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_bufsz)*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_bufsz,
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_bufsz;
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 dprintk("svc: releasing dead socket\n");
304 sock_release(svsk->sk_sock);
310 svc_sock_release(struct svc_rqst *rqstp)
312 struct svc_sock *svsk = rqstp->rq_sock;
314 svc_release_skb(rqstp);
316 svc_free_allpages(rqstp);
317 rqstp->rq_res.page_len = 0;
318 rqstp->rq_res.page_base = 0;
321 /* Reset response buffer and release
323 * But first, check that enough space was reserved
324 * for the reply, otherwise we have a bug!
326 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
327 printk(KERN_ERR "RPC request reserved %d but used %d\n",
331 rqstp->rq_res.head[0].iov_len = 0;
332 svc_reserve(rqstp, 0);
333 rqstp->rq_sock = NULL;
339 * External function to wake up a server waiting for data
340 * This really only makes sense for services like lockd
341 * which have exactly one thread anyway.
344 svc_wake_up(struct svc_serv *serv)
346 struct svc_rqst *rqstp;
348 struct svc_pool *pool;
350 for (i = 0; i < serv->sv_nrpools; i++) {
351 pool = &serv->sv_pools[i];
353 spin_lock_bh(&pool->sp_lock);
354 if (!list_empty(&pool->sp_threads)) {
355 rqstp = list_entry(pool->sp_threads.next,
358 dprintk("svc: daemon %p woken up.\n", rqstp);
360 svc_thread_dequeue(pool, rqstp);
361 rqstp->rq_sock = NULL;
363 wake_up(&rqstp->rq_wait);
365 spin_unlock_bh(&pool->sp_lock);
370 * Generic sendto routine
373 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
375 struct svc_sock *svsk = rqstp->rq_sock;
376 struct socket *sock = svsk->sk_sock;
378 char buffer[CMSG_SPACE(sizeof(struct in_pktinfo))];
379 struct cmsghdr *cmh = (struct cmsghdr *)buffer;
380 struct in_pktinfo *pki = (struct in_pktinfo *)CMSG_DATA(cmh);
384 struct page **ppage = xdr->pages;
385 size_t base = xdr->page_base;
386 unsigned int pglen = xdr->page_len;
387 unsigned int flags = MSG_MORE;
391 if (rqstp->rq_prot == IPPROTO_UDP) {
392 /* set the source and destination */
394 msg.msg_name = &rqstp->rq_addr;
395 msg.msg_namelen = sizeof(rqstp->rq_addr);
398 msg.msg_flags = MSG_MORE;
400 msg.msg_control = cmh;
401 msg.msg_controllen = sizeof(buffer);
402 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
403 cmh->cmsg_level = SOL_IP;
404 cmh->cmsg_type = IP_PKTINFO;
405 pki->ipi_ifindex = 0;
406 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr;
408 if (sock_sendmsg(sock, &msg, 0) < 0)
413 if (slen == xdr->head[0].iov_len)
415 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0, xdr->head[0].iov_len, flags);
416 if (len != xdr->head[0].iov_len)
418 slen -= xdr->head[0].iov_len;
423 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
427 result = kernel_sendpage(sock, *ppage, base, size, flags);
434 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
439 if (xdr->tail[0].iov_len) {
440 result = kernel_sendpage(sock, rqstp->rq_respages[rqstp->rq_restailpage],
441 ((unsigned long)xdr->tail[0].iov_base)& (PAGE_SIZE-1),
442 xdr->tail[0].iov_len, 0);
448 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %x)\n",
449 rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len, xdr->len, len,
450 rqstp->rq_addr.sin_addr.s_addr);
456 * Report socket names for nfsdfs
458 static int one_sock_name(char *buf, struct svc_sock *svsk)
462 switch(svsk->sk_sk->sk_family) {
464 len = sprintf(buf, "ipv4 %s %u.%u.%u.%u %d\n",
465 svsk->sk_sk->sk_protocol==IPPROTO_UDP?
467 NIPQUAD(inet_sk(svsk->sk_sk)->rcv_saddr),
468 inet_sk(svsk->sk_sk)->num);
471 len = sprintf(buf, "*unknown-%d*\n",
472 svsk->sk_sk->sk_family);
478 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
480 struct svc_sock *svsk, *closesk = NULL;
485 spin_lock(&serv->sv_lock);
486 list_for_each_entry(svsk, &serv->sv_permsocks, sk_list) {
487 int onelen = one_sock_name(buf+len, svsk);
488 if (toclose && strcmp(toclose, buf+len) == 0)
493 spin_unlock(&serv->sv_lock);
495 svc_delete_socket(closesk);
498 EXPORT_SYMBOL(svc_sock_names);
501 * Check input queue length
504 svc_recv_available(struct svc_sock *svsk)
506 struct socket *sock = svsk->sk_sock;
509 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
511 return (err >= 0)? avail : err;
515 * Generic recvfrom routine.
