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, if a reply is expected,
  16  *      it installs a timer that is run after the packet's timeout has
  17  *      expired.
  18  *  -   When a packet arrives, the data_ready handler walks the list of
  19  *      pending requests for that transport. If a matching XID is found, the
  20  *      caller is woken up, and the timer removed.
  21  *  -   When no reply arrives within the timeout interval, the timer is
  22  *      fired by the kernel and runs xprt_timer(). It either adjusts the
  23  *      timeout values (minor timeout) or wakes up the caller with a status
  24  *      of -ETIMEDOUT.
  25  *  -   When the caller receives a notification from RPC that a reply arrived,
  26  *      it should release the RPC slot, and process the reply.
  27  *      If the call timed out, it may choose to retry the operation by
  28  *      adjusting the initial timeout value, and simply calling rpc_call
  29  *      again.
  30  *
  31  *  Support for async RPC is done through a set of RPC-specific scheduling
  32  *  primitives that `transparently' work for processes as well as async
  33  *  tasks that rely on callbacks.
  34  *
  35  *  Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de>
  36  *
  37  *  Transport switch API copyright (C) 2005, Chuck Lever <cel@netapp.com>
  38  */
  39
  40 #include <linux/module.h>
  41
  42 #include <linux/types.h>
  43 #include <linux/interrupt.h>
  44 #include <linux/workqueue.h>
  45 #include <linux/net.h>
  46 #include <linux/ktime.h>
  47
  48 #include <linux/sunrpc/clnt.h>
  49 #include <linux/sunrpc/metrics.h>
  50 #include <linux/sunrpc/bc_xprt.h>
  51
  52 #include "sunrpc.h"
  53
  54 /*
  55  * Local variables
  56  */
  57
  58 #ifdef RPC_DEBUG
  59 # define RPCDBG_FACILITY        RPCDBG_XPRT
  60 #endif
  61
  62 /*
  63  * Local functions
  64  */
  65 static void      xprt_init(struct rpc_xprt *xprt, struct net *net);
  66 static void     xprt_request_init(struct rpc_task *, struct rpc_xprt *);
  67 static void     xprt_connect_status(struct rpc_task *task);
  68 static int      __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
  69 static void      xprt_destroy(struct rpc_xprt *xprt);
  70
  71 static DEFINE_SPINLOCK(xprt_list_lock);
  72 static LIST_HEAD(xprt_list);
  73
  74 /*
  75  * The transport code maintains an estimate on the maximum number of out-
  76  * standing RPC requests, using a smoothed version of the congestion
  77  * avoidance implemented in 44BSD. This is basically the Van Jacobson
  78  * congestion algorithm: If a retransmit occurs, the congestion window is
  79  * halved; otherwise, it is incremented by 1/cwnd when
  80  *
  81  *      -       a reply is received and
  82  *      -       a full number of requests are outstanding and
  83  *      -       the congestion window hasn't been updated recently.
  84  */
  85 #define RPC_CWNDSHIFT           (8U)
  86 #define RPC_CWNDSCALE           (1U << RPC_CWNDSHIFT)
  87 #define RPC_INITCWND            RPC_CWNDSCALE
  88 #define RPC_MAXCWND(xprt)       ((xprt)->max_reqs << RPC_CWNDSHIFT)
  89
  90 #define RPCXPRT_CONGESTED(xprt) ((xprt)->cong >= (xprt)->cwnd)
  91
  92 /**
  93  * xprt_register_transport - register a transport implementation
  94  * @transport: transport to register
  95  *
  96  * If a transport implementation is loaded as a kernel module, it can
  97  * call this interface to make itself known to the RPC client.
  98  *
  99  * Returns:
 100  * 0:           transport successfully registered
 101  * -EEXIST:     transport already registered
 102  * -EINVAL:     transport module being unloaded
 103  */
 104 int xprt_register_transport(struct xprt_class *transport)
 105 {
 106         struct xprt_class *t;
 107         int result;
 108
 109         result = -EEXIST;
 110         spin_lock(&xprt_list_lock);
 111         list_for_each_entry(t, &xprt_list, list) {
 112                 /* don't register the same transport class twice */
 113                 if (t->ident == transport->ident)
 114                         goto out;
 115         }
 116
 117         list_add_tail(&transport->list, &xprt_list);
 118         printk(KERN_INFO "RPC: Registered %s transport module.\n",
 119                transport->name);
 120         result = 0;
 121
 122 out:
 123         spin_unlock(&xprt_list_lock);
 124         return result;
 125 }
 126 EXPORT_SYMBOL_GPL(xprt_register_transport);
 127
 128 /**
 129  * xprt_unregister_transport - unregister a transport implementation
 130  * @transport: transport to unregister
 131  *
 132  * Returns:
 133  * 0:           transport successfully unregistered
 134  * -ENOENT:     transport never registered
 135  */
 136 int xprt_unregister_transport(struct xprt_class *transport)
 137 {
 138         struct xprt_class *t;
 139         int result;
 140
 141         result = 0;
 142         spin_lock(&xprt_list_lock);
 143         list_for_each_entry(t, &xprt_list, list) {
 144                 if (t == transport) {
 145                         printk(KERN_INFO
 146                                 "RPC: Unregistered %s transport module.\n",
 147                                 transport->name);
 148                         list_del_init(&transport->list);
 149                         goto out;
 150                 }
 151         }
 152         result = -ENOENT;
 153
 154 out:
 155         spin_unlock(&xprt_list_lock);
 156         return result;
 157 }
 158 EXPORT_SYMBOL_GPL(xprt_unregister_transport);
 159
 160 /**
 161  * xprt_load_transport - load a transport implementation
 162  * @transport_name: transport to load
 163  *
 164  * Returns:
 165  * 0:           transport successfully loaded
 166  * -ENOENT:     transport module not available
 167  */
 168 int xprt_load_transport(const char *transport_name)
 169 {
 170         struct xprt_class *t;
 171         int result;
 172
 173         result = 0;
 174         spin_lock(&xprt_list_lock);
 175         list_for_each_entry(t, &xprt_list, list) {
 176                 if (strcmp(t->name, transport_name) == 0) {
 177                         spin_unlock(&xprt_list_lock);
 178                         goto out;
 179                 }
 180         }
 181         spin_unlock(&xprt_list_lock);
 182         result = request_module("xprt%s", transport_name);
 183 out:
 184         return result;
 185 }
 186 EXPORT_SYMBOL_GPL(xprt_load_transport);
 187
 188 /**
 189  * xprt_reserve_xprt - serialize write access to transports
 190  * @task: task that is requesting access to the transport
 191  * @xprt: pointer to the target transport
 192  *
 193  * This prevents mixing the payload of separate requests, and prevents
 194  * transport connects from colliding with writes.  No congestion control
 195  * is provided.
