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_sock_free(struct svc_xprt *);
64 static struct svc_xprt *svc_create_socket(struct svc_serv *, int,
65 struct sockaddr *, int, int);
66 #ifdef CONFIG_DEBUG_LOCK_ALLOC
67 static struct lock_class_key svc_key[2];
68 static struct lock_class_key svc_slock_key[2];
70 static void svc_reclassify_socket(struct socket *sock)
72 struct sock *sk = sock->sk;
73 BUG_ON(sock_owned_by_user(sk));
74 switch (sk->sk_family) {
76 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFSD",
78 "sk_xprt.xpt_lock-AF_INET-NFSD",
83 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFSD",
85 "sk_xprt.xpt_lock-AF_INET6-NFSD",
94 static void svc_reclassify_socket(struct socket *sock)
100 * Release an skbuff after use
102 static void svc_release_skb(struct svc_rqst *rqstp)
104 struct sk_buff *skb = rqstp->rq_xprt_ctxt;
105 struct svc_deferred_req *dr = rqstp->rq_deferred;
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);
116 rqstp->rq_deferred = NULL;
121 union svc_pktinfo_u {
122 struct in_pktinfo pkti;
123 struct in6_pktinfo pkti6;
125 #define SVC_PKTINFO_SPACE \
126 CMSG_SPACE(sizeof(union svc_pktinfo_u))
128 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
130 struct svc_sock *svsk =
131 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
132 switch (svsk->sk_sk->sk_family) {
134 struct in_pktinfo *pki = CMSG_DATA(cmh);
136 cmh->cmsg_level = SOL_IP;
137 cmh->cmsg_type = IP_PKTINFO;
138 pki->ipi_ifindex = 0;
139 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr.addr.s_addr;
140 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
145 struct in6_pktinfo *pki = CMSG_DATA(cmh);
147 cmh->cmsg_level = SOL_IPV6;
148 cmh->cmsg_type = IPV6_PKTINFO;
149 pki->ipi6_ifindex = 0;
150 ipv6_addr_copy(&pki->ipi6_addr,
151 &rqstp->rq_daddr.addr6);
152 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
160 * Generic sendto routine
162 static int svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
164 struct svc_sock *svsk =
165 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
166 struct socket *sock = svsk->sk_sock;
170 long all[SVC_PKTINFO_SPACE / sizeof(long)];
172 struct cmsghdr *cmh = &buffer.hdr;
176 struct page **ppage = xdr->pages;
177 size_t base = xdr->page_base;
178 unsigned int pglen = xdr->page_len;
179 unsigned int flags = MSG_MORE;
180 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
184 if (rqstp->rq_prot == IPPROTO_UDP) {
185 struct msghdr msg = {
186 .msg_name = &rqstp->rq_addr,
187 .msg_namelen = rqstp->rq_addrlen,
189 .msg_controllen = sizeof(buffer),
190 .msg_flags = MSG_MORE,
193 svc_set_cmsg_data(rqstp, cmh);
195 if (sock_sendmsg(sock, &msg, 0) < 0)
200 if (slen == xdr->head[0].iov_len)
202 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0,
203 xdr->head[0].iov_len, flags);
204 if (len != xdr->head[0].iov_len)
206 slen -= xdr->head[0].iov_len;
211 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
215 result = kernel_sendpage(sock, *ppage, base, size, flags);
222 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
227 if (xdr->tail[0].iov_len) {
228 result = kernel_sendpage(sock, rqstp->rq_respages[0],
229 ((unsigned long)xdr->tail[0].iov_base)
231 xdr->tail[0].iov_len, 0);
237 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
238 svsk, xdr->head[0].iov_base, xdr->head[0].iov_len,
239 xdr->len, len, svc_print_addr(rqstp, buf, sizeof(buf)));
245 * Report socket names for nfsdfs
247 static int one_sock_name(char *buf, struct svc_sock *svsk)
251 switch(svsk->sk_sk->sk_family) {
253 len = sprintf(buf, "ipv4 %s %u.%u.%u.%u %d\n",
254 svsk->sk_sk->sk_protocol==IPPROTO_UDP?
