2 * linux/net/sunrpc/svc_xprt.c
4 * Author: Tom Tucker <tom@opengridcomputing.com>
7 #include <linux/sched.h>
8 #include <linux/smp_lock.h>
9 #include <linux/errno.h>
10 #include <linux/freezer.h>
11 #include <linux/kthread.h>
13 #include <linux/sunrpc/stats.h>
14 #include <linux/sunrpc/svc_xprt.h>
15 #include <linux/sunrpc/svcsock.h>
17 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
19 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
);
20 static int svc_deferred_recv(struct svc_rqst
*rqstp
);
21 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
);
22 static void svc_age_temp_xprts(unsigned long closure
);
24 /* apparently the "standard" is that clients close
25 * idle connections after 5 minutes, servers after
27 * http://www.connectathon.org/talks96/nfstcp.pdf
29 static int svc_conn_age_period
= 6*60;
31 /* List of registered transport classes */
32 static DEFINE_SPINLOCK(svc_xprt_class_lock
);
33 static LIST_HEAD(svc_xprt_class_list
);
35 /* SMP locking strategy:
37 * svc_pool->sp_lock protects most of the fields of that pool.
38 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
39 * when both need to be taken (rare), svc_serv->sv_lock is first.
40 * BKL protects svc_serv->sv_nrthread.
41 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
42 * and the ->sk_info_authunix cache.
44 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
45 * enqueued multiply. During normal transport processing this bit
46 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
47 * Providers should not manipulate this bit directly.
49 * Some flags can be set to certain values at any time
50 * providing that certain rules are followed:
53 * - Can be set or cleared at any time.
54 * - After a set, svc_xprt_enqueue must be called to enqueue
55 * the transport for processing.
56 * - After a clear, the transport must be read/accepted.
57 * If this succeeds, it must be set again.
59 * - Can set at any time. It is never cleared.
61 * - Can only be set while XPT_BUSY is held which ensures
62 * that no other thread will be using the transport or will
63 * try to set XPT_DEAD.
66 int svc_reg_xprt_class(struct svc_xprt_class
*xcl
)
68 struct svc_xprt_class
*cl
;
71 dprintk("svc: Adding svc transport class '%s'\n", xcl
->xcl_name
);
73 INIT_LIST_HEAD(&xcl
->xcl_list
);
74 spin_lock(&svc_xprt_class_lock
);
75 /* Make sure there isn't already a class with the same name */
76 list_for_each_entry(cl
, &svc_xprt_class_list
, xcl_list
) {
77 if (strcmp(xcl
->xcl_name
, cl
->xcl_name
) == 0)
80 list_add_tail(&xcl
->xcl_list
, &svc_xprt_class_list
);
83 spin_unlock(&svc_xprt_class_lock
);
86 EXPORT_SYMBOL_GPL(svc_reg_xprt_class
);
88 void svc_unreg_xprt_class(struct svc_xprt_class
*xcl
)
90 dprintk("svc: Removing svc transport class '%s'\n", xcl
->xcl_name
);
91 spin_lock(&svc_xprt_class_lock
);
92 list_del_init(&xcl
->xcl_list
);
93 spin_unlock(&svc_xprt_class_lock
);
95 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class
);
98 * Format the transport list for printing
100 int svc_print_xprts(char *buf
, int maxlen
)
102 struct list_head
*le
;
107 spin_lock(&svc_xprt_class_lock
);
108 list_for_each(le
, &svc_xprt_class_list
) {
110 struct svc_xprt_class
*xcl
=
111 list_entry(le
, struct svc_xprt_class
, xcl_list
);
113 sprintf(tmpstr
, "%s %d\n", xcl
->xcl_name
, xcl
->xcl_max_payload
);
114 slen
= strlen(tmpstr
);
115 if (len
+ slen
> maxlen
)
120 spin_unlock(&svc_xprt_class_lock
);
125 static void svc_xprt_free(struct kref
*kref
)
127 struct svc_xprt
*xprt
=
128 container_of(kref
, struct svc_xprt
, xpt_ref
);
129 struct module
*owner
= xprt
->xpt_class
->xcl_owner
;
130 if (test_bit(XPT_CACHE_AUTH
, &xprt
->xpt_flags
) &&
131 xprt
->xpt_auth_cache
!= NULL
)
132 svcauth_unix_info_release(xprt
->xpt_auth_cache
);
133 xprt
->xpt_ops
->xpo_free(xprt
);
137 void svc_xprt_put(struct svc_xprt
*xprt
)
139 kref_put(&xprt
->xpt_ref
, svc_xprt_free
);
141 EXPORT_SYMBOL_GPL(svc_xprt_put
);
144 * Called by transport drivers to initialize the transport independent
145 * portion of the transport instance.
