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>
12 #include <linux/slab.h>
14 #include <linux/sunrpc/stats.h>
15 #include <linux/sunrpc/svc_xprt.h>
16 #include <linux/sunrpc/svcsock.h>
18 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
20 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
);
21 static int svc_deferred_recv(struct svc_rqst
*rqstp
);
22 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
);
23 static void svc_age_temp_xprts(unsigned long closure
);
25 /* apparently the "standard" is that clients close
26 * idle connections after 5 minutes, servers after
28 * http://www.connectathon.org/talks96/nfstcp.pdf
30 static int svc_conn_age_period
= 6*60;
32 /* List of registered transport classes */
33 static DEFINE_SPINLOCK(svc_xprt_class_lock
);
34 static LIST_HEAD(svc_xprt_class_list
);
36 /* SMP locking strategy:
38 * svc_pool->sp_lock protects most of the fields of that pool.
39 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
40 * when both need to be taken (rare), svc_serv->sv_lock is first.
41 * BKL protects svc_serv->sv_nrthread.
42 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
43 * and the ->sk_info_authunix cache.
45 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
46 * enqueued multiply. During normal transport processing this bit
47 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
48 * Providers should not manipulate this bit directly.
50 * Some flags can be set to certain values at any time
51 * providing that certain rules are followed:
54 * - Can be set or cleared at any time.
55 * - After a set, svc_xprt_enqueue must be called to enqueue
56 * the transport for processing.
57 * - After a clear, the transport must be read/accepted.
58 * If this succeeds, it must be set again.
60 * - Can set at any time. It is never cleared.
62 * - Can only be set while XPT_BUSY is held which ensures
63 * that no other thread will be using the transport or will
64 * try to set XPT_DEAD.
67 int svc_reg_xprt_class(struct svc_xprt_class
*xcl
)
69 struct svc_xprt_class
*cl
;
72 dprintk("svc: Adding svc transport class '%s'\n", xcl
->xcl_name
);
74 INIT_LIST_HEAD(&xcl
->xcl_list
);
75 spin_lock(&svc_xprt_class_lock
);
76 /* Make sure there isn't already a class with the same name */
77 list_for_each_entry(cl
, &svc_xprt_class_list
, xcl_list
) {
78 if (strcmp(xcl
->xcl_name
, cl
->xcl_name
) == 0)
81 list_add_tail(&xcl
->xcl_list
, &svc_xprt_class_list
);
84 spin_unlock(&svc_xprt_class_lock
);
87 EXPORT_SYMBOL_GPL(svc_reg_xprt_class
);
89 void svc_unreg_xprt_class(struct svc_xprt_class
*xcl
)
91 dprintk("svc: Removing svc transport class '%s'\n", xcl
->xcl_name
);
92 spin_lock(&svc_xprt_class_lock
);
93 list_del_init(&xcl
->xcl_list
);
94 spin_unlock(&svc_xprt_class_lock
);
96 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class
);
99 * Format the transport list for printing
101 int svc_print_xprts(char *buf
, int maxlen
)
103 struct svc_xprt_class
*xcl
;
108 spin_lock(&svc_xprt_class_lock
);
109 list_for_each_entry(xcl
, &svc_xprt_class_list
, xcl_list
) {
112 sprintf(tmpstr
, "%s %d\n", xcl
->xcl_name
, xcl
->xcl_max_payload
);
113 slen
= strlen(tmpstr
);
114 if (len
+ slen
> maxlen
)
119 spin_unlock(&svc_xprt_class_lock
);
124 static void svc_xprt_free(struct kref
*kref
)
126 struct svc_xprt
*xprt
=
127 container_of(kref
, struct svc_xprt
, xpt_ref
);
128 struct module
*owner
= xprt
->xpt_class
->xcl_owner
;
129 if (test_bit(XPT_CACHE_AUTH
, &xprt
->xpt_flags
))
130 svcauth_unix_info_release(xprt
);
131 put_net(xprt
->xpt_net
);
132 xprt
->xpt_ops
->xpo_free(xprt
);
136 void svc_xprt_put(struct svc_xprt
*xprt
)
138 kref_put(&xprt
->xpt_ref
, svc_xprt_free
);
140 EXPORT_SYMBOL_GPL(svc_xprt_put
);
143 * Called by transport drivers to initialize the transport independent
144 * portion of the transport instance.
