lowcomms.c revision afb853fb4eec380b492a3c369f837359359c28e8
1/****************************************************************************** 2******************************************************************************* 3** 4** Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved. 5** Copyright (C) 2004-2007 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 it's 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 wither 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 shouldbe 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/idr.h> 52#include <linux/file.h> 53#include <linux/sctp.h> 54#include <net/sctp/user.h> 55 56#include "dlm_internal.h" 57#include "lowcomms.h" 58#include "midcomms.h" 59#include "config.h" 60 61#define NEEDED_RMEM (4*1024*1024) 62 63struct cbuf { 64 unsigned int base; 65 unsigned int len; 66 unsigned int mask; 67}; 68 69static void cbuf_add(struct cbuf *cb, int n) 70{ 71 cb->len += n; 72} 73 74static int cbuf_data(struct cbuf *cb) 75{ 76 return ((cb->base + cb->len) & cb->mask); 77} 78 79static void cbuf_init(struct cbuf *cb, int size) 80{ 81 cb->base = cb->len = 0; 82 cb->mask = size-1; 83} 84 85static void cbuf_eat(struct cbuf *cb, int n) 86{ 87 cb->len -= n; 88 cb->base += n; 89 cb->base &= cb->mask; 90} 91 92static bool cbuf_empty(struct cbuf *cb) 93{ 94 return cb->len == 0; 95} 96 97struct connection { 98 struct socket *sock; /* NULL if not connected */ 99 uint32_t nodeid; /* So we know who we are in the list */ 100 struct mutex sock_mutex; 101 unsigned long flags; 102#define CF_READ_PENDING 1 103#define CF_WRITE_PENDING 2 104#define CF_CONNECT_PENDING 3 105#define CF_INIT_PENDING 4 106#define CF_IS_OTHERCON 5 107 struct list_head writequeue; /* List of outgoing writequeue_entries */ 108 spinlock_t writequeue_lock; 109 int (*rx_action) (struct connection *); /* What to do when active */ 110 void (*connect_action) (struct connection *); /* What to do to connect */ 111 struct page *rx_page; 112 struct cbuf cb; 113 int retries; 114#define MAX_CONNECT_RETRIES 3 115 int sctp_assoc; 116 struct connection *othercon; 117 struct work_struct rwork; /* Receive workqueue */ 118 struct work_struct swork; /* Send workqueue */ 119}; 120#define sock2con(x) ((struct connection *)(x)->sk_user_data) 121 122/* An entry waiting to be sent */ 123struct writequeue_entry { 124 struct list_head list; 125 struct page *page; 126 int offset; 127 int len; 128 int end; 129 int users; 130 struct connection *con; 131}; 132 133static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT]; 134static int dlm_local_count; 135 136/* Work queues */ 137static struct workqueue_struct *recv_workqueue; 138static struct workqueue_struct *send_workqueue; 139 140static DEFINE_IDR(connections_idr); 141static DECLARE_MUTEX(connections_lock); 142static int max_nodeid; 143static struct kmem_cache *con_cache; 144 145static void process_recv_sockets(struct work_struct *work); 146static void process_send_sockets(struct work_struct *work); 147 148/* 149 * If 'allocation' is zero then we don't attempt to create a new 150 * connection structure for this node. 151 */ 152static struct connection *__nodeid2con(int nodeid, gfp_t alloc) 153{ 154 struct connection *con = NULL; 155 int r; 156 int n; 157 158 con = idr_find(&connections_idr, nodeid); 159 if (con || !alloc) 160 return con; 161 162 r = idr_pre_get(&connections_idr, alloc); 163 if (!r) 164 return NULL; 165 166 con = kmem_cache_zalloc(con_cache, alloc); 167 if (!con) 168 return NULL; 169 170 r = idr_get_new_above(&connections_idr, con, nodeid, &n); 171 if (r) { 172 kmem_cache_free(con_cache, con); 173 return NULL; 174 } 175 176 if (n != nodeid) { 177 idr_remove(&connections_idr, n); 178 kmem_cache_free(con_cache, con); 179 return NULL; 180 } 181 182 con->nodeid = nodeid; 183 mutex_init(&con->sock_mutex); 184 INIT_LIST_HEAD(&con->writequeue); 185 spin_lock_init(&con->writequeue_lock); 186 INIT_WORK(&con->swork, process_send_sockets); 187 INIT_WORK(&con->rwork, process_recv_sockets); 188 189 /* Setup action pointers for child sockets */ 190 if (con->nodeid) { 191 struct connection *zerocon = idr_find(&connections_idr, 0); 192 193 con->connect_action = zerocon->connect_action; 194 if (!con->rx_action) 195 con->rx_action = zerocon->rx_action; 196 } 197 198 if (nodeid > max_nodeid) 199 max_nodeid = nodeid; 200 201 return con; 202} 203 204static struct connection *nodeid2con(int nodeid, gfp_t allocation) 205{ 206 struct connection *con; 207 208 down(&connections_lock); 209 con = __nodeid2con(nodeid, allocation); 210 up(&connections_lock); 211 212 return con; 213} 214 215/* This is a bit drastic, but only called when things go wrong */ 216static struct connection *assoc2con(int assoc_id) 217{ 218 int i; 219 struct connection *con; 220 221 down(&connections_lock); 222 for (i=0; i<=max_nodeid; i++) { 223 