xref: /openbmc/linux/net/sunrpc/svcsock.c (revision 87c2ce3b)
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 <net/sock.h>
35 #include <net/checksum.h>
36 #include <net/ip.h>
37 #include <net/tcp_states.h>
38 #include <asm/uaccess.h>
39 #include <asm/ioctls.h>
40 
41 #include <linux/sunrpc/types.h>
42 #include <linux/sunrpc/xdr.h>
43 #include <linux/sunrpc/svcsock.h>
44 #include <linux/sunrpc/stats.h>
45 
46 /* SMP locking strategy:
47  *
48  * 	svc_serv->sv_lock protects most stuff for that service.
49  *
50  *	Some flags can be set to certain values at any time
51  *	providing that certain rules are followed:
52  *
53  *	SK_BUSY  can be set to 0 at any time.
54  *		svc_sock_enqueue must be called afterwards
55  *	SK_CONN, SK_DATA, can be set or cleared at any time.
56  *		after a set, svc_sock_enqueue must be called.
57  *		after a clear, the socket must be read/accepted
58  *		 if this succeeds, it must be set again.
59  *	SK_CLOSE can set at any time. It is never cleared.
60  *
61  */
62 
63 #define RPCDBG_FACILITY	RPCDBG_SVCSOCK
64 
65 
66 static struct svc_sock *svc_setup_socket(struct svc_serv *, struct socket *,
67 					 int *errp, int pmap_reg);
68 static void		svc_udp_data_ready(struct sock *, int);
69 static int		svc_udp_recvfrom(struct svc_rqst *);
70 static int		svc_udp_sendto(struct svc_rqst *);
71 
72 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk);
73 static int svc_deferred_recv(struct svc_rqst *rqstp);
74 static struct cache_deferred_req *svc_defer(struct cache_req *req);
75 
76 /*
77  * Queue up an idle server thread.  Must have serv->sv_lock held.
78  * Note: this is really a stack rather than a queue, so that we only
79  * use as many different threads as we need, and the rest don't polute
80  * the cache.
81  */
82 static inline void
83 svc_serv_enqueue(struct svc_serv *serv, struct svc_rqst *rqstp)
84 {
85 	list_add(&rqstp->rq_list, &serv->sv_threads);
86 }
87 
88 /*
89  * Dequeue an nfsd thread.  Must have serv->sv_lock held.
90  */
91 static inline void
92 svc_serv_dequeue(struct svc_serv *serv, struct svc_rqst *rqstp)
93 {
94 	list_del(&rqstp->rq_list);
95 }
96 
97 /*
98  * Release an skbuff after use
99  */
100 static inline void
101 svc_release_skb(struct svc_rqst *rqstp)
102 {
103 	struct sk_buff *skb = rqstp->rq_skbuff;
104 	struct svc_deferred_req *dr = rqstp->rq_deferred;
105 
106 	if (skb) {
107 		rqstp->rq_skbuff = NULL;
108 
109 		dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
110 		skb_free_datagram(rqstp->rq_sock->sk_sk, skb);
111 	}
112 	if (dr) {
113 		rqstp->rq_deferred = NULL;
114 		kfree(dr);
115 	}
116 }
117 
118 /*
119  * Any space to write?
120  */
121 static inline unsigned long
122 svc_sock_wspace(struct svc_sock *svsk)
123 {
124 	int wspace;
125 
126 	if (svsk->sk_sock->type == SOCK_STREAM)
127 		wspace = sk_stream_wspace(svsk->sk_sk);
128 	else
129 		wspace = sock_wspace(svsk->sk_sk);
130 
131 	return wspace;
132 }
133 
134 /*
135  * Queue up a socket with data pending. If there are idle nfsd
136  * processes, wake 'em up.
137  *
138  */
139 static void
140 svc_sock_enqueue(struct svc_sock *svsk)
141 {
142 	struct svc_serv	*serv = svsk->sk_server;
143 	struct svc_rqst	*rqstp;
144 
145 	if (!(svsk->sk_flags &
146 	      ( (1<<SK_CONN)|(1<<SK_DATA)|(1<<SK_CLOSE)|(1<<SK_DEFERRED)) ))
147 		return;
148 	if (test_bit(SK_DEAD, &svsk->sk_flags))
149 		return;
150 
151 	spin_lock_bh(&serv->sv_lock);
152 
153 	if (!list_empty(&serv->sv_threads) &&
154 	    !list_empty(&serv->sv_sockets))
155 		printk(KERN_ERR
156 			"svc_sock_enqueue: threads and sockets both waiting??\n");
157 
158 	if (test_bit(SK_DEAD, &svsk->sk_flags)) {
159 		/* Don't enqueue dead sockets */
160 		dprintk("svc: socket %p is dead, not enqueued\n", svsk->sk_sk);
161 		goto out_unlock;
162 	}
163 
164 	if (test_bit(SK_BUSY, &svsk->sk_flags)) {
165 		/* Don't enqueue socket while daemon is receiving */
166 		dprintk("svc: socket %p busy, not enqueued\n", svsk->sk_sk);
167 		goto out_unlock;
168 	}
169 
170 	set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
171 	if (((svsk->sk_reserved + serv->sv_bufsz)*2
172 	     > svc_sock_wspace(svsk))
173 	    && !test_bit(SK_CLOSE, &svsk->sk_flags)
174 	    && !test_bit(SK_CONN, &svsk->sk_flags)) {
175 		/* Don't enqueue while not enough space for reply */
176 		dprintk("svc: socket %p  no space, %d*2 > %ld, not enqueued\n",
177 			svsk->sk_sk, svsk->sk_reserved+serv->sv_bufsz,
178 			svc_sock_wspace(svsk));
179 		goto out_unlock;
180 	}
181 	clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
182 
183 	/* Mark socket as busy. It will remain in this state until the
184 	 * server has processed all pending data and put the socket back
185 	 * on the idle list.
186 	 */
187 	set_bit(SK_BUSY, &svsk->sk_flags);
188 
189 	if (!list_empty(&serv->sv_threads)) {
190 		rqstp = list_entry(serv->sv_threads.next,
191 				   struct svc_rqst,
192 				   rq_list);
193 		dprintk("svc: socket %p served by daemon %p\n",
194 			svsk->sk_sk, rqstp);
195 		svc_serv_dequeue(serv, rqstp);
196 		if (rqstp->rq_sock)
197 			printk(KERN_ERR
198 				"svc_sock_enqueue: server %p, rq_sock=%p!\n",
199 				rqstp, rqstp->rq_sock);
200 		rqstp->rq_sock = svsk;
201 		svsk->sk_inuse++;
202 		rqstp->rq_reserved = serv->sv_bufsz;
203 		svsk->sk_reserved += rqstp->rq_reserved;
204 		wake_up(&rqstp->rq_wait);
205 	} else {
206 		dprintk("svc: socket %p put into queue\n", svsk->sk_sk);
207 		list_add_tail(&svsk->sk_ready, &serv->sv_sockets);
208 	}
209 
210 out_unlock:
211 	spin_unlock_bh(&serv->sv_lock);
212 }
213 
214 /*
215  * Dequeue the first socket.  Must be called with the serv->sv_lock held.