 196  */
 197 int xprt_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 198 {
 199         struct rpc_rqst *req = task->tk_rqstp;
 200         int priority;
 201
 202         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 203                 if (task == xprt->snd_task)
 204                         return 1;
 205                 goto out_sleep;
 206         }
 207         xprt->snd_task = task;
 208         if (req != NULL) {
 209                 req->rq_bytes_sent = 0;
 210                 req->rq_ntrans++;
 211         }
 212
 213         return 1;
 214
 215 out_sleep:
 216         dprintk("RPC: %5u failed to lock transport %p\n",
 217                         task->tk_pid, xprt);
 218         task->tk_timeout = 0;
 219         task->tk_status = -EAGAIN;
 220         if (req == NULL)
 221                 priority = RPC_PRIORITY_LOW;
 222         else if (!req->rq_ntrans)
 223                 priority = RPC_PRIORITY_NORMAL;
 224         else
 225                 priority = RPC_PRIORITY_HIGH;
 226         rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
 227         return 0;
 228 }
 229 EXPORT_SYMBOL_GPL(xprt_reserve_xprt);
 230
 231 static void xprt_clear_locked(struct rpc_xprt *xprt)
 232 {
 233         xprt->snd_task = NULL;
 234         if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state)) {
 235                 smp_mb__before_clear_bit();
 236                 clear_bit(XPRT_LOCKED, &xprt->state);
 237                 smp_mb__after_clear_bit();
 238         } else
 239                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 240 }
 241
 242 /*
 243  * xprt_reserve_xprt_cong - serialize write access to transports
 244  * @task: task that is requesting access to the transport
 245  *
 246  * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
 247  * integrated into the decision of whether a request is allowed to be
 248  * woken up and given access to the transport.
 249  */
 250 int xprt_reserve_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 251 {
 252         struct rpc_rqst *req = task->tk_rqstp;
 253         int priority;
 254
 255         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 256                 if (task == xprt->snd_task)
 257                         return 1;
 258                 goto out_sleep;
 259         }
 260         if (req == NULL) {
 261                 xprt->snd_task = task;
 262                 return 1;
 263         }
 264         if (__xprt_get_cong(xprt, task)) {
 265                 xprt->snd_task = task;
 266                 req->rq_bytes_sent = 0;
 267                 req->rq_ntrans++;
 268                 return 1;
 269         }
 270         xprt_clear_locked(xprt);
 271 out_sleep:
 272         dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
 273         task->tk_timeout = 0;
 274         task->tk_status = -EAGAIN;
 275         if (req == NULL)
 276                 priority = RPC_PRIORITY_LOW;
 277         else if (!req->rq_ntrans)
 278                 priority = RPC_PRIORITY_NORMAL;
 279         else
 280                 priority = RPC_PRIORITY_HIGH;
 281         rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
 282         return 0;
 283 }
 284 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
 285
 286 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 287 {
 288         int retval;
 289
 290         spin_lock_bh(&xprt->transport_lock);
 291         retval = xprt->ops->reserve_xprt(xprt, task);
 292         spin_unlock_bh(&xprt->transport_lock);
 293         return retval;
 294 }
 295
 296 static bool __xprt_lock_write_func(struct rpc_task *task, void *data)
 297 {
 298         struct rpc_xprt *xprt = data;
 299         struct rpc_rqst *req;
 300
 301         req = task->tk_rqstp;
 302         xprt->snd_task = task;
 303         if (req) {
 304                 req->rq_bytes_sent = 0;
 305                 req->rq_ntrans++;
 306         }
 307         return true;
 308 }
 309
 310 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
 311 {
 312         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 313                 return;
 314
 315         if (rpc_wake_up_first(&xprt->sending, __xprt_lock_write_func, xprt))
 316                 return;
 317         xprt_clear_locked(xprt);
 318 }
 319
 320 static bool __xprt_lock_write_cong_func(struct rpc_task *task, void *data)
 321 {
 322         struct rpc_xprt *xprt = data;
 323         struct rpc_rqst *req;
 324
 325         req = task->tk_rqstp;
 326         if (req == NULL) {
 327                 xprt->snd_task = task;
 328                 return true;
 329         }
 330         if (__xprt_get_cong(xprt, task)) {
 331                 xprt->snd_task = task;
 332                 req->rq_bytes_sent = 0;
 333                 req->rq_ntrans++;
 334                 return true;
 335         }
 336         return false;
 337 }
 338
 339 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
 340 {
 341         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 342                 return;
 343         if (RPCXPRT_CONGESTED(xprt))
 344                 goto out_unlock;
 345         if (rpc_wake_up_first(&xprt->sending, __xprt_lock_write_cong_func, xprt))
 346                 return;
 347 out_unlock:
 348         xprt_clear_locked(xprt);
 349 }
 350
 351 /**
 352  * xprt_release_xprt - allow other requests to use a transport
 353  * @xprt: transport with other tasks potentially waiting
 354  * @task: task that is releasing access to the transport
 355  *
 356  * Note that "task" can be NULL.  No congestion control is provided.
