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 #include <linux/rcupdate.h>
  52
  53 #include <trace/events/sunrpc.h>
  54
  55 #include "sunrpc.h"
  56
  57 /*
  58  * Local variables
  59  */
  60
  61 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  62 # define RPCDBG_FACILITY        RPCDBG_XPRT
  63 #endif
  64
  65 /*
  66  * Local functions
  67  */
  68 static void      xprt_init(struct rpc_xprt *xprt, struct net *net);
  69 static __be32   xprt_alloc_xid(struct rpc_xprt *xprt);
  70 static void     xprt_connect_status(struct rpc_task *task);
  71 static int      __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
  72 static void     __xprt_put_cong(struct rpc_xprt *, struct rpc_rqst *);
  73 static void      xprt_destroy(struct rpc_xprt *xprt);
  74
  75 static DEFINE_SPINLOCK(xprt_list_lock);
  76 static LIST_HEAD(xprt_list);
  77
  78 /**
  79  * xprt_register_transport - register a transport implementation
  80  * @transport: transport to register
  81  *
  82  * If a transport implementation is loaded as a kernel module, it can
  83  * call this interface to make itself known to the RPC client.
  84  *
  85  * Returns:
  86  * 0:           transport successfully registered
  87  * -EEXIST:     transport already registered
  88  * -EINVAL:     transport module being unloaded
  89  */
  90 int xprt_register_transport(struct xprt_class *transport)
  91 {
  92         struct xprt_class *t;
  93         int result;
  94
  95         result = -EEXIST;
  96         spin_lock(&xprt_list_lock);
  97         list_for_each_entry(t, &xprt_list, list) {
  98                 /* don't register the same transport class twice */
  99                 if (t->ident == transport->ident)
 100                         goto out;
 101         }
 102
 103         list_add_tail(&transport->list, &xprt_list);
 104         printk(KERN_INFO "RPC: Registered %s transport module.\n",
 105                transport->name);
 106         result = 0;
 107
 108 out:
 109         spin_unlock(&xprt_list_lock);
 110         return result;
 111 }
 112 EXPORT_SYMBOL_GPL(xprt_register_transport);
 113
 114 /**
 115  * xprt_unregister_transport - unregister a transport implementation
 116  * @transport: transport to unregister
 117  *
 118  * Returns:
 119  * 0:           transport successfully unregistered
 120  * -ENOENT:     transport never registered
 121  */
 122 int xprt_unregister_transport(struct xprt_class *transport)
 123 {
 124         struct xprt_class *t;
 125         int result;
 126
 127         result = 0;
 128         spin_lock(&xprt_list_lock);
 129         list_for_each_entry(t, &xprt_list, list) {
 130                 if (t == transport) {
 131                         printk(KERN_INFO
 132                                 "RPC: Unregistered %s transport module.\n",
 133                                 transport->name);
 134                         list_del_init(&transport->list);
 135                         goto out;
 136                 }
 137         }
 138         result = -ENOENT;
 139
 140 out:
 141         spin_unlock(&xprt_list_lock);
 142         return result;
 143 }
 144 EXPORT_SYMBOL_GPL(xprt_unregister_transport);
 145
 146 /**
 147  * xprt_load_transport - load a transport implementation
 148  * @transport_name: transport to load
 149  *
 150  * Returns:
 151  * 0:           transport successfully loaded
 152  * -ENOENT:     transport module not available
 153  */
 154 int xprt_load_transport(const char *transport_name)
 155 {
 156         struct xprt_class *t;
 157         int result;
 158
 159         result = 0;
 160         spin_lock(&xprt_list_lock);
 161         list_for_each_entry(t, &xprt_list, list) {
 162                 if (strcmp(t->name, transport_name) == 0) {
 163                         spin_unlock(&xprt_list_lock);
 164                         goto out;
 165                 }
 166         }
 167         spin_unlock(&xprt_list_lock);
 168         result = request_module("xprt%s", transport_name);
 169 out:
 170         return result;
 171 }
 172 EXPORT_SYMBOL_GPL(xprt_load_transport);
 173
 174 /**
 175  * xprt_reserve_xprt - serialize write access to transports
 176  * @task: task that is requesting access to the transport
 177  * @xprt: pointer to the target transport
 178  *
 179  * This prevents mixing the payload of separate requests, and prevents
 180  * transport connects from colliding with writes.  No congestion control
 181  * is provided.
 182  */
 183 int xprt_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 184 {
 185         struct rpc_rqst *req = task->tk_rqstp;
 186         int priority;
 187
 188         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 189                 if (task == xprt->snd_task)
 190                         return 1;
 191                 goto out_sleep;
 192         }
 193         xprt->snd_task = task;
 194         if (req != NULL)
 195                 req->rq_ntrans++;
 196
 197         return 1;
 198
 199 out_sleep:
 200         dprintk("RPC: %5u failed to lock transport %p\n",
 201                         task->tk_pid, xprt);
 202         task->tk_timeout = 0;
 203         task->tk_status = -EAGAIN;
 204         if (req == NULL)
 205                 priority = RPC_PRIORITY_LOW;
 206         else if (!req->rq_ntrans)
 207                 priority = RPC_PRIORITY_NORMAL;
 208         else
 209                 priority = RPC_PRIORITY_HIGH;
 210         rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
 211         return 0;
 212 }
 213 EXPORT_SYMBOL_GPL(xprt_reserve_xprt);
 214
 215 static void xprt_clear_locked(struct rpc_xprt *xprt)
 216 {
 217         xprt->snd_task = NULL;
 218         if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state)) {
 219                 smp_mb__before_atomic();
 220                 clear_bit(XPRT_LOCKED, &xprt->state);
 221                 smp_mb__after_atomic();
 222         } else
 223                 queue_work(xprtiod_workqueue, &xprt->task_cleanup);
 224 }
 225
 226 /*
 227  * xprt_reserve_xprt_cong - serialize write access to transports
 228  * @task: task that is requesting access to the transport
 229  *
 230  * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
 231  * integrated into the decision of whether a request is allowed to be
 232  * woken up and given access to the transport.