256 NIPQUAD(inet_sk(svsk->sk_sk)->rcv_saddr),
257 inet_sk(svsk->sk_sk)->num);
260 len = sprintf(buf, "*unknown-%d*\n",
261 svsk->sk_sk->sk_family);
267 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
269 struct svc_sock *svsk, *closesk = NULL;
274 spin_lock_bh(&serv->sv_lock);
275 list_for_each_entry(svsk, &serv->sv_permsocks, sk_xprt.xpt_list) {
276 int onelen = one_sock_name(buf+len, svsk);
277 if (toclose && strcmp(toclose, buf+len) == 0)
282 spin_unlock_bh(&serv->sv_lock);
284 /* Should unregister with portmap, but you cannot
285 * unregister just one protocol...
287 svc_close_xprt(&closesk->sk_xprt);
292 EXPORT_SYMBOL(svc_sock_names);
295 * Check input queue length
297 static int svc_recv_available(struct svc_sock *svsk)
299 struct socket *sock = svsk->sk_sock;
302 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
304 return (err >= 0)? avail : err;
308 * Generic recvfrom routine.
310 static int svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr,
313 struct svc_sock *svsk =
314 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
315 struct msghdr msg = {
316 .msg_flags = MSG_DONTWAIT,
320 rqstp->rq_xprt_hlen = 0;
322 len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
325 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
326 svsk, iov[0].iov_base, iov[0].iov_len, len);
331 * Set socket snd and rcv buffer lengths
333 static void svc_sock_setbufsize(struct socket *sock, unsigned int snd,
338 oldfs = get_fs(); set_fs(KERNEL_DS);
339 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
340 (char*)&snd, sizeof(snd));
341 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
342 (char*)&rcv, sizeof(rcv));
344 /* sock_setsockopt limits use to sysctl_?mem_max,
345 * which isn't acceptable. Until that is made conditional
346 * on not having CAP_SYS_RESOURCE or similar, we go direct...
347 * DaveM said I could!
350 sock->sk->sk_sndbuf = snd * 2;
351 sock->sk->sk_rcvbuf = rcv * 2;
352 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
353 release_sock(sock->sk);
357 * INET callback when data has been received on the socket.
359 static void svc_udp_data_ready(struct sock *sk, int count)
361 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
364 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
366 test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
367 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
368 svc_xprt_enqueue(&svsk->sk_xprt);
370 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
371 wake_up_interruptible(sk->sk_sleep);
375 * INET callback when space is newly available on the socket.
377 static void svc_write_space(struct sock *sk)
379 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
382 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
383 svsk, sk, test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
384 svc_xprt_enqueue(&svsk->sk_xprt);
387 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
388 dprintk("RPC svc_write_space: someone sleeping on %p\n",
390 wake_up_interruptible(sk->sk_sleep);
395 * Copy the UDP datagram's destination address to the rqstp structure.
396 * The 'destination' address in this case is the address to which the
397 * peer sent the datagram, i.e. our local address. For multihomed
398 * hosts, this can change from msg to msg. Note that only the IP
399 * address changes, the port number should remain the same.
401 static void svc_udp_get_dest_address(struct svc_rqst *rqstp,
404 struct svc_sock *svsk =
405 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
406 switch (svsk->sk_sk->sk_family) {
408 struct in_pktinfo *pki = CMSG_DATA(cmh);
409 rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
413 struct in6_pktinfo *pki = CMSG_DATA(cmh);
414 ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
421 * Receive a datagram from a UDP socket.
423 static int svc_udp_recvfrom(struct svc_rqst *rqstp)
425 struct svc_sock *svsk =
426 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
427 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
431 long all[SVC_PKTINFO_SPACE / sizeof(long)];
433 struct cmsghdr *cmh = &buffer.hdr;
435 struct msghdr msg = {
436 .msg_name = svc_addr(rqstp),
438 .msg_controllen = sizeof(buffer),
439 .msg_flags = MSG_DONTWAIT,
442 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
443 /* udp sockets need large rcvbuf as all pending
444 * requests are still in that buffer. sndbuf must
445 * also be large enough that there is enough space
446 * for one reply per thread. We count all threads
447 * rather than threads in a particular pool, which
448 * provides an upper bound on the number of threads
449 * which will access the socket.
451 svc_sock_setbufsize(svsk->sk_sock,
452 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
453 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
455 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
457 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
458 0, 0, MSG_PEEK | MSG_DONTWAIT);
460 skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err);
463 if (err != -EAGAIN) {
464 /* possibly an icmp error */
465 dprintk("svc: recvfrom returned error %d\n", -err);
466 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
468 svc_xprt_received(&svsk->sk_xprt);
471 len = svc_addr_len(svc_addr(rqstp));
474 rqstp->rq_addrlen = len;
475 if (skb->tstamp.tv64 == 0) {
476 skb->tstamp = ktime_get_real();
477 /* Don't enable netstamp, sunrpc doesn't
478 need that much accuracy */
480 svsk->sk_sk->sk_stamp = skb->tstamp;
481 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* there may be more data... */
484 * Maybe more packets - kick another thread ASAP.
