net/ceph/messenger.c

   1 #include <linux/ceph/ceph_debug.h>
   2
   3 #include <linux/crc32c.h>
   4 #include <linux/ctype.h>
   5 #include <linux/highmem.h>
   6 #include <linux/inet.h>
   7 #include <linux/kthread.h>
   8 #include <linux/net.h>
   9 #include <linux/slab.h>
  10 #include <linux/socket.h>
  11 #include <linux/string.h>
  12 #include <linux/bio.h>
  13 #include <linux/blkdev.h>
  14 #include <net/tcp.h>
  15
  16 #include <linux/ceph/libceph.h>
  17 #include <linux/ceph/messenger.h>
  18 #include <linux/ceph/decode.h>
  19 #include <linux/ceph/pagelist.h>
  20
  21 /*
  22  * Ceph uses the messenger to exchange ceph_msg messages with other
  23  * hosts in the system.  The messenger provides ordered and reliable
  24  * delivery.  We tolerate TCP disconnects by reconnecting (with
  25  * exponential backoff) in the case of a fault (disconnection, bad
  26  * crc, protocol error).  Acks allow sent messages to be discarded by
  27  * the sender.
  28  */
  29
  30 /* static tag bytes (protocol control messages) */
  31 static char tag_msg = CEPH_MSGR_TAG_MSG;
  32 static char tag_ack = CEPH_MSGR_TAG_ACK;
  33 static char tag_keepalive = CEPH_MSGR_TAG_KEEPALIVE;
  34
  35 #ifdef CONFIG_LOCKDEP
  36 static struct lock_class_key socket_class;
  37 #endif
  38
  39
  40 static void queue_con(struct ceph_connection *con);
  41 static void con_work(struct work_struct *);
  42 static void ceph_fault(struct ceph_connection *con);
  43
  44 /*
  45  * nicely render a sockaddr as a string.
  46  */
  47 #define MAX_ADDR_STR 20
  48 #define MAX_ADDR_STR_LEN 60
  49 static char addr_str[MAX_ADDR_STR][MAX_ADDR_STR_LEN];
  50 static DEFINE_SPINLOCK(addr_str_lock);
  51 static int last_addr_str;
  52
  53 const char *ceph_pr_addr(const struct sockaddr_storage *ss)
  54 {
  55         int i;
  56         char *s;
  57         struct sockaddr_in *in4 = (void *)ss;
  58         struct sockaddr_in6 *in6 = (void *)ss;
  59
  60         spin_lock(&addr_str_lock);
  61         i = last_addr_str++;
  62         if (last_addr_str == MAX_ADDR_STR)
  63                 last_addr_str = 0;
  64         spin_unlock(&addr_str_lock);
  65         s = addr_str[i];
  66
  67         switch (ss->ss_family) {
  68         case AF_INET:
  69                 snprintf(s, MAX_ADDR_STR_LEN, "%pI4:%u", &in4->sin_addr,
  70                          (unsigned int)ntohs(in4->sin_port));
  71                 break;
  72
  73         case AF_INET6:
  74                 snprintf(s, MAX_ADDR_STR_LEN, "[%pI6c]:%u", &in6->sin6_addr,
  75                          (unsigned int)ntohs(in6->sin6_port));
  76                 break;
  77
  78         default:
  79                 sprintf(s, "(unknown sockaddr family %d)", (int)ss->ss_family);
  80         }
  81
  82         return s;
  83 }
  84 EXPORT_SYMBOL(ceph_pr_addr);
  85
  86 static void encode_my_addr(struct ceph_messenger *msgr)
  87 {
  88         memcpy(&msgr->my_enc_addr, &msgr->inst.addr, sizeof(msgr->my_enc_addr));
  89         ceph_encode_addr(&msgr->my_enc_addr);
  90 }
  91
  92 /*
  93  * work queue for all reading and writing to/from the socket.
  94  */
  95 struct workqueue_struct *ceph_msgr_wq;
  96
  97 int ceph_msgr_init(void)
  98 {
  99         ceph_msgr_wq = alloc_workqueue("ceph-msgr", WQ_NON_REENTRANT, 0);
 100         if (!ceph_msgr_wq) {
 101                 pr_err("msgr_init failed to create workqueue\n");
 102                 return -ENOMEM;
 103         }
 104         return 0;
 105 }
 106 EXPORT_SYMBOL(ceph_msgr_init);
 107
 108 void ceph_msgr_exit(void)
 109 {
 110         destroy_workqueue(ceph_msgr_wq);
 111 }
 112 EXPORT_SYMBOL(ceph_msgr_exit);
 113
 114 void ceph_msgr_flush(void)
 115 {
 116         flush_workqueue(ceph_msgr_wq);
 117 }
 118 EXPORT_SYMBOL(ceph_msgr_flush);
 119
 120
 121 /*
 122  * socket callback functions
 123  */
 124
 125 /* data available on socket, or listen socket received a connect */
 126 static void ceph_data_ready(struct sock *sk, int count_unused)
 127 {
 128         struct ceph_connection *con =
 129                 (struct ceph_connection *)sk->sk_user_data;
 130         if (sk->sk_state != TCP_CLOSE_WAIT) {
 131                 dout("ceph_data_ready on %p state = %lu, queueing work\n",
 132                      con, con->state);
 133                 queue_con(con);
 134         }
 135 }
 136
 137 /* socket has buffer space for writing */
 138 static void ceph_write_space(struct sock *sk)
 139 {
 140         struct ceph_connection *con =
 141                 (struct ceph_connection *)sk->sk_user_data;
 142
 143         /* only queue to workqueue if there is data we want to write. */
 144         if (test_bit(WRITE_PENDING, &con->state)) {
 145                 dout("ceph_write_space %p queueing write work\n", con);
 146                 queue_con(con);
 147         } else {
 148                 dout("ceph_write_space %p nothing to write\n", con);
 149         }
 150
 151         /* since we have our own write_space, clear the SOCK_NOSPACE flag */
 152         clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 153 }
 154
 155 /* socket's state has changed */
 156 static void ceph_state_change(struct sock *sk)
 157 {
 158         struct ceph_connection *con =
 159                 (struct ceph_connection *)sk->sk_user_data;
 160
 161         dout("ceph_state_change %p state = %lu sk_state = %u\n",
 162              con, con->state, sk->sk_state);
 163
 164         if (test_bit(CLOSED, &con->state))
 165                 return;
 166
 167         switch (sk->sk_state) {
 168         case TCP_CLOSE:
 169                 dout("ceph_state_change TCP_CLOSE\n");
 170         case TCP_CLOSE_WAIT:
 171                 dout("ceph_state_change TCP_CLOSE_WAIT\n");
 172                 if (test_and_set_bit(SOCK_CLOSED, &con->state) == 0) {
 173                         if (test_bit(CONNECTING, &con->state))
 174                                 con->error_msg = "connection failed";
 175                         else
 176                                 con->error_msg = "socket closed";
 177                         queue_con(con);
 178                 }
 179                 break;
 180         case TCP_ESTABLISHED:
 181                 dout("ceph_state_change TCP_ESTABLISHED\n");
 182                 queue_con(con);
 183                 break;
 184         }
 185 }
 186
 187 /*
 188  * set up socket callbacks
 189  */
 190 static void set_sock_callbacks(struct socket *sock,
 191                                struct ceph_connection *con)
 192 {
 193         struct sock *sk = sock->sk;
 194         sk->sk_user_data = (void *)con;
 195         sk->sk_data_ready = ceph_data_ready;
 196         sk->sk_write_space = ceph_write_space;
 197         sk->sk_state_change = ceph_state_change;
 198 }
 199
 200
 201 /*
 202  * socket helpers
 203  */
 204
 205 /*
 206  * initiate connection to a remote socket.
 207  */
 208 static struct socket *ceph_tcp_connect(struct ceph_connection *con)
 209 {
 210         struct sockaddr_storage *paddr = &con->peer_addr.in_addr;
 211         struct socket *sock;
 212         int ret;
 213
 214         BUG_ON(con->sock);
 215         ret = sock_create_kern(con->peer_addr.in_addr.ss_family, SOCK_STREAM,
 216                                IPPROTO_TCP, &sock);
 217         if (ret)
 218                 return ERR_PTR(ret);
 219         con->sock = sock;
 220         sock->sk->sk_allocation = GFP_NOFS;
 221
 222 #ifdef CONFIG_LOCKDEP
 223         lockdep_set_class(&sock->sk->sk_lock, &socket_class);
 224 #endif
 225
 226         set_sock_callbacks(sock, con);
 227
 228         dout("connect %s\n", ceph_pr_addr(&con->peer_addr.in_addr));
 229
 230         ret = sock->ops->connect(sock, (struct sockaddr *)paddr, sizeof(*paddr),
 231                                  O_NONBLOCK);
 232         if (ret == -EINPROGRESS) {
 233                 dout("connect %s EINPROGRESS sk_state = %u\n",
 234                      ceph_pr_addr(&con->peer_addr.in_addr),
 235                      sock->sk->sk_state);
 236                 ret = 0;
 237         }
 238         if (ret < 0) {
 239                 pr_err("connect %s error %d\n",
 240                        ceph_pr_addr(&con->peer_addr.in_addr), ret);
 241                 sock_release(sock);
 242                 con->sock = NULL;
 243                 con->error_msg = "connect error";
 244         }
 245
 246         if (ret < 0)
 247                 return ERR_PTR(ret);
 248         return sock;
 249 }
 250
 251 static int ceph_tcp_recvmsg(struct socket *sock, void *buf, size_t len)
 252 {
 253         struct kvec iov = {buf, len};
 254         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
 255         int r;
 256
 257         r = kernel_recvmsg(sock, &msg, &iov, 1, len, msg.msg_flags);
 258         if (r == -EAGAIN)
 259                 r = 0;
 260         return r;
 261 }
 262
 263 /*
 264  * write something.  @more is true if caller will be sending more data
 265  * shortly.
 266  */
 267 static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov,
 268                      size_t kvlen, size_t len, int more)
 269 {
 270         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
 271         int r;
 272
 273         if (more)
 274                 msg.msg_flags |= MSG_MORE;
 275         else
 276                 msg.msg_flags |= MSG_EOR;  /* superfluous, but what the hell */
 277
 278         r = kernel_sendmsg(sock, &msg, iov, kvlen, len);
 279         if (r == -EAGAIN)
 280                 r = 0;
 281         return r;
 282 }
 283
 284
 285 /*
 286  * Shutdown/close the socket for the given connection.
 287  */
 288 static int con_close_socket(struct ceph_connection *con)
 289 {
 290         int rc;
 291
 292         dout("con_close_socket on %p sock %p\n", con, con->sock);
 293         if (!con->sock)
 294                 return 0;
 295         set_bit(SOCK_CLOSED, &con->state);
 296         rc = con->sock->ops->shutdown(con->sock, SHUT_RDWR);
 297         sock_release(con->sock);
 298         con->sock = NULL;
 299         clear_bit(SOCK_CLOSED, &con->state);
 300         return rc;
 301 }
 302
 303 /*
 304  * Reset a connection.  Discard all incoming and outgoing messages
 305  * and clear *_seq state.