con = __nodeid2con(i, 0); 224 if (con && con->sctp_assoc == assoc_id) { 225 up(&connections_lock); 226 return con; 227 } 228 } 229 up(&connections_lock); 230 return NULL; 231} 232 233static int nodeid_to_addr(int nodeid, struct sockaddr *retaddr) 234{ 235 struct sockaddr_storage addr; 236 int error; 237 238 if (!dlm_local_count) 239 return -1; 240 241 error = dlm_nodeid_to_addr(nodeid, &addr); 242 if (error) 243 return error; 244 245 if (dlm_local_addr[0]->ss_family == AF_INET) { 246 struct sockaddr_in *in4 = (struct sockaddr_in *) &addr; 247 struct sockaddr_in *ret4 = (struct sockaddr_in *) retaddr; 248 ret4->sin_addr.s_addr = in4->sin_addr.s_addr; 249 } else { 250 struct sockaddr_in6 *in6 = (struct sockaddr_in6 *) &addr; 251 struct sockaddr_in6 *ret6 = (struct sockaddr_in6 *) retaddr; 252 memcpy(&ret6->sin6_addr, &in6->sin6_addr, 253 sizeof(in6->sin6_addr)); 254 } 255 256 return 0; 257} 258 259/* Data available on socket or listen socket received a connect */ 260static void lowcomms_data_ready(struct sock *sk, int count_unused) 261{ 262 struct connection *con = sock2con(sk); 263 if (con && !test_and_set_bit(CF_READ_PENDING, &con->flags)) 264 queue_work(recv_workqueue, &con->rwork); 265} 266 267static void lowcomms_write_space(struct sock *sk) 268{ 269 struct connection *con = sock2con(sk); 270 271 if (con && !test_and_set_bit(CF_WRITE_PENDING, &con->flags)) 272 queue_work(send_workqueue, &con->swork); 273} 274 275static inline void lowcomms_connect_sock(struct connection *con) 276{ 277 if (!test_and_set_bit(CF_CONNECT_PENDING, &con->flags)) 278 queue_work(send_workqueue, &con->swork); 279} 280 281static void lowcomms_state_change(struct sock *sk) 282{ 283 if (sk->sk_state == TCP_ESTABLISHED) 284 lowcomms_write_space(sk); 285} 286 287/* Make a socket active */ 288static int add_sock(struct socket *sock, struct connection *con) 289{ 290 con->sock = sock; 291 292 /* Install a data_ready callback */ 293 con->sock->sk->sk_data_ready = lowcomms_data_ready; 294 con->sock->sk->sk_write_space = lowcomms_write_space; 295 con->sock->sk->sk_state_change = lowcomms_state_change; 296 con->sock->sk->sk_user_data = con; 297 return 0; 298} 299 300/* Add the port number to an IPv6 or 4 sockaddr and return the address 301 length */ 302static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port, 303 int *addr_len) 304{ 305 saddr->ss_family = dlm_local_addr[0]->ss_family; 306 if (saddr->ss_family == AF_INET) { 307 struct sockaddr_in *in4_addr = (struct sockaddr_in *)saddr; 308 in4_addr->sin_port = cpu_to_be16(port); 309 *addr_len = sizeof(struct sockaddr_in); 310 memset(&in4_addr->sin_zero, 0, sizeof(in4_addr->sin_zero)); 311 } else { 312 struct sockaddr_in6 *in6_addr = (struct sockaddr_in6 *)saddr; 313 in6_addr->sin6_port = cpu_to_be16(port); 314 *addr_len = sizeof(struct sockaddr_in6); 315 } 316} 317 318/* Close a remote connection and tidy up */ 319static void close_connection(struct connection *con, bool and_other) 320{ 321 mutex_lock(&con->sock_mutex); 322 323 if (con->sock) { 324 sock_release(con->sock); 325 con->sock = NULL; 326 } 327 if (con->othercon && and_other) { 328 /* Will only re-enter once. */ 329 close_connection(con->othercon, false); 330 } 331 if (con->rx_page) { 332 __free_page(con->rx_page); 333 con->rx_page = NULL; 334 } 335 con->retries = 0; 336 mutex_unlock(&con->sock_mutex); 337} 338 339/* We only send shutdown messages to nodes that are not part of the cluster */ 340static void sctp_send_shutdown(sctp_assoc_t associd) 341{ 342 static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))]; 343 struct msghdr outmessage; 344 struct cmsghdr *cmsg; 345 struct sctp_sndrcvinfo *sinfo; 346 int ret; 347 struct connection *con; 348 349 con = nodeid2con(0,0); 350 BUG_ON(con == NULL); 351 352 outmessage.msg_name = NULL; 353 outmessage.msg_namelen = 0; 354 outmessage.msg_control = outcmsg; 355 outmessage.msg_controllen = sizeof(outcmsg); 356 outmessage.msg_flags = MSG_EOR; 357 358 cmsg = CMSG_FIRSTHDR(&outmessage); 359 cmsg->cmsg_level = IPPROTO_SCTP; 360 cmsg->cmsg_type = SCTP_SNDRCV; 361 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo)); 362 outmessage.msg_controllen = cmsg->cmsg_len; 363 sinfo = CMSG_DATA(cmsg); 364 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo)); 365 366 sinfo->sinfo_flags |= MSG_EOF; 367 sinfo->sinfo_assoc_id = associd; 368 369 ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0); 370 371 if (ret != 0) 372 log_print("send EOF to node failed: %d", ret); 373} 374 375/* INIT failed but we don't know which node... 376 restart INIT on all pending nodes */ 377static void sctp_init_failed(void) 378{ 379 int i; 380 struct connection *con; 381 382 down(&connections_lock); 383 for (i=1; i<=max_nodeid; i++) { 384 con = __nodeid2con(i, 0); 385 if (!con) 386 continue; 387 con->sctp_assoc = 0; 388 