git.oblomov.eu Git - linux-2.6/blob - net/sunrpc/xprt.c

   1 /*
   2  *  linux/net/sunrpc/xprt.c
   3  *
   4  *  This is a generic RPC call interface supporting congestion avoidance,
   5  *  and asynchronous calls.
   6  *
   7  *  The interface works like this:
   8  *
   9  *  -   When a process places a call, it allocates a request slot if
  10  *      one is available. Otherwise, it sleeps on the backlog queue
  11  *      (xprt_reserve).
  12  *  -   Next, the caller puts together the RPC message, stuffs it into
  13  *      the request struct, and calls xprt_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/net.h>
  45
  46 #include <linux/sunrpc/clnt.h>
  47 #include <linux/sunrpc/metrics.h>
  48
  49 /*
  50  * Local variables
  51  */
  52
  53 #ifdef RPC_DEBUG
  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 spinlock_t xprt_list_lock = SPIN_LOCK_UNLOCKED;
  66 static LIST_HEAD(xprt_list);
  67
  68 /*
  69  * The transport code maintains an estimate on the maximum number of out-
  70  * standing RPC requests, using a smoothed version of the congestion
  71  * avoidance implemented in 44BSD. This is basically the Van Jacobson
  72  * congestion algorithm: If a retransmit occurs, the congestion window is
  73  * halved; otherwise, it is incremented by 1/cwnd when
  74  *
  75  *      -       a reply is received and
  76  *      -       a full number of requests are outstanding and
  77  *      -       the congestion window hasn't been updated recently.
  78  */
  79 #define RPC_CWNDSHIFT           (8U)
  80 #define RPC_CWNDSCALE           (1U << RPC_CWNDSHIFT)
  81 #define RPC_INITCWND            RPC_CWNDSCALE
  82 #define RPC_MAXCWND(xprt)       ((xprt)->max_reqs << RPC_CWNDSHIFT)
  83
  84 #define RPCXPRT_CONGESTED(xprt) ((xprt)->cong >= (xprt)->cwnd)
  85
  86 /**
  87  * xprt_register_transport - register a transport implementation
  88  * @transport: transport to register
  89  *
  90  * If a transport implementation is loaded as a kernel module, it can
  91  * call this interface to make itself known to the RPC client.
  92  *
  93  * Returns:
  94  * 0:           transport successfully registered
  95  * -EEXIST:     transport already registered
  96  * -EINVAL:     transport module being unloaded
  97  */
  98 int xprt_register_transport(struct xprt_class *transport)
  99 {
 100         struct xprt_class *t;
 101         int result;
 102
 103         result = -EEXIST;
 104         spin_lock(&xprt_list_lock);
 105         list_for_each_entry(t, &xprt_list, list) {
 106                 /* don't register the same transport class twice */
 107                 if (t == transport)
 108                         goto out;
 109         }
 110
 111         result = -EINVAL;
 112         if (try_module_get(THIS_MODULE)) {
 113                 list_add_tail(&transport->list, &xprt_list);
 114                 printk(KERN_INFO "RPC: Registered %s transport module.\n",
 115                         transport->name);
 116                 result = 0;
 117         }
 118
 119 out:
 120         spin_unlock(&xprt_list_lock);
 121         return result;
 122 }
 123 EXPORT_SYMBOL_GPL(xprt_register_transport);
 124
 125 /**
 126  * xprt_unregister_transport - unregister a transport implementation
 127  * transport: transport to unregister
 128  *
 129  * Returns:
 130  * 0:           transport successfully unregistered
 131  * -ENOENT:     transport never registered
 132  */
 133 int xprt_unregister_transport(struct xprt_class *transport)
 134 {
 135         struct xprt_class *t;
 136         int result;
 137
 138         result = 0;
 139         spin_lock(&xprt_list_lock);
 140         list_for_each_entry(t, &xprt_list, list) {
 141                 if (t == transport) {
 142                         printk(KERN_INFO
 143                                 "RPC: Unregistered %s transport module.\n",
 144                                 transport->name);
 145                         list_del_init(&transport->list);
 146                         module_put(THIS_MODULE);
 147                         goto out;
 148                 }
 149         }
 150         result = -ENOENT;
 151
 152 out:
 153         spin_unlock(&xprt_list_lock);
 154         return result;
 155 }
 156 EXPORT_SYMBOL_GPL(xprt_unregister_transport);
 157
 158 /**
 159  * xprt_reserve_xprt - serialize write access to transports
 160  * @task: task that is requesting access to the transport
 161  *
 162  * This prevents mixing the payload of separate requests, and prevents
 163  * transport connects from colliding with writes.  No congestion control
 164  * is provided.
