2 * linux/net/sunrpc/svcsock.c
4 * These are the RPC server socket internals.
6 * The server scheduling algorithm does not always distribute the load
7 * evenly when servicing a single client. May need to modify the
8 * svc_xprt_enqueue procedure...
10 * TCP support is largely untested and may be a little slow. The problem
11 * is that we currently do two separate recvfrom's, one for the 4-byte
12 * record length, and the second for the actual record. This could possibly
13 * be improved by always reading a minimum size of around 100 bytes and
14 * tucking any superfluous bytes away in a temporary store. Still, that
15 * leaves write requests out in the rain. An alternative may be to peek at
16 * the first skb in the queue, and if it matches the next TCP sequence
17 * number, to extract the record marker. Yuck.
19 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
22 #include <linux/kernel.h>
23 #include <linux/sched.h>
24 #include <linux/errno.h>
25 #include <linux/fcntl.h>
26 #include <linux/net.h>
28 #include <linux/inet.h>
29 #include <linux/udp.h>
30 #include <linux/tcp.h>
31 #include <linux/unistd.h>
32 #include <linux/slab.h>
33 #include <linux/netdevice.h>
34 #include <linux/skbuff.h>
35 #include <linux/file.h>
36 #include <linux/freezer.h>
38 #include <net/checksum.h>
41 #include <net/tcp_states.h>
42 #include <asm/uaccess.h>
43 #include <asm/ioctls.h>
45 #include <linux/sunrpc/types.h>
46 #include <linux/sunrpc/clnt.h>
47 #include <linux/sunrpc/xdr.h>
48 #include <linux/sunrpc/svcsock.h>
49 #include <linux/sunrpc/stats.h>
51 /* SMP locking strategy:
53 * svc_pool->sp_lock protects most of the fields of that pool.
54 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
55 * when both need to be taken (rare), svc_serv->sv_lock is first.
56 * BKL protects svc_serv->sv_nrthread.
57 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
58 * and the ->sk_info_authunix cache.
59 * svc_sock->sk_xprt.xpt_flags.XPT_BUSY prevents a svc_sock being
62 * Some flags can be set to certain values at any time
63 * providing that certain rules are followed:
65 * XPT_CONN, XPT_DATA, can be set or cleared at any time.
66 * after a set, svc_xprt_enqueue must be called.
67 * after a clear, the socket must be read/accepted
68 * if this succeeds, it must be set again.
69 * XPT_CLOSE can set at any time. It is never cleared.
70 * xpt_ref contains a bias of '1' until XPT_DEAD is set.
71 * so when xprt_ref hits zero, we know the transport is dead
72 * and no-one is using it.
73 * XPT_DEAD can only be set while XPT_BUSY is held which ensures
74 * no other thread will be using the socket or will try to
79 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
82 static struct svc_sock *svc_setup_socket(struct svc_serv *, struct socket *,
83 int *errp, int flags);
84 static void svc_delete_xprt(struct svc_xprt *xprt);
85 static void svc_udp_data_ready(struct sock *, int);
86 static int svc_udp_recvfrom(struct svc_rqst *);
87 static int svc_udp_sendto(struct svc_rqst *);
88 static void svc_close_xprt(struct svc_xprt *xprt);
89 static void svc_sock_detach(struct svc_xprt *);
90 static void svc_sock_free(struct svc_xprt *);
92 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
93 static int svc_deferred_recv(struct svc_rqst *rqstp);
94 static struct cache_deferred_req *svc_defer(struct cache_req *req);
95 static struct svc_xprt *svc_create_socket(struct svc_serv *, int,
96 struct sockaddr *, int, int);
98 /* apparently the "standard" is that clients close
99 * idle connections after 5 minutes, servers after
101 * http://www.connectathon.org/talks96/nfstcp.pdf
103 static int svc_conn_age_period = 6*60;
105 #ifdef CONFIG_DEBUG_LOCK_ALLOC
106 static struct lock_class_key svc_key[2];
107 static struct lock_class_key svc_slock_key[2];
109 static inline void svc_reclassify_socket(struct socket *sock)
111 struct sock *sk = sock->sk;
112 BUG_ON(sock_owned_by_user(sk));
113 switch (sk->sk_family) {
115 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFSD",
117 "sk_xprt.xpt_lock-AF_INET-NFSD",
122 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFSD",
124 "sk_xprt.xpt_lock-AF_INET6-NFSD",
133 static inline void svc_reclassify_socket(struct socket *sock)
138 static char *__svc_print_addr(struct sockaddr *addr, char *buf, size_t len)
140 switch (addr->sa_family) {
142 snprintf(buf, len, "%u.%u.%u.%u, port=%u",
143 NIPQUAD(((struct sockaddr_in *) addr)->sin_addr),
144 ntohs(((struct sockaddr_in *) addr)->sin_port));
148 snprintf(buf, len, "%x:%x:%x:%x:%x:%x:%x:%x, port=%u",
149 NIP6(((struct sockaddr_in6 *) addr)->sin6_addr),
150 ntohs(((struct sockaddr_in6 *) addr)->sin6_port));
154 snprintf(buf, len, "unknown address type: %d", addr->sa_family);
161 * svc_print_addr - Format rq_addr field for printing
162 * @rqstp: svc_rqst struct containing address to print
163 * @buf: target buffer for formatted address
164 * @len: length of target buffer
167 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
169 return __svc_print_addr(svc_addr(rqstp), buf, len);
171 EXPORT_SYMBOL_GPL(svc_print_addr);
174 * Queue up an idle server thread. Must have pool->sp_lock held.
175 * Note: this is really a stack rather than a queue, so that we only
176 * use as many different threads as we need, and the rest don't pollute
180 svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
182 list_add(&rqstp->rq_list, &pool->sp_threads);
186 * Dequeue an nfsd thread. Must have pool->sp_lock held.
189 svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
191 list_del(&rqstp->rq_list);
195 * Release an skbuff after use
197 static void svc_release_skb(struct svc_rqst *rqstp)
199 struct sk_buff *skb = rqstp->rq_xprt_ctxt;
200 struct svc_deferred_req *dr = rqstp->rq_deferred;
203 rqstp->rq_xprt_ctxt = NULL;
205 dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
206 skb_free_datagram(rqstp->rq_sock->sk_sk, skb);
209 rqstp->rq_deferred = NULL;
215 * Queue up a socket with data pending. If there are idle nfsd
216 * processes, wake 'em up.
