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