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/socket.h>
  10 #include <linux/string.h>
  11 #include <net/tcp.h>
  12
  13 #include "super.h"
  14 #include "messenger.h"
  15 #include "decode.h"
  16
  17 /*
  18  * Ceph uses the messenger to exchange ceph_msg messages with other
  19  * hosts in the system.  The messenger provides ordered and reliable
  20  * delivery.  We tolerate TCP disconnects by reconnecting (with
  21  * exponential backoff) in the case of a fault (disconnection, bad
  22  * crc, protocol error).  Acks allow sent messages to be discarded by
  23  * the sender.
  24  */
  25
  26 /* static tag bytes (protocol control messages) */
  27 static char tag_msg = CEPH_MSGR_TAG_MSG;
  28 static char tag_ack = CEPH_MSGR_TAG_ACK;
  29 static char tag_keepalive = CEPH_MSGR_TAG_KEEPALIVE;
  30
  31
  32 static void queue_con(struct ceph_connection *con);
  33 static void con_work(struct work_struct *);
  34 static void ceph_fault(struct ceph_connection *con);
  35
  36 const char *ceph_name_type_str(int t)
  37 {
  38         switch (t) {
  39         case CEPH_ENTITY_TYPE_MON: return "mon";
  40         case CEPH_ENTITY_TYPE_MDS: return "mds";
  41         case CEPH_ENTITY_TYPE_OSD: return "osd";
  42         case CEPH_ENTITY_TYPE_CLIENT: return "client";
  43         case CEPH_ENTITY_TYPE_ADMIN: return "admin";
  44         default: return "???";
  45         }
  46 }
  47
  48 /*
  49  * nicely render a sockaddr as a string.
  50  */
  51 #define MAX_ADDR_STR 20
  52 static char addr_str[MAX_ADDR_STR][40];
  53 static DEFINE_SPINLOCK(addr_str_lock);
  54 static int last_addr_str;
  55
  56 const char *pr_addr(const struct sockaddr_storage *ss)
  57 {
  58         int i;
  59         char *s;
  60         struct sockaddr_in *in4 = (void *)ss;
  61         unsigned char *quad = (void *)&in4->sin_addr.s_addr;
  62         struct sockaddr_in6 *in6 = (void *)ss;
  63
  64         spin_lock(&addr_str_lock);
  65         i = last_addr_str++;
  66         if (last_addr_str == MAX_ADDR_STR)
  67                 last_addr_str = 0;
  68         spin_unlock(&addr_str_lock);
  69         s = addr_str[i];
  70
  71         switch (ss->ss_family) {
  72         case AF_INET:
  73                 sprintf(s, "%u.%u.%u.%u:%u",
  74                         (unsigned int)quad[0],
  75                         (unsigned int)quad[1],
  76                         (unsigned int)quad[2],
  77                         (unsigned int)quad[3],
  78                         (unsigned int)ntohs(in4->sin_port));
  79                 break;
  80
  81         case AF_INET6:
  82                 sprintf(s, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x:%u",
  83                         in6->sin6_addr.s6_addr16[0],
  84                         in6->sin6_addr.s6_addr16[1],
  85                         in6->sin6_addr.s6_addr16[2],
  86                         in6->sin6_addr.s6_addr16[3],
  87                         in6->sin6_addr.s6_addr16[4],
  88                         in6->sin6_addr.s6_addr16[5],
  89                         in6->sin6_addr.s6_addr16[6],
  90                         in6->sin6_addr.s6_addr16[7],
  91                         (unsigned int)ntohs(in6->sin6_port));
  92                 break;
  93
  94         default:
  95                 sprintf(s, "(unknown sockaddr family %d)", (int)ss->ss_family);
  96         }
  97
  98         return s;
  99 }
 100
 101 static void encode_my_addr(struct ceph_messenger *msgr)
 102 {
 103         memcpy(&msgr->my_enc_addr, &msgr->inst.addr, sizeof(msgr->my_enc_addr));
 104         ceph_encode_addr(&msgr->my_enc_addr);
 105 }
 106
 107 /*
 108  * work queue for all reading and writing to/from the socket.
 109  */
 110 struct workqueue_struct *ceph_msgr_wq;
 111
 112 int __init ceph_msgr_init(void)
 113 {
 114         ceph_msgr_wq = create_workqueue("ceph-msgr");
 115         if (IS_ERR(ceph_msgr_wq)) {
 116                 int ret = PTR_ERR(ceph_msgr_wq);
 117                 pr_err("msgr_init failed to create workqueue: %d\n", ret);
 118                 ceph_msgr_wq = NULL;
 119                 return ret;
 120         }
 121         return 0;
 122 }
 123
 124 void ceph_msgr_exit(void)
 125 {
 126         destroy_workqueue(ceph_msgr_wq);
 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         set_sock_callbacks(sock, con);
 230
 231         dout("connect %s\n", pr_addr(&con->peer_addr.in_addr));
 232
 233         ret = sock->ops->connect(sock, paddr, sizeof(*paddr), O_NONBLOCK);
 234         if (ret == -EINPROGRESS) {
 235                 dout("connect %s EINPROGRESS sk_state = %u\n",
 236                      pr_addr(&con->peer_addr.in_addr),
 237                      sock->sk->sk_state);
 238                 ret = 0;
 239         }
 240         if (ret < 0) {
 241                 pr_err("connect %s error %d\n",
 242                        pr_addr(&con->peer_addr.in_addr), ret);
 243                 sock_release(sock);
 244                 con->sock = NULL;
 245                 con->error_msg = "connect error";
 246         }
 247
 248         if (ret < 0)
 249                 return ERR_PTR(ret);
 250         return sock;
 251 }
 252
 253 static int ceph_tcp_recvmsg(struct socket *sock, void *buf, size_t len)
 254 {
 255         struct kvec iov = {buf, len};
 256         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
 257
 258         return kernel_recvmsg(sock, &msg, &iov, 1, len, msg.msg_flags);
 259 }
 260
 261 /*
 262  * write something.  @more is true if caller will be sending more data
 263  * shortly.
 264  */
 265 static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov,
 266                      size_t kvlen, size_t len, int more)
 267 {
 268         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
 269
 270         if (more)
 271                 msg.msg_flags |= MSG_MORE;
 272         else
 273                 msg.msg_flags |= MSG_EOR;  /* superfluous, but what the hell */
 274
 275         return kernel_sendmsg(sock, &msg, iov, kvlen, len);
 276 }
 277
 278
 279 /*
 280  * Shutdown/close the socket for the given connection.
 281  */
 282 static int con_close_socket(struct ceph_connection *con)
 283 {
 284         int rc;
 285
 286         dout("con_close_socket on %p sock %p\n", con, con->sock);
 287         if (!con->sock)
 288                 return 0;
 289         set_bit(SOCK_CLOSED, &con->state);
 290         rc = con->sock->ops->shutdown(con->sock, SHUT_RDWR);
 291         sock_release(con->sock);
 292         con->sock = NULL;
 293         clear_bit(SOCK_CLOSED, &con->state);
 294         return rc;
 295 }
 296
 297 /*
 298  * Reset a connection.  Discard all incoming and outgoing messages
 299  * and clear *_seq state.
 300  */
 301 static void ceph_msg_remove(struct ceph_msg *msg)
 302 {
 303         list_del_init(&msg->list_head);
 304         ceph_msg_put(msg);
 305 }
 306 static void ceph_msg_remove_list(struct list_head *head)
 307 {
 308         while (!list_empty(head)) {
 309                 struct ceph_msg *msg = list_first_entry(head, struct ceph_msg,
 310                                                         list_head);
 311                 ceph_msg_remove(msg);
 312         }
 313 }
 314
 315 static void reset_connection(struct ceph_connection *con)
 316 {
 317         /* reset connection, out_queue, msg_ and connect_seq */
 318         /* discard existing out_queue and msg_seq */
 319         mutex_lock(&con->out_mutex);
 320         ceph_msg_remove_list(&con->out_queue);
 321         ceph_msg_remove_list(&con->out_sent);
 322
 323         con->connect_seq = 0;
 324         con->out_seq = 0;
 325         if (con->out_msg) {
 326                 ceph_msg_put(con->out_msg);
 327                 con->out_msg = NULL;
 328         }
 329         con->in_seq = 0;
 330         mutex_unlock(&con->out_mutex);
 331 }
 332
 333 /*
 334  * mark a peer down.  drop any open connections.
 335  */
 336 void ceph_con_close(struct ceph_connection *con)
 337 {
 338         dout("con_close %p peer %s\n", con, pr_addr(&con->peer_addr.in_addr));
 339         set_bit(CLOSED, &con->state);  /* in case there's queued work */
 340         clear_bit(STANDBY, &con->state);  /* avoid connect_seq bump */
 341         reset_connection(con);
 342         queue_con(con);
 343 }
 344
 345 /*
 346  * Reopen a closed connection, with a new peer address.
