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 one_sock_name(char *buf, struct svc_sock *svsk)
247 switch(svsk->sk_sk->sk_family) {
249 len = sprintf(buf, "ipv4 %s %pI4 %d\n",
250 svsk->sk_sk->sk_protocol == IPPROTO_UDP ?
252 &inet_sk(svsk->sk_sk)->rcv_saddr,
253 inet_sk(svsk->sk_sk)->num);
256 len = sprintf(buf, "*unknown-%d*\n",
257 svsk->sk_sk->sk_family);
263 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
265 struct svc_sock *svsk, *closesk = NULL;
270 spin_lock_bh(&serv->sv_lock);
271 list_for_each_entry(svsk, &serv->sv_permsocks, sk_xprt.xpt_list) {
272 int onelen = one_sock_name(buf+len, svsk);
273 if (toclose && strcmp(toclose, buf+len) == 0)
278 spin_unlock_bh(&serv->sv_lock);
280 /* Should unregister with portmap, but you cannot
281 * unregister just one protocol...
283 svc_close_xprt(&closesk->sk_xprt);
288 EXPORT_SYMBOL_GPL(svc_sock_names);
291 * Check input queue length
293 static int svc_recv_available(struct svc_sock *svsk)
295 struct socket *sock = svsk->sk_sock;
298 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
300 return (err >= 0)? avail : err;
304 * Generic recvfrom routine.
306 static int svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr,
309 struct svc_sock *svsk =
310 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
311 struct msghdr msg = {
312 .msg_flags = MSG_DONTWAIT,
316 rqstp->rq_xprt_hlen = 0;
318 len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
321 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
322 svsk, iov[0].iov_base, iov[0].iov_len, len);
327 * Set socket snd and rcv buffer lengths
329 static void svc_sock_setbufsize(struct socket *sock, unsigned int snd,
334 oldfs = get_fs(); set_fs(KERNEL_DS);
335 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
336 (char*)&snd, sizeof(snd));
337 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
338 (char*)&rcv, sizeof(rcv));
340 /* sock_setsockopt limits use to sysctl_?mem_max,
341 * which isn't acceptable. Until that is made conditional
342 * on not having CAP_SYS_RESOURCE or similar, we go direct...
343 * DaveM said I could!
346 sock->sk->sk_sndbuf = snd * 2;
347 sock->sk->sk_rcvbuf = rcv * 2;
348 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
349 release_sock(sock->sk);
353 * INET callback when data has been received on the socket.
355 static void svc_udp_data_ready(struct sock *sk, int count)
357 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
360 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
362 test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
363 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
364 svc_xprt_enqueue(&svsk->sk_xprt);
366 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
367 wake_up_interruptible(sk->sk_sleep);
371 * INET callback when space is newly available on the socket.
373 static void svc_write_space(struct sock *sk)
375 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
378 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
379 svsk, sk, test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
380 svc_xprt_enqueue(&svsk->sk_xprt);
383 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
384 dprintk("RPC svc_write_space: someone sleeping on %p\n",
386 wake_up_interruptible(sk->sk_sleep);
391 * Copy the UDP datagram's destination address to the rqstp structure.
392 * The 'destination' address in this case is the address to which the
393 * peer sent the datagram, i.e. our local address. For multihomed
394 * hosts, this can change from msg to msg. Note that only the IP
395 * address changes, the port number should remain the same.
397 static void svc_udp_get_dest_address(struct svc_rqst *rqstp,
400 struct svc_sock *svsk =
401 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
402 switch (svsk->sk_sk->sk_family) {
404 struct in_pktinfo *pki = CMSG_DATA(cmh);
405 rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
409 struct in6_pktinfo *pki = CMSG_DATA(cmh);
410 ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
417 * Receive a datagram from a UDP socket.
