1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/net/sunrpc/svcsock.c
5 * These are the RPC server socket internals.
7 * The server scheduling algorithm does not always distribute the load
8 * evenly when servicing a single client. May need to modify the
9 * svc_xprt_enqueue procedure...
11 * TCP support is largely untested and may be a little slow. The problem
12 * is that we currently do two separate recvfrom's, one for the 4-byte
13 * record length, and the second for the actual record. This could possibly
14 * be improved by always reading a minimum size of around 100 bytes and
15 * tucking any superfluous bytes away in a temporary store. Still, that
16 * leaves write requests out in the rain. An alternative may be to peek at
17 * the first skb in the queue, and if it matches the next TCP sequence
18 * number, to extract the record marker. Yuck.
20 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
23 #include <linux/kernel.h>
24 #include <linux/sched.h>
25 #include <linux/module.h>
26 #include <linux/errno.h>
27 #include <linux/fcntl.h>
28 #include <linux/net.h>
30 #include <linux/inet.h>
31 #include <linux/udp.h>
32 #include <linux/tcp.h>
33 #include <linux/unistd.h>
34 #include <linux/slab.h>
35 #include <linux/netdevice.h>
36 #include <linux/skbuff.h>
37 #include <linux/file.h>
38 #include <linux/freezer.h>
40 #include <net/checksum.h>
45 #include <net/tcp_states.h>
46 #include <linux/uaccess.h>
47 #include <linux/highmem.h>
48 #include <asm/ioctls.h>
50 #include <linux/sunrpc/types.h>
51 #include <linux/sunrpc/clnt.h>
52 #include <linux/sunrpc/xdr.h>
53 #include <linux/sunrpc/msg_prot.h>
54 #include <linux/sunrpc/svcsock.h>
55 #include <linux/sunrpc/stats.h>
56 #include <linux/sunrpc/xprt.h>
58 #include <trace/events/sunrpc.h>
63 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
66 static struct svc_sock *svc_setup_socket(struct svc_serv *, struct socket *,
68 static int svc_udp_recvfrom(struct svc_rqst *);
69 static int svc_udp_sendto(struct svc_rqst *);
70 static void svc_sock_detach(struct svc_xprt *);
71 static void svc_tcp_sock_detach(struct svc_xprt *);
72 static void svc_sock_free(struct svc_xprt *);
74 static struct svc_xprt *svc_create_socket(struct svc_serv *, int,
75 struct net *, struct sockaddr *,
77 #ifdef CONFIG_DEBUG_LOCK_ALLOC
78 static struct lock_class_key svc_key[2];
79 static struct lock_class_key svc_slock_key[2];
81 static void svc_reclassify_socket(struct socket *sock)
83 struct sock *sk = sock->sk;
85 if (WARN_ON_ONCE(!sock_allow_reclassification(sk)))
88 switch (sk->sk_family) {
90 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFSD",
92 "sk_xprt.xpt_lock-AF_INET-NFSD",
97 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFSD",
99 "sk_xprt.xpt_lock-AF_INET6-NFSD",
108 static void svc_reclassify_socket(struct socket *sock)
114 * svc_tcp_release_rqst - Release transport-related resources
115 * @rqstp: request structure with resources to be released
118 static void svc_tcp_release_rqst(struct svc_rqst *rqstp)
123 * svc_udp_release_rqst - Release transport-related resources
124 * @rqstp: request structure with resources to be released
127 static void svc_udp_release_rqst(struct svc_rqst *rqstp)
129 struct sk_buff *skb = rqstp->rq_xprt_ctxt;
132 rqstp->rq_xprt_ctxt = NULL;
137 union svc_pktinfo_u {
138 struct in_pktinfo pkti;
139 struct in6_pktinfo pkti6;
141 #define SVC_PKTINFO_SPACE \
142 CMSG_SPACE(sizeof(union svc_pktinfo_u))
144 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
146 struct svc_sock *svsk =
147 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
148 switch (svsk->sk_sk->sk_family) {
150 struct in_pktinfo *pki = CMSG_DATA(cmh);
152 cmh->cmsg_level = SOL_IP;
153 cmh->cmsg_type = IP_PKTINFO;
154 pki->ipi_ifindex = 0;
155 pki->ipi_spec_dst.s_addr =
156 svc_daddr_in(rqstp)->sin_addr.s_addr;
157 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
162 struct in6_pktinfo *pki = CMSG_DATA(cmh);
163 struct sockaddr_in6 *daddr = svc_daddr_in6(rqstp);
165 cmh->cmsg_level = SOL_IPV6;
166 cmh->cmsg_type = IPV6_PKTINFO;
167 pki->ipi6_ifindex = daddr->sin6_scope_id;
168 pki->ipi6_addr = daddr->sin6_addr;
169 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
175 static int svc_sock_result_payload(struct svc_rqst *rqstp, unsigned int offset,
182 * Report socket names for nfsdfs
184 static int svc_one_sock_name(struct svc_sock *svsk, char *buf, int remaining)
186 const struct sock *sk = svsk->sk_sk;
187 const char *proto_name = sk->sk_protocol == IPPROTO_UDP ?
