net/ceph/messenger.c

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