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 struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
315 struct svc_xprt *xprt;
317 if (list_empty(&pool->sp_sockets))
320 xprt = list_entry(pool->sp_sockets.next,
321 struct svc_xprt, xpt_ready);
322 list_del_init(&xprt->xpt_ready);
324 dprintk("svc: transport %p dequeued, inuse=%d\n",
325 xprt, atomic_read(&xprt->xpt_ref.refcount));
331 * svc_xprt_received conditionally queues the transport for processing
332 * by another thread. The caller must hold the XPT_BUSY bit and must
333 * not thereafter touch transport data.
335 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
336 * insufficient) data.
338 void svc_xprt_received(struct svc_xprt *xprt)
340 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
341 xprt->xpt_pool = NULL;
342 clear_bit(XPT_BUSY, &xprt->xpt_flags);
343 svc_xprt_enqueue(xprt);
345 EXPORT_SYMBOL_GPL(svc_xprt_received);
348 * svc_reserve - change the space reserved for the reply to a request.
349 * @rqstp: The request in question
350 * @space: new max space to reserve
352 * Each request reserves some space on the output queue of the socket
353 * to make sure the reply fits. This function reduces that reserved
354 * space to be the amount of space used already, plus @space.
357 void svc_reserve(struct svc_rqst *rqstp, int space)
359 space += rqstp->rq_res.head[0].iov_len;
361 if (space < rqstp->rq_reserved) {
362 struct svc_xprt *xprt = rqstp->rq_xprt;
363 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
364 rqstp->rq_reserved = space;
366 svc_xprt_enqueue(xprt);
370 static void svc_xprt_release(struct svc_rqst *rqstp)
372 struct svc_xprt *xprt = rqstp->rq_xprt;
374 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
376 svc_free_res_pages(rqstp);
377 rqstp->rq_res.page_len = 0;
378 rqstp->rq_res.page_base = 0;
380 /* Reset response buffer and release
382 * But first, check that enough space was reserved
383 * for the reply, otherwise we have a bug!
385 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
386 printk(KERN_ERR "RPC request reserved %d but used %d\n",
390 rqstp->rq_res.head[0].iov_len = 0;
391 svc_reserve(rqstp, 0);
392 rqstp->rq_xprt = NULL;
398 * External function to wake up a server waiting for data
399 * This really only makes sense for services like lockd
400 * which have exactly one thread anyway.
403 svc_wake_up(struct svc_serv *serv)
405 struct svc_rqst *rqstp;
407 struct svc_pool *pool;
409 for (i = 0; i < serv->sv_nrpools; i++) {
410 pool = &serv->sv_pools[i];
412 spin_lock_bh(&pool->sp_lock);
413 if (!list_empty(&pool->sp_threads)) {
414 rqstp = list_entry(pool->sp_threads.next,
417 dprintk("svc: daemon %p woken up.\n", rqstp);
419 svc_thread_dequeue(pool, rqstp);
420 rqstp->rq_sock = NULL;
422 wake_up(&rqstp->rq_wait);
424 spin_unlock_bh(&pool->sp_lock);
428 union svc_pktinfo_u {
429 struct in_pktinfo pkti;
430 struct in6_pktinfo pkti6;
432 #define SVC_PKTINFO_SPACE \
433 CMSG_SPACE(sizeof(union svc_pktinfo_u))
435 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
437 switch (rqstp->rq_sock->sk_sk->sk_family) {
439 struct in_pktinfo *pki = CMSG_DATA(cmh);
441 cmh->cmsg_level = SOL_IP;
442 cmh->cmsg_type = IP_PKTINFO;
443 pki->ipi_ifindex = 0;
444 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr.addr.s_addr;
445 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
450 struct in6_pktinfo *pki = CMSG_DATA(cmh);
452 cmh->cmsg_level = SOL_IPV6;
453 cmh->cmsg_type = IPV6_PKTINFO;
454 pki->ipi6_ifindex = 0;
455 ipv6_addr_copy(&pki->ipi6_addr,
456 &rqstp->rq_daddr.addr6);
457 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
465 * Generic sendto routine
468 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
470 struct svc_sock *svsk = rqstp->rq_sock;
471 struct socket *sock = svsk->sk_sock;
475 long all[SVC_PKTINFO_SPACE / sizeof(long)];
477 struct cmsghdr *cmh = &buffer.hdr;
481 struct page **ppage = xdr->pages;
482 size_t base = xdr->page_base;
483 unsigned int pglen = xdr->page_len;
484 unsigned int flags = MSG_MORE;
485 char buf[RPC_MAX_ADDRBUFLEN];
489 if (rqstp->rq_prot == IPPROTO_UDP) {
490 struct msghdr msg = {
491 .msg_name = &rqstp->rq_addr,
492 .msg_namelen = rqstp->rq_addrlen,
494 .msg_controllen = sizeof(buffer),
495 .msg_flags = MSG_MORE,
498 svc_set_cmsg_data(rqstp, cmh);
500 if (sock_sendmsg(sock, &msg, 0) < 0)
505 if (slen == xdr->head[0].iov_len)
507 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0,
508 xdr->head[0].iov_len, flags);
509 if (len != xdr->head[0].iov_len)
511 slen -= xdr->head[0].iov_len;
516 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
520 result = kernel_sendpage(sock, *ppage, base, size, flags);
527 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
532 if (xdr->tail[0].iov_len) {
533 result = kernel_sendpage(sock, rqstp->rq_respages[0],
534 ((unsigned long)xdr->tail[0].iov_base)
536 xdr->tail[0].iov_len, 0);
542 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
543 rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len,
544 xdr->len, len, svc_print_addr(rqstp, buf, sizeof(buf)));
550 * Report socket names for nfsdfs
552 static int one_sock_name(char *buf, struct svc_sock *svsk)
556 switch(svsk->sk_sk->sk_family) {
558 len = sprintf(buf, "ipv4 %s %u.%u.%u.%u %d\n",
559 svsk->sk_sk->sk_protocol==IPPROTO_UDP?