518 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
524 rqstp->rq_addrlen = sizeof(rqstp->rq_addr);
525 sock = rqstp->rq_sock->sk_sock;
527 msg.msg_name = &rqstp->rq_addr;
528 msg.msg_namelen = sizeof(rqstp->rq_addr);
529 msg.msg_control = NULL;
530 msg.msg_controllen = 0;
532 msg.msg_flags = MSG_DONTWAIT;
534 len = kernel_recvmsg(sock, &msg, iov, nr, buflen, MSG_DONTWAIT);
536 /* sock_recvmsg doesn't fill in the name/namelen, so we must..
537 * possibly we should cache this in the svc_sock structure
538 * at accept time. FIXME
540 alen = sizeof(rqstp->rq_addr);
541 kernel_getpeername(sock, (struct sockaddr *)&rqstp->rq_addr, &alen);
543 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
544 rqstp->rq_sock, iov[0].iov_base, iov[0].iov_len, len);
550 * Set socket snd and rcv buffer lengths
553 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
557 oldfs = get_fs(); set_fs(KERNEL_DS);
558 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
559 (char*)&snd, sizeof(snd));
560 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
561 (char*)&rcv, sizeof(rcv));
563 /* sock_setsockopt limits use to sysctl_?mem_max,
564 * which isn't acceptable. Until that is made conditional
565 * on not having CAP_SYS_RESOURCE or similar, we go direct...
566 * DaveM said I could!
569 sock->sk->sk_sndbuf = snd * 2;
570 sock->sk->sk_rcvbuf = rcv * 2;
571 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
572 release_sock(sock->sk);
576 * INET callback when data has been received on the socket.
579 svc_udp_data_ready(struct sock *sk, int count)
581 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
584 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
585 svsk, sk, count, test_bit(SK_BUSY, &svsk->sk_flags));
586 set_bit(SK_DATA, &svsk->sk_flags);
587 svc_sock_enqueue(svsk);
589 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
590 wake_up_interruptible(sk->sk_sleep);
594 * INET callback when space is newly available on the socket.
597 svc_write_space(struct sock *sk)
599 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
602 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
603 svsk, sk, test_bit(SK_BUSY, &svsk->sk_flags));
604 svc_sock_enqueue(svsk);
607 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
608 dprintk("RPC svc_write_space: someone sleeping on %p\n",
610 wake_up_interruptible(sk->sk_sleep);
615 * Receive a datagram from a UDP socket.
618 svc_udp_recvfrom(struct svc_rqst *rqstp)
620 struct svc_sock *svsk = rqstp->rq_sock;
621 struct svc_serv *serv = svsk->sk_server;
625 if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
626 /* udp sockets need large rcvbuf as all pending
627 * requests are still in that buffer. sndbuf must
628 * also be large enough that there is enough space
629 * for one reply per thread. We count all threads
630 * rather than threads in a particular pool, which
631 * provides an upper bound on the number of threads
632 * which will access the socket.
634 svc_sock_setbufsize(svsk->sk_sock,
635 (serv->sv_nrthreads+3) * serv->sv_bufsz,
636 (serv->sv_nrthreads+3) * serv->sv_bufsz);
638 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
639 svc_sock_received(svsk);
640 return svc_deferred_recv(rqstp);
643 clear_bit(SK_DATA, &svsk->sk_flags);
644 while ((skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err)) == NULL) {
645 if (err == -EAGAIN) {
646 svc_sock_received(svsk);
649 /* possibly an icmp error */
650 dprintk("svc: recvfrom returned error %d\n", -err);
652 if (skb->tstamp.off_sec == 0) {
655 tv.tv_sec = xtime.tv_sec;
656 tv.tv_usec = xtime.tv_nsec / NSEC_PER_USEC;
657 skb_set_timestamp(skb, &tv);
658 /* Don't enable netstamp, sunrpc doesn't
659 need that much accuracy */
661 skb_get_timestamp(skb, &svsk->sk_sk->sk_stamp);
662 set_bit(SK_DATA, &svsk->sk_flags); /* there may be more data... */
665 * Maybe more packets - kick another thread ASAP.