147 void svc_xprt_init(struct svc_xprt_class
*xcl
, struct svc_xprt
*xprt
,
148 struct svc_serv
*serv
)
150 memset(xprt
, 0, sizeof(*xprt
));
151 xprt
->xpt_class
= xcl
;
152 xprt
->xpt_ops
= xcl
->xcl_ops
;
153 kref_init(&xprt
->xpt_ref
);
154 xprt
->xpt_server
= serv
;
155 INIT_LIST_HEAD(&xprt
->xpt_list
);
156 INIT_LIST_HEAD(&xprt
->xpt_ready
);
157 INIT_LIST_HEAD(&xprt
->xpt_deferred
);
158 mutex_init(&xprt
->xpt_mutex
);
159 spin_lock_init(&xprt
->xpt_lock
);
160 set_bit(XPT_BUSY
, &xprt
->xpt_flags
);
161 rpc_init_wait_queue(&xprt
->xpt_bc_pending
, "xpt_bc_pending");
163 EXPORT_SYMBOL_GPL(svc_xprt_init
);
165 static struct svc_xprt
*__svc_xpo_create(struct svc_xprt_class
*xcl
,
166 struct svc_serv
*serv
,
168 const unsigned short port
,
171 struct sockaddr_in sin
= {
172 .sin_family
= AF_INET
,
173 .sin_addr
.s_addr
= htonl(INADDR_ANY
),
174 .sin_port
= htons(port
),
176 struct sockaddr_in6 sin6
= {
177 .sin6_family
= AF_INET6
,
178 .sin6_addr
= IN6ADDR_ANY_INIT
,
179 .sin6_port
= htons(port
),
181 struct sockaddr
*sap
;
186 sap
= (struct sockaddr
*)&sin
;
190 sap
= (struct sockaddr
*)&sin6
;
194 return ERR_PTR(-EAFNOSUPPORT
);
197 return xcl
->xcl_ops
->xpo_create(serv
, sap
, len
, flags
);
200 int svc_create_xprt(struct svc_serv
*serv
, const char *xprt_name
,
201 const int family
, const unsigned short port
,
204 struct svc_xprt_class
*xcl
;
206 dprintk("svc: creating transport %s[%d]\n", xprt_name
, port
);
207 spin_lock(&svc_xprt_class_lock
);
208 list_for_each_entry(xcl
, &svc_xprt_class_list
, xcl_list
) {
209 struct svc_xprt
*newxprt
;
211 if (strcmp(xprt_name
, xcl
->xcl_name
))
214 if (!try_module_get(xcl
->xcl_owner
))
217 spin_unlock(&svc_xprt_class_lock
);
218 newxprt
= __svc_xpo_create(xcl
, serv
, family
, port
, flags
);
219 if (IS_ERR(newxprt
)) {
220 module_put(xcl
->xcl_owner
);
221 return PTR_ERR(newxprt
);
224 clear_bit(XPT_TEMP
, &newxprt
->xpt_flags
);
225 spin_lock_bh(&serv
->sv_lock
);
226 list_add(&newxprt
->xpt_list
, &serv
->sv_permsocks
);
227 spin_unlock_bh(&serv
->sv_lock
);
228 clear_bit(XPT_BUSY
, &newxprt
->xpt_flags
);
229 return svc_xprt_local_port(newxprt
);
232 spin_unlock(&svc_xprt_class_lock
);
233 dprintk("svc: transport %s not found\n", xprt_name
);
236 EXPORT_SYMBOL_GPL(svc_create_xprt
);
239 * Copy the local and remote xprt addresses to the rqstp structure
241 void svc_xprt_copy_addrs(struct svc_rqst
*rqstp
, struct svc_xprt
*xprt
)
243 struct sockaddr
*sin
;
245 memcpy(&rqstp
->rq_addr
, &xprt
->xpt_remote
, xprt
->xpt_remotelen
);
246 rqstp
->rq_addrlen
= xprt
->xpt_remotelen
;
249 * Destination address in request is needed for binding the
250 * source address in RPC replies/callbacks later.
252 sin
= (struct sockaddr
*)&xprt
->xpt_local
;
253 switch (sin
->sa_family
) {
255 rqstp
->rq_daddr
.addr
= ((struct sockaddr_in
*)sin
)->sin_addr
;
258 rqstp
->rq_daddr
.addr6
= ((struct sockaddr_in6
*)sin
)->sin6_addr
;
262 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs
);
265 * svc_print_addr - Format rq_addr field for printing
266 * @rqstp: svc_rqst struct containing address to print
267 * @buf: target buffer for formatted address
268 * @len: length of target buffer
271 char *svc_print_addr(struct svc_rqst
*rqstp
, char *buf
, size_t len
)
273 return __svc_print_addr(svc_addr(rqstp
), buf
, len
);
275 EXPORT_SYMBOL_GPL(svc_print_addr
);
278 * Queue up an idle server thread. Must have pool->sp_lock held.
279 * Note: this is really a stack rather than a queue, so that we only
280 * use as many different threads as we need, and the rest don't pollute
283 static void svc_thread_enqueue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
285 list_add(&rqstp
->rq_list
, &pool
->sp_threads
);
289 * Dequeue an nfsd thread. Must have pool->sp_lock held.
