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>
35 #include <net/checksum.h>
37 #include <net/tcp_states.h>
38 #include <asm/uaccess.h>
39 #include <asm/ioctls.h>
41 #include <linux/sunrpc/types.h>
42 #include <linux/sunrpc/xdr.h>
43 #include <linux/sunrpc/svcsock.h>
44 #include <linux/sunrpc/stats.h>
46 /* SMP locking strategy:
48 * svc_serv->sv_lock protects most stuff for that service.
50 * Some flags can be set to certain values at any time
51 * providing that certain rules are followed:
53 * SK_BUSY can be set to 0 at any time.
54 * svc_sock_enqueue must be called afterwards
55 * SK_CONN, SK_DATA, can be set or cleared at any time.
56 * after a set, svc_sock_enqueue must be called.
57 * after a clear, the socket must be read/accepted
58 * if this succeeds, it must be set again.
59 * SK_CLOSE can set at any time. It is never cleared.
63 #define RPCDBG_FACILITY RPCDBG_SVCSOCK
66 static struct svc_sock *svc_setup_socket(struct svc_serv *, struct socket *,
67 int *errp, int pmap_reg);
68 static void svc_udp_data_ready(struct sock *, int);
69 static int svc_udp_recvfrom(struct svc_rqst *);
70 static int svc_udp_sendto(struct svc_rqst *);
72 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk);
73 static int svc_deferred_recv(struct svc_rqst *rqstp);
74 static struct cache_deferred_req *svc_defer(struct cache_req *req);
77 * Queue up an idle server thread. Must have serv->sv_lock held.
78 * Note: this is really a stack rather than a queue, so that we only
79 * use as many different threads as we need, and the rest don't polute
83 svc_serv_enqueue(struct svc_serv *serv, struct svc_rqst *rqstp)
85 list_add(&rqstp->rq_list, &serv->sv_threads);
89 * Dequeue an nfsd thread. Must have serv->sv_lock held.
92 svc_serv_dequeue(struct svc_serv *serv, struct svc_rqst *rqstp)
94 list_del(&rqstp->rq_list);
98 * Release an skbuff after use
101 svc_release_skb(struct svc_rqst *rqstp)
103 struct sk_buff *skb = rqstp->rq_skbuff;
104 struct svc_deferred_req *dr = rqstp->rq_deferred;
107 rqstp->rq_skbuff = NULL;
109 dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
110 skb_free_datagram(rqstp->rq_sock->sk_sk, skb);
113 rqstp->rq_deferred = NULL;
119 * Any space to write?
121 static inline unsigned long
122 svc_sock_wspace(struct svc_sock *svsk)
126 if (svsk->sk_sock->type == SOCK_STREAM)
127 wspace = sk_stream_wspace(svsk->sk_sk);
129 wspace = sock_wspace(svsk->sk_sk);
135 * Queue up a socket with data pending. If there are idle nfsd
136 * processes, wake 'em up.
140 svc_sock_enqueue(struct svc_sock *svsk)
142 struct svc_serv *serv = svsk->sk_server;
143 struct svc_rqst *rqstp;
145 if (!(svsk->sk_flags &
146 ( (1<<SK_CONN)|(1<<SK_DATA)|(1<<SK_CLOSE)|(1<<SK_DEFERRED)) ))
148 if (test_bit(SK_DEAD, &svsk->sk_flags))
151 spin_lock_bh(&serv->sv_lock);
153 if (!list_empty(&serv->sv_threads) &&
154 !list_empty(&serv->sv_sockets))
156 "svc_sock_enqueue: threads and sockets both waiting??\n");
158 if (test_bit(SK_DEAD, &svsk->sk_flags)) {
159 /* Don't enqueue dead sockets */
160 dprintk("svc: socket %p is dead, not enqueued\n", svsk->sk_sk);
164 if (test_bit(SK_BUSY, &svsk->sk_flags)) {
165 /* Don't enqueue socket while daemon is receiving */
166 dprintk("svc: socket %p busy, not enqueued\n", svsk->sk_sk);
170 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
171 if (((svsk->sk_reserved + serv->sv_bufsz)*2
172 > svc_sock_wspace(svsk))
173 && !test_bit(SK_CLOSE, &svsk->sk_flags)
174 && !test_bit(SK_CONN, &svsk->sk_flags)) {
175 /* Don't enqueue while not enough space for reply */
176 dprintk("svc: socket %p no space, %d*2 > %ld, not enqueued\n",
177 svsk->sk_sk, svsk->sk_reserved+serv->sv_bufsz,
178 svc_sock_wspace(svsk));
181 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
183 /* Mark socket as busy. It will remain in this state until the
184 * server has processed all pending data and put the socket back
187 set_bit(SK_BUSY, &svsk->sk_flags);
189 if (!list_empty(&serv->sv_threads)) {
190 rqstp = list_entry(serv->sv_threads.next,
193 dprintk("svc: socket %p served by daemon %p\n",
195 svc_serv_dequeue(serv, rqstp);
198 "svc_sock_enqueue: server %p, rq_sock=%p!\n",
199 rqstp, rqstp->rq_sock);
200 rqstp->rq_sock = svsk;
202 rqstp->rq_reserved = serv->sv_bufsz;
203 svsk->sk_reserved += rqstp->rq_reserved;
204 wake_up(&rqstp->rq_wait);
206 dprintk("svc: socket %p put into queue\n", svsk->sk_sk);
207 list_add_tail(&svsk->sk_ready, &serv->sv_sockets);
211 spin_unlock_bh(&serv->sv_lock);
215 * Dequeue the first socket. Must be called with the serv->sv_lock held.
