect_to_sock(struct connection *con) 1025 { 1026 struct sockaddr_storage daddr; 1027 int one = 1; 1028 int result; 1029 int addr_len; 1030 struct socket *sock; 1031 1032 if (con->nodeid == 0) { 1033 log_print("attempt to connect sock 0 foiled"); 1034 return; 1035 } 1036 1037 mutex_lock(&con->sock_mutex); 1038 1039 /* Some odd races can cause double-connects, ignore them */ 1040 if (con->retries++ > MAX_CONNECT_RETRIES) 1041 goto out; 1042 1043 if (con->sock) { 1044 log_print("node %d already connected.", con->nodeid); 1045 goto out; 1046 } 1047 1048 memset(&daddr, 0, sizeof(daddr)); 1049 result = nodeid_to_addr(con->nodeid, &daddr, NULL, true); 1050 if (result < 0) { 1051 log_print("no address for nodeid %d", con->nodeid); 1052 goto out; 1053 } 1054 1055 /* Create a socket to communicate with */ 1056 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 1057 SOCK_STREAM, IPPROTO_SCTP, &sock); 1058 if (result < 0) 1059 goto socket_err; 1060 1061 sock->sk->sk_user_data = con; 1062 con->rx_action = receive_from_sock; 1063 con->connect_action = sctp_connect_to_sock; 1064 add_sock(sock, con, true); 1065 1066 /* Bind to all addresses. */ 1067 if (sctp_bind_addrs(con, 0)) 1068 goto bind_err; 1069 1070 make_sockaddr(&daddr, dlm_config.ci_tcp_port, &addr_len); 1071 1072 log_print("connecting to %d", con->nodeid); 1073 1074 /* Turn off Nagle's algorithm */ 1075 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one, 1076 sizeof(one)); 1077 1078 result = sock->ops->connect(sock, (struct sockaddr *)&daddr, addr_len, 1079 O_NONBLOCK); 1080 if (result == -EINPROGRESS) 1081 result = 0; 1082 if (result == 0) 1083 goto out; 1084 1085 1086 bind_err: 1087 con->sock = NULL; 1088 sock_release(sock); 1089 1090 socket_err: 1091 /* 1092 * Some errors are fatal and this list might need adjusting. For other 1093 * errors we try again until the max number of retries is reached. 1094 */ 1095 if (result != -EHOSTUNREACH && 1096 result != -ENETUNREACH && 1097 result != -ENETDOWN && 1098 result != -EINVAL && 1099 result != -EPROTONOSUPPORT) { 1100 log_print("connect %d try %d error %d", con->nodeid, 1101 con->retries, result); 1102 mutex_unlock(&con->sock_mutex); 1103 msleep(1000); 1104 clear_bit(CF_CONNECT_PENDING, &con->flags); 1105 lowcomms_connect_sock(con); 1106 return; 1107 } 1108 1109 out: 1110 mutex_unlock(&con->sock_mutex); 1111 set_bit(CF_WRITE_PENDING, &con->flags); 1112 } 1113 1114 /* Connect a new socket to its peer */ 1115 static void tcp_connect_to_sock(struct connection *con) 1116 { 1117 struct sockaddr_storage saddr, src_addr; 1118 int addr_len; 1119 struct socket *sock = NULL; 1120 int one = 1; 1121 int result; 1122 1123 if (con->nodeid == 0) { 1124 log_print("attempt to connect sock 0 foiled"); 1125 return; 1126 } 1127 1128 mutex_lock(&con->sock_mutex); 1129 if (con->retries++ > MAX_CONNECT_RETRIES) 1130 goto out; 1131 1132 /* Some odd races can cause double-connects, ignore them */ 1133 if (con->sock) 1134 goto out; 1135 1136 /* Create a socket to communicate with */ 1137 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 1138 SOCK_STREAM, IPPROTO_TCP, &sock); 1139 if (result < 0) 1140 goto out_err; 1141 1142 memset(&saddr, 0, sizeof(saddr)); 1143 result = nodeid_to_addr(con->nodeid, &saddr, NULL, false); 1144 if (result < 0) { 1145 log_print("no address for nodeid %d", con->nodeid); 1146 goto out_err; 1147 } 1148 1149 sock->sk->sk_user_data = con; 1150 con->rx_action = receive_from_sock; 1151 con->connect_action = tcp_connect_to_sock; 1152 add_sock(sock, con, true); 1153 1154 /* Bind to our cluster-known address connecting to avoid 1155 routing problems */ 1156 memcpy(&src_addr, dlm_local_addr[0], sizeof(src_addr)); 1157 make_sockaddr(&src_addr, 0, &addr_len); 