486 svc_xprt_received(&svsk->sk_xprt);
488 len = skb->len - sizeof(struct udphdr);
489 rqstp->rq_arg.len = len;
491 rqstp->rq_prot = IPPROTO_UDP;
493 if (cmh->cmsg_level != IPPROTO_IP ||
494 cmh->cmsg_type != IP_PKTINFO) {
496 printk("rpcsvc: received unknown control message:"
498 cmh->cmsg_level, cmh->cmsg_type);
499 skb_free_datagram(svsk->sk_sk, skb);
502 svc_udp_get_dest_address(rqstp, cmh);
504 if (skb_is_nonlinear(skb)) {
505 /* we have to copy */
507 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
510 skb_free_datagram(svsk->sk_sk, skb);
514 skb_free_datagram(svsk->sk_sk, skb);
516 /* we can use it in-place */
517 rqstp->rq_arg.head[0].iov_base = skb->data +
518 sizeof(struct udphdr);
519 rqstp->rq_arg.head[0].iov_len = len;
520 if (skb_checksum_complete(skb)) {
521 skb_free_datagram(svsk->sk_sk, skb);
524 rqstp->rq_xprt_ctxt = skb;
527 rqstp->rq_arg.page_base = 0;
528 if (len <= rqstp->rq_arg.head[0].iov_len) {
529 rqstp->rq_arg.head[0].iov_len = len;
530 rqstp->rq_arg.page_len = 0;
531 rqstp->rq_respages = rqstp->rq_pages+1;
533 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
534 rqstp->rq_respages = rqstp->rq_pages + 1 +
535 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
539 serv->sv_stats->netudpcnt++;
545 svc_udp_sendto(struct svc_rqst *rqstp)
549 error = svc_sendto(rqstp, &rqstp->rq_res);
550 if (error == -ECONNREFUSED)
551 /* ICMP error on earlier request. */
552 error = svc_sendto(rqstp, &rqstp->rq_res);
557 static void svc_udp_prep_reply_hdr(struct svc_rqst *rqstp)
561 static int svc_udp_has_wspace(struct svc_xprt *xprt)
563 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
564 struct svc_serv *serv = xprt->xpt_server;
565 unsigned long required;
568 * Set the SOCK_NOSPACE flag before checking the available
571 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
572 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
573 if (required*2 > sock_wspace(svsk->sk_sk))
575 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
579 static struct svc_xprt *svc_udp_accept(struct svc_xprt *xprt)
585 static struct svc_xprt *svc_udp_create(struct svc_serv *serv,
586 struct sockaddr *sa, int salen,
589 return svc_create_socket(serv, IPPROTO_UDP, sa, salen, flags);
592 static struct svc_xprt_ops svc_udp_ops = {
593 .xpo_create = svc_udp_create,
594 .xpo_recvfrom = svc_udp_recvfrom,
595 .xpo_sendto = svc_udp_sendto,
596 .xpo_release_rqst = svc_release_skb,
597 .xpo_detach = svc_sock_detach,
598 .xpo_free = svc_sock_free,
599 .xpo_prep_reply_hdr = svc_udp_prep_reply_hdr,
600 .xpo_has_wspace = svc_udp_has_wspace,
601 .xpo_accept = svc_udp_accept,
604 static struct svc_xprt_class svc_udp_class = {
606 .xcl_owner = THIS_MODULE,
607 .xcl_ops = &svc_udp_ops,
608 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
611 static void svc_udp_init(struct svc_sock *svsk, struct svc_serv *serv)
616 svc_xprt_init(&svc_udp_class, &svsk->sk_xprt, serv);
617 clear_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
618 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
619 svsk->sk_sk->sk_write_space = svc_write_space;
621 /* initialise setting must have enough space to
622 * receive and respond to one request.