 306  */
 307 static void ceph_msg_remove(struct ceph_msg *msg)
 308 {
 309         list_del_init(&msg->list_head);
 310         ceph_msg_put(msg);
 311 }
 312 static void ceph_msg_remove_list(struct list_head *head)
 313 {
 314         while (!list_empty(head)) {
 315                 struct ceph_msg *msg = list_first_entry(head, struct ceph_msg,
 316                                                         list_head);
 317                 ceph_msg_remove(msg);
 318         }
 319 }
 320
 321 static void reset_connection(struct ceph_connection *con)
 322 {
 323         /* reset connection, out_queue, msg_ and connect_seq */
 324         /* discard existing out_queue and msg_seq */
 325         ceph_msg_remove_list(&con->out_queue);
 326         ceph_msg_remove_list(&con->out_sent);
 327
 328         if (con->in_msg) {
 329                 ceph_msg_put(con->in_msg);
 330                 con->in_msg = NULL;
 331         }
 332
 333         con->connect_seq = 0;
 334         con->out_seq = 0;
 335         if (con->out_msg) {
 336                 ceph_msg_put(con->out_msg);
 337                 con->out_msg = NULL;
 338         }
 339         con->out_keepalive_pending = false;
 340         con->in_seq = 0;
 341         con->in_seq_acked = 0;
 342 }
 343
 344 /*
 345  * mark a peer down.  drop any open connections.
 346  */
 347 void ceph_con_close(struct ceph_connection *con)
 348 {
 349         dout("con_close %p peer %s\n", con,
 350              ceph_pr_addr(&con->peer_addr.in_addr));
 351         set_bit(CLOSED, &con->state);  /* in case there's queued work */
 352         clear_bit(STANDBY, &con->state);  /* avoid connect_seq bump */
 353         clear_bit(LOSSYTX, &con->state);  /* so we retry next connect */
 354         clear_bit(KEEPALIVE_PENDING, &con->state);
 355         clear_bit(WRITE_PENDING, &con->state);
 356         mutex_lock(&con->mutex);
 357         reset_connection(con);
 358         con->peer_global_seq = 0;
 359         cancel_delayed_work(&con->work);
 360         mutex_unlock(&con->mutex);
 361         queue_con(con);
 362 }
 363 EXPORT_SYMBOL(ceph_con_close);
 364
 365 /*
 366  * Reopen a closed connection, with a new peer address.
 367  */
 368 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
 369 {
 370         dout("con_open %p %s\n", con, ceph_pr_addr(&addr->in_addr));
 371         set_bit(OPENING, &con->state);
 372         clear_bit(CLOSED, &con->state);
 373         memcpy(&con->peer_addr, addr, sizeof(*addr));
 374         con->delay = 0;      /* reset backoff memory */
 375         queue_con(con);
 376 }
 377 EXPORT_SYMBOL(ceph_con_open);
 378
 379 /*
 380  * return true if this connection ever successfully opened
 381  */
 382 bool ceph_con_opened(struct ceph_connection *con)
 383 {
 384         return con->connect_seq > 0;
 385 }
 386
 387 /*
 388  * generic get/put
 389  */
 390 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
 391 {
 392         dout("con_get %p nref = %d -> %d\n", con,
 393              atomic_read(&con->nref), atomic_read(&con->nref) + 1);
 394         if (atomic_inc_not_zero(&con->nref))
 395                 return con;
 396         return NULL;
 397 }
 398
 399 void ceph_con_put(struct ceph_connection *con)
 400 {
 401         dout("con_put %p nref = %d -> %d\n", con,
 402              atomic_read(&con->nref), atomic_read(&con->nref) - 1);
 403         BUG_ON(atomic_read(&con->nref) == 0);
 404         if (atomic_dec_and_test(&con->nref)) {
 405                 BUG_ON(con->sock);
 406                 kfree(con);
 407         }
 408 }
 409
 410 /*
 411  * initialize a new connection.
 412  */
 413 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
 414 {
 415         dout("con_init %p\n", con);
 416         memset(con, 0, sizeof(*con));
 417         atomic_set(&con->nref, 1);
 418         con->msgr = msgr;
 419         mutex_init(&con->mutex);
 420         INIT_LIST_HEAD(&con->out_queue);
 421         INIT_LIST_HEAD(&con->out_sent);
 422         INIT_DELAYED_WORK(&con->work, con_work);
 423 }
 424 EXPORT_SYMBOL(ceph_con_init);
 425
 426
 427 /*
 428  * We maintain a global counter to order connection attempts.  Get
 429  * a unique seq greater than @gt.
 430  */
 431 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
 432 {
 433         u32 ret;
 434
 435         spin_lock(&msgr->global_seq_lock);
 436         if (msgr->global_seq < gt)
 437                 msgr->global_seq = gt;
 438         ret = ++msgr->global_seq;
 439         spin_unlock(&msgr->global_seq_lock);
 440         return ret;
 441 }
 442
 443
 444 /*
 445  * Prepare footer for currently outgoing message, and finish things
 446  * off.  Assumes out_kvec* are already valid.. we just add on to the end.
 447  */
 448 static void prepare_write_message_footer(struct ceph_connection *con, int v)
 449 {
 450         struct ceph_msg *m = con->out_msg;
 451
 452         dout("prepare_write_message_footer %p\n", con);
 453         con->out_kvec_is_msg = true;
 454         con->out_kvec[v].iov_base = &m->footer;
 455         con->out_kvec[v].iov_len = sizeof(m->footer);
 456         con->out_kvec_bytes += sizeof(m->footer);
 457         con->out_kvec_left++;
 458         con->out_more = m->more_to_follow;
 459         con->out_msg_done = true;
 460 }
 461
 462 /*
 463  * Prepare headers for the next outgoing message.
 464  */
 465 static void prepare_write_message(struct ceph_connection *con)
 466 {
 467         struct ceph_msg *m;
 468         int v = 0;
 469
 470         con->out_kvec_bytes = 0;
 471         con->out_kvec_is_msg = true;
 472         con->out_msg_done = false;
 473
 474         /* Sneak an ack in there first?  If we can get it into the same
 475          * TCP packet that's a good thing. */
 476         if (con->in_seq > con->in_seq_acked) {
 477                 con->in_seq_acked = con->in_seq;
 478                 con->out_kvec[v].iov_base = &tag_ack;
 479                 con->out_kvec[v++].iov_len = 1;
 480                 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 481                 con->out_kvec[v].iov_base = &con->out_temp_ack;
 482                 con->out_kvec[v++].iov_len = sizeof(con->out_temp_ack);
 483                 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 484         }
 485
 486         m = list_first_entry(&con->out_queue,
 487                        struct ceph_msg, list_head);
 488         con->out_msg = m;
 489         if (test_bit(LOSSYTX, &con->state)) {
 490                 list_del_init(&m->list_head);
 491         } else {
 492                 /* put message on sent list */
 493                 ceph_msg_get(m);
 494                 list_move_tail(&m->list_head, &con->out_sent);
 495         }
 496
 497         /*
 498          * only assign outgoing seq # if we haven't sent this message
 499          * yet.  if it is requeued, resend with it's original seq.
 500          */
 501         if (m->needs_out_seq) {
 502                 m->hdr.seq = cpu_to_le64(++con->out_seq);
 503                 m->needs_out_seq = false;
 504         }
 505
 506         dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
 507              m, con->out_seq, le16_to_cpu(m->hdr.type),
 508              le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
 509              le32_to_cpu(m->hdr.data_len),
 510              m->nr_pages);
 511         BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
 512
 513         /* tag + hdr + front + middle */
 514         con->out_kvec[v].iov_base = &tag_msg;
 515         con->out_kvec[v++].iov_len = 1;
 516         con->out_kvec[v].iov_base = &m->hdr;
 517         con->out_kvec[v++].iov_len = sizeof(m->hdr);
 518         con->out_kvec[v++] = m->front;
 519         if (m->middle)
 520                 con->out_kvec[v++] = m->middle->vec;
 521         con->out_kvec_left = v;
 522         con->out_kvec_bytes += 1 + sizeof(m->hdr) + m->front.iov_len +
 523                 (m->middle ? m->middle->vec.iov_len : 0);
 524         con->out_kvec_cur = con->out_kvec;
 525
 526         /* fill in crc (except data pages), footer */
 527         con->out_msg->hdr.crc =
 528                 cpu_to_le32(crc32c(0, (void *)&m->hdr,
 529                                       sizeof(m->hdr) - sizeof(m->hdr.crc)));
 530         con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
 531         con->out_msg->footer.front_crc =
 532                 cpu_to_le32(crc32c(0, m->front.iov_base, m->front.iov_len));
 533         if (m->middle)
 534                 con->out_msg->footer.middle_crc =
 535                         cpu_to_le32(crc32c(0, m->middle->vec.iov_base,
 536                                            m->middle->vec.iov_len));
 537         else
 538                 con->out_msg->footer.middle_crc = 0;
 539         con->out_msg->footer.data_crc = 0;
 540         dout("prepare_write_message front_crc %u data_crc %u\n",
 541              le32_to_cpu(con->out_msg->footer.front_crc),
 542              le32_to_cpu(con->out_msg->footer.middle_crc));
 543
 544         /* is there a data payload? */
 545         if (le32_to_cpu(m->hdr.data_len) > 0) {
 546                 /* initialize page iterator */
 547                 con->out_msg_pos.page = 0;
 548                 if (m->pages)
 549                         con->out_msg_pos.page_pos = m->page_alignment;
 550                 else
 551                         con->out_msg_pos.page_pos = 0;
 552                 con->out_msg_pos.data_pos = 0;
 553                 con->out_msg_pos.did_page_crc = 0;
 554                 con->out_more = 1;  /* data + footer will follow */
 555         } else {
 556                 /* no, queue up footer too and be done */
 557                 prepare_write_message_footer(con, v);
 558         }
 559
 560         set_bit(WRITE_PENDING, &con->state);
 561 }
 562
 563 /*
 564  * Prepare an ack.
 565  */
 566 static void prepare_write_ack(struct ceph_connection *con)
 567 {
 568         dout("prepare_write_ack %p %llu -> %llu\n", con,
 569              con->in_seq_acked, con->in_seq);
 570         con->in_seq_acked = con->in_seq;
 571
 572         con->out_kvec[0].iov_base = &tag_ack;
 573         con->out_kvec[0].iov_len = 1;
 574         con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 575         con->out_kvec[1].iov_base = &con->out_temp_ack;
 576         con->out_kvec[1].iov_len = sizeof(con->out_temp_ack);
 577         con->out_kvec_left = 2;
 578         con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 579         con->out_kvec_cur = con->out_kvec;
 580         con->out_more = 1;  /* more will follow.. eventually.. */
 581         set_bit(WRITE_PENDING, &con->state);
 582 }
 583
 584 /*
 585  * Prepare to write keepalive byte.