146 void svc_xprt_init(struct svc_xprt_class
*xcl
, struct svc_xprt
*xprt
,
147 struct svc_serv
*serv
)
149 memset(xprt
, 0, sizeof(*xprt
));
150 xprt
->xpt_class
= xcl
;
151 xprt
->xpt_ops
= xcl
->xcl_ops
;
152 kref_init(&xprt
->xpt_ref
);
153 xprt
->xpt_server
= serv
;
154 INIT_LIST_HEAD(&xprt
->xpt_list
);
155 INIT_LIST_HEAD(&xprt
->xpt_ready
);
156 INIT_LIST_HEAD(&xprt
->xpt_deferred
);
157 INIT_LIST_HEAD(&xprt
->xpt_users
);
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");
162 xprt
->xpt_net
= get_net(&init_net
);
164 EXPORT_SYMBOL_GPL(svc_xprt_init
);
166 static struct svc_xprt
*__svc_xpo_create(struct svc_xprt_class
*xcl
,
167 struct svc_serv
*serv
,
170 const unsigned short port
,
173 struct sockaddr_in sin
= {
174 .sin_family
= AF_INET
,
175 .sin_addr
.s_addr
= htonl(INADDR_ANY
),
176 .sin_port
= htons(port
),
178 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
179 struct sockaddr_in6 sin6
= {
180 .sin6_family
= AF_INET6
,
181 .sin6_addr
= IN6ADDR_ANY_INIT
,
182 .sin6_port
= htons(port
),
184 #endif /* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
185 struct sockaddr
*sap
;
190 sap
= (struct sockaddr
*)&sin
;
193 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
195 sap
= (struct sockaddr
*)&sin6
;
198 #endif /* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
200 return ERR_PTR(-EAFNOSUPPORT
);
203 return xcl
->xcl_ops
->xpo_create(serv
, net
, sap
, len
, flags
);
206 int svc_create_xprt(struct svc_serv
*serv
, const char *xprt_name
,
207 struct net
*net
, const int family
,
208 const unsigned short port
, int flags
)
210 struct svc_xprt_class
*xcl
;
212 dprintk("svc: creating transport %s[%d]\n", xprt_name
, port
);
213 spin_lock(&svc_xprt_class_lock
);
214 list_for_each_entry(xcl
, &svc_xprt_class_list
, xcl_list
) {
215 struct svc_xprt
*newxprt
;
217 if (strcmp(xprt_name
, xcl
->xcl_name
))
220 if (!try_module_get(xcl
->xcl_owner
))
223 spin_unlock(&svc_xprt_class_lock
);
224 newxprt
= __svc_xpo_create(xcl
, serv
, net
, family
, port
, flags
);
225 if (IS_ERR(newxprt
)) {
226 module_put(xcl
->xcl_owner
);
227 return PTR_ERR(newxprt
);
230 clear_bit(XPT_TEMP
, &newxprt
->xpt_flags
);
231 spin_lock_bh(&serv
->sv_lock
);
232 list_add(&newxprt
->xpt_list
, &serv
->sv_permsocks
);
233 spin_unlock_bh(&serv
->sv_lock
);
234 clear_bit(XPT_BUSY
, &newxprt
->xpt_flags
);
235 return svc_xprt_local_port(newxprt
);
238 spin_unlock(&svc_xprt_class_lock
);
239 dprintk("svc: transport %s not found\n", xprt_name
);
241 /* This errno is exposed to user space. Provide a reasonable
242 * perror msg for a bad transport. */
243 return -EPROTONOSUPPORT
;
245 EXPORT_SYMBOL_GPL(svc_create_xprt
);
248 * Copy the local and remote xprt addresses to the rqstp structure
250 void svc_xprt_copy_addrs(struct svc_rqst
*rqstp
, struct svc_xprt
*xprt
)
252 struct sockaddr
*sin
;
254 memcpy(&rqstp
->rq_addr
, &xprt
->xpt_remote
, xprt
->xpt_remotelen
);
255 rqstp
->rq_addrlen
= xprt
->xpt_remotelen
;
258 * Destination address in request is needed for binding the
259 * source address in RPC replies/callbacks later.
