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  * Serialize write access to transports, in order to prevent different
  69  * requests from interfering with each other.
  70  * Also prevents transport connects from colliding with writes.
  71  */
  72 static int
  73 __xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
  74 {
  75         struct rpc_rqst *req = task->tk_rqstp;
  76
  77         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
  78                 if (task == xprt->snd_task)
  79                         return 1;
  80                 goto out_sleep;
  81         }
  82         if (xprt->nocong || __xprt_get_cong(xprt, task)) {
  83                 xprt->snd_task = task;
  84                 if (req) {
  85                         req->rq_bytes_sent = 0;
  86                         req->rq_ntrans++;
  87                 }
  88                 return 1;
  89         }
  90         smp_mb__before_clear_bit();
  91         clear_bit(XPRT_LOCKED, &xprt->state);
  92         smp_mb__after_clear_bit();
  93 out_sleep:
  94         dprintk("RPC: %4d failed to lock transport %p\n", task->tk_pid, xprt);
  95         task->tk_timeout = 0;
  96         task->tk_status = -EAGAIN;
  97         if (req && req->rq_ntrans)
  98                 rpc_sleep_on(&xprt->resend, task, NULL, NULL);
  99         else
 100                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
 101         return 0;
 102 }
 103
 104 static inline int
 105 xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 106 {
 107         int retval;
 108
 109         spin_lock_bh(&xprt->transport_lock);
 110         retval = __xprt_lock_write(xprt, task);
 111         spin_unlock_bh(&xprt->transport_lock);
 112         return retval;
 113 }
 114
 115
 116 static void
 117 __xprt_lock_write_next(struct rpc_xprt *xprt)
 118 {
 119         struct rpc_task *task;
 120
 121         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 122                 return;
 123         if (!xprt->nocong && RPCXPRT_CONGESTED(xprt))
 124                 goto out_unlock;
 125         task = rpc_wake_up_next(&xprt->resend);
 126         if (!task) {
 127                 task = rpc_wake_up_next(&xprt->sending);
 128                 if (!task)
 129                         goto out_unlock;
 130         }
 131         if (xprt->nocong || __xprt_get_cong(xprt, task)) {
 132                 struct rpc_rqst *req = task->tk_rqstp;
 133                 xprt->snd_task = task;
 134                 if (req) {
 135                         req->rq_bytes_sent = 0;
 136                         req->rq_ntrans++;
 137                 }
 138                 return;
 139         }
 140 out_unlock:
 141         smp_mb__before_clear_bit();
 142         clear_bit(XPRT_LOCKED, &xprt->state);
 143         smp_mb__after_clear_bit();
 144 }
 145
 146 /*
 147  * Releases the transport for use by other requests.
 148  */
 149 static void
 150 __xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 151 {
 152         if (xprt->snd_task == task) {
 153                 xprt->snd_task = NULL;
 154                 smp_mb__before_clear_bit();
 155                 clear_bit(XPRT_LOCKED, &xprt->state);
 156                 smp_mb__after_clear_bit();
 157                 __xprt_lock_write_next(xprt);
 158         }
 159 }
 160
 161 static inline void
 162 xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 163 {
 164         spin_lock_bh(&xprt->transport_lock);
 165         __xprt_release_write(xprt, task);
 166         spin_unlock_bh(&xprt->transport_lock);
 167 }
 168
 169 /*
 170  * Van Jacobson congestion avoidance. Check if the congestion window
 171  * overflowed. Put the task to sleep if this is the case.