219 void svc_xprt_enqueue(struct svc_xprt *xprt)
221 struct svc_serv *serv = xprt->xpt_server;
222 struct svc_pool *pool;
223 struct svc_rqst *rqstp;
226 if (!(xprt->xpt_flags &
227 ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
229 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
233 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
236 spin_lock_bh(&pool->sp_lock);
238 if (!list_empty(&pool->sp_threads) &&
239 !list_empty(&pool->sp_sockets))
242 "threads and transports both waiting??\n");
244 if (test_bit(XPT_DEAD, &xprt->xpt_flags)) {
245 /* Don't enqueue dead sockets */
246 dprintk("svc: transport %p is dead, not enqueued\n", xprt);
250 /* Mark socket as busy. It will remain in this state until the
251 * server has processed all pending data and put the socket back
252 * on the idle list. We update XPT_BUSY atomically because
253 * it also guards against trying to enqueue the svc_sock twice.
255 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
256 /* Don't enqueue socket while already enqueued */
257 dprintk("svc: transport %p busy, not enqueued\n", xprt);
260 BUG_ON(xprt->xpt_pool != NULL);
261 xprt->xpt_pool = pool;
263 /* Handle pending connection */
264 if (test_bit(XPT_CONN, &xprt->xpt_flags))
267 /* Handle close in-progress */
268 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
271 /* Check if we have space to reply to a request */
272 if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
273 /* Don't enqueue while not enough space for reply */
274 dprintk("svc: no write space, transport %p not enqueued\n",
276 xprt->xpt_pool = NULL;
277 clear_bit(XPT_BUSY, &xprt->xpt_flags);
282 if (!list_empty(&pool->sp_threads)) {
283 rqstp = list_entry(pool->sp_threads.next,
286 dprintk("svc: transport %p served by daemon %p\n",
288 svc_thread_dequeue(pool, rqstp);
291 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
292 rqstp, rqstp->rq_xprt);
293 rqstp->rq_xprt = xprt;
295 rqstp->rq_reserved = serv->sv_max_mesg;
296 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
297 BUG_ON(xprt->xpt_pool != pool);
298 wake_up(&rqstp->rq_wait);
300 dprintk("svc: transport %p put into queue\n", xprt);
301 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
302 BUG_ON(xprt->xpt_pool != pool);
306 spin_unlock_bh(&pool->sp_lock);
308 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
311 * Dequeue the first socket. Must be called with the pool->sp_lock held.
313 static inline struct svc_sock *
314 svc_sock_dequeue(struct svc_pool *pool)
316 struct svc_sock *svsk;
318 if (list_empty(&pool->sp_sockets))
321 svsk = list_entry(pool->sp_sockets.next,
322 struct svc_sock, sk_xprt.xpt_ready);
323 list_del_init(&svsk->sk_xprt.xpt_ready);
325 dprintk("svc: socket %p dequeued, inuse=%d\n",
326 svsk->sk_sk, atomic_read(&svsk->sk_xprt.xpt_ref.refcount));
332 * svc_xprt_received conditionally queues the transport for processing
333 * by another thread. The caller must hold the XPT_BUSY bit and must
334 * not thereafter touch transport data.
336 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
337 * insufficient) data.
339 void svc_xprt_received(struct svc_xprt *xprt)
341 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
342 xprt->xpt_pool = NULL;
343 clear_bit(XPT_BUSY, &xprt->xpt_flags);
344 svc_xprt_enqueue(xprt);
346 EXPORT_SYMBOL_GPL(svc_xprt_received);
349 * svc_reserve - change the space reserved for the reply to a request.
350 * @rqstp: The request in question
351 * @space: new max space to reserve
353 * Each request reserves some space on the output queue of the socket
354 * to make sure the reply fits. This function reduces that reserved
355 * space to be the amount of space used already, plus @space.
358 void svc_reserve(struct svc_rqst *rqstp, int space)
360 space += rqstp->rq_res.head[0].iov_len;
362 if (space < rqstp->rq_reserved) {
363 struct svc_xprt *xprt = rqstp->rq_xprt;
364 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
365 rqstp->rq_reserved = space;
367 svc_xprt_enqueue(xprt);
372 svc_sock_release(struct svc_rqst *rqstp)
374 struct svc_sock *svsk = rqstp->rq_sock;
376 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
378 svc_free_res_pages(rqstp);
379 rqstp->rq_res.page_len = 0;
380 rqstp->rq_res.page_base = 0;
383 /* Reset response buffer and release
385 * But first, check that enough space was reserved
386 * for the reply, otherwise we have a bug!
388 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
389 printk(KERN_ERR "RPC request reserved %d but used %d\n",
393 rqstp->rq_res.head[0].iov_len = 0;
394 svc_reserve(rqstp, 0);
395 rqstp->rq_sock = NULL;
397 svc_xprt_put(&svsk->sk_xprt);
401 * External function to wake up a server waiting for data
402 * This really only makes sense for services like lockd
403 * which have exactly one thread anyway.
406 svc_wake_up(struct svc_serv *serv)
408 struct svc_rqst *rqstp;
410 struct svc_pool *pool;
412 for (i = 0; i < serv->sv_nrpools; i++) {
413 pool = &serv->sv_pools[i];
415 spin_lock_bh(&pool->sp_lock);
416 if (!list_empty(&pool->sp_threads)) {
417 rqstp = list_entry(pool->sp_threads.next,
420 dprintk("svc: daemon %p woken up.\n", rqstp);
422 svc_thread_dequeue(pool, rqstp);
423 rqstp->rq_sock = NULL;
425 wake_up(&rqstp->rq_wait);
427 spin_unlock_bh(&pool->sp_lock);
431 union svc_pktinfo_u {
432 struct in_pktinfo pkti;
433 struct in6_pktinfo pkti6;
435 #define SVC_PKTINFO_SPACE \
436 CMSG_SPACE(sizeof(union svc_pktinfo_u))
438 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
440 switch (rqstp->rq_sock->sk_sk->sk_family) {
442 struct in_pktinfo *pki = CMSG_DATA(cmh);
444 cmh->cmsg_level = SOL_IP;
445 cmh->cmsg_type = IP_PKTINFO;
446 pki->ipi_ifindex = 0;
447 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr.addr.s_addr;
448 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
453 struct in6_pktinfo *pki = CMSG_DATA(cmh);
455 cmh->cmsg_level = SOL_IPV6;
456 cmh->cmsg_type = IPV6_PKTINFO;
457 pki->ipi6_ifindex = 0;
458 ipv6_addr_copy(&pki->ipi6_addr,
459 &rqstp->rq_daddr.addr6);
460 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
468 * Generic sendto routine
471 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
473 struct svc_sock *svsk = rqstp->rq_sock;
474 struct socket *sock = svsk->sk_sock;
478 long all[SVC_PKTINFO_SPACE / sizeof(long)];
480 struct cmsghdr *cmh = &buffer.hdr;
484 struct page **ppage = xdr->pages;
485 size_t base = xdr->page_base;
486 unsigned int pglen = xdr->page_len;
487 unsigned int flags = MSG_MORE;
488 char buf[RPC_MAX_ADDRBUFLEN];
492 if (rqstp->rq_prot == IPPROTO_UDP) {
493 struct msghdr msg = {
494 .msg_name = &rqstp->rq_addr,
495 .msg_namelen = rqstp->rq_addrlen,
497 .msg_controllen = sizeof(buffer),
498 .msg_flags = MSG_MORE,
501 svc_set_cmsg_data(rqstp, cmh);
503 if (sock_sendmsg(sock, &msg, 0) < 0)
508 if (slen == xdr->head[0].iov_len)
510 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0,
511 xdr->head[0].iov_len, flags);
512 if (len != xdr->head[0].iov_len)
514 slen -= xdr->head[0].iov_len;
519 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
523 result = kernel_sendpage(sock, *ppage, base, size, flags);
530 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
535 if (xdr->tail[0].iov_len) {
536 result = kernel_sendpage(sock, rqstp->rq_respages[0],
537 ((unsigned long)xdr->tail[0].iov_base)
539 xdr->tail[0].iov_len, 0);
545 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
546 rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len,
547 xdr->len, len, svc_print_addr(rqstp, buf, sizeof(buf)));
553 * Report socket names for nfsdfs
555 static int one_sock_name(char *buf, struct svc_sock *svsk)
559 switch(svsk->sk_sk->sk_family) {
561 len = sprintf(buf, "ipv4 %s %u.%u.%u.%u %d\n",
562 svsk->sk_sk->sk_protocol==IPPROTO_UDP?
