]> nv-tegra.nvidia Code Review - linux-3.10.git/blob - net/sunrpc/xprt.c
SUNRPC: constify rpc_program->name
[linux-3.10.git] / net / sunrpc / xprt.c
1 /*
2  *  linux/net/sunrpc/xprt.c
3  *
4  *  This is a generic RPC call interface supporting congestion avoidance,
5  *  and asynchronous calls.
6  *
7  *  The interface works like this:
8  *
9  *  -   When a process places a call, it allocates a request slot if
10  *      one is available. Otherwise, it sleeps on the backlog queue
11  *      (xprt_reserve).
12  *  -   Next, the caller puts together the RPC message, stuffs it into
13  *      the request struct, and calls xprt_transmit().
14  *  -   xprt_transmit sends the message and installs the caller on the
15  *      transport's wait list. At the same time, if a reply is expected,
16  *      it installs a timer that is run after the packet's timeout has
17  *      expired.
18  *  -   When a packet arrives, the data_ready handler walks the list of
19  *      pending requests for that transport. If a matching XID is found, the
20  *      caller is woken up, and the timer removed.
21  *  -   When no reply arrives within the timeout interval, the timer is
22  *      fired by the kernel and runs xprt_timer(). It either adjusts the
23  *      timeout values (minor timeout) or wakes up the caller with a status
24  *      of -ETIMEDOUT.
25  *  -   When the caller receives a notification from RPC that a reply arrived,
26  *      it should release the RPC slot, and process the reply.
27  *      If the call timed out, it may choose to retry the operation by
28  *      adjusting the initial timeout value, and simply calling rpc_call
29  *      again.
30  *
31  *  Support for async RPC is done through a set of RPC-specific scheduling
32  *  primitives that `transparently' work for processes as well as async
33  *  tasks that rely on callbacks.
34  *
35  *  Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de>
36  *
37  *  Transport switch API copyright (C) 2005, Chuck Lever <cel@netapp.com>
38  */
39
40 #include <linux/module.h>
41
42 #include <linux/types.h>
43 #include <linux/interrupt.h>
44 #include <linux/workqueue.h>
45 #include <linux/net.h>
46 #include <linux/ktime.h>
47
48 #include <linux/sunrpc/clnt.h>
49 #include <linux/sunrpc/metrics.h>
50 #include <linux/sunrpc/bc_xprt.h>
51
52 #include "sunrpc.h"
53
54 /*
55  * Local variables
56  */
57
58 #ifdef RPC_DEBUG
59 # define RPCDBG_FACILITY        RPCDBG_XPRT
60 #endif
61
62 /*
63  * Local functions
64  */
65 static void      xprt_init(struct rpc_xprt *xprt, struct net *net);
66 static void     xprt_request_init(struct rpc_task *, struct rpc_xprt *);
67 static void     xprt_connect_status(struct rpc_task *task);
68 static int      __xprt_get_cong(struct rpc_xprt *, struct rpc_task *);
69
70 static DEFINE_SPINLOCK(xprt_list_lock);
71 static LIST_HEAD(xprt_list);
72
73 /*
74  * The transport code maintains an estimate on the maximum number of out-
75  * standing RPC requests, using a smoothed version of the congestion
76  * avoidance implemented in 44BSD. This is basically the Van Jacobson
77  * congestion algorithm: If a retransmit occurs, the congestion window is
78  * halved; otherwise, it is incremented by 1/cwnd when
79  *
80  *      -       a reply is received and
81  *      -       a full number of requests are outstanding and
82  *      -       the congestion window hasn't been updated recently.
83  */
84 #define RPC_CWNDSHIFT           (8U)
85 #define RPC_CWNDSCALE           (1U << RPC_CWNDSHIFT)
86 #define RPC_INITCWND            RPC_CWNDSCALE
87 #define RPC_MAXCWND(xprt)       ((xprt)->max_reqs << RPC_CWNDSHIFT)
88
89 #define RPCXPRT_CONGESTED(xprt) ((xprt)->cong >= (xprt)->cwnd)
90
91 /**
92  * xprt_register_transport - register a transport implementation
93  * @transport: transport to register
94  *
95  * If a transport implementation is loaded as a kernel module, it can
96  * call this interface to make itself known to the RPC client.
