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 #include <linux/random.h>
  61
  62 #include <net/sock.h>
  63 #include <net/checksum.h>
  64 #include <net/udp.h>
  65 #include <net/tcp.h>
  66
  67 /*
  68  * Local variables
  69  */
  70
  71 #ifdef RPC_DEBUG
  72 # undef  RPC_DEBUG_DATA
  73 # define RPCDBG_FACILITY        RPCDBG_XPRT
  74 #endif
  75
  76 #define XPRT_MAX_BACKOFF        (8)
  77 #define XPRT_IDLE_TIMEOUT       (5*60*HZ)
  78 #define XPRT_MAX_RESVPORT       (800)
  79
  80 /*
  81  * Local functions
  82  */
  83 static void     xprt_request_init(struct rpc_task *, struct rpc_xprt *);
  84 static inline void      do_xprt_reserve(struct rpc_task *);
  85 static void     xprt_disconnect(struct rpc_xprt *);
  86 static void     xprt_connect_status(struct rpc_task *task);
  87 static struct rpc_xprt * xprt_setup(int proto, struct sockaddr_in *ap,
  88                                                 struct rpc_timeout *to);
  89 static struct socket *xprt_create_socket(struct rpc_xprt *, int, int);
  90 static void     xprt_bind_socket(struct rpc_xprt *, struct socket *);
  91 static int      __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
  92
  93 static int      xprt_clear_backlog(struct rpc_xprt *xprt);
  94
  95 #ifdef RPC_DEBUG_DATA
  96 /*
  97  * Print the buffer contents (first 128 bytes only--just enough for
  98  * diropres return).
  99  */
 100 static void
 101 xprt_pktdump(char *msg, u32 *packet, unsigned int count)
 102 {
 103         u8      *buf = (u8 *) packet;
 104         int     j;
 105
 106         dprintk("RPC:      %s\n", msg);
 107         for (j = 0; j < count && j < 128; j += 4) {
 108                 if (!(j & 31)) {
 109                         if (j)
 110                                 dprintk("\n");
 111                         dprintk("0x%04x ", j);
 112                 }
 113                 dprintk("%02x%02x%02x%02x ",
 114                         buf[j], buf[j+1], buf[j+2], buf[j+3]);
 115         }
 116         dprintk("\n");
 117 }
 118 #else
 119 static inline void
 120 xprt_pktdump(char *msg, u32 *packet, unsigned int count)
 121 {
 122         /* NOP */
 123 }
 124 #endif
 125
 126 /*
 127  * Look up RPC transport given an INET socket
 128  */
 129 static inline struct rpc_xprt *
 130 xprt_from_sock(struct sock *sk)
 131 {
 132         return (struct rpc_xprt *) sk->sk_user_data;
 133 }
 134
 135 /*
 136  * Serialize write access to sockets, in order to prevent different
 137  * requests from interfering with each other.
 138  * Also prevents TCP socket connects from colliding with writes.
 139  */
 140 static int
 141 __xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 142 {
 143         struct rpc_rqst *req = task->tk_rqstp;
 144
 145         if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate)) {
 146                 if (task == xprt->snd_task)
 147                         return 1;
 148                 if (task == NULL)
 149                         return 0;
 150                 goto out_sleep;
 151         }
 152         if (xprt->nocong || __xprt_get_cong(xprt, task)) {
 153                 xprt->snd_task = task;
 154                 if (req) {
 155                         req->rq_bytes_sent = 0;
 156                         req->rq_ntrans++;
 157                 }
 158                 return 1;
 159         }
 160         smp_mb__before_clear_bit();
 161         clear_bit(XPRT_LOCKED, &xprt->sockstate);
 162         smp_mb__after_clear_bit();
 163 out_sleep:
 164         dprintk("RPC: %4d failed to lock socket %p\n", task->tk_pid, xprt);
 165         task->tk_timeout = 0;
 166         task->tk_status = -EAGAIN;
 167         if (req && req->rq_ntrans)
 168                 rpc_sleep_on(&xprt->resend, task, NULL, NULL);
 169         else
 170                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
 171         return 0;
 172 }
 173
 174 static inline int
 175 xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 176 {
 177         int retval;
 178
 179         spin_lock_bh(&xprt->sock_lock);
 180         retval = __xprt_lock_write(xprt, task);
 181         spin_unlock_bh(&xprt->sock_lock);
 182         return retval;
 183 }
 184
 185
 186 static void
 187 __xprt_lock_write_next(struct rpc_xprt *xprt)
 188 {
 189         struct rpc_task *task;
 190
 191         if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
 192                 return;
 193         if (!xprt->nocong && RPCXPRT_CONGESTED(xprt))
 194                 goto out_unlock;
 195         task = rpc_wake_up_next(&xprt->resend);
 196         if (!task) {
 197                 task = rpc_wake_up_next(&xprt->sending);
 198                 if (!task)
 199                         goto out_unlock;
 200         }
 201         if (xprt->nocong || __xprt_get_cong(xprt, task)) {
 202                 struct rpc_rqst *req = task->tk_rqstp;
 203                 xprt->snd_task = task;
 204                 if (req) {
 205                         req->rq_bytes_sent = 0;
 206                         req->rq_ntrans++;
 207                 }
 208                 return;
 209         }
 210 out_unlock:
 211         smp_mb__before_clear_bit();
 212         clear_bit(XPRT_LOCKED, &xprt->sockstate);
 213         smp_mb__after_clear_bit();
 214 }
 215
 216 /*
 217  * Releases the socket for use by other requests.
 218  */
 219 static void
 220 __xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 221 {
 222         if (xprt->snd_task == task) {
 223                 xprt->snd_task = NULL;
 224                 smp_mb__before_clear_bit();
 225                 clear_bit(XPRT_LOCKED, &xprt->sockstate);
 226                 smp_mb__after_clear_bit();
 227                 __xprt_lock_write_next(xprt);
 228         }
 229 }
 230
 231 static inline void
 232 xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 233 {
 234         spin_lock_bh(&xprt->sock_lock);
 235         __xprt_release_write(xprt, task);
 236         spin_unlock_bh(&xprt->sock_lock);
 237 }
 238
 239 /*
 240  * Write data to socket.
 241  */
 242 static inline int
 243 xprt_sendmsg(struct rpc_xprt *xprt, struct rpc_rqst *req)
 244 {
 245         struct socket   *sock = xprt->sock;
 246         struct xdr_buf  *xdr = &req->rq_snd_buf;
 247         struct sockaddr *addr = NULL;
 248         int addrlen = 0;
 249         unsigned int    skip;
 250         int             result;
 251
 252         if (!sock)
 253                 return -ENOTCONN;
 254
 255         xprt_pktdump("packet data:",
 256                                 req->rq_svec->iov_base,
 257                                 req->rq_svec->iov_len);
 258
 259         /* For UDP, we need to provide an address */
 260         if (!xprt->stream) {
 261                 addr = (struct sockaddr *) &xprt->addr;
 262                 addrlen = sizeof(xprt->addr);
 263         }
 264         /* Dont repeat bytes */
 265         skip = req->rq_bytes_sent;
 266
 267         clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 268         result = xdr_sendpages(sock, addr, addrlen, xdr, skip, MSG_DONTWAIT);
 269
 270         dprintk("RPC:      xprt_sendmsg(%d) = %d\n", xdr->len - skip, result);
 271
 272         if (result >= 0)
 273                 return result;
 274
 275         switch (result) {
 276         case -ECONNREFUSED:
 277                 /* When the server has died, an ICMP port unreachable message
 278                  * prompts ECONNREFUSED.
