SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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->out_keepalive_pending = false;
 344         con->in_seq = 0;
 345         con->in_seq_acked = 0;
 346 }
 347
 348 /*
 349  * mark a peer down.  drop any open connections.
 350  */
 351 void ceph_con_close(struct ceph_connection *con)
 352 {
 353         dout("con_close %p peer %s\n", con, pr_addr(&con->peer_addr.in_addr));
 354         set_bit(CLOSED, &con->state);  /* in case there's queued work */
 355         clear_bit(STANDBY, &con->state);  /* avoid connect_seq bump */
 356         clear_bit(LOSSYTX, &con->state);  /* so we retry next connect */
 357         clear_bit(KEEPALIVE_PENDING, &con->state);
 358         clear_bit(WRITE_PENDING, &con->state);
 359         mutex_lock(&con->mutex);
 360         reset_connection(con);
 361         con->peer_global_seq = 0;
 362         cancel_delayed_work(&con->work);
 363         mutex_unlock(&con->mutex);
 364         queue_con(con);
 365 }
 366
 367 /*
 368  * Reopen a closed connection, with a new peer address.
 369  */
 370 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
 371 {
 372         dout("con_open %p %s\n", con, pr_addr(&addr->in_addr));
 373         set_bit(OPENING, &con->state);
 374         clear_bit(CLOSED, &con->state);
 375         memcpy(&con->peer_addr, addr, sizeof(*addr));
 376         con->delay = 0;      /* reset backoff memory */
 377         queue_con(con);
 378 }
 379
 380 /*
 381  * return true if this connection ever successfully opened
 382  */
 383 bool ceph_con_opened(struct ceph_connection *con)
 384 {
 385         return con->connect_seq > 0;
 386 }
 387
 388 /*
 389  * generic get/put
 390  */
 391 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
 392 {
 393         dout("con_get %p nref = %d -> %d\n", con,
 394              atomic_read(&con->nref), atomic_read(&con->nref) + 1);
 395         if (atomic_inc_not_zero(&con->nref))
 396                 return con;
 397         return NULL;
 398 }
 399
 400 void ceph_con_put(struct ceph_connection *con)
 401 {
 402         dout("con_put %p nref = %d -> %d\n", con,
 403              atomic_read(&con->nref), atomic_read(&con->nref) - 1);
 404         BUG_ON(atomic_read(&con->nref) == 0);
 405         if (atomic_dec_and_test(&con->nref)) {
 406                 BUG_ON(con->sock);
 407                 kfree(con);
 408         }
 409 }
 410
 411 /*
 412  * initialize a new connection.
 413  */
 414 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
 415 {
 416         dout("con_init %p\n", con);
 417         memset(con, 0, sizeof(*con));
 418         atomic_set(&con->nref, 1);
 419         con->msgr = msgr;
 420         mutex_init(&con->mutex);
 421         INIT_LIST_HEAD(&con->out_queue);
 422         INIT_LIST_HEAD(&con->out_sent);
 423         INIT_DELAYED_WORK(&con->work, con_work);
 424 }
 425
 426
 427 /*
 428  * We maintain a global counter to order connection attempts.  Get
 429  * a unique seq greater than @gt.
 430  */
 431 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
 432 {
 433         u32 ret;
 434
 435         spin_lock(&msgr->global_seq_lock);
 436         if (msgr->global_seq < gt)
 437                 msgr->global_seq = gt;
 438         ret = ++msgr->global_seq;
 439         spin_unlock(&msgr->global_seq_lock);
 440         return ret;
 441 }
 442
 443
 444 /*
 445  * Prepare footer for currently outgoing message, and finish things
 446  * off.  Assumes out_kvec* are already valid.. we just add on to the end.
 447  */
 448 static void prepare_write_message_footer(struct ceph_connection *con, int v)
 449 {
 450         struct ceph_msg *m = con->out_msg;
 451
 452         dout("prepare_write_message_footer %p\n", con);
 453         con->out_kvec_is_msg = true;
 454         con->out_kvec[v].iov_base = &m->footer;
 455         con->out_kvec[v].iov_len = sizeof(m->footer);
 456         con->out_kvec_bytes += sizeof(m->footer);
 457         con->out_kvec_left++;
 458         con->out_more = m->more_to_follow;
 459         con->out_msg_done = true;
 460 }
 461
 462 /*
 463  * Prepare headers for the next outgoing message.
 464  */
 465 static void prepare_write_message(struct ceph_connection *con)
 466 {
 467         struct ceph_msg *m;
 468         int v = 0;
 469
 470         con->out_kvec_bytes = 0;
 471         con->out_kvec_is_msg = true;
 472         con->out_msg_done = false;
 473
 474         /* Sneak an ack in there first?  If we can get it into the same
 475          * TCP packet that's a good thing. */
 476         if (con->in_seq > con->in_seq_acked) {
 477                 con->in_seq_acked = con->in_seq;
 478                 con->out_kvec[v].iov_base = &tag_ack;
 479                 con->out_kvec[v++].iov_len = 1;
 480                 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 481                 con->out_kvec[v].iov_base = &con->out_temp_ack;
 482                 con->out_kvec[v++].iov_len = sizeof(con->out_temp_ack);
 483                 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 484         }
 485
 486         m = list_first_entry(&con->out_queue,
 487                        struct ceph_msg, list_head);
 488         con->out_msg = m;
 489         if (test_bit(LOSSYTX, &con->state)) {
 490                 list_del_init(&m->list_head);
 491         } else {
 492                 /* put message on sent list */
 493                 ceph_msg_get(m);
 494                 list_move_tail(&m->list_head, &con->out_sent);
 495         }
 496
 497         /*
 498          * only assign outgoing seq # if we haven't sent this message
 499          * yet.  if it is requeued, resend with it's original seq.
 500          */
 501         if (m->needs_out_seq) {
 502                 m->hdr.seq = cpu_to_le64(++con->out_seq);
 503                 m->needs_out_seq = false;
 504         }
 505
 506         dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
 507              m, con->out_seq, le16_to_cpu(m->hdr.type),
 508              le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
 509              le32_to_cpu(m->hdr.data_len),
 510              m->nr_pages);
 511         BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
 512
 513         /* tag + hdr + front + middle */
 514         con->out_kvec[v].iov_base = &tag_msg;
 515         con->out_kvec[v++].iov_len = 1;
 516         con->out_kvec[v].iov_base = &m->hdr;
 517         con->out_kvec[v++].iov_len = sizeof(m->hdr);
 518         con->out_kvec[v++] = m->front;
 519         if (m->middle)
 520                 con->out_kvec[v++] = m->middle->vec;
 521         con->out_kvec_left = v;
 522         con->out_kvec_bytes += 1 + sizeof(m->hdr) + m->front.iov_len +
 523                 (m->middle ? m->middle->vec.iov_len : 0);
 524         con->out_kvec_cur = con->out_kvec;
 525
 526         /* fill in crc (except data pages), footer */
 527         con->out_msg->hdr.crc =
 528                 cpu_to_le32(crc32c(0, (void *)&m->hdr,
 529                                       sizeof(m->hdr) - sizeof(m->hdr.crc)));
 530         con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
 531         con->out_msg->footer.front_crc =
 532                 cpu_to_le32(crc32c(0, m->front.iov_base, m->front.iov_len));
 533         if (m->middle)
 534                 con->out_msg->footer.middle_crc =
 535                         cpu_to_le32(crc32c(0, m->middle->vec.iov_base,
 536                                            m->middle->vec.iov_len));
 537         else
 538                 con->out_msg->footer.middle_crc = 0;
 539         con->out_msg->footer.data_crc = 0;
 540         dout("prepare_write_message front_crc %u data_crc %u\n",
 541              le32_to_cpu(con->out_msg->footer.front_crc),
 542              le32_to_cpu(con->out_msg->footer.middle_crc));
 543
 544         /* is there a data payload? */
 545         if (le32_to_cpu(m->hdr.data_len) > 0) {
 546                 /* initialize page iterator */
 547                 con->out_msg_pos.page = 0;
 548                 con->out_msg_pos.page_pos =
 549                         le16_to_cpu(m->hdr.data_off) & ~PAGE_MASK;
 550                 con->out_msg_pos.data_pos = 0;
 551                 con->out_msg_pos.did_page_crc = 0;
 552                 con->out_more = 1;  /* data + footer will follow */
 553         } else {
 554                 /* no, queue up footer too and be done */
 555                 prepare_write_message_footer(con, v);
 556         }
 557
 558         set_bit(WRITE_PENDING, &con->state);
 559 }
 560
 561 /*
 562  * Prepare an ack.