 172  */
 173 static int
 174 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 175 {
 176         struct rpc_rqst *req = task->tk_rqstp;
 177
 178         if (req->rq_cong)
 179                 return 1;
 180         dprintk("RPC: %4d xprt_cwnd_limited cong = %ld cwnd = %ld\n",
 181                         task->tk_pid, xprt->cong, xprt->cwnd);
 182         if (RPCXPRT_CONGESTED(xprt))
 183                 return 0;
 184         req->rq_cong = 1;
 185         xprt->cong += RPC_CWNDSCALE;
 186         return 1;
 187 }
 188
 189 /*
 190  * Adjust the congestion window, and wake up the next task
 191  * that has been sleeping due to congestion
 192  */
 193 static void
 194 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 195 {
 196         if (!req->rq_cong)
 197                 return;
 198         req->rq_cong = 0;
 199         xprt->cong -= RPC_CWNDSCALE;
 200         __xprt_lock_write_next(xprt);
 201 }
 202
 203 /*
 204  * Adjust RPC congestion window
 205  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
 206  */
 207 static void
 208 xprt_adjust_cwnd(struct rpc_xprt *xprt, int result)
 209 {
 210         unsigned long   cwnd;
 211
 212         cwnd = xprt->cwnd;
 213         if (result >= 0 && cwnd <= xprt->cong) {
 214                 /* The (cwnd >> 1) term makes sure
 215                  * the result gets rounded properly. */
 216                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 217                 if (cwnd > RPC_MAXCWND(xprt))
 218                         cwnd = RPC_MAXCWND(xprt);
 219                 __xprt_lock_write_next(xprt);
 220         } else if (result == -ETIMEDOUT) {
 221                 cwnd >>= 1;
 222                 if (cwnd < RPC_CWNDSCALE)
 223                         cwnd = RPC_CWNDSCALE;
 224         }
 225         dprintk("RPC:      cong %ld, cwnd was %ld, now %ld\n",
 226                         xprt->cong, xprt->cwnd, cwnd);
 227         xprt->cwnd = cwnd;
 228 }
 229
 230 /**
 231  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
 232  * @xprt: transport with waiting tasks
 233  * @status: result code to plant in each task before waking it
 234  *
 235  */
 236 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
 237 {
 238         if (status < 0)
 239                 rpc_wake_up_status(&xprt->pending, status);
 240         else
 241                 rpc_wake_up(&xprt->pending);
 242 }
 243
 244 /**
 245  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
 246  * @task: task to be put to sleep
 247  *
 248  */
 249 void xprt_wait_for_buffer_space(struct rpc_task *task)
 250 {
 251         struct rpc_rqst *req = task->tk_rqstp;
 252         struct rpc_xprt *xprt = req->rq_xprt;
 253
 254         task->tk_timeout = req->rq_timeout;
 255         rpc_sleep_on(&xprt->pending, task, NULL, NULL);
 256 }
 257
 258 /**
 259  * xprt_write_space - wake the task waiting for transport output buffer space
 260  * @xprt: transport with waiting tasks
 261  *
 262  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
 263  */
 264 void xprt_write_space(struct rpc_xprt *xprt)
 265 {
 266         if (unlikely(xprt->shutdown))
 267                 return;
 268
 269         spin_lock_bh(&xprt->transport_lock);
 270         if (xprt->snd_task) {
 271                 dprintk("RPC:      write space: waking waiting task on xprt %p\n",
 272                                 xprt);
 273                 rpc_wake_up_task(xprt->snd_task);
 274         }
 275         spin_unlock_bh(&xprt->transport_lock);
 276 }
 277
 278 /**
 279  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
 280  * @task: task whose timeout is to be set
 281  *
 282  * Set a request's retransmit timeout based on the transport's
 283  * default timeout parameters.  Used by transports that don't adjust
 284  * the retransmit timeout based on round-trip time estimation.
 285  */
 286 void xprt_set_retrans_timeout_def(struct rpc_task *task)
 287 {
 288         task->tk_timeout = task->tk_rqstp->rq_timeout;
 289 }
 290
 291 /*
 292  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
 293  * @task: task whose timeout is to be set
 294  *
 295  * Set a request's retransmit timeout using the RTT estimator.
