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_sock_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_flags.SK_BUSY prevents a svc_sock being enqueued multiply.
61 * Some flags can be set to certain values at any time
62 * providing that certain rules are followed:
64 * SK_CONN, SK_DATA, can be set or cleared at any time.
65 * after a set, svc_sock_enqueue must be called.
66 * after a clear, the socket must be read/accepted
67 * if this succeeds, it must be set again.
68 * SK_CLOSE can set at any time. It is never cleared.
69 * xpt_ref contains a bias of '1' until SK_DEAD is set.
70 * so when xprt_ref hits zero, we know the transport is dead
71 * and no-one is using it.
72 * SK_DEAD can only be set while SK_BUSY is held which ensures
73 * no other thread will be using the socket or will try to
78 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
81 static struct svc_sock
*svc_setup_socket(struct svc_serv
*, struct socket
*,
82 int *errp
, int flags
);
83 static void svc_delete_socket(struct svc_sock
*svsk
);
84 static void svc_udp_data_ready(struct sock
*, int);
85 static int svc_udp_recvfrom(struct svc_rqst
*);
86 static int svc_udp_sendto(struct svc_rqst
*);
87 static void svc_close_socket(struct svc_sock
*svsk
);
88 static void svc_sock_detach(struct svc_xprt
*);
89 static void svc_sock_free(struct svc_xprt
*);
91 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_sock
*svsk
);
92 static int svc_deferred_recv(struct svc_rqst
*rqstp
);
93 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
);
94 static struct svc_xprt
*svc_create_socket(struct svc_serv
*, int,
95 struct sockaddr
*, int, int);
97 /* apparently the "standard" is that clients close
98 * idle connections after 5 minutes, servers after
100 * http://www.connectathon.org/talks96/nfstcp.pdf
102 static int svc_conn_age_period
= 6*60;
104 #ifdef CONFIG_DEBUG_LOCK_ALLOC
105 static struct lock_class_key svc_key
[2];
106 static struct lock_class_key svc_slock_key
[2];
108 static inline void svc_reclassify_socket(struct socket
*sock
)
110 struct sock
*sk
= sock
->sk
;
111 BUG_ON(sock_owned_by_user(sk
));
112 switch (sk
->sk_family
) {
114 sock_lock_init_class_and_name(sk
, "slock-AF_INET-NFSD",
115 &svc_slock_key
[0], "sk_lock-AF_INET-NFSD", &svc_key
[0]);
119 sock_lock_init_class_and_name(sk
, "slock-AF_INET6-NFSD",
120 &svc_slock_key
[1], "sk_lock-AF_INET6-NFSD", &svc_key
[1]);
128 static inline void svc_reclassify_socket(struct socket
*sock
)
133 static char *__svc_print_addr(struct sockaddr
*addr
, char *buf
, size_t len
)
135 switch (addr
->sa_family
) {
137 snprintf(buf
, len
, "%u.%u.%u.%u, port=%u",
138 NIPQUAD(((struct sockaddr_in
*) addr
)->sin_addr
),
139 ntohs(((struct sockaddr_in
*) addr
)->sin_port
));
143 snprintf(buf
, len
, "%x:%x:%x:%x:%x:%x:%x:%x, port=%u",
144 NIP6(((struct sockaddr_in6
*) addr
)->sin6_addr
),
145 ntohs(((struct sockaddr_in6
*) addr
)->sin6_port
));
149 snprintf(buf
, len
, "unknown address type: %d", addr
->sa_family
);
156 * svc_print_addr - Format rq_addr field for printing
157 * @rqstp: svc_rqst struct containing address to print
158 * @buf: target buffer for formatted address
159 * @len: length of target buffer
162 char *svc_print_addr(struct svc_rqst
*rqstp
, char *buf
, size_t len
)
164 return __svc_print_addr(svc_addr(rqstp
), buf
, len
);
166 EXPORT_SYMBOL_GPL(svc_print_addr
);
169 * Queue up an idle server thread. Must have pool->sp_lock held.
170 * Note: this is really a stack rather than a queue, so that we only
171 * use as many different threads as we need, and the rest don't pollute
175 svc_thread_enqueue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
177 list_add(&rqstp
->rq_list
, &pool
->sp_threads
);
181 * Dequeue an nfsd thread. Must have pool->sp_lock held.
184 svc_thread_dequeue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
186 list_del(&rqstp
->rq_list
);
190 * Release an skbuff after use
192 static void svc_release_skb(struct svc_rqst
*rqstp
)
194 struct sk_buff
*skb
= rqstp
->rq_xprt_ctxt
;
195 struct svc_deferred_req
*dr
= rqstp
->rq_deferred
;
198 rqstp
->rq_xprt_ctxt
= NULL
;
200 dprintk("svc: service %p, releasing skb %p\n", rqstp
, skb
);
201 skb_free_datagram(rqstp
->rq_sock
->sk_sk
, skb
);
204 rqstp
->rq_deferred
= NULL
;
210 * Queue up a socket with data pending. If there are idle nfsd
211 * processes, wake 'em up.
215 svc_sock_enqueue(struct svc_sock
*svsk
)
217 struct svc_serv
*serv
= svsk
->sk_server
;
218 struct svc_pool
*pool
;
219 struct svc_rqst
*rqstp
;
222 if (!(svsk
->sk_flags
&
223 ( (1<<SK_CONN
)|(1<<SK_DATA
)|(1<<SK_CLOSE
)|(1<<SK_DEFERRED
)) ))
225 if (test_bit(SK_DEAD
, &svsk
->sk_flags
))
229 pool
= svc_pool_for_cpu(svsk
->sk_server
, cpu
);
232 spin_lock_bh(&pool
->sp_lock
);
234 if (!list_empty(&pool
->sp_threads
) &&
235 !list_empty(&pool
->sp_sockets
))
237 "svc_sock_enqueue: threads and sockets both waiting??\n");
239 if (test_bit(SK_DEAD
, &svsk
->sk_flags
)) {
240 /* Don't enqueue dead sockets */
241 dprintk("svc: socket %p is dead, not enqueued\n", svsk
->sk_sk
);
245 /* Mark socket as busy. It will remain in this state until the
246 * server has processed all pending data and put the socket back
247 * on the idle list. We update SK_BUSY atomically because
248 * it also guards against trying to enqueue the svc_sock twice.
250 if (test_and_set_bit(SK_BUSY
, &svsk
->sk_flags
)) {
251 /* Don't enqueue socket while already enqueued */
252 dprintk("svc: socket %p busy, not enqueued\n", svsk
->sk_sk
);
255 BUG_ON(svsk
->sk_pool
!= NULL
);
256 svsk
->sk_pool
= pool
;
258 /* Handle pending connection */
259 if (test_bit(SK_CONN
, &svsk
->sk_flags
))
262 /* Handle close in-progress */
263 if (test_bit(SK_CLOSE
, &svsk
->sk_flags
))
266 /* Check if we have space to reply to a request */
267 if (!svsk
->sk_xprt
.xpt_ops
->xpo_has_wspace(&svsk
->sk_xprt
)) {
268 /* Don't enqueue while not enough space for reply */
269 dprintk("svc: no write space, socket %p not enqueued\n", svsk
);
270 svsk
->sk_pool
= NULL
;
271 clear_bit(SK_BUSY
, &svsk
->sk_flags
);
276 if (!list_empty(&pool
->sp_threads
)) {
277 rqstp
= list_entry(pool
->sp_threads
.next
,
280 dprintk("svc: socket %p served by daemon %p\n",
282 svc_thread_dequeue(pool
, rqstp
);
285 "svc_sock_enqueue: server %p, rq_sock=%p!\n",
286 rqstp
, rqstp
->rq_sock
);
287 rqstp
->rq_sock
= svsk
;
288 svc_xprt_get(&svsk
->sk_xprt
);
289 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
290 atomic_add(rqstp
->rq_reserved
, &svsk
->sk_reserved
);
291 BUG_ON(svsk
->sk_pool
!= pool
);
292 wake_up(&rqstp
->rq_wait
);
294 dprintk("svc: socket %p put into queue\n", svsk
->sk_sk
);
295 list_add_tail(&svsk
->sk_ready
, &pool
->sp_sockets
);
296 BUG_ON(svsk
->sk_pool
!= pool
);
300 spin_unlock_bh(&pool
->sp_lock
);
304 * Dequeue the first socket. Must be called with the pool->sp_lock held.
306 static inline struct svc_sock
*
307 svc_sock_dequeue(struct svc_pool
*pool
)
309 struct svc_sock
*svsk
;
311 if (list_empty(&pool
->sp_sockets
))
314 svsk
= list_entry(pool
->sp_sockets
.next
,
315 struct svc_sock
, sk_ready
);
316 list_del_init(&svsk
->sk_ready
);
318 dprintk("svc: socket %p dequeued, inuse=%d\n",
319 svsk
->sk_sk
, atomic_read(&svsk
->sk_xprt
.xpt_ref
.refcount
));
325 * Having read something from a socket, check whether it
326 * needs to be re-enqueued.