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) { 389 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) { 390 queue_work(send_workqueue, &con->swork); 391 } 392 } 393 } 394 up(&connections_lock); 395} 396 397/* Something happened to an association */ 398static void process_sctp_notification(struct connection *con, 399 struct msghdr *msg, char *buf) 400{ 401 union sctp_notification *sn = (union sctp_notification *)buf; 402 403 if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) { 404 switch (sn->sn_assoc_change.sac_state) { 405 406 case SCTP_COMM_UP: 407 case SCTP_RESTART: 408 { 409 /* Check that the new node is in the lockspace */ 410 struct sctp_prim prim; 411 int nodeid; 412 int prim_len, ret; 413 int addr_len; 414 struct connection *new_con; 415 struct file *file; 416 sctp_peeloff_arg_t parg; 417 int parglen = sizeof(parg); 418 419 /* 420 * We get this before any data for an association. 421 * We verify that the node is in the cluster and 422 * then peel off a socket for it. 423 */ 424 if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) { 425 log_print("COMM_UP for invalid assoc ID %d", 426 (int)sn->sn_assoc_change.sac_assoc_id); 427 sctp_init_failed(); 428 return; 429 } 430 memset(&prim, 0, sizeof(struct sctp_prim)); 431 prim_len = sizeof(struct sctp_prim); 432 prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id; 433 434 ret = kernel_getsockopt(con->sock, 435 IPPROTO_SCTP, 436 SCTP_PRIMARY_ADDR, 437 (char*)&prim, 438 &prim_len); 439 if (ret < 0) { 440 log_print("getsockopt/sctp_primary_addr on " 441 "new assoc %d failed : %d", 442 (int)sn->sn_assoc_change.sac_assoc_id, 443 ret); 444 445 /* Retry INIT later */ 446 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id); 447 if (new_con) 448 clear_bit(CF_CONNECT_PENDING, &con->flags); 449 return; 450 } 451 make_sockaddr(&prim.ssp_addr, 0, &addr_len); 452 if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) { 453 int i; 454 unsigned char *b=(unsigned char *)&prim.ssp_addr; 455 log_print("reject connect from unknown addr"); 456 for (i=0; i<sizeof(struct sockaddr_storage);i++) 457 printk("%02x ", b[i]); 458 printk("\n"); 459 sctp_send_shutdown(prim.ssp_assoc_id); 460 return; 461 } 462 463 new_con = nodeid2con(nodeid, GFP_KERNEL); 464 if (!new_con) 465 return; 466 467 /* Peel off a new sock */ 468 parg.associd = sn->sn_assoc_change.sac_assoc_id; 469 ret = kernel_getsockopt(con->sock, IPPROTO_SCTP, 470 SCTP_SOCKOPT_PEELOFF, 471 (void *)&parg, &parglen); 472 if (ret) { 473 log_print("Can't peel off a socket for " 474 "connection %d to node %d: err=%d\n", 475 parg.associd, nodeid, ret); 476 } 477 file = fget(parg.sd); 478 new_con->sock = SOCKET_I(file->f_dentry->d_inode); 479 add_sock(new_con->sock, new_con); 480 fput(file); 481 put_unused_fd(parg.sd); 482 483 log_print("got new/restarted association %d nodeid %d", 484 (int)sn->sn_assoc_change.sac_assoc_id, nodeid); 485 486 /* Send any pending writes */ 487 clear_bit(CF_CONNECT_PENDING, &new_con->flags); 488 clear_bit(CF_INIT_PENDING, &con->flags); 489 if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) { 490 queue_work(send_workqueue, &new_con->swork); 491 } 492 if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags)) 493 queue_work(recv_workqueue, &new_con->rwork); 494 } 495 break; 496 497 case SCTP_COMM_LOST: 498 case SCTP_SHUTDOWN_COMP: 499 { 500 con = assoc2con(sn->sn_assoc_change.sac_assoc_id); 501 if (con) { 502 con->sctp_assoc = 0; 503 } 504 } 505 break; 506 507 /* We don't know which INIT failed, so clear the PENDING flags 508 * on them all. if assoc_id is zero then it will then try 509 * again */ 510 511 case SCTP_CANT_STR_ASSOC: 512 { 513 log_print("Can't start SCTP association - retrying"); 514 sctp_init_failed(); 515 } 516 break; 517 518 default: 519 log_print("unexpected SCTP assoc change id=%d state=%d", 520 (int)sn->sn_assoc_change.sac_assoc_id, 521 sn->sn_assoc_change.sac_state); 522 } 523 } 524} 525 526/* Data received from remote end */ 527static int receive_from_sock(struct connection *con) 528{ 529 int ret = 0; 530 struct msghdr msg = {}; 531 struct kvec iov[2]; 532 unsigned len; 533 int r; 534 int call_again_soon = 0; 535 int nvec; 536 char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))]; 537 538 mutex_lock(&con->sock_mutex); 539 540 if (con->sock == NULL) { 541 ret = -EAGAIN; 542 goto out_close; 543 } 544 545 if (con->rx_page == NULL) { 546 /* 547 * This doesn't need to be atomic, but I think it should 548 * improve performance if it is. 549 */ 550 con->rx_page = alloc_page(GFP_ATOMIC); 551 if (con->rx_page == NULL) 552 goto out_resched; 553 cbuf_init(&con->cb, PAGE_CACHE_SIZE); 554 } 555 556 /* Only SCTP needs these really */ 557 memset(&incmsg, 0, sizeof(incmsg)); 558 msg.msg_control = incmsg; 559 msg.msg_controllen = sizeof(incmsg); 560 561 /* 562 * iov[0] is the bit of the circular buffer between the current end 563 * point (cb.base + cb.len) and the end of the buffer. 