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_xprt *xprt, struct rpc_task *task)
 195 {
 196         struct rpc_rqst *req = task->tk_rqstp;
 197
 198         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 199                 if (task == xprt->snd_task)
 200                         return 1;
 201                 goto out_sleep;
 202         }
 203         xprt->snd_task = task;
 204         if (req != NULL) {
 205                 req->rq_bytes_sent = 0;
 206                 req->rq_ntrans++;
 207         }
 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 != NULL && 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_xprt *xprt, struct rpc_task *task)
 244 {
 245         struct rpc_rqst *req = task->tk_rqstp;
 246
 247         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
 248                 if (task == xprt->snd_task)
 249                         return 1;
 250                 goto out_sleep;
 251         }
 252         if (req == NULL) {
 253                 xprt->snd_task = task;
 254                 return 1;
 255         }
 256         if (__xprt_get_cong(xprt, task)) {
 257                 xprt->snd_task = task;
 258                 req->rq_bytes_sent = 0;
 259                 req->rq_ntrans++;
 260                 return 1;
 261         }
 262         xprt_clear_locked(xprt);
 263 out_sleep:
 264         dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
 265         task->tk_timeout = 0;
 266         task->tk_status = -EAGAIN;
 267         if (req != NULL && req->rq_ntrans)
 268                 rpc_sleep_on(&xprt->resend, task, NULL);
 269         else
 270                 rpc_sleep_on(&xprt->sending, task, NULL);
 271         return 0;
 272 }
 273 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
 274
 275 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
 276 {
 277         int retval;
 278
 279         spin_lock_bh(&xprt->transport_lock);
 280         retval = xprt->ops->reserve_xprt(xprt, task);
 281         spin_unlock_bh(&xprt->transport_lock);
 282         return retval;
 283 }
 284
 285 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
 286 {
 287         struct rpc_task *task;
 288         struct rpc_rqst *req;
 289
 290         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 291                 return;
 292
 293         task = rpc_wake_up_next(&xprt->resend);
 294         if (!task) {
 295                 task = rpc_wake_up_next(&xprt->sending);
 296                 if (task == NULL)
 297                         goto out_unlock;
 298         }
 299
 300         req = task->tk_rqstp;
 301         xprt->snd_task = task;
 302         if (req) {
 303                 req->rq_bytes_sent = 0;
 304                 req->rq_ntrans++;
 305         }
 306         return;
 307
 308 out_unlock:
 309         xprt_clear_locked(xprt);
 310 }
 311
 312 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
 313 {
 314         struct rpc_task *task;
 315         struct rpc_rqst *req;
 316
 317         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 318                 return;
 319         if (RPCXPRT_CONGESTED(xprt))
 320                 goto out_unlock;
 321         task = rpc_wake_up_next(&xprt->resend);
 322         if (!task) {
 323                 task = rpc_wake_up_next(&xprt->sending);
 324                 if (task == NULL)
 325                         goto out_unlock;
 326         }
 327
 328         req = task->tk_rqstp;
 329         if (req == NULL) {
 330                 xprt->snd_task = task;
 331                 return;
 332         }
 333         if (__xprt_get_cong(xprt, task)) {
 334                 xprt->snd_task = task;
 335                 req->rq_bytes_sent = 0;
 336                 req->rq_ntrans++;
 337                 return;
 338         }
 339 out_unlock:
 340         xprt_clear_locked(xprt);
 341 }
 342
 343 /**
 344  * xprt_release_xprt - allow other requests to use a transport
 345  * @xprt: transport with other tasks potentially waiting
 346  * @task: task that is releasing access to the transport
 347  *
 348  * Note that "task" can be NULL.  No congestion control is provided.
 349  */
 350 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 351 {
 352         if (xprt->snd_task == task) {
 353                 xprt_clear_locked(xprt);
 354                 __xprt_lock_write_next(xprt);
 355         }
 356 }
 357 EXPORT_SYMBOL_GPL(xprt_release_xprt);
 358
 359 /**
 360  * xprt_release_xprt_cong - allow other requests to use a transport
 361  * @xprt: transport with other tasks potentially waiting
 362  * @task: task that is releasing access to the transport
 363  *
 364  * Note that "task" can be NULL.  Another task is awoken to use the
 365  * transport if the transport's congestion window allows it.
