err.no Git - linux-2.6/blob - 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                 goto out_sleep;
 149         }
 150         if (xprt->nocong || __xprt_get_cong(xprt, task)) {
 151                 xprt->snd_task = task;
 152                 if (req) {
 153                         req->rq_bytes_sent = 0;
 154                         req->rq_ntrans++;
 155                 }
 156                 return 1;
 157         }
 158         smp_mb__before_clear_bit();
 159         clear_bit(XPRT_LOCKED, &xprt->sockstate);
 160         smp_mb__after_clear_bit();
 161 out_sleep:
 162         dprintk("RPC: %4d failed to lock socket %p\n", task->tk_pid, xprt);
 163         task->tk_timeout = 0;
 164         task->tk_status = -EAGAIN;
 165         if (req && req->rq_ntrans)
 166                 rpc_sleep_on(&xprt->resend, task, NULL, NULL);
 167         else
 168                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
 169         return 0;
 170 }
 171
 172 static inline int
 173 xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 174 {
 175         int retval;
 176
 177         spin_lock_bh(&xprt->sock_lock);
 178         retval = __xprt_lock_write(xprt, task);
 179         spin_unlock_bh(&xprt->sock_lock);
 180         return retval;
 181 }
 182
 183
 184 static void
 185 __xprt_lock_write_next(struct rpc_xprt *xprt)
 186 {
 187         struct rpc_task *task;
 188
 189         if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
 190                 return;
 191         if (!xprt->nocong && RPCXPRT_CONGESTED(xprt))
 192                 goto out_unlock;
 193         task = rpc_wake_up_next(&xprt->resend);
 194         if (!task) {
 195                 task = rpc_wake_up_next(&xprt->sending);
 196                 if (!task)
 197                         goto out_unlock;
 198         }
 199         if (xprt->nocong || __xprt_get_cong(xprt, task)) {
 200                 struct rpc_rqst *req = task->tk_rqstp;
 201                 xprt->snd_task = task;
 202                 if (req) {
 203                         req->rq_bytes_sent = 0;
 204                         req->rq_ntrans++;
 205                 }
 206                 return;
 207         }
 208 out_unlock:
 209         smp_mb__before_clear_bit();
 210         clear_bit(XPRT_LOCKED, &xprt->sockstate);
 211         smp_mb__after_clear_bit();
 212 }
 213
 214 /*
 215  * Releases the socket for use by other requests.
 216  */
 217 static void
 218 __xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 219 {
 220         if (xprt->snd_task == task) {
 221                 xprt->snd_task = NULL;
 222                 smp_mb__before_clear_bit();
 223                 clear_bit(XPRT_LOCKED, &xprt->sockstate);
 224                 smp_mb__after_clear_bit();
 225                 __xprt_lock_write_next(xprt);
 226         }
 227 }
 228
 229 static inline void
 230 xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 231 {
 232         spin_lock_bh(&xprt->sock_lock);
 233         __xprt_release_write(xprt, task);
 234         spin_unlock_bh(&xprt->sock_lock);
 235 }
 236
 237 /*
 238  * Write data to socket.
 239  */
 240 static inline int
 241 xprt_sendmsg(struct rpc_xprt *xprt, struct rpc_rqst *req)
 242 {
 243         struct socket   *sock = xprt->sock;
 244         struct xdr_buf  *xdr = &req->rq_snd_buf;
 245         struct sockaddr *addr = NULL;
 246         int addrlen = 0;
 247         unsigned int    skip;
 248         int             result;
 249
 250         if (!sock)
 251                 return -ENOTCONN;
 252
 253         xprt_pktdump("packet data:",
 254                                 req->rq_svec->iov_base,
 255                                 req->rq_svec->iov_len);
 256
 257         /* For UDP, we need to provide an address */
 258         if (!xprt->stream) {
 259                 addr = (struct sockaddr *) &xprt->addr;
 260                 addrlen = sizeof(xprt->addr);
 261         }
 262         /* Dont repeat bytes */
 263         skip = req->rq_bytes_sent;
 264
 265         clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 266         result = xdr_sendpages(sock, addr, addrlen, xdr, skip, MSG_DONTWAIT);
 267
 268         dprintk("RPC:      xprt_sendmsg(%d) = %d\n", xdr->len - skip, result);
 269
 270         if (result >= 0)
 271                 return result;
 272
 273         switch (result) {
 274         case -ECONNREFUSED:
 275                 /* When the server has died, an ICMP port unreachable message
 276                  * prompts ECONNREFUSED.
 277                  */
 278         case -EAGAIN:
 279                 break;
 280         case -ECONNRESET:
 281         case -ENOTCONN:
 282         case -EPIPE:
 283                 /* connection broken */
 284                 if (xprt->stream)
 285                         result = -ENOTCONN;
 286                 break;
 287         default:
 288                 printk(KERN_NOTICE "RPC: sendmsg returned error %d\n", -result);
 289         }
 290         return result;
 291 }
 292
 293 /*
 294  * Van Jacobson congestion avoidance. Check if the congestion window
 295  * overflowed. Put the task to sleep if this is the case.