 347  */
 348 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
 349 {
 350         dout("con_open %p %s\n", con, pr_addr(&addr->in_addr));
 351         set_bit(OPENING, &con->state);
 352         clear_bit(CLOSED, &con->state);
 353         memcpy(&con->peer_addr, addr, sizeof(*addr));
 354         con->delay = 0;      /* reset backoff memory */
 355         queue_con(con);
 356 }
 357
 358 /*
 359  * generic get/put
 360  */
 361 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
 362 {
 363         dout("con_get %p nref = %d -> %d\n", con,
 364              atomic_read(&con->nref), atomic_read(&con->nref) + 1);
 365         if (atomic_inc_not_zero(&con->nref))
 366                 return con;
 367         return NULL;
 368 }
 369
 370 void ceph_con_put(struct ceph_connection *con)
 371 {
 372         dout("con_put %p nref = %d -> %d\n", con,
 373              atomic_read(&con->nref), atomic_read(&con->nref) - 1);
 374         BUG_ON(atomic_read(&con->nref) == 0);
 375         if (atomic_dec_and_test(&con->nref)) {
 376                 BUG_ON(con->sock);
 377                 kfree(con);
 378         }
 379 }
 380
 381 /*
 382  * initialize a new connection.
 383  */
 384 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
 385 {
 386         dout("con_init %p\n", con);
 387         memset(con, 0, sizeof(*con));
 388         atomic_set(&con->nref, 1);
 389         con->msgr = msgr;
 390         mutex_init(&con->out_mutex);
 391         INIT_LIST_HEAD(&con->out_queue);
 392         INIT_LIST_HEAD(&con->out_sent);
 393         INIT_DELAYED_WORK(&con->work, con_work);
 394 }
 395
 396
 397 /*
 398  * We maintain a global counter to order connection attempts.  Get
 399  * a unique seq greater than @gt.
 400  */
 401 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
 402 {
 403         u32 ret;
 404
 405         spin_lock(&msgr->global_seq_lock);
 406         if (msgr->global_seq < gt)
 407                 msgr->global_seq = gt;
 408         ret = ++msgr->global_seq;
 409         spin_unlock(&msgr->global_seq_lock);
 410         return ret;
 411 }
 412
 413
 414 /*
 415  * Prepare footer for currently outgoing message, and finish things
 416  * off.  Assumes out_kvec* are already valid.. we just add on to the end.
 417  */
 418 static void prepare_write_message_footer(struct ceph_connection *con, int v)
 419 {
 420         struct ceph_msg *m = con->out_msg;
 421
 422         dout("prepare_write_message_footer %p\n", con);
 423         con->out_kvec_is_msg = true;
 424         con->out_kvec[v].iov_base = &m->footer;
 425         con->out_kvec[v].iov_len = sizeof(m->footer);
 426         con->out_kvec_bytes += sizeof(m->footer);
 427         con->out_kvec_left++;
 428         con->out_more = m->more_to_follow;
 429         con->out_msg_done = true;
 430 }
 431
 432 /*
 433  * Prepare headers for the next outgoing message.
 434  */
 435 static void prepare_write_message(struct ceph_connection *con)
 436 {
 437         struct ceph_msg *m;
 438         int v = 0;
 439
 440         con->out_kvec_bytes = 0;
 441         con->out_kvec_is_msg = true;
 442         con->out_msg_done = false;
 443
 444         /* Sneak an ack in there first?  If we can get it into the same
 445          * TCP packet that's a good thing. */
 446         if (con->in_seq > con->in_seq_acked) {
 447                 con->in_seq_acked = con->in_seq;
 448                 con->out_kvec[v].iov_base = &tag_ack;
 449                 con->out_kvec[v++].iov_len = 1;
 450                 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 451                 con->out_kvec[v].iov_base = &con->out_temp_ack;
 452                 con->out_kvec[v++].iov_len = sizeof(con->out_temp_ack);
 453                 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 454         }
 455
 456         m = list_first_entry(&con->out_queue,
 457                        struct ceph_msg, list_head);
 458         con->out_msg = m;
 459         if (test_bit(LOSSYTX, &con->state)) {
 460                 /* put message on sent list */
 461                 ceph_msg_get(m);
 462                 list_move_tail(&m->list_head, &con->out_sent);
 463         } else {
 464                 list_del_init(&m->list_head);
 465         }
 466
 467         m->hdr.seq = cpu_to_le64(++con->out_seq);
 468
 469         dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
 470              m, con->out_seq, le16_to_cpu(m->hdr.type),
 471              le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
 472              le32_to_cpu(m->hdr.data_len),
 473              m->nr_pages);
 474         BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
 475
 476         /* tag + hdr + front + middle */
 477         con->out_kvec[v].iov_base = &tag_msg;
 478         con->out_kvec[v++].iov_len = 1;
 479         con->out_kvec[v].iov_base = &m->hdr;
 480         con->out_kvec[v++].iov_len = sizeof(m->hdr);
 481         con->out_kvec[v++] = m->front;
 482         if (m->middle)
 483                 con->out_kvec[v++] = m->middle->vec;
 484         con->out_kvec_left = v;
 485         con->out_kvec_bytes += 1 + sizeof(m->hdr) + m->front.iov_len +
 486                 (m->middle ? m->middle->vec.iov_len : 0);
 487         con->out_kvec_cur = con->out_kvec;
 488
 489         /* fill in crc (except data pages), footer */
 490         con->out_msg->hdr.crc =
 491                 cpu_to_le32(crc32c(0, (void *)&m->hdr,
 492                                       sizeof(m->hdr) - sizeof(m->hdr.crc)));
 493         con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
 494         con->out_msg->footer.front_crc =
 495                 cpu_to_le32(crc32c(0, m->front.iov_base, m->front.iov_len));
 496         if (m->middle)
 497                 con->out_msg->footer.middle_crc =
 498                         cpu_to_le32(crc32c(0, m->middle->vec.iov_base,
 499                                            m->middle->vec.iov_len));
 500         else
 501                 con->out_msg->footer.middle_crc = 0;
 502         con->out_msg->footer.data_crc = 0;
 503         dout("prepare_write_message front_crc %u data_crc %u\n",
 504              le32_to_cpu(con->out_msg->footer.front_crc),
 505              le32_to_cpu(con->out_msg->footer.middle_crc));
 506
 507         /* is there a data payload? */
 508         if (le32_to_cpu(m->hdr.data_len) > 0) {
 509                 /* initialize page iterator */
 510                 con->out_msg_pos.page = 0;
 511                 con->out_msg_pos.page_pos =
 512                         le16_to_cpu(m->hdr.data_off) & ~PAGE_MASK;
 513                 con->out_msg_pos.data_pos = 0;
 514                 con->out_msg_pos.did_page_crc = 0;
 515                 con->out_more = 1;  /* data + footer will follow */
 516         } else {
 517                 /* no, queue up footer too and be done */
 518                 prepare_write_message_footer(con, v);
 519         }
 520
 521         set_bit(WRITE_PENDING, &con->state);
 522 }
 523
 524 /*
 525  * Prepare an ack.
 526  */
 527 static void prepare_write_ack(struct ceph_connection *con)
 528 {
 529         dout("prepare_write_ack %p %llu -> %llu\n", con,
 530              con->in_seq_acked, con->in_seq);
 531         con->in_seq_acked = con->in_seq;
 532
 533         con->out_kvec[0].iov_base = &tag_ack;
 534         con->out_kvec[0].iov_len = 1;
 535         con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 536         con->out_kvec[1].iov_base = &con->out_temp_ack;
 537         con->out_kvec[1].iov_len = sizeof(con->out_temp_ack);
 538         con->out_kvec_left = 2;
 539         con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 540         con->out_kvec_cur = con->out_kvec;
 541         con->out_more = 1;  /* more will follow.. eventually.. */
 542         set_bit(WRITE_PENDING, &con->state);
 543 }
 544
 545 /*
 546  * Prepare to write keepalive byte.
 547  */
 548 static void prepare_write_keepalive(struct ceph_connection *con)
 549 {
 550         dout("prepare_write_keepalive %p\n", con);
 551         con->out_kvec[0].iov_base = &tag_keepalive;
 552         con->out_kvec[0].iov_len = 1;
 553         con->out_kvec_left = 1;
 554         con->out_kvec_bytes = 1;
 555         con->out_kvec_cur = con->out_kvec;
 556         set_bit(WRITE_PENDING, &con->state);
 557 }
 558
 559 /*
 560  * Connection negotiation.
 561  */
 562
 563 static void prepare_connect_authorizer(struct ceph_connection *con)
 564 {
 565         void *auth_buf;
 566         int auth_len = 0;
 567         int auth_protocol = 0;
 568
 569         if (con->ops->get_authorizer)
 570                 con->ops->get_authorizer(con, &auth_buf, &auth_len,
 571                                          &auth_protocol, &con->auth_reply_buf,
 572                                          &con->auth_reply_buf_len,
 573                                          con->auth_retry);
 574
 575         con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
 576         con->out_connect.authorizer_len = cpu_to_le32(auth_len);
 577
 578         con->out_kvec[con->out_kvec_left].iov_base = auth_buf;
 579         con->out_kvec[con->out_kvec_left].iov_len = auth_len;
 580         con->out_kvec_left++;
 581         con->out_kvec_bytes += auth_len;
 582 }
 583
 584 /*
 585  * We connected to a peer and are saying hello.
