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

   1 /*
   2  *  linux/net/sunrpc/xprt.c
   3  *
   4  *  This is a generic RPC call interface supporting congestion avoidance,
   5  *  and asynchronous calls.
   6  *
   7  *  The interface works like this:
   8  *
   9  *  -   When a process places a call, it allocates a request slot if
  10  *      one is available. Otherwise, it sleeps on the backlog queue
  11  *      (xprt_reserve).
  12  *  -   Next, the caller puts together the RPC message, stuffs it into
  13  *      the request struct, and calls xprt_transmit().
  14  *  -   xprt_transmit sends the message and installs the caller on the
  15  *      transport's wait list. At the same time, it installs a timer that
  16  *      is run after the packet's timeout has expired.
  17  *  -   When a packet arrives, the data_ready handler walks the list of
  18  *      pending requests for that transport. If a matching XID is found, the
  19  *      caller is woken up, and the timer removed.
  20  *  -   When no reply arrives within the timeout interval, the timer is
  21  *      fired by the kernel and runs xprt_timer(). It either adjusts the
  22  *      timeout values (minor timeout) or wakes up the caller with a status
  23  *      of -ETIMEDOUT.
  24  *  -   When the caller receives a notification from RPC that a reply arrived,
  25  *      it should release the RPC slot, and process the reply.
  26  *      If the call timed out, it may choose to retry the operation by
  27  *      adjusting the initial timeout value, and simply calling rpc_call
  28  *      again.
  29  *
  30  *  Support for async RPC is done through a set of RPC-specific scheduling
  31  *  primitives that `transparently' work for processes as well as async
  32  *  tasks that rely on callbacks.
  33  *
  34  *  Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de>
  35  *
  36  *  Transport switch API copyright (C) 2005, Chuck Lever <cel@netapp.com>
  37  */
  38
  39 #include <linux/module.h>
  40
  41 #include <linux/types.h>
  42 #include <linux/interrupt.h>
  43 #include <linux/workqueue.h>
  44 #include <linux/net.h>
  45
  46 #include <linux/sunrpc/clnt.h>
  47 #include <linux/sunrpc/metrics.h>
  48
  49 /*
  50  * Local variables
  51  */
  52
  53 #ifdef RPC_DEBUG
  54 # define RPCDBG_FACILITY        RPCDBG_XPRT
  55 #endif
  56
  57 /*
  58  * Local functions
  59  */
  60 static void     xprt_request_init(struct rpc_task *, struct rpc_xprt *);
  61 static inline void      do_xprt_reserve(struct rpc_task *);
  62 static void     xprt_connect_status(struct rpc_task *task);
  63 static int      __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
  64
  65 static DEFINE_SPINLOCK(xprt_list_lock);
  66 static LIST_HEAD(xprt_list);
  67
  68 /*
  69  * The transport code maintains an estimate on the maximum number of out-
  70  * standing RPC requests, using a smoothed version of the congestion
  71  * avoidance implemented in 44BSD. This is basically the Van Jacobson
  72  * congestion algorithm: If a retransmit occurs, the congestion window is
  73  * halved; otherwise, it is incremented by 1/cwnd when
  74  *
  75  *      -       a reply is received and
  76  *      -       a full number of requests are outstanding and
  77  *      -       the congestion window hasn't been updated recently.
  78  */
  79 #define RPC_CWNDSHIFT           (8U)
  80 #define RPC_CWNDSCALE           (1U << RPC_CWNDSHIFT)
  81 #define RPC_INITCWND            RPC_CWNDSCALE
  82 #define RPC_MAXCWND(xprt)       ((xprt)->max_reqs << RPC_CWNDSHIFT)
  83
  84 #define RPCXPRT_CONGESTED(xprt) ((xprt)->cong >= (xprt)->cwnd)
  85
  86 /**
  87  * xprt_register_transport - register a transport implementation
  88  * @transport: transport to register
  89  *
  90  * If a transport implementation is loaded as a kernel module, it can
  91  * call this interface to make itself known to the RPC client.
