2 * linux/net/sunrpc/svcsock.c
4 * These are the RPC server socket internals.
6 * The server scheduling algorithm does not always distribute the load
7 * evenly when servicing a single client. May need to modify the
8 * svc_xprt_enqueue procedure...
10 * TCP support is largely untested and may be a little slow. The problem
11 * is that we currently do two separate recvfrom's, one for the 4-byte
12 * record length, and the second for the actual record. This could possibly
13 * be improved by always reading a minimum size of around 100 bytes and
14 * tucking any superfluous bytes away in a temporary store. Still, that
15 * leaves write requests out in the rain. An alternative may be to peek at
16 * the first skb in the queue, and if it matches the next TCP sequence
17 * number, to extract the record marker. Yuck.
19 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
22 #include <linux/kernel.h>
23 #include <linux/sched.h>
24 #include <linux/errno.h>
25 #include <linux/fcntl.h>
26 #include <linux/net.h>
28 #include <linux/inet.h>
29 #include <linux/udp.h>
30 #include <linux/tcp.h>
31 #include <linux/unistd.h>
32 #include <linux/slab.h>
33 #include <linux/netdevice.h>
34 #include <linux/skbuff.h>
35 #include <linux/file.h>
36 #include <linux/freezer.h>
38 #include <net/checksum.h>
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_xprt.xpt_flags.XPT_BUSY prevents a svc_sock being
62 * Some flags can be set to certain values at any time
63 * providing that certain rules are followed:
65 * XPT_CONN, XPT_DATA, can be set or cleared at any time.
66 * after a set, svc_xprt_enqueue must be called.
67 * after a clear, the socket must be read/accepted
68 * if this succeeds, it must be set again.
69 * XPT_CLOSE can set at any time. It is never cleared.
70 * xpt_ref contains a bias of '1' until XPT_DEAD is set.
71 * so when xprt_ref hits zero, we know the transport is dead
72 * and no-one is using it.
73 * XPT_DEAD can only be set while XPT_BUSY is held which ensures
74 * no other thread will be using the socket or will try to
79 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
82 static struct svc_sock *svc_setup_socket(struct svc_serv *, struct socket *,
83 int *errp, int flags);
84 static void svc_delete_xprt(struct svc_xprt *xprt);
85 static void svc_udp_data_ready(struct sock *, int);
86 static int svc_udp_recvfrom(struct svc_rqst *);
87 static int svc_udp_sendto(struct svc_rqst *);
88 static void svc_close_xprt(struct svc_xprt *xprt);
89 static void svc_sock_detach(struct svc_xprt *);
90 static void svc_sock_free(struct svc_xprt *);
92 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
93 static int svc_deferred_recv(struct svc_rqst *rqstp);
94 static struct cache_deferred_req *svc_defer(struct cache_req *req);
95 static struct svc_xprt *svc_create_socket(struct svc_serv *, int,
96 struct sockaddr *, int, int);
98 /* apparently the "standard" is that clients close
99 * idle connections after 5 minutes, servers after
101 * http://www.connectathon.org/talks96/nfstcp.pdf
103 static int svc_conn_age_period = 6*60;
105 #ifdef CONFIG_DEBUG_LOCK_ALLOC
106 static struct lock_class_key svc_key[2];
107 static struct lock_class_key svc_slock_key[2];
109 static inline void svc_reclassify_socket(struct socket *sock)
111 struct sock *sk = sock->sk;
112 BUG_ON(sock_owned_by_user(sk));
113 switch (sk->sk_family) {
115 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFSD",
117 "sk_xprt.xpt_lock-AF_INET-NFSD",
122 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFSD",
124 "sk_xprt.xpt_lock-AF_INET6-NFSD",
133 static inline void svc_reclassify_socket(struct socket *sock)
138 static char *__svc_print_addr(struct sockaddr *addr, char *buf, size_t len)
140 switch (addr->sa_family) {
142 snprintf(buf, len, "%u.%u.%u.%u, port=%u",
143 NIPQUAD(((struct sockaddr_in *) addr)->sin_addr),
144 ntohs(((struct sockaddr_in *) addr)->sin_port));
148 snprintf(buf, len, "%x:%x:%x:%x:%x:%x:%x:%x, port=%u",
149 NIP6(((struct sockaddr_in6 *) addr)->sin6_addr),
150 ntohs(((struct sockaddr_in6 *) addr)->sin6_port));
154 snprintf(buf, len, "unknown address type: %d", addr->sa_family);
161 * svc_print_addr - Format rq_addr field for printing
162 * @rqstp: svc_rqst struct containing address to print
163 * @buf: target buffer for formatted address
164 * @len: length of target buffer
167 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
169 return __svc_print_addr(svc_addr(rqstp), buf, len);
171 EXPORT_SYMBOL_GPL(svc_print_addr);
174 * Queue up an idle server thread. Must have pool->sp_lock held.
175 * Note: this is really a stack rather than a queue, so that we only
176 * use as many different threads as we need, and the rest don't pollute
180 svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
182 list_add(&rqstp->rq_list, &pool->sp_threads);
186 * Dequeue an nfsd thread. Must have pool->sp_lock held.
189 svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
191 list_del(&rqstp->rq_list);
195 * Release an skbuff after use
197 static void svc_release_skb(struct svc_rqst *rqstp)
199 struct sk_buff *skb = rqstp->rq_xprt_ctxt;
200 struct svc_deferred_req *dr = rqstp->rq_deferred;
203 rqstp->rq_xprt_ctxt = NULL;
205 dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
206 skb_free_datagram(rqstp->rq_sock->sk_sk, skb);
209 rqstp->rq_deferred = NULL;
215 * Queue up a socket with data pending. If there are idle nfsd
216 * processes, wake 'em up.
219 void svc_xprt_enqueue(struct svc_xprt *xprt)
221 struct svc_serv *serv = xprt->xpt_server;
222 struct svc_pool *pool;
223 struct svc_rqst *rqstp;
226 if (!(xprt->xpt_flags &
227 ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
229 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
233 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
236 spin_lock_bh(&pool->sp_lock);
238 if (!list_empty(&pool->sp_threads) &&
239 !list_empty(&pool->sp_sockets))
242 "threads and transports both waiting??\n");
244 if (test_bit(XPT_DEAD, &xprt->xpt_flags)) {
245 /* Don't enqueue dead sockets */
246 dprintk("svc: transport %p is dead, not enqueued\n", xprt);
250 /* Mark socket as busy. It will remain in this state until the
251 * server has processed all pending data and put the socket back
252 * on the idle list. We update XPT_BUSY atomically because
253 * it also guards against trying to enqueue the svc_sock twice.
