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 <linux/dns_resolver.h>
  15 #include <net/tcp.h>
  16
  17 #include <linux/ceph/libceph.h>
  18 #include <linux/ceph/messenger.h>
  19 #include <linux/ceph/decode.h>
  20 #include <linux/ceph/pagelist.h>
  21
  22 /*
  23  * Ceph uses the messenger to exchange ceph_msg messages with other
  24  * hosts in the system.  The messenger provides ordered and reliable
  25  * delivery.  We tolerate TCP disconnects by reconnecting (with
  26  * exponential backoff) in the case of a fault (disconnection, bad
  27  * crc, protocol error).  Acks allow sent messages to be discarded by
  28  * the sender.
  29  */
  30
  31 /* static tag bytes (protocol control messages) */
  32 static char tag_msg = CEPH_MSGR_TAG_MSG;
  33 static char tag_ack = CEPH_MSGR_TAG_ACK;
  34 static char tag_keepalive = CEPH_MSGR_TAG_KEEPALIVE;
  35
  36 #ifdef CONFIG_LOCKDEP
  37 static struct lock_class_key socket_class;
  38 #endif
  39
  40
  41 static void queue_con(struct ceph_connection *con);
  42 static void con_work(struct work_struct *);
  43 static void ceph_fault(struct ceph_connection *con);
  44
  45 /*
  46  * nicely render a sockaddr as a string.
  47  */
  48 #define MAX_ADDR_STR 20
  49 #define MAX_ADDR_STR_LEN 60
  50 static char addr_str[MAX_ADDR_STR][MAX_ADDR_STR_LEN];
  51 static DEFINE_SPINLOCK(addr_str_lock);
  52 static int last_addr_str;
  53
  54 const char *ceph_pr_addr(const struct sockaddr_storage *ss)
  55 {
  56         int i;
  57         char *s;
  58         struct sockaddr_in *in4 = (void *)ss;
  59         struct sockaddr_in6 *in6 = (void *)ss;
  60
  61         spin_lock(&addr_str_lock);
  62         i = last_addr_str++;
  63         if (last_addr_str == MAX_ADDR_STR)
  64                 last_addr_str = 0;
  65         spin_unlock(&addr_str_lock);
  66         s = addr_str[i];
  67
  68         switch (ss->ss_family) {
  69         case AF_INET:
  70                 snprintf(s, MAX_ADDR_STR_LEN, "%pI4:%u", &in4->sin_addr,
  71                          (unsigned int)ntohs(in4->sin_port));
  72                 break;
  73
  74         case AF_INET6:
  75                 snprintf(s, MAX_ADDR_STR_LEN, "[%pI6c]:%u", &in6->sin6_addr,
  76                          (unsigned int)ntohs(in6->sin6_port));
  77                 break;
  78
  79         default:
  80                 snprintf(s, MAX_ADDR_STR_LEN, "(unknown sockaddr family %d)",
  81                          (int)ss->ss_family);
  82         }
  83
  84         return s;
  85 }
  86 EXPORT_SYMBOL(ceph_pr_addr);
  87
  88 static void encode_my_addr(struct ceph_messenger *msgr)
  89 {
  90         memcpy(&msgr->my_enc_addr, &msgr->inst.addr, sizeof(msgr->my_enc_addr));
  91         ceph_encode_addr(&msgr->my_enc_addr);
  92 }
  93
  94 /*
  95  * work queue for all reading and writing to/from the socket.
  96  */
  97 struct workqueue_struct *ceph_msgr_wq;
  98
  99 int ceph_msgr_init(void)
 100 {
 101         ceph_msgr_wq = alloc_workqueue("ceph-msgr", WQ_NON_REENTRANT, 0);
 102         if (!ceph_msgr_wq) {
 103                 pr_err("msgr_init failed to create workqueue\n");
 104                 return -ENOMEM;
 105         }
 106         return 0;
 107 }
 108 EXPORT_SYMBOL(ceph_msgr_init);
 109
 110 void ceph_msgr_exit(void)
 111 {
 112         destroy_workqueue(ceph_msgr_wq);
 113 }
 114 EXPORT_SYMBOL(ceph_msgr_exit);
 115
 116 void ceph_msgr_flush(void)
 117 {
 118         flush_workqueue(ceph_msgr_wq);
 119 }
 120 EXPORT_SYMBOL(ceph_msgr_flush);
 121
 122
 123 /*
 124  * socket callback functions
 125  */
 126
 127 /* data available on socket, or listen socket received a connect */
 128 static void ceph_data_ready(struct sock *sk, int count_unused)
 129 {
 130         struct ceph_connection *con =
 131                 (struct ceph_connection *)sk->sk_user_data;
 132         if (sk->sk_state != TCP_CLOSE_WAIT) {
 133                 dout("ceph_data_ready on %p state = %lu, queueing work\n",
 134                      con, con->state);
 135                 queue_con(con);
 136         }
 137 }
 138
 139 /* socket has buffer space for writing */
 140 static void ceph_write_space(struct sock *sk)
 141 {
 142         struct ceph_connection *con =
 143                 (struct ceph_connection *)sk->sk_user_data;
 144
 145         /* only queue to workqueue if there is data we want to write. */
 146         if (test_bit(WRITE_PENDING, &con->state)) {
 147                 dout("ceph_write_space %p queueing write work\n", con);
 148                 queue_con(con);
 149         } else {
 150                 dout("ceph_write_space %p nothing to write\n", con);
 151         }
 152
 153         /* since we have our own write_space, clear the SOCK_NOSPACE flag */
 154         clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 155 }
 156
 157 /* socket's state has changed */
 158 static void ceph_state_change(struct sock *sk)
 159 {
 160         struct ceph_connection *con =
 161                 (struct ceph_connection *)sk->sk_user_data;
 162
 163         dout("ceph_state_change %p state = %lu sk_state = %u\n",
 164              con, con->state, sk->sk_state);
 165
 166         if (test_bit(CLOSED, &con->state))
 167                 return;
 168
 169         switch (sk->sk_state) {
 170         case TCP_CLOSE:
 171                 dout("ceph_state_change TCP_CLOSE\n");
 172         case TCP_CLOSE_WAIT:
 173                 dout("ceph_state_change TCP_CLOSE_WAIT\n");
 174                 if (test_and_set_bit(SOCK_CLOSED, &con->state) == 0) {
 175                         if (test_bit(CONNECTING, &con->state))
 176                                 con->error_msg = "connection failed";
 177                         else
 178                                 con->error_msg = "socket closed";
 179                         queue_con(con);
 180                 }
 181                 break;
 182         case TCP_ESTABLISHED:
 183                 dout("ceph_state_change TCP_ESTABLISHED\n");
 184                 queue_con(con);
 185                 break;
 186         }
 187 }
 188
 189 /*
 190  * set up socket callbacks
 191  */
 192 static void set_sock_callbacks(struct socket *sock,
 193                                struct ceph_connection *con)
 194 {
 195         struct sock *sk = sock->sk;
 196         sk->sk_user_data = (void *)con;
 197         sk->sk_data_ready = ceph_data_ready;
 198         sk->sk_write_space = ceph_write_space;
 199         sk->sk_state_change = ceph_state_change;
 200 }
 201
 202
 203 /*
 204  * socket helpers
 205  */
 206
 207 /*
 208  * initiate connection to a remote socket.
 209  */
 210 static struct socket *ceph_tcp_connect(struct ceph_connection *con)
 211 {
 212         struct sockaddr_storage *paddr = &con->peer_addr.in_addr;
 213         struct socket *sock;
 214         int ret;
 215
 216         BUG_ON(con->sock);
 217         ret = sock_create_kern(con->peer_addr.in_addr.ss_family, SOCK_STREAM,
 218                                IPPROTO_TCP, &sock);
 219         if (ret)
 220                 return ERR_PTR(ret);
 221         con->sock = sock;
 222         sock->sk->sk_allocation = GFP_NOFS;
 223
 224 #ifdef CONFIG_LOCKDEP
 225         lockdep_set_class(&sock->sk->sk_lock, &socket_class);
 226 #endif
 227
 228         set_sock_callbacks(sock, con);
 229
 230         dout("connect %s\n", ceph_pr_addr(&con->peer_addr.in_addr));
 231
 232         ret = sock->ops->connect(sock, (struct sockaddr *)paddr, sizeof(*paddr),
 233                                  O_NONBLOCK);
 234         if (ret == -EINPROGRESS) {
 235                 dout("connect %s EINPROGRESS sk_state = %u\n",
 236                      ceph_pr_addr(&con->peer_addr.in_addr),
 237                      sock->sk->sk_state);
 238                 ret = 0;
 239         }
 240         if (ret < 0) {
 241                 pr_err("connect %s error %d\n",
 242                        ceph_pr_addr(&con->peer_addr.in_addr), ret);
 243                 sock_release(sock);
 244                 con->sock = NULL;
 245                 con->error_msg = "connect error";
 246         }
 247
 248         if (ret < 0)
 249                 return ERR_PTR(ret);
 250         return sock;
 251 }
 252
 253 static int ceph_tcp_recvmsg(struct socket *sock, void *buf, size_t len)
 254 {
 255         struct kvec iov = {buf, len};
 256         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
 257         int r;
 258
 259         r = kernel_recvmsg(sock, &msg, &iov, 1, len, msg.msg_flags);
 260         if (r == -EAGAIN)
 261                 r = 0;
 262         return r;
 263 }
 264
 265 /*
 266  * write something.  @more is true if caller will be sending more data
 267  * shortly.
 268  */
 269 static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov,
 270                      size_t kvlen, size_t len, int more)
 271 {
 272         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
 273         int r;
 274
 275         if (more)
 276                 msg.msg_flags |= MSG_MORE;
 277         else
 278                 msg.msg_flags |= MSG_EOR;  /* superfluous, but what the hell */
 279
 280         r = kernel_sendmsg(sock, &msg, iov, kvlen, len);
 281         if (r == -EAGAIN)
 282                 r = 0;
 283         return r;
 284 }
 285
 286
 287 /*
 288  * Shutdown/close the socket for the given connection.
 289  */
 290 static int con_close_socket(struct ceph_connection *con)
 291 {
 292         int rc;
 293
 294         dout("con_close_socket on %p sock %p\n", con, con->sock);
 295         if (!con->sock)
 296                 return 0;
 297         set_bit(SOCK_CLOSED, &con->state);
 298         rc = con->sock->ops->shutdown(con->sock, SHUT_RDWR);
 299         sock_release(con->sock);
 300         con->sock = NULL;
 301         clear_bit(SOCK_CLOSED, &con->state);
 302         return rc;
 303 }
 304
 305 /*
 306  * Reset a connection.  Discard all incoming and outgoing messages
 307  * and clear *_seq state.
 308  */
 309 static void ceph_msg_remove(struct ceph_msg *msg)
 310 {
 311         list_del_init(&msg->list_head);
 312         ceph_msg_put(msg);
 313 }
 314 static void ceph_msg_remove_list(struct list_head *head)
 315 {
 316         while (!list_empty(head)) {
 317                 struct ceph_msg *msg = list_first_entry(head, struct ceph_msg,
 318                                                         list_head);
 319                 ceph_msg_remove(msg);
 320         }
 321 }
 322
 323 static void reset_connection(struct ceph_connection *con)
 324 {
 325         /* reset connection, out_queue, msg_ and connect_seq */
 326         /* discard existing out_queue and msg_seq */
 327         ceph_msg_remove_list(&con->out_queue);
 328         ceph_msg_remove_list(&con->out_sent);
 329
 330         if (con->in_msg) {
 331                 ceph_msg_put(con->in_msg);
 332                 con->in_msg = NULL;
 333         }
 334
 335         con->connect_seq = 0;
 336         con->out_seq = 0;
 337         if (con->out_msg) {
 338                 ceph_msg_put(con->out_msg);
 339                 con->out_msg = NULL;
 340         }
 341         con->in_seq = 0;
 342         con->in_seq_acked = 0;
 343 }
 344
 345 /*
 346  * mark a peer down.  drop any open connections.
