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);
88 static void svc_sock_detach(struct svc_xprt *);
89 static void svc_sock_free(struct svc_xprt *);
91 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk);
92 static int svc_deferred_recv(struct svc_rqst *rqstp);
93 static struct cache_deferred_req *svc_defer(struct cache_req *req);
94 static struct svc_xprt *svc_create_socket(struct svc_serv *, int,
95 struct sockaddr *, int, int);
97 /* apparently the "standard" is that clients close
98 * idle connections after 5 minutes, servers after
100 * http://www.connectathon.org/talks96/nfstcp.pdf
102 static int svc_conn_age_period = 6*60;
104 #ifdef CONFIG_DEBUG_LOCK_ALLOC
105 static struct lock_class_key svc_key[2];
106 static struct lock_class_key svc_slock_key[2];
108 static inline void svc_reclassify_socket(struct socket *sock)
110 struct sock *sk = sock->sk;
111 BUG_ON(sock_owned_by_user(sk));
112 switch (sk->sk_family) {
114 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFSD",
115 &svc_slock_key[0], "sk_lock-AF_INET-NFSD", &svc_key[0]);
119 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFSD",
120 &svc_slock_key[1], "sk_lock-AF_INET6-NFSD", &svc_key[1]);
128 static inline void svc_reclassify_socket(struct socket *sock)
133 static char *__svc_print_addr(struct sockaddr *addr, char *buf, size_t len)
135 switch (addr->sa_family) {
137 snprintf(buf, len, "%u.%u.%u.%u, port=%u",
138 NIPQUAD(((struct sockaddr_in *) addr)->sin_addr),
139 ntohs(((struct sockaddr_in *) addr)->sin_port));
143 snprintf(buf, len, "%x:%x:%x:%x:%x:%x:%x:%x, port=%u",
144 NIP6(((struct sockaddr_in6 *) addr)->sin6_addr),
145 ntohs(((struct sockaddr_in6 *) addr)->sin6_port));
149 snprintf(buf, len, "unknown address type: %d", addr->sa_family);
156 * svc_print_addr - Format rq_addr field for printing
157 * @rqstp: svc_rqst struct containing address to print
158 * @buf: target buffer for formatted address
159 * @len: length of target buffer
162 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
164 return __svc_print_addr(svc_addr(rqstp), buf, len);
166 EXPORT_SYMBOL_GPL(svc_print_addr);
169 * Queue up an idle server thread. Must have pool->sp_lock held.
170 * Note: this is really a stack rather than a queue, so that we only
171 * use as many different threads as we need, and the rest don't pollute
175 svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
177 list_add(&rqstp->rq_list, &pool->sp_threads);
181 * Dequeue an nfsd thread. Must have pool->sp_lock held.
184 svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
186 list_del(&rqstp->rq_list);
190 * Release an skbuff after use
192 static void svc_release_skb(struct svc_rqst *rqstp)
194 struct sk_buff *skb = rqstp->rq_xprt_ctxt;
195 struct svc_deferred_req *dr = rqstp->rq_deferred;
198 rqstp->rq_xprt_ctxt = NULL;
200 dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
201 skb_free_datagram(rqstp->rq_sock->sk_sk, skb);
204 rqstp->rq_deferred = NULL;
210 * Queue up a socket with data pending. If there are idle nfsd
211 * processes, wake 'em up.
215 svc_sock_enqueue(struct svc_sock *svsk)
217 struct svc_serv *serv = svsk->sk_server;
218 struct svc_pool *pool;
219 struct svc_rqst *rqstp;
222 if (!(svsk->sk_flags &
223 ( (1<<SK_CONN)|(1<<SK_DATA)|(1<<SK_CLOSE)|(1<<SK_DEFERRED)) ))
225 if (test_bit(SK_DEAD, &svsk->sk_flags))
229 pool = svc_pool_for_cpu(svsk->sk_server, cpu);
232 spin_lock_bh(&pool->sp_lock);
234 if (!list_empty(&pool->sp_threads) &&
235 !list_empty(&pool->sp_sockets))
237 "svc_sock_enqueue: threads and sockets both waiting??\n");
239 if (test_bit(SK_DEAD, &svsk->sk_flags)) {
240 /* Don't enqueue dead sockets */
241 dprintk("svc: socket %p is dead, not enqueued\n", svsk->sk_sk);
245 /* Mark socket as busy. It will remain in this state until the
246 * server has processed all pending data and put the socket back
247 * on the idle list. We update SK_BUSY atomically because
248 * it also guards against trying to enqueue the svc_sock twice.
250 if (test_and_set_bit(SK_BUSY, &svsk->sk_flags)) {
251 /* Don't enqueue socket while already enqueued */
252 dprintk("svc: socket %p busy, not enqueued\n", svsk->sk_sk);
255 BUG_ON(svsk->sk_pool != NULL);
256 svsk->sk_pool = pool;
258 /* Handle pending connection */
259 if (test_bit(SK_CONN, &svsk->sk_flags))
262 /* Handle close in-progress */
263 if (test_bit(SK_CLOSE, &svsk->sk_flags))
266 /* Check if we have space to reply to a request */
267 if (!svsk->sk_xprt.xpt_ops->xpo_has_wspace(&svsk->sk_xprt)) {
268 /* Don't enqueue while not enough space for reply */
269 dprintk("svc: no write space, socket %p not enqueued\n", svsk);
270 svsk->sk_pool = NULL;
271 clear_bit(SK_BUSY, &svsk->sk_flags);
276 if (!list_empty(&pool->sp_threads)) {
277 rqstp = list_entry(pool->sp_threads.next,
280 dprintk("svc: socket %p served by daemon %p\n",
282 svc_thread_dequeue(pool, rqstp);
285 "svc_sock_enqueue: server %p, rq_sock=%p!\n",
286 rqstp, rqstp->rq_sock);
287 rqstp->rq_sock = svsk;
288 atomic_inc(&svsk->sk_inuse);
289 rqstp->rq_reserved = serv->sv_max_mesg;
290 atomic_add(rqstp->rq_reserved, &svsk->sk_reserved);
291 BUG_ON(svsk->sk_pool != pool);
292 wake_up(&rqstp->rq_wait);
294 dprintk("svc: socket %p put into queue\n", svsk->sk_sk);
295 list_add_tail(&svsk->sk_ready, &pool->sp_sockets);
296 BUG_ON(svsk->sk_pool != pool);
300 spin_unlock_bh(&pool->sp_lock);
304 * Dequeue the first socket. Must be called with the pool->sp_lock held.
306 static inline struct svc_sock *
307 svc_sock_dequeue(struct svc_pool *pool)
309 struct svc_sock *svsk;
311 if (list_empty(&pool->sp_sockets))
314 svsk = list_entry(pool->sp_sockets.next,
315 struct svc_sock, sk_ready);
316 list_del_init(&svsk->sk_ready);
318 dprintk("svc: socket %p dequeued, inuse=%d\n",
319 svsk->sk_sk, atomic_read(&svsk->sk_inuse));
325 * Having read something from a socket, check whether it
326 * needs to be re-enqueued.
