make sure that we always have a return path from kernel_execve()
[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 static void      xprt_destroy(struct rpc_xprt *xprt);
70
71 static DEFINE_SPINLOCK(xprt_list_lock);
72 static LIST_HEAD(xprt_list);
73
74 /*
75  * The transport code maintains an estimate on the maximum number of out-
76  * standing RPC requests, using a smoothed version of the congestion
77  * avoidance implemented in 44BSD. This is basically the Van Jacobson
78  * congestion algorithm: If a retransmit occurs, the congestion window is
79  * halved; otherwise, it is incremented by 1/cwnd when
80  *
81  *      -       a reply is received and
82  *      -       a full number of requests are outstanding and
83  *      -       the congestion window hasn't been updated recently.
84  */
85 #define RPC_CWNDSHIFT           (8U)
86 #define RPC_CWNDSCALE           (1U << RPC_CWNDSHIFT)
87 #define RPC_INITCWND            RPC_CWNDSCALE
88 #define RPC_MAXCWND(xprt)       ((xprt)->max_reqs << RPC_CWNDSHIFT)
89
90 #define RPCXPRT_CONGESTED(xprt) ((xprt)->cong >= (xprt)->cwnd)
91
92 /**
93  * xprt_register_transport - register a transport implementation
94  * @transport: transport to register
95  *
96  * If a transport implementation is loaded as a kernel module, it can
97  * call this interface to make itself known to the RPC client.
98  *
99  * Returns:
100  * 0:           transport successfully registered
101  * -EEXIST:     transport already registered
102  * -EINVAL:     transport module being unloaded
103  */
104 int xprt_register_transport(struct xprt_class *transport)
105 {
106         struct xprt_class *t;
107         int result;
108
109         result = -EEXIST;
110         spin_lock(&xprt_list_lock);
111         list_for_each_entry(t, &xprt_list, list) {
112                 /* don't register the same transport class twice */
113                 if (t->ident == transport->ident)
114                         goto out;
115         }
116
117         list_add_tail(&transport->list, &xprt_list);
118         printk(KERN_INFO "RPC: Registered %s transport module.\n",
119                transport->name);
120         result = 0;
121
122 out:
123         spin_unlock(&xprt_list_lock);
124         return result;
125 }
126 EXPORT_SYMBOL_GPL(xprt_register_transport);
127
128 /**
129  * xprt_unregister_transport - unregister a transport implementation
130  * @transport: transport to unregister
131  *
132  * Returns:
133  * 0:           transport successfully unregistered
134  * -ENOENT:     transport never registered
135  */
136 int xprt_unregister_transport(struct xprt_class *transport)
137 {
138         struct xprt_class *t;
139         int result;
140
141         result = 0;
142         spin_lock(&xprt_list_lock);
143         list_for_each_entry(t, &xprt_list, list) {
144                 if (t == transport) {
145                         printk(KERN_INFO
146                                 "RPC: Unregistered %s transport module.\n",
147                                 transport->name);
148                         list_del_init(&transport->list);
149                         goto out;
150                 }
151         }
152         result = -ENOENT;
153
154 out:
155         spin_unlock(&xprt_list_lock);
156         return result;
157 }
158 EXPORT_SYMBOL_GPL(xprt_unregister_transport);
159
160 /**
161  * xprt_load_transport - load a transport implementation
162  * @transport_name: transport to load
163  *
164  * Returns:
165  * 0:           transport successfully loaded
166  * -ENOENT:     transport module not available
167  */
168 int xprt_load_transport(const char *transport_name)
169 {
170         struct xprt_class *t;
171         int result;
172
173         result = 0;
174         spin_lock(&xprt_list_lock);
175         list_for_each_entry(t, &xprt_list, list) {
176                 if (strcmp(t->name, transport_name) == 0) {
177                         spin_unlock(&xprt_list_lock);
178                         goto out;
179                 }
180         }
181         spin_unlock(&xprt_list_lock);
182         result = request_module("xprt%s", transport_name);
183 out:
184         return result;
185 }
186 EXPORT_SYMBOL_GPL(xprt_load_transport);
187
188 /**
189  * xprt_reserve_xprt - serialize write access to transports
190  * @task: task that is requesting access to the transport
191  * @xprt: pointer to the target transport
192  *
193  * This prevents mixing the payload of separate requests, and prevents
194  * transport connects from colliding with writes.  No congestion control
195  * is provided.
