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_xprt_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/kernel.h>
23 #include <linux/sched.h>
24 #include <linux/errno.h>
25 #include <linux/fcntl.h>
26 #include <linux/net.h>
28 #include <linux/inet.h>
29 #include <linux/udp.h>
30 #include <linux/tcp.h>
31 #include <linux/unistd.h>
32 #include <linux/slab.h>
33 #include <linux/netdevice.h>
34 #include <linux/skbuff.h>
35 #include <linux/file.h>
36 #include <linux/freezer.h>
38 #include <net/checksum.h>
42 #include <net/tcp_states.h>
43 #include <asm/uaccess.h>
44 #include <asm/ioctls.h>
46 #include <linux/sunrpc/types.h>
47 #include <linux/sunrpc/clnt.h>
48 #include <linux/sunrpc/xdr.h>
49 #include <linux/sunrpc/msg_prot.h>
50 #include <linux/sunrpc/svcsock.h>
51 #include <linux/sunrpc/stats.h>
53 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
56 static struct svc_sock *svc_setup_socket(struct svc_serv *, struct socket *,
57 int *errp, int flags);
58 static void svc_udp_data_ready(struct sock *, int);
59 static int svc_udp_recvfrom(struct svc_rqst *);
60 static int svc_udp_sendto(struct svc_rqst *);
61 static void svc_sock_detach(struct svc_xprt *);
62 static void svc_tcp_sock_detach(struct svc_xprt *);
63 static void svc_sock_free(struct svc_xprt *);
65 static struct svc_xprt *svc_create_socket(struct svc_serv *, int,
66 struct sockaddr *, int, int);
67 #ifdef CONFIG_DEBUG_LOCK_ALLOC
68 static struct lock_class_key svc_key[2];
69 static struct lock_class_key svc_slock_key[2];
71 static void svc_reclassify_socket(struct socket *sock)
73 struct sock *sk = sock->sk;
74 BUG_ON(sock_owned_by_user(sk));
75 switch (sk->sk_family) {
77 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFSD",
79 "sk_xprt.xpt_lock-AF_INET-NFSD",
84 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFSD",
86 "sk_xprt.xpt_lock-AF_INET6-NFSD",
95 static void svc_reclassify_socket(struct socket *sock)
101 * Release an skbuff after use
103 static void svc_release_skb(struct svc_rqst *rqstp)
105 struct sk_buff *skb = rqstp->rq_xprt_ctxt;
108 struct svc_sock *svsk =
109 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
110 rqstp->rq_xprt_ctxt = NULL;
112 dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
113 skb_free_datagram(svsk->sk_sk, skb);
117 union svc_pktinfo_u {
118 struct in_pktinfo pkti;
119 struct in6_pktinfo pkti6;
121 #define SVC_PKTINFO_SPACE \
122 CMSG_SPACE(sizeof(union svc_pktinfo_u))
124 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
126 struct svc_sock *svsk =
127 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
128 switch (svsk->sk_sk->sk_family) {
130 struct in_pktinfo *pki = CMSG_DATA(cmh);
132 cmh->cmsg_level = SOL_IP;
133 cmh->cmsg_type = IP_PKTINFO;
134 pki->ipi_ifindex = 0;
135 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr.addr.s_addr;
136 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
141 struct in6_pktinfo *pki = CMSG_DATA(cmh);
143 cmh->cmsg_level = SOL_IPV6;
144 cmh->cmsg_type = IPV6_PKTINFO;
145 pki->ipi6_ifindex = 0;
146 ipv6_addr_copy(&pki->ipi6_addr,
147 &rqstp->rq_daddr.addr6);
148 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
156 * Generic sendto routine
158 static int svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
160 struct svc_sock *svsk =
161 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
162 struct socket *sock = svsk->sk_sock;
166 long all[SVC_PKTINFO_SPACE / sizeof(long)];
168 struct cmsghdr *cmh = &buffer.hdr;
172 struct page **ppage = xdr->pages;
173 size_t base = xdr->page_base;
174 unsigned int pglen = xdr->page_len;
175 unsigned int flags = MSG_MORE;
176 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
180 if (rqstp->rq_prot == IPPROTO_UDP) {
181 struct msghdr msg = {
182 .msg_name = &rqstp->rq_addr,
