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);
97 static void svc_age_temp_xprts(unsigned long closure);
99 /* apparently the "standard" is that clients close
100 * idle connections after 5 minutes, servers after
102 * http://www.connectathon.org/talks96/nfstcp.pdf
104 static int svc_conn_age_period = 6*60;
106 #ifdef CONFIG_DEBUG_LOCK_ALLOC
107 static struct lock_class_key svc_key[2];
108 static struct lock_class_key svc_slock_key[2];
110 static inline void svc_reclassify_socket(struct socket *sock)
112 struct sock *sk = sock->sk;
113 BUG_ON(sock_owned_by_user(sk));
114 switch (sk->sk_family) {
116 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFSD",
118 "sk_xprt.xpt_lock-AF_INET-NFSD",
123 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFSD",
125 "sk_xprt.xpt_lock-AF_INET6-NFSD",
134 static inline void svc_reclassify_socket(struct socket *sock)
139 static char *__svc_print_addr(struct sockaddr *addr, char *buf, size_t len)
141 switch (addr->sa_family) {
143 snprintf(buf, len, "%u.%u.%u.%u, port=%u",
144 NIPQUAD(((struct sockaddr_in *) addr)->sin_addr),
145 ntohs(((struct sockaddr_in *) addr)->sin_port));
149 snprintf(buf, len, "%x:%x:%x:%x:%x:%x:%x:%x, port=%u",
150 NIP6(((struct sockaddr_in6 *) addr)->sin6_addr),
151 ntohs(((struct sockaddr_in6 *) addr)->sin6_port));
155 snprintf(buf, len, "unknown address type: %d", addr->sa_family);
162 * svc_print_addr - Format rq_addr field for printing
163 * @rqstp: svc_rqst struct containing address to print
164 * @buf: target buffer for formatted address
165 * @len: length of target buffer
168 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
170 return __svc_print_addr(svc_addr(rqstp), buf, len);
172 EXPORT_SYMBOL_GPL(svc_print_addr);
175 * Queue up an idle server thread. Must have pool->sp_lock held.
176 * Note: this is really a stack rather than a queue, so that we only
177 * use as many different threads as we need, and the rest don't pollute
181 svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
183 list_add(&rqstp->rq_list, &pool->sp_threads);
187 * Dequeue an nfsd thread. Must have pool->sp_lock held.
190 svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
192 list_del(&rqstp->rq_list);
196 * Release an skbuff after use
198 static void svc_release_skb(struct svc_rqst *rqstp)
200 struct sk_buff *skb = rqstp->rq_xprt_ctxt;
201 struct svc_deferred_req *dr = rqstp->rq_deferred;
204 rqstp->rq_xprt_ctxt = NULL;
206 dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
207 skb_free_datagram(rqstp->rq_sock->sk_sk, skb);
210 rqstp->rq_deferred = NULL;
216 * Queue up a socket with data pending. If there are idle nfsd
217 * processes, wake 'em up.
220 void svc_xprt_enqueue(struct svc_xprt *xprt)
222 struct svc_serv *serv = xprt->xpt_server;
223 struct svc_pool *pool;
224 struct svc_rqst *rqstp;
227 if (!(xprt->xpt_flags &
228 ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
230 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
234 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
237 spin_lock_bh(&pool->sp_lock);
239 if (!list_empty(&pool->sp_threads) &&
240 !list_empty(&pool->sp_sockets))
243 "threads and transports both waiting??\n");
245 if (test_bit(XPT_DEAD, &xprt->xpt_flags)) {
246 /* Don't enqueue dead sockets */
247 dprintk("svc: transport %p is dead, not enqueued\n", xprt);
251 /* Mark socket as busy. It will remain in this state until the
252 * server has processed all pending data and put the socket back
253 * on the idle list. We update XPT_BUSY atomically because
254 * it also guards against trying to enqueue the svc_sock twice.
256 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
257 /* Don't enqueue socket while already enqueued */
258 dprintk("svc: transport %p busy, not enqueued\n", xprt);
261 BUG_ON(xprt->xpt_pool != NULL);
262 xprt->xpt_pool = pool;
264 /* Handle pending connection */
265 if (test_bit(XPT_CONN, &xprt->xpt_flags))
268 /* Handle close in-progress */
269 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
272 /* Check if we have space to reply to a request */
273 if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
274 /* Don't enqueue while not enough space for reply */
275 dprintk("svc: no write space, transport %p not enqueued\n",
277 xprt->xpt_pool = NULL;
278 clear_bit(XPT_BUSY, &xprt->xpt_flags);
283 if (!list_empty(&pool->sp_threads)) {
284 rqstp = list_entry(pool->sp_threads.next,
287 dprintk("svc: transport %p served by daemon %p\n",
289 svc_thread_dequeue(pool, rqstp);
292 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
293 rqstp, rqstp->rq_xprt);
294 rqstp->rq_xprt = xprt;
296 rqstp->rq_reserved = serv->sv_max_mesg;
297 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
298 BUG_ON(xprt->xpt_pool != pool);
299 wake_up(&rqstp->rq_wait);
301 dprintk("svc: transport %p put into queue\n", xprt);
302 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
303 BUG_ON(xprt->xpt_pool != pool);
307 spin_unlock_bh(&pool->sp_lock);
309 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
312 * Dequeue the first socket. Must be called with the pool->sp_lock held.
314 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
316 struct svc_xprt *xprt;
318 if (list_empty(&pool->sp_sockets))
321 xprt = list_entry(pool->sp_sockets.next,
322 struct svc_xprt, xpt_ready);
323 list_del_init(&xprt->xpt_ready);
325 dprintk("svc: transport %p dequeued, inuse=%d\n",
326 xprt, atomic_read(&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);
371 static void svc_xprt_release(struct svc_rqst *rqstp)
373 struct svc_xprt *xprt = rqstp->rq_xprt;
375 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
377 svc_free_res_pages(rqstp);
378 rqstp->rq_res.page_len = 0;
379 rqstp->rq_res.page_base = 0;
381 /* Reset response buffer and release
383 * But first, check that enough space was reserved
384 * for the reply, otherwise we have a bug!
