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 #ifdef RPC_DEBUG_DATA
  94 /*
  95  * Print the buffer contents (first 128 bytes only--just enough for
  96  * diropres return).
  97  */
  98 static void
  99 xprt_pktdump(char *msg, u32 *packet, unsigned int count)
 100 {
 101         u8      *buf = (u8 *) packet;
 102         int     j;
 103
 104         dprintk("RPC:      %s\n", msg);
 105         for (j = 0; j < count && j < 128; j += 4) {
 106                 if (!(j & 31)) {
 107                         if (j)
 108                                 dprintk("\n");
 109                         dprintk("0x%04x ", j);
 110                 }
 111                 dprintk("%02x%02x%02x%02x ",
 112                         buf[j], buf[j+1], buf[j+2], buf[j+3]);
 113         }
 114         dprintk("\n");
 115 }
 116 #else
 117 static inline void
 118 xprt_pktdump(char *msg, u32 *packet, unsigned int count)
 119 {
 120         /* NOP */
 121 }
 122 #endif
 123
 124 /*
 125  * Look up RPC transport given an INET socket
 126  */
 127 static inline struct rpc_xprt *
 128 xprt_from_sock(struct sock *sk)
 129 {
 130         return (struct rpc_xprt *) sk->sk_user_data;
 131 }
 132
 133 /*
 134  * Serialize write access to sockets, in order to prevent different
 135  * requests from interfering with each other.
 136  * Also prevents TCP socket connects from colliding with writes.
 137  */
 138 static int
 139 __xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 140 {
 141         struct rpc_rqst *req = task->tk_rqstp;
 142
 143         if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate)) {
 144                 if (task == xprt->snd_task)
 145                         return 1;
 146                 if (task == NULL)
 147                         return 0;
 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  * Adjust timeout values etc for next retransmit
 356  */
 357 int
 358 xprt_adjust_timeout(struct rpc_timeout *to)
 359 {
 360         if (to->to_retries > 0) {
 361                 if (to->to_exponential)
 362                         to->to_current <<= 1;
 363                 else
 364                         to->to_current += to->to_increment;
 365                 if (to->to_maxval && to->to_current >= to->to_maxval)
 366                         to->to_current = to->to_maxval;
 367         } else {
 368                 if (to->to_exponential)
 369                         to->to_initval <<= 1;
 370                 else
 371                         to->to_initval += to->to_increment;
 372                 if (to->to_maxval && to->to_initval >= to->to_maxval)
 373                         to->to_initval = to->to_maxval;
 374                 to->to_current = to->to_initval;
 375         }
 376
 377         if (!to->to_current) {
 378                 printk(KERN_WARNING "xprt_adjust_timeout: to_current = 0!\n");
 379                 to->to_current = 5 * HZ;
 380         }
 381         pprintk("RPC: %lu %s\n", jiffies,
 382                         to->to_retries? "retrans" : "timeout");
 383         return to->to_retries-- > 0;
 384 }
 385
 386 /*
 387  * Close down a transport socket
 388  */
 389 static void
 390 xprt_close(struct rpc_xprt *xprt)
 391 {
 392         struct socket   *sock = xprt->sock;
 393         struct sock     *sk = xprt->inet;
 394
 395         if (!sk)
 396                 return;
 397
 398         write_lock_bh(&sk->sk_callback_lock);
 399         xprt->inet = NULL;
 400         xprt->sock = NULL;
 401
 402         sk->sk_user_data    = NULL;
 403         sk->sk_data_ready   = xprt->old_data_ready;
 404         sk->sk_state_change = xprt->old_state_change;
 405         sk->sk_write_space  = xprt->old_write_space;
 406         write_unlock_bh(&sk->sk_callback_lock);
 407
 408         sk->sk_no_check  = 0;
 409
 410         sock_release(sock);
 411 }
 412
 413 static void
 414 xprt_socket_autoclose(void *args)
 415 {
 416         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
 417
 418         xprt_disconnect(xprt);
 419         xprt_close(xprt);
 420         xprt_release_write(xprt, NULL);
 421 }
 422
 423 /*
 424  * Mark a transport as disconnected
 425  */
 426 static void
 427 xprt_disconnect(struct rpc_xprt *xprt)
 428 {
 429         dprintk("RPC:      disconnected transport %p\n", xprt);
 430         spin_lock_bh(&xprt->sock_lock);
 431         xprt_clear_connected(xprt);
 432         rpc_wake_up_status(&xprt->pending, -ENOTCONN);
 433         spin_unlock_bh(&xprt->sock_lock);
 434 }
 435
 436 /*
 437  * Used to allow disconnection when we've been idle
 438  */
 439 static void
 440 xprt_init_autodisconnect(unsigned long data)
 441 {
 442         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 443
 444         spin_lock(&xprt->sock_lock);
 445         if (!list_empty(&xprt->recv) || xprt->shutdown)
 446                 goto out_abort;
 447         if (test_and_set_bit(XPRT_LOCKED, &xprt->sockstate))
 448                 goto out_abort;
 449         spin_unlock(&xprt->sock_lock);
 450         /* Let keventd close the socket */
 451         if (test_bit(XPRT_CONNECTING, &xprt->sockstate) != 0)
 452                 xprt_release_write(xprt, NULL);
 453         else
 454                 schedule_work(&xprt->task_cleanup);
 455         return;
 456 out_abort:
 457         spin_unlock(&xprt->sock_lock);
 458 }
 459
 460 static void xprt_socket_connect(void *args)
 461 {
 462         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
 463         struct socket *sock = xprt->sock;
 464         int status = -EIO;
 465
 466         if (xprt->shutdown || xprt->addr.sin_port == 0)
 467                 goto out;
 468
 469         /*
 470          * Start by resetting any existing state
 471          */
 472         xprt_close(xprt);
 473         sock = xprt_create_socket(xprt, xprt->prot, xprt->resvport);
 474         if (sock == NULL) {
 475                 /* couldn't create socket or bind to reserved port;
 476                  * this is likely a permanent error, so cause an abort */
 477                 goto out;
 478         }
 479         xprt_bind_socket(xprt, sock);
 480         xprt_sock_setbufsize(xprt);
 481
 482         status = 0;
 483         if (!xprt->stream)
 484                 goto out;
 485
 486         /*
 487          * Tell the socket layer to start connecting...