97  *
98  * Returns:
99  * 0:           transport successfully registered
100  * -EEXIST:     transport already registered
101  * -EINVAL:     transport module being unloaded
102  */
103 int xprt_register_transport(struct xprt_class *transport)
104 {
105         struct xprt_class *t;
106         int result;
107
108         result = -EEXIST;
109         spin_lock(&xprt_list_lock);
110         list_for_each_entry(t, &xprt_list, list) {
111                 /* don't register the same transport class twice */
112                 if (t->ident == transport->ident)
113                         goto out;
114         }
115
116         list_add_tail(&transport->list, &xprt_list);
117         printk(KERN_INFO "RPC: Registered %s transport module.\n",
118                transport->name);
119         result = 0;
120
121 out:
122         spin_unlock(&xprt_list_lock);
123         return result;
124 }
125 EXPORT_SYMBOL_GPL(xprt_register_transport);
126
127 /**
128  * xprt_unregister_transport - unregister a transport implementation
129  * @transport: transport to unregister
130  *
131  * Returns:
132  * 0:           transport successfully unregistered
133  * -ENOENT:     transport never registered
134  */
135 int xprt_unregister_transport(struct xprt_class *transport)
136 {
137         struct xprt_class *t;
138         int result;
139
140         result = 0;
141         spin_lock(&xprt_list_lock);
142         list_for_each_entry(t, &xprt_list, list) {
143                 if (t == transport) {
144                         printk(KERN_INFO
145                                 "RPC: Unregistered %s transport module.\n",
146                                 transport->name);
147                         list_del_init(&transport->list);
148                         goto out;
149                 }
150         }
151         result = -ENOENT;
152
153 out:
154         spin_unlock(&xprt_list_lock);
155         return result;
156 }
157 EXPORT_SYMBOL_GPL(xprt_unregister_transport);
158
159 /**
160  * xprt_load_transport - load a transport implementation
161  * @transport_name: transport to load
162  *
163  * Returns:
164  * 0:           transport successfully loaded
165  * -ENOENT:     transport module not available
166  */
167 int xprt_load_transport(const char *transport_name)
168 {
169         struct xprt_class *t;
170         int result;
171
172         result = 0;
173         spin_lock(&xprt_list_lock);
174         list_for_each_entry(t, &xprt_list, list) {
175                 if (strcmp(t->name, transport_name) == 0) {
176                         spin_unlock(&xprt_list_lock);
177                         goto out;
178                 }
179         }
180         spin_unlock(&xprt_list_lock);
181         result = request_module("xprt%s", transport_name);
182 out:
183         return result;
184 }
185 EXPORT_SYMBOL_GPL(xprt_load_transport);
186
187 /**
188  * xprt_reserve_xprt - serialize write access to transports
189  * @task: task that is requesting access to the transport
190  * @xprt: pointer to the target transport
191  *
192  * This prevents mixing the payload of separate requests, and prevents
193  * transport connects from colliding with writes.  No congestion control
194  * is provided.
195  */
196 int xprt_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
197 {
198         struct rpc_rqst *req = task->tk_rqstp;
199         int priority;
200
201         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
202                 if (task == xprt->snd_task)
203                         return 1;
204                 goto out_sleep;
205         }
206         xprt->snd_task = task;
207         if (req != NULL) {
208                 req->rq_bytes_sent = 0;
209                 req->rq_ntrans++;
210         }
211
212         return 1;
213
214 out_sleep:
215         dprintk("RPC: %5u failed to lock transport %p\n",
216                         task->tk_pid, xprt);
217         task->tk_timeout = 0;
218         task->tk_status = -EAGAIN;
219         if (req == NULL)
220                 priority = RPC_PRIORITY_LOW;
221         else if (!req->rq_ntrans)
222                 priority = RPC_PRIORITY_NORMAL;
223         else
224                 priority = RPC_PRIORITY_HIGH;
225         rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
226         return 0;
227 }
228 EXPORT_SYMBOL_GPL(xprt_reserve_xprt);
229
230 static void xprt_clear_locked(struct rpc_xprt *xprt)
231 {
232         xprt->snd_task = NULL;
233         if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state) || xprt->shutdown) {
234                 smp_mb__before_clear_bit();
235                 clear_bit(XPRT_LOCKED, &xprt->state);
236                 smp_mb__after_clear_bit();
237         } else
238                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
239 }
240
241 /*
242  * xprt_reserve_xprt_cong - serialize write access to transports
243  * @task: task that is requesting access to the transport
244  *
245  * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
246  * integrated into the decision of whether a request is allowed to be
247  * woken up and given access to the transport.
248  */
249 int xprt_reserve_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
250 {
251         struct rpc_rqst *req = task->tk_rqstp;
252         int priority;
253
254         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
255                 if (task == xprt->snd_task)
256                         return 1;
257                 goto out_sleep;
258         }
259         if (req == NULL) {
260                 xprt->snd_task = task;
261                 return 1;
262         }
263         if (__xprt_get_cong(xprt, task)) {
264                 xprt->snd_task = task;
265                 req->rq_bytes_sent = 0;
266                 req->rq_ntrans++;
267                 return 1;
268         }
269         xprt_clear_locked(xprt);
270 out_sleep:
271         dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
272         task->tk_timeout = 0;
273         task->tk_status = -EAGAIN;
274         if (req == NULL)
275                 priority = RPC_PRIORITY_LOW;
276         else if (!req->rq_ntrans)
277                 priority = RPC_PRIORITY_NORMAL;
278         else
279                 priority = RPC_PRIORITY_HIGH;
280         rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
281         return 0;
282 }
283 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
284
285 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
286 {
287         int retval;
288
289         spin_lock_bh(&xprt->transport_lock);
290         retval = xprt->ops->reserve_xprt(xprt, task);
291         spin_unlock_bh(&xprt->transport_lock);
292         return retval;
293 }
294
295 static bool __xprt_lock_write_func(struct rpc_task *task, void *data)
296 {
297         struct rpc_xprt *xprt = data;
298         struct rpc_rqst *req;
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 true;
307 }
308
309 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
310 {
311         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
312                 return;
313
314         if (rpc_wake_up_first(&xprt->sending, __xprt_lock_write_func, xprt))
315                 return;
316         xprt_clear_locked(xprt);
317 }
318
319 static bool __xprt_lock_write_cong_func(struct rpc_task *task, void *data)
320 {
321         struct rpc_xprt *xprt = data;
322         struct rpc_rqst *req;
323
324         req = task->tk_rqstp;
325         if (req == NULL) {
326                 xprt->snd_task = task;
327                 return true;
328         }
329         if (__xprt_get_cong(xprt, task)) {
330                 xprt->snd_task = task;
331                 req->rq_bytes_sent = 0;
332                 req->rq_ntrans++;
333                 return true;
334         }
335         return false;
336 }
337
338 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
339 {
340         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
341                 return;
342         if (RPCXPRT_CONGESTED(xprt))
343                 goto out_unlock;
344         if (rpc_wake_up_first(&xprt->sending, __xprt_lock_write_cong_func, xprt))
345                 return;
346 out_unlock:
347         xprt_clear_locked(xprt);
348 }
349
350 /**
351  * xprt_release_xprt - allow other requests to use a transport
352  * @xprt: transport with other tasks potentially waiting
353  * @task: task that is releasing access to the transport
354  *
355  * Note that "task" can be NULL.  No congestion control is provided.