 366  */
 367 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 368 {
 369         if (xprt->snd_task == task) {
 370                 xprt_clear_locked(xprt);
 371                 __xprt_lock_write_next_cong(xprt);
 372         }
 373 }
 374 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
 375
 376 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
 377 {
 378         spin_lock_bh(&xprt->transport_lock);
 379         xprt->ops->release_xprt(xprt, task);
 380         spin_unlock_bh(&xprt->transport_lock);
 381 }
 382
 383 /*
 384  * Van Jacobson congestion avoidance. Check if the congestion window
 385  * overflowed. Put the task to sleep if this is the case.
 386  */
 387 static int
 388 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
 389 {
 390         struct rpc_rqst *req = task->tk_rqstp;
 391
 392         if (req->rq_cong)
 393                 return 1;
 394         dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
 395                         task->tk_pid, xprt->cong, xprt->cwnd);
 396         if (RPCXPRT_CONGESTED(xprt))
 397                 return 0;
 398         req->rq_cong = 1;
 399         xprt->cong += RPC_CWNDSCALE;
 400         return 1;
 401 }
 402
 403 /*
 404  * Adjust the congestion window, and wake up the next task
 405  * that has been sleeping due to congestion
 406  */
 407 static void
 408 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
 409 {
 410         if (!req->rq_cong)
 411                 return;
 412         req->rq_cong = 0;
 413         xprt->cong -= RPC_CWNDSCALE;
 414         __xprt_lock_write_next_cong(xprt);
 415 }
 416
 417 /**
 418  * xprt_release_rqst_cong - housekeeping when request is complete
 419  * @task: RPC request that recently completed
 420  *
 421  * Useful for transports that require congestion control.
 422  */
 423 void xprt_release_rqst_cong(struct rpc_task *task)
 424 {
 425         __xprt_put_cong(task->tk_xprt, task->tk_rqstp);
 426 }
 427 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
 428
 429 /**
 430  * xprt_adjust_cwnd - adjust transport congestion window
 431  * @task: recently completed RPC request used to adjust window
 432  * @result: result code of completed RPC request
 433  *
 434  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
 435  */
 436 void xprt_adjust_cwnd(struct rpc_task *task, int result)
 437 {
 438         struct rpc_rqst *req = task->tk_rqstp;
 439         struct rpc_xprt *xprt = task->tk_xprt;
 440         unsigned long cwnd = xprt->cwnd;
 441
 442         if (result >= 0 && cwnd <= xprt->cong) {
 443                 /* The (cwnd >> 1) term makes sure
 444                  * the result gets rounded properly. */
 445                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
 446                 if (cwnd > RPC_MAXCWND(xprt))
 447                         cwnd = RPC_MAXCWND(xprt);
 448                 __xprt_lock_write_next_cong(xprt);
 449         } else if (result == -ETIMEDOUT) {
 450                 cwnd >>= 1;
 451                 if (cwnd < RPC_CWNDSCALE)
 452                         cwnd = RPC_CWNDSCALE;
 453         }
 454         dprintk("RPC:       cong %ld, cwnd was %ld, now %ld\n",
 455                         xprt->cong, xprt->cwnd, cwnd);
 456         xprt->cwnd = cwnd;
 457         __xprt_put_cong(xprt, req);
 458 }
 459 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
 460
 461 /**
 462  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
 463  * @xprt: transport with waiting tasks
 464  * @status: result code to plant in each task before waking it
 465  *
 466  */
 467 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
 468 {
 469         if (status < 0)
 470                 rpc_wake_up_status(&xprt->pending, status);
 471         else
 472                 rpc_wake_up(&xprt->pending);
 473 }
 474 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
 475
 476 /**
 477  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
 478  * @task: task to be put to sleep
 479  * @action: function pointer to be executed after wait
 480  */
 481 void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action)
 482 {
 483         struct rpc_rqst *req = task->tk_rqstp;
 484         struct rpc_xprt *xprt = req->rq_xprt;
 485
 486         task->tk_timeout = req->rq_timeout;
 487         rpc_sleep_on(&xprt->pending, task, action);
 488 }
 489 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
 490
 491 /**
 492  * xprt_write_space - wake the task waiting for transport output buffer space
 493  * @xprt: transport with waiting tasks
 494  *
 495  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
 496  */
 497 void xprt_write_space(struct rpc_xprt *xprt)
 498 {
 499         if (unlikely(xprt->shutdown))
 500                 return;
 501
 502         spin_lock_bh(&xprt->transport_lock);
 503         if (xprt->snd_task) {
 504                 dprintk("RPC:       write space: waking waiting task on "
 505                                 "xprt %p\n", xprt);
 506                 rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task);
 507         }
 508         spin_unlock_bh(&xprt->transport_lock);
 509 }
 510 EXPORT_SYMBOL_GPL(xprt_write_space);
 511
 512 /**
 513  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
 514  * @task: task whose timeout is to be set
 515  *
 516  * Set a request's retransmit timeout based on the transport's
 517  * default timeout parameters.  Used by transports that don't adjust
 518  * the retransmit timeout based on round-trip time estimation.