191 switch (sk->sk_family) {
193 len = snprintf(buf, remaining, "ipv4 %s %pI4 %d\n",
195 &inet_sk(sk)->inet_rcv_saddr,
196 inet_sk(sk)->inet_num);
198 #if IS_ENABLED(CONFIG_IPV6)
200 len = snprintf(buf, remaining, "ipv6 %s %pI6 %d\n",
202 &sk->sk_v6_rcv_saddr,
203 inet_sk(sk)->inet_num);
207 len = snprintf(buf, remaining, "*unknown-%d*\n",
211 if (len >= remaining) {
213 return -ENAMETOOLONG;
218 #if ARCH_IMPLEMENTS_FLUSH_DCACHE_PAGE
219 static void svc_flush_bvec(const struct bio_vec *bvec, size_t size, size_t seek)
221 struct bvec_iter bi = {
222 .bi_size = size + seek,
226 bvec_iter_advance(bvec, &bi, seek & PAGE_MASK);
227 for_each_bvec(bv, bvec, bi, bi)
228 flush_dcache_page(bv.bv_page);
231 static inline void svc_flush_bvec(const struct bio_vec *bvec, size_t size,
238 * Read from @rqstp's transport socket. The incoming message fills whole
239 * pages in @rqstp's rq_pages array until the last page of the message
240 * has been received into a partial page.
242 static ssize_t svc_tcp_read_msg(struct svc_rqst *rqstp, size_t buflen,
245 struct svc_sock *svsk =
246 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
247 struct bio_vec *bvec = rqstp->rq_bvec;
248 struct msghdr msg = { NULL };
253 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
255 for (i = 0, t = 0; t < buflen; i++, t += PAGE_SIZE) {
256 bvec[i].bv_page = rqstp->rq_pages[i];
257 bvec[i].bv_len = PAGE_SIZE;
258 bvec[i].bv_offset = 0;
260 rqstp->rq_respages = &rqstp->rq_pages[i];
261 rqstp->rq_next_page = rqstp->rq_respages + 1;
263 iov_iter_bvec(&msg.msg_iter, READ, bvec, i, buflen);
265 iov_iter_advance(&msg.msg_iter, seek);
268 len = sock_recvmsg(svsk->sk_sock, &msg, MSG_DONTWAIT);
270 svc_flush_bvec(bvec, len, seek);
272 /* If we read a full record, then assume there may be more
273 * data to read (stream based sockets only!)
276 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
282 * Set socket snd and rcv buffer lengths
284 static void svc_sock_setbufsize(struct svc_sock *svsk, unsigned int nreqs)
286 unsigned int max_mesg = svsk->sk_xprt.xpt_server->sv_max_mesg;
287 struct socket *sock = svsk->sk_sock;
289 nreqs = min(nreqs, INT_MAX / 2 / max_mesg);
292 sock->sk->sk_sndbuf = nreqs * max_mesg * 2;
293 sock->sk->sk_rcvbuf = nreqs * max_mesg * 2;
294 sock->sk->sk_write_space(sock->sk);
295 release_sock(sock->sk);
298 static void svc_sock_secure_port(struct svc_rqst *rqstp)
300 if (svc_port_is_privileged(svc_addr(rqstp)))
301 __set_bit(RQ_SECURE, &rqstp->rq_flags);
303 __clear_bit(RQ_SECURE, &rqstp->rq_flags);
307 * INET callback when data has been received on the socket.
309 static void svc_data_ready(struct sock *sk)
311 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
314 /* Refer to svc_setup_socket() for details. */
317 trace_svcsock_data_ready(&svsk->sk_xprt, 0);
318 if (!test_and_set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags))
319 svc_xprt_enqueue(&svsk->sk_xprt);
324 * INET callback when space is newly available on the socket.
326 static void svc_write_space(struct sock *sk)
328 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
331 /* Refer to svc_setup_socket() for details. */
333 trace_svcsock_write_space(&svsk->sk_xprt, 0);
334 svsk->sk_owspace(sk);
335 svc_xprt_enqueue(&svsk->sk_xprt);
339 static int svc_tcp_has_wspace(struct svc_xprt *xprt)
341 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
343 if (test_bit(XPT_LISTENER, &xprt->xpt_flags))
345 return !test_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
348 static void svc_tcp_kill_temp_xprt(struct svc_xprt *xprt)
350 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
352 sock_no_linger(svsk->sk_sock->sk);
356 * See net/ipv6/ip_sockglue.c : ip_cmsg_recv_pktinfo
358 static int svc_udp_get_dest_address4(struct svc_rqst *rqstp,
361 struct in_pktinfo *pki = CMSG_DATA(cmh);
362 struct sockaddr_in *daddr = svc_daddr_in(rqstp);
364 if (cmh->cmsg_type != IP_PKTINFO)
367 daddr->sin_family = AF_INET;
368 daddr->sin_addr.s_addr = pki->ipi_spec_dst.s_addr;
373 * See net/ipv6/datagram.c : ip6_datagram_recv_ctl
375 static int svc_udp_get_dest_address6(struct svc_rqst *rqstp,
378 struct in6_pktinfo *pki = CMSG_DATA(cmh);
379 struct sockaddr_in6 *daddr = svc_daddr_in6(rqstp);
381 if (cmh->cmsg_type != IPV6_PKTINFO)
384 daddr->sin6_family = AF_INET6;
385 daddr->sin6_addr = pki->ipi6_addr;
386 daddr->sin6_scope_id = pki->ipi6_ifindex;
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 int svc_udp_get_dest_address(struct svc_rqst *rqstp,
400 switch (cmh->cmsg_level) {
402 return svc_udp_get_dest_address4(rqstp, cmh);
404 return svc_udp_get_dest_address6(rqstp, cmh);
411 * svc_udp_recvfrom - Receive a datagram from a UDP socket.
