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_transmit().
  14  *  -   xprt_transmit sends the message and installs the caller on the
  15  *      transport'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 transport. 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  *  Transport switch API copyright (C) 2005, Chuck Lever <cel@netapp.com>
  37  */
  38
  39 #include <linux/module.h>
  40
  41 #include <linux/types.h>
  42 #include <linux/interrupt.h>
  43 #include <linux/workqueue.h>
  44 #include <linux/random.h>
  45
  46 #include <linux/sunrpc/clnt.h>
  47
  48 /*
  49  * Local variables
  50  */
  51
  52 #ifdef RPC_DEBUG
  53 # undef  RPC_DEBUG_DATA
  54 # define RPCDBG_FACILITY        RPCDBG_XPRT
  55 #endif
  56
  57 /*
  58  * Local functions
  59  */
  60 static void     xprt_request_init(struct rpc_task *, struct rpc_xprt *);
  61 static inline void      do_xprt_reserve(struct rpc_task *);
  62 static void     xprt_connect_status(struct rpc_task *task);
  63 static int      __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
  64
  65 static int      xprt_clear_backlog(struct rpc_xprt *xprt);
  66
  67 /**
  68  * xprt_reserve_xprt - serialize write access to transports
  69  * @task: task that is requesting access to the transport
  70  *
  71  * This prevents mixing the payload of separate requests, and prevents
  72  * transport connects from colliding with writes.  No congestion control
  73  * is provided.
  74  */
  75 int xprt_reserve_xprt(struct rpc_task *task)
  76 {
  77         struct rpc_xprt *xprt = task->tk_xprt;
  78         struct rpc_rqst *req = task->tk_rqstp;
  79
  80         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
  81                 if (task == xprt->snd_task)
  82                         return 1;
  83                 if (task == NULL)
  84                         return 0;
  85                 goto out_sleep;
  86         }
  87         xprt->snd_task = task;
  88         if (req) {
  89                 req->rq_bytes_sent = 0;
  90                 req->rq_ntrans++;
  91         }
  92         return 1;
  93
  94 out_sleep:
  95         dprintk("RPC: %4d failed to lock transport %p\n",
  96                         task->tk_pid, xprt);
  97         task->tk_timeout = 0;
  98         task->tk_status = -EAGAIN;
  99         if (req && req->rq_ntrans)
 100                 rpc_sleep_on(&xprt->resend, task, NULL, NULL);
 101         else
 102                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
 103         return 0;
 104 }
 105
 106 /*
 107  * xprt_reserve_xprt_cong - serialize write access to transports
 108  * @task: task that is requesting access to the transport
 109  *
 110  * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
 111  * integrated into the decision of whether a request is allowed to be
 112  * woken up and given access to the transport.
 113  */
 114 int xprt_reserve_xprt_cong(struct rpc_task *task)
 115 {
 116         struct rpc_xprt *xprt = task->tk_xprt;
 117         struct rpc_rqst *req = task->tk_rqstp;
 118
 119         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 120                 if (task == xprt->snd_task)
 121                         return 1;
 122                 goto out_sleep;
 123         }
 124         if (__xprt_get_cong(xprt, task)) {
 125                 xprt->snd_task = task;
 126                 if (req) {
 127                         req->rq_bytes_sent = 0;
 128                         req->rq_ntrans++;
 129                 }
 130                 return 1;
 131         }
 132         smp_mb__before_clear_bit();
 133         clear_bit(XPRT_LOCKED, &xprt->state);
 134         smp_mb__after_clear_bit();
 135 out_sleep:
 136         dprintk("RPC: %4d failed to lock transport %p\n", task->tk_pid, xprt);
 137         task->tk_timeout = 0;
 138         task->tk_status = -EAGAIN;
 139         if (req && req->rq_ntrans)
 140                 rpc_sleep_on(&xprt->resend, task, NULL, NULL);
