2 * linux/net/sunrpc/svcsock.c
4 * These are the RPC server socket internals.
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...
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.
19 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
22 #include <linux/sched.h>
23 #include <linux/errno.h>
24 #include <linux/fcntl.h>
25 #include <linux/net.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>
36 #include <net/checksum.h>
38 #include <net/tcp_states.h>
39 #include <asm/uaccess.h>
40 #include <asm/ioctls.h>
42 #include <linux/sunrpc/types.h>
43 #include <linux/sunrpc/xdr.h>
44 #include <linux/sunrpc/svcsock.h>
45 #include <linux/sunrpc/stats.h>
47 /* SMP locking strategy:
49 * svc_pool->sp_lock protects most of the fields of that pool.
50 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
51 * when both need to be taken (rare), svc_serv->sv_lock is first.
52 * BKL protects svc_serv->sv_nrthread.
53 * svc_sock->sk_defer_lock protects the svc_sock->sk_deferred list
54 * svc_sock->sk_flags.SK_BUSY prevents a svc_sock being enqueued multiply.
56 * Some flags can be set to certain values at any time
57 * providing that certain rules are followed:
59 * SK_CONN, SK_DATA, can be set or cleared at any time.
60 * after a set, svc_sock_enqueue must be called.
61 * after a clear, the socket must be read/accepted
62 * if this succeeds, it must be set again.
63 * SK_CLOSE can set at any time. It is never cleared.
67 #define RPCDBG_FACILITY RPCDBG_SVCSOCK
70 static struct svc_sock *svc_setup_socket(struct svc_serv *, struct socket *,
71 int *errp, int pmap_reg);
72 static void svc_udp_data_ready(struct sock *, int);
73 static int svc_udp_recvfrom(struct svc_rqst *);
74 static int svc_udp_sendto(struct svc_rqst *);
76 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk);
77 static int svc_deferred_recv(struct svc_rqst *rqstp);
78 static struct cache_deferred_req *svc_defer(struct cache_req *req);
80 /* apparently the "standard" is that clients close
81 * idle connections after 5 minutes, servers after
83 * http://www.connectathon.org/talks96/nfstcp.pdf
85 static int svc_conn_age_period = 6*60;
88 * Queue up an idle server thread. Must have pool->sp_lock held.
89 * Note: this is really a stack rather than a queue, so that we only
90 * use as many different threads as we need, and the rest don't pollute
94 svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
96 list_add(&rqstp->rq_list, &pool->sp_threads);
100 * Dequeue an nfsd thread. Must have pool->sp_lock held.
103 svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
105 list_del(&rqstp->rq_list);
109 * Release an skbuff after use
112 svc_release_skb(struct svc_rqst *rqstp)
114 struct sk_buff *skb = rqstp->rq_skbuff;
115 struct svc_deferred_req *dr = rqstp->rq_deferred;
118 rqstp->rq_skbuff = NULL;
120 dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
121 skb_free_datagram(rqstp->rq_sock->sk_sk, skb);
124 rqstp->rq_deferred = NULL;
130 * Any space to write?
132 static inline unsigned long
133 svc_sock_wspace(struct svc_sock *svsk)
137 if (svsk->sk_sock->type == SOCK_STREAM)
138 wspace = sk_stream_wspace(svsk->sk_sk);
140 wspace = sock_wspace(svsk->sk_sk);
146 * Queue up a socket with data pending. If there are idle nfsd
147 * processes, wake 'em up.
151 svc_sock_enqueue(struct svc_sock *svsk)
153 struct svc_serv *serv = svsk->sk_server;
154 struct svc_pool *pool;
155 struct svc_rqst *rqstp;
158 if (!(svsk->sk_flags &
159 ( (1<<SK_CONN)|(1<<SK_DATA)|(1<<SK_CLOSE)|(1<<SK_DEFERRED)) ))
161 if (test_bit(SK_DEAD, &svsk->sk_flags))
165 pool = svc_pool_for_cpu(svsk->sk_server, cpu);
168 spin_lock_bh(&pool->sp_lock);
170 if (!list_empty(&pool->sp_threads) &&
171 !list_empty(&pool->sp_sockets))
173 "svc_sock_enqueue: threads and sockets both waiting??\n");
175 if (test_bit(SK_DEAD, &svsk->sk_flags)) {
176 /* Don't enqueue dead sockets */
177 dprintk("svc: socket %p is dead, not enqueued\n", svsk->sk_sk);
181 /* Mark socket as busy. It will remain in this state until the
182 * server has processed all pending data and put the socket back
183 * on the idle list. We update SK_BUSY atomically because
184 * it also guards against trying to enqueue the svc_sock twice.
186 if (test_and_set_bit(SK_BUSY, &svsk->sk_flags)) {
187 /* Don't enqueue socket while already enqueued */
188 dprintk("svc: socket %p busy, not enqueued\n", svsk->sk_sk);
191 BUG_ON(svsk->sk_pool != NULL);
192 svsk->sk_pool = pool;
194 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
195 if (((atomic_read(&svsk->sk_reserved) + serv->sv_max_mesg)*2
196 > svc_sock_wspace(svsk))
197 && !test_bit(SK_CLOSE, &svsk->sk_flags)
198 && !test_bit(SK_CONN, &svsk->sk_flags)) {
199 /* Don't enqueue while not enough space for reply */
200 dprintk("svc: socket %p no space, %d*2 > %ld, not enqueued\n",
201 svsk->sk_sk, atomic_read(&svsk->sk_reserved)+serv->sv_max_mesg,
202 svc_sock_wspace(svsk));
203 svsk->sk_pool = NULL;
204 clear_bit(SK_BUSY, &svsk->sk_flags);
207 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
210 if (!list_empty(&pool->sp_threads)) {
211 rqstp = list_entry(pool->sp_threads.next,
214 dprintk("svc: socket %p served by daemon %p\n",
216 svc_thread_dequeue(pool, rqstp);
219 "svc_sock_enqueue: server %p, rq_sock=%p!\n",
220 rqstp, rqstp->rq_sock);
221 rqstp->rq_sock = svsk;
222 atomic_inc(&svsk->sk_inuse);
223 rqstp->rq_reserved = serv->sv_max_mesg;
224 atomic_add(rqstp->rq_reserved, &svsk->sk_reserved);
225 BUG_ON(svsk->sk_pool != pool);
226 wake_up(&rqstp->rq_wait);
228 dprintk("svc: socket %p put into queue\n", svsk->sk_sk);
229 list_add_tail(&svsk->sk_ready, &pool->sp_sockets);
230 BUG_ON(svsk->sk_pool != pool);
234 spin_unlock_bh(&pool->sp_lock);
238 * Dequeue the first socket. Must be called with the pool->sp_lock held.
