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