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/unistd.h>
30 #include <linux/slab.h>
31 #include <linux/netdevice.h>
32 #include <linux/skbuff.h>
34 #include <net/checksum.h>
36 #include <net/tcp_states.h>
37 #include <asm/uaccess.h>
38 #include <asm/ioctls.h>
40 #include <linux/sunrpc/types.h>
41 #include <linux/sunrpc/xdr.h>
42 #include <linux/sunrpc/svcsock.h>
43 #include <linux/sunrpc/stats.h>
45 /* SMP locking strategy:
47 * svc_serv->sv_lock protects most stuff for that service.
49 * Some flags can be set to certain values at any time
50 * providing that certain rules are followed:
52 * SK_BUSY can be set to 0 at any time.
53 * svc_sock_enqueue must be called afterwards
54 * SK_CONN, SK_DATA, can be set or cleared at any time.
55 * after a set, svc_sock_enqueue must be called.
56 * after a clear, the socket must be read/accepted
57 * if this succeeds, it must be set again.
58 * SK_CLOSE can set at any time. It is never cleared.
62 #define RPCDBG_FACILITY RPCDBG_SVCSOCK
65 static struct svc_sock *svc_setup_socket(struct svc_serv *, struct socket *,
66 int *errp, int pmap_reg);
67 static void svc_udp_data_ready(struct sock *, int);
68 static int svc_udp_recvfrom(struct svc_rqst *);
69 static int svc_udp_sendto(struct svc_rqst *);
71 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk);
72 static int svc_deferred_recv(struct svc_rqst *rqstp);
73 static struct cache_deferred_req *svc_defer(struct cache_req *req);
76 * Queue up an idle server thread. Must have serv->sv_lock held.
77 * Note: this is really a stack rather than a queue, so that we only
78 * use as many different threads as we need, and the rest don't polute
82 svc_serv_enqueue(struct svc_serv *serv, struct svc_rqst *rqstp)
84 list_add(&rqstp->rq_list, &serv->sv_threads);
88 * Dequeue an nfsd thread. Must have serv->sv_lock held.
91 svc_serv_dequeue(struct svc_serv *serv, struct svc_rqst *rqstp)
93 list_del(&rqstp->rq_list);
97 * Release an skbuff after use
100 svc_release_skb(struct svc_rqst *rqstp)
102 struct sk_buff *skb = rqstp->rq_skbuff;
103 struct svc_deferred_req *dr = rqstp->rq_deferred;
106 rqstp->rq_skbuff = NULL;
108 dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
109 skb_free_datagram(rqstp->rq_sock->sk_sk, skb);
112 rqstp->rq_deferred = NULL;
118 * Any space to write?
120 static inline unsigned long
121 svc_sock_wspace(struct svc_sock *svsk)
125 if (svsk->sk_sock->type == SOCK_STREAM)
126 wspace = sk_stream_wspace(svsk->sk_sk);
128 wspace = sock_wspace(svsk->sk_sk);
134 * Queue up a socket with data pending. If there are idle nfsd
135 * processes, wake 'em up.
139 svc_sock_enqueue(struct svc_sock *svsk)
141 struct svc_serv *serv = svsk->sk_server;
142 struct svc_rqst *rqstp;
144 if (!(svsk->sk_flags &
145 ( (1<<SK_CONN)|(1<<SK_DATA)|(1<<SK_CLOSE)|(1<<SK_DEFERRED)) ))
147 if (test_bit(SK_DEAD, &svsk->sk_flags))
150 spin_lock_bh(&serv->sv_lock);
152 if (!list_empty(&serv->sv_threads) &&
153 !list_empty(&serv->sv_sockets))
155 "svc_sock_enqueue: threads and sockets both waiting??\n");
157 if (test_bit(SK_DEAD, &svsk->sk_flags)) {
158 /* Don't enqueue dead sockets */
159 dprintk("svc: socket %p is dead, not enqueued\n", svsk->sk_sk);
163 if (test_bit(SK_BUSY, &svsk->sk_flags)) {
164 /* Don't enqueue socket while daemon is receiving */
165 dprintk("svc: socket %p busy, not enqueued\n", svsk->sk_sk);
169 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
170 if (((svsk->sk_reserved + serv->sv_bufsz)*2
171 > svc_sock_wspace(svsk))
172 && !test_bit(SK_CLOSE, &svsk->sk_flags)
173 && !test_bit(SK_CONN, &svsk->sk_flags)) {
174 /* Don't enqueue while not enough space for reply */
175 dprintk("svc: socket %p no space, %d*2 > %ld, not enqueued\n",
176 svsk->sk_sk, svsk->sk_reserved+serv->sv_bufsz,
177 svc_sock_wspace(svsk));
180 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
182 /* Mark socket as busy. It will remain in this state until the
183 * server has processed all pending data and put the socket back
186 set_bit(SK_BUSY, &svsk->sk_flags);
188 if (!list_empty(&serv->sv_threads)) {
189 rqstp = list_entry(serv->sv_threads.next,
192 dprintk("svc: socket %p served by daemon %p\n",
194 svc_serv_dequeue(serv, rqstp);
197 "svc_sock_enqueue: server %p, rq_sock=%p!\n",
198 rqstp, rqstp->rq_sock);
199 rqstp->rq_sock = svsk;
201 rqstp->rq_reserved = serv->sv_bufsz;
202 svsk->sk_reserved += rqstp->rq_reserved;
203 wake_up(&rqstp->rq_wait);
205 dprintk("svc: socket %p put into queue\n", svsk->sk_sk);
206 list_add_tail(&svsk->sk_ready, &serv->sv_sockets);
210 spin_unlock_bh(&serv->sv_lock);
214 * Dequeue the first socket. Must be called with the serv->sv_lock held.