255 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
256 /* Don't enqueue socket while already enqueued */
257 dprintk("svc: transport %p busy, not enqueued\n", xprt);
260 BUG_ON(xprt->xpt_pool != NULL);
261 xprt->xpt_pool = pool;
263 /* Handle pending connection */
264 if (test_bit(XPT_CONN, &xprt->xpt_flags))
267 /* Handle close in-progress */
268 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
271 /* Check if we have space to reply to a request */
272 if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
273 /* Don't enqueue while not enough space for reply */
274 dprintk("svc: no write space, transport %p not enqueued\n",
276 xprt->xpt_pool = NULL;
277 clear_bit(XPT_BUSY, &xprt->xpt_flags);
282 if (!list_empty(&pool->sp_threads)) {
283 rqstp = list_entry(pool->sp_threads.next,
286 dprintk("svc: transport %p served by daemon %p\n",
288 svc_thread_dequeue(pool, rqstp);
291 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
292 rqstp, rqstp->rq_xprt);
293 rqstp->rq_xprt = xprt;
295 rqstp->rq_reserved = serv->sv_max_mesg;
296 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
297 BUG_ON(xprt->xpt_pool != pool);
298 wake_up(&rqstp->rq_wait);
300 dprintk("svc: transport %p put into queue\n", xprt);
301 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
302 BUG_ON(xprt->xpt_pool != pool);
306 spin_unlock_bh(&pool->sp_lock);
308 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
311 * Dequeue the first socket. Must be called with the pool->sp_lock held.
313 static inline struct svc_sock *
314 svc_sock_dequeue(struct svc_pool *pool)
316 struct svc_sock *svsk;
318 if (list_empty(&pool->sp_sockets))
321 svsk = list_entry(pool->sp_sockets.next,
322 struct svc_sock, sk_xprt.xpt_ready);
323 list_del_init(&svsk->sk_xprt.xpt_ready);
325 dprintk("svc: socket %p dequeued, inuse=%d\n",
326 svsk->sk_sk, atomic_read(&svsk->sk_xprt.xpt_ref.refcount));
332 * svc_xprt_received conditionally queues the transport for processing
333 * by another thread. The caller must hold the XPT_BUSY bit and must
334 * not thereafter touch transport data.
336 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
337 * insufficient) data.
339 void svc_xprt_received(struct svc_xprt *xprt)
341 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
342 xprt->xpt_pool = NULL;
343 clear_bit(XPT_BUSY, &xprt->xpt_flags);
344 svc_xprt_enqueue(xprt);
346 EXPORT_SYMBOL_GPL(svc_xprt_received);
349 * svc_reserve - change the space reserved for the reply to a request.
350 * @rqstp: The request in question
351 * @space: new max space to reserve
353 * Each request reserves some space on the output queue of the socket
354 * to make sure the reply fits. This function reduces that reserved
355 * space to be the amount of space used already, plus @space.
358 void svc_reserve(struct svc_rqst *rqstp, int space)
360 space += rqstp->rq_res.head[0].iov_len;
362 if (space < rqstp->rq_reserved) {
363 struct svc_xprt *xprt = rqstp->rq_xprt;
364 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
365 rqstp->rq_reserved = space;
367 svc_xprt_enqueue(xprt);
372 svc_sock_release(struct svc_rqst *rqstp)
374 struct svc_sock *svsk = rqstp->rq_sock;
376 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
378 svc_free_res_pages(rqstp);
379 rqstp->rq_res.page_len = 0;
380 rqstp->rq_res.page_base = 0;
383 /* Reset response buffer and release
385 * But first, check that enough space was reserved
386 * for the reply, otherwise we have a bug!
388 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
389 printk(KERN_ERR "RPC request reserved %d but used %d\n",
393 rqstp->rq_res.head[0].iov_len = 0;
394 svc_reserve(rqstp, 0);
395 rqstp->rq_sock = NULL;
397 svc_xprt_put(&svsk->sk_xprt);
401 * External function to wake up a server waiting for data
402 * This really only makes sense for services like lockd
403 * which have exactly one thread anyway.
406 svc_wake_up(struct svc_serv *serv)
408 struct svc_rqst *rqstp;
410 struct svc_pool *pool;
412 for (i = 0; i < serv->sv_nrpools; i++) {
413 pool = &serv->sv_pools[i];
415 spin_lock_bh(&pool->sp_lock);
416 if (!list_empty(&pool->sp_threads)) {
417 rqstp = list_entry(pool->sp_threads.next,
420 dprintk("svc: daemon %p woken up.\n", rqstp);
422 svc_thread_dequeue(pool, rqstp);
423 rqstp->rq_sock = NULL;
425 wake_up(&rqstp->rq_wait);
427 spin_unlock_bh(&pool->sp_lock);
431 union svc_pktinfo_u {
432 struct in_pktinfo pkti;
433 struct in6_pktinfo pkti6;
435 #define SVC_PKTINFO_SPACE \
436 CMSG_SPACE(sizeof(union svc_pktinfo_u))
438 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
440 switch (rqstp->rq_sock->sk_sk->sk_family) {
442 struct in_pktinfo *pki = CMSG_DATA(cmh);
444 cmh->cmsg_level = SOL_IP;
445 cmh->cmsg_type = IP_PKTINFO;
446 pki->ipi_ifindex = 0;
447 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr.addr.s_addr;
448 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
453 struct in6_pktinfo *pki = CMSG_DATA(cmh);
455 cmh->cmsg_level = SOL_IPV6;
456 cmh->cmsg_type = IPV6_PKTINFO;
457 pki->ipi6_ifindex = 0;
458 ipv6_addr_copy(&pki->ipi6_addr,
459 &rqstp->rq_daddr.addr6);
460 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
468 * Generic sendto routine
471 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
473 struct svc_sock *svsk = rqstp->rq_sock;
474 struct socket *sock = svsk->sk_sock;
478 long all[SVC_PKTINFO_SPACE / sizeof(long)];
480 struct cmsghdr *cmh = &buffer.hdr;
484 struct page **ppage = xdr->pages;
485 size_t base = xdr->page_base;
486 unsigned int pglen = xdr->page_len;
487 unsigned int flags = MSG_MORE;
488 char buf[RPC_MAX_ADDRBUFLEN];
492 if (rqstp->rq_prot == IPPROTO_UDP) {
493 struct msghdr msg = {
494 .msg_name = &rqstp->rq_addr,
495 .msg_namelen = rqstp->rq_addrlen,
497 .msg_controllen = sizeof(buffer),
498 .msg_flags = MSG_MORE,
501 svc_set_cmsg_data(rqstp, cmh);
503 if (sock_sendmsg(sock, &msg, 0) < 0)
508 if (slen == xdr->head[0].iov_len)
510 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0,
511 xdr->head[0].iov_len, flags);
512 if (len != xdr->head[0].iov_len)
514 slen -= xdr->head[0].iov_len;
519 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
523 result = kernel_sendpage(sock, *ppage, base, size, flags);
530 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
535 if (xdr->tail[0].iov_len) {
536 result = kernel_sendpage(sock, rqstp->rq_respages[0],
537 ((unsigned long)xdr->tail[0].iov_base)
539 xdr->tail[0].iov_len, 0);
545 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
546 rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len,
547 xdr->len, len, svc_print_addr(rqstp, buf, sizeof(buf)));
553 * Report socket names for nfsdfs
555 static int one_sock_name(char *buf, struct svc_sock *svsk)
559 switch(svsk->sk_sk->sk_family) {
561 len = sprintf(buf, "ipv4 %s %u.%u.%u.%u %d\n",
562 svsk->sk_sk->sk_protocol==IPPROTO_UDP?