196  */
197 int xprt_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
198 {
199         struct rpc_rqst *req = task->tk_rqstp;
200         int priority;
201
202         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
203                 if (task == xprt->snd_task)
204                         return 1;
205                 goto out_sleep;
206         }
207         xprt->snd_task = task;
208         if (req != NULL) {
209                 req->rq_bytes_sent = 0;
210                 req->rq_ntrans++;
211         }
212
213         return 1;
214
215 out_sleep:
216         dprintk("RPC: %5u failed to lock transport %p\n",
217                         task->tk_pid, xprt);
218         task->tk_timeout = 0;
219         task->tk_status = -EAGAIN;
220         if (req == NULL)
221                 priority = RPC_PRIORITY_LOW;
222         else if (!req->rq_ntrans)
223                 priority = RPC_PRIORITY_NORMAL;
224         else
225                 priority = RPC_PRIORITY_HIGH;
226         rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
227         return 0;
228 }
229 EXPORT_SYMBOL_GPL(xprt_reserve_xprt);
230
231 static void xprt_clear_locked(struct rpc_xprt *xprt)
232 {
233         xprt->snd_task = NULL;
234         if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state) || xprt->shutdown) {
235                 smp_mb__before_clear_bit();
236                 clear_bit(XPRT_LOCKED, &xprt->state);
237                 smp_mb__after_clear_bit();
238         } else
239                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
240 }
241
242 /*
243  * xprt_reserve_xprt_cong - serialize write access to transports
244  * @task: task that is requesting access to the transport
245  *
246  * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
247  * integrated into the decision of whether a request is allowed to be
248  * woken up and given access to the transport.
249  */
250 int xprt_reserve_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
251 {
252         struct rpc_rqst *req = task->tk_rqstp;
253         int priority;
254
255         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
256                 if (task == xprt->snd_task)
257                         return 1;
258                 goto out_sleep;
259         }
260         if (req == NULL) {
261                 xprt->snd_task = task;
262                 return 1;
263         }
264         if (__xprt_get_cong(xprt, task)) {
265                 xprt->snd_task = task;
266                 req->rq_bytes_sent = 0;
267                 req->rq_ntrans++;
268                 return 1;
269         }
270         xprt_clear_locked(xprt);
271 out_sleep:
272         dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
273         task->tk_timeout = 0;
274         task->tk_status = -EAGAIN;
275         if (req == NULL)
276                 priority = RPC_PRIORITY_LOW;
277         else if (!req->rq_ntrans)
278                 priority = RPC_PRIORITY_NORMAL;
279         else
280                 priority = RPC_PRIORITY_HIGH;
281         rpc_sleep_on_priority(&xprt->sending, task, NULL, priority);
282         return 0;
283 }
284 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
285
286 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
287 {
288         int retval;
289
290         spin_lock_bh(&xprt->transport_lock);
291         retval = xprt->ops->reserve_xprt(xprt, task);
292         spin_unlock_bh(&xprt->transport_lock);
293         return retval;
294 }
295
296 static bool __xprt_lock_write_func(struct rpc_task *task, void *data)
297 {
298         struct rpc_xprt *xprt = data;
299         struct rpc_rqst *req;
300
301         req = task->tk_rqstp;
302         xprt->snd_task = task;
303         if (req) {
304                 req->rq_bytes_sent = 0;
305                 req->rq_ntrans++;
306         }
307         return true;
308 }
309
310 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
311 {
312         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
313                 return;
314
315         if (rpc_wake_up_first(&xprt->sending, __xprt_lock_write_func, xprt))
316                 return;
317         xprt_clear_locked(xprt);
318 }
319
320 static bool __xprt_lock_write_cong_func(struct rpc_task *task, void *data)
321 {
322         struct rpc_xprt *xprt = data;
323         struct rpc_rqst *req;
324
325         req = task->tk_rqstp;
326         if (req == NULL) {
327                 xprt->snd_task = task;
328                 return true;
329         }
330         if (__xprt_get_cong(xprt, task)) {
331                 xprt->snd_task = task;
332                 req->rq_bytes_sent = 0;
333                 req->rq_ntrans++;
334                 return true;
335         }
336         return false;
337 }
338
339 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
340 {
341         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
342                 return;
343         if (RPCXPRT_CONGESTED(xprt))
344                 goto out_unlock;
345         if (rpc_wake_up_first(&xprt->sending, __xprt_lock_write_cong_func, xprt))
346                 return;
347 out_unlock:
348         xprt_clear_locked(xprt);
349 }
350
351 /**
352  * xprt_release_xprt - allow other requests to use a transport
353  * @xprt: transport with other tasks potentially waiting
354  * @task: task that is releasing access to the transport
355  *
356  * Note that "task" can be NULL.  No congestion control is provided.
357  */
358 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
359 {
360         if (xprt->snd_task == task) {
361                 xprt_clear_locked(xprt);
362                 __xprt_lock_write_next(xprt);
363         }
364 }
365 EXPORT_SYMBOL_GPL(xprt_release_xprt);
366
367 /**
368  * xprt_release_xprt_cong - allow other requests to use a transport
369  * @xprt: transport with other tasks potentially waiting
370  * @task: task that is releasing access to the transport
371  *
372  * Note that "task" can be NULL.  Another task is awoken to use the
373  * transport if the transport's congestion window allows it.
374  */
375 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
376 {
377         if (xprt->snd_task == task) {
378                 xprt_clear_locked(xprt);
379                 __xprt_lock_write_next_cong(xprt);
380         }
381 }
382 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
383
384 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
385 {
386         spin_lock_bh(&xprt->transport_lock);
387         xprt->ops->release_xprt(xprt, task);
388         spin_unlock_bh(&xprt->transport_lock);
389 }
390
391 /*
392  * Van Jacobson congestion avoidance. Check if the congestion window
393  * overflowed. Put the task to sleep if this is the case.
