xref: /linux/net/sunrpc/svcsock.c (revision c537b994505099b7197e7d3125b942ecbcc51eb6)
1 /*
2  * linux/net/sunrpc/svcsock.c
3  *
4  * These are the RPC server socket internals.
5  *
6  * The server scheduling algorithm does not always distribute the load
7  * evenly when servicing a single client. May need to modify the
8  * svc_sock_enqueue procedure...
9  *
10  * TCP support is largely untested and may be a little slow. The problem
11  * is that we currently do two separate recvfrom's, one for the 4-byte
12  * record length, and the second for the actual record. This could possibly
13  * be improved by always reading a minimum size of around 100 bytes and
14  * tucking any superfluous bytes away in a temporary store. Still, that
15  * leaves write requests out in the rain. An alternative may be to peek at
16  * the first skb in the queue, and if it matches the next TCP sequence
17  * number, to extract the record marker. Yuck.
18  *
19  * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
20  */
21 
22 #include <linux/sched.h>
23 #include <linux/errno.h>
24 #include <linux/fcntl.h>
25 #include <linux/net.h>
26 #include <linux/in.h>
27 #include <linux/inet.h>
28 #include <linux/udp.h>
29 #include <linux/tcp.h>
30 #include <linux/unistd.h>
31 #include <linux/slab.h>
32 #include <linux/netdevice.h>
33 #include <linux/skbuff.h>
34 #include <linux/file.h>
35 #include <linux/freezer.h>
36 #include <net/sock.h>
37 #include <net/checksum.h>
38 #include <net/ip.h>
39 #include <net/ipv6.h>
40 #include <net/tcp_states.h>
41 #include <asm/uaccess.h>
42 #include <asm/ioctls.h>
43 
44 #include <linux/sunrpc/types.h>
45 #include <linux/sunrpc/clnt.h>
46 #include <linux/sunrpc/xdr.h>
47 #include <linux/sunrpc/svcsock.h>
48 #include <linux/sunrpc/stats.h>
49 
50 /* SMP locking strategy:
51  *
52  *	svc_pool->sp_lock protects most of the fields of that pool.
53  * 	svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
54  *	when both need to be taken (rare), svc_serv->sv_lock is first.
55  *	BKL protects svc_serv->sv_nrthread.
56  *	svc_sock->sk_defer_lock protects the svc_sock->sk_deferred list
57  *	svc_sock->sk_flags.SK_BUSY prevents a svc_sock being enqueued multiply.
58  *
59  *	Some flags can be set to certain values at any time
60  *	providing that certain rules are followed:
61  *
62  *	SK_CONN, SK_DATA, can be set or cleared at any time.
63  *		after a set, svc_sock_enqueue must be called.
64  *		after a clear, the socket must be read/accepted
65  *		 if this succeeds, it must be set again.
66  *	SK_CLOSE can set at any time. It is never cleared.
67  *      sk_inuse contains a bias of '1' until SK_DEAD is set.
68  *             so when sk_inuse hits zero, we know the socket is dead
69  *             and no-one is using it.
70  *      SK_DEAD can only be set while SK_BUSY is held which ensures
71  *             no other thread will be using the socket or will try to
72  *	       set SK_DEAD.
73  *
74  */
75 
76 #define RPCDBG_FACILITY	RPCDBG_SVCSOCK
77 
78 
79 static struct svc_sock *svc_setup_socket(struct svc_serv *, struct socket *,
80 					 int *errp, int flags);
81 static void		svc_delete_socket(struct svc_sock *svsk);
82 static void		svc_udp_data_ready(struct sock *, int);
83 static int		svc_udp_recvfrom(struct svc_rqst *);
84 static int		svc_udp_sendto(struct svc_rqst *);
85 static void		svc_close_socket(struct svc_sock *svsk);
86 
87 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk);
88 static int svc_deferred_recv(struct svc_rqst *rqstp);
89 static struct cache_deferred_req *svc_defer(struct cache_req *req);
90 
91 /* apparently the "standard" is that clients close
92  * idle connections after 5 minutes, servers after
93  * 6 minutes
94  *   http://www.connectathon.org/talks96/nfstcp.pdf
95  */
96 static int svc_conn_age_period = 6*60;
97 
98 #ifdef CONFIG_DEBUG_LOCK_ALLOC
99 static struct lock_class_key svc_key[2];
100 static struct lock_class_key svc_slock_key[2];
101 
102 static inline void svc_reclassify_socket(struct socket *sock)
103 {
104 	struct sock *sk = sock->sk;
105 	BUG_ON(sk->sk_lock.owner != NULL);
106 	switch (sk->sk_family) {
107 	case AF_INET:
108 		sock_lock_init_class_and_name(sk, "slock-AF_INET-NFSD",
109 		    &svc_slock_key[0], "sk_lock-AF_INET-NFSD", &svc_key[0]);
110 		break;
111 
112 	case AF_INET6:
113 		sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFSD",
114 		    &svc_slock_key[1], "sk_lock-AF_INET6-NFSD", &svc_key[1]);
115 		break;
116 
117 	default:
118 		BUG();
119 	}
120 }
121 #else
122 static inline void svc_reclassify_socket(struct socket *sock)
123 {
124 }
125 #endif
126 
127 static char *__svc_print_addr(struct sockaddr *addr, char *buf, size_t len)
128 {
129 	switch (addr->sa_family) {
130 	case AF_INET:
131 		snprintf(buf, len, "%u.%u.%u.%u, port=%u",
132 			NIPQUAD(((struct sockaddr_in *) addr)->sin_addr),
133 			htons(((struct sockaddr_in *) addr)->sin_port));
134 		break;
135 
136 	case AF_INET6:
137 		snprintf(buf, len, "%x:%x:%x:%x:%x:%x:%x:%x, port=%u",
138 			NIP6(((struct sockaddr_in6 *) addr)->sin6_addr),
139 			htons(((struct sockaddr_in6 *) addr)->sin6_port));
140 		break;
141 
142 	default:
143 		snprintf(buf, len, "unknown address type: %d", addr->sa_family);
144 		break;
145 	}
146 	return buf;
147 }
148 
149 /**
150  * svc_print_addr - Format rq_addr field for printing
151  * @rqstp: svc_rqst struct containing address to print
152  * @buf: target buffer for formatted address
153  * @len: length of target buffer
154  *
155  */
156 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
157 {
158 	return __svc_print_addr(svc_addr(rqstp), buf, len);
159 }
160 EXPORT_SYMBOL_GPL(svc_print_addr);
161 
162 /*
163  * Queue up an idle server thread.  Must have pool->sp_lock held.
164  * Note: this is really a stack rather than a queue, so that we only
165  * use as many different threads as we need, and the rest don't pollute
166  * the cache.
167  */
168 static inline void
169 svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
170 {
171 	list_add(&rqstp->rq_list, &pool->sp_threads);
172 }
173 
174 /*
175  * Dequeue an nfsd thread.  Must have pool->sp_lock held.
176  */
177 static inline void
178 svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
179 {
180 	list_del(&rqstp->rq_list);
181 }
182 
183 /*
184  * Release an skbuff after use
185  */
186 static inline void
187 svc_release_skb(struct svc_rqst *rqstp)
188 {
189 	struct sk_buff *skb = rqstp->rq_skbuff;
190 	struct svc_deferred_req *dr = rqstp->rq_deferred;
191 
192 	if (skb) {
193 		rqstp->rq_skbuff = NULL;
194 
195 		dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
196 		skb_free_datagram(rqstp->rq_sock->sk_sk, skb);
197 	}
198 	if (dr) {
199 		rqstp->rq_deferred = NULL;
200 		kfree(dr);
201 	}
202 }
203 
204 /*
205  * Any space to write?
206  */
207 static inline unsigned long
208 svc_sock_wspace(struct svc_sock *svsk)
209 {
210 	int wspace;
211 
212 	if (svsk->sk_sock->type == SOCK_STREAM)
213 		wspace = sk_stream_wspace(svsk->sk_sk);
214 	else
215 		wspace = sock_wspace(svsk->sk_sk);
216 
217 	return wspace;
218 }
219 
220 /*
221  * Queue up a socket with data pending. If there are idle nfsd
222  * processes, wake 'em up.
223  *
224  */
225 static void
226 svc_sock_enqueue(struct svc_sock *svsk)
227 {
228 	struct svc_serv	*serv = svsk->sk_server;
229 	struct svc_pool *pool;
230 	struct svc_rqst	*rqstp;
231 	int cpu;
232 
233 	if (!(svsk->sk_flags &
234 	      ( (1<<SK_CONN)|(1<<SK_DATA)|(1<<SK_CLOSE)|(1<<SK_DEFERRED)) ))
235 		return;
236 	if (test_bit(SK_DEAD, &svsk->sk_flags))
237 		return;
238 
239 	cpu = get_cpu();
240 	pool = svc_pool_for_cpu(svsk->sk_server, cpu);
241 	put_cpu();
242 
243 	spin_lock_bh(&pool->sp_lock);
244 
245 	if (!list_empty(&pool->sp_threads) &&
246 	    !list_empty(&pool->sp_sockets))
247 		printk(KERN_ERR
248 			"svc_sock_enqueue: threads and sockets both waiting??\n");
249 
250 	if (test_bit(SK_DEAD, &svsk->sk_flags)) {
251 		/* Don't enqueue dead sockets */
252 		dprintk("svc: socket %p is dead, not enqueued\n", svsk->sk_sk);
253 		goto out_unlock;
254 	}
255 
256 	/* Mark socket as busy. It will remain in this state until the
257 	 * server has processed all pending data and put the socket back
258 	 * on the idle list.  We update SK_BUSY atomically because
259 	 * it also guards against trying to enqueue the svc_sock twice.