261 sin
= (struct sockaddr
*)&xprt
->xpt_local
;
262 switch (sin
->sa_family
) {
264 rqstp
->rq_daddr
.addr
= ((struct sockaddr_in
*)sin
)->sin_addr
;
267 rqstp
->rq_daddr
.addr6
= ((struct sockaddr_in6
*)sin
)->sin6_addr
;
271 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs
);
274 * svc_print_addr - Format rq_addr field for printing
275 * @rqstp: svc_rqst struct containing address to print
276 * @buf: target buffer for formatted address
277 * @len: length of target buffer
280 char *svc_print_addr(struct svc_rqst
*rqstp
, char *buf
, size_t len
)
282 return __svc_print_addr(svc_addr(rqstp
), buf
, len
);
284 EXPORT_SYMBOL_GPL(svc_print_addr
);
287 * Queue up an idle server thread. Must have pool->sp_lock held.
288 * Note: this is really a stack rather than a queue, so that we only
289 * use as many different threads as we need, and the rest don't pollute
292 static void svc_thread_enqueue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
294 list_add(&rqstp
->rq_list
, &pool
->sp_threads
);
298 * Dequeue an nfsd thread. Must have pool->sp_lock held.
300 static void svc_thread_dequeue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
302 list_del(&rqstp
->rq_list
);
306 * Queue up a transport with data pending. If there are idle nfsd
307 * processes, wake 'em up.
310 void svc_xprt_enqueue(struct svc_xprt
*xprt
)
312 struct svc_serv
*serv
= xprt
->xpt_server
;
313 struct svc_pool
*pool
;
314 struct svc_rqst
*rqstp
;
317 if (!(xprt
->xpt_flags
&
318 ((1<<XPT_CONN
)|(1<<XPT_DATA
)|(1<<XPT_CLOSE
)|(1<<XPT_DEFERRED
))))
322 pool
= svc_pool_for_cpu(xprt
->xpt_server
, cpu
);
325 spin_lock_bh(&pool
->sp_lock
);
327 if (!list_empty(&pool
->sp_threads
) &&
328 !list_empty(&pool
->sp_sockets
))
331 "threads and transports both waiting??\n");
333 pool
->sp_stats
.packets
++;
335 /* Mark transport as busy. It will remain in this state until
336 * the provider calls svc_xprt_received. We update XPT_BUSY
337 * atomically because it also guards against trying to enqueue
338 * the transport twice.