291 static void svc_thread_dequeue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
293 list_del(&rqstp
->rq_list
);
297 * Queue up a transport with data pending. If there are idle nfsd
298 * processes, wake 'em up.
301 void svc_xprt_enqueue(struct svc_xprt
*xprt
)
303 struct svc_serv
*serv
= xprt
->xpt_server
;
304 struct svc_pool
*pool
;
305 struct svc_rqst
*rqstp
;
308 if (!(xprt
->xpt_flags
&
309 ((1<<XPT_CONN
)|(1<<XPT_DATA
)|(1<<XPT_CLOSE
)|(1<<XPT_DEFERRED
))))
313 pool
= svc_pool_for_cpu(xprt
->xpt_server
, cpu
);
316 spin_lock_bh(&pool
->sp_lock
);
318 if (!list_empty(&pool
->sp_threads
) &&
319 !list_empty(&pool
->sp_sockets
))
322 "threads and transports both waiting??\n");
324 if (test_bit(XPT_DEAD
, &xprt
->xpt_flags
)) {
325 /* Don't enqueue dead transports */
326 dprintk("svc: transport %p is dead, not enqueued\n", xprt
);
330 pool
->sp_stats
.packets
++;
332 /* Mark transport as busy. It will remain in this state until
333 * the provider calls svc_xprt_received. We update XPT_BUSY
334 * atomically because it also guards against trying to enqueue
335 * the transport twice.
337 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
338 /* Don't enqueue transport while already enqueued */
339 dprintk("svc: transport %p busy, not enqueued\n", xprt
);
342 BUG_ON(xprt
->xpt_pool
!= NULL
);
343 xprt
->xpt_pool
= pool
;
345 /* Handle pending connection */
346 if (test_bit(XPT_CONN
, &xprt
->xpt_flags
))
349 /* Handle close in-progress */
350 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
))
353 /* Check if we have space to reply to a request */
354 if (!xprt
->xpt_ops
->xpo_has_wspace(xprt
)) {
355 /* Don't enqueue while not enough space for reply */
356 dprintk("svc: no write space, transport %p not enqueued\n",
358 xprt
->xpt_pool
= NULL
;
359 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
364 if (!list_empty(&pool
->sp_threads
)) {
365 rqstp
= list_entry(pool
->sp_threads
.next
,
368 dprintk("svc: transport %p served by daemon %p\n",
370 svc_thread_dequeue(pool
, rqstp
);
373 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
374 rqstp
, rqstp
->rq_xprt
);
375 rqstp
->rq_xprt
= xprt
;
377 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
378 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
379 pool
->sp_stats
.threads_woken
++;
380 BUG_ON(xprt
->xpt_pool
!= pool
);
381 wake_up(&rqstp
->rq_wait
);
383 dprintk("svc: transport %p put into queue\n", xprt
);
384 list_add_tail(&xprt
->xpt_ready
, &pool
->sp_sockets
);
385 pool
->sp_stats
.sockets_queued
++;
386 BUG_ON(xprt
->xpt_pool
!= pool
);
390 spin_unlock_bh(&pool
->sp_lock
);
392 EXPORT_SYMBOL_GPL(svc_xprt_enqueue
);
395 * Dequeue the first transport. Must be called with the pool->sp_lock held.
397 static struct svc_xprt
*svc_xprt_dequeue(struct svc_pool
*pool
)
399 struct svc_xprt
*xprt
;
401 if (list_empty(&pool
->sp_sockets
))
404 xprt
= list_entry(pool
->sp_sockets
.next
,
405 struct svc_xprt
, xpt_ready
);
406 list_del_init(&xprt
->xpt_ready
);
408 dprintk("svc: transport %p dequeued, inuse=%d\n",
409 xprt
, atomic_read(&xprt
->xpt_ref
.refcount
));
415 * svc_xprt_received conditionally queues the transport for processing
416 * by another thread. The caller must hold the XPT_BUSY bit and must
417 * not thereafter touch transport data.
419 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
420 * insufficient) data.
422 void svc_xprt_received(struct svc_xprt
*xprt
)
424 BUG_ON(!test_bit(XPT_BUSY
, &xprt
->xpt_flags
));
425 xprt
->xpt_pool
= NULL
;
426 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
427 svc_xprt_enqueue(xprt
);
429 EXPORT_SYMBOL_GPL(svc_xprt_received
);
432 * svc_reserve - change the space reserved for the reply to a request.
433 * @rqstp: The request in question
434 * @space: new max space to reserve
436 * Each request reserves some space on the output queue of the transport
437 * to make sure the reply fits. This function reduces that reserved
438 * space to be the amount of space used already, plus @space.