 233  */
 234 int xprt_reserve_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 235 {
 236         struct rpc_rqst *req = task->tk_rqstp;
 237         int priority;
 238
 239         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 240                 if (task == xprt->snd_task)
 241                         return 1;
 242                 goto out_sleep;
 243         }
 244         if (req == NULL) {
 245                 xprt->snd_task = task;
 246                 return 1;
 247         }
 248         if (__xprt_get_cong(xprt, task)) {
 249                 xprt->snd_task = task;
 250                 req->rq_ntrans++;
 251                 return 1;
 252         }
 253         xprt_clear_locked(xprt);
 254 out_sleep:
 255         if (req)
 256                 __xprt_put_cong(xprt, req);
 257         dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
 258         task->tk_timeout = 0;
 259         task->tk_status = -EAGAIN;
 260         if (req == NULL)
 261                 priority = RPC_PRIORITY_LOW;
 262         else if (!req->rq_ntrans)
 263                 priority = RPC_PRIORITY_NORMAL;
 264         else
 265                 priority = RPC_PRIORITY_HIGH;
 266         rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
 267         return 0;
 268 }
 269 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
 270
 271 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 272 {
 273         int retval;
 274
 275         spin_lock_bh(&xprt->transport_lock);
 276         retval = xprt->ops->reserve_xprt(xprt, task);
 277         spin_unlock_bh(&xprt->transport_lock);
 278         return retval;
 279 }
 280
 281 static bool __xprt_lock_write_func(struct rpc_task *task, void *data)
 282 {
 283         struct rpc_xprt *xprt = data;
 284         struct rpc_rqst *req;
 285
 286         req = task->tk_rqstp;
 287         xprt->snd_task = task;
 288         if (req)
 289                 req->rq_ntrans++;
 290         return true;
 291 }
 292
 293 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
 294 {
 295         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 296                 return;
 297
 298         if (rpc_wake_up_first_on_wq(xprtiod_workqueue, &xprt->sending,
 299                                 __xprt_lock_write_func, xprt))
 300                 return;
 301         xprt_clear_locked(xprt);
 302 }
 303
 304 static bool __xprt_lock_write_cong_func(struct rpc_task *task, void *data)
 305 {
 306         struct rpc_xprt *xprt = data;
 307         struct rpc_rqst *req;
 308
 309         req = task->tk_rqstp;
 310         if (req == NULL) {
 311                 xprt->snd_task = task;
 312                 return true;
 313         }
 314         if (__xprt_get_cong(xprt, task)) {
 315                 xprt->snd_task = task;
 316                 req->rq_ntrans++;
 317                 return true;
 318         }
 319         return false;
 320 }
 321
 322 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
 323 {
 324         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 325                 return;
 326         if (RPCXPRT_CONGESTED(xprt))
 327                 goto out_unlock;
 328         if (rpc_wake_up_first_on_wq(xprtiod_workqueue, &xprt->sending,
 329                                 __xprt_lock_write_cong_func, xprt))
 330                 return;
 331 out_unlock:
 332         xprt_clear_locked(xprt);
 333 }
 334
 335 static void xprt_task_clear_bytes_sent(struct rpc_task *task)
 336 {
 337         if (task != NULL) {
 338                 struct rpc_rqst *req = task->tk_rqstp;
 339                 if (req != NULL)
 340                         req->rq_bytes_sent = 0;
 341         }
 342 }
 343
 344 /**
 345  * xprt_release_xprt - allow other requests to use a transport
 346  * @xprt: transport with other tasks potentially waiting
 347  * @task: task that is releasing access to the transport
 348  *
 349  * Note that "task" can be NULL.  No congestion control is provided.
 350  */
 351 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 352 {
 353         if (xprt->snd_task == task) {
 354                 xprt_task_clear_bytes_sent(task);
 355                 xprt_clear_locked(xprt);
 356                 __xprt_lock_write_next(xprt);
 357         }
 358 }
 359 EXPORT_SYMBOL_GPL(xprt_release_xprt);
 360
 361 /**
 362  * xprt_release_xprt_cong - allow other requests to use a transport
 363  * @xprt: transport with other tasks potentially waiting
 364  * @task: task that is releasing access to the transport
 365  *
 366  * Note that "task" can be NULL.  Another task is awoken to use the
 367  * transport if the transport's congestion window allows it.
 368  */
 369 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 370 {
 371         if (xprt->snd_task == task) {
 372                 xprt_task_clear_bytes_sent(task);
 373                 xprt_clear_locked(xprt);
 374                 __xprt_lock_write_next_cong(xprt);
 375         }
 376 }
 377 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
 378
 379 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 380 {
 381         spin_lock_bh(&xprt->transport_lock);
 382         xprt->ops->release_xprt(xprt, task);
 383         spin_unlock_bh(&xprt->transport_lock);
 384 }
 385
 386 /*
 387  * Van Jacobson congestion avoidance. Check if the congestion window
 388  * overflowed. Put the task to sleep if this is the case.
 389  */
 390 static int
 391 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 392 {
 393         struct rpc_rqst *req = task->tk_rqstp;
 394
 395         if (req->rq_cong)
 396                 return 1;
 397         dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
 398                         task->tk_pid, xprt->cong, xprt->cwnd);
 399         if (RPCXPRT_CONGESTED(xprt))
 400                 return 0;
 401         req->rq_cong = 1;
 402         xprt->cong += RPC_CWNDSCALE;
 403         return 1;
 404 }
 405
 406 /*
 407  * Adjust the congestion window, and wake up the next task
 408  * that has been sleeping due to congestion
 409  */
 410 static void
 411 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 412 {
 413         if (!req->rq_cong)
 414                 return;
 415         req->rq_cong = 0;
 416         xprt->cong -= RPC_CWNDSCALE;
 417         __xprt_lock_write_next_cong(xprt);
 418 }
 419
 420 /**
 421  * xprt_release_rqst_cong - housekeeping when request is complete
 422  * @task: RPC request that recently completed
 423  *
 424  * Useful for transports that require congestion control.
 425  */
 426 void xprt_release_rqst_cong(struct rpc_task *task)
 427 {
 428         struct rpc_rqst *req = task->tk_rqstp;
 429
 430         __xprt_put_cong(req->rq_xprt, req);
 431 }
 432 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
 433
 434 /**
 435  * xprt_adjust_cwnd - adjust transport congestion window
 436  * @xprt: pointer to xprt
 437  * @task: recently completed RPC request used to adjust window
 438  * @result: result code of completed RPC request
 439  *
 440  * The transport code maintains an estimate on the maximum number of out-
 441  * standing RPC requests, using a smoothed version of the congestion
 442  * avoidance implemented in 44BSD. This is basically the Van Jacobson
 443  * congestion algorithm: If a retransmit occurs, the congestion window is
 444  * halved; otherwise, it is incremented by 1/cwnd when
 445  *
 446  *      -       a reply is received and
 447  *      -       a full number of requests are outstanding and
 448  *      -       the congestion window hasn't been updated recently.
