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/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
 157 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
 158 {
 159         struct rpc_task *task;
 160
 161         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 162                 return;
 163         if (!xprt->nocong && RPCXPRT_CONGESTED(xprt))
 164                 goto out_unlock;
 165         task = rpc_wake_up_next(&xprt->resend);
 166         if (!task) {
 167                 task = rpc_wake_up_next(&xprt->sending);
 168                 if (!task)
 169                         goto out_unlock;
 170         }
 171         if (xprt->nocong || __xprt_get_cong(xprt, task)) {
 172                 struct rpc_rqst *req = task->tk_rqstp;
 173                 xprt->snd_task = task;
 174                 if (req) {
 175                         req->rq_bytes_sent = 0;
 176                         req->rq_ntrans++;
 177                 }
 178                 return;
 179         }
 180 out_unlock:
 181         smp_mb__before_clear_bit();
 182         clear_bit(XPRT_LOCKED, &xprt->state);
 183         smp_mb__after_clear_bit();
 184 }
 185
 186 /*
 187  * Releases the transport for use by other requests.
 188  */
 189 static void
 190 __xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 191 {
 192         if (xprt->snd_task == task) {
 193                 xprt->snd_task = NULL;
 194                 smp_mb__before_clear_bit();
 195                 clear_bit(XPRT_LOCKED, &xprt->state);
 196                 smp_mb__after_clear_bit();
 197                 __xprt_lock_write_next(xprt);
 198         }
 199 }
 200
 201 static inline void
 202 xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 203 {
 204         spin_lock_bh(&xprt->transport_lock);
 205         __xprt_release_write(xprt, task);
 206         spin_unlock_bh(&xprt->transport_lock);
 207 }
 208
 209 /*
 210  * Van Jacobson congestion avoidance. Check if the congestion window
 211  * overflowed. Put the task to sleep if this is the case.
 212  */
 213 static int
 214 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 215 {
 216         struct rpc_rqst *req = task->tk_rqstp;
 217
 218         if (req->rq_cong)
 219                 return 1;
 220         dprintk("RPC: %4d xprt_cwnd_limited cong = %ld cwnd = %ld\n",
 221                         task->tk_pid, xprt->cong, xprt->cwnd);
 222         if (RPCXPRT_CONGESTED(xprt))
 223                 return 0;
 224         req->rq_cong = 1;
 225         xprt->cong += RPC_CWNDSCALE;
 226         return 1;
 227 }
 228
 229 /*
 230  * Adjust the congestion window, and wake up the next task
 231  * that has been sleeping due to congestion
 232  */
 233 static void
 234 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 235 {
 236         if (!req->rq_cong)
 237                 return;
 238         req->rq_cong = 0;
 239         xprt->cong -= RPC_CWNDSCALE;
 240         __xprt_lock_write_next(xprt);
 241 }
 242
 243 /*
 244  * Adjust RPC congestion window
 245  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
 246  */
 247 static void
 248 xprt_adjust_cwnd(struct rpc_xprt *xprt, int result)
 249 {
 250         unsigned long   cwnd;
 251
 252         cwnd = xprt->cwnd;
 253         if (result >= 0 && cwnd <= xprt->cong) {
 254                 /* The (cwnd >> 1) term makes sure
 255                  * the result gets rounded properly. */
 256                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 257                 if (cwnd > RPC_MAXCWND(xprt))
 258                         cwnd = RPC_MAXCWND(xprt);
 259                 __xprt_lock_write_next(xprt);
 260         } else if (result == -ETIMEDOUT) {
 261                 cwnd >>= 1;
 262                 if (cwnd < RPC_CWNDSCALE)
 263                         cwnd = RPC_CWNDSCALE;
 264         }
 265         dprintk("RPC:      cong %ld, cwnd was %ld, now %ld\n",
 266                         xprt->cong, xprt->cwnd, cwnd);
 267         xprt->cwnd = cwnd;
 268 }
 269
 270 /**
 271  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