561 NIPQUAD(inet_sk(svsk->sk_sk)->rcv_saddr),
562 inet_sk(svsk->sk_sk)->num);
565 len = sprintf(buf, "*unknown-%d*\n",
566 svsk->sk_sk->sk_family);
572 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
574 struct svc_sock *svsk, *closesk = NULL;
579 spin_lock_bh(&serv->sv_lock);
580 list_for_each_entry(svsk, &serv->sv_permsocks, sk_xprt.xpt_list) {
581 int onelen = one_sock_name(buf+len, svsk);
582 if (toclose && strcmp(toclose, buf+len) == 0)
587 spin_unlock_bh(&serv->sv_lock);
589 /* Should unregister with portmap, but you cannot
590 * unregister just one protocol...
592 svc_close_xprt(&closesk->sk_xprt);
597 EXPORT_SYMBOL(svc_sock_names);
600 * Check input queue length
603 svc_recv_available(struct svc_sock *svsk)
605 struct socket *sock = svsk->sk_sock;
608 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
610 return (err >= 0)? avail : err;
614 * Generic recvfrom routine.
617 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
619 struct svc_sock *svsk = rqstp->rq_sock;
620 struct msghdr msg = {
621 .msg_flags = MSG_DONTWAIT,
625 len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
628 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
629 svsk, iov[0].iov_base, iov[0].iov_len, len);
634 * Set socket snd and rcv buffer lengths
637 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
641 oldfs = get_fs(); set_fs(KERNEL_DS);
642 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
643 (char*)&snd, sizeof(snd));
644 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
645 (char*)&rcv, sizeof(rcv));
647 /* sock_setsockopt limits use to sysctl_?mem_max,
648 * which isn't acceptable. Until that is made conditional
649 * on not having CAP_SYS_RESOURCE or similar, we go direct...
650 * DaveM said I could!
653 sock->sk->sk_sndbuf = snd * 2;
654 sock->sk->sk_rcvbuf = rcv * 2;
655 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
656 release_sock(sock->sk);
660 * INET callback when data has been received on the socket.
663 svc_udp_data_ready(struct sock *sk, int count)
665 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
668 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
670 test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
671 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
672 svc_xprt_enqueue(&svsk->sk_xprt);
674 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
675 wake_up_interruptible(sk->sk_sleep);
679 * INET callback when space is newly available on the socket.
682 svc_write_space(struct sock *sk)
684 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
687 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
688 svsk, sk, test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
689 svc_xprt_enqueue(&svsk->sk_xprt);
692 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
693 dprintk("RPC svc_write_space: someone sleeping on %p\n",
695 wake_up_interruptible(sk->sk_sleep);
700 * Copy the UDP datagram's destination address to the rqstp structure.
701 * The 'destination' address in this case is the address to which the
702 * peer sent the datagram, i.e. our local address. For multihomed
703 * hosts, this can change from msg to msg. Note that only the IP
704 * address changes, the port number should remain the same.
706 static void svc_udp_get_dest_address(struct svc_rqst *rqstp,
709 switch (rqstp->rq_sock->sk_sk->sk_family) {
711 struct in_pktinfo *pki = CMSG_DATA(cmh);
712 rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
716 struct in6_pktinfo *pki = CMSG_DATA(cmh);
717 ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
724 * Receive a datagram from a UDP socket.
727 svc_udp_recvfrom(struct svc_rqst *rqstp)
729 struct svc_sock *svsk = rqstp->rq_sock;
730 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
734 long all[SVC_PKTINFO_SPACE / sizeof(long)];
736 struct cmsghdr *cmh = &buffer.hdr;
738 struct msghdr msg = {
739 .msg_name = svc_addr(rqstp),
741 .msg_controllen = sizeof(buffer),
742 .msg_flags = MSG_DONTWAIT,
745 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
746 /* udp sockets need large rcvbuf as all pending
747 * requests are still in that buffer. sndbuf must
748 * also be large enough that there is enough space
749 * for one reply per thread. We count all threads
750 * rather than threads in a particular pool, which
751 * provides an upper bound on the number of threads
752 * which will access the socket.
754 svc_sock_setbufsize(svsk->sk_sock,
755 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
756 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
758 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
760 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
761 0, 0, MSG_PEEK | MSG_DONTWAIT);
763 skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err);
766 if (err != -EAGAIN) {
767 /* possibly an icmp error */
768 dprintk("svc: recvfrom returned error %d\n", -err);
769 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
771 svc_xprt_received(&svsk->sk_xprt);
774 len = svc_addr_len(svc_addr(rqstp));
777 rqstp->rq_addrlen = len;
778 if (skb->tstamp.tv64 == 0) {
779 skb->tstamp = ktime_get_real();
780 /* Don't enable netstamp, sunrpc doesn't
781 need that much accuracy */
783 svsk->sk_sk->sk_stamp = skb->tstamp;
784 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* there may be more data... */
787 * Maybe more packets - kick another thread ASAP.
