net/sunrpc/xprt.c

   1 /*
   2  *  linux/net/sunrpc/xprt.c
   3  *
   4  *  This is a generic RPC call interface supporting congestion avoidance,
   5  *  and asynchronous calls.
   6  *
   7  *  The interface works like this:
   8  *
   9  *  -   When a process places a call, it allocates a request slot if
  10  *      one is available. Otherwise, it sleeps on the backlog queue
  11  *      (xprt_reserve).
  12  *  -   Next, the caller puts together the RPC message, stuffs it into
  13  *      the request struct, and calls xprt_call().
  14  *  -   xprt_call transmits the message and installs the caller on the
  15  *      socket's wait list. At the same time, it installs a timer that
  16  *      is run after the packet's timeout has expired.
  17  *  -   When a packet arrives, the data_ready handler walks the list of
  18  *      pending requests for that socket. If a matching XID is found, the
  19  *      caller is woken up, and the timer removed.
  20  *  -   When no reply arrives within the timeout interval, the timer is
  21  *      fired by the kernel and runs xprt_timer(). It either adjusts the
  22  *      timeout values (minor timeout) or wakes up the caller with a status
  23  *      of -ETIMEDOUT.
  24  *  -   When the caller receives a notification from RPC that a reply arrived,
  25  *      it should release the RPC slot, and process the reply.
  26  *      If the call timed out, it may choose to retry the operation by
  27  *      adjusting the initial timeout value, and simply calling rpc_call
  28  *      again.
  29  *
  30  *  Support for async RPC is done through a set of RPC-specific scheduling
  31  *  primitives that `transparently' work for processes as well as async
  32  *  tasks that rely on callbacks.
  33  *
  34  *  Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de>
  35  *
  36  *  TCP callback races fixes (C) 1998 Red Hat Software <alan@redhat.com>
  37  *  TCP send fixes (C) 1998 Red Hat Software <alan@redhat.com>
  38  *  TCP NFS related read + write fixes
  39  *   (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
  40  *
  41  *  Rewrite of larges part of the code in order to stabilize TCP stuff.
  42  *  Fix behaviour when socket buffer is full.
  43  *   (C) 1999 Trond Myklebust <trond.myklebust@fys.uio.no>
  44  */
  45
  46 #include <linux/types.h>
  47 #include <linux/slab.h>
  48 #include <linux/capability.h>
  49 #include <linux/sched.h>
  50 #include <linux/errno.h>
  51 #include <linux/socket.h>
  52 #include <linux/in.h>
  53 #include <linux/net.h>
  54 #include <linux/mm.h>
  55 #include <linux/udp.h>
  56 #include <linux/tcp.h>
  57 #include <linux/sunrpc/clnt.h>
  58 #include <linux/file.h>
  59 #include <linux/workqueue.h>
  60
  61 #include <net/sock.h>
  62 #include <net/checksum.h>
  63 #include <net/udp.h>
  64 #include <net/tcp.h>
  65
  66 /*
  67  * Local variables
  68  */
  69
  70 #ifdef RPC_DEBUG
  71 # undef  RPC_DEBUG_DATA
  72 # define RPCDBG_FACILITY        RPCDBG_XPRT
  73 #endif
  74
  75 #define XPRT_MAX_BACKOFF        (8)
  76 #define XPRT_IDLE_TIMEOUT       (5*60*HZ)
  77
  78 /*
  79  * Local functions
  80  */
  81 static void     xprt_request_init(struct rpc_task *, struct rpc_xprt *);
  82 static inline void      do_xprt_reserve(struct rpc_task *);
  83 static void     xprt_disconnect(struct rpc_xprt *);
  84 static void     xprt_connect_status(struct rpc_task *task);
  85 static struct rpc_xprt * xprt_setup(int proto, struct sockaddr_in *ap,
  86                                                 struct rpc_timeout *to);
  87 static struct socket *xprt_create_socket(int, struct rpc_timeout *, int);
  88 static void     xprt_bind_socket(struct rpc_xprt *, struct socket *);
  89 static int      __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
  90
  91 #ifdef RPC_DEBUG_DATA
  92 /*
  93  * Print the buffer contents (first 128 bytes only--just enough for
  94  * diropres return).
  95  */
  96 static void
  97 xprt_pktdump(char *msg, u32 *packet, unsigned int count)
  98 {
  99         u8      *buf = (u8 *) packet;
 100         int     j;
 101
 102         dprintk("RPC:      %s\n", msg);
 103         for (j = 0; j < count && j < 128; j += 4) {
 104                 if (!(j & 31)) {
 105                         if (j)
 106                                 dprintk("\n");
 107                         dprintk("0x%04x ", j);
 108                 }
 109                 dprintk("%02x%02x%02x%02x ",
 110                         buf[j], buf[j+1], buf[j+2], buf[j+3]);
 111         }
 112         dprintk("\n");
 113 }
 114 #else
 115 static inline void
 116 xprt_pktdump(char *msg, u32 *packet, unsigned int count)
 117 {
 118         /* NOP */
 119 }
 120 #endif
 121
 122 /*
 123  * Look up RPC transport given an INET socket
 124  */
 125 static inline struct rpc_xprt *
 126 xprt_from_sock(struct sock *sk)
 127 {
 128         return (struct rpc_xprt *) sk->sk_user_data;
 129 }
 130
 131 /*
 132  * Serialize write access to sockets, in order to prevent different
 133  * requests from interfering with each other.
 134  * Also prevents TCP socket connects from colliding with writes.
 135  */
 136 static int
 137 __xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 138 {
 139         struct rpc_rqst *req = task->tk_rqstp;
 140
 141         if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate)) {
 142                 if (task == xprt->snd_task)
 143                         return 1;
 144                 if (task == NULL)
 145                         return 0;
 146                 goto out_sleep;
 147         }
 148         if (xprt->nocong || __xprt_get_cong(xprt, task)) {
 149                 xprt->snd_task = task;
 150                 if (req) {
 151                         req->rq_bytes_sent = 0;
 152                         req->rq_ntrans++;
 153                 }
 154                 return 1;
 155         }
 156         smp_mb__before_clear_bit();
 157         clear_bit(XPRT_LOCKED, &xprt->sockstate);
 158         smp_mb__after_clear_bit();
 159 out_sleep:
 160         dprintk("RPC: %4d failed to lock socket %p\n", task->tk_pid, xprt);
 161         task->tk_timeout = 0;
 162         task->tk_status = -EAGAIN;
 163         if (req && req->rq_ntrans)
 164                 rpc_sleep_on(&xprt->resend, task, NULL, NULL);
 165         else
 166                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
 167         return 0;
 168 }
 169
 170 static inline int
 171 xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 172 {
 173         int retval;
 174
 175         spin_lock_bh(&xprt->sock_lock);
 176         retval = __xprt_lock_write(xprt, task);
 177         spin_unlock_bh(&xprt->sock_lock);
 178         return retval;
 179 }
 180
 181
 182 static void
 183 __xprt_lock_write_next(struct rpc_xprt *xprt)
 184 {
 185         struct rpc_task *task;
 186
 187         if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
 188                 return;
 189         if (!xprt->nocong && RPCXPRT_CONGESTED(xprt))
 190                 goto out_unlock;
 191         task = rpc_wake_up_next(&xprt->resend);
 192         if (!task) {
 193                 task = rpc_wake_up_next(&xprt->sending);
 194                 if (!task)
 195                         goto out_unlock;
 196         }
 197         if (xprt->nocong || __xprt_get_cong(xprt, task)) {
 198                 struct rpc_rqst *req = task->tk_rqstp;
 199                 xprt->snd_task = task;
 200                 if (req) {
 201                         req->rq_bytes_sent = 0;
 202                         req->rq_ntrans++;
 203                 }
 204                 return;
 205         }
 206 out_unlock:
 207         smp_mb__before_clear_bit();
 208         clear_bit(XPRT_LOCKED, &xprt->sockstate);
 209         smp_mb__after_clear_bit();
 210 }
 211
 212 /*
 213  * Releases the socket for use by other requests.