394  */
395 static int
396 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_task *task)
397 {
398         struct rpc_rqst *req = task->tk_rqstp;
399
400         if (req->rq_cong)
401                 return 1;
402         dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
403                         task->tk_pid, xprt->cong, xprt->cwnd);
404         if (RPCXPRT_CONGESTED(xprt))
405                 return 0;
406         req->rq_cong = 1;
407         xprt->cong += RPC_CWNDSCALE;
408         return 1;
409 }
410
411 /*
412  * Adjust the congestion window, and wake up the next task
413  * that has been sleeping due to congestion
414  */
415 static void
416 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
417 {
418         if (!req->rq_cong)
419                 return;
420         req->rq_cong = 0;
421         xprt->cong -= RPC_CWNDSCALE;
422         __xprt_lock_write_next_cong(xprt);
423 }
424
425 /**
426  * xprt_release_rqst_cong - housekeeping when request is complete
427  * @task: RPC request that recently completed
428  *
429  * Useful for transports that require congestion control.
430  */
431 void xprt_release_rqst_cong(struct rpc_task *task)
432 {
433         __xprt_put_cong(task->tk_xprt, task->tk_rqstp);
434 }
435 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
436
437 /**
438  * xprt_adjust_cwnd - adjust transport congestion window
439  * @task: recently completed RPC request used to adjust window
440  * @result: result code of completed RPC request
441  *
442  * We use a time-smoothed congestion estimator to avoid heavy oscillation.
443  */
444 void xprt_adjust_cwnd(struct rpc_task *task, int result)
445 {
446         struct rpc_rqst *req = task->tk_rqstp;
447         struct rpc_xprt *xprt = task->tk_xprt;
448         unsigned long cwnd = xprt->cwnd;
449
450         if (result >= 0 && cwnd <= xprt->cong) {
451                 /* The (cwnd >> 1) term makes sure
452                  * the result gets rounded properly. */
453                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
454                 if (cwnd > RPC_MAXCWND(xprt))
455                         cwnd = RPC_MAXCWND(xprt);
456                 __xprt_lock_write_next_cong(xprt);
457         } else if (result == -ETIMEDOUT) {
458                 cwnd >>= 1;
459                 if (cwnd < RPC_CWNDSCALE)
460                         cwnd = RPC_CWNDSCALE;
461         }
462         dprintk("RPC:       cong %ld, cwnd was %ld, now %ld\n",
463                         xprt->cong, xprt->cwnd, cwnd);
464         xprt->cwnd = cwnd;
465         __xprt_put_cong(xprt, req);
466 }
467 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
468
469 /**
470  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
471  * @xprt: transport with waiting tasks
472  * @status: result code to plant in each task before waking it
473  *
474  */
475 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
476 {
477         if (status < 0)
478                 rpc_wake_up_status(&xprt->pending, status);
479         else
480                 rpc_wake_up(&xprt->pending);
481 }
482 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
483
484 /**
485  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
486  * @task: task to be put to sleep
487  * @action: function pointer to be executed after wait
488  */
489 void xprt_wait_for_buffer_space(struct rpc_task *task, rpc_action action)
490 {
491         struct rpc_rqst *req = task->tk_rqstp;
492         struct rpc_xprt *xprt = req->rq_xprt;
493
494         task->tk_timeout = req->rq_timeout;
495         rpc_sleep_on(&xprt->pending, task, action);
496 }
497 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
498
499 /**
500  * xprt_write_space - wake the task waiting for transport output buffer space
501  * @xprt: transport with waiting tasks
502  *
503  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
504  */
505 void xprt_write_space(struct rpc_xprt *xprt)
506 {
507         if (unlikely(xprt->shutdown))
508                 return;
509
510         spin_lock_bh(&xprt->transport_lock);
511         if (xprt->snd_task) {
512                 dprintk("RPC:       write space: waking waiting task on "
513                                 "xprt %p\n", xprt);
514                 rpc_wake_up_queued_task(&xprt->pending, xprt->snd_task);
515         }
516         spin_unlock_bh(&xprt->transport_lock);
517 }
518 EXPORT_SYMBOL_GPL(xprt_write_space);
519
520 /**
521  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
522  * @task: task whose timeout is to be set
523  *
524  * Set a request's retransmit timeout based on the transport's
525  * default timeout parameters.  Used by transports that don't adjust
526  * the retransmit timeout based on round-trip time estimation.
527  */
528 void xprt_set_retrans_timeout_def(struct rpc_task *task)
529 {
530         task->tk_timeout = task->tk_rqstp->rq_timeout;
531 }
532 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
533
534 /**
535  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
536  * @task: task whose timeout is to be set
537  *
538  * Set a request's retransmit timeout using the RTT estimator.
