xref: /linux/net/sunrpc/svc_xprt.c (revision d257f9bf06129613de539ea71ecea60848b662cd)
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 unsigned int svc_rpc_per_connection_limit __read_mostly;
25 module_param(svc_rpc_per_connection_limit, uint, 0644);
26 
27 
28 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
29 static int svc_deferred_recv(struct svc_rqst *rqstp);
30 static struct cache_deferred_req *svc_defer(struct cache_req *req);
31 static void svc_age_temp_xprts(unsigned long closure);
32 static void svc_delete_xprt(struct svc_xprt *xprt);
33 
34 /* apparently the "standard" is that clients close
35  * idle connections after 5 minutes, servers after
36  * 6 minutes
37  *   http://www.connectathon.org/talks96/nfstcp.pdf
38  */
39 static int svc_conn_age_period = 6*60;
40 
41 /* List of registered transport classes */
42 static DEFINE_SPINLOCK(svc_xprt_class_lock);
43 static LIST_HEAD(svc_xprt_class_list);
44 
45 /* SMP locking strategy:
46  *
47  *	svc_pool->sp_lock protects most of the fields of that pool.
48  *	svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
49  *	when both need to be taken (rare), svc_serv->sv_lock is first.
50  *	The "service mutex" protects svc_serv->sv_nrthread.
51  *	svc_sock->sk_lock protects the svc_sock->sk_deferred list
52  *             and the ->sk_info_authunix cache.
53  *
54  *	The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
55  *	enqueued multiply. During normal transport processing this bit
56  *	is set by svc_xprt_enqueue and cleared by svc_xprt_received.
57  *	Providers should not manipulate this bit directly.
58  *
59  *	Some flags can be set to certain values at any time
60  *	providing that certain rules are followed:
61  *
62  *	XPT_CONN, XPT_DATA:
63  *		- Can be set or cleared at any time.
64  *		- After a set, svc_xprt_enqueue must be called to enqueue
65  *		  the transport for processing.
66  *		- After a clear, the transport must be read/accepted.
67  *		  If this succeeds, it must be set again.
68  *	XPT_CLOSE:
69  *		- Can set at any time. It is never cleared.
70  *      XPT_DEAD:
71  *		- Can only be set while XPT_BUSY is held which ensures
72  *		  that no other thread will be using the transport or will
73  *		  try to set XPT_DEAD.
74  */
75 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
76 {
77 	struct svc_xprt_class *cl;
78 	int res = -EEXIST;
79 
80 	dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
81 
82 	INIT_LIST_HEAD(&xcl->xcl_list);
83 	spin_lock(&svc_xprt_class_lock);
84 	/* Make sure there isn't already a class with the same name */
85 	list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
86 		if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
87 			goto out;
88 	}
89 	list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
90 	res = 0;
91 out:
92 	spin_unlock(&svc_xprt_class_lock);
93 	return res;
94 }
95 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
96 
97 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
98 {
99 	dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
100 	spin_lock(&svc_xprt_class_lock);
101 	list_del_init(&xcl->xcl_list);
102 	spin_unlock(&svc_xprt_class_lock);
103 }
104 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
105 
106 /*
107  * Format the transport list for printing
108  */
109 int svc_print_xprts(char *buf, int maxlen)
110 {
111 	struct svc_xprt_class *xcl;
112 	char tmpstr[80];
113 	int len = 0;
114 	buf[0] = '\0';
115 
116 	spin_lock(&svc_xprt_class_lock);
117 	list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
118 		int slen;
119 
120 		sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
121 		slen = strlen(tmpstr);
122 		if (len + slen > maxlen)
123 			break;
124 		len += slen;
125 		strcat(buf, tmpstr);
126 	}
127 	spin_unlock(&svc_xprt_class_lock);
128 
129 	return len;
130 }
131 
132 static void svc_xprt_free(struct kref *kref)
133 {
134 	struct svc_xprt *xprt =
135 		container_of(kref, struct svc_xprt, xpt_ref);
136 	struct module *owner = xprt->xpt_class->xcl_owner;
137 	if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
138 		svcauth_unix_info_release(xprt);
139 	put_net(xprt->xpt_net);
140 	/* See comment on corresponding get in xs_setup_bc_tcp(): */
141 	if (xprt->xpt_bc_xprt)
142 		xprt_put(xprt->xpt_bc_xprt);
143 	if (xprt->xpt_bc_xps)
144 		xprt_switch_put(xprt->xpt_bc_xps);
145 	xprt->xpt_ops->xpo_free(xprt);
146 	module_put(owner);
147 }
148 
149 void svc_xprt_put(struct svc_xprt *xprt)
150 {
151 	kref_put(&xprt->xpt_ref, svc_xprt_free);
152 }
153 EXPORT_SYMBOL_GPL(svc_xprt_put);
154 
155 /*
156  * Called by transport drivers to initialize the transport independent
157  * portion of the transport instance.
158  */
159 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
160 		   struct svc_xprt *xprt, struct svc_serv *serv)
161 {
162 	memset(xprt, 0, sizeof(*xprt));
163 	xprt->xpt_class = xcl;
164 	xprt->xpt_ops = xcl->xcl_ops;
165 	kref_init(&xprt->xpt_ref);
166 	xprt->xpt_server = serv;
167 	INIT_LIST_HEAD(&xprt->xpt_list);
168 	INIT_LIST_HEAD(&xprt->xpt_ready);
169 	INIT_LIST_HEAD(&xprt->xpt_deferred);
170 	INIT_LIST_HEAD(&xprt->xpt_users);
171 	mutex_init(&xprt->xpt_mutex);
172 	spin_lock_init(&xprt->xpt_lock);
173 	set_bit(XPT_BUSY, &xprt->xpt_flags);
174 	rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
175 	xprt->xpt_net = get_net(net);
176 }
177 EXPORT_SYMBOL_GPL(svc_xprt_init);
178 
179 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
180 					 struct svc_serv *serv,
181 					 struct net *net,
182 					 const int family,
183 					 const unsigned short port,
184 					 int flags)
185 {
186 	struct sockaddr_in sin = {
187 		.sin_family		= AF_INET,
188 		.sin_addr.s_addr	= htonl(INADDR_ANY),
189 		.sin_port		= htons(port),
190 	};
191 #if IS_ENABLED(CONFIG_IPV6)
192 	struct sockaddr_in6 sin6 = {
193 		.sin6_family		= AF_INET6,
194 		.sin6_addr		= IN6ADDR_ANY_INIT,
195 		.sin6_port		= htons(port),
196 	};
197 #endif
198 	struct sockaddr *sap;
199 	size_t len;
200 
201 	switch (family) {
202 	case PF_INET:
203 		sap = (struct sockaddr *)&sin;
204 		len = sizeof(sin);
205 		break;
206 #if IS_ENABLED(CONFIG_IPV6)
207 	case PF_INET6:
208 		sap = (struct sockaddr *)&sin6;
209 		len = sizeof(sin6);
210 		break;
211 #endif
212 	default:
213 		return ERR_PTR(-EAFNOSUPPORT);
214 	}
215 
216 	return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
217 }
218 
219 /*
220  * svc_xprt_received conditionally queues the transport for processing
221  * by another thread. The caller must hold the XPT_BUSY bit and must
222  * not thereafter touch transport data.