564 */ 565 iov[0].iov_len = con->cb.base - cbuf_data(&con->cb); 566 iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb); 567 iov[1].iov_len = 0; 568 nvec = 1; 569 570 /* 571 * iov[1] is the bit of the circular buffer between the start of the 572 * buffer and the start of the currently used section (cb.base) 573 */ 574 if (cbuf_data(&con->cb) >= con->cb.base) { 575 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb); 576 iov[1].iov_len = con->cb.base; 577 iov[1].iov_base = page_address(con->rx_page); 578 nvec = 2; 579 } 580 len = iov[0].iov_len + iov[1].iov_len; 581 582 r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len, 583 MSG_DONTWAIT | MSG_NOSIGNAL); 584 if (ret <= 0) 585 goto out_close; 586 587 /* Process SCTP notifications */ 588 if (msg.msg_flags & MSG_NOTIFICATION) { 589 msg.msg_control = incmsg; 590 msg.msg_controllen = sizeof(incmsg); 591 592 process_sctp_notification(con, &msg, 593 page_address(con->rx_page) + con->cb.base); 594 mutex_unlock(&con->sock_mutex); 595 return 0; 596 } 597 BUG_ON(con->nodeid == 0); 598 599 if (ret == len) 600 call_again_soon = 1; 601 cbuf_add(&con->cb, ret); 602 ret = dlm_process_incoming_buffer(con->nodeid, 603 page_address(con->rx_page), 604 con->cb.base, con->cb.len, 605 PAGE_CACHE_SIZE); 606 if (ret == -EBADMSG) { 607 log_print("lowcomms: addr=%p, base=%u, len=%u, " 608 "iov_len=%u, iov_base[0]=%p, read=%d", 609 page_address(con->rx_page), con->cb.base, con->cb.len, 610 len, iov[0].iov_base, r); 611 } 612 if (ret < 0) 613 goto out_close; 614 cbuf_eat(&con->cb, ret); 615 616 if (cbuf_empty(&con->cb) && !call_again_soon) { 617 __free_page(con->rx_page); 618 con->rx_page = NULL; 619 } 620 621 if (call_again_soon) 622 goto out_resched; 623 mutex_unlock(&con->sock_mutex); 624 return 0; 625 626out_resched: 627 if (!test_and_set_bit(CF_READ_PENDING, &con->flags)) 628 queue_work(recv_workqueue, &con->rwork); 629 mutex_unlock(&con->sock_mutex); 630 return -EAGAIN; 631 632out_close: 633 mutex_unlock(&con->sock_mutex); 634 if (ret != -EAGAIN && !test_bit(CF_IS_OTHERCON, &con->flags)) { 635 close_connection(con, false); 636 /* Reconnect when there is something to send */ 637 } 638 /* Don't return success if we really got EOF */ 639 if (ret == 0) 640 ret = -EAGAIN; 641 642 return ret; 643} 644 645/* Listening socket is busy, accept a connection */ 646static int tcp_accept_from_sock(struct connection *con) 647{ 648 int result; 649 struct sockaddr_storage peeraddr; 650 struct socket *newsock; 651 int len; 652 int nodeid; 653 struct connection *newcon; 654 struct connection *addcon; 655 656 memset(&peeraddr, 0, sizeof(peeraddr)); 657 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM, 658 IPPROTO_TCP, &newsock); 659 if (result < 0) 660 return -ENOMEM; 661 662 mutex_lock_nested(&con->sock_mutex, 0); 663 664 result = -ENOTCONN; 665 if (con->sock == NULL) 666 goto accept_err; 667 668 newsock->type = con->sock->type; 669 newsock->ops = con->sock->ops; 670 671 result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK); 672 if (result < 0) 673 goto accept_err; 674 675 /* Get the connected socket's peer */ 676 memset(&peeraddr, 0, sizeof(peeraddr)); 677 if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr, 678 &len, 2)) { 679 result = -ECONNABORTED; 680 goto accept_err; 681 } 682 683 /* Get the new node's NODEID */ 684 make_sockaddr(&peeraddr, 0, &len); 685 if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) { 686 log_print("connect from non cluster node"); 687 sock_release(newsock); 688 mutex_unlock(&con->sock_mutex); 689 return -1; 690 } 691 692 log_print("got connection from %d", nodeid); 693 694 /* Check to see if we already have a connection to this node. This 695 * could happen if the two nodes initiate a connection at roughly 696 * the same time and the connections cross on the wire. 697 * In this case we store the incoming one in "othercon" 698 */ 699 newcon = nodeid2con(nodeid, GFP_KERNEL); 700 if (!newcon) { 701 result = -ENOMEM; 702 goto accept_err; 703 } 704 mutex_lock_nested(&newcon->sock_mutex, 1); 705 if (newcon->sock) { 706 struct connection *othercon = newcon->othercon; 707 708 if (!othercon) { 709 othercon = kmem_cache_zalloc(con_cache, GFP_KERNEL); 710 if (!othercon) { 711 log_print("failed to allocate incoming socket"); 712 mutex_unlock(&newcon->sock_mutex); 713 result = -ENOMEM; 714 goto accept_err; 715 } 716 othercon->nodeid = nodeid; 717 othercon->rx_action = receive_from_sock; 718 mutex_init(&othercon->sock_mutex); 719 INIT_WORK(&othercon->swork, process_send_sockets); 720 INIT_WORK(&othercon->rwork, process_recv_sockets); 721 set_bit(CF_IS_OTHERCON, &othercon->flags); 722 newcon->othercon = othercon; 723 } 724 othercon->sock = newsock; 725 newsock->sk->sk_user_data = othercon; 726 