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 void     xprt_request_init(struct rpc_task *, struct rpc_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(&xprt->pending, xprt->snd_task);
 521         }
 522         spin_unlock_bh(&xprt->transport_lock);
 523 }
 524 EXPORT_SYMBOL_GPL(xprt_write_space);
 525
 526 /**
 527  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
 528  * @task: task whose timeout is to be set
 529  *
 530  * Set a request's retransmit timeout based on the transport's
 531  * default timeout parameters.  Used by transports that don't adjust
 532  * the retransmit timeout based on round-trip time estimation.
 533  */
 534 void xprt_set_retrans_timeout_def(struct rpc_task *task)
 535 {
 536         task->tk_timeout = task->tk_rqstp->rq_timeout;
 537 }
 538 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
 539
 540 /**
 541  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
 542  * @task: task whose timeout is to be set
 543  *
 544  * Set a request's retransmit timeout using the RTT estimator.
 545  */
 546 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
 547 {
 548         int timer = task->tk_msg.rpc_proc->p_timer;
 549         struct rpc_clnt *clnt = task->tk_client;
 550         struct rpc_rtt *rtt = clnt->cl_rtt;
 551         struct rpc_rqst *req = task->tk_rqstp;
 552         unsigned long max_timeout = clnt->cl_timeout->to_maxval;
 553
 554         task->tk_timeout = rpc_calc_rto(rtt, timer);
 555         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
 556         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
 557                 task->tk_timeout = max_timeout;
 558 }
 559 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
 560
 561 static void xprt_reset_majortimeo(struct rpc_rqst *req)
 562 {
 563         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 564
 565         req->rq_majortimeo = req->rq_timeout;
 566         if (to->to_exponential)
 567                 req->rq_majortimeo <<= to->to_retries;
 568         else
 569                 req->rq_majortimeo += to->to_increment * to->to_retries;
 570         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
 571                 req->rq_majortimeo = to->to_maxval;
 572         req->rq_majortimeo += jiffies;
 573 }
 574
 575 /**
 576  * xprt_adjust_timeout - adjust timeout values for next retransmit
 577  * @req: RPC request containing parameters to use for the adjustment
 578  *
 579  */
 580 int xprt_adjust_timeout(struct rpc_rqst *req)
 581 {
 582         struct rpc_xprt *xprt = req->rq_xprt;
 583         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
 584         int status = 0;
 585
 586         if (time_before(jiffies, req->rq_majortimeo)) {
 587                 if (to->to_exponential)
 588                         req->rq_timeout <<= 1;
 589                 else
 590                         req->rq_timeout += to->to_increment;
 591                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
 592                         req->rq_timeout = to->to_maxval;
 593                 req->rq_retries++;
 594         } else {
 595                 req->rq_timeout = to->to_initval;
 596                 req->rq_retries = 0;
 597                 xprt_reset_majortimeo(req);
 598                 /* Reset the RTT counters == "slow start" */
 599                 spin_lock_bh(&xprt->transport_lock);
 600                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
 601                 spin_unlock_bh(&xprt->transport_lock);
 602                 status = -ETIMEDOUT;
 603         }
 604
 605         if (req->rq_timeout == 0) {
 606                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
 607                 req->rq_timeout = 5 * HZ;
 608         }
 609         return status;
 610 }
 611
 612 static void xprt_autoclose(struct work_struct *work)
 613 {
 614         struct rpc_xprt *xprt =
 615                 container_of(work, struct rpc_xprt, task_cleanup);
 616
 617         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
 618         xprt->ops->close(xprt);
 619         xprt_release_write(xprt, NULL);
 620         wake_up_bit(&xprt->state, XPRT_LOCKED);
 621 }
 622
 623 /**
 624  * xprt_disconnect_done - mark a transport as disconnected
 625  * @xprt: transport to flag for disconnect
 626  *
 627  */
 628 void xprt_disconnect_done(struct rpc_xprt *xprt)
 629 {
 630         dprintk("RPC:       disconnected transport %p\n", xprt);
 631         spin_lock_bh(&xprt->transport_lock);
 632         xprt_clear_connected(xprt);
 633         xprt_wake_pending_tasks(xprt, -EAGAIN);
 634         spin_unlock_bh(&xprt->transport_lock);
 635 }
 636 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
 637
 638 /**
 639  * xprt_force_disconnect - force a transport to disconnect
 640  * @xprt: transport to disconnect
 641  *
 642  */
 643 void xprt_force_disconnect(struct rpc_xprt *xprt)
 644 {
 645         /* Don't race with the test_bit() in xprt_clear_locked() */
 646         spin_lock_bh(&xprt->transport_lock);
 647         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 648         /* Try to schedule an autoclose RPC call */
 649         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 650                 queue_work(xprtiod_workqueue, &xprt->task_cleanup);
 651         xprt_wake_pending_tasks(xprt, -EAGAIN);
 652         spin_unlock_bh(&xprt->transport_lock);
 653 }
 654 EXPORT_SYMBOL_GPL(xprt_force_disconnect);
 655
 656 /**
 657  * xprt_conditional_disconnect - force a transport to disconnect
 658  * @xprt: transport to disconnect
 659  * @cookie: 'connection cookie'
 660  *
 661  * This attempts to break the connection if and only if 'cookie' matches
 662  * the current transport 'connection cookie'. It ensures that we don't
 663  * try to break the connection more than once when we need to retransmit
 664  * a batch of RPC requests.