327 * Note: SK_DATA only gets cleared when a read-attempt finds
328 * no (or insufficient) data.
331 svc_sock_received(struct svc_sock
*svsk
)
333 svsk
->sk_pool
= NULL
;
334 clear_bit(SK_BUSY
, &svsk
->sk_flags
);
335 svc_sock_enqueue(svsk
);
340 * svc_reserve - change the space reserved for the reply to a request.
341 * @rqstp: The request in question
342 * @space: new max space to reserve
344 * Each request reserves some space on the output queue of the socket
345 * to make sure the reply fits. This function reduces that reserved
346 * space to be the amount of space used already, plus @space.
349 void svc_reserve(struct svc_rqst
*rqstp
, int space
)
351 space
+= rqstp
->rq_res
.head
[0].iov_len
;
353 if (space
< rqstp
->rq_reserved
) {
354 struct svc_sock
*svsk
= rqstp
->rq_sock
;
355 atomic_sub((rqstp
->rq_reserved
- space
), &svsk
->sk_reserved
);
356 rqstp
->rq_reserved
= space
;
358 svc_sock_enqueue(svsk
);
363 svc_sock_release(struct svc_rqst
*rqstp
)
365 struct svc_sock
*svsk
= rqstp
->rq_sock
;
367 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
369 svc_free_res_pages(rqstp
);
370 rqstp
->rq_res
.page_len
= 0;
371 rqstp
->rq_res
.page_base
= 0;
374 /* Reset response buffer and release
376 * But first, check that enough space was reserved
377 * for the reply, otherwise we have a bug!
379 if ((rqstp
->rq_res
.len
) > rqstp
->rq_reserved
)
380 printk(KERN_ERR
"RPC request reserved %d but used %d\n",
384 rqstp
->rq_res
.head
[0].iov_len
= 0;
385 svc_reserve(rqstp
, 0);
386 rqstp
->rq_sock
= NULL
;
388 svc_xprt_put(&svsk
->sk_xprt
);
392 * External function to wake up a server waiting for data
393 * This really only makes sense for services like lockd
394 * which have exactly one thread anyway.
397 svc_wake_up(struct svc_serv
*serv
)
399 struct svc_rqst
*rqstp
;
401 struct svc_pool
*pool
;
403 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
404 pool
= &serv
->sv_pools
[i
];
406 spin_lock_bh(&pool
->sp_lock
);
407 if (!list_empty(&pool
->sp_threads
)) {
408 rqstp
= list_entry(pool
->sp_threads
.next
,
411 dprintk("svc: daemon %p woken up.\n", rqstp
);
413 svc_thread_dequeue(pool, rqstp);
414 rqstp->rq_sock = NULL;
416 wake_up(&rqstp
->rq_wait
);
418 spin_unlock_bh(&pool
->sp_lock
);
422 union svc_pktinfo_u
{
423 struct in_pktinfo pkti
;
424 struct in6_pktinfo pkti6
;
426 #define SVC_PKTINFO_SPACE \
427 CMSG_SPACE(sizeof(union svc_pktinfo_u))
429 static void svc_set_cmsg_data(struct svc_rqst
*rqstp
, struct cmsghdr
*cmh
)
431 switch (rqstp
->rq_sock
->sk_sk
->sk_family
) {
433 struct in_pktinfo
*pki
= CMSG_DATA(cmh
);
435 cmh
->cmsg_level
= SOL_IP
;
436 cmh
->cmsg_type
= IP_PKTINFO
;
437 pki
->ipi_ifindex
= 0;
438 pki
->ipi_spec_dst
.s_addr
= rqstp
->rq_daddr
.addr
.s_addr
;
439 cmh
->cmsg_len
= CMSG_LEN(sizeof(*pki
));
444 struct in6_pktinfo
*pki
= CMSG_DATA(cmh
);
446 cmh
->cmsg_level
= SOL_IPV6
;
447 cmh
->cmsg_type
= IPV6_PKTINFO
;
448 pki
->ipi6_ifindex
= 0;
449 ipv6_addr_copy(&pki
->ipi6_addr
,
450 &rqstp
->rq_daddr
.addr6
);
451 cmh
->cmsg_len
= CMSG_LEN(sizeof(*pki
));
459 * Generic sendto routine
462 svc_sendto(struct svc_rqst
*rqstp
, struct xdr_buf
*xdr
)
464 struct svc_sock
*svsk
= rqstp
->rq_sock
;
465 struct socket
*sock
= svsk
->sk_sock
;
469 long all
[SVC_PKTINFO_SPACE
/ sizeof(long)];
471 struct cmsghdr
*cmh
= &buffer
.hdr
;
475 struct page
**ppage
= xdr
->pages
;
476 size_t base
= xdr
->page_base
;
477 unsigned int pglen
= xdr
->page_len
;
478 unsigned int flags
= MSG_MORE
;
479 char buf
[RPC_MAX_ADDRBUFLEN
];
483 if (rqstp
->rq_prot
== IPPROTO_UDP
) {
484 struct msghdr msg
= {
485 .msg_name
= &rqstp
->rq_addr
,
486 .msg_namelen
= rqstp
->rq_addrlen
,
488 .msg_controllen
= sizeof(buffer
),
489 .msg_flags
= MSG_MORE
,
492 svc_set_cmsg_data(rqstp
, cmh
);
494 if (sock_sendmsg(sock
, &msg
, 0) < 0)
499 if (slen
== xdr
->head
[0].iov_len
)
501 len
= kernel_sendpage(sock
, rqstp
->rq_respages
[0], 0,
502 xdr
->head
[0].iov_len
, flags
);
503 if (len
!= xdr
->head
[0].iov_len
)
505 slen
-= xdr
->head
[0].iov_len
;
510 size
= PAGE_SIZE
- base
< pglen
? PAGE_SIZE
- base
: pglen
;
514 result
= kernel_sendpage(sock
, *ppage
, base
, size
, flags
);
521 size
= PAGE_SIZE
< pglen
? PAGE_SIZE
: pglen
;
526 if (xdr
->tail
[0].iov_len
) {
527 result
= kernel_sendpage(sock
, rqstp
->rq_respages
[0],
528 ((unsigned long)xdr
->tail
[0].iov_base
)
530 xdr
->tail
[0].iov_len
, 0);
536 dprintk("svc: socket %p sendto([%p %Zu... ], %d) = %d (addr %s)\n",
537 rqstp
->rq_sock
, xdr
->head
[0].iov_base
, xdr
->head
[0].iov_len
,
538 xdr
->len
, len
, svc_print_addr(rqstp
, buf
, sizeof(buf
)));
544 * Report socket names for nfsdfs
546 static int one_sock_name(char *buf
, struct svc_sock
*svsk
)
550 switch(svsk
->sk_sk
->sk_family
) {
552 len
= sprintf(buf
, "ipv4 %s %u.%u.%u.%u %d\n",
553 svsk
->sk_sk
->sk_protocol
==IPPROTO_UDP
?
555 NIPQUAD(inet_sk(svsk
->sk_sk
)->rcv_saddr
),
556 inet_sk(svsk
->sk_sk
)->num
);
559 len
= sprintf(buf
, "*unknown-%d*\n",
560 svsk
->sk_sk
->sk_family
);
566 svc_sock_names(char *buf
, struct svc_serv
*serv
, char *toclose
)
568 struct svc_sock
*svsk
, *closesk
= NULL
;
573 spin_lock_bh(&serv
->sv_lock
);
574 list_for_each_entry(svsk
, &serv
->sv_permsocks
, sk_list
) {
575 int onelen
= one_sock_name(buf
+len
, svsk
);
576 if (toclose
&& strcmp(toclose
, buf
+len
) == 0)
581 spin_unlock_bh(&serv
->sv_lock
);
583 /* Should unregister with portmap, but you cannot
584 * unregister just one protocol...
586 svc_close_socket(closesk
);
591 EXPORT_SYMBOL(svc_sock_names
);
594 * Check input queue length
597 svc_recv_available(struct svc_sock
*svsk
)
599 struct socket
*sock
= svsk
->sk_sock
;
602 err
= kernel_sock_ioctl(sock
, TIOCINQ
, (unsigned long) &avail
);
604 return (err
>= 0)? avail
: err
;
608 * Generic recvfrom routine.
611 svc_recvfrom(struct svc_rqst
*rqstp
, struct kvec
*iov
, int nr
, int buflen
)
613 struct svc_sock
*svsk
= rqstp
->rq_sock
;
614 struct msghdr msg
= {
615 .msg_flags
= MSG_DONTWAIT
,
617 struct sockaddr
*sin
;
620 len
= kernel_recvmsg(svsk
->sk_sock
, &msg
, iov
, nr
, buflen
,
623 /* sock_recvmsg doesn't fill in the name/namelen, so we must..
