2 * linux/net/sunrpc/svc_xprt.c
4 * Author: Tom Tucker <tom@opengridcomputing.com>
7 #include <linux/sched.h>
8 #include <linux/errno.h>
9 #include <linux/freezer.h>
10 #include <linux/kthread.h>
12 #include <linux/sunrpc/stats.h>
13 #include <linux/sunrpc/svc_xprt.h>
15 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
17 #define SVC_MAX_WAKING 5
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
);
162 EXPORT_SYMBOL_GPL(svc_xprt_init
);
164 static struct svc_xprt
*__svc_xpo_create(struct svc_xprt_class
*xcl
,
165 struct svc_serv
*serv
,
166 unsigned short port
, int flags
)
168 struct sockaddr_in sin
= {
169 .sin_family
= AF_INET
,
170 .sin_addr
.s_addr
= htonl(INADDR_ANY
),
171 .sin_port
= htons(port
),
173 struct sockaddr_in6 sin6
= {
174 .sin6_family
= AF_INET6
,
175 .sin6_addr
= IN6ADDR_ANY_INIT
,
176 .sin6_port
= htons(port
),
178 struct sockaddr
*sap
;
181 switch (serv
->sv_family
) {
183 sap
= (struct sockaddr
*)&sin
;
187 sap
= (struct sockaddr
*)&sin6
;
191 return ERR_PTR(-EAFNOSUPPORT
);
194 return xcl
->xcl_ops
->xpo_create(serv
, sap
, len
, flags
);
197 int svc_create_xprt(struct svc_serv
*serv
, char *xprt_name
, unsigned short port
,
200 struct svc_xprt_class
*xcl
;
202 dprintk("svc: creating transport %s[%d]\n", xprt_name
, port
);
203 spin_lock(&svc_xprt_class_lock
);
204 list_for_each_entry(xcl
, &svc_xprt_class_list
, xcl_list
) {
205 struct svc_xprt
*newxprt
;
207 if (strcmp(xprt_name
, xcl
->xcl_name
))
210 if (!try_module_get(xcl
->xcl_owner
))
213 spin_unlock(&svc_xprt_class_lock
);
214 newxprt
= __svc_xpo_create(xcl
, serv
, port
, flags
);
215 if (IS_ERR(newxprt
)) {
216 module_put(xcl
->xcl_owner
);
217 return PTR_ERR(newxprt
);
220 clear_bit(XPT_TEMP
, &newxprt
->xpt_flags
);
221 spin_lock_bh(&serv
->sv_lock
);
222 list_add(&newxprt
->xpt_list
, &serv
->sv_permsocks
);
223 spin_unlock_bh(&serv
->sv_lock
);
224 clear_bit(XPT_BUSY
, &newxprt
->xpt_flags
);
225 return svc_xprt_local_port(newxprt
);
228 spin_unlock(&svc_xprt_class_lock
);
229 dprintk("svc: transport %s not found\n", xprt_name
);
232 EXPORT_SYMBOL_GPL(svc_create_xprt
);
235 * Copy the local and remote xprt addresses to the rqstp structure
237 void svc_xprt_copy_addrs(struct svc_rqst
*rqstp
, struct svc_xprt
*xprt
)
239 struct sockaddr
*sin
;
241 memcpy(&rqstp
->rq_addr
, &xprt
->xpt_remote
, xprt
->xpt_remotelen
);
242 rqstp
->rq_addrlen
= xprt
->xpt_remotelen
;
245 * Destination address in request is needed for binding the
246 * source address in RPC replies/callbacks later.
248 sin
= (struct sockaddr
*)&xprt
->xpt_local
;
249 switch (sin
->sa_family
) {
251 rqstp
->rq_daddr
.addr
= ((struct sockaddr_in
*)sin
)->sin_addr
;
254 rqstp
->rq_daddr
.addr6
= ((struct sockaddr_in6
*)sin
)->sin6_addr
;
258 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs
);
261 * svc_print_addr - Format rq_addr field for printing
262 * @rqstp: svc_rqst struct containing address to print
263 * @buf: target buffer for formatted address
264 * @len: length of target buffer
267 char *svc_print_addr(struct svc_rqst
*rqstp
, char *buf
, size_t len
)
269 return __svc_print_addr(svc_addr(rqstp
), buf
, len
);
271 EXPORT_SYMBOL_GPL(svc_print_addr
);
274 * Queue up an idle server thread. Must have pool->sp_lock held.