539  */
540 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
541 {
542         int timer = task->tk_msg.rpc_proc->p_timer;
543         struct rpc_clnt *clnt = task->tk_client;
544         struct rpc_rtt *rtt = clnt->cl_rtt;
545         struct rpc_rqst *req = task->tk_rqstp;
546         unsigned long max_timeout = clnt->cl_timeout->to_maxval;
547
548         task->tk_timeout = rpc_calc_rto(rtt, timer);
549         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
550         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
551                 task->tk_timeout = max_timeout;
552 }
553 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
554
555 static void xprt_reset_majortimeo(struct rpc_rqst *req)
556 {
557         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
558
559         req->rq_majortimeo = req->rq_timeout;
560         if (to->to_exponential)
561                 req->rq_majortimeo <<= to->to_retries;
562         else
563                 req->rq_majortimeo += to->to_increment * to->to_retries;
564         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
565                 req->rq_majortimeo = to->to_maxval;
566         req->rq_majortimeo += jiffies;
567 }
568
569 /**
570  * xprt_adjust_timeout - adjust timeout values for next retransmit
571  * @req: RPC request containing parameters to use for the adjustment
572  *
573  */
574 int xprt_adjust_timeout(struct rpc_rqst *req)
575 {
576         struct rpc_xprt *xprt = req->rq_xprt;
577         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
578         int status = 0;
579
580         if (time_before(jiffies, req->rq_majortimeo)) {
581                 if (to->to_exponential)
582                         req->rq_timeout <<= 1;
583                 else
584                         req->rq_timeout += to->to_increment;
585                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
586                         req->rq_timeout = to->to_maxval;
587                 req->rq_retries++;
588         } else {
589                 req->rq_timeout = to->to_initval;
590                 req->rq_retries = 0;
591                 xprt_reset_majortimeo(req);
592                 /* Reset the RTT counters == "slow start" */
593                 spin_lock_bh(&xprt->transport_lock);
594                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
595                 spin_unlock_bh(&xprt->transport_lock);
596                 status = -ETIMEDOUT;
597         }
598
599         if (req->rq_timeout == 0) {
600                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
601                 req->rq_timeout = 5 * HZ;
602         }
603         return status;
604 }
605
606 static void xprt_autoclose(struct work_struct *work)
607 {
608         struct rpc_xprt *xprt =
609                 container_of(work, struct rpc_xprt, task_cleanup);
610
611         xprt->ops->close(xprt);
612         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
613         xprt_release_write(xprt, NULL);
614 }
615
616 /**
617  * xprt_disconnect_done - mark a transport as disconnected
618  * @xprt: transport to flag for disconnect
619  *
620  */
621 void xprt_disconnect_done(struct rpc_xprt *xprt)
622 {
623         dprintk("RPC:       disconnected transport %p\n", xprt);
624         spin_lock_bh(&xprt->transport_lock);
625         xprt_clear_connected(xprt);
626         xprt_wake_pending_tasks(xprt, -EAGAIN);
627         spin_unlock_bh(&xprt->transport_lock);
628 }
629 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
630
631 /**
632  * xprt_force_disconnect - force a transport to disconnect
633  * @xprt: transport to disconnect
634  *
635  */
636 void xprt_force_disconnect(struct rpc_xprt *xprt)
637 {
638         /* Don't race with the test_bit() in xprt_clear_locked() */
639         spin_lock_bh(&xprt->transport_lock);
640         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
641         /* Try to schedule an autoclose RPC call */
642         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
643                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
644         xprt_wake_pending_tasks(xprt, -EAGAIN);
645         spin_unlock_bh(&xprt->transport_lock);
646 }
647
648 /**
649  * xprt_conditional_disconnect - force a transport to disconnect
650  * @xprt: transport to disconnect
651  * @cookie: 'connection cookie'
652  *
653  * This attempts to break the connection if and only if 'cookie' matches
654  * the current transport 'connection cookie'. It ensures that we don't
655  * try to break the connection more than once when we need to retransmit
656  * a batch of RPC requests.