625 memcpy(&rqstp
->rq_addr
, &svsk
->sk_remote
, svsk
->sk_remotelen
);
626 rqstp
->rq_addrlen
= svsk
->sk_remotelen
;
628 /* Destination address in request is needed for binding the
629 * source address in RPC callbacks later.
631 sin
= (struct sockaddr
*)&svsk
->sk_local
;
632 switch (sin
->sa_family
) {
634 rqstp
->rq_daddr
.addr
= ((struct sockaddr_in
*)sin
)->sin_addr
;
637 rqstp
->rq_daddr
.addr6
= ((struct sockaddr_in6
*)sin
)->sin6_addr
;
641 dprintk("svc: socket %p recvfrom(%p, %Zu) = %d\n",
642 svsk
, iov
[0].iov_base
, iov
[0].iov_len
, len
);
648 * Set socket snd and rcv buffer lengths
651 svc_sock_setbufsize(struct socket
*sock
, unsigned int snd
, unsigned int rcv
)
655 oldfs
= get_fs(); set_fs(KERNEL_DS
);
656 sock_setsockopt(sock
, SOL_SOCKET
, SO_SNDBUF
,
657 (char*)&snd
, sizeof(snd
));
658 sock_setsockopt(sock
, SOL_SOCKET
, SO_RCVBUF
,
659 (char*)&rcv
, sizeof(rcv
));
661 /* sock_setsockopt limits use to sysctl_?mem_max,
662 * which isn't acceptable. Until that is made conditional
663 * on not having CAP_SYS_RESOURCE or similar, we go direct...
664 * DaveM said I could!
667 sock
->sk
->sk_sndbuf
= snd
* 2;
668 sock
->sk
->sk_rcvbuf
= rcv
* 2;
669 sock
->sk
->sk_userlocks
|= SOCK_SNDBUF_LOCK
|SOCK_RCVBUF_LOCK
;
670 release_sock(sock
->sk
);
674 * INET callback when data has been received on the socket.
677 svc_udp_data_ready(struct sock
*sk
, int count
)
679 struct svc_sock
*svsk
= (struct svc_sock
*)sk
->sk_user_data
;
682 dprintk("svc: socket %p(inet %p), count=%d, busy=%d\n",
683 svsk
, sk
, count
, test_bit(SK_BUSY
, &svsk
->sk_flags
));
684 set_bit(SK_DATA
, &svsk
->sk_flags
);
685 svc_sock_enqueue(svsk
);
687 if (sk
->sk_sleep
&& waitqueue_active(sk
->sk_sleep
))
688 wake_up_interruptible(sk
->sk_sleep
);
692 * INET callback when space is newly available on the socket.
695 svc_write_space(struct sock
*sk
)
697 struct svc_sock
*svsk
= (struct svc_sock
*)(sk
->sk_user_data
);
700 dprintk("svc: socket %p(inet %p), write_space busy=%d\n",
701 svsk
, sk
, test_bit(SK_BUSY
, &svsk
->sk_flags
));
702 svc_sock_enqueue(svsk
);
705 if (sk
->sk_sleep
&& waitqueue_active(sk
->sk_sleep
)) {
706 dprintk("RPC svc_write_space: someone sleeping on %p\n",
708 wake_up_interruptible(sk
->sk_sleep
);
712 static inline void svc_udp_get_dest_address(struct svc_rqst
*rqstp
,
715 switch (rqstp
->rq_sock
->sk_sk
->sk_family
) {
717 struct in_pktinfo
*pki
= CMSG_DATA(cmh
);
718 rqstp
->rq_daddr
.addr
.s_addr
= pki
->ipi_spec_dst
.s_addr
;
722 struct in6_pktinfo
*pki
= CMSG_DATA(cmh
);
723 ipv6_addr_copy(&rqstp
->rq_daddr
.addr6
, &pki
->ipi6_addr
);
730 * Receive a datagram from a UDP socket.
733 svc_udp_recvfrom(struct svc_rqst
*rqstp
)
735 struct svc_sock
*svsk
= rqstp
->rq_sock
;
736 struct svc_serv
*serv
= svsk
->sk_server
;
740 long all
[SVC_PKTINFO_SPACE
/ sizeof(long)];
742 struct cmsghdr
*cmh
= &buffer
.hdr
;
744 struct msghdr msg
= {
745 .msg_name
= svc_addr(rqstp
),
747 .msg_controllen
= sizeof(buffer
),
748 .msg_flags
= MSG_DONTWAIT
,
751 if (test_and_clear_bit(SK_CHNGBUF
, &svsk
->sk_flags
))
752 /* udp sockets need large rcvbuf as all pending
753 * requests are still in that buffer. sndbuf must
754 * also be large enough that there is enough space
755 * for one reply per thread. We count all threads
756 * rather than threads in a particular pool, which
757 * provides an upper bound on the number of threads
758 * which will access the socket.
760 svc_sock_setbufsize(svsk
->sk_sock
,
761 (serv
->sv_nrthreads
+3) * serv
->sv_max_mesg
,
762 (serv
->sv_nrthreads
+3) * serv
->sv_max_mesg
);
764 if ((rqstp
->rq_deferred
= svc_deferred_dequeue(svsk
))) {
765 svc_sock_received(svsk
);
766 return svc_deferred_recv(rqstp
);
769 clear_bit(SK_DATA
, &svsk
->sk_flags
);
771 err
= kernel_recvmsg(svsk
->sk_sock
, &msg
, NULL
,
772 0, 0, MSG_PEEK
| MSG_DONTWAIT
);
774 skb
= skb_recv_datagram(svsk
->sk_sk
, 0, 1, &err
);
777 if (err
!= -EAGAIN
) {
778 /* possibly an icmp error */
779 dprintk("svc: recvfrom returned error %d\n", -err
);
780 set_bit(SK_DATA
, &svsk
->sk_flags
);
782 svc_sock_received(svsk
);
785 rqstp
->rq_addrlen
= sizeof(rqstp
->rq_addr
);
786 if (skb
->tstamp
.tv64
== 0) {
787 skb
->tstamp
= ktime_get_real();
788 /* Don't enable netstamp, sunrpc doesn't
789 need that much accuracy */
791 svsk
->sk_sk
->sk_stamp
= skb
->tstamp
;
792 set_bit(SK_DATA
, &svsk
->sk_flags
); /* there may be more data... */
795 * Maybe more packets - kick another thread ASAP.
797 svc_sock_received(svsk
);
799 len
= skb
->len
- sizeof(struct udphdr
);
800 rqstp
->rq_arg
.len
= len
;
802 rqstp
->rq_prot
= IPPROTO_UDP
;
804 if (cmh
->cmsg_level
!= IPPROTO_IP
||
805 cmh
->cmsg_type
!= IP_PKTINFO
) {
807 printk("rpcsvc: received unknown control message:"
809 cmh
->cmsg_level
, cmh
->cmsg_type
);
810 skb_free_datagram(svsk
->sk_sk
, skb
);
813 svc_udp_get_dest_address(rqstp
, cmh
);
815 if (skb_is_nonlinear(skb
)) {
816 /* we have to copy */
818 if (csum_partial_copy_to_xdr(&rqstp
->rq_arg
, skb
)) {
821 skb_free_datagram(svsk
->sk_sk
, skb
);
825 skb_free_datagram(svsk
->sk_sk
, skb
);
827 /* we can use it in-place */
828 rqstp
->rq_arg
.head
[0].iov_base
= skb
->data
+ sizeof(struct udphdr
);
829 rqstp
->rq_arg
.head
[0].iov_len
= len
;
830 if (skb_checksum_complete(skb
)) {
831 skb_free_datagram(svsk
->sk_sk
, skb
);
834 rqstp
->rq_xprt_ctxt
= skb
;
837 rqstp
->rq_arg
.page_base
= 0;
838 if (len
<= rqstp
->rq_arg
.head
[0].iov_len
) {
839 rqstp
->rq_arg
.head
[0].iov_len
= len
;
840 rqstp
->rq_arg
.page_len
= 0;
841 rqstp
->rq_respages
= rqstp
->rq_pages
+1;
843 rqstp
->rq_arg
.page_len
= len
- rqstp
->rq_arg
.head
[0].iov_len
;
844 rqstp
->rq_respages
= rqstp
->rq_pages
+ 1 +
845 DIV_ROUND_UP(rqstp
->rq_arg
.page_len
, PAGE_SIZE
);
849 serv
->sv_stats
->netudpcnt
++;
855 svc_udp_sendto(struct svc_rqst
*rqstp
)
859 error
= svc_sendto(rqstp
, &rqstp
->rq_res
);
860 if (error
== -ECONNREFUSED
)
861 /* ICMP error on earlier request. */
862 error
= svc_sendto(rqstp
, &rqstp
->rq_res
);
867 static void svc_udp_prep_reply_hdr(struct svc_rqst
*rqstp
)
871 static int svc_udp_has_wspace(struct svc_xprt
*xprt
)
873 struct svc_sock
*svsk
= container_of(xprt
, struct svc_sock
, sk_xprt
);
874 struct svc_serv
*serv
= svsk
->sk_server
;
875 unsigned long required
;
878 * Set the SOCK_NOSPACE flag before checking the available
881 set_bit(SOCK_NOSPACE
, &svsk
->sk_sock
->flags
);
882 required
= atomic_read(&svsk
->sk_reserved
) + serv
->sv_max_mesg
;
883 if (required
*2 > sock_wspace(svsk
->sk_sk
))
885 clear_bit(SOCK_NOSPACE
, &svsk
->sk_sock
->flags
);
889 static struct svc_xprt
*svc_udp_accept(struct svc_xprt
*xprt
)
895 static struct svc_xprt
*svc_udp_create(struct svc_serv
*serv
,
896 struct sockaddr
*sa
, int salen
,
899 return svc_create_socket(serv
, IPPROTO_UDP
, sa
, salen
, flags
);
902 static struct svc_xprt_ops svc_udp_ops
= {
903 .xpo_create
= svc_udp_create
,
904 .xpo_recvfrom
= svc_udp_recvfrom
,
905 .xpo_sendto
= svc_udp_sendto
,
906 .xpo_release_rqst
= svc_release_skb
,
907 .xpo_detach
= svc_sock_detach
,
908 .xpo_free
= svc_sock_free
,
909 .xpo_prep_reply_hdr
= svc_udp_prep_reply_hdr
,
910 .xpo_has_wspace
= svc_udp_has_wspace
,
911 .xpo_accept
= svc_udp_accept
,
914 static struct svc_xprt_class svc_udp_class
= {
916 .xcl_owner
= THIS_MODULE
,
917 .xcl_ops
= &svc_udp_ops
,
918 .xcl_max_payload
= RPCSVC_MAXPAYLOAD_UDP
,
922 svc_udp_init(struct svc_sock
*svsk
)
927 svc_xprt_init(&svc_udp_class
, &svsk
->sk_xprt
);
928 svsk
->sk_sk
->sk_data_ready
= svc_udp_data_ready
;
929 svsk
->sk_sk
->sk_write_space
= svc_write_space
;
931 /* initialise setting must have enough space to
932 * receive and respond to one request.