275 * Note: this is really a stack rather than a queue, so that we only
276 * use as many different threads as we need, and the rest don't pollute
279 static void svc_thread_enqueue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
281 list_add(&rqstp
->rq_list
, &pool
->sp_threads
);
285 * Dequeue an nfsd thread. Must have pool->sp_lock held.
287 static void svc_thread_dequeue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
289 list_del(&rqstp
->rq_list
);
293 * Queue up a transport with data pending. If there are idle nfsd
294 * processes, wake 'em up.
297 void svc_xprt_enqueue(struct svc_xprt
*xprt
)
299 struct svc_serv
*serv
= xprt
->xpt_server
;
300 struct svc_pool
*pool
;
301 struct svc_rqst
*rqstp
;
305 if (!(xprt
->xpt_flags
&
306 ((1<<XPT_CONN
)|(1<<XPT_DATA
)|(1<<XPT_CLOSE
)|(1<<XPT_DEFERRED
))))
310 pool
= svc_pool_for_cpu(xprt
->xpt_server
, cpu
);
313 spin_lock_bh(&pool
->sp_lock
);
315 if (test_bit(XPT_DEAD
, &xprt
->xpt_flags
)) {
316 /* Don't enqueue dead transports */
317 dprintk("svc: transport %p is dead, not enqueued\n", xprt
);
321 pool
->sp_stats
.packets
++;
323 /* Mark transport as busy. It will remain in this state until
324 * the provider calls svc_xprt_received. We update XPT_BUSY
325 * atomically because it also guards against trying to enqueue
326 * the transport twice.
328 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
329 /* Don't enqueue transport while already enqueued */
330 dprintk("svc: transport %p busy, not enqueued\n", xprt
);
333 BUG_ON(xprt
->xpt_pool
!= NULL
);
334 xprt
->xpt_pool
= pool
;
336 /* Handle pending connection */
337 if (test_bit(XPT_CONN
, &xprt
->xpt_flags
))
340 /* Handle close in-progress */
341 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
))
344 /* Check if we have space to reply to a request */
345 if (!xprt
->xpt_ops
->xpo_has_wspace(xprt
)) {
346 /* Don't enqueue while not enough space for reply */
347 dprintk("svc: no write space, transport %p not enqueued\n",
349 xprt
->xpt_pool
= NULL
;
350 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
355 /* Work out whether threads are available */
356 thread_avail
= !list_empty(&pool
->sp_threads
); /* threads are asleep */
357 if (pool
->sp_nwaking
>= SVC_MAX_WAKING
) {
358 /* too many threads are runnable and trying to wake up */
360 pool
->sp_stats
.overloads_avoided
++;
364 rqstp
= list_entry(pool
->sp_threads
.next
,
367 dprintk("svc: transport %p served by daemon %p\n",
369 svc_thread_dequeue(pool
, rqstp
);
372 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
373 rqstp
, rqstp
->rq_xprt
);
374 rqstp
->rq_xprt
= xprt
;
376 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
377 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
378 rqstp
->rq_waking
= 1;
380 pool
->sp_stats
.threads_woken
++;
381 BUG_ON(xprt
->xpt_pool
!= pool
);
382 wake_up(&rqstp
->rq_wait
);
384 dprintk("svc: transport %p put into queue\n", xprt
);
385 list_add_tail(&xprt
->xpt_ready
, &pool
->sp_sockets
);
386 pool
->sp_stats
.sockets_queued
++;
387 BUG_ON(xprt
->xpt_pool
!= pool
);
391 spin_unlock_bh(&pool
->sp_lock
);
393 EXPORT_SYMBOL_GPL(svc_xprt_enqueue
);
396 * Dequeue the first transport. Must be called with the pool->sp_lock held.
398 static struct svc_xprt
*svc_xprt_dequeue(struct svc_pool
*pool
)
400 struct svc_xprt
*xprt
;
402 if (list_empty(&pool
->sp_sockets
))
405 xprt
= list_entry(pool
->sp_sockets
.next
,
406 struct svc_xprt
, xpt_ready
);
407 list_del_init(&xprt
->xpt_ready
);
409 dprintk("svc: transport %p dequeued, inuse=%d\n",
410 xprt
, atomic_read(&xprt
->xpt_ref
.refcount
));
416 * svc_xprt_received conditionally queues the transport for processing
417 * by another thread. The caller must hold the XPT_BUSY bit and must
418 * not thereafter touch transport data.