 357  */
 358 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 359 {
 360         if (xprt->snd_task == task) {
 361                 xprt_clear_locked(xprt);
 362                 __xprt_lock_write_next(xprt);
 363         }
 364 }
 365 EXPORT_SYMBOL_GPL(xprt_release_xprt);
 366
 367 /**
 368  * xprt_release_xprt_cong - allow other requests to use a transport
 369  * @xprt: transport with other tasks potentially waiting
 370  * @task: task that is releasing access to the transport
 371  *
 372  * Note that "task" can be NULL.  Another task is awoken to use the
 373  * transport if the transport's congestion window allows it.
 374  */
 375 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 376 {
 377         if (xprt->snd_task == task) {
 378                 xprt_clear_locked(xprt);
 379                 __xprt_lock_write_next_cong(xprt);
 380         }
 381 }
 382 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
 383
 384 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 385 {
 386         spin_lock_bh(&xprt->transport_lock);
 387         xprt->ops->release_xprt(xprt, task);
 388         spin_unlock_bh(&xprt->transport_lock);
 389 }
 390
 391 /*
 392  * Van Jacobson congestion avoidance. Check if the congestion window
 393  * overflowed. Put the task to sleep if this is the case.
 394  */
 395 static int
 396 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 397 {
 398         struct rpc_rqst *req = task->tk_rqstp;
 399
 400         if (req->rq_cong)
 401                 return 1;
 402         dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
 403                         task->tk_pid, xprt->cong, xprt->cwnd);
 404         if (RPCXPRT_CONGESTED(xprt))
 405                 return 0;
 406         req->rq_cong = 1;
 407         xprt->cong += RPC_CWNDSCALE;
 408         return 1;
 409 }
 410
 411 /*
 412  * Adjust the congestion window, and wake up the next task
 413  * that has been sleeping due to congestion
 414  */
 415 static void
 416 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 417 {
 418         if (!req->rq_cong)
 419                 return;
 420         req->rq_cong = 0;
 421         xprt->cong -= RPC_CWNDSCALE;
 422         __xprt_lock_write_next_cong(xprt);
 423 }
 424
 425 /**
 426  * xprt_release_rqst_cong - housekeeping when request is complete
 427  * @task: RPC request that recently completed
 428  *
 429  * Useful for transports that require congestion control.
 430  */
 431 void xprt_release_rqst_cong(struct rpc_task *task)
 432 {
 433         struct rpc_rqst *req = task->tk_rqstp;
 434
 435         __xprt_put_cong(req->rq_xprt, req);
 436 }
 437 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
 438
 439 /**
 440  * xprt_adjust_cwnd - adjust transport congestion window
 441  * @xprt: pointer to xprt
 442  * @task: recently completed RPC request used to adjust window
 443  * @result: result code of completed RPC request
 444  *
 445  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
 446  */
 447 void xprt_adjust_cwnd(struct rpc_xprt *xprt, struct rpc_task *task, int result)
 448 {
 449         struct rpc_rqst *req = task->tk_rqstp;
 450         unsigned long cwnd = xprt->cwnd;
 451
 452         if (result >= 0 && cwnd <= xprt->cong) {
 453                 /* The (cwnd >> 1) term makes sure
 454                  * the result gets rounded properly. */
 455                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 456                 if (cwnd > RPC_MAXCWND(xprt))
 457                         cwnd = RPC_MAXCWND(xprt);
 458                 __xprt_lock_write_next_cong(xprt);
 459         } else if (result == -ETIMEDOUT) {
 460                 cwnd >>= 1;
 461                 if (cwnd < RPC_CWNDSCALE)
 462                         cwnd = RPC_CWNDSCALE;
 463         }
 464         dprintk("RPC:       cong %ld, cwnd was %ld, now %ld\n",
 465                         xprt->cong, xprt->cwnd, cwnd);
 466         xprt->cwnd = cwnd;
 467         __xprt_put_cong(xprt, req);
 468 }
 469 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
 470
 471 /**
 472  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
 473  * @xprt: transport with waiting tasks
 474  * @status: result code to plant in each task before waking it
 475  *
 476  */
 477 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
 478 {
 479         if (status < 0)
 480                 rpc_wake_up_status(&xprt->pending, status);
 481         else
 482                 rpc_wake_up(&xprt->pending);
 483 }
 484 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
 485
 486 /**
 487  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
 488  * @task: task to be put to sleep
 489  * @action: function pointer to be executed after wait
 490  *
 491  * Note that we only set the timer for the case of RPC_IS_SOFT(), since
 492  * we don't in general want to force a socket disconnection due to
 493  * an incomplete RPC call transmission.
 494  */
 495 void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action)
 496 {
 497         struct rpc_rqst *req = task->tk_rqstp;
 498         struct rpc_xprt *xprt = req->rq_xprt;
 499
 500         task->tk_timeout = RPC_IS_SOFT(task) ? req->rq_timeout : 0;
 501         rpc_sleep_on(&xprt->pending, task, action);
 502 }
 503 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
 504
 505 /**
 506  * xprt_write_space - wake the task waiting for transport output buffer space
 507  * @xprt: transport with waiting tasks
 508  *
 509  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
 510  */
 511 void xprt_write_space(struct rpc_xprt *xprt)
 512 {
 513         spin_lock_bh(&xprt->transport_lock);
 514         if (xprt->snd_task) {
 515                 dprintk("RPC:       write space: waking waiting task on "
 516                                 "xprt %p\n", xprt);
 517                 rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task);
 518         }
 519         spin_unlock_bh(&xprt->transport_lock);
 520 }
 521 EXPORT_SYMBOL_GPL(xprt_write_space);
 522
 523 /**
 524  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
 525  * @task: task whose timeout is to be set
 526  *
 527  * Set a request's retransmit timeout based on the transport's
 528  * default timeout parameters.  Used by transports that don't adjust
 529  * the retransmit timeout based on round-trip time estimation.
