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>
52 #include <linux/sunrpc/xprt.h>
54 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
57 static struct svc_sock *svc_setup_socket(struct svc_serv *, struct socket *,
58 int *errp, int flags);
59 static void svc_udp_data_ready(struct sock *, int);
60 static int svc_udp_recvfrom(struct svc_rqst *);
61 static int svc_udp_sendto(struct svc_rqst *);
62 static void svc_sock_detach(struct svc_xprt *);
63 static void svc_tcp_sock_detach(struct svc_xprt *);
64 static void svc_sock_free(struct svc_xprt *);
66 static struct svc_xprt *svc_create_socket(struct svc_serv *, int,
67 struct sockaddr *, int, int);
68 #ifdef CONFIG_DEBUG_LOCK_ALLOC
69 static struct lock_class_key svc_key[2];
70 static struct lock_class_key svc_slock_key[2];
72 static void svc_reclassify_socket(struct socket *sock)
74 struct sock *sk = sock->sk;
75 BUG_ON(sock_owned_by_user(sk));
76 switch (sk->sk_family) {
78 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFSD",
80 "sk_xprt.xpt_lock-AF_INET-NFSD",
85 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFSD",
87 "sk_xprt.xpt_lock-AF_INET6-NFSD",
96 static void svc_reclassify_socket(struct socket *sock)
102 * Release an skbuff after use
104 static void svc_release_skb(struct svc_rqst *rqstp)
106 struct sk_buff *skb = rqstp->rq_xprt_ctxt;
109 struct svc_sock *svsk =
110 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
111 rqstp->rq_xprt_ctxt = NULL;
113 dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
114 skb_free_datagram_locked(svsk->sk_sk, skb);
118 union svc_pktinfo_u {
119 struct in_pktinfo pkti;
120 struct in6_pktinfo pkti6;
122 #define SVC_PKTINFO_SPACE \
123 CMSG_SPACE(sizeof(union svc_pktinfo_u))
125 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
127 struct svc_sock *svsk =
128 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
129 switch (svsk->sk_sk->sk_family) {
131 struct in_pktinfo *pki = CMSG_DATA(cmh);
133 cmh->cmsg_level = SOL_IP;
134 cmh->cmsg_type = IP_PKTINFO;
135 pki->ipi_ifindex = 0;
136 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr.addr.s_addr;
137 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
142 struct in6_pktinfo *pki = CMSG_DATA(cmh);
144 cmh->cmsg_level = SOL_IPV6;
145 cmh->cmsg_type = IPV6_PKTINFO;
146 pki->ipi6_ifindex = 0;
147 ipv6_addr_copy(&pki->ipi6_addr,
148 &rqstp->rq_daddr.addr6);
149 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
157 * send routine intended to be shared by the fore- and back-channel
159 int svc_send_common(struct socket *sock, struct xdr_buf *xdr,
160 struct page *headpage, unsigned long headoffset,
161 struct page *tailpage, unsigned long tailoffset)
165 struct page **ppage = xdr->pages;
166 size_t base = xdr->page_base;
167 unsigned int pglen = xdr->page_len;
168 unsigned int flags = MSG_MORE;
175 if (slen == xdr->head[0].iov_len)
177 len = kernel_sendpage(sock, headpage, headoffset,
178 xdr->head[0].iov_len, flags);
179 if (len != xdr->head[0].iov_len)
181 slen -= xdr->head[0].iov_len;
186 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
190 result = kernel_sendpage(sock, *ppage, base, size, flags);
197 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
203 if (xdr->tail[0].iov_len) {
204 result = kernel_sendpage(sock, tailpage, tailoffset,
205 xdr->tail[0].iov_len, 0);
216 * Generic sendto routine
218 static int svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
220 struct svc_sock *svsk =
221 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
222 struct socket *sock = svsk->sk_sock;
225 long all[SVC_PKTINFO_SPACE / sizeof(long)];
227 struct cmsghdr *cmh = &buffer.hdr;
229 unsigned long tailoff;
230 unsigned long headoff;
231 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
233 if (rqstp->rq_prot == IPPROTO_UDP) {
234 struct msghdr msg = {
235 .msg_name = &rqstp->rq_addr,
236 .msg_namelen = rqstp->rq_addrlen,
238 .msg_controllen = sizeof(buffer),
239 .msg_flags = MSG_MORE,
242 svc_set_cmsg_data(rqstp, cmh);
244 if (sock_sendmsg(sock, &msg, 0) < 0)
248 tailoff = ((unsigned long)xdr->tail[0].iov_base) & (PAGE_SIZE-1);
250 len = svc_send_common(sock, xdr, rqstp->rq_respages[0], headoff,
251 rqstp->rq_respages[0], tailoff);
254 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
255 svsk, xdr->head[0].iov_base, xdr->head[0].iov_len,
256 xdr->len, len, svc_print_addr(rqstp, buf, sizeof(buf)));
262 * Report socket names for nfsdfs
264 static int svc_one_sock_name(struct svc_sock *svsk, char *buf, int remaining)
266 const struct sock *sk = svsk->sk_sk;
267 const char *proto_name = sk->sk_protocol == IPPROTO_UDP ?