 488          */
 489         status = sock->ops->connect(sock, (struct sockaddr *) &xprt->addr,
 490                         sizeof(xprt->addr), O_NONBLOCK);
 491         dprintk("RPC: %p  connect status %d connected %d sock state %d\n",
 492                         xprt, -status, xprt_connected(xprt), sock->sk->sk_state);
 493         if (status < 0) {
 494                 switch (status) {
 495                         case -EINPROGRESS:
 496                         case -EALREADY:
 497                                 goto out_clear;
 498                 }
 499         }
 500 out:
 501         if (status < 0)
 502                 rpc_wake_up_status(&xprt->pending, status);
 503         else
 504                 rpc_wake_up(&xprt->pending);
 505 out_clear:
 506         smp_mb__before_clear_bit();
 507         clear_bit(XPRT_CONNECTING, &xprt->sockstate);
 508         smp_mb__after_clear_bit();
 509 }
 510
 511 /*
 512  * Attempt to connect a TCP socket.
 513  *
 514  */
 515 void xprt_connect(struct rpc_task *task)
 516 {
 517         struct rpc_xprt *xprt = task->tk_xprt;
 518
 519         dprintk("RPC: %4d xprt_connect xprt %p %s connected\n", task->tk_pid,
 520                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 521
 522         if (xprt->shutdown) {
 523                 task->tk_status = -EIO;
 524                 return;
 525         }
 526         if (!xprt->addr.sin_port) {
 527                 task->tk_status = -EIO;
 528                 return;
 529         }
 530         if (!xprt_lock_write(xprt, task))
 531                 return;
 532         if (xprt_connected(xprt))
 533                 goto out_write;
 534
 535         if (task->tk_rqstp)
 536                 task->tk_rqstp->rq_bytes_sent = 0;
 537
 538         task->tk_timeout = RPC_CONNECT_TIMEOUT;
 539         rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
 540         if (!test_and_set_bit(XPRT_CONNECTING, &xprt->sockstate))
 541                 schedule_work(&xprt->sock_connect);
 542         return;
 543  out_write:
 544         xprt_release_write(xprt, task);
 545 }
 546
 547 /*
 548  * We arrive here when awoken from waiting on connection establishment.
 549  */
 550 static void
 551 xprt_connect_status(struct rpc_task *task)
 552 {
 553         struct rpc_xprt *xprt = task->tk_xprt;
 554
 555         if (task->tk_status >= 0) {
 556                 dprintk("RPC: %4d xprt_connect_status: connection established\n",
 557                                 task->tk_pid);
 558                 return;
 559         }
 560
 561         /* if soft mounted, just cause this RPC to fail */
 562         if (RPC_IS_SOFT(task))
 563                 task->tk_status = -EIO;
 564
 565         switch (task->tk_status) {
 566         case -ECONNREFUSED:
 567         case -ECONNRESET:
 568         case -ENOTCONN:
 569                 rpc_delay(task, RPC_REESTABLISH_TIMEOUT);
 570                 return;
 571         case -ETIMEDOUT:
 572                 dprintk("RPC: %4d xprt_connect_status: timed out\n",
 573                                 task->tk_pid);
 574                 break;
 575         default:
 576                 printk(KERN_ERR "RPC: error %d connecting to server %s\n",
 577                                 -task->tk_status, task->tk_client->cl_server);
 578         }
 579         xprt_release_write(xprt, task);
 580 }
 581
 582 /*
 583  * Look up the RPC request corresponding to a reply, and then lock it.
 584  */
 585 static inline struct rpc_rqst *
 586 xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
 587 {
 588         struct list_head *pos;
 589         struct rpc_rqst *req = NULL;
 590
 591         list_for_each(pos, &xprt->recv) {
 592                 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
 593                 if (entry->rq_xid == xid) {
 594                         req = entry;
 595                         break;
 596                 }
 597         }
 598         return req;
 599 }
 600
 601 /*
 602  * Complete reply received.
 603  * The TCP code relies on us to remove the request from xprt->pending.
 604  */
 605 static void
 606 xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
 607 {
 608         struct rpc_task *task = req->rq_task;
 609         struct rpc_clnt *clnt = task->tk_client;
 610
 611         /* Adjust congestion window */
 612         if (!xprt->nocong) {
 613                 unsigned timer = task->tk_msg.rpc_proc->p_timer;
 614                 xprt_adjust_cwnd(xprt, copied);
 615                 __xprt_put_cong(xprt, req);
 616                 if (timer) {
 617                         if (req->rq_ntrans == 1)
 618                                 rpc_update_rtt(clnt->cl_rtt, timer,
 619                                                 (long)jiffies - req->rq_xtime);
 620                         rpc_set_timeo(clnt->cl_rtt, timer, req->rq_ntrans - 1);
 621                 }
 622         }
 623
 624 #ifdef RPC_PROFILE
 625         /* Profile only reads for now */
 626         if (copied > 1024) {
 627                 static unsigned long    nextstat;
 628                 static unsigned long    pkt_rtt, pkt_len, pkt_cnt;
 629
 630                 pkt_cnt++;
 631                 pkt_len += req->rq_slen + copied;
 632                 pkt_rtt += jiffies - req->rq_xtime;