 530  */
 531 void xprt_set_retrans_timeout_def(struct rpc_task *task)
 532 {
 533         task->tk_timeout = task->tk_rqstp->rq_timeout;
 534 }
 535 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
 536
 537 /**
 538  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
 539  * @task: task whose timeout is to be set
 540  *
 541  * Set a request's retransmit timeout using the RTT estimator.
 542  */
 543 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
 544 {
 545         int timer = task->tk_msg.rpc_proc->p_timer;
 546         struct rpc_clnt *clnt = task->tk_client;
 547         struct rpc_rtt *rtt = clnt->cl_rtt;
 548         struct rpc_rqst *req = task->tk_rqstp;
 549         unsigned long max_timeout = clnt->cl_timeout->to_maxval;
 550
 551         task->tk_timeout = rpc_calc_rto(rtt, timer);
 552         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
 553         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
 554                 task->tk_timeout = max_timeout;
 555 }
 556 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
 557
 558 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 559 {
 560         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 561
 562         req->rq_majortimeo = req->rq_timeout;
 563         if (to->to_exponential)
 564                 req->rq_majortimeo <<= to->to_retries;
 565         else
 566                 req->rq_majortimeo += to->to_increment * to->to_retries;
 567         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 568                 req->rq_majortimeo = to->to_maxval;
 569         req->rq_majortimeo += jiffies;
 570 }
 571
 572 /**
 573  * xprt_adjust_timeout - adjust timeout values for next retransmit
 574  * @req: RPC request containing parameters to use for the adjustment
 575  *
 576  */
 577 int xprt_adjust_timeout(struct rpc_rqst *req)
 578 {
 579         struct rpc_xprt *xprt = req->rq_xprt;
 580         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 581         int status = 0;
 582
 583         if (time_before(jiffies, req->rq_majortimeo)) {
 584                 if (to->to_exponential)
 585                         req->rq_timeout <<= 1;
 586                 else
 587                         req->rq_timeout += to->to_increment;
 588                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 589                         req->rq_timeout = to->to_maxval;
 590                 req->rq_retries++;
 591         } else {
 592                 req->rq_timeout = to->to_initval;
 593                 req->rq_retries = 0;
 594                 xprt_reset_majortimeo(req);
 595                 /* Reset the RTT counters == "slow start" */
 596                 spin_lock_bh(&xprt->transport_lock);
 597                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 598                 spin_unlock_bh(&xprt->transport_lock);
 599                 status = -ETIMEDOUT;
 600         }
 601
 602         if (req->rq_timeout == 0) {
 603                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 604                 req->rq_timeout = 5 * HZ;
 605         }
 606         return status;
 607 }
 608
 609 static void xprt_autoclose(struct work_struct *work)
 610 {
 611         struct rpc_xprt *xprt =
 612                 container_of(work, struct rpc_xprt, task_cleanup);
 613
 614         xprt->ops->close(xprt);
 615         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
 616         xprt_release_write(xprt, NULL);
 617 }
 618
 619 /**
 620  * xprt_disconnect_done - mark a transport as disconnected
 621  * @xprt: transport to flag for disconnect
 622  *
 623  */
 624 void xprt_disconnect_done(struct rpc_xprt *xprt)
 625 {
 626         dprintk("RPC:       disconnected transport %p\n", xprt);
 627         spin_lock_bh(&xprt->transport_lock);
 628         xprt_clear_connected(xprt);
 629         xprt_wake_pending_tasks(xprt, -EAGAIN);
 630         spin_unlock_bh(&xprt->transport_lock);
 631 }
 632 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
 633
 634 /**
 635  * xprt_force_disconnect - force a transport to disconnect
 636  * @xprt: transport to disconnect
 637  *
 638  */
 639 void xprt_force_disconnect(struct rpc_xprt *xprt)
 640 {
 641         /* Don't race with the test_bit() in xprt_clear_locked() */
 642         spin_lock_bh(&xprt->transport_lock);
 643         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 644         /* Try to schedule an autoclose RPC call */
 645         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 646                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 647         xprt_wake_pending_tasks(xprt, -EAGAIN);
 648         spin_unlock_bh(&xprt->transport_lock);
 649 }
 650
 651 /**
 652  * xprt_conditional_disconnect - force a transport to disconnect
 653  * @xprt: transport to disconnect
 654  * @cookie: 'connection cookie'
 655  *
 656  * This attempts to break the connection if and only if 'cookie' matches
 657  * the current transport 'connection cookie'. It ensures that we don't
 658  * try to break the connection more than once when we need to retransmit
 659  * a batch of RPC requests.