 296  */
 297 static int
 298 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 299 {
 300         struct rpc_rqst *req = task->tk_rqstp;
 301
 302         if (req->rq_cong)
 303                 return 1;
 304         dprintk("RPC: %4d xprt_cwnd_limited cong = %ld cwnd = %ld\n",
 305                         task->tk_pid, xprt->cong, xprt->cwnd);
 306         if (RPCXPRT_CONGESTED(xprt))
 307                 return 0;
 308         req->rq_cong = 1;
 309         xprt->cong += RPC_CWNDSCALE;
 310         return 1;
 311 }
 312
 313 /*
 314  * Adjust the congestion window, and wake up the next task
 315  * that has been sleeping due to congestion
 316  */
 317 static void
 318 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 319 {
 320         if (!req->rq_cong)
 321                 return;
 322         req->rq_cong = 0;
 323         xprt->cong -= RPC_CWNDSCALE;
 324         __xprt_lock_write_next(xprt);
 325 }
 326
 327 /*
 328  * Adjust RPC congestion window
 329  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
 330  */
 331 static void
 332 xprt_adjust_cwnd(struct rpc_xprt *xprt, int result)
 333 {
 334         unsigned long   cwnd;
 335
 336         cwnd = xprt->cwnd;
 337         if (result >= 0 && cwnd <= xprt->cong) {
 338                 /* The (cwnd >> 1) term makes sure
 339                  * the result gets rounded properly. */
 340                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 341                 if (cwnd > RPC_MAXCWND(xprt))
 342                         cwnd = RPC_MAXCWND(xprt);
 343                 __xprt_lock_write_next(xprt);
 344         } else if (result == -ETIMEDOUT) {
 345                 cwnd >>= 1;
 346                 if (cwnd < RPC_CWNDSCALE)
 347                         cwnd = RPC_CWNDSCALE;
 348         }
 349         dprintk("RPC:      cong %ld, cwnd was %ld, now %ld\n",
 350                         xprt->cong, xprt->cwnd, cwnd);
 351         xprt->cwnd = cwnd;
 352 }
 353
 354 /*
 355  * Reset the major timeout value
 356  */
 357 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 358 {
 359         struct rpc_timeout *to = &req->rq_xprt->timeout;
 360
 361         req->rq_majortimeo = req->rq_timeout;
 362         if (to->to_exponential)
 363                 req->rq_majortimeo <<= to->to_retries;
 364         else
 365                 req->rq_majortimeo += to->to_increment * to->to_retries;
 366         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 367                 req->rq_majortimeo = to->to_maxval;
 368         req->rq_majortimeo += jiffies;
 369 }
 370
 371 /*
 372  * Adjust timeout values etc for next retransmit
 373  */
 374 int xprt_adjust_timeout(struct rpc_rqst *req)
 375 {
 376         struct rpc_xprt *xprt = req->rq_xprt;
 377         struct rpc_timeout *to = &xprt->timeout;
 378         int status = 0;
 379
 380         if (time_before(jiffies, req->rq_majortimeo)) {
 381                 if (to->to_exponential)
 382                         req->rq_timeout <<= 1;
 383                 else
 384                         req->rq_timeout += to->to_increment;
 385                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 386                         req->rq_timeout = to->to_maxval;
 387                 req->rq_retries++;
 388                 pprintk("RPC: %lu retrans\n", jiffies);
 389         } else {
 390                 req->rq_timeout = to->to_initval;
 391                 req->rq_retries = 0;
 392                 xprt_reset_majortimeo(req);
 393                 /* Reset the RTT counters == "slow start" */
 394                 spin_lock_bh(&xprt->sock_lock);
 395                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 396                 spin_unlock_bh(&xprt->sock_lock);
 397                 pprintk("RPC: %lu timeout\n", jiffies);
 398                 status = -ETIMEDOUT;
 399         }
 400
 401         if (req->rq_timeout == 0) {
 402                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 403                 req->rq_timeout = 5 * HZ;
 404         }
 405         return status;
 406 }
 407
 408 /*
 409  * Close down a transport socket
 410  */
 411 static void
 412 xprt_close(struct rpc_xprt *xprt)
 413 {
 414         struct socket   *sock = xprt->sock;
 415         struct sock     *sk = xprt->inet;
 416
 417         if (!sk)
 418                 return;
 419
 420         write_lock_bh(&sk->sk_callback_lock);
 421         xprt->inet = NULL;
 422         xprt->sock = NULL;
 423
 424         sk->sk_user_data    = NULL;
 425         sk->sk_data_ready   = xprt->old_data_ready;
 426         sk->sk_state_change = xprt->old_state_change;
 427         sk->sk_write_space  = xprt->old_write_space;
 428         write_unlock_bh(&sk->sk_callback_lock);
 429
 430         sk->sk_no_check  = 0;
 431
 432         sock_release(sock);
 433 }
 434
 435 static void
 436 xprt_socket_autoclose(void *args)
 437 {
 438         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
 439
 440         xprt_disconnect(xprt);
 441         xprt_close(xprt);
 442         xprt_release_write(xprt, NULL);
 443 }
 444
 445 /*
 446  * Mark a transport as disconnected
 447  */
 448 static void
 449 xprt_disconnect(struct rpc_xprt *xprt)
 450 {
 451         dprintk("RPC:      disconnected transport %p\n", xprt);
 452         spin_lock_bh(&xprt->sock_lock);
 453         xprt_clear_connected(xprt);
 454         rpc_wake_up_status(&xprt->pending, -ENOTCONN);
 455         spin_unlock_bh(&xprt->sock_lock);
 456 }
 457
 458 /*
 459  * Used to allow disconnection when we've been idle
 460  */
 461 static void
 462 xprt_init_autodisconnect(unsigned long data)
 463 {
 464         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 465
 466         spin_lock(&xprt->sock_lock);
 467         if (!list_empty(&xprt->recv) || xprt->shutdown)
 468                 goto out_abort;
 469         if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
 470                 goto out_abort;
 471         spin_unlock(&xprt->sock_lock);
 472         /* Let keventd close the socket */
 473         if (test_bit(XPRT_CONNECTING, &xprt->sockstate) != 0)
 474                 xprt_release_write(xprt, NULL);
 475         else
 476                 schedule_work(&xprt->task_cleanup);
 477         return;
 478 out_abort:
 479         spin_unlock(&xprt->sock_lock);
 480 }
 481
 482 static void xprt_socket_connect(void *args)
 483 {
 484         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
 485         struct socket *sock = xprt->sock;
 486         int status = -EIO;
 487
 488         if (xprt->shutdown || xprt->addr.sin_port == 0)
 489                 goto out;
 490
 491         /*
 492          * Start by resetting any existing state
 493          */
 494         xprt_close(xprt);
 495         sock = xprt_create_socket(xprt, xprt->prot, xprt->resvport);
 496         if (sock == NULL) {
 497                 /* couldn't create socket or bind to reserved port;
 498                  * this is likely a permanent error, so cause an abort */
 499                 goto out;
 500         }
 501         xprt_bind_socket(xprt, sock);
 502         xprt_sock_setbufsize(xprt);
 503
 504         status = 0;
 505         if (!xprt->stream)
 506                 goto out;
 507
 508         /*
 509          * Tell the socket layer to start connecting...
 510          */
 511         status = sock->ops->connect(sock, (struct sockaddr *) &xprt->addr,
 512                         sizeof(xprt->addr), O_NONBLOCK);
 513         dprintk("RPC: %p  connect status %d connected %d sock state %d\n",
 514                         xprt, -status, xprt_connected(xprt), sock->sk->sk_state);
 515         if (status < 0) {
 516                 switch (status) {
 517                         case -EINPROGRESS:
 518                         case -EALREADY:
 519                                 goto out_clear;
 520                 }
 521         }
 522 out:
 523         if (status < 0)
 524                 rpc_wake_up_status(&xprt->pending, status);
 525         else
 526                 rpc_wake_up(&xprt->pending);
 527 out_clear:
 528         smp_mb__before_clear_bit();
 529         clear_bit(XPRT_CONNECTING, &xprt->sockstate);
 530         smp_mb__after_clear_bit();
 531 }
 532
 533 /*
 534  * Attempt to connect a TCP socket.
 535  *
 536  */
 537 void xprt_connect(struct rpc_task *task)
 538 {
 539         struct rpc_xprt *xprt = task->tk_xprt;
 540
 541         dprintk("RPC: %4d xprt_connect xprt %p %s connected\n", task->tk_pid,
 542                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 543
 544         if (xprt->shutdown) {
 545                 task->tk_status = -EIO;
 546                 return;
 547         }
 548         if (!xprt->addr.sin_port) {
 549                 task->tk_status = -EIO;
 550                 return;
 551         }
 552         if (!xprt_lock_write(xprt, task))
 553                 return;
 554         if (xprt_connected(xprt))
 555                 goto out_write;
 556
 557         if (task->tk_rqstp)
 558                 task->tk_rqstp->rq_bytes_sent = 0;
 559
 560         task->tk_timeout = RPC_CONNECT_TIMEOUT;
 561         rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
 562         if (!test_and_set_bit(XPRT_CONNECTING, &xprt->sockstate)) {
 563                 /* Note: if we are here due to a dropped connection
 564                  *       we delay reconnecting by RPC_REESTABLISH_TIMEOUT/HZ
 565                  *       seconds
 566                  */
 567                 if (xprt->sock != NULL)
 568                         schedule_delayed_work(&xprt->sock_connect,
 569                                         RPC_REESTABLISH_TIMEOUT);
 570                 else {
 571                         schedule_work(&xprt->sock_connect);
 572                         if (!RPC_IS_ASYNC(task))
 573                                 flush_scheduled_work();
 574                 }
 575         }
 576         return;
 577  out_write:
 578         xprt_release_write(xprt, task);
 579 }
 580
 581 /*
 582  * We arrive here when awoken from waiting on connection establishment.