 347  */
 348 void ceph_con_close(struct ceph_connection *con)
 349 {
 350         dout("con_close %p peer %s\n", con,
 351              ceph_pr_addr(&con->peer_addr.in_addr));
 352         set_bit(CLOSED, &con->state);  /* in case there's queued work */
 353         clear_bit(STANDBY, &con->state);  /* avoid connect_seq bump */
 354         clear_bit(LOSSYTX, &con->state);  /* so we retry next connect */
 355         clear_bit(KEEPALIVE_PENDING, &con->state);
 356         clear_bit(WRITE_PENDING, &con->state);
 357         mutex_lock(&con->mutex);
 358         reset_connection(con);
 359         con->peer_global_seq = 0;
 360         cancel_delayed_work(&con->work);
 361         mutex_unlock(&con->mutex);
 362         queue_con(con);
 363 }
 364 EXPORT_SYMBOL(ceph_con_close);
 365
 366 /*
 367  * Reopen a closed connection, with a new peer address.
 368  */
 369 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
 370 {
 371         dout("con_open %p %s\n", con, ceph_pr_addr(&addr->in_addr));
 372         set_bit(OPENING, &con->state);
 373         clear_bit(CLOSED, &con->state);
 374         memcpy(&con->peer_addr, addr, sizeof(*addr));
 375         con->delay = 0;      /* reset backoff memory */
 376         queue_con(con);
 377 }
 378 EXPORT_SYMBOL(ceph_con_open);
 379
 380 /*
 381  * return true if this connection ever successfully opened
 382  */
 383 bool ceph_con_opened(struct ceph_connection *con)
 384 {
 385         return con->connect_seq > 0;
 386 }
 387
 388 /*
 389  * generic get/put
 390  */
 391 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
 392 {
 393         dout("con_get %p nref = %d -> %d\n", con,
 394              atomic_read(&con->nref), atomic_read(&con->nref) + 1);
 395         if (atomic_inc_not_zero(&con->nref))
 396                 return con;
 397         return NULL;
 398 }
 399
 400 void ceph_con_put(struct ceph_connection *con)
 401 {
 402         dout("con_put %p nref = %d -> %d\n", con,
 403              atomic_read(&con->nref), atomic_read(&con->nref) - 1);
 404         BUG_ON(atomic_read(&con->nref) == 0);
 405         if (atomic_dec_and_test(&con->nref)) {
 406                 BUG_ON(con->sock);
 407                 kfree(con);
 408         }
 409 }
 410
 411 /*
 412  * initialize a new connection.
 413  */
 414 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
 415 {
 416         dout("con_init %p\n", con);
 417         memset(con, 0, sizeof(*con));
 418         atomic_set(&con->nref, 1);
 419         con->msgr = msgr;
 420         mutex_init(&con->mutex);
 421         INIT_LIST_HEAD(&con->out_queue);
 422         INIT_LIST_HEAD(&con->out_sent);
 423         INIT_DELAYED_WORK(&con->work, con_work);
 424 }
 425 EXPORT_SYMBOL(ceph_con_init);
 426
 427
 428 /*
 429  * We maintain a global counter to order connection attempts.  Get
 430  * a unique seq greater than @gt.
 431  */
 432 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
 433 {
 434         u32 ret;
 435
 436         spin_lock(&msgr->global_seq_lock);
 437         if (msgr->global_seq < gt)
 438                 msgr->global_seq = gt;
 439         ret = ++msgr->global_seq;
 440         spin_unlock(&msgr->global_seq_lock);
 441         return ret;
 442 }
 443
 444
 445 /*
 446  * Prepare footer for currently outgoing message, and finish things
 447  * off.  Assumes out_kvec* are already valid.. we just add on to the end.
 448  */
 449 static void prepare_write_message_footer(struct ceph_connection *con, int v)
 450 {
 451         struct ceph_msg *m = con->out_msg;
 452
 453         dout("prepare_write_message_footer %p\n", con);
 454         con->out_kvec_is_msg = true;
 455         con->out_kvec[v].iov_base = &m->footer;
 456         con->out_kvec[v].iov_len = sizeof(m->footer);
 457         con->out_kvec_bytes += sizeof(m->footer);
 458         con->out_kvec_left++;
 459         con->out_more = m->more_to_follow;
 460         con->out_msg_done = true;
 461 }
 462
 463 /*
 464  * Prepare headers for the next outgoing message.
 465  */
 466 static void prepare_write_message(struct ceph_connection *con)
 467 {
 468         struct ceph_msg *m;
 469         int v = 0;
 470
 471         con->out_kvec_bytes = 0;
 472         con->out_kvec_is_msg = true;
 473         con->out_msg_done = false;
 474
 475         /* Sneak an ack in there first?  If we can get it into the same
 476          * TCP packet that's a good thing. */
 477         if (con->in_seq > con->in_seq_acked) {
 478                 con->in_seq_acked = con->in_seq;
 479                 con->out_kvec[v].iov_base = &tag_ack;
 480                 con->out_kvec[v++].iov_len = 1;
 481                 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 482                 con->out_kvec[v].iov_base = &con->out_temp_ack;
 483                 con->out_kvec[v++].iov_len = sizeof(con->out_temp_ack);
 484                 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 485         }
 486
 487         m = list_first_entry(&con->out_queue,
 488                        struct ceph_msg, list_head);
 489         con->out_msg = m;
 490
 491         /* put message on sent list */
 492         ceph_msg_get(m);
 493         list_move_tail(&m->list_head, &con->out_sent);
 494
 495         /*
 496          * only assign outgoing seq # if we haven't sent this message
 497          * yet.  if it is requeued, resend with it's original seq.
 498          */
 499         if (m->needs_out_seq) {
 500                 m->hdr.seq = cpu_to_le64(++con->out_seq);
 501                 m->needs_out_seq = false;
 502         }
 503
 504         dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
 505              m, con->out_seq, le16_to_cpu(m->hdr.type),
 506              le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
 507              le32_to_cpu(m->hdr.data_len),
 508              m->nr_pages);
 509         BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
 510
 511         /* tag + hdr + front + middle */
 512         con->out_kvec[v].iov_base = &tag_msg;
 513         con->out_kvec[v++].iov_len = 1;
 514         con->out_kvec[v].iov_base = &m->hdr;
 515         con->out_kvec[v++].iov_len = sizeof(m->hdr);
 516         con->out_kvec[v++] = m->front;
 517         if (m->middle)
 518                 con->out_kvec[v++] = m->middle->vec;
 519         con->out_kvec_left = v;
 520         con->out_kvec_bytes += 1 + sizeof(m->hdr) + m->front.iov_len +
 521                 (m->middle ? m->middle->vec.iov_len : 0);
 522         con->out_kvec_cur = con->out_kvec;
 523
 524         /* fill in crc (except data pages), footer */
 525         con->out_msg->hdr.crc =
 526                 cpu_to_le32(crc32c(0, (void *)&m->hdr,
 527                                       sizeof(m->hdr) - sizeof(m->hdr.crc)));
 528         con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
 529         con->out_msg->footer.front_crc =
 530                 cpu_to_le32(crc32c(0, m->front.iov_base, m->front.iov_len));
 531         if (m->middle)
 532                 con->out_msg->footer.middle_crc =
 533                         cpu_to_le32(crc32c(0, m->middle->vec.iov_base,
 534                                            m->middle->vec.iov_len));
 535         else
 536                 con->out_msg->footer.middle_crc = 0;
 537         con->out_msg->footer.data_crc = 0;
 538         dout("prepare_write_message front_crc %u data_crc %u\n",
 539              le32_to_cpu(con->out_msg->footer.front_crc),
 540              le32_to_cpu(con->out_msg->footer.middle_crc));
 541
 542         /* is there a data payload? */
 543         if (le32_to_cpu(m->hdr.data_len) > 0) {
 544                 /* initialize page iterator */
 545                 con->out_msg_pos.page = 0;
 546                 if (m->pages)
 547                         con->out_msg_pos.page_pos = m->page_alignment;
 548                 else
 549                         con->out_msg_pos.page_pos = 0;
 550                 con->out_msg_pos.data_pos = 0;
 551                 con->out_msg_pos.did_page_crc = 0;
 552                 con->out_more = 1;  /* data + footer will follow */
 553         } else {
 554                 /* no, queue up footer too and be done */
 555                 prepare_write_message_footer(con, v);
 556         }
 557
 558         set_bit(WRITE_PENDING, &con->state);
 559 }
 560
 561 /*
 562  * Prepare an ack.
 563  */
 564 static void prepare_write_ack(struct ceph_connection *con)
 565 {
 566         dout("prepare_write_ack %p %llu -> %llu\n", con,
 567              con->in_seq_acked, con->in_seq);
 568         con->in_seq_acked = con->in_seq;
 569
 570         con->out_kvec[0].iov_base = &tag_ack;
 571         con->out_kvec[0].iov_len = 1;
 572         con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 573         con->out_kvec[1].iov_base = &con->out_temp_ack;
 574         con->out_kvec[1].iov_len = sizeof(con->out_temp_ack);
 575         con->out_kvec_left = 2;
 576         con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 577         con->out_kvec_cur = con->out_kvec;
 578         con->out_more = 1;  /* more will follow.. eventually.. */
 579         set_bit(WRITE_PENDING, &con->state);
 580 }
 581
 582 /*
 583  * Prepare to write keepalive byte.
 584  */
 585 static void prepare_write_keepalive(struct ceph_connection *con)
 586 {
 587         dout("prepare_write_keepalive %p\n", con);
 588         con->out_kvec[0].iov_base = &tag_keepalive;
 589         con->out_kvec[0].iov_len = 1;
 590         con->out_kvec_left = 1;
 591         con->out_kvec_bytes = 1;
 592         con->out_kvec_cur = con->out_kvec;
 593         set_bit(WRITE_PENDING, &con->state);
 594 }
 595
 596 /*
 597  * Connection negotiation.
 598  */
 599
 600 static int prepare_connect_authorizer(struct ceph_connection *con)
 601 {
 602         void *auth_buf;
 603         int auth_len = 0;
 604         int auth_protocol = 0;
 605
 606         mutex_unlock(&con->mutex);
 607         if (con->ops->get_authorizer)
 608                 con->ops->get_authorizer(con, &auth_buf, &auth_len,
 609                                          &auth_protocol, &con->auth_reply_buf,
 610                                          &con->auth_reply_buf_len,
 611                                          con->auth_retry);
 612         mutex_lock(&con->mutex);
 613
 614         if (test_bit(CLOSED, &con->state) ||
 615             test_bit(OPENING, &con->state))
 616                 return -EAGAIN;
 617
 618         con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
 619         con->out_connect.authorizer_len = cpu_to_le32(auth_len);
 620
 621         if (auth_len) {
 622                 con->out_kvec[con->out_kvec_left].iov_base = auth_buf;
 623                 con->out_kvec[con->out_kvec_left].iov_len = auth_len;
 624                 con->out_kvec_left++;
 625                 con->out_kvec_bytes += auth_len;
 626         }
 627         return 0;
 628 }
 629
 630 /*
 631  * We connected to a peer and are saying hello.
