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/socket.h>
  10 #include <linux/string.h>
  11 #include <net/tcp.h>
  12
  13 #include "super.h"
  14 #include "messenger.h"
  15 #include "decode.h"
  16 #include "pagelist.h"
  17
  18 /*
  19  * Ceph uses the messenger to exchange ceph_msg messages with other
  20  * hosts in the system.  The messenger provides ordered and reliable
  21  * delivery.  We tolerate TCP disconnects by reconnecting (with
  22  * exponential backoff) in the case of a fault (disconnection, bad
  23  * crc, protocol error).  Acks allow sent messages to be discarded by
  24  * the sender.
  25  */
  26
  27 /* static tag bytes (protocol control messages) */
  28 static char tag_msg = CEPH_MSGR_TAG_MSG;
  29 static char tag_ack = CEPH_MSGR_TAG_ACK;
  30 static char tag_keepalive = CEPH_MSGR_TAG_KEEPALIVE;
  31
  32
  33 static void queue_con(struct ceph_connection *con);
  34 static void con_work(struct work_struct *);
  35 static void ceph_fault(struct ceph_connection *con);
  36
  37 const char *ceph_name_type_str(int t)
  38 {
  39         switch (t) {
  40         case CEPH_ENTITY_TYPE_MON: return "mon";
  41         case CEPH_ENTITY_TYPE_MDS: return "mds";
  42         case CEPH_ENTITY_TYPE_OSD: return "osd";
  43         case CEPH_ENTITY_TYPE_CLIENT: return "client";
  44         case CEPH_ENTITY_TYPE_ADMIN: return "admin";
  45         default: return "???";
  46         }
  47 }
  48
  49 /*
  50  * nicely render a sockaddr as a string.
  51  */
  52 #define MAX_ADDR_STR 20
  53 static char addr_str[MAX_ADDR_STR][40];
  54 static DEFINE_SPINLOCK(addr_str_lock);
  55 static int last_addr_str;
  56
  57 const char *pr_addr(const struct sockaddr_storage *ss)
  58 {
  59         int i;
  60         char *s;
  61         struct sockaddr_in *in4 = (void *)ss;
  62         unsigned char *quad = (void *)&in4->sin_addr.s_addr;
  63         struct sockaddr_in6 *in6 = (void *)ss;
  64
  65         spin_lock(&addr_str_lock);
  66         i = last_addr_str++;
  67         if (last_addr_str == MAX_ADDR_STR)
  68                 last_addr_str = 0;
  69         spin_unlock(&addr_str_lock);
  70         s = addr_str[i];
  71
  72         switch (ss->ss_family) {
  73         case AF_INET:
  74                 sprintf(s, "%u.%u.%u.%u:%u",
  75                         (unsigned int)quad[0],
  76                         (unsigned int)quad[1],
  77                         (unsigned int)quad[2],
  78                         (unsigned int)quad[3],
  79                         (unsigned int)ntohs(in4->sin_port));
  80                 break;
  81
  82         case AF_INET6:
  83                 sprintf(s, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x:%u",
  84                         in6->sin6_addr.s6_addr16[0],
  85                         in6->sin6_addr.s6_addr16[1],
  86                         in6->sin6_addr.s6_addr16[2],
  87                         in6->sin6_addr.s6_addr16[3],
  88                         in6->sin6_addr.s6_addr16[4],
  89                         in6->sin6_addr.s6_addr16[5],
  90                         in6->sin6_addr.s6_addr16[6],
  91                         in6->sin6_addr.s6_addr16[7],
  92                         (unsigned int)ntohs(in6->sin6_port));
  93                 break;
  94
  95         default:
  96                 sprintf(s, "(unknown sockaddr family %d)", (int)ss->ss_family);
  97         }
  98
  99         return s;
 100 }
 101
 102 static void encode_my_addr(struct ceph_messenger *msgr)
 103 {
 104         memcpy(&msgr->my_enc_addr, &msgr->inst.addr, sizeof(msgr->my_enc_addr));
 105         ceph_encode_addr(&msgr->my_enc_addr);
 106 }
 107
 108 /*
 109  * work queue for all reading and writing to/from the socket.
 110  */
 111 struct workqueue_struct *ceph_msgr_wq;
 112
 113 int __init ceph_msgr_init(void)
 114 {
 115         ceph_msgr_wq = create_workqueue("ceph-msgr");
 116         if (IS_ERR(ceph_msgr_wq)) {
 117                 int ret = PTR_ERR(ceph_msgr_wq);
 118                 pr_err("msgr_init failed to create workqueue: %d\n", ret);
 119                 ceph_msgr_wq = NULL;
 120                 return ret;
 121         }
 122         return 0;
 123 }
 124
 125 void ceph_msgr_exit(void)
 126 {
 127         destroy_workqueue(ceph_msgr_wq);
 128 }
 129
 130 /*
 131  * socket callback functions
 132  */
 133
 134 /* data available on socket, or listen socket received a connect */
 135 static void ceph_data_ready(struct sock *sk, int count_unused)
 136 {
 137         struct ceph_connection *con =
 138                 (struct ceph_connection *)sk->sk_user_data;
 139         if (sk->sk_state != TCP_CLOSE_WAIT) {
 140                 dout("ceph_data_ready on %p state = %lu, queueing work\n",
 141                      con, con->state);
 142                 queue_con(con);
 143         }
 144 }
 145
 146 /* socket has buffer space for writing */
 147 static void ceph_write_space(struct sock *sk)
 148 {
 149         struct ceph_connection *con =
 150                 (struct ceph_connection *)sk->sk_user_data;
 151
 152         /* only queue to workqueue if there is data we want to write. */
 153         if (test_bit(WRITE_PENDING, &con->state)) {
 154                 dout("ceph_write_space %p queueing write work\n", con);
 155                 queue_con(con);
 156         } else {
 157                 dout("ceph_write_space %p nothing to write\n", con);
 158         }
 159
 160         /* since we have our own write_space, clear the SOCK_NOSPACE flag */
 161         clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 162 }
 163
 164 /* socket's state has changed */
 165 static void ceph_state_change(struct sock *sk)
 166 {
 167         struct ceph_connection *con =
 168                 (struct ceph_connection *)sk->sk_user_data;
 169
 170         dout("ceph_state_change %p state = %lu sk_state = %u\n",
 171              con, con->state, sk->sk_state);
 172
 173         if (test_bit(CLOSED, &con->state))
 174                 return;
 175
 176         switch (sk->sk_state) {
 177         case TCP_CLOSE:
 178                 dout("ceph_state_change TCP_CLOSE\n");
 179         case TCP_CLOSE_WAIT:
 180                 dout("ceph_state_change TCP_CLOSE_WAIT\n");
 181                 if (test_and_set_bit(SOCK_CLOSED, &con->state) == 0) {
 182                         if (test_bit(CONNECTING, &con->state))
 183                                 con->error_msg = "connection failed";
 184                         else
 185                                 con->error_msg = "socket closed";
 186                         queue_con(con);
 187                 }
 188                 break;
 189         case TCP_ESTABLISHED:
 190                 dout("ceph_state_change TCP_ESTABLISHED\n");
 191                 queue_con(con);
 192                 break;
 193         }
 194 }
 195
 196 /*
 197  * set up socket callbacks
 198  */
 199 static void set_sock_callbacks(struct socket *sock,
 200                                struct ceph_connection *con)
 201 {
 202         struct sock *sk = sock->sk;
 203         sk->sk_user_data = (void *)con;
 204         sk->sk_data_ready = ceph_data_ready;
 205         sk->sk_write_space = ceph_write_space;
 206         sk->sk_state_change = ceph_state_change;
 207 }
 208
 209
 210 /*
 211  * socket helpers
 212  */
 213
 214 /*
 215  * initiate connection to a remote socket.
 216  */
 217 static struct socket *ceph_tcp_connect(struct ceph_connection *con)
 218 {
 219         struct sockaddr *paddr = (struct sockaddr *)&con->peer_addr.in_addr;
 220         struct socket *sock;
 221         int ret;
 222
 223         BUG_ON(con->sock);
 224         ret = sock_create_kern(AF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
 225         if (ret)
 226                 return ERR_PTR(ret);
 227         con->sock = sock;
 228         sock->sk->sk_allocation = GFP_NOFS;
 229
 230         set_sock_callbacks(sock, con);
 231
 232         dout("connect %s\n", pr_addr(&con->peer_addr.in_addr));
 233
 234         ret = sock->ops->connect(sock, paddr, sizeof(*paddr), O_NONBLOCK);
 235         if (ret == -EINPROGRESS) {
 236                 dout("connect %s EINPROGRESS sk_state = %u\n",
 237                      pr_addr(&con->peer_addr.in_addr),
 238                      sock->sk->sk_state);
 239                 ret = 0;
 240         }
 241         if (ret < 0) {
 242                 pr_err("connect %s error %d\n",
 243                        pr_addr(&con->peer_addr.in_addr), ret);
 244                 sock_release(sock);
 245                 con->sock = NULL;
 246                 con->error_msg = "connect error";
 247         }
 248
 249         if (ret < 0)
 250                 return ERR_PTR(ret);
 251         return sock;
 252 }
 253
 254 static int ceph_tcp_recvmsg(struct socket *sock, void *buf, size_t len)
 255 {
 256         struct kvec iov = {buf, len};
 257         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
 258
 259         return kernel_recvmsg(sock, &msg, &iov, 1, len, msg.msg_flags);
 260 }
 261
 262 /*
 263  * write something.  @more is true if caller will be sending more data
 264  * shortly.
 265  */
 266 static int ceph_tcp_sendmsg(struct socket *sock, struct kvec *iov,
 267                      size_t kvlen, size_t len, int more)
 268 {
 269         struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_NOSIGNAL };
 270
 271         if (more)
 272                 msg.msg_flags |= MSG_MORE;
 273         else
 274                 msg.msg_flags |= MSG_EOR;  /* superfluous, but what the hell */
 275
 276         return kernel_sendmsg(sock, &msg, iov, kvlen, len);
 277 }
 278
 279
 280 /*
 281  * Shutdown/close the socket for the given connection.
 282  */
 283 static int con_close_socket(struct ceph_connection *con)
 284 {
 285         int rc;
 286
 287         dout("con_close_socket on %p sock %p\n", con, con->sock);
 288         if (!con->sock)
 289                 return 0;
 290         set_bit(SOCK_CLOSED, &con->state);
 291         rc = con->sock->ops->shutdown(con->sock, SHUT_RDWR);
 292         sock_release(con->sock);
 293         con->sock = NULL;
 294         clear_bit(SOCK_CLOSED, &con->state);
 295         return rc;
 296 }
 297
 298 /*
 299  * Reset a connection.  Discard all incoming and outgoing messages
 300  * and clear *_seq state.