 141         else
 142                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
 143         return 0;
 144 }
 145
 146 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 147 {
 148         int retval;
 149
 150         spin_lock_bh(&xprt->transport_lock);
 151         retval = xprt->ops->reserve_xprt(task);
 152         spin_unlock_bh(&xprt->transport_lock);
 153         return retval;
 154 }
 155
 156 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
 157 {
 158         struct rpc_task *task;
 159         struct rpc_rqst *req;
 160
 161         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 162                 return;
 163
 164         task = rpc_wake_up_next(&xprt->resend);
 165         if (!task) {
 166                 task = rpc_wake_up_next(&xprt->sending);
 167                 if (!task)
 168                         goto out_unlock;
 169         }
 170
 171         req = task->tk_rqstp;
 172         xprt->snd_task = task;
 173         if (req) {
 174                 req->rq_bytes_sent = 0;
 175                 req->rq_ntrans++;
 176         }
 177         return;
 178
 179 out_unlock:
 180         smp_mb__before_clear_bit();
 181         clear_bit(XPRT_LOCKED, &xprt->state);
 182         smp_mb__after_clear_bit();
 183 }
 184
 185 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
 186 {
 187         struct rpc_task *task;
 188
 189         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 190                 return;
 191         if (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_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->state);
 211         smp_mb__after_clear_bit();
 212 }
 213
 214 /**
 215  * xprt_release_xprt - allow other requests to use a transport
 216  * @xprt: transport with other tasks potentially waiting
 217  * @task: task that is releasing access to the transport
 218  *
 219  * Note that "task" can be NULL.  No congestion control is provided.
 220  */
 221 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 222 {
 223         if (xprt->snd_task == task) {
 224                 xprt->snd_task = NULL;
 225                 smp_mb__before_clear_bit();
 226                 clear_bit(XPRT_LOCKED, &xprt->state);
 227                 smp_mb__after_clear_bit();
 228                 __xprt_lock_write_next(xprt);
 229         }
 230 }
 231
 232 /**
 233  * xprt_release_xprt_cong - allow other requests to use a transport
 234  * @xprt: transport with other tasks potentially waiting
 235  * @task: task that is releasing access to the transport
 236  *
 237  * Note that "task" can be NULL.  Another task is awoken to use the
 238  * transport if the transport's congestion window allows it.
 239  */
 240 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 241 {
 242         if (xprt->snd_task == task) {
 243                 xprt->snd_task = NULL;
 244                 smp_mb__before_clear_bit();
 245                 clear_bit(XPRT_LOCKED, &xprt->state);
 246                 smp_mb__after_clear_bit();
 247                 __xprt_lock_write_next_cong(xprt);
 248         }
 249 }
 250
 251 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 252 {
 253         spin_lock_bh(&xprt->transport_lock);
 254         xprt->ops->release_xprt(xprt, task);
 255         spin_unlock_bh(&xprt->transport_lock);
 256 }
 257
 258 /*
 259  * Van Jacobson congestion avoidance. Check if the congestion window
 260  * overflowed. Put the task to sleep if this is the case.
 261  */
 262 static int
 263 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 264 {
 265         struct rpc_rqst *req = task->tk_rqstp;
 266
 267         if (req->rq_cong)
 268                 return 1;
 269         dprintk("RPC: %4d xprt_cwnd_limited cong = %ld cwnd = %ld\n",
 270                         task->tk_pid, xprt->cong, xprt->cwnd);
 271         if (RPCXPRT_CONGESTED(xprt))
 272                 return 0;
 273         req->rq_cong = 1;
 274         xprt->cong += RPC_CWNDSCALE;