386 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
387 printk(KERN_ERR "RPC request reserved %d but used %d\n",
391 rqstp->rq_res.head[0].iov_len = 0;
392 svc_reserve(rqstp, 0);
393 rqstp->rq_xprt = NULL;
399 * External function to wake up a server waiting for data
400 * This really only makes sense for services like lockd
401 * which have exactly one thread anyway.
404 svc_wake_up(struct svc_serv *serv)
406 struct svc_rqst *rqstp;
408 struct svc_pool *pool;
410 for (i = 0; i < serv->sv_nrpools; i++) {
411 pool = &serv->sv_pools[i];
413 spin_lock_bh(&pool->sp_lock);
414 if (!list_empty(&pool->sp_threads)) {
415 rqstp = list_entry(pool->sp_threads.next,
418 dprintk("svc: daemon %p woken up.\n", rqstp);
420 svc_thread_dequeue(pool, rqstp);
421 rqstp->rq_sock = NULL;
423 wake_up(&rqstp->rq_wait);
425 spin_unlock_bh(&pool->sp_lock);
429 union svc_pktinfo_u {
430 struct in_pktinfo pkti;
431 struct in6_pktinfo pkti6;
433 #define SVC_PKTINFO_SPACE \
434 CMSG_SPACE(sizeof(union svc_pktinfo_u))
436 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
438 switch (rqstp->rq_sock->sk_sk->sk_family) {
440 struct in_pktinfo *pki = CMSG_DATA(cmh);
442 cmh->cmsg_level = SOL_IP;
443 cmh->cmsg_type = IP_PKTINFO;
444 pki->ipi_ifindex = 0;
445 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr.addr.s_addr;
446 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
451 struct in6_pktinfo *pki = CMSG_DATA(cmh);
453 cmh->cmsg_level = SOL_IPV6;
454 cmh->cmsg_type = IPV6_PKTINFO;
455 pki->ipi6_ifindex = 0;
456 ipv6_addr_copy(&pki->ipi6_addr,
457 &rqstp->rq_daddr.addr6);
458 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
466 * Generic sendto routine
469 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
471 struct svc_sock *svsk = rqstp->rq_sock;
472 struct socket *sock = svsk->sk_sock;
476 long all[SVC_PKTINFO_SPACE / sizeof(long)];
478 struct cmsghdr *cmh = &buffer.hdr;
482 struct page **ppage = xdr->pages;
483 size_t base = xdr->page_base;
484 unsigned int pglen = xdr->page_len;
485 unsigned int flags = MSG_MORE;
486 char buf[RPC_MAX_ADDRBUFLEN];
490 if (rqstp->rq_prot == IPPROTO_UDP) {
491 struct msghdr msg = {
492 .msg_name = &rqstp->rq_addr,
493 .msg_namelen = rqstp->rq_addrlen,
495 .msg_controllen = sizeof(buffer),
496 .msg_flags = MSG_MORE,
499 svc_set_cmsg_data(rqstp, cmh);
501 if (sock_sendmsg(sock, &msg, 0) < 0)
506 if (slen == xdr->head[0].iov_len)
508 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0,
509 xdr->head[0].iov_len, flags);
510 if (len != xdr->head[0].iov_len)
512 slen -= xdr->head[0].iov_len;
517 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
521 result = kernel_sendpage(sock, *ppage, base, size, flags);
528 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
533 if (xdr->tail[0].iov_len) {
534 result = kernel_sendpage(sock, rqstp->rq_respages[0],
535 ((unsigned long)xdr->tail[0].iov_base)
537 xdr->tail[0].iov_len, 0);
543 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
544 rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len,
545 xdr->len, len, svc_print_addr(rqstp, buf, sizeof(buf)));
551 * Report socket names for nfsdfs
553 static int one_sock_name(char *buf, struct svc_sock *svsk)
557 switch(svsk->sk_sk->sk_family) {
559 len = sprintf(buf, "ipv4 %s %u.%u.%u.%u %d\n",
560 svsk->sk_sk->sk_protocol==IPPROTO_UDP?
562 NIPQUAD(inet_sk(svsk->sk_sk)->rcv_saddr),
563 inet_sk(svsk->sk_sk)->num);
566 len = sprintf(buf, "*unknown-%d*\n",
567 svsk->sk_sk->sk_family);
573 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
575 struct svc_sock *svsk, *closesk = NULL;
580 spin_lock_bh(&serv->sv_lock);
581 list_for_each_entry(svsk, &serv->sv_permsocks, sk_xprt.xpt_list) {
582 int onelen = one_sock_name(buf+len, svsk);
583 if (toclose && strcmp(toclose, buf+len) == 0)
588 spin_unlock_bh(&serv->sv_lock);
590 /* Should unregister with portmap, but you cannot
591 * unregister just one protocol...
593 svc_close_xprt(&closesk->sk_xprt);
598 EXPORT_SYMBOL(svc_sock_names);
601 * Check input queue length
604 svc_recv_available(struct svc_sock *svsk)
606 struct socket *sock = svsk->sk_sock;
609 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
611 return (err >= 0)? avail : err;
615 * Generic recvfrom routine.
618 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
620 struct svc_sock *svsk = rqstp->rq_sock;
621 struct msghdr msg = {
622 .msg_flags = MSG_DONTWAIT,
626 len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
629 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
630 svsk, iov[0].iov_base, iov[0].iov_len, len);
635 * Set socket snd and rcv buffer lengths
638 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
642 oldfs = get_fs(); set_fs(KERNEL_DS);
643 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
644 (char*)&snd, sizeof(snd));
645 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
646 (char*)&rcv, sizeof(rcv));
648 /* sock_setsockopt limits use to sysctl_?mem_max,
649 * which isn't acceptable. Until that is made conditional
650 * on not having CAP_SYS_RESOURCE or similar, we go direct...
651 * DaveM said I could!
654 sock->sk->sk_sndbuf = snd * 2;
655 sock->sk->sk_rcvbuf = rcv * 2;
656 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
657 release_sock(sock->sk);
661 * INET callback when data has been received on the socket.
