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