 563  */
 564 static void prepare_write_ack(struct ceph_connection *con)
 565 {
 566         dout("prepare_write_ack %p %llu -> %llu\n", con,
 567              con->in_seq_acked, con->in_seq);
 568         con->in_seq_acked = con->in_seq;
 569
 570         con->out_kvec[0].iov_base = &tag_ack;
 571         con->out_kvec[0].iov_len = 1;
 572         con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 573         con->out_kvec[1].iov_base = &con->out_temp_ack;
 574         con->out_kvec[1].iov_len = sizeof(con->out_temp_ack);
 575         con->out_kvec_left = 2;
 576         con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 577         con->out_kvec_cur = con->out_kvec;
 578         con->out_more = 1;  /* more will follow.. eventually.. */
 579         set_bit(WRITE_PENDING, &con->state);
 580 }
 581
 582 /*
 583  * Prepare to write keepalive byte.
 584  */
 585 static void prepare_write_keepalive(struct ceph_connection *con)
 586 {
 587         dout("prepare_write_keepalive %p\n", con);
 588         con->out_kvec[0].iov_base = &tag_keepalive;
 589         con->out_kvec[0].iov_len = 1;
 590         con->out_kvec_left = 1;
 591         con->out_kvec_bytes = 1;
 592         con->out_kvec_cur = con->out_kvec;
 593         set_bit(WRITE_PENDING, &con->state);
 594 }
 595
 596 /*
 597  * Connection negotiation.
 598  */
 599
 600 static void prepare_connect_authorizer(struct ceph_connection *con)
 601 {
 602         void *auth_buf;
 603         int auth_len = 0;
 604         int auth_protocol = 0;
 605
 606         mutex_unlock(&con->mutex);
 607         if (con->ops->get_authorizer)
 608                 con->ops->get_authorizer(con, &auth_buf, &auth_len,
 609                                          &auth_protocol, &con->auth_reply_buf,
 610                                          &con->auth_reply_buf_len,
 611                                          con->auth_retry);
 612         mutex_lock(&con->mutex);
 613
 614         con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
 615         con->out_connect.authorizer_len = cpu_to_le32(auth_len);
 616
 617         con->out_kvec[con->out_kvec_left].iov_base = auth_buf;
 618         con->out_kvec[con->out_kvec_left].iov_len = auth_len;
 619         con->out_kvec_left++;
 620         con->out_kvec_bytes += auth_len;
 621 }
 622
 623 /*
 624  * We connected to a peer and are saying hello.
 625  */
 626 static void prepare_write_banner(struct ceph_messenger *msgr,
 627                                  struct ceph_connection *con)
 628 {
 629         int len = strlen(CEPH_BANNER);
 630
 631         con->out_kvec[0].iov_base = CEPH_BANNER;
 632         con->out_kvec[0].iov_len = len;
 633         con->out_kvec[1].iov_base = &msgr->my_enc_addr;
 634         con->out_kvec[1].iov_len = sizeof(msgr->my_enc_addr);
 635         con->out_kvec_left = 2;
 636         con->out_kvec_bytes = len + sizeof(msgr->my_enc_addr);
 637         con->out_kvec_cur = con->out_kvec;
 638         con->out_more = 0;
 639         set_bit(WRITE_PENDING, &con->state);
 640 }
 641
 642 static void prepare_write_connect(struct ceph_messenger *msgr,
 643                                   struct ceph_connection *con,
 644                                   int after_banner)
 645 {
 646         unsigned global_seq = get_global_seq(con->msgr, 0);
 647         int proto;
 648
 649         switch (con->peer_name.type) {
 650         case CEPH_ENTITY_TYPE_MON:
 651                 proto = CEPH_MONC_PROTOCOL;
 652                 break;
 653         case CEPH_ENTITY_TYPE_OSD:
 654                 proto = CEPH_OSDC_PROTOCOL;
 655                 break;
 656         case CEPH_ENTITY_TYPE_MDS:
 657                 proto = CEPH_MDSC_PROTOCOL;
 658                 break;
 659         default:
 660                 BUG();
 661         }
 662
 663         dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
 664              con->connect_seq, global_seq, proto);
 665
 666         con->out_connect.features = CEPH_FEATURE_SUPPORTED_CLIENT;
 667         con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
 668         con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
 669         con->out_connect.global_seq = cpu_to_le32(global_seq);
 670         con->out_connect.protocol_version = cpu_to_le32(proto);
 671         con->out_connect.flags = 0;
 672
 673         if (!after_banner) {
 674                 con->out_kvec_left = 0;
 675                 con->out_kvec_bytes = 0;
 676         }
 677         con->out_kvec[con->out_kvec_left].iov_base = &con->out_connect;
 678         con->out_kvec[con->out_kvec_left].iov_len = sizeof(con->out_connect);
 679         con->out_kvec_left++;
 680         con->out_kvec_bytes += sizeof(con->out_connect);
 681         con->out_kvec_cur = con->out_kvec;
 682         con->out_more = 0;
 683         set_bit(WRITE_PENDING, &con->state);
 684
 685         prepare_connect_authorizer(con);
 686 }
 687
 688
 689 /*
 690  * write as much of pending kvecs to the socket as we can.
 691  *  1 -> done
 692  *  0 -> socket full, but more to do
 693  * <0 -> error
 694  */
 695 static int write_partial_kvec(struct ceph_connection *con)
 696 {
 697         int ret;
 698
 699         dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
 700         while (con->out_kvec_bytes > 0) {
 701                 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
 702                                        con->out_kvec_left, con->out_kvec_bytes,
 703                                        con->out_more);
 704                 if (ret <= 0)
 705                         goto out;
 706                 con->out_kvec_bytes -= ret;
 707                 if (con->out_kvec_bytes == 0)
 708                         break;            /* done */
 709                 while (ret > 0) {
 710                         if (ret >= con->out_kvec_cur->iov_len) {
 711                                 ret -= con->out_kvec_cur->iov_len;
 712                                 con->out_kvec_cur++;
 713                                 con->out_kvec_left--;
 714                         } else {
 715                                 con->out_kvec_cur->iov_len -= ret;
 716                                 con->out_kvec_cur->iov_base += ret;
 717                                 ret = 0;
 718                                 break;
 719                         }
 720                 }
 721         }
 722         con->out_kvec_left = 0;
 723         con->out_kvec_is_msg = false;
 724         ret = 1;
 725 out:
 726         dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
 727              con->out_kvec_bytes, con->out_kvec_left, ret);
 728         return ret;  /* done! */
 729 }
 730
 731 /*
 732  * Write as much message data payload as we can.  If we finish, queue
 733  * up the footer.
 734  *  1 -> done, footer is now queued in out_kvec[].
 735  *  0 -> socket full, but more to do
 736  * <0 -> error
 737  */
 738 static int write_partial_msg_pages(struct ceph_connection *con)
 739 {
 740         struct ceph_msg *msg = con->out_msg;
 741         unsigned data_len = le32_to_cpu(msg->hdr.data_len);
 742         size_t len;
 743         int crc = con->msgr->nocrc;
 744         int ret;
 745
 746         dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
 747              con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
 748              con->out_msg_pos.page_pos);
 749
 750         while (con->out_msg_pos.page < con->out_msg->nr_pages) {
 751                 struct page *page = NULL;
 752                 void *kaddr = NULL;
 753
 754                 /*
 755                  * if we are calculating the data crc (the default), we need
 756                  * to map the page.  if our pages[] has been revoked, use the
 757                  * zero page.
