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 const char *ceph_name_type_str(int t)
  43 {
  44         switch (t) {
  45         case CEPH_ENTITY_TYPE_MON: return "mon";
  46         case CEPH_ENTITY_TYPE_MDS: return "mds";
  47         case CEPH_ENTITY_TYPE_OSD: return "osd";
  48         case CEPH_ENTITY_TYPE_CLIENT: return "client";
  49         case CEPH_ENTITY_TYPE_ADMIN: return "admin";
  50         default: return "???";
  51         }
  52 }
  53
  54 /*
  55  * nicely render a sockaddr as a string.
  56  */
  57 #define MAX_ADDR_STR 20
  58 static char addr_str[MAX_ADDR_STR][40];
  59 static DEFINE_SPINLOCK(addr_str_lock);
  60 static int last_addr_str;
  61
  62 const char *pr_addr(const struct sockaddr_storage *ss)
  63 {
  64         int i;
  65         char *s;
  66         struct sockaddr_in *in4 = (void *)ss;
  67         unsigned char *quad = (void *)&in4->sin_addr.s_addr;
  68         struct sockaddr_in6 *in6 = (void *)ss;
  69
  70         spin_lock(&addr_str_lock);
  71         i = last_addr_str++;
  72         if (last_addr_str == MAX_ADDR_STR)
  73                 last_addr_str = 0;
  74         spin_unlock(&addr_str_lock);
  75         s = addr_str[i];
  76
  77         switch (ss->ss_family) {
  78         case AF_INET:
  79                 sprintf(s, "%u.%u.%u.%u:%u",
  80                         (unsigned int)quad[0],
  81                         (unsigned int)quad[1],
  82                         (unsigned int)quad[2],
  83                         (unsigned int)quad[3],
  84                         (unsigned int)ntohs(in4->sin_port));
  85                 break;
  86
  87         case AF_INET6:
  88                 sprintf(s, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x:%u",
  89                         in6->sin6_addr.s6_addr16[0],
  90                         in6->sin6_addr.s6_addr16[1],
  91                         in6->sin6_addr.s6_addr16[2],
  92                         in6->sin6_addr.s6_addr16[3],
  93                         in6->sin6_addr.s6_addr16[4],
  94                         in6->sin6_addr.s6_addr16[5],
  95                         in6->sin6_addr.s6_addr16[6],
  96                         in6->sin6_addr.s6_addr16[7],
  97                         (unsigned int)ntohs(in6->sin6_port));
  98                 break;
  99
 100         default:
 101                 sprintf(s, "(unknown sockaddr family %d)", (int)ss->ss_family);
 102         }
 103
 104         return s;
 105 }
 106
 107 static void encode_my_addr(struct ceph_messenger *msgr)
 108 {
 109         memcpy(&msgr->my_enc_addr, &msgr->inst.addr, sizeof(msgr->my_enc_addr));
 110         ceph_encode_addr(&msgr->my_enc_addr);
 111 }
 112
 113 /*
 114  * work queue for all reading and writing to/from the socket.
 115  */
 116 struct workqueue_struct *ceph_msgr_wq;
 117
 118 int __init ceph_msgr_init(void)
 119 {
 120         ceph_msgr_wq = create_workqueue("ceph-msgr");
 121         if (IS_ERR(ceph_msgr_wq)) {
 122                 int ret = PTR_ERR(ceph_msgr_wq);
 123                 pr_err("msgr_init failed to create workqueue: %d\n", ret);
 124                 ceph_msgr_wq = NULL;
 125                 return ret;
 126         }
 127         return 0;
 128 }
 129
 130 void ceph_msgr_exit(void)
 131 {
 132         destroy_workqueue(ceph_msgr_wq);
 133 }
 134
 135 /*
 136  * socket callback functions
 137  */
 138
 139 /* data available on socket, or listen socket received a connect */
 140 static void ceph_data_ready(struct sock *sk, int count_unused)
 141 {
 142         struct ceph_connection *con =
 143                 (struct ceph_connection *)sk->sk_user_data;
 144         if (sk->sk_state != TCP_CLOSE_WAIT) {
 145                 dout("ceph_data_ready on %p state = %lu, queueing work\n",
 146                      con, con->state);
 147                 queue_con(con);
 148         }
 149 }
 150
 151 /* socket has buffer space for writing */
 152 static void ceph_write_space(struct sock *sk)
 153 {
 154         struct ceph_connection *con =
 155                 (struct ceph_connection *)sk->sk_user_data;
 156
 157         /* only queue to workqueue if there is data we want to write. */
 158         if (test_bit(WRITE_PENDING, &con->state)) {
 159                 dout("ceph_write_space %p queueing write work\n", con);
 160                 queue_con(con);
 161         } else {
 162                 dout("ceph_write_space %p nothing to write\n", con);
 163         }
 164
 165         /* since we have our own write_space, clear the SOCK_NOSPACE flag */
 166         clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 167 }
 168
 169 /* socket's state has changed */
 170 static void ceph_state_change(struct sock *sk)
 171 {
 172         struct ceph_connection *con =
 173                 (struct ceph_connection *)sk->sk_user_data;
 174
 175         dout("ceph_state_change %p state = %lu sk_state = %u\n",
 176              con, con->state, sk->sk_state);
 177
 178         if (test_bit(CLOSED, &con->state))
 179                 return;
 180
 181         switch (sk->sk_state) {
 182         case TCP_CLOSE:
 183                 dout("ceph_state_change TCP_CLOSE\n");
 184         case TCP_CLOSE_WAIT:
 185                 dout("ceph_state_change TCP_CLOSE_WAIT\n");
 186                 if (test_and_set_bit(SOCK_CLOSED, &con->state) == 0) {
 187                         if (test_bit(CONNECTING, &con->state))
 188                                 con->error_msg = "connection failed";
 189                         else
 190                                 con->error_msg = "socket closed";
 191                         queue_con(con);
 192                 }
 193                 break;
 194         case TCP_ESTABLISHED:
 195                 dout("ceph_state_change TCP_ESTABLISHED\n");
 196                 queue_con(con);
 197                 break;
 198         }
 199 }
 200
 201 /*
 202  * set up socket callbacks
 203  */
 204 static void set_sock_callbacks(struct socket *sock,
 205                                struct ceph_connection *con)
 206 {
 207         struct sock *sk = sock->sk;
 208         sk->sk_user_data = (void *)con;
 209         sk->sk_data_ready = ceph_data_ready;
 210         sk->sk_write_space = ceph_write_space;
 211         sk->sk_state_change = ceph_state_change;
 212 }
 213
 214
 215 /*
 216  * socket helpers
 217  */
 218
 219 /*
 220  * initiate connection to a remote socket.
 221  */
 222 static struct socket *ceph_tcp_connect(struct ceph_connection *con)
 223 {
 224         struct sockaddr *paddr = (struct sockaddr *)&con->peer_addr.in_addr;
 225         struct socket *sock;
 226         int ret;
 227
 228         BUG_ON(con->sock);
 229         ret = sock_create_kern(AF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
 230         if (ret)
 231                 return ERR_PTR(ret);
 232         con->sock = sock;
 233         sock->sk->sk_allocation = GFP_NOFS;
 234
 235 #ifdef CONFIG_LOCKDEP
 236         lockdep_set_class(&sock->sk->sk_lock, &socket_class);
 237 #endif
 238
 239         set_sock_callbacks(sock, con);
 240
 241         dout("connect %s\n", pr_addr(&con->peer_addr.in_addr));
 242
 243         ret = sock->ops->connect(sock, paddr, sizeof(*paddr), O_NONBLOCK);
 244         if (ret == -EINPROGRESS) {
 245                 dout("connect %s EINPROGRESS sk_state = %u\n",
 246                      pr_addr(&con->peer_addr.in_addr),
 247                      sock->sk->sk_state);
 248                 ret = 0;
 249         }
 250         if (ret < 0) {
 251                 pr_err("connect %s error %d\n",
 252                        pr_addr(&con->peer_addr.in_addr), ret);
 253                 sock_release(sock);
 254                 con->sock = NULL;
 255                 con->error_msg = "connect error";
 256         }
 257
 258         if (ret < 0)
 259                 return ERR_PTR(ret);
 260         return sock;
 261 }
 262
 263 static int ceph_tcp_recvmsg(struct socket *sock, void *buf, size_t len)
 264 {
 265         struct kvec iov = {buf, len};
 266         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
 267
 268         return kernel_recvmsg(sock, &msg, &iov, 1, len, msg.msg_flags);
 269 }
 270
 271 /*
 272  * write something.  @more is true if caller will be sending more data
 273  * shortly.
 274  */
 275 static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov,
 276                      size_t kvlen, size_t len, int more)
 277 {
 278         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
 279
 280         if (more)
 281                 msg.msg_flags |= MSG_MORE;
 282         else
 283                 msg.msg_flags |= MSG_EOR;  /* superfluous, but what the hell */
 284
 285         return kernel_sendmsg(sock, &msg, iov, kvlen, len);
 286 }
 287
 288
 289 /*
 290  * Shutdown/close the socket for the given connection.
 291  */
 292 static int con_close_socket(struct ceph_connection *con)
 293 {
 294         int rc;
 295
 296         dout("con_close_socket on %p sock %p\n", con, con->sock);
 297         if (!con->sock)
 298                 return 0;
 299         set_bit(SOCK_CLOSED, &con->state);
 300         rc = con->sock->ops->shutdown(con->sock, SHUT_RDWR);
 301         sock_release(con->sock);
 302         con->sock = NULL;
 303         clear_bit(SOCK_CLOSED, &con->state);
 304         return rc;
 305 }
 306
 307 /*
 308  * Reset a connection.  Discard all incoming and outgoing messages
 309  * and clear *_seq state.
