fs/ceph/messenger.c

   1 #include "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 <net/tcp.h>
  13
  14 #include "super.h"
  15 #include "messenger.h"
  16 #include "decode.h"
  17 #include "pagelist.h"
  18
  19 /*
  20  * Ceph uses the messenger to exchange ceph_msg messages with other
  21  * hosts in the system.  The messenger provides ordered and reliable
  22  * delivery.  We tolerate TCP disconnects by reconnecting (with
  23  * exponential backoff) in the case of a fault (disconnection, bad
  24  * crc, protocol error).  Acks allow sent messages to be discarded by
  25  * the sender.
  26  */
  27
  28 /* static tag bytes (protocol control messages) */
  29 static char tag_msg = CEPH_MSGR_TAG_MSG;
  30 static char tag_ack = CEPH_MSGR_TAG_ACK;
  31 static char tag_keepalive = CEPH_MSGR_TAG_KEEPALIVE;
  32
  33 #ifdef CONFIG_LOCKDEP
  34 static struct lock_class_key socket_class;
  35 #endif
  36
  37
  38 static void queue_con(struct ceph_connection *con);
  39 static void con_work(struct work_struct *);
  40 static void ceph_fault(struct ceph_connection *con);
  41
  42 /*
  43  * nicely render a sockaddr as a string.
  44  */
  45 #define MAX_ADDR_STR 20
  46 static char addr_str[MAX_ADDR_STR][40];
  47 static DEFINE_SPINLOCK(addr_str_lock);
  48 static int last_addr_str;
  49
  50 const char *pr_addr(const struct sockaddr_storage *ss)
  51 {
  52         int i;
  53         char *s;
  54         struct sockaddr_in *in4 = (void *)ss;
  55         unsigned char *quad = (void *)&in4->sin_addr.s_addr;
  56         struct sockaddr_in6 *in6 = (void *)ss;
  57
  58         spin_lock(&addr_str_lock);
  59         i = last_addr_str++;
  60         if (last_addr_str == MAX_ADDR_STR)
  61                 last_addr_str = 0;
  62         spin_unlock(&addr_str_lock);
  63         s = addr_str[i];
  64
  65         switch (ss->ss_family) {
  66         case AF_INET:
  67                 sprintf(s, "%u.%u.%u.%u:%u",
  68                         (unsigned int)quad[0],
  69                         (unsigned int)quad[1],
  70                         (unsigned int)quad[2],
  71                         (unsigned int)quad[3],
  72                         (unsigned int)ntohs(in4->sin_port));
  73                 break;
  74
  75         case AF_INET6:
  76                 sprintf(s, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x:%u",
  77                         in6->sin6_addr.s6_addr16[0],
  78                         in6->sin6_addr.s6_addr16[1],
  79                         in6->sin6_addr.s6_addr16[2],
  80                         in6->sin6_addr.s6_addr16[3],
  81                         in6->sin6_addr.s6_addr16[4],
  82                         in6->sin6_addr.s6_addr16[5],
  83                         in6->sin6_addr.s6_addr16[6],
  84                         in6->sin6_addr.s6_addr16[7],
  85                         (unsigned int)ntohs(in6->sin6_port));
  86                 break;
  87
  88         default:
  89                 sprintf(s, "(unknown sockaddr family %d)", (int)ss->ss_family);
  90         }
  91
  92         return s;
  93 }
  94
  95 static void encode_my_addr(struct ceph_messenger *msgr)
  96 {
  97         memcpy(&msgr->my_enc_addr, &msgr->inst.addr, sizeof(msgr->my_enc_addr));
  98         ceph_encode_addr(&msgr->my_enc_addr);
  99 }
 100
 101 /*
 102  * work queue for all reading and writing to/from the socket.
 103  */
 104 struct workqueue_struct *ceph_msgr_wq;
 105
 106 int __init ceph_msgr_init(void)
 107 {
 108         ceph_msgr_wq = create_workqueue("ceph-msgr");
 109         if (IS_ERR(ceph_msgr_wq)) {
 110                 int ret = PTR_ERR(ceph_msgr_wq);
 111                 pr_err("msgr_init failed to create workqueue: %d\n", ret);
 112                 ceph_msgr_wq = NULL;
 113                 return ret;
 114         }
 115         return 0;
 116 }
 117
 118 void ceph_msgr_exit(void)
 119 {
 120         destroy_workqueue(ceph_msgr_wq);
 121 }
 122
 123 /*
 124  * socket callback functions
 125  */
 126
 127 /* data available on socket, or listen socket received a connect */
 128 static void ceph_data_ready(struct sock *sk, int count_unused)
 129 {
 130         struct ceph_connection *con =
 131                 (struct ceph_connection *)sk->sk_user_data;
 132         if (sk->sk_state != TCP_CLOSE_WAIT) {
 133                 dout("ceph_data_ready on %p state = %lu, queueing work\n",
 134                      con, con->state);
 135                 queue_con(con);
 136         }
 137 }
 138
 139 /* socket has buffer space for writing */
 140 static void ceph_write_space(struct sock *sk)
 141 {
 142         struct ceph_connection *con =
 143                 (struct ceph_connection *)sk->sk_user_data;
 144
 145         /* only queue to workqueue if there is data we want to write. */
 146         if (test_bit(WRITE_PENDING, &con->state)) {
 147                 dout("ceph_write_space %p queueing write work\n", con);
 148                 queue_con(con);
 149         } else {
 150                 dout("ceph_write_space %p nothing to write\n", con);
 151         }
 152
 153         /* since we have our own write_space, clear the SOCK_NOSPACE flag */
 154         clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 155 }
 156
 157 /* socket's state has changed */
 158 static void ceph_state_change(struct sock *sk)
 159 {
 160         struct ceph_connection *con =
 161                 (struct ceph_connection *)sk->sk_user_data;
 162
 163         dout("ceph_state_change %p state = %lu sk_state = %u\n",
 164              con, con->state, sk->sk_state);
 165
 166         if (test_bit(CLOSED, &con->state))
 167                 return;
 168
 169         switch (sk->sk_state) {
 170         case TCP_CLOSE:
 171                 dout("ceph_state_change TCP_CLOSE\n");
 172         case TCP_CLOSE_WAIT:
 173                 dout("ceph_state_change TCP_CLOSE_WAIT\n");
 174                 if (test_and_set_bit(SOCK_CLOSED, &con->state) == 0) {
 175                         if (test_bit(CONNECTING, &con->state))
 176                                 con->error_msg = "connection failed";
 177                         else
 178                                 con->error_msg = "socket closed";
 179                         queue_con(con);
 180                 }
 181                 break;
 182         case TCP_ESTABLISHED:
 183                 dout("ceph_state_change TCP_ESTABLISHED\n");
 184                 queue_con(con);
 185                 break;
 186         }
 187 }
 188
 189 /*
 190  * set up socket callbacks
 191  */
 192 static void set_sock_callbacks(struct socket *sock,
 193                                struct ceph_connection *con)
 194 {
 195         struct sock *sk = sock->sk;
 196         sk->sk_user_data = (void *)con;
 197         sk->sk_data_ready = ceph_data_ready;
 198         sk->sk_write_space = ceph_write_space;
 199         sk->sk_state_change = ceph_state_change;
 200 }
 201
 202
 203 /*
 204  * socket helpers
 205  */
 206
 207 /*
 208  * initiate connection to a remote socket.
 209  */
 210 static struct socket *ceph_tcp_connect(struct ceph_connection *con)
 211 {
 212         struct sockaddr *paddr = (struct sockaddr *)&con->peer_addr.in_addr;
 213         struct socket *sock;
 214         int ret;
 215
 216         BUG_ON(con->sock);
 217         ret = sock_create_kern(AF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
 218         if (ret)
 219                 return ERR_PTR(ret);
 220         con->sock = sock;
 221         sock->sk->sk_allocation = GFP_NOFS;
 222
 223 #ifdef CONFIG_LOCKDEP
 224         lockdep_set_class(&sock->sk->sk_lock, &socket_class);
 225 #endif
 226
 227         set_sock_callbacks(sock, con);
 228
 229         dout("connect %s\n", pr_addr(&con->peer_addr.in_addr));
 230
 231         ret = sock->ops->connect(sock, paddr, sizeof(*paddr), O_NONBLOCK);
 232         if (ret == -EINPROGRESS) {
 233                 dout("connect %s EINPROGRESS sk_state = %u\n",
 234                      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                        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, pr_addr(&con->peer_addr.in_addr));
 342         set_bit(CLOSED, &con->state);  /* in case there's queued work */
 343         clear_bit(STANDBY, &con->state);  /* avoid connect_seq bump */
 344         clear_bit(LOSSYTX, &con->state);  /* so we retry next connect */
 345         clear_bit(KEEPALIVE_PENDING, &con->state);
 346         clear_bit(WRITE_PENDING, &con->state);
 347         mutex_lock(&con->mutex);
 348         reset_connection(con);
 349         con->peer_global_seq = 0;
 350         cancel_delayed_work(&con->work);
 351         mutex_unlock(&con->mutex);
 352         queue_con(con);
 353 }
 354
 355 /*
 356  * Reopen a closed connection, with a new peer address.