217 static inline struct svc_sock *
218 svc_sock_dequeue(struct svc_serv *serv)
220 struct svc_sock *svsk;
222 if (list_empty(&serv->sv_sockets))
225 svsk = list_entry(serv->sv_sockets.next,
226 struct svc_sock, sk_ready);
227 list_del_init(&svsk->sk_ready);
229 dprintk("svc: socket %p dequeued, inuse=%d\n",
230 svsk->sk_sk, svsk->sk_inuse);
236 * Having read something from a socket, check whether it
237 * needs to be re-enqueued.
238 * Note: SK_DATA only gets cleared when a read-attempt finds
239 * no (or insufficient) data.
242 svc_sock_received(struct svc_sock *svsk)
244 clear_bit(SK_BUSY, &svsk->sk_flags);
245 svc_sock_enqueue(svsk);
250 * svc_reserve - change the space reserved for the reply to a request.
251 * @rqstp: The request in question
252 * @space: new max space to reserve
254 * Each request reserves some space on the output queue of the socket
255 * to make sure the reply fits. This function reduces that reserved
256 * space to be the amount of space used already, plus @space.
259 void svc_reserve(struct svc_rqst *rqstp, int space)
261 space += rqstp->rq_res.head[0].iov_len;
263 if (space < rqstp->rq_reserved) {
264 struct svc_sock *svsk = rqstp->rq_sock;
265 spin_lock_bh(&svsk->sk_server->sv_lock);
266 svsk->sk_reserved -= (rqstp->rq_reserved - space);
267 rqstp->rq_reserved = space;
268 spin_unlock_bh(&svsk->sk_server->sv_lock);
270 svc_sock_enqueue(svsk);
275 * Release a socket after use.
278 svc_sock_put(struct svc_sock *svsk)
280 struct svc_serv *serv = svsk->sk_server;
282 spin_lock_bh(&serv->sv_lock);
283 if (!--(svsk->sk_inuse) && test_bit(SK_DEAD, &svsk->sk_flags)) {
284 spin_unlock_bh(&serv->sv_lock);
285 dprintk("svc: releasing dead socket\n");
286 sock_release(svsk->sk_sock);
290 spin_unlock_bh(&serv->sv_lock);
294 svc_sock_release(struct svc_rqst *rqstp)
296 struct svc_sock *svsk = rqstp->rq_sock;
298 svc_release_skb(rqstp);
300 svc_free_allpages(rqstp);
301 rqstp->rq_res.page_len = 0;
302 rqstp->rq_res.page_base = 0;
305 /* Reset response buffer and release
307 * But first, check that enough space was reserved
308 * for the reply, otherwise we have a bug!
310 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
311 printk(KERN_ERR "RPC request reserved %d but used %d\n",
315 rqstp->rq_res.head[0].iov_len = 0;
316 svc_reserve(rqstp, 0);
317 rqstp->rq_sock = NULL;
323 * External function to wake up a server waiting for data
326 svc_wake_up(struct svc_serv *serv)
328 struct svc_rqst *rqstp;
330 spin_lock_bh(&serv->sv_lock);
331 if (!list_empty(&serv->sv_threads)) {
332 rqstp = list_entry(serv->sv_threads.next,
335 dprintk("svc: daemon %p woken up.\n", rqstp);
337 svc_serv_dequeue(serv, rqstp);
338 rqstp->rq_sock = NULL;
340 wake_up(&rqstp->rq_wait);
342 spin_unlock_bh(&serv->sv_lock);
346 * Generic sendto routine
349 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
351 struct svc_sock *svsk = rqstp->rq_sock;
352 struct socket *sock = svsk->sk_sock;
354 char buffer[CMSG_SPACE(sizeof(struct in_pktinfo))];
355 struct cmsghdr *cmh = (struct cmsghdr *)buffer;
356 struct in_pktinfo *pki = (struct in_pktinfo *)CMSG_DATA(cmh);
360 struct page **ppage = xdr->pages;
361 size_t base = xdr->page_base;
362 unsigned int pglen = xdr->page_len;
363 unsigned int flags = MSG_MORE;
367 if (rqstp->rq_prot == IPPROTO_UDP) {
368 /* set the source and destination */
370 msg.msg_name = &rqstp->rq_addr;
371 msg.msg_namelen = sizeof(rqstp->rq_addr);
374 msg.msg_flags = MSG_MORE;
376 msg.msg_control = cmh;
377 msg.msg_controllen = sizeof(buffer);
378 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
379 cmh->cmsg_level = SOL_IP;
380 cmh->cmsg_type = IP_PKTINFO;
381 pki->ipi_ifindex = 0;
382 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr;
384 if (sock_sendmsg(sock, &msg, 0) < 0)
389 if (slen == xdr->head[0].iov_len)
391 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0, xdr->head[0].iov_len, flags);
392 if (len != xdr->head[0].iov_len)
394 slen -= xdr->head[0].iov_len;
399 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
403 result = kernel_sendpage(sock, *ppage, base, size, flags);
410 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
415 if (xdr->tail[0].iov_len) {
416 result = kernel_sendpage(sock, rqstp->rq_respages[rqstp->rq_restailpage],
417 ((unsigned long)xdr->tail[0].iov_base)& (PAGE_SIZE-1),
418 xdr->tail[0].iov_len, 0);
424 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %x)\n",
425 rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len, xdr->len, len,
426 rqstp->rq_addr.sin_addr.s_addr);
432 * Check input queue length
435 svc_recv_available(struct svc_sock *svsk)
437 struct socket *sock = svsk->sk_sock;
440 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
442 return (err >= 0)? avail : err;
446 * Generic recvfrom routine.