 758                  */
 759                 if (msg->pages) {
 760                         page = msg->pages[con->out_msg_pos.page];
 761                         if (crc)
 762                                 kaddr = kmap(page);
 763                 } else if (msg->pagelist) {
 764                         page = list_first_entry(&msg->pagelist->head,
 765                                                 struct page, lru);
 766                         if (crc)
 767                                 kaddr = kmap(page);
 768                 } else {
 769                         page = con->msgr->zero_page;
 770                         if (crc)
 771                                 kaddr = page_address(con->msgr->zero_page);
 772                 }
 773                 len = min((int)(PAGE_SIZE - con->out_msg_pos.page_pos),
 774                           (int)(data_len - con->out_msg_pos.data_pos));
 775                 if (crc && !con->out_msg_pos.did_page_crc) {
 776                         void *base = kaddr + con->out_msg_pos.page_pos;
 777                         u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
 778
 779                         BUG_ON(kaddr == NULL);
 780                         con->out_msg->footer.data_crc =
 781                                 cpu_to_le32(crc32c(tmpcrc, base, len));
 782                         con->out_msg_pos.did_page_crc = 1;
 783                 }
 784
 785                 ret = kernel_sendpage(con->sock, page,
 786                                       con->out_msg_pos.page_pos, len,
 787                                       MSG_DONTWAIT | MSG_NOSIGNAL |
 788                                       MSG_MORE);
 789
 790                 if (crc && (msg->pages || msg->pagelist))
 791                         kunmap(page);
 792
 793                 if (ret <= 0)
 794                         goto out;
 795
 796                 con->out_msg_pos.data_pos += ret;
 797                 con->out_msg_pos.page_pos += ret;
 798                 if (ret == len) {
 799                         con->out_msg_pos.page_pos = 0;
 800                         con->out_msg_pos.page++;
 801                         con->out_msg_pos.did_page_crc = 0;
 802                         if (msg->pagelist)
 803                                 list_move_tail(&page->lru,
 804                                                &msg->pagelist->head);
 805                 }
 806         }
 807
 808         dout("write_partial_msg_pages %p msg %p done\n", con, msg);
 809
 810         /* prepare and queue up footer, too */
 811         if (!crc)
 812                 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
 813         con->out_kvec_bytes = 0;
 814         con->out_kvec_left = 0;
 815         con->out_kvec_cur = con->out_kvec;
 816         prepare_write_message_footer(con, 0);
 817         ret = 1;
 818 out:
 819         return ret;
 820 }
 821
 822 /*
 823  * write some zeros
 824  */
 825 static int write_partial_skip(struct ceph_connection *con)
 826 {
 827         int ret;
 828
 829         while (con->out_skip > 0) {
 830                 struct kvec iov = {
 831                         .iov_base = page_address(con->msgr->zero_page),
 832                         .iov_len = min(con->out_skip, (int)PAGE_CACHE_SIZE)
 833                 };
 834
 835                 ret = ceph_tcp_sendmsg(con->sock, &iov, 1, iov.iov_len, 1);
 836                 if (ret <= 0)
 837                         goto out;
 838                 con->out_skip -= ret;
 839         }
 840         ret = 1;
 841 out:
 842         return ret;
 843 }
 844
 845 /*
 846  * Prepare to read connection handshake, or an ack.
 847  */
 848 static void prepare_read_banner(struct ceph_connection *con)
 849 {
 850         dout("prepare_read_banner %p\n", con);
 851         con->in_base_pos = 0;
 852 }
 853
 854 static void prepare_read_connect(struct ceph_connection *con)
 855 {
 856         dout("prepare_read_connect %p\n", con);
 857         con->in_base_pos = 0;
 858 }
 859
 860 static void prepare_read_ack(struct ceph_connection *con)
 861 {
 862         dout("prepare_read_ack %p\n", con);
 863         con->in_base_pos = 0;
 864 }
 865
 866 static void prepare_read_tag(struct ceph_connection *con)
 867 {
 868         dout("prepare_read_tag %p\n", con);
 869         con->in_base_pos = 0;
 870         con->in_tag = CEPH_MSGR_TAG_READY;
 871 }
 872
 873 /*
 874  * Prepare to read a message.
 875  */
 876 static int prepare_read_message(struct ceph_connection *con)
 877 {
 878         dout("prepare_read_message %p\n", con);
 879         BUG_ON(con->in_msg != NULL);
 880         con->in_base_pos = 0;
 881         con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
 882         return 0;
 883 }
 884
 885
 886 static int read_partial(struct ceph_connection *con,
 887                         int *to, int size, void *object)
 888 {
 889         *to += size;
 890         while (con->in_base_pos < *to) {
 891                 int left = *to - con->in_base_pos;
 892                 int have = size - left;
 893                 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
 894                 if (ret <= 0)
 895                         return ret;
 896                 con->in_base_pos += ret;
 897         }
 898         return 1;
 899 }
 900
 901
 902 /*
 903  * Read all or part of the connect-side handshake on a new connection
 904  */
 905 static int read_partial_banner(struct ceph_connection *con)
 906 {
 907         int ret, to = 0;
 908
 909         dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
 910
 911         /* peer's banner */
 912         ret = read_partial(con, &to, strlen(CEPH_BANNER), con->in_banner);
 913         if (ret <= 0)
 914                 goto out;
 915         ret = read_partial(con, &to, sizeof(con->actual_peer_addr),
 916                            &con->actual_peer_addr);
 917         if (ret <= 0)
 918                 goto out;
 919         ret = read_partial(con, &to, sizeof(con->peer_addr_for_me),
 920                            &con->peer_addr_for_me);
 921         if (ret <= 0)
 922                 goto out;
 923 out:
 924         return ret;
 925 }
 926
 927 static int read_partial_connect(struct ceph_connection *con)
 928 {
 929         int ret, to = 0;
 930
 931         dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
 932
 933         ret = read_partial(con, &to, sizeof(con->in_reply), &con->in_reply);
 934         if (ret <= 0)
 935                 goto out;
 936         ret = read_partial(con, &to, le32_to_cpu(con->in_reply.authorizer_len),
 937                            con->auth_reply_buf);
 938         if (ret <= 0)
 939                 goto out;
 940
 941         dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
 942              con, (int)con->in_reply.tag,
 943              le32_to_cpu(con->in_reply.connect_seq),
 944              le32_to_cpu(con->in_reply.global_seq));
 945 out:
 946         return ret;
 947
 948 }
 949
 950 /*
 951  * Verify the hello banner looks okay.
 952  */
 953 static int verify_hello(struct ceph_connection *con)
 954 {
 955         if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
 956                 pr_err("connect to %s got bad banner\n",
 957                        pr_addr(&con->peer_addr.in_addr));
 958                 con->error_msg = "protocol error, bad banner";
 959                 return -1;
 960         }
 961         return 0;
 962 }
 963
 964 static bool addr_is_blank(struct sockaddr_storage *ss)
 965 {
 966         switch (ss->ss_family) {
 967         case AF_INET:
 968                 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
 969         case AF_INET6:
 970                 return
 971                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
 972                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
 973                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
 974                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
 975         }
 976         return false;
 977 }
 978
 979 static int addr_port(struct sockaddr_storage *ss)
 980 {
 981         switch (ss->ss_family) {
 982         case AF_INET:
 983                 return ntohs(((struct sockaddr_in *)ss)->sin_port);
 984         case AF_INET6:
 985                 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
 986         }
 987         return 0;
 988 }
 989
 990 static void addr_set_port(struct sockaddr_storage *ss, int p)
 991 {
 992         switch (ss->ss_family) {
 993         case AF_INET:
 994                 ((struct sockaddr_in *)ss)->sin_port = htons(p);
 995         case AF_INET6:
 996                 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
 997         }
 998 }
 999
1000 /*
1001  * Parse an ip[:port] list into an addr array.  Use the default
1002  * monitor port if a port isn't specified.
1003  */
1004 int ceph_parse_ips(const char *c, const char *end,
1005                    struct ceph_entity_addr *addr,
1006                    int max_count, int *count)
1007 {
1008         int i;
1009         const char *p = c;
1010
1011         dout("parse_ips on '%.*s'\n", (int)(end-c), c);
1012         for (i = 0; i < max_count; i++) {
1013                 const char *ipend;
1014                 struct sockaddr_storage *ss = &addr[i].in_addr;
1015                 struct sockaddr_in *in4 = (void *)ss;
1016                 struct sockaddr_in6 *in6 = (void *)ss;
1017                 int port;
1018
1019                 memset(ss, 0, sizeof(*ss));
1020                 if (in4_pton(p, end - p, (u8 *)&in4->sin_addr.s_addr,
1021                              ',', &ipend)) {
1022                         ss->ss_family = AF_INET;
1023                 } else if (in6_pton(p, end - p, (u8 *)&in6->sin6_addr.s6_addr,
1024                                     ',', &ipend)) {
1025                         ss->ss_family = AF_INET6;
1026                 } else {
1027                         goto bad;
1028                 }
1029                 p = ipend;
1030
1031                 /* port? */
1032                 if (p < end && *p == ':') {
1033                         port = 0;
1034                         p++;
1035                         while (p < end && *p >= '0' && *p <= '9') {
1036                                 port = (port * 10) + (*p - '0');
1037                                 p++;
1038                         }
1039                         if (port > 65535 || port == 0)
1040                                 goto bad;
1041                 } else {
1042                         port = CEPH_MON_PORT;
1043                 }
1044
1045                 addr_set_port(ss, port);
1046
1047                 dout("parse_ips got %s\n", pr_addr(ss));
1048
1049                 if (p == end)
1050                         break;
1051                 if (*p != ',')
1052                         goto bad;
1053                 p++;
1054         }
1055
1056         if (p != end)
1057                 goto bad;
1058
1059         if (count)
1060                 *count = i + 1;
1061         return 0;
1062
1063 bad:
1064         pr_err("parse_ips bad ip '%s'\n", c);
1065         return -EINVAL;
1066 }
1067
1068 static int process_banner(struct ceph_connection *con)
1069 {
1070         dout("process_banner on %p\n", con);
1071
1072         if (verify_hello(con) < 0)
1073                 return -1;
1074
1075         ceph_decode_addr(&con->actual_peer_addr);
1076         ceph_decode_addr(&con->peer_addr_for_me);
1077
1078         /*
1079          * Make sure the other end is who we wanted.  note that the other
1080          * end may not yet know their ip address, so if it's 0.0.0.0, give
1081          * them the benefit of the doubt.