 586  */
 587 static void prepare_write_keepalive(struct ceph_connection *con)
 588 {
 589         dout("prepare_write_keepalive %p\n", con);
 590         con->out_kvec[0].iov_base = &tag_keepalive;
 591         con->out_kvec[0].iov_len = 1;
 592         con->out_kvec_left = 1;
 593         con->out_kvec_bytes = 1;
 594         con->out_kvec_cur = con->out_kvec;
 595         set_bit(WRITE_PENDING, &con->state);
 596 }
 597
 598 /*
 599  * Connection negotiation.
 600  */
 601
 602 static void prepare_connect_authorizer(struct ceph_connection *con)
 603 {
 604         void *auth_buf;
 605         int auth_len = 0;
 606         int auth_protocol = 0;
 607
 608         mutex_unlock(&con->mutex);
 609         if (con->ops->get_authorizer)
 610                 con->ops->get_authorizer(con, &auth_buf, &auth_len,
 611                                          &auth_protocol, &con->auth_reply_buf,
 612                                          &con->auth_reply_buf_len,
 613                                          con->auth_retry);
 614         mutex_lock(&con->mutex);
 615
 616         con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
 617         con->out_connect.authorizer_len = cpu_to_le32(auth_len);
 618
 619         con->out_kvec[con->out_kvec_left].iov_base = auth_buf;
 620         con->out_kvec[con->out_kvec_left].iov_len = auth_len;
 621         con->out_kvec_left++;
 622         con->out_kvec_bytes += auth_len;
 623 }
 624
 625 /*
 626  * We connected to a peer and are saying hello.
 627  */
 628 static void prepare_write_banner(struct ceph_messenger *msgr,
 629                                  struct ceph_connection *con)
 630 {
 631         int len = strlen(CEPH_BANNER);
 632
 633         con->out_kvec[0].iov_base = CEPH_BANNER;
 634         con->out_kvec[0].iov_len = len;
 635         con->out_kvec[1].iov_base = &msgr->my_enc_addr;
 636         con->out_kvec[1].iov_len = sizeof(msgr->my_enc_addr);
 637         con->out_kvec_left = 2;
 638         con->out_kvec_bytes = len + sizeof(msgr->my_enc_addr);
 639         con->out_kvec_cur = con->out_kvec;
 640         con->out_more = 0;
 641         set_bit(WRITE_PENDING, &con->state);
 642 }
 643
 644 static void prepare_write_connect(struct ceph_messenger *msgr,
 645                                   struct ceph_connection *con,
 646                                   int after_banner)
 647 {
 648         unsigned global_seq = get_global_seq(con->msgr, 0);
 649         int proto;
 650
 651         switch (con->peer_name.type) {
 652         case CEPH_ENTITY_TYPE_MON:
 653                 proto = CEPH_MONC_PROTOCOL;
 654                 break;
 655         case CEPH_ENTITY_TYPE_OSD:
 656                 proto = CEPH_OSDC_PROTOCOL;
 657                 break;
 658         case CEPH_ENTITY_TYPE_MDS:
 659                 proto = CEPH_MDSC_PROTOCOL;
 660                 break;
 661         default:
 662                 BUG();
 663         }
 664
 665         dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
 666              con->connect_seq, global_seq, proto);
 667
 668         con->out_connect.features = cpu_to_le64(msgr->supported_features);
 669         con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
 670         con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
 671         con->out_connect.global_seq = cpu_to_le32(global_seq);
 672         con->out_connect.protocol_version = cpu_to_le32(proto);
 673         con->out_connect.flags = 0;
 674
 675         if (!after_banner) {
 676                 con->out_kvec_left = 0;
 677                 con->out_kvec_bytes = 0;
 678         }
 679         con->out_kvec[con->out_kvec_left].iov_base = &con->out_connect;
 680         con->out_kvec[con->out_kvec_left].iov_len = sizeof(con->out_connect);
 681         con->out_kvec_left++;
 682         con->out_kvec_bytes += sizeof(con->out_connect);
 683         con->out_kvec_cur = con->out_kvec;
 684         con->out_more = 0;
 685         set_bit(WRITE_PENDING, &con->state);
 686
 687         prepare_connect_authorizer(con);
 688 }
 689
 690
 691 /*
 692  * write as much of pending kvecs to the socket as we can.
 693  *  1 -> done
 694  *  0 -> socket full, but more to do
 695  * <0 -> error
 696  */
 697 static int write_partial_kvec(struct ceph_connection *con)
 698 {
 699         int ret;
 700
 701         dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
 702         while (con->out_kvec_bytes > 0) {
 703                 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
 704                                        con->out_kvec_left, con->out_kvec_bytes,
 705                                        con->out_more);
 706                 if (ret <= 0)
 707                         goto out;
 708                 con->out_kvec_bytes -= ret;
 709                 if (con->out_kvec_bytes == 0)
 710                         break;            /* done */
 711                 while (ret > 0) {
 712                         if (ret >= con->out_kvec_cur->iov_len) {
 713                                 ret -= con->out_kvec_cur->iov_len;
 714                                 con->out_kvec_cur++;
 715                                 con->out_kvec_left--;
 716                         } else {
 717                                 con->out_kvec_cur->iov_len -= ret;
 718                                 con->out_kvec_cur->iov_base += ret;
 719                                 ret = 0;
 720                                 break;
 721                         }
 722                 }
 723         }
 724         con->out_kvec_left = 0;
 725         con->out_kvec_is_msg = false;
 726         ret = 1;
 727 out:
 728         dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
 729              con->out_kvec_bytes, con->out_kvec_left, ret);
 730         return ret;  /* done! */
 731 }
 732
 733 #ifdef CONFIG_BLOCK
 734 static void init_bio_iter(struct bio *bio, struct bio **iter, int *seg)
 735 {
 736         if (!bio) {
 737                 *iter = NULL;
 738                 *seg = 0;
 739                 return;
 740         }
 741         *iter = bio;
 742         *seg = bio->bi_idx;
 743 }
 744
 745 static void iter_bio_next(struct bio **bio_iter, int *seg)
 746 {
 747         if (*bio_iter == NULL)
 748                 return;
 749
 750         BUG_ON(*seg >= (*bio_iter)->bi_vcnt);
 751
 752         (*seg)++;
 753         if (*seg == (*bio_iter)->bi_vcnt)
 754                 init_bio_iter((*bio_iter)->bi_next, bio_iter, seg);
 755 }
 756 #endif
 757
 758 /*
 759  * Write as much message data payload as we can.  If we finish, queue
 760  * up the footer.
 761  *  1 -> done, footer is now queued in out_kvec[].
 762  *  0 -> socket full, but more to do
 763  * <0 -> error
 764  */
 765 static int write_partial_msg_pages(struct ceph_connection *con)
 766 {
 767         struct ceph_msg *msg = con->out_msg;
 768         unsigned data_len = le32_to_cpu(msg->hdr.data_len);
 769         size_t len;
 770         int crc = con->msgr->nocrc;
 771         int ret;
 772         int total_max_write;
 773         int in_trail = 0;
 774         size_t trail_len = (msg->trail ? msg->trail->length : 0);
 775
 776         dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
 777              con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
 778              con->out_msg_pos.page_pos);
 779
 780 #ifdef CONFIG_BLOCK
 781         if (msg->bio && !msg->bio_iter)
 782                 init_bio_iter(msg->bio, &msg->bio_iter, &msg->bio_seg);
 783 #endif
 784
 785         while (data_len > con->out_msg_pos.data_pos) {
 786                 struct page *page = NULL;
 787                 void *kaddr = NULL;
 788                 int max_write = PAGE_SIZE;
 789                 int page_shift = 0;
 790
 791                 total_max_write = data_len - trail_len -
 792                         con->out_msg_pos.data_pos;
 793
 794                 /*
 795                  * if we are calculating the data crc (the default), we need
 796                  * to map the page.  if our pages[] has been revoked, use the
 797                  * zero page.
 798                  */
 799
 800                 /* have we reached the trail part of the data? */
 801                 if (con->out_msg_pos.data_pos >= data_len - trail_len) {
 802                         in_trail = 1;
 803
 804                         total_max_write = data_len - con->out_msg_pos.data_pos;
 805
 806                         page = list_first_entry(&msg->trail->head,
 807                                                 struct page, lru);
 808                         if (crc)
 809                                 kaddr = kmap(page);
 810                         max_write = PAGE_SIZE;
 811                 } else if (msg->pages) {
 812                         page = msg->pages[con->out_msg_pos.page];
 813                         if (crc)
 814                                 kaddr = kmap(page);
 815                 } else if (msg->pagelist) {
 816                         page = list_first_entry(&msg->pagelist->head,
 817                                                 struct page, lru);
 818                         if (crc)
 819                                 kaddr = kmap(page);
 820 #ifdef CONFIG_BLOCK
 821                 } else if (msg->bio) {
 822                         struct bio_vec *bv;
 823
 824                         bv = bio_iovec_idx(msg->bio_iter, msg->bio_seg);
 825                         page = bv->bv_page;
 826                         page_shift = bv->bv_offset;
 827                         if (crc)
 828                                 kaddr = kmap(page) + page_shift;
 829                         max_write = bv->bv_len;
 830 #endif
 831                 } else {
 832                         page = con->msgr->zero_page;
 833                         if (crc)
 834                                 kaddr = page_address(con->msgr->zero_page);
 835                 }
 836                 len = min_t(int, max_write - con->out_msg_pos.page_pos,
 837                             total_max_write);
 838
 839                 if (crc && !con->out_msg_pos.did_page_crc) {
 840                         void *base = kaddr + con->out_msg_pos.page_pos;
 841                         u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
 842
 843                         BUG_ON(kaddr == NULL);
 844                         con->out_msg->footer.data_crc =
 845                                 cpu_to_le32(crc32c(tmpcrc, base, len));
 846                         con->out_msg_pos.did_page_crc = 1;
 847                 }
 848                 ret = kernel_sendpage(con->sock, page,
 849                                       con->out_msg_pos.page_pos + page_shift,
 850                                       len,
 851                                       MSG_DONTWAIT | MSG_NOSIGNAL |
 852                                       MSG_MORE);
 853
 854                 if (crc &&
 855                     (msg->pages || msg->pagelist || msg->bio || in_trail))
 856                         kunmap(page);
 857
 858                 if (ret == -EAGAIN)
 859                         ret = 0;
 860                 if (ret <= 0)
 861                         goto out;
 862
 863                 con->out_msg_pos.data_pos += ret;
 864                 con->out_msg_pos.page_pos += ret;
 865                 if (ret == len) {
 866                         con->out_msg_pos.page_pos = 0;
 867                         con->out_msg_pos.page++;
 868                         con->out_msg_pos.did_page_crc = 0;
 869                         if (in_trail)
 870                                 list_move_tail(&page->lru,
 871                                                &msg->trail->head);
 872                         else if (msg->pagelist)
 873                                 list_move_tail(&page->lru,
 874                                                &msg->pagelist->head);
 875 #ifdef CONFIG_BLOCK
 876                         else if (msg->bio)
 877                                 iter_bio_next(&msg->bio_iter, &msg->bio_seg);
 878 #endif
 879                 }
 880         }
 881
 882         dout("write_partial_msg_pages %p msg %p done\n", con, msg);
 883
 884         /* prepare and queue up footer, too */
 885         if (!crc)
 886                 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
 887         con->out_kvec_bytes = 0;
 888         con->out_kvec_left = 0;
 889         con->out_kvec_cur = con->out_kvec;
 890         prepare_write_message_footer(con, 0);
 891         ret = 1;
 892 out:
 893         return ret;
 894 }
 895
 896 /*
 897  * write some zeros
 898  */
 899 static int write_partial_skip(struct ceph_connection *con)
 900 {
 901         int ret;
 902
 903         while (con->out_skip > 0) {
 904                 struct kvec iov = {
 905                         .iov_base = page_address(con->msgr->zero_page),
 906                         .iov_len = min(con->out_skip, (int)PAGE_CACHE_SIZE)
 907                 };
 908
 909                 ret = ceph_tcp_sendmsg(con->sock, &iov, 1, iov.iov_len, 1);
 910                 if (ret <= 0)
 911                         goto out;
 912                 con->out_skip -= ret;
 913         }
 914         ret = 1;
 915 out:
 916         return ret;
 917 }
 918
 919 /*
 920  * Prepare to read connection handshake, or an ack.