  92  *
  93  * Returns:
  94  * 0:           transport successfully registered
  95  * -EEXIST:     transport already registered
  96  * -EINVAL:     transport module being unloaded
  97  */
  98 int xprt_register_transport(struct xprt_class *transport)
  99 {
 100         struct xprt_class *t;
 101         int result;
 102
 103         result = -EEXIST;
 104         spin_lock(&xprt_list_lock);
 105         list_for_each_entry(t, &xprt_list, list) {
 106                 /* don't register the same transport class twice */
 107                 if (t->ident == transport->ident)
 108                         goto out;
 109         }
 110
 111         list_add_tail(&transport->list, &xprt_list);
 112         printk(KERN_INFO "RPC: Registered %s transport module.\n",
 113                transport->name);
 114         result = 0;
 115
 116 out:
 117         spin_unlock(&xprt_list_lock);
 118         return result;
 119 }
 120 EXPORT_SYMBOL_GPL(xprt_register_transport);
 121
 122 /**
 123  * xprt_unregister_transport - unregister a transport implementation
 124  * @transport: transport to unregister
 125  *
 126  * Returns:
 127  * 0:           transport successfully unregistered
 128  * -ENOENT:     transport never registered
 129  */
 130 int xprt_unregister_transport(struct xprt_class *transport)
 131 {
 132         struct xprt_class *t;
 133         int result;
 134
 135         result = 0;
 136         spin_lock(&xprt_list_lock);
 137         list_for_each_entry(t, &xprt_list, list) {
 138                 if (t == transport) {
 139                         printk(KERN_INFO
 140                                 "RPC: Unregistered %s transport module.\n",
 141                                 transport->name);
 142                         list_del_init(&transport->list);
 143                         goto out;
 144                 }
 145         }
 146         result = -ENOENT;
 147
 148 out:
 149         spin_unlock(&xprt_list_lock);
 150         return result;
 151 }
 152 EXPORT_SYMBOL_GPL(xprt_unregister_transport);
 153
 154 /**
 155  * xprt_reserve_xprt - serialize write access to transports
 156  * @task: task that is requesting access to the transport
 157  *
 158  * This prevents mixing the payload of separate requests, and prevents
 159  * transport connects from colliding with writes.  No congestion control
 160  * is provided.
 161  */
 162 int xprt_reserve_xprt(struct rpc_task *task)
 163 {
 164         struct rpc_xprt *xprt = task->tk_xprt;
 165         struct rpc_rqst *req = task->tk_rqstp;
 166
 167         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 168                 if (task == xprt->snd_task)
 169                         return 1;
 170                 if (task == NULL)
 171                         return 0;
 172                 goto out_sleep;
 173         }
 174         xprt->snd_task = task;
 175         if (req) {
 176                 req->rq_bytes_sent = 0;
 177                 req->rq_ntrans++;
 178         }
 179         return 1;
 180
 181 out_sleep:
 182         dprintk("RPC: %5u failed to lock transport %p\n",
 183                         task->tk_pid, xprt);
 184         task->tk_timeout = 0;
 185         task->tk_status = -EAGAIN;
 186         if (req && req->rq_ntrans)
 187                 rpc_sleep_on(&xprt->resend, task, NULL);
 188         else
 189                 rpc_sleep_on(&xprt->sending, task, NULL);
 190         return 0;
 191 }
 192 EXPORT_SYMBOL_GPL(xprt_reserve_xprt);
 193
 194 static void xprt_clear_locked(struct rpc_xprt *xprt)
 195 {
 196         xprt->snd_task = NULL;
 197         if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state) || xprt->shutdown) {
 198                 smp_mb__before_clear_bit();
 199                 clear_bit(XPRT_LOCKED, &xprt->state);
 200                 smp_mb__after_clear_bit();
 201         } else
 202                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 203 }
 204
 205 /*
 206  * xprt_reserve_xprt_cong - serialize write access to transports
 207  * @task: task that is requesting access to the transport
 208  *
 209  * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
 210  * integrated into the decision of whether a request is allowed to be
 211  * woken up and given access to the transport.
 212  */
 213 int xprt_reserve_xprt_cong(struct rpc_task *task)
 214 {
 215         struct rpc_xprt *xprt = task->tk_xprt;
 216         struct rpc_rqst *req = task->tk_rqstp;
 217
 218         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 219                 if (task == xprt->snd_task)
 220                         return 1;
 221                 goto out_sleep;
 222         }
 223         if (__xprt_get_cong(xprt, task)) {
 224                 xprt->snd_task = task;
 225                 if (req) {
 226                         req->rq_bytes_sent = 0;
 227                         req->rq_ntrans++;
 228                 }
 229                 return 1;
 230         }
 231         xprt_clear_locked(xprt);
 232 out_sleep:
 233         dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