223  *
224  * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
225  * insufficient) data.
226  */
227 static void svc_xprt_received(struct svc_xprt *xprt)
228 {
229 	if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
230 		WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
231 		return;
232 	}
233 
234 	/* As soon as we clear busy, the xprt could be closed and
235 	 * 'put', so we need a reference to call svc_enqueue_xprt with:
236 	 */
237 	svc_xprt_get(xprt);
238 	smp_mb__before_atomic();
239 	clear_bit(XPT_BUSY, &xprt->xpt_flags);
240 	xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
241 	svc_xprt_put(xprt);
242 }
243 
244 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
245 {
246 	clear_bit(XPT_TEMP, &new->xpt_flags);
247 	spin_lock_bh(&serv->sv_lock);
248 	list_add(&new->xpt_list, &serv->sv_permsocks);
249 	spin_unlock_bh(&serv->sv_lock);
250 	svc_xprt_received(new);
251 }
252 
253 int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
254 		    struct net *net, const int family,
255 		    const unsigned short port, int flags)
256 {
257 	struct svc_xprt_class *xcl;
258 
259 	spin_lock(&svc_xprt_class_lock);
260 	list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
261 		struct svc_xprt *newxprt;
262 		unsigned short newport;
263 
264 		if (strcmp(xprt_name, xcl->xcl_name))
265 			continue;
266 
267 		if (!try_module_get(xcl->xcl_owner))
268 			goto err;
269 
270 		spin_unlock(&svc_xprt_class_lock);
271 		newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
272 		if (IS_ERR(newxprt)) {
273 			module_put(xcl->xcl_owner);
274 			return PTR_ERR(newxprt);
275 		}
276 		svc_add_new_perm_xprt(serv, newxprt);
277 		newport = svc_xprt_local_port(newxprt);
278 		return newport;
279 	}
280  err:
281 	spin_unlock(&svc_xprt_class_lock);
282 	/* This errno is exposed to user space.  Provide a reasonable
283 	 * perror msg for a bad transport. */
284 	return -EPROTONOSUPPORT;
285 }
286 
287 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
288 		    struct net *net, const int family,
289 		    const unsigned short port, int flags)
290 {
291 	int err;
292 
293 	dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
294 	err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
295 	if (err == -EPROTONOSUPPORT) {
296 		request_module("svc%s", xprt_name);
297 		err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
298 	}
299 	if (err)
300 		dprintk("svc: transport %s not found, err %d\n",
301 			xprt_name, err);
302 	return err;
303 }
304 EXPORT_SYMBOL_GPL(svc_create_xprt);
305 
306 /*
307  * Copy the local and remote xprt addresses to the rqstp structure
308  */
309 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
310 {
311 	memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
312 	rqstp->rq_addrlen = xprt->xpt_remotelen;
313 
314 	/*
315 	 * Destination address in request is needed for binding the
316 	 * source address in RPC replies/callbacks later.
317 	 */
318 	memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
319 	rqstp->rq_daddrlen = xprt->xpt_locallen;
320 }
321 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
322 
323 /**
324  * svc_print_addr - Format rq_addr field for printing
325  * @rqstp: svc_rqst struct containing address to print
326  * @buf: target buffer for formatted address
327  * @len: length of target buffer
328  *
329  */
330 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
331 {
332 	return __svc_print_addr(svc_addr(rqstp), buf, len);
333 }
334 EXPORT_SYMBOL_GPL(svc_print_addr);
335 
336 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
337 {
338 	unsigned int limit = svc_rpc_per_connection_limit;
339 	int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
340 
341 	return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
342 }
343 
344 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
345 {
346 	if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
347 		if (!svc_xprt_slots_in_range(xprt))
348 			return false;
349 		atomic_inc(&xprt->xpt_nr_rqsts);
350 		set_bit(RQ_DATA, &rqstp->rq_flags);
351 	}
352 	return true;
353 }
354 
355 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
356 {
357 	struct svc_xprt	*xprt = rqstp->rq_xprt;
358 	if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
359 		atomic_dec(&xprt->xpt_nr_rqsts);
360 		svc_xprt_enqueue(xprt);
361 	}
362 }
363 
364 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
365 {
366 	if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
367 		return true;
368 	if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED))) {
369 		if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
370 		    svc_xprt_slots_in_range(xprt))
371 			return true;
372 		trace_svc_xprt_no_write_space(xprt);
373 		return false;
374 	}
375 	return false;
376 }
377 
378 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
379 {
380 	struct svc_pool *pool;
381 	struct svc_rqst	*rqstp = NULL;
382 	int cpu;
383 	bool queued = false;
384 
385 	if (!svc_xprt_has_something_to_do(xprt))
386 		goto out;
387 
388 	/* Mark transport as busy. It will remain in this state until
389 	 * the provider calls svc_xprt_received. We update XPT_BUSY
390 	 * atomically because it also guards against trying to enqueue
391 	 * the transport twice.
