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;
106 struct svc_deferred_req *dr = rqstp->rq_deferred;
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(svsk->sk_sk, skb);
117 rqstp->rq_deferred = NULL;
122 union svc_pktinfo_u {
123 struct in_pktinfo pkti;
124 struct in6_pktinfo pkti6;
126 #define SVC_PKTINFO_SPACE \
127 CMSG_SPACE(sizeof(union svc_pktinfo_u))
129 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
131 struct svc_sock *svsk =
132 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
133 switch (svsk->sk_sk->sk_family) {
135 struct in_pktinfo *pki = CMSG_DATA(cmh);
137 cmh->cmsg_level = SOL_IP;
138 cmh->cmsg_type = IP_PKTINFO;
139 pki->ipi_ifindex = 0;
140 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr.addr.s_addr;
141 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
146 struct in6_pktinfo *pki = CMSG_DATA(cmh);
148 cmh->cmsg_level = SOL_IPV6;
149 cmh->cmsg_type = IPV6_PKTINFO;
150 pki->ipi6_ifindex = 0;
151 ipv6_addr_copy(&pki->ipi6_addr,
152 &rqstp->rq_daddr.addr6);
153 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
161 * Generic sendto routine
163 static int svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
165 struct svc_sock *svsk =
166 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
167 struct socket *sock = svsk->sk_sock;
171 long all[SVC_PKTINFO_SPACE / sizeof(long)];
173 struct cmsghdr *cmh = &buffer.hdr;
177 struct page **ppage = xdr->pages;
178 size_t base = xdr->page_base;
179 unsigned int pglen = xdr->page_len;
180 unsigned int flags = MSG_MORE;
181 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
185 if (rqstp->rq_prot == IPPROTO_UDP) {
186 struct msghdr msg = {
187 .msg_name = &rqstp->rq_addr,
188 .msg_namelen = rqstp->rq_addrlen,
190 .msg_controllen = sizeof(buffer),
191 .msg_flags = MSG_MORE,
194 svc_set_cmsg_data(rqstp, cmh);
196 if (sock_sendmsg(sock, &msg, 0) < 0)
201 if (slen == xdr->head[0].iov_len)
203 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0,
204 xdr->head[0].iov_len, flags);
205 if (len != xdr->head[0].iov_len)
207 slen -= xdr->head[0].iov_len;
212 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
216 result = kernel_sendpage(sock, *ppage, base, size, flags);
223 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
228 if (xdr->tail[0].iov_len) {
229 result = kernel_sendpage(sock, rqstp->rq_respages[0],
230 ((unsigned long)xdr->tail[0].iov_base)
232 xdr->tail[0].iov_len, 0);
238 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
239 svsk, xdr->head[0].iov_base, xdr->head[0].iov_len,
240 xdr->len, len, svc_print_addr(rqstp, buf, sizeof(buf)));
246 * Report socket names for nfsdfs
248 static int one_sock_name(char *buf, struct svc_sock *svsk)
252 switch(svsk->sk_sk->sk_family) {
254 len = sprintf(buf, "ipv4 %s %pI4 %d\n",
255 svsk->sk_sk->sk_protocol == IPPROTO_UDP ?
257 &inet_sk(svsk->sk_sk)->rcv_saddr,
258 inet_sk(svsk->sk_sk)->num);
261 len = sprintf(buf, "*unknown-%d*\n",
262 svsk->sk_sk->sk_family);
268 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
270 struct svc_sock *svsk, *closesk = NULL;
275 spin_lock_bh(&serv->sv_lock);
276 list_for_each_entry(svsk, &serv->sv_permsocks, sk_xprt.xpt_list) {
277 int onelen = one_sock_name(buf+len, svsk);
278 if (toclose && strcmp(toclose, buf+len) == 0)
283 spin_unlock_bh(&serv->sv_lock);
285 /* Should unregister with portmap, but you cannot
286 * unregister just one protocol...
288 svc_close_xprt(&closesk->sk_xprt);
293 EXPORT_SYMBOL(svc_sock_names);
296 * Check input queue length
298 static int svc_recv_available(struct svc_sock *svsk)
300 struct socket *sock = svsk->sk_sock;
303 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
305 return (err >= 0)? avail : err;
309 * Generic recvfrom routine.