240 static inline struct svc_sock *
241 svc_sock_dequeue(struct svc_pool *pool)
243 struct svc_sock *svsk;
245 if (list_empty(&pool->sp_sockets))
248 svsk = list_entry(pool->sp_sockets.next,
249 struct svc_sock, sk_ready);
250 list_del_init(&svsk->sk_ready);
252 dprintk("svc: socket %p dequeued, inuse=%d\n",
253 svsk->sk_sk, atomic_read(&svsk->sk_inuse));
259 * Having read something from a socket, check whether it
260 * needs to be re-enqueued.
261 * Note: SK_DATA only gets cleared when a read-attempt finds
262 * no (or insufficient) data.
265 svc_sock_received(struct svc_sock *svsk)
267 svsk->sk_pool = NULL;
268 clear_bit(SK_BUSY, &svsk->sk_flags);
269 svc_sock_enqueue(svsk);
274 * svc_reserve - change the space reserved for the reply to a request.
275 * @rqstp: The request in question
276 * @space: new max space to reserve
278 * Each request reserves some space on the output queue of the socket
279 * to make sure the reply fits. This function reduces that reserved
280 * space to be the amount of space used already, plus @space.
283 void svc_reserve(struct svc_rqst *rqstp, int space)
285 space += rqstp->rq_res.head[0].iov_len;
287 if (space < rqstp->rq_reserved) {
288 struct svc_sock *svsk = rqstp->rq_sock;
289 atomic_sub((rqstp->rq_reserved - space), &svsk->sk_reserved);
290 rqstp->rq_reserved = space;
292 svc_sock_enqueue(svsk);
297 * Release a socket after use.
300 svc_sock_put(struct svc_sock *svsk)
302 if (atomic_dec_and_test(&svsk->sk_inuse) && test_bit(SK_DEAD, &svsk->sk_flags)) {
303 printk("svc: releasing dead socket\n");
304 if (svsk->sk_sock->file)
305 sockfd_put(svsk->sk_sock);
307 sock_release(svsk->sk_sock);
308 if (svsk->sk_info_authunix != NULL)
309 svcauth_unix_info_release(svsk->sk_info_authunix);
315 svc_sock_release(struct svc_rqst *rqstp)
317 struct svc_sock *svsk = rqstp->rq_sock;
319 svc_release_skb(rqstp);
321 svc_free_res_pages(rqstp);
322 rqstp->rq_res.page_len = 0;
323 rqstp->rq_res.page_base = 0;
326 /* Reset response buffer and release
328 * But first, check that enough space was reserved
329 * for the reply, otherwise we have a bug!
331 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
332 printk(KERN_ERR "RPC request reserved %d but used %d\n",
336 rqstp->rq_res.head[0].iov_len = 0;
337 svc_reserve(rqstp, 0);
338 rqstp->rq_sock = NULL;
344 * External function to wake up a server waiting for data
345 * This really only makes sense for services like lockd
346 * which have exactly one thread anyway.
349 svc_wake_up(struct svc_serv *serv)
351 struct svc_rqst *rqstp;
353 struct svc_pool *pool;
355 for (i = 0; i < serv->sv_nrpools; i++) {
356 pool = &serv->sv_pools[i];
358 spin_lock_bh(&pool->sp_lock);
359 if (!list_empty(&pool->sp_threads)) {
360 rqstp = list_entry(pool->sp_threads.next,
363 dprintk("svc: daemon %p woken up.\n", rqstp);
365 svc_thread_dequeue(pool, rqstp);
366 rqstp->rq_sock = NULL;
368 wake_up(&rqstp->rq_wait);
370 spin_unlock_bh(&pool->sp_lock);
375 * Generic sendto routine
378 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
380 struct svc_sock *svsk = rqstp->rq_sock;
381 struct socket *sock = svsk->sk_sock;
383 char buffer[CMSG_SPACE(sizeof(struct in_pktinfo))];
384 struct cmsghdr *cmh = (struct cmsghdr *)buffer;
385 struct in_pktinfo *pki = (struct in_pktinfo *)CMSG_DATA(cmh);
389 struct page **ppage = xdr->pages;
390 size_t base = xdr->page_base;
391 unsigned int pglen = xdr->page_len;
392 unsigned int flags = MSG_MORE;
396 if (rqstp->rq_prot == IPPROTO_UDP) {
397 /* set the source and destination */
399 msg.msg_name = &rqstp->rq_addr;
400 msg.msg_namelen = sizeof(rqstp->rq_addr);
403 msg.msg_flags = MSG_MORE;
405 msg.msg_control = cmh;
406 msg.msg_controllen = sizeof(buffer);
407 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
408 cmh->cmsg_level = SOL_IP;
409 cmh->cmsg_type = IP_PKTINFO;
410 pki->ipi_ifindex = 0;
411 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr;
413 if (sock_sendmsg(sock, &msg, 0) < 0)
418 if (slen == xdr->head[0].iov_len)
420 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0,
421 xdr->head[0].iov_len, flags);
422 if (len != xdr->head[0].iov_len)
424 slen -= xdr->head[0].iov_len;
429 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
433 result = kernel_sendpage(sock, *ppage, base, size, flags);
440 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
445 if (xdr->tail[0].iov_len) {
446 result = kernel_sendpage(sock, rqstp->rq_respages[0],
447 ((unsigned long)xdr->tail[0].iov_base)
449 xdr->tail[0].iov_len, 0);
455 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %x)\n",
456 rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len, xdr->len, len,
457 rqstp->rq_addr.sin_addr.s_addr);
463 * Report socket names for nfsdfs
465 static int one_sock_name(char *buf, struct svc_sock *svsk)
469 switch(svsk->sk_sk->sk_family) {
471 len = sprintf(buf, "ipv4 %s %u.%u.%u.%u %d\n",
472 svsk->sk_sk->sk_protocol==IPPROTO_UDP?