449 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
455 rqstp->rq_addrlen = sizeof(rqstp->rq_addr);
456 sock = rqstp->rq_sock->sk_sock;
458 msg.msg_name = &rqstp->rq_addr;
459 msg.msg_namelen = sizeof(rqstp->rq_addr);
460 msg.msg_control = NULL;
461 msg.msg_controllen = 0;
463 msg.msg_flags = MSG_DONTWAIT;
465 len = kernel_recvmsg(sock, &msg, iov, nr, buflen, MSG_DONTWAIT);
467 /* sock_recvmsg doesn't fill in the name/namelen, so we must..
468 * possibly we should cache this in the svc_sock structure
469 * at accept time. FIXME
471 alen = sizeof(rqstp->rq_addr);
472 kernel_getpeername(sock, (struct sockaddr *)&rqstp->rq_addr, &alen);
474 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
475 rqstp->rq_sock, iov[0].iov_base, iov[0].iov_len, len);
481 * Set socket snd and rcv buffer lengths
484 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
488 oldfs = get_fs(); set_fs(KERNEL_DS);
489 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
490 (char*)&snd, sizeof(snd));
491 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
492 (char*)&rcv, sizeof(rcv));
494 /* sock_setsockopt limits use to sysctl_?mem_max,
495 * which isn't acceptable. Until that is made conditional
496 * on not having CAP_SYS_RESOURCE or similar, we go direct...
497 * DaveM said I could!
500 sock->sk->sk_sndbuf = snd * 2;
501 sock->sk->sk_rcvbuf = rcv * 2;
502 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
503 release_sock(sock->sk);
507 * INET callback when data has been received on the socket.
510 svc_udp_data_ready(struct sock *sk, int count)
512 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
515 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
516 svsk, sk, count, test_bit(SK_BUSY, &svsk->sk_flags));
517 set_bit(SK_DATA, &svsk->sk_flags);
518 svc_sock_enqueue(svsk);
520 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
521 wake_up_interruptible(sk->sk_sleep);
525 * INET callback when space is newly available on the socket.
528 svc_write_space(struct sock *sk)
530 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
533 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
534 svsk, sk, test_bit(SK_BUSY, &svsk->sk_flags));
535 svc_sock_enqueue(svsk);
538 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
539 dprintk("RPC svc_write_space: someone sleeping on %p\n",
541 wake_up_interruptible(sk->sk_sleep);
546 * Receive a datagram from a UDP socket.
549 svc_udp_recvfrom(struct svc_rqst *rqstp)
551 struct svc_sock *svsk = rqstp->rq_sock;
552 struct svc_serv *serv = svsk->sk_server;
556 if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
557 /* udp sockets need large rcvbuf as all pending
558 * requests are still in that buffer. sndbuf must
559 * also be large enough that there is enough space
560 * for one reply per thread.
562 svc_sock_setbufsize(svsk->sk_sock,
563 (serv->sv_nrthreads+3) * serv->sv_bufsz,
564 (serv->sv_nrthreads+3) * serv->sv_bufsz);
566 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
567 svc_sock_received(svsk);
568 return svc_deferred_recv(rqstp);
571 clear_bit(SK_DATA, &svsk->sk_flags);
572 while ((skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err)) == NULL) {
573 if (err == -EAGAIN) {
574 svc_sock_received(svsk);
577 /* possibly an icmp error */
578 dprintk("svc: recvfrom returned error %d\n", -err);
580 if (skb->tstamp.off_sec == 0) {
583 tv.tv_sec = xtime.tv_sec;
584 tv.tv_usec = xtime.tv_nsec / NSEC_PER_USEC;
585 skb_set_timestamp(skb, &tv);
586 /* Don't enable netstamp, sunrpc doesn't
587 need that much accuracy */
589 skb_get_timestamp(skb, &svsk->sk_sk->sk_stamp);
590 set_bit(SK_DATA, &svsk->sk_flags); /* there may be more data... */
593 * Maybe more packets - kick another thread ASAP.