 665  *
 666  */
 667 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
 668 {
 669         /* Don't race with the test_bit() in xprt_clear_locked() */
 670         spin_lock_bh(&xprt->transport_lock);
 671         if (cookie != xprt->connect_cookie)
 672                 goto out;
 673         if (test_bit(XPRT_CLOSING, &xprt->state))
 674                 goto out;
 675         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 676         /* Try to schedule an autoclose RPC call */
 677         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 678                 queue_work(xprtiod_workqueue, &xprt->task_cleanup);
 679         xprt_wake_pending_tasks(xprt, -EAGAIN);
 680 out:
 681         spin_unlock_bh(&xprt->transport_lock);
 682 }
 683
 684 static bool
 685 xprt_has_timer(const struct rpc_xprt *xprt)
 686 {
 687         return xprt->idle_timeout != 0;
 688 }
 689
 690 static void
 691 xprt_schedule_autodisconnect(struct rpc_xprt *xprt)
 692         __must_hold(&xprt->transport_lock)
 693 {
 694         if (list_empty(&xprt->recv) && xprt_has_timer(xprt))
 695                 mod_timer(&xprt->timer, xprt->last_used + xprt->idle_timeout);
 696 }
 697
 698 static void
 699 xprt_init_autodisconnect(struct timer_list *t)
 700 {
 701         struct rpc_xprt *xprt = from_timer(xprt, t, timer);
 702
 703         spin_lock(&xprt->transport_lock);
 704         if (!list_empty(&xprt->recv))
 705                 goto out_abort;
 706         /* Reset xprt->last_used to avoid connect/autodisconnect cycling */
 707         xprt->last_used = jiffies;
 708         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
 709                 goto out_abort;
 710         spin_unlock(&xprt->transport_lock);
 711         queue_work(xprtiod_workqueue, &xprt->task_cleanup);
 712         return;
 713 out_abort:
 714         spin_unlock(&xprt->transport_lock);
 715 }
 716
 717 bool xprt_lock_connect(struct rpc_xprt *xprt,
 718                 struct rpc_task *task,
 719                 void *cookie)
 720 {
 721         bool ret = false;
 722
 723         spin_lock_bh(&xprt->transport_lock);
 724         if (!test_bit(XPRT_LOCKED, &xprt->state))
 725                 goto out;
 726         if (xprt->snd_task != task)
 727                 goto out;
 728         xprt_task_clear_bytes_sent(task);
 729         xprt->snd_task = cookie;
 730         ret = true;
 731 out:
 732         spin_unlock_bh(&xprt->transport_lock);
 733         return ret;
 734 }
 735
 736 void xprt_unlock_connect(struct rpc_xprt *xprt, void *cookie)
 737 {
 738         spin_lock_bh(&xprt->transport_lock);
 739         if (xprt->snd_task != cookie)
 740                 goto out;
 741         if (!test_bit(XPRT_LOCKED, &xprt->state))
 742                 goto out;
 743         xprt->snd_task =NULL;
 744         xprt->ops->release_xprt(xprt, NULL);
 745         xprt_schedule_autodisconnect(xprt);
 746 out:
 747         spin_unlock_bh(&xprt->transport_lock);
 748         wake_up_bit(&xprt->state, XPRT_LOCKED);
 749 }
 750
 751 /**
 752  * xprt_connect - schedule a transport connect operation
 753  * @task: RPC task that is requesting the connect
 754  *
 755  */
 756 void xprt_connect(struct rpc_task *task)
 757 {
 758         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 759
 760         dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
 761                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
 762
 763         if (!xprt_bound(xprt)) {
 764                 task->tk_status = -EAGAIN;
 765                 return;
 766         }
 767         if (!xprt_lock_write(xprt, task))
 768                 return;
 769
 770         if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
 771                 xprt->ops->close(xprt);
 772
 773         if (!xprt_connected(xprt)) {
 774                 task->tk_rqstp->rq_bytes_sent = 0;
 775                 task->tk_timeout = task->tk_rqstp->rq_timeout;
 776                 task->tk_rqstp->rq_connect_cookie = xprt->connect_cookie;
 777                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
 778
 779                 if (test_bit(XPRT_CLOSING, &xprt->state))
 780                         return;
 781                 if (xprt_test_and_set_connecting(xprt))
 782                         return;
 783                 xprt->stat.connect_start = jiffies;
 784                 xprt->ops->connect(xprt, task);
 785         }
 786         xprt_release_write(xprt, task);
 787 }
 788
 789 static void xprt_connect_status(struct rpc_task *task)
 790 {
 791         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