 632  */
 633 static void prepare_write_banner(struct ceph_messenger *msgr,
 634                                  struct ceph_connection *con)
 635 {
 636         int len = strlen(CEPH_BANNER);
 637
 638         con->out_kvec[0].iov_base = CEPH_BANNER;
 639         con->out_kvec[0].iov_len = len;
 640         con->out_kvec[1].iov_base = &msgr->my_enc_addr;
 641         con->out_kvec[1].iov_len = sizeof(msgr->my_enc_addr);
 642         con->out_kvec_left = 2;
 643         con->out_kvec_bytes = len + sizeof(msgr->my_enc_addr);
 644         con->out_kvec_cur = con->out_kvec;
 645         con->out_more = 0;
 646         set_bit(WRITE_PENDING, &con->state);
 647 }
 648
 649 static int prepare_write_connect(struct ceph_messenger *msgr,
 650                                  struct ceph_connection *con,
 651                                  int after_banner)
 652 {
 653         unsigned global_seq = get_global_seq(con->msgr, 0);
 654         int proto;
 655
 656         switch (con->peer_name.type) {
 657         case CEPH_ENTITY_TYPE_MON:
 658                 proto = CEPH_MONC_PROTOCOL;
 659                 break;
 660         case CEPH_ENTITY_TYPE_OSD:
 661                 proto = CEPH_OSDC_PROTOCOL;
 662                 break;
 663         case CEPH_ENTITY_TYPE_MDS:
 664                 proto = CEPH_MDSC_PROTOCOL;
 665                 break;
 666         default:
 667                 BUG();
 668         }
 669
 670         dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
 671              con->connect_seq, global_seq, proto);
 672
 673         con->out_connect.features = cpu_to_le64(msgr->supported_features);
 674         con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
 675         con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
 676         con->out_connect.global_seq = cpu_to_le32(global_seq);
 677         con->out_connect.protocol_version = cpu_to_le32(proto);
 678         con->out_connect.flags = 0;
 679
 680         if (!after_banner) {
 681                 con->out_kvec_left = 0;
 682                 con->out_kvec_bytes = 0;
 683         }
 684         con->out_kvec[con->out_kvec_left].iov_base = &con->out_connect;
 685         con->out_kvec[con->out_kvec_left].iov_len = sizeof(con->out_connect);
 686         con->out_kvec_left++;
 687         con->out_kvec_bytes += sizeof(con->out_connect);
 688         con->out_kvec_cur = con->out_kvec;
 689         con->out_more = 0;
 690         set_bit(WRITE_PENDING, &con->state);
 691
 692         return prepare_connect_authorizer(con);
 693 }
 694
 695
 696 /*
 697  * write as much of pending kvecs to the socket as we can.
 698  *  1 -> done
 699  *  0 -> socket full, but more to do
 700  * <0 -> error
 701  */
 702 static int write_partial_kvec(struct ceph_connection *con)
 703 {
 704         int ret;
 705
 706         dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
 707         while (con->out_kvec_bytes > 0) {
 708                 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
 709                                        con->out_kvec_left, con->out_kvec_bytes,
 710                                        con->out_more);
 711                 if (ret <= 0)
 712                         goto out;
 713                 con->out_kvec_bytes -= ret;
 714                 if (con->out_kvec_bytes == 0)
 715                         break;            /* done */
 716                 while (ret > 0) {
 717                         if (ret >= con->out_kvec_cur->iov_len) {
 718                                 ret -= con->out_kvec_cur->iov_len;
 719                                 con->out_kvec_cur++;
 720                                 con->out_kvec_left--;
 721                         } else {
 722                                 con->out_kvec_cur->iov_len -= ret;
 723                                 con->out_kvec_cur->iov_base += ret;
 724                                 ret = 0;
 725                                 break;
 726                         }
 727                 }
 728         }
 729         con->out_kvec_left = 0;
 730         con->out_kvec_is_msg = false;
 731         ret = 1;
 732 out:
 733         dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
 734              con->out_kvec_bytes, con->out_kvec_left, ret);
 735         return ret;  /* done! */
 736 }
 737
 738 #ifdef CONFIG_BLOCK
 739 static void init_bio_iter(struct bio *bio, struct bio **iter, int *seg)
 740 {
 741         if (!bio) {
 742                 *iter = NULL;
 743                 *seg = 0;
 744                 return;
 745         }
 746         *iter = bio;
 747         *seg = bio->bi_idx;
 748 }
 749
 750 static void iter_bio_next(struct bio **bio_iter, int *seg)
 751 {
 752         if (*bio_iter == NULL)
 753                 return;
 754
 755         BUG_ON(*seg >= (*bio_iter)->bi_vcnt);
 756
 757         (*seg)++;
 758         if (*seg == (*bio_iter)->bi_vcnt)
 759                 init_bio_iter((*bio_iter)->bi_next, bio_iter, seg);
 760 }
 761 #endif
 762
 763 /*
 764  * Write as much message data payload as we can.  If we finish, queue
 765  * up the footer.
 766  *  1 -> done, footer is now queued in out_kvec[].
 767  *  0 -> socket full, but more to do
 768  * <0 -> error
 769  */
 770 static int write_partial_msg_pages(struct ceph_connection *con)
 771 {
 772         struct ceph_msg *msg = con->out_msg;
 773         unsigned data_len = le32_to_cpu(msg->hdr.data_len);
 774         size_t len;
 775         int crc = con->msgr->nocrc;
 776         int ret;
 777         int total_max_write;
 778         int in_trail = 0;
 779         size_t trail_len = (msg->trail ? msg->trail->length : 0);
 780
 781         dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
 782              con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
 783              con->out_msg_pos.page_pos);
 784
 785 #ifdef CONFIG_BLOCK
 786         if (msg->bio && !msg->bio_iter)
 787                 init_bio_iter(msg->bio, &msg->bio_iter, &msg->bio_seg);
 788 #endif
 789
 790         while (data_len > con->out_msg_pos.data_pos) {
 791                 struct page *page = NULL;
 792                 void *kaddr = NULL;
 793                 int max_write = PAGE_SIZE;
 794                 int page_shift = 0;
 795
 796                 total_max_write = data_len - trail_len -
 797                         con->out_msg_pos.data_pos;
 798
 799                 /*
 800                  * if we are calculating the data crc (the default), we need
 801                  * to map the page.  if our pages[] has been revoked, use the
 802                  * zero page.
 803                  */
 804
 805                 /* have we reached the trail part of the data? */
 806                 if (con->out_msg_pos.data_pos >= data_len - trail_len) {
 807                         in_trail = 1;
 808
 809                         total_max_write = data_len - con->out_msg_pos.data_pos;
 810
 811                         page = list_first_entry(&msg->trail->head,
 812                                                 struct page, lru);
 813                         if (crc)
 814                                 kaddr = kmap(page);
 815                         max_write = PAGE_SIZE;
 816                 } else if (msg->pages) {
 817                         page = msg->pages[con->out_msg_pos.page];
 818                         if (crc)
 819                                 kaddr = kmap(page);
 820                 } else if (msg->pagelist) {
 821                         page = list_first_entry(&msg->pagelist->head,
 822                                                 struct page, lru);
 823                         if (crc)
 824                                 kaddr = kmap(page);
 825 #ifdef CONFIG_BLOCK
 826                 } else if (msg->bio) {
 827                         struct bio_vec *bv;
 828
 829                         bv = bio_iovec_idx(msg->bio_iter, msg->bio_seg);
 830                         page = bv->bv_page;
 831                         page_shift = bv->bv_offset;
 832                         if (crc)
 833                                 kaddr = kmap(page) + page_shift;
 834                         max_write = bv->bv_len;
 835 #endif
 836                 } else {
 837                         page = con->msgr->zero_page;
 838                         if (crc)
 839                                 kaddr = page_address(con->msgr->zero_page);
 840                 }
 841                 len = min_t(int, max_write - con->out_msg_pos.page_pos,
 842                             total_max_write);
 843
 844                 if (crc && !con->out_msg_pos.did_page_crc) {
 845                         void *base = kaddr + con->out_msg_pos.page_pos;
 846                         u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
 847
 848                         BUG_ON(kaddr == NULL);
 849                         con->out_msg->footer.data_crc =
 850                                 cpu_to_le32(crc32c(tmpcrc, base, len));
 851                         con->out_msg_pos.did_page_crc = 1;
 852                 }
 853                 ret = kernel_sendpage(con->sock, page,
 854                                       con->out_msg_pos.page_pos + page_shift,
 855                                       len,
 856                                       MSG_DONTWAIT | MSG_NOSIGNAL |
 857                                       MSG_MORE);
 858
 859                 if (crc &&
 860                     (msg->pages || msg->pagelist || msg->bio || in_trail))
 861                         kunmap(page);
 862
 863                 if (ret == -EAGAIN)
 864                         ret = 0;
 865                 if (ret <= 0)
 866                         goto out;
 867
 868                 con->out_msg_pos.data_pos += ret;
 869                 con->out_msg_pos.page_pos += ret;
 870                 if (ret == len) {
 871                         con->out_msg_pos.page_pos = 0;
 872                         con->out_msg_pos.page++;
 873                         con->out_msg_pos.did_page_crc = 0;
 874                         if (in_trail)
 875                                 list_move_tail(&page->lru,
 876                                                &msg->trail->head);
 877                         else if (msg->pagelist)
 878                                 list_move_tail(&page->lru,
 879                                                &msg->pagelist->head);
 880 #ifdef CONFIG_BLOCK
 881                         else if (msg->bio)
 882                                 iter_bio_next(&msg->bio_iter, &msg->bio_seg);
 883 #endif
 884                 }
 885         }
 886
 887         dout("write_partial_msg_pages %p msg %p done\n", con, msg);
 888
 889         /* prepare and queue up footer, too */
 890         if (!crc)
 891                 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
 892         con->out_kvec_bytes = 0;
 893         con->out_kvec_left = 0;
 894         con->out_kvec_cur = con->out_kvec;
 895         prepare_write_message_footer(con, 0);
 896         ret = 1;
 897 out:
 898         return ret;
 899 }
 900
 901 /*
 902  * write some zeros
 903  */
 904 static int write_partial_skip(struct ceph_connection *con)
 905 {
 906         int ret;
 907
 908         while (con->out_skip > 0) {
 909                 struct kvec iov = {
 910                         .iov_base = page_address(con->msgr->zero_page),
 911                         .iov_len = min(con->out_skip, (int)PAGE_CACHE_SIZE)
 912                 };
 913
 914                 ret = ceph_tcp_sendmsg(con->sock, &iov, 1, iov.iov_len, 1);
 915                 if (ret <= 0)
 916                         goto out;
 917                 con->out_skip -= ret;
 918         }
 919         ret = 1;
 920 out:
 921         return ret;
 922 }
 923
 924 /*
 925  * Prepare to read connection handshake, or an ack.