 275         return 1;
 276 }
 277
 278 /*
 279  * Adjust the congestion window, and wake up the next task
 280  * that has been sleeping due to congestion
 281  */
 282 static void
 283 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 284 {
 285         if (!req->rq_cong)
 286                 return;
 287         req->rq_cong = 0;
 288         xprt->cong -= RPC_CWNDSCALE;
 289         __xprt_lock_write_next_cong(xprt);
 290 }
 291
 292 /**
 293  * xprt_adjust_cwnd - adjust transport congestion window
 294  * @task: recently completed RPC request used to adjust window
 295  * @result: result code of completed RPC request
 296  *
 297  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
 298  */
 299 void xprt_adjust_cwnd(struct rpc_task *task, int result)
 300 {
 301         struct rpc_rqst *req = task->tk_rqstp;
 302         struct rpc_xprt *xprt = task->tk_xprt;
 303         unsigned long cwnd = xprt->cwnd;
 304
 305         if (result >= 0 && cwnd <= xprt->cong) {
 306                 /* The (cwnd >> 1) term makes sure
 307                  * the result gets rounded properly. */
 308                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 309                 if (cwnd > RPC_MAXCWND(xprt))
 310                         cwnd = RPC_MAXCWND(xprt);
 311                 __xprt_lock_write_next_cong(xprt);
 312         } else if (result == -ETIMEDOUT) {
 313                 cwnd >>= 1;
 314                 if (cwnd < RPC_CWNDSCALE)
 315                         cwnd = RPC_CWNDSCALE;
 316         }
 317         dprintk("RPC:      cong %ld, cwnd was %ld, now %ld\n",
 318                         xprt->cong, xprt->cwnd, cwnd);
 319         xprt->cwnd = cwnd;
 320         __xprt_put_cong(xprt, req);
 321 }
 322
 323 /**
 324  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
 325  * @xprt: transport with waiting tasks
 326  * @status: result code to plant in each task before waking it
 327  *
 328  */
 329 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
 330 {
 331         if (status < 0)
 332                 rpc_wake_up_status(&xprt->pending, status);
 333         else
 334                 rpc_wake_up(&xprt->pending);
 335 }
 336
 337 /**
 338  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
 339  * @task: task to be put to sleep
 340  *
 341  */
 342 void xprt_wait_for_buffer_space(struct rpc_task *task)
 343 {
 344         struct rpc_rqst *req = task->tk_rqstp;
 345         struct rpc_xprt *xprt = req->rq_xprt;
 346
 347         task->tk_timeout = req->rq_timeout;
 348         rpc_sleep_on(&xprt->pending, task, NULL, NULL);
 349 }
 350
 351 /**
 352  * xprt_write_space - wake the task waiting for transport output buffer space
 353  * @xprt: transport with waiting tasks
 354  *
 355  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
 356  */
 357 void xprt_write_space(struct rpc_xprt *xprt)
 358 {
 359         if (unlikely(xprt->shutdown))
 360                 return;
 361
 362         spin_lock_bh(&xprt->transport_lock);
 363         if (xprt->snd_task) {
 364                 dprintk("RPC:      write space: waking waiting task on xprt %p\n",
 365                                 xprt);
 366                 rpc_wake_up_task(xprt->snd_task);
 367         }
 368         spin_unlock_bh(&xprt->transport_lock);
 369 }
 370
 371 /**
 372  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
 373  * @task: task whose timeout is to be set
 374  *
 375  * Set a request's retransmit timeout based on the transport's
 376  * default timeout parameters.  Used by transports that don't adjust
 377  * the retransmit timeout based on round-trip time estimation.
 378  */
 379 void xprt_set_retrans_timeout_def(struct rpc_task *task)
 380 {
 381         task->tk_timeout = task->tk_rqstp->rq_timeout;
 382 }
 383
 384 /*
 385  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
 386  * @task: task whose timeout is to be set
 387  *
 388  * Set a request's retransmit timeout using the RTT estimator.
