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_sock_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>
41 #include <net/tcp_states.h>
42 #include <asm/uaccess.h>
43 #include <asm/ioctls.h>
45 #include <linux/sunrpc/types.h>
46 #include <linux/sunrpc/clnt.h>
47 #include <linux/sunrpc/xdr.h>
48 #include <linux/sunrpc/svcsock.h>
49 #include <linux/sunrpc/stats.h>
51 /* SMP locking strategy:
53 * svc_pool->sp_lock protects most of the fields of that pool.
54 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
55 * when both need to be taken (rare), svc_serv->sv_lock is first.
56 * BKL protects svc_serv->sv_nrthread.
57 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
58 * and the ->sk_info_authunix cache.
59 * svc_sock->sk_flags.SK_BUSY prevents a svc_sock being enqueued multiply.
61 * Some flags can be set to certain values at any time
62 * providing that certain rules are followed:
64 * SK_CONN, SK_DATA, can be set or cleared at any time.
65 * after a set, svc_sock_enqueue must be called.
66 * after a clear, the socket must be read/accepted
67 * if this succeeds, it must be set again.
68 * SK_CLOSE can set at any time. It is never cleared.
69 * sk_inuse contains a bias of '1' until SK_DEAD is set.
70 * so when sk_inuse hits zero, we know the socket is dead
71 * and no-one is using it.
72 * SK_DEAD can only be set while SK_BUSY is held which ensures
73 * no other thread will be using the socket or will try to
78 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
81 static struct svc_sock *svc_setup_socket(struct svc_serv *, struct socket *,
82 int *errp, int flags);
83 static void svc_delete_socket(struct svc_sock *svsk);
84 static void svc_udp_data_ready(struct sock *, int);
85 static int svc_udp_recvfrom(struct svc_rqst *);
86 static int svc_udp_sendto(struct svc_rqst *);
87 static void svc_close_socket(struct svc_sock *svsk);
89 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk);
90 static int svc_deferred_recv(struct svc_rqst *rqstp);
91 static struct cache_deferred_req *svc_defer(struct cache_req *req);
93 /* apparently the "standard" is that clients close
94 * idle connections after 5 minutes, servers after
96 * http://www.connectathon.org/talks96/nfstcp.pdf
98 static int svc_conn_age_period = 6*60;
100 #ifdef CONFIG_DEBUG_LOCK_ALLOC
101 static struct lock_class_key svc_key[2];
102 static struct lock_class_key svc_slock_key[2];
104 static inline void svc_reclassify_socket(struct socket *sock)
106 struct sock *sk = sock->sk;
107 BUG_ON(sock_owned_by_user(sk));
108 switch (sk->sk_family) {
110 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFSD",
111 &svc_slock_key[0], "sk_lock-AF_INET-NFSD", &svc_key[0]);
115 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFSD",
116 &svc_slock_key[1], "sk_lock-AF_INET6-NFSD", &svc_key[1]);
124 static inline void svc_reclassify_socket(struct socket *sock)
129 static char *__svc_print_addr(struct sockaddr *addr, char *buf, size_t len)
131 switch (addr->sa_family) {
133 snprintf(buf, len, "%u.%u.%u.%u, port=%u",
134 NIPQUAD(((struct sockaddr_in *) addr)->sin_addr),
135 ntohs(((struct sockaddr_in *) addr)->sin_port));
139 snprintf(buf, len, "%x:%x:%x:%x:%x:%x:%x:%x, port=%u",
140 NIP6(((struct sockaddr_in6 *) addr)->sin6_addr),
141 ntohs(((struct sockaddr_in6 *) addr)->sin6_port));
145 snprintf(buf, len, "unknown address type: %d", addr->sa_family);
152 * svc_print_addr - Format rq_addr field for printing
153 * @rqstp: svc_rqst struct containing address to print
154 * @buf: target buffer for formatted address
155 * @len: length of target buffer
158 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
160 return __svc_print_addr(svc_addr(rqstp), buf, len);
162 EXPORT_SYMBOL_GPL(svc_print_addr);
165 * Queue up an idle server thread. Must have pool->sp_lock held.
166 * Note: this is really a stack rather than a queue, so that we only
167 * use as many different threads as we need, and the rest don't pollute
171 svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
173 list_add(&rqstp->rq_list, &pool->sp_threads);
177 * Dequeue an nfsd thread. Must have pool->sp_lock held.
180 svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
182 list_del(&rqstp->rq_list);
186 * Release an skbuff after use
188 static void svc_release_skb(struct svc_rqst *rqstp)
190 struct sk_buff *skb = rqstp->rq_xprt_ctxt;
191 struct svc_deferred_req *dr = rqstp->rq_deferred;
194 rqstp->rq_xprt_ctxt = NULL;
196 dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
197 skb_free_datagram(rqstp->rq_sock->sk_sk, skb);
200 rqstp->rq_deferred = NULL;
206 * Any space to write?
208 static inline unsigned long
209 svc_sock_wspace(struct svc_sock *svsk)
213 if (svsk->sk_sock->type == SOCK_STREAM)
214 wspace = sk_stream_wspace(svsk->sk_sk);
216 wspace = sock_wspace(svsk->sk_sk);
222 * Queue up a socket with data pending. If there are idle nfsd
223 * processes, wake 'em up.
227 svc_sock_enqueue(struct svc_sock *svsk)
229 struct svc_serv *serv = svsk->sk_server;
230 struct svc_pool *pool;
231 struct svc_rqst *rqstp;
234 if (!(svsk->sk_flags &
235 ( (1<<SK_CONN)|(1<<SK_DATA)|(1<<SK_CLOSE)|(1<<SK_DEFERRED)) ))
237 if (test_bit(SK_DEAD, &svsk->sk_flags))
241 pool = svc_pool_for_cpu(svsk->sk_server, cpu);
244 spin_lock_bh(&pool->sp_lock);
246 if (!list_empty(&pool->sp_threads) &&
247 !list_empty(&pool->sp_sockets))
249 "svc_sock_enqueue: threads and sockets both waiting??\n");
251 if (test_bit(SK_DEAD, &svsk->sk_flags)) {
252 /* Don't enqueue dead sockets */
253 dprintk("svc: socket %p is dead, not enqueued\n", svsk->sk_sk);
257 /* Mark socket as busy. It will remain in this state until the
258 * server has processed all pending data and put the socket back
259 * on the idle list. We update SK_BUSY atomically because
260 * it also guards against trying to enqueue the svc_sock twice.
