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         xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_and_csum_bits);
 729         if (desc.offset != skb->len) {
 730                 unsigned int csum2;
 731                 csum2 = skb_checksum(skb, desc.offset, skb->len - desc.offset, 0);
 732                 desc.csum = csum_block_add(desc.csum, csum2, desc.offset);
 733         }
 734         if (desc.count)
 735                 return -1;
 736         if ((unsigned short)csum_fold(desc.csum))
 737                 return -1;
 738         return 0;
 739 no_checksum:
 740         xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_bits);
 741         if (desc.count)
 742                 return -1;
 743         return 0;
 744 }
 745
 746 /*
 747  * Input handler for RPC replies. Called from a bottom half and hence
 748  * atomic.
 749  */
 750 static void
 751 udp_data_ready(struct sock *sk, int len)
 752 {
 753         struct rpc_task *task;
 754         struct rpc_xprt *xprt;
 755         struct rpc_rqst *rovr;
 756         struct sk_buff  *skb;
 757         int err, repsize, copied;
 758         u32 _xid, *xp;
 759
 760         read_lock(&sk->sk_callback_lock);
 761         dprintk("RPC:      udp_data_ready...\n");
 762         if (!(xprt = xprt_from_sock(sk))) {
 763                 printk("RPC:      udp_data_ready request not found!\n");
 764                 goto out;
 765         }
 766
 767         dprintk("RPC:      udp_data_ready client %p\n", xprt);
 768
 769         if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
 770                 goto out;
 771
 772         if (xprt->shutdown)
 773                 goto dropit;
 774
 775         repsize = skb->len - sizeof(struct udphdr);
 776         if (repsize < 4) {
 777                 printk("RPC: impossible RPC reply size %d!\n", repsize);
 778                 goto dropit;
 779         }
 780
 781         /* Copy the XID from the skb... */
 782         xp = skb_header_pointer(skb, sizeof(struct udphdr),
 783                                 sizeof(_xid), &_xid);
 784         if (xp == NULL)
 785                 goto dropit;
 786
 787         /* Look up and lock the request corresponding to the given XID */
 788         spin_lock(&xprt->sock_lock);
 789         rovr = xprt_lookup_rqst(xprt, *xp);
 790         if (!rovr)
 791                 goto out_unlock;
 792         task = rovr->rq_task;
 793
 794         dprintk("RPC: %4d received reply\n", task->tk_pid);
 795
 796         if ((copied = rovr->rq_private_buf.buflen) > repsize)
 797                 copied = repsize;
 798
 799         /* Suck it into the iovec, verify checksum if not done by hw. */
 800         if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
 801                 goto out_unlock;
 802
 803         /* Something worked... */
 804         dst_confirm(skb->dst);
 805
 806         xprt_complete_rqst(xprt, rovr, copied);
 807
 808  out_unlock:
 809         spin_unlock(&xprt->sock_lock);
 810  dropit:
 811         skb_free_datagram(sk, skb);
 812  out:
 813         read_unlock(&sk->sk_callback_lock);
 814 }
 815
 816 /*
 817  * Copy from an skb into memory and shrink the skb.
