xref: /linux/net/sunrpc/svc_xprt.c (revision e7c22eeaff8565d9a8374f320238c251ca31480b)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * linux/net/sunrpc/svc_xprt.c
4  *
5  * Author: Tom Tucker <tom@opengridcomputing.com>
6  */
7 
8 #include <linux/sched.h>
9 #include <linux/errno.h>
10 #include <linux/freezer.h>
11 #include <linux/kthread.h>
12 #include <linux/slab.h>
13 #include <net/sock.h>
14 #include <linux/sunrpc/addr.h>
15 #include <linux/sunrpc/stats.h>
16 #include <linux/sunrpc/svc_xprt.h>
17 #include <linux/sunrpc/svcsock.h>
18 #include <linux/sunrpc/xprt.h>
19 #include <linux/module.h>
20 #include <linux/netdevice.h>
21 #include <trace/events/sunrpc.h>
22 
23 #define RPCDBG_FACILITY	RPCDBG_SVCXPRT
24 
25 static unsigned int svc_rpc_per_connection_limit __read_mostly;
26 module_param(svc_rpc_per_connection_limit, uint, 0644);
27 
28 
29 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
30 static int svc_deferred_recv(struct svc_rqst *rqstp);
31 static struct cache_deferred_req *svc_defer(struct cache_req *req);
32 static void svc_age_temp_xprts(struct timer_list *t);
33 static void svc_delete_xprt(struct svc_xprt *xprt);
34 
35 /* apparently the "standard" is that clients close
36  * idle connections after 5 minutes, servers after
37  * 6 minutes
38  *   http://nfsv4bat.org/Documents/ConnectAThon/1996/nfstcp.pdf
39  */
40 static int svc_conn_age_period = 6*60;
41 
42 /* List of registered transport classes */
43 static DEFINE_SPINLOCK(svc_xprt_class_lock);
44 static LIST_HEAD(svc_xprt_class_list);
45 
46 /* SMP locking strategy:
47  *
48  *	svc_pool->sp_lock protects most of the fields of that pool.
49  *	svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
50  *	when both need to be taken (rare), svc_serv->sv_lock is first.
51  *	The "service mutex" protects svc_serv->sv_nrthread.
52  *	svc_sock->sk_lock protects the svc_sock->sk_deferred list
53  *             and the ->sk_info_authunix cache.
54  *
55  *	The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
56  *	enqueued multiply. During normal transport processing this bit
57  *	is set by svc_xprt_enqueue and cleared by svc_xprt_received.
58  *	Providers should not manipulate this bit directly.
59  *
60  *	Some flags can be set to certain values at any time
61  *	providing that certain rules are followed:
62  *
63  *	XPT_CONN, XPT_DATA:
64  *		- Can be set or cleared at any time.
65  *		- After a set, svc_xprt_enqueue must be called to enqueue
66  *		  the transport for processing.
67  *		- After a clear, the transport must be read/accepted.
68  *		  If this succeeds, it must be set again.
69  *	XPT_CLOSE:
70  *		- Can set at any time. It is never cleared.
71  *      XPT_DEAD:
72  *		- Can only be set while XPT_BUSY is held which ensures
73  *		  that no other thread will be using the transport or will
74  *		  try to set XPT_DEAD.
75  */
76 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
77 {
78 	struct svc_xprt_class *cl;
79 	int res = -EEXIST;
80 
81 	dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
82 
83 	INIT_LIST_HEAD(&xcl->xcl_list);
84 	spin_lock(&svc_xprt_class_lock);
85 	/* Make sure there isn't already a class with the same name */
86 	list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
87 		if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
88 			goto out;
89 	}
90 	list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
91 	res = 0;
92 out:
93 	spin_unlock(&svc_xprt_class_lock);
94 	return res;
95 }
96 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
97 
98 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
99 {
100 	dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
101 	spin_lock(&svc_xprt_class_lock);
102 	list_del_init(&xcl->xcl_list);
103 	spin_unlock(&svc_xprt_class_lock);
104 }
105 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
106 
107 /**
108  * svc_print_xprts - Format the transport list for printing
109  * @buf: target buffer for formatted address
110  * @maxlen: length of target buffer
111  *
112  * Fills in @buf with a string containing a list of transport names, each name
113  * terminated with '\n'. If the buffer is too small, some entries may be
114  * missing, but it is guaranteed that all lines in the output buffer are
115  * complete.
116  *
117  * Returns positive length of the filled-in string.
118  */
119 int svc_print_xprts(char *buf, int maxlen)
120 {
121 	struct svc_xprt_class *xcl;
122 	char tmpstr[80];
123 	int len = 0;
124 	buf[0] = '\0';
125 
126 	spin_lock(&svc_xprt_class_lock);
127 	list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
128 		int slen;
129 
130 		slen = snprintf(tmpstr, sizeof(tmpstr), "%s %d\n",
131 				xcl->xcl_name, xcl->xcl_max_payload);
132 		if (slen >= sizeof(tmpstr) || len + slen >= maxlen)
133 			break;
134 		len += slen;
135 		strcat(buf, tmpstr);
136 	}
137 	spin_unlock(&svc_xprt_class_lock);
138 
139 	return len;
140 }
141 
142 static void svc_xprt_free(struct kref *kref)
143 {
144 	struct svc_xprt *xprt =
145 		container_of(kref, struct svc_xprt, xpt_ref);
146 	struct module *owner = xprt->xpt_class->xcl_owner;
147 	if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
148 		svcauth_unix_info_release(xprt);
149 	put_cred(xprt->xpt_cred);
150 	put_net(xprt->xpt_net);
151 	/* See comment on corresponding get in xs_setup_bc_tcp(): */
152 	if (xprt->xpt_bc_xprt)
153 		xprt_put(xprt->xpt_bc_xprt);
154 	if (xprt->xpt_bc_xps)
155 		xprt_switch_put(xprt->xpt_bc_xps);
156 	trace_svc_xprt_free(xprt);
157 	xprt->xpt_ops->xpo_free(xprt);
158 	module_put(owner);
159 }
160 
161 void svc_xprt_put(struct svc_xprt *xprt)
162 {
163 	kref_put(&xprt->xpt_ref, svc_xprt_free);
164 }
165 EXPORT_SYMBOL_GPL(svc_xprt_put);
166 
167 /*
168  * Called by transport drivers to initialize the transport independent
169  * portion of the transport instance.