 214  */
 215 static void
 216 __xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 217 {
 218         if (xprt->snd_task == task) {
 219                 xprt->snd_task = NULL;
 220                 smp_mb__before_clear_bit();
 221                 clear_bit(XPRT_LOCKED, &xprt->sockstate);
 222                 smp_mb__after_clear_bit();
 223                 __xprt_lock_write_next(xprt);
 224         }
 225 }
 226
 227 static inline void
 228 xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 229 {
 230         spin_lock_bh(&xprt->sock_lock);
 231         __xprt_release_write(xprt, task);
 232         spin_unlock_bh(&xprt->sock_lock);
 233 }
 234
 235 /*
 236  * Write data to socket.
 237  */
 238 static inline int
 239 xprt_sendmsg(struct rpc_xprt *xprt, struct rpc_rqst *req)
 240 {
 241         struct socket   *sock = xprt->sock;
 242         struct xdr_buf  *xdr = &req->rq_snd_buf;
 243         struct sockaddr *addr = NULL;
 244         int addrlen = 0;
 245         unsigned int    skip;
 246         int             result;
 247
 248         if (!sock)
 249                 return -ENOTCONN;
 250
 251         xprt_pktdump("packet data:",
 252                                 req->rq_svec->iov_base,
 253                                 req->rq_svec->iov_len);
 254
 255         /* For UDP, we need to provide an address */
 256         if (!xprt->stream) {
 257                 addr = (struct sockaddr *) &xprt->addr;
 258                 addrlen = sizeof(xprt->addr);
 259         }
 260         /* Dont repeat bytes */
 261         skip = req->rq_bytes_sent;
 262
 263         clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 264         result = xdr_sendpages(sock, addr, addrlen, xdr, skip, MSG_DONTWAIT);
 265
 266         dprintk("RPC:      xprt_sendmsg(%d) = %d\n", xdr->len - skip, result);
 267
 268         if (result >= 0)
 269                 return result;
 270
 271         switch (result) {
 272         case -ECONNREFUSED:
 273                 /* When the server has died, an ICMP port unreachable message
 274                  * prompts ECONNREFUSED.
 275                  */
 276         case -EAGAIN:
 277                 break;
 278         case -ECONNRESET:
 279         case -ENOTCONN:
 280         case -EPIPE:
 281                 /* connection broken */
 282                 if (xprt->stream)
 283                         result = -ENOTCONN;
 284                 break;
 285         default:
 286                 printk(KERN_NOTICE "RPC: sendmsg returned error %d\n", -result);
 287         }
 288         return result;
 289 }
 290
 291 /*
 292  * Van Jacobson congestion avoidance. Check if the congestion window
 293  * overflowed. Put the task to sleep if this is the case.
 294  */
 295 static int
 296 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 297 {
 298         struct rpc_rqst *req = task->tk_rqstp;
 299
 300         if (req->rq_cong)
 301                 return 1;
 302         dprintk("RPC: %4d xprt_cwnd_limited cong = %ld cwnd = %ld\n",
 303                         task->tk_pid, xprt->cong, xprt->cwnd);
 304         if (RPCXPRT_CONGESTED(xprt))
 305                 return 0;
 306         req->rq_cong = 1;
 307         xprt->cong += RPC_CWNDSCALE;
 308         return 1;
 309 }
 310
 311 /*
 312  * Adjust the congestion window, and wake up the next task
 313  * that has been sleeping due to congestion
 314  */
 315 static void
 316 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 317 {
 318         if (!req->rq_cong)
 319                 return;
 320         req->rq_cong = 0;
 321         xprt->cong -= RPC_CWNDSCALE;
 322         __xprt_lock_write_next(xprt);
 323 }
 324
 325 /*
 326  * Adjust RPC congestion window
 327  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
 328  */
 329 static void
 330 xprt_adjust_cwnd(struct rpc_xprt *xprt, int result)
 331 {
 332         unsigned long   cwnd;
 333
 334         cwnd = xprt->cwnd;
 335         if (result >= 0 && cwnd <= xprt->cong) {
 336                 /* The (cwnd >> 1) term makes sure
 337                  * the result gets rounded properly. */
 338                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 339                 if (cwnd > RPC_MAXCWND(xprt))
 340                         cwnd = RPC_MAXCWND(xprt);
 341                 __xprt_lock_write_next(xprt);
 342         } else if (result == -ETIMEDOUT) {
 343                 cwnd >>= 1;
 344                 if (cwnd < RPC_CWNDSCALE)
 345                         cwnd = RPC_CWNDSCALE;
 346         }
 347         dprintk("RPC:      cong %ld, cwnd was %ld, now %ld\n",
 348                         xprt->cong, xprt->cwnd, cwnd);
 349         xprt->cwnd = cwnd;
 350 }
 351
 352 /*
 353  * Adjust timeout values etc for next retransmit
 354  */
 355 int
 356 xprt_adjust_timeout(struct rpc_timeout *to)
 357 {
 358         if (to->to_retries > 0) {
 359                 if (to->to_exponential)
 360                         to->to_current <<= 1;
 361                 else
 362                         to->to_current += to->to_increment;
 363                 if (to->to_maxval && to->to_current >= to->to_maxval)
 364                         to->to_current = to->to_maxval;
 365         } else {
 366                 if (to->to_exponential)
 367                         to->to_initval <<= 1;
 368                 else
 369                         to->to_initval += to->to_increment;
 370                 if (to->to_maxval && to->to_initval >= to->to_maxval)
 371                         to->to_initval = to->to_maxval;
 372                 to->to_current = to->to_initval;
 373         }
 374
 375         if (!to->to_current) {
 376                 printk(KERN_WARNING "xprt_adjust_timeout: to_current = 0!\n");
 377                 to->to_current = 5 * HZ;
 378         }
 379         pprintk("RPC: %lu %s\n", jiffies,
 380                         to->to_retries? "retrans" : "timeout");
 381         return to->to_retries-- > 0;
 382 }
 383
 384 /*
 385  * Close down a transport socket
 386  */
 387 static void
 388 xprt_close(struct rpc_xprt *xprt)
 389 {
 390         struct socket   *sock = xprt->sock;
 391         struct sock     *sk = xprt->inet;
 392
 393         if (!sk)
 394                 return;
 395
 396         write_lock_bh(&sk->sk_callback_lock);
 397         xprt->inet = NULL;
 398         xprt->sock = NULL;
 399
 400         sk->sk_user_data    = NULL;
 401         sk->sk_data_ready   = xprt->old_data_ready;
 402         sk->sk_state_change = xprt->old_state_change;
 403         sk->sk_write_space  = xprt->old_write_space;
 404         write_unlock_bh(&sk->sk_callback_lock);
 405
 406         xprt_disconnect(xprt);
 407         sk->sk_no_check  = 0;
 408
 409         sock_release(sock);
 410 }
 411
 412 static void
 413 xprt_socket_autoclose(void *args)
 414 {
 415         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
 416
 417         xprt_close(xprt);
 418         xprt_release_write(xprt, NULL);
 419 }
 420
 421 /*
 422  * Mark a transport as disconnected
 423  */
 424 static void
 425 xprt_disconnect(struct rpc_xprt *xprt)
 426 {
 427         dprintk("RPC:      disconnected transport %p\n", xprt);
 428         spin_lock_bh(&xprt->sock_lock);
 429         xprt_clear_connected(xprt);
 430         rpc_wake_up_status(&xprt->pending, -ENOTCONN);
 431         spin_unlock_bh(&xprt->sock_lock);
 432 }
 433
 434 /*
 435  * Used to allow disconnection when we've been idle
 436  */
 437 static void
 438 xprt_init_autodisconnect(unsigned long data)
 439 {
 440         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 441
 442         spin_lock(&xprt->sock_lock);
 443         if (!list_empty(&xprt->recv) || xprt->shutdown)
 444                 goto out_abort;
 445         if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
 446                 goto out_abort;
 447         spin_unlock(&xprt->sock_lock);
 448         /* Let keventd close the socket */
 449         schedule_work(&xprt->task_cleanup);
 450         return;
 451 out_abort:
 452         spin_unlock(&xprt->sock_lock);
 453 }
 454
 455 /*
 456  * Attempt to connect a TCP socket.