789 svc_xprt_received(&svsk->sk_xprt);
791 len = skb->len - sizeof(struct udphdr);
792 rqstp->rq_arg.len = len;
794 rqstp->rq_prot = IPPROTO_UDP;
796 if (cmh->cmsg_level != IPPROTO_IP ||
797 cmh->cmsg_type != IP_PKTINFO) {
799 printk("rpcsvc: received unknown control message:"
801 cmh->cmsg_level, cmh->cmsg_type);
802 skb_free_datagram(svsk->sk_sk, skb);
805 svc_udp_get_dest_address(rqstp, cmh);
807 if (skb_is_nonlinear(skb)) {
808 /* we have to copy */
810 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
813 skb_free_datagram(svsk->sk_sk, skb);
817 skb_free_datagram(svsk->sk_sk, skb);
819 /* we can use it in-place */
820 rqstp->rq_arg.head[0].iov_base = skb->data + sizeof(struct udphdr);
821 rqstp->rq_arg.head[0].iov_len = len;
822 if (skb_checksum_complete(skb)) {
823 skb_free_datagram(svsk->sk_sk, skb);
826 rqstp->rq_xprt_ctxt = skb;
829 rqstp->rq_arg.page_base = 0;
830 if (len <= rqstp->rq_arg.head[0].iov_len) {
831 rqstp->rq_arg.head[0].iov_len = len;
832 rqstp->rq_arg.page_len = 0;
833 rqstp->rq_respages = rqstp->rq_pages+1;
835 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
836 rqstp->rq_respages = rqstp->rq_pages + 1 +
837 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
841 serv->sv_stats->netudpcnt++;
847 svc_udp_sendto(struct svc_rqst *rqstp)
851 error = svc_sendto(rqstp, &rqstp->rq_res);
852 if (error == -ECONNREFUSED)
853 /* ICMP error on earlier request. */
854 error = svc_sendto(rqstp, &rqstp->rq_res);
859 static void svc_udp_prep_reply_hdr(struct svc_rqst *rqstp)
863 static int svc_udp_has_wspace(struct svc_xprt *xprt)
865 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
866 struct svc_serv *serv = xprt->xpt_server;
867 unsigned long required;
870 * Set the SOCK_NOSPACE flag before checking the available
873 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
874 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
875 if (required*2 > sock_wspace(svsk->sk_sk))
877 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
881 static struct svc_xprt *svc_udp_accept(struct svc_xprt *xprt)
887 static struct svc_xprt *svc_udp_create(struct svc_serv *serv,
888 struct sockaddr *sa, int salen,
891 return svc_create_socket(serv, IPPROTO_UDP, sa, salen, flags);
894 static struct svc_xprt_ops svc_udp_ops = {
895 .xpo_create = svc_udp_create,
896 .xpo_recvfrom = svc_udp_recvfrom,
897 .xpo_sendto = svc_udp_sendto,
898 .xpo_release_rqst = svc_release_skb,
899 .xpo_detach = svc_sock_detach,
900 .xpo_free = svc_sock_free,
901 .xpo_prep_reply_hdr = svc_udp_prep_reply_hdr,
902 .xpo_has_wspace = svc_udp_has_wspace,
903 .xpo_accept = svc_udp_accept,
906 static struct svc_xprt_class svc_udp_class = {
908 .xcl_owner = THIS_MODULE,
909 .xcl_ops = &svc_udp_ops,
910 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
913 static void svc_udp_init(struct svc_sock *svsk, struct svc_serv *serv)
918 svc_xprt_init(&svc_udp_class, &svsk->sk_xprt, serv);
919 clear_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
920 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
921 svsk->sk_sk->sk_write_space = svc_write_space;
923 /* initialise setting must have enough space to
924 * receive and respond to one request.
925 * svc_udp_recvfrom will re-adjust if necessary
927 svc_sock_setbufsize(svsk->sk_sock,
928 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
929 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
931 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* might have come in before data_ready set up */
932 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
936 /* make sure we get destination address info */
937 svsk->sk_sock->ops->setsockopt(svsk->sk_sock, IPPROTO_IP, IP_PKTINFO,
938 (char __user *)&one, sizeof(one));
943 * A data_ready event on a listening socket means there's a connection
944 * pending. Do not use state_change as a substitute for it.
947 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
949 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
951 dprintk("svc: socket %p TCP (listen) state change %d\n",
955 * This callback may called twice when a new connection
956 * is established as a child socket inherits everything
957 * from a parent LISTEN socket.
958 * 1) data_ready method of the parent socket will be called
959 * when one of child sockets become ESTABLISHED.
960 * 2) data_ready method of the child socket may be called
961 * when it receives data before the socket is accepted.
962 * In case of 2, we should ignore it silently.
964 if (sk->sk_state == TCP_LISTEN) {
966 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
967 svc_xprt_enqueue(&svsk->sk_xprt);
969 printk("svc: socket %p: no user data\n", sk);
972 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
973 wake_up_interruptible_all(sk->sk_sleep);
977 * A state change on a connected socket means it's dying or dead.