623 * svc_udp_recvfrom will re-adjust if necessary
625 svc_sock_setbufsize(svsk->sk_sock,
626 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
627 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
629 /* data might have come in before data_ready set up */
630 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
631 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
635 /* make sure we get destination address info */
636 svsk->sk_sock->ops->setsockopt(svsk->sk_sock, IPPROTO_IP, IP_PKTINFO,
637 (char __user *)&one, sizeof(one));
642 * A data_ready event on a listening socket means there's a connection
643 * pending. Do not use state_change as a substitute for it.
645 static void svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
647 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
649 dprintk("svc: socket %p TCP (listen) state change %d\n",
653 * This callback may called twice when a new connection
654 * is established as a child socket inherits everything
655 * from a parent LISTEN socket.
656 * 1) data_ready method of the parent socket will be called
657 * when one of child sockets become ESTABLISHED.
658 * 2) data_ready method of the child socket may be called
659 * when it receives data before the socket is accepted.
660 * In case of 2, we should ignore it silently.
662 if (sk->sk_state == TCP_LISTEN) {
664 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
665 svc_xprt_enqueue(&svsk->sk_xprt);
667 printk("svc: socket %p: no user data\n", sk);
670 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
671 wake_up_interruptible_all(sk->sk_sleep);
675 * A state change on a connected socket means it's dying or dead.
677 static void svc_tcp_state_change(struct sock *sk)
679 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
681 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
682 sk, sk->sk_state, sk->sk_user_data);
685 printk("svc: socket %p: no user data\n", sk);
687 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
688 svc_xprt_enqueue(&svsk->sk_xprt);
690 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
691 wake_up_interruptible_all(sk->sk_sleep);
694 static void svc_tcp_data_ready(struct sock *sk, int count)
696 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
698 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
699 sk, sk->sk_user_data);
701 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
702 svc_xprt_enqueue(&svsk->sk_xprt);
704 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
705 wake_up_interruptible(sk->sk_sleep);
709 * Accept a TCP connection
711 static struct svc_xprt *svc_tcp_accept(struct svc_xprt *xprt)
713 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
714 struct sockaddr_storage addr;
715 struct sockaddr *sin = (struct sockaddr *) &addr;
716 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
717 struct socket *sock = svsk->sk_sock;
718 struct socket *newsock;
719 struct svc_sock *newsvsk;
721 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
723 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
727 clear_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
728 err = kernel_accept(sock, &newsock, O_NONBLOCK);
731 printk(KERN_WARNING "%s: no more sockets!\n",
733 else if (err != -EAGAIN && net_ratelimit())
734 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
735 serv->sv_name, -err);
738 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
740 err = kernel_getpeername(newsock, sin, &slen);
743 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
744 serv->sv_name, -err);
745 goto failed; /* aborted connection or whatever */
748 /* Ideally, we would want to reject connections from unauthorized
749 * hosts here, but when we get encryption, the IP of the host won't
750 * tell us anything. For now just warn about unpriv connections.
752 if (!svc_port_is_privileged(sin)) {
754 "%s: connect from unprivileged port: %s\n",
756 __svc_print_addr(sin, buf, sizeof(buf)));
758 dprintk("%s: connect from %s\n", serv->sv_name,
759 __svc_print_addr(sin, buf, sizeof(buf)));
761 /* make sure that a write doesn't block forever when
764 newsock->sk->sk_sndtimeo = HZ*30;
766 if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
767 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
769 svc_xprt_set_remote(&newsvsk->sk_xprt, sin, slen);
770 err = kernel_getsockname(newsock, sin, &slen);
771 if (unlikely(err < 0)) {
772 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err);
773 slen = offsetof(struct sockaddr, sa_data);
775 svc_xprt_set_local(&newsvsk->sk_xprt, sin, slen);
778 serv->sv_stats->nettcpconn++;
780 return &newsvsk->sk_xprt;
783 sock_release(newsock);
788 * Receive data from a TCP socket.
790 static int svc_tcp_recvfrom(struct svc_rqst *rqstp)
792 struct svc_sock *svsk =
793 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
794 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
799 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
800 svsk, test_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags),
801 test_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags),
802 test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags));
804 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
805 /* sndbuf needs to have room for one request
806 * per thread, otherwise we can stall even when the
807 * network isn't a bottleneck.
809 * We count all threads rather than threads in a
810 * particular pool, which provides an upper bound
811 * on the number of threads which will access the socket.
813 * rcvbuf just needs to be able to hold a few requests.
814 * Normally they will be removed from the queue
815 * as soon a a complete request arrives.