 357  */
 358 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
 359 {
 360         dout("con_open %p %s\n", con, pr_addr(&addr->in_addr));
 361         set_bit(OPENING, &con->state);
 362         clear_bit(CLOSED, &con->state);
 363         memcpy(&con->peer_addr, addr, sizeof(*addr));
 364         con->delay = 0;      /* reset backoff memory */
 365         queue_con(con);
 366 }
 367
 368 /*
 369  * return true if this connection ever successfully opened
 370  */
 371 bool ceph_con_opened(struct ceph_connection *con)
 372 {
 373         return con->connect_seq > 0;
 374 }
 375
 376 /*
 377  * generic get/put
 378  */
 379 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
 380 {
 381         dout("con_get %p nref = %d -> %d\n", con,
 382              atomic_read(&con->nref), atomic_read(&con->nref) + 1);
 383         if (atomic_inc_not_zero(&con->nref))
 384                 return con;
 385         return NULL;
 386 }
 387
 388 void ceph_con_put(struct ceph_connection *con)
 389 {
 390         dout("con_put %p nref = %d -> %d\n", con,
 391              atomic_read(&con->nref), atomic_read(&con->nref) - 1);
 392         BUG_ON(atomic_read(&con->nref) == 0);
 393         if (atomic_dec_and_test(&con->nref)) {
 394                 BUG_ON(con->sock);
 395                 kfree(con);
 396         }
 397 }
 398
 399 /*
 400  * initialize a new connection.
 401  */
 402 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
 403 {
 404         dout("con_init %p\n", con);
 405         memset(con, 0, sizeof(*con));
 406         atomic_set(&con->nref, 1);
 407         con->msgr = msgr;
 408         mutex_init(&con->mutex);
 409         INIT_LIST_HEAD(&con->out_queue);
 410         INIT_LIST_HEAD(&con->out_sent);
 411         INIT_DELAYED_WORK(&con->work, con_work);
 412 }
 413
 414
 415 /*
 416  * We maintain a global counter to order connection attempts.  Get
 417  * a unique seq greater than @gt.
 418  */
 419 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
 420 {
 421         u32 ret;
 422
 423         spin_lock(&msgr->global_seq_lock);
 424         if (msgr->global_seq < gt)
 425                 msgr->global_seq = gt;
 426         ret = ++msgr->global_seq;
 427         spin_unlock(&msgr->global_seq_lock);
 428         return ret;
 429 }
 430
 431
 432 /*
 433  * Prepare footer for currently outgoing message, and finish things
 434  * off.  Assumes out_kvec* are already valid.. we just add on to the end.
 435  */
 436 static void prepare_write_message_footer(struct ceph_connection *con, int v)
 437 {
 438         struct ceph_msg *m = con->out_msg;
 439
 440         dout("prepare_write_message_footer %p\n", con);
 441         con->out_kvec_is_msg = true;
 442         con->out_kvec[v].iov_base = &m->footer;
 443         con->out_kvec[v].iov_len = sizeof(m->footer);
 444         con->out_kvec_bytes += sizeof(m->footer);
 445         con->out_kvec_left++;
 446         con->out_more = m->more_to_follow;
 447         con->out_msg_done = true;
 448 }
 449
 450 /*
 451  * Prepare headers for the next outgoing message.
 452  */
 453 static void prepare_write_message(struct ceph_connection *con)
 454 {
 455         struct ceph_msg *m;
 456         int v = 0;
 457
 458         con->out_kvec_bytes = 0;
 459         con->out_kvec_is_msg = true;
 460         con->out_msg_done = false;
 461
 462         /* Sneak an ack in there first?  If we can get it into the same
 463          * TCP packet that's a good thing. */
 464         if (con->in_seq > con->in_seq_acked) {
 465                 con->in_seq_acked = con->in_seq;
 466                 con->out_kvec[v].iov_base = &tag_ack;
 467                 con->out_kvec[v++].iov_len = 1;
 468                 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 469                 con->out_kvec[v].iov_base = &con->out_temp_ack;
 470                 con->out_kvec[v++].iov_len = sizeof(con->out_temp_ack);
 471                 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 472         }
 473
 474         m = list_first_entry(&con->out_queue,
 475                        struct ceph_msg, list_head);
 476         con->out_msg = m;
 477         if (test_bit(LOSSYTX, &con->state)) {
 478                 list_del_init(&m->list_head);
 479         } else {
 480                 /* put message on sent list */
 481                 ceph_msg_get(m);
 482                 list_move_tail(&m->list_head, &con->out_sent);
 483         }
 484
 485         /*
 486          * only assign outgoing seq # if we haven't sent this message
 487          * yet.  if it is requeued, resend with it's original seq.
 488          */
 489         if (m->needs_out_seq) {
 490                 m->hdr.seq = cpu_to_le64(++con->out_seq);
 491                 m->needs_out_seq = false;
 492         }
 493
 494         dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
 495              m, con->out_seq, le16_to_cpu(m->hdr.type),
 496              le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
 497              le32_to_cpu(m->hdr.data_len),
 498              m->nr_pages);
 499         BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
 500
 501         /* tag + hdr + front + middle */
 502         con->out_kvec[v].iov_base = &tag_msg;
 503         con->out_kvec[v++].iov_len = 1;
 504         con->out_kvec[v].iov_base = &m->hdr;
 505         con->out_kvec[v++].iov_len = sizeof(m->hdr);
 506         con->out_kvec[v++] = m->front;
 507         if (m->middle)
 508                 con->out_kvec[v++] = m->middle->vec;
 509         con->out_kvec_left = v;
 510         con->out_kvec_bytes += 1 + sizeof(m->hdr) + m->front.iov_len +
 511                 (m->middle ? m->middle->vec.iov_len : 0);
 512         con->out_kvec_cur = con->out_kvec;
 513
 514         /* fill in crc (except data pages), footer */
 515         con->out_msg->hdr.crc =
 516                 cpu_to_le32(crc32c(0, (void *)&m->hdr,
 517                                       sizeof(m->hdr) - sizeof(m->hdr.crc)));
 518         con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
 519         con->out_msg->footer.front_crc =
 520                 cpu_to_le32(crc32c(0, m->front.iov_base, m->front.iov_len));
 521         if (m->middle)
 522                 con->out_msg->footer.middle_crc =
 523                         cpu_to_le32(crc32c(0, m->middle->vec.iov_base,
 524                                            m->middle->vec.iov_len));
 525         else
 526                 con->out_msg->footer.middle_crc = 0;
 527         con->out_msg->footer.data_crc = 0;
 528         dout("prepare_write_message front_crc %u data_crc %u\n",
 529              le32_to_cpu(con->out_msg->footer.front_crc),
 530              le32_to_cpu(con->out_msg->footer.middle_crc));
 531
 532         /* is there a data payload? */
 533         if (le32_to_cpu(m->hdr.data_len) > 0) {
 534                 /* initialize page iterator */
 535                 con->out_msg_pos.page = 0;
 536                 con->out_msg_pos.page_pos =
 537                         le16_to_cpu(m->hdr.data_off) & ~PAGE_MASK;
 538                 con->out_msg_pos.data_pos = 0;
 539                 con->out_msg_pos.did_page_crc = 0;
 540                 con->out_more = 1;  /* data + footer will follow */
 541         } else {
 542                 /* no, queue up footer too and be done */
 543                 prepare_write_message_footer(con, v);
 544         }
 545
 546         set_bit(WRITE_PENDING, &con->state);
 547 }
 548
 549 /*
 550  * Prepare an ack.
 551  */
 552 static void prepare_write_ack(struct ceph_connection *con)
 553 {
 554         dout("prepare_write_ack %p %llu -> %llu\n", con,
 555              con->in_seq_acked, con->in_seq);
 556         con->in_seq_acked = con->in_seq;
 557
 558         con->out_kvec[0].iov_base = &tag_ack;
 559         con->out_kvec[0].iov_len = 1;
 560         con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 561         con->out_kvec[1].iov_base = &con->out_temp_ack;
 562         con->out_kvec[1].iov_len = sizeof(con->out_temp_ack);
 563         con->out_kvec_left = 2;
 564         con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 565         con->out_kvec_cur = con->out_kvec;
 566         con->out_more = 1;  /* more will follow.. eventually.. */
 567         set_bit(WRITE_PENDING, &con->state);
 568 }
 569
 570 /*
 571  * Prepare to write keepalive byte.
 572  */
 573 static void prepare_write_keepalive(struct ceph_connection *con)
 574 {
 575         dout("prepare_write_keepalive %p\n", con);
 576         con->out_kvec[0].iov_base = &tag_keepalive;
 577         con->out_kvec[0].iov_len = 1;
 578         con->out_kvec_left = 1;
 579         con->out_kvec_bytes = 1;
 580         con->out_kvec_cur = con->out_kvec;
 581         set_bit(WRITE_PENDING, &con->state);
 582 }
 583
 584 /*
 585  * Connection negotiation.
 586  */
 587
 588 static void prepare_connect_authorizer(struct ceph_connection *con)
 589 {
 590         void *auth_buf;
 591         int auth_len = 0;
 592         int auth_protocol = 0;
 593
 594         mutex_unlock(&con->mutex);
 595         if (con->ops->get_authorizer)
 596                 con->ops->get_authorizer(con, &auth_buf, &auth_len,
 597                                          &auth_protocol, &con->auth_reply_buf,
 598                                          &con->auth_reply_buf_len,
 599                                          con->auth_retry);
 600         mutex_lock(&con->mutex);
 601
 602         con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
 603         con->out_connect.authorizer_len = cpu_to_le32(auth_len);
 604
 605         con->out_kvec[con->out_kvec_left].iov_base = auth_buf;
 606         con->out_kvec[con->out_kvec_left].iov_len = auth_len;
 607         con->out_kvec_left++;
 608         con->out_kvec_bytes += auth_len;
 609 }
 610
 611 /*
 612  * We connected to a peer and are saying hello.