183 .msg_namelen = rqstp->rq_addrlen,
185 .msg_controllen = sizeof(buffer),
186 .msg_flags = MSG_MORE,
189 svc_set_cmsg_data(rqstp, cmh);
191 if (sock_sendmsg(sock, &msg, 0) < 0)
196 if (slen == xdr->head[0].iov_len)
198 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0,
199 xdr->head[0].iov_len, flags);
200 if (len != xdr->head[0].iov_len)
202 slen -= xdr->head[0].iov_len;
207 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
211 result = kernel_sendpage(sock, *ppage, base, size, flags);
218 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
223 if (xdr->tail[0].iov_len) {
224 result = kernel_sendpage(sock, rqstp->rq_respages[0],
225 ((unsigned long)xdr->tail[0].iov_base)
227 xdr->tail[0].iov_len, 0);
233 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
234 svsk, xdr->head[0].iov_base, xdr->head[0].iov_len,
235 xdr->len, len, svc_print_addr(rqstp, buf, sizeof(buf)));
241 * Report socket names for nfsdfs
243 static int svc_one_sock_name(struct svc_sock *svsk, char *buf, int remaining)
245 const struct sock *sk = svsk->sk_sk;
246 const char *proto_name = sk->sk_protocol == IPPROTO_UDP ?
250 switch (sk->sk_family) {
252 len = snprintf(buf, remaining, "ipv4 %s %pI4 %d\n",
254 &inet_sk(sk)->rcv_saddr,
258 len = snprintf(buf, remaining, "ipv6 %s %pI6 %d\n",
260 &inet6_sk(sk)->rcv_saddr,
264 len = snprintf(buf, remaining, "*unknown-%d*\n",
268 if (len >= remaining) {
270 return -ENAMETOOLONG;
276 * svc_sock_names - construct a list of listener names in a string
277 * @serv: pointer to RPC service
278 * @buf: pointer to a buffer to fill in with socket names
279 * @buflen: size of the buffer to be filled
280 * @toclose: pointer to '\0'-terminated C string containing the name
281 * of a listener to be closed
283 * Fills in @buf with a '\n'-separated list of names of listener
284 * sockets. If @toclose is not NULL, the socket named by @toclose
285 * is closed, and is not included in the output list.
287 * Returns positive length of the socket name string, or a negative
288 * errno value on error.
290 int svc_sock_names(struct svc_serv *serv, char *buf, const size_t buflen,
293 struct svc_sock *svsk, *closesk = NULL;
299 spin_lock_bh(&serv->sv_lock);
300 list_for_each_entry(svsk, &serv->sv_permsocks, sk_xprt.xpt_list) {
301 int onelen = svc_one_sock_name(svsk, buf + len, buflen - len);
306 if (toclose && strcmp(toclose, buf + len) == 0)
311 spin_unlock_bh(&serv->sv_lock);
314 /* Should unregister with portmap, but you cannot
315 * unregister just one protocol...
317 svc_close_xprt(&closesk->sk_xprt);
322 EXPORT_SYMBOL_GPL(svc_sock_names);
325 * Check input queue length
327 static int svc_recv_available(struct svc_sock *svsk)
329 struct socket *sock = svsk->sk_sock;
332 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
334 return (err >= 0)? avail : err;
338 * Generic recvfrom routine.
340 static int svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr,
343 struct svc_sock *svsk =
344 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
345 struct msghdr msg = {
346 .msg_flags = MSG_DONTWAIT,
350 rqstp->rq_xprt_hlen = 0;
352 len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
355 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
356 svsk, iov[0].iov_base, iov[0].iov_len, len);
361 * Set socket snd and rcv buffer lengths
363 static void svc_sock_setbufsize(struct socket *sock, unsigned int snd,
368 oldfs = get_fs(); set_fs(KERNEL_DS);
369 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
370 (char*)&snd, sizeof(snd));
371 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
372 (char*)&rcv, sizeof(rcv));
374 /* sock_setsockopt limits use to sysctl_?mem_max,
375 * which isn't acceptable. Until that is made conditional
376 * on not having CAP_SYS_RESOURCE or similar, we go direct...
377 * DaveM said I could!