1082          */
1083         if (memcmp(&con->peer_addr, &con->actual_peer_addr,
1084                    sizeof(con->peer_addr)) != 0 &&
1085             !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1086               con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1087                 pr_warning("wrong peer, want %s/%lld, got %s/%lld\n",
1088                            pr_addr(&con->peer_addr.in_addr),
1089                            le64_to_cpu(con->peer_addr.nonce),
1090                            pr_addr(&con->actual_peer_addr.in_addr),
1091                            le64_to_cpu(con->actual_peer_addr.nonce));
1092                 con->error_msg = "wrong peer at address";
1093                 return -1;
1094         }
1095
1096         /*
1097          * did we learn our address?
1098          */
1099         if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1100                 int port = addr_port(&con->msgr->inst.addr.in_addr);
1101
1102                 memcpy(&con->msgr->inst.addr.in_addr,
1103                        &con->peer_addr_for_me.in_addr,
1104                        sizeof(con->peer_addr_for_me.in_addr));
1105                 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1106                 encode_my_addr(con->msgr);
1107                 dout("process_banner learned my addr is %s\n",
1108                      pr_addr(&con->msgr->inst.addr.in_addr));
1109         }
1110
1111         set_bit(NEGOTIATING, &con->state);
1112         prepare_read_connect(con);
1113         return 0;
1114 }
1115
1116 static void fail_protocol(struct ceph_connection *con)
1117 {
1118         reset_connection(con);
1119         set_bit(CLOSED, &con->state);  /* in case there's queued work */
1120
1121         mutex_unlock(&con->mutex);
1122         if (con->ops->bad_proto)
1123                 con->ops->bad_proto(con);
1124         mutex_lock(&con->mutex);
1125 }
1126
1127 static int process_connect(struct ceph_connection *con)
1128 {
1129         u64 sup_feat = CEPH_FEATURE_SUPPORTED_CLIENT;
1130         u64 req_feat = CEPH_FEATURE_REQUIRED_CLIENT;
1131         u64 server_feat = le64_to_cpu(con->in_reply.features);
1132
1133         dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1134
1135         switch (con->in_reply.tag) {
1136         case CEPH_MSGR_TAG_FEATURES:
1137                 pr_err("%s%lld %s feature set mismatch,"
1138                        " my %llx < server's %llx, missing %llx\n",
1139                        ENTITY_NAME(con->peer_name),
1140                        pr_addr(&con->peer_addr.in_addr),
1141                        sup_feat, server_feat, server_feat & ~sup_feat);
1142                 con->error_msg = "missing required protocol features";
1143                 fail_protocol(con);
1144                 return -1;
1145
1146         case CEPH_MSGR_TAG_BADPROTOVER:
1147                 pr_err("%s%lld %s protocol version mismatch,"
1148                        " my %d != server's %d\n",
1149                        ENTITY_NAME(con->peer_name),
1150                        pr_addr(&con->peer_addr.in_addr),
1151                        le32_to_cpu(con->out_connect.protocol_version),
1152                        le32_to_cpu(con->in_reply.protocol_version));
1153                 con->error_msg = "protocol version mismatch";
1154                 fail_protocol(con);
1155                 return -1;
1156
1157         case CEPH_MSGR_TAG_BADAUTHORIZER:
1158                 con->auth_retry++;
1159                 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1160                      con->auth_retry);
1161                 if (con->auth_retry == 2) {
1162                         con->error_msg = "connect authorization failure";
1163                         reset_connection(con);
1164                         set_bit(CLOSED, &con->state);
1165                         return -1;
1166                 }
1167                 con->auth_retry = 1;
1168                 prepare_write_connect(con->msgr, con, 0);
1169                 prepare_read_connect(con);
1170                 break;
1171
1172         case CEPH_MSGR_TAG_RESETSESSION:
1173                 /*
1174                  * If we connected with a large connect_seq but the peer
1175                  * has no record of a session with us (no connection, or
1176                  * connect_seq == 0), they will send RESETSESION to indicate
1177                  * that they must have reset their session, and may have
1178                  * dropped messages.
1179                  */
1180                 dout("process_connect got RESET peer seq %u\n",
1181                      le32_to_cpu(con->in_connect.connect_seq));
1182                 pr_err("%s%lld %s connection reset\n",
1183                        ENTITY_NAME(con->peer_name),
1184                        pr_addr(&con->peer_addr.in_addr));
1185                 reset_connection(con);
1186                 prepare_write_connect(con->msgr, con, 0);
1187                 prepare_read_connect(con);
1188
1189                 /* Tell ceph about it. */
1190                 mutex_unlock(&con->mutex);
1191                 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1192                 if (con->ops->peer_reset)
1193                         con->ops->peer_reset(con);
1194                 mutex_lock(&con->mutex);
1195                 break;
1196
1197         case CEPH_MSGR_TAG_RETRY_SESSION:
1198                 /*
1199                  * If we sent a smaller connect_seq than the peer has, try
1200                  * again with a larger value.
1201                  */
1202                 dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
1203                      le32_to_cpu(con->out_connect.connect_seq),
1204                      le32_to_cpu(con->in_connect.connect_seq));
1205                 con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
1206                 prepare_write_connect(con->msgr, con, 0);
1207                 prepare_read_connect(con);
1208                 break;
1209
1210         case CEPH_MSGR_TAG_RETRY_GLOBAL:
1211                 /*
1212                  * If we sent a smaller global_seq than the peer has, try
1213                  * again with a larger value.
1214                  */
1215                 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1216                      con->peer_global_seq,
1217                      le32_to_cpu(con->in_connect.global_seq));
1218                 get_global_seq(con->msgr,
1219                                le32_to_cpu(con->in_connect.global_seq));
1220                 prepare_write_connect(con->msgr, con, 0);
1221                 prepare_read_connect(con);
1222                 break;
1223
1224         case CEPH_MSGR_TAG_READY:
1225                 if (req_feat & ~server_feat) {
1226                         pr_err("%s%lld %s protocol feature mismatch,"
1227                                " my required %llx > server's %llx, need %llx\n",
1228                                ENTITY_NAME(con->peer_name),
1229                                pr_addr(&con->peer_addr.in_addr),
1230                                req_feat, server_feat, req_feat & ~server_feat);
1231                         con->error_msg = "missing required protocol features";
1232                         fail_protocol(con);
1233                         return -1;
1234                 }
1235                 clear_bit(CONNECTING, &con->state);
1236                 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1237                 con->connect_seq++;
1238                 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1239                      con->peer_global_seq,
1240                      le32_to_cpu(con->in_reply.connect_seq),
1241                      con->connect_seq);
1242                 WARN_ON(con->connect_seq !=
1243                         le32_to_cpu(con->in_reply.connect_seq));
1244
1245                 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1246                         set_bit(LOSSYTX, &con->state);
1247
1248                 prepare_read_tag(con);
1249                 break;
1250
1251         case CEPH_MSGR_TAG_WAIT:
1252                 /*
1253                  * If there is a connection race (we are opening
1254                  * connections to each other), one of us may just have
1255                  * to WAIT.  This shouldn't happen if we are the
1256                  * client.
1257                  */
1258                 pr_err("process_connect peer connecting WAIT\n");
1259
1260         default:
1261                 pr_err("connect protocol error, will retry\n");
1262                 con->error_msg = "protocol error, garbage tag during connect";
1263                 return -1;
1264         }
1265         return 0;
1266 }
1267
1268
1269 /*
1270  * read (part of) an ack
1271  */
1272 static int read_partial_ack(struct ceph_connection *con)
1273 {
1274         int to = 0;
1275
1276         return read_partial(con, &to, sizeof(con->in_temp_ack),
1277                             &con->in_temp_ack);
1278 }
1279
1280
1281 /*
1282  * We can finally discard anything that's been acked.