add_sock(newsock, othercon); 727 addcon = othercon; 728 } 729 else { 730 newsock->sk->sk_user_data = newcon; 731 newcon->rx_action = receive_from_sock; 732 add_sock(newsock, newcon); 733 addcon = newcon; 734 } 735 736 mutex_unlock(&newcon->sock_mutex); 737 738 /* 739 * Add it to the active queue in case we got data 740 * beween processing the accept adding the socket 741 * to the read_sockets list 742 */ 743 if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags)) 744 queue_work(recv_workqueue, &addcon->rwork); 745 mutex_unlock(&con->sock_mutex); 746 747 return 0; 748 749accept_err: 750 mutex_unlock(&con->sock_mutex); 751 sock_release(newsock); 752 753 if (result != -EAGAIN) 754 log_print("error accepting connection from node: %d", result); 755 return result; 756} 757 758static void free_entry(struct writequeue_entry *e) 759{ 760 __free_page(e->page); 761 kfree(e); 762} 763 764/* Initiate an SCTP association. 765 This is a special case of send_to_sock() in that we don't yet have a 766 peeled-off socket for this association, so we use the listening socket 767 and add the primary IP address of the remote node. 768 */ 769static void sctp_init_assoc(struct connection *con) 770{ 771 struct sockaddr_storage rem_addr; 772 char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))]; 773 struct msghdr outmessage; 774 struct cmsghdr *cmsg; 775 struct sctp_sndrcvinfo *sinfo; 776 struct connection *base_con; 777 struct writequeue_entry *e; 778 int len, offset; 779 int ret; 780 int addrlen; 781 struct kvec iov[1]; 782 783 if (test_and_set_bit(CF_INIT_PENDING, &con->flags)) 784 return; 785 786 if (con->retries++ > MAX_CONNECT_RETRIES) 787 return; 788 789 log_print("Initiating association with node %d", con->nodeid); 790 791 if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) { 792 log_print("no address for nodeid %d", con->nodeid); 793 return; 794 } 795 base_con = nodeid2con(0, 0); 796 BUG_ON(base_con == NULL); 797 798 make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen); 799 800 outmessage.msg_name = &rem_addr; 801 outmessage.msg_namelen = addrlen; 802 outmessage.msg_control = outcmsg; 803 outmessage.msg_controllen = sizeof(outcmsg); 804 outmessage.msg_flags = MSG_EOR; 805 806 spin_lock(&con->writequeue_lock); 807 e = list_entry(con->writequeue.next, struct writequeue_entry, 808 list); 809 810 BUG_ON((struct list_head *) e == &con->writequeue); 811 812 len = e->len; 813 offset = e->offset; 814 spin_unlock(&con->writequeue_lock); 815 kmap(e->page); 816 817 /* Send the first block off the write queue */ 818 iov[0].iov_base = page_address(e->page)+offset; 819 iov[0].iov_len = len; 820 821 cmsg = CMSG_FIRSTHDR(&outmessage); 822 cmsg->cmsg_level = IPPROTO_SCTP; 823 cmsg->cmsg_type = SCTP_SNDRCV; 824 cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo)); 825 sinfo = CMSG_DATA(cmsg); 826 memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo)); 827 sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid()); 828 outmessage.msg_controllen = cmsg->cmsg_len; 829 830 ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len); 831 if (ret < 0) { 832 log_print("Send first packet to node %d failed: %d", 833 con->nodeid, ret); 834 835 /* Try again later */ 836 clear_bit(CF_CONNECT_PENDING, &con->flags); 837 clear_bit(CF_INIT_PENDING, &con->flags); 838 } 839 else { 840 spin_lock(&con->writequeue_lock); 841 e->offset += ret; 842 e->len -= ret; 843 844 if (e->len == 0 && e->users == 0) { 845 list_del(&e->list); 846 kunmap(e->page); 847 free_entry(e); 848 } 849 spin_unlock(&con->writequeue_lock); 850 } 851} 852 853/* Connect a new socket to its peer */ 854static void tcp_connect_to_sock(struct connection *con) 855{ 856 int result = -EHOSTUNREACH; 857 struct sockaddr_storage saddr; 858 int addr_len; 859 struct socket *sock; 860 861 if (con->nodeid == 0) { 862 log_print("attempt to connect sock 0 foiled"); 863 return; 864 } 865 866 mutex_lock(&con->sock_mutex); 867 if (con->retries++ > MAX_CONNECT_RETRIES) 868 goto out; 869 870 /* Some odd races can cause double-connects, ignore them */ 871 if (con->sock) { 872 result = 0; 873 goto out; 874 } 875 876 /* Create a socket to communicate with */ 877 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM, 878 IPPROTO_TCP, &sock); 879 if (result < 0) 880 goto out_err; 881 882 memset(&saddr, 0, sizeof(saddr)); 883 if (dlm_nodeid_to_addr(con->nodeid, &saddr)) 884 goto out_err; 885 886 sock->sk->sk_user_data = con; 887 con->rx_action = receive_from_sock; 888 con->connect_action = tcp_connect_to_sock; 889 add_sock(sock, con); 890 891 make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len); 892 893 log_print("connecting to %d", con->nodeid); 894 result = 895 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len, 896 O_NONBLOCK); 897 if (result == -EINPROGRESS) 898 result = 0; 899 if (result == 0) 900 goto out; 901 902out_err: 903 if (con->sock) { 904 sock_release(con->sock); 905 con->sock = NULL; 906 } 907 /* 908 * Some errors are fatal and this list might need adjusting. For other 909 * errors we try again until the max number of retries is reached. 