380 sock->sk->sk_sndbuf = snd * 2;
381 sock->sk->sk_rcvbuf = rcv * 2;
382 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
383 release_sock(sock->sk);
387 * INET callback when data has been received on the socket.
389 static void svc_udp_data_ready(struct sock *sk, int count)
391 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
394 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
396 test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
397 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
398 svc_xprt_enqueue(&svsk->sk_xprt);
400 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
401 wake_up_interruptible(sk->sk_sleep);
405 * INET callback when space is newly available on the socket.
407 static void svc_write_space(struct sock *sk)
409 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
412 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
413 svsk, sk, test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
414 svc_xprt_enqueue(&svsk->sk_xprt);
417 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
418 dprintk("RPC svc_write_space: someone sleeping on %p\n",
420 wake_up_interruptible(sk->sk_sleep);
425 * Copy the UDP datagram's destination address to the rqstp structure.
426 * The 'destination' address in this case is the address to which the
427 * peer sent the datagram, i.e. our local address. For multihomed
428 * hosts, this can change from msg to msg. Note that only the IP
429 * address changes, the port number should remain the same.
431 static void svc_udp_get_dest_address(struct svc_rqst *rqstp,
434 struct svc_sock *svsk =
435 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
436 switch (svsk->sk_sk->sk_family) {
438 struct in_pktinfo *pki = CMSG_DATA(cmh);
439 rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
443 struct in6_pktinfo *pki = CMSG_DATA(cmh);
444 ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
451 * Receive a datagram from a UDP socket.
453 static int svc_udp_recvfrom(struct svc_rqst *rqstp)
455 struct svc_sock *svsk =
456 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
457 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
461 long all[SVC_PKTINFO_SPACE / sizeof(long)];
463 struct cmsghdr *cmh = &buffer.hdr;
464 struct msghdr msg = {
465 .msg_name = svc_addr(rqstp),
467 .msg_controllen = sizeof(buffer),
468 .msg_flags = MSG_DONTWAIT,
473 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
474 /* udp sockets need large rcvbuf as all pending
475 * requests are still in that buffer. sndbuf must
476 * also be large enough that there is enough space
477 * for one reply per thread. We count all threads
478 * rather than threads in a particular pool, which
479 * provides an upper bound on the number of threads
480 * which will access the socket.
482 svc_sock_setbufsize(svsk->sk_sock,
483 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
484 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
486 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
488 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
489 0, 0, MSG_PEEK | MSG_DONTWAIT);
491 skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err);
494 if (err != -EAGAIN) {
495 /* possibly an icmp error */
496 dprintk("svc: recvfrom returned error %d\n", -err);
497 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
499 svc_xprt_received(&svsk->sk_xprt);
502 len = svc_addr_len(svc_addr(rqstp));
504 return -EAFNOSUPPORT;
505 rqstp->rq_addrlen = len;
506 if (skb->tstamp.tv64 == 0) {
507 skb->tstamp = ktime_get_real();
508 /* Don't enable netstamp, sunrpc doesn't
509 need that much accuracy */
511 svsk->sk_sk->sk_stamp = skb->tstamp;
512 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* there may be more data... */
515 * Maybe more packets - kick another thread ASAP.