933 * svc_udp_recvfrom will re-adjust if necessary
935 svc_sock_setbufsize(svsk
->sk_sock
,
936 3 * svsk
->sk_server
->sv_max_mesg
,
937 3 * svsk
->sk_server
->sv_max_mesg
);
939 set_bit(SK_DATA
, &svsk
->sk_flags
); /* might have come in before data_ready set up */
940 set_bit(SK_CHNGBUF
, &svsk
->sk_flags
);
944 /* make sure we get destination address info */
945 svsk
->sk_sock
->ops
->setsockopt(svsk
->sk_sock
, IPPROTO_IP
, IP_PKTINFO
,
946 (char __user
*)&one
, sizeof(one
));
951 * A data_ready event on a listening socket means there's a connection
952 * pending. Do not use state_change as a substitute for it.
955 svc_tcp_listen_data_ready(struct sock
*sk
, int count_unused
)
957 struct svc_sock
*svsk
= (struct svc_sock
*)sk
->sk_user_data
;
959 dprintk("svc: socket %p TCP (listen) state change %d\n",
963 * This callback may called twice when a new connection
964 * is established as a child socket inherits everything
965 * from a parent LISTEN socket.
966 * 1) data_ready method of the parent socket will be called
967 * when one of child sockets become ESTABLISHED.
968 * 2) data_ready method of the child socket may be called
969 * when it receives data before the socket is accepted.
970 * In case of 2, we should ignore it silently.
972 if (sk
->sk_state
== TCP_LISTEN
) {
974 set_bit(SK_CONN
, &svsk
->sk_flags
);
975 svc_sock_enqueue(svsk
);
977 printk("svc: socket %p: no user data\n", sk
);
980 if (sk
->sk_sleep
&& waitqueue_active(sk
->sk_sleep
))
981 wake_up_interruptible_all(sk
->sk_sleep
);
985 * A state change on a connected socket means it's dying or dead.
988 svc_tcp_state_change(struct sock
*sk
)
990 struct svc_sock
*svsk
= (struct svc_sock
*)sk
->sk_user_data
;
992 dprintk("svc: socket %p TCP (connected) state change %d (svsk %p)\n",
993 sk
, sk
->sk_state
, sk
->sk_user_data
);
996 printk("svc: socket %p: no user data\n", sk
);
998 set_bit(SK_CLOSE
, &svsk
->sk_flags
);
999 svc_sock_enqueue(svsk
);
1001 if (sk
->sk_sleep
&& waitqueue_active(sk
->sk_sleep
))
1002 wake_up_interruptible_all(sk
->sk_sleep
);
1006 svc_tcp_data_ready(struct sock
*sk
, int count
)
1008 struct svc_sock
*svsk
= (struct svc_sock
*)sk
->sk_user_data
;
1010 dprintk("svc: socket %p TCP data ready (svsk %p)\n",
1011 sk
, sk
->sk_user_data
);
1013 set_bit(SK_DATA
, &svsk
->sk_flags
);
1014 svc_sock_enqueue(svsk
);
1016 if (sk
->sk_sleep
&& waitqueue_active(sk
->sk_sleep
))
1017 wake_up_interruptible(sk
->sk_sleep
);
1020 static inline int svc_port_is_privileged(struct sockaddr
*sin
)
1022 switch (sin
->sa_family
) {
1024 return ntohs(((struct sockaddr_in
*)sin
)->sin_port
)
1027 return ntohs(((struct sockaddr_in6
*)sin
)->sin6_port
)
1035 * Accept a TCP connection
1037 static struct svc_xprt
*svc_tcp_accept(struct svc_xprt
*xprt
)
1039 struct svc_sock
*svsk
= container_of(xprt
, struct svc_sock
, sk_xprt
);
1040 struct sockaddr_storage addr
;
1041 struct sockaddr
*sin
= (struct sockaddr
*) &addr
;
1042 struct svc_serv
*serv
= svsk
->sk_server
;
1043 struct socket
*sock
= svsk
->sk_sock
;
1044 struct socket
*newsock
;
1045 struct svc_sock
*newsvsk
;
1047 char buf
[RPC_MAX_ADDRBUFLEN
];
1049 dprintk("svc: tcp_accept %p sock %p\n", svsk
, sock
);
1053 clear_bit(SK_CONN
, &svsk
->sk_flags
);
1054 err
= kernel_accept(sock
, &newsock
, O_NONBLOCK
);
1057 printk(KERN_WARNING
"%s: no more sockets!\n",
1059 else if (err
!= -EAGAIN
&& net_ratelimit())
1060 printk(KERN_WARNING
"%s: accept failed (err %d)!\n",
1061 serv
->sv_name
, -err
);
1065 set_bit(SK_CONN
, &svsk
->sk_flags
);
1067 err
= kernel_getpeername(newsock
, sin
, &slen
);
1069 if (net_ratelimit())
1070 printk(KERN_WARNING
"%s: peername failed (err %d)!\n",
1071 serv
->sv_name
, -err
);
1072 goto failed
; /* aborted connection or whatever */
1075 /* Ideally, we would want to reject connections from unauthorized
1076 * hosts here, but when we get encryption, the IP of the host won't
1077 * tell us anything. For now just warn about unpriv connections.
1079 if (!svc_port_is_privileged(sin
)) {
1080 dprintk(KERN_WARNING
1081 "%s: connect from unprivileged port: %s\n",
1083 __svc_print_addr(sin
, buf
, sizeof(buf
)));
1085 dprintk("%s: connect from %s\n", serv
->sv_name
,
1086 __svc_print_addr(sin
, buf
, sizeof(buf
)));
1088 /* make sure that a write doesn't block forever when
1091 newsock
->sk
->sk_sndtimeo
= HZ
*30;
1093 if (!(newsvsk
= svc_setup_socket(serv
, newsock
, &err
,
1094 (SVC_SOCK_ANONYMOUS
| SVC_SOCK_TEMPORARY
))))
1096 memcpy(&newsvsk
->sk_remote
, sin
, slen
);
1097 newsvsk
->sk_remotelen
= slen
;
1098 err
= kernel_getsockname(newsock
, sin
, &slen
);
1099 if (unlikely(err
< 0)) {
1100 dprintk("svc_tcp_accept: kernel_getsockname error %d\n", -err
);
1101 slen
= offsetof(struct sockaddr
, sa_data
);
1103 memcpy(&newsvsk
->sk_local
, sin
, slen
);
1105 svc_sock_received(newsvsk
);
1108 serv
->sv_stats
->nettcpconn
++;
1110 return &newsvsk
->sk_xprt
;
1113 sock_release(newsock
);
1118 * Receive data from a TCP socket.