 519  */
 520 void xprt_set_retrans_timeout_def(struct rpc_task *task)
 521 {
 522         task->tk_timeout = task->tk_rqstp->rq_timeout;
 523 }
 524 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
 525
 526 /*
 527  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
 528  * @task: task whose timeout is to be set
 529  *
 530  * Set a request's retransmit timeout using the RTT estimator.
 531  */
 532 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
 533 {
 534         int timer = task->tk_msg.rpc_proc->p_timer;
 535         struct rpc_clnt *clnt = task->tk_client;
 536         struct rpc_rtt *rtt = clnt->cl_rtt;
 537         struct rpc_rqst *req = task->tk_rqstp;
 538         unsigned long max_timeout = clnt->cl_timeout->to_maxval;
 539
 540         task->tk_timeout = rpc_calc_rto(rtt, timer);
 541         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
 542         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
 543                 task->tk_timeout = max_timeout;
 544 }
 545 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
 546
 547 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 548 {
 549         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 550
 551         req->rq_majortimeo = req->rq_timeout;
 552         if (to->to_exponential)
 553                 req->rq_majortimeo <<= to->to_retries;
 554         else
 555                 req->rq_majortimeo += to->to_increment * to->to_retries;
 556         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 557                 req->rq_majortimeo = to->to_maxval;
 558         req->rq_majortimeo += jiffies;
 559 }
 560
 561 /**
 562  * xprt_adjust_timeout - adjust timeout values for next retransmit
 563  * @req: RPC request containing parameters to use for the adjustment
 564  *
 565  */
 566 int xprt_adjust_timeout(struct rpc_rqst *req)
 567 {
 568         struct rpc_xprt *xprt = req->rq_xprt;
 569         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 570         int status = 0;
 571
 572         if (time_before(jiffies, req->rq_majortimeo)) {
 573                 if (to->to_exponential)
 574                         req->rq_timeout <<= 1;
 575                 else
 576                         req->rq_timeout += to->to_increment;
 577                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 578                         req->rq_timeout = to->to_maxval;
 579                 req->rq_retries++;
 580         } else {
 581                 req->rq_timeout = to->to_initval;
 582                 req->rq_retries = 0;
 583                 xprt_reset_majortimeo(req);
 584                 /* Reset the RTT counters == "slow start" */
 585                 spin_lock_bh(&xprt->transport_lock);
 586                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 587                 spin_unlock_bh(&xprt->transport_lock);
 588                 status = -ETIMEDOUT;
 589         }
 590
 591         if (req->rq_timeout == 0) {
 592                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 593                 req->rq_timeout = 5 * HZ;
 594         }
 595         return status;
 596 }
 597
 598 static void xprt_autoclose(struct work_struct *work)
 599 {
 600         struct rpc_xprt *xprt =
 601                 container_of(work, struct rpc_xprt, task_cleanup);
 602
 603         xprt->ops->close(xprt);
 604         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
 605         xprt_release_write(xprt, NULL);
 606 }
 607
 608 /**
 609  * xprt_disconnect_done - mark a transport as disconnected
 610  * @xprt: transport to flag for disconnect
 611  *
 612  */
 613 void xprt_disconnect_done(struct rpc_xprt *xprt)
 614 {
 615         dprintk("RPC:       disconnected transport %p\n", xprt);
 616         spin_lock_bh(&xprt->transport_lock);
 617         xprt_clear_connected(xprt);
 618         xprt_wake_pending_tasks(xprt, -EAGAIN);
 619         spin_unlock_bh(&xprt->transport_lock);
 620 }
 621 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
 622
 623 /**
 624  * xprt_force_disconnect - force a transport to disconnect
 625  * @xprt: transport to disconnect
 626  *
 627  */
 628 void xprt_force_disconnect(struct rpc_xprt *xprt)
 629 {
 630         /* Don't race with the test_bit() in xprt_clear_locked() */
 631         spin_lock_bh(&xprt->transport_lock);
 632         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 633         /* Try to schedule an autoclose RPC call */