412 * @rqstp: request structure into which to receive an RPC Call
414 * Called in a loop when XPT_DATA has been set.
417 * On success, the number of bytes in a received RPC Call, or
418 * %0 if a complete RPC Call message was not ready to return
420 static int svc_udp_recvfrom(struct svc_rqst *rqstp)
422 struct svc_sock *svsk =
423 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
424 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
428 long all[SVC_PKTINFO_SPACE / sizeof(long)];
430 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,
440 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
441 /* udp sockets need large rcvbuf as all pending
442 * requests are still in that buffer. sndbuf must
443 * also be large enough that there is enough space
444 * for one reply per thread. We count all threads
445 * rather than threads in a particular pool, which
446 * provides an upper bound on the number of threads
447 * which will access the socket.
449 svc_sock_setbufsize(svsk, serv->sv_nrthreads + 3);
451 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
452 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
453 0, 0, MSG_PEEK | MSG_DONTWAIT);
456 skb = skb_recv_udp(svsk->sk_sk, MSG_DONTWAIT, &err);
460 len = svc_addr_len(svc_addr(rqstp));
461 rqstp->rq_addrlen = len;
462 if (skb->tstamp == 0) {
463 skb->tstamp = ktime_get_real();
464 /* Don't enable netstamp, sunrpc doesn't
465 need that much accuracy */
467 sock_write_timestamp(svsk->sk_sk, skb->tstamp);
468 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* there may be more data... */
471 rqstp->rq_arg.len = len;
472 trace_svcsock_udp_recv(&svsk->sk_xprt, len);
474 rqstp->rq_prot = IPPROTO_UDP;
476 if (!svc_udp_get_dest_address(rqstp, cmh))
478 rqstp->rq_daddrlen = svc_addr_len(svc_daddr(rqstp));
480 if (skb_is_nonlinear(skb)) {
481 /* we have to copy */
483 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb))
488 /* we can use it in-place */
489 rqstp->rq_arg.head[0].iov_base = skb->data;
490 rqstp->rq_arg.head[0].iov_len = len;
491 if (skb_checksum_complete(skb))
493 rqstp->rq_xprt_ctxt = skb;
496 rqstp->rq_arg.page_base = 0;
497 if (len <= rqstp->rq_arg.head[0].iov_len) {
498 rqstp->rq_arg.head[0].iov_len = len;
499 rqstp->rq_arg.page_len = 0;
500 rqstp->rq_respages = rqstp->rq_pages+1;
502 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
503 rqstp->rq_respages = rqstp->rq_pages + 1 +
504 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
506 rqstp->rq_next_page = rqstp->rq_respages+1;
509 serv->sv_stats->netudpcnt++;
511 svc_xprt_received(rqstp->rq_xprt);
515 if (err != -EAGAIN) {
516 /* possibly an icmp error */
517 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
519 trace_svcsock_udp_recv_err(&svsk->sk_xprt, err);
522 net_warn_ratelimited("svc: received unknown control message %d/%d; dropping RPC reply datagram\n",
523 cmh->cmsg_level, cmh->cmsg_type);
530 svc_xprt_received(rqstp->rq_xprt);
535 * svc_udp_sendto - Send out a reply on a UDP socket
536 * @rqstp: completed svc_rqst
538 * xpt_mutex ensures @rqstp's whole message is written to the socket
539 * without interruption.
541 * Returns the number of bytes sent, or a negative errno.
543 static int svc_udp_sendto(struct svc_rqst *rqstp)
545 struct svc_xprt *xprt = rqstp->rq_xprt;
546 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
547 struct xdr_buf *xdr = &rqstp->rq_res;
550 long all[SVC_PKTINFO_SPACE / sizeof(long)];
552 struct cmsghdr *cmh = &buffer.hdr;
553 struct msghdr msg = {
554 .msg_name = &rqstp->rq_addr,
555 .msg_namelen = rqstp->rq_addrlen,
557 .msg_controllen = sizeof(buffer),
562 svc_udp_release_rqst(rqstp);
564 svc_set_cmsg_data(rqstp, cmh);
566 mutex_lock(&xprt->xpt_mutex);
568 if (svc_xprt_is_dead(xprt))
571 err = xdr_alloc_bvec(xdr, GFP_KERNEL);
575 err = xprt_sock_sendmsg(svsk->sk_sock, &msg, xdr, 0, 0, &sent);
576 if (err == -ECONNREFUSED) {
577 /* ICMP error on earlier request. */
578 err = xprt_sock_sendmsg(svsk->sk_sock, &msg, xdr, 0, 0, &sent);
581 trace_svcsock_udp_send(xprt, err);
583 mutex_unlock(&xprt->xpt_mutex);
589 mutex_unlock(&xprt->xpt_mutex);
593 static int svc_udp_has_wspace(struct svc_xprt *xprt)
595 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
596 struct svc_serv *serv = xprt->xpt_server;
597 unsigned long required;
600 * Set the SOCK_NOSPACE flag before checking the available
603 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
604 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
605 if (required*2 > sock_wspace(svsk->sk_sk))
607 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