1121 svc_tcp_recvfrom(struct svc_rqst
*rqstp
)
1123 struct svc_sock
*svsk
= rqstp
->rq_sock
;
1124 struct svc_serv
*serv
= svsk
->sk_server
;
1129 dprintk("svc: tcp_recv %p data %d conn %d close %d\n",
1130 svsk
, test_bit(SK_DATA
, &svsk
->sk_flags
),
1131 test_bit(SK_CONN
, &svsk
->sk_flags
),
1132 test_bit(SK_CLOSE
, &svsk
->sk_flags
));
1134 if ((rqstp
->rq_deferred
= svc_deferred_dequeue(svsk
))) {
1135 svc_sock_received(svsk
);
1136 return svc_deferred_recv(rqstp
);
1139 if (test_and_clear_bit(SK_CHNGBUF
, &svsk
->sk_flags
))
1140 /* sndbuf needs to have room for one request
1141 * per thread, otherwise we can stall even when the
1142 * network isn't a bottleneck.
1144 * We count all threads rather than threads in a
1145 * particular pool, which provides an upper bound
1146 * on the number of threads which will access the socket.
1148 * rcvbuf just needs to be able to hold a few requests.
1149 * Normally they will be removed from the queue
1150 * as soon a a complete request arrives.
1152 svc_sock_setbufsize(svsk
->sk_sock
,
1153 (serv
->sv_nrthreads
+3) * serv
->sv_max_mesg
,
1154 3 * serv
->sv_max_mesg
);
1156 clear_bit(SK_DATA
, &svsk
->sk_flags
);
1158 /* Receive data. If we haven't got the record length yet, get
1159 * the next four bytes. Otherwise try to gobble up as much as
1160 * possible up to the complete record length.
1162 if (svsk
->sk_tcplen
< 4) {
1163 unsigned long want
= 4 - svsk
->sk_tcplen
;
1166 iov
.iov_base
= ((char *) &svsk
->sk_reclen
) + svsk
->sk_tcplen
;
1168 if ((len
= svc_recvfrom(rqstp
, &iov
, 1, want
)) < 0)
1170 svsk
->sk_tcplen
+= len
;
1173 dprintk("svc: short recvfrom while reading record length (%d of %lu)\n",
1175 svc_sock_received(svsk
);
1176 return -EAGAIN
; /* record header not complete */
1179 svsk
->sk_reclen
= ntohl(svsk
->sk_reclen
);
1180 if (!(svsk
->sk_reclen
& 0x80000000)) {
1181 /* FIXME: technically, a record can be fragmented,
1182 * and non-terminal fragments will not have the top
1183 * bit set in the fragment length header.
1184 * But apparently no known nfs clients send fragmented
1186 if (net_ratelimit())
1187 printk(KERN_NOTICE
"RPC: bad TCP reclen 0x%08lx"
1188 " (non-terminal)\n",
1189 (unsigned long) svsk
->sk_reclen
);
1192 svsk
->sk_reclen
&= 0x7fffffff;
1193 dprintk("svc: TCP record, %d bytes\n", svsk
->sk_reclen
);
1194 if (svsk
->sk_reclen
> serv
->sv_max_mesg
) {
1195 if (net_ratelimit())
1196 printk(KERN_NOTICE
"RPC: bad TCP reclen 0x%08lx"
1198 (unsigned long) svsk
->sk_reclen
);
1203 /* Check whether enough data is available */
1204 len
= svc_recv_available(svsk
);
1208 if (len
< svsk
->sk_reclen
) {
1209 dprintk("svc: incomplete TCP record (%d of %d)\n",
1210 len
, svsk
->sk_reclen
);
1211 svc_sock_received(svsk
);
1212 return -EAGAIN
; /* record not complete */
1214 len
= svsk
->sk_reclen
;
1215 set_bit(SK_DATA
, &svsk
->sk_flags
);
1217 vec
= rqstp
->rq_vec
;
1218 vec
[0] = rqstp
->rq_arg
.head
[0];
1221 while (vlen
< len
) {
1222 vec
[pnum
].iov_base
= page_address(rqstp
->rq_pages
[pnum
]);
1223 vec
[pnum
].iov_len
= PAGE_SIZE
;
1227 rqstp
->rq_respages
= &rqstp
->rq_pages
[pnum
];
1229 /* Now receive data */
1230 len
= svc_recvfrom(rqstp
, vec
, pnum
, len
);
1234 dprintk("svc: TCP complete record (%d bytes)\n", len
);
1235 rqstp
->rq_arg
.len
= len
;
1236 rqstp
->rq_arg
.page_base
= 0;
1237 if (len
<= rqstp
->rq_arg
.head
[0].iov_len
) {
1238 rqstp
->rq_arg
.head
[0].iov_len
= len
;
1239 rqstp
->rq_arg
.page_len
= 0;
1241 rqstp
->rq_arg
.page_len
= len
- rqstp
->rq_arg
.head
[0].iov_len
;
1244 rqstp
->rq_xprt_ctxt
= NULL
;
1245 rqstp
->rq_prot
= IPPROTO_TCP
;
1247 /* Reset TCP read info */
1248 svsk
->sk_reclen
= 0;
1249 svsk
->sk_tcplen
= 0;
1251 svc_sock_received(svsk
);
1253 serv
->sv_stats
->nettcpcnt
++;
1258 set_bit(SK_CLOSE
, &svsk
->sk_flags
);
1262 if (len
== -EAGAIN
) {
1263 dprintk("RPC: TCP recvfrom got EAGAIN\n");
1264 svc_sock_received(svsk
);
1266 printk(KERN_NOTICE
"%s: recvfrom returned errno %d\n",
1267 svsk
->sk_server
->sv_name
, -len
);
1275 * Send out data on TCP socket.
1278 svc_tcp_sendto(struct svc_rqst
*rqstp
)
1280 struct xdr_buf
*xbufp
= &rqstp
->rq_res
;
1284 /* Set up the first element of the reply kvec.
1285 * Any other kvecs that may be in use have been taken
1286 * care of by the server implementation itself.
1288 reclen
= htonl(0x80000000|((xbufp
->len
) - 4));
1289 memcpy(xbufp
->head
[0].iov_base
, &reclen
, 4);
1291 if (test_bit(SK_DEAD
, &rqstp
->rq_sock
->sk_flags
))
1294 sent
= svc_sendto(rqstp
, &rqstp
->rq_res
);
1295 if (sent
!= xbufp
->len
) {
1296 printk(KERN_NOTICE
"rpc-srv/tcp: %s: %s %d when sending %d bytes - shutting down socket\n",
1297 rqstp
->rq_sock
->sk_server
->sv_name
,
1298 (sent
<0)?"got error":"sent only",
1300 set_bit(SK_CLOSE
, &rqstp
->rq_sock
->sk_flags
);
1301 svc_sock_enqueue(rqstp
->rq_sock
);
1308 * Setup response header. TCP has a 4B record length field.
1310 static void svc_tcp_prep_reply_hdr(struct svc_rqst
*rqstp
)
1312 struct kvec
*resv
= &rqstp
->rq_res
.head
[0];
1314 /* tcp needs a space for the record length... */
1318 static int svc_tcp_has_wspace(struct svc_xprt
*xprt
)
1320 struct svc_sock
*svsk
= container_of(xprt
, struct svc_sock
, sk_xprt
);
1321 struct svc_serv
*serv
= svsk
->sk_server
;
1326 * Set the SOCK_NOSPACE flag before checking the available
1329 set_bit(SOCK_NOSPACE
, &svsk
->sk_sock
->flags
);
1330 required
= atomic_read(&svsk
->sk_reserved
) + serv
->sv_max_mesg
;
1331 wspace
= sk_stream_wspace(svsk
->sk_sk
);
1333 if (wspace
< sk_stream_min_wspace(svsk
->sk_sk
))
1335 if (required
* 2 > wspace
)
1338 clear_bit(SOCK_NOSPACE
, &svsk
->sk_sock
->flags
);
1342 static struct svc_xprt
*svc_tcp_create(struct svc_serv
*serv
,
1343 struct sockaddr
*sa
, int salen
,
1346 return svc_create_socket(serv
, IPPROTO_TCP
, sa
, salen
, flags
);
1349 static struct svc_xprt_ops svc_tcp_ops
= {
1350 .xpo_create
= svc_tcp_create
,
1351 .xpo_recvfrom
= svc_tcp_recvfrom
,
1352 .xpo_sendto
= svc_tcp_sendto
,
1353 .xpo_release_rqst
= svc_release_skb
,
1354 .xpo_detach
= svc_sock_detach
,
1355 .xpo_free
= svc_sock_free
,
1356 .xpo_prep_reply_hdr
= svc_tcp_prep_reply_hdr
,
1357 .xpo_has_wspace
= svc_tcp_has_wspace
,
1358 .xpo_accept
= svc_tcp_accept
,
1361 static struct svc_xprt_class svc_tcp_class
= {
1363 .xcl_owner
= THIS_MODULE
,
1364 .xcl_ops
= &svc_tcp_ops
,
1365 .xcl_max_payload
= RPCSVC_MAXPAYLOAD_TCP
,
1368 void svc_init_xprt_sock(void)
1370 svc_reg_xprt_class(&svc_tcp_class
);
1371 svc_reg_xprt_class(&svc_udp_class
);
1374 void svc_cleanup_xprt_sock(void)
1376 svc_unreg_xprt_class(&svc_tcp_class
);
1377 svc_unreg_xprt_class(&svc_udp_class
);
1381 svc_tcp_init(struct svc_sock
*svsk
)
1383 struct sock
*sk
= svsk
->sk_sk
;
1384 struct tcp_sock
*tp
= tcp_sk(sk
);
1386 svc_xprt_init(&svc_tcp_class
, &svsk
->sk_xprt
);
1388 if (sk
->sk_state
== TCP_LISTEN
) {
1389 dprintk("setting up TCP socket for listening\n");
1390 set_bit(SK_LISTENER
, &svsk
->sk_flags
);
1391 sk
->sk_data_ready
= svc_tcp_listen_data_ready
;
1392 set_bit(SK_CONN
, &svsk
->sk_flags
);
1394 dprintk("setting up TCP socket for reading\n");
1395 sk
->sk_state_change
= svc_tcp_state_change
;
1396 sk
->sk_data_ready
= svc_tcp_data_ready
;
1397 sk
->sk_write_space
= svc_write_space
;
1399 svsk
->sk_reclen
= 0;
1400 svsk
->sk_tcplen
= 0;
1402 tp
->nonagle
= 1; /* disable Nagle's algorithm */
1404 /* initialise setting must have enough space to
1405 * receive and respond to one request.