327 * Note: SK_DATA only gets cleared when a read-attempt finds
328 * no (or insufficient) data.
331 svc_sock_received(struct svc_sock *svsk)
333 svsk->sk_pool = NULL;
334 clear_bit(SK_BUSY, &svsk->sk_flags);
335 svc_sock_enqueue(svsk);
340 * svc_reserve - change the space reserved for the reply to a request.
341 * @rqstp: The request in question
342 * @space: new max space to reserve
344 * Each request reserves some space on the output queue of the socket
345 * to make sure the reply fits. This function reduces that reserved
346 * space to be the amount of space used already, plus @space.
349 void svc_reserve(struct svc_rqst *rqstp, int space)
351 space += rqstp->rq_res.head[0].iov_len;
353 if (space < rqstp->rq_reserved) {
354 struct svc_sock *svsk = rqstp->rq_sock;
355 atomic_sub((rqstp->rq_reserved - space), &svsk->sk_reserved);
356 rqstp->rq_reserved = space;
358 svc_sock_enqueue(svsk);
363 * Release a socket after use.
366 svc_sock_put(struct svc_sock *svsk)
368 if (atomic_dec_and_test(&svsk->sk_inuse)) {
369 BUG_ON(!test_bit(SK_DEAD, &svsk->sk_flags));
370 module_put(svsk->sk_xprt.xpt_class->xcl_owner);
371 svsk->sk_xprt.xpt_ops->xpo_free(&svsk->sk_xprt);
376 svc_sock_release(struct svc_rqst *rqstp)
378 struct svc_sock *svsk = rqstp->rq_sock;
380 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
382 svc_free_res_pages(rqstp);
383 rqstp->rq_res.page_len = 0;
384 rqstp->rq_res.page_base = 0;
387 /* Reset response buffer and release
389 * But first, check that enough space was reserved
390 * for the reply, otherwise we have a bug!
392 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
393 printk(KERN_ERR "RPC request reserved %d but used %d\n",
397 rqstp->rq_res.head[0].iov_len = 0;
398 svc_reserve(rqstp, 0);
399 rqstp->rq_sock = NULL;
405 * External function to wake up a server waiting for data
406 * This really only makes sense for services like lockd
407 * which have exactly one thread anyway.
410 svc_wake_up(struct svc_serv *serv)
412 struct svc_rqst *rqstp;
414 struct svc_pool *pool;
416 for (i = 0; i < serv->sv_nrpools; i++) {
417 pool = &serv->sv_pools[i];
419 spin_lock_bh(&pool->sp_lock);
420 if (!list_empty(&pool->sp_threads)) {
421 rqstp = list_entry(pool->sp_threads.next,
424 dprintk("svc: daemon %p woken up.\n", rqstp);
426 svc_thread_dequeue(pool, rqstp);
427 rqstp->rq_sock = NULL;
429 wake_up(&rqstp->rq_wait);
431 spin_unlock_bh(&pool->sp_lock);
435 union svc_pktinfo_u {
436 struct in_pktinfo pkti;
437 struct in6_pktinfo pkti6;
439 #define SVC_PKTINFO_SPACE \
440 CMSG_SPACE(sizeof(union svc_pktinfo_u))
442 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
444 switch (rqstp->rq_sock->sk_sk->sk_family) {
446 struct in_pktinfo *pki = CMSG_DATA(cmh);
448 cmh->cmsg_level = SOL_IP;
449 cmh->cmsg_type = IP_PKTINFO;
450 pki->ipi_ifindex = 0;
451 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr.addr.s_addr;
452 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
457 struct in6_pktinfo *pki = CMSG_DATA(cmh);
459 cmh->cmsg_level = SOL_IPV6;
460 cmh->cmsg_type = IPV6_PKTINFO;
461 pki->ipi6_ifindex = 0;
462 ipv6_addr_copy(&pki->ipi6_addr,
463 &rqstp->rq_daddr.addr6);
464 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
472 * Generic sendto routine
475 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
477 struct svc_sock *svsk = rqstp->rq_sock;
478 struct socket *sock = svsk->sk_sock;
482 long all[SVC_PKTINFO_SPACE / sizeof(long)];
484 struct cmsghdr *cmh = &buffer.hdr;
488 struct page **ppage = xdr->pages;
489 size_t base = xdr->page_base;
490 unsigned int pglen = xdr->page_len;
491 unsigned int flags = MSG_MORE;
492 char buf[RPC_MAX_ADDRBUFLEN];
496 if (rqstp->rq_prot == IPPROTO_UDP) {
497 struct msghdr msg = {
498 .msg_name = &rqstp->rq_addr,
499 .msg_namelen = rqstp->rq_addrlen,
501 .msg_controllen = sizeof(buffer),
502 .msg_flags = MSG_MORE,
505 svc_set_cmsg_data(rqstp, cmh);
507 if (sock_sendmsg(sock, &msg, 0) < 0)
512 if (slen == xdr->head[0].iov_len)
514 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0,
515 xdr->head[0].iov_len, flags);
516 if (len != xdr->head[0].iov_len)
518 slen -= xdr->head[0].iov_len;
523 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
527 result = kernel_sendpage(sock, *ppage, base, size, flags);
534 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
539 if (xdr->tail[0].iov_len) {
540 result = kernel_sendpage(sock, rqstp->rq_respages[0],
541 ((unsigned long)xdr->tail[0].iov_base)
543 xdr->tail[0].iov_len, 0);
549 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
550 rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len,
551 xdr->len, len, svc_print_addr(rqstp, buf, sizeof(buf)));
557 * Report socket names for nfsdfs
559 static int one_sock_name(char *buf, struct svc_sock *svsk)
563 switch(svsk->sk_sk->sk_family) {
565 len = sprintf(buf, "ipv4 %s %u.%u.%u.%u %d\n",
566 svsk->sk_sk->sk_protocol==IPPROTO_UDP?
568 NIPQUAD(inet_sk(svsk->sk_sk)->rcv_saddr),
569 inet_sk(svsk->sk_sk)->num);
572 len = sprintf(buf, "*unknown-%d*\n",
573 svsk->sk_sk->sk_family);
579 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
581 struct svc_sock *svsk, *closesk = NULL;
586 spin_lock_bh(&serv->sv_lock);
587 list_for_each_entry(svsk, &serv->sv_permsocks, sk_list) {
588 int onelen = one_sock_name(buf+len, svsk);
589 if (toclose && strcmp(toclose, buf+len) == 0)
594 spin_unlock_bh(&serv->sv_lock);
596 /* Should unregister with portmap, but you cannot
597 * unregister just one protocol...
599 svc_close_socket(closesk);
604 EXPORT_SYMBOL(svc_sock_names);
607 * Check input queue length
610 svc_recv_available(struct svc_sock *svsk)
612 struct socket *sock = svsk->sk_sock;
615 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
617 return (err >= 0)? avail : err;
621 * Generic recvfrom routine.