667 svc_sock_received(svsk);
669 len = skb->len - sizeof(struct udphdr);
670 rqstp->rq_arg.len = len;
672 rqstp->rq_prot = IPPROTO_UDP;
674 /* Get sender address */
675 rqstp->rq_addr.sin_family = AF_INET;
676 rqstp->rq_addr.sin_port = skb->h.uh->source;
677 rqstp->rq_addr.sin_addr.s_addr = skb->nh.iph->saddr;
678 rqstp->rq_daddr = skb->nh.iph->daddr;
680 if (skb_is_nonlinear(skb)) {
681 /* we have to copy */
683 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
686 skb_free_datagram(svsk->sk_sk, skb);
690 skb_free_datagram(svsk->sk_sk, skb);
692 /* we can use it in-place */
693 rqstp->rq_arg.head[0].iov_base = skb->data + sizeof(struct udphdr);
694 rqstp->rq_arg.head[0].iov_len = len;
695 if (skb_checksum_complete(skb)) {
696 skb_free_datagram(svsk->sk_sk, skb);
699 rqstp->rq_skbuff = skb;
702 rqstp->rq_arg.page_base = 0;
703 if (len <= rqstp->rq_arg.head[0].iov_len) {
704 rqstp->rq_arg.head[0].iov_len = len;
705 rqstp->rq_arg.page_len = 0;
707 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
708 rqstp->rq_argused += (rqstp->rq_arg.page_len + PAGE_SIZE - 1)/ PAGE_SIZE;
712 serv->sv_stats->netudpcnt++;
718 svc_udp_sendto(struct svc_rqst *rqstp)
722 error = svc_sendto(rqstp, &rqstp->rq_res);
723 if (error == -ECONNREFUSED)
724 /* ICMP error on earlier request. */
725 error = svc_sendto(rqstp, &rqstp->rq_res);
731 svc_udp_init(struct svc_sock *svsk)
733 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
734 svsk->sk_sk->sk_write_space = svc_write_space;
735 svsk->sk_recvfrom = svc_udp_recvfrom;
736 svsk->sk_sendto = svc_udp_sendto;
738 /* initialise setting must have enough space to
739 * receive and respond to one request.
740 * svc_udp_recvfrom will re-adjust if necessary
742 svc_sock_setbufsize(svsk->sk_sock,
743 3 * svsk->sk_server->sv_bufsz,
744 3 * svsk->sk_server->sv_bufsz);
746 set_bit(SK_DATA, &svsk->sk_flags); /* might have come in before data_ready set up */
747 set_bit(SK_CHNGBUF, &svsk->sk_flags);
751 * A data_ready event on a listening socket means there's a connection
752 * pending. Do not use state_change as a substitute for it.
755 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
757 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
759 dprintk("svc: socket %p TCP (listen) state change %d\n",
763 * This callback may called twice when a new connection
764 * is established as a child socket inherits everything
765 * from a parent LISTEN socket.
766 * 1) data_ready method of the parent socket will be called
767 * when one of child sockets become ESTABLISHED.
768 * 2) data_ready method of the child socket may be called
769 * when it receives data before the socket is accepted.
770 * In case of 2, we should ignore it silently.
772 if (sk->sk_state == TCP_LISTEN) {
774 set_bit(SK_CONN, &svsk->sk_flags);
775 svc_sock_enqueue(svsk);
777 printk("svc: socket %p: no user data\n", sk);
780 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
781 wake_up_interruptible_all(sk->sk_sleep);
785 * A state change on a connected socket means it's dying or dead.
788 svc_tcp_state_change(struct sock *sk)
790 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
792 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
793 sk, sk->sk_state, sk->sk_user_data);
796 printk("svc: socket %p: no user data\n", sk);
798 set_bit(SK_CLOSE, &svsk->sk_flags);
799 svc_sock_enqueue(svsk);
801 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
802 wake_up_interruptible_all(sk->sk_sleep);
806 svc_tcp_data_ready(struct sock *sk, int count)
808 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
810 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
811 sk, sk->sk_user_data);
813 set_bit(SK_DATA, &svsk->sk_flags);
814 svc_sock_enqueue(svsk);
816 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
817 wake_up_interruptible(sk->sk_sleep);
821 * Accept a TCP connection
824 svc_tcp_accept(struct svc_sock *svsk)
826 struct sockaddr_in sin;
827 struct svc_serv *serv = svsk->sk_server;
828 struct socket *sock = svsk->sk_sock;
829 struct socket *newsock;
830 struct svc_sock *newsvsk;
833 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
837 clear_bit(SK_CONN, &svsk->sk_flags);
838 err = kernel_accept(sock, &newsock, O_NONBLOCK);
841 printk(KERN_WARNING "%s: no more sockets!\n",
843 else if (err != -EAGAIN && net_ratelimit())
844 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
845 serv->sv_name, -err);
849 set_bit(SK_CONN, &svsk->sk_flags);
850 svc_sock_enqueue(svsk);
853 err = kernel_getpeername(newsock, (struct sockaddr *) &sin, &slen);
856 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
857 serv->sv_name, -err);
858 goto failed; /* aborted connection or whatever */
861 /* Ideally, we would want to reject connections from unauthorized
862 * hosts here, but when we get encription, the IP of the host won't
863 * tell us anything. For now just warn about unpriv connections.
865 if (ntohs(sin.sin_port) >= 1024) {
867 "%s: connect from unprivileged port: %u.%u.%u.%u:%d\n",
869 NIPQUAD(sin.sin_addr.s_addr), ntohs(sin.sin_port));
872 dprintk("%s: connect from %u.%u.%u.%u:%04x\n", serv->sv_name,
873 NIPQUAD(sin.sin_addr.s_addr), ntohs(sin.sin_port));
875 /* make sure that a write doesn't block forever when
878 newsock->sk->sk_sndtimeo = HZ*30;
880 if (!(newsvsk = svc_setup_socket(serv, newsock, &err, 0)))
884 /* make sure that we don't have too many active connections.