 586  */
 587 static void prepare_write_banner(struct ceph_messenger *msgr,
 588                                  struct ceph_connection *con)
 589 {
 590         int len = strlen(CEPH_BANNER);
 591
 592         con->out_kvec[0].iov_base = CEPH_BANNER;
 593         con->out_kvec[0].iov_len = len;
 594         con->out_kvec[1].iov_base = &msgr->my_enc_addr;
 595         con->out_kvec[1].iov_len = sizeof(msgr->my_enc_addr);
 596         con->out_kvec_left = 2;
 597         con->out_kvec_bytes = len + sizeof(msgr->my_enc_addr);
 598         con->out_kvec_cur = con->out_kvec;
 599         con->out_more = 0;
 600         set_bit(WRITE_PENDING, &con->state);
 601 }
 602
 603 static void prepare_write_connect(struct ceph_messenger *msgr,
 604                                   struct ceph_connection *con,
 605                                   int after_banner)
 606 {
 607         unsigned global_seq = get_global_seq(con->msgr, 0);
 608         int proto;
 609
 610         switch (con->peer_name.type) {
 611         case CEPH_ENTITY_TYPE_MON:
 612                 proto = CEPH_MONC_PROTOCOL;
 613                 break;
 614         case CEPH_ENTITY_TYPE_OSD:
 615                 proto = CEPH_OSDC_PROTOCOL;
 616                 break;
 617         case CEPH_ENTITY_TYPE_MDS:
 618                 proto = CEPH_MDSC_PROTOCOL;
 619                 break;
 620         default:
 621                 BUG();
 622         }
 623
 624         dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
 625              con->connect_seq, global_seq, proto);
 626
 627         con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
 628         con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
 629         con->out_connect.global_seq = cpu_to_le32(global_seq);
 630         con->out_connect.protocol_version = cpu_to_le32(proto);
 631         con->out_connect.flags = 0;
 632
 633         if (!after_banner) {
 634                 con->out_kvec_left = 0;
 635                 con->out_kvec_bytes = 0;
 636         }
 637         con->out_kvec[con->out_kvec_left].iov_base = &con->out_connect;
 638         con->out_kvec[con->out_kvec_left].iov_len = sizeof(con->out_connect);
 639         con->out_kvec_left++;
 640         con->out_kvec_bytes += sizeof(con->out_connect);
 641         con->out_kvec_cur = con->out_kvec;
 642         con->out_more = 0;
 643         set_bit(WRITE_PENDING, &con->state);
 644
 645         prepare_connect_authorizer(con);
 646 }
 647
 648
 649 /*
 650  * write as much of pending kvecs to the socket as we can.
 651  *  1 -> done
 652  *  0 -> socket full, but more to do
 653  * <0 -> error
 654  */
 655 static int write_partial_kvec(struct ceph_connection *con)
 656 {
 657         int ret;
 658
 659         dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
 660         while (con->out_kvec_bytes > 0) {
 661                 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
 662                                        con->out_kvec_left, con->out_kvec_bytes,
 663                                        con->out_more);
 664                 if (ret <= 0)
 665                         goto out;
 666                 con->out_kvec_bytes -= ret;
 667                 if (con->out_kvec_bytes == 0)
 668                         break;            /* done */
 669                 while (ret > 0) {
 670                         if (ret >= con->out_kvec_cur->iov_len) {
 671                                 ret -= con->out_kvec_cur->iov_len;
 672                                 con->out_kvec_cur++;
 673                                 con->out_kvec_left--;
 674                         } else {
 675                                 con->out_kvec_cur->iov_len -= ret;
 676                                 con->out_kvec_cur->iov_base += ret;
 677                                 ret = 0;
 678                                 break;
 679                         }
 680                 }
 681         }
 682         con->out_kvec_left = 0;
 683         con->out_kvec_is_msg = false;
 684         ret = 1;
 685 out:
 686         dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
 687              con->out_kvec_bytes, con->out_kvec_left, ret);
 688         return ret;  /* done! */
 689 }
 690
 691 /*
 692  * Write as much message data payload as we can.  If we finish, queue
 693  * up the footer.
 694  *  1 -> done, footer is now queued in out_kvec[].
 695  *  0 -> socket full, but more to do
 696  * <0 -> error
 697  */
 698 static int write_partial_msg_pages(struct ceph_connection *con)
 699 {
 700         struct ceph_msg *msg = con->out_msg;
 701         unsigned data_len = le32_to_cpu(msg->hdr.data_len);
 702         size_t len;
 703         int crc = con->msgr->nocrc;
 704         int ret;
 705
 706         dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
 707              con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
 708              con->out_msg_pos.page_pos);
 709
 710         while (con->out_msg_pos.page < con->out_msg->nr_pages) {
 711                 struct page *page = NULL;
 712                 void *kaddr = NULL;
 713
 714                 /*
 715                  * if we are calculating the data crc (the default), we need
 716                  * to map the page.  if our pages[] has been revoked, use the
 717                  * zero page.
 718                  */
 719                 if (msg->pages) {
 720                         page = msg->pages[con->out_msg_pos.page];
 721                         if (crc)
 722                                 kaddr = kmap(page);
 723                 } else {
 724                         page = con->msgr->zero_page;
 725                         if (crc)
 726                                 kaddr = page_address(con->msgr->zero_page);
 727                 }
 728                 len = min((int)(PAGE_SIZE - con->out_msg_pos.page_pos),
 729                           (int)(data_len - con->out_msg_pos.data_pos));
 730                 if (crc && !con->out_msg_pos.did_page_crc) {
 731                         void *base = kaddr + con->out_msg_pos.page_pos;
 732                         u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
 733
 734                         BUG_ON(kaddr == NULL);
 735                         con->out_msg->footer.data_crc =
 736                                 cpu_to_le32(crc32c(tmpcrc, base, len));
 737                         con->out_msg_pos.did_page_crc = 1;
 738                 }
 739
 740                 ret = kernel_sendpage(con->sock, page,
 741                                       con->out_msg_pos.page_pos, len,
 742                                       MSG_DONTWAIT | MSG_NOSIGNAL |
 743                                       MSG_MORE);
 744
 745                 if (crc && msg->pages)
 746                         kunmap(page);
 747
 748                 if (ret <= 0)
 749                         goto out;
 750
 751                 con->out_msg_pos.data_pos += ret;
 752                 con->out_msg_pos.page_pos += ret;
 753                 if (ret == len) {
 754                         con->out_msg_pos.page_pos = 0;
 755                         con->out_msg_pos.page++;
 756                         con->out_msg_pos.did_page_crc = 0;
 757                 }
 758         }
 759
 760         dout("write_partial_msg_pages %p msg %p done\n", con, msg);
 761
 762         /* prepare and queue up footer, too */
 763         if (!crc)
 764                 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
 765         con->out_kvec_bytes = 0;
 766         con->out_kvec_left = 0;
 767         con->out_kvec_cur = con->out_kvec;
 768         prepare_write_message_footer(con, 0);
 769         ret = 1;
 770 out:
 771         return ret;
 772 }
 773
 774 /*
 775  * write some zeros
 776  */
 777 static int write_partial_skip(struct ceph_connection *con)
 778 {
 779         int ret;
 780
 781         while (con->out_skip > 0) {
 782                 struct kvec iov = {
 783                         .iov_base = page_address(con->msgr->zero_page),
 784                         .iov_len = min(con->out_skip, (int)PAGE_CACHE_SIZE)
 785                 };
 786
 787                 ret = ceph_tcp_sendmsg(con->sock, &iov, 1, iov.iov_len, 1);
 788                 if (ret <= 0)
 789                         goto out;
 790                 con->out_skip -= ret;
 791         }
 792         ret = 1;
 793 out:
 794         return ret;
 795 }
 796
 797 /*
 798  * Prepare to read connection handshake, or an ack.
 799  */
 800 static void prepare_read_banner(struct ceph_connection *con)
 801 {
 802         dout("prepare_read_banner %p\n", con);
 803         con->in_base_pos = 0;
 804 }
 805
 806 static void prepare_read_connect(struct ceph_connection *con)
 807 {
 808         dout("prepare_read_connect %p\n", con);
 809         con->in_base_pos = 0;
 810 }
 811
 812 static void prepare_read_connect_retry(struct ceph_connection *con)
 813 {
 814         dout("prepare_read_connect_retry %p\n", con);
 815         con->in_base_pos = strlen(CEPH_BANNER) + sizeof(con->actual_peer_addr)
 816                 + sizeof(con->peer_addr_for_me);
 817 }
 818
 819 static void prepare_read_ack(struct ceph_connection *con)
 820 {
 821         dout("prepare_read_ack %p\n", con);
 822         con->in_base_pos = 0;
 823 }
 824
 825 static void prepare_read_tag(struct ceph_connection *con)
 826 {
 827         dout("prepare_read_tag %p\n", con);
 828         con->in_base_pos = 0;
 829         con->in_tag = CEPH_MSGR_TAG_READY;
 830 }
 831
 832 /*
 833  * Prepare to read a message.