624 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
626 struct svc_sock *svsk = rqstp->rq_sock;
627 struct msghdr msg = {
628 .msg_flags = MSG_DONTWAIT,
630 struct sockaddr *sin;
633 len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
636 /* sock_recvmsg doesn't fill in the name/namelen, so we must..
638 memcpy(&rqstp->rq_addr, &svsk->sk_remote, svsk->sk_remotelen);
639 rqstp->rq_addrlen = svsk->sk_remotelen;
641 /* Destination address in request is needed for binding the
642 * source address in RPC callbacks later.
644 sin = (struct sockaddr *)&svsk->sk_local;
645 switch (sin->sa_family) {
647 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
650 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
654 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
655 svsk, iov[0].iov_base, iov[0].iov_len, len);
661 * Set socket snd and rcv buffer lengths
664 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
668 oldfs = get_fs(); set_fs(KERNEL_DS);
669 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
670 (char*)&snd, sizeof(snd));
671 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
672 (char*)&rcv, sizeof(rcv));
674 /* sock_setsockopt limits use to sysctl_?mem_max,
675 * which isn't acceptable. Until that is made conditional
676 * on not having CAP_SYS_RESOURCE or similar, we go direct...
677 * DaveM said I could!
680 sock->sk->sk_sndbuf = snd * 2;
681 sock->sk->sk_rcvbuf = rcv * 2;
682 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
683 release_sock(sock->sk);
687 * INET callback when data has been received on the socket.
690 svc_udp_data_ready(struct sock *sk, int count)
692 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
695 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
696 svsk, sk, count, test_bit(SK_BUSY, &svsk->sk_flags));
697 set_bit(SK_DATA, &svsk->sk_flags);
698 svc_sock_enqueue(svsk);
700 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
701 wake_up_interruptible(sk->sk_sleep);
705 * INET callback when space is newly available on the socket.
708 svc_write_space(struct sock *sk)
710 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
713 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
714 svsk, sk, test_bit(SK_BUSY, &svsk->sk_flags));
715 svc_sock_enqueue(svsk);
718 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
719 dprintk("RPC svc_write_space: someone sleeping on %p\n",
721 wake_up_interruptible(sk->sk_sleep);
725 static inline void svc_udp_get_dest_address(struct svc_rqst *rqstp,
728 switch (rqstp->rq_sock->sk_sk->sk_family) {
730 struct in_pktinfo *pki = CMSG_DATA(cmh);
731 rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
735 struct in6_pktinfo *pki = CMSG_DATA(cmh);
736 ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
743 * Receive a datagram from a UDP socket.
746 svc_udp_recvfrom(struct svc_rqst *rqstp)
748 struct svc_sock *svsk = rqstp->rq_sock;
749 struct svc_serv *serv = svsk->sk_server;
753 long all[SVC_PKTINFO_SPACE / sizeof(long)];
755 struct cmsghdr *cmh = &buffer.hdr;
757 struct msghdr msg = {
758 .msg_name = svc_addr(rqstp),
760 .msg_controllen = sizeof(buffer),
761 .msg_flags = MSG_DONTWAIT,
764 if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
765 /* udp sockets need large rcvbuf as all pending
766 * requests are still in that buffer. sndbuf must
767 * also be large enough that there is enough space
768 * for one reply per thread. We count all threads
769 * rather than threads in a particular pool, which
770 * provides an upper bound on the number of threads
771 * which will access the socket.
773 svc_sock_setbufsize(svsk->sk_sock,
774 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
775 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
777 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
778 svc_sock_received(svsk);
779 return svc_deferred_recv(rqstp);
782 clear_bit(SK_DATA, &svsk->sk_flags);
784 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
785 0, 0, MSG_PEEK | MSG_DONTWAIT);
787 skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err);
790 if (err != -EAGAIN) {
791 /* possibly an icmp error */
792 dprintk("svc: recvfrom returned error %d\n", -err);
793 set_bit(SK_DATA, &svsk->sk_flags);
795 svc_sock_received(svsk);
798 rqstp->rq_addrlen = sizeof(rqstp->rq_addr);
799 if (skb->tstamp.tv64 == 0) {
800 skb->tstamp = ktime_get_real();
801 /* Don't enable netstamp, sunrpc doesn't
802 need that much accuracy */
804 svsk->sk_sk->sk_stamp = skb->tstamp;
805 set_bit(SK_DATA, &svsk->sk_flags); /* there may be more data... */
808 * Maybe more packets - kick another thread ASAP.
810 svc_sock_received(svsk);
812 len = skb->len - sizeof(struct udphdr);
813 rqstp->rq_arg.len = len;
815 rqstp->rq_prot = IPPROTO_UDP;
817 if (cmh->cmsg_level != IPPROTO_IP ||
818 cmh->cmsg_type != IP_PKTINFO) {
820 printk("rpcsvc: received unknown control message:"
822 cmh->cmsg_level, cmh->cmsg_type);
823 skb_free_datagram(svsk->sk_sk, skb);
826 svc_udp_get_dest_address(rqstp, cmh);
828 if (skb_is_nonlinear(skb)) {
829 /* we have to copy */
831 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
834 skb_free_datagram(svsk->sk_sk, skb);
838 skb_free_datagram(svsk->sk_sk, skb);
840 /* we can use it in-place */
841 rqstp->rq_arg.head[0].iov_base = skb->data + sizeof(struct udphdr);
842 rqstp->rq_arg.head[0].iov_len = len;
843 if (skb_checksum_complete(skb)) {
844 skb_free_datagram(svsk->sk_sk, skb);
847 rqstp->rq_xprt_ctxt = skb;
850 rqstp->rq_arg.page_base = 0;
851 if (len <= rqstp->rq_arg.head[0].iov_len) {
852 rqstp->rq_arg.head[0].iov_len = len;
853 rqstp->rq_arg.page_len = 0;
854 rqstp->rq_respages = rqstp->rq_pages+1;
856 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
857 rqstp->rq_respages = rqstp->rq_pages + 1 +
858 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
862 serv->sv_stats->netudpcnt++;
868 svc_udp_sendto(struct svc_rqst *rqstp)
872 error = svc_sendto(rqstp, &rqstp->rq_res);
873 if (error == -ECONNREFUSED)
874 /* ICMP error on earlier request. */
875 error = svc_sendto(rqstp, &rqstp->rq_res);
880 static void svc_udp_prep_reply_hdr(struct svc_rqst *rqstp)
884 static int svc_udp_has_wspace(struct svc_xprt *xprt)
886 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
887 struct svc_serv *serv = svsk->sk_server;
888 unsigned long required;
891 * Set the SOCK_NOSPACE flag before checking the available
894 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
895 required = atomic_read(&svsk->sk_reserved) + serv->sv_max_mesg;
896 if (required*2 > sock_wspace(svsk->sk_sk))
898 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
902 static struct svc_xprt *svc_udp_accept(struct svc_xprt *xprt)
908 static struct svc_xprt *svc_udp_create(struct svc_serv *serv,
909 struct sockaddr *sa, int salen,
912 return svc_create_socket(serv, IPPROTO_UDP, sa, salen, flags);
915 static struct svc_xprt_ops svc_udp_ops = {
916 .xpo_create = svc_udp_create,
917 .xpo_recvfrom = svc_udp_recvfrom,
918 .xpo_sendto = svc_udp_sendto,
919 .xpo_release_rqst = svc_release_skb,
920 .xpo_detach = svc_sock_detach,
921 .xpo_free = svc_sock_free,
922 .xpo_prep_reply_hdr = svc_udp_prep_reply_hdr,
923 .xpo_has_wspace = svc_udp_has_wspace,
924 .xpo_accept = svc_udp_accept,
927 static struct svc_xprt_class svc_udp_class = {
929 .xcl_owner = THIS_MODULE,
930 .xcl_ops = &svc_udp_ops,
931 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
935 svc_udp_init(struct svc_sock *svsk)
940 svc_xprt_init(&svc_udp_class, &svsk->sk_xprt);
941 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
942 svsk->sk_sk->sk_write_space = svc_write_space;
944 /* initialise setting must have enough space to
945 * receive and respond to one request.
