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_request_init(struct rpc_task *, struct rpc_xprt *);
  66 static void     xprt_connect_status(struct rpc_task *task);
  67 static int      __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
  68
  69 static DEFINE_SPINLOCK(xprt_list_lock);
  70 static LIST_HEAD(xprt_list);
  71
  72 /*
  73  * The transport code maintains an estimate on the maximum number of out-
  74  * standing RPC requests, using a smoothed version of the congestion
  75  * avoidance implemented in 44BSD. This is basically the Van Jacobson
  76  * congestion algorithm: If a retransmit occurs, the congestion window is
  77  * halved; otherwise, it is incremented by 1/cwnd when
  78  *
  79  *      -       a reply is received and
  80  *      -       a full number of requests are outstanding and
  81  *      -       the congestion window hasn't been updated recently.
  82  */
  83 #define RPC_CWNDSHIFT           (8U)
  84 #define RPC_CWNDSCALE           (1U << RPC_CWNDSHIFT)
  85 #define RPC_INITCWND            RPC_CWNDSCALE
  86 #define RPC_MAXCWND(xprt)       ((xprt)->max_reqs << RPC_CWNDSHIFT)
  87
  88 #define RPCXPRT_CONGESTED(xprt) ((xprt)->cong >= (xprt)->cwnd)
  89
  90 /**
  91  * xprt_register_transport - register a transport implementation
  92  * @transport: transport to register
  93  *
  94  * If a transport implementation is loaded as a kernel module, it can
  95  * call this interface to make itself known to the RPC client.
  96  *
  97  * Returns:
  98  * 0:           transport successfully registered
  99  * -EEXIST:     transport already registered
 100  * -EINVAL:     transport module being unloaded
 101  */
 102 int xprt_register_transport(struct xprt_class *transport)
 103 {
 104         struct xprt_class *t;
 105         int result;
 106
 107         result = -EEXIST;
 108         spin_lock(&xprt_list_lock);
 109         list_for_each_entry(t, &xprt_list, list) {
 110                 /* don't register the same transport class twice */
 111                 if (t->ident == transport->ident)
 112                         goto out;
 113         }
 114
 115         list_add_tail(&transport->list, &xprt_list);
 116         printk(KERN_INFO "RPC: Registered %s transport module.\n",
 117                transport->name);
 118         result = 0;
 119
 120 out:
 121         spin_unlock(&xprt_list_lock);
 122         return result;
 123 }
 124 EXPORT_SYMBOL_GPL(xprt_register_transport);
 125
 126 /**
 127  * xprt_unregister_transport - unregister a transport implementation
 128  * @transport: transport to unregister
 129  *
 130  * Returns:
 131  * 0:           transport successfully unregistered
 132  * -ENOENT:     transport never registered
 133  */
 134 int xprt_unregister_transport(struct xprt_class *transport)
 135 {
 136         struct xprt_class *t;
 137         int result;
 138
 139         result = 0;
 140         spin_lock(&xprt_list_lock);
 141         list_for_each_entry(t, &xprt_list, list) {
 142                 if (t == transport) {
 143                         printk(KERN_INFO
 144                                 "RPC: Unregistered %s transport module.\n",
 145                                 transport->name);
 146                         list_del_init(&transport->list);
 147                         goto out;
 148                 }
 149         }
 150         result = -ENOENT;
 151
 152 out:
 153         spin_unlock(&xprt_list_lock);
 154         return result;
 155 }
 156 EXPORT_SYMBOL_GPL(xprt_unregister_transport);
 157
 158 /**
 159  * xprt_load_transport - load a transport implementation
 160  * @transport_name: transport to load
 161  *
 162  * Returns:
 163  * 0:           transport successfully loaded
 164  * -ENOENT:     transport module not available
 165  */
 166 int xprt_load_transport(const char *transport_name)
 167 {
 168         struct xprt_class *t;
 169         int result;
 170
 171         result = 0;
 172         spin_lock(&xprt_list_lock);
 173         list_for_each_entry(t, &xprt_list, list) {
 174                 if (strcmp(t->name, transport_name) == 0) {
 175                         spin_unlock(&xprt_list_lock);
 176                         goto out;
 177                 }
 178         }
 179         spin_unlock(&xprt_list_lock);
 180         result = request_module("xprt%s", transport_name);
 181 out:
 182         return result;
 183 }
 184 EXPORT_SYMBOL_GPL(xprt_load_transport);
 185
 186 /**
 187  * xprt_reserve_xprt - serialize write access to transports
 188  * @task: task that is requesting access to the transport
 189  *
 190  * This prevents mixing the payload of separate requests, and prevents
 191  * transport connects from colliding with writes.  No congestion control
 192  * is provided.