 272  * @xprt: transport with waiting tasks
 273  * @status: result code to plant in each task before waking it
 274  *
 275  */
 276 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
 277 {
 278         if (status < 0)
 279                 rpc_wake_up_status(&xprt->pending, status);
 280         else
 281                 rpc_wake_up(&xprt->pending);
 282 }
 283
 284 /**
 285  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
 286  * @task: task to be put to sleep
 287  *
 288  */
 289 void xprt_wait_for_buffer_space(struct rpc_task *task)
 290 {
 291         struct rpc_rqst *req = task->tk_rqstp;
 292         struct rpc_xprt *xprt = req->rq_xprt;
 293
 294         task->tk_timeout = req->rq_timeout;
 295         rpc_sleep_on(&xprt->pending, task, NULL, NULL);
 296 }
 297
 298 /**
 299  * xprt_write_space - wake the task waiting for transport output buffer space
 300  * @xprt: transport with waiting tasks
 301  *
 302  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
 303  */
 304 void xprt_write_space(struct rpc_xprt *xprt)
 305 {
 306         if (unlikely(xprt->shutdown))
 307                 return;
 308
 309         spin_lock_bh(&xprt->transport_lock);
 310         if (xprt->snd_task) {
 311                 dprintk("RPC:      write space: waking waiting task on xprt %p\n",
 312                                 xprt);
 313                 rpc_wake_up_task(xprt->snd_task);
 314         }
 315         spin_unlock_bh(&xprt->transport_lock);
 316 }
 317
 318 /**
 319  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
 320  * @task: task whose timeout is to be set
 321  *
 322  * Set a request's retransmit timeout based on the transport's
 323  * default timeout parameters.  Used by transports that don't adjust
 324  * the retransmit timeout based on round-trip time estimation.
 325  */
 326 void xprt_set_retrans_timeout_def(struct rpc_task *task)
 327 {
 328         task->tk_timeout = task->tk_rqstp->rq_timeout;
 329 }
 330
 331 /*
 332  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
 333  * @task: task whose timeout is to be set
 334  *
 335  * Set a request's retransmit timeout using the RTT estimator.
 336  */
 337 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
 338 {
 339         int timer = task->tk_msg.rpc_proc->p_timer;
 340         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
 341         struct rpc_rqst *req = task->tk_rqstp;
 342         unsigned long max_timeout = req->rq_xprt->timeout.to_maxval;
 343
 344         task->tk_timeout = rpc_calc_rto(rtt, timer);
 345         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
 346         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
 347                 task->tk_timeout = max_timeout;
 348 }
 349
 350 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 351 {
 352         struct rpc_timeout *to = &req->rq_xprt->timeout;
 353
 354         req->rq_majortimeo = req->rq_timeout;
 355         if (to->to_exponential)
 356                 req->rq_majortimeo <<= to->to_retries;
 357         else
 358                 req->rq_majortimeo += to->to_increment * to->to_retries;
 359         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 360                 req->rq_majortimeo = to->to_maxval;
 361         req->rq_majortimeo += jiffies;
 362 }
 363
 364 /**
 365  * xprt_adjust_timeout - adjust timeout values for next retransmit
 366  * @req: RPC request containing parameters to use for the adjustment
 367  *
 368  */
 369 int xprt_adjust_timeout(struct rpc_rqst *req)
 370 {
 371         struct rpc_xprt *xprt = req->rq_xprt;
 372         struct rpc_timeout *to = &xprt->timeout;
 373         int status = 0;
 374
 375         if (time_before(jiffies, req->rq_majortimeo)) {
 376                 if (to->to_exponential)
 377                         req->rq_timeout <<= 1;
 378                 else
 379                         req->rq_timeout += to->to_increment;
 380                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 381                         req->rq_timeout = to->to_maxval;