260 	 */
261 	if (test_and_set_bit(SK_BUSY, &svsk->sk_flags)) {
262 		/* Don't enqueue socket while already enqueued */
263 		dprintk("svc: socket %p busy, not enqueued\n", svsk->sk_sk);
264 		goto out_unlock;
265 	}
266 	BUG_ON(svsk->sk_pool != NULL);
267 	svsk->sk_pool = pool;
268 
269 	set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
270 	if (((atomic_read(&svsk->sk_reserved) + serv->sv_max_mesg)*2
271 	     > svc_sock_wspace(svsk))
272 	    && !test_bit(SK_CLOSE, &svsk->sk_flags)
273 	    && !test_bit(SK_CONN, &svsk->sk_flags)) {
274 		/* Don't enqueue while not enough space for reply */
275 		dprintk("svc: socket %p  no space, %d*2 > %ld, not enqueued\n",
276 			svsk->sk_sk, atomic_read(&svsk->sk_reserved)+serv->sv_max_mesg,
277 			svc_sock_wspace(svsk));
278 		svsk->sk_pool = NULL;
279 		clear_bit(SK_BUSY, &svsk->sk_flags);
280 		goto out_unlock;
281 	}
282 	clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
283 
284 
285 	if (!list_empty(&pool->sp_threads)) {
286 		rqstp = list_entry(pool->sp_threads.next,
287 				   struct svc_rqst,
288 				   rq_list);
289 		dprintk("svc: socket %p served by daemon %p\n",
290 			svsk->sk_sk, rqstp);
291 		svc_thread_dequeue(pool, rqstp);
292 		if (rqstp->rq_sock)
293 			printk(KERN_ERR
294 				"svc_sock_enqueue: server %p, rq_sock=%p!\n",
295 				rqstp, rqstp->rq_sock);
296 		rqstp->rq_sock = svsk;
297 		atomic_inc(&svsk->sk_inuse);
298 		rqstp->rq_reserved = serv->sv_max_mesg;
299 		atomic_add(rqstp->rq_reserved, &svsk->sk_reserved);
300 		BUG_ON(svsk->sk_pool != pool);
301 		wake_up(&rqstp->rq_wait);
302 	} else {
303 		dprintk("svc: socket %p put into queue\n", svsk->sk_sk);
304 		list_add_tail(&svsk->sk_ready, &pool->sp_sockets);
305 		BUG_ON(svsk->sk_pool != pool);
306 	}
307 
308 out_unlock:
309 	spin_unlock_bh(&pool->sp_lock);
310 }
311 
312 /*
313  * Dequeue the first socket.  Must be called with the pool->sp_lock held.
314  */
315 static inline struct svc_sock *
316 svc_sock_dequeue(struct svc_pool *pool)
317 {
318 	struct svc_sock	*svsk;
319 
320 	if (list_empty(&pool->sp_sockets))
321 		return NULL;
322 
323 	svsk = list_entry(pool->sp_sockets.next,
324 			  struct svc_sock, sk_ready);
325 	list_del_init(&svsk->sk_ready);
326 
327 	dprintk("svc: socket %p dequeued, inuse=%d\n",
328 		svsk->sk_sk, atomic_read(&svsk->sk_inuse));
329 
330 	return svsk;
331 }
332 
333 /*
334  * Having read something from a socket, check whether it
335  * needs to be re-enqueued.
336  * Note: SK_DATA only gets cleared when a read-attempt finds
337  * no (or insufficient) data.
338  */
339 static inline void
340 svc_sock_received(struct svc_sock *svsk)
341 {
342 	svsk->sk_pool = NULL;
343 	clear_bit(SK_BUSY, &svsk->sk_flags);
344 	svc_sock_enqueue(svsk);
345 }
346 
347 
348 /**
349  * svc_reserve - change the space reserved for the reply to a request.
350  * @rqstp:  The request in question
351  * @space: new max space to reserve
352  *
353  * Each request reserves some space on the output queue of the socket
354  * to make sure the reply fits.  This function reduces that reserved
355  * space to be the amount of space used already, plus @space.
356  *
357  */
358 void svc_reserve(struct svc_rqst *rqstp, int space)
359 {
360 	space += rqstp->rq_res.head[0].iov_len;
361 
362 	if (space < rqstp->rq_reserved) {
363 		struct svc_sock *svsk = rqstp->rq_sock;
364 		atomic_sub((rqstp->rq_reserved - space), &svsk->sk_reserved);
365 		rqstp->rq_reserved = space;
366 
367 		svc_sock_enqueue(svsk);
368 	}
369 }
370 
371 /*
372  * Release a socket after use.
373  */
374 static inline void
375 svc_sock_put(struct svc_sock *svsk)
376 {
377 	if (atomic_dec_and_test(&svsk->sk_inuse)) {
378 		BUG_ON(! test_bit(SK_DEAD, &svsk->sk_flags));
379 
380 		dprintk("svc: releasing dead socket\n");
381 		if (svsk->sk_sock->file)
382 			sockfd_put(svsk->sk_sock);
383 		else
384 			sock_release(svsk->sk_sock);
385 		if (svsk->sk_info_authunix != NULL)
386 			svcauth_unix_info_release(svsk->sk_info_authunix);
387 		kfree(svsk);
388 	}
389 }
390 
391 static void
392 svc_sock_release(struct svc_rqst *rqstp)
393 {
394 	struct svc_sock	*svsk = rqstp->rq_sock;
395 
396 	svc_release_skb(rqstp);
397 
398 	svc_free_res_pages(rqstp);
399 	rqstp->rq_res.page_len = 0;
400 	rqstp->rq_res.page_base = 0;
401 
402 
403 	/* Reset response buffer and release
404 	 * the reservation.
405 	 * But first, check that enough space was reserved
406 	 * for the reply, otherwise we have a bug!
407 	 */
408 	if ((rqstp->rq_res.len) >  rqstp->rq_reserved)
409 		printk(KERN_ERR "RPC request reserved %d but used %d\n",
410 		       rqstp->rq_reserved,
411 		       rqstp->rq_res.len);
412 
413 	rqstp->rq_res.head[0].iov_len = 0;
414 	svc_reserve(rqstp, 0);
415 	rqstp->rq_sock = NULL;
416 
417 	svc_sock_put(svsk);
418 }
419 
420 /*
421  * External function to wake up a server waiting for data
422  * This really only makes sense for services like lockd
423  * which have exactly one thread anyway.
