xref: /linux/net/sunrpc/svc_xprt.c (revision 86a8280a7fe007d61b05fa8a352edc0595283dad)
1 /*
2  * linux/net/sunrpc/svc_xprt.c
3  *
4  * Author: Tom Tucker <tom@opengridcomputing.com>
5  */
6 
7 #include <linux/sched.h>
8 #include <linux/errno.h>
9 #include <linux/freezer.h>
10 #include <linux/kthread.h>
11 #include <linux/slab.h>
12 #include <net/sock.h>
13 #include <linux/sunrpc/addr.h>
14 #include <linux/sunrpc/stats.h>
15 #include <linux/sunrpc/svc_xprt.h>
16 #include <linux/sunrpc/svcsock.h>
17 #include <linux/sunrpc/xprt.h>
18 #include <linux/module.h>
19 #include <linux/netdevice.h>
20 #include <trace/events/sunrpc.h>
21 
22 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
23 
24 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
25 static int svc_deferred_recv(struct svc_rqst *rqstp);
26 static struct cache_deferred_req *svc_defer(struct cache_req *req);
27 static void svc_age_temp_xprts(unsigned long closure);
28 static void svc_delete_xprt(struct svc_xprt *xprt);
29 
30 /* apparently the "standard" is that clients close
31  * idle connections after 5 minutes, servers after
32  * 6 minutes
33  *   http://www.connectathon.org/talks96/nfstcp.pdf
34  */
35 static int svc_conn_age_period = 6*60;
36 
37 /* List of registered transport classes */
38 static DEFINE_SPINLOCK(svc_xprt_class_lock);
39 static LIST_HEAD(svc_xprt_class_list);
40 
41 /* SMP locking strategy:
42  *
43  *	svc_pool->sp_lock protects most of the fields of that pool.
44  *	svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
45  *	when both need to be taken (rare), svc_serv->sv_lock is first.
46  *	The "service mutex" protects svc_serv->sv_nrthread.
47  *	svc_sock->sk_lock protects the svc_sock->sk_deferred list
48  *             and the ->sk_info_authunix cache.
49  *
50  *	The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
51  *	enqueued multiply. During normal transport processing this bit
52  *	is set by svc_xprt_enqueue and cleared by svc_xprt_received.
53  *	Providers should not manipulate this bit directly.
54  *
55  *	Some flags can be set to certain values at any time
56  *	providing that certain rules are followed:
57  *
58  *	XPT_CONN, XPT_DATA:
59  *		- Can be set or cleared at any time.
60  *		- After a set, svc_xprt_enqueue must be called to enqueue
61  *		  the transport for processing.
62  *		- After a clear, the transport must be read/accepted.
63  *		  If this succeeds, it must be set again.
64  *	XPT_CLOSE:
65  *		- Can set at any time. It is never cleared.
66  *      XPT_DEAD:
67  *		- Can only be set while XPT_BUSY is held which ensures
68  *		  that no other thread will be using the transport or will
69  *		  try to set XPT_DEAD.
70  */
71 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
72 {
73 	struct svc_xprt_class *cl;
74 	int res = -EEXIST;
75 
76 	dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
77 
78 	INIT_LIST_HEAD(&xcl->xcl_list);
79 	spin_lock(&svc_xprt_class_lock);
80 	/* Make sure there isn't already a class with the same name */
81 	list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
82 		if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
83 			goto out;
84 	}
85 	list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
86 	res = 0;
87 out:
88 	spin_unlock(&svc_xprt_class_lock);
89 	return res;
90 }
91 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
92 
93 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
94 {
95 	dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
96 	spin_lock(&svc_xprt_class_lock);
97 	list_del_init(&xcl->xcl_list);
98 	spin_unlock(&svc_xprt_class_lock);
99 }
100 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
101 
102 /*
103  * Format the transport list for printing
104  */
105 int svc_print_xprts(char *buf, int maxlen)
106 {
107 	struct svc_xprt_class *xcl;
108 	char tmpstr[80];
109 	int len = 0;
110 	buf[0] = '\0';
111 
112 	spin_lock(&svc_xprt_class_lock);
113 	list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
114 		int slen;
115 
116 		sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
117 		slen = strlen(tmpstr);
118 		if (len + slen > maxlen)
119 			break;
120 		len += slen;
121 		strcat(buf, tmpstr);
122 	}
123 	spin_unlock(&svc_xprt_class_lock);
124 
125 	return len;
126 }
127 
128 static void svc_xprt_free(struct kref *kref)
129 {
130 	struct svc_xprt *xprt =
131 		container_of(kref, struct svc_xprt, xpt_ref);
132 	struct module *owner = xprt->xpt_class->xcl_owner;
133 	if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
134 		svcauth_unix_info_release(xprt);
135 	put_net(xprt->xpt_net);
136 	/* See comment on corresponding get in xs_setup_bc_tcp(): */
137 	if (xprt->xpt_bc_xprt)
138 		xprt_put(xprt->xpt_bc_xprt);
139 	xprt->xpt_ops->xpo_free(xprt);
140 	module_put(owner);
141 }
142 
143 void svc_xprt_put(struct svc_xprt *xprt)
144 {
145 	kref_put(&xprt->xpt_ref, svc_xprt_free);
146 }
147 EXPORT_SYMBOL_GPL(svc_xprt_put);
148 
149 /*
150  * Called by transport drivers to initialize the transport independent
151  * portion of the transport instance.