 818  */
 819 static inline size_t
 820 tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
 821 {
 822         if (len > desc->count)
 823                 len = desc->count;
 824         if (skb_copy_bits(desc->skb, desc->offset, p, len))
 825                 return 0;
 826         desc->offset += len;
 827         desc->count -= len;
 828         return len;
 829 }
 830
 831 /*
 832  * TCP read fragment marker
 833  */
 834 static inline void
 835 tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
 836 {
 837         size_t len, used;
 838         char *p;
 839
 840         p = ((char *) &xprt->tcp_recm) + xprt->tcp_offset;
 841         len = sizeof(xprt->tcp_recm) - xprt->tcp_offset;
 842         used = tcp_copy_data(desc, p, len);
 843         xprt->tcp_offset += used;
 844         if (used != len)
 845                 return;
 846         xprt->tcp_reclen = ntohl(xprt->tcp_recm);
 847         if (xprt->tcp_reclen & 0x80000000)
 848                 xprt->tcp_flags |= XPRT_LAST_FRAG;
 849         else
 850                 xprt->tcp_flags &= ~XPRT_LAST_FRAG;
 851         xprt->tcp_reclen &= 0x7fffffff;
 852         xprt->tcp_flags &= ~XPRT_COPY_RECM;
 853         xprt->tcp_offset = 0;
 854         /* Sanity check of the record length */
 855         if (xprt->tcp_reclen < 4) {
 856                 printk(KERN_ERR "RPC: Invalid TCP record fragment length\n");
 857                 xprt_disconnect(xprt);
 858         }
 859         dprintk("RPC:      reading TCP record fragment of length %d\n",
 860                         xprt->tcp_reclen);
 861 }
 862
 863 static void
 864 tcp_check_recm(struct rpc_xprt *xprt)
 865 {
 866         if (xprt->tcp_offset == xprt->tcp_reclen) {
 867                 xprt->tcp_flags |= XPRT_COPY_RECM;
 868                 xprt->tcp_offset = 0;
 869                 if (xprt->tcp_flags & XPRT_LAST_FRAG) {
 870                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
 871                         xprt->tcp_flags |= XPRT_COPY_XID;
 872                         xprt->tcp_copied = 0;
 873                 }
 874         }
 875 }
 876
 877 /*
 878  * TCP read xid
 879  */
 880 static inline void
 881 tcp_read_xid(struct rpc_xprt *xprt, skb_reader_t *desc)
 882 {
 883         size_t len, used;
 884         char *p;
 885
 886         len = sizeof(xprt->tcp_xid) - xprt->tcp_offset;
 887         dprintk("RPC:      reading XID (%Zu bytes)\n", len);
 888         p = ((char *) &xprt->tcp_xid) + xprt->tcp_offset;
 889         used = tcp_copy_data(desc, p, len);
 890         xprt->tcp_offset += used;
 891         if (used != len)
 892                 return;
 893         xprt->tcp_flags &= ~XPRT_COPY_XID;
 894         xprt->tcp_flags |= XPRT_COPY_DATA;
 895         xprt->tcp_copied = 4;
 896         dprintk("RPC:      reading reply for XID %08x\n",
 897                                                 ntohl(xprt->tcp_xid));
 898         tcp_check_recm(xprt);
 899 }
 900
 901 /*
 902  * TCP read and complete request
 903  */
 904 static inline void
 905 tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
 906 {
 907         struct rpc_rqst *req;
 908         struct xdr_buf *rcvbuf;
 909         size_t len;
 910
 911         /* Find and lock the request corresponding to this xid */
 912         spin_lock(&xprt->sock_lock);
 913         req = xprt_lookup_rqst(xprt, xprt->tcp_xid);
 914         if (!req) {
 915                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 916                 dprintk("RPC:      XID %08x request not found!\n",
 917                                 ntohl(xprt->tcp_xid));
 918                 spin_unlock(&xprt->sock_lock);
 919                 return;
 920         }
 921
 922         rcvbuf = &req->rq_private_buf;
 923         len = desc->count;
 924         if (len > xprt->tcp_reclen - xprt->tcp_offset) {
 925                 skb_reader_t my_desc;
 926
 927                 len = xprt->tcp_reclen - xprt->tcp_offset;
 928                 memcpy(&my_desc, desc, sizeof(my_desc));
 929                 my_desc.count = len;
 930                 xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
 931                                           &my_desc, tcp_copy_data);
 932                 desc->count -= len;
 933                 desc->offset += len;
 934         } else
 935                 xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
 936                                           desc, tcp_copy_data);
 937         xprt->tcp_copied += len;
 938         xprt->tcp_offset += len;
 939
 940         if (xprt->tcp_copied == req->rq_private_buf.buflen)
 941                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 942         else if (xprt->tcp_offset == xprt->tcp_reclen) {
 943                 if (xprt->tcp_flags & XPRT_LAST_FRAG)
 944                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
 945         }
 946
 947         if (!(xprt->tcp_flags & XPRT_COPY_DATA)) {
 948                 dprintk("RPC: %4d received reply complete\n",
 949                                 req->rq_task->tk_pid);
 950                 xprt_complete_rqst(xprt, req, xprt->tcp_copied);
 951         }
 952         spin_unlock(&xprt->sock_lock);
 953         tcp_check_recm(xprt);
 954 }
 955
 956 /*
 957  * TCP discard extra bytes from a short read
 958  */
 959 static inline void
 960 tcp_read_discard(struct rpc_xprt *xprt, skb_reader_t *desc)
 961 {
 962         size_t len;
 963
 964         len = xprt->tcp_reclen - xprt->tcp_offset;
 965         if (len > desc->count)
 966                 len = desc->count;
 967         desc->count -= len;
 968         desc->offset += len;
 969         xprt->tcp_offset += len;
 970         tcp_check_recm(xprt);
 971 }
 972
 973 /*
 974  * TCP record receive routine
 975  * We first have to grab the record marker, then the XID, then the data.