 449  */
 450 void xprt_adjust_cwnd(struct rpc_xprt *xprt, struct rpc_task *task, int result)
 451 {
 452         struct rpc_rqst *req = task->tk_rqstp;
 453         unsigned long cwnd = xprt->cwnd;
 454
 455         if (result >= 0 && cwnd <= xprt->cong) {
 456                 /* The (cwnd >> 1) term makes sure
 457                  * the result gets rounded properly. */
 458                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 459                 if (cwnd > RPC_MAXCWND(xprt))
 460                         cwnd = RPC_MAXCWND(xprt);
 461                 __xprt_lock_write_next_cong(xprt);
 462         } else if (result == -ETIMEDOUT) {
 463                 cwnd >>= 1;
 464                 if (cwnd < RPC_CWNDSCALE)
 465                         cwnd = RPC_CWNDSCALE;
 466         }
 467         dprintk("RPC:       cong %ld, cwnd was %ld, now %ld\n",
 468                         xprt->cong, xprt->cwnd, cwnd);
 469         xprt->cwnd = cwnd;
 470         __xprt_put_cong(xprt, req);
 471 }
 472 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
 473
 474 /**
 475  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
 476  * @xprt: transport with waiting tasks
 477  * @status: result code to plant in each task before waking it
 478  *
 479  */
 480 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
 481 {
 482         if (status < 0)
 483                 rpc_wake_up_status(&xprt->pending, status);
 484         else
 485                 rpc_wake_up(&xprt->pending);
 486 }
 487 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
 488
 489 /**
 490  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
 491  * @task: task to be put to sleep
 492  * @action: function pointer to be executed after wait
 493  *
 494  * Note that we only set the timer for the case of RPC_IS_SOFT(), since
 495  * we don't in general want to force a socket disconnection due to
 496  * an incomplete RPC call transmission.
 497  */
 498 void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action)
 499 {
 500         struct rpc_rqst *req = task->tk_rqstp;
 501         struct rpc_xprt *xprt = req->rq_xprt;
 502
 503         task->tk_timeout = RPC_IS_SOFT(task) ? req->rq_timeout : 0;
 504         rpc_sleep_on(&xprt->pending, task, action);
 505 }
 506 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
 507
 508 /**
 509  * xprt_write_space - wake the task waiting for transport output buffer space
 510  * @xprt: transport with waiting tasks
 511  *
 512  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
 513  */
 514 void xprt_write_space(struct rpc_xprt *xprt)
 515 {
 516         spin_lock_bh(&xprt->transport_lock);
 517         if (xprt->snd_task) {
 518                 dprintk("RPC:       write space: waking waiting task on "
 519                                 "xprt %p\n", xprt);
 520                 rpc_wake_up_queued_task_on_wq(xprtiod_workqueue,
 521                                 &xprt->pending, xprt->snd_task);
 522         }
 523         spin_unlock_bh(&xprt->transport_lock);
 524 }
 525 EXPORT_SYMBOL_GPL(xprt_write_space);
 526
 527 /**
 528  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
 529  * @task: task whose timeout is to be set
 530  *
 531  * Set a request's retransmit timeout based on the transport's
 532  * default timeout parameters.  Used by transports that don't adjust
 533  * the retransmit timeout based on round-trip time estimation.
 534  */
 535 void xprt_set_retrans_timeout_def(struct rpc_task *task)
 536 {
 537         task->tk_timeout = task->tk_rqstp->rq_timeout;
 538 }
 539 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
 540
 541 /**
 542  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
 543  * @task: task whose timeout is to be set
 544  *
 545  * Set a request's retransmit timeout using the RTT estimator.
 546  */
 547 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
 548 {
 549         int timer = task->tk_msg.rpc_proc->p_timer;
 550         struct rpc_clnt *clnt = task->tk_client;
 551         struct rpc_rtt *rtt = clnt->cl_rtt;
 552         struct rpc_rqst *req = task->tk_rqstp;
 553         unsigned long max_timeout = clnt->cl_timeout->to_maxval;
 554
 555         task->tk_timeout = rpc_calc_rto(rtt, timer);
 556         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
 557         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
 558                 task->tk_timeout = max_timeout;
 559 }
 560 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
 561
 562 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 563 {
 564         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 565
 566         req->rq_majortimeo = req->rq_timeout;
 567         if (to->to_exponential)
 568                 req->rq_majortimeo <<= to->to_retries;
 569         else
 570                 req->rq_majortimeo += to->to_increment * to->to_retries;
 571         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 572                 req->rq_majortimeo = to->to_maxval;
 573         req->rq_majortimeo += jiffies;
 574 }
 575
 576 /**
 577  * xprt_adjust_timeout - adjust timeout values for next retransmit
 578  * @req: RPC request containing parameters to use for the adjustment
 579  *
 580  */
 581 int xprt_adjust_timeout(struct rpc_rqst *req)
 582 {
 583         struct rpc_xprt *xprt = req->rq_xprt;
 584         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 585         int status = 0;
 586
 587         if (time_before(jiffies, req->rq_majortimeo)) {
 588                 if (to->to_exponential)
 589                         req->rq_timeout <<= 1;
 590                 else
 591                         req->rq_timeout += to->to_increment;
 592                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 593                         req->rq_timeout = to->to_maxval;
 594                 req->rq_retries++;
 595         } else {
 596                 req->rq_timeout = to->to_initval;
 597                 req->rq_retries = 0;
 598                 xprt_reset_majortimeo(req);
 599                 /* Reset the RTT counters == "slow start" */
 600                 spin_lock_bh(&xprt->transport_lock);
 601                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 602                 spin_unlock_bh(&xprt->transport_lock);
 603                 status = -ETIMEDOUT;
 604         }
 605
 606         if (req->rq_timeout == 0) {
 607                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 608                 req->rq_timeout = 5 * HZ;
 609         }
 610         return status;
 611 }
 612
 613 static void xprt_autoclose(struct work_struct *work)
 614 {
 615         struct rpc_xprt *xprt =
 616                 container_of(work, struct rpc_xprt, task_cleanup);
 617
 618         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
 619         xprt->ops->close(xprt);
 620         xprt_release_write(xprt, NULL);
 621         wake_up_bit(&xprt->state, XPRT_LOCKED);
 622 }
 623
 624 /**
 625  * xprt_disconnect_done - mark a transport as disconnected
 626  * @xprt: transport to flag for disconnect
 627  *
 628  */
 629 void xprt_disconnect_done(struct rpc_xprt *xprt)
 630 {
 631         dprintk("RPC:       disconnected transport %p\n", xprt);
 632         spin_lock_bh(&xprt->transport_lock);
 633         xprt_clear_connected(xprt);
 634         xprt_wake_pending_tasks(xprt, -EAGAIN);
 635         spin_unlock_bh(&xprt->transport_lock);
 636 }
 637 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
 638
 639 /**
 640  * xprt_force_disconnect - force a transport to disconnect
 641  * @xprt: transport to disconnect
 642  *
 643  */
 644 void xprt_force_disconnect(struct rpc_xprt *xprt)
 645 {
 646         /* Don't race with the test_bit() in xprt_clear_locked() */
 647         spin_lock_bh(&xprt->transport_lock);
 648         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 649         /* Try to schedule an autoclose RPC call */
 650         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 651                 queue_work(xprtiod_workqueue, &xprt->task_cleanup);
 652         xprt_wake_pending_tasks(xprt, -EAGAIN);
 653         spin_unlock_bh(&xprt->transport_lock);
 654 }
 655 EXPORT_SYMBOL_GPL(xprt_force_disconnect);
 656
 657 /**
 658  * xprt_conditional_disconnect - force a transport to disconnect
 659  * @xprt: transport to disconnect
 660  * @cookie: 'connection cookie'
 661  *
 662  * This attempts to break the connection if and only if 'cookie' matches
 663  * the current transport 'connection cookie'. It ensures that we don't
 664  * try to break the connection more than once when we need to retransmit
 665  * a batch of RPC requests.
