git.oblomov.eu Git - linux-2.6/blob - fs/dlm/lowcomms.c

   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
  63 struct cbuf {
  64         unsigned int base;
  65         unsigned int len;
  66         unsigned int mask;
  67 };
  68
  69 static void cbuf_add(struct cbuf *cb, int n)
  70 {
  71         cb->len += n;
  72 }
  73
  74 static int cbuf_data(struct cbuf *cb)
  75 {
  76         return ((cb->base + cb->len) & cb->mask);
  77 }
  78
  79 static void cbuf_init(struct cbuf *cb, int size)
  80 {
  81         cb->base = cb->len = 0;
  82         cb->mask = size-1;
  83 }
  84
  85 static void cbuf_eat(struct cbuf *cb, int n)
  86 {
  87         cb->len  -= n;
  88         cb->base += n;
  89         cb->base &= cb->mask;
  90 }
  91
  92 static bool cbuf_empty(struct cbuf *cb)
  93 {
  94         return cb->len == 0;
  95 }
  96
  97 struct 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 */
 123 struct 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
 133 static struct sockaddr_storage *dlm_local_addr[DLM_MAX_ADDR_COUNT];
 134 static int dlm_local_count;
 135
 136 /* Work queues */
 137 static struct workqueue_struct *recv_workqueue;
 138 static struct workqueue_struct *send_workqueue;
 139
 140 static DEFINE_IDR(connections_idr);
 141 static DECLARE_MUTEX(connections_lock);
 142 static int max_nodeid;
 143 static struct kmem_cache *con_cache;
 144
 145 static void process_recv_sockets(struct work_struct *work);
 146 static 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  */
 152 static 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
 204 static 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 */
 216 static 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
 233 static 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 */
 260 static 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
 267 static 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
 275 static 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
 281 static 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 */
 288 static 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 */
 302 static 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         memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 317 }
 318
 319 /* Close a remote connection and tidy up */
 320 static void close_connection(struct connection *con, bool and_other)
 321 {
 322         mutex_lock(&con->sock_mutex);
 323
 324         if (con->sock) {
 325                 sock_release(con->sock);
 326                 con->sock = NULL;
 327         }
 328         if (con->othercon && and_other) {
 329                 /* Will only re-enter once. */
 330                 close_connection(con->othercon, false);
 331         }
 332         if (con->rx_page) {
 333                 __free_page(con->rx_page);
 334                 con->rx_page = NULL;
 335         }
 336
 337         /* If we are an 'othercon' then NULL the pointer to us
 338            from the parent and tidy ourself up */
 339         if (test_bit(CF_IS_OTHERCON, &con->flags)) {
 340                 struct connection *parent = __nodeid2con(con->nodeid, 0);
 341                 parent->othercon = NULL;
 342                 kmem_cache_free(con_cache, con);
 343         }
 344         else {
 345                 /* Parent connections get reused */
 346                 con->retries = 0;
 347                 mutex_unlock(&con->sock_mutex);
 348         }
 349 }
 350
 351 /* We only send shutdown messages to nodes that are not part of the cluster */
 352 static void sctp_send_shutdown(sctp_assoc_t associd)
 353 {
 354         static char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 355         struct msghdr outmessage;
 356         struct cmsghdr *cmsg;
 357         struct sctp_sndrcvinfo *sinfo;
 358         int ret;
 359         struct connection *con;
 360
 361         con = nodeid2con(0,0);
 362         BUG_ON(con == NULL);
 363
 364         outmessage.msg_name = NULL;
 365         outmessage.msg_namelen = 0;
 366         outmessage.msg_control = outcmsg;
 367         outmessage.msg_controllen = sizeof(outcmsg);
 368         outmessage.msg_flags = MSG_EOR;
 369
 370         cmsg = CMSG_FIRSTHDR(&outmessage);
 371         cmsg->cmsg_level = IPPROTO_SCTP;
 372         cmsg->cmsg_type = SCTP_SNDRCV;
 373         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 374         outmessage.msg_controllen = cmsg->cmsg_len;
 375         sinfo = CMSG_DATA(cmsg);
 376         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 377
 378         sinfo->sinfo_flags |= MSG_EOF;
 379         sinfo->sinfo_assoc_id = associd;
 380
 381         ret = kernel_sendmsg(con->sock, &outmessage, NULL, 0, 0);
 382
 383         if (ret != 0)
 384                 log_print("send EOF to node failed: %d", ret);
 385 }
 386
 387 /* INIT failed but we don't know which node...
 388    restart INIT on all pending nodes */
 389 static void sctp_init_failed(void)
 390 {
 391         int i;
 392         struct connection *con;
 393
 394         down(&connections_lock);
 395         for (i=1; i<=max_nodeid; i++) {
 396                 con = __nodeid2con(i, 0);
 397                 if (!con)
 398                         continue;
 399                 con->sctp_assoc = 0;
 400                 if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
 401                         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
 402                                 queue_work(send_workqueue, &con->swork);
 403                         }
 404                 }
 405         }
 406         up(&connections_lock);
 407 }
 408
 409 /* Something happened to an association */
 410 static void process_sctp_notification(struct connection *con,
 411                                       struct msghdr *msg, char *buf)
 412 {
 413         union sctp_notification *sn = (union sctp_notification *)buf;
 414
 415         if (sn->sn_header.sn_type == SCTP_ASSOC_CHANGE) {
 416                 switch (sn->sn_assoc_change.sac_state) {
 417
 418                 case SCTP_COMM_UP:
 419                 case SCTP_RESTART:
 420                 {
 421                         /* Check that the new node is in the lockspace */
 422                         struct sctp_prim prim;
 423                         int nodeid;
 424                         int prim_len, ret;
 425                         int addr_len;
 426                         struct connection *new_con;
 427                         struct file *file;
 428                         sctp_peeloff_arg_t parg;
 429                         int parglen = sizeof(parg);
 430
 431                         /*
 432                          * We get this before any data for an association.