424  */
425 void
426 svc_wake_up(struct svc_serv *serv)
427 {
428 	struct svc_rqst	*rqstp;
429 	unsigned int i;
430 	struct svc_pool *pool;
431 
432 	for (i = 0; i < serv->sv_nrpools; i++) {
433 		pool = &serv->sv_pools[i];
434 
435 		spin_lock_bh(&pool->sp_lock);
436 		if (!list_empty(&pool->sp_threads)) {
437 			rqstp = list_entry(pool->sp_threads.next,
438 					   struct svc_rqst,
439 					   rq_list);
440 			dprintk("svc: daemon %p woken up.\n", rqstp);
441 			/*
442 			svc_thread_dequeue(pool, rqstp);
443 			rqstp->rq_sock = NULL;
444 			 */
445 			wake_up(&rqstp->rq_wait);
446 		}
447 		spin_unlock_bh(&pool->sp_lock);
448 	}
449 }
450 
451 union svc_pktinfo_u {
452 	struct in_pktinfo pkti;
453 	struct in6_pktinfo pkti6;
454 };
455 
456 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
457 {
458 	switch (rqstp->rq_sock->sk_sk->sk_family) {
459 	case AF_INET: {
460 			struct in_pktinfo *pki = CMSG_DATA(cmh);
461 
462 			cmh->cmsg_level = SOL_IP;
463 			cmh->cmsg_type = IP_PKTINFO;
464 			pki->ipi_ifindex = 0;
465 			pki->ipi_spec_dst.s_addr = rqstp->rq_daddr.addr.s_addr;
466 			cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
467 		}
468 		break;
469 
470 	case AF_INET6: {
471 			struct in6_pktinfo *pki = CMSG_DATA(cmh);
472 
473 			cmh->cmsg_level = SOL_IPV6;
474 			cmh->cmsg_type = IPV6_PKTINFO;
475 			pki->ipi6_ifindex = 0;
476 			ipv6_addr_copy(&pki->ipi6_addr,
477 					&rqstp->rq_daddr.addr6);
478 			cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
479 		}
480 		break;
481 	}
482 	return;
483 }
484 
485 /*
486  * Generic sendto routine
487  */
488 static int
489 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
490 {
491 	struct svc_sock	*svsk = rqstp->rq_sock;
492 	struct socket	*sock = svsk->sk_sock;
493 	int		slen;
494 	char 		buffer[CMSG_SPACE(sizeof(union svc_pktinfo_u))];
495 	struct cmsghdr *cmh = (struct cmsghdr *)buffer;
496 	int		len = 0;
497 	int		result;
498 	int		size;
499 	struct page	**ppage = xdr->pages;
500 	size_t		base = xdr->page_base;
501 	unsigned int	pglen = xdr->page_len;
502 	unsigned int	flags = MSG_MORE;
503 	char		buf[RPC_MAX_ADDRBUFLEN];
504 
505 	slen = xdr->len;
506 
507 	if (rqstp->rq_prot == IPPROTO_UDP) {
508 		struct msghdr msg = {
509 			.msg_name	= &rqstp->rq_addr,
510 			.msg_namelen	= rqstp->rq_addrlen,
511 			.msg_control	= cmh,
512 			.msg_controllen	= sizeof(buffer),
513 			.msg_flags	= MSG_MORE,
514 		};
515 
516 		svc_set_cmsg_data(rqstp, cmh);
517 
518 		if (sock_sendmsg(sock, &msg, 0) < 0)
519 			goto out;
520 	}
521 
522 	/* send head */
523 	if (slen == xdr->head[0].iov_len)
524 		flags = 0;
525 	len = kernel_sendpage(sock, rqstp->rq_respages[0], 0,
526 				  xdr->head[0].iov_len, flags);
527 	if (len != xdr->head[0].iov_len)
528 		goto out;
529 	slen -= xdr->head[0].iov_len;
530 	if (slen == 0)
531 		goto out;
532 
533 	/* send page data */
534 	size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
535 	while (pglen > 0) {
536 		if (slen == size)
537 			flags = 0;
538 		result = kernel_sendpage(sock, *ppage, base, size, flags);
539 		if (result > 0)
540 			len += result;
541 		if (result != size)
542 			goto out;
543 		slen -= size;
544 		pglen -= size;
545 		size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
546 		base = 0;
547 		ppage++;
548 	}
549 	/* send tail */
550 	if (xdr->tail[0].iov_len) {
551 		result = kernel_sendpage(sock, rqstp->rq_respages[0],
552 					     ((unsigned long)xdr->tail[0].iov_base)
553 						& (PAGE_SIZE-1),
554 					     xdr->tail[0].iov_len, 0);
555 
556 		if (result > 0)
557 			len += result;
558 	}
559 out:
560 	dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
561 		rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len,
562 		xdr->len, len, svc_print_addr(rqstp, buf, sizeof(buf)));
563 
564 	return len;
565 }
566 
567 /*
568  * Report socket names for nfsdfs
569  */
570 static int one_sock_name(char *buf, struct svc_sock *svsk)
571 {
572 	int len;
573 
574 	switch(svsk->sk_sk->sk_family) {
575 	case AF_INET:
576 		len = sprintf(buf, "ipv4 %s %u.%u.%u.%u %d\n",
577 			      svsk->sk_sk->sk_protocol==IPPROTO_UDP?
578 			      "udp" : "tcp",
579 			      NIPQUAD(inet_sk(svsk->sk_sk)->rcv_saddr),
580 			      inet_sk(svsk->sk_sk)->num);
581 		break;
582 	default:
583 		len = sprintf(buf, "*unknown-%d*\n",
584 			       svsk->sk_sk->sk_family);
585 	}
586 	return len;
587 }
588 
589 int
590 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
591 {
592 	struct svc_sock *svsk, *closesk = NULL;
593 	int len = 0;
594 
595 	if (!serv)
596 		return 0;
597 	spin_lock_bh(&serv->sv_lock);
598 	list_for_each_entry(svsk, &serv->sv_permsocks, sk_list) {
599 		int onelen = one_sock_name(buf+len, svsk);
600 		if (toclose && strcmp(toclose, buf+len) == 0)
601 			closesk = svsk;
602 		else
603 			len += onelen;
604 	}
605 	spin_unlock_bh(&serv->sv_lock);
606 	if (closesk)
607 		/* Should unregister with portmap, but you cannot
608 		 * unregister just one protocol...
609 		 */
610 		svc_close_socket(closesk);
611 	else if (toclose)
612 		return -ENOENT;
613 	return len;
614 }
615 EXPORT_SYMBOL(svc_sock_names);
616 
617 /*
618  * Check input queue length
619  */
620 static int
621 svc_recv_available(struct svc_sock *svsk)
622 {
623 	struct socket	*sock = svsk->sk_sock;
624 	int		avail, err;
625 
626 	err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
627 
628 	return (err >= 0)? avail : err;
629 }
630 
631 /*
632  * Generic recvfrom routine.
633  */
634 static int
635 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
636 {
637 	struct svc_sock *svsk = rqstp->rq_sock;
638 	struct msghdr msg = {
639 		.msg_flags	= MSG_DONTWAIT,
640 	};
641 	int len;
642 
643 	len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
644 				msg.msg_flags);
645 
646 	/* sock_recvmsg doesn't fill in the name/namelen, so we must..
647 	 */
648 	memcpy(&rqstp->rq_addr, &svsk->sk_remote, svsk->sk_remotelen);
649 	rqstp->rq_addrlen = svsk->sk_remotelen;
650 
651 	dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
652 		svsk, iov[0].iov_base, iov[0].iov_len, len);
653 
654 	return len;
655 }
656 
657 /*
658  * Set socket snd and rcv buffer lengths
659  */
660 static inline void
661 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
662 {
663 #if 0
664 	mm_segment_t	oldfs;
665 	oldfs = get_fs(); set_fs(KERNEL_DS);
666 	sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
667 			(char*)&snd, sizeof(snd));
668 	sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
669 			(char*)&rcv, sizeof(rcv));
670 #else
671 	/* sock_setsockopt limits use to sysctl_?mem_max,
672 	 * which isn't acceptable.  Until that is made conditional
673 	 * on not having CAP_SYS_RESOURCE or similar, we go direct...
674 	 * DaveM said I could!
675 	 */
676 	lock_sock(sock->sk);
677 	sock->sk->sk_sndbuf = snd * 2;
678 	sock->sk->sk_rcvbuf = rcv * 2;
679 	sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
680 	release_sock(sock->sk);
681 #endif
682 }
683 /*
684  * INET callback when data has been received on the socket.
685  */
686 static void
687 svc_udp_data_ready(struct sock *sk, int count)
688 {
689 	struct svc_sock	*svsk = (struct svc_sock *)sk->sk_user_data;
690 
691 	if (svsk) {
692 		dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
693 			svsk, sk, count, test_bit(SK_BUSY, &svsk->sk_flags));
694 		set_bit(SK_DATA, &svsk->sk_flags);
695 		svc_sock_enqueue(svsk);
696 	}
697 	if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
698 		wake_up_interruptible(sk->sk_sleep);
699 }
700 
701 /*
702  * INET callback when space is newly available on the socket.
703  */
704 static void
705 svc_write_space(struct sock *sk)
706 {
707 	struct svc_sock	*svsk = (struct svc_sock *)(sk->sk_user_data);
708 
709 	if (svsk) {
710 		dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
711 			svsk, sk, test_bit(SK_BUSY, &svsk->sk_flags));
712 		svc_sock_enqueue(svsk);
713 	}
714 
715 	if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
716 		dprintk("RPC svc_write_space: someone sleeping on %p\n",
717 		       svsk);
718 		wake_up_interruptible(sk->sk_sleep);
719 	}
720 }
721 
722 static inline void svc_udp_get_dest_address(struct svc_rqst *rqstp,
723 					    struct cmsghdr *cmh)
724 {
725 	switch (rqstp->rq_sock->sk_sk->sk_family) {
726 	case AF_INET: {
727 		struct in_pktinfo *pki = CMSG_DATA(cmh);
728 		rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
729 		break;
730 		}
731 	case AF_INET6: {
732 		struct in6_pktinfo *pki = CMSG_DATA(cmh);
733 		ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
734 		break;
735 		}
736 	}
737 }
738 
739 /*
740  * Receive a datagram from a UDP socket.
741  */
742 static int
743 svc_udp_recvfrom(struct svc_rqst *rqstp)
744 {
745 	struct svc_sock	*svsk = rqstp->rq_sock;
746 	struct svc_serv	*serv = svsk->sk_server;
747 	struct sk_buff	*skb;
748 	char		buffer[CMSG_SPACE(sizeof(union svc_pktinfo_u))];
749 	struct cmsghdr *cmh = (struct cmsghdr *)buffer;
750 	int		err, len;
751 	struct msghdr msg = {
752 		.msg_name = svc_addr(rqstp),
753 		.msg_control = cmh,
754 		.msg_controllen = sizeof(buffer),
755 		.msg_flags = MSG_DONTWAIT,
756 	};
757 
758 	if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
759 	    /* udp sockets need large rcvbuf as all pending
760 	     * requests are still in that buffer.  sndbuf must
761 	     * also be large enough that there is enough space
762 	     * for one reply per thread.  We count all threads
763 	     * rather than threads in a particular pool, which
764 	     * provides an upper bound on the number of threads
765 	     * which will access the socket.