 666  *
 667  */
 668 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
 669 {
 670         /* Don't race with the test_bit() in xprt_clear_locked() */
 671         spin_lock_bh(&xprt->transport_lock);
 672         if (cookie != xprt->connect_cookie)
 673                 goto out;
 674         if (test_bit(XPRT_CLOSING, &xprt->state))
 675                 goto out;
 676         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 677         /* Try to schedule an autoclose RPC call */
 678         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 679                 queue_work(xprtiod_workqueue, &xprt->task_cleanup);
 680         xprt_wake_pending_tasks(xprt, -EAGAIN);
 681 out:
 682         spin_unlock_bh(&xprt->transport_lock);
 683 }
 684
 685 static bool
 686 xprt_has_timer(const struct rpc_xprt *xprt)
 687 {
 688         return xprt->idle_timeout != 0;
 689 }
 690
 691 static void
 692 xprt_schedule_autodisconnect(struct rpc_xprt *xprt)
 693         __must_hold(&xprt->transport_lock)
 694 {
 695         if (list_empty(&xprt->recv) && xprt_has_timer(xprt))
 696                 mod_timer(&xprt->timer, xprt->last_used + xprt->idle_timeout);
 697 }
 698
 699 static void
 700 xprt_init_autodisconnect(struct timer_list *t)
 701 {
 702         struct rpc_xprt *xprt = from_timer(xprt, t, timer);
 703
 704         spin_lock(&xprt->transport_lock);
 705         if (!list_empty(&xprt->recv))
 706                 goto out_abort;
 707         /* Reset xprt->last_used to avoid connect/autodisconnect cycling */
 708         xprt->last_used = jiffies;
 709         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 710                 goto out_abort;
 711         spin_unlock(&xprt->transport_lock);
 712         queue_work(xprtiod_workqueue, &xprt->task_cleanup);
 713         return;
 714 out_abort:
 715         spin_unlock(&xprt->transport_lock);
 716 }
 717
 718 bool xprt_lock_connect(struct rpc_xprt *xprt,
 719                 struct rpc_task *task,
 720                 void *cookie)
 721 {
 722         bool ret = false;
 723
 724         spin_lock_bh(&xprt->transport_lock);
 725         if (!test_bit(XPRT_LOCKED, &xprt->state))
 726                 goto out;
 727         if (xprt->snd_task != task)
 728                 goto out;
 729         xprt_task_clear_bytes_sent(task);
 730         xprt->snd_task = cookie;
 731         ret = true;
 732 out:
 733         spin_unlock_bh(&xprt->transport_lock);
 734         return ret;
 735 }
 736
 737 void xprt_unlock_connect(struct rpc_xprt *xprt, void *cookie)
 738 {
 739         spin_lock_bh(&xprt->transport_lock);
 740         if (xprt->snd_task != cookie)
 741                 goto out;
 742         if (!test_bit(XPRT_LOCKED, &xprt->state))
 743                 goto out;
 744         xprt->snd_task =NULL;
 745         xprt->ops->release_xprt(xprt, NULL);
 746         xprt_schedule_autodisconnect(xprt);
 747 out:
 748         spin_unlock_bh(&xprt->transport_lock);
 749         wake_up_bit(&xprt->state, XPRT_LOCKED);
 750 }
 751
 752 /**
 753  * xprt_connect - schedule a transport connect operation
 754  * @task: RPC task that is requesting the connect
 755  *
 756  */
 757 void xprt_connect(struct rpc_task *task)
 758 {
 759         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 760
 761         dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
 762                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 763
 764         if (!xprt_bound(xprt)) {
 765                 task->tk_status = -EAGAIN;
 766                 return;
 767         }
 768         if (!xprt_lock_write(xprt, task))
 769                 return;
 770
 771         if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
 772                 xprt->ops->close(xprt);
 773
 774         if (!xprt_connected(xprt)) {
 775                 task->tk_rqstp->rq_bytes_sent = 0;
 776                 task->tk_timeout = task->tk_rqstp->rq_timeout;
 777                 task->tk_rqstp->rq_connect_cookie = xprt->connect_cookie;
 778                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
 779
 780                 if (test_bit(XPRT_CLOSING, &xprt->state))
 781                         return;
 782                 if (xprt_test_and_set_connecting(xprt))
 783                         return;
 784                 xprt->stat.connect_start = jiffies;
 785                 xprt->ops->connect(xprt, task);
 786         }
 787         xprt_release_write(xprt, task);
 788 }
 789
 790 static void xprt_connect_status(struct rpc_task *task)
 791 {
 792         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 793
 794         if (task->tk_status == 0) {
 795                 xprt->stat.connect_count++;
 796                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
 797                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
 798                                 task->tk_pid);
 799                 return;
 800         }
 801
 802         switch (task->tk_status) {
 803         case -ECONNREFUSED:
 804         case -ECONNRESET:
 805         case -ECONNABORTED:
 806         case -ENETUNREACH:
 807         case -EHOSTUNREACH:
 808         case -EPIPE:
 809         case -EAGAIN:
 810                 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
 811                 break;
 812         case -ETIMEDOUT:
 813                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
 814                                 "out\n", task->tk_pid);
 815                 break;
 816         default:
 817                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
 818                                 "server %s\n", task->tk_pid, -task->tk_status,
 819                                 xprt->servername);
 820                 task->tk_status = -EIO;
 821         }
 822 }
 823
 824 /**
 825  * xprt_lookup_rqst - find an RPC request corresponding to an XID
 826  * @xprt: transport on which the original request was transmitted
 827  * @xid: RPC XID of incoming reply
 828  *
 829  * Caller holds xprt->recv_lock.