 457  *
 458  */
 459 void
 460 xprt_connect(struct rpc_task *task)
 461 {
 462         struct rpc_xprt *xprt = task->tk_xprt;
 463         struct socket   *sock = xprt->sock;
 464         struct sock     *inet;
 465         int             status;
 466
 467         dprintk("RPC: %4d xprt_connect xprt %p %s connected\n", task->tk_pid,
 468                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 469
 470         if (xprt->shutdown) {
 471                 task->tk_status = -EIO;
 472                 return;
 473         }
 474         if (!xprt->addr.sin_port) {
 475                 task->tk_status = -EIO;
 476                 return;
 477         }
 478         if (!xprt_lock_write(xprt, task))
 479                 return;
 480         if (xprt_connected(xprt))
 481                 goto out_write;
 482
 483         if (task->tk_rqstp)
 484                 task->tk_rqstp->rq_bytes_sent = 0;
 485
 486         /*
 487          * We're here because the xprt was marked disconnected.
 488          * Start by resetting any existing state.
 489          */
 490         xprt_close(xprt);
 491         if (!(sock = xprt_create_socket(xprt->prot, &xprt->timeout, xprt->resvport))) {
 492                 /* couldn't create socket or bind to reserved port;
 493                  * this is likely a permanent error, so cause an abort */
 494                 task->tk_status = -EIO;
 495                 goto out_write;
 496         }
 497         xprt_bind_socket(xprt, sock);
 498         xprt_sock_setbufsize(xprt);
 499
 500         if (!xprt->stream)
 501                 goto out_write;
 502
 503         inet = sock->sk;
 504
 505         /*
 506          * Tell the socket layer to start connecting...
 507          */
 508         status = sock->ops->connect(sock, (struct sockaddr *) &xprt->addr,
 509                                 sizeof(xprt->addr), O_NONBLOCK);
 510         dprintk("RPC: %4d  connect status %d connected %d sock state %d\n",
 511                 task->tk_pid, -status, xprt_connected(xprt), inet->sk_state);
 512
 513         if (status >= 0)
 514                 return;
 515
 516         switch (status) {
 517         case -EINPROGRESS:
 518         case -EALREADY:
 519                 /* Protect against TCP socket state changes */
 520                 lock_sock(inet);
 521                 if (inet->sk_state != TCP_ESTABLISHED) {
 522                         dprintk("RPC: %4d  waiting for connection\n",
 523                                         task->tk_pid);
 524                         task->tk_timeout = RPC_CONNECT_TIMEOUT;
 525                         /* if the socket is already closing, delay briefly */
 526                         if ((1 << inet->sk_state) &
 527                             ~(TCPF_SYN_SENT | TCPF_SYN_RECV))
 528                                 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
 529                         rpc_sleep_on(&xprt->pending, task, xprt_connect_status,
 530                                                                         NULL);
 531                 }
 532                 release_sock(inet);
 533                 break;
 534         case -ECONNREFUSED:
 535         case -ECONNRESET:
 536         case -ENOTCONN:
 537                 if (!RPC_IS_SOFT(task)) {
 538                         rpc_delay(task, RPC_REESTABLISH_TIMEOUT);
 539                         task->tk_status = -ENOTCONN;
 540                         break;
 541                 }
 542         default:
 543                 /* Report myriad other possible returns.  If this file
 544                  * system is soft mounted, just error out, like Solaris.  */
 545                 if (RPC_IS_SOFT(task)) {
 546                         printk(KERN_WARNING
 547                         "RPC: error %d connecting to server %s, exiting\n",
 548                                         -status, task->tk_client->cl_server);
 549                         task->tk_status = -EIO;
 550                         goto out_write;
 551                 }
 552                 printk(KERN_WARNING "RPC: error %d connecting to server %s\n",
 553                                 -status, task->tk_client->cl_server);
 554                 /* This will prevent anybody else from reconnecting */
 555                 rpc_delay(task, RPC_REESTABLISH_TIMEOUT);
 556                 task->tk_status = status;
 557                 break;
 558         }
 559         return;
 560  out_write:
 561         xprt_release_write(xprt, task);
 562 }
 563
 564 /*
 565  * We arrive here when awoken from waiting on connection establishment.
 566  */
 567 static void
 568 xprt_connect_status(struct rpc_task *task)
 569 {
 570         struct rpc_xprt *xprt = task->tk_xprt;
 571
 572         if (task->tk_status >= 0) {
 573                 dprintk("RPC: %4d xprt_connect_status: connection established\n",
 574                                 task->tk_pid);
 575                 return;
 576         }
 577
 578         /* if soft mounted, just cause this RPC to fail */
 579         if (RPC_IS_SOFT(task))
 580                 task->tk_status = -EIO;
 581
 582         switch (task->tk_status) {
 583         case -ENOTCONN:
 584                 rpc_delay(task, RPC_REESTABLISH_TIMEOUT);
 585                 return;
 586         case -ETIMEDOUT:
 587                 dprintk("RPC: %4d xprt_connect_status: timed out\n",
 588                                 task->tk_pid);
 589                 break;
 590         default:
 591                 printk(KERN_ERR "RPC: error %d connecting to server %s\n",
 592                                 -task->tk_status, task->tk_client->cl_server);
 593         }
 594         xprt_release_write(xprt, task);
 595 }
 596
 597 /*
 598  * Look up the RPC request corresponding to a reply, and then lock it.
 599  */
 600 static inline struct rpc_rqst *
 601 xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
 602 {
 603         struct list_head *pos;
 604         struct rpc_rqst *req = NULL;
 605
 606         list_for_each(pos, &xprt->recv) {
 607                 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
 608                 if (entry->rq_xid == xid) {
 609                         req = entry;
 610                         break;
 611                 }
 612         }
 613         return req;
 614 }
 615
 616 /*
 617  * Complete reply received.
 618  * The TCP code relies on us to remove the request from xprt->pending.