 834  */
 835 static int prepare_read_message(struct ceph_connection *con)
 836 {
 837         dout("prepare_read_message %p\n", con);
 838         BUG_ON(con->in_msg != NULL);
 839         con->in_base_pos = 0;
 840         con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
 841         return 0;
 842 }
 843
 844
 845 static int read_partial(struct ceph_connection *con,
 846                         int *to, int size, void *object)
 847 {
 848         *to += size;
 849         while (con->in_base_pos < *to) {
 850                 int left = *to - con->in_base_pos;
 851                 int have = size - left;
 852                 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
 853                 if (ret <= 0)
 854                         return ret;
 855                 con->in_base_pos += ret;
 856         }
 857         return 1;
 858 }
 859
 860
 861 /*
 862  * Read all or part of the connect-side handshake on a new connection
 863  */
 864 static int read_partial_banner(struct ceph_connection *con)
 865 {
 866         int ret, to = 0;
 867
 868         dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
 869
 870         /* peer's banner */
 871         ret = read_partial(con, &to, strlen(CEPH_BANNER), con->in_banner);
 872         if (ret <= 0)
 873                 goto out;
 874         ret = read_partial(con, &to, sizeof(con->actual_peer_addr),
 875                            &con->actual_peer_addr);
 876         if (ret <= 0)
 877                 goto out;
 878         ret = read_partial(con, &to, sizeof(con->peer_addr_for_me),
 879                            &con->peer_addr_for_me);
 880         if (ret <= 0)
 881                 goto out;
 882 out:
 883         return ret;
 884 }
 885
 886 static int read_partial_connect(struct ceph_connection *con)
 887 {
 888         int ret, to = 0;
 889
 890         dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
 891
 892         ret = read_partial(con, &to, sizeof(con->in_reply), &con->in_reply);
 893         if (ret <= 0)
 894                 goto out;
 895         ret = read_partial(con, &to, le32_to_cpu(con->in_reply.authorizer_len),
 896                            con->auth_reply_buf);
 897         if (ret <= 0)
 898                 goto out;
 899
 900         dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
 901              con, (int)con->in_reply.tag,
 902              le32_to_cpu(con->in_reply.connect_seq),
 903              le32_to_cpu(con->in_reply.global_seq));
 904 out:
 905         return ret;
 906
 907 }
 908
 909 /*
 910  * Verify the hello banner looks okay.
 911  */
 912 static int verify_hello(struct ceph_connection *con)
 913 {
 914         if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
 915                 pr_err("connect to %s got bad banner\n",
 916                        pr_addr(&con->peer_addr.in_addr));
 917                 con->error_msg = "protocol error, bad banner";
 918                 return -1;
 919         }
 920         return 0;
 921 }
 922
 923 static bool addr_is_blank(struct sockaddr_storage *ss)
 924 {
 925         switch (ss->ss_family) {
 926         case AF_INET:
 927                 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
 928         case AF_INET6:
 929                 return
 930                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
 931                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
 932                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
 933                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
 934         }
 935         return false;
 936 }
 937
 938 static int addr_port(struct sockaddr_storage *ss)
 939 {
 940         switch (ss->ss_family) {
 941         case AF_INET:
 942                 return ntohs(((struct sockaddr_in *)ss)->sin_port);
 943         case AF_INET6:
 944                 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
 945         }
 946         return 0;
 947 }
 948
 949 static void addr_set_port(struct sockaddr_storage *ss, int p)
 950 {
 951         switch (ss->ss_family) {
 952         case AF_INET:
 953                 ((struct sockaddr_in *)ss)->sin_port = htons(p);
 954         case AF_INET6:
 955                 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
 956         }
 957 }
 958
 959 /*
 960  * Parse an ip[:port] list into an addr array.  Use the default
 961  * monitor port if a port isn't specified.
 962  */
 963 int ceph_parse_ips(const char *c, const char *end,
 964                    struct ceph_entity_addr *addr,
 965                    int max_count, int *count)
 966 {
 967         int i;
 968         const char *p = c;
 969
 970         dout("parse_ips on '%.*s'\n", (int)(end-c), c);
 971         for (i = 0; i < max_count; i++) {
 972                 const char *ipend;
 973                 struct sockaddr_storage *ss = &addr[i].in_addr;
 974                 struct sockaddr_in *in4 = (void *)ss;
 975                 struct sockaddr_in6 *in6 = (void *)ss;
 976                 int port;
 977
 978                 memset(ss, 0, sizeof(*ss));
 979                 if (in4_pton(p, end - p, (u8 *)&in4->sin_addr.s_addr,
 980                              ',', &ipend)) {
 981                         ss->ss_family = AF_INET;
 982                 } else if (in6_pton(p, end - p, (u8 *)&in6->sin6_addr.s6_addr,
 983                                     ',', &ipend)) {
 984                         ss->ss_family = AF_INET6;
 985                 } else {
 986                         goto bad;
 987                 }
 988                 p = ipend;
 989
 990                 /* port? */
 991                 if (p < end && *p == ':') {
 992                         port = 0;
 993                         p++;
 994                         while (p < end && *p >= '0' && *p <= '9') {
 995                                 port = (port * 10) + (*p - '0');
 996                                 p++;
 997                         }
 998                         if (port > 65535 || port == 0)
 999                                 goto bad;
1000                 } else {
1001                         port = CEPH_MON_PORT;
1002                 }
1003
1004                 addr_set_port(ss, port);
1005
1006                 dout("parse_ips got %s\n", pr_addr(ss));
1007
1008                 if (p == end)
1009                         break;
1010                 if (*p != ',')
1011                         goto bad;
1012                 p++;
1013         }
1014
1015         if (p != end)
1016                 goto bad;
1017
1018         if (count)
1019                 *count = i + 1;
1020         return 0;
1021
1022 bad:
1023         pr_err("parse_ips bad ip '%s'\n", c);
1024         return -EINVAL;
1025 }
1026
1027 static int process_banner(struct ceph_connection *con)
1028 {
1029         dout("process_banner on %p\n", con);
1030
1031         if (verify_hello(con) < 0)
1032                 return -1;
1033
1034         ceph_decode_addr(&con->actual_peer_addr);
1035         ceph_decode_addr(&con->peer_addr_for_me);
1036
1037         /*
1038          * Make sure the other end is who we wanted.  note that the other
1039          * end may not yet know their ip address, so if it's 0.0.0.0, give
1040          * them the benefit of the doubt.
1041          */
1042         if (!ceph_entity_addr_is_local(&con->peer_addr,
1043                                        &con->actual_peer_addr) &&
1044             !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1045               con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1046                 pr_err("wrong peer, want %s/%d, "
1047                        "got %s/%d, wtf\n",
1048                        pr_addr(&con->peer_addr.in_addr),
1049                        con->peer_addr.nonce,
1050                        pr_addr(&con->actual_peer_addr.in_addr),
1051                        con->actual_peer_addr.nonce);
1052                 con->error_msg = "protocol error, wrong peer";
1053                 return -1;
1054         }
1055
1056         /*
1057          * did we learn our address?
1058          */
1059         if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1060                 int port = addr_port(&con->msgr->inst.addr.in_addr);
1061
1062                 memcpy(&con->msgr->inst.addr.in_addr,
1063                        &con->peer_addr_for_me.in_addr,
1064                        sizeof(con->peer_addr_for_me.in_addr));
1065                 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1066                 encode_my_addr(con->msgr);
1067                 dout("process_banner learned my addr is %s\n",
1068                      pr_addr(&con->msgr->inst.addr.in_addr));
1069         }
1070
1071         set_bit(NEGOTIATING, &con->state);
1072         prepare_read_connect(con);
1073         return 0;
1074 }
1075
1076 static int process_connect(struct ceph_connection *con)
1077 {
1078         dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1079
1080         switch (con->in_reply.tag) {
1081         case CEPH_MSGR_TAG_BADPROTOVER:
1082                 dout("process_connect got BADPROTOVER my %d != their %d\n",
1083                      le32_to_cpu(con->out_connect.protocol_version),
1084                      le32_to_cpu(con->in_reply.protocol_version));
1085                 pr_err("%s%lld %s protocol version mismatch,"
1086                        " my %d != server's %d\n",
1087                        ENTITY_NAME(con->peer_name),
1088                        pr_addr(&con->peer_addr.in_addr),
1089                        le32_to_cpu(con->out_connect.protocol_version),
1090                        le32_to_cpu(con->in_reply.protocol_version));
1091                 con->error_msg = "protocol version mismatch";
1092                 if (con->ops->bad_proto)
1093                         con->ops->bad_proto(con);
1094                 reset_connection(con);
1095                 set_bit(CLOSED, &con->state);  /* in case there's queued work */
1096                 return -1;
1097
1098         case CEPH_MSGR_TAG_BADAUTHORIZER:
1099                 con->auth_retry++;
1100                 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1101                      con->auth_retry);
1102                 if (con->auth_retry == 2) {
1103                         con->error_msg = "connect authorization failure";
1104                         reset_connection(con);
1105                         set_bit(CLOSED, &con->state);
1106                         return -1;
1107                 }
1108                 con->auth_retry = 1;
1109                 prepare_write_connect(con->msgr, con, 0);
1110                 prepare_read_connect_retry(con);
1111                 break;
1112
1113         case CEPH_MSGR_TAG_RESETSESSION:
1114                 /*
1115                  * If we connected with a large connect_seq but the peer
1116                  * has no record of a session with us (no connection, or
1117                  * connect_seq == 0), they will send RESETSESION to indicate
1118                  * that they must have reset their session, and may have
1119                  * dropped messages.
