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_sock *svsk);
93 static int svc_deferred_recv(struct svc_rqst *rqstp);
94 static struct cache_deferred_req *svc_defer(struct cache_req *req);
95 static struct svc_xprt *svc_create_socket(struct svc_serv *, int,
96 struct sockaddr *, int, int);
98 /* apparently the "standard" is that clients close
99 * idle connections after 5 minutes, servers after
101 * http://www.connectathon.org/talks96/nfstcp.pdf
103 static int svc_conn_age_period = 6*60;
105 #ifdef CONFIG_DEBUG_LOCK_ALLOC
106 static struct lock_class_key svc_key[2];
107 static struct lock_class_key svc_slock_key[2];
109 static inline void svc_reclassify_socket(struct socket *sock)
111 struct sock *sk = sock->sk;
112 BUG_ON(sock_owned_by_user(sk));
113 switch (sk->sk_family) {
115 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFSD",
117 "sk_xprt.xpt_lock-AF_INET-NFSD",
122 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFSD",
124 "sk_xprt.xpt_lock-AF_INET6-NFSD",
133 static inline void svc_reclassify_socket(struct socket *sock)
138 static char *__svc_print_addr(struct sockaddr *addr, char *buf, size_t len)
140 switch (addr->sa_family) {
142 snprintf(buf, len, "%u.%u.%u.%u, port=%u",
143 NIPQUAD(((struct sockaddr_in *) addr)->sin_addr),
144 ntohs(((struct sockaddr_in *) addr)->sin_port));
148 snprintf(buf, len, "%x:%x:%x:%x:%x:%x:%x:%x, port=%u",
149 NIP6(((struct sockaddr_in6 *) addr)->sin6_addr),
150 ntohs(((struct sockaddr_in6 *) addr)->sin6_port));
154 snprintf(buf, len, "unknown address type: %d", addr->sa_family);
161 * svc_print_addr - Format rq_addr field for printing
162 * @rqstp: svc_rqst struct containing address to print
163 * @buf: target buffer for formatted address
164 * @len: length of target buffer
167 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
169 return __svc_print_addr(svc_addr(rqstp), buf, len);
171 EXPORT_SYMBOL_GPL(svc_print_addr);
174 * Queue up an idle server thread. Must have pool->sp_lock held.
175 * Note: this is really a stack rather than a queue, so that we only
176 * use as many different threads as we need, and the rest don't pollute
180 svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
182 list_add(&rqstp->rq_list, &pool->sp_threads);
186 * Dequeue an nfsd thread. Must have pool->sp_lock held.
189 svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
191 list_del(&rqstp->rq_list);
195 * Release an skbuff after use
197 static void svc_release_skb(struct svc_rqst *rqstp)
199 struct sk_buff *skb = rqstp->rq_xprt_ctxt;
200 struct svc_deferred_req *dr = rqstp->rq_deferred;
203 rqstp->rq_xprt_ctxt = NULL;
205 dprintk("svc: service %p, releasing skb %p\n", rqstp, skb);
206 skb_free_datagram(rqstp->rq_sock->sk_sk, skb);
209 rqstp->rq_deferred = NULL;
215 * Queue up a socket with data pending. If there are idle nfsd
216 * processes, wake 'em up.
219 void svc_xprt_enqueue(struct svc_xprt *xprt)
221 struct svc_serv *serv = xprt->xpt_server;
222 struct svc_pool *pool;
223 struct svc_rqst *rqstp;
226 if (!(xprt->xpt_flags &
227 ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
229 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
233 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
236 spin_lock_bh(&pool->sp_lock);
238 if (!list_empty(&pool->sp_threads) &&
239 !list_empty(&pool->sp_sockets))
242 "threads and transports both waiting??\n");
244 if (test_bit(XPT_DEAD, &xprt->xpt_flags)) {
245 /* Don't enqueue dead sockets */
246 dprintk("svc: transport %p is dead, not enqueued\n", xprt);
250 /* Mark socket as busy. It will remain in this state until the
251 * server has processed all pending data and put the socket back
252 * on the idle list. We update XPT_BUSY atomically because
253 * it also guards against trying to enqueue the svc_sock twice.
255 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
256 /* Don't enqueue socket while already enqueued */
257 dprintk("svc: transport %p busy, not enqueued\n", xprt);
260 BUG_ON(xprt->xpt_pool != NULL);
261 xprt->xpt_pool = pool;
263 /* Handle pending connection */
264 if (test_bit(XPT_CONN, &xprt->xpt_flags))
267 /* Handle close in-progress */
268 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
271 /* Check if we have space to reply to a request */
272 if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
273 /* Don't enqueue while not enough space for reply */
274 dprintk("svc: no write space, transport %p not enqueued\n",
276 xprt->xpt_pool = NULL;
277 clear_bit(XPT_BUSY, &xprt->xpt_flags);
282 if (!list_empty(&pool->sp_threads)) {
283 rqstp = list_entry(pool->sp_threads.next,
286 dprintk("svc: transport %p served by daemon %p\n",
288 svc_thread_dequeue(pool, rqstp);
291 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
292 rqstp, rqstp->rq_xprt);
293 rqstp->rq_xprt = xprt;
295 rqstp->rq_reserved = serv->sv_max_mesg;
296 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
297 BUG_ON(xprt->xpt_pool != pool);
298 wake_up(&rqstp->rq_wait);
300 dprintk("svc: transport %p put into queue\n", xprt);
301 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
302 BUG_ON(xprt->xpt_pool != pool);
306 spin_unlock_bh(&pool->sp_lock);
308 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
311 * Dequeue the first socket. Must be called with the pool->sp_lock held.
313 static inline struct svc_sock *
314 svc_sock_dequeue(struct svc_pool *pool)
316 struct svc_sock *svsk;
318 if (list_empty(&pool->sp_sockets))
321 svsk = list_entry(pool->sp_sockets.next,
322 struct svc_sock, sk_xprt.xpt_ready);
323 list_del_init(&svsk->sk_xprt.xpt_ready);
325 dprintk("svc: socket %p dequeued, inuse=%d\n",
326 svsk->sk_sk, atomic_read(&svsk->sk_xprt.xpt_ref.refcount));
332 * svc_xprt_received conditionally queues the transport for processing
333 * by another thread. The caller must hold the XPT_BUSY bit and must
334 * not thereafter touch transport data.
336 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
337 * insufficient) data.
339 void svc_xprt_received(struct svc_xprt *xprt)
341 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
342 xprt->xpt_pool = NULL;
343 clear_bit(XPT_BUSY, &xprt->xpt_flags);
344 svc_xprt_enqueue(xprt);
346 EXPORT_SYMBOL_GPL(svc_xprt_received);
349 * svc_reserve - change the space reserved for the reply to a request.
350 * @rqstp: The request in question
351 * @space: new max space to reserve
353 * Each request reserves some space on the output queue of the socket
354 * to make sure the reply fits. This function reduces that reserved
355 * space to be the amount of space used already, plus @space.