 619  */
 620 static void
 621 xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
 622 {
 623         struct rpc_task *task = req->rq_task;
 624         struct rpc_clnt *clnt = task->tk_client;
 625
 626         /* Adjust congestion window */
 627         if (!xprt->nocong) {
 628                 unsigned timer = task->tk_msg.rpc_proc->p_timer;
 629                 xprt_adjust_cwnd(xprt, copied);
 630                 __xprt_put_cong(xprt, req);
 631                 if (timer) {
 632                         if (req->rq_ntrans == 1)
 633                                 rpc_update_rtt(clnt->cl_rtt, timer,
 634                                                 (long)jiffies - req->rq_xtime);
 635                         rpc_set_timeo(clnt->cl_rtt, timer, req->rq_ntrans - 1);
 636                 }
 637         }
 638
 639 #ifdef RPC_PROFILE
 640         /* Profile only reads for now */
 641         if (copied > 1024) {
 642                 static unsigned long    nextstat;
 643                 static unsigned long    pkt_rtt, pkt_len, pkt_cnt;
 644
 645                 pkt_cnt++;
 646                 pkt_len += req->rq_slen + copied;
 647                 pkt_rtt += jiffies - req->rq_xtime;
 648                 if (time_before(nextstat, jiffies)) {
 649                         printk("RPC: %lu %ld cwnd\n", jiffies, xprt->cwnd);
 650                         printk("RPC: %ld %ld %ld %ld stat\n",
 651                                         jiffies, pkt_cnt, pkt_len, pkt_rtt);
 652                         pkt_rtt = pkt_len = pkt_cnt = 0;
 653                         nextstat = jiffies + 5 * HZ;
 654                 }
 655         }
 656 #endif
 657
 658         dprintk("RPC: %4d has input (%d bytes)\n", task->tk_pid, copied);
 659         req->rq_received = copied;
 660         list_del_init(&req->rq_list);
 661
 662         /* ... and wake up the process. */
 663         rpc_wake_up_task(task);
 664         return;
 665 }
 666
 667 static size_t
 668 skb_read_bits(skb_reader_t *desc, void *to, size_t len)
 669 {
 670         if (len > desc->count)
 671                 len = desc->count;
 672         if (skb_copy_bits(desc->skb, desc->offset, to, len))
 673                 return 0;
 674         desc->count -= len;
 675         desc->offset += len;
 676         return len;
 677 }
 678
 679 static size_t
 680 skb_read_and_csum_bits(skb_reader_t *desc, void *to, size_t len)
 681 {
 682         unsigned int csum2, pos;
 683
 684         if (len > desc->count)
 685                 len = desc->count;
 686         pos = desc->offset;
 687         csum2 = skb_copy_and_csum_bits(desc->skb, pos, to, len, 0);
 688         desc->csum = csum_block_add(desc->csum, csum2, pos);
 689         desc->count -= len;
 690         desc->offset += len;
 691         return len;
 692 }
 693
 694 /*
 695  * We have set things up such that we perform the checksum of the UDP
 696  * packet in parallel with the copies into the RPC client iovec.  -DaveM
 697  */
 698 int
 699 csum_partial_copy_to_xdr(struct xdr_buf *xdr, struct sk_buff *skb)
 700 {
 701         skb_reader_t desc;
 702
 703         desc.skb = skb;
 704         desc.offset = sizeof(struct udphdr);
 705         desc.count = skb->len - desc.offset;
 706
 707         if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 708                 goto no_checksum;
 709
 710         desc.csum = csum_partial(skb->data, desc.offset, skb->csum);
 711         xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_and_csum_bits);
 712         if (desc.offset != skb->len) {
 713                 unsigned int csum2;
 714                 csum2 = skb_checksum(skb, desc.offset, skb->len - desc.offset, 0);
 715                 desc.csum = csum_block_add(desc.csum, csum2, desc.offset);
 716         }
 717         if (desc.count)
 718                 return -1;
 719         if ((unsigned short)csum_fold(desc.csum))
 720                 return -1;
 721         return 0;
 722 no_checksum:
 723         xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_bits);
 724         if (desc.count)
 725                 return -1;
 726         return 0;
 727 }
 728
 729 /*
 730  * Input handler for RPC replies. Called from a bottom half and hence
 731  * atomic.
 732  */
 733 static void
 734 udp_data_ready(struct sock *sk, int len)
 735 {
 736         struct rpc_task *task;
 737         struct rpc_xprt *xprt;
 738         struct rpc_rqst *rovr;
 739         struct sk_buff  *skb;
 740         int err, repsize, copied;
 741         u32 xid;
 742
 743         read_lock(&sk->sk_callback_lock);
 744         dprintk("RPC:      udp_data_ready...\n");
 745         if (!(xprt = xprt_from_sock(sk))) {
 746                 printk("RPC:      udp_data_ready request not found!\n");
 747                 goto out;
 748         }
 749
 750         dprintk("RPC:      udp_data_ready client %p\n", xprt);
 751
 752         if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
 753                 goto out;
 754
 755         if (xprt->shutdown)
 756                 goto dropit;
 757
 758         repsize = skb->len - sizeof(struct udphdr);
 759         if (repsize < 4) {
 760                 printk("RPC: impossible RPC reply size %d!\n", repsize);
 761                 goto dropit;
 762         }
 763
 764         /* Copy the XID from the skb... */
 765         if (skb_copy_bits(skb, sizeof(struct udphdr), &xid, sizeof(xid)) < 0)
 766                 goto dropit;
 767
 768         /* Look up and lock the request corresponding to the given XID */
 769         spin_lock(&xprt->sock_lock);
 770         rovr = xprt_lookup_rqst(xprt, xid);
 771         if (!rovr)
 772                 goto out_unlock;
 773         task = rovr->rq_task;
 774
 775         dprintk("RPC: %4d received reply\n", task->tk_pid);
 776
 777         if ((copied = rovr->rq_private_buf.len) > repsize)
 778                 copied = repsize;
 779
 780         /* Suck it into the iovec, verify checksum if not done by hw. */
 781         if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
 782                 goto out_unlock;
 783
 784         /* Something worked... */
 785         dst_confirm(skb->dst);
 786
 787         xprt_complete_rqst(xprt, rovr, copied);
 788
 789  out_unlock:
 790         spin_unlock(&xprt->sock_lock);
 791  dropit:
 792         skb_free_datagram(sk, skb);
 793  out:
 794         if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
 795                 wake_up_interruptible(sk->sk_sleep);
 796         read_unlock(&sk->sk_callback_lock);
 797 }
 798
 799 /*
 800  * Copy from an skb into memory and shrink the skb.