910 */ 911 if (result != -EHOSTUNREACH && result != -ENETUNREACH && 912 result != -ENETDOWN && result != EINVAL 913 && result != -EPROTONOSUPPORT) { 914 lowcomms_connect_sock(con); 915 result = 0; 916 } 917out: 918 mutex_unlock(&con->sock_mutex); 919 return; 920} 921 922static struct socket *tcp_create_listen_sock(struct connection *con, 923 struct sockaddr_storage *saddr) 924{ 925 struct socket *sock = NULL; 926 int result = 0; 927 int one = 1; 928 int addr_len; 929 930 if (dlm_local_addr[0]->ss_family == AF_INET) 931 addr_len = sizeof(struct sockaddr_in); 932 else 933 addr_len = sizeof(struct sockaddr_in6); 934 935 /* Create a socket to communicate with */ 936 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM, 937 IPPROTO_TCP, &sock); 938 if (result < 0) { 939 log_print("Can't create listening comms socket"); 940 goto create_out; 941 } 942 943 result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR, 944 (char *)&one, sizeof(one)); 945 946 if (result < 0) { 947 log_print("Failed to set SO_REUSEADDR on socket: %d", result); 948 } 949 sock->sk->sk_user_data = con; 950 con->rx_action = tcp_accept_from_sock; 951 con->connect_action = tcp_connect_to_sock; 952 con->sock = sock; 953 954 /* Bind to our port */ 955 make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len); 956 result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len); 957 if (result < 0) { 958 log_print("Can't bind to port %d", dlm_config.ci_tcp_port); 959 sock_release(sock); 960 sock = NULL; 961 con->sock = NULL; 962 goto create_out; 963 } 964 result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE, 965 (char *)&one, sizeof(one)); 966 if (result < 0) { 967 log_print("Set keepalive failed: %d", result); 968 } 969 970 result = sock->ops->listen(sock, 5); 971 if (result < 0) { 972 log_print("Can't listen on port %d", dlm_config.ci_tcp_port); 973 sock_release(sock); 974 sock = NULL; 975 goto create_out; 976 } 977 978create_out: 979 return sock; 980} 981 982/* Get local addresses */ 983static void init_local(void) 984{ 985 struct sockaddr_storage sas, *addr; 986 int i; 987 988 dlm_local_count = 0; 989 for (i = 0; i < DLM_MAX_ADDR_COUNT - 1; i++) { 990 if (dlm_our_addr(&sas, i)) 991 break; 992 993 addr = kmalloc(sizeof(*addr), GFP_KERNEL); 994 if (!addr) 995 break; 996 memcpy(addr, &sas, sizeof(*addr)); 997 dlm_local_addr[dlm_local_count++] = addr; 998 } 999} 1000 1001/* Bind to an IP address. SCTP allows multiple address so it can do 1002 multi-homing */ 1003static int add_sctp_bind_addr(struct connection *sctp_con, 1004 struct sockaddr_storage *addr, 1005 int addr_len, int num) 1006{ 1007 int result = 0; 1008 1009 if (num == 1) 1010 result = kernel_bind(sctp_con->sock, 1011 (struct sockaddr *) addr, 1012 addr_len); 1013 else 1014 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP, 1015 SCTP_SOCKOPT_BINDX_ADD, 1016 (char *)addr, addr_len); 1017 1018 if (result < 0) 1019 log_print("Can't bind to port %d addr number %d", 1020 dlm_config.ci_tcp_port, num); 1021 1022 return result; 1023} 1024 1025/* Initialise SCTP socket and bind to all interfaces */ 1026static int sctp_listen_for_all(void) 1027{ 1028 struct socket *sock = NULL; 1029 struct sockaddr_storage localaddr; 1030 struct sctp_event_subscribe subscribe; 1031 int result = -EINVAL, num = 1, i, addr_len; 1032 struct connection *con = nodeid2con(0, GFP_KERNEL); 1033 int bufsize = NEEDED_RMEM; 1034 1035 if (!con) 1036 return -ENOMEM; 1037 1038 log_print("Using SCTP for communications"); 1039 1040 result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET, 1041 IPPROTO_SCTP, &sock); 1042 if (result < 0) { 1043 log_print("Can't create comms socket, check SCTP is loaded"); 1044 goto out; 1045 } 1046 1047 /* Listen for events */ 1048 memset(&subscribe, 0, sizeof(subscribe)); 1049 subscribe.sctp_data_io_event = 1; 1050 subscribe.sctp_association_event = 1; 1051 subscribe.sctp_send_failure_event = 1; 1052 subscribe.sctp_shutdown_event = 1; 1053 subscribe.sctp_partial_delivery_event = 1; 1054 1055 result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUF, 1056 (char *)&bufsize, sizeof(bufsize)); 1057 if (result) 1058 log_print("Error increasing buffer space on socket %d", result); 1059 1060 result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS, 1061 (char *)&subscribe, sizeof(subscribe)); 1062 if (result < 0) { 1063 log_print("Failed to set SCTP_EVENTS on socket: result=%d", 1064 result); 1065 goto create_delsock; 1066 } 1067 1068 /* Init con struct */ 1069 sock->sk->sk_user_data = con; 1070 con->sock = sock; 1071 con->sock->sk->sk_data_ready = lowcomms_data_ready; 