358 void svc_reserve(struct svc_rqst *rqstp, int space)
360 space += rqstp->rq_res.head[0].iov_len;
362 if (space < rqstp->rq_reserved) {
363 struct svc_xprt *xprt = rqstp->rq_xprt;
364 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
365 rqstp->rq_reserved = space;
367 svc_xprt_enqueue(xprt);
372 svc_sock_release(struct svc_rqst *rqstp)
374 struct svc_sock *svsk = rqstp->rq_sock;
376 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
378 svc_free_res_pages(rqstp);
379 rqstp->rq_res.page_len = 0;
380 rqstp->rq_res.page_base = 0;
383 /* Reset response buffer and release
385 * But first, check that enough space was reserved
386 * for the reply, otherwise we have a bug!
388 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
389 printk(KERN_ERR "RPC request reserved %d but used %d\n",
393 rqstp->rq_res.head[0].iov_len = 0;
394 svc_reserve(rqstp, 0);
395 rqstp->rq_sock = NULL;
397 svc_xprt_put(&svsk->sk_xprt);
401 * External function to wake up a server waiting for data
402 * This really only makes sense for services like lockd
403 * which have exactly one thread anyway.
406 svc_wake_up(struct svc_serv *serv)
408 struct svc_rqst *rqstp;
410 struct svc_pool *pool;
412 for (i = 0; i < serv->sv_nrpools; i++) {
413 pool = &serv->sv_pools[i];
415 spin_lock_bh(&pool->sp_lock);
416 if (!list_empty(&pool->sp_threads)) {
417 rqstp = list_entry(pool->sp_threads.next,
420 dprintk("svc: daemon %p woken up.\n", rqstp);
422 svc_thread_dequeue(pool, rqstp);
423 rqstp->rq_sock = NULL;
425 wake_up(&rqstp->rq_wait);
427 spin_unlock_bh(&pool->sp_lock);
431 union svc_pktinfo_u {
432 struct in_pktinfo pkti;
433 struct in6_pktinfo pkti6;
435 #define SVC_PKTINFO_SPACE \
436 CMSG_SPACE(sizeof(union svc_pktinfo_u))
438 static void svc_set_cmsg_data(struct svc_rqst *rqstp, struct cmsghdr *cmh)
440 switch (rqstp->rq_sock->sk_sk->sk_family) {
442 struct in_pktinfo *pki = CMSG_DATA(cmh);
444 cmh->cmsg_level = SOL_IP;
445 cmh->cmsg_type = IP_PKTINFO;
446 pki->ipi_ifindex = 0;
447 pki->ipi_spec_dst.s_addr = rqstp->rq_daddr.addr.s_addr;
448 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
453 struct in6_pktinfo *pki = CMSG_DATA(cmh);
455 cmh->cmsg_level = SOL_IPV6;
456 cmh->cmsg_type = IPV6_PKTINFO;
457 pki->ipi6_ifindex = 0;
458 ipv6_addr_copy(&pki->ipi6_addr,
459 &rqstp->rq_daddr.addr6);
460 cmh->cmsg_len = CMSG_LEN(sizeof(*pki));
468 * Generic sendto routine
471 svc_sendto(struct svc_rqst *rqstp, struct xdr_buf *xdr)
473 struct svc_sock *svsk = rqstp->rq_sock;
474 struct socket *sock = svsk->sk_sock;
478 long all[SVC_PKTINFO_SPACE / sizeof(long)];
480 struct cmsghdr *cmh = &buffer.hdr;
484 struct page **ppage = xdr->pages;
485 size_t base = xdr->page_base;
486 unsigned int pglen = xdr->page_len;
487 unsigned int flags = MSG_MORE;
488 char buf[RPC_MAX_ADDRBUFLEN];
492 if (rqstp->rq_prot == IPPROTO_UDP) {
493 struct msghdr msg = {
494 .msg_name = &rqstp->rq_addr,
495 .msg_namelen = rqstp->rq_addrlen,
497 .msg_controllen = sizeof(buffer),
498 .msg_flags = MSG_MORE,
501 svc_set_cmsg_data(rqstp, cmh);
503 if (sock_sendmsg(sock, &msg, 0) < 0)
508 if (slen == xdr->head[0].iov_len)
510 len = kernel_sendpage(sock, rqstp->rq_respages[0], 0,
511 xdr->head[0].iov_len, flags);
512 if (len != xdr->head[0].iov_len)
514 slen -= xdr->head[0].iov_len;
519 size = PAGE_SIZE - base < pglen ? PAGE_SIZE - base : pglen;
523 result = kernel_sendpage(sock, *ppage, base, size, flags);
530 size = PAGE_SIZE < pglen ? PAGE_SIZE : pglen;
535 if (xdr->tail[0].iov_len) {
536 result = kernel_sendpage(sock, rqstp->rq_respages[0],
537 ((unsigned long)xdr->tail[0].iov_base)
539 xdr->tail[0].iov_len, 0);
545 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
546 rqstp->rq_sock, xdr->head[0].iov_base, xdr->head[0].iov_len,
547 xdr->len, len, svc_print_addr(rqstp, buf, sizeof(buf)));
553 * Report socket names for nfsdfs
555 static int one_sock_name(char *buf, struct svc_sock *svsk)
559 switch(svsk->sk_sk->sk_family) {
561 len = sprintf(buf, "ipv4 %s %u.%u.%u.%u %d\n",
562 svsk->sk_sk->sk_protocol==IPPROTO_UDP?
564 NIPQUAD(inet_sk(svsk->sk_sk)->rcv_saddr),
565 inet_sk(svsk->sk_sk)->num);
568 len = sprintf(buf, "*unknown-%d*\n",
569 svsk->sk_sk->sk_family);
575 svc_sock_names(char *buf, struct svc_serv *serv, char *toclose)
577 struct svc_sock *svsk, *closesk = NULL;
582 spin_lock_bh(&serv->sv_lock);
583 list_for_each_entry(svsk, &serv->sv_permsocks, sk_xprt.xpt_list) {
584 int onelen = one_sock_name(buf+len, svsk);
585 if (toclose && strcmp(toclose, buf+len) == 0)
590 spin_unlock_bh(&serv->sv_lock);
592 /* Should unregister with portmap, but you cannot
593 * unregister just one protocol...
595 svc_close_xprt(&closesk->sk_xprt);
600 EXPORT_SYMBOL(svc_sock_names);
603 * Check input queue length
606 svc_recv_available(struct svc_sock *svsk)
608 struct socket *sock = svsk->sk_sock;
611 err = kernel_sock_ioctl(sock, TIOCINQ, (unsigned long) &avail);
613 return (err >= 0)? avail : err;
617 * Generic recvfrom routine.
620 svc_recvfrom(struct svc_rqst *rqstp, struct kvec *iov, int nr, int buflen)
622 struct svc_sock *svsk = rqstp->rq_sock;
623 struct msghdr msg = {
624 .msg_flags = MSG_DONTWAIT,
626 struct sockaddr *sin;
629 len = kernel_recvmsg(svsk->sk_sock, &msg, iov, nr, buflen,
632 /* sock_recvmsg doesn't fill in the name/namelen, so we must..
634 memcpy(&rqstp->rq_addr, &svsk->sk_remote, svsk->sk_remotelen);
635 rqstp->rq_addrlen = svsk->sk_remotelen;
637 /* Destination address in request is needed for binding the
638 * source address in RPC callbacks later.