 382                 req->rq_retries++;
 383                 pprintk("RPC: %lu retrans\n", jiffies);
 384         } else {
 385                 req->rq_timeout = to->to_initval;
 386                 req->rq_retries = 0;
 387                 xprt_reset_majortimeo(req);
 388                 /* Reset the RTT counters == "slow start" */
 389                 spin_lock_bh(&xprt->transport_lock);
 390                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 391                 spin_unlock_bh(&xprt->transport_lock);
 392                 pprintk("RPC: %lu timeout\n", jiffies);
 393                 status = -ETIMEDOUT;
 394         }
 395
 396         if (req->rq_timeout == 0) {
 397                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 398                 req->rq_timeout = 5 * HZ;
 399         }
 400         return status;
 401 }
 402
 403 static void xprt_autoclose(void *args)
 404 {
 405         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
 406
 407         xprt_disconnect(xprt);
 408         xprt->ops->close(xprt);
 409         xprt_release_write(xprt, NULL);
 410 }
 411
 412 /**
 413  * xprt_disconnect - mark a transport as disconnected
 414  * @xprt: transport to flag for disconnect
 415  *
 416  */
 417 void xprt_disconnect(struct rpc_xprt *xprt)
 418 {
 419         dprintk("RPC:      disconnected transport %p\n", xprt);
 420         spin_lock_bh(&xprt->transport_lock);
 421         xprt_clear_connected(xprt);
 422         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 423         spin_unlock_bh(&xprt->transport_lock);
 424 }
 425
 426 static void
 427 xprt_init_autodisconnect(unsigned long data)
 428 {
 429         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 430
 431         spin_lock(&xprt->transport_lock);
 432         if (!list_empty(&xprt->recv) || xprt->shutdown)
 433                 goto out_abort;
 434         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 435                 goto out_abort;
 436         spin_unlock(&xprt->transport_lock);
 437         if (xprt_connecting(xprt))
 438                 xprt_release_write(xprt, NULL);
 439         else
 440                 schedule_work(&xprt->task_cleanup);
 441         return;
 442 out_abort:
 443         spin_unlock(&xprt->transport_lock);
 444 }
 445
 446 /**
 447  * xprt_connect - schedule a transport connect operation
 448  * @task: RPC task that is requesting the connect
 449  *
 450  */
 451 void xprt_connect(struct rpc_task *task)
 452 {
 453         struct rpc_xprt *xprt = task->tk_xprt;
 454
 455         dprintk("RPC: %4d xprt_connect xprt %p %s connected\n", task->tk_pid,
 456                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 457
 458         if (xprt->shutdown) {
 459                 task->tk_status = -EIO;
 460                 return;
 461         }
 462         if (!xprt->addr.sin_port) {
 463                 task->tk_status = -EIO;
 464                 return;
 465         }
 466         if (!xprt_lock_write(xprt, task))
 467                 return;
 468         if (xprt_connected(xprt))
 469                 xprt_release_write(xprt, task);
 470         else {
 471                 if (task->tk_rqstp)
 472                         task->tk_rqstp->rq_bytes_sent = 0;
 473
 474                 task->tk_timeout = RPC_CONNECT_TIMEOUT;
 475                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
 476                 xprt->ops->connect(task);
 477         }
 478         return;
 479 }
 480
 481 static void xprt_connect_status(struct rpc_task *task)
 482 {
 483         struct rpc_xprt *xprt = task->tk_xprt;
 484
 485         if (task->tk_status >= 0) {
 486                 dprintk("RPC: %4d xprt_connect_status: connection established\n",
 487                                 task->tk_pid);
 488                 return;
 489         }
 490
 491         switch (task->tk_status) {
 492         case -ECONNREFUSED:
 493         case -ECONNRESET:
 494                 dprintk("RPC: %4d xprt_connect_status: server %s refused connection\n",