1120                  */
1121                 dout("process_connect got RESET peer seq %u\n",
1122                      le32_to_cpu(con->in_connect.connect_seq));
1123                 pr_err("%s%lld %s connection reset\n",
1124                        ENTITY_NAME(con->peer_name),
1125                        pr_addr(&con->peer_addr.in_addr));
1126                 reset_connection(con);
1127                 prepare_write_connect(con->msgr, con, 0);
1128                 prepare_read_connect(con);
1129
1130                 /* Tell ceph about it. */
1131                 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1132                 if (con->ops->peer_reset)
1133                         con->ops->peer_reset(con);
1134                 break;
1135
1136         case CEPH_MSGR_TAG_RETRY_SESSION:
1137                 /*
1138                  * If we sent a smaller connect_seq than the peer has, try
1139                  * again with a larger value.
1140                  */
1141                 dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
1142                      le32_to_cpu(con->out_connect.connect_seq),
1143                      le32_to_cpu(con->in_connect.connect_seq));
1144                 con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
1145                 prepare_write_connect(con->msgr, con, 0);
1146                 prepare_read_connect(con);
1147                 break;
1148
1149         case CEPH_MSGR_TAG_RETRY_GLOBAL:
1150                 /*
1151                  * If we sent a smaller global_seq than the peer has, try
1152                  * again with a larger value.
1153                  */
1154                 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1155                      con->peer_global_seq,
1156                      le32_to_cpu(con->in_connect.global_seq));
1157                 get_global_seq(con->msgr,
1158                                le32_to_cpu(con->in_connect.global_seq));
1159                 prepare_write_connect(con->msgr, con, 0);
1160                 prepare_read_connect(con);
1161                 break;
1162
1163         case CEPH_MSGR_TAG_READY:
1164                 clear_bit(CONNECTING, &con->state);
1165                 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1166                 con->connect_seq++;
1167                 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1168                      con->peer_global_seq,
1169                      le32_to_cpu(con->in_reply.connect_seq),
1170                      con->connect_seq);
1171                 WARN_ON(con->connect_seq !=
1172                         le32_to_cpu(con->in_reply.connect_seq));
1173
1174                 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1175                         set_bit(LOSSYTX, &con->state);
1176
1177                 prepare_read_tag(con);
1178                 break;
1179
1180         case CEPH_MSGR_TAG_WAIT:
1181                 /*
1182                  * If there is a connection race (we are opening
1183                  * connections to each other), one of us may just have
1184                  * to WAIT.  This shouldn't happen if we are the
1185                  * client.
1186                  */
1187                 pr_err("process_connect peer connecting WAIT\n");
1188
1189         default:
1190                 pr_err("connect protocol error, will retry\n");
1191                 con->error_msg = "protocol error, garbage tag during connect";
1192                 return -1;
1193         }
1194         return 0;
1195 }
1196
1197
1198 /*
1199  * read (part of) an ack
1200  */
1201 static int read_partial_ack(struct ceph_connection *con)
1202 {
1203         int to = 0;
1204
1205         return read_partial(con, &to, sizeof(con->in_temp_ack),
1206                             &con->in_temp_ack);
1207 }
1208
1209
1210 /*
1211  * We can finally discard anything that's been acked.
1212  */
1213 static void process_ack(struct ceph_connection *con)
1214 {
1215         struct ceph_msg *m;
1216         u64 ack = le64_to_cpu(con->in_temp_ack);
1217         u64 seq;
1218
1219         mutex_lock(&con->out_mutex);
1220         while (!list_empty(&con->out_sent)) {
1221                 m = list_first_entry(&con->out_sent, struct ceph_msg,
1222                                      list_head);
1223                 seq = le64_to_cpu(m->hdr.seq);
1224                 if (seq > ack)
1225                         break;
1226                 dout("got ack for seq %llu type %d at %p\n", seq,
1227                      le16_to_cpu(m->hdr.type), m);
1228                 ceph_msg_remove(m);
1229         }
1230         mutex_unlock(&con->out_mutex);
1231         prepare_read_tag(con);
1232 }
1233
1234
1235
1236
1237
1238
1239 /*
1240  * read (part of) a message.
1241  */
1242 static int read_partial_message(struct ceph_connection *con)
1243 {
1244         struct ceph_msg *m = con->in_msg;
1245         void *p;
1246         int ret;
1247         int to, want, left;
1248         unsigned front_len, middle_len, data_len, data_off;
1249         int datacrc = con->msgr->nocrc;
1250
1251         dout("read_partial_message con %p msg %p\n", con, m);
1252
1253         /* header */
1254         while (con->in_base_pos < sizeof(con->in_hdr)) {
1255                 left = sizeof(con->in_hdr) - con->in_base_pos;
1256                 ret = ceph_tcp_recvmsg(con->sock,
1257                                        (char *)&con->in_hdr + con->in_base_pos,
1258                                        left);
1259                 if (ret <= 0)
1260                         return ret;
1261                 con->in_base_pos += ret;
1262                 if (con->in_base_pos == sizeof(con->in_hdr)) {
1263                         u32 crc = crc32c(0, (void *)&con->in_hdr,
1264                                  sizeof(con->in_hdr) - sizeof(con->in_hdr.crc));
1265                         if (crc != le32_to_cpu(con->in_hdr.crc)) {
1266                                 pr_err("read_partial_message bad hdr "
1267                                        " crc %u != expected %u\n",
1268                                        crc, con->in_hdr.crc);
1269                                 return -EBADMSG;
1270                         }
1271                 }
1272         }
1273
1274         front_len = le32_to_cpu(con->in_hdr.front_len);
1275         if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1276                 return -EIO;
1277         middle_len = le32_to_cpu(con->in_hdr.middle_len);
1278         if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1279                 return -EIO;
1280         data_len = le32_to_cpu(con->in_hdr.data_len);
1281         if (data_len > CEPH_MSG_MAX_DATA_LEN)
1282                 return -EIO;
1283
1284         /* allocate message? */
1285         if (!con->in_msg) {
1286                 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1287                      con->in_hdr.front_len, con->in_hdr.data_len);
1288                 con->in_msg = con->ops->alloc_msg(con, &con->in_hdr);
1289                 if (!con->in_msg) {
1290                         /* skip this message */
1291                         dout("alloc_msg returned NULL, skipping message\n");
1292                         con->in_base_pos = -front_len - middle_len - data_len -
1293                                 sizeof(m->footer);
1294                         con->in_tag = CEPH_MSGR_TAG_READY;
1295                         return 0;
1296                 }
1297                 if (IS_ERR(con->in_msg)) {
1298                         ret = PTR_ERR(con->in_msg);
1299                         con->in_msg = NULL;
1300                         con->error_msg = "out of memory for incoming message";
1301                         return ret;
1302                 }
1303                 m = con->in_msg;
1304                 m->front.iov_len = 0;    /* haven't read it yet */
1305                 memcpy(&m->hdr, &con->in_hdr, sizeof(con->in_hdr));
1306         }
1307
1308         /* front */
1309         while (m->front.iov_len < front_len) {
1310                 BUG_ON(m->front.iov_base == NULL);
1311                 left = front_len - m->front.iov_len;
1312                 ret = ceph_tcp_recvmsg(con->sock, (char *)m->front.iov_base +
1313                                        m->front.iov_len, left);
1314                 if (ret <= 0)
1315                         return ret;
1316                 m->front.iov_len += ret;
1317                 if (m->front.iov_len == front_len)
1318                         con->in_front_crc = crc32c(0, m->front.iov_base,
1319                                                       m->front.iov_len);
1320         }
1321
1322         /* middle */
1323         while (middle_len > 0 && (!m->middle ||
1324                                   m->middle->vec.iov_len < middle_len)) {
1325                 if (m->middle == NULL) {
1326                         ret = -EOPNOTSUPP;
1327                         if (con->ops->alloc_middle)
1328                                 ret = con->ops->alloc_middle(con, m);
1329                         if (ret < 0) {
1330                                 dout("alloc_middle failed, skipping payload\n");
1331                                 con->in_base_pos = -middle_len - data_len
1332                                         - sizeof(m->footer);
1333                                 ceph_msg_put(con->in_msg);
1334                                 con->in_msg = NULL;
1335                                 con->in_tag = CEPH_MSGR_TAG_READY;
1336                                 return 0;
1337                         }
1338                         m->middle->vec.iov_len = 0;
1339                 }
1340                 left = middle_len - m->middle->vec.iov_len;
1341                 ret = ceph_tcp_recvmsg(con->sock,
1342                                        (char *)m->middle->vec.iov_base +
1343                                        m->middle->vec.iov_len, left);
1344                 if (ret <= 0)
1345                         return ret;
1346                 m->middle->vec.iov_len += ret;
1347                 if (m->middle->vec.iov_len == middle_len)
1348                         con->in_middle_crc = crc32c(0, m->middle->vec.iov_base,
1349                                                       m->middle->vec.iov_len);
1350         }
1351
1352         /* (page) data */
1353         data_off = le16_to_cpu(m->hdr.data_off);
1354         if (data_len == 0)
1355                 goto no_data;
1356
1357         if (m->nr_pages == 0) {
1358                 con->in_msg_pos.page = 0;
1359                 con->in_msg_pos.page_pos = data_off & ~PAGE_MASK;
1360                 con->in_msg_pos.data_pos = 0;
1361                 /* find pages for data payload */