340 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
341 /* Don't enqueue transport while already enqueued */
342 dprintk("svc: transport %p busy, not enqueued\n", xprt
);
345 BUG_ON(xprt
->xpt_pool
!= NULL
);
346 xprt
->xpt_pool
= pool
;
348 /* Handle pending connection */
349 if (test_bit(XPT_CONN
, &xprt
->xpt_flags
))
352 /* Handle close in-progress */
353 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
))
356 /* Check if we have space to reply to a request */
357 if (!xprt
->xpt_ops
->xpo_has_wspace(xprt
)) {
358 /* Don't enqueue while not enough space for reply */
359 dprintk("svc: no write space, transport %p not enqueued\n",
361 xprt
->xpt_pool
= NULL
;
362 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
367 if (!list_empty(&pool
->sp_threads
)) {
368 rqstp
= list_entry(pool
->sp_threads
.next
,
371 dprintk("svc: transport %p served by daemon %p\n",
373 svc_thread_dequeue(pool
, rqstp
);
376 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
377 rqstp
, rqstp
->rq_xprt
);
378 rqstp
->rq_xprt
= xprt
;
380 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
381 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
382 pool
->sp_stats
.threads_woken
++;
383 BUG_ON(xprt
->xpt_pool
!= pool
);
384 wake_up(&rqstp
->rq_wait
);
386 dprintk("svc: transport %p put into queue\n", xprt
);
387 list_add_tail(&xprt
->xpt_ready
, &pool
->sp_sockets
);
388 pool
->sp_stats
.sockets_queued
++;
389 BUG_ON(xprt
->xpt_pool
!= pool
);
393 spin_unlock_bh(&pool
->sp_lock
);
395 EXPORT_SYMBOL_GPL(svc_xprt_enqueue
);
398 * Dequeue the first transport. Must be called with the pool->sp_lock held.
400 static struct svc_xprt
*svc_xprt_dequeue(struct svc_pool
*pool
)
402 struct svc_xprt
*xprt
;
404 if (list_empty(&pool
->sp_sockets
))
407 xprt
= list_entry(pool
->sp_sockets
.next
,
408 struct svc_xprt
, xpt_ready
);
409 list_del_init(&xprt
->xpt_ready
);
411 dprintk("svc: transport %p dequeued, inuse=%d\n",
412 xprt
, atomic_read(&xprt
->xpt_ref
.refcount
));
418 * svc_xprt_received conditionally queues the transport for processing
419 * by another thread. The caller must hold the XPT_BUSY bit and must
420 * not thereafter touch transport data.
422 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
423 * insufficient) data.
425 void svc_xprt_received(struct svc_xprt
*xprt
)
427 BUG_ON(!test_bit(XPT_BUSY
, &xprt
->xpt_flags
));
428 xprt
->xpt_pool
= NULL
;
429 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
430 svc_xprt_enqueue(xprt
);
432 EXPORT_SYMBOL_GPL(svc_xprt_received
);
435 * svc_reserve - change the space reserved for the reply to a request.
436 * @rqstp: The request in question
437 * @space: new max space to reserve
439 * Each request reserves some space on the output queue of the transport
440 * to make sure the reply fits. This function reduces that reserved
441 * space to be the amount of space used already, plus @space.
444 void svc_reserve(struct svc_rqst
*rqstp
, int space
)
446 space
+= rqstp
->rq_res
.head
[0].iov_len
;
448 if (space
< rqstp
->rq_reserved
) {
449 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
450 atomic_sub((rqstp
->rq_reserved
- space
), &xprt
->xpt_reserved
);
451 rqstp
->rq_reserved
= space
;
453 svc_xprt_enqueue(xprt
);
456 EXPORT_SYMBOL_GPL(svc_reserve
);
458 static void svc_xprt_release(struct svc_rqst
*rqstp
)
460 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
462 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
464 kfree(rqstp
->rq_deferred
);
465 rqstp
->rq_deferred
= NULL
;
467 svc_free_res_pages(rqstp
);
468 rqstp
->rq_res
.page_len
= 0;
469 rqstp
->rq_res
.page_base
= 0;
471 /* Reset response buffer and release
473 * But first, check that enough space was reserved
474 * for the reply, otherwise we have a bug!
476 if ((rqstp
->rq_res
.len
) > rqstp
->rq_reserved
)
477 printk(KERN_ERR
"RPC request reserved %d but used %d\n",
481 rqstp
->rq_res
.head
[0].iov_len
= 0;
482 svc_reserve(rqstp
, 0);
483 rqstp
->rq_xprt
= NULL
;
489 * External function to wake up a server waiting for data
490 * This really only makes sense for services like lockd
491 * which have exactly one thread anyway.