216 static inline struct svc_sock *
217 svc_sock_dequeue(struct svc_serv *serv)
219 struct svc_sock *svsk;
221 if (list_empty(&serv->sv_sockets))
224 svsk = list_entry(serv->sv_sockets.next,
225 struct svc_sock, sk_ready);
226 list_del_init(&svsk->sk_ready);
228 dprintk("svc: socket %p dequeued, inuse=%d\n",
229 svsk->sk_sk, svsk->sk_inuse);
235 * Having read something from a socket, check whether it
236 * needs to be re-enqueued.
237 * Note: SK_DATA only gets cleared when a read-attempt finds
238 * no (or insufficient) data.
241 svc_sock_received(struct svc_sock *svsk)
243 clear_bit(SK_BUSY, &svsk->sk_flags);
244 svc_sock_enqueue(svsk);
249 * svc_reserve - change the space reserved for the reply to a request.
250 * @rqstp: The request in question
251 * @space: new max space to reserve
253 * Each request reserves some space on the output queue of the socket
254 * to make sure the reply fits. This function reduces that reserved
255 * space to be the amount of space used already, plus @space.
258 void svc_reserve(struct svc_rqst *rqstp, int space)
260 space += rqstp->rq_res.head[0].iov_len;
262 if (space < rqstp->rq_reserved) {
263 struct svc_sock *svsk = rqstp->rq_sock;
264 spin_lock_bh(&svsk->sk_server->sv_lock);
265 svsk->sk_reserved -= (rqstp->rq_reserved - space);
266 rqstp->rq_reserved = space;
267 spin_unlock_bh(&svsk->sk_server->sv_lock);
269 svc_sock_enqueue(svsk);
274 * Release a socket after use.
277 svc_sock_put(struct svc_sock *svsk)
279 struct svc_serv *serv = svsk->sk_server;
281 spin_lock_bh(&serv->sv_lock);
282 if (!--(svsk->sk_inuse) && test_bit(SK_DEAD, &svsk->sk_flags)) {
283 spin_unlock_bh(&serv->sv_lock);
284 dprintk("svc: releasing dead socket\n");
285 sock_release(svsk->sk_sock);
289 spin_unlock_bh(&serv->sv_lock);
293 svc_sock_release(struct svc_rqst *rqstp)
295 struct svc_sock *svsk = rqstp->rq_sock;
297 svc_release_skb(rqstp);
299 svc_free_allpages(rqstp);
300 rqstp->rq_res.page_len = 0;
301 rqstp->rq_res.page_base = 0;
304 /* Reset response buffer and release
306 * But first, check that enough space was reserved
307 * for the reply, otherwise we have a bug!
309 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
310 printk(KERN_ERR "RPC request reserved %d but used %d\n",
314 rqstp->rq_res.head[0].iov_len = 0;
315 svc_reserve(rqstp, 0);
316 rqstp->rq_sock = NULL;
322 * External function to wake up a server waiting for data
325 svc_wake_up(struct svc_serv *serv)
327 struct svc_rqst *rqstp;
329 spin_lock_bh(&serv->sv_lock);
330 if (!list_empty(&serv->sv_threads)) {
331 rqstp = list_entry(serv->sv_threads.next,
334 dprintk("svc: daemon %p woken up.\n", rqstp);
336 svc_serv_dequeue(serv, rqstp);
337 rqstp->rq_sock = NULL;
339 wake_up(&rqstp->rq_wait);
341 spin_unlock_bh(&serv->sv_lock);
345 * Generic sendto routine
348 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
350 struct svc_sock *svsk = rqstp->rq_sock;
351 struct socket *sock = svsk->sk_sock;
353 char buffer[CMSG_SPACE(sizeof(struct in_pktinfo))];
354 struct cmsghdr *cmh = (struct cmsghdr *)buffer;
355 struct in_pktinfo *pki = (struct in_pktinfo *)CMSG_DATA(cmh);
359 struct page **ppage = xdr->pages;
360 size_t base = xdr->page_base;
361 unsigned int pglen = xdr->page_len;
362 unsigned int flags = MSG_MORE;
366 if (rqstp->rq_prot == IPPROTO_UDP) {
367 /* set the source and destination */
369 msg.msg_name = &rqstp->rq_addr;
370 msg.msg_namelen = sizeof(rqstp->rq_addr);
373 msg.msg_flags = MSG_MORE;
375 msg.msg_control = cmh;
376 msg.msg_controllen = sizeof(buffer);
377 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
378 cmh->cmsg_level = SOL_IP;
379 cmh->cmsg_type = IP_PKTINFO;
380 pki->ipi_ifindex = 0;
381 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr;
383 if (sock_sendmsg(sock, &msg, 0) < 0)
388 if (slen == xdr->head[0].iov_len)
390 len = sock->ops->sendpage(sock, rqstp->rq_respages[0], 0, xdr->head[0].iov_len, flags);
391 if (len != xdr->head[0].iov_len)
393 slen -= xdr->head[0].iov_len;
398 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
402 result = sock->ops->sendpage(sock, *ppage, base, size, flags);
409 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
414 if (xdr->tail[0].iov_len) {
415 result = sock->ops->sendpage(sock, rqstp->rq_respages[rqstp->rq_restailpage],
416 ((unsigned long)xdr->tail[0].iov_base)& (PAGE_SIZE-1),
417 xdr->tail[0].iov_len, 0);
423 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %x)\n",
424 rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len, xdr->len, len,
425 rqstp->rq_addr.sin_addr.s_addr);
431 * Check input queue length
434 svc_recv_available(struct svc_sock *svsk)
437 struct socket *sock = svsk->sk_sock;
440 oldfs = get_fs(); set_fs(KERNEL_DS);
441 err = sock->ops->ioctl(sock, TIOCINQ, (unsigned long) &avail);
444 return (err >= 0)? avail : err;
448 * Generic recvfrom routine.