 976  */
 977 static int
 978 tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 979                 unsigned int offset, size_t len)
 980 {
 981         struct rpc_xprt *xprt = rd_desc->arg.data;
 982         skb_reader_t desc = {
 983                 .skb    = skb,
 984                 .offset = offset,
 985                 .count  = len,
 986                 .csum   = 0
 987         };
 988
 989         dprintk("RPC:      tcp_data_recv\n");
 990         do {
 991                 /* Read in a new fragment marker if necessary */
 992                 /* Can we ever really expect to get completely empty fragments? */
 993                 if (xprt->tcp_flags & XPRT_COPY_RECM) {
 994                         tcp_read_fraghdr(xprt, &desc);
 995                         continue;
 996                 }
 997                 /* Read in the xid if necessary */
 998                 if (xprt->tcp_flags & XPRT_COPY_XID) {
 999                         tcp_read_xid(xprt, &desc);
1000                         continue;
1001                 }
1002                 /* Read in the request data */
1003                 if (xprt->tcp_flags & XPRT_COPY_DATA) {
1004                         tcp_read_request(xprt, &desc);
1005                         continue;
1006                 }
1007                 /* Skip over any trailing bytes on short reads */
1008                 tcp_read_discard(xprt, &desc);
1009         } while (desc.count);
1010         dprintk("RPC:      tcp_data_recv done\n");
1011         return len - desc.count;
1012 }
1013
1014 static void tcp_data_ready(struct sock *sk, int bytes)
1015 {
1016         struct rpc_xprt *xprt;
1017         read_descriptor_t rd_desc;
1018
1019         read_lock(&sk->sk_callback_lock);
1020         dprintk("RPC:      tcp_data_ready...\n");
1021         if (!(xprt = xprt_from_sock(sk))) {
1022                 printk("RPC:      tcp_data_ready socket info not found!\n");
1023                 goto out;
1024         }
1025         if (xprt->shutdown)
1026                 goto out;
1027
1028         /* We use rd_desc to pass struct xprt to tcp_data_recv */
1029         rd_desc.arg.data = xprt;
1030         rd_desc.count = 65536;
1031         tcp_read_sock(sk, &rd_desc, tcp_data_recv);
1032 out:
1033         read_unlock(&sk->sk_callback_lock);
1034 }
1035
1036 static void
1037 tcp_state_change(struct sock *sk)
1038 {
1039         struct rpc_xprt *xprt;
1040
1041         read_lock(&sk->sk_callback_lock);
1042         if (!(xprt = xprt_from_sock(sk)))
1043                 goto out;
1044         dprintk("RPC:      tcp_state_change client %p...\n", xprt);
1045         dprintk("RPC:      state %x conn %d dead %d zapped %d\n",
1046                                 sk->sk_state, xprt_connected(xprt),
1047                                 sock_flag(sk, SOCK_DEAD), sk->sk_zapped);
1048
1049         switch (sk->sk_state) {
1050         case TCP_ESTABLISHED:
1051                 spin_lock_bh(&xprt->sock_lock);
1052                 if (!xprt_test_and_set_connected(xprt)) {
1053                         /* Reset TCP record info */
1054                         xprt->tcp_offset = 0;
1055                         xprt->tcp_reclen = 0;
1056                         xprt->tcp_copied = 0;
1057                         xprt->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
1058                         rpc_wake_up(&xprt->pending);
1059                 }
1060                 spin_unlock_bh(&xprt->sock_lock);
1061                 break;
1062         case TCP_SYN_SENT:
1063         case TCP_SYN_RECV:
1064                 break;
1065         default:
1066                 if (xprt_test_and_clear_connected(xprt))
1067                         rpc_wake_up_status(&xprt->pending, -ENOTCONN);
1068                 break;
1069         }
1070  out:
1071         read_unlock(&sk->sk_callback_lock);
1072 }
1073
1074 /*
1075  * Called when more output buffer space is available for this socket.