474 NIPQUAD(inet_sk(svsk->sk_sk)->rcv_saddr),
475 inet_sk(svsk->sk_sk)->num);
478 len = sprintf(buf, "*unknown-%d*\n",
479 svsk->sk_sk->sk_family);
485 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
487 struct svc_sock *svsk, *closesk = NULL;
492 spin_lock(&serv->sv_lock);
493 list_for_each_entry(svsk, &serv->sv_permsocks, sk_list) {
494 int onelen = one_sock_name(buf+len, svsk);
495 if (toclose && strcmp(toclose, buf+len) == 0)
500 spin_unlock(&serv->sv_lock);
502 /* Should unregister with portmap, but you cannot
503 * unregister just one protocol...
505 svc_delete_socket(closesk);
510 EXPORT_SYMBOL(svc_sock_names);
513 * Check input queue length
516 svc_recv_available(struct svc_sock *svsk)
518 struct socket *sock = svsk->sk_sock;
521 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
523 return (err >= 0)? avail : err;
527 * Generic recvfrom routine.
530 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
536 rqstp->rq_addrlen = sizeof(rqstp->rq_addr);
537 sock = rqstp->rq_sock->sk_sock;
539 msg.msg_name = &rqstp->rq_addr;
540 msg.msg_namelen = sizeof(rqstp->rq_addr);
541 msg.msg_control = NULL;
542 msg.msg_controllen = 0;
544 msg.msg_flags = MSG_DONTWAIT;
546 len = kernel_recvmsg(sock, &msg, iov, nr, buflen, MSG_DONTWAIT);
548 /* sock_recvmsg doesn't fill in the name/namelen, so we must..
549 * possibly we should cache this in the svc_sock structure
550 * at accept time. FIXME
552 alen = sizeof(rqstp->rq_addr);
553 kernel_getpeername(sock, (struct sockaddr *)&rqstp->rq_addr, &alen);
555 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
556 rqstp->rq_sock, iov[0].iov_base, iov[0].iov_len, len);
562 * Set socket snd and rcv buffer lengths
565 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
569 oldfs = get_fs(); set_fs(KERNEL_DS);
570 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
571 (char*)&snd, sizeof(snd));
572 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
573 (char*)&rcv, sizeof(rcv));
575 /* sock_setsockopt limits use to sysctl_?mem_max,
576 * which isn't acceptable. Until that is made conditional
577 * on not having CAP_SYS_RESOURCE or similar, we go direct...
578 * DaveM said I could!
581 sock->sk->sk_sndbuf = snd * 2;
582 sock->sk->sk_rcvbuf = rcv * 2;
583 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
584 release_sock(sock->sk);
588 * INET callback when data has been received on the socket.
591 svc_udp_data_ready(struct sock *sk, int count)
593 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
596 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
597 svsk, sk, count, test_bit(SK_BUSY, &svsk->sk_flags));
598 set_bit(SK_DATA, &svsk->sk_flags);
599 svc_sock_enqueue(svsk);
601 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
602 wake_up_interruptible(sk->sk_sleep);
606 * INET callback when space is newly available on the socket.
609 svc_write_space(struct sock *sk)
611 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
614 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
615 svsk, sk, test_bit(SK_BUSY, &svsk->sk_flags));
616 svc_sock_enqueue(svsk);
619 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
620 dprintk("RPC svc_write_space: someone sleeping on %p\n",
622 wake_up_interruptible(sk->sk_sleep);
627 * Receive a datagram from a UDP socket.
630 svc_udp_recvfrom(struct svc_rqst *rqstp)
632 struct svc_sock *svsk = rqstp->rq_sock;
633 struct svc_serv *serv = svsk->sk_server;
637 if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
638 /* udp sockets need large rcvbuf as all pending
639 * requests are still in that buffer. sndbuf must
640 * also be large enough that there is enough space
641 * for one reply per thread. We count all threads
642 * rather than threads in a particular pool, which
643 * provides an upper bound on the number of threads
644 * which will access the socket.
646 svc_sock_setbufsize(svsk->sk_sock,
647 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
648 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
650 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
651 svc_sock_received(svsk);
652 return svc_deferred_recv(rqstp);
655 clear_bit(SK_DATA, &svsk->sk_flags);
656 while ((skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err)) == NULL) {
657 if (err == -EAGAIN) {
658 svc_sock_received(svsk);
661 /* possibly an icmp error */
662 dprintk("svc: recvfrom returned error %d\n", -err);
664 if (skb->tstamp.off_sec == 0) {
667 tv.tv_sec = xtime.tv_sec;
668 tv.tv_usec = xtime.tv_nsec / NSEC_PER_USEC;
669 skb_set_timestamp(skb, &tv);
670 /* Don't enable netstamp, sunrpc doesn't
671 need that much accuracy */
673 skb_get_timestamp(skb, &svsk->sk_sk->sk_stamp);
674 set_bit(SK_DATA, &svsk->sk_flags); /* there may be more data... */
677 * Maybe more packets - kick another thread ASAP.