 921  */
 922 static void prepare_read_banner(struct ceph_connection *con)
 923 {
 924         dout("prepare_read_banner %p\n", con);
 925         con->in_base_pos = 0;
 926 }
 927
 928 static void prepare_read_connect(struct ceph_connection *con)
 929 {
 930         dout("prepare_read_connect %p\n", con);
 931         con->in_base_pos = 0;
 932 }
 933
 934 static void prepare_read_ack(struct ceph_connection *con)
 935 {
 936         dout("prepare_read_ack %p\n", con);
 937         con->in_base_pos = 0;
 938 }
 939
 940 static void prepare_read_tag(struct ceph_connection *con)
 941 {
 942         dout("prepare_read_tag %p\n", con);
 943         con->in_base_pos = 0;
 944         con->in_tag = CEPH_MSGR_TAG_READY;
 945 }
 946
 947 /*
 948  * Prepare to read a message.
 949  */
 950 static int prepare_read_message(struct ceph_connection *con)
 951 {
 952         dout("prepare_read_message %p\n", con);
 953         BUG_ON(con->in_msg != NULL);
 954         con->in_base_pos = 0;
 955         con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
 956         return 0;
 957 }
 958
 959
 960 static int read_partial(struct ceph_connection *con,
 961                         int *to, int size, void *object)
 962 {
 963         *to += size;
 964         while (con->in_base_pos < *to) {
 965                 int left = *to - con->in_base_pos;
 966                 int have = size - left;
 967                 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
 968                 if (ret <= 0)
 969                         return ret;
 970                 con->in_base_pos += ret;
 971         }
 972         return 1;
 973 }
 974
 975
 976 /*
 977  * Read all or part of the connect-side handshake on a new connection
 978  */
 979 static int read_partial_banner(struct ceph_connection *con)
 980 {
 981         int ret, to = 0;
 982
 983         dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
 984
 985         /* peer's banner */
 986         ret = read_partial(con, &to, strlen(CEPH_BANNER), con->in_banner);
 987         if (ret <= 0)
 988                 goto out;
 989         ret = read_partial(con, &to, sizeof(con->actual_peer_addr),
 990                            &con->actual_peer_addr);
 991         if (ret <= 0)
 992                 goto out;
 993         ret = read_partial(con, &to, sizeof(con->peer_addr_for_me),
 994                            &con->peer_addr_for_me);
 995         if (ret <= 0)
 996                 goto out;
 997 out:
 998         return ret;
 999 }
1000
1001 static int read_partial_connect(struct ceph_connection *con)
1002 {
1003         int ret, to = 0;
1004
1005         dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
1006
1007         ret = read_partial(con, &to, sizeof(con->in_reply), &con->in_reply);
1008         if (ret <= 0)
1009                 goto out;
1010         ret = read_partial(con, &to, le32_to_cpu(con->in_reply.authorizer_len),
1011                            con->auth_reply_buf);
1012         if (ret <= 0)
1013                 goto out;
1014
1015         dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
1016              con, (int)con->in_reply.tag,
1017              le32_to_cpu(con->in_reply.connect_seq),
1018              le32_to_cpu(con->in_reply.global_seq));
1019 out:
1020         return ret;
1021
1022 }
1023
1024 /*
1025  * Verify the hello banner looks okay.
1026  */
1027 static int verify_hello(struct ceph_connection *con)
1028 {
1029         if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
1030                 pr_err("connect to %s got bad banner\n",
1031                        ceph_pr_addr(&con->peer_addr.in_addr));
1032                 con->error_msg = "protocol error, bad banner";
1033                 return -1;
1034         }
1035         return 0;
1036 }
1037
1038 static bool addr_is_blank(struct sockaddr_storage *ss)
1039 {
1040         switch (ss->ss_family) {
1041         case AF_INET:
1042                 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
1043         case AF_INET6:
1044                 return
1045                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
1046                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
1047                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
1048                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
1049         }
1050         return false;
1051 }
1052
1053 static int addr_port(struct sockaddr_storage *ss)
1054 {
1055         switch (ss->ss_family) {
1056         case AF_INET:
1057                 return ntohs(((struct sockaddr_in *)ss)->sin_port);
1058         case AF_INET6:
1059                 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
1060         }
1061         return 0;
1062 }
1063
1064 static void addr_set_port(struct sockaddr_storage *ss, int p)
1065 {
1066         switch (ss->ss_family) {
1067         case AF_INET:
1068                 ((struct sockaddr_in *)ss)->sin_port = htons(p);
1069         case AF_INET6:
1070                 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
1071         }
1072 }
1073
1074 /*
1075  * Parse an ip[:port] list into an addr array.  Use the default
1076  * monitor port if a port isn't specified.
1077  */
1078 int ceph_parse_ips(const char *c, const char *end,
1079                    struct ceph_entity_addr *addr,
1080                    int max_count, int *count)
1081 {
1082         int i;
1083         const char *p = c;
1084
1085         dout("parse_ips on '%.*s'\n", (int)(end-c), c);
1086         for (i = 0; i < max_count; i++) {
1087                 const char *ipend;
1088                 struct sockaddr_storage *ss = &addr[i].in_addr;
1089                 struct sockaddr_in *in4 = (void *)ss;
1090                 struct sockaddr_in6 *in6 = (void *)ss;
1091                 int port;
1092                 char delim = ',';
1093
1094                 if (*p == '[') {
1095                         delim = ']';
1096                         p++;
1097                 }
1098
1099                 memset(ss, 0, sizeof(*ss));
1100                 if (in4_pton(p, end - p, (u8 *)&in4->sin_addr.s_addr,
1101                              delim, &ipend))
1102                         ss->ss_family = AF_INET;
1103                 else if (in6_pton(p, end - p, (u8 *)&in6->sin6_addr.s6_addr,
1104                                   delim, &ipend))
1105                         ss->ss_family = AF_INET6;
1106                 else
1107                         goto bad;
1108                 p = ipend;
1109
1110                 if (delim == ']') {
1111                         if (*p != ']') {
1112                                 dout("missing matching ']'\n");
1113                                 goto bad;
1114                         }
1115                         p++;
1116                 }
1117
1118                 /* port? */
1119                 if (p < end && *p == ':') {
1120                         port = 0;
1121                         p++;
1122                         while (p < end && *p >= '0' && *p <= '9') {
1123                                 port = (port * 10) + (*p - '0');
1124                                 p++;
1125                         }
1126                         if (port > 65535 || port == 0)
1127                                 goto bad;
1128                 } else {
1129                         port = CEPH_MON_PORT;
1130                 }
1131
1132                 addr_set_port(ss, port);
1133
1134                 dout("parse_ips got %s\n", ceph_pr_addr(ss));
1135
1136                 if (p == end)
1137                         break;
1138                 if (*p != ',')
1139                         goto bad;
1140                 p++;
1141         }
1142
1143         if (p != end)
1144                 goto bad;
1145
1146         if (count)
1147                 *count = i + 1;
1148         return 0;
1149
1150 bad:
1151         pr_err("parse_ips bad ip '%.*s'\n", (int)(end - c), c);
1152         return -EINVAL;
1153 }
1154 EXPORT_SYMBOL(ceph_parse_ips);
1155
1156 static int process_banner(struct ceph_connection *con)
1157 {
1158         dout("process_banner on %p\n", con);
1159
1160         if (verify_hello(con) < 0)
1161                 return -1;
1162
1163         ceph_decode_addr(&con->actual_peer_addr);
1164         ceph_decode_addr(&con->peer_addr_for_me);
1165
1166         /*
1167          * Make sure the other end is who we wanted.  note that the other
1168          * end may not yet know their ip address, so if it's 0.0.0.0, give
1169          * them the benefit of the doubt.
1170          */
1171         if (memcmp(&con->peer_addr, &con->actual_peer_addr,
1172                    sizeof(con->peer_addr)) != 0 &&
1173             !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1174               con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1175                 pr_warning("wrong peer, want %s/%d, got %s/%d\n",
1176                            ceph_pr_addr(&con->peer_addr.in_addr),
1177                            (int)le32_to_cpu(con->peer_addr.nonce),
1178                            ceph_pr_addr(&con->actual_peer_addr.in_addr),
1179                            (int)le32_to_cpu(con->actual_peer_addr.nonce));
1180                 con->error_msg = "wrong peer at address";
1181                 return -1;
1182         }
1183
1184         /*
1185          * did we learn our address?