 193  */
 194 int xprt_reserve_xprt(struct rpc_task *task)
 195 {
 196         struct rpc_rqst *req = task->tk_rqstp;
 197         struct rpc_xprt *xprt = req->rq_xprt;
 198
 199         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 200                 if (task == xprt->snd_task)
 201                         return 1;
 202                 goto out_sleep;
 203         }
 204         xprt->snd_task = task;
 205         if (req) {
 206                 req->rq_bytes_sent = 0;
 207                 req->rq_ntrans++;
 208         }
 209         return 1;
 210
 211 out_sleep:
 212         dprintk("RPC: %5u failed to lock transport %p\n",
 213                         task->tk_pid, xprt);
 214         task->tk_timeout = 0;
 215         task->tk_status = -EAGAIN;
 216         if (req && req->rq_ntrans)
 217                 rpc_sleep_on(&xprt->resend, task, NULL);
 218         else
 219                 rpc_sleep_on(&xprt->sending, task, NULL);
 220         return 0;
 221 }
 222 EXPORT_SYMBOL_GPL(xprt_reserve_xprt);
 223
 224 static void xprt_clear_locked(struct rpc_xprt *xprt)
 225 {
 226         xprt->snd_task = NULL;
 227         if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state) || xprt->shutdown) {
 228                 smp_mb__before_clear_bit();
 229                 clear_bit(XPRT_LOCKED, &xprt->state);
 230                 smp_mb__after_clear_bit();
 231         } else
 232                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 233 }
 234
 235 /*
 236  * xprt_reserve_xprt_cong - serialize write access to transports
 237  * @task: task that is requesting access to the transport
 238  *
 239  * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
 240  * integrated into the decision of whether a request is allowed to be
 241  * woken up and given access to the transport.
 242  */
 243 int xprt_reserve_xprt_cong(struct rpc_task *task)
 244 {
 245         struct rpc_xprt *xprt = task->tk_xprt;
 246         struct rpc_rqst *req = task->tk_rqstp;
 247
 248         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 249                 if (task == xprt->snd_task)
 250                         return 1;
 251                 goto out_sleep;
 252         }
 253         if (__xprt_get_cong(xprt, task)) {
 254                 xprt->snd_task = task;
 255                 if (req) {
 256                         req->rq_bytes_sent = 0;
 257                         req->rq_ntrans++;
 258                 }
 259                 return 1;
 260         }
 261         xprt_clear_locked(xprt);
 262 out_sleep:
 263         dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
 264         task->tk_timeout = 0;
 265         task->tk_status = -EAGAIN;
 266         if (req && req->rq_ntrans)
 267                 rpc_sleep_on(&xprt->resend, task, NULL);
 268         else
 269                 rpc_sleep_on(&xprt->sending, task, NULL);
 270         return 0;
 271 }
 272 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
 273
 274 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 275 {
 276         int retval;
 277
 278         spin_lock_bh(&xprt->transport_lock);
 279         retval = xprt->ops->reserve_xprt(task);
 280         spin_unlock_bh(&xprt->transport_lock);
 281         return retval;
 282 }
 283
 284 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
 285 {
 286         struct rpc_task *task;
 287         struct rpc_rqst *req;
 288
 289         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 290                 return;
 291
 292         task = rpc_wake_up_next(&xprt->resend);
 293         if (!task) {
 294                 task = rpc_wake_up_next(&xprt->sending);
 295                 if (!task)
 296                         goto out_unlock;
 297         }
 298
 299         req = task->tk_rqstp;
 300         xprt->snd_task = task;
 301         if (req) {
 302                 req->rq_bytes_sent = 0;
 303                 req->rq_ntrans++;
 304         }
 305         return;
 306
 307 out_unlock:
 308         xprt_clear_locked(xprt);
 309 }
 310
 311 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
 312 {
 313         struct rpc_task *task;
 314
 315         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 316                 return;
 317         if (RPCXPRT_CONGESTED(xprt))
 318                 goto out_unlock;
 319         task = rpc_wake_up_next(&xprt->resend);
 320         if (!task) {
 321                 task = rpc_wake_up_next(&xprt->sending);
 322                 if (!task)
 323                         goto out_unlock;
 324         }
 325         if (__xprt_get_cong(xprt, task)) {
 326                 struct rpc_rqst *req = task->tk_rqstp;
 327                 xprt->snd_task = task;
 328                 if (req) {
 329                         req->rq_bytes_sent = 0;
 330                         req->rq_ntrans++;
 331                 }
 332                 return;
 333         }
 334 out_unlock:
 335         xprt_clear_locked(xprt);
 336 }
 337
 338 /**
 339  * xprt_release_xprt - allow other requests to use a transport
 340  * @xprt: transport with other tasks potentially waiting
 341  * @task: task that is releasing access to the transport
 342  *
 343  * Note that "task" can be NULL.  No congestion control is provided.