 792
 793         if (task->tk_status == 0) {
 794                 xprt->stat.connect_count++;
 795                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
 796                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
 797                                 task->tk_pid);
 798                 return;
 799         }
 800
 801         switch (task->tk_status) {
 802         case -ECONNREFUSED:
 803         case -ECONNRESET:
 804         case -ECONNABORTED:
 805         case -ENETUNREACH:
 806         case -EHOSTUNREACH:
 807         case -EPIPE:
 808         case -EAGAIN:
 809                 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
 810                 break;
 811         case -ETIMEDOUT:
 812                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
 813                                 "out\n", task->tk_pid);
 814                 break;
 815         default:
 816                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
 817                                 "server %s\n", task->tk_pid, -task->tk_status,
 818                                 xprt->servername);
 819                 task->tk_status = -EIO;
 820         }
 821 }
 822
 823 /**
 824  * xprt_lookup_rqst - find an RPC request corresponding to an XID
 825  * @xprt: transport on which the original request was transmitted
 826  * @xid: RPC XID of incoming reply
 827  *
 828  */
 829 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
 830 {
 831         struct rpc_rqst *entry;
 832
 833         list_for_each_entry(entry, &xprt->recv, rq_list)
 834                 if (entry->rq_xid == xid) {
 835                         trace_xprt_lookup_rqst(xprt, xid, 0);
 836                         return entry;
 837                 }
 838
 839         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
 840                         ntohl(xid));
 841         trace_xprt_lookup_rqst(xprt, xid, -ENOENT);
 842         xprt->stat.bad_xids++;
 843         return NULL;
 844 }
 845 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
 846
 847 /**
 848  * xprt_pin_rqst - Pin a request on the transport receive list
 849  * @req: Request to pin
 850  *
 851  * Caller must ensure this is atomic with the call to xprt_lookup_rqst()
 852  * so should be holding the xprt transport lock.
 853  */
 854 void xprt_pin_rqst(struct rpc_rqst *req)
 855 {
 856         set_bit(RPC_TASK_MSG_RECV, &req->rq_task->tk_runstate);
 857 }
 858 EXPORT_SYMBOL_GPL(xprt_pin_rqst);
 859
 860 /**
 861  * xprt_unpin_rqst - Unpin a request on the transport receive list
 862  * @req: Request to pin
 863  *
 864  * Caller should be holding the xprt transport lock.
 865  */
 866 void xprt_unpin_rqst(struct rpc_rqst *req)
 867 {
 868         struct rpc_task *task = req->rq_task;
 869
 870         clear_bit(RPC_TASK_MSG_RECV, &task->tk_runstate);
 871         if (test_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate))
 872                 wake_up_bit(&task->tk_runstate, RPC_TASK_MSG_RECV);
 873 }
 874 EXPORT_SYMBOL_GPL(xprt_unpin_rqst);
 875
 876 static void xprt_wait_on_pinned_rqst(struct rpc_rqst *req)
 877 __must_hold(&req->rq_xprt->recv_lock)
 878 {
 879         struct rpc_task *task = req->rq_task;
 880
 881         if (task && test_bit(RPC_TASK_MSG_RECV, &task->tk_runstate)) {
 882                 spin_unlock(&req->rq_xprt->recv_lock);
 883                 set_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate);
 884                 wait_on_bit(&task->tk_runstate, RPC_TASK_MSG_RECV,
 885                                 TASK_UNINTERRUPTIBLE);
 886                 clear_bit(RPC_TASK_MSG_RECV_WAIT, &task->tk_runstate);
 887                 spin_lock(&req->rq_xprt->recv_lock);
 888         }
 889 }
 890
 891 static void xprt_update_rtt(struct rpc_task *task)
 892 {
 893         struct rpc_rqst *req = task->tk_rqstp;
 894         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
 895         unsigned int timer = task->tk_msg.rpc_proc->p_timer;
 896         long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
 897
 898         if (timer) {
 899                 if (req->rq_ntrans == 1)
 900                         rpc_update_rtt(rtt, timer, m);
 901                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
 902         }
 903 }
 904
 905 /**
 906  * xprt_complete_rqst - called when reply processing is complete
 907  * @task: RPC request that recently completed
 908  * @copied: actual number of bytes received from the transport
 909  *
 910  * Caller holds transport lock.