 583  */
 584 static void
 585 xprt_connect_status(struct rpc_task *task)
 586 {
 587         struct rpc_xprt *xprt = task->tk_xprt;
 588
 589         if (task->tk_status >= 0) {
 590                 dprintk("RPC: %4d xprt_connect_status: connection established\n",
 591                                 task->tk_pid);
 592                 return;
 593         }
 594
 595         switch (task->tk_status) {
 596         case -ECONNREFUSED:
 597         case -ECONNRESET:
 598                 dprintk("RPC: %4d xprt_connect_status: server %s refused connection\n",
 599                                 task->tk_pid, task->tk_client->cl_server);
 600                 break;
 601         case -ENOTCONN:
 602                 dprintk("RPC: %4d xprt_connect_status: connection broken\n",
 603                                 task->tk_pid);
 604                 break;
 605         case -ETIMEDOUT:
 606                 dprintk("RPC: %4d xprt_connect_status: connect attempt timed out\n",
 607                                 task->tk_pid);
 608                 break;
 609         default:
 610                 dprintk("RPC: %4d xprt_connect_status: error %d connecting to server %s\n",
 611                                 task->tk_pid, -task->tk_status, task->tk_client->cl_server);
 612                 xprt_release_write(xprt, task);
 613                 task->tk_status = -EIO;
 614                 return;
 615         }
 616
 617         /* if soft mounted, just cause this RPC to fail */
 618         if (RPC_IS_SOFT(task)) {
 619                 xprt_release_write(xprt, task);
 620                 task->tk_status = -EIO;
 621         }
 622 }
 623
 624 /*
 625  * Look up the RPC request corresponding to a reply, and then lock it.
 626  */
 627 static inline struct rpc_rqst *
 628 xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
 629 {
 630         struct list_head *pos;
 631         struct rpc_rqst *req = NULL;
 632
 633         list_for_each(pos, &xprt->recv) {
 634                 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
 635                 if (entry->rq_xid == xid) {
 636                         req = entry;
 637                         break;
 638                 }
 639         }
 640         return req;
 641 }
 642
 643 /*
 644  * Complete reply received.
 645  * The TCP code relies on us to remove the request from xprt->pending.
 646  */
 647 static void
 648 xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
 649 {
 650         struct rpc_task *task = req->rq_task;
 651         struct rpc_clnt *clnt = task->tk_client;
 652
 653         /* Adjust congestion window */
 654         if (!xprt->nocong) {
 655                 unsigned timer = task->tk_msg.rpc_proc->p_timer;
 656                 xprt_adjust_cwnd(xprt, copied);
 657                 __xprt_put_cong(xprt, req);
 658                 if (timer) {
 659                         if (req->rq_ntrans == 1)
 660                                 rpc_update_rtt(clnt->cl_rtt, timer,
 661                                                 (long)jiffies - req->rq_xtime);
 662                         rpc_set_timeo(clnt->cl_rtt, timer, req->rq_ntrans - 1);
 663                 }
 664         }
 665
 666 #ifdef RPC_PROFILE
 667         /* Profile only reads for now */
 668         if (copied > 1024) {
 669                 static unsigned long    nextstat;
 670                 static unsigned long    pkt_rtt, pkt_len, pkt_cnt;
 671
 672                 pkt_cnt++;
 673                 pkt_len += req->rq_slen + copied;
 674                 pkt_rtt += jiffies - req->rq_xtime;
 675                 if (time_before(nextstat, jiffies)) {
 676                         printk("RPC: %lu %ld cwnd\n", jiffies, xprt->cwnd);
 677                         printk("RPC: %ld %ld %ld %ld stat\n",
 678                                         jiffies, pkt_cnt, pkt_len, pkt_rtt);
 679                         pkt_rtt = pkt_len = pkt_cnt = 0;
 680                         nextstat = jiffies + 5 * HZ;
 681                 }
 682         }
 683 #endif
 684
 685         dprintk("RPC: %4d has input (%d bytes)\n", task->tk_pid, copied);
 686         list_del_init(&req->rq_list);
 687         req->rq_received = req->rq_private_buf.len = copied;
 688
 689         /* ... and wake up the process. */
 690         rpc_wake_up_task(task);
 691         return;
 692 }
 693
 694 static size_t
 695 skb_read_bits(skb_reader_t *desc, void *to, size_t len)
 696 {
 697         if (len > desc->count)
 698                 len = desc->count;
 699         if (skb_copy_bits(desc->skb, desc->offset, to, len))
 700                 return 0;
 701         desc->count -= len;
 702         desc->offset += len;
 703         return len;
 704 }
 705
 706 static size_t
 707 skb_read_and_csum_bits(skb_reader_t *desc, void *to, size_t len)
 708 {
 709         unsigned int csum2, pos;
 710
 711         if (len > desc->count)
 712                 len = desc->count;
 713         pos = desc->offset;
 714         csum2 = skb_copy_and_csum_bits(desc->skb, pos, to, len, 0);
 715         desc->csum = csum_block_add(desc->csum, csum2, pos);
 716         desc->count -= len;
 717         desc->offset += len;
 718         return len;
 719 }
 720
 721 /*
 722  * We have set things up such that we perform the checksum of the UDP
 723  * packet in parallel with the copies into the RPC client iovec.  -DaveM
 724  */
 725 int
 726 csum_partial_copy_to_xdr(struct xdr_buf *xdr, struct sk_buff *skb)
 727 {
 728         skb_reader_t desc;
 729
 730         desc.skb = skb;
 731         desc.offset = sizeof(struct udphdr);
 732         desc.count = skb->len - desc.offset;
 733
 734         if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 735                 goto no_checksum;
 736
 737         desc.csum = csum_partial(skb->data, desc.offset, skb->csum);
 738         if (xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_and_csum_bits) < 0)
 739                 return -1;
 740         if (desc.offset != skb->len) {
 741                 unsigned int csum2;
 742                 csum2 = skb_checksum(skb, desc.offset, skb->len - desc.offset, 0);
 743                 desc.csum = csum_block_add(desc.csum, csum2, desc.offset);
 744         }
 745         if (desc.count)
 746                 return -1;
 747         if ((unsigned short)csum_fold(desc.csum))
 748                 return -1;
 749         return 0;
 750 no_checksum:
 751         if (xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_bits) < 0)
 752                 return -1;
 753         if (desc.count)
 754                 return -1;
 755         return 0;
 756 }
 757
 758 /*
 759  * Input handler for RPC replies. Called from a bottom half and hence
 760  * atomic.