 301  */
 302 static void ceph_msg_remove(struct ceph_msg *msg)
 303 {
 304         list_del_init(&msg->list_head);
 305         ceph_msg_put(msg);
 306 }
 307 static void ceph_msg_remove_list(struct list_head *head)
 308 {
 309         while (!list_empty(head)) {
 310                 struct ceph_msg *msg = list_first_entry(head, struct ceph_msg,
 311                                                         list_head);
 312                 ceph_msg_remove(msg);
 313         }
 314 }
 315
 316 static void reset_connection(struct ceph_connection *con)
 317 {
 318         /* reset connection, out_queue, msg_ and connect_seq */
 319         /* discard existing out_queue and msg_seq */
 320         ceph_msg_remove_list(&con->out_queue);
 321         ceph_msg_remove_list(&con->out_sent);
 322
 323         if (con->in_msg) {
 324                 ceph_msg_put(con->in_msg);
 325                 con->in_msg = NULL;
 326         }
 327
 328         con->connect_seq = 0;
 329         con->out_seq = 0;
 330         if (con->out_msg) {
 331                 ceph_msg_put(con->out_msg);
 332                 con->out_msg = NULL;
 333         }
 334         con->in_seq = 0;
 335 }
 336
 337 /*
 338  * mark a peer down.  drop any open connections.
 339  */
 340 void ceph_con_close(struct ceph_connection *con)
 341 {
 342         dout("con_close %p peer %s\n", con, pr_addr(&con->peer_addr.in_addr));
 343         set_bit(CLOSED, &con->state);  /* in case there's queued work */
 344         clear_bit(STANDBY, &con->state);  /* avoid connect_seq bump */
 345         mutex_lock(&con->mutex);
 346         reset_connection(con);
 347         cancel_delayed_work(&con->work);
 348         mutex_unlock(&con->mutex);
 349         queue_con(con);
 350 }
 351
 352 /*
 353  * Reopen a closed connection, with a new peer address.
 354  */
 355 void ceph_con_open(struct ceph_connection *con, struct ceph_entity_addr *addr)
 356 {
 357         dout("con_open %p %s\n", con, pr_addr(&addr->in_addr));
 358         set_bit(OPENING, &con->state);
 359         clear_bit(CLOSED, &con->state);
 360         memcpy(&con->peer_addr, addr, sizeof(*addr));
 361         con->delay = 0;      /* reset backoff memory */
 362         queue_con(con);
 363 }
 364
 365 /*
 366  * generic get/put
 367  */
 368 struct ceph_connection *ceph_con_get(struct ceph_connection *con)
 369 {
 370         dout("con_get %p nref = %d -> %d\n", con,
 371              atomic_read(&con->nref), atomic_read(&con->nref) + 1);
 372         if (atomic_inc_not_zero(&con->nref))
 373                 return con;
 374         return NULL;
 375 }
 376
 377 void ceph_con_put(struct ceph_connection *con)
 378 {
 379         dout("con_put %p nref = %d -> %d\n", con,
 380              atomic_read(&con->nref), atomic_read(&con->nref) - 1);
 381         BUG_ON(atomic_read(&con->nref) == 0);
 382         if (atomic_dec_and_test(&con->nref)) {
 383                 BUG_ON(con->sock);
 384                 kfree(con);
 385         }
 386 }
 387
 388 /*
 389  * initialize a new connection.
 390  */
 391 void ceph_con_init(struct ceph_messenger *msgr, struct ceph_connection *con)
 392 {
 393         dout("con_init %p\n", con);
 394         memset(con, 0, sizeof(*con));
 395         atomic_set(&con->nref, 1);
 396         con->msgr = msgr;
 397         mutex_init(&con->mutex);
 398         INIT_LIST_HEAD(&con->out_queue);
 399         INIT_LIST_HEAD(&con->out_sent);
 400         INIT_DELAYED_WORK(&con->work, con_work);
 401 }
 402
 403
 404 /*
 405  * We maintain a global counter to order connection attempts.  Get
 406  * a unique seq greater than @gt.
 407  */
 408 static u32 get_global_seq(struct ceph_messenger *msgr, u32 gt)
 409 {
 410         u32 ret;
 411
 412         spin_lock(&msgr->global_seq_lock);
 413         if (msgr->global_seq < gt)
 414                 msgr->global_seq = gt;
 415         ret = ++msgr->global_seq;
 416         spin_unlock(&msgr->global_seq_lock);
 417         return ret;
 418 }
 419
 420
 421 /*
 422  * Prepare footer for currently outgoing message, and finish things
 423  * off.  Assumes out_kvec* are already valid.. we just add on to the end.
 424  */
 425 static void prepare_write_message_footer(struct ceph_connection *con, int v)
 426 {
 427         struct ceph_msg *m = con->out_msg;
 428
 429         dout("prepare_write_message_footer %p\n", con);
 430         con->out_kvec_is_msg = true;
 431         con->out_kvec[v].iov_base = &m->footer;
 432         con->out_kvec[v].iov_len = sizeof(m->footer);
 433         con->out_kvec_bytes += sizeof(m->footer);
 434         con->out_kvec_left++;
 435         con->out_more = m->more_to_follow;
 436         con->out_msg_done = true;
 437 }
 438
 439 /*
 440  * Prepare headers for the next outgoing message.
 441  */
 442 static void prepare_write_message(struct ceph_connection *con)
 443 {
 444         struct ceph_msg *m;
 445         int v = 0;
 446
 447         con->out_kvec_bytes = 0;
 448         con->out_kvec_is_msg = true;
 449         con->out_msg_done = false;
 450
 451         /* Sneak an ack in there first?  If we can get it into the same
 452          * TCP packet that's a good thing. */
 453         if (con->in_seq > con->in_seq_acked) {
 454                 con->in_seq_acked = con->in_seq;
 455                 con->out_kvec[v].iov_base = &tag_ack;
 456                 con->out_kvec[v++].iov_len = 1;
 457                 con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 458                 con->out_kvec[v].iov_base = &con->out_temp_ack;
 459                 con->out_kvec[v++].iov_len = sizeof(con->out_temp_ack);
 460                 con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 461         }
 462
 463         m = list_first_entry(&con->out_queue,
 464                        struct ceph_msg, list_head);
 465         con->out_msg = m;
 466         if (test_bit(LOSSYTX, &con->state)) {
 467                 list_del_init(&m->list_head);
 468         } else {
 469                 /* put message on sent list */
 470                 ceph_msg_get(m);
 471                 list_move_tail(&m->list_head, &con->out_sent);
 472         }
 473
 474         m->hdr.seq = cpu_to_le64(++con->out_seq);
 475
 476         dout("prepare_write_message %p seq %lld type %d len %d+%d+%d %d pgs\n",
 477              m, con->out_seq, le16_to_cpu(m->hdr.type),
 478              le32_to_cpu(m->hdr.front_len), le32_to_cpu(m->hdr.middle_len),
 479              le32_to_cpu(m->hdr.data_len),
 480              m->nr_pages);
 481         BUG_ON(le32_to_cpu(m->hdr.front_len) != m->front.iov_len);
 482
 483         /* tag + hdr + front + middle */
 484         con->out_kvec[v].iov_base = &tag_msg;
 485         con->out_kvec[v++].iov_len = 1;
 486         con->out_kvec[v].iov_base = &m->hdr;
 487         con->out_kvec[v++].iov_len = sizeof(m->hdr);
 488         con->out_kvec[v++] = m->front;
 489         if (m->middle)
 490                 con->out_kvec[v++] = m->middle->vec;
 491         con->out_kvec_left = v;
 492         con->out_kvec_bytes += 1 + sizeof(m->hdr) + m->front.iov_len +
 493                 (m->middle ? m->middle->vec.iov_len : 0);
 494         con->out_kvec_cur = con->out_kvec;
 495
 496         /* fill in crc (except data pages), footer */
 497         con->out_msg->hdr.crc =
 498                 cpu_to_le32(crc32c(0, (void *)&m->hdr,
 499                                       sizeof(m->hdr) - sizeof(m->hdr.crc)));
 500         con->out_msg->footer.flags = CEPH_MSG_FOOTER_COMPLETE;
 501         con->out_msg->footer.front_crc =
 502                 cpu_to_le32(crc32c(0, m->front.iov_base, m->front.iov_len));
 503         if (m->middle)
 504                 con->out_msg->footer.middle_crc =
 505                         cpu_to_le32(crc32c(0, m->middle->vec.iov_base,
 506                                            m->middle->vec.iov_len));
 507         else
 508                 con->out_msg->footer.middle_crc = 0;
 509         con->out_msg->footer.data_crc = 0;
 510         dout("prepare_write_message front_crc %u data_crc %u\n",
 511              le32_to_cpu(con->out_msg->footer.front_crc),
 512              le32_to_cpu(con->out_msg->footer.middle_crc));
 513
 514         /* is there a data payload? */
 515         if (le32_to_cpu(m->hdr.data_len) > 0) {
 516                 /* initialize page iterator */
 517                 con->out_msg_pos.page = 0;
 518                 con->out_msg_pos.page_pos =
 519                         le16_to_cpu(m->hdr.data_off) & ~PAGE_MASK;
 520                 con->out_msg_pos.data_pos = 0;
 521                 con->out_msg_pos.did_page_crc = 0;
 522                 con->out_more = 1;  /* data + footer will follow */
 523         } else {
 524                 /* no, queue up footer too and be done */
 525                 prepare_write_message_footer(con, v);
 526         }
 527
 528         set_bit(WRITE_PENDING, &con->state);
 529 }
 530
 531 /*
 532  * Prepare an ack.
 533  */
 534 static void prepare_write_ack(struct ceph_connection *con)
 535 {
 536         dout("prepare_write_ack %p %llu -> %llu\n", con,
 537              con->in_seq_acked, con->in_seq);
 538         con->in_seq_acked = con->in_seq;
 539
 540         con->out_kvec[0].iov_base = &tag_ack;
 541         con->out_kvec[0].iov_len = 1;
 542         con->out_temp_ack = cpu_to_le64(con->in_seq_acked);
 543         con->out_kvec[1].iov_base = &con->out_temp_ack;
 544         con->out_kvec[1].iov_len = sizeof(con->out_temp_ack);
 545         con->out_kvec_left = 2;
 546         con->out_kvec_bytes = 1 + sizeof(con->out_temp_ack);
 547         con->out_kvec_cur = con->out_kvec;
 548         con->out_more = 1;  /* more will follow.. eventually.. */
 549         set_bit(WRITE_PENDING, &con->state);
 550 }
 551
 552 /*
 553  * Prepare to write keepalive byte.