 344  */
 345 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 346 {
 347         if (xprt->snd_task == task) {
 348                 xprt_clear_locked(xprt);
 349                 __xprt_lock_write_next(xprt);
 350         }
 351 }
 352 EXPORT_SYMBOL_GPL(xprt_release_xprt);
 353
 354 /**
 355  * xprt_release_xprt_cong - allow other requests to use a transport
 356  * @xprt: transport with other tasks potentially waiting
 357  * @task: task that is releasing access to the transport
 358  *
 359  * Note that "task" can be NULL.  Another task is awoken to use the
 360  * transport if the transport's congestion window allows it.
 361  */
 362 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 363 {
 364         if (xprt->snd_task == task) {
 365                 xprt_clear_locked(xprt);
 366                 __xprt_lock_write_next_cong(xprt);
 367         }
 368 }
 369 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
 370
 371 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 372 {
 373         spin_lock_bh(&xprt->transport_lock);
 374         xprt->ops->release_xprt(xprt, task);
 375         spin_unlock_bh(&xprt->transport_lock);
 376 }
 377
 378 /*
 379  * Van Jacobson congestion avoidance. Check if the congestion window
 380  * overflowed. Put the task to sleep if this is the case.
 381  */
 382 static int
 383 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 384 {
 385         struct rpc_rqst *req = task->tk_rqstp;
 386
 387         if (req->rq_cong)
 388                 return 1;
 389         dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
 390                         task->tk_pid, xprt->cong, xprt->cwnd);
 391         if (RPCXPRT_CONGESTED(xprt))
 392                 return 0;
 393         req->rq_cong = 1;
 394         xprt->cong += RPC_CWNDSCALE;
 395         return 1;
 396 }
 397
 398 /*
 399  * Adjust the congestion window, and wake up the next task
 400  * that has been sleeping due to congestion
 401  */
 402 static void
 403 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 404 {
 405         if (!req->rq_cong)
 406                 return;
 407         req->rq_cong = 0;
 408         xprt->cong -= RPC_CWNDSCALE;
 409         __xprt_lock_write_next_cong(xprt);
 410 }
 411
 412 /**
 413  * xprt_release_rqst_cong - housekeeping when request is complete
 414  * @task: RPC request that recently completed
 415  *
 416  * Useful for transports that require congestion control.
 417  */
 418 void xprt_release_rqst_cong(struct rpc_task *task)
 419 {
 420         __xprt_put_cong(task->tk_xprt, task->tk_rqstp);
 421 }
 422 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
 423
 424 /**
 425  * xprt_adjust_cwnd - adjust transport congestion window
 426  * @task: recently completed RPC request used to adjust window
 427  * @result: result code of completed RPC request
 428  *
 429  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
 430  */
 431 void xprt_adjust_cwnd(struct rpc_task *task, int result)
 432 {
 433         struct rpc_rqst *req = task->tk_rqstp;
 434         struct rpc_xprt *xprt = task->tk_xprt;
 435         unsigned long cwnd = xprt->cwnd;
 436
 437         if (result >= 0 && cwnd <= xprt->cong) {
 438                 /* The (cwnd >> 1) term makes sure
 439                  * the result gets rounded properly. */
 440                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 441                 if (cwnd > RPC_MAXCWND(xprt))
 442                         cwnd = RPC_MAXCWND(xprt);
 443                 __xprt_lock_write_next_cong(xprt);
 444         } else if (result == -ETIMEDOUT) {
 445                 cwnd >>= 1;
 446                 if (cwnd < RPC_CWNDSCALE)
 447                         cwnd = RPC_CWNDSCALE;
 448         }
 449         dprintk("RPC:       cong %ld, cwnd was %ld, now %ld\n",
 450                         xprt->cong, xprt->cwnd, cwnd);
 451         xprt->cwnd = cwnd;
 452         __xprt_put_cong(xprt, req);
 453 }
 454 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
 455
 456 /**
 457  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
 458  * @xprt: transport with waiting tasks
 459  * @status: result code to plant in each task before waking it
 460  *
 461  */
 462 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
 463 {
 464         if (status < 0)
 465                 rpc_wake_up_status(&xprt->pending, status);
 466         else
 467                 rpc_wake_up(&xprt->pending);
 468 }
 469 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
 470
 471 /**
 472  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
 473  * @task: task to be put to sleep
 474  * @action: function pointer to be executed after wait
 475  */
 476 void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action)
 477 {
 478         struct rpc_rqst *req = task->tk_rqstp;
 479         struct rpc_xprt *xprt = req->rq_xprt;
 480
 481         task->tk_timeout = req->rq_timeout;
 482         rpc_sleep_on(&xprt->pending, task, action);
 483 }
 484 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
 485
 486 /**
 487  * xprt_write_space - wake the task waiting for transport output buffer space
 488  * @xprt: transport with waiting tasks
 489  *
 490  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
 491  */
 492 void xprt_write_space(struct rpc_xprt *xprt)
 493 {
 494         if (unlikely(xprt->shutdown))
 495                 return;
 496
 497         spin_lock_bh(&xprt->transport_lock);
 498         if (xprt->snd_task) {
 499                 dprintk("RPC:       write space: waking waiting task on "
 500                                 "xprt %p\n", xprt);
 501                 rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task);
 502         }
 503         spin_unlock_bh(&xprt->transport_lock);
 504 }
 505 EXPORT_SYMBOL_GPL(xprt_write_space);
 506
 507 /**
 508  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
 509  * @task: task whose timeout is to be set
 510  *
 511  * Set a request's retransmit timeout based on the transport's
 512  * default timeout parameters.  Used by transports that don't adjust
 513  * the retransmit timeout based on round-trip time estimation.