 613  */
 614 static void prepare_write_banner(struct ceph_messenger *msgr,
 615                                  struct ceph_connection *con)
 616 {
 617         int len = strlen(CEPH_BANNER);
 618
 619         con->out_kvec[0].iov_base = CEPH_BANNER;
 620         con->out_kvec[0].iov_len = len;
 621         con->out_kvec[1].iov_base = &msgr->my_enc_addr;
 622         con->out_kvec[1].iov_len = sizeof(msgr->my_enc_addr);
 623         con->out_kvec_left = 2;
 624         con->out_kvec_bytes = len + sizeof(msgr->my_enc_addr);
 625         con->out_kvec_cur = con->out_kvec;
 626         con->out_more = 0;
 627         set_bit(WRITE_PENDING, &con->state);
 628 }
 629
 630 static void prepare_write_connect(struct ceph_messenger *msgr,
 631                                   struct ceph_connection *con,
 632                                   int after_banner)
 633 {
 634         unsigned global_seq = get_global_seq(con->msgr, 0);
 635         int proto;
 636
 637         switch (con->peer_name.type) {
 638         case CEPH_ENTITY_TYPE_MON:
 639                 proto = CEPH_MONC_PROTOCOL;
 640                 break;
 641         case CEPH_ENTITY_TYPE_OSD:
 642                 proto = CEPH_OSDC_PROTOCOL;
 643                 break;
 644         case CEPH_ENTITY_TYPE_MDS:
 645                 proto = CEPH_MDSC_PROTOCOL;
 646                 break;
 647         default:
 648                 BUG();
 649         }
 650
 651         dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
 652              con->connect_seq, global_seq, proto);
 653
 654         con->out_connect.features = CEPH_FEATURE_SUPPORTED_CLIENT;
 655         con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
 656         con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
 657         con->out_connect.global_seq = cpu_to_le32(global_seq);
 658         con->out_connect.protocol_version = cpu_to_le32(proto);
 659         con->out_connect.flags = 0;
 660
 661         if (!after_banner) {
 662                 con->out_kvec_left = 0;
 663                 con->out_kvec_bytes = 0;
 664         }
 665         con->out_kvec[con->out_kvec_left].iov_base = &con->out_connect;
 666         con->out_kvec[con->out_kvec_left].iov_len = sizeof(con->out_connect);
 667         con->out_kvec_left++;
 668         con->out_kvec_bytes += sizeof(con->out_connect);
 669         con->out_kvec_cur = con->out_kvec;
 670         con->out_more = 0;
 671         set_bit(WRITE_PENDING, &con->state);
 672
 673         prepare_connect_authorizer(con);
 674 }
 675
 676
 677 /*
 678  * write as much of pending kvecs to the socket as we can.
 679  *  1 -> done
 680  *  0 -> socket full, but more to do
 681  * <0 -> error
 682  */
 683 static int write_partial_kvec(struct ceph_connection *con)
 684 {
 685         int ret;
 686
 687         dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
 688         while (con->out_kvec_bytes > 0) {
 689                 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
 690                                        con->out_kvec_left, con->out_kvec_bytes,
 691                                        con->out_more);
 692                 if (ret <= 0)
 693                         goto out;
 694                 con->out_kvec_bytes -= ret;
 695                 if (con->out_kvec_bytes == 0)
 696                         break;            /* done */
 697                 while (ret > 0) {
 698                         if (ret >= con->out_kvec_cur->iov_len) {
 699                                 ret -= con->out_kvec_cur->iov_len;
 700                                 con->out_kvec_cur++;
 701                                 con->out_kvec_left--;
 702                         } else {
 703                                 con->out_kvec_cur->iov_len -= ret;
 704                                 con->out_kvec_cur->iov_base += ret;
 705                                 ret = 0;
 706                                 break;
 707                         }
 708                 }
 709         }
 710         con->out_kvec_left = 0;
 711         con->out_kvec_is_msg = false;
 712         ret = 1;
 713 out:
 714         dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
 715              con->out_kvec_bytes, con->out_kvec_left, ret);
 716         return ret;  /* done! */
 717 }
 718
 719 /*
 720  * Write as much message data payload as we can.  If we finish, queue
 721  * up the footer.
 722  *  1 -> done, footer is now queued in out_kvec[].
 723  *  0 -> socket full, but more to do
 724  * <0 -> error
 725  */
 726 static int write_partial_msg_pages(struct ceph_connection *con)
 727 {
 728         struct ceph_msg *msg = con->out_msg;
 729         unsigned data_len = le32_to_cpu(msg->hdr.data_len);
 730         size_t len;
 731         int crc = con->msgr->nocrc;
 732         int ret;
 733
 734         dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
 735              con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
 736              con->out_msg_pos.page_pos);
 737
 738         while (con->out_msg_pos.page < con->out_msg->nr_pages) {
 739                 struct page *page = NULL;
 740                 void *kaddr = NULL;
 741
 742                 /*
 743                  * if we are calculating the data crc (the default), we need
 744                  * to map the page.  if our pages[] has been revoked, use the
 745                  * zero page.
 746                  */
 747                 if (msg->pages) {
 748                         page = msg->pages[con->out_msg_pos.page];
 749                         if (crc)
 750                                 kaddr = kmap(page);
 751                 } else if (msg->pagelist) {
 752                         page = list_first_entry(&msg->pagelist->head,
 753                                                 struct page, lru);
 754                         if (crc)
 755                                 kaddr = kmap(page);
 756                 } else {
 757                         page = con->msgr->zero_page;
 758                         if (crc)
 759                                 kaddr = page_address(con->msgr->zero_page);
 760                 }
 761                 len = min((int)(PAGE_SIZE - con->out_msg_pos.page_pos),
 762                           (int)(data_len - con->out_msg_pos.data_pos));
 763                 if (crc && !con->out_msg_pos.did_page_crc) {
 764                         void *base = kaddr + con->out_msg_pos.page_pos;
 765                         u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
 766
 767                         BUG_ON(kaddr == NULL);
 768                         con->out_msg->footer.data_crc =
 769                                 cpu_to_le32(crc32c(tmpcrc, base, len));
 770                         con->out_msg_pos.did_page_crc = 1;
 771                 }
 772
 773                 ret = kernel_sendpage(con->sock, page,
 774                                       con->out_msg_pos.page_pos, len,
 775                                       MSG_DONTWAIT | MSG_NOSIGNAL |
 776                                       MSG_MORE);
 777
 778                 if (crc && (msg->pages || msg->pagelist))
 779                         kunmap(page);
 780
 781                 if (ret <= 0)
 782                         goto out;
 783
 784                 con->out_msg_pos.data_pos += ret;
 785                 con->out_msg_pos.page_pos += ret;
 786                 if (ret == len) {
 787                         con->out_msg_pos.page_pos = 0;
 788                         con->out_msg_pos.page++;
 789                         con->out_msg_pos.did_page_crc = 0;
 790                         if (msg->pagelist)
 791                                 list_move_tail(&page->lru,
 792                                                &msg->pagelist->head);
 793                 }
 794         }
 795
 796         dout("write_partial_msg_pages %p msg %p done\n", con, msg);
 797
 798         /* prepare and queue up footer, too */
 799         if (!crc)
 800                 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
 801         con->out_kvec_bytes = 0;
 802         con->out_kvec_left = 0;
 803         con->out_kvec_cur = con->out_kvec;
 804         prepare_write_message_footer(con, 0);
 805         ret = 1;
 806 out:
 807         return ret;
 808 }
 809
 810 /*
 811  * write some zeros
 812  */
 813 static int write_partial_skip(struct ceph_connection *con)
 814 {
 815         int ret;
 816
 817         while (con->out_skip > 0) {
 818                 struct kvec iov = {
 819                         .iov_base = page_address(con->msgr->zero_page),
 820                         .iov_len = min(con->out_skip, (int)PAGE_CACHE_SIZE)
 821                 };
 822
 823                 ret = ceph_tcp_sendmsg(con->sock, &iov, 1, iov.iov_len, 1);
 824                 if (ret <= 0)
 825                         goto out;
 826                 con->out_skip -= ret;
 827         }
 828         ret = 1;
 829 out:
 830         return ret;
 831 }
 832
 833 /*
 834  * Prepare to read connection handshake, or an ack.
 835  */
 836 static void prepare_read_banner(struct ceph_connection *con)
 837 {
 838         dout("prepare_read_banner %p\n", con);
 839         con->in_base_pos = 0;
 840 }
 841
 842 static void prepare_read_connect(struct ceph_connection *con)
 843 {
 844         dout("prepare_read_connect %p\n", con);
 845         con->in_base_pos = 0;
 846 }
 847
 848 static void prepare_read_ack(struct ceph_connection *con)
 849 {
 850         dout("prepare_read_ack %p\n", con);
 851         con->in_base_pos = 0;
 852 }
 853
 854 static void prepare_read_tag(struct ceph_connection *con)
 855 {
 856         dout("prepare_read_tag %p\n", con);
 857         con->in_base_pos = 0;
 858         con->in_tag = CEPH_MSGR_TAG_READY;
 859 }
 860
 861 /*
 862  * Prepare to read a message.