 554  */
 555 static void prepare_write_keepalive(struct ceph_connection *con)
 556 {
 557         dout("prepare_write_keepalive %p\n", con);
 558         con->out_kvec[0].iov_base = &tag_keepalive;
 559         con->out_kvec[0].iov_len = 1;
 560         con->out_kvec_left = 1;
 561         con->out_kvec_bytes = 1;
 562         con->out_kvec_cur = con->out_kvec;
 563         set_bit(WRITE_PENDING, &con->state);
 564 }
 565
 566 /*
 567  * Connection negotiation.
 568  */
 569
 570 static void prepare_connect_authorizer(struct ceph_connection *con)
 571 {
 572         void *auth_buf;
 573         int auth_len = 0;
 574         int auth_protocol = 0;
 575
 576         mutex_unlock(&con->mutex);
 577         if (con->ops->get_authorizer)
 578                 con->ops->get_authorizer(con, &auth_buf, &auth_len,
 579                                          &auth_protocol, &con->auth_reply_buf,
 580                                          &con->auth_reply_buf_len,
 581                                          con->auth_retry);
 582         mutex_lock(&con->mutex);
 583
 584         con->out_connect.authorizer_protocol = cpu_to_le32(auth_protocol);
 585         con->out_connect.authorizer_len = cpu_to_le32(auth_len);
 586
 587         con->out_kvec[con->out_kvec_left].iov_base = auth_buf;
 588         con->out_kvec[con->out_kvec_left].iov_len = auth_len;
 589         con->out_kvec_left++;
 590         con->out_kvec_bytes += auth_len;
 591 }
 592
 593 /*
 594  * We connected to a peer and are saying hello.
 595  */
 596 static void prepare_write_banner(struct ceph_messenger *msgr,
 597                                  struct ceph_connection *con)
 598 {
 599         int len = strlen(CEPH_BANNER);
 600
 601         con->out_kvec[0].iov_base = CEPH_BANNER;
 602         con->out_kvec[0].iov_len = len;
 603         con->out_kvec[1].iov_base = &msgr->my_enc_addr;
 604         con->out_kvec[1].iov_len = sizeof(msgr->my_enc_addr);
 605         con->out_kvec_left = 2;
 606         con->out_kvec_bytes = len + sizeof(msgr->my_enc_addr);
 607         con->out_kvec_cur = con->out_kvec;
 608         con->out_more = 0;
 609         set_bit(WRITE_PENDING, &con->state);
 610 }
 611
 612 static void prepare_write_connect(struct ceph_messenger *msgr,
 613                                   struct ceph_connection *con,
 614                                   int after_banner)
 615 {
 616         unsigned global_seq = get_global_seq(con->msgr, 0);
 617         int proto;
 618
 619         switch (con->peer_name.type) {
 620         case CEPH_ENTITY_TYPE_MON:
 621                 proto = CEPH_MONC_PROTOCOL;
 622                 break;
 623         case CEPH_ENTITY_TYPE_OSD:
 624                 proto = CEPH_OSDC_PROTOCOL;
 625                 break;
 626         case CEPH_ENTITY_TYPE_MDS:
 627                 proto = CEPH_MDSC_PROTOCOL;
 628                 break;
 629         default:
 630                 BUG();
 631         }
 632
 633         dout("prepare_write_connect %p cseq=%d gseq=%d proto=%d\n", con,
 634              con->connect_seq, global_seq, proto);
 635
 636         con->out_connect.features = CEPH_FEATURE_SUPPORTED;
 637         con->out_connect.host_type = cpu_to_le32(CEPH_ENTITY_TYPE_CLIENT);
 638         con->out_connect.connect_seq = cpu_to_le32(con->connect_seq);
 639         con->out_connect.global_seq = cpu_to_le32(global_seq);
 640         con->out_connect.protocol_version = cpu_to_le32(proto);
 641         con->out_connect.flags = 0;
 642
 643         if (!after_banner) {
 644                 con->out_kvec_left = 0;
 645                 con->out_kvec_bytes = 0;
 646         }
 647         con->out_kvec[con->out_kvec_left].iov_base = &con->out_connect;
 648         con->out_kvec[con->out_kvec_left].iov_len = sizeof(con->out_connect);
 649         con->out_kvec_left++;
 650         con->out_kvec_bytes += sizeof(con->out_connect);
 651         con->out_kvec_cur = con->out_kvec;
 652         con->out_more = 0;
 653         set_bit(WRITE_PENDING, &con->state);
 654
 655         prepare_connect_authorizer(con);
 656 }
 657
 658
 659 /*
 660  * write as much of pending kvecs to the socket as we can.
 661  *  1 -> done
 662  *  0 -> socket full, but more to do
 663  * <0 -> error
 664  */
 665 static int write_partial_kvec(struct ceph_connection *con)
 666 {
 667         int ret;
 668
 669         dout("write_partial_kvec %p %d left\n", con, con->out_kvec_bytes);
 670         while (con->out_kvec_bytes > 0) {
 671                 ret = ceph_tcp_sendmsg(con->sock, con->out_kvec_cur,
 672                                        con->out_kvec_left, con->out_kvec_bytes,
 673                                        con->out_more);
 674                 if (ret <= 0)
 675                         goto out;
 676                 con->out_kvec_bytes -= ret;
 677                 if (con->out_kvec_bytes == 0)
 678                         break;            /* done */
 679                 while (ret > 0) {
 680                         if (ret >= con->out_kvec_cur->iov_len) {
 681                                 ret -= con->out_kvec_cur->iov_len;
 682                                 con->out_kvec_cur++;
 683                                 con->out_kvec_left--;
 684                         } else {
 685                                 con->out_kvec_cur->iov_len -= ret;
 686                                 con->out_kvec_cur->iov_base += ret;
 687                                 ret = 0;
 688                                 break;
 689                         }
 690                 }
 691         }
 692         con->out_kvec_left = 0;
 693         con->out_kvec_is_msg = false;
 694         ret = 1;
 695 out:
 696         dout("write_partial_kvec %p %d left in %d kvecs ret = %d\n", con,
 697              con->out_kvec_bytes, con->out_kvec_left, ret);
 698         return ret;  /* done! */
 699 }
 700
 701 /*
 702  * Write as much message data payload as we can.  If we finish, queue
 703  * up the footer.
 704  *  1 -> done, footer is now queued in out_kvec[].
 705  *  0 -> socket full, but more to do
 706  * <0 -> error
 707  */
 708 static int write_partial_msg_pages(struct ceph_connection *con)
 709 {
 710         struct ceph_msg *msg = con->out_msg;
 711         unsigned data_len = le32_to_cpu(msg->hdr.data_len);
 712         size_t len;
 713         int crc = con->msgr->nocrc;
 714         int ret;
 715
 716         dout("write_partial_msg_pages %p msg %p page %d/%d offset %d\n",
 717              con, con->out_msg, con->out_msg_pos.page, con->out_msg->nr_pages,
 718              con->out_msg_pos.page_pos);
 719
 720         while (con->out_msg_pos.page < con->out_msg->nr_pages) {
 721                 struct page *page = NULL;
 722                 void *kaddr = NULL;
 723
 724                 /*
 725                  * if we are calculating the data crc (the default), we need
 726                  * to map the page.  if our pages[] has been revoked, use the
 727                  * zero page.
 728                  */
 729                 if (msg->pages) {
 730                         page = msg->pages[con->out_msg_pos.page];
 731                         if (crc)
 732                                 kaddr = kmap(page);
 733                 } else if (msg->pagelist) {
 734                         page = list_first_entry(&msg->pagelist->head,
 735                                                 struct page, lru);
 736                         if (crc)
 737                                 kaddr = kmap(page);
 738                 } else {
 739                         page = con->msgr->zero_page;
 740                         if (crc)
 741                                 kaddr = page_address(con->msgr->zero_page);
 742                 }
 743                 len = min((int)(PAGE_SIZE - con->out_msg_pos.page_pos),
 744                           (int)(data_len - con->out_msg_pos.data_pos));
 745                 if (crc && !con->out_msg_pos.did_page_crc) {
 746                         void *base = kaddr + con->out_msg_pos.page_pos;
 747                         u32 tmpcrc = le32_to_cpu(con->out_msg->footer.data_crc);
 748
 749                         BUG_ON(kaddr == NULL);
 750                         con->out_msg->footer.data_crc =
 751                                 cpu_to_le32(crc32c(tmpcrc, base, len));
 752                         con->out_msg_pos.did_page_crc = 1;
 753                 }
 754
 755                 ret = kernel_sendpage(con->sock, page,
 756                                       con->out_msg_pos.page_pos, len,
 757                                       MSG_DONTWAIT | MSG_NOSIGNAL |
 758                                       MSG_MORE);
 759
 760                 if (crc && (msg->pages || msg->pagelist))
 761                         kunmap(page);
 762
 763                 if (ret <= 0)
 764                         goto out;
 765
 766                 con->out_msg_pos.data_pos += ret;
 767                 con->out_msg_pos.page_pos += ret;
 768                 if (ret == len) {
 769                         con->out_msg_pos.page_pos = 0;
 770                         con->out_msg_pos.page++;
 771                         con->out_msg_pos.did_page_crc = 0;
 772                         if (msg->pagelist)
 773                                 list_move_tail(&page->lru,
 774                                                &msg->pagelist->head);
 775                 }
 776         }
 777
 778         dout("write_partial_msg_pages %p msg %p done\n", con, msg);
 779
 780         /* prepare and queue up footer, too */
 781         if (!crc)
 782                 con->out_msg->footer.flags |= CEPH_MSG_FOOTER_NOCRC;
 783         con->out_kvec_bytes = 0;
 784         con->out_kvec_left = 0;
 785         con->out_kvec_cur = con->out_kvec;
 786         prepare_write_message_footer(con, 0);
 787         ret = 1;
 788 out:
 789         return ret;
 790 }
 791
 792 /*
 793  * write some zeros
 794  */
 795 static int write_partial_skip(struct ceph_connection *con)
 796 {
 797         int ret;
 798
 799         while (con->out_skip > 0) {
 800                 struct kvec iov = {
 801                         .iov_base = page_address(con->msgr->zero_page),
 802                         .iov_len = min(con->out_skip, (int)PAGE_CACHE_SIZE)
 803                 };
 804
 805                 ret = ceph_tcp_sendmsg(con->sock, &iov, 1, iov.iov_len, 1);
 806                 if (ret <= 0)
 807                         goto out;
 808                 con->out_skip -= ret;
 809         }
 810         ret = 1;
 811 out:
 812         return ret;
 813 }
 814
 815 /*
 816  * Prepare to read connection handshake, or an ack.