517 svc_xprt_received(&svsk->sk_xprt);
519 len = skb->len - sizeof(struct udphdr);
520 rqstp->rq_arg.len = len;
522 rqstp->rq_prot = IPPROTO_UDP;
524 if (cmh->cmsg_level != IPPROTO_IP ||
525 cmh->cmsg_type != IP_PKTINFO) {
527 printk("rpcsvc: received unknown control message:"
529 cmh->cmsg_level, cmh->cmsg_type);
530 skb_free_datagram(svsk->sk_sk, skb);
533 svc_udp_get_dest_address(rqstp, cmh);
535 if (skb_is_nonlinear(skb)) {
536 /* we have to copy */
538 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
541 skb_free_datagram(svsk->sk_sk, skb);
545 skb_free_datagram(svsk->sk_sk, skb);
547 /* we can use it in-place */
548 rqstp->rq_arg.head[0].iov_base = skb->data +
549 sizeof(struct udphdr);
550 rqstp->rq_arg.head[0].iov_len = len;
551 if (skb_checksum_complete(skb)) {
552 skb_free_datagram(svsk->sk_sk, skb);
555 rqstp->rq_xprt_ctxt = skb;
558 rqstp->rq_arg.page_base = 0;
559 if (len <= rqstp->rq_arg.head[0].iov_len) {
560 rqstp->rq_arg.head[0].iov_len = len;
561 rqstp->rq_arg.page_len = 0;
562 rqstp->rq_respages = rqstp->rq_pages+1;
564 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
565 rqstp->rq_respages = rqstp->rq_pages + 1 +
566 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
570 serv->sv_stats->netudpcnt++;
576 svc_udp_sendto(struct svc_rqst *rqstp)
580 error = svc_sendto(rqstp, &rqstp->rq_res);
581 if (error == -ECONNREFUSED)
582 /* ICMP error on earlier request. */
583 error = svc_sendto(rqstp, &rqstp->rq_res);
588 static void svc_udp_prep_reply_hdr(struct svc_rqst *rqstp)
592 static int svc_udp_has_wspace(struct svc_xprt *xprt)
594 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
595 struct svc_serv *serv = xprt->xpt_server;
596 unsigned long required;
599 * Set the SOCK_NOSPACE flag before checking the available
602 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
603 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
604 if (required*2 > sock_wspace(svsk->sk_sk))
606 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
610 static struct svc_xprt *svc_udp_accept(struct svc_xprt *xprt)
616 static struct svc_xprt *svc_udp_create(struct svc_serv *serv,
617 struct sockaddr *sa, int salen,
620 return svc_create_socket(serv, IPPROTO_UDP, sa, salen, flags);
623 static struct svc_xprt_ops svc_udp_ops = {
624 .xpo_create = svc_udp_create,
625 .xpo_recvfrom = svc_udp_recvfrom,
626 .xpo_sendto = svc_udp_sendto,
627 .xpo_release_rqst = svc_release_skb,
628 .xpo_detach = svc_sock_detach,
629 .xpo_free = svc_sock_free,
630 .xpo_prep_reply_hdr = svc_udp_prep_reply_hdr,
631 .xpo_has_wspace = svc_udp_has_wspace,
632 .xpo_accept = svc_udp_accept,
635 static struct svc_xprt_class svc_udp_class = {
637 .xcl_owner = THIS_MODULE,
638 .xcl_ops = &svc_udp_ops,
639 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
642 static void svc_udp_init(struct svc_sock *svsk, struct svc_serv *serv)
647 svc_xprt_init(&svc_udp_class, &svsk->sk_xprt, serv);
648 clear_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
649 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
650 svsk->sk_sk->sk_write_space = svc_write_space;
652 /* initialise setting must have enough space to
653 * receive and respond to one request.
654 * svc_udp_recvfrom will re-adjust if necessary
656 svc_sock_setbufsize(svsk->sk_sock,
657 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
658 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
660 /* data might have come in before data_ready set up */
661 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
662 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
666 /* make sure we get destination address info */
667 svsk->sk_sock->ops->setsockopt(svsk->sk_sock, IPPROTO_IP, IP_PKTINFO,
668 (char __user *)&one, sizeof(one));
673 * A data_ready event on a listening socket means there's a connection
674 * pending. Do not use state_change as a substitute for it.
676 static void svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
678 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
680 dprintk("svc: socket %p TCP (listen) state change %d\n",
684 * This callback may called twice when a new connection
685 * is established as a child socket inherits everything
686 * from a parent LISTEN socket.
687 * 1) data_ready method of the parent socket will be called
688 * when one of child sockets become ESTABLISHED.
689 * 2) data_ready method of the child socket may be called
690 * when it receives data before the socket is accepted.
691 * In case of 2, we should ignore it silently.
693 if (sk->sk_state == TCP_LISTEN) {
695 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
696 svc_xprt_enqueue(&svsk->sk_xprt);
698 printk("svc: socket %p: no user data\n", sk);
701 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
702 wake_up_interruptible_all(sk->sk_sleep);
706 * A state change on a connected socket means it's dying or dead.