271 switch (sk->sk_family) {
273 len = snprintf(buf, remaining, "ipv4 %s %pI4 %d\n",
275 &inet_sk(sk)->inet_rcv_saddr,
276 inet_sk(sk)->inet_num);
279 len = snprintf(buf, remaining, "ipv6 %s %pI6 %d\n",
281 &inet6_sk(sk)->rcv_saddr,
282 inet_sk(sk)->inet_num);
285 len = snprintf(buf, remaining, "*unknown-%d*\n",
289 if (len >= remaining) {
291 return -ENAMETOOLONG;
297 * svc_sock_names - construct a list of listener names in a string
298 * @serv: pointer to RPC service
299 * @buf: pointer to a buffer to fill in with socket names
300 * @buflen: size of the buffer to be filled
301 * @toclose: pointer to '\0'-terminated C string containing the name
302 * of a listener to be closed
304 * Fills in @buf with a '\n'-separated list of names of listener
305 * sockets. If @toclose is not NULL, the socket named by @toclose
306 * is closed, and is not included in the output list.
308 * Returns positive length of the socket name string, or a negative
309 * errno value on error.
311 int svc_sock_names(struct svc_serv *serv, char *buf, const size_t buflen,
314 struct svc_sock *svsk, *closesk = NULL;
320 spin_lock_bh(&serv->sv_lock);
321 list_for_each_entry(svsk, &serv->sv_permsocks, sk_xprt.xpt_list) {
322 int onelen = svc_one_sock_name(svsk, buf + len, buflen - len);
327 if (toclose && strcmp(toclose, buf + len) == 0)
332 spin_unlock_bh(&serv->sv_lock);
335 /* Should unregister with portmap, but you cannot
336 * unregister just one protocol...
338 svc_close_xprt(&closesk->sk_xprt);
343 EXPORT_SYMBOL_GPL(svc_sock_names);
346 * Check input queue length
348 static int svc_recv_available(struct svc_sock *svsk)
350 struct socket *sock = svsk->sk_sock;
353 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
355 return (err >= 0)? avail : err;
359 * Generic recvfrom routine.
361 static int svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr,
364 struct svc_sock *svsk =
365 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
366 struct msghdr msg = {
367 .msg_flags = MSG_DONTWAIT,
371 rqstp->rq_xprt_hlen = 0;
373 len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
376 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
377 svsk, iov[0].iov_base, iov[0].iov_len, len);
382 * Set socket snd and rcv buffer lengths
384 static void svc_sock_setbufsize(struct socket *sock, unsigned int snd,
389 oldfs = get_fs(); set_fs(KERNEL_DS);
390 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
391 (char*)&snd, sizeof(snd));
392 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
393 (char*)&rcv, sizeof(rcv));
395 /* sock_setsockopt limits use to sysctl_?mem_max,
396 * which isn't acceptable. Until that is made conditional
397 * on not having CAP_SYS_RESOURCE or similar, we go direct...
398 * DaveM said I could!
401 sock->sk->sk_sndbuf = snd * 2;
402 sock->sk->sk_rcvbuf = rcv * 2;
403 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
404 sock->sk->sk_write_space(sock->sk);
405 release_sock(sock->sk);
409 * INET callback when data has been received on the socket.
411 static void svc_udp_data_ready(struct sock *sk, int count)
413 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
416 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
418 test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
419 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
420 svc_xprt_enqueue(&svsk->sk_xprt);
422 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
423 wake_up_interruptible(sk->sk_sleep);
427 * INET callback when space is newly available on the socket.
429 static void svc_write_space(struct sock *sk)
431 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
434 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
435 svsk, sk, test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
436 svc_xprt_enqueue(&svsk->sk_xprt);
439 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
440 dprintk("RPC svc_write_space: someone sleeping on %p\n",
442 wake_up_interruptible(sk->sk_sleep);
446 static void svc_tcp_write_space(struct sock *sk)
448 struct socket *sock = sk->sk_socket;
450 if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk) && sock)
451 clear_bit(SOCK_NOSPACE, &sock->flags);
456 * See net/ipv6/ip_sockglue.c : ip_cmsg_recv_pktinfo
458 static int svc_udp_get_dest_address4(struct svc_rqst *rqstp,
461 struct in_pktinfo *pki = CMSG_DATA(cmh);
462 if (cmh->cmsg_type != IP_PKTINFO)
464 rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
469 * See net/ipv6/datagram.c : datagram_recv_ctl
471 static int svc_udp_get_dest_address6(struct svc_rqst *rqstp,
474 struct in6_pktinfo *pki = CMSG_DATA(cmh);
475 if (cmh->cmsg_type != IPV6_PKTINFO)
477 ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
482 * Copy the UDP datagram's destination address to the rqstp structure.