564 NIPQUAD(inet_sk(svsk->sk_sk)->rcv_saddr),
565 inet_sk(svsk->sk_sk)->num);
568 len = sprintf(buf, "*unknown-%d*\n",
569 svsk->sk_sk->sk_family);
575 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
577 struct svc_sock *svsk, *closesk = NULL;
582 spin_lock_bh(&serv->sv_lock);
583 list_for_each_entry(svsk, &serv->sv_permsocks, sk_xprt.xpt_list) {
584 int onelen = one_sock_name(buf+len, svsk);
585 if (toclose && strcmp(toclose, buf+len) == 0)
590 spin_unlock_bh(&serv->sv_lock);
592 /* Should unregister with portmap, but you cannot
593 * unregister just one protocol...
595 svc_close_xprt(&closesk->sk_xprt);
600 EXPORT_SYMBOL(svc_sock_names);
603 * Check input queue length
606 svc_recv_available(struct svc_sock *svsk)
608 struct socket *sock = svsk->sk_sock;
611 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
613 return (err >= 0)? avail : err;
617 * Generic recvfrom routine.
620 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
622 struct svc_sock *svsk = rqstp->rq_sock;
623 struct msghdr msg = {
624 .msg_flags = MSG_DONTWAIT,
626 struct sockaddr *sin;
629 len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
632 /* sock_recvmsg doesn't fill in the name/namelen, so we must..
634 memcpy(&rqstp->rq_addr, &svsk->sk_remote, svsk->sk_remotelen);
635 rqstp->rq_addrlen = svsk->sk_remotelen;
637 /* Destination address in request is needed for binding the
638 * source address in RPC callbacks later.
640 sin = (struct sockaddr *)&svsk->sk_local;
641 switch (sin->sa_family) {
643 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
646 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
650 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
651 svsk, iov[0].iov_base, iov[0].iov_len, len);
657 * Set socket snd and rcv buffer lengths
660 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
664 oldfs = get_fs(); set_fs(KERNEL_DS);
665 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
666 (char*)&snd, sizeof(snd));
667 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
668 (char*)&rcv, sizeof(rcv));
670 /* sock_setsockopt limits use to sysctl_?mem_max,
671 * which isn't acceptable. Until that is made conditional
672 * on not having CAP_SYS_RESOURCE or similar, we go direct...
673 * DaveM said I could!
676 sock->sk->sk_sndbuf = snd * 2;
677 sock->sk->sk_rcvbuf = rcv * 2;
678 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
679 release_sock(sock->sk);
683 * INET callback when data has been received on the socket.
686 svc_udp_data_ready(struct sock *sk, int count)
688 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
691 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
693 test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
694 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
695 svc_xprt_enqueue(&svsk->sk_xprt);
697 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
698 wake_up_interruptible(sk->sk_sleep);
702 * INET callback when space is newly available on the socket.
705 svc_write_space(struct sock *sk)
707 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
710 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
711 svsk, sk, test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
712 svc_xprt_enqueue(&svsk->sk_xprt);
715 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
716 dprintk("RPC svc_write_space: someone sleeping on %p\n",
718 wake_up_interruptible(sk->sk_sleep);
722 static inline void svc_udp_get_dest_address(struct svc_rqst *rqstp,
725 switch (rqstp->rq_sock->sk_sk->sk_family) {
727 struct in_pktinfo *pki = CMSG_DATA(cmh);
728 rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
732 struct in6_pktinfo *pki = CMSG_DATA(cmh);
733 ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
740 * Receive a datagram from a UDP socket.
743 svc_udp_recvfrom(struct svc_rqst *rqstp)
745 struct svc_sock *svsk = rqstp->rq_sock;
746 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
750 long all[SVC_PKTINFO_SPACE / sizeof(long)];
752 struct cmsghdr *cmh = &buffer.hdr;
754 struct msghdr msg = {
755 .msg_name = svc_addr(rqstp),
757 .msg_controllen = sizeof(buffer),
758 .msg_flags = MSG_DONTWAIT,
761 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
762 /* udp sockets need large rcvbuf as all pending
763 * requests are still in that buffer. sndbuf must
764 * also be large enough that there is enough space
765 * for one reply per thread. We count all threads
766 * rather than threads in a particular pool, which
767 * provides an upper bound on the number of threads
768 * which will access the socket.
770 svc_sock_setbufsize(svsk->sk_sock,
771 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
772 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
774 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
776 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
777 0, 0, MSG_PEEK | MSG_DONTWAIT);
779 skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err);
782 if (err != -EAGAIN) {
783 /* possibly an icmp error */
784 dprintk("svc: recvfrom returned error %d\n", -err);
785 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
787 svc_xprt_received(&svsk->sk_xprt);
790 rqstp->rq_addrlen = sizeof(rqstp->rq_addr);
791 if (skb->tstamp.tv64 == 0) {
792 skb->tstamp = ktime_get_real();
793 /* Don't enable netstamp, sunrpc doesn't
794 need that much accuracy */
796 svsk->sk_sk->sk_stamp = skb->tstamp;
797 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* there may be more data... */
800 * Maybe more packets - kick another thread ASAP.