611 static struct svc_xprt *svc_udp_accept(struct svc_xprt *xprt)
617 static void svc_udp_kill_temp_xprt(struct svc_xprt *xprt)
621 static struct svc_xprt *svc_udp_create(struct svc_serv *serv,
623 struct sockaddr *sa, int salen,
626 return svc_create_socket(serv, IPPROTO_UDP, net, sa, salen, flags);
629 static const struct svc_xprt_ops svc_udp_ops = {
630 .xpo_create = svc_udp_create,
631 .xpo_recvfrom = svc_udp_recvfrom,
632 .xpo_sendto = svc_udp_sendto,
633 .xpo_result_payload = svc_sock_result_payload,
634 .xpo_release_rqst = svc_udp_release_rqst,
635 .xpo_detach = svc_sock_detach,
636 .xpo_free = svc_sock_free,
637 .xpo_has_wspace = svc_udp_has_wspace,
638 .xpo_accept = svc_udp_accept,
639 .xpo_secure_port = svc_sock_secure_port,
640 .xpo_kill_temp_xprt = svc_udp_kill_temp_xprt,
643 static struct svc_xprt_class svc_udp_class = {
645 .xcl_owner = THIS_MODULE,
646 .xcl_ops = &svc_udp_ops,
647 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
648 .xcl_ident = XPRT_TRANSPORT_UDP,
651 static void svc_udp_init(struct svc_sock *svsk, struct svc_serv *serv)
653 svc_xprt_init(sock_net(svsk->sk_sock->sk), &svc_udp_class,
654 &svsk->sk_xprt, serv);
655 clear_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
656 svsk->sk_sk->sk_data_ready = svc_data_ready;
657 svsk->sk_sk->sk_write_space = svc_write_space;
659 /* initialise setting must have enough space to
660 * receive and respond to one request.
661 * svc_udp_recvfrom will re-adjust if necessary
663 svc_sock_setbufsize(svsk, 3);
665 /* data might have come in before data_ready set up */
666 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
667 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
669 /* make sure we get destination address info */
670 switch (svsk->sk_sk->sk_family) {
672 ip_sock_set_pktinfo(svsk->sk_sock->sk);
675 ip6_sock_set_recvpktinfo(svsk->sk_sock->sk);
683 * A data_ready event on a listening socket means there's a connection
684 * pending. Do not use state_change as a substitute for it.
686 static void svc_tcp_listen_data_ready(struct sock *sk)
688 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
691 /* Refer to svc_setup_socket() for details. */
697 * This callback may called twice when a new connection
698 * is established as a child socket inherits everything
699 * from a parent LISTEN socket.
700 * 1) data_ready method of the parent socket will be called
701 * when one of child sockets become ESTABLISHED.
702 * 2) data_ready method of the child socket may be called
703 * when it receives data before the socket is accepted.
704 * In case of 2, we should ignore it silently.
706 if (sk->sk_state == TCP_LISTEN) {
708 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
709 svc_xprt_enqueue(&svsk->sk_xprt);
715 * A state change on a connected socket means it's dying or dead.
717 static void svc_tcp_state_change(struct sock *sk)
719 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
722 /* Refer to svc_setup_socket() for details. */
725 trace_svcsock_tcp_state(&svsk->sk_xprt, svsk->sk_sock);
726 if (sk->sk_state != TCP_ESTABLISHED)
727 svc_xprt_deferred_close(&svsk->sk_xprt);
732 * Accept a TCP connection
734 static struct svc_xprt *svc_tcp_accept(struct svc_xprt *xprt)
736 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
737 struct sockaddr_storage addr;
738 struct sockaddr *sin = (struct sockaddr *) &addr;
739 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
740 struct socket *sock = svsk->sk_sock;
741 struct socket *newsock;
742 struct svc_sock *newsvsk;
748 clear_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
749 err = kernel_accept(sock, &newsock, O_NONBLOCK);
752 printk(KERN_WARNING "%s: no more sockets!\n",
754 else if (err != -EAGAIN)
755 net_warn_ratelimited("%s: accept failed (err %d)!\n",
756 serv->sv_name, -err);
757 trace_svcsock_accept_err(xprt, serv->sv_name, err);
760 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
762 err = kernel_getpeername(newsock, sin);
764 trace_svcsock_getpeername_err(xprt, serv->sv_name, err);
765 goto failed; /* aborted connection or whatever */
769 /* Reset the inherited callbacks before calling svc_setup_socket */
770 newsock->sk->sk_state_change = svsk->sk_ostate;
771 newsock->sk->sk_data_ready = svsk->sk_odata;
772 newsock->sk->sk_write_space = svsk->sk_owspace;
774 /* make sure that a write doesn't block forever when
777 newsock->sk->sk_sndtimeo = HZ*30;
779 newsvsk = svc_setup_socket(serv, newsock,
780 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY));
783 svc_xprt_set_remote(&newsvsk->sk_xprt, sin, slen);