420 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
421 * insufficient) data.
423 void svc_xprt_received(struct svc_xprt
*xprt
)
425 BUG_ON(!test_bit(XPT_BUSY
, &xprt
->xpt_flags
));
426 xprt
->xpt_pool
= NULL
;
427 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
428 svc_xprt_enqueue(xprt
);
430 EXPORT_SYMBOL_GPL(svc_xprt_received
);
433 * svc_reserve - change the space reserved for the reply to a request.
434 * @rqstp: The request in question
435 * @space: new max space to reserve
437 * Each request reserves some space on the output queue of the transport
438 * to make sure the reply fits. This function reduces that reserved
439 * space to be the amount of space used already, plus @space.
442 void svc_reserve(struct svc_rqst
*rqstp
, int space
)
444 space
+= rqstp
->rq_res
.head
[0].iov_len
;
446 if (space
< rqstp
->rq_reserved
) {
447 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
448 atomic_sub((rqstp
->rq_reserved
- space
), &xprt
->xpt_reserved
);
449 rqstp
->rq_reserved
= space
;
451 svc_xprt_enqueue(xprt
);
454 EXPORT_SYMBOL_GPL(svc_reserve
);
456 static void svc_xprt_release(struct svc_rqst
*rqstp
)
458 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
460 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
462 kfree(rqstp
->rq_deferred
);
463 rqstp
->rq_deferred
= NULL
;
465 svc_free_res_pages(rqstp
);
466 rqstp
->rq_res
.page_len
= 0;
467 rqstp
->rq_res
.page_base
= 0;
469 /* Reset response buffer and release
471 * But first, check that enough space was reserved
472 * for the reply, otherwise we have a bug!
474 if ((rqstp
->rq_res
.len
) > rqstp
->rq_reserved
)
475 printk(KERN_ERR
"RPC request reserved %d but used %d\n",
479 rqstp
->rq_res
.head
[0].iov_len
= 0;
480 svc_reserve(rqstp
, 0);
481 rqstp
->rq_xprt
= NULL
;
487 * External function to wake up a server waiting for data
488 * This really only makes sense for services like lockd
489 * which have exactly one thread anyway.
491 void svc_wake_up(struct svc_serv
*serv
)
493 struct svc_rqst
*rqstp
;
495 struct svc_pool
*pool
;
497 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
498 pool
= &serv
->sv_pools
[i
];
500 spin_lock_bh(&pool
->sp_lock
);
501 if (!list_empty(&pool
->sp_threads
)) {
502 rqstp
= list_entry(pool
->sp_threads
.next
,
505 dprintk("svc: daemon %p woken up.\n", rqstp
);
507 svc_thread_dequeue(pool, rqstp);
508 rqstp->rq_xprt = NULL;
510 wake_up(&rqstp
->rq_wait
);
512 spin_unlock_bh(&pool
->sp_lock
);
515 EXPORT_SYMBOL_GPL(svc_wake_up
);
517 int svc_port_is_privileged(struct sockaddr
*sin
)
519 switch (sin
->sa_family
) {
521 return ntohs(((struct sockaddr_in
*)sin
)->sin_port
)
524 return ntohs(((struct sockaddr_in6
*)sin
)->sin6_port
)
532 * Make sure that we don't have too many active connections. If we have,
533 * something must be dropped. It's not clear what will happen if we allow
534 * "too many" connections, but when dealing with network-facing software,
535 * we have to code defensively. Here we do that by imposing hard limits.
537 * There's no point in trying to do random drop here for DoS
538 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
539 * attacker can easily beat that.
541 * The only somewhat efficient mechanism would be if drop old
542 * connections from the same IP first. But right now we don't even
543 * record the client IP in svc_sock.