 830  */
 831 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
 832 {
 833         struct rpc_rqst *entry;
 834
 835         list_for_each_entry(entry, &xprt->recv, rq_list)
 836                 if (entry->rq_xid == xid) {
 837                         trace_xprt_lookup_rqst(xprt, xid, 0);
 838                         entry->rq_rtt = ktime_sub(ktime_get(), entry->rq_xtime);
 839                         return entry;
 840                 }
 841
 842         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
 843                         ntohl(xid));
 844         trace_xprt_lookup_rqst(xprt, xid, -ENOENT);
 845         xprt->stat.bad_xids++;
 846         return NULL;
 847 }
 848 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
 849
 850 /**
 851  * xprt_pin_rqst - Pin a request on the transport receive list
 852  * @req: Request to pin
 853  *
 854  * Caller must ensure this is atomic with the call to xprt_lookup_rqst()
 855  * so should be holding the xprt transport lock.
 856  */
 857 void xprt_pin_rqst(struct rpc_rqst *req)
 858 {
 859         set_bit(RPC_TASK_MSG_RECV, &req->rq_task->tk_runstate);
 860 }
 861 EXPORT_SYMBOL_GPL(xprt_pin_rqst);
 862
 863 /**
 864  * xprt_unpin_rqst - Unpin a request on the transport receive list
 865  * @req: Request to pin
 866  *
 867  * Caller should be holding the xprt transport lock.
 868  */
 869 void xprt_unpin_rqst(struct rpc_rqst *req)
 870 {
 871         struct rpc_task *task = req->rq_task;
 872
 873         clear_bit(RPC_TASK_MSG_RECV, &task->tk_runstate);
 874         if (test_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate))
 875                 wake_up_bit(&task->tk_runstate, RPC_TASK_MSG_RECV);
 876 }
 877 EXPORT_SYMBOL_GPL(xprt_unpin_rqst);
 878
 879 static void xprt_wait_on_pinned_rqst(struct rpc_rqst *req)
 880 __must_hold(&req->rq_xprt->recv_lock)
 881 {
 882         struct rpc_task *task = req->rq_task;
 883
 884         if (task && test_bit(RPC_TASK_MSG_RECV, &task->tk_runstate)) {
 885                 spin_unlock(&req->rq_xprt->recv_lock);
 886                 set_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate);
 887                 wait_on_bit(&task->tk_runstate, RPC_TASK_MSG_RECV,
 888                                 TASK_UNINTERRUPTIBLE);
 889                 clear_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate);
 890                 spin_lock(&req->rq_xprt->recv_lock);
 891         }
 892 }
 893
 894 /**
 895  * xprt_update_rtt - Update RPC RTT statistics
 896  * @task: RPC request that recently completed
 897  *
 898  * Caller holds xprt->recv_lock.
 899  */
 900 void xprt_update_rtt(struct rpc_task *task)
 901 {
 902         struct rpc_rqst *req = task->tk_rqstp;
 903         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
 904         unsigned int timer = task->tk_msg.rpc_proc->p_timer;
 905         long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
 906
 907         if (timer) {
 908                 if (req->rq_ntrans == 1)
 909                         rpc_update_rtt(rtt, timer, m);
 910                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
 911         }
 912 }
 913 EXPORT_SYMBOL_GPL(xprt_update_rtt);
 914
 915 /**
 916  * xprt_complete_rqst - called when reply processing is complete
 917  * @task: RPC request that recently completed
 918  * @copied: actual number of bytes received from the transport
 919  *
 920  * Caller holds xprt->recv_lock.
 921  */
 922 void xprt_complete_rqst(struct rpc_task *task, int copied)
 923 {
 924         struct rpc_rqst *req = task->tk_rqstp;
 925         struct rpc_xprt *xprt = req->rq_xprt;
 926
 927         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
 928                         task->tk_pid, ntohl(req->rq_xid), copied);
 929         trace_xprt_complete_rqst(xprt, req->rq_xid, copied);
 930
 931         xprt->stat.recvs++;
 932
 933         list_del_init(&req->rq_list);
 934         req->rq_private_buf.len = copied;
 935         /* Ensure all writes are done before we update */
 936         /* req->rq_reply_bytes_recvd */
 937         smp_wmb();
 938         req->rq_reply_bytes_recvd = copied;
 939         rpc_wake_up_queued_task(&xprt->pending, task);
 940 }
 941 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
 942
 943 static void xprt_timer(struct rpc_task *task)
 944 {
 945         struct rpc_rqst *req = task->tk_rqstp;
 946         struct rpc_xprt *xprt = req->rq_xprt;
 947
 948         if (task->tk_status != -ETIMEDOUT)
 949                 return;
 950
 951         trace_xprt_timer(xprt, req->rq_xid, task->tk_status);
 952         if (!req->rq_reply_bytes_recvd) {
 953                 if (xprt->ops->timer)
 954                         xprt->ops->timer(xprt, task);
 955         } else
 956                 task->tk_status = 0;
 957 }
 958
 959 /**
 960  * xprt_prepare_transmit - reserve the transport before sending a request
 961  * @task: RPC task about to send a request
 962  *
 963  */
 964 bool xprt_prepare_transmit(struct rpc_task *task)
 965 {
 966         struct rpc_rqst *req = task->tk_rqstp;
 967         struct rpc_xprt *xprt = req->rq_xprt;
 968         bool ret = false;
 969
 970         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
 971
 972         spin_lock_bh(&xprt->transport_lock);
 973         if (!req->rq_bytes_sent) {
 974                 if (req->rq_reply_bytes_recvd) {
 975                         task->tk_status = req->rq_reply_bytes_recvd;
 976                         goto out_unlock;
 977                 }
 978                 if ((task->tk_flags & RPC_TASK_NO_RETRANS_TIMEOUT)
 979                     && xprt_connected(xprt)
 980                     && req->rq_connect_cookie == xprt->connect_cookie) {
 981                         xprt->ops->set_retrans_timeout(task);
 982                         rpc_sleep_on(&xprt->pending, task, xprt_timer);
 983                         goto out_unlock;
 984                 }
 985         }
 986         if (!xprt->ops->reserve_xprt(xprt, task)) {
 987                 task->tk_status = -EAGAIN;
 988                 goto out_unlock;
 989         }
 990         if (!bc_prealloc(req) && !req->rq_xmit_bytes_sent)
 991                 req->rq_xid = xprt_alloc_xid(xprt);
 992         ret = true;
 993 out_unlock:
 994         spin_unlock_bh(&xprt->transport_lock);
 995         return ret;
 996 }
 997
 998 void xprt_end_transmit(struct rpc_task *task)
 999 {
1000         xprt_release_write(task->tk_rqstp->rq_xprt, task);
1001 }
1002
1003 /**
1004  * xprt_transmit - send an RPC request on a transport
1005  * @task: controlling RPC task
1006  *
1007  * We have to copy the iovec because sendmsg fiddles with its contents.