 633                 if (time_before(nextstat, jiffies)) {
 634                         printk("RPC: %lu %ld cwnd\n", jiffies, xprt->cwnd);
 635                         printk("RPC: %ld %ld %ld %ld stat\n",
 636                                         jiffies, pkt_cnt, pkt_len, pkt_rtt);
 637                         pkt_rtt = pkt_len = pkt_cnt = 0;
 638                         nextstat = jiffies + 5 * HZ;
 639                 }
 640         }
 641 #endif
 642
 643         dprintk("RPC: %4d has input (%d bytes)\n", task->tk_pid, copied);
 644         list_del_init(&req->rq_list);
 645         req->rq_received = req->rq_private_buf.len = copied;
 646
 647         /* ... and wake up the process. */
 648         rpc_wake_up_task(task);
 649         return;
 650 }
 651
 652 static size_t
 653 skb_read_bits(skb_reader_t *desc, void *to, size_t len)
 654 {
 655         if (len > desc->count)
 656                 len = desc->count;
 657         if (skb_copy_bits(desc->skb, desc->offset, to, len))
 658                 return 0;
 659         desc->count -= len;
 660         desc->offset += len;
 661         return len;
 662 }
 663
 664 static size_t
 665 skb_read_and_csum_bits(skb_reader_t *desc, void *to, size_t len)
 666 {
 667         unsigned int csum2, pos;
 668
 669         if (len > desc->count)
 670                 len = desc->count;
 671         pos = desc->offset;
 672         csum2 = skb_copy_and_csum_bits(desc->skb, pos, to, len, 0);
 673         desc->csum = csum_block_add(desc->csum, csum2, pos);
 674         desc->count -= len;
 675         desc->offset += len;
 676         return len;
 677 }
 678
 679 /*
 680  * We have set things up such that we perform the checksum of the UDP
 681  * packet in parallel with the copies into the RPC client iovec.  -DaveM
 682  */
 683 int
 684 csum_partial_copy_to_xdr(struct xdr_buf *xdr, struct sk_buff *skb)
 685 {
 686         skb_reader_t desc;
 687
 688         desc.skb = skb;
 689         desc.offset = sizeof(struct udphdr);
 690         desc.count = skb->len - desc.offset;
 691
 692         if (skb->ip_summed == CHECKSUM_UNNECESSARY)
 693                 goto no_checksum;
 694
 695         desc.csum = csum_partial(skb->data, desc.offset, skb->csum);
 696         xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_and_csum_bits);
 697         if (desc.offset != skb->len) {
 698                 unsigned int csum2;
 699                 csum2 = skb_checksum(skb, desc.offset, skb->len - desc.offset, 0);
 700                 desc.csum = csum_block_add(desc.csum, csum2, desc.offset);
 701         }
 702         if (desc.count)
 703                 return -1;
 704         if ((unsigned short)csum_fold(desc.csum))
 705                 return -1;
 706         return 0;
 707 no_checksum:
 708         xdr_partial_copy_from_skb(xdr, 0, &desc, skb_read_bits);
 709         if (desc.count)
 710                 return -1;
 711         return 0;
 712 }
 713
 714 /*
 715  * Input handler for RPC replies. Called from a bottom half and hence
 716  * atomic.
 717  */
 718 static void
 719 udp_data_ready(struct sock *sk, int len)
 720 {
 721         struct rpc_task *task;
 722         struct rpc_xprt *xprt;
 723         struct rpc_rqst *rovr;
 724         struct sk_buff  *skb;
 725         int err, repsize, copied;
 726         u32 xid;
 727
 728         read_lock(&sk->sk_callback_lock);
 729         dprintk("RPC:      udp_data_ready...\n");
 730         if (!(xprt = xprt_from_sock(sk))) {
 731                 printk("RPC:      udp_data_ready request not found!\n");
 732                 goto out;
 733         }
 734
 735         dprintk("RPC:      udp_data_ready client %p\n", xprt);
 736
 737         if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
 738                 goto out;
 739
 740         if (xprt->shutdown)
 741                 goto dropit;
 742
 743         repsize = skb->len - sizeof(struct udphdr);
 744         if (repsize < 4) {
 745                 printk("RPC: impossible RPC reply size %d!\n", repsize);
 746                 goto dropit;
 747         }
 748
 749         /* Copy the XID from the skb... */
 750         if (skb_copy_bits(skb, sizeof(struct udphdr), &xid, sizeof(xid)) < 0)
 751                 goto dropit;
 752
 753         /* Look up and lock the request corresponding to the given XID */
 754         spin_lock(&xprt->sock_lock);
 755         rovr = xprt_lookup_rqst(xprt, xid);
 756         if (!rovr)
 757                 goto out_unlock;
 758         task = rovr->rq_task;
 759
 760         dprintk("RPC: %4d received reply\n", task->tk_pid);
 761
 762         if ((copied = rovr->rq_private_buf.buflen) > repsize)
 763                 copied = repsize;
 764
 765         /* Suck it into the iovec, verify checksum if not done by hw. */
 766         if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
 767                 goto out_unlock;
 768
 769         /* Something worked... */
 770         dst_confirm(skb->dst);
 771
 772         xprt_complete_rqst(xprt, rovr, copied);
 773
 774  out_unlock:
 775         spin_unlock(&xprt->sock_lock);
 776  dropit:
 777         skb_free_datagram(sk, skb);
 778  out:
 779         read_unlock(&sk->sk_callback_lock);
 780 }
 781
 782 /*
 783  * Copy from an skb into memory and shrink the skb.