 863  */
 864 static int prepare_read_message(struct ceph_connection *con)
 865 {
 866         dout("prepare_read_message %p\n", con);
 867         BUG_ON(con->in_msg != NULL);
 868         con->in_base_pos = 0;
 869         con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
 870         return 0;
 871 }
 872
 873
 874 static int read_partial(struct ceph_connection *con,
 875                         int *to, int size, void *object)
 876 {
 877         *to += size;
 878         while (con->in_base_pos < *to) {
 879                 int left = *to - con->in_base_pos;
 880                 int have = size - left;
 881                 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
 882                 if (ret <= 0)
 883                         return ret;
 884                 con->in_base_pos += ret;
 885         }
 886         return 1;
 887 }
 888
 889
 890 /*
 891  * Read all or part of the connect-side handshake on a new connection
 892  */
 893 static int read_partial_banner(struct ceph_connection *con)
 894 {
 895         int ret, to = 0;
 896
 897         dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
 898
 899         /* peer's banner */
 900         ret = read_partial(con, &to, strlen(CEPH_BANNER), con->in_banner);
 901         if (ret <= 0)
 902                 goto out;
 903         ret = read_partial(con, &to, sizeof(con->actual_peer_addr),
 904                            &con->actual_peer_addr);
 905         if (ret <= 0)
 906                 goto out;
 907         ret = read_partial(con, &to, sizeof(con->peer_addr_for_me),
 908                            &con->peer_addr_for_me);
 909         if (ret <= 0)
 910                 goto out;
 911 out:
 912         return ret;
 913 }
 914
 915 static int read_partial_connect(struct ceph_connection *con)
 916 {
 917         int ret, to = 0;
 918
 919         dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
 920
 921         ret = read_partial(con, &to, sizeof(con->in_reply), &con->in_reply);
 922         if (ret <= 0)
 923                 goto out;
 924         ret = read_partial(con, &to, le32_to_cpu(con->in_reply.authorizer_len),
 925                            con->auth_reply_buf);
 926         if (ret <= 0)
 927                 goto out;
 928
 929         dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
 930              con, (int)con->in_reply.tag,
 931              le32_to_cpu(con->in_reply.connect_seq),
 932              le32_to_cpu(con->in_reply.global_seq));
 933 out:
 934         return ret;
 935
 936 }
 937
 938 /*
 939  * Verify the hello banner looks okay.
 940  */
 941 static int verify_hello(struct ceph_connection *con)
 942 {
 943         if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
 944                 pr_err("connect to %s got bad banner\n",
 945                        pr_addr(&con->peer_addr.in_addr));
 946                 con->error_msg = "protocol error, bad banner";
 947                 return -1;
 948         }
 949         return 0;
 950 }
 951
 952 static bool addr_is_blank(struct sockaddr_storage *ss)
 953 {
 954         switch (ss->ss_family) {
 955         case AF_INET:
 956                 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
 957         case AF_INET6:
 958                 return
 959                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
 960                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
 961                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
 962                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
 963         }
 964         return false;
 965 }
 966
 967 static int addr_port(struct sockaddr_storage *ss)
 968 {
 969         switch (ss->ss_family) {
 970         case AF_INET:
 971                 return ntohs(((struct sockaddr_in *)ss)->sin_port);
 972         case AF_INET6:
 973                 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
 974         }
 975         return 0;
 976 }
 977
 978 static void addr_set_port(struct sockaddr_storage *ss, int p)
 979 {
 980         switch (ss->ss_family) {
 981         case AF_INET:
 982                 ((struct sockaddr_in *)ss)->sin_port = htons(p);
 983         case AF_INET6:
 984                 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
 985         }
 986 }
 987
 988 /*
 989  * Parse an ip[:port] list into an addr array.  Use the default
 990  * monitor port if a port isn't specified.
 991  */
 992 int ceph_parse_ips(const char *c, const char *end,
 993                    struct ceph_entity_addr *addr,
 994                    int max_count, int *count)
 995 {
 996         int i;
 997         const char *p = c;
 998
 999         dout("parse_ips on '%.*s'\n", (int)(end-c), c);
1000         for (i = 0; i < max_count; i++) {
1001                 const char *ipend;
1002                 struct sockaddr_storage *ss = &addr[i].in_addr;
1003                 struct sockaddr_in *in4 = (void *)ss;
1004                 struct sockaddr_in6 *in6 = (void *)ss;
1005                 int port;
1006
1007                 memset(ss, 0, sizeof(*ss));
1008                 if (in4_pton(p, end - p, (u8 *)&in4->sin_addr.s_addr,
1009                              ',', &ipend)) {
1010                         ss->ss_family = AF_INET;
1011                 } else if (in6_pton(p, end - p, (u8 *)&in6->sin6_addr.s6_addr,
1012                                     ',', &ipend)) {
1013                         ss->ss_family = AF_INET6;
1014                 } else {
1015                         goto bad;
1016                 }
1017                 p = ipend;
1018
1019                 /* port? */
1020                 if (p < end && *p == ':') {
1021                         port = 0;
1022                         p++;
1023                         while (p < end && *p >= '0' && *p <= '9') {
1024                                 port = (port * 10) + (*p - '0');
1025                                 p++;
1026                         }
1027                         if (port > 65535 || port == 0)
1028                                 goto bad;
1029                 } else {
1030                         port = CEPH_MON_PORT;
1031                 }
1032
1033                 addr_set_port(ss, port);
1034
1035                 dout("parse_ips got %s\n", pr_addr(ss));
1036
1037                 if (p == end)
1038                         break;
1039                 if (*p != ',')
1040                         goto bad;
1041                 p++;
1042         }
1043
1044         if (p != end)
1045                 goto bad;
1046
1047         if (count)
1048                 *count = i + 1;
1049         return 0;
1050
1051 bad:
1052         pr_err("parse_ips bad ip '%s'\n", c);
1053         return -EINVAL;
1054 }
1055
1056 static int process_banner(struct ceph_connection *con)
1057 {
1058         dout("process_banner on %p\n", con);
1059
1060         if (verify_hello(con) < 0)
1061                 return -1;
1062
1063         ceph_decode_addr(&con->actual_peer_addr);
1064         ceph_decode_addr(&con->peer_addr_for_me);
1065
1066         /*
1067          * Make sure the other end is who we wanted.  note that the other
1068          * end may not yet know their ip address, so if it's 0.0.0.0, give
1069          * them the benefit of the doubt.
1070          */
1071         if (memcmp(&con->peer_addr, &con->actual_peer_addr,
1072                    sizeof(con->peer_addr)) != 0 &&
1073             !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1074               con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1075                 pr_warning("wrong peer, want %s/%lld, got %s/%lld\n",
1076                            pr_addr(&con->peer_addr.in_addr),
1077                            le64_to_cpu(con->peer_addr.nonce),
1078                            pr_addr(&con->actual_peer_addr.in_addr),
1079                            le64_to_cpu(con->actual_peer_addr.nonce));
1080                 con->error_msg = "wrong peer at address";
1081                 return -1;
1082         }
1083
1084         /*
1085          * did we learn our address?
1086          */
1087         if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1088                 int port = addr_port(&con->msgr->inst.addr.in_addr);
1089
1090                 memcpy(&con->msgr->inst.addr.in_addr,
1091                        &con->peer_addr_for_me.in_addr,
1092                        sizeof(con->peer_addr_for_me.in_addr));
1093                 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1094                 encode_my_addr(con->msgr);
1095                 dout("process_banner learned my addr is %s\n",
1096                      pr_addr(&con->msgr->inst.addr.in_addr));
1097         }
1098
1099         set_bit(NEGOTIATING, &con->state);
1100         prepare_read_connect(con);
1101         return 0;
1102 }
1103
1104 static void fail_protocol(struct ceph_connection *con)
1105 {
1106         reset_connection(con);
1107         set_bit(CLOSED, &con->state);  /* in case there's queued work */
1108
1109         mutex_unlock(&con->mutex);
1110         if (con->ops->bad_proto)
1111                 con->ops->bad_proto(con);
1112         mutex_lock(&con->mutex);
1113 }
1114
1115 static int process_connect(struct ceph_connection *con)
1116 {
1117         u64 sup_feat = CEPH_FEATURE_SUPPORTED_CLIENT;
1118         u64 req_feat = CEPH_FEATURE_REQUIRED_CLIENT;
1119         u64 server_feat = le64_to_cpu(con->in_reply.features);
1120
1121         dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1122
1123         switch (con->in_reply.tag) {
1124         case CEPH_MSGR_TAG_FEATURES:
1125                 pr_err("%s%lld %s feature set mismatch,"
1126                        " my %llx < server's %llx, missing %llx\n",
1127                        ENTITY_NAME(con->peer_name),
1128                        pr_addr(&con->peer_addr.in_addr),
1129                        sup_feat, server_feat, server_feat & ~sup_feat);
1130                 con->error_msg = "missing required protocol features";
1131                 fail_protocol(con);
1132                 return -1;
1133
1134         case CEPH_MSGR_TAG_BADPROTOVER:
1135                 pr_err("%s%lld %s protocol version mismatch,"
1136                        " my %d != server's %d\n",
1137                        ENTITY_NAME(con->peer_name),
1138                        pr_addr(&con->peer_addr.in_addr),
1139                        le32_to_cpu(con->out_connect.protocol_version),
1140                        le32_to_cpu(con->in_reply.protocol_version));
1141                 con->error_msg = "protocol version mismatch";
1142                 fail_protocol(con);
1143                 return -1;
1144
1145         case CEPH_MSGR_TAG_BADAUTHORIZER:
1146                 con->auth_retry++;
1147                 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1148                      con->auth_retry);
1149                 if (con->auth_retry == 2) {
1150                         con->error_msg = "connect authorization failure";
1151                         reset_connection(con);
1152                         set_bit(CLOSED, &con->state);
1153                         return -1;
1154                 }
1155                 con->auth_retry = 1;
1156                 prepare_write_connect(con->msgr, con, 0);
1157                 prepare_read_connect(con);
1158                 break;
1159
1160         case CEPH_MSGR_TAG_RESETSESSION:
1161                 /*
1162                  * If we connected with a large connect_seq but the peer
1163                  * has no record of a session with us (no connection, or
1164                  * connect_seq == 0), they will send RESETSESION to indicate
1165                  * that they must have reset their session, and may have
1166                  * dropped messages.