1283  */
1284 static void process_ack(struct ceph_connection *con)
1285 {
1286         struct ceph_msg *m;
1287         u64 ack = le64_to_cpu(con->in_temp_ack);
1288         u64 seq;
1289
1290         while (!list_empty(&con->out_sent)) {
1291                 m = list_first_entry(&con->out_sent, struct ceph_msg,
1292                                      list_head);
1293                 seq = le64_to_cpu(m->hdr.seq);
1294                 if (seq > ack)
1295                         break;
1296                 dout("got ack for seq %llu type %d at %p\n", seq,
1297                      le16_to_cpu(m->hdr.type), m);
1298                 ceph_msg_remove(m);
1299         }
1300         prepare_read_tag(con);
1301 }
1302
1303
1304
1305
1306 static int read_partial_message_section(struct ceph_connection *con,
1307                                         struct kvec *section, unsigned int sec_len,
1308                                         u32 *crc)
1309 {
1310         int left;
1311         int ret;
1312
1313         BUG_ON(!section);
1314
1315         while (section->iov_len < sec_len) {
1316                 BUG_ON(section->iov_base == NULL);
1317                 left = sec_len - section->iov_len;
1318                 ret = ceph_tcp_recvmsg(con->sock, (char *)section->iov_base +
1319                                        section->iov_len, left);
1320                 if (ret <= 0)
1321                         return ret;
1322                 section->iov_len += ret;
1323                 if (section->iov_len == sec_len)
1324                         *crc = crc32c(0, section->iov_base,
1325                                       section->iov_len);
1326         }
1327
1328         return 1;
1329 }
1330
1331 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
1332                                 struct ceph_msg_header *hdr,
1333                                 int *skip);
1334 /*
1335  * read (part of) a message.
1336  */
1337 static int read_partial_message(struct ceph_connection *con)
1338 {
1339         struct ceph_msg *m = con->in_msg;
1340         void *p;
1341         int ret;
1342         int to, left;
1343         unsigned front_len, middle_len, data_len, data_off;
1344         int datacrc = con->msgr->nocrc;
1345         int skip;
1346         u64 seq;
1347
1348         dout("read_partial_message con %p msg %p\n", con, m);
1349
1350         /* header */
1351         while (con->in_base_pos < sizeof(con->in_hdr)) {
1352                 left = sizeof(con->in_hdr) - con->in_base_pos;
1353                 ret = ceph_tcp_recvmsg(con->sock,
1354                                        (char *)&con->in_hdr + con->in_base_pos,
1355                                        left);
1356                 if (ret <= 0)
1357                         return ret;
1358                 con->in_base_pos += ret;
1359                 if (con->in_base_pos == sizeof(con->in_hdr)) {
1360                         u32 crc = crc32c(0, (void *)&con->in_hdr,
1361                                  sizeof(con->in_hdr) - sizeof(con->in_hdr.crc));
1362                         if (crc != le32_to_cpu(con->in_hdr.crc)) {
1363                                 pr_err("read_partial_message bad hdr "
1364                                        " crc %u != expected %u\n",
1365                                        crc, con->in_hdr.crc);
1366                                 return -EBADMSG;
1367                         }
1368                 }
1369         }
1370         front_len = le32_to_cpu(con->in_hdr.front_len);
1371         if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1372                 return -EIO;
1373         middle_len = le32_to_cpu(con->in_hdr.middle_len);
1374         if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1375                 return -EIO;
1376         data_len = le32_to_cpu(con->in_hdr.data_len);
1377         if (data_len > CEPH_MSG_MAX_DATA_LEN)
1378                 return -EIO;
1379         data_off = le16_to_cpu(con->in_hdr.data_off);
1380
1381         /* verify seq# */
1382         seq = le64_to_cpu(con->in_hdr.seq);
1383         if ((s64)seq - (s64)con->in_seq < 1) {
1384                 pr_info("skipping %s%lld %s seq %lld, expected %lld\n",
1385                         ENTITY_NAME(con->peer_name),
1386                         pr_addr(&con->peer_addr.in_addr),
1387                         seq, con->in_seq + 1);
1388                 con->in_base_pos = -front_len - middle_len - data_len -
1389                         sizeof(m->footer);
1390                 con->in_tag = CEPH_MSGR_TAG_READY;
1391                 con->in_seq++;
1392                 return 0;
1393         } else if ((s64)seq - (s64)con->in_seq > 1) {
1394                 pr_err("read_partial_message bad seq %lld expected %lld\n",
1395                        seq, con->in_seq + 1);
1396                 con->error_msg = "bad message sequence # for incoming message";
1397                 return -EBADMSG;
1398         }
1399
1400         /* allocate message? */
1401         if (!con->in_msg) {
1402                 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1403                      con->in_hdr.front_len, con->in_hdr.data_len);
1404                 con->in_msg = ceph_alloc_msg(con, &con->in_hdr, &skip);
1405                 if (skip) {
1406                         /* skip this message */
1407                         dout("alloc_msg said skip message\n");
1408                         con->in_base_pos = -front_len - middle_len - data_len -
1409                                 sizeof(m->footer);
1410                         con->in_tag = CEPH_MSGR_TAG_READY;
1411                         con->in_seq++;
1412                         return 0;
1413                 }
1414                 if (!con->in_msg) {
1415                         con->error_msg =
1416                                 "error allocating memory for incoming message";
1417                         return -ENOMEM;
1418                 }
1419                 m = con->in_msg;
1420                 m->front.iov_len = 0;    /* haven't read it yet */
1421                 if (m->middle)
1422                         m->middle->vec.iov_len = 0;
1423
1424                 con->in_msg_pos.page = 0;
1425                 con->in_msg_pos.page_pos = data_off & ~PAGE_MASK;
1426                 con->in_msg_pos.data_pos = 0;
1427         }
1428
1429         /* front */
1430         ret = read_partial_message_section(con, &m->front, front_len,
1431                                            &con->in_front_crc);
1432         if (ret <= 0)
1433                 return ret;
1434
1435         /* middle */
1436         if (m->middle) {
1437                 ret = read_partial_message_section(con, &m->middle->vec, middle_len,
1438                                                    &con->in_middle_crc);
1439                 if (ret <= 0)
1440                         return ret;
1441         }
1442
1443         /* (page) data */
1444         while (con->in_msg_pos.data_pos < data_len) {
1445                 left = min((int)(data_len - con->in_msg_pos.data_pos),
1446                            (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1447                 BUG_ON(m->pages == NULL);
1448                 p = kmap(m->pages[con->in_msg_pos.page]);
1449                 ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1450                                        left);
1451                 if (ret > 0 && datacrc)
1452                         con->in_data_crc =
1453                                 crc32c(con->in_data_crc,
1454                                           p + con->in_msg_pos.page_pos, ret);
1455                 kunmap(m->pages[con->in_msg_pos.page]);
1456                 if (ret <= 0)
1457                         return ret;
1458                 con->in_msg_pos.data_pos += ret;
1459                 con->in_msg_pos.page_pos += ret;
1460                 if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1461                         con->in_msg_pos.page_pos = 0;
1462                         con->in_msg_pos.page++;
1463                 }
1464         }
1465
1466         /* footer */
1467         to = sizeof(m->hdr) + sizeof(m->footer);
1468         while (con->in_base_pos < to) {
1469                 left = to - con->in_base_pos;
1470                 ret = ceph_tcp_recvmsg(con->sock, (char *)&m->footer +
1471                                        (con->in_base_pos - sizeof(m->hdr)),
1472                                        left);
1473                 if (ret <= 0)
1474                         return ret;
1475                 con->in_base_pos += ret;
1476         }
1477         dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1478              m, front_len, m->footer.front_crc, middle_len,
1479              m->footer.middle_crc, data_len, m->footer.data_crc);
1480
1481         /* crc ok? */
1482         if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1483                 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1484                        m, con->in_front_crc, m->footer.front_crc);
1485                 return -EBADMSG;
1486         }
1487         if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1488                 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1489                        m, con->in_middle_crc, m->footer.middle_crc);
1490                 return -EBADMSG;
1491         }
1492         if (datacrc &&
1493             (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1494             con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1495                 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1496                        con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1497                 return -EBADMSG;
1498         }
1499
1500         return 1; /* done! */
1501 }
1502
1503 /*
1504  * Process message.  This happens in the worker thread.  The callback should
1505  * be careful not to do anything that waits on other incoming messages or it
1506  * may deadlock.