262 if (test_and_set_bit(SK_BUSY, &svsk->sk_flags)) {
263 /* Don't enqueue socket while already enqueued */
264 dprintk("svc: socket %p busy, not enqueued\n", svsk->sk_sk);
267 BUG_ON(svsk->sk_pool != NULL);
268 svsk->sk_pool = pool;
270 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
271 if (((atomic_read(&svsk->sk_reserved) + serv->sv_max_mesg)*2
272 > svc_sock_wspace(svsk))
273 && !test_bit(SK_CLOSE, &svsk->sk_flags)
274 && !test_bit(SK_CONN, &svsk->sk_flags)) {
275 /* Don't enqueue while not enough space for reply */
276 dprintk("svc: socket %p no space, %d*2 > %ld, not enqueued\n",
277 svsk->sk_sk, atomic_read(&svsk->sk_reserved)+serv->sv_max_mesg,
278 svc_sock_wspace(svsk));
279 svsk->sk_pool = NULL;
280 clear_bit(SK_BUSY, &svsk->sk_flags);
283 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
286 if (!list_empty(&pool->sp_threads)) {
287 rqstp = list_entry(pool->sp_threads.next,
290 dprintk("svc: socket %p served by daemon %p\n",
292 svc_thread_dequeue(pool, rqstp);
295 "svc_sock_enqueue: server %p, rq_sock=%p!\n",
296 rqstp, rqstp->rq_sock);
297 rqstp->rq_sock = svsk;
298 atomic_inc(&svsk->sk_inuse);
299 rqstp->rq_reserved = serv->sv_max_mesg;
300 atomic_add(rqstp->rq_reserved, &svsk->sk_reserved);
301 BUG_ON(svsk->sk_pool != pool);
302 wake_up(&rqstp->rq_wait);
304 dprintk("svc: socket %p put into queue\n", svsk->sk_sk);
305 list_add_tail(&svsk->sk_ready, &pool->sp_sockets);
306 BUG_ON(svsk->sk_pool != pool);
310 spin_unlock_bh(&pool->sp_lock);
314 * Dequeue the first socket. Must be called with the pool->sp_lock held.
316 static inline struct svc_sock *
317 svc_sock_dequeue(struct svc_pool *pool)
319 struct svc_sock *svsk;
321 if (list_empty(&pool->sp_sockets))
324 svsk = list_entry(pool->sp_sockets.next,
325 struct svc_sock, sk_ready);
326 list_del_init(&svsk->sk_ready);
328 dprintk("svc: socket %p dequeued, inuse=%d\n",
329 svsk->sk_sk, atomic_read(&svsk->sk_inuse));
335 * Having read something from a socket, check whether it
336 * needs to be re-enqueued.
337 * Note: SK_DATA only gets cleared when a read-attempt finds
338 * no (or insufficient) data.
341 svc_sock_received(struct svc_sock *svsk)
343 svsk->sk_pool = NULL;
344 clear_bit(SK_BUSY, &svsk->sk_flags);
345 svc_sock_enqueue(svsk);
350 * svc_reserve - change the space reserved for the reply to a request.
351 * @rqstp: The request in question
352 * @space: new max space to reserve
354 * Each request reserves some space on the output queue of the socket
355 * to make sure the reply fits. This function reduces that reserved
356 * space to be the amount of space used already, plus @space.
359 void svc_reserve(struct svc_rqst *rqstp, int space)
361 space += rqstp->rq_res.head[0].iov_len;
363 if (space < rqstp->rq_reserved) {
364 struct svc_sock *svsk = rqstp->rq_sock;
365 atomic_sub((rqstp->rq_reserved - space), &svsk->sk_reserved);
366 rqstp->rq_reserved = space;
368 svc_sock_enqueue(svsk);
373 * Release a socket after use.
376 svc_sock_put(struct svc_sock *svsk)
378 if (atomic_dec_and_test(&svsk->sk_inuse)) {
379 BUG_ON(! test_bit(SK_DEAD, &svsk->sk_flags));
381 dprintk("svc: releasing dead socket\n");
382 if (svsk->sk_sock->file)
383 sockfd_put(svsk->sk_sock);
385 sock_release(svsk->sk_sock);
386 if (svsk->sk_info_authunix != NULL)
387 svcauth_unix_info_release(svsk->sk_info_authunix);
393 svc_sock_release(struct svc_rqst *rqstp)
395 struct svc_sock *svsk = rqstp->rq_sock;
397 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
399 svc_free_res_pages(rqstp);
400 rqstp->rq_res.page_len = 0;
401 rqstp->rq_res.page_base = 0;
404 /* Reset response buffer and release
406 * But first, check that enough space was reserved
407 * for the reply, otherwise we have a bug!
409 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
410 printk(KERN_ERR "RPC request reserved %d but used %d\n",
414 rqstp->rq_res.head[0].iov_len = 0;
415 svc_reserve(rqstp, 0);
416 rqstp->rq_sock = NULL;
422 * External function to wake up a server waiting for data
423 * This really only makes sense for services like lockd
424 * which have exactly one thread anyway.
427 svc_wake_up(struct svc_serv *serv)
429 struct svc_rqst *rqstp;
431 struct svc_pool *pool;
433 for (i = 0; i < serv->sv_nrpools; i++) {
434 pool = &serv->sv_pools[i];
436 spin_lock_bh(&pool->sp_lock);
437 if (!list_empty(&pool->sp_threads)) {
438 rqstp = list_entry(pool->sp_threads.next,
441 dprintk("svc: daemon %p woken up.\n", rqstp);
443 svc_thread_dequeue(pool, rqstp);
444 rqstp->rq_sock = NULL;
446 wake_up(&rqstp->rq_wait);
448 spin_unlock_bh(&pool->sp_lock);
452 union svc_pktinfo_u {
453 struct in_pktinfo pkti;
454 struct in6_pktinfo pkti6;
456 #define SVC_PKTINFO_SPACE \
457 CMSG_SPACE(sizeof(union svc_pktinfo_u))
459 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
461 switch (rqstp->rq_sock->sk_sk->sk_family) {
463 struct in_pktinfo *pki = CMSG_DATA(cmh);
465 cmh->cmsg_level = SOL_IP;
466 cmh->cmsg_type = IP_PKTINFO;
467 pki->ipi_ifindex = 0;
468 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr.addr.s_addr;
469 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
474 struct in6_pktinfo *pki = CMSG_DATA(cmh);
476 cmh->cmsg_level = SOL_IPV6;
477 cmh->cmsg_type = IPV6_PKTINFO;
478 pki->ipi6_ifindex = 0;
479 ipv6_addr_copy(&pki->ipi6_addr,
480 &rqstp->rq_daddr.addr6);
481 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
489 * Generic sendto routine
492 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
494 struct svc_sock *svsk = rqstp->rq_sock;
495 struct socket *sock = svsk->sk_sock;
499 long all[SVC_PKTINFO_SPACE / sizeof(long)];
501 struct cmsghdr *cmh = &buffer.hdr;
505 struct page **ppage = xdr->pages;
506 size_t base = xdr->page_base;
507 unsigned int pglen = xdr->page_len;
508 unsigned int flags = MSG_MORE;
509 char buf[RPC_MAX_ADDRBUFLEN];
513 if (rqstp->rq_prot == IPPROTO_UDP) {
514 struct msghdr msg = {
515 .msg_name = &rqstp->rq_addr,
516 .msg_namelen = rqstp->rq_addrlen,
518 .msg_controllen = sizeof(buffer),
519 .msg_flags = MSG_MORE,
522 svc_set_cmsg_data(rqstp, cmh);
524 if (sock_sendmsg(sock, &msg, 0) < 0)
529 if (slen == xdr->head[0].iov_len)
531 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0,
532 xdr->head[0].iov_len, flags);
533 if (len != xdr->head[0].iov_len)
535 slen -= xdr->head[0].iov_len;
540 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
544 result = kernel_sendpage(sock, *ppage, base, size, flags);
551 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
556 if (xdr->tail[0].iov_len) {
557 result = kernel_sendpage(sock, rqstp->rq_respages[0],
558 ((unsigned long)xdr->tail[0].iov_base)
560 xdr->tail[0].iov_len, 0);
566 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
567 rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len,
568 xdr->len, len, svc_print_addr(rqstp, buf, sizeof(buf)));
574 * Report socket names for nfsdfs
576 static int one_sock_name(char *buf, struct svc_sock *svsk)
580 switch(svsk->sk_sk->sk_family) {
582 len = sprintf(buf, "ipv4 %s %u.%u.%u.%u %d\n",
583 svsk->sk_sk->sk_protocol==IPPROTO_UDP?