152  */
153 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
154 		   struct svc_xprt *xprt, struct svc_serv *serv)
155 {
156 	memset(xprt, 0, sizeof(*xprt));
157 	xprt->xpt_class = xcl;
158 	xprt->xpt_ops = xcl->xcl_ops;
159 	kref_init(&xprt->xpt_ref);
160 	xprt->xpt_server = serv;
161 	INIT_LIST_HEAD(&xprt->xpt_list);
162 	INIT_LIST_HEAD(&xprt->xpt_ready);
163 	INIT_LIST_HEAD(&xprt->xpt_deferred);
164 	INIT_LIST_HEAD(&xprt->xpt_users);
165 	mutex_init(&xprt->xpt_mutex);
166 	spin_lock_init(&xprt->xpt_lock);
167 	set_bit(XPT_BUSY, &xprt->xpt_flags);
168 	rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
169 	xprt->xpt_net = get_net(net);
170 }
171 EXPORT_SYMBOL_GPL(svc_xprt_init);
172 
173 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
174 					 struct svc_serv *serv,
175 					 struct net *net,
176 					 const int family,
177 					 const unsigned short port,
178 					 int flags)
179 {
180 	struct sockaddr_in sin = {
181 		.sin_family		= AF_INET,
182 		.sin_addr.s_addr	= htonl(INADDR_ANY),
183 		.sin_port		= htons(port),
184 	};
185 #if IS_ENABLED(CONFIG_IPV6)
186 	struct sockaddr_in6 sin6 = {
187 		.sin6_family		= AF_INET6,
188 		.sin6_addr		= IN6ADDR_ANY_INIT,
189 		.sin6_port		= htons(port),
190 	};
191 #endif
192 	struct sockaddr *sap;
193 	size_t len;
194 
195 	switch (family) {
196 	case PF_INET:
197 		sap = (struct sockaddr *)&sin;
198 		len = sizeof(sin);
199 		break;
200 #if IS_ENABLED(CONFIG_IPV6)
201 	case PF_INET6:
202 		sap = (struct sockaddr *)&sin6;
203 		len = sizeof(sin6);
204 		break;
205 #endif
206 	default:
207 		return ERR_PTR(-EAFNOSUPPORT);
208 	}
209 
210 	return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
211 }
212 
213 /*
214  * svc_xprt_received conditionally queues the transport for processing
215  * by another thread. The caller must hold the XPT_BUSY bit and must
216  * not thereafter touch transport data.
217  *
218  * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
219  * insufficient) data.
220  */
221 static void svc_xprt_received(struct svc_xprt *xprt)
222 {
223 	if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
224 		WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
225 		return;
226 	}
227 
228 	/* As soon as we clear busy, the xprt could be closed and
229 	 * 'put', so we need a reference to call svc_enqueue_xprt with:
230 	 */
231 	svc_xprt_get(xprt);
232 	smp_mb__before_atomic();
233 	clear_bit(XPT_BUSY, &xprt->xpt_flags);
234 	xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
235 	svc_xprt_put(xprt);
236 }
237 
238 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
239 {
240 	clear_bit(XPT_TEMP, &new->xpt_flags);
241 	spin_lock_bh(&serv->sv_lock);
242 	list_add(&new->xpt_list, &serv->sv_permsocks);
243 	spin_unlock_bh(&serv->sv_lock);
244 	svc_xprt_received(new);
245 }
246 
247 int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
248 		    struct net *net, const int family,
249 		    const unsigned short port, int flags)
250 {
251 	struct svc_xprt_class *xcl;
252 
253 	spin_lock(&svc_xprt_class_lock);
254 	list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
255 		struct svc_xprt *newxprt;
256 		unsigned short newport;
257 
258 		if (strcmp(xprt_name, xcl->xcl_name))
259 			continue;
260 
261 		if (!try_module_get(xcl->xcl_owner))
262 			goto err;
263 
264 		spin_unlock(&svc_xprt_class_lock);
265 		newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
266 		if (IS_ERR(newxprt)) {
267 			module_put(xcl->xcl_owner);
268 			return PTR_ERR(newxprt);
269 		}
270 		svc_add_new_perm_xprt(serv, newxprt);
271 		newport = svc_xprt_local_port(newxprt);
272 		return newport;
273 	}
274  err:
275 	spin_unlock(&svc_xprt_class_lock);
276 	/* This errno is exposed to user space.  Provide a reasonable
277 	 * perror msg for a bad transport. */
278 	return -EPROTONOSUPPORT;
279 }
280 
281 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
282 		    struct net *net, const int family,
283 		    const unsigned short port, int flags)
284 {
285 	int err;
286 
287 	dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
288 	err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
289 	if (err == -EPROTONOSUPPORT) {
290 		request_module("svc%s", xprt_name);
291 		err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
292 	}
293 	if (err)
294 		dprintk("svc: transport %s not found, err %d\n",
295 			xprt_name, err);
296 	return err;
297 }
298 EXPORT_SYMBOL_GPL(svc_create_xprt);
299 
300 /*
301  * Copy the local and remote xprt addresses to the rqstp structure
302  */
303 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
304 {
305 	memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
306 	rqstp->rq_addrlen = xprt->xpt_remotelen;
307 
308 	/*
309 	 * Destination address in request is needed for binding the
310 	 * source address in RPC replies/callbacks later.
311 	 */
312 	memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
313 	rqstp->rq_daddrlen = xprt->xpt_locallen;
314 }
315 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
316 
317 /**
318  * svc_print_addr - Format rq_addr field for printing
319  * @rqstp: svc_rqst struct containing address to print
320  * @buf: target buffer for formatted address
321  * @len: length of target buffer
322  *
323  */
324 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
325 {
326 	return __svc_print_addr(svc_addr(rqstp), buf, len);
327 }
328 EXPORT_SYMBOL_GPL(svc_print_addr);
329 
330 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
331 {
332 	if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
333 		return true;
334 	if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
335 		return xprt->xpt_ops->xpo_has_wspace(xprt);
336 	return false;
337 }
338 
339 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
340 {
341 	struct svc_pool *pool;
342 	struct svc_rqst	*rqstp = NULL;
343 	int cpu;
344 	bool queued = false;
345 
346 	if (!svc_xprt_has_something_to_do(xprt))
347 		goto out;
348 
349 	/* Mark transport as busy. It will remain in this state until
350 	 * the provider calls svc_xprt_received. We update XPT_BUSY
351 	 * atomically because it also guards against trying to enqueue
352 	 * the transport twice.