392 	 */
393 	if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
394 		/* Don't enqueue transport while already enqueued */
395 		dprintk("svc: transport %p busy, not enqueued\n", xprt);
396 		goto out;
397 	}
398 
399 	cpu = get_cpu();
400 	pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
401 
402 	atomic_long_inc(&pool->sp_stats.packets);
403 
404 redo_search:
405 	/* find a thread for this xprt */
406 	rcu_read_lock();
407 	list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
408 		/* Do a lockless check first */
409 		if (test_bit(RQ_BUSY, &rqstp->rq_flags))
410 			continue;
411 
412 		/*
413 		 * Once the xprt has been queued, it can only be dequeued by
414 		 * the task that intends to service it. All we can do at that
415 		 * point is to try to wake this thread back up so that it can
416 		 * do so.
417 		 */
418 		if (!queued) {
419 			spin_lock_bh(&rqstp->rq_lock);
420 			if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags)) {
421 				/* already busy, move on... */
422 				spin_unlock_bh(&rqstp->rq_lock);
423 				continue;
424 			}
425 
426 			/* this one will do */
427 			rqstp->rq_xprt = xprt;
428 			svc_xprt_get(xprt);
429 			spin_unlock_bh(&rqstp->rq_lock);
430 		}
431 		rcu_read_unlock();
432 
433 		atomic_long_inc(&pool->sp_stats.threads_woken);
434 		wake_up_process(rqstp->rq_task);
435 		put_cpu();
436 		goto out;
437 	}
438 	rcu_read_unlock();
439 
440 	/*
441 	 * We didn't find an idle thread to use, so we need to queue the xprt.
442 	 * Do so and then search again. If we find one, we can't hook this one
443 	 * up to it directly but we can wake the thread up in the hopes that it
444 	 * will pick it up once it searches for a xprt to service.
445 	 */
446 	if (!queued) {
447 		queued = true;
448 		dprintk("svc: transport %p put into queue\n", xprt);
449 		spin_lock_bh(&pool->sp_lock);
450 		list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
451 		pool->sp_stats.sockets_queued++;
452 		spin_unlock_bh(&pool->sp_lock);
453 		goto redo_search;
454 	}
455 	rqstp = NULL;
456 	put_cpu();
457 out:
458 	trace_svc_xprt_do_enqueue(xprt, rqstp);
459 }
460 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
461 
462 /*
463  * Queue up a transport with data pending. If there are idle nfsd
464  * processes, wake 'em up.
465  *
466  */
467 void svc_xprt_enqueue(struct svc_xprt *xprt)
468 {
469 	if (test_bit(XPT_BUSY, &xprt->xpt_flags))
470 		return;
471 	xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
472 }
473 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
474 
475 /*
476  * Dequeue the first transport, if there is one.
477  */
478 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
479 {
480 	struct svc_xprt	*xprt = NULL;
481 
482 	if (list_empty(&pool->sp_sockets))
483 		goto out;
484 
485 	spin_lock_bh(&pool->sp_lock);
486 	if (likely(!list_empty(&pool->sp_sockets))) {
487 		xprt = list_first_entry(&pool->sp_sockets,
488 					struct svc_xprt, xpt_ready);
489 		list_del_init(&xprt->xpt_ready);
490 		svc_xprt_get(xprt);
491 
492 		dprintk("svc: transport %p dequeued, inuse=%d\n",
493 			xprt, kref_read(&xprt->xpt_ref));
494 	}
495 	spin_unlock_bh(&pool->sp_lock);
496 out:
497 	trace_svc_xprt_dequeue(xprt);
498 	return xprt;
499 }
500 
501 /**
502  * svc_reserve - change the space reserved for the reply to a request.
503  * @rqstp:  The request in question
504  * @space: new max space to reserve
505  *
506  * Each request reserves some space on the output queue of the transport
507  * to make sure the reply fits.  This function reduces that reserved
508  * space to be the amount of space used already, plus @space.
509  *
510  */
511 void svc_reserve(struct svc_rqst *rqstp, int space)
512 {
513 	space += rqstp->rq_res.head[0].iov_len;
514 
515 	if (space < rqstp->rq_reserved) {
516 		struct svc_xprt *xprt = rqstp->rq_xprt;
517 		atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
518 		rqstp->rq_reserved = space;
519 
520 		svc_xprt_enqueue(xprt);
521 	}
522 }
523 EXPORT_SYMBOL_GPL(svc_reserve);
524 
525 static void svc_xprt_release(struct svc_rqst *rqstp)
526 {
527 	struct svc_xprt	*xprt = rqstp->rq_xprt;
528 
529 	rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
530 
531 	kfree(rqstp->rq_deferred);
532 	rqstp->rq_deferred = NULL;
533 
534 	svc_free_res_pages(rqstp);
535 	rqstp->rq_res.page_len = 0;
536 	rqstp->rq_res.page_base = 0;
537 
538 	/* Reset response buffer and release
539 	 * the reservation.
540 	 * But first, check that enough space was reserved
541 	 * for the reply, otherwise we have a bug!
542 	 */
543 	if ((rqstp->rq_res.len) >  rqstp->rq_reserved)
544 		printk(KERN_ERR "RPC request reserved %d but used %d\n",
545 		       rqstp->rq_reserved,
546 		       rqstp->rq_res.len);
547 
548 	rqstp->rq_res.head[0].iov_len = 0;
549 	svc_reserve(rqstp, 0);
550 	svc_xprt_release_slot(rqstp);
551 	rqstp->rq_xprt = NULL;
552 	svc_xprt_put(xprt);
553 }
554 
555 /*
556  * Some svc_serv's will have occasional work to do, even when a xprt is not
557  * waiting to be serviced. This function is there to "kick" a task in one of
558  * those services so that it can wake up and do that work. Note that we only
559  * bother with pool 0 as we don't need to wake up more than one thread for
560  * this purpose.