1167                  */
1168                 dout("process_connect got RESET peer seq %u\n",
1169                      le32_to_cpu(con->in_connect.connect_seq));
1170                 pr_err("%s%lld %s connection reset\n",
1171                        ENTITY_NAME(con->peer_name),
1172                        pr_addr(&con->peer_addr.in_addr));
1173                 reset_connection(con);
1174                 prepare_write_connect(con->msgr, con, 0);
1175                 prepare_read_connect(con);
1176
1177                 /* Tell ceph about it. */
1178                 mutex_unlock(&con->mutex);
1179                 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1180                 if (con->ops->peer_reset)
1181                         con->ops->peer_reset(con);
1182                 mutex_lock(&con->mutex);
1183                 break;
1184
1185         case CEPH_MSGR_TAG_RETRY_SESSION:
1186                 /*
1187                  * If we sent a smaller connect_seq than the peer has, try
1188                  * again with a larger value.
1189                  */
1190                 dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
1191                      le32_to_cpu(con->out_connect.connect_seq),
1192                      le32_to_cpu(con->in_connect.connect_seq));
1193                 con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
1194                 prepare_write_connect(con->msgr, con, 0);
1195                 prepare_read_connect(con);
1196                 break;
1197
1198         case CEPH_MSGR_TAG_RETRY_GLOBAL:
1199                 /*
1200                  * If we sent a smaller global_seq than the peer has, try
1201                  * again with a larger value.
1202                  */
1203                 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1204                      con->peer_global_seq,
1205                      le32_to_cpu(con->in_connect.global_seq));
1206                 get_global_seq(con->msgr,
1207                                le32_to_cpu(con->in_connect.global_seq));
1208                 prepare_write_connect(con->msgr, con, 0);
1209                 prepare_read_connect(con);
1210                 break;
1211
1212         case CEPH_MSGR_TAG_READY:
1213                 if (req_feat & ~server_feat) {
1214                         pr_err("%s%lld %s protocol feature mismatch,"
1215                                " my required %llx > server's %llx, need %llx\n",
1216                                ENTITY_NAME(con->peer_name),
1217                                pr_addr(&con->peer_addr.in_addr),
1218                                req_feat, server_feat, req_feat & ~server_feat);
1219                         con->error_msg = "missing required protocol features";
1220                         fail_protocol(con);
1221                         return -1;
1222                 }
1223                 clear_bit(CONNECTING, &con->state);
1224                 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1225                 con->connect_seq++;
1226                 con->peer_features = server_feat;
1227                 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1228                      con->peer_global_seq,
1229                      le32_to_cpu(con->in_reply.connect_seq),
1230                      con->connect_seq);
1231                 WARN_ON(con->connect_seq !=
1232                         le32_to_cpu(con->in_reply.connect_seq));
1233
1234                 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1235                         set_bit(LOSSYTX, &con->state);
1236
1237                 prepare_read_tag(con);
1238                 break;
1239
1240         case CEPH_MSGR_TAG_WAIT:
1241                 /*
1242                  * If there is a connection race (we are opening
1243                  * connections to each other), one of us may just have
1244                  * to WAIT.  This shouldn't happen if we are the
1245                  * client.
1246                  */
1247                 pr_err("process_connect peer connecting WAIT\n");
1248
1249         default:
1250                 pr_err("connect protocol error, will retry\n");
1251                 con->error_msg = "protocol error, garbage tag during connect";
1252                 return -1;
1253         }
1254         return 0;
1255 }
1256
1257
1258 /*
1259  * read (part of) an ack
1260  */
1261 static int read_partial_ack(struct ceph_connection *con)
1262 {
1263         int to = 0;
1264
1265         return read_partial(con, &to, sizeof(con->in_temp_ack),
1266                             &con->in_temp_ack);
1267 }
1268
1269
1270 /*
1271  * We can finally discard anything that's been acked.
1272  */
1273 static void process_ack(struct ceph_connection *con)
1274 {
1275         struct ceph_msg *m;
1276         u64 ack = le64_to_cpu(con->in_temp_ack);
1277         u64 seq;
1278
1279         while (!list_empty(&con->out_sent)) {
1280                 m = list_first_entry(&con->out_sent, struct ceph_msg,
1281                                      list_head);
1282                 seq = le64_to_cpu(m->hdr.seq);
1283                 if (seq > ack)
1284                         break;
1285                 dout("got ack for seq %llu type %d at %p\n", seq,
1286                      le16_to_cpu(m->hdr.type), m);
1287                 ceph_msg_remove(m);
1288         }
1289         prepare_read_tag(con);
1290 }
1291
1292
1293
1294
1295 static int read_partial_message_section(struct ceph_connection *con,
1296                                         struct kvec *section, unsigned int sec_len,
1297                                         u32 *crc)
1298 {
1299         int left;
1300         int ret;
1301
1302         BUG_ON(!section);
1303
1304         while (section->iov_len < sec_len) {
1305                 BUG_ON(section->iov_base == NULL);
1306                 left = sec_len - section->iov_len;
1307                 ret = ceph_tcp_recvmsg(con->sock, (char *)section->iov_base +
1308                                        section->iov_len, left);
1309                 if (ret <= 0)
1310                         return ret;
1311                 section->iov_len += ret;
1312                 if (section->iov_len == sec_len)
1313                         *crc = crc32c(0, section->iov_base,
1314                                       section->iov_len);
1315         }
1316
1317         return 1;
1318 }
1319
1320 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
1321                                 struct ceph_msg_header *hdr,
1322                                 int *skip);
1323 /*
1324  * read (part of) a message.
1325  */
1326 static int read_partial_message(struct ceph_connection *con)
1327 {
1328         struct ceph_msg *m = con->in_msg;
1329         void *p;
1330         int ret;
1331         int to, left;
1332         unsigned front_len, middle_len, data_len, data_off;
1333         int datacrc = con->msgr->nocrc;
1334         int skip;
1335         u64 seq;
1336
1337         dout("read_partial_message con %p msg %p\n", con, m);
1338
1339         /* header */
1340         while (con->in_base_pos < sizeof(con->in_hdr)) {
1341                 left = sizeof(con->in_hdr) - con->in_base_pos;
1342                 ret = ceph_tcp_recvmsg(con->sock,
1343                                        (char *)&con->in_hdr + con->in_base_pos,
1344                                        left);
1345                 if (ret <= 0)
1346                         return ret;
1347                 con->in_base_pos += ret;
1348                 if (con->in_base_pos == sizeof(con->in_hdr)) {
1349                         u32 crc = crc32c(0, (void *)&con->in_hdr,
1350                                  sizeof(con->in_hdr) - sizeof(con->in_hdr.crc));
1351                         if (crc != le32_to_cpu(con->in_hdr.crc)) {
1352                                 pr_err("read_partial_message bad hdr "
1353                                        " crc %u != expected %u\n",
1354                                        crc, con->in_hdr.crc);
1355                                 return -EBADMSG;
1356                         }
1357                 }
1358         }
1359         front_len = le32_to_cpu(con->in_hdr.front_len);
1360         if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1361                 return -EIO;
1362         middle_len = le32_to_cpu(con->in_hdr.middle_len);
1363         if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1364                 return -EIO;
1365         data_len = le32_to_cpu(con->in_hdr.data_len);
1366         if (data_len > CEPH_MSG_MAX_DATA_LEN)
1367                 return -EIO;
1368         data_off = le16_to_cpu(con->in_hdr.data_off);
1369
1370         /* verify seq# */
1371         seq = le64_to_cpu(con->in_hdr.seq);
1372         if ((s64)seq - (s64)con->in_seq < 1) {
1373                 pr_info("skipping %s%lld %s seq %lld, expected %lld\n",
1374                         ENTITY_NAME(con->peer_name),
1375                         pr_addr(&con->peer_addr.in_addr),
1376                         seq, con->in_seq + 1);
1377                 con->in_base_pos = -front_len - middle_len - data_len -
1378                         sizeof(m->footer);
1379                 con->in_tag = CEPH_MSGR_TAG_READY;
1380                 con->in_seq++;
1381                 return 0;
1382         } else if ((s64)seq - (s64)con->in_seq > 1) {
1383                 pr_err("read_partial_message bad seq %lld expected %lld\n",
1384                        seq, con->in_seq + 1);
1385                 con->error_msg = "bad message sequence # for incoming message";
1386                 return -EBADMSG;
1387         }
1388
1389         /* allocate message? */
1390         if (!con->in_msg) {
1391                 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1392                      con->in_hdr.front_len, con->in_hdr.data_len);
1393                 con->in_msg = ceph_alloc_msg(con, &con->in_hdr, &skip);
1394                 if (skip) {
1395                         /* skip this message */
1396                         dout("alloc_msg said skip message\n");
1397                         con->in_base_pos = -front_len - middle_len - data_len -
1398                                 sizeof(m->footer);
1399                         con->in_tag = CEPH_MSGR_TAG_READY;
1400                         con->in_seq++;
1401                         return 0;
1402                 }
1403                 if (!con->in_msg) {
1404                         con->error_msg =
1405                                 "error allocating memory for incoming message";
1406                         return -ENOMEM;
1407                 }
1408                 m = con->in_msg;
1409                 m->front.iov_len = 0;    /* haven't read it yet */
1410                 if (m->middle)
1411                         m->middle->vec.iov_len = 0;
1412
1413                 con->in_msg_pos.page = 0;
1414                 con->in_msg_pos.page_pos = data_off & ~PAGE_MASK;
1415                 con->in_msg_pos.data_pos = 0;
1416         }
1417
1418         /* front */
1419         ret = read_partial_message_section(con, &m->front, front_len,
1420                                            &con->in_front_crc);
1421         if (ret <= 0)
1422                 return ret;
1423
1424         /* middle */
1425         if (m->middle) {
1426                 ret = read_partial_message_section(con, &m->middle->vec, middle_len,
1427                                                    &con->in_middle_crc);
1428                 if (ret <= 0)
1429                         return ret;
1430         }
1431
1432         /* (page) data */
1433         while (con->in_msg_pos.data_pos < data_len) {
1434                 left = min((int)(data_len - con->in_msg_pos.data_pos),
1435                            (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1436                 BUG_ON(m->pages == NULL);
1437                 p = kmap(m->pages[con->in_msg_pos.page]);
1438                 ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1439                                        left);
1440                 if (ret > 0 && datacrc)
1441                         con->in_data_crc =
1442                                 crc32c(con->in_data_crc,
1443                                           p + con->in_msg_pos.page_pos, ret);
1444                 kunmap(m->pages[con->in_msg_pos.page]);
1445                 if (ret <= 0)
1446                         return ret;
1447                 con->in_msg_pos.data_pos += ret;
1448                 con->in_msg_pos.page_pos += ret;
1449                 if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1450                         con->in_msg_pos.page_pos = 0;
1451                         con->in_msg_pos.page++;
1452                 }
1453         }
1454
1455         /* footer */
1456         to = sizeof(m->hdr) + sizeof(m->footer);
1457         while (con->in_base_pos < to) {
1458                 left = to - con->in_base_pos;
1459                 ret = ceph_tcp_recvmsg(con->sock, (char *)&m->footer +
1460                                        (con->in_base_pos - sizeof(m->hdr)),
1461                                        left);
1462                 if (ret <= 0)
1463                         return ret;
1464                 con->in_base_pos += ret;
1465         }
1466         dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1467              m, front_len, m->footer.front_crc, middle_len,
1468              m->footer.middle_crc, data_len, m->footer.data_crc);
1469
1470         /* crc ok? */
1471         if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1472                 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1473                        m, con->in_front_crc, m->footer.front_crc);
1474                 return -EBADMSG;
1475         }
1476         if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1477                 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1478                        m, con->in_middle_crc, m->footer.middle_crc);
1479                 return -EBADMSG;
1480         }
1481         if (datacrc &&
1482             (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1483             con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1484                 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1485                        con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1486                 return -EBADMSG;
1487         }
1488
1489         return 1; /* done! */
1490 }
1491
1492 /*
1493  * Process message.  This happens in the worker thread.  The callback should
1494  * be careful not to do anything that waits on other incoming messages or it
1495  * may deadlock.