595 svc_sock_received(svsk);
597 len = skb->len - sizeof(struct udphdr);
598 rqstp->rq_arg.len = len;
600 rqstp->rq_prot = IPPROTO_UDP;
602 /* Get sender address */
603 rqstp->rq_addr.sin_family = AF_INET;
604 rqstp->rq_addr.sin_port = skb->h.uh->source;
605 rqstp->rq_addr.sin_addr.s_addr = skb->nh.iph->saddr;
606 rqstp->rq_daddr = skb->nh.iph->daddr;
608 if (skb_is_nonlinear(skb)) {
609 /* we have to copy */
611 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
614 skb_free_datagram(svsk->sk_sk, skb);
618 skb_free_datagram(svsk->sk_sk, skb);
620 /* we can use it in-place */
621 rqstp->rq_arg.head[0].iov_base = skb->data + sizeof(struct udphdr);
622 rqstp->rq_arg.head[0].iov_len = len;
623 if (skb_checksum_complete(skb)) {
624 skb_free_datagram(svsk->sk_sk, skb);
627 rqstp->rq_skbuff = skb;
630 rqstp->rq_arg.page_base = 0;
631 if (len <= rqstp->rq_arg.head[0].iov_len) {
632 rqstp->rq_arg.head[0].iov_len = len;
633 rqstp->rq_arg.page_len = 0;
635 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
636 rqstp->rq_argused += (rqstp->rq_arg.page_len + PAGE_SIZE - 1)/ PAGE_SIZE;
640 serv->sv_stats->netudpcnt++;
646 svc_udp_sendto(struct svc_rqst *rqstp)
650 error = svc_sendto(rqstp, &rqstp->rq_res);
651 if (error == -ECONNREFUSED)
652 /* ICMP error on earlier request. */
653 error = svc_sendto(rqstp, &rqstp->rq_res);
659 svc_udp_init(struct svc_sock *svsk)
661 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
662 svsk->sk_sk->sk_write_space = svc_write_space;
663 svsk->sk_recvfrom = svc_udp_recvfrom;
664 svsk->sk_sendto = svc_udp_sendto;
666 /* initialise setting must have enough space to
667 * receive and respond to one request.
668 * svc_udp_recvfrom will re-adjust if necessary
670 svc_sock_setbufsize(svsk->sk_sock,
671 3 * svsk->sk_server->sv_bufsz,
672 3 * svsk->sk_server->sv_bufsz);
674 set_bit(SK_DATA, &svsk->sk_flags); /* might have come in before data_ready set up */
675 set_bit(SK_CHNGBUF, &svsk->sk_flags);
679 * A data_ready event on a listening socket means there's a connection
680 * pending. Do not use state_change as a substitute for it.
683 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
685 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
687 dprintk("svc: socket %p TCP (listen) state change %d\n",
691 * This callback may called twice when a new connection
692 * is established as a child socket inherits everything
693 * from a parent LISTEN socket.
694 * 1) data_ready method of the parent socket will be called
695 * when one of child sockets become ESTABLISHED.
696 * 2) data_ready method of the child socket may be called
697 * when it receives data before the socket is accepted.
698 * In case of 2, we should ignore it silently.
700 if (sk->sk_state == TCP_LISTEN) {
702 set_bit(SK_CONN, &svsk->sk_flags);
703 svc_sock_enqueue(svsk);
705 printk("svc: socket %p: no user data\n", sk);
708 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
709 wake_up_interruptible_all(sk->sk_sleep);
713 * A state change on a connected socket means it's dying or dead.
716 svc_tcp_state_change(struct sock *sk)
718 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
720 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
721 sk, sk->sk_state, sk->sk_user_data);
724 printk("svc: socket %p: no user data\n", sk);
726 set_bit(SK_CLOSE, &svsk->sk_flags);
727 svc_sock_enqueue(svsk);
729 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
730 wake_up_interruptible_all(sk->sk_sleep);
734 svc_tcp_data_ready(struct sock *sk, int count)
736 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
738 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
739 sk, sk->sk_user_data);
741 set_bit(SK_DATA, &svsk->sk_flags);
742 svc_sock_enqueue(svsk);
744 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
745 wake_up_interruptible(sk->sk_sleep);
749 * Accept a TCP connection
752 svc_tcp_accept(struct svc_sock *svsk)
754 struct sockaddr_in sin;
755 struct svc_serv *serv = svsk->sk_server;
756 struct socket *sock = svsk->sk_sock;
757 struct socket *newsock;
758 struct svc_sock *newsvsk;
761 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
765 clear_bit(SK_CONN, &svsk->sk_flags);
766 err = kernel_accept(sock, &newsock, O_NONBLOCK);
769 printk(KERN_WARNING "%s: no more sockets!\n",
771 else if (err != -EAGAIN && net_ratelimit())
772 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
773 serv->sv_name, -err);
777 set_bit(SK_CONN, &svsk->sk_flags);
778 svc_sock_enqueue(svsk);
781 err = kernel_getpeername(newsock, (struct sockaddr *) &sin, &slen);
784 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
785 serv->sv_name, -err);
786 goto failed; /* aborted connection or whatever */
789 /* Ideally, we would want to reject connections from unauthorized
790 * hosts here, but when we get encription, the IP of the host won't
791 * tell us anything. For now just warn about unpriv connections.
793 if (ntohs(sin.sin_port) >= 1024) {
795 "%s: connect from unprivileged port: %u.%u.%u.%u:%d\n",
797 NIPQUAD(sin.sin_addr.s_addr), ntohs(sin.sin_port));
800 dprintk("%s: connect from %u.%u.%u.%u:%04x\n", serv->sv_name,
801 NIPQUAD(sin.sin_addr.s_addr), ntohs(sin.sin_port));
803 /* make sure that a write doesn't block forever when
806 newsock->sk->sk_sndtimeo = HZ*30;
808 if (!(newsvsk = svc_setup_socket(serv, newsock, &err, 0)))
812 /* make sure that we don't have too many active connections.
813 * If we have, something must be dropped.
815 * There's no point in trying to do random drop here for
816 * DoS prevention. The NFS clients does 1 reconnect in 15
817 * seconds. An attacker can easily beat that.
819 * The only somewhat efficient mechanism would be if drop
820 * old connections from the same IP first. But right now
821 * we don't even record the client IP in svc_sock.