 296  */
 297 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
 298 {
 299         int timer = task->tk_msg.rpc_proc->p_timer;
 300         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
 301         struct rpc_rqst *req = task->tk_rqstp;
 302         unsigned long max_timeout = req->rq_xprt->timeout.to_maxval;
 303
 304         task->tk_timeout = rpc_calc_rto(rtt, timer);
 305         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
 306         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
 307                 task->tk_timeout = max_timeout;
 308 }
 309
 310 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 311 {
 312         struct rpc_timeout *to = &req->rq_xprt->timeout;
 313
 314         req->rq_majortimeo = req->rq_timeout;
 315         if (to->to_exponential)
 316                 req->rq_majortimeo <<= to->to_retries;
 317         else
 318                 req->rq_majortimeo += to->to_increment * to->to_retries;
 319         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 320                 req->rq_majortimeo = to->to_maxval;
 321         req->rq_majortimeo += jiffies;
 322 }
 323
 324 /**
 325  * xprt_adjust_timeout - adjust timeout values for next retransmit
 326  * @req: RPC request containing parameters to use for the adjustment
 327  *
 328  */
 329 int xprt_adjust_timeout(struct rpc_rqst *req)
 330 {
 331         struct rpc_xprt *xprt = req->rq_xprt;
 332         struct rpc_timeout *to = &xprt->timeout;
 333         int status = 0;
 334
 335         if (time_before(jiffies, req->rq_majortimeo)) {
 336                 if (to->to_exponential)
 337                         req->rq_timeout <<= 1;
 338                 else
 339                         req->rq_timeout += to->to_increment;
 340                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 341                         req->rq_timeout = to->to_maxval;
 342                 req->rq_retries++;
 343                 pprintk("RPC: %lu retrans\n", jiffies);
 344         } else {
 345                 req->rq_timeout = to->to_initval;
 346                 req->rq_retries = 0;
 347                 xprt_reset_majortimeo(req);
 348                 /* Reset the RTT counters == "slow start" */
 349                 spin_lock_bh(&xprt->transport_lock);
 350                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 351                 spin_unlock_bh(&xprt->transport_lock);
 352                 pprintk("RPC: %lu timeout\n", jiffies);
 353                 status = -ETIMEDOUT;
 354         }
 355
 356         if (req->rq_timeout == 0) {
 357                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 358                 req->rq_timeout = 5 * HZ;
 359         }
 360         return status;
 361 }
 362
 363 static void xprt_autoclose(void *args)
 364 {
 365         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
 366
 367         xprt_disconnect(xprt);
 368         xprt->ops->close(xprt);
 369         xprt_release_write(xprt, NULL);
 370 }
 371
 372 /**
 373  * xprt_disconnect - mark a transport as disconnected
 374  * @xprt: transport to flag for disconnect
 375  *
 376  */
 377 void xprt_disconnect(struct rpc_xprt *xprt)
 378 {
 379         dprintk("RPC:      disconnected transport %p\n", xprt);
 380         spin_lock_bh(&xprt->transport_lock);
 381         xprt_clear_connected(xprt);
 382         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 383         spin_unlock_bh(&xprt->transport_lock);
 384 }
 385
 386 static void
 387 xprt_init_autodisconnect(unsigned long data)
 388 {
 389         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 390
 391         spin_lock(&xprt->transport_lock);
 392         if (!list_empty(&xprt->recv) || xprt->shutdown)
 393                 goto out_abort;
 394         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 395                 goto out_abort;
 396         spin_unlock(&xprt->transport_lock);
 397         if (xprt_connecting(xprt))
 398                 xprt_release_write(xprt, NULL);
 399         else
 400                 schedule_work(&xprt->task_cleanup);
 401         return;
 402 out_abort:
 403         spin_unlock(&xprt->transport_lock);
 404 }
 405
 406 /**
 407  * xprt_connect - schedule a transport connect operation
 408  * @task: RPC task that is requesting the connect
 409  *
 410  */
 411 void xprt_connect(struct rpc_task *task)
 412 {
 413         struct rpc_xprt *xprt = task->tk_xprt;
 414
 415         dprintk("RPC: %4d xprt_connect xprt %p %s connected\n", task->tk_pid,
 416                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 417
 418         if (xprt->shutdown) {
 419                 task->tk_status = -EIO;
 420                 return;
 421         }
 422         if (!xprt->addr.sin_port) {
 423                 task->tk_status = -EIO;
 424                 return;
 425         }
 426         if (!xprt_lock_write(xprt, task))
 427                 return;
 428         if (xprt_connected(xprt))