441 void svc_reserve(struct svc_rqst
*rqstp
, int space
)
443 space
+= rqstp
->rq_res
.head
[0].iov_len
;
445 if (space
< rqstp
->rq_reserved
) {
446 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
447 atomic_sub((rqstp
->rq_reserved
- space
), &xprt
->xpt_reserved
);
448 rqstp
->rq_reserved
= space
;
450 svc_xprt_enqueue(xprt
);
453 EXPORT_SYMBOL_GPL(svc_reserve
);
455 static void svc_xprt_release(struct svc_rqst
*rqstp
)
457 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
459 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
461 kfree(rqstp
->rq_deferred
);
462 rqstp
->rq_deferred
= NULL
;
464 svc_free_res_pages(rqstp
);
465 rqstp
->rq_res
.page_len
= 0;
466 rqstp
->rq_res
.page_base
= 0;
468 /* Reset response buffer and release
470 * But first, check that enough space was reserved
471 * for the reply, otherwise we have a bug!
473 if ((rqstp
->rq_res
.len
) > rqstp
->rq_reserved
)
474 printk(KERN_ERR
"RPC request reserved %d but used %d\n",
478 rqstp
->rq_res
.head
[0].iov_len
= 0;
479 svc_reserve(rqstp
, 0);
480 rqstp
->rq_xprt
= NULL
;
486 * External function to wake up a server waiting for data
487 * This really only makes sense for services like lockd
488 * which have exactly one thread anyway.
490 void svc_wake_up(struct svc_serv
*serv
)
492 struct svc_rqst
*rqstp
;
494 struct svc_pool
*pool
;
496 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
497 pool
= &serv
->sv_pools
[i
];
499 spin_lock_bh(&pool
->sp_lock
);
500 if (!list_empty(&pool
->sp_threads
)) {
501 rqstp
= list_entry(pool
->sp_threads
.next
,
504 dprintk("svc: daemon %p woken up.\n", rqstp
);
506 svc_thread_dequeue(pool, rqstp);
507 rqstp->rq_xprt = NULL;
509 wake_up(&rqstp
->rq_wait
);
511 spin_unlock_bh(&pool
->sp_lock
);
514 EXPORT_SYMBOL_GPL(svc_wake_up
);
516 int svc_port_is_privileged(struct sockaddr
*sin
)
518 switch (sin
->sa_family
) {
520 return ntohs(((struct sockaddr_in
*)sin
)->sin_port
)
523 return ntohs(((struct sockaddr_in6
*)sin
)->sin6_port
)
531 * Make sure that we don't have too many active connections. If we have,
532 * something must be dropped. It's not clear what will happen if we allow
533 * "too many" connections, but when dealing with network-facing software,
534 * we have to code defensively. Here we do that by imposing hard limits.
536 * There's no point in trying to do random drop here for DoS
537 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
538 * attacker can easily beat that.
540 * The only somewhat efficient mechanism would be if drop old
541 * connections from the same IP first. But right now we don't even
542 * record the client IP in svc_sock.
544 * single-threaded services that expect a lot of clients will probably
545 * need to set sv_maxconn to override the default value which is based
546 * on the number of threads
548 static void svc_check_conn_limits(struct svc_serv
*serv
)
550 unsigned int limit
= serv
->sv_maxconn
? serv
->sv_maxconn
:
551 (serv
->sv_nrthreads
+3) * 20;
553 if (serv
->sv_tmpcnt
> limit
) {
554 struct svc_xprt
*xprt
= NULL
;
555 spin_lock_bh(&serv
->sv_lock
);
556 if (!list_empty(&serv
->sv_tempsocks
)) {
557 if (net_ratelimit()) {
558 /* Try to help the admin */
559 printk(KERN_NOTICE
"%s: too many open "
560 "connections, consider increasing %s\n",
561 serv
->sv_name
, serv
->sv_maxconn
?
562 "the max number of connections." :
563 "the number of threads.");
566 * Always select the oldest connection. It's not fair,
569 xprt
= list_entry(serv
->sv_tempsocks
.prev
,
572 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
575 spin_unlock_bh(&serv
->sv_lock
);
578 svc_xprt_enqueue(xprt
);
585 * Receive the next request on any transport. This code is carefully
586 * organised not to touch any cachelines in the shared svc_serv
587 * structure, only cachelines in the local svc_pool.