493 void svc_wake_up(struct svc_serv
*serv
)
495 struct svc_rqst
*rqstp
;
497 struct svc_pool
*pool
;
499 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
500 pool
= &serv
->sv_pools
[i
];
502 spin_lock_bh(&pool
->sp_lock
);
503 if (!list_empty(&pool
->sp_threads
)) {
504 rqstp
= list_entry(pool
->sp_threads
.next
,
507 dprintk("svc: daemon %p woken up.\n", rqstp
);
509 svc_thread_dequeue(pool, rqstp);
510 rqstp->rq_xprt = NULL;
512 wake_up(&rqstp
->rq_wait
);
514 spin_unlock_bh(&pool
->sp_lock
);
517 EXPORT_SYMBOL_GPL(svc_wake_up
);
519 int svc_port_is_privileged(struct sockaddr
*sin
)
521 switch (sin
->sa_family
) {
523 return ntohs(((struct sockaddr_in
*)sin
)->sin_port
)
526 return ntohs(((struct sockaddr_in6
*)sin
)->sin6_port
)
534 * Make sure that we don't have too many active connections. If we have,
535 * something must be dropped. It's not clear what will happen if we allow
536 * "too many" connections, but when dealing with network-facing software,
537 * we have to code defensively. Here we do that by imposing hard limits.
539 * There's no point in trying to do random drop here for DoS
540 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
541 * attacker can easily beat that.
543 * The only somewhat efficient mechanism would be if drop old
544 * connections from the same IP first. But right now we don't even
545 * record the client IP in svc_sock.
547 * single-threaded services that expect a lot of clients will probably
548 * need to set sv_maxconn to override the default value which is based
549 * on the number of threads
551 static void svc_check_conn_limits(struct svc_serv
*serv
)
553 unsigned int limit
= serv
->sv_maxconn
? serv
->sv_maxconn
:
554 (serv
->sv_nrthreads
+3) * 20;
556 if (serv
->sv_tmpcnt
> limit
) {
557 struct svc_xprt
*xprt
= NULL
;
558 spin_lock_bh(&serv
->sv_lock
);
559 if (!list_empty(&serv
->sv_tempsocks
)) {
560 if (net_ratelimit()) {
561 /* Try to help the admin */
562 printk(KERN_NOTICE
"%s: too many open "
563 "connections, consider increasing %s\n",
564 serv
->sv_name
, serv
->sv_maxconn
?
565 "the max number of connections." :
566 "the number of threads.");
569 * Always select the oldest connection. It's not fair,
572 xprt
= list_entry(serv
->sv_tempsocks
.prev
,
575 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
578 spin_unlock_bh(&serv
->sv_lock
);
581 svc_xprt_enqueue(xprt
);
588 * Receive the next request on any transport. This code is carefully
589 * organised not to touch any cachelines in the shared svc_serv
590 * structure, only cachelines in the local svc_pool.