1072 con->rx_action = receive_from_sock; 1073 con->connect_action = sctp_init_assoc; 1074 1075 /* Bind to all interfaces. */ 1076 for (i = 0; i < dlm_local_count; i++) { 1077 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr)); 1078 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len); 1079 1080 result = add_sctp_bind_addr(con, &localaddr, addr_len, num); 1081 if (result) 1082 goto create_delsock; 1083 ++num; 1084 } 1085 1086 result = sock->ops->listen(sock, 5); 1087 if (result < 0) { 1088 log_print("Can't set socket listening"); 1089 goto create_delsock; 1090 } 1091 1092 return 0; 1093 1094create_delsock: 1095 sock_release(sock); 1096 con->sock = NULL; 1097out: 1098 return result; 1099} 1100 1101static int tcp_listen_for_all(void) 1102{ 1103 struct socket *sock = NULL; 1104 struct connection *con = nodeid2con(0, GFP_KERNEL); 1105 int result = -EINVAL; 1106 1107 if (!con) 1108 return -ENOMEM; 1109 1110 /* We don't support multi-homed hosts */ 1111 if (dlm_local_addr[1] != NULL) { 1112 log_print("TCP protocol can't handle multi-homed hosts, " 1113 "try SCTP"); 1114 return -EINVAL; 1115 } 1116 1117 log_print("Using TCP for communications"); 1118 1119 set_bit(CF_IS_OTHERCON, &con->flags); 1120 1121 sock = tcp_create_listen_sock(con, dlm_local_addr[0]); 1122 if (sock) { 1123 add_sock(sock, con); 1124 result = 0; 1125 } 1126 else { 1127 result = -EADDRINUSE; 1128 } 1129 1130 return result; 1131} 1132 1133 1134 1135static struct writequeue_entry *new_writequeue_entry(struct connection *con, 1136 gfp_t allocation) 1137{ 1138 struct writequeue_entry *entry; 1139 1140 entry = kmalloc(sizeof(struct writequeue_entry), allocation); 1141 if (!entry) 1142 return NULL; 1143 1144 entry->page = alloc_page(allocation); 1145 if (!entry->page) { 1146 kfree(entry); 1147 return NULL; 1148 } 1149 1150 entry->offset = 0; 1151 entry->len = 0; 1152 entry->end = 0; 1153 entry->users = 0; 1154 entry->con = con; 1155 1156 return entry; 1157} 1158 1159void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc) 1160{ 1161 struct connection *con; 1162 struct writequeue_entry *e; 1163 int offset = 0; 1164 int users = 0; 1165 1166 con = nodeid2con(nodeid, allocation); 1167 if (!con) 1168 return NULL; 1169 1170 spin_lock(&con->writequeue_lock); 1171 e = list_entry(con->writequeue.prev, struct writequeue_entry, list); 1172 if ((&e->list == &con->writequeue) || 1173 (PAGE_CACHE_SIZE - e->end < len)) { 1174 e = NULL; 1175 } else { 1176 offset = e->end; 1177 e->end += len; 1178 users = e->users++; 1179 } 1180 spin_unlock(&con->writequeue_lock); 1181 1182 if (e) { 1183 got_one: 1184 if (users == 0) 1185 kmap(e->page); 1186 *ppc = page_address(e->page) + offset; 1187 return e; 1188 } 1189 1190 e = new_writequeue_entry(con, allocation); 1191 if (e) { 1192 spin_lock(&con->writequeue_lock); 1193 offset = e->end; 1194 e->end += len; 1195 users = e->users++; 1196 list_add_tail(&e->list, &con->writequeue); 1197 spin_unlock(&con->writequeue_lock); 1198 goto got_one; 1199 } 1200 return NULL; 1201} 1202 1203void dlm_lowcomms_commit_buffer(void *mh) 1204{ 1205 struct writequeue_entry *e = (struct writequeue_entry *)mh; 1206 struct connection *con = e->con; 1207 int users; 1208 1209 spin_lock(&con->writequeue_lock); 1210 users = --e->users; 1211 if (users) 1212 goto out; 1213 e->len = e->end - e->offset; 1214 kunmap(e->page); 1215 spin_unlock(&con->writequeue_lock); 1216 1217 if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) { 1218 queue_work(send_workqueue, &con->swork); 1219 } 1220 return; 1221 1222out: 1223 spin_unlock(&con->writequeue_lock); 1224 return; 1225} 1226 1227/* Send a message */ 1228static void send_to_sock(struct connection *con) 1229{ 1230 int ret = 0; 1231 ssize_t(*sendpage) (struct socket *, struct page *, int, size_t, int); 1232 const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL; 1233 struct writequeue_entry *e; 1234 int len, offset; 1235 1236 mutex_lock(&con->sock_mutex); 1237 if (con->sock == NULL) 1238 goto out_connect; 1239 1240 sendpage = con->sock->ops->sendpage; 1241 1242 spin_lock(&con->writequeue_lock); 1243 for (;;) { 1244 e = list_entry(con->writequeue.next, struct writequeue_entry, 1245 list); 1246 if ((struct list_head *) e == &con->writequeue) 1247 break; 1248 1249 len = e->len; 1250 offset = e->offset; 1251 BUG_ON(len == 0 && e->users == 0); 1252 spin_unlock(&con->writequeue_lock); 1253 kmap(e->page); 1254 1255 ret = 0; 1256 if (len) { 1257 ret = sendpage(con->sock, e->page, offset, len, 1258 msg_flags); 1259 if (ret == -EAGAIN || ret == 0) 1260 goto out; 1261 if (ret <= 0) 1262 goto send_error; 1263 } else { 1264 /* Don't starve people filling buffers */ 1265 cond_resched(); 1266 } 1267 1268 spin_lock(&con->writequeue_lock); 1269 e->offset += ret; 1270 e->len -= ret; 1271 1272 if (e->len == 0 && e->users == 0) { 1273 