585 NIPQUAD(inet_sk(svsk->sk_sk)->rcv_saddr),
586 inet_sk(svsk->sk_sk)->num);
589 len = sprintf(buf, "*unknown-%d*\n",
590 svsk->sk_sk->sk_family);
596 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
598 struct svc_sock *svsk, *closesk = NULL;
603 spin_lock_bh(&serv->sv_lock);
604 list_for_each_entry(svsk, &serv->sv_permsocks, sk_list) {
605 int onelen = one_sock_name(buf+len, svsk);
606 if (toclose && strcmp(toclose, buf+len) == 0)
611 spin_unlock_bh(&serv->sv_lock);
613 /* Should unregister with portmap, but you cannot
614 * unregister just one protocol...
616 svc_close_socket(closesk);
621 EXPORT_SYMBOL(svc_sock_names);
624 * Check input queue length
627 svc_recv_available(struct svc_sock *svsk)
629 struct socket *sock = svsk->sk_sock;
632 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
634 return (err >= 0)? avail : err;
638 * Generic recvfrom routine.
641 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
643 struct svc_sock *svsk = rqstp->rq_sock;
644 struct msghdr msg = {
645 .msg_flags = MSG_DONTWAIT,
647 struct sockaddr *sin;
650 len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
653 /* sock_recvmsg doesn't fill in the name/namelen, so we must..
655 memcpy(&rqstp->rq_addr, &svsk->sk_remote, svsk->sk_remotelen);
656 rqstp->rq_addrlen = svsk->sk_remotelen;
658 /* Destination address in request is needed for binding the
659 * source address in RPC callbacks later.
661 sin = (struct sockaddr *)&svsk->sk_local;
662 switch (sin->sa_family) {
664 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
667 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
671 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
672 svsk, iov[0].iov_base, iov[0].iov_len, len);
678 * Set socket snd and rcv buffer lengths
681 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
685 oldfs = get_fs(); set_fs(KERNEL_DS);
686 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
687 (char*)&snd, sizeof(snd));
688 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
689 (char*)&rcv, sizeof(rcv));
691 /* sock_setsockopt limits use to sysctl_?mem_max,
692 * which isn't acceptable. Until that is made conditional
693 * on not having CAP_SYS_RESOURCE or similar, we go direct...
694 * DaveM said I could!
697 sock->sk->sk_sndbuf = snd * 2;
698 sock->sk->sk_rcvbuf = rcv * 2;
699 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
700 release_sock(sock->sk);
704 * INET callback when data has been received on the socket.
707 svc_udp_data_ready(struct sock *sk, int count)
709 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
712 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
713 svsk, sk, count, test_bit(SK_BUSY, &svsk->sk_flags));
714 set_bit(SK_DATA, &svsk->sk_flags);
715 svc_sock_enqueue(svsk);
717 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
718 wake_up_interruptible(sk->sk_sleep);
722 * INET callback when space is newly available on the socket.
725 svc_write_space(struct sock *sk)
727 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
730 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
731 svsk, sk, test_bit(SK_BUSY, &svsk->sk_flags));
732 svc_sock_enqueue(svsk);
735 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
736 dprintk("RPC svc_write_space: someone sleeping on %p\n",
738 wake_up_interruptible(sk->sk_sleep);
742 static inline void svc_udp_get_dest_address(struct svc_rqst *rqstp,
745 switch (rqstp->rq_sock->sk_sk->sk_family) {
747 struct in_pktinfo *pki = CMSG_DATA(cmh);
748 rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
752 struct in6_pktinfo *pki = CMSG_DATA(cmh);
753 ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
760 * Receive a datagram from a UDP socket.
763 svc_udp_recvfrom(struct svc_rqst *rqstp)
765 struct svc_sock *svsk = rqstp->rq_sock;
766 struct svc_serv *serv = svsk->sk_server;
770 long all[SVC_PKTINFO_SPACE / sizeof(long)];
772 struct cmsghdr *cmh = &buffer.hdr;
774 struct msghdr msg = {
775 .msg_name = svc_addr(rqstp),
777 .msg_controllen = sizeof(buffer),
778 .msg_flags = MSG_DONTWAIT,
781 if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
782 /* udp sockets need large rcvbuf as all pending
783 * requests are still in that buffer. sndbuf must
784 * also be large enough that there is enough space
785 * for one reply per thread. We count all threads
786 * rather than threads in a particular pool, which
787 * provides an upper bound on the number of threads
788 * which will access the socket.
790 svc_sock_setbufsize(svsk->sk_sock,
791 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
792 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
794 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
795 svc_sock_received(svsk);
796 return svc_deferred_recv(rqstp);
799 if (test_bit(SK_CLOSE, &svsk->sk_flags)) {
800 svc_delete_socket(svsk);
804 clear_bit(SK_DATA, &svsk->sk_flags);
806 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
807 0, 0, MSG_PEEK | MSG_DONTWAIT);
809 skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err);
812 if (err != -EAGAIN) {
813 /* possibly an icmp error */
814 dprintk("svc: recvfrom returned error %d\n", -err);
815 set_bit(SK_DATA, &svsk->sk_flags);
817 svc_sock_received(svsk);
820 rqstp->rq_addrlen = sizeof(rqstp->rq_addr);
821 if (skb->tstamp.tv64 == 0) {
822 skb->tstamp = ktime_get_real();
823 /* Don't enable netstamp, sunrpc doesn't
824 need that much accuracy */
826 svsk->sk_sk->sk_stamp = skb->tstamp;
827 set_bit(SK_DATA, &svsk->sk_flags); /* there may be more data... */
830 * Maybe more packets - kick another thread ASAP.