 279                  */
 280         case -EAGAIN:
 281                 break;
 282         case -ECONNRESET:
 283         case -ENOTCONN:
 284         case -EPIPE:
 285                 /* connection broken */
 286                 if (xprt->stream)
 287                         result = -ENOTCONN;
 288                 break;
 289         default:
 290                 printk(KERN_NOTICE "RPC: sendmsg returned error %d\n", -result);
 291         }
 292         return result;
 293 }
 294
 295 /*
 296  * Van Jacobson congestion avoidance. Check if the congestion window
 297  * overflowed. Put the task to sleep if this is the case.
 298  */
 299 static int
 300 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 301 {
 302         struct rpc_rqst *req = task->tk_rqstp;
 303
 304         if (req->rq_cong)
 305                 return 1;
 306         dprintk("RPC: %4d xprt_cwnd_limited cong = %ld cwnd = %ld\n",
 307                         task->tk_pid, xprt->cong, xprt->cwnd);
 308         if (RPCXPRT_CONGESTED(xprt))
 309                 return 0;
 310         req->rq_cong = 1;
 311         xprt->cong += RPC_CWNDSCALE;
 312         return 1;
 313 }
 314
 315 /*
 316  * Adjust the congestion window, and wake up the next task
 317  * that has been sleeping due to congestion
 318  */
 319 static void
 320 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 321 {
 322         if (!req->rq_cong)
 323                 return;
 324         req->rq_cong = 0;
 325         xprt->cong -= RPC_CWNDSCALE;
 326         __xprt_lock_write_next(xprt);
 327 }
 328
 329 /*
 330  * Adjust RPC congestion window
 331  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
 332  */
 333 static void
 334 xprt_adjust_cwnd(struct rpc_xprt *xprt, int result)
 335 {
 336         unsigned long   cwnd;
 337
 338         cwnd = xprt->cwnd;
 339         if (result >= 0 && cwnd <= xprt->cong) {
 340                 /* The (cwnd >> 1) term makes sure
 341                  * the result gets rounded properly. */
 342                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 343                 if (cwnd > RPC_MAXCWND(xprt))
 344                         cwnd = RPC_MAXCWND(xprt);
 345                 __xprt_lock_write_next(xprt);
 346         } else if (result == -ETIMEDOUT) {
 347                 cwnd >>= 1;
 348                 if (cwnd < RPC_CWNDSCALE)
 349                         cwnd = RPC_CWNDSCALE;
 350         }
 351         dprintk("RPC:      cong %ld, cwnd was %ld, now %ld\n",
 352                         xprt->cong, xprt->cwnd, cwnd);
 353         xprt->cwnd = cwnd;
 354 }
 355
 356 /*
 357  * Reset the major timeout value
 358  */
 359 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 360 {
 361         struct rpc_timeout *to = &req->rq_xprt->timeout;
 362
 363         req->rq_majortimeo = req->rq_timeout;
 364         if (to->to_exponential)
 365                 req->rq_majortimeo <<= to->to_retries;
 366         else
 367                 req->rq_majortimeo += to->to_increment * to->to_retries;
 368         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 369                 req->rq_majortimeo = to->to_maxval;
 370         req->rq_majortimeo += jiffies;
 371 }
 372
 373 /*
 374  * Adjust timeout values etc for next retransmit
 375  */
 376 int xprt_adjust_timeout(struct rpc_rqst *req)
 377 {
 378         struct rpc_xprt *xprt = req->rq_xprt;
 379         struct rpc_timeout *to = &xprt->timeout;
 380         int status = 0;
 381
 382         if (time_before(jiffies, req->rq_majortimeo)) {
 383                 if (to->to_exponential)
 384                         req->rq_timeout <<= 1;
 385                 else
 386                         req->rq_timeout += to->to_increment;
 387                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 388                         req->rq_timeout = to->to_maxval;
 389                 req->rq_retries++;
 390                 pprintk("RPC: %lu retrans\n", jiffies);
 391         } else {
 392                 req->rq_timeout = to->to_initval;
 393                 req->rq_retries = 0;
 394                 xprt_reset_majortimeo(req);
 395                 /* Reset the RTT counters == "slow start" */
 396                 spin_lock_bh(&xprt->sock_lock);
 397                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 398                 spin_unlock_bh(&xprt->sock_lock);
 399                 pprintk("RPC: %lu timeout\n", jiffies);
 400                 status = -ETIMEDOUT;
 401         }
 402
 403         if (req->rq_timeout == 0) {
 404                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 405                 req->rq_timeout = 5 * HZ;
 406         }
 407         return status;
 408 }
 409
 410 /*
 411  * Close down a transport socket
 412  */
 413 static void
 414 xprt_close(struct rpc_xprt *xprt)
 415 {
 416         struct socket   *sock = xprt->sock;
 417         struct sock     *sk = xprt->inet;
 418
 419         if (!sk)
 420                 return;
 421
 422         write_lock_bh(&sk->sk_callback_lock);
 423         xprt->inet = NULL;
 424         xprt->sock = NULL;
 425
 426         sk->sk_user_data    = NULL;
 427         sk->sk_data_ready   = xprt->old_data_ready;
 428         sk->sk_state_change = xprt->old_state_change;
 429         sk->sk_write_space  = xprt->old_write_space;
 430         write_unlock_bh(&sk->sk_callback_lock);
 431
 432         sk->sk_no_check  = 0;
 433
 434         sock_release(sock);
 435 }
 436
 437 static void
 438 xprt_socket_autoclose(void *args)
 439 {
 440         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
 441
 442         xprt_disconnect(xprt);
 443         xprt_close(xprt);
 444         xprt_release_write(xprt, NULL);
 445 }
 446
 447 /*
 448  * Mark a transport as disconnected
 449  */
 450 static void
 451 xprt_disconnect(struct rpc_xprt *xprt)
 452 {
 453         dprintk("RPC:      disconnected transport %p\n", xprt);
 454         spin_lock_bh(&xprt->sock_lock);
 455         xprt_clear_connected(xprt);
 456         rpc_wake_up_status(&xprt->pending, -ENOTCONN);
 457         spin_unlock_bh(&xprt->sock_lock);
 458 }
 459
 460 /*
 461  * Used to allow disconnection when we've been idle
 462  */
 463 static void
 464 xprt_init_autodisconnect(unsigned long data)
 465 {
 466         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 467
 468         spin_lock(&xprt->sock_lock);
 469         if (!list_empty(&xprt->recv) || xprt->shutdown)
 470                 goto out_abort;
 471         if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
 472                 goto out_abort;
 473         spin_unlock(&xprt->sock_lock);
 474         /* Let keventd close the socket */
 475         if (test_bit(XPRT_CONNECTING, &xprt->sockstate) != 0)
 476                 xprt_release_write(xprt, NULL);
 477         else
 478                 schedule_work(&xprt->task_cleanup);
 479         return;
 480 out_abort:
 481         spin_unlock(&xprt->sock_lock);
 482 }
 483
 484 static void xprt_socket_connect(void *args)
 485 {
 486         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
 487         struct socket *sock = xprt->sock;
 488         int status = -EIO;
 489
 490         if (xprt->shutdown || xprt->addr.sin_port == 0)
 491                 goto out;
 492
 493         /*
 494          * Start by resetting any existing state
 495          */
 496         xprt_close(xprt);
 497         sock = xprt_create_socket(xprt, xprt->prot, xprt->resvport);
 498         if (sock == NULL) {
 499                 /* couldn't create socket or bind to reserved port;
 500                  * this is likely a permanent error, so cause an abort */
 501                 goto out;
 502         }
 503         xprt_bind_socket(xprt, sock);
 504         xprt_sock_setbufsize(xprt);
 505
 506         status = 0;
 507         if (!xprt->stream)
 508                 goto out;
 509
 510         /*
 511          * Tell the socket layer to start connecting...