353 	 */
354 	if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
355 		/* Don't enqueue transport while already enqueued */
356 		dprintk("svc: transport %p busy, not enqueued\n", xprt);
357 		goto out;
358 	}
359 
360 	cpu = get_cpu();
361 	pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
362 
363 	atomic_long_inc(&pool->sp_stats.packets);
364 
365 redo_search:
366 	/* find a thread for this xprt */
367 	rcu_read_lock();
368 	list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
369 		/* Do a lockless check first */
370 		if (test_bit(RQ_BUSY, &rqstp->rq_flags))
371 			continue;
372 
373 		/*
374 		 * Once the xprt has been queued, it can only be dequeued by
375 		 * the task that intends to service it. All we can do at that
376 		 * point is to try to wake this thread back up so that it can
377 		 * do so.
378 		 */
379 		if (!queued) {
380 			spin_lock_bh(&rqstp->rq_lock);
381 			if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags)) {
382 				/* already busy, move on... */
383 				spin_unlock_bh(&rqstp->rq_lock);
384 				continue;
385 			}
386 
387 			/* this one will do */
388 			rqstp->rq_xprt = xprt;
389 			svc_xprt_get(xprt);
390 			spin_unlock_bh(&rqstp->rq_lock);
391 		}
392 		rcu_read_unlock();
393 
394 		atomic_long_inc(&pool->sp_stats.threads_woken);
395 		wake_up_process(rqstp->rq_task);
396 		put_cpu();
397 		goto out;
398 	}
399 	rcu_read_unlock();
400 
401 	/*
402 	 * We didn't find an idle thread to use, so we need to queue the xprt.
403 	 * Do so and then search again. If we find one, we can't hook this one
404 	 * up to it directly but we can wake the thread up in the hopes that it
405 	 * will pick it up once it searches for a xprt to service.
406 	 */
407 	if (!queued) {
408 		queued = true;
409 		dprintk("svc: transport %p put into queue\n", xprt);
410 		spin_lock_bh(&pool->sp_lock);
411 		list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
412 		pool->sp_stats.sockets_queued++;
413 		spin_unlock_bh(&pool->sp_lock);
414 		goto redo_search;
415 	}
416 	rqstp = NULL;
417 	put_cpu();
418 out:
419 	trace_svc_xprt_do_enqueue(xprt, rqstp);
420 }
421 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
422 
423 /*
424  * Queue up a transport with data pending. If there are idle nfsd
425  * processes, wake 'em up.
426  *
427  */
428 void svc_xprt_enqueue(struct svc_xprt *xprt)
429 {
430 	if (test_bit(XPT_BUSY, &xprt->xpt_flags))
431 		return;
432 	xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
433 }
434 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
435 
436 /*
437  * Dequeue the first transport, if there is one.
438  */
439 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
440 {
441 	struct svc_xprt	*xprt = NULL;
442 
443 	if (list_empty(&pool->sp_sockets))
444 		goto out;
445 
446 	spin_lock_bh(&pool->sp_lock);
447 	if (likely(!list_empty(&pool->sp_sockets))) {
448 		xprt = list_first_entry(&pool->sp_sockets,
449 					struct svc_xprt, xpt_ready);
450 		list_del_init(&xprt->xpt_ready);
451 		svc_xprt_get(xprt);
452 
453 		dprintk("svc: transport %p dequeued, inuse=%d\n",
454 			xprt, atomic_read(&xprt->xpt_ref.refcount));
455 	}
456 	spin_unlock_bh(&pool->sp_lock);
457 out:
458 	trace_svc_xprt_dequeue(xprt);
459 	return xprt;
460 }
461 
462 /**
463  * svc_reserve - change the space reserved for the reply to a request.
464  * @rqstp:  The request in question
465  * @space: new max space to reserve
466  *
467  * Each request reserves some space on the output queue of the transport
468  * to make sure the reply fits.  This function reduces that reserved
469  * space to be the amount of space used already, plus @space.
470  *
471  */
472 void svc_reserve(struct svc_rqst *rqstp, int space)
473 {
474 	space += rqstp->rq_res.head[0].iov_len;
475 
476 	if (space < rqstp->rq_reserved) {
477 		struct svc_xprt *xprt = rqstp->rq_xprt;
478 		atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
479 		rqstp->rq_reserved = space;
480 
481 		if (xprt->xpt_ops->xpo_adjust_wspace)
482 			xprt->xpt_ops->xpo_adjust_wspace(xprt);
483 		svc_xprt_enqueue(xprt);
484 	}
485 }
486 EXPORT_SYMBOL_GPL(svc_reserve);
487 
488 static void svc_xprt_release(struct svc_rqst *rqstp)
489 {
490 	struct svc_xprt	*xprt = rqstp->rq_xprt;
491 
492 	rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
493 
494 	kfree(rqstp->rq_deferred);
495 	rqstp->rq_deferred = NULL;
496 
497 	svc_free_res_pages(rqstp);
498 	rqstp->rq_res.page_len = 0;
499 	rqstp->rq_res.page_base = 0;
500 
501 	/* Reset response buffer and release
502 	 * the reservation.
503 	 * But first, check that enough space was reserved
504 	 * for the reply, otherwise we have a bug!
505 	 */
506 	if ((rqstp->rq_res.len) >  rqstp->rq_reserved)
507 		printk(KERN_ERR "RPC request reserved %d but used %d\n",
508 		       rqstp->rq_reserved,
509 		       rqstp->rq_res.len);
510 
511 	rqstp->rq_res.head[0].iov_len = 0;
512 	svc_reserve(rqstp, 0);
513 	rqstp->rq_xprt = NULL;
514 
515 	svc_xprt_put(xprt);
516 }
517 
518 /*
519  * Some svc_serv's will have occasional work to do, even when a xprt is not
520  * waiting to be serviced. This function is there to "kick" a task in one of
521  * those services so that it can wake up and do that work. Note that we only
522  * bother with pool 0 as we don't need to wake up more than one thread for
523  * this purpose.