657  *
658  */
659 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
660 {
661         /* Don't race with the test_bit() in xprt_clear_locked() */
662         spin_lock_bh(&xprt->transport_lock);
663         if (cookie != xprt->connect_cookie)
664                 goto out;
665         if (test_bit(XPRT_CLOSING, &xprt->state) || !xprt_connected(xprt))
666                 goto out;
667         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
668         /* Try to schedule an autoclose RPC call */
669         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
670                 queue_work(rpciod_workqueue, &xprt->task_cleanup);
671         xprt_wake_pending_tasks(xprt, -EAGAIN);
672 out:
673         spin_unlock_bh(&xprt->transport_lock);
674 }
675
676 static void
677 xprt_init_autodisconnect(unsigned long data)
678 {
679         struct rpc_xprt *xprt = (struct rpc_xprt *)data;
680
681         spin_lock(&xprt->transport_lock);
682         if (!list_empty(&xprt->recv) || xprt->shutdown)
683                 goto out_abort;
684         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
685                 goto out_abort;
686         spin_unlock(&xprt->transport_lock);
687         set_bit(XPRT_CONNECTION_CLOSE, &xprt->state);
688         queue_work(rpciod_workqueue, &xprt->task_cleanup);
689         return;
690 out_abort:
691         spin_unlock(&xprt->transport_lock);
692 }
693
694 /**
695  * xprt_connect - schedule a transport connect operation
696  * @task: RPC task that is requesting the connect
697  *
698  */
699 void xprt_connect(struct rpc_task *task)
700 {
701         struct rpc_xprt *xprt = task->tk_xprt;
702
703         dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
704                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
705
706         if (!xprt_bound(xprt)) {
707                 task->tk_status = -EAGAIN;
708                 return;
709         }
710         if (!xprt_lock_write(xprt, task))
711                 return;
712
713         if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
714                 xprt->ops->close(xprt);
715
716         if (xprt_connected(xprt))
717                 xprt_release_write(xprt, task);
718         else {
719                 task->tk_rqstp->rq_bytes_sent = 0;
720                 task->tk_timeout = task->tk_rqstp->rq_timeout;
721                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
722
723                 if (test_bit(XPRT_CLOSING, &xprt->state))
724                         return;
725                 if (xprt_test_and_set_connecting(xprt))
726                         return;
727                 xprt->stat.connect_start = jiffies;
728                 xprt->ops->connect(task);
729         }
730 }
731
732 static void xprt_connect_status(struct rpc_task *task)
733 {
734         struct rpc_xprt *xprt = task->tk_xprt;
735
736         if (task->tk_status == 0) {
737                 xprt->stat.connect_count++;
738                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
739                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
740                                 task->tk_pid);
741                 return;
742         }
743
744         switch (task->tk_status) {
745         case -EAGAIN:
746                 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
747                 break;
748         case -ETIMEDOUT:
749                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
750                                 "out\n", task->tk_pid);
751                 break;
752         default:
753                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
754                                 "server %s\n", task->tk_pid, -task->tk_status,
755                                 xprt->servername);
756                 xprt_release_write(xprt, task);
757                 task->tk_status = -EIO;
758         }
759 }
760
761 /**
762  * xprt_lookup_rqst - find an RPC request corresponding to an XID
763  * @xprt: transport on which the original request was transmitted
764  * @xid: RPC XID of incoming reply
765  *
766  */
767 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
768 {
769         struct rpc_rqst *entry;
770
771         list_for_each_entry(entry, &xprt->recv, rq_list)
772                 if (entry->rq_xid == xid)
773                         return entry;
774
775         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
776                         ntohl(xid));
777         xprt->stat.bad_xids++;
778         return NULL;
779 }
780 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
781
782 static void xprt_update_rtt(struct rpc_task *task)
783 {
784         struct rpc_rqst *req = task->tk_rqstp;
785         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
786         unsigned int timer = task->tk_msg.rpc_proc->p_timer;
787         long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
788
789         if (timer) {
790                 if (req->rq_ntrans == 1)
791                         rpc_update_rtt(rtt, timer, m);
792                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
793         }
794 }
795
796 /**
797  * xprt_complete_rqst - called when reply processing is complete
798  * @task: RPC request that recently completed
799  * @copied: actual number of bytes received from the transport
800  *
801  * Caller holds transport lock.
802  */
803 void xprt_complete_rqst(struct rpc_task *task, int copied)
804 {
805         struct rpc_rqst *req = task->tk_rqstp;
806         struct rpc_xprt *xprt = req->rq_xprt;
807
808         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
809                         task->tk_pid, ntohl(req->rq_xid), copied);
810
811         xprt->stat.recvs++;
812         req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime);
813         if (xprt->ops->timer != NULL)
814                 xprt_update_rtt(task);
815
816         list_del_init(&req->rq_list);
817         req->rq_private_buf.len = copied;
818         /* Ensure all writes are done before we update */
819         /* req->rq_reply_bytes_recvd */
820         smp_wmb();
821         req->rq_reply_bytes_recvd = copied;
822         rpc_wake_up_queued_task(&xprt->pending, task);
823 }
824 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
825
826 static void xprt_timer(struct rpc_task *task)
827 {
828         struct rpc_rqst *req = task->tk_rqstp;
829         struct rpc_xprt *xprt = req->rq_xprt;
830
831         if (task->tk_status != -ETIMEDOUT)
832                 return;
833         dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
834
835         spin_lock_bh(&xprt->transport_lock);
836         if (!req->rq_reply_bytes_recvd) {
837                 if (xprt->ops->timer)
838                         xprt->ops->timer(task);
839         } else
840                 task->tk_status = 0;
841         spin_unlock_bh(&xprt->transport_lock);
842 }
843
844 static inline int xprt_has_timer(struct rpc_xprt *xprt)
845 {
846         return xprt->idle_timeout != 0;
847 }
848
849 /**
850  * xprt_prepare_transmit - reserve the transport before sending a request
851  * @task: RPC task about to send a request
852  *
853  */
854 int xprt_prepare_transmit(struct rpc_task *task)
855 {
856         struct rpc_rqst *req = task->tk_rqstp;
857         struct rpc_xprt *xprt = req->rq_xprt;
858         int err = 0;
859
860         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
861
862         spin_lock_bh(&xprt->transport_lock);
863         if (req->rq_reply_bytes_recvd && !req->rq_bytes_sent) {
864                 err = req->rq_reply_bytes_recvd;
865                 goto out_unlock;
866         }
867         if (!xprt->ops->reserve_xprt(xprt, task))
868                 err = -EAGAIN;
869 out_unlock:
870         spin_unlock_bh(&xprt->transport_lock);
871         return err;
872 }
873
874 void xprt_end_transmit(struct rpc_task *task)
875 {
876         xprt_release_write(task->tk_rqstp->rq_xprt, task);
877 }
878
879 /**
880  * xprt_transmit - send an RPC request on a transport
881  * @task: controlling RPC task
882  *
883  * We have to copy the iovec because sendmsg fiddles with its contents.