 761  */
 762 static void
 763 udp_data_ready(struct sock *sk, int len)
 764 {
 765         struct rpc_task *task;
 766         struct rpc_xprt *xprt;
 767         struct rpc_rqst *rovr;
 768         struct sk_buff  *skb;
 769         int err, repsize, copied;
 770         u32 _xid, *xp;
 771
 772         read_lock(&sk->sk_callback_lock);
 773         dprintk("RPC:      udp_data_ready...\n");
 774         if (!(xprt = xprt_from_sock(sk))) {
 775                 printk("RPC:      udp_data_ready request not found!\n");
 776                 goto out;
 777         }
 778
 779         dprintk("RPC:      udp_data_ready client %p\n", xprt);
 780
 781         if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
 782                 goto out;
 783
 784         if (xprt->shutdown)
 785                 goto dropit;
 786
 787         repsize = skb->len - sizeof(struct udphdr);
 788         if (repsize < 4) {
 789                 printk("RPC: impossible RPC reply size %d!\n", repsize);
 790                 goto dropit;
 791         }
 792
 793         /* Copy the XID from the skb... */
 794         xp = skb_header_pointer(skb, sizeof(struct udphdr),
 795                                 sizeof(_xid), &_xid);
 796         if (xp == NULL)
 797                 goto dropit;
 798
 799         /* Look up and lock the request corresponding to the given XID */
 800         spin_lock(&xprt->sock_lock);
 801         rovr = xprt_lookup_rqst(xprt, *xp);
 802         if (!rovr)
 803                 goto out_unlock;
 804         task = rovr->rq_task;
 805
 806         dprintk("RPC: %4d received reply\n", task->tk_pid);
 807
 808         if ((copied = rovr->rq_private_buf.buflen) > repsize)
 809                 copied = repsize;
 810
 811         /* Suck it into the iovec, verify checksum if not done by hw. */
 812         if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
 813                 goto out_unlock;
 814
 815         /* Something worked... */
 816         dst_confirm(skb->dst);
 817
 818         xprt_complete_rqst(xprt, rovr, copied);
 819
 820  out_unlock:
 821         spin_unlock(&xprt->sock_lock);
 822  dropit:
 823         skb_free_datagram(sk, skb);
 824  out:
 825         read_unlock(&sk->sk_callback_lock);
 826 }
 827
 828 /*
 829  * Copy from an skb into memory and shrink the skb.
 830  */
 831 static inline size_t
 832 tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
 833 {
 834         if (len > desc->count)
 835                 len = desc->count;
 836         if (skb_copy_bits(desc->skb, desc->offset, p, len)) {
 837                 dprintk("RPC:      failed to copy %zu bytes from skb. %zu bytes remain\n",
 838                                 len, desc->count);
 839                 return 0;
 840         }
 841         desc->offset += len;
 842         desc->count -= len;
 843         dprintk("RPC:      copied %zu bytes from skb. %zu bytes remain\n",
 844                         len, desc->count);
 845         return len;
 846 }
 847
 848 /*
 849  * TCP read fragment marker
 850  */
 851 static inline void
 852 tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
 853 {
 854         size_t len, used;
 855         char *p;
 856
 857         p = ((char *) &xprt->tcp_recm) + xprt->tcp_offset;
 858         len = sizeof(xprt->tcp_recm) - xprt->tcp_offset;
 859         used = tcp_copy_data(desc, p, len);
 860         xprt->tcp_offset += used;
 861         if (used != len)
 862                 return;
 863         xprt->tcp_reclen = ntohl(xprt->tcp_recm);
 864         if (xprt->tcp_reclen & 0x80000000)
 865                 xprt->tcp_flags |= XPRT_LAST_FRAG;
 866         else
 867                 xprt->tcp_flags &= ~XPRT_LAST_FRAG;
 868         xprt->tcp_reclen &= 0x7fffffff;
 869         xprt->tcp_flags &= ~XPRT_COPY_RECM;
 870         xprt->tcp_offset = 0;
 871         /* Sanity check of the record length */
 872         if (xprt->tcp_reclen < 4) {
 873                 printk(KERN_ERR "RPC: Invalid TCP record fragment length\n");
 874                 xprt_disconnect(xprt);
 875         }
 876         dprintk("RPC:      reading TCP record fragment of length %d\n",
 877                         xprt->tcp_reclen);
 878 }
 879
 880 static void
 881 tcp_check_recm(struct rpc_xprt *xprt)
 882 {
 883         dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u, tcp_flags = %lx\n",
 884                         xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen, xprt->tcp_flags);
 885         if (xprt->tcp_offset == xprt->tcp_reclen) {
 886                 xprt->tcp_flags |= XPRT_COPY_RECM;
 887                 xprt->tcp_offset = 0;
 888                 if (xprt->tcp_flags & XPRT_LAST_FRAG) {
 889                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
 890                         xprt->tcp_flags |= XPRT_COPY_XID;
 891                         xprt->tcp_copied = 0;
 892                 }
 893         }
 894 }
 895
 896 /*
 897  * TCP read xid
 898  */
 899 static inline void
 900 tcp_read_xid(struct rpc_xprt *xprt, skb_reader_t *desc)
 901 {
 902         size_t len, used;
 903         char *p;
 904
 905         len = sizeof(xprt->tcp_xid) - xprt->tcp_offset;
 906         dprintk("RPC:      reading XID (%Zu bytes)\n", len);
 907         p = ((char *) &xprt->tcp_xid) + xprt->tcp_offset;
 908         used = tcp_copy_data(desc, p, len);
 909         xprt->tcp_offset += used;
 910         if (used != len)
 911                 return;
 912         xprt->tcp_flags &= ~XPRT_COPY_XID;
 913         xprt->tcp_flags |= XPRT_COPY_DATA;
 914         xprt->tcp_copied = 4;
 915         dprintk("RPC:      reading reply for XID %08x\n",
 916                                                 ntohl(xprt->tcp_xid));
 917         tcp_check_recm(xprt);
 918 }
 919
 920 /*
 921  * TCP read and complete request
 922  */
 923 static inline void
 924 tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
 925 {
 926         struct rpc_rqst *req;
 927         struct xdr_buf *rcvbuf;
 928         size_t len;
 929         ssize_t r;
 930
 931         /* Find and lock the request corresponding to this xid */
 932         spin_lock(&xprt->sock_lock);
 933         req = xprt_lookup_rqst(xprt, xprt->tcp_xid);
 934         if (!req) {
 935                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 936                 dprintk("RPC:      XID %08x request not found!\n",
 937                                 ntohl(xprt->tcp_xid));
 938                 spin_unlock(&xprt->sock_lock);
 939                 return;
 940         }
 941
 942         rcvbuf = &req->rq_private_buf;
 943         len = desc->count;
 944         if (len > xprt->tcp_reclen - xprt->tcp_offset) {
 945                 skb_reader_t my_desc;
 946
 947                 len = xprt->tcp_reclen - xprt->tcp_offset;
 948                 memcpy(&my_desc, desc, sizeof(my_desc));
 949                 my_desc.count = len;
 950                 r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
 951                                           &my_desc, tcp_copy_data);
 952                 desc->count -= r;
 953                 desc->offset += r;
 954         } else
 955                 r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
 956                                           desc, tcp_copy_data);
 957
 958         if (r > 0) {
 959                 xprt->tcp_copied += r;
 960                 xprt->tcp_offset += r;
 961         }
 962         if (r != len) {
 963                 /* Error when copying to the receive buffer,
 964                  * usually because we weren't able to allocate
 965                  * additional buffer pages. All we can do now
 966                  * is turn off XPRT_COPY_DATA, so the request
 967                  * will not receive any additional updates,
 968                  * and time out.