1008  */
1009 void xprt_transmit(struct rpc_task *task)
1010 {
1011         struct rpc_rqst *req = task->tk_rqstp;
1012         struct rpc_xprt *xprt = req->rq_xprt;
1013         unsigned int connect_cookie;
1014         int status;
1015
1016         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
1017
1018         if (!req->rq_reply_bytes_recvd) {
1019                 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) {
1020                         /*
1021                          * Add to the list only if we're expecting a reply
1022                          */
1023                         /* Update the softirq receive buffer */
1024                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
1025                                         sizeof(req->rq_private_buf));
1026                         /* Add request to the receive list */
1027                         spin_lock(&xprt->recv_lock);
1028                         list_add_tail(&req->rq_list, &xprt->recv);
1029                         spin_unlock(&xprt->recv_lock);
1030                         xprt_reset_majortimeo(req);
1031                         /* Turn off autodisconnect */
1032                         del_singleshot_timer_sync(&xprt->timer);
1033                 }
1034         } else if (!req->rq_bytes_sent)
1035                 return;
1036
1037         connect_cookie = xprt->connect_cookie;
1038         status = xprt->ops->send_request(task);
1039         trace_xprt_transmit(xprt, req->rq_xid, status);
1040         if (status != 0) {
1041                 task->tk_status = status;
1042                 return;
1043         }
1044         xprt_inject_disconnect(xprt);
1045
1046         dprintk("RPC: %5u xmit complete\n", task->tk_pid);
1047         task->tk_flags |= RPC_TASK_SENT;
1048         spin_lock_bh(&xprt->transport_lock);
1049
1050         xprt->ops->set_retrans_timeout(task);
1051
1052         xprt->stat.sends++;
1053         xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
1054         xprt->stat.bklog_u += xprt->backlog.qlen;
1055         xprt->stat.sending_u += xprt->sending.qlen;
1056         xprt->stat.pending_u += xprt->pending.qlen;
1057         spin_unlock_bh(&xprt->transport_lock);
1058
1059         req->rq_connect_cookie = connect_cookie;
1060         if (rpc_reply_expected(task) && !READ_ONCE(req->rq_reply_bytes_recvd)) {
1061                 /*
1062                  * Sleep on the pending queue if we're expecting a reply.
1063                  * The spinlock ensures atomicity between the test of
1064                  * req->rq_reply_bytes_recvd, and the call to rpc_sleep_on().
1065                  */
1066                 spin_lock(&xprt->recv_lock);
1067                 if (!req->rq_reply_bytes_recvd) {
1068                         rpc_sleep_on(&xprt->pending, task, xprt_timer);
1069                         /*
1070                          * Send an extra queue wakeup call if the
1071                          * connection was dropped in case the call to
1072                          * rpc_sleep_on() raced.
1073                          */
1074                         if (!xprt_connected(xprt))
1075                                 xprt_wake_pending_tasks(xprt, -ENOTCONN);
1076                 }
1077                 spin_unlock(&xprt->recv_lock);
1078         }
1079 }
1080
1081 static void xprt_add_backlog(struct rpc_xprt *xprt, struct rpc_task *task)
1082 {
1083         set_bit(XPRT_CONGESTED, &xprt->state);
1084         rpc_sleep_on(&xprt->backlog, task, NULL);
1085 }
1086
1087 static void xprt_wake_up_backlog(struct rpc_xprt *xprt)
1088 {
1089         if (rpc_wake_up_next(&xprt->backlog) == NULL)
1090                 clear_bit(XPRT_CONGESTED, &xprt->state);
1091 }
1092
1093 static bool xprt_throttle_congested(struct rpc_xprt *xprt, struct rpc_task *task)
1094 {
1095         bool ret = false;
1096
1097         if (!test_bit(XPRT_CONGESTED, &xprt->state))
1098                 goto out;
1099         spin_lock(&xprt->reserve_lock);
1100         if (test_bit(XPRT_CONGESTED, &xprt->state)) {
1101                 rpc_sleep_on(&xprt->backlog, task, NULL);
1102                 ret = true;
1103         }
1104         spin_unlock(&xprt->reserve_lock);
1105 out:
1106         return ret;
1107 }
1108
1109 static struct rpc_rqst *xprt_dynamic_alloc_slot(struct rpc_xprt *xprt)
1110 {
1111         struct rpc_rqst *req = ERR_PTR(-EAGAIN);
1112
1113         if (xprt->num_reqs >= xprt->max_reqs)
1114                 goto out;
1115         ++xprt->num_reqs;
1116         spin_unlock(&xprt->reserve_lock);
1117         req = kzalloc(sizeof(struct rpc_rqst), GFP_NOFS);
1118         spin_lock(&xprt->reserve_lock);
1119         if (req != NULL)
1120                 goto out;
1121         --xprt->num_reqs;
1122         req = ERR_PTR(-ENOMEM);
1123 out:
1124         return req;
1125 }
1126
1127 static bool xprt_dynamic_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1128 {
1129         if (xprt->num_reqs > xprt->min_reqs) {
1130                 --xprt->num_reqs;
1131                 kfree(req);
1132                 return true;
1133         }
1134         return false;
1135 }
1136
1137 void xprt_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1138 {
1139         struct rpc_rqst *req;
1140
1141         spin_lock(&xprt->reserve_lock);
1142         if (!list_empty(&xprt->free)) {
1143                 req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1144                 list_del(&req->rq_list);
1145                 goto out_init_req;
1146         }
1147         req = xprt_dynamic_alloc_slot(xprt);
1148         if (!IS_ERR(req))
1149                 goto out_init_req;
1150         switch (PTR_ERR(req)) {
1151         case -ENOMEM:
1152                 dprintk("RPC:       dynamic allocation of request slot "
1153                                 "failed! Retrying\n");
1154                 task->tk_status = -ENOMEM;
1155                 break;
1156         case -EAGAIN:
1157                 xprt_add_backlog(xprt, task);
1158                 dprintk("RPC:       waiting for request slot\n");
1159                 /* fall through */
1160         default:
1161                 task->tk_status = -EAGAIN;
1162         }
1163         spin_unlock(&xprt->reserve_lock);
1164         return;
1165 out_init_req:
1166         xprt->stat.max_slots = max_t(unsigned int, xprt->stat.max_slots,
1167                                      xprt->num_reqs);
1168         spin_unlock(&xprt->reserve_lock);
1169
1170         task->tk_status = 0;
1171         task->tk_rqstp = req;
1172 }
1173 EXPORT_SYMBOL_GPL(xprt_alloc_slot);
1174
1175 void xprt_lock_and_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1176 {
1177         /* Note: grabbing the xprt_lock_write() ensures that we throttle
1178          * new slot allocation if the transport is congested (i.e. when
1179          * reconnecting a stream transport or when out of socket write
1180          * buffer space).