885 * If we have, something must be dropped.
887 * There's no point in trying to do random drop here for
888 * DoS prevention. The NFS clients does 1 reconnect in 15
889 * seconds. An attacker can easily beat that.
891 * The only somewhat efficient mechanism would be if drop
892 * old connections from the same IP first. But right now
893 * we don't even record the client IP in svc_sock.
895 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
896 struct svc_sock *svsk = NULL;
897 spin_lock_bh(&serv->sv_lock);
898 if (!list_empty(&serv->sv_tempsocks)) {
899 if (net_ratelimit()) {
900 /* Try to help the admin */
901 printk(KERN_NOTICE "%s: too many open TCP "
902 "sockets, consider increasing the "
903 "number of nfsd threads\n",
905 printk(KERN_NOTICE "%s: last TCP connect from "
908 NIPQUAD(sin.sin_addr.s_addr),
909 ntohs(sin.sin_port));
912 * Always select the oldest socket. It's not fair,
915 svsk = list_entry(serv->sv_tempsocks.prev,
918 set_bit(SK_CLOSE, &svsk->sk_flags);
919 atomic_inc(&svsk->sk_inuse);
921 spin_unlock_bh(&serv->sv_lock);
924 svc_sock_enqueue(svsk);
931 serv->sv_stats->nettcpconn++;
936 sock_release(newsock);
941 * Receive data from a TCP socket.
944 svc_tcp_recvfrom(struct svc_rqst *rqstp)
946 struct svc_sock *svsk = rqstp->rq_sock;
947 struct svc_serv *serv = svsk->sk_server;
949 struct kvec vec[RPCSVC_MAXPAGES];
952 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
953 svsk, test_bit(SK_DATA, &svsk->sk_flags),
954 test_bit(SK_CONN, &svsk->sk_flags),
955 test_bit(SK_CLOSE, &svsk->sk_flags));
957 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
958 svc_sock_received(svsk);
959 return svc_deferred_recv(rqstp);
962 if (test_bit(SK_CLOSE, &svsk->sk_flags)) {
963 svc_delete_socket(svsk);
967 if (test_bit(SK_CONN, &svsk->sk_flags)) {
968 svc_tcp_accept(svsk);
969 svc_sock_received(svsk);
973 if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
974 /* sndbuf needs to have room for one request
975 * per thread, otherwise we can stall even when the
976 * network isn't a bottleneck.
978 * We count all threads rather than threads in a
979 * particular pool, which provides an upper bound
980 * on the number of threads which will access the socket.
982 * rcvbuf just needs to be able to hold a few requests.
983 * Normally they will be removed from the queue
984 * as soon a a complete request arrives.
986 svc_sock_setbufsize(svsk->sk_sock,
987 (serv->sv_nrthreads+3) * serv->sv_bufsz,
990 clear_bit(SK_DATA, &svsk->sk_flags);
992 /* Receive data. If we haven't got the record length yet, get
993 * the next four bytes. Otherwise try to gobble up as much as
994 * possible up to the complete record length.
996 if (svsk->sk_tcplen < 4) {
997 unsigned long want = 4 - svsk->sk_tcplen;
1000 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
1002 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
1004 svsk->sk_tcplen += len;
1007 dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
1009 svc_sock_received(svsk);
1010 return -EAGAIN; /* record header not complete */
1013 svsk->sk_reclen = ntohl(svsk->sk_reclen);
1014 if (!(svsk->sk_reclen & 0x80000000)) {
1015 /* FIXME: technically, a record can be fragmented,
1016 * and non-terminal fragments will not have the top
1017 * bit set in the fragment length header.
1018 * But apparently no known nfs clients send fragmented
1020 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx (non-terminal)\n",
1021 (unsigned long) svsk->sk_reclen);
1024 svsk->sk_reclen &= 0x7fffffff;
1025 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
1026 if (svsk->sk_reclen > serv->sv_bufsz) {
1027 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx (large)\n",
1028 (unsigned long) svsk->sk_reclen);
1033 /* Check whether enough data is available */
1034 len = svc_recv_available(svsk);
1038 if (len < svsk->sk_reclen) {
1039 dprintk("svc: incomplete TCP record (%d of %d)\n",
1040 len, svsk->sk_reclen);
1041 svc_sock_received(svsk);
1042 return -EAGAIN; /* record not complete */
1044 len = svsk->sk_reclen;
1045 set_bit(SK_DATA, &svsk->sk_flags);
1047 vec[0] = rqstp->rq_arg.head[0];
1050 while (vlen < len) {
1051 vec[pnum].iov_base = page_address(rqstp->rq_argpages[rqstp->rq_argused++]);
1052 vec[pnum].iov_len = PAGE_SIZE;
1057 /* Now receive data */
1058 len = svc_recvfrom(rqstp, vec, pnum, len);
1062 dprintk("svc: TCP complete record (%d bytes)\n", len);
1063 rqstp->rq_arg.len = len;
1064 rqstp->rq_arg.page_base = 0;
1065 if (len <= rqstp->rq_arg.head[0].iov_len) {
1066 rqstp->rq_arg.head[0].iov_len = len;
1067 rqstp->rq_arg.page_len = 0;
1069 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1072 rqstp->rq_skbuff = NULL;
1073 rqstp->rq_prot = IPPROTO_TCP;
1075 /* Reset TCP read info */
1076 svsk->sk_reclen = 0;
1077 svsk->sk_tcplen = 0;
1079 svc_sock_received(svsk);
1081 serv->sv_stats->nettcpcnt++;
1086 svc_delete_socket(svsk);
1090 if (len == -EAGAIN) {
1091 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1092 svc_sock_received(svsk);
1094 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1095 svsk->sk_server->sv_name, -len);
1103 * Send out data on TCP socket.