802 svc_xprt_received(&svsk->sk_xprt);
804 len = skb->len - sizeof(struct udphdr);
805 rqstp->rq_arg.len = len;
807 rqstp->rq_prot = IPPROTO_UDP;
809 if (cmh->cmsg_level != IPPROTO_IP ||
810 cmh->cmsg_type != IP_PKTINFO) {
812 printk("rpcsvc: received unknown control message:"
814 cmh->cmsg_level, cmh->cmsg_type);
815 skb_free_datagram(svsk->sk_sk, skb);
818 svc_udp_get_dest_address(rqstp, cmh);
820 if (skb_is_nonlinear(skb)) {
821 /* we have to copy */
823 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
826 skb_free_datagram(svsk->sk_sk, skb);
830 skb_free_datagram(svsk->sk_sk, skb);
832 /* we can use it in-place */
833 rqstp->rq_arg.head[0].iov_base = skb->data + sizeof(struct udphdr);
834 rqstp->rq_arg.head[0].iov_len = len;
835 if (skb_checksum_complete(skb)) {
836 skb_free_datagram(svsk->sk_sk, skb);
839 rqstp->rq_xprt_ctxt = skb;
842 rqstp->rq_arg.page_base = 0;
843 if (len <= rqstp->rq_arg.head[0].iov_len) {
844 rqstp->rq_arg.head[0].iov_len = len;
845 rqstp->rq_arg.page_len = 0;
846 rqstp->rq_respages = rqstp->rq_pages+1;
848 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
849 rqstp->rq_respages = rqstp->rq_pages + 1 +
850 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
854 serv->sv_stats->netudpcnt++;
860 svc_udp_sendto(struct svc_rqst *rqstp)
864 error = svc_sendto(rqstp, &rqstp->rq_res);
865 if (error == -ECONNREFUSED)
866 /* ICMP error on earlier request. */
867 error = svc_sendto(rqstp, &rqstp->rq_res);
872 static void svc_udp_prep_reply_hdr(struct svc_rqst *rqstp)
876 static int svc_udp_has_wspace(struct svc_xprt *xprt)
878 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
879 struct svc_serv *serv = xprt->xpt_server;
880 unsigned long required;
883 * Set the SOCK_NOSPACE flag before checking the available
886 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
887 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
888 if (required*2 > sock_wspace(svsk->sk_sk))
890 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
894 static struct svc_xprt *svc_udp_accept(struct svc_xprt *xprt)
900 static struct svc_xprt *svc_udp_create(struct svc_serv *serv,
901 struct sockaddr *sa, int salen,
904 return svc_create_socket(serv, IPPROTO_UDP, sa, salen, flags);
907 static struct svc_xprt_ops svc_udp_ops = {
908 .xpo_create = svc_udp_create,
909 .xpo_recvfrom = svc_udp_recvfrom,
910 .xpo_sendto = svc_udp_sendto,
911 .xpo_release_rqst = svc_release_skb,
912 .xpo_detach = svc_sock_detach,
913 .xpo_free = svc_sock_free,
914 .xpo_prep_reply_hdr = svc_udp_prep_reply_hdr,
915 .xpo_has_wspace = svc_udp_has_wspace,
916 .xpo_accept = svc_udp_accept,
919 static struct svc_xprt_class svc_udp_class = {
921 .xcl_owner = THIS_MODULE,
922 .xcl_ops = &svc_udp_ops,
923 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
926 static void svc_udp_init(struct svc_sock *svsk, struct svc_serv *serv)
931 svc_xprt_init(&svc_udp_class, &svsk->sk_xprt, serv);
932 clear_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
933 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
934 svsk->sk_sk->sk_write_space = svc_write_space;
936 /* initialise setting must have enough space to
937 * receive and respond to one request.
938 * svc_udp_recvfrom will re-adjust if necessary
940 svc_sock_setbufsize(svsk->sk_sock,
941 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
942 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
944 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* might have come in before data_ready set up */
945 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
949 /* make sure we get destination address info */
950 svsk->sk_sock->ops->setsockopt(svsk->sk_sock, IPPROTO_IP, IP_PKTINFO,
951 (char __user *)&one, sizeof(one));
956 * A data_ready event on a listening socket means there's a connection
957 * pending. Do not use state_change as a substitute for it.
960 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
962 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
964 dprintk("svc: socket %p TCP (listen) state change %d\n",
968 * This callback may called twice when a new connection
969 * is established as a child socket inherits everything
970 * from a parent LISTEN socket.
971 * 1) data_ready method of the parent socket will be called
972 * when one of child sockets become ESTABLISHED.
973 * 2) data_ready method of the child socket may be called
974 * when it receives data before the socket is accepted.
975 * In case of 2, we should ignore it silently.
977 if (sk->sk_state == TCP_LISTEN) {
979 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
980 svc_xprt_enqueue(&svsk->sk_xprt);
982 printk("svc: socket %p: no user data\n", sk);
985 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
986 wake_up_interruptible_all(sk->sk_sleep);
990 * A state change on a connected socket means it's dying or dead.
993 svc_tcp_state_change(struct sock *sk)
995 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
997 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
998 sk, sk->sk_state, sk->sk_user_data);
1001 printk("svc: socket %p: no user data\n", sk);
1003 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1004 svc_xprt_enqueue(&svsk->sk_xprt);
1006 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1007 wake_up_interruptible_all(sk->sk_sleep);
1011 svc_tcp_data_ready(struct sock *sk, int count)
1013 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
1015 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
1016 sk, sk->sk_user_data);
1018 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1019 svc_xprt_enqueue(&svsk->sk_xprt);
1021 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1022 wake_up_interruptible(sk->sk_sleep);
1025 static inline int svc_port_is_privileged(struct sockaddr *sin)
1027 switch (sin->sa_family) {
1029 return ntohs(((struct sockaddr_in *)sin)->sin_port)
1032 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
1040 * Accept a TCP connection
1042 static struct svc_xprt *svc_tcp_accept(struct svc_xprt *xprt)
1044 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1045 struct sockaddr_storage addr;
1046 struct sockaddr *sin = (struct sockaddr *) &addr;
1047 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1048 struct socket *sock = svsk->sk_sock;
1049 struct socket *newsock;
1050 struct svc_sock *newsvsk;
1052 char buf[RPC_MAX_ADDRBUFLEN];
1054 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
1058 clear_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1059 err = kernel_accept(sock, &newsock, O_NONBLOCK);
1062 printk(KERN_WARNING "%s: no more sockets!\n",
1064 else if (err != -EAGAIN && net_ratelimit())
1065 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
1066 serv->sv_name, -err);
1069 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1071 err = kernel_getpeername(newsock, sin, &slen);
1073 if (net_ratelimit())
1074 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
1075 serv->sv_name, -err);
1076 goto failed; /* aborted connection or whatever */
1079 /* Ideally, we would want to reject connections from unauthorized
1080 * hosts here, but when we get encryption, the IP of the host won't
1081 * tell us anything. For now just warn about unpriv connections.