545 * single-threaded services that expect a lot of clients will probably
546 * need to set sv_maxconn to override the default value which is based
547 * on the number of threads
549 static void svc_check_conn_limits(struct svc_serv
*serv
)
551 unsigned int limit
= serv
->sv_maxconn
? serv
->sv_maxconn
:
552 (serv
->sv_nrthreads
+3) * 20;
554 if (serv
->sv_tmpcnt
> limit
) {
555 struct svc_xprt
*xprt
= NULL
;
556 spin_lock_bh(&serv
->sv_lock
);
557 if (!list_empty(&serv
->sv_tempsocks
)) {
558 if (net_ratelimit()) {
559 /* Try to help the admin */
560 printk(KERN_NOTICE
"%s: too many open "
561 "connections, consider increasing %s\n",
562 serv
->sv_name
, serv
->sv_maxconn
?
563 "the max number of connections." :
564 "the number of threads.");
567 * Always select the oldest connection. It's not fair,
570 xprt
= list_entry(serv
->sv_tempsocks
.prev
,
573 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
576 spin_unlock_bh(&serv
->sv_lock
);
579 svc_xprt_enqueue(xprt
);
586 * Receive the next request on any transport. This code is carefully
587 * organised not to touch any cachelines in the shared svc_serv
588 * structure, only cachelines in the local svc_pool.
590 int svc_recv(struct svc_rqst
*rqstp
, long timeout
)
592 struct svc_xprt
*xprt
= NULL
;
593 struct svc_serv
*serv
= rqstp
->rq_server
;
594 struct svc_pool
*pool
= rqstp
->rq_pool
;
598 DECLARE_WAITQUEUE(wait
, current
);
601 dprintk("svc: server %p waiting for data (to = %ld)\n",
606 "svc_recv: service %p, transport not NULL!\n",
608 if (waitqueue_active(&rqstp
->rq_wait
))
610 "svc_recv: service %p, wait queue active!\n",
613 /* now allocate needed pages. If we get a failure, sleep briefly */
614 pages
= (serv
->sv_max_mesg
+ PAGE_SIZE
) / PAGE_SIZE
;
615 for (i
= 0; i
< pages
; i
++)
616 while (rqstp
->rq_pages
[i
] == NULL
) {
617 struct page
*p
= alloc_page(GFP_KERNEL
);
619 set_current_state(TASK_INTERRUPTIBLE
);
620 if (signalled() || kthread_should_stop()) {
621 set_current_state(TASK_RUNNING
);
624 schedule_timeout(msecs_to_jiffies(500));
626 rqstp
->rq_pages
[i
] = p
;
628 rqstp
->rq_pages
[i
++] = NULL
; /* this might be seen in nfs_read_actor */
629 BUG_ON(pages
>= RPCSVC_MAXPAGES
);
631 /* Make arg->head point to first page and arg->pages point to rest */
632 arg
= &rqstp
->rq_arg
;
633 arg
->head
[0].iov_base
= page_address(rqstp
->rq_pages
[0]);
634 arg
->head
[0].iov_len
= PAGE_SIZE
;
635 arg
->pages
= rqstp
->rq_pages
+ 1;
637 /* save at least one page for response */
638 arg
->page_len
= (pages
-2)*PAGE_SIZE
;
639 arg
->len
= (pages
-1)*PAGE_SIZE
;
640 arg
->tail
[0].iov_len
= 0;
644 if (signalled() || kthread_should_stop())
647 spin_lock_bh(&pool
->sp_lock
);
648 if (rqstp
->rq_waking
) {
649 rqstp
->rq_waking
= 0;
651 BUG_ON(pool
->sp_nwaking
< 0);
653 xprt
= svc_xprt_dequeue(pool
);
655 rqstp
->rq_xprt
= xprt
;
657 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
658 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
660 /* No data pending. Go to sleep */
661 svc_thread_enqueue(pool
, rqstp
);
664 * We have to be able to interrupt this wait
665 * to bring down the daemons ...