 310  */
 311 static void ceph_msg_remove(struct ceph_msg *msg)
 312 {
 313         list_del_init(&msg->list_head);
 314         ceph_msg_put(msg);
 315 }
 316 static void ceph_msg_remove_list(struct list_head *head)
 317 {
 318         while (!list_empty(head)) {
 319                 struct ceph_msg *msg = list_first_entry(head, struct ceph_msg,
 320                                                         list_head);
 321                 ceph_msg_remove(msg);
 322         }
 323 }
 324
 325 static void reset_connection(struct ceph_connection *con)
 326 {
 327         /* reset connection, out_queue, msg_ and connect_seq */
 328         /* discard existing out_queue and msg_seq */
 329         ceph_msg_remove_list(&con->out_queue);
 330         ceph_msg_remove_list(&con->out_sent);
 331
 332         if (con->in_msg) {
 333                 ceph_msg_put(con->in_msg);
 334                 con->in_msg = NULL;
 335         }
 336
 337         con->connect_seq = 0;
 338         con->out_seq = 0;
 339         if (con->out_msg) {
 340                 ceph_msg_put(con->out_msg);
 341                 con->out_msg = NULL;
 342         }
 343         con->in_seq = 0;
 344         con->in_seq_acked = 0;
 345 }
 346
 347 /*
 348  * mark a peer down.  drop any open connections.
 349  */
 350 void ceph_con_close(struct ceph_connection *con)
 351 {
 352         dout("con_close %p peer %s\n", con, pr_addr(&con->peer_addr.in_addr));
 353         set_bit(CLOSED, &con->state);  /* in case there's queued work */
 354         clear_bit(STANDBY, &con->state);  /* avoid connect_seq bump */
 355         clear_bit(LOSSYTX, &con->state);  /* so we retry next connect */
 356         clear_bit(KEEPALIVE_PENDING, &con->state);
 357         clear_bit(WRITE_PENDING, &con->state);
 358         mutex_lock(&con->mutex);
 359         reset_connection(con);
 360         cancel_delayed_work(&con->work);
 361         mutex_unlock(&con->mutex);
 362         queue_con(con);
 363 }
 364
 365 /*
 366  * Reopen a closed connection, with a new peer address.
 367  */
 368 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
 369 {
 370         dout("con_open %p %s\n", con, pr_addr(&addr->in_addr));
 371         set_bit(OPENING, &con->state);
 372         clear_bit(CLOSED, &con->state);
 373         memcpy(&con->peer_addr, addr, sizeof(*addr));
 374         con->delay = 0;      /* reset backoff memory */
 375         queue_con(con);
 376 }
 377
 378 /*
 379  * return true if this connection ever successfully opened
 380  */
 381 bool ceph_con_opened(struct ceph_connection *con)
 382 {
 383         return con->connect_seq > 0;
 384 }
 385
 386 /*
 387  * generic get/put
 388  */
 389 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
 390 {
 391         dout("con_get %p nref = %d -> %d\n", con,
 392              atomic_read(&con->nref), atomic_read(&con->nref) + 1);
 393         if (atomic_inc_not_zero(&con->nref))
 394                 return con;
 395         return NULL;
 396 }
 397
 398 void ceph_con_put(struct ceph_connection *con)
 399 {
 400         dout("con_put %p nref = %d -> %d\n", con,
 401              atomic_read(&con->nref), atomic_read(&con->nref) - 1);
 402         BUG_ON(atomic_read(&con->nref) == 0);
 403         if (atomic_dec_and_test(&con->nref)) {
 404                 BUG_ON(con->sock);
 405                 kfree(con);
 406         }
 407 }
 408
 409 /*
 410  * initialize a new connection.
 411  */
 412 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
 413 {
 414         dout("con_init %p\n", con);
 415         memset(con, 0, sizeof(*con));
 416         atomic_set(&con->nref, 1);
 417         con->msgr = msgr;
 418         mutex_init(&con->mutex);
 419         INIT_LIST_HEAD(&con->out_queue);
 420         INIT_LIST_HEAD(&con->out_sent);
 421         INIT_DELAYED_WORK(&con->work, con_work);
 422 }
 423
 424
 425 /*
 426  * We maintain a global counter to order connection attempts.  Get
 427  * a unique seq greater than @gt.
 428  */
 429 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
 430 {
 431         u32 ret;
 432
 433         spin_lock(&msgr->global_seq_lock);
 434         if (msgr->global_seq < gt)
 435                 msgr->global_seq = gt;
 436         ret = ++msgr->global_seq;
 437         spin_unlock(&msgr->global_seq_lock);
 438         return ret;
 439 }
 440
 441
 442 /*
 443  * Prepare footer for currently outgoing message, and finish things
 444  * off.  Assumes out_kvec* are already valid.. we just add on to the end.
 445  */
 446 static void prepare_write_message_footer(struct ceph_connection *con, int v)
 447 {
 448         struct ceph_msg *m = con->out_msg;
 449
 450         dout("prepare_write_message_footer %p\n", con);
 451         con->out_kvec_is_msg = true;
 452         con->out_kvec[v].iov_base = &m->footer;
 453         con->out_kvec[v].iov_len = sizeof(m->footer);
 454         con->out_kvec_bytes += sizeof(m->footer);
 455         con->out_kvec_left++;
 456         con->out_more = m->more_to_follow;
 457         con->out_msg_done = true;
 458 }
 459
 460 /*
 461  * Prepare headers for the next outgoing message.
 462  */
 463 static void prepare_write_message(struct ceph_connection *con)
 464 {
 465         struct ceph_msg *m;
 466         int v = 0;
 467
 468         con->out_kvec_bytes = 0;
 469         con->out_kvec_is_msg = true;
 470         con->out_msg_done = false;
 471
 472         /* Sneak an ack in there first?  If we can get it into the same
 473          * TCP packet that's a good thing. */
 474         if (con->in_seq > con->in_seq_acked) {
 475                 con->in_seq_acked = con->in_seq;
 476                 con->out_kvec[v].iov_base = &tag_ack;
 477                 con->out_kvec[v++].iov_len = 1;
 478                 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 479                 con->out_kvec[v].iov_base = &con->out_temp_ack;
 480                 con->out_kvec[v++].iov_len = sizeof(con->out_temp_ack);
 481                 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 482         }
 483
 484         m = list_first_entry(&con->out_queue,
 485                        struct ceph_msg, list_head);
 486         con->out_msg = m;
 487         if (test_bit(LOSSYTX, &con->state)) {
 488                 list_del_init(&m->list_head);
 489         } else {
 490                 /* put message on sent list */
 491                 ceph_msg_get(m);
 492                 list_move_tail(&m->list_head, &con->out_sent);
 493         }
 494
 495         m->hdr.seq = cpu_to_le64(++con->out_seq);
 496
 497         dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
 498              m, con->out_seq, le16_to_cpu(m->hdr.type),
 499              le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
 500              le32_to_cpu(m->hdr.data_len),
 501              m->nr_pages);
 502         BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
 503
 504         /* tag + hdr + front + middle */
 505         con->out_kvec[v].iov_base = &tag_msg;
 506         con->out_kvec[v++].iov_len = 1;
 507         con->out_kvec[v].iov_base = &m->hdr;
 508         con->out_kvec[v++].iov_len = sizeof(m->hdr);
 509         con->out_kvec[v++] = m->front;
 510         if (m->middle)
 511                 con->out_kvec[v++] = m->middle->vec;
 512         con->out_kvec_left = v;
 513         con->out_kvec_bytes += 1 + sizeof(m->hdr) + m->front.iov_len +
 514                 (m->middle ? m->middle->vec.iov_len : 0);
 515         con->out_kvec_cur = con->out_kvec;
 516
 517         /* fill in crc (except data pages), footer */
 518         con->out_msg->hdr.crc =
 519                 cpu_to_le32(crc32c(0, (void *)&m->hdr,
 520                                       sizeof(m->hdr) - sizeof(m->hdr.crc)));
 521         con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
 522         con->out_msg->footer.front_crc =
 523                 cpu_to_le32(crc32c(0, m->front.iov_base, m->front.iov_len));
 524         if (m->middle)
 525                 con->out_msg->footer.middle_crc =
 526                         cpu_to_le32(crc32c(0, m->middle->vec.iov_base,
 527                                            m->middle->vec.iov_len));
 528         else
 529                 con->out_msg->footer.middle_crc = 0;
 530         con->out_msg->footer.data_crc = 0;
 531         dout("prepare_write_message front_crc %u data_crc %u\n",
 532              le32_to_cpu(con->out_msg->footer.front_crc),
 533              le32_to_cpu(con->out_msg->footer.middle_crc));
 534
 535         /* is there a data payload? */
 536         if (le32_to_cpu(m->hdr.data_len) > 0) {
 537                 /* initialize page iterator */
 538                 con->out_msg_pos.page = 0;
 539                 con->out_msg_pos.page_pos =
 540                         le16_to_cpu(m->hdr.data_off) & ~PAGE_MASK;
 541                 con->out_msg_pos.data_pos = 0;
 542                 con->out_msg_pos.did_page_crc = 0;
 543                 con->out_more = 1;  /* data + footer will follow */
 544         } else {
 545                 /* no, queue up footer too and be done */
 546                 prepare_write_message_footer(con, v);
 547         }
 548
 549         set_bit(WRITE_PENDING, &con->state);
 550 }
 551
 552 /*
 553  * Prepare an ack.
 554  */
 555 static void prepare_write_ack(struct ceph_connection *con)
 556 {
 557         dout("prepare_write_ack %p %llu -> %llu\n", con,
 558              con->in_seq_acked, con->in_seq);
 559         con->in_seq_acked = con->in_seq;
 560
 561         con->out_kvec[0].iov_base = &tag_ack;
 562         con->out_kvec[0].iov_len = 1;
 563         con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 564         con->out_kvec[1].iov_base = &con->out_temp_ack;
 565         con->out_kvec[1].iov_len = sizeof(con->out_temp_ack);
 566         con->out_kvec_left = 2;
 567         con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 568         con->out_kvec_cur = con->out_kvec;
 569         con->out_more = 1;  /* more will follow.. eventually.. */
 570         set_bit(WRITE_PENDING, &con->state);
 571 }
 572
 573 /*
 574  * Prepare to write keepalive byte.