980 svc_tcp_state_change(struct sock *sk)
982 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
984 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
985 sk, sk->sk_state, sk->sk_user_data);
988 printk("svc: socket %p: no user data\n", sk);
990 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
991 svc_xprt_enqueue(&svsk->sk_xprt);
993 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
994 wake_up_interruptible_all(sk->sk_sleep);
998 svc_tcp_data_ready(struct sock *sk, int count)
1000 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
1002 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
1003 sk, sk->sk_user_data);
1005 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1006 svc_xprt_enqueue(&svsk->sk_xprt);
1008 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1009 wake_up_interruptible(sk->sk_sleep);
1012 static inline int svc_port_is_privileged(struct sockaddr *sin)
1014 switch (sin->sa_family) {
1016 return ntohs(((struct sockaddr_in *)sin)->sin_port)
1019 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
1027 * Accept a TCP connection
1029 static struct svc_xprt *svc_tcp_accept(struct svc_xprt *xprt)
1031 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1032 struct sockaddr_storage addr;
1033 struct sockaddr *sin = (struct sockaddr *) &addr;
1034 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1035 struct socket *sock = svsk->sk_sock;
1036 struct socket *newsock;
1037 struct svc_sock *newsvsk;
1039 char buf[RPC_MAX_ADDRBUFLEN];
1041 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
1045 clear_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1046 err = kernel_accept(sock, &newsock, O_NONBLOCK);
1049 printk(KERN_WARNING "%s: no more sockets!\n",
1051 else if (err != -EAGAIN && net_ratelimit())
1052 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
1053 serv->sv_name, -err);
1056 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1058 err = kernel_getpeername(newsock, sin, &slen);
1060 if (net_ratelimit())
1061 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
1062 serv->sv_name, -err);
1063 goto failed; /* aborted connection or whatever */
1066 /* Ideally, we would want to reject connections from unauthorized
1067 * hosts here, but when we get encryption, the IP of the host won't
1068 * tell us anything. For now just warn about unpriv connections.
1070 if (!svc_port_is_privileged(sin)) {
1071 dprintk(KERN_WARNING
1072 "%s: connect from unprivileged port: %s\n",
1074 __svc_print_addr(sin, buf, sizeof(buf)));
1076 dprintk("%s: connect from %s\n", serv->sv_name,
1077 __svc_print_addr(sin, buf, sizeof(buf)));
1079 /* make sure that a write doesn't block forever when
1082 newsock->sk->sk_sndtimeo = HZ*30;
1084 if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
1085 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
1087 svc_xprt_set_remote(&newsvsk->sk_xprt, sin, slen);
1088 err = kernel_getsockname(newsock, sin, &slen);
1089 if (unlikely(err < 0)) {
1090 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err);
1091 slen = offsetof(struct sockaddr, sa_data);
1093 svc_xprt_set_local(&newsvsk->sk_xprt, sin, slen);
1096 serv->sv_stats->nettcpconn++;
1098 return &newsvsk->sk_xprt;
1101 sock_release(newsock);
1106 * Receive data from a TCP socket.
1109 svc_tcp_recvfrom(struct svc_rqst *rqstp)
1111 struct svc_sock *svsk = rqstp->rq_sock;
1112 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1117 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
1118 svsk, test_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags),
1119 test_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags),
1120 test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags));
1122 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
1123 /* sndbuf needs to have room for one request
1124 * per thread, otherwise we can stall even when the
1125 * network isn't a bottleneck.
1127 * We count all threads rather than threads in a
1128 * particular pool, which provides an upper bound
1129 * on the number of threads which will access the socket.
1131 * rcvbuf just needs to be able to hold a few requests.
1132 * Normally they will be removed from the queue
1133 * as soon a a complete request arrives.
1135 svc_sock_setbufsize(svsk->sk_sock,
1136 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
1137 3 * serv->sv_max_mesg);
1139 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1141 /* Receive data. If we haven't got the record length yet, get
1142 * the next four bytes. Otherwise try to gobble up as much as
1143 * possible up to the complete record length.
1145 if (svsk->sk_tcplen < 4) {
1146 unsigned long want = 4 - svsk->sk_tcplen;
1149 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
1151 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
1153 svsk->sk_tcplen += len;
1156 dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
1158 svc_xprt_received(&svsk->sk_xprt);
1159 return -EAGAIN; /* record header not complete */
1162 svsk->sk_reclen = ntohl(svsk->sk_reclen);
1163 if (!(svsk->sk_reclen & 0x80000000)) {
1164 /* FIXME: technically, a record can be fragmented,
1165 * and non-terminal fragments will not have the top
1166 * bit set in the fragment length header.
1167 * But apparently no known nfs clients send fragmented
1169 if (net_ratelimit())
1170 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1171 " (non-terminal)\n",
1172 (unsigned long) svsk->sk_reclen);
1175 svsk->sk_reclen &= 0x7fffffff;
1176 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
1177 if (svsk->sk_reclen > serv->sv_max_mesg) {
1178 if (net_ratelimit())
1179 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1181 (unsigned long) svsk->sk_reclen);
1186 /* Check whether enough data is available */
1187 len = svc_recv_available(svsk);
1191 if (len < svsk->sk_reclen) {
1192 dprintk("svc: incomplete TCP record (%d of %d)\n",
1193 len, svsk->sk_reclen);
1194 svc_xprt_received(&svsk->sk_xprt);
1195 return -EAGAIN; /* record not complete */
1197 len = svsk->sk_reclen;
1198 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1200 vec = rqstp->rq_vec;
1201 vec[0] = rqstp->rq_arg.head[0];
1204 while (vlen < len) {
1205 vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
1206 vec[pnum].iov_len = PAGE_SIZE;
1210 rqstp->rq_respages = &rqstp->rq_pages[pnum];
1212 /* Now receive data */
1213 len = svc_recvfrom(rqstp, vec, pnum, len);
1217 dprintk("svc: TCP complete record (%d bytes)\n", len);
1218 rqstp->rq_arg.len = len;
1219 rqstp->rq_arg.page_base = 0;
1220 if (len <= rqstp->rq_arg.head[0].iov_len) {
1221 rqstp->rq_arg.head[0].iov_len = len;
1222 rqstp->rq_arg.page_len = 0;
1224 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1227 rqstp->rq_xprt_ctxt = NULL;
1228 rqstp->rq_prot = IPPROTO_TCP;
1230 /* Reset TCP read info */
1231 svsk->sk_reclen = 0;
1232 svsk->sk_tcplen = 0;
1234 svc_xprt_copy_addrs(rqstp, &svsk->sk_xprt);
1235 svc_xprt_received(&svsk->sk_xprt);
1237 serv->sv_stats->nettcpcnt++;
1242 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1246 if (len == -EAGAIN) {
1247 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1248 svc_xprt_received(&svsk->sk_xprt);
1250 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1251 svsk->sk_xprt.xpt_server->sv_name, -len);
1259 * Send out data on TCP socket.
