]> Pileus Git - ~andy/linux/blob - net/sunrpc/xprt.c
SUNRPC: add sending,pending queue and max slot to xprt stats
[~andy/linux] / 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                 task->tk_rqstp->rq_bytes_sent = 0;
719                 task->tk_timeout = task->tk_rqstp->rq_timeout;
720                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
721
722                 if (test_bit(XPRT_CLOSING, &xprt->state))
723                         return;
724                 if (xprt_test_and_set_connecting(xprt))
725                         return;
726                 xprt->stat.connect_start = jiffies;
727                 xprt->ops->connect(task);
728         }
729 }
730
731 static void xprt_connect_status(struct rpc_task *task)
732 {
733         struct rpc_xprt *xprt = task->tk_xprt;
734
735         if (task->tk_status == 0) {
736                 xprt->stat.connect_count++;
737                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
738                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
739                                 task->tk_pid);
740                 return;
741         }
742
743         switch (task->tk_status) {
744         case -EAGAIN:
745                 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
746                 break;
747         case -ETIMEDOUT:
748                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
749                                 "out\n", task->tk_pid);
750                 break;
751         default:
752                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
753                                 "server %s\n", task->tk_pid, -task->tk_status,
754                                 task->tk_client->cl_server);
755                 xprt_release_write(xprt, task);
756                 task->tk_status = -EIO;
757         }
758 }
759
760 /**
761  * xprt_lookup_rqst - find an RPC request corresponding to an XID
762  * @xprt: transport on which the original request was transmitted
763  * @xid: RPC XID of incoming reply
764  *
765  */
766 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
767 {
768         struct rpc_rqst *entry;
769
770         list_for_each_entry(entry, &xprt->recv, rq_list)
771                 if (entry->rq_xid == xid)
772                         return entry;
773
774         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
775                         ntohl(xid));
776         xprt->stat.bad_xids++;
777         return NULL;
778 }
779 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
780
781 static void xprt_update_rtt(struct rpc_task *task)
782 {
783         struct rpc_rqst *req = task->tk_rqstp;
784         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
785         unsigned timer = task->tk_msg.rpc_proc->p_timer;
786         long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
787
788         if (timer) {
789                 if (req->rq_ntrans == 1)
790                         rpc_update_rtt(rtt, timer, m);
791                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
792         }
793 }
794
795 /**
796  * xprt_complete_rqst - called when reply processing is complete
797  * @task: RPC request that recently completed
798  * @copied: actual number of bytes received from the transport
799  *
800  * Caller holds transport lock.
801  */
802 void xprt_complete_rqst(struct rpc_task *task, int copied)
803 {
804         struct rpc_rqst *req = task->tk_rqstp;
805         struct rpc_xprt *xprt = req->rq_xprt;
806
807         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
808                         task->tk_pid, ntohl(req->rq_xid), copied);
809
810         xprt->stat.recvs++;
811         req->rq_rtt = ktime_sub(ktime_get(), req->rq_xtime);
812         if (xprt->ops->timer != NULL)
813                 xprt_update_rtt(task);
814
815         list_del_init(&req->rq_list);
816         req->rq_private_buf.len = copied;
817         /* Ensure all writes are done before we update */
818         /* req->rq_reply_bytes_recvd */
819         smp_wmb();
820         req->rq_reply_bytes_recvd = copied;
821         rpc_wake_up_queued_task(&xprt->pending, task);
822 }
823 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
824
825 static void xprt_timer(struct rpc_task *task)
826 {
827         struct rpc_rqst *req = task->tk_rqstp;