1362                 want = calc_pages_for(data_off & ~PAGE_MASK, data_len);
1363                 ret = -1;
1364                 if (con->ops->prepare_pages)
1365                         ret = con->ops->prepare_pages(con, m, want);
1366                 if (ret < 0) {
1367                         dout("%p prepare_pages failed, skipping payload\n", m);
1368                         con->in_base_pos = -data_len - sizeof(m->footer);
1369                         ceph_msg_put(con->in_msg);
1370                         con->in_msg = NULL;
1371                         con->in_tag = CEPH_MSGR_TAG_READY;
1372                         return 0;
1373                 }
1374                 BUG_ON(m->nr_pages < want);
1375         }
1376         while (con->in_msg_pos.data_pos < data_len) {
1377                 left = min((int)(data_len - con->in_msg_pos.data_pos),
1378                            (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1379                 BUG_ON(m->pages == NULL);
1380                 p = kmap(m->pages[con->in_msg_pos.page]);
1381                 ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1382                                        left);
1383                 if (ret > 0 && datacrc)
1384                         con->in_data_crc =
1385                                 crc32c(con->in_data_crc,
1386                                           p + con->in_msg_pos.page_pos, ret);
1387                 kunmap(m->pages[con->in_msg_pos.page]);
1388                 if (ret <= 0)
1389                         return ret;
1390                 con->in_msg_pos.data_pos += ret;
1391                 con->in_msg_pos.page_pos += ret;
1392                 if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1393                         con->in_msg_pos.page_pos = 0;
1394                         con->in_msg_pos.page++;
1395                 }
1396         }
1397
1398 no_data:
1399         /* footer */
1400         to = sizeof(m->hdr) + sizeof(m->footer);
1401         while (con->in_base_pos < to) {
1402                 left = to - con->in_base_pos;
1403                 ret = ceph_tcp_recvmsg(con->sock, (char *)&m->footer +
1404                                        (con->in_base_pos - sizeof(m->hdr)),
1405                                        left);
1406                 if (ret <= 0)
1407                         return ret;
1408                 con->in_base_pos += ret;
1409         }
1410         dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1411              m, front_len, m->footer.front_crc, middle_len,
1412              m->footer.middle_crc, data_len, m->footer.data_crc);
1413
1414         /* crc ok? */
1415         if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1416                 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1417                        m, con->in_front_crc, m->footer.front_crc);
1418                 return -EBADMSG;
1419         }
1420         if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1421                 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1422                        m, con->in_middle_crc, m->footer.middle_crc);
1423                 return -EBADMSG;
1424         }
1425         if (datacrc &&
1426             (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1427             con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1428                 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1429                        con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1430                 return -EBADMSG;
1431         }
1432
1433         return 1; /* done! */
1434 }
1435
1436 /*
1437  * Process message.  This happens in the worker thread.  The callback should
1438  * be careful not to do anything that waits on other incoming messages or it
1439  * may deadlock.
1440  */
1441 static void process_message(struct ceph_connection *con)
1442 {
1443         struct ceph_msg *msg;
1444
1445         msg = con->in_msg;
1446         con->in_msg = NULL;
1447
1448         /* if first message, set peer_name */
1449         if (con->peer_name.type == 0)
1450                 con->peer_name = msg->hdr.src.name;
1451
1452         mutex_lock(&con->out_mutex);
1453         con->in_seq++;
1454         mutex_unlock(&con->out_mutex);
1455
1456         dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1457              msg, le64_to_cpu(msg->hdr.seq),
1458              ENTITY_NAME(msg->hdr.src.name),
1459              le16_to_cpu(msg->hdr.type),
1460              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1461              le32_to_cpu(msg->hdr.front_len),
1462              le32_to_cpu(msg->hdr.data_len),
1463              con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1464         con->ops->dispatch(con, msg);
1465         prepare_read_tag(con);
1466 }
1467
1468
1469 /*
1470  * Write something to the socket.  Called in a worker thread when the
1471  * socket appears to be writeable and we have something ready to send.
1472  */
1473 static int try_write(struct ceph_connection *con)
1474 {
1475         struct ceph_messenger *msgr = con->msgr;
1476         int ret = 1;
1477
1478         dout("try_write start %p state %lu nref %d\n", con, con->state,
1479              atomic_read(&con->nref));
1480
1481         mutex_lock(&con->out_mutex);
1482 more:
1483         dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1484
1485         /* open the socket first? */
1486         if (con->sock == NULL) {
1487                 /*
1488                  * if we were STANDBY and are reconnecting _this_
1489                  * connection, bump connect_seq now.  Always bump
1490                  * global_seq.
1491                  */
1492                 if (test_and_clear_bit(STANDBY, &con->state))
1493                         con->connect_seq++;
1494
1495                 prepare_write_banner(msgr, con);
1496                 prepare_write_connect(msgr, con, 1);
1497                 prepare_read_banner(con);
1498                 set_bit(CONNECTING, &con->state);
1499                 clear_bit(NEGOTIATING, &con->state);
1500
1501                 con->in_tag = CEPH_MSGR_TAG_READY;
1502                 dout("try_write initiating connect on %p new state %lu\n",
1503                      con, con->state);
1504                 con->sock = ceph_tcp_connect(con);
1505                 if (IS_ERR(con->sock)) {
1506                         con->sock = NULL;
1507                         con->error_msg = "connect error";
1508                         ret = -1;
1509                         goto out;
1510                 }
1511         }
1512
1513 more_kvec:
1514         /* kvec data queued? */
1515         if (con->out_skip) {
1516                 ret = write_partial_skip(con);
1517                 if (ret <= 0)
1518                         goto done;
1519                 if (ret < 0) {
1520                         dout("try_write write_partial_skip err %d\n", ret);
1521                         goto done;
1522                 }
1523         }
1524         if (con->out_kvec_left) {
1525                 ret = write_partial_kvec(con);
1526                 if (ret <= 0)
1527                         goto done;
1528                 if (ret < 0) {
1529                         dout("try_write write_partial_kvec err %d\n", ret);
1530                         goto done;
1531                 }
1532         }
1533
1534         /* msg pages? */
1535         if (con->out_msg) {
1536                 if (con->out_msg_done) {
1537                         ceph_msg_put(con->out_msg);
1538                         con->out_msg = NULL;   /* we're done with this one */
1539                         goto do_next;
1540                 }
1541
1542                 ret = write_partial_msg_pages(con);
1543                 if (ret == 1)
1544                         goto more_kvec;  /* we need to send the footer, too! */
1545                 if (ret == 0)
1546                         goto done;
1547                 if (ret < 0) {
1548                         dout("try_write write_partial_msg_pages err %d\n",
1549                              ret);
1550                         goto done;
1551                 }
1552         }
1553
1554 do_next:
1555         if (!test_bit(CONNECTING, &con->state)) {
1556                 /* is anything else pending? */
1557                 if (!list_empty(&con->out_queue)) {
1558                         prepare_write_message(con);
1559                         goto more;
1560                 }
1561                 if (con->in_seq > con->in_seq_acked) {
1562                         prepare_write_ack(con);
1563                         goto more;
1564                 }
1565                 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1566                         prepare_write_keepalive(con);
1567                         goto more;
1568                 }
1569         }
1570
1571         /* Nothing to do! */
1572         clear_bit(WRITE_PENDING, &con->state);
1573         dout("try_write nothing else to write.\n");
1574 done:
1575         ret = 0;
1576 out:
1577         mutex_unlock(&con->out_mutex);
1578         dout("try_write done on %p\n", con);
1579         return ret;
1580 }
1581
1582
1583
1584 /*
1585  * Read what we can from the socket.
1586  */
1587 static int try_read(struct ceph_connection *con)
1588 {
1589         struct ceph_messenger *msgr;
1590         int ret = -1;
1591
1592         if (!con->sock)
1593                 return 0;
1594
1595         if (test_bit(STANDBY, &con->state))
1596                 return 0;
1597
1598         dout("try_read start on %p\n", con);
1599         msgr = con->msgr;
1600
1601 more:
1602         dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1603              con->in_base_pos);
1604         if (test_bit(CONNECTING, &con->state)) {
1605                 if (!test_bit(NEGOTIATING, &con->state)) {
1606                         dout("try_read connecting\n");
1607                         ret = read_partial_banner(con);
1608                         if (ret <= 0)
1609                                 goto done;
1610                         if (process_banner(con) < 0) {
1611                                 ret = -1;
1612                                 goto out;
1613                         }
1614                 }
1615                 ret = read_partial_connect(con);
1616                 if (ret <= 0)
1617                         goto done;
1618                 if (process_connect(con) < 0) {
1619                         ret = -1;
1620                         goto out;
1621                 }
1622                 goto more;
1623         }
1624
1625         if (con->in_base_pos < 0) {
1626                 /*
1627                  * skipping + discarding content.