 495                                 task->tk_pid, task->tk_client->cl_server);
 496                 break;
 497         case -ENOTCONN:
 498                 dprintk("RPC: %4d xprt_connect_status: connection broken\n",
 499                                 task->tk_pid);
 500                 break;
 501         case -ETIMEDOUT:
 502                 dprintk("RPC: %4d xprt_connect_status: connect attempt timed out\n",
 503                                 task->tk_pid);
 504                 break;
 505         default:
 506                 dprintk("RPC: %4d xprt_connect_status: error %d connecting to server %s\n",
 507                                 task->tk_pid, -task->tk_status, task->tk_client->cl_server);
 508                 xprt_release_write(xprt, task);
 509                 task->tk_status = -EIO;
 510                 return;
 511         }
 512
 513         /* if soft mounted, just cause this RPC to fail */
 514         if (RPC_IS_SOFT(task)) {
 515                 xprt_release_write(xprt, task);
 516                 task->tk_status = -EIO;
 517         }
 518 }
 519
 520 /**
 521  * xprt_lookup_rqst - find an RPC request corresponding to an XID
 522  * @xprt: transport on which the original request was transmitted
 523  * @xid: RPC XID of incoming reply
 524  *
 525  */
 526 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
 527 {
 528         struct list_head *pos;
 529         struct rpc_rqst *req = NULL;
 530
 531         list_for_each(pos, &xprt->recv) {
 532                 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
 533                 if (entry->rq_xid == xid) {
 534                         req = entry;
 535                         break;
 536                 }
 537         }
 538         return req;
 539 }
 540
 541 /**
 542  * xprt_complete_rqst - called when reply processing is complete
 543  * @xprt: controlling transport
 544  * @req: RPC request that just completed
 545  * @copied: actual number of bytes received from the transport
 546  *
 547  */
 548 void xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
 549 {
 550         struct rpc_task *task = req->rq_task;
 551         struct rpc_clnt *clnt = task->tk_client;
 552
 553         /* Adjust congestion window */
 554         if (!xprt->nocong) {
 555                 unsigned timer = task->tk_msg.rpc_proc->p_timer;
 556                 xprt_adjust_cwnd(xprt, copied);
 557                 __xprt_put_cong(xprt, req);
 558                 if (timer) {
 559                         if (req->rq_ntrans == 1)
 560                                 rpc_update_rtt(clnt->cl_rtt, timer,
 561                                                 (long)jiffies - req->rq_xtime);
 562                         rpc_set_timeo(clnt->cl_rtt, timer, req->rq_ntrans - 1);
 563                 }
 564         }
 565
 566 #ifdef RPC_PROFILE
 567         /* Profile only reads for now */
 568         if (copied > 1024) {
 569                 static unsigned long    nextstat;
 570                 static unsigned long    pkt_rtt, pkt_len, pkt_cnt;
 571
 572                 pkt_cnt++;
 573                 pkt_len += req->rq_slen + copied;
 574                 pkt_rtt += jiffies - req->rq_xtime;
 575                 if (time_before(nextstat, jiffies)) {
 576                         printk("RPC: %lu %ld cwnd\n", jiffies, xprt->cwnd);
 577                         printk("RPC: %ld %ld %ld %ld stat\n",
 578                                         jiffies, pkt_cnt, pkt_len, pkt_rtt);
 579                         pkt_rtt = pkt_len = pkt_cnt = 0;
 580                         nextstat = jiffies + 5 * HZ;
 581                 }
 582         }
 583 #endif
 584
 585         dprintk("RPC: %4d has input (%d bytes)\n", task->tk_pid, copied);
 586         list_del_init(&req->rq_list);
 587         req->rq_received = req->rq_private_buf.len = copied;
 588
 589         /* ... and wake up the process. */
 590         rpc_wake_up_task(task);
 591         return;
 592 }
 593
 594 /*
 595  * RPC receive timeout handler.