 969                  * Any remaining data from this record will
 970                  * be discarded.
 971                  */
 972                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 973                 dprintk("RPC:      XID %08x truncated request\n",
 974                                 ntohl(xprt->tcp_xid));
 975                 dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
 976                                 xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
 977                 goto out;
 978         }
 979
 980         dprintk("RPC:      XID %08x read %Zd bytes\n",
 981                         ntohl(xprt->tcp_xid), r);
 982         dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
 983                         xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
 984
 985         if (xprt->tcp_copied == req->rq_private_buf.buflen)
 986                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 987         else if (xprt->tcp_offset == xprt->tcp_reclen) {
 988                 if (xprt->tcp_flags & XPRT_LAST_FRAG)
 989                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
 990         }
 991
 992 out:
 993         if (!(xprt->tcp_flags & XPRT_COPY_DATA)) {
 994                 dprintk("RPC: %4d received reply complete\n",
 995                                 req->rq_task->tk_pid);
 996                 xprt_complete_rqst(xprt, req, xprt->tcp_copied);
 997         }
 998         spin_unlock(&xprt->sock_lock);
 999         tcp_check_recm(xprt);
1000 }
1001
1002 /*
1003  * TCP discard extra bytes from a short read
1004  */
1005 static inline void
1006 tcp_read_discard(struct rpc_xprt *xprt, skb_reader_t *desc)
1007 {
1008         size_t len;
1009
1010         len = xprt->tcp_reclen - xprt->tcp_offset;
1011         if (len > desc->count)
1012                 len = desc->count;
1013         desc->count -= len;
1014         desc->offset += len;
1015         xprt->tcp_offset += len;
1016         dprintk("RPC:      discarded %Zu bytes\n", len);
1017         tcp_check_recm(xprt);
1018 }
1019
1020 /*
1021  * TCP record receive routine
1022  * We first have to grab the record marker, then the XID, then the data.
1023  */
1024 static int
1025 tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
1026                 unsigned int offset, size_t len)
1027 {
1028         struct rpc_xprt *xprt = rd_desc->arg.data;
1029         skb_reader_t desc = {
1030                 .skb    = skb,
1031                 .offset = offset,
1032                 .count  = len,
1033                 .csum   = 0
1034         };
1035
1036         dprintk("RPC:      tcp_data_recv\n");
1037         do {
1038                 /* Read in a new fragment marker if necessary */
1039                 /* Can we ever really expect to get completely empty fragments? */
1040                 if (xprt->tcp_flags & XPRT_COPY_RECM) {
1041                         tcp_read_fraghdr(xprt, &desc);
1042                         continue;
1043                 }
1044                 /* Read in the xid if necessary */
1045                 if (xprt->tcp_flags & XPRT_COPY_XID) {
1046                         tcp_read_xid(xprt, &desc);
1047                         continue;
1048                 }
1049                 /* Read in the request data */
1050                 if (xprt->tcp_flags & XPRT_COPY_DATA) {
1051                         tcp_read_request(xprt, &desc);
1052                         continue;
1053                 }
1054                 /* Skip over any trailing bytes on short reads */
1055                 tcp_read_discard(xprt, &desc);
1056         } while (desc.count);
1057         dprintk("RPC:      tcp_data_recv done\n");
1058         return len - desc.count;
1059 }
1060
1061 static void tcp_data_ready(struct sock *sk, int bytes)
1062 {
1063         struct rpc_xprt *xprt;
1064         read_descriptor_t rd_desc;
1065
1066         read_lock(&sk->sk_callback_lock);
1067         dprintk("RPC:      tcp_data_ready...\n");
1068         if (!(xprt = xprt_from_sock(sk))) {
1069                 printk("RPC:      tcp_data_ready socket info not found!\n");
1070                 goto out;
1071         }
1072         if (xprt->shutdown)
1073                 goto out;
1074
1075         /* We use rd_desc to pass struct xprt to tcp_data_recv */
1076         rd_desc.arg.data = xprt;
1077         rd_desc.count = 65536;
1078         tcp_read_sock(sk, &rd_desc, tcp_data_recv);
1079 out:
1080         read_unlock(&sk->sk_callback_lock);
1081 }
1082
1083 static void
1084 tcp_state_change(struct sock *sk)
1085 {
1086         struct rpc_xprt *xprt;
1087
1088         read_lock(&sk->sk_callback_lock);
1089         if (!(xprt = xprt_from_sock(sk)))
1090                 goto out;
1091         dprintk("RPC:      tcp_state_change client %p...\n", xprt);
1092         dprintk("RPC:      state %x conn %d dead %d zapped %d\n",
1093                                 sk->sk_state, xprt_connected(xprt),
1094                                 sock_flag(sk, SOCK_DEAD),
1095                                 sock_flag(sk, SOCK_ZAPPED));
1096
1097         switch (sk->sk_state) {
1098         case TCP_ESTABLISHED:
1099                 spin_lock_bh(&xprt->sock_lock);
1100                 if (!xprt_test_and_set_connected(xprt)) {
1101                         /* Reset TCP record info */
1102                         xprt->tcp_offset = 0;
1103                         xprt->tcp_reclen = 0;
1104                         xprt->tcp_copied = 0;
1105                         xprt->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
1106                         rpc_wake_up(&xprt->pending);
1107                 }
1108                 spin_unlock_bh(&xprt->sock_lock);
1109                 break;
1110         case TCP_SYN_SENT:
1111         case TCP_SYN_RECV:
1112                 break;
1113         default:
1114                 xprt_disconnect(xprt);
1115                 break;
1116         }
1117  out:
1118         read_unlock(&sk->sk_callback_lock);
1119 }
1120
1121 /*
1122  * Called when more output buffer space is available for this socket.
1123  * We try not to wake our writers until they can make "significant"
1124  * progress, otherwise we'll waste resources thrashing sock_sendmsg
1125  * with a bunch of small requests.