 634         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 635                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 636         xprt_wake_pending_tasks(xprt, -EAGAIN);
 637         spin_unlock_bh(&xprt->transport_lock);
 638 }
 639
 640 /**
 641  * xprt_conditional_disconnect - force a transport to disconnect
 642  * @xprt: transport to disconnect
 643  * @cookie: 'connection cookie'
 644  *
 645  * This attempts to break the connection if and only if 'cookie' matches
 646  * the current transport 'connection cookie'. It ensures that we don't
 647  * try to break the connection more than once when we need to retransmit
 648  * a batch of RPC requests.
 649  *
 650  */
 651 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
 652 {
 653         /* Don't race with the test_bit() in xprt_clear_locked() */
 654         spin_lock_bh(&xprt->transport_lock);
 655         if (cookie != xprt->connect_cookie)
 656                 goto out;
 657         if (test_bit(XPRT_CLOSING, &xprt->state) || !xprt_connected(xprt))
 658                 goto out;
 659         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 660         /* Try to schedule an autoclose RPC call */
 661         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 662                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
 663         xprt_wake_pending_tasks(xprt, -EAGAIN);
 664 out:
 665         spin_unlock_bh(&xprt->transport_lock);
 666 }
 667
 668 static void
 669 xprt_init_autodisconnect(unsigned long data)
 670 {
 671         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
 672
 673         spin_lock(&xprt->transport_lock);
 674         if (!list_empty(&xprt->recv) || xprt->shutdown)
 675                 goto out_abort;
 676         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 677                 goto out_abort;
 678         spin_unlock(&xprt->transport_lock);
 679         set_bit(XPRT_CONNECTION_CLOSE, &xprt->state);
 680         queue_work(rpciod_workqueue, &xprt->task_cleanup);
 681         return;
 682 out_abort:
 683         spin_unlock(&xprt->transport_lock);
 684 }
 685
 686 /**
 687  * xprt_connect - schedule a transport connect operation
 688  * @task: RPC task that is requesting the connect
 689  *
 690  */
 691 void xprt_connect(struct rpc_task *task)
 692 {
 693         struct rpc_xprt *xprt = task->tk_xprt;
 694
 695         dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
 696                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 697
 698         if (!xprt_bound(xprt)) {
 699                 task->tk_status = -EAGAIN;
 700                 return;
 701         }
 702         if (!xprt_lock_write(xprt, task))
 703                 return;
 704
 705         if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
 706                 xprt->ops->close(xprt);
 707
 708         if (xprt_connected(xprt))
 709                 xprt_release_write(xprt, task);
 710         else {
 711                 if (task->tk_rqstp)
 712                         task->tk_rqstp->rq_bytes_sent = 0;
 713
 714                 task->tk_timeout = task->tk_rqstp->rq_timeout;
 715                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
 716
 717                 if (test_bit(XPRT_CLOSING, &xprt->state))
 718                         return;
 719                 if (xprt_test_and_set_connecting(xprt))
 720                         return;
 721                 xprt->stat.connect_start = jiffies;
 722                 xprt->ops->connect(task);
 723         }
 724 }
 725
 726 static void xprt_connect_status(struct rpc_task *task)
 727 {
 728         struct rpc_xprt *xprt = task->tk_xprt;
 729
 730         if (task->tk_status == 0) {
 731                 xprt->stat.connect_count++;
 732                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
 733                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
 734                                 task->tk_pid);
 735                 return;
 736         }
 737
 738         switch (task->tk_status) {
 739         case -EAGAIN:
 740                 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
 741                 break;
 742         case -ETIMEDOUT:
 743                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
 744                                 "out\n", task->tk_pid);
 745                 break;
 746         default:
 747                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
 748                                 "server %s\n", task->tk_pid, -task->tk_status,