1186          */
1187         if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1188                 int port = addr_port(&con->msgr->inst.addr.in_addr);
1189
1190                 memcpy(&con->msgr->inst.addr.in_addr,
1191                        &con->peer_addr_for_me.in_addr,
1192                        sizeof(con->peer_addr_for_me.in_addr));
1193                 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1194                 encode_my_addr(con->msgr);
1195                 dout("process_banner learned my addr is %s\n",
1196                      ceph_pr_addr(&con->msgr->inst.addr.in_addr));
1197         }
1198
1199         set_bit(NEGOTIATING, &con->state);
1200         prepare_read_connect(con);
1201         return 0;
1202 }
1203
1204 static void fail_protocol(struct ceph_connection *con)
1205 {
1206         reset_connection(con);
1207         set_bit(CLOSED, &con->state);  /* in case there's queued work */
1208
1209         mutex_unlock(&con->mutex);
1210         if (con->ops->bad_proto)
1211                 con->ops->bad_proto(con);
1212         mutex_lock(&con->mutex);
1213 }
1214
1215 static int process_connect(struct ceph_connection *con)
1216 {
1217         u64 sup_feat = con->msgr->supported_features;
1218         u64 req_feat = con->msgr->required_features;
1219         u64 server_feat = le64_to_cpu(con->in_reply.features);
1220
1221         dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1222
1223         switch (con->in_reply.tag) {
1224         case CEPH_MSGR_TAG_FEATURES:
1225                 pr_err("%s%lld %s feature set mismatch,"
1226                        " my %llx < server's %llx, missing %llx\n",
1227                        ENTITY_NAME(con->peer_name),
1228                        ceph_pr_addr(&con->peer_addr.in_addr),
1229                        sup_feat, server_feat, server_feat & ~sup_feat);
1230                 con->error_msg = "missing required protocol features";
1231                 fail_protocol(con);
1232                 return -1;
1233
1234         case CEPH_MSGR_TAG_BADPROTOVER:
1235                 pr_err("%s%lld %s protocol version mismatch,"
1236                        " my %d != server's %d\n",
1237                        ENTITY_NAME(con->peer_name),
1238                        ceph_pr_addr(&con->peer_addr.in_addr),
1239                        le32_to_cpu(con->out_connect.protocol_version),
1240                        le32_to_cpu(con->in_reply.protocol_version));
1241                 con->error_msg = "protocol version mismatch";
1242                 fail_protocol(con);
1243                 return -1;
1244
1245         case CEPH_MSGR_TAG_BADAUTHORIZER:
1246                 con->auth_retry++;
1247                 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1248                      con->auth_retry);
1249                 if (con->auth_retry == 2) {
1250                         con->error_msg = "connect authorization failure";
1251                         return -1;
1252                 }
1253                 con->auth_retry = 1;
1254                 prepare_write_connect(con->msgr, con, 0);
1255                 prepare_read_connect(con);
1256                 break;
1257
1258         case CEPH_MSGR_TAG_RESETSESSION:
1259                 /*
1260                  * If we connected with a large connect_seq but the peer
1261                  * has no record of a session with us (no connection, or
1262                  * connect_seq == 0), they will send RESETSESION to indicate
1263                  * that they must have reset their session, and may have
1264                  * dropped messages.
1265                  */
1266                 dout("process_connect got RESET peer seq %u\n",
1267                      le32_to_cpu(con->in_connect.connect_seq));
1268                 pr_err("%s%lld %s connection reset\n",
1269                        ENTITY_NAME(con->peer_name),
1270                        ceph_pr_addr(&con->peer_addr.in_addr));
1271                 reset_connection(con);
1272                 prepare_write_connect(con->msgr, con, 0);
1273                 prepare_read_connect(con);
1274
1275                 /* Tell ceph about it. */
1276                 mutex_unlock(&con->mutex);
1277                 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1278                 if (con->ops->peer_reset)
1279                         con->ops->peer_reset(con);
1280                 mutex_lock(&con->mutex);
1281                 break;
1282
1283         case CEPH_MSGR_TAG_RETRY_SESSION:
1284                 /*
1285                  * If we sent a smaller connect_seq than the peer has, try
1286                  * again with a larger value.
1287                  */
1288                 dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
1289                      le32_to_cpu(con->out_connect.connect_seq),
1290                      le32_to_cpu(con->in_connect.connect_seq));
1291                 con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
1292                 prepare_write_connect(con->msgr, con, 0);
1293                 prepare_read_connect(con);
1294                 break;
1295
1296         case CEPH_MSGR_TAG_RETRY_GLOBAL:
1297                 /*
1298                  * If we sent a smaller global_seq than the peer has, try
1299                  * again with a larger value.
1300                  */
1301                 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1302                      con->peer_global_seq,
1303                      le32_to_cpu(con->in_connect.global_seq));
1304                 get_global_seq(con->msgr,
1305                                le32_to_cpu(con->in_connect.global_seq));
1306                 prepare_write_connect(con->msgr, con, 0);
1307                 prepare_read_connect(con);
1308                 break;
1309
1310         case CEPH_MSGR_TAG_READY:
1311                 if (req_feat & ~server_feat) {
1312                         pr_err("%s%lld %s protocol feature mismatch,"
1313                                " my required %llx > server's %llx, need %llx\n",
1314                                ENTITY_NAME(con->peer_name),
1315                                ceph_pr_addr(&con->peer_addr.in_addr),
1316                                req_feat, server_feat, req_feat & ~server_feat);
1317                         con->error_msg = "missing required protocol features";
1318                         fail_protocol(con);
1319                         return -1;
1320                 }
1321                 clear_bit(CONNECTING, &con->state);
1322                 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1323                 con->connect_seq++;
1324                 con->peer_features = server_feat;
1325                 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1326                      con->peer_global_seq,
1327                      le32_to_cpu(con->in_reply.connect_seq),
1328                      con->connect_seq);
1329                 WARN_ON(con->connect_seq !=
1330                         le32_to_cpu(con->in_reply.connect_seq));
1331
1332                 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1333                         set_bit(LOSSYTX, &con->state);
1334
1335                 prepare_read_tag(con);
1336                 break;
1337
1338         case CEPH_MSGR_TAG_WAIT:
1339                 /*
1340                  * If there is a connection race (we are opening
1341                  * connections to each other), one of us may just have
1342                  * to WAIT.  This shouldn't happen if we are the
1343                  * client.
1344                  */
1345                 pr_err("process_connect peer connecting WAIT\n");
1346
1347         default:
1348                 pr_err("connect protocol error, will retry\n");
1349                 con->error_msg = "protocol error, garbage tag during connect";
1350                 return -1;
1351         }
1352         return 0;
1353 }
1354
1355
1356 /*
1357  * read (part of) an ack
1358  */
1359 static int read_partial_ack(struct ceph_connection *con)
1360 {
1361         int to = 0;
1362
1363         return read_partial(con, &to, sizeof(con->in_temp_ack),
1364                             &con->in_temp_ack);
1365 }
1366
1367
1368 /*
1369  * We can finally discard anything that's been acked.
1370  */
1371 static void process_ack(struct ceph_connection *con)
1372 {
1373         struct ceph_msg *m;
1374         u64 ack = le64_to_cpu(con->in_temp_ack);
1375         u64 seq;
1376
1377         while (!list_empty(&con->out_sent)) {
1378                 m = list_first_entry(&con->out_sent, struct ceph_msg,
1379                                      list_head);
1380                 seq = le64_to_cpu(m->hdr.seq);
1381                 if (seq > ack)
1382                         break;
1383                 dout("got ack for seq %llu type %d at %p\n", seq,
1384                      le16_to_cpu(m->hdr.type), m);
1385                 ceph_msg_remove(m);
1386         }
1387         prepare_read_tag(con);
1388 }
1389
1390
1391
1392
1393 static int read_partial_message_section(struct ceph_connection *con,
1394                                         struct kvec *section,
1395                                         unsigned int sec_len, u32 *crc)
1396 {
1397         int ret, left;
1398
1399         BUG_ON(!section);
1400
1401         while (section->iov_len < sec_len) {
1402                 BUG_ON(section->iov_base == NULL);
1403                 left = sec_len - section->iov_len;
1404                 ret = ceph_tcp_recvmsg(con->sock, (char *)section->iov_base +
1405                                        section->iov_len, left);
1406                 if (ret <= 0)
1407                         return ret;
1408                 section->iov_len += ret;
1409                 if (section->iov_len == sec_len)
1410                         *crc = crc32c(0, section->iov_base,
1411                                       section->iov_len);
1412         }
1413
1414         return 1;
1415 }
1416
1417 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
1418                                 struct ceph_msg_header *hdr,
1419                                 int *skip);
1420
1421
1422 static int read_partial_message_pages(struct ceph_connection *con,
1423                                       struct page **pages,
1424                                       unsigned data_len, int datacrc)
1425 {
1426         void *p;
1427         int ret;
1428         int left;
1429
1430         left = min((int)(data_len - con->in_msg_pos.data_pos),
1431                    (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1432         /* (page) data */
1433         BUG_ON(pages == NULL);
1434         p = kmap(pages[con->in_msg_pos.page]);
1435         ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1436                                left);
1437         if (ret > 0 && datacrc)
1438                 con->in_data_crc =
1439                         crc32c(con->in_data_crc,
1440                                   p + con->in_msg_pos.page_pos, ret);
1441         kunmap(pages[con->in_msg_pos.page]);
1442         if (ret <= 0)
1443                 return ret;
1444         con->in_msg_pos.data_pos += ret;
1445         con->in_msg_pos.page_pos += ret;
1446         if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1447                 con->in_msg_pos.page_pos = 0;
1448                 con->in_msg_pos.page++;
1449         }
1450
1451         return ret;
1452 }
1453
1454 #ifdef CONFIG_BLOCK
1455 static int read_partial_message_bio(struct ceph_connection *con,
1456                                     struct bio **bio_iter, int *bio_seg,
1457                                     unsigned data_len, int datacrc)
1458 {
1459         struct bio_vec *bv = bio_iovec_idx(*bio_iter, *bio_seg);
1460         void *p;
1461         int ret, left;
1462
1463         if (IS_ERR(bv))
1464                 return PTR_ERR(bv);
1465
1466         left = min((int)(data_len - con->in_msg_pos.data_pos),
1467                    (int)(bv->bv_len - con->in_msg_pos.page_pos));
1468
1469         p = kmap(bv->bv_page) + bv->bv_offset;
1470
1471         ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1472                                left);
1473         if (ret > 0 && datacrc)
1474                 con->in_data_crc =
1475                         crc32c(con->in_data_crc,
1476                                   p + con->in_msg_pos.page_pos, ret);
1477         kunmap(bv->bv_page);
1478         if (ret <= 0)
1479                 return ret;
1480         con->in_msg_pos.data_pos += ret;
1481         con->in_msg_pos.page_pos += ret;
1482         if (con->in_msg_pos.page_pos == bv->bv_len) {
1483                 con->in_msg_pos.page_pos = 0;
1484                 iter_bio_next(bio_iter, bio_seg);
1485         }
1486
1487         return ret;
1488 }
1489 #endif
1490
1491 /*
1492  * read (part of) a message.