708 static void svc_tcp_state_change(struct sock *sk)
710 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
712 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
713 sk, sk->sk_state, sk->sk_user_data);
716 printk("svc: socket %p: no user data\n", sk);
718 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
719 svc_xprt_enqueue(&svsk->sk_xprt);
721 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
722 wake_up_interruptible_all(sk->sk_sleep);
725 static void svc_tcp_data_ready(struct sock *sk, int count)
727 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
729 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
730 sk, sk->sk_user_data);
732 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
733 svc_xprt_enqueue(&svsk->sk_xprt);
735 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
736 wake_up_interruptible(sk->sk_sleep);
740 * Accept a TCP connection
742 static struct svc_xprt *svc_tcp_accept(struct svc_xprt *xprt)
744 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
745 struct sockaddr_storage addr;
746 struct sockaddr *sin = (struct sockaddr *) &addr;
747 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
748 struct socket *sock = svsk->sk_sock;
749 struct socket *newsock;
750 struct svc_sock *newsvsk;
752 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
754 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
758 clear_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
759 err = kernel_accept(sock, &newsock, O_NONBLOCK);
762 printk(KERN_WARNING "%s: no more sockets!\n",
764 else if (err != -EAGAIN && net_ratelimit())
765 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
766 serv->sv_name, -err);
769 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
771 err = kernel_getpeername(newsock, sin, &slen);
774 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
775 serv->sv_name, -err);
776 goto failed; /* aborted connection or whatever */
779 /* Ideally, we would want to reject connections from unauthorized
780 * hosts here, but when we get encryption, the IP of the host won't
781 * tell us anything. For now just warn about unpriv connections.
783 if (!svc_port_is_privileged(sin)) {
785 "%s: connect from unprivileged port: %s\n",
787 __svc_print_addr(sin, buf, sizeof(buf)));
789 dprintk("%s: connect from %s\n", serv->sv_name,
790 __svc_print_addr(sin, buf, sizeof(buf)));
792 /* make sure that a write doesn't block forever when
795 newsock->sk->sk_sndtimeo = HZ*30;
797 if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
798 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
800 svc_xprt_set_remote(&newsvsk->sk_xprt, sin, slen);
801 err = kernel_getsockname(newsock, sin, &slen);
802 if (unlikely(err < 0)) {
803 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err);
804 slen = offsetof(struct sockaddr, sa_data);
806 svc_xprt_set_local(&newsvsk->sk_xprt, sin, slen);
809 serv->sv_stats->nettcpconn++;
811 return &newsvsk->sk_xprt;
814 sock_release(newsock);
819 * Receive data from a TCP socket.
821 static int svc_tcp_recvfrom(struct svc_rqst *rqstp)
823 struct svc_sock *svsk =
824 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
825 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
830 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
831 svsk, test_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags),
832 test_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags),
833 test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags));
835 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
836 /* sndbuf needs to have room for one request
837 * per thread, otherwise we can stall even when the
838 * network isn't a bottleneck.
840 * We count all threads rather than threads in a
841 * particular pool, which provides an upper bound
842 * on the number of threads which will access the socket.
844 * rcvbuf just needs to be able to hold a few requests.
845 * Normally they will be removed from the queue
846 * as soon a a complete request arrives.
848 svc_sock_setbufsize(svsk->sk_sock,
849 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
850 3 * serv->sv_max_mesg);
852 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
854 /* Receive data. If we haven't got the record length yet, get
855 * the next four bytes. Otherwise try to gobble up as much as
856 * possible up to the complete record length.
858 if (svsk->sk_tcplen < sizeof(rpc_fraghdr)) {
859 int want = sizeof(rpc_fraghdr) - svsk->sk_tcplen;
862 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
864 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
866 svsk->sk_tcplen += len;
869 dprintk("svc: short recvfrom while reading record "
870 "length (%d of %d)\n", len, want);
871 svc_xprt_received(&svsk->sk_xprt);
872 return -EAGAIN; /* record header not complete */
875 svsk->sk_reclen = ntohl(svsk->sk_reclen);
876 if (!(svsk->sk_reclen & RPC_LAST_STREAM_FRAGMENT)) {
877 /* FIXME: technically, a record can be fragmented,
878 * and non-terminal fragments will not have the top
879 * bit set in the fragment length header.