 749                                 task->tk_client->cl_server);
 750                 xprt_release_write(xprt, task);
 751                 task->tk_status = -EIO;
 752         }
 753 }
 754
 755 /**
 756  * xprt_lookup_rqst - find an RPC request corresponding to an XID
 757  * @xprt: transport on which the original request was transmitted
 758  * @xid: RPC XID of incoming reply
 759  *
 760  */
 761 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
 762 {
 763         struct rpc_rqst *entry;
 764
 765         list_for_each_entry(entry, &xprt->recv, rq_list)
 766                 if (entry->rq_xid == xid)
 767                         return entry;
 768
 769         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
 770                         ntohl(xid));
 771         xprt->stat.bad_xids++;
 772         return NULL;
 773 }
 774 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
 775
 776 static void xprt_update_rtt(struct rpc_task *task)
 777 {
 778         struct rpc_rqst *req = task->tk_rqstp;
 779         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
 780         unsigned timer = task->tk_msg.rpc_proc->p_timer;
 781         long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
 782
 783         if (timer) {
 784                 if (req->rq_ntrans == 1)
 785                         rpc_update_rtt(rtt, timer, m);
 786                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
 787         }
 788 }
 789
 790 /**
 791  * xprt_complete_rqst - called when reply processing is complete
 792  * @task: RPC request that recently completed
 793  * @copied: actual number of bytes received from the transport
 794  *
 795  * Caller holds transport lock.
 796  */
 797 void xprt_complete_rqst(struct rpc_task *task, int copied)
 798 {
 799         struct rpc_rqst *req = task->tk_rqstp;
 800         struct rpc_xprt *xprt = req->rq_xprt;
 801
 802         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
 803                         task->tk_pid, ntohl(req->rq_xid), copied);
 804
 805         xprt->stat.recvs++;
 806         req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime);
 807         if (xprt->ops->timer != NULL)
 808                 xprt_update_rtt(task);
 809
 810         list_del_init(&req->rq_list);
 811         req->rq_private_buf.len = copied;
 812         /* Ensure all writes are done before we update */
 813         /* req->rq_reply_bytes_recvd */
 814         smp_wmb();
 815         req->rq_reply_bytes_recvd = copied;
 816         rpc_wake_up_queued_task(&xprt->pending, task);
 817 }
 818 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
 819
 820 static void xprt_timer(struct rpc_task *task)
 821 {
 822         struct rpc_rqst *req = task->tk_rqstp;
 823         struct rpc_xprt *xprt = req->rq_xprt;
 824
 825         if (task->tk_status != -ETIMEDOUT)
 826                 return;
 827         dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
 828
 829         spin_lock_bh(&xprt->transport_lock);
 830         if (!req->rq_reply_bytes_recvd) {
 831                 if (xprt->ops->timer)
 832                         xprt->ops->timer(task);
 833         } else
 834                 task->tk_status = 0;
 835         spin_unlock_bh(&xprt->transport_lock);
 836 }
 837
 838 static inline int xprt_has_timer(struct rpc_xprt *xprt)
 839 {
 840         return xprt->idle_timeout != 0;
 841 }
 842
 843 /**
 844  * xprt_prepare_transmit - reserve the transport before sending a request
 845  * @task: RPC task about to send a request
 846  *
 847  */
 848 int xprt_prepare_transmit(struct rpc_task *task)
 849 {
 850         struct rpc_rqst *req = task->tk_rqstp;
 851         struct rpc_xprt *xprt = req->rq_xprt;
 852         int err = 0;
 853
 854         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
 855
 856         spin_lock_bh(&xprt->transport_lock);
 857         if (req->rq_reply_bytes_recvd && !req->rq_bytes_sent) {
 858                 err = req->rq_reply_bytes_recvd;
 859                 goto out_unlock;
 860         }
 861         if (!xprt->ops->reserve_xprt(xprt, task))
 862                 err = -EAGAIN;
 863 out_unlock:
 864         spin_unlock_bh(&xprt->transport_lock);
 865         return err;
 866 }
 867
 868 void xprt_end_transmit(struct rpc_task *task)
 869 {
 870         xprt_release_write(task->tk_rqstp->rq_xprt, task);
 871 }
 872
 873 /**
 874  * xprt_transmit - send an RPC request on a transport
 875  * @task: controlling RPC task
 876  *
 877  * We have to copy the iovec because sendmsg fiddles with its contents.