1507  */
1508 static void process_message(struct ceph_connection *con)
1509 {
1510         struct ceph_msg *msg;
1511
1512         msg = con->in_msg;
1513         con->in_msg = NULL;
1514
1515         /* if first message, set peer_name */
1516         if (con->peer_name.type == 0)
1517                 con->peer_name = msg->hdr.src;
1518
1519         con->in_seq++;
1520         mutex_unlock(&con->mutex);
1521
1522         dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1523              msg, le64_to_cpu(msg->hdr.seq),
1524              ENTITY_NAME(msg->hdr.src),
1525              le16_to_cpu(msg->hdr.type),
1526              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1527              le32_to_cpu(msg->hdr.front_len),
1528              le32_to_cpu(msg->hdr.data_len),
1529              con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1530         con->ops->dispatch(con, msg);
1531
1532         mutex_lock(&con->mutex);
1533         prepare_read_tag(con);
1534 }
1535
1536
1537 /*
1538  * Write something to the socket.  Called in a worker thread when the
1539  * socket appears to be writeable and we have something ready to send.
1540  */
1541 static int try_write(struct ceph_connection *con)
1542 {
1543         struct ceph_messenger *msgr = con->msgr;
1544         int ret = 1;
1545
1546         dout("try_write start %p state %lu nref %d\n", con, con->state,
1547              atomic_read(&con->nref));
1548
1549 more:
1550         dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1551
1552         /* open the socket first? */
1553         if (con->sock == NULL) {
1554                 /*
1555                  * if we were STANDBY and are reconnecting _this_
1556                  * connection, bump connect_seq now.  Always bump
1557                  * global_seq.
1558                  */
1559                 if (test_and_clear_bit(STANDBY, &con->state))
1560                         con->connect_seq++;
1561
1562                 prepare_write_banner(msgr, con);
1563                 prepare_write_connect(msgr, con, 1);
1564                 prepare_read_banner(con);
1565                 set_bit(CONNECTING, &con->state);
1566                 clear_bit(NEGOTIATING, &con->state);
1567
1568                 BUG_ON(con->in_msg);
1569                 con->in_tag = CEPH_MSGR_TAG_READY;
1570                 dout("try_write initiating connect on %p new state %lu\n",
1571                      con, con->state);
1572                 con->sock = ceph_tcp_connect(con);
1573                 if (IS_ERR(con->sock)) {
1574                         con->sock = NULL;
1575                         con->error_msg = "connect error";
1576                         ret = -1;
1577                         goto out;
1578                 }
1579         }
1580
1581 more_kvec:
1582         /* kvec data queued? */
1583         if (con->out_skip) {
1584                 ret = write_partial_skip(con);
1585                 if (ret <= 0)
1586                         goto done;
1587                 if (ret < 0) {
1588                         dout("try_write write_partial_skip err %d\n", ret);
1589                         goto done;
1590                 }
1591         }
1592         if (con->out_kvec_left) {
1593                 ret = write_partial_kvec(con);
1594                 if (ret <= 0)
1595                         goto done;
1596         }
1597
1598         /* msg pages? */
1599         if (con->out_msg) {
1600                 if (con->out_msg_done) {
1601                         ceph_msg_put(con->out_msg);
1602                         con->out_msg = NULL;   /* we're done with this one */
1603                         goto do_next;
1604                 }
1605
1606                 ret = write_partial_msg_pages(con);
1607                 if (ret == 1)
1608                         goto more_kvec;  /* we need to send the footer, too! */
1609                 if (ret == 0)
1610                         goto done;
1611                 if (ret < 0) {
1612                         dout("try_write write_partial_msg_pages err %d\n",
1613                              ret);
1614                         goto done;
1615                 }
1616         }
1617
1618 do_next:
1619         if (!test_bit(CONNECTING, &con->state)) {
1620                 /* is anything else pending? */
1621                 if (!list_empty(&con->out_queue)) {
1622                         prepare_write_message(con);
1623                         goto more;
1624                 }
1625                 if (con->in_seq > con->in_seq_acked) {
1626                         prepare_write_ack(con);
1627                         goto more;
1628                 }
1629                 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1630                         prepare_write_keepalive(con);
1631                         goto more;
1632                 }
1633         }
1634
1635         /* Nothing to do! */
1636         clear_bit(WRITE_PENDING, &con->state);
1637         dout("try_write nothing else to write.\n");
1638 done:
1639         ret = 0;
1640 out:
1641         dout("try_write done on %p\n", con);
1642         return ret;
1643 }
1644
1645
1646
1647 /*
1648  * Read what we can from the socket.
1649  */
1650 static int try_read(struct ceph_connection *con)
1651 {
1652         int ret = -1;
1653
1654         if (!con->sock)
1655                 return 0;
1656
1657         if (test_bit(STANDBY, &con->state))
1658                 return 0;
1659
1660         dout("try_read start on %p\n", con);
1661
1662 more:
1663         dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1664              con->in_base_pos);
1665         if (test_bit(CONNECTING, &con->state)) {
1666                 if (!test_bit(NEGOTIATING, &con->state)) {
1667                         dout("try_read connecting\n");
1668                         ret = read_partial_banner(con);
1669                         if (ret <= 0)
1670                                 goto done;
1671                         if (process_banner(con) < 0) {
1672                                 ret = -1;
1673                                 goto out;
1674                         }
1675                 }
1676                 ret = read_partial_connect(con);
1677                 if (ret <= 0)
1678                         goto done;
1679                 if (process_connect(con) < 0) {
1680                         ret = -1;
1681                         goto out;
1682                 }
1683                 goto more;
1684         }
1685
1686         if (con->in_base_pos < 0) {
1687                 /*
1688                  * skipping + discarding content.
1689                  *
1690                  * FIXME: there must be a better way to do this!
1691                  */
1692                 static char buf[1024];
1693                 int skip = min(1024, -con->in_base_pos);
1694                 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
1695                 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
1696                 if (ret <= 0)
1697                         goto done;
1698                 con->in_base_pos += ret;
1699                 if (con->in_base_pos)
1700                         goto more;
1701         }
1702         if (con->in_tag == CEPH_MSGR_TAG_READY) {
1703                 /*
1704                  * what's next?
1705                  */
1706                 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
1707                 if (ret <= 0)
1708                         goto done;
1709                 dout("try_read got tag %d\n", (int)con->in_tag);
1710                 switch (con->in_tag) {
1711                 case CEPH_MSGR_TAG_MSG:
1712                         prepare_read_message(con);
1713                         break;
1714                 case CEPH_MSGR_TAG_ACK:
1715                         prepare_read_ack(con);
1716                         break;
1717                 case CEPH_MSGR_TAG_CLOSE:
1718                         set_bit(CLOSED, &con->state);   /* fixme */
1719                         goto done;
1720                 default:
1721                         goto bad_tag;
1722                 }
1723         }
1724         if (con->in_tag == CEPH_MSGR_TAG_MSG) {
1725                 ret = read_partial_message(con);
1726                 if (ret <= 0) {
1727                         switch (ret) {
1728                         case -EBADMSG:
1729                                 con->error_msg = "bad crc";
1730                                 ret = -EIO;
1731                                 goto out;
1732                         case -EIO:
1733                                 con->error_msg = "io error";
1734                                 goto out;
1735                         default:
1736                                 goto done;
1737                         }
1738                 }
1739                 if (con->in_tag == CEPH_MSGR_TAG_READY)
1740                         goto more;
1741                 process_message(con);
1742                 goto more;
1743         }
1744         if (con->in_tag == CEPH_MSGR_TAG_ACK) {
1745                 ret = read_partial_ack(con);
1746                 if (ret <= 0)
1747                         goto done;
1748                 process_ack(con);
1749                 goto more;
1750         }
1751
1752 done:
1753         ret = 0;
1754 out:
1755         dout("try_read done on %p\n", con);
1756         return ret;
1757
1758 bad_tag:
1759         pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
1760         con->error_msg = "protocol error, garbage tag";
1761         ret = -1;
1762         goto out;
1763 }
1764
1765
1766 /*
1767  * Atomically queue work on a connection.  Bump @con reference to
1768  * avoid races with connection teardown.
1769  *
1770  * There is some trickery going on with QUEUED and BUSY because we
1771  * only want a _single_ thread operating on each connection at any
1772  * point in time, but we want to use all available CPUs.
1773  *
1774  * The worker thread only proceeds if it can atomically set BUSY.  It
1775  * clears QUEUED and does it's thing.  When it thinks it's done, it
1776  * clears BUSY, then rechecks QUEUED.. if it's set again, it loops
1777  * (tries again to set BUSY).
1778  *
1779  * To queue work, we first set QUEUED, _then_ if BUSY isn't set, we
1780  * try to queue work.  If that fails (work is already queued, or BUSY)
1781  * we give up (work also already being done or is queued) but leave QUEUED
1782  * set so that the worker thread will loop if necessary.