 911  */
 912 void xprt_complete_rqst(struct rpc_task *task, int copied)
 913 {
 914         struct rpc_rqst *req = task->tk_rqstp;
 915         struct rpc_xprt *xprt = req->rq_xprt;
 916
 917         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
 918                         task->tk_pid, ntohl(req->rq_xid), copied);
 919         trace_xprt_complete_rqst(xprt, req->rq_xid, copied);
 920
 921         xprt->stat.recvs++;
 922         req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime);
 923         if (xprt->ops->timer != NULL)
 924                 xprt_update_rtt(task);
 925
 926         list_del_init(&req->rq_list);
 927         req->rq_private_buf.len = copied;
 928         /* Ensure all writes are done before we update */
 929         /* req->rq_reply_bytes_recvd */
 930         smp_wmb();
 931         req->rq_reply_bytes_recvd = copied;
 932         rpc_wake_up_queued_task(&xprt->pending, task);
 933 }
 934 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
 935
 936 static void xprt_timer(struct rpc_task *task)
 937 {
 938         struct rpc_rqst *req = task->tk_rqstp;
 939         struct rpc_xprt *xprt = req->rq_xprt;
 940
 941         if (task->tk_status != -ETIMEDOUT)
 942                 return;
 943         dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
 944
 945         if (!req->rq_reply_bytes_recvd) {
 946                 if (xprt->ops->timer)
 947                         xprt->ops->timer(xprt, task);
 948         } else
 949                 task->tk_status = 0;
 950 }
 951
 952 /**
 953  * xprt_prepare_transmit - reserve the transport before sending a request
 954  * @task: RPC task about to send a request
 955  *
 956  */
 957 bool xprt_prepare_transmit(struct rpc_task *task)
 958 {
 959         struct rpc_rqst *req = task->tk_rqstp;
 960         struct rpc_xprt *xprt = req->rq_xprt;
 961         bool ret = false;
 962
 963         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
 964
 965         spin_lock_bh(&xprt->transport_lock);
 966         if (!req->rq_bytes_sent) {
 967                 if (req->rq_reply_bytes_recvd) {
 968                         task->tk_status = req->rq_reply_bytes_recvd;
 969                         goto out_unlock;
 970                 }
 971                 if ((task->tk_flags & RPC_TASK_NO_RETRANS_TIMEOUT)
 972                     && xprt_connected(xprt)
 973                     && req->rq_connect_cookie == xprt->connect_cookie) {
 974                         xprt->ops->set_retrans_timeout(task);
 975                         rpc_sleep_on(&xprt->pending, task, xprt_timer);
 976                         goto out_unlock;
 977                 }
 978         }
 979         if (!xprt->ops->reserve_xprt(xprt, task)) {
 980                 task->tk_status = -EAGAIN;
 981                 goto out_unlock;
 982         }
 983         ret = true;
 984 out_unlock:
 985         spin_unlock_bh(&xprt->transport_lock);
 986         return ret;
 987 }
 988
 989 void xprt_end_transmit(struct rpc_task *task)
 990 {
 991         xprt_release_write(task->tk_rqstp->rq_xprt, task);
 992 }
 993
 994 /**
 995  * xprt_transmit - send an RPC request on a transport
 996  * @task: controlling RPC task
 997  *
 998  * We have to copy the iovec because sendmsg fiddles with its contents.