170  */
171 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
172 		   struct svc_xprt *xprt, struct svc_serv *serv)
173 {
174 	memset(xprt, 0, sizeof(*xprt));
175 	xprt->xpt_class = xcl;
176 	xprt->xpt_ops = xcl->xcl_ops;
177 	kref_init(&xprt->xpt_ref);
178 	xprt->xpt_server = serv;
179 	INIT_LIST_HEAD(&xprt->xpt_list);
180 	INIT_LIST_HEAD(&xprt->xpt_ready);
181 	INIT_LIST_HEAD(&xprt->xpt_deferred);
182 	INIT_LIST_HEAD(&xprt->xpt_users);
183 	mutex_init(&xprt->xpt_mutex);
184 	spin_lock_init(&xprt->xpt_lock);
185 	set_bit(XPT_BUSY, &xprt->xpt_flags);
186 	xprt->xpt_net = get_net(net);
187 	strcpy(xprt->xpt_remotebuf, "uninitialized");
188 }
189 EXPORT_SYMBOL_GPL(svc_xprt_init);
190 
191 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
192 					 struct svc_serv *serv,
193 					 struct net *net,
194 					 const int family,
195 					 const unsigned short port,
196 					 int flags)
197 {
198 	struct sockaddr_in sin = {
199 		.sin_family		= AF_INET,
200 		.sin_addr.s_addr	= htonl(INADDR_ANY),
201 		.sin_port		= htons(port),
202 	};
203 #if IS_ENABLED(CONFIG_IPV6)
204 	struct sockaddr_in6 sin6 = {
205 		.sin6_family		= AF_INET6,
206 		.sin6_addr		= IN6ADDR_ANY_INIT,
207 		.sin6_port		= htons(port),
208 	};
209 #endif
210 	struct svc_xprt *xprt;
211 	struct sockaddr *sap;
212 	size_t len;
213 
214 	switch (family) {
215 	case PF_INET:
216 		sap = (struct sockaddr *)&sin;
217 		len = sizeof(sin);
218 		break;
219 #if IS_ENABLED(CONFIG_IPV6)
220 	case PF_INET6:
221 		sap = (struct sockaddr *)&sin6;
222 		len = sizeof(sin6);
223 		break;
224 #endif
225 	default:
226 		return ERR_PTR(-EAFNOSUPPORT);
227 	}
228 
229 	xprt = xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
230 	if (IS_ERR(xprt))
231 		trace_svc_xprt_create_err(serv->sv_program->pg_name,
232 					  xcl->xcl_name, sap, xprt);
233 	return xprt;
234 }
235 
236 /*
237  * svc_xprt_received conditionally queues the transport for processing
238  * by another thread. The caller must hold the XPT_BUSY bit and must
239  * not thereafter touch transport data.
240  *
241  * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
242  * insufficient) data.
243  */
244 static void svc_xprt_received(struct svc_xprt *xprt)
245 {
246 	if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
247 		WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
248 		return;
249 	}
250 
251 	/* As soon as we clear busy, the xprt could be closed and
252 	 * 'put', so we need a reference to call svc_enqueue_xprt with:
253 	 */
254 	svc_xprt_get(xprt);
255 	smp_mb__before_atomic();
256 	clear_bit(XPT_BUSY, &xprt->xpt_flags);
257 	xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
258 	svc_xprt_put(xprt);
259 }
260 
261 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
262 {
263 	clear_bit(XPT_TEMP, &new->xpt_flags);
264 	spin_lock_bh(&serv->sv_lock);
265 	list_add(&new->xpt_list, &serv->sv_permsocks);
266 	spin_unlock_bh(&serv->sv_lock);
267 	svc_xprt_received(new);
268 }
269 
270 static int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
271 			    struct net *net, const int family,
272 			    const unsigned short port, int flags,
273 			    const struct cred *cred)
274 {
275 	struct svc_xprt_class *xcl;
276 
277 	spin_lock(&svc_xprt_class_lock);
278 	list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
279 		struct svc_xprt *newxprt;
280 		unsigned short newport;
281 
282 		if (strcmp(xprt_name, xcl->xcl_name))
283 			continue;
284 
285 		if (!try_module_get(xcl->xcl_owner))
286 			goto err;
287 
288 		spin_unlock(&svc_xprt_class_lock);
289 		newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
290 		if (IS_ERR(newxprt)) {
291 			module_put(xcl->xcl_owner);
292 			return PTR_ERR(newxprt);
293 		}
294 		newxprt->xpt_cred = get_cred(cred);
295 		svc_add_new_perm_xprt(serv, newxprt);
296 		newport = svc_xprt_local_port(newxprt);
297 		return newport;
298 	}
299  err:
300 	spin_unlock(&svc_xprt_class_lock);
301 	/* This errno is exposed to user space.  Provide a reasonable
302 	 * perror msg for a bad transport. */
303 	return -EPROTONOSUPPORT;
304 }
305 
306 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
307 		    struct net *net, const int family,
308 		    const unsigned short port, int flags,
309 		    const struct cred *cred)
310 {
311 	int err;
312 
313 	err = _svc_create_xprt(serv, xprt_name, net, family, port, flags, cred);
314 	if (err == -EPROTONOSUPPORT) {
315 		request_module("svc%s", xprt_name);
316 		err = _svc_create_xprt(serv, xprt_name, net, family, port, flags, cred);
317 	}
318 	return err;
319 }
320 EXPORT_SYMBOL_GPL(svc_create_xprt);
321 
322 /*
323  * Copy the local and remote xprt addresses to the rqstp structure
324  */
325 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
326 {
327 	memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
328 	rqstp->rq_addrlen = xprt->xpt_remotelen;
329 
330 	/*
331 	 * Destination address in request is needed for binding the
332 	 * source address in RPC replies/callbacks later.