564 NIPQUAD(inet_sk(svsk->sk_sk)->rcv_saddr),
565 inet_sk(svsk->sk_sk)->num);
568 len = sprintf(buf, "*unknown-%d*\n",
569 svsk->sk_sk->sk_family);
575 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
577 struct svc_sock *svsk, *closesk = NULL;
582 spin_lock_bh(&serv->sv_lock);
583 list_for_each_entry(svsk, &serv->sv_permsocks, sk_xprt.xpt_list) {
584 int onelen = one_sock_name(buf+len, svsk);
585 if (toclose && strcmp(toclose, buf+len) == 0)
590 spin_unlock_bh(&serv->sv_lock);
592 /* Should unregister with portmap, but you cannot
593 * unregister just one protocol...
595 svc_close_xprt(&closesk->sk_xprt);
600 EXPORT_SYMBOL(svc_sock_names);
603 * Check input queue length
606 svc_recv_available(struct svc_sock *svsk)
608 struct socket *sock = svsk->sk_sock;
611 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
613 return (err >= 0)? avail : err;
617 * Generic recvfrom routine.
620 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
622 struct svc_sock *svsk = rqstp->rq_sock;
623 struct msghdr msg = {
624 .msg_flags = MSG_DONTWAIT,
628 len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
631 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
632 svsk, iov[0].iov_base, iov[0].iov_len, len);
637 * Set socket snd and rcv buffer lengths
640 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
644 oldfs = get_fs(); set_fs(KERNEL_DS);
645 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
646 (char*)&snd, sizeof(snd));
647 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
648 (char*)&rcv, sizeof(rcv));
650 /* sock_setsockopt limits use to sysctl_?mem_max,
651 * which isn't acceptable. Until that is made conditional
652 * on not having CAP_SYS_RESOURCE or similar, we go direct...
653 * DaveM said I could!
656 sock->sk->sk_sndbuf = snd * 2;
657 sock->sk->sk_rcvbuf = rcv * 2;
658 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
659 release_sock(sock->sk);
663 * INET callback when data has been received on the socket.
666 svc_udp_data_ready(struct sock *sk, int count)
668 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
671 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
673 test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
674 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
675 svc_xprt_enqueue(&svsk->sk_xprt);
677 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
678 wake_up_interruptible(sk->sk_sleep);
682 * INET callback when space is newly available on the socket.
685 svc_write_space(struct sock *sk)
687 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
690 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
691 svsk, sk, test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
692 svc_xprt_enqueue(&svsk->sk_xprt);
695 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
696 dprintk("RPC svc_write_space: someone sleeping on %p\n",
698 wake_up_interruptible(sk->sk_sleep);
703 * Copy the UDP datagram's destination address to the rqstp structure.
704 * The 'destination' address in this case is the address to which the
705 * peer sent the datagram, i.e. our local address. For multihomed
706 * hosts, this can change from msg to msg. Note that only the IP
707 * address changes, the port number should remain the same.
709 static void svc_udp_get_dest_address(struct svc_rqst *rqstp,
712 switch (rqstp->rq_sock->sk_sk->sk_family) {
714 struct in_pktinfo *pki = CMSG_DATA(cmh);
715 rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
719 struct in6_pktinfo *pki = CMSG_DATA(cmh);
720 ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
727 * Receive a datagram from a UDP socket.
730 svc_udp_recvfrom(struct svc_rqst *rqstp)
732 struct svc_sock *svsk = rqstp->rq_sock;
733 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
737 long all[SVC_PKTINFO_SPACE / sizeof(long)];
739 struct cmsghdr *cmh = &buffer.hdr;
741 struct msghdr msg = {
742 .msg_name = svc_addr(rqstp),
744 .msg_controllen = sizeof(buffer),
745 .msg_flags = MSG_DONTWAIT,
748 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
749 /* udp sockets need large rcvbuf as all pending
750 * requests are still in that buffer. sndbuf must
751 * also be large enough that there is enough space
752 * for one reply per thread. We count all threads
753 * rather than threads in a particular pool, which
754 * provides an upper bound on the number of threads
755 * which will access the socket.
757 svc_sock_setbufsize(svsk->sk_sock,
758 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
759 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
761 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
763 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
764 0, 0, MSG_PEEK | MSG_DONTWAIT);
766 skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err);
769 if (err != -EAGAIN) {
770 /* possibly an icmp error */
771 dprintk("svc: recvfrom returned error %d\n", -err);
772 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
774 svc_xprt_received(&svsk->sk_xprt);
777 len = svc_addr_len(svc_addr(rqstp));
780 rqstp->rq_addrlen = len;
781 if (skb->tstamp.tv64 == 0) {
782 skb->tstamp = ktime_get_real();
783 /* Don't enable netstamp, sunrpc doesn't
784 need that much accuracy */
786 svsk->sk_sk->sk_stamp = skb->tstamp;
787 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* there may be more data... */
790 * Maybe more packets - kick another thread ASAP.