 999  */
1000 void xprt_transmit(struct rpc_task *task)
1001 {
1002         struct rpc_rqst *req = task->tk_rqstp;
1003         struct rpc_xprt *xprt = req->rq_xprt;
1004         unsigned int connect_cookie;
1005         int status, numreqs;
1006
1007         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
1008
1009         if (!req->rq_reply_bytes_recvd) {
1010                 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) {
1011                         /*
1012                          * Add to the list only if we're expecting a reply
1013                          */
1014                         /* Update the softirq receive buffer */
1015                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
1016                                         sizeof(req->rq_private_buf));
1017                         /* Add request to the receive list */
1018                         spin_lock(&xprt->recv_lock);
1019                         list_add_tail(&req->rq_list, &xprt->recv);
1020                         spin_unlock(&xprt->recv_lock);
1021                         xprt_reset_majortimeo(req);
1022                         /* Turn off autodisconnect */
1023                         del_singleshot_timer_sync(&xprt->timer);
1024                 }
1025         } else if (!req->rq_bytes_sent)
1026                 return;
1027
1028         connect_cookie = xprt->connect_cookie;
1029         req->rq_xtime = ktime_get();
1030         status = xprt->ops->send_request(task);
1031         trace_xprt_transmit(xprt, req->rq_xid, status);
1032         if (status != 0) {
1033                 task->tk_status = status;
1034                 return;
1035         }
1036         xprt_inject_disconnect(xprt);
1037
1038         dprintk("RPC: %5u xmit complete\n", task->tk_pid);
1039         task->tk_flags |= RPC_TASK_SENT;
1040         spin_lock_bh(&xprt->transport_lock);
1041
1042         xprt->ops->set_retrans_timeout(task);
1043
1044         numreqs = atomic_read(&xprt->num_reqs);
1045         if (numreqs > xprt->stat.max_slots)
1046                 xprt->stat.max_slots = numreqs;
1047         xprt->stat.sends++;
1048         xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
1049         xprt->stat.bklog_u += xprt->backlog.qlen;
1050         xprt->stat.sending_u += xprt->sending.qlen;
1051         xprt->stat.pending_u += xprt->pending.qlen;
1052         spin_unlock_bh(&xprt->transport_lock);
1053
1054         req->rq_connect_cookie = connect_cookie;
1055         if (rpc_reply_expected(task) && !READ_ONCE(req->rq_reply_bytes_recvd)) {
1056                 /*
1057                  * Sleep on the pending queue if we're expecting a reply.
1058                  * The spinlock ensures atomicity between the test of
1059                  * req->rq_reply_bytes_recvd, and the call to rpc_sleep_on().
1060                  */
1061                 spin_lock(&xprt->recv_lock);
1062                 if (!req->rq_reply_bytes_recvd) {
1063                         rpc_sleep_on(&xprt->pending, task, xprt_timer);
1064                         /*
1065                          * Send an extra queue wakeup call if the
1066                          * connection was dropped in case the call to
1067                          * rpc_sleep_on() raced.
1068                          */
1069                         if (!xprt_connected(xprt))
1070                                 xprt_wake_pending_tasks(xprt, -ENOTCONN);
1071                 }
1072                 spin_unlock(&xprt->recv_lock);
1073         }
1074 }
1075
1076 static void xprt_add_backlog(struct rpc_xprt *xprt, struct rpc_task *task)
1077 {
1078         set_bit(XPRT_CONGESTED, &xprt->state);
1079         rpc_sleep_on(&xprt->backlog, task, NULL);
1080 }
1081
1082 static void xprt_wake_up_backlog(struct rpc_xprt *xprt)
1083 {
1084         if (rpc_wake_up_next(&xprt->backlog) == NULL)
1085                 clear_bit(XPRT_CONGESTED, &xprt->state);
1086 }
1087
1088 static bool xprt_throttle_congested(struct rpc_xprt *xprt, struct rpc_task *task)
1089 {
1090         bool ret = false;
1091
1092         if (!test_bit(XPRT_CONGESTED, &xprt->state))
1093                 goto out;
1094         spin_lock(&xprt->reserve_lock);
1095         if (test_bit(XPRT_CONGESTED, &xprt->state)) {
1096                 rpc_sleep_on(&xprt->backlog, task, NULL);
1097                 ret = true;
1098         }
1099         spin_unlock(&xprt->reserve_lock);
1100 out:
1101         return ret;
1102 }
1103
1104 static struct rpc_rqst *xprt_dynamic_alloc_slot(struct rpc_xprt *xprt)
1105 {
1106         struct rpc_rqst *req = ERR_PTR(-EAGAIN);
1107
1108         if (!atomic_add_unless(&xprt->num_reqs, 1, xprt->max_reqs))
1109                 goto out;
1110         spin_unlock(&xprt->reserve_lock);
1111         req = kzalloc(sizeof(struct rpc_rqst), GFP_NOFS);
1112         spin_lock(&xprt->reserve_lock);
1113         if (req != NULL)
1114                 goto out;
1115         atomic_dec(&xprt->num_reqs);
1116         req = ERR_PTR(-ENOMEM);
1117 out:
1118         return req;
1119 }
1120
1121 static bool xprt_dynamic_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1122 {
1123         if (atomic_add_unless(&xprt->num_reqs, -1, xprt->min_reqs)) {
1124                 kfree(req);
1125                 return true;
1126         }
1127         return false;
1128 }
1129
1130 void xprt_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1131 {
1132         struct rpc_rqst *req;
1133
1134         spin_lock(&xprt->reserve_lock);
1135         if (!list_empty(&xprt->free)) {
1136                 req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1137                 list_del(&req->rq_list);
1138                 goto out_init_req;
1139         }
1140         req = xprt_dynamic_alloc_slot(xprt);
1141         if (!IS_ERR(req))
1142                 goto out_init_req;
1143         switch (PTR_ERR(req)) {
1144         case -ENOMEM:
1145                 dprintk("RPC:       dynamic allocation of request slot "
1146                                 "failed! Retrying\n");
1147                 task->tk_status = -ENOMEM;
1148                 break;
1149         case -EAGAIN:
1150                 xprt_add_backlog(xprt, task);
1151                 dprintk("RPC:       waiting for request slot\n");
1152                 /* fall through */
1153         default:
1154                 task->tk_status = -EAGAIN;
1155         }
1156         spin_unlock(&xprt->reserve_lock);
1157         return;
1158 out_init_req:
1159         task->tk_status = 0;
1160         task->tk_rqstp = req;
1161         xprt_request_init(task, xprt);
1162         spin_unlock(&xprt->reserve_lock);
1163 }
1164 EXPORT_SYMBOL_GPL(xprt_alloc_slot);
1165
1166 void xprt_lock_and_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1167 {
1168         /* Note: grabbing the xprt_lock_write() ensures that we throttle
1169          * new slot allocation if the transport is congested (i.e. when
1170          * reconnecting a stream transport or when out of socket write
1171          * buffer space).