1126  */
1127 static void
1128 xprt_write_space(struct sock *sk)
1129 {
1130         struct rpc_xprt *xprt;
1131         struct socket   *sock;
1132
1133         read_lock(&sk->sk_callback_lock);
1134         if (!(xprt = xprt_from_sock(sk)) || !(sock = sk->sk_socket))
1135                 goto out;
1136         if (xprt->shutdown)
1137                 goto out;
1138
1139         /* Wait until we have enough socket memory */
1140         if (xprt->stream) {
1141                 /* from net/core/stream.c:sk_stream_write_space */
1142                 if (sk_stream_wspace(sk) < sk_stream_min_wspace(sk))
1143                         goto out;
1144         } else {
1145                 /* from net/core/sock.c:sock_def_write_space */
1146                 if (!sock_writeable(sk))
1147                         goto out;
1148         }
1149
1150         if (!test_and_clear_bit(SOCK_NOSPACE, &sock->flags))
1151                 goto out;
1152
1153         spin_lock_bh(&xprt->sock_lock);
1154         if (xprt->snd_task)
1155                 rpc_wake_up_task(xprt->snd_task);
1156         spin_unlock_bh(&xprt->sock_lock);
1157 out:
1158         read_unlock(&sk->sk_callback_lock);
1159 }
1160
1161 /*
1162  * RPC receive timeout handler.
1163  */
1164 static void
1165 xprt_timer(struct rpc_task *task)
1166 {
1167         struct rpc_rqst *req = task->tk_rqstp;
1168         struct rpc_xprt *xprt = req->rq_xprt;
1169
1170         spin_lock(&xprt->sock_lock);
1171         if (req->rq_received)
1172                 goto out;
1173
1174         xprt_adjust_cwnd(req->rq_xprt, -ETIMEDOUT);
1175         __xprt_put_cong(xprt, req);
1176
1177         dprintk("RPC: %4d xprt_timer (%s request)\n",
1178                 task->tk_pid, req ? "pending" : "backlogged");
1179
1180         task->tk_status  = -ETIMEDOUT;
1181 out:
1182         task->tk_timeout = 0;
1183         rpc_wake_up_task(task);
1184         spin_unlock(&xprt->sock_lock);
1185 }
1186
1187 /*
1188  * Place the actual RPC call.
1189  * We have to copy the iovec because sendmsg fiddles with its contents.
1190  */
1191 int
1192 xprt_prepare_transmit(struct rpc_task *task)
1193 {
1194         struct rpc_rqst *req = task->tk_rqstp;
1195         struct rpc_xprt *xprt = req->rq_xprt;
1196         int err = 0;
1197
1198         dprintk("RPC: %4d xprt_prepare_transmit\n", task->tk_pid);
1199
1200         if (xprt->shutdown)
1201                 return -EIO;
1202
1203         spin_lock_bh(&xprt->sock_lock);
1204         if (req->rq_received && !req->rq_bytes_sent) {
1205                 err = req->rq_received;
1206                 goto out_unlock;
1207         }
1208         if (!__xprt_lock_write(xprt, task)) {
1209                 err = -EAGAIN;
1210                 goto out_unlock;
1211         }
1212
1213         if (!xprt_connected(xprt)) {
1214                 err = -ENOTCONN;
1215                 goto out_unlock;
1216         }
1217 out_unlock:
1218         spin_unlock_bh(&xprt->sock_lock);
1219         return err;
1220 }
1221
1222 void
1223 xprt_transmit(struct rpc_task *task)
1224 {
1225         struct rpc_clnt *clnt = task->tk_client;
1226         struct rpc_rqst *req = task->tk_rqstp;
1227         struct rpc_xprt *xprt = req->rq_xprt;
1228         int status, retry = 0;
1229
1230
1231         dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
1232
1233         /* set up everything as needed. */
1234         /* Write the record marker */
1235         if (xprt->stream) {
1236                 u32     *marker = req->rq_svec[0].iov_base;
1237
1238                 *marker = htonl(0x80000000|(req->rq_slen-sizeof(*marker)));
1239         }
1240
1241         smp_rmb();
1242         if (!req->rq_received) {
1243                 if (list_empty(&req->rq_list)) {
1244                         spin_lock_bh(&xprt->sock_lock);
1245                         /* Update the softirq receive buffer */
1246                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
1247                                         sizeof(req->rq_private_buf));
1248                         /* Add request to the receive list */
1249                         list_add_tail(&req->rq_list, &xprt->recv);
1250                         spin_unlock_bh(&xprt->sock_lock);
1251                         xprt_reset_majortimeo(req);
1252                         /* Turn off autodisconnect */
1253                         del_singleshot_timer_sync(&xprt->timer);
1254                 }
1255         } else if (!req->rq_bytes_sent)
1256                 return;
1257
1258         /* Continue transmitting the packet/record. We must be careful
1259          * to cope with writespace callbacks arriving _after_ we have
1260          * called xprt_sendmsg().
1261          */
1262         while (1) {
1263                 req->rq_xtime = jiffies;
1264                 status = xprt_sendmsg(xprt, req);
1265
1266                 if (status < 0)
1267                         break;
1268
1269                 if (xprt->stream) {
1270                         req->rq_bytes_sent += status;
1271
1272                         /* If we've sent the entire packet, immediately
1273                          * reset the count of bytes sent. */
1274                         if (req->rq_bytes_sent >= req->rq_slen) {
1275                                 req->rq_bytes_sent = 0;
1276                                 goto out_receive;
1277                         }
1278                 } else {
1279                         if (status >= req->rq_slen)
1280                                 goto out_receive;
1281                         status = -EAGAIN;
1282                         break;
1283                 }
1284
1285                 dprintk("RPC: %4d xmit incomplete (%d left of %d)\n",
1286                                 task->tk_pid, req->rq_slen - req->rq_bytes_sent,
1287                                 req->rq_slen);
1288
1289                 status = -EAGAIN;
1290                 if (retry++ > 50)
1291                         break;
1292         }
1293
1294         /* Note: at this point, task->tk_sleeping has not yet been set,
1295          *       hence there is no danger of the waking up task being put on
1296          *       schedq, and being picked up by a parallel run of rpciod().
1297          */
1298         task->tk_status = status;
1299
1300         switch (status) {
1301         case -EAGAIN:
1302                 if (test_bit(SOCK_ASYNC_NOSPACE, &xprt->sock->flags)) {
1303                         /* Protect against races with xprt_write_space */
1304                         spin_lock_bh(&xprt->sock_lock);
1305                         /* Don't race with disconnect */
1306                         if (!xprt_connected(xprt))
1307                                 task->tk_status = -ENOTCONN;
1308                         else if (test_bit(SOCK_NOSPACE, &xprt->sock->flags)) {
1309                                 task->tk_timeout = req->rq_timeout;
1310                                 rpc_sleep_on(&xprt->pending, task, NULL, NULL);
1311                         }
1312                         spin_unlock_bh(&xprt->sock_lock);
1313                         return;
1314                 }
1315                 /* Keep holding the socket if it is blocked */
1316                 rpc_delay(task, HZ>>4);
1317                 return;
1318         case -ECONNREFUSED:
1319                 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
1320                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
1321         case -ENOTCONN:
1322                 return;
1323         default:
1324                 if (xprt->stream)
1325                         xprt_disconnect(xprt);
1326         }
1327         xprt_release_write(xprt, task);
1328         return;
1329  out_receive:
1330         dprintk("RPC: %4d xmit complete\n", task->tk_pid);
1331         /* Set the task's receive timeout value */
1332         spin_lock_bh(&xprt->sock_lock);
1333         if (!xprt->nocong) {
1334                 int timer = task->tk_msg.rpc_proc->p_timer;
1335                 task->tk_timeout = rpc_calc_rto(clnt->cl_rtt, timer);
1336                 task->tk_timeout <<= rpc_ntimeo(clnt->cl_rtt, timer) + req->rq_retries;
1337                 if (task->tk_timeout > xprt->timeout.to_maxval || task->tk_timeout == 0)
1338                         task->tk_timeout = xprt->timeout.to_maxval;
1339         } else
1340                 task->tk_timeout = req->rq_timeout;
1341         /* Don't race with disconnect */
1342         if (!xprt_connected(xprt))
1343                 task->tk_status = -ENOTCONN;
1344         else if (!req->rq_received)
1345                 rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
1346         __xprt_release_write(xprt, task);
1347         spin_unlock_bh(&xprt->sock_lock);
1348 }
1349
1350 /*
1351  * Reserve an RPC call slot.