333 	 */
334 	memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
335 	rqstp->rq_daddrlen = xprt->xpt_locallen;
336 }
337 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
338 
339 /**
340  * svc_print_addr - Format rq_addr field for printing
341  * @rqstp: svc_rqst struct containing address to print
342  * @buf: target buffer for formatted address
343  * @len: length of target buffer
344  *
345  */
346 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
347 {
348 	return __svc_print_addr(svc_addr(rqstp), buf, len);
349 }
350 EXPORT_SYMBOL_GPL(svc_print_addr);
351 
352 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
353 {
354 	unsigned int limit = svc_rpc_per_connection_limit;
355 	int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
356 
357 	return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
358 }
359 
360 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
361 {
362 	if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
363 		if (!svc_xprt_slots_in_range(xprt))
364 			return false;
365 		atomic_inc(&xprt->xpt_nr_rqsts);
366 		set_bit(RQ_DATA, &rqstp->rq_flags);
367 	}
368 	return true;
369 }
370 
371 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
372 {
373 	struct svc_xprt	*xprt = rqstp->rq_xprt;
374 	if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
375 		atomic_dec(&xprt->xpt_nr_rqsts);
376 		smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
377 		svc_xprt_enqueue(xprt);
378 	}
379 }
380 
381 static bool svc_xprt_ready(struct svc_xprt *xprt)
382 {
383 	unsigned long xpt_flags;
384 
385 	/*
386 	 * If another cpu has recently updated xpt_flags,
387 	 * sk_sock->flags, xpt_reserved, or xpt_nr_rqsts, we need to
388 	 * know about it; otherwise it's possible that both that cpu and
389 	 * this one could call svc_xprt_enqueue() without either
390 	 * svc_xprt_enqueue() recognizing that the conditions below
391 	 * are satisfied, and we could stall indefinitely:
392 	 */
393 	smp_rmb();
394 	xpt_flags = READ_ONCE(xprt->xpt_flags);
395 
396 	if (xpt_flags & (BIT(XPT_CONN) | BIT(XPT_CLOSE)))
397 		return true;
398 	if (xpt_flags & (BIT(XPT_DATA) | BIT(XPT_DEFERRED))) {
399 		if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
400 		    svc_xprt_slots_in_range(xprt))
401 			return true;
402 		trace_svc_xprt_no_write_space(xprt);
403 		return false;
404 	}
405 	return false;
406 }
407 
408 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
409 {
410 	struct svc_pool *pool;
411 	struct svc_rqst	*rqstp = NULL;
412 	int cpu;
413 
414 	if (!svc_xprt_ready(xprt))
415 		return;
416 
417 	/* Mark transport as busy. It will remain in this state until
418 	 * the provider calls svc_xprt_received. We update XPT_BUSY
419 	 * atomically because it also guards against trying to enqueue
420 	 * the transport twice.
421 	 */
422 	if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
423 		return;
424 
425 	cpu = get_cpu();
426 	pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
427 
428 	atomic_long_inc(&pool->sp_stats.packets);
429 
430 	spin_lock_bh(&pool->sp_lock);
431 	list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
432 	pool->sp_stats.sockets_queued++;
433 	spin_unlock_bh(&pool->sp_lock);
434 
435 	/* find a thread for this xprt */
436 	rcu_read_lock();
437 	list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
438 		if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
439 			continue;
440 		atomic_long_inc(&pool->sp_stats.threads_woken);
441 		rqstp->rq_qtime = ktime_get();
442 		wake_up_process(rqstp->rq_task);
443 		goto out_unlock;
444 	}
445 	set_bit(SP_CONGESTED, &pool->sp_flags);
446 	rqstp = NULL;
447 out_unlock:
448 	rcu_read_unlock();
449 	put_cpu();
450 	trace_svc_xprt_do_enqueue(xprt, rqstp);
451 }
452 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
453 
454 /*
455  * Queue up a transport with data pending. If there are idle nfsd
456  * processes, wake 'em up.
457  *
458  */
459 void svc_xprt_enqueue(struct svc_xprt *xprt)
460 {
461 	if (test_bit(XPT_BUSY, &xprt->xpt_flags))
462 		return;
463 	xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
464 }
465 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
466 
467 /*
468  * Dequeue the first transport, if there is one.
469  */
470 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
471 {
472 	struct svc_xprt	*xprt = NULL;
473 
474 	if (list_empty(&pool->sp_sockets))
475 		goto out;
476 
477 	spin_lock_bh(&pool->sp_lock);
478 	if (likely(!list_empty(&pool->sp_sockets))) {
479 		xprt = list_first_entry(&pool->sp_sockets,
480 					struct svc_xprt, xpt_ready);
481 		list_del_init(&xprt->xpt_ready);
482 		svc_xprt_get(xprt);
483 	}
484 	spin_unlock_bh(&pool->sp_lock);
485 out:
486 	return xprt;
487 }
488 
489 /**
490  * svc_reserve - change the space reserved for the reply to a request.
491  * @rqstp:  The request in question
492  * @space: new max space to reserve
493  *
494  * Each request reserves some space on the output queue of the transport
495  * to make sure the reply fits.  This function reduces that reserved
496  * space to be the amount of space used already, plus @space.
497  *
498  */
499 void svc_reserve(struct svc_rqst *rqstp, int space)
500 {
501 	struct svc_xprt *xprt = rqstp->rq_xprt;
502 
503 	space += rqstp->rq_res.head[0].iov_len;
504 
505 	if (xprt && space < rqstp->rq_reserved) {
506 		atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
507 		rqstp->rq_reserved = space;
508 		smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
509 		svc_xprt_enqueue(xprt);
510 	}
511 }
512 EXPORT_SYMBOL_GPL(svc_reserve);
513 
514 static void svc_xprt_release(struct svc_rqst *rqstp)
515 {
516 	struct svc_xprt	*xprt = rqstp->rq_xprt;
517 
518 	xprt->xpt_ops->xpo_release_rqst(rqstp);
519 
520 	kfree(rqstp->rq_deferred);
521 	rqstp->rq_deferred = NULL;
522 
523 	svc_free_res_pages(rqstp);
524 	rqstp->rq_res.page_len = 0;
525 	rqstp->rq_res.page_base = 0;
526 
527 	/* Reset response buffer and release
528 	 * the reservation.
529 	 * But first, check that enough space was reserved
530 	 * for the reply, otherwise we have a bug!