 784  */
 785 static inline size_t
 786 tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
 787 {
 788         if (len > desc->count)
 789                 len = desc->count;
 790         if (skb_copy_bits(desc->skb, desc->offset, p, len))
 791                 return 0;
 792         desc->offset += len;
 793         desc->count -= len;
 794         return len;
 795 }
 796
 797 /*
 798  * TCP read fragment marker
 799  */
 800 static inline void
 801 tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
 802 {
 803         size_t len, used;
 804         char *p;
 805
 806         p = ((char *) &xprt->tcp_recm) + xprt->tcp_offset;
 807         len = sizeof(xprt->tcp_recm) - xprt->tcp_offset;
 808         used = tcp_copy_data(desc, p, len);
 809         xprt->tcp_offset += used;
 810         if (used != len)
 811                 return;
 812         xprt->tcp_reclen = ntohl(xprt->tcp_recm);
 813         if (xprt->tcp_reclen & 0x80000000)
 814                 xprt->tcp_flags |= XPRT_LAST_FRAG;
 815         else
 816                 xprt->tcp_flags &= ~XPRT_LAST_FRAG;
 817         xprt->tcp_reclen &= 0x7fffffff;
 818         xprt->tcp_flags &= ~XPRT_COPY_RECM;
 819         xprt->tcp_offset = 0;
 820         /* Sanity check of the record length */
 821         if (xprt->tcp_reclen < 4) {
 822                 printk(KERN_ERR "RPC: Invalid TCP record fragment length\n");
 823                 xprt_disconnect(xprt);
 824         }
 825         dprintk("RPC:      reading TCP record fragment of length %d\n",
 826                         xprt->tcp_reclen);
 827 }
 828
 829 static void
 830 tcp_check_recm(struct rpc_xprt *xprt)
 831 {
 832         if (xprt->tcp_offset == xprt->tcp_reclen) {
 833                 xprt->tcp_flags |= XPRT_COPY_RECM;
 834                 xprt->tcp_offset = 0;
 835                 if (xprt->tcp_flags & XPRT_LAST_FRAG) {
 836                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
 837                         xprt->tcp_flags |= XPRT_COPY_XID;
 838                         xprt->tcp_copied = 0;
 839                 }
 840         }
 841 }
 842
 843 /*
 844  * TCP read xid
 845  */
 846 static inline void
 847 tcp_read_xid(struct rpc_xprt *xprt, skb_reader_t *desc)
 848 {
 849         size_t len, used;
 850         char *p;
 851
 852         len = sizeof(xprt->tcp_xid) - xprt->tcp_offset;
 853         dprintk("RPC:      reading XID (%Zu bytes)\n", len);
 854         p = ((char *) &xprt->tcp_xid) + xprt->tcp_offset;
 855         used = tcp_copy_data(desc, p, len);
 856         xprt->tcp_offset += used;
 857         if (used != len)
 858                 return;
 859         xprt->tcp_flags &= ~XPRT_COPY_XID;
 860         xprt->tcp_flags |= XPRT_COPY_DATA;
 861         xprt->tcp_copied = 4;
 862         dprintk("RPC:      reading reply for XID %08x\n", xprt->tcp_xid);
 863         tcp_check_recm(xprt);
 864 }
 865
 866 /*
 867  * TCP read and complete request
 868  */
 869 static inline void
 870 tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
 871 {
 872         struct rpc_rqst *req;
 873         struct xdr_buf *rcvbuf;
 874         size_t len;
 875
 876         /* Find and lock the request corresponding to this xid */
 877         spin_lock(&xprt->sock_lock);
 878         req = xprt_lookup_rqst(xprt, xprt->tcp_xid);
 879         if (!req) {
 880                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 881                 dprintk("RPC:      XID %08x request not found!\n",
 882                                 xprt->tcp_xid);
 883                 spin_unlock(&xprt->sock_lock);
 884                 return;
 885         }
 886
 887         rcvbuf = &req->rq_private_buf;
 888         len = desc->count;
 889         if (len > xprt->tcp_reclen - xprt->tcp_offset) {
 890                 skb_reader_t my_desc;
 891
 892                 len = xprt->tcp_reclen - xprt->tcp_offset;
 893                 memcpy(&my_desc, desc, sizeof(my_desc));
 894                 my_desc.count = len;
 895                 xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
 896                                           &my_desc, tcp_copy_data);
 897                 desc->count -= len;
 898                 desc->offset += len;
 899         } else
 900                 xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
 901                                           desc, tcp_copy_data);
 902         xprt->tcp_copied += len;
 903         xprt->tcp_offset += len;
 904
 905         if (xprt->tcp_copied == req->rq_private_buf.buflen)
 906                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 907         else if (xprt->tcp_offset == xprt->tcp_reclen) {
 908                 if (xprt->tcp_flags & XPRT_LAST_FRAG)
 909                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
 910         }
 911
 912         if (!(xprt->tcp_flags & XPRT_COPY_DATA)) {
 913                 dprintk("RPC: %4d received reply complete\n",
 914                                 req->rq_task->tk_pid);
 915                 xprt_complete_rqst(xprt, req, xprt->tcp_copied);
 916         }
 917         spin_unlock(&xprt->sock_lock);
 918         tcp_check_recm(xprt);
 919 }
 920
 921 /*
 922  * TCP discard extra bytes from a short read
 923  */
 924 static inline void
 925 tcp_read_discard(struct rpc_xprt *xprt, skb_reader_t *desc)
 926 {
 927         size_t len;
 928
 929         len = xprt->tcp_reclen - xprt->tcp_offset;
 930         if (len > desc->count)
 931                 len = desc->count;
 932         desc->count -= len;
 933         desc->offset += len;
 934         xprt->tcp_offset += len;
 935         tcp_check_recm(xprt);
 936 }
 937
 938 /*
 939  * TCP record receive routine
 940  * We first have to grab the record marker, then the XID, then the data.