1181          */
1182         if (xprt_lock_write(xprt, task)) {
1183                 xprt_alloc_slot(xprt, task);
1184                 xprt_release_write(xprt, task);
1185         }
1186 }
1187 EXPORT_SYMBOL_GPL(xprt_lock_and_alloc_slot);
1188
1189 void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1190 {
1191         spin_lock(&xprt->reserve_lock);
1192         if (!xprt_dynamic_free_slot(xprt, req)) {
1193                 memset(req, 0, sizeof(*req));   /* mark unused */
1194                 list_add(&req->rq_list, &xprt->free);
1195         }
1196         xprt_wake_up_backlog(xprt);
1197         spin_unlock(&xprt->reserve_lock);
1198 }
1199 EXPORT_SYMBOL_GPL(xprt_free_slot);
1200
1201 static void xprt_free_all_slots(struct rpc_xprt *xprt)
1202 {
1203         struct rpc_rqst *req;
1204         while (!list_empty(&xprt->free)) {
1205                 req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list);
1206                 list_del(&req->rq_list);
1207                 kfree(req);
1208         }
1209 }
1210
1211 struct rpc_xprt *xprt_alloc(struct net *net, size_t size,
1212                 unsigned int num_prealloc,
1213                 unsigned int max_alloc)
1214 {
1215         struct rpc_xprt *xprt;
1216         struct rpc_rqst *req;
1217         int i;
1218
1219         xprt = kzalloc(size, GFP_KERNEL);
1220         if (xprt == NULL)
1221                 goto out;
1222
1223         xprt_init(xprt, net);
1224
1225         for (i = 0; i < num_prealloc; i++) {
1226                 req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL);
1227                 if (!req)
1228                         goto out_free;
1229                 list_add(&req->rq_list, &xprt->free);
1230         }
1231         if (max_alloc > num_prealloc)
1232                 xprt->max_reqs = max_alloc;
1233         else
1234                 xprt->max_reqs = num_prealloc;
1235         xprt->min_reqs = num_prealloc;
1236         xprt->num_reqs = num_prealloc;
1237
1238         return xprt;
1239
1240 out_free:
1241         xprt_free(xprt);
1242 out:
1243         return NULL;
1244 }
1245 EXPORT_SYMBOL_GPL(xprt_alloc);
1246
1247 void xprt_free(struct rpc_xprt *xprt)
1248 {
1249         put_net(xprt->xprt_net);
1250         xprt_free_all_slots(xprt);
1251         kfree_rcu(xprt, rcu);
1252 }
1253 EXPORT_SYMBOL_GPL(xprt_free);
1254
1255 /**
1256  * xprt_reserve - allocate an RPC request slot
1257  * @task: RPC task requesting a slot allocation
1258  *
1259  * If the transport is marked as being congested, or if no more
1260  * slots are available, place the task on the transport's
1261  * backlog queue.
1262  */
1263 void xprt_reserve(struct rpc_task *task)
1264 {
1265         struct rpc_xprt *xprt = task->tk_xprt;
1266
1267         task->tk_status = 0;
1268         if (task->tk_rqstp != NULL)
1269                 return;
1270
1271         task->tk_timeout = 0;
1272         task->tk_status = -EAGAIN;
1273         if (!xprt_throttle_congested(xprt, task))
1274                 xprt->ops->alloc_slot(xprt, task);
1275 }
1276
1277 /**
1278  * xprt_retry_reserve - allocate an RPC request slot
1279  * @task: RPC task requesting a slot allocation
1280  *
1281  * If no more slots are available, place the task on the transport's
1282  * backlog queue.
1283  * Note that the only difference with xprt_reserve is that we now
1284  * ignore the value of the XPRT_CONGESTED flag.
1285  */
1286 void xprt_retry_reserve(struct rpc_task *task)
1287 {
1288         struct rpc_xprt *xprt = task->tk_xprt;
1289
1290         task->tk_status = 0;
1291         if (task->tk_rqstp != NULL)
1292                 return;
1293
1294         task->tk_timeout = 0;
1295         task->tk_status = -EAGAIN;
1296         xprt->ops->alloc_slot(xprt, task);
1297 }
1298
1299 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
1300 {
1301         return (__force __be32)xprt->xid++;
1302 }
1303
1304 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1305 {
1306         xprt->xid = prandom_u32();
1307 }
1308
1309 void xprt_request_init(struct rpc_task *task)
1310 {
1311         struct rpc_xprt *xprt = task->tk_xprt;
1312         struct rpc_rqst *req = task->tk_rqstp;
1313
1314         INIT_LIST_HEAD(&req->rq_list);
1315         req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1316         req->rq_task    = task;
1317         req->rq_xprt    = xprt;
1318         req->rq_buffer  = NULL;
1319         req->rq_connect_cookie = xprt->connect_cookie - 1;
1320         req->rq_bytes_sent = 0;
1321         req->rq_snd_buf.len = 0;
1322         req->rq_snd_buf.buflen = 0;
1323         req->rq_rcv_buf.len = 0;
1324         req->rq_rcv_buf.buflen = 0;
1325         req->rq_release_snd_buf = NULL;
1326         xprt_reset_majortimeo(req);
1327         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1328                         req, ntohl(req->rq_xid));
1329 }
1330
1331 /**
1332  * xprt_release - release an RPC request slot
1333  * @task: task which is finished with the slot
1334  *
1335  */
1336 void xprt_release(struct rpc_task *task)
1337 {
1338         struct rpc_xprt *xprt;
1339         struct rpc_rqst *req = task->tk_rqstp;
1340
1341         if (req == NULL) {
1342                 if (task->tk_client) {
1343                         xprt = task->tk_xprt;
1344                         if (xprt->snd_task == task)
1345                                 xprt_release_write(xprt, task);
1346                 }
1347                 return;
1348         }
1349
1350         xprt = req->rq_xprt;
1351         if (task->tk_ops->rpc_count_stats != NULL)
1352                 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
1353         else if (task->tk_client)
1354                 rpc_count_iostats(task, task->tk_client->cl_metrics);
1355         spin_lock(&xprt->recv_lock);
1356         if (!list_empty(&req->rq_list)) {
1357                 list_del_init(&req->rq_list);
1358                 xprt_wait_on_pinned_rqst(req);
1359         }
1360         spin_unlock(&xprt->recv_lock);
1361         spin_lock_bh(&xprt->transport_lock);
1362         xprt->ops->release_xprt(xprt, task);