 801  */
 802 static inline size_t
 803 tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
 804 {
 805         if (len > desc->count)
 806                 len = desc->count;
 807         if (skb_copy_bits(desc->skb, desc->offset, p, len))
 808                 return 0;
 809         desc->offset += len;
 810         desc->count -= len;
 811         return len;
 812 }
 813
 814 /*
 815  * TCP read fragment marker
 816  */
 817 static inline void
 818 tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
 819 {
 820         size_t len, used;
 821         char *p;
 822
 823         p = ((char *) &xprt->tcp_recm) + xprt->tcp_offset;
 824         len = sizeof(xprt->tcp_recm) - xprt->tcp_offset;
 825         used = tcp_copy_data(desc, p, len);
 826         xprt->tcp_offset += used;
 827         if (used != len)
 828                 return;
 829         xprt->tcp_reclen = ntohl(xprt->tcp_recm);
 830         if (xprt->tcp_reclen & 0x80000000)
 831                 xprt->tcp_flags |= XPRT_LAST_FRAG;
 832         else
 833                 xprt->tcp_flags &= ~XPRT_LAST_FRAG;
 834         xprt->tcp_reclen &= 0x7fffffff;
 835         xprt->tcp_flags &= ~XPRT_COPY_RECM;
 836         xprt->tcp_offset = 0;
 837         /* Sanity check of the record length */
 838         if (xprt->tcp_reclen < 4) {
 839                 printk(KERN_ERR "RPC: Invalid TCP record fragment length\n");
 840                 xprt_disconnect(xprt);
 841         }
 842         dprintk("RPC:      reading TCP record fragment of length %d\n",
 843                         xprt->tcp_reclen);
 844 }
 845
 846 static void
 847 tcp_check_recm(struct rpc_xprt *xprt)
 848 {
 849         if (xprt->tcp_offset == xprt->tcp_reclen) {
 850                 xprt->tcp_flags |= XPRT_COPY_RECM;
 851                 xprt->tcp_offset = 0;
 852                 if (xprt->tcp_flags & XPRT_LAST_FRAG) {
 853                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
 854                         xprt->tcp_flags |= XPRT_COPY_XID;
 855                         xprt->tcp_copied = 0;
 856                 }
 857         }
 858 }
 859
 860 /*
 861  * TCP read xid
 862  */
 863 static inline void
 864 tcp_read_xid(struct rpc_xprt *xprt, skb_reader_t *desc)
 865 {
 866         size_t len, used;
 867         char *p;
 868
 869         len = sizeof(xprt->tcp_xid) - xprt->tcp_offset;
 870         dprintk("RPC:      reading XID (%Zu bytes)\n", len);
 871         p = ((char *) &xprt->tcp_xid) + xprt->tcp_offset;
 872         used = tcp_copy_data(desc, p, len);
 873         xprt->tcp_offset += used;
 874         if (used != len)
 875                 return;
 876         xprt->tcp_flags &= ~XPRT_COPY_XID;
 877         xprt->tcp_flags |= XPRT_COPY_DATA;
 878         xprt->tcp_copied = 4;
 879         dprintk("RPC:      reading reply for XID %08x\n", xprt->tcp_xid);
 880         tcp_check_recm(xprt);
 881 }
 882
 883 /*
 884  * TCP read and complete request
 885  */
 886 static inline void
 887 tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
 888 {
 889         struct rpc_rqst *req;
 890         struct xdr_buf *rcvbuf;
 891         size_t len;
 892
 893         /* Find and lock the request corresponding to this xid */
 894         spin_lock(&xprt->sock_lock);
 895         req = xprt_lookup_rqst(xprt, xprt->tcp_xid);
 896         if (!req) {
 897                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 898                 dprintk("RPC:      XID %08x request not found!\n",
 899                                 xprt->tcp_xid);
 900                 spin_unlock(&xprt->sock_lock);
 901                 return;
 902         }
 903
 904         rcvbuf = &req->rq_private_buf;
 905         len = desc->count;
 906         if (len > xprt->tcp_reclen - xprt->tcp_offset) {
 907                 skb_reader_t my_desc;
 908
 909                 len = xprt->tcp_reclen - xprt->tcp_offset;
 910                 memcpy(&my_desc, desc, sizeof(my_desc));
 911                 my_desc.count = len;
 912                 xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
 913                                           &my_desc, tcp_copy_data);
 914                 desc->count -= len;
 915                 desc->offset += len;
 916         } else
 917                 xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
 918                                           desc, tcp_copy_data);
 919         xprt->tcp_copied += len;
 920         xprt->tcp_offset += len;
 921
 922         if (xprt->tcp_copied == req->rq_private_buf.len)
 923                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 924         else if (xprt->tcp_offset == xprt->tcp_reclen) {
 925                 if (xprt->tcp_flags & XPRT_LAST_FRAG)
 926                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
 927         }
 928
 929         if (!(xprt->tcp_flags & XPRT_COPY_DATA)) {
 930                 dprintk("RPC: %4d received reply complete\n",
 931                                 req->rq_task->tk_pid);
 932                 xprt_complete_rqst(xprt, req, xprt->tcp_copied);
 933         }
 934         spin_unlock(&xprt->sock_lock);
 935         tcp_check_recm(xprt);
 936 }
 937
 938 /*
 939  * TCP discard extra bytes from a short read
 940  */
 941 static inline void
 942 tcp_read_discard(struct rpc_xprt *xprt, skb_reader_t *desc)
 943 {
 944         size_t len;
 945
 946         len = xprt->tcp_reclen - xprt->tcp_offset;
 947         if (len > desc->count)
 948                 len = desc->count;
 949         desc->count -= len;
 950         desc->offset += len;
 951         xprt->tcp_offset += len;
 952         tcp_check_recm(xprt);
 953 }
 954
 955 /*
 956  * TCP record receive routine
 957  * We first have to grab the record marker, then the XID, then the data.
 958  */
 959 static int
 960 tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 961                 unsigned int offset, size_t len)
 962 {
 963         struct rpc_xprt *xprt = (struct rpc_xprt *)rd_desc->buf;
 964         skb_reader_t desc = {
 965                 .skb    = skb,
 966                 .offset = offset,
 967                 .count  = len,
 968                 .csum   = 0
 969         };
 970
 971         dprintk("RPC:      tcp_data_recv\n");
 972         do {
 973                 /* Read in a new fragment marker if necessary */
 974                 /* Can we ever really expect to get completely empty fragments? */
 975                 if (xprt->tcp_flags & XPRT_COPY_RECM) {
 976                         tcp_read_fraghdr(xprt, &desc);
 977                         continue;
 978                 }
 979                 /* Read in the xid if necessary */
 980                 if (xprt->tcp_flags & XPRT_COPY_XID) {
 981                         tcp_read_xid(xprt, &desc);
 982                         continue;
 983                 }
 984                 /* Read in the request data */
 985                 if (xprt->tcp_flags & XPRT_COPY_DATA) {
 986                         tcp_read_request(xprt, &desc);
 987                         continue;
 988                 }
 989                 /* Skip over any trailing bytes on short reads */
 990                 tcp_read_discard(xprt, &desc);
 991         } while (desc.count);
 992         dprintk("RPC:      tcp_data_recv done\n");
 993         return len - desc.count;
 994 }
 995
 996 static void tcp_data_ready(struct sock *sk, int bytes)
 997 {
 998         struct rpc_xprt *xprt;
 999         read_descriptor_t rd_desc;
1000
1001         read_lock(&sk->sk_callback_lock);
1002         dprintk("RPC:      tcp_data_ready...\n");
1003         if (!(xprt = xprt_from_sock(sk))) {
1004                 printk("RPC:      tcp_data_ready socket info not found!\n");
1005                 goto out;
1006         }
1007         if (xprt->shutdown)
1008                 goto out;
1009
1010         /* We use rd_desc to pass struct xprt to tcp_data_recv */
1011         rd_desc.buf = (char *)xprt;
1012         rd_desc.count = 65536;
1013         tcp_read_sock(sk, &rd_desc, tcp_data_recv);
1014 out:
1015         read_unlock(&sk->sk_callback_lock);
1016 }
1017
1018 static void
1019 tcp_state_change(struct sock *sk)
1020 {
1021         struct rpc_xprt *xprt;
1022
1023         read_lock(&sk->sk_callback_lock);
1024         if (!(xprt = xprt_from_sock(sk)))
1025                 goto out;
1026         dprintk("RPC:      tcp_state_change client %p...\n", xprt);
1027         dprintk("RPC:      state %x conn %d dead %d zapped %d\n",
1028                                 sk->sk_state, xprt_connected(xprt),
1029                                 sock_flag(sk, SOCK_DEAD), sk->sk_zapped);
1030
1031         switch (sk->sk_state) {
1032         case TCP_ESTABLISHED:
1033                 spin_lock_bh(&xprt->sock_lock);
1034                 if (!xprt_test_and_set_connected(xprt)) {
1035                         /* Reset TCP record info */
1036                         xprt->tcp_offset = 0;
1037                         xprt->tcp_reclen = 0;
1038                         xprt->tcp_copied = 0;
1039                         xprt->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
1040
1041                         if (xprt->snd_task)
1042                                 rpc_wake_up_task(xprt->snd_task);
1043                         rpc_wake_up(&xprt->pending);
1044                 }
1045                 spin_unlock_bh(&xprt->sock_lock);
1046                 break;
1047         case TCP_SYN_SENT:
1048         case TCP_SYN_RECV:
1049                 break;
1050         default:
1051                 xprt_disconnect(xprt);
1052                 break;
1053         }
1054  out:
1055         if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1056                 wake_up_interruptible_all(sk->sk_sleep);
1057         read_unlock(&sk->sk_callback_lock);
1058 }
1059
1060 /*
1061  * Called when more output buffer space is available for this socket.