1352  */
1353 static inline void
1354 do_xprt_reserve(struct rpc_task *task)
1355 {
1356         struct rpc_xprt *xprt = task->tk_xprt;
1357
1358         task->tk_status = 0;
1359         if (task->tk_rqstp)
1360                 return;
1361         if (!list_empty(&xprt->free)) {
1362                 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1363                 list_del_init(&req->rq_list);
1364                 task->tk_rqstp = req;
1365                 xprt_request_init(task, xprt);
1366                 return;
1367         }
1368         dprintk("RPC:      waiting for request slot\n");
1369         task->tk_status = -EAGAIN;
1370         task->tk_timeout = 0;
1371         rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
1372 }
1373
1374 void
1375 xprt_reserve(struct rpc_task *task)
1376 {
1377         struct rpc_xprt *xprt = task->tk_xprt;
1378
1379         task->tk_status = -EIO;
1380         if (!xprt->shutdown) {
1381                 spin_lock(&xprt->xprt_lock);
1382                 do_xprt_reserve(task);
1383                 spin_unlock(&xprt->xprt_lock);
1384         }
1385 }
1386
1387 /*
1388  * Allocate a 'unique' XID
1389  */
1390 static inline u32 xprt_alloc_xid(struct rpc_xprt *xprt)
1391 {
1392         return xprt->xid++;
1393 }
1394
1395 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1396 {
1397         get_random_bytes(&xprt->xid, sizeof(xprt->xid));
1398 }
1399
1400 /*
1401  * Initialize RPC request
1402  */
1403 static void
1404 xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1405 {
1406         struct rpc_rqst *req = task->tk_rqstp;
1407
1408         req->rq_timeout = xprt->timeout.to_initval;
1409         req->rq_task    = task;
1410         req->rq_xprt    = xprt;
1411         req->rq_xid     = xprt_alloc_xid(xprt);
1412         dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
1413                         req, ntohl(req->rq_xid));
1414 }
1415
1416 /*
1417  * Release an RPC call slot
1418  */
1419 void
1420 xprt_release(struct rpc_task *task)
1421 {
1422         struct rpc_xprt *xprt = task->tk_xprt;
1423         struct rpc_rqst *req;
1424
1425         if (!(req = task->tk_rqstp))
1426                 return;
1427         spin_lock_bh(&xprt->sock_lock);
1428         __xprt_release_write(xprt, task);
1429         __xprt_put_cong(xprt, req);
1430         if (!list_empty(&req->rq_list))
1431                 list_del(&req->rq_list);
1432         xprt->last_used = jiffies;
1433         if (list_empty(&xprt->recv) && !xprt->shutdown)
1434                 mod_timer(&xprt->timer, xprt->last_used + XPRT_IDLE_TIMEOUT);
1435         spin_unlock_bh(&xprt->sock_lock);
1436         task->tk_rqstp = NULL;
1437         memset(req, 0, sizeof(*req));   /* mark unused */
1438
1439         dprintk("RPC: %4d release request %p\n", task->tk_pid, req);
1440
1441         spin_lock(&xprt->xprt_lock);
1442         list_add(&req->rq_list, &xprt->free);
1443         xprt_clear_backlog(xprt);
1444         spin_unlock(&xprt->xprt_lock);
1445 }
1446
1447 /*
1448  * Set default timeout parameters
1449  */
1450 static void
1451 xprt_default_timeout(struct rpc_timeout *to, int proto)
1452 {
1453         if (proto == IPPROTO_UDP)
1454                 xprt_set_timeout(to, 5,  5 * HZ);
1455         else
1456                 xprt_set_timeout(to, 5, 60 * HZ);
1457 }
1458
1459 /*
1460  * Set constant timeout
1461  */
1462 void
1463 xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
1464 {
1465         to->to_initval   =
1466         to->to_increment = incr;
1467         to->to_maxval    = incr * retr;
1468         to->to_retries   = retr;
1469         to->to_exponential = 0;
1470 }
1471
1472 unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1473 unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1474
1475 /*
1476  * Initialize an RPC client
1477  */
1478 static struct rpc_xprt *
1479 xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
1480 {
1481         struct rpc_xprt *xprt;
1482         unsigned int entries;
1483         size_t slot_table_size;
1484         struct rpc_rqst *req;
1485
1486         dprintk("RPC:      setting up %s transport...\n",
1487                                 proto == IPPROTO_UDP? "UDP" : "TCP");
1488
1489         entries = (proto == IPPROTO_TCP)?