 165  */
 166 int xprt_reserve_xprt(struct rpc_task *task)
 167 {
 168         struct rpc_xprt *xprt = task->tk_xprt;
 169         struct rpc_rqst *req = task->tk_rqstp;
 170
 171         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 172                 if (task == xprt->snd_task)
 173                         return 1;
 174                 if (task == NULL)
 175                         return 0;
 176                 goto out_sleep;
 177         }
 178         xprt->snd_task = task;
 179         if (req) {
 180                 req->rq_bytes_sent = 0;
 181                 req->rq_ntrans++;
 182         }
 183         return 1;
 184
 185 out_sleep:
 186         dprintk("RPC: %5u failed to lock transport %p\n",
 187                         task->tk_pid, xprt);
 188         task->tk_timeout = 0;
 189         task->tk_status = -EAGAIN;
 190         if (req && req->rq_ntrans)
 191                 rpc_sleep_on(&xprt->resend, task, NULL, NULL);
 192         else
 193                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
 194         return 0;
 195 }
 196 EXPORT_SYMBOL_GPL(xprt_reserve_xprt);
 197
 198 static void xprt_clear_locked(struct rpc_xprt *xprt)
 199 {
 200         xprt->snd_task = NULL;
 201         if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state) || xprt->shutdown) {
 202                 smp_mb__before_clear_bit();
 203                 clear_bit(XPRT_LOCKED, &xprt->state);
 204                 smp_mb__after_clear_bit();
 205         } else
 206                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 207 }
 208
 209 /*
 210  * xprt_reserve_xprt_cong - serialize write access to transports
 211  * @task: task that is requesting access to the transport
 212  *
 213  * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
 214  * integrated into the decision of whether a request is allowed to be
 215  * woken up and given access to the transport.
 216  */
 217 int xprt_reserve_xprt_cong(struct rpc_task *task)
 218 {
 219         struct rpc_xprt *xprt = task->tk_xprt;
 220         struct rpc_rqst *req = task->tk_rqstp;
 221
 222         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 223                 if (task == xprt->snd_task)
 224                         return 1;
 225                 goto out_sleep;
 226         }
 227         if (__xprt_get_cong(xprt, task)) {
 228                 xprt->snd_task = task;
 229                 if (req) {
 230                         req->rq_bytes_sent = 0;
 231                         req->rq_ntrans++;
 232                 }
 233                 return 1;
 234         }
 235         xprt_clear_locked(xprt);
 236 out_sleep:
 237         dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
 238         task->tk_timeout = 0;
 239         task->tk_status = -EAGAIN;
 240         if (req && req->rq_ntrans)
 241                 rpc_sleep_on(&xprt->resend, task, NULL, NULL);
 242         else
 243                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
 244         return 0;
 245 }
 246 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
 247
 248 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 249 {
 250         int retval;
 251
 252         spin_lock_bh(&xprt->transport_lock);
 253         retval = xprt->ops->reserve_xprt(task);
 254         spin_unlock_bh(&xprt->transport_lock);
 255         return retval;
 256 }
 257
 258 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
 259 {
 260         struct rpc_task *task;
 261         struct rpc_rqst *req;
 262
 263         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 264                 return;
 265
 266         task = rpc_wake_up_next(&xprt->resend);
 267         if (!task) {
 268                 task = rpc_wake_up_next(&xprt->sending);
 269                 if (!task)
 270                         goto out_unlock;
 271         }
 272
 273         req = task->tk_rqstp;
 274         xprt->snd_task = task;
 275         if (req) {
 276                 req->rq_bytes_sent = 0;
 277                 req->rq_ntrans++;
 278         }
 279         return;
 280
 281 out_unlock:
 282         xprt_clear_locked(xprt);
 283 }
 284
 285 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
 286 {
 287         struct rpc_task *task;
 288
 289         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 290                 return;
 291         if (RPCXPRT_CONGESTED(xprt))
 292                 goto out_unlock;
 293         task = rpc_wake_up_next(&xprt->resend);
 294         if (!task) {
 295                 task = rpc_wake_up_next(&xprt->sending);
 296                 if (!task)
 297                         goto out_unlock;
 298         }
 299         if (__xprt_get_cong(xprt, task)) {
 300                 struct rpc_rqst *req = task->tk_rqstp;
 301                 xprt->snd_task = task;
 302                 if (req) {
 303                         req->rq_bytes_sent = 0;
 304                         req->rq_ntrans++;
 305                 }
 306                 return;
 307         }
 308 out_unlock:
 309         xprt_clear_locked(xprt);
 310 }
 311
 312 /**
 313  * xprt_release_xprt - allow other requests to use a transport
 314  * @xprt: transport with other tasks potentially waiting
 315  * @task: task that is releasing access to the transport
 316  *
 317  * Note that "task" can be NULL.  No congestion control is provided.