 926  */
 927 static void prepare_read_banner(struct ceph_connection *con)
 928 {
 929         dout("prepare_read_banner %p\n", con);
 930         con->in_base_pos = 0;
 931 }
 932
 933 static void prepare_read_connect(struct ceph_connection *con)
 934 {
 935         dout("prepare_read_connect %p\n", con);
 936         con->in_base_pos = 0;
 937 }
 938
 939 static void prepare_read_ack(struct ceph_connection *con)
 940 {
 941         dout("prepare_read_ack %p\n", con);
 942         con->in_base_pos = 0;
 943 }
 944
 945 static void prepare_read_tag(struct ceph_connection *con)
 946 {
 947         dout("prepare_read_tag %p\n", con);
 948         con->in_base_pos = 0;
 949         con->in_tag = CEPH_MSGR_TAG_READY;
 950 }
 951
 952 /*
 953  * Prepare to read a message.
 954  */
 955 static int prepare_read_message(struct ceph_connection *con)
 956 {
 957         dout("prepare_read_message %p\n", con);
 958         BUG_ON(con->in_msg != NULL);
 959         con->in_base_pos = 0;
 960         con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
 961         return 0;
 962 }
 963
 964
 965 static int read_partial(struct ceph_connection *con,
 966                         int *to, int size, void *object)
 967 {
 968         *to += size;
 969         while (con->in_base_pos < *to) {
 970                 int left = *to - con->in_base_pos;
 971                 int have = size - left;
 972                 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
 973                 if (ret <= 0)
 974                         return ret;
 975                 con->in_base_pos += ret;
 976         }
 977         return 1;
 978 }
 979
 980
 981 /*
 982  * Read all or part of the connect-side handshake on a new connection
 983  */
 984 static int read_partial_banner(struct ceph_connection *con)
 985 {
 986         int ret, to = 0;
 987
 988         dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
 989
 990         /* peer's banner */
 991         ret = read_partial(con, &to, strlen(CEPH_BANNER), con->in_banner);
 992         if (ret <= 0)
 993                 goto out;
 994         ret = read_partial(con, &to, sizeof(con->actual_peer_addr),
 995                            &con->actual_peer_addr);
 996         if (ret <= 0)
 997                 goto out;
 998         ret = read_partial(con, &to, sizeof(con->peer_addr_for_me),
 999                            &con->peer_addr_for_me);
1000         if (ret <= 0)
1001                 goto out;
1002 out:
1003         return ret;
1004 }
1005
1006 static int read_partial_connect(struct ceph_connection *con)
1007 {
1008         int ret, to = 0;
1009
1010         dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
1011
1012         ret = read_partial(con, &to, sizeof(con->in_reply), &con->in_reply);
1013         if (ret <= 0)
1014                 goto out;
1015         ret = read_partial(con, &to, le32_to_cpu(con->in_reply.authorizer_len),
1016                            con->auth_reply_buf);
1017         if (ret <= 0)
1018                 goto out;
1019
1020         dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
1021              con, (int)con->in_reply.tag,
1022              le32_to_cpu(con->in_reply.connect_seq),
1023              le32_to_cpu(con->in_reply.global_seq));
1024 out:
1025         return ret;
1026
1027 }
1028
1029 /*
1030  * Verify the hello banner looks okay.
1031  */
1032 static int verify_hello(struct ceph_connection *con)
1033 {
1034         if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
1035                 pr_err("connect to %s got bad banner\n",
1036                        ceph_pr_addr(&con->peer_addr.in_addr));
1037                 con->error_msg = "protocol error, bad banner";
1038                 return -1;
1039         }
1040         return 0;
1041 }
1042
1043 static bool addr_is_blank(struct sockaddr_storage *ss)
1044 {
1045         switch (ss->ss_family) {
1046         case AF_INET:
1047                 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
1048         case AF_INET6:
1049                 return
1050                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
1051                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
1052                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
1053                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
1054         }
1055         return false;
1056 }
1057
1058 static int addr_port(struct sockaddr_storage *ss)
1059 {
1060         switch (ss->ss_family) {
1061         case AF_INET:
1062                 return ntohs(((struct sockaddr_in *)ss)->sin_port);
1063         case AF_INET6:
1064                 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
1065         }
1066         return 0;
1067 }
1068
1069 static void addr_set_port(struct sockaddr_storage *ss, int p)
1070 {
1071         switch (ss->ss_family) {
1072         case AF_INET:
1073                 ((struct sockaddr_in *)ss)->sin_port = htons(p);
1074                 break;
1075         case AF_INET6:
1076                 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
1077                 break;
1078         }
1079 }
1080
1081 /*
1082  * Unlike other *_pton function semantics, zero indicates success.
1083  */
1084 static int ceph_pton(const char *str, size_t len, struct sockaddr_storage *ss,
1085                 char delim, const char **ipend)
1086 {
1087         struct sockaddr_in *in4 = (void *)ss;
1088         struct sockaddr_in6 *in6 = (void *)ss;
1089
1090         memset(ss, 0, sizeof(*ss));
1091
1092         if (in4_pton(str, len, (u8 *)&in4->sin_addr.s_addr, delim, ipend)) {
1093                 ss->ss_family = AF_INET;
1094                 return 0;
1095         }
1096
1097         if (in6_pton(str, len, (u8 *)&in6->sin6_addr.s6_addr, delim, ipend)) {
1098                 ss->ss_family = AF_INET6;
1099                 return 0;
1100         }
1101
1102         return -EINVAL;
1103 }
1104
1105 /*
1106  * Extract hostname string and resolve using kernel DNS facility.
1107  */
1108 #ifdef CONFIG_CEPH_LIB_USE_DNS_RESOLVER
1109 static int ceph_dns_resolve_name(const char *name, size_t namelen,
1110                 struct sockaddr_storage *ss, char delim, const char **ipend)
1111 {
1112         const char *end, *delim_p;
1113         char *colon_p, *ip_addr = NULL;
1114         int ip_len, ret;
1115
1116         /*
1117          * The end of the hostname occurs immediately preceding the delimiter or
1118          * the port marker (':') where the delimiter takes precedence.
1119          */
1120         delim_p = memchr(name, delim, namelen);
1121         colon_p = memchr(name, ':', namelen);
1122
1123         if (delim_p && colon_p)
1124                 end = delim_p < colon_p ? delim_p : colon_p;
1125         else if (!delim_p && colon_p)
1126                 end = colon_p;
1127         else {
1128                 end = delim_p;
1129                 if (!end) /* case: hostname:/ */
1130                         end = name + namelen;
1131         }
1132
1133         if (end <= name)
1134                 return -EINVAL;
1135
1136         /* do dns_resolve upcall */
1137         ip_len = dns_query(NULL, name, end - name, NULL, &ip_addr, NULL);
1138         if (ip_len > 0)
1139                 ret = ceph_pton(ip_addr, ip_len, ss, -1, NULL);
1140         else
1141                 ret = -ESRCH;
1142
1143         kfree(ip_addr);
1144
1145         *ipend = end;
1146
1147         pr_info("resolve '%.*s' (ret=%d): %s\n", (int)(end - name), name,
1148                         ret, ret ? "failed" : ceph_pr_addr(ss));
1149
1150         return ret;
1151 }
1152 #else
1153 static inline int ceph_dns_resolve_name(const char *name, size_t namelen,
1154                 struct sockaddr_storage *ss, char delim, const char **ipend)
1155 {
1156         return -EINVAL;
1157 }
1158 #endif
1159
1160 /*
1161  * Parse a server name (IP or hostname). If a valid IP address is not found
1162  * then try to extract a hostname to resolve using userspace DNS upcall.
1163  */
1164 static int ceph_parse_server_name(const char *name, size_t namelen,
1165                         struct sockaddr_storage *ss, char delim, const char **ipend)
1166 {
1167         int ret;
1168
1169         ret = ceph_pton(name, namelen, ss, delim, ipend);
1170         if (ret)
1171                 ret = ceph_dns_resolve_name(name, namelen, ss, delim, ipend);
1172
1173         return ret;
1174 }
1175
1176 /*
1177  * Parse an ip[:port] list into an addr array.  Use the default
1178  * monitor port if a port isn't specified.
1179  */
1180 int ceph_parse_ips(const char *c, const char *end,
1181                    struct ceph_entity_addr *addr,
1182                    int max_count, int *count)
1183 {
1184         int i, ret = -EINVAL;
1185         const char *p = c;
1186
1187         dout("parse_ips on '%.*s'\n", (int)(end-c), c);
1188         for (i = 0; i < max_count; i++) {
1189                 const char *ipend;
1190                 struct sockaddr_storage *ss = &addr[i].in_addr;
1191                 int port;
1192                 char delim = ',';
1193
1194                 if (*p == '[') {
1195                         delim = ']';
1196                         p++;
1197                 }
1198
1199                 ret = ceph_parse_server_name(p, end - p, ss, delim, &ipend);
1200                 if (ret)
1201                         goto bad;
1202                 ret = -EINVAL;
1203
1204                 p = ipend;
1205
1206                 if (delim == ']') {
1207                         if (*p != ']') {
1208                                 dout("missing matching ']'\n");
1209                                 goto bad;
1210                         }
1211                         p++;
1212                 }
1213
1214                 /* port? */
1215                 if (p < end && *p == ':') {
1216                         port = 0;
1217                         p++;
1218                         while (p < end && *p >= '0' && *p <= '9') {
1219                                 port = (port * 10) + (*p - '0');
1220                                 p++;
1221                         }
1222                         if (port > 65535 || port == 0)
1223                                 goto bad;
1224                 } else {
1225                         port = CEPH_MON_PORT;
1226                 }
1227
1228                 addr_set_port(ss, port);
1229
1230                 dout("parse_ips got %s\n", ceph_pr_addr(ss));
1231
1232                 if (p == end)
1233                         break;
1234                 if (*p != ',')
1235                         goto bad;
1236                 p++;
1237         }
1238
1239         if (p != end)
1240                 goto bad;
1241
1242         if (count)
1243                 *count = i + 1;
1244         return 0;
1245
1246 bad:
1247         pr_err("parse_ips bad ip '%.*s'\n", (int)(end - c), c);
1248         return ret;
1249 }
1250 EXPORT_SYMBOL(ceph_parse_ips);
1251
1252 static int process_banner(struct ceph_connection *con)
1253 {
1254         dout("process_banner on %p\n", con);
1255
1256         if (verify_hello(con) < 0)
1257                 return -1;
1258
1259         ceph_decode_addr(&con->actual_peer_addr);
1260         ceph_decode_addr(&con->peer_addr_for_me);
1261
1262         /*
1263          * Make sure the other end is who we wanted.  note that the other
1264          * end may not yet know their ip address, so if it's 0.0.0.0, give
1265          * them the benefit of the doubt.
1266          */
1267         if (memcmp(&con->peer_addr, &con->actual_peer_addr,
1268                    sizeof(con->peer_addr)) != 0 &&
1269             !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1270               con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1271                 pr_warning("wrong peer, want %s/%d, got %s/%d\n",
1272                            ceph_pr_addr(&con->peer_addr.in_addr),
1273                            (int)le32_to_cpu(con->peer_addr.nonce),
1274                            ceph_pr_addr(&con->actual_peer_addr.in_addr),
1275                            (int)le32_to_cpu(con->actual_peer_addr.nonce));
1276                 con->error_msg = "wrong peer at address";
1277                 return -1;
1278         }
1279
1280         /*
1281          * did we learn our address?