419 static int svc_udp_recvfrom(struct svc_rqst *rqstp)
421 struct svc_sock *svsk =
422 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
423 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
427 long all[SVC_PKTINFO_SPACE / sizeof(long)];
429 struct cmsghdr *cmh = &buffer.hdr;
431 struct msghdr msg = {
432 .msg_name = svc_addr(rqstp),
434 .msg_controllen = sizeof(buffer),
435 .msg_flags = MSG_DONTWAIT,
438 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
439 /* udp sockets need large rcvbuf as all pending
440 * requests are still in that buffer. sndbuf must
441 * also be large enough that there is enough space
442 * for one reply per thread. We count all threads
443 * rather than threads in a particular pool, which
444 * provides an upper bound on the number of threads
445 * which will access the socket.
447 svc_sock_setbufsize(svsk->sk_sock,
448 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
449 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
451 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
453 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
454 0, 0, MSG_PEEK | MSG_DONTWAIT);
456 skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err);
459 if (err != -EAGAIN) {
460 /* possibly an icmp error */
461 dprintk("svc: recvfrom returned error %d\n", -err);
462 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
464 svc_xprt_received(&svsk->sk_xprt);
467 len = svc_addr_len(svc_addr(rqstp));
470 rqstp->rq_addrlen = len;
471 if (skb->tstamp.tv64 == 0) {
472 skb->tstamp = ktime_get_real();
473 /* Don't enable netstamp, sunrpc doesn't
474 need that much accuracy */
476 svsk->sk_sk->sk_stamp = skb->tstamp;
477 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* there may be more data... */
480 * Maybe more packets - kick another thread ASAP.
482 svc_xprt_received(&svsk->sk_xprt);
484 len = skb->len - sizeof(struct udphdr);
485 rqstp->rq_arg.len = len;
487 rqstp->rq_prot = IPPROTO_UDP;
489 if (cmh->cmsg_level != IPPROTO_IP ||
490 cmh->cmsg_type != IP_PKTINFO) {
492 printk("rpcsvc: received unknown control message:"
494 cmh->cmsg_level, cmh->cmsg_type);
495 skb_free_datagram(svsk->sk_sk, skb);
498 svc_udp_get_dest_address(rqstp, cmh);
500 if (skb_is_nonlinear(skb)) {
501 /* we have to copy */
503 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
506 skb_free_datagram(svsk->sk_sk, skb);
510 skb_free_datagram(svsk->sk_sk, skb);
512 /* we can use it in-place */
513 rqstp->rq_arg.head[0].iov_base = skb->data +
514 sizeof(struct udphdr);
515 rqstp->rq_arg.head[0].iov_len = len;
516 if (skb_checksum_complete(skb)) {
517 skb_free_datagram(svsk->sk_sk, skb);
520 rqstp->rq_xprt_ctxt = skb;
523 rqstp->rq_arg.page_base = 0;
524 if (len <= rqstp->rq_arg.head[0].iov_len) {
525 rqstp->rq_arg.head[0].iov_len = len;
526 rqstp->rq_arg.page_len = 0;
527 rqstp->rq_respages = rqstp->rq_pages+1;
529 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
530 rqstp->rq_respages = rqstp->rq_pages + 1 +
531 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
535 serv->sv_stats->netudpcnt++;
541 svc_udp_sendto(struct svc_rqst *rqstp)
545 error = svc_sendto(rqstp, &rqstp->rq_res);
546 if (error == -ECONNREFUSED)
547 /* ICMP error on earlier request. */
548 error = svc_sendto(rqstp, &rqstp->rq_res);
553 static void svc_udp_prep_reply_hdr(struct svc_rqst *rqstp)
557 static int svc_udp_has_wspace(struct svc_xprt *xprt)
559 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
560 struct svc_serv *serv = xprt->xpt_server;
561 unsigned long required;
564 * Set the SOCK_NOSPACE flag before checking the available
567 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
568 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
569 if (required*2 > sock_wspace(svsk->sk_sk))
571 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
575 static struct svc_xprt *svc_udp_accept(struct svc_xprt *xprt)
581 static struct svc_xprt *svc_udp_create(struct svc_serv *serv,
582 struct sockaddr *sa, int salen,
585 return svc_create_socket(serv, IPPROTO_UDP, sa, salen, flags);
588 static struct svc_xprt_ops svc_udp_ops = {
589 .xpo_create = svc_udp_create,
590 .xpo_recvfrom = svc_udp_recvfrom,
591 .xpo_sendto = svc_udp_sendto,
592 .xpo_release_rqst = svc_release_skb,
593 .xpo_detach = svc_sock_detach,
594 .xpo_free = svc_sock_free,
595 .xpo_prep_reply_hdr = svc_udp_prep_reply_hdr,
596 .xpo_has_wspace = svc_udp_has_wspace,