 318  */
 319 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 320 {
 321         if (xprt->snd_task == task) {
 322                 xprt_clear_locked(xprt);
 323                 __xprt_lock_write_next(xprt);
 324         }
 325 }
 326 EXPORT_SYMBOL_GPL(xprt_release_xprt);
 327
 328 /**
 329  * xprt_release_xprt_cong - allow other requests to use a transport
 330  * @xprt: transport with other tasks potentially waiting
 331  * @task: task that is releasing access to the transport
 332  *
 333  * Note that "task" can be NULL.  Another task is awoken to use the
 334  * transport if the transport's congestion window allows it.
 335  */
 336 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 337 {
 338         if (xprt->snd_task == task) {
 339                 xprt_clear_locked(xprt);
 340                 __xprt_lock_write_next_cong(xprt);
 341         }
 342 }
 343 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
 344
 345 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 346 {
 347         spin_lock_bh(&xprt->transport_lock);
 348         xprt->ops->release_xprt(xprt, task);
 349         spin_unlock_bh(&xprt->transport_lock);
 350 }
 351
 352 /*
 353  * Van Jacobson congestion avoidance. Check if the congestion window
 354  * overflowed. Put the task to sleep if this is the case.
 355  */
 356 static int
 357 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 358 {
 359         struct rpc_rqst *req = task->tk_rqstp;
 360
 361         if (req->rq_cong)
 362                 return 1;
 363         dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
 364                         task->tk_pid, xprt->cong, xprt->cwnd);
 365         if (RPCXPRT_CONGESTED(xprt))
 366                 return 0;
 367         req->rq_cong = 1;
 368         xprt->cong += RPC_CWNDSCALE;
 369         return 1;
 370 }
 371
 372 /*
 373  * Adjust the congestion window, and wake up the next task
 374  * that has been sleeping due to congestion
 375  */
 376 static void
 377 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 378 {
 379         if (!req->rq_cong)
 380                 return;
 381         req->rq_cong = 0;
 382         xprt->cong -= RPC_CWNDSCALE;
 383         __xprt_lock_write_next_cong(xprt);
 384 }
 385
 386 /**
 387  * xprt_release_rqst_cong - housekeeping when request is complete
 388  * @task: RPC request that recently completed
 389  *
 390  * Useful for transports that require congestion control.
 391  */
 392 void xprt_release_rqst_cong(struct rpc_task *task)
 393 {
 394         __xprt_put_cong(task->tk_xprt, task->tk_rqstp);
 395 }
 396 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
 397
 398 /**
 399  * xprt_adjust_cwnd - adjust transport congestion window
 400  * @task: recently completed RPC request used to adjust window
 401  * @result: result code of completed RPC request
 402  *
 403  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
 404  */
 405 void xprt_adjust_cwnd(struct rpc_task *task, int result)
 406 {
 407         struct rpc_rqst *req = task->tk_rqstp;
 408         struct rpc_xprt *xprt = task->tk_xprt;
 409         unsigned long cwnd = xprt->cwnd;
 410
 411         if (result >= 0 && cwnd <= xprt->cong) {
 412                 /* The (cwnd >> 1) term makes sure
 413                  * the result gets rounded properly. */
 414                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 415                 if (cwnd > RPC_MAXCWND(xprt))
 416                         cwnd = RPC_MAXCWND(xprt);
 417                 __xprt_lock_write_next_cong(xprt);
 418         } else if (result == -ETIMEDOUT) {
 419                 cwnd >>= 1;
 420                 if (cwnd < RPC_CWNDSCALE)
 421                         cwnd = RPC_CWNDSCALE;
 422         }
 423         dprintk("RPC:       cong %ld, cwnd was %ld, now %ld\n",
 424                         xprt->cong, xprt->cwnd, cwnd);
 425         xprt->cwnd = cwnd;
 426         __xprt_put_cong(xprt, req);
 427 }
 428 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
 429
 430 /**
 431  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
 432  * @xprt: transport with waiting tasks
 433  * @status: result code to plant in each task before waking it
 434  *
 435  */
 436 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
 437 {
 438         if (status < 0)
 439                 rpc_wake_up_status(&xprt->pending, status);
 440         else
 441                 rpc_wake_up(&xprt->pending);
 442 }
 443 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
 444
 445 /**
 446  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
 447  * @task: task to be put to sleep
 448  *
 449  */
 450 void xprt_wait_for_buffer_space(struct rpc_task *task)
 451 {
 452         struct rpc_rqst *req = task->tk_rqstp;
 453         struct rpc_xprt *xprt = req->rq_xprt;
 454
 455         task->tk_timeout = req->rq_timeout;
 456         rpc_sleep_on(&xprt->pending, task, NULL, NULL);
 457 }
 458 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
 459
 460 /**
 461  * xprt_write_space - wake the task waiting for transport output buffer space