list_del(&e->list); 1274 kunmap(e->page); 1275 free_entry(e); 1276 continue; 1277 } 1278 } 1279 spin_unlock(&con->writequeue_lock); 1280out: 1281 mutex_unlock(&con->sock_mutex); 1282 return; 1283 1284send_error: 1285 mutex_unlock(&con->sock_mutex); 1286 close_connection(con, false); 1287 lowcomms_connect_sock(con); 1288 return; 1289 1290out_connect: 1291 mutex_unlock(&con->sock_mutex); 1292 if (!test_bit(CF_INIT_PENDING, &con->flags)) 1293 lowcomms_connect_sock(con); 1294 return; 1295} 1296 1297static void clean_one_writequeue(struct connection *con) 1298{ 1299 struct list_head *list; 1300 struct list_head *temp; 1301 1302 spin_lock(&con->writequeue_lock); 1303 list_for_each_safe(list, temp, &con->writequeue) { 1304 struct writequeue_entry *e = 1305 list_entry(list, struct writequeue_entry, list); 1306 list_del(&e->list); 1307 free_entry(e); 1308 } 1309 spin_unlock(&con->writequeue_lock); 1310} 1311 1312/* Called from recovery when it knows that a node has 1313 left the cluster */ 1314int dlm_lowcomms_close(int nodeid) 1315{ 1316 struct connection *con; 1317 1318 log_print("closing connection to node %d", nodeid); 1319 con = nodeid2con(nodeid, 0); 1320 if (con) { 1321 clean_one_writequeue(con); 1322 close_connection(con, true); 1323 } 1324 return 0; 1325} 1326 1327/* Receive workqueue function */ 1328static void process_recv_sockets(struct work_struct *work) 1329{ 1330 struct connection *con = container_of(work, struct connection, rwork); 1331 int err; 1332 1333 clear_bit(CF_READ_PENDING, &con->flags); 1334 do { 1335 err = con->rx_action(con); 1336 } while (!err); 1337} 1338 1339/* Send workqueue function */ 1340static void process_send_sockets(struct work_struct *work) 1341{ 1342 struct connection *con = container_of(work, struct connection, swork); 1343 1344 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) { 1345 con->connect_action(con); 1346 } 1347 clear_bit(CF_WRITE_PENDING, &con->flags); 1348 send_to_sock(con); 1349} 1350 1351 1352/* Discard all entries on the write queues */ 1353static void clean_writequeues(void) 1354{ 1355 int nodeid; 1356 1357 for (nodeid = 1; nodeid <= max_nodeid; nodeid++) { 1358 struct connection *con = __nodeid2con(nodeid, 0); 1359 1360 if (con) 1361 clean_one_writequeue(con); 1362 } 1363} 1364 1365static void work_stop(void) 1366{ 1367 destroy_workqueue(recv_workqueue); 1368 destroy_workqueue(send_workqueue); 1369} 1370 1371static int work_start(void) 1372{ 1373 int error; 1374 recv_workqueue = create_workqueue("dlm_recv"); 1375 error = IS_ERR(recv_workqueue); 1376 if (error) { 1377 log_print("can't start dlm_recv %d", error); 1378 return error; 1379 } 1380 1381 send_workqueue = create_singlethread_workqueue("dlm_send"); 1382 error = IS_ERR(send_workqueue); 1383 if (error) { 1384 log_print("can't start dlm_send %d", error); 1385 destroy_workqueue(recv_workqueue); 1386 return error; 1387 } 1388 1389 return 0; 1390} 1391 1392void dlm_lowcomms_stop(void) 1393{ 1394 int i; 1395 struct connection *con; 1396 1397 /* Set all the flags to prevent any 1398 socket activity. 1399 */ 1400 down(&connections_lock); 1401 for (i = 0; i <= max_nodeid; i++) { 1402 con = __nodeid2con(i, 0); 1403 if (con) { 1404 con->flags |= 0xFF; 1405 if (con->sock) 1406 con->sock->sk->sk_user_data = NULL; 1407 } 1408 } 1409 up(&connections_lock); 1410 1411 work_stop(); 1412 1413 down(&connections_lock); 1414 clean_writequeues(); 1415 1416 for (i = 0; i <= max_nodeid; i++) { 1417 con = __nodeid2con(i, 0); 1418 if (con) { 1419 close_connection(con, true); 1420 if (con->othercon) 1421 kmem_cache_free(con_cache, con->othercon); 1422 kmem_cache_free(con_cache, con); 1423 } 1424 } 1425 max_nodeid = 0; 1426 up(&connections_lock); 1427 kmem_cache_destroy(con_cache); 1428 idr_init(&connections_idr); 1429} 1430 1431int dlm_lowcomms_start(void) 1432{ 1433 int error = -EINVAL; 1434 struct connection *con; 1435 1436 init_local(); 1437 if (!dlm_local_count) { 1438 error = -ENOTCONN; 1439 log_print("no local IP address has been set"); 1440 goto out; 1441 } 1442 1443 error = -ENOMEM; 1444 con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection), 1445 __alignof__(struct connection), 0, 1446 NULL, NULL); 1447 if (!con_cache) 1448 goto out; 1449 1450 /* Set some sysctl minima */ 1451 if (sysctl_rmem_max < NEEDED_RMEM) 1452 sysctl_rmem_max = NEEDED_RMEM; 1453 1454 /* Start listening */ 1455 if (dlm_config.ci_protocol == 0) 1456 error = tcp_listen_for_all(); 1457 else 1458 error = sctp_listen_for_all(); 1459 if (error) 1460 goto fail_unlisten; 1461 1462 error = work_start(); 1463 if (error) 1464 goto fail_unlisten; 1465 1466 return 0; 1467 1468fail_unlisten: 1469 con = nodeid2con(0,0); 1470 if (con) { 1471 close_connection(con, false); 1472 kmem_cache_free(con_cache, con); 1473 } 1474 kmem_cache_destroy(con_cache); 1475 1476out: 1477 return error; 1478} 1479