524  */
525 void svc_wake_up(struct svc_serv *serv)
526 {
527 	struct svc_rqst	*rqstp;
528 	struct svc_pool *pool;
529 
530 	pool = &serv->sv_pools[0];
531 
532 	rcu_read_lock();
533 	list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
534 		/* skip any that aren't queued */
535 		if (test_bit(RQ_BUSY, &rqstp->rq_flags))
536 			continue;
537 		rcu_read_unlock();
538 		dprintk("svc: daemon %p woken up.\n", rqstp);
539 		wake_up_process(rqstp->rq_task);
540 		trace_svc_wake_up(rqstp->rq_task->pid);
541 		return;
542 	}
543 	rcu_read_unlock();
544 
545 	/* No free entries available */
546 	set_bit(SP_TASK_PENDING, &pool->sp_flags);
547 	smp_wmb();
548 	trace_svc_wake_up(0);
549 }
550 EXPORT_SYMBOL_GPL(svc_wake_up);
551 
552 int svc_port_is_privileged(struct sockaddr *sin)
553 {
554 	switch (sin->sa_family) {
555 	case AF_INET:
556 		return ntohs(((struct sockaddr_in *)sin)->sin_port)
557 			< PROT_SOCK;
558 	case AF_INET6:
559 		return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
560 			< PROT_SOCK;
561 	default:
562 		return 0;
563 	}
564 }
565 
566 /*
567  * Make sure that we don't have too many active connections. If we have,
568  * something must be dropped. It's not clear what will happen if we allow
569  * "too many" connections, but when dealing with network-facing software,
570  * we have to code defensively. Here we do that by imposing hard limits.
571  *
572  * There's no point in trying to do random drop here for DoS
573  * prevention. The NFS clients does 1 reconnect in 15 seconds. An
574  * attacker can easily beat that.
575  *
576  * The only somewhat efficient mechanism would be if drop old
577  * connections from the same IP first. But right now we don't even
578  * record the client IP in svc_sock.
579  *
580  * single-threaded services that expect a lot of clients will probably
581  * need to set sv_maxconn to override the default value which is based
582  * on the number of threads
583  */
584 static void svc_check_conn_limits(struct svc_serv *serv)
585 {
586 	unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
587 				(serv->sv_nrthreads+3) * 20;
588 
589 	if (serv->sv_tmpcnt > limit) {
590 		struct svc_xprt *xprt = NULL;
591 		spin_lock_bh(&serv->sv_lock);
592 		if (!list_empty(&serv->sv_tempsocks)) {
593 			/* Try to help the admin */
594 			net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
595 					       serv->sv_name, serv->sv_maxconn ?
596 					       "max number of connections" :
597 					       "number of threads");
598 			/*
599 			 * Always select the oldest connection. It's not fair,
600 			 * but so is life
601 			 */
602 			xprt = list_entry(serv->sv_tempsocks.prev,
603 					  struct svc_xprt,
604 					  xpt_list);
605 			set_bit(XPT_CLOSE, &xprt->xpt_flags);
606 			svc_xprt_get(xprt);
607 		}
608 		spin_unlock_bh(&serv->sv_lock);
609 
610 		if (xprt) {
611 			svc_xprt_enqueue(xprt);
612 			svc_xprt_put(xprt);
613 		}
614 	}
615 }
616 
617 static int svc_alloc_arg(struct svc_rqst *rqstp)
618 {
619 	struct svc_serv *serv = rqstp->rq_server;
620 	struct xdr_buf *arg;
621 	int pages;
622 	int i;
623 
624 	/* now allocate needed pages.  If we get a failure, sleep briefly */
625 	pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
626 	WARN_ON_ONCE(pages >= RPCSVC_MAXPAGES);
627 	if (pages >= RPCSVC_MAXPAGES)
628 		/* use as many pages as possible */
629 		pages = RPCSVC_MAXPAGES - 1;
630 	for (i = 0; i < pages ; i++)
631 		while (rqstp->rq_pages[i] == NULL) {
632 			struct page *p = alloc_page(GFP_KERNEL);
633 			if (!p) {
634 				set_current_state(TASK_INTERRUPTIBLE);
635 				if (signalled() || kthread_should_stop()) {
636 					set_current_state(TASK_RUNNING);
637 					return -EINTR;
638 				}
639 				schedule_timeout(msecs_to_jiffies(500));
640 			}
641 			rqstp->rq_pages[i] = p;
642 		}
643 	rqstp->rq_page_end = &rqstp->rq_pages[i];
644 	rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
645 
646 	/* Make arg->head point to first page and arg->pages point to rest */
647 	arg = &rqstp->rq_arg;
648 	arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
649 	arg->head[0].iov_len = PAGE_SIZE;
650 	arg->pages = rqstp->rq_pages + 1;
651 	arg->page_base = 0;
652 	/* save at least one page for response */
653 	arg->page_len = (pages-2)*PAGE_SIZE;
654 	arg->len = (pages-1)*PAGE_SIZE;
655 	arg->tail[0].iov_len = 0;
656 	return 0;
657 }
658 
659 static bool
660 rqst_should_sleep(struct svc_rqst *rqstp)
661 {
662 	struct svc_pool		*pool = rqstp->rq_pool;
663 
664 	/* did someone call svc_wake_up? */
665 	if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
666 		return false;
667 
668 	/* was a socket queued? */
669 	if (!list_empty(&pool->sp_sockets))
670 		return false;
671 
672 	/* are we shutting down? */
673 	if (signalled() || kthread_should_stop())
674 		return false;
675 
676 	/* are we freezing? */
677 	if (freezing(current))
678 		return false;
679 
680 	return true;
681 }
682 
683 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
684 {
685 	struct svc_xprt *xprt;
686 	struct svc_pool		*pool = rqstp->rq_pool;
687 	long			time_left = 0;
688 
689 	/* rq_xprt should be clear on entry */
690 	WARN_ON_ONCE(rqstp->rq_xprt);
691 
692 	/* Normally we will wait up to 5 seconds for any required
693 	 * cache information to be provided.
694 	 */
695 	rqstp->rq_chandle.thread_wait = 5*HZ;
696 
697 	xprt = svc_xprt_dequeue(pool);
698 	if (xprt) {
699 		rqstp->rq_xprt = xprt;
700 
701 		/* As there is a shortage of threads and this request
702 		 * had to be queued, don't allow the thread to wait so
703 		 * long for cache updates.
704 		 */
705 		rqstp->rq_chandle.thread_wait = 1*HZ;
706 		clear_bit(SP_TASK_PENDING, &pool->sp_flags);
707 		return xprt;
708 	}
709 
710 	/*
711 	 * We have to be able to interrupt this wait
712 	 * to bring down the daemons ...