451 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
457 rqstp->rq_addrlen = sizeof(rqstp->rq_addr);
458 sock = rqstp->rq_sock->sk_sock;
460 msg.msg_name = &rqstp->rq_addr;
461 msg.msg_namelen = sizeof(rqstp->rq_addr);
462 msg.msg_control = NULL;
463 msg.msg_controllen = 0;
465 msg.msg_flags = MSG_DONTWAIT;
467 len = kernel_recvmsg(sock, &msg, iov, nr, buflen, MSG_DONTWAIT);
469 /* sock_recvmsg doesn't fill in the name/namelen, so we must..
470 * possibly we should cache this in the svc_sock structure
471 * at accept time. FIXME
473 alen = sizeof(rqstp->rq_addr);
474 sock->ops->getname(sock, (struct sockaddr *)&rqstp->rq_addr, &alen, 1);
476 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
477 rqstp->rq_sock, iov[0].iov_base, iov[0].iov_len, len);
483 * Set socket snd and rcv buffer lengths
486 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
490 oldfs = get_fs(); set_fs(KERNEL_DS);
491 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
492 (char*)&snd, sizeof(snd));
493 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
494 (char*)&rcv, sizeof(rcv));
496 /* sock_setsockopt limits use to sysctl_?mem_max,
497 * which isn't acceptable. Until that is made conditional
498 * on not having CAP_SYS_RESOURCE or similar, we go direct...
499 * DaveM said I could!
502 sock->sk->sk_sndbuf = snd * 2;
503 sock->sk->sk_rcvbuf = rcv * 2;
504 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
505 release_sock(sock->sk);
509 * INET callback when data has been received on the socket.
512 svc_udp_data_ready(struct sock *sk, int count)
514 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
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);
523 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
524 wake_up_interruptible(sk->sk_sleep);
528 * INET callback when space is newly available on the socket.
531 svc_write_space(struct sock *sk)
533 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
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);
541 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
542 printk(KERN_WARNING "RPC svc_write_space: some sleeping on %p\n",
544 wake_up_interruptible(sk->sk_sleep);
549 * Receive a datagram from a UDP socket.
552 csum_partial_copy_to_xdr(struct xdr_buf *xdr, struct sk_buff *skb);
555 svc_udp_recvfrom(struct svc_rqst *rqstp)
557 struct svc_sock *svsk = rqstp->rq_sock;
558 struct svc_serv *serv = svsk->sk_server;
562 if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
563 /* udp sockets need large rcvbuf as all pending
564 * requests are still in that buffer. sndbuf must
565 * also be large enough that there is enough space
566 * for one reply per thread.
568 svc_sock_setbufsize(svsk->sk_sock,
569 (serv->sv_nrthreads+3) * serv->sv_bufsz,
570 (serv->sv_nrthreads+3) * serv->sv_bufsz);
572 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
573 svc_sock_received(svsk);
574 return svc_deferred_recv(rqstp);
577 clear_bit(SK_DATA, &svsk->sk_flags);
578 while ((skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err)) == NULL) {
579 if (err == -EAGAIN) {
580 svc_sock_received(svsk);
583 /* possibly an icmp error */
584 dprintk("svc: recvfrom returned error %d\n", -err);
586 if (skb->stamp.tv_sec == 0) {
587 skb->stamp.tv_sec = xtime.tv_sec;
588 skb->stamp.tv_usec = xtime.tv_nsec / NSEC_PER_USEC;
589 /* Don't enable netstamp, sunrpc doesn't
590 need that much accuracy */
592 svsk->sk_sk->sk_stamp = skb->stamp;
593 set_bit(SK_DATA, &svsk->sk_flags); /* there may be more data... */
596 * Maybe more packets - kick another thread ASAP.
598 svc_sock_received(svsk);
600 len = skb->len - sizeof(struct udphdr);
601 rqstp->rq_arg.len = len;
603 rqstp->rq_prot = IPPROTO_UDP;
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;
611 if (skb_is_nonlinear(skb)) {
612 /* we have to copy */
614 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
617 skb_free_datagram(svsk->sk_sk, skb);
621 skb_free_datagram(svsk->sk_sk, skb);
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->ip_summed != CHECKSUM_UNNECESSARY) {
627 if ((unsigned short)csum_fold(skb_checksum(skb, 0, skb->len, skb->csum))) {
628 skb_free_datagram(svsk->sk_sk, skb);
631 skb->ip_summed = CHECKSUM_UNNECESSARY;
633 rqstp->rq_skbuff = skb;
636 rqstp->rq_arg.page_base = 0;
637 if (len <= rqstp->rq_arg.head[0].iov_len) {
638 rqstp->rq_arg.head[0].iov_len = len;
639 rqstp->rq_arg.page_len = 0;
641 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
642 rqstp->rq_argused += (rqstp->rq_arg.page_len + PAGE_SIZE - 1)/ PAGE_SIZE;
646 serv->sv_stats->netudpcnt++;
652 svc_udp_sendto(struct svc_rqst *rqstp)
656 error = svc_sendto(rqstp, &rqstp->rq_res);
657 if (error == -ECONNREFUSED)
658 /* ICMP error on earlier request. */
659 error = svc_sendto(rqstp, &rqstp->rq_res);
665 svc_udp_init(struct svc_sock *svsk)
667 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
668 svsk->sk_sk->sk_write_space = svc_write_space;
669 svsk->sk_recvfrom = svc_udp_recvfrom;
670 svsk->sk_sendto = svc_udp_sendto;
672 /* initialise setting must have enough space to
673 * receive and respond to one request.