1406 * svc_tcp_recvfrom will re-adjust if necessary
1408 svc_sock_setbufsize(svsk
->sk_sock
,
1409 3 * svsk
->sk_server
->sv_max_mesg
,
1410 3 * svsk
->sk_server
->sv_max_mesg
);
1412 set_bit(SK_CHNGBUF
, &svsk
->sk_flags
);
1413 set_bit(SK_DATA
, &svsk
->sk_flags
);
1414 if (sk
->sk_state
!= TCP_ESTABLISHED
)
1415 set_bit(SK_CLOSE
, &svsk
->sk_flags
);
1420 svc_sock_update_bufs(struct svc_serv
*serv
)
1423 * The number of server threads has changed. Update
1424 * rcvbuf and sndbuf accordingly on all sockets
1426 struct list_head
*le
;
1428 spin_lock_bh(&serv
->sv_lock
);
1429 list_for_each(le
, &serv
->sv_permsocks
) {
1430 struct svc_sock
*svsk
=
1431 list_entry(le
, struct svc_sock
, sk_list
);
1432 set_bit(SK_CHNGBUF
, &svsk
->sk_flags
);
1434 list_for_each(le
, &serv
->sv_tempsocks
) {
1435 struct svc_sock
*svsk
=
1436 list_entry(le
, struct svc_sock
, sk_list
);
1437 set_bit(SK_CHNGBUF
, &svsk
->sk_flags
);
1439 spin_unlock_bh(&serv
->sv_lock
);
1443 * Make sure that we don't have too many active connections. If we
1444 * have, something must be dropped.
1446 * There's no point in trying to do random drop here for DoS
1447 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
1448 * attacker can easily beat that.
1450 * The only somewhat efficient mechanism would be if drop old
1451 * connections from the same IP first. But right now we don't even
1452 * record the client IP in svc_sock.
1454 static void svc_check_conn_limits(struct svc_serv
*serv
)
1456 if (serv
->sv_tmpcnt
> (serv
->sv_nrthreads
+3)*20) {
1457 struct svc_sock
*svsk
= NULL
;
1458 spin_lock_bh(&serv
->sv_lock
);
1459 if (!list_empty(&serv
->sv_tempsocks
)) {
1460 if (net_ratelimit()) {
1461 /* Try to help the admin */
1462 printk(KERN_NOTICE
"%s: too many open TCP "
1463 "sockets, consider increasing the "
1464 "number of nfsd threads\n",
1468 * Always select the oldest socket. It's not fair,
1471 svsk
= list_entry(serv
->sv_tempsocks
.prev
,
1474 set_bit(SK_CLOSE
, &svsk
->sk_flags
);
1475 svc_xprt_get(&svsk
->sk_xprt
);
1477 spin_unlock_bh(&serv
->sv_lock
);
1480 svc_sock_enqueue(svsk
);
1481 svc_xprt_put(&svsk
->sk_xprt
);
1487 * Receive the next request on any socket. This code is carefully
1488 * organised not to touch any cachelines in the shared svc_serv
1489 * structure, only cachelines in the local svc_pool.
1492 svc_recv(struct svc_rqst
*rqstp
, long timeout
)
1494 struct svc_sock
*svsk
= NULL
;
1495 struct svc_serv
*serv
= rqstp
->rq_server
;
1496 struct svc_pool
*pool
= rqstp
->rq_pool
;
1499 struct xdr_buf
*arg
;
1500 DECLARE_WAITQUEUE(wait
, current
);
1502 dprintk("svc: server %p waiting for data (to = %ld)\n",
1507 "svc_recv: service %p, socket not NULL!\n",
1509 if (waitqueue_active(&rqstp
->rq_wait
))
1511 "svc_recv: service %p, wait queue active!\n",
1515 /* now allocate needed pages. If we get a failure, sleep briefly */
1516 pages
= (serv
->sv_max_mesg
+ PAGE_SIZE
) / PAGE_SIZE
;
1517 for (i
=0; i
< pages
; i
++)
1518 while (rqstp
->rq_pages
[i
] == NULL
) {
1519 struct page
*p
= alloc_page(GFP_KERNEL
);
1521 schedule_timeout_uninterruptible(msecs_to_jiffies(500));
1522 rqstp
->rq_pages
[i
] = p
;
1524 rqstp
->rq_pages
[i
++] = NULL
; /* this might be seen in nfs_read_actor */
1525 BUG_ON(pages
>= RPCSVC_MAXPAGES
);
1527 /* Make arg->head point to first page and arg->pages point to rest */
1528 arg
= &rqstp
->rq_arg
;
1529 arg
->head
[0].iov_base
= page_address(rqstp
->rq_pages
[0]);
1530 arg
->head
[0].iov_len
= PAGE_SIZE
;
1531 arg
->pages
= rqstp
->rq_pages
+ 1;
1533 /* save at least one page for response */
1534 arg
->page_len
= (pages
-2)*PAGE_SIZE
;
1535 arg
->len
= (pages
-1)*PAGE_SIZE
;
1536 arg
->tail
[0].iov_len
= 0;
1543 spin_lock_bh(&pool
->sp_lock
);
1544 if ((svsk
= svc_sock_dequeue(pool
)) != NULL
) {
1545 rqstp
->rq_sock
= svsk
;
1546 svc_xprt_get(&svsk
->sk_xprt
);
1547 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
1548 atomic_add(rqstp
->rq_reserved
, &svsk
->sk_reserved
);
1550 /* No data pending. Go to sleep */
1551 svc_thread_enqueue(pool
, rqstp
);
1554 * We have to be able to interrupt this wait
1555 * to bring down the daemons ...
1557 set_current_state(TASK_INTERRUPTIBLE
);
1558 add_wait_queue(&rqstp
->rq_wait
, &wait
);
1559 spin_unlock_bh(&pool
->sp_lock
);
1561 schedule_timeout(timeout
);
1565 spin_lock_bh(&pool
->sp_lock
);
1566 remove_wait_queue(&rqstp
->rq_wait
, &wait
);
1568 if (!(svsk
= rqstp
->rq_sock
)) {
1569 svc_thread_dequeue(pool
, rqstp
);
1570 spin_unlock_bh(&pool
->sp_lock
);
1571 dprintk("svc: server %p, no data yet\n", rqstp
);
1572 return signalled()? -EINTR
: -EAGAIN
;
1575 spin_unlock_bh(&pool
->sp_lock
);
1578 if (test_bit(SK_CLOSE
, &svsk
->sk_flags
)) {
1579 dprintk("svc_recv: found SK_CLOSE\n");
1580 svc_delete_socket(svsk
);
1581 } else if (test_bit(SK_LISTENER
, &svsk
->sk_flags
)) {
1582 struct svc_xprt
*newxpt
;
1583 newxpt
= svsk
->sk_xprt
.xpt_ops
->xpo_accept(&svsk
->sk_xprt
);
1586 * We know this module_get will succeed because the
1587 * listener holds a reference too
1589 __module_get(newxpt
->xpt_class
->xcl_owner
);
1590 svc_check_conn_limits(svsk
->sk_server
);
1592 svc_sock_received(svsk
);
1594 dprintk("svc: server %p, pool %u, socket %p, inuse=%d\n",
1595 rqstp
, pool
->sp_id
, svsk
,
1596 atomic_read(&svsk
->sk_xprt
.xpt_ref
.refcount
));
1597 len
= svsk
->sk_xprt
.xpt_ops
->xpo_recvfrom(rqstp
);
1598 dprintk("svc: got len=%d\n", len
);
1601 /* No data, incomplete (TCP) read, or accept() */
1602 if (len
== 0 || len
== -EAGAIN
) {
1603 rqstp
->rq_res
.len
= 0;
1604 svc_sock_release(rqstp
);
1607 svsk
->sk_lastrecv
= get_seconds();
1608 clear_bit(SK_OLD
, &svsk
->sk_flags
);
1610 rqstp
->rq_secure
= svc_port_is_privileged(svc_addr(rqstp
));
1611 rqstp
->rq_chandle
.defer
= svc_defer
;
1614 serv
->sv_stats
->netcnt
++;
1622 svc_drop(struct svc_rqst
*rqstp
)
1624 dprintk("svc: socket %p dropped request\n", rqstp
->rq_sock
);
1625 svc_sock_release(rqstp
);
1629 * Return reply to client.