832 svc_sock_received(svsk);
834 len = skb->len - sizeof(struct udphdr);
835 rqstp->rq_arg.len = len;
837 rqstp->rq_prot = IPPROTO_UDP;
839 if (cmh->cmsg_level != IPPROTO_IP ||
840 cmh->cmsg_type != IP_PKTINFO) {
842 printk("rpcsvc: received unknown control message:"
844 cmh->cmsg_level, cmh->cmsg_type);
845 skb_free_datagram(svsk->sk_sk, skb);
848 svc_udp_get_dest_address(rqstp, cmh);
850 if (skb_is_nonlinear(skb)) {
851 /* we have to copy */
853 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
856 skb_free_datagram(svsk->sk_sk, skb);
860 skb_free_datagram(svsk->sk_sk, skb);
862 /* we can use it in-place */
863 rqstp->rq_arg.head[0].iov_base = skb->data + sizeof(struct udphdr);
864 rqstp->rq_arg.head[0].iov_len = len;
865 if (skb_checksum_complete(skb)) {
866 skb_free_datagram(svsk->sk_sk, skb);
869 rqstp->rq_xprt_ctxt = skb;
872 rqstp->rq_arg.page_base = 0;
873 if (len <= rqstp->rq_arg.head[0].iov_len) {
874 rqstp->rq_arg.head[0].iov_len = len;
875 rqstp->rq_arg.page_len = 0;
876 rqstp->rq_respages = rqstp->rq_pages+1;
878 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
879 rqstp->rq_respages = rqstp->rq_pages + 1 +
880 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
884 serv->sv_stats->netudpcnt++;
890 svc_udp_sendto(struct svc_rqst *rqstp)
894 error = svc_sendto(rqstp, &rqstp->rq_res);
895 if (error == -ECONNREFUSED)
896 /* ICMP error on earlier request. */
897 error = svc_sendto(rqstp, &rqstp->rq_res);
902 static struct svc_xprt_ops svc_udp_ops = {
903 .xpo_recvfrom = svc_udp_recvfrom,
904 .xpo_sendto = svc_udp_sendto,
905 .xpo_release_rqst = svc_release_skb,
908 static struct svc_xprt_class svc_udp_class = {
910 .xcl_ops = &svc_udp_ops,
911 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
915 svc_udp_init(struct svc_sock *svsk)
920 svc_xprt_init(&svc_udp_class, &svsk->sk_xprt);
921 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
922 svsk->sk_sk->sk_write_space = svc_write_space;
924 /* initialise setting must have enough space to
925 * receive and respond to one request.
926 * svc_udp_recvfrom will re-adjust if necessary
928 svc_sock_setbufsize(svsk->sk_sock,
929 3 * svsk->sk_server->sv_max_mesg,
930 3 * svsk->sk_server->sv_max_mesg);
932 set_bit(SK_DATA, &svsk->sk_flags); /* might have come in before data_ready set up */
933 set_bit(SK_CHNGBUF, &svsk->sk_flags);
937 /* make sure we get destination address info */
938 svsk->sk_sock->ops->setsockopt(svsk->sk_sock, IPPROTO_IP, IP_PKTINFO,
939 (char __user *)&one, sizeof(one));
944 * A data_ready event on a listening socket means there's a connection
945 * pending. Do not use state_change as a substitute for it.
948 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
950 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
952 dprintk("svc: socket %p TCP (listen) state change %d\n",
956 * This callback may called twice when a new connection
957 * is established as a child socket inherits everything
958 * from a parent LISTEN socket.
959 * 1) data_ready method of the parent socket will be called
960 * when one of child sockets become ESTABLISHED.
961 * 2) data_ready method of the child socket may be called
962 * when it receives data before the socket is accepted.
963 * In case of 2, we should ignore it silently.
965 if (sk->sk_state == TCP_LISTEN) {
967 set_bit(SK_CONN, &svsk->sk_flags);
968 svc_sock_enqueue(svsk);
970 printk("svc: socket %p: no user data\n", sk);
973 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
974 wake_up_interruptible_all(sk->sk_sleep);
978 * A state change on a connected socket means it's dying or dead.
981 svc_tcp_state_change(struct sock *sk)
983 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
985 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
986 sk, sk->sk_state, sk->sk_user_data);
989 printk("svc: socket %p: no user data\n", sk);
991 set_bit(SK_CLOSE, &svsk->sk_flags);
992 svc_sock_enqueue(svsk);
994 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
995 wake_up_interruptible_all(sk->sk_sleep);
999 svc_tcp_data_ready(struct sock *sk, int count)
1001 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
1003 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
1004 sk, sk->sk_user_data);
1006 set_bit(SK_DATA, &svsk->sk_flags);
1007 svc_sock_enqueue(svsk);
1009 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1010 wake_up_interruptible(sk->sk_sleep);
1013 static inline int svc_port_is_privileged(struct sockaddr *sin)
1015 switch (sin->sa_family) {
1017 return ntohs(((struct sockaddr_in *)sin)->sin_port)
1020 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
1028 * Accept a TCP connection
1031 svc_tcp_accept(struct svc_sock *svsk)
1033 struct sockaddr_storage addr;
1034 struct sockaddr *sin = (struct sockaddr *) &addr;
1035 struct svc_serv *serv = svsk->sk_server;
1036 struct socket *sock = svsk->sk_sock;
1037 struct socket *newsock;
1038 struct svc_sock *newsvsk;
1040 char buf[RPC_MAX_ADDRBUFLEN];
1042 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
1046 clear_bit(SK_CONN, &svsk->sk_flags);
1047 err = kernel_accept(sock, &newsock, O_NONBLOCK);
1050 printk(KERN_WARNING "%s: no more sockets!\n",
1052 else if (err != -EAGAIN && net_ratelimit())
1053 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
1054 serv->sv_name, -err);
1058 set_bit(SK_CONN, &svsk->sk_flags);
1059 svc_sock_enqueue(svsk);
1061 err = kernel_getpeername(newsock, sin, &slen);
1063 if (net_ratelimit())
1064 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
1065 serv->sv_name, -err);
1066 goto failed; /* aborted connection or whatever */
1069 /* Ideally, we would want to reject connections from unauthorized
1070 * hosts here, but when we get encryption, the IP of the host won't
1071 * tell us anything. For now just warn about unpriv connections.