 660  *
 661  */
 662 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
 663 {
 664         /* Don't race with the test_bit() in xprt_clear_locked() */
 665         spin_lock_bh(&xprt->transport_lock);
 666         if (cookie != xprt->connect_cookie)
 667                 goto out;
 668         if (test_bit(XPRT_CLOSING, &xprt->state) || !xprt_connected(xprt))
 669                 goto out;
 670         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 671         /* Try to schedule an autoclose RPC call */
 672         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 673                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 674         xprt_wake_pending_tasks(xprt, -EAGAIN);
 675 out:
 676         spin_unlock_bh(&xprt->transport_lock);
 677 }
 678
 679 static void
 680 xprt_init_autodisconnect(unsigned long data)
 681 {
 682         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 683
 684         spin_lock(&xprt->transport_lock);
 685         if (!list_empty(&xprt->recv))
 686                 goto out_abort;
 687         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 688                 goto out_abort;
 689         spin_unlock(&xprt->transport_lock);
 690         set_bit(XPRT_CONNECTION_CLOSE, &xprt->state);
 691         queue_work(rpciod_workqueue, &xprt->task_cleanup);
 692         return;
 693 out_abort:
 694         spin_unlock(&xprt->transport_lock);
 695 }
 696
 697 /**
 698  * xprt_connect - schedule a transport connect operation
 699  * @task: RPC task that is requesting the connect
 700  *
 701  */
 702 void xprt_connect(struct rpc_task *task)
 703 {
 704         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 705
 706         dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
 707                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 708
 709         if (!xprt_bound(xprt)) {
 710                 task->tk_status = -EAGAIN;
 711                 return;
 712         }
 713         if (!xprt_lock_write(xprt, task))
 714                 return;
 715
 716         if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
 717                 xprt->ops->close(xprt);
 718
 719         if (xprt_connected(xprt))
 720                 xprt_release_write(xprt, task);
 721         else {
 722                 task->tk_rqstp->rq_bytes_sent = 0;
 723                 task->tk_timeout = task->tk_rqstp->rq_timeout;
 724                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
 725
 726                 if (test_bit(XPRT_CLOSING, &xprt->state))
 727                         return;
 728                 if (xprt_test_and_set_connecting(xprt))
 729                         return;
 730                 xprt->stat.connect_start = jiffies;
 731                 xprt->ops->connect(xprt, task);
 732         }
 733 }
 734
 735 static void xprt_connect_status(struct rpc_task *task)
 736 {
 737         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 738
 739         if (task->tk_status == 0) {
 740                 xprt->stat.connect_count++;
 741                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
 742                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
 743                                 task->tk_pid);
 744                 return;
 745         }
 746
 747         switch (task->tk_status) {
 748         case -EAGAIN:
 749                 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
 750                 break;
 751         case -ETIMEDOUT:
 752                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
 753                                 "out\n", task->tk_pid);
 754                 break;
 755         default:
 756                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
 757                                 "server %s\n", task->tk_pid, -task->tk_status,
 758                                 xprt->servername);
 759                 xprt_release_write(xprt, task);
 760                 task->tk_status = -EIO;
 761         }
 762 }
 763
 764 /**
 765  * xprt_lookup_rqst - find an RPC request corresponding to an XID
 766  * @xprt: transport on which the original request was transmitted
 767  * @xid: RPC XID of incoming reply
 768  *
 769  */
 770 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
 771 {
 772         struct rpc_rqst *entry;
 773
 774         list_for_each_entry(entry, &xprt->recv, rq_list)
 775                 if (entry->rq_xid == xid)
 776                         return entry;
 777
 778         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
 779                         ntohl(xid));
 780         xprt->stat.bad_xids++;
 781         return NULL;
 782 }
 783 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
 784
 785 static void xprt_update_rtt(struct rpc_task *task)
 786 {
 787         struct rpc_rqst *req = task->tk_rqstp;
 788         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
 789         unsigned int timer = task->tk_msg.rpc_proc->p_timer;
 790         long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
 791
 792         if (timer) {
 793                 if (req->rq_ntrans == 1)
 794                         rpc_update_rtt(rtt, timer, m);
 795                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
 796         }
 797 }
 798
 799 /**
 800  * xprt_complete_rqst - called when reply processing is complete
 801  * @task: RPC request that recently completed
 802  * @copied: actual number of bytes received from the transport
 803  *
 804  * Caller holds transport lock.
 805  */
 806 void xprt_complete_rqst(struct rpc_task *task, int copied)
 807 {
 808         struct rpc_rqst *req = task->tk_rqstp;
 809         struct rpc_xprt *xprt = req->rq_xprt;
 810
 811         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
 812                         task->tk_pid, ntohl(req->rq_xid), copied);
 813
 814         xprt->stat.recvs++;
 815         req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime);
 816         if (xprt->ops->timer != NULL)
 817                 xprt_update_rtt(task);
 818
 819         list_del_init(&req->rq_list);
 820         req->rq_private_buf.len = copied;
 821         /* Ensure all writes are done before we update */
 822         /* req->rq_reply_bytes_recvd */
 823         smp_wmb();
 824         req->rq_reply_bytes_recvd = copied;
 825         rpc_wake_up_queued_task(&xprt->pending, task);
 826 }
 827 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
 828
 829 static void xprt_timer(struct rpc_task *task)
 830 {
 831         struct rpc_rqst *req = task->tk_rqstp;
 832         struct rpc_xprt *xprt = req->rq_xprt;
 833
 834         if (task->tk_status != -ETIMEDOUT)
 835                 return;
 836         dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
 837
 838         spin_lock_bh(&xprt->transport_lock);
 839         if (!req->rq_reply_bytes_recvd) {
 840                 if (xprt->ops->timer)
 841                         xprt->ops->timer(xprt, task);
 842         } else
 843                 task->tk_status = 0;
 844         spin_unlock_bh(&xprt->transport_lock);
 845 }
 846
 847 static inline int xprt_has_timer(struct rpc_xprt *xprt)
 848 {
 849         return xprt->idle_timeout != 0;
 850 }
 851
 852 /**
 853  * xprt_prepare_transmit - reserve the transport before sending a request
 854  * @task: RPC task about to send a request
 855  *
 856  */
 857 int xprt_prepare_transmit(struct rpc_task *task)
 858 {
 859         struct rpc_rqst *req = task->tk_rqstp;
 860         struct rpc_xprt *xprt = req->rq_xprt;
 861         int err = 0;
 862
 863         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
 864
 865         spin_lock_bh(&xprt->transport_lock);
 866         if (req->rq_reply_bytes_recvd && !req->rq_bytes_sent) {
 867                 err = req->rq_reply_bytes_recvd;
 868                 goto out_unlock;
 869         }
 870         if (!xprt->ops->reserve_xprt(xprt, task))
 871                 err = -EAGAIN;
 872 out_unlock:
 873         spin_unlock_bh(&xprt->transport_lock);
 874         return err;
 875 }
 876
 877 void xprt_end_transmit(struct rpc_task *task)
 878 {
 879         xprt_release_write(task->tk_rqstp->rq_xprt, task);
 880 }
 881
 882 /**
 883  * xprt_transmit - send an RPC request on a transport
 884  * @task: controlling RPC task
 885  *
 886  * We have to copy the iovec because sendmsg fiddles with its contents.