713 	 */
714 	set_current_state(TASK_INTERRUPTIBLE);
715 	clear_bit(RQ_BUSY, &rqstp->rq_flags);
716 	smp_mb();
717 
718 	if (likely(rqst_should_sleep(rqstp)))
719 		time_left = schedule_timeout(timeout);
720 	else
721 		__set_current_state(TASK_RUNNING);
722 
723 	try_to_freeze();
724 
725 	spin_lock_bh(&rqstp->rq_lock);
726 	set_bit(RQ_BUSY, &rqstp->rq_flags);
727 	spin_unlock_bh(&rqstp->rq_lock);
728 
729 	xprt = rqstp->rq_xprt;
730 	if (xprt != NULL)
731 		return xprt;
732 
733 	if (!time_left)
734 		atomic_long_inc(&pool->sp_stats.threads_timedout);
735 
736 	if (signalled() || kthread_should_stop())
737 		return ERR_PTR(-EINTR);
738 	return ERR_PTR(-EAGAIN);
739 }
740 
741 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
742 {
743 	spin_lock_bh(&serv->sv_lock);
744 	set_bit(XPT_TEMP, &newxpt->xpt_flags);
745 	list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
746 	serv->sv_tmpcnt++;
747 	if (serv->sv_temptimer.function == NULL) {
748 		/* setup timer to age temp transports */
749 		setup_timer(&serv->sv_temptimer, svc_age_temp_xprts,
750 			    (unsigned long)serv);
751 		mod_timer(&serv->sv_temptimer,
752 			  jiffies + svc_conn_age_period * HZ);
753 	}
754 	spin_unlock_bh(&serv->sv_lock);
755 	svc_xprt_received(newxpt);
756 }
757 
758 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
759 {
760 	struct svc_serv *serv = rqstp->rq_server;
761 	int len = 0;
762 
763 	if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
764 		dprintk("svc_recv: found XPT_CLOSE\n");
765 		svc_delete_xprt(xprt);
766 		/* Leave XPT_BUSY set on the dead xprt: */
767 		goto out;
768 	}
769 	if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
770 		struct svc_xprt *newxpt;
771 		/*
772 		 * We know this module_get will succeed because the
773 		 * listener holds a reference too
774 		 */
775 		__module_get(xprt->xpt_class->xcl_owner);
776 		svc_check_conn_limits(xprt->xpt_server);
777 		newxpt = xprt->xpt_ops->xpo_accept(xprt);
778 		if (newxpt)
779 			svc_add_new_temp_xprt(serv, newxpt);
780 		else
781 			module_put(xprt->xpt_class->xcl_owner);
782 	} else {
783 		/* XPT_DATA|XPT_DEFERRED case: */
784 		dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
785 			rqstp, rqstp->rq_pool->sp_id, xprt,
786 			atomic_read(&xprt->xpt_ref.refcount));
787 		rqstp->rq_deferred = svc_deferred_dequeue(xprt);
788 		if (rqstp->rq_deferred)
789 			len = svc_deferred_recv(rqstp);
790 		else
791 			len = xprt->xpt_ops->xpo_recvfrom(rqstp);
792 		dprintk("svc: got len=%d\n", len);
793 		rqstp->rq_reserved = serv->sv_max_mesg;
794 		atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
795 	}
796 	/* clear XPT_BUSY: */
797 	svc_xprt_received(xprt);
798 out:
799 	trace_svc_handle_xprt(xprt, len);
800 	return len;
801 }
802 
803 /*
804  * Receive the next request on any transport.  This code is carefully
805  * organised not to touch any cachelines in the shared svc_serv
806  * structure, only cachelines in the local svc_pool.
807  */
808 int svc_recv(struct svc_rqst *rqstp, long timeout)
809 {
810 	struct svc_xprt		*xprt = NULL;
811 	struct svc_serv		*serv = rqstp->rq_server;
812 	int			len, err;
813 
814 	dprintk("svc: server %p waiting for data (to = %ld)\n",
815 		rqstp, timeout);
816 
817 	if (rqstp->rq_xprt)
818 		printk(KERN_ERR
819 			"svc_recv: service %p, transport not NULL!\n",
820 			 rqstp);
821 
822 	err = svc_alloc_arg(rqstp);
823 	if (err)
824 		goto out;
825 
826 	try_to_freeze();
827 	cond_resched();
828 	err = -EINTR;
829 	if (signalled() || kthread_should_stop())
830 		goto out;
831 
832 	xprt = svc_get_next_xprt(rqstp, timeout);
833 	if (IS_ERR(xprt)) {
834 		err = PTR_ERR(xprt);
835 		goto out;
836 	}
837 
838 	len = svc_handle_xprt(rqstp, xprt);
839 
840 	/* No data, incomplete (TCP) read, or accept() */
841 	err = -EAGAIN;
842 	if (len <= 0)
843 		goto out_release;
844 
845 	clear_bit(XPT_OLD, &xprt->xpt_flags);
846 
847 	if (xprt->xpt_ops->xpo_secure_port(rqstp))
848 		set_bit(RQ_SECURE, &rqstp->rq_flags);
849 	else
850 		clear_bit(RQ_SECURE, &rqstp->rq_flags);
851 	rqstp->rq_chandle.defer = svc_defer;
852 	rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
853 
854 	if (serv->sv_stats)
855 		serv->sv_stats->netcnt++;
856 	trace_svc_recv(rqstp, len);
857 	return len;
858 out_release:
859 	rqstp->rq_res.len = 0;
860 	svc_xprt_release(rqstp);
861 out:
862 	trace_svc_recv(rqstp, err);
863 	return err;
864 }
865 EXPORT_SYMBOL_GPL(svc_recv);
866 
867 /*
868  * Drop request
869  */
870 void svc_drop(struct svc_rqst *rqstp)
871 {
872 	dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
873 	svc_xprt_release(rqstp);
874 }
875 EXPORT_SYMBOL_GPL(svc_drop);
876 
877 /*
878  * Return reply to client.