561  */
562 void svc_wake_up(struct svc_serv *serv)
563 {
564 	struct svc_rqst	*rqstp;
565 	struct svc_pool *pool;
566 
567 	pool = &serv->sv_pools[0];
568 
569 	rcu_read_lock();
570 	list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
571 		/* skip any that aren't queued */
572 		if (test_bit(RQ_BUSY, &rqstp->rq_flags))
573 			continue;
574 		rcu_read_unlock();
575 		dprintk("svc: daemon %p woken up.\n", rqstp);
576 		wake_up_process(rqstp->rq_task);
577 		trace_svc_wake_up(rqstp->rq_task->pid);
578 		return;
579 	}
580 	rcu_read_unlock();
581 
582 	/* No free entries available */
583 	set_bit(SP_TASK_PENDING, &pool->sp_flags);
584 	smp_wmb();
585 	trace_svc_wake_up(0);
586 }
587 EXPORT_SYMBOL_GPL(svc_wake_up);
588 
589 int svc_port_is_privileged(struct sockaddr *sin)
590 {
591 	switch (sin->sa_family) {
592 	case AF_INET:
593 		return ntohs(((struct sockaddr_in *)sin)->sin_port)
594 			< PROT_SOCK;
595 	case AF_INET6:
596 		return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
597 			< PROT_SOCK;
598 	default:
599 		return 0;
600 	}
601 }
602 
603 /*
604  * Make sure that we don't have too many active connections. If we have,
605  * something must be dropped. It's not clear what will happen if we allow
606  * "too many" connections, but when dealing with network-facing software,
607  * we have to code defensively. Here we do that by imposing hard limits.
608  *
609  * There's no point in trying to do random drop here for DoS
610  * prevention. The NFS clients does 1 reconnect in 15 seconds. An
611  * attacker can easily beat that.
612  *
613  * The only somewhat efficient mechanism would be if drop old
614  * connections from the same IP first. But right now we don't even
615  * record the client IP in svc_sock.
616  *
617  * single-threaded services that expect a lot of clients will probably
618  * need to set sv_maxconn to override the default value which is based
619  * on the number of threads
620  */
621 static void svc_check_conn_limits(struct svc_serv *serv)
622 {
623 	unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
624 				(serv->sv_nrthreads+3) * 20;
625 
626 	if (serv->sv_tmpcnt > limit) {
627 		struct svc_xprt *xprt = NULL;
628 		spin_lock_bh(&serv->sv_lock);
629 		if (!list_empty(&serv->sv_tempsocks)) {
630 			/* Try to help the admin */
631 			net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
632 					       serv->sv_name, serv->sv_maxconn ?
633 					       "max number of connections" :
634 					       "number of threads");
635 			/*
636 			 * Always select the oldest connection. It's not fair,
637 			 * but so is life
638 			 */
639 			xprt = list_entry(serv->sv_tempsocks.prev,
640 					  struct svc_xprt,
641 					  xpt_list);
642 			set_bit(XPT_CLOSE, &xprt->xpt_flags);
643 			svc_xprt_get(xprt);
644 		}
645 		spin_unlock_bh(&serv->sv_lock);
646 
647 		if (xprt) {
648 			svc_xprt_enqueue(xprt);
649 			svc_xprt_put(xprt);
650 		}
651 	}
652 }
653 
654 static int svc_alloc_arg(struct svc_rqst *rqstp)
655 {
656 	struct svc_serv *serv = rqstp->rq_server;
657 	struct xdr_buf *arg;
658 	int pages;
659 	int i;
660 
661 	/* now allocate needed pages.  If we get a failure, sleep briefly */
662 	pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
663 	if (pages > RPCSVC_MAXPAGES) {
664 		pr_warn_once("svc: warning: pages=%u > RPCSVC_MAXPAGES=%lu\n",
665 			     pages, RPCSVC_MAXPAGES);
666 		/* use as many pages as possible */
667 		pages = RPCSVC_MAXPAGES;
668 	}
669 	for (i = 0; i < pages ; i++)
670 		while (rqstp->rq_pages[i] == NULL) {
671 			struct page *p = alloc_page(GFP_KERNEL);
672 			if (!p) {
673 				set_current_state(TASK_INTERRUPTIBLE);
674 				if (signalled() || kthread_should_stop()) {
675 					set_current_state(TASK_RUNNING);
676 					return -EINTR;
677 				}
678 				schedule_timeout(msecs_to_jiffies(500));
679 			}
680 			rqstp->rq_pages[i] = p;
681 		}
682 	rqstp->rq_page_end = &rqstp->rq_pages[i];
683 	rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
684 
685 	/* Make arg->head point to first page and arg->pages point to rest */
686 	arg = &rqstp->rq_arg;
687 	arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
688 	arg->head[0].iov_len = PAGE_SIZE;
689 	arg->pages = rqstp->rq_pages + 1;
690 	arg->page_base = 0;
691 	/* save at least one page for response */
692 	arg->page_len = (pages-2)*PAGE_SIZE;
693 	arg->len = (pages-1)*PAGE_SIZE;
694 	arg->tail[0].iov_len = 0;
695 	return 0;
696 }
697 
698 static bool
699 rqst_should_sleep(struct svc_rqst *rqstp)
700 {
701 	struct svc_pool		*pool = rqstp->rq_pool;
702 
703 	/* did someone call svc_wake_up? */
704 	if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
705 		return false;
706 
707 	/* was a socket queued? */
708 	if (!list_empty(&pool->sp_sockets))
709 		return false;
710 
711 	/* are we shutting down? */
712 	if (signalled() || kthread_should_stop())
713 		return false;
714 
715 	/* are we freezing? */
716 	if (freezing(current))
717 		return false;
718 
719 	return true;
720 }
721 
722 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
723 {
724 	struct svc_xprt *xprt;
725 	struct svc_pool		*pool = rqstp->rq_pool;
726 	long			time_left = 0;
727 
728 	/* rq_xprt should be clear on entry */
729 	WARN_ON_ONCE(rqstp->rq_xprt);
730 
731 	/* Normally we will wait up to 5 seconds for any required
732 	 * cache information to be provided.
733 	 */
734 	rqstp->rq_chandle.thread_wait = 5*HZ;
735 
736 	xprt = svc_xprt_dequeue(pool);
737 	if (xprt) {
738 		rqstp->rq_xprt = xprt;
739 
740 		/* As there is a shortage of threads and this request
741 		 * had to be queued, don't allow the thread to wait so
742 		 * long for cache updates.