592 int svc_recv(struct svc_rqst
*rqstp
, long timeout
)
594 struct svc_xprt
*xprt
= NULL
;
595 struct svc_serv
*serv
= rqstp
->rq_server
;
596 struct svc_pool
*pool
= rqstp
->rq_pool
;
600 DECLARE_WAITQUEUE(wait
, current
);
603 dprintk("svc: server %p waiting for data (to = %ld)\n",
608 "svc_recv: service %p, transport not NULL!\n",
610 if (waitqueue_active(&rqstp
->rq_wait
))
612 "svc_recv: service %p, wait queue active!\n",
615 /* now allocate needed pages. If we get a failure, sleep briefly */
616 pages
= (serv
->sv_max_mesg
+ PAGE_SIZE
) / PAGE_SIZE
;
617 for (i
= 0; i
< pages
; i
++)
618 while (rqstp
->rq_pages
[i
] == NULL
) {
619 struct page
*p
= alloc_page(GFP_KERNEL
);
621 set_current_state(TASK_INTERRUPTIBLE
);
622 if (signalled() || kthread_should_stop()) {
623 set_current_state(TASK_RUNNING
);
626 schedule_timeout(msecs_to_jiffies(500));
628 rqstp
->rq_pages
[i
] = p
;
630 rqstp
->rq_pages
[i
++] = NULL
; /* this might be seen in nfs_read_actor */
631 BUG_ON(pages
>= RPCSVC_MAXPAGES
);
633 /* Make arg->head point to first page and arg->pages point to rest */
634 arg
= &rqstp
->rq_arg
;
635 arg
->head
[0].iov_base
= page_address(rqstp
->rq_pages
[0]);
636 arg
->head
[0].iov_len
= PAGE_SIZE
;
637 arg
->pages
= rqstp
->rq_pages
+ 1;
639 /* save at least one page for response */
640 arg
->page_len
= (pages
-2)*PAGE_SIZE
;
641 arg
->len
= (pages
-1)*PAGE_SIZE
;
642 arg
->tail
[0].iov_len
= 0;
646 if (signalled() || kthread_should_stop())
649 /* Normally we will wait up to 5 seconds for any required
650 * cache information to be provided.
652 rqstp
->rq_chandle
.thread_wait
= 5*HZ
;
654 spin_lock_bh(&pool
->sp_lock
);
655 xprt
= svc_xprt_dequeue(pool
);
657 rqstp
->rq_xprt
= xprt
;
659 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
660 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
662 /* As there is a shortage of threads and this request
663 * had to be queued, don't allow the thread to wait so
664 * long for cache updates.
666 rqstp
->rq_chandle
.thread_wait
= 1*HZ
;
668 /* No data pending. Go to sleep */
669 svc_thread_enqueue(pool
, rqstp
);
672 * We have to be able to interrupt this wait
673 * to bring down the daemons ...
675 set_current_state(TASK_INTERRUPTIBLE
);
678 * checking kthread_should_stop() here allows us to avoid
679 * locking and signalling when stopping kthreads that call
680 * svc_recv. If the thread has already been woken up, then
681 * we can exit here without sleeping. If not, then it
682 * it'll be woken up quickly during the schedule_timeout
684 if (kthread_should_stop()) {
685 set_current_state(TASK_RUNNING
);
686 spin_unlock_bh(&pool
->sp_lock
);
690 add_wait_queue(&rqstp
->rq_wait
, &wait
);
691 spin_unlock_bh(&pool
->sp_lock
);
693 time_left
= schedule_timeout(timeout
);
697 spin_lock_bh(&pool
->sp_lock
);
698 remove_wait_queue(&rqstp
->rq_wait
, &wait
);
700 pool
->sp_stats
.threads_timedout
++;
702 xprt