817 svc_sock_setbufsize(svsk->sk_sock,
818 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
819 3 * serv->sv_max_mesg);
821 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
823 /* Receive data. If we haven't got the record length yet, get
824 * the next four bytes. Otherwise try to gobble up as much as
825 * possible up to the complete record length.
827 if (svsk->sk_tcplen < sizeof(rpc_fraghdr)) {
828 int want = sizeof(rpc_fraghdr) - svsk->sk_tcplen;
831 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
833 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
835 svsk->sk_tcplen += len;
838 dprintk("svc: short recvfrom while reading record "
839 "length (%d of %d)\n", len, want);
840 svc_xprt_received(&svsk->sk_xprt);
841 return -EAGAIN; /* record header not complete */
844 svsk->sk_reclen = ntohl(svsk->sk_reclen);
845 if (!(svsk->sk_reclen & RPC_LAST_STREAM_FRAGMENT)) {
846 /* FIXME: technically, a record can be fragmented,
847 * and non-terminal fragments will not have the top
848 * bit set in the fragment length header.
849 * But apparently no known nfs clients send fragmented
852 printk(KERN_NOTICE "RPC: multiple fragments "
853 "per record not supported\n");
856 svsk->sk_reclen &= RPC_FRAGMENT_SIZE_MASK;
857 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
858 if (svsk->sk_reclen > serv->sv_max_mesg) {
860 printk(KERN_NOTICE "RPC: "
861 "fragment too large: 0x%08lx\n",
862 (unsigned long)svsk->sk_reclen);
867 /* Check whether enough data is available */
868 len = svc_recv_available(svsk);
872 if (len < svsk->sk_reclen) {
873 dprintk("svc: incomplete TCP record (%d of %d)\n",
874 len, svsk->sk_reclen);
875 svc_xprt_received(&svsk->sk_xprt);
876 return -EAGAIN; /* record not complete */
878 len = svsk->sk_reclen;
879 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
882 vec[0] = rqstp->rq_arg.head[0];
886 vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
887 vec[pnum].iov_len = PAGE_SIZE;
891 rqstp->rq_respages = &rqstp->rq_pages[pnum];
893 /* Now receive data */
894 len = svc_recvfrom(rqstp, vec, pnum, len);
898 dprintk("svc: TCP complete record (%d bytes)\n", len);
899 rqstp->rq_arg.len = len;
900 rqstp->rq_arg.page_base = 0;
901 if (len <= rqstp->rq_arg.head[0].iov_len) {
902 rqstp->rq_arg.head[0].iov_len = len;
903 rqstp->rq_arg.page_len = 0;
905 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
908 rqstp->rq_xprt_ctxt = NULL;
909 rqstp->rq_prot = IPPROTO_TCP;
911 /* Reset TCP read info */
915 svc_xprt_copy_addrs(rqstp, &svsk->sk_xprt);
916 svc_xprt_received(&svsk->sk_xprt);
918 serv->sv_stats->nettcpcnt++;
923 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
927 if (len == -EAGAIN) {
928 dprintk("RPC: TCP recvfrom got EAGAIN\n");
929 svc_xprt_received(&svsk->sk_xprt);
931 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
932 svsk->sk_xprt.xpt_server->sv_name, -len);
940 * Send out data on TCP socket.
942 static int svc_tcp_sendto(struct svc_rqst *rqstp)
944 struct xdr_buf *xbufp = &rqstp->rq_res;
948 /* Set up the first element of the reply kvec.
949 * Any other kvecs that may be in use have been taken
950 * care of by the server implementation itself.
952 reclen = htonl(0x80000000|((xbufp->len ) - 4));
953 memcpy(xbufp->head[0].iov_base, &reclen, 4);
955 if (test_bit(XPT_DEAD, &rqstp->rq_xprt->xpt_flags))
958 sent = svc_sendto(rqstp, &rqstp->rq_res);
959 if (sent != xbufp->len) {
961 "rpc-srv/tcp: %s: %s %d when sending %d bytes "
962 "- shutting down socket\n",
963 rqstp->rq_xprt->xpt_server->sv_name,
964 (sent<0)?"got error":"sent only",
966 set_bit(XPT_CLOSE, &rqstp->rq_xprt->xpt_flags);
967 svc_xprt_enqueue(rqstp->rq_xprt);
974 * Setup response header. TCP has a 4B record length field.