 389  */
 390 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
 391 {
 392         int timer = task->tk_msg.rpc_proc->p_timer;
 393         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
 394         struct rpc_rqst *req = task->tk_rqstp;
 395         unsigned long max_timeout = req->rq_xprt->timeout.to_maxval;
 396
 397         task->tk_timeout = rpc_calc_rto(rtt, timer);
 398         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
 399         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
 400                 task->tk_timeout = max_timeout;
 401 }
 402
 403 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 404 {
 405         struct rpc_timeout *to = &req->rq_xprt->timeout;
 406
 407         req->rq_majortimeo = req->rq_timeout;
 408         if (to->to_exponential)
 409                 req->rq_majortimeo <<= to->to_retries;
 410         else
 411                 req->rq_majortimeo += to->to_increment * to->to_retries;
 412         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 413                 req->rq_majortimeo = to->to_maxval;
 414         req->rq_majortimeo += jiffies;
 415 }
 416
 417 /**
 418  * xprt_adjust_timeout - adjust timeout values for next retransmit
 419  * @req: RPC request containing parameters to use for the adjustment
 420  *
 421  */
 422 int xprt_adjust_timeout(struct rpc_rqst *req)
 423 {
 424         struct rpc_xprt *xprt = req->rq_xprt;
 425         struct rpc_timeout *to = &xprt->timeout;
 426         int status = 0;
 427
 428         if (time_before(jiffies, req->rq_majortimeo)) {
 429                 if (to->to_exponential)
 430                         req->rq_timeout <<= 1;
 431                 else
 432                         req->rq_timeout += to->to_increment;
 433                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 434                         req->rq_timeout = to->to_maxval;
 435                 req->rq_retries++;
 436                 pprintk("RPC: %lu retrans\n", jiffies);
 437         } else {
 438                 req->rq_timeout = to->to_initval;
 439                 req->rq_retries = 0;
 440                 xprt_reset_majortimeo(req);
 441                 /* Reset the RTT counters == "slow start" */
 442                 spin_lock_bh(&xprt->transport_lock);
 443                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 444                 spin_unlock_bh(&xprt->transport_lock);
 445                 pprintk("RPC: %lu timeout\n", jiffies);
 446                 status = -ETIMEDOUT;
 447         }
 448
 449         if (req->rq_timeout == 0) {
 450                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 451                 req->rq_timeout = 5 * HZ;
 452         }
 453         return status;
 454 }
 455
 456 static void xprt_autoclose(void *args)
 457 {
 458         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
 459
 460         xprt_disconnect(xprt);
 461         xprt->ops->close(xprt);
 462         xprt_release_write(xprt, NULL);
 463 }
 464
 465 /**
 466  * xprt_disconnect - mark a transport as disconnected
 467  * @xprt: transport to flag for disconnect
 468  *
 469  */
 470 void xprt_disconnect(struct rpc_xprt *xprt)
 471 {
 472         dprintk("RPC:      disconnected transport %p\n", xprt);
 473         spin_lock_bh(&xprt->transport_lock);
 474         xprt_clear_connected(xprt);
 475         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 476         spin_unlock_bh(&xprt->transport_lock);
 477 }
 478
 479 static void
 480 xprt_init_autodisconnect(unsigned long data)
 481 {
 482         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 483
 484         spin_lock(&xprt->transport_lock);
 485         if (!list_empty(&xprt->recv) || xprt->shutdown)
 486                 goto out_abort;
 487         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 488                 goto out_abort;
 489         spin_unlock(&xprt->transport_lock);
 490         if (xprt_connecting(xprt))
 491                 xprt_release_write(xprt, NULL);
 492         else
 493                 schedule_work(&xprt->task_cleanup);
 494         return;
 495 out_abort:
 496         spin_unlock(&xprt->transport_lock);
 497 }
 498
 499 /**
 500  * xprt_connect - schedule a transport connect operation
 501  * @task: RPC task that is requesting the connect
 502  *
 503  */
 504 void xprt_connect(struct rpc_task *task)
 505 {
 506         struct rpc_xprt *xprt = task->tk_xprt;
 507
 508         dprintk("RPC: %4d xprt_connect xprt %p %s connected\n", task->tk_pid,
 509                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 510
 511         if (xprt->shutdown) {