784 err = kernel_getsockname(newsock, sin);
786 if (unlikely(err < 0))
787 slen = offsetof(struct sockaddr, sa_data);
788 svc_xprt_set_local(&newsvsk->sk_xprt, sin, slen);
790 if (sock_is_loopback(newsock->sk))
791 set_bit(XPT_LOCAL, &newsvsk->sk_xprt.xpt_flags);
793 clear_bit(XPT_LOCAL, &newsvsk->sk_xprt.xpt_flags);
795 serv->sv_stats->nettcpconn++;
797 return &newsvsk->sk_xprt;
800 sock_release(newsock);
804 static size_t svc_tcp_restore_pages(struct svc_sock *svsk,
805 struct svc_rqst *rqstp)
807 size_t len = svsk->sk_datalen;
808 unsigned int i, npages;
812 npages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
813 for (i = 0; i < npages; i++) {
814 if (rqstp->rq_pages[i] != NULL)
815 put_page(rqstp->rq_pages[i]);
816 BUG_ON(svsk->sk_pages[i] == NULL);
817 rqstp->rq_pages[i] = svsk->sk_pages[i];
818 svsk->sk_pages[i] = NULL;
820 rqstp->rq_arg.head[0].iov_base = page_address(rqstp->rq_pages[0]);
824 static void svc_tcp_save_pages(struct svc_sock *svsk, struct svc_rqst *rqstp)
826 unsigned int i, len, npages;
828 if (svsk->sk_datalen == 0)
830 len = svsk->sk_datalen;
831 npages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
832 for (i = 0; i < npages; i++) {
833 svsk->sk_pages[i] = rqstp->rq_pages[i];
834 rqstp->rq_pages[i] = NULL;
838 static void svc_tcp_clear_pages(struct svc_sock *svsk)
840 unsigned int i, len, npages;
842 if (svsk->sk_datalen == 0)
844 len = svsk->sk_datalen;
845 npages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
846 for (i = 0; i < npages; i++) {
847 if (svsk->sk_pages[i] == NULL) {
851 put_page(svsk->sk_pages[i]);
852 svsk->sk_pages[i] = NULL;
856 svsk->sk_datalen = 0;
860 * Receive fragment record header into sk_marker.
862 static ssize_t svc_tcp_read_marker(struct svc_sock *svsk,
863 struct svc_rqst *rqstp)
867 /* If we haven't gotten the record length yet,
868 * get the next four bytes.
870 if (svsk->sk_tcplen < sizeof(rpc_fraghdr)) {
871 struct msghdr msg = { NULL };
874 want = sizeof(rpc_fraghdr) - svsk->sk_tcplen;
875 iov.iov_base = ((char *)&svsk->sk_marker) + svsk->sk_tcplen;
877 iov_iter_kvec(&msg.msg_iter, READ, &iov, 1, want);
878 len = sock_recvmsg(svsk->sk_sock, &msg, MSG_DONTWAIT);
881 svsk->sk_tcplen += len;
883 /* call again to read the remaining bytes */
886 trace_svcsock_marker(&svsk->sk_xprt, svsk->sk_marker);
887 if (svc_sock_reclen(svsk) + svsk->sk_datalen >
888 svsk->sk_xprt.xpt_server->sv_max_mesg)
891 return svc_sock_reclen(svsk);
894 net_notice_ratelimited("svc: %s %s RPC fragment too large: %d\n",
895 __func__, svsk->sk_xprt.xpt_server->sv_name,
896 svc_sock_reclen(svsk));
897 svc_xprt_deferred_close(&svsk->sk_xprt);
902 static int receive_cb_reply(struct svc_sock *svsk, struct svc_rqst *rqstp)
904 struct rpc_xprt *bc_xprt = svsk->sk_xprt.xpt_bc_xprt;
905 struct rpc_rqst *req = NULL;
906 struct kvec *src, *dst;
907 __be32 *p = (__be32 *)rqstp->rq_arg.head[0].iov_base;
916 spin_lock(&bc_xprt->queue_lock);
917 req = xprt_lookup_rqst(bc_xprt, xid);
919 goto unlock_notfound;
921 memcpy(&req->rq_private_buf, &req->rq_rcv_buf, sizeof(struct xdr_buf));
923 * XXX!: cheating for now! Only copying HEAD.
924 * But we know this is good enough for now (in fact, for any
925 * callback reply in the forseeable future).
927 dst = &req->rq_private_buf.head[0];
928 src = &rqstp->rq_arg.head[0];
929 if (dst->iov_len < src->iov_len)
930 goto unlock_eagain; /* whatever; just giving up. */
931 memcpy(dst->iov_base, src->iov_base, src->iov_len);
932 xprt_complete_rqst(req->rq_task, rqstp->rq_arg.len);
933 rqstp->rq_arg.len = 0;
934 spin_unlock(&bc_xprt->queue_lock);
938 "%s: Got unrecognized reply: "
939 "calldir 0x%x xpt_bc_xprt %p xid %08x\n",
940 __func__, ntohl(calldir),
941 bc_xprt, ntohl(xid));
943 spin_unlock(&bc_xprt->queue_lock);
947 static void svc_tcp_fragment_received(struct svc_sock *svsk)
949 /* If we have more data, signal svc_xprt_enqueue() to try again */
951 svsk->sk_marker = xdr_zero;
955 * svc_tcp_recvfrom - Receive data from a TCP socket
956 * @rqstp: request structure into which to receive an RPC Call
958 * Called in a loop when XPT_DATA has been set.
960 * Read the 4-byte stream record marker, then use the record length
961 * in that marker to set up exactly the resources needed to receive
962 * the next RPC message into @rqstp.
965 * On success, the number of bytes in a received RPC Call, or
966 * %0 if a complete RPC Call message was not ready to return
968 * The zero return case handles partial receives and callback Replies.
969 * The state of a partial receive is preserved in the svc_sock for
970 * the next call to svc_tcp_recvfrom.