766 	     */
767 	    svc_sock_setbufsize(svsk->sk_sock,
768 				(serv->sv_nrthreads+3) * serv->sv_max_mesg,
769 				(serv->sv_nrthreads+3) * serv->sv_max_mesg);
770 
771 	if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
772 		svc_sock_received(svsk);
773 		return svc_deferred_recv(rqstp);
774 	}
775 
776 	if (test_bit(SK_CLOSE, &svsk->sk_flags)) {
777 		svc_delete_socket(svsk);
778 		return 0;
779 	}
780 
781 	clear_bit(SK_DATA, &svsk->sk_flags);
782 	while ((err == kernel_recvmsg(svsk->sk_sock, &msg, NULL,
783 				      0, 0, MSG_PEEK | MSG_DONTWAIT)) < 0 ||
784 	       (skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err)) == NULL) {
785 		if (err == -EAGAIN) {
786 			svc_sock_received(svsk);
787 			return err;
788 		}
789 		/* possibly an icmp error */
790 		dprintk("svc: recvfrom returned error %d\n", -err);
791 	}
792 	rqstp->rq_addrlen = sizeof(rqstp->rq_addr);
793 	if (skb->tstamp.off_sec == 0) {
794 		struct timeval tv;
795 
796 		tv.tv_sec = xtime.tv_sec;
797 		tv.tv_usec = xtime.tv_nsec / NSEC_PER_USEC;
798 		skb_set_timestamp(skb, &tv);
799 		/* Don't enable netstamp, sunrpc doesn't
800 		   need that much accuracy */
801 	}
802 	skb_get_timestamp(skb, &svsk->sk_sk->sk_stamp);
803 	set_bit(SK_DATA, &svsk->sk_flags); /* there may be more data... */
804 
805 	/*
806 	 * Maybe more packets - kick another thread ASAP.
807 	 */
808 	svc_sock_received(svsk);
809 
810 	len  = skb->len - sizeof(struct udphdr);
811 	rqstp->rq_arg.len = len;
812 
813 	rqstp->rq_prot = IPPROTO_UDP;
814 
815 	if (cmh->cmsg_level != IPPROTO_IP ||
816 	    cmh->cmsg_type != IP_PKTINFO) {
817 		if (net_ratelimit())
818 			printk("rpcsvc: received unknown control message:"
819 			       "%d/%d\n",
820 			       cmh->cmsg_level, cmh->cmsg_type);
821 		skb_free_datagram(svsk->sk_sk, skb);
822 		return 0;
823 	}
824 	svc_udp_get_dest_address(rqstp, cmh);
825 
826 	if (skb_is_nonlinear(skb)) {
827 		/* we have to copy */
828 		local_bh_disable();
829 		if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
830 			local_bh_enable();
831 			/* checksum error */
832 			skb_free_datagram(svsk->sk_sk, skb);
833 			return 0;
834 		}
835 		local_bh_enable();
836 		skb_free_datagram(svsk->sk_sk, skb);
837 	} else {
838 		/* we can use it in-place */
839 		rqstp->rq_arg.head[0].iov_base = skb->data + sizeof(struct udphdr);
840 		rqstp->rq_arg.head[0].iov_len = len;
841 		if (skb_checksum_complete(skb)) {
842 			skb_free_datagram(svsk->sk_sk, skb);
843 			return 0;
844 		}
845 		rqstp->rq_skbuff = skb;
846 	}
847 
848 	rqstp->rq_arg.page_base = 0;
849 	if (len <= rqstp->rq_arg.head[0].iov_len) {
850 		rqstp->rq_arg.head[0].iov_len = len;
851 		rqstp->rq_arg.page_len = 0;
852 		rqstp->rq_respages = rqstp->rq_pages+1;
853 	} else {
854 		rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
855 		rqstp->rq_respages = rqstp->rq_pages + 1 +
856 			(rqstp->rq_arg.page_len + PAGE_SIZE - 1)/ PAGE_SIZE;
857 	}
858 
859 	if (serv->sv_stats)
860 		serv->sv_stats->netudpcnt++;
861 
862 	return len;
863 }
864 
865 static int
866 svc_udp_sendto(struct svc_rqst *rqstp)
867 {
868 	int		error;
869 
870 	error = svc_sendto(rqstp, &rqstp->rq_res);
871 	if (error == -ECONNREFUSED)
872 		/* ICMP error on earlier request. */
873 		error = svc_sendto(rqstp, &rqstp->rq_res);
874 
875 	return error;
876 }
877 
878 static void
879 svc_udp_init(struct svc_sock *svsk)
880 {
881 	int one = 1;
882 	mm_segment_t oldfs;
883 
884 	svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
885 	svsk->sk_sk->sk_write_space = svc_write_space;
886 	svsk->sk_recvfrom = svc_udp_recvfrom;
887 	svsk->sk_sendto = svc_udp_sendto;
888 
889 	/* initialise setting must have enough space to
890 	 * receive and respond to one request.
891 	 * svc_udp_recvfrom will re-adjust if necessary
892 	 */
893 	svc_sock_setbufsize(svsk->sk_sock,
894 			    3 * svsk->sk_server->sv_max_mesg,
895 			    3 * svsk->sk_server->sv_max_mesg);
896 
897 	set_bit(SK_DATA, &svsk->sk_flags); /* might have come in before data_ready set up */
898 	set_bit(SK_CHNGBUF, &svsk->sk_flags);
899 
900 	oldfs = get_fs();
901 	set_fs(KERNEL_DS);
902 	/* make sure we get destination address info */
903 	svsk->sk_sock->ops->setsockopt(svsk->sk_sock, IPPROTO_IP, IP_PKTINFO,
904 				       (char __user *)&one, sizeof(one));
905 	set_fs(oldfs);
906 }
907 
908 /*
909  * A data_ready event on a listening socket means there's a connection
910  * pending. Do not use state_change as a substitute for it.
911  */
912 static void
913 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
914 {
915 	struct svc_sock	*svsk = (struct svc_sock *)sk->sk_user_data;
916 
917 	dprintk("svc: socket %p TCP (listen) state change %d\n",
918 		sk, sk->sk_state);
919 
920 	/*
921 	 * This callback may called twice when a new connection
922 	 * is established as a child socket inherits everything
923 	 * from a parent LISTEN socket.
924 	 * 1) data_ready method of the parent socket will be called
925 	 *    when one of child sockets become ESTABLISHED.
926 	 * 2) data_ready method of the child socket may be called
927 	 *    when it receives data before the socket is accepted.
928 	 * In case of 2, we should ignore it silently.
929 	 */
930 	if (sk->sk_state == TCP_LISTEN) {
931 		if (svsk) {
932 			set_bit(SK_CONN, &svsk->sk_flags);
933 			svc_sock_enqueue(svsk);
934 		} else
935 			printk("svc: socket %p: no user data\n", sk);
936 	}
937 
938 	if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
939 		wake_up_interruptible_all(sk->sk_sleep);
940 }
941 
942 /*
943  * A state change on a connected socket means it's dying or dead.
944  */
945 static void
946 svc_tcp_state_change(struct sock *sk)
947 {
948 	struct svc_sock	*svsk = (struct svc_sock *)sk->sk_user_data;
949 
950 	dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
951 		sk, sk->sk_state, sk->sk_user_data);
952 
953 	if (!svsk)
954 		printk("svc: socket %p: no user data\n", sk);
955 	else {
956 		set_bit(SK_CLOSE, &svsk->sk_flags);
957 		svc_sock_enqueue(svsk);
958 	}
959 	if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
960 		wake_up_interruptible_all(sk->sk_sleep);
961 }
962 
963 static void
964 svc_tcp_data_ready(struct sock *sk, int count)
965 {
966 	struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
967 
968 	dprintk("svc: socket %p TCP data ready (svsk %p)\n",
969 		sk, sk->sk_user_data);
970 	if (svsk) {
971 		set_bit(SK_DATA, &svsk->sk_flags);
972 		svc_sock_enqueue(svsk);
973 	}
974 	if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
975 		wake_up_interruptible(sk->sk_sleep);
976 }
977 
978 static inline int svc_port_is_privileged(struct sockaddr *sin)
979 {
980 	switch (sin->sa_family) {
981 	case AF_INET:
982 		return ntohs(((struct sockaddr_in *)sin)->sin_port)
983 			< PROT_SOCK;
984 	case AF_INET6:
985 		return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
986 			< PROT_SOCK;
987 	default:
988 		return 0;
989 	}
990 }
991 
992 /*
993  * Accept a TCP connection
994  */
995 static void
996 svc_tcp_accept(struct svc_sock *svsk)
997 {
998 	struct sockaddr_storage addr;
999 	struct sockaddr	*sin = (struct sockaddr *) &addr;
1000 	struct svc_serv	*serv = svsk->sk_server;
1001 	struct socket	*sock = svsk->sk_sock;
1002 	struct socket	*newsock;
1003 	struct svc_sock	*newsvsk;
1004 	int		err, slen;
1005 	char		buf[RPC_MAX_ADDRBUFLEN];
1006 
1007 	dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
1008 	if (!sock)
1009 		return;
1010 
1011 	clear_bit(SK_CONN, &svsk->sk_flags);
1012 	err = kernel_accept(sock, &newsock, O_NONBLOCK);
1013 	if (err < 0) {
1014 		if (err == -ENOMEM)
1015 			printk(KERN_WARNING "%s: no more sockets!\n",
1016 			       serv->sv_name);
1017 		else if (err != -EAGAIN && net_ratelimit())
1018 			printk(KERN_WARNING "%s: accept failed (err %d)!\n",
1019 				   serv->sv_name, -err);
1020 		return;
1021 	}
1022 
1023 	set_bit(SK_CONN, &svsk->sk_flags);
1024 	svc_sock_enqueue(svsk);
1025 
1026 	err = kernel_getpeername(newsock, sin, &slen);
1027 	if (err < 0) {
1028 		if (net_ratelimit())
1029 			printk(KERN_WARNING "%s: peername failed (err %d)!\n",
1030 				   serv->sv_name, -err);
1031 		goto failed;		/* aborted connection or whatever */
1032 	}
1033 
1034 	/* Ideally, we would want to reject connections from unauthorized
1035 	 * hosts here, but when we get encryption, the IP of the host won't
1036 	 * tell us anything.  For now just warn about unpriv connections.