356  */
357 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
358 {
359         if (xprt->snd_task == task) {
360                 xprt_clear_locked(xprt);
361                 __xprt_lock_write_next(xprt);
362         }
363 }
364 EXPORT_SYMBOL_GPL(xprt_release_xprt);
365
366 /**
367  * xprt_release_xprt_cong - allow other requests to use a transport
368  * @xprt: transport with other tasks potentially waiting
369  * @task: task that is releasing access to the transport
370  *
371  * Note that "task" can be NULL.  Another task is awoken to use the
372  * transport if the transport's congestion window allows it.
373  */
374 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
375 {
376         if (xprt->snd_task == task) {
377                 xprt_clear_locked(xprt);
378                 __xprt_lock_write_next_cong(xprt);
379         }
380 }
381 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
382
383 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
384 {
385         spin_lock_bh(&xprt->transport_lock);
386         xprt->ops->release_xprt(xprt, task);
387         spin_unlock_bh(&xprt->transport_lock);
388 }
389
390 /*
391  * Van Jacobson congestion avoidance. Check if the congestion window
392  * overflowed. Put the task to sleep if this is the case.
393  */
394 static int
395 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
396 {
397         struct rpc_rqst *req = task->tk_rqstp;
398
399         if (req->rq_cong)
400                 return 1;
401         dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
402                         task->tk_pid, xprt->cong, xprt->cwnd);
403         if (RPCXPRT_CONGESTED(xprt))
404                 return 0;
405         req->rq_cong = 1;
406         xprt->cong += RPC_CWNDSCALE;
407         return 1;
408 }
409
410 /*
411  * Adjust the congestion window, and wake up the next task
412  * that has been sleeping due to congestion
413  */
414 static void
415 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
416 {
417         if (!req->rq_cong)
418                 return;
419         req->rq_cong = 0;
420         xprt->cong -= RPC_CWNDSCALE;
421         __xprt_lock_write_next_cong(xprt);
422 }
423
424 /**
425  * xprt_release_rqst_cong - housekeeping when request is complete
426  * @task: RPC request that recently completed
427  *
428  * Useful for transports that require congestion control.
429  */
430 void xprt_release_rqst_cong(struct rpc_task *task)
431 {
432         __xprt_put_cong(task->tk_xprt, task->tk_rqstp);
433 }
434 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
435
436 /**
437  * xprt_adjust_cwnd - adjust transport congestion window
438  * @task: recently completed RPC request used to adjust window
439  * @result: result code of completed RPC request
440  *
441  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
442  */
443 void xprt_adjust_cwnd(struct rpc_task *task, int result)
444 {
445         struct rpc_rqst *req = task->tk_rqstp;
446         struct rpc_xprt *xprt = task->tk_xprt;
447         unsigned long cwnd = xprt->cwnd;
448
449         if (result >= 0 && cwnd <= xprt->cong) {
450                 /* The (cwnd >> 1) term makes sure
451                  * the result gets rounded properly. */
452                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
453                 if (cwnd > RPC_MAXCWND(xprt))
454                         cwnd = RPC_MAXCWND(xprt);
455                 __xprt_lock_write_next_cong(xprt);
456         } else if (result == -ETIMEDOUT) {
457                 cwnd >>= 1;
458                 if (cwnd < RPC_CWNDSCALE)
459                         cwnd = RPC_CWNDSCALE;
460         }
461         dprintk("RPC:       cong %ld, cwnd was %ld, now %ld\n",
462                         xprt->cong, xprt->cwnd, cwnd);
463         xprt->cwnd = cwnd;
464         __xprt_put_cong(xprt, req);
465 }
466 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
467
468 /**
469  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
470  * @xprt: transport with waiting tasks
471  * @status: result code to plant in each task before waking it
472  *
473  */
474 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
475 {
476         if (status < 0)
477                 rpc_wake_up_status(&xprt->pending, status);
478         else
479                 rpc_wake_up(&xprt->pending);
480 }
481 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
482
483 /**
484  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
485  * @task: task to be put to sleep
486  * @action: function pointer to be executed after wait
487  */
488 void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action)
489 {
490         struct rpc_rqst *req = task->tk_rqstp;
491         struct rpc_xprt *xprt = req->rq_xprt;
492
493         task->tk_timeout = req->rq_timeout;
494         rpc_sleep_on(&xprt->pending, task, action);
495 }
496 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
497
498 /**
499  * xprt_write_space - wake the task waiting for transport output buffer space
500  * @xprt: transport with waiting tasks
501  *
502  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
503  */
504 void xprt_write_space(struct rpc_xprt *xprt)
505 {
506         if (unlikely(xprt->shutdown))
507                 return;
508
509         spin_lock_bh(&xprt->transport_lock);
510         if (xprt->snd_task) {
511                 dprintk("RPC:       write space: waking waiting task on "
512                                 "xprt %p\n", xprt);
513                 rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task);
514         }
515         spin_unlock_bh(&xprt->transport_lock);
516 }
517 EXPORT_SYMBOL_GPL(xprt_write_space);
518
519 /**
520  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
521  * @task: task whose timeout is to be set
522  *
523  * Set a request's retransmit timeout based on the transport's
524  * default timeout parameters.  Used by transports that don't adjust
525  * the retransmit timeout based on round-trip time estimation.