792 svc_xprt_received(&svsk->sk_xprt);
794 len = skb->len - sizeof(struct udphdr);
795 rqstp->rq_arg.len = len;
797 rqstp->rq_prot = IPPROTO_UDP;
799 if (cmh->cmsg_level != IPPROTO_IP ||
800 cmh->cmsg_type != IP_PKTINFO) {
802 printk("rpcsvc: received unknown control message:"
804 cmh->cmsg_level, cmh->cmsg_type);
805 skb_free_datagram(svsk->sk_sk, skb);
808 svc_udp_get_dest_address(rqstp, cmh);
810 if (skb_is_nonlinear(skb)) {
811 /* we have to copy */
813 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
816 skb_free_datagram(svsk->sk_sk, skb);
820 skb_free_datagram(svsk->sk_sk, skb);
822 /* we can use it in-place */
823 rqstp->rq_arg.head[0].iov_base = skb->data + sizeof(struct udphdr);
824 rqstp->rq_arg.head[0].iov_len = len;
825 if (skb_checksum_complete(skb)) {
826 skb_free_datagram(svsk->sk_sk, skb);
829 rqstp->rq_xprt_ctxt = skb;
832 rqstp->rq_arg.page_base = 0;
833 if (len <= rqstp->rq_arg.head[0].iov_len) {
834 rqstp->rq_arg.head[0].iov_len = len;
835 rqstp->rq_arg.page_len = 0;
836 rqstp->rq_respages = rqstp->rq_pages+1;
838 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
839 rqstp->rq_respages = rqstp->rq_pages + 1 +
840 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
844 serv->sv_stats->netudpcnt++;
850 svc_udp_sendto(struct svc_rqst *rqstp)
854 error = svc_sendto(rqstp, &rqstp->rq_res);
855 if (error == -ECONNREFUSED)
856 /* ICMP error on earlier request. */
857 error = svc_sendto(rqstp, &rqstp->rq_res);
862 static void svc_udp_prep_reply_hdr(struct svc_rqst *rqstp)
866 static int svc_udp_has_wspace(struct svc_xprt *xprt)
868 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
869 struct svc_serv *serv = xprt->xpt_server;
870 unsigned long required;
873 * Set the SOCK_NOSPACE flag before checking the available
876 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
877 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
878 if (required*2 > sock_wspace(svsk->sk_sk))
880 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
884 static struct svc_xprt *svc_udp_accept(struct svc_xprt *xprt)
890 static struct svc_xprt *svc_udp_create(struct svc_serv *serv,
891 struct sockaddr *sa, int salen,
894 return svc_create_socket(serv, IPPROTO_UDP, sa, salen, flags);
897 static struct svc_xprt_ops svc_udp_ops = {
898 .xpo_create = svc_udp_create,
899 .xpo_recvfrom = svc_udp_recvfrom,
900 .xpo_sendto = svc_udp_sendto,
901 .xpo_release_rqst = svc_release_skb,
902 .xpo_detach = svc_sock_detach,
903 .xpo_free = svc_sock_free,
904 .xpo_prep_reply_hdr = svc_udp_prep_reply_hdr,
905 .xpo_has_wspace = svc_udp_has_wspace,
906 .xpo_accept = svc_udp_accept,
909 static struct svc_xprt_class svc_udp_class = {
911 .xcl_owner = THIS_MODULE,
912 .xcl_ops = &svc_udp_ops,
913 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
916 static void svc_udp_init(struct svc_sock *svsk, struct svc_serv *serv)
921 svc_xprt_init(&svc_udp_class, &svsk->sk_xprt, serv);
922 clear_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
923 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
924 svsk->sk_sk->sk_write_space = svc_write_space;
926 /* initialise setting must have enough space to
927 * receive and respond to one request.
928 * svc_udp_recvfrom will re-adjust if necessary
930 svc_sock_setbufsize(svsk->sk_sock,
931 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
932 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
934 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* might have come in before data_ready set up */
935 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
939 /* make sure we get destination address info */
940 svsk->sk_sock->ops->setsockopt(svsk->sk_sock, IPPROTO_IP, IP_PKTINFO,
941 (char __user *)&one, sizeof(one));
946 * A data_ready event on a listening socket means there's a connection
947 * pending. Do not use state_change as a substitute for it.
950 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
952 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
954 dprintk("svc: socket %p TCP (listen) state change %d\n",
958 * This callback may called twice when a new connection
959 * is established as a child socket inherits everything
960 * from a parent LISTEN socket.
961 * 1) data_ready method of the parent socket will be called
962 * when one of child sockets become ESTABLISHED.
963 * 2) data_ready method of the child socket may be called
964 * when it receives data before the socket is accepted.
965 * In case of 2, we should ignore it silently.
967 if (sk->sk_state == TCP_LISTEN) {
969 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
970 svc_xprt_enqueue(&svsk->sk_xprt);
972 printk("svc: socket %p: no user data\n", sk);
975 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
976 wake_up_interruptible_all(sk->sk_sleep);
980 * A state change on a connected socket means it's dying or dead.
983 svc_tcp_state_change(struct sock *sk)
985 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
987 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
988 sk, sk->sk_state, sk->sk_user_data);
991 printk("svc: socket %p: no user data\n", sk);
993 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
994 svc_xprt_enqueue(&svsk->sk_xprt);
996 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
997 wake_up_interruptible_all(sk->sk_sleep);
1001 svc_tcp_data_ready(struct sock *sk, int count)
1003 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
1005 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
1006 sk, sk->sk_user_data);
1008 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1009 svc_xprt_enqueue(&svsk->sk_xprt);
1011 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1012 wake_up_interruptible(sk->sk_sleep);
1015 static inline int svc_port_is_privileged(struct sockaddr *sin)
1017 switch (sin->sa_family) {
1019 return ntohs(((struct sockaddr_in *)sin)->sin_port)
1022 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
1030 * Accept a TCP connection
1032 static struct svc_xprt *svc_tcp_accept(struct svc_xprt *xprt)
1034 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1035 struct sockaddr_storage addr;
1036 struct sockaddr *sin = (struct sockaddr *) &addr;
1037 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1038 struct socket *sock = svsk->sk_sock;
1039 struct socket *newsock;
1040 struct svc_sock *newsvsk;
1042 char buf[RPC_MAX_ADDRBUFLEN];
1044 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
1048 clear_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1049 err = kernel_accept(sock, &newsock, O_NONBLOCK);
1052 printk(KERN_WARNING "%s: no more sockets!\n",
1054 else if (err != -EAGAIN && net_ratelimit())
1055 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
1056 serv->sv_name, -err);
1059 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1061 err = kernel_getpeername(newsock, sin, &slen);
1063 if (net_ratelimit())
1064 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
1065 serv->sv_name, -err);
1066 goto failed; /* aborted connection or whatever */
1069 /* Ideally, we would want to reject connections from unauthorized
1070 * hosts here, but when we get encryption, the IP of the host won't
1071 * tell us anything. For now just warn about unpriv connections.