640 sin = (struct sockaddr *)&svsk->sk_local;
641 switch (sin->sa_family) {
643 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
646 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
650 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
651 svsk, iov[0].iov_base, iov[0].iov_len, len);
657 * Set socket snd and rcv buffer lengths
660 svc_sock_setbufsize(struct socket *sock, unsigned int snd, unsigned int rcv)
664 oldfs = get_fs(); set_fs(KERNEL_DS);
665 sock_setsockopt(sock, SOL_SOCKET, SO_SNDBUF,
666 (char*)&snd, sizeof(snd));
667 sock_setsockopt(sock, SOL_SOCKET, SO_RCVBUF,
668 (char*)&rcv, sizeof(rcv));
670 /* sock_setsockopt limits use to sysctl_?mem_max,
671 * which isn't acceptable. Until that is made conditional
672 * on not having CAP_SYS_RESOURCE or similar, we go direct...
673 * DaveM said I could!
676 sock->sk->sk_sndbuf = snd * 2;
677 sock->sk->sk_rcvbuf = rcv * 2;
678 sock->sk->sk_userlocks |= SOCK_SNDBUF_LOCK|SOCK_RCVBUF_LOCK;
679 release_sock(sock->sk);
683 * INET callback when data has been received on the socket.
686 svc_udp_data_ready(struct sock *sk, int count)
688 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
691 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
693 test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
694 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
695 svc_xprt_enqueue(&svsk->sk_xprt);
697 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
698 wake_up_interruptible(sk->sk_sleep);
702 * INET callback when space is newly available on the socket.
705 svc_write_space(struct sock *sk)
707 struct svc_sock *svsk = (struct svc_sock *)(sk->sk_user_data);
710 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
711 svsk, sk, test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags));
712 svc_xprt_enqueue(&svsk->sk_xprt);
715 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep)) {
716 dprintk("RPC svc_write_space: someone sleeping on %p\n",
718 wake_up_interruptible(sk->sk_sleep);
722 static inline void svc_udp_get_dest_address(struct svc_rqst *rqstp,
725 switch (rqstp->rq_sock->sk_sk->sk_family) {
727 struct in_pktinfo *pki = CMSG_DATA(cmh);
728 rqstp->rq_daddr.addr.s_addr = pki->ipi_spec_dst.s_addr;
732 struct in6_pktinfo *pki = CMSG_DATA(cmh);
733 ipv6_addr_copy(&rqstp->rq_daddr.addr6, &pki->ipi6_addr);
740 * Receive a datagram from a UDP socket.
743 svc_udp_recvfrom(struct svc_rqst *rqstp)
745 struct svc_sock *svsk = rqstp->rq_sock;
746 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
750 long all[SVC_PKTINFO_SPACE / sizeof(long)];
752 struct cmsghdr *cmh = &buffer.hdr;
754 struct msghdr msg = {
755 .msg_name = svc_addr(rqstp),
757 .msg_controllen = sizeof(buffer),
758 .msg_flags = MSG_DONTWAIT,
761 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
762 /* udp sockets need large rcvbuf as all pending
763 * requests are still in that buffer. sndbuf must
764 * also be large enough that there is enough space
765 * for one reply per thread. We count all threads
766 * rather than threads in a particular pool, which
767 * provides an upper bound on the number of threads
768 * which will access the socket.
770 svc_sock_setbufsize(svsk->sk_sock,
771 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
772 (serv->sv_nrthreads+3) * serv->sv_max_mesg);
774 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
775 svc_xprt_received(&svsk->sk_xprt);
776 return svc_deferred_recv(rqstp);
779 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
781 err = kernel_recvmsg(svsk->sk_sock, &msg, NULL,
782 0, 0, MSG_PEEK | MSG_DONTWAIT);
784 skb = skb_recv_datagram(svsk->sk_sk, 0, 1, &err);
787 if (err != -EAGAIN) {
788 /* possibly an icmp error */
789 dprintk("svc: recvfrom returned error %d\n", -err);
790 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
792 svc_xprt_received(&svsk->sk_xprt);
795 rqstp->rq_addrlen = sizeof(rqstp->rq_addr);
796 if (skb->tstamp.tv64 == 0) {
797 skb->tstamp = ktime_get_real();
798 /* Don't enable netstamp, sunrpc doesn't
799 need that much accuracy */
801 svsk->sk_sk->sk_stamp = skb->tstamp;
802 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* there may be more data... */
805 * Maybe more packets - kick another thread ASAP.
807 svc_xprt_received(&svsk->sk_xprt);
809 len = skb->len - sizeof(struct udphdr);
810 rqstp->rq_arg.len = len;
812 rqstp->rq_prot = IPPROTO_UDP;
814 if (cmh->cmsg_level != IPPROTO_IP ||
815 cmh->cmsg_type != IP_PKTINFO) {
817 printk("rpcsvc: received unknown control message:"
819 cmh->cmsg_level, cmh->cmsg_type);
820 skb_free_datagram(svsk->sk_sk, skb);
823 svc_udp_get_dest_address(rqstp, cmh);
825 if (skb_is_nonlinear(skb)) {
826 /* we have to copy */
828 if (csum_partial_copy_to_xdr(&rqstp->rq_arg, skb)) {
831 skb_free_datagram(svsk->sk_sk, skb);
835 skb_free_datagram(svsk->sk_sk, skb);
837 /* we can use it in-place */
838 rqstp->rq_arg.head[0].iov_base = skb->data + sizeof(struct udphdr);
839 rqstp->rq_arg.head[0].iov_len = len;
840 if (skb_checksum_complete(skb)) {
841 skb_free_datagram(svsk->sk_sk, skb);
844 rqstp->rq_xprt_ctxt = skb;
847 rqstp->rq_arg.page_base = 0;
848 if (len <= rqstp->rq_arg.head[0].iov_len) {
849 rqstp->rq_arg.head[0].iov_len = len;
850 rqstp->rq_arg.page_len = 0;
851 rqstp->rq_respages = rqstp->rq_pages+1;
853 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
854 rqstp->rq_respages = rqstp->rq_pages + 1 +
855 DIV_ROUND_UP(rqstp->rq_arg.page_len, PAGE_SIZE);
859 serv->sv_stats->netudpcnt++;
865 svc_udp_sendto(struct svc_rqst *rqstp)
869 error = svc_sendto(rqstp, &rqstp->rq_res);
870 if (error == -ECONNREFUSED)
871 /* ICMP error on earlier request. */
872 error = svc_sendto(rqstp, &rqstp->rq_res);
877 static void svc_udp_prep_reply_hdr(struct svc_rqst *rqstp)
881 static int svc_udp_has_wspace(struct svc_xprt *xprt)
883 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
884 struct svc_serv *serv = xprt->xpt_server;
885 unsigned long required;
888 * Set the SOCK_NOSPACE flag before checking the available
891 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
892 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
893 if (required*2 > sock_wspace(svsk->sk_sk))
895 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
899 static struct svc_xprt *svc_udp_accept(struct svc_xprt *xprt)
905 static struct svc_xprt *svc_udp_create(struct svc_serv *serv,
906 struct sockaddr *sa, int salen,
909 return svc_create_socket(serv, IPPROTO_UDP, sa, salen, flags);
912 static struct svc_xprt_ops svc_udp_ops = {
913 .xpo_create = svc_udp_create,
914 .xpo_recvfrom = svc_udp_recvfrom,
915 .xpo_sendto = svc_udp_sendto,
916 .xpo_release_rqst = svc_release_skb,
917 .xpo_detach = svc_sock_detach,
918 .xpo_free = svc_sock_free,
919 .xpo_prep_reply_hdr = svc_udp_prep_reply_hdr,
920 .xpo_has_wspace = svc_udp_has_wspace,
921 .xpo_accept = svc_udp_accept,
924 static struct svc_xprt_class svc_udp_class = {
926 .xcl_owner = THIS_MODULE,
927 .xcl_ops = &svc_udp_ops,
928 .xcl_max_payload = RPCSVC_MAXPAYLOAD_UDP,
931 static void svc_udp_init(struct svc_sock *svsk, struct svc_serv *serv)
936 svc_xprt_init(&svc_udp_class, &svsk->sk_xprt, serv);
937 clear_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
938 svsk->sk_sk->sk_data_ready = svc_udp_data_ready;
939 svsk->sk_sk->sk_write_space = svc_write_space;
941 /* initialise setting must have enough space to
942 * receive and respond to one request.