 433                          * We verify that the node is in the cluster and
 434                          * then peel off a socket for it.
 435                          */
 436                         if ((int)sn->sn_assoc_change.sac_assoc_id <= 0) {
 437                                 log_print("COMM_UP for invalid assoc ID %d",
 438                                          (int)sn->sn_assoc_change.sac_assoc_id);
 439                                 sctp_init_failed();
 440                                 return;
 441                         }
 442                         memset(&prim, 0, sizeof(struct sctp_prim));
 443                         prim_len = sizeof(struct sctp_prim);
 444                         prim.ssp_assoc_id = sn->sn_assoc_change.sac_assoc_id;
 445
 446                         ret = kernel_getsockopt(con->sock,
 447                                                 IPPROTO_SCTP,
 448                                                 SCTP_PRIMARY_ADDR,
 449                                                 (char*)&prim,
 450                                                 &prim_len);
 451                         if (ret < 0) {
 452                                 log_print("getsockopt/sctp_primary_addr on "
 453                                           "new assoc %d failed : %d",
 454                                           (int)sn->sn_assoc_change.sac_assoc_id,
 455                                           ret);
 456
 457                                 /* Retry INIT later */
 458                                 new_con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 459                                 if (new_con)
 460                                         clear_bit(CF_CONNECT_PENDING, &con->flags);
 461                                 return;
 462                         }
 463                         make_sockaddr(&prim.ssp_addr, 0, &addr_len);
 464                         if (dlm_addr_to_nodeid(&prim.ssp_addr, &nodeid)) {
 465                                 int i;
 466                                 unsigned char *b=(unsigned char *)&prim.ssp_addr;
 467                                 log_print("reject connect from unknown addr");
 468                                 for (i=0; i<sizeof(struct sockaddr_storage);i++)
 469                                         printk("%02x ", b[i]);
 470                                 printk("\n");
 471                                 sctp_send_shutdown(prim.ssp_assoc_id);
 472                                 return;
 473                         }
 474
 475                         new_con = nodeid2con(nodeid, GFP_KERNEL);
 476                         if (!new_con)
 477                                 return;
 478
 479                         /* Peel off a new sock */
 480                         parg.associd = sn->sn_assoc_change.sac_assoc_id;
 481                         ret = kernel_getsockopt(con->sock, IPPROTO_SCTP,
 482                                                 SCTP_SOCKOPT_PEELOFF,
 483                                                 (void *)&parg, &parglen);
 484                         if (ret) {
 485                                 log_print("Can't peel off a socket for "
 486                                           "connection %d to node %d: err=%d\n",
 487                                           parg.associd, nodeid, ret);
 488                         }
 489                         file = fget(parg.sd);
 490                         new_con->sock = SOCKET_I(file->f_dentry->d_inode);
 491                         add_sock(new_con->sock, new_con);
 492                         fput(file);
 493                         put_unused_fd(parg.sd);
 494
 495                         log_print("got new/restarted association %d nodeid %d",
 496                                  (int)sn->sn_assoc_change.sac_assoc_id, nodeid);
 497
 498                         /* Send any pending writes */
 499                         clear_bit(CF_CONNECT_PENDING, &new_con->flags);
 500                         clear_bit(CF_INIT_PENDING, &con->flags);
 501                         if (!test_and_set_bit(CF_WRITE_PENDING, &new_con->flags)) {
 502                                 queue_work(send_workqueue, &new_con->swork);
 503                         }
 504                         if (!test_and_set_bit(CF_READ_PENDING, &new_con->flags))
 505                                 queue_work(recv_workqueue, &new_con->rwork);
 506                 }
 507                 break;
 508
 509                 case SCTP_COMM_LOST:
 510                 case SCTP_SHUTDOWN_COMP:
 511                 {
 512                         con = assoc2con(sn->sn_assoc_change.sac_assoc_id);
 513                         if (con) {
 514                                 con->sctp_assoc = 0;
 515                         }
 516                 }
 517                 break;
 518
 519                 /* We don't know which INIT failed, so clear the PENDING flags
 520                  * on them all.  if assoc_id is zero then it will then try
 521                  * again */
 522
 523                 case SCTP_CANT_STR_ASSOC:
 524                 {
 525                         log_print("Can't start SCTP association - retrying");
 526                         sctp_init_failed();
 527                 }
 528                 break;
 529
 530                 default:
 531                         log_print("unexpected SCTP assoc change id=%d state=%d",
 532                                   (int)sn->sn_assoc_change.sac_assoc_id,
 533                                   sn->sn_assoc_change.sac_state);
 534                 }
 535         }
 536 }
 537
 538 /* Data received from remote end */
 539 static int receive_from_sock(struct connection *con)
 540 {
 541         int ret = 0;
 542         struct msghdr msg = {};
 543         struct kvec iov[2];
 544         unsigned len;
 545         int r;
 546         int call_again_soon = 0;
 547         int nvec;
 548         char incmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 549
 550         mutex_lock(&con->sock_mutex);
 551
 552         if (con->sock == NULL) {
 553                 ret = -EAGAIN;
 554                 goto out_close;
 555         }
 556
 557         if (con->rx_page == NULL) {
 558                 /*
 559                  * This doesn't need to be atomic, but I think it should
 560                  * improve performance if it is.