 514  */
 515 void xprt_set_retrans_timeout_def(struct rpc_task *task)
 516 {
 517         task->tk_timeout = task->tk_rqstp->rq_timeout;
 518 }
 519 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
 520
 521 /*
 522  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
 523  * @task: task whose timeout is to be set
 524  *
 525  * Set a request's retransmit timeout using the RTT estimator.
 526  */
 527 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
 528 {
 529         int timer = task->tk_msg.rpc_proc->p_timer;
 530         struct rpc_clnt *clnt = task->tk_client;
 531         struct rpc_rtt *rtt = clnt->cl_rtt;
 532         struct rpc_rqst *req = task->tk_rqstp;
 533         unsigned long max_timeout = clnt->cl_timeout->to_maxval;
 534
 535         task->tk_timeout = rpc_calc_rto(rtt, timer);
 536         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
 537         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
 538                 task->tk_timeout = max_timeout;
 539 }
 540 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
 541
 542 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 543 {
 544         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 545
 546         req->rq_majortimeo = req->rq_timeout;
 547         if (to->to_exponential)
 548                 req->rq_majortimeo <<= to->to_retries;
 549         else
 550                 req->rq_majortimeo += to->to_increment * to->to_retries;
 551         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 552                 req->rq_majortimeo = to->to_maxval;
 553         req->rq_majortimeo += jiffies;
 554 }
 555
 556 /**
 557  * xprt_adjust_timeout - adjust timeout values for next retransmit
 558  * @req: RPC request containing parameters to use for the adjustment
 559  *
 560  */
 561 int xprt_adjust_timeout(struct rpc_rqst *req)
 562 {
 563         struct rpc_xprt *xprt = req->rq_xprt;
 564         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 565         int status = 0;
 566
 567         if (time_before(jiffies, req->rq_majortimeo)) {
 568                 if (to->to_exponential)
 569                         req->rq_timeout <<= 1;
 570                 else
 571                         req->rq_timeout += to->to_increment;
 572                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 573                         req->rq_timeout = to->to_maxval;
 574                 req->rq_retries++;
 575         } else {
 576                 req->rq_timeout = to->to_initval;
 577                 req->rq_retries = 0;
 578                 xprt_reset_majortimeo(req);
 579                 /* Reset the RTT counters == "slow start" */
 580                 spin_lock_bh(&xprt->transport_lock);
 581                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 582                 spin_unlock_bh(&xprt->transport_lock);
 583                 status = -ETIMEDOUT;
 584         }
 585
 586         if (req->rq_timeout == 0) {
 587                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 588                 req->rq_timeout = 5 * HZ;
 589         }
 590         return status;
 591 }
 592
 593 static void xprt_autoclose(struct work_struct *work)
 594 {
 595         struct rpc_xprt *xprt =
 596                 container_of(work, struct rpc_xprt, task_cleanup);
 597
 598         xprt->ops->close(xprt);
 599         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
 600         xprt_release_write(xprt, NULL);
 601 }
 602
 603 /**
 604  * xprt_disconnect_done - mark a transport as disconnected
 605  * @xprt: transport to flag for disconnect
 606  *
 607  */
 608 void xprt_disconnect_done(struct rpc_xprt *xprt)
 609 {
 610         dprintk("RPC:       disconnected transport %p\n", xprt);
 611         spin_lock_bh(&xprt->transport_lock);
 612         xprt_clear_connected(xprt);
 613         xprt_wake_pending_tasks(xprt, -EAGAIN);
 614         spin_unlock_bh(&xprt->transport_lock);
 615 }
 616 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
 617
 618 /**
 619  * xprt_force_disconnect - force a transport to disconnect
 620  * @xprt: transport to disconnect
 621  *
 622  */
 623 void xprt_force_disconnect(struct rpc_xprt *xprt)
 624 {
 625         /* Don't race with the test_bit() in xprt_clear_locked() */
 626         spin_lock_bh(&xprt->transport_lock);
 627         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 628         /* Try to schedule an autoclose RPC call */
 629         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 630                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 631         xprt_wake_pending_tasks(xprt, -EAGAIN);
 632         spin_unlock_bh(&xprt->transport_lock);
 633 }
 634
 635 /**
 636  * xprt_conditional_disconnect - force a transport to disconnect
 637  * @xprt: transport to disconnect
 638  * @cookie: 'connection cookie'
 639  *
 640  * This attempts to break the connection if and only if 'cookie' matches
 641  * the current transport 'connection cookie'. It ensures that we don't
 642  * try to break the connection more than once when we need to retransmit
 643  * a batch of RPC requests.