1073 if (!svc_port_is_privileged(sin)) {
1074 dprintk(KERN_WARNING
1075 "%s: connect from unprivileged port: %s\n",
1077 __svc_print_addr(sin, buf, sizeof(buf)));
1079 dprintk("%s: connect from %s\n", serv->sv_name,
1080 __svc_print_addr(sin, buf, sizeof(buf)));
1082 /* make sure that a write doesn't block forever when
1085 newsock->sk->sk_sndtimeo = HZ*30;
1087 if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
1088 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
1090 memcpy(&newsvsk->sk_remote, sin, slen);
1091 newsvsk->sk_remotelen = slen;
1092 err = kernel_getsockname(newsock, sin, &slen);
1093 if (unlikely(err < 0)) {
1094 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err);
1095 slen = offsetof(struct sockaddr, sa_data);
1097 memcpy(&newsvsk->sk_local, sin, slen);
1099 svc_sock_received(newsvsk);
1101 /* make sure that we don't have too many active connections.
1102 * If we have, something must be dropped.
1104 * There's no point in trying to do random drop here for
1105 * DoS prevention. The NFS clients does 1 reconnect in 15
1106 * seconds. An attacker can easily beat that.
1108 * The only somewhat efficient mechanism would be if drop
1109 * old connections from the same IP first. But right now
1110 * we don't even record the client IP in svc_sock.
1112 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
1113 struct svc_sock *svsk = NULL;
1114 spin_lock_bh(&serv->sv_lock);
1115 if (!list_empty(&serv->sv_tempsocks)) {
1116 if (net_ratelimit()) {
1117 /* Try to help the admin */
1118 printk(KERN_NOTICE "%s: too many open TCP "
1119 "sockets, consider increasing the "
1120 "number of nfsd threads\n",
1123 "%s: last TCP connect from %s\n",
1124 serv->sv_name, __svc_print_addr(sin,
1128 * Always select the oldest socket. It's not fair,
1131 svsk = list_entry(serv->sv_tempsocks.prev,
1134 set_bit(SK_CLOSE, &svsk->sk_flags);
1135 atomic_inc(&svsk->sk_inuse);
1137 spin_unlock_bh(&serv->sv_lock);
1140 svc_sock_enqueue(svsk);
1147 serv->sv_stats->nettcpconn++;
1152 sock_release(newsock);
1157 * Receive data from a TCP socket.
1160 svc_tcp_recvfrom(struct svc_rqst *rqstp)
1162 struct svc_sock *svsk = rqstp->rq_sock;
1163 struct svc_serv *serv = svsk->sk_server;
1168 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
1169 svsk, test_bit(SK_DATA, &svsk->sk_flags),
1170 test_bit(SK_CONN, &svsk->sk_flags),
1171 test_bit(SK_CLOSE, &svsk->sk_flags));
1173 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
1174 svc_sock_received(svsk);
1175 return svc_deferred_recv(rqstp);
1178 if (test_bit(SK_CLOSE, &svsk->sk_flags)) {
1179 svc_delete_socket(svsk);
1183 if (svsk->sk_sk->sk_state == TCP_LISTEN) {
1184 svc_tcp_accept(svsk);
1185 svc_sock_received(svsk);
1189 if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
1190 /* sndbuf needs to have room for one request
1191 * per thread, otherwise we can stall even when the
1192 * network isn't a bottleneck.
1194 * We count all threads rather than threads in a
1195 * particular pool, which provides an upper bound
1196 * on the number of threads which will access the socket.
1198 * rcvbuf just needs to be able to hold a few requests.
1199 * Normally they will be removed from the queue
1200 * as soon a a complete request arrives.
1202 svc_sock_setbufsize(svsk->sk_sock,
1203 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
1204 3 * serv->sv_max_mesg);
1206 clear_bit(SK_DATA, &svsk->sk_flags);
1208 /* Receive data. If we haven't got the record length yet, get
1209 * the next four bytes. Otherwise try to gobble up as much as
1210 * possible up to the complete record length.
1212 if (svsk->sk_tcplen < 4) {
1213 unsigned long want = 4 - svsk->sk_tcplen;
1216 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
1218 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
1220 svsk->sk_tcplen += len;
1223 dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
1225 svc_sock_received(svsk);
1226 return -EAGAIN; /* record header not complete */
1229 svsk->sk_reclen = ntohl(svsk->sk_reclen);
1230 if (!(svsk->sk_reclen & 0x80000000)) {
1231 /* FIXME: technically, a record can be fragmented,
1232 * and non-terminal fragments will not have the top
1233 * bit set in the fragment length header.
1234 * But apparently no known nfs clients send fragmented
1236 if (net_ratelimit())
1237 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1238 " (non-terminal)\n",
1239 (unsigned long) svsk->sk_reclen);
1242 svsk->sk_reclen &= 0x7fffffff;
1243 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
1244 if (svsk->sk_reclen > serv->sv_max_mesg) {
1245 if (net_ratelimit())
1246 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1248 (unsigned long) svsk->sk_reclen);
1253 /* Check whether enough data is available */
1254 len = svc_recv_available(svsk);
1258 if (len < svsk->sk_reclen) {
1259 dprintk("svc: incomplete TCP record (%d of %d)\n",
1260 len, svsk->sk_reclen);
1261 svc_sock_received(svsk);
1262 return -EAGAIN; /* record not complete */
1264 len = svsk->sk_reclen;
1265 set_bit(SK_DATA, &svsk->sk_flags);
1267 vec = rqstp->rq_vec;
1268 vec[0] = rqstp->rq_arg.head[0];
1271 while (vlen < len) {
1272 vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
1273 vec[pnum].iov_len = PAGE_SIZE;
1277 rqstp->rq_respages = &rqstp->rq_pages[pnum];
1279 /* Now receive data */
1280 len = svc_recvfrom(rqstp, vec, pnum, len);
1284 dprintk("svc: TCP complete record (%d bytes)\n", len);
1285 rqstp->rq_arg.len = len;
1286 rqstp->rq_arg.page_base = 0;
1287 if (len <= rqstp->rq_arg.head[0].iov_len) {
1288 rqstp->rq_arg.head[0].iov_len = len;
1289 rqstp->rq_arg.page_len = 0;
1291 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1294 rqstp->rq_xprt_ctxt = NULL;
1295 rqstp->rq_prot = IPPROTO_TCP;
1297 /* Reset TCP read info */
1298 svsk->sk_reclen = 0;
1299 svsk->sk_tcplen = 0;
1301 svc_sock_received(svsk);
1303 serv->sv_stats->nettcpcnt++;
1308 svc_delete_socket(svsk);
1312 if (len == -EAGAIN) {
1313 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1314 svc_sock_received(svsk);
1316 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1317 svsk->sk_server->sv_name, -len);
1325 * Send out data on TCP socket.