 561                  */
 562                 con->rx_page = alloc_page(GFP_ATOMIC);
 563                 if (con->rx_page == NULL)
 564                         goto out_resched;
 565                 cbuf_init(&con->cb, PAGE_CACHE_SIZE);
 566         }
 567
 568         /* Only SCTP needs these really */
 569         memset(&incmsg, 0, sizeof(incmsg));
 570         msg.msg_control = incmsg;
 571         msg.msg_controllen = sizeof(incmsg);
 572
 573         /*
 574          * iov[0] is the bit of the circular buffer between the current end
 575          * point (cb.base + cb.len) and the end of the buffer.
 576          */
 577         iov[0].iov_len = con->cb.base - cbuf_data(&con->cb);
 578         iov[0].iov_base = page_address(con->rx_page) + cbuf_data(&con->cb);
 579         iov[1].iov_len = 0;
 580         nvec = 1;
 581
 582         /*
 583          * iov[1] is the bit of the circular buffer between the start of the
 584          * buffer and the start of the currently used section (cb.base)
 585          */
 586         if (cbuf_data(&con->cb) >= con->cb.base) {
 587                 iov[0].iov_len = PAGE_CACHE_SIZE - cbuf_data(&con->cb);
 588                 iov[1].iov_len = con->cb.base;
 589                 iov[1].iov_base = page_address(con->rx_page);
 590                 nvec = 2;
 591         }
 592         len = iov[0].iov_len + iov[1].iov_len;
 593
 594         r = ret = kernel_recvmsg(con->sock, &msg, iov, nvec, len,
 595                                MSG_DONTWAIT | MSG_NOSIGNAL);
 596         if (ret <= 0)
 597                 goto out_close;
 598
 599         /* Process SCTP notifications */
 600         if (msg.msg_flags & MSG_NOTIFICATION) {
 601                 msg.msg_control = incmsg;
 602                 msg.msg_controllen = sizeof(incmsg);
 603
 604                 process_sctp_notification(con, &msg,
 605                                 page_address(con->rx_page) + con->cb.base);
 606                 mutex_unlock(&con->sock_mutex);
 607                 return 0;
 608         }
 609         BUG_ON(con->nodeid == 0);
 610
 611         if (ret == len)
 612                 call_again_soon = 1;
 613         cbuf_add(&con->cb, ret);
 614         ret = dlm_process_incoming_buffer(con->nodeid,
 615                                           page_address(con->rx_page),
 616                                           con->cb.base, con->cb.len,
 617                                           PAGE_CACHE_SIZE);
 618         if (ret == -EBADMSG) {
 619                 log_print("lowcomms: addr=%p, base=%u, len=%u, "
 620                           "iov_len=%u, iov_base[0]=%p, read=%d",
 621                           page_address(con->rx_page), con->cb.base, con->cb.len,
 622                           len, iov[0].iov_base, r);
 623         }
 624         if (ret < 0)
 625                 goto out_close;
 626         cbuf_eat(&con->cb, ret);
 627
 628         if (cbuf_empty(&con->cb) && !call_again_soon) {
 629                 __free_page(con->rx_page);
 630                 con->rx_page = NULL;
 631         }
 632
 633         if (call_again_soon)
 634                 goto out_resched;
 635         mutex_unlock(&con->sock_mutex);
 636         return 0;
 637
 638 out_resched:
 639         if (!test_and_set_bit(CF_READ_PENDING, &con->flags))
 640                 queue_work(recv_workqueue, &con->rwork);
 641         mutex_unlock(&con->sock_mutex);
 642         return -EAGAIN;
 643
 644 out_close:
 645         mutex_unlock(&con->sock_mutex);
 646         if (ret != -EAGAIN) {
 647                 close_connection(con, false);
 648                 /* Reconnect when there is something to send */
 649         }
 650         /* Don't return success if we really got EOF */
 651         if (ret == 0)
 652                 ret = -EAGAIN;
 653
 654         return ret;
 655 }
 656
 657 /* Listening socket is busy, accept a connection */
 658 static int tcp_accept_from_sock(struct connection *con)
 659 {
 660         int result;
 661         struct sockaddr_storage peeraddr;
 662         struct socket *newsock;
 663         int len;
 664         int nodeid;
 665         struct connection *newcon;
 666         struct connection *addcon;
 667
 668         memset(&peeraddr, 0, sizeof(peeraddr));
 669         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 670                                   IPPROTO_TCP, &newsock);
 671         if (result < 0)
 672                 return -ENOMEM;
 673
 674         mutex_lock_nested(&con->sock_mutex, 0);
 675
 676         result = -ENOTCONN;
 677         if (con->sock == NULL)
 678                 goto accept_err;
 679
 680         newsock->type = con->sock->type;
 681         newsock->ops = con->sock->ops;
 682
 683         result = con->sock->ops->accept(con->sock, newsock, O_NONBLOCK);
 684         if (result < 0)
 685                 goto accept_err;
 686
 687         /* Get the connected socket's peer */
 688         memset(&peeraddr, 0, sizeof(peeraddr));
 689         if (newsock->ops->getname(newsock, (struct sockaddr *)&peeraddr,
 690                                   &len, 2)) {
 691                 result = -ECONNABORTED;
 692                 goto accept_err;
 693         }
 694
 695         /* Get the new node's NODEID */
 696         make_sockaddr(&peeraddr, 0, &len);
 697         if (dlm_addr_to_nodeid(&peeraddr, &nodeid)) {
 698                 log_print("connect from non cluster node");
 699                 sock_release(newsock);
 700                 mutex_unlock(&con->sock_mutex);
 701                 return -1;
 702         }
 703
 704         log_print("got connection from %d", nodeid);
 705
 706         /*  Check to see if we already have a connection to this node. This
 707          *  could happen if the two nodes initiate a connection at roughly
 708          *  the same time and the connections cross on the wire.