526  */
527 void xprt_set_retrans_timeout_def(struct rpc_task *task)
528 {
529         task->tk_timeout = task->tk_rqstp->rq_timeout;
530 }
531 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
532
533 /*
534  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
535  * @task: task whose timeout is to be set
536  *
537  * Set a request's retransmit timeout using the RTT estimator.
538  */
539 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
540 {
541         int timer = task->tk_msg.rpc_proc->p_timer;
542         struct rpc_clnt *clnt = task->tk_client;
543         struct rpc_rtt *rtt = clnt->cl_rtt;
544         struct rpc_rqst *req = task->tk_rqstp;
545         unsigned long max_timeout = clnt->cl_timeout->to_maxval;
546
547         task->tk_timeout = rpc_calc_rto(rtt, timer);
548         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
549         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
550                 task->tk_timeout = max_timeout;
551 }
552 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
553
554 static void xprt_reset_majortimeo(struct rpc_rqst *req)
555 {
556         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
557
558         req->rq_majortimeo = req->rq_timeout;
559         if (to->to_exponential)
560                 req->rq_majortimeo <<= to->to_retries;
561         else
562                 req->rq_majortimeo += to->to_increment * to->to_retries;
563         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
564                 req->rq_majortimeo = to->to_maxval;
565         req->rq_majortimeo += jiffies;
566 }
567
568 /**
569  * xprt_adjust_timeout - adjust timeout values for next retransmit
570  * @req: RPC request containing parameters to use for the adjustment
571  *
572  */
573 int xprt_adjust_timeout(struct rpc_rqst *req)
574 {
575         struct rpc_xprt *xprt = req->rq_xprt;
576         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
577         int status = 0;
578
579         if (time_before(jiffies, req->rq_majortimeo)) {
580                 if (to->to_exponential)
581                         req->rq_timeout <<= 1;
582                 else
583                         req->rq_timeout += to->to_increment;
584                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
585                         req->rq_timeout = to->to_maxval;
586                 req->rq_retries++;
587         } else {
588                 req->rq_timeout = to->to_initval;
589                 req->rq_retries = 0;
590                 xprt_reset_majortimeo(req);
591                 /* Reset the RTT counters == "slow start" */
592                 spin_lock_bh(&xprt->transport_lock);
593                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
594                 spin_unlock_bh(&xprt->transport_lock);
595                 status = -ETIMEDOUT;
596         }
597
598         if (req->rq_timeout == 0) {
599                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
600                 req->rq_timeout = 5 * HZ;
601         }
602         return status;
603 }
604
605 static void xprt_autoclose(struct work_struct *work)
606 {
607         struct rpc_xprt *xprt =
608                 container_of(work, struct rpc_xprt, task_cleanup);
609
610         xprt->ops->close(xprt);
611         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
612         xprt_release_write(xprt, NULL);
613 }
614
615 /**
616  * xprt_disconnect_done - mark a transport as disconnected
617  * @xprt: transport to flag for disconnect
618  *
619  */
620 void xprt_disconnect_done(struct rpc_xprt *xprt)
621 {
622         dprintk("RPC:       disconnected transport %p\n", xprt);
623         spin_lock_bh(&xprt->transport_lock);
624         xprt_clear_connected(xprt);
625         xprt_wake_pending_tasks(xprt, -EAGAIN);
626         spin_unlock_bh(&xprt->transport_lock);
627 }
628 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
629
630 /**
631  * xprt_force_disconnect - force a transport to disconnect
632  * @xprt: transport to disconnect
633  *
634  */
635 void xprt_force_disconnect(struct rpc_xprt *xprt)
636 {
637         /* Don't race with the test_bit() in xprt_clear_locked() */
638         spin_lock_bh(&xprt->transport_lock);
639         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
640         /* Try to schedule an autoclose RPC call */
641         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
642                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
643         xprt_wake_pending_tasks(xprt, -EAGAIN);
644         spin_unlock_bh(&xprt->transport_lock);
645 }
646
647 /**
648  * xprt_conditional_disconnect - force a transport to disconnect
649  * @xprt: transport to disconnect
650  * @cookie: 'connection cookie'
651  *
652  * This attempts to break the connection if and only if 'cookie' matches
653  * the current transport 'connection cookie'. It ensures that we don't
654  * try to break the connection more than once when we need to retransmit
655  * a batch of RPC requests.