1083 if (!svc_port_is_privileged(sin)) {
1084 dprintk(KERN_WARNING
1085 "%s: connect from unprivileged port: %s\n",
1087 __svc_print_addr(sin, buf, sizeof(buf)));
1089 dprintk("%s: connect from %s\n", serv->sv_name,
1090 __svc_print_addr(sin, buf, sizeof(buf)));
1092 /* make sure that a write doesn't block forever when
1095 newsock->sk->sk_sndtimeo = HZ*30;
1097 if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
1098 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
1100 memcpy(&newsvsk->sk_remote, sin, slen);
1101 newsvsk->sk_remotelen = slen;
1102 err = kernel_getsockname(newsock, sin, &slen);
1103 if (unlikely(err < 0)) {
1104 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err);
1105 slen = offsetof(struct sockaddr, sa_data);
1107 memcpy(&newsvsk->sk_local, sin, slen);
1110 serv->sv_stats->nettcpconn++;
1112 return &newsvsk->sk_xprt;
1115 sock_release(newsock);
1120 * Receive data from a TCP socket.
1123 svc_tcp_recvfrom(struct svc_rqst *rqstp)
1125 struct svc_sock *svsk = rqstp->rq_sock;
1126 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1131 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
1132 svsk, test_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags),
1133 test_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags),
1134 test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags));
1136 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
1137 /* sndbuf needs to have room for one request
1138 * per thread, otherwise we can stall even when the
1139 * network isn't a bottleneck.
1141 * We count all threads rather than threads in a
1142 * particular pool, which provides an upper bound
1143 * on the number of threads which will access the socket.
1145 * rcvbuf just needs to be able to hold a few requests.
1146 * Normally they will be removed from the queue
1147 * as soon a a complete request arrives.
1149 svc_sock_setbufsize(svsk->sk_sock,
1150 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
1151 3 * serv->sv_max_mesg);
1153 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1155 /* Receive data. If we haven't got the record length yet, get
1156 * the next four bytes. Otherwise try to gobble up as much as
1157 * possible up to the complete record length.
1159 if (svsk->sk_tcplen < 4) {
1160 unsigned long want = 4 - svsk->sk_tcplen;
1163 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
1165 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
1167 svsk->sk_tcplen += len;
1170 dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
1172 svc_xprt_received(&svsk->sk_xprt);
1173 return -EAGAIN; /* record header not complete */
1176 svsk->sk_reclen = ntohl(svsk->sk_reclen);
1177 if (!(svsk->sk_reclen & 0x80000000)) {
1178 /* FIXME: technically, a record can be fragmented,
1179 * and non-terminal fragments will not have the top
1180 * bit set in the fragment length header.
1181 * But apparently no known nfs clients send fragmented
1183 if (net_ratelimit())
1184 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1185 " (non-terminal)\n",
1186 (unsigned long) svsk->sk_reclen);
1189 svsk->sk_reclen &= 0x7fffffff;
1190 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
1191 if (svsk->sk_reclen > serv->sv_max_mesg) {
1192 if (net_ratelimit())
1193 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1195 (unsigned long) svsk->sk_reclen);
1200 /* Check whether enough data is available */
1201 len = svc_recv_available(svsk);
1205 if (len < svsk->sk_reclen) {
1206 dprintk("svc: incomplete TCP record (%d of %d)\n",
1207 len, svsk->sk_reclen);
1208 svc_xprt_received(&svsk->sk_xprt);
1209 return -EAGAIN; /* record not complete */
1211 len = svsk->sk_reclen;
1212 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1214 vec = rqstp->rq_vec;
1215 vec[0] = rqstp->rq_arg.head[0];
1218 while (vlen < len) {
1219 vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
1220 vec[pnum].iov_len = PAGE_SIZE;
1224 rqstp->rq_respages = &rqstp->rq_pages[pnum];
1226 /* Now receive data */
1227 len = svc_recvfrom(rqstp, vec, pnum, len);
1231 dprintk("svc: TCP complete record (%d bytes)\n", len);
1232 rqstp->rq_arg.len = len;
1233 rqstp->rq_arg.page_base = 0;
1234 if (len <= rqstp->rq_arg.head[0].iov_len) {
1235 rqstp->rq_arg.head[0].iov_len = len;
1236 rqstp->rq_arg.page_len = 0;
1238 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1241 rqstp->rq_xprt_ctxt = NULL;
1242 rqstp->rq_prot = IPPROTO_TCP;
1244 /* Reset TCP read info */
1245 svsk->sk_reclen = 0;
1246 svsk->sk_tcplen = 0;
1248 svc_xprt_received(&svsk->sk_xprt);
1250 serv->sv_stats->nettcpcnt++;
1255 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1259 if (len == -EAGAIN) {
1260 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1261 svc_xprt_received(&svsk->sk_xprt);
1263 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1264 svsk->sk_xprt.xpt_server->sv_name, -len);
1272 * Send out data on TCP socket.
1275 svc_tcp_sendto(struct svc_rqst *rqstp)
1277 struct xdr_buf *xbufp = &rqstp->rq_res;
1281 /* Set up the first element of the reply kvec.
1282 * Any other kvecs that may be in use have been taken
1283 * care of by the server implementation itself.
1285 reclen = htonl(0x80000000|((xbufp->len ) - 4));
1286 memcpy(xbufp->head[0].iov_base, &reclen, 4);
1288 if (test_bit(XPT_DEAD, &rqstp->rq_sock->sk_xprt.xpt_flags))
1291 sent = svc_sendto(rqstp, &rqstp->rq_res);
1292 if (sent != xbufp->len) {
1293 printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1294 rqstp->rq_sock->sk_xprt.xpt_server->sv_name,
1295 (sent<0)?"got error":"sent only",
1297 set_bit(XPT_CLOSE, &rqstp->rq_sock->sk_xprt.xpt_flags);
1298 svc_xprt_enqueue(rqstp->rq_xprt);
1305 * Setup response header. TCP has a 4B record length field.