1076  * We try not to wake our writers until they can make "significant"
1077  * progress, otherwise we'll waste resources thrashing sock_sendmsg
1078  * with a bunch of small requests.
1079  */
1080 static void
1081 xprt_write_space(struct sock *sk)
1082 {
1083         struct rpc_xprt *xprt;
1084         struct socket   *sock;
1085
1086         read_lock(&sk->sk_callback_lock);
1087         if (!(xprt = xprt_from_sock(sk)) || !(sock = sk->sk_socket))
1088                 goto out;
1089         if (xprt->shutdown)
1090                 goto out;
1091
1092         /* Wait until we have enough socket memory */
1093         if (xprt->stream) {
1094                 /* from net/core/stream.c:sk_stream_write_space */
1095                 if (sk_stream_wspace(sk) < sk_stream_min_wspace(sk))
1096                         goto out;
1097         } else {
1098                 /* from net/core/sock.c:sock_def_write_space */
1099                 if (!sock_writeable(sk))
1100                         goto out;
1101         }
1102
1103         if (!test_and_clear_bit(SOCK_NOSPACE, &sock->flags))
1104                 goto out;
1105
1106         spin_lock_bh(&xprt->sock_lock);
1107         if (xprt->snd_task)
1108                 rpc_wake_up_task(xprt->snd_task);
1109         spin_unlock_bh(&xprt->sock_lock);
1110 out:
1111         read_unlock(&sk->sk_callback_lock);
1112 }
1113
1114 /*
1115  * RPC receive timeout handler.
1116  */
1117 static void
1118 xprt_timer(struct rpc_task *task)
1119 {
1120         struct rpc_rqst *req = task->tk_rqstp;
1121         struct rpc_xprt *xprt = req->rq_xprt;
1122
1123         spin_lock(&xprt->sock_lock);
1124         if (req->rq_received)
1125                 goto out;
1126
1127         xprt_adjust_cwnd(req->rq_xprt, -ETIMEDOUT);
1128         __xprt_put_cong(xprt, req);
1129
1130         dprintk("RPC: %4d xprt_timer (%s request)\n",
1131                 task->tk_pid, req ? "pending" : "backlogged");
1132
1133         task->tk_status  = -ETIMEDOUT;
1134 out:
1135         task->tk_timeout = 0;
1136         rpc_wake_up_task(task);
1137         spin_unlock(&xprt->sock_lock);
1138 }
1139
1140 /*
1141  * Place the actual RPC call.
1142  * We have to copy the iovec because sendmsg fiddles with its contents.
1143  */
1144 int
1145 xprt_prepare_transmit(struct rpc_task *task)
1146 {
1147         struct rpc_rqst *req = task->tk_rqstp;
1148         struct rpc_xprt *xprt = req->rq_xprt;
1149         int err = 0;
1150
1151         dprintk("RPC: %4d xprt_prepare_transmit\n", task->tk_pid);
1152
1153         if (xprt->shutdown)
1154                 return -EIO;
1155
1156         spin_lock_bh(&xprt->sock_lock);
1157         if (req->rq_received && !req->rq_bytes_sent) {
1158                 err = req->rq_received;
1159                 goto out_unlock;
1160         }
1161         if (!__xprt_lock_write(xprt, task)) {
1162                 err = -EAGAIN;
1163                 goto out_unlock;
1164         }
1165
1166         if (!xprt_connected(xprt)) {
1167                 err = -ENOTCONN;
1168                 goto out_unlock;
1169         }
1170 out_unlock:
1171         spin_unlock_bh(&xprt->sock_lock);
1172         return err;
1173 }
1174
1175 void
1176 xprt_transmit(struct rpc_task *task)
1177 {
1178         struct rpc_clnt *clnt = task->tk_client;
1179         struct rpc_rqst *req = task->tk_rqstp;
1180         struct rpc_xprt *xprt = req->rq_xprt;
1181         int status, retry = 0;
1182
1183
1184         dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
1185
1186         /* set up everything as needed. */
1187         /* Write the record marker */
1188         if (xprt->stream) {
1189                 u32     *marker = req->rq_svec[0].iov_base;
1190
1191                 *marker = htonl(0x80000000|(req->rq_slen-sizeof(*marker)));
1192         }
1193
1194         smp_rmb();
1195         if (!req->rq_received) {
1196                 if (list_empty(&req->rq_list)) {
1197                         spin_lock_bh(&xprt->sock_lock);
1198                         /* Update the softirq receive buffer */
1199                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
1200                                         sizeof(req->rq_private_buf));
1201                         /* Add request to the receive list */
1202                         list_add_tail(&req->rq_list, &xprt->recv);
1203                         spin_unlock_bh(&xprt->sock_lock);
1204                         xprt_reset_majortimeo(req);
1205                 }
1206         } else if (!req->rq_bytes_sent)
1207                 return;
1208
1209         /* Continue transmitting the packet/record. We must be careful
1210          * to cope with writespace callbacks arriving _after_ we have
1211          * called xprt_sendmsg().