 462  * @xprt: transport with waiting tasks
 463  *
 464  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
 465  */
 466 void xprt_write_space(struct rpc_xprt *xprt)
 467 {
 468         if (unlikely(xprt->shutdown))
 469                 return;
 470
 471         spin_lock_bh(&xprt->transport_lock);
 472         if (xprt->snd_task) {
 473                 dprintk("RPC:       write space: waking waiting task on "
 474                                 "xprt %p\n", xprt);
 475                 rpc_wake_up_task(xprt->snd_task);
 476         }
 477         spin_unlock_bh(&xprt->transport_lock);
 478 }
 479 EXPORT_SYMBOL_GPL(xprt_write_space);
 480
 481 /**
 482  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
 483  * @task: task whose timeout is to be set
 484  *
 485  * Set a request's retransmit timeout based on the transport's
 486  * default timeout parameters.  Used by transports that don't adjust
 487  * the retransmit timeout based on round-trip time estimation.
 488  */
 489 void xprt_set_retrans_timeout_def(struct rpc_task *task)
 490 {
 491         task->tk_timeout = task->tk_rqstp->rq_timeout;
 492 }
 493 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
 494
 495 /*
 496  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
 497  * @task: task whose timeout is to be set
 498  *
 499  * Set a request's retransmit timeout using the RTT estimator.
 500  */
 501 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
 502 {
 503         int timer = task->tk_msg.rpc_proc->p_timer;
 504         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
 505         struct rpc_rqst *req = task->tk_rqstp;
 506         unsigned long max_timeout = req->rq_xprt->timeout.to_maxval;
 507
 508         task->tk_timeout = rpc_calc_rto(rtt, timer);
 509         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
 510         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
 511                 task->tk_timeout = max_timeout;
 512 }
 513 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
 514
 515 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 516 {
 517         struct rpc_timeout *to = &req->rq_xprt->timeout;
 518
 519         req->rq_majortimeo = req->rq_timeout;
 520         if (to->to_exponential)
 521                 req->rq_majortimeo <<= to->to_retries;
 522         else
 523                 req->rq_majortimeo += to->to_increment * to->to_retries;
 524         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 525                 req->rq_majortimeo = to->to_maxval;
 526         req->rq_majortimeo += jiffies;
 527 }
 528
 529 /**
 530  * xprt_adjust_timeout - adjust timeout values for next retransmit
 531  * @req: RPC request containing parameters to use for the adjustment
 532  *
 533  */
 534 int xprt_adjust_timeout(struct rpc_rqst *req)
 535 {
 536         struct rpc_xprt *xprt = req->rq_xprt;
 537         struct rpc_timeout *to = &xprt->timeout;
 538         int status = 0;
 539
 540         if (time_before(jiffies, req->rq_majortimeo)) {
 541                 if (to->to_exponential)
 542                         req->rq_timeout <<= 1;
 543                 else
 544                         req->rq_timeout += to->to_increment;
 545                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 546                         req->rq_timeout = to->to_maxval;
 547                 req->rq_retries++;
 548         } else {
 549                 req->rq_timeout = to->to_initval;
 550                 req->rq_retries = 0;
 551                 xprt_reset_majortimeo(req);
 552                 /* Reset the RTT counters == "slow start" */
 553                 spin_lock_bh(&xprt->transport_lock);
 554                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 555                 spin_unlock_bh(&xprt->transport_lock);
 556                 status = -ETIMEDOUT;
 557         }
 558
 559         if (req->rq_timeout == 0) {
 560                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 561                 req->rq_timeout = 5 * HZ;
 562         }
 563         return status;
 564 }
 565
 566 static void xprt_autoclose(struct work_struct *work)
 567 {
 568         struct rpc_xprt *xprt =
 569                 container_of(work, struct rpc_xprt, task_cleanup);
 570
 571         xprt_disconnect(xprt);
 572         xprt->ops->close(xprt);
 573         xprt_release_write(xprt, NULL);
 574 }
 575
 576 /**
 577  * xprt_disconnect - mark a transport as disconnected
 578  * @xprt: transport to flag for disconnect
 579  *
 580  */
 581 void xprt_disconnect(struct rpc_xprt *xprt)
 582 {
 583         dprintk("RPC:       disconnected transport %p\n", xprt);
 584         spin_lock_bh(&xprt->transport_lock);
 585         xprt_clear_connected(xprt);
 586         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 587         spin_unlock_bh(&xprt->transport_lock);
 588 }
 589 EXPORT_SYMBOL_GPL(xprt_disconnect);
 590
 591 static void
 592 xprt_init_autodisconnect(unsigned long data)
 593 {
 594         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 595