946 * svc_udp_recvfrom will re-adjust if necessary
948 svc_sock_setbufsize(svsk->sk_sock,
949 3 * svsk->sk_server->sv_max_mesg,
950 3 * svsk->sk_server->sv_max_mesg);
952 set_bit(SK_DATA, &svsk->sk_flags); /* might have come in before data_ready set up */
953 set_bit(SK_CHNGBUF, &svsk->sk_flags);
957 /* make sure we get destination address info */
958 svsk->sk_sock->ops->setsockopt(svsk->sk_sock, IPPROTO_IP, IP_PKTINFO,
959 (char __user *)&one, sizeof(one));
964 * A data_ready event on a listening socket means there's a connection
965 * pending. Do not use state_change as a substitute for it.
968 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
970 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
972 dprintk("svc: socket %p TCP (listen) state change %d\n",
976 * This callback may called twice when a new connection
977 * is established as a child socket inherits everything
978 * from a parent LISTEN socket.
979 * 1) data_ready method of the parent socket will be called
980 * when one of child sockets become ESTABLISHED.
981 * 2) data_ready method of the child socket may be called
982 * when it receives data before the socket is accepted.
983 * In case of 2, we should ignore it silently.
985 if (sk->sk_state == TCP_LISTEN) {
987 set_bit(SK_CONN, &svsk->sk_flags);
988 svc_sock_enqueue(svsk);
990 printk("svc: socket %p: no user data\n", sk);
993 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
994 wake_up_interruptible_all(sk->sk_sleep);
998 * A state change on a connected socket means it's dying or dead.
1001 svc_tcp_state_change(struct sock *sk)
1003 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
1005 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
1006 sk, sk->sk_state, sk->sk_user_data);
1009 printk("svc: socket %p: no user data\n", sk);
1011 set_bit(SK_CLOSE, &svsk->sk_flags);
1012 svc_sock_enqueue(svsk);
1014 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1015 wake_up_interruptible_all(sk->sk_sleep);
1019 svc_tcp_data_ready(struct sock *sk, int count)
1021 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
1023 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
1024 sk, sk->sk_user_data);
1026 set_bit(SK_DATA, &svsk->sk_flags);
1027 svc_sock_enqueue(svsk);
1029 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1030 wake_up_interruptible(sk->sk_sleep);
1033 static inline int svc_port_is_privileged(struct sockaddr *sin)
1035 switch (sin->sa_family) {
1037 return ntohs(((struct sockaddr_in *)sin)->sin_port)
1040 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
1048 * Accept a TCP connection
1050 static struct svc_xprt *svc_tcp_accept(struct svc_xprt *xprt)
1052 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1053 struct sockaddr_storage addr;
1054 struct sockaddr *sin = (struct sockaddr *) &addr;
1055 struct svc_serv *serv = svsk->sk_server;
1056 struct socket *sock = svsk->sk_sock;
1057 struct socket *newsock;
1058 struct svc_sock *newsvsk;
1060 char buf[RPC_MAX_ADDRBUFLEN];
1062 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
1066 clear_bit(SK_CONN, &svsk->sk_flags);
1067 err = kernel_accept(sock, &newsock, O_NONBLOCK);
1070 printk(KERN_WARNING "%s: no more sockets!\n",
1072 else if (err != -EAGAIN && net_ratelimit())
1073 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
1074 serv->sv_name, -err);
1078 set_bit(SK_CONN, &svsk->sk_flags);
1080 err = kernel_getpeername(newsock, sin, &slen);
1082 if (net_ratelimit())
1083 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
1084 serv->sv_name, -err);
1085 goto failed; /* aborted connection or whatever */
1088 /* Ideally, we would want to reject connections from unauthorized
1089 * hosts here, but when we get encryption, the IP of the host won't
1090 * tell us anything. For now just warn about unpriv connections.
1092 if (!svc_port_is_privileged(sin)) {
1093 dprintk(KERN_WARNING
1094 "%s: connect from unprivileged port: %s\n",
1096 __svc_print_addr(sin, buf, sizeof(buf)));
1098 dprintk("%s: connect from %s\n", serv->sv_name,
1099 __svc_print_addr(sin, buf, sizeof(buf)));
1101 /* make sure that a write doesn't block forever when
1104 newsock->sk->sk_sndtimeo = HZ*30;
1106 if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
1107 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
1109 memcpy(&newsvsk->sk_remote, sin, slen);
1110 newsvsk->sk_remotelen = slen;
1111 err = kernel_getsockname(newsock, sin, &slen);
1112 if (unlikely(err < 0)) {
1113 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err);
1114 slen = offsetof(struct sockaddr, sa_data);
1116 memcpy(&newsvsk->sk_local, sin, slen);
1118 svc_sock_received(newsvsk);
1121 serv->sv_stats->nettcpconn++;
1123 return &newsvsk->sk_xprt;
1126 sock_release(newsock);
1131 * Make sure that we don't have too many active connections. If we
1132 * have, something must be dropped.
1134 * There's no point in trying to do random drop here for DoS
1135 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
1136 * attacker can easily beat that.
1138 * The only somewhat efficient mechanism would be if drop old
1139 * connections from the same IP first. But right now we don't even
1140 * record the client IP in svc_sock.