1158 result = sock->ops->bind(sock, (struct sockaddr *) &src_addr, 1159 addr_len); 1160 if (result < 0) { 1161 log_print("could not bind for connect: %d", result); 1162 /* This *may* not indicate a critical error */ 1163 } 1164 1165 make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len); 1166 1167 log_print("connecting to %d", con->nodeid); 1168 1169 /* Turn off Nagle's algorithm */ 1170 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one, 1171 sizeof(one)); 1172 1173 result = sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len, 1174 O_NONBLOCK); 1175 if (result == -EINPROGRESS) 1176 result = 0; 1177 if (result == 0) 1178 goto out; 1179 1180 out_err: 1181 if (con->sock) { 1182 sock_release(con->sock); 1183 con->sock = NULL; 1184 } else if (sock) { 1185 sock_release(sock); 1186 } 1187 /* 1188 * Some errors are fatal and this list might need adjusting. For other 1189 * errors we try again until the max number of retries is reached. 1190 */ 1191 if (result != -EHOSTUNREACH && 1192 result != -ENETUNREACH && 1193 result != -ENETDOWN && 1194 result != -EINVAL && 1195 result != -EPROTONOSUPPORT) { 1196 log_print("connect %d try %d error %d", con->nodeid, 1197 con->retries, result); 1198 mutex_unlock(&con->sock_mutex); 1199 msleep(1000); 1200 clear_bit(CF_CONNECT_PENDING, &con->flags); 1201 lowcomms_connect_sock(con); 1202 return; 1203 } 1204 out: 1205 mutex_unlock(&con->sock_mutex); 1206 set_bit(CF_WRITE_PENDING, &con->flags); 1207 return; 1208 } 1209 1210 static struct socket *tcp_create_listen_sock(struct connection *con, 1211 struct sockaddr_storage *saddr) 1212 { 1213 struct socket *sock = NULL; 1214 int result = 0; 1215 int one = 1; 1216 int addr_len; 1217 1218 if (dlm_local_addr[0]->ss_family == AF_INET) 1219 addr_len = sizeof(struct sockaddr_in); 1220 else 1221 addr_len = sizeof(struct sockaddr_in6); 1222 1223 /* Create a socket to communicate with */ 1224 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 1225 SOCK_STREAM, IPPROTO_TCP, &sock); 1226 if (result < 0) { 1227 log_print("Can't create listening comms socket"); 1228 goto create_out; 1229 } 1230 1231 /* Turn off Nagle's algorithm */ 1232 kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY, (char *)&one, 1233 sizeof(one)); 1234 1235 result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, 1236 (char *)&one, sizeof(one)); 1237 1238 if (result < 0) { 1239 log_print("Failed to set SO_REUSEADDR on socket: %d", result); 1240 } 1241 sock->sk->sk_user_data = con; 1242 1243 con->rx_action = tcp_accept_from_sock; 1244 con->connect_action = tcp_connect_to_sock; 1245 1246 /* Bind to our port */ 1247 make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len); 1248 result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len); 1249 if (result < 0) { 1250 log_print("Can't bind to port %d", dlm_config.ci_tcp_port); 1251 sock_release(sock); 1252 sock = NULL; 1253 con->sock = NULL; 1254 goto create_out; 1255 } 1256 result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE, 1257 (char *)&one, sizeof(one)); 1258 if (result < 0) { 1259 log_print("Set keepalive failed: %d", result); 1260 } 1261 1262 result = sock->ops->listen(sock, 5); 1263 if (result < 0) { 1264 log_print("Can't listen on port %d", dlm_config.ci_tcp_port); 1265 sock_release(sock); 1266 sock = NULL; 1267 goto create_out; 1268 } 1269 1270 create_out: 1271 return sock; 1272 } 1273 1274 /* Get local addresses */ 1275 static void init_local(void) 1276 { 1277 struct sockaddr_storage sas, *addr; 1278 int i; 1279 1280 dlm_local_count = 0; 1281 for (i = 0; i < DLM_MAX_ADDR_COUNT; i++) { 1282 if (dlm_our_addr(&sas, i)) 1283 break; 1284 1285 addr = kmemdup(&sas, sizeof(*addr), GFP_NOFS); 1286 if (!addr) 1287 break; 1288 