1212          */
1213         while (1) {
1214                 req->rq_xtime = jiffies;
1215                 status = xprt_sendmsg(xprt, req);
1216
1217                 if (status < 0)
1218                         break;
1219
1220                 if (xprt->stream) {
1221                         req->rq_bytes_sent += status;
1222
1223                         /* If we've sent the entire packet, immediately
1224                          * reset the count of bytes sent. */
1225                         if (req->rq_bytes_sent >= req->rq_slen) {
1226                                 req->rq_bytes_sent = 0;
1227                                 goto out_receive;
1228                         }
1229                 } else {
1230                         if (status >= req->rq_slen)
1231                                 goto out_receive;
1232                         status = -EAGAIN;
1233                         break;
1234                 }
1235
1236                 dprintk("RPC: %4d xmit incomplete (%d left of %d)\n",
1237                                 task->tk_pid, req->rq_slen - req->rq_bytes_sent,
1238                                 req->rq_slen);
1239
1240                 status = -EAGAIN;
1241                 if (retry++ > 50)
1242                         break;
1243         }
1244
1245         /* Note: at this point, task->tk_sleeping has not yet been set,
1246          *       hence there is no danger of the waking up task being put on
1247          *       schedq, and being picked up by a parallel run of rpciod().
1248          */
1249         task->tk_status = status;
1250
1251         switch (status) {
1252         case -EAGAIN:
1253                 if (test_bit(SOCK_ASYNC_NOSPACE, &xprt->sock->flags)) {
1254                         /* Protect against races with xprt_write_space */
1255                         spin_lock_bh(&xprt->sock_lock);
1256                         /* Don't race with disconnect */
1257                         if (!xprt_connected(xprt))
1258                                 task->tk_status = -ENOTCONN;
1259                         else if (test_bit(SOCK_NOSPACE, &xprt->sock->flags)) {
1260                                 task->tk_timeout = req->rq_timeout;
1261                                 rpc_sleep_on(&xprt->pending, task, NULL, NULL);
1262                         }
1263                         spin_unlock_bh(&xprt->sock_lock);
1264                         return;
1265                 }
1266                 /* Keep holding the socket if it is blocked */
1267                 rpc_delay(task, HZ>>4);
1268                 return;
1269         case -ECONNREFUSED:
1270                 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
1271                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
1272         case -ENOTCONN:
1273                 return;
1274         default:
1275                 if (xprt->stream)
1276                         xprt_disconnect(xprt);
1277         }
1278         xprt_release_write(xprt, task);
1279         return;
1280  out_receive:
1281         dprintk("RPC: %4d xmit complete\n", task->tk_pid);
1282         /* Set the task's receive timeout value */
1283         spin_lock_bh(&xprt->sock_lock);
1284         if (!xprt->nocong) {
1285                 int timer = task->tk_msg.rpc_proc->p_timer;
1286                 task->tk_timeout = rpc_calc_rto(clnt->cl_rtt, timer);
1287                 task->tk_timeout <<= rpc_ntimeo(clnt->cl_rtt, timer) + req->rq_retries;
1288                 if (task->tk_timeout > xprt->timeout.to_maxval || task->tk_timeout == 0)
1289                         task->tk_timeout = xprt->timeout.to_maxval;
1290         } else
1291                 task->tk_timeout = req->rq_timeout;
1292         /* Don't race with disconnect */
1293         if (!xprt_connected(xprt))
1294                 task->tk_status = -ENOTCONN;
1295         else if (!req->rq_received)
1296                 rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
1297         __xprt_release_write(xprt, task);
1298         spin_unlock_bh(&xprt->sock_lock);
1299 }
1300
1301 /*
1302  * Reserve an RPC call slot.