 644  *
 645  */
 646 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
 647 {
 648         /* Don't race with the test_bit() in xprt_clear_locked() */
 649         spin_lock_bh(&xprt->transport_lock);
 650         if (cookie != xprt->connect_cookie)
 651                 goto out;
 652         if (test_bit(XPRT_CLOSING, &xprt->state) || !xprt_connected(xprt))
 653                 goto out;
 654         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 655         /* Try to schedule an autoclose RPC call */
 656         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 657                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 658         xprt_wake_pending_tasks(xprt, -EAGAIN);
 659 out:
 660         spin_unlock_bh(&xprt->transport_lock);
 661 }
 662
 663 static void
 664 xprt_init_autodisconnect(unsigned long data)
 665 {
 666         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 667
 668         spin_lock(&xprt->transport_lock);
 669         if (!list_empty(&xprt->recv) || xprt->shutdown)
 670                 goto out_abort;
 671         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 672                 goto out_abort;
 673         spin_unlock(&xprt->transport_lock);
 674         set_bit(XPRT_CONNECTION_CLOSE, &xprt->state);
 675         queue_work(rpciod_workqueue, &xprt->task_cleanup);
 676         return;
 677 out_abort:
 678         spin_unlock(&xprt->transport_lock);
 679 }
 680
 681 /**
 682  * xprt_connect - schedule a transport connect operation
 683  * @task: RPC task that is requesting the connect
 684  *
 685  */
 686 void xprt_connect(struct rpc_task *task)
 687 {
 688         struct rpc_xprt *xprt = task->tk_xprt;
 689
 690         dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
 691                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 692
 693         if (!xprt_bound(xprt)) {
 694                 task->tk_status = -EAGAIN;
 695                 return;
 696         }
 697         if (!xprt_lock_write(xprt, task))
 698                 return;
 699
 700         if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
 701                 xprt->ops->close(xprt);
 702
 703         if (xprt_connected(xprt))
 704                 xprt_release_write(xprt, task);
 705         else {
 706                 if (task->tk_rqstp)
 707                         task->tk_rqstp->rq_bytes_sent = 0;
 708
 709                 task->tk_timeout = task->tk_rqstp->rq_timeout;
 710                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
 711
 712                 if (test_bit(XPRT_CLOSING, &xprt->state))
 713                         return;
 714                 if (xprt_test_and_set_connecting(xprt))
 715                         return;
 716                 xprt->stat.connect_start = jiffies;
 717                 xprt->ops->connect(task);
 718         }
 719 }
 720
 721 static void xprt_connect_status(struct rpc_task *task)
 722 {
 723         struct rpc_xprt *xprt = task->tk_xprt;
 724
 725         if (task->tk_status == 0) {
 726                 xprt->stat.connect_count++;
 727                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
 728                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
 729                                 task->tk_pid);
 730                 return;
 731         }
 732
 733         switch (task->tk_status) {
 734         case -EAGAIN:
 735                 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
 736                 break;
 737         case -ETIMEDOUT:
 738                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
 739                                 "out\n", task->tk_pid);
 740                 break;
 741         default:
 742                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
 743                                 "server %s\n", task->tk_pid, -task->tk_status,
 744                                 task->tk_client->cl_server);
 745                 xprt_release_write(xprt, task);
 746                 task->tk_status = -EIO;
 747         }
 748 }
 749
 750 /**
 751  * xprt_lookup_rqst - find an RPC request corresponding to an XID
 752  * @xprt: transport on which the original request was transmitted
 753  * @xid: RPC XID of incoming reply
 754  *
 755  */
 756 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
 757 {
 758         struct rpc_rqst *entry;
 759
 760         list_for_each_entry(entry, &xprt->recv, rq_list)
 761                 if (entry->rq_xid == xid)
 762                         return entry;
 763
 764         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
 765                         ntohl(xid));
 766         xprt->stat.bad_xids++;
 767         return NULL;
 768 }
 769 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
 770
 771 static void xprt_update_rtt(struct rpc_task *task)
 772 {
 773         struct rpc_rqst *req = task->tk_rqstp;
 774         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
 775         unsigned timer = task->tk_msg.rpc_proc->p_timer;
 776         long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
 777
 778         if (timer) {
 779                 if (req->rq_ntrans == 1)
 780                         rpc_update_rtt(rtt, timer, m);
 781                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
 782         }
 783 }
 784
 785 /**
 786  * xprt_complete_rqst - called when reply processing is complete
 787  * @task: RPC request that recently completed
 788  * @copied: actual number of bytes received from the transport
 789  *
 790  * Caller holds transport lock.