1037 	 */
1038 	if (!svc_port_is_privileged(sin)) {
1039 		dprintk(KERN_WARNING
1040 			"%s: connect from unprivileged port: %s\n",
1041 			serv->sv_name,
1042 			__svc_print_addr(sin, buf, sizeof(buf)));
1043 	}
1044 	dprintk("%s: connect from %s\n", serv->sv_name,
1045 		__svc_print_addr(sin, buf, sizeof(buf)));
1046 
1047 	/* make sure that a write doesn't block forever when
1048 	 * low on memory
1049 	 */
1050 	newsock->sk->sk_sndtimeo = HZ*30;
1051 
1052 	if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
1053 				 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
1054 		goto failed;
1055 	memcpy(&newsvsk->sk_remote, sin, slen);
1056 	newsvsk->sk_remotelen = slen;
1057 
1058 	svc_sock_received(newsvsk);
1059 
1060 	/* make sure that we don't have too many active connections.
1061 	 * If we have, something must be dropped.
1062 	 *
1063 	 * There's no point in trying to do random drop here for
1064 	 * DoS prevention. The NFS clients does 1 reconnect in 15
1065 	 * seconds. An attacker can easily beat that.
1066 	 *
1067 	 * The only somewhat efficient mechanism would be if drop
1068 	 * old connections from the same IP first. But right now
1069 	 * we don't even record the client IP in svc_sock.
1070 	 */
1071 	if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
1072 		struct svc_sock *svsk = NULL;
1073 		spin_lock_bh(&serv->sv_lock);
1074 		if (!list_empty(&serv->sv_tempsocks)) {
1075 			if (net_ratelimit()) {
1076 				/* Try to help the admin */
1077 				printk(KERN_NOTICE "%s: too many open TCP "
1078 					"sockets, consider increasing the "
1079 					"number of nfsd threads\n",
1080 						   serv->sv_name);
1081 				printk(KERN_NOTICE
1082 				       "%s: last TCP connect from %s\n",
1083 				       serv->sv_name, buf);
1084 			}
1085 			/*
1086 			 * Always select the oldest socket. It's not fair,
1087 			 * but so is life
1088 			 */
1089 			svsk = list_entry(serv->sv_tempsocks.prev,
1090 					  struct svc_sock,
1091 					  sk_list);
1092 			set_bit(SK_CLOSE, &svsk->sk_flags);
1093 			atomic_inc(&svsk->sk_inuse);
1094 		}
1095 		spin_unlock_bh(&serv->sv_lock);
1096 
1097 		if (svsk) {
1098 			svc_sock_enqueue(svsk);
1099 			svc_sock_put(svsk);
1100 		}
1101 
1102 	}
1103 
1104 	if (serv->sv_stats)
1105 		serv->sv_stats->nettcpconn++;
1106 
1107 	return;
1108 
1109 failed:
1110 	sock_release(newsock);
1111 	return;
1112 }
1113 
1114 /*
1115  * Receive data from a TCP socket.
1116  */
1117 static int
1118 svc_tcp_recvfrom(struct svc_rqst *rqstp)
1119 {
1120 	struct svc_sock	*svsk = rqstp->rq_sock;
1121 	struct svc_serv	*serv = svsk->sk_server;
1122 	int		len;
1123 	struct kvec *vec;
1124 	int pnum, vlen;
1125 
1126 	dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
1127 		svsk, test_bit(SK_DATA, &svsk->sk_flags),
1128 		test_bit(SK_CONN, &svsk->sk_flags),
1129 		test_bit(SK_CLOSE, &svsk->sk_flags));
1130 
1131 	if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
1132 		svc_sock_received(svsk);
1133 		return svc_deferred_recv(rqstp);
1134 	}
1135 
1136 	if (test_bit(SK_CLOSE, &svsk->sk_flags)) {
1137 		svc_delete_socket(svsk);
1138 		return 0;
1139 	}
1140 
1141 	if (svsk->sk_sk->sk_state == TCP_LISTEN) {
1142 		svc_tcp_accept(svsk);
1143 		svc_sock_received(svsk);
1144 		return 0;
1145 	}
1146 
1147 	if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
1148 		/* sndbuf needs to have room for one request
1149 		 * per thread, otherwise we can stall even when the
1150 		 * network isn't a bottleneck.
1151 		 *
1152 		 * We count all threads rather than threads in a
1153 		 * particular pool, which provides an upper bound
1154 		 * on the number of threads which will access the socket.
1155 		 *
1156 		 * rcvbuf just needs to be able to hold a few requests.
1157 		 * Normally they will be removed from the queue
1158 		 * as soon a a complete request arrives.
1159 		 */
1160 		svc_sock_setbufsize(svsk->sk_sock,
1161 				    (serv->sv_nrthreads+3) * serv->sv_max_mesg,
1162 				    3 * serv->sv_max_mesg);
1163 
1164 	clear_bit(SK_DATA, &svsk->sk_flags);
1165 
1166 	/* Receive data. If we haven't got the record length yet, get
1167 	 * the next four bytes. Otherwise try to gobble up as much as
1168 	 * possible up to the complete record length.
1169 	 */
1170 	if (svsk->sk_tcplen < 4) {
1171 		unsigned long	want = 4 - svsk->sk_tcplen;
1172 		struct kvec	iov;
1173 
1174 		iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
1175 		iov.iov_len  = want;
1176 		if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
1177 			goto error;
1178 		svsk->sk_tcplen += len;
1179 
1180 		if (len < want) {
1181 			dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
1182 				len, want);
1183 			svc_sock_received(svsk);
1184 			return -EAGAIN; /* record header not complete */
1185 		}
1186 
1187 		svsk->sk_reclen = ntohl(svsk->sk_reclen);
1188 		if (!(svsk->sk_reclen & 0x80000000)) {
1189 			/* FIXME: technically, a record can be fragmented,
1190 			 *  and non-terminal fragments will not have the top
1191 			 *  bit set in the fragment length header.
1192 			 *  But apparently no known nfs clients send fragmented
1193 			 *  records. */
1194 			if (net_ratelimit())
1195 				printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1196 				       " (non-terminal)\n",
1197 				       (unsigned long) svsk->sk_reclen);
1198 			goto err_delete;
1199 		}
1200 		svsk->sk_reclen &= 0x7fffffff;
1201 		dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
1202 		if (svsk->sk_reclen > serv->sv_max_mesg) {
1203 			if (net_ratelimit())
1204 				printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1205 				       " (large)\n",
1206 				       (unsigned long) svsk->sk_reclen);
1207 			goto err_delete;
1208 		}
1209 	}
1210 
1211 	/* Check whether enough data is available */
1212 	len = svc_recv_available(svsk);
1213 	if (len < 0)
1214 		goto error;
1215 
1216 	if (len < svsk->sk_reclen) {
1217 		dprintk("svc: incomplete TCP record (%d of %d)\n",
1218 			len, svsk->sk_reclen);
1219 		svc_sock_received(svsk);
1220 		return -EAGAIN;	/* record not complete */
1221 	}
1222 	len = svsk->sk_reclen;
1223 	set_bit(SK_DATA, &svsk->sk_flags);
1224 
1225 	vec = rqstp->rq_vec;
1226 	vec[0] = rqstp->rq_arg.head[0];
1227 	vlen = PAGE_SIZE;
1228 	pnum = 1;
1229 	while (vlen < len) {
1230 		vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
1231 		vec[pnum].iov_len = PAGE_SIZE;
1232 		pnum++;
1233 		vlen += PAGE_SIZE;
1234 	}
1235 	rqstp->rq_respages = &rqstp->rq_pages[pnum];
1236 
1237 	/* Now receive data */
1238 	len = svc_recvfrom(rqstp, vec, pnum, len);
1239 	if (len < 0)
1240 		goto error;
1241 
1242 	dprintk("svc: TCP complete record (%d bytes)\n", len);
1243 	rqstp->rq_arg.len = len;
1244 	rqstp->rq_arg.page_base = 0;
1245 	if (len <= rqstp->rq_arg.head[0].iov_len) {
1246 		rqstp->rq_arg.head[0].iov_len = len;
1247 		rqstp->rq_arg.page_len = 0;
1248 	} else {
1249 		rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1250 	}
1251 
1252 	rqstp->rq_skbuff      = NULL;
1253 	rqstp->rq_prot	      = IPPROTO_TCP;
1254 
1255 	/* Reset TCP read info */
1256 	svsk->sk_reclen = 0;
1257 	svsk->sk_tcplen = 0;
1258 
1259 	svc_sock_received(svsk);
1260 	if (serv->sv_stats)
1261 		serv->sv_stats->nettcpcnt++;
1262 
1263 	return len;
1264 
1265  err_delete:
1266 	svc_delete_socket(svsk);
1267 	return -EAGAIN;
1268 
1269  error:
1270 	if (len == -EAGAIN) {
1271 		dprintk("RPC: TCP recvfrom got EAGAIN\n");
1272 		svc_sock_received(svsk);
1273 	} else {
1274 		printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1275 					svsk->sk_server->sv_name, -len);
1276 		goto err_delete;
1277 	}
1278 
1279 	return len;
1280 }
1281 
1282 /*
1283  * Send out data on TCP socket.