880 * But apparently no known nfs clients send fragmented
883 printk(KERN_NOTICE "RPC: multiple fragments "
884 "per record not supported\n");
887 svsk->sk_reclen &= RPC_FRAGMENT_SIZE_MASK;
888 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
889 if (svsk->sk_reclen > serv->sv_max_mesg) {
891 printk(KERN_NOTICE "RPC: "
892 "fragment too large: 0x%08lx\n",
893 (unsigned long)svsk->sk_reclen);
898 /* Check whether enough data is available */
899 len = svc_recv_available(svsk);
903 if (len < svsk->sk_reclen) {
904 dprintk("svc: incomplete TCP record (%d of %d)\n",
905 len, svsk->sk_reclen);
906 svc_xprt_received(&svsk->sk_xprt);
907 return -EAGAIN; /* record not complete */
909 len = svsk->sk_reclen;
910 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
913 vec[0] = rqstp->rq_arg.head[0];
917 vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
918 vec[pnum].iov_len = PAGE_SIZE;
922 rqstp->rq_respages = &rqstp->rq_pages[pnum];
924 /* Now receive data */
925 len = svc_recvfrom(rqstp, vec, pnum, len);
929 dprintk("svc: TCP complete record (%d bytes)\n", len);
930 rqstp->rq_arg.len = len;
931 rqstp->rq_arg.page_base = 0;
932 if (len <= rqstp->rq_arg.head[0].iov_len) {
933 rqstp->rq_arg.head[0].iov_len = len;
934 rqstp->rq_arg.page_len = 0;
936 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
939 rqstp->rq_xprt_ctxt = NULL;
940 rqstp->rq_prot = IPPROTO_TCP;
942 /* Reset TCP read info */
946 svc_xprt_copy_addrs(rqstp, &svsk->sk_xprt);
947 svc_xprt_received(&svsk->sk_xprt);
949 serv->sv_stats->nettcpcnt++;
954 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
958 if (len == -EAGAIN) {
959 dprintk("RPC: TCP recvfrom got EAGAIN\n");
960 svc_xprt_received(&svsk->sk_xprt);
962 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
963 svsk->sk_xprt.xpt_server->sv_name, -len);
971 * Send out data on TCP socket.
973 static int svc_tcp_sendto(struct svc_rqst *rqstp)
975 struct xdr_buf *xbufp = &rqstp->rq_res;
979 /* Set up the first element of the reply kvec.
980 * Any other kvecs that may be in use have been taken
981 * care of by the server implementation itself.
983 reclen = htonl(0x80000000|((xbufp->len ) - 4));
984 memcpy(xbufp->head[0].iov_base, &reclen, 4);
986 if (test_bit(XPT_DEAD, &rqstp->rq_xprt->xpt_flags))
989 sent = svc_sendto(rqstp, &rqstp->rq_res);
990 if (sent != xbufp->len) {
992 "rpc-srv/tcp: %s: %s %d when sending %d bytes "
993 "- shutting down socket\n",
994 rqstp->rq_xprt->xpt_server->sv_name,
995 (sent<0)?"got error":"sent only",
997 set_bit(XPT_CLOSE, &rqstp->rq_xprt->xpt_flags);
998 svc_xprt_enqueue(rqstp->rq_xprt);
1005 * Setup response header. TCP has a 4B record length field.
