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