1282          */
1283         if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1284                 int port = addr_port(&con->msgr->inst.addr.in_addr);
1285
1286                 memcpy(&con->msgr->inst.addr.in_addr,
1287                        &con->peer_addr_for_me.in_addr,
1288                        sizeof(con->peer_addr_for_me.in_addr));
1289                 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1290                 encode_my_addr(con->msgr);
1291                 dout("process_banner learned my addr is %s\n",
1292                      ceph_pr_addr(&con->msgr->inst.addr.in_addr));
1293         }
1294
1295         set_bit(NEGOTIATING, &con->state);
1296         prepare_read_connect(con);
1297         return 0;
1298 }
1299
1300 static void fail_protocol(struct ceph_connection *con)
1301 {
1302         reset_connection(con);
1303         set_bit(CLOSED, &con->state);  /* in case there's queued work */
1304
1305         mutex_unlock(&con->mutex);
1306         if (con->ops->bad_proto)
1307                 con->ops->bad_proto(con);
1308         mutex_lock(&con->mutex);
1309 }
1310
1311 static int process_connect(struct ceph_connection *con)
1312 {
1313         u64 sup_feat = con->msgr->supported_features;
1314         u64 req_feat = con->msgr->required_features;
1315         u64 server_feat = le64_to_cpu(con->in_reply.features);
1316         int ret;
1317
1318         dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1319
1320         switch (con->in_reply.tag) {
1321         case CEPH_MSGR_TAG_FEATURES:
1322                 pr_err("%s%lld %s feature set mismatch,"
1323                        " my %llx < server's %llx, missing %llx\n",
1324                        ENTITY_NAME(con->peer_name),
1325                        ceph_pr_addr(&con->peer_addr.in_addr),
1326                        sup_feat, server_feat, server_feat & ~sup_feat);
1327                 con->error_msg = "missing required protocol features";
1328                 fail_protocol(con);
1329                 return -1;
1330
1331         case CEPH_MSGR_TAG_BADPROTOVER:
1332                 pr_err("%s%lld %s protocol version mismatch,"
1333                        " my %d != server's %d\n",
1334                        ENTITY_NAME(con->peer_name),
1335                        ceph_pr_addr(&con->peer_addr.in_addr),
1336                        le32_to_cpu(con->out_connect.protocol_version),
1337                        le32_to_cpu(con->in_reply.protocol_version));
1338                 con->error_msg = "protocol version mismatch";
1339                 fail_protocol(con);
1340                 return -1;
1341
1342         case CEPH_MSGR_TAG_BADAUTHORIZER:
1343                 con->auth_retry++;
1344                 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1345                      con->auth_retry);
1346                 if (con->auth_retry == 2) {
1347                         con->error_msg = "connect authorization failure";
1348                         return -1;
1349                 }
1350                 con->auth_retry = 1;
1351                 ret = prepare_write_connect(con->msgr, con, 0);
1352                 if (ret < 0)
1353                         return ret;
1354                 prepare_read_connect(con);
1355                 break;
1356
1357         case CEPH_MSGR_TAG_RESETSESSION:
1358                 /*
1359                  * If we connected with a large connect_seq but the peer
1360                  * has no record of a session with us (no connection, or
1361                  * connect_seq == 0), they will send RESETSESION to indicate
1362                  * that they must have reset their session, and may have
1363                  * dropped messages.
1364                  */
1365                 dout("process_connect got RESET peer seq %u\n",
1366                      le32_to_cpu(con->in_connect.connect_seq));
1367                 pr_err("%s%lld %s connection reset\n",
1368                        ENTITY_NAME(con->peer_name),
1369                        ceph_pr_addr(&con->peer_addr.in_addr));
1370                 reset_connection(con);
1371                 prepare_write_connect(con->msgr, con, 0);
1372                 prepare_read_connect(con);
1373
1374                 /* Tell ceph about it. */
1375                 mutex_unlock(&con->mutex);
1376                 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1377                 if (con->ops->peer_reset)
1378                         con->ops->peer_reset(con);
1379                 mutex_lock(&con->mutex);
1380                 if (test_bit(CLOSED, &con->state) ||
1381                     test_bit(OPENING, &con->state))
1382                         return -EAGAIN;
1383                 break;
1384
1385         case CEPH_MSGR_TAG_RETRY_SESSION:
1386                 /*
1387                  * If we sent a smaller connect_seq than the peer has, try
1388                  * again with a larger value.
1389                  */
1390                 dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
1391                      le32_to_cpu(con->out_connect.connect_seq),
1392                      le32_to_cpu(con->in_connect.connect_seq));
1393                 con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
1394                 prepare_write_connect(con->msgr, con, 0);
1395                 prepare_read_connect(con);
1396                 break;
1397
1398         case CEPH_MSGR_TAG_RETRY_GLOBAL:
1399                 /*
1400                  * If we sent a smaller global_seq than the peer has, try
1401                  * again with a larger value.
1402                  */
1403                 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1404                      con->peer_global_seq,
1405                      le32_to_cpu(con->in_connect.global_seq));
1406                 get_global_seq(con->msgr,
1407                                le32_to_cpu(con->in_connect.global_seq));
1408                 prepare_write_connect(con->msgr, con, 0);
1409                 prepare_read_connect(con);
1410                 break;
1411
1412         case CEPH_MSGR_TAG_READY:
1413                 if (req_feat & ~server_feat) {
1414                         pr_err("%s%lld %s protocol feature mismatch,"
1415                                " my required %llx > server's %llx, need %llx\n",
1416                                ENTITY_NAME(con->peer_name),
1417                                ceph_pr_addr(&con->peer_addr.in_addr),
1418                                req_feat, server_feat, req_feat & ~server_feat);
1419                         con->error_msg = "missing required protocol features";
1420                         fail_protocol(con);
1421                         return -1;
1422                 }
1423                 clear_bit(CONNECTING, &con->state);
1424                 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1425                 con->connect_seq++;
1426                 con->peer_features = server_feat;
1427                 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1428                      con->peer_global_seq,
1429                      le32_to_cpu(con->in_reply.connect_seq),
1430                      con->connect_seq);
1431                 WARN_ON(con->connect_seq !=
1432                         le32_to_cpu(con->in_reply.connect_seq));
1433
1434                 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1435                         set_bit(LOSSYTX, &con->state);
1436
1437                 prepare_read_tag(con);
1438                 break;
1439
1440         case CEPH_MSGR_TAG_WAIT:
1441                 /*
1442                  * If there is a connection race (we are opening
1443                  * connections to each other), one of us may just have
1444                  * to WAIT.  This shouldn't happen if we are the
1445                  * client.
1446                  */
1447                 pr_err("process_connect got WAIT as client\n");
1448                 con->error_msg = "protocol error, got WAIT as client";
1449                 return -1;
1450
1451         default:
1452                 pr_err("connect protocol error, will retry\n");
1453                 con->error_msg = "protocol error, garbage tag during connect";
1454                 return -1;
1455         }
1456         return 0;
1457 }
1458
1459
1460 /*
1461  * read (part of) an ack
1462  */
1463 static int read_partial_ack(struct ceph_connection *con)
1464 {
1465         int to = 0;
1466
1467         return read_partial(con, &to, sizeof(con->in_temp_ack),
1468                             &con->in_temp_ack);
1469 }
1470
1471
1472 /*
1473  * We can finally discard anything that's been acked.
1474  */
1475 static void process_ack(struct ceph_connection *con)
1476 {
1477         struct ceph_msg *m;
1478         u64 ack = le64_to_cpu(con->in_temp_ack);
1479         u64 seq;
1480
1481         while (!list_empty(&con->out_sent)) {
1482                 m = list_first_entry(&con->out_sent, struct ceph_msg,
1483                                      list_head);
1484                 seq = le64_to_cpu(m->hdr.seq);
1485                 if (seq > ack)
1486                         break;
1487                 dout("got ack for seq %llu type %d at %p\n", seq,
1488                      le16_to_cpu(m->hdr.type), m);
1489                 m->ack_stamp = jiffies;
1490                 ceph_msg_remove(m);
1491         }
1492         prepare_read_tag(con);
1493 }
1494
1495
1496
1497
1498 static int read_partial_message_section(struct ceph_connection *con,
1499                                         struct kvec *section,
1500                                         unsigned int sec_len, u32 *crc)
1501 {
1502         int ret, left;
1503
1504         BUG_ON(!section);
1505
1506         while (section->iov_len < sec_len) {
1507                 BUG_ON(section->iov_base == NULL);
1508                 left = sec_len - section->iov_len;
1509                 ret = ceph_tcp_recvmsg(con->sock, (char *)section->iov_base +
1510                                        section->iov_len, left);
1511                 if (ret <= 0)
1512                         return ret;
1513                 section->iov_len += ret;
1514                 if (section->iov_len == sec_len)
1515                         *crc = crc32c(0, section->iov_base,
1516                                       section->iov_len);
1517         }
1518
1519         return 1;
1520 }
1521
1522 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
1523                                 struct ceph_msg_header *hdr,
1524                                 int *skip);
1525
1526
1527 static int read_partial_message_pages(struct ceph_connection *con,
1528                                       struct page **pages,
1529                                       unsigned data_len, int datacrc)
1530 {
1531         void *p;
1532         int ret;
1533         int left;
1534
1535         left = min((int)(data_len - con->in_msg_pos.data_pos),
1536                    (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1537         /* (page) data */
1538         BUG_ON(pages == NULL);
1539         p = kmap(pages[con->in_msg_pos.page]);
1540         ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1541                                left);
1542         if (ret > 0 && datacrc)
1543                 con->in_data_crc =
1544                         crc32c(con->in_data_crc,
1545                                   p + con->in_msg_pos.page_pos, ret);
1546         kunmap(pages[con->in_msg_pos.page]);
1547         if (ret <= 0)
1548                 return ret;
1549         con->in_msg_pos.data_pos += ret;
1550         con->in_msg_pos.page_pos += ret;
1551         if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1552                 con->in_msg_pos.page_pos = 0;
1553                 con->in_msg_pos.page++;
1554         }
1555
1556         return ret;
1557 }
1558
1559 #ifdef CONFIG_BLOCK
1560 static int read_partial_message_bio(struct ceph_connection *con,
1561                                     struct bio **bio_iter, int *bio_seg,
1562                                     unsigned data_len, int datacrc)
1563 {
1564         struct bio_vec *bv = bio_iovec_idx(*bio_iter, *bio_seg);
1565         void *p;
1566         int ret, left;
1567
1568         if (IS_ERR(bv))
1569                 return PTR_ERR(bv);
1570
1571         left = min((int)(data_len - con->in_msg_pos.data_pos),
1572                    (int)(bv->bv_len - con->in_msg_pos.page_pos));
1573
1574         p = kmap(bv->bv_page) + bv->bv_offset;
1575
1576         ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1577                                left);
1578         if (ret > 0 && datacrc)
1579                 con->in_data_crc =
1580                         crc32c(con->in_data_crc,
1581                                   p + con->in_msg_pos.page_pos, ret);
1582         kunmap(bv->bv_page);
1583         if (ret <= 0)
1584                 return ret;
1585         con->in_msg_pos.data_pos += ret;
1586         con->in_msg_pos.page_pos += ret;
1587         if (con->in_msg_pos.page_pos == bv->bv_len) {
1588                 con->in_msg_pos.page_pos = 0;
1589                 iter_bio_next(bio_iter, bio_seg);
1590         }
1591
1592         return ret;
1593 }
1594 #endif
1595
1596 /*
1597  * read (part of) a message.