 234         task->tk_timeout = 0;
 235         task->tk_status = -EAGAIN;
 236         if (req && req->rq_ntrans)
 237                 rpc_sleep_on(&xprt->resend, task, NULL);
 238         else
 239                 rpc_sleep_on(&xprt->sending, task, NULL);
 240         return 0;
 241 }
 242 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
 243
 244 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 245 {
 246         int retval;
 247
 248         spin_lock_bh(&xprt->transport_lock);
 249         retval = xprt->ops->reserve_xprt(task);
 250         spin_unlock_bh(&xprt->transport_lock);
 251         return retval;
 252 }
 253
 254 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
 255 {
 256         struct rpc_task *task;
 257         struct rpc_rqst *req;
 258
 259         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 260                 return;
 261
 262         task = rpc_wake_up_next(&xprt->resend);
 263         if (!task) {
 264                 task = rpc_wake_up_next(&xprt->sending);
 265                 if (!task)
 266                         goto out_unlock;
 267         }
 268
 269         req = task->tk_rqstp;
 270         xprt->snd_task = task;
 271         if (req) {
 272                 req->rq_bytes_sent = 0;
 273                 req->rq_ntrans++;
 274         }
 275         return;
 276
 277 out_unlock:
 278         xprt_clear_locked(xprt);
 279 }
 280
 281 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
 282 {
 283         struct rpc_task *task;
 284
 285         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 286                 return;
 287         if (RPCXPRT_CONGESTED(xprt))
 288                 goto out_unlock;
 289         task = rpc_wake_up_next(&xprt->resend);
 290         if (!task) {
 291                 task = rpc_wake_up_next(&xprt->sending);
 292                 if (!task)
 293                         goto out_unlock;
 294         }
 295         if (__xprt_get_cong(xprt, task)) {
 296                 struct rpc_rqst *req = task->tk_rqstp;
 297                 xprt->snd_task = task;
 298                 if (req) {
 299                         req->rq_bytes_sent = 0;
 300                         req->rq_ntrans++;
 301                 }
 302                 return;
 303         }
 304 out_unlock:
 305         xprt_clear_locked(xprt);
 306 }
 307
 308 /**
 309  * xprt_release_xprt - allow other requests to use a transport
 310  * @xprt: transport with other tasks potentially waiting
 311  * @task: task that is releasing access to the transport
 312  *
 313  * Note that "task" can be NULL.  No congestion control is provided.
 314  */
 315 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 316 {
 317         if (xprt->snd_task == task) {
 318                 xprt_clear_locked(xprt);
 319                 __xprt_lock_write_next(xprt);
 320         }
 321 }
 322 EXPORT_SYMBOL_GPL(xprt_release_xprt);
 323
 324 /**
 325  * xprt_release_xprt_cong - allow other requests to use a transport
 326  * @xprt: transport with other tasks potentially waiting
 327  * @task: task that is releasing access to the transport
 328  *
 329  * Note that "task" can be NULL.  Another task is awoken to use the
 330  * transport if the transport's congestion window allows it.
 331  */
 332 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 333 {
 334         if (xprt->snd_task == task) {
 335                 xprt_clear_locked(xprt);
 336                 __xprt_lock_write_next_cong(xprt);
 337         }
 338 }
 339 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
 340
 341 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 342 {
 343         spin_lock_bh(&xprt->transport_lock);
 344         xprt->ops->release_xprt(xprt, task);
 345         spin_unlock_bh(&xprt->transport_lock);
 346 }
 347
 348 /*
 349  * Van Jacobson congestion avoidance. Check if the congestion window
 350  * overflowed. Put the task to sleep if this is the case.
 351  */
 352 static int
 353 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 354 {
 355         struct rpc_rqst *req = task->tk_rqstp;
 356
 357         if (req->rq_cong)
 358                 return 1;
 359         dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
 360                         task->tk_pid, xprt->cong, xprt->cwnd);
 361         if (RPCXPRT_CONGESTED(xprt))
 362                 return 0;
 363         req->rq_cong = 1;
 364         xprt->cong += RPC_CWNDSCALE;
 365         return 1;
 366 }
 367
 368 /*
 369  * Adjust the congestion window, and wake up the next task
 370  * that has been sleeping due to congestion
 371  */
 372 static void
 373 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 374 {
 375         if (!req->rq_cong)
 376                 return;
 377         req->rq_cong = 0;
 378         xprt->cong -= RPC_CWNDSCALE;
 379         __xprt_lock_write_next_cong(xprt);
 380 }
 381
 382 /**
 383  * xprt_release_rqst_cong - housekeeping when request is complete
 384  * @task: RPC request that recently completed
 385  *
 386  * Useful for transports that require congestion control.