1307 static void svc_tcp_prep_reply_hdr(struct svc_rqst *rqstp)
1309 struct kvec *resv = &rqstp->rq_res.head[0];
1311 /* tcp needs a space for the record length... */
1315 static int svc_tcp_has_wspace(struct svc_xprt *xprt)
1317 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1318 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1323 * Set the SOCK_NOSPACE flag before checking the available
1326 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1327 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
1328 wspace = sk_stream_wspace(svsk->sk_sk);
1330 if (wspace < sk_stream_min_wspace(svsk->sk_sk))
1332 if (required * 2 > wspace)
1335 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1339 static struct svc_xprt *svc_tcp_create(struct svc_serv *serv,
1340 struct sockaddr *sa, int salen,
1343 return svc_create_socket(serv, IPPROTO_TCP, sa, salen, flags);
1346 static struct svc_xprt_ops svc_tcp_ops = {
1347 .xpo_create = svc_tcp_create,
1348 .xpo_recvfrom = svc_tcp_recvfrom,
1349 .xpo_sendto = svc_tcp_sendto,
1350 .xpo_release_rqst = svc_release_skb,
1351 .xpo_detach = svc_sock_detach,
1352 .xpo_free = svc_sock_free,
1353 .xpo_prep_reply_hdr = svc_tcp_prep_reply_hdr,
1354 .xpo_has_wspace = svc_tcp_has_wspace,
1355 .xpo_accept = svc_tcp_accept,
1358 static struct svc_xprt_class svc_tcp_class = {
1360 .xcl_owner = THIS_MODULE,
1361 .xcl_ops = &svc_tcp_ops,
1362 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1365 void svc_init_xprt_sock(void)
1367 svc_reg_xprt_class(&svc_tcp_class);
1368 svc_reg_xprt_class(&svc_udp_class);
1371 void svc_cleanup_xprt_sock(void)
1373 svc_unreg_xprt_class(&svc_tcp_class);
1374 svc_unreg_xprt_class(&svc_udp_class);
1377 static void svc_tcp_init(struct svc_sock *svsk, struct svc_serv *serv)
1379 struct sock *sk = svsk->sk_sk;
1380 struct tcp_sock *tp = tcp_sk(sk);
1382 svc_xprt_init(&svc_tcp_class, &svsk->sk_xprt, serv);
1383 set_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
1384 if (sk->sk_state == TCP_LISTEN) {
1385 dprintk("setting up TCP socket for listening\n");
1386 set_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags);
1387 sk->sk_data_ready = svc_tcp_listen_data_ready;
1388 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1390 dprintk("setting up TCP socket for reading\n");
1391 sk->sk_state_change = svc_tcp_state_change;
1392 sk->sk_data_ready = svc_tcp_data_ready;
1393 sk->sk_write_space = svc_write_space;
1395 svsk->sk_reclen = 0;
1396 svsk->sk_tcplen = 0;
1398 tp->nonagle = 1; /* disable Nagle's algorithm */
1400 /* initialise setting must have enough space to
1401 * receive and respond to one request.
1402 * svc_tcp_recvfrom will re-adjust if necessary
1404 svc_sock_setbufsize(svsk->sk_sock,
1405 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
1406 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
1408 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1409 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1410 if (sk->sk_state != TCP_ESTABLISHED)
1411 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1416 svc_sock_update_bufs(struct svc_serv *serv)
1419 * The number of server threads has changed. Update
1420 * rcvbuf and sndbuf accordingly on all sockets
1422 struct list_head *le;
1424 spin_lock_bh(&serv->sv_lock);
1425 list_for_each(le, &serv->sv_permsocks) {
1426 struct svc_sock *svsk =
1427 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1428 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1430 list_for_each(le, &serv->sv_tempsocks) {
1431 struct svc_sock *svsk =
1432 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1433 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1435 spin_unlock_bh(&serv->sv_lock);
1439 * Make sure that we don't have too many active connections. If we
1440 * have, something must be dropped.
1442 * There's no point in trying to do random drop here for DoS
1443 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
1444 * attacker can easily beat that.
1446 * The only somewhat efficient mechanism would be if drop old
1447 * connections from the same IP first. But right now we don't even
1448 * record the client IP in svc_sock.
1450 static void svc_check_conn_limits(struct svc_serv *serv)
1452 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
1453 struct svc_sock *svsk = NULL;
1454 spin_lock_bh(&serv->sv_lock);
1455 if (!list_empty(&serv->sv_tempsocks)) {
1456 if (net_ratelimit()) {
1457 /* Try to help the admin */
1458 printk(KERN_NOTICE "%s: too many open TCP "
1459 "sockets, consider increasing the "
1460 "number of nfsd threads\n",
1464 * Always select the oldest socket. It's not fair,
1467 svsk = list_entry(serv->sv_tempsocks.prev,
1470 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1471 svc_xprt_get(&svsk->sk_xprt);
1473 spin_unlock_bh(&serv->sv_lock);
1476 svc_xprt_enqueue(&svsk->sk_xprt);
1477 svc_xprt_put(&svsk->sk_xprt);
1483 * Receive the next request on any socket. This code is carefully
1484 * organised not to touch any cachelines in the shared svc_serv
1485 * structure, only cachelines in the local svc_pool.
1488 svc_recv(struct svc_rqst *rqstp, long timeout)
1490 struct svc_sock *svsk = NULL;
1491 struct svc_serv *serv = rqstp->rq_server;
1492 struct svc_pool *pool = rqstp->rq_pool;
1495 struct xdr_buf *arg;
1496 DECLARE_WAITQUEUE(wait, current);
1498 dprintk("svc: server %p waiting for data (to = %ld)\n",
1503 "svc_recv: service %p, socket not NULL!\n",
1505 if (waitqueue_active(&rqstp->rq_wait))
1507 "svc_recv: service %p, wait queue active!\n",
1511 /* now allocate needed pages. If we get a failure, sleep briefly */
1512 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
1513 for (i=0; i < pages ; i++)
1514 while (rqstp->rq_pages[i] == NULL) {
1515 struct page *p = alloc_page(GFP_KERNEL);
1517 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1518 rqstp->rq_pages[i] = p;
1520 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
1521 BUG_ON(pages >= RPCSVC_MAXPAGES);
1523 /* Make arg->head point to first page and arg->pages point to rest */
1524 arg = &rqstp->rq_arg;
1525 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
1526 arg->head[0].iov_len = PAGE_SIZE;
1527 arg->pages = rqstp->rq_pages + 1;
1529 /* save at least one page for response */
1530 arg->page_len = (pages-2)*PAGE_SIZE;
1531 arg->len = (pages-1)*PAGE_SIZE;
1532 arg->tail[0].iov_len = 0;
1539 spin_lock_bh(&pool->sp_lock);
1540 if ((svsk = svc_sock_dequeue(pool)) != NULL) {
1541 rqstp->rq_sock = svsk;
1542 svc_xprt_get(&svsk->sk_xprt);
1543 rqstp->rq_reserved = serv->sv_max_mesg;
1544 atomic_add(rqstp->rq_reserved, &svsk->sk_xprt.xpt_reserved);
1546 /* No data pending. Go to sleep */
1547 svc_thread_enqueue(pool, rqstp);
1550 * We have to be able to interrupt this wait
1551 * to bring down the daemons ...