1062  * We try not to wake our writers until they can make "significant"
1063  * progress, otherwise we'll waste resources thrashing sock_sendmsg
1064  * with a bunch of small requests.
1065  */
1066 static void
1067 xprt_write_space(struct sock *sk)
1068 {
1069         struct rpc_xprt *xprt;
1070         struct socket   *sock;
1071
1072         read_lock(&sk->sk_callback_lock);
1073         if (!(xprt = xprt_from_sock(sk)) || !(sock = sk->sk_socket))
1074                 goto out;
1075         if (xprt->shutdown)
1076                 goto out;
1077
1078         /* Wait until we have enough socket memory */
1079         if (xprt->stream) {
1080                 /* from net/ipv4/tcp.c:tcp_write_space */
1081                 if (tcp_wspace(sk) < tcp_min_write_space(sk))
1082                         goto out;
1083         } else {
1084                 /* from net/core/sock.c:sock_def_write_space */
1085                 if (!sock_writeable(sk))
1086                         goto out;
1087         }
1088
1089         if (!test_and_clear_bit(SOCK_NOSPACE, &sock->flags))
1090                 goto out;
1091
1092         spin_lock_bh(&xprt->sock_lock);
1093         if (xprt->snd_task && xprt->snd_task->tk_rpcwait == &xprt->pending)
1094                 rpc_wake_up_task(xprt->snd_task);
1095         spin_unlock_bh(&xprt->sock_lock);
1096         if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1097                 wake_up_interruptible(sk->sk_sleep);
1098 out:
1099         read_unlock(&sk->sk_callback_lock);
1100 }
1101
1102 /*
1103  * RPC receive timeout handler.
1104  */
1105 static void
1106 xprt_timer(struct rpc_task *task)
1107 {
1108         struct rpc_rqst *req = task->tk_rqstp;
1109         struct rpc_xprt *xprt = req->rq_xprt;
1110
1111         spin_lock(&xprt->sock_lock);
1112         if (req->rq_received)
1113                 goto out;
1114
1115         xprt_adjust_cwnd(req->rq_xprt, -ETIMEDOUT);
1116         __xprt_put_cong(xprt, req);
1117
1118         dprintk("RPC: %4d xprt_timer (%s request)\n",
1119                 task->tk_pid, req ? "pending" : "backlogged");
1120
1121         task->tk_status  = -ETIMEDOUT;
1122 out:
1123         task->tk_timeout = 0;
1124         rpc_wake_up_task(task);
1125         spin_unlock(&xprt->sock_lock);
1126 }
1127
1128 /*
1129  * Place the actual RPC call.
1130  * We have to copy the iovec because sendmsg fiddles with its contents.
1131  */
1132 int
1133 xprt_prepare_transmit(struct rpc_task *task)
1134 {
1135         struct rpc_rqst *req = task->tk_rqstp;
1136         struct rpc_xprt *xprt = req->rq_xprt;
1137         int err = 0;
1138
1139         dprintk("RPC: %4d xprt_prepare_transmit\n", task->tk_pid);
1140
1141         if (xprt->shutdown)
1142                 return -EIO;
1143
1144         spin_lock_bh(&xprt->sock_lock);
1145         if (req->rq_received && !req->rq_bytes_sent) {
1146                 err = req->rq_received;
1147                 goto out_unlock;
1148         }
1149         if (!__xprt_lock_write(xprt, task)) {
1150                 err = -EAGAIN;
1151                 goto out_unlock;
1152         }
1153
1154         if (!xprt_connected(xprt)) {
1155                 err = -ENOTCONN;
1156                 goto out_unlock;
1157         }
1158 out_unlock:
1159         spin_unlock_bh(&xprt->sock_lock);
1160         return err;
1161 }
1162
1163 void
1164 xprt_transmit(struct rpc_task *task)
1165 {
1166         struct rpc_clnt *clnt = task->tk_client;
1167         struct rpc_rqst *req = task->tk_rqstp;
1168         struct rpc_xprt *xprt = req->rq_xprt;
1169         int status, retry = 0;
1170
1171
1172         dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
1173
1174         /* set up everything as needed. */
1175         /* Write the record marker */
1176         if (xprt->stream) {
1177                 u32     *marker = req->rq_svec[0].iov_base;
1178
1179                 *marker = htonl(0x80000000|(req->rq_slen-sizeof(*marker)));
1180         }
1181
1182         smp_rmb();
1183         if (!req->rq_received) {
1184                 if (list_empty(&req->rq_list)) {
1185                         spin_lock_bh(&xprt->sock_lock);
1186                         /* Update the softirq receive buffer */
1187                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
1188                                         sizeof(req->rq_private_buf));
1189                         /* Add request to the receive list */
1190                         list_add_tail(&req->rq_list, &xprt->recv);
1191                         spin_unlock_bh(&xprt->sock_lock);
1192                 }
1193         } else if (!req->rq_bytes_sent)
1194                 return;
1195
1196         /* Continue transmitting the packet/record. We must be careful
1197          * to cope with writespace callbacks arriving _after_ we have
1198          * called xprt_sendmsg().
1199          */
1200         while (1) {
1201                 req->rq_xtime = jiffies;
1202                 status = xprt_sendmsg(xprt, req);
1203
1204                 if (status < 0)
1205                         break;
1206
1207                 if (xprt->stream) {
1208                         req->rq_bytes_sent += status;
1209
1210                         /* If we've sent the entire packet, immediately
1211                          * reset the count of bytes sent. */
1212                         if (req->rq_bytes_sent >= req->rq_slen) {
1213                                 req->rq_bytes_sent = 0;
1214                                 goto out_receive;
1215                         }
1216                 } else {
1217                         if (status >= req->rq_slen)
1218                                 goto out_receive;
1219                         status = -EAGAIN;
1220                         break;
1221                 }
1222
1223                 dprintk("RPC: %4d xmit incomplete (%d left of %d)\n",
1224                                 task->tk_pid, req->rq_slen - req->rq_bytes_sent,
1225                                 req->rq_slen);
1226
1227                 status = -EAGAIN;
1228                 if (retry++ > 50)
1229                         break;
1230         }
1231
1232         /* Note: at this point, task->tk_sleeping has not yet been set,
1233          *       hence there is no danger of the waking up task being put on
1234          *       schedq, and being picked up by a parallel run of rpciod().