667 set_current_state(TASK_INTERRUPTIBLE
);
670 * checking kthread_should_stop() here allows us to avoid
671 * locking and signalling when stopping kthreads that call
672 * svc_recv. If the thread has already been woken up, then
673 * we can exit here without sleeping. If not, then it
674 * it'll be woken up quickly during the schedule_timeout
676 if (kthread_should_stop()) {
677 set_current_state(TASK_RUNNING
);
678 spin_unlock_bh(&pool
->sp_lock
);
682 add_wait_queue(&rqstp
->rq_wait
, &wait
);
683 spin_unlock_bh(&pool
->sp_lock
);
685 time_left
= schedule_timeout(timeout
);
689 spin_lock_bh(&pool
->sp_lock
);
690 remove_wait_queue(&rqstp
->rq_wait
, &wait
);
692 pool
->sp_stats
.threads_timedout
++;
694 xprt
= rqstp
->rq_xprt
;
696 svc_thread_dequeue(pool
, rqstp
);
697 spin_unlock_bh(&pool
->sp_lock
);
698 dprintk("svc: server %p, no data yet\n", rqstp
);
699 if (signalled() || kthread_should_stop())
705 spin_unlock_bh(&pool
->sp_lock
);
708 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
)) {
709 dprintk("svc_recv: found XPT_CLOSE\n");
710 svc_delete_xprt(xprt
);
711 } else if (test_bit(XPT_LISTENER
, &xprt
->xpt_flags
)) {
712 struct svc_xprt
*newxpt
;
713 newxpt
= xprt
->xpt_ops
->xpo_accept(xprt
);
716 * We know this module_get will succeed because the
717 * listener holds a reference too
719 __module_get(newxpt
->xpt_class
->xcl_owner
);
720 svc_check_conn_limits(xprt
->xpt_server
);
721 spin_lock_bh(&serv
->sv_lock
);
722 set_bit(XPT_TEMP
, &newxpt
->xpt_flags
);
723 list_add(&newxpt
->xpt_list
, &serv
->sv_tempsocks
);
725 if (serv
->sv_temptimer
.function
== NULL
) {
726 /* setup timer to age temp transports */
727 setup_timer(&serv
->sv_temptimer
,
729 (unsigned long)serv
);
730 mod_timer(&serv
->sv_temptimer
,
731 jiffies
+ svc_conn_age_period
* HZ
);
733 spin_unlock_bh(&serv
->sv_lock
);
734 svc_xprt_received(newxpt
);
736 svc_xprt_received(xprt
);
738 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
739 rqstp
, pool
->sp_id
, xprt
,
740 atomic_read(&xprt
->xpt_ref
.refcount
));
741 rqstp
->rq_deferred
= svc_deferred_dequeue(xprt
);
742 if (rqstp
->rq_deferred
) {
743 svc_xprt_received(xprt
);
744 len
= svc_deferred_recv(rqstp
);
746 len
= xprt
->xpt_ops
->xpo_recvfrom(rqstp
);
747 dprintk("svc: got len=%d\n", len
);
750 /* No data, incomplete (TCP) read, or accept() */
751 if (len
== 0 || len
== -EAGAIN
) {
752 rqstp
->rq_res
.len
= 0;
753 svc_xprt_release(rqstp
);
756 clear_bit(XPT_OLD
, &xprt
->xpt_flags
);
758 rqstp
->rq_secure
= svc_port_is_privileged(svc_addr(rqstp
));
759 rqstp
->rq_chandle
.defer
= svc_defer
;
762 serv
->sv_stats
->netcnt
++;
765 EXPORT_SYMBOL_GPL(svc_recv
);
770 void svc_drop(struct svc_rqst
*rqstp
)
772 dprintk("svc: xprt %p dropped request\n", rqstp
->rq_xprt
);
773 svc_xprt_release(rqstp
);
775 EXPORT_SYMBOL_GPL(svc_drop
);
778 * Return reply to client.
780 int svc_send(struct svc_rqst
*rqstp
)
782 struct svc_xprt
*xprt
;
786 xprt
= rqstp
->rq_xprt
;
790 /* release the receive skb before sending the reply */
791 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
793 /* calculate over-all length */
795 xb
->len
= xb
->head
[0].iov_len
+
799 /* Grab mutex to serialize outgoing data. */
800 mutex_lock(&xprt
->xpt_mutex
);
801 if (test_bit(XPT_DEAD
, &xprt
->xpt_flags
))
804 len
= xprt
->xpt_ops
->xpo_sendto(rqstp
);
805 mutex_unlock(&xprt
->xpt_mutex
);
806 svc_xprt_release(rqstp
);
808 if (len
== -ECONNREFUSED
|| len
== -ENOTCONN
|| len
== -EAGAIN
)
814 * Timer function to close old temporary transports, using
815 * a mark-and-sweep algorithm.