1553 set_current_state(TASK_INTERRUPTIBLE);
1554 add_wait_queue(&rqstp->rq_wait, &wait);
1555 spin_unlock_bh(&pool->sp_lock);
1557 schedule_timeout(timeout);
1561 spin_lock_bh(&pool->sp_lock);
1562 remove_wait_queue(&rqstp->rq_wait, &wait);
1564 if (!(svsk = rqstp->rq_sock)) {
1565 svc_thread_dequeue(pool, rqstp);
1566 spin_unlock_bh(&pool->sp_lock);
1567 dprintk("svc: server %p, no data yet\n", rqstp);
1568 return signalled()? -EINTR : -EAGAIN;
1571 spin_unlock_bh(&pool->sp_lock);
1574 if (test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags)) {
1575 dprintk("svc_recv: found XPT_CLOSE\n");
1576 svc_delete_xprt(&svsk->sk_xprt);
1577 } else if (test_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags)) {
1578 struct svc_xprt *newxpt;
1579 newxpt = svsk->sk_xprt.xpt_ops->xpo_accept(&svsk->sk_xprt);
1582 * We know this module_get will succeed because the
1583 * listener holds a reference too
1585 __module_get(newxpt->xpt_class->xcl_owner);
1586 svc_check_conn_limits(svsk->sk_xprt.xpt_server);
1587 svc_xprt_received(newxpt);
1589 svc_xprt_received(&svsk->sk_xprt);
1591 dprintk("svc: server %p, pool %u, socket %p, inuse=%d\n",
1592 rqstp, pool->sp_id, svsk,
1593 atomic_read(&svsk->sk_xprt.xpt_ref.refcount));
1594 rqstp->rq_deferred = svc_deferred_dequeue(&svsk->sk_xprt);
1595 if (rqstp->rq_deferred) {
1596 svc_xprt_received(&svsk->sk_xprt);
1597 len = svc_deferred_recv(rqstp);
1599 len = svsk->sk_xprt.xpt_ops->xpo_recvfrom(rqstp);
1600 dprintk("svc: got len=%d\n", len);
1603 /* No data, incomplete (TCP) read, or accept() */
1604 if (len == 0 || len == -EAGAIN) {
1605 rqstp->rq_res.len = 0;
1606 svc_sock_release(rqstp);
1609 clear_bit(XPT_OLD, &svsk->sk_xprt.xpt_flags);
1611 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
1612 rqstp->rq_chandle.defer = svc_defer;
1615 serv->sv_stats->netcnt++;
1623 svc_drop(struct svc_rqst *rqstp)
1625 dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1626 svc_sock_release(rqstp);
1630 * Return reply to client.
1633 svc_send(struct svc_rqst *rqstp)
1635 struct svc_xprt *xprt;
1639 xprt = rqstp->rq_xprt;
1643 /* release the receive skb before sending the reply */
1644 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
1646 /* calculate over-all length */
1647 xb = & rqstp->rq_res;
1648 xb->len = xb->head[0].iov_len +
1650 xb->tail[0].iov_len;
1652 /* Grab mutex to serialize outgoing data. */
1653 mutex_lock(&xprt->xpt_mutex);
1654 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
1657 len = xprt->xpt_ops->xpo_sendto(rqstp);
1658 mutex_unlock(&xprt->xpt_mutex);
1659 svc_sock_release(rqstp);
1661 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1667 * Timer function to close old temporary sockets, using
1668 * a mark-and-sweep algorithm.
1671 svc_age_temp_sockets(unsigned long closure)
1673 struct svc_serv *serv = (struct svc_serv *)closure;
1674 struct svc_sock *svsk;
1675 struct list_head *le, *next;
1676 LIST_HEAD(to_be_aged);
1678 dprintk("svc_age_temp_sockets\n");
1680 if (!spin_trylock_bh(&serv->sv_lock)) {
1681 /* busy, try again 1 sec later */
1682 dprintk("svc_age_temp_sockets: busy\n");
1683 mod_timer(&serv->sv_temptimer, jiffies + HZ);
1687 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1688 svsk = list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1690 if (!test_and_set_bit(XPT_OLD, &svsk->sk_xprt.xpt_flags))
1692 if (atomic_read(&svsk->sk_xprt.xpt_ref.refcount) > 1
1693 || test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags))
1695 svc_xprt_get(&svsk->sk_xprt);
1696 list_move(le, &to_be_aged);
1697 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1698 set_bit(XPT_DETACHED, &svsk->sk_xprt.xpt_flags);
1700 spin_unlock_bh(&serv->sv_lock);
1702 while (!list_empty(&to_be_aged)) {
1703 le = to_be_aged.next;
1704 /* fiddling the sk_xprt.xpt_list node is safe 'cos we're XPT_DETACHED */
1706 svsk = list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1708 dprintk("queuing svsk %p for closing\n", svsk);
1710 /* a thread will dequeue and close it soon */
1711 svc_xprt_enqueue(&svsk->sk_xprt);
1712 svc_xprt_put(&svsk->sk_xprt);
1715 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
1719 * Initialize socket for RPC use and create svc_sock struct
1720 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1722 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1723 struct socket *sock,
1724 int *errp, int flags)
1726 struct svc_sock *svsk;
1728 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1729 int is_temporary = flags & SVC_SOCK_TEMPORARY;
1731 dprintk("svc: svc_setup_socket %p\n", sock);
1732 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1739 /* Register socket with portmapper */
1740 if (*errp >= 0 && pmap_register)
1741 *errp = svc_register(serv, inet->sk_protocol,
1742 ntohs(inet_sk(inet)->sport));
1749 set_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags);
1750 inet->sk_user_data = svsk;
1751 svsk->sk_sock = sock;
1753 svsk->sk_ostate = inet->sk_state_change;
1754 svsk->sk_odata = inet->sk_data_ready;
1755 svsk->sk_owspace = inet->sk_write_space;
1757 /* Initialize the socket */
1758 if (sock->type == SOCK_DGRAM)
1759 svc_udp_init(svsk, serv);
1761 svc_tcp_init(svsk, serv);
1763 spin_lock_bh(&serv->sv_lock);
1765 set_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1766 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_tempsocks);
1768 if (serv->sv_temptimer.function == NULL) {
1769 /* setup timer to age temp sockets */
1770 setup_timer(&serv->sv_temptimer, svc_age_temp_sockets,
1771 (unsigned long)serv);
1772 mod_timer(&serv->sv_temptimer,
1773 jiffies + svc_conn_age_period * HZ);
1776 clear_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1777 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_permsocks);
1779 spin_unlock_bh(&serv->sv_lock);
1781 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1787 int svc_addsock(struct svc_serv *serv,
1793 struct socket *so = sockfd_lookup(fd, &err);
1794 struct svc_sock *svsk = NULL;
1798 if (so->sk->sk_family != AF_INET)
1799 err = -EAFNOSUPPORT;
1800 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1801 so->sk->sk_protocol != IPPROTO_UDP)
1802 err = -EPROTONOSUPPORT;
1803 else if (so->state > SS_UNCONNECTED)
1806 svsk = svc_setup_socket(serv, so, &err, SVC_SOCK_DEFAULTS);
1808 svc_xprt_received(&svsk->sk_xprt);
1816 if (proto) *proto = so->sk->sk_protocol;
1817 return one_sock_name(name_return, svsk);
1819 EXPORT_SYMBOL_GPL(svc_addsock);
1822 * Create socket for RPC service.