 512          */
 513         status = sock->ops->connect(sock, (struct sockaddr *) &xprt->addr,
 514                         sizeof(xprt->addr), O_NONBLOCK);
 515         dprintk("RPC: %p  connect status %d connected %d sock state %d\n",
 516                         xprt, -status, xprt_connected(xprt), sock->sk->sk_state);
 517         if (status < 0) {
 518                 switch (status) {
 519                         case -EINPROGRESS:
 520                         case -EALREADY:
 521                                 goto out_clear;
 522                 }
 523         }
 524 out:
 525         if (status < 0)
 526                 rpc_wake_up_status(&xprt->pending, status);
 527         else
 528                 rpc_wake_up(&xprt->pending);
 529 out_clear:
 530         smp_mb__before_clear_bit();
 531         clear_bit(XPRT_CONNECTING, &xprt->sockstate);
 532         smp_mb__after_clear_bit();
 533 }
 534
 535 /*
 536  * Attempt to connect a TCP socket.
 537  *
 538  */
 539 void xprt_connect(struct rpc_task *task)
 540 {
 541         struct rpc_xprt *xprt = task->tk_xprt;
 542
 543         dprintk("RPC: %4d xprt_connect xprt %p %s connected\n", task->tk_pid,
 544                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 545
 546         if (xprt->shutdown) {
 547                 task->tk_status = -EIO;
 548                 return;
 549         }
 550         if (!xprt->addr.sin_port) {
 551                 task->tk_status = -EIO;
 552                 return;
 553         }
 554         if (!xprt_lock_write(xprt, task))
 555                 return;
 556         if (xprt_connected(xprt))
 557                 goto out_write;
 558
 559         if (task->tk_rqstp)
 560                 task->tk_rqstp->rq_bytes_sent = 0;
 561
 562         task->tk_timeout = RPC_CONNECT_TIMEOUT;
 563         rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
 564         if (!test_and_set_bit(XPRT_CONNECTING, &xprt->sockstate)) {
 565                 /* Note: if we are here due to a dropped connection
 566                  *       we delay reconnecting by RPC_REESTABLISH_TIMEOUT/HZ
 567                  *       seconds
 568                  */
 569                 if (xprt->sock != NULL)
 570                         schedule_delayed_work(&xprt->sock_connect,
 571                                         RPC_REESTABLISH_TIMEOUT);
 572                 else
 573                         schedule_work(&xprt->sock_connect);
 574         }
 575         return;
 576  out_write:
 577         xprt_release_write(xprt, task);
 578 }
 579
 580 /*
 581  * We arrive here when awoken from waiting on connection establishment.
 582  */
 583 static void
 584 xprt_connect_status(struct rpc_task *task)
 585 {
 586         struct rpc_xprt *xprt = task->tk_xprt;
 587
 588         if (task->tk_status >= 0) {
 589                 dprintk("RPC: %4d xprt_connect_status: connection established\n",
 590                                 task->tk_pid);
 591                 return;
 592         }
 593
 594         /* if soft mounted, just cause this RPC to fail */
 595         if (RPC_IS_SOFT(task))
 596                 task->tk_status = -EIO;
 597
 598         switch (task->tk_status) {
 599         case -ECONNREFUSED:
 600         case -ECONNRESET:
 601         case -ENOTCONN:
 602                 return;
 603         case -ETIMEDOUT:
 604                 dprintk("RPC: %4d xprt_connect_status: timed out\n",
 605                                 task->tk_pid);
 606                 break;
 607         default:
 608                 printk(KERN_ERR "RPC: error %d connecting to server %s\n",
 609                                 -task->tk_status, task->tk_client->cl_server);
 610         }
 611         xprt_release_write(xprt, task);
 612 }
 613
 614 /*
 615  * Look up the RPC request corresponding to a reply, and then lock it.
 616  */
 617 static inline struct rpc_rqst *
 618 xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
 619 {
 620         struct list_head *pos;
 621         struct rpc_rqst *req = NULL;
 622
 623         list_for_each(pos, &xprt->recv) {
 624                 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
 625                 if (entry->rq_xid == xid) {
 626                         req = entry;
 627                         break;
 628                 }
 629         }
 630         return req;
 631 }
 632
 633 /*
 634  * Complete reply received.
 635  * The TCP code relies on us to remove the request from xprt->pending.
 636  */
 637 static void
 638 xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
 639 {
 640         struct rpc_task *task = req->rq_task;
 641         struct rpc_clnt *clnt = task->tk_client;
 642
 643         /* Adjust congestion window */
 644         if (!xprt->nocong) {
 645                 unsigned timer = task->tk_msg.rpc_proc->p_timer;
 646                 xprt_adjust_cwnd(xprt, copied);
 647                 __xprt_put_cong(xprt, req);
 648                 if (timer) {
 649                         if (req->rq_ntrans == 1)
 650                                 rpc_update_rtt(clnt->cl_rtt, timer,
 651                                                 (long)jiffies - req->rq_xtime);
 652                         rpc_set_timeo(clnt->cl_rtt, timer, req->rq_ntrans - 1);
 653                 }
 654         }
 655
 656 #ifdef RPC_PROFILE
 657         /* Profile only reads for now */
 658         if (copied > 1024) {
 659                 static unsigned long    nextstat;
 660                 static unsigned long    pkt_rtt, pkt_len, pkt_cnt;
 661
 662                 pkt_cnt++;
 663                 pkt_len += req->rq_slen + copied;
 664                 pkt_rtt += jiffies - req->rq_xtime;
 665                 if (time_before(nextstat, jiffies)) {
 666                         printk("RPC: %lu %ld cwnd\n", jiffies, xprt->cwnd);
 667                         printk("RPC: %ld %ld %ld %ld stat\n",
 668                                         jiffies, pkt_cnt, pkt_len, pkt_rtt);
 669                         pkt_rtt = pkt_len = pkt_cnt = 0;
 670                         nextstat = jiffies + 5 * HZ;
 671                 }
 672         }
 673 #endif
 674
 675         dprintk("RPC: %4d has input (%d bytes)\n", task->tk_pid, copied);
 676         list_del_init(&req->rq_list);
 677         req->rq_received = req->rq_private_buf.len = copied;
 678
 679         /* ... and wake up the process. */
 680         rpc_wake_up_task(task);
 681         return;
 682 }
 683
 684 static size_t
 685 skb_read_bits(skb_reader_t *desc, void *to, size_t len)
 686 {
 687         if (len > desc->count)
 688                 len = desc->count;
 689         if (skb_copy_bits(desc->skb, desc->offset, to, len))
 690                 return 0;
 691         desc->count -= len;
 692         desc->offset += len;
 693         return len;
 694 }
 695
 696 static size_t
 697 skb_read_and_csum_bits(skb_reader_t *desc, void *to, size_t len)
 698 {
 699         unsigned int csum2, pos;
 700
 701         if (len > desc->count)
 702                 len = desc->count;
 703         pos = desc->offset;
 704         csum2 = skb_copy_and_csum_bits(desc->skb, pos, to, len, 0);
 705         desc->csum = csum_block_add(desc->csum, csum2, pos);
 706         desc->count -= len;
 707         desc->offset += len;
 708         return len;
 709 }
 710
 711 /*
 712  * We have set things up such that we perform the checksum of the UDP
 713  * packet in parallel with the copies into the RPC client iovec.  -DaveM
 714  */
 715 int
 716 csum_partial_copy_to_xdr(struct xdr_buf *xdr, struct sk_buff *skb)
 717 {
 718         skb_reader_t desc;
 719
 720         desc.skb = skb;
 721         desc.offset = sizeof(struct udphdr);
 722         desc.count = skb->len - desc.offset;
 723
 724         if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 725                 goto no_checksum;
 726
 727         desc.csum = csum_partial(skb->data, desc.offset, skb->csum);
 728         if (xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_and_csum_bits) < 0)
 729                 return -1;
 730         if (desc.offset != skb->len) {
 731                 unsigned int csum2;
 732                 csum2 = skb_checksum(skb, desc.offset, skb->len - desc.offset, 0);
 733                 desc.csum = csum_block_add(desc.csum, csum2, desc.offset);
 734         }
 735         if (desc.count)
 736                 return -1;
 737         if ((unsigned short)csum_fold(desc.csum))
 738                 return -1;
 739         return 0;
 740 no_checksum:
 741         if (xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_bits) < 0)
 742                 return -1;
 743         if (desc.count)
 744                 return -1;
 745         return 0;
 746 }
 747
 748 /*
 749  * Input handler for RPC replies. Called from a bottom half and hence
 750  * atomic.