1303  */
1304 static inline void
1305 do_xprt_reserve(struct rpc_task *task)
1306 {
1307         struct rpc_xprt *xprt = task->tk_xprt;
1308
1309         task->tk_status = 0;
1310         if (task->tk_rqstp)
1311                 return;
1312         if (!list_empty(&xprt->free)) {
1313                 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1314                 list_del_init(&req->rq_list);
1315                 task->tk_rqstp = req;
1316                 xprt_request_init(task, xprt);
1317                 return;
1318         }
1319         dprintk("RPC:      waiting for request slot\n");
1320         task->tk_status = -EAGAIN;
1321         task->tk_timeout = 0;
1322         rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
1323 }
1324
1325 void
1326 xprt_reserve(struct rpc_task *task)
1327 {
1328         struct rpc_xprt *xprt = task->tk_xprt;
1329
1330         task->tk_status = -EIO;
1331         if (!xprt->shutdown) {
1332                 spin_lock(&xprt->xprt_lock);
1333                 do_xprt_reserve(task);
1334                 spin_unlock(&xprt->xprt_lock);
1335                 if (task->tk_rqstp)
1336                         del_timer_sync(&xprt->timer);
1337         }
1338 }
1339
1340 /*
1341  * Allocate a 'unique' XID
1342  */
1343 static inline u32 xprt_alloc_xid(struct rpc_xprt *xprt)
1344 {
1345         return xprt->xid++;
1346 }
1347
1348 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1349 {
1350         get_random_bytes(&xprt->xid, sizeof(xprt->xid));
1351 }
1352
1353 /*
1354  * Initialize RPC request
1355  */
1356 static void
1357 xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1358 {
1359         struct rpc_rqst *req = task->tk_rqstp;
1360
1361         req->rq_timeout = xprt->timeout.to_initval;
1362         req->rq_task    = task;
1363         req->rq_xprt    = xprt;
1364         req->rq_xid     = xprt_alloc_xid(xprt);
1365         dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
1366                         req, ntohl(req->rq_xid));
1367 }
1368
1369 /*
1370  * Release an RPC call slot
1371  */
1372 void
1373 xprt_release(struct rpc_task *task)
1374 {
1375         struct rpc_xprt *xprt = task->tk_xprt;
1376         struct rpc_rqst *req;
1377
1378         if (!(req = task->tk_rqstp))
1379                 return;
1380         spin_lock_bh(&xprt->sock_lock);
1381         __xprt_release_write(xprt, task);
1382         __xprt_put_cong(xprt, req);
1383         if (!list_empty(&req->rq_list))
1384                 list_del(&req->rq_list);
1385         xprt->last_used = jiffies;
1386         if (list_empty(&xprt->recv) && !xprt->shutdown)
1387                 mod_timer(&xprt->timer, xprt->last_used + XPRT_IDLE_TIMEOUT);
1388         spin_unlock_bh(&xprt->sock_lock);
1389         task->tk_rqstp = NULL;
1390         memset(req, 0, sizeof(*req));   /* mark unused */
1391
1392         dprintk("RPC: %4d release request %p\n", task->tk_pid, req);
1393
1394         spin_lock(&xprt->xprt_lock);
1395         list_add(&req->rq_list, &xprt->free);
1396         xprt_clear_backlog(xprt);
1397         spin_unlock(&xprt->xprt_lock);
1398 }
1399
1400 /*
1401  * Set default timeout parameters
1402  */
1403 static void
1404 xprt_default_timeout(struct rpc_timeout *to, int proto)
1405 {
1406         if (proto == IPPROTO_UDP)
1407                 xprt_set_timeout(to, 5,  5 * HZ);
1408         else
1409                 xprt_set_timeout(to, 5, 60 * HZ);
1410 }
1411
1412 /*
1413  * Set constant timeout
1414  */
1415 void
1416 xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
1417 {
1418         to->to_initval   =
1419         to->to_increment = incr;
1420         to->to_maxval    = incr * retr;
1421         to->to_retries   = retr;
1422         to->to_exponential = 0;
1423 }
1424
1425 unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1426 unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1427
1428 /*
1429  * Initialize an RPC client
1430  */
1431 static struct rpc_xprt *
1432 xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
1433 {
1434         struct rpc_xprt *xprt;
1435         unsigned int entries;
1436         size_t slot_table_size;
1437         struct rpc_rqst *req;
1438
1439         dprintk("RPC:      setting up %s transport...\n",
1440                                 proto == IPPROTO_UDP? "UDP" : "TCP");
1441
1442         entries = (proto == IPPROTO_TCP)?