1598  */
1599 static int read_partial_message(struct ceph_connection *con)
1600 {
1601         struct ceph_msg *m = con->in_msg;
1602         int ret;
1603         int to, left;
1604         unsigned front_len, middle_len, data_len;
1605         int datacrc = con->msgr->nocrc;
1606         int skip;
1607         u64 seq;
1608
1609         dout("read_partial_message con %p msg %p\n", con, m);
1610
1611         /* header */
1612         while (con->in_base_pos < sizeof(con->in_hdr)) {
1613                 left = sizeof(con->in_hdr) - con->in_base_pos;
1614                 ret = ceph_tcp_recvmsg(con->sock,
1615                                        (char *)&con->in_hdr + con->in_base_pos,
1616                                        left);
1617                 if (ret <= 0)
1618                         return ret;
1619                 con->in_base_pos += ret;
1620                 if (con->in_base_pos == sizeof(con->in_hdr)) {
1621                         u32 crc = crc32c(0, (void *)&con->in_hdr,
1622                                  sizeof(con->in_hdr) - sizeof(con->in_hdr.crc));
1623                         if (crc != le32_to_cpu(con->in_hdr.crc)) {
1624                                 pr_err("read_partial_message bad hdr "
1625                                        " crc %u != expected %u\n",
1626                                        crc, con->in_hdr.crc);
1627                                 return -EBADMSG;
1628                         }
1629                 }
1630         }
1631         front_len = le32_to_cpu(con->in_hdr.front_len);
1632         if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1633                 return -EIO;
1634         middle_len = le32_to_cpu(con->in_hdr.middle_len);
1635         if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1636                 return -EIO;
1637         data_len = le32_to_cpu(con->in_hdr.data_len);
1638         if (data_len > CEPH_MSG_MAX_DATA_LEN)
1639                 return -EIO;
1640
1641         /* verify seq# */
1642         seq = le64_to_cpu(con->in_hdr.seq);
1643         if ((s64)seq - (s64)con->in_seq < 1) {
1644                 pr_info("skipping %s%lld %s seq %lld expected %lld\n",
1645                         ENTITY_NAME(con->peer_name),
1646                         ceph_pr_addr(&con->peer_addr.in_addr),
1647                         seq, con->in_seq + 1);
1648                 con->in_base_pos = -front_len - middle_len - data_len -
1649                         sizeof(m->footer);
1650                 con->in_tag = CEPH_MSGR_TAG_READY;
1651                 return 0;
1652         } else if ((s64)seq - (s64)con->in_seq > 1) {
1653                 pr_err("read_partial_message bad seq %lld expected %lld\n",
1654                        seq, con->in_seq + 1);
1655                 con->error_msg = "bad message sequence # for incoming message";
1656                 return -EBADMSG;
1657         }
1658
1659         /* allocate message? */
1660         if (!con->in_msg) {
1661                 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1662                      con->in_hdr.front_len, con->in_hdr.data_len);
1663                 skip = 0;
1664                 con->in_msg = ceph_alloc_msg(con, &con->in_hdr, &skip);
1665                 if (skip) {
1666                         /* skip this message */
1667                         dout("alloc_msg said skip message\n");
1668                         BUG_ON(con->in_msg);
1669                         con->in_base_pos = -front_len - middle_len - data_len -
1670                                 sizeof(m->footer);
1671                         con->in_tag = CEPH_MSGR_TAG_READY;
1672                         con->in_seq++;
1673                         return 0;
1674                 }
1675                 if (!con->in_msg) {
1676                         con->error_msg =
1677                                 "error allocating memory for incoming message";
1678                         return -ENOMEM;
1679                 }
1680                 m = con->in_msg;
1681                 m->front.iov_len = 0;    /* haven't read it yet */
1682                 if (m->middle)
1683                         m->middle->vec.iov_len = 0;
1684
1685                 con->in_msg_pos.page = 0;
1686                 if (m->pages)
1687                         con->in_msg_pos.page_pos = m->page_alignment;
1688                 else
1689                         con->in_msg_pos.page_pos = 0;
1690                 con->in_msg_pos.data_pos = 0;
1691         }
1692
1693         /* front */
1694         ret = read_partial_message_section(con, &m->front, front_len,
1695                                            &con->in_front_crc);
1696         if (ret <= 0)
1697                 return ret;
1698
1699         /* middle */
1700         if (m->middle) {
1701                 ret = read_partial_message_section(con, &m->middle->vec,
1702                                                    middle_len,
1703                                                    &con->in_middle_crc);
1704                 if (ret <= 0)
1705                         return ret;
1706         }
1707 #ifdef CONFIG_BLOCK
1708         if (m->bio && !m->bio_iter)
1709                 init_bio_iter(m->bio, &m->bio_iter, &m->bio_seg);
1710 #endif
1711
1712         /* (page) data */
1713         while (con->in_msg_pos.data_pos < data_len) {
1714                 if (m->pages) {
1715                         ret = read_partial_message_pages(con, m->pages,
1716                                                  data_len, datacrc);
1717                         if (ret <= 0)
1718                                 return ret;
1719 #ifdef CONFIG_BLOCK
1720                 } else if (m->bio) {
1721
1722                         ret = read_partial_message_bio(con,
1723                                                  &m->bio_iter, &m->bio_seg,
1724                                                  data_len, datacrc);
1725                         if (ret <= 0)
1726                                 return ret;
1727 #endif
1728                 } else {
1729                         BUG_ON(1);
1730                 }
1731         }
1732
1733         /* footer */
1734         to = sizeof(m->hdr) + sizeof(m->footer);
1735         while (con->in_base_pos < to) {
1736                 left = to - con->in_base_pos;
1737                 ret = ceph_tcp_recvmsg(con->sock, (char *)&m->footer +
1738                                        (con->in_base_pos - sizeof(m->hdr)),
1739                                        left);
1740                 if (ret <= 0)
1741                         return ret;
1742                 con->in_base_pos += ret;
1743         }
1744         dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1745              m, front_len, m->footer.front_crc, middle_len,
1746              m->footer.middle_crc, data_len, m->footer.data_crc);
1747
1748         /* crc ok? */
1749         if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1750                 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1751                        m, con->in_front_crc, m->footer.front_crc);
1752                 return -EBADMSG;
1753         }
1754         if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1755                 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1756                        m, con->in_middle_crc, m->footer.middle_crc);
1757                 return -EBADMSG;
1758         }
1759         if (datacrc &&
1760             (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1761             con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1762                 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1763                        con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1764                 return -EBADMSG;
1765         }
1766
1767         return 1; /* done! */
1768 }
1769
1770 /*
1771  * Process message.  This happens in the worker thread.  The callback should
1772  * be careful not to do anything that waits on other incoming messages or it
1773  * may deadlock.
1774  */
1775 static void process_message(struct ceph_connection *con)
1776 {
1777         struct ceph_msg *msg;
1778
1779         msg = con->in_msg;
1780         con->in_msg = NULL;
1781
1782         /* if first message, set peer_name */
1783         if (con->peer_name.type == 0)
1784                 con->peer_name = msg->hdr.src;
1785
1786         con->in_seq++;
1787         mutex_unlock(&con->mutex);
1788
1789         dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1790              msg, le64_to_cpu(msg->hdr.seq),
1791              ENTITY_NAME(msg->hdr.src),
1792              le16_to_cpu(msg->hdr.type),
1793              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1794              le32_to_cpu(msg->hdr.front_len),
1795              le32_to_cpu(msg->hdr.data_len),
1796              con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1797         con->ops->dispatch(con, msg);
1798
1799         mutex_lock(&con->mutex);
1800         prepare_read_tag(con);
1801 }
1802
1803
1804 /*
1805  * Write something to the socket.  Called in a worker thread when the
1806  * socket appears to be writeable and we have something ready to send.
1807  */
1808 static int try_write(struct ceph_connection *con)
1809 {
1810         struct ceph_messenger *msgr = con->msgr;
1811         int ret = 1;
1812
1813         dout("try_write start %p state %lu nref %d\n", con, con->state,
1814              atomic_read(&con->nref));
1815
1816 more:
1817         dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1818
1819         /* open the socket first? */
1820         if (con->sock == NULL) {
1821                 prepare_write_banner(msgr, con);
1822                 prepare_write_connect(msgr, con, 1);
1823                 prepare_read_banner(con);
1824                 set_bit(CONNECTING, &con->state);
1825                 clear_bit(NEGOTIATING, &con->state);
1826
1827                 BUG_ON(con->in_msg);
1828                 con->in_tag = CEPH_MSGR_TAG_READY;
1829                 dout("try_write initiating connect on %p new state %lu\n",
1830                      con, con->state);
1831                 con->sock = ceph_tcp_connect(con);
1832                 if (IS_ERR(con->sock)) {
1833                         con->sock = NULL;
1834                         con->error_msg = "connect error";
1835                         ret = -1;
1836                         goto out;
1837                 }
1838         }
1839
1840 more_kvec:
1841         /* kvec data queued? */
1842         if (con->out_skip) {
1843                 ret = write_partial_skip(con);
1844                 if (ret <= 0)
1845                         goto out;
1846         }
1847         if (con->out_kvec_left) {
1848                 ret = write_partial_kvec(con);
1849                 if (ret <= 0)
1850                         goto out;
1851         }
1852
1853         /* msg pages? */
1854         if (con->out_msg) {
1855                 if (con->out_msg_done) {
1856                         ceph_msg_put(con->out_msg);
1857                         con->out_msg = NULL;   /* we're done with this one */
1858                         goto do_next;
1859                 }
1860
1861                 ret = write_partial_msg_pages(con);
1862                 if (ret == 1)
1863                         goto more_kvec;  /* we need to send the footer, too! */
1864                 if (ret == 0)
1865                         goto out;
1866                 if (ret < 0) {
1867                         dout("try_write write_partial_msg_pages err %d\n",
1868                              ret);
1869                         goto out;
1870                 }
1871         }
1872
1873 do_next:
1874         if (!test_bit(CONNECTING, &con->state)) {
1875                 /* is anything else pending? */
1876                 if (!list_empty(&con->out_queue)) {
1877                         prepare_write_message(con);
1878                         goto more;
1879                 }
1880                 if (con->in_seq > con->in_seq_acked) {
1881                         prepare_write_ack(con);
1882                         goto more;
1883                 }
1884                 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1885                         prepare_write_keepalive(con);
1886                         goto more;
1887                 }
1888         }
1889
1890         /* Nothing to do! */
1891         clear_bit(WRITE_PENDING, &con->state);
1892         dout("try_write nothing else to write.\n");
1893         ret = 0;
1894 out:
1895         dout("try_write done on %p ret %d\n", con, ret);
1896         return ret;
1897 }
1898
1899
1900
1901 /*
1902  * Read what we can from the socket.
1903  */
1904 static int try_read(struct ceph_connection *con)
1905 {
1906         int ret = -1;
1907
1908         if (!con->sock)
1909                 return 0;
1910
1911         if (test_bit(STANDBY, &con->state))
1912                 return 0;
1913
1914         dout("try_read start on %p\n", con);
1915
1916 more:
1917         dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1918              con->in_base_pos);
1919
1920         /*
1921          * process_connect and process_message drop and re-take
1922          * con->mutex.  make sure we handle a racing close or reopen.