943 * svc_udp_recvfrom will re-adjust if necessary
945 svc_sock_setbufsize(svsk->sk_sock,
946 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
947 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
949 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags); /* might have come in before data_ready set up */
950 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
954 /* make sure we get destination address info */
955 svsk->sk_sock->ops->setsockopt(svsk->sk_sock, IPPROTO_IP, IP_PKTINFO,
956 (char __user *)&one, sizeof(one));
961 * A data_ready event on a listening socket means there's a connection
962 * pending. Do not use state_change as a substitute for it.
965 svc_tcp_listen_data_ready(struct sock *sk, int count_unused)
967 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
969 dprintk("svc: socket %p TCP (listen) state change %d\n",
973 * This callback may called twice when a new connection
974 * is established as a child socket inherits everything
975 * from a parent LISTEN socket.
976 * 1) data_ready method of the parent socket will be called
977 * when one of child sockets become ESTABLISHED.
978 * 2) data_ready method of the child socket may be called
979 * when it receives data before the socket is accepted.
980 * In case of 2, we should ignore it silently.
982 if (sk->sk_state == TCP_LISTEN) {
984 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
985 svc_xprt_enqueue(&svsk->sk_xprt);
987 printk("svc: socket %p: no user data\n", sk);
990 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
991 wake_up_interruptible_all(sk->sk_sleep);
995 * A state change on a connected socket means it's dying or dead.
998 svc_tcp_state_change(struct sock *sk)
1000 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
1002 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
1003 sk, sk->sk_state, sk->sk_user_data);
1006 printk("svc: socket %p: no user data\n", sk);
1008 set_bit(XPT_CLOSE, &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_all(sk->sk_sleep);
1016 svc_tcp_data_ready(struct sock *sk, int count)
1018 struct svc_sock *svsk = (struct svc_sock *)sk->sk_user_data;
1020 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
1021 sk, sk->sk_user_data);
1023 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1024 svc_xprt_enqueue(&svsk->sk_xprt);
1026 if (sk->sk_sleep && waitqueue_active(sk->sk_sleep))
1027 wake_up_interruptible(sk->sk_sleep);
1030 static inline int svc_port_is_privileged(struct sockaddr *sin)
1032 switch (sin->sa_family) {
1034 return ntohs(((struct sockaddr_in *)sin)->sin_port)
1037 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
1045 * Accept a TCP connection
1047 static struct svc_xprt *svc_tcp_accept(struct svc_xprt *xprt)
1049 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1050 struct sockaddr_storage addr;
1051 struct sockaddr *sin = (struct sockaddr *) &addr;
1052 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1053 struct socket *sock = svsk->sk_sock;
1054 struct socket *newsock;
1055 struct svc_sock *newsvsk;
1057 char buf[RPC_MAX_ADDRBUFLEN];
1059 dprintk("svc: tcp_accept %p sock %p\n", svsk, sock);
1063 clear_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1064 err = kernel_accept(sock, &newsock, O_NONBLOCK);
1067 printk(KERN_WARNING "%s: no more sockets!\n",
1069 else if (err != -EAGAIN && net_ratelimit())
1070 printk(KERN_WARNING "%s: accept failed (err %d)!\n",
1071 serv->sv_name, -err);
1074 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1076 err = kernel_getpeername(newsock, sin, &slen);
1078 if (net_ratelimit())
1079 printk(KERN_WARNING "%s: peername failed (err %d)!\n",
1080 serv->sv_name, -err);
1081 goto failed; /* aborted connection or whatever */
1084 /* Ideally, we would want to reject connections from unauthorized
1085 * hosts here, but when we get encryption, the IP of the host won't
1086 * tell us anything. For now just warn about unpriv connections.
1088 if (!svc_port_is_privileged(sin)) {
1089 dprintk(KERN_WARNING
1090 "%s: connect from unprivileged port: %s\n",
1092 __svc_print_addr(sin, buf, sizeof(buf)));
1094 dprintk("%s: connect from %s\n", serv->sv_name,
1095 __svc_print_addr(sin, buf, sizeof(buf)));
1097 /* make sure that a write doesn't block forever when
1100 newsock->sk->sk_sndtimeo = HZ*30;
1102 if (!(newsvsk = svc_setup_socket(serv, newsock, &err,
1103 (SVC_SOCK_ANONYMOUS | SVC_SOCK_TEMPORARY))))
1105 memcpy(&newsvsk->sk_remote, sin, slen);
1106 newsvsk->sk_remotelen = slen;
1107 err = kernel_getsockname(newsock, sin, &slen);
1108 if (unlikely(err < 0)) {
1109 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err);
1110 slen = offsetof(struct sockaddr, sa_data);
1112 memcpy(&newsvsk->sk_local, sin, slen);
1115 serv->sv_stats->nettcpconn++;
1117 return &newsvsk->sk_xprt;
1120 sock_release(newsock);
1125 * Receive data from a TCP socket.
1128 svc_tcp_recvfrom(struct svc_rqst *rqstp)
1130 struct svc_sock *svsk = rqstp->rq_sock;
1131 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1136 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
1137 svsk, test_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags),
1138 test_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags),
1139 test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags));
1141 if ((rqstp->rq_deferred = svc_deferred_dequeue(svsk))) {
1142 svc_xprt_received(&svsk->sk_xprt);
1143 return svc_deferred_recv(rqstp);
1146 if (test_and_clear_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags))
1147 /* sndbuf needs to have room for one request
1148 * per thread, otherwise we can stall even when the
1149 * network isn't a bottleneck.
1151 * We count all threads rather than threads in a
1152 * particular pool, which provides an upper bound
1153 * on the number of threads which will access the socket.
1155 * rcvbuf just needs to be able to hold a few requests.
1156 * Normally they will be removed from the queue
1157 * as soon a a complete request arrives.