 878  */
 879 void xprt_transmit(struct rpc_task *task)
 880 {
 881         struct rpc_rqst *req = task->tk_rqstp;
 882         struct rpc_xprt *xprt = req->rq_xprt;
 883         int status;
 884
 885         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
 886
 887         if (!req->rq_reply_bytes_recvd) {
 888                 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) {
 889                         /*
 890                          * Add to the list only if we're expecting a reply
 891                          */
 892                         spin_lock_bh(&xprt->transport_lock);
 893                         /* Update the softirq receive buffer */
 894                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
 895                                         sizeof(req->rq_private_buf));
 896                         /* Add request to the receive list */
 897                         list_add_tail(&req->rq_list, &xprt->recv);
 898                         spin_unlock_bh(&xprt->transport_lock);
 899                         xprt_reset_majortimeo(req);
 900                         /* Turn off autodisconnect */
 901                         del_singleshot_timer_sync(&xprt->timer);
 902                 }
 903         } else if (!req->rq_bytes_sent)
 904                 return;
 905
 906         req->rq_connect_cookie = xprt->connect_cookie;
 907         req->rq_xtime = ktime_get();
 908         status = xprt->ops->send_request(task);
 909         if (status != 0) {
 910                 task->tk_status = status;
 911                 return;
 912         }
 913
 914         dprintk("RPC: %5u xmit complete\n", task->tk_pid);
 915         task->tk_flags |= RPC_TASK_SENT;
 916         spin_lock_bh(&xprt->transport_lock);
 917
 918         xprt->ops->set_retrans_timeout(task);
 919
 920         xprt->stat.sends++;
 921         xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
 922         xprt->stat.bklog_u += xprt->backlog.qlen;
 923
 924         /* Don't race with disconnect */
 925         if (!xprt_connected(xprt))
 926                 task->tk_status = -ENOTCONN;
 927         else if (!req->rq_reply_bytes_recvd && rpc_reply_expected(task)) {
 928                 /*
 929                  * Sleep on the pending queue since
 930                  * we're expecting a reply.
 931                  */
 932                 rpc_sleep_on(&xprt->pending, task, xprt_timer);
 933         }
 934         spin_unlock_bh(&xprt->transport_lock);
 935 }
 936
 937 static void xprt_alloc_slot(struct rpc_task *task)
 938 {
 939         struct rpc_xprt *xprt = task->tk_xprt;
 940
 941         task->tk_status = 0;
 942         if (!list_empty(&xprt->free)) {
 943                 struct rpc_rqst *req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
 944                 list_del_init(&req->rq_list);
 945                 task->tk_rqstp = req;
 946                 xprt_request_init(task, xprt);
 947                 return;
 948         }
 949         dprintk("RPC:       waiting for request slot\n");
 950         task->tk_status = -EAGAIN;
 951         rpc_sleep_on(&xprt->backlog, task, NULL);
 952 }
 953
 954 static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
 955 {
 956         memset(req, 0, sizeof(*req));   /* mark unused */
 957
 958         spin_lock(&xprt->reserve_lock);
 959         list_add(&req->rq_list, &xprt->free);
 960         rpc_wake_up_next(&xprt->backlog);
 961         spin_unlock(&xprt->reserve_lock);
 962 }
 963
 964 struct rpc_xprt *xprt_alloc(struct net *net, int size, int max_req)
 965 {
 966         struct rpc_xprt *xprt;
 967
 968         xprt = kzalloc(size, GFP_KERNEL);
 969         if (xprt == NULL)
 970                 goto out;
 971         atomic_set(&xprt->count, 1);
 972
 973         xprt->max_reqs = max_req;
 974         xprt->slot = kcalloc(max_req, sizeof(struct rpc_rqst), GFP_KERNEL);
 975         if (xprt->slot == NULL)
 976                 goto out_free;
 977
 978         xprt->xprt_net = get_net(net);
 979         return xprt;
 980
 981 out_free:
 982         kfree(xprt);
 983 out:
 984         return NULL;
 985 }
 986 EXPORT_SYMBOL_GPL(xprt_alloc);
 987
 988 void xprt_free(struct rpc_xprt *xprt)
 989 {
 990         put_net(xprt->xprt_net);
 991         kfree(xprt->slot);
 992         kfree(xprt);
 993 }
 994 EXPORT_SYMBOL_GPL(xprt_free);
 995
 996 /**
 997  * xprt_reserve - allocate an RPC request slot
 998  * @task: RPC task requesting a slot allocation
 999  *
1000  * If no more slots are available, place the task on the transport's
1001  * backlog queue.