817 static void svc_age_temp_xprts(unsigned long closure
)
819 struct svc_serv
*serv
= (struct svc_serv
*)closure
;
820 struct svc_xprt
*xprt
;
821 struct list_head
*le
, *next
;
822 LIST_HEAD(to_be_aged
);
824 dprintk("svc_age_temp_xprts\n");
826 if (!spin_trylock_bh(&serv
->sv_lock
)) {
827 /* busy, try again 1 sec later */
828 dprintk("svc_age_temp_xprts: busy\n");
829 mod_timer(&serv
->sv_temptimer
, jiffies
+ HZ
);
833 list_for_each_safe(le
, next
, &serv
->sv_tempsocks
) {
834 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
836 /* First time through, just mark it OLD. Second time
837 * through, close it. */
838 if (!test_and_set_bit(XPT_OLD
, &xprt
->xpt_flags
))
840 if (atomic_read(&xprt
->xpt_ref
.refcount
) > 1
841 || test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
844 list_move(le
, &to_be_aged
);
845 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
846 set_bit(XPT_DETACHED
, &xprt
->xpt_flags
);
848 spin_unlock_bh(&serv
->sv_lock
);
850 while (!list_empty(&to_be_aged
)) {
851 le
= to_be_aged
.next
;
852 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
854 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
856 dprintk("queuing xprt %p for closing\n", xprt
);
858 /* a thread will dequeue and close it soon */
859 svc_xprt_enqueue(xprt
);
863 mod_timer(&serv
->sv_temptimer
, jiffies
+ svc_conn_age_period
* HZ
);
867 * Remove a dead transport
869 void svc_delete_xprt(struct svc_xprt
*xprt
)
871 struct svc_serv
*serv
= xprt
->xpt_server
;
872 struct svc_deferred_req
*dr
;
874 /* Only do this once */
875 if (test_and_set_bit(XPT_DEAD
, &xprt
->xpt_flags
))
878 dprintk("svc: svc_delete_xprt(%p)\n", xprt
);
879 xprt
->xpt_ops
->xpo_detach(xprt
);
881 spin_lock_bh(&serv
->sv_lock
);
882 if (!test_and_set_bit(XPT_DETACHED
, &xprt
->xpt_flags
))
883 list_del_init(&xprt
->xpt_list
);
885 * We used to delete the transport from whichever list
886 * it's sk_xprt.xpt_ready node was on, but we don't actually
887 * need to. This is because the only time we're called
888 * while still attached to a queue, the queue itself
889 * is about to be destroyed (in svc_destroy).
891 if (test_bit(XPT_TEMP
, &xprt
->xpt_flags
))
894 for (dr
= svc_deferred_dequeue(xprt
); dr
;
895 dr
= svc_deferred_dequeue(xprt
)) {
901 spin_unlock_bh(&serv
->sv_lock
);
904 void svc_close_xprt(struct svc_xprt
*xprt
)
906 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
907 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
))
908 /* someone else will have to effect the close */
912 svc_delete_xprt(xprt
);
913 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
916 EXPORT_SYMBOL_GPL(svc_close_xprt
);
918 void svc_close_all(struct list_head
*xprt_list
)
920 struct svc_xprt
*xprt
;
921 struct svc_xprt
*tmp
;
923 list_for_each_entry_safe(xprt
, tmp
, xprt_list
, xpt_list
) {
924 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
925 if (test_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
926 /* Waiting to be processed, but no threads left,
927 * So just remove it from the waiting list
929 list_del_init(&xprt
->xpt_ready
);
930 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
932 svc_close_xprt(xprt
);
937 * Handle defer and revisit of requests
940 static void svc_revisit(struct cache_deferred_req
*dreq
, int too_many
)
942 struct svc_deferred_req
*dr
=
943 container_of(dreq
, struct svc_deferred_req
, handle
);
944 struct svc_xprt
*xprt
= dr
->xprt
;
946 spin_lock(&xprt
->xpt_lock
);
947 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