 387  */
 388 void xprt_release_rqst_cong(struct rpc_task *task)
 389 {
 390         __xprt_put_cong(task->tk_xprt, task->tk_rqstp);
 391 }
 392 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
 393
 394 /**
 395  * xprt_adjust_cwnd - adjust transport congestion window
 396  * @task: recently completed RPC request used to adjust window
 397  * @result: result code of completed RPC request
 398  *
 399  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
 400  */
 401 void xprt_adjust_cwnd(struct rpc_task *task, int result)
 402 {
 403         struct rpc_rqst *req = task->tk_rqstp;
 404         struct rpc_xprt *xprt = task->tk_xprt;
 405         unsigned long cwnd = xprt->cwnd;
 406
 407         if (result >= 0 && cwnd <= xprt->cong) {
 408                 /* The (cwnd >> 1) term makes sure
 409                  * the result gets rounded properly. */
 410                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 411                 if (cwnd > RPC_MAXCWND(xprt))
 412                         cwnd = RPC_MAXCWND(xprt);
 413                 __xprt_lock_write_next_cong(xprt);
 414         } else if (result == -ETIMEDOUT) {
 415                 cwnd >>= 1;
 416                 if (cwnd < RPC_CWNDSCALE)
 417                         cwnd = RPC_CWNDSCALE;
 418         }
 419         dprintk("RPC:       cong %ld, cwnd was %ld, now %ld\n",
 420                         xprt->cong, xprt->cwnd, cwnd);
 421         xprt->cwnd = cwnd;
 422         __xprt_put_cong(xprt, req);
 423 }
 424 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
 425
 426 /**
 427  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
 428  * @xprt: transport with waiting tasks
 429  * @status: result code to plant in each task before waking it
 430  *
 431  */
 432 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
 433 {
 434         if (status < 0)
 435                 rpc_wake_up_status(&xprt->pending, status);
 436         else
 437                 rpc_wake_up(&xprt->pending);
 438 }
 439 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
 440
 441 /**
 442  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
 443  * @task: task to be put to sleep
 444  * @action: function pointer to be executed after wait
 445  */
 446 void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action)
 447 {
 448         struct rpc_rqst *req = task->tk_rqstp;
 449         struct rpc_xprt *xprt = req->rq_xprt;
 450
 451         task->tk_timeout = req->rq_timeout;
 452         rpc_sleep_on(&xprt->pending, task, action);
 453 }
 454 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
 455
 456 /**
 457  * xprt_write_space - wake the task waiting for transport output buffer space
 458  * @xprt: transport with waiting tasks
 459  *
 460  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
 461  */
 462 void xprt_write_space(struct rpc_xprt *xprt)
 463 {
 464         if (unlikely(xprt->shutdown))
 465                 return;
 466
 467         spin_lock_bh(&xprt->transport_lock);
 468         if (xprt->snd_task) {
 469                 dprintk("RPC:       write space: waking waiting task on "
 470                                 "xprt %p\n", xprt);
 471                 rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task);
 472         }
 473         spin_unlock_bh(&xprt->transport_lock);
 474 }
 475 EXPORT_SYMBOL_GPL(xprt_write_space);
 476
 477 /**
 478  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
 479  * @task: task whose timeout is to be set
 480  *
 481  * Set a request's retransmit timeout based on the transport's
 482  * default timeout parameters.  Used by transports that don't adjust
 483  * the retransmit timeout based on round-trip time estimation.
 484  */
 485 void xprt_set_retrans_timeout_def(struct rpc_task *task)
 486 {
 487         task->tk_timeout = task->tk_rqstp->rq_timeout;
 488 }
 489 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
 490
 491 /*
 492  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
 493  * @task: task whose timeout is to be set
 494  *
 495  * Set a request's retransmit timeout using the RTT estimator.
 496  */
 497 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
 498 {
 499         int timer = task->tk_msg.rpc_proc->p_timer;
 500         struct rpc_clnt *clnt = task->tk_client;
 501         struct rpc_rtt *rtt = clnt->cl_rtt;
 502         struct rpc_rqst *req = task->tk_rqstp;
 503         unsigned long max_timeout = clnt->cl_timeout->to_maxval;
 504
 505         task->tk_timeout = rpc_calc_rto(rtt, timer);
 506         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
 507         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
 508                 task->tk_timeout = max_timeout;
 509 }
 510 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
 511
 512 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 513 {
 514         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 515
 516         req->rq_majortimeo = req->rq_timeout;
 517         if (to->to_exponential)
 518                 req->rq_majortimeo <<= to->to_retries;
 519         else
 520                 req->rq_majortimeo += to->to_increment * to->to_retries;
 521         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 522                 req->rq_majortimeo = to->to_maxval;
 523         req->rq_majortimeo += jiffies;
 524 }
 525
 526 /**
 527  * xprt_adjust_timeout - adjust timeout values for next retransmit
 528  * @req: RPC request containing parameters to use for the adjustment
 529  *
 530  */
 531 int xprt_adjust_timeout(struct rpc_rqst *req)
 532 {
 533         struct rpc_xprt *xprt = req->rq_xprt;
 534         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 535         int status = 0;
 536
 537         if (time_before(jiffies, req->rq_majortimeo)) {