311 static int svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr,
314 struct svc_sock *svsk =
315 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
316 struct msghdr msg = {
317 .msg_flags = MSG_DONTWAIT,
321 rqstp->rq_xprt_hlen = 0;
323 len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
326 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
327 svsk, iov[0].iov_base, iov[0].iov_len, len);
332 * Set socket snd and rcv buffer lengths
334 static void svc_sock_setbufsize(struct socket *sock, unsigned int snd,
339 oldfs = get_fs(); set_fs(KERNEL_DS);
340 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
341 (char*)&snd, sizeof(snd));
342 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
343 (char*)&rcv, sizeof(rcv));
345 /* sock_setsockopt limits use to sysctl_?mem_max,
346 * which isn't acceptable. Until that is made conditional
347 * on not having CAP_SYS_RESOURCE or similar, we go direct...
348 * DaveM said I could!
351 sock->sk->sk_sndbuf = snd * 2;
352 sock->sk->sk_rcvbuf = rcv * 2;
353 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
354 release_sock(sock->sk);
358 * INET callback when data has been received on the socket.
360 static void svc_udp_data_ready(struct sock *sk, int count)
362 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
365 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
367 test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
368 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
369 svc_xprt_enqueue(&svsk->sk_xprt);
371 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
372 wake_up_interruptible(sk->sk_sleep);
376 * INET callback when space is newly available on the socket.
378 static void svc_write_space(struct sock *sk)
380 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
383 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
384 svsk, sk, test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
385 svc_xprt_enqueue(&svsk->sk_xprt);
388 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
389 dprintk("RPC svc_write_space: someone sleeping on %p\n",
391 wake_up_interruptible(sk->sk_sleep);
396 * Copy the UDP datagram's destination address to the rqstp structure.
397 * The 'destination' address in this case is the address to which the
398 * peer sent the datagram, i.e. our local address. For multihomed
399 * hosts, this can change from msg to msg. Note that only the IP
400 * address changes, the port number should remain the same.
402 static void svc_udp_get_dest_address(struct svc_rqst *rqstp,
405 struct svc_sock *svsk =
406 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
407 switch (svsk->sk_sk->sk_family) {
409 struct in_pktinfo *pki = CMSG_DATA(cmh);
410 rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
414 struct in6_pktinfo *pki = CMSG_DATA(cmh);
415 ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
422 * Receive a datagram from a UDP socket.
424 static int svc_udp_recvfrom(struct svc_rqst *rqstp)
426 struct svc_sock *svsk =
427 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
428 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
432 long all[SVC_PKTINFO_SPACE / sizeof(long)];
434 struct cmsghdr *cmh = &buffer.hdr;
436 struct msghdr msg = {
437 .msg_name = svc_addr(rqstp),
439 .msg_controllen = sizeof(buffer),
440 .msg_flags = MSG_DONTWAIT,
443 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
444 /* udp sockets need large rcvbuf as all pending
445 * requests are still in that buffer. sndbuf must
446 * also be large enough that there is enough space
447 * for one reply per thread. We count all threads
448 * rather than threads in a particular pool, which
449 * provides an upper bound on the number of threads
450 * which will access the socket.
452 svc_sock_setbufsize(svsk->sk_sock,
453 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
454 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
456 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
458 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
459 0, 0, MSG_PEEK | MSG_DONTWAIT);
461 skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err);
464 if (err != -EAGAIN) {
465 /* possibly an icmp error */
466 dprintk("svc: recvfrom returned error %d\n", -err);
467 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
469 svc_xprt_received(&svsk->sk_xprt);
472 len = svc_addr_len(svc_addr(rqstp));
475 rqstp->rq_addrlen = len;
476 if (skb->tstamp.tv64 == 0) {
477 skb->tstamp = ktime_get_real();
478 /* Don't enable netstamp, sunrpc doesn't
479 need that much accuracy */
481 svsk->sk_sk->sk_stamp = skb->tstamp;
482 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* there may be more data... */
485 * Maybe more packets - kick another thread ASAP.