 751  */
 752 static void
 753 udp_data_ready(struct sock *sk, int len)
 754 {
 755         struct rpc_task *task;
 756         struct rpc_xprt *xprt;
 757         struct rpc_rqst *rovr;
 758         struct sk_buff  *skb;
 759         int err, repsize, copied;
 760         u32 _xid, *xp;
 761
 762         read_lock(&sk->sk_callback_lock);
 763         dprintk("RPC:      udp_data_ready...\n");
 764         if (!(xprt = xprt_from_sock(sk))) {
 765                 printk("RPC:      udp_data_ready request not found!\n");
 766                 goto out;
 767         }
 768
 769         dprintk("RPC:      udp_data_ready client %p\n", xprt);
 770
 771         if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
 772                 goto out;
 773
 774         if (xprt->shutdown)
 775                 goto dropit;
 776
 777         repsize = skb->len - sizeof(struct udphdr);
 778         if (repsize < 4) {
 779                 printk("RPC: impossible RPC reply size %d!\n", repsize);
 780                 goto dropit;
 781         }
 782
 783         /* Copy the XID from the skb... */
 784         xp = skb_header_pointer(skb, sizeof(struct udphdr),
 785                                 sizeof(_xid), &_xid);
 786         if (xp == NULL)
 787                 goto dropit;
 788
 789         /* Look up and lock the request corresponding to the given XID */
 790         spin_lock(&xprt->sock_lock);
 791         rovr = xprt_lookup_rqst(xprt, *xp);
 792         if (!rovr)
 793                 goto out_unlock;
 794         task = rovr->rq_task;
 795
 796         dprintk("RPC: %4d received reply\n", task->tk_pid);
 797
 798         if ((copied = rovr->rq_private_buf.buflen) > repsize)
 799                 copied = repsize;
 800
 801         /* Suck it into the iovec, verify checksum if not done by hw. */
 802         if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
 803                 goto out_unlock;
 804
 805         /* Something worked... */
 806         dst_confirm(skb->dst);
 807
 808         xprt_complete_rqst(xprt, rovr, copied);
 809
 810  out_unlock:
 811         spin_unlock(&xprt->sock_lock);
 812  dropit:
 813         skb_free_datagram(sk, skb);
 814  out:
 815         read_unlock(&sk->sk_callback_lock);
 816 }
 817
 818 /*
 819  * Copy from an skb into memory and shrink the skb.
 820  */
 821 static inline size_t
 822 tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
 823 {
 824         if (len > desc->count)
 825                 len = desc->count;
 826         if (skb_copy_bits(desc->skb, desc->offset, p, len))
 827                 return 0;
 828         desc->offset += len;
 829         desc->count -= len;
 830         return len;
 831 }
 832
 833 /*
 834  * TCP read fragment marker
 835  */
 836 static inline void
 837 tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
 838 {
 839         size_t len, used;
 840         char *p;
 841
 842         p = ((char *) &xprt->tcp_recm) + xprt->tcp_offset;
 843         len = sizeof(xprt->tcp_recm) - xprt->tcp_offset;
 844         used = tcp_copy_data(desc, p, len);
 845         xprt->tcp_offset += used;
 846         if (used != len)
 847                 return;
 848         xprt->tcp_reclen = ntohl(xprt->tcp_recm);
 849         if (xprt->tcp_reclen & 0x80000000)
 850                 xprt->tcp_flags |= XPRT_LAST_FRAG;
 851         else
 852                 xprt->tcp_flags &= ~XPRT_LAST_FRAG;
 853         xprt->tcp_reclen &= 0x7fffffff;
 854         xprt->tcp_flags &= ~XPRT_COPY_RECM;
 855         xprt->tcp_offset = 0;
 856         /* Sanity check of the record length */
 857         if (xprt->tcp_reclen < 4) {
 858                 printk(KERN_ERR "RPC: Invalid TCP record fragment length\n");
 859                 xprt_disconnect(xprt);
 860         }
 861         dprintk("RPC:      reading TCP record fragment of length %d\n",
 862                         xprt->tcp_reclen);
 863 }
 864
 865 static void
 866 tcp_check_recm(struct rpc_xprt *xprt)
 867 {
 868         if (xprt->tcp_offset == xprt->tcp_reclen) {
 869                 xprt->tcp_flags |= XPRT_COPY_RECM;
 870                 xprt->tcp_offset = 0;
 871                 if (xprt->tcp_flags & XPRT_LAST_FRAG) {
 872                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
 873                         xprt->tcp_flags |= XPRT_COPY_XID;
 874                         xprt->tcp_copied = 0;
 875                 }
 876         }
 877 }
 878
 879 /*
 880  * TCP read xid
 881  */
 882 static inline void
 883 tcp_read_xid(struct rpc_xprt *xprt, skb_reader_t *desc)
 884 {
 885         size_t len, used;
 886         char *p;
 887
 888         len = sizeof(xprt->tcp_xid) - xprt->tcp_offset;
 889         dprintk("RPC:      reading XID (%Zu bytes)\n", len);
 890         p = ((char *) &xprt->tcp_xid) + xprt->tcp_offset;
 891         used = tcp_copy_data(desc, p, len);
 892         xprt->tcp_offset += used;
 893         if (used != len)
 894                 return;
 895         xprt->tcp_flags &= ~XPRT_COPY_XID;
 896         xprt->tcp_flags |= XPRT_COPY_DATA;
 897         xprt->tcp_copied = 4;
 898         dprintk("RPC:      reading reply for XID %08x\n",
 899                                                 ntohl(xprt->tcp_xid));
 900         tcp_check_recm(xprt);
 901 }
 902
 903 /*
 904  * TCP read and complete request
 905  */
 906 static inline void
 907 tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
 908 {
 909         struct rpc_rqst *req;
 910         struct xdr_buf *rcvbuf;
 911         size_t len;
 912         int r;
 913
 914         /* Find and lock the request corresponding to this xid */
 915         spin_lock(&xprt->sock_lock);
 916         req = xprt_lookup_rqst(xprt, xprt->tcp_xid);
 917         if (!req) {
 918                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 919                 dprintk("RPC:      XID %08x request not found!\n",
 920                                 ntohl(xprt->tcp_xid));
 921                 spin_unlock(&xprt->sock_lock);
 922                 return;
 923         }
 924
 925         rcvbuf = &req->rq_private_buf;
 926         len = desc->count;
 927         if (len > xprt->tcp_reclen - xprt->tcp_offset) {
 928                 skb_reader_t my_desc;
 929
 930                 len = xprt->tcp_reclen - xprt->tcp_offset;
 931                 memcpy(&my_desc, desc, sizeof(my_desc));
 932                 my_desc.count = len;
 933                 r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
 934                                           &my_desc, tcp_copy_data);
 935                 desc->count -= len;
 936                 desc->offset += len;
 937         } else
 938                 r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
 939                                           desc, tcp_copy_data);
 940         xprt->tcp_copied += len;
 941         xprt->tcp_offset += len;
 942
 943         if (r < 0) {
 944                 /* Error when copying to the receive buffer,
 945                  * usually because we weren't able to allocate
 946                  * additional buffer pages. All we can do now
 947                  * is turn off XPRT_COPY_DATA, so the request
 948                  * will not receive any additional updates,
 949                  * and time out.