656  *
657  */
658 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
659 {
660         /* Don't race with the test_bit() in xprt_clear_locked() */
661         spin_lock_bh(&xprt->transport_lock);
662         if (cookie != xprt->connect_cookie)
663                 goto out;
664         if (test_bit(XPRT_CLOSING, &xprt->state) || !xprt_connected(xprt))
665                 goto out;
666         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
667         /* Try to schedule an autoclose RPC call */
668         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
669                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
670         xprt_wake_pending_tasks(xprt, -EAGAIN);
671 out:
672         spin_unlock_bh(&xprt->transport_lock);
673 }
674
675 static void
676 xprt_init_autodisconnect(unsigned long data)
677 {
678         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
679
680         spin_lock(&xprt->transport_lock);
681         if (!list_empty(&xprt->recv) || xprt->shutdown)
682                 goto out_abort;
683         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
684                 goto out_abort;
685         spin_unlock(&xprt->transport_lock);
686         set_bit(XPRT_CONNECTION_CLOSE, &xprt->state);
687         queue_work(rpciod_workqueue, &xprt->task_cleanup);
688         return;
689 out_abort:
690         spin_unlock(&xprt->transport_lock);
691 }
692
693 /**
694  * xprt_connect - schedule a transport connect operation
695  * @task: RPC task that is requesting the connect
696  *
697  */
698 void xprt_connect(struct rpc_task *task)
699 {
700         struct rpc_xprt *xprt = task->tk_xprt;
701
702         dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
703                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
704
705         if (!xprt_bound(xprt)) {
706                 task->tk_status = -EAGAIN;
707                 return;
708         }
709         if (!xprt_lock_write(xprt, task))
710                 return;
711
712         if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
713                 xprt->ops->close(xprt);
714
715         if (xprt_connected(xprt))
716                 xprt_release_write(xprt, task);
717         else {
718                 if (task->tk_rqstp)
719                         task->tk_rqstp->rq_bytes_sent = 0;
720
721                 task->tk_timeout = task->tk_rqstp->rq_timeout;
722                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
723
724                 if (test_bit(XPRT_CLOSING, &xprt->state))
725                         return;
726                 if (xprt_test_and_set_connecting(xprt))
727                         return;
728                 xprt->stat.connect_start = jiffies;
729                 xprt->ops->connect(task);
730         }
731 }
732
733 static void xprt_connect_status(struct rpc_task *task)
734 {
735         struct rpc_xprt *xprt = task->tk_xprt;
736
737         if (task->tk_status == 0) {
738                 xprt->stat.connect_count++;
739                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
740                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
741                                 task->tk_pid);
742                 return;
743         }
744
745         switch (task->tk_status) {
746         case -EAGAIN:
747                 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
748                 break;
749         case -ETIMEDOUT:
750                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
751                                 "out\n", task->tk_pid);
752                 break;
753         default:
754                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
755                                 "server %s\n", task->tk_pid, -task->tk_status,
756                                 task->tk_client->cl_server);
757                 xprt_release_write(xprt, task);
758                 task->tk_status = -EIO;
759         }
760 }
761
762 /**
763  * xprt_lookup_rqst - find an RPC request corresponding to an XID
764  * @xprt: transport on which the original request was transmitted
765  * @xid: RPC XID of incoming reply
766  *
767  */
768 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
769 {
770         struct rpc_rqst *entry;
771
772         list_for_each_entry(entry, &xprt->recv, rq_list)
773                 if (entry->rq_xid == xid)
774                         return entry;
775
776         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
777                         ntohl(xid));
778         xprt->stat.bad_xids++;
779         return NULL;
780 }
781 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
782
783 static void xprt_update_rtt(struct rpc_task *task)
784 {
785         struct rpc_rqst *req = task->tk_rqstp;
786         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
787         unsigned timer = task->tk_msg.rpc_proc->p_timer;
788         long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
789
790         if (timer) {
791                 if (req->rq_ntrans == 1)
792                         rpc_update_rtt(rtt, timer, m);
793                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
794         }
795 }
796
797 /**
798  * xprt_complete_rqst - called when reply processing is complete
799  * @task: RPC request that recently completed
800  * @copied: actual number of bytes received from the transport
801  *
802  * Caller holds transport lock.