 817  */
 818 static void prepare_read_banner(struct ceph_connection *con)
 819 {
 820         dout("prepare_read_banner %p\n", con);
 821         con->in_base_pos = 0;
 822 }
 823
 824 static void prepare_read_connect(struct ceph_connection *con)
 825 {
 826         dout("prepare_read_connect %p\n", con);
 827         con->in_base_pos = 0;
 828 }
 829
 830 static void prepare_read_connect_retry(struct ceph_connection *con)
 831 {
 832         dout("prepare_read_connect_retry %p\n", con);
 833         con->in_base_pos = strlen(CEPH_BANNER) + sizeof(con->actual_peer_addr)
 834                 + sizeof(con->peer_addr_for_me);
 835 }
 836
 837 static void prepare_read_ack(struct ceph_connection *con)
 838 {
 839         dout("prepare_read_ack %p\n", con);
 840         con->in_base_pos = 0;
 841 }
 842
 843 static void prepare_read_tag(struct ceph_connection *con)
 844 {
 845         dout("prepare_read_tag %p\n", con);
 846         con->in_base_pos = 0;
 847         con->in_tag = CEPH_MSGR_TAG_READY;
 848 }
 849
 850 /*
 851  * Prepare to read a message.
 852  */
 853 static int prepare_read_message(struct ceph_connection *con)
 854 {
 855         dout("prepare_read_message %p\n", con);
 856         BUG_ON(con->in_msg != NULL);
 857         con->in_base_pos = 0;
 858         con->in_front_crc = con->in_middle_crc = con->in_data_crc = 0;
 859         return 0;
 860 }
 861
 862
 863 static int read_partial(struct ceph_connection *con,
 864                         int *to, int size, void *object)
 865 {
 866         *to += size;
 867         while (con->in_base_pos < *to) {
 868                 int left = *to - con->in_base_pos;
 869                 int have = size - left;
 870                 int ret = ceph_tcp_recvmsg(con->sock, object + have, left);
 871                 if (ret <= 0)
 872                         return ret;
 873                 con->in_base_pos += ret;
 874         }
 875         return 1;
 876 }
 877
 878
 879 /*
 880  * Read all or part of the connect-side handshake on a new connection
 881  */
 882 static int read_partial_banner(struct ceph_connection *con)
 883 {
 884         int ret, to = 0;
 885
 886         dout("read_partial_banner %p at %d\n", con, con->in_base_pos);
 887
 888         /* peer's banner */
 889         ret = read_partial(con, &to, strlen(CEPH_BANNER), con->in_banner);
 890         if (ret <= 0)
 891                 goto out;
 892         ret = read_partial(con, &to, sizeof(con->actual_peer_addr),
 893                            &con->actual_peer_addr);
 894         if (ret <= 0)
 895                 goto out;
 896         ret = read_partial(con, &to, sizeof(con->peer_addr_for_me),
 897                            &con->peer_addr_for_me);
 898         if (ret <= 0)
 899                 goto out;
 900 out:
 901         return ret;
 902 }
 903
 904 static int read_partial_connect(struct ceph_connection *con)
 905 {
 906         int ret, to = 0;
 907
 908         dout("read_partial_connect %p at %d\n", con, con->in_base_pos);
 909
 910         ret = read_partial(con, &to, sizeof(con->in_reply), &con->in_reply);
 911         if (ret <= 0)
 912                 goto out;
 913         ret = read_partial(con, &to, le32_to_cpu(con->in_reply.authorizer_len),
 914                            con->auth_reply_buf);
 915         if (ret <= 0)
 916                 goto out;
 917
 918         dout("read_partial_connect %p tag %d, con_seq = %u, g_seq = %u\n",
 919              con, (int)con->in_reply.tag,
 920              le32_to_cpu(con->in_reply.connect_seq),
 921              le32_to_cpu(con->in_reply.global_seq));
 922 out:
 923         return ret;
 924
 925 }
 926
 927 /*
 928  * Verify the hello banner looks okay.
 929  */
 930 static int verify_hello(struct ceph_connection *con)
 931 {
 932         if (memcmp(con->in_banner, CEPH_BANNER, strlen(CEPH_BANNER))) {
 933                 pr_err("connect to %s got bad banner\n",
 934                        pr_addr(&con->peer_addr.in_addr));
 935                 con->error_msg = "protocol error, bad banner";
 936                 return -1;
 937         }
 938         return 0;
 939 }
 940
 941 static bool addr_is_blank(struct sockaddr_storage *ss)
 942 {
 943         switch (ss->ss_family) {
 944         case AF_INET:
 945                 return ((struct sockaddr_in *)ss)->sin_addr.s_addr == 0;
 946         case AF_INET6:
 947                 return
 948                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[0] == 0 &&
 949                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[1] == 0 &&
 950                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[2] == 0 &&
 951                      ((struct sockaddr_in6 *)ss)->sin6_addr.s6_addr32[3] == 0;
 952         }
 953         return false;
 954 }
 955
 956 static int addr_port(struct sockaddr_storage *ss)
 957 {
 958         switch (ss->ss_family) {
 959         case AF_INET:
 960                 return ntohs(((struct sockaddr_in *)ss)->sin_port);
 961         case AF_INET6:
 962                 return ntohs(((struct sockaddr_in6 *)ss)->sin6_port);
 963         }
 964         return 0;
 965 }
 966
 967 static void addr_set_port(struct sockaddr_storage *ss, int p)
 968 {
 969         switch (ss->ss_family) {
 970         case AF_INET:
 971                 ((struct sockaddr_in *)ss)->sin_port = htons(p);
 972         case AF_INET6:
 973                 ((struct sockaddr_in6 *)ss)->sin6_port = htons(p);
 974         }
 975 }
 976
 977 /*
 978  * Parse an ip[:port] list into an addr array.  Use the default
 979  * monitor port if a port isn't specified.
 980  */
 981 int ceph_parse_ips(const char *c, const char *end,
 982                    struct ceph_entity_addr *addr,
 983                    int max_count, int *count)
 984 {
 985         int i;
 986         const char *p = c;
 987
 988         dout("parse_ips on '%.*s'\n", (int)(end-c), c);
 989         for (i = 0; i < max_count; i++) {
 990                 const char *ipend;
 991                 struct sockaddr_storage *ss = &addr[i].in_addr;
 992                 struct sockaddr_in *in4 = (void *)ss;
 993                 struct sockaddr_in6 *in6 = (void *)ss;
 994                 int port;
 995
 996                 memset(ss, 0, sizeof(*ss));
 997                 if (in4_pton(p, end - p, (u8 *)&in4->sin_addr.s_addr,
 998                              ',', &ipend)) {
 999                         ss->ss_family = AF_INET;
1000                 } else if (in6_pton(p, end - p, (u8 *)&in6->sin6_addr.s6_addr,
1001                                     ',', &ipend)) {
1002                         ss->ss_family = AF_INET6;
1003                 } else {
1004                         goto bad;
1005                 }
1006                 p = ipend;
1007
1008                 /* port? */
1009                 if (p < end && *p == ':') {
1010                         port = 0;
1011                         p++;
1012                         while (p < end && *p >= '0' && *p <= '9') {
1013                                 port = (port * 10) + (*p - '0');
1014                                 p++;
1015                         }
1016                         if (port > 65535 || port == 0)
1017                                 goto bad;
1018                 } else {
1019                         port = CEPH_MON_PORT;
1020                 }
1021
1022                 addr_set_port(ss, port);
1023
1024                 dout("parse_ips got %s\n", pr_addr(ss));
1025
1026                 if (p == end)
1027                         break;
1028                 if (*p != ',')
1029                         goto bad;
1030                 p++;
1031         }
1032
1033         if (p != end)
1034                 goto bad;
1035
1036         if (count)
1037                 *count = i + 1;
1038         return 0;
1039
1040 bad:
1041         pr_err("parse_ips bad ip '%s'\n", c);
1042         return -EINVAL;
1043 }
1044
1045 static int process_banner(struct ceph_connection *con)
1046 {
1047         dout("process_banner on %p\n", con);
1048
1049         if (verify_hello(con) < 0)
1050                 return -1;
1051
1052         ceph_decode_addr(&con->actual_peer_addr);
1053         ceph_decode_addr(&con->peer_addr_for_me);
1054
1055         /*
1056          * Make sure the other end is who we wanted.  note that the other
1057          * end may not yet know their ip address, so if it's 0.0.0.0, give
1058          * them the benefit of the doubt.
1059          */
1060         if (memcmp(&con->peer_addr, &con->actual_peer_addr,
1061                    sizeof(con->peer_addr)) != 0 &&
1062             !(addr_is_blank(&con->actual_peer_addr.in_addr) &&
1063               con->actual_peer_addr.nonce == con->peer_addr.nonce)) {
1064                 pr_warning("wrong peer, want %s/%lld, got %s/%lld\n",
1065                            pr_addr(&con->peer_addr.in_addr),
1066                            le64_to_cpu(con->peer_addr.nonce),
1067                            pr_addr(&con->actual_peer_addr.in_addr),
1068                            le64_to_cpu(con->actual_peer_addr.nonce));
1069                 con->error_msg = "wrong peer at address";
1070                 return -1;
1071         }
1072
1073         /*
1074          * did we learn our address?
1075          */
1076         if (addr_is_blank(&con->msgr->inst.addr.in_addr)) {
1077                 int port = addr_port(&con->msgr->inst.addr.in_addr);
1078
1079                 memcpy(&con->msgr->inst.addr.in_addr,
1080                        &con->peer_addr_for_me.in_addr,
1081                        sizeof(con->peer_addr_for_me.in_addr));
1082                 addr_set_port(&con->msgr->inst.addr.in_addr, port);
1083                 encode_my_addr(con->msgr);
1084                 dout("process_banner learned my addr is %s\n",
1085                      pr_addr(&con->msgr->inst.addr.in_addr));
1086         }
1087
1088         set_bit(NEGOTIATING, &con->state);
1089         prepare_read_connect(con);
1090         return 0;
1091 }
1092
1093 static void fail_protocol(struct ceph_connection *con)
1094 {
1095         reset_connection(con);
1096         set_bit(CLOSED, &con->state);  /* in case there's queued work */
1097
1098         mutex_unlock(&con->mutex);
1099         if (con->ops->bad_proto)
1100                 con->ops->bad_proto(con);
1101         mutex_lock(&con->mutex);
1102 }
1103
1104 static int process_connect(struct ceph_connection *con)
1105 {
1106         u64 sup_feat = CEPH_FEATURE_SUPPORTED;
1107         u64 req_feat = CEPH_FEATURE_REQUIRED;
1108         u64 server_feat = le64_to_cpu(con->in_reply.features);
1109
1110         dout("process_connect on %p tag %d\n", con, (int)con->in_tag);
1111
1112         switch (con->in_reply.tag) {
1113         case CEPH_MSGR_TAG_FEATURES:
1114                 pr_err("%s%lld %s feature set mismatch,"
1115                        " my %llx < server's %llx, missing %llx\n",
1116                        ENTITY_NAME(con->peer_name),
1117                        pr_addr(&con->peer_addr.in_addr),
1118                        sup_feat, server_feat, server_feat & ~sup_feat);
1119                 con->error_msg = "missing required protocol features";
1120                 fail_protocol(con);
1121                 return -1;
1122
1123         case CEPH_MSGR_TAG_BADPROTOVER:
1124                 pr_err("%s%lld %s protocol version mismatch,"
1125                        " my %d != server's %d\n",
1126                        ENTITY_NAME(con->peer_name),
1127                        pr_addr(&con->peer_addr.in_addr),
1128                        le32_to_cpu(con->out_connect.protocol_version),
1129                        le32_to_cpu(con->in_reply.protocol_version));
1130                 con->error_msg = "protocol version mismatch";
1131                 fail_protocol(con);
1132                 return -1;
1133
1134         case CEPH_MSGR_TAG_BADAUTHORIZER:
1135                 con->auth_retry++;
1136                 dout("process_connect %p got BADAUTHORIZER attempt %d\n", con,
1137                      con->auth_retry);
1138                 if (con->auth_retry == 2) {
1139                         con->error_msg = "connect authorization failure";
1140                         reset_connection(con);
1141                         set_bit(CLOSED, &con->state);
1142                         return -1;
1143                 }
1144                 con->auth_retry = 1;
1145                 prepare_write_connect(con->msgr, con, 0);
1146                 prepare_read_connect_retry(con);
1147                 break;
1148
1149         case CEPH_MSGR_TAG_RESETSESSION:
1150                 /*
1151                  * If we connected with a large connect_seq but the peer
1152                  * has no record of a session with us (no connection, or
1153                  * connect_seq == 0), they will send RESETSESION to indicate
1154                  * that they must have reset their session, and may have
1155                  * dropped messages.