674 * svc_udp_recvfrom will re-adjust if necessary
676 svc_sock_setbufsize(svsk->sk_sock,
677 3 * svsk->sk_server->sv_bufsz,
678 3 * svsk->sk_server->sv_bufsz);
680 set_bit(SK_DATA, &svsk->sk_flags); /* might have come in before data_ready set up */
681 set_bit(SK_CHNGBUF, &svsk->sk_flags);
685 * A data_ready event on a listening socket means there's a connection
686 * pending. Do not use state_change as a substitute for it.
689 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
691 struct svc_sock *svsk;
693 dprintk("svc: socket %p TCP (listen) state change %d\n",
696 if (sk->sk_state != TCP_LISTEN) {
698 * This callback may called twice when a new connection
699 * is established as a child socket inherits everything
700 * from a parent LISTEN socket.
701 * 1) data_ready method of the parent socket will be called
702 * when one of child sockets become ESTABLISHED.
703 * 2) data_ready method of the child socket may be called
704 * when it receives data before the socket is accepted.
705 * In case of 2, we should ignore it silently.
709 if (!(svsk = (struct svc_sock *) sk->sk_user_data)) {
710 printk("svc: socket %p: no user data\n", sk);
713 set_bit(SK_CONN, &svsk->sk_flags);
714 svc_sock_enqueue(svsk);
716 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
717 wake_up_interruptible_all(sk->sk_sleep);
721 * A state change on a connected socket means it's dying or dead.
724 svc_tcp_state_change(struct sock *sk)
726 struct svc_sock *svsk;
728 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
729 sk, sk->sk_state, sk->sk_user_data);
731 if (!(svsk = (struct svc_sock *) sk->sk_user_data)) {
732 printk("svc: socket %p: no user data\n", sk);
735 set_bit(SK_CLOSE, &svsk->sk_flags);
736 svc_sock_enqueue(svsk);
738 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
739 wake_up_interruptible_all(sk->sk_sleep);
743 svc_tcp_data_ready(struct sock *sk, int count)
745 struct svc_sock * svsk;
747 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
748 sk, sk->sk_user_data);
749 if (!(svsk = (struct svc_sock *)(sk->sk_user_data)))
751 set_bit(SK_DATA, &svsk->sk_flags);
752 svc_sock_enqueue(svsk);
754 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
755 wake_up_interruptible(sk->sk_sleep);
759 * Accept a TCP connection
762 svc_tcp_accept(struct svc_sock *svsk)
764 struct sockaddr_in sin;
765 struct svc_serv *serv = svsk->sk_server;
766 struct socket *sock = svsk->sk_sock;
767 struct socket *newsock;
768 struct proto_ops *ops;
769 struct svc_sock *newsvsk;
772 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
776 err = sock_create_lite(PF_INET, SOCK_STREAM, IPPROTO_TCP, &newsock);
779 printk(KERN_WARNING "%s: no more sockets!\n",
784 dprintk("svc: tcp_accept %p allocated\n", newsock);
785 newsock->ops = ops = sock->ops;
787 clear_bit(SK_CONN, &svsk->sk_flags);
788 if ((err = ops->accept(sock, newsock, O_NONBLOCK)) < 0) {
789 if (err != -EAGAIN && net_ratelimit())
790 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
791 serv->sv_name, -err);
792 goto failed; /* aborted connection or whatever */
794 set_bit(SK_CONN, &svsk->sk_flags);
795 svc_sock_enqueue(svsk);
798 err = ops->getname(newsock, (struct sockaddr *) &sin, &slen, 1);
801 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
802 serv->sv_name, -err);
803 goto failed; /* aborted connection or whatever */
806 /* Ideally, we would want to reject connections from unauthorized
807 * hosts here, but when we get encription, the IP of the host won't
808 * tell us anything. For now just warn about unpriv connections.
810 if (ntohs(sin.sin_port) >= 1024) {
812 "%s: connect from unprivileged port: %u.%u.%u.%u:%d\n",
814 NIPQUAD(sin.sin_addr.s_addr), ntohs(sin.sin_port));
817 dprintk("%s: connect from %u.%u.%u.%u:%04x\n", serv->sv_name,
818 NIPQUAD(sin.sin_addr.s_addr), ntohs(sin.sin_port));
820 /* make sure that a write doesn't block forever when
823 newsock->sk->sk_sndtimeo = HZ*30;
825 if (!(newsvsk = svc_setup_socket(serv, newsock, &err, 0)))
829 /* make sure that we don't have too many active connections.
830 * If we have, something must be dropped.
832 * There's no point in trying to do random drop here for
833 * DoS prevention. The NFS clients does 1 reconnect in 15
834 * seconds. An attacker can easily beat that.