 596         spin_lock(&xprt->transport_lock);
 597         if (!list_empty(&xprt->recv) || xprt->shutdown)
 598                 goto out_abort;
 599         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 600                 goto out_abort;
 601         spin_unlock(&xprt->transport_lock);
 602         if (xprt_connecting(xprt))
 603                 xprt_release_write(xprt, NULL);
 604         else
 605                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 606         return;
 607 out_abort:
 608         spin_unlock(&xprt->transport_lock);
 609 }
 610
 611 /**
 612  * xprt_connect - schedule a transport connect operation
 613  * @task: RPC task that is requesting the connect
 614  *
 615  */
 616 void xprt_connect(struct rpc_task *task)
 617 {
 618         struct rpc_xprt *xprt = task->tk_xprt;
 619
 620         dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
 621                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 622
 623         if (!xprt_bound(xprt)) {
 624                 task->tk_status = -EIO;
 625                 return;
 626         }
 627         if (!xprt_lock_write(xprt, task))
 628                 return;
 629         if (xprt_connected(xprt))
 630                 xprt_release_write(xprt, task);
 631         else {
 632                 if (task->tk_rqstp)
 633                         task->tk_rqstp->rq_bytes_sent = 0;
 634
 635                 task->tk_timeout = xprt->connect_timeout;
 636                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
 637                 xprt->stat.connect_start = jiffies;
 638                 xprt->ops->connect(task);
 639         }
 640         return;
 641 }
 642
 643 static void xprt_connect_status(struct rpc_task *task)
 644 {
 645         struct rpc_xprt *xprt = task->tk_xprt;
 646
 647         if (task->tk_status >= 0) {
 648                 xprt->stat.connect_count++;
 649                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
 650                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
 651                                 task->tk_pid);
 652                 return;
 653         }
 654
 655         switch (task->tk_status) {
 656         case -ECONNREFUSED:
 657         case -ECONNRESET:
 658                 dprintk("RPC: %5u xprt_connect_status: server %s refused "
 659                                 "connection\n", task->tk_pid,
 660                                 task->tk_client->cl_server);
 661                 break;
 662         case -ENOTCONN:
 663                 dprintk("RPC: %5u xprt_connect_status: connection broken\n",
 664                                 task->tk_pid);
 665                 break;
 666         case -ETIMEDOUT:
 667                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
 668                                 "out\n", task->tk_pid);
 669                 break;
 670         default:
 671                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
 672                                 "server %s\n", task->tk_pid, -task->tk_status,
 673                                 task->tk_client->cl_server);
 674                 xprt_release_write(xprt, task);
 675                 task->tk_status = -EIO;
 676         }
 677 }
 678
 679 /**
 680  * xprt_lookup_rqst - find an RPC request corresponding to an XID
 681  * @xprt: transport on which the original request was transmitted
 682  * @xid: RPC XID of incoming reply
 683  *
 684  */
 685 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
 686 {
 687         struct list_head *pos;
 688
 689         list_for_each(pos, &xprt->recv) {
 690                 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
 691                 if (entry->rq_xid == xid)
 692                         return entry;
 693         }
 694
 695         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
 696                         ntohl(xid));
 697         xprt->stat.bad_xids++;
 698         return NULL;
 699 }
 700 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
 701
 702 /**
 703  * xprt_update_rtt - update an RPC client's RTT state after receiving a reply
 704  * @task: RPC request that recently completed
 705  *
 706  */
 707 void xprt_update_rtt(struct rpc_task *task)
 708 {
 709         struct rpc_rqst *req = task->tk_rqstp;
 710         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
 711         unsigned timer = task->tk_msg.rpc_proc->p_timer;
 712
 713         if (timer) {
 714                 if (req->rq_ntrans == 1)
 715                         rpc_update_rtt(rtt, timer,
 716                                         (long)jiffies - req->rq_xtime);
 717                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
 718         }
 719 }
 720 EXPORT_SYMBOL_GPL(xprt_update_rtt);
 721
 722 /**
 723  * xprt_complete_rqst - called when reply processing is complete
 724  * @task: RPC request that recently completed
 725  * @copied: actual number of bytes received from the transport
 726  *
 727  * Caller holds transport lock.