 709          *  In this case we store the incoming one in "othercon"
 710          */
 711         newcon = nodeid2con(nodeid, GFP_KERNEL);
 712         if (!newcon) {
 713                 result = -ENOMEM;
 714                 goto accept_err;
 715         }
 716         mutex_lock_nested(&newcon->sock_mutex, 1);
 717         if (newcon->sock) {
 718                 struct connection *othercon = newcon->othercon;
 719
 720                 if (!othercon) {
 721                         othercon = kmem_cache_zalloc(con_cache, GFP_KERNEL);
 722                         if (!othercon) {
 723                                 log_print("failed to allocate incoming socket");
 724                                 mutex_unlock(&newcon->sock_mutex);
 725                                 result = -ENOMEM;
 726                                 goto accept_err;
 727                         }
 728                         othercon->nodeid = nodeid;
 729                         othercon->rx_action = receive_from_sock;
 730                         mutex_init(&othercon->sock_mutex);
 731                         INIT_WORK(&othercon->swork, process_send_sockets);
 732                         INIT_WORK(&othercon->rwork, process_recv_sockets);
 733                         set_bit(CF_IS_OTHERCON, &othercon->flags);
 734                         newcon->othercon = othercon;
 735                         othercon->sock = newsock;
 736                         newsock->sk->sk_user_data = othercon;
 737                         add_sock(newsock, othercon);
 738                         addcon = othercon;
 739                 }
 740                 else {
 741                         printk("Extra connection from node %d attempted\n", nodeid);
 742                         result = -EAGAIN;
 743                         mutex_unlock(&newcon->sock_mutex);
 744                         goto accept_err;
 745                 }
 746         }
 747         else {
 748                 newsock->sk->sk_user_data = newcon;
 749                 newcon->rx_action = receive_from_sock;
 750                 add_sock(newsock, newcon);
 751                 addcon = newcon;
 752         }
 753
 754         mutex_unlock(&newcon->sock_mutex);
 755
 756         /*
 757          * Add it to the active queue in case we got data
 758          * beween processing the accept adding the socket
 759          * to the read_sockets list
 760          */
 761         if (!test_and_set_bit(CF_READ_PENDING, &addcon->flags))
 762                 queue_work(recv_workqueue, &addcon->rwork);
 763         mutex_unlock(&con->sock_mutex);
 764
 765         return 0;
 766
 767 accept_err:
 768         mutex_unlock(&con->sock_mutex);
 769         sock_release(newsock);
 770
 771         if (result != -EAGAIN)
 772                 log_print("error accepting connection from node: %d", result);
 773         return result;
 774 }
 775
 776 static void free_entry(struct writequeue_entry *e)
 777 {
 778         __free_page(e->page);
 779         kfree(e);
 780 }
 781
 782 /* Initiate an SCTP association.
 783    This is a special case of send_to_sock() in that we don't yet have a
 784    peeled-off socket for this association, so we use the listening socket
 785    and add the primary IP address of the remote node.
 786  */
 787 static void sctp_init_assoc(struct connection *con)
 788 {
 789         struct sockaddr_storage rem_addr;
 790         char outcmsg[CMSG_SPACE(sizeof(struct sctp_sndrcvinfo))];
 791         struct msghdr outmessage;
 792         struct cmsghdr *cmsg;
 793         struct sctp_sndrcvinfo *sinfo;
 794         struct connection *base_con;
 795         struct writequeue_entry *e;
 796         int len, offset;
 797         int ret;
 798         int addrlen;
 799         struct kvec iov[1];
 800
 801         if (test_and_set_bit(CF_INIT_PENDING, &con->flags))
 802                 return;
 803
 804         if (con->retries++ > MAX_CONNECT_RETRIES)
 805                 return;
 806
 807         log_print("Initiating association with node %d", con->nodeid);
 808
 809         if (nodeid_to_addr(con->nodeid, (struct sockaddr *)&rem_addr)) {
 810                 log_print("no address for nodeid %d", con->nodeid);
 811                 return;
 812         }
 813         base_con = nodeid2con(0, 0);
 814         BUG_ON(base_con == NULL);
 815
 816         make_sockaddr(&rem_addr, dlm_config.ci_tcp_port, &addrlen);
 817
 818         outmessage.msg_name = &rem_addr;
 819         outmessage.msg_namelen = addrlen;
 820         outmessage.msg_control = outcmsg;
 821         outmessage.msg_controllen = sizeof(outcmsg);
 822         outmessage.msg_flags = MSG_EOR;
 823
 824         spin_lock(&con->writequeue_lock);
 825         e = list_entry(con->writequeue.next, struct writequeue_entry,
 826                        list);
 827
 828         BUG_ON((struct list_head *) e == &con->writequeue);
 829
 830         len = e->len;
 831         offset = e->offset;
 832         spin_unlock(&con->writequeue_lock);
 833         kmap(e->page);
 834
 835         /* Send the first block off the write queue */
 836         iov[0].iov_base = page_address(e->page)+offset;
 837         iov[0].iov_len = len;
 838
 839         cmsg = CMSG_FIRSTHDR(&outmessage);
 840         cmsg->cmsg_level = IPPROTO_SCTP;
 841         cmsg->cmsg_type = SCTP_SNDRCV;
 842         cmsg->cmsg_len = CMSG_LEN(sizeof(struct sctp_sndrcvinfo));
 843         sinfo = CMSG_DATA(cmsg);
 844         memset(sinfo, 0x00, sizeof(struct sctp_sndrcvinfo));
 845         sinfo->sinfo_ppid = cpu_to_le32(dlm_our_nodeid());