823 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
824 struct svc_sock *svsk = NULL;
825 spin_lock_bh(&serv->sv_lock);
826 if (!list_empty(&serv->sv_tempsocks)) {
827 if (net_ratelimit()) {
828 /* Try to help the admin */
829 printk(KERN_NOTICE "%s: too many open TCP "
830 "sockets, consider increasing the "
831 "number of nfsd threads\n",
833 printk(KERN_NOTICE "%s: last TCP connect from "
836 NIPQUAD(sin.sin_addr.s_addr),
837 ntohs(sin.sin_port));
840 * Always select the oldest socket. It's not fair,
843 svsk = list_entry(serv->sv_tempsocks.prev,
846 set_bit(SK_CLOSE, &svsk->sk_flags);
849 spin_unlock_bh(&serv->sv_lock);
852 svc_sock_enqueue(svsk);
859 serv->sv_stats->nettcpconn++;
864 sock_release(newsock);
869 * Receive data from a TCP socket.
872 svc_tcp_recvfrom(struct svc_rqst *rqstp)
874 struct svc_sock *svsk = rqstp->rq_sock;
875 struct svc_serv *serv = svsk->sk_server;
877 struct kvec vec[RPCSVC_MAXPAGES];
880 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
881 svsk, test_bit(SK_DATA, &svsk->sk_flags),
882 test_bit(SK_CONN, &svsk->sk_flags),
883 test_bit(SK_CLOSE, &svsk->sk_flags));
885 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
886 svc_sock_received(svsk);
887 return svc_deferred_recv(rqstp);
890 if (test_bit(SK_CLOSE, &svsk->sk_flags)) {
891 svc_delete_socket(svsk);
895 if (test_bit(SK_CONN, &svsk->sk_flags)) {
896 svc_tcp_accept(svsk);
897 svc_sock_received(svsk);
901 if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
902 /* sndbuf needs to have room for one request
903 * per thread, otherwise we can stall even when the
904 * network isn't a bottleneck.
905 * rcvbuf just needs to be able to hold a few requests.
906 * Normally they will be removed from the queue
907 * as soon a a complete request arrives.
909 svc_sock_setbufsize(svsk->sk_sock,
910 (serv->sv_nrthreads+3) * serv->sv_bufsz,
913 clear_bit(SK_DATA, &svsk->sk_flags);
915 /* Receive data. If we haven't got the record length yet, get
916 * the next four bytes. Otherwise try to gobble up as much as
917 * possible up to the complete record length.
919 if (svsk->sk_tcplen < 4) {
920 unsigned long want = 4 - svsk->sk_tcplen;
923 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
925 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
927 svsk->sk_tcplen += len;
930 dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
932 svc_sock_received(svsk);
933 return -EAGAIN; /* record header not complete */
936 svsk->sk_reclen = ntohl(svsk->sk_reclen);
937 if (!(svsk->sk_reclen & 0x80000000)) {
938 /* FIXME: technically, a record can be fragmented,
939 * and non-terminal fragments will not have the top
940 * bit set in the fragment length header.
941 * But apparently no known nfs clients send fragmented
943 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx (non-terminal)\n",
944 (unsigned long) svsk->sk_reclen);
947 svsk->sk_reclen &= 0x7fffffff;
948 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
949 if (svsk->sk_reclen > serv->sv_bufsz) {
950 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx (large)\n",
951 (unsigned long) svsk->sk_reclen);
956 /* Check whether enough data is available */
957 len = svc_recv_available(svsk);
961 if (len < svsk->sk_reclen) {
962 dprintk("svc: incomplete TCP record (%d of %d)\n",
963 len, svsk->sk_reclen);
964 svc_sock_received(svsk);
965 return -EAGAIN; /* record not complete */
967 len = svsk->sk_reclen;
968 set_bit(SK_DATA, &svsk->sk_flags);
970 vec[0] = rqstp->rq_arg.head[0];
974 vec[pnum].iov_base = page_address(rqstp->rq_argpages[rqstp->rq_argused++]);
975 vec[pnum].iov_len = PAGE_SIZE;
980 /* Now receive data */
981 len = svc_recvfrom(rqstp, vec, pnum, len);
985 dprintk("svc: TCP complete record (%d bytes)\n", len);
986 rqstp->rq_arg.len = len;
987 rqstp->rq_arg.page_base = 0;
988 if (len <= rqstp->rq_arg.head[0].iov_len) {
989 rqstp->rq_arg.head[0].iov_len = len;
990 rqstp->rq_arg.page_len = 0;
992 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
995 rqstp->rq_skbuff = NULL;
996 rqstp->rq_prot = IPPROTO_TCP;
998 /* Reset TCP read info */
1000 svsk->sk_tcplen = 0;
1002 svc_sock_received(svsk);
1004 serv->sv_stats->nettcpcnt++;
1009 svc_delete_socket(svsk);
1013 if (len == -EAGAIN) {
1014 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1015 svc_sock_received(svsk);
1017 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1018 svsk->sk_server->sv_name, -len);
1026 * Send out data on TCP socket.
1029 svc_tcp_sendto(struct svc_rqst *rqstp)
1031 struct xdr_buf *xbufp = &rqstp->rq_res;
1035 /* Set up the first element of the reply kvec.
1036 * Any other kvecs that may be in use have been taken
1037 * care of by the server implementation itself.