743 		 */
744 		rqstp->rq_chandle.thread_wait = 1*HZ;
745 		clear_bit(SP_TASK_PENDING, &pool->sp_flags);
746 		return xprt;
747 	}
748 
749 	/*
750 	 * We have to be able to interrupt this wait
751 	 * to bring down the daemons ...
752 	 */
753 	set_current_state(TASK_INTERRUPTIBLE);
754 	clear_bit(RQ_BUSY, &rqstp->rq_flags);
755 	smp_mb();
756 
757 	if (likely(rqst_should_sleep(rqstp)))
758 		time_left = schedule_timeout(timeout);
759 	else
760 		__set_current_state(TASK_RUNNING);
761 
762 	try_to_freeze();
763 
764 	spin_lock_bh(&rqstp->rq_lock);
765 	set_bit(RQ_BUSY, &rqstp->rq_flags);
766 	spin_unlock_bh(&rqstp->rq_lock);
767 
768 	xprt = rqstp->rq_xprt;
769 	if (xprt != NULL)
770 		return xprt;
771 
772 	if (!time_left)
773 		atomic_long_inc(&pool->sp_stats.threads_timedout);
774 
775 	if (signalled() || kthread_should_stop())
776 		return ERR_PTR(-EINTR);
777 	return ERR_PTR(-EAGAIN);
778 }
779 
780 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
781 {
782 	spin_lock_bh(&serv->sv_lock);
783 	set_bit(XPT_TEMP, &newxpt->xpt_flags);
784 	list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
785 	serv->sv_tmpcnt++;
786 	if (serv->sv_temptimer.function == NULL) {
787 		/* setup timer to age temp transports */
788 		setup_timer(&serv->sv_temptimer, svc_age_temp_xprts,
789 			    (unsigned long)serv);
790 		mod_timer(&serv->sv_temptimer,
791 			  jiffies + svc_conn_age_period * HZ);
792 	}
793 	spin_unlock_bh(&serv->sv_lock);
794 	svc_xprt_received(newxpt);
795 }
796 
797 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
798 {
799 	struct svc_serv *serv = rqstp->rq_server;
800 	int len = 0;
801 
802 	if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
803 		dprintk("svc_recv: found XPT_CLOSE\n");
804 		if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
805 			xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
806 		svc_delete_xprt(xprt);
807 		/* Leave XPT_BUSY set on the dead xprt: */
808 		goto out;
809 	}
810 	if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
811 		struct svc_xprt *newxpt;
812 		/*
813 		 * We know this module_get will succeed because the
814 		 * listener holds a reference too
815 		 */
816 		__module_get(xprt->xpt_class->xcl_owner);
817 		svc_check_conn_limits(xprt->xpt_server);
818 		newxpt = xprt->xpt_ops->xpo_accept(xprt);
819 		if (newxpt)
820 			svc_add_new_temp_xprt(serv, newxpt);
821 		else
822 			module_put(xprt->xpt_class->xcl_owner);
823 	} else if (svc_xprt_reserve_slot(rqstp, xprt)) {
824 		/* XPT_DATA|XPT_DEFERRED case: */
825 		dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
826 			rqstp, rqstp->rq_pool->sp_id, xprt,
827 			kref_read(&xprt->xpt_ref));
828 		rqstp->rq_deferred = svc_deferred_dequeue(xprt);
829 		if (rqstp->rq_deferred)
830 			len = svc_deferred_recv(rqstp);
831 		else
832 			len = xprt->xpt_ops->xpo_recvfrom(rqstp);
833 		dprintk("svc: got len=%d\n", len);
834 		rqstp->rq_reserved = serv->sv_max_mesg;
835 		atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
836 	}
837 	/* clear XPT_BUSY: */
838 	svc_xprt_received(xprt);
839 out:
840 	trace_svc_handle_xprt(xprt, len);
841 	return len;
842 }
843 
844 /*
845  * Receive the next request on any transport.  This code is carefully
846  * organised not to touch any cachelines in the shared svc_serv
847  * structure, only cachelines in the local svc_pool.
848  */
849 int svc_recv(struct svc_rqst *rqstp, long timeout)
850 {
851 	struct svc_xprt		*xprt = NULL;
852 	struct svc_serv		*serv = rqstp->rq_server;
853 	int			len, err;
854 
855 	dprintk("svc: server %p waiting for data (to = %ld)\n",
856 		rqstp, timeout);
857 
858 	if (rqstp->rq_xprt)
859 		printk(KERN_ERR
860 			"svc_recv: service %p, transport not NULL!\n",
861 			 rqstp);
862 
863 	err = svc_alloc_arg(rqstp);
864 	if (err)
865 		goto out;
866 
867 	try_to_freeze();
868 	cond_resched();
869 	err = -EINTR;
870 	if (signalled() || kthread_should_stop())
871 		goto out;
872 
873 	xprt = svc_get_next_xprt(rqstp, timeout);
874 	if (IS_ERR(xprt)) {
875 		err = PTR_ERR(xprt);
876 		goto out;
877 	}
878 
879 	len = svc_handle_xprt(rqstp, xprt);
880 
881 	/* No data, incomplete (TCP) read, or accept() */
882 	err = -EAGAIN;
883 	if (len <= 0)
884 		goto out_release;
885 
886 	clear_bit(XPT_OLD, &xprt->xpt_flags);
887 
888 	if (xprt->xpt_ops->xpo_secure_port(rqstp))
889 		set_bit(RQ_SECURE, &rqstp->rq_flags);
890 	else
891 		clear_bit(RQ_SECURE, &rqstp->rq_flags);
892 	rqstp->rq_chandle.defer = svc_defer;
893 	rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
894 
895 	if (serv->sv_stats)
896 		serv->sv_stats->netcnt++;
897 	trace_svc_recv(rqstp, len);
898 	return len;
899 out_release:
900 	rqstp->rq_res.len = 0;
901 	svc_xprt_release(rqstp);
902 out:
903 	trace_svc_recv(rqstp, err);
904 	return err;
905 }
906 EXPORT_SYMBOL_GPL(svc_recv);
907 
908 /*
909  * Drop request
910  */
911 void svc_drop(struct svc_rqst *rqstp)
912 {
913 	trace_svc_drop(rqstp);
914 	dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
915 	svc_xprt_release(rqstp);
916 }
917 EXPORT_SYMBOL_GPL(svc_drop);
918 
919 /*
920  * Return reply to client.