 538                 if (to->to_exponential)
 539                         req->rq_timeout <<= 1;
 540                 else
 541                         req->rq_timeout += to->to_increment;
 542                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 543                         req->rq_timeout = to->to_maxval;
 544                 req->rq_retries++;
 545         } else {
 546                 req->rq_timeout = to->to_initval;
 547                 req->rq_retries = 0;
 548                 xprt_reset_majortimeo(req);
 549                 /* Reset the RTT counters == "slow start" */
 550                 spin_lock_bh(&xprt->transport_lock);
 551                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 552                 spin_unlock_bh(&xprt->transport_lock);
 553                 status = -ETIMEDOUT;
 554         }
 555
 556         if (req->rq_timeout == 0) {
 557                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 558                 req->rq_timeout = 5 * HZ;
 559         }
 560         return status;
 561 }
 562
 563 static void xprt_autoclose(struct work_struct *work)
 564 {
 565         struct rpc_xprt *xprt =
 566                 container_of(work, struct rpc_xprt, task_cleanup);
 567
 568         xprt->ops->close(xprt);
 569         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
 570         xprt_release_write(xprt, NULL);
 571 }
 572
 573 /**
 574  * xprt_disconnect_done - mark a transport as disconnected
 575  * @xprt: transport to flag for disconnect
 576  *
 577  */
 578 void xprt_disconnect_done(struct rpc_xprt *xprt)
 579 {
 580         dprintk("RPC:       disconnected transport %p\n", xprt);
 581         spin_lock_bh(&xprt->transport_lock);
 582         xprt_clear_connected(xprt);
 583         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 584         spin_unlock_bh(&xprt->transport_lock);
 585 }
 586 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
 587
 588 /**
 589  * xprt_force_disconnect - force a transport to disconnect
 590  * @xprt: transport to disconnect
 591  *
 592  */
 593 void xprt_force_disconnect(struct rpc_xprt *xprt)
 594 {
 595         /* Don't race with the test_bit() in xprt_clear_locked() */
 596         spin_lock_bh(&xprt->transport_lock);
 597         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 598         /* Try to schedule an autoclose RPC call */
 599         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 600                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 601         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 602         spin_unlock_bh(&xprt->transport_lock);
 603 }
 604
 605 /**
 606  * xprt_conditional_disconnect - force a transport to disconnect
 607  * @xprt: transport to disconnect
 608  * @cookie: 'connection cookie'
 609  *
 610  * This attempts to break the connection if and only if 'cookie' matches
 611  * the current transport 'connection cookie'. It ensures that we don't
 612  * try to break the connection more than once when we need to retransmit
 613  * a batch of RPC requests.
 614  *
 615  */
 616 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
 617 {
 618         /* Don't race with the test_bit() in xprt_clear_locked() */
 619         spin_lock_bh(&xprt->transport_lock);
 620         if (cookie != xprt->connect_cookie)
 621                 goto out;
 622         if (test_bit(XPRT_CLOSING, &xprt->state) || !xprt_connected(xprt))
 623                 goto out;
 624         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 625         /* Try to schedule an autoclose RPC call */
 626         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 627                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 628         xprt_wake_pending_tasks(xprt, -ENOTCONN);
 629 out:
 630         spin_unlock_bh(&xprt->transport_lock);
 631 }
 632
 633 static void
 634 xprt_init_autodisconnect(unsigned long data)
 635 {
 636         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 637
 638         spin_lock(&xprt->transport_lock);
 639         if (!list_empty(&xprt->recv) || xprt->shutdown)
 640                 goto out_abort;
 641         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 642                 goto out_abort;
 643         spin_unlock(&xprt->transport_lock);
 644         if (xprt_connecting(xprt))
 645                 xprt_release_write(xprt, NULL);
 646         else
 647                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 648         return;
 649 out_abort:
 650         spin_unlock(&xprt->transport_lock);
 651 }
 652
 653 /**
 654  * xprt_connect - schedule a transport connect operation
 655  * @task: RPC task that is requesting the connect
 656  *
 657  */
 658 void xprt_connect(struct rpc_task *task)
 659 {
 660         struct rpc_xprt *xprt = task->tk_xprt;
 661
 662         dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
 663                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 664
 665         if (!xprt_bound(xprt)) {
 666                 task->tk_status = -EAGAIN;
 667                 return;
 668         }
 669         if (!xprt_lock_write(xprt, task))
 670                 return;
 671         if (xprt_connected(xprt))
 672                 xprt_release_write(xprt, task);
 673         else {
 674                 if (task->tk_rqstp)
 675                         task->tk_rqstp->rq_bytes_sent = 0;
 676
 677                 task->tk_timeout = xprt->connect_timeout;
 678                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
 679                 xprt->stat.connect_start = jiffies;
 680                 xprt->ops->connect(task);
 681         }
 682         return;
 683 }
 684