836 * The only somewhat efficient mechanism would be if drop
837 * old connections from the same IP first. But right now
838 * we don't even record the client IP in svc_sock.
840 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
841 struct svc_sock *svsk = NULL;
842 spin_lock_bh(&serv->sv_lock);
843 if (!list_empty(&serv->sv_tempsocks)) {
844 if (net_ratelimit()) {
845 /* Try to help the admin */
846 printk(KERN_NOTICE "%s: too many open TCP "
847 "sockets, consider increasing the "
848 "number of nfsd threads\n",
850 printk(KERN_NOTICE "%s: last TCP connect from "
853 NIPQUAD(sin.sin_addr.s_addr),
854 ntohs(sin.sin_port));
857 * Always select the oldest socket. It's not fair,
860 svsk = list_entry(serv->sv_tempsocks.prev,
863 set_bit(SK_CLOSE, &svsk->sk_flags);
866 spin_unlock_bh(&serv->sv_lock);
869 svc_sock_enqueue(svsk);
876 serv->sv_stats->nettcpconn++;
881 sock_release(newsock);
886 * Receive data from a TCP socket.
889 svc_tcp_recvfrom(struct svc_rqst *rqstp)
891 struct svc_sock *svsk = rqstp->rq_sock;
892 struct svc_serv *serv = svsk->sk_server;
894 struct kvec vec[RPCSVC_MAXPAGES];
897 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
898 svsk, test_bit(SK_DATA, &svsk->sk_flags),
899 test_bit(SK_CONN, &svsk->sk_flags),
900 test_bit(SK_CLOSE, &svsk->sk_flags));
902 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
903 svc_sock_received(svsk);
904 return svc_deferred_recv(rqstp);
907 if (test_bit(SK_CLOSE, &svsk->sk_flags)) {
908 svc_delete_socket(svsk);
912 if (test_bit(SK_CONN, &svsk->sk_flags)) {
913 svc_tcp_accept(svsk);
914 svc_sock_received(svsk);
918 if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
919 /* sndbuf needs to have room for one request
920 * per thread, otherwise we can stall even when the
921 * network isn't a bottleneck.
922 * rcvbuf just needs to be able to hold a few requests.
923 * Normally they will be removed from the queue
924 * as soon a a complete request arrives.
926 svc_sock_setbufsize(svsk->sk_sock,
927 (serv->sv_nrthreads+3) * serv->sv_bufsz,
930 clear_bit(SK_DATA, &svsk->sk_flags);
932 /* Receive data. If we haven't got the record length yet, get
933 * the next four bytes. Otherwise try to gobble up as much as
934 * possible up to the complete record length.
936 if (svsk->sk_tcplen < 4) {
937 unsigned long want = 4 - svsk->sk_tcplen;
940 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
942 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
944 svsk->sk_tcplen += len;
947 dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
949 svc_sock_received(svsk);
950 return -EAGAIN; /* record header not complete */
953 svsk->sk_reclen = ntohl(svsk->sk_reclen);
954 if (!(svsk->sk_reclen & 0x80000000)) {
955 /* FIXME: technically, a record can be fragmented,
956 * and non-terminal fragments will not have the top
957 * bit set in the fragment length header.
958 * But apparently no known nfs clients send fragmented
960 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx (non-terminal)\n",
961 (unsigned long) svsk->sk_reclen);
964 svsk->sk_reclen &= 0x7fffffff;
965 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
966 if (svsk->sk_reclen > serv->sv_bufsz) {
967 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx (large)\n",
968 (unsigned long) svsk->sk_reclen);
973 /* Check whether enough data is available */
974 len = svc_recv_available(svsk);
978 if (len < svsk->sk_reclen) {
979 dprintk("svc: incomplete TCP record (%d of %d)\n",
980 len, svsk->sk_reclen);
981 svc_sock_received(svsk);
982 return -EAGAIN; /* record not complete */
984 len = svsk->sk_reclen;
985 set_bit(SK_DATA, &svsk->sk_flags);
987 vec[0] = rqstp->rq_arg.head[0];
991 vec[pnum].iov_base = page_address(rqstp->rq_argpages[rqstp->rq_argused++]);
992 vec[pnum].iov_len = PAGE_SIZE;
997 /* Now receive data */
998 len = svc_recvfrom(rqstp, vec, pnum, len);
1002 dprintk("svc: TCP complete record (%d bytes)\n", len);
1003 rqstp->rq_arg.len = len;
1004 rqstp->rq_arg.page_base = 0;
1005 if (len <= rqstp->rq_arg.head[0].iov_len) {
1006 rqstp->rq_arg.head[0].iov_len = len;
1007 rqstp->rq_arg.page_len = 0;
1009 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1012 rqstp->rq_skbuff = NULL;
1013 rqstp->rq_prot = IPPROTO_TCP;
1015 /* Reset TCP read info */
1016 svsk->sk_reclen = 0;
1017 svsk->sk_tcplen = 0;
1019 svc_sock_received(svsk);
1021 serv->sv_stats->nettcpcnt++;
1026 svc_delete_socket(svsk);
1030 if (len == -EAGAIN) {
1031 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1032 svc_sock_received(svsk);
1034 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1035 svsk->sk_server->sv_name, -len);
1036 svc_sock_received(svsk);
1043 * Send out data on TCP socket.
1046 svc_tcp_sendto(struct svc_rqst *rqstp)
1048 struct xdr_buf *xbufp = &rqstp->rq_res;
1052 /* Set up the first element of the reply kvec.