1923          */
1924         if (test_bit(CLOSED, &con->state) ||
1925             test_bit(OPENING, &con->state)) {
1926                 ret = -EAGAIN;
1927                 goto out;
1928         }
1929
1930         if (test_bit(CONNECTING, &con->state)) {
1931                 if (!test_bit(NEGOTIATING, &con->state)) {
1932                         dout("try_read connecting\n");
1933                         ret = read_partial_banner(con);
1934                         if (ret <= 0)
1935                                 goto out;
1936                         ret = process_banner(con);
1937                         if (ret < 0)
1938                                 goto out;
1939                 }
1940                 ret = read_partial_connect(con);
1941                 if (ret <= 0)
1942                         goto out;
1943                 ret = process_connect(con);
1944                 if (ret < 0)
1945                         goto out;
1946                 goto more;
1947         }
1948
1949         if (con->in_base_pos < 0) {
1950                 /*
1951                  * skipping + discarding content.
1952                  *
1953                  * FIXME: there must be a better way to do this!
1954                  */
1955                 static char buf[1024];
1956                 int skip = min(1024, -con->in_base_pos);
1957                 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
1958                 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
1959                 if (ret <= 0)
1960                         goto out;
1961                 con->in_base_pos += ret;
1962                 if (con->in_base_pos)
1963                         goto more;
1964         }
1965         if (con->in_tag == CEPH_MSGR_TAG_READY) {
1966                 /*
1967                  * what's next?
1968                  */
1969                 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
1970                 if (ret <= 0)
1971                         goto out;
1972                 dout("try_read got tag %d\n", (int)con->in_tag);
1973                 switch (con->in_tag) {
1974                 case CEPH_MSGR_TAG_MSG:
1975                         prepare_read_message(con);
1976                         break;
1977                 case CEPH_MSGR_TAG_ACK:
1978                         prepare_read_ack(con);
1979                         break;
1980                 case CEPH_MSGR_TAG_CLOSE:
1981                         set_bit(CLOSED, &con->state);   /* fixme */
1982                         goto out;
1983                 default:
1984                         goto bad_tag;
1985                 }
1986         }
1987         if (con->in_tag == CEPH_MSGR_TAG_MSG) {
1988                 ret = read_partial_message(con);
1989                 if (ret <= 0) {
1990                         switch (ret) {
1991                         case -EBADMSG:
1992                                 con->error_msg = "bad crc";
1993                                 ret = -EIO;
1994                                 break;
1995                         case -EIO:
1996                                 con->error_msg = "io error";
1997                                 break;
1998                         }
1999                         goto out;
2000                 }
2001                 if (con->in_tag == CEPH_MSGR_TAG_READY)
2002                         goto more;
2003                 process_message(con);
2004                 goto more;
2005         }
2006         if (con->in_tag == CEPH_MSGR_TAG_ACK) {
2007                 ret = read_partial_ack(con);
2008                 if (ret <= 0)
2009                         goto out;
2010                 process_ack(con);
2011                 goto more;
2012         }
2013
2014 out:
2015         dout("try_read done on %p ret %d\n", con, ret);
2016         return ret;
2017
2018 bad_tag:
2019         pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
2020         con->error_msg = "protocol error, garbage tag";
2021         ret = -1;
2022         goto out;
2023 }
2024
2025
2026 /*
2027  * Atomically queue work on a connection.  Bump @con reference to
2028  * avoid races with connection teardown.
2029  */
2030 static void queue_con(struct ceph_connection *con)
2031 {
2032         if (test_bit(DEAD, &con->state)) {
2033                 dout("queue_con %p ignoring: DEAD\n",
2034                      con);
2035                 return;
2036         }
2037
2038         if (!con->ops->get(con)) {
2039                 dout("queue_con %p ref count 0\n", con);
2040                 return;
2041         }
2042
2043         if (!queue_delayed_work(ceph_msgr_wq, &con->work, 0)) {
2044                 dout("queue_con %p - already queued\n", con);
2045                 con->ops->put(con);
2046         } else {
2047                 dout("queue_con %p\n", con);
2048         }
2049 }
2050
2051 /*
2052  * Do some work on a connection.  Drop a connection ref when we're done.
2053  */
2054 static void con_work(struct work_struct *work)
2055 {
2056         struct ceph_connection *con = container_of(work, struct ceph_connection,
2057                                                    work.work);
2058         int ret;
2059
2060         mutex_lock(&con->mutex);
2061 restart:
2062         if (test_and_clear_bit(BACKOFF, &con->state)) {
2063                 dout("con_work %p backing off\n", con);
2064                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
2065                                        round_jiffies_relative(con->delay))) {
2066                         dout("con_work %p backoff %lu\n", con, con->delay);
2067                         mutex_unlock(&con->mutex);
2068                         return;
2069                 } else {
2070                         con->ops->put(con);
2071                         dout("con_work %p FAILED to back off %lu\n", con,
2072                              con->delay);
2073                 }
2074         }
2075
2076         if (test_bit(STANDBY, &con->state)) {
2077                 dout("con_work %p STANDBY\n", con);
2078                 goto done;
2079         }
2080         if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
2081                 dout("con_work CLOSED\n");
2082                 con_close_socket(con);
2083                 goto done;
2084         }
2085         if (test_and_clear_bit(OPENING, &con->state)) {
2086                 /* reopen w/ new peer */
2087                 dout("con_work OPENING\n");
2088                 con_close_socket(con);
2089         }
2090
2091         if (test_and_clear_bit(SOCK_CLOSED, &con->state))
2092                 goto fault;
2093
2094         ret = try_read(con);
2095         if (ret == -EAGAIN)
2096                 goto restart;
2097         if (ret < 0)
2098                 goto fault;
2099
2100         ret = try_write(con);
2101         if (ret == -EAGAIN)
2102                 goto restart;
2103         if (ret < 0)
2104                 goto fault;
2105
2106 done:
2107         mutex_unlock(&con->mutex);
2108 done_unlocked:
2109         con->ops->put(con);
2110         return;
2111
2112 fault:
2113         mutex_unlock(&con->mutex);
2114         ceph_fault(con);     /* error/fault path */
2115         goto done_unlocked;
2116 }
2117
2118
2119 /*
2120  * Generic error/fault handler.  A retry mechanism is used with
2121  * exponential backoff
2122  */
2123 static void ceph_fault(struct ceph_connection *con)
2124 {
2125         pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
2126                ceph_pr_addr(&con->peer_addr.in_addr), con->error_msg);
2127         dout("fault %p state %lu to peer %s\n",
2128              con, con->state, ceph_pr_addr(&con->peer_addr.in_addr));
2129
2130         if (test_bit(LOSSYTX, &con->state)) {
2131                 dout("fault on LOSSYTX channel\n");
2132                 goto out;
2133         }
2134
2135         mutex_lock(&con->mutex);
2136         if (test_bit(CLOSED, &con->state))
2137                 goto out_unlock;
2138
2139         con_close_socket(con);
2140
2141         if (con->in_msg) {
2142                 ceph_msg_put(con->in_msg);
2143                 con->in_msg = NULL;
2144         }
2145
2146         /* Requeue anything that hasn't been acked */
2147         list_splice_init(&con->out_sent, &con->out_queue);
2148
2149         /* If there are no messages queued or keepalive pending, place
2150          * the connection in a STANDBY state */
2151         if (list_empty(&con->out_queue) &&
2152             !test_bit(KEEPALIVE_PENDING, &con->state)) {
2153                 dout("fault %p setting STANDBY clearing WRITE_PENDING\n", con);
2154                 clear_bit(WRITE_PENDING, &con->state);
2155                 set_bit(STANDBY, &con->state);
2156         } else {
2157                 /* retry after a delay. */
2158                 if (con->delay == 0)
2159                         con->delay = BASE_DELAY_INTERVAL;
2160                 else if (con->delay < MAX_DELAY_INTERVAL)
2161                         con->delay *= 2;
2162                 con->ops->get(con);
2163                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
2164                                        round_jiffies_relative(con->delay))) {
2165                         dout("fault queued %p delay %lu\n", con, con->delay);
2166                 } else {
2167                         con->ops->put(con);
2168                         dout("fault failed to queue %p delay %lu, backoff\n",
2169                              con, con->delay);
2170                         /*
2171                          * In many cases we see a socket state change
2172                          * while con_work is running and end up
2173                          * queuing (non-delayed) work, such that we
2174                          * can't backoff with a delay.  Set a flag so
2175                          * that when con_work restarts we schedule the
2176                          * delay then.
2177                          */
2178                         set_bit(BACKOFF, &con->state);
2179                 }
2180         }
2181
2182 out_unlock:
2183         mutex_unlock(&con->mutex);
2184 out:
2185         /*
2186          * in case we faulted due to authentication, invalidate our
2187          * current tickets so that we can get new ones.
2188          */
2189         if (con->auth_retry && con->ops->invalidate_authorizer) {
2190                 dout("calling invalidate_authorizer()\n");
2191                 con->ops->invalidate_authorizer(con);
2192         }
2193
2194         if (con->ops->fault)
2195                 con->ops->fault(con);
2196 }
2197
2198
2199
2200 /*
2201  * create a new messenger instance
2202  */
2203 struct ceph_messenger *ceph_messenger_create(struct ceph_entity_addr *myaddr,
2204                                              u32 supported_features,
2205                                              u32 required_features)
2206 {
2207         struct ceph_messenger *msgr;
2208
2209         msgr = kzalloc(sizeof(*msgr), GFP_KERNEL);
2210         if (msgr == NULL)
2211                 return ERR_PTR(-ENOMEM);
2212
2213         msgr->supported_features = supported_features;
2214         msgr->required_features = required_features;
2215
2216         spin_lock_init(&msgr->global_seq_lock);
2217
2218         /* the zero page is needed if a request is "canceled" while the message
2219          * is being written over the socket */
2220         msgr->zero_page = __page_cache_alloc(GFP_KERNEL | __GFP_ZERO);
2221         if (!msgr->zero_page) {
2222                 kfree(msgr);
2223                 return ERR_PTR(-ENOMEM);
2224         }
2225         kmap(msgr->zero_page);
2226
2227         if (myaddr)
2228                 msgr->inst.addr = *myaddr;
2229
2230         /* select a random nonce */
2231         msgr->inst.addr.type = 0;
2232         get_random_bytes(&msgr->inst.addr.nonce, sizeof(msgr->inst.addr.nonce));
2233         encode_my_addr(msgr);
2234
2235         dout("messenger_create %p\n", msgr);
2236         return msgr;
2237 }
2238 EXPORT_SYMBOL(ceph_messenger_create);
2239
2240 void ceph_messenger_destroy(struct ceph_messenger *msgr)
2241 {
2242         dout("destroy %p\n", msgr);
2243         kunmap(msgr->zero_page);
2244         __free_page(msgr->zero_page);
2245         kfree(msgr);
2246         dout("destroyed messenger %p\n", msgr);
2247 }
2248 EXPORT_SYMBOL(ceph_messenger_destroy);
2249
2250 static void clear_standby(struct ceph_connection *con)
2251 {
2252         /* come back from STANDBY? */
2253         if (test_and_clear_bit(STANDBY, &con->state)) {
2254                 mutex_lock(&con->mutex);
2255                 dout("clear_standby %p and ++connect_seq\n", con);
2256                 con->connect_seq++;
2257                 WARN_ON(test_bit(WRITE_PENDING, &con->state));
2258                 WARN_ON(test_bit(KEEPALIVE_PENDING, &con->state));
2259                 mutex_unlock(&con->mutex);
2260         }
2261 }
2262
2263 /*
2264  * Queue up an outgoing message on the given connection.