 512                 task->tk_status = -EIO;
 513                 return;
 514         }
 515         if (!xprt->addr.sin_port) {
 516                 task->tk_status = -EIO;
 517                 return;
 518         }
 519         if (!xprt_lock_write(xprt, task))
 520                 return;
 521         if (xprt_connected(xprt))
 522                 xprt_release_write(xprt, task);
 523         else {
 524                 if (task->tk_rqstp)
 525                         task->tk_rqstp->rq_bytes_sent = 0;
 526
 527                 task->tk_timeout = RPC_CONNECT_TIMEOUT;
 528                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
 529                 xprt->ops->connect(task);
 530         }
 531         return;
 532 }
 533
 534 static void xprt_connect_status(struct rpc_task *task)
 535 {
 536         struct rpc_xprt *xprt = task->tk_xprt;
 537
 538         if (task->tk_status >= 0) {
 539                 dprintk("RPC: %4d xprt_connect_status: connection established\n",
 540                                 task->tk_pid);
 541                 return;
 542         }
 543
 544         switch (task->tk_status) {
 545         case -ECONNREFUSED:
 546         case -ECONNRESET:
 547                 dprintk("RPC: %4d xprt_connect_status: server %s refused connection\n",
 548                                 task->tk_pid, task->tk_client->cl_server);
 549                 break;
 550         case -ENOTCONN:
 551                 dprintk("RPC: %4d xprt_connect_status: connection broken\n",
 552                                 task->tk_pid);
 553                 break;
 554         case -ETIMEDOUT:
 555                 dprintk("RPC: %4d xprt_connect_status: connect attempt timed out\n",
 556                                 task->tk_pid);
 557                 break;
 558         default:
 559                 dprintk("RPC: %4d xprt_connect_status: error %d connecting to server %s\n",
 560                                 task->tk_pid, -task->tk_status, task->tk_client->cl_server);
 561                 xprt_release_write(xprt, task);
 562                 task->tk_status = -EIO;
 563                 return;
 564         }
 565
 566         /* if soft mounted, just cause this RPC to fail */
 567         if (RPC_IS_SOFT(task)) {
 568                 xprt_release_write(xprt, task);
 569                 task->tk_status = -EIO;
 570         }
 571 }
 572
 573 /**
 574  * xprt_lookup_rqst - find an RPC request corresponding to an XID
 575  * @xprt: transport on which the original request was transmitted
 576  * @xid: RPC XID of incoming reply
 577  *
 578  */
 579 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
 580 {
 581         struct list_head *pos;
 582         struct rpc_rqst *req = NULL;
 583
 584         list_for_each(pos, &xprt->recv) {
 585                 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
 586                 if (entry->rq_xid == xid) {
 587                         req = entry;
 588                         break;
 589                 }
 590         }
 591         return req;
 592 }
 593
 594 /**
 595  * xprt_complete_rqst - called when reply processing is complete
 596  * @xprt: controlling transport
 597  * @req: RPC request that just completed
 598  * @copied: actual number of bytes received from the transport
 599  *
 600  */
 601 void xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
 602 {
 603         struct rpc_task *task = req->rq_task;
 604         struct rpc_clnt *clnt = task->tk_client;
 605
 606         /* Adjust congestion window */
 607         if (!xprt->nocong) {
 608                 unsigned timer = task->tk_msg.rpc_proc->p_timer;
 609                 xprt_adjust_cwnd(task, copied);
 610                 if (timer) {
 611                         if (req->rq_ntrans == 1)
 612                                 rpc_update_rtt(clnt->cl_rtt, timer,
 613                                                 (long)jiffies - req->rq_xtime);
 614                         rpc_set_timeo(clnt->cl_rtt, timer, req->rq_ntrans - 1);
 615                 }
 616         }
 617
 618 #ifdef RPC_PROFILE
 619         /* Profile only reads for now */
 620         if (copied > 1024) {
 621                 static unsigned long    nextstat;
 622                 static unsigned long    pkt_rtt, pkt_len, pkt_cnt;
 623
 624                 pkt_cnt++;
 625                 pkt_len += req->rq_slen + copied;
 626                 pkt_rtt += jiffies - req->rq_xtime;
 627                 if (time_before(nextstat, jiffies)) {
 628                         printk("RPC: %lu %ld cwnd\n", jiffies, xprt->cwnd);