1490                 xprt_tcp_slot_table_entries : xprt_udp_slot_table_entries;
1491
1492         if ((xprt = kmalloc(sizeof(struct rpc_xprt), GFP_KERNEL)) == NULL)
1493                 return ERR_PTR(-ENOMEM);
1494         memset(xprt, 0, sizeof(*xprt)); /* Nnnngh! */
1495         xprt->max_reqs = entries;
1496         slot_table_size = entries * sizeof(xprt->slot[0]);
1497         xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
1498         if (xprt->slot == NULL) {
1499                 kfree(xprt);
1500                 return ERR_PTR(-ENOMEM);
1501         }
1502         memset(xprt->slot, 0, slot_table_size);
1503
1504         xprt->addr = *ap;
1505         xprt->prot = proto;
1506         xprt->stream = (proto == IPPROTO_TCP)? 1 : 0;
1507         if (xprt->stream) {
1508                 xprt->cwnd = RPC_MAXCWND(xprt);
1509                 xprt->nocong = 1;
1510                 xprt->max_payload = (1U << 31) - 1;
1511         } else {
1512                 xprt->cwnd = RPC_INITCWND;
1513                 xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
1514         }
1515         spin_lock_init(&xprt->sock_lock);
1516         spin_lock_init(&xprt->xprt_lock);
1517         init_waitqueue_head(&xprt->cong_wait);
1518
1519         INIT_LIST_HEAD(&xprt->free);
1520         INIT_LIST_HEAD(&xprt->recv);
1521         INIT_WORK(&xprt->sock_connect, xprt_socket_connect, xprt);
1522         INIT_WORK(&xprt->task_cleanup, xprt_socket_autoclose, xprt);
1523         init_timer(&xprt->timer);
1524         xprt->timer.function = xprt_init_autodisconnect;
1525         xprt->timer.data = (unsigned long) xprt;
1526         xprt->last_used = jiffies;
1527         xprt->port = XPRT_MAX_RESVPORT;
1528
1529         /* Set timeout parameters */
1530         if (to) {
1531                 xprt->timeout = *to;
1532         } else
1533                 xprt_default_timeout(&xprt->timeout, xprt->prot);
1534
1535         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1536         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
1537         rpc_init_wait_queue(&xprt->resend, "xprt_resend");
1538         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1539
1540         /* initialize free list */
1541         for (req = &xprt->slot[entries-1]; req >= &xprt->slot[0]; req--)
1542                 list_add(&req->rq_list, &xprt->free);
1543
1544         xprt_init_xid(xprt);
1545
1546         /* Check whether we want to use a reserved port */
1547         xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1548
1549         dprintk("RPC:      created transport %p with %u slots\n", xprt,
1550                         xprt->max_reqs);
1551
1552         return xprt;
1553 }
1554
1555 /*
1556  * Bind to a reserved port
1557  */
1558 static inline int xprt_bindresvport(struct rpc_xprt *xprt, struct socket *sock)
1559 {
1560         struct sockaddr_in myaddr = {
1561                 .sin_family = AF_INET,
1562         };
1563         int             err, port;
1564
1565         /* Were we already bound to a given port? Try to reuse it */
1566         port = xprt->port;
1567         do {
1568                 myaddr.sin_port = htons(port);
1569                 err = sock->ops->bind(sock, (struct sockaddr *) &myaddr,
1570                                                 sizeof(myaddr));
1571                 if (err == 0) {
1572                         xprt->port = port;
1573                         return 0;
1574                 }
1575                 if (--port == 0)
1576                         port = XPRT_MAX_RESVPORT;
1577         } while (err == -EADDRINUSE && port != xprt->port);
1578
1579         printk("RPC: Can't bind to reserved port (%d).\n", -err);
1580         return err;
1581 }
1582
1583 static void
1584 xprt_bind_socket(struct rpc_xprt *xprt, struct socket *sock)
1585 {
1586         struct sock     *sk = sock->sk;
1587
1588         if (xprt->inet)
1589                 return;
1590
1591         write_lock_bh(&sk->sk_callback_lock);
1592         sk->sk_user_data = xprt;
1593         xprt->old_data_ready = sk->sk_data_ready;
1594         xprt->old_state_change = sk->sk_state_change;
1595         xprt->old_write_space = sk->sk_write_space;
1596         if (xprt->prot == IPPROTO_UDP) {
1597                 sk->sk_data_ready = udp_data_ready;
1598                 sk->sk_no_check = UDP_CSUM_NORCV;
1599                 xprt_set_connected(xprt);
1600         } else {
1601                 tcp_sk(sk)->nonagle = 1;        /* disable Nagle's algorithm */
1602                 sk->sk_data_ready = tcp_data_ready;
1603                 sk->sk_state_change = tcp_state_change;
1604                 xprt_clear_connected(xprt);
1605         }
1606         sk->sk_write_space = xprt_write_space;
1607
1608         /* Reset to new socket */
1609         xprt->sock = sock;
1610         xprt->inet = sk;
1611         write_unlock_bh(&sk->sk_callback_lock);
1612
1613         return;
1614 }
1615
1616 /*
1617  * Set socket buffer length
1618  */
1619 void
1620 xprt_sock_setbufsize(struct rpc_xprt *xprt)
1621 {
1622         struct sock *sk = xprt->inet;
1623
1624         if (xprt->stream)
1625                 return;
1626         if (xprt->rcvsize) {
1627                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
1628                 sk->sk_rcvbuf = xprt->rcvsize * xprt->max_reqs *  2;
1629         }
1630         if (xprt->sndsize) {
1631                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
1632                 sk->sk_sndbuf = xprt->sndsize * xprt->max_reqs * 2;
1633                 sk->sk_write_space(sk);
1634         }
1635 }
1636
1637 /*
1638  * Datastream sockets are created here, but xprt_connect will create
1639  * and connect stream sockets.
1640  */
1641 static struct socket * xprt_create_socket(struct rpc_xprt *xprt, int proto, int resvport)
1642 {
1643         struct socket   *sock;
1644         int             type, err;
1645
1646         dprintk("RPC:      xprt_create_socket(%s %d)\n",
1647                            (proto == IPPROTO_UDP)? "udp" : "tcp", proto);
1648
1649         type = (proto == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1650
1651         if ((err = sock_create_kern(PF_INET, type, proto, &sock)) < 0) {
1652                 printk("RPC: can't create socket (%d).\n", -err);
1653                 return NULL;
1654         }
1655
1656         /* If the caller has the capability, bind to a reserved port */
1657         if (resvport && xprt_bindresvport(xprt, sock) < 0) {
1658                 printk("RPC: can't bind to reserved port.\n");
1659                 goto failed;
1660         }
1661
1662         return sock;
1663
1664 failed:
1665         sock_release(sock);
1666         return NULL;
1667 }
1668
1669 /*
1670  * Create an RPC client transport given the protocol and peer address.
1671  */
1672 struct rpc_xprt *
1673 xprt_create_proto(int proto, struct sockaddr_in *sap, struct rpc_timeout *to)
1674 {
1675         struct rpc_xprt *xprt;
1676
1677         xprt = xprt_setup(proto, sap, to);
1678         if (IS_ERR(xprt))
1679                 dprintk("RPC:      xprt_create_proto failed\n");
1680         else
1681                 dprintk("RPC:      xprt_create_proto created xprt %p\n", xprt);
1682         return xprt;
1683 }
1684
1685 /*
1686  * Prepare for transport shutdown.
1687  */
1688 static void
1689 xprt_shutdown(struct rpc_xprt *xprt)
1690 {
1691         xprt->shutdown = 1;
1692         rpc_wake_up(&xprt->sending);
1693         rpc_wake_up(&xprt->resend);
1694         rpc_wake_up(&xprt->pending);
1695         rpc_wake_up(&xprt->backlog);
1696         wake_up(&xprt->cong_wait);
1697         del_timer_sync(&xprt->timer);
1698
1699         /* synchronously wait for connect worker to finish */
1700         cancel_delayed_work(&xprt->sock_connect);
1701         flush_scheduled_work();
1702 }
1703
1704 /*
1705  * Clear the xprt backlog queue
1706  */
1707 static int
1708 xprt_clear_backlog(struct rpc_xprt *xprt) {
1709         rpc_wake_up_next(&xprt->backlog);
1710         wake_up(&xprt->cong_wait);
1711         return 1;
1712 }
1713
1714 /*
1715  * Destroy an RPC transport, killing off all requests.
1716  */
1717 int
1718 xprt_destroy(struct rpc_xprt *xprt)
1719 {
1720         dprintk("RPC:      destroying transport %p\n", xprt);
1721         xprt_shutdown(xprt);
1722         xprt_disconnect(xprt);
1723         xprt_close(xprt);
1724         kfree(xprt->slot);
1725         kfree(xprt);
1726
1727         return 0;
1728 }