1002  */
1003 void xprt_reserve(struct rpc_task *task)
1004 {
1005         struct rpc_xprt *xprt = task->tk_xprt;
1006
1007         task->tk_status = 0;
1008         if (task->tk_rqstp != NULL)
1009                 return;
1010
1011         /* Note: grabbing the xprt_lock_write() here is not strictly needed,
1012          * but ensures that we throttle new slot allocation if the transport
1013          * is congested (e.g. if reconnecting or if we're out of socket
1014          * write buffer space).
1015          */
1016         task->tk_timeout = 0;
1017         task->tk_status = -EAGAIN;
1018         if (!xprt_lock_write(xprt, task))
1019                 return;
1020
1021         task->tk_status = -EIO;
1022         spin_lock(&xprt->reserve_lock);
1023         xprt_alloc_slot(task);
1024         spin_unlock(&xprt->reserve_lock);
1025         xprt_release_write(xprt, task);
1026 }
1027
1028 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
1029 {
1030         return (__force __be32)xprt->xid++;
1031 }
1032
1033 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1034 {
1035         xprt->xid = net_random();
1036 }
1037
1038 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1039 {
1040         struct rpc_rqst *req = task->tk_rqstp;
1041
1042         req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1043         req->rq_task    = task;
1044         req->rq_xprt    = xprt;
1045         req->rq_buffer  = NULL;
1046         req->rq_xid     = xprt_alloc_xid(xprt);
1047         req->rq_release_snd_buf = NULL;
1048         xprt_reset_majortimeo(req);
1049         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1050                         req, ntohl(req->rq_xid));
1051 }
1052
1053 /**
1054  * xprt_release - release an RPC request slot
1055  * @task: task which is finished with the slot
1056  *
1057  */
1058 void xprt_release(struct rpc_task *task)
1059 {
1060         struct rpc_xprt *xprt;
1061         struct rpc_rqst *req;
1062
1063         if (!(req = task->tk_rqstp))
1064                 return;
1065
1066         xprt = req->rq_xprt;
1067         rpc_count_iostats(task);
1068         spin_lock_bh(&xprt->transport_lock);
1069         xprt->ops->release_xprt(xprt, task);
1070         if (xprt->ops->release_request)
1071                 xprt->ops->release_request(task);
1072         if (!list_empty(&req->rq_list))
1073                 list_del(&req->rq_list);
1074         xprt->last_used = jiffies;
1075         if (list_empty(&xprt->recv) && xprt_has_timer(xprt))
1076                 mod_timer(&xprt->timer,
1077                                 xprt->last_used + xprt->idle_timeout);
1078         spin_unlock_bh(&xprt->transport_lock);
1079         if (req->rq_buffer)
1080                 xprt->ops->buf_free(req->rq_buffer);
1081         if (req->rq_cred != NULL)
1082                 put_rpccred(req->rq_cred);
1083         task->tk_rqstp = NULL;
1084         if (req->rq_release_snd_buf)
1085                 req->rq_release_snd_buf(req);
1086
1087         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1088         if (likely(!bc_prealloc(req)))
1089                 xprt_free_slot(xprt, req);
1090         else
1091                 xprt_free_bc_request(req);
1092 }
1093
1094 static void xprt_init(struct rpc_xprt *xprt)
1095 {
1096         struct rpc_rqst *req;
1097
1098         spin_lock_init(&xprt->transport_lock);
1099         spin_lock_init(&xprt->reserve_lock);
1100
1101         INIT_LIST_HEAD(&xprt->free);
1102         INIT_LIST_HEAD(&xprt->recv);
1103 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1104         spin_lock_init(&xprt->bc_pa_lock);
1105         INIT_LIST_HEAD(&xprt->bc_pa_list);
1106 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1107