1053 * Any other kvecs that may be in use have been taken
1054 * care of by the server implementation itself.
1056 reclen = htonl(0x80000000|((xbufp->len ) - 4));
1057 memcpy(xbufp->head[0].iov_base, &reclen, 4);
1059 if (test_bit(SK_DEAD, &rqstp->rq_sock->sk_flags))
1062 sent = svc_sendto(rqstp, &rqstp->rq_res);
1063 if (sent != xbufp->len) {
1064 printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1065 rqstp->rq_sock->sk_server->sv_name,
1066 (sent<0)?"got error":"sent only",
1068 svc_delete_socket(rqstp->rq_sock);
1075 svc_tcp_init(struct svc_sock *svsk)
1077 struct sock *sk = svsk->sk_sk;
1078 struct tcp_sock *tp = tcp_sk(sk);
1080 svsk->sk_recvfrom = svc_tcp_recvfrom;
1081 svsk->sk_sendto = svc_tcp_sendto;
1083 if (sk->sk_state == TCP_LISTEN) {
1084 dprintk("setting up TCP socket for listening\n");
1085 sk->sk_data_ready = svc_tcp_listen_data_ready;
1086 set_bit(SK_CONN, &svsk->sk_flags);
1088 dprintk("setting up TCP socket for reading\n");
1089 sk->sk_state_change = svc_tcp_state_change;
1090 sk->sk_data_ready = svc_tcp_data_ready;
1091 sk->sk_write_space = svc_write_space;
1093 svsk->sk_reclen = 0;
1094 svsk->sk_tcplen = 0;
1096 tp->nonagle = 1; /* disable Nagle's algorithm */
1098 /* initialise setting must have enough space to
1099 * receive and respond to one request.
1100 * svc_tcp_recvfrom will re-adjust if necessary
1102 svc_sock_setbufsize(svsk->sk_sock,
1103 3 * svsk->sk_server->sv_bufsz,
1104 3 * svsk->sk_server->sv_bufsz);
1106 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1107 set_bit(SK_DATA, &svsk->sk_flags);
1108 if (sk->sk_state != TCP_ESTABLISHED)
1109 set_bit(SK_CLOSE, &svsk->sk_flags);
1114 svc_sock_update_bufs(struct svc_serv *serv)
1117 * The number of server threads has changed. Update
1118 * rcvbuf and sndbuf accordingly on all sockets
1120 struct list_head *le;
1122 spin_lock_bh(&serv->sv_lock);
1123 list_for_each(le, &serv->sv_permsocks) {
1124 struct svc_sock *svsk =
1125 list_entry(le, struct svc_sock, sk_list);
1126 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1128 list_for_each(le, &serv->sv_tempsocks) {
1129 struct svc_sock *svsk =
1130 list_entry(le, struct svc_sock, sk_list);
1131 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1133 spin_unlock_bh(&serv->sv_lock);
1137 * Receive the next request on any socket.
1140 svc_recv(struct svc_serv *serv, struct svc_rqst *rqstp, long timeout)
1142 struct svc_sock *svsk =NULL;
1145 struct xdr_buf *arg;
1146 DECLARE_WAITQUEUE(wait, current);
1148 dprintk("svc: server %p waiting for data (to = %ld)\n",
1153 "svc_recv: service %p, socket not NULL!\n",
1155 if (waitqueue_active(&rqstp->rq_wait))
1157 "svc_recv: service %p, wait queue active!\n",
1160 /* Initialize the buffers */
1161 /* first reclaim pages that were moved to response list */
1162 svc_pushback_allpages(rqstp);
1164 /* now allocate needed pages. If we get a failure, sleep briefly */
1165 pages = 2 + (serv->sv_bufsz + PAGE_SIZE -1) / PAGE_SIZE;
1166 while (rqstp->rq_arghi < pages) {
1167 struct page *p = alloc_page(GFP_KERNEL);
1169 set_current_state(TASK_UNINTERRUPTIBLE);
1170 schedule_timeout(HZ/2);
1173 rqstp->rq_argpages[rqstp->rq_arghi++] = p;
1176 /* Make arg->head point to first page and arg->pages point to rest */
1177 arg = &rqstp->rq_arg;
1178 arg->head[0].iov_base = page_address(rqstp->rq_argpages[0]);
1179 arg->head[0].iov_len = PAGE_SIZE;
1180 rqstp->rq_argused = 1;
1181 arg->pages = rqstp->rq_argpages + 1;
1183 /* save at least one page for response */
1184 arg->page_len = (pages-2)*PAGE_SIZE;
1185 arg->len = (pages-1)*PAGE_SIZE;
1186 arg->tail[0].iov_len = 0;
1192 spin_lock_bh(&serv->sv_lock);
1193 if (!list_empty(&serv->sv_tempsocks)) {
1194 svsk = list_entry(serv->sv_tempsocks.next,
1195 struct svc_sock, sk_list);
1196 /* apparently the "standard" is that clients close
1197 * idle connections after 5 minutes, servers after
1199 * http://www.connectathon.org/talks96/nfstcp.pdf
1201 if (get_seconds() - svsk->sk_lastrecv < 6*60
1202 || test_bit(SK_BUSY, &svsk->sk_flags))
1206 set_bit(SK_BUSY, &svsk->sk_flags);
1207 set_bit(SK_CLOSE, &svsk->sk_flags);
1208 rqstp->rq_sock = svsk;
1210 } else if ((svsk = svc_sock_dequeue(serv)) != NULL) {
1211 rqstp->rq_sock = svsk;
1213 rqstp->rq_reserved = serv->sv_bufsz;
1214 svsk->sk_reserved += rqstp->rq_reserved;
1216 /* No data pending. Go to sleep */
1217 svc_serv_enqueue(serv, rqstp);
1220 * We have to be able to interrupt this wait
1221 * to bring down the daemons ...