679 svc_sock_received(svsk);
681 len = skb->len - sizeof(struct udphdr);
682 rqstp->rq_arg.len = len;
684 rqstp->rq_prot = IPPROTO_UDP;
686 /* Get sender address */
687 rqstp->rq_addr.sin_family = AF_INET;
688 rqstp->rq_addr.sin_port = skb->h.uh->source;
689 rqstp->rq_addr.sin_addr.s_addr = skb->nh.iph->saddr;
690 rqstp->rq_daddr = skb->nh.iph->daddr;
692 if (skb_is_nonlinear(skb)) {
693 /* we have to copy */
695 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
698 skb_free_datagram(svsk->sk_sk, skb);
702 skb_free_datagram(svsk->sk_sk, skb);
704 /* we can use it in-place */
705 rqstp->rq_arg.head[0].iov_base = skb->data + sizeof(struct udphdr);
706 rqstp->rq_arg.head[0].iov_len = len;
707 if (skb_checksum_complete(skb)) {
708 skb_free_datagram(svsk->sk_sk, skb);
711 rqstp->rq_skbuff = skb;
714 rqstp->rq_arg.page_base = 0;
715 if (len <= rqstp->rq_arg.head[0].iov_len) {
716 rqstp->rq_arg.head[0].iov_len = len;
717 rqstp->rq_arg.page_len = 0;
718 rqstp->rq_respages = rqstp->rq_pages+1;
720 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
721 rqstp->rq_respages = rqstp->rq_pages + 1 +
722 (rqstp->rq_arg.page_len + PAGE_SIZE - 1)/ PAGE_SIZE;
726 serv->sv_stats->netudpcnt++;
732 svc_udp_sendto(struct svc_rqst *rqstp)
736 error = svc_sendto(rqstp, &rqstp->rq_res);
737 if (error == -ECONNREFUSED)
738 /* ICMP error on earlier request. */
739 error = svc_sendto(rqstp, &rqstp->rq_res);
745 svc_udp_init(struct svc_sock *svsk)
747 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
748 svsk->sk_sk->sk_write_space = svc_write_space;
749 svsk->sk_recvfrom = svc_udp_recvfrom;
750 svsk->sk_sendto = svc_udp_sendto;
752 /* initialise setting must have enough space to
753 * receive and respond to one request.
754 * svc_udp_recvfrom will re-adjust if necessary
756 svc_sock_setbufsize(svsk->sk_sock,
757 3 * svsk->sk_server->sv_max_mesg,
758 3 * svsk->sk_server->sv_max_mesg);
760 set_bit(SK_DATA, &svsk->sk_flags); /* might have come in before data_ready set up */
761 set_bit(SK_CHNGBUF, &svsk->sk_flags);
765 * A data_ready event on a listening socket means there's a connection
766 * pending. Do not use state_change as a substitute for it.
769 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
771 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
773 dprintk("svc: socket %p TCP (listen) state change %d\n",
777 * This callback may called twice when a new connection
778 * is established as a child socket inherits everything
779 * from a parent LISTEN socket.
780 * 1) data_ready method of the parent socket will be called
781 * when one of child sockets become ESTABLISHED.
782 * 2) data_ready method of the child socket may be called
783 * when it receives data before the socket is accepted.
784 * In case of 2, we should ignore it silently.
786 if (sk->sk_state == TCP_LISTEN) {
788 set_bit(SK_CONN, &svsk->sk_flags);
789 svc_sock_enqueue(svsk);
791 printk("svc: socket %p: no user data\n", sk);
794 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
795 wake_up_interruptible_all(sk->sk_sleep);
799 * A state change on a connected socket means it's dying or dead.
802 svc_tcp_state_change(struct sock *sk)
804 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
806 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
807 sk, sk->sk_state, sk->sk_user_data);
810 printk("svc: socket %p: no user data\n", sk);
812 set_bit(SK_CLOSE, &svsk->sk_flags);
813 svc_sock_enqueue(svsk);
815 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
816 wake_up_interruptible_all(sk->sk_sleep);
820 svc_tcp_data_ready(struct sock *sk, int count)
822 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
824 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
825 sk, sk->sk_user_data);
827 set_bit(SK_DATA, &svsk->sk_flags);
828 svc_sock_enqueue(svsk);
830 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
831 wake_up_interruptible(sk->sk_sleep);
835 * Accept a TCP connection
838 svc_tcp_accept(struct svc_sock *svsk)
840 struct sockaddr_in sin;
841 struct svc_serv *serv = svsk->sk_server;
842 struct socket *sock = svsk->sk_sock;
843 struct socket *newsock;
844 struct svc_sock *newsvsk;
847 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
851 clear_bit(SK_CONN, &svsk->sk_flags);
852 err = kernel_accept(sock, &newsock, O_NONBLOCK);
855 printk(KERN_WARNING "%s: no more sockets!\n",
857 else if (err != -EAGAIN && net_ratelimit())
858 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
859 serv->sv_name, -err);
863 set_bit(SK_CONN, &svsk->sk_flags);
864 svc_sock_enqueue(svsk);
867 err = kernel_getpeername(newsock, (struct sockaddr *) &sin, &slen);
870 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
871 serv->sv_name, -err);
872 goto failed; /* aborted connection or whatever */
875 /* Ideally, we would want to reject connections from unauthorized
876 * hosts here, but when we get encription, the IP of the host won't
877 * tell us anything. For now just warn about unpriv connections.
879 if (ntohs(sin.sin_port) >= 1024) {
881 "%s: connect from unprivileged port: %u.%u.%u.%u:%d\n",
883 NIPQUAD(sin.sin_addr.s_addr), ntohs(sin.sin_port));
886 dprintk("%s: connect from %u.%u.%u.%u:%04x\n", serv->sv_name,
887 NIPQUAD(sin.sin_addr.s_addr), ntohs(sin.sin_port));
889 /* make sure that a write doesn't block forever when
892 newsock->sk->sk_sndtimeo = HZ*30;
894 if (!(newsvsk = svc_setup_socket(serv, newsock, &err, 0)))
898 /* make sure that we don't have too many active connections.
899 * If we have, something must be dropped.
901 * There's no point in trying to do random drop here for
902 * DoS prevention. The NFS clients does 1 reconnect in 15
903 * seconds. An attacker can easily beat that.