531 	 */
532 	if ((rqstp->rq_res.len) >  rqstp->rq_reserved)
533 		printk(KERN_ERR "RPC request reserved %d but used %d\n",
534 		       rqstp->rq_reserved,
535 		       rqstp->rq_res.len);
536 
537 	rqstp->rq_res.head[0].iov_len = 0;
538 	svc_reserve(rqstp, 0);
539 	svc_xprt_release_slot(rqstp);
540 	rqstp->rq_xprt = NULL;
541 	svc_xprt_put(xprt);
542 }
543 
544 /*
545  * Some svc_serv's will have occasional work to do, even when a xprt is not
546  * waiting to be serviced. This function is there to "kick" a task in one of
547  * those services so that it can wake up and do that work. Note that we only
548  * bother with pool 0 as we don't need to wake up more than one thread for
549  * this purpose.
550  */
551 void svc_wake_up(struct svc_serv *serv)
552 {
553 	struct svc_rqst	*rqstp;
554 	struct svc_pool *pool;
555 
556 	pool = &serv->sv_pools[0];
557 
558 	rcu_read_lock();
559 	list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
560 		/* skip any that aren't queued */
561 		if (test_bit(RQ_BUSY, &rqstp->rq_flags))
562 			continue;
563 		rcu_read_unlock();
564 		wake_up_process(rqstp->rq_task);
565 		trace_svc_wake_up(rqstp->rq_task->pid);
566 		return;
567 	}
568 	rcu_read_unlock();
569 
570 	/* No free entries available */
571 	set_bit(SP_TASK_PENDING, &pool->sp_flags);
572 	smp_wmb();
573 	trace_svc_wake_up(0);
574 }
575 EXPORT_SYMBOL_GPL(svc_wake_up);
576 
577 int svc_port_is_privileged(struct sockaddr *sin)
578 {
579 	switch (sin->sa_family) {
580 	case AF_INET:
581 		return ntohs(((struct sockaddr_in *)sin)->sin_port)
582 			< PROT_SOCK;
583 	case AF_INET6:
584 		return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
585 			< PROT_SOCK;
586 	default:
587 		return 0;
588 	}
589 }
590 
591 /*
592  * Make sure that we don't have too many active connections. If we have,
593  * something must be dropped. It's not clear what will happen if we allow
594  * "too many" connections, but when dealing with network-facing software,
595  * we have to code defensively. Here we do that by imposing hard limits.
596  *
597  * There's no point in trying to do random drop here for DoS
598  * prevention. The NFS clients does 1 reconnect in 15 seconds. An
599  * attacker can easily beat that.
600  *
601  * The only somewhat efficient mechanism would be if drop old
602  * connections from the same IP first. But right now we don't even
603  * record the client IP in svc_sock.
604  *
605  * single-threaded services that expect a lot of clients will probably
606  * need to set sv_maxconn to override the default value which is based
607  * on the number of threads
608  */
609 static void svc_check_conn_limits(struct svc_serv *serv)
610 {
611 	unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
612 				(serv->sv_nrthreads+3) * 20;
613 
614 	if (serv->sv_tmpcnt > limit) {
615 		struct svc_xprt *xprt = NULL;
616 		spin_lock_bh(&serv->sv_lock);
617 		if (!list_empty(&serv->sv_tempsocks)) {
618 			/* Try to help the admin */
619 			net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
620 					       serv->sv_name, serv->sv_maxconn ?
621 					       "max number of connections" :
622 					       "number of threads");
623 			/*
624 			 * Always select the oldest connection. It's not fair,
625 			 * but so is life
626 			 */
627 			xprt = list_entry(serv->sv_tempsocks.prev,
628 					  struct svc_xprt,
629 					  xpt_list);
630 			set_bit(XPT_CLOSE, &xprt->xpt_flags);
631 			svc_xprt_get(xprt);
632 		}
633 		spin_unlock_bh(&serv->sv_lock);
634 
635 		if (xprt) {
636 			svc_xprt_enqueue(xprt);
637 			svc_xprt_put(xprt);
638 		}
639 	}
640 }
641 
642 static int svc_alloc_arg(struct svc_rqst *rqstp)
643 {
644 	struct svc_serv *serv = rqstp->rq_server;
645 	struct xdr_buf *arg;
646 	int pages;
647 	int i;
648 
649 	/* now allocate needed pages.  If we get a failure, sleep briefly */
650 	pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
651 	if (pages > RPCSVC_MAXPAGES) {
652 		pr_warn_once("svc: warning: pages=%u > RPCSVC_MAXPAGES=%lu\n",
653 			     pages, RPCSVC_MAXPAGES);
654 		/* use as many pages as possible */
655 		pages = RPCSVC_MAXPAGES;
656 	}
657 	for (i = 0; i < pages ; i++)
658 		while (rqstp->rq_pages[i] == NULL) {
659 			struct page *p = alloc_page(GFP_KERNEL);
660 			if (!p) {
661 				set_current_state(TASK_INTERRUPTIBLE);
662 				if (signalled() || kthread_should_stop()) {
663 					set_current_state(TASK_RUNNING);
664 					return -EINTR;
665 				}
666 				schedule_timeout(msecs_to_jiffies(500));
667 			}
668 			rqstp->rq_pages[i] = p;
669 		}
670 	rqstp->rq_page_end = &rqstp->rq_pages[i];
671 	rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
672 
673 	/* Make arg->head point to first page and arg->pages point to rest */
674 	arg = &rqstp->rq_arg;
675 	arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
676 	arg->head[0].iov_len = PAGE_SIZE;
677 	arg->pages = rqstp->rq_pages + 1;
678 	arg->page_base = 0;
679 	/* save at least one page for response */
680 	arg->page_len = (pages-2)*PAGE_SIZE;
681 	arg->len = (pages-1)*PAGE_SIZE;
682 	arg->tail[0].iov_len = 0;
683 	return 0;
684 }
685 
686 static bool
687 rqst_should_sleep(struct svc_rqst *rqstp)
688 {
689 	struct svc_pool		*pool = rqstp->rq_pool;
690 
691 	/* did someone call svc_wake_up? */
692 	if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
693 		return false;
694 
695 	/* was a socket queued? */
696 	if (!list_empty(&pool->sp_sockets))
697 		return false;
698 
699 	/* are we shutting down? */
700 	if (signalled() || kthread_should_stop())
701 		return false;
702 
703 	/* are we freezing? */
704 	if (freezing(current))
705 		return false;
706 
707 	return true;
708 }
709 
710 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
711 {
712 	struct svc_pool		*pool = rqstp->rq_pool;
713 	long			time_left = 0;
714 
715 	/* rq_xprt should be clear on entry */
716 	WARN_ON_ONCE(rqstp->rq_xprt);
717 
718 	rqstp->rq_xprt = svc_xprt_dequeue(pool);
719 	if (rqstp->rq_xprt)
720 		goto out_found;
721 
722 	/*
723 	 * We have to be able to interrupt this wait
724 	 * to bring down the daemons ...