1106 svc_tcp_sendto(struct svc_rqst *rqstp)
1108 struct xdr_buf *xbufp = &rqstp->rq_res;
1112 /* Set up the first element of the reply kvec.
1113 * Any other kvecs that may be in use have been taken
1114 * care of by the server implementation itself.
1116 reclen = htonl(0x80000000|((xbufp->len ) - 4));
1117 memcpy(xbufp->head[0].iov_base, &reclen, 4);
1119 if (test_bit(SK_DEAD, &rqstp->rq_sock->sk_flags))
1122 sent = svc_sendto(rqstp, &rqstp->rq_res);
1123 if (sent != xbufp->len) {
1124 printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1125 rqstp->rq_sock->sk_server->sv_name,
1126 (sent<0)?"got error":"sent only",
1128 svc_delete_socket(rqstp->rq_sock);
1135 svc_tcp_init(struct svc_sock *svsk)
1137 struct sock *sk = svsk->sk_sk;
1138 struct tcp_sock *tp = tcp_sk(sk);
1140 svsk->sk_recvfrom = svc_tcp_recvfrom;
1141 svsk->sk_sendto = svc_tcp_sendto;
1143 if (sk->sk_state == TCP_LISTEN) {
1144 dprintk("setting up TCP socket for listening\n");
1145 sk->sk_data_ready = svc_tcp_listen_data_ready;
1146 set_bit(SK_CONN, &svsk->sk_flags);
1148 dprintk("setting up TCP socket for reading\n");
1149 sk->sk_state_change = svc_tcp_state_change;
1150 sk->sk_data_ready = svc_tcp_data_ready;
1151 sk->sk_write_space = svc_write_space;
1153 svsk->sk_reclen = 0;
1154 svsk->sk_tcplen = 0;
1156 tp->nonagle = 1; /* disable Nagle's algorithm */
1158 /* initialise setting must have enough space to
1159 * receive and respond to one request.
1160 * svc_tcp_recvfrom will re-adjust if necessary
1162 svc_sock_setbufsize(svsk->sk_sock,
1163 3 * svsk->sk_server->sv_bufsz,
1164 3 * svsk->sk_server->sv_bufsz);
1166 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1167 set_bit(SK_DATA, &svsk->sk_flags);
1168 if (sk->sk_state != TCP_ESTABLISHED)
1169 set_bit(SK_CLOSE, &svsk->sk_flags);
1174 svc_sock_update_bufs(struct svc_serv *serv)
1177 * The number of server threads has changed. Update
1178 * rcvbuf and sndbuf accordingly on all sockets
1180 struct list_head *le;
1182 spin_lock_bh(&serv->sv_lock);
1183 list_for_each(le, &serv->sv_permsocks) {
1184 struct svc_sock *svsk =
1185 list_entry(le, struct svc_sock, sk_list);
1186 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1188 list_for_each(le, &serv->sv_tempsocks) {
1189 struct svc_sock *svsk =
1190 list_entry(le, struct svc_sock, sk_list);
1191 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1193 spin_unlock_bh(&serv->sv_lock);
1197 * Receive the next request on any socket. This code is carefully
1198 * organised not to touch any cachelines in the shared svc_serv
1199 * structure, only cachelines in the local svc_pool.