 596  */
 597 static void
 598 xprt_timer(struct rpc_task *task)
 599 {
 600         struct rpc_rqst *req = task->tk_rqstp;
 601         struct rpc_xprt *xprt = req->rq_xprt;
 602
 603         spin_lock(&xprt->transport_lock);
 604         if (req->rq_received)
 605                 goto out;
 606
 607         xprt_adjust_cwnd(req->rq_xprt, -ETIMEDOUT);
 608         __xprt_put_cong(xprt, req);
 609
 610         dprintk("RPC: %4d xprt_timer (%s request)\n",
 611                 task->tk_pid, req ? "pending" : "backlogged");
 612
 613         task->tk_status  = -ETIMEDOUT;
 614 out:
 615         task->tk_timeout = 0;
 616         rpc_wake_up_task(task);
 617         spin_unlock(&xprt->transport_lock);
 618 }
 619
 620 /**
 621  * xprt_prepare_transmit - reserve the transport before sending a request
 622  * @task: RPC task about to send a request
 623  *
 624  */
 625 int xprt_prepare_transmit(struct rpc_task *task)
 626 {
 627         struct rpc_rqst *req = task->tk_rqstp;
 628         struct rpc_xprt *xprt = req->rq_xprt;
 629         int err = 0;
 630
 631         dprintk("RPC: %4d xprt_prepare_transmit\n", task->tk_pid);
 632
 633         if (xprt->shutdown)
 634                 return -EIO;
 635
 636         spin_lock_bh(&xprt->transport_lock);
 637         if (req->rq_received && !req->rq_bytes_sent) {
 638                 err = req->rq_received;
 639                 goto out_unlock;
 640         }
 641         if (!xprt->ops->reserve_xprt(task)) {
 642                 err = -EAGAIN;
 643                 goto out_unlock;
 644         }
 645
 646         if (!xprt_connected(xprt)) {
 647                 err = -ENOTCONN;
 648                 goto out_unlock;
 649         }
 650 out_unlock:
 651         spin_unlock_bh(&xprt->transport_lock);
 652         return err;
 653 }
 654
 655 /**
 656  * xprt_transmit - send an RPC request on a transport
 657  * @task: controlling RPC task
 658  *
 659  * We have to copy the iovec because sendmsg fiddles with its contents.
 660  */
 661 void xprt_transmit(struct rpc_task *task)
 662 {
 663         struct rpc_rqst *req = task->tk_rqstp;
 664         struct rpc_xprt *xprt = req->rq_xprt;
 665         int status;
 666
 667         dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
 668
 669         smp_rmb();
 670         if (!req->rq_received) {
 671                 if (list_empty(&req->rq_list)) {
 672                         spin_lock_bh(&xprt->transport_lock);
 673                         /* Update the softirq receive buffer */
 674                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
 675                                         sizeof(req->rq_private_buf));
 676                         /* Add request to the receive list */
 677                         list_add_tail(&req->rq_list, &xprt->recv);
 678                         spin_unlock_bh(&xprt->transport_lock);
 679                         xprt_reset_majortimeo(req);
 680                         /* Turn off autodisconnect */
 681                         del_singleshot_timer_sync(&xprt->timer);
 682                 }
 683         } else if (!req->rq_bytes_sent)
 684                 return;
 685
 686         status = xprt->ops->send_request(task);
 687         if (status == 0) {
 688                 dprintk("RPC: %4d xmit complete\n", task->tk_pid);
 689                 spin_lock_bh(&xprt->transport_lock);
 690                 xprt->ops->set_retrans_timeout(task);
 691                 /* Don't race with disconnect */
 692                 if (!xprt_connected(xprt))
 693                         task->tk_status = -ENOTCONN;
 694                 else if (!req->rq_received)
 695                         rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
 696                 __xprt_release_write(xprt, task);
 697                 spin_unlock_bh(&xprt->transport_lock);
 698                 return;
 699         }
 700
 701         /* Note: at this point, task->tk_sleeping has not yet been set,
 702          *       hence there is no danger of the waking up task being put on
 703          *       schedq, and being picked up by a parallel run of rpciod().