976 static void svc_tcp_prep_reply_hdr(struct svc_rqst *rqstp)
978 struct kvec *resv = &rqstp->rq_res.head[0];
980 /* tcp needs a space for the record length... */
984 static int svc_tcp_has_wspace(struct svc_xprt *xprt)
986 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
987 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
992 * Set the SOCK_NOSPACE flag before checking the available
995 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
996 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
997 wspace = sk_stream_wspace(svsk->sk_sk);
999 if (wspace < sk_stream_min_wspace(svsk->sk_sk))
1001 if (required * 2 > wspace)
1004 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1008 static struct svc_xprt *svc_tcp_create(struct svc_serv *serv,
1009 struct sockaddr *sa, int salen,
1012 return svc_create_socket(serv, IPPROTO_TCP, sa, salen, flags);
1015 static struct svc_xprt_ops svc_tcp_ops = {
1016 .xpo_create = svc_tcp_create,
1017 .xpo_recvfrom = svc_tcp_recvfrom,
1018 .xpo_sendto = svc_tcp_sendto,
1019 .xpo_release_rqst = svc_release_skb,
1020 .xpo_detach = svc_sock_detach,
1021 .xpo_free = svc_sock_free,
1022 .xpo_prep_reply_hdr = svc_tcp_prep_reply_hdr,
1023 .xpo_has_wspace = svc_tcp_has_wspace,
1024 .xpo_accept = svc_tcp_accept,
1027 static struct svc_xprt_class svc_tcp_class = {
1029 .xcl_owner = THIS_MODULE,
1030 .xcl_ops = &svc_tcp_ops,
1031 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1034 void svc_init_xprt_sock(void)
1036 svc_reg_xprt_class(&svc_tcp_class);
1037 svc_reg_xprt_class(&svc_udp_class);
1040 void svc_cleanup_xprt_sock(void)
1042 svc_unreg_xprt_class(&svc_tcp_class);
1043 svc_unreg_xprt_class(&svc_udp_class);
1046 static void svc_tcp_init(struct svc_sock *svsk, struct svc_serv *serv)
1048 struct sock *sk = svsk->sk_sk;
1050 svc_xprt_init(&svc_tcp_class, &svsk->sk_xprt, serv);
1051 set_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
1052 if (sk->sk_state == TCP_LISTEN) {
1053 dprintk("setting up TCP socket for listening\n");
1054 set_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags);
1055 sk->sk_data_ready = svc_tcp_listen_data_ready;
1056 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1058 dprintk("setting up TCP socket for reading\n");
1059 sk->sk_state_change = svc_tcp_state_change;
1060 sk->sk_data_ready = svc_tcp_data_ready;
1061 sk->sk_write_space = svc_write_space;
1063 svsk->sk_reclen = 0;
1064 svsk->sk_tcplen = 0;
1066 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1068 /* initialise setting must have enough space to
1069 * receive and respond to one request.
1070 * svc_tcp_recvfrom will re-adjust if necessary
1072 svc_sock_setbufsize(svsk->sk_sock,
1073 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
1074 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
1076 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1077 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1078 if (sk->sk_state != TCP_ESTABLISHED)
1079 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1083 void svc_sock_update_bufs(struct svc_serv *serv)
1086 * The number of server threads has changed. Update
1087 * rcvbuf and sndbuf accordingly on all sockets
1089 struct list_head *le;
1091 spin_lock_bh(&serv->sv_lock);
1092 list_for_each(le, &serv->sv_permsocks) {
1093 struct svc_sock *svsk =
1094 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1095 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1097 list_for_each(le, &serv->sv_tempsocks) {
1098 struct svc_sock *svsk =
1099 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1100 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1102 spin_unlock_bh(&serv->sv_lock);
1104 EXPORT_SYMBOL(svc_sock_update_bufs);
1107 * Initialize socket for RPC use and create svc_sock struct
1108 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1110 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1111 struct socket *sock,
1112 int *errp, int flags)
1114 struct svc_sock *svsk;
1116 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1119 dprintk("svc: svc_setup_socket %p\n", sock);
1120 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1127 /* Register socket with portmapper */
1128 if (*errp >= 0 && pmap_register)
1129 *errp = svc_register(serv, inet->sk_protocol,
1130 ntohs(inet_sk(inet)->sport));
1137 inet->sk_user_data = svsk;
1138 svsk->sk_sock = sock;
1140 svsk->sk_ostate = inet->sk_state_change;
1141 svsk->sk_odata = inet->sk_data_ready;
1142 svsk->sk_owspace = inet->sk_write_space;
1144 /* Initialize the socket */
1145 if (sock->type == SOCK_DGRAM)
1146 svc_udp_init(svsk, serv);
1148 svc_tcp_init(svsk, serv);
1151 * We start one listener per sv_serv. We want AF_INET
1152 * requests to be automatically shunted to our AF_INET6
1153 * listener using a mapped IPv4 address. Make sure
1154 * no-one starts an equivalent IPv4 listener, which
1155 * would steal our incoming connections.