589 int svc_recv(struct svc_rqst
*rqstp
, long timeout
)
591 struct svc_xprt
*xprt
= NULL
;
592 struct svc_serv
*serv
= rqstp
->rq_server
;
593 struct svc_pool
*pool
= rqstp
->rq_pool
;
597 DECLARE_WAITQUEUE(wait
, current
);
600 dprintk("svc: server %p waiting for data (to = %ld)\n",
605 "svc_recv: service %p, transport not NULL!\n",
607 if (waitqueue_active(&rqstp
->rq_wait
))
609 "svc_recv: service %p, wait queue active!\n",
612 /* now allocate needed pages. If we get a failure, sleep briefly */
613 pages
= (serv
->sv_max_mesg
+ PAGE_SIZE
) / PAGE_SIZE
;
614 for (i
= 0; i
< pages
; i
++)
615 while (rqstp
->rq_pages
[i
] == NULL
) {
616 struct page
*p
= alloc_page(GFP_KERNEL
);
618 set_current_state(TASK_INTERRUPTIBLE
);
619 if (signalled() || kthread_should_stop()) {
620 set_current_state(TASK_RUNNING
);
623 schedule_timeout(msecs_to_jiffies(500));
625 rqstp
->rq_pages
[i
] = p
;
627 rqstp
->rq_pages
[i
++] = NULL
; /* this might be seen in nfs_read_actor */
628 BUG_ON(pages
>= RPCSVC_MAXPAGES
);
630 /* Make arg->head point to first page and arg->pages point to rest */
631 arg
= &rqstp
->rq_arg
;
632 arg
->head
[0].iov_base
= page_address(rqstp
->rq_pages
[0]);
633 arg
->head
[0].iov_len
= PAGE_SIZE
;
634 arg
->pages
= rqstp
->rq_pages
+ 1;
636 /* save at least one page for response */
637 arg
->page_len
= (pages
-2)*PAGE_SIZE
;
638 arg
->len
= (pages
-1)*PAGE_SIZE
;
639 arg
->tail
[0].iov_len
= 0;
643 if (signalled() || kthread_should_stop())
646 spin_lock_bh(&pool
->sp_lock
);
647 xprt
= svc_xprt_dequeue(pool
);
649 rqstp
->rq_xprt
= xprt
;
651 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
652 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
654 /* No data pending. Go to sleep */
655 svc_thread_enqueue(pool
, rqstp
);
658 * We have to be able to interrupt this wait
659 * to bring down the daemons ...
661 set_current_state(TASK_INTERRUPTIBLE
);
664 * checking kthread_should_stop() here allows us to avoid
665 * locking and signalling when stopping kthreads that call
666 * svc_recv. If the thread has already been woken up, then
667 * we can exit here without sleeping. If not, then it
668 * it'll be woken up quickly during the schedule_timeout
670 if (kthread_should_stop()) {
671 set_current_state(TASK_RUNNING
);
672 spin_unlock_bh(&pool
->sp_lock
);
676 add_wait_queue(&rqstp
->rq_wait
, &wait
);
677 spin_unlock_bh(&pool
->sp_lock
);
679 time_left
= schedule_timeout(timeout
);
683 spin_lock_bh(&pool
->sp_lock
);
684 remove_wait_queue(&rqstp
->rq_wait
, &wait
);
686 pool
->sp_stats
.threads_timedout
++;
688 xprt
= rqstp
->rq_xprt
;
690 svc_thread_dequeue(pool
, rqstp
);
691 spin_unlock_bh(&pool
->sp_lock
);
692 dprintk("svc: server %p, no data yet\n", rqstp
);
693 if (signalled() || kthread_should_stop())
699 spin_unlock_bh(&pool
->sp_lock
);
702 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
)) {
703 dprintk("svc_recv: found XPT_CLOSE\n");
704 svc_delete_xprt(xprt
);
705 } else if (test_bit(XPT_LISTENER
, &xprt
->xpt_flags
)) {
706 struct svc_xprt
*newxpt
;
707 newxpt
= xprt
->xpt_ops
->xpo_accept(xprt
);
710 * We know this module_get will succeed because the
711 * listener holds a reference too
713 __module_get(newxpt
->xpt_class
->xcl_owner
);
714 svc_check_conn_limits(xprt
->xpt_server
);
715 spin_lock_bh(&serv
->sv_lock
);
716 set_bit(XPT_TEMP
, &newxpt
->xpt_flags
);
717 list_add(&newxpt
->xpt_list
, &serv
->sv_tempsocks
);
719 if (serv
->sv_temptimer
.function
== NULL
) {
720 /* setup timer to age temp transports */
721 setup_timer(&serv
->sv_temptimer
,
723 (unsigned long)serv
);
724 mod_timer(&serv
->sv_temptimer
,
725 jiffies
+ svc_conn_age_period
* HZ
);
727 spin_unlock_bh(&serv
->sv_lock
);
728 svc_xprt_received(newxpt
);
730 svc_xprt_received(xprt
);
732 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
733 rqstp
, pool
->sp_id
, xprt
,
734 atomic_read(&xprt
->xpt_ref
.refcount
));
735 rqstp
->rq_deferred
= svc_deferred_dequeue(xprt
);
736 if (rqstp
->rq_deferred
) {
737 svc_xprt_received(xprt
);
738 len
= svc_deferred_recv(rqstp
);
740 len
= xprt
->xpt_ops
->xpo_recvfrom(rqstp
);
741 dprintk("svc: got len=%d\n", len
);
744 /* No data, incomplete (TCP) read, or accept() */
745 if (len
== 0 || len
== -EAGAIN
) {
746 rqstp
->rq_res
.len
= 0;
747 svc_xprt_release(rqstp
);
750 clear_bit(XPT_OLD
, &xprt
->xpt_flags
);
752 rqstp
->rq_secure
= svc_port_is_privileged(svc_addr(rqstp
));
753 rqstp
->rq_chandle
.defer
= svc_defer
;
756 serv
->sv_stats
->netcnt
++;
759 EXPORT_SYMBOL_GPL(svc_recv
);
764 void svc_drop(struct svc_rqst
*rqstp
)
766 dprintk("svc: xprt %p dropped request\n", rqstp
->rq_xprt
);
767 svc_xprt_release(rqstp
);
769 EXPORT_SYMBOL_GPL(svc_drop
);
772 * Return reply to client.