1284  */
1285 static int
1286 svc_tcp_sendto(struct svc_rqst *rqstp)
1287 {
1288 	struct xdr_buf	*xbufp = &rqstp->rq_res;
1289 	int sent;
1290 	__be32 reclen;
1291 
1292 	/* Set up the first element of the reply kvec.
1293 	 * Any other kvecs that may be in use have been taken
1294 	 * care of by the server implementation itself.
1295 	 */
1296 	reclen = htonl(0x80000000|((xbufp->len ) - 4));
1297 	memcpy(xbufp->head[0].iov_base, &reclen, 4);
1298 
1299 	if (test_bit(SK_DEAD, &rqstp->rq_sock->sk_flags))
1300 		return -ENOTCONN;
1301 
1302 	sent = svc_sendto(rqstp, &rqstp->rq_res);
1303 	if (sent != xbufp->len) {
1304 		printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1305 		       rqstp->rq_sock->sk_server->sv_name,
1306 		       (sent<0)?"got error":"sent only",
1307 		       sent, xbufp->len);
1308 		set_bit(SK_CLOSE, &rqstp->rq_sock->sk_flags);
1309 		svc_sock_enqueue(rqstp->rq_sock);
1310 		sent = -EAGAIN;
1311 	}
1312 	return sent;
1313 }
1314 
1315 static void
1316 svc_tcp_init(struct svc_sock *svsk)
1317 {
1318 	struct sock	*sk = svsk->sk_sk;
1319 	struct tcp_sock *tp = tcp_sk(sk);
1320 
1321 	svsk->sk_recvfrom = svc_tcp_recvfrom;
1322 	svsk->sk_sendto = svc_tcp_sendto;
1323 
1324 	if (sk->sk_state == TCP_LISTEN) {
1325 		dprintk("setting up TCP socket for listening\n");
1326 		sk->sk_data_ready = svc_tcp_listen_data_ready;
1327 		set_bit(SK_CONN, &svsk->sk_flags);
1328 	} else {
1329 		dprintk("setting up TCP socket for reading\n");
1330 		sk->sk_state_change = svc_tcp_state_change;
1331 		sk->sk_data_ready = svc_tcp_data_ready;
1332 		sk->sk_write_space = svc_write_space;
1333 
1334 		svsk->sk_reclen = 0;
1335 		svsk->sk_tcplen = 0;
1336 
1337 		tp->nonagle = 1;        /* disable Nagle's algorithm */
1338 
1339 		/* initialise setting must have enough space to
1340 		 * receive and respond to one request.
1341 		 * svc_tcp_recvfrom will re-adjust if necessary
1342 		 */
1343 		svc_sock_setbufsize(svsk->sk_sock,
1344 				    3 * svsk->sk_server->sv_max_mesg,
1345 				    3 * svsk->sk_server->sv_max_mesg);
1346 
1347 		set_bit(SK_CHNGBUF, &svsk->sk_flags);
1348 		set_bit(SK_DATA, &svsk->sk_flags);
1349 		if (sk->sk_state != TCP_ESTABLISHED)
1350 			set_bit(SK_CLOSE, &svsk->sk_flags);
1351 	}
1352 }
1353 
1354 void
1355 svc_sock_update_bufs(struct svc_serv *serv)
1356 {
1357 	/*
1358 	 * The number of server threads has changed. Update
1359 	 * rcvbuf and sndbuf accordingly on all sockets
1360 	 */
1361 	struct list_head *le;
1362 
1363 	spin_lock_bh(&serv->sv_lock);
1364 	list_for_each(le, &serv->sv_permsocks) {
1365 		struct svc_sock *svsk =
1366 			list_entry(le, struct svc_sock, sk_list);
1367 		set_bit(SK_CHNGBUF, &svsk->sk_flags);
1368 	}
1369 	list_for_each(le, &serv->sv_tempsocks) {
1370 		struct svc_sock *svsk =
1371 			list_entry(le, struct svc_sock, sk_list);
1372 		set_bit(SK_CHNGBUF, &svsk->sk_flags);
1373 	}
1374 	spin_unlock_bh(&serv->sv_lock);
1375 }
1376 
1377 /*
1378  * Receive the next request on any socket.  This code is carefully
1379  * organised not to touch any cachelines in the shared svc_serv
1380  * structure, only cachelines in the local svc_pool.
1381  */
1382 int
1383 svc_recv(struct svc_rqst *rqstp, long timeout)
1384 {
1385 	struct svc_sock		*svsk = NULL;
1386 	struct svc_serv		*serv = rqstp->rq_server;
1387 	struct svc_pool		*pool = rqstp->rq_pool;
1388 	int			len, i;
1389 	int 			pages;
1390 	struct xdr_buf		*arg;
1391 	DECLARE_WAITQUEUE(wait, current);
1392 
1393 	dprintk("svc: server %p waiting for data (to = %ld)\n",
1394 		rqstp, timeout);
1395 
1396 	if (rqstp->rq_sock)
1397 		printk(KERN_ERR
1398 			"svc_recv: service %p, socket not NULL!\n",
1399 			 rqstp);
1400 	if (waitqueue_active(&rqstp->rq_wait))
1401 		printk(KERN_ERR
1402 			"svc_recv: service %p, wait queue active!\n",
1403 			 rqstp);
1404 
1405 
1406 	/* now allocate needed pages.  If we get a failure, sleep briefly */
1407 	pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
1408 	for (i=0; i < pages ; i++)
1409 		while (rqstp->rq_pages[i] == NULL) {
1410 			struct page *p = alloc_page(GFP_KERNEL);
1411 			if (!p)
1412 				schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1413 			rqstp->rq_pages[i] = p;
1414 		}
1415 	rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
1416 	BUG_ON(pages >= RPCSVC_MAXPAGES);
1417 
1418 	/* Make arg->head point to first page and arg->pages point to rest */
1419 	arg = &rqstp->rq_arg;
1420 	arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
1421 	arg->head[0].iov_len = PAGE_SIZE;
1422 	arg->pages = rqstp->rq_pages + 1;
1423 	arg->page_base = 0;
1424 	/* save at least one page for response */
1425 	arg->page_len = (pages-2)*PAGE_SIZE;
1426 	arg->len = (pages-1)*PAGE_SIZE;
1427 	arg->tail[0].iov_len = 0;
1428 
1429 	try_to_freeze();
1430 	cond_resched();
1431 	if (signalled())
1432 		return -EINTR;
1433 
1434 	spin_lock_bh(&pool->sp_lock);
1435 	if ((svsk = svc_sock_dequeue(pool)) != NULL) {
1436 		rqstp->rq_sock = svsk;
1437 		atomic_inc(&svsk->sk_inuse);
1438 		rqstp->rq_reserved = serv->sv_max_mesg;
1439 		atomic_add(rqstp->rq_reserved, &svsk->sk_reserved);
1440 	} else {
1441 		/* No data pending. Go to sleep */
1442 		svc_thread_enqueue(pool, rqstp);
1443 
1444 		/*
1445 		 * We have to be able to interrupt this wait
1446 		 * to bring down the daemons ...