1158 if (serv->sv_family == AF_INET6)
1159 kernel_setsockopt(sock, SOL_IPV6, IPV6_V6ONLY,
1160 (char *)&val, sizeof(val));
1162 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1168 int svc_addsock(struct svc_serv *serv,
1173 struct socket *so = sockfd_lookup(fd, &err);
1174 struct svc_sock *svsk = NULL;
1178 if (so->sk->sk_family != AF_INET)
1179 err = -EAFNOSUPPORT;
1180 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1181 so->sk->sk_protocol != IPPROTO_UDP)
1182 err = -EPROTONOSUPPORT;
1183 else if (so->state > SS_UNCONNECTED)
1186 svsk = svc_setup_socket(serv, so, &err, SVC_SOCK_DEFAULTS);
1188 struct sockaddr_storage addr;
1189 struct sockaddr *sin = (struct sockaddr *)&addr;
1191 if (kernel_getsockname(svsk->sk_sock, sin, &salen) == 0)
1192 svc_xprt_set_local(&svsk->sk_xprt, sin, salen);
1193 clear_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1194 spin_lock_bh(&serv->sv_lock);
1195 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_permsocks);
1196 spin_unlock_bh(&serv->sv_lock);
1197 svc_xprt_received(&svsk->sk_xprt);
1205 return one_sock_name(name_return, svsk);
1207 EXPORT_SYMBOL_GPL(svc_addsock);
1210 * Create socket for RPC service.
1212 static struct svc_xprt *svc_create_socket(struct svc_serv *serv,
1214 struct sockaddr *sin, int len,
1217 struct svc_sock *svsk;
1218 struct socket *sock;
1221 struct sockaddr_storage addr;
1222 struct sockaddr *newsin = (struct sockaddr *)&addr;
1224 RPC_IFDEBUG(char buf[RPC_MAX_ADDRBUFLEN]);
1226 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1227 serv->sv_program->pg_name, protocol,
1228 __svc_print_addr(sin, buf, sizeof(buf)));
1230 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1231 printk(KERN_WARNING "svc: only UDP and TCP "
1232 "sockets supported\n");
1233 return ERR_PTR(-EINVAL);
1235 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1237 error = sock_create_kern(sin->sa_family, type, protocol, &sock);
1239 return ERR_PTR(error);
1241 svc_reclassify_socket(sock);
1243 if (type == SOCK_STREAM)
1244 sock->sk->sk_reuse = 1; /* allow address reuse */
1245 error = kernel_bind(sock, sin, len);
1250 error = kernel_getsockname(sock, newsin, &newlen);
1254 if (protocol == IPPROTO_TCP) {
1255 if ((error = kernel_listen(sock, 64)) < 0)
1259 if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1260 svc_xprt_set_local(&svsk->sk_xprt, newsin, newlen);
1261 return (struct svc_xprt *)svsk;
1265 dprintk("svc: svc_create_socket error = %d\n", -error);
1267 return ERR_PTR(error);
1271 * Detach the svc_sock from the socket so that no
1272 * more callbacks occur.
1274 static void svc_sock_detach(struct svc_xprt *xprt)
1276 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1277 struct sock *sk = svsk->sk_sk;
1279 dprintk("svc: svc_sock_detach(%p)\n", svsk);
1281 /* put back the old socket callbacks */
1282 sk->sk_state_change = svsk->sk_ostate;
1283 sk->sk_data_ready = svsk->sk_odata;
1284 sk->sk_write_space = svsk->sk_owspace;
1288 * Free the svc_sock's socket resources and the svc_sock itself.
1290 static void svc_sock_free(struct svc_xprt *xprt)
1292 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1293 dprintk("svc: svc_sock_free(%p)\n", svsk);
1295 if (svsk->sk_sock->file)
1296 sockfd_put(svsk->sk_sock);
1298 sock_release(svsk->sk_sock);