 791  */
 792 void xprt_complete_rqst(struct rpc_task *task, int copied)
 793 {
 794         struct rpc_rqst *req = task->tk_rqstp;
 795         struct rpc_xprt *xprt = req->rq_xprt;
 796
 797         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
 798                         task->tk_pid, ntohl(req->rq_xid), copied);
 799
 800         xprt->stat.recvs++;
 801         req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime);
 802         if (xprt->ops->timer != NULL)
 803                 xprt_update_rtt(task);
 804
 805         list_del_init(&req->rq_list);
 806         req->rq_private_buf.len = copied;
 807         /* Ensure all writes are done before we update */
 808         /* req->rq_reply_bytes_recvd */
 809         smp_wmb();
 810         req->rq_reply_bytes_recvd = copied;
 811         rpc_wake_up_queued_task(&xprt->pending, task);
 812 }
 813 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
 814
 815 static void xprt_timer(struct rpc_task *task)
 816 {
 817         struct rpc_rqst *req = task->tk_rqstp;
 818         struct rpc_xprt *xprt = req->rq_xprt;
 819
 820         if (task->tk_status != -ETIMEDOUT)
 821                 return;
 822         dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
 823
 824         spin_lock_bh(&xprt->transport_lock);
 825         if (!req->rq_reply_bytes_recvd) {
 826                 if (xprt->ops->timer)
 827                         xprt->ops->timer(task);
 828         } else
 829                 task->tk_status = 0;
 830         spin_unlock_bh(&xprt->transport_lock);
 831 }
 832
 833 static inline int xprt_has_timer(struct rpc_xprt *xprt)
 834 {
 835         return xprt->idle_timeout != 0;
 836 }
 837
 838 /**
 839  * xprt_prepare_transmit - reserve the transport before sending a request
 840  * @task: RPC task about to send a request
 841  *
 842  */
 843 int xprt_prepare_transmit(struct rpc_task *task)
 844 {
 845         struct rpc_rqst *req = task->tk_rqstp;
 846         struct rpc_xprt *xprt = req->rq_xprt;
 847         int err = 0;
 848
 849         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
 850
 851         spin_lock_bh(&xprt->transport_lock);
 852         if (req->rq_reply_bytes_recvd && !req->rq_bytes_sent) {
 853                 err = req->rq_reply_bytes_recvd;
 854                 goto out_unlock;
 855         }
 856         if (!xprt->ops->reserve_xprt(task))
 857                 err = -EAGAIN;
 858 out_unlock:
 859         spin_unlock_bh(&xprt->transport_lock);
 860         return err;
 861 }
 862
 863 void xprt_end_transmit(struct rpc_task *task)
 864 {
 865         xprt_release_write(task->tk_rqstp->rq_xprt, task);
 866 }
 867
 868 /**
 869  * xprt_transmit - send an RPC request on a transport
 870  * @task: controlling RPC task
 871  *
 872  * We have to copy the iovec because sendmsg fiddles with its contents.
 873  */
 874 void xprt_transmit(struct rpc_task *task)
 875 {
 876         struct rpc_rqst *req = task->tk_rqstp;
 877         struct rpc_xprt *xprt = req->rq_xprt;
 878         int status;
 879
 880         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
 881
 882         if (!req->rq_reply_bytes_recvd) {
 883                 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) {
 884                         /*
 885                          * Add to the list only if we're expecting a reply
 886                          */
 887                         spin_lock_bh(&xprt->transport_lock);
 888                         /* Update the softirq receive buffer */
 889                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
 890                                         sizeof(req->rq_private_buf));
 891                         /* Add request to the receive list */
 892                         list_add_tail(&req->rq_list, &xprt->recv);
 893                         spin_unlock_bh(&xprt->transport_lock);
 894                         xprt_reset_majortimeo(req);
 895                         /* Turn off autodisconnect */
 896                         del_singleshot_timer_sync(&xprt->timer);
 897                 }
 898         } else if (!req->rq_bytes_sent)
 899                 return;
 900
 901         req->rq_connect_cookie = xprt->connect_cookie;
 902         req->rq_xtime = ktime_get();
 903         status = xprt->ops->send_request(task);
 904         if (status != 0) {
 905                 task->tk_status = status;
 906                 return;
 907         }
 908
 909         dprintk("RPC: %5u xmit complete\n", task->tk_pid);
 910         spin_lock_bh(&xprt->transport_lock);
 911
 912         xprt->ops->set_retrans_timeout(task);
 913
 914         xprt->stat.sends++;
 915         xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
 916         xprt->stat.bklog_u += xprt->backlog.qlen;
 917
 918         /* Don't race with disconnect */
 919         if (!xprt_connected(xprt))
 920                 task->tk_status = -ENOTCONN;
 921         else if (!req->rq_reply_bytes_recvd && rpc_reply_expected(task)) {
 922                 /*
 923                  * Sleep on the pending queue since
 924                  * we're expecting a reply.