1443                 xprt_tcp_slot_table_entries : xprt_udp_slot_table_entries;
1444
1445         if ((xprt = kmalloc(sizeof(struct rpc_xprt), GFP_KERNEL)) == NULL)
1446                 return ERR_PTR(-ENOMEM);
1447         memset(xprt, 0, sizeof(*xprt)); /* Nnnngh! */
1448         xprt->max_reqs = entries;
1449         slot_table_size = entries * sizeof(xprt->slot[0]);
1450         xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
1451         if (xprt->slot == NULL) {
1452                 kfree(xprt);
1453                 return ERR_PTR(-ENOMEM);
1454         }
1455         memset(xprt->slot, 0, slot_table_size);
1456
1457         xprt->addr = *ap;
1458         xprt->prot = proto;
1459         xprt->stream = (proto == IPPROTO_TCP)? 1 : 0;
1460         if (xprt->stream) {
1461                 xprt->cwnd = RPC_MAXCWND(xprt);
1462                 xprt->nocong = 1;
1463         } else
1464                 xprt->cwnd = RPC_INITCWND;
1465         spin_lock_init(&xprt->sock_lock);
1466         spin_lock_init(&xprt->xprt_lock);
1467         init_waitqueue_head(&xprt->cong_wait);
1468
1469         INIT_LIST_HEAD(&xprt->free);
1470         INIT_LIST_HEAD(&xprt->recv);
1471         INIT_WORK(&xprt->sock_connect, xprt_socket_connect, xprt);
1472         INIT_WORK(&xprt->task_cleanup, xprt_socket_autoclose, xprt);
1473         init_timer(&xprt->timer);
1474         xprt->timer.function = xprt_init_autodisconnect;
1475         xprt->timer.data = (unsigned long) xprt;
1476         xprt->last_used = jiffies;
1477         xprt->port = XPRT_MAX_RESVPORT;
1478
1479         /* Set timeout parameters */
1480         if (to) {
1481                 xprt->timeout = *to;
1482         } else
1483                 xprt_default_timeout(&xprt->timeout, xprt->prot);
1484
1485         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1486         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
1487         rpc_init_wait_queue(&xprt->resend, "xprt_resend");
1488         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1489
1490         /* initialize free list */
1491         for (req = &xprt->slot[entries-1]; req >= &xprt->slot[0]; req--)
1492                 list_add(&req->rq_list, &xprt->free);
1493
1494         xprt_init_xid(xprt);
1495
1496         /* Check whether we want to use a reserved port */
1497         xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1498
1499         dprintk("RPC:      created transport %p with %u slots\n", xprt,
1500                         xprt->max_reqs);
1501
1502         return xprt;
1503 }
1504
1505 /*
1506  * Bind to a reserved port
1507  */
1508 static inline int xprt_bindresvport(struct rpc_xprt *xprt, struct socket *sock)
1509 {
1510         struct sockaddr_in myaddr = {
1511                 .sin_family = AF_INET,
1512         };
1513         int             err, port;
1514
1515         /* Were we already bound to a given port? Try to reuse it */
1516         port = xprt->port;
1517         do {
1518                 myaddr.sin_port = htons(port);
1519                 err = sock->ops->bind(sock, (struct sockaddr *) &myaddr,
1520                                                 sizeof(myaddr));
1521                 if (err == 0) {
1522                         xprt->port = port;
1523                         return 0;
1524                 }
1525                 if (--port == 0)
1526                         port = XPRT_MAX_RESVPORT;
1527         } while (err == -EADDRINUSE && port != xprt->port);