1328 svc_tcp_sendto(struct svc_rqst *rqstp)
1330 struct xdr_buf *xbufp = &rqstp->rq_res;
1334 /* Set up the first element of the reply kvec.
1335 * Any other kvecs that may be in use have been taken
1336 * care of by the server implementation itself.
1338 reclen = htonl(0x80000000|((xbufp->len ) - 4));
1339 memcpy(xbufp->head[0].iov_base, &reclen, 4);
1341 if (test_bit(SK_DEAD, &rqstp->rq_sock->sk_flags))
1344 sent = svc_sendto(rqstp, &rqstp->rq_res);
1345 if (sent != xbufp->len) {
1346 printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1347 rqstp->rq_sock->sk_server->sv_name,
1348 (sent<0)?"got error":"sent only",
1350 set_bit(SK_CLOSE, &rqstp->rq_sock->sk_flags);
1351 svc_sock_enqueue(rqstp->rq_sock);
1357 static struct svc_xprt_ops svc_tcp_ops = {
1358 .xpo_recvfrom = svc_tcp_recvfrom,
1359 .xpo_sendto = svc_tcp_sendto,
1360 .xpo_release_rqst = svc_release_skb,
1363 static struct svc_xprt_class svc_tcp_class = {
1365 .xcl_ops = &svc_tcp_ops,
1366 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1369 void svc_init_xprt_sock(void)
1371 svc_reg_xprt_class(&svc_tcp_class);
1372 svc_reg_xprt_class(&svc_udp_class);
1375 void svc_cleanup_xprt_sock(void)
1377 svc_unreg_xprt_class(&svc_tcp_class);
1378 svc_unreg_xprt_class(&svc_udp_class);
1382 svc_tcp_init(struct svc_sock *svsk)
1384 struct sock *sk = svsk->sk_sk;
1385 struct tcp_sock *tp = tcp_sk(sk);
1387 svc_xprt_init(&svc_tcp_class, &svsk->sk_xprt);
1389 if (sk->sk_state == TCP_LISTEN) {
1390 dprintk("setting up TCP socket for listening\n");
1391 sk->sk_data_ready = svc_tcp_listen_data_ready;
1392 set_bit(SK_CONN, &svsk->sk_flags);
1394 dprintk("setting up TCP socket for reading\n");
1395 sk->sk_state_change = svc_tcp_state_change;
1396 sk->sk_data_ready = svc_tcp_data_ready;
1397 sk->sk_write_space = svc_write_space;
1399 svsk->sk_reclen = 0;
1400 svsk->sk_tcplen = 0;
1402 tp->nonagle = 1; /* disable Nagle's algorithm */
1404 /* initialise setting must have enough space to
1405 * receive and respond to one request.
1406 * svc_tcp_recvfrom will re-adjust if necessary
1408 svc_sock_setbufsize(svsk->sk_sock,
1409 3 * svsk->sk_server->sv_max_mesg,
1410 3 * svsk->sk_server->sv_max_mesg);
1412 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1413 set_bit(SK_DATA, &svsk->sk_flags);
1414 if (sk->sk_state != TCP_ESTABLISHED)
1415 set_bit(SK_CLOSE, &svsk->sk_flags);
1420 svc_sock_update_bufs(struct svc_serv *serv)
1423 * The number of server threads has changed. Update
1424 * rcvbuf and sndbuf accordingly on all sockets
1426 struct list_head *le;
1428 spin_lock_bh(&serv->sv_lock);
1429 list_for_each(le, &serv->sv_permsocks) {
1430 struct svc_sock *svsk =
1431 list_entry(le, struct svc_sock, sk_list);
1432 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1434 list_for_each(le, &serv->sv_tempsocks) {
1435 struct svc_sock *svsk =
1436 list_entry(le, struct svc_sock, sk_list);
1437 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1439 spin_unlock_bh(&serv->sv_lock);
1443 * Receive the next request on any socket. This code is carefully
1444 * organised not to touch any cachelines in the shared svc_serv
1445 * structure, only cachelines in the local svc_pool.
1448 svc_recv(struct svc_rqst *rqstp, long timeout)
1450 struct svc_sock *svsk = NULL;
1451 struct svc_serv *serv = rqstp->rq_server;
1452 struct svc_pool *pool = rqstp->rq_pool;
1455 struct xdr_buf *arg;
1456 DECLARE_WAITQUEUE(wait, current);
1458 dprintk("svc: server %p waiting for data (to = %ld)\n",
1463 "svc_recv: service %p, socket not NULL!\n",
1465 if (waitqueue_active(&rqstp->rq_wait))
1467 "svc_recv: service %p, wait queue active!\n",
1471 /* now allocate needed pages. If we get a failure, sleep briefly */
1472 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
1473 for (i=0; i < pages ; i++)
1474 while (rqstp->rq_pages[i] == NULL) {
1475 struct page *p = alloc_page(GFP_KERNEL);
1477 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1478 rqstp->rq_pages[i] = p;
1480 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
1481 BUG_ON(pages >= RPCSVC_MAXPAGES);
1483 /* Make arg->head point to first page and arg->pages point to rest */
1484 arg = &rqstp->rq_arg;
1485 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
1486 arg->head[0].iov_len = PAGE_SIZE;
1487 arg->pages = rqstp->rq_pages + 1;
1489 /* save at least one page for response */
1490 arg->page_len = (pages-2)*PAGE_SIZE;
1491 arg->len = (pages-1)*PAGE_SIZE;
1492 arg->tail[0].iov_len = 0;
1499 spin_lock_bh(&pool->sp_lock);
1500 if ((svsk = svc_sock_dequeue(pool)) != NULL) {
1501 rqstp->rq_sock = svsk;
1502 atomic_inc(&svsk->sk_inuse);
1503 rqstp->rq_reserved = serv->sv_max_mesg;
1504 atomic_add(rqstp->rq_reserved, &svsk->sk_reserved);
1506 /* No data pending. Go to sleep */
1507 svc_thread_enqueue(pool, rqstp);
1510 * We have to be able to interrupt this wait
1511 * to bring down the daemons ...