905 * The only somewhat efficient mechanism would be if drop
906 * old connections from the same IP first. But right now
907 * we don't even record the client IP in svc_sock.
909 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
910 struct svc_sock *svsk = NULL;
911 spin_lock_bh(&serv->sv_lock);
912 if (!list_empty(&serv->sv_tempsocks)) {
913 if (net_ratelimit()) {
914 /* Try to help the admin */
915 printk(KERN_NOTICE "%s: too many open TCP "
916 "sockets, consider increasing the "
917 "number of nfsd threads\n",
919 printk(KERN_NOTICE "%s: last TCP connect from "
922 NIPQUAD(sin.sin_addr.s_addr),
923 ntohs(sin.sin_port));
926 * Always select the oldest socket. It's not fair,
929 svsk = list_entry(serv->sv_tempsocks.prev,
932 set_bit(SK_CLOSE, &svsk->sk_flags);
933 atomic_inc(&svsk->sk_inuse);
935 spin_unlock_bh(&serv->sv_lock);
938 svc_sock_enqueue(svsk);
945 serv->sv_stats->nettcpconn++;
950 sock_release(newsock);
955 * Receive data from a TCP socket.
958 svc_tcp_recvfrom(struct svc_rqst *rqstp)
960 struct svc_sock *svsk = rqstp->rq_sock;
961 struct svc_serv *serv = svsk->sk_server;
966 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
967 svsk, test_bit(SK_DATA, &svsk->sk_flags),
968 test_bit(SK_CONN, &svsk->sk_flags),
969 test_bit(SK_CLOSE, &svsk->sk_flags));
971 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
972 svc_sock_received(svsk);
973 return svc_deferred_recv(rqstp);
976 if (test_bit(SK_CLOSE, &svsk->sk_flags)) {
977 svc_delete_socket(svsk);
981 if (svsk->sk_sk->sk_state == TCP_LISTEN) {
982 svc_tcp_accept(svsk);
983 svc_sock_received(svsk);
987 if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
988 /* sndbuf needs to have room for one request
989 * per thread, otherwise we can stall even when the
990 * network isn't a bottleneck.
992 * We count all threads rather than threads in a
993 * particular pool, which provides an upper bound
994 * on the number of threads which will access the socket.
996 * rcvbuf just needs to be able to hold a few requests.
997 * Normally they will be removed from the queue
998 * as soon a a complete request arrives.
1000 svc_sock_setbufsize(svsk->sk_sock,
1001 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
1002 3 * serv->sv_max_mesg);
1004 clear_bit(SK_DATA, &svsk->sk_flags);
1006 /* Receive data. If we haven't got the record length yet, get
1007 * the next four bytes. Otherwise try to gobble up as much as
1008 * possible up to the complete record length.
1010 if (svsk->sk_tcplen < 4) {
1011 unsigned long want = 4 - svsk->sk_tcplen;
1014 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
1016 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
1018 svsk->sk_tcplen += len;
1021 dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
1023 svc_sock_received(svsk);
1024 return -EAGAIN; /* record header not complete */
1027 svsk->sk_reclen = ntohl(svsk->sk_reclen);
1028 if (!(svsk->sk_reclen & 0x80000000)) {
1029 /* FIXME: technically, a record can be fragmented,
1030 * and non-terminal fragments will not have the top
1031 * bit set in the fragment length header.
1032 * But apparently no known nfs clients send fragmented
1034 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx (non-terminal)\n",
1035 (unsigned long) svsk->sk_reclen);
1038 svsk->sk_reclen &= 0x7fffffff;
1039 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
1040 if (svsk->sk_reclen > serv->sv_max_mesg) {
1041 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx (large)\n",
1042 (unsigned long) svsk->sk_reclen);
1047 /* Check whether enough data is available */
1048 len = svc_recv_available(svsk);
1052 if (len < svsk->sk_reclen) {
1053 dprintk("svc: incomplete TCP record (%d of %d)\n",
1054 len, svsk->sk_reclen);
1055 svc_sock_received(svsk);
1056 return -EAGAIN; /* record not complete */
1058 len = svsk->sk_reclen;
1059 set_bit(SK_DATA, &svsk->sk_flags);
1061 vec = rqstp->rq_vec;
1062 vec[0] = rqstp->rq_arg.head[0];
1065 while (vlen < len) {
1066 vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
1067 vec[pnum].iov_len = PAGE_SIZE;
1071 rqstp->rq_respages = &rqstp->rq_pages[pnum];
1073 /* Now receive data */
1074 len = svc_recvfrom(rqstp, vec, pnum, len);
1078 dprintk("svc: TCP complete record (%d bytes)\n", len);
1079 rqstp->rq_arg.len = len;
1080 rqstp->rq_arg.page_base = 0;
1081 if (len <= rqstp->rq_arg.head[0].iov_len) {
1082 rqstp->rq_arg.head[0].iov_len = len;
1083 rqstp->rq_arg.page_len = 0;
1085 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1088 rqstp->rq_skbuff = NULL;
1089 rqstp->rq_prot = IPPROTO_TCP;
1091 /* Reset TCP read info */
1092 svsk->sk_reclen = 0;
1093 svsk->sk_tcplen = 0;
1095 svc_sock_received(svsk);
1097 serv->sv_stats->nettcpcnt++;
1102 svc_delete_socket(svsk);
1106 if (len == -EAGAIN) {
1107 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1108 svc_sock_received(svsk);
1110 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1111 svsk->sk_server->sv_name, -len);
1119 * Send out data on TCP socket.
1122 svc_tcp_sendto(struct svc_rqst *rqstp)
1124 struct xdr_buf *xbufp = &rqstp->rq_res;
1128 /* Set up the first element of the reply kvec.
1129 * Any other kvecs that may be in use have been taken
1130 * care of by the server implementation itself.