216  */
217 static inline struct svc_sock *
218 svc_sock_dequeue(struct svc_serv *serv)
219 {
220 	struct svc_sock	*svsk;
221 
222 	if (list_empty(&serv->sv_sockets))
223 		return NULL;
224 
225 	svsk = list_entry(serv->sv_sockets.next,
226 			  struct svc_sock, sk_ready);
227 	list_del_init(&svsk->sk_ready);
228 
229 	dprintk("svc: socket %p dequeued, inuse=%d\n",
230 		svsk->sk_sk, svsk->sk_inuse);
231 
232 	return svsk;
233 }
234 
235 /*
236  * Having read something from a socket, check whether it
237  * needs to be re-enqueued.
238  * Note: SK_DATA only gets cleared when a read-attempt finds
239  * no (or insufficient) data.
240  */
241 static inline void
242 svc_sock_received(struct svc_sock *svsk)
243 {
244 	clear_bit(SK_BUSY, &svsk->sk_flags);
245 	svc_sock_enqueue(svsk);
246 }
247 
248 
249 /**
250  * svc_reserve - change the space reserved for the reply to a request.
251  * @rqstp:  The request in question
252  * @space: new max space to reserve
253  *
254  * Each request reserves some space on the output queue of the socket
255  * to make sure the reply fits.  This function reduces that reserved
256  * space to be the amount of space used already, plus @space.
257  *
258  */
259 void svc_reserve(struct svc_rqst *rqstp, int space)
260 {
261 	space += rqstp->rq_res.head[0].iov_len;
262 
263 	if (space < rqstp->rq_reserved) {
264 		struct svc_sock *svsk = rqstp->rq_sock;
265 		spin_lock_bh(&svsk->sk_server->sv_lock);
266 		svsk->sk_reserved -= (rqstp->rq_reserved - space);
267 		rqstp->rq_reserved = space;
268 		spin_unlock_bh(&svsk->sk_server->sv_lock);
269 
270 		svc_sock_enqueue(svsk);
271 	}
272 }
273 
274 /*
275  * Release a socket after use.
276  */
277 static inline void
278 svc_sock_put(struct svc_sock *svsk)
279 {
280 	struct svc_serv *serv = svsk->sk_server;
281 
282 	spin_lock_bh(&serv->sv_lock);
283 	if (!--(svsk->sk_inuse) && test_bit(SK_DEAD, &svsk->sk_flags)) {
284 		spin_unlock_bh(&serv->sv_lock);
285 		dprintk("svc: releasing dead socket\n");
286 		sock_release(svsk->sk_sock);
287 		kfree(svsk);
288 	}
289 	else
290 		spin_unlock_bh(&serv->sv_lock);
291 }
292 
293 static void
294 svc_sock_release(struct svc_rqst *rqstp)
295 {
296 	struct svc_sock	*svsk = rqstp->rq_sock;
297 
298 	svc_release_skb(rqstp);
299 
300 	svc_free_allpages(rqstp);
301 	rqstp->rq_res.page_len = 0;
302 	rqstp->rq_res.page_base = 0;
303 
304 
305 	/* Reset response buffer and release
306 	 * the reservation.
307 	 * But first, check that enough space was reserved
308 	 * for the reply, otherwise we have a bug!
309 	 */
310 	if ((rqstp->rq_res.len) >  rqstp->rq_reserved)
311 		printk(KERN_ERR "RPC request reserved %d but used %d\n",
312 		       rqstp->rq_reserved,
313 		       rqstp->rq_res.len);
314 
315 	rqstp->rq_res.head[0].iov_len = 0;
316 	svc_reserve(rqstp, 0);
317 	rqstp->rq_sock = NULL;
318 
319 	svc_sock_put(svsk);
320 }
321 
322 /*
323  * External function to wake up a server waiting for data
324  */
325 void
326 svc_wake_up(struct svc_serv *serv)
327 {
328 	struct svc_rqst	*rqstp;
329 
330 	spin_lock_bh(&serv->sv_lock);
331 	if (!list_empty(&serv->sv_threads)) {
332 		rqstp = list_entry(serv->sv_threads.next,
333 				   struct svc_rqst,
334 				   rq_list);
335 		dprintk("svc: daemon %p woken up.\n", rqstp);
336 		/*
337 		svc_serv_dequeue(serv, rqstp);
338 		rqstp->rq_sock = NULL;
339 		 */
340 		wake_up(&rqstp->rq_wait);
341 	}
342 	spin_unlock_bh(&serv->sv_lock);
343 }
344 
345 /*
346  * Generic sendto routine
347  */
348 static int
349 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
350 {
351 	struct svc_sock	*svsk = rqstp->rq_sock;
352 	struct socket	*sock = svsk->sk_sock;
353 	int		slen;
354 	char 		buffer[CMSG_SPACE(sizeof(struct in_pktinfo))];
355 	struct cmsghdr *cmh = (struct cmsghdr *)buffer;
356 	struct in_pktinfo *pki = (struct in_pktinfo *)CMSG_DATA(cmh);
357 	int		len = 0;
358 	int		result;
359 	int		size;
360 	struct page	**ppage = xdr->pages;
361 	size_t		base = xdr->page_base;
362 	unsigned int	pglen = xdr->page_len;
363 	unsigned int	flags = MSG_MORE;
364 
365 	slen = xdr->len;
366 
367 	if (rqstp->rq_prot == IPPROTO_UDP) {
368 		/* set the source and destination */
369 		struct msghdr	msg;
370 		msg.msg_name    = &rqstp->rq_addr;
371 		msg.msg_namelen = sizeof(rqstp->rq_addr);
372 		msg.msg_iov     = NULL;
373 		msg.msg_iovlen  = 0;
374 		msg.msg_flags	= MSG_MORE;
375 
376 		msg.msg_control = cmh;
377 		msg.msg_controllen = sizeof(buffer);
378 		cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
379 		cmh->cmsg_level = SOL_IP;
380 		cmh->cmsg_type = IP_PKTINFO;
381 		pki->ipi_ifindex = 0;
382 		pki->ipi_spec_dst.s_addr = rqstp->rq_daddr;
383 
384 		if (sock_sendmsg(sock, &msg, 0) < 0)
385 			goto out;
386 	}
387 
388 	/* send head */
389 	if (slen == xdr->head[0].iov_len)
390 		flags = 0;
391 	len = sock->ops->sendpage(sock, rqstp->rq_respages[0], 0, xdr->head[0].iov_len, flags);
392 	if (len != xdr->head[0].iov_len)
393 		goto out;
394 	slen -= xdr->head[0].iov_len;
395 	if (slen == 0)
396 		goto out;
397 
398 	/* send page data */
399 	size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
400 	while (pglen > 0) {
401 		if (slen == size)
402 			flags = 0;
403 		result = sock->ops->sendpage(sock, *ppage, base, size, flags);
404 		if (result > 0)
405 			len += result;
406 		if (result != size)
407 			goto out;
408 		slen -= size;
409 		pglen -= size;
410 		size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
411 		base = 0;
412 		ppage++;
413 	}
414 	/* send tail */
415 	if (xdr->tail[0].iov_len) {
416 		result = sock->ops->sendpage(sock, rqstp->rq_respages[rqstp->rq_restailpage],
417 					     ((unsigned long)xdr->tail[0].iov_base)& (PAGE_SIZE-1),
418 					     xdr->tail[0].iov_len, 0);
419 
420 		if (result > 0)
421 			len += result;
422 	}
423 out:
424 	dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %x)\n",
425 			rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len, xdr->len, len,
426 		rqstp->rq_addr.sin_addr.s_addr);
427 
428 	return len;
429 }
430 
431 /*
432  * Check input queue length
433  */
434 static int
435 svc_recv_available(struct svc_sock *svsk)
436 {
437 	mm_segment_t	oldfs;
438 	struct socket	*sock = svsk->sk_sock;
439 	int		avail, err;
440 
441 	oldfs = get_fs(); set_fs(KERNEL_DS);
442 	err = sock->ops->ioctl(sock, TIOCINQ, (unsigned long) &avail);
443 	set_fs(oldfs);
444 
445 	return (err >= 0)? avail : err;
446 }
447 
448 /*
449  * Generic recvfrom routine.