1159 svc_sock_setbufsize(svsk->sk_sock,
1160 (serv->sv_nrthreads+3) * serv->sv_max_mesg,
1161 3 * serv->sv_max_mesg);
1163 clear_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1165 /* Receive data. If we haven't got the record length yet, get
1166 * the next four bytes. Otherwise try to gobble up as much as
1167 * possible up to the complete record length.
1169 if (svsk->sk_tcplen < 4) {
1170 unsigned long want = 4 - svsk->sk_tcplen;
1173 iov.iov_base = ((char *) &svsk->sk_reclen) + svsk->sk_tcplen;
1175 if ((len = svc_recvfrom(rqstp, &iov, 1, want)) < 0)
1177 svsk->sk_tcplen += len;
1180 dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
1182 svc_xprt_received(&svsk->sk_xprt);
1183 return -EAGAIN; /* record header not complete */
1186 svsk->sk_reclen = ntohl(svsk->sk_reclen);
1187 if (!(svsk->sk_reclen & 0x80000000)) {
1188 /* FIXME: technically, a record can be fragmented,
1189 * and non-terminal fragments will not have the top
1190 * bit set in the fragment length header.
1191 * But apparently no known nfs clients send fragmented
1193 if (net_ratelimit())
1194 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1195 " (non-terminal)\n",
1196 (unsigned long) svsk->sk_reclen);
1199 svsk->sk_reclen &= 0x7fffffff;
1200 dprintk("svc: TCP record, %d bytes\n", svsk->sk_reclen);
1201 if (svsk->sk_reclen > serv->sv_max_mesg) {
1202 if (net_ratelimit())
1203 printk(KERN_NOTICE "RPC: bad TCP reclen 0x%08lx"
1205 (unsigned long) svsk->sk_reclen);
1210 /* Check whether enough data is available */
1211 len = svc_recv_available(svsk);
1215 if (len < svsk->sk_reclen) {
1216 dprintk("svc: incomplete TCP record (%d of %d)\n",
1217 len, svsk->sk_reclen);
1218 svc_xprt_received(&svsk->sk_xprt);
1219 return -EAGAIN; /* record not complete */
1221 len = svsk->sk_reclen;
1222 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1224 vec = rqstp->rq_vec;
1225 vec[0] = rqstp->rq_arg.head[0];
1228 while (vlen < len) {
1229 vec[pnum].iov_base = page_address(rqstp->rq_pages[pnum]);
1230 vec[pnum].iov_len = PAGE_SIZE;
1234 rqstp->rq_respages = &rqstp->rq_pages[pnum];
1236 /* Now receive data */
1237 len = svc_recvfrom(rqstp, vec, pnum, len);
1241 dprintk("svc: TCP complete record (%d bytes)\n", len);
1242 rqstp->rq_arg.len = len;
1243 rqstp->rq_arg.page_base = 0;
1244 if (len <= rqstp->rq_arg.head[0].iov_len) {
1245 rqstp->rq_arg.head[0].iov_len = len;
1246 rqstp->rq_arg.page_len = 0;
1248 rqstp->rq_arg.page_len = len - rqstp->rq_arg.head[0].iov_len;
1251 rqstp->rq_xprt_ctxt = NULL;
1252 rqstp->rq_prot = IPPROTO_TCP;
1254 /* Reset TCP read info */
1255 svsk->sk_reclen = 0;
1256 svsk->sk_tcplen = 0;
1258 svc_xprt_received(&svsk->sk_xprt);
1260 serv->sv_stats->nettcpcnt++;
1265 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1269 if (len == -EAGAIN) {
1270 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1271 svc_xprt_received(&svsk->sk_xprt);
1273 printk(KERN_NOTICE "%s: recvfrom returned errno %d\n",
1274 svsk->sk_xprt.xpt_server->sv_name, -len);
1282 * Send out data on TCP socket.
1285 svc_tcp_sendto(struct svc_rqst *rqstp)
1287 struct xdr_buf *xbufp = &rqstp->rq_res;
1291 /* Set up the first element of the reply kvec.
1292 * Any other kvecs that may be in use have been taken
1293 * care of by the server implementation itself.
1295 reclen = htonl(0x80000000|((xbufp->len ) - 4));
1296 memcpy(xbufp->head[0].iov_base, &reclen, 4);
1298 if (test_bit(XPT_DEAD, &rqstp->rq_sock->sk_xprt.xpt_flags))
1301 sent = svc_sendto(rqstp, &rqstp->rq_res);
1302 if (sent != xbufp->len) {
1303 printk(KERN_NOTICE "rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1304 rqstp->rq_sock->sk_xprt.xpt_server->sv_name,
1305 (sent<0)?"got error":"sent only",
1307 set_bit(XPT_CLOSE, &rqstp->rq_sock->sk_xprt.xpt_flags);
1308 svc_xprt_enqueue(rqstp->rq_xprt);
1315 * Setup response header. TCP has a 4B record length field.
1317 static void svc_tcp_prep_reply_hdr(struct svc_rqst *rqstp)
1319 struct kvec *resv = &rqstp->rq_res.head[0];
1321 /* tcp needs a space for the record length... */
1325 static int svc_tcp_has_wspace(struct svc_xprt *xprt)
1327 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1328 struct svc_serv *serv = svsk->sk_xprt.xpt_server;
1333 * Set the SOCK_NOSPACE flag before checking the available
1336 set_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1337 required = atomic_read(&svsk->sk_xprt.xpt_reserved) + serv->sv_max_mesg;
1338 wspace = sk_stream_wspace(svsk->sk_sk);
1340 if (wspace < sk_stream_min_wspace(svsk->sk_sk))
1342 if (required * 2 > wspace)
1345 clear_bit(SOCK_NOSPACE, &svsk->sk_sock->flags);
1349 static struct svc_xprt *svc_tcp_create(struct svc_serv *serv,
1350 struct sockaddr *sa, int salen,
1353 return svc_create_socket(serv, IPPROTO_TCP, sa, salen, flags);
1356 static struct svc_xprt_ops svc_tcp_ops = {
1357 .xpo_create = svc_tcp_create,
1358 .xpo_recvfrom = svc_tcp_recvfrom,
1359 .xpo_sendto = svc_tcp_sendto,
1360 .xpo_release_rqst = svc_release_skb,
1361 .xpo_detach = svc_sock_detach,
1362 .xpo_free = svc_sock_free,
1363 .xpo_prep_reply_hdr = svc_tcp_prep_reply_hdr,
1364 .xpo_has_wspace = svc_tcp_has_wspace,
1365 .xpo_accept = svc_tcp_accept,
1368 static struct svc_xprt_class svc_tcp_class = {
1370 .xcl_owner = THIS_MODULE,
1371 .xcl_ops = &svc_tcp_ops,
1372 .xcl_max_payload = RPCSVC_MAXPAYLOAD_TCP,
1375 void svc_init_xprt_sock(void)
1377 svc_reg_xprt_class(&svc_tcp_class);
1378 svc_reg_xprt_class(&svc_udp_class);
1381 void svc_cleanup_xprt_sock(void)
1383 svc_unreg_xprt_class(&svc_tcp_class);
1384 svc_unreg_xprt_class(&svc_udp_class);
1387 static void svc_tcp_init(struct svc_sock *svsk, struct svc_serv *serv)
1389 struct sock *sk = svsk->sk_sk;
1390 struct tcp_sock *tp = tcp_sk(sk);
1392 svc_xprt_init(&svc_tcp_class, &svsk->sk_xprt, serv);
1393 set_bit(XPT_CACHE_AUTH, &svsk->sk_xprt.xpt_flags);
1394 if (sk->sk_state == TCP_LISTEN) {
1395 dprintk("setting up TCP socket for listening\n");
1396 set_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags);
1397 sk->sk_data_ready = svc_tcp_listen_data_ready;
1398 set_bit(XPT_CONN, &svsk->sk_xprt.xpt_flags);
1400 dprintk("setting up TCP socket for reading\n");
1401 sk->sk_state_change = svc_tcp_state_change;
1402 sk->sk_data_ready = svc_tcp_data_ready;
1403 sk->sk_write_space = svc_write_space;
1405 svsk->sk_reclen = 0;
1406 svsk->sk_tcplen = 0;
1408 tp->nonagle = 1; /* disable Nagle's algorithm */
1410 /* initialise setting must have enough space to
1411 * receive and respond to one request.