597 .xpo_accept = svc_udp_accept,
600 static struct svc_xprt_class svc_udp_class = {
602 .xcl_owner = THIS_MODULE,
603 .xcl_ops = &svc_udp_ops,
604 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
607 static void svc_udp_init(struct svc_sock *svsk, struct svc_serv *serv)
612 svc_xprt_init(&svc_udp_class, &svsk->sk_xprt, serv);
613 clear_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
614 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
615 svsk->sk_sk->sk_write_space = svc_write_space;
617 /* initialise setting must have enough space to
618 * receive and respond to one request.
619 * svc_udp_recvfrom will re-adjust if necessary
621 svc_sock_setbufsize(svsk->sk_sock,
622 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
623 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
625 /* data might have come in before data_ready set up */
626 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
627 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
631 /* make sure we get destination address info */
632 svsk->sk_sock->ops->setsockopt(svsk->sk_sock, IPPROTO_IP, IP_PKTINFO,
633 (char __user *)&one, sizeof(one));
638 * A data_ready event on a listening socket means there's a connection
639 * pending. Do not use state_change as a substitute for it.
641 static void svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
643 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
645 dprintk("svc: socket %p TCP (listen) state change %d\n",
649 * This callback may called twice when a new connection
650 * is established as a child socket inherits everything
651 * from a parent LISTEN socket.
652 * 1) data_ready method of the parent socket will be called
653 * when one of child sockets become ESTABLISHED.
654 * 2) data_ready method of the child socket may be called
655 * when it receives data before the socket is accepted.
656 * In case of 2, we should ignore it silently.
658 if (sk->sk_state == TCP_LISTEN) {
660 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
661 svc_xprt_enqueue(&svsk->sk_xprt);
663 printk("svc: socket %p: no user data\n", sk);
666 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
667 wake_up_interruptible_all(sk->sk_sleep);
671 * A state change on a connected socket means it's dying or dead.
673 static void svc_tcp_state_change(struct sock *sk)
675 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
677 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
678 sk, sk->sk_state, sk->sk_user_data);
681 printk("svc: socket %p: no user data\n", sk);
683 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
684 svc_xprt_enqueue(&svsk->sk_xprt);
686 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
687 wake_up_interruptible_all(sk->sk_sleep);
690 static void svc_tcp_data_ready(struct sock *sk, int count)
692 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
694 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
695 sk, sk->sk_user_data);
697 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
698 svc_xprt_enqueue(&svsk->sk_xprt);
700 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
701 wake_up_interruptible(sk->sk_sleep);
705 * Accept a TCP connection
707 static struct svc_xprt *svc_tcp_accept(struct svc_xprt *xprt)
709 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
710 struct sockaddr_storage addr;
711 struct sockaddr *sin = (struct sockaddr *) &addr;
712 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
713 struct socket *sock = svsk->sk_sock;
714 struct socket *newsock;
715 struct svc_sock *newsvsk;
717 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
719 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
723 clear_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
724 err = kernel_accept(sock, &newsock, O_NONBLOCK);
727 printk(KERN_WARNING "%s: no more sockets!\n",
729 else if (err != -EAGAIN && net_ratelimit())
730 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
731 serv->sv_name, -err);
734 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
736 err = kernel_getpeername(newsock, sin, &slen);
739 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
740 serv->sv_name, -err);
741 goto failed; /* aborted connection or whatever */
744 /* Ideally, we would want to reject connections from unauthorized
745 * hosts here, but when we get encryption, the IP of the host won't
746 * tell us anything. For now just warn about unpriv connections.