921  */
922 int svc_send(struct svc_rqst *rqstp)
923 {
924 	struct svc_xprt	*xprt;
925 	int		len = -EFAULT;
926 	struct xdr_buf	*xb;
927 
928 	xprt = rqstp->rq_xprt;
929 	if (!xprt)
930 		goto out;
931 
932 	/* release the receive skb before sending the reply */
933 	rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
934 
935 	/* calculate over-all length */
936 	xb = &rqstp->rq_res;
937 	xb->len = xb->head[0].iov_len +
938 		xb->page_len +
939 		xb->tail[0].iov_len;
940 
941 	/* Grab mutex to serialize outgoing data. */
942 	mutex_lock(&xprt->xpt_mutex);
943 	if (test_bit(XPT_DEAD, &xprt->xpt_flags)
944 			|| test_bit(XPT_CLOSE, &xprt->xpt_flags))
945 		len = -ENOTCONN;
946 	else
947 		len = xprt->xpt_ops->xpo_sendto(rqstp);
948 	mutex_unlock(&xprt->xpt_mutex);
949 	rpc_wake_up(&xprt->xpt_bc_pending);
950 	svc_xprt_release(rqstp);
951 
952 	if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
953 		len = 0;
954 out:
955 	trace_svc_send(rqstp, len);
956 	return len;
957 }
958 
959 /*
960  * Timer function to close old temporary transports, using
961  * a mark-and-sweep algorithm.
962  */
963 static void svc_age_temp_xprts(unsigned long closure)
964 {
965 	struct svc_serv *serv = (struct svc_serv *)closure;
966 	struct svc_xprt *xprt;
967 	struct list_head *le, *next;
968 
969 	dprintk("svc_age_temp_xprts\n");
970 
971 	if (!spin_trylock_bh(&serv->sv_lock)) {
972 		/* busy, try again 1 sec later */
973 		dprintk("svc_age_temp_xprts: busy\n");
974 		mod_timer(&serv->sv_temptimer, jiffies + HZ);
975 		return;
976 	}
977 
978 	list_for_each_safe(le, next, &serv->sv_tempsocks) {
979 		xprt = list_entry(le, struct svc_xprt, xpt_list);
980 
981 		/* First time through, just mark it OLD. Second time
982 		 * through, close it. */
983 		if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
984 			continue;
985 		if (kref_read(&xprt->xpt_ref) > 1 ||
986 		    test_bit(XPT_BUSY, &xprt->xpt_flags))
987 			continue;
988 		list_del_init(le);
989 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
990 		dprintk("queuing xprt %p for closing\n", xprt);
991 
992 		/* a thread will dequeue and close it soon */
993 		svc_xprt_enqueue(xprt);
994 	}
995 	spin_unlock_bh(&serv->sv_lock);
996 
997 	mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
998 }
999 
1000 /* Close temporary transports whose xpt_local matches server_addr immediately
1001  * instead of waiting for them to be picked up by the timer.
1002  *
1003  * This is meant to be called from a notifier_block that runs when an ip
1004  * address is deleted.
1005  */
1006 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
1007 {
1008 	struct svc_xprt *xprt;
1009 	struct list_head *le, *next;
1010 	LIST_HEAD(to_be_closed);
1011 
1012 	spin_lock_bh(&serv->sv_lock);
1013 	list_for_each_safe(le, next, &serv->sv_tempsocks) {
1014 		xprt = list_entry(le, struct svc_xprt, xpt_list);
1015 		if (rpc_cmp_addr(server_addr, (struct sockaddr *)
1016 				&xprt->xpt_local)) {
1017 			dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
1018 			list_move(le, &to_be_closed);
1019 		}
1020 	}
1021 	spin_unlock_bh(&serv->sv_lock);
1022 
1023 	while (!list_empty(&to_be_closed)) {
1024 		le = to_be_closed.next;
1025 		list_del_init(le);
1026 		xprt = list_entry(le, struct svc_xprt, xpt_list);
1027 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
1028 		set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
1029 		dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
1030 				xprt);
1031 		svc_xprt_enqueue(xprt);
1032 	}
1033 }
1034 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1035 
1036 static void call_xpt_users(struct svc_xprt *xprt)
1037 {
1038 	struct svc_xpt_user *u;
1039 
1040 	spin_lock(&xprt->xpt_lock);
1041 	while (!list_empty(&xprt->xpt_users)) {
1042 		u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1043 		list_del(&u->list);
1044 		u->callback(u);
1045 	}
1046 	spin_unlock(&xprt->xpt_lock);
1047 }
1048 
1049 /*
1050  * Remove a dead transport
1051  */
1052 static void svc_delete_xprt(struct svc_xprt *xprt)
1053 {
1054 	struct svc_serv	*serv = xprt->xpt_server;
1055 	struct svc_deferred_req *dr;
1056 
1057 	/* Only do this once */
1058 	if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1059 		BUG();
1060 
1061 	dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1062 	xprt->xpt_ops->xpo_detach(xprt);
1063 
1064 	spin_lock_bh(&serv->sv_lock);
1065 	list_del_init(&xprt->xpt_list);
1066 	WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1067 	if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1068 		serv->sv_tmpcnt--;
1069 	spin_unlock_bh(&serv->sv_lock);
1070 
1071 	while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1072 		kfree(dr);
1073 
1074 	call_xpt_users(xprt);
1075 	svc_xprt_put(xprt);
1076 }
1077 
1078 void svc_close_xprt(struct svc_xprt *xprt)
1079 {
1080 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
1081 	if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1082 		/* someone else will have to effect the close */
1083 		return;
1084 	/*
1085 	 * We expect svc_close_xprt() to work even when no threads are
1086 	 * running (e.g., while configuring the server before starting
1087 	 * any threads), so if the transport isn't busy, we delete
1088 	 * it ourself:
1089 	 */
1090 	svc_delete_xprt(xprt);
1091 }
1092 EXPORT_SYMBOL_GPL(svc_close_xprt);
1093 
1094 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1095 {
1096 	struct svc_xprt *xprt;
1097 	int ret = 0;
1098 
1099 	spin_lock(&serv->sv_lock);
1100 	list_for_each_entry(xprt, xprt_list, xpt_list) {
1101 		if (xprt->xpt_net != net)
1102 			continue;
1103 		ret++;
1104 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
1105 		svc_xprt_enqueue(xprt);
1106 	}
1107 	spin_unlock(&serv->sv_lock);
1108 	return ret;
1109 }
1110 
1111 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1112 {
1113 	struct svc_pool *pool;
1114 	struct svc_xprt *xprt;
1115 	struct svc_xprt *tmp;
1116 	int i;
1117 
1118 	for (i = 0; i < serv->sv_nrpools; i++) {
1119 		pool = &serv->sv_pools[i];
1120 
1121 		spin_lock_bh(&pool->sp_lock);
1122 		list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1123 			if (xprt->xpt_net != net)
1124 				continue;
1125 			list_del_init(&xprt->xpt_ready);
1126 			spin_unlock_bh(&pool->sp_lock);
1127 			return xprt;
1128 		}
1129 		spin_unlock_bh(&pool->sp_lock);
1130 	}
1131 	return NULL;
1132 }
1133 
1134 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1135 {
1136 	struct svc_xprt *xprt;
1137 
1138 	while ((xprt = svc_dequeue_net(serv, net))) {
1139 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
1140 		svc_delete_xprt(xprt);
1141 	}
1142 }
1143 
1144 /*
1145  * Server threads may still be running (especially in the case where the
1146  * service is still running in other network namespaces).