487 svc_xprt_received(&svsk->sk_xprt);
489 len = skb->len - sizeof(struct udphdr);
490 rqstp->rq_arg.len = len;
492 rqstp->rq_prot = IPPROTO_UDP;
494 if (cmh->cmsg_level != IPPROTO_IP ||
495 cmh->cmsg_type != IP_PKTINFO) {
497 printk("rpcsvc: received unknown control message:"
499 cmh->cmsg_level, cmh->cmsg_type);
500 skb_free_datagram(svsk->sk_sk, skb);
503 svc_udp_get_dest_address(rqstp, cmh);
505 if (skb_is_nonlinear(skb)) {
506 /* we have to copy */
508 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
511 skb_free_datagram(svsk->sk_sk, skb);
515 skb_free_datagram(svsk->sk_sk, skb);
517 /* we can use it in-place */
518 rqstp->rq_arg.head[0].iov_base = skb->data +
519 sizeof(struct udphdr);
520 rqstp->rq_arg.head[0].iov_len = len;
521 if (skb_checksum_complete(skb)) {
522 skb_free_datagram(svsk->sk_sk, skb);
525 rqstp->rq_xprt_ctxt = skb;
528 rqstp->rq_arg.page_base = 0;
529 if (len <= rqstp->rq_arg.head[0].iov_len) {
530 rqstp->rq_arg.head[0].iov_len = len;
531 rqstp->rq_arg.page_len = 0;
532 rqstp->rq_respages = rqstp->rq_pages+1;
534 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
535 rqstp->rq_respages = rqstp->rq_pages + 1 +
536 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
540 serv->sv_stats->netudpcnt++;
546 svc_udp_sendto(struct svc_rqst *rqstp)
550 error = svc_sendto(rqstp, &rqstp->rq_res);
551 if (error == -ECONNREFUSED)
552 /* ICMP error on earlier request. */
553 error = svc_sendto(rqstp, &rqstp->rq_res);
558 static void svc_udp_prep_reply_hdr(struct svc_rqst *rqstp)
562 static int svc_udp_has_wspace(struct svc_xprt *xprt)
564 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
565 struct svc_serv *serv = xprt->xpt_server;
566 unsigned long required;
569 * Set the SOCK_NOSPACE flag before checking the available
572 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
573 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
574 if (required*2 > sock_wspace(svsk->sk_sk))
576 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
580 static struct svc_xprt *svc_udp_accept(struct svc_xprt *xprt)
586 static struct svc_xprt *svc_udp_create(struct svc_serv *serv,
587 struct sockaddr *sa, int salen,
590 return svc_create_socket(serv, IPPROTO_UDP, sa, salen, flags);
593 static struct svc_xprt_ops svc_udp_ops = {
594 .xpo_create = svc_udp_create,
595 .xpo_recvfrom = svc_udp_recvfrom,
596 .xpo_sendto = svc_udp_sendto,
597 .xpo_release_rqst = svc_release_skb,
598 .xpo_detach = svc_sock_detach,
599 .xpo_free = svc_sock_free,
600 .xpo_prep_reply_hdr = svc_udp_prep_reply_hdr,
601 .xpo_has_wspace = svc_udp_has_wspace,
602 .xpo_accept = svc_udp_accept,
605 static struct svc_xprt_class svc_udp_class = {
607 .xcl_owner = THIS_MODULE,
608 .xcl_ops = &svc_udp_ops,
609 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
612 static void svc_udp_init(struct svc_sock *svsk, struct svc_serv *serv)
617 svc_xprt_init(&svc_udp_class, &svsk->sk_xprt, serv);
618 clear_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
619 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
620 svsk->sk_sk->sk_write_space = svc_write_space;
622 /* initialise setting must have enough space to
623 * receive and respond to one request.
624 * svc_udp_recvfrom will re-adjust if necessary
626 svc_sock_setbufsize(svsk->sk_sock,
627 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
628 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
630 /* data might have come in before data_ready set up */
631 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
632 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
636 /* make sure we get destination address info */
637 svsk->sk_sock->ops->setsockopt(svsk->sk_sock, IPPROTO_IP, IP_PKTINFO,
638 (char __user *)&one, sizeof(one));
643 * A data_ready event on a listening socket means there's a connection
644 * pending. Do not use state_change as a substitute for it.
646 static void svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
648 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
650 dprintk("svc: socket %p TCP (listen) state change %d\n",
654 * This callback may called twice when a new connection
655 * is established as a child socket inherits everything
656 * from a parent LISTEN socket.
657 * 1) data_ready method of the parent socket will be called
658 * when one of child sockets become ESTABLISHED.
659 * 2) data_ready method of the child socket may be called
660 * when it receives data before the socket is accepted.