1132 reclen = htonl(0x80000000|((xbufp->len ) - 4));
1133 memcpy(xbufp->head[0].iov_base, &reclen, 4);
1135 if (test_bit(SK_DEAD, &rqstp->rq_sock->sk_flags))
1138 sent = svc_sendto(rqstp, &rqstp->rq_res);
1139 if (sent != xbufp->len) {
1140 printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1141 rqstp->rq_sock->sk_server->sv_name,
1142 (sent<0)?"got error":"sent only",
1144 svc_delete_socket(rqstp->rq_sock);
1151 svc_tcp_init(struct svc_sock *svsk)
1153 struct sock *sk = svsk->sk_sk;
1154 struct tcp_sock *tp = tcp_sk(sk);
1156 svsk->sk_recvfrom = svc_tcp_recvfrom;
1157 svsk->sk_sendto = svc_tcp_sendto;
1159 if (sk->sk_state == TCP_LISTEN) {
1160 dprintk("setting up TCP socket for listening\n");
1161 sk->sk_data_ready = svc_tcp_listen_data_ready;
1162 set_bit(SK_CONN, &svsk->sk_flags);
1164 dprintk("setting up TCP socket for reading\n");
1165 sk->sk_state_change = svc_tcp_state_change;
1166 sk->sk_data_ready = svc_tcp_data_ready;
1167 sk->sk_write_space = svc_write_space;
1169 svsk->sk_reclen = 0;
1170 svsk->sk_tcplen = 0;
1172 tp->nonagle = 1; /* disable Nagle's algorithm */
1174 /* initialise setting must have enough space to
1175 * receive and respond to one request.
1176 * svc_tcp_recvfrom will re-adjust if necessary
1178 svc_sock_setbufsize(svsk->sk_sock,
1179 3 * svsk->sk_server->sv_max_mesg,
1180 3 * svsk->sk_server->sv_max_mesg);
1182 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1183 set_bit(SK_DATA, &svsk->sk_flags);
1184 if (sk->sk_state != TCP_ESTABLISHED)
1185 set_bit(SK_CLOSE, &svsk->sk_flags);
1190 svc_sock_update_bufs(struct svc_serv *serv)
1193 * The number of server threads has changed. Update
1194 * rcvbuf and sndbuf accordingly on all sockets
1196 struct list_head *le;
1198 spin_lock_bh(&serv->sv_lock);
1199 list_for_each(le, &serv->sv_permsocks) {
1200 struct svc_sock *svsk =
1201 list_entry(le, struct svc_sock, sk_list);
1202 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1204 list_for_each(le, &serv->sv_tempsocks) {
1205 struct svc_sock *svsk =
1206 list_entry(le, struct svc_sock, sk_list);
1207 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1209 spin_unlock_bh(&serv->sv_lock);
1213 * Receive the next request on any socket. This code is carefully
1214 * organised not to touch any cachelines in the shared svc_serv
1215 * structure, only cachelines in the local svc_pool.
1218 svc_recv(struct svc_rqst *rqstp, long timeout)
1220 struct svc_sock *svsk =NULL;
1221 struct svc_serv *serv = rqstp->rq_server;
1222 struct svc_pool *pool = rqstp->rq_pool;
1225 struct xdr_buf *arg;
1226 DECLARE_WAITQUEUE(wait, current);
1228 dprintk("svc: server %p waiting for data (to = %ld)\n",
1233 "svc_recv: service %p, socket not NULL!\n",
1235 if (waitqueue_active(&rqstp->rq_wait))
1237 "svc_recv: service %p, wait queue active!\n",
1241 /* now allocate needed pages. If we get a failure, sleep briefly */
1242 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
1243 for (i=0; i < pages ; i++)
1244 while (rqstp->rq_pages[i] == NULL) {
1245 struct page *p = alloc_page(GFP_KERNEL);
1247 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1248 rqstp->rq_pages[i] = p;
1251 /* Make arg->head point to first page and arg->pages point to rest */
1252 arg = &rqstp->rq_arg;
1253 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
1254 arg->head[0].iov_len = PAGE_SIZE;
1255 arg->pages = rqstp->rq_pages + 1;
1257 /* save at least one page for response */
1258 arg->page_len = (pages-2)*PAGE_SIZE;
1259 arg->len = (pages-1)*PAGE_SIZE;
1260 arg->tail[0].iov_len = 0;
1267 spin_lock_bh(&pool->sp_lock);
1268 if ((svsk = svc_sock_dequeue(pool)) != NULL) {
1269 rqstp->rq_sock = svsk;
1270 atomic_inc(&svsk->sk_inuse);
1271 rqstp->rq_reserved = serv->sv_max_mesg;
1272 atomic_add(rqstp->rq_reserved, &svsk->sk_reserved);
1274 /* No data pending. Go to sleep */
1275 svc_thread_enqueue(pool, rqstp);
1278 * We have to be able to interrupt this wait
1279 * to bring down the daemons ...
1281 set_current_state(TASK_INTERRUPTIBLE);
1282 add_wait_queue(&rqstp->rq_wait, &wait);
1283 spin_unlock_bh(&pool->sp_lock);
1285 schedule_timeout(timeout);
1289 spin_lock_bh(&pool->sp_lock);
1290 remove_wait_queue(&rqstp->rq_wait, &wait);
1292 if (!(svsk = rqstp->rq_sock)) {
1293 svc_thread_dequeue(pool, rqstp);
1294 spin_unlock_bh(&pool->sp_lock);
1295 dprintk("svc: server %p, no data yet\n", rqstp);
1296 return signalled()? -EINTR : -EAGAIN;
1299 spin_unlock_bh(&pool->sp_lock);
1301 dprintk("svc: server %p, pool %u, socket %p, inuse=%d\n",
1302 rqstp, pool->sp_id, svsk, atomic_read(&svsk->sk_inuse));
1303 len = svsk->sk_recvfrom(rqstp);
1304 dprintk("svc: got len=%d\n", len);
1306 /* No data, incomplete (TCP) read, or accept() */
1307 if (len == 0 || len == -EAGAIN) {
1308 rqstp->rq_res.len = 0;
1309 svc_sock_release(rqstp);
1312 svsk->sk_lastrecv = get_seconds();
1313 clear_bit(SK_OLD, &svsk->sk_flags);
1315 rqstp->rq_secure = ntohs(rqstp->rq_addr.sin_port) < 1024;
1316 rqstp->rq_chandle.defer = svc_defer;
1319 serv->sv_stats->netcnt++;
1327 svc_drop(struct svc_rqst *rqstp)
1329 dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1330 svc_sock_release(rqstp);
1334 * Return reply to client.