1493  */
1494 static int read_partial_message(struct ceph_connection *con)
1495 {
1496         struct ceph_msg *m = con->in_msg;
1497         int ret;
1498         int to, left;
1499         unsigned front_len, middle_len, data_len;
1500         int datacrc = con->msgr->nocrc;
1501         int skip;
1502         u64 seq;
1503
1504         dout("read_partial_message con %p msg %p\n", con, m);
1505
1506         /* header */
1507         while (con->in_base_pos < sizeof(con->in_hdr)) {
1508                 left = sizeof(con->in_hdr) - con->in_base_pos;
1509                 ret = ceph_tcp_recvmsg(con->sock,
1510                                        (char *)&con->in_hdr + con->in_base_pos,
1511                                        left);
1512                 if (ret <= 0)
1513                         return ret;
1514                 con->in_base_pos += ret;
1515                 if (con->in_base_pos == sizeof(con->in_hdr)) {
1516                         u32 crc = crc32c(0, (void *)&con->in_hdr,
1517                                  sizeof(con->in_hdr) - sizeof(con->in_hdr.crc));
1518                         if (crc != le32_to_cpu(con->in_hdr.crc)) {
1519                                 pr_err("read_partial_message bad hdr "
1520                                        " crc %u != expected %u\n",
1521                                        crc, con->in_hdr.crc);
1522                                 return -EBADMSG;
1523                         }
1524                 }
1525         }
1526         front_len = le32_to_cpu(con->in_hdr.front_len);
1527         if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1528                 return -EIO;
1529         middle_len = le32_to_cpu(con->in_hdr.middle_len);
1530         if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1531                 return -EIO;
1532         data_len = le32_to_cpu(con->in_hdr.data_len);
1533         if (data_len > CEPH_MSG_MAX_DATA_LEN)
1534                 return -EIO;
1535
1536         /* verify seq# */
1537         seq = le64_to_cpu(con->in_hdr.seq);
1538         if ((s64)seq - (s64)con->in_seq < 1) {
1539                 pr_info("skipping %s%lld %s seq %lld expected %lld\n",
1540                         ENTITY_NAME(con->peer_name),
1541                         ceph_pr_addr(&con->peer_addr.in_addr),
1542                         seq, con->in_seq + 1);
1543                 con->in_base_pos = -front_len - middle_len - data_len -
1544                         sizeof(m->footer);
1545                 con->in_tag = CEPH_MSGR_TAG_READY;
1546                 return 0;
1547         } else if ((s64)seq - (s64)con->in_seq > 1) {
1548                 pr_err("read_partial_message bad seq %lld expected %lld\n",
1549                        seq, con->in_seq + 1);
1550                 con->error_msg = "bad message sequence # for incoming message";
1551                 return -EBADMSG;
1552         }
1553
1554         /* allocate message? */
1555         if (!con->in_msg) {
1556                 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1557                      con->in_hdr.front_len, con->in_hdr.data_len);
1558                 skip = 0;
1559                 con->in_msg = ceph_alloc_msg(con, &con->in_hdr, &skip);
1560                 if (skip) {
1561                         /* skip this message */
1562                         dout("alloc_msg said skip message\n");
1563                         BUG_ON(con->in_msg);
1564                         con->in_base_pos = -front_len - middle_len - data_len -
1565                                 sizeof(m->footer);
1566                         con->in_tag = CEPH_MSGR_TAG_READY;
1567                         con->in_seq++;
1568                         return 0;
1569                 }
1570                 if (!con->in_msg) {
1571                         con->error_msg =
1572                                 "error allocating memory for incoming message";
1573                         return -ENOMEM;
1574                 }
1575                 m = con->in_msg;
1576                 m->front.iov_len = 0;    /* haven't read it yet */
1577                 if (m->middle)
1578                         m->middle->vec.iov_len = 0;
1579
1580                 con->in_msg_pos.page = 0;
1581                 if (m->pages)
1582                         con->in_msg_pos.page_pos = m->page_alignment;
1583                 else
1584                         con->in_msg_pos.page_pos = 0;
1585                 con->in_msg_pos.data_pos = 0;
1586         }
1587
1588         /* front */
1589         ret = read_partial_message_section(con, &m->front, front_len,
1590                                            &con->in_front_crc);
1591         if (ret <= 0)
1592                 return ret;
1593
1594         /* middle */
1595         if (m->middle) {
1596                 ret = read_partial_message_section(con, &m->middle->vec,
1597                                                    middle_len,
1598                                                    &con->in_middle_crc);
1599                 if (ret <= 0)
1600                         return ret;
1601         }
1602 #ifdef CONFIG_BLOCK
1603         if (m->bio && !m->bio_iter)
1604                 init_bio_iter(m->bio, &m->bio_iter, &m->bio_seg);
1605 #endif
1606
1607         /* (page) data */
1608         while (con->in_msg_pos.data_pos < data_len) {
1609                 if (m->pages) {
1610                         ret = read_partial_message_pages(con, m->pages,
1611                                                  data_len, datacrc);
1612                         if (ret <= 0)
1613                                 return ret;
1614 #ifdef CONFIG_BLOCK
1615                 } else if (m->bio) {
1616
1617                         ret = read_partial_message_bio(con,
1618                                                  &m->bio_iter, &m->bio_seg,
1619                                                  data_len, datacrc);
1620                         if (ret <= 0)
1621                                 return ret;
1622 #endif
1623                 } else {
1624                         BUG_ON(1);
1625                 }
1626         }
1627
1628         /* footer */
1629         to = sizeof(m->hdr) + sizeof(m->footer);
1630         while (con->in_base_pos < to) {
1631                 left = to - con->in_base_pos;
1632                 ret = ceph_tcp_recvmsg(con->sock, (char *)&m->footer +
1633                                        (con->in_base_pos - sizeof(m->hdr)),
1634                                        left);
1635                 if (ret <= 0)
1636                         return ret;
1637                 con->in_base_pos += ret;
1638         }
1639         dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1640              m, front_len, m->footer.front_crc, middle_len,
1641              m->footer.middle_crc, data_len, m->footer.data_crc);
1642
1643         /* crc ok? */
1644         if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1645                 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1646                        m, con->in_front_crc, m->footer.front_crc);
1647                 return -EBADMSG;
1648         }
1649         if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1650                 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1651                        m, con->in_middle_crc, m->footer.middle_crc);
1652                 return -EBADMSG;
1653         }
1654         if (datacrc &&
1655             (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1656             con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1657                 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1658                        con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1659                 return -EBADMSG;
1660         }
1661
1662         return 1; /* done! */
1663 }
1664
1665 /*
1666  * Process message.  This happens in the worker thread.  The callback should
1667  * be careful not to do anything that waits on other incoming messages or it
1668  * may deadlock.
1669  */
1670 static void process_message(struct ceph_connection *con)
1671 {
1672         struct ceph_msg *msg;
1673
1674         msg = con->in_msg;
1675         con->in_msg = NULL;
1676
1677         /* if first message, set peer_name */
1678         if (con->peer_name.type == 0)
1679                 con->peer_name = msg->hdr.src;
1680
1681         con->in_seq++;
1682         mutex_unlock(&con->mutex);
1683
1684         dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1685              msg, le64_to_cpu(msg->hdr.seq),
1686              ENTITY_NAME(msg->hdr.src),
1687              le16_to_cpu(msg->hdr.type),
1688              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1689              le32_to_cpu(msg->hdr.front_len),
1690              le32_to_cpu(msg->hdr.data_len),
1691              con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1692         con->ops->dispatch(con, msg);
1693
1694         mutex_lock(&con->mutex);
1695         prepare_read_tag(con);
1696 }
1697
1698
1699 /*
1700  * Write something to the socket.  Called in a worker thread when the
1701  * socket appears to be writeable and we have something ready to send.
1702  */
1703 static int try_write(struct ceph_connection *con)
1704 {
1705         struct ceph_messenger *msgr = con->msgr;
1706         int ret = 1;
1707
1708         dout("try_write start %p state %lu nref %d\n", con, con->state,
1709              atomic_read(&con->nref));
1710
1711 more:
1712         dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1713
1714         /* open the socket first? */
1715         if (con->sock == NULL) {
1716                 /*
1717                  * if we were STANDBY and are reconnecting _this_
1718                  * connection, bump connect_seq now.  Always bump
1719                  * global_seq.
1720                  */
1721                 if (test_and_clear_bit(STANDBY, &con->state))
1722                         con->connect_seq++;
1723
1724                 prepare_write_banner(msgr, con);
1725                 prepare_write_connect(msgr, con, 1);
1726                 prepare_read_banner(con);
1727                 set_bit(CONNECTING, &con->state);
1728                 clear_bit(NEGOTIATING, &con->state);
1729
1730                 BUG_ON(con->in_msg);
1731                 con->in_tag = CEPH_MSGR_TAG_READY;
1732                 dout("try_write initiating connect on %p new state %lu\n",
1733                      con, con->state);
1734                 con->sock = ceph_tcp_connect(con);
1735                 if (IS_ERR(con->sock)) {
1736                         con->sock = NULL;
1737                         con->error_msg = "connect error";
1738                         ret = -1;
1739                         goto out;
1740                 }
1741         }
1742
1743 more_kvec:
1744         /* kvec data queued? */
1745         if (con->out_skip) {
1746                 ret = write_partial_skip(con);
1747                 if (ret <= 0)
1748                         goto out;
1749         }
1750         if (con->out_kvec_left) {
1751                 ret = write_partial_kvec(con);
1752                 if (ret <= 0)
1753                         goto out;
1754         }
1755
1756         /* msg pages? */
1757         if (con->out_msg) {
1758                 if (con->out_msg_done) {
1759                         ceph_msg_put(con->out_msg);
1760                         con->out_msg = NULL;   /* we're done with this one */
1761                         goto do_next;
1762                 }
1763
1764                 ret = write_partial_msg_pages(con);
1765                 if (ret == 1)
1766                         goto more_kvec;  /* we need to send the footer, too! */
1767                 if (ret == 0)
1768                         goto out;
1769                 if (ret < 0) {
1770                         dout("try_write write_partial_msg_pages err %d\n",
1771                              ret);
1772                         goto out;
1773                 }
1774         }
1775
1776 do_next:
1777         if (!test_bit(CONNECTING, &con->state)) {
1778                 /* is anything else pending? */
1779                 if (!list_empty(&con->out_queue)) {
1780                         prepare_write_message(con);
1781                         goto more;
1782                 }
1783                 if (con->in_seq > con->in_seq_acked) {
1784                         prepare_write_ack(con);
1785                         goto more;
1786                 }
1787                 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1788                         prepare_write_keepalive(con);
1789                         goto more;
1790                 }
1791         }
1792
1793         /* Nothing to do! */
1794         clear_bit(WRITE_PENDING, &con->state);
1795         dout("try_write nothing else to write.\n");
1796         ret = 0;
1797 out:
1798         dout("try_write done on %p ret %d\n", con, ret);
1799         return ret;
1800 }
1801
1802
1803
1804 /*
1805  * Read what we can from the socket.