 887  */
 888 void xprt_transmit(struct rpc_task *task)
 889 {
 890         struct rpc_rqst *req = task->tk_rqstp;
 891         struct rpc_xprt *xprt = req->rq_xprt;
 892         int status, numreqs;
 893
 894         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
 895
 896         if (!req->rq_reply_bytes_recvd) {
 897                 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) {
 898                         /*
 899                          * Add to the list only if we're expecting a reply
 900                          */
 901                         spin_lock_bh(&xprt->transport_lock);
 902                         /* Update the softirq receive buffer */
 903                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
 904                                         sizeof(req->rq_private_buf));
 905                         /* Add request to the receive list */
 906                         list_add_tail(&req->rq_list, &xprt->recv);
 907                         spin_unlock_bh(&xprt->transport_lock);
 908                         xprt_reset_majortimeo(req);
 909                         /* Turn off autodisconnect */
 910                         del_singleshot_timer_sync(&xprt->timer);
 911                 }
 912         } else if (!req->rq_bytes_sent)
 913                 return;
 914
 915         req->rq_connect_cookie = xprt->connect_cookie;
 916         req->rq_xtime = ktime_get();
 917         status = xprt->ops->send_request(task);
 918         if (status != 0) {
 919                 task->tk_status = status;
 920                 return;
 921         }
 922
 923         dprintk("RPC: %5u xmit complete\n", task->tk_pid);
 924         task->tk_flags |= RPC_TASK_SENT;
 925         spin_lock_bh(&xprt->transport_lock);
 926
 927         xprt->ops->set_retrans_timeout(task);
 928
 929         numreqs = atomic_read(&xprt->num_reqs);
 930         if (numreqs > xprt->stat.max_slots)
 931                 xprt->stat.max_slots = numreqs;
 932         xprt->stat.sends++;
 933         xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
 934         xprt->stat.bklog_u += xprt->backlog.qlen;
 935         xprt->stat.sending_u += xprt->sending.qlen;
 936         xprt->stat.pending_u += xprt->pending.qlen;
 937
 938         /* Don't race with disconnect */
 939         if (!xprt_connected(xprt))
 940                 task->tk_status = -ENOTCONN;
 941         else if (!req->rq_reply_bytes_recvd && rpc_reply_expected(task)) {
 942                 /*
 943                  * Sleep on the pending queue since
 944                  * we're expecting a reply.
 945                  */
 946                 rpc_sleep_on(&xprt->pending, task, xprt_timer);
 947         }
 948         spin_unlock_bh(&xprt->transport_lock);
 949 }
 950
 951 static void xprt_add_backlog(struct rpc_xprt *xprt, struct rpc_task *task)
 952 {
 953         set_bit(XPRT_CONGESTED, &xprt->state);
 954         rpc_sleep_on(&xprt->backlog, task, NULL);
 955 }
 956
 957 static void xprt_wake_up_backlog(struct rpc_xprt *xprt)
 958 {
 959         if (rpc_wake_up_next(&xprt->backlog) == NULL)
 960                 clear_bit(XPRT_CONGESTED, &xprt->state);
 961 }
 962
 963 static bool xprt_throttle_congested(struct rpc_xprt *xprt, struct rpc_task *task)
 964 {
 965         bool ret = false;
 966
 967         if (!test_bit(XPRT_CONGESTED, &xprt->state))
 968                 goto out;
 969         spin_lock(&xprt->reserve_lock);
 970         if (test_bit(XPRT_CONGESTED, &xprt->state)) {
 971                 rpc_sleep_on(&xprt->backlog, task, NULL);
 972                 ret = true;
 973         }
 974         spin_unlock(&xprt->reserve_lock);
 975 out:
 976         return ret;
 977 }
 978
 979 static struct rpc_rqst *xprt_dynamic_alloc_slot(struct rpc_xprt *xprt, gfp_t gfp_flags)
 980 {
 981         struct rpc_rqst *req = ERR_PTR(-EAGAIN);
 982
 983         if (!atomic_add_unless(&xprt->num_reqs, 1, xprt->max_reqs))
 984                 goto out;
 985         req = kzalloc(sizeof(struct rpc_rqst), gfp_flags);
 986         if (req != NULL)
 987                 goto out;
 988         atomic_dec(&xprt->num_reqs);
 989         req = ERR_PTR(-ENOMEM);
 990 out:
 991         return req;
 992 }
 993
 994 static bool xprt_dynamic_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
 995 {
 996         if (atomic_add_unless(&xprt->num_reqs, -1, xprt->min_reqs)) {
 997                 kfree(req);
 998                 return true;
 999         }
1000         return false;
1001 }
1002
1003 void xprt_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1004 {
1005         struct rpc_rqst *req;
1006
1007         spin_lock(&xprt->reserve_lock);
1008         if (!list_empty(&xprt->free)) {
1009                 req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1010                 list_del(&req->rq_list);
1011                 goto out_init_req;
1012         }
1013         req = xprt_dynamic_alloc_slot(xprt, GFP_NOWAIT|__GFP_NOWARN);
1014         if (!IS_ERR(req))
1015                 goto out_init_req;
1016         switch (PTR_ERR(req)) {
1017         case -ENOMEM:
1018                 dprintk("RPC:       dynamic allocation of request slot "
1019                                 "failed! Retrying\n");
1020                 task->tk_status = -ENOMEM;
1021                 break;
1022         case -EAGAIN:
1023                 xprt_add_backlog(xprt, task);
1024                 dprintk("RPC:       waiting for request slot\n");
1025         default:
1026                 task->tk_status = -EAGAIN;
1027         }
1028         spin_unlock(&xprt->reserve_lock);
1029         return;
1030 out_init_req:
1031         task->tk_status = 0;
1032         task->tk_rqstp = req;
1033         xprt_request_init(task, xprt);
1034         spin_unlock(&xprt->reserve_lock);
1035 }
1036 EXPORT_SYMBOL_GPL(xprt_alloc_slot);
1037
1038 void xprt_lock_and_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1039 {
1040         /* Note: grabbing the xprt_lock_write() ensures that we throttle
1041          * new slot allocation if the transport is congested (i.e. when
1042          * reconnecting a stream transport or when out of socket write
1043          * buffer space).