1412 * svc_tcp_recvfrom will re-adjust if necessary
1414 svc_sock_setbufsize(svsk->sk_sock,
1415 3 * svsk->sk_xprt.xpt_server->sv_max_mesg,
1416 3 * svsk->sk_xprt.xpt_server->sv_max_mesg);
1418 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1419 set_bit(XPT_DATA, &svsk->sk_xprt.xpt_flags);
1420 if (sk->sk_state != TCP_ESTABLISHED)
1421 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1426 svc_sock_update_bufs(struct svc_serv *serv)
1429 * The number of server threads has changed. Update
1430 * rcvbuf and sndbuf accordingly on all sockets
1432 struct list_head *le;
1434 spin_lock_bh(&serv->sv_lock);
1435 list_for_each(le, &serv->sv_permsocks) {
1436 struct svc_sock *svsk =
1437 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1438 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1440 list_for_each(le, &serv->sv_tempsocks) {
1441 struct svc_sock *svsk =
1442 list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1443 set_bit(XPT_CHNGBUF, &svsk->sk_xprt.xpt_flags);
1445 spin_unlock_bh(&serv->sv_lock);
1449 * Make sure that we don't have too many active connections. If we
1450 * have, something must be dropped.
1452 * There's no point in trying to do random drop here for DoS
1453 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
1454 * attacker can easily beat that.
1456 * The only somewhat efficient mechanism would be if drop old
1457 * connections from the same IP first. But right now we don't even
1458 * record the client IP in svc_sock.
1460 static void svc_check_conn_limits(struct svc_serv *serv)
1462 if (serv->sv_tmpcnt > (serv->sv_nrthreads+3)*20) {
1463 struct svc_sock *svsk = NULL;
1464 spin_lock_bh(&serv->sv_lock);
1465 if (!list_empty(&serv->sv_tempsocks)) {
1466 if (net_ratelimit()) {
1467 /* Try to help the admin */
1468 printk(KERN_NOTICE "%s: too many open TCP "
1469 "sockets, consider increasing the "
1470 "number of nfsd threads\n",
1474 * Always select the oldest socket. It's not fair,
1477 svsk = list_entry(serv->sv_tempsocks.prev,
1480 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1481 svc_xprt_get(&svsk->sk_xprt);
1483 spin_unlock_bh(&serv->sv_lock);
1486 svc_xprt_enqueue(&svsk->sk_xprt);
1487 svc_xprt_put(&svsk->sk_xprt);
1493 * Receive the next request on any socket. This code is carefully
1494 * organised not to touch any cachelines in the shared svc_serv
1495 * structure, only cachelines in the local svc_pool.
1498 svc_recv(struct svc_rqst *rqstp, long timeout)
1500 struct svc_sock *svsk = NULL;
1501 struct svc_serv *serv = rqstp->rq_server;
1502 struct svc_pool *pool = rqstp->rq_pool;
1505 struct xdr_buf *arg;
1506 DECLARE_WAITQUEUE(wait, current);
1508 dprintk("svc: server %p waiting for data (to = %ld)\n",
1513 "svc_recv: service %p, socket not NULL!\n",
1515 if (waitqueue_active(&rqstp->rq_wait))
1517 "svc_recv: service %p, wait queue active!\n",
1521 /* now allocate needed pages. If we get a failure, sleep briefly */
1522 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
1523 for (i=0; i < pages ; i++)
1524 while (rqstp->rq_pages[i] == NULL) {
1525 struct page *p = alloc_page(GFP_KERNEL);
1527 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1528 rqstp->rq_pages[i] = p;
1530 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
1531 BUG_ON(pages >= RPCSVC_MAXPAGES);
1533 /* Make arg->head point to first page and arg->pages point to rest */
1534 arg = &rqstp->rq_arg;
1535 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
1536 arg->head[0].iov_len = PAGE_SIZE;
1537 arg->pages = rqstp->rq_pages + 1;
1539 /* save at least one page for response */
1540 arg->page_len = (pages-2)*PAGE_SIZE;
1541 arg->len = (pages-1)*PAGE_SIZE;
1542 arg->tail[0].iov_len = 0;
1549 spin_lock_bh(&pool->sp_lock);
1550 if ((svsk = svc_sock_dequeue(pool)) != NULL) {
1551 rqstp->rq_sock = svsk;
1552 svc_xprt_get(&svsk->sk_xprt);
1553 rqstp->rq_reserved = serv->sv_max_mesg;
1554 atomic_add(rqstp->rq_reserved, &svsk->sk_xprt.xpt_reserved);
1556 /* No data pending. Go to sleep */
1557 svc_thread_enqueue(pool, rqstp);
1560 * We have to be able to interrupt this wait
1561 * to bring down the daemons ...