1073 if (!svc_port_is_privileged(sin)) {
1074 dprintk(KERN_WARNING
1075 "%s: connect from unprivileged port: %s\n",
1077 __svc_print_addr(sin, buf, sizeof(buf)));
1079 dprintk("%s: connect from %s\n", serv->sv_name,
1080 __svc_print_addr(sin, buf, sizeof(buf)));
1082 /* make sure that a write doesn't block forever when
1085 newsock->sk->sk_sndtimeo = HZ*30;
1087 if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
1088 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
1090 svc_xprt_set_remote(&newsvsk->sk_xprt, sin, slen);
1091 err = kernel_getsockname(newsock, sin, &slen);
1092 if (unlikely(err < 0)) {
1093 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err);
1094 slen = offsetof(struct sockaddr, sa_data);
1096 svc_xprt_set_local(&newsvsk->sk_xprt, sin, slen);
1099 serv->sv_stats->nettcpconn++;
1101 return &newsvsk->sk_xprt;
1104 sock_release(newsock);
1109 * Receive data from a TCP socket.
1112 svc_tcp_recvfrom(struct svc_rqst *rqstp)
1114 struct svc_sock *svsk = rqstp->rq_sock;
1115 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1120 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
1121 svsk, test_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags),
1122 test_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags),
1123 test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags));
1125 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
1126 /* sndbuf needs to have room for one request
1127 * per thread, otherwise we can stall even when the
1128 * network isn't a bottleneck.
1130 * We count all threads rather than threads in a
1131 * particular pool, which provides an upper bound
1132 * on the number of threads which will access the socket.
1134 * rcvbuf just needs to be able to hold a few requests.
1135 * Normally they will be removed from the queue
1136 * as soon a a complete request arrives.
1138 svc_sock_setbufsize(svsk->sk_sock,
1139 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
1140 3 * serv->sv_max_mesg);
1142 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1144 /* Receive data. If we haven't got the record length yet, get
1145 * the next four bytes. Otherwise try to gobble up as much as
1146 * possible up to the complete record length.
1148 if (svsk->sk_tcplen < 4) {
1149 unsigned long want = 4 - svsk->sk_tcplen;
1152 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
1154 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
1156 svsk->sk_tcplen += len;
1159 dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
1161 svc_xprt_received(&svsk->sk_xprt);
1162 return -EAGAIN; /* record header not complete */
1165 svsk->sk_reclen = ntohl(svsk->sk_reclen);
1166 if (!(svsk->sk_reclen & 0x80000000)) {
1167 /* FIXME: technically, a record can be fragmented,
1168 * and non-terminal fragments will not have the top
1169 * bit set in the fragment length header.
1170 * But apparently no known nfs clients send fragmented
1172 if (net_ratelimit())
1173 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1174 " (non-terminal)\n",
1175 (unsigned long) svsk->sk_reclen);
1178 svsk->sk_reclen &= 0x7fffffff;
1179 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
1180 if (svsk->sk_reclen > serv->sv_max_mesg) {
1181 if (net_ratelimit())
1182 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1184 (unsigned long) svsk->sk_reclen);
1189 /* Check whether enough data is available */
1190 len = svc_recv_available(svsk);
1194 if (len < svsk->sk_reclen) {
1195 dprintk("svc: incomplete TCP record (%d of %d)\n",
1196 len, svsk->sk_reclen);
1197 svc_xprt_received(&svsk->sk_xprt);
1198 return -EAGAIN; /* record not complete */
1200 len = svsk->sk_reclen;
1201 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1203 vec = rqstp->rq_vec;
1204 vec[0] = rqstp->rq_arg.head[0];
1207 while (vlen < len) {
1208 vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
1209 vec[pnum].iov_len = PAGE_SIZE;
1213 rqstp->rq_respages = &rqstp->rq_pages[pnum];
1215 /* Now receive data */
1216 len = svc_recvfrom(rqstp, vec, pnum, len);
1220 dprintk("svc: TCP complete record (%d bytes)\n", len);
1221 rqstp->rq_arg.len = len;
1222 rqstp->rq_arg.page_base = 0;
1223 if (len <= rqstp->rq_arg.head[0].iov_len) {
1224 rqstp->rq_arg.head[0].iov_len = len;
1225 rqstp->rq_arg.page_len = 0;
1227 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1230 rqstp->rq_xprt_ctxt = NULL;
1231 rqstp->rq_prot = IPPROTO_TCP;
1233 /* Reset TCP read info */
1234 svsk->sk_reclen = 0;
1235 svsk->sk_tcplen = 0;
1237 svc_xprt_copy_addrs(rqstp, &svsk->sk_xprt);
1238 svc_xprt_received(&svsk->sk_xprt);
1240 serv->sv_stats->nettcpcnt++;
1245 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1249 if (len == -EAGAIN) {
1250 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1251 svc_xprt_received(&svsk->sk_xprt);
1253 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1254 svsk->sk_xprt.xpt_server->sv_name, -len);
1262 * Send out data on TCP socket.
1265 svc_tcp_sendto(struct svc_rqst *rqstp)
1267 struct xdr_buf *xbufp = &rqstp->rq_res;
1271 /* Set up the first element of the reply kvec.
1272 * Any other kvecs that may be in use have been taken
1273 * care of by the server implementation itself.
1275 reclen = htonl(0x80000000|((xbufp->len ) - 4));
1276 memcpy(xbufp->head[0].iov_base, &reclen, 4);
1278 if (test_bit(XPT_DEAD, &rqstp->rq_sock->sk_xprt.xpt_flags))
1281 sent = svc_sendto(rqstp, &rqstp->rq_res);
1282 if (sent != xbufp->len) {
1283 printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1284 rqstp->rq_sock->sk_xprt.xpt_server->sv_name,
1285 (sent<0)?"got error":"sent only",
1287 set_bit(XPT_CLOSE, &rqstp->rq_sock->sk_xprt.xpt_flags);
1288 svc_xprt_enqueue(rqstp->rq_xprt);
1295 * Setup response header. TCP has a 4B record length field.