450  */
451 static int
452 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
453 {
454 	struct msghdr	msg;
455 	struct socket	*sock;
456 	int		len, alen;
457 
458 	rqstp->rq_addrlen = sizeof(rqstp->rq_addr);
459 	sock = rqstp->rq_sock->sk_sock;
460 
461 	msg.msg_name    = &rqstp->rq_addr;
462 	msg.msg_namelen = sizeof(rqstp->rq_addr);
463 	msg.msg_control = NULL;
464 	msg.msg_controllen = 0;
465 
466 	msg.msg_flags	= MSG_DONTWAIT;
467 
468 	len = kernel_recvmsg(sock, &msg, iov, nr, buflen, MSG_DONTWAIT);
469 
470 	/* sock_recvmsg doesn't fill in the name/namelen, so we must..
471 	 * possibly we should cache this in the svc_sock structure
472 	 * at accept time. FIXME
473 	 */
474 	alen = sizeof(rqstp->rq_addr);
475 	sock->ops->getname(sock, (struct sockaddr *)&rqstp->rq_addr, &alen, 1);
476 
477 	dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
478 		rqstp->rq_sock, iov[0].iov_base, iov[0].iov_len, len);
479 
480 	return len;
481 }
482 
483 /*
484  * Set socket snd and rcv buffer lengths
485  */
486 static inline void
487 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
488 {
489 #if 0
490 	mm_segment_t	oldfs;
491 	oldfs = get_fs(); set_fs(KERNEL_DS);
492 	sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
493 			(char*)&snd, sizeof(snd));
494 	sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
495 			(char*)&rcv, sizeof(rcv));
496 #else
497 	/* sock_setsockopt limits use to sysctl_?mem_max,
498 	 * which isn't acceptable.  Until that is made conditional
499 	 * on not having CAP_SYS_RESOURCE or similar, we go direct...
500 	 * DaveM said I could!
501 	 */
502 	lock_sock(sock->sk);
503 	sock->sk->sk_sndbuf = snd * 2;
504 	sock->sk->sk_rcvbuf = rcv * 2;
505 	sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
506 	release_sock(sock->sk);
507 #endif
508 }
509 /*
510  * INET callback when data has been received on the socket.
511  */
512 static void
513 svc_udp_data_ready(struct sock *sk, int count)
514 {
515 	struct svc_sock	*svsk = (struct svc_sock *)sk->sk_user_data;
516 
517 	if (svsk) {
518 		dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
519 			svsk, sk, count, test_bit(SK_BUSY, &svsk->sk_flags));
520 		set_bit(SK_DATA, &svsk->sk_flags);
521 		svc_sock_enqueue(svsk);
522 	}
523 	if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
524 		wake_up_interruptible(sk->sk_sleep);
525 }
526 
527 /*
528  * INET callback when space is newly available on the socket.
529  */
530 static void
531 svc_write_space(struct sock *sk)
532 {
533 	struct svc_sock	*svsk = (struct svc_sock *)(sk->sk_user_data);
534 
535 	if (svsk) {
536 		dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
537 			svsk, sk, test_bit(SK_BUSY, &svsk->sk_flags));
538 		svc_sock_enqueue(svsk);
539 	}
540 
541 	if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
542 		dprintk("RPC svc_write_space: someone sleeping on %p\n",
543 		       svsk);
544 		wake_up_interruptible(sk->sk_sleep);
545 	}
546 }
547 
548 /*
549  * Receive a datagram from a UDP socket.
550  */
551 static int
552 svc_udp_recvfrom(struct svc_rqst *rqstp)
553 {
554 	struct svc_sock	*svsk = rqstp->rq_sock;
555 	struct svc_serv	*serv = svsk->sk_server;
556 	struct sk_buff	*skb;
557 	int		err, len;
558 
559 	if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
560 	    /* udp sockets need large rcvbuf as all pending
561 	     * requests are still in that buffer.  sndbuf must
562 	     * also be large enough that there is enough space
563 	     * for one reply per thread.
564 	     */
565 	    svc_sock_setbufsize(svsk->sk_sock,
566 				(serv->sv_nrthreads+3) * serv->sv_bufsz,
567 				(serv->sv_nrthreads+3) * serv->sv_bufsz);
568 
569 	if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
570 		svc_sock_received(svsk);
571 		return svc_deferred_recv(rqstp);
572 	}
573 
574 	clear_bit(SK_DATA, &svsk->sk_flags);
575 	while ((skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err)) == NULL) {
576 		if (err == -EAGAIN) {
577 			svc_sock_received(svsk);
578 			return err;
579 		}
580 		/* possibly an icmp error */
581 		dprintk("svc: recvfrom returned error %d\n", -err);
582 	}
583 	if (skb->tstamp.off_sec == 0) {
584 		struct timeval tv;
585 
586 		tv.tv_sec = xtime.tv_sec;
587 		tv.tv_usec = xtime.tv_nsec / NSEC_PER_USEC;
588 		skb_set_timestamp(skb, &tv);
589 		/* Don't enable netstamp, sunrpc doesn't
590 		   need that much accuracy */
591 	}
592 	skb_get_timestamp(skb, &svsk->sk_sk->sk_stamp);
593 	set_bit(SK_DATA, &svsk->sk_flags); /* there may be more data... */
594 
595 	/*
596 	 * Maybe more packets - kick another thread ASAP.
597 	 */
598 	svc_sock_received(svsk);
599 
600 	len  = skb->len - sizeof(struct udphdr);
601 	rqstp->rq_arg.len = len;
602 
603 	rqstp->rq_prot        = IPPROTO_UDP;
604 
605 	/* Get sender address */
606 	rqstp->rq_addr.sin_family = AF_INET;
607 	rqstp->rq_addr.sin_port = skb->h.uh->source;
608 	rqstp->rq_addr.sin_addr.s_addr = skb->nh.iph->saddr;
609 	rqstp->rq_daddr = skb->nh.iph->daddr;
610 
611 	if (skb_is_nonlinear(skb)) {
612 		/* we have to copy */
613 		local_bh_disable();
614 		if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
615 			local_bh_enable();
616 			/* checksum error */
617 			skb_free_datagram(svsk->sk_sk, skb);
618 			return 0;
619 		}
620 		local_bh_enable();
621 		skb_free_datagram(svsk->sk_sk, skb);
622 	} else {
623 		/* we can use it in-place */
624 		rqstp->rq_arg.head[0].iov_base = skb->data + sizeof(struct udphdr);
625 		rqstp->rq_arg.head[0].iov_len = len;
626 		if (skb_checksum_complete(skb)) {
627 			skb_free_datagram(svsk->sk_sk, skb);
628 			return 0;
629 		}
630 		rqstp->rq_skbuff = skb;
631 	}
632 
633 	rqstp->rq_arg.page_base = 0;
634 	if (len <= rqstp->rq_arg.head[0].iov_len) {
635 		rqstp->rq_arg.head[0].iov_len = len;
636 		rqstp->rq_arg.page_len = 0;
637 	} else {
638 		rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
639 		rqstp->rq_argused += (rqstp->rq_arg.page_len + PAGE_SIZE - 1)/ PAGE_SIZE;
640 	}
641 
642 	if (serv->sv_stats)
643 		serv->sv_stats->netudpcnt++;
644 
645 	return len;
646 }
647 
648 static int
649 svc_udp_sendto(struct svc_rqst *rqstp)
650 {
651 	int		error;
652 
653 	error = svc_sendto(rqstp, &rqstp->rq_res);
654 	if (error == -ECONNREFUSED)
655 		/* ICMP error on earlier request. */
656 		error = svc_sendto(rqstp, &rqstp->rq_res);
657 
658 	return error;
659 }
660 
661 static void
662 svc_udp_init(struct svc_sock *svsk)
663 {
664 	svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
665 	svsk->sk_sk->sk_write_space = svc_write_space;
666 	svsk->sk_recvfrom = svc_udp_recvfrom;
667 	svsk->sk_sendto = svc_udp_sendto;
668 
669 	/* initialise setting must have enough space to
670 	 * receive and respond to one request.