dlm_local_addr[dlm_local_count++] = addr; 1289 } 1290 } 1291 1292 /* Initialise SCTP socket and bind to all interfaces */ 1293 static int sctp_listen_for_all(void) 1294 { 1295 struct socket *sock = NULL; 1296 int result = -EINVAL; 1297 struct connection *con = nodeid2con(0, GFP_NOFS); 1298 int bufsize = NEEDED_RMEM; 1299 int one = 1; 1300 1301 if (!con) 1302 return -ENOMEM; 1303 1304 log_print("Using SCTP for communications"); 1305 1306 result = sock_create_kern(&init_net, dlm_local_addr[0]->ss_family, 1307 SOCK_STREAM, IPPROTO_SCTP, &sock); 1308 if (result < 0) { 1309 log_print("Can't create comms socket, check SCTP is loaded"); 1310 goto out; 1311 } 1312 1313 result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUFFORCE, 1314 (char *)&bufsize, sizeof(bufsize)); 1315 if (result) 1316 log_print("Error increasing buffer space on socket %d", result); 1317 1318 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_NODELAY, (char *)&one, 1319 sizeof(one)); 1320 if (result < 0) 1321 log_print("Could not set SCTP NODELAY error %d\n", result); 1322 1323 write_lock_bh(&sock->sk->sk_callback_lock); 1324 /* Init con struct */ 1325 sock->sk->sk_user_data = con; 1326 con->sock = sock; 1327 con->sock->sk->sk_data_ready = lowcomms_data_ready; 1328 con->rx_action = sctp_accept_from_sock; 1329 con->connect_action = sctp_connect_to_sock; 1330 1331 write_unlock_bh(&sock->sk->sk_callback_lock); 1332 1333 /* Bind to all addresses. */ 1334 if (sctp_bind_addrs(con, dlm_config.ci_tcp_port)) 1335 goto create_delsock; 1336 1337 result = sock->ops->listen(sock, 5); 1338 if (result < 0) { 1339 log_print("Can't set socket listening"); 1340 goto create_delsock; 1341 } 1342 1343 return 0; 1344 1345 create_delsock: 1346 sock_release(sock); 1347 con->sock = NULL; 1348 out: 1349 return result; 1350 } 1351 1352 static int tcp_listen_for_all(void) 1353 { 1354 struct socket *sock = NULL; 1355 struct connection *con = nodeid2con(0, GFP_NOFS); 1356 int result = -EINVAL; 1357 1358 if (!con) 1359 return -ENOMEM; 1360 1361 /* We don't support multi-homed hosts */ 1362 if (dlm_local_addr[1] != NULL) { 1363 log_print("TCP protocol can't handle multi-homed hosts, " 1364 "try SCTP"); 1365 return -EINVAL; 1366 } 1367 1368 log_print("Using TCP for communications"); 1369 1370 sock = tcp_create_listen_sock(con, dlm_local_addr[0]); 1371 if (sock) { 1372 add_sock(sock, con, true); 1373 result = 0; 1374 } 1375 else { 1376 result = -EADDRINUSE; 1377 } 1378 1379 return result; 1380 } 1381 1382 1383 1384 static struct writequeue_entry *new_writequeue_entry(struct connection *con, 1385 gfp_t allocation) 1386 { 1387 struct writequeue_entry *entry; 1388 1389 entry = kmalloc(sizeof(struct writequeue_entry), allocation); 1390 if (!entry) 1391 return NULL; 1392 1393 entry->page = alloc_page(allocation); 1394 if (!entry->page) { 1395 kfree(entry); 1396 return NULL; 1397 } 1398 1399 entry->offset = 0; 1400 entry->len = 0; 1401 entry->end = 0; 1402 entry->users = 0; 1403 entry->con = con; 1404 1405 return entry; 1406 } 1407 1408 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc) 1409 { 1410 struct connection *con; 1411 struct writequeue_entry *e; 1412 int offset = 0; 1413 1414 con = nodeid2con(nodeid, allocation); 1415 if (!con) 1416 return NULL; 1417 1418 spin_lock(&con->writequeue_lock); 1419 e = list_entry(con->writequeue.prev, struct writequeue_entry, list); 1420 if ((&e->list == &con->writequeue) || 1421 (PAGE_SIZE - e->end < len)) { 1422 e = NULL; 1423 } else { 1424 offset = e->end; 1425 e->end += len; 1426 e->users++; 1427 } 1428 spin_unlock(&con->writequeue_lock); 1429 1430 if (e) { 1431 got_one: 1432 *ppc = page_address(e->page) + offset; 1433 return e; 1434 } 1435 1436 e = new_writequeue_entry(con, allocation); 