1628                  *
1629                  * FIXME: there must be a better way to do this!
1630                  */
1631                 static char buf[1024];
1632                 int skip = min(1024, -con->in_base_pos);
1633                 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
1634                 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
1635                 if (ret <= 0)
1636                         goto done;
1637                 con->in_base_pos += ret;
1638                 if (con->in_base_pos)
1639                         goto more;
1640         }
1641         if (con->in_tag == CEPH_MSGR_TAG_READY) {
1642                 /*
1643                  * what's next?
1644                  */
1645                 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
1646                 if (ret <= 0)
1647                         goto done;
1648                 dout("try_read got tag %d\n", (int)con->in_tag);
1649                 switch (con->in_tag) {
1650                 case CEPH_MSGR_TAG_MSG:
1651                         prepare_read_message(con);
1652                         break;
1653                 case CEPH_MSGR_TAG_ACK:
1654                         prepare_read_ack(con);
1655                         break;
1656                 case CEPH_MSGR_TAG_CLOSE:
1657                         set_bit(CLOSED, &con->state);   /* fixme */
1658                         goto done;
1659                 default:
1660                         goto bad_tag;
1661                 }
1662         }
1663         if (con->in_tag == CEPH_MSGR_TAG_MSG) {
1664                 ret = read_partial_message(con);
1665                 if (ret <= 0) {
1666                         switch (ret) {
1667                         case -EBADMSG:
1668                                 con->error_msg = "bad crc";
1669                                 ret = -EIO;
1670                                 goto out;
1671                         case -EIO:
1672                                 con->error_msg = "io error";
1673                                 goto out;
1674                         default:
1675                                 goto done;
1676                         }
1677                 }
1678                 if (con->in_tag == CEPH_MSGR_TAG_READY)
1679                         goto more;
1680                 process_message(con);
1681                 goto more;
1682         }
1683         if (con->in_tag == CEPH_MSGR_TAG_ACK) {
1684                 ret = read_partial_ack(con);
1685                 if (ret <= 0)
1686                         goto done;
1687                 process_ack(con);
1688                 goto more;
1689         }
1690
1691 done:
1692         ret = 0;
1693 out:
1694         dout("try_read done on %p\n", con);
1695         return ret;
1696
1697 bad_tag:
1698         pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
1699         con->error_msg = "protocol error, garbage tag";
1700         ret = -1;
1701         goto out;
1702 }
1703
1704
1705 /*
1706  * Atomically queue work on a connection.  Bump @con reference to
1707  * avoid races with connection teardown.
1708  *
1709  * There is some trickery going on with QUEUED and BUSY because we
1710  * only want a _single_ thread operating on each connection at any
1711  * point in time, but we want to use all available CPUs.
1712  *
1713  * The worker thread only proceeds if it can atomically set BUSY.  It
1714  * clears QUEUED and does it's thing.  When it thinks it's done, it
1715  * clears BUSY, then rechecks QUEUED.. if it's set again, it loops
1716  * (tries again to set BUSY).
1717  *
1718  * To queue work, we first set QUEUED, _then_ if BUSY isn't set, we
1719  * try to queue work.  If that fails (work is already queued, or BUSY)
1720  * we give up (work also already being done or is queued) but leave QUEUED
1721  * set so that the worker thread will loop if necessary.
1722  */
1723 static void queue_con(struct ceph_connection *con)
1724 {
1725         if (test_bit(DEAD, &con->state)) {
1726                 dout("queue_con %p ignoring: DEAD\n",
1727                      con);
1728                 return;
1729         }
1730
1731         if (!con->ops->get(con)) {
1732                 dout("queue_con %p ref count 0\n", con);
1733                 return;
1734         }
1735
1736         set_bit(QUEUED, &con->state);
1737         if (test_bit(BUSY, &con->state)) {
1738                 dout("queue_con %p - already BUSY\n", con);
1739                 con->ops->put(con);
1740         } else if (!queue_work(ceph_msgr_wq, &con->work.work)) {
1741                 dout("queue_con %p - already queued\n", con);
1742                 con->ops->put(con);
1743         } else {
1744                 dout("queue_con %p\n", con);
1745         }
1746 }
1747
1748 /*
1749  * Do some work on a connection.  Drop a connection ref when we're done.
1750  */
1751 static void con_work(struct work_struct *work)
1752 {
1753         struct ceph_connection *con = container_of(work, struct ceph_connection,
1754                                                    work.work);
1755         int backoff = 0;
1756
1757 more:
1758         if (test_and_set_bit(BUSY, &con->state) != 0) {
1759                 dout("con_work %p BUSY already set\n", con);
1760                 goto out;
1761         }
1762         dout("con_work %p start, clearing QUEUED\n", con);
1763         clear_bit(QUEUED, &con->state);
1764
1765         if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
1766                 dout("con_work CLOSED\n");
1767                 con_close_socket(con);
1768                 goto done;
1769         }
1770         if (test_and_clear_bit(OPENING, &con->state)) {
1771                 /* reopen w/ new peer */
1772                 dout("con_work OPENING\n");
1773                 con_close_socket(con);
1774         }
1775
1776         if (test_and_clear_bit(SOCK_CLOSED, &con->state) ||
1777             try_read(con) < 0 ||
1778             try_write(con) < 0) {
1779                 backoff = 1;
1780                 ceph_fault(con);     /* error/fault path */
1781         }
1782
1783 done:
1784         clear_bit(BUSY, &con->state);
1785         dout("con->state=%lu\n", con->state);
1786         if (test_bit(QUEUED, &con->state)) {
1787                 if (!backoff) {
1788                         dout("con_work %p QUEUED reset, looping\n", con);
1789                         goto more;
1790                 }
1791                 dout("con_work %p QUEUED reset, but just faulted\n", con);
1792                 clear_bit(QUEUED, &con->state);
1793         }
1794         dout("con_work %p done\n", con);
1795
1796 out:
1797         con->ops->put(con);
1798 }
1799
1800
1801 /*
1802  * Generic error/fault handler.  A retry mechanism is used with
1803  * exponential backoff
1804  */
1805 static void ceph_fault(struct ceph_connection *con)
1806 {
1807         pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
1808                pr_addr(&con->peer_addr.in_addr), con->error_msg);
1809         dout("fault %p state %lu to peer %s\n",
1810              con, con->state, pr_addr(&con->peer_addr.in_addr));
1811
1812         if (test_bit(LOSSYTX, &con->state)) {
1813                 dout("fault on LOSSYTX channel\n");
1814                 goto out;
1815         }
1816
1817         clear_bit(BUSY, &con->state);  /* to avoid an improbable race */
1818
1819         con_close_socket(con);
1820
1821         if (con->in_msg) {
1822                 ceph_msg_put(con->in_msg);
1823                 con->in_msg = NULL;
1824         }
1825
1826         /* If there are no messages in the queue, place the connection
1827          * in a STANDBY state (i.e., don't try to reconnect just yet). */
1828         mutex_lock(&con->out_mutex);
1829         if (list_empty(&con->out_queue) && !con->out_keepalive_pending) {
1830                 dout("fault setting STANDBY\n");
1831                 set_bit(STANDBY, &con->state);
1832                 mutex_unlock(&con->out_mutex);
1833                 goto out;
1834         }
1835
1836         /* Requeue anything that hasn't been acked, and retry after a
1837          * delay. */
1838         list_splice_init(&con->out_sent, &con->out_queue);
1839         mutex_unlock(&con->out_mutex);
1840
1841         if (con->delay == 0)
1842                 con->delay = BASE_DELAY_INTERVAL;
1843         else if (con->delay < MAX_DELAY_INTERVAL)
1844                 con->delay *= 2;
1845
1846         /* explicitly schedule work to try to reconnect again later. */
1847         dout("fault queueing %p delay %lu\n", con, con->delay);
1848         con->ops->get(con);
1849         if (queue_delayed_work(ceph_msgr_wq, &con->work,
1850                                round_jiffies_relative(con->delay)) == 0)
1851                 con->ops->put(con);
1852
1853 out:
1854         if (con->ops->fault)
1855                 con->ops->fault(con);
1856 }
1857
1858
1859
1860 /*
1861  * create a new messenger instance
1862  */
1863 struct ceph_messenger *ceph_messenger_create(struct ceph_entity_addr *myaddr)
1864 {
1865         struct ceph_messenger *msgr;
1866
1867         msgr = kzalloc(sizeof(*msgr), GFP_KERNEL);
1868         if (msgr == NULL)
1869                 return ERR_PTR(-ENOMEM);
1870
1871         spin_lock_init(&msgr->global_seq_lock);
1872
1873         /* the zero page is needed if a request is "canceled" while the message
1874          * is being written over the socket */
1875         msgr->zero_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1876         if (!msgr->zero_page) {
1877                 kfree(msgr);
1878                 return ERR_PTR(-ENOMEM);
1879         }
1880         kmap(msgr->zero_page);
1881
1882         if (myaddr)
1883                 msgr->inst.addr = *myaddr;
1884
1885         /* select a random nonce */
1886         get_random_bytes(&msgr->inst.addr.nonce,
1887                          sizeof(msgr->inst.addr.nonce));
1888         encode_my_addr(msgr);
1889
1890         dout("messenger_create %p\n", msgr);
1891         return msgr;
1892 }
1893
1894 void ceph_messenger_destroy(struct ceph_messenger *msgr)
1895 {
1896         dout("destroy %p\n", msgr);
1897         kunmap(msgr->zero_page);
1898         __free_page(msgr->zero_page);
1899         kfree(msgr);
1900         dout("destroyed messenger %p\n", msgr);
1901 }
1902
1903 /*
1904  * Queue up an outgoing message on the given connection.