1202 svc_recv(struct svc_rqst *rqstp, long timeout)
1204 struct svc_sock *svsk =NULL;
1205 struct svc_serv *serv = rqstp->rq_server;
1206 struct svc_pool *pool = rqstp->rq_pool;
1209 struct xdr_buf *arg;
1210 DECLARE_WAITQUEUE(wait, current);
1212 dprintk("svc: server %p waiting for data (to = %ld)\n",
1217 "svc_recv: service %p, socket not NULL!\n",
1219 if (waitqueue_active(&rqstp->rq_wait))
1221 "svc_recv: service %p, wait queue active!\n",
1224 /* Initialize the buffers */
1225 /* first reclaim pages that were moved to response list */
1226 svc_pushback_allpages(rqstp);
1228 /* now allocate needed pages. If we get a failure, sleep briefly */
1229 pages = 2 + (serv->sv_bufsz + PAGE_SIZE -1) / PAGE_SIZE;
1230 while (rqstp->rq_arghi < pages) {
1231 struct page *p = alloc_page(GFP_KERNEL);
1233 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1236 rqstp->rq_argpages[rqstp->rq_arghi++] = p;
1239 /* Make arg->head point to first page and arg->pages point to rest */
1240 arg = &rqstp->rq_arg;
1241 arg->head[0].iov_base = page_address(rqstp->rq_argpages[0]);
1242 arg->head[0].iov_len = PAGE_SIZE;
1243 rqstp->rq_argused = 1;
1244 arg->pages = rqstp->rq_argpages + 1;
1246 /* save at least one page for response */
1247 arg->page_len = (pages-2)*PAGE_SIZE;
1248 arg->len = (pages-1)*PAGE_SIZE;
1249 arg->tail[0].iov_len = 0;
1256 spin_lock_bh(&pool->sp_lock);
1257 if ((svsk = svc_sock_dequeue(pool)) != NULL) {
1258 rqstp->rq_sock = svsk;
1259 atomic_inc(&svsk->sk_inuse);
1260 rqstp->rq_reserved = serv->sv_bufsz;
1261 atomic_add(rqstp->rq_reserved, &svsk->sk_reserved);
1263 /* No data pending. Go to sleep */
1264 svc_thread_enqueue(pool, rqstp);
1267 * We have to be able to interrupt this wait
1268 * to bring down the daemons ...
1270 set_current_state(TASK_INTERRUPTIBLE);
1271 add_wait_queue(&rqstp->rq_wait, &wait);
1272 spin_unlock_bh(&pool->sp_lock);
1274 schedule_timeout(timeout);
1278 spin_lock_bh(&pool->sp_lock);
1279 remove_wait_queue(&rqstp->rq_wait, &wait);
1281 if (!(svsk = rqstp->rq_sock)) {
1282 svc_thread_dequeue(pool, rqstp);
1283 spin_unlock_bh(&pool->sp_lock);
1284 dprintk("svc: server %p, no data yet\n", rqstp);
1285 return signalled()? -EINTR : -EAGAIN;
1288 spin_unlock_bh(&pool->sp_lock);
1290 dprintk("svc: server %p, pool %u, socket %p, inuse=%d\n",
1291 rqstp, pool->sp_id, svsk, atomic_read(&svsk->sk_inuse));
1292 len = svsk->sk_recvfrom(rqstp);
1293 dprintk("svc: got len=%d\n", len);
1295 /* No data, incomplete (TCP) read, or accept() */
1296 if (len == 0 || len == -EAGAIN) {
1297 rqstp->rq_res.len = 0;
1298 svc_sock_release(rqstp);
1301 svsk->sk_lastrecv = get_seconds();
1302 clear_bit(SK_OLD, &svsk->sk_flags);
1304 rqstp->rq_secure = ntohs(rqstp->rq_addr.sin_port) < 1024;
1305 rqstp->rq_chandle.defer = svc_defer;
1308 serv->sv_stats->netcnt++;
1316 svc_drop(struct svc_rqst *rqstp)
1318 dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1319 svc_sock_release(rqstp);
1323 * Return reply to client.
1326 svc_send(struct svc_rqst *rqstp)
1328 struct svc_sock *svsk;
1332 if ((svsk = rqstp->rq_sock) == NULL) {
1333 printk(KERN_WARNING "NULL socket pointer in %s:%d\n",
1334 __FILE__, __LINE__);
1338 /* release the receive skb before sending the reply */
1339 svc_release_skb(rqstp);
1341 /* calculate over-all length */
1342 xb = & rqstp->rq_res;
1343 xb->len = xb->head[0].iov_len +
1345 xb->tail[0].iov_len;
1347 /* Grab svsk->sk_mutex to serialize outgoing data. */
1348 mutex_lock(&svsk->sk_mutex);
1349 if (test_bit(SK_DEAD, &svsk->sk_flags))
1352 len = svsk->sk_sendto(rqstp);
1353 mutex_unlock(&svsk->sk_mutex);
1354 svc_sock_release(rqstp);
1356 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1362 * Timer function to close old temporary sockets, using
1363 * a mark-and-sweep algorithm.