884  */
885 void xprt_transmit(struct rpc_task *task)
886 {
887         struct rpc_rqst *req = task->tk_rqstp;
888         struct rpc_xprt *xprt = req->rq_xprt;
889         int status, numreqs;
890
891         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
892
893         if (!req->rq_reply_bytes_recvd) {
894                 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) {
895                         /*
896                          * Add to the list only if we're expecting a reply
897                          */
898                         spin_lock_bh(&xprt->transport_lock);
899                         /* Update the softirq receive buffer */
900                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
901                                         sizeof(req->rq_private_buf));
902                         /* Add request to the receive list */
903                         list_add_tail(&req->rq_list, &xprt->recv);
904                         spin_unlock_bh(&xprt->transport_lock);
905                         xprt_reset_majortimeo(req);
906                         /* Turn off autodisconnect */
907                         del_singleshot_timer_sync(&xprt->timer);
908                 }
909         } else if (!req->rq_bytes_sent)
910                 return;
911
912         req->rq_connect_cookie = xprt->connect_cookie;
913         req->rq_xtime = ktime_get();
914         status = xprt->ops->send_request(task);
915         if (status != 0) {
916                 task->tk_status = status;
917                 return;
918         }
919
920         dprintk("RPC: %5u xmit complete\n", task->tk_pid);
921         task->tk_flags |= RPC_TASK_SENT;
922         spin_lock_bh(&xprt->transport_lock);
923
924         xprt->ops->set_retrans_timeout(task);
925
926         numreqs = atomic_read(&xprt->num_reqs);
927         if (numreqs > xprt->stat.max_slots)
928                 xprt->stat.max_slots = numreqs;
929         xprt->stat.sends++;
930         xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
931         xprt->stat.bklog_u += xprt->backlog.qlen;
932         xprt->stat.sending_u += xprt->sending.qlen;
933         xprt->stat.pending_u += xprt->pending.qlen;
934
935         /* Don't race with disconnect */
936         if (!xprt_connected(xprt))
937                 task->tk_status = -ENOTCONN;
938         else if (!req->rq_reply_bytes_recvd && rpc_reply_expected(task)) {
939                 /*
940                  * Sleep on the pending queue since
941                  * we're expecting a reply.
942                  */
943                 rpc_sleep_on(&xprt->pending, task, xprt_timer);
944         }
945         spin_unlock_bh(&xprt->transport_lock);
946 }
947
948 static struct rpc_rqst *xprt_dynamic_alloc_slot(struct rpc_xprt *xprt, gfp_t gfp_flags)
949 {
950         struct rpc_rqst *req = ERR_PTR(-EAGAIN);
951
952         if (!atomic_add_unless(&xprt->num_reqs, 1, xprt->max_reqs))
953                 goto out;
954         req = kzalloc(sizeof(struct rpc_rqst), gfp_flags);
955         if (req != NULL)
956                 goto out;
957         atomic_dec(&xprt->num_reqs);
958         req = ERR_PTR(-ENOMEM);
959 out:
960         return req;
961 }
962
963 static bool xprt_dynamic_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
964 {
965         if (atomic_add_unless(&xprt->num_reqs, -1, xprt->min_reqs)) {
966                 kfree(req);
967                 return true;
968         }
969         return false;
970 }
971
972 void xprt_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
973 {
974         struct rpc_rqst *req;
975
976         spin_lock(&xprt->reserve_lock);
977         if (!list_empty(&xprt->free)) {
978                 req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
979                 list_del(&req->rq_list);
980                 goto out_init_req;
981         }
982         req = xprt_dynamic_alloc_slot(xprt, GFP_NOWAIT|__GFP_NOWARN);
983         if (!IS_ERR(req))
984                 goto out_init_req;
985         switch (PTR_ERR(req)) {
986         case -ENOMEM:
987                 dprintk("RPC:       dynamic allocation of request slot "
988                                 "failed! Retrying\n");
989                 task->tk_status = -ENOMEM;
990                 break;
991         case -EAGAIN:
992                 rpc_sleep_on(&xprt->backlog, task, NULL);
993                 dprintk("RPC:       waiting for request slot\n");
994         default:
995                 task->tk_status = -EAGAIN;
996         }
997         spin_unlock(&xprt->reserve_lock);
998         return;
999 out_init_req:
1000         task->tk_status = 0;
1001         task->tk_rqstp = req;
1002         xprt_request_init(task, xprt);
1003         spin_unlock(&xprt->reserve_lock);
1004 }
1005 EXPORT_SYMBOL_GPL(xprt_alloc_slot);
1006
1007 void xprt_lock_and_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1008 {
1009         /* Note: grabbing the xprt_lock_write() ensures that we throttle
1010          * new slot allocation if the transport is congested (i.e. when
1011          * reconnecting a stream transport or when out of socket write
1012          * buffer space).