664 svc_udp_data_ready(struct sock *sk, int count)
666 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
669 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
671 test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
672 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
673 svc_xprt_enqueue(&svsk->sk_xprt);
675 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
676 wake_up_interruptible(sk->sk_sleep);
680 * INET callback when space is newly available on the socket.
683 svc_write_space(struct sock *sk)
685 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
688 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
689 svsk, sk, test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
690 svc_xprt_enqueue(&svsk->sk_xprt);
693 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
694 dprintk("RPC svc_write_space: someone sleeping on %p\n",
696 wake_up_interruptible(sk->sk_sleep);
701 * Copy the UDP datagram's destination address to the rqstp structure.
702 * The 'destination' address in this case is the address to which the
703 * peer sent the datagram, i.e. our local address. For multihomed
704 * hosts, this can change from msg to msg. Note that only the IP
705 * address changes, the port number should remain the same.
707 static void svc_udp_get_dest_address(struct svc_rqst *rqstp,
710 switch (rqstp->rq_sock->sk_sk->sk_family) {
712 struct in_pktinfo *pki = CMSG_DATA(cmh);
713 rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
717 struct in6_pktinfo *pki = CMSG_DATA(cmh);
718 ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
725 * Receive a datagram from a UDP socket.
728 svc_udp_recvfrom(struct svc_rqst *rqstp)
730 struct svc_sock *svsk = rqstp->rq_sock;
731 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
735 long all[SVC_PKTINFO_SPACE / sizeof(long)];
737 struct cmsghdr *cmh = &buffer.hdr;
739 struct msghdr msg = {
740 .msg_name = svc_addr(rqstp),
742 .msg_controllen = sizeof(buffer),
743 .msg_flags = MSG_DONTWAIT,
746 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
747 /* udp sockets need large rcvbuf as all pending
748 * requests are still in that buffer. sndbuf must
749 * also be large enough that there is enough space
750 * for one reply per thread. We count all threads
751 * rather than threads in a particular pool, which
752 * provides an upper bound on the number of threads
753 * which will access the socket.
755 svc_sock_setbufsize(svsk->sk_sock,
756 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
757 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
759 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
761 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
762 0, 0, MSG_PEEK | MSG_DONTWAIT);
764 skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err);
767 if (err != -EAGAIN) {
768 /* possibly an icmp error */
769 dprintk("svc: recvfrom returned error %d\n", -err);
770 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
772 svc_xprt_received(&svsk->sk_xprt);
775 len = svc_addr_len(svc_addr(rqstp));
778 rqstp->rq_addrlen = len;
779 if (skb->tstamp.tv64 == 0) {
780 skb->tstamp = ktime_get_real();
781 /* Don't enable netstamp, sunrpc doesn't
782 need that much accuracy */
784 svsk->sk_sk->sk_stamp = skb->tstamp;
785 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* there may be more data... */
788 * Maybe more packets - kick another thread ASAP.
790 svc_xprt_received(&svsk->sk_xprt);
792 len = skb->len - sizeof(struct udphdr);
793 rqstp->rq_arg.len = len;
795 rqstp->rq_prot = IPPROTO_UDP;
797 if (cmh->cmsg_level != IPPROTO_IP ||
798 cmh->cmsg_type != IP_PKTINFO) {
800 printk("rpcsvc: received unknown control message:"
802 cmh->cmsg_level, cmh->cmsg_type);
803 skb_free_datagram(svsk->sk_sk, skb);
806 svc_udp_get_dest_address(rqstp, cmh);
808 if (skb_is_nonlinear(skb)) {
809 /* we have to copy */
811 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
814 skb_free_datagram(svsk->sk_sk, skb);
818 skb_free_datagram(svsk->sk_sk, skb);
820 /* we can use it in-place */
821 rqstp->rq_arg.head[0].iov_base = skb->data + sizeof(struct udphdr);
822 rqstp->rq_arg.head[0].iov_len = len;
823 if (skb_checksum_complete(skb)) {
824 skb_free_datagram(svsk->sk_sk, skb);
827 rqstp->rq_xprt_ctxt = skb;
830 rqstp->rq_arg.page_base = 0;
831 if (len <= rqstp->rq_arg.head[0].iov_len) {
832 rqstp->rq_arg.head[0].iov_len = len;
833 rqstp->rq_arg.page_len = 0;
834 rqstp->rq_respages = rqstp->rq_pages+1;
836 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
837 rqstp->rq_respages = rqstp->rq_pages + 1 +
838 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
842 serv->sv_stats->netudpcnt++;
848 svc_udp_sendto(struct svc_rqst *rqstp)
852 error = svc_sendto(rqstp, &rqstp->rq_res);
853 if (error == -ECONNREFUSED)
854 /* ICMP error on earlier request. */
855 error = svc_sendto(rqstp, &rqstp->rq_res);
860 static void svc_udp_prep_reply_hdr(struct svc_rqst *rqstp)
864 static int svc_udp_has_wspace(struct svc_xprt *xprt)
866 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
867 struct svc_serv *serv = xprt->xpt_server;
868 unsigned long required;
871 * Set the SOCK_NOSPACE flag before checking the available
874 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
875 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
876 if (required*2 > sock_wspace(svsk->sk_sk))
878 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
882 static struct svc_xprt *svc_udp_accept(struct svc_xprt *xprt)
888 static struct svc_xprt *svc_udp_create(struct svc_serv *serv,
889 struct sockaddr *sa, int salen,
892 return svc_create_socket(serv, IPPROTO_UDP, sa, salen, flags);
895 static struct svc_xprt_ops svc_udp_ops = {
896 .xpo_create = svc_udp_create,
897 .xpo_recvfrom = svc_udp_recvfrom,
898 .xpo_sendto = svc_udp_sendto,
899 .xpo_release_rqst = svc_release_skb,
900 .xpo_detach = svc_sock_detach,
901 .xpo_free = svc_sock_free,
902 .xpo_prep_reply_hdr = svc_udp_prep_reply_hdr,
903 .xpo_has_wspace = svc_udp_has_wspace,
904 .xpo_accept = svc_udp_accept,
907 static struct svc_xprt_class svc_udp_class = {
909 .xcl_owner = THIS_MODULE,
910 .xcl_ops = &svc_udp_ops,
911 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
914 static void svc_udp_init(struct svc_sock *svsk, struct svc_serv *serv)
919 svc_xprt_init(&svc_udp_class, &svsk->sk_xprt, serv);
920 clear_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
921 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
922 svsk->sk_sk->sk_write_space = svc_write_space;
924 /* initialise setting must have enough space to
925 * receive and respond to one request.