828         struct rpc_xprt *xprt = req->rq_xprt;
829
830         if (task->tk_status != -ETIMEDOUT)
831                 return;
832         dprintk("RPC: %5u xprt_timer\n", task->tk_pid);
833
834         spin_lock_bh(&xprt->transport_lock);
835         if (!req->rq_reply_bytes_recvd) {
836                 if (xprt->ops->timer)
837                         xprt->ops->timer(task);
838         } else
839                 task->tk_status = 0;
840         spin_unlock_bh(&xprt->transport_lock);
841 }
842
843 static inline int xprt_has_timer(struct rpc_xprt *xprt)
844 {
845         return xprt->idle_timeout != 0;
846 }
847
848 /**
849  * xprt_prepare_transmit - reserve the transport before sending a request
850  * @task: RPC task about to send a request
851  *
852  */
853 int xprt_prepare_transmit(struct rpc_task *task)
854 {
855         struct rpc_rqst *req = task->tk_rqstp;
856         struct rpc_xprt *xprt = req->rq_xprt;
857         int err = 0;
858
859         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
860
861         spin_lock_bh(&xprt->transport_lock);
862         if (req->rq_reply_bytes_recvd && !req->rq_bytes_sent) {
863                 err = req->rq_reply_bytes_recvd;
864                 goto out_unlock;
865         }
866         if (!xprt->ops->reserve_xprt(xprt, task))
867                 err = -EAGAIN;
868 out_unlock:
869         spin_unlock_bh(&xprt->transport_lock);
870         return err;
871 }
872
873 void xprt_end_transmit(struct rpc_task *task)
874 {
875         xprt_release_write(task->tk_rqstp->rq_xprt, task);
876 }
877
878 /**
879  * xprt_transmit - send an RPC request on a transport
880  * @task: controlling RPC task
881  *
882  * We have to copy the iovec because sendmsg fiddles with its contents.
883  */
884 void xprt_transmit(struct rpc_task *task)
885 {
886         struct rpc_rqst *req = task->tk_rqstp;
887         struct rpc_xprt *xprt = req->rq_xprt;
888         int status, numreqs;
889
890         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
891
892         if (!req->rq_reply_bytes_recvd) {
893                 if (list_empty(&req->rq_list) && rpc_reply_expected(task)) {
894                         /*
895                          * Add to the list only if we're expecting a reply
896                          */
897                         spin_lock_bh(&xprt->transport_lock);
898                         /* Update the softirq receive buffer */
899                         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
900                                         sizeof(req->rq_private_buf));
901                         /* Add request to the receive list */
902                         list_add_tail(&req->rq_list, &xprt->recv);
903                         spin_unlock_bh(&xprt->transport_lock);
904                         xprt_reset_majortimeo(req);
905                         /* Turn off autodisconnect */
906                         del_singleshot_timer_sync(&xprt->timer);
907                 }
908         } else if (!req->rq_bytes_sent)
909                 return;
910
911         req->rq_connect_cookie = xprt->connect_cookie;
912         req->rq_xtime = ktime_get();
913         status = xprt->ops->send_request(task);
914         if (status != 0) {
915                 task->tk_status = status;
916                 return;
917         }
918
919         dprintk("RPC: %5u xmit complete\n", task->tk_pid);
920         task->tk_flags |= RPC_TASK_SENT;
921         spin_lock_bh(&xprt->transport_lock);
922
923         xprt->ops->set_retrans_timeout(task);
924
925         numreqs = atomic_read(&xprt->num_reqs);
926         if (numreqs > xprt->stat.max_slots)
927                 xprt->stat.max_slots = numreqs;
928         xprt->stat.sends++;
929         xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
930         xprt->stat.bklog_u += xprt->backlog.qlen;
931         xprt->stat.sending_u += xprt->sending.qlen;
932         xprt->stat.pending_u += xprt->pending.qlen;
933
934         /* Don't race with disconnect */
935         if (!xprt_connected(xprt))
936                 task->tk_status = -ENOTCONN;
937         else if (!req->rq_reply_bytes_recvd && rpc_reply_expected(task)) {
938                 /*
939                  * Sleep on the pending queue since
940                  * we're expecting a reply.