774 int svc_send(struct svc_rqst
*rqstp
)
776 struct svc_xprt
*xprt
;
780 xprt
= rqstp
->rq_xprt
;
784 /* release the receive skb before sending the reply */
785 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
787 /* calculate over-all length */
789 xb
->len
= xb
->head
[0].iov_len
+
793 /* Grab mutex to serialize outgoing data. */
794 mutex_lock(&xprt
->xpt_mutex
);
795 if (test_bit(XPT_DEAD
, &xprt
->xpt_flags
))
798 len
= xprt
->xpt_ops
->xpo_sendto(rqstp
);
799 mutex_unlock(&xprt
->xpt_mutex
);
800 rpc_wake_up(&xprt
->xpt_bc_pending
);
801 svc_xprt_release(rqstp
);
803 if (len
== -ECONNREFUSED
|| len
== -ENOTCONN
|| len
== -EAGAIN
)
809 * Timer function to close old temporary transports, using
810 * a mark-and-sweep algorithm.
812 static void svc_age_temp_xprts(unsigned long closure
)
814 struct svc_serv
*serv
= (struct svc_serv
*)closure
;
815 struct svc_xprt
*xprt
;
816 struct list_head
*le
, *next
;
817 LIST_HEAD(to_be_aged
);
819 dprintk("svc_age_temp_xprts\n");
821 if (!spin_trylock_bh(&serv
->sv_lock
)) {
822 /* busy, try again 1 sec later */
823 dprintk("svc_age_temp_xprts: busy\n");
824 mod_timer(&serv
->sv_temptimer
, jiffies
+ HZ
);
828 list_for_each_safe(le
, next
, &serv
->sv_tempsocks
) {
829 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
831 /* First time through, just mark it OLD. Second time
832 * through, close it. */
833 if (!test_and_set_bit(XPT_OLD
, &xprt
->xpt_flags
))
835 if (atomic_read(&xprt
->xpt_ref
.refcount
) > 1 ||
836 test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
839 list_move(le
, &to_be_aged
);
840 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
841 set_bit(XPT_DETACHED
, &xprt
->xpt_flags
);
843 spin_unlock_bh(&serv
->sv_lock
);
845 while (!list_empty(&to_be_aged
)) {
846 le
= to_be_aged
.next
;
847 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
849 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
851 dprintk("queuing xprt %p for closing\n", xprt
);
853 /* a thread will dequeue and close it soon */
854 svc_xprt_enqueue(xprt
);
858 mod_timer(&serv
->sv_temptimer
, jiffies
+ svc_conn_age_period
* HZ
);
862 * Remove a dead transport
864 void svc_delete_xprt(struct svc_xprt
*xprt
)
866 struct svc_serv
*serv
= xprt
->xpt_server
;
867 struct svc_deferred_req
*dr
;
869 /* Only do this once */
870 if (test_and_set_bit(XPT_DEAD
, &xprt
->xpt_flags
))
873 dprintk("svc: svc_delete_xprt(%p)\n", xprt
);
874 xprt
->xpt_ops
->xpo_detach(xprt
);
876 spin_lock_bh(&serv
->sv_lock
);
877 if (!test_and_set_bit(XPT_DETACHED
, &xprt
->xpt_flags
))
878 list_del_init(&xprt
->xpt_list
);
880 * We used to delete the transport from whichever list
881 * it's sk_xprt.xpt_ready node was on, but we don't actually
882 * need to. This is because the only time we're called
883 * while still attached to a queue, the queue itself
884 * is about to be destroyed (in svc_destroy).