 629                         printk("RPC: %ld %ld %ld %ld stat\n",
 630                                         jiffies, pkt_cnt, pkt_len, pkt_rtt);
 631                         pkt_rtt = pkt_len = pkt_cnt = 0;
 632                         nextstat = jiffies + 5 * HZ;
 633                 }
 634         }
 635 #endif
 636
 637         dprintk("RPC: %4d has input (%d bytes)\n", task->tk_pid, copied);
 638         list_del_init(&req->rq_list);
 639         req->rq_received = req->rq_private_buf.len = copied;
 640
 641         /* ... and wake up the process. */
 642         rpc_wake_up_task(task);
 643         return;
 644 }
 645
 646 static void xprt_timer(struct rpc_task *task)
 647 {
 648         struct rpc_rqst *req = task->tk_rqstp;
 649         struct rpc_xprt *xprt = req->rq_xprt;
 650
 651         dprintk("RPC: %4d xprt_timer\n", task->tk_pid);
 652
 653         spin_lock(&xprt->transport_lock);
 654         if (!req->rq_received) {
 655                 if (xprt->ops->timer)
 656                         xprt->ops->timer(task);
 657                 task->tk_status = -ETIMEDOUT;
 658         }
 659         task->tk_timeout = 0;
 660         rpc_wake_up_task(task);
 661         spin_unlock(&xprt->transport_lock);
 662 }
 663
 664 /**
 665  * xprt_prepare_transmit - reserve the transport before sending a request
 666  * @task: RPC task about to send a request
 667  *
 668  */
 669 int xprt_prepare_transmit(struct rpc_task *task)
 670 {
 671         struct rpc_rqst *req = task->tk_rqstp;
 672         struct rpc_xprt *xprt = req->rq_xprt;
 673         int err = 0;
 674
 675         dprintk("RPC: %4d xprt_prepare_transmit\n", task->tk_pid);
 676
 677         if (xprt->shutdown)
 678                 return -EIO;
 679
 680         spin_lock_bh(&xprt->transport_lock);
 681         if (req->rq_received && !req->rq_bytes_sent) {
 682                 err = req->rq_received;
 683                 goto out_unlock;
 684         }
 685         if (!xprt->ops->reserve_xprt(task)) {
 686                 err = -EAGAIN;
 687                 goto out_unlock;
 688         }
 689
 690         if (!xprt_connected(xprt)) {
 691                 err = -ENOTCONN;
 692                 goto out_unlock;
 693         }
 694 out_unlock:
 695         spin_unlock_bh(&xprt->transport_lock);
 696         return err;
 697 }
 698
 699 /**
 700  * xprt_transmit - send an RPC request on a transport
 701  * @task: controlling RPC task
 702  *
 703  * We have to copy the iovec because sendmsg fiddles with its contents.
 704  */
 705 void xprt_transmit(struct rpc_task *task)
 706 {
 707         struct rpc_rqst *req = task->tk_rqstp;
 708         struct rpc_xprt *xprt = req->rq_xprt;
 709         int status;
 710
 711         dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
 712
 713         smp_rmb();
 714         if (!req->rq_received) {
 715                 if (list_empty(&req->rq_list)) {
 716                         spin_lock_bh(&xprt->transport_lock);
 717                         /* Update the softirq receive buffer */
 718                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
 719                                         sizeof(req->rq_private_buf));
 720                         /* Add request to the receive list */
 721                         list_add_tail(&req->rq_list, &xprt->recv);
 722                         spin_unlock_bh(&xprt->transport_lock);
 723                         xprt_reset_majortimeo(req);
 724                         /* Turn off autodisconnect */
 725                         del_singleshot_timer_sync(&xprt->timer);
 726                 }
 727         } else if (!req->rq_bytes_sent)
 728                 return;
 729
 730         status = xprt->ops->send_request(task);
 731         if (status == 0) {
 732                 dprintk("RPC: %4d xmit complete\n", task->tk_pid);
 733                 spin_lock_bh(&xprt->transport_lock);
 734                 xprt->ops->set_retrans_timeout(task);
 735                 /* Don't race with disconnect */
 736                 if (!xprt_connected(xprt))
 737                         task->tk_status = -ENOTCONN;
 738                 else if (!req->rq_received)
 739                         rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
 740                 xprt->ops->release_xprt(xprt, task);
 741                 spin_unlock_bh(&xprt->transport_lock);
 742                 return;
 743         }
 744
 745         /* Note: at this point, task->tk_sleeping has not yet been set,
 746          *       hence there is no danger of the waking up task being put on
 747          *       schedq, and being picked up by a parallel run of rpciod().