941                  */
942                 rpc_sleep_on(&xprt->pending, task, xprt_timer);
943         }
944         spin_unlock_bh(&xprt->transport_lock);
945 }
946
947 static struct rpc_rqst *xprt_dynamic_alloc_slot(struct rpc_xprt *xprt, gfp_t gfp_flags)
948 {
949         struct rpc_rqst *req = ERR_PTR(-EAGAIN);
950
951         if (!atomic_add_unless(&xprt->num_reqs, 1, xprt->max_reqs))
952                 goto out;
953         req = kzalloc(sizeof(struct rpc_rqst), gfp_flags);
954         if (req != NULL)
955                 goto out;
956         atomic_dec(&xprt->num_reqs);
957         req = ERR_PTR(-ENOMEM);
958 out:
959         return req;
960 }
961
962 static bool xprt_dynamic_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
963 {
964         if (atomic_add_unless(&xprt->num_reqs, -1, xprt->min_reqs)) {
965                 kfree(req);
966                 return true;
967         }
968         return false;
969 }
970
971 static void xprt_alloc_slot(struct rpc_task *task)
972 {
973         struct rpc_xprt *xprt = task->tk_xprt;
974         struct rpc_rqst *req;
975
976         if (!list_empty(&xprt->free)) {
977                 req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
978                 list_del(&req->rq_list);
979                 goto out_init_req;
980         }
981         req = xprt_dynamic_alloc_slot(xprt, GFP_NOWAIT);
982         if (!IS_ERR(req))
983                 goto out_init_req;
984         switch (PTR_ERR(req)) {
985         case -ENOMEM:
986                 rpc_delay(task, HZ >> 2);
987                 dprintk("RPC:       dynamic allocation of request slot "
988                                 "failed! Retrying\n");
989                 break;
990         case -EAGAIN:
991                 rpc_sleep_on(&xprt->backlog, task, NULL);
992                 dprintk("RPC:       waiting for request slot\n");
993         }
994         task->tk_status = -EAGAIN;
995         return;
996 out_init_req:
997         task->tk_status = 0;
998         task->tk_rqstp = req;
999         xprt_request_init(task, xprt);
1000 }
1001
1002 static void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1003 {
1004         spin_lock(&xprt->reserve_lock);
1005         if (!xprt_dynamic_free_slot(xprt, req)) {
1006                 memset(req, 0, sizeof(*req));   /* mark unused */
1007                 list_add(&req->rq_list, &xprt->free);
1008         }
1009         rpc_wake_up_next(&xprt->backlog);
1010         spin_unlock(&xprt->reserve_lock);
1011 }
1012
1013 static void xprt_free_all_slots(struct rpc_xprt *xprt)
1014 {
1015         struct rpc_rqst *req;
1016         while (!list_empty(&xprt->free)) {
1017                 req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list);
1018                 list_del(&req->rq_list);
1019                 kfree(req);
1020         }
1021 }
1022
1023 struct rpc_xprt *xprt_alloc(struct net *net, size_t size,
1024                 unsigned int num_prealloc,
1025                 unsigned int max_alloc)
1026 {
1027         struct rpc_xprt *xprt;
1028         struct rpc_rqst *req;
1029         int i;
1030
1031         xprt = kzalloc(size, GFP_KERNEL);
1032         if (xprt == NULL)
1033                 goto out;
1034
1035         xprt_init(xprt, net);
1036
1037         for (i = 0; i < num_prealloc; i++) {
1038                 req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL);
1039                 if (!req)
1040                         break;
1041                 list_add(&req->rq_list, &xprt->free);
1042         }
1043         if (i < num_prealloc)
1044                 goto out_free;
1045         if (max_alloc > num_prealloc)
1046                 xprt->max_reqs = max_alloc;
1047         else
1048                 xprt->max_reqs = num_prealloc;
1049         xprt->min_reqs = num_prealloc;
1050         atomic_set(&xprt->num_reqs, num_prealloc);
1051
1052         return xprt;
1053
1054 out_free:
1055         xprt_free(xprt);
1056 out:
1057         return NULL;
1058 }
1059 EXPORT_SYMBOL_GPL(xprt_alloc);
1060
1061 void xprt_free(struct rpc_xprt *xprt)
1062 {
1063         put_net(xprt->xprt_net);
1064         xprt_free_all_slots(xprt);
1065         kfree(xprt);
1066 }
1067 EXPORT_SYMBOL_GPL(xprt_free);
1068
1069 /**
1070  * xprt_reserve - allocate an RPC request slot
1071  * @task: RPC task requesting a slot allocation
1072  *
1073  * If no more slots are available, place the task on the transport's
1074  * backlog queue.