 925                  */
 926                 rpc_sleep_on(&xprt->pending, task, xprt_timer);
 927         }
 928         spin_unlock_bh(&xprt->transport_lock);
 929 }
 930
 931 static void xprt_alloc_slot(struct rpc_task *task)
 932 {
 933         struct rpc_xprt *xprt = task->tk_xprt;
 934
 935         task->tk_status = 0;
 936         if (task->tk_rqstp)
 937                 return;
 938         if (!list_empty(&xprt->free)) {
 939                 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
 940                 list_del_init(&req->rq_list);
 941                 task->tk_rqstp = req;
 942                 xprt_request_init(task, xprt);
 943                 return;
 944         }
 945         dprintk("RPC:       waiting for request slot\n");
 946         task->tk_status = -EAGAIN;
 947         task->tk_timeout = 0;
 948         rpc_sleep_on(&xprt->backlog, task, NULL);
 949 }
 950
 951 static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
 952 {
 953         memset(req, 0, sizeof(*req));   /* mark unused */
 954
 955         spin_lock(&xprt->reserve_lock);
 956         list_add(&req->rq_list, &xprt->free);
 957         rpc_wake_up_next(&xprt->backlog);
 958         spin_unlock(&xprt->reserve_lock);
 959 }
 960
 961 struct rpc_xprt *xprt_alloc(struct net *net, int size, int max_req)
 962 {
 963         struct rpc_xprt *xprt;
 964
 965         xprt = kzalloc(size, GFP_KERNEL);
 966         if (xprt == NULL)
 967                 goto out;
 968         kref_init(&xprt->kref);
 969
 970         xprt->max_reqs = max_req;
 971         xprt->slot = kcalloc(max_req, sizeof(struct rpc_rqst), GFP_KERNEL);
 972         if (xprt->slot == NULL)
 973                 goto out_free;
 974
 975         xprt->xprt_net = get_net(net);
 976         return xprt;
 977
 978 out_free:
 979         kfree(xprt);
 980 out:
 981         return NULL;
 982 }
 983 EXPORT_SYMBOL_GPL(xprt_alloc);
 984
 985 void xprt_free(struct rpc_xprt *xprt)
 986 {
 987         put_net(xprt->xprt_net);
 988         kfree(xprt->slot);
 989         kfree(xprt);
 990 }
 991 EXPORT_SYMBOL_GPL(xprt_free);
 992
 993 /**
 994  * xprt_reserve - allocate an RPC request slot
 995  * @task: RPC task requesting a slot allocation
 996  *
 997  * If no more slots are available, place the task on the transport's
 998  * backlog queue.
 999  */
1000 void xprt_reserve(struct rpc_task *task)
1001 {
1002         struct rpc_xprt *xprt = task->tk_xprt;
1003
1004         task->tk_status = -EIO;
1005         spin_lock(&xprt->reserve_lock);
1006         xprt_alloc_slot(task);
1007         spin_unlock(&xprt->reserve_lock);
1008 }
1009
1010 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
1011 {
1012         return (__force __be32)xprt->xid++;
1013 }
1014
1015 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1016 {
1017         xprt->xid = net_random();
1018 }
1019
1020 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1021 {
1022         struct rpc_rqst *req = task->tk_rqstp;
1023
1024         req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1025         req->rq_task    = task;
1026         req->rq_xprt    = xprt;
1027         req->rq_buffer  = NULL;
1028         req->rq_xid     = xprt_alloc_xid(xprt);
1029         req->rq_release_snd_buf = NULL;
1030         xprt_reset_majortimeo(req);
1031         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1032                         req, ntohl(req->rq_xid));
1033 }
1034
1035 /**
1036  * xprt_release - release an RPC request slot
1037  * @task: task which is finished with the slot
1038  *
1039  */
1040 void xprt_release(struct rpc_task *task)
1041 {
1042         struct rpc_xprt *xprt;
1043         struct rpc_rqst *req;
1044
1045         if (!(req = task->tk_rqstp))
1046                 return;
1047
1048         xprt = req->rq_xprt;
1049         rpc_count_iostats(task);
1050         spin_lock_bh(&xprt->transport_lock);
1051         xprt->ops->release_xprt(xprt, task);
1052         if (xprt->ops->release_request)
1053                 xprt->ops->release_request(task);
1054         if (!list_empty(&req->rq_list))
1055                 list_del(&req->rq_list);
1056         xprt->last_used = jiffies;
1057         if (list_empty(&xprt->recv) && xprt_has_timer(xprt))
1058                 mod_timer(&xprt->timer,
1059                                 xprt->last_used + xprt->idle_timeout);
1060         spin_unlock_bh(&xprt->transport_lock);
1061         if (req->rq_buffer)
1062                 xprt->ops->buf_free(req->rq_buffer);