1142 static void svc_check_conn_limits(struct svc_serv *serv)
1144 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
1145 struct svc_sock *svsk = NULL;
1146 spin_lock_bh(&serv->sv_lock);
1147 if (!list_empty(&serv->sv_tempsocks)) {
1148 if (net_ratelimit()) {
1149 /* Try to help the admin */
1150 printk(KERN_NOTICE "%s: too many open TCP "
1151 "sockets, consider increasing the "
1152 "number of nfsd threads\n",
1156 * Always select the oldest socket. It's not fair,
1159 svsk = list_entry(serv->sv_tempsocks.prev,
1162 set_bit(SK_CLOSE, &svsk->sk_flags);
1163 atomic_inc(&svsk->sk_inuse);
1165 spin_unlock_bh(&serv->sv_lock);
1168 svc_sock_enqueue(svsk);
1175 * Receive data from a TCP socket.
1178 svc_tcp_recvfrom(struct svc_rqst *rqstp)
1180 struct svc_sock *svsk = rqstp->rq_sock;
1181 struct svc_serv *serv = svsk->sk_server;
1186 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
1187 svsk, test_bit(SK_DATA, &svsk->sk_flags),
1188 test_bit(SK_CONN, &svsk->sk_flags),
1189 test_bit(SK_CLOSE, &svsk->sk_flags));
1191 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
1192 svc_sock_received(svsk);
1193 return svc_deferred_recv(rqstp);
1196 if (test_and_clear_bit(SK_CHNGBUF, &svsk->sk_flags))
1197 /* sndbuf needs to have room for one request
1198 * per thread, otherwise we can stall even when the
1199 * network isn't a bottleneck.
1201 * We count all threads rather than threads in a
1202 * particular pool, which provides an upper bound
1203 * on the number of threads which will access the socket.
1205 * rcvbuf just needs to be able to hold a few requests.
1206 * Normally they will be removed from the queue
1207 * as soon a a complete request arrives.
1209 svc_sock_setbufsize(svsk->sk_sock,
1210 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
1211 3 * serv->sv_max_mesg);
1213 clear_bit(SK_DATA, &svsk->sk_flags);
1215 /* Receive data. If we haven't got the record length yet, get
1216 * the next four bytes. Otherwise try to gobble up as much as
1217 * possible up to the complete record length.
1219 if (svsk->sk_tcplen < 4) {
1220 unsigned long want = 4 - svsk->sk_tcplen;
1223 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
1225 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
1227 svsk->sk_tcplen += len;
1230 dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
1232 svc_sock_received(svsk);
1233 return -EAGAIN; /* record header not complete */
1236 svsk->sk_reclen = ntohl(svsk->sk_reclen);
1237 if (!(svsk->sk_reclen & 0x80000000)) {
1238 /* FIXME: technically, a record can be fragmented,
1239 * and non-terminal fragments will not have the top
1240 * bit set in the fragment length header.
1241 * But apparently no known nfs clients send fragmented
1243 if (net_ratelimit())
1244 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1245 " (non-terminal)\n",
1246 (unsigned long) svsk->sk_reclen);
1249 svsk->sk_reclen &= 0x7fffffff;
1250 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
1251 if (svsk->sk_reclen > serv->sv_max_mesg) {
1252 if (net_ratelimit())
1253 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1255 (unsigned long) svsk->sk_reclen);
1260 /* Check whether enough data is available */
1261 len = svc_recv_available(svsk);
1265 if (len < svsk->sk_reclen) {
1266 dprintk("svc: incomplete TCP record (%d of %d)\n",
1267 len, svsk->sk_reclen);
1268 svc_sock_received(svsk);
1269 return -EAGAIN; /* record not complete */
1271 len = svsk->sk_reclen;
1272 set_bit(SK_DATA, &svsk->sk_flags);
1274 vec = rqstp->rq_vec;
1275 vec[0] = rqstp->rq_arg.head[0];
1278 while (vlen < len) {
1279 vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
1280 vec[pnum].iov_len = PAGE_SIZE;
1284 rqstp->rq_respages = &rqstp->rq_pages[pnum];
1286 /* Now receive data */
1287 len = svc_recvfrom(rqstp, vec, pnum, len);
1291 dprintk("svc: TCP complete record (%d bytes)\n", len);
1292 rqstp->rq_arg.len = len;
1293 rqstp->rq_arg.page_base = 0;
1294 if (len <= rqstp->rq_arg.head[0].iov_len) {
1295 rqstp->rq_arg.head[0].iov_len = len;
1296 rqstp->rq_arg.page_len = 0;
1298 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1301 rqstp->rq_xprt_ctxt = NULL;
1302 rqstp->rq_prot = IPPROTO_TCP;
1304 /* Reset TCP read info */
1305 svsk->sk_reclen = 0;
1306 svsk->sk_tcplen = 0;
1308 svc_sock_received(svsk);
1310 serv->sv_stats->nettcpcnt++;
1315 set_bit(SK_CLOSE, &svsk->sk_flags);
1319 if (len == -EAGAIN) {
1320 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1321 svc_sock_received(svsk);
1323 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1324 svsk->sk_server->sv_name, -len);
1332 * Send out data on TCP socket.
1335 svc_tcp_sendto(struct svc_rqst *rqstp)
1337 struct xdr_buf *xbufp = &rqstp->rq_res;
1341 /* Set up the first element of the reply kvec.
1342 * Any other kvecs that may be in use have been taken
1343 * care of by the server implementation itself.
1345 reclen = htonl(0x80000000|((xbufp->len ) - 4));
1346 memcpy(xbufp->head[0].iov_base, &reclen, 4);
1348 if (test_bit(SK_DEAD, &rqstp->rq_sock->sk_flags))
1351 sent = svc_sendto(rqstp, &rqstp->rq_res);
1352 if (sent != xbufp->len) {
1353 printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1354 rqstp->rq_sock->sk_server->sv_name,
1355 (sent<0)?"got error":"sent only",
1357 set_bit(SK_CLOSE, &rqstp->rq_sock->sk_flags);
1358 svc_sock_enqueue(rqstp->rq_sock);
1365 * Setup response header. TCP has a 4B record length field.