926 * svc_udp_recvfrom will re-adjust if necessary
928 svc_sock_setbufsize(svsk->sk_sock,
929 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
930 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
932 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* might have come in before data_ready set up */
933 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
937 /* make sure we get destination address info */
938 svsk->sk_sock->ops->setsockopt(svsk->sk_sock, IPPROTO_IP, IP_PKTINFO,
939 (char __user *)&one, sizeof(one));
944 * A data_ready event on a listening socket means there's a connection
945 * pending. Do not use state_change as a substitute for it.
948 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
950 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
952 dprintk("svc: socket %p TCP (listen) state change %d\n",
956 * This callback may called twice when a new connection
957 * is established as a child socket inherits everything
958 * from a parent LISTEN socket.
959 * 1) data_ready method of the parent socket will be called
960 * when one of child sockets become ESTABLISHED.
961 * 2) data_ready method of the child socket may be called
962 * when it receives data before the socket is accepted.
963 * In case of 2, we should ignore it silently.
965 if (sk->sk_state == TCP_LISTEN) {
967 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
968 svc_xprt_enqueue(&svsk->sk_xprt);
970 printk("svc: socket %p: no user data\n", sk);
973 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
974 wake_up_interruptible_all(sk->sk_sleep);
978 * A state change on a connected socket means it's dying or dead.
981 svc_tcp_state_change(struct sock *sk)
983 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
985 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
986 sk, sk->sk_state, sk->sk_user_data);
989 printk("svc: socket %p: no user data\n", sk);
991 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
992 svc_xprt_enqueue(&svsk->sk_xprt);
994 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
995 wake_up_interruptible_all(sk->sk_sleep);
999 svc_tcp_data_ready(struct sock *sk, int count)
1001 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
1003 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
1004 sk, sk->sk_user_data);
1006 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1007 svc_xprt_enqueue(&svsk->sk_xprt);
1009 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1010 wake_up_interruptible(sk->sk_sleep);
1013 static inline int svc_port_is_privileged(struct sockaddr *sin)
1015 switch (sin->sa_family) {
1017 return ntohs(((struct sockaddr_in *)sin)->sin_port)
1020 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
1028 * Accept a TCP connection
1030 static struct svc_xprt *svc_tcp_accept(struct svc_xprt *xprt)
1032 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1033 struct sockaddr_storage addr;
1034 struct sockaddr *sin = (struct sockaddr *) &addr;
1035 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1036 struct socket *sock = svsk->sk_sock;
1037 struct socket *newsock;
1038 struct svc_sock *newsvsk;
1040 char buf[RPC_MAX_ADDRBUFLEN];
1042 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
1046 clear_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1047 err = kernel_accept(sock, &newsock, O_NONBLOCK);
1050 printk(KERN_WARNING "%s: no more sockets!\n",
1052 else if (err != -EAGAIN && net_ratelimit())
1053 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
1054 serv->sv_name, -err);
1057 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1059 err = kernel_getpeername(newsock, sin, &slen);
1061 if (net_ratelimit())
1062 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
1063 serv->sv_name, -err);
1064 goto failed; /* aborted connection or whatever */
1067 /* Ideally, we would want to reject connections from unauthorized
1068 * hosts here, but when we get encryption, the IP of the host won't
1069 * tell us anything. For now just warn about unpriv connections.
1071 if (!svc_port_is_privileged(sin)) {
1072 dprintk(KERN_WARNING
1073 "%s: connect from unprivileged port: %s\n",
1075 __svc_print_addr(sin, buf, sizeof(buf)));
1077 dprintk("%s: connect from %s\n", serv->sv_name,
1078 __svc_print_addr(sin, buf, sizeof(buf)));
1080 /* make sure that a write doesn't block forever when
1083 newsock->sk->sk_sndtimeo = HZ*30;
1085 if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
1086 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
1088 svc_xprt_set_remote(&newsvsk->sk_xprt, sin, slen);
1089 err = kernel_getsockname(newsock, sin, &slen);
1090 if (unlikely(err < 0)) {
1091 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err);
1092 slen = offsetof(struct sockaddr, sa_data);
1094 svc_xprt_set_local(&newsvsk->sk_xprt, sin, slen);
1097 serv->sv_stats->nettcpconn++;
1099 return &newsvsk->sk_xprt;
1102 sock_release(newsock);
1107 * Receive data from a TCP socket.
1110 svc_tcp_recvfrom(struct svc_rqst *rqstp)
1112 struct svc_sock *svsk = rqstp->rq_sock;
1113 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1118 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
1119 svsk, test_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags),
1120 test_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags),
1121 test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags));
1123 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
1124 /* sndbuf needs to have room for one request
1125 * per thread, otherwise we can stall even when the
1126 * network isn't a bottleneck.
1128 * We count all threads rather than threads in a
1129 * particular pool, which provides an upper bound
1130 * on the number of threads which will access the socket.
1132 * rcvbuf just needs to be able to hold a few requests.
1133 * Normally they will be removed from the queue
1134 * as soon a a complete request arrives.
1136 svc_sock_setbufsize(svsk->sk_sock,
1137 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
1138 3 * serv->sv_max_mesg);
1140 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1142 /* Receive data. If we haven't got the record length yet, get
1143 * the next four bytes. Otherwise try to gobble up as much as
1144 * possible up to the complete record length.
1146 if (svsk->sk_tcplen < 4) {
1147 unsigned long want = 4 - svsk->sk_tcplen;
1150 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
1152 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
1154 svsk->sk_tcplen += len;
1157 dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
1159 svc_xprt_received(&svsk->sk_xprt);
1160 return -EAGAIN; /* record header not complete */
1163 svsk->sk_reclen = ntohl(svsk->sk_reclen);
1164 if (!(svsk->sk_reclen & 0x80000000)) {
1165 /* FIXME: technically, a record can be fragmented,
1166 * and non-terminal fragments will not have the top
1167 * bit set in the fragment length header.