483 * The 'destination' address in this case is the address to which the
484 * peer sent the datagram, i.e. our local address. For multihomed
485 * hosts, this can change from msg to msg. Note that only the IP
486 * address changes, the port number should remain the same.
488 static int svc_udp_get_dest_address(struct svc_rqst *rqstp,
491 switch (cmh->cmsg_level) {
493 return svc_udp_get_dest_address4(rqstp, cmh);
495 return svc_udp_get_dest_address6(rqstp, cmh);
502 * Receive a datagram from a UDP socket.
504 static int svc_udp_recvfrom(struct svc_rqst *rqstp)
506 struct svc_sock *svsk =
507 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
508 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
512 long all[SVC_PKTINFO_SPACE / sizeof(long)];
514 struct cmsghdr *cmh = &buffer.hdr;
515 struct msghdr msg = {
516 .msg_name = svc_addr(rqstp),
518 .msg_controllen = sizeof(buffer),
519 .msg_flags = MSG_DONTWAIT,
524 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
525 /* udp sockets need large rcvbuf as all pending
526 * requests are still in that buffer. sndbuf must
527 * also be large enough that there is enough space
528 * for one reply per thread. We count all threads
529 * rather than threads in a particular pool, which
530 * provides an upper bound on the number of threads
531 * which will access the socket.
533 svc_sock_setbufsize(svsk->sk_sock,
534 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
535 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
537 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
539 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
540 0, 0, MSG_PEEK | MSG_DONTWAIT);
542 skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err);
545 if (err != -EAGAIN) {
546 /* possibly an icmp error */
547 dprintk("svc: recvfrom returned error %d\n", -err);
548 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
552 len = svc_addr_len(svc_addr(rqstp));
554 return -EAFNOSUPPORT;
555 rqstp->rq_addrlen = len;
556 if (skb->tstamp.tv64 == 0) {
557 skb->tstamp = ktime_get_real();
558 /* Don't enable netstamp, sunrpc doesn't
559 need that much accuracy */
561 svsk->sk_sk->sk_stamp = skb->tstamp;
562 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* there may be more data... */
564 len = skb->len - sizeof(struct udphdr);
565 rqstp->rq_arg.len = len;
567 rqstp->rq_prot = IPPROTO_UDP;
569 if (!svc_udp_get_dest_address(rqstp, cmh)) {
572 "svc: received unknown control message %d/%d; "
573 "dropping RPC reply datagram\n",
574 cmh->cmsg_level, cmh->cmsg_type);
575 skb_free_datagram_locked(svsk->sk_sk, skb);
579 if (skb_is_nonlinear(skb)) {
580 /* we have to copy */
582 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
585 skb_free_datagram_locked(svsk->sk_sk, skb);
589 skb_free_datagram_locked(svsk->sk_sk, skb);
591 /* we can use it in-place */
592 rqstp->rq_arg.head[0].iov_base = skb->data +
593 sizeof(struct udphdr);
594 rqstp->rq_arg.head[0].iov_len = len;
595 if (skb_checksum_complete(skb)) {
596 skb_free_datagram_locked(svsk->sk_sk, skb);
599 rqstp->rq_xprt_ctxt = skb;
602 rqstp->rq_arg.page_base = 0;
603 if (len <= rqstp->rq_arg.head[0].iov_len) {
604 rqstp->rq_arg.head[0].iov_len = len;
605 rqstp->rq_arg.page_len = 0;
606 rqstp->rq_respages = rqstp->rq_pages+1;
608 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
609 rqstp->rq_respages = rqstp->rq_pages + 1 +
610 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
614 serv->sv_stats->netudpcnt++;
620 svc_udp_sendto(struct svc_rqst *rqstp)
624 error = svc_sendto(rqstp, &rqstp->rq_res);
625 if (error == -ECONNREFUSED)
626 /* ICMP error on earlier request. */
627 error = svc_sendto(rqstp, &rqstp->rq_res);
632 static void svc_udp_prep_reply_hdr(struct svc_rqst *rqstp)
636 static int svc_udp_has_wspace(struct svc_xprt *xprt)
638 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
639 struct svc_serv *serv = xprt->xpt_server;
640 unsigned long required;
643 * Set the SOCK_NOSPACE flag before checking the available
646 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
647 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
648 if (required*2 > sock_wspace(svsk->sk_sk))
650 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
654 static struct svc_xprt *svc_udp_accept(struct svc_xprt *xprt)
660 static struct svc_xprt *svc_udp_create(struct svc_serv *serv,
661 struct sockaddr *sa, int salen,
664 return svc_create_socket(serv, IPPROTO_UDP, sa, salen, flags);
667 static struct svc_xprt_ops svc_udp_ops = {
668 .xpo_create = svc_udp_create,
669 .xpo_recvfrom = svc_udp_recvfrom,
670 .xpo_sendto = svc_udp_sendto,
671 .xpo_release_rqst = svc_release_skb,
672 .xpo_detach = svc_sock_detach,
673 .xpo_free = svc_sock_free,
674 .xpo_prep_reply_hdr = svc_udp_prep_reply_hdr,
675 .xpo_has_wspace = svc_udp_has_wspace,
676 .xpo_accept = svc_udp_accept,
679 static struct svc_xprt_class svc_udp_class = {
681 .xcl_owner = THIS_MODULE,
682 .xcl_ops = &svc_udp_ops,
683 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
686 static void svc_udp_init(struct svc_sock *svsk, struct svc_serv *serv)
688 int err, level, optname, one = 1;
690 svc_xprt_init(&svc_udp_class, &svsk->sk_xprt, serv);
691 clear_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
692 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
693 svsk->sk_sk->sk_write_space = svc_write_space;
695 /* initialise setting must have enough space to
696 * receive and respond to one request.