948 if (too_many
|| test_bit(XPT_DEAD
, &xprt
->xpt_flags
)) {
949 spin_unlock(&xprt
->xpt_lock
);
950 dprintk("revisit canceled\n");
955 dprintk("revisit queued\n");
957 list_add(&dr
->handle
.recent
, &xprt
->xpt_deferred
);
958 spin_unlock(&xprt
->xpt_lock
);
959 svc_xprt_enqueue(xprt
);
964 * Save the request off for later processing. The request buffer looks
967 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
969 * This code can only handle requests that consist of an xprt-header
972 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
)
974 struct svc_rqst
*rqstp
= container_of(req
, struct svc_rqst
, rq_chandle
);
975 struct svc_deferred_req
*dr
;
977 if (rqstp
->rq_arg
.page_len
|| !rqstp
->rq_usedeferral
)
978 return NULL
; /* if more than a page, give up FIXME */
979 if (rqstp
->rq_deferred
) {
980 dr
= rqstp
->rq_deferred
;
981 rqstp
->rq_deferred
= NULL
;
985 /* FIXME maybe discard if size too large */
986 size
= sizeof(struct svc_deferred_req
) + rqstp
->rq_arg
.len
;
987 dr
= kmalloc(size
, GFP_KERNEL
);
991 dr
->handle
.owner
= rqstp
->rq_server
;
992 dr
->prot
= rqstp
->rq_prot
;
993 memcpy(&dr
->addr
, &rqstp
->rq_addr
, rqstp
->rq_addrlen
);
994 dr
->addrlen
= rqstp
->rq_addrlen
;
995 dr
->daddr
= rqstp
->rq_daddr
;
996 dr
->argslen
= rqstp
->rq_arg
.len
>> 2;
997 dr
->xprt_hlen
= rqstp
->rq_xprt_hlen
;
999 /* back up head to the start of the buffer and copy */
1000 skip
= rqstp
->rq_arg
.len
- rqstp
->rq_arg
.head
[0].iov_len
;
1001 memcpy(dr
->args
, rqstp
->rq_arg
.head
[0].iov_base
- skip
,
1004 svc_xprt_get(rqstp
->rq_xprt
);
1005 dr
->xprt
= rqstp
->rq_xprt
;
1007 dr
->handle
.revisit
= svc_revisit
;
1012 * recv data from a deferred request into an active one
1014 static int svc_deferred_recv(struct svc_rqst
*rqstp
)
1016 struct svc_deferred_req
*dr
= rqstp
->rq_deferred
;
1018 /* setup iov_base past transport header */
1019 rqstp
->rq_arg
.head
[0].iov_base
= dr
->args
+ (dr
->xprt_hlen
>>2);
1020 /* The iov_len does not include the transport header bytes */
1021 rqstp
->rq_arg
.head
[0].iov_len
= (dr
->argslen
<<2) - dr
->xprt_hlen
;
1022 rqstp
->rq_arg
.page_len
= 0;
1023 /* The rq_arg.len includes the transport header bytes */
1024 rqstp
->rq_arg
.len
= dr
->argslen
<<2;
1025 rqstp
->rq_prot
= dr
->prot
;
1026 memcpy(&rqstp
->rq_addr
, &dr
->addr
, dr
->addrlen
);
1027 rqstp
->rq_addrlen
= dr
->addrlen
;
1028 /* Save off transport header len in case we get deferred again */
1029 rqstp
->rq_xprt_hlen
= dr
->xprt_hlen
;
1030 rqstp
->rq_daddr
= dr
->daddr
;
1031 rqstp
->rq_respages
= rqstp
->rq_pages
;
1032 return (dr
->argslen
<<2) - dr
->xprt_hlen
;
1036 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
)
1038 struct svc_deferred_req
*dr
= NULL
;
1040 if (!test_bit(XPT_DEFERRED
, &xprt
->xpt_flags
))
1042 spin_lock(&xprt
->xpt_lock
);
1043 clear_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1044 if (!list_empty(&xprt
->xpt_deferred
)) {
1045 dr
= list_entry(xprt
->xpt_deferred
.next
,
1046 struct svc_deferred_req
,
1048 list_del_init(&dr
->handle
.recent
);
1049 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1051 spin_unlock(&xprt
->xpt_lock
);
1056 * Return the transport instance pointer for the endpoint accepting
1057 * connections/peer traffic from the specified transport class,
1058 * address family and port.
1060 * Specifying 0 for the address family or port is effectively a
1061 * wild-card, and will result in matching the first transport in the
1062 * service's list that has a matching class name.