 575  */
 576 static void prepare_write_keepalive(struct ceph_connection *con)
 577 {
 578         dout("prepare_write_keepalive %p\n", con);
 579         con->out_kvec[0].iov_base = &tag_keepalive;
 580         con->out_kvec[0].iov_len = 1;
 581         con->out_kvec_left = 1;
 582         con->out_kvec_bytes = 1;
 583         con->out_kvec_cur = con->out_kvec;
 584         set_bit(WRITE_PENDING, &con->state);
 585 }
 586
 587 /*
 588  * Connection negotiation.
 589  */
 590
 591 static void prepare_connect_authorizer(struct ceph_connection *con)
 592 {
 593         void *auth_buf;
 594         int auth_len = 0;
 595         int auth_protocol = 0;
 596
 597         mutex_unlock(&con->mutex);
 598         if (con->ops->get_authorizer)
 599                 con->ops->get_authorizer(con, &auth_buf, &auth_len,
 600                                          &auth_protocol, &con->auth_reply_buf,
 601                                          &con->auth_reply_buf_len,
 602                                          con->auth_retry);
 603         mutex_lock(&con->mutex);
 604
 605         con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
 606         con->out_connect.authorizer_len = cpu_to_le32(auth_len);
 607
 608         con->out_kvec[con->out_kvec_left].iov_base = auth_buf;
 609         con->out_kvec[con->out_kvec_left].iov_len = auth_len;
 610         con->out_kvec_left++;
 611         con->out_kvec_bytes += auth_len;
 612 }
 613
 614 /*
 615  * We connected to a peer and are saying hello.
 616  */
 617 static void prepare_write_banner(struct ceph_messenger *msgr,
 618                                  struct ceph_connection *con)
 619 {
 620         int len = strlen(CEPH_BANNER);
 621
 622         con->out_kvec[0].iov_base = CEPH_BANNER;
 623         con->out_kvec[0].iov_len = len;
 624         con->out_kvec[1].iov_base = &msgr->my_enc_addr;
 625         con->out_kvec[1].iov_len = sizeof(msgr->my_enc_addr);
 626         con->out_kvec_left = 2;
 627         con->out_kvec_bytes = len + sizeof(msgr->my_enc_addr);
 628         con->out_kvec_cur = con->out_kvec;
 629         con->out_more = 0;
 630         set_bit(WRITE_PENDING, &con->state);
 631 }
 632
 633 static void prepare_write_connect(struct ceph_messenger *msgr,
 634                                   struct ceph_connection *con,
 635                                   int after_banner)
 636 {
 637         unsigned global_seq = get_global_seq(con->msgr, 0);
 638         int proto;
 639
 640         switch (con->peer_name.type) {
 641         case CEPH_ENTITY_TYPE_MON:
 642                 proto = CEPH_MONC_PROTOCOL;
 643                 break;
 644         case CEPH_ENTITY_TYPE_OSD:
 645                 proto = CEPH_OSDC_PROTOCOL;
 646                 break;
 647         case CEPH_ENTITY_TYPE_MDS:
 648                 proto = CEPH_MDSC_PROTOCOL;
 649                 break;
 650         default:
 651                 BUG();
 652         }
 653
 654         dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
 655              con->connect_seq, global_seq, proto);
 656
 657         con->out_connect.features = CEPH_FEATURE_SUPPORTED;
 658         con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
 659         con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
 660         con->out_connect.global_seq = cpu_to_le32(global_seq);
 661         con->out_connect.protocol_version = cpu_to_le32(proto);
 662         con->out_connect.flags = 0;
 663
 664         if (!after_banner) {
 665                 con->out_kvec_left = 0;
 666                 con->out_kvec_bytes = 0;
 667         }
 668         con->out_kvec[con->out_kvec_left].iov_base = &con->out_connect;
 669         con->out_kvec[con->out_kvec_left].iov_len = sizeof(con->out_connect);
 670         con->out_kvec_left++;
 671         con->out_kvec_bytes += sizeof(con->out_connect);
 672         con->out_kvec_cur = con->out_kvec;
 673         con->out_more = 0;
 674         set_bit(WRITE_PENDING, &con->state);
 675
 676         prepare_connect_authorizer(con);
 677 }
 678
 679
 680 /*
 681  * write as much of pending kvecs to the socket as we can.
 682  *  1 -> done
 683  *  0 -> socket full, but more to do
 684  * <0 -> error
 685  */
 686 static int write_partial_kvec(struct ceph_connection *con)
 687 {
 688         int ret;
 689
 690         dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
 691         while (con->out_kvec_bytes > 0) {
 692                 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
 693                                        con->out_kvec_left, con->out_kvec_bytes,
 694                                        con->out_more);
 695                 if (ret <= 0)
 696                         goto out;
 697                 con->out_kvec_bytes -= ret;
 698                 if (con->out_kvec_bytes == 0)
 699                         break;            /* done */
 700                 while (ret > 0) {
 701                         if (ret >= con->out_kvec_cur->iov_len) {
 702                                 ret -= con->out_kvec_cur->iov_len;
 703                                 con->out_kvec_cur++;
 704                                 con->out_kvec_left--;
 705                         } else {
 706                                 con->out_kvec_cur->iov_len -= ret;
 707                                 con->out_kvec_cur->iov_base += ret;
 708                                 ret = 0;
 709                                 break;
 710                         }
 711                 }
 712         }
 713         con->out_kvec_left = 0;
 714         con->out_kvec_is_msg = false;
 715         ret = 1;
 716 out:
 717         dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
 718              con->out_kvec_bytes, con->out_kvec_left, ret);
 719         return ret;  /* done! */
 720 }
 721
 722 /*
 723  * Write as much message data payload as we can.  If we finish, queue
 724  * up the footer.
 725  *  1 -> done, footer is now queued in out_kvec[].
 726  *  0 -> socket full, but more to do
 727  * <0 -> error
 728  */
 729 static int write_partial_msg_pages(struct ceph_connection *con)
 730 {
 731         struct ceph_msg *msg = con->out_msg;
 732         unsigned data_len = le32_to_cpu(msg->hdr.data_len);
 733         size_t len;
 734         int crc = con->msgr->nocrc;
 735         int ret;
 736
 737         dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
 738              con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
 739              con->out_msg_pos.page_pos);
 740
 741         while (con->out_msg_pos.page < con->out_msg->nr_pages) {
 742                 struct page *page = NULL;
 743                 void *kaddr = NULL;
 744
 745                 /*
 746                  * if we are calculating the data crc (the default), we need
 747                  * to map the page.  if our pages[] has been revoked, use the
 748                  * zero page.
 749                  */
 750                 if (msg->pages) {
 751                         page = msg->pages[con->out_msg_pos.page];
 752                         if (crc)
 753                                 kaddr = kmap(page);
 754                 } else if (msg->pagelist) {
 755                         page = list_first_entry(&msg->pagelist->head,
 756                                                 struct page, lru);
 757                         if (crc)
 758                                 kaddr = kmap(page);
 759                 } else {
 760                         page = con->msgr->zero_page;
 761                         if (crc)
 762                                 kaddr = page_address(con->msgr->zero_page);
 763                 }
 764                 len = min((int)(PAGE_SIZE - con->out_msg_pos.page_pos),
 765                           (int)(data_len - con->out_msg_pos.data_pos));
 766                 if (crc && !con->out_msg_pos.did_page_crc) {
 767                         void *base = kaddr + con->out_msg_pos.page_pos;
 768                         u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
 769
 770                         BUG_ON(kaddr == NULL);
 771                         con->out_msg->footer.data_crc =
 772                                 cpu_to_le32(crc32c(tmpcrc, base, len));
 773                         con->out_msg_pos.did_page_crc = 1;
 774                 }
 775
 776                 ret = kernel_sendpage(con->sock, page,
 777                                       con->out_msg_pos.page_pos, len,
 778                                       MSG_DONTWAIT | MSG_NOSIGNAL |
 779                                       MSG_MORE);
 780
 781                 if (crc && (msg->pages || msg->pagelist))
 782                         kunmap(page);
 783
 784                 if (ret <= 0)
 785                         goto out;
 786
 787                 con->out_msg_pos.data_pos += ret;
 788                 con->out_msg_pos.page_pos += ret;
 789                 if (ret == len) {
 790                         con->out_msg_pos.page_pos = 0;
 791                         con->out_msg_pos.page++;
 792                         con->out_msg_pos.did_page_crc = 0;
 793                         if (msg->pagelist)
 794                                 list_move_tail(&page->lru,
 795                                                &msg->pagelist->head);
 796                 }
 797         }
 798
 799         dout("write_partial_msg_pages %p msg %p done\n", con, msg);
 800
 801         /* prepare and queue up footer, too */
 802         if (!crc)
 803                 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
 804         con->out_kvec_bytes = 0;
 805         con->out_kvec_left = 0;
 806         con->out_kvec_cur = con->out_kvec;
 807         prepare_write_message_footer(con, 0);
 808         ret = 1;
 809 out:
 810         return ret;
 811 }
 812
 813 /*
 814  * write some zeros
 815  */
 816 static int write_partial_skip(struct ceph_connection *con)
 817 {
 818         int ret;
 819
 820         while (con->out_skip > 0) {
 821                 struct kvec iov = {
 822                         .iov_base = page_address(con->msgr->zero_page),
 823                         .iov_len = min(con->out_skip, (int)PAGE_CACHE_SIZE)
 824                 };
 825
 826                 ret = ceph_tcp_sendmsg(con->sock, &iov, 1, iov.iov_len, 1);
 827                 if (ret <= 0)
 828                         goto out;
 829                 con->out_skip -= ret;
 830         }
 831         ret = 1;
 832 out:
 833         return ret;
 834 }
 835
 836 /*
 837  * Prepare to read connection handshake, or an ack.