 950                  * Any remaining data from this record will
 951                  * be discarded.
 952                  */
 953                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 954                 goto out;
 955         }
 956
 957         if (xprt->tcp_copied == req->rq_private_buf.buflen)
 958                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 959         else if (xprt->tcp_offset == xprt->tcp_reclen) {
 960                 if (xprt->tcp_flags & XPRT_LAST_FRAG)
 961                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
 962         }
 963
 964         if (!(xprt->tcp_flags & XPRT_COPY_DATA)) {
 965                 dprintk("RPC: %4d received reply complete\n",
 966                                 req->rq_task->tk_pid);
 967                 xprt_complete_rqst(xprt, req, xprt->tcp_copied);
 968         }
 969 out:
 970         spin_unlock(&xprt->sock_lock);
 971         tcp_check_recm(xprt);
 972 }
 973
 974 /*
 975  * TCP discard extra bytes from a short read
 976  */
 977 static inline void
 978 tcp_read_discard(struct rpc_xprt *xprt, skb_reader_t *desc)
 979 {
 980         size_t len;
 981
 982         len = xprt->tcp_reclen - xprt->tcp_offset;
 983         if (len > desc->count)
 984                 len = desc->count;
 985         desc->count -= len;
 986         desc->offset += len;
 987         xprt->tcp_offset += len;
 988         tcp_check_recm(xprt);
 989 }
 990
 991 /*
 992  * TCP record receive routine
 993  * We first have to grab the record marker, then the XID, then the data.
 994  */
 995 static int
 996 tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 997                 unsigned int offset, size_t len)
 998 {
 999         struct rpc_xprt *xprt = rd_desc->arg.data;
1000         skb_reader_t desc = {
1001                 .skb    = skb,
1002                 .offset = offset,
1003                 .count  = len,
1004                 .csum   = 0
1005         };
1006
1007         dprintk("RPC:      tcp_data_recv\n");
1008         do {
1009                 /* Read in a new fragment marker if necessary */
1010                 /* Can we ever really expect to get completely empty fragments? */
1011                 if (xprt->tcp_flags & XPRT_COPY_RECM) {
1012                         tcp_read_fraghdr(xprt, &desc);
1013                         continue;
1014                 }
1015                 /* Read in the xid if necessary */
1016                 if (xprt->tcp_flags & XPRT_COPY_XID) {
1017                         tcp_read_xid(xprt, &desc);
1018                         continue;
1019                 }
1020                 /* Read in the request data */
1021                 if (xprt->tcp_flags & XPRT_COPY_DATA) {
1022                         tcp_read_request(xprt, &desc);
1023                         continue;
1024                 }
1025                 /* Skip over any trailing bytes on short reads */
1026                 tcp_read_discard(xprt, &desc);
1027         } while (desc.count);
1028         dprintk("RPC:      tcp_data_recv done\n");
1029         return len - desc.count;
1030 }
1031
1032 static void tcp_data_ready(struct sock *sk, int bytes)
1033 {
1034         struct rpc_xprt *xprt;
1035         read_descriptor_t rd_desc;
1036
1037         read_lock(&sk->sk_callback_lock);
1038         dprintk("RPC:      tcp_data_ready...\n");
1039         if (!(xprt = xprt_from_sock(sk))) {
1040                 printk("RPC:      tcp_data_ready socket info not found!\n");
1041                 goto out;
1042         }
1043         if (xprt->shutdown)
1044                 goto out;
1045
1046         /* We use rd_desc to pass struct xprt to tcp_data_recv */
1047         rd_desc.arg.data = xprt;
1048         rd_desc.count = 65536;
1049         tcp_read_sock(sk, &rd_desc, tcp_data_recv);
1050 out:
1051         read_unlock(&sk->sk_callback_lock);
1052 }
1053
1054 static void
1055 tcp_state_change(struct sock *sk)
1056 {
1057         struct rpc_xprt *xprt;
1058
1059         read_lock(&sk->sk_callback_lock);
1060         if (!(xprt = xprt_from_sock(sk)))
1061                 goto out;
1062         dprintk("RPC:      tcp_state_change client %p...\n", xprt);
1063         dprintk("RPC:      state %x conn %d dead %d zapped %d\n",
1064                                 sk->sk_state, xprt_connected(xprt),
1065                                 sock_flag(sk, SOCK_DEAD), sk->sk_zapped);
1066
1067         switch (sk->sk_state) {
1068         case TCP_ESTABLISHED:
1069                 spin_lock_bh(&xprt->sock_lock);
1070                 if (!xprt_test_and_set_connected(xprt)) {
1071                         /* Reset TCP record info */
1072                         xprt->tcp_offset = 0;
1073                         xprt->tcp_reclen = 0;
1074                         xprt->tcp_copied = 0;
1075                         xprt->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
1076                         rpc_wake_up(&xprt->pending);
1077                 }
1078                 spin_unlock_bh(&xprt->sock_lock);
1079                 break;
1080         case TCP_SYN_SENT:
1081         case TCP_SYN_RECV:
1082                 break;
1083         default:
1084                 if (xprt_test_and_clear_connected(xprt))
1085                         rpc_wake_up_status(&xprt->pending, -ENOTCONN);
1086                 break;
1087         }
1088  out:
1089         read_unlock(&sk->sk_callback_lock);
1090 }
1091
1092 /*
1093  * Called when more output buffer space is available for this socket.