748 if (!svc_port_is_privileged(sin)) {
750 "%s: connect from unprivileged port: %s\n",
752 __svc_print_addr(sin, buf, sizeof(buf)));
754 dprintk("%s: connect from %s\n", serv->sv_name,
755 __svc_print_addr(sin, buf, sizeof(buf)));
757 /* make sure that a write doesn't block forever when
760 newsock->sk->sk_sndtimeo = HZ*30;
762 if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
763 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
765 svc_xprt_set_remote(&newsvsk->sk_xprt, sin, slen);
766 err = kernel_getsockname(newsock, sin, &slen);
767 if (unlikely(err < 0)) {
768 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err);
769 slen = offsetof(struct sockaddr, sa_data);
771 svc_xprt_set_local(&newsvsk->sk_xprt, sin, slen);
774 serv->sv_stats->nettcpconn++;
776 return &newsvsk->sk_xprt;
779 sock_release(newsock);
784 * Receive data from a TCP socket.
786 static int svc_tcp_recvfrom(struct svc_rqst *rqstp)
788 struct svc_sock *svsk =
789 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
790 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
795 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
796 svsk, test_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags),
797 test_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags),
798 test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags));
800 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
801 /* sndbuf needs to have room for one request
802 * per thread, otherwise we can stall even when the
803 * network isn't a bottleneck.
805 * We count all threads rather than threads in a
806 * particular pool, which provides an upper bound
807 * on the number of threads which will access the socket.
809 * rcvbuf just needs to be able to hold a few requests.
810 * Normally they will be removed from the queue
811 * as soon a a complete request arrives.
813 svc_sock_setbufsize(svsk->sk_sock,
814 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
815 3 * serv->sv_max_mesg);
817 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
819 /* Receive data. If we haven't got the record length yet, get
820 * the next four bytes. Otherwise try to gobble up as much as
821 * possible up to the complete record length.
823 if (svsk->sk_tcplen < sizeof(rpc_fraghdr)) {
824 int want = sizeof(rpc_fraghdr) - svsk->sk_tcplen;
827 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
829 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
831 svsk->sk_tcplen += len;
834 dprintk("svc: short recvfrom while reading record "
835 "length (%d of %d)\n", len, want);
836 svc_xprt_received(&svsk->sk_xprt);
837 return -EAGAIN; /* record header not complete */
840 svsk->sk_reclen = ntohl(svsk->sk_reclen);
841 if (!(svsk->sk_reclen & RPC_LAST_STREAM_FRAGMENT)) {
842 /* FIXME: technically, a record can be fragmented,
843 * and non-terminal fragments will not have the top
844 * bit set in the fragment length header.