1063         if (req->rq_cred != NULL)
1064                 put_rpccred(req->rq_cred);
1065         task->tk_rqstp = NULL;
1066         if (req->rq_release_snd_buf)
1067                 req->rq_release_snd_buf(req);
1068
1069         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1070         if (likely(!bc_prealloc(req)))
1071                 xprt_free_slot(xprt, req);
1072         else
1073                 xprt_free_bc_request(req);
1074 }
1075
1076 /**
1077  * xprt_create_transport - create an RPC transport
1078  * @args: rpc transport creation arguments
1079  *
1080  */
1081 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1082 {
1083         struct rpc_xprt *xprt;
1084         struct rpc_rqst *req;
1085         struct xprt_class *t;
1086
1087         spin_lock(&xprt_list_lock);
1088         list_for_each_entry(t, &xprt_list, list) {
1089                 if (t->ident == args->ident) {
1090                         spin_unlock(&xprt_list_lock);
1091                         goto found;
1092                 }
1093         }
1094         spin_unlock(&xprt_list_lock);
1095         printk(KERN_ERR "RPC: transport (%d) not supported\n", args->ident);
1096         return ERR_PTR(-EIO);
1097
1098 found:
1099         xprt = t->setup(args);
1100         if (IS_ERR(xprt)) {
1101                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1102                                 -PTR_ERR(xprt));
1103                 return xprt;
1104         }
1105         if (test_and_set_bit(XPRT_INITIALIZED, &xprt->state))
1106                 /* ->setup returned a pre-initialized xprt: */
1107                 return xprt;
1108
1109         spin_lock_init(&xprt->transport_lock);
1110         spin_lock_init(&xprt->reserve_lock);
1111
1112         INIT_LIST_HEAD(&xprt->free);
1113         INIT_LIST_HEAD(&xprt->recv);
1114 #if defined(CONFIG_NFS_V4_1)
1115         spin_lock_init(&xprt->bc_pa_lock);
1116         INIT_LIST_HEAD(&xprt->bc_pa_list);
1117 #endif /* CONFIG_NFS_V4_1 */
1118
1119         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1120         if (xprt_has_timer(xprt))
1121                 setup_timer(&xprt->timer, xprt_init_autodisconnect,
1122                             (unsigned long)xprt);
1123         else
1124                 init_timer(&xprt->timer);
1125         xprt->last_used = jiffies;
1126         xprt->cwnd = RPC_INITCWND;
1127         xprt->bind_index = 0;
1128
1129         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1130         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1131         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
1132         rpc_init_wait_queue(&xprt->resend, "xprt_resend");
1133         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1134
1135         /* initialize free list */
1136         for (req = &xprt->slot[xprt->max_reqs-1]; req >= &xprt->slot[0]; req--)
1137                 list_add(&req->rq_list, &xprt->free);
1138
1139         xprt_init_xid(xprt);
1140
1141         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1142                         xprt->max_reqs);
1143         return xprt;
1144 }
1145
1146 /**
1147  * xprt_destroy - destroy an RPC transport, killing off all requests.
1148  * @kref: kref for the transport to destroy
1149  *
1150  */
1151 static void xprt_destroy(struct kref *kref)
1152 {
1153         struct rpc_xprt *xprt = container_of(kref, struct rpc_xprt, kref);
1154
1155         dprintk("RPC:       destroying transport %p\n", xprt);
1156         xprt->shutdown = 1;
1157         del_timer_sync(&xprt->timer);
1158
1159         rpc_destroy_wait_queue(&xprt->binding);
1160         rpc_destroy_wait_queue(&xprt->pending);
1161         rpc_destroy_wait_queue(&xprt->sending);
1162         rpc_destroy_wait_queue(&xprt->resend);
1163         rpc_destroy_wait_queue(&xprt->backlog);
1164         cancel_work_sync(&xprt->task_cleanup);
1165         /*
1166          * Tear down transport state and free the rpc_xprt
1167          */
1168         xprt->ops->destroy(xprt);
1169 }
1170
1171 /**
1172  * xprt_put - release a reference to an RPC transport.
1173  * @xprt: pointer to the transport
1174  *
1175  */
1176 void xprt_put(struct rpc_xprt *xprt)
1177 {
1178         kref_put(&xprt->kref, xprt_destroy);
1179 }
1180
1181 /**
1182  * xprt_get - return a reference to an RPC transport.
1183  * @xprt: pointer to the transport
1184  *
1185  */
1186 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1187 {
1188         kref_get(&xprt->kref);
1189         return xprt;
1190 }