972 static int svc_tcp_recvfrom(struct svc_rqst *rqstp)
974 struct svc_sock *svsk =
975 container_of(rqstp->rq_xprt, struct svc_sock, sk_xprt);
976 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
982 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
983 len = svc_tcp_read_marker(svsk, rqstp);
987 base = svc_tcp_restore_pages(svsk, rqstp);
988 want = len - (svsk->sk_tcplen - sizeof(rpc_fraghdr));
989 len = svc_tcp_read_msg(rqstp, base + want, base);
991 trace_svcsock_tcp_recv(&svsk->sk_xprt, len);
992 svsk->sk_tcplen += len;
993 svsk->sk_datalen += len;
995 if (len != want || !svc_sock_final_rec(svsk))
997 if (svsk->sk_datalen < 8)
1000 rqstp->rq_arg.len = svsk->sk_datalen;
1001 rqstp->rq_arg.page_base = 0;
1002 if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len) {
1003 rqstp->rq_arg.head[0].iov_len = rqstp->rq_arg.len;
1004 rqstp->rq_arg.page_len = 0;
1006 rqstp->rq_arg.page_len = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1008 rqstp->rq_xprt_ctxt = NULL;
1009 rqstp->rq_prot = IPPROTO_TCP;
1010 if (test_bit(XPT_LOCAL, &svsk->sk_xprt.xpt_flags))
1011 __set_bit(RQ_LOCAL, &rqstp->rq_flags);
1013 __clear_bit(RQ_LOCAL, &rqstp->rq_flags);
1015 p = (__be32 *)rqstp->rq_arg.head[0].iov_base;
1018 len = receive_cb_reply(svsk, rqstp);
1020 /* Reset TCP read info */
1021 svsk->sk_datalen = 0;
1022 svc_tcp_fragment_received(svsk);
1027 svc_xprt_copy_addrs(rqstp, &svsk->sk_xprt);
1029 serv->sv_stats->nettcpcnt++;
1031 svc_xprt_received(rqstp->rq_xprt);
1032 return rqstp->rq_arg.len;
1035 svc_tcp_save_pages(svsk, rqstp);
1036 if (len < 0 && len != -EAGAIN)
1039 svc_tcp_fragment_received(svsk);
1041 trace_svcsock_tcp_recv_short(&svsk->sk_xprt,
1042 svc_sock_reclen(svsk),
1043 svsk->sk_tcplen - sizeof(rpc_fraghdr));
1048 trace_svcsock_tcp_recv_eagain(&svsk->sk_xprt, 0);
1051 svsk->sk_datalen = 0;
1053 trace_svcsock_tcp_recv_err(&svsk->sk_xprt, len);
1054 svc_xprt_deferred_close(&svsk->sk_xprt);
1056 svc_xprt_received(rqstp->rq_xprt);
1057 return 0; /* record not complete */
1060 static int svc_tcp_send_kvec(struct socket *sock, const struct kvec *vec,
1063 return kernel_sendpage(sock, virt_to_page(vec->iov_base),
1064 offset_in_page(vec->iov_base),
1065 vec->iov_len, flags);
1069 * kernel_sendpage() is used exclusively to reduce the number of
1070 * copy operations in this path. Therefore the caller must ensure
1071 * that the pages backing @xdr are unchanging.
1073 * In addition, the logic assumes that * .bv_len is never larger
1076 static int svc_tcp_sendmsg(struct socket *sock, struct xdr_buf *xdr,
1077 rpc_fraghdr marker, unsigned int *sentp)
1079 const struct kvec *head = xdr->head;
1080 const struct kvec *tail = xdr->tail;
1082 .iov_base = &marker,
1083 .iov_len = sizeof(marker),
1085 struct msghdr msg = {
1091 ret = xdr_alloc_bvec(xdr, GFP_KERNEL);
1095 ret = kernel_sendmsg(sock, &msg, &rm, 1, rm.iov_len);
1099 if (ret != rm.iov_len)
1102 ret = svc_tcp_send_kvec(sock, head, 0);
1106 if (ret != head->iov_len)
1109 if (xdr->page_len) {
1110 unsigned int offset, len, remaining;
1111 struct bio_vec *bvec;
1113 bvec = xdr->bvec + (xdr->page_base >> PAGE_SHIFT);
1114 offset = offset_in_page(xdr->page_base);
1115 remaining = xdr->page_len;
1116 while (remaining > 0) {
1117 len = min(remaining, bvec->bv_len - offset);
1118 ret = kernel_sendpage(sock, bvec->bv_page,
1119 bvec->bv_offset + offset,
1132 if (tail->iov_len) {
1133 ret = svc_tcp_send_kvec(sock, tail, 0);
1144 * svc_tcp_sendto - Send out a reply on a TCP socket
1145 * @rqstp: completed svc_rqst
1147 * xpt_mutex ensures @rqstp's whole message is written to the socket
1148 * without interruption.
1150 * Returns the number of bytes sent, or a negative errno.