1905  */
1906 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
1907 {
1908         if (test_bit(CLOSED, &con->state)) {
1909                 dout("con_send %p closed, dropping %p\n", con, msg);
1910                 ceph_msg_put(msg);
1911                 return;
1912         }
1913
1914         /* set src+dst */
1915         msg->hdr.src.name = con->msgr->inst.name;
1916         msg->hdr.src.addr = con->msgr->my_enc_addr;
1917         msg->hdr.orig_src = msg->hdr.src;
1918         msg->hdr.dst_erank = con->peer_addr.erank;
1919
1920         /* queue */
1921         mutex_lock(&con->out_mutex);
1922         BUG_ON(!list_empty(&msg->list_head));
1923         list_add_tail(&msg->list_head, &con->out_queue);
1924         dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
1925              ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
1926              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1927              le32_to_cpu(msg->hdr.front_len),
1928              le32_to_cpu(msg->hdr.middle_len),
1929              le32_to_cpu(msg->hdr.data_len));
1930         mutex_unlock(&con->out_mutex);
1931
1932         /* if there wasn't anything waiting to send before, queue
1933          * new work */
1934         if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
1935                 queue_con(con);
1936 }
1937
1938 /*
1939  * Revoke a message that was previously queued for send
1940  */
1941 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
1942 {
1943         mutex_lock(&con->out_mutex);
1944         if (!list_empty(&msg->list_head)) {
1945                 dout("con_revoke %p msg %p\n", con, msg);
1946                 list_del_init(&msg->list_head);
1947                 ceph_msg_put(msg);
1948                 msg->hdr.seq = 0;
1949                 if (con->out_msg == msg) {
1950                         ceph_msg_put(con->out_msg);
1951                         con->out_msg = NULL;
1952                 }
1953                 if (con->out_kvec_is_msg) {
1954                         con->out_skip = con->out_kvec_bytes;
1955                         con->out_kvec_is_msg = false;
1956                 }
1957         } else {
1958                 dout("con_revoke %p msg %p - not queued (sent?)\n", con, msg);
1959         }
1960         mutex_unlock(&con->out_mutex);
1961 }
1962
1963 /*
1964  * Queue a keepalive byte to ensure the tcp connection is alive.
1965  */
1966 void ceph_con_keepalive(struct ceph_connection *con)
1967 {
1968         if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
1969             test_and_set_bit(WRITE_PENDING, &con->state) == 0)
1970                 queue_con(con);
1971 }
1972
1973
1974 /*
1975  * construct a new message with given type, size
1976  * the new msg has a ref count of 1.
1977  */
1978 struct ceph_msg *ceph_msg_new(int type, int front_len,
1979                               int page_len, int page_off, struct page **pages)
1980 {
1981         struct ceph_msg *m;
1982
1983         m = kmalloc(sizeof(*m), GFP_NOFS);
1984         if (m == NULL)
1985                 goto out;
1986         kref_init(&m->kref);
1987         INIT_LIST_HEAD(&m->list_head);
1988
1989         m->hdr.type = cpu_to_le16(type);
1990         m->hdr.front_len = cpu_to_le32(front_len);
1991         m->hdr.middle_len = 0;
1992         m->hdr.data_len = cpu_to_le32(page_len);
1993         m->hdr.data_off = cpu_to_le16(page_off);
1994         m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
1995         m->footer.front_crc = 0;
1996         m->footer.middle_crc = 0;
1997         m->footer.data_crc = 0;
1998         m->front_max = front_len;
1999         m->front_is_vmalloc = false;
2000         m->more_to_follow = false;
2001         m->pool = NULL;
2002
2003         /* front */
2004         if (front_len) {
2005                 if (front_len > PAGE_CACHE_SIZE) {
2006                         m->front.iov_base = __vmalloc(front_len, GFP_NOFS,
2007                                                       PAGE_KERNEL);
2008                         m->front_is_vmalloc = true;
2009                 } else {
2010                         m->front.iov_base = kmalloc(front_len, GFP_NOFS);
2011                 }
2012                 if (m->front.iov_base == NULL) {
2013                         pr_err("msg_new can't allocate %d bytes\n",
2014                              front_len);
2015                         goto out2;
2016                 }
2017         } else {
2018                 m->front.iov_base = NULL;
2019         }
2020         m->front.iov_len = front_len;
2021
2022         /* middle */
2023         m->middle = NULL;
2024
2025         /* data */
2026         m->nr_pages = calc_pages_for(page_off, page_len);
2027         m->pages = pages;
2028
2029         dout("ceph_msg_new %p page %d~%d -> %d\n", m, page_off, page_len,
2030              m->nr_pages);
2031         return m;
2032
2033 out2:
2034         ceph_msg_put(m);
2035 out:
2036         pr_err("msg_new can't create type %d len %d\n", type, front_len);
2037         return ERR_PTR(-ENOMEM);
2038 }
2039
2040 /*
2041  * Generic message allocator, for incoming messages.
2042  */
2043 struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2044                                 struct ceph_msg_header *hdr)
2045 {
2046         int type = le16_to_cpu(hdr->type);
2047         int front_len = le32_to_cpu(hdr->front_len);
2048         struct ceph_msg *msg = ceph_msg_new(type, front_len, 0, 0, NULL);
2049
2050         if (!msg) {
2051                 pr_err("unable to allocate msg type %d len %d\n",
2052                        type, front_len);
2053                 return ERR_PTR(-ENOMEM);
2054         }
2055         return msg;
2056 }
2057
2058 /*
2059  * Allocate "middle" portion of a message, if it is needed and wasn't
2060  * allocated by alloc_msg.  This allows us to read a small fixed-size
2061  * per-type header in the front and then gracefully fail (i.e.,
2062  * propagate the error to the caller based on info in the front) when
2063  * the middle is too large.
2064  */
2065 int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2066 {
2067         int type = le16_to_cpu(msg->hdr.type);
2068         int middle_len = le32_to_cpu(msg->hdr.middle_len);
2069
2070         dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2071              ceph_msg_type_name(type), middle_len);
2072         BUG_ON(!middle_len);
2073         BUG_ON(msg->middle);
2074
2075         msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2076         if (!msg->middle)
2077                 return -ENOMEM;
2078         return 0;
2079 }
2080
2081
2082 /*
2083  * Free a generically kmalloc'd message.
2084  */
2085 void ceph_msg_kfree(struct ceph_msg *m)
2086 {
2087         dout("msg_kfree %p\n", m);
2088         if (m->front_is_vmalloc)
2089                 vfree(m->front.iov_base);
2090         else
2091                 kfree(m->front.iov_base);
2092         kfree(m);
2093 }
2094
2095 /*
2096  * Drop a msg ref.  Destroy as needed.
2097  */
2098 void ceph_msg_last_put(struct kref *kref)
2099 {
2100         struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2101
2102         dout("ceph_msg_put last one on %p\n", m);
2103         WARN_ON(!list_empty(&m->list_head));
2104
2105         /* drop middle, data, if any */
2106         if (m->middle) {
2107                 ceph_buffer_put(m->middle);
2108                 m->middle = NULL;
2109         }
2110         m->nr_pages = 0;
2111         m->pages = NULL;
2112
2113         if (m->pool)
2114                 ceph_msgpool_put(m->pool, m);
2115         else
2116                 ceph_msg_kfree(m);
2117 }
2118
2119 void ceph_msg_dump(struct ceph_msg *msg)
2120 {
2121         pr_debug("msg_dump %p (front_max %d nr_pages %d)\n", msg,
2122                  msg->front_max, msg->nr_pages);
2123         print_hex_dump(KERN_DEBUG, "header: ",
2124                        DUMP_PREFIX_OFFSET, 16, 1,
2125                        &msg->hdr, sizeof(msg->hdr), true);
2126         print_hex_dump(KERN_DEBUG, " front: ",
2127                        DUMP_PREFIX_OFFSET, 16, 1,
2128                        msg->front.iov_base, msg->front.iov_len, true);
2129         if (msg->middle)
2130                 print_hex_dump(KERN_DEBUG, "middle: ",
2131                                DUMP_PREFIX_OFFSET, 16, 1,
2132                                msg->middle->vec.iov_base,
2133                                msg->middle->vec.iov_len, true);
2134         print_hex_dump(KERN_DEBUG, "footer: ",
2135                        DUMP_PREFIX_OFFSET, 16, 1,
2136                        &msg->footer, sizeof(msg->footer), true);
2137 }