1064 struct svc_xprt
*svc_find_xprt(struct svc_serv
*serv
, char *xcl_name
,
1067 struct svc_xprt
*xprt
;
1068 struct svc_xprt
*found
= NULL
;
1070 /* Sanity check the args */
1071 if (!serv
|| !xcl_name
)
1074 spin_lock_bh(&serv
->sv_lock
);
1075 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1076 if (strcmp(xprt
->xpt_class
->xcl_name
, xcl_name
))
1078 if (af
!= AF_UNSPEC
&& af
!= xprt
->xpt_local
.ss_family
)
1080 if (port
&& port
!= svc_xprt_local_port(xprt
))
1086 spin_unlock_bh(&serv
->sv_lock
);
1089 EXPORT_SYMBOL_GPL(svc_find_xprt
);
1092 * Format a buffer with a list of the active transports. A zero for
1093 * the buflen parameter disables target buffer overflow checking.
1095 int svc_xprt_names(struct svc_serv
*serv
, char *buf
, int buflen
)
1097 struct svc_xprt
*xprt
;
1102 /* Sanity check args */
1106 spin_lock_bh(&serv
->sv_lock
);
1107 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1108 len
= snprintf(xprt_str
, sizeof(xprt_str
),
1109 "%s %d\n", xprt
->xpt_class
->xcl_name
,
1110 svc_xprt_local_port(xprt
));
1111 /* If the string was truncated, replace with error string */
1112 if (len
>= sizeof(xprt_str
))
1113 strcpy(xprt_str
, "name-too-long\n");
1114 /* Don't overflow buffer */
1115 len
= strlen(xprt_str
);
1116 if (buflen
&& (len
+ totlen
>= buflen
))
1118 strcpy(buf
+totlen
, xprt_str
);
1121 spin_unlock_bh(&serv
->sv_lock
);
1124 EXPORT_SYMBOL_GPL(svc_xprt_names
);
1127 /*----------------------------------------------------------------------------*/
1129 static void *svc_pool_stats_start(struct seq_file
*m
, loff_t
*pos
)
1131 unsigned int pidx
= (unsigned int)*pos
;
1132 struct svc_serv
*serv
= m
->private;
1134 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx
);
1137 /* bump up the pseudo refcount while traversing */
1142 return SEQ_START_TOKEN
;
1143 return (pidx
> serv
->sv_nrpools
? NULL
: &serv
->sv_pools
[pidx
-1]);
1146 static void *svc_pool_stats_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
1148 struct svc_pool
*pool
= p
;
1149 struct svc_serv
*serv
= m
->private;
1151 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos
);
1153 if (p
== SEQ_START_TOKEN
) {
1154 pool
= &serv
->sv_pools
[0];
1156 unsigned int pidx
= (pool
- &serv
->sv_pools
[0]);
1157 if (pidx
< serv
->sv_nrpools
-1)
1158 pool
= &serv
->sv_pools
[pidx
+1];
1166 static void svc_pool_stats_stop(struct seq_file
*m
, void *p
)
1168 struct svc_serv
*serv
= m
->private;
1171 /* this function really, really should have been called svc_put() */
1176 static int svc_pool_stats_show(struct seq_file
*m
, void *p
)
1178 struct svc_pool
*pool
= p
;
1180 if (p
== SEQ_START_TOKEN
) {
1181 seq_puts(m
, "# pool packets-arrived sockets-enqueued threads-woken overloads-avoided threads-timedout\n");
1185 seq_printf(m
, "%u %lu %lu %lu %lu %lu\n",
1187 pool
->sp_stats
.packets
,
1188 pool
->sp_stats
.sockets_queued
,
1189 pool
->sp_stats
.threads_woken
,
1190 pool
->sp_stats
.overloads_avoided
,
1191 pool
->sp_stats
.threads_timedout
);
1196 static const struct seq_operations svc_pool_stats_seq_ops
= {
1197 .start
= svc_pool_stats_start
,
1198 .next
= svc_pool_stats_next
,
1199 .stop
= svc_pool_stats_stop
,
1200 .show
= svc_pool_stats_show
,
1203 int svc_pool_stats_open(struct svc_serv
*serv
, struct file
*file
)
1207 err
= seq_open(file
, &svc_pool_stats_seq_ops
);
1209 ((struct seq_file
*) file
->private_data
)->private = serv
;
1212 EXPORT_SYMBOL(svc_pool_stats_open
);
1214 /*----------------------------------------------------------------------------*/