1223 set_current_state(TASK_INTERRUPTIBLE);
1224 add_wait_queue(&rqstp->rq_wait, &wait);
1225 spin_unlock_bh(&serv->sv_lock);
1227 schedule_timeout(timeout);
1231 spin_lock_bh(&serv->sv_lock);
1232 remove_wait_queue(&rqstp->rq_wait, &wait);
1234 if (!(svsk = rqstp->rq_sock)) {
1235 svc_serv_dequeue(serv, rqstp);
1236 spin_unlock_bh(&serv->sv_lock);
1237 dprintk("svc: server %p, no data yet\n", rqstp);
1238 return signalled()? -EINTR : -EAGAIN;
1241 spin_unlock_bh(&serv->sv_lock);
1243 dprintk("svc: server %p, socket %p, inuse=%d\n",
1244 rqstp, svsk, svsk->sk_inuse);
1245 len = svsk->sk_recvfrom(rqstp);
1246 dprintk("svc: got len=%d\n", len);
1248 /* No data, incomplete (TCP) read, or accept() */
1249 if (len == 0 || len == -EAGAIN) {
1250 rqstp->rq_res.len = 0;
1251 svc_sock_release(rqstp);
1254 svsk->sk_lastrecv = get_seconds();
1255 if (test_bit(SK_TEMP, &svsk->sk_flags)) {
1256 /* push active sockets to end of list */
1257 spin_lock_bh(&serv->sv_lock);
1258 if (!list_empty(&svsk->sk_list))
1259 list_move_tail(&svsk->sk_list, &serv->sv_tempsocks);
1260 spin_unlock_bh(&serv->sv_lock);
1263 rqstp->rq_secure = ntohs(rqstp->rq_addr.sin_port) < 1024;
1264 rqstp->rq_chandle.defer = svc_defer;
1267 serv->sv_stats->netcnt++;
1275 svc_drop(struct svc_rqst *rqstp)
1277 dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1278 svc_sock_release(rqstp);
1282 * Return reply to client.
1285 svc_send(struct svc_rqst *rqstp)
1287 struct svc_sock *svsk;
1291 if ((svsk = rqstp->rq_sock) == NULL) {
1292 printk(KERN_WARNING "NULL socket pointer in %s:%d\n",
1293 __FILE__, __LINE__);
1297 /* release the receive skb before sending the reply */
1298 svc_release_skb(rqstp);
1300 /* calculate over-all length */
1301 xb = & rqstp->rq_res;
1302 xb->len = xb->head[0].iov_len +
1304 xb->tail[0].iov_len;
1306 /* Grab svsk->sk_sem to serialize outgoing data. */
1307 down(&svsk->sk_sem);
1308 if (test_bit(SK_DEAD, &svsk->sk_flags))
1311 len = svsk->sk_sendto(rqstp);
1313 svc_sock_release(rqstp);
1315 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1321 * Initialize socket for RPC use and create svc_sock struct
1322 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1324 static struct svc_sock *
1325 svc_setup_socket(struct svc_serv *serv, struct socket *sock,
1326 int *errp, int pmap_register)
1328 struct svc_sock *svsk;
1331 dprintk("svc: svc_setup_socket %p\n", sock);
1332 if (!(svsk = kmalloc(sizeof(*svsk), GFP_KERNEL))) {
1336 memset(svsk, 0, sizeof(*svsk));
1340 /* Register socket with portmapper */
1341 if (*errp >= 0 && pmap_register)
1342 *errp = svc_register(serv, inet->sk_protocol,
1343 ntohs(inet_sk(inet)->sport));
1350 set_bit(SK_BUSY, &svsk->sk_flags);
1351 inet->sk_user_data = svsk;
1352 svsk->sk_sock = sock;
1354 svsk->sk_ostate = inet->sk_state_change;
1355 svsk->sk_odata = inet->sk_data_ready;
1356 svsk->sk_owspace = inet->sk_write_space;
1357 svsk->sk_server = serv;
1358 svsk->sk_lastrecv = get_seconds();
1359 INIT_LIST_HEAD(&svsk->sk_deferred);
1360 INIT_LIST_HEAD(&svsk->sk_ready);
1361 sema_init(&svsk->sk_sem, 1);
1363 /* Initialize the socket */
1364 if (sock->type == SOCK_DGRAM)
1369 spin_lock_bh(&serv->sv_lock);
1370 if (!pmap_register) {
1371 set_bit(SK_TEMP, &svsk->sk_flags);
1372 list_add(&svsk->sk_list, &serv->sv_tempsocks);
1375 clear_bit(SK_TEMP, &svsk->sk_flags);
1376 list_add(&svsk->sk_list, &serv->sv_permsocks);
1378 spin_unlock_bh(&serv->sv_lock);
1380 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1383 clear_bit(SK_BUSY, &svsk->sk_flags);
1384 svc_sock_enqueue(svsk);
1389 * Create socket for RPC service.