1806  */
1807 static int try_read(struct ceph_connection *con)
1808 {
1809         int ret = -1;
1810
1811         if (!con->sock)
1812                 return 0;
1813
1814         if (test_bit(STANDBY, &con->state))
1815                 return 0;
1816
1817         dout("try_read start on %p\n", con);
1818
1819 more:
1820         dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1821              con->in_base_pos);
1822         if (test_bit(CONNECTING, &con->state)) {
1823                 if (!test_bit(NEGOTIATING, &con->state)) {
1824                         dout("try_read connecting\n");
1825                         ret = read_partial_banner(con);
1826                         if (ret <= 0)
1827                                 goto out;
1828                         ret = process_banner(con);
1829                         if (ret < 0)
1830                                 goto out;
1831                 }
1832                 ret = read_partial_connect(con);
1833                 if (ret <= 0)
1834                         goto out;
1835                 ret = process_connect(con);
1836                 if (ret < 0)
1837                         goto out;
1838                 goto more;
1839         }
1840
1841         if (con->in_base_pos < 0) {
1842                 /*
1843                  * skipping + discarding content.
1844                  *
1845                  * FIXME: there must be a better way to do this!
1846                  */
1847                 static char buf[1024];
1848                 int skip = min(1024, -con->in_base_pos);
1849                 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
1850                 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
1851                 if (ret <= 0)
1852                         goto out;
1853                 con->in_base_pos += ret;
1854                 if (con->in_base_pos)
1855                         goto more;
1856         }
1857         if (con->in_tag == CEPH_MSGR_TAG_READY) {
1858                 /*
1859                  * what's next?
1860                  */
1861                 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
1862                 if (ret <= 0)
1863                         goto out;
1864                 dout("try_read got tag %d\n", (int)con->in_tag);
1865                 switch (con->in_tag) {
1866                 case CEPH_MSGR_TAG_MSG:
1867                         prepare_read_message(con);
1868                         break;
1869                 case CEPH_MSGR_TAG_ACK:
1870                         prepare_read_ack(con);
1871                         break;
1872                 case CEPH_MSGR_TAG_CLOSE:
1873                         set_bit(CLOSED, &con->state);   /* fixme */
1874                         goto out;
1875                 default:
1876                         goto bad_tag;
1877                 }
1878         }
1879         if (con->in_tag == CEPH_MSGR_TAG_MSG) {
1880                 ret = read_partial_message(con);
1881                 if (ret <= 0) {
1882                         switch (ret) {
1883                         case -EBADMSG:
1884                                 con->error_msg = "bad crc";
1885                                 ret = -EIO;
1886                                 break;
1887                         case -EIO:
1888                                 con->error_msg = "io error";
1889                                 break;
1890                         }
1891                         goto out;
1892                 }
1893                 if (con->in_tag == CEPH_MSGR_TAG_READY)
1894                         goto more;
1895                 process_message(con);
1896                 goto more;
1897         }
1898         if (con->in_tag == CEPH_MSGR_TAG_ACK) {
1899                 ret = read_partial_ack(con);
1900                 if (ret <= 0)
1901                         goto out;
1902                 process_ack(con);
1903                 goto more;
1904         }
1905
1906 out:
1907         dout("try_read done on %p ret %d\n", con, ret);
1908         return ret;
1909
1910 bad_tag:
1911         pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
1912         con->error_msg = "protocol error, garbage tag";
1913         ret = -1;
1914         goto out;
1915 }
1916
1917
1918 /*
1919  * Atomically queue work on a connection.  Bump @con reference to
1920  * avoid races with connection teardown.
1921  */
1922 static void queue_con(struct ceph_connection *con)
1923 {
1924         if (test_bit(DEAD, &con->state)) {
1925                 dout("queue_con %p ignoring: DEAD\n",
1926                      con);
1927                 return;
1928         }
1929
1930         if (!con->ops->get(con)) {
1931                 dout("queue_con %p ref count 0\n", con);
1932                 return;
1933         }
1934
1935         if (!queue_delayed_work(ceph_msgr_wq, &con->work, 0)) {
1936                 dout("queue_con %p - already queued\n", con);
1937                 con->ops->put(con);
1938         } else {
1939                 dout("queue_con %p\n", con);
1940         }
1941 }
1942
1943 /*
1944  * Do some work on a connection.  Drop a connection ref when we're done.
1945  */
1946 static void con_work(struct work_struct *work)
1947 {
1948         struct ceph_connection *con = container_of(work, struct ceph_connection,
1949                                                    work.work);
1950
1951         mutex_lock(&con->mutex);
1952         if (test_and_clear_bit(BACKOFF, &con->state)) {
1953                 dout("con_work %p backing off\n", con);
1954                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
1955                                        round_jiffies_relative(con->delay))) {
1956                         dout("con_work %p backoff %lu\n", con, con->delay);
1957                         mutex_unlock(&con->mutex);
1958                         return;
1959                 } else {
1960                         con->ops->put(con);
1961                         dout("con_work %p FAILED to back off %lu\n", con,
1962                              con->delay);
1963                 }
1964         }
1965
1966         if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
1967                 dout("con_work CLOSED\n");
1968                 con_close_socket(con);
1969                 goto done;
1970         }
1971         if (test_and_clear_bit(OPENING, &con->state)) {
1972                 /* reopen w/ new peer */
1973                 dout("con_work OPENING\n");
1974                 con_close_socket(con);
1975         }
1976
1977         if (test_and_clear_bit(SOCK_CLOSED, &con->state) ||
1978             try_read(con) < 0 ||
1979             try_write(con) < 0) {
1980                 mutex_unlock(&con->mutex);
1981                 ceph_fault(con);     /* error/fault path */
1982                 goto done_unlocked;
1983         }
1984
1985 done:
1986         mutex_unlock(&con->mutex);
1987 done_unlocked:
1988         con->ops->put(con);
1989 }
1990
1991
1992 /*
1993  * Generic error/fault handler.  A retry mechanism is used with
1994  * exponential backoff
1995  */
1996 static void ceph_fault(struct ceph_connection *con)
1997 {
1998         pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
1999                ceph_pr_addr(&con->peer_addr.in_addr), con->error_msg);
2000         dout("fault %p state %lu to peer %s\n",
2001              con, con->state, ceph_pr_addr(&con->peer_addr.in_addr));
2002
2003         if (test_bit(LOSSYTX, &con->state)) {
2004                 dout("fault on LOSSYTX channel\n");
2005                 goto out;
2006         }
2007
2008         mutex_lock(&con->mutex);
2009         if (test_bit(CLOSED, &con->state))
2010                 goto out_unlock;
2011
2012         con_close_socket(con);
2013
2014         if (con->in_msg) {
2015                 ceph_msg_put(con->in_msg);
2016                 con->in_msg = NULL;
2017         }
2018
2019         /* Requeue anything that hasn't been acked */
2020         list_splice_init(&con->out_sent, &con->out_queue);
2021
2022         /* If there are no messages in the queue, place the connection
2023          * in a STANDBY state (i.e., don't try to reconnect just yet). */
2024         if (list_empty(&con->out_queue) && !con->out_keepalive_pending) {
2025                 dout("fault setting STANDBY\n");
2026                 set_bit(STANDBY, &con->state);
2027         } else {
2028                 /* retry after a delay. */
2029                 if (con->delay == 0)
2030                         con->delay = BASE_DELAY_INTERVAL;
2031                 else if (con->delay < MAX_DELAY_INTERVAL)
2032                         con->delay *= 2;
2033                 con->ops->get(con);
2034                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
2035                                        round_jiffies_relative(con->delay))) {
2036                         dout("fault queued %p delay %lu\n", con, con->delay);
2037                 } else {
2038                         con->ops->put(con);
2039                         dout("fault failed to queue %p delay %lu, backoff\n",
2040                              con, con->delay);
2041                         /*
2042                          * In many cases we see a socket state change
2043                          * while con_work is running and end up
2044                          * queuing (non-delayed) work, such that we
2045                          * can't backoff with a delay.  Set a flag so
2046                          * that when con_work restarts we schedule the
2047                          * delay then.
2048                          */
2049                         set_bit(BACKOFF, &con->state);
2050                 }
2051         }
2052
2053 out_unlock:
2054         mutex_unlock(&con->mutex);
2055 out:
2056         /*
2057          * in case we faulted due to authentication, invalidate our
2058          * current tickets so that we can get new ones.
2059          */
2060         if (con->auth_retry && con->ops->invalidate_authorizer) {
2061                 dout("calling invalidate_authorizer()\n");
2062                 con->ops->invalidate_authorizer(con);
2063         }
2064
2065         if (con->ops->fault)
2066                 con->ops->fault(con);
2067 }
2068
2069
2070
2071 /*
2072  * create a new messenger instance
2073  */
2074 struct ceph_messenger *ceph_messenger_create(struct ceph_entity_addr *myaddr,
2075                                              u32 supported_features,
2076                                              u32 required_features)
2077 {
2078         struct ceph_messenger *msgr;
2079
2080         msgr = kzalloc(sizeof(*msgr), GFP_KERNEL);
2081         if (msgr == NULL)
2082                 return ERR_PTR(-ENOMEM);
2083
2084         msgr->supported_features = supported_features;
2085         msgr->required_features = required_features;
2086
2087         spin_lock_init(&msgr->global_seq_lock);
2088
2089         /* the zero page is needed if a request is "canceled" while the message
2090          * is being written over the socket */
2091         msgr->zero_page = __page_cache_alloc(GFP_KERNEL | __GFP_ZERO);
2092         if (!msgr->zero_page) {
2093                 kfree(msgr);
2094                 return ERR_PTR(-ENOMEM);
2095         }
2096         kmap(msgr->zero_page);
2097
2098         if (myaddr)
2099                 msgr->inst.addr = *myaddr;
2100
2101         /* select a random nonce */
2102         msgr->inst.addr.type = 0;
2103         get_random_bytes(&msgr->inst.addr.nonce, sizeof(msgr->inst.addr.nonce));
2104         encode_my_addr(msgr);
2105
2106         dout("messenger_create %p\n", msgr);
2107         return msgr;
2108 }
2109 EXPORT_SYMBOL(ceph_messenger_create);
2110
2111 void ceph_messenger_destroy(struct ceph_messenger *msgr)
2112 {
2113         dout("destroy %p\n", msgr);
2114         kunmap(msgr->zero_page);
2115         __free_page(msgr->zero_page);
2116         kfree(msgr);
2117         dout("destroyed messenger %p\n", msgr);
2118 }
2119 EXPORT_SYMBOL(ceph_messenger_destroy);
2120
2121 /*
2122  * Queue up an outgoing message on the given connection.
