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