= rqstp
->rq_xprt
;
704 svc_thread_dequeue(pool
, rqstp
);
705 spin_unlock_bh(&pool
->sp_lock
);
706 dprintk("svc: server %p, no data yet\n", rqstp
);
707 if (signalled() || kthread_should_stop())
713 spin_unlock_bh(&pool
->sp_lock
);
716 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
)) {
717 dprintk("svc_recv: found XPT_CLOSE\n");
718 svc_delete_xprt(xprt
);
719 } else if (test_bit(XPT_LISTENER
, &xprt
->xpt_flags
)) {
720 struct svc_xprt
*newxpt
;
721 newxpt
= xprt
->xpt_ops
->xpo_accept(xprt
);
724 * We know this module_get will succeed because the
725 * listener holds a reference too
727 __module_get(newxpt
->xpt_class
->xcl_owner
);
728 svc_check_conn_limits(xprt
->xpt_server
);
729 spin_lock_bh(&serv
->sv_lock
);
730 set_bit(XPT_TEMP
, &newxpt
->xpt_flags
);
731 list_add(&newxpt
->xpt_list
, &serv
->sv_tempsocks
);
733 if (serv
->sv_temptimer
.function
== NULL
) {
734 /* setup timer to age temp transports */
735 setup_timer(&serv
->sv_temptimer
,
737 (unsigned long)serv
);
738 mod_timer(&serv
->sv_temptimer
,
739 jiffies
+ svc_conn_age_period
* HZ
);
741 spin_unlock_bh(&serv
->sv_lock
);
742 svc_xprt_received(newxpt
);
744 svc_xprt_received(xprt
);
746 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
747 rqstp
, pool
->sp_id
, xprt
,
748 atomic_read(&xprt
->xpt_ref
.refcount
));
749 rqstp
->rq_deferred
= svc_deferred_dequeue(xprt
);
750 if (rqstp
->rq_deferred
) {
751 svc_xprt_received(xprt
);
752 len
= svc_deferred_recv(rqstp
);
754 len
= xprt
->xpt_ops
->xpo_recvfrom(rqstp
);
755 svc_xprt_received(xprt
);
757 dprintk("svc: got len=%d\n", len
);
760 /* No data, incomplete (TCP) read, or accept() */
761 if (len
== 0 || len
== -EAGAIN
) {
762 rqstp
->rq_res
.len
= 0;
763 svc_xprt_release(rqstp
);
766 clear_bit(XPT_OLD
, &xprt
->xpt_flags
);
768 rqstp
->rq_secure
= svc_port_is_privileged(svc_addr(rqstp
));
769 rqstp
->rq_chandle
.defer
= svc_defer
;
772 serv
->sv_stats
->netcnt
++;
775 EXPORT_SYMBOL_GPL(svc_recv
);
780 void svc_drop(struct svc_rqst
*rqstp
)
782 dprintk("svc: xprt %p dropped request\n", rqstp
->rq_xprt
);
783 svc_xprt_release(rqstp
);
785 EXPORT_SYMBOL_GPL(svc_drop
);
788 * Return reply to client.
790 int svc_send(struct svc_rqst
*rqstp
)
792 struct svc_xprt
*xprt
;
796 xprt
= rqstp
->rq_xprt
;
800 /* release the receive skb before sending the reply */
801 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
803 /* calculate over-all length */
805 xb
->len
= xb
->head
[0].iov_len
+
809 /* Grab mutex to serialize outgoing data. */
810 mutex_lock(&xprt
->xpt_mutex
);
811 if (test_bit(XPT_DEAD
, &xprt
->xpt_flags
))
814 len
= xprt
->xpt_ops
->xpo_sendto(rqstp
);
815 mutex_unlock(&xprt
->xpt_mutex
);
816 rpc_wake_up(&xprt
->xpt_bc_pending
);
817 svc_xprt_release(rqstp
);
819 if (len
== -ECONNREFUSED
|| len
== -ENOTCONN
|| len
== -EAGAIN
)
825 * Timer function to close old temporary transports, using
826 * a mark-and-sweep algorithm.