803  */
804 void xprt_complete_rqst(struct rpc_task *task, int copied)
805 {
806         struct rpc_rqst *req = task->tk_rqstp;
807         struct rpc_xprt *xprt = req->rq_xprt;
808
809         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
810                         task->tk_pid, ntohl(req->rq_xid), copied);
811
812         xprt->stat.recvs++;
813         req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime);
814         if (xprt->ops->timer != NULL)
815                 xprt_update_rtt(task);
816
817         list_del_init(&req->rq_list);
818         req->rq_private_buf.len = copied;
819         /* Ensure all writes are done before we update */
820         /* req->rq_reply_bytes_recvd */
821         smp_wmb();
822         req->rq_reply_bytes_recvd = copied;
823         rpc_wake_up_queued_task(&xprt->pending, task);
824 }
825 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
826
827 static void xprt_timer(struct rpc_task *task)
828 {
829         struct rpc_rqst *req = task->tk_rqstp;
830         struct rpc_xprt *xprt = req->rq_xprt;
831
832         if (task->tk_status != -ETIMEDOUT)
833                 return;
834         dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
835
836         spin_lock_bh(&xprt->transport_lock);
837         if (!req->rq_reply_bytes_recvd) {
838                 if (xprt->ops->timer)
839                         xprt->ops->timer(task);
840         } else
841                 task->tk_status = 0;
842         spin_unlock_bh(&xprt->transport_lock);
843 }
844
845 static inline int xprt_has_timer(struct rpc_xprt *xprt)
846 {
847         return xprt->idle_timeout != 0;
848 }
849
850 /**
851  * xprt_prepare_transmit - reserve the transport before sending a request
852  * @task: RPC task about to send a request
853  *
854  */
855 int xprt_prepare_transmit(struct rpc_task *task)
856 {
857         struct rpc_rqst *req = task->tk_rqstp;
858         struct rpc_xprt *xprt = req->rq_xprt;
859         int err = 0;
860
861         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
862
863         spin_lock_bh(&xprt->transport_lock);
864         if (req->rq_reply_bytes_recvd && !req->rq_bytes_sent) {
865                 err = req->rq_reply_bytes_recvd;
866                 goto out_unlock;
867         }
868         if (!xprt->ops->reserve_xprt(xprt, task))
869                 err = -EAGAIN;
870 out_unlock:
871         spin_unlock_bh(&xprt->transport_lock);
872         return err;
873 }
874
875 void xprt_end_transmit(struct rpc_task *task)
876 {
877         xprt_release_write(task->tk_rqstp->rq_xprt, task);
878 }
879
880 /**
881  * xprt_transmit - send an RPC request on a transport
882  * @task: controlling RPC task
883  *
884  * We have to copy the iovec because sendmsg fiddles with its contents.
885  */
886 void xprt_transmit(struct rpc_task *task)
887 {
888         struct rpc_rqst *req = task->tk_rqstp;
889         struct rpc_xprt *xprt = req->rq_xprt;
890         int status;
891
892         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
893
894         if (!req->rq_reply_bytes_recvd) {
895                 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) {
896                         /*
897                          * Add to the list only if we're expecting a reply
898                          */
899                         spin_lock_bh(&xprt->transport_lock);
900                         /* Update the softirq receive buffer */
901                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
902                                         sizeof(req->rq_private_buf));
903                         /* Add request to the receive list */
904                         list_add_tail(&req->rq_list, &xprt->recv);
905                         spin_unlock_bh(&xprt->transport_lock);
906                         xprt_reset_majortimeo(req);
907                         /* Turn off autodisconnect */
908                         del_singleshot_timer_sync(&xprt->timer);
909                 }
910         } else if (!req->rq_bytes_sent)
911                 return;
912
913         req->rq_connect_cookie = xprt->connect_cookie;
914         req->rq_xtime = ktime_get();
915         status = xprt->ops->send_request(task);
916         if (status != 0) {
917                 task->tk_status = status;
918                 return;
919         }
920
921         dprintk("RPC: %5u xmit complete\n", task->tk_pid);
922         task->tk_flags |= RPC_TASK_SENT;
923         spin_lock_bh(&xprt->transport_lock);
924
925         xprt->ops->set_retrans_timeout(task);
926
927         xprt->stat.sends++;
928         xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
929         xprt->stat.bklog_u += xprt->backlog.qlen;
930
931         /* Don't race with disconnect */
932         if (!xprt_connected(xprt))
933                 task->tk_status = -ENOTCONN;
934         else if (!req->rq_reply_bytes_recvd && rpc_reply_expected(task)) {
935                 /*
936                  * Sleep on the pending queue since
937                  * we're expecting a reply.