 728  */
 729 void xprt_complete_rqst(struct rpc_task *task, int copied)
 730 {
 731         struct rpc_rqst *req = task->tk_rqstp;
 732
 733         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
 734                         task->tk_pid, ntohl(req->rq_xid), copied);
 735
 736         task->tk_xprt->stat.recvs++;
 737         task->tk_rtt = (long)jiffies - req->rq_xtime;
 738
 739         list_del_init(&req->rq_list);
 740         /* Ensure all writes are done before we update req->rq_received */
 741         smp_wmb();
 742         req->rq_received = req->rq_private_buf.len = copied;
 743         rpc_wake_up_task(task);
 744 }
 745 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
 746
 747 static void xprt_timer(struct rpc_task *task)
 748 {
 749         struct rpc_rqst *req = task->tk_rqstp;
 750         struct rpc_xprt *xprt = req->rq_xprt;
 751
 752         dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
 753
 754         spin_lock(&xprt->transport_lock);
 755         if (!req->rq_received) {
 756                 if (xprt->ops->timer)
 757                         xprt->ops->timer(task);
 758                 task->tk_status = -ETIMEDOUT;
 759         }
 760         task->tk_timeout = 0;
 761         rpc_wake_up_task(task);
 762         spin_unlock(&xprt->transport_lock);
 763 }
 764
 765 /**
 766  * xprt_prepare_transmit - reserve the transport before sending a request
 767  * @task: RPC task about to send a request
 768  *
 769  */
 770 int xprt_prepare_transmit(struct rpc_task *task)
 771 {
 772         struct rpc_rqst *req = task->tk_rqstp;
 773         struct rpc_xprt *xprt = req->rq_xprt;
 774         int err = 0;
 775
 776         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
 777
 778         spin_lock_bh(&xprt->transport_lock);
 779         if (req->rq_received && !req->rq_bytes_sent) {
 780                 err = req->rq_received;
 781                 goto out_unlock;
 782         }
 783         if (!xprt->ops->reserve_xprt(task)) {
 784                 err = -EAGAIN;
 785                 goto out_unlock;
 786         }
 787
 788         if (!xprt_connected(xprt)) {
 789                 err = -ENOTCONN;
 790                 goto out_unlock;
 791         }
 792 out_unlock:
 793         spin_unlock_bh(&xprt->transport_lock);
 794         return err;
 795 }
 796
 797 void xprt_end_transmit(struct rpc_task *task)
 798 {
 799         xprt_release_write(task->tk_xprt, task);
 800 }
 801
 802 /**
 803  * xprt_transmit - send an RPC request on a transport
 804  * @task: controlling RPC task
 805  *
 806  * We have to copy the iovec because sendmsg fiddles with its contents.