 846         outmessage.msg_controllen = cmsg->cmsg_len;
 847
 848         ret = kernel_sendmsg(base_con->sock, &outmessage, iov, 1, len);
 849         if (ret < 0) {
 850                 log_print("Send first packet to node %d failed: %d",
 851                           con->nodeid, ret);
 852
 853                 /* Try again later */
 854                 clear_bit(CF_CONNECT_PENDING, &con->flags);
 855                 clear_bit(CF_INIT_PENDING, &con->flags);
 856         }
 857         else {
 858                 spin_lock(&con->writequeue_lock);
 859                 e->offset += ret;
 860                 e->len -= ret;
 861
 862                 if (e->len == 0 && e->users == 0) {
 863                         list_del(&e->list);
 864                         kunmap(e->page);
 865                         free_entry(e);
 866                 }
 867                 spin_unlock(&con->writequeue_lock);
 868         }
 869 }
 870
 871 /* Connect a new socket to its peer */
 872 static void tcp_connect_to_sock(struct connection *con)
 873 {
 874         int result = -EHOSTUNREACH;
 875         struct sockaddr_storage saddr;
 876         int addr_len;
 877         struct socket *sock;
 878
 879         if (con->nodeid == 0) {
 880                 log_print("attempt to connect sock 0 foiled");
 881                 return;
 882         }
 883
 884         mutex_lock(&con->sock_mutex);
 885         if (con->retries++ > MAX_CONNECT_RETRIES)
 886                 goto out;
 887
 888         /* Some odd races can cause double-connects, ignore them */
 889         if (con->sock) {
 890                 result = 0;
 891                 goto out;
 892         }
 893
 894         /* Create a socket to communicate with */
 895         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 896                                   IPPROTO_TCP, &sock);
 897         if (result < 0)
 898                 goto out_err;
 899
 900         memset(&saddr, 0, sizeof(saddr));
 901         if (dlm_nodeid_to_addr(con->nodeid, &saddr))
 902                 goto out_err;
 903
 904         sock->sk->sk_user_data = con;
 905         con->rx_action = receive_from_sock;
 906         con->connect_action = tcp_connect_to_sock;
 907         add_sock(sock, con);
 908
 909         make_sockaddr(&saddr, dlm_config.ci_tcp_port, &addr_len);
 910
 911         log_print("connecting to %d", con->nodeid);
 912         result =
 913                 sock->ops->connect(sock, (struct sockaddr *)&saddr, addr_len,
 914                                    O_NONBLOCK);
 915         if (result == -EINPROGRESS)
 916                 result = 0;
 917         if (result == 0)
 918                 goto out;
 919
 920 out_err:
 921         if (con->sock) {
 922                 sock_release(con->sock);
 923                 con->sock = NULL;
 924         }
 925         /*
 926          * Some errors are fatal and this list might need adjusting. For other
 927          * errors we try again until the max number of retries is reached.
 928          */
 929         if (result != -EHOSTUNREACH && result != -ENETUNREACH &&
 930             result != -ENETDOWN && result != EINVAL
 931             && result != -EPROTONOSUPPORT) {
 932                 lowcomms_connect_sock(con);
 933                 result = 0;
 934         }
 935 out:
 936         mutex_unlock(&con->sock_mutex);
 937         return;
 938 }
 939
 940 static struct socket *tcp_create_listen_sock(struct connection *con,
 941                                              struct sockaddr_storage *saddr)
 942 {
 943         struct socket *sock = NULL;
 944         int result = 0;
 945         int one = 1;
 946         int addr_len;
 947
 948         if (dlm_local_addr[0]->ss_family == AF_INET)
 949                 addr_len = sizeof(struct sockaddr_in);
 950         else
 951                 addr_len = sizeof(struct sockaddr_in6);
 952
 953         /* Create a socket to communicate with */
 954         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_STREAM,
 955                                   IPPROTO_TCP, &sock);
 956         if (result < 0) {
 957                 log_print("Can't create listening comms socket");
 958                 goto create_out;
 959         }
 960
 961         result = kernel_setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
 962                                    (char *)&one, sizeof(one));
 963
 964         if (result < 0) {
 965                 log_print("Failed to set SO_REUSEADDR on socket: %d", result);
 966         }
 967         sock->sk->sk_user_data = con;
 968         con->rx_action = tcp_accept_from_sock;
 969         con->connect_action = tcp_connect_to_sock;
 970         con->sock = sock;
 971
 972         /* Bind to our port */
 973         make_sockaddr(saddr, dlm_config.ci_tcp_port, &addr_len);
 974         result = sock->ops->bind(sock, (struct sockaddr *) saddr, addr_len);
 975         if (result < 0) {
 976                 log_print("Can't bind to port %d", dlm_config.ci_tcp_port);
 977                 sock_release(sock);
 978                 sock = NULL;
 979                 con->sock = NULL;
 980                 goto create_out;
 981         }
 982         result = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE,
 983                                  (char *)&one, sizeof(one));
 984         if (result < 0) {
 985                 log_print("Set keepalive failed: %d", result);
 986         }
 987
 988         result = sock->ops->listen(sock, 5);
 989         if (result < 0) {
 990                 log_print("Can't listen on port %d", dlm_config.ci_tcp_port);