1172          */
1173         if (xprt_lock_write(xprt, task)) {
1174                 xprt_alloc_slot(xprt, task);
1175                 xprt_release_write(xprt, task);
1176         }
1177 }
1178 EXPORT_SYMBOL_GPL(xprt_lock_and_alloc_slot);
1179
1180 static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1181 {
1182         spin_lock(&xprt->reserve_lock);
1183         if (!xprt_dynamic_free_slot(xprt, req)) {
1184                 memset(req, 0, sizeof(*req));   /* mark unused */
1185                 list_add(&req->rq_list, &xprt->free);
1186         }
1187         xprt_wake_up_backlog(xprt);
1188         spin_unlock(&xprt->reserve_lock);
1189 }
1190
1191 static void xprt_free_all_slots(struct rpc_xprt *xprt)
1192 {
1193         struct rpc_rqst *req;
1194         while (!list_empty(&xprt->free)) {
1195                 req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list);
1196                 list_del(&req->rq_list);
1197                 kfree(req);
1198         }
1199 }
1200
1201 struct rpc_xprt *xprt_alloc(struct net *net, size_t size,
1202                 unsigned int num_prealloc,
1203                 unsigned int max_alloc)
1204 {
1205         struct rpc_xprt *xprt;
1206         struct rpc_rqst *req;
1207         int i;
1208
1209         xprt = kzalloc(size, GFP_KERNEL);
1210         if (xprt == NULL)
1211                 goto out;
1212
1213         xprt_init(xprt, net);
1214
1215         for (i = 0; i < num_prealloc; i++) {
1216                 req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL);
1217                 if (!req)
1218                         goto out_free;
1219                 list_add(&req->rq_list, &xprt->free);
1220         }
1221         if (max_alloc > num_prealloc)
1222                 xprt->max_reqs = max_alloc;
1223         else
1224                 xprt->max_reqs = num_prealloc;
1225         xprt->min_reqs = num_prealloc;
1226         atomic_set(&xprt->num_reqs, num_prealloc);
1227
1228         return xprt;
1229
1230 out_free:
1231         xprt_free(xprt);
1232 out:
1233         return NULL;
1234 }
1235 EXPORT_SYMBOL_GPL(xprt_alloc);
1236
1237 void xprt_free(struct rpc_xprt *xprt)
1238 {
1239         put_net(xprt->xprt_net);
1240         xprt_free_all_slots(xprt);
1241         kfree_rcu(xprt, rcu);
1242 }
1243 EXPORT_SYMBOL_GPL(xprt_free);
1244
1245 /**
1246  * xprt_reserve - allocate an RPC request slot
1247  * @task: RPC task requesting a slot allocation
1248  *
1249  * If the transport is marked as being congested, or if no more
1250  * slots are available, place the task on the transport's
1251  * backlog queue.
1252  */
1253 void xprt_reserve(struct rpc_task *task)
1254 {
1255         struct rpc_xprt *xprt = task->tk_xprt;
1256
1257         task->tk_status = 0;
1258         if (task->tk_rqstp != NULL)
1259                 return;
1260
1261         task->tk_timeout = 0;
1262         task->tk_status = -EAGAIN;
1263         if (!xprt_throttle_congested(xprt, task))
1264                 xprt->ops->alloc_slot(xprt, task);
1265 }
1266
1267 /**
1268  * xprt_retry_reserve - allocate an RPC request slot
1269  * @task: RPC task requesting a slot allocation
1270  *
1271  * If no more slots are available, place the task on the transport's
1272  * backlog queue.