1156                  */
1157                 dout("process_connect got RESET peer seq %u\n",
1158                      le32_to_cpu(con->in_connect.connect_seq));
1159                 pr_err("%s%lld %s connection reset\n",
1160                        ENTITY_NAME(con->peer_name),
1161                        pr_addr(&con->peer_addr.in_addr));
1162                 reset_connection(con);
1163                 prepare_write_connect(con->msgr, con, 0);
1164                 prepare_read_connect(con);
1165
1166                 /* Tell ceph about it. */
1167                 mutex_unlock(&con->mutex);
1168                 pr_info("reset on %s%lld\n", ENTITY_NAME(con->peer_name));
1169                 if (con->ops->peer_reset)
1170                         con->ops->peer_reset(con);
1171                 mutex_lock(&con->mutex);
1172                 break;
1173
1174         case CEPH_MSGR_TAG_RETRY_SESSION:
1175                 /*
1176                  * If we sent a smaller connect_seq than the peer has, try
1177                  * again with a larger value.
1178                  */
1179                 dout("process_connect got RETRY my seq = %u, peer_seq = %u\n",
1180                      le32_to_cpu(con->out_connect.connect_seq),
1181                      le32_to_cpu(con->in_connect.connect_seq));
1182                 con->connect_seq = le32_to_cpu(con->in_connect.connect_seq);
1183                 prepare_write_connect(con->msgr, con, 0);
1184                 prepare_read_connect(con);
1185                 break;
1186
1187         case CEPH_MSGR_TAG_RETRY_GLOBAL:
1188                 /*
1189                  * If we sent a smaller global_seq than the peer has, try
1190                  * again with a larger value.
1191                  */
1192                 dout("process_connect got RETRY_GLOBAL my %u peer_gseq %u\n",
1193                      con->peer_global_seq,
1194                      le32_to_cpu(con->in_connect.global_seq));
1195                 get_global_seq(con->msgr,
1196                                le32_to_cpu(con->in_connect.global_seq));
1197                 prepare_write_connect(con->msgr, con, 0);
1198                 prepare_read_connect(con);
1199                 break;
1200
1201         case CEPH_MSGR_TAG_READY:
1202                 if (req_feat & ~server_feat) {
1203                         pr_err("%s%lld %s protocol feature mismatch,"
1204                                " my required %llx > server's %llx, need %llx\n",
1205                                ENTITY_NAME(con->peer_name),
1206                                pr_addr(&con->peer_addr.in_addr),
1207                                req_feat, server_feat, req_feat & ~server_feat);
1208                         con->error_msg = "missing required protocol features";
1209                         fail_protocol(con);
1210                         return -1;
1211                 }
1212                 clear_bit(CONNECTING, &con->state);
1213                 con->peer_global_seq = le32_to_cpu(con->in_reply.global_seq);
1214                 con->connect_seq++;
1215                 dout("process_connect got READY gseq %d cseq %d (%d)\n",
1216                      con->peer_global_seq,
1217                      le32_to_cpu(con->in_reply.connect_seq),
1218                      con->connect_seq);
1219                 WARN_ON(con->connect_seq !=
1220                         le32_to_cpu(con->in_reply.connect_seq));
1221
1222                 if (con->in_reply.flags & CEPH_MSG_CONNECT_LOSSY)
1223                         set_bit(LOSSYTX, &con->state);
1224
1225                 prepare_read_tag(con);
1226                 break;
1227
1228         case CEPH_MSGR_TAG_WAIT:
1229                 /*
1230                  * If there is a connection race (we are opening
1231                  * connections to each other), one of us may just have
1232                  * to WAIT.  This shouldn't happen if we are the
1233                  * client.
1234                  */
1235                 pr_err("process_connect peer connecting WAIT\n");
1236
1237         default:
1238                 pr_err("connect protocol error, will retry\n");
1239                 con->error_msg = "protocol error, garbage tag during connect";
1240                 return -1;
1241         }
1242         return 0;
1243 }
1244
1245
1246 /*
1247  * read (part of) an ack
1248  */
1249 static int read_partial_ack(struct ceph_connection *con)
1250 {
1251         int to = 0;
1252
1253         return read_partial(con, &to, sizeof(con->in_temp_ack),
1254                             &con->in_temp_ack);
1255 }
1256
1257
1258 /*
1259  * We can finally discard anything that's been acked.
1260  */
1261 static void process_ack(struct ceph_connection *con)
1262 {
1263         struct ceph_msg *m;
1264         u64 ack = le64_to_cpu(con->in_temp_ack);
1265         u64 seq;
1266
1267         while (!list_empty(&con->out_sent)) {
1268                 m = list_first_entry(&con->out_sent, struct ceph_msg,
1269                                      list_head);
1270                 seq = le64_to_cpu(m->hdr.seq);
1271                 if (seq > ack)
1272                         break;
1273                 dout("got ack for seq %llu type %d at %p\n", seq,
1274                      le16_to_cpu(m->hdr.type), m);
1275                 ceph_msg_remove(m);
1276         }
1277         prepare_read_tag(con);
1278 }
1279
1280
1281
1282
1283 static int read_partial_message_section(struct ceph_connection *con,
1284                                         struct kvec *section, unsigned int sec_len,
1285                                         u32 *crc)
1286 {
1287         int left;
1288         int ret;
1289
1290         BUG_ON(!section);
1291
1292         while (section->iov_len < sec_len) {
1293                 BUG_ON(section->iov_base == NULL);
1294                 left = sec_len - section->iov_len;
1295                 ret = ceph_tcp_recvmsg(con->sock, (char *)section->iov_base +
1296                                        section->iov_len, left);
1297                 if (ret <= 0)
1298                         return ret;
1299                 section->iov_len += ret;
1300                 if (section->iov_len == sec_len)
1301                         *crc = crc32c(0, section->iov_base,
1302                                       section->iov_len);
1303         }
1304
1305         return 1;
1306 }
1307
1308 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
1309                                 struct ceph_msg_header *hdr,
1310                                 int *skip);
1311 /*
1312  * read (part of) a message.