697 * svc_udp_recvfrom will re-adjust if necessary
699 svc_sock_setbufsize(svsk->sk_sock,
700 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
701 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
703 /* data might have come in before data_ready set up */
704 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
705 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
707 /* make sure we get destination address info */
708 switch (svsk->sk_sk->sk_family) {
711 optname = IP_PKTINFO;
715 optname = IPV6_RECVPKTINFO;
720 err = kernel_setsockopt(svsk->sk_sock, level, optname,
721 (char *)&one, sizeof(one));
722 dprintk("svc: kernel_setsockopt returned %d\n", err);
726 * A data_ready event on a listening socket means there's a connection
727 * pending. Do not use state_change as a substitute for it.
729 static void svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
731 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
733 dprintk("svc: socket %p TCP (listen) state change %d\n",
737 * This callback may called twice when a new connection
738 * is established as a child socket inherits everything
739 * from a parent LISTEN socket.
740 * 1) data_ready method of the parent socket will be called
741 * when one of child sockets become ESTABLISHED.
742 * 2) data_ready method of the child socket may be called
743 * when it receives data before the socket is accepted.
744 * In case of 2, we should ignore it silently.
746 if (sk->sk_state == TCP_LISTEN) {
748 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
749 svc_xprt_enqueue(&svsk->sk_xprt);
751 printk("svc: socket %p: no user data\n", sk);
754 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
755 wake_up_interruptible_all(sk->sk_sleep);
759 * A state change on a connected socket means it's dying or dead.
761 static void svc_tcp_state_change(struct sock *sk)
763 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
765 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
766 sk, sk->sk_state, sk->sk_user_data);
769 printk("svc: socket %p: no user data\n", sk);
771 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
772 svc_xprt_enqueue(&svsk->sk_xprt);
774 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
775 wake_up_interruptible_all(sk->sk_sleep);
778 static void svc_tcp_data_ready(struct sock *sk, int count)
780 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
782 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
783 sk, sk->sk_user_data);
785 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
786 svc_xprt_enqueue(&svsk->sk_xprt);
788 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
789 wake_up_interruptible(sk->sk_sleep);
793 * Accept a TCP connection
795 static struct svc_xprt *svc_tcp_accept(struct svc_xprt *xprt)
797 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
798 struct sockaddr_storage addr;
799 struct sockaddr *sin = (struct sockaddr *) &addr;
800 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
801 struct socket *sock = svsk->sk_sock;
802 struct socket *newsock;
803 struct svc_sock *newsvsk;
805 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
807 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
811 clear_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
812 err = kernel_accept(sock, &newsock, O_NONBLOCK);
815 printk(KERN_WARNING "%s: no more sockets!\n",
817 else if (err != -EAGAIN && net_ratelimit())
818 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
819 serv->sv_name, -err);
822 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
824 err = kernel_getpeername(newsock, sin, &slen);
827 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
828 serv->sv_name, -err);
829 goto failed; /* aborted connection or whatever */
832 /* Ideally, we would want to reject connections from unauthorized
833 * hosts here, but when we get encryption, the IP of the host won't
834 * tell us anything. For now just warn about unpriv connections.