1007 static void svc_tcp_prep_reply_hdr(struct svc_rqst *rqstp)
1009 struct kvec *resv = &rqstp->rq_res.head[0];
1011 /* tcp needs a space for the record length... */
1015 static int svc_tcp_has_wspace(struct svc_xprt *xprt)
1017 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1018 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1023 * Set the SOCK_NOSPACE flag before checking the available
1026 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1027 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
1028 wspace = sk_stream_wspace(svsk->sk_sk);
1030 if (wspace < sk_stream_min_wspace(svsk->sk_sk))
1032 if (required * 2 > wspace)
1035 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1039 static struct svc_xprt *svc_tcp_create(struct svc_serv *serv,
1040 struct sockaddr *sa, int salen,
1043 return svc_create_socket(serv, IPPROTO_TCP, sa, salen, flags);
1046 static struct svc_xprt_ops svc_tcp_ops = {
1047 .xpo_create = svc_tcp_create,
1048 .xpo_recvfrom = svc_tcp_recvfrom,
1049 .xpo_sendto = svc_tcp_sendto,
1050 .xpo_release_rqst = svc_release_skb,
1051 .xpo_detach = svc_tcp_sock_detach,
1052 .xpo_free = svc_sock_free,
1053 .xpo_prep_reply_hdr = svc_tcp_prep_reply_hdr,
1054 .xpo_has_wspace = svc_tcp_has_wspace,
1055 .xpo_accept = svc_tcp_accept,
1058 static struct svc_xprt_class svc_tcp_class = {
1060 .xcl_owner = THIS_MODULE,
1061 .xcl_ops = &svc_tcp_ops,
1062 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1065 void svc_init_xprt_sock(void)
1067 svc_reg_xprt_class(&svc_tcp_class);
1068 svc_reg_xprt_class(&svc_udp_class);
1071 void svc_cleanup_xprt_sock(void)
1073 svc_unreg_xprt_class(&svc_tcp_class);
1074 svc_unreg_xprt_class(&svc_udp_class);
1077 static void svc_tcp_init(struct svc_sock *svsk, struct svc_serv *serv)
1079 struct sock *sk = svsk->sk_sk;
1081 svc_xprt_init(&svc_tcp_class, &svsk->sk_xprt, serv);
1082 set_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
1083 if (sk->sk_state == TCP_LISTEN) {
1084 dprintk("setting up TCP socket for listening\n");
1085 set_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags);
1086 sk->sk_data_ready = svc_tcp_listen_data_ready;
1087 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1089 dprintk("setting up TCP socket for reading\n");
1090 sk->sk_state_change = svc_tcp_state_change;
1091 sk->sk_data_ready = svc_tcp_data_ready;
1092 sk->sk_write_space = svc_write_space;
1094 svsk->sk_reclen = 0;
1095 svsk->sk_tcplen = 0;
1097 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1099 /* initialise setting must have enough space to
1100 * receive and respond to one request.
1101 * svc_tcp_recvfrom will re-adjust if necessary
1103 svc_sock_setbufsize(svsk->sk_sock,
1104 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
1105 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
1107 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1108 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1109 if (sk->sk_state != TCP_ESTABLISHED)
1110 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1114 void 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_xprt.xpt_list);
1126 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1128 list_for_each(le, &serv->sv_tempsocks) {
1129 struct svc_sock *svsk =
1130 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1131 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1133 spin_unlock_bh(&serv->sv_lock);
1135 EXPORT_SYMBOL_GPL(svc_sock_update_bufs);
1138 * Initialize socket for RPC use and create svc_sock struct
1139 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1141 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1142 struct socket *sock,
1143 int *errp, int flags)
1145 struct svc_sock *svsk;
1147 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1149 dprintk("svc: svc_setup_socket %p\n", sock);
1150 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1157 /* Register socket with portmapper */
1158 if (*errp >= 0 && pmap_register)
1159 *errp = svc_register(serv, inet->sk_family, inet->sk_protocol,
1160 ntohs(inet_sk(inet)->sport));
1167 inet->sk_user_data = svsk;
1168 svsk->sk_sock = sock;
1170 svsk->sk_ostate = inet->sk_state_change;
1171 svsk->sk_odata = inet->sk_data_ready;
1172 svsk->sk_owspace = inet->sk_write_space;
1174 /* Initialize the socket */
1175 if (sock->type == SOCK_DGRAM)
1176 svc_udp_init(svsk, serv);
1178 svc_tcp_init(svsk, serv);
1180 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1187 * svc_addsock - add a listener socket to an RPC service
1188 * @serv: pointer to RPC service to which to add a new listener
1189 * @fd: file descriptor of the new listener
1190 * @name_return: pointer to buffer to fill in with name of listener
1191 * @len: size of the buffer
1193 * Fills in socket name and returns positive length of name if successful.