1313  */
1314 static int read_partial_message(struct ceph_connection *con)
1315 {
1316         struct ceph_msg *m = con->in_msg;
1317         void *p;
1318         int ret;
1319         int to, left;
1320         unsigned front_len, middle_len, data_len, data_off;
1321         int datacrc = con->msgr->nocrc;
1322         int skip;
1323
1324         dout("read_partial_message con %p msg %p\n", con, m);
1325
1326         /* header */
1327         while (con->in_base_pos < sizeof(con->in_hdr)) {
1328                 left = sizeof(con->in_hdr) - con->in_base_pos;
1329                 ret = ceph_tcp_recvmsg(con->sock,
1330                                        (char *)&con->in_hdr + con->in_base_pos,
1331                                        left);
1332                 if (ret <= 0)
1333                         return ret;
1334                 con->in_base_pos += ret;
1335                 if (con->in_base_pos == sizeof(con->in_hdr)) {
1336                         u32 crc = crc32c(0, (void *)&con->in_hdr,
1337                                  sizeof(con->in_hdr) - sizeof(con->in_hdr.crc));
1338                         if (crc != le32_to_cpu(con->in_hdr.crc)) {
1339                                 pr_err("read_partial_message bad hdr "
1340                                        " crc %u != expected %u\n",
1341                                        crc, con->in_hdr.crc);
1342                                 return -EBADMSG;
1343                         }
1344                 }
1345         }
1346         front_len = le32_to_cpu(con->in_hdr.front_len);
1347         if (front_len > CEPH_MSG_MAX_FRONT_LEN)
1348                 return -EIO;
1349         middle_len = le32_to_cpu(con->in_hdr.middle_len);
1350         if (middle_len > CEPH_MSG_MAX_DATA_LEN)
1351                 return -EIO;
1352         data_len = le32_to_cpu(con->in_hdr.data_len);
1353         if (data_len > CEPH_MSG_MAX_DATA_LEN)
1354                 return -EIO;
1355         data_off = le16_to_cpu(con->in_hdr.data_off);
1356
1357         /* allocate message? */
1358         if (!con->in_msg) {
1359                 dout("got hdr type %d front %d data %d\n", con->in_hdr.type,
1360                      con->in_hdr.front_len, con->in_hdr.data_len);
1361                 con->in_msg = ceph_alloc_msg(con, &con->in_hdr, &skip);
1362                 if (skip) {
1363                         /* skip this message */
1364                         dout("alloc_msg returned NULL, skipping message\n");
1365                         con->in_base_pos = -front_len - middle_len - data_len -
1366                                 sizeof(m->footer);
1367                         con->in_tag = CEPH_MSGR_TAG_READY;
1368                         return 0;
1369                 }
1370                 if (IS_ERR(con->in_msg)) {
1371                         ret = PTR_ERR(con->in_msg);
1372                         con->in_msg = NULL;
1373                         con->error_msg =
1374                                 "error allocating memory for incoming message";
1375                         return ret;
1376                 }
1377                 m = con->in_msg;
1378                 m->front.iov_len = 0;    /* haven't read it yet */
1379                 if (m->middle)
1380                         m->middle->vec.iov_len = 0;
1381
1382                 con->in_msg_pos.page = 0;
1383                 con->in_msg_pos.page_pos = data_off & ~PAGE_MASK;
1384                 con->in_msg_pos.data_pos = 0;
1385         }
1386
1387         /* front */
1388         ret = read_partial_message_section(con, &m->front, front_len,
1389                                            &con->in_front_crc);
1390         if (ret <= 0)
1391                 return ret;
1392
1393         /* middle */
1394         if (m->middle) {
1395                 ret = read_partial_message_section(con, &m->middle->vec, middle_len,
1396                                                    &con->in_middle_crc);
1397                 if (ret <= 0)
1398                         return ret;
1399         }
1400
1401         /* (page) data */
1402         while (con->in_msg_pos.data_pos < data_len) {
1403                 left = min((int)(data_len - con->in_msg_pos.data_pos),
1404                            (int)(PAGE_SIZE - con->in_msg_pos.page_pos));
1405                 BUG_ON(m->pages == NULL);
1406                 p = kmap(m->pages[con->in_msg_pos.page]);
1407                 ret = ceph_tcp_recvmsg(con->sock, p + con->in_msg_pos.page_pos,
1408                                        left);
1409                 if (ret > 0 && datacrc)
1410                         con->in_data_crc =
1411                                 crc32c(con->in_data_crc,
1412                                           p + con->in_msg_pos.page_pos, ret);
1413                 kunmap(m->pages[con->in_msg_pos.page]);
1414                 if (ret <= 0)
1415                         return ret;
1416                 con->in_msg_pos.data_pos += ret;
1417                 con->in_msg_pos.page_pos += ret;
1418                 if (con->in_msg_pos.page_pos == PAGE_SIZE) {
1419                         con->in_msg_pos.page_pos = 0;
1420                         con->in_msg_pos.page++;
1421                 }
1422         }
1423
1424         /* footer */
1425         to = sizeof(m->hdr) + sizeof(m->footer);
1426         while (con->in_base_pos < to) {
1427                 left = to - con->in_base_pos;
1428                 ret = ceph_tcp_recvmsg(con->sock, (char *)&m->footer +
1429                                        (con->in_base_pos - sizeof(m->hdr)),
1430                                        left);
1431                 if (ret <= 0)
1432                         return ret;
1433                 con->in_base_pos += ret;
1434         }
1435         dout("read_partial_message got msg %p %d (%u) + %d (%u) + %d (%u)\n",
1436              m, front_len, m->footer.front_crc, middle_len,
1437              m->footer.middle_crc, data_len, m->footer.data_crc);
1438
1439         /* crc ok? */
1440         if (con->in_front_crc != le32_to_cpu(m->footer.front_crc)) {
1441                 pr_err("read_partial_message %p front crc %u != exp. %u\n",
1442                        m, con->in_front_crc, m->footer.front_crc);
1443                 return -EBADMSG;
1444         }
1445         if (con->in_middle_crc != le32_to_cpu(m->footer.middle_crc)) {
1446                 pr_err("read_partial_message %p middle crc %u != exp %u\n",
1447                        m, con->in_middle_crc, m->footer.middle_crc);
1448                 return -EBADMSG;
1449         }
1450         if (datacrc &&
1451             (m->footer.flags & CEPH_MSG_FOOTER_NOCRC) == 0 &&
1452             con->in_data_crc != le32_to_cpu(m->footer.data_crc)) {
1453                 pr_err("read_partial_message %p data crc %u != exp. %u\n", m,
1454                        con->in_data_crc, le32_to_cpu(m->footer.data_crc));
1455                 return -EBADMSG;
1456         }
1457
1458         return 1; /* done! */
1459 }
1460
1461 /*
1462  * Process message.  This happens in the worker thread.  The callback should
1463  * be careful not to do anything that waits on other incoming messages or it
1464  * may deadlock.
1465  */
1466 static void process_message(struct ceph_connection *con)
1467 {
1468         struct ceph_msg *msg;
1469
1470         msg = con->in_msg;
1471         con->in_msg = NULL;
1472
1473         /* if first message, set peer_name */
1474         if (con->peer_name.type == 0)
1475                 con->peer_name = msg->hdr.src.name;
1476
1477         con->in_seq++;
1478         mutex_unlock(&con->mutex);
1479
1480         dout("===== %p %llu from %s%lld %d=%s len %d+%d (%u %u %u) =====\n",
1481              msg, le64_to_cpu(msg->hdr.seq),
1482              ENTITY_NAME(msg->hdr.src.name),
1483              le16_to_cpu(msg->hdr.type),
1484              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1485              le32_to_cpu(msg->hdr.front_len),
1486              le32_to_cpu(msg->hdr.data_len),
1487              con->in_front_crc, con->in_middle_crc, con->in_data_crc);
1488         con->ops->dispatch(con, msg);
1489
1490         mutex_lock(&con->mutex);
1491         prepare_read_tag(con);
1492 }
1493
1494
1495 /*
1496  * Write something to the socket.  Called in a worker thread when the
1497  * socket appears to be writeable and we have something ready to send.
1498  */
1499 static int try_write(struct ceph_connection *con)
1500 {
1501         struct ceph_messenger *msgr = con->msgr;
1502         int ret = 1;
1503
1504         dout("try_write start %p state %lu nref %d\n", con, con->state,
1505              atomic_read(&con->nref));
1506
1507         mutex_lock(&con->mutex);
1508 more:
1509         dout("try_write out_kvec_bytes %d\n", con->out_kvec_bytes);
1510
1511         /* open the socket first? */
1512         if (con->sock == NULL) {
1513                 /*
1514                  * if we were STANDBY and are reconnecting _this_
1515                  * connection, bump connect_seq now.  Always bump
1516                  * global_seq.
1517                  */
1518                 if (test_and_clear_bit(STANDBY, &con->state))
1519                         con->connect_seq++;
1520
1521                 prepare_write_banner(msgr, con);
1522                 prepare_write_connect(msgr, con, 1);
1523                 prepare_read_banner(con);
1524                 set_bit(CONNECTING, &con->state);
1525                 clear_bit(NEGOTIATING, &con->state);
1526
1527                 BUG_ON(con->in_msg);
1528                 con->in_tag = CEPH_MSGR_TAG_READY;
1529                 dout("try_write initiating connect on %p new state %lu\n",
1530                      con, con->state);
1531                 con->sock = ceph_tcp_connect(con);
1532                 if (IS_ERR(con->sock)) {
1533                         con->sock = NULL;
1534                         con->error_msg = "connect error";
1535                         ret = -1;
1536                         goto out;
1537                 }
1538         }
1539
1540 more_kvec:
1541         /* kvec data queued? */
1542         if (con->out_skip) {
1543                 ret = write_partial_skip(con);
1544                 if (ret <= 0)
1545                         goto done;
1546                 if (ret < 0) {
1547                         dout("try_write write_partial_skip err %d\n", ret);
1548                         goto done;
1549                 }
1550         }
1551         if (con->out_kvec_left) {
1552                 ret = write_partial_kvec(con);
1553                 if (ret <= 0)
1554                         goto done;
1555         }
1556
1557         /* msg pages? */
1558         if (con->out_msg) {
1559                 if (con->out_msg_done) {
1560                         ceph_msg_put(con->out_msg);
1561                         con->out_msg = NULL;   /* we're done with this one */
1562                         goto do_next;
1563                 }
1564
1565                 ret = write_partial_msg_pages(con);
1566                 if (ret == 1)
1567                         goto more_kvec;  /* we need to send the footer, too! */
1568                 if (ret == 0)
1569                         goto done;
1570                 if (ret < 0) {
1571                         dout("try_write write_partial_msg_pages err %d\n",
1572                              ret);
1573                         goto done;
1574                 }
1575         }
1576
1577 do_next:
1578         if (!test_bit(CONNECTING, &con->state)) {
1579                 /* is anything else pending? */
1580                 if (!list_empty(&con->out_queue)) {
1581                         prepare_write_message(con);
1582                         goto more;
1583                 }
1584                 if (con->in_seq > con->in_seq_acked) {
1585                         prepare_write_ack(con);
1586                         goto more;
1587                 }
1588                 if (test_and_clear_bit(KEEPALIVE_PENDING, &con->state)) {
1589                         prepare_write_keepalive(con);
1590                         goto more;
1591                 }
1592         }
1593
1594         /* Nothing to do! */
1595         clear_bit(WRITE_PENDING, &con->state);
1596         dout("try_write nothing else to write.\n");
1597 done:
1598         ret = 0;
1599 out:
1600         mutex_unlock(&con->mutex);
1601         dout("try_write done on %p\n", con);
1602         return ret;
1603 }
1604
1605
1606
1607 /*
1608  * Read what we can from the socket.
1609  */
1610 static int try_read(struct ceph_connection *con)
1611 {
1612         struct ceph_messenger *msgr;
1613         int ret = -1;
1614
1615         if (!con->sock)
1616                 return 0;
1617
1618         if (test_bit(STANDBY, &con->state))
1619                 return 0;
1620
1621         dout("try_read start on %p\n", con);
1622         msgr = con->msgr;
1623
1624         mutex_lock(&con->mutex);
1625
1626 more:
1627         dout("try_read tag %d in_base_pos %d\n", (int)con->in_tag,
1628              con->in_base_pos);
1629         if (test_bit(CONNECTING, &con->state)) {
1630                 if (!test_bit(NEGOTIATING, &con->state)) {
1631                         dout("try_read connecting\n");
1632                         ret = read_partial_banner(con);
1633                         if (ret <= 0)
1634                                 goto done;
1635                         if (process_banner(con) < 0) {
1636                                 ret = -1;
1637                                 goto out;
1638                         }
1639                 }
1640                 ret = read_partial_connect(con);
1641                 if (ret <= 0)
1642                         goto done;
1643                 if (process_connect(con) < 0) {
1644                         ret = -1;
1645                         goto out;
1646                 }
1647                 goto more;
1648         }
1649
1650         if (con->in_base_pos < 0) {
1651                 /*
1652                  * skipping + discarding content.
1653                  *
1654                  * FIXME: there must be a better way to do this!
1655                  */
1656                 static char buf[1024];
1657                 int skip = min(1024, -con->in_base_pos);
1658                 dout("skipping %d / %d bytes\n", skip, -con->in_base_pos);
1659                 ret = ceph_tcp_recvmsg(con->sock, buf, skip);
1660                 if (ret <= 0)
1661                         goto done;
1662                 con->in_base_pos += ret;
1663                 if (con->in_base_pos)
1664                         goto more;
1665         }
1666         if (con->in_tag == CEPH_MSGR_TAG_READY) {
1667                 /*
1668                  * what's next?