 748          */
 749         task->tk_status = status;
 750
 751         switch (status) {
 752         case -ECONNREFUSED:
 753                 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
 754                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
 755         case -EAGAIN:
 756         case -ENOTCONN:
 757                 return;
 758         default:
 759                 break;
 760         }
 761         xprt_release_write(xprt, task);
 762         return;
 763 }
 764
 765 static inline void do_xprt_reserve(struct rpc_task *task)
 766 {
 767         struct rpc_xprt *xprt = task->tk_xprt;
 768
 769         task->tk_status = 0;
 770         if (task->tk_rqstp)
 771                 return;
 772         if (!list_empty(&xprt->free)) {
 773                 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
 774                 list_del_init(&req->rq_list);
 775                 task->tk_rqstp = req;
 776                 xprt_request_init(task, xprt);
 777                 return;
 778         }
 779         dprintk("RPC:      waiting for request slot\n");
 780         task->tk_status = -EAGAIN;
 781         task->tk_timeout = 0;
 782         rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
 783 }
 784
 785 /**
 786  * xprt_reserve - allocate an RPC request slot
 787  * @task: RPC task requesting a slot allocation
 788  *
 789  * If no more slots are available, place the task on the transport's
 790  * backlog queue.
 791  */
 792 void xprt_reserve(struct rpc_task *task)
 793 {
 794         struct rpc_xprt *xprt = task->tk_xprt;
 795
 796         task->tk_status = -EIO;
 797         if (!xprt->shutdown) {
 798                 spin_lock(&xprt->reserve_lock);
 799                 do_xprt_reserve(task);
 800                 spin_unlock(&xprt->reserve_lock);
 801         }
 802 }
 803
 804 static inline u32 xprt_alloc_xid(struct rpc_xprt *xprt)
 805 {
 806         return xprt->xid++;
 807 }
 808
 809 static inline void xprt_init_xid(struct rpc_xprt *xprt)
 810 {
 811         get_random_bytes(&xprt->xid, sizeof(xprt->xid));
 812 }
 813
 814 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
 815 {
 816         struct rpc_rqst *req = task->tk_rqstp;
 817
 818         req->rq_timeout = xprt->timeout.to_initval;
 819         req->rq_task    = task;
 820         req->rq_xprt    = xprt;
 821         req->rq_xid     = xprt_alloc_xid(xprt);
 822         dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
 823                         req, ntohl(req->rq_xid));
 824 }
 825
 826 /**
 827  * xprt_release - release an RPC request slot
 828  * @task: task which is finished with the slot
 829  *
 830  */
 831 void xprt_release(struct rpc_task *task)
 832 {
 833         struct rpc_xprt *xprt = task->tk_xprt;
 834         struct rpc_rqst *req;
 835
 836         if (!(req = task->tk_rqstp))
 837                 return;
 838         spin_lock_bh(&xprt->transport_lock);
 839         xprt->ops->release_xprt(xprt, task);
 840         __xprt_put_cong(xprt, req);
 841         if (!list_empty(&req->rq_list))
 842                 list_del(&req->rq_list);
 843         xprt->last_used = jiffies;
 844         if (list_empty(&xprt->recv) && !xprt->shutdown)
 845                 mod_timer(&xprt->timer,
 846                                 xprt->last_used + RPC_IDLE_DISCONNECT_TIMEOUT);
 847         spin_unlock_bh(&xprt->transport_lock);
 848         task->tk_rqstp = NULL;
 849         memset(req, 0, sizeof(*req));   /* mark unused */
 850
 851         dprintk("RPC: %4d release request %p\n", task->tk_pid, req);
 852
 853         spin_lock(&xprt->reserve_lock);
 854         list_add(&req->rq_list, &xprt->free);
 855         xprt_clear_backlog(xprt);
 856         spin_unlock(&xprt->reserve_lock);
 857 }
 858
 859 /**
 860  * xprt_set_timeout - set constant RPC timeout
 861  * @to: RPC timeout parameters to set up
 862  * @retr: number of retries
 863  * @incr: amount of increase after each retry
 864  *
 865  */
 866 void xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