 941  */
 942 static int
 943 tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 944                 unsigned int offset, size_t len)
 945 {
 946         struct rpc_xprt *xprt = (struct rpc_xprt *)rd_desc->buf;
 947         skb_reader_t desc = {
 948                 .skb    = skb,
 949                 .offset = offset,
 950                 .count  = len,
 951                 .csum   = 0
 952         };
 953
 954         dprintk("RPC:      tcp_data_recv\n");
 955         do {
 956                 /* Read in a new fragment marker if necessary */
 957                 /* Can we ever really expect to get completely empty fragments? */
 958                 if (xprt->tcp_flags & XPRT_COPY_RECM) {
 959                         tcp_read_fraghdr(xprt, &desc);
 960                         continue;
 961                 }
 962                 /* Read in the xid if necessary */
 963                 if (xprt->tcp_flags & XPRT_COPY_XID) {
 964                         tcp_read_xid(xprt, &desc);
 965                         continue;
 966                 }
 967                 /* Read in the request data */
 968                 if (xprt->tcp_flags & XPRT_COPY_DATA) {
 969                         tcp_read_request(xprt, &desc);
 970                         continue;
 971                 }
 972                 /* Skip over any trailing bytes on short reads */
 973                 tcp_read_discard(xprt, &desc);
 974         } while (desc.count);
 975         dprintk("RPC:      tcp_data_recv done\n");
 976         return len - desc.count;
 977 }
 978
 979 static void tcp_data_ready(struct sock *sk, int bytes)
 980 {
 981         struct rpc_xprt *xprt;
 982         read_descriptor_t rd_desc;
 983
 984         read_lock(&sk->sk_callback_lock);
 985         dprintk("RPC:      tcp_data_ready...\n");
 986         if (!(xprt = xprt_from_sock(sk))) {
 987                 printk("RPC:      tcp_data_ready socket info not found!\n");
 988                 goto out;
 989         }
 990         if (xprt->shutdown)
 991                 goto out;
 992
 993         /* We use rd_desc to pass struct xprt to tcp_data_recv */
 994         rd_desc.buf = (char *)xprt;
 995         rd_desc.count = 65536;
 996         tcp_read_sock(sk, &rd_desc, tcp_data_recv);
 997 out:
 998         read_unlock(&sk->sk_callback_lock);
 999 }
1000
1001 static void
1002 tcp_state_change(struct sock *sk)
1003 {
1004         struct rpc_xprt *xprt;
1005
1006         read_lock(&sk->sk_callback_lock);
1007         if (!(xprt = xprt_from_sock(sk)))
1008                 goto out;
1009         dprintk("RPC:      tcp_state_change client %p...\n", xprt);
1010         dprintk("RPC:      state %x conn %d dead %d zapped %d\n",
1011                                 sk->sk_state, xprt_connected(xprt),
1012                                 sock_flag(sk, SOCK_DEAD), sk->sk_zapped);
1013
1014         switch (sk->sk_state) {
1015         case TCP_ESTABLISHED:
1016                 spin_lock_bh(&xprt->sock_lock);
1017                 if (!xprt_test_and_set_connected(xprt)) {
1018                         /* Reset TCP record info */
1019                         xprt->tcp_offset = 0;
1020                         xprt->tcp_reclen = 0;
1021                         xprt->tcp_copied = 0;
1022                         xprt->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
1023                         rpc_wake_up(&xprt->pending);
1024                 }
1025                 spin_unlock_bh(&xprt->sock_lock);
1026                 break;
1027         case TCP_SYN_SENT:
1028         case TCP_SYN_RECV:
1029                 break;
1030         default:
1031                 if (xprt_test_and_clear_connected(xprt))
1032                         rpc_wake_up_status(&xprt->pending, -ENOTCONN);
1033                 break;
1034         }
1035  out:
1036         read_unlock(&sk->sk_callback_lock);
1037 }
1038
1039 /*
1040  * Called when more output buffer space is available for this socket.
1041  * We try not to wake our writers until they can make "significant"
1042  * progress, otherwise we'll waste resources thrashing sock_sendmsg
1043  * with a bunch of small requests.
1044  */
1045 static void
1046 xprt_write_space(struct sock *sk)
1047 {
1048         struct rpc_xprt *xprt;
1049         struct socket   *sock;
1050
1051         read_lock(&sk->sk_callback_lock);
1052         if (!(xprt = xprt_from_sock(sk)) || !(sock = sk->sk_socket))
1053                 goto out;
1054         if (xprt->shutdown)
1055                 goto out;
1056
1057         /* Wait until we have enough socket memory */
1058         if (xprt->stream) {
1059                 /* from net/ipv4/tcp.c:tcp_write_space */
1060                 if (tcp_wspace(sk) < tcp_min_write_space(sk))
1061                         goto out;
1062         } else {
1063                 /* from net/core/sock.c:sock_def_write_space */
1064                 if (!sock_writeable(sk))
1065                         goto out;
1066         }
1067
1068         if (!test_and_clear_bit(SOCK_NOSPACE, &sock->flags))
1069                 goto out;
1070
1071         spin_lock_bh(&xprt->sock_lock);
1072         if (xprt->snd_task && xprt->snd_task->tk_rpcwait == &xprt->pending)
1073                 rpc_wake_up_task(xprt->snd_task);
1074         spin_unlock_bh(&xprt->sock_lock);
1075 out:
1076         read_unlock(&sk->sk_callback_lock);
1077 }
1078
1079 /*
1080  * RPC receive timeout handler.
1081  */
1082 static void
1083 xprt_timer(struct rpc_task *task)
1084 {
1085         struct rpc_rqst *req = task->tk_rqstp;
1086         struct rpc_xprt *xprt = req->rq_xprt;
1087
1088         spin_lock(&xprt->sock_lock);
1089         if (req->rq_received)
1090                 goto out;
1091
1092         xprt_adjust_cwnd(req->rq_xprt, -ETIMEDOUT);
1093         __xprt_put_cong(xprt, req);
1094
1095         dprintk("RPC: %4d xprt_timer (%s request)\n",
1096                 task->tk_pid, req ? "pending" : "backlogged");
1097
1098         task->tk_status  = -ETIMEDOUT;
1099 out:
1100         task->tk_timeout = 0;
1101         rpc_wake_up_task(task);
1102         spin_unlock(&xprt->sock_lock);
1103 }
1104
1105 /*
1106  * Place the actual RPC call.
1107  * We have to copy the iovec because sendmsg fiddles with its contents.