 429                 xprt_release_write(xprt, task);
 430         else {
 431                 if (task->tk_rqstp)
 432                         task->tk_rqstp->rq_bytes_sent = 0;
 433
 434                 task->tk_timeout = RPC_CONNECT_TIMEOUT;
 435                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status, NULL);
 436                 xprt->ops->connect(task);
 437         }
 438         return;
 439 }
 440
 441 static void xprt_connect_status(struct rpc_task *task)
 442 {
 443         struct rpc_xprt *xprt = task->tk_xprt;
 444
 445         if (task->tk_status >= 0) {
 446                 dprintk("RPC: %4d xprt_connect_status: connection established\n",
 447                                 task->tk_pid);
 448                 return;
 449         }
 450
 451         switch (task->tk_status) {
 452         case -ECONNREFUSED:
 453         case -ECONNRESET:
 454                 dprintk("RPC: %4d xprt_connect_status: server %s refused connection\n",
 455                                 task->tk_pid, task->tk_client->cl_server);
 456                 break;
 457         case -ENOTCONN:
 458                 dprintk("RPC: %4d xprt_connect_status: connection broken\n",
 459                                 task->tk_pid);
 460                 break;
 461         case -ETIMEDOUT:
 462                 dprintk("RPC: %4d xprt_connect_status: connect attempt timed out\n",
 463                                 task->tk_pid);
 464                 break;
 465         default:
 466                 dprintk("RPC: %4d xprt_connect_status: error %d connecting to server %s\n",
 467                                 task->tk_pid, -task->tk_status, task->tk_client->cl_server);
 468                 xprt_release_write(xprt, task);
 469                 task->tk_status = -EIO;
 470                 return;
 471         }
 472
 473         /* if soft mounted, just cause this RPC to fail */
 474         if (RPC_IS_SOFT(task)) {
 475                 xprt_release_write(xprt, task);
 476                 task->tk_status = -EIO;
 477         }
 478 }
 479
 480 /**
 481  * xprt_lookup_rqst - find an RPC request corresponding to an XID
 482  * @xprt: transport on which the original request was transmitted
 483  * @xid: RPC XID of incoming reply
 484  *
 485  */
 486 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, u32 xid)
 487 {
 488         struct list_head *pos;
 489         struct rpc_rqst *req = NULL;
 490
 491         list_for_each(pos, &xprt->recv) {
 492                 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
 493                 if (entry->rq_xid == xid) {
 494                         req = entry;
 495                         break;
 496                 }
 497         }
 498         return req;
 499 }
 500
 501 /**
 502  * xprt_complete_rqst - called when reply processing is complete
 503  * @xprt: controlling transport
 504  * @req: RPC request that just completed
 505  * @copied: actual number of bytes received from the transport
 506  *
 507  */
 508 void xprt_complete_rqst(struct rpc_xprt *xprt, struct rpc_rqst *req, int copied)
 509 {
 510         struct rpc_task *task = req->rq_task;
 511         struct rpc_clnt *clnt = task->tk_client;
 512
 513         /* Adjust congestion window */
 514         if (!xprt->nocong) {
 515                 unsigned timer = task->tk_msg.rpc_proc->p_timer;
 516                 xprt_adjust_cwnd(xprt, copied);
 517                 __xprt_put_cong(xprt, req);
 518                 if (timer) {
 519                         if (req->rq_ntrans == 1)
 520                                 rpc_update_rtt(clnt->cl_rtt, timer,
 521                                                 (long)jiffies - req->rq_xtime);
 522                         rpc_set_timeo(clnt->cl_rtt, timer, req->rq_ntrans - 1);
 523                 }
 524         }
 525
 526 #ifdef RPC_PROFILE
 527         /* Profile only reads for now */
 528         if (copied > 1024) {
 529                 static unsigned long    nextstat;
 530                 static unsigned long    pkt_rtt, pkt_len, pkt_cnt;
 531
 532                 pkt_cnt++;
 533                 pkt_len += req->rq_slen + copied;
 534                 pkt_rtt += jiffies - req->rq_xtime;
 535                 if (time_before(nextstat, jiffies)) {
 536                         printk("RPC: %lu %ld cwnd\n", jiffies, xprt->cwnd);
 537                         printk("RPC: %ld %ld %ld %ld stat\n",
 538                                         jiffies, pkt_cnt, pkt_len, pkt_rtt);
 539                         pkt_rtt = pkt_len = pkt_cnt = 0;
 540                         nextstat = jiffies + 5 * HZ;
 541                 }
 542         }
 543 #endif
 544
 545         dprintk("RPC: %4d has input (%d bytes)\n", task->tk_pid, copied);
 546         list_del_init(&req->rq_list);
 547         req->rq_received = req->rq_private_buf.len = copied;
 548
 549         /* ... and wake up the process. */
 550         rpc_wake_up_task(task);
 551         return;
 552 }
 553
 554 /*
 555  * RPC receive timeout handler.