836 if (!svc_port_is_privileged(sin)) {
838 "%s: connect from unprivileged port: %s\n",
840 __svc_print_addr(sin, buf, sizeof(buf)));
842 dprintk("%s: connect from %s\n", serv->sv_name,
843 __svc_print_addr(sin, buf, sizeof(buf)));
845 /* make sure that a write doesn't block forever when
848 newsock->sk->sk_sndtimeo = HZ*30;
850 if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
851 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
853 svc_xprt_set_remote(&newsvsk->sk_xprt, sin, slen);
854 err = kernel_getsockname(newsock, sin, &slen);
855 if (unlikely(err < 0)) {
856 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err);
857 slen = offsetof(struct sockaddr, sa_data);
859 svc_xprt_set_local(&newsvsk->sk_xprt, sin, slen);
862 serv->sv_stats->nettcpconn++;
864 return &newsvsk->sk_xprt;
867 sock_release(newsock);
873 * If we haven't gotten the record length yet, get the next four bytes.
874 * Otherwise try to gobble up as much as possible up to the complete
877 static int svc_tcp_recv_record(struct svc_sock *svsk, struct svc_rqst *rqstp)
879 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
882 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
883 /* sndbuf needs to have room for one request
884 * per thread, otherwise we can stall even when the
885 * network isn't a bottleneck.
887 * We count all threads rather than threads in a
888 * particular pool, which provides an upper bound
889 * on the number of threads which will access the socket.
891 * rcvbuf just needs to be able to hold a few requests.
892 * Normally they will be removed from the queue
893 * as soon a a complete request arrives.
895 svc_sock_setbufsize(svsk->sk_sock,
896 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
897 3 * serv->sv_max_mesg);
899 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
901 if (svsk->sk_tcplen < sizeof(rpc_fraghdr)) {
902 int want = sizeof(rpc_fraghdr) - svsk->sk_tcplen;
905 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
907 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
909 svsk->sk_tcplen += len;
912 dprintk("svc: short recvfrom while reading record "
913 "length (%d of %d)\n", len, want);
914 goto err_again; /* record header not complete */
917 svsk->sk_reclen = ntohl(svsk->sk_reclen);
918 if (!(svsk->sk_reclen & RPC_LAST_STREAM_FRAGMENT)) {
919 /* FIXME: technically, a record can be fragmented,
920 * and non-terminal fragments will not have the top
921 * bit set in the fragment length header.
922 * But apparently no known nfs clients send fragmented
925 printk(KERN_NOTICE "RPC: multiple fragments "
926 "per record not supported\n");
930 svsk->sk_reclen &= RPC_FRAGMENT_SIZE_MASK;
931 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
932 if (svsk->sk_reclen > serv->sv_max_mesg) {
934 printk(KERN_NOTICE "RPC: "
935 "fragment too large: 0x%08lx\n",
936 (unsigned long)svsk->sk_reclen);
941 /* Check whether enough data is available */
942 len = svc_recv_available(svsk);
946 if (len < svsk->sk_reclen) {
947 dprintk("svc: incomplete TCP record (%d of %d)\n",
948 len, svsk->sk_reclen);
949 goto err_again; /* record not complete */
951 len = svsk->sk_reclen;
952 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
957 dprintk("RPC: TCP recv_record got EAGAIN\n");
960 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
965 static int svc_process_calldir(struct svc_sock *svsk, struct svc_rqst *rqstp,
966 struct rpc_rqst **reqpp, struct kvec *vec)
968 struct rpc_rqst *req = NULL;
974 len = svc_recvfrom(rqstp, vec, 1, 8);
978 p = (u32 *)rqstp->rq_arg.head[0].iov_base;
983 /* REQUEST is the most common case */
984 vec[0] = rqstp->rq_arg.head[0];
987 if (svsk->sk_bc_xprt)
988 req = xprt_lookup_rqst(svsk->sk_bc_xprt, xid);
992 "%s: Got unrecognized reply: "
993 "calldir 0x%x sk_bc_xprt %p xid %08x\n",
994 __func__, ntohl(calldir),
995 svsk->sk_bc_xprt, xid);
996 vec[0] = rqstp->rq_arg.head[0];
1000 memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
1001 sizeof(struct xdr_buf));
1002 /* copy the xid and call direction */
1003 memcpy(req->rq_private_buf.head[0].iov_base,
1004 rqstp->rq_arg.head[0].iov_base, 8);
1005 vec[0] = req->rq_private_buf.head[0];
1008 vec[0].iov_base += 8;
1009 vec[0].iov_len -= 8;
1010 len = svsk->sk_reclen - 8;
1017 * Receive data from a TCP socket.
1019 static int svc_tcp_recvfrom(struct svc_rqst *rqstp)
1021 struct svc_sock *svsk =
1022 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
1023 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1027 struct rpc_rqst *req = NULL;
1029 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
1030 svsk, test_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags),
1031 test_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags),
1032 test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags));
1034 len = svc_tcp_recv_record(svsk, rqstp);
1038 vec = rqstp->rq_vec;
1039 vec[0] = rqstp->rq_arg.head[0];
1043 * We have enough data for the whole tcp record. Let's try and read the
1044 * first 8 bytes to get the xid and the call direction. We can use this
1045 * to figure out if this is a call or a reply to a callback. If
1046 * sk_reclen is < 8 (xid and calldir), then this is a malformed packet.