938                  */
939                 rpc_sleep_on(&xprt->pending, task, xprt_timer);
940         }
941         spin_unlock_bh(&xprt->transport_lock);
942 }
943
944 static struct rpc_rqst *xprt_dynamic_alloc_slot(struct rpc_xprt *xprt, gfp_t gfp_flags)
945 {
946         struct rpc_rqst *req = ERR_PTR(-EAGAIN);
947
948         if (!atomic_add_unless(&xprt->num_reqs, 1, xprt->max_reqs))
949                 goto out;
950         req = kzalloc(sizeof(struct rpc_rqst), gfp_flags);
951         if (req != NULL)
952                 goto out;
953         atomic_dec(&xprt->num_reqs);
954         req = ERR_PTR(-ENOMEM);
955 out:
956         return req;
957 }
958
959 static bool xprt_dynamic_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
960 {
961         if (atomic_add_unless(&xprt->num_reqs, -1, xprt->min_reqs)) {
962                 kfree(req);
963                 return true;
964         }
965         return false;
966 }
967
968 static void xprt_alloc_slot(struct rpc_task *task)
969 {
970         struct rpc_xprt *xprt = task->tk_xprt;
971         struct rpc_rqst *req;
972
973         if (!list_empty(&xprt->free)) {
974                 req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
975                 list_del(&req->rq_list);
976                 goto out_init_req;
977         }
978         req = xprt_dynamic_alloc_slot(xprt, GFP_NOWAIT);
979         if (!IS_ERR(req))
980                 goto out_init_req;
981         switch (PTR_ERR(req)) {
982         case -ENOMEM:
983                 rpc_delay(task, HZ >> 2);
984                 dprintk("RPC:       dynamic allocation of request slot "
985                                 "failed! Retrying\n");
986                 break;
987         case -EAGAIN:
988                 rpc_sleep_on(&xprt->backlog, task, NULL);
989                 dprintk("RPC:       waiting for request slot\n");
990         }
991         task->tk_status = -EAGAIN;
992         return;
993 out_init_req:
994         task->tk_status = 0;
995         task->tk_rqstp = req;
996         xprt_request_init(task, xprt);
997 }
998
999 static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1000 {
1001         spin_lock(&xprt->reserve_lock);
1002         if (!xprt_dynamic_free_slot(xprt, req)) {
1003                 memset(req, 0, sizeof(*req));   /* mark unused */
1004                 list_add(&req->rq_list, &xprt->free);
1005         }
1006         rpc_wake_up_next(&xprt->backlog);
1007         spin_unlock(&xprt->reserve_lock);
1008 }
1009
1010 static void xprt_free_all_slots(struct rpc_xprt *xprt)
1011 {
1012         struct rpc_rqst *req;
1013         while (!list_empty(&xprt->free)) {
1014                 req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list);
1015                 list_del(&req->rq_list);
1016                 kfree(req);
1017         }
1018 }
1019
1020 struct rpc_xprt *xprt_alloc(struct net *net, size_t size,
1021                 unsigned int num_prealloc,
1022                 unsigned int max_alloc)
1023 {
1024         struct rpc_xprt *xprt;
1025         struct rpc_rqst *req;
1026         int i;
1027
1028         xprt = kzalloc(size, GFP_KERNEL);
1029         if (xprt == NULL)
1030                 goto out;
1031
1032         xprt_init(xprt, net);
1033
1034         for (i = 0; i < num_prealloc; i++) {
1035                 req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL);
1036                 if (!req)
1037                         break;
1038                 list_add(&req->rq_list, &xprt->free);
1039         }
1040         if (i < num_prealloc)
1041                 goto out_free;
1042         if (max_alloc > num_prealloc)
1043                 xprt->max_reqs = max_alloc;
1044         else
1045                 xprt->max_reqs = num_prealloc;
1046         xprt->min_reqs = num_prealloc;
1047         atomic_set(&xprt->num_reqs, num_prealloc);
1048
1049         return xprt;
1050
1051 out_free:
1052         xprt_free(xprt);
1053 out:
1054         return NULL;
1055 }
1056 EXPORT_SYMBOL_GPL(xprt_alloc);
1057
1058 void xprt_free(struct rpc_xprt *xprt)
1059 {
1060         put_net(xprt->xprt_net);
1061         xprt_free_all_slots(xprt);
1062         kfree(xprt);
1063 }
1064 EXPORT_SYMBOL_GPL(xprt_free);
1065
1066 /**
1067  * xprt_reserve - allocate an RPC request slot
1068  * @task: RPC task requesting a slot allocation
1069  *
1070  * If no more slots are available, place the task on the transport's
1071  * backlog queue.
1072  */
1073 void xprt_reserve(struct rpc_task *task)
1074 {
1075         struct rpc_xprt *xprt = task->tk_xprt;
1076
1077         task->tk_status = 0;
1078         if (task->tk_rqstp != NULL)
1079                 return;
1080
1081         /* Note: grabbing the xprt_lock_write() here is not strictly needed,
1082          * but ensures that we throttle new slot allocation if the transport
1083          * is congested (e.g. if reconnecting or if we're out of socket
1084          * write buffer space).