 685 static void xprt_connect_status(struct rpc_task *task)
 686 {
 687         struct rpc_xprt *xprt = task->tk_xprt;
 688
 689         if (task->tk_status == 0) {
 690                 xprt->stat.connect_count++;
 691                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
 692                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
 693                                 task->tk_pid);
 694                 return;
 695         }
 696
 697         switch (task->tk_status) {
 698         case -ENOTCONN:
 699                 dprintk("RPC: %5u xprt_connect_status: connection broken\n",
 700                                 task->tk_pid);
 701                 break;
 702         case -ETIMEDOUT:
 703                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
 704                                 "out\n", task->tk_pid);
 705                 break;
 706         default:
 707                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
 708                                 "server %s\n", task->tk_pid, -task->tk_status,
 709                                 task->tk_client->cl_server);
 710                 xprt_release_write(xprt, task);
 711                 task->tk_status = -EIO;
 712         }
 713 }
 714
 715 /**
 716  * xprt_lookup_rqst - find an RPC request corresponding to an XID
 717  * @xprt: transport on which the original request was transmitted
 718  * @xid: RPC XID of incoming reply
 719  *
 720  */
 721 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
 722 {
 723         struct list_head *pos;
 724
 725         list_for_each(pos, &xprt->recv) {
 726                 struct rpc_rqst *entry = list_entry(pos, struct rpc_rqst, rq_list);
 727                 if (entry->rq_xid == xid)
 728                         return entry;
 729         }
 730
 731         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
 732                         ntohl(xid));
 733         xprt->stat.bad_xids++;
 734         return NULL;
 735 }
 736 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
 737
 738 /**
 739  * xprt_update_rtt - update an RPC client's RTT state after receiving a reply
 740  * @task: RPC request that recently completed
 741  *
 742  */
 743 void xprt_update_rtt(struct rpc_task *task)
 744 {
 745         struct rpc_rqst *req = task->tk_rqstp;
 746         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
 747         unsigned timer = task->tk_msg.rpc_proc->p_timer;
 748
 749         if (timer) {
 750                 if (req->rq_ntrans == 1)
 751                         rpc_update_rtt(rtt, timer,
 752                                         (long)jiffies - req->rq_xtime);
 753                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
 754         }
 755 }
 756 EXPORT_SYMBOL_GPL(xprt_update_rtt);
 757
 758 /**
 759  * xprt_complete_rqst - called when reply processing is complete
 760  * @task: RPC request that recently completed
 761  * @copied: actual number of bytes received from the transport
 762  *
 763  * Caller holds transport lock.
 764  */
 765 void xprt_complete_rqst(struct rpc_task *task, int copied)
 766 {
 767         struct rpc_rqst *req = task->tk_rqstp;
 768         struct rpc_xprt *xprt = req->rq_xprt;
 769
 770         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
 771                         task->tk_pid, ntohl(req->rq_xid), copied);
 772
 773         xprt->stat.recvs++;
 774         task->tk_rtt = (long)jiffies - req->rq_xtime;
 775
 776         list_del_init(&req->rq_list);
 777         req->rq_private_buf.len = copied;
 778         /* Ensure all writes are done before we update req->rq_received */
 779         smp_wmb();
 780         req->rq_received = copied;
 781         rpc_wake_up_queued_task(&xprt->pending, task);
 782 }
 783 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
 784
 785 static void xprt_timer(struct rpc_task *task)
 786 {
 787         struct rpc_rqst *req = task->tk_rqstp;
 788         struct rpc_xprt *xprt = req->rq_xprt;
 789
 790         if (task->tk_status != -ETIMEDOUT)
 791                 return;
 792         dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
 793
 794         spin_lock_bh(&xprt->transport_lock);
 795         if (!req->rq_received) {
 796                 if (xprt->ops->timer)
 797                         xprt->ops->timer(task);
 798         } else
 799                 task->tk_status = 0;
 800         spin_unlock_bh(&xprt->transport_lock);
 801 }
 802
 803 /**
 804  * xprt_prepare_transmit - reserve the transport before sending a request
 805  * @task: RPC task about to send a request
 806  *
 807  */
 808 int xprt_prepare_transmit(struct rpc_task *task)
 809 {
 810         struct rpc_rqst *req = task->tk_rqstp;
 811         struct rpc_xprt *xprt = req->rq_xprt;
 812         int err = 0;
 813
 814         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
 815
 816         spin_lock_bh(&xprt->transport_lock);
 817         if (req->rq_received && !req->rq_bytes_sent) {
 818                 err = req->rq_received;
 819                 goto out_unlock;
 820         }
 821         if (!xprt->ops->reserve_xprt(task)) {
 822                 err = -EAGAIN;
 823                 goto out_unlock;
 824         }
 825
 826         if (!xprt_connected(xprt)) {
 827                 err = -ENOTCONN;
 828                 goto out_unlock;
 829         }
 830 out_unlock:
 831         spin_unlock_bh(&xprt->transport_lock);
 832         return err;
 833 }
 834
 835 void xprt_end_transmit(struct rpc_task *task)
 836 {
 837         xprt_release_write(task->tk_xprt, task);
 838 }
 839
 840 /**
 841  * xprt_transmit - send an RPC request on a transport
 842  * @task: controlling RPC task
 843  *
 844  * We have to copy the iovec because sendmsg fiddles with its contents.