1632 svc_send(struct svc_rqst
*rqstp
)
1634 struct svc_sock
*svsk
;
1638 if ((svsk
= rqstp
->rq_sock
) == NULL
) {
1639 printk(KERN_WARNING
"NULL socket pointer in %s:%d\n",
1640 __FILE__
, __LINE__
);
1644 /* release the receive skb before sending the reply */
1645 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
1647 /* calculate over-all length */
1648 xb
= & rqstp
->rq_res
;
1649 xb
->len
= xb
->head
[0].iov_len
+
1651 xb
->tail
[0].iov_len
;
1653 /* Grab svsk->sk_mutex to serialize outgoing data. */
1654 mutex_lock(&svsk
->sk_mutex
);
1655 if (test_bit(SK_DEAD
, &svsk
->sk_flags
))
1658 len
= svsk
->sk_xprt
.xpt_ops
->xpo_sendto(rqstp
);
1659 mutex_unlock(&svsk
->sk_mutex
);
1660 svc_sock_release(rqstp
);
1662 if (len
== -ECONNREFUSED
|| len
== -ENOTCONN
|| len
== -EAGAIN
)
1668 * Timer function to close old temporary sockets, using
1669 * a mark-and-sweep algorithm.
1672 svc_age_temp_sockets(unsigned long closure
)
1674 struct svc_serv
*serv
= (struct svc_serv
*)closure
;
1675 struct svc_sock
*svsk
;
1676 struct list_head
*le
, *next
;
1677 LIST_HEAD(to_be_aged
);
1679 dprintk("svc_age_temp_sockets\n");
1681 if (!spin_trylock_bh(&serv
->sv_lock
)) {
1682 /* busy, try again 1 sec later */
1683 dprintk("svc_age_temp_sockets: busy\n");
1684 mod_timer(&serv
->sv_temptimer
, jiffies
+ HZ
);
1688 list_for_each_safe(le
, next
, &serv
->sv_tempsocks
) {
1689 svsk
= list_entry(le
, struct svc_sock
, sk_list
);
1691 if (!test_and_set_bit(SK_OLD
, &svsk
->sk_flags
))
1693 if (atomic_read(&svsk
->sk_xprt
.xpt_ref
.refcount
) > 1
1694 || test_bit(SK_BUSY
, &svsk
->sk_flags
))
1696 svc_xprt_get(&svsk
->sk_xprt
);
1697 list_move(le
, &to_be_aged
);
1698 set_bit(SK_CLOSE
, &svsk
->sk_flags
);
1699 set_bit(SK_DETACHED
, &svsk
->sk_flags
);
1701 spin_unlock_bh(&serv
->sv_lock
);
1703 while (!list_empty(&to_be_aged
)) {
1704 le
= to_be_aged
.next
;
1705 /* fiddling the sk_list node is safe 'cos we're SK_DETACHED */
1707 svsk
= list_entry(le
, struct svc_sock
, sk_list
);
1709 dprintk("queuing svsk %p for closing, %lu seconds old\n",
1710 svsk
, get_seconds() - svsk
->sk_lastrecv
);
1712 /* a thread will dequeue and close it soon */
1713 svc_sock_enqueue(svsk
);
1714 svc_xprt_put(&svsk
->sk_xprt
);
1717 mod_timer(&serv
->sv_temptimer
, jiffies
+ svc_conn_age_period
* HZ
);
1721 * Initialize socket for RPC use and create svc_sock struct
1722 * XXX: May want to setsockopt SO_SNDBUF and SO_RCVBUF.
1724 static struct svc_sock
*svc_setup_socket(struct svc_serv
*serv
,
1725 struct socket
*sock
,
1726 int *errp
, int flags
)
1728 struct svc_sock
*svsk
;
1730 int pmap_register
= !(flags
& SVC_SOCK_ANONYMOUS
);
1731 int is_temporary
= flags
& SVC_SOCK_TEMPORARY
;
1733 dprintk("svc: svc_setup_socket %p\n", sock
);
1734 if (!(svsk
= kzalloc(sizeof(*svsk
), GFP_KERNEL
))) {
1741 /* Register socket with portmapper */
1742 if (*errp
>= 0 && pmap_register
)
1743 *errp
= svc_register(serv
, inet
->sk_protocol
,
1744 ntohs(inet_sk(inet
)->sport
));
1751 set_bit(SK_BUSY
, &svsk
->sk_flags
);
1752 inet
->sk_user_data
= svsk
;
1753 svsk
->sk_sock
= sock
;
1755 svsk
->sk_ostate
= inet
->sk_state_change
;
1756 svsk
->sk_odata
= inet
->sk_data_ready
;
1757 svsk
->sk_owspace
= inet
->sk_write_space
;
1758 svsk
->sk_server
= serv
;
1759 svsk
->sk_lastrecv
= get_seconds();
1760 spin_lock_init(&svsk
->sk_lock
);
1761 INIT_LIST_HEAD(&svsk
->sk_deferred
);
1762 INIT_LIST_HEAD(&svsk
->sk_ready
);
1763 mutex_init(&svsk
->sk_mutex
);
1765 /* Initialize the socket */
1766 if (sock
->type
== SOCK_DGRAM
)
1771 spin_lock_bh(&serv
->sv_lock
);
1773 set_bit(SK_TEMP
, &svsk
->sk_flags
);
1774 list_add(&svsk
->sk_list
, &serv
->sv_tempsocks
);
1776 if (serv
->sv_temptimer
.function
== NULL
) {
1777 /* setup timer to age temp sockets */
1778 setup_timer(&serv
->sv_temptimer
, svc_age_temp_sockets
,
1779 (unsigned long)serv
);
1780 mod_timer(&serv
->sv_temptimer
,
1781 jiffies
+ svc_conn_age_period
* HZ
);
1784 clear_bit(SK_TEMP
, &svsk
->sk_flags
);
1785 list_add(&svsk
->sk_list
, &serv
->sv_permsocks
);
1787 spin_unlock_bh(&serv
->sv_lock
);
1789 dprintk("svc: svc_setup_socket created %p (inet %p)\n",
1795 int svc_addsock(struct svc_serv
*serv
,
1801 struct socket
*so
= sockfd_lookup(fd
, &err
);
1802 struct svc_sock
*svsk
= NULL
;
1806 if (so
->sk
->sk_family
!= AF_INET
)
1807 err
= -EAFNOSUPPORT
;
1808 else if (so
->sk
->sk_protocol
!= IPPROTO_TCP
&&
1809 so
->sk
->sk_protocol
!= IPPROTO_UDP
)
1810 err
= -EPROTONOSUPPORT
;
1811 else if (so
->state
> SS_UNCONNECTED
)
1814 svsk
= svc_setup_socket(serv
, so
, &err
, SVC_SOCK_DEFAULTS
);
1816 svc_sock_received(svsk
);
1824 if (proto
) *proto
= so
->sk
->sk_protocol
;
1825 return one_sock_name(name_return
, svsk
);
1827 EXPORT_SYMBOL_GPL(svc_addsock
);
1830 * Create socket for RPC service.