1437 if (e) { 1438 spin_lock(&con->writequeue_lock); 1439 offset = e->end; 1440 e->end += len; 1441 e->users++; 1442 list_add_tail(&e->list, &con->writequeue); 1443 spin_unlock(&con->writequeue_lock); 1444 goto got_one; 1445 } 1446 return NULL; 1447 } 1448 1449 void dlm_lowcomms_commit_buffer(void *mh) 1450 { 1451 struct writequeue_entry *e = (struct writequeue_entry *)mh; 1452 struct connection *con = e->con; 1453 int users; 1454 1455 spin_lock(&con->writequeue_lock); 1456 users = --e->users; 1457 if (users) 1458 goto out; 1459 e->len = e->end - e->offset; 1460 spin_unlock(&con->writequeue_lock); 1461 1462 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) { 1463 queue_work(send_workqueue, &con->swork); 1464 } 1465 return; 1466 1467 out: 1468 spin_unlock(&con->writequeue_lock); 1469 return; 1470 } 1471 1472 /* Send a message */ 1473 static void send_to_sock(struct connection *con) 1474 { 1475 int ret = 0; 1476 const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL; 1477 struct writequeue_entry *e; 1478 int len, offset; 1479 int count = 0; 1480 1481 mutex_lock(&con->sock_mutex); 1482 if (con->sock == NULL) 1483 goto out_connect; 1484 1485 spin_lock(&con->writequeue_lock); 1486 for (;;) { 1487 e = list_entry(con->writequeue.next, struct writequeue_entry, 1488 list); 1489 if ((struct list_head *) e == &con->writequeue) 1490 break; 1491 1492 len = e->len; 1493 offset = e->offset; 1494 BUG_ON(len == 0 && e->users == 0); 1495 spin_unlock(&con->writequeue_lock); 1496 1497 ret = 0; 1498 if (len) { 1499 ret = kernel_sendpage(con->sock, e->page, offset, len, 1500 msg_flags); 1501 if (ret == -EAGAIN || ret == 0) { 1502 if (ret == -EAGAIN && 1503 test_bit(SOCKWQ_ASYNC_NOSPACE, &con->sock->flags) && 1504 !test_and_set_bit(CF_APP_LIMITED, &con->flags)) { 1505 /* Notify TCP that we're limited by the 1506 * application window size. 1507 */ 1508 set_bit(SOCK_NOSPACE, &con->sock->flags); 1509 con->sock->sk->sk_write_pending++; 1510 } 1511 cond_resched(); 1512 goto out; 1513 } else if (ret < 0) 1514 goto send_error; 1515 } 1516 1517 /* Don't starve people filling buffers */ 1518 if (++count >= MAX_SEND_MSG_COUNT) { 1519 cond_resched(); 1520 count = 0; 1521 } 1522 1523 spin_lock(&con->writequeue_lock); 1524 writequeue_entry_complete(e, ret); 1525 } 1526 spin_unlock(&con->writequeue_lock); 1527 out: 1528 mutex_unlock(&con->sock_mutex); 1529 return; 1530 1531 send_error: 1532 mutex_unlock(&con->sock_mutex); 1533 close_connection(con, false, false, true); 1534 lowcomms_connect_sock(con); 1535 return; 1536 1537 out_connect: 1538 mutex_unlock(&con->sock_mutex); 1539 lowcomms_connect_sock(con); 1540 } 1541 1542 static void clean_one_writequeue(struct connection *con) 1543 { 1544 struct writequeue_entry *e, *safe; 1545 1546 spin_lock(&con->writequeue_lock); 1547 list_for_each_entry_safe(e, safe, &con->writequeue, list) { 1548 list_del(&e->list); 1549 free_entry(e); 1550 } 1551 spin_unlock(&con->writequeue_lock); 1552 } 1553 1554 /* Called from recovery when it knows that a node has 1555 left the cluster */ 1556 int dlm_lowcomms_close(int nodeid) 1557 { 1558 struct connection *con; 1559 struct dlm_node_addr *na; 1560 1561 log_print("closing connection to node %d", nodeid); 1562 con = nodeid2con(nodeid, 0); 1563 if (con) { 1564 set_bit(CF_CLOSE, &con->flags); 1565 close_connection(con, true, true, true); 1566 clean_one_writequeue(con); 1567 } 1568 1569 spin_lock(&dlm_node_addrs_spin); 1570 na = find_node_addr(nodeid); 1571 if (na) { 1572 list_del(&na->list); 1573 while (na->addr_count--) 1574 kfree(na->addr[na->addr_count]); 1575 kfree(na); 1576 } 1577 spin_unlock(&dlm_node_addrs_spin); 1578 1579 return 0; 1580 } 1581 1582 /* Receive workqueue function */ 