1262 svc_tcp_sendto(struct svc_rqst *rqstp)
1264 struct xdr_buf *xbufp = &rqstp->rq_res;
1268 /* Set up the first element of the reply kvec.
1269 * Any other kvecs that may be in use have been taken
1270 * care of by the server implementation itself.
1272 reclen = htonl(0x80000000|((xbufp->len ) - 4));
1273 memcpy(xbufp->head[0].iov_base, &reclen, 4);
1275 if (test_bit(XPT_DEAD, &rqstp->rq_sock->sk_xprt.xpt_flags))
1278 sent = svc_sendto(rqstp, &rqstp->rq_res);
1279 if (sent != xbufp->len) {
1280 printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1281 rqstp->rq_sock->sk_xprt.xpt_server->sv_name,
1282 (sent<0)?"got error":"sent only",
1284 set_bit(XPT_CLOSE, &rqstp->rq_sock->sk_xprt.xpt_flags);
1285 svc_xprt_enqueue(rqstp->rq_xprt);
1292 * Setup response header. TCP has a 4B record length field.
1294 static void svc_tcp_prep_reply_hdr(struct svc_rqst *rqstp)
1296 struct kvec *resv = &rqstp->rq_res.head[0];
1298 /* tcp needs a space for the record length... */
1302 static int svc_tcp_has_wspace(struct svc_xprt *xprt)
1304 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1305 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1310 * Set the SOCK_NOSPACE flag before checking the available
1313 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1314 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
1315 wspace = sk_stream_wspace(svsk->sk_sk);
1317 if (wspace < sk_stream_min_wspace(svsk->sk_sk))
1319 if (required * 2 > wspace)
1322 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1326 static struct svc_xprt *svc_tcp_create(struct svc_serv *serv,
1327 struct sockaddr *sa, int salen,
1330 return svc_create_socket(serv, IPPROTO_TCP, sa, salen, flags);
1333 static struct svc_xprt_ops svc_tcp_ops = {
1334 .xpo_create = svc_tcp_create,
1335 .xpo_recvfrom = svc_tcp_recvfrom,
1336 .xpo_sendto = svc_tcp_sendto,
1337 .xpo_release_rqst = svc_release_skb,
1338 .xpo_detach = svc_sock_detach,
1339 .xpo_free = svc_sock_free,
1340 .xpo_prep_reply_hdr = svc_tcp_prep_reply_hdr,
1341 .xpo_has_wspace = svc_tcp_has_wspace,
1342 .xpo_accept = svc_tcp_accept,
1345 static struct svc_xprt_class svc_tcp_class = {
1347 .xcl_owner = THIS_MODULE,
1348 .xcl_ops = &svc_tcp_ops,
1349 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1352 void svc_init_xprt_sock(void)
1354 svc_reg_xprt_class(&svc_tcp_class);
1355 svc_reg_xprt_class(&svc_udp_class);
1358 void svc_cleanup_xprt_sock(void)
1360 svc_unreg_xprt_class(&svc_tcp_class);
1361 svc_unreg_xprt_class(&svc_udp_class);
1364 static void svc_tcp_init(struct svc_sock *svsk, struct svc_serv *serv)
1366 struct sock *sk = svsk->sk_sk;
1367 struct tcp_sock *tp = tcp_sk(sk);
1369 svc_xprt_init(&svc_tcp_class, &svsk->sk_xprt, serv);
1370 set_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
1371 if (sk->sk_state == TCP_LISTEN) {
1372 dprintk("setting up TCP socket for listening\n");
1373 set_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags);
1374 sk->sk_data_ready = svc_tcp_listen_data_ready;
1375 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1377 dprintk("setting up TCP socket for reading\n");
1378 sk->sk_state_change = svc_tcp_state_change;
1379 sk->sk_data_ready = svc_tcp_data_ready;
1380 sk->sk_write_space = svc_write_space;
1382 svsk->sk_reclen = 0;
1383 svsk->sk_tcplen = 0;
1385 tp->nonagle = 1; /* disable Nagle's algorithm */
1387 /* initialise setting must have enough space to
1388 * receive and respond to one request.
1389 * svc_tcp_recvfrom will re-adjust if necessary
1391 svc_sock_setbufsize(svsk->sk_sock,
1392 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
1393 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
1395 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1396 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1397 if (sk->sk_state != TCP_ESTABLISHED)
1398 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1403 svc_sock_update_bufs(struct svc_serv *serv)
1406 * The number of server threads has changed. Update
1407 * rcvbuf and sndbuf accordingly on all sockets
1409 struct list_head *le;
1411 spin_lock_bh(&serv->sv_lock);
1412 list_for_each(le, &serv->sv_permsocks) {
1413 struct svc_sock *svsk =
1414 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1415 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1417 list_for_each(le, &serv->sv_tempsocks) {
1418 struct svc_sock *svsk =
1419 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1420 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1422 spin_unlock_bh(&serv->sv_lock);
1426 * Make sure that we don't have too many active connections. If we
1427 * have, something must be dropped.
1429 * There's no point in trying to do random drop here for DoS
1430 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
1431 * attacker can easily beat that.
1433 * The only somewhat efficient mechanism would be if drop old
1434 * connections from the same IP first. But right now we don't even
1435 * record the client IP in svc_sock.