1496  */
1497 static void process_message(struct ceph_connection *con)
1498 {
1499         struct ceph_msg *msg;
1500
1501         msg = con->in_msg;
1502         con->in_msg = NULL;
1503
1504         /* if first message, set peer_name */
1505         if (con->peer_name.type == 0)
1506                 con->peer_name = msg->hdr.src;
1507
1508         con->in_seq++;
1509         mutex_unlock(&con->mutex);
1510
1511         dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1512              msg, le64_to_cpu(msg->hdr.seq),
1513              ENTITY_NAME(msg->hdr.src),
1514              le16_to_cpu(msg->hdr.type),
1515              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1516              le32_to_cpu(msg->hdr.front_len),
1517              le32_to_cpu(msg->hdr.data_len),
1518              con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1519         con->ops->dispatch(con, msg);
1520
1521         mutex_lock(&con->mutex);
1522         prepare_read_tag(con);
1523 }
1524
1525
1526 /*
1527  * Write something to the socket.  Called in a worker thread when the
1528  * socket appears to be writeable and we have something ready to send.
1529  */
1530 static int try_write(struct ceph_connection *con)
1531 {
1532         struct ceph_messenger *msgr = con->msgr;
1533         int ret = 1;
1534
1535         dout("try_write start %p state %lu nref %d\n", con, con->state,
1536              atomic_read(&con->nref));
1537
1538 more:
1539         dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1540
1541         /* open the socket first? */
1542         if (con->sock == NULL) {
1543                 /*
1544                  * if we were STANDBY and are reconnecting _this_
1545                  * connection, bump connect_seq now.  Always bump
1546                  * global_seq.
1547                  */
1548                 if (test_and_clear_bit(STANDBY, &con->state))
1549                         con->connect_seq++;
1550
1551                 prepare_write_banner(msgr, con);
1552                 prepare_write_connect(msgr, con, 1);
1553                 prepare_read_banner(con);
1554                 set_bit(CONNECTING, &con->state);
1555                 clear_bit(NEGOTIATING, &con->state);
1556
1557                 BUG_ON(con->in_msg);
1558                 con->in_tag = CEPH_MSGR_TAG_READY;
1559                 dout("try_write initiating connect on %p new state %lu\n",
1560                      con, con->state);
1561                 con->sock = ceph_tcp_connect(con);
1562                 if (IS_ERR(con->sock)) {
1563                         con->sock = NULL;
1564                         con->error_msg = "connect error";
1565                         ret = -1;
1566                         goto out;
1567                 }
1568         }
1569
1570 more_kvec:
1571         /* kvec data queued? */
1572         if (con->out_skip) {
1573                 ret = write_partial_skip(con);
1574                 if (ret <= 0)
1575                         goto done;
1576                 if (ret < 0) {
1577                         dout("try_write write_partial_skip err %d\n", ret);
1578                         goto done;
1579                 }
1580         }
1581         if (con->out_kvec_left) {
1582                 ret = write_partial_kvec(con);
1583                 if (ret <= 0)
1584                         goto done;
1585         }
1586
1587         /* msg pages? */
1588         if (con->out_msg) {
1589                 if (con->out_msg_done) {
1590                         ceph_msg_put(con->out_msg);
1591                         con->out_msg = NULL;   /* we're done with this one */
1592                         goto do_next;
1593                 }
1594
1595                 ret = write_partial_msg_pages(con);
1596                 if (ret == 1)
1597                         goto more_kvec;  /* we need to send the footer, too! */
1598                 if (ret == 0)
1599                         goto done;
1600                 if (ret < 0) {
1601                         dout("try_write write_partial_msg_pages err %d\n",
1602                              ret);
1603                         goto done;
1604                 }
1605         }
1606
1607 do_next:
1608         if (!test_bit(CONNECTING, &con->state)) {
1609                 /* is anything else pending? */
1610                 if (!list_empty(&con->out_queue)) {
1611                         prepare_write_message(con);
1612                         goto more;
1613                 }
1614                 if (con->in_seq > con->in_seq_acked) {
1615                         prepare_write_ack(con);
1616                         goto more;
1617                 }
1618                 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1619                         prepare_write_keepalive(con);
1620                         goto more;
1621                 }
1622         }
1623
1624         /* Nothing to do! */
1625         clear_bit(WRITE_PENDING, &con->state);
1626         dout("try_write nothing else to write.\n");
1627 done:
1628         ret = 0;
1629 out:
1630         dout("try_write done on %p\n", con);
1631         return ret;
1632 }
1633
1634
1635
1636 /*
1637  * Read what we can from the socket.
1638  */
1639 static int try_read(struct ceph_connection *con)
1640 {
1641         int ret = -1;
1642
1643         if (!con->sock)
1644                 return 0;
1645
1646         if (test_bit(STANDBY, &con->state))
1647                 return 0;
1648
1649         dout("try_read start on %p\n", con);
1650
1651 more:
1652         dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1653              con->in_base_pos);
1654         if (test_bit(CONNECTING, &con->state)) {
1655                 if (!test_bit(NEGOTIATING, &con->state)) {
1656                         dout("try_read connecting\n");
1657                         ret = read_partial_banner(con);
1658                         if (ret <= 0)
1659                                 goto done;
1660                         if (process_banner(con) < 0) {
1661                                 ret = -1;
1662                                 goto out;
1663                         }
1664                 }
1665                 ret = read_partial_connect(con);
1666                 if (ret <= 0)
1667                         goto done;
1668                 if (process_connect(con) < 0) {
1669                         ret = -1;
1670                         goto out;
1671                 }
1672                 goto more;
1673         }
1674
1675         if (con->in_base_pos < 0) {
1676                 /*
1677                  * skipping + discarding content.
1678                  *
1679                  * FIXME: there must be a better way to do this!
1680                  */
1681                 static char buf[1024];
1682                 int skip = min(1024, -con->in_base_pos);
1683                 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
1684                 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
1685                 if (ret <= 0)
1686                         goto done;
1687                 con->in_base_pos += ret;
1688                 if (con->in_base_pos)
1689                         goto more;
1690         }
1691         if (con->in_tag == CEPH_MSGR_TAG_READY) {
1692                 /*
1693                  * what's next?