1297 static void svc_tcp_prep_reply_hdr(struct svc_rqst *rqstp)
1299 struct kvec *resv = &rqstp->rq_res.head[0];
1301 /* tcp needs a space for the record length... */
1305 static int svc_tcp_has_wspace(struct svc_xprt *xprt)
1307 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1308 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1313 * Set the SOCK_NOSPACE flag before checking the available
1316 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1317 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
1318 wspace = sk_stream_wspace(svsk->sk_sk);
1320 if (wspace < sk_stream_min_wspace(svsk->sk_sk))
1322 if (required * 2 > wspace)
1325 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1329 static struct svc_xprt *svc_tcp_create(struct svc_serv *serv,
1330 struct sockaddr *sa, int salen,
1333 return svc_create_socket(serv, IPPROTO_TCP, sa, salen, flags);
1336 static struct svc_xprt_ops svc_tcp_ops = {
1337 .xpo_create = svc_tcp_create,
1338 .xpo_recvfrom = svc_tcp_recvfrom,
1339 .xpo_sendto = svc_tcp_sendto,
1340 .xpo_release_rqst = svc_release_skb,
1341 .xpo_detach = svc_sock_detach,
1342 .xpo_free = svc_sock_free,
1343 .xpo_prep_reply_hdr = svc_tcp_prep_reply_hdr,
1344 .xpo_has_wspace = svc_tcp_has_wspace,
1345 .xpo_accept = svc_tcp_accept,
1348 static struct svc_xprt_class svc_tcp_class = {
1350 .xcl_owner = THIS_MODULE,
1351 .xcl_ops = &svc_tcp_ops,
1352 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1355 void svc_init_xprt_sock(void)
1357 svc_reg_xprt_class(&svc_tcp_class);
1358 svc_reg_xprt_class(&svc_udp_class);
1361 void svc_cleanup_xprt_sock(void)
1363 svc_unreg_xprt_class(&svc_tcp_class);
1364 svc_unreg_xprt_class(&svc_udp_class);
1367 static void svc_tcp_init(struct svc_sock *svsk, struct svc_serv *serv)
1369 struct sock *sk = svsk->sk_sk;
1370 struct tcp_sock *tp = tcp_sk(sk);
1372 svc_xprt_init(&svc_tcp_class, &svsk->sk_xprt, serv);
1373 set_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
1374 if (sk->sk_state == TCP_LISTEN) {
1375 dprintk("setting up TCP socket for listening\n");
1376 set_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags);
1377 sk->sk_data_ready = svc_tcp_listen_data_ready;
1378 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1380 dprintk("setting up TCP socket for reading\n");
1381 sk->sk_state_change = svc_tcp_state_change;
1382 sk->sk_data_ready = svc_tcp_data_ready;
1383 sk->sk_write_space = svc_write_space;
1385 svsk->sk_reclen = 0;
1386 svsk->sk_tcplen = 0;
1388 tp->nonagle = 1; /* disable Nagle's algorithm */
1390 /* initialise setting must have enough space to
1391 * receive and respond to one request.
1392 * svc_tcp_recvfrom will re-adjust if necessary
1394 svc_sock_setbufsize(svsk->sk_sock,
1395 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
1396 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
1398 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1399 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1400 if (sk->sk_state != TCP_ESTABLISHED)
1401 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1406 svc_sock_update_bufs(struct svc_serv *serv)
1409 * The number of server threads has changed. Update
1410 * rcvbuf and sndbuf accordingly on all sockets
1412 struct list_head *le;
1414 spin_lock_bh(&serv->sv_lock);
1415 list_for_each(le, &serv->sv_permsocks) {
1416 struct svc_sock *svsk =
1417 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1418 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1420 list_for_each(le, &serv->sv_tempsocks) {
1421 struct svc_sock *svsk =
1422 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1423 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1425 spin_unlock_bh(&serv->sv_lock);
1429 * Make sure that we don't have too many active connections. If we
1430 * have, something must be dropped.
1432 * There's no point in trying to do random drop here for DoS
1433 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
1434 * attacker can easily beat that.
1436 * The only somewhat efficient mechanism would be if drop old
1437 * connections from the same IP first. But right now we don't even
1438 * record the client IP in svc_sock.
1440 static void svc_check_conn_limits(struct svc_serv *serv)
1442 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
1443 struct svc_sock *svsk = NULL;
1444 spin_lock_bh(&serv->sv_lock);
1445 if (!list_empty(&serv->sv_tempsocks)) {
1446 if (net_ratelimit()) {
1447 /* Try to help the admin */
1448 printk(KERN_NOTICE "%s: too many open TCP "
1449 "sockets, consider increasing the "
1450 "number of nfsd threads\n",
1454 * Always select the oldest socket. It's not fair,
1457 svsk = list_entry(serv->sv_tempsocks.prev,
1460 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1461 svc_xprt_get(&svsk->sk_xprt);
1463 spin_unlock_bh(&serv->sv_lock);
1466 svc_xprt_enqueue(&svsk->sk_xprt);
1467 svc_xprt_put(&svsk->sk_xprt);
1473 * Receive the next request on any socket. This code is carefully
1474 * organised not to touch any cachelines in the shared svc_serv
1475 * structure, only cachelines in the local svc_pool.
1478 svc_recv(struct svc_rqst *rqstp, long timeout)
1480 struct svc_sock *svsk = NULL;
1481 struct svc_serv *serv = rqstp->rq_server;
1482 struct svc_pool *pool = rqstp->rq_pool;
1485 struct xdr_buf *arg;
1486 DECLARE_WAITQUEUE(wait, current);
1488 dprintk("svc: server %p waiting for data (to = %ld)\n",
1493 "svc_recv: service %p, socket not NULL!\n",
1495 if (waitqueue_active(&rqstp->rq_wait))
1497 "svc_recv: service %p, wait queue active!\n",
1501 /* now allocate needed pages. If we get a failure, sleep briefly */
1502 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
1503 for (i=0; i < pages ; i++)
1504 while (rqstp->rq_pages[i] == NULL) {
1505 struct page *p = alloc_page(GFP_KERNEL);
1507 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1508 rqstp->rq_pages[i] = p;
1510 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
1511 BUG_ON(pages >= RPCSVC_MAXPAGES);
1513 /* Make arg->head point to first page and arg->pages point to rest */
1514 arg = &rqstp->rq_arg;
1515 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
1516 arg->head[0].iov_len = PAGE_SIZE;
1517 arg->pages = rqstp->rq_pages + 1;
1519 /* save at least one page for response */
1520 arg->page_len = (pages-2)*PAGE_SIZE;
1521 arg->len = (pages-1)*PAGE_SIZE;
1522 arg->tail[0].iov_len = 0;
1529 spin_lock_bh(&pool->sp_lock);
1530 if ((svsk = svc_sock_dequeue(pool)) != NULL) {
1531 rqstp->rq_sock = svsk;
1532 svc_xprt_get(&svsk->sk_xprt);
1533 rqstp->rq_reserved = serv->sv_max_mesg;
1534 atomic_add(rqstp->rq_reserved, &svsk->sk_xprt.xpt_reserved);
1536 /* No data pending. Go to sleep */
1537 svc_thread_enqueue(pool, rqstp);
1540 * We have to be able to interrupt this wait
1541 * to bring down the daemons ...