886 if (test_bit(XPT_TEMP
, &xprt
->xpt_flags
))
889 while ((dr
= svc_deferred_dequeue(xprt
)) != NULL
)
893 spin_unlock_bh(&serv
->sv_lock
);
896 void svc_close_xprt(struct svc_xprt
*xprt
)
898 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
899 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
))
900 /* someone else will have to effect the close */
904 svc_delete_xprt(xprt
);
905 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
908 EXPORT_SYMBOL_GPL(svc_close_xprt
);
910 void svc_close_all(struct list_head
*xprt_list
)
912 struct svc_xprt
*xprt
;
913 struct svc_xprt
*tmp
;
915 list_for_each_entry_safe(xprt
, tmp
, xprt_list
, xpt_list
) {
916 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
917 if (test_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
918 /* Waiting to be processed, but no threads left,
919 * So just remove it from the waiting list
921 list_del_init(&xprt
->xpt_ready
);
922 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
924 svc_close_xprt(xprt
);
929 * Handle defer and revisit of requests
932 static void svc_revisit(struct cache_deferred_req
*dreq
, int too_many
)
934 struct svc_deferred_req
*dr
=
935 container_of(dreq
, struct svc_deferred_req
, handle
);
936 struct svc_xprt
*xprt
= dr
->xprt
;
938 spin_lock(&xprt
->xpt_lock
);
939 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
940 if (too_many
|| test_bit(XPT_DEAD
, &xprt
->xpt_flags
)) {
941 spin_unlock(&xprt
->xpt_lock
);
942 dprintk("revisit canceled\n");
947 dprintk("revisit queued\n");
949 list_add(&dr
->handle
.recent
, &xprt
->xpt_deferred
);
950 spin_unlock(&xprt
->xpt_lock
);
951 svc_xprt_enqueue(xprt
);
956 * Save the request off for later processing. The request buffer looks
959 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
961 * This code can only handle requests that consist of an xprt-header
964 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
)
966 struct svc_rqst
*rqstp
= container_of(req
, struct svc_rqst
, rq_chandle
);
967 struct svc_deferred_req
*dr
;
969 if (rqstp
->rq_arg
.page_len
|| !rqstp
->rq_usedeferral
)
970 return NULL
; /* if more than a page, give up FIXME */
971 if (rqstp
->rq_deferred
) {
972 dr
= rqstp
->rq_deferred
;
973 rqstp
->rq_deferred
= NULL
;
977 /* FIXME maybe discard if size too large */
978 size
= sizeof(struct svc_deferred_req
) + rqstp
->rq_arg
.len
;
979 dr
= kmalloc(size
, GFP_KERNEL
);
983 dr
->handle
.owner
= rqstp
->rq_server
;
984 dr
->prot
= rqstp
->rq_prot
;
985 memcpy(&dr
->addr
, &rqstp
->rq_addr
, rqstp
->rq_addrlen
);
986 dr
->addrlen
= rqstp
->rq_addrlen
;
987 dr
->daddr
= rqstp
->rq_daddr
;
988 dr
->argslen
= rqstp
->rq_arg
.len
>> 2;
989 dr
->xprt_hlen
= rqstp
->rq_xprt_hlen
;
991 /* back up head to the start of the buffer and copy */
992 skip
= rqstp
->rq_arg
.len
- rqstp
->rq_arg
.head
[0].iov_len
;
993 memcpy(dr
->args
, rqstp
->rq_arg
.head
[0].iov_base
- skip
,
996 svc_xprt_get(rqstp
->rq_xprt
);
997 dr
->xprt
= rqstp
->rq_xprt
;
999 dr
->handle
.revisit
= svc_revisit
;
1004 * recv data from a deferred request into an active one
1006 static int svc_deferred_recv(struct svc_rqst
*rqstp
)
1008 struct svc_deferred_req
*dr
= rqstp
->rq_deferred
;
1010 /* setup iov_base past transport header */
1011 rqstp
->rq_arg
.head
[0].iov_base
= dr
->args
+ (dr
->xprt_hlen
>>2);
1012 /* The iov_len does not include the transport header bytes */
1013 rqstp
->rq_arg
.head
[0].iov_len
= (dr
->argslen
<<2) - dr
->xprt_hlen
;
1014 rqstp
->rq_arg
.page_len
= 0;
1015 /* The rq_arg.len includes the transport header bytes */
1016 rqstp
->rq_arg
.len
= dr
->argslen
<<2;
1017 rqstp
->rq_prot
= dr
->prot
;
1018 memcpy(&rqstp
->rq_addr
, &dr
->addr
, dr
->addrlen
);
1019 rqstp
->rq_addrlen
= dr
->addrlen
;
1020 /* Save off transport header len in case we get deferred again */
1021 rqstp
->rq_xprt_hlen
= dr
->xprt_hlen
;
1022 rqstp
->rq_daddr
= dr
->daddr
;
1023 rqstp
->rq_respages
= rqstp
->rq_pages
;
1024 return (dr
->argslen
<<2) - dr
->xprt_hlen
;
1028 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
)
1030 struct svc_deferred_req
*dr
= NULL
;
1032 if (!test_bit(XPT_DEFERRED
, &xprt
->xpt_flags
))
1034 spin_lock(&xprt
->xpt_lock
);
1035 clear_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1036 if (!list_empty(&xprt
->xpt_deferred
)) {
1037 dr
= list_entry(xprt
->xpt_deferred
.next
,
1038 struct svc_deferred_req
,
1040 list_del_init(&dr
->handle
.recent
);
1041 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1043 spin_unlock(&xprt
->xpt_lock
);
1048 * svc_find_xprt - find an RPC transport instance
1049 * @serv: pointer to svc_serv to search
1050 * @xcl_name: C string containing transport's class name
1051 * @af: Address family of transport's local address
1052 * @port: transport's IP port number
1054 * Return the transport instance pointer for the endpoint accepting
1055 * connections/peer traffic from the specified transport class,
1056 * address family and port.