661 * In case of 2, we should ignore it silently.
663 if (sk->sk_state == TCP_LISTEN) {
665 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
666 svc_xprt_enqueue(&svsk->sk_xprt);
668 printk("svc: socket %p: no user data\n", sk);
671 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
672 wake_up_interruptible_all(sk->sk_sleep);
676 * A state change on a connected socket means it's dying or dead.
678 static void svc_tcp_state_change(struct sock *sk)
680 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
682 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
683 sk, sk->sk_state, sk->sk_user_data);
686 printk("svc: socket %p: no user data\n", sk);
688 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
689 svc_xprt_enqueue(&svsk->sk_xprt);
691 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
692 wake_up_interruptible_all(sk->sk_sleep);
695 static void svc_tcp_data_ready(struct sock *sk, int count)
697 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
699 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
700 sk, sk->sk_user_data);
702 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
703 svc_xprt_enqueue(&svsk->sk_xprt);
705 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
706 wake_up_interruptible(sk->sk_sleep);
710 * Accept a TCP connection
712 static struct svc_xprt *svc_tcp_accept(struct svc_xprt *xprt)
714 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
715 struct sockaddr_storage addr;
716 struct sockaddr *sin = (struct sockaddr *) &addr;
717 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
718 struct socket *sock = svsk->sk_sock;
719 struct socket *newsock;
720 struct svc_sock *newsvsk;
722 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
724 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
728 clear_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
729 err = kernel_accept(sock, &newsock, O_NONBLOCK);
732 printk(KERN_WARNING "%s: no more sockets!\n",
734 else if (err != -EAGAIN && net_ratelimit())
735 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
736 serv->sv_name, -err);
739 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
741 err = kernel_getpeername(newsock, sin, &slen);
744 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
745 serv->sv_name, -err);
746 goto failed; /* aborted connection or whatever */
749 /* Ideally, we would want to reject connections from unauthorized
750 * hosts here, but when we get encryption, the IP of the host won't
751 * tell us anything. For now just warn about unpriv connections.
753 if (!svc_port_is_privileged(sin)) {
755 "%s: connect from unprivileged port: %s\n",
757 __svc_print_addr(sin, buf, sizeof(buf)));
759 dprintk("%s: connect from %s\n", serv->sv_name,
760 __svc_print_addr(sin, buf, sizeof(buf)));
762 /* make sure that a write doesn't block forever when
765 newsock->sk->sk_sndtimeo = HZ*30;
767 if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
768 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
770 svc_xprt_set_remote(&newsvsk->sk_xprt, sin, slen);
771 err = kernel_getsockname(newsock, sin, &slen);
772 if (unlikely(err < 0)) {
773 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err);
774 slen = offsetof(struct sockaddr, sa_data);
776 svc_xprt_set_local(&newsvsk->sk_xprt, sin, slen);
779 serv->sv_stats->nettcpconn++;
781 return &newsvsk->sk_xprt;
784 sock_release(newsock);
789 * Receive data from a TCP socket.
791 static int svc_tcp_recvfrom(struct svc_rqst *rqstp)
793 struct svc_sock *svsk =
794 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
795 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
800 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
801 svsk, test_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags),
802 test_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags),
803 test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags));
805 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
806 /* sndbuf needs to have room for one request
807 * per thread, otherwise we can stall even when the
808 * network isn't a bottleneck.
810 * We count all threads rather than threads in a
811 * particular pool, which provides an upper bound
812 * on the number of threads which will access the socket.
814 * rcvbuf just needs to be able to hold a few requests.
815 * Normally they will be removed from the queue
816 * as soon a a complete request arrives.
818 svc_sock_setbufsize(svsk->sk_sock,
819 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
820 3 * serv->sv_max_mesg);
822 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
824 /* Receive data. If we haven't got the record length yet, get
825 * the next four bytes. Otherwise try to gobble up as much as
826 * possible up to the complete record length.
828 if (svsk->sk_tcplen < sizeof(rpc_fraghdr)) {
829 int want = sizeof(rpc_fraghdr) - svsk->sk_tcplen;
832 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
834 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
836 svsk->sk_tcplen += len;
839 dprintk("svc: short recvfrom while reading record "
840 "length (%d of %d)\n", len, want);
841 svc_xprt_received(&svsk->sk_xprt);
842 return -EAGAIN; /* record header not complete */
845 svsk->sk_reclen = ntohl(svsk->sk_reclen);
846 if (!(svsk->sk_reclen & RPC_LAST_STREAM_FRAGMENT)) {
847 /* FIXME: technically, a record can be fragmented,
848 * and non-terminal fragments will not have the top
849 * bit set in the fragment length header.