1235          */
1236         task->tk_status = status;
1237
1238         switch (status) {
1239         case -EAGAIN:
1240                 if (test_bit(SOCK_ASYNC_NOSPACE, &xprt->sock->flags)) {
1241                         /* Protect against races with xprt_write_space */
1242                         spin_lock_bh(&xprt->sock_lock);
1243                         /* Don't race with disconnect */
1244                         if (!xprt_connected(xprt))
1245                                 task->tk_status = -ENOTCONN;
1246                         else if (test_bit(SOCK_NOSPACE, &xprt->sock->flags)) {
1247                                 task->tk_timeout = req->rq_timeout.to_current;
1248                                 rpc_sleep_on(&xprt->pending, task, NULL, NULL);
1249                         }
1250                         spin_unlock_bh(&xprt->sock_lock);
1251                         return;
1252                 }
1253                 /* Keep holding the socket if it is blocked */
1254                 rpc_delay(task, HZ>>4);
1255                 return;
1256         case -ECONNREFUSED:
1257                 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
1258                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
1259         case -ENOTCONN:
1260                 return;
1261         default:
1262                 if (xprt->stream)
1263                         xprt_disconnect(xprt);
1264         }
1265         xprt_release_write(xprt, task);
1266         return;
1267  out_receive:
1268         dprintk("RPC: %4d xmit complete\n", task->tk_pid);
1269         /* Set the task's receive timeout value */
1270         spin_lock_bh(&xprt->sock_lock);
1271         if (!xprt->nocong) {
1272                 int timer = task->tk_msg.rpc_proc->p_timer;
1273                 task->tk_timeout = rpc_calc_rto(clnt->cl_rtt, timer);
1274                 task->tk_timeout <<= rpc_ntimeo(clnt->cl_rtt, timer);
1275                 task->tk_timeout <<= clnt->cl_timeout.to_retries
1276                         - req->rq_timeout.to_retries;
1277                 if (!task->tk_timeout || task->tk_timeout > req->rq_timeout.to_maxval)
1278                         task->tk_timeout = req->rq_timeout.to_maxval;
1279         } else
1280                 task->tk_timeout = req->rq_timeout.to_current;
1281         /* Don't race with disconnect */
1282         if (!xprt_connected(xprt))
1283                 task->tk_status = -ENOTCONN;
1284         else if (!req->rq_received)
1285                 rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
1286         __xprt_release_write(xprt, task);
1287         spin_unlock_bh(&xprt->sock_lock);
1288 }
1289
1290 /*
1291  * Reserve an RPC call slot.
1292  */
1293 void
1294 xprt_reserve(struct rpc_task *task)
1295 {
1296         struct rpc_xprt *xprt = task->tk_xprt;
1297
1298         task->tk_status = -EIO;
1299         if (!xprt->shutdown) {
1300                 spin_lock(&xprt->xprt_lock);
1301                 do_xprt_reserve(task);
1302                 spin_unlock(&xprt->xprt_lock);
1303                 if (task->tk_rqstp)
1304                         del_timer_sync(&xprt->timer);
1305         }
1306 }
1307
1308 static inline void
1309 do_xprt_reserve(struct rpc_task *task)
1310 {
1311         struct rpc_xprt *xprt = task->tk_xprt;
1312
1313         task->tk_status = 0;
1314         if (task->tk_rqstp)
1315                 return;
1316         if (xprt->free) {
1317                 struct rpc_rqst *req = xprt->free;
1318                 xprt->free = req->rq_next;
1319                 req->rq_next = NULL;
1320                 task->tk_rqstp = req;
1321                 xprt_request_init(task, xprt);
1322                 return;
1323         }
1324         dprintk("RPC:      waiting for request slot\n");
1325         task->tk_status = -EAGAIN;
1326         task->tk_timeout = 0;
1327         rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
1328 }
1329
1330 /*
1331  * Allocate a 'unique' XID
1332  */
1333 static u32
1334 xprt_alloc_xid(void)
1335 {
1336         static spinlock_t xid_lock = SPIN_LOCK_UNLOCKED;
1337         static int need_init = 1;
1338         static u32 xid;
1339         u32 ret;
1340
1341         spin_lock(&xid_lock);
1342         if (unlikely(need_init)) {
1343                 xid = get_seconds() << 12;
1344                 need_init = 0;
1345         }
1346         ret = xid++;
1347         spin_unlock(&xid_lock);
1348         return ret;
1349 }
1350
1351 /*
1352  * Initialize RPC request
1353  */
1354 static void
1355 xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1356 {
1357         struct rpc_rqst *req = task->tk_rqstp;
1358
1359         req->rq_timeout = xprt->timeout;
1360         req->rq_task    = task;
1361         req->rq_xprt    = xprt;
1362         req->rq_xid     = xprt_alloc_xid();
1363         INIT_LIST_HEAD(&req->rq_list);
1364         dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
1365                         req, req->rq_xid);
1366 }
1367
1368 /*
1369  * Release an RPC call slot
1370  */
1371 void
1372 xprt_release(struct rpc_task *task)
1373 {
1374         struct rpc_xprt *xprt = task->tk_xprt;
1375         struct rpc_rqst *req;
1376
1377         if (!(req = task->tk_rqstp))
1378                 return;
1379         spin_lock_bh(&xprt->sock_lock);
1380         __xprt_release_write(xprt, task);
1381         __xprt_put_cong(xprt, req);
1382         if (!list_empty(&req->rq_list))
1383                 list_del(&req->rq_list);
1384         xprt->last_used = jiffies;
1385         if (list_empty(&xprt->recv) && !xprt->shutdown)
1386                 mod_timer(&xprt->timer, xprt->last_used + XPRT_IDLE_TIMEOUT);
1387         spin_unlock_bh(&xprt->sock_lock);
1388         task->tk_rqstp = NULL;
1389         memset(req, 0, sizeof(*req));   /* mark unused */
1390
1391         dprintk("RPC: %4d release request %p\n", task->tk_pid, req);
1392
1393         spin_lock(&xprt->xprt_lock);
1394         req->rq_next = xprt->free;
1395         xprt->free   = req;
1396
1397         xprt_clear_backlog(xprt);
1398         spin_unlock(&xprt->xprt_lock);
1399 }
1400
1401 /*
1402  * Set default timeout parameters
1403  */
1404 void
1405 xprt_default_timeout(struct rpc_timeout *to, int proto)
1406 {
1407         if (proto == IPPROTO_UDP)
1408                 xprt_set_timeout(to, 5,  5 * HZ);
1409         else
1410                 xprt_set_timeout(to, 5, 60 * HZ);
1411 }
1412
1413 /*
1414  * Set constant timeout
1415  */
1416 void
1417 xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
1418 {
1419         to->to_current   =
1420         to->to_initval   =
1421         to->to_increment = incr;
1422         to->to_maxval    = incr * retr;
1423         to->to_retries   = retr;
1424         to->to_exponential = 0;
1425 }
1426
1427 unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1428 unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE << 2;
1429
1430 /*
1431  * Initialize an RPC client
1432  */
1433 static struct rpc_xprt *
1434 xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
1435 {
1436         struct rpc_xprt *xprt;
1437         unsigned int entries, xprt_size;
1438         struct rpc_rqst *req;
1439         int             i;
1440
1441         dprintk("RPC:      setting up %s transport...\n",
1442                                 proto == IPPROTO_UDP? "UDP" : "TCP");
1443
1444         entries = (proto == IPPROTO_TCP)?