671 	 * svc_udp_recvfrom will re-adjust if necessary
672 	 */
673 	svc_sock_setbufsize(svsk->sk_sock,
674 			    3 * svsk->sk_server->sv_bufsz,
675 			    3 * svsk->sk_server->sv_bufsz);
676 
677 	set_bit(SK_DATA, &svsk->sk_flags); /* might have come in before data_ready set up */
678 	set_bit(SK_CHNGBUF, &svsk->sk_flags);
679 }
680 
681 /*
682  * A data_ready event on a listening socket means there's a connection
683  * pending. Do not use state_change as a substitute for it.
684  */
685 static void
686 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
687 {
688 	struct svc_sock	*svsk = (struct svc_sock *)sk->sk_user_data;
689 
690 	dprintk("svc: socket %p TCP (listen) state change %d\n",
691 		sk, sk->sk_state);
692 
693 	/*
694 	 * This callback may called twice when a new connection
695 	 * is established as a child socket inherits everything
696 	 * from a parent LISTEN socket.
697 	 * 1) data_ready method of the parent socket will be called
698 	 *    when one of child sockets become ESTABLISHED.
699 	 * 2) data_ready method of the child socket may be called
700 	 *    when it receives data before the socket is accepted.
701 	 * In case of 2, we should ignore it silently.
702 	 */
703 	if (sk->sk_state == TCP_LISTEN) {
704 		if (svsk) {
705 			set_bit(SK_CONN, &svsk->sk_flags);
706 			svc_sock_enqueue(svsk);
707 		} else
708 			printk("svc: socket %p: no user data\n", sk);
709 	}
710 
711 	if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
712 		wake_up_interruptible_all(sk->sk_sleep);
713 }
714 
715 /*
716  * A state change on a connected socket means it's dying or dead.
717  */
718 static void
719 svc_tcp_state_change(struct sock *sk)
720 {
721 	struct svc_sock	*svsk = (struct svc_sock *)sk->sk_user_data;
722 
723 	dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
724 		sk, sk->sk_state, sk->sk_user_data);
725 
726 	if (!svsk)
727 		printk("svc: socket %p: no user data\n", sk);
728 	else {
729 		set_bit(SK_CLOSE, &svsk->sk_flags);
730 		svc_sock_enqueue(svsk);
731 	}
732 	if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
733 		wake_up_interruptible_all(sk->sk_sleep);
734 }
735 
736 static void
737 svc_tcp_data_ready(struct sock *sk, int count)
738 {
739 	struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
740 
741 	dprintk("svc: socket %p TCP data ready (svsk %p)\n",
742 		sk, sk->sk_user_data);
743 	if (svsk) {
744 		set_bit(SK_DATA, &svsk->sk_flags);
745 		svc_sock_enqueue(svsk);
746 	}
747 	if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
748 		wake_up_interruptible(sk->sk_sleep);
749 }
750 
751 /*
752  * Accept a TCP connection
753  */
754 static void
755 svc_tcp_accept(struct svc_sock *svsk)
756 {
757 	struct sockaddr_in sin;
758 	struct svc_serv	*serv = svsk->sk_server;
759 	struct socket	*sock = svsk->sk_sock;
760 	struct socket	*newsock;
761 	const struct proto_ops *ops;
762 	struct svc_sock	*newsvsk;
763 	int		err, slen;
764 
765 	dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
766 	if (!sock)
767 		return;
768 
769 	err = sock_create_lite(PF_INET, SOCK_STREAM, IPPROTO_TCP, &newsock);
770 	if (err) {
771 		if (err == -ENOMEM)
772 			printk(KERN_WARNING "%s: no more sockets!\n",
773 			       serv->sv_name);
774 		return;
775 	}
776 
777 	dprintk("svc: tcp_accept %p allocated\n", newsock);
778 	newsock->ops = ops = sock->ops;
779 
780 	clear_bit(SK_CONN, &svsk->sk_flags);
781 	if ((err = ops->accept(sock, newsock, O_NONBLOCK)) < 0) {
782 		if (err != -EAGAIN && net_ratelimit())
783 			printk(KERN_WARNING "%s: accept failed (err %d)!\n",
784 				   serv->sv_name, -err);
785 		goto failed;		/* aborted connection or whatever */
786 	}
787 	set_bit(SK_CONN, &svsk->sk_flags);
788 	svc_sock_enqueue(svsk);
789 
790 	slen = sizeof(sin);
791 	err = ops->getname(newsock, (struct sockaddr *) &sin, &slen, 1);
792 	if (err < 0) {
793 		if (net_ratelimit())
794 			printk(KERN_WARNING "%s: peername failed (err %d)!\n",
795 				   serv->sv_name, -err);
796 		goto failed;		/* aborted connection or whatever */
797 	}
798 
799 	/* Ideally, we would want to reject connections from unauthorized
800 	 * hosts here, but when we get encription, the IP of the host won't
801 	 * tell us anything. For now just warn about unpriv connections.
802 	 */
803 	if (ntohs(sin.sin_port) >= 1024) {
804 		dprintk(KERN_WARNING
805 			"%s: connect from unprivileged port: %u.%u.%u.%u:%d\n",
806 			serv->sv_name,
807 			NIPQUAD(sin.sin_addr.s_addr), ntohs(sin.sin_port));
808 	}
809 
810 	dprintk("%s: connect from %u.%u.%u.%u:%04x\n", serv->sv_name,
811 			NIPQUAD(sin.sin_addr.s_addr), ntohs(sin.sin_port));
812 
813 	/* make sure that a write doesn't block forever when
814 	 * low on memory
815 	 */
816 	newsock->sk->sk_sndtimeo = HZ*30;
817 
818 	if (!(newsvsk = svc_setup_socket(serv, newsock, &err, 0)))
819 		goto failed;
820 
821 
822 	/* make sure that we don't have too many active connections.
823 	 * If we have, something must be dropped.
824 	 *
825 	 * There's no point in trying to do random drop here for
826 	 * DoS prevention. The NFS clients does 1 reconnect in 15
827 	 * seconds. An attacker can easily beat that.