1047 * In that case, don't bother with the calldir and just read the data.
1048 * It will be rejected in svc_process.
1051 len = svc_process_calldir(svsk, rqstp, &req, vec);
1058 while (vlen < len) {
1059 vec[pnum].iov_base = (req) ?
1060 page_address(req->rq_private_buf.pages[pnum - 1]) :
1061 page_address(rqstp->rq_pages[pnum]);
1062 vec[pnum].iov_len = PAGE_SIZE;
1066 rqstp->rq_respages = &rqstp->rq_pages[pnum];
1068 /* Now receive data */
1069 len = svc_recvfrom(rqstp, vec, pnum, len);
1074 * Account for the 8 bytes we read earlier
1079 xprt_complete_rqst(req->rq_task, len);
1083 dprintk("svc: TCP complete record (%d bytes)\n", len);
1084 rqstp->rq_arg.len = len;
1085 rqstp->rq_arg.page_base = 0;
1086 if (len <= rqstp->rq_arg.head[0].iov_len) {
1087 rqstp->rq_arg.head[0].iov_len = len;
1088 rqstp->rq_arg.page_len = 0;
1090 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1093 rqstp->rq_xprt_ctxt = NULL;
1094 rqstp->rq_prot = IPPROTO_TCP;
1097 /* Reset TCP read info */
1098 svsk->sk_reclen = 0;
1099 svsk->sk_tcplen = 0;
1101 svc_xprt_copy_addrs(rqstp, &svsk->sk_xprt);
1103 serv->sv_stats->nettcpcnt++;
1108 if (len == -EAGAIN) {
1109 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1113 if (len != -EAGAIN) {
1114 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1115 svsk->sk_xprt.xpt_server->sv_name, -len);
1116 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1122 * Send out data on TCP socket.
1124 static int svc_tcp_sendto(struct svc_rqst *rqstp)
1126 struct xdr_buf *xbufp = &rqstp->rq_res;
1130 /* Set up the first element of the reply kvec.
1131 * Any other kvecs that may be in use have been taken
1132 * care of by the server implementation itself.
1134 reclen = htonl(0x80000000|((xbufp->len ) - 4));
1135 memcpy(xbufp->head[0].iov_base, &reclen, 4);
1137 if (test_bit(XPT_DEAD, &rqstp->rq_xprt->xpt_flags))
1140 sent = svc_sendto(rqstp, &rqstp->rq_res);
1141 if (sent != xbufp->len) {
1143 "rpc-srv/tcp: %s: %s %d when sending %d bytes "
1144 "- shutting down socket\n",
1145 rqstp->rq_xprt->xpt_server->sv_name,
1146 (sent<0)?"got error":"sent only",
1148 set_bit(XPT_CLOSE, &rqstp->rq_xprt->xpt_flags);
1149 svc_xprt_enqueue(rqstp->rq_xprt);
1156 * Setup response header. TCP has a 4B record length field.
1158 static void svc_tcp_prep_reply_hdr(struct svc_rqst *rqstp)
1160 struct kvec *resv = &rqstp->rq_res.head[0];
1162 /* tcp needs a space for the record length... */
1166 static int svc_tcp_has_wspace(struct svc_xprt *xprt)
1168 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1169 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1172 if (test_bit(XPT_LISTENER, &xprt->xpt_flags))
1174 required = atomic_read(&xprt->xpt_reserved) + serv->sv_max_mesg;
1175 if (sk_stream_wspace(svsk->sk_sk) >= required)
1177 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1181 static struct svc_xprt *svc_tcp_create(struct svc_serv *serv,
1182 struct sockaddr *sa, int salen,
1185 return svc_create_socket(serv, IPPROTO_TCP, sa, salen, flags);
1188 static struct svc_xprt_ops svc_tcp_ops = {
1189 .xpo_create = svc_tcp_create,
1190 .xpo_recvfrom = svc_tcp_recvfrom,
1191 .xpo_sendto = svc_tcp_sendto,
1192 .xpo_release_rqst = svc_release_skb,
1193 .xpo_detach = svc_tcp_sock_detach,
1194 .xpo_free = svc_sock_free,
1195 .xpo_prep_reply_hdr = svc_tcp_prep_reply_hdr,
1196 .xpo_has_wspace = svc_tcp_has_wspace,
1197 .xpo_accept = svc_tcp_accept,
1200 static struct svc_xprt_class svc_tcp_class = {
1202 .xcl_owner = THIS_MODULE,
1203 .xcl_ops = &svc_tcp_ops,
1204 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1207 void svc_init_xprt_sock(void)
1209 svc_reg_xprt_class(&svc_tcp_class);
1210 svc_reg_xprt_class(&svc_udp_class);
1213 void svc_cleanup_xprt_sock(void)
1215 svc_unreg_xprt_class(&svc_tcp_class);
1216 svc_unreg_xprt_class(&svc_udp_class);
1219 static void svc_tcp_init(struct svc_sock *svsk, struct svc_serv *serv)
1221 struct sock *sk = svsk->sk_sk;
1223 svc_xprt_init(&svc_tcp_class, &svsk->sk_xprt, serv);
1224 set_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
1225 if (sk->sk_state == TCP_LISTEN) {
1226 dprintk("setting up TCP socket for listening\n");
1227 set_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags);
1228 sk->sk_data_ready = svc_tcp_listen_data_ready;
1229 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1231 dprintk("setting up TCP socket for reading\n");
1232 sk->sk_state_change = svc_tcp_state_change;
1233 sk->sk_data_ready = svc_tcp_data_ready;
1234 sk->sk_write_space = svc_tcp_write_space;
1236 svsk->sk_reclen = 0;
1237 svsk->sk_tcplen = 0;
1239 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1241 /* initialise setting must have enough space to
1242 * receive and respond to one request.