1437 static void svc_check_conn_limits(struct svc_serv *serv)
1439 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
1440 struct svc_sock *svsk = NULL;
1441 spin_lock_bh(&serv->sv_lock);
1442 if (!list_empty(&serv->sv_tempsocks)) {
1443 if (net_ratelimit()) {
1444 /* Try to help the admin */
1445 printk(KERN_NOTICE "%s: too many open TCP "
1446 "sockets, consider increasing the "
1447 "number of nfsd threads\n",
1451 * Always select the oldest socket. It's not fair,
1454 svsk = list_entry(serv->sv_tempsocks.prev,
1457 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1458 svc_xprt_get(&svsk->sk_xprt);
1460 spin_unlock_bh(&serv->sv_lock);
1463 svc_xprt_enqueue(&svsk->sk_xprt);
1464 svc_xprt_put(&svsk->sk_xprt);
1470 * Receive the next request on any socket. This code is carefully
1471 * organised not to touch any cachelines in the shared svc_serv
1472 * structure, only cachelines in the local svc_pool.
1475 svc_recv(struct svc_rqst *rqstp, long timeout)
1477 struct svc_xprt *xprt = NULL;
1478 struct svc_serv *serv = rqstp->rq_server;
1479 struct svc_pool *pool = rqstp->rq_pool;
1482 struct xdr_buf *arg;
1483 DECLARE_WAITQUEUE(wait, current);
1485 dprintk("svc: server %p waiting for data (to = %ld)\n",
1490 "svc_recv: service %p, transport not NULL!\n",
1492 if (waitqueue_active(&rqstp->rq_wait))
1494 "svc_recv: service %p, wait queue active!\n",
1498 /* now allocate needed pages. If we get a failure, sleep briefly */
1499 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
1500 for (i=0; i < pages ; i++)
1501 while (rqstp->rq_pages[i] == NULL) {
1502 struct page *p = alloc_page(GFP_KERNEL);
1504 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1505 rqstp->rq_pages[i] = p;
1507 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
1508 BUG_ON(pages >= RPCSVC_MAXPAGES);
1510 /* Make arg->head point to first page and arg->pages point to rest */
1511 arg = &rqstp->rq_arg;
1512 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
1513 arg->head[0].iov_len = PAGE_SIZE;
1514 arg->pages = rqstp->rq_pages + 1;
1516 /* save at least one page for response */
1517 arg->page_len = (pages-2)*PAGE_SIZE;
1518 arg->len = (pages-1)*PAGE_SIZE;
1519 arg->tail[0].iov_len = 0;
1526 spin_lock_bh(&pool->sp_lock);
1527 xprt = svc_xprt_dequeue(pool);
1529 rqstp->rq_xprt = xprt;
1531 rqstp->rq_reserved = serv->sv_max_mesg;
1532 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
1534 /* No data pending. Go to sleep */
1535 svc_thread_enqueue(pool, rqstp);
1538 * We have to be able to interrupt this wait
1539 * to bring down the daemons ...
1541 set_current_state(TASK_INTERRUPTIBLE);
1542 add_wait_queue(&rqstp->rq_wait, &wait);
1543 spin_unlock_bh(&pool->sp_lock);
1545 schedule_timeout(timeout);
1549 spin_lock_bh(&pool->sp_lock);
1550 remove_wait_queue(&rqstp->rq_wait, &wait);
1552 xprt = rqstp->rq_xprt;
1554 svc_thread_dequeue(pool, rqstp);
1555 spin_unlock_bh(&pool->sp_lock);
1556 dprintk("svc: server %p, no data yet\n", rqstp);
1557 return signalled()? -EINTR : -EAGAIN;
1560 spin_unlock_bh(&pool->sp_lock);
1563 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
1564 dprintk("svc_recv: found XPT_CLOSE\n");
1565 svc_delete_xprt(xprt);
1566 } else if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
1567 struct svc_xprt *newxpt;
1568 newxpt = xprt->xpt_ops->xpo_accept(xprt);
1571 * We know this module_get will succeed because the
1572 * listener holds a reference too
1574 __module_get(newxpt->xpt_class->xcl_owner);
1575 svc_check_conn_limits(xprt->xpt_server);
1576 svc_xprt_received(newxpt);
1578 svc_xprt_received(xprt);
1580 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
1581 rqstp, pool->sp_id, xprt,
1582 atomic_read(&xprt->xpt_ref.refcount));
1583 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
1584 if (rqstp->rq_deferred) {
1585 svc_xprt_received(xprt);
1586 len = svc_deferred_recv(rqstp);
1588 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
1589 dprintk("svc: got len=%d\n", len);
1592 /* No data, incomplete (TCP) read, or accept() */
1593 if (len == 0 || len == -EAGAIN) {
1594 rqstp->rq_res.len = 0;
1595 svc_xprt_release(rqstp);
1598 clear_bit(XPT_OLD, &xprt->xpt_flags);
1600 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
1601 rqstp->rq_chandle.defer = svc_defer;
1604 serv->sv_stats->netcnt++;
1612 svc_drop(struct svc_rqst *rqstp)
1614 dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1615 svc_xprt_release(rqstp);
1619 * Return reply to client.
1622 svc_send(struct svc_rqst *rqstp)
1624 struct svc_xprt *xprt;
1628 xprt = rqstp->rq_xprt;
1632 /* release the receive skb before sending the reply */
1633 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
1635 /* calculate over-all length */
1636 xb = & rqstp->rq_res;
1637 xb->len = xb->head[0].iov_len +
1639 xb->tail[0].iov_len;
1641 /* Grab mutex to serialize outgoing data. */
1642 mutex_lock(&xprt->xpt_mutex);
1643 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
1646 len = xprt->xpt_ops->xpo_sendto(rqstp);
1647 mutex_unlock(&xprt->xpt_mutex);
1648 svc_xprt_release(rqstp);
1650 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1656 * Timer function to close old temporary sockets, using
1657 * a mark-and-sweep algorithm.