 845  */
 846 void xprt_transmit(struct rpc_task *task)
 847 {
 848         struct rpc_rqst *req = task->tk_rqstp;
 849         struct rpc_xprt *xprt = req->rq_xprt;
 850         int status;
 851
 852         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
 853
 854         if (!req->rq_received) {
 855                 if (list_empty(&req->rq_list)) {
 856                         spin_lock_bh(&xprt->transport_lock);
 857                         /* Update the softirq receive buffer */
 858                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
 859                                         sizeof(req->rq_private_buf));
 860                         /* Add request to the receive list */
 861                         list_add_tail(&req->rq_list, &xprt->recv);
 862                         spin_unlock_bh(&xprt->transport_lock);
 863                         xprt_reset_majortimeo(req);
 864                         /* Turn off autodisconnect */
 865                         del_singleshot_timer_sync(&xprt->timer);
 866                 }
 867         } else if (!req->rq_bytes_sent)
 868                 return;
 869
 870         req->rq_connect_cookie = xprt->connect_cookie;
 871         req->rq_xtime = jiffies;
 872         status = xprt->ops->send_request(task);
 873         if (status == 0) {
 874                 dprintk("RPC: %5u xmit complete\n", task->tk_pid);
 875                 spin_lock_bh(&xprt->transport_lock);
 876
 877                 xprt->ops->set_retrans_timeout(task);
 878
 879                 xprt->stat.sends++;
 880                 xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
 881                 xprt->stat.bklog_u += xprt->backlog.qlen;
 882
 883                 /* Don't race with disconnect */
 884                 if (!xprt_connected(xprt))
 885                         task->tk_status = -ENOTCONN;
 886                 else if (!req->rq_received)
 887                         rpc_sleep_on(&xprt->pending, task, xprt_timer);
 888                 spin_unlock_bh(&xprt->transport_lock);
 889                 return;
 890         }
 891
 892         /* Note: at this point, task->tk_sleeping has not yet been set,
 893          *       hence there is no danger of the waking up task being put on
 894          *       schedq, and being picked up by a parallel run of rpciod().
 895          */
 896         task->tk_status = status;
 897         if (status == -ECONNREFUSED)
 898                 rpc_sleep_on(&xprt->sending, task, NULL);
 899 }
 900
 901 static inline void do_xprt_reserve(struct rpc_task *task)
 902 {
 903         struct rpc_xprt *xprt = task->tk_xprt;
 904
 905         task->tk_status = 0;
 906         if (task->tk_rqstp)
 907                 return;
 908         if (!list_empty(&xprt->free)) {
 909                 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
 910                 list_del_init(&req->rq_list);
 911                 task->tk_rqstp = req;
 912                 xprt_request_init(task, xprt);
 913                 return;
 914         }
 915         dprintk("RPC:       waiting for request slot\n");
 916         task->tk_status = -EAGAIN;
 917         task->tk_timeout = 0;
 918         rpc_sleep_on(&xprt->backlog, task, NULL);
 919 }
 920
 921 /**
 922  * xprt_reserve - allocate an RPC request slot
 923  * @task: RPC task requesting a slot allocation
 924  *
 925  * If no more slots are available, place the task on the transport's
 926  * backlog queue.