1528
1529         printk("RPC: Can't bind to reserved port (%d).\n", -err);
1530         return err;
1531 }
1532
1533 static void
1534 xprt_bind_socket(struct rpc_xprt *xprt, struct socket *sock)
1535 {
1536         struct sock     *sk = sock->sk;
1537
1538         if (xprt->inet)
1539                 return;
1540
1541         write_lock_bh(&sk->sk_callback_lock);
1542         sk->sk_user_data = xprt;
1543         xprt->old_data_ready = sk->sk_data_ready;
1544         xprt->old_state_change = sk->sk_state_change;
1545         xprt->old_write_space = sk->sk_write_space;
1546         if (xprt->prot == IPPROTO_UDP) {
1547                 sk->sk_data_ready = udp_data_ready;
1548                 sk->sk_no_check = UDP_CSUM_NORCV;
1549                 xprt_set_connected(xprt);
1550         } else {
1551                 tcp_sk(sk)->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 * xprt_create_socket(struct rpc_xprt *xprt, int proto, int resvport)
1592 {
1593         struct socket   *sock;
1594         int             type, err;
1595
1596         dprintk("RPC:      xprt_create_socket(%s %d)\n",
1597                            (proto == IPPROTO_UDP)? "udp" : "tcp", proto);
1598
1599         type = (proto == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1600
1601         if ((err = sock_create_kern(PF_INET, type, proto, &sock)) < 0) {
1602                 printk("RPC: can't create socket (%d).\n", -err);
1603                 return NULL;
1604         }
1605
1606         /* If the caller has the capability, bind to a reserved port */
1607         if (resvport && xprt_bindresvport(xprt, sock) < 0) {
1608                 printk("RPC: can't bind to reserved port.\n");
1609                 goto failed;
1610         }
1611
1612         return sock;
1613
1614 failed:
1615         sock_release(sock);
1616         return NULL;
1617 }
1618
1619 /*
1620  * Create an RPC client transport given the protocol and peer address.
1621  */
1622 struct rpc_xprt *
1623 xprt_create_proto(int proto, struct sockaddr_in *sap, struct rpc_timeout *to)
1624 {
1625         struct rpc_xprt *xprt;
1626
1627         xprt = xprt_setup(proto, sap, to);
1628         if (IS_ERR(xprt))
1629                 dprintk("RPC:      xprt_create_proto failed\n");
1630         else
1631                 dprintk("RPC:      xprt_create_proto created xprt %p\n", xprt);
1632         return xprt;
1633 }
1634
1635 /*
1636  * Prepare for transport shutdown.
1637  */
1638 static void
1639 xprt_shutdown(struct rpc_xprt *xprt)
1640 {
1641         xprt->shutdown = 1;
1642         rpc_wake_up(&xprt->sending);
1643         rpc_wake_up(&xprt->resend);
1644         rpc_wake_up(&xprt->pending);
1645         rpc_wake_up(&xprt->backlog);
1646         wake_up(&xprt->cong_wait);
1647         del_timer_sync(&xprt->timer);
1648 }
1649
1650 /*
1651  * Clear the xprt backlog queue
1652  */
1653 static int
1654 xprt_clear_backlog(struct rpc_xprt *xprt) {
1655         rpc_wake_up_next(&xprt->backlog);
1656         wake_up(&xprt->cong_wait);
1657         return 1;
1658 }
1659
1660 /*
1661  * Destroy an RPC transport, killing off all requests.
1662  */
1663 int
1664 xprt_destroy(struct rpc_xprt *xprt)
1665 {
1666         dprintk("RPC:      destroying transport %p\n", xprt);
1667         xprt_shutdown(xprt);
1668         xprt_disconnect(xprt);
1669         xprt_close(xprt);
1670         kfree(xprt->slot);
1671         kfree(xprt);
1672
1673         return 0;
1674 }