1337 svc_send(struct svc_rqst *rqstp)
1339 struct svc_sock *svsk;
1343 if ((svsk = rqstp->rq_sock) == NULL) {
1344 printk(KERN_WARNING "NULL socket pointer in %s:%d\n",
1345 __FILE__, __LINE__);
1349 /* release the receive skb before sending the reply */
1350 svc_release_skb(rqstp);
1352 /* calculate over-all length */
1353 xb = & rqstp->rq_res;
1354 xb->len = xb->head[0].iov_len +
1356 xb->tail[0].iov_len;
1358 /* Grab svsk->sk_mutex to serialize outgoing data. */
1359 mutex_lock(&svsk->sk_mutex);
1360 if (test_bit(SK_DEAD, &svsk->sk_flags))
1363 len = svsk->sk_sendto(rqstp);
1364 mutex_unlock(&svsk->sk_mutex);
1365 svc_sock_release(rqstp);
1367 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1373 * Timer function to close old temporary sockets, using
1374 * a mark-and-sweep algorithm.
1377 svc_age_temp_sockets(unsigned long closure)
1379 struct svc_serv *serv = (struct svc_serv *)closure;
1380 struct svc_sock *svsk;
1381 struct list_head *le, *next;
1382 LIST_HEAD(to_be_aged);
1384 dprintk("svc_age_temp_sockets\n");
1386 if (!spin_trylock_bh(&serv->sv_lock)) {
1387 /* busy, try again 1 sec later */
1388 dprintk("svc_age_temp_sockets: busy\n");
1389 mod_timer(&serv->sv_temptimer, jiffies + HZ);
1393 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1394 svsk = list_entry(le, struct svc_sock, sk_list);
1396 if (!test_and_set_bit(SK_OLD, &svsk->sk_flags))
1398 if (atomic_read(&svsk->sk_inuse) || test_bit(SK_BUSY, &svsk->sk_flags))
1400 atomic_inc(&svsk->sk_inuse);
1401 list_move(le, &to_be_aged);
1402 set_bit(SK_CLOSE, &svsk->sk_flags);
1403 set_bit(SK_DETACHED, &svsk->sk_flags);
1405 spin_unlock_bh(&serv->sv_lock);
1407 while (!list_empty(&to_be_aged)) {
1408 le = to_be_aged.next;
1409 /* fiddling the sk_list node is safe 'cos we're SK_DETACHED */
1411 svsk = list_entry(le, struct svc_sock, sk_list);
1413 dprintk("queuing svsk %p for closing, %lu seconds old\n",
1414 svsk, get_seconds() - svsk->sk_lastrecv);
1416 /* a thread will dequeue and close it soon */
1417 svc_sock_enqueue(svsk);
1421 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
1425 * Initialize socket for RPC use and create svc_sock struct
1426 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1428 static struct svc_sock *
1429 svc_setup_socket(struct svc_serv *serv, struct socket *sock,
1430 int *errp, int pmap_register)
1432 struct svc_sock *svsk;
1435 dprintk("svc: svc_setup_socket %p\n", sock);
1436 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1443 /* Register socket with portmapper */
1444 if (*errp >= 0 && pmap_register)
1445 *errp = svc_register(serv, inet->sk_protocol,
1446 ntohs(inet_sk(inet)->sport));
1453 set_bit(SK_BUSY, &svsk->sk_flags);
1454 inet->sk_user_data = svsk;
1455 svsk->sk_sock = sock;
1457 svsk->sk_ostate = inet->sk_state_change;
1458 svsk->sk_odata = inet->sk_data_ready;
1459 svsk->sk_owspace = inet->sk_write_space;
1460 svsk->sk_server = serv;
1461 atomic_set(&svsk->sk_inuse, 0);
1462 svsk->sk_lastrecv = get_seconds();
1463 spin_lock_init(&svsk->sk_defer_lock);
1464 INIT_LIST_HEAD(&svsk->sk_deferred);
1465 INIT_LIST_HEAD(&svsk->sk_ready);
1466 mutex_init(&svsk->sk_mutex);
1468 /* Initialize the socket */
1469 if (sock->type == SOCK_DGRAM)
1474 spin_lock_bh(&serv->sv_lock);
1475 if (!pmap_register) {
1476 set_bit(SK_TEMP, &svsk->sk_flags);
1477 list_add(&svsk->sk_list, &serv->sv_tempsocks);
1479 if (serv->sv_temptimer.function == NULL) {
1480 /* setup timer to age temp sockets */
1481 setup_timer(&serv->sv_temptimer, svc_age_temp_sockets,
1482 (unsigned long)serv);
1483 mod_timer(&serv->sv_temptimer,
1484 jiffies + svc_conn_age_period * HZ);
1487 clear_bit(SK_TEMP, &svsk->sk_flags);
1488 list_add(&svsk->sk_list, &serv->sv_permsocks);
1490 spin_unlock_bh(&serv->sv_lock);
1492 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1495 clear_bit(SK_BUSY, &svsk->sk_flags);
1496 svc_sock_enqueue(svsk);
1500 int svc_addsock(struct svc_serv *serv,
1506 struct socket *so = sockfd_lookup(fd, &err);
1507 struct svc_sock *svsk = NULL;
1511 if (so->sk->sk_family != AF_INET)
1512 err = -EAFNOSUPPORT;
1513 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1514 so->sk->sk_protocol != IPPROTO_UDP)
1515 err = -EPROTONOSUPPORT;
1516 else if (so->state > SS_UNCONNECTED)
1519 svsk = svc_setup_socket(serv, so, &err, 1);
1527 if (proto) *proto = so->sk->sk_protocol;
1528 return one_sock_name(name_return, svsk);
1530 EXPORT_SYMBOL_GPL(svc_addsock);
1533 * Create socket for RPC service.