 556  */
 557 static void
 558 xprt_timer(struct rpc_task *task)
 559 {
 560         struct rpc_rqst *req = task->tk_rqstp;
 561         struct rpc_xprt *xprt = req->rq_xprt;
 562
 563         spin_lock(&xprt->transport_lock);
 564         if (req->rq_received)
 565                 goto out;
 566
 567         xprt_adjust_cwnd(req->rq_xprt, -ETIMEDOUT);
 568         __xprt_put_cong(xprt, req);
 569
 570         dprintk("RPC: %4d xprt_timer (%s request)\n",
 571                 task->tk_pid, req ? "pending" : "backlogged");
 572
 573         task->tk_status  = -ETIMEDOUT;
 574 out:
 575         task->tk_timeout = 0;
 576         rpc_wake_up_task(task);
 577         spin_unlock(&xprt->transport_lock);
 578 }
 579
 580 /**
 581  * xprt_prepare_transmit - reserve the transport before sending a request
 582  * @task: RPC task about to send a request
 583  *
 584  */
 585 int xprt_prepare_transmit(struct rpc_task *task)
 586 {
 587         struct rpc_rqst *req = task->tk_rqstp;
 588         struct rpc_xprt *xprt = req->rq_xprt;
 589         int err = 0;
 590
 591         dprintk("RPC: %4d xprt_prepare_transmit\n", task->tk_pid);
 592
 593         if (xprt->shutdown)
 594                 return -EIO;
 595
 596         spin_lock_bh(&xprt->transport_lock);
 597         if (req->rq_received && !req->rq_bytes_sent) {
 598                 err = req->rq_received;
 599                 goto out_unlock;
 600         }
 601         if (!__xprt_lock_write(xprt, task)) {
 602                 err = -EAGAIN;
 603                 goto out_unlock;
 604         }
 605
 606         if (!xprt_connected(xprt)) {
 607                 err = -ENOTCONN;
 608                 goto out_unlock;
 609         }
 610 out_unlock:
 611         spin_unlock_bh(&xprt->transport_lock);
 612         return err;
 613 }
 614
 615 /**
 616  * xprt_transmit - send an RPC request on a transport
 617  * @task: controlling RPC task
 618  *
 619  * We have to copy the iovec because sendmsg fiddles with its contents.
 620  */
 621 void xprt_transmit(struct rpc_task *task)
 622 {
 623         struct rpc_rqst *req = task->tk_rqstp;
 624         struct rpc_xprt *xprt = req->rq_xprt;
 625         int status;
 626
 627         dprintk("RPC: %4d xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
 628
 629         smp_rmb();
 630         if (!req->rq_received) {
 631                 if (list_empty(&req->rq_list)) {
 632                         spin_lock_bh(&xprt->transport_lock);
 633                         /* Update the softirq receive buffer */
 634                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
 635                                         sizeof(req->rq_private_buf));
 636                         /* Add request to the receive list */
 637                         list_add_tail(&req->rq_list, &xprt->recv);
 638                         spin_unlock_bh(&xprt->transport_lock);
 639                         xprt_reset_majortimeo(req);
 640                         /* Turn off autodisconnect */
 641                         del_singleshot_timer_sync(&xprt->timer);
 642                 }
 643         } else if (!req->rq_bytes_sent)
 644                 return;
 645
 646         status = xprt->ops->send_request(task);
 647         if (status == 0) {
 648                 dprintk("RPC: %4d xmit complete\n", task->tk_pid);
 649                 spin_lock_bh(&xprt->transport_lock);
 650                 xprt->ops->set_retrans_timeout(task);
 651                 /* Don't race with disconnect */
 652                 if (!xprt_connected(xprt))
 653                         task->tk_status = -ENOTCONN;
 654                 else if (!req->rq_received)
 655                         rpc_sleep_on(&xprt->pending, task, NULL, xprt_timer);
 656                 __xprt_release_write(xprt, task);
 657                 spin_unlock_bh(&xprt->transport_lock);
 658                 return;
 659         }
 660
 661         /* Note: at this point, task->tk_sleeping has not yet been set,
 662          *       hence there is no danger of the waking up task being put on
 663          *       schedq, and being picked up by a parallel run of rpciod().