725 	 */
726 	set_current_state(TASK_INTERRUPTIBLE);
727 	smp_mb__before_atomic();
728 	clear_bit(SP_CONGESTED, &pool->sp_flags);
729 	clear_bit(RQ_BUSY, &rqstp->rq_flags);
730 	smp_mb__after_atomic();
731 
732 	if (likely(rqst_should_sleep(rqstp)))
733 		time_left = schedule_timeout(timeout);
734 	else
735 		__set_current_state(TASK_RUNNING);
736 
737 	try_to_freeze();
738 
739 	set_bit(RQ_BUSY, &rqstp->rq_flags);
740 	smp_mb__after_atomic();
741 	rqstp->rq_xprt = svc_xprt_dequeue(pool);
742 	if (rqstp->rq_xprt)
743 		goto out_found;
744 
745 	if (!time_left)
746 		atomic_long_inc(&pool->sp_stats.threads_timedout);
747 
748 	if (signalled() || kthread_should_stop())
749 		return ERR_PTR(-EINTR);
750 	return ERR_PTR(-EAGAIN);
751 out_found:
752 	/* Normally we will wait up to 5 seconds for any required
753 	 * cache information to be provided.
754 	 */
755 	if (!test_bit(SP_CONGESTED, &pool->sp_flags))
756 		rqstp->rq_chandle.thread_wait = 5*HZ;
757 	else
758 		rqstp->rq_chandle.thread_wait = 1*HZ;
759 	trace_svc_xprt_dequeue(rqstp);
760 	return rqstp->rq_xprt;
761 }
762 
763 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
764 {
765 	spin_lock_bh(&serv->sv_lock);
766 	set_bit(XPT_TEMP, &newxpt->xpt_flags);
767 	list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
768 	serv->sv_tmpcnt++;
769 	if (serv->sv_temptimer.function == NULL) {
770 		/* setup timer to age temp transports */
771 		serv->sv_temptimer.function = svc_age_temp_xprts;
772 		mod_timer(&serv->sv_temptimer,
773 			  jiffies + svc_conn_age_period * HZ);
774 	}
775 	spin_unlock_bh(&serv->sv_lock);
776 	svc_xprt_received(newxpt);
777 }
778 
779 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
780 {
781 	struct svc_serv *serv = rqstp->rq_server;
782 	int len = 0;
783 
784 	if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
785 		if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
786 			xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
787 		svc_delete_xprt(xprt);
788 		/* Leave XPT_BUSY set on the dead xprt: */
789 		goto out;
790 	}
791 	if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
792 		struct svc_xprt *newxpt;
793 		/*
794 		 * We know this module_get will succeed because the
795 		 * listener holds a reference too
796 		 */
797 		__module_get(xprt->xpt_class->xcl_owner);
798 		svc_check_conn_limits(xprt->xpt_server);
799 		newxpt = xprt->xpt_ops->xpo_accept(xprt);
800 		if (newxpt) {
801 			newxpt->xpt_cred = get_cred(xprt->xpt_cred);
802 			svc_add_new_temp_xprt(serv, newxpt);
803 			trace_svc_xprt_accept(newxpt, serv->sv_name);
804 		} else
805 			module_put(xprt->xpt_class->xcl_owner);
806 	} else if (svc_xprt_reserve_slot(rqstp, xprt)) {
807 		/* XPT_DATA|XPT_DEFERRED case: */
808 		dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
809 			rqstp, rqstp->rq_pool->sp_id, xprt,
810 			kref_read(&xprt->xpt_ref));
811 		rqstp->rq_deferred = svc_deferred_dequeue(xprt);
812 		if (rqstp->rq_deferred)
813 			len = svc_deferred_recv(rqstp);
814 		else
815 			len = xprt->xpt_ops->xpo_recvfrom(rqstp);
816 		rqstp->rq_stime = ktime_get();
817 		rqstp->rq_reserved = serv->sv_max_mesg;
818 		atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
819 	}
820 	/* clear XPT_BUSY: */
821 	svc_xprt_received(xprt);
822 out:
823 	trace_svc_handle_xprt(xprt, len);
824 	return len;
825 }
826 
827 /*
828  * Receive the next request on any transport.  This code is carefully
829  * organised not to touch any cachelines in the shared svc_serv
830  * structure, only cachelines in the local svc_pool.
831  */
832 int svc_recv(struct svc_rqst *rqstp, long timeout)
833 {
834 	struct svc_xprt		*xprt = NULL;
835 	struct svc_serv		*serv = rqstp->rq_server;
836 	int			len, err;
837 
838 	err = svc_alloc_arg(rqstp);
839 	if (err)
840 		goto out;
841 
842 	try_to_freeze();
843 	cond_resched();
844 	err = -EINTR;
845 	if (signalled() || kthread_should_stop())
846 		goto out;
847 
848 	xprt = svc_get_next_xprt(rqstp, timeout);
849 	if (IS_ERR(xprt)) {
850 		err = PTR_ERR(xprt);
851 		goto out;
852 	}
853 
854 	len = svc_handle_xprt(rqstp, xprt);
855 
856 	/* No data, incomplete (TCP) read, or accept() */
857 	err = -EAGAIN;
858 	if (len <= 0)
859 		goto out_release;
860 
861 	clear_bit(XPT_OLD, &xprt->xpt_flags);
862 
863 	xprt->xpt_ops->xpo_secure_port(rqstp);
864 	rqstp->rq_chandle.defer = svc_defer;
865 	rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
866 
867 	if (serv->sv_stats)
868 		serv->sv_stats->netcnt++;
869 	trace_svc_xdr_recvfrom(rqstp, &rqstp->rq_arg);
870 	return len;
871 out_release:
872 	rqstp->rq_res.len = 0;
873 	svc_xprt_release(rqstp);
874 out:
875 	return err;
876 }
877 EXPORT_SYMBOL_GPL(svc_recv);
878 
879 /*
880  * Drop request
881  */
882 void svc_drop(struct svc_rqst *rqstp)
883 {
884 	trace_svc_drop(rqstp);
885 	svc_xprt_release(rqstp);
886 }
887 EXPORT_SYMBOL_GPL(svc_drop);
888 
889 /*
890  * Return reply to client.