1563 set_current_state(TASK_INTERRUPTIBLE);
1564 add_wait_queue(&rqstp->rq_wait, &wait);
1565 spin_unlock_bh(&pool->sp_lock);
1567 schedule_timeout(timeout);
1571 spin_lock_bh(&pool->sp_lock);
1572 remove_wait_queue(&rqstp->rq_wait, &wait);
1574 if (!(svsk = rqstp->rq_sock)) {
1575 svc_thread_dequeue(pool, rqstp);
1576 spin_unlock_bh(&pool->sp_lock);
1577 dprintk("svc: server %p, no data yet\n", rqstp);
1578 return signalled()? -EINTR : -EAGAIN;
1581 spin_unlock_bh(&pool->sp_lock);
1584 if (test_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags)) {
1585 dprintk("svc_recv: found XPT_CLOSE\n");
1586 svc_delete_xprt(&svsk->sk_xprt);
1587 } else if (test_bit(XPT_LISTENER, &svsk->sk_xprt.xpt_flags)) {
1588 struct svc_xprt *newxpt;
1589 newxpt = svsk->sk_xprt.xpt_ops->xpo_accept(&svsk->sk_xprt);
1592 * We know this module_get will succeed because the
1593 * listener holds a reference too
1595 __module_get(newxpt->xpt_class->xcl_owner);
1596 svc_check_conn_limits(svsk->sk_xprt.xpt_server);
1597 svc_xprt_received(newxpt);
1599 svc_xprt_received(&svsk->sk_xprt);
1601 dprintk("svc: server %p, pool %u, socket %p, inuse=%d\n",
1602 rqstp, pool->sp_id, svsk,
1603 atomic_read(&svsk->sk_xprt.xpt_ref.refcount));
1604 len = svsk->sk_xprt.xpt_ops->xpo_recvfrom(rqstp);
1605 dprintk("svc: got len=%d\n", len);
1608 /* No data, incomplete (TCP) read, or accept() */
1609 if (len == 0 || len == -EAGAIN) {
1610 rqstp->rq_res.len = 0;
1611 svc_sock_release(rqstp);
1614 clear_bit(XPT_OLD, &svsk->sk_xprt.xpt_flags);
1616 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
1617 rqstp->rq_chandle.defer = svc_defer;
1620 serv->sv_stats->netcnt++;
1628 svc_drop(struct svc_rqst *rqstp)
1630 dprintk("svc: socket %p dropped request\n", rqstp->rq_sock);
1631 svc_sock_release(rqstp);
1635 * Return reply to client.
1638 svc_send(struct svc_rqst *rqstp)
1640 struct svc_xprt *xprt;
1644 xprt = rqstp->rq_xprt;
1648 /* release the receive skb before sending the reply */
1649 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
1651 /* calculate over-all length */
1652 xb = & rqstp->rq_res;
1653 xb->len = xb->head[0].iov_len +
1655 xb->tail[0].iov_len;
1657 /* Grab mutex to serialize outgoing data. */
1658 mutex_lock(&xprt->xpt_mutex);
1659 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
1662 len = xprt->xpt_ops->xpo_sendto(rqstp);
1663 mutex_unlock(&xprt->xpt_mutex);
1664 svc_sock_release(rqstp);
1666 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
1672 * Timer function to close old temporary sockets, using
1673 * a mark-and-sweep algorithm.
1676 svc_age_temp_sockets(unsigned long closure)
1678 struct svc_serv *serv = (struct svc_serv *)closure;
1679 struct svc_sock *svsk;
1680 struct list_head *le, *next;
1681 LIST_HEAD(to_be_aged);
1683 dprintk("svc_age_temp_sockets\n");
1685 if (!spin_trylock_bh(&serv->sv_lock)) {
1686 /* busy, try again 1 sec later */
1687 dprintk("svc_age_temp_sockets: busy\n");
1688 mod_timer(&serv->sv_temptimer, jiffies + HZ);
1692 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1693 svsk = list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1695 if (!test_and_set_bit(XPT_OLD, &svsk->sk_xprt.xpt_flags))
1697 if (atomic_read(&svsk->sk_xprt.xpt_ref.refcount) > 1
1698 || test_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags))
1700 svc_xprt_get(&svsk->sk_xprt);
1701 list_move(le, &to_be_aged);
1702 set_bit(XPT_CLOSE, &svsk->sk_xprt.xpt_flags);
1703 set_bit(XPT_DETACHED, &svsk->sk_xprt.xpt_flags);
1705 spin_unlock_bh(&serv->sv_lock);
1707 while (!list_empty(&to_be_aged)) {
1708 le = to_be_aged.next;
1709 /* fiddling the sk_xprt.xpt_list node is safe 'cos we're XPT_DETACHED */
1711 svsk = list_entry(le, struct svc_sock, sk_xprt.xpt_list);
1713 dprintk("queuing svsk %p for closing\n", svsk);
1715 /* a thread will dequeue and close it soon */
1716 svc_xprt_enqueue(&svsk->sk_xprt);
1717 svc_xprt_put(&svsk->sk_xprt);
1720 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
1724 * Initialize socket for RPC use and create svc_sock struct
1725 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1727 static struct svc_sock *svc_setup_socket(struct svc_serv *serv,
1728 struct socket *sock,
1729 int *errp, int flags)
1731 struct svc_sock *svsk;
1733 int pmap_register = !(flags & SVC_SOCK_ANONYMOUS);
1734 int is_temporary = flags & SVC_SOCK_TEMPORARY;
1736 dprintk("svc: svc_setup_socket %p\n", sock);
1737 if (!(svsk = kzalloc(sizeof(*svsk), GFP_KERNEL))) {
1744 /* Register socket with portmapper */
1745 if (*errp >= 0 && pmap_register)
1746 *errp = svc_register(serv, inet->sk_protocol,
1747 ntohs(inet_sk(inet)->sport));
1754 set_bit(XPT_BUSY, &svsk->sk_xprt.xpt_flags);
1755 inet->sk_user_data = svsk;
1756 svsk->sk_sock = sock;
1758 svsk->sk_ostate = inet->sk_state_change;
1759 svsk->sk_odata = inet->sk_data_ready;
1760 svsk->sk_owspace = inet->sk_write_space;
1761 INIT_LIST_HEAD(&svsk->sk_deferred);
1763 /* Initialize the socket */
1764 if (sock->type == SOCK_DGRAM)
1765 svc_udp_init(svsk, serv);
1767 svc_tcp_init(svsk, serv);
1769 spin_lock_bh(&serv->sv_lock);
1771 set_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1772 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_tempsocks);
1774 if (serv->sv_temptimer.function == NULL) {
1775 /* setup timer to age temp sockets */
1776 setup_timer(&serv->sv_temptimer, svc_age_temp_sockets,
1777 (unsigned long)serv);
1778 mod_timer(&serv->sv_temptimer,
1779 jiffies + svc_conn_age_period * HZ);
1782 clear_bit(XPT_TEMP, &svsk->sk_xprt.xpt_flags);
1783 list_add(&svsk->sk_xprt.xpt_list, &serv->sv_permsocks);
1785 spin_unlock_bh(&serv->sv_lock);
1787 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1793 int svc_addsock(struct svc_serv *serv,
1799 struct socket *so = sockfd_lookup(fd, &err);
1800 struct svc_sock *svsk = NULL;
1804 if (so->sk->sk_family != AF_INET)
1805 err = -EAFNOSUPPORT;
1806 else if (so->sk->sk_protocol != IPPROTO_TCP &&
1807 so->sk->sk_protocol != IPPROTO_UDP)
1808 err = -EPROTONOSUPPORT;
1809 else if (so->state > SS_UNCONNECTED)
1812 svsk = svc_setup_socket(serv, so, &err, SVC_SOCK_DEFAULTS);
1814 svc_xprt_received(&svsk->sk_xprt);
1822 if (proto) *proto = so->sk->sk_protocol;
1823 return one_sock_name(name_return, svsk);
1825 EXPORT_SYMBOL_GPL(svc_addsock);
1828 * Create socket for RPC service.