1694                  */
1695                 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
1696                 if (ret <= 0)
1697                         goto done;
1698                 dout("try_read got tag %d\n", (int)con->in_tag);
1699                 switch (con->in_tag) {
1700                 case CEPH_MSGR_TAG_MSG:
1701                         prepare_read_message(con);
1702                         break;
1703                 case CEPH_MSGR_TAG_ACK:
1704                         prepare_read_ack(con);
1705                         break;
1706                 case CEPH_MSGR_TAG_CLOSE:
1707                         set_bit(CLOSED, &con->state);   /* fixme */
1708                         goto done;
1709                 default:
1710                         goto bad_tag;
1711                 }
1712         }
1713         if (con->in_tag == CEPH_MSGR_TAG_MSG) {
1714                 ret = read_partial_message(con);
1715                 if (ret <= 0) {
1716                         switch (ret) {
1717                         case -EBADMSG:
1718                                 con->error_msg = "bad crc";
1719                                 ret = -EIO;
1720                                 goto out;
1721                         case -EIO:
1722                                 con->error_msg = "io error";
1723                                 goto out;
1724                         default:
1725                                 goto done;
1726                         }
1727                 }
1728                 if (con->in_tag == CEPH_MSGR_TAG_READY)
1729                         goto more;
1730                 process_message(con);
1731                 goto more;
1732         }
1733         if (con->in_tag == CEPH_MSGR_TAG_ACK) {
1734                 ret = read_partial_ack(con);
1735                 if (ret <= 0)
1736                         goto done;
1737                 process_ack(con);
1738                 goto more;
1739         }
1740
1741 done:
1742         ret = 0;
1743 out:
1744         dout("try_read done on %p\n", con);
1745         return ret;
1746
1747 bad_tag:
1748         pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
1749         con->error_msg = "protocol error, garbage tag";
1750         ret = -1;
1751         goto out;
1752 }
1753
1754
1755 /*
1756  * Atomically queue work on a connection.  Bump @con reference to
1757  * avoid races with connection teardown.
1758  *
1759  * There is some trickery going on with QUEUED and BUSY because we
1760  * only want a _single_ thread operating on each connection at any
1761  * point in time, but we want to use all available CPUs.
1762  *
1763  * The worker thread only proceeds if it can atomically set BUSY.  It
1764  * clears QUEUED and does it's thing.  When it thinks it's done, it
1765  * clears BUSY, then rechecks QUEUED.. if it's set again, it loops
1766  * (tries again to set BUSY).
1767  *
1768  * To queue work, we first set QUEUED, _then_ if BUSY isn't set, we
1769  * try to queue work.  If that fails (work is already queued, or BUSY)
1770  * we give up (work also already being done or is queued) but leave QUEUED
1771  * set so that the worker thread will loop if necessary.
1772  */
1773 static void queue_con(struct ceph_connection *con)
1774 {
1775         if (test_bit(DEAD, &con->state)) {
1776                 dout("queue_con %p ignoring: DEAD\n",
1777                      con);
1778                 return;
1779         }
1780
1781         if (!con->ops->get(con)) {
1782                 dout("queue_con %p ref count 0\n", con);
1783                 return;
1784         }
1785
1786         set_bit(QUEUED, &con->state);
1787         if (test_bit(BUSY, &con->state)) {
1788                 dout("queue_con %p - already BUSY\n", con);
1789                 con->ops->put(con);
1790         } else if (!queue_work(ceph_msgr_wq, &con->work.work)) {
1791                 dout("queue_con %p - already queued\n", con);
1792                 con->ops->put(con);
1793         } else {
1794                 dout("queue_con %p\n", con);
1795         }
1796 }
1797
1798 /*
1799  * Do some work on a connection.  Drop a connection ref when we're done.
1800  */
1801 static void con_work(struct work_struct *work)
1802 {
1803         struct ceph_connection *con = container_of(work, struct ceph_connection,
1804                                                    work.work);
1805         int backoff = 0;
1806
1807 more:
1808         if (test_and_set_bit(BUSY, &con->state) != 0) {
1809                 dout("con_work %p BUSY already set\n", con);
1810                 goto out;
1811         }
1812         dout("con_work %p start, clearing QUEUED\n", con);
1813         clear_bit(QUEUED, &con->state);
1814
1815         mutex_lock(&con->mutex);
1816
1817         if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
1818                 dout("con_work CLOSED\n");
1819                 con_close_socket(con);
1820                 goto done;
1821         }
1822         if (test_and_clear_bit(OPENING, &con->state)) {
1823                 /* reopen w/ new peer */
1824                 dout("con_work OPENING\n");
1825                 con_close_socket(con);
1826         }
1827
1828         if (test_and_clear_bit(SOCK_CLOSED, &con->state) ||
1829             try_read(con) < 0 ||
1830             try_write(con) < 0) {
1831                 mutex_unlock(&con->mutex);
1832                 backoff = 1;
1833                 ceph_fault(con);     /* error/fault path */
1834                 goto done_unlocked;
1835         }
1836
1837 done:
1838         mutex_unlock(&con->mutex);
1839
1840 done_unlocked:
1841         clear_bit(BUSY, &con->state);
1842         dout("con->state=%lu\n", con->state);
1843         if (test_bit(QUEUED, &con->state)) {
1844                 if (!backoff || test_bit(OPENING, &con->state)) {
1845                         dout("con_work %p QUEUED reset, looping\n", con);
1846                         goto more;
1847                 }
1848                 dout("con_work %p QUEUED reset, but just faulted\n", con);
1849                 clear_bit(QUEUED, &con->state);
1850         }
1851         dout("con_work %p done\n", con);
1852
1853 out:
1854         con->ops->put(con);
1855 }
1856
1857
1858 /*
1859  * Generic error/fault handler.  A retry mechanism is used with
1860  * exponential backoff
1861  */
1862 static void ceph_fault(struct ceph_connection *con)
1863 {
1864         pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
1865                pr_addr(&con->peer_addr.in_addr), con->error_msg);
1866         dout("fault %p state %lu to peer %s\n",
1867              con, con->state, pr_addr(&con->peer_addr.in_addr));
1868
1869         if (test_bit(LOSSYTX, &con->state)) {
1870                 dout("fault on LOSSYTX channel\n");
1871                 goto out;
1872         }
1873
1874         mutex_lock(&con->mutex);
1875         if (test_bit(CLOSED, &con->state))
1876                 goto out_unlock;
1877
1878         con_close_socket(con);
1879
1880         if (con->in_msg) {
1881                 ceph_msg_put(con->in_msg);
1882                 con->in_msg = NULL;
1883         }
1884
1885         /* Requeue anything that hasn't been acked */
1886         list_splice_init(&con->out_sent, &con->out_queue);
1887
1888         /* If there are no messages in the queue, place the connection
1889          * in a STANDBY state (i.e., don't try to reconnect just yet). */
1890         if (list_empty(&con->out_queue) && !con->out_keepalive_pending) {
1891                 dout("fault setting STANDBY\n");
1892                 set_bit(STANDBY, &con->state);
1893         } else {
1894                 /* retry after a delay. */
1895                 if (con->delay == 0)
1896                         con->delay = BASE_DELAY_INTERVAL;
1897                 else if (con->delay < MAX_DELAY_INTERVAL)
1898                         con->delay *= 2;
1899                 dout("fault queueing %p delay %lu\n", con, con->delay);
1900                 con->ops->get(con);
1901                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
1902                                        round_jiffies_relative(con->delay)) == 0)
1903                         con->ops->put(con);
1904         }
1905
1906 out_unlock:
1907         mutex_unlock(&con->mutex);
1908 out:
1909         /*
1910          * in case we faulted due to authentication, invalidate our
1911          * current tickets so that we can get new ones.
1912          */
1913         if (con->auth_retry && con->ops->invalidate_authorizer) {
1914                 dout("calling invalidate_authorizer()\n");
1915                 con->ops->invalidate_authorizer(con);
1916         }
1917
1918         if (con->ops->fault)
1919                 con->ops->fault(con);
1920 }
1921
1922
1923
1924 /*
1925  * create a new messenger instance
1926  */
1927 struct ceph_messenger *ceph_messenger_create(struct ceph_entity_addr *myaddr)
1928 {
1929         struct ceph_messenger *msgr;
1930
1931         msgr = kzalloc(sizeof(*msgr), GFP_KERNEL);
1932         if (msgr == NULL)
1933                 return ERR_PTR(-ENOMEM);
1934
1935         spin_lock_init(&msgr->global_seq_lock);
1936
1937         /* the zero page is needed if a request is "canceled" while the message
1938          * is being written over the socket */
1939         msgr->zero_page = __page_cache_alloc(GFP_KERNEL | __GFP_ZERO);
1940         if (!msgr->zero_page) {
1941                 kfree(msgr);
1942                 return ERR_PTR(-ENOMEM);
1943         }
1944         kmap(msgr->zero_page);
1945
1946         if (myaddr)
1947                 msgr->inst.addr = *myaddr;
1948
1949         /* select a random nonce */
1950         msgr->inst.addr.type = 0;
1951         get_random_bytes(&msgr->inst.addr.nonce, sizeof(msgr->inst.addr.nonce));
1952         encode_my_addr(msgr);
1953
1954         dout("messenger_create %p\n", msgr);
1955         return msgr;
1956 }
1957
1958 void ceph_messenger_destroy(struct ceph_messenger *msgr)
1959 {
1960         dout("destroy %p\n", msgr);
1961         kunmap(msgr->zero_page);
1962         __free_page(msgr->zero_page);
1963         kfree(msgr);
1964         dout("destroyed messenger %p\n", msgr);
1965 }
1966
1967 /*
1968  * Queue up an outgoing message on the given connection.