1660 svc_age_temp_sockets(unsigned long closure)
1662 struct svc_serv *serv = (struct svc_serv *)closure;
1663 struct svc_sock *svsk;
1664 struct list_head *le, *next;
1665 LIST_HEAD(to_be_aged);
1667 dprintk("svc_age_temp_sockets\n");
1669 if (!spin_trylock_bh(&serv->sv_lock)) {
1670 /* busy, try again 1 sec later */
1671 dprintk("svc_age_temp_sockets: busy\n");
1672 mod_timer(&serv->sv_temptimer, jiffies + HZ);
1676 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1677 svsk = list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1679 if (!test_and_set_bit(XPT_OLD, &svsk->sk_xprt.xpt_flags))
1681 if (atomic_read(&svsk->sk_xprt.xpt_ref.refcount) > 1
1682 || test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags))
1684 svc_xprt_get(&svsk->sk_xprt);
1685 list_move(le, &to_be_aged);
1686 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1687 set_bit(XPT_DETACHED, &svsk->sk_xprt.xpt_flags);
1689 spin_unlock_bh(&serv->sv_lock);
1691 while (!list_empty(&to_be_aged)) {
1692 le = to_be_aged.next;
1693 /* fiddling the sk_xprt.xpt_list node is safe 'cos we're XPT_DETACHED */
1695 svsk = list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1697 dprintk("queuing svsk %p for closing\n", svsk);
1699 /* a thread will dequeue and close it soon */
1700 svc_xprt_enqueue(&svsk->sk_xprt);
1701 svc_xprt_put(&svsk->sk_xprt);
1704 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
1708 * Initialize socket for RPC use and create svc_sock struct
1709 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1711 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1712 struct socket *sock,
1713 int *errp, int flags)
1715 struct svc_sock *svsk;
1717 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1718 int is_temporary = flags & SVC_SOCK_TEMPORARY;
1720 dprintk("svc: svc_setup_socket %p\n", sock);
1721 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1728 /* Register socket with portmapper */
1729 if (*errp >= 0 && pmap_register)
1730 *errp = svc_register(serv, inet->sk_protocol,
1731 ntohs(inet_sk(inet)->sport));
1738 set_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags);
1739 inet->sk_user_data = svsk;
1740 svsk->sk_sock = sock;
1742 svsk->sk_ostate = inet->sk_state_change;
1743 svsk->sk_odata = inet->sk_data_ready;
1744 svsk->sk_owspace = inet->sk_write_space;
1746 /* Initialize the socket */
1747 if (sock->type == SOCK_DGRAM)
1748 svc_udp_init(svsk, serv);
1750 svc_tcp_init(svsk, serv);
1752 spin_lock_bh(&serv->sv_lock);
1754 set_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1755 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_tempsocks);
1757 if (serv->sv_temptimer.function == NULL) {
1758 /* setup timer to age temp sockets */
1759 setup_timer(&serv->sv_temptimer, svc_age_temp_sockets,
1760 (unsigned long)serv);
1761 mod_timer(&serv->sv_temptimer,
1762 jiffies + svc_conn_age_period * HZ);
1765 clear_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1766 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_permsocks);
1768 spin_unlock_bh(&serv->sv_lock);
1770 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1776 int svc_addsock(struct svc_serv *serv,
1782 struct socket *so = sockfd_lookup(fd, &err);
1783 struct svc_sock *svsk = NULL;
1787 if (so->sk->sk_family != AF_INET)
1788 err = -EAFNOSUPPORT;
1789 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1790 so->sk->sk_protocol != IPPROTO_UDP)
1791 err = -EPROTONOSUPPORT;
1792 else if (so->state > SS_UNCONNECTED)
1795 svsk = svc_setup_socket(serv, so, &err, SVC_SOCK_DEFAULTS);
1797 struct sockaddr_storage addr;
1798 struct sockaddr *sin = (struct sockaddr *)&addr;
1800 if (kernel_getsockname(svsk->sk_sock, sin, &salen) == 0)
1801 svc_xprt_set_local(&svsk->sk_xprt, sin, salen);
1802 svc_xprt_received(&svsk->sk_xprt);
1810 if (proto) *proto = so->sk->sk_protocol;
1811 return one_sock_name(name_return, svsk);
1813 EXPORT_SYMBOL_GPL(svc_addsock);
1816 * Create socket for RPC service.
1818 static struct svc_xprt *svc_create_socket(struct svc_serv *serv,
1820 struct sockaddr *sin, int len,
1823 struct svc_sock *svsk;
1824 struct socket *sock;
1827 char buf[RPC_MAX_ADDRBUFLEN];
1828 struct sockaddr_storage addr;
1829 struct sockaddr *newsin = (struct sockaddr *)&addr;
1832 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1833 serv->sv_program->pg_name, protocol,
1834 __svc_print_addr(sin, buf, sizeof(buf)));
1836 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1837 printk(KERN_WARNING "svc: only UDP and TCP "
1838 "sockets supported\n");
1839 return ERR_PTR(-EINVAL);
1841 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1843 error = sock_create_kern(sin->sa_family, type, protocol, &sock);
1845 return ERR_PTR(error);
1847 svc_reclassify_socket(sock);
1849 if (type == SOCK_STREAM)
1850 sock->sk->sk_reuse = 1; /* allow address reuse */
1851 error = kernel_bind(sock, sin, len);
1856 error = kernel_getsockname(sock, newsin, &newlen);
1860 if (protocol == IPPROTO_TCP) {
1861 if ((error = kernel_listen(sock, 64)) < 0)
1865 if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1866 svc_xprt_set_local(&svsk->sk_xprt, newsin, newlen);
1867 svc_xprt_received(&svsk->sk_xprt);
1868 return (struct svc_xprt *)svsk;
1872 dprintk("svc: svc_create_socket error = %d\n", -error);
1874 return ERR_PTR(error);
1878 * Detach the svc_sock from the socket so that no
1879 * more callbacks occur.