 838  */
 839 static void prepare_read_banner(struct ceph_connection *con)
 840 {
 841         dout("prepare_read_banner %p\n", con);
 842         con->in_base_pos = 0;
 843 }
 844
 845 static void prepare_read_connect(struct ceph_connection *con)
 846 {
 847         dout("prepare_read_connect %p\n", con);
 848         con->in_base_pos = 0;
 849 }
 850
 851 static void prepare_read_ack(struct ceph_connection *con)
 852 {
 853         dout("prepare_read_ack %p\n", con);
 854         con->in_base_pos = 0;
 855 }
 856
 857 static void prepare_read_tag(struct ceph_connection *con)
 858 {
 859         dout("prepare_read_tag %p\n", con);
 860         con->in_base_pos = 0;
 861         con->in_tag = CEPH_MSGR_TAG_READY;
 862 }
 863
 864 /*
 865  * Prepare to read a message.
 866  */
 867 static int prepare_read_message(struct ceph_connection *con)
 868 {
 869         dout("prepare_read_message %p\n", con);
 870         BUG_ON(con->in_msg != NULL);
 871         con->in_base_pos = 0;
 872         con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
 873         return 0;
 874 }
 875
 876
 877 static int read_partial(struct ceph_connection *con,
 878                         int *to, int size, void *object)
 879 {
 880         *to += size;
 881         while (con->in_base_pos < *to) {
 882                 int left = *to - con->in_base_pos;
 883                 int have = size - left;
 884                 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
 885                 if (ret <= 0)
 886                         return ret;
 887                 con->in_base_pos += ret;
 888         }
 889         return 1;
 890 }
 891
 892
 893 /*
 894  * Read all or part of the connect-side handshake on a new connection
 895  */
 896 static int read_partial_banner(struct ceph_connection *con)
 897 {
 898         int ret, to = 0;
 899
 900         dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
 901
 902         /* peer's banner */
 903         ret = read_partial(con, &to, strlen(CEPH_BANNER), con->in_banner);
 904         if (ret <= 0)
 905                 goto out;
 906         ret = read_partial(con, &to, sizeof(con->actual_peer_addr),
 907                            &con->actual_peer_addr);
 908         if (ret <= 0)
 909                 goto out;
 910         ret = read_partial(con, &to, sizeof(con->peer_addr_for_me),
 911                            &con->peer_addr_for_me);
 912         if (ret <= 0)
 913                 goto out;
 914 out:
 915         return ret;
 916 }
 917
 918 static int read_partial_connect(struct ceph_connection *con)
 919 {
 920         int ret, to = 0;
 921
 922         dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
 923
 924         ret = read_partial(con, &to, sizeof(con->in_reply), &con->in_reply);
 925         if (ret <= 0)
 926                 goto out;
 927         ret = read_partial(con, &to, le32_to_cpu(con->in_reply.authorizer_len),
 928                            con->auth_reply_buf);
 929         if (ret <= 0)
 930                 goto out;
 931
 932         dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
 933              con, (int)con->in_reply.tag,
 934              le32_to_cpu(con->in_reply.connect_seq),
 935              le32_to_cpu(con->in_reply.global_seq));
 936 out:
 937         return ret;
 938
 939 }
 940
 941 /*
 942  * Verify the hello banner looks okay.
 943  */
 944 static int verify_hello(struct ceph_connection *con)
 945 {
 946         if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
 947                 pr_err("connect to %s got bad banner\n",
 948                        pr_addr(&con->peer_addr.in_addr));
 949                 con->error_msg = "protocol error, bad banner";
 950                 return -1;
 951         }
 952         return 0;
 953 }
 954
 955 static bool addr_is_blank(struct sockaddr_storage *ss)
 956 {
 957         switch (ss->ss_family) {
 958         case AF_INET:
 959                 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
 960         case AF_INET6:
 961                 return
 962                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
 963                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
 964                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
 965                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
 966         }
 967         return false;
 968 }
 969
 970 static int addr_port(struct sockaddr_storage *ss)
 971 {
 972         switch (ss->ss_family) {
 973         case AF_INET:
 974                 return ntohs(((struct sockaddr_in *)ss)->sin_port);
 975         case AF_INET6:
 976                 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
 977         }
 978         return 0;
 979 }
 980
 981 static void addr_set_port(struct sockaddr_storage *ss, int p)
 982 {
 983         switch (ss->ss_family) {
 984         case AF_INET:
 985                 ((struct sockaddr_in *)ss)->sin_port = htons(p);
 986         case AF_INET6:
 987                 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
 988         }
 989 }
 990
 991 /*
 992  * Parse an ip[:port] list into an addr array.  Use the default
 993  * monitor port if a port isn't specified.
 994  */
 995 int ceph_parse_ips(const char *c, const char *end,
 996                    struct ceph_entity_addr *addr,
 997                    int max_count, int *count)
 998 {
 999         int i;
1000         const char *p = c;
1001
1002         dout("parse_ips on '%.*s'\n", (int)(end-c), c);
1003         for (i = 0; i < max_count; i++) {
1004                 const char *ipend;
1005                 struct sockaddr_storage *ss = &addr[i].in_addr;
1006                 struct sockaddr_in *in4 = (void *)ss;
1007                 struct sockaddr_in6 *in6 = (void *)ss;
1008                 int port;
1009
1010                 memset(ss, 0, sizeof(*ss));
1011                 if (in4_pton(p, end - p, (u8 *)&in4->sin_addr.s_addr,
1012                              ',', &ipend)) {
1013                         ss->ss_family = AF_INET;
1014                 } else if (in6_pton(p, end - p, (u8 *)&in6->sin6_addr.s6_addr,
1015                                     ',', &ipend)) {
1016                         ss->ss_family = AF_INET6;
1017                 } else {
1018                         goto bad;
1019                 }
1020                 p = ipend;
1021
1022                 /* port? */
1023                 if (p < end && *p == ':') {
1024                         port = 0;
1025                         p++;
1026                         while (p < end && *p >= '0' && *p <= '9') {
1027                                 port = (port * 10) + (*p - '0');
1028                                 p++;
1029                         }
1030                         if (port > 65535 || port == 0)
1031                                 goto bad;
1032                 } else {
1033                         port = CEPH_MON_PORT;
1034                 }
1035
1036                 addr_set_port(ss, port);
1037
1038                 dout("parse_ips got %s\n", pr_addr(ss));
1039
1040                 if (p == end)
1041                         break;
1042                 if (*p != ',')
1043                         goto bad;
1044                 p++;
1045         }
1046
1047         if (p != end)
1048                 goto bad;
1049
1050         if (count)
1051                 *count = i + 1;
1052         return 0;
1053
1054 bad:
1055         pr_err("parse_ips bad ip '%s'\n", c);
1056         return -EINVAL;
1057 }
1058
1059 static int process_banner(struct ceph_connection *con)
1060 {
1061         dout("process_banner on %p\n", con);
1062
1063         if (verify_hello(con) < 0)
1064                 return -1;
1065
1066         ceph_decode_addr(&con->actual_peer_addr);
1067         ceph_decode_addr(&con->peer_addr_for_me);
1068
1069         /*
1070          * Make sure the other end is who we wanted.  note that the other
1071          * end may not yet know their ip address, so if it's 0.0.0.0, give
1072          * them the benefit of the doubt.
1073          */
1074         if (memcmp(&con->peer_addr, &con->actual_peer_addr,
1075                    sizeof(con->peer_addr)) != 0 &&
1076             !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1077               con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1078                 pr_warning("wrong peer, want %s/%lld, got %s/%lld\n",
1079                            pr_addr(&con->peer_addr.in_addr),
1080                            le64_to_cpu(con->peer_addr.nonce),
1081                            pr_addr(&con->actual_peer_addr.in_addr),
1082                            le64_to_cpu(con->actual_peer_addr.nonce));
1083                 con->error_msg = "wrong peer at address";
1084                 return -1;
1085         }
1086
1087         /*
1088          * did we learn our address?
1089          */
1090         if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1091                 int port = addr_port(&con->msgr->inst.addr.in_addr);
1092
1093                 memcpy(&con->msgr->inst.addr.in_addr,
1094                        &con->peer_addr_for_me.in_addr,
1095                        sizeof(con->peer_addr_for_me.in_addr));
1096                 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1097                 encode_my_addr(con->msgr);
1098                 dout("process_banner learned my addr is %s\n",
1099                      pr_addr(&con->msgr->inst.addr.in_addr));
1100         }
1101
1102         set_bit(NEGOTIATING, &con->state);
1103         prepare_read_connect(con);
1104         return 0;
1105 }
1106
1107 static void fail_protocol(struct ceph_connection *con)
1108 {
1109         reset_connection(con);
1110         set_bit(CLOSED, &con->state);  /* in case there's queued work */
1111
1112         mutex_unlock(&con->mutex);
1113         if (con->ops->bad_proto)
1114                 con->ops->bad_proto(con);
1115         mutex_lock(&con->mutex);
1116 }
1117
1118 static int process_connect(struct ceph_connection *con)
1119 {
1120         u64 sup_feat = CEPH_FEATURE_SUPPORTED;
1121         u64 req_feat = CEPH_FEATURE_REQUIRED;
1122         u64 server_feat = le64_to_cpu(con->in_reply.features);
1123
1124         dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1125
1126         switch (con->in_reply.tag) {
1127         case CEPH_MSGR_TAG_FEATURES:
1128                 pr_err("%s%lld %s feature set mismatch,"
1129                        " my %llx < server's %llx, missing %llx\n",
1130                        ENTITY_NAME(con->peer_name),
1131                        pr_addr(&con->peer_addr.in_addr),
1132                        sup_feat, server_feat, server_feat & ~sup_feat);
1133                 con->error_msg = "missing required protocol features";
1134                 fail_protocol(con);
1135                 return -1;
1136
1137         case CEPH_MSGR_TAG_BADPROTOVER:
1138                 pr_err("%s%lld %s protocol version mismatch,"
1139                        " my %d != server's %d\n",
1140                        ENTITY_NAME(con->peer_name),
1141                        pr_addr(&con->peer_addr.in_addr),
1142                        le32_to_cpu(con->out_connect.protocol_version),
1143                        le32_to_cpu(con->in_reply.protocol_version));
1144                 con->error_msg = "protocol version mismatch";
1145                 fail_protocol(con);
1146                 return -1;
1147
1148         case CEPH_MSGR_TAG_BADAUTHORIZER:
1149                 con->auth_retry++;
1150                 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1151                      con->auth_retry);
1152                 if (con->auth_retry == 2) {
1153                         con->error_msg = "connect authorization failure";
1154                         reset_connection(con);
1155                         set_bit(CLOSED, &con->state);
1156                         return -1;
1157                 }
1158                 con->auth_retry = 1;
1159                 prepare_write_connect(con->msgr, con, 0);
1160                 prepare_read_connect(con);
1161                 break;
1162
1163         case CEPH_MSGR_TAG_RESETSESSION:
1164                 /*
1165                  * If we connected with a large connect_seq but the peer
1166                  * has no record of a session with us (no connection, or
1167                  * connect_seq == 0), they will send RESETSESION to indicate
1168                  * that they must have reset their session, and may have
1169                  * dropped messages.