845 * But apparently no known nfs clients send fragmented
848 printk(KERN_NOTICE "RPC: multiple fragments "
849 "per record not supported\n");
852 svsk->sk_reclen &= RPC_FRAGMENT_SIZE_MASK;
853 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
854 if (svsk->sk_reclen > serv->sv_max_mesg) {
856 printk(KERN_NOTICE "RPC: "
857 "fragment too large: 0x%08lx\n",
858 (unsigned long)svsk->sk_reclen);
863 /* Check whether enough data is available */
864 len = svc_recv_available(svsk);
868 if (len < svsk->sk_reclen) {
869 dprintk("svc: incomplete TCP record (%d of %d)\n",
870 len, svsk->sk_reclen);
871 svc_xprt_received(&svsk->sk_xprt);
872 return -EAGAIN; /* record not complete */
874 len = svsk->sk_reclen;
875 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
878 vec[0] = rqstp->rq_arg.head[0];
882 vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
883 vec[pnum].iov_len = PAGE_SIZE;
887 rqstp->rq_respages = &rqstp->rq_pages[pnum];
889 /* Now receive data */
890 len = svc_recvfrom(rqstp, vec, pnum, len);
894 dprintk("svc: TCP complete record (%d bytes)\n", len);
895 rqstp->rq_arg.len = len;
896 rqstp->rq_arg.page_base = 0;
897 if (len <= rqstp->rq_arg.head[0].iov_len) {
898 rqstp->rq_arg.head[0].iov_len = len;
899 rqstp->rq_arg.page_len = 0;
901 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
904 rqstp->rq_xprt_ctxt = NULL;
905 rqstp->rq_prot = IPPROTO_TCP;
907 /* Reset TCP read info */
911 svc_xprt_copy_addrs(rqstp, &svsk->sk_xprt);
912 svc_xprt_received(&svsk->sk_xprt);
914 serv->sv_stats->nettcpcnt++;
919 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
923 if (len == -EAGAIN) {
924 dprintk("RPC: TCP recvfrom got EAGAIN\n");
925 svc_xprt_received(&svsk->sk_xprt);
927 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
928 svsk->sk_xprt.xpt_server->sv_name, -len);
936 * Send out data on TCP socket.
938 static int svc_tcp_sendto(struct svc_rqst *rqstp)
940 struct xdr_buf *xbufp = &rqstp->rq_res;
944 /* Set up the first element of the reply kvec.
945 * Any other kvecs that may be in use have been taken
946 * care of by the server implementation itself.
948 reclen = htonl(0x80000000|((xbufp->len ) - 4));
949 memcpy(xbufp->head[0].iov_base, &reclen, 4);
951 if (test_bit(XPT_DEAD, &rqstp->rq_xprt->xpt_flags))
954 sent = svc_sendto(rqstp, &rqstp->rq_res);
955 if (sent != xbufp->len) {
957 "rpc-srv/tcp: %s: %s %d when sending %d bytes "
958 "- shutting down socket\n",
959 rqstp->rq_xprt->xpt_server->sv_name,
960 (sent<0)?"got error":"sent only",
962 set_bit(XPT_CLOSE, &rqstp->rq_xprt->xpt_flags);
963 svc_xprt_enqueue(rqstp->rq_xprt);
970 * Setup response header. TCP has a 4B record length field.
972 static void svc_tcp_prep_reply_hdr(struct svc_rqst *rqstp)
974 struct kvec *resv = &rqstp->rq_res.head[0];
976 /* tcp needs a space for the record length... */
980 static int svc_tcp_has_wspace(struct svc_xprt *xprt)
982 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
983 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
988 * Set the SOCK_NOSPACE flag before checking the available
991 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
992 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
993 wspace = sk_stream_wspace(svsk->sk_sk);
995 if (wspace < sk_stream_min_wspace(svsk->sk_sk))
997 if (required * 2 > wspace)
1000 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1004 static struct svc_xprt *svc_tcp_create(struct svc_serv *serv,
1005 struct sockaddr *sa, int salen,
1008 return svc_create_socket(serv, IPPROTO_TCP, sa, salen, flags);
1011 static struct svc_xprt_ops svc_tcp_ops = {
1012 .xpo_create = svc_tcp_create,
1013 .xpo_recvfrom = svc_tcp_recvfrom,
1014 .xpo_sendto = svc_tcp_sendto,
1015 .xpo_release_rqst = svc_release_skb,
1016 .xpo_detach = svc_tcp_sock_detach,
1017 .xpo_free = svc_sock_free,
1018 .xpo_prep_reply_hdr = svc_tcp_prep_reply_hdr,
1019 .xpo_has_wspace = svc_tcp_has_wspace,
1020 .xpo_accept = svc_tcp_accept,
1023 static struct svc_xprt_class svc_tcp_class = {
1025 .xcl_owner = THIS_MODULE,
1026 .xcl_ops = &svc_tcp_ops,
1027 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1030 void svc_init_xprt_sock(void)