828 	 *
829 	 * The only somewhat efficient mechanism would be if drop
830 	 * old connections from the same IP first. But right now
831 	 * we don't even record the client IP in svc_sock.
832 	 */
833 	if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
834 		struct svc_sock *svsk = NULL;
835 		spin_lock_bh(&serv->sv_lock);
836 		if (!list_empty(&serv->sv_tempsocks)) {
837 			if (net_ratelimit()) {
838 				/* Try to help the admin */
839 				printk(KERN_NOTICE "%s: too many open TCP "
840 					"sockets, consider increasing the "
841 					"number of nfsd threads\n",
842 						   serv->sv_name);
843 				printk(KERN_NOTICE "%s: last TCP connect from "
844 					"%u.%u.%u.%u:%d\n",
845 					serv->sv_name,
846 					NIPQUAD(sin.sin_addr.s_addr),
847 					ntohs(sin.sin_port));
848 			}
849 			/*
850 			 * Always select the oldest socket. It's not fair,
851 			 * but so is life
852 			 */
853 			svsk = list_entry(serv->sv_tempsocks.prev,
854 					  struct svc_sock,
855 					  sk_list);
856 			set_bit(SK_CLOSE, &svsk->sk_flags);
857 			svsk->sk_inuse ++;
858 		}
859 		spin_unlock_bh(&serv->sv_lock);
860 
861 		if (svsk) {
862 			svc_sock_enqueue(svsk);
863 			svc_sock_put(svsk);
864 		}
865 
866 	}
867 
868 	if (serv->sv_stats)
869 		serv->sv_stats->nettcpconn++;
870 
871 	return;
872 
873 failed:
874 	sock_release(newsock);
875 	return;
876 }
877 
878 /*
879  * Receive data from a TCP socket.
880  */
881 static int
882 svc_tcp_recvfrom(struct svc_rqst *rqstp)
883 {
884 	struct svc_sock	*svsk = rqstp->rq_sock;
885 	struct svc_serv	*serv = svsk->sk_server;
886 	int		len;
887 	struct kvec vec[RPCSVC_MAXPAGES];
888 	int pnum, vlen;
889 
890 	dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
891 		svsk, test_bit(SK_DATA, &svsk->sk_flags),
892 		test_bit(SK_CONN, &svsk->sk_flags),
893 		test_bit(SK_CLOSE, &svsk->sk_flags));
894 
895 	if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
896 		svc_sock_received(svsk);
897 		return svc_deferred_recv(rqstp);
898 	}
899 
900 	if (test_bit(SK_CLOSE, &svsk->sk_flags)) {
901 		svc_delete_socket(svsk);
902 		return 0;
903 	}
904 
905 	if (test_bit(SK_CONN, &svsk->sk_flags)) {
906 		svc_tcp_accept(svsk);
907 		svc_sock_received(svsk);
908 		return 0;
909 	}
910 
911 	if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
912 		/* sndbuf needs to have room for one request
913 		 * per thread, otherwise we can stall even when the
914 		 * network isn't a bottleneck.
915 		 * rcvbuf just needs to be able to hold a few requests.
916 		 * Normally they will be removed from the queue
917 		 * as soon a a complete request arrives.
918 		 */
919 		svc_sock_setbufsize(svsk->sk_sock,
920 				    (serv->sv_nrthreads+3) * serv->sv_bufsz,
921 				    3 * serv->sv_bufsz);
922 
923 	clear_bit(SK_DATA, &svsk->sk_flags);
924 
925 	/* Receive data. If we haven't got the record length yet, get
926 	 * the next four bytes. Otherwise try to gobble up as much as
927 	 * possible up to the complete record length.
928 	 */
929 	if (svsk->sk_tcplen < 4) {
930 		unsigned long	want = 4 - svsk->sk_tcplen;
931 		struct kvec	iov;
932 
933 		iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
934 		iov.iov_len  = want;
935 		if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
936 			goto error;
937 		svsk->sk_tcplen += len;
938 
939 		if (len < want) {
940 			dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
941 			        len, want);
942 			svc_sock_received(svsk);
943 			return -EAGAIN; /* record header not complete */
944 		}
945 
946 		svsk->sk_reclen = ntohl(svsk->sk_reclen);
947 		if (!(svsk->sk_reclen & 0x80000000)) {
948 			/* FIXME: technically, a record can be fragmented,
949 			 *  and non-terminal fragments will not have the top
950 			 *  bit set in the fragment length header.
951 			 *  But apparently no known nfs clients send fragmented
952 			 *  records. */
953 			printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx (non-terminal)\n",
954 			       (unsigned long) svsk->sk_reclen);
955 			goto err_delete;
956 		}
957 		svsk->sk_reclen &= 0x7fffffff;
958 		dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
959 		if (svsk->sk_reclen > serv->sv_bufsz) {
960 			printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx (large)\n",
961 			       (unsigned long) svsk->sk_reclen);
962 			goto err_delete;
963 		}
964 	}
965 
966 	/* Check whether enough data is available */
967 	len = svc_recv_available(svsk);
968 	if (len < 0)
969 		goto error;
970 
971 	if (len < svsk->sk_reclen) {
972 		dprintk("svc: incomplete TCP record (%d of %d)\n",
973 			len, svsk->sk_reclen);
974 		svc_sock_received(svsk);
975 		return -EAGAIN;	/* record not complete */
976 	}
977 	len = svsk->sk_reclen;
978 	set_bit(SK_DATA, &svsk->sk_flags);
979 
980 	vec[0] = rqstp->rq_arg.head[0];
981 	vlen = PAGE_SIZE;
982 	pnum = 1;
983 	while (vlen < len) {
984 		vec[pnum].iov_base = page_address(rqstp->rq_argpages[rqstp->rq_argused++]);
985 		vec[pnum].iov_len = PAGE_SIZE;
986 		pnum++;
987 		vlen += PAGE_SIZE;
988 	}
989 
990 	/* Now receive data */
991 	len = svc_recvfrom(rqstp, vec, pnum, len);
992 	if (len < 0)
993 		goto error;
994 
995 	dprintk("svc: TCP complete record (%d bytes)\n", len);
996 	rqstp->rq_arg.len = len;
997 	rqstp->rq_arg.page_base = 0;
998 	if (len <= rqstp->rq_arg.head[0].iov_len) {
999 		rqstp->rq_arg.head[0].iov_len = len;
1000 		rqstp->rq_arg.page_len = 0;
1001 	} else {
1002 		rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1003 	}
1004 
1005 	rqstp->rq_skbuff      = NULL;
1006 	rqstp->rq_prot	      = IPPROTO_TCP;
1007 
1008 	/* Reset TCP read info */
1009 	svsk->sk_reclen = 0;
1010 	svsk->sk_tcplen = 0;
1011 
1012 	svc_sock_received(svsk);
1013 	if (serv->sv_stats)
1014 		serv->sv_stats->nettcpcnt++;
1015 
1016 	return len;
1017 
1018  err_delete:
1019 	svc_delete_socket(svsk);
1020 	return -EAGAIN;
1021 
1022  error:
1023 	if (len == -EAGAIN) {
1024 		dprintk("RPC: TCP recvfrom got EAGAIN\n");
1025 		svc_sock_received(svsk);
1026 	} else {
1027 		printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1028 					svsk->sk_server->sv_name, -len);
1029 		goto err_delete;
1030 	}
1031 
1032 	return len;
1033 }
1034 
1035 /*
1036  * Send out data on TCP socket.