1363         if (xprt->ops->release_request)
1364                 xprt->ops->release_request(task);
1365         xprt->last_used = jiffies;
1366         xprt_schedule_autodisconnect(xprt);
1367         spin_unlock_bh(&xprt->transport_lock);
1368         if (req->rq_buffer)
1369                 xprt->ops->buf_free(task);
1370         xprt_inject_disconnect(xprt);
1371         if (req->rq_cred != NULL)
1372                 put_rpccred(req->rq_cred);
1373         task->tk_rqstp = NULL;
1374         if (req->rq_release_snd_buf)
1375                 req->rq_release_snd_buf(req);
1376
1377         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1378         if (likely(!bc_prealloc(req)))
1379                 xprt->ops->free_slot(xprt, req);
1380         else
1381                 xprt_free_bc_request(req);
1382 }
1383
1384 static void xprt_init(struct rpc_xprt *xprt, struct net *net)
1385 {
1386         kref_init(&xprt->kref);
1387
1388         spin_lock_init(&xprt->transport_lock);
1389         spin_lock_init(&xprt->reserve_lock);
1390         spin_lock_init(&xprt->recv_lock);
1391
1392         INIT_LIST_HEAD(&xprt->free);
1393         INIT_LIST_HEAD(&xprt->recv);
1394 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1395         spin_lock_init(&xprt->bc_pa_lock);
1396         INIT_LIST_HEAD(&xprt->bc_pa_list);
1397 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1398         INIT_LIST_HEAD(&xprt->xprt_switch);
1399
1400         xprt->last_used = jiffies;
1401         xprt->cwnd = RPC_INITCWND;
1402         xprt->bind_index = 0;
1403
1404         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1405         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1406         rpc_init_priority_wait_queue(&xprt->sending, "xprt_sending");
1407         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1408
1409         xprt_init_xid(xprt);
1410
1411         xprt->xprt_net = get_net(net);
1412 }
1413
1414 /**
1415  * xprt_create_transport - create an RPC transport
1416  * @args: rpc transport creation arguments
1417  *
1418  */
1419 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1420 {
1421         struct rpc_xprt *xprt;
1422         struct xprt_class *t;
1423
1424         spin_lock(&xprt_list_lock);
1425         list_for_each_entry(t, &xprt_list, list) {
1426                 if (t->ident == args->ident) {
1427                         spin_unlock(&xprt_list_lock);
1428                         goto found;
1429                 }
1430         }
1431         spin_unlock(&xprt_list_lock);
1432         dprintk("RPC: transport (%d) not supported\n", args->ident);
1433         return ERR_PTR(-EIO);
1434
1435 found:
1436         xprt = t->setup(args);
1437         if (IS_ERR(xprt)) {
1438                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1439                                 -PTR_ERR(xprt));
1440                 goto out;
1441         }
1442         if (args->flags & XPRT_CREATE_NO_IDLE_TIMEOUT)
1443                 xprt->idle_timeout = 0;
1444         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1445         if (xprt_has_timer(xprt))
1446                 timer_setup(&xprt->timer, xprt_init_autodisconnect, 0);
1447         else
1448                 timer_setup(&xprt->timer, NULL, 0);
1449
1450         if (strlen(args->servername) > RPC_MAXNETNAMELEN) {
1451                 xprt_destroy(xprt);
1452                 return ERR_PTR(-EINVAL);
1453         }
1454         xprt->servername = kstrdup(args->servername, GFP_KERNEL);
1455         if (xprt->servername == NULL) {
1456                 xprt_destroy(xprt);
1457                 return ERR_PTR(-ENOMEM);
1458         }
1459
1460         rpc_xprt_debugfs_register(xprt);
1461
1462         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1463                         xprt->max_reqs);
1464 out:
1465         return xprt;
1466 }
1467
1468 static void xprt_destroy_cb(struct work_struct *work)
1469 {
1470         struct rpc_xprt *xprt =
1471                 container_of(work, struct rpc_xprt, task_cleanup);
1472
1473         rpc_xprt_debugfs_unregister(xprt);
1474         rpc_destroy_wait_queue(&xprt->binding);
1475         rpc_destroy_wait_queue(&xprt->pending);
1476         rpc_destroy_wait_queue(&xprt->sending);
1477         rpc_destroy_wait_queue(&xprt->backlog);
1478         kfree(xprt->servername);
1479         /*
1480          * Tear down transport state and free the rpc_xprt
1481          */
1482         xprt->ops->destroy(xprt);
1483 }
1484
1485 /**
1486  * xprt_destroy - destroy an RPC transport, killing off all requests.
1487  * @xprt: transport to destroy
1488  *
1489  */
1490 static void xprt_destroy(struct rpc_xprt *xprt)
1491 {
1492         dprintk("RPC:       destroying transport %p\n", xprt);
1493
1494         /*
1495          * Exclude transport connect/disconnect handlers and autoclose
1496          */
1497         wait_on_bit_lock(&xprt->state, XPRT_LOCKED, TASK_UNINTERRUPTIBLE);
1498
1499         del_timer_sync(&xprt->timer);
1500
1501         /*
1502          * Destroy sockets etc from the system workqueue so they can
1503          * safely flush receive work running on rpciod.
1504          */
1505         INIT_WORK(&xprt->task_cleanup, xprt_destroy_cb);
1506         schedule_work(&xprt->task_cleanup);
1507 }
1508
1509 static void xprt_destroy_kref(struct kref *kref)
1510 {
1511         xprt_destroy(container_of(kref, struct rpc_xprt, kref));
1512 }
1513
1514 /**
1515  * xprt_get - return a reference to an RPC transport.
1516  * @xprt: pointer to the transport
1517  *
1518  */
1519 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1520 {
1521         if (xprt != NULL && kref_get_unless_zero(&xprt->kref))
1522                 return xprt;
1523         return NULL;
1524 }
1525 EXPORT_SYMBOL_GPL(xprt_get);
1526
1527 /**
1528  * xprt_put - release a reference to an RPC transport.
1529  * @xprt: pointer to the transport
1530  *
1531  */
1532 void xprt_put(struct rpc_xprt *xprt)
1533 {
1534         if (xprt != NULL)
1535                 kref_put(&xprt->kref, xprt_destroy_kref);
1536 }
1537 EXPORT_SYMBOL_GPL(xprt_put);