1168 * But apparently no known nfs clients send fragmented
1170 if (net_ratelimit())
1171 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1172 " (non-terminal)\n",
1173 (unsigned long) svsk->sk_reclen);
1176 svsk->sk_reclen &= 0x7fffffff;
1177 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
1178 if (svsk->sk_reclen > serv->sv_max_mesg) {
1179 if (net_ratelimit())
1180 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1182 (unsigned long) svsk->sk_reclen);
1187 /* Check whether enough data is available */
1188 len = svc_recv_available(svsk);
1192 if (len < svsk->sk_reclen) {
1193 dprintk("svc: incomplete TCP record (%d of %d)\n",
1194 len, svsk->sk_reclen);
1195 svc_xprt_received(&svsk->sk_xprt);
1196 return -EAGAIN; /* record not complete */
1198 len = svsk->sk_reclen;
1199 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1201 vec = rqstp->rq_vec;
1202 vec[0] = rqstp->rq_arg.head[0];
1205 while (vlen < len) {
1206 vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
1207 vec[pnum].iov_len = PAGE_SIZE;
1211 rqstp->rq_respages = &rqstp->rq_pages[pnum];
1213 /* Now receive data */
1214 len = svc_recvfrom(rqstp, vec, pnum, len);
1218 dprintk("svc: TCP complete record (%d bytes)\n", len);
1219 rqstp->rq_arg.len = len;
1220 rqstp->rq_arg.page_base = 0;
1221 if (len <= rqstp->rq_arg.head[0].iov_len) {
1222 rqstp->rq_arg.head[0].iov_len = len;
1223 rqstp->rq_arg.page_len = 0;
1225 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1228 rqstp->rq_xprt_ctxt = NULL;
1229 rqstp->rq_prot = IPPROTO_TCP;
1231 /* Reset TCP read info */
1232 svsk->sk_reclen = 0;
1233 svsk->sk_tcplen = 0;
1235 svc_xprt_copy_addrs(rqstp, &svsk->sk_xprt);
1236 svc_xprt_received(&svsk->sk_xprt);
1238 serv->sv_stats->nettcpcnt++;
1243 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1247 if (len == -EAGAIN) {
1248 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1249 svc_xprt_received(&svsk->sk_xprt);
1251 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1252 svsk->sk_xprt.xpt_server->sv_name, -len);
1260 * Send out data on TCP socket.
1263 svc_tcp_sendto(struct svc_rqst *rqstp)
1265 struct xdr_buf *xbufp = &rqstp->rq_res;
1269 /* Set up the first element of the reply kvec.
1270 * Any other kvecs that may be in use have been taken
1271 * care of by the server implementation itself.
1273 reclen = htonl(0x80000000|((xbufp->len ) - 4));
1274 memcpy(xbufp->head[0].iov_base, &reclen, 4);
1276 if (test_bit(XPT_DEAD, &rqstp->rq_sock->sk_xprt.xpt_flags))
1279 sent = svc_sendto(rqstp, &rqstp->rq_res);
1280 if (sent != xbufp->len) {
1281 printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1282 rqstp->rq_sock->sk_xprt.xpt_server->sv_name,
1283 (sent<0)?"got error":"sent only",
1285 set_bit(XPT_CLOSE, &rqstp->rq_sock->sk_xprt.xpt_flags);
1286 svc_xprt_enqueue(rqstp->rq_xprt);
1293 * Setup response header. TCP has a 4B record length field.
1295 static void svc_tcp_prep_reply_hdr(struct svc_rqst *rqstp)
1297 struct kvec *resv = &rqstp->rq_res.head[0];
1299 /* tcp needs a space for the record length... */
1303 static int svc_tcp_has_wspace(struct svc_xprt *xprt)
1305 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1306 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1311 * Set the SOCK_NOSPACE flag before checking the available
1314 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1315 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
1316 wspace = sk_stream_wspace(svsk->sk_sk);
1318 if (wspace < sk_stream_min_wspace(svsk->sk_sk))
1320 if (required * 2 > wspace)
1323 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1327 static struct svc_xprt *svc_tcp_create(struct svc_serv *serv,
1328 struct sockaddr *sa, int salen,
1331 return svc_create_socket(serv, IPPROTO_TCP, sa, salen, flags);
1334 static struct svc_xprt_ops svc_tcp_ops = {
1335 .xpo_create = svc_tcp_create,
1336 .xpo_recvfrom = svc_tcp_recvfrom,
1337 .xpo_sendto = svc_tcp_sendto,
1338 .xpo_release_rqst = svc_release_skb,
1339 .xpo_detach = svc_sock_detach,
1340 .xpo_free = svc_sock_free,
1341 .xpo_prep_reply_hdr = svc_tcp_prep_reply_hdr,
1342 .xpo_has_wspace = svc_tcp_has_wspace,
1343 .xpo_accept = svc_tcp_accept,
1346 static struct svc_xprt_class svc_tcp_class = {
1348 .xcl_owner = THIS_MODULE,
1349 .xcl_ops = &svc_tcp_ops,
1350 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1353 void svc_init_xprt_sock(void)
1355 svc_reg_xprt_class(&svc_tcp_class);
1356 svc_reg_xprt_class(&svc_udp_class);
1359 void svc_cleanup_xprt_sock(void)
1361 svc_unreg_xprt_class(&svc_tcp_class);
1362 svc_unreg_xprt_class(&svc_udp_class);
1365 static void svc_tcp_init(struct svc_sock *svsk, struct svc_serv *serv)
1367 struct sock *sk = svsk->sk_sk;
1368 struct tcp_sock *tp = tcp_sk(sk);
1370 svc_xprt_init(&svc_tcp_class, &svsk->sk_xprt, serv);
1371 set_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
1372 if (sk->sk_state == TCP_LISTEN) {
1373 dprintk("setting up TCP socket for listening\n");
1374 set_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags);
1375 sk->sk_data_ready = svc_tcp_listen_data_ready;
1376 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1378 dprintk("setting up TCP socket for reading\n");
1379 sk->sk_state_change = svc_tcp_state_change;
1380 sk->sk_data_ready = svc_tcp_data_ready;
1381 sk->sk_write_space = svc_write_space;
1383 svsk->sk_reclen = 0;
1384 svsk->sk_tcplen = 0;
1386 tp->nonagle = 1; /* disable Nagle's algorithm */
1388 /* initialise setting must have enough space to
1389 * receive and respond to one request.