850 * But apparently no known nfs clients send fragmented
853 printk(KERN_NOTICE "RPC: multiple fragments "
854 "per record not supported\n");
857 svsk->sk_reclen &= RPC_FRAGMENT_SIZE_MASK;
858 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
859 if (svsk->sk_reclen > serv->sv_max_mesg) {
861 printk(KERN_NOTICE "RPC: "
862 "fragment too large: 0x%08lx\n",
863 (unsigned long)svsk->sk_reclen);
868 /* Check whether enough data is available */
869 len = svc_recv_available(svsk);
873 if (len < svsk->sk_reclen) {
874 dprintk("svc: incomplete TCP record (%d of %d)\n",
875 len, svsk->sk_reclen);
876 svc_xprt_received(&svsk->sk_xprt);
877 return -EAGAIN; /* record not complete */
879 len = svsk->sk_reclen;
880 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
883 vec[0] = rqstp->rq_arg.head[0];
887 vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
888 vec[pnum].iov_len = PAGE_SIZE;
892 rqstp->rq_respages = &rqstp->rq_pages[pnum];
894 /* Now receive data */
895 len = svc_recvfrom(rqstp, vec, pnum, len);
899 dprintk("svc: TCP complete record (%d bytes)\n", len);
900 rqstp->rq_arg.len = len;
901 rqstp->rq_arg.page_base = 0;
902 if (len <= rqstp->rq_arg.head[0].iov_len) {
903 rqstp->rq_arg.head[0].iov_len = len;
904 rqstp->rq_arg.page_len = 0;
906 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
909 rqstp->rq_xprt_ctxt = NULL;
910 rqstp->rq_prot = IPPROTO_TCP;
912 /* Reset TCP read info */
916 svc_xprt_copy_addrs(rqstp, &svsk->sk_xprt);
917 svc_xprt_received(&svsk->sk_xprt);
919 serv->sv_stats->nettcpcnt++;
924 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
928 if (len == -EAGAIN) {
929 dprintk("RPC: TCP recvfrom got EAGAIN\n");
930 svc_xprt_received(&svsk->sk_xprt);
932 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
933 svsk->sk_xprt.xpt_server->sv_name, -len);
941 * Send out data on TCP socket.
943 static int svc_tcp_sendto(struct svc_rqst *rqstp)
945 struct xdr_buf *xbufp = &rqstp->rq_res;
949 /* Set up the first element of the reply kvec.
950 * Any other kvecs that may be in use have been taken
951 * care of by the server implementation itself.
953 reclen = htonl(0x80000000|((xbufp->len ) - 4));
954 memcpy(xbufp->head[0].iov_base, &reclen, 4);
956 if (test_bit(XPT_DEAD, &rqstp->rq_xprt->xpt_flags))
959 sent = svc_sendto(rqstp, &rqstp->rq_res);
960 if (sent != xbufp->len) {
962 "rpc-srv/tcp: %s: %s %d when sending %d bytes "
963 "- shutting down socket\n",
964 rqstp->rq_xprt->xpt_server->sv_name,
965 (sent<0)?"got error":"sent only",
967 set_bit(XPT_CLOSE, &rqstp->rq_xprt->xpt_flags);
968 svc_xprt_enqueue(rqstp->rq_xprt);
975 * Setup response header. TCP has a 4B record length field.