1669                  */
1670                 ret = ceph_tcp_recvmsg(con->sock, &con->in_tag, 1);
1671                 if (ret <= 0)
1672                         goto done;
1673                 dout("try_read got tag %d\n", (int)con->in_tag);
1674                 switch (con->in_tag) {
1675                 case CEPH_MSGR_TAG_MSG:
1676                         prepare_read_message(con);
1677                         break;
1678                 case CEPH_MSGR_TAG_ACK:
1679                         prepare_read_ack(con);
1680                         break;
1681                 case CEPH_MSGR_TAG_CLOSE:
1682                         set_bit(CLOSED, &con->state);   /* fixme */
1683                         goto done;
1684                 default:
1685                         goto bad_tag;
1686                 }
1687         }
1688         if (con->in_tag == CEPH_MSGR_TAG_MSG) {
1689                 ret = read_partial_message(con);
1690                 if (ret <= 0) {
1691                         switch (ret) {
1692                         case -EBADMSG:
1693                                 con->error_msg = "bad crc";
1694                                 ret = -EIO;
1695                                 goto out;
1696                         case -EIO:
1697                                 con->error_msg = "io error";
1698                                 goto out;
1699                         default:
1700                                 goto done;
1701                         }
1702                 }
1703                 if (con->in_tag == CEPH_MSGR_TAG_READY)
1704                         goto more;
1705                 process_message(con);
1706                 goto more;
1707         }
1708         if (con->in_tag == CEPH_MSGR_TAG_ACK) {
1709                 ret = read_partial_ack(con);
1710                 if (ret <= 0)
1711                         goto done;
1712                 process_ack(con);
1713                 goto more;
1714         }
1715
1716 done:
1717         ret = 0;
1718 out:
1719         mutex_unlock(&con->mutex);
1720         dout("try_read done on %p\n", con);
1721         return ret;
1722
1723 bad_tag:
1724         pr_err("try_read bad con->in_tag = %d\n", (int)con->in_tag);
1725         con->error_msg = "protocol error, garbage tag";
1726         ret = -1;
1727         goto out;
1728 }
1729
1730
1731 /*
1732  * Atomically queue work on a connection.  Bump @con reference to
1733  * avoid races with connection teardown.
1734  *
1735  * There is some trickery going on with QUEUED and BUSY because we
1736  * only want a _single_ thread operating on each connection at any
1737  * point in time, but we want to use all available CPUs.
1738  *
1739  * The worker thread only proceeds if it can atomically set BUSY.  It
1740  * clears QUEUED and does it's thing.  When it thinks it's done, it
1741  * clears BUSY, then rechecks QUEUED.. if it's set again, it loops
1742  * (tries again to set BUSY).
1743  *
1744  * To queue work, we first set QUEUED, _then_ if BUSY isn't set, we
1745  * try to queue work.  If that fails (work is already queued, or BUSY)
1746  * we give up (work also already being done or is queued) but leave QUEUED
1747  * set so that the worker thread will loop if necessary.
1748  */
1749 static void queue_con(struct ceph_connection *con)
1750 {
1751         if (test_bit(DEAD, &con->state)) {
1752                 dout("queue_con %p ignoring: DEAD\n",
1753                      con);
1754                 return;
1755         }
1756
1757         if (!con->ops->get(con)) {
1758                 dout("queue_con %p ref count 0\n", con);
1759                 return;
1760         }
1761
1762         set_bit(QUEUED, &con->state);
1763         if (test_bit(BUSY, &con->state)) {
1764                 dout("queue_con %p - already BUSY\n", con);
1765                 con->ops->put(con);
1766         } else if (!queue_work(ceph_msgr_wq, &con->work.work)) {
1767                 dout("queue_con %p - already queued\n", con);
1768                 con->ops->put(con);
1769         } else {
1770                 dout("queue_con %p\n", con);
1771         }
1772 }
1773
1774 /*
1775  * Do some work on a connection.  Drop a connection ref when we're done.
1776  */
1777 static void con_work(struct work_struct *work)
1778 {
1779         struct ceph_connection *con = container_of(work, struct ceph_connection,
1780                                                    work.work);
1781         int backoff = 0;
1782
1783 more:
1784         if (test_and_set_bit(BUSY, &con->state) != 0) {
1785                 dout("con_work %p BUSY already set\n", con);
1786                 goto out;
1787         }
1788         dout("con_work %p start, clearing QUEUED\n", con);
1789         clear_bit(QUEUED, &con->state);
1790
1791         if (test_bit(CLOSED, &con->state)) { /* e.g. if we are replaced */
1792                 dout("con_work CLOSED\n");
1793                 con_close_socket(con);
1794                 goto done;
1795         }
1796         if (test_and_clear_bit(OPENING, &con->state)) {
1797                 /* reopen w/ new peer */
1798                 dout("con_work OPENING\n");
1799                 con_close_socket(con);
1800         }
1801
1802         if (test_and_clear_bit(SOCK_CLOSED, &con->state) ||
1803             try_read(con) < 0 ||
1804             try_write(con) < 0) {
1805                 backoff = 1;
1806                 ceph_fault(con);     /* error/fault path */
1807         }
1808
1809 done:
1810         clear_bit(BUSY, &con->state);
1811         dout("con->state=%lu\n", con->state);
1812         if (test_bit(QUEUED, &con->state)) {
1813                 if (!backoff || test_bit(OPENING, &con->state)) {
1814                         dout("con_work %p QUEUED reset, looping\n", con);
1815                         goto more;
1816                 }
1817                 dout("con_work %p QUEUED reset, but just faulted\n", con);
1818                 clear_bit(QUEUED, &con->state);
1819         }
1820         dout("con_work %p done\n", con);
1821
1822 out:
1823         con->ops->put(con);
1824 }
1825
1826
1827 /*
1828  * Generic error/fault handler.  A retry mechanism is used with
1829  * exponential backoff
1830  */
1831 static void ceph_fault(struct ceph_connection *con)
1832 {
1833         pr_err("%s%lld %s %s\n", ENTITY_NAME(con->peer_name),
1834                pr_addr(&con->peer_addr.in_addr), con->error_msg);
1835         dout("fault %p state %lu to peer %s\n",
1836              con, con->state, pr_addr(&con->peer_addr.in_addr));
1837
1838         if (test_bit(LOSSYTX, &con->state)) {
1839                 dout("fault on LOSSYTX channel\n");
1840                 goto out;
1841         }
1842
1843         clear_bit(BUSY, &con->state);  /* to avoid an improbable race */
1844
1845         mutex_lock(&con->mutex);
1846         if (test_bit(CLOSED, &con->state))
1847                 goto out_unlock;
1848
1849         con_close_socket(con);
1850
1851         if (con->in_msg) {
1852                 ceph_msg_put(con->in_msg);
1853                 con->in_msg = NULL;
1854         }
1855
1856         /* Requeue anything that hasn't been acked */
1857         list_splice_init(&con->out_sent, &con->out_queue);
1858
1859         /* If there are no messages in the queue, place the connection
1860          * in a STANDBY state (i.e., don't try to reconnect just yet). */
1861         if (list_empty(&con->out_queue) && !con->out_keepalive_pending) {
1862                 dout("fault setting STANDBY\n");
1863                 set_bit(STANDBY, &con->state);
1864         } else {
1865                 /* retry after a delay. */
1866                 if (con->delay == 0)
1867                         con->delay = BASE_DELAY_INTERVAL;
1868                 else if (con->delay < MAX_DELAY_INTERVAL)
1869                         con->delay *= 2;
1870                 dout("fault queueing %p delay %lu\n", con, con->delay);
1871                 con->ops->get(con);
1872                 if (queue_delayed_work(ceph_msgr_wq, &con->work,
1873                                        round_jiffies_relative(con->delay)) == 0)
1874                         con->ops->put(con);
1875         }
1876
1877 out_unlock:
1878         mutex_unlock(&con->mutex);
1879 out:
1880         /*
1881          * in case we faulted due to authentication, invalidate our
1882          * current tickets so that we can get new ones.
1883          */
1884         if (con->auth_retry && con->ops->invalidate_authorizer) {
1885                 dout("calling invalidate_authorizer()\n");
1886                 con->ops->invalidate_authorizer(con);
1887         }
1888
1889         if (con->ops->fault)
1890                 con->ops->fault(con);
1891 }
1892
1893
1894
1895 /*
1896  * create a new messenger instance
1897  */
1898 struct ceph_messenger *ceph_messenger_create(struct ceph_entity_addr *myaddr)
1899 {
1900         struct ceph_messenger *msgr;
1901
1902         msgr = kzalloc(sizeof(*msgr), GFP_KERNEL);
1903         if (msgr == NULL)
1904                 return ERR_PTR(-ENOMEM);
1905
1906         spin_lock_init(&msgr->global_seq_lock);
1907
1908         /* the zero page is needed if a request is "canceled" while the message
1909          * is being written over the socket */
1910         msgr->zero_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
1911         if (!msgr->zero_page) {
1912                 kfree(msgr);
1913                 return ERR_PTR(-ENOMEM);
1914         }
1915         kmap(msgr->zero_page);
1916
1917         if (myaddr)
1918                 msgr->inst.addr = *myaddr;
1919
1920         /* select a random nonce */
1921         msgr->inst.addr.type = 0;
1922         get_random_bytes(&msgr->inst.addr.nonce, sizeof(msgr->inst.addr.nonce));
1923         encode_my_addr(msgr);
1924
1925         dout("messenger_create %p\n", msgr);
1926         return msgr;
1927 }
1928
1929 void ceph_messenger_destroy(struct ceph_messenger *msgr)
1930 {
1931         dout("destroy %p\n", msgr);
1932         kunmap(msgr->zero_page);
1933         __free_page(msgr->zero_page);
1934         kfree(msgr);
1935         dout("destroyed messenger %p\n", msgr);
1936 }
1937
1938 /*
1939  * Queue up an outgoing message on the given connection.