1367 static void svc_tcp_prep_reply_hdr(struct svc_rqst *rqstp)
1369 struct kvec *resv = &rqstp->rq_res.head[0];
1371 /* tcp needs a space for the record length... */
1375 static int svc_tcp_has_wspace(struct svc_xprt *xprt)
1377 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1378 struct svc_serv *serv = svsk->sk_server;
1383 * Set the SOCK_NOSPACE flag before checking the available
1386 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1387 required = atomic_read(&svsk->sk_reserved) + serv->sv_max_mesg;
1388 wspace = sk_stream_wspace(svsk->sk_sk);
1390 if (wspace < sk_stream_min_wspace(svsk->sk_sk))
1392 if (required * 2 > wspace)
1395 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1399 static struct svc_xprt *svc_tcp_create(struct svc_serv *serv,
1400 struct sockaddr *sa, int salen,
1403 return svc_create_socket(serv, IPPROTO_TCP, sa, salen, flags);
1406 static struct svc_xprt_ops svc_tcp_ops = {
1407 .xpo_create = svc_tcp_create,
1408 .xpo_recvfrom = svc_tcp_recvfrom,
1409 .xpo_sendto = svc_tcp_sendto,
1410 .xpo_release_rqst = svc_release_skb,
1411 .xpo_detach = svc_sock_detach,
1412 .xpo_free = svc_sock_free,
1413 .xpo_prep_reply_hdr = svc_tcp_prep_reply_hdr,
1414 .xpo_has_wspace = svc_tcp_has_wspace,
1415 .xpo_accept = svc_tcp_accept,
1418 static struct svc_xprt_class svc_tcp_class = {
1420 .xcl_owner = THIS_MODULE,
1421 .xcl_ops = &svc_tcp_ops,
1422 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1425 void svc_init_xprt_sock(void)
1427 svc_reg_xprt_class(&svc_tcp_class);
1428 svc_reg_xprt_class(&svc_udp_class);
1431 void svc_cleanup_xprt_sock(void)
1433 svc_unreg_xprt_class(&svc_tcp_class);
1434 svc_unreg_xprt_class(&svc_udp_class);
1438 svc_tcp_init(struct svc_sock *svsk)
1440 struct sock *sk = svsk->sk_sk;
1441 struct tcp_sock *tp = tcp_sk(sk);
1443 svc_xprt_init(&svc_tcp_class, &svsk->sk_xprt);
1445 if (sk->sk_state == TCP_LISTEN) {
1446 dprintk("setting up TCP socket for listening\n");
1447 set_bit(SK_LISTENER, &svsk->sk_flags);
1448 sk->sk_data_ready = svc_tcp_listen_data_ready;
1449 set_bit(SK_CONN, &svsk->sk_flags);
1451 dprintk("setting up TCP socket for reading\n");
1452 sk->sk_state_change = svc_tcp_state_change;
1453 sk->sk_data_ready = svc_tcp_data_ready;
1454 sk->sk_write_space = svc_write_space;
1456 svsk->sk_reclen = 0;
1457 svsk->sk_tcplen = 0;
1459 tp->nonagle = 1; /* disable Nagle's algorithm */
1461 /* initialise setting must have enough space to
1462 * receive and respond to one request.
1463 * svc_tcp_recvfrom will re-adjust if necessary
1465 svc_sock_setbufsize(svsk->sk_sock,
1466 3 * svsk->sk_server->sv_max_mesg,
1467 3 * svsk->sk_server->sv_max_mesg);
1469 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1470 set_bit(SK_DATA, &svsk->sk_flags);
1471 if (sk->sk_state != TCP_ESTABLISHED)
1472 set_bit(SK_CLOSE, &svsk->sk_flags);
1477 svc_sock_update_bufs(struct svc_serv *serv)
1480 * The number of server threads has changed. Update
1481 * rcvbuf and sndbuf accordingly on all sockets
1483 struct list_head *le;
1485 spin_lock_bh(&serv->sv_lock);
1486 list_for_each(le, &serv->sv_permsocks) {
1487 struct svc_sock *svsk =
1488 list_entry(le, struct svc_sock, sk_list);
1489 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1491 list_for_each(le, &serv->sv_tempsocks) {
1492 struct svc_sock *svsk =
1493 list_entry(le, struct svc_sock, sk_list);
1494 set_bit(SK_CHNGBUF, &svsk->sk_flags);
1496 spin_unlock_bh(&serv->sv_lock);
1500 * Receive the next request on any socket. This code is carefully
1501 * organised not to touch any cachelines in the shared svc_serv
1502 * structure, only cachelines in the local svc_pool.
1505 svc_recv(struct svc_rqst *rqstp, long timeout)
1507 struct svc_sock *svsk = NULL;
1508 struct svc_serv *serv = rqstp->rq_server;
1509 struct svc_pool *pool = rqstp->rq_pool;
1512 struct xdr_buf *arg;
1513 DECLARE_WAITQUEUE(wait, current);
1515 dprintk("svc: server %p waiting for data (to = %ld)\n",
1520 "svc_recv: service %p, socket not NULL!\n",
1522 if (waitqueue_active(&rqstp->rq_wait))
1524 "svc_recv: service %p, wait queue active!\n",
1528 /* now allocate needed pages. If we get a failure, sleep briefly */
1529 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
1530 for (i=0; i < pages ; i++)
1531 while (rqstp->rq_pages[i] == NULL) {
1532 struct page *p = alloc_page(GFP_KERNEL);
1534 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1535 rqstp->rq_pages[i] = p;
1537 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
1538 BUG_ON(pages >= RPCSVC_MAXPAGES);
1540 /* Make arg->head point to first page and arg->pages point to rest */
1541 arg = &rqstp->rq_arg;
1542 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
1543 arg->head[0].iov_len = PAGE_SIZE;
1544 arg->pages = rqstp->rq_pages + 1;
1546 /* save at least one page for response */
1547 arg->page_len = (pages-2)*PAGE_SIZE;
1548 arg->len = (pages-1)*PAGE_SIZE;
1549 arg->tail[0].iov_len = 0;
1556 spin_lock_bh(&pool->sp_lock);
1557 if ((svsk = svc_sock_dequeue(pool)) != NULL) {
1558 rqstp->rq_sock = svsk;
1559 atomic_inc(&svsk->sk_inuse);
1560 rqstp->rq_reserved = serv->sv_max_mesg;
1561 atomic_add(rqstp->rq_reserved, &svsk->sk_reserved);
1563 /* No data pending. Go to sleep */
1564 svc_thread_enqueue(pool, rqstp);
1567 * We have to be able to interrupt this wait
1568 * to bring down the daemons ...
1570 set_current_state(TASK_INTERRUPTIBLE);
1571 add_wait_queue(&rqstp->rq_wait, &wait);
1572 spin_unlock_bh(&pool->sp_lock);
1574 schedule_timeout(timeout);
1578 spin_lock_bh(&pool->sp_lock);
1579 remove_wait_queue(&rqstp->rq_wait, &wait);
1581 if (!(svsk = rqstp->rq_sock)) {
1582 svc_thread_dequeue(pool, rqstp);
1583 spin_unlock_bh(&pool->sp_lock);
1584 dprintk("svc: server %p, no data yet\n", rqstp);
1585 return signalled()? -EINTR : -EAGAIN;
1588 spin_unlock_bh(&pool->sp_lock);
1591 if (test_bit(SK_CLOSE, &svsk->sk_flags)) {
1592 dprintk("svc_recv: found SK_CLOSE\n");
1593 svc_delete_socket(svsk);
1594 } else if (test_bit(SK_LISTENER, &svsk->sk_flags)) {
1595 struct svc_xprt *newxpt;
1596 newxpt = svsk->sk_xprt.xpt_ops->xpo_accept(&svsk->sk_xprt);
1599 * We know this module_get will succeed because the
1600 * listener holds a reference too
1602 __module_get(newxpt->xpt_class->xcl_owner);
1603 svc_check_conn_limits(svsk->sk_server);
1605 svc_sock_received(svsk);
1607 dprintk("svc: server %p, pool %u, socket %p, inuse=%d\n",
1608 rqstp, pool->sp_id, svsk, atomic_read(&svsk->sk_inuse));
1609 len = svsk->sk_xprt.xpt_ops->xpo_recvfrom(rqstp);
1610 dprintk("svc: got len=%d\n", len);
1613 /* No data, incomplete (TCP) read, or accept() */
1614 if (len == 0 || len == -EAGAIN) {
1615 rqstp->rq_res.len = 0;
1616 svc_sock_release(rqstp);
1619 svsk->sk_lastrecv = get_seconds();
1620 clear_bit(SK_OLD, &svsk->sk_flags);
1622 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
1623 rqstp->rq_chandle.defer = svc_defer;
1626 serv->sv_stats->netcnt++;
1634 svc_drop(struct svc_rqst *rqstp)
1636 dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1637 svc_sock_release(rqstp);
1641 * Return reply to client.