1085          */
1086         task->tk_timeout = 0;
1087         task->tk_status = -EAGAIN;
1088         if (!xprt_lock_write(xprt, task))
1089                 return;
1090
1091         spin_lock(&xprt->reserve_lock);
1092         xprt_alloc_slot(task);
1093         spin_unlock(&xprt->reserve_lock);
1094         xprt_release_write(xprt, task);
1095 }
1096
1097 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
1098 {
1099         return (__force __be32)xprt->xid++;
1100 }
1101
1102 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1103 {
1104         xprt->xid = net_random();
1105 }
1106
1107 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1108 {
1109         struct rpc_rqst *req = task->tk_rqstp;
1110
1111         INIT_LIST_HEAD(&req->rq_list);
1112         req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1113         req->rq_task    = task;
1114         req->rq_xprt    = xprt;
1115         req->rq_buffer  = NULL;
1116         req->rq_xid     = xprt_alloc_xid(xprt);
1117         req->rq_release_snd_buf = NULL;
1118         xprt_reset_majortimeo(req);
1119         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1120                         req, ntohl(req->rq_xid));
1121 }
1122
1123 /**
1124  * xprt_release - release an RPC request slot
1125  * @task: task which is finished with the slot
1126  *
1127  */
1128 void xprt_release(struct rpc_task *task)
1129 {
1130         struct rpc_xprt *xprt;
1131         struct rpc_rqst *req;
1132
1133         if (!(req = task->tk_rqstp))
1134                 return;
1135
1136         xprt = req->rq_xprt;
1137         rpc_count_iostats(task);
1138         spin_lock_bh(&xprt->transport_lock);
1139         xprt->ops->release_xprt(xprt, task);
1140         if (xprt->ops->release_request)
1141                 xprt->ops->release_request(task);
1142         if (!list_empty(&req->rq_list))
1143                 list_del(&req->rq_list);
1144         xprt->last_used = jiffies;
1145         if (list_empty(&xprt->recv) && xprt_has_timer(xprt))
1146                 mod_timer(&xprt->timer,
1147                                 xprt->last_used + xprt->idle_timeout);
1148         spin_unlock_bh(&xprt->transport_lock);
1149         if (req->rq_buffer)
1150                 xprt->ops->buf_free(req->rq_buffer);
1151         if (req->rq_cred != NULL)
1152                 put_rpccred(req->rq_cred);
1153         task->tk_rqstp = NULL;
1154         if (req->rq_release_snd_buf)
1155                 req->rq_release_snd_buf(req);
1156
1157         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1158         if (likely(!bc_prealloc(req)))
1159                 xprt_free_slot(xprt, req);
1160         else
1161                 xprt_free_bc_request(req);
1162 }
1163
1164 static void xprt_init(struct rpc_xprt *xprt, struct net *net)
1165 {
1166         atomic_set(&xprt->count, 1);
1167
1168         spin_lock_init(&xprt->transport_lock);
1169         spin_lock_init(&xprt->reserve_lock);
1170
1171         INIT_LIST_HEAD(&xprt->free);
1172         INIT_LIST_HEAD(&xprt->recv);
1173 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1174         spin_lock_init(&xprt->bc_pa_lock);
1175         INIT_LIST_HEAD(&xprt->bc_pa_list);
1176 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1177
1178         xprt->last_used = jiffies;
1179         xprt->cwnd = RPC_INITCWND;
1180         xprt->bind_index = 0;
1181
1182         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1183         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1184         rpc_init_priority_wait_queue(&xprt->sending, "xprt_sending");
1185         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1186
1187         xprt_init_xid(xprt);
1188
1189         xprt->xprt_net = get_net(net);
1190 }
1191
1192 /**
1193  * xprt_create_transport - create an RPC transport
1194  * @args: rpc transport creation arguments
1195  *
1196  */
1197 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1198 {
1199         struct rpc_xprt *xprt;
1200         struct xprt_class *t;
1201
1202         spin_lock(&xprt_list_lock);
1203         list_for_each_entry(t, &xprt_list, list) {
1204                 if (t->ident == args->ident) {
1205                         spin_unlock(&xprt_list_lock);
1206                         goto found;
1207                 }
1208         }
1209         spin_unlock(&xprt_list_lock);
1210         printk(KERN_ERR "RPC: transport (%d) not supported\n", args->ident);
1211         return ERR_PTR(-EIO);
1212
1213 found:
1214         xprt = t->setup(args);
1215         if (IS_ERR(xprt)) {
1216                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1217                                 -PTR_ERR(xprt));
1218                 goto out;
1219         }
1220         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1221         if (xprt_has_timer(xprt))
1222                 setup_timer(&xprt->timer, xprt_init_autodisconnect,
1223                             (unsigned long)xprt);
1224         else
1225                 init_timer(&xprt->timer);
1226         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1227                         xprt->max_reqs);
1228 out:
1229         return xprt;
1230 }
1231
1232 /**
1233  * xprt_destroy - destroy an RPC transport, killing off all requests.
1234  * @xprt: transport to destroy
1235  *
1236  */
1237 static void xprt_destroy(struct rpc_xprt *xprt)
1238 {
1239         dprintk("RPC:       destroying transport %p\n", xprt);
1240         xprt->shutdown = 1;
1241         del_timer_sync(&xprt->timer);
1242
1243         rpc_destroy_wait_queue(&xprt->binding);
1244         rpc_destroy_wait_queue(&xprt->pending);
1245         rpc_destroy_wait_queue(&xprt->sending);
1246         rpc_destroy_wait_queue(&xprt->backlog);
1247         cancel_work_sync(&xprt->task_cleanup);
1248         /*
1249          * Tear down transport state and free the rpc_xprt
1250          */
1251         xprt->ops->destroy(xprt);
1252 }
1253
1254 /**
1255  * xprt_put - release a reference to an RPC transport.
1256  * @xprt: pointer to the transport
1257  *
1258  */
1259 void xprt_put(struct rpc_xprt *xprt)
1260 {
1261         if (atomic_dec_and_test(&xprt->count))
1262                 xprt_destroy(xprt);
1263 }
1264
1265 /**
1266  * xprt_get - return a reference to an RPC transport.
1267  * @xprt: pointer to the transport
1268  *
1269  */
1270 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1271 {
1272         if (atomic_inc_not_zero(&xprt->count))
1273                 return xprt;
1274         return NULL;
1275 }