1194 * Name is terminated with '\n'. On error, returns a negative errno
1197 int svc_addsock(struct svc_serv *serv, const int fd, char *name_return,
1201 struct socket *so = sockfd_lookup(fd, &err);
1202 struct svc_sock *svsk = NULL;
1206 if (so->sk->sk_family != AF_INET)
1207 err = -EAFNOSUPPORT;
1208 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1209 so->sk->sk_protocol != IPPROTO_UDP)
1210 err = -EPROTONOSUPPORT;
1211 else if (so->state > SS_UNCONNECTED)
1214 if (!try_module_get(THIS_MODULE))
1217 svsk = svc_setup_socket(serv, so, &err,
1220 struct sockaddr_storage addr;
1221 struct sockaddr *sin = (struct sockaddr *)&addr;
1223 if (kernel_getsockname(svsk->sk_sock, sin, &salen) == 0)
1224 svc_xprt_set_local(&svsk->sk_xprt, sin, salen);
1225 clear_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1226 spin_lock_bh(&serv->sv_lock);
1227 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_permsocks);
1228 spin_unlock_bh(&serv->sv_lock);
1229 svc_xprt_received(&svsk->sk_xprt);
1232 module_put(THIS_MODULE);
1238 return svc_one_sock_name(svsk, name_return, len);
1240 EXPORT_SYMBOL_GPL(svc_addsock);
1243 * Create socket for RPC service.
1245 static struct svc_xprt *svc_create_socket(struct svc_serv *serv,
1247 struct sockaddr *sin, int len,
1250 struct svc_sock *svsk;
1251 struct socket *sock;
1254 struct sockaddr_storage addr;
1255 struct sockaddr *newsin = (struct sockaddr *)&addr;
1259 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
1261 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1262 serv->sv_program->pg_name, protocol,
1263 __svc_print_addr(sin, buf, sizeof(buf)));
1265 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1266 printk(KERN_WARNING "svc: only UDP and TCP "
1267 "sockets supported\n");
1268 return ERR_PTR(-EINVAL);
1271 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1272 switch (sin->sa_family) {
1280 return ERR_PTR(-EINVAL);
1283 error = sock_create_kern(family, type, protocol, &sock);
1285 return ERR_PTR(error);
1287 svc_reclassify_socket(sock);
1290 * If this is an PF_INET6 listener, we want to avoid
1291 * getting requests from IPv4 remotes. Those should
1292 * be shunted to a PF_INET listener via rpcbind.
1295 if (family == PF_INET6)
1296 kernel_setsockopt(sock, SOL_IPV6, IPV6_V6ONLY,
1297 (char *)&val, sizeof(val));
1299 if (type == SOCK_STREAM)
1300 sock->sk->sk_reuse = 1; /* allow address reuse */
1301 error = kernel_bind(sock, sin, len);
1306 error = kernel_getsockname(sock, newsin, &newlen);
1310 if (protocol == IPPROTO_TCP) {
1311 if ((error = kernel_listen(sock, 64)) < 0)
1315 if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1316 svc_xprt_set_local(&svsk->sk_xprt, newsin, newlen);
1317 return (struct svc_xprt *)svsk;
1321 dprintk("svc: svc_create_socket error = %d\n", -error);
1323 return ERR_PTR(error);
1327 * Detach the svc_sock from the socket so that no
1328 * more callbacks occur.
1330 static void svc_sock_detach(struct svc_xprt *xprt)
1332 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1333 struct sock *sk = svsk->sk_sk;
1335 dprintk("svc: svc_sock_detach(%p)\n", svsk);
1337 /* put back the old socket callbacks */
1338 sk->sk_state_change = svsk->sk_ostate;
1339 sk->sk_data_ready = svsk->sk_odata;
1340 sk->sk_write_space = svsk->sk_owspace;
1342 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1343 wake_up_interruptible(sk->sk_sleep);
1347 * Disconnect the socket, and reset the callbacks
1349 static void svc_tcp_sock_detach(struct svc_xprt *xprt)
1351 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1353 dprintk("svc: svc_tcp_sock_detach(%p)\n", svsk);
1355 svc_sock_detach(xprt);
1357 if (!test_bit(XPT_LISTENER, &xprt->xpt_flags))
1358 kernel_sock_shutdown(svsk->sk_sock, SHUT_RDWR);
1362 * Free the svc_sock's socket resources and the svc_sock itself.
1364 static void svc_sock_free(struct svc_xprt *xprt)
1366 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1367 dprintk("svc: svc_sock_free(%p)\n", svsk);
1369 if (svsk->sk_sock->file)
1370 sockfd_put(svsk->sk_sock);
1372 sock_release(svsk->sk_sock);