1830 static struct svc_xprt *svc_create_socket(struct svc_serv *serv,
1832 struct sockaddr *sin, int len,
1835 struct svc_sock *svsk;
1836 struct socket *sock;
1839 char buf[RPC_MAX_ADDRBUFLEN];
1841 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1842 serv->sv_program->pg_name, protocol,
1843 __svc_print_addr(sin, buf, sizeof(buf)));
1845 if (protocol != IPPROTO_UDP && protocol != IPPROTO_TCP) {
1846 printk(KERN_WARNING "svc: only UDP and TCP "
1847 "sockets supported\n");
1848 return ERR_PTR(-EINVAL);
1850 type = (protocol == IPPROTO_UDP)? SOCK_DGRAM : SOCK_STREAM;
1852 error = sock_create_kern(sin->sa_family, type, protocol, &sock);
1854 return ERR_PTR(error);
1856 svc_reclassify_socket(sock);
1858 if (type == SOCK_STREAM)
1859 sock->sk->sk_reuse = 1; /* allow address reuse */
1860 error = kernel_bind(sock, sin, len);
1864 if (protocol == IPPROTO_TCP) {
1865 if ((error = kernel_listen(sock, 64)) < 0)
1869 if ((svsk = svc_setup_socket(serv, sock, &error, flags)) != NULL) {
1870 svc_xprt_received(&svsk->sk_xprt);
1871 return (struct svc_xprt *)svsk;
1875 dprintk("svc: svc_create_socket error = %d\n", -error);
1877 return ERR_PTR(error);
1881 * Detach the svc_sock from the socket so that no
1882 * more callbacks occur.
1884 static void svc_sock_detach(struct svc_xprt *xprt)
1886 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1887 struct sock *sk = svsk->sk_sk;
1889 dprintk("svc: svc_sock_detach(%p)\n", svsk);
1891 /* put back the old socket callbacks */
1892 sk->sk_state_change = svsk->sk_ostate;
1893 sk->sk_data_ready = svsk->sk_odata;
1894 sk->sk_write_space = svsk->sk_owspace;
1898 * Free the svc_sock's socket resources and the svc_sock itself.
1900 static void svc_sock_free(struct svc_xprt *xprt)
1902 struct svc_sock *svsk = container_of(xprt, struct svc_sock, sk_xprt);
1903 dprintk("svc: svc_sock_free(%p)\n", svsk);
1905 if (svsk->sk_sock->file)
1906 sockfd_put(svsk->sk_sock);
1908 sock_release(svsk->sk_sock);
1913 * Remove a dead transport
1915 static void svc_delete_xprt(struct svc_xprt *xprt)
1917 struct svc_serv *serv = xprt->xpt_server;
1919 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1920 xprt->xpt_ops->xpo_detach(xprt);
1922 spin_lock_bh(&serv->sv_lock);
1923 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
1924 list_del_init(&xprt->xpt_list);
1926 * We used to delete the transport from whichever list
1927 * it's sk_xprt.xpt_ready node was on, but we don't actually
1928 * need to. This is because the only time we're called
1929 * while still attached to a queue, the queue itself
1930 * is about to be destroyed (in svc_destroy).
1932 if (!test_and_set_bit(XPT_DEAD, &xprt->xpt_flags)) {
1933 BUG_ON(atomic_read(&xprt->xpt_ref.refcount) < 2);
1934 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1938 spin_unlock_bh(&serv->sv_lock);
1941 static void svc_close_xprt(struct svc_xprt *xprt)
1943 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1944 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1945 /* someone else will have to effect the close */
1949 svc_delete_xprt(xprt);
1950 clear_bit(XPT_BUSY, &xprt->xpt_flags);
1954 void svc_close_all(struct list_head *xprt_list)
1956 struct svc_xprt *xprt;
1957 struct svc_xprt *tmp;
1959 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
1960 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1961 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
1962 /* Waiting to be processed, but no threads left,
1963 * So just remove it from the waiting list
1965 list_del_init(&xprt->xpt_ready);
1966 clear_bit(XPT_BUSY, &xprt->xpt_flags);
1968 svc_close_xprt(xprt);
1973 * Handle defer and revisit of requests
1976 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1978 struct svc_deferred_req *dr = container_of(dreq, struct svc_deferred_req, handle);
1979 struct svc_sock *svsk;
1982 svc_xprt_put(&dr->svsk->sk_xprt);
1986 dprintk("revisit queued\n");
1989 spin_lock(&svsk->sk_xprt.xpt_lock);
1990 list_add(&dr->handle.recent, &svsk->sk_deferred);
1991 spin_unlock(&svsk->sk_xprt.xpt_lock);
1992 set_bit(XPT_DEFERRED, &svsk->sk_xprt.xpt_flags);
1993 svc_xprt_enqueue(&svsk->sk_xprt);
1994 svc_xprt_put(&svsk->sk_xprt);
1997 static struct cache_deferred_req *
1998 svc_defer(struct cache_req *req)
2000 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
2001 int size = sizeof(struct svc_deferred_req) + (rqstp->rq_arg.len);
2002 struct svc_deferred_req *dr;
2004 if (rqstp->rq_arg.page_len)
2005 return NULL; /* if more than a page, give up FIXME */
2006 if (rqstp->rq_deferred) {
2007 dr = rqstp->rq_deferred;
2008 rqstp->rq_deferred = NULL;
2010 int skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
2011 /* FIXME maybe discard if size too large */
2012 dr = kmalloc(size, GFP_KERNEL);
2016 dr->handle.owner = rqstp->rq_server;
2017 dr->prot = rqstp->rq_prot;
2018 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
2019 dr->addrlen = rqstp->rq_addrlen;
2020 dr->daddr = rqstp->rq_daddr;
2021 dr->argslen = rqstp->rq_arg.len >> 2;
2022 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base-skip, dr->argslen<<2);
2024 svc_xprt_get(rqstp->rq_xprt);
2025 dr->svsk = rqstp->rq_sock;
2027 dr->handle.revisit = svc_revisit;
2032 * recv data from a deferred request into an active one
2034 static int svc_deferred_recv(struct svc_rqst *rqstp)
2036 struct svc_deferred_req *dr = rqstp->rq_deferred;
2038 rqstp->rq_arg.head[0].iov_base = dr->args;
2039 rqstp->rq_arg.head[0].iov_len = dr->argslen<<2;
2040 rqstp->rq_arg.page_len = 0;
2041 rqstp->rq_arg.len = dr->argslen<<2;
2042 rqstp->rq_prot = dr->prot;
2043 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
2044 rqstp->rq_addrlen = dr->addrlen;
2045 rqstp->rq_daddr = dr->daddr;
2046 rqstp->rq_respages = rqstp->rq_pages;
2047 return dr->argslen<<2;
2051 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_sock *svsk)
2053 struct svc_deferred_req *dr = NULL;
2055 if (!test_bit(XPT_DEFERRED, &svsk->sk_xprt.xpt_flags))
2057 spin_lock(&svsk->sk_xprt.xpt_lock);
2058 clear_bit(XPT_DEFERRED, &svsk->sk_xprt.xpt_flags);
2059 if (!list_empty(&svsk->sk_deferred)) {
2060 dr = list_entry(svsk->sk_deferred.next,
2061 struct svc_deferred_req,
2063 list_del_init(&dr->handle.recent);
2064 set_bit(XPT_DEFERRED, &svsk->sk_xprt.xpt_flags);
2066 spin_unlock(&svsk->sk_xprt.xpt_lock);