879  */
880 int svc_send(struct svc_rqst *rqstp)
881 {
882 	struct svc_xprt	*xprt;
883 	int		len = -EFAULT;
884 	struct xdr_buf	*xb;
885 
886 	xprt = rqstp->rq_xprt;
887 	if (!xprt)
888 		goto out;
889 
890 	/* release the receive skb before sending the reply */
891 	rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
892 
893 	/* calculate over-all length */
894 	xb = &rqstp->rq_res;
895 	xb->len = xb->head[0].iov_len +
896 		xb->page_len +
897 		xb->tail[0].iov_len;
898 
899 	/* Grab mutex to serialize outgoing data. */
900 	mutex_lock(&xprt->xpt_mutex);
901 	if (test_bit(XPT_DEAD, &xprt->xpt_flags)
902 			|| test_bit(XPT_CLOSE, &xprt->xpt_flags))
903 		len = -ENOTCONN;
904 	else
905 		len = xprt->xpt_ops->xpo_sendto(rqstp);
906 	mutex_unlock(&xprt->xpt_mutex);
907 	rpc_wake_up(&xprt->xpt_bc_pending);
908 	svc_xprt_release(rqstp);
909 
910 	if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
911 		len = 0;
912 out:
913 	trace_svc_send(rqstp, len);
914 	return len;
915 }
916 
917 /*
918  * Timer function to close old temporary transports, using
919  * a mark-and-sweep algorithm.
920  */
921 static void svc_age_temp_xprts(unsigned long closure)
922 {
923 	struct svc_serv *serv = (struct svc_serv *)closure;
924 	struct svc_xprt *xprt;
925 	struct list_head *le, *next;
926 
927 	dprintk("svc_age_temp_xprts\n");
928 
929 	if (!spin_trylock_bh(&serv->sv_lock)) {
930 		/* busy, try again 1 sec later */
931 		dprintk("svc_age_temp_xprts: busy\n");
932 		mod_timer(&serv->sv_temptimer, jiffies + HZ);
933 		return;
934 	}
935 
936 	list_for_each_safe(le, next, &serv->sv_tempsocks) {
937 		xprt = list_entry(le, struct svc_xprt, xpt_list);
938 
939 		/* First time through, just mark it OLD. Second time
940 		 * through, close it. */
941 		if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
942 			continue;
943 		if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
944 		    test_bit(XPT_BUSY, &xprt->xpt_flags))
945 			continue;
946 		list_del_init(le);
947 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
948 		dprintk("queuing xprt %p for closing\n", xprt);
949 
950 		/* a thread will dequeue and close it soon */
951 		svc_xprt_enqueue(xprt);
952 	}
953 	spin_unlock_bh(&serv->sv_lock);
954 
955 	mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
956 }
957 
958 /* Close temporary transports whose xpt_local matches server_addr immediately
959  * instead of waiting for them to be picked up by the timer.
960  *
961  * This is meant to be called from a notifier_block that runs when an ip
962  * address is deleted.
963  */
964 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
965 {
966 	struct svc_xprt *xprt;
967 	struct svc_sock *svsk;
968 	struct socket *sock;
969 	struct list_head *le, *next;
970 	LIST_HEAD(to_be_closed);
971 	struct linger no_linger = {
972 		.l_onoff = 1,
973 		.l_linger = 0,
974 	};
975 
976 	spin_lock_bh(&serv->sv_lock);
977 	list_for_each_safe(le, next, &serv->sv_tempsocks) {
978 		xprt = list_entry(le, struct svc_xprt, xpt_list);
979 		if (rpc_cmp_addr(server_addr, (struct sockaddr *)
980 				&xprt->xpt_local)) {
981 			dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
982 			list_move(le, &to_be_closed);
983 		}
984 	}
985 	spin_unlock_bh(&serv->sv_lock);
986 
987 	while (!list_empty(&to_be_closed)) {
988 		le = to_be_closed.next;
989 		list_del_init(le);
990 		xprt = list_entry(le, struct svc_xprt, xpt_list);
991 		dprintk("svc_age_temp_xprts_now: closing %p\n", xprt);
992 		svsk = container_of(xprt, struct svc_sock, sk_xprt);
993 		sock = svsk->sk_sock;
994 		kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
995 				  (char *)&no_linger, sizeof(no_linger));
996 		svc_close_xprt(xprt);
997 	}
998 }
999 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1000 
1001 static void call_xpt_users(struct svc_xprt *xprt)
1002 {
1003 	struct svc_xpt_user *u;
1004 
1005 	spin_lock(&xprt->xpt_lock);
1006 	while (!list_empty(&xprt->xpt_users)) {
1007 		u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1008 		list_del(&u->list);
1009 		u->callback(u);
1010 	}
1011 	spin_unlock(&xprt->xpt_lock);
1012 }
1013 
1014 /*
1015  * Remove a dead transport
1016  */
1017 static void svc_delete_xprt(struct svc_xprt *xprt)
1018 {
1019 	struct svc_serv	*serv = xprt->xpt_server;
1020 	struct svc_deferred_req *dr;
1021 
1022 	/* Only do this once */
1023 	if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1024 		BUG();
1025 
1026 	dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1027 	xprt->xpt_ops->xpo_detach(xprt);
1028 
1029 	spin_lock_bh(&serv->sv_lock);
1030 	list_del_init(&xprt->xpt_list);
1031 	WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1032 	if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1033 		serv->sv_tmpcnt--;
1034 	spin_unlock_bh(&serv->sv_lock);
1035 
1036 	while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1037 		kfree(dr);
1038 
1039 	call_xpt_users(xprt);
1040 	svc_xprt_put(xprt);
1041 }
1042 
1043 void svc_close_xprt(struct svc_xprt *xprt)
1044 {
1045 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
1046 	if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1047 		/* someone else will have to effect the close */
1048 		return;
1049 	/*
1050 	 * We expect svc_close_xprt() to work even when no threads are
1051 	 * running (e.g., while configuring the server before starting
1052 	 * any threads), so if the transport isn't busy, we delete
1053 	 * it ourself:
1054 	 */
1055 	svc_delete_xprt(xprt);
1056 }
1057 EXPORT_SYMBOL_GPL(svc_close_xprt);
1058 
1059 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1060 {
1061 	struct svc_xprt *xprt;
1062 	int ret = 0;
1063 
1064 	spin_lock(&serv->sv_lock);
1065 	list_for_each_entry(xprt, xprt_list, xpt_list) {
1066 		if (xprt->xpt_net != net)
1067 			continue;
1068 		ret++;
1069 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
1070 		svc_xprt_enqueue(xprt);
1071 	}
1072 	spin_unlock(&serv->sv_lock);
1073 	return ret;
1074 }
1075 
1076 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1077 {
1078 	struct svc_pool *pool;
1079 	struct svc_xprt *xprt;
1080 	struct svc_xprt *tmp;
1081 	int i;
1082 
1083 	for (i = 0; i < serv->sv_nrpools; i++) {
1084 		pool = &serv->sv_pools[i];
1085 
1086 		spin_lock_bh(&pool->sp_lock);
1087 		list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1088 			if (xprt->xpt_net != net)
1089 				continue;
1090 			list_del_init(&xprt->xpt_ready);
1091 			spin_unlock_bh(&pool->sp_lock);
1092 			return xprt;
1093 		}
1094 		spin_unlock_bh(&pool->sp_lock);
1095 	}
1096 	return NULL;
1097 }
1098 
1099 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1100 {
1101 	struct svc_xprt *xprt;
1102 
1103 	while ((xprt = svc_dequeue_net(serv, net))) {
1104 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
1105 		svc_delete_xprt(xprt);
1106 	}
1107 }
1108 
1109 /*
1110  * Server threads may still be running (especially in the case where the
1111  * service is still running in other network namespaces).