1366 svc_age_temp_sockets(unsigned long closure)
1368 struct svc_serv *serv = (struct svc_serv *)closure;
1369 struct svc_sock *svsk;
1370 struct list_head *le, *next;
1371 LIST_HEAD(to_be_aged);
1373 dprintk("svc_age_temp_sockets\n");
1375 if (!spin_trylock_bh(&serv->sv_lock)) {
1376 /* busy, try again 1 sec later */
1377 dprintk("svc_age_temp_sockets: busy\n");
1378 mod_timer(&serv->sv_temptimer, jiffies + HZ);
1382 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1383 svsk = list_entry(le, struct svc_sock, sk_list);
1385 if (!test_and_set_bit(SK_OLD, &svsk->sk_flags))
1387 if (atomic_read(&svsk->sk_inuse) || test_bit(SK_BUSY, &svsk->sk_flags))
1389 atomic_inc(&svsk->sk_inuse);
1390 list_move(le, &to_be_aged);
1391 set_bit(SK_CLOSE, &svsk->sk_flags);
1392 set_bit(SK_DETACHED, &svsk->sk_flags);
1394 spin_unlock_bh(&serv->sv_lock);
1396 while (!list_empty(&to_be_aged)) {
1397 le = to_be_aged.next;
1398 /* fiddling the sk_list node is safe 'cos we're SK_DETACHED */
1400 svsk = list_entry(le, struct svc_sock, sk_list);
1402 dprintk("queuing svsk %p for closing, %lu seconds old\n",
1403 svsk, get_seconds() - svsk->sk_lastrecv);
1405 /* a thread will dequeue and close it soon */
1406 svc_sock_enqueue(svsk);
1410 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
1414 * Initialize socket for RPC use and create svc_sock struct
1415 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1417 static struct svc_sock *
1418 svc_setup_socket(struct svc_serv *serv, struct socket *sock,
1419 int *errp, int pmap_register)
1421 struct svc_sock *svsk;
1424 dprintk("svc: svc_setup_socket %p\n", sock);
1425 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1432 /* Register socket with portmapper */
1433 if (*errp >= 0 && pmap_register)
1434 *errp = svc_register(serv, inet->sk_protocol,
1435 ntohs(inet_sk(inet)->sport));
1442 set_bit(SK_BUSY, &svsk->sk_flags);
1443 inet->sk_user_data = svsk;
1444 svsk->sk_sock = sock;
1446 svsk->sk_ostate = inet->sk_state_change;
1447 svsk->sk_odata = inet->sk_data_ready;
1448 svsk->sk_owspace = inet->sk_write_space;
1449 svsk->sk_server = serv;
1450 atomic_set(&svsk->sk_inuse, 0);
1451 svsk->sk_lastrecv = get_seconds();
1452 spin_lock_init(&svsk->sk_defer_lock);
1453 INIT_LIST_HEAD(&svsk->sk_deferred);
1454 INIT_LIST_HEAD(&svsk->sk_ready);
1455 mutex_init(&svsk->sk_mutex);
1457 /* Initialize the socket */
1458 if (sock->type == SOCK_DGRAM)
1463 spin_lock_bh(&serv->sv_lock);
1464 if (!pmap_register) {
1465 set_bit(SK_TEMP, &svsk->sk_flags);
1466 list_add(&svsk->sk_list, &serv->sv_tempsocks);
1468 if (serv->sv_temptimer.function == NULL) {
1469 /* setup timer to age temp sockets */
1470 setup_timer(&serv->sv_temptimer, svc_age_temp_sockets,
1471 (unsigned long)serv);
1472 mod_timer(&serv->sv_temptimer,
1473 jiffies + svc_conn_age_period * HZ);
1476 clear_bit(SK_TEMP, &svsk->sk_flags);
1477 list_add(&svsk->sk_list, &serv->sv_permsocks);
1479 spin_unlock_bh(&serv->sv_lock);
1481 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1484 clear_bit(SK_BUSY, &svsk->sk_flags);
1485 svc_sock_enqueue(svsk);
1489 int svc_addsock(struct svc_serv *serv,
1495 struct socket *so = sockfd_lookup(fd, &err);
1496 struct svc_sock *svsk = NULL;
1500 if (so->sk->sk_family != AF_INET)
1501 err = -EAFNOSUPPORT;
1502 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1503 so->sk->sk_protocol != IPPROTO_UDP)
1504 err = -EPROTONOSUPPORT;
1505 else if (so->state > SS_UNCONNECTED)
1508 svsk = svc_setup_socket(serv, so, &err, 1);
1516 if (proto) *proto = so->sk->sk_protocol;
1517 return one_sock_name(name_return, svsk);
1519 EXPORT_SYMBOL_GPL(svc_addsock);
1522 * Create socket for RPC service.