 704          */
 705         task->tk_status = status;
 706
 707         switch (status) {
 708         case -ECONNREFUSED:
 709                 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
 710                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
 711         case -EAGAIN:
 712         case -ENOTCONN:
 713                 return;
 714         default:
 715                 break;
 716         }
 717         xprt_release_write(xprt, task);
 718         return;
 719 }
 720
 721 static inline void do_xprt_reserve(struct rpc_task *task)
 722 {
 723         struct rpc_xprt *xprt = task->tk_xprt;
 724
 725         task->tk_status = 0;
 726         if (task->tk_rqstp)
 727                 return;
 728         if (!list_empty(&xprt->free)) {
 729                 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
 730                 list_del_init(&req->rq_list);
 731                 task->tk_rqstp = req;
 732                 xprt_request_init(task, xprt);
 733                 return;
 734         }
 735         dprintk("RPC:      waiting for request slot\n");
 736         task->tk_status = -EAGAIN;
 737         task->tk_timeout = 0;
 738         rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
 739 }
 740
 741 /**
 742  * xprt_reserve - allocate an RPC request slot
 743  * @task: RPC task requesting a slot allocation
 744  *
 745  * If no more slots are available, place the task on the transport's
 746  * backlog queue.
 747  */
 748 void xprt_reserve(struct rpc_task *task)
 749 {
 750         struct rpc_xprt *xprt = task->tk_xprt;
 751
 752         task->tk_status = -EIO;
 753         if (!xprt->shutdown) {
 754                 spin_lock(&xprt->reserve_lock);
 755                 do_xprt_reserve(task);
 756                 spin_unlock(&xprt->reserve_lock);
 757         }
 758 }
 759
 760 static inline u32 xprt_alloc_xid(struct rpc_xprt *xprt)
 761 {
 762         return xprt->xid++;
 763 }
 764
 765 static inline void xprt_init_xid(struct rpc_xprt *xprt)
 766 {
 767         get_random_bytes(&xprt->xid, sizeof(xprt->xid));
 768 }
 769
 770 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
 771 {
 772         struct rpc_rqst *req = task->tk_rqstp;
 773
 774         req->rq_timeout = xprt->timeout.to_initval;
 775         req->rq_task    = task;
 776         req->rq_xprt    = xprt;
 777         req->rq_xid     = xprt_alloc_xid(xprt);
 778         dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
 779                         req, ntohl(req->rq_xid));
 780 }
 781
 782 /**
 783  * xprt_release - release an RPC request slot
 784  * @task: task which is finished with the slot
 785  *
 786  */
 787 void xprt_release(struct rpc_task *task)
 788 {
 789         struct rpc_xprt *xprt = task->tk_xprt;
 790         struct rpc_rqst *req;
 791
 792         if (!(req = task->tk_rqstp))
 793                 return;
 794         spin_lock_bh(&xprt->transport_lock);
 795         __xprt_release_write(xprt, task);
 796         __xprt_put_cong(xprt, req);
 797         if (!list_empty(&req->rq_list))
 798                 list_del(&req->rq_list);
 799         xprt->last_used = jiffies;
 800         if (list_empty(&xprt->recv) && !xprt->shutdown)
 801                 mod_timer(&xprt->timer,
 802                                 xprt->last_used + RPC_IDLE_DISCONNECT_TIMEOUT);
 803         spin_unlock_bh(&xprt->transport_lock);
 804         task->tk_rqstp = NULL;
 805         memset(req, 0, sizeof(*req));   /* mark unused */
 806
 807         dprintk("RPC: %4d release request %p\n", task->tk_pid, req);
 808
 809         spin_lock(&xprt->reserve_lock);
 810         list_add(&req->rq_list, &xprt->free);
 811         xprt_clear_backlog(xprt);
 812         spin_unlock(&xprt->reserve_lock);
 813 }
 814
 815 /**
 816  * xprt_set_timeout - set constant RPC timeout
 817  * @to: RPC timeout parameters to set up
 818  * @retr: number of retries
 819  * @incr: amount of increase after each retry
 820  *
 821  */
 822 void xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