1832 static struct svc_xprt
*svc_create_socket(struct svc_serv
*serv
,
1834 struct sockaddr
*sin
, int len
,
1837 struct svc_sock
*svsk
;
1838 struct socket
*sock
;
1841 char buf
[RPC_MAX_ADDRBUFLEN
];
1843 dprintk("svc: svc_create_socket(%s, %d, %s)\n",
1844 serv
->sv_program
->pg_name
, protocol
,
1845 __svc_print_addr(sin
, buf
, sizeof(buf
)));
1847 if (protocol
!= IPPROTO_UDP
&& protocol
!= IPPROTO_TCP
) {
1848 printk(KERN_WARNING
"svc: only UDP and TCP "
1849 "sockets supported\n");
1850 return ERR_PTR(-EINVAL
);
1852 type
= (protocol
== IPPROTO_UDP
)? SOCK_DGRAM
: SOCK_STREAM
;
1854 error
= sock_create_kern(sin
->sa_family
, type
, protocol
, &sock
);
1856 return ERR_PTR(error
);
1858 svc_reclassify_socket(sock
);
1860 if (type
== SOCK_STREAM
)
1861 sock
->sk
->sk_reuse
= 1; /* allow address reuse */
1862 error
= kernel_bind(sock
, sin
, len
);
1866 if (protocol
== IPPROTO_TCP
) {
1867 if ((error
= kernel_listen(sock
, 64)) < 0)
1871 if ((svsk
= svc_setup_socket(serv
, sock
, &error
, flags
)) != NULL
) {
1872 svc_sock_received(svsk
);
1873 return (struct svc_xprt
*)svsk
;
1877 dprintk("svc: svc_create_socket error = %d\n", -error
);
1879 return ERR_PTR(error
);
1883 * Detach the svc_sock from the socket so that no
1884 * more callbacks occur.
1886 static void svc_sock_detach(struct svc_xprt
*xprt
)
1888 struct svc_sock
*svsk
= container_of(xprt
, struct svc_sock
, sk_xprt
);
1889 struct sock
*sk
= svsk
->sk_sk
;
1891 dprintk("svc: svc_sock_detach(%p)\n", svsk
);
1893 /* put back the old socket callbacks */
1894 sk
->sk_state_change
= svsk
->sk_ostate
;
1895 sk
->sk_data_ready
= svsk
->sk_odata
;
1896 sk
->sk_write_space
= svsk
->sk_owspace
;
1900 * Free the svc_sock's socket resources and the svc_sock itself.
1902 static void svc_sock_free(struct svc_xprt
*xprt
)
1904 struct svc_sock
*svsk
= container_of(xprt
, struct svc_sock
, sk_xprt
);
1905 dprintk("svc: svc_sock_free(%p)\n", svsk
);
1907 if (svsk
->sk_info_authunix
!= NULL
)
1908 svcauth_unix_info_release(svsk
->sk_info_authunix
);
1909 if (svsk
->sk_sock
->file
)
1910 sockfd_put(svsk
->sk_sock
);
1912 sock_release(svsk
->sk_sock
);
1917 * Remove a dead socket
1920 svc_delete_socket(struct svc_sock
*svsk
)
1922 struct svc_serv
*serv
;
1925 dprintk("svc: svc_delete_socket(%p)\n", svsk
);
1927 serv
= svsk
->sk_server
;
1930 svsk
->sk_xprt
.xpt_ops
->xpo_detach(&svsk
->sk_xprt
);
1932 spin_lock_bh(&serv
->sv_lock
);
1934 if (!test_and_set_bit(SK_DETACHED
, &svsk
->sk_flags
))
1935 list_del_init(&svsk
->sk_list
);
1937 * We used to delete the svc_sock from whichever list
1938 * it's sk_ready node was on, but we don't actually
1939 * need to. This is because the only time we're called
1940 * while still attached to a queue, the queue itself
1941 * is about to be destroyed (in svc_destroy).
1943 if (!test_and_set_bit(SK_DEAD
, &svsk
->sk_flags
)) {
1944 BUG_ON(atomic_read(&svsk
->sk_xprt
.xpt_ref
.refcount
) < 2);
1945 if (test_bit(SK_TEMP
, &svsk
->sk_flags
))
1947 svc_xprt_put(&svsk
->sk_xprt
);
1950 spin_unlock_bh(&serv
->sv_lock
);
1953 static void svc_close_socket(struct svc_sock
*svsk
)
1955 set_bit(SK_CLOSE
, &svsk
->sk_flags
);
1956 if (test_and_set_bit(SK_BUSY
, &svsk
->sk_flags
))
1957 /* someone else will have to effect the close */
1960 svc_xprt_get(&svsk
->sk_xprt
);
1961 svc_delete_socket(svsk
);
1962 clear_bit(SK_BUSY
, &svsk
->sk_flags
);
1963 svc_xprt_put(&svsk
->sk_xprt
);
1966 void svc_force_close_socket(struct svc_sock
*svsk
)
1968 set_bit(SK_CLOSE
, &svsk
->sk_flags
);
1969 if (test_bit(SK_BUSY
, &svsk
->sk_flags
)) {
1970 /* Waiting to be processed, but no threads left,
1971 * So just remove it from the waiting list
1973 list_del_init(&svsk
->sk_ready
);
1974 clear_bit(SK_BUSY
, &svsk
->sk_flags
);
1976 svc_close_socket(svsk
);
1980 * Handle defer and revisit of requests
1983 static void svc_revisit(struct cache_deferred_req
*dreq
, int too_many
)
1985 struct svc_deferred_req
*dr
= container_of(dreq
, struct svc_deferred_req
, handle
);
1986 struct svc_sock
*svsk
;
1989 svc_xprt_put(&dr
->svsk
->sk_xprt
);
1993 dprintk("revisit queued\n");
1996 spin_lock(&svsk
->sk_lock
);
1997 list_add(&dr
->handle
.recent
, &svsk
->sk_deferred
);
1998 spin_unlock(&svsk
->sk_lock
);
1999 set_bit(SK_DEFERRED
, &svsk
->sk_flags
);
2000 svc_sock_enqueue(svsk
);
2001 svc_xprt_put(&svsk
->sk_xprt
);
2004 static struct cache_deferred_req
*
2005 svc_defer(struct cache_req
*req
)
2007 struct svc_rqst
*rqstp
= container_of(req
, struct svc_rqst
, rq_chandle
);
2008 int size
= sizeof(struct svc_deferred_req
) + (rqstp
->rq_arg
.len
);
2009 struct svc_deferred_req
*dr
;
2011 if (rqstp
->rq_arg
.page_len
)
2012 return NULL
; /* if more than a page, give up FIXME */
2013 if (rqstp
->rq_deferred
) {
2014 dr
= rqstp
->rq_deferred
;
2015 rqstp
->rq_deferred
= NULL
;
2017 int skip
= rqstp
->rq_arg
.len
- rqstp
->rq_arg
.head
[0].iov_len
;
2018 /* FIXME maybe discard if size too large */
2019 dr
= kmalloc(size
, GFP_KERNEL
);
2023 dr
->handle
.owner
= rqstp
->rq_server
;
2024 dr
->prot
= rqstp
->rq_prot
;
2025 memcpy(&dr
->addr
, &rqstp
->rq_addr
, rqstp
->rq_addrlen
);
2026 dr
->addrlen
= rqstp
->rq_addrlen
;
2027 dr
->daddr
= rqstp
->rq_daddr
;
2028 dr
->argslen
= rqstp
->rq_arg
.len
>> 2;
2029 memcpy(dr
->args
, rqstp
->rq_arg
.head
[0].iov_base
-skip
, dr
->argslen
<<2);
2031 svc_xprt_get(rqstp
->rq_xprt
);
2032 dr
->svsk
= rqstp
->rq_sock
;
2034 dr
->handle
.revisit
= svc_revisit
;
2039 * recv data from a deferred request into an active one
2041 static int svc_deferred_recv(struct svc_rqst
*rqstp
)
2043 struct svc_deferred_req
*dr
= rqstp
->rq_deferred
;
2045 rqstp
->rq_arg
.head
[0].iov_base
= dr
->args
;
2046 rqstp
->rq_arg
.head
[0].iov_len
= dr
->argslen
<<2;
2047 rqstp
->rq_arg
.page_len
= 0;
2048 rqstp
->rq_arg
.len
= dr
->argslen
<<2;
2049 rqstp
->rq_prot
= dr
->prot
;
2050 memcpy(&rqstp
->rq_addr
, &dr
->addr
, dr
->addrlen
);
2051 rqstp
->rq_addrlen
= dr
->addrlen
;
2052 rqstp
->rq_daddr
= dr
->daddr
;
2053 rqstp
->rq_respages
= rqstp
->rq_pages
;
2054 return dr
->argslen
<<2;
2058 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_sock
*svsk
)
2060 struct svc_deferred_req
*dr
= NULL
;
2062 if (!test_bit(SK_DEFERRED
, &svsk
->sk_flags
))
2064 spin_lock(&svsk
->sk_lock
);
2065 clear_bit(SK_DEFERRED
, &svsk
->sk_flags
);
2066 if (!list_empty(&svsk
->sk_deferred
)) {
2067 dr
= list_entry(svsk
->sk_deferred
.next
,
2068 struct svc_deferred_req
,
2070 list_del_init(&dr
->handle
.recent
);
2071 set_bit(SK_DEFERRED
, &svsk
->sk_flags
);
2073 spin_unlock(&svsk
->sk_lock
);