891  */
892 int svc_send(struct svc_rqst *rqstp)
893 {
894 	struct svc_xprt	*xprt;
895 	int		len = -EFAULT;
896 	struct xdr_buf	*xb;
897 
898 	xprt = rqstp->rq_xprt;
899 	if (!xprt)
900 		goto out;
901 
902 	/* calculate over-all length */
903 	xb = &rqstp->rq_res;
904 	xb->len = xb->head[0].iov_len +
905 		xb->page_len +
906 		xb->tail[0].iov_len;
907 	trace_svc_xdr_sendto(rqstp, xb);
908 	trace_svc_stats_latency(rqstp);
909 
910 	len = xprt->xpt_ops->xpo_sendto(rqstp);
911 
912 	trace_svc_send(rqstp, len);
913 	svc_xprt_release(rqstp);
914 
915 	if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
916 		len = 0;
917 out:
918 	return len;
919 }
920 
921 /*
922  * Timer function to close old temporary transports, using
923  * a mark-and-sweep algorithm.
924  */
925 static void svc_age_temp_xprts(struct timer_list *t)
926 {
927 	struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
928 	struct svc_xprt *xprt;
929 	struct list_head *le, *next;
930 
931 	dprintk("svc_age_temp_xprts\n");
932 
933 	if (!spin_trylock_bh(&serv->sv_lock)) {
934 		/* busy, try again 1 sec later */
935 		dprintk("svc_age_temp_xprts: busy\n");
936 		mod_timer(&serv->sv_temptimer, jiffies + HZ);
937 		return;
938 	}
939 
940 	list_for_each_safe(le, next, &serv->sv_tempsocks) {
941 		xprt = list_entry(le, struct svc_xprt, xpt_list);
942 
943 		/* First time through, just mark it OLD. Second time
944 		 * through, close it. */
945 		if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
946 			continue;
947 		if (kref_read(&xprt->xpt_ref) > 1 ||
948 		    test_bit(XPT_BUSY, &xprt->xpt_flags))
949 			continue;
950 		list_del_init(le);
951 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
952 		dprintk("queuing xprt %p for closing\n", xprt);
953 
954 		/* a thread will dequeue and close it soon */
955 		svc_xprt_enqueue(xprt);
956 	}
957 	spin_unlock_bh(&serv->sv_lock);
958 
959 	mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
960 }
961 
962 /* Close temporary transports whose xpt_local matches server_addr immediately
963  * instead of waiting for them to be picked up by the timer.
964  *
965  * This is meant to be called from a notifier_block that runs when an ip
966  * address is deleted.
967  */
968 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
969 {
970 	struct svc_xprt *xprt;
971 	struct list_head *le, *next;
972 	LIST_HEAD(to_be_closed);
973 
974 	spin_lock_bh(&serv->sv_lock);
975 	list_for_each_safe(le, next, &serv->sv_tempsocks) {
976 		xprt = list_entry(le, struct svc_xprt, xpt_list);
977 		if (rpc_cmp_addr(server_addr, (struct sockaddr *)
978 				&xprt->xpt_local)) {
979 			dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
980 			list_move(le, &to_be_closed);
981 		}
982 	}
983 	spin_unlock_bh(&serv->sv_lock);
984 
985 	while (!list_empty(&to_be_closed)) {
986 		le = to_be_closed.next;
987 		list_del_init(le);
988 		xprt = list_entry(le, struct svc_xprt, xpt_list);
989 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
990 		set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
991 		dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
992 				xprt);
993 		svc_xprt_enqueue(xprt);
994 	}
995 }
996 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
997 
998 static void call_xpt_users(struct svc_xprt *xprt)
999 {
1000 	struct svc_xpt_user *u;
1001 
1002 	spin_lock(&xprt->xpt_lock);
1003 	while (!list_empty(&xprt->xpt_users)) {
1004 		u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1005 		list_del_init(&u->list);
1006 		u->callback(u);
1007 	}
1008 	spin_unlock(&xprt->xpt_lock);
1009 }
1010 
1011 /*
1012  * Remove a dead transport
1013  */
1014 static void svc_delete_xprt(struct svc_xprt *xprt)
1015 {
1016 	struct svc_serv	*serv = xprt->xpt_server;
1017 	struct svc_deferred_req *dr;
1018 
1019 	if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1020 		return;
1021 
1022 	trace_svc_xprt_detach(xprt);
1023 	xprt->xpt_ops->xpo_detach(xprt);
1024 	if (xprt->xpt_bc_xprt)
1025 		xprt->xpt_bc_xprt->ops->close(xprt->xpt_bc_xprt);
1026 
1027 	spin_lock_bh(&serv->sv_lock);
1028 	list_del_init(&xprt->xpt_list);
1029 	WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1030 	if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1031 		serv->sv_tmpcnt--;
1032 	spin_unlock_bh(&serv->sv_lock);
1033 
1034 	while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1035 		kfree(dr);
1036 
1037 	call_xpt_users(xprt);
1038 	svc_xprt_put(xprt);
1039 }
1040 
1041 void svc_close_xprt(struct svc_xprt *xprt)
1042 {
1043 	trace_svc_xprt_close(xprt);
1044 	set_bit(XPT_CLOSE, &xprt->xpt_flags);
1045 	if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1046 		/* someone else will have to effect the close */
1047 		return;
1048 	/*
1049 	 * We expect svc_close_xprt() to work even when no threads are
1050 	 * running (e.g., while configuring the server before starting
1051 	 * any threads), so if the transport isn't busy, we delete
1052 	 * it ourself:
1053 	 */
1054 	svc_delete_xprt(xprt);
1055 }
1056 EXPORT_SYMBOL_GPL(svc_close_xprt);
1057 
1058 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1059 {
1060 	struct svc_xprt *xprt;
1061 	int ret = 0;
1062 
1063 	spin_lock(&serv->sv_lock);
1064 	list_for_each_entry(xprt, xprt_list, xpt_list) {
1065 		if (xprt->xpt_net != net)
1066 			continue;
1067 		ret++;
1068 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
1069 		svc_xprt_enqueue(xprt);
1070 	}
1071 	spin_unlock(&serv->sv_lock);
1072 	return ret;
1073 }
1074 
1075 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1076 {
1077 	struct svc_pool *pool;
1078 	struct svc_xprt *xprt;
1079 	struct svc_xprt *tmp;
1080 	int i;
1081 
1082 	for (i = 0; i < serv->sv_nrpools; i++) {
1083 		pool = &serv->sv_pools[i];
1084 
1085 		spin_lock_bh(&pool->sp_lock);