1881 static void svc_sock_detach(struct svc_xprt *xprt)
1883 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1884 struct sock *sk = svsk->sk_sk;
1886 dprintk("svc: svc_sock_detach(%p)\n", svsk);
1888 /* put back the old socket callbacks */
1889 sk->sk_state_change = svsk->sk_ostate;
1890 sk->sk_data_ready = svsk->sk_odata;
1891 sk->sk_write_space = svsk->sk_owspace;
1895 * Free the svc_sock's socket resources and the svc_sock itself.
1897 static void svc_sock_free(struct svc_xprt *xprt)
1899 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1900 dprintk("svc: svc_sock_free(%p)\n", svsk);
1902 if (svsk->sk_sock->file)
1903 sockfd_put(svsk->sk_sock);
1905 sock_release(svsk->sk_sock);
1910 * Remove a dead transport
1912 static void svc_delete_xprt(struct svc_xprt *xprt)
1914 struct svc_serv *serv = xprt->xpt_server;
1916 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1917 xprt->xpt_ops->xpo_detach(xprt);
1919 spin_lock_bh(&serv->sv_lock);
1920 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
1921 list_del_init(&xprt->xpt_list);
1923 * We used to delete the transport from whichever list
1924 * it's sk_xprt.xpt_ready node was on, but we don't actually
1925 * need to. This is because the only time we're called
1926 * while still attached to a queue, the queue itself
1927 * is about to be destroyed (in svc_destroy).
1929 if (!test_and_set_bit(XPT_DEAD, &xprt->xpt_flags)) {
1930 BUG_ON(atomic_read(&xprt->xpt_ref.refcount) < 2);
1931 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1935 spin_unlock_bh(&serv->sv_lock);
1938 static void svc_close_xprt(struct svc_xprt *xprt)
1940 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1941 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1942 /* someone else will have to effect the close */
1946 svc_delete_xprt(xprt);
1947 clear_bit(XPT_BUSY, &xprt->xpt_flags);
1951 void svc_close_all(struct list_head *xprt_list)
1953 struct svc_xprt *xprt;
1954 struct svc_xprt *tmp;
1956 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
1957 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1958 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
1959 /* Waiting to be processed, but no threads left,
1960 * So just remove it from the waiting list
1962 list_del_init(&xprt->xpt_ready);
1963 clear_bit(XPT_BUSY, &xprt->xpt_flags);
1965 svc_close_xprt(xprt);
1970 * Handle defer and revisit of requests
1973 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1975 struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1976 struct svc_xprt *xprt = dr->xprt;
1983 dprintk("revisit queued\n");
1985 spin_lock(&xprt->xpt_lock);
1986 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1987 spin_unlock(&xprt->xpt_lock);
1988 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1989 svc_xprt_enqueue(xprt);
1993 static struct cache_deferred_req *
1994 svc_defer(struct cache_req *req)
1996 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1997 int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
1998 struct svc_deferred_req *dr;
2000 if (rqstp->rq_arg.page_len)
2001 return NULL; /* if more than a page, give up FIXME */
2002 if (rqstp->rq_deferred) {
2003 dr = rqstp->rq_deferred;
2004 rqstp->rq_deferred = NULL;
2006 int skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
2007 /* FIXME maybe discard if size too large */
2008 dr = kmalloc(size, GFP_KERNEL);
2012 dr->handle.owner = rqstp->rq_server;
2013 dr->prot = rqstp->rq_prot;
2014 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
2015 dr->addrlen = rqstp->rq_addrlen;
2016 dr->daddr = rqstp->rq_daddr;
2017 dr->argslen = rqstp->rq_arg.len >> 2;
2018 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
2020 svc_xprt_get(rqstp->rq_xprt);
2021 dr->xprt = rqstp->rq_xprt;
2023 dr->handle.revisit = svc_revisit;
2028 * recv data from a deferred request into an active one
2030 static int svc_deferred_recv(struct svc_rqst *rqstp)
2032 struct svc_deferred_req *dr = rqstp->rq_deferred;
2034 rqstp->rq_arg.head[0].iov_base = dr->args;
2035 rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
2036 rqstp->rq_arg.page_len = 0;
2037 rqstp->rq_arg.len = dr->argslen<<2;
2038 rqstp->rq_prot = dr->prot;
2039 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
2040 rqstp->rq_addrlen = dr->addrlen;
2041 rqstp->rq_daddr = dr->daddr;
2042 rqstp->rq_respages = rqstp->rq_pages;
2043 return dr->argslen<<2;
2047 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
2049 struct svc_deferred_req *dr = NULL;
2051 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
2053 spin_lock(&xprt->xpt_lock);
2054 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
2055 if (!list_empty(&xprt->xpt_deferred)) {
2056 dr = list_entry(xprt->xpt_deferred.next,
2057 struct svc_deferred_req,
2059 list_del_init(&dr->handle.recent);
2060 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
2062 spin_unlock(&xprt->xpt_lock);