1969  */
1970 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
1971 {
1972         if (test_bit(CLOSED, &con->state)) {
1973                 dout("con_send %p closed, dropping %p\n", con, msg);
1974                 ceph_msg_put(msg);
1975                 return;
1976         }
1977
1978         /* set src+dst */
1979         msg->hdr.src = con->msgr->inst.name;
1980
1981         BUG_ON(msg->front.iov_len != le32_to_cpu(msg->hdr.front_len));
1982
1983         msg->needs_out_seq = true;
1984
1985         /* queue */
1986         mutex_lock(&con->mutex);
1987         BUG_ON(!list_empty(&msg->list_head));
1988         list_add_tail(&msg->list_head, &con->out_queue);
1989         dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
1990              ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
1991              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1992              le32_to_cpu(msg->hdr.front_len),
1993              le32_to_cpu(msg->hdr.middle_len),
1994              le32_to_cpu(msg->hdr.data_len));
1995         mutex_unlock(&con->mutex);
1996
1997         /* if there wasn't anything waiting to send before, queue
1998          * new work */
1999         if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2000                 queue_con(con);
2001 }
2002
2003 /*
2004  * Revoke a message that was previously queued for send
2005  */
2006 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
2007 {
2008         mutex_lock(&con->mutex);
2009         if (!list_empty(&msg->list_head)) {
2010                 dout("con_revoke %p msg %p\n", con, msg);
2011                 list_del_init(&msg->list_head);
2012                 ceph_msg_put(msg);
2013                 msg->hdr.seq = 0;
2014                 if (con->out_msg == msg) {
2015                         ceph_msg_put(con->out_msg);
2016                         con->out_msg = NULL;
2017                 }
2018                 if (con->out_kvec_is_msg) {
2019                         con->out_skip = con->out_kvec_bytes;
2020                         con->out_kvec_is_msg = false;
2021                 }
2022         } else {
2023                 dout("con_revoke %p msg %p - not queued (sent?)\n", con, msg);
2024         }
2025         mutex_unlock(&con->mutex);
2026 }
2027
2028 /*
2029  * Revoke a message that we may be reading data into
2030  */
2031 void ceph_con_revoke_message(struct ceph_connection *con, struct ceph_msg *msg)
2032 {
2033         mutex_lock(&con->mutex);
2034         if (con->in_msg && con->in_msg == msg) {
2035                 unsigned front_len = le32_to_cpu(con->in_hdr.front_len);
2036                 unsigned middle_len = le32_to_cpu(con->in_hdr.middle_len);
2037                 unsigned data_len = le32_to_cpu(con->in_hdr.data_len);
2038
2039                 /* skip rest of message */
2040                 dout("con_revoke_pages %p msg %p revoked\n", con, msg);
2041                         con->in_base_pos = con->in_base_pos -
2042                                 sizeof(struct ceph_msg_header) -
2043                                 front_len -
2044                                 middle_len -
2045                                 data_len -
2046                                 sizeof(struct ceph_msg_footer);
2047                 ceph_msg_put(con->in_msg);
2048                 con->in_msg = NULL;
2049                 con->in_tag = CEPH_MSGR_TAG_READY;
2050                 con->in_seq++;
2051         } else {
2052                 dout("con_revoke_pages %p msg %p pages %p no-op\n",
2053                      con, con->in_msg, msg);
2054         }
2055         mutex_unlock(&con->mutex);
2056 }
2057
2058 /*
2059  * Queue a keepalive byte to ensure the tcp connection is alive.
2060  */
2061 void ceph_con_keepalive(struct ceph_connection *con)
2062 {
2063         if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
2064             test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2065                 queue_con(con);
2066 }
2067
2068
2069 /*
2070  * construct a new message with given type, size
2071  * the new msg has a ref count of 1.
2072  */
2073 struct ceph_msg *ceph_msg_new(int type, int front_len, gfp_t flags)
2074 {
2075         struct ceph_msg *m;
2076
2077         m = kmalloc(sizeof(*m), flags);
2078         if (m == NULL)
2079                 goto out;
2080         kref_init(&m->kref);
2081         INIT_LIST_HEAD(&m->list_head);
2082
2083         m->hdr.tid = 0;
2084         m->hdr.type = cpu_to_le16(type);
2085         m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
2086         m->hdr.version = 0;
2087         m->hdr.front_len = cpu_to_le32(front_len);
2088         m->hdr.middle_len = 0;
2089         m->hdr.data_len = 0;
2090         m->hdr.data_off = 0;
2091         m->hdr.reserved = 0;
2092         m->footer.front_crc = 0;
2093         m->footer.middle_crc = 0;
2094         m->footer.data_crc = 0;
2095         m->footer.flags = 0;
2096         m->front_max = front_len;
2097         m->front_is_vmalloc = false;
2098         m->more_to_follow = false;
2099         m->pool = NULL;
2100
2101         /* front */
2102         if (front_len) {
2103                 if (front_len > PAGE_CACHE_SIZE) {
2104                         m->front.iov_base = __vmalloc(front_len, flags,
2105                                                       PAGE_KERNEL);
2106                         m->front_is_vmalloc = true;
2107                 } else {
2108                         m->front.iov_base = kmalloc(front_len, flags);
2109                 }
2110                 if (m->front.iov_base == NULL) {
2111                         pr_err("msg_new can't allocate %d bytes\n",
2112                              front_len);
2113                         goto out2;
2114                 }
2115         } else {
2116                 m->front.iov_base = NULL;
2117         }
2118         m->front.iov_len = front_len;
2119
2120         /* middle */
2121         m->middle = NULL;
2122
2123         /* data */
2124         m->nr_pages = 0;
2125         m->pages = NULL;
2126         m->pagelist = NULL;
2127
2128         dout("ceph_msg_new %p front %d\n", m, front_len);
2129         return m;
2130
2131 out2:
2132         ceph_msg_put(m);
2133 out:
2134         pr_err("msg_new can't create type %d front %d\n", type, front_len);
2135         return NULL;
2136 }
2137
2138 /*
2139  * Allocate "middle" portion of a message, if it is needed and wasn't
2140  * allocated by alloc_msg.  This allows us to read a small fixed-size
2141  * per-type header in the front and then gracefully fail (i.e.,
2142  * propagate the error to the caller based on info in the front) when
2143  * the middle is too large.
2144  */
2145 static int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2146 {
2147         int type = le16_to_cpu(msg->hdr.type);
2148         int middle_len = le32_to_cpu(msg->hdr.middle_len);
2149
2150         dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2151              ceph_msg_type_name(type), middle_len);
2152         BUG_ON(!middle_len);
2153         BUG_ON(msg->middle);
2154
2155         msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2156         if (!msg->middle)
2157                 return -ENOMEM;
2158         return 0;
2159 }
2160
2161 /*
2162  * Generic message allocator, for incoming messages.
2163  */
2164 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2165                                 struct ceph_msg_header *hdr,
2166                                 int *skip)
2167 {
2168         int type = le16_to_cpu(hdr->type);
2169         int front_len = le32_to_cpu(hdr->front_len);
2170         int middle_len = le32_to_cpu(hdr->middle_len);
2171         struct ceph_msg *msg = NULL;
2172         int ret;
2173
2174         if (con->ops->alloc_msg) {
2175                 mutex_unlock(&con->mutex);
2176                 msg = con->ops->alloc_msg(con, hdr, skip);
2177                 mutex_lock(&con->mutex);
2178                 if (!msg || *skip)
2179                         return NULL;
2180         }
2181         if (!msg) {
2182                 *skip = 0;
2183                 msg = ceph_msg_new(type, front_len, GFP_NOFS);
2184                 if (!msg) {
2185                         pr_err("unable to allocate msg type %d len %d\n",
2186                                type, front_len);
2187                         return NULL;
2188                 }
2189         }
2190         memcpy(&msg->hdr, &con->in_hdr, sizeof(con->in_hdr));
2191
2192         if (middle_len && !msg->middle) {
2193                 ret = ceph_alloc_middle(con, msg);
2194                 if (ret < 0) {
2195                         ceph_msg_put(msg);
2196                         return NULL;
2197                 }
2198         }
2199
2200         return msg;
2201 }
2202
2203
2204 /*
2205  * Free a generically kmalloc'd message.
2206  */
2207 void ceph_msg_kfree(struct ceph_msg *m)
2208 {
2209         dout("msg_kfree %p\n", m);
2210         if (m->front_is_vmalloc)
2211                 vfree(m->front.iov_base);
2212         else
2213                 kfree(m->front.iov_base);
2214         kfree(m);
2215 }
2216
2217 /*
2218  * Drop a msg ref.  Destroy as needed.
2219  */
2220 void ceph_msg_last_put(struct kref *kref)
2221 {
2222         struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2223
2224         dout("ceph_msg_put last one on %p\n", m);
2225         WARN_ON(!list_empty(&m->list_head));
2226
2227         /* drop middle, data, if any */
2228         if (m->middle) {
2229                 ceph_buffer_put(m->middle);
2230                 m->middle = NULL;
2231         }
2232         m->nr_pages = 0;
2233         m->pages = NULL;
2234
2235         if (m->pagelist) {
2236                 ceph_pagelist_release(m->pagelist);
2237                 kfree(m->pagelist);
2238                 m->pagelist = NULL;
2239         }
2240
2241         if (m->pool)
2242                 ceph_msgpool_put(m->pool, m);
2243         else
2244                 ceph_msg_kfree(m);
2245 }
2246
2247 void ceph_msg_dump(struct ceph_msg *msg)
2248 {
2249         pr_debug("msg_dump %p (front_max %d nr_pages %d)\n", msg,
2250                  msg->front_max, msg->nr_pages);
2251         print_hex_dump(KERN_DEBUG, "header: ",
2252                        DUMP_PREFIX_OFFSET, 16, 1,
2253                        &msg->hdr, sizeof(msg->hdr), true);
2254         print_hex_dump(KERN_DEBUG, " front: ",
2255                        DUMP_PREFIX_OFFSET, 16, 1,
2256                        msg->front.iov_base, msg->front.iov_len, true);
2257         if (msg->middle)
2258                 print_hex_dump(KERN_DEBUG, "middle: ",
2259                                DUMP_PREFIX_OFFSET, 16, 1,
2260                                msg->middle->vec.iov_base,
2261                                msg->middle->vec.iov_len, true);
2262         print_hex_dump(KERN_DEBUG, "footer: ",
2263                        DUMP_PREFIX_OFFSET, 16, 1,
2264                        &msg->footer, sizeof(msg->footer), true);
2265 }