1112  *
1113  * So we shut down sockets the same way we would on a running server, by
1114  * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1115  * the close.  In the case there are no such other threads,
1116  * threads running, svc_clean_up_xprts() does a simple version of a
1117  * server's main event loop, and in the case where there are other
1118  * threads, we may need to wait a little while and then check again to
1119  * see if they're done.
1120  */
1121 void svc_close_net(struct svc_serv *serv, struct net *net)
1122 {
1123 	int delay = 0;
1124 
1125 	while (svc_close_list(serv, &serv->sv_permsocks, net) +
1126 	       svc_close_list(serv, &serv->sv_tempsocks, net)) {
1127 
1128 		svc_clean_up_xprts(serv, net);
1129 		msleep(delay++);
1130 	}
1131 }
1132 
1133 /*
1134  * Handle defer and revisit of requests
1135  */
1136 
1137 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1138 {
1139 	struct svc_deferred_req *dr =
1140 		container_of(dreq, struct svc_deferred_req, handle);
1141 	struct svc_xprt *xprt = dr->xprt;
1142 
1143 	spin_lock(&xprt->xpt_lock);
1144 	set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1145 	if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1146 		spin_unlock(&xprt->xpt_lock);
1147 		dprintk("revisit canceled\n");
1148 		svc_xprt_put(xprt);
1149 		kfree(dr);
1150 		return;
1151 	}
1152 	dprintk("revisit queued\n");
1153 	dr->xprt = NULL;
1154 	list_add(&dr->handle.recent, &xprt->xpt_deferred);
1155 	spin_unlock(&xprt->xpt_lock);
1156 	svc_xprt_enqueue(xprt);
1157 	svc_xprt_put(xprt);
1158 }
1159 
1160 /*
1161  * Save the request off for later processing. The request buffer looks
1162  * like this:
1163  *
1164  * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1165  *
1166  * This code can only handle requests that consist of an xprt-header
1167  * and rpc-header.
1168  */
1169 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1170 {
1171 	struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1172 	struct svc_deferred_req *dr;
1173 
1174 	if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1175 		return NULL; /* if more than a page, give up FIXME */
1176 	if (rqstp->rq_deferred) {
1177 		dr = rqstp->rq_deferred;
1178 		rqstp->rq_deferred = NULL;
1179 	} else {
1180 		size_t skip;
1181 		size_t size;
1182 		/* FIXME maybe discard if size too large */
1183 		size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1184 		dr = kmalloc(size, GFP_KERNEL);
1185 		if (dr == NULL)
1186 			return NULL;
1187 
1188 		dr->handle.owner = rqstp->rq_server;
1189 		dr->prot = rqstp->rq_prot;
1190 		memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1191 		dr->addrlen = rqstp->rq_addrlen;
1192 		dr->daddr = rqstp->rq_daddr;
1193 		dr->argslen = rqstp->rq_arg.len >> 2;
1194 		dr->xprt_hlen = rqstp->rq_xprt_hlen;
1195 
1196 		/* back up head to the start of the buffer and copy */
1197 		skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1198 		memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1199 		       dr->argslen << 2);
1200 	}
1201 	svc_xprt_get(rqstp->rq_xprt);
1202 	dr->xprt = rqstp->rq_xprt;
1203 	set_bit(RQ_DROPME, &rqstp->rq_flags);
1204 
1205 	dr->handle.revisit = svc_revisit;
1206 	return &dr->handle;
1207 }
1208 
1209 /*
1210  * recv data from a deferred request into an active one
1211  */
1212 static int svc_deferred_recv(struct svc_rqst *rqstp)
1213 {
1214 	struct svc_deferred_req *dr = rqstp->rq_deferred;
1215 
1216 	/* setup iov_base past transport header */
1217 	rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1218 	/* The iov_len does not include the transport header bytes */
1219 	rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1220 	rqstp->rq_arg.page_len = 0;
1221 	/* The rq_arg.len includes the transport header bytes */
1222 	rqstp->rq_arg.len     = dr->argslen<<2;
1223 	rqstp->rq_prot        = dr->prot;
1224 	memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1225 	rqstp->rq_addrlen     = dr->addrlen;
1226 	/* Save off transport header len in case we get deferred again */
1227 	rqstp->rq_xprt_hlen   = dr->xprt_hlen;
1228 	rqstp->rq_daddr       = dr->daddr;
1229 	rqstp->rq_respages    = rqstp->rq_pages;
1230 	return (dr->argslen<<2) - dr->xprt_hlen;
1231 }
1232 
1233 
1234 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1235 {
1236 	struct svc_deferred_req *dr = NULL;
1237 
1238 	if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1239 		return NULL;
1240 	spin_lock(&xprt->xpt_lock);
1241 	if (!list_empty(&xprt->xpt_deferred)) {
1242 		dr = list_entry(xprt->xpt_deferred.next,
1243 				struct svc_deferred_req,
1244 				handle.recent);
1245 		list_del_init(&dr->handle.recent);
1246 	} else
1247 		clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1248 	spin_unlock(&xprt->xpt_lock);
1249 	return dr;
1250 }
1251 
1252 /**
1253  * svc_find_xprt - find an RPC transport instance
1254  * @serv: pointer to svc_serv to search
1255  * @xcl_name: C string containing transport's class name
1256  * @net: owner net pointer
1257  * @af: Address family of transport's local address
1258  * @port: transport's IP port number
1259  *
1260  * Return the transport instance pointer for the endpoint accepting
1261  * connections/peer traffic from the specified transport class,
1262  * address family and port.