1075  */
1076 void xprt_reserve(struct rpc_task *task)
1077 {
1078         struct rpc_xprt *xprt = task->tk_xprt;
1079
1080         task->tk_status = 0;
1081         if (task->tk_rqstp != NULL)
1082                 return;
1083
1084         /* Note: grabbing the xprt_lock_write() here is not strictly needed,
1085          * but ensures that we throttle new slot allocation if the transport
1086          * is congested (e.g. if reconnecting or if we're out of socket
1087          * write buffer space).
1088          */
1089         task->tk_timeout = 0;
1090         task->tk_status = -EAGAIN;
1091         if (!xprt_lock_write(xprt, task))
1092                 return;
1093
1094         spin_lock(&xprt->reserve_lock);
1095         xprt_alloc_slot(task);
1096         spin_unlock(&xprt->reserve_lock);
1097         xprt_release_write(xprt, task);
1098 }
1099
1100 static inline __be32 xprt_alloc_xid(struct rpc_xprt *xprt)
1101 {
1102         return (__force __be32)xprt->xid++;
1103 }
1104
1105 static inline void xprt_init_xid(struct rpc_xprt *xprt)
1106 {
1107         xprt->xid = net_random();
1108 }
1109
1110 static void xprt_request_init(struct rpc_task *task, struct rpc_xprt *xprt)
1111 {
1112         struct rpc_rqst *req = task->tk_rqstp;
1113
1114         INIT_LIST_HEAD(&req->rq_list);
1115         req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1116         req->rq_task    = task;
1117         req->rq_xprt    = xprt;
1118         req->rq_buffer  = NULL;
1119         req->rq_xid     = xprt_alloc_xid(xprt);
1120         req->rq_release_snd_buf = NULL;
1121         xprt_reset_majortimeo(req);
1122         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1123                         req, ntohl(req->rq_xid));
1124 }
1125
1126 /**
1127  * xprt_release - release an RPC request slot
1128  * @task: task which is finished with the slot
1129  *
1130  */
1131 void xprt_release(struct rpc_task *task)
1132 {
1133         struct rpc_xprt *xprt;
1134         struct rpc_rqst *req;
1135
1136         if (!(req = task->tk_rqstp))
1137                 return;
1138
1139         xprt = req->rq_xprt;
1140         rpc_count_iostats(task);
1141         spin_lock_bh(&xprt->transport_lock);
1142         xprt->ops->release_xprt(xprt, task);
1143         if (xprt->ops->release_request)
1144                 xprt->ops->release_request(task);
1145         if (!list_empty(&req->rq_list))
1146                 list_del(&req->rq_list);
1147         xprt->last_used = jiffies;
1148         if (list_empty(&xprt->recv) && xprt_has_timer(xprt))
1149                 mod_timer(&xprt->timer,
1150                                 xprt->last_used + xprt->idle_timeout);
1151         spin_unlock_bh(&xprt->transport_lock);
1152         if (req->rq_buffer)
1153                 xprt->ops->buf_free(req->rq_buffer);
1154         if (req->rq_cred != NULL)
1155                 put_rpccred(req->rq_cred);
1156         task->tk_rqstp = NULL;
1157         if (req->rq_release_snd_buf)
1158                 req->rq_release_snd_buf(req);
1159
1160         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1161         if (likely(!bc_prealloc(req)))
1162                 xprt_free_slot(xprt, req);
1163         else
1164                 xprt_free_bc_request(req);
1165 }
1166
1167 static void xprt_init(struct rpc_xprt *xprt, struct net *net)
1168 {
1169         atomic_set(&xprt->count, 1);
1170
1171         spin_lock_init(&xprt->transport_lock);
1172         spin_lock_init(&xprt->reserve_lock);
1173
1174         INIT_LIST_HEAD(&xprt->free);