1044          */
1045         if (xprt_lock_write(xprt, task)) {
1046                 xprt_alloc_slot(xprt, task);
1047                 xprt_release_write(xprt, task);
1048         }
1049 }
1050 EXPORT_SYMBOL_GPL(xprt_lock_and_alloc_slot);
1051
1052 static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1053 {
1054         spin_lock(&xprt->reserve_lock);
1055         if (!xprt_dynamic_free_slot(xprt, req)) {
1056                 memset(req, 0, sizeof(*req));   /* mark unused */
1057                 list_add(&req->rq_list, &xprt->free);
1058         }
1059         xprt_wake_up_backlog(xprt);
1060         spin_unlock(&xprt->reserve_lock);
1061 }
1062
1063 static void xprt_free_all_slots(struct rpc_xprt *xprt)
1064 {
1065         struct rpc_rqst *req;
1066         while (!list_empty(&xprt->free)) {
1067                 req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list);
1068                 list_del(&req->rq_list);
1069                 kfree(req);
1070         }
1071 }
1072
1073 struct rpc_xprt *xprt_alloc(struct net *net, size_t size,
1074                 unsigned int num_prealloc,
1075                 unsigned int max_alloc)
1076 {
1077         struct rpc_xprt *xprt;
1078         struct rpc_rqst *req;
1079         int i;
1080
1081         xprt = kzalloc(size, GFP_KERNEL);
1082         if (xprt == NULL)
1083                 goto out;
1084
1085         xprt_init(xprt, net);
1086
1087         for (i = 0; i < num_prealloc; i++) {
1088                 req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL);
1089                 if (!req)
1090                         break;
1091                 list_add(&req->rq_list, &xprt->free);
1092         }
1093         if (i < num_prealloc)
1094                 goto out_free;
1095         if (max_alloc > num_prealloc)
1096                 xprt->max_reqs = max_alloc;
1097         else
1098                 xprt->max_reqs = num_prealloc;
1099         xprt->min_reqs = num_prealloc;
1100         atomic_set(&xprt->num_reqs, num_prealloc);
1101
1102         return xprt;
1103
1104 out_free:
1105         xprt_free(xprt);
1106 out:
1107         return NULL;
1108 }
1109 EXPORT_SYMBOL_GPL(xprt_alloc);
1110
1111 void xprt_free(struct rpc_xprt *xprt)
1112 {
1113         put_net(xprt->xprt_net);
1114         xprt_free_all_slots(xprt);
1115         kfree(xprt);
1116 }
1117 EXPORT_SYMBOL_GPL(xprt_free);
1118
1119 /**
1120  * xprt_reserve - allocate an RPC request slot
1121  * @task: RPC task requesting a slot allocation
1122  *
1123  * If the transport is marked as being congested, or if no more
1124  * slots are available, place the task on the transport's
1125  * backlog queue.
1126  */
1127 void xprt_reserve(struct rpc_task *task)
1128 {
1129         struct rpc_xprt *xprt;
1130
1131         task->tk_status = 0;
1132         if (task->tk_rqstp != NULL)
1133                 return;
1134
1135         task->tk_timeout = 0;
1136         task->tk_status = -EAGAIN;
1137         rcu_read_lock();
1138         xprt = rcu_dereference(task->tk_client->cl_xprt);
1139         if (!xprt_throttle_congested(xprt, task))
1140                 xprt->ops->alloc_slot(xprt, task);
1141         rcu_read_unlock();
1142 }
1143
1144 /**
1145  * xprt_retry_reserve - allocate an RPC request slot
1146  * @task: RPC task requesting a slot allocation
1147  *
1148  * If no more slots are available, place the task on the transport's
1149  * backlog queue.
1150  * Note that the only difference with xprt_reserve is that we now
1151  * ignore the value of the XPRT_CONGESTED flag.