977 static void svc_tcp_prep_reply_hdr(struct svc_rqst *rqstp)
979 struct kvec *resv = &rqstp->rq_res.head[0];
981 /* tcp needs a space for the record length... */
985 static int svc_tcp_has_wspace(struct svc_xprt *xprt)
987 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
988 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
993 * Set the SOCK_NOSPACE flag before checking the available
996 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
997 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
998 wspace = sk_stream_wspace(svsk->sk_sk);
1000 if (wspace < sk_stream_min_wspace(svsk->sk_sk))
1002 if (required * 2 > wspace)
1005 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1009 static struct svc_xprt *svc_tcp_create(struct svc_serv *serv,
1010 struct sockaddr *sa, int salen,
1013 return svc_create_socket(serv, IPPROTO_TCP, sa, salen, flags);
1016 static struct svc_xprt_ops svc_tcp_ops = {
1017 .xpo_create = svc_tcp_create,
1018 .xpo_recvfrom = svc_tcp_recvfrom,
1019 .xpo_sendto = svc_tcp_sendto,
1020 .xpo_release_rqst = svc_release_skb,
1021 .xpo_detach = svc_tcp_sock_detach,
1022 .xpo_free = svc_sock_free,
1023 .xpo_prep_reply_hdr = svc_tcp_prep_reply_hdr,
1024 .xpo_has_wspace = svc_tcp_has_wspace,
1025 .xpo_accept = svc_tcp_accept,
1028 static struct svc_xprt_class svc_tcp_class = {
1030 .xcl_owner = THIS_MODULE,
1031 .xcl_ops = &svc_tcp_ops,
1032 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1035 void svc_init_xprt_sock(void)
1037 svc_reg_xprt_class(&svc_tcp_class);
1038 svc_reg_xprt_class(&svc_udp_class);
1041 void svc_cleanup_xprt_sock(void)
1043 svc_unreg_xprt_class(&svc_tcp_class);
1044 svc_unreg_xprt_class(&svc_udp_class);
1047 static void svc_tcp_init(struct svc_sock *svsk, struct svc_serv *serv)
1049 struct sock *sk = svsk->sk_sk;
1051 svc_xprt_init(&svc_tcp_class, &svsk->sk_xprt, serv);
1052 set_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
1053 if (sk->sk_state == TCP_LISTEN) {
1054 dprintk("setting up TCP socket for listening\n");
1055 set_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags);
1056 sk->sk_data_ready = svc_tcp_listen_data_ready;
1057 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1059 dprintk("setting up TCP socket for reading\n");
1060 sk->sk_state_change = svc_tcp_state_change;
1061 sk->sk_data_ready = svc_tcp_data_ready;
1062 sk->sk_write_space = svc_write_space;
1064 svsk->sk_reclen = 0;
1065 svsk->sk_tcplen = 0;
1067 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1069 /* initialise setting must have enough space to
1070 * receive and respond to one request.
1071 * svc_tcp_recvfrom will re-adjust if necessary
1073 svc_sock_setbufsize(svsk->sk_sock,
1074 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
1075 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
1077 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1078 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1079 if (sk->sk_state != TCP_ESTABLISHED)
1080 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1084 void svc_sock_update_bufs(struct svc_serv *serv)
1087 * The number of server threads has changed. Update
1088 * rcvbuf and sndbuf accordingly on all sockets
1090 struct list_head *le;
1092 spin_lock_bh(&serv->sv_lock);
1093 list_for_each(le, &serv->sv_permsocks) {
1094 struct svc_sock *svsk =
1095 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1096 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1098 list_for_each(le, &serv->sv_tempsocks) {
1099 struct svc_sock *svsk =
1100 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1101 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1103 spin_unlock_bh(&serv->sv_lock);
1105 EXPORT_SYMBOL(svc_sock_update_bufs);
1108 * Initialize socket for RPC use and create svc_sock struct
1109 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1111 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1112 struct socket *sock,
1113 int *errp, int flags)
1115 struct svc_sock *svsk;
1117 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1120 dprintk("svc: svc_setup_socket %p\n", sock);
1121 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1128 /* Register socket with portmapper */
1129 if (*errp >= 0 && pmap_register)
1130 *errp = svc_register(serv, inet->sk_protocol,
1131 ntohs(inet_sk(inet)->sport));
1138 inet->sk_user_data = svsk;
1139 svsk->sk_sock = sock;
1141 svsk->sk_ostate = inet->sk_state_change;
1142 svsk->sk_odata = inet->sk_data_ready;
1143 svsk->sk_owspace = inet->sk_write_space;
1145 /* Initialize the socket */
1146 if (sock->type == SOCK_DGRAM)
1147 svc_udp_init(svsk, serv);
1149 svc_tcp_init(svsk, serv);
1152 * We start one listener per sv_serv. We want AF_INET
1153 * requests to be automatically shunted to our AF_INET6
1154 * listener using a mapped IPv4 address. Make sure
1155 * no-one starts an equivalent IPv4 listener, which
1156 * would steal our incoming connections.