 927  */
 928 void xprt_reserve(struct rpc_task *task)
 929 {
 930         struct rpc_xprt *xprt = task->tk_xprt;
 931
 932         task->tk_status = -EIO;
 933         spin_lock(&xprt->reserve_lock);
 934         do_xprt_reserve(task);
 935         spin_unlock(&xprt->reserve_lock);
 936 }
 937
 938 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
 939 {
 940         return xprt->xid++;
 941 }
 942
 943 static inline void xprt_init_xid(struct rpc_xprt *xprt)
 944 {
 945         xprt->xid = net_random();
 946 }
 947
 948 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
 949 {
 950         struct rpc_rqst *req = task->tk_rqstp;
 951
 952         req->rq_timeout = task->tk_client->cl_timeout->to_initval;
 953         req->rq_task    = task;
 954         req->rq_xprt    = xprt;
 955         req->rq_buffer  = NULL;
 956         req->rq_xid     = xprt_alloc_xid(xprt);
 957         req->rq_release_snd_buf = NULL;
 958         xprt_reset_majortimeo(req);
 959         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
 960                         req, ntohl(req->rq_xid));
 961 }
 962
 963 /**
 964  * xprt_release - release an RPC request slot
 965  * @task: task which is finished with the slot
 966  *
 967  */
 968 void xprt_release(struct rpc_task *task)
 969 {
 970         struct rpc_xprt *xprt = task->tk_xprt;
 971         struct rpc_rqst *req;
 972
 973         if (!(req = task->tk_rqstp))
 974                 return;
 975         rpc_count_iostats(task);
 976         spin_lock_bh(&xprt->transport_lock);
 977         xprt->ops->release_xprt(xprt, task);
 978         if (xprt->ops->release_request)
 979                 xprt->ops->release_request(task);
 980         if (!list_empty(&req->rq_list))
 981                 list_del(&req->rq_list);
 982         xprt->last_used = jiffies;
 983         if (list_empty(&xprt->recv))
 984                 mod_timer(&xprt->timer,
 985                                 xprt->last_used + xprt->idle_timeout);
 986         spin_unlock_bh(&xprt->transport_lock);
 987         xprt->ops->buf_free(req->rq_buffer);
 988         task->tk_rqstp = NULL;
 989         if (req->rq_release_snd_buf)
 990                 req->rq_release_snd_buf(req);
 991         memset(req, 0, sizeof(*req));   /* mark unused */
 992
 993         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
 994
 995         spin_lock(&xprt->reserve_lock);
 996         list_add(&req->rq_list, &xprt->free);
 997         rpc_wake_up_next(&xprt->backlog);
 998         spin_unlock(&xprt->reserve_lock);
 999 }
1000
1001 /**
1002  * xprt_create_transport - create an RPC transport
1003  * @args: rpc transport creation arguments
1004  *
1005  */
1006 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1007 {
1008         struct rpc_xprt *xprt;
1009         struct rpc_rqst *req;
1010         struct xprt_class *t;
1011
1012         spin_lock(&xprt_list_lock);
1013         list_for_each_entry(t, &xprt_list, list) {
1014                 if (t->ident == args->ident) {
1015                         spin_unlock(&xprt_list_lock);
1016                         goto found;
1017                 }
1018         }
1019         spin_unlock(&xprt_list_lock);
1020         printk(KERN_ERR "RPC: transport (%d) not supported\n", args->ident);
1021         return ERR_PTR(-EIO);
1022
1023 found:
1024         xprt = t->setup(args);
1025         if (IS_ERR(xprt)) {
1026                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1027                                 -PTR_ERR(xprt));
1028                 return xprt;
1029         }
1030
1031         kref_init(&xprt->kref);
1032         spin_lock_init(&xprt->transport_lock);
1033         spin_lock_init(&xprt->reserve_lock);
1034
1035         INIT_LIST_HEAD(&xprt->free);
1036         INIT_LIST_HEAD(&xprt->recv);
1037         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1038         setup_timer(&xprt->timer, xprt_init_autodisconnect,
1039                         (unsigned long)xprt);
1040         xprt->last_used = jiffies;
1041         xprt->cwnd = RPC_INITCWND;
1042         xprt->bind_index = 0;
1043
1044         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1045         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1046         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
1047         rpc_init_wait_queue(&xprt->resend, "xprt_resend");
1048         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1049
1050         /* initialize free list */
1051         for (req = &xprt->slot[xprt->max_reqs-1]; req >= &xprt->slot[0]; req--)
1052                 list_add(&req->rq_list, &xprt->free);
1053
1054         xprt_init_xid(xprt);
1055
1056         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1057                         xprt->max_reqs);
1058
1059         return xprt;
1060 }
1061
1062 /**
1063  * xprt_destroy - destroy an RPC transport, killing off all requests.
1064  * @kref: kref for the transport to destroy
1065  *
1066  */
1067 static void xprt_destroy(struct kref *kref)
1068 {
1069         struct rpc_xprt *xprt = container_of(kref, struct rpc_xprt, kref);
1070
1071         dprintk("RPC:       destroying transport %p\n", xprt);
1072         xprt->shutdown = 1;
1073         del_timer_sync(&xprt->timer);
1074
1075         rpc_destroy_wait_queue(&xprt->binding);
1076         rpc_destroy_wait_queue(&xprt->pending);
1077         rpc_destroy_wait_queue(&xprt->sending);
1078         rpc_destroy_wait_queue(&xprt->resend);
1079         rpc_destroy_wait_queue(&xprt->backlog);
1080         /*
1081          * Tear down transport state and free the rpc_xprt
1082          */
1083         xprt->ops->destroy(xprt);
1084 }
1085
1086 /**
1087  * xprt_put - release a reference to an RPC transport.
1088  * @xprt: pointer to the transport
1089  *
1090  */
1091 void xprt_put(struct rpc_xprt *xprt)
1092 {
1093         kref_put(&xprt->kref, xprt_destroy);
1094 }
1095
1096 /**
1097  * xprt_get - return a reference to an RPC transport.
1098  * @xprt: pointer to the transport
1099  *
1100  */
1101 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1102 {
1103         kref_get(&xprt->kref);
1104         return xprt;
1105 }