2265  */
2266 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
2267 {
2268         if (test_bit(CLOSED, &con->state)) {
2269                 dout("con_send %p closed, dropping %p\n", con, msg);
2270                 ceph_msg_put(msg);
2271                 return;
2272         }
2273
2274         /* set src+dst */
2275         msg->hdr.src = con->msgr->inst.name;
2276
2277         BUG_ON(msg->front.iov_len != le32_to_cpu(msg->hdr.front_len));
2278
2279         msg->needs_out_seq = true;
2280
2281         /* queue */
2282         mutex_lock(&con->mutex);
2283         BUG_ON(!list_empty(&msg->list_head));
2284         list_add_tail(&msg->list_head, &con->out_queue);
2285         dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
2286              ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
2287              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
2288              le32_to_cpu(msg->hdr.front_len),
2289              le32_to_cpu(msg->hdr.middle_len),
2290              le32_to_cpu(msg->hdr.data_len));
2291         mutex_unlock(&con->mutex);
2292
2293         /* if there wasn't anything waiting to send before, queue
2294          * new work */
2295         clear_standby(con);
2296         if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2297                 queue_con(con);
2298 }
2299 EXPORT_SYMBOL(ceph_con_send);
2300
2301 /*
2302  * Revoke a message that was previously queued for send
2303  */
2304 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
2305 {
2306         mutex_lock(&con->mutex);
2307         if (!list_empty(&msg->list_head)) {
2308                 dout("con_revoke %p msg %p - was on queue\n", con, msg);
2309                 list_del_init(&msg->list_head);
2310                 ceph_msg_put(msg);
2311                 msg->hdr.seq = 0;
2312         }
2313         if (con->out_msg == msg) {
2314                 dout("con_revoke %p msg %p - was sending\n", con, msg);
2315                 con->out_msg = NULL;
2316                 if (con->out_kvec_is_msg) {
2317                         con->out_skip = con->out_kvec_bytes;
2318                         con->out_kvec_is_msg = false;
2319                 }
2320                 ceph_msg_put(msg);
2321                 msg->hdr.seq = 0;
2322         }
2323         mutex_unlock(&con->mutex);
2324 }
2325
2326 /*
2327  * Revoke a message that we may be reading data into
2328  */
2329 void ceph_con_revoke_message(struct ceph_connection *con, struct ceph_msg *msg)
2330 {
2331         mutex_lock(&con->mutex);
2332         if (con->in_msg && con->in_msg == msg) {
2333                 unsigned front_len = le32_to_cpu(con->in_hdr.front_len);
2334                 unsigned middle_len = le32_to_cpu(con->in_hdr.middle_len);
2335                 unsigned data_len = le32_to_cpu(con->in_hdr.data_len);
2336
2337                 /* skip rest of message */
2338                 dout("con_revoke_pages %p msg %p revoked\n", con, msg);
2339                         con->in_base_pos = con->in_base_pos -
2340                                 sizeof(struct ceph_msg_header) -
2341                                 front_len -
2342                                 middle_len -
2343                                 data_len -
2344                                 sizeof(struct ceph_msg_footer);
2345                 ceph_msg_put(con->in_msg);
2346                 con->in_msg = NULL;
2347                 con->in_tag = CEPH_MSGR_TAG_READY;
2348                 con->in_seq++;
2349         } else {
2350                 dout("con_revoke_pages %p msg %p pages %p no-op\n",
2351                      con, con->in_msg, msg);
2352         }
2353         mutex_unlock(&con->mutex);
2354 }
2355
2356 /*
2357  * Queue a keepalive byte to ensure the tcp connection is alive.
2358  */
2359 void ceph_con_keepalive(struct ceph_connection *con)
2360 {
2361         dout("con_keepalive %p\n", con);
2362         clear_standby(con);
2363         if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
2364             test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2365                 queue_con(con);
2366 }
2367 EXPORT_SYMBOL(ceph_con_keepalive);
2368
2369
2370 /*
2371  * construct a new message with given type, size
2372  * the new msg has a ref count of 1.
2373  */
2374 struct ceph_msg *ceph_msg_new(int type, int front_len, gfp_t flags,
2375                               bool can_fail)
2376 {
2377         struct ceph_msg *m;
2378
2379         m = kmalloc(sizeof(*m), flags);
2380         if (m == NULL)
2381                 goto out;
2382         kref_init(&m->kref);
2383         INIT_LIST_HEAD(&m->list_head);
2384
2385         m->hdr.tid = 0;
2386         m->hdr.type = cpu_to_le16(type);
2387         m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
2388         m->hdr.version = 0;
2389         m->hdr.front_len = cpu_to_le32(front_len);
2390         m->hdr.middle_len = 0;
2391         m->hdr.data_len = 0;
2392         m->hdr.data_off = 0;
2393         m->hdr.reserved = 0;
2394         m->footer.front_crc = 0;
2395         m->footer.middle_crc = 0;
2396         m->footer.data_crc = 0;
2397         m->footer.flags = 0;
2398         m->front_max = front_len;
2399         m->front_is_vmalloc = false;
2400         m->more_to_follow = false;
2401         m->ack_stamp = 0;
2402         m->pool = NULL;
2403
2404         /* middle */
2405         m->middle = NULL;
2406
2407         /* data */
2408         m->nr_pages = 0;
2409         m->page_alignment = 0;
2410         m->pages = NULL;
2411         m->pagelist = NULL;
2412         m->bio = NULL;
2413         m->bio_iter = NULL;
2414         m->bio_seg = 0;
2415         m->trail = NULL;
2416
2417         /* front */
2418         if (front_len) {
2419                 if (front_len > PAGE_CACHE_SIZE) {
2420                         m->front.iov_base = __vmalloc(front_len, flags,
2421                                                       PAGE_KERNEL);
2422                         m->front_is_vmalloc = true;
2423                 } else {
2424                         m->front.iov_base = kmalloc(front_len, flags);
2425                 }
2426                 if (m->front.iov_base == NULL) {
2427                         dout("ceph_msg_new can't allocate %d bytes\n",
2428                              front_len);
2429                         goto out2;
2430                 }
2431         } else {
2432                 m->front.iov_base = NULL;
2433         }
2434         m->front.iov_len = front_len;
2435
2436         dout("ceph_msg_new %p front %d\n", m, front_len);
2437         return m;
2438
2439 out2:
2440         ceph_msg_put(m);
2441 out:
2442         if (!can_fail) {
2443                 pr_err("msg_new can't create type %d front %d\n", type,
2444                        front_len);
2445                 WARN_ON(1);
2446         } else {
2447                 dout("msg_new can't create type %d front %d\n", type,
2448                      front_len);
2449         }
2450         return NULL;
2451 }
2452 EXPORT_SYMBOL(ceph_msg_new);
2453
2454 /*
2455  * Allocate "middle" portion of a message, if it is needed and wasn't
2456  * allocated by alloc_msg.  This allows us to read a small fixed-size
2457  * per-type header in the front and then gracefully fail (i.e.,
2458  * propagate the error to the caller based on info in the front) when
2459  * the middle is too large.
2460  */
2461 static int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2462 {
2463         int type = le16_to_cpu(msg->hdr.type);
2464         int middle_len = le32_to_cpu(msg->hdr.middle_len);
2465
2466         dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2467              ceph_msg_type_name(type), middle_len);
2468         BUG_ON(!middle_len);
2469         BUG_ON(msg->middle);
2470
2471         msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2472         if (!msg->middle)
2473                 return -ENOMEM;
2474         return 0;
2475 }
2476
2477 /*
2478  * Generic message allocator, for incoming messages.
2479  */
2480 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2481                                 struct ceph_msg_header *hdr,
2482                                 int *skip)
2483 {
2484         int type = le16_to_cpu(hdr->type);
2485         int front_len = le32_to_cpu(hdr->front_len);
2486         int middle_len = le32_to_cpu(hdr->middle_len);
2487         struct ceph_msg *msg = NULL;
2488         int ret;
2489
2490         if (con->ops->alloc_msg) {
2491                 mutex_unlock(&con->mutex);
2492                 msg = con->ops->alloc_msg(con, hdr, skip);
2493                 mutex_lock(&con->mutex);
2494                 if (!msg || *skip)
2495                         return NULL;
2496         }
2497         if (!msg) {
2498                 *skip = 0;
2499                 msg = ceph_msg_new(type, front_len, GFP_NOFS, false);
2500                 if (!msg) {
2501                         pr_err("unable to allocate msg type %d len %d\n",
2502                                type, front_len);
2503                         return NULL;
2504                 }
2505                 msg->page_alignment = le16_to_cpu(hdr->data_off);
2506         }
2507         memcpy(&msg->hdr, &con->in_hdr, sizeof(con->in_hdr));
2508
2509         if (middle_len && !msg->middle) {
2510                 ret = ceph_alloc_middle(con, msg);
2511                 if (ret < 0) {
2512                         ceph_msg_put(msg);
2513                         return NULL;
2514                 }
2515         }
2516
2517         return msg;
2518 }
2519
2520
2521 /*
2522  * Free a generically kmalloc'd message.
2523  */
2524 void ceph_msg_kfree(struct ceph_msg *m)
2525 {
2526         dout("msg_kfree %p\n", m);
2527         if (m->front_is_vmalloc)
2528                 vfree(m->front.iov_base);
2529         else
2530                 kfree(m->front.iov_base);
2531         kfree(m);
2532 }
2533
2534 /*
2535  * Drop a msg ref.  Destroy as needed.
2536  */
2537 void ceph_msg_last_put(struct kref *kref)
2538 {
2539         struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2540
2541         dout("ceph_msg_put last one on %p\n", m);
2542         WARN_ON(!list_empty(&m->list_head));
2543
2544         /* drop middle, data, if any */
2545         if (m->middle) {
2546                 ceph_buffer_put(m->middle);
2547                 m->middle = NULL;
2548         }
2549         m->nr_pages = 0;
2550         m->pages = NULL;
2551
2552         if (m->pagelist) {
2553                 ceph_pagelist_release(m->pagelist);
2554                 kfree(m->pagelist);
2555                 m->pagelist = NULL;
2556         }
2557
2558         m->trail = NULL;
2559
2560         if (m->pool)
2561                 ceph_msgpool_put(m->pool, m);
2562         else
2563                 ceph_msg_kfree(m);
2564 }
2565 EXPORT_SYMBOL(ceph_msg_last_put);
2566
2567 void ceph_msg_dump(struct ceph_msg *msg)
2568 {
2569         pr_debug("msg_dump %p (front_max %d nr_pages %d)\n", msg,
2570                  msg->front_max, msg->nr_pages);
2571         print_hex_dump(KERN_DEBUG, "header: ",
2572                        DUMP_PREFIX_OFFSET, 16, 1,
2573                        &msg->hdr, sizeof(msg->hdr), true);
2574         print_hex_dump(KERN_DEBUG, " front: ",
2575                        DUMP_PREFIX_OFFSET, 16, 1,
2576                        msg->front.iov_base, msg->front.iov_len, true);
2577         if (msg->middle)
2578                 print_hex_dump(KERN_DEBUG, "middle: ",
2579                                DUMP_PREFIX_OFFSET, 16, 1,
2580                                msg->middle->vec.iov_base,
2581                                msg->middle->vec.iov_len, true);
2582         print_hex_dump(KERN_DEBUG, "footer: ",
2583                        DUMP_PREFIX_OFFSET, 16, 1,
2584                        &msg->footer, sizeof(msg->footer), true);
2585 }
2586 EXPORT_SYMBOL(ceph_msg_dump);