1032 svc_reg_xprt_class(&svc_tcp_class);
1033 svc_reg_xprt_class(&svc_udp_class);
1036 void svc_cleanup_xprt_sock(void)
1038 svc_unreg_xprt_class(&svc_tcp_class);
1039 svc_unreg_xprt_class(&svc_udp_class);
1042 static void svc_tcp_init(struct svc_sock *svsk, struct svc_serv *serv)
1044 struct sock *sk = svsk->sk_sk;
1046 svc_xprt_init(&svc_tcp_class, &svsk->sk_xprt, serv);
1047 set_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
1048 if (sk->sk_state == TCP_LISTEN) {
1049 dprintk("setting up TCP socket for listening\n");
1050 set_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags);
1051 sk->sk_data_ready = svc_tcp_listen_data_ready;
1052 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1054 dprintk("setting up TCP socket for reading\n");
1055 sk->sk_state_change = svc_tcp_state_change;
1056 sk->sk_data_ready = svc_tcp_data_ready;
1057 sk->sk_write_space = svc_write_space;
1059 svsk->sk_reclen = 0;
1060 svsk->sk_tcplen = 0;
1062 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1064 /* initialise setting must have enough space to
1065 * receive and respond to one request.
1066 * svc_tcp_recvfrom will re-adjust if necessary
1068 svc_sock_setbufsize(svsk->sk_sock,
1069 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
1070 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
1072 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1073 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1074 if (sk->sk_state != TCP_ESTABLISHED)
1075 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1079 void svc_sock_update_bufs(struct svc_serv *serv)
1082 * The number of server threads has changed. Update
1083 * rcvbuf and sndbuf accordingly on all sockets
1085 struct list_head *le;
1087 spin_lock_bh(&serv->sv_lock);
1088 list_for_each(le, &serv->sv_permsocks) {
1089 struct svc_sock *svsk =
1090 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1091 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1093 list_for_each(le, &serv->sv_tempsocks) {
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 spin_unlock_bh(&serv->sv_lock);
1100 EXPORT_SYMBOL_GPL(svc_sock_update_bufs);
1103 * Initialize socket for RPC use and create svc_sock struct
1104 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1106 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1107 struct socket *sock,
1108 int *errp, int flags)
1110 struct svc_sock *svsk;
1112 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1114 dprintk("svc: svc_setup_socket %p\n", sock);
1115 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1122 /* Register socket with portmapper */
1123 if (*errp >= 0 && pmap_register)
1124 *errp = svc_register(serv, inet->sk_family, inet->sk_protocol,
1125 ntohs(inet_sk(inet)->sport));
1132 inet->sk_user_data = svsk;
1133 svsk->sk_sock = sock;
1135 svsk->sk_ostate = inet->sk_state_change;
1136 svsk->sk_odata = inet->sk_data_ready;
1137 svsk->sk_owspace = inet->sk_write_space;
1139 /* Initialize the socket */
1140 if (sock->type == SOCK_DGRAM)
1141 svc_udp_init(svsk, serv);
1143 svc_tcp_init(svsk, serv);
1145 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1151 int svc_addsock(struct svc_serv *serv,
1156 struct socket *so = sockfd_lookup(fd, &err);
1157 struct svc_sock *svsk = NULL;
1161 if (so->sk->sk_family != AF_INET)
1162 err = -EAFNOSUPPORT;
1163 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1164 so->sk->sk_protocol != IPPROTO_UDP)
1165 err = -EPROTONOSUPPORT;
1166 else if (so->state > SS_UNCONNECTED)
1169 if (!try_module_get(THIS_MODULE))
1172 svsk = svc_setup_socket(serv, so, &err,
1175 struct sockaddr_storage addr;
1176 struct sockaddr *sin = (struct sockaddr *)&addr;
1178 if (kernel_getsockname(svsk->sk_sock, sin, &salen) == 0)
1179 svc_xprt_set_local(&svsk->sk_xprt, sin, salen);
1180 clear_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1181 spin_lock_bh(&serv->sv_lock);
1182 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_permsocks);
1183 spin_unlock_bh(&serv->sv_lock);
1184 svc_xprt_received(&svsk->sk_xprt);
1187 module_put(THIS_MODULE);
1193 return one_sock_name(name_return, svsk);
1195 EXPORT_SYMBOL_GPL(svc_addsock);
1198 * Create socket for RPC service.