1039 reclen = htonl(0x80000000|((xbufp->len ) - 4));
1040 memcpy(xbufp->head[0].iov_base, &reclen, 4);
1042 if (test_bit(SK_DEAD, &rqstp->rq_sock->sk_flags))
1045 sent = svc_sendto(rqstp, &rqstp->rq_res);
1046 if (sent != xbufp->len) {
1047 printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1048 rqstp->rq_sock->sk_server->sv_name,
1049 (sent<0)?"got error":"sent only",
1051 svc_delete_socket(rqstp->rq_sock);
1058 svc_tcp_init(struct svc_sock *svsk)
1060 struct sock *sk = svsk->sk_sk;
1061 struct tcp_sock *tp = tcp_sk(sk);
1063 svsk->sk_recvfrom = svc_tcp_recvfrom;
1064 svsk->sk_sendto = svc_tcp_sendto;
1066 if (sk->sk_state == TCP_LISTEN) {
1067 dprintk("setting up TCP socket for listening\n");
1068 sk->sk_data_ready = svc_tcp_listen_data_ready;
1069 set_bit(SK_CONN, &svsk->sk_flags);
1071 dprintk("setting up TCP socket for reading\n");
1072 sk->sk_state_change = svc_tcp_state_change;
1073 sk->sk_data_ready = svc_tcp_data_ready;
1074 sk->sk_write_space = svc_write_space;
1076 svsk->sk_reclen = 0;
1077 svsk->sk_tcplen = 0;
1079 tp->nonagle = 1; /* disable Nagle's algorithm */
1081 /* initialise setting must have enough space to
1082 * receive and respond to one request.
1083 * svc_tcp_recvfrom will re-adjust if necessary
1085 svc_sock_setbufsize(svsk->sk_sock,
1086 3 * svsk->sk_server->sv_bufsz,
1087 3 * svsk->sk_server->sv_bufsz);
1089 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1090 set_bit(SK_DATA, &svsk->sk_flags);
1091 if (sk->sk_state != TCP_ESTABLISHED)
1092 set_bit(SK_CLOSE, &svsk->sk_flags);
1097 svc_sock_update_bufs(struct svc_serv *serv)
1100 * The number of server threads has changed. Update
1101 * rcvbuf and sndbuf accordingly on all sockets
1103 struct list_head *le;
1105 spin_lock_bh(&serv->sv_lock);
1106 list_for_each(le, &serv->sv_permsocks) {
1107 struct svc_sock *svsk =
1108 list_entry(le, struct svc_sock, sk_list);
1109 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1111 list_for_each(le, &serv->sv_tempsocks) {
1112 struct svc_sock *svsk =
1113 list_entry(le, struct svc_sock, sk_list);
1114 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1116 spin_unlock_bh(&serv->sv_lock);
1120 * Receive the next request on any socket.
1123 svc_recv(struct svc_serv *serv, struct svc_rqst *rqstp, long timeout)
1125 struct svc_sock *svsk =NULL;
1128 struct xdr_buf *arg;
1129 DECLARE_WAITQUEUE(wait, current);
1131 dprintk("svc: server %p waiting for data (to = %ld)\n",
1136 "svc_recv: service %p, socket not NULL!\n",
1138 if (waitqueue_active(&rqstp->rq_wait))
1140 "svc_recv: service %p, wait queue active!\n",
1143 /* Initialize the buffers */
1144 /* first reclaim pages that were moved to response list */
1145 svc_pushback_allpages(rqstp);
1147 /* now allocate needed pages. If we get a failure, sleep briefly */
1148 pages = 2 + (serv->sv_bufsz + PAGE_SIZE -1) / PAGE_SIZE;
1149 while (rqstp->rq_arghi < pages) {
1150 struct page *p = alloc_page(GFP_KERNEL);
1152 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1155 rqstp->rq_argpages[rqstp->rq_arghi++] = p;
1158 /* Make arg->head point to first page and arg->pages point to rest */
1159 arg = &rqstp->rq_arg;
1160 arg->head[0].iov_base = page_address(rqstp->rq_argpages[0]);
1161 arg->head[0].iov_len = PAGE_SIZE;
1162 rqstp->rq_argused = 1;
1163 arg->pages = rqstp->rq_argpages + 1;
1165 /* save at least one page for response */
1166 arg->page_len = (pages-2)*PAGE_SIZE;
1167 arg->len = (pages-1)*PAGE_SIZE;
1168 arg->tail[0].iov_len = 0;
1175 spin_lock_bh(&serv->sv_lock);
1176 if (!list_empty(&serv->sv_tempsocks)) {
1177 svsk = list_entry(serv->sv_tempsocks.next,
1178 struct svc_sock, sk_list);
1179 /* apparently the "standard" is that clients close
1180 * idle connections after 5 minutes, servers after
1182 * http://www.connectathon.org/talks96/nfstcp.pdf
1184 if (get_seconds() - svsk->sk_lastrecv < 6*60
1185 || test_bit(SK_BUSY, &svsk->sk_flags))
1189 set_bit(SK_BUSY, &svsk->sk_flags);
1190 set_bit(SK_CLOSE, &svsk->sk_flags);
1191 rqstp->rq_sock = svsk;
1193 } else if ((svsk = svc_sock_dequeue(serv)) != NULL) {
1194 rqstp->rq_sock = svsk;
1196 rqstp->rq_reserved = serv->sv_bufsz;
1197 svsk->sk_reserved += rqstp->rq_reserved;
1199 /* No data pending. Go to sleep */
1200 svc_serv_enqueue(serv, rqstp);
1203 * We have to be able to interrupt this wait
1204 * to bring down the daemons ...