 807  */
 808 void xprt_transmit(struct rpc_task *task)
 809 {
 810         struct rpc_rqst *req = task->tk_rqstp;
 811         struct rpc_xprt *xprt = req->rq_xprt;
 812         int status;
 813
 814         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
 815
 816         if (!req->rq_received) {
 817                 if (list_empty(&req->rq_list)) {
 818                         spin_lock_bh(&xprt->transport_lock);
 819                         /* Update the softirq receive buffer */
 820                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
 821                                         sizeof(req->rq_private_buf));
 822                         /* Add request to the receive list */
 823                         list_add_tail(&req->rq_list, &xprt->recv);
 824                         spin_unlock_bh(&xprt->transport_lock);
 825                         xprt_reset_majortimeo(req);
 826                         /* Turn off autodisconnect */
 827                         del_singleshot_timer_sync(&xprt->timer);
 828                 }
 829         } else if (!req->rq_bytes_sent)
 830                 return;
 831
 832         status = xprt->ops->send_request(task);
 833         if (status == 0) {
 834                 dprintk("RPC: %5u xmit complete\n", task->tk_pid);
 835                 spin_lock_bh(&xprt->transport_lock);
 836
 837                 xprt->ops->set_retrans_timeout(task);
 838
 839                 xprt->stat.sends++;
 840                 xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
 841                 xprt->stat.bklog_u += xprt->backlog.qlen;
 842
 843                 /* Don't race with disconnect */
 844                 if (!xprt_connected(xprt))
 845                         task->tk_status = -ENOTCONN;
 846                 else if (!req->rq_received)
 847                         rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
 848                 spin_unlock_bh(&xprt->transport_lock);
 849                 return;
 850         }
 851
 852         /* Note: at this point, task->tk_sleeping has not yet been set,
 853          *       hence there is no danger of the waking up task being put on
 854          *       schedq, and being picked up by a parallel run of rpciod().
 855          */
 856         task->tk_status = status;
 857         if (status == -ECONNREFUSED)
 858                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
 859 }
 860
 861 static inline void do_xprt_reserve(struct rpc_task *task)
 862 {
 863         struct rpc_xprt *xprt = task->tk_xprt;
 864
 865         task->tk_status = 0;
 866         if (task->tk_rqstp)
 867                 return;
 868         if (!list_empty(&xprt->free)) {
 869                 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
 870                 list_del_init(&req->rq_list);
 871                 task->tk_rqstp = req;
 872                 xprt_request_init(task, xprt);
 873                 return;
 874         }
 875         dprintk("RPC:       waiting for request slot\n");
 876         task->tk_status = -EAGAIN;
 877         task->tk_timeout = 0;
 878         rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
 879 }
 880
 881 /**
 882  * xprt_reserve - allocate an RPC request slot
 883  * @task: RPC task requesting a slot allocation
 884  *
 885  * If no more slots are available, place the task on the transport's
 886  * backlog queue.
 887  */
 888 void xprt_reserve(struct rpc_task *task)
 889 {
 890         struct rpc_xprt *xprt = task->tk_xprt;
 891
 892         task->tk_status = -EIO;
 893         spin_lock(&xprt->reserve_lock);
 894         do_xprt_reserve(task);
 895         spin_unlock(&xprt->reserve_lock);
 896 }
 897
 898 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
 899 {
 900         return xprt->xid++;
 901 }
 902
 903 static inline void xprt_init_xid(struct rpc_xprt *xprt)
 904 {
 905         xprt->xid = net_random();
 906 }
 907
 908 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
 909 {
 910         struct rpc_rqst *req = task->tk_rqstp;
 911
 912         req->rq_timeout = xprt->timeout.to_initval;
 913         req->rq_task    = task;
 914         req->rq_xprt    = xprt;
 915         req->rq_buffer  = NULL;
 916         req->rq_xid     = xprt_alloc_xid(xprt);
 917         req->rq_release_snd_buf = NULL;
 918         xprt_reset_majortimeo(req);
 919         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
 920                         req, ntohl(req->rq_xid));
 921 }
 922
 923 /**
 924  * xprt_release - release an RPC request slot
 925  * @task: task which is finished with the slot
 926  *
 927  */
 928 void xprt_release(struct rpc_task *task)
 929 {
 930         struct rpc_xprt *xprt = task->tk_xprt;
 931         struct rpc_rqst *req;
 932
 933         if (!(req = task->tk_rqstp))
 934                 return;
 935         rpc_count_iostats(task);
 936         spin_lock_bh(&xprt->transport_lock);
 937         xprt->ops->release_xprt(xprt, task);