1037  */
1038 static int
1039 svc_tcp_sendto(struct svc_rqst *rqstp)
1040 {
1041 	struct xdr_buf	*xbufp = &rqstp->rq_res;
1042 	int sent;
1043 	u32 reclen;
1044 
1045 	/* Set up the first element of the reply kvec.
1046 	 * Any other kvecs that may be in use have been taken
1047 	 * care of by the server implementation itself.
1048 	 */
1049 	reclen = htonl(0x80000000|((xbufp->len ) - 4));
1050 	memcpy(xbufp->head[0].iov_base, &reclen, 4);
1051 
1052 	if (test_bit(SK_DEAD, &rqstp->rq_sock->sk_flags))
1053 		return -ENOTCONN;
1054 
1055 	sent = svc_sendto(rqstp, &rqstp->rq_res);
1056 	if (sent != xbufp->len) {
1057 		printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1058 		       rqstp->rq_sock->sk_server->sv_name,
1059 		       (sent<0)?"got error":"sent only",
1060 		       sent, xbufp->len);
1061 		svc_delete_socket(rqstp->rq_sock);
1062 		sent = -EAGAIN;
1063 	}
1064 	return sent;
1065 }
1066 
1067 static void
1068 svc_tcp_init(struct svc_sock *svsk)
1069 {
1070 	struct sock	*sk = svsk->sk_sk;
1071 	struct tcp_sock *tp = tcp_sk(sk);
1072 
1073 	svsk->sk_recvfrom = svc_tcp_recvfrom;
1074 	svsk->sk_sendto = svc_tcp_sendto;
1075 
1076 	if (sk->sk_state == TCP_LISTEN) {
1077 		dprintk("setting up TCP socket for listening\n");
1078 		sk->sk_data_ready = svc_tcp_listen_data_ready;
1079 		set_bit(SK_CONN, &svsk->sk_flags);
1080 	} else {
1081 		dprintk("setting up TCP socket for reading\n");
1082 		sk->sk_state_change = svc_tcp_state_change;
1083 		sk->sk_data_ready = svc_tcp_data_ready;
1084 		sk->sk_write_space = svc_write_space;
1085 
1086 		svsk->sk_reclen = 0;
1087 		svsk->sk_tcplen = 0;
1088 
1089 		tp->nonagle = 1;        /* disable Nagle's algorithm */
1090 
1091 		/* initialise setting must have enough space to
1092 		 * receive and respond to one request.
1093 		 * svc_tcp_recvfrom will re-adjust if necessary
1094 		 */
1095 		svc_sock_setbufsize(svsk->sk_sock,
1096 				    3 * svsk->sk_server->sv_bufsz,
1097 				    3 * svsk->sk_server->sv_bufsz);
1098 
1099 		set_bit(SK_CHNGBUF, &svsk->sk_flags);
1100 		set_bit(SK_DATA, &svsk->sk_flags);
1101 		if (sk->sk_state != TCP_ESTABLISHED)
1102 			set_bit(SK_CLOSE, &svsk->sk_flags);
1103 	}
1104 }
1105 
1106 void
1107 svc_sock_update_bufs(struct svc_serv *serv)
1108 {
1109 	/*
1110 	 * The number of server threads has changed. Update
1111 	 * rcvbuf and sndbuf accordingly on all sockets
1112 	 */
1113 	struct list_head *le;
1114 
1115 	spin_lock_bh(&serv->sv_lock);
1116 	list_for_each(le, &serv->sv_permsocks) {
1117 		struct svc_sock *svsk =
1118 			list_entry(le, struct svc_sock, sk_list);
1119 		set_bit(SK_CHNGBUF, &svsk->sk_flags);
1120 	}
1121 	list_for_each(le, &serv->sv_tempsocks) {
1122 		struct svc_sock *svsk =
1123 			list_entry(le, struct svc_sock, sk_list);
1124 		set_bit(SK_CHNGBUF, &svsk->sk_flags);
1125 	}
1126 	spin_unlock_bh(&serv->sv_lock);
1127 }
1128 
1129 /*
1130  * Receive the next request on any socket.
1131  */
1132 int
1133 svc_recv(struct svc_serv *serv, struct svc_rqst *rqstp, long timeout)
1134 {
1135 	struct svc_sock		*svsk =NULL;
1136 	int			len;
1137 	int 			pages;
1138 	struct xdr_buf		*arg;
1139 	DECLARE_WAITQUEUE(wait, current);
1140 
1141 	dprintk("svc: server %p waiting for data (to = %ld)\n",
1142 		rqstp, timeout);
1143 
1144 	if (rqstp->rq_sock)
1145 		printk(KERN_ERR
1146 			"svc_recv: service %p, socket not NULL!\n",
1147 			 rqstp);
1148 	if (waitqueue_active(&rqstp->rq_wait))
1149 		printk(KERN_ERR
1150 			"svc_recv: service %p, wait queue active!\n",
1151 			 rqstp);
1152 
1153 	/* Initialize the buffers */
1154 	/* first reclaim pages that were moved to response list */
1155 	svc_pushback_allpages(rqstp);
1156 
1157 	/* now allocate needed pages.  If we get a failure, sleep briefly */
1158 	pages = 2 + (serv->sv_bufsz + PAGE_SIZE -1) / PAGE_SIZE;
1159 	while (rqstp->rq_arghi < pages) {
1160 		struct page *p = alloc_page(GFP_KERNEL);
1161 		if (!p) {
1162 			schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1163 			continue;
1164 		}
1165 		rqstp->rq_argpages[rqstp->rq_arghi++] = p;
1166 	}
1167 
1168 	/* Make arg->head point to first page and arg->pages point to rest */
1169 	arg = &rqstp->rq_arg;
1170 	arg->head[0].iov_base = page_address(rqstp->rq_argpages[0]);
1171 	arg->head[0].iov_len = PAGE_SIZE;
1172 	rqstp->rq_argused = 1;
1173 	arg->pages = rqstp->rq_argpages + 1;
1174 	arg->page_base = 0;
1175 	/* save at least one page for response */
1176 	arg->page_len = (pages-2)*PAGE_SIZE;
1177 	arg->len = (pages-1)*PAGE_SIZE;
1178 	arg->tail[0].iov_len = 0;
1179 
1180 	try_to_freeze();
1181 	cond_resched();
1182 	if (signalled())
1183 		return -EINTR;
1184 
1185 	spin_lock_bh(&serv->sv_lock);
1186 	if (!list_empty(&serv->sv_tempsocks)) {
1187 		svsk = list_entry(serv->sv_tempsocks.next,
1188 				  struct svc_sock, sk_list);
1189 		/* apparently the "standard" is that clients close
1190 		 * idle connections after 5 minutes, servers after
1191 		 * 6 minutes
1192 		 *   http://www.connectathon.org/talks96/nfstcp.pdf
1193 		 */
1194 		if (get_seconds() - svsk->sk_lastrecv < 6*60
1195 		    || test_bit(SK_BUSY, &svsk->sk_flags))
1196 			svsk = NULL;
1197 	}
1198 	if (svsk) {
1199 		set_bit(SK_BUSY, &svsk->sk_flags);
1200 		set_bit(SK_CLOSE, &svsk->sk_flags);
1201 		rqstp->rq_sock = svsk;
1202 		svsk->sk_inuse++;
1203 	} else if ((svsk = svc_sock_dequeue(serv)) != NULL) {
1204 		rqstp->rq_sock = svsk;
1205 		svsk->sk_inuse++;
1206 		rqstp->rq_reserved = serv->sv_bufsz;
1207 		svsk->sk_reserved += rqstp->rq_reserved;
1208 	} else {
1209 		/* No data pending. Go to sleep */
1210 		svc_serv_enqueue(serv, rqstp);
1211 
1212 		/*
1213 		 * We have to be able to interrupt this wait
1214 		 * to bring down the daemons ...