1094  * We try not to wake our writers until they can make "significant"
1095  * progress, otherwise we'll waste resources thrashing sock_sendmsg
1096  * with a bunch of small requests.
1097  */
1098 static void
1099 xprt_write_space(struct sock *sk)
1100 {
1101         struct rpc_xprt *xprt;
1102         struct socket   *sock;
1103
1104         read_lock(&sk->sk_callback_lock);
1105         if (!(xprt = xprt_from_sock(sk)) || !(sock = sk->sk_socket))
1106                 goto out;
1107         if (xprt->shutdown)
1108                 goto out;
1109
1110         /* Wait until we have enough socket memory */
1111         if (xprt->stream) {
1112                 /* from net/core/stream.c:sk_stream_write_space */
1113                 if (sk_stream_wspace(sk) < sk_stream_min_wspace(sk))
1114                         goto out;
1115         } else {
1116                 /* from net/core/sock.c:sock_def_write_space */
1117                 if (!sock_writeable(sk))
1118                         goto out;
1119         }
1120
1121         if (!test_and_clear_bit(SOCK_NOSPACE, &sock->flags))
1122                 goto out;
1123
1124         spin_lock_bh(&xprt->sock_lock);
1125         if (xprt->snd_task)
1126                 rpc_wake_up_task(xprt->snd_task);
1127         spin_unlock_bh(&xprt->sock_lock);
1128 out:
1129         read_unlock(&sk->sk_callback_lock);
1130 }
1131
1132 /*
1133  * RPC receive timeout handler.
1134  */
1135 static void
1136 xprt_timer(struct rpc_task *task)
1137 {
1138         struct rpc_rqst *req = task->tk_rqstp;
1139         struct rpc_xprt *xprt = req->rq_xprt;
1140
1141         spin_lock(&xprt->sock_lock);
1142         if (req->rq_received)
1143                 goto out;
1144
1145         xprt_adjust_cwnd(req->rq_xprt, -ETIMEDOUT);
1146         __xprt_put_cong(xprt, req);
1147
1148         dprintk("RPC: %4d xprt_timer (%s request)\n",
1149                 task->tk_pid, req ? "pending" : "backlogged");
1150
1151         task->tk_status  = -ETIMEDOUT;
1152 out:
1153         task->tk_timeout = 0;
1154         rpc_wake_up_task(task);
1155         spin_unlock(&xprt->sock_lock);
1156 }
1157
1158 /*
1159  * Place the actual RPC call.
1160  * We have to copy the iovec because sendmsg fiddles with its contents.
1161  */
1162 int
1163 xprt_prepare_transmit(struct rpc_task *task)
1164 {
1165         struct rpc_rqst *req = task->tk_rqstp;
1166         struct rpc_xprt *xprt = req->rq_xprt;
1167         int err = 0;
1168
1169         dprintk("RPC: %4d xprt_prepare_transmit\n", task->tk_pid);
1170
1171         if (xprt->shutdown)
1172                 return -EIO;
1173
1174         spin_lock_bh(&xprt->sock_lock);
1175         if (req->rq_received && !req->rq_bytes_sent) {
1176                 err = req->rq_received;
1177                 goto out_unlock;
1178         }
1179         if (!__xprt_lock_write(xprt, task)) {
1180                 err = -EAGAIN;
1181                 goto out_unlock;
1182         }
1183
1184         if (!xprt_connected(xprt)) {
1185                 err = -ENOTCONN;
1186                 goto out_unlock;
1187         }
1188 out_unlock:
1189         spin_unlock_bh(&xprt->sock_lock);
1190         return err;
1191 }
1192
1193 void
1194 xprt_transmit(struct rpc_task *task)
1195 {
1196         struct rpc_clnt *clnt = task->tk_client;
1197         struct rpc_rqst *req = task->tk_rqstp;
1198         struct rpc_xprt *xprt = req->rq_xprt;
1199         int status, retry = 0;
1200
1201
1202         dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
1203
1204         /* set up everything as needed. */
1205         /* Write the record marker */
1206         if (xprt->stream) {
1207                 u32     *marker = req->rq_svec[0].iov_base;
1208
1209                 *marker = htonl(0x80000000|(req->rq_slen-sizeof(*marker)));
1210         }
1211
1212         smp_rmb();
1213         if (!req->rq_received) {
1214                 if (list_empty(&req->rq_list)) {
1215                         spin_lock_bh(&xprt->sock_lock);
1216                         /* Update the softirq receive buffer */
1217                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
1218                                         sizeof(req->rq_private_buf));
1219                         /* Add request to the receive list */
1220                         list_add_tail(&req->rq_list, &xprt->recv);
1221                         spin_unlock_bh(&xprt->sock_lock);
1222                         xprt_reset_majortimeo(req);
1223                 }
1224         } else if (!req->rq_bytes_sent)
1225                 return;
1226
1227         /* Continue transmitting the packet/record. We must be careful
1228          * to cope with writespace callbacks arriving _after_ we have
1229          * called xprt_sendmsg().
1230          */
1231         while (1) {
1232                 req->rq_xtime = jiffies;
1233                 status = xprt_sendmsg(xprt, req);
1234
1235                 if (status < 0)
1236                         break;
1237
1238                 if (xprt->stream) {
1239                         req->rq_bytes_sent += status;
1240
1241                         /* If we've sent the entire packet, immediately
1242                          * reset the count of bytes sent. */
1243                         if (req->rq_bytes_sent >= req->rq_slen) {
1244                                 req->rq_bytes_sent = 0;
1245                                 goto out_receive;
1246                         }
1247                 } else {
1248                         if (status >= req->rq_slen)
1249                                 goto out_receive;
1250                         status = -EAGAIN;
1251                         break;
1252                 }
1253
1254                 dprintk("RPC: %4d xmit incomplete (%d left of %d)\n",
1255                                 task->tk_pid, req->rq_slen - req->rq_bytes_sent,
1256                                 req->rq_slen);
1257
1258                 status = -EAGAIN;
1259                 if (retry++ > 50)
1260                         break;
1261         }
1262
1263         /* Note: at this point, task->tk_sleeping has not yet been set,
1264          *       hence there is no danger of the waking up task being put on
1265          *       schedq, and being picked up by a parallel run of rpciod().