1940  */
1941 void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg)
1942 {
1943         if (test_bit(CLOSED, &con->state)) {
1944                 dout("con_send %p closed, dropping %p\n", con, msg);
1945                 ceph_msg_put(msg);
1946                 return;
1947         }
1948
1949         /* set src+dst */
1950         msg->hdr.src.name = con->msgr->inst.name;
1951         msg->hdr.src.addr = con->msgr->my_enc_addr;
1952         msg->hdr.orig_src = msg->hdr.src;
1953
1954         /* queue */
1955         mutex_lock(&con->mutex);
1956         BUG_ON(!list_empty(&msg->list_head));
1957         list_add_tail(&msg->list_head, &con->out_queue);
1958         dout("----- %p to %s%lld %d=%s len %d+%d+%d -----\n", msg,
1959              ENTITY_NAME(con->peer_name), le16_to_cpu(msg->hdr.type),
1960              ceph_msg_type_name(le16_to_cpu(msg->hdr.type)),
1961              le32_to_cpu(msg->hdr.front_len),
1962              le32_to_cpu(msg->hdr.middle_len),
1963              le32_to_cpu(msg->hdr.data_len));
1964         mutex_unlock(&con->mutex);
1965
1966         /* if there wasn't anything waiting to send before, queue
1967          * new work */
1968         if (test_and_set_bit(WRITE_PENDING, &con->state) == 0)
1969                 queue_con(con);
1970 }
1971
1972 /*
1973  * Revoke a message that was previously queued for send
1974  */
1975 void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg)
1976 {
1977         mutex_lock(&con->mutex);
1978         if (!list_empty(&msg->list_head)) {
1979                 dout("con_revoke %p msg %p\n", con, msg);
1980                 list_del_init(&msg->list_head);
1981                 ceph_msg_put(msg);
1982                 msg->hdr.seq = 0;
1983                 if (con->out_msg == msg) {
1984                         ceph_msg_put(con->out_msg);
1985                         con->out_msg = NULL;
1986                 }
1987                 if (con->out_kvec_is_msg) {
1988                         con->out_skip = con->out_kvec_bytes;
1989                         con->out_kvec_is_msg = false;
1990                 }
1991         } else {
1992                 dout("con_revoke %p msg %p - not queued (sent?)\n", con, msg);
1993         }
1994         mutex_unlock(&con->mutex);
1995 }
1996
1997 /*
1998  * Revoke a message that we may be reading data into
1999  */
2000 void ceph_con_revoke_message(struct ceph_connection *con, struct ceph_msg *msg)
2001 {
2002         mutex_lock(&con->mutex);
2003         if (con->in_msg && con->in_msg == msg) {
2004                 unsigned front_len = le32_to_cpu(con->in_hdr.front_len);
2005                 unsigned middle_len = le32_to_cpu(con->in_hdr.middle_len);
2006                 unsigned data_len = le32_to_cpu(con->in_hdr.data_len);
2007
2008                 /* skip rest of message */
2009                 dout("con_revoke_pages %p msg %p revoked\n", con, msg);
2010                         con->in_base_pos = con->in_base_pos -
2011                                 sizeof(struct ceph_msg_header) -
2012                                 front_len -
2013                                 middle_len -
2014                                 data_len -
2015                                 sizeof(struct ceph_msg_footer);
2016                 ceph_msg_put(con->in_msg);
2017                 con->in_msg = NULL;
2018                 con->in_tag = CEPH_MSGR_TAG_READY;
2019         } else {
2020                 dout("con_revoke_pages %p msg %p pages %p no-op\n",
2021                      con, con->in_msg, msg);
2022         }
2023         mutex_unlock(&con->mutex);
2024 }
2025
2026 /*
2027  * Queue a keepalive byte to ensure the tcp connection is alive.
2028  */
2029 void ceph_con_keepalive(struct ceph_connection *con)
2030 {
2031         if (test_and_set_bit(KEEPALIVE_PENDING, &con->state) == 0 &&
2032             test_and_set_bit(WRITE_PENDING, &con->state) == 0)
2033                 queue_con(con);
2034 }
2035
2036
2037 /*
2038  * construct a new message with given type, size
2039  * the new msg has a ref count of 1.
2040  */
2041 struct ceph_msg *ceph_msg_new(int type, int front_len,
2042                               int page_len, int page_off, struct page **pages)
2043 {
2044         struct ceph_msg *m;
2045
2046         m = kmalloc(sizeof(*m), GFP_NOFS);
2047         if (m == NULL)
2048                 goto out;
2049         kref_init(&m->kref);
2050         INIT_LIST_HEAD(&m->list_head);
2051
2052         m->hdr.type = cpu_to_le16(type);
2053         m->hdr.front_len = cpu_to_le32(front_len);
2054         m->hdr.middle_len = 0;
2055         m->hdr.data_len = cpu_to_le32(page_len);
2056         m->hdr.data_off = cpu_to_le16(page_off);
2057         m->hdr.priority = cpu_to_le16(CEPH_MSG_PRIO_DEFAULT);
2058         m->footer.front_crc = 0;
2059         m->footer.middle_crc = 0;
2060         m->footer.data_crc = 0;
2061         m->front_max = front_len;
2062         m->front_is_vmalloc = false;
2063         m->more_to_follow = false;
2064         m->pool = NULL;
2065
2066         /* front */
2067         if (front_len) {
2068                 if (front_len > PAGE_CACHE_SIZE) {
2069                         m->front.iov_base = __vmalloc(front_len, GFP_NOFS,
2070                                                       PAGE_KERNEL);
2071                         m->front_is_vmalloc = true;
2072                 } else {
2073                         m->front.iov_base = kmalloc(front_len, GFP_NOFS);
2074                 }
2075                 if (m->front.iov_base == NULL) {
2076                         pr_err("msg_new can't allocate %d bytes\n",
2077                              front_len);
2078                         goto out2;
2079                 }
2080         } else {
2081                 m->front.iov_base = NULL;
2082         }
2083         m->front.iov_len = front_len;
2084
2085         /* middle */
2086         m->middle = NULL;
2087
2088         /* data */
2089         m->nr_pages = calc_pages_for(page_off, page_len);
2090         m->pages = pages;
2091         m->pagelist = NULL;
2092
2093         dout("ceph_msg_new %p page %d~%d -> %d\n", m, page_off, page_len,
2094              m->nr_pages);
2095         return m;
2096
2097 out2:
2098         ceph_msg_put(m);
2099 out:
2100         pr_err("msg_new can't create type %d len %d\n", type, front_len);
2101         return ERR_PTR(-ENOMEM);
2102 }
2103
2104 /*
2105  * Allocate "middle" portion of a message, if it is needed and wasn't
2106  * allocated by alloc_msg.  This allows us to read a small fixed-size
2107  * per-type header in the front and then gracefully fail (i.e.,
2108  * propagate the error to the caller based on info in the front) when
2109  * the middle is too large.
2110  */
2111 static int ceph_alloc_middle(struct ceph_connection *con, struct ceph_msg *msg)
2112 {
2113         int type = le16_to_cpu(msg->hdr.type);
2114         int middle_len = le32_to_cpu(msg->hdr.middle_len);
2115
2116         dout("alloc_middle %p type %d %s middle_len %d\n", msg, type,
2117              ceph_msg_type_name(type), middle_len);
2118         BUG_ON(!middle_len);
2119         BUG_ON(msg->middle);
2120
2121         msg->middle = ceph_buffer_new(middle_len, GFP_NOFS);
2122         if (!msg->middle)
2123                 return -ENOMEM;
2124         return 0;
2125 }
2126
2127 /*
2128  * Generic message allocator, for incoming messages.
2129  */
2130 static struct ceph_msg *ceph_alloc_msg(struct ceph_connection *con,
2131                                 struct ceph_msg_header *hdr,
2132                                 int *skip)
2133 {
2134         int type = le16_to_cpu(hdr->type);
2135         int front_len = le32_to_cpu(hdr->front_len);
2136         int middle_len = le32_to_cpu(hdr->middle_len);
2137         struct ceph_msg *msg = NULL;
2138         int ret;
2139
2140         if (con->ops->alloc_msg) {
2141                 mutex_unlock(&con->mutex);
2142                 msg = con->ops->alloc_msg(con, hdr, skip);
2143                 mutex_lock(&con->mutex);
2144                 if (IS_ERR(msg))
2145                         return msg;
2146
2147                 if (*skip)
2148                         return NULL;
2149         }
2150         if (!msg) {
2151                 *skip = 0;
2152                 msg = ceph_msg_new(type, front_len, 0, 0, NULL);
2153                 if (!msg) {
2154                         pr_err("unable to allocate msg type %d len %d\n",
2155                                type, front_len);
2156                         return ERR_PTR(-ENOMEM);
2157                 }
2158         }
2159         memcpy(&msg->hdr, &con->in_hdr, sizeof(con->in_hdr));
2160
2161         if (middle_len) {
2162                 ret = ceph_alloc_middle(con, msg);
2163
2164                 if (ret < 0) {
2165                         ceph_msg_put(msg);
2166                         return msg;
2167                 }
2168         }
2169
2170         return msg;
2171 }
2172
2173
2174 /*
2175  * Free a generically kmalloc'd message.
2176  */
2177 void ceph_msg_kfree(struct ceph_msg *m)
2178 {
2179         dout("msg_kfree %p\n", m);
2180         if (m->front_is_vmalloc)
2181                 vfree(m->front.iov_base);
2182         else
2183                 kfree(m->front.iov_base);
2184         kfree(m);
2185 }
2186
2187 /*
2188  * Drop a msg ref.  Destroy as needed.
2189  */
2190 void ceph_msg_last_put(struct kref *kref)
2191 {
2192         struct ceph_msg *m = container_of(kref, struct ceph_msg, kref);
2193
2194         dout("ceph_msg_put last one on %p\n", m);
2195         WARN_ON(!list_empty(&m->list_head));
2196
2197         /* drop middle, data, if any */
2198         if (m->middle) {
2199                 ceph_buffer_put(m->middle);
2200                 m->middle = NULL;
2201         }
2202         m->nr_pages = 0;
2203         m->pages = NULL;
2204
2205         if (m->pagelist) {
2206                 ceph_pagelist_release(m->pagelist);
2207                 kfree(m->pagelist);
2208                 m->pagelist = NULL;
2209         }
2210
2211         if (m->pool)
2212                 ceph_msgpool_put(m->pool, m);
2213         else
2214                 ceph_msg_kfree(m);
2215 }
2216
2217 void ceph_msg_dump(struct ceph_msg *msg)
2218 {
2219         pr_debug("msg_dump %p (front_max %d nr_pages %d)\n", msg,
2220                  msg->front_max, msg->nr_pages);
2221         print_hex_dump(KERN_DEBUG, "header: ",
2222                        DUMP_PREFIX_OFFSET, 16, 1,
2223                        &msg->hdr, sizeof(msg->hdr), true);
2224         print_hex_dump(KERN_DEBUG, " front: ",
2225                        DUMP_PREFIX_OFFSET, 16, 1,
2226                        msg->front.iov_base, msg->front.iov_len, true);
2227         if (msg->middle)
2228                 print_hex_dump(KERN_DEBUG, "middle: ",
2229                                DUMP_PREFIX_OFFSET, 16, 1,
2230                                msg->middle->vec.iov_base,
2231                                msg->middle->vec.iov_len, true);
2232         print_hex_dump(KERN_DEBUG, "footer: ",
2233                        DUMP_PREFIX_OFFSET, 16, 1,
2234                        &msg->footer, sizeof(msg->footer), true);
2235 }