1824 static struct svc_xprt *svc_create_socket(struct svc_serv *serv,
1826 struct sockaddr *sin, int len,
1829 struct svc_sock *svsk;
1830 struct socket *sock;
1833 char buf[RPC_MAX_ADDRBUFLEN];
1835 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1836 serv->sv_program->pg_name, protocol,
1837 __svc_print_addr(sin, buf, sizeof(buf)));
1839 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1840 printk(KERN_WARNING "svc: only UDP and TCP "
1841 "sockets supported\n");
1842 return ERR_PTR(-EINVAL);
1844 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1846 error = sock_create_kern(sin->sa_family, type, protocol, &sock);
1848 return ERR_PTR(error);
1850 svc_reclassify_socket(sock);
1852 if (type == SOCK_STREAM)
1853 sock->sk->sk_reuse = 1; /* allow address reuse */
1854 error = kernel_bind(sock, sin, len);
1858 if (protocol == IPPROTO_TCP) {
1859 if ((error = kernel_listen(sock, 64)) < 0)
1863 if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1864 svc_xprt_received(&svsk->sk_xprt);
1865 return (struct svc_xprt *)svsk;
1869 dprintk("svc: svc_create_socket error = %d\n", -error);
1871 return ERR_PTR(error);
1875 * Detach the svc_sock from the socket so that no
1876 * more callbacks occur.
1878 static void svc_sock_detach(struct svc_xprt *xprt)
1880 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1881 struct sock *sk = svsk->sk_sk;
1883 dprintk("svc: svc_sock_detach(%p)\n", svsk);
1885 /* put back the old socket callbacks */
1886 sk->sk_state_change = svsk->sk_ostate;
1887 sk->sk_data_ready = svsk->sk_odata;
1888 sk->sk_write_space = svsk->sk_owspace;
1892 * Free the svc_sock's socket resources and the svc_sock itself.
1894 static void svc_sock_free(struct svc_xprt *xprt)
1896 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1897 dprintk("svc: svc_sock_free(%p)\n", svsk);
1899 if (svsk->sk_sock->file)
1900 sockfd_put(svsk->sk_sock);
1902 sock_release(svsk->sk_sock);
1907 * Remove a dead transport
1909 static void svc_delete_xprt(struct svc_xprt *xprt)
1911 struct svc_serv *serv = xprt->xpt_server;
1913 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1914 xprt->xpt_ops->xpo_detach(xprt);
1916 spin_lock_bh(&serv->sv_lock);
1917 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
1918 list_del_init(&xprt->xpt_list);
1920 * We used to delete the transport from whichever list
1921 * it's sk_xprt.xpt_ready node was on, but we don't actually
1922 * need to. This is because the only time we're called
1923 * while still attached to a queue, the queue itself
1924 * is about to be destroyed (in svc_destroy).
1926 if (!test_and_set_bit(XPT_DEAD, &xprt->xpt_flags)) {
1927 BUG_ON(atomic_read(&xprt->xpt_ref.refcount) < 2);
1928 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1932 spin_unlock_bh(&serv->sv_lock);
1935 static void svc_close_xprt(struct svc_xprt *xprt)
1937 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1938 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1939 /* someone else will have to effect the close */
1943 svc_delete_xprt(xprt);
1944 clear_bit(XPT_BUSY, &xprt->xpt_flags);
1948 void svc_close_all(struct list_head *xprt_list)
1950 struct svc_xprt *xprt;
1951 struct svc_xprt *tmp;
1953 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
1954 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1955 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
1956 /* Waiting to be processed, but no threads left,
1957 * So just remove it from the waiting list
1959 list_del_init(&xprt->xpt_ready);
1960 clear_bit(XPT_BUSY, &xprt->xpt_flags);
1962 svc_close_xprt(xprt);
1967 * Handle defer and revisit of requests
1970 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1972 struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1973 struct svc_xprt *xprt = dr->xprt;
1980 dprintk("revisit queued\n");
1982 spin_lock(&xprt->xpt_lock);
1983 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1984 spin_unlock(&xprt->xpt_lock);
1985 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1986 svc_xprt_enqueue(xprt);
1990 static struct cache_deferred_req *
1991 svc_defer(struct cache_req *req)
1993 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1994 int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
1995 struct svc_deferred_req *dr;
1997 if (rqstp->rq_arg.page_len)
1998 return NULL; /* if more than a page, give up FIXME */
1999 if (rqstp->rq_deferred) {
2000 dr = rqstp->rq_deferred;
2001 rqstp->rq_deferred = NULL;
2003 int skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
2004 /* FIXME maybe discard if size too large */
2005 dr = kmalloc(size, GFP_KERNEL);
2009 dr->handle.owner = rqstp->rq_server;
2010 dr->prot = rqstp->rq_prot;
2011 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
2012 dr->addrlen = rqstp->rq_addrlen;
2013 dr->daddr = rqstp->rq_daddr;
2014 dr->argslen = rqstp->rq_arg.len >> 2;
2015 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
2017 svc_xprt_get(rqstp->rq_xprt);
2018 dr->xprt = rqstp->rq_xprt;
2020 dr->handle.revisit = svc_revisit;
2025 * recv data from a deferred request into an active one
2027 static int svc_deferred_recv(struct svc_rqst *rqstp)
2029 struct svc_deferred_req *dr = rqstp->rq_deferred;
2031 rqstp->rq_arg.head[0].iov_base = dr->args;
2032 rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
2033 rqstp->rq_arg.page_len = 0;
2034 rqstp->rq_arg.len = dr->argslen<<2;
2035 rqstp->rq_prot = dr->prot;
2036 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
2037 rqstp->rq_addrlen = dr->addrlen;
2038 rqstp->rq_daddr = dr->daddr;
2039 rqstp->rq_respages = rqstp->rq_pages;
2040 return dr->argslen<<2;
2044 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
2046 struct svc_deferred_req *dr = NULL;
2048 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
2050 spin_lock(&xprt->xpt_lock);
2051 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
2052 if (!list_empty(&xprt->xpt_deferred)) {
2053 dr = list_entry(xprt->xpt_deferred.next,
2054 struct svc_deferred_req,
2056 list_del_init(&dr->handle.recent);
2057 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
2059 spin_unlock(&xprt->xpt_lock);