1170                  */
1171                 dout("process_connect got RESET peer seq %u\n",
1172                      le32_to_cpu(con->in_connect.connect_seq));
1173                 pr_err("%s%lld %s connection reset\n",
1174                        ENTITY_NAME(con->peer_name),
1175                        pr_addr(&con->peer_addr.in_addr));
1176                 reset_connection(con);
1177                 prepare_write_connect(con->msgr, con, 0);
1178                 prepare_read_connect(con);
1179
1180                 /* Tell ceph about it. */
1181                 mutex_unlock(&con->mutex);
1182                 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1183                 if (con->ops->peer_reset)
1184                         con->ops->peer_reset(con);
1185                 mutex_lock(&con->mutex);
1186                 break;
1187
1188         case CEPH_MSGR_TAG_RETRY_SESSION:
1189                 /*
1190                  * If we sent a smaller connect_seq than the peer has, try
1191                  * again with a larger value.
1192                  */
1193                 dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
1194                      le32_to_cpu(con->out_connect.connect_seq),
1195                      le32_to_cpu(con->in_connect.connect_seq));
1196                 con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
1197                 prepare_write_connect(con->msgr, con, 0);
1198                 prepare_read_connect(con);
1199                 break;
1200
1201         case CEPH_MSGR_TAG_RETRY_GLOBAL:
1202                 /*
1203                  * If we sent a smaller global_seq than the peer has, try
1204                  * again with a larger value.
1205                  */
1206                 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1207                      con->peer_global_seq,
1208                      le32_to_cpu(con->in_connect.global_seq));
1209                 get_global_seq(con->msgr,
1210                                le32_to_cpu(con->in_connect.global_seq));
1211                 prepare_write_connect(con->msgr, con, 0);
1212                 prepare_read_connect(con);
1213                 break;
1214
1215         case CEPH_MSGR_TAG_READY:
1216                 if (req_feat & ~server_feat) {
1217                         pr_err("%s%lld %s protocol feature mismatch,"
1218                                " my required %llx > server's %llx, need %llx\n",
1219                                ENTITY_NAME(con->peer_name),
1220                                pr_addr(&con->peer_addr.in_addr),
1221                                req_feat, server_feat, req_feat & ~server_feat);
1222                         con->error_msg = "missing required protocol features";
1223                         fail_protocol(con);
1224                         return -1;
1225                 }
1226                 clear_bit(CONNECTING, &con->state);
1227                 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1228                 con->connect_seq++;
1229                 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1230                      con->peer_global_seq,
1231                      le32_to_cpu(con->in_reply.connect_seq),
1232                      con->connect_seq);
1233                 WARN_ON(con->connect_seq !=
1234                         le32_to_cpu(con->in_reply.connect_seq));
1235
1236                 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1237                         set_bit(LOSSYTX, &con->state);
1238
1239                 prepare_read_tag(con);
1240                 break;
1241
1242         case CEPH_MSGR_TAG_WAIT:
1243                 /*
1244                  * If there is a connection race (we are opening
1245                  * connections to each other), one of us may just have
1246                  * to WAIT.  This shouldn't happen if we are the
1247                  * client.
1248                  */
1249                 pr_err("process_connect peer connecting WAIT\n");
1250
1251         default:
1252                 pr_err("connect protocol error, will retry\n");
1253                 con->error_msg = "protocol error, garbage tag during connect";
1254                 return -1;
1255         }
1256         return 0;
1257 }
1258
1259
1260 /*
1261  * read (part of) an ack
1262  */
1263 static int read_partial_ack(struct ceph_connection *con)
1264 {
1265         int to = 0;
1266
1267         return read_partial(con, &to, sizeof(con->in_temp_ack),
1268                             &con->in_temp_ack);
1269 }
1270
1271
1272 /*
1273  * We can finally discard anything that's been acked.
1274  */
1275 static void process_ack(struct ceph_connection *con)
1276 {
1277         struct ceph_msg *m;
1278         u64 ack = le64_to_cpu(con->in_temp_ack);
1279         u64 seq;
1280
1281         while (!list_empty(&con->out_sent)) {
1282                 m = list_first_entry(&con->out_sent, struct ceph_msg,
1283                                      list_head);
1284                 seq = le64_to_cpu(m->hdr.seq);
1285                 if (seq > ack)
1286                         break;
1287                 dout("got ack for seq %llu type %d at %p\n", seq,
1288                      le16_to_cpu(m->hdr.type), m);
1289                 ceph_msg_remove(m);
1290         }
1291         prepare_read_tag(con);
1292 }
1293
1294
1295
1296
1297 static int read_partial_message_section(struct ceph_connection *con,
1298                                         struct kvec *section, unsigned int sec_len,
1299                                         u32 *crc)
1300 {
1301         int left;
1302         int ret;
1303
1304         BUG_ON(!section);
1305
1306         while (section->iov_len < sec_len) {
1307                 BUG_ON(section->iov_base == NULL);
1308                 left = sec_len - section->iov_len;
1309                 ret = ceph_tcp_recvmsg(con->sock, (char *)section->iov_base +
1310                                        section->iov_len, left);
1311                 if (ret <= 0)
1312                         return ret;
1313                 section->iov_len += ret;
1314                 if (section->iov_len == sec_len)
1315                         *crc = crc32c(0, section->iov_base,
1316                                       section->iov_len);
1317         }
1318
1319         return 1;
1320 }
1321
1322 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
1323                                 struct ceph_msg_header *hdr,
1324                                 int *skip);
1325 /*
1326  * read (part of) a message.