1215 		 */
1216 		set_current_state(TASK_INTERRUPTIBLE);
1217 		add_wait_queue(&rqstp->rq_wait, &wait);
1218 		spin_unlock_bh(&serv->sv_lock);
1219 
1220 		schedule_timeout(timeout);
1221 
1222 		try_to_freeze();
1223 
1224 		spin_lock_bh(&serv->sv_lock);
1225 		remove_wait_queue(&rqstp->rq_wait, &wait);
1226 
1227 		if (!(svsk = rqstp->rq_sock)) {
1228 			svc_serv_dequeue(serv, rqstp);
1229 			spin_unlock_bh(&serv->sv_lock);
1230 			dprintk("svc: server %p, no data yet\n", rqstp);
1231 			return signalled()? -EINTR : -EAGAIN;
1232 		}
1233 	}
1234 	spin_unlock_bh(&serv->sv_lock);
1235 
1236 	dprintk("svc: server %p, socket %p, inuse=%d\n",
1237 		 rqstp, svsk, svsk->sk_inuse);
1238 	len = svsk->sk_recvfrom(rqstp);
1239 	dprintk("svc: got len=%d\n", len);
1240 
1241 	/* No data, incomplete (TCP) read, or accept() */
1242 	if (len == 0 || len == -EAGAIN) {
1243 		rqstp->rq_res.len = 0;
1244 		svc_sock_release(rqstp);
1245 		return -EAGAIN;
1246 	}
1247 	svsk->sk_lastrecv = get_seconds();
1248 	if (test_bit(SK_TEMP, &svsk->sk_flags)) {
1249 		/* push active sockets to end of list */
1250 		spin_lock_bh(&serv->sv_lock);
1251 		if (!list_empty(&svsk->sk_list))
1252 			list_move_tail(&svsk->sk_list, &serv->sv_tempsocks);
1253 		spin_unlock_bh(&serv->sv_lock);
1254 	}
1255 
1256 	rqstp->rq_secure  = ntohs(rqstp->rq_addr.sin_port) < 1024;
1257 	rqstp->rq_chandle.defer = svc_defer;
1258 
1259 	if (serv->sv_stats)
1260 		serv->sv_stats->netcnt++;
1261 	return len;
1262 }
1263 
1264 /*
1265  * Drop request
1266  */
1267 void
1268 svc_drop(struct svc_rqst *rqstp)
1269 {
1270 	dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1271 	svc_sock_release(rqstp);
1272 }
1273 
1274 /*
1275  * Return reply to client.
1276  */
1277 int
1278 svc_send(struct svc_rqst *rqstp)
1279 {
1280 	struct svc_sock	*svsk;
1281 	int		len;
1282 	struct xdr_buf	*xb;
1283 
1284 	if ((svsk = rqstp->rq_sock) == NULL) {
1285 		printk(KERN_WARNING "NULL socket pointer in %s:%d\n",
1286 				__FILE__, __LINE__);
1287 		return -EFAULT;
1288 	}
1289 
1290 	/* release the receive skb before sending the reply */
1291 	svc_release_skb(rqstp);
1292 
1293 	/* calculate over-all length */
1294 	xb = & rqstp->rq_res;
1295 	xb->len = xb->head[0].iov_len +
1296 		xb->page_len +
1297 		xb->tail[0].iov_len;
1298 
1299 	/* Grab svsk->sk_sem to serialize outgoing data. */
1300 	down(&svsk->sk_sem);
1301 	if (test_bit(SK_DEAD, &svsk->sk_flags))
1302 		len = -ENOTCONN;
1303 	else
1304 		len = svsk->sk_sendto(rqstp);
1305 	up(&svsk->sk_sem);
1306 	svc_sock_release(rqstp);
1307 
1308 	if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1309 		return 0;
1310 	return len;
1311 }
1312 
1313 /*
1314  * Initialize socket for RPC use and create svc_sock struct
1315  * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1316  */
1317 static struct svc_sock *
1318 svc_setup_socket(struct svc_serv *serv, struct socket *sock,
1319 					int *errp, int pmap_register)
1320 {
1321 	struct svc_sock	*svsk;
1322 	struct sock	*inet;
1323 
1324 	dprintk("svc: svc_setup_socket %p\n", sock);
1325 	if (!(svsk = kmalloc(sizeof(*svsk), GFP_KERNEL))) {
1326 		*errp = -ENOMEM;
1327 		return NULL;
1328 	}
1329 	memset(svsk, 0, sizeof(*svsk));
1330 
1331 	inet = sock->sk;
1332 
1333 	/* Register socket with portmapper */
1334 	if (*errp >= 0 && pmap_register)
1335 		*errp = svc_register(serv, inet->sk_protocol,
1336 				     ntohs(inet_sk(inet)->sport));
1337 
1338 	if (*errp < 0) {
1339 		kfree(svsk);
1340 		return NULL;
1341 	}
1342 
1343 	set_bit(SK_BUSY, &svsk->sk_flags);
1344 	inet->sk_user_data = svsk;
1345 	svsk->sk_sock = sock;
1346 	svsk->sk_sk = inet;
1347 	svsk->sk_ostate = inet->sk_state_change;
1348 	svsk->sk_odata = inet->sk_data_ready;
1349 	svsk->sk_owspace = inet->sk_write_space;
1350 	svsk->sk_server = serv;
1351 	svsk->sk_lastrecv = get_seconds();
1352 	INIT_LIST_HEAD(&svsk->sk_deferred);
1353 	INIT_LIST_HEAD(&svsk->sk_ready);
1354 	sema_init(&svsk->sk_sem, 1);
1355 
1356 	/* Initialize the socket */
1357 	if (sock->type == SOCK_DGRAM)
1358 		svc_udp_init(svsk);
1359 	else
1360 		svc_tcp_init(svsk);
1361 
1362 	spin_lock_bh(&serv->sv_lock);
1363 	if (!pmap_register) {
1364 		set_bit(SK_TEMP, &svsk->sk_flags);
1365 		list_add(&svsk->sk_list, &serv->sv_tempsocks);
1366 		serv->sv_tmpcnt++;
1367 	} else {
1368 		clear_bit(SK_TEMP, &svsk->sk_flags);
1369 		list_add(&svsk->sk_list, &serv->sv_permsocks);
1370 	}
1371 	spin_unlock_bh(&serv->sv_lock);
1372 
1373 	dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1374 				svsk, svsk->sk_sk);
1375 
1376 	clear_bit(SK_BUSY, &svsk->sk_flags);
1377 	svc_sock_enqueue(svsk);
1378 	return svsk;
1379 }
1380 
1381 /*
1382  * Create socket for RPC service.