1108  */
1109 int
1110 xprt_prepare_transmit(struct rpc_task *task)
1111 {
1112         struct rpc_rqst *req = task->tk_rqstp;
1113         struct rpc_xprt *xprt = req->rq_xprt;
1114         int err = 0;
1115
1116         dprintk("RPC: %4d xprt_prepare_transmit\n", task->tk_pid);
1117
1118         if (xprt->shutdown)
1119                 return -EIO;
1120
1121         spin_lock_bh(&xprt->sock_lock);
1122         if (req->rq_received && !req->rq_bytes_sent) {
1123                 err = req->rq_received;
1124                 goto out_unlock;
1125         }
1126         if (!__xprt_lock_write(xprt, task)) {
1127                 err = -EAGAIN;
1128                 goto out_unlock;
1129         }
1130
1131         if (!xprt_connected(xprt)) {
1132                 err = -ENOTCONN;
1133                 goto out_unlock;
1134         }
1135 out_unlock:
1136         spin_unlock_bh(&xprt->sock_lock);
1137         return err;
1138 }
1139
1140 void
1141 xprt_transmit(struct rpc_task *task)
1142 {
1143         struct rpc_clnt *clnt = task->tk_client;
1144         struct rpc_rqst *req = task->tk_rqstp;
1145         struct rpc_xprt *xprt = req->rq_xprt;
1146         int status, retry = 0;
1147
1148
1149         dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
1150
1151         /* set up everything as needed. */
1152         /* Write the record marker */
1153         if (xprt->stream) {
1154                 u32     *marker = req->rq_svec[0].iov_base;
1155
1156                 *marker = htonl(0x80000000|(req->rq_slen-sizeof(*marker)));
1157         }
1158
1159         smp_rmb();
1160         if (!req->rq_received) {
1161                 if (list_empty(&req->rq_list)) {
1162                         spin_lock_bh(&xprt->sock_lock);
1163                         /* Update the softirq receive buffer */
1164                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
1165                                         sizeof(req->rq_private_buf));
1166                         /* Add request to the receive list */
1167                         list_add_tail(&req->rq_list, &xprt->recv);
1168                         spin_unlock_bh(&xprt->sock_lock);
1169                 }
1170         } else if (!req->rq_bytes_sent)
1171                 return;
1172
1173         /* Continue transmitting the packet/record. We must be careful
1174          * to cope with writespace callbacks arriving _after_ we have
1175          * called xprt_sendmsg().
1176          */
1177         while (1) {
1178                 req->rq_xtime = jiffies;
1179                 status = xprt_sendmsg(xprt, req);
1180
1181                 if (status < 0)
1182                         break;
1183
1184                 if (xprt->stream) {
1185                         req->rq_bytes_sent += status;
1186
1187                         /* If we've sent the entire packet, immediately
1188                          * reset the count of bytes sent. */
1189                         if (req->rq_bytes_sent >= req->rq_slen) {
1190                                 req->rq_bytes_sent = 0;
1191                                 goto out_receive;
1192                         }
1193                 } else {
1194                         if (status >= req->rq_slen)
1195                                 goto out_receive;
1196                         status = -EAGAIN;
1197                         break;
1198                 }
1199
1200                 dprintk("RPC: %4d xmit incomplete (%d left of %d)\n",
1201                                 task->tk_pid, req->rq_slen - req->rq_bytes_sent,
1202                                 req->rq_slen);
1203
1204                 status = -EAGAIN;
1205                 if (retry++ > 50)
1206                         break;
1207         }
1208
1209         /* Note: at this point, task->tk_sleeping has not yet been set,
1210          *       hence there is no danger of the waking up task being put on
1211          *       schedq, and being picked up by a parallel run of rpciod().
1212          */
1213         task->tk_status = status;
1214
1215         switch (status) {
1216         case -EAGAIN:
1217                 if (test_bit(SOCK_ASYNC_NOSPACE, &xprt->sock->flags)) {
1218                         /* Protect against races with xprt_write_space */
1219                         spin_lock_bh(&xprt->sock_lock);
1220                         /* Don't race with disconnect */
1221                         if (!xprt_connected(xprt))
1222                                 task->tk_status = -ENOTCONN;
1223                         else if (test_bit(SOCK_NOSPACE, &xprt->sock->flags)) {
1224                                 task->tk_timeout = req->rq_timeout.to_current;
1225                                 rpc_sleep_on(&xprt->pending, task, NULL, NULL);
1226                         }
1227                         spin_unlock_bh(&xprt->sock_lock);
1228                         return;
1229                 }
1230                 /* Keep holding the socket if it is blocked */
1231                 rpc_delay(task, HZ>>4);
1232                 return;
1233         case -ECONNREFUSED:
1234                 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
1235                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
1236         case -ENOTCONN:
1237                 return;
1238         default:
1239                 if (xprt->stream)
1240                         xprt_disconnect(xprt);
1241         }
1242         xprt_release_write(xprt, task);
1243         return;
1244  out_receive:
1245         dprintk("RPC: %4d xmit complete\n", task->tk_pid);
1246         /* Set the task's receive timeout value */
1247         spin_lock_bh(&xprt->sock_lock);
1248         if (!xprt->nocong) {
1249                 int timer = task->tk_msg.rpc_proc->p_timer;
1250                 task->tk_timeout = rpc_calc_rto(clnt->cl_rtt, timer);
1251                 task->tk_timeout <<= rpc_ntimeo(clnt->cl_rtt, timer);
1252                 task->tk_timeout <<= clnt->cl_timeout.to_retries
1253                         - req->rq_timeout.to_retries;
1254                 if (task->tk_timeout > req->rq_timeout.to_maxval)
1255                         task->tk_timeout = req->rq_timeout.to_maxval;
1256         } else
1257                 task->tk_timeout = req->rq_timeout.to_current;
1258         /* Don't race with disconnect */
1259         if (!xprt_connected(xprt))
1260                 task->tk_status = -ENOTCONN;
1261         else if (!req->rq_received)
1262                 rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
1263         __xprt_release_write(xprt, task);
1264         spin_unlock_bh(&xprt->sock_lock);
1265 }
1266
1267 /*
1268  * Reserve an RPC call slot.