 867 {
 868         to->to_initval   =
 869         to->to_increment = incr;
 870         to->to_maxval    = to->to_initval + (incr * retr);
 871         to->to_retries   = retr;
 872         to->to_exponential = 0;
 873 }
 874
 875 static struct rpc_xprt *xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
 876 {
 877         int result;
 878         struct rpc_xprt *xprt;
 879         struct rpc_rqst *req;
 880
 881         if ((xprt = kmalloc(sizeof(struct rpc_xprt), GFP_KERNEL)) == NULL)
 882                 return ERR_PTR(-ENOMEM);
 883         memset(xprt, 0, sizeof(*xprt)); /* Nnnngh! */
 884
 885         xprt->addr = *ap;
 886
 887         switch (proto) {
 888         case IPPROTO_UDP:
 889                 result = xs_setup_udp(xprt, to);
 890                 break;
 891         case IPPROTO_TCP:
 892                 result = xs_setup_tcp(xprt, to);
 893                 break;
 894         default:
 895                 printk(KERN_ERR "RPC: unrecognized transport protocol: %d\n",
 896                                 proto);
 897                 result = -EIO;
 898                 break;
 899         }
 900         if (result) {
 901                 kfree(xprt);
 902                 return ERR_PTR(result);
 903         }
 904
 905         spin_lock_init(&xprt->transport_lock);
 906         spin_lock_init(&xprt->reserve_lock);
 907         init_waitqueue_head(&xprt->cong_wait);
 908
 909         INIT_LIST_HEAD(&xprt->free);
 910         INIT_LIST_HEAD(&xprt->recv);
 911         INIT_WORK(&xprt->task_cleanup, xprt_autoclose, xprt);
 912         init_timer(&xprt->timer);
 913         xprt->timer.function = xprt_init_autodisconnect;
 914         xprt->timer.data = (unsigned long) xprt;
 915         xprt->last_used = jiffies;
 916
 917         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
 918         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
 919         rpc_init_wait_queue(&xprt->resend, "xprt_resend");
 920         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
 921
 922         /* initialize free list */
 923         for (req = &xprt->slot[xprt->max_reqs-1]; req >= &xprt->slot[0]; req--)
 924                 list_add(&req->rq_list, &xprt->free);
 925
 926         xprt_init_xid(xprt);
 927
 928         dprintk("RPC:      created transport %p with %u slots\n", xprt,
 929                         xprt->max_reqs);
 930
 931         return xprt;
 932 }
 933
 934 /**
 935  * xprt_create_proto - create an RPC client transport
 936  * @proto: requested transport protocol
 937  * @sap: remote peer's address
 938  * @to: timeout parameters for new transport
 939  *
 940  */
 941 struct rpc_xprt *xprt_create_proto(int proto, struct sockaddr_in *sap, struct rpc_timeout *to)
 942 {
 943         struct rpc_xprt *xprt;
 944
 945         xprt = xprt_setup(proto, sap, to);
 946         if (IS_ERR(xprt))
 947                 dprintk("RPC:      xprt_create_proto failed\n");
 948         else
 949                 dprintk("RPC:      xprt_create_proto created xprt %p\n", xprt);
 950         return xprt;
 951 }
 952
 953 static void xprt_shutdown(struct rpc_xprt *xprt)
 954 {
 955         xprt->shutdown = 1;
 956         rpc_wake_up(&xprt->sending);
 957         rpc_wake_up(&xprt->resend);
 958         xprt_wake_pending_tasks(xprt, -EIO);
 959         rpc_wake_up(&xprt->backlog);
 960         wake_up(&xprt->cong_wait);
 961         del_timer_sync(&xprt->timer);
 962 }
 963
 964 static int xprt_clear_backlog(struct rpc_xprt *xprt) {
 965         rpc_wake_up_next(&xprt->backlog);
 966         wake_up(&xprt->cong_wait);
 967         return 1;
 968 }
 969
 970 /**
 971  * xprt_destroy - destroy an RPC transport, killing off all requests.
 972  * @xprt: transport to destroy
 973  *
 974  */
 975 int xprt_destroy(struct rpc_xprt *xprt)
 976 {
 977         dprintk("RPC:      destroying transport %p\n", xprt);
 978         xprt_shutdown(xprt);
 979         xprt->ops->destroy(xprt);
 980         kfree(xprt);
 981
 982         return 0;
 983 }