1447 		 */
1448 		set_current_state(TASK_INTERRUPTIBLE);
1449 		add_wait_queue(&rqstp->rq_wait, &wait);
1450 		spin_unlock_bh(&pool->sp_lock);
1451 
1452 		schedule_timeout(timeout);
1453 
1454 		try_to_freeze();
1455 
1456 		spin_lock_bh(&pool->sp_lock);
1457 		remove_wait_queue(&rqstp->rq_wait, &wait);
1458 
1459 		if (!(svsk = rqstp->rq_sock)) {
1460 			svc_thread_dequeue(pool, rqstp);
1461 			spin_unlock_bh(&pool->sp_lock);
1462 			dprintk("svc: server %p, no data yet\n", rqstp);
1463 			return signalled()? -EINTR : -EAGAIN;
1464 		}
1465 	}
1466 	spin_unlock_bh(&pool->sp_lock);
1467 
1468 	dprintk("svc: server %p, pool %u, socket %p, inuse=%d\n",
1469 		 rqstp, pool->sp_id, svsk, atomic_read(&svsk->sk_inuse));
1470 	len = svsk->sk_recvfrom(rqstp);
1471 	dprintk("svc: got len=%d\n", len);
1472 
1473 	/* No data, incomplete (TCP) read, or accept() */
1474 	if (len == 0 || len == -EAGAIN) {
1475 		rqstp->rq_res.len = 0;
1476 		svc_sock_release(rqstp);
1477 		return -EAGAIN;
1478 	}
1479 	svsk->sk_lastrecv = get_seconds();
1480 	clear_bit(SK_OLD, &svsk->sk_flags);
1481 
1482 	rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
1483 	rqstp->rq_chandle.defer = svc_defer;
1484 
1485 	if (serv->sv_stats)
1486 		serv->sv_stats->netcnt++;
1487 	return len;
1488 }
1489 
1490 /*
1491  * Drop request
1492  */
1493 void
1494 svc_drop(struct svc_rqst *rqstp)
1495 {
1496 	dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1497 	svc_sock_release(rqstp);
1498 }
1499 
1500 /*
1501  * Return reply to client.
1502  */
1503 int
1504 svc_send(struct svc_rqst *rqstp)
1505 {
1506 	struct svc_sock	*svsk;
1507 	int		len;
1508 	struct xdr_buf	*xb;
1509 
1510 	if ((svsk = rqstp->rq_sock) == NULL) {
1511 		printk(KERN_WARNING "NULL socket pointer in %s:%d\n",
1512 				__FILE__, __LINE__);
1513 		return -EFAULT;
1514 	}
1515 
1516 	/* release the receive skb before sending the reply */
1517 	svc_release_skb(rqstp);
1518 
1519 	/* calculate over-all length */
1520 	xb = & rqstp->rq_res;
1521 	xb->len = xb->head[0].iov_len +
1522 		xb->page_len +
1523 		xb->tail[0].iov_len;
1524 
1525 	/* Grab svsk->sk_mutex to serialize outgoing data. */
1526 	mutex_lock(&svsk->sk_mutex);
1527 	if (test_bit(SK_DEAD, &svsk->sk_flags))
1528 		len = -ENOTCONN;
1529 	else
1530 		len = svsk->sk_sendto(rqstp);
1531 	mutex_unlock(&svsk->sk_mutex);
1532 	svc_sock_release(rqstp);
1533 
1534 	if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1535 		return 0;
1536 	return len;
1537 }
1538 
1539 /*
1540  * Timer function to close old temporary sockets, using
1541  * a mark-and-sweep algorithm.
1542  */
1543 static void
1544 svc_age_temp_sockets(unsigned long closure)
1545 {
1546 	struct svc_serv *serv = (struct svc_serv *)closure;
1547 	struct svc_sock *svsk;
1548 	struct list_head *le, *next;
1549 	LIST_HEAD(to_be_aged);
1550 
1551 	dprintk("svc_age_temp_sockets\n");
1552 
1553 	if (!spin_trylock_bh(&serv->sv_lock)) {
1554 		/* busy, try again 1 sec later */
1555 		dprintk("svc_age_temp_sockets: busy\n");
1556 		mod_timer(&serv->sv_temptimer, jiffies + HZ);
1557 		return;
1558 	}
1559 
1560 	list_for_each_safe(le, next, &serv->sv_tempsocks) {
1561 		svsk = list_entry(le, struct svc_sock, sk_list);
1562 
1563 		if (!test_and_set_bit(SK_OLD, &svsk->sk_flags))
1564 			continue;
1565 		if (atomic_read(&svsk->sk_inuse) || test_bit(SK_BUSY, &svsk->sk_flags))
1566 			continue;
1567 		atomic_inc(&svsk->sk_inuse);
1568 		list_move(le, &to_be_aged);
1569 		set_bit(SK_CLOSE, &svsk->sk_flags);
1570 		set_bit(SK_DETACHED, &svsk->sk_flags);
1571 	}
1572 	spin_unlock_bh(&serv->sv_lock);
1573 
1574 	while (!list_empty(&to_be_aged)) {
1575 		le = to_be_aged.next;
1576 		/* fiddling the sk_list node is safe 'cos we're SK_DETACHED */
1577 		list_del_init(le);
1578 		svsk = list_entry(le, struct svc_sock, sk_list);
1579 
1580 		dprintk("queuing svsk %p for closing, %lu seconds old\n",
1581 			svsk, get_seconds() - svsk->sk_lastrecv);
1582 
1583 		/* a thread will dequeue and close it soon */
1584 		svc_sock_enqueue(svsk);
1585 		svc_sock_put(svsk);
1586 	}
1587 
1588 	mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
1589 }
1590 
1591 /*
1592  * Initialize socket for RPC use and create svc_sock struct
1593  * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1594  */
1595 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1596 						struct socket *sock,
1597 						int *errp, int flags)
1598 {
1599 	struct svc_sock	*svsk;
1600 	struct sock	*inet;
1601 	int		pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1602 	int		is_temporary = flags & SVC_SOCK_TEMPORARY;
1603 
1604 	dprintk("svc: svc_setup_socket %p\n", sock);
1605 	if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1606 		*errp = -ENOMEM;
1607 		return NULL;
1608 	}
1609 
1610 	inet = sock->sk;
1611 
1612 	/* Register socket with portmapper */
1613 	if (*errp >= 0 && pmap_register)
1614 		*errp = svc_register(serv, inet->sk_protocol,
1615 				     ntohs(inet_sk(inet)->sport));
1616 
1617 	if (*errp < 0) {
1618 		kfree(svsk);
1619 		return NULL;
1620 	}
1621 
1622 	set_bit(SK_BUSY, &svsk->sk_flags);
1623 	inet->sk_user_data = svsk;
1624 	svsk->sk_sock = sock;
1625 	svsk->sk_sk = inet;
1626 	svsk->sk_ostate = inet->sk_state_change;
1627 	svsk->sk_odata = inet->sk_data_ready;
1628 	svsk->sk_owspace = inet->sk_write_space;
1629 	svsk->sk_server = serv;
1630 	atomic_set(&svsk->sk_inuse, 1);
1631 	svsk->sk_lastrecv = get_seconds();
1632 	spin_lock_init(&svsk->sk_defer_lock);
1633 	INIT_LIST_HEAD(&svsk->sk_deferred);
1634 	INIT_LIST_HEAD(&svsk->sk_ready);
1635 	mutex_init(&svsk->sk_mutex);
1636 
1637 	/* Initialize the socket */
1638 	if (sock->type == SOCK_DGRAM)
1639 		svc_udp_init(svsk);
1640 	else
1641 		svc_tcp_init(svsk);
1642 
1643 	spin_lock_bh(&serv->sv_lock);
1644 	if (is_temporary) {
1645 		set_bit(SK_TEMP, &svsk->sk_flags);
1646 		list_add(&svsk->sk_list, &serv->sv_tempsocks);
1647 		serv->sv_tmpcnt++;
1648 		if (serv->sv_temptimer.function == NULL) {
1649 			/* setup timer to age temp sockets */
1650 			setup_timer(&serv->sv_temptimer, svc_age_temp_sockets,
1651 					(unsigned long)serv);
1652 			mod_timer(&serv->sv_temptimer,
1653 					jiffies + svc_conn_age_period * HZ);
1654 		}
1655 	} else {
1656 		clear_bit(SK_TEMP, &svsk->sk_flags);
1657 		list_add(&svsk->sk_list, &serv->sv_permsocks);
1658 	}
1659 	spin_unlock_bh(&serv->sv_lock);
1660 
1661 	dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1662 				svsk, svsk->sk_sk);
1663 
1664 	return svsk;
1665 }
1666 
1667 int svc_addsock(struct svc_serv *serv,
1668 		int fd,
1669 		char *name_return,
1670 		int *proto)
1671 {
1672 	int err = 0;
1673 	struct socket *so = sockfd_lookup(fd, &err);
1674 	struct svc_sock *svsk = NULL;
1675 
1676 	if (!so)
1677 		return err;
1678 	if (so->sk->sk_family != AF_INET)
1679 		err =  -EAFNOSUPPORT;
1680 	else if (so->sk->sk_protocol != IPPROTO_TCP &&
1681 	    so->sk->sk_protocol != IPPROTO_UDP)
1682 		err =  -EPROTONOSUPPORT;
1683 	else if (so->state > SS_UNCONNECTED)
1684 		err = -EISCONN;
1685 	else {
1686 		svsk = svc_setup_socket(serv, so, &err, SVC_SOCK_DEFAULTS);
1687 		if (svsk) {
1688 			svc_sock_received(svsk);
1689 			err = 0;
1690 		}
1691 	}
1692 	if (err) {
1693 		sockfd_put(so);
1694 		return err;
1695 	}
1696 	if (proto) *proto = so->sk->sk_protocol;
1697 	return one_sock_name(name_return, svsk);
1698 }
1699 EXPORT_SYMBOL_GPL(svc_addsock);
1700 
1701 /*
1702  * Create socket for RPC service.