1390 * svc_tcp_recvfrom will re-adjust if necessary
1392 svc_sock_setbufsize(svsk->sk_sock,
1393 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
1394 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
1396 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1397 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1398 if (sk->sk_state != TCP_ESTABLISHED)
1399 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1404 svc_sock_update_bufs(struct svc_serv *serv)
1407 * The number of server threads has changed. Update
1408 * rcvbuf and sndbuf accordingly on all sockets
1410 struct list_head *le;
1412 spin_lock_bh(&serv->sv_lock);
1413 list_for_each(le, &serv->sv_permsocks) {
1414 struct svc_sock *svsk =
1415 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1416 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1418 list_for_each(le, &serv->sv_tempsocks) {
1419 struct svc_sock *svsk =
1420 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1421 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1423 spin_unlock_bh(&serv->sv_lock);
1427 * Make sure that we don't have too many active connections. If we
1428 * have, something must be dropped.
1430 * There's no point in trying to do random drop here for DoS
1431 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
1432 * attacker can easily beat that.
1434 * The only somewhat efficient mechanism would be if drop old
1435 * connections from the same IP first. But right now we don't even
1436 * record the client IP in svc_sock.
1438 static void svc_check_conn_limits(struct svc_serv *serv)
1440 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
1441 struct svc_sock *svsk = NULL;
1442 spin_lock_bh(&serv->sv_lock);
1443 if (!list_empty(&serv->sv_tempsocks)) {
1444 if (net_ratelimit()) {
1445 /* Try to help the admin */
1446 printk(KERN_NOTICE "%s: too many open TCP "
1447 "sockets, consider increasing the "
1448 "number of nfsd threads\n",
1452 * Always select the oldest socket. It's not fair,
1455 svsk = list_entry(serv->sv_tempsocks.prev,
1458 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1459 svc_xprt_get(&svsk->sk_xprt);
1461 spin_unlock_bh(&serv->sv_lock);
1464 svc_xprt_enqueue(&svsk->sk_xprt);
1465 svc_xprt_put(&svsk->sk_xprt);
1471 * Receive the next request on any socket. This code is carefully
1472 * organised not to touch any cachelines in the shared svc_serv
1473 * structure, only cachelines in the local svc_pool.
1476 svc_recv(struct svc_rqst *rqstp, long timeout)
1478 struct svc_xprt *xprt = NULL;
1479 struct svc_serv *serv = rqstp->rq_server;
1480 struct svc_pool *pool = rqstp->rq_pool;
1483 struct xdr_buf *arg;
1484 DECLARE_WAITQUEUE(wait, current);
1486 dprintk("svc: server %p waiting for data (to = %ld)\n",
1491 "svc_recv: service %p, transport not NULL!\n",
1493 if (waitqueue_active(&rqstp->rq_wait))
1495 "svc_recv: service %p, wait queue active!\n",
1499 /* now allocate needed pages. If we get a failure, sleep briefly */
1500 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
1501 for (i=0; i < pages ; i++)
1502 while (rqstp->rq_pages[i] == NULL) {
1503 struct page *p = alloc_page(GFP_KERNEL);
1505 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1506 rqstp->rq_pages[i] = p;
1508 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
1509 BUG_ON(pages >= RPCSVC_MAXPAGES);
1511 /* Make arg->head point to first page and arg->pages point to rest */
1512 arg = &rqstp->rq_arg;
1513 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
1514 arg->head[0].iov_len = PAGE_SIZE;
1515 arg->pages = rqstp->rq_pages + 1;
1517 /* save at least one page for response */
1518 arg->page_len = (pages-2)*PAGE_SIZE;
1519 arg->len = (pages-1)*PAGE_SIZE;
1520 arg->tail[0].iov_len = 0;
1527 spin_lock_bh(&pool->sp_lock);
1528 xprt = svc_xprt_dequeue(pool);
1530 rqstp->rq_xprt = xprt;
1532 rqstp->rq_reserved = serv->sv_max_mesg;
1533 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
1535 /* No data pending. Go to sleep */
1536 svc_thread_enqueue(pool, rqstp);
1539 * We have to be able to interrupt this wait
1540 * to bring down the daemons ...
1542 set_current_state(TASK_INTERRUPTIBLE);
1543 add_wait_queue(&rqstp->rq_wait, &wait);
1544 spin_unlock_bh(&pool->sp_lock);
1546 schedule_timeout(timeout);
1550 spin_lock_bh(&pool->sp_lock);
1551 remove_wait_queue(&rqstp->rq_wait, &wait);
1553 xprt = rqstp->rq_xprt;
1555 svc_thread_dequeue(pool, rqstp);
1556 spin_unlock_bh(&pool->sp_lock);
1557 dprintk("svc: server %p, no data yet\n", rqstp);
1558 return signalled()? -EINTR : -EAGAIN;
1561 spin_unlock_bh(&pool->sp_lock);
1564 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
1565 dprintk("svc_recv: found XPT_CLOSE\n");
1566 svc_delete_xprt(xprt);
1567 } else if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
1568 struct svc_xprt *newxpt;
1569 newxpt = xprt->xpt_ops->xpo_accept(xprt);
1572 * We know this module_get will succeed because the
1573 * listener holds a reference too
1575 __module_get(newxpt->xpt_class->xcl_owner);
1576 svc_check_conn_limits(xprt->xpt_server);
1577 spin_lock_bh(&serv->sv_lock);
1578 set_bit(XPT_TEMP, &newxpt->xpt_flags);
1579 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
1581 if (serv->sv_temptimer.function == NULL) {
1582 /* setup timer to age temp sockets */
1583 setup_timer(&serv->sv_temptimer,
1585 (unsigned long)serv);
1586 mod_timer(&serv->sv_temptimer,
1587 jiffies + svc_conn_age_period * HZ);
1589 spin_unlock_bh(&serv->sv_lock);
1590 svc_xprt_received(newxpt);
1592 svc_xprt_received(xprt);
1594 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
1595 rqstp, pool->sp_id, xprt,
1596 atomic_read(&xprt->xpt_ref.refcount));
1597 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
1598 if (rqstp->rq_deferred) {
1599 svc_xprt_received(xprt);
1600 len = svc_deferred_recv(rqstp);
1602 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
1603 dprintk("svc: got len=%d\n", len);
1606 /* No data, incomplete (TCP) read, or accept() */
1607 if (len == 0 || len == -EAGAIN) {
1608 rqstp->rq_res.len = 0;
1609 svc_xprt_release(rqstp);
1612 clear_bit(XPT_OLD, &xprt->xpt_flags);
1614 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
1615 rqstp->rq_chandle.defer = svc_defer;
1618 serv->sv_stats->netcnt++;
1626 svc_drop(struct svc_rqst *rqstp)
1628 dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1629 svc_xprt_release(rqstp);
1633 * Return reply to client.