1152 static int svc_tcp_sendto(struct svc_rqst *rqstp)
1154 struct svc_xprt *xprt = rqstp->rq_xprt;
1155 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1156 struct xdr_buf *xdr = &rqstp->rq_res;
1157 rpc_fraghdr marker = cpu_to_be32(RPC_LAST_STREAM_FRAGMENT |
1162 svc_tcp_release_rqst(rqstp);
1164 atomic_inc(&svsk->sk_sendqlen);
1165 mutex_lock(&xprt->xpt_mutex);
1166 if (svc_xprt_is_dead(xprt))
1168 tcp_sock_set_cork(svsk->sk_sk, true);
1169 err = svc_tcp_sendmsg(svsk->sk_sock, xdr, marker, &sent);
1171 trace_svcsock_tcp_send(xprt, err < 0 ? (long)err : sent);
1172 if (err < 0 || sent != (xdr->len + sizeof(marker)))
1174 if (atomic_dec_and_test(&svsk->sk_sendqlen))
1175 tcp_sock_set_cork(svsk->sk_sk, false);
1176 mutex_unlock(&xprt->xpt_mutex);
1180 atomic_dec(&svsk->sk_sendqlen);
1181 mutex_unlock(&xprt->xpt_mutex);
1184 pr_notice("rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1185 xprt->xpt_server->sv_name,
1186 (err < 0) ? "got error" : "sent",
1187 (err < 0) ? err : sent, xdr->len);
1188 svc_xprt_deferred_close(xprt);
1189 atomic_dec(&svsk->sk_sendqlen);
1190 mutex_unlock(&xprt->xpt_mutex);
1194 static struct svc_xprt *svc_tcp_create(struct svc_serv *serv,
1196 struct sockaddr *sa, int salen,
1199 return svc_create_socket(serv, IPPROTO_TCP, net, sa, salen, flags);
1202 static const struct svc_xprt_ops svc_tcp_ops = {
1203 .xpo_create = svc_tcp_create,
1204 .xpo_recvfrom = svc_tcp_recvfrom,
1205 .xpo_sendto = svc_tcp_sendto,
1206 .xpo_result_payload = svc_sock_result_payload,
1207 .xpo_release_rqst = svc_tcp_release_rqst,
1208 .xpo_detach = svc_tcp_sock_detach,
1209 .xpo_free = svc_sock_free,
1210 .xpo_has_wspace = svc_tcp_has_wspace,
1211 .xpo_accept = svc_tcp_accept,
1212 .xpo_secure_port = svc_sock_secure_port,
1213 .xpo_kill_temp_xprt = svc_tcp_kill_temp_xprt,
1216 static struct svc_xprt_class svc_tcp_class = {
1218 .xcl_owner = THIS_MODULE,
1219 .xcl_ops = &svc_tcp_ops,
1220 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1221 .xcl_ident = XPRT_TRANSPORT_TCP,
1224 void svc_init_xprt_sock(void)
1226 svc_reg_xprt_class(&svc_tcp_class);
1227 svc_reg_xprt_class(&svc_udp_class);
1230 void svc_cleanup_xprt_sock(void)
1232 svc_unreg_xprt_class(&svc_tcp_class);
1233 svc_unreg_xprt_class(&svc_udp_class);
1236 static void svc_tcp_init(struct svc_sock *svsk, struct svc_serv *serv)
1238 struct sock *sk = svsk->sk_sk;
1240 svc_xprt_init(sock_net(svsk->sk_sock->sk), &svc_tcp_class,
1241 &svsk->sk_xprt, serv);
1242 set_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
1243 set_bit(XPT_CONG_CTRL, &svsk->sk_xprt.xpt_flags);
1244 if (sk->sk_state == TCP_LISTEN) {
1245 strcpy(svsk->sk_xprt.xpt_remotebuf, "listener");
1246 set_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags);
1247 sk->sk_data_ready = svc_tcp_listen_data_ready;
1248 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1250 sk->sk_state_change = svc_tcp_state_change;
1251 sk->sk_data_ready = svc_data_ready;
1252 sk->sk_write_space = svc_write_space;
1254 svsk->sk_marker = xdr_zero;
1255 svsk->sk_tcplen = 0;
1256 svsk->sk_datalen = 0;
1257 memset(&svsk->sk_pages[0], 0, sizeof(svsk->sk_pages));
1259 tcp_sock_set_nodelay(sk);
1261 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1262 switch (sk->sk_state) {
1264 case TCP_ESTABLISHED:
1267 svc_xprt_deferred_close(&svsk->sk_xprt);
1272 void svc_sock_update_bufs(struct svc_serv *serv)
1275 * The number of server threads has changed. Update
1276 * rcvbuf and sndbuf accordingly on all sockets
1278 struct svc_sock *svsk;
1280 spin_lock_bh(&serv->sv_lock);
1281 list_for_each_entry(svsk, &serv->sv_permsocks, sk_xprt.xpt_list)
1282 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1283 spin_unlock_bh(&serv->sv_lock);
1285 EXPORT_SYMBOL_GPL(svc_sock_update_bufs);
1288 * Initialize socket for RPC use and create svc_sock struct
1290 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1291 struct socket *sock,
1294 struct svc_sock *svsk;
1296 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1299 svsk = kzalloc(sizeof(*svsk), GFP_KERNEL);
1301 return ERR_PTR(-ENOMEM);
1305 /* Register socket with portmapper */
1307 err = svc_register(serv, sock_net(sock->sk), inet->sk_family,
1309 ntohs(inet_sk(inet)->inet_sport));
1313 return ERR_PTR(err);
1316 svsk->sk_sock = sock;
1318 svsk->sk_ostate = inet->sk_state_change;
1319 svsk->sk_odata = inet->sk_data_ready;
1320 svsk->sk_owspace = inet->sk_write_space;
1322 * This barrier is necessary in order to prevent race condition
1323 * with svc_data_ready(), svc_listen_data_ready() and others
1324 * when calling callbacks above.