1269  */
1270 void
1271 xprt_reserve(struct rpc_task *task)
1272 {
1273         struct rpc_xprt *xprt = task->tk_xprt;
1274
1275         task->tk_status = -EIO;
1276         if (!xprt->shutdown) {
1277                 spin_lock(&xprt->xprt_lock);
1278                 do_xprt_reserve(task);
1279                 spin_unlock(&xprt->xprt_lock);
1280                 if (task->tk_rqstp)
1281                         del_timer_sync(&xprt->timer);
1282         }
1283 }
1284
1285 static inline void
1286 do_xprt_reserve(struct rpc_task *task)
1287 {
1288         struct rpc_xprt *xprt = task->tk_xprt;
1289
1290         task->tk_status = 0;
1291         if (task->tk_rqstp)
1292                 return;
1293         if (!list_empty(&xprt->free)) {
1294                 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1295                 list_del_init(&req->rq_list);
1296                 task->tk_rqstp = req;
1297                 xprt_request_init(task, xprt);
1298                 return;
1299         }
1300         dprintk("RPC:      waiting for request slot\n");
1301         task->tk_status = -EAGAIN;
1302         task->tk_timeout = 0;
1303         rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
1304 }
1305
1306 /*
1307  * Allocate a 'unique' XID
1308  */
1309 static inline u32 xprt_alloc_xid(struct rpc_xprt *xprt)
1310 {
1311         return xprt->xid++;
1312 }
1313
1314 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1315 {
1316         get_random_bytes(&xprt->xid, sizeof(xprt->xid));
1317 }
1318
1319 /*
1320  * Initialize RPC request
1321  */
1322 static void
1323 xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1324 {
1325         struct rpc_rqst *req = task->tk_rqstp;
1326
1327         req->rq_timeout = xprt->timeout;
1328         req->rq_task    = task;
1329         req->rq_xprt    = xprt;
1330         req->rq_xid     = xprt_alloc_xid(xprt);
1331         dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
1332                         req, req->rq_xid);
1333 }
1334
1335 /*
1336  * Release an RPC call slot
1337  */
1338 void
1339 xprt_release(struct rpc_task *task)
1340 {
1341         struct rpc_xprt *xprt = task->tk_xprt;
1342         struct rpc_rqst *req;
1343
1344         if (!(req = task->tk_rqstp))
1345                 return;
1346         spin_lock_bh(&xprt->sock_lock);
1347         __xprt_release_write(xprt, task);
1348         __xprt_put_cong(xprt, req);
1349         if (!list_empty(&req->rq_list))
1350                 list_del(&req->rq_list);
1351         xprt->last_used = jiffies;
1352         if (list_empty(&xprt->recv) && !xprt->shutdown)
1353                 mod_timer(&xprt->timer, xprt->last_used + XPRT_IDLE_TIMEOUT);
1354         spin_unlock_bh(&xprt->sock_lock);
1355         task->tk_rqstp = NULL;
1356         memset(req, 0, sizeof(*req));   /* mark unused */
1357
1358         dprintk("RPC: %4d release request %p\n", task->tk_pid, req);
1359
1360         spin_lock(&xprt->xprt_lock);
1361         list_add(&req->rq_list, &xprt->free);
1362         xprt_clear_backlog(xprt);
1363         spin_unlock(&xprt->xprt_lock);
1364 }
1365
1366 /*
1367  * Set default timeout parameters
1368  */
1369 void
1370 xprt_default_timeout(struct rpc_timeout *to, int proto)
1371 {
1372         if (proto == IPPROTO_UDP)
1373                 xprt_set_timeout(to, 5,  5 * HZ);
1374         else
1375                 xprt_set_timeout(to, 5, 60 * HZ);
1376 }
1377
1378 /*
1379  * Set constant timeout
1380  */
1381 void
1382 xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
1383 {
1384         to->to_current   =
1385         to->to_initval   =
1386         to->to_increment = incr;
1387         to->to_maxval    = incr * retr;
1388         to->to_retries   = retr;
1389         to->to_exponential = 0;
1390 }
1391
1392 unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1393 unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
1394
1395 /*
1396  * Initialize an RPC client
1397  */
1398 static struct rpc_xprt *
1399 xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
1400 {
1401         struct rpc_xprt *xprt;
1402         unsigned int entries;
1403         size_t slot_table_size;
1404         struct rpc_rqst *req;
1405
1406         dprintk("RPC:      setting up %s transport...\n",
1407                                 proto == IPPROTO_UDP? "UDP" : "TCP");
1408
1409         entries = (proto == IPPROTO_TCP)?
1410                 xprt_tcp_slot_table_entries : xprt_udp_slot_table_entries;
1411
1412         if ((xprt = kmalloc(sizeof(struct rpc_xprt), GFP_KERNEL)) == NULL)
1413                 return ERR_PTR(-ENOMEM);
1414         memset(xprt, 0, sizeof(*xprt)); /* Nnnngh! */
1415         xprt->max_reqs = entries;
1416         slot_table_size = entries * sizeof(xprt->slot[0]);
1417         xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
1418         if (xprt->slot == NULL) {
1419                 kfree(xprt);
1420                 return ERR_PTR(-ENOMEM);
1421         }
1422         memset(xprt->slot, 0, slot_table_size);
1423
1424         xprt->addr = *ap;
1425         xprt->prot = proto;
1426         xprt->stream = (proto == IPPROTO_TCP)? 1 : 0;
1427         if (xprt->stream) {
1428                 xprt->cwnd = RPC_MAXCWND(xprt);
1429                 xprt->nocong = 1;
1430         } else
1431                 xprt->cwnd = RPC_INITCWND;
1432         spin_lock_init(&xprt->sock_lock);
1433         spin_lock_init(&xprt->xprt_lock);
1434         init_waitqueue_head(&xprt->cong_wait);
1435
1436         INIT_LIST_HEAD(&xprt->free);
1437         INIT_LIST_HEAD(&xprt->recv);
1438         INIT_WORK(&xprt->sock_connect, xprt_socket_connect, xprt);
1439         INIT_WORK(&xprt->task_cleanup, xprt_socket_autoclose, xprt);
1440         init_timer(&xprt->timer);
1441         xprt->timer.function = xprt_init_autodisconnect;
1442         xprt->timer.data = (unsigned long) xprt;
1443         xprt->last_used = jiffies;
1444         xprt->port = XPRT_MAX_RESVPORT;
1445
1446         /* Set timeout parameters */
1447         if (to) {
1448                 xprt->timeout = *to;
1449                 xprt->timeout.to_current = to->to_initval;
1450         } else
1451                 xprt_default_timeout(&xprt->timeout, xprt->prot);
1452
1453         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1454         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