 991                 sock_release(sock);
 992                 sock = NULL;
 993                 goto create_out;
 994         }
 995
 996 create_out:
 997         return sock;
 998 }
 999
1000 /* Get local addresses */
1001 static void init_local(void)
1002 {
1003         struct sockaddr_storage sas, *addr;
1004         int i;
1005
1006         dlm_local_count = 0;
1007         for (i = 0; i < DLM_MAX_ADDR_COUNT - 1; i++) {
1008                 if (dlm_our_addr(&sas, i))
1009                         break;
1010
1011                 addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1012                 if (!addr)
1013                         break;
1014                 memcpy(addr, &sas, sizeof(*addr));
1015                 dlm_local_addr[dlm_local_count++] = addr;
1016         }
1017 }
1018
1019 /* Bind to an IP address. SCTP allows multiple address so it can do
1020    multi-homing */
1021 static int add_sctp_bind_addr(struct connection *sctp_con,
1022                               struct sockaddr_storage *addr,
1023                               int addr_len, int num)
1024 {
1025         int result = 0;
1026
1027         if (num == 1)
1028                 result = kernel_bind(sctp_con->sock,
1029                                      (struct sockaddr *) addr,
1030                                      addr_len);
1031         else
1032                 result = kernel_setsockopt(sctp_con->sock, SOL_SCTP,
1033                                            SCTP_SOCKOPT_BINDX_ADD,
1034                                            (char *)addr, addr_len);
1035
1036         if (result < 0)
1037                 log_print("Can't bind to port %d addr number %d",
1038                           dlm_config.ci_tcp_port, num);
1039
1040         return result;
1041 }
1042
1043 /* Initialise SCTP socket and bind to all interfaces */
1044 static int sctp_listen_for_all(void)
1045 {
1046         struct socket *sock = NULL;
1047         struct sockaddr_storage localaddr;
1048         struct sctp_event_subscribe subscribe;
1049         int result = -EINVAL, num = 1, i, addr_len;
1050         struct connection *con = nodeid2con(0, GFP_KERNEL);
1051         int bufsize = NEEDED_RMEM;
1052
1053         if (!con)
1054                 return -ENOMEM;
1055
1056         log_print("Using SCTP for communications");
1057
1058         result = sock_create_kern(dlm_local_addr[0]->ss_family, SOCK_SEQPACKET,
1059                                   IPPROTO_SCTP, &sock);
1060         if (result < 0) {
1061                 log_print("Can't create comms socket, check SCTP is loaded");
1062                 goto out;
1063         }
1064
1065         /* Listen for events */
1066         memset(&subscribe, 0, sizeof(subscribe));
1067         subscribe.sctp_data_io_event = 1;
1068         subscribe.sctp_association_event = 1;
1069         subscribe.sctp_send_failure_event = 1;
1070         subscribe.sctp_shutdown_event = 1;
1071         subscribe.sctp_partial_delivery_event = 1;
1072
1073         result = kernel_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
1074                                  (char *)&bufsize, sizeof(bufsize));
1075         if (result)
1076                 log_print("Error increasing buffer space on socket %d", result);
1077
1078         result = kernel_setsockopt(sock, SOL_SCTP, SCTP_EVENTS,
1079                                    (char *)&subscribe, sizeof(subscribe));
1080         if (result < 0) {
1081                 log_print("Failed to set SCTP_EVENTS on socket: result=%d",
1082                           result);
1083                 goto create_delsock;
1084         }
1085
1086         /* Init con struct */
1087         sock->sk->sk_user_data = con;
1088         con->sock = sock;
1089         con->sock->sk->sk_data_ready = lowcomms_data_ready;
1090         con->rx_action = receive_from_sock;
1091         con->connect_action = sctp_init_assoc;
1092
1093         /* Bind to all interfaces. */
1094         for (i = 0; i < dlm_local_count; i++) {
1095                 memcpy(&localaddr, dlm_local_addr[i], sizeof(localaddr));
1096                 make_sockaddr(&localaddr, dlm_config.ci_tcp_port, &addr_len);
1097
1098                 result = add_sctp_bind_addr(con, &localaddr, addr_len, num);
1099                 if (result)
1100                         goto create_delsock;
1101                 ++num;
1102         }
1103
1104         result = sock->ops->listen(sock, 5);
1105         if (result < 0) {
1106                 log_print("Can't set socket listening");
1107                 goto create_delsock;
1108         }
1109
1110         return 0;
1111
1112 create_delsock:
1113         sock_release(sock);
1114         con->sock = NULL;
1115 out:
1116         return result;
1117 }
1118
1119 static int tcp_listen_for_all(void)
1120 {
1121         struct socket *sock = NULL;
1122         struct connection *con = nodeid2con(0, GFP_KERNEL);
1123         int result = -EINVAL;
1124
1125         if (!con)
1126                 return -ENOMEM;
1127
1128         /* We don't support multi-homed hosts */
1129         if (dlm_local_addr[1] != NULL) {
1130                 log_print("TCP protocol can't handle multi-homed hosts, "
1131                           "try SCTP");
1132                 return -EINVAL;
1133         }
1134
1135         log_print("Using TCP for communications");
1136
1137         sock = tcp_create_listen_sock(con, dlm_local_addr[0]);
1138         if (sock) {
1139                 add_sock(sock, con);
1140                 result = 0;
1141         }
1142         else {
1143                 result = -EADDRINUSE;
1144         }
1145