1513 set_current_state(TASK_INTERRUPTIBLE);
1514 add_wait_queue(&rqstp->rq_wait, &wait);
1515 spin_unlock_bh(&pool->sp_lock);
1517 schedule_timeout(timeout);
1521 spin_lock_bh(&pool->sp_lock);
1522 remove_wait_queue(&rqstp->rq_wait, &wait);
1524 if (!(svsk = rqstp->rq_sock)) {
1525 svc_thread_dequeue(pool, rqstp);
1526 spin_unlock_bh(&pool->sp_lock);
1527 dprintk("svc: server %p, no data yet\n", rqstp);
1528 return signalled()? -EINTR : -EAGAIN;
1531 spin_unlock_bh(&pool->sp_lock);
1533 dprintk("svc: server %p, pool %u, socket %p, inuse=%d\n",
1534 rqstp, pool->sp_id, svsk, atomic_read(&svsk->sk_inuse));
1535 len = svsk->sk_xprt.xpt_ops->xpo_recvfrom(rqstp);
1536 dprintk("svc: got len=%d\n", len);
1538 /* No data, incomplete (TCP) read, or accept() */
1539 if (len == 0 || len == -EAGAIN) {
1540 rqstp->rq_res.len = 0;
1541 svc_sock_release(rqstp);
1544 svsk->sk_lastrecv = get_seconds();
1545 clear_bit(SK_OLD, &svsk->sk_flags);
1547 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
1548 rqstp->rq_chandle.defer = svc_defer;
1551 serv->sv_stats->netcnt++;
1559 svc_drop(struct svc_rqst *rqstp)
1561 dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1562 svc_sock_release(rqstp);
1566 * Return reply to client.
1569 svc_send(struct svc_rqst *rqstp)
1571 struct svc_sock *svsk;
1575 if ((svsk = rqstp->rq_sock) == NULL) {
1576 printk(KERN_WARNING "NULL socket pointer in %s:%d\n",
1577 __FILE__, __LINE__);
1581 /* release the receive skb before sending the reply */
1582 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
1584 /* calculate over-all length */
1585 xb = & rqstp->rq_res;
1586 xb->len = xb->head[0].iov_len +
1588 xb->tail[0].iov_len;
1590 /* Grab svsk->sk_mutex to serialize outgoing data. */
1591 mutex_lock(&svsk->sk_mutex);
1592 if (test_bit(SK_DEAD, &svsk->sk_flags))
1595 len = svsk->sk_xprt.xpt_ops->xpo_sendto(rqstp);
1596 mutex_unlock(&svsk->sk_mutex);
1597 svc_sock_release(rqstp);
1599 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1605 * Timer function to close old temporary sockets, using
1606 * a mark-and-sweep algorithm.
1609 svc_age_temp_sockets(unsigned long closure)
1611 struct svc_serv *serv = (struct svc_serv *)closure;
1612 struct svc_sock *svsk;
1613 struct list_head *le, *next;
1614 LIST_HEAD(to_be_aged);
1616 dprintk("svc_age_temp_sockets\n");
1618 if (!spin_trylock_bh(&serv->sv_lock)) {
1619 /* busy, try again 1 sec later */
1620 dprintk("svc_age_temp_sockets: busy\n");
1621 mod_timer(&serv->sv_temptimer, jiffies + HZ);
1625 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1626 svsk = list_entry(le, struct svc_sock, sk_list);
1628 if (!test_and_set_bit(SK_OLD, &svsk->sk_flags))
1630 if (atomic_read(&svsk->sk_inuse) > 1 || test_bit(SK_BUSY, &svsk->sk_flags))
1632 atomic_inc(&svsk->sk_inuse);
1633 list_move(le, &to_be_aged);
1634 set_bit(SK_CLOSE, &svsk->sk_flags);
1635 set_bit(SK_DETACHED, &svsk->sk_flags);
1637 spin_unlock_bh(&serv->sv_lock);
1639 while (!list_empty(&to_be_aged)) {
1640 le = to_be_aged.next;
1641 /* fiddling the sk_list node is safe 'cos we're SK_DETACHED */
1643 svsk = list_entry(le, struct svc_sock, sk_list);
1645 dprintk("queuing svsk %p for closing, %lu seconds old\n",
1646 svsk, get_seconds() - svsk->sk_lastrecv);
1648 /* a thread will dequeue and close it soon */
1649 svc_sock_enqueue(svsk);
1653 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
1657 * Initialize socket for RPC use and create svc_sock struct
1658 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1660 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1661 struct socket *sock,
1662 int *errp, int flags)
1664 struct svc_sock *svsk;
1666 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1667 int is_temporary = flags & SVC_SOCK_TEMPORARY;
1669 dprintk("svc: svc_setup_socket %p\n", sock);
1670 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1677 /* Register socket with portmapper */
1678 if (*errp >= 0 && pmap_register)
1679 *errp = svc_register(serv, inet->sk_protocol,
1680 ntohs(inet_sk(inet)->sport));
1687 set_bit(SK_BUSY, &svsk->sk_flags);
1688 inet->sk_user_data = svsk;
1689 svsk->sk_sock = sock;
1691 svsk->sk_ostate = inet->sk_state_change;
1692 svsk->sk_odata = inet->sk_data_ready;
1693 svsk->sk_owspace = inet->sk_write_space;
1694 svsk->sk_server = serv;
1695 atomic_set(&svsk->sk_inuse, 1);
1696 svsk->sk_lastrecv = get_seconds();
1697 spin_lock_init(&svsk->sk_lock);
1698 INIT_LIST_HEAD(&svsk->sk_deferred);
1699 INIT_LIST_HEAD(&svsk->sk_ready);
1700 mutex_init(&svsk->sk_mutex);
1702 /* Initialize the socket */
1703 if (sock->type == SOCK_DGRAM)
1708 spin_lock_bh(&serv->sv_lock);
1710 set_bit(SK_TEMP, &svsk->sk_flags);
1711 list_add(&svsk->sk_list, &serv->sv_tempsocks);
1713 if (serv->sv_temptimer.function == NULL) {
1714 /* setup timer to age temp sockets */
1715 setup_timer(&serv->sv_temptimer, svc_age_temp_sockets,
1716 (unsigned long)serv);
1717 mod_timer(&serv->sv_temptimer,
1718 jiffies + svc_conn_age_period * HZ);
1721 clear_bit(SK_TEMP, &svsk->sk_flags);
1722 list_add(&svsk->sk_list, &serv->sv_permsocks);
1724 spin_unlock_bh(&serv->sv_lock);
1726 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1732 int svc_addsock(struct svc_serv *serv,
1738 struct socket *so = sockfd_lookup(fd, &err);
1739 struct svc_sock *svsk = NULL;
1743 if (so->sk->sk_family != AF_INET)
1744 err = -EAFNOSUPPORT;
1745 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1746 so->sk->sk_protocol != IPPROTO_UDP)
1747 err = -EPROTONOSUPPORT;
1748 else if (so->state > SS_UNCONNECTED)
1751 svsk = svc_setup_socket(serv, so, &err, SVC_SOCK_DEFAULTS);
1753 svc_sock_received(svsk);
1761 if (proto) *proto = so->sk->sk_protocol;
1762 return one_sock_name(name_return, svsk);
1764 EXPORT_SYMBOL_GPL(svc_addsock);
1767 * Create socket for RPC service.