1243 * svc_tcp_recvfrom will re-adjust if necessary
1245 svc_sock_setbufsize(svsk->sk_sock,
1246 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
1247 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
1249 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1250 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1251 if (sk->sk_state != TCP_ESTABLISHED)
1252 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1256 void svc_sock_update_bufs(struct svc_serv *serv)
1259 * The number of server threads has changed. Update
1260 * rcvbuf and sndbuf accordingly on all sockets
1262 struct list_head *le;
1264 spin_lock_bh(&serv->sv_lock);
1265 list_for_each(le, &serv->sv_permsocks) {
1266 struct svc_sock *svsk =
1267 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1268 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1270 list_for_each(le, &serv->sv_tempsocks) {
1271 struct svc_sock *svsk =
1272 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1273 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1275 spin_unlock_bh(&serv->sv_lock);
1277 EXPORT_SYMBOL_GPL(svc_sock_update_bufs);
1280 * Initialize socket for RPC use and create svc_sock struct
1281 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1283 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1284 struct socket *sock,
1285 int *errp, int flags)
1287 struct svc_sock *svsk;
1289 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1291 dprintk("svc: svc_setup_socket %p\n", sock);
1292 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1299 /* Register socket with portmapper */
1300 if (*errp >= 0 && pmap_register)
1301 *errp = svc_register(serv, inet->sk_family, inet->sk_protocol,
1302 ntohs(inet_sk(inet)->inet_sport));
1309 inet->sk_user_data = svsk;
1310 svsk->sk_sock = sock;
1312 svsk->sk_ostate = inet->sk_state_change;
1313 svsk->sk_odata = inet->sk_data_ready;
1314 svsk->sk_owspace = inet->sk_write_space;
1316 /* Initialize the socket */
1317 if (sock->type == SOCK_DGRAM)
1318 svc_udp_init(svsk, serv);
1320 svc_tcp_init(svsk, serv);
1322 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1329 * svc_addsock - add a listener socket to an RPC service
1330 * @serv: pointer to RPC service to which to add a new listener
1331 * @fd: file descriptor of the new listener
1332 * @name_return: pointer to buffer to fill in with name of listener
1333 * @len: size of the buffer
1335 * Fills in socket name and returns positive length of name if successful.
1336 * Name is terminated with '\n'. On error, returns a negative errno
1339 int svc_addsock(struct svc_serv *serv, const int fd, char *name_return,
1343 struct socket *so = sockfd_lookup(fd, &err);
1344 struct svc_sock *svsk = NULL;
1348 if ((so->sk->sk_family != PF_INET) && (so->sk->sk_family != PF_INET6))
1349 err = -EAFNOSUPPORT;
1350 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1351 so->sk->sk_protocol != IPPROTO_UDP)
1352 err = -EPROTONOSUPPORT;
1353 else if (so->state > SS_UNCONNECTED)
1356 if (!try_module_get(THIS_MODULE))
1359 svsk = svc_setup_socket(serv, so, &err,
1362 struct sockaddr_storage addr;
1363 struct sockaddr *sin = (struct sockaddr *)&addr;
1365 if (kernel_getsockname(svsk->sk_sock, sin, &salen) == 0)
1366 svc_xprt_set_local(&svsk->sk_xprt, sin, salen);
1367 clear_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1368 spin_lock_bh(&serv->sv_lock);
1369 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_permsocks);
1370 spin_unlock_bh(&serv->sv_lock);
1371 svc_xprt_received(&svsk->sk_xprt);
1374 module_put(THIS_MODULE);
1380 return svc_one_sock_name(svsk, name_return, len);
1382 EXPORT_SYMBOL_GPL(svc_addsock);
1385 * Create socket for RPC service.