1206 set_current_state(TASK_INTERRUPTIBLE);
1207 add_wait_queue(&rqstp->rq_wait, &wait);
1208 spin_unlock_bh(&serv->sv_lock);
1210 schedule_timeout(timeout);
1214 spin_lock_bh(&serv->sv_lock);
1215 remove_wait_queue(&rqstp->rq_wait, &wait);
1217 if (!(svsk = rqstp->rq_sock)) {
1218 svc_serv_dequeue(serv, rqstp);
1219 spin_unlock_bh(&serv->sv_lock);
1220 dprintk("svc: server %p, no data yet\n", rqstp);
1221 return signalled()? -EINTR : -EAGAIN;
1224 spin_unlock_bh(&serv->sv_lock);
1226 dprintk("svc: server %p, socket %p, inuse=%d\n",
1227 rqstp, svsk, svsk->sk_inuse);
1228 len = svsk->sk_recvfrom(rqstp);
1229 dprintk("svc: got len=%d\n", len);
1231 /* No data, incomplete (TCP) read, or accept() */
1232 if (len == 0 || len == -EAGAIN) {
1233 rqstp->rq_res.len = 0;
1234 svc_sock_release(rqstp);
1237 svsk->sk_lastrecv = get_seconds();
1238 if (test_bit(SK_TEMP, &svsk->sk_flags)) {
1239 /* push active sockets to end of list */
1240 spin_lock_bh(&serv->sv_lock);
1241 if (!list_empty(&svsk->sk_list))
1242 list_move_tail(&svsk->sk_list, &serv->sv_tempsocks);
1243 spin_unlock_bh(&serv->sv_lock);
1246 rqstp->rq_secure = ntohs(rqstp->rq_addr.sin_port) < 1024;
1247 rqstp->rq_chandle.defer = svc_defer;
1250 serv->sv_stats->netcnt++;
1258 svc_drop(struct svc_rqst *rqstp)
1260 dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1261 svc_sock_release(rqstp);
1265 * Return reply to client.
1268 svc_send(struct svc_rqst *rqstp)
1270 struct svc_sock *svsk;
1274 if ((svsk = rqstp->rq_sock) == NULL) {
1275 printk(KERN_WARNING "NULL socket pointer in %s:%d\n",
1276 __FILE__, __LINE__);
1280 /* release the receive skb before sending the reply */
1281 svc_release_skb(rqstp);
1283 /* calculate over-all length */
1284 xb = & rqstp->rq_res;
1285 xb->len = xb->head[0].iov_len +
1287 xb->tail[0].iov_len;
1289 /* Grab svsk->sk_mutex to serialize outgoing data. */
1290 mutex_lock(&svsk->sk_mutex);
1291 if (test_bit(SK_DEAD, &svsk->sk_flags))
1294 len = svsk->sk_sendto(rqstp);
1295 mutex_unlock(&svsk->sk_mutex);
1296 svc_sock_release(rqstp);
1298 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1304 * Initialize socket for RPC use and create svc_sock struct
1305 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1307 static struct svc_sock *
1308 svc_setup_socket(struct svc_serv *serv, struct socket *sock,
1309 int *errp, int pmap_register)
1311 struct svc_sock *svsk;
1314 dprintk("svc: svc_setup_socket %p\n", sock);
1315 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1322 /* Register socket with portmapper */
1323 if (*errp >= 0 && pmap_register)
1324 *errp = svc_register(serv, inet->sk_protocol,
1325 ntohs(inet_sk(inet)->sport));
1332 set_bit(SK_BUSY, &svsk->sk_flags);
1333 inet->sk_user_data = svsk;
1334 svsk->sk_sock = sock;
1336 svsk->sk_ostate = inet->sk_state_change;
1337 svsk->sk_odata = inet->sk_data_ready;
1338 svsk->sk_owspace = inet->sk_write_space;
1339 svsk->sk_server = serv;
1340 svsk->sk_lastrecv = get_seconds();
1341 INIT_LIST_HEAD(&svsk->sk_deferred);
1342 INIT_LIST_HEAD(&svsk->sk_ready);
1343 mutex_init(&svsk->sk_mutex);
1345 /* Initialize the socket */
1346 if (sock->type == SOCK_DGRAM)
1351 spin_lock_bh(&serv->sv_lock);
1352 if (!pmap_register) {
1353 set_bit(SK_TEMP, &svsk->sk_flags);
1354 list_add(&svsk->sk_list, &serv->sv_tempsocks);
1357 clear_bit(SK_TEMP, &svsk->sk_flags);
1358 list_add(&svsk->sk_list, &serv->sv_permsocks);
1360 spin_unlock_bh(&serv->sv_lock);
1362 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1365 clear_bit(SK_BUSY, &svsk->sk_flags);
1366 svc_sock_enqueue(svsk);
1371 * Create socket for RPC service.