1644 svc_send(struct svc_rqst *rqstp)
1646 struct svc_sock *svsk;
1650 if ((svsk = rqstp->rq_sock) == NULL) {
1651 printk(KERN_WARNING "NULL socket pointer in %s:%d\n",
1652 __FILE__, __LINE__);
1656 /* release the receive skb before sending the reply */
1657 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
1659 /* calculate over-all length */
1660 xb = & rqstp->rq_res;
1661 xb->len = xb->head[0].iov_len +
1663 xb->tail[0].iov_len;
1665 /* Grab svsk->sk_mutex to serialize outgoing data. */
1666 mutex_lock(&svsk->sk_mutex);
1667 if (test_bit(SK_DEAD, &svsk->sk_flags))
1670 len = svsk->sk_xprt.xpt_ops->xpo_sendto(rqstp);
1671 mutex_unlock(&svsk->sk_mutex);
1672 svc_sock_release(rqstp);
1674 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1680 * Timer function to close old temporary sockets, using
1681 * a mark-and-sweep algorithm.
1684 svc_age_temp_sockets(unsigned long closure)
1686 struct svc_serv *serv = (struct svc_serv *)closure;
1687 struct svc_sock *svsk;
1688 struct list_head *le, *next;
1689 LIST_HEAD(to_be_aged);
1691 dprintk("svc_age_temp_sockets\n");
1693 if (!spin_trylock_bh(&serv->sv_lock)) {
1694 /* busy, try again 1 sec later */
1695 dprintk("svc_age_temp_sockets: busy\n");
1696 mod_timer(&serv->sv_temptimer, jiffies + HZ);
1700 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1701 svsk = list_entry(le, struct svc_sock, sk_list);
1703 if (!test_and_set_bit(SK_OLD, &svsk->sk_flags))
1705 if (atomic_read(&svsk->sk_inuse) > 1 || test_bit(SK_BUSY, &svsk->sk_flags))
1707 atomic_inc(&svsk->sk_inuse);
1708 list_move(le, &to_be_aged);
1709 set_bit(SK_CLOSE, &svsk->sk_flags);
1710 set_bit(SK_DETACHED, &svsk->sk_flags);
1712 spin_unlock_bh(&serv->sv_lock);
1714 while (!list_empty(&to_be_aged)) {
1715 le = to_be_aged.next;
1716 /* fiddling the sk_list node is safe 'cos we're SK_DETACHED */
1718 svsk = list_entry(le, struct svc_sock, sk_list);
1720 dprintk("queuing svsk %p for closing, %lu seconds old\n",
1721 svsk, get_seconds() - svsk->sk_lastrecv);
1723 /* a thread will dequeue and close it soon */
1724 svc_sock_enqueue(svsk);
1728 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
1732 * Initialize socket for RPC use and create svc_sock struct
1733 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1735 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1736 struct socket *sock,
1737 int *errp, int flags)
1739 struct svc_sock *svsk;
1741 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1742 int is_temporary = flags & SVC_SOCK_TEMPORARY;
1744 dprintk("svc: svc_setup_socket %p\n", sock);
1745 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1752 /* Register socket with portmapper */
1753 if (*errp >= 0 && pmap_register)
1754 *errp = svc_register(serv, inet->sk_protocol,
1755 ntohs(inet_sk(inet)->sport));
1762 set_bit(SK_BUSY, &svsk->sk_flags);
1763 inet->sk_user_data = svsk;
1764 svsk->sk_sock = sock;
1766 svsk->sk_ostate = inet->sk_state_change;
1767 svsk->sk_odata = inet->sk_data_ready;
1768 svsk->sk_owspace = inet->sk_write_space;
1769 svsk->sk_server = serv;
1770 atomic_set(&svsk->sk_inuse, 1);
1771 svsk->sk_lastrecv = get_seconds();
1772 spin_lock_init(&svsk->sk_lock);
1773 INIT_LIST_HEAD(&svsk->sk_deferred);
1774 INIT_LIST_HEAD(&svsk->sk_ready);
1775 mutex_init(&svsk->sk_mutex);
1777 /* Initialize the socket */
1778 if (sock->type == SOCK_DGRAM)
1783 spin_lock_bh(&serv->sv_lock);
1785 set_bit(SK_TEMP, &svsk->sk_flags);
1786 list_add(&svsk->sk_list, &serv->sv_tempsocks);
1788 if (serv->sv_temptimer.function == NULL) {
1789 /* setup timer to age temp sockets */
1790 setup_timer(&serv->sv_temptimer, svc_age_temp_sockets,
1791 (unsigned long)serv);
1792 mod_timer(&serv->sv_temptimer,
1793 jiffies + svc_conn_age_period * HZ);
1796 clear_bit(SK_TEMP, &svsk->sk_flags);
1797 list_add(&svsk->sk_list, &serv->sv_permsocks);
1799 spin_unlock_bh(&serv->sv_lock);
1801 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1807 int svc_addsock(struct svc_serv *serv,
1813 struct socket *so = sockfd_lookup(fd, &err);
1814 struct svc_sock *svsk = NULL;
1818 if (so->sk->sk_family != AF_INET)
1819 err = -EAFNOSUPPORT;
1820 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1821 so->sk->sk_protocol != IPPROTO_UDP)
1822 err = -EPROTONOSUPPORT;
1823 else if (so->state > SS_UNCONNECTED)
1826 svsk = svc_setup_socket(serv, so, &err, SVC_SOCK_DEFAULTS);
1828 svc_sock_received(svsk);
1836 if (proto) *proto = so->sk->sk_protocol;
1837 return one_sock_name(name_return, svsk);
1839 EXPORT_SYMBOL_GPL(svc_addsock);
1842 * Create socket for RPC service.