2123  */
2124 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
2125 {
2126         if (test_bit(CLOSED, &con->state)) {
2127                 dout("con_send %p closed, dropping %p\n", con, msg);
2128                 ceph_msg_put(msg);
2129                 return;
2130         }
2131
2132         /* set src+dst */
2133         msg->hdr.src = con->msgr->inst.name;
2134
2135         BUG_ON(msg->front.iov_len != le32_to_cpu(msg->hdr.front_len));
2136
2137         msg->needs_out_seq = true;
2138
2139         /* queue */
2140         mutex_lock(&con->mutex);
2141         BUG_ON(!list_empty(&msg->list_head));
2142         list_add_tail(&msg->list_head, &con->out_queue);
2143         dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
2144              ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
2145              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
2146              le32_to_cpu(msg->hdr.front_len),
2147              le32_to_cpu(msg->hdr.middle_len),
2148              le32_to_cpu(msg->hdr.data_len));
2149         mutex_unlock(&con->mutex);
2150
2151         /* if there wasn't anything waiting to send before, queue
2152          * new work */
2153         if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2154                 queue_con(con);
2155 }
2156 EXPORT_SYMBOL(ceph_con_send);
2157
2158 /*
2159  * Revoke a message that was previously queued for send
2160  */
2161 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
2162 {
2163         mutex_lock(&con->mutex);
2164         if (!list_empty(&msg->list_head)) {
2165                 dout("con_revoke %p msg %p - was on queue\n", con, msg);
2166                 list_del_init(&msg->list_head);
2167                 ceph_msg_put(msg);
2168                 msg->hdr.seq = 0;
2169         }
2170         if (con->out_msg == msg) {
2171                 dout("con_revoke %p msg %p - was sending\n", con, msg);
2172                 con->out_msg = NULL;
2173                 if (con->out_kvec_is_msg) {
2174                         con->out_skip = con->out_kvec_bytes;
2175                         con->out_kvec_is_msg = false;
2176                 }
2177                 ceph_msg_put(msg);
2178                 msg->hdr.seq = 0;
2179         }
2180         mutex_unlock(&con->mutex);
2181 }
2182
2183 /*
2184  * Revoke a message that we may be reading data into
2185  */
2186 void ceph_con_revoke_message(struct ceph_connection *con, struct ceph_msg *msg)
2187 {
2188         mutex_lock(&con->mutex);
2189         if (con->in_msg && con->in_msg == msg) {
2190                 unsigned front_len = le32_to_cpu(con->in_hdr.front_len);
2191                 unsigned middle_len = le32_to_cpu(con->in_hdr.middle_len);
2192                 unsigned data_len = le32_to_cpu(con->in_hdr.data_len);
2193
2194                 /* skip rest of message */
2195                 dout("con_revoke_pages %p msg %p revoked\n", con, msg);
2196                         con->in_base_pos = con->in_base_pos -
2197                                 sizeof(struct ceph_msg_header) -
2198                                 front_len -
2199                                 middle_len -
2200                                 data_len -
2201                                 sizeof(struct ceph_msg_footer);
2202                 ceph_msg_put(con->in_msg);
2203                 con->in_msg = NULL;
2204                 con->in_tag = CEPH_MSGR_TAG_READY;
2205                 con->in_seq++;
2206         } else {
2207                 dout("con_revoke_pages %p msg %p pages %p no-op\n",
2208                      con, con->in_msg, msg);
2209         }
2210         mutex_unlock(&con->mutex);
2211 }
2212
2213 /*
2214  * Queue a keepalive byte to ensure the tcp connection is alive.
2215  */
2216 void ceph_con_keepalive(struct ceph_connection *con)
2217 {
2218         if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
2219             test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2220                 queue_con(con);
2221 }
2222 EXPORT_SYMBOL(ceph_con_keepalive);
2223
2224
2225 /*
2226  * construct a new message with given type, size
2227  * the new msg has a ref count of 1.
2228  */
2229 struct ceph_msg *ceph_msg_new(int type, int front_len, gfp_t flags)
2230 {
2231         struct ceph_msg *m;
2232
2233         m = kmalloc(sizeof(*m), flags);
2234         if (m == NULL)
2235                 goto out;
2236         kref_init(&m->kref);
2237         INIT_LIST_HEAD(&m->list_head);
2238
2239         m->hdr.tid = 0;
2240         m->hdr.type = cpu_to_le16(type);
2241         m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
2242         m->hdr.version = 0;
2243         m->hdr.front_len = cpu_to_le32(front_len);
2244         m->hdr.middle_len = 0;
2245         m->hdr.data_len = 0;
2246         m->hdr.data_off = 0;
2247         m->hdr.reserved = 0;
2248         m->footer.front_crc = 0;
2249         m->footer.middle_crc = 0;
2250         m->footer.data_crc = 0;
2251         m->footer.flags = 0;
2252         m->front_max = front_len;
2253         m->front_is_vmalloc = false;
2254         m->more_to_follow = false;
2255         m->pool = NULL;
2256
2257         /* front */
2258         if (front_len) {
2259                 if (front_len > PAGE_CACHE_SIZE) {
2260                         m->front.iov_base = __vmalloc(front_len, flags,
2261                                                       PAGE_KERNEL);
2262                         m->front_is_vmalloc = true;
2263                 } else {
2264                         m->front.iov_base = kmalloc(front_len, flags);
2265                 }
2266                 if (m->front.iov_base == NULL) {
2267                         pr_err("msg_new can't allocate %d bytes\n",
2268                              front_len);
2269                         goto out2;
2270                 }
2271         } else {
2272                 m->front.iov_base = NULL;
2273         }
2274         m->front.iov_len = front_len;
2275
2276         /* middle */
2277         m->middle = NULL;
2278
2279         /* data */
2280         m->nr_pages = 0;
2281         m->page_alignment = 0;
2282         m->pages = NULL;
2283         m->pagelist = NULL;
2284         m->bio = NULL;
2285         m->bio_iter = NULL;
2286         m->bio_seg = 0;
2287         m->trail = NULL;
2288
2289         dout("ceph_msg_new %p front %d\n", m, front_len);
2290         return m;
2291
2292 out2:
2293         ceph_msg_put(m);
2294 out:
2295         pr_err("msg_new can't create type %d front %d\n", type, front_len);
2296         return NULL;
2297 }
2298 EXPORT_SYMBOL(ceph_msg_new);
2299
2300 /*
2301  * Allocate "middle" portion of a message, if it is needed and wasn't
2302  * allocated by alloc_msg.  This allows us to read a small fixed-size
2303  * per-type header in the front and then gracefully fail (i.e.,
2304  * propagate the error to the caller based on info in the front) when
2305  * the middle is too large.
2306  */
2307 static int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2308 {
2309         int type = le16_to_cpu(msg->hdr.type);
2310         int middle_len = le32_to_cpu(msg->hdr.middle_len);
2311
2312         dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2313              ceph_msg_type_name(type), middle_len);
2314         BUG_ON(!middle_len);
2315         BUG_ON(msg->middle);
2316
2317         msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2318         if (!msg->middle)
2319                 return -ENOMEM;
2320         return 0;
2321 }
2322
2323 /*
2324  * Generic message allocator, for incoming messages.
2325  */
2326 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2327                                 struct ceph_msg_header *hdr,
2328                                 int *skip)
2329 {
2330         int type = le16_to_cpu(hdr->type);
2331         int front_len = le32_to_cpu(hdr->front_len);
2332         int middle_len = le32_to_cpu(hdr->middle_len);
2333         struct ceph_msg *msg = NULL;
2334         int ret;
2335
2336         if (con->ops->alloc_msg) {
2337                 mutex_unlock(&con->mutex);
2338                 msg = con->ops->alloc_msg(con, hdr, skip);
2339                 mutex_lock(&con->mutex);
2340                 if (!msg || *skip)
2341                         return NULL;
2342         }
2343         if (!msg) {
2344                 *skip = 0;
2345                 msg = ceph_msg_new(type, front_len, GFP_NOFS);
2346                 if (!msg) {
2347                         pr_err("unable to allocate msg type %d len %d\n",
2348                                type, front_len);
2349                         return NULL;
2350                 }
2351                 msg->page_alignment = le16_to_cpu(hdr->data_off);
2352         }
2353         memcpy(&msg->hdr, &con->in_hdr, sizeof(con->in_hdr));
2354
2355         if (middle_len && !msg->middle) {
2356                 ret = ceph_alloc_middle(con, msg);
2357                 if (ret < 0) {
2358                         ceph_msg_put(msg);
2359                         return NULL;
2360                 }
2361         }
2362
2363         return msg;
2364 }
2365
2366
2367 /*
2368  * Free a generically kmalloc'd message.
2369  */
2370 void ceph_msg_kfree(struct ceph_msg *m)
2371 {
2372         dout("msg_kfree %p\n", m);
2373         if (m->front_is_vmalloc)
2374                 vfree(m->front.iov_base);
2375         else
2376                 kfree(m->front.iov_base);
2377         kfree(m);
2378 }
2379
2380 /*
2381  * Drop a msg ref.  Destroy as needed.
2382  */
2383 void ceph_msg_last_put(struct kref *kref)
2384 {
2385         struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2386
2387         dout("ceph_msg_put last one on %p\n", m);
2388         WARN_ON(!list_empty(&m->list_head));
2389
2390         /* drop middle, data, if any */
2391         if (m->middle) {
2392                 ceph_buffer_put(m->middle);
2393                 m->middle = NULL;
2394         }
2395         m->nr_pages = 0;
2396         m->pages = NULL;
2397
2398         if (m->pagelist) {
2399                 ceph_pagelist_release(m->pagelist);
2400                 kfree(m->pagelist);
2401                 m->pagelist = NULL;
2402         }
2403
2404         m->trail = NULL;
2405
2406         if (m->pool)
2407                 ceph_msgpool_put(m->pool, m);
2408         else
2409                 ceph_msg_kfree(m);
2410 }
2411 EXPORT_SYMBOL(ceph_msg_last_put);
2412
2413 void ceph_msg_dump(struct ceph_msg *msg)
2414 {
2415         pr_debug("msg_dump %p (front_max %d nr_pages %d)\n", msg,
2416                  msg->front_max, msg->nr_pages);
2417         print_hex_dump(KERN_DEBUG, "header: ",
2418                        DUMP_PREFIX_OFFSET, 16, 1,
2419                        &msg->hdr, sizeof(msg->hdr), true);
2420         print_hex_dump(KERN_DEBUG, " front: ",
2421                        DUMP_PREFIX_OFFSET, 16, 1,
2422                        msg->front.iov_base, msg->front.iov_len, true);
2423         if (msg->middle)
2424                 print_hex_dump(KERN_DEBUG, "middle: ",
2425                                DUMP_PREFIX_OFFSET, 16, 1,
2426                                msg->middle->vec.iov_base,
2427                                msg->middle->vec.iov_len, true);
2428         print_hex_dump(KERN_DEBUG, "footer: ",
2429                        DUMP_PREFIX_OFFSET, 16, 1,
2430                        &msg->footer, sizeof(msg->footer), true);
2431 }
2432 EXPORT_SYMBOL(ceph_msg_dump);