 823 {
 824         to->to_initval   =
 825         to->to_increment = incr;
 826         to->to_maxval    = to->to_initval + (incr * retr);
 827         to->to_retries   = retr;
 828         to->to_exponential = 0;
 829 }
 830
 831 static struct rpc_xprt *xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
 832 {
 833         int result;
 834         struct rpc_xprt *xprt;
 835         struct rpc_rqst *req;
 836
 837         if ((xprt = kmalloc(sizeof(struct rpc_xprt), GFP_KERNEL)) == NULL)
 838                 return ERR_PTR(-ENOMEM);
 839         memset(xprt, 0, sizeof(*xprt)); /* Nnnngh! */
 840
 841         xprt->addr = *ap;
 842
 843         switch (proto) {
 844         case IPPROTO_UDP:
 845                 result = xs_setup_udp(xprt, to);
 846                 break;
 847         case IPPROTO_TCP:
 848                 result = xs_setup_tcp(xprt, to);
 849                 break;
 850         default:
 851                 printk(KERN_ERR "RPC: unrecognized transport protocol: %d\n",
 852                                 proto);
 853                 result = -EIO;
 854                 break;
 855         }
 856         if (result) {
 857                 kfree(xprt);
 858                 return ERR_PTR(result);
 859         }
 860
 861         spin_lock_init(&xprt->transport_lock);
 862         spin_lock_init(&xprt->reserve_lock);
 863         init_waitqueue_head(&xprt->cong_wait);
 864
 865         INIT_LIST_HEAD(&xprt->free);
 866         INIT_LIST_HEAD(&xprt->recv);
 867         INIT_WORK(&xprt->task_cleanup, xprt_autoclose, xprt);
 868         init_timer(&xprt->timer);
 869         xprt->timer.function = xprt_init_autodisconnect;
 870         xprt->timer.data = (unsigned long) xprt;
 871         xprt->last_used = jiffies;
 872
 873         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
 874         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
 875         rpc_init_wait_queue(&xprt->resend, "xprt_resend");
 876         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
 877
 878         /* initialize free list */
 879         for (req = &xprt->slot[xprt->max_reqs-1]; req >= &xprt->slot[0]; req--)
 880                 list_add(&req->rq_list, &xprt->free);
 881
 882         xprt_init_xid(xprt);
 883
 884         dprintk("RPC:      created transport %p with %u slots\n", xprt,
 885                         xprt->max_reqs);
 886
 887         return xprt;
 888 }
 889
 890 /**
 891  * xprt_create_proto - create an RPC client transport
 892  * @proto: requested transport protocol
 893  * @sap: remote peer's address
 894  * @to: timeout parameters for new transport
 895  *
 896  */
 897 struct rpc_xprt *xprt_create_proto(int proto, struct sockaddr_in *sap, struct rpc_timeout *to)
 898 {
 899         struct rpc_xprt *xprt;
 900
 901         xprt = xprt_setup(proto, sap, to);
 902         if (IS_ERR(xprt))
 903                 dprintk("RPC:      xprt_create_proto failed\n");
 904         else
 905                 dprintk("RPC:      xprt_create_proto created xprt %p\n", xprt);
 906         return xprt;
 907 }
 908
 909 static void xprt_shutdown(struct rpc_xprt *xprt)
 910 {
 911         xprt->shutdown = 1;
 912         rpc_wake_up(&xprt->sending);
 913         rpc_wake_up(&xprt->resend);
 914         xprt_wake_pending_tasks(xprt, -EIO);
 915         rpc_wake_up(&xprt->backlog);
 916         wake_up(&xprt->cong_wait);
 917         del_timer_sync(&xprt->timer);
 918 }
 919
 920 static int xprt_clear_backlog(struct rpc_xprt *xprt) {
 921         rpc_wake_up_next(&xprt->backlog);
 922         wake_up(&xprt->cong_wait);
 923         return 1;
 924 }
 925
 926 /**
 927  * xprt_destroy - destroy an RPC transport, killing off all requests.
 928  * @xprt: transport to destroy
 929  *
 930  */
 931 int xprt_destroy(struct rpc_xprt *xprt)
 932 {
 933         dprintk("RPC:      destroying transport %p\n", xprt);
 934         xprt_shutdown(xprt);
 935         xprt->ops->destroy(xprt);
 936         kfree(xprt);
 937
 938         return 0;
 939 }