1844 static struct svc_xprt *svc_create_socket(struct svc_serv *serv,
1846 struct sockaddr *sin, int len,
1849 struct svc_sock *svsk;
1850 struct socket *sock;
1853 char buf[RPC_MAX_ADDRBUFLEN];
1855 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1856 serv->sv_program->pg_name, protocol,
1857 __svc_print_addr(sin, buf, sizeof(buf)));
1859 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1860 printk(KERN_WARNING "svc: only UDP and TCP "
1861 "sockets supported\n");
1862 return ERR_PTR(-EINVAL);
1864 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1866 error = sock_create_kern(sin->sa_family, type, protocol, &sock);
1868 return ERR_PTR(error);
1870 svc_reclassify_socket(sock);
1872 if (type == SOCK_STREAM)
1873 sock->sk->sk_reuse = 1; /* allow address reuse */
1874 error = kernel_bind(sock, sin, len);
1878 if (protocol == IPPROTO_TCP) {
1879 if ((error = kernel_listen(sock, 64)) < 0)
1883 if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1884 svc_sock_received(svsk);
1885 return (struct svc_xprt *)svsk;
1889 dprintk("svc: svc_create_socket error = %d\n", -error);
1891 return ERR_PTR(error);
1895 * Detach the svc_sock from the socket so that no
1896 * more callbacks occur.
1898 static void svc_sock_detach(struct svc_xprt *xprt)
1900 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1901 struct sock *sk = svsk->sk_sk;
1903 dprintk("svc: svc_sock_detach(%p)\n", svsk);
1905 /* put back the old socket callbacks */
1906 sk->sk_state_change = svsk->sk_ostate;
1907 sk->sk_data_ready = svsk->sk_odata;
1908 sk->sk_write_space = svsk->sk_owspace;
1912 * Free the svc_sock's socket resources and the svc_sock itself.
1914 static void svc_sock_free(struct svc_xprt *xprt)
1916 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1917 dprintk("svc: svc_sock_free(%p)\n", svsk);
1919 if (svsk->sk_info_authunix != NULL)
1920 svcauth_unix_info_release(svsk->sk_info_authunix);
1921 if (svsk->sk_sock->file)
1922 sockfd_put(svsk->sk_sock);
1924 sock_release(svsk->sk_sock);
1929 * Remove a dead socket
1932 svc_delete_socket(struct svc_sock *svsk)
1934 struct svc_serv *serv;
1937 dprintk("svc: svc_delete_socket(%p)\n", svsk);
1939 serv = svsk->sk_server;
1942 svsk->sk_xprt.xpt_ops->xpo_detach(&svsk->sk_xprt);
1944 spin_lock_bh(&serv->sv_lock);
1946 if (!test_and_set_bit(SK_DETACHED, &svsk->sk_flags))
1947 list_del_init(&svsk->sk_list);
1949 * We used to delete the svc_sock from whichever list
1950 * it's sk_ready node was on, but we don't actually
1951 * need to. This is because the only time we're called
1952 * while still attached to a queue, the queue itself
1953 * is about to be destroyed (in svc_destroy).
1955 if (!test_and_set_bit(SK_DEAD, &svsk->sk_flags)) {
1956 BUG_ON(atomic_read(&svsk->sk_inuse)<2);
1957 atomic_dec(&svsk->sk_inuse);
1958 if (test_bit(SK_TEMP, &svsk->sk_flags))
1962 spin_unlock_bh(&serv->sv_lock);
1965 static void svc_close_socket(struct svc_sock *svsk)
1967 set_bit(SK_CLOSE, &svsk->sk_flags);
1968 if (test_and_set_bit(SK_BUSY, &svsk->sk_flags))
1969 /* someone else will have to effect the close */
1972 atomic_inc(&svsk->sk_inuse);
1973 svc_delete_socket(svsk);
1974 clear_bit(SK_BUSY, &svsk->sk_flags);
1978 void svc_force_close_socket(struct svc_sock *svsk)
1980 set_bit(SK_CLOSE, &svsk->sk_flags);
1981 if (test_bit(SK_BUSY, &svsk->sk_flags)) {
1982 /* Waiting to be processed, but no threads left,
1983 * So just remove it from the waiting list
1985 list_del_init(&svsk->sk_ready);
1986 clear_bit(SK_BUSY, &svsk->sk_flags);
1988 svc_close_socket(svsk);
1992 * Handle defer and revisit of requests
1995 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1997 struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1998 struct svc_sock *svsk;
2001 svc_sock_put(dr->svsk);
2005 dprintk("revisit queued\n");
2008 spin_lock(&svsk->sk_lock);
2009 list_add(&dr->handle.recent, &svsk->sk_deferred);
2010 spin_unlock(&svsk->sk_lock);
2011 set_bit(SK_DEFERRED, &svsk->sk_flags);
2012 svc_sock_enqueue(svsk);
2016 static struct cache_deferred_req *
2017 svc_defer(struct cache_req *req)
2019 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
2020 int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
2021 struct svc_deferred_req *dr;
2023 if (rqstp->rq_arg.page_len)
2024 return NULL; /* if more than a page, give up FIXME */
2025 if (rqstp->rq_deferred) {
2026 dr = rqstp->rq_deferred;
2027 rqstp->rq_deferred = NULL;
2029 int skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
2030 /* FIXME maybe discard if size too large */
2031 dr = kmalloc(size, GFP_KERNEL);
2035 dr->handle.owner = rqstp->rq_server;
2036 dr->prot = rqstp->rq_prot;
2037 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
2038 dr->addrlen = rqstp->rq_addrlen;
2039 dr->daddr = rqstp->rq_daddr;
2040 dr->argslen = rqstp->rq_arg.len >> 2;
2041 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
2043 atomic_inc(&rqstp->rq_sock->sk_inuse);
2044 dr->svsk = rqstp->rq_sock;
2046 dr->handle.revisit = svc_revisit;
2051 * recv data from a deferred request into an active one
2053 static int svc_deferred_recv(struct svc_rqst *rqstp)
2055 struct svc_deferred_req *dr = rqstp->rq_deferred;
2057 rqstp->rq_arg.head[0].iov_base = dr->args;
2058 rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
2059 rqstp->rq_arg.page_len = 0;
2060 rqstp->rq_arg.len = dr->argslen<<2;
2061 rqstp->rq_prot = dr->prot;
2062 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
2063 rqstp->rq_addrlen = dr->addrlen;
2064 rqstp->rq_daddr = dr->daddr;
2065 rqstp->rq_respages = rqstp->rq_pages;
2066 return dr->argslen<<2;
2070 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk)
2072 struct svc_deferred_req *dr = NULL;
2074 if (!test_bit(SK_DEFERRED, &svsk->sk_flags))
2076 spin_lock(&svsk->sk_lock);
2077 clear_bit(SK_DEFERRED, &svsk->sk_flags);
2078 if (!list_empty(&svsk->sk_deferred)) {
2079 dr = list_entry(svsk->sk_deferred.next,
2080 struct svc_deferred_req,
2082 list_del_init(&dr->handle.recent);
2083 set_bit(SK_DEFERRED, &svsk->sk_flags);
2085 spin_unlock(&svsk->sk_lock);