1525 svc_create_socket(struct svc_serv *serv, int protocol, struct sockaddr_in *sin)
1527 struct svc_sock *svsk;
1528 struct socket *sock;
1532 dprintk("svc: svc_create_socket(%s, %d, %u.%u.%u.%u:%d)\n",
1533 serv->sv_program->pg_name, protocol,
1534 NIPQUAD(sin->sin_addr.s_addr),
1535 ntohs(sin->sin_port));
1537 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1538 printk(KERN_WARNING "svc: only UDP and TCP "
1539 "sockets supported\n");
1542 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1544 if ((error = sock_create_kern(PF_INET, type, protocol, &sock)) < 0)
1547 if (type == SOCK_STREAM)
1548 sock->sk->sk_reuse = 1; /* allow address reuse */
1549 error = kernel_bind(sock, (struct sockaddr *) sin,
1554 if (protocol == IPPROTO_TCP) {
1555 if ((error = kernel_listen(sock, 64)) < 0)
1559 if ((svsk = svc_setup_socket(serv, sock, &error, 1)) != NULL)
1563 dprintk("svc: svc_create_socket error = %d\n", -error);
1569 * Remove a dead socket
1572 svc_delete_socket(struct svc_sock *svsk)
1574 struct svc_serv *serv;
1577 dprintk("svc: svc_delete_socket(%p)\n", svsk);
1579 serv = svsk->sk_server;
1582 sk->sk_state_change = svsk->sk_ostate;
1583 sk->sk_data_ready = svsk->sk_odata;
1584 sk->sk_write_space = svsk->sk_owspace;
1586 spin_lock_bh(&serv->sv_lock);
1588 if (!test_and_set_bit(SK_DETACHED, &svsk->sk_flags))
1589 list_del_init(&svsk->sk_list);
1591 * We used to delete the svc_sock from whichever list
1592 * it's sk_ready node was on, but we don't actually
1593 * need to. This is because the only time we're called
1594 * while still attached to a queue, the queue itself
1595 * is about to be destroyed (in svc_destroy).
1597 if (!test_and_set_bit(SK_DEAD, &svsk->sk_flags))
1598 if (test_bit(SK_TEMP, &svsk->sk_flags))
1601 if (!atomic_read(&svsk->sk_inuse)) {
1602 spin_unlock_bh(&serv->sv_lock);
1603 if (svsk->sk_sock->file)
1604 sockfd_put(svsk->sk_sock);
1606 sock_release(svsk->sk_sock);
1609 spin_unlock_bh(&serv->sv_lock);
1610 dprintk(KERN_NOTICE "svc: server socket destroy delayed\n");
1611 /* svsk->sk_server = NULL; */
1616 * Make a socket for nfsd and lockd
1619 svc_makesock(struct svc_serv *serv, int protocol, unsigned short port)
1621 struct sockaddr_in sin;
1623 dprintk("svc: creating socket proto = %d\n", protocol);
1624 sin.sin_family = AF_INET;
1625 sin.sin_addr.s_addr = INADDR_ANY;
1626 sin.sin_port = htons(port);
1627 return svc_create_socket(serv, protocol, &sin);
1631 * Handle defer and revisit of requests
1634 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1636 struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1637 struct svc_sock *svsk;
1640 svc_sock_put(dr->svsk);
1644 dprintk("revisit queued\n");
1647 spin_lock_bh(&svsk->sk_defer_lock);
1648 list_add(&dr->handle.recent, &svsk->sk_deferred);
1649 spin_unlock_bh(&svsk->sk_defer_lock);
1650 set_bit(SK_DEFERRED, &svsk->sk_flags);
1651 svc_sock_enqueue(svsk);
1655 static struct cache_deferred_req *
1656 svc_defer(struct cache_req *req)
1658 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1659 int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
1660 struct svc_deferred_req *dr;
1662 if (rqstp->rq_arg.page_len)
1663 return NULL; /* if more than a page, give up FIXME */
1664 if (rqstp->rq_deferred) {
1665 dr = rqstp->rq_deferred;
1666 rqstp->rq_deferred = NULL;
1668 int skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1669 /* FIXME maybe discard if size too large */
1670 dr = kmalloc(size, GFP_KERNEL);
1674 dr->handle.owner = rqstp->rq_server;
1675 dr->prot = rqstp->rq_prot;
1676 dr->addr = rqstp->rq_addr;
1677 dr->daddr = rqstp->rq_daddr;
1678 dr->argslen = rqstp->rq_arg.len >> 2;
1679 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
1681 atomic_inc(&rqstp->rq_sock->sk_inuse);
1682 dr->svsk = rqstp->rq_sock;
1684 dr->handle.revisit = svc_revisit;
1689 * recv data from a deferred request into an active one
1691 static int svc_deferred_recv(struct svc_rqst *rqstp)
1693 struct svc_deferred_req *dr = rqstp->rq_deferred;
1695 rqstp->rq_arg.head[0].iov_base = dr->args;
1696 rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
1697 rqstp->rq_arg.page_len = 0;
1698 rqstp->rq_arg.len = dr->argslen<<2;
1699 rqstp->rq_prot = dr->prot;
1700 rqstp->rq_addr = dr->addr;
1701 rqstp->rq_daddr = dr->daddr;
1702 return dr->argslen<<2;
1706 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk)
1708 struct svc_deferred_req *dr = NULL;
1710 if (!test_bit(SK_DEFERRED, &svsk->sk_flags))
1712 spin_lock_bh(&svsk->sk_defer_lock);
1713 clear_bit(SK_DEFERRED, &svsk->sk_flags);
1714 if (!list_empty(&svsk->sk_deferred)) {
1715 dr = list_entry(svsk->sk_deferred.next,
1716 struct svc_deferred_req,
1718 list_del_init(&dr->handle.recent);
1719 set_bit(SK_DEFERRED, &svsk->sk_flags);
1721 spin_unlock_bh(&svsk->sk_defer_lock);