1086 		list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1087 			if (xprt->xpt_net != net)
1088 				continue;
1089 			list_del_init(&xprt->xpt_ready);
1090 			spin_unlock_bh(&pool->sp_lock);
1091 			return xprt;
1092 		}
1093 		spin_unlock_bh(&pool->sp_lock);
1094 	}
1095 	return NULL;
1096 }
1097 
1098 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1099 {
1100 	struct svc_xprt *xprt;
1101 
1102 	while ((xprt = svc_dequeue_net(serv, net))) {
1103 		set_bit(XPT_CLOSE, &xprt->xpt_flags);
1104 		svc_delete_xprt(xprt);
1105 	}
1106 }
1107 
1108 /*
1109  * Server threads may still be running (especially in the case where the
1110  * service is still running in other network namespaces).
1111  *
1112  * So we shut down sockets the same way we would on a running server, by
1113  * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1114  * the close.  In the case there are no such other threads,
1115  * threads running, svc_clean_up_xprts() does a simple version of a
1116  * server's main event loop, and in the case where there are other
1117  * threads, we may need to wait a little while and then check again to
1118  * see if they're done.
1119  */
1120 void svc_close_net(struct svc_serv *serv, struct net *net)
1121 {
1122 	int delay = 0;
1123 
1124 	while (svc_close_list(serv, &serv->sv_permsocks, net) +
1125 	       svc_close_list(serv, &serv->sv_tempsocks, net)) {
1126 
1127 		svc_clean_up_xprts(serv, net);
1128 		msleep(delay++);
1129 	}
1130 }
1131 
1132 /*
1133  * Handle defer and revisit of requests
1134  */
1135 
1136 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1137 {
1138 	struct svc_deferred_req *dr =
1139 		container_of(dreq, struct svc_deferred_req, handle);
1140 	struct svc_xprt *xprt = dr->xprt;
1141 
1142 	spin_lock(&xprt->xpt_lock);
1143 	set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1144 	if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1145 		spin_unlock(&xprt->xpt_lock);
1146 		trace_svc_defer_drop(dr);
1147 		svc_xprt_put(xprt);
1148 		kfree(dr);
1149 		return;
1150 	}
1151 	dr->xprt = NULL;
1152 	list_add(&dr->handle.recent, &xprt->xpt_deferred);
1153 	spin_unlock(&xprt->xpt_lock);
1154 	trace_svc_defer_queue(dr);
1155 	svc_xprt_enqueue(xprt);
1156 	svc_xprt_put(xprt);
1157 }
1158 
1159 /*
1160  * Save the request off for later processing. The request buffer looks
1161  * like this:
1162  *
1163  * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1164  *
1165  * This code can only handle requests that consist of an xprt-header
1166  * and rpc-header.
1167  */
1168 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1169 {
1170 	struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1171 	struct svc_deferred_req *dr;
1172 
1173 	if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1174 		return NULL; /* if more than a page, give up FIXME */
1175 	if (rqstp->rq_deferred) {
1176 		dr = rqstp->rq_deferred;
1177 		rqstp->rq_deferred = NULL;
1178 	} else {
1179 		size_t skip;
1180 		size_t size;
1181 		/* FIXME maybe discard if size too large */
1182 		size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1183 		dr = kmalloc(size, GFP_KERNEL);
1184 		if (dr == NULL)
1185 			return NULL;
1186 
1187 		dr->handle.owner = rqstp->rq_server;
1188 		dr->prot = rqstp->rq_prot;
1189 		memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1190 		dr->addrlen = rqstp->rq_addrlen;
1191 		dr->daddr = rqstp->rq_daddr;
1192 		dr->argslen = rqstp->rq_arg.len >> 2;
1193 		dr->xprt_hlen = rqstp->rq_xprt_hlen;
1194 
1195 		/* back up head to the start of the buffer and copy */
1196 		skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1197 		memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1198 		       dr->argslen << 2);
1199 	}
1200 	trace_svc_defer(rqstp);
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 noinline int svc_deferred_recv(struct svc_rqst *rqstp)
1213 {
1214 	struct svc_deferred_req *dr = rqstp->rq_deferred;
1215 
1216 	trace_svc_defer_recv(dr);
1217 
1218 	/* setup iov_base past transport header */
1219 	rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1220 	/* The iov_len does not include the transport header bytes */
1221 	rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1222 	rqstp->rq_arg.page_len = 0;
1223 	/* The rq_arg.len includes the transport header bytes */
1224 	rqstp->rq_arg.len     = dr->argslen<<2;
1225 	rqstp->rq_prot        = dr->prot;
1226 	memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1227 	rqstp->rq_addrlen     = dr->addrlen;
1228 	/* Save off transport header len in case we get deferred again */
1229 	rqstp->rq_xprt_hlen   = dr->xprt_hlen;
1230 	rqstp->rq_daddr       = dr->daddr;
1231 	rqstp->rq_respages    = rqstp->rq_pages;
1232 	return (dr->argslen<<2) - dr->xprt_hlen;
1233 }
1234 
1235 
1236 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1237 {
1238 	struct svc_deferred_req *dr = NULL;
1239 
1240 	if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1241 		return NULL;
1242 	spin_lock(&xprt->xpt_lock);
1243 	if (!list_empty(&xprt->xpt_deferred)) {
1244 		dr = list_entry(xprt->xpt_deferred.next,
1245 				struct svc_deferred_req,
1246 				handle.recent);
1247 		list_del_init(&dr->handle.recent);
1248 	} else
1249 		clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1250 	spin_unlock(&xprt->xpt_lock);
1251 	return dr;
1252 }
1253 
1254 /**
1255  * svc_find_xprt - find an RPC transport instance
1256  * @serv: pointer to svc_serv to search
1257  * @xcl_name: C string containing transport's class name
1258  * @net: owner net pointer
1259  * @af: Address family of transport's local address
1260  * @port: transport's IP port number
1261  *
1262  * Return the transport instance pointer for the endpoint accepting
1263  * connections/peer traffic from the specified transport class,
1264  * address family and port.