1327  */
1328 static int read_partial_message(struct ceph_connection *con)
1329 {
1330         struct ceph_msg *m = con->in_msg;
1331         void *p;
1332         int ret;
1333         int to, left;
1334         unsigned front_len, middle_len, data_len, data_off;
1335         int datacrc = con->msgr->nocrc;
1336         int skip;
1337
1338         dout("read_partial_message con %p msg %p\n", con, m);
1339
1340         /* header */
1341         while (con->in_base_pos < sizeof(con->in_hdr)) {
1342                 left = sizeof(con->in_hdr) - con->in_base_pos;
1343                 ret = ceph_tcp_recvmsg(con->sock,
1344                                        (char *)&con->in_hdr + con->in_base_pos,
1345                                        left);
1346                 if (ret <= 0)
1347                         return ret;
1348                 con->in_base_pos += ret;
1349                 if (con->in_base_pos == sizeof(con->in_hdr)) {
1350                         u32 crc = crc32c(0, (void *)&con->in_hdr,
1351                                  sizeof(con->in_hdr) - sizeof(con->in_hdr.crc));
1352                         if (crc != le32_to_cpu(con->in_hdr.crc)) {
1353                                 pr_err("read_partial_message bad hdr "
1354                                        " crc %u != expected %u\n",
1355                                        crc, con->in_hdr.crc);
1356                                 return -EBADMSG;
1357                         }
1358                 }
1359         }
1360         front_len = le32_to_cpu(con->in_hdr.front_len);
1361         if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1362                 return -EIO;
1363         middle_len = le32_to_cpu(con->in_hdr.middle_len);
1364         if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1365                 return -EIO;
1366         data_len = le32_to_cpu(con->in_hdr.data_len);
1367         if (data_len > CEPH_MSG_MAX_DATA_LEN)
1368                 return -EIO;
1369         data_off = le16_to_cpu(con->in_hdr.data_off);
1370
1371         /* allocate message? */
1372         if (!con->in_msg) {
1373                 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1374                      con->in_hdr.front_len, con->in_hdr.data_len);
1375                 con->in_msg = ceph_alloc_msg(con, &con->in_hdr, &skip);
1376                 if (skip) {
1377                         /* skip this message */
1378                         dout("alloc_msg returned NULL, skipping message\n");
1379                         con->in_base_pos = -front_len - middle_len - data_len -
1380                                 sizeof(m->footer);
1381                         con->in_tag = CEPH_MSGR_TAG_READY;
1382                         return 0;
1383                 }
1384                 if (IS_ERR(con->in_msg)) {
1385                         ret = PTR_ERR(con->in_msg);
1386                         con->in_msg = NULL;
1387                         con->error_msg =
1388                                 "error allocating memory for incoming message";
1389                         return ret;
1390                 }
1391                 m = con->in_msg;
1392                 m->front.iov_len = 0;    /* haven't read it yet */
1393                 if (m->middle)
1394                         m->middle->vec.iov_len = 0;
1395
1396                 con->in_msg_pos.page = 0;
1397                 con->in_msg_pos.page_pos = data_off & ~PAGE_MASK;
1398                 con->in_msg_pos.data_pos = 0;
1399         }
1400
1401         /* front */
1402         ret = read_partial_message_section(con, &m->front, front_len,
1403                                            &con->in_front_crc);
1404         if (ret <= 0)
1405                 return ret;
1406
1407         /* middle */
1408         if (m->middle) {
1409                 ret = read_partial_message_section(con, &m->middle->vec, middle_len,
1410                                                    &con->in_middle_crc);
1411                 if (ret <= 0)
1412                         return ret;
1413         }
1414
1415         /* (page) data */
1416         while (con->in_msg_pos.data_pos < data_len) {
1417                 left = min((int)(data_len - con->in_msg_pos.data_pos),
1418                            (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1419                 BUG_ON(m->pages == NULL);
1420                 p = kmap(m->pages[con->in_msg_pos.page]);
1421                 ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1422                                        left);
1423                 if (ret > 0 && datacrc)
1424                         con->in_data_crc =
1425                                 crc32c(con->in_data_crc,
1426                                           p + con->in_msg_pos.page_pos, ret);
1427                 kunmap(m->pages[con->in_msg_pos.page]);
1428                 if (ret <= 0)
1429                         return ret;
1430                 con->in_msg_pos.data_pos += ret;
1431                 con->in_msg_pos.page_pos += ret;
1432                 if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1433                         con->in_msg_pos.page_pos = 0;
1434                         con->in_msg_pos.page++;
1435                 }
1436         }
1437
1438         /* footer */
1439         to = sizeof(m->hdr) + sizeof(m->footer);
1440         while (con->in_base_pos < to) {
1441                 left = to - con->in_base_pos;
1442                 ret = ceph_tcp_recvmsg(con->sock, (char *)&m->footer +
1443                                        (con->in_base_pos - sizeof(m->hdr)),
1444                                        left);
1445                 if (ret <= 0)
1446                         return ret;
1447                 con->in_base_pos += ret;
1448         }
1449         dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1450              m, front_len, m->footer.front_crc, middle_len,
1451              m->footer.middle_crc, data_len, m->footer.data_crc);
1452
1453         /* crc ok? */
1454         if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1455                 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1456                        m, con->in_front_crc, m->footer.front_crc);
1457                 return -EBADMSG;
1458         }
1459         if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1460                 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1461                        m, con->in_middle_crc, m->footer.middle_crc);
1462                 return -EBADMSG;
1463         }
1464         if (datacrc &&
1465             (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1466             con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1467                 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1468                        con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1469                 return -EBADMSG;
1470         }
1471
1472         return 1; /* done! */
1473 }
1474
1475 /*
1476  * Process message.  This happens in the worker thread.  The callback should
1477  * be careful not to do anything that waits on other incoming messages or it
1478  * may deadlock.
1479  */
1480 static void process_message(struct ceph_connection *con)
1481 {
1482         struct ceph_msg *msg;
1483
1484         msg = con->in_msg;
1485         con->in_msg = NULL;
1486
1487         /* if first message, set peer_name */
1488         if (con->peer_name.type == 0)
1489                 con->peer_name = msg->hdr.src.name;
1490
1491         con->in_seq++;
1492         mutex_unlock(&con->mutex);
1493
1494         dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1495              msg, le64_to_cpu(msg->hdr.seq),
1496              ENTITY_NAME(msg->hdr.src.name),
1497              le16_to_cpu(msg->hdr.type),
1498              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1499              le32_to_cpu(msg->hdr.front_len),
1500              le32_to_cpu(msg->hdr.data_len),
1501              con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1502         con->ops->dispatch(con, msg);
1503
1504         mutex_lock(&con->mutex);
1505         prepare_read_tag(con);
1506 }
1507
1508
1509 /*
1510  * Write something to the socket.  Called in a worker thread when the
1511  * socket appears to be writeable and we have something ready to send.
1512  */
1513 static int try_write(struct ceph_connection *con)
1514 {
1515         struct ceph_messenger *msgr = con->msgr;
1516         int ret = 1;
1517
1518         dout("try_write start %p state %lu nref %d\n", con, con->state,
1519              atomic_read(&con->nref));
1520
1521         mutex_lock(&con->mutex);
1522 more:
1523         dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1524
1525         /* open the socket first? */
1526         if (con->sock == NULL) {
1527                 /*
1528                  * if we were STANDBY and are reconnecting _this_
1529                  * connection, bump connect_seq now.  Always bump
1530                  * global_seq.
1531                  */
1532                 if (test_and_clear_bit(STANDBY, &con->state))
1533                         con->connect_seq++;
1534
1535                 prepare_write_banner(msgr, con);
1536                 prepare_write_connect(msgr, con, 1);
1537                 prepare_read_banner(con);
1538                 set_bit(CONNECTING, &con->state);
1539                 clear_bit(NEGOTIATING, &con->state);
1540
1541                 BUG_ON(con->in_msg);
1542                 con->in_tag = CEPH_MSGR_TAG_READY;
1543                 dout("try_write initiating connect on %p new state %lu\n",
1544                      con, con->state);
1545                 con->sock = ceph_tcp_connect(con);
1546                 if (IS_ERR(con->sock)) {
1547                         con->sock = NULL;
1548                         con->error_msg = "connect error";
1549                         ret = -1;
1550                         goto out;
1551                 }
1552         }
1553
1554 more_kvec:
1555         /* kvec data queued? */
1556         if (con->out_skip) {
1557                 ret = write_partial_skip(con);
1558                 if (ret <= 0)
1559                         goto done;
1560                 if (ret < 0) {
1561                         dout("try_write write_partial_skip err %d\n", ret);
1562                         goto done;
1563                 }
1564         }
1565         if (con->out_kvec_left) {
1566                 ret = write_partial_kvec(con);
1567                 if (ret <= 0)
1568                         goto done;
1569         }
1570
1571         /* msg pages? */
1572         if (con->out_msg) {
1573                 if (con->out_msg_done) {
1574                         ceph_msg_put(con->out_msg);
1575                         con->out_msg = NULL;   /* we're done with this one */
1576                         goto do_next;
1577                 }
1578
1579                 ret = write_partial_msg_pages(con);
1580                 if (ret == 1)
1581                         goto more_kvec;  /* we need to send the footer, too! */
1582                 if (ret == 0)
1583                         goto done;
1584                 if (ret < 0) {
1585                         dout("try_write write_partial_msg_pages err %d\n",
1586                              ret);
1587                         goto done;
1588                 }
1589         }
1590
1591 do_next:
1592         if (!test_bit(CONNECTING, &con->state)) {
1593                 /* is anything else pending? */
1594                 if (!list_empty(&con->out_queue)) {
1595                         prepare_write_message(con);
1596                         goto more;
1597                 }
1598                 if (con->in_seq > con->in_seq_acked) {
1599                         prepare_write_ack(con);
1600                         goto more;
1601                 }
1602                 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1603                         prepare_write_keepalive(con);
1604                         goto more;
1605                 }
1606         }
1607
1608         /* Nothing to do! */
1609         clear_bit(WRITE_PENDING, &con->state);
1610         dout("try_write nothing else to write.\n");
1611 done:
1612         ret = 0;
1613 out:
1614         mutex_unlock(&con->mutex);
1615         dout("try_write done on %p\n", con);
1616         return ret;
1617 }
1618
1619
1620
1621 /*
1622  * Read what we can from the socket.
1623  */
1624 static int try_read(struct ceph_connection *con)
1625 {
1626         struct ceph_messenger *msgr;
1627         int ret = -1;
1628
1629         if (!con->sock)
1630                 return 0;
1631
1632         if (test_bit(STANDBY, &con->state))
1633                 return 0;
1634
1635         dout("try_read start on %p\n", con);
1636         msgr = con->msgr;
1637
1638         mutex_lock(&con->mutex);
1639
1640 more:
1641         dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1642              con->in_base_pos);
1643         if (test_bit(CONNECTING, &con->state)) {
1644                 if (!test_bit(NEGOTIATING, &con->state)) {
1645                         dout("try_read connecting\n");
1646                         ret = read_partial_banner(con);
1647                         if (ret <= 0)
1648                                 goto done;
1649                         if (process_banner(con) < 0) {
1650                                 ret = -1;
1651                                 goto out;
1652                         }
1653                 }
1654                 ret = read_partial_connect(con);
1655                 if (ret <= 0)
1656                         goto done;
1657                 if (process_connect(con) < 0) {
1658                         ret = -1;
1659                         goto out;
1660                 }
1661                 goto more;
1662         }
1663
1664         if (con->in_base_pos < 0) {
1665                 /*
1666                  * skipping + discarding content.
1667                  *
1668                  * FIXME: there must be a better way to do this!
1669                  */
1670                 static char buf[1024];
1671                 int skip = min(1024, -con->in_base_pos);
1672                 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
1673                 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
1674                 if (ret <= 0)
1675                         goto done;
1676                 con->in_base_pos += ret;
1677                 if (con->in_base_pos)
1678                         goto more;
1679         }
1680         if (con->in_tag == CEPH_MSGR_TAG_READY) {
1681                 /*
1682                  * what's next?