1147  *
1148  * So we shut down sockets the same way we would on a running server, by
1149  * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1150  * the close.  In the case there are no such other threads,
1151  * threads running, svc_clean_up_xprts() does a simple version of a
1152  * server's main event loop, and in the case where there are other
1153  * threads, we may need to wait a little while and then check again to
1154  * see if they're done.
1155  */
1156 void svc_close_net(struct svc_serv *serv, struct net *net)
1157 {
1158 	int delay = 0;
1159 
1160 	while (svc_close_list(serv, &serv->sv_permsocks, net) +
1161 	       svc_close_list(serv, &serv->sv_tempsocks, net)) {
1162 
1163 		svc_clean_up_xprts(serv, net);
1164 		msleep(delay++);
1165 	}
1166 }
1167 
1168 /*
1169  * Handle defer and revisit of requests
1170  */
1171 
1172 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1173 {
1174 	struct svc_deferred_req *dr =
1175 		container_of(dreq, struct svc_deferred_req, handle);
1176 	struct svc_xprt *xprt = dr->xprt;
1177 
1178 	spin_lock(&xprt->xpt_lock);
1179 	set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1180 	if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1181 		spin_unlock(&xprt->xpt_lock);
1182 		dprintk("revisit canceled\n");
1183 		svc_xprt_put(xprt);
1184 		trace_svc_drop_deferred(dr);
1185 		kfree(dr);
1186 		return;
1187 	}
1188 	dprintk("revisit queued\n");
1189 	dr->xprt = NULL;
1190 	list_add(&dr->handle.recent, &xprt->xpt_deferred);
1191 	spin_unlock(&xprt->xpt_lock);
1192 	svc_xprt_enqueue(xprt);
1193 	svc_xprt_put(xprt);
1194 }
1195 
1196 /*
1197  * Save the request off for later processing. The request buffer looks
1198  * like this:
1199  *
1200  * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1201  *
1202  * This code can only handle requests that consist of an xprt-header
1203  * and rpc-header.
1204  */
1205 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1206 {
1207 	struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1208 	struct svc_deferred_req *dr;
1209 
1210 	if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1211 		return NULL; /* if more than a page, give up FIXME */
1212 	if (rqstp->rq_deferred) {
1213 		dr = rqstp->rq_deferred;
1214 		rqstp->rq_deferred = NULL;
1215 	} else {
1216 		size_t skip;
1217 		size_t size;
1218 		/* FIXME maybe discard if size too large */
1219 		size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1220 		dr = kmalloc(size, GFP_KERNEL);
1221 		if (dr == NULL)
1222 			return NULL;
1223 
1224 		dr->handle.owner = rqstp->rq_server;
1225 		dr->prot = rqstp->rq_prot;
1226 		memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1227 		dr->addrlen = rqstp->rq_addrlen;
1228 		dr->daddr = rqstp->rq_daddr;
1229 		dr->argslen = rqstp->rq_arg.len >> 2;
1230 		dr->xprt_hlen = rqstp->rq_xprt_hlen;
1231 
1232 		/* back up head to the start of the buffer and copy */
1233 		skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1234 		memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1235 		       dr->argslen << 2);
1236 	}
1237 	svc_xprt_get(rqstp->rq_xprt);
1238 	dr->xprt = rqstp->rq_xprt;
1239 	set_bit(RQ_DROPME, &rqstp->rq_flags);
1240 
1241 	dr->handle.revisit = svc_revisit;
1242 	trace_svc_defer(rqstp);
1243 	return &dr->handle;
1244 }
1245 
1246 /*
1247  * recv data from a deferred request into an active one
1248  */
1249 static int svc_deferred_recv(struct svc_rqst *rqstp)
1250 {
1251 	struct svc_deferred_req *dr = rqstp->rq_deferred;
1252 
1253 	/* setup iov_base past transport header */
1254 	rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1255 	/* The iov_len does not include the transport header bytes */
1256 	rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1257 	rqstp->rq_arg.page_len = 0;
1258 	/* The rq_arg.len includes the transport header bytes */
1259 	rqstp->rq_arg.len     = dr->argslen<<2;
1260 	rqstp->rq_prot        = dr->prot;
1261 	memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1262 	rqstp->rq_addrlen     = dr->addrlen;
1263 	/* Save off transport header len in case we get deferred again */
1264 	rqstp->rq_xprt_hlen   = dr->xprt_hlen;
1265 	rqstp->rq_daddr       = dr->daddr;
1266 	rqstp->rq_respages    = rqstp->rq_pages;
1267 	return (dr->argslen<<2) - dr->xprt_hlen;
1268 }
1269 
1270 
1271 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1272 {
1273 	struct svc_deferred_req *dr = NULL;
1274 
1275 	if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1276 		return NULL;
1277 	spin_lock(&xprt->xpt_lock);
1278 	if (!list_empty(&xprt->xpt_deferred)) {
1279 		dr = list_entry(xprt->xpt_deferred.next,
1280 				struct svc_deferred_req,
1281 				handle.recent);
1282 		list_del_init(&dr->handle.recent);
1283 		trace_svc_revisit_deferred(dr);
1284 	} else
1285 		clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1286 	spin_unlock(&xprt->xpt_lock);
1287 	return dr;
1288 }
1289 
1290 /**
1291  * svc_find_xprt - find an RPC transport instance
1292  * @serv: pointer to svc_serv to search
1293  * @xcl_name: C string containing transport's class name
1294  * @net: owner net pointer
1295  * @af: Address family of transport's local address
1296  * @port: transport's IP port number
1297  *
1298  * Return the transport instance pointer for the endpoint accepting
1299  * connections/peer traffic from the specified transport class,
1300  * address family and port.