1200 static struct svc_xprt *svc_create_socket(struct svc_serv *serv,
1202 struct sockaddr *sin, int len,
1205 struct svc_sock *svsk;
1206 struct socket *sock;
1209 struct sockaddr_storage addr;
1210 struct sockaddr *newsin = (struct sockaddr *)&addr;
1214 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
1216 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1217 serv->sv_program->pg_name, protocol,
1218 __svc_print_addr(sin, buf, sizeof(buf)));
1220 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1221 printk(KERN_WARNING "svc: only UDP and TCP "
1222 "sockets supported\n");
1223 return ERR_PTR(-EINVAL);
1226 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1227 switch (sin->sa_family) {
1235 return ERR_PTR(-EINVAL);
1238 error = sock_create_kern(family, type, protocol, &sock);
1240 return ERR_PTR(error);
1242 svc_reclassify_socket(sock);
1245 * If this is an PF_INET6 listener, we want to avoid
1246 * getting requests from IPv4 remotes. Those should
1247 * be shunted to a PF_INET listener via rpcbind.
1250 if (family == PF_INET6)
1251 kernel_setsockopt(sock, SOL_IPV6, IPV6_V6ONLY,
1252 (char *)&val, sizeof(val));
1254 if (type == SOCK_STREAM)
1255 sock->sk->sk_reuse = 1; /* allow address reuse */
1256 error = kernel_bind(sock, sin, len);
1261 error = kernel_getsockname(sock, newsin, &newlen);
1265 if (protocol == IPPROTO_TCP) {
1266 if ((error = kernel_listen(sock, 64)) < 0)
1270 if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1271 svc_xprt_set_local(&svsk->sk_xprt, newsin, newlen);
1272 return (struct svc_xprt *)svsk;
1276 dprintk("svc: svc_create_socket error = %d\n", -error);
1278 return ERR_PTR(error);
1282 * Detach the svc_sock from the socket so that no
1283 * more callbacks occur.
1285 static void svc_sock_detach(struct svc_xprt *xprt)
1287 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1288 struct sock *sk = svsk->sk_sk;
1290 dprintk("svc: svc_sock_detach(%p)\n", svsk);
1292 /* put back the old socket callbacks */
1293 sk->sk_state_change = svsk->sk_ostate;
1294 sk->sk_data_ready = svsk->sk_odata;
1295 sk->sk_write_space = svsk->sk_owspace;
1297 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1298 wake_up_interruptible(sk->sk_sleep);
1302 * Disconnect the socket, and reset the callbacks
1304 static void svc_tcp_sock_detach(struct svc_xprt *xprt)
1306 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1308 dprintk("svc: svc_tcp_sock_detach(%p)\n", svsk);
1310 svc_sock_detach(xprt);
1312 if (!test_bit(XPT_LISTENER, &xprt->xpt_flags))
1313 kernel_sock_shutdown(svsk->sk_sock, SHUT_RDWR);
1317 * Free the svc_sock's socket resources and the svc_sock itself.
1319 static void svc_sock_free(struct svc_xprt *xprt)
1321 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1322 dprintk("svc: svc_sock_free(%p)\n", svsk);
1324 if (svsk->sk_sock->file)
1325 sockfd_put(svsk->sk_sock);
1327 sock_release(svsk->sk_sock);