1374 svc_create_socket(struct svc_serv *serv, int protocol, struct sockaddr_in *sin)
1376 struct svc_sock *svsk;
1377 struct socket *sock;
1381 dprintk("svc: svc_create_socket(%s, %d, %u.%u.%u.%u:%d)\n",
1382 serv->sv_program->pg_name, protocol,
1383 NIPQUAD(sin->sin_addr.s_addr),
1384 ntohs(sin->sin_port));
1386 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1387 printk(KERN_WARNING "svc: only UDP and TCP "
1388 "sockets supported\n");
1391 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1393 if ((error = sock_create_kern(PF_INET, type, protocol, &sock)) < 0)
1397 if (type == SOCK_STREAM)
1398 sock->sk->sk_reuse = 1; /* allow address reuse */
1399 error = kernel_bind(sock, (struct sockaddr *) sin,
1405 if (protocol == IPPROTO_TCP) {
1406 if ((error = kernel_listen(sock, 64)) < 0)
1410 if ((svsk = svc_setup_socket(serv, sock, &error, 1)) != NULL)
1414 dprintk("svc: svc_create_socket error = %d\n", -error);
1420 * Remove a dead socket
1423 svc_delete_socket(struct svc_sock *svsk)
1425 struct svc_serv *serv;
1428 dprintk("svc: svc_delete_socket(%p)\n", svsk);
1430 serv = svsk->sk_server;
1433 sk->sk_state_change = svsk->sk_ostate;
1434 sk->sk_data_ready = svsk->sk_odata;
1435 sk->sk_write_space = svsk->sk_owspace;
1437 spin_lock_bh(&serv->sv_lock);
1439 list_del_init(&svsk->sk_list);
1440 list_del_init(&svsk->sk_ready);
1441 if (!test_and_set_bit(SK_DEAD, &svsk->sk_flags))
1442 if (test_bit(SK_TEMP, &svsk->sk_flags))
1445 if (!svsk->sk_inuse) {
1446 spin_unlock_bh(&serv->sv_lock);
1447 sock_release(svsk->sk_sock);
1450 spin_unlock_bh(&serv->sv_lock);
1451 dprintk(KERN_NOTICE "svc: server socket destroy delayed\n");
1452 /* svsk->sk_server = NULL; */
1457 * Make a socket for nfsd and lockd
1460 svc_makesock(struct svc_serv *serv, int protocol, unsigned short port)
1462 struct sockaddr_in sin;
1464 dprintk("svc: creating socket proto = %d\n", protocol);
1465 sin.sin_family = AF_INET;
1466 sin.sin_addr.s_addr = INADDR_ANY;
1467 sin.sin_port = htons(port);
1468 return svc_create_socket(serv, protocol, &sin);
1472 * Handle defer and revisit of requests
1475 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1477 struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1478 struct svc_serv *serv = dreq->owner;
1479 struct svc_sock *svsk;
1482 svc_sock_put(dr->svsk);
1486 dprintk("revisit queued\n");
1489 spin_lock_bh(&serv->sv_lock);
1490 list_add(&dr->handle.recent, &svsk->sk_deferred);
1491 spin_unlock_bh(&serv->sv_lock);
1492 set_bit(SK_DEFERRED, &svsk->sk_flags);
1493 svc_sock_enqueue(svsk);
1497 static struct cache_deferred_req *
1498 svc_defer(struct cache_req *req)
1500 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1501 int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
1502 struct svc_deferred_req *dr;
1504 if (rqstp->rq_arg.page_len)
1505 return NULL; /* if more than a page, give up FIXME */
1506 if (rqstp->rq_deferred) {
1507 dr = rqstp->rq_deferred;
1508 rqstp->rq_deferred = NULL;
1510 int skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1511 /* FIXME maybe discard if size too large */
1512 dr = kmalloc(size, GFP_KERNEL);
1516 dr->handle.owner = rqstp->rq_server;
1517 dr->prot = rqstp->rq_prot;
1518 dr->addr = rqstp->rq_addr;
1519 dr->daddr = rqstp->rq_daddr;
1520 dr->argslen = rqstp->rq_arg.len >> 2;
1521 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
1523 spin_lock_bh(&rqstp->rq_server->sv_lock);
1524 rqstp->rq_sock->sk_inuse++;
1525 dr->svsk = rqstp->rq_sock;
1526 spin_unlock_bh(&rqstp->rq_server->sv_lock);
1528 dr->handle.revisit = svc_revisit;
1533 * recv data from a deferred request into an active one
1535 static int svc_deferred_recv(struct svc_rqst *rqstp)
1537 struct svc_deferred_req *dr = rqstp->rq_deferred;
1539 rqstp->rq_arg.head[0].iov_base = dr->args;
1540 rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
1541 rqstp->rq_arg.page_len = 0;
1542 rqstp->rq_arg.len = dr->argslen<<2;
1543 rqstp->rq_prot = dr->prot;
1544 rqstp->rq_addr = dr->addr;
1545 rqstp->rq_daddr = dr->daddr;
1546 return dr->argslen<<2;
1550 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk)
1552 struct svc_deferred_req *dr = NULL;
1553 struct svc_serv *serv = svsk->sk_server;
1555 if (!test_bit(SK_DEFERRED, &svsk->sk_flags))
1557 spin_lock_bh(&serv->sv_lock);
1558 clear_bit(SK_DEFERRED, &svsk->sk_flags);
1559 if (!list_empty(&svsk->sk_deferred)) {
1560 dr = list_entry(svsk->sk_deferred.next,
1561 struct svc_deferred_req,
1563 list_del_init(&dr->handle.recent);
1564 set_bit(SK_DEFERRED, &svsk->sk_flags);
1566 spin_unlock_bh(&serv->sv_lock);