1058 * Specifying 0 for the address family or port is effectively a
1059 * wild-card, and will result in matching the first transport in the
1060 * service's list that has a matching class name.
1062 struct svc_xprt
*svc_find_xprt(struct svc_serv
*serv
, const char *xcl_name
,
1063 const sa_family_t af
, const unsigned short port
)
1065 struct svc_xprt
*xprt
;
1066 struct svc_xprt
*found
= NULL
;
1068 /* Sanity check the args */
1069 if (serv
== NULL
|| xcl_name
== NULL
)
1072 spin_lock_bh(&serv
->sv_lock
);
1073 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1074 if (strcmp(xprt
->xpt_class
->xcl_name
, xcl_name
))
1076 if (af
!= AF_UNSPEC
&& af
!= xprt
->xpt_local
.ss_family
)
1078 if (port
!= 0 && port
!= svc_xprt_local_port(xprt
))
1084 spin_unlock_bh(&serv
->sv_lock
);
1087 EXPORT_SYMBOL_GPL(svc_find_xprt
);
1089 static int svc_one_xprt_name(const struct svc_xprt
*xprt
,
1090 char *pos
, int remaining
)
1094 len
= snprintf(pos
, remaining
, "%s %u\n",
1095 xprt
->xpt_class
->xcl_name
,
1096 svc_xprt_local_port(xprt
));
1097 if (len
>= remaining
)
1098 return -ENAMETOOLONG
;
1103 * svc_xprt_names - format a buffer with a list of transport names
1104 * @serv: pointer to an RPC service
1105 * @buf: pointer to a buffer to be filled in
1106 * @buflen: length of buffer to be filled in
1108 * Fills in @buf with a string containing a list of transport names,
1109 * each name terminated with '\n'.
1111 * Returns positive length of the filled-in string on success; otherwise
1112 * a negative errno value is returned if an error occurs.
1114 int svc_xprt_names(struct svc_serv
*serv
, char *buf
, const int buflen
)
1116 struct svc_xprt
*xprt
;
1120 /* Sanity check args */
1124 spin_lock_bh(&serv
->sv_lock
);
1128 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1129 len
= svc_one_xprt_name(xprt
, pos
, buflen
- totlen
);
1141 spin_unlock_bh(&serv
->sv_lock
);
1144 EXPORT_SYMBOL_GPL(svc_xprt_names
);
1147 /*----------------------------------------------------------------------------*/
1149 static void *svc_pool_stats_start(struct seq_file
*m
, loff_t
*pos
)
1151 unsigned int pidx
= (unsigned int)*pos
;
1152 struct svc_serv
*serv
= m
->private;
1154 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx
);
1157 return SEQ_START_TOKEN
;
1158 return (pidx
> serv
->sv_nrpools
? NULL
: &serv
->sv_pools
[pidx
-1]);
1161 static void *svc_pool_stats_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
1163 struct svc_pool
*pool
= p
;
1164 struct svc_serv
*serv
= m
->private;
1166 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos
);
1168 if (p
== SEQ_START_TOKEN
) {
1169 pool
= &serv
->sv_pools
[0];
1171 unsigned int pidx
= (pool
- &serv
->sv_pools
[0]);
1172 if (pidx
< serv
->sv_nrpools
-1)
1173 pool
= &serv
->sv_pools
[pidx
+1];
1181 static void svc_pool_stats_stop(struct seq_file
*m
, void *p
)
1185 static int svc_pool_stats_show(struct seq_file
*m
, void *p
)
1187 struct svc_pool
*pool
= p
;
1189 if (p
== SEQ_START_TOKEN
) {
1190 seq_puts(m
, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1194 seq_printf(m
, "%u %lu %lu %lu %lu\n",
1196 pool
->sp_stats
.packets
,
1197 pool
->sp_stats
.sockets_queued
,
1198 pool
->sp_stats
.threads_woken
,
1199 pool
->sp_stats
.threads_timedout
);
1204 static const struct seq_operations svc_pool_stats_seq_ops
= {
1205 .start
= svc_pool_stats_start
,
1206 .next
= svc_pool_stats_next
,
1207 .stop
= svc_pool_stats_stop
,
1208 .show
= svc_pool_stats_show
,
1211 int svc_pool_stats_open(struct svc_serv
*serv
, struct file
*file
)
1215 err
= seq_open(file
, &svc_pool_stats_seq_ops
);
1217 ((struct seq_file
*) file
->private_data
)->private = serv
;
1220 EXPORT_SYMBOL(svc_pool_stats_open
);
1222 /*----------------------------------------------------------------------------*/