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