1703  */
1704 static int svc_create_socket(struct svc_serv *serv, int protocol,
1705 				struct sockaddr *sin, int len, int flags)
1706 {
1707 	struct svc_sock	*svsk;
1708 	struct socket	*sock;
1709 	int		error;
1710 	int		type;
1711 	char		buf[RPC_MAX_ADDRBUFLEN];
1712 
1713 	dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1714 			serv->sv_program->pg_name, protocol,
1715 			__svc_print_addr(sin, buf, sizeof(buf)));
1716 
1717 	if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1718 		printk(KERN_WARNING "svc: only UDP and TCP "
1719 				"sockets supported\n");
1720 		return -EINVAL;
1721 	}
1722 	type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1723 
1724 	error = sock_create_kern(sin->sa_family, type, protocol, &sock);
1725 	if (error < 0)
1726 		return error;
1727 
1728 	svc_reclassify_socket(sock);
1729 
1730 	if (type == SOCK_STREAM)
1731 		sock->sk->sk_reuse = 1;		/* allow address reuse */
1732 	error = kernel_bind(sock, sin, len);
1733 	if (error < 0)
1734 		goto bummer;
1735 
1736 	if (protocol == IPPROTO_TCP) {
1737 		if ((error = kernel_listen(sock, 64)) < 0)
1738 			goto bummer;
1739 	}
1740 
1741 	if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1742 		svc_sock_received(svsk);
1743 		return ntohs(inet_sk(svsk->sk_sk)->sport);
1744 	}
1745 
1746 bummer:
1747 	dprintk("svc: svc_create_socket error = %d\n", -error);
1748 	sock_release(sock);
1749 	return error;
1750 }
1751 
1752 /*
1753  * Remove a dead socket
1754  */
1755 static void
1756 svc_delete_socket(struct svc_sock *svsk)
1757 {
1758 	struct svc_serv	*serv;
1759 	struct sock	*sk;
1760 
1761 	dprintk("svc: svc_delete_socket(%p)\n", svsk);
1762 
1763 	serv = svsk->sk_server;
1764 	sk = svsk->sk_sk;
1765 
1766 	sk->sk_state_change = svsk->sk_ostate;
1767 	sk->sk_data_ready = svsk->sk_odata;
1768 	sk->sk_write_space = svsk->sk_owspace;
1769 
1770 	spin_lock_bh(&serv->sv_lock);
1771 
1772 	if (!test_and_set_bit(SK_DETACHED, &svsk->sk_flags))
1773 		list_del_init(&svsk->sk_list);
1774 	/*
1775 	 * We used to delete the svc_sock from whichever list
1776 	 * it's sk_ready node was on, but we don't actually
1777 	 * need to.  This is because the only time we're called
1778 	 * while still attached to a queue, the queue itself
1779 	 * is about to be destroyed (in svc_destroy).
1780 	 */
1781 	if (!test_and_set_bit(SK_DEAD, &svsk->sk_flags)) {
1782 		BUG_ON(atomic_read(&svsk->sk_inuse)<2);
1783 		atomic_dec(&svsk->sk_inuse);
1784 		if (test_bit(SK_TEMP, &svsk->sk_flags))
1785 			serv->sv_tmpcnt--;
1786 	}
1787 
1788 	spin_unlock_bh(&serv->sv_lock);
1789 }
1790 
1791 static void svc_close_socket(struct svc_sock *svsk)
1792 {
1793 	set_bit(SK_CLOSE, &svsk->sk_flags);
1794 	if (test_and_set_bit(SK_BUSY, &svsk->sk_flags))
1795 		/* someone else will have to effect the close */
1796 		return;
1797 
1798 	atomic_inc(&svsk->sk_inuse);
1799 	svc_delete_socket(svsk);
1800 	clear_bit(SK_BUSY, &svsk->sk_flags);
1801 	svc_sock_put(svsk);
1802 }
1803 
1804 void svc_force_close_socket(struct svc_sock *svsk)
1805 {
1806 	set_bit(SK_CLOSE, &svsk->sk_flags);
1807 	if (test_bit(SK_BUSY, &svsk->sk_flags)) {
1808 		/* Waiting to be processed, but no threads left,
1809 		 * So just remove it from the waiting list
1810 		 */
1811 		list_del_init(&svsk->sk_ready);
1812 		clear_bit(SK_BUSY, &svsk->sk_flags);
1813 	}
1814 	svc_close_socket(svsk);
1815 }
1816 
1817 /**
1818  * svc_makesock - Make a socket for nfsd and lockd
1819  * @serv: RPC server structure
1820  * @protocol: transport protocol to use
1821  * @port: port to use
1822  * @flags: requested socket characteristics
1823  *
1824  */
1825 int svc_makesock(struct svc_serv *serv, int protocol, unsigned short port,
1826 			int flags)
1827 {
1828 	struct sockaddr_in sin = {
1829 		.sin_family		= AF_INET,
1830 		.sin_addr.s_addr	= INADDR_ANY,
1831 		.sin_port		= htons(port),
1832 	};
1833 
1834 	dprintk("svc: creating socket proto = %d\n", protocol);
1835 	return svc_create_socket(serv, protocol, (struct sockaddr *) &sin,
1836 							sizeof(sin), flags);
1837 }
1838 
1839 /*
1840  * Handle defer and revisit of requests
1841  */
1842 
1843 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1844 {
1845 	struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1846 	struct svc_sock *svsk;
1847 
1848 	if (too_many) {
1849 		svc_sock_put(dr->svsk);
1850 		kfree(dr);
1851 		return;
1852 	}
1853 	dprintk("revisit queued\n");
1854 	svsk = dr->svsk;
1855 	dr->svsk = NULL;
1856 	spin_lock_bh(&svsk->sk_defer_lock);
1857 	list_add(&dr->handle.recent, &svsk->sk_deferred);
1858 	spin_unlock_bh(&svsk->sk_defer_lock);
1859 	set_bit(SK_DEFERRED, &svsk->sk_flags);
1860 	svc_sock_enqueue(svsk);
1861 	svc_sock_put(svsk);
1862 }
1863 
1864 static struct cache_deferred_req *
1865 svc_defer(struct cache_req *req)
1866 {
1867 	struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1868 	int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
1869 	struct svc_deferred_req *dr;
1870 
1871 	if (rqstp->rq_arg.page_len)
1872 		return NULL; /* if more than a page, give up FIXME */
1873 	if (rqstp->rq_deferred) {
1874 		dr = rqstp->rq_deferred;
1875 		rqstp->rq_deferred = NULL;
1876 	} else {
1877 		int skip  = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1878 		/* FIXME maybe discard if size too large */
1879 		dr = kmalloc(size, GFP_KERNEL);
1880 		if (dr == NULL)
1881 			return NULL;
1882 
1883 		dr->handle.owner = rqstp->rq_server;
1884 		dr->prot = rqstp->rq_prot;
1885 		memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1886 		dr->addrlen = rqstp->rq_addrlen;
1887 		dr->daddr = rqstp->rq_daddr;
1888 		dr->argslen = rqstp->rq_arg.len >> 2;
1889 		memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
1890 	}
1891 	atomic_inc(&rqstp->rq_sock->sk_inuse);
1892 	dr->svsk = rqstp->rq_sock;
1893 
1894 	dr->handle.revisit = svc_revisit;
1895 	return &dr->handle;
1896 }
1897 
1898 /*
1899  * recv data from a deferred request into an active one
1900  */
1901 static int svc_deferred_recv(struct svc_rqst *rqstp)
1902 {
1903 	struct svc_deferred_req *dr = rqstp->rq_deferred;
1904 
1905 	rqstp->rq_arg.head[0].iov_base = dr->args;
1906 	rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
1907 	rqstp->rq_arg.page_len = 0;
1908 	rqstp->rq_arg.len = dr->argslen<<2;
1909 	rqstp->rq_prot        = dr->prot;
1910 	memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1911 	rqstp->rq_addrlen     = dr->addrlen;
1912 	rqstp->rq_daddr       = dr->daddr;
1913 	rqstp->rq_respages    = rqstp->rq_pages;
1914 	return dr->argslen<<2;
1915 }
1916 
1917 
1918 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk)
1919 {
1920 	struct svc_deferred_req *dr = NULL;
1921 
1922 	if (!test_bit(SK_DEFERRED, &svsk->sk_flags))
1923 		return NULL;
1924 	spin_lock_bh(&svsk->sk_defer_lock);
1925 	clear_bit(SK_DEFERRED, &svsk->sk_flags);
1926 	if (!list_empty(&svsk->sk_deferred)) {
1927 		dr = list_entry(svsk->sk_deferred.next,
1928 				struct svc_deferred_req,
1929 				handle.recent);
1930 		list_del_init(&dr->handle.recent);
1931 		set_bit(SK_DEFERRED, &svsk->sk_flags);
1932 	}
1933 	spin_unlock_bh(&svsk->sk_defer_lock);
1934 	return dr;
1935 }
1936