1583 static void process_recv_sockets(struct work_struct *work) 1584 { 1585 struct connection *con = container_of(work, struct connection, rwork); 1586 int err; 1587 1588 clear_bit(CF_READ_PENDING, &con->flags); 1589 do { 1590 err = con->rx_action(con); 1591 } while (!err); 1592 } 1593 1594 /* Send workqueue function */ 1595 static void process_send_sockets(struct work_struct *work) 1596 { 1597 struct connection *con = container_of(work, struct connection, swork); 1598 1599 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) 1600 con->connect_action(con); 1601 if (test_and_clear_bit(CF_WRITE_PENDING, &con->flags)) 1602 send_to_sock(con); 1603 } 1604 1605 1606 /* Discard all entries on the write queues */ 1607 static void clean_writequeues(void) 1608 { 1609 foreach_conn(clean_one_writequeue); 1610 } 1611 1612 static void work_stop(void) 1613 { 1614 destroy_workqueue(recv_workqueue); 1615 destroy_workqueue(send_workqueue); 1616 } 1617 1618 static int work_start(void) 1619 { 1620 recv_workqueue = alloc_workqueue("dlm_recv", 1621 WQ_UNBOUND | WQ_MEM_RECLAIM, 1); 1622 if (!recv_workqueue) { 1623 log_print("can't start dlm_recv"); 1624 return -ENOMEM; 1625 } 1626 1627 send_workqueue = alloc_workqueue("dlm_send", 1628 WQ_UNBOUND | WQ_MEM_RECLAIM, 1); 1629 if (!send_workqueue) { 1630 log_print("can't start dlm_send"); 1631 destroy_workqueue(recv_workqueue); 1632 return -ENOMEM; 1633 } 1634 1635 return 0; 1636 } 1637 1638 static void stop_conn(struct connection *con) 1639 { 1640 con->flags |= 0x0F; 1641 if (con->sock && con->sock->sk) 1642 con->sock->sk->sk_user_data = NULL; 1643 } 1644 1645 static void free_conn(struct connection *con) 1646 { 1647 close_connection(con, true, true, true); 1648 if (con->othercon) 1649 kmem_cache_free(con_cache, con->othercon); 1650 hlist_del(&con->list); 1651 kmem_cache_free(con_cache, con); 1652 } 1653 1654 void dlm_lowcomms_stop(void) 1655 { 1656 /* Set all the flags to prevent any 1657 socket activity. 1658 */ 1659 mutex_lock(&connections_lock); 1660 dlm_allow_conn = 0; 1661 foreach_conn(stop_conn); 1662 clean_writequeues(); 1663 foreach_conn(free_conn); 1664 mutex_unlock(&connections_lock); 1665 1666 work_stop(); 1667 1668 kmem_cache_destroy(con_cache); 1669 } 1670 1671 int dlm_lowcomms_start(void) 1672 { 1673 int error = -EINVAL; 1674 struct connection *con; 1675 int i; 1676 1677 for (i = 0; i < CONN_HASH_SIZE; i++) 1678 INIT_HLIST_HEAD(&connection_hash[i]); 1679 1680 init_local(); 1681 if (!dlm_local_count) { 1682 error = -ENOTCONN; 1683 log_print("no local IP address has been set"); 1684 goto fail; 1685 } 1686 1687 error = -ENOMEM; 1688 con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection), 1689 __alignof__(struct connection), 0, 1690 NULL); 1691 if (!con_cache) 1692 goto fail; 1693 1694 error = work_start(); 1695 if (error) 1696 goto fail_destroy; 1697 1698 dlm_allow_conn = 1; 1699 1700 /* Start listening */ 1701 if (dlm_config.ci_protocol == 0) 1702 error = tcp_listen_for_all(); 1703 else 1704 error = sctp_listen_for_all(); 1705 if (error) 1706 goto fail_unlisten; 1707 1708 return 0; 1709 1710 fail_unlisten: 1711 dlm_allow_conn = 0; 1712 con = nodeid2con(0,0); 1713 if (con) { 1714 close_connection(con, false, true, true); 1715 kmem_cache_free(con_cache, con); 1716 } 1717 fail_destroy: 1718 kmem_cache_destroy(con_cache); 1719 fail: 1720 return error; 1721 } 1722 1723 void dlm_lowcomms_exit(void) 1724 { 1725 struct dlm_node_addr *na, *safe; 1726 1727 spin_lock(&dlm_node_addrs_spin); 1728 list_for_each_entry_safe(na, safe, &dlm_node_addrs, list) { 1729 list_del(&na->list); 1730 while (na->addr_count--) 1731 kfree(na->addr[na->addr_count]); 1732 kfree(na); 1733 } 1734 spin_unlock(&dlm_node_addrs_spin); 1735 } 1736