1301  *
1302  * Specifying 0 for the address family or port is effectively a
1303  * wild-card, and will result in matching the first transport in the
1304  * service's list that has a matching class name.
1305  */
1306 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1307 			       struct net *net, const sa_family_t af,
1308 			       const unsigned short port)
1309 {
1310 	struct svc_xprt *xprt;
1311 	struct svc_xprt *found = NULL;
1312 
1313 	/* Sanity check the args */
1314 	if (serv == NULL || xcl_name == NULL)
1315 		return found;
1316 
1317 	spin_lock_bh(&serv->sv_lock);
1318 	list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1319 		if (xprt->xpt_net != net)
1320 			continue;
1321 		if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1322 			continue;
1323 		if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1324 			continue;
1325 		if (port != 0 && port != svc_xprt_local_port(xprt))
1326 			continue;
1327 		found = xprt;
1328 		svc_xprt_get(xprt);
1329 		break;
1330 	}
1331 	spin_unlock_bh(&serv->sv_lock);
1332 	return found;
1333 }
1334 EXPORT_SYMBOL_GPL(svc_find_xprt);
1335 
1336 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1337 			     char *pos, int remaining)
1338 {
1339 	int len;
1340 
1341 	len = snprintf(pos, remaining, "%s %u\n",
1342 			xprt->xpt_class->xcl_name,
1343 			svc_xprt_local_port(xprt));
1344 	if (len >= remaining)
1345 		return -ENAMETOOLONG;
1346 	return len;
1347 }
1348 
1349 /**
1350  * svc_xprt_names - format a buffer with a list of transport names
1351  * @serv: pointer to an RPC service
1352  * @buf: pointer to a buffer to be filled in
1353  * @buflen: length of buffer to be filled in
1354  *
1355  * Fills in @buf with a string containing a list of transport names,
1356  * each name terminated with '\n'.
1357  *
1358  * Returns positive length of the filled-in string on success; otherwise
1359  * a negative errno value is returned if an error occurs.
1360  */
1361 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1362 {
1363 	struct svc_xprt *xprt;
1364 	int len, totlen;
1365 	char *pos;
1366 
1367 	/* Sanity check args */
1368 	if (!serv)
1369 		return 0;
1370 
1371 	spin_lock_bh(&serv->sv_lock);
1372 
1373 	pos = buf;
1374 	totlen = 0;
1375 	list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1376 		len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1377 		if (len < 0) {
1378 			*buf = '\0';
1379 			totlen = len;
1380 		}
1381 		if (len <= 0)
1382 			break;
1383 
1384 		pos += len;
1385 		totlen += len;
1386 	}
1387 
1388 	spin_unlock_bh(&serv->sv_lock);
1389 	return totlen;
1390 }
1391 EXPORT_SYMBOL_GPL(svc_xprt_names);
1392 
1393 
1394 /*----------------------------------------------------------------------------*/
1395 
1396 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1397 {
1398 	unsigned int pidx = (unsigned int)*pos;
1399 	struct svc_serv *serv = m->private;
1400 
1401 	dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1402 
1403 	if (!pidx)
1404 		return SEQ_START_TOKEN;
1405 	return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1406 }
1407 
1408 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1409 {
1410 	struct svc_pool *pool = p;
1411 	struct svc_serv *serv = m->private;
1412 
1413 	dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1414 
1415 	if (p == SEQ_START_TOKEN) {
1416 		pool = &serv->sv_pools[0];
1417 	} else {
1418 		unsigned int pidx = (pool - &serv->sv_pools[0]);
1419 		if (pidx < serv->sv_nrpools-1)
1420 			pool = &serv->sv_pools[pidx+1];
1421 		else
1422 			pool = NULL;
1423 	}
1424 	++*pos;
1425 	return pool;
1426 }
1427 
1428 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1429 {
1430 }
1431 
1432 static int svc_pool_stats_show(struct seq_file *m, void *p)
1433 {
1434 	struct svc_pool *pool = p;
1435 
1436 	if (p == SEQ_START_TOKEN) {
1437 		seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1438 		return 0;
1439 	}
1440 
1441 	seq_printf(m, "%u %lu %lu %lu %lu\n",
1442 		pool->sp_id,
1443 		(unsigned long)atomic_long_read(&pool->sp_stats.packets),
1444 		pool->sp_stats.sockets_queued,
1445 		(unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1446 		(unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1447 
1448 	return 0;
1449 }
1450 
1451 static const struct seq_operations svc_pool_stats_seq_ops = {
1452 	.start	= svc_pool_stats_start,
1453 	.next	= svc_pool_stats_next,
1454 	.stop	= svc_pool_stats_stop,
1455 	.show	= svc_pool_stats_show,
1456 };
1457 
1458 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1459 {
1460 	int err;
1461 
1462 	err = seq_open(file, &svc_pool_stats_seq_ops);
1463 	if (!err)
1464 		((struct seq_file *) file->private_data)->private = serv;
1465 	return err;
1466 }
1467 EXPORT_SYMBOL(svc_pool_stats_open);
1468 
1469 /*----------------------------------------------------------------------------*/
1470