1266          */
1267         task->tk_status = status;
1268
1269         switch (status) {
1270         case -EAGAIN:
1271                 if (test_bit(SOCK_ASYNC_NOSPACE, &xprt->sock->flags)) {
1272                         /* Protect against races with xprt_write_space */
1273                         spin_lock_bh(&xprt->sock_lock);
1274                         /* Don't race with disconnect */
1275                         if (!xprt_connected(xprt))
1276                                 task->tk_status = -ENOTCONN;
1277                         else if (test_bit(SOCK_NOSPACE, &xprt->sock->flags)) {
1278                                 task->tk_timeout = req->rq_timeout;
1279                                 rpc_sleep_on(&xprt->pending, task, NULL, NULL);
1280                         }
1281                         spin_unlock_bh(&xprt->sock_lock);
1282                         return;
1283                 }
1284                 /* Keep holding the socket if it is blocked */
1285                 rpc_delay(task, HZ>>4);
1286                 return;
1287         case -ECONNREFUSED:
1288                 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
1289                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
1290         case -ENOTCONN:
1291                 return;
1292         default:
1293                 if (xprt->stream)
1294                         xprt_disconnect(xprt);
1295         }
1296         xprt_release_write(xprt, task);
1297         return;
1298  out_receive:
1299         dprintk("RPC: %4d xmit complete\n", task->tk_pid);
1300         /* Set the task's receive timeout value */
1301         spin_lock_bh(&xprt->sock_lock);
1302         if (!xprt->nocong) {
1303                 int timer = task->tk_msg.rpc_proc->p_timer;
1304                 task->tk_timeout = rpc_calc_rto(clnt->cl_rtt, timer);
1305                 task->tk_timeout <<= rpc_ntimeo(clnt->cl_rtt, timer) + req->rq_retries;
1306                 if (task->tk_timeout > xprt->timeout.to_maxval || task->tk_timeout == 0)
1307                         task->tk_timeout = xprt->timeout.to_maxval;
1308         } else
1309                 task->tk_timeout = req->rq_timeout;
1310         /* Don't race with disconnect */
1311         if (!xprt_connected(xprt))
1312                 task->tk_status = -ENOTCONN;
1313         else if (!req->rq_received)
1314                 rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
1315         __xprt_release_write(xprt, task);
1316         spin_unlock_bh(&xprt->sock_lock);
1317 }
1318
1319 /*
1320  * Reserve an RPC call slot.
1321  */
1322 static inline void
1323 do_xprt_reserve(struct rpc_task *task)
1324 {
1325         struct rpc_xprt *xprt = task->tk_xprt;
1326
1327         task->tk_status = 0;
1328         if (task->tk_rqstp)
1329                 return;
1330         if (!list_empty(&xprt->free)) {
1331                 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1332                 list_del_init(&req->rq_list);
1333                 task->tk_rqstp = req;
1334                 xprt_request_init(task, xprt);
1335                 return;
1336         }
1337         dprintk("RPC:      waiting for request slot\n");
1338         task->tk_status = -EAGAIN;
1339         task->tk_timeout = 0;
1340         rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
1341 }
1342
1343 void
1344 xprt_reserve(struct rpc_task *task)
1345 {
1346         struct rpc_xprt *xprt = task->tk_xprt;
1347
1348         task->tk_status = -EIO;
1349         if (!xprt->shutdown) {
1350                 spin_lock(&xprt->xprt_lock);
1351                 do_xprt_reserve(task);
1352                 spin_unlock(&xprt->xprt_lock);
1353                 if (task->tk_rqstp)
1354                         del_timer_sync(&xprt->timer);
1355         }
1356 }
1357
1358 /*
1359  * Allocate a 'unique' XID
1360  */
1361 static inline u32 xprt_alloc_xid(struct rpc_xprt *xprt)
1362 {
1363         return xprt->xid++;
1364 }
1365
1366 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1367 {
1368         get_random_bytes(&xprt->xid, sizeof(xprt->xid));
1369 }
1370
1371 /*
1372  * Initialize RPC request
1373  */
1374 static void
1375 xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1376 {
1377         struct rpc_rqst *req = task->tk_rqstp;
1378
1379         req->rq_timeout = xprt->timeout.to_initval;
1380         req->rq_task    = task;
1381         req->rq_xprt    = xprt;
1382         req->rq_xid     = xprt_alloc_xid(xprt);
1383         dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
1384                         req, ntohl(req->rq_xid));
1385 }
1386
1387 /*
1388  * Release an RPC call slot
1389  */
1390 void
1391 xprt_release(struct rpc_task *task)
1392 {
1393         struct rpc_xprt *xprt = task->tk_xprt;
1394         struct rpc_rqst *req;
1395
1396         if (!(req = task->tk_rqstp))
1397                 return;
1398         spin_lock_bh(&xprt->sock_lock);
1399         __xprt_release_write(xprt, task);
1400         __xprt_put_cong(xprt, req);
1401         if (!list_empty(&req->rq_list))
1402                 list_del(&req->rq_list);
1403         xprt->last_used = jiffies;
1404         if (list_empty(&xprt->recv) && !xprt->shutdown)
1405                 mod_timer(&xprt->timer, xprt->last_used + XPRT_IDLE_TIMEOUT);
1406         spin_unlock_bh(&xprt->sock_lock);
1407         task->tk_rqstp = NULL;
1408         memset(req, 0, sizeof(*req));   /* mark unused */
1409
1410         dprintk("RPC: %4d release request %p\n", task->tk_pid, req);
1411
1412         spin_lock(&xprt->xprt_lock);
1413         list_add(&req->rq_list, &xprt->free);
1414         xprt_clear_backlog(xprt);
1415         spin_unlock(&xprt->xprt_lock);
1416 }
1417
1418 /*
1419  * Set default timeout parameters
1420  */
1421 static void
1422 xprt_default_timeout(struct rpc_timeout *to, int proto)
1423 {
1424         if (proto == IPPROTO_UDP)
1425                 xprt_set_timeout(to, 5,  5 * HZ);
1426         else
1427                 xprt_set_timeout(to, 5, 60 * HZ);
1428 }
1429
1430 /*
1431  * Set constant timeout
1432  */
1433 void
1434 xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
1435 {
1436         to->to_initval   =
1437         to->to_increment = incr;
1438         to->to_maxval    = incr * retr;
1439         to->to_retries   = retr;
1440         to->to_exponential = 0;
1441 }
1442
1443 unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1444 unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1445
1446 /*
1447  * Initialize an RPC client
1448  */
1449 static struct rpc_xprt *
1450 xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
1451 {
1452         struct rpc_xprt *xprt;
1453         unsigned int entries;
1454         size_t slot_table_size;
1455         struct rpc_rqst *req;
1456
1457         dprintk("RPC:      setting up %s transport...\n",
1458                                 proto == IPPROTO_UDP? "UDP" : "TCP");
1459
1460         entries = (proto == IPPROTO_TCP)?