1636 svc_send(struct svc_rqst *rqstp)
1638 struct svc_xprt *xprt;
1642 xprt = rqstp->rq_xprt;
1646 /* release the receive skb before sending the reply */
1647 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
1649 /* calculate over-all length */
1650 xb = & rqstp->rq_res;
1651 xb->len = xb->head[0].iov_len +
1653 xb->tail[0].iov_len;
1655 /* Grab mutex to serialize outgoing data. */
1656 mutex_lock(&xprt->xpt_mutex);
1657 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
1660 len = xprt->xpt_ops->xpo_sendto(rqstp);
1661 mutex_unlock(&xprt->xpt_mutex);
1662 svc_xprt_release(rqstp);
1664 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1670 * Timer function to close old temporary sockets, using
1671 * a mark-and-sweep algorithm.
1673 static void svc_age_temp_xprts(unsigned long closure)
1675 struct svc_serv *serv = (struct svc_serv *)closure;
1676 struct svc_xprt *xprt;
1677 struct list_head *le, *next;
1678 LIST_HEAD(to_be_aged);
1680 dprintk("svc_age_temp_xprts\n");
1682 if (!spin_trylock_bh(&serv->sv_lock)) {
1683 /* busy, try again 1 sec later */
1684 dprintk("svc_age_temp_xprts: busy\n");
1685 mod_timer(&serv->sv_temptimer, jiffies + HZ);
1689 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1690 xprt = list_entry(le, struct svc_xprt, xpt_list);
1692 /* First time through, just mark it OLD. Second time
1693 * through, close it. */
1694 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
1696 if (atomic_read(&xprt->xpt_ref.refcount) > 1
1697 || test_bit(XPT_BUSY, &xprt->xpt_flags))
1700 list_move(le, &to_be_aged);
1701 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1702 set_bit(XPT_DETACHED, &xprt->xpt_flags);
1704 spin_unlock_bh(&serv->sv_lock);
1706 while (!list_empty(&to_be_aged)) {
1707 le = to_be_aged.next;
1708 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
1710 xprt = list_entry(le, struct svc_xprt, xpt_list);
1712 dprintk("queuing xprt %p for closing\n", xprt);
1714 /* a thread will dequeue and close it soon */
1715 svc_xprt_enqueue(xprt);
1719 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
1723 * Initialize socket for RPC use and create svc_sock struct
1724 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1726 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1727 struct socket *sock,
1728 int *errp, int flags)
1730 struct svc_sock *svsk;
1732 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1734 dprintk("svc: svc_setup_socket %p\n", sock);
1735 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1742 /* Register socket with portmapper */
1743 if (*errp >= 0 && pmap_register)
1744 *errp = svc_register(serv, inet->sk_protocol,
1745 ntohs(inet_sk(inet)->sport));
1752 inet->sk_user_data = svsk;
1753 svsk->sk_sock = sock;
1755 svsk->sk_ostate = inet->sk_state_change;
1756 svsk->sk_odata = inet->sk_data_ready;
1757 svsk->sk_owspace = inet->sk_write_space;
1759 /* Initialize the socket */
1760 if (sock->type == SOCK_DGRAM)
1761 svc_udp_init(svsk, serv);
1763 svc_tcp_init(svsk, serv);
1765 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1771 int svc_addsock(struct svc_serv *serv,
1777 struct socket *so = sockfd_lookup(fd, &err);
1778 struct svc_sock *svsk = NULL;
1782 if (so->sk->sk_family != AF_INET)
1783 err = -EAFNOSUPPORT;
1784 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1785 so->sk->sk_protocol != IPPROTO_UDP)
1786 err = -EPROTONOSUPPORT;
1787 else if (so->state > SS_UNCONNECTED)
1790 svsk = svc_setup_socket(serv, so, &err, SVC_SOCK_DEFAULTS);
1792 struct sockaddr_storage addr;
1793 struct sockaddr *sin = (struct sockaddr *)&addr;
1795 if (kernel_getsockname(svsk->sk_sock, sin, &salen) == 0)
1796 svc_xprt_set_local(&svsk->sk_xprt, sin, salen);
1797 clear_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1798 spin_lock_bh(&serv->sv_lock);
1799 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_permsocks);
1800 spin_unlock_bh(&serv->sv_lock);
1801 svc_xprt_received(&svsk->sk_xprt);
1809 if (proto) *proto = so->sk->sk_protocol;
1810 return one_sock_name(name_return, svsk);
1812 EXPORT_SYMBOL_GPL(svc_addsock);
1815 * Create socket for RPC service.