1783  */
1784 static void queue_con(struct ceph_connection *con)
1785 {
1786         if (test_bit(DEAD, &con->state)) {
1787                 dout("queue_con %p ignoring: DEAD\n",
1788                      con);
1789                 return;
1790         }
1791
1792         if (!con->ops->get(con)) {
1793                 dout("queue_con %p ref count 0\n", con);
1794                 return;
1795         }
1796
1797         set_bit(QUEUED, &con->state);
1798         if (test_bit(BUSY, &con->state)) {
1799                 dout("queue_con %p - already BUSY\n", con);
1800                 con->ops->put(con);
1801         } else if (!queue_work(ceph_msgr_wq, &con->work.work)) {
1802                 dout("queue_con %p - already queued\n", con);
1803                 con->ops->put(con);
1804         } else {
1805                 dout("queue_con %p\n", con);
1806         }
1807 }
1808
1809 /*
1810  * Do some work on a connection.  Drop a connection ref when we're done.
1811  */
1812 static void con_work(struct work_struct *work)
1813 {
1814         struct ceph_connection *con = container_of(work, struct ceph_connection,
1815                                                    work.work);
1816         int backoff = 0;
1817
1818 more:
1819         if (test_and_set_bit(BUSY, &con->state) != 0) {
1820                 dout("con_work %p BUSY already set\n", con);
1821                 goto out;
1822         }
1823         dout("con_work %p start, clearing QUEUED\n", con);
1824         clear_bit(QUEUED, &con->state);
1825
1826         mutex_lock(&con->mutex);
1827
1828         if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
1829                 dout("con_work CLOSED\n");
1830                 con_close_socket(con);
1831                 goto done;
1832         }
1833         if (test_and_clear_bit(OPENING, &con->state)) {
1834                 /* reopen w/ new peer */
1835                 dout("con_work OPENING\n");
1836                 con_close_socket(con);
1837         }
1838
1839         if (test_and_clear_bit(SOCK_CLOSED, &con->state) ||
1840             try_read(con) < 0 ||
1841             try_write(con) < 0) {
1842                 mutex_unlock(&con->mutex);
1843                 backoff = 1;
1844                 ceph_fault(con);     /* error/fault path */
1845                 goto done_unlocked;
1846         }
1847
1848 done:
1849         mutex_unlock(&con->mutex);
1850
1851 done_unlocked:
1852         clear_bit(BUSY, &con->state);
1853         dout("con->state=%lu\n", con->state);
1854         if (test_bit(QUEUED, &con->state)) {
1855                 if (!backoff || test_bit(OPENING, &con->state)) {
1856                         dout("con_work %p QUEUED reset, looping\n", con);
1857                         goto more;
1858                 }
1859                 dout("con_work %p QUEUED reset, but just faulted\n", con);
1860                 clear_bit(QUEUED, &con->state);
1861         }
1862         dout("con_work %p done\n", con);
1863
1864 out:
1865         con->ops->put(con);
1866 }
1867
1868
1869 /*
1870  * Generic error/fault handler.  A retry mechanism is used with
1871  * exponential backoff
1872  */
1873 static void ceph_fault(struct ceph_connection *con)
1874 {
1875         pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
1876                pr_addr(&con->peer_addr.in_addr), con->error_msg);
1877         dout("fault %p state %lu to peer %s\n",
1878              con, con->state, pr_addr(&con->peer_addr.in_addr));
1879
1880         if (test_bit(LOSSYTX, &con->state)) {
1881                 dout("fault on LOSSYTX channel\n");
1882                 goto out;
1883         }
1884
1885         mutex_lock(&con->mutex);
1886         if (test_bit(CLOSED, &con->state))
1887                 goto out_unlock;
1888
1889         con_close_socket(con);
1890
1891         if (con->in_msg) {
1892                 ceph_msg_put(con->in_msg);
1893                 con->in_msg = NULL;
1894         }
1895
1896         /* Requeue anything that hasn't been acked */
1897         list_splice_init(&con->out_sent, &con->out_queue);
1898
1899         /* If there are no messages in the queue, place the connection
1900          * in a STANDBY state (i.e., don't try to reconnect just yet). */
1901         if (list_empty(&con->out_queue) && !con->out_keepalive_pending) {
1902                 dout("fault setting STANDBY\n");
1903                 set_bit(STANDBY, &con->state);
1904         } else {
1905                 /* retry after a delay. */
1906                 if (con->delay == 0)
1907                         con->delay = BASE_DELAY_INTERVAL;
1908                 else if (con->delay < MAX_DELAY_INTERVAL)
1909                         con->delay *= 2;
1910                 dout("fault queueing %p delay %lu\n", con, con->delay);
1911                 con->ops->get(con);
1912                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
1913                                        round_jiffies_relative(con->delay)) == 0)
1914                         con->ops->put(con);
1915         }
1916
1917 out_unlock:
1918         mutex_unlock(&con->mutex);
1919 out:
1920         /*
1921          * in case we faulted due to authentication, invalidate our
1922          * current tickets so that we can get new ones.
1923          */
1924         if (con->auth_retry && con->ops->invalidate_authorizer) {
1925                 dout("calling invalidate_authorizer()\n");
1926                 con->ops->invalidate_authorizer(con);
1927         }
1928
1929         if (con->ops->fault)
1930                 con->ops->fault(con);
1931 }
1932
1933
1934
1935 /*
1936  * create a new messenger instance
1937  */
1938 struct ceph_messenger *ceph_messenger_create(struct ceph_entity_addr *myaddr)
1939 {
1940         struct ceph_messenger *msgr;
1941
1942         msgr = kzalloc(sizeof(*msgr), GFP_KERNEL);
1943         if (msgr == NULL)
1944                 return ERR_PTR(-ENOMEM);
1945
1946         spin_lock_init(&msgr->global_seq_lock);
1947
1948         /* the zero page is needed if a request is "canceled" while the message
1949          * is being written over the socket */
1950         msgr->zero_page = __page_cache_alloc(GFP_KERNEL | __GFP_ZERO);
1951         if (!msgr->zero_page) {
1952                 kfree(msgr);
1953                 return ERR_PTR(-ENOMEM);
1954         }
1955         kmap(msgr->zero_page);
1956
1957         if (myaddr)
1958                 msgr->inst.addr = *myaddr;
1959
1960         /* select a random nonce */
1961         msgr->inst.addr.type = 0;
1962         get_random_bytes(&msgr->inst.addr.nonce, sizeof(msgr->inst.addr.nonce));
1963         encode_my_addr(msgr);
1964
1965         dout("messenger_create %p\n", msgr);
1966         return msgr;
1967 }
1968
1969 void ceph_messenger_destroy(struct ceph_messenger *msgr)
1970 {
1971         dout("destroy %p\n", msgr);
1972         kunmap(msgr->zero_page);
1973         __free_page(msgr->zero_page);
1974         kfree(msgr);
1975         dout("destroyed messenger %p\n", msgr);
1976 }
1977
1978 /*
1979  * Queue up an outgoing message on the given connection.
1980  */
1981 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
1982 {
1983         if (test_bit(CLOSED, &con->state)) {
1984                 dout("con_send %p closed, dropping %p\n", con, msg);
1985                 ceph_msg_put(msg);
1986                 return;
1987         }
1988
1989         /* set src+dst */
1990         msg->hdr.src = con->msgr->inst.name;
1991
1992         BUG_ON(msg->front.iov_len != le32_to_cpu(msg->hdr.front_len));
1993
1994         msg->needs_out_seq = true;
1995
1996         /* queue */
1997         mutex_lock(&con->mutex);
1998         BUG_ON(!list_empty(&msg->list_head));
1999         list_add_tail(&msg->list_head, &con->out_queue);
2000         dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
2001              ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
2002              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
2003              le32_to_cpu(msg->hdr.front_len),
2004              le32_to_cpu(msg->hdr.middle_len),
2005              le32_to_cpu(msg->hdr.data_len));
2006         mutex_unlock(&con->mutex);
2007
2008         /* if there wasn't anything waiting to send before, queue
2009          * new work */
2010         if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2011                 queue_con(con);
2012 }
2013
2014 /*
2015  * Revoke a message that was previously queued for send
2016  */
2017 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
2018 {
2019         mutex_lock(&con->mutex);
2020         if (!list_empty(&msg->list_head)) {
2021                 dout("con_revoke %p msg %p\n", con, msg);
2022                 list_del_init(&msg->list_head);
2023                 ceph_msg_put(msg);
2024                 msg->hdr.seq = 0;
2025                 if (con->out_msg == msg) {
2026                         ceph_msg_put(con->out_msg);
2027                         con->out_msg = NULL;
2028                 }
2029                 if (con->out_kvec_is_msg) {
2030                         con->out_skip = con->out_kvec_bytes;
2031                         con->out_kvec_is_msg = false;
2032                 }
2033         } else {
2034                 dout("con_revoke %p msg %p - not queued (sent?)\n", con, msg);
2035         }
2036         mutex_unlock(&con->mutex);
2037 }
2038
2039 /*
2040  * Revoke a message that we may be reading data into
2041  */
2042 void ceph_con_revoke_message(struct ceph_connection *con, struct ceph_msg *msg)
2043 {
2044         mutex_lock(&con->mutex);
2045         if (con->in_msg && con->in_msg == msg) {
2046                 unsigned front_len = le32_to_cpu(con->in_hdr.front_len);
2047                 unsigned middle_len = le32_to_cpu(con->in_hdr.middle_len);
2048                 unsigned data_len = le32_to_cpu(con->in_hdr.data_len);
2049
2050                 /* skip rest of message */
2051                 dout("con_revoke_pages %p msg %p revoked\n", con, msg);
2052                         con->in_base_pos = con->in_base_pos -
2053                                 sizeof(struct ceph_msg_header) -
2054                                 front_len -
2055                                 middle_len -
2056                                 data_len -
2057                                 sizeof(struct ceph_msg_footer);
2058                 ceph_msg_put(con->in_msg);
2059                 con->in_msg = NULL;
2060                 con->in_tag = CEPH_MSGR_TAG_READY;
2061                 con->in_seq++;
2062         } else {
2063                 dout("con_revoke_pages %p msg %p pages %p no-op\n",
2064                      con, con->in_msg, msg);
2065         }
2066         mutex_unlock(&con->mutex);
2067 }
2068
2069 /*
2070  * Queue a keepalive byte to ensure the tcp connection is alive.