1146         return result;
1147 }
1148
1149
1150
1151 static struct writequeue_entry *new_writequeue_entry(struct connection *con,
1152                                                      gfp_t allocation)
1153 {
1154         struct writequeue_entry *entry;
1155
1156         entry = kmalloc(sizeof(struct writequeue_entry), allocation);
1157         if (!entry)
1158                 return NULL;
1159
1160         entry->page = alloc_page(allocation);
1161         if (!entry->page) {
1162                 kfree(entry);
1163                 return NULL;
1164         }
1165
1166         entry->offset = 0;
1167         entry->len = 0;
1168         entry->end = 0;
1169         entry->users = 0;
1170         entry->con = con;
1171
1172         return entry;
1173 }
1174
1175 void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
1176 {
1177         struct connection *con;
1178         struct writequeue_entry *e;
1179         int offset = 0;
1180         int users = 0;
1181
1182         con = nodeid2con(nodeid, allocation);
1183         if (!con)
1184                 return NULL;
1185
1186         spin_lock(&con->writequeue_lock);
1187         e = list_entry(con->writequeue.prev, struct writequeue_entry, list);
1188         if ((&e->list == &con->writequeue) ||
1189             (PAGE_CACHE_SIZE - e->end < len)) {
1190                 e = NULL;
1191         } else {
1192                 offset = e->end;
1193                 e->end += len;
1194                 users = e->users++;
1195         }
1196         spin_unlock(&con->writequeue_lock);
1197
1198         if (e) {
1199         got_one:
1200                 if (users == 0)
1201                         kmap(e->page);
1202                 *ppc = page_address(e->page) + offset;
1203                 return e;
1204         }
1205
1206         e = new_writequeue_entry(con, allocation);
1207         if (e) {
1208                 spin_lock(&con->writequeue_lock);
1209                 offset = e->end;
1210                 e->end += len;
1211                 users = e->users++;
1212                 list_add_tail(&e->list, &con->writequeue);
1213                 spin_unlock(&con->writequeue_lock);
1214                 goto got_one;
1215         }
1216         return NULL;
1217 }
1218
1219 void dlm_lowcomms_commit_buffer(void *mh)
1220 {
1221         struct writequeue_entry *e = (struct writequeue_entry *)mh;
1222         struct connection *con = e->con;
1223         int users;
1224
1225         spin_lock(&con->writequeue_lock);
1226         users = --e->users;
1227         if (users)
1228                 goto out;
1229         e->len = e->end - e->offset;
1230         kunmap(e->page);
1231         spin_unlock(&con->writequeue_lock);
1232
1233         if (!test_and_set_bit(CF_WRITE_PENDING, &con->flags)) {
1234                 queue_work(send_workqueue, &con->swork);
1235         }
1236         return;
1237
1238 out:
1239         spin_unlock(&con->writequeue_lock);
1240         return;
1241 }
1242
1243 /* Send a message */
1244 static void send_to_sock(struct connection *con)
1245 {
1246         int ret = 0;
1247         ssize_t(*sendpage) (struct socket *, struct page *, int, size_t, int);
1248         const int msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL;
1249         struct writequeue_entry *e;
1250         int len, offset;
1251
1252         mutex_lock(&con->sock_mutex);
1253         if (con->sock == NULL)
1254                 goto out_connect;
1255
1256         sendpage = con->sock->ops->sendpage;
1257
1258         spin_lock(&con->writequeue_lock);
1259         for (;;) {
1260                 e = list_entry(con->writequeue.next, struct writequeue_entry,
1261                                list);
1262                 if ((struct list_head *) e == &con->writequeue)
1263                         break;
1264
1265                 len = e->len;
1266                 offset = e->offset;
1267                 BUG_ON(len == 0 && e->users == 0);
1268                 spin_unlock(&con->writequeue_lock);
1269                 kmap(e->page);
1270
1271                 ret = 0;
1272                 if (len) {
1273                         ret = sendpage(con->sock, e->page, offset, len,
1274                                        msg_flags);
1275                         if (ret == -EAGAIN || ret == 0)
1276                                 goto out;
1277                         if (ret <= 0)
1278                                 goto send_error;
1279                 } else {
1280                         /* Don't starve people filling buffers */
1281                         cond_resched();
1282                 }
1283
1284                 spin_lock(&con->writequeue_lock);
1285                 e->offset += ret;
1286                 e->len -= ret;
1287
1288                 if (e->len == 0 && e->users == 0) {
1289                         list_del(&e->list);
1290                         kunmap(e->page);
1291                         free_entry(e);
1292                         continue;
1293                 }
1294         }
1295         spin_unlock(&con->writequeue_lock);
1296 out:
1297         mutex_unlock(&con->sock_mutex);
1298         return;
1299
1300 send_error:
1301         mutex_unlock(&con->sock_mutex);
1302         close_connection(con, false);
1303         lowcomms_connect_sock(con);
1304         return;
1305
1306 out_connect:
1307         mutex_unlock(&con->sock_mutex);
1308         if (!test_bit(CF_INIT_PENDING, &con->flags))
1309                 lowcomms_connect_sock(con);
1310         return;
1311 }
1312
1313 static void clean_one_writequeue(struct connection *con)