1383  */
1384 static int
1385 svc_create_socket(struct svc_serv *serv, int protocol, struct sockaddr_in *sin)
1386 {
1387 	struct svc_sock	*svsk;
1388 	struct socket	*sock;
1389 	int		error;
1390 	int		type;
1391 
1392 	dprintk("svc: svc_create_socket(%s, %d, %u.%u.%u.%u:%d)\n",
1393 				serv->sv_program->pg_name, protocol,
1394 				NIPQUAD(sin->sin_addr.s_addr),
1395 				ntohs(sin->sin_port));
1396 
1397 	if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1398 		printk(KERN_WARNING "svc: only UDP and TCP "
1399 				"sockets supported\n");
1400 		return -EINVAL;
1401 	}
1402 	type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1403 
1404 	if ((error = sock_create_kern(PF_INET, type, protocol, &sock)) < 0)
1405 		return error;
1406 
1407 	if (sin != NULL) {
1408 		if (type == SOCK_STREAM)
1409 			sock->sk->sk_reuse = 1; /* allow address reuse */
1410 		error = sock->ops->bind(sock, (struct sockaddr *) sin,
1411 						sizeof(*sin));
1412 		if (error < 0)
1413 			goto bummer;
1414 	}
1415 
1416 	if (protocol == IPPROTO_TCP) {
1417 		if ((error = sock->ops->listen(sock, 64)) < 0)
1418 			goto bummer;
1419 	}
1420 
1421 	if ((svsk = svc_setup_socket(serv, sock, &error, 1)) != NULL)
1422 		return 0;
1423 
1424 bummer:
1425 	dprintk("svc: svc_create_socket error = %d\n", -error);
1426 	sock_release(sock);
1427 	return error;
1428 }
1429 
1430 /*
1431  * Remove a dead socket
1432  */
1433 void
1434 svc_delete_socket(struct svc_sock *svsk)
1435 {
1436 	struct svc_serv	*serv;
1437 	struct sock	*sk;
1438 
1439 	dprintk("svc: svc_delete_socket(%p)\n", svsk);
1440 
1441 	serv = svsk->sk_server;
1442 	sk = svsk->sk_sk;
1443 
1444 	sk->sk_state_change = svsk->sk_ostate;
1445 	sk->sk_data_ready = svsk->sk_odata;
1446 	sk->sk_write_space = svsk->sk_owspace;
1447 
1448 	spin_lock_bh(&serv->sv_lock);
1449 
1450 	list_del_init(&svsk->sk_list);
1451 	list_del_init(&svsk->sk_ready);
1452 	if (!test_and_set_bit(SK_DEAD, &svsk->sk_flags))
1453 		if (test_bit(SK_TEMP, &svsk->sk_flags))
1454 			serv->sv_tmpcnt--;
1455 
1456 	if (!svsk->sk_inuse) {
1457 		spin_unlock_bh(&serv->sv_lock);
1458 		sock_release(svsk->sk_sock);
1459 		kfree(svsk);
1460 	} else {
1461 		spin_unlock_bh(&serv->sv_lock);
1462 		dprintk(KERN_NOTICE "svc: server socket destroy delayed\n");
1463 		/* svsk->sk_server = NULL; */
1464 	}
1465 }
1466 
1467 /*
1468  * Make a socket for nfsd and lockd
1469  */
1470 int
1471 svc_makesock(struct svc_serv *serv, int protocol, unsigned short port)
1472 {
1473 	struct sockaddr_in	sin;
1474 
1475 	dprintk("svc: creating socket proto = %d\n", protocol);
1476 	sin.sin_family      = AF_INET;
1477 	sin.sin_addr.s_addr = INADDR_ANY;
1478 	sin.sin_port        = htons(port);
1479 	return svc_create_socket(serv, protocol, &sin);
1480 }
1481 
1482 /*
1483  * Handle defer and revisit of requests
1484  */
1485 
1486 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1487 {
1488 	struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1489 	struct svc_serv *serv = dreq->owner;
1490 	struct svc_sock *svsk;
1491 
1492 	if (too_many) {
1493 		svc_sock_put(dr->svsk);
1494 		kfree(dr);
1495 		return;
1496 	}
1497 	dprintk("revisit queued\n");
1498 	svsk = dr->svsk;
1499 	dr->svsk = NULL;
1500 	spin_lock_bh(&serv->sv_lock);
1501 	list_add(&dr->handle.recent, &svsk->sk_deferred);
1502 	spin_unlock_bh(&serv->sv_lock);
1503 	set_bit(SK_DEFERRED, &svsk->sk_flags);
1504 	svc_sock_enqueue(svsk);
1505 	svc_sock_put(svsk);
1506 }
1507 
1508 static struct cache_deferred_req *
1509 svc_defer(struct cache_req *req)
1510 {
1511 	struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1512 	int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
1513 	struct svc_deferred_req *dr;
1514 
1515 	if (rqstp->rq_arg.page_len)
1516 		return NULL; /* if more than a page, give up FIXME */
1517 	if (rqstp->rq_deferred) {
1518 		dr = rqstp->rq_deferred;
1519 		rqstp->rq_deferred = NULL;
1520 	} else {
1521 		int skip  = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1522 		/* FIXME maybe discard if size too large */
1523 		dr = kmalloc(size, GFP_KERNEL);
1524 		if (dr == NULL)
1525 			return NULL;
1526 
1527 		dr->handle.owner = rqstp->rq_server;
1528 		dr->prot = rqstp->rq_prot;
1529 		dr->addr = rqstp->rq_addr;
1530 		dr->argslen = rqstp->rq_arg.len >> 2;
1531 		memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
1532 	}
1533 	spin_lock_bh(&rqstp->rq_server->sv_lock);
1534 	rqstp->rq_sock->sk_inuse++;
1535 	dr->svsk = rqstp->rq_sock;
1536 	spin_unlock_bh(&rqstp->rq_server->sv_lock);
1537 
1538 	dr->handle.revisit = svc_revisit;
1539 	return &dr->handle;
1540 }
1541 
1542 /*
1543  * recv data from a deferred request into an active one
1544  */
1545 static int svc_deferred_recv(struct svc_rqst *rqstp)
1546 {
1547 	struct svc_deferred_req *dr = rqstp->rq_deferred;
1548 
1549 	rqstp->rq_arg.head[0].iov_base = dr->args;
1550 	rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
1551 	rqstp->rq_arg.page_len = 0;
1552 	rqstp->rq_arg.len = dr->argslen<<2;
1553 	rqstp->rq_prot        = dr->prot;
1554 	rqstp->rq_addr        = dr->addr;
1555 	return dr->argslen<<2;
1556 }
1557 
1558 
1559 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk)
1560 {
1561 	struct svc_deferred_req *dr = NULL;
1562 	struct svc_serv	*serv = svsk->sk_server;
1563 
1564 	if (!test_bit(SK_DEFERRED, &svsk->sk_flags))
1565 		return NULL;
1566 	spin_lock_bh(&serv->sv_lock);
1567 	clear_bit(SK_DEFERRED, &svsk->sk_flags);
1568 	if (!list_empty(&svsk->sk_deferred)) {
1569 		dr = list_entry(svsk->sk_deferred.next,
1570 				struct svc_deferred_req,
1571 				handle.recent);
1572 		list_del_init(&dr->handle.recent);
1573 		set_bit(SK_DEFERRED, &svsk->sk_flags);
1574 	}
1575 	spin_unlock_bh(&serv->sv_lock);
1576 	return dr;
1577 }
1578