2071  */
2072 void ceph_con_keepalive(struct ceph_connection *con)
2073 {
2074         if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
2075             test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2076                 queue_con(con);
2077 }
2078
2079
2080 /*
2081  * construct a new message with given type, size
2082  * the new msg has a ref count of 1.
2083  */
2084 struct ceph_msg *ceph_msg_new(int type, int front_len)
2085 {
2086         struct ceph_msg *m;
2087
2088         m = kmalloc(sizeof(*m), GFP_NOFS);
2089         if (m == NULL)
2090                 goto out;
2091         kref_init(&m->kref);
2092         INIT_LIST_HEAD(&m->list_head);
2093
2094         m->hdr.tid = 0;
2095         m->hdr.type = cpu_to_le16(type);
2096         m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
2097         m->hdr.version = 0;
2098         m->hdr.front_len = cpu_to_le32(front_len);
2099         m->hdr.middle_len = 0;
2100         m->hdr.data_len = 0;
2101         m->hdr.data_off = 0;
2102         m->hdr.reserved = 0;
2103         m->footer.front_crc = 0;
2104         m->footer.middle_crc = 0;
2105         m->footer.data_crc = 0;
2106         m->footer.flags = 0;
2107         m->front_max = front_len;
2108         m->front_is_vmalloc = false;
2109         m->more_to_follow = false;
2110         m->pool = NULL;
2111
2112         /* front */
2113         if (front_len) {
2114                 if (front_len > PAGE_CACHE_SIZE) {
2115                         m->front.iov_base = __vmalloc(front_len, GFP_NOFS,
2116                                                       PAGE_KERNEL);
2117                         m->front_is_vmalloc = true;
2118                 } else {
2119                         m->front.iov_base = kmalloc(front_len, GFP_NOFS);
2120                 }
2121                 if (m->front.iov_base == NULL) {
2122                         pr_err("msg_new can't allocate %d bytes\n",
2123                              front_len);
2124                         goto out2;
2125                 }
2126         } else {
2127                 m->front.iov_base = NULL;
2128         }
2129         m->front.iov_len = front_len;
2130
2131         /* middle */
2132         m->middle = NULL;
2133
2134         /* data */
2135         m->nr_pages = 0;
2136         m->pages = NULL;
2137         m->pagelist = NULL;
2138
2139         dout("ceph_msg_new %p front %d\n", m, front_len);
2140         return m;
2141
2142 out2:
2143         ceph_msg_put(m);
2144 out:
2145         pr_err("msg_new can't create type %d front %d\n", type, front_len);
2146         return NULL;
2147 }
2148
2149 /*
2150  * Allocate "middle" portion of a message, if it is needed and wasn't
2151  * allocated by alloc_msg.  This allows us to read a small fixed-size
2152  * per-type header in the front and then gracefully fail (i.e.,
2153  * propagate the error to the caller based on info in the front) when
2154  * the middle is too large.
2155  */
2156 static int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2157 {
2158         int type = le16_to_cpu(msg->hdr.type);
2159         int middle_len = le32_to_cpu(msg->hdr.middle_len);
2160
2161         dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2162              ceph_msg_type_name(type), middle_len);
2163         BUG_ON(!middle_len);
2164         BUG_ON(msg->middle);
2165
2166         msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2167         if (!msg->middle)
2168                 return -ENOMEM;
2169         return 0;
2170 }
2171
2172 /*
2173  * Generic message allocator, for incoming messages.
2174  */
2175 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2176                                 struct ceph_msg_header *hdr,
2177                                 int *skip)
2178 {
2179         int type = le16_to_cpu(hdr->type);
2180         int front_len = le32_to_cpu(hdr->front_len);
2181         int middle_len = le32_to_cpu(hdr->middle_len);
2182         struct ceph_msg *msg = NULL;
2183         int ret;
2184
2185         if (con->ops->alloc_msg) {
2186                 mutex_unlock(&con->mutex);
2187                 msg = con->ops->alloc_msg(con, hdr, skip);
2188                 mutex_lock(&con->mutex);
2189                 if (!msg || *skip)
2190                         return NULL;
2191         }
2192         if (!msg) {
2193                 *skip = 0;
2194                 msg = ceph_msg_new(type, front_len);
2195                 if (!msg) {
2196                         pr_err("unable to allocate msg type %d len %d\n",
2197                                type, front_len);
2198                         return NULL;
2199                 }
2200         }
2201         memcpy(&msg->hdr, &con->in_hdr, sizeof(con->in_hdr));
2202
2203         if (middle_len && !msg->middle) {
2204                 ret = ceph_alloc_middle(con, msg);
2205                 if (ret < 0) {
2206                         ceph_msg_put(msg);
2207                         return NULL;
2208                 }
2209         }
2210
2211         return msg;
2212 }
2213
2214
2215 /*
2216  * Free a generically kmalloc'd message.
2217  */
2218 void ceph_msg_kfree(struct ceph_msg *m)
2219 {
2220         dout("msg_kfree %p\n", m);
2221         if (m->front_is_vmalloc)
2222                 vfree(m->front.iov_base);
2223         else
2224                 kfree(m->front.iov_base);
2225         kfree(m);
2226 }
2227
2228 /*
2229  * Drop a msg ref.  Destroy as needed.
2230  */
2231 void ceph_msg_last_put(struct kref *kref)
2232 {
2233         struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2234
2235         dout("ceph_msg_put last one on %p\n", m);
2236         WARN_ON(!list_empty(&m->list_head));
2237
2238         /* drop middle, data, if any */
2239         if (m->middle) {
2240                 ceph_buffer_put(m->middle);
2241                 m->middle = NULL;
2242         }
2243         m->nr_pages = 0;
2244         m->pages = NULL;
2245
2246         if (m->pagelist) {
2247                 ceph_pagelist_release(m->pagelist);
2248                 kfree(m->pagelist);
2249                 m->pagelist = NULL;
2250         }
2251
2252         if (m->pool)
2253                 ceph_msgpool_put(m->pool, m);
2254         else
2255                 ceph_msg_kfree(m);
2256 }
2257
2258 void ceph_msg_dump(struct ceph_msg *msg)
2259 {
2260         pr_debug("msg_dump %p (front_max %d nr_pages %d)\n", msg,
2261                  msg->front_max, msg->nr_pages);
2262         print_hex_dump(KERN_DEBUG, "header: ",
2263                        DUMP_PREFIX_OFFSET, 16, 1,
2264                        &msg->hdr, sizeof(msg->hdr), true);
2265         print_hex_dump(KERN_DEBUG, " front: ",
2266                        DUMP_PREFIX_OFFSET, 16, 1,
2267                        msg->front.iov_base, msg->front.iov_len, true);
2268         if (msg->middle)
2269                 print_hex_dump(KERN_DEBUG, "middle: ",
2270                                DUMP_PREFIX_OFFSET, 16, 1,
2271                                msg->middle->vec.iov_base,
2272                                msg->middle->vec.iov_len, true);
2273         print_hex_dump(KERN_DEBUG, "footer: ",
2274                        DUMP_PREFIX_OFFSET, 16, 1,
2275                        &msg->footer, sizeof(msg->footer), true);
2276 }