 938         if (xprt->ops->release_request)
 939                 xprt->ops->release_request(task);
 940         if (!list_empty(&req->rq_list))
 941                 list_del(&req->rq_list);
 942         xprt->last_used = jiffies;
 943         if (list_empty(&xprt->recv))
 944                 mod_timer(&xprt->timer,
 945                                 xprt->last_used + xprt->idle_timeout);
 946         spin_unlock_bh(&xprt->transport_lock);
 947         xprt->ops->buf_free(req->rq_buffer);
 948         task->tk_rqstp = NULL;
 949         if (req->rq_release_snd_buf)
 950                 req->rq_release_snd_buf(req);
 951         memset(req, 0, sizeof(*req));   /* mark unused */
 952
 953         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
 954
 955         spin_lock(&xprt->reserve_lock);
 956         list_add(&req->rq_list, &xprt->free);
 957         rpc_wake_up_next(&xprt->backlog);
 958         spin_unlock(&xprt->reserve_lock);
 959 }
 960
 961 /**
 962  * xprt_set_timeout - set constant RPC timeout
 963  * @to: RPC timeout parameters to set up
 964  * @retr: number of retries
 965  * @incr: amount of increase after each retry
 966  *
 967  */
 968 void xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
 969 {
 970         to->to_initval   =
 971         to->to_increment = incr;
 972         to->to_maxval    = to->to_initval + (incr * retr);
 973         to->to_retries   = retr;
 974         to->to_exponential = 0;
 975 }
 976
 977 /**
 978  * xprt_create_transport - create an RPC transport
 979  * @args: rpc transport creation arguments
 980  *
 981  */
 982 struct rpc_xprt *xprt_create_transport(struct rpc_xprtsock_create *args)
 983 {
 984         struct rpc_xprt *xprt;
 985         struct rpc_rqst *req;
 986         struct xprt_class *t;
 987
 988         spin_lock(&xprt_list_lock);
 989         list_for_each_entry(t, &xprt_list, list) {
 990                 if ((t->family == args->dstaddr->sa_family) &&
 991                     (t->protocol == args->proto)) {
 992                         spin_unlock(&xprt_list_lock);
 993                         goto found;
 994                 }
 995         }
 996         spin_unlock(&xprt_list_lock);
 997         printk(KERN_ERR "RPC: transport (%u/%d) not supported\n",
 998                         args->dstaddr->sa_family, args->proto);
 999         return ERR_PTR(-EIO);
1000
1001 found:
1002         xprt = t->setup(args);
1003         if (IS_ERR(xprt)) {
1004                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1005                                 -PTR_ERR(xprt));
1006                 return xprt;
1007         }
1008
1009         kref_init(&xprt->kref);
1010         spin_lock_init(&xprt->transport_lock);
1011         spin_lock_init(&xprt->reserve_lock);
1012
1013         INIT_LIST_HEAD(&xprt->free);
1014         INIT_LIST_HEAD(&xprt->recv);
1015         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1016         init_timer(&xprt->timer);
1017         xprt->timer.function = xprt_init_autodisconnect;
1018         xprt->timer.data = (unsigned long) xprt;
1019         xprt->last_used = jiffies;
1020         xprt->cwnd = RPC_INITCWND;
1021         xprt->bind_index = 0;
1022
1023         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1024         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1025         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
1026         rpc_init_wait_queue(&xprt->resend, "xprt_resend");
1027         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1028
1029         /* initialize free list */
1030         for (req = &xprt->slot[xprt->max_reqs-1]; req >= &xprt->slot[0]; req--)
1031                 list_add(&req->rq_list, &xprt->free);
1032
1033         xprt_init_xid(xprt);
1034
1035         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1036                         xprt->max_reqs);
1037
1038         return xprt;
1039 }
1040
1041 /**
1042  * xprt_destroy - destroy an RPC transport, killing off all requests.
1043  * @kref: kref for the transport to destroy
1044  *
1045  */
1046 static void xprt_destroy(struct kref *kref)
1047 {
1048         struct rpc_xprt *xprt = container_of(kref, struct rpc_xprt, kref);
1049
1050         dprintk("RPC:       destroying transport %p\n", xprt);
1051         xprt->shutdown = 1;
1052         del_timer_sync(&xprt->timer);
1053
1054         /*
1055          * Tear down transport state and free the rpc_xprt
1056          */
1057         xprt->ops->destroy(xprt);
1058 }
1059
1060 /**
1061  * xprt_put - release a reference to an RPC transport.
1062  * @xprt: pointer to the transport
1063  *
1064  */
1065 void xprt_put(struct rpc_xprt *xprt)
1066 {
1067         kref_put(&xprt->kref, xprt_destroy);
1068 }
1069
1070 /**
1071  * xprt_get - return a reference to an RPC transport.
1072  * @xprt: pointer to the transport
1073  *
1074  */
1075 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1076 {
1077         kref_get(&xprt->kref);
1078         return xprt;
1079 }