1314 {
1315         struct list_head *list;
1316         struct list_head *temp;
1317
1318         spin_lock(&con->writequeue_lock);
1319         list_for_each_safe(list, temp, &con->writequeue) {
1320                 struct writequeue_entry *e =
1321                         list_entry(list, struct writequeue_entry, list);
1322                 list_del(&e->list);
1323                 free_entry(e);
1324         }
1325         spin_unlock(&con->writequeue_lock);
1326 }
1327
1328 /* Called from recovery when it knows that a node has
1329    left the cluster */
1330 int dlm_lowcomms_close(int nodeid)
1331 {
1332         struct connection *con;
1333
1334         log_print("closing connection to node %d", nodeid);
1335         con = nodeid2con(nodeid, 0);
1336         if (con) {
1337                 clean_one_writequeue(con);
1338                 close_connection(con, true);
1339         }
1340         return 0;
1341 }
1342
1343 /* Receive workqueue function */
1344 static void process_recv_sockets(struct work_struct *work)
1345 {
1346         struct connection *con = container_of(work, struct connection, rwork);
1347         int err;
1348
1349         clear_bit(CF_READ_PENDING, &con->flags);
1350         do {
1351                 err = con->rx_action(con);
1352         } while (!err);
1353 }
1354
1355 /* Send workqueue function */
1356 static void process_send_sockets(struct work_struct *work)
1357 {
1358         struct connection *con = container_of(work, struct connection, swork);
1359
1360         if (test_and_clear_bit(CF_CONNECT_PENDING, &con->flags)) {
1361                 con->connect_action(con);
1362         }
1363         clear_bit(CF_WRITE_PENDING, &con->flags);
1364         send_to_sock(con);
1365 }
1366
1367
1368 /* Discard all entries on the write queues */
1369 static void clean_writequeues(void)
1370 {
1371         int nodeid;
1372
1373         for (nodeid = 1; nodeid <= max_nodeid; nodeid++) {
1374                 struct connection *con = __nodeid2con(nodeid, 0);
1375
1376                 if (con)
1377                         clean_one_writequeue(con);
1378         }
1379 }
1380
1381 static void work_stop(void)
1382 {
1383         destroy_workqueue(recv_workqueue);
1384         destroy_workqueue(send_workqueue);
1385 }
1386
1387 static int work_start(void)
1388 {
1389         int error;
1390         recv_workqueue = create_workqueue("dlm_recv");
1391         error = IS_ERR(recv_workqueue);
1392         if (error) {
1393                 log_print("can't start dlm_recv %d", error);
1394                 return error;
1395         }
1396
1397         send_workqueue = create_singlethread_workqueue("dlm_send");
1398         error = IS_ERR(send_workqueue);
1399         if (error) {
1400                 log_print("can't start dlm_send %d", error);
1401                 destroy_workqueue(recv_workqueue);
1402                 return error;
1403         }
1404
1405         return 0;
1406 }
1407
1408 void dlm_lowcomms_stop(void)
1409 {
1410         int i;
1411         struct connection *con;
1412
1413         /* Set all the flags to prevent any
1414            socket activity.
1415         */
1416         down(&connections_lock);
1417         for (i = 0; i <= max_nodeid; i++) {
1418                 con = __nodeid2con(i, 0);
1419                 if (con) {
1420                         con->flags |= 0x0F;
1421                         if (con->sock)
1422                                 con->sock->sk->sk_user_data = NULL;
1423                 }
1424         }
1425         up(&connections_lock);
1426
1427         work_stop();
1428
1429         down(&connections_lock);
1430         clean_writequeues();
1431
1432         for (i = 0; i <= max_nodeid; i++) {
1433                 con = __nodeid2con(i, 0);
1434                 if (con) {
1435                         close_connection(con, true);
1436                         kmem_cache_free(con_cache, con);
1437                 }
1438         }
1439         max_nodeid = 0;
1440         up(&connections_lock);
1441         kmem_cache_destroy(con_cache);
1442         idr_init(&connections_idr);
1443 }
1444
1445 int dlm_lowcomms_start(void)
1446 {
1447         int error = -EINVAL;
1448         struct connection *con;
1449
1450         init_local();
1451         if (!dlm_local_count) {
1452                 error = -ENOTCONN;
1453                 log_print("no local IP address has been set");
1454                 goto out;
1455         }
1456
1457         error = -ENOMEM;
1458         con_cache = kmem_cache_create("dlm_conn", sizeof(struct connection),
1459                                       __alignof__(struct connection), 0,
1460                                       NULL);
1461         if (!con_cache)
1462                 goto out;
1463
1464         /* Set some sysctl minima */
1465         if (sysctl_rmem_max < NEEDED_RMEM)
1466                 sysctl_rmem_max = NEEDED_RMEM;
1467
1468         /* Start listening */
1469         if (dlm_config.ci_protocol == 0)
1470                 error = tcp_listen_for_all();
1471         else
1472                 error = sctp_listen_for_all();
1473         if (error)
1474                 goto fail_unlisten;
1475
1476         error = work_start();
1477         if (error)
1478                 goto fail_unlisten;
1479
1480         return 0;
1481
1482 fail_unlisten:
1483         con = nodeid2con(0,0);
1484         if (con) {
1485                 close_connection(con, false);
1486                 kmem_cache_free(con_cache, con);
1487         }
1488         kmem_cache_destroy(con_cache);
1489
1490 out:
1491         return error;
1492 }