1013          */
1014         if (xprt_lock_write(xprt, task)) {
1015                 xprt_alloc_slot(xprt, task);
1016                 xprt_release_write(xprt, task);
1017         }
1018 }
1019 EXPORT_SYMBOL_GPL(xprt_lock_and_alloc_slot);
1020
1021 static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1022 {
1023         spin_lock(&xprt->reserve_lock);
1024         if (!xprt_dynamic_free_slot(xprt, req)) {
1025                 memset(req, 0, sizeof(*req));   /* mark unused */
1026                 list_add(&req->rq_list, &xprt->free);
1027         }
1028         rpc_wake_up_next(&xprt->backlog);
1029         spin_unlock(&xprt->reserve_lock);
1030 }
1031
1032 static void xprt_free_all_slots(struct rpc_xprt *xprt)
1033 {
1034         struct rpc_rqst *req;
1035         while (!list_empty(&xprt->free)) {
1036                 req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list);
1037                 list_del(&req->rq_list);
1038                 kfree(req);
1039         }
1040 }
1041
1042 struct rpc_xprt *xprt_alloc(struct net *net, size_t size,
1043                 unsigned int num_prealloc,
1044                 unsigned int max_alloc)
1045 {
1046         struct rpc_xprt *xprt;
1047         struct rpc_rqst *req;
1048         int i;
1049
1050         xprt = kzalloc(size, GFP_KERNEL);
1051         if (xprt == NULL)
1052                 goto out;
1053
1054         xprt_init(xprt, net);
1055
1056         for (i = 0; i < num_prealloc; i++) {
1057                 req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL);
1058                 if (!req)
1059                         break;
1060                 list_add(&req->rq_list, &xprt->free);
1061         }
1062         if (i < num_prealloc)
1063                 goto out_free;
1064         if (max_alloc > num_prealloc)
1065                 xprt->max_reqs = max_alloc;
1066         else
1067                 xprt->max_reqs = num_prealloc;
1068         xprt->min_reqs = num_prealloc;
1069         atomic_set(&xprt->num_reqs, num_prealloc);
1070
1071         return xprt;
1072
1073 out_free:
1074         xprt_free(xprt);
1075 out:
1076         return NULL;
1077 }
1078 EXPORT_SYMBOL_GPL(xprt_alloc);
1079
1080 void xprt_free(struct rpc_xprt *xprt)
1081 {
1082         put_net(xprt->xprt_net);
1083         xprt_free_all_slots(xprt);
1084         kfree(xprt);
1085 }
1086 EXPORT_SYMBOL_GPL(xprt_free);
1087
1088 /**
1089  * xprt_reserve - allocate an RPC request slot
1090  * @task: RPC task requesting a slot allocation
1091  *
1092  * If no more slots are available, place the task on the transport's
1093  * backlog queue.
1094  */
1095 void xprt_reserve(struct rpc_task *task)
1096 {
1097         struct rpc_xprt *xprt = task->tk_xprt;
1098
1099         task->tk_status = 0;
1100         if (task->tk_rqstp != NULL)
1101                 return;
1102
1103         task->tk_timeout = 0;
1104         task->tk_status = -EAGAIN;
1105         xprt->ops->alloc_slot(xprt, task);
1106 }
1107
1108 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
1109 {
1110         return (__force __be32)xprt->xid++;
1111 }
1112
1113 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1114 {
1115         xprt->xid = net_random();
1116 }
1117
1118 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1119 {
1120         struct rpc_rqst *req = task->tk_rqstp;
1121
1122         INIT_LIST_HEAD(&req->rq_list);
1123         req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1124         req->rq_task    = task;
1125         req->rq_xprt    = xprt;
1126         req->rq_buffer  = NULL;
1127         req->rq_xid     = xprt_alloc_xid(xprt);
1128         req->rq_release_snd_buf = NULL;
1129         xprt_reset_majortimeo(req);
1130         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1131                         req, ntohl(req->rq_xid));
1132 }
1133
1134 /**
1135  * xprt_release - release an RPC request slot
1136  * @task: task which is finished with the slot
1137  *
1138  */
1139 void xprt_release(struct rpc_task *task)
1140 {
1141         struct rpc_xprt *xprt;
1142         struct rpc_rqst *req;
1143
1144         if (!(req = task->tk_rqstp))
1145                 return;
1146
1147         xprt = req->rq_xprt;
1148         if (task->tk_ops->rpc_count_stats != NULL)
1149                 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
1150         else if (task->tk_client)
1151                 rpc_count_iostats(task, task->tk_client->cl_metrics);
1152         spin_lock_bh(&xprt->transport_lock);
1153         xprt->ops->release_xprt(xprt, task);
1154         if (xprt->ops->release_request)
1155                 xprt->ops->release_request(task);
1156         if (!list_empty(&req->rq_list))
1157                 list_del(&req->rq_list);
1158         xprt->last_used = jiffies;
1159         if (list_empty(&xprt->recv) && xprt_has_timer(xprt))
1160                 mod_timer(&xprt->timer,