1445                 xprt_tcp_slot_table_entries : xprt_udp_slot_table_entries;
1446         xprt_size = sizeof(struct rpc_xprt) + sizeof(struct rpc_rqst) * entries;
1447
1448         if ((xprt = kmalloc(xprt_size, GFP_KERNEL)) == NULL)
1449                 return NULL;
1450         memset(xprt, 0, xprt_size);
1451         xprt->free = (struct rpc_rqst *)(xprt + 1);
1452         xprt->max_reqs = entries;
1453
1454         xprt->addr = *ap;
1455         xprt->prot = proto;
1456         xprt->stream = (proto == IPPROTO_TCP)? 1 : 0;
1457         if (xprt->stream) {
1458                 xprt->cwnd = RPC_MAXCWND(xprt);
1459                 xprt->nocong = 1;
1460         } else
1461                 xprt->cwnd = RPC_INITCWND;
1462         spin_lock_init(&xprt->sock_lock);
1463         spin_lock_init(&xprt->xprt_lock);
1464         init_waitqueue_head(&xprt->cong_wait);
1465
1466         INIT_LIST_HEAD(&xprt->recv);
1467         INIT_WORK(&xprt->task_cleanup, xprt_socket_autoclose, xprt);
1468         init_timer(&xprt->timer);
1469         xprt->timer.function = xprt_init_autodisconnect;
1470         xprt->timer.data = (unsigned long) xprt;
1471         xprt->last_used = jiffies;
1472
1473         /* Set timeout parameters */
1474         if (to) {
1475                 xprt->timeout = *to;
1476                 xprt->timeout.to_current = to->to_initval;
1477         } else
1478                 xprt_default_timeout(&xprt->timeout, xprt->prot);
1479
1480         INIT_RPC_WAITQ(&xprt->pending, "xprt_pending");
1481         INIT_RPC_WAITQ(&xprt->sending, "xprt_sending");
1482         INIT_RPC_WAITQ(&xprt->resend, "xprt_resend");
1483         INIT_RPC_WAITQ(&xprt->backlog, "xprt_backlog");
1484
1485         /* initialize free list */
1486         entries--;
1487         req = xprt->free;
1488         for (i = 0; i < entries; i++, req++)
1489                 req->rq_next = req + 1;
1490         req->rq_next = NULL;
1491
1492         /* Check whether we want to use a reserved port */
1493         xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1494
1495         dprintk("RPC:      created transport %p with %u slots\n", xprt,
1496                         xprt->max_reqs);
1497
1498         return xprt;
1499 }
1500
1501 /*
1502  * Bind to a reserved port
1503  */
1504 static inline int
1505 xprt_bindresvport(struct socket *sock)
1506 {
1507         struct sockaddr_in myaddr;
1508         int             err, port;
1509         kernel_cap_t saved_cap = current->cap_effective;
1510
1511         /* Override capabilities.
1512          * They were checked in xprt_create_proto i.e. at mount time
1513          */
1514         cap_raise(current->cap_effective, CAP_NET_BIND_SERVICE);
1515
1516         memset(&myaddr, 0, sizeof(myaddr));
1517         myaddr.sin_family = AF_INET;
1518         port = 800;
1519         do {
1520                 myaddr.sin_port = htons(port);
1521                 err = sock->ops->bind(sock, (struct sockaddr *) &myaddr,
1522                                                 sizeof(myaddr));
1523         } while (err == -EADDRINUSE && --port > 0);
1524         current->cap_effective = saved_cap;
1525
1526         if (err < 0)
1527                 printk("RPC: Can't bind to reserved port (%d).\n", -err);
1528
1529         return err;
1530 }
1531
1532 static void
1533 xprt_bind_socket(struct rpc_xprt *xprt, struct socket *sock)
1534 {
1535         struct sock     *sk = sock->sk;
1536
1537         if (xprt->inet)
1538                 return;
1539
1540         write_lock_bh(&sk->sk_callback_lock);
1541         sk->sk_user_data = xprt;
1542         xprt->old_data_ready = sk->sk_data_ready;
1543         xprt->old_state_change = sk->sk_state_change;
1544         xprt->old_write_space = sk->sk_write_space;
1545         if (xprt->prot == IPPROTO_UDP) {
1546                 sk->sk_data_ready = udp_data_ready;
1547                 sk->sk_no_check = UDP_CSUM_NORCV;
1548                 xprt_set_connected(xprt);
1549         } else {
1550                 struct tcp_opt *tp = tcp_sk(sk);
1551                 tp->nonagle = 1;        /* disable Nagle's algorithm */
1552                 sk->sk_data_ready = tcp_data_ready;
1553                 sk->sk_state_change = tcp_state_change;
1554                 xprt_clear_connected(xprt);
1555         }
1556         sk->sk_write_space = xprt_write_space;
1557
1558         /* Reset to new socket */
1559         xprt->sock = sock;
1560         xprt->inet = sk;
1561         write_unlock_bh(&sk->sk_callback_lock);
1562
1563         return;
1564 }
1565
1566 /*
1567  * Set socket buffer length
1568  */
1569 void
1570 xprt_sock_setbufsize(struct rpc_xprt *xprt)
1571 {
1572         struct sock *sk = xprt->inet;
1573
1574         if (xprt->stream)
1575                 return;
1576         if (xprt->rcvsize) {
1577                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
1578                 sk->sk_rcvbuf = xprt->rcvsize * xprt->max_reqs *  2;
1579         }
1580         if (xprt->sndsize) {
1581                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
1582                 sk->sk_sndbuf = xprt->sndsize * xprt->max_reqs * 2;
1583                 sk->sk_write_space(sk);
1584         }
1585 }
1586
1587 /*
1588  * Datastream sockets are created here, but xprt_connect will create
1589  * and connect stream sockets.
1590  */
1591 static struct socket *
1592 xprt_create_socket(int proto, struct rpc_timeout *to, int resvport)
1593 {
1594         struct socket   *sock;
1595         int             type, err;
1596
1597         dprintk("RPC:      xprt_create_socket(%s %d)\n",
1598                            (proto == IPPROTO_UDP)? "udp" : "tcp", proto);
1599
1600         type = (proto == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1601
1602         if ((err = sock_create(PF_INET, type, proto, &sock)) < 0) {
1603                 printk("RPC: can't create socket (%d).\n", -err);
1604                 return NULL;
1605         }
1606
1607         /* If the caller has the capability, bind to a reserved port */
1608         if (resvport && xprt_bindresvport(sock) < 0) {
1609                 printk("RPC: can't bind to reserved port.\n");
1610                 goto failed;
1611         }
1612
1613         return sock;
1614
1615 failed:
1616         sock_release(sock);
1617         return NULL;
1618 }
1619
1620 /*
1621  * Create an RPC client transport given the protocol and peer address.
1622  */
1623 struct rpc_xprt *
1624 xprt_create_proto(int proto, struct sockaddr_in *sap, struct rpc_timeout *to)
1625 {
1626         struct rpc_xprt *xprt;
1627
1628         xprt = xprt_setup(proto, sap, to);
1629         if (!xprt)
1630                 goto out_bad;
1631
1632         dprintk("RPC:      xprt_create_proto created xprt %p\n", xprt);
1633         return xprt;
1634  out_bad:
1635         dprintk("RPC:      xprt_create_proto failed\n");
1636         if (xprt)
1637                 kfree(xprt);
1638         return NULL;
1639 }
1640
1641 /*
1642  * Prepare for transport shutdown.
1643  */
1644 void
1645 xprt_shutdown(struct rpc_xprt *xprt)
1646 {
1647         xprt->shutdown = 1;
1648         rpc_wake_up(&xprt->sending);
1649         rpc_wake_up(&xprt->resend);
1650         rpc_wake_up(&xprt->pending);
1651         rpc_wake_up(&xprt->backlog);
1652         if (waitqueue_active(&xprt->cong_wait))
1653                 wake_up(&xprt->cong_wait);
1654         del_timer_sync(&xprt->timer);
1655 }
1656
1657 /*
1658  * Clear the xprt backlog queue
1659  */
1660 int
1661 xprt_clear_backlog(struct rpc_xprt *xprt) {
1662         rpc_wake_up_next(&xprt->backlog);
1663         if (waitqueue_active(&xprt->cong_wait))
1664                 wake_up(&xprt->cong_wait);
1665         return 1;
1666 }
1667
1668 /*
1669  * Destroy an RPC transport, killing off all requests.
1670  */
1671 int
1672 xprt_destroy(struct rpc_xprt *xprt)
1673 {
1674         dprintk("RPC:      destroying transport %p\n", xprt);
1675         xprt_shutdown(xprt);
1676         xprt_close(xprt);
1677         kfree(xprt);
1678
1679         return 0;
1680 }