1265  *
1266  * Specifying 0 for the address family or port is effectively a
1267  * wild-card, and will result in matching the first transport in the
1268  * service's list that has a matching class name.
1269  */
1270 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1271 			       struct net *net, const sa_family_t af,
1272 			       const unsigned short port)
1273 {
1274 	struct svc_xprt *xprt;
1275 	struct svc_xprt *found = NULL;
1276 
1277 	/* Sanity check the args */
1278 	if (serv == NULL || xcl_name == NULL)
1279 		return found;
1280 
1281 	spin_lock_bh(&serv->sv_lock);
1282 	list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1283 		if (xprt->xpt_net != net)
1284 			continue;
1285 		if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1286 			continue;
1287 		if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1288 			continue;
1289 		if (port != 0 && port != svc_xprt_local_port(xprt))
1290 			continue;
1291 		found = xprt;
1292 		svc_xprt_get(xprt);
1293 		break;
1294 	}
1295 	spin_unlock_bh(&serv->sv_lock);
1296 	return found;
1297 }
1298 EXPORT_SYMBOL_GPL(svc_find_xprt);
1299 
1300 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1301 			     char *pos, int remaining)
1302 {
1303 	int len;
1304 
1305 	len = snprintf(pos, remaining, "%s %u\n",
1306 			xprt->xpt_class->xcl_name,
1307 			svc_xprt_local_port(xprt));
1308 	if (len >= remaining)
1309 		return -ENAMETOOLONG;
1310 	return len;
1311 }
1312 
1313 /**
1314  * svc_xprt_names - format a buffer with a list of transport names
1315  * @serv: pointer to an RPC service
1316  * @buf: pointer to a buffer to be filled in
1317  * @buflen: length of buffer to be filled in
1318  *
1319  * Fills in @buf with a string containing a list of transport names,
1320  * each name terminated with '\n'.
1321  *
1322  * Returns positive length of the filled-in string on success; otherwise
1323  * a negative errno value is returned if an error occurs.
1324  */
1325 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1326 {
1327 	struct svc_xprt *xprt;
1328 	int len, totlen;
1329 	char *pos;
1330 
1331 	/* Sanity check args */
1332 	if (!serv)
1333 		return 0;
1334 
1335 	spin_lock_bh(&serv->sv_lock);
1336 
1337 	pos = buf;
1338 	totlen = 0;
1339 	list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1340 		len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1341 		if (len < 0) {
1342 			*buf = '\0';
1343 			totlen = len;
1344 		}
1345 		if (len <= 0)
1346 			break;
1347 
1348 		pos += len;
1349 		totlen += len;
1350 	}
1351 
1352 	spin_unlock_bh(&serv->sv_lock);
1353 	return totlen;
1354 }
1355 EXPORT_SYMBOL_GPL(svc_xprt_names);
1356 
1357 
1358 /*----------------------------------------------------------------------------*/
1359 
1360 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1361 {
1362 	unsigned int pidx = (unsigned int)*pos;
1363 	struct svc_serv *serv = m->private;
1364 
1365 	dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1366 
1367 	if (!pidx)
1368 		return SEQ_START_TOKEN;
1369 	return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1370 }
1371 
1372 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1373 {
1374 	struct svc_pool *pool = p;
1375 	struct svc_serv *serv = m->private;
1376 
1377 	dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1378 
1379 	if (p == SEQ_START_TOKEN) {
1380 		pool = &serv->sv_pools[0];
1381 	} else {
1382 		unsigned int pidx = (pool - &serv->sv_pools[0]);
1383 		if (pidx < serv->sv_nrpools-1)
1384 			pool = &serv->sv_pools[pidx+1];
1385 		else
1386 			pool = NULL;
1387 	}
1388 	++*pos;
1389 	return pool;
1390 }
1391 
1392 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1393 {
1394 }
1395 
1396 static int svc_pool_stats_show(struct seq_file *m, void *p)
1397 {
1398 	struct svc_pool *pool = p;
1399 
1400 	if (p == SEQ_START_TOKEN) {
1401 		seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1402 		return 0;
1403 	}
1404 
1405 	seq_printf(m, "%u %lu %lu %lu %lu\n",
1406 		pool->sp_id,
1407 		(unsigned long)atomic_long_read(&pool->sp_stats.packets),
1408 		pool->sp_stats.sockets_queued,
1409 		(unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1410 		(unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1411 
1412 	return 0;
1413 }
1414 
1415 static const struct seq_operations svc_pool_stats_seq_ops = {
1416 	.start	= svc_pool_stats_start,
1417 	.next	= svc_pool_stats_next,
1418 	.stop	= svc_pool_stats_stop,
1419 	.show	= svc_pool_stats_show,
1420 };
1421 
1422 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1423 {
1424 	int err;
1425 
1426 	err = seq_open(file, &svc_pool_stats_seq_ops);
1427 	if (!err)
1428 		((struct seq_file *) file->private_data)->private = serv;
1429 	return err;
1430 }
1431 EXPORT_SYMBOL(svc_pool_stats_open);
1432 
1433 /*----------------------------------------------------------------------------*/
1434