1273  * Note that the only difference with xprt_reserve is that we now
1274  * ignore the value of the XPRT_CONGESTED flag.
1275  */
1276 void xprt_retry_reserve(struct rpc_task *task)
1277 {
1278         struct rpc_xprt *xprt = task->tk_xprt;
1279
1280         task->tk_status = 0;
1281         if (task->tk_rqstp != NULL)
1282                 return;
1283
1284         task->tk_timeout = 0;
1285         task->tk_status = -EAGAIN;
1286         xprt->ops->alloc_slot(xprt, task);
1287 }
1288
1289 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
1290 {
1291         return (__force __be32)xprt->xid++;
1292 }
1293
1294 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1295 {
1296         xprt->xid = prandom_u32();
1297 }
1298
1299 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1300 {
1301         struct rpc_rqst *req = task->tk_rqstp;
1302
1303         INIT_LIST_HEAD(&req->rq_list);
1304         req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1305         req->rq_task    = task;
1306         req->rq_xprt    = xprt;
1307         req->rq_buffer  = NULL;
1308         req->rq_xid     = xprt_alloc_xid(xprt);
1309         req->rq_connect_cookie = xprt->connect_cookie - 1;
1310         req->rq_bytes_sent = 0;
1311         req->rq_snd_buf.len = 0;
1312         req->rq_snd_buf.buflen = 0;
1313         req->rq_rcv_buf.len = 0;
1314         req->rq_rcv_buf.buflen = 0;
1315         req->rq_release_snd_buf = NULL;
1316         xprt_reset_majortimeo(req);
1317         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1318                         req, ntohl(req->rq_xid));
1319 }
1320
1321 /**
1322  * xprt_release - release an RPC request slot
1323  * @task: task which is finished with the slot
1324  *
1325  */
1326 void xprt_release(struct rpc_task *task)
1327 {
1328         struct rpc_xprt *xprt;
1329         struct rpc_rqst *req = task->tk_rqstp;
1330
1331         if (req == NULL) {
1332                 if (task->tk_client) {
1333                         xprt = task->tk_xprt;
1334                         if (xprt->snd_task == task)
1335                                 xprt_release_write(xprt, task);
1336                 }
1337                 return;
1338         }
1339
1340         xprt = req->rq_xprt;
1341         if (task->tk_ops->rpc_count_stats != NULL)
1342                 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
1343         else if (task->tk_client)
1344                 rpc_count_iostats(task, task->tk_client->cl_metrics);
1345         spin_lock(&xprt->recv_lock);
1346         if (!list_empty(&req->rq_list)) {
1347                 list_del_init(&req->rq_list);
1348                 xprt_wait_on_pinned_rqst(req);
1349         }
1350         spin_unlock(&xprt->recv_lock);
1351         spin_lock_bh(&xprt->transport_lock);
1352         xprt->ops->release_xprt(xprt, task);
1353         if (xprt->ops->release_request)
1354                 xprt->ops->release_request(task);
1355         xprt->last_used = jiffies;
1356         xprt_schedule_autodisconnect(xprt);
1357         spin_unlock_bh(&xprt->transport_lock);
1358         if (req->rq_buffer)
1359                 xprt->ops->buf_free(task);
1360         xprt_inject_disconnect(xprt);
1361         if (req->rq_cred != NULL)
1362                 put_rpccred(req->rq_cred);
1363         task->tk_rqstp = NULL;
1364         if (req->rq_release_snd_buf)
1365                 req->rq_release_snd_buf(req);
1366
1367         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1368         if (likely(!bc_prealloc(req)))
1369                 xprt_free_slot(xprt, req);
1370         else
1371                 xprt_free_bc_request(req);
1372 }
1373
1374 static void xprt_init(struct rpc_xprt *xprt, struct net *net)
1375 {
1376         kref_init(&xprt->kref);
1377
1378         spin_lock_init(&xprt->transport_lock);
1379         spin_lock_init(&xprt->reserve_lock);
1380         spin_lock_init(&xprt->recv_lock);
1381
1382         INIT_LIST_HEAD(&xprt->free);
1383         INIT_LIST_HEAD(&xprt->recv);
1384 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1385         spin_lock_init(&xprt->bc_pa_lock);
1386         INIT_LIST_HEAD(&xprt->bc_pa_list);
1387 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1388         INIT_LIST_HEAD(&xprt->xprt_switch);
1389
1390         xprt->last_used = jiffies;
1391         xprt->cwnd = RPC_INITCWND;
1392         xprt->bind_index = 0;
1393