 664          */
 665         task->tk_status = status;
 666
 667         switch (status) {
 668         case -ECONNREFUSED:
 669                 task->tk_timeout = RPC_REESTABLISH_TIMEOUT;
 670                 rpc_sleep_on(&xprt->sending, task, NULL, NULL);
 671         case -EAGAIN:
 672         case -ENOTCONN:
 673                 return;
 674         default:
 675                 break;
 676         }
 677         xprt_release_write(xprt, task);
 678         return;
 679 }
 680
 681 static inline void do_xprt_reserve(struct rpc_task *task)
 682 {
 683         struct rpc_xprt *xprt = task->tk_xprt;
 684
 685         task->tk_status = 0;
 686         if (task->tk_rqstp)
 687                 return;
 688         if (!list_empty(&xprt->free)) {
 689                 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
 690                 list_del_init(&req->rq_list);
 691                 task->tk_rqstp = req;
 692                 xprt_request_init(task, xprt);
 693                 return;
 694         }
 695         dprintk("RPC:      waiting for request slot\n");
 696         task->tk_status = -EAGAIN;
 697         task->tk_timeout = 0;
 698         rpc_sleep_on(&xprt->backlog, task, NULL, NULL);
 699 }
 700
 701 /**
 702  * xprt_reserve - allocate an RPC request slot
 703  * @task: RPC task requesting a slot allocation
 704  *
 705  * If no more slots are available, place the task on the transport's
 706  * backlog queue.
 707  */
 708 void xprt_reserve(struct rpc_task *task)
 709 {
 710         struct rpc_xprt *xprt = task->tk_xprt;
 711
 712         task->tk_status = -EIO;
 713         if (!xprt->shutdown) {
 714                 spin_lock(&xprt->reserve_lock);
 715                 do_xprt_reserve(task);
 716                 spin_unlock(&xprt->reserve_lock);
 717         }
 718 }
 719
 720 static inline u32 xprt_alloc_xid(struct rpc_xprt *xprt)
 721 {
 722         return xprt->xid++;
 723 }
 724
 725 static inline void xprt_init_xid(struct rpc_xprt *xprt)
 726 {
 727         get_random_bytes(&xprt->xid, sizeof(xprt->xid));
 728 }
 729
 730 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
 731 {
 732         struct rpc_rqst *req = task->tk_rqstp;
 733
 734         req->rq_timeout = xprt->timeout.to_initval;
 735         req->rq_task    = task;
 736         req->rq_xprt    = xprt;
 737         req->rq_xid     = xprt_alloc_xid(xprt);
 738         dprintk("RPC: %4d reserved req %p xid %08x\n", task->tk_pid,
 739                         req, ntohl(req->rq_xid));
 740 }
 741
 742 /**
 743  * xprt_release - release an RPC request slot
 744  * @task: task which is finished with the slot
 745  *
 746  */
 747 void xprt_release(struct rpc_task *task)
 748 {
 749         struct rpc_xprt *xprt = task->tk_xprt;
 750         struct rpc_rqst *req;
 751
 752         if (!(req = task->tk_rqstp))
 753                 return;
 754         spin_lock_bh(&xprt->transport_lock);
 755         __xprt_release_write(xprt, task);
 756         __xprt_put_cong(xprt, req);
 757         if (!list_empty(&req->rq_list))
 758                 list_del(&req->rq_list);
 759         xprt->last_used = jiffies;
 760         if (list_empty(&xprt->recv) && !xprt->shutdown)
 761                 mod_timer(&xprt->timer,
 762                                 xprt->last_used + RPC_IDLE_DISCONNECT_TIMEOUT);
 763         spin_unlock_bh(&xprt->transport_lock);
 764         task->tk_rqstp = NULL;
 765         memset(req, 0, sizeof(*req));   /* mark unused */
 766
 767         dprintk("RPC: %4d release request %p\n", task->tk_pid, req);
 768
 769         spin_lock(&xprt->reserve_lock);
 770         list_add(&req->rq_list, &xprt->free);
 771         xprt_clear_backlog(xprt);
 772         spin_unlock(&xprt->reserve_lock);
 773 }
 774
 775 /**
 776  * xprt_set_timeout - set constant RPC timeout
 777  * @to: RPC timeout parameters to set up
 778  * @retr: number of retries
 779  * @incr: amount of increase after each retry
 780  *
 781  */
 782 void xprt_set_timeout(struct rpc_timeout *to, unsigned int retr, unsigned long incr)