1683                  */
1684                 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
1685                 if (ret <= 0)
1686                         goto done;
1687                 dout("try_read got tag %d\n", (int)con->in_tag);
1688                 switch (con->in_tag) {
1689                 case CEPH_MSGR_TAG_MSG:
1690                         prepare_read_message(con);
1691                         break;
1692                 case CEPH_MSGR_TAG_ACK:
1693                         prepare_read_ack(con);
1694                         break;
1695                 case CEPH_MSGR_TAG_CLOSE:
1696                         set_bit(CLOSED, &con->state);   /* fixme */
1697                         goto done;
1698                 default:
1699                         goto bad_tag;
1700                 }
1701         }
1702         if (con->in_tag == CEPH_MSGR_TAG_MSG) {
1703                 ret = read_partial_message(con);
1704                 if (ret <= 0) {
1705                         switch (ret) {
1706                         case -EBADMSG:
1707                                 con->error_msg = "bad crc";
1708                                 ret = -EIO;
1709                                 goto out;
1710                         case -EIO:
1711                                 con->error_msg = "io error";
1712                                 goto out;
1713                         default:
1714                                 goto done;
1715                         }
1716                 }
1717                 if (con->in_tag == CEPH_MSGR_TAG_READY)
1718                         goto more;
1719                 process_message(con);
1720                 goto more;
1721         }
1722         if (con->in_tag == CEPH_MSGR_TAG_ACK) {
1723                 ret = read_partial_ack(con);
1724                 if (ret <= 0)
1725                         goto done;
1726                 process_ack(con);
1727                 goto more;
1728         }
1729
1730 done:
1731         ret = 0;
1732 out:
1733         mutex_unlock(&con->mutex);
1734         dout("try_read done on %p\n", con);
1735         return ret;
1736
1737 bad_tag:
1738         pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
1739         con->error_msg = "protocol error, garbage tag";
1740         ret = -1;
1741         goto out;
1742 }
1743
1744
1745 /*
1746  * Atomically queue work on a connection.  Bump @con reference to
1747  * avoid races with connection teardown.
1748  *
1749  * There is some trickery going on with QUEUED and BUSY because we
1750  * only want a _single_ thread operating on each connection at any
1751  * point in time, but we want to use all available CPUs.
1752  *
1753  * The worker thread only proceeds if it can atomically set BUSY.  It
1754  * clears QUEUED and does it's thing.  When it thinks it's done, it
1755  * clears BUSY, then rechecks QUEUED.. if it's set again, it loops
1756  * (tries again to set BUSY).
1757  *
1758  * To queue work, we first set QUEUED, _then_ if BUSY isn't set, we
1759  * try to queue work.  If that fails (work is already queued, or BUSY)
1760  * we give up (work also already being done or is queued) but leave QUEUED
1761  * set so that the worker thread will loop if necessary.
1762  */
1763 static void queue_con(struct ceph_connection *con)
1764 {
1765         if (test_bit(DEAD, &con->state)) {
1766                 dout("queue_con %p ignoring: DEAD\n",
1767                      con);
1768                 return;
1769         }
1770
1771         if (!con->ops->get(con)) {
1772                 dout("queue_con %p ref count 0\n", con);
1773                 return;
1774         }
1775
1776         set_bit(QUEUED, &con->state);
1777         if (test_bit(BUSY, &con->state)) {
1778                 dout("queue_con %p - already BUSY\n", con);
1779                 con->ops->put(con);
1780         } else if (!queue_work(ceph_msgr_wq, &con->work.work)) {
1781                 dout("queue_con %p - already queued\n", con);
1782                 con->ops->put(con);
1783         } else {
1784                 dout("queue_con %p\n", con);
1785         }
1786 }
1787
1788 /*
1789  * Do some work on a connection.  Drop a connection ref when we're done.
1790  */
1791 static void con_work(struct work_struct *work)
1792 {
1793         struct ceph_connection *con = container_of(work, struct ceph_connection,
1794                                                    work.work);
1795         int backoff = 0;
1796
1797 more:
1798         if (test_and_set_bit(BUSY, &con->state) != 0) {
1799                 dout("con_work %p BUSY already set\n", con);
1800                 goto out;
1801         }
1802         dout("con_work %p start, clearing QUEUED\n", con);
1803         clear_bit(QUEUED, &con->state);
1804
1805         if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
1806                 dout("con_work CLOSED\n");
1807                 con_close_socket(con);
1808                 goto done;
1809         }
1810         if (test_and_clear_bit(OPENING, &con->state)) {
1811                 /* reopen w/ new peer */
1812                 dout("con_work OPENING\n");
1813                 con_close_socket(con);
1814         }
1815
1816         if (test_and_clear_bit(SOCK_CLOSED, &con->state) ||
1817             try_read(con) < 0 ||
1818             try_write(con) < 0) {
1819                 backoff = 1;
1820                 ceph_fault(con);     /* error/fault path */
1821         }
1822
1823 done:
1824         clear_bit(BUSY, &con->state);
1825         dout("con->state=%lu\n", con->state);
1826         if (test_bit(QUEUED, &con->state)) {
1827                 if (!backoff || test_bit(OPENING, &con->state)) {
1828                         dout("con_work %p QUEUED reset, looping\n", con);
1829                         goto more;
1830                 }
1831                 dout("con_work %p QUEUED reset, but just faulted\n", con);
1832                 clear_bit(QUEUED, &con->state);
1833         }
1834         dout("con_work %p done\n", con);
1835
1836 out:
1837         con->ops->put(con);
1838 }
1839
1840
1841 /*
1842  * Generic error/fault handler.  A retry mechanism is used with
1843  * exponential backoff
1844  */
1845 static void ceph_fault(struct ceph_connection *con)
1846 {
1847         pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
1848                pr_addr(&con->peer_addr.in_addr), con->error_msg);
1849         dout("fault %p state %lu to peer %s\n",
1850              con, con->state, pr_addr(&con->peer_addr.in_addr));
1851
1852         if (test_bit(LOSSYTX, &con->state)) {
1853                 dout("fault on LOSSYTX channel\n");
1854                 goto out;
1855         }
1856
1857         mutex_lock(&con->mutex);
1858         if (test_bit(CLOSED, &con->state))
1859                 goto out_unlock;
1860
1861         con_close_socket(con);
1862
1863         if (con->in_msg) {
1864                 ceph_msg_put(con->in_msg);
1865                 con->in_msg = NULL;
1866         }
1867
1868         /* Requeue anything that hasn't been acked */
1869         list_splice_init(&con->out_sent, &con->out_queue);
1870
1871         /* If there are no messages in the queue, place the connection
1872          * in a STANDBY state (i.e., don't try to reconnect just yet). */
1873         if (list_empty(&con->out_queue) && !con->out_keepalive_pending) {
1874                 dout("fault setting STANDBY\n");
1875                 set_bit(STANDBY, &con->state);
1876         } else {
1877                 /* retry after a delay. */
1878                 if (con->delay == 0)
1879                         con->delay = BASE_DELAY_INTERVAL;
1880                 else if (con->delay < MAX_DELAY_INTERVAL)
1881                         con->delay *= 2;
1882                 dout("fault queueing %p delay %lu\n", con, con->delay);
1883                 con->ops->get(con);
1884                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
1885                                        round_jiffies_relative(con->delay)) == 0)
1886                         con->ops->put(con);
1887         }
1888
1889 out_unlock:
1890         mutex_unlock(&con->mutex);
1891 out:
1892         /*
1893          * in case we faulted due to authentication, invalidate our
1894          * current tickets so that we can get new ones.
1895          */
1896         if (con->auth_retry && con->ops->invalidate_authorizer) {
1897                 dout("calling invalidate_authorizer()\n");
1898                 con->ops->invalidate_authorizer(con);
1899         }
1900
1901         if (con->ops->fault)
1902                 con->ops->fault(con);
1903 }
1904
1905
1906
1907 /*
1908  * create a new messenger instance
1909  */
1910 struct ceph_messenger *ceph_messenger_create(struct ceph_entity_addr *myaddr)
1911 {
1912         struct ceph_messenger *msgr;
1913
1914         msgr = kzalloc(sizeof(*msgr), GFP_KERNEL);
1915         if (msgr == NULL)
1916                 return ERR_PTR(-ENOMEM);
1917
1918         spin_lock_init(&msgr->global_seq_lock);
1919
1920         /* the zero page is needed if a request is "canceled" while the message
1921          * is being written over the socket */
1922         msgr->zero_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1923         if (!msgr->zero_page) {
1924                 kfree(msgr);
1925                 return ERR_PTR(-ENOMEM);
1926         }
1927         kmap(msgr->zero_page);
1928
1929         if (myaddr)
1930                 msgr->inst.addr = *myaddr;
1931
1932         /* select a random nonce */
1933         msgr->inst.addr.type = 0;
1934         get_random_bytes(&msgr->inst.addr.nonce, sizeof(msgr->inst.addr.nonce));
1935         encode_my_addr(msgr);
1936
1937         dout("messenger_create %p\n", msgr);
1938         return msgr;
1939 }
1940
1941 void ceph_messenger_destroy(struct ceph_messenger *msgr)
1942 {
1943         dout("destroy %p\n", msgr);
1944         kunmap(msgr->zero_page);
1945         __free_page(msgr->zero_page);
1946         kfree(msgr);
1947         dout("destroyed messenger %p\n", msgr);
1948 }
1949
1950 /*
1951  * Queue up an outgoing message on the given connection.