1394         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1395         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1396         rpc_init_priority_wait_queue(&xprt->sending, "xprt_sending");
1397         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1398
1399         xprt_init_xid(xprt);
1400
1401         xprt->xprt_net = get_net(net);
1402 }
1403
1404 /**
1405  * xprt_create_transport - create an RPC transport
1406  * @args: rpc transport creation arguments
1407  *
1408  */
1409 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1410 {
1411         struct rpc_xprt *xprt;
1412         struct xprt_class *t;
1413
1414         spin_lock(&xprt_list_lock);
1415         list_for_each_entry(t, &xprt_list, list) {
1416                 if (t->ident == args->ident) {
1417                         spin_unlock(&xprt_list_lock);
1418                         goto found;
1419                 }
1420         }
1421         spin_unlock(&xprt_list_lock);
1422         dprintk("RPC: transport (%d) not supported\n", args->ident);
1423         return ERR_PTR(-EIO);
1424
1425 found:
1426         xprt = t->setup(args);
1427         if (IS_ERR(xprt)) {
1428                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1429                                 -PTR_ERR(xprt));
1430                 goto out;
1431         }
1432         if (args->flags & XPRT_CREATE_NO_IDLE_TIMEOUT)
1433                 xprt->idle_timeout = 0;
1434         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1435         if (xprt_has_timer(xprt))
1436                 timer_setup(&xprt->timer, xprt_init_autodisconnect, 0);
1437         else
1438                 timer_setup(&xprt->timer, NULL, 0);
1439
1440         if (strlen(args->servername) > RPC_MAXNETNAMELEN) {
1441                 xprt_destroy(xprt);
1442                 return ERR_PTR(-EINVAL);
1443         }
1444         xprt->servername = kstrdup(args->servername, GFP_KERNEL);
1445         if (xprt->servername == NULL) {
1446                 xprt_destroy(xprt);
1447                 return ERR_PTR(-ENOMEM);
1448         }
1449
1450         rpc_xprt_debugfs_register(xprt);
1451
1452         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1453                         xprt->max_reqs);
1454 out:
1455         return xprt;
1456 }
1457
1458 static void xprt_destroy_cb(struct work_struct *work)
1459 {
1460         struct rpc_xprt *xprt =
1461                 container_of(work, struct rpc_xprt, task_cleanup);
1462
1463         rpc_xprt_debugfs_unregister(xprt);
1464         rpc_destroy_wait_queue(&xprt->binding);
1465         rpc_destroy_wait_queue(&xprt->pending);
1466         rpc_destroy_wait_queue(&xprt->sending);
1467         rpc_destroy_wait_queue(&xprt->backlog);
1468         kfree(xprt->servername);
1469         /*
1470          * Tear down transport state and free the rpc_xprt
1471          */
1472         xprt->ops->destroy(xprt);
1473 }
1474
1475 /**
1476  * xprt_destroy - destroy an RPC transport, killing off all requests.
1477  * @xprt: transport to destroy
1478  *
1479  */
1480 static void xprt_destroy(struct rpc_xprt *xprt)
1481 {
1482         dprintk("RPC:       destroying transport %p\n", xprt);
1483
1484         /*
1485          * Exclude transport connect/disconnect handlers and autoclose
1486          */
1487         wait_on_bit_lock(&xprt->state, XPRT_LOCKED, TASK_UNINTERRUPTIBLE);
1488
1489         del_timer_sync(&xprt->timer);
1490
1491         /*
1492          * Destroy sockets etc from the system workqueue so they can
1493          * safely flush receive work running on rpciod.
1494          */
1495         INIT_WORK(&xprt->task_cleanup, xprt_destroy_cb);
1496         schedule_work(&xprt->task_cleanup);
1497 }
1498
1499 static void xprt_destroy_kref(struct kref *kref)
1500 {
1501         xprt_destroy(container_of(kref, struct rpc_xprt, kref));
1502 }
1503
1504 /**
1505  * xprt_get - return a reference to an RPC transport.
1506  * @xprt: pointer to the transport
1507  *
1508  */
1509 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1510 {
1511         if (xprt != NULL && kref_get_unless_zero(&xprt->kref))
1512                 return xprt;
1513         return NULL;
1514 }
1515 EXPORT_SYMBOL_GPL(xprt_get);
1516
1517 /**
1518  * xprt_put - release a reference to an RPC transport.
1519  * @xprt: pointer to the transport
1520  *
1521  */
1522 void xprt_put(struct rpc_xprt *xprt)
1523 {
1524         if (xprt != NULL)
1525                 kref_put(&xprt->kref, xprt_destroy_kref);
1526 }
1527 EXPORT_SYMBOL_GPL(xprt_put);