1161                                 xprt->last_used + xprt->idle_timeout);
1162         spin_unlock_bh(&xprt->transport_lock);
1163         if (req->rq_buffer)
1164                 xprt->ops->buf_free(req->rq_buffer);
1165         if (req->rq_cred != NULL)
1166                 put_rpccred(req->rq_cred);
1167         task->tk_rqstp = NULL;
1168         if (req->rq_release_snd_buf)
1169                 req->rq_release_snd_buf(req);
1170
1171         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1172         if (likely(!bc_prealloc(req)))
1173                 xprt_free_slot(xprt, req);
1174         else
1175                 xprt_free_bc_request(req);
1176 }
1177
1178 static void xprt_init(struct rpc_xprt *xprt, struct net *net)
1179 {
1180         atomic_set(&xprt->count, 1);
1181
1182         spin_lock_init(&xprt->transport_lock);
1183         spin_lock_init(&xprt->reserve_lock);
1184
1185         INIT_LIST_HEAD(&xprt->free);
1186         INIT_LIST_HEAD(&xprt->recv);
1187 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1188         spin_lock_init(&xprt->bc_pa_lock);
1189         INIT_LIST_HEAD(&xprt->bc_pa_list);
1190 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1191
1192         xprt->last_used = jiffies;
1193         xprt->cwnd = RPC_INITCWND;
1194         xprt->bind_index = 0;
1195
1196         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1197         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1198         rpc_init_priority_wait_queue(&xprt->sending, "xprt_sending");
1199         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1200
1201         xprt_init_xid(xprt);
1202
1203         xprt->xprt_net = get_net(net);
1204 }
1205
1206 /**
1207  * xprt_create_transport - create an RPC transport
1208  * @args: rpc transport creation arguments
1209  *
1210  */
1211 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1212 {
1213         struct rpc_xprt *xprt;
1214         struct xprt_class *t;
1215
1216         spin_lock(&xprt_list_lock);
1217         list_for_each_entry(t, &xprt_list, list) {
1218                 if (t->ident == args->ident) {
1219                         spin_unlock(&xprt_list_lock);
1220                         goto found;
1221                 }
1222         }
1223         spin_unlock(&xprt_list_lock);
1224         printk(KERN_ERR "RPC: transport (%d) not supported\n", args->ident);
1225         return ERR_PTR(-EIO);
1226
1227 found:
1228         xprt = t->setup(args);
1229         if (IS_ERR(xprt)) {
1230                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1231                                 -PTR_ERR(xprt));
1232                 goto out;
1233         }
1234         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1235         if (xprt_has_timer(xprt))
1236                 setup_timer(&xprt->timer, xprt_init_autodisconnect,
1237                             (unsigned long)xprt);
1238         else
1239                 init_timer(&xprt->timer);
1240
1241         if (strlen(args->servername) > RPC_MAXNETNAMELEN) {
1242                 xprt_destroy(xprt);
1243                 return ERR_PTR(-EINVAL);
1244         }
1245         xprt->servername = kstrdup(args->servername, GFP_KERNEL);
1246         if (xprt->servername == NULL) {
1247                 xprt_destroy(xprt);
1248                 return ERR_PTR(-ENOMEM);
1249         }
1250
1251         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1252                         xprt->max_reqs);
1253 out:
1254         return xprt;
1255 }
1256
1257 /**
1258  * xprt_destroy - destroy an RPC transport, killing off all requests.
1259  * @xprt: transport to destroy
1260  *
1261  */
1262 static void xprt_destroy(struct rpc_xprt *xprt)
1263 {
1264         dprintk("RPC:       destroying transport %p\n", xprt);
1265         xprt->shutdown = 1;
1266         del_timer_sync(&xprt->timer);
1267
1268         rpc_destroy_wait_queue(&xprt->binding);
1269         rpc_destroy_wait_queue(&xprt->pending);
1270         rpc_destroy_wait_queue(&xprt->sending);
1271         rpc_destroy_wait_queue(&xprt->backlog);
1272         cancel_work_sync(&xprt->task_cleanup);
1273         kfree(xprt->servername);
1274         /*
1275          * Tear down transport state and free the rpc_xprt
1276          */
1277         xprt->ops->destroy(xprt);
1278 }
1279
1280 /**
1281  * xprt_put - release a reference to an RPC transport.
1282  * @xprt: pointer to the transport
1283  *
1284  */
1285 void xprt_put(struct rpc_xprt *xprt)
1286 {
1287         if (atomic_dec_and_test(&xprt->count))
1288                 xprt_destroy(xprt);
1289 }
1290
1291 /**
1292  * xprt_get - return a reference to an RPC transport.
1293  * @xprt: pointer to the transport
1294  *
1295  */
1296 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1297 {
1298         if (atomic_inc_not_zero(&xprt->count))
1299                 return xprt;
1300         return NULL;
1301 }