1108         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1109         if (xprt_has_timer(xprt))
1110                 setup_timer(&xprt->timer, xprt_init_autodisconnect,
1111                             (unsigned long)xprt);
1112         else
1113                 init_timer(&xprt->timer);
1114         xprt->last_used = jiffies;
1115         xprt->cwnd = RPC_INITCWND;
1116         xprt->bind_index = 0;
1117
1118         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1119         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1120         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
1121         rpc_init_wait_queue(&xprt->resend, "xprt_resend");
1122         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1123
1124         /* initialize free list */
1125         for (req = &xprt->slot[xprt->max_reqs-1]; req >= &xprt->slot[0]; req--)
1126                 list_add(&req->rq_list, &xprt->free);
1127
1128         xprt_init_xid(xprt);
1129
1130 }
1131
1132 /**
1133  * xprt_create_transport - create an RPC transport
1134  * @args: rpc transport creation arguments
1135  *
1136  */
1137 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1138 {
1139         struct rpc_xprt *xprt;
1140         struct xprt_class *t;
1141
1142         spin_lock(&xprt_list_lock);
1143         list_for_each_entry(t, &xprt_list, list) {
1144                 if (t->ident == args->ident) {
1145                         spin_unlock(&xprt_list_lock);
1146                         goto found;
1147                 }
1148         }
1149         spin_unlock(&xprt_list_lock);
1150         printk(KERN_ERR "RPC: transport (%d) not supported\n", args->ident);
1151         return ERR_PTR(-EIO);
1152
1153 found:
1154         xprt = t->setup(args);
1155         if (IS_ERR(xprt)) {
1156                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1157                                 -PTR_ERR(xprt));
1158                 return xprt;
1159         }
1160         if (test_and_set_bit(XPRT_INITIALIZED, &xprt->state))
1161                 /* ->setup returned a pre-initialized xprt: */
1162                 return xprt;
1163
1164         xprt_init(xprt);
1165
1166         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1167                         xprt->max_reqs);
1168         return xprt;
1169 }
1170
1171 /**
1172  * xprt_destroy - destroy an RPC transport, killing off all requests.
1173  * @xprt: transport to destroy
1174  *
1175  */
1176 static void xprt_destroy(struct rpc_xprt *xprt)
1177 {
1178         dprintk("RPC:       destroying transport %p\n", xprt);
1179         xprt->shutdown = 1;
1180         del_timer_sync(&xprt->timer);
1181
1182         rpc_destroy_wait_queue(&xprt->binding);
1183         rpc_destroy_wait_queue(&xprt->pending);
1184         rpc_destroy_wait_queue(&xprt->sending);
1185         rpc_destroy_wait_queue(&xprt->resend);
1186         rpc_destroy_wait_queue(&xprt->backlog);
1187         cancel_work_sync(&xprt->task_cleanup);
1188         /*
1189          * Tear down transport state and free the rpc_xprt
1190          */
1191         xprt->ops->destroy(xprt);
1192 }
1193
1194 /**
1195  * xprt_put - release a reference to an RPC transport.
1196  * @xprt: pointer to the transport
1197  *
1198  */
1199 void xprt_put(struct rpc_xprt *xprt)
1200 {
1201         if (atomic_dec_and_test(&xprt->count))
1202                 xprt_destroy(xprt);
1203 }
1204
1205 /**
1206  * xprt_get - return a reference to an RPC transport.
1207  * @xprt: pointer to the transport
1208  *
1209  */
1210 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1211 {
1212         if (atomic_inc_not_zero(&xprt->count))
1213                 return xprt;
1214         return NULL;
1215 }