1769 static int svc_create_socket(struct svc_serv *serv, int protocol,
1770 struct sockaddr *sin, int len, int flags)
1772 struct svc_sock *svsk;
1773 struct socket *sock;
1776 char buf[RPC_MAX_ADDRBUFLEN];
1778 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1779 serv->sv_program->pg_name, protocol,
1780 __svc_print_addr(sin, buf, sizeof(buf)));
1782 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1783 printk(KERN_WARNING "svc: only UDP and TCP "
1784 "sockets supported\n");
1787 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1789 error = sock_create_kern(sin->sa_family, type, protocol, &sock);
1793 svc_reclassify_socket(sock);
1795 if (type == SOCK_STREAM)
1796 sock->sk->sk_reuse = 1; /* allow address reuse */
1797 error = kernel_bind(sock, sin, len);
1801 if (protocol == IPPROTO_TCP) {
1802 if ((error = kernel_listen(sock, 64)) < 0)
1806 if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1807 svc_sock_received(svsk);
1808 return ntohs(inet_sk(svsk->sk_sk)->sport);
1812 dprintk("svc: svc_create_socket error = %d\n", -error);
1818 * Remove a dead socket
1821 svc_delete_socket(struct svc_sock *svsk)
1823 struct svc_serv *serv;
1826 dprintk("svc: svc_delete_socket(%p)\n", svsk);
1828 serv = svsk->sk_server;
1831 sk->sk_state_change = svsk->sk_ostate;
1832 sk->sk_data_ready = svsk->sk_odata;
1833 sk->sk_write_space = svsk->sk_owspace;
1835 spin_lock_bh(&serv->sv_lock);
1837 if (!test_and_set_bit(SK_DETACHED, &svsk->sk_flags))
1838 list_del_init(&svsk->sk_list);
1840 * We used to delete the svc_sock from whichever list
1841 * it's sk_ready node was on, but we don't actually
1842 * need to. This is because the only time we're called
1843 * while still attached to a queue, the queue itself
1844 * is about to be destroyed (in svc_destroy).
1846 if (!test_and_set_bit(SK_DEAD, &svsk->sk_flags)) {
1847 BUG_ON(atomic_read(&svsk->sk_inuse)<2);
1848 atomic_dec(&svsk->sk_inuse);
1849 if (test_bit(SK_TEMP, &svsk->sk_flags))
1853 spin_unlock_bh(&serv->sv_lock);
1856 static void svc_close_socket(struct svc_sock *svsk)
1858 set_bit(SK_CLOSE, &svsk->sk_flags);
1859 if (test_and_set_bit(SK_BUSY, &svsk->sk_flags))
1860 /* someone else will have to effect the close */
1863 atomic_inc(&svsk->sk_inuse);
1864 svc_delete_socket(svsk);
1865 clear_bit(SK_BUSY, &svsk->sk_flags);
1869 void svc_force_close_socket(struct svc_sock *svsk)
1871 set_bit(SK_CLOSE, &svsk->sk_flags);
1872 if (test_bit(SK_BUSY, &svsk->sk_flags)) {
1873 /* Waiting to be processed, but no threads left,
1874 * So just remove it from the waiting list
1876 list_del_init(&svsk->sk_ready);
1877 clear_bit(SK_BUSY, &svsk->sk_flags);
1879 svc_close_socket(svsk);
1883 * svc_makesock - Make a socket for nfsd and lockd
1884 * @serv: RPC server structure
1885 * @protocol: transport protocol to use
1886 * @port: port to use
1887 * @flags: requested socket characteristics
1890 int svc_makesock(struct svc_serv *serv, int protocol, unsigned short port,
1893 struct sockaddr_in sin = {
1894 .sin_family = AF_INET,
1895 .sin_addr.s_addr = INADDR_ANY,
1896 .sin_port = htons(port),
1899 dprintk("svc: creating socket proto = %d\n", protocol);
1900 return svc_create_socket(serv, protocol, (struct sockaddr *) &sin,
1901 sizeof(sin), flags);
1905 * Handle defer and revisit of requests
1908 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1910 struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1911 struct svc_sock *svsk;
1914 svc_sock_put(dr->svsk);
1918 dprintk("revisit queued\n");
1921 spin_lock(&svsk->sk_lock);
1922 list_add(&dr->handle.recent, &svsk->sk_deferred);
1923 spin_unlock(&svsk->sk_lock);
1924 set_bit(SK_DEFERRED, &svsk->sk_flags);
1925 svc_sock_enqueue(svsk);
1929 static struct cache_deferred_req *
1930 svc_defer(struct cache_req *req)
1932 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1933 int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
1934 struct svc_deferred_req *dr;
1936 if (rqstp->rq_arg.page_len)
1937 return NULL; /* if more than a page, give up FIXME */
1938 if (rqstp->rq_deferred) {
1939 dr = rqstp->rq_deferred;
1940 rqstp->rq_deferred = NULL;
1942 int skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1943 /* FIXME maybe discard if size too large */
1944 dr = kmalloc(size, GFP_KERNEL);
1948 dr->handle.owner = rqstp->rq_server;
1949 dr->prot = rqstp->rq_prot;
1950 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1951 dr->addrlen = rqstp->rq_addrlen;
1952 dr->daddr = rqstp->rq_daddr;
1953 dr->argslen = rqstp->rq_arg.len >> 2;
1954 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
1956 atomic_inc(&rqstp->rq_sock->sk_inuse);
1957 dr->svsk = rqstp->rq_sock;
1959 dr->handle.revisit = svc_revisit;
1964 * recv data from a deferred request into an active one
1966 static int svc_deferred_recv(struct svc_rqst *rqstp)
1968 struct svc_deferred_req *dr = rqstp->rq_deferred;
1970 rqstp->rq_arg.head[0].iov_base = dr->args;
1971 rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
1972 rqstp->rq_arg.page_len = 0;
1973 rqstp->rq_arg.len = dr->argslen<<2;
1974 rqstp->rq_prot = dr->prot;
1975 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1976 rqstp->rq_addrlen = dr->addrlen;
1977 rqstp->rq_daddr = dr->daddr;
1978 rqstp->rq_respages = rqstp->rq_pages;
1979 return dr->argslen<<2;
1983 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk)
1985 struct svc_deferred_req *dr = NULL;
1987 if (!test_bit(SK_DEFERRED, &svsk->sk_flags))
1989 spin_lock(&svsk->sk_lock);
1990 clear_bit(SK_DEFERRED, &svsk->sk_flags);
1991 if (!list_empty(&svsk->sk_deferred)) {
1992 dr = list_entry(svsk->sk_deferred.next,
1993 struct svc_deferred_req,
1995 list_del_init(&dr->handle.recent);
1996 set_bit(SK_DEFERRED, &svsk->sk_flags);
1998 spin_unlock(&svsk->sk_lock);