1536 svc_create_socket(struct svc_serv *serv, int protocol, struct sockaddr_in *sin)
1538 struct svc_sock *svsk;
1539 struct socket *sock;
1543 dprintk("svc: svc_create_socket(%s, %d, %u.%u.%u.%u:%d)\n",
1544 serv->sv_program->pg_name, protocol,
1545 NIPQUAD(sin->sin_addr.s_addr),
1546 ntohs(sin->sin_port));
1548 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1549 printk(KERN_WARNING "svc: only UDP and TCP "
1550 "sockets supported\n");
1553 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1555 if ((error = sock_create_kern(PF_INET, type, protocol, &sock)) < 0)
1558 if (type == SOCK_STREAM)
1559 sock->sk->sk_reuse = 1; /* allow address reuse */
1560 error = kernel_bind(sock, (struct sockaddr *) sin,
1565 if (protocol == IPPROTO_TCP) {
1566 if ((error = kernel_listen(sock, 64)) < 0)
1570 if ((svsk = svc_setup_socket(serv, sock, &error, 1)) != NULL)
1574 dprintk("svc: svc_create_socket error = %d\n", -error);
1580 * Remove a dead socket
1583 svc_delete_socket(struct svc_sock *svsk)
1585 struct svc_serv *serv;
1588 dprintk("svc: svc_delete_socket(%p)\n", svsk);
1590 serv = svsk->sk_server;
1593 sk->sk_state_change = svsk->sk_ostate;
1594 sk->sk_data_ready = svsk->sk_odata;
1595 sk->sk_write_space = svsk->sk_owspace;
1597 spin_lock_bh(&serv->sv_lock);
1599 if (!test_and_set_bit(SK_DETACHED, &svsk->sk_flags))
1600 list_del_init(&svsk->sk_list);
1602 * We used to delete the svc_sock from whichever list
1603 * it's sk_ready node was on, but we don't actually
1604 * need to. This is because the only time we're called
1605 * while still attached to a queue, the queue itself
1606 * is about to be destroyed (in svc_destroy).
1608 if (!test_and_set_bit(SK_DEAD, &svsk->sk_flags))
1609 if (test_bit(SK_TEMP, &svsk->sk_flags))
1612 /* This atomic_inc should be needed - svc_delete_socket
1613 * should have the semantic of dropping a reference.
1614 * But it doesn't yet....
1616 atomic_inc(&svsk->sk_inuse);
1617 spin_unlock_bh(&serv->sv_lock);
1622 * Make a socket for nfsd and lockd
1625 svc_makesock(struct svc_serv *serv, int protocol, unsigned short port)
1627 struct sockaddr_in sin;
1629 dprintk("svc: creating socket proto = %d\n", protocol);
1630 sin.sin_family = AF_INET;
1631 sin.sin_addr.s_addr = INADDR_ANY;
1632 sin.sin_port = htons(port);
1633 return svc_create_socket(serv, protocol, &sin);
1637 * Handle defer and revisit of requests
1640 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1642 struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1643 struct svc_sock *svsk;
1646 svc_sock_put(dr->svsk);
1650 dprintk("revisit queued\n");
1653 spin_lock_bh(&svsk->sk_defer_lock);
1654 list_add(&dr->handle.recent, &svsk->sk_deferred);
1655 spin_unlock_bh(&svsk->sk_defer_lock);
1656 set_bit(SK_DEFERRED, &svsk->sk_flags);
1657 svc_sock_enqueue(svsk);
1661 static struct cache_deferred_req *
1662 svc_defer(struct cache_req *req)
1664 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1665 int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
1666 struct svc_deferred_req *dr;
1668 if (rqstp->rq_arg.page_len)
1669 return NULL; /* if more than a page, give up FIXME */
1670 if (rqstp->rq_deferred) {
1671 dr = rqstp->rq_deferred;
1672 rqstp->rq_deferred = NULL;
1674 int skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1675 /* FIXME maybe discard if size too large */
1676 dr = kmalloc(size, GFP_KERNEL);
1680 dr->handle.owner = rqstp->rq_server;
1681 dr->prot = rqstp->rq_prot;
1682 dr->addr = rqstp->rq_addr;
1683 dr->daddr = rqstp->rq_daddr;
1684 dr->argslen = rqstp->rq_arg.len >> 2;
1685 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
1687 atomic_inc(&rqstp->rq_sock->sk_inuse);
1688 dr->svsk = rqstp->rq_sock;
1690 dr->handle.revisit = svc_revisit;
1695 * recv data from a deferred request into an active one
1697 static int svc_deferred_recv(struct svc_rqst *rqstp)
1699 struct svc_deferred_req *dr = rqstp->rq_deferred;
1701 rqstp->rq_arg.head[0].iov_base = dr->args;
1702 rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
1703 rqstp->rq_arg.page_len = 0;
1704 rqstp->rq_arg.len = dr->argslen<<2;
1705 rqstp->rq_prot = dr->prot;
1706 rqstp->rq_addr = dr->addr;
1707 rqstp->rq_daddr = dr->daddr;
1708 rqstp->rq_respages = rqstp->rq_pages;
1709 return dr->argslen<<2;
1713 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk)
1715 struct svc_deferred_req *dr = NULL;
1717 if (!test_bit(SK_DEFERRED, &svsk->sk_flags))
1719 spin_lock_bh(&svsk->sk_defer_lock);
1720 clear_bit(SK_DEFERRED, &svsk->sk_flags);
1721 if (!list_empty(&svsk->sk_deferred)) {
1722 dr = list_entry(svsk->sk_deferred.next,
1723 struct svc_deferred_req,
1725 list_del_init(&dr->handle.recent);
1726 set_bit(SK_DEFERRED, &svsk->sk_flags);
1728 spin_unlock_bh(&svsk->sk_defer_lock);