1152  */
1153 void xprt_retry_reserve(struct rpc_task *task)
1154 {
1155         struct rpc_xprt *xprt;
1156
1157         task->tk_status = 0;
1158         if (task->tk_rqstp != NULL)
1159                 return;
1160
1161         task->tk_timeout = 0;
1162         task->tk_status = -EAGAIN;
1163         rcu_read_lock();
1164         xprt = rcu_dereference(task->tk_client->cl_xprt);
1165         xprt->ops->alloc_slot(xprt, task);
1166         rcu_read_unlock();
1167 }
1168
1169 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
1170 {
1171         return (__force __be32)xprt->xid++;
1172 }
1173
1174 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1175 {
1176         xprt->xid = net_random();
1177 }
1178
1179 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1180 {
1181         struct rpc_rqst *req = task->tk_rqstp;
1182
1183         INIT_LIST_HEAD(&req->rq_list);
1184         req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1185         req->rq_task    = task;
1186         req->rq_xprt    = xprt;
1187         req->rq_buffer  = NULL;
1188         req->rq_xid     = xprt_alloc_xid(xprt);
1189         req->rq_release_snd_buf = NULL;
1190         xprt_reset_majortimeo(req);
1191         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1192                         req, ntohl(req->rq_xid));
1193 }
1194
1195 /**
1196  * xprt_release - release an RPC request slot
1197  * @task: task which is finished with the slot
1198  *
1199  */
1200 void xprt_release(struct rpc_task *task)
1201 {
1202         struct rpc_xprt *xprt;
1203         struct rpc_rqst *req = task->tk_rqstp;
1204
1205         if (req == NULL) {
1206                 if (task->tk_client) {
1207                         rcu_read_lock();
1208                         xprt = rcu_dereference(task->tk_client->cl_xprt);
1209                         if (xprt->snd_task == task)
1210                                 xprt_release_write(xprt, task);
1211                         rcu_read_unlock();
1212                 }
1213                 return;
1214         }
1215
1216         xprt = req->rq_xprt;
1217         if (task->tk_ops->rpc_count_stats != NULL)
1218                 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
1219         else if (task->tk_client)
1220                 rpc_count_iostats(task, task->tk_client->cl_metrics);
1221         spin_lock_bh(&xprt->transport_lock);
1222         xprt->ops->release_xprt(xprt, task);
1223         if (xprt->ops->release_request)
1224                 xprt->ops->release_request(task);
1225         if (!list_empty(&req->rq_list))
1226                 list_del(&req->rq_list);
1227         xprt->last_used = jiffies;
1228         if (list_empty(&xprt->recv) && xprt_has_timer(xprt))
1229                 mod_timer(&xprt->timer,
1230                                 xprt->last_used + xprt->idle_timeout);
1231         spin_unlock_bh(&xprt->transport_lock);
1232         if (req->rq_buffer)
1233                 xprt->ops->buf_free(req->rq_buffer);
1234         if (req->rq_cred != NULL)
1235                 put_rpccred(req->rq_cred);
1236         task->tk_rqstp = NULL;
1237         if (req->rq_release_snd_buf)
1238                 req->rq_release_snd_buf(req);
1239
1240         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1241         if (likely(!bc_prealloc(req)))
1242                 xprt_free_slot(xprt, req);
1243         else
1244                 xprt_free_bc_request(req);
1245 }
1246
1247 static void xprt_init(struct rpc_xprt *xprt, struct net *net)
1248 {
1249         atomic_set(&xprt->count, 1);
1250
1251         spin_lock_init(&xprt->transport_lock);
1252         spin_lock_init(&xprt->reserve_lock);
1253
1254         INIT_LIST_HEAD(&xprt->free);
1255         INIT_LIST_HEAD(&xprt->recv);
1256 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1257         spin_lock_init(&xprt->bc_pa_lock);
1258         INIT_LIST_HEAD(&xprt->bc_pa_list);
1259 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1260
1261         xprt->last_used = jiffies;
1262         xprt->cwnd = RPC_INITCWND;
1263         xprt->bind_index = 0;
1264
1265         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1266         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1267         rpc_init_priority_wait_queue(&xprt->sending, "xprt_sending");
1268         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1269
1270         xprt_init_xid(xprt);
1271
1272         xprt->xprt_net = get_net(net);
1273 }
1274
1275 /**
1276  * xprt_create_transport - create an RPC transport
1277  * @args: rpc transport creation arguments
1278  *
1279  */
1280 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1281 {
1282         struct rpc_xprt *xprt;
1283         struct xprt_class *t;
1284
1285         spin_lock(&xprt_list_lock);
1286         list_for_each_entry(t, &xprt_list, list) {
1287                 if (t->ident == args->ident) {
1288                         spin_unlock(&xprt_list_lock);
1289                         goto found;
1290                 }
1291         }
1292         spin_unlock(&xprt_list_lock);
1293         printk(KERN_ERR "RPC: transport (%d) not supported\n", args->ident);
1294         return ERR_PTR(-EIO);
1295
1296 found:
1297         xprt = t->setup(args);
1298         if (IS_ERR(xprt)) {
1299                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1300                                 -PTR_ERR(xprt));
1301                 goto out;
1302         }
1303         if (args->flags & XPRT_CREATE_NO_IDLE_TIMEOUT)
1304                 xprt->idle_timeout = 0;
1305         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1306         if (xprt_has_timer(xprt))
1307                 setup_timer(&xprt->timer, xprt_init_autodisconnect,
1308                             (unsigned long)xprt);
1309         else
1310                 init_timer(&xprt->timer);
1311
1312         if (strlen(args->servername) > RPC_MAXNETNAMELEN) {
1313                 xprt_destroy(xprt);
1314                 return ERR_PTR(-EINVAL);
1315         }
1316         xprt->servername = kstrdup(args->servername, GFP_KERNEL);
1317         if (xprt->servername == NULL) {
1318                 xprt_destroy(xprt);
1319                 return ERR_PTR(-ENOMEM);
1320         }
1321
1322         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1323                         xprt->max_reqs);
1324 out:
1325         return xprt;
1326 }
1327
1328 /**
1329  * xprt_destroy - destroy an RPC transport, killing off all requests.
1330  * @xprt: transport to destroy
1331  *
1332  */
1333 static void xprt_destroy(struct rpc_xprt *xprt)
1334 {
1335         dprintk("RPC:       destroying transport %p\n", xprt);
1336         del_timer_sync(&xprt->timer);
1337
1338         rpc_destroy_wait_queue(&xprt->binding);
1339         rpc_destroy_wait_queue(&xprt->pending);
1340         rpc_destroy_wait_queue(&xprt->sending);
1341         rpc_destroy_wait_queue(&xprt->backlog);
1342         cancel_work_sync(&xprt->task_cleanup);
1343         kfree(xprt->servername);
1344         /*
1345          * Tear down transport state and free the rpc_xprt
1346          */
1347         xprt->ops->destroy(xprt);
1348 }
1349
1350 /**
1351  * xprt_put - release a reference to an RPC transport.
1352  * @xprt: pointer to the transport
1353  *
1354  */
1355 void xprt_put(struct rpc_xprt *xprt)
1356 {
1357         if (atomic_dec_and_test(&xprt->count))
1358                 xprt_destroy(xprt);
1359 }
1360
1361 /**
1362  * xprt_get - return a reference to an RPC transport.
1363  * @xprt: pointer to the transport
1364  *
1365  */
1366 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1367 {
1368         if (atomic_inc_not_zero(&xprt->count))
1369                 return xprt;
1370         return NULL;
1371 }