1952  */
1953 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
1954 {
1955         if (test_bit(CLOSED, &con->state)) {
1956                 dout("con_send %p closed, dropping %p\n", con, msg);
1957                 ceph_msg_put(msg);
1958                 return;
1959         }
1960
1961         /* set src+dst */
1962         msg->hdr.src.name = con->msgr->inst.name;
1963         msg->hdr.src.addr = con->msgr->my_enc_addr;
1964         msg->hdr.orig_src = msg->hdr.src;
1965
1966         BUG_ON(msg->front.iov_len != le32_to_cpu(msg->hdr.front_len));
1967
1968         /* queue */
1969         mutex_lock(&con->mutex);
1970         BUG_ON(!list_empty(&msg->list_head));
1971         list_add_tail(&msg->list_head, &con->out_queue);
1972         dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
1973              ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
1974              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1975              le32_to_cpu(msg->hdr.front_len),
1976              le32_to_cpu(msg->hdr.middle_len),
1977              le32_to_cpu(msg->hdr.data_len));
1978         mutex_unlock(&con->mutex);
1979
1980         /* if there wasn't anything waiting to send before, queue
1981          * new work */
1982         if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
1983                 queue_con(con);
1984 }
1985
1986 /*
1987  * Revoke a message that was previously queued for send
1988  */
1989 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
1990 {
1991         mutex_lock(&con->mutex);
1992         if (!list_empty(&msg->list_head)) {
1993                 dout("con_revoke %p msg %p\n", con, msg);
1994                 list_del_init(&msg->list_head);
1995                 ceph_msg_put(msg);
1996                 msg->hdr.seq = 0;
1997                 if (con->out_msg == msg) {
1998                         ceph_msg_put(con->out_msg);
1999                         con->out_msg = NULL;
2000                 }
2001                 if (con->out_kvec_is_msg) {
2002                         con->out_skip = con->out_kvec_bytes;
2003                         con->out_kvec_is_msg = false;
2004                 }
2005         } else {
2006                 dout("con_revoke %p msg %p - not queued (sent?)\n", con, msg);
2007         }
2008         mutex_unlock(&con->mutex);
2009 }
2010
2011 /*
2012  * Revoke a message that we may be reading data into
2013  */
2014 void ceph_con_revoke_message(struct ceph_connection *con, struct ceph_msg *msg)
2015 {
2016         mutex_lock(&con->mutex);
2017         if (con->in_msg && con->in_msg == msg) {
2018                 unsigned front_len = le32_to_cpu(con->in_hdr.front_len);
2019                 unsigned middle_len = le32_to_cpu(con->in_hdr.middle_len);
2020                 unsigned data_len = le32_to_cpu(con->in_hdr.data_len);
2021
2022                 /* skip rest of message */
2023                 dout("con_revoke_pages %p msg %p revoked\n", con, msg);
2024                         con->in_base_pos = con->in_base_pos -
2025                                 sizeof(struct ceph_msg_header) -
2026                                 front_len -
2027                                 middle_len -
2028                                 data_len -
2029                                 sizeof(struct ceph_msg_footer);
2030                 ceph_msg_put(con->in_msg);
2031                 con->in_msg = NULL;
2032                 con->in_tag = CEPH_MSGR_TAG_READY;
2033         } else {
2034                 dout("con_revoke_pages %p msg %p pages %p no-op\n",
2035                      con, con->in_msg, msg);
2036         }
2037         mutex_unlock(&con->mutex);
2038 }
2039
2040 /*
2041  * Queue a keepalive byte to ensure the tcp connection is alive.
2042  */
2043 void ceph_con_keepalive(struct ceph_connection *con)
2044 {
2045         if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
2046             test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2047                 queue_con(con);
2048 }
2049
2050
2051 /*
2052  * construct a new message with given type, size
2053  * the new msg has a ref count of 1.
2054  */
2055 struct ceph_msg *ceph_msg_new(int type, int front_len,
2056                               int page_len, int page_off, struct page **pages)
2057 {
2058         struct ceph_msg *m;
2059
2060         m = kmalloc(sizeof(*m), GFP_NOFS);
2061         if (m == NULL)
2062                 goto out;
2063         kref_init(&m->kref);
2064         INIT_LIST_HEAD(&m->list_head);
2065
2066         m->hdr.type = cpu_to_le16(type);
2067         m->hdr.front_len = cpu_to_le32(front_len);
2068         m->hdr.middle_len = 0;
2069         m->hdr.data_len = cpu_to_le32(page_len);
2070         m->hdr.data_off = cpu_to_le16(page_off);
2071         m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
2072         m->footer.front_crc = 0;
2073         m->footer.middle_crc = 0;
2074         m->footer.data_crc = 0;
2075         m->front_max = front_len;
2076         m->front_is_vmalloc = false;
2077         m->more_to_follow = false;
2078         m->pool = NULL;
2079
2080         /* front */
2081         if (front_len) {
2082                 if (front_len > PAGE_CACHE_SIZE) {
2083                         m->front.iov_base = __vmalloc(front_len, GFP_NOFS,
2084                                                       PAGE_KERNEL);
2085                         m->front_is_vmalloc = true;
2086                 } else {
2087                         m->front.iov_base = kmalloc(front_len, GFP_NOFS);
2088                 }
2089                 if (m->front.iov_base == NULL) {
2090                         pr_err("msg_new can't allocate %d bytes\n",
2091                              front_len);
2092                         goto out2;
2093                 }
2094         } else {
2095                 m->front.iov_base = NULL;
2096         }
2097         m->front.iov_len = front_len;
2098
2099         /* middle */
2100         m->middle = NULL;
2101
2102         /* data */
2103         m->nr_pages = calc_pages_for(page_off, page_len);
2104         m->pages = pages;
2105         m->pagelist = NULL;
2106
2107         dout("ceph_msg_new %p page %d~%d -> %d\n", m, page_off, page_len,
2108              m->nr_pages);
2109         return m;
2110
2111 out2:
2112         ceph_msg_put(m);
2113 out:
2114         pr_err("msg_new can't create type %d len %d\n", type, front_len);
2115         return ERR_PTR(-ENOMEM);
2116 }
2117
2118 /*
2119  * Allocate "middle" portion of a message, if it is needed and wasn't
2120  * allocated by alloc_msg.  This allows us to read a small fixed-size
2121  * per-type header in the front and then gracefully fail (i.e.,
2122  * propagate the error to the caller based on info in the front) when
2123  * the middle is too large.
2124  */
2125 static int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2126 {
2127         int type = le16_to_cpu(msg->hdr.type);
2128         int middle_len = le32_to_cpu(msg->hdr.middle_len);
2129
2130         dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2131              ceph_msg_type_name(type), middle_len);
2132         BUG_ON(!middle_len);
2133         BUG_ON(msg->middle);
2134
2135         msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2136         if (!msg->middle)
2137                 return -ENOMEM;
2138         return 0;
2139 }
2140
2141 /*
2142  * Generic message allocator, for incoming messages.
2143  */
2144 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2145                                 struct ceph_msg_header *hdr,
2146                                 int *skip)
2147 {
2148         int type = le16_to_cpu(hdr->type);
2149         int front_len = le32_to_cpu(hdr->front_len);
2150         int middle_len = le32_to_cpu(hdr->middle_len);
2151         struct ceph_msg *msg = NULL;
2152         int ret;
2153
2154         if (con->ops->alloc_msg) {
2155                 mutex_unlock(&con->mutex);
2156                 msg = con->ops->alloc_msg(con, hdr, skip);
2157                 mutex_lock(&con->mutex);
2158                 if (IS_ERR(msg))
2159                         return msg;
2160
2161                 if (*skip)
2162                         return NULL;
2163         }
2164         if (!msg) {
2165                 *skip = 0;
2166                 msg = ceph_msg_new(type, front_len, 0, 0, NULL);
2167                 if (!msg) {
2168                         pr_err("unable to allocate msg type %d len %d\n",
2169                                type, front_len);
2170                         return ERR_PTR(-ENOMEM);
2171                 }
2172         }
2173         memcpy(&msg->hdr, &con->in_hdr, sizeof(con->in_hdr));
2174
2175         if (middle_len) {
2176                 ret = ceph_alloc_middle(con, msg);
2177
2178                 if (ret < 0) {
2179                         ceph_msg_put(msg);
2180                         return msg;
2181                 }
2182         }
2183
2184         return msg;
2185 }
2186
2187
2188 /*
2189  * Free a generically kmalloc'd message.
2190  */
2191 void ceph_msg_kfree(struct ceph_msg *m)
2192 {
2193         dout("msg_kfree %p\n", m);
2194         if (m->front_is_vmalloc)
2195                 vfree(m->front.iov_base);
2196         else
2197                 kfree(m->front.iov_base);
2198         kfree(m);
2199 }
2200
2201 /*
2202  * Drop a msg ref.  Destroy as needed.
2203  */
2204 void ceph_msg_last_put(struct kref *kref)
2205 {
2206         struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2207
2208         dout("ceph_msg_put last one on %p\n", m);
2209         WARN_ON(!list_empty(&m->list_head));
2210
2211         /* drop middle, data, if any */
2212         if (m->middle) {
2213                 ceph_buffer_put(m->middle);
2214                 m->middle = NULL;
2215         }
2216         m->nr_pages = 0;
2217         m->pages = NULL;
2218
2219         if (m->pagelist) {
2220                 ceph_pagelist_release(m->pagelist);
2221                 kfree(m->pagelist);
2222                 m->pagelist = NULL;
2223         }
2224
2225         if (m->pool)
2226                 ceph_msgpool_put(m->pool, m);
2227         else
2228                 ceph_msg_kfree(m);
2229 }
2230
2231 void ceph_msg_dump(struct ceph_msg *msg)
2232 {
2233         pr_debug("msg_dump %p (front_max %d nr_pages %d)\n", msg,
2234                  msg->front_max, msg->nr_pages);
2235         print_hex_dump(KERN_DEBUG, "header: ",
2236                        DUMP_PREFIX_OFFSET, 16, 1,
2237                        &msg->hdr, sizeof(msg->hdr), true);
2238         print_hex_dump(KERN_DEBUG, " front: ",
2239                        DUMP_PREFIX_OFFSET, 16, 1,
2240                        msg->front.iov_base, msg->front.iov_len, true);
2241         if (msg->middle)
2242                 print_hex_dump(KERN_DEBUG, "middle: ",
2243                                DUMP_PREFIX_OFFSET, 16, 1,
2244                                msg->middle->vec.iov_base,
2245                                msg->middle->vec.iov_len, true);
2246         print_hex_dump(KERN_DEBUG, "footer: ",
2247                        DUMP_PREFIX_OFFSET, 16, 1,
2248                        &msg->footer, sizeof(msg->footer), true);
2249 }