1175         INIT_LIST_HEAD(&xprt->recv);
1176 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1177         spin_lock_init(&xprt->bc_pa_lock);
1178         INIT_LIST_HEAD(&xprt->bc_pa_list);
1179 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1180
1181         xprt->last_used = jiffies;
1182         xprt->cwnd = RPC_INITCWND;
1183         xprt->bind_index = 0;
1184
1185         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1186         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1187         rpc_init_priority_wait_queue(&xprt->sending, "xprt_sending");
1188         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1189
1190         xprt_init_xid(xprt);
1191
1192         xprt->xprt_net = get_net(net);
1193 }
1194
1195 /**
1196  * xprt_create_transport - create an RPC transport
1197  * @args: rpc transport creation arguments
1198  *
1199  */
1200 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1201 {
1202         struct rpc_xprt *xprt;
1203         struct xprt_class *t;
1204
1205         spin_lock(&xprt_list_lock);
1206         list_for_each_entry(t, &xprt_list, list) {
1207                 if (t->ident == args->ident) {
1208                         spin_unlock(&xprt_list_lock);
1209                         goto found;
1210                 }
1211         }
1212         spin_unlock(&xprt_list_lock);
1213         printk(KERN_ERR "RPC: transport (%d) not supported\n", args->ident);
1214         return ERR_PTR(-EIO);
1215
1216 found:
1217         xprt = t->setup(args);
1218         if (IS_ERR(xprt)) {
1219                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1220                                 -PTR_ERR(xprt));
1221                 goto out;
1222         }
1223         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1224         if (xprt_has_timer(xprt))
1225                 setup_timer(&xprt->timer, xprt_init_autodisconnect,
1226                             (unsigned long)xprt);
1227         else
1228                 init_timer(&xprt->timer);
1229         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1230                         xprt->max_reqs);
1231 out:
1232         return xprt;
1233 }
1234
1235 /**
1236  * xprt_destroy - destroy an RPC transport, killing off all requests.
1237  * @xprt: transport to destroy
1238  *
1239  */
1240 static void xprt_destroy(struct rpc_xprt *xprt)
1241 {
1242         dprintk("RPC:       destroying transport %p\n", xprt);
1243         xprt->shutdown = 1;
1244         del_timer_sync(&xprt->timer);
1245
1246         rpc_destroy_wait_queue(&xprt->binding);
1247         rpc_destroy_wait_queue(&xprt->pending);
1248         rpc_destroy_wait_queue(&xprt->sending);
1249         rpc_destroy_wait_queue(&xprt->backlog);
1250         cancel_work_sync(&xprt->task_cleanup);
1251         /*
1252          * Tear down transport state and free the rpc_xprt
1253          */
1254         xprt->ops->destroy(xprt);
1255 }
1256
1257 /**
1258  * xprt_put - release a reference to an RPC transport.
1259  * @xprt: pointer to the transport
1260  *
1261  */
1262 void xprt_put(struct rpc_xprt *xprt)
1263 {
1264         if (atomic_dec_and_test(&xprt->count))
1265                 xprt_destroy(xprt);
1266 }
1267
1268 /**
1269  * xprt_get - return a reference to an RPC transport.
1270  * @xprt: pointer to the transport
1271  *
1272  */
1273 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1274 {
1275         if (atomic_inc_not_zero(&xprt->count))
1276                 return xprt;
1277         return NULL;
1278 }