net/ceph/messenger.c

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