1543 set_current_state(TASK_INTERRUPTIBLE);
1544 add_wait_queue(&rqstp->rq_wait, &wait);
1545 spin_unlock_bh(&pool->sp_lock);
1547 schedule_timeout(timeout);
1551 spin_lock_bh(&pool->sp_lock);
1552 remove_wait_queue(&rqstp->rq_wait, &wait);
1554 if (!(svsk = rqstp->rq_sock)) {
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, &svsk->sk_xprt.xpt_flags)) {
1565 dprintk("svc_recv: found XPT_CLOSE\n");
1566 svc_delete_xprt(&svsk->sk_xprt);
1567 } else if (test_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags)) {
1568 struct svc_xprt *newxpt;
1569 newxpt = svsk->sk_xprt.xpt_ops->xpo_accept(&svsk->sk_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(svsk->sk_xprt.xpt_server);
1577 svc_xprt_received(newxpt);
1579 svc_xprt_received(&svsk->sk_xprt);
1581 dprintk("svc: server %p, pool %u, socket %p, inuse=%d\n",
1582 rqstp, pool->sp_id, svsk,
1583 atomic_read(&svsk->sk_xprt.xpt_ref.refcount));
1584 rqstp->rq_deferred = svc_deferred_dequeue(&svsk->sk_xprt);
1585 if (rqstp->rq_deferred) {
1586 svc_xprt_received(&svsk->sk_xprt);
1587 len = svc_deferred_recv(rqstp);
1589 len = svsk->sk_xprt.xpt_ops->xpo_recvfrom(rqstp);
1590 dprintk("svc: got len=%d\n", len);
1593 /* No data, incomplete (TCP) read, or accept() */
1594 if (len == 0 || len == -EAGAIN) {
1595 rqstp->rq_res.len = 0;
1596 svc_sock_release(rqstp);
1599 clear_bit(XPT_OLD, &svsk->sk_xprt.xpt_flags);
1601 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
1602 rqstp->rq_chandle.defer = svc_defer;
1605 serv->sv_stats->netcnt++;
1613 svc_drop(struct svc_rqst *rqstp)
1615 dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1616 svc_sock_release(rqstp);
1620 * Return reply to client.
1623 svc_send(struct svc_rqst *rqstp)
1625 struct svc_xprt *xprt;
1629 xprt = rqstp->rq_xprt;
1633 /* release the receive skb before sending the reply */
1634 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
1636 /* calculate over-all length */
1637 xb = & rqstp->rq_res;
1638 xb->len = xb->head[0].iov_len +
1640 xb->tail[0].iov_len;
1642 /* Grab mutex to serialize outgoing data. */
1643 mutex_lock(&xprt->xpt_mutex);
1644 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
1647 len = xprt->xpt_ops->xpo_sendto(rqstp);
1648 mutex_unlock(&xprt->xpt_mutex);
1649 svc_sock_release(rqstp);
1651 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1657 * Timer function to close old temporary sockets, using
1658 * a mark-and-sweep algorithm.
1661 svc_age_temp_sockets(unsigned long closure)
1663 struct svc_serv *serv = (struct svc_serv *)closure;
1664 struct svc_sock *svsk;
1665 struct list_head *le, *next;
1666 LIST_HEAD(to_be_aged);
1668 dprintk("svc_age_temp_sockets\n");
1670 if (!spin_trylock_bh(&serv->sv_lock)) {
1671 /* busy, try again 1 sec later */
1672 dprintk("svc_age_temp_sockets: busy\n");
1673 mod_timer(&serv->sv_temptimer, jiffies + HZ);
1677 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1678 svsk = list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1680 if (!test_and_set_bit(XPT_OLD, &svsk->sk_xprt.xpt_flags))
1682 if (atomic_read(&svsk->sk_xprt.xpt_ref.refcount) > 1
1683 || test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags))
1685 svc_xprt_get(&svsk->sk_xprt);
1686 list_move(le, &to_be_aged);
1687 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1688 set_bit(XPT_DETACHED, &svsk->sk_xprt.xpt_flags);
1690 spin_unlock_bh(&serv->sv_lock);
1692 while (!list_empty(&to_be_aged)) {
1693 le = to_be_aged.next;
1694 /* fiddling the sk_xprt.xpt_list node is safe 'cos we're XPT_DETACHED */
1696 svsk = list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1698 dprintk("queuing svsk %p for closing\n", svsk);
1700 /* a thread will dequeue and close it soon */
1701 svc_xprt_enqueue(&svsk->sk_xprt);
1702 svc_xprt_put(&svsk->sk_xprt);
1705 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
1709 * Initialize socket for RPC use and create svc_sock struct
1710 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1712 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1713 struct socket *sock,
1714 int *errp, int flags)
1716 struct svc_sock *svsk;
1718 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1719 int is_temporary = flags & SVC_SOCK_TEMPORARY;
1721 dprintk("svc: svc_setup_socket %p\n", sock);
1722 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1729 /* Register socket with portmapper */
1730 if (*errp >= 0 && pmap_register)
1731 *errp = svc_register(serv, inet->sk_protocol,
1732 ntohs(inet_sk(inet)->sport));
1739 set_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags);
1740 inet->sk_user_data = svsk;
1741 svsk->sk_sock = sock;
1743 svsk->sk_ostate = inet->sk_state_change;
1744 svsk->sk_odata = inet->sk_data_ready;
1745 svsk->sk_owspace = inet->sk_write_space;
1747 /* Initialize the socket */
1748 if (sock->type == SOCK_DGRAM)
1749 svc_udp_init(svsk, serv);
1751 svc_tcp_init(svsk, serv);
1753 spin_lock_bh(&serv->sv_lock);
1755 set_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1756 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_tempsocks);
1758 if (serv->sv_temptimer.function == NULL) {
1759 /* setup timer to age temp sockets */
1760 setup_timer(&serv->sv_temptimer, svc_age_temp_sockets,
1761 (unsigned long)serv);
1762 mod_timer(&serv->sv_temptimer,
1763 jiffies + svc_conn_age_period * HZ);
1766 clear_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1767 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_permsocks);
1769 spin_unlock_bh(&serv->sv_lock);
1771 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1777 int svc_addsock(struct svc_serv *serv,
1783 struct socket *so = sockfd_lookup(fd, &err);
1784 struct svc_sock *svsk = NULL;
1788 if (so->sk->sk_family != AF_INET)
1789 err = -EAFNOSUPPORT;
1790 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1791 so->sk->sk_protocol != IPPROTO_UDP)
1792 err = -EPROTONOSUPPORT;
1793 else if (so->state > SS_UNCONNECTED)
1796 svsk = svc_setup_socket(serv, so, &err, SVC_SOCK_DEFAULTS);
1798 struct sockaddr_storage addr;
1799 struct sockaddr *sin = (struct sockaddr *)&addr;
1801 if (kernel_getsockname(svsk->sk_sock, sin, &salen) == 0)
1802 svc_xprt_set_local(&svsk->sk_xprt, sin, salen);
1803 svc_xprt_received(&svsk->sk_xprt);
1811 if (proto) *proto = so->sk->sk_protocol;
1812 return one_sock_name(name_return, svsk);
1814 EXPORT_SYMBOL_GPL(svc_addsock);
1817 * Create socket for RPC service.