1392 svc_create_socket(struct svc_serv *serv, int protocol, struct sockaddr_in *sin)
1394 struct svc_sock *svsk;
1395 struct socket *sock;
1399 dprintk("svc: svc_create_socket(%s, %d, %u.%u.%u.%u:%d)\n",
1400 serv->sv_program->pg_name, protocol,
1401 NIPQUAD(sin->sin_addr.s_addr),
1402 ntohs(sin->sin_port));
1404 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1405 printk(KERN_WARNING "svc: only UDP and TCP "
1406 "sockets supported\n");
1409 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1411 if ((error = sock_create_kern(PF_INET, type, protocol, &sock)) < 0)
1415 if (type == SOCK_STREAM)
1416 sock->sk->sk_reuse = 1; /* allow address reuse */
1417 error = sock->ops->bind(sock, (struct sockaddr *) sin,
1423 if (protocol == IPPROTO_TCP) {
1424 if ((error = sock->ops->listen(sock, 64)) < 0)
1428 if ((svsk = svc_setup_socket(serv, sock, &error, 1)) != NULL)
1432 dprintk("svc: svc_create_socket error = %d\n", -error);
1438 * Remove a dead socket
1441 svc_delete_socket(struct svc_sock *svsk)
1443 struct svc_serv *serv;
1446 dprintk("svc: svc_delete_socket(%p)\n", svsk);
1448 serv = svsk->sk_server;
1451 sk->sk_state_change = svsk->sk_ostate;
1452 sk->sk_data_ready = svsk->sk_odata;
1453 sk->sk_write_space = svsk->sk_owspace;
1455 spin_lock_bh(&serv->sv_lock);
1457 list_del_init(&svsk->sk_list);
1458 list_del_init(&svsk->sk_ready);
1459 if (!test_and_set_bit(SK_DEAD, &svsk->sk_flags))
1460 if (test_bit(SK_TEMP, &svsk->sk_flags))
1463 if (!svsk->sk_inuse) {
1464 spin_unlock_bh(&serv->sv_lock);
1465 sock_release(svsk->sk_sock);
1468 spin_unlock_bh(&serv->sv_lock);
1469 dprintk(KERN_NOTICE "svc: server socket destroy delayed\n");
1470 /* svsk->sk_server = NULL; */
1475 * Make a socket for nfsd and lockd
1478 svc_makesock(struct svc_serv *serv, int protocol, unsigned short port)
1480 struct sockaddr_in sin;
1482 dprintk("svc: creating socket proto = %d\n", protocol);
1483 sin.sin_family = AF_INET;
1484 sin.sin_addr.s_addr = INADDR_ANY;
1485 sin.sin_port = htons(port);
1486 return svc_create_socket(serv, protocol, &sin);
1490 * Handle defer and revisit of requests
1493 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1495 struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1496 struct svc_serv *serv = dreq->owner;
1497 struct svc_sock *svsk;
1500 svc_sock_put(dr->svsk);
1504 dprintk("revisit queued\n");
1507 spin_lock_bh(&serv->sv_lock);
1508 list_add(&dr->handle.recent, &svsk->sk_deferred);
1509 spin_unlock_bh(&serv->sv_lock);
1510 set_bit(SK_DEFERRED, &svsk->sk_flags);
1511 svc_sock_enqueue(svsk);
1515 static struct cache_deferred_req *
1516 svc_defer(struct cache_req *req)
1518 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1519 int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
1520 struct svc_deferred_req *dr;
1522 if (rqstp->rq_arg.page_len)
1523 return NULL; /* if more than a page, give up FIXME */
1524 if (rqstp->rq_deferred) {
1525 dr = rqstp->rq_deferred;
1526 rqstp->rq_deferred = NULL;
1528 int skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1529 /* FIXME maybe discard if size too large */
1530 dr = kmalloc(size, GFP_KERNEL);
1534 dr->handle.owner = rqstp->rq_server;
1535 dr->prot = rqstp->rq_prot;
1536 dr->addr = rqstp->rq_addr;
1537 dr->argslen = rqstp->rq_arg.len >> 2;
1538 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
1540 spin_lock_bh(&rqstp->rq_server->sv_lock);
1541 rqstp->rq_sock->sk_inuse++;
1542 dr->svsk = rqstp->rq_sock;
1543 spin_unlock_bh(&rqstp->rq_server->sv_lock);
1545 dr->handle.revisit = svc_revisit;
1550 * recv data from a deferred request into an active one
1552 static int svc_deferred_recv(struct svc_rqst *rqstp)
1554 struct svc_deferred_req *dr = rqstp->rq_deferred;
1556 rqstp->rq_arg.head[0].iov_base = dr->args;
1557 rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
1558 rqstp->rq_arg.page_len = 0;
1559 rqstp->rq_arg.len = dr->argslen<<2;
1560 rqstp->rq_prot = dr->prot;
1561 rqstp->rq_addr = dr->addr;
1562 return dr->argslen<<2;
1566 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk)
1568 struct svc_deferred_req *dr = NULL;
1569 struct svc_serv *serv = svsk->sk_server;
1571 if (!test_bit(SK_DEFERRED, &svsk->sk_flags))
1573 spin_lock_bh(&serv->sv_lock);
1574 clear_bit(SK_DEFERRED, &svsk->sk_flags);
1575 if (!list_empty(&svsk->sk_deferred)) {
1576 dr = list_entry(svsk->sk_deferred.next,
1577 struct svc_deferred_req,
1579 list_del_init(&dr->handle.recent);
1580 set_bit(SK_DEFERRED, &svsk->sk_flags);
1582 spin_unlock_bh(&serv->sv_lock);