1461                 xprt_tcp_slot_table_entries : xprt_udp_slot_table_entries;
1462
1463         if ((xprt = kmalloc(sizeof(struct rpc_xprt), GFP_KERNEL)) == NULL)
1464                 return ERR_PTR(-ENOMEM);
1465         memset(xprt, 0, sizeof(*xprt)); /* Nnnngh! */
1466         xprt->max_reqs = entries;
1467         slot_table_size = entries * sizeof(xprt->slot[0]);
1468         xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
1469         if (xprt->slot == NULL) {
1470                 kfree(xprt);
1471                 return ERR_PTR(-ENOMEM);
1472         }
1473         memset(xprt->slot, 0, slot_table_size);
1474
1475         xprt->addr = *ap;
1476         xprt->prot = proto;
1477         xprt->stream = (proto == IPPROTO_TCP)? 1 : 0;
1478         if (xprt->stream) {
1479                 xprt->cwnd = RPC_MAXCWND(xprt);
1480                 xprt->nocong = 1;
1481         } else
1482                 xprt->cwnd = RPC_INITCWND;
1483         spin_lock_init(&xprt->sock_lock);
1484         spin_lock_init(&xprt->xprt_lock);
1485         init_waitqueue_head(&xprt->cong_wait);
1486
1487         INIT_LIST_HEAD(&xprt->free);
1488         INIT_LIST_HEAD(&xprt->recv);
1489         INIT_WORK(&xprt->sock_connect, xprt_socket_connect, xprt);
1490         INIT_WORK(&xprt->task_cleanup, xprt_socket_autoclose, xprt);
1491         init_timer(&xprt->timer);
1492         xprt->timer.function = xprt_init_autodisconnect;
1493         xprt->timer.data = (unsigned long) xprt;
1494         xprt->last_used = jiffies;
1495         xprt->port = XPRT_MAX_RESVPORT;
1496
1497         /* Set timeout parameters */
1498         if (to) {
1499                 xprt->timeout = *to;
1500         } else
1501                 xprt_default_timeout(&xprt->timeout, xprt->prot);
1502
1503         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1504         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
1505         rpc_init_wait_queue(&xprt->resend, "xprt_resend");
1506         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1507
1508         /* initialize free list */
1509         for (req = &xprt->slot[entries-1]; req >= &xprt->slot[0]; req--)
1510                 list_add(&req->rq_list, &xprt->free);
1511
1512         xprt_init_xid(xprt);
1513
1514         /* Check whether we want to use a reserved port */
1515         xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1516
1517         dprintk("RPC:      created transport %p with %u slots\n", xprt,
1518                         xprt->max_reqs);
1519
1520         return xprt;
1521 }
1522
1523 /*
1524  * Bind to a reserved port
1525  */
1526 static inline int xprt_bindresvport(struct rpc_xprt *xprt, struct socket *sock)
1527 {
1528         struct sockaddr_in myaddr = {
1529                 .sin_family = AF_INET,
1530         };
1531         int             err, port;
1532
1533         /* Were we already bound to a given port? Try to reuse it */
1534         port = xprt->port;
1535         do {
1536                 myaddr.sin_port = htons(port);
1537                 err = sock->ops->bind(sock, (struct sockaddr *) &myaddr,
1538                                                 sizeof(myaddr));
1539                 if (err == 0) {
1540                         xprt->port = port;
1541                         return 0;
1542                 }
1543                 if (--port == 0)
1544                         port = XPRT_MAX_RESVPORT;
1545         } while (err == -EADDRINUSE && port != xprt->port);
1546
1547         printk("RPC: Can't bind to reserved port (%d).\n", -err);
1548         return err;
1549 }
1550
1551 static void
1552 xprt_bind_socket(struct rpc_xprt *xprt, struct socket *sock)
1553 {
1554         struct sock     *sk = sock->sk;
1555
1556         if (xprt->inet)
1557                 return;
1558
1559         write_lock_bh(&sk->sk_callback_lock);
1560         sk->sk_user_data = xprt;
1561         xprt->old_data_ready = sk->sk_data_ready;
1562         xprt->old_state_change = sk->sk_state_change;
1563         xprt->old_write_space = sk->sk_write_space;
1564         if (xprt->prot == IPPROTO_UDP) {
1565                 sk->sk_data_ready = udp_data_ready;
1566                 sk->sk_no_check = UDP_CSUM_NORCV;
1567                 xprt_set_connected(xprt);
1568         } else {
1569                 tcp_sk(sk)->nonagle = 1;        /* disable Nagle's algorithm */
1570                 sk->sk_data_ready = tcp_data_ready;
1571                 sk->sk_state_change = tcp_state_change;
1572                 xprt_clear_connected(xprt);
1573         }
1574         sk->sk_write_space = xprt_write_space;
1575
1576         /* Reset to new socket */
1577         xprt->sock = sock;
1578         xprt->inet = sk;
1579         write_unlock_bh(&sk->sk_callback_lock);
1580
1581         return;
1582 }
1583
1584 /*
1585  * Set socket buffer length
1586  */
1587 void
1588 xprt_sock_setbufsize(struct rpc_xprt *xprt)
1589 {
1590         struct sock *sk = xprt->inet;
1591
1592         if (xprt->stream)
1593                 return;
1594         if (xprt->rcvsize) {
1595                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
1596                 sk->sk_rcvbuf = xprt->rcvsize * xprt->max_reqs *  2;
1597         }
1598         if (xprt->sndsize) {
1599                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
1600                 sk->sk_sndbuf = xprt->sndsize * xprt->max_reqs * 2;
1601                 sk->sk_write_space(sk);
1602         }
1603 }
1604
1605 /*
1606  * Datastream sockets are created here, but xprt_connect will create
1607  * and connect stream sockets.
1608  */
1609 static struct socket * xprt_create_socket(struct rpc_xprt *xprt, int proto, int resvport)
1610 {
1611         struct socket   *sock;
1612         int             type, err;
1613
1614         dprintk("RPC:      xprt_create_socket(%s %d)\n",
1615                            (proto == IPPROTO_UDP)? "udp" : "tcp", proto);
1616
1617         type = (proto == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1618
1619         if ((err = sock_create_kern(PF_INET, type, proto, &sock)) < 0) {
1620                 printk("RPC: can't create socket (%d).\n", -err);
1621                 return NULL;
1622         }
1623
1624         /* If the caller has the capability, bind to a reserved port */
1625         if (resvport && xprt_bindresvport(xprt, sock) < 0) {
1626                 printk("RPC: can't bind to reserved port.\n");
1627                 goto failed;
1628         }
1629
1630         return sock;
1631
1632 failed:
1633         sock_release(sock);
1634         return NULL;
1635 }
1636
1637 /*
1638  * Create an RPC client transport given the protocol and peer address.
1639  */
1640 struct rpc_xprt *
1641 xprt_create_proto(int proto, struct sockaddr_in *sap, struct rpc_timeout *to)
1642 {
1643         struct rpc_xprt *xprt;
1644
1645         xprt = xprt_setup(proto, sap, to);
1646         if (IS_ERR(xprt))
1647                 dprintk("RPC:      xprt_create_proto failed\n");
1648         else
1649                 dprintk("RPC:      xprt_create_proto created xprt %p\n", xprt);
1650         return xprt;
1651 }
1652
1653 /*
1654  * Prepare for transport shutdown.
1655  */
1656 static void
1657 xprt_shutdown(struct rpc_xprt *xprt)
1658 {
1659         xprt->shutdown = 1;
1660         rpc_wake_up(&xprt->sending);
1661         rpc_wake_up(&xprt->resend);
1662         rpc_wake_up(&xprt->pending);
1663         rpc_wake_up(&xprt->backlog);
1664         wake_up(&xprt->cong_wait);
1665         del_timer_sync(&xprt->timer);
1666 }
1667
1668 /*
1669  * Clear the xprt backlog queue
1670  */
1671 static int
1672 xprt_clear_backlog(struct rpc_xprt *xprt) {
1673         rpc_wake_up_next(&xprt->backlog);
1674         wake_up(&xprt->cong_wait);
1675         return 1;
1676 }
1677
1678 /*
1679  * Destroy an RPC transport, killing off all requests.
1680  */
1681 int
1682 xprt_destroy(struct rpc_xprt *xprt)
1683 {
1684         dprintk("RPC:      destroying transport %p\n", xprt);
1685         xprt_shutdown(xprt);
1686         xprt_disconnect(xprt);
1687         xprt_close(xprt);
1688         kfree(xprt->slot);
1689         kfree(xprt);
1690
1691         return 0;
1692 }