1819 static struct svc_xprt *svc_create_socket(struct svc_serv *serv,
1821 struct sockaddr *sin, int len,
1824 struct svc_sock *svsk;
1825 struct socket *sock;
1828 char buf[RPC_MAX_ADDRBUFLEN];
1829 struct sockaddr_storage addr;
1830 struct sockaddr *newsin = (struct sockaddr *)&addr;
1833 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1834 serv->sv_program->pg_name, protocol,
1835 __svc_print_addr(sin, buf, sizeof(buf)));
1837 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1838 printk(KERN_WARNING "svc: only UDP and TCP "
1839 "sockets supported\n");
1840 return ERR_PTR(-EINVAL);
1842 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1844 error = sock_create_kern(sin->sa_family, type, protocol, &sock);
1846 return ERR_PTR(error);
1848 svc_reclassify_socket(sock);
1850 if (type == SOCK_STREAM)
1851 sock->sk->sk_reuse = 1; /* allow address reuse */
1852 error = kernel_bind(sock, sin, len);
1857 error = kernel_getsockname(sock, newsin, &newlen);
1861 if (protocol == IPPROTO_TCP) {
1862 if ((error = kernel_listen(sock, 64)) < 0)
1866 if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1867 svc_xprt_set_local(&svsk->sk_xprt, newsin, newlen);
1868 svc_xprt_received(&svsk->sk_xprt);
1869 return (struct svc_xprt *)svsk;
1873 dprintk("svc: svc_create_socket error = %d\n", -error);
1875 return ERR_PTR(error);
1879 * Detach the svc_sock from the socket so that no
1880 * more callbacks occur.
1882 static void svc_sock_detach(struct svc_xprt *xprt)
1884 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1885 struct sock *sk = svsk->sk_sk;
1887 dprintk("svc: svc_sock_detach(%p)\n", svsk);
1889 /* put back the old socket callbacks */
1890 sk->sk_state_change = svsk->sk_ostate;
1891 sk->sk_data_ready = svsk->sk_odata;
1892 sk->sk_write_space = svsk->sk_owspace;
1896 * Free the svc_sock's socket resources and the svc_sock itself.
1898 static void svc_sock_free(struct svc_xprt *xprt)
1900 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1901 dprintk("svc: svc_sock_free(%p)\n", svsk);
1903 if (svsk->sk_sock->file)
1904 sockfd_put(svsk->sk_sock);
1906 sock_release(svsk->sk_sock);
1911 * Remove a dead transport
1913 static void svc_delete_xprt(struct svc_xprt *xprt)
1915 struct svc_serv *serv = xprt->xpt_server;
1917 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1918 xprt->xpt_ops->xpo_detach(xprt);
1920 spin_lock_bh(&serv->sv_lock);
1921 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
1922 list_del_init(&xprt->xpt_list);
1924 * We used to delete the transport from whichever list
1925 * it's sk_xprt.xpt_ready node was on, but we don't actually
1926 * need to. This is because the only time we're called
1927 * while still attached to a queue, the queue itself
1928 * is about to be destroyed (in svc_destroy).
1930 if (!test_and_set_bit(XPT_DEAD, &xprt->xpt_flags)) {
1931 BUG_ON(atomic_read(&xprt->xpt_ref.refcount) < 2);
1932 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1936 spin_unlock_bh(&serv->sv_lock);
1939 static void svc_close_xprt(struct svc_xprt *xprt)
1941 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1942 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1943 /* someone else will have to effect the close */
1947 svc_delete_xprt(xprt);
1948 clear_bit(XPT_BUSY, &xprt->xpt_flags);
1952 void svc_close_all(struct list_head *xprt_list)
1954 struct svc_xprt *xprt;
1955 struct svc_xprt *tmp;
1957 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
1958 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1959 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
1960 /* Waiting to be processed, but no threads left,
1961 * So just remove it from the waiting list
1963 list_del_init(&xprt->xpt_ready);
1964 clear_bit(XPT_BUSY, &xprt->xpt_flags);
1966 svc_close_xprt(xprt);
1971 * Handle defer and revisit of requests
1974 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1976 struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1977 struct svc_xprt *xprt = dr->xprt;
1984 dprintk("revisit queued\n");
1986 spin_lock(&xprt->xpt_lock);
1987 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1988 spin_unlock(&xprt->xpt_lock);
1989 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1990 svc_xprt_enqueue(xprt);
1994 static struct cache_deferred_req *
1995 svc_defer(struct cache_req *req)
1997 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1998 int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
1999 struct svc_deferred_req *dr;
2001 if (rqstp->rq_arg.page_len)
2002 return NULL; /* if more than a page, give up FIXME */
2003 if (rqstp->rq_deferred) {
2004 dr = rqstp->rq_deferred;
2005 rqstp->rq_deferred = NULL;
2007 int skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
2008 /* FIXME maybe discard if size too large */
2009 dr = kmalloc(size, GFP_KERNEL);
2013 dr->handle.owner = rqstp->rq_server;
2014 dr->prot = rqstp->rq_prot;
2015 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
2016 dr->addrlen = rqstp->rq_addrlen;
2017 dr->daddr = rqstp->rq_daddr;
2018 dr->argslen = rqstp->rq_arg.len >> 2;
2019 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
2021 svc_xprt_get(rqstp->rq_xprt);
2022 dr->xprt = rqstp->rq_xprt;
2024 dr->handle.revisit = svc_revisit;
2029 * recv data from a deferred request into an active one
2031 static int svc_deferred_recv(struct svc_rqst *rqstp)
2033 struct svc_deferred_req *dr = rqstp->rq_deferred;
2035 rqstp->rq_arg.head[0].iov_base = dr->args;
2036 rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
2037 rqstp->rq_arg.page_len = 0;
2038 rqstp->rq_arg.len = dr->argslen<<2;
2039 rqstp->rq_prot = dr->prot;
2040 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
2041 rqstp->rq_addrlen = dr->addrlen;
2042 rqstp->rq_daddr = dr->daddr;
2043 rqstp->rq_respages = rqstp->rq_pages;
2044 return dr->argslen<<2;
2048 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
2050 struct svc_deferred_req *dr = NULL;
2052 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
2054 spin_lock(&xprt->xpt_lock);
2055 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
2056 if (!list_empty(&xprt->xpt_deferred)) {
2057 dr = list_entry(xprt->xpt_deferred.next,
2058 struct svc_deferred_req,
2060 list_del_init(&dr->handle.recent);
2061 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
2063 spin_unlock(&xprt->xpt_lock);