1455         rpc_init_wait_queue(&xprt->resend, "xprt_resend");
1456         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1457
1458         /* initialize free list */
1459         for (req = &xprt->slot[entries-1]; req >= &xprt->slot[0]; req--)
1460                 list_add(&req->rq_list, &xprt->free);
1461
1462         xprt_init_xid(xprt);
1463
1464         /* Check whether we want to use a reserved port */
1465         xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1466
1467         dprintk("RPC:      created transport %p with %u slots\n", xprt,
1468                         xprt->max_reqs);
1469
1470         return xprt;
1471 }
1472
1473 /*
1474  * Bind to a reserved port
1475  */
1476 static inline int xprt_bindresvport(struct rpc_xprt *xprt, struct socket *sock)
1477 {
1478         struct sockaddr_in myaddr = {
1479                 .sin_family = AF_INET,
1480         };
1481         int             err, port;
1482
1483         /* Were we already bound to a given port? Try to reuse it */
1484         port = xprt->port;
1485         do {
1486                 myaddr.sin_port = htons(port);
1487                 err = sock->ops->bind(sock, (struct sockaddr *) &myaddr,
1488                                                 sizeof(myaddr));
1489                 if (err == 0) {
1490                         xprt->port = port;
1491                         return 0;
1492                 }
1493                 if (--port == 0)
1494                         port = XPRT_MAX_RESVPORT;
1495         } while (err == -EADDRINUSE && port != xprt->port);
1496
1497         printk("RPC: Can't bind to reserved port (%d).\n", -err);
1498         return err;
1499 }
1500
1501 static void
1502 xprt_bind_socket(struct rpc_xprt *xprt, struct socket *sock)
1503 {
1504         struct sock     *sk = sock->sk;
1505
1506         if (xprt->inet)
1507                 return;
1508
1509         write_lock_bh(&sk->sk_callback_lock);
1510         sk->sk_user_data = xprt;
1511         xprt->old_data_ready = sk->sk_data_ready;
1512         xprt->old_state_change = sk->sk_state_change;
1513         xprt->old_write_space = sk->sk_write_space;
1514         if (xprt->prot == IPPROTO_UDP) {
1515                 sk->sk_data_ready = udp_data_ready;
1516                 sk->sk_no_check = UDP_CSUM_NORCV;
1517                 xprt_set_connected(xprt);
1518         } else {
1519                 struct tcp_opt *tp = tcp_sk(sk);
1520                 tp->nonagle = 1;        /* disable Nagle's algorithm */
1521                 sk->sk_data_ready = tcp_data_ready;
1522                 sk->sk_state_change = tcp_state_change;
1523                 xprt_clear_connected(xprt);
1524         }
1525         sk->sk_write_space = xprt_write_space;
1526
1527         /* Reset to new socket */
1528         xprt->sock = sock;
1529         xprt->inet = sk;
1530         write_unlock_bh(&sk->sk_callback_lock);
1531
1532         return;
1533 }
1534
1535 /*
1536  * Set socket buffer length
1537  */
1538 void
1539 xprt_sock_setbufsize(struct rpc_xprt *xprt)
1540 {
1541         struct sock *sk = xprt->inet;
1542
1543         if (xprt->stream)
1544                 return;
1545         if (xprt->rcvsize) {
1546                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
1547                 sk->sk_rcvbuf = xprt->rcvsize * xprt->max_reqs *  2;
1548         }
1549         if (xprt->sndsize) {
1550                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
1551                 sk->sk_sndbuf = xprt->sndsize * xprt->max_reqs * 2;
1552                 sk->sk_write_space(sk);
1553         }
1554 }
1555
1556 /*
1557  * Datastream sockets are created here, but xprt_connect will create
1558  * and connect stream sockets.
1559  */
1560 static struct socket * xprt_create_socket(struct rpc_xprt *xprt, int proto, int resvport)
1561 {
1562         struct socket   *sock;
1563         int             type, err;
1564
1565         dprintk("RPC:      xprt_create_socket(%s %d)\n",
1566                            (proto == IPPROTO_UDP)? "udp" : "tcp", proto);
1567
1568         type = (proto == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1569
1570         if ((err = sock_create_kern(PF_INET, type, proto, &sock)) < 0) {
1571                 printk("RPC: can't create socket (%d).\n", -err);
1572                 return NULL;
1573         }
1574
1575         /* If the caller has the capability, bind to a reserved port */
1576         if (resvport && xprt_bindresvport(xprt, sock) < 0) {
1577                 printk("RPC: can't bind to reserved port.\n");
1578                 goto failed;
1579         }
1580
1581         return sock;
1582
1583 failed:
1584         sock_release(sock);
1585         return NULL;
1586 }
1587
1588 /*
1589  * Create an RPC client transport given the protocol and peer address.
1590  */
1591 struct rpc_xprt *
1592 xprt_create_proto(int proto, struct sockaddr_in *sap, struct rpc_timeout *to)
1593 {
1594         struct rpc_xprt *xprt;
1595
1596         xprt = xprt_setup(proto, sap, to);
1597         if (IS_ERR(xprt))
1598                 dprintk("RPC:      xprt_create_proto failed\n");
1599         else
1600                 dprintk("RPC:      xprt_create_proto created xprt %p\n", xprt);
1601         return xprt;
1602 }
1603
1604 /*
1605  * Prepare for transport shutdown.
1606  */
1607 void
1608 xprt_shutdown(struct rpc_xprt *xprt)
1609 {
1610         xprt->shutdown = 1;
1611         rpc_wake_up(&xprt->sending);
1612         rpc_wake_up(&xprt->resend);
1613         rpc_wake_up(&xprt->pending);
1614         rpc_wake_up(&xprt->backlog);
1615         wake_up(&xprt->cong_wait);
1616         del_timer_sync(&xprt->timer);
1617 }
1618
1619 /*
1620  * Clear the xprt backlog queue
1621  */
1622 int
1623 xprt_clear_backlog(struct rpc_xprt *xprt) {
1624         rpc_wake_up_next(&xprt->backlog);
1625         wake_up(&xprt->cong_wait);
1626         return 1;
1627 }
1628
1629 /*
1630  * Destroy an RPC transport, killing off all requests.
1631  */
1632 int
1633 xprt_destroy(struct rpc_xprt *xprt)
1634 {
1635         dprintk("RPC:      destroying transport %p\n", xprt);
1636         xprt_shutdown(xprt);
1637         xprt_disconnect(xprt);
1638         xprt_close(xprt);
1639         kfree(xprt->slot);
1640         kfree(xprt);
1641
1642         return 0;
1643 }