1159 if (serv->sv_family == AF_INET6)
1160 kernel_setsockopt(sock, SOL_IPV6, IPV6_V6ONLY,
1161 (char *)&val, sizeof(val));
1163 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1169 int svc_addsock(struct svc_serv *serv,
1174 struct socket *so = sockfd_lookup(fd, &err);
1175 struct svc_sock *svsk = NULL;
1179 if (so->sk->sk_family != AF_INET)
1180 err = -EAFNOSUPPORT;
1181 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1182 so->sk->sk_protocol != IPPROTO_UDP)
1183 err = -EPROTONOSUPPORT;
1184 else if (so->state > SS_UNCONNECTED)
1187 if (!try_module_get(THIS_MODULE))
1190 svsk = svc_setup_socket(serv, so, &err,
1193 struct sockaddr_storage addr;
1194 struct sockaddr *sin = (struct sockaddr *)&addr;
1196 if (kernel_getsockname(svsk->sk_sock, sin, &salen) == 0)
1197 svc_xprt_set_local(&svsk->sk_xprt, sin, salen);
1198 clear_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1199 spin_lock_bh(&serv->sv_lock);
1200 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_permsocks);
1201 spin_unlock_bh(&serv->sv_lock);
1202 svc_xprt_received(&svsk->sk_xprt);
1205 module_put(THIS_MODULE);
1211 return one_sock_name(name_return, svsk);
1213 EXPORT_SYMBOL_GPL(svc_addsock);
1216 * Create socket for RPC service.
1218 static struct svc_xprt *svc_create_socket(struct svc_serv *serv,
1220 struct sockaddr *sin, int len,
1223 struct svc_sock *svsk;
1224 struct socket *sock;
1227 struct sockaddr_storage addr;
1228 struct sockaddr *newsin = (struct sockaddr *)&addr;
1230 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
1232 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1233 serv->sv_program->pg_name, protocol,
1234 __svc_print_addr(sin, buf, sizeof(buf)));
1236 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1237 printk(KERN_WARNING "svc: only UDP and TCP "
1238 "sockets supported\n");
1239 return ERR_PTR(-EINVAL);
1241 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1243 error = sock_create_kern(sin->sa_family, type, protocol, &sock);
1245 return ERR_PTR(error);
1247 svc_reclassify_socket(sock);
1249 if (type == SOCK_STREAM)
1250 sock->sk->sk_reuse = 1; /* allow address reuse */
1251 error = kernel_bind(sock, sin, len);
1256 error = kernel_getsockname(sock, newsin, &newlen);
1260 if (protocol == IPPROTO_TCP) {
1261 if ((error = kernel_listen(sock, 64)) < 0)
1265 if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1266 svc_xprt_set_local(&svsk->sk_xprt, newsin, newlen);
1267 return (struct svc_xprt *)svsk;
1271 dprintk("svc: svc_create_socket error = %d\n", -error);
1273 return ERR_PTR(error);
1277 * Detach the svc_sock from the socket so that no
1278 * more callbacks occur.
1280 static void svc_sock_detach(struct svc_xprt *xprt)
1282 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1283 struct sock *sk = svsk->sk_sk;
1285 dprintk("svc: svc_sock_detach(%p)\n", svsk);
1287 /* put back the old socket callbacks */
1288 sk->sk_state_change = svsk->sk_ostate;
1289 sk->sk_data_ready = svsk->sk_odata;
1290 sk->sk_write_space = svsk->sk_owspace;
1292 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1293 wake_up_interruptible(sk->sk_sleep);
1297 * Disconnect the socket, and reset the callbacks
1299 static void svc_tcp_sock_detach(struct svc_xprt *xprt)
1301 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1303 dprintk("svc: svc_tcp_sock_detach(%p)\n", svsk);
1305 svc_sock_detach(xprt);
1307 if (!test_bit(XPT_LISTENER, &xprt->xpt_flags))
1308 kernel_sock_shutdown(svsk->sk_sock, SHUT_RDWR);
1312 * Free the svc_sock's socket resources and the svc_sock itself.
1314 static void svc_sock_free(struct svc_xprt *xprt)
1316 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1317 dprintk("svc: svc_sock_free(%p)\n", svsk);
1319 if (svsk->sk_sock->file)
1320 sockfd_put(svsk->sk_sock);
1322 sock_release(svsk->sk_sock);