1387 static struct svc_xprt *svc_create_socket(struct svc_serv *serv,
1389 struct sockaddr *sin, int len,
1392 struct svc_sock *svsk;
1393 struct socket *sock;
1396 struct sockaddr_storage addr;
1397 struct sockaddr *newsin = (struct sockaddr *)&addr;
1401 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
1403 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1404 serv->sv_program->pg_name, protocol,
1405 __svc_print_addr(sin, buf, sizeof(buf)));
1407 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1408 printk(KERN_WARNING "svc: only UDP and TCP "
1409 "sockets supported\n");
1410 return ERR_PTR(-EINVAL);
1413 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1414 switch (sin->sa_family) {
1422 return ERR_PTR(-EINVAL);
1425 error = sock_create_kern(family, type, protocol, &sock);
1427 return ERR_PTR(error);
1429 svc_reclassify_socket(sock);
1432 * If this is an PF_INET6 listener, we want to avoid
1433 * getting requests from IPv4 remotes. Those should
1434 * be shunted to a PF_INET listener via rpcbind.
1437 if (family == PF_INET6)
1438 kernel_setsockopt(sock, SOL_IPV6, IPV6_V6ONLY,
1439 (char *)&val, sizeof(val));
1441 if (type == SOCK_STREAM)
1442 sock->sk->sk_reuse = 1; /* allow address reuse */
1443 error = kernel_bind(sock, sin, len);
1448 error = kernel_getsockname(sock, newsin, &newlen);
1452 if (protocol == IPPROTO_TCP) {
1453 if ((error = kernel_listen(sock, 64)) < 0)
1457 if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1458 svc_xprt_set_local(&svsk->sk_xprt, newsin, newlen);
1459 return (struct svc_xprt *)svsk;
1463 dprintk("svc: svc_create_socket error = %d\n", -error);
1465 return ERR_PTR(error);
1469 * Detach the svc_sock from the socket so that no
1470 * more callbacks occur.
1472 static void svc_sock_detach(struct svc_xprt *xprt)
1474 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1475 struct sock *sk = svsk->sk_sk;
1477 dprintk("svc: svc_sock_detach(%p)\n", svsk);
1479 /* put back the old socket callbacks */
1480 sk->sk_state_change = svsk->sk_ostate;
1481 sk->sk_data_ready = svsk->sk_odata;
1482 sk->sk_write_space = svsk->sk_owspace;
1484 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1485 wake_up_interruptible(sk->sk_sleep);
1489 * Disconnect the socket, and reset the callbacks
1491 static void svc_tcp_sock_detach(struct svc_xprt *xprt)
1493 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1495 dprintk("svc: svc_tcp_sock_detach(%p)\n", svsk);
1497 svc_sock_detach(xprt);
1499 if (!test_bit(XPT_LISTENER, &xprt->xpt_flags))
1500 kernel_sock_shutdown(svsk->sk_sock, SHUT_RDWR);
1504 * Free the svc_sock's socket resources and the svc_sock itself.
1506 static void svc_sock_free(struct svc_xprt *xprt)
1508 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1509 dprintk("svc: svc_sock_free(%p)\n", svsk);
1511 if (svsk->sk_sock->file)
1512 sockfd_put(svsk->sk_sock);
1514 sock_release(svsk->sk_sock);
1519 * Create a svc_xprt.
1521 * For internal use only (e.g. nfsv4.1 backchannel).
1522 * Callers should typically use the xpo_create() method.
1524 struct svc_xprt *svc_sock_create(struct svc_serv *serv, int prot)
1526 struct svc_sock *svsk;
1527 struct svc_xprt *xprt = NULL;
1529 dprintk("svc: %s\n", __func__);
1530 svsk = kzalloc(sizeof(*svsk), GFP_KERNEL);
1534 xprt = &svsk->sk_xprt;
1535 if (prot == IPPROTO_TCP)
1536 svc_xprt_init(&svc_tcp_class, xprt, serv);
1537 else if (prot == IPPROTO_UDP)
1538 svc_xprt_init(&svc_udp_class, xprt, serv);
1542 dprintk("svc: %s return %p\n", __func__, xprt);
1545 EXPORT_SYMBOL_GPL(svc_sock_create);
1548 * Destroy a svc_sock.
1550 void svc_sock_destroy(struct svc_xprt *xprt)
1553 kfree(container_of(xprt, struct svc_sock, sk_xprt));
1555 EXPORT_SYMBOL_GPL(svc_sock_destroy);