1263  *
1264  * Specifying 0 for the address family or port is effectively a
1265  * wild-card, and will result in matching the first transport in the
1266  * service's list that has a matching class name.
1267  */
1268 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1269 			       struct net *net, const sa_family_t af,
1270 			       const unsigned short port)
1271 {
1272 	struct svc_xprt *xprt;
1273 	struct svc_xprt *found = NULL;
1274 
1275 	/* Sanity check the args */
1276 	if (serv == NULL || xcl_name == NULL)
1277 		return found;
1278 
1279 	spin_lock_bh(&serv->sv_lock);
1280 	list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1281 		if (xprt->xpt_net != net)
1282 			continue;
1283 		if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1284 			continue;
1285 		if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1286 			continue;
1287 		if (port != 0 && port != svc_xprt_local_port(xprt))
1288 			continue;
1289 		found = xprt;
1290 		svc_xprt_get(xprt);
1291 		break;
1292 	}
1293 	spin_unlock_bh(&serv->sv_lock);
1294 	return found;
1295 }
1296 EXPORT_SYMBOL_GPL(svc_find_xprt);
1297 
1298 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1299 			     char *pos, int remaining)
1300 {
1301 	int len;
1302 
1303 	len = snprintf(pos, remaining, "%s %u\n",
1304 			xprt->xpt_class->xcl_name,
1305 			svc_xprt_local_port(xprt));
1306 	if (len >= remaining)
1307 		return -ENAMETOOLONG;
1308 	return len;
1309 }
1310 
1311 /**
1312  * svc_xprt_names - format a buffer with a list of transport names
1313  * @serv: pointer to an RPC service
1314  * @buf: pointer to a buffer to be filled in
1315  * @buflen: length of buffer to be filled in
1316  *
1317  * Fills in @buf with a string containing a list of transport names,
1318  * each name terminated with '\n'.
1319  *
1320  * Returns positive length of the filled-in string on success; otherwise
1321  * a negative errno value is returned if an error occurs.
1322  */
1323 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1324 {
1325 	struct svc_xprt *xprt;
1326 	int len, totlen;
1327 	char *pos;
1328 
1329 	/* Sanity check args */
1330 	if (!serv)
1331 		return 0;
1332 
1333 	spin_lock_bh(&serv->sv_lock);
1334 
1335 	pos = buf;
1336 	totlen = 0;
1337 	list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1338 		len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1339 		if (len < 0) {
1340 			*buf = '\0';
1341 			totlen = len;
1342 		}
1343 		if (len <= 0)
1344 			break;
1345 
1346 		pos += len;
1347 		totlen += len;
1348 	}
1349 
1350 	spin_unlock_bh(&serv->sv_lock);
1351 	return totlen;
1352 }
1353 EXPORT_SYMBOL_GPL(svc_xprt_names);
1354 
1355 
1356 /*----------------------------------------------------------------------------*/
1357 
1358 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1359 {
1360 	unsigned int pidx = (unsigned int)*pos;
1361 	struct svc_serv *serv = m->private;
1362 
1363 	dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1364 
1365 	if (!pidx)
1366 		return SEQ_START_TOKEN;
1367 	return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1368 }
1369 
1370 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1371 {
1372 	struct svc_pool *pool = p;
1373 	struct svc_serv *serv = m->private;
1374 
1375 	dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1376 
1377 	if (p == SEQ_START_TOKEN) {
1378 		pool = &serv->sv_pools[0];
1379 	} else {
1380 		unsigned int pidx = (pool - &serv->sv_pools[0]);
1381 		if (pidx < serv->sv_nrpools-1)
1382 			pool = &serv->sv_pools[pidx+1];
1383 		else
1384 			pool = NULL;
1385 	}
1386 	++*pos;
1387 	return pool;
1388 }
1389 
1390 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1391 {
1392 }
1393 
1394 static int svc_pool_stats_show(struct seq_file *m, void *p)
1395 {
1396 	struct svc_pool *pool = p;
1397 
1398 	if (p == SEQ_START_TOKEN) {
1399 		seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1400 		return 0;
1401 	}
1402 
1403 	seq_printf(m, "%u %lu %lu %lu %lu\n",
1404 		pool->sp_id,
1405 		(unsigned long)atomic_long_read(&pool->sp_stats.packets),
1406 		pool->sp_stats.sockets_queued,
1407 		(unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1408 		(unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1409 
1410 	return 0;
1411 }
1412 
1413 static const struct seq_operations svc_pool_stats_seq_ops = {
1414 	.start	= svc_pool_stats_start,
1415 	.next	= svc_pool_stats_next,
1416 	.stop	= svc_pool_stats_stop,
1417 	.show	= svc_pool_stats_show,
1418 };
1419 
1420 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1421 {
1422 	int err;
1423 
1424 	err = seq_open(file, &svc_pool_stats_seq_ops);
1425 	if (!err)
1426 		((struct seq_file *) file->private_data)->private = serv;
1427 	return err;
1428 }
1429 EXPORT_SYMBOL(svc_pool_stats_open);
1430 
1431 /*----------------------------------------------------------------------------*/
1432