1817 static struct svc_xprt *svc_create_socket(struct svc_serv *serv,
1819 struct sockaddr *sin, int len,
1822 struct svc_sock *svsk;
1823 struct socket *sock;
1826 char buf[RPC_MAX_ADDRBUFLEN];
1827 struct sockaddr_storage addr;
1828 struct sockaddr *newsin = (struct sockaddr *)&addr;
1831 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1832 serv->sv_program->pg_name, protocol,
1833 __svc_print_addr(sin, buf, sizeof(buf)));
1835 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1836 printk(KERN_WARNING "svc: only UDP and TCP "
1837 "sockets supported\n");
1838 return ERR_PTR(-EINVAL);
1840 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1842 error = sock_create_kern(sin->sa_family, type, protocol, &sock);
1844 return ERR_PTR(error);
1846 svc_reclassify_socket(sock);
1848 if (type == SOCK_STREAM)
1849 sock->sk->sk_reuse = 1; /* allow address reuse */
1850 error = kernel_bind(sock, sin, len);
1855 error = kernel_getsockname(sock, newsin, &newlen);
1859 if (protocol == IPPROTO_TCP) {
1860 if ((error = kernel_listen(sock, 64)) < 0)
1864 if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1865 svc_xprt_set_local(&svsk->sk_xprt, newsin, newlen);
1866 return (struct svc_xprt *)svsk;
1870 dprintk("svc: svc_create_socket error = %d\n", -error);
1872 return ERR_PTR(error);
1876 * Detach the svc_sock from the socket so that no
1877 * more callbacks occur.
1879 static void svc_sock_detach(struct svc_xprt *xprt)
1881 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1882 struct sock *sk = svsk->sk_sk;
1884 dprintk("svc: svc_sock_detach(%p)\n", svsk);
1886 /* put back the old socket callbacks */
1887 sk->sk_state_change = svsk->sk_ostate;
1888 sk->sk_data_ready = svsk->sk_odata;
1889 sk->sk_write_space = svsk->sk_owspace;
1893 * Free the svc_sock's socket resources and the svc_sock itself.
1895 static void svc_sock_free(struct svc_xprt *xprt)
1897 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1898 dprintk("svc: svc_sock_free(%p)\n", svsk);
1900 if (svsk->sk_sock->file)
1901 sockfd_put(svsk->sk_sock);
1903 sock_release(svsk->sk_sock);
1908 * Remove a dead transport
1910 static void svc_delete_xprt(struct svc_xprt *xprt)
1912 struct svc_serv *serv = xprt->xpt_server;
1914 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1915 xprt->xpt_ops->xpo_detach(xprt);
1917 spin_lock_bh(&serv->sv_lock);
1918 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
1919 list_del_init(&xprt->xpt_list);
1921 * We used to delete the transport from whichever list
1922 * it's sk_xprt.xpt_ready node was on, but we don't actually
1923 * need to. This is because the only time we're called
1924 * while still attached to a queue, the queue itself
1925 * is about to be destroyed (in svc_destroy).
1927 if (!test_and_set_bit(XPT_DEAD, &xprt->xpt_flags)) {
1928 BUG_ON(atomic_read(&xprt->xpt_ref.refcount) < 2);
1929 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1933 spin_unlock_bh(&serv->sv_lock);
1936 static void svc_close_xprt(struct svc_xprt *xprt)
1938 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1939 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1940 /* someone else will have to effect the close */
1944 svc_delete_xprt(xprt);
1945 clear_bit(XPT_BUSY, &xprt->xpt_flags);
1949 void svc_close_all(struct list_head *xprt_list)
1951 struct svc_xprt *xprt;
1952 struct svc_xprt *tmp;
1954 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
1955 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1956 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
1957 /* Waiting to be processed, but no threads left,
1958 * So just remove it from the waiting list
1960 list_del_init(&xprt->xpt_ready);
1961 clear_bit(XPT_BUSY, &xprt->xpt_flags);
1963 svc_close_xprt(xprt);
1968 * Handle defer and revisit of requests
1971 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1973 struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1974 struct svc_xprt *xprt = dr->xprt;
1981 dprintk("revisit queued\n");
1983 spin_lock(&xprt->xpt_lock);
1984 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1985 spin_unlock(&xprt->xpt_lock);
1986 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1987 svc_xprt_enqueue(xprt);
1991 static struct cache_deferred_req *
1992 svc_defer(struct cache_req *req)
1994 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1995 int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
1996 struct svc_deferred_req *dr;
1998 if (rqstp->rq_arg.page_len)
1999 return NULL; /* if more than a page, give up FIXME */
2000 if (rqstp->rq_deferred) {
2001 dr = rqstp->rq_deferred;
2002 rqstp->rq_deferred = NULL;
2004 int skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
2005 /* FIXME maybe discard if size too large */
2006 dr = kmalloc(size, GFP_KERNEL);
2010 dr->handle.owner = rqstp->rq_server;
2011 dr->prot = rqstp->rq_prot;
2012 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
2013 dr->addrlen = rqstp->rq_addrlen;
2014 dr->daddr = rqstp->rq_daddr;
2015 dr->argslen = rqstp->rq_arg.len >> 2;
2016 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
2018 svc_xprt_get(rqstp->rq_xprt);
2019 dr->xprt = rqstp->rq_xprt;
2021 dr->handle.revisit = svc_revisit;
2026 * recv data from a deferred request into an active one
2028 static int svc_deferred_recv(struct svc_rqst *rqstp)
2030 struct svc_deferred_req *dr = rqstp->rq_deferred;
2032 rqstp->rq_arg.head[0].iov_base = dr->args;
2033 rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
2034 rqstp->rq_arg.page_len = 0;
2035 rqstp->rq_arg.len = dr->argslen<<2;
2036 rqstp->rq_prot = dr->prot;
2037 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
2038 rqstp->rq_addrlen = dr->addrlen;
2039 rqstp->rq_daddr = dr->daddr;
2040 rqstp->rq_respages = rqstp->rq_pages;
2041 return dr->argslen<<2;
2045 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
2047 struct svc_deferred_req *dr = NULL;
2049 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
2051 spin_lock(&xprt->xpt_lock);
2052 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
2053 if (!list_empty(&xprt->xpt_deferred)) {
2054 dr = list_entry(xprt->xpt_deferred.next,
2055 struct svc_deferred_req,
2057 list_del_init(&dr->handle.recent);
2058 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
2060 spin_unlock(&xprt->xpt_lock);