 783 {
 784         to->to_initval   =
 785         to->to_increment = incr;
 786         to->to_maxval    = to->to_initval + (incr * retr);
 787         to->to_retries   = retr;
 788         to->to_exponential = 0;
 789 }
 790
 791 static struct rpc_xprt *xprt_setup(int proto, struct sockaddr_in *ap, struct rpc_timeout *to)
 792 {
 793         int result;
 794         struct rpc_xprt *xprt;
 795         struct rpc_rqst *req;
 796
 797         if ((xprt = kmalloc(sizeof(struct rpc_xprt), GFP_KERNEL)) == NULL)
 798                 return ERR_PTR(-ENOMEM);
 799         memset(xprt, 0, sizeof(*xprt)); /* Nnnngh! */
 800
 801         xprt->addr = *ap;
 802
 803         switch (proto) {
 804         case IPPROTO_UDP:
 805                 result = xs_setup_udp(xprt, to);
 806                 break;
 807         case IPPROTO_TCP:
 808                 result = xs_setup_tcp(xprt, to);
 809                 break;
 810         default:
 811                 printk(KERN_ERR "RPC: unrecognized transport protocol: %d\n",
 812                                 proto);
 813                 result = -EIO;
 814                 break;
 815         }
 816         if (result) {
 817                 kfree(xprt);
 818                 return ERR_PTR(result);
 819         }
 820
 821         spin_lock_init(&xprt->transport_lock);
 822         spin_lock_init(&xprt->reserve_lock);
 823         init_waitqueue_head(&xprt->cong_wait);
 824
 825         INIT_LIST_HEAD(&xprt->free);
 826         INIT_LIST_HEAD(&xprt->recv);
 827         INIT_WORK(&xprt->task_cleanup, xprt_autoclose, xprt);
 828         init_timer(&xprt->timer);
 829         xprt->timer.function = xprt_init_autodisconnect;
 830         xprt->timer.data = (unsigned long) xprt;
 831         xprt->last_used = jiffies;
 832
 833         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
 834         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
 835         rpc_init_wait_queue(&xprt->resend, "xprt_resend");
 836         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
 837
 838         /* initialize free list */
 839         for (req = &xprt->slot[xprt->max_reqs-1]; req >= &xprt->slot[0]; req--)
 840                 list_add(&req->rq_list, &xprt->free);
 841
 842         xprt_init_xid(xprt);
 843
 844         dprintk("RPC:      created transport %p with %u slots\n", xprt,
 845                         xprt->max_reqs);
 846
 847         return xprt;
 848 }
 849
 850 /**
 851  * xprt_create_proto - create an RPC client transport
 852  * @proto: requested transport protocol
 853  * @sap: remote peer's address
 854  * @to: timeout parameters for new transport
 855  *
 856  */
 857 struct rpc_xprt *xprt_create_proto(int proto, struct sockaddr_in *sap, struct rpc_timeout *to)
 858 {
 859         struct rpc_xprt *xprt;
 860
 861         xprt = xprt_setup(proto, sap, to);
 862         if (IS_ERR(xprt))
 863                 dprintk("RPC:      xprt_create_proto failed\n");
 864         else
 865                 dprintk("RPC:      xprt_create_proto created xprt %p\n", xprt);
 866         return xprt;
 867 }
 868
 869 static void xprt_shutdown(struct rpc_xprt *xprt)
 870 {
 871         xprt->shutdown = 1;
 872         rpc_wake_up(&xprt->sending);
 873         rpc_wake_up(&xprt->resend);
 874         xprt_wake_pending_tasks(xprt, -EIO);
 875         rpc_wake_up(&xprt->backlog);
 876         wake_up(&xprt->cong_wait);
 877         del_timer_sync(&xprt->timer);
 878 }
 879
 880 static int xprt_clear_backlog(struct rpc_xprt *xprt) {
 881         rpc_wake_up_next(&xprt->backlog);
 882         wake_up(&xprt->cong_wait);
 883         return 1;
 884 }
 885
 886 /**
 887  * xprt_destroy - destroy an RPC transport, killing off all requests.
 888  * @xprt: transport to destroy
 889  *
 890  */
 891 int xprt_destroy(struct rpc_xprt *xprt)
 892 {
 893         dprintk("RPC:      destroying transport %p\n", xprt);
 894         xprt_shutdown(xprt);
 895         xprt->ops->destroy(xprt);
 896         kfree(xprt);
 897
 898         return 0;
 899 }