1327 inet->sk_user_data = svsk;
1329 /* Initialize the socket */
1330 if (sock->type == SOCK_DGRAM)
1331 svc_udp_init(svsk, serv);
1333 svc_tcp_init(svsk, serv);
1335 trace_svcsock_new_socket(sock);
1339 bool svc_alien_sock(struct net *net, int fd)
1342 struct socket *sock = sockfd_lookup(fd, &err);
1347 if (sock_net(sock->sk) != net)
1353 EXPORT_SYMBOL_GPL(svc_alien_sock);
1356 * svc_addsock - add a listener socket to an RPC service
1357 * @serv: pointer to RPC service to which to add a new listener
1358 * @fd: file descriptor of the new listener
1359 * @name_return: pointer to buffer to fill in with name of listener
1360 * @len: size of the buffer
1363 * Fills in socket name and returns positive length of name if successful.
1364 * Name is terminated with '\n'. On error, returns a negative errno
1367 int svc_addsock(struct svc_serv *serv, const int fd, char *name_return,
1368 const size_t len, const struct cred *cred)
1371 struct socket *so = sockfd_lookup(fd, &err);
1372 struct svc_sock *svsk = NULL;
1373 struct sockaddr_storage addr;
1374 struct sockaddr *sin = (struct sockaddr *)&addr;
1379 err = -EAFNOSUPPORT;
1380 if ((so->sk->sk_family != PF_INET) && (so->sk->sk_family != PF_INET6))
1382 err = -EPROTONOSUPPORT;
1383 if (so->sk->sk_protocol != IPPROTO_TCP &&
1384 so->sk->sk_protocol != IPPROTO_UDP)
1387 if (so->state > SS_UNCONNECTED)
1390 if (!try_module_get(THIS_MODULE))
1392 svsk = svc_setup_socket(serv, so, SVC_SOCK_DEFAULTS);
1394 module_put(THIS_MODULE);
1395 err = PTR_ERR(svsk);
1398 salen = kernel_getsockname(svsk->sk_sock, sin);
1400 svc_xprt_set_local(&svsk->sk_xprt, sin, salen);
1401 svsk->sk_xprt.xpt_cred = get_cred(cred);
1402 svc_add_new_perm_xprt(serv, &svsk->sk_xprt);
1403 return svc_one_sock_name(svsk, name_return, len);
1408 EXPORT_SYMBOL_GPL(svc_addsock);
1411 * Create socket for RPC service.
1413 static struct svc_xprt *svc_create_socket(struct svc_serv *serv,
1416 struct sockaddr *sin, int len,
1419 struct svc_sock *svsk;
1420 struct socket *sock;
1423 struct sockaddr_storage addr;
1424 struct sockaddr *newsin = (struct sockaddr *)&addr;
1428 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1429 printk(KERN_WARNING "svc: only UDP and TCP "
1430 "sockets supported\n");
1431 return ERR_PTR(-EINVAL);
1434 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1435 switch (sin->sa_family) {
1443 return ERR_PTR(-EINVAL);
1446 error = __sock_create(net, family, type, protocol, &sock, 1);
1448 return ERR_PTR(error);
1450 svc_reclassify_socket(sock);
1453 * If this is an PF_INET6 listener, we want to avoid
1454 * getting requests from IPv4 remotes. Those should
1455 * be shunted to a PF_INET listener via rpcbind.
1457 if (family == PF_INET6)
1458 ip6_sock_set_v6only(sock->sk);
1459 if (type == SOCK_STREAM)
1460 sock->sk->sk_reuse = SK_CAN_REUSE; /* allow address reuse */
1461 error = kernel_bind(sock, sin, len);
1465 error = kernel_getsockname(sock, newsin);
1470 if (protocol == IPPROTO_TCP) {
1471 if ((error = kernel_listen(sock, 64)) < 0)
1475 svsk = svc_setup_socket(serv, sock, flags);
1477 error = PTR_ERR(svsk);
1480 svc_xprt_set_local(&svsk->sk_xprt, newsin, newlen);
1481 return (struct svc_xprt *)svsk;
1484 return ERR_PTR(error);
1488 * Detach the svc_sock from the socket so that no
1489 * more callbacks occur.
1491 static void svc_sock_detach(struct svc_xprt *xprt)
1493 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1494 struct sock *sk = svsk->sk_sk;
1496 /* put back the old socket callbacks */
1498 sk->sk_state_change = svsk->sk_ostate;
1499 sk->sk_data_ready = svsk->sk_odata;
1500 sk->sk_write_space = svsk->sk_owspace;
1501 sk->sk_user_data = NULL;
1506 * Disconnect the socket, and reset the callbacks
1508 static void svc_tcp_sock_detach(struct svc_xprt *xprt)
1510 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1512 svc_sock_detach(xprt);
1514 if (!test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
1515 svc_tcp_clear_pages(svsk);
1516 kernel_sock_shutdown(svsk->sk_sock, SHUT_RDWR);
1521 * Free the svc_sock's socket resources and the svc_sock itself.
1523 static void svc_sock_free(struct svc_xprt *xprt)
1525 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1527 if (svsk->sk_sock->file)
1528 sockfd_put(svsk->sk_sock);
1530 sock_release(svsk->sk_sock);