828 static void svc_age_temp_xprts(unsigned long closure
)
830 struct svc_serv
*serv
= (struct svc_serv
*)closure
;
831 struct svc_xprt
*xprt
;
832 struct list_head
*le
, *next
;
833 LIST_HEAD(to_be_aged
);
835 dprintk("svc_age_temp_xprts\n");
837 if (!spin_trylock_bh(&serv
->sv_lock
)) {
838 /* busy, try again 1 sec later */
839 dprintk("svc_age_temp_xprts: busy\n");
840 mod_timer(&serv
->sv_temptimer
, jiffies
+ HZ
);
844 list_for_each_safe(le
, next
, &serv
->sv_tempsocks
) {
845 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
847 /* First time through, just mark it OLD. Second time
848 * through, close it. */
849 if (!test_and_set_bit(XPT_OLD
, &xprt
->xpt_flags
))
851 if (atomic_read(&xprt
->xpt_ref
.refcount
) > 1 ||
852 test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
855 list_move(le
, &to_be_aged
);
856 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
857 set_bit(XPT_DETACHED
, &xprt
->xpt_flags
);
859 spin_unlock_bh(&serv
->sv_lock
);
861 while (!list_empty(&to_be_aged
)) {
862 le
= to_be_aged
.next
;
863 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
865 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
867 dprintk("queuing xprt %p for closing\n", xprt
);
869 /* a thread will dequeue and close it soon */
870 svc_xprt_enqueue(xprt
);
874 mod_timer(&serv
->sv_temptimer
, jiffies
+ svc_conn_age_period
* HZ
);
877 static void call_xpt_users(struct svc_xprt
*xprt
)
879 struct svc_xpt_user
*u
;
881 spin_lock(&xprt
->xpt_lock
);
882 while (!list_empty(&xprt
->xpt_users
)) {
883 u
= list_first_entry(&xprt
->xpt_users
, struct svc_xpt_user
, list
);
887 spin_unlock(&xprt
->xpt_lock
);
891 * Remove a dead transport
893 void svc_delete_xprt(struct svc_xprt
*xprt
)
895 struct svc_serv
*serv
= xprt
->xpt_server
;
896 struct svc_deferred_req
*dr
;
898 /* Only do this once */
899 if (test_and_set_bit(XPT_DEAD
, &xprt
->xpt_flags
))
902 dprintk("svc: svc_delete_xprt(%p)\n", xprt
);
903 xprt
->xpt_ops
->xpo_detach(xprt
);
905 spin_lock_bh(&serv
->sv_lock
);
906 if (!test_and_set_bit(XPT_DETACHED
, &xprt
->xpt_flags
))
907 list_del_init(&xprt
->xpt_list
);
909 * We used to delete the transport from whichever list
910 * it's sk_xprt.xpt_ready node was on, but we don't actually
911 * need to. This is because the only time we're called
912 * while still attached to a queue, the queue itself
913 * is about to be destroyed (in svc_destroy).
915 if (test_bit(XPT_TEMP
, &xprt
->xpt_flags
))
917 spin_unlock_bh(&serv
->sv_lock
);
919 while ((dr
= svc_deferred_dequeue(xprt
)) != NULL
)
922 call_xpt_users(xprt
);
926 void svc_close_xprt(struct svc_xprt
*xprt
)
928 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
929 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
))
930 /* someone else will have to effect the close */
933 svc_delete_xprt(xprt
);
935 EXPORT_SYMBOL_GPL(svc_close_xprt
);
937 void svc_close_all(struct list_head
*xprt_list
)
939 struct svc_xprt
*xprt
;
940 struct svc_xprt
*tmp
;
942 list_for_each_entry_safe(xprt
, tmp
, xprt_list
, xpt_list
) {
943 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
944 if (test_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
945 /* Waiting to be processed, but no threads left,
946 * So just remove it from the waiting list
948 list_del_init(&xprt
->xpt_ready
);
949 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
951 svc_close_xprt(xprt
);
956 * Handle defer and revisit of requests
959 static void svc_revisit(struct cache_deferred_req
*dreq
, int too_many
)
961 struct svc_deferred_req
*dr
=
962 container_of(dreq
, struct svc_deferred_req
, handle
);
963 struct svc_xprt
*xprt
= dr
->xprt
;
965 spin_lock(&xprt
->xpt_lock
);
966 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
967 if (too_many
|| test_bit(XPT_DEAD
, &xprt
->xpt_flags
)) {
968 spin_unlock(&xprt
->xpt_lock
);
969 dprintk("revisit canceled\n");
974 dprintk("revisit queued\n");
976 list_add(&dr
->handle
.recent
, &xprt
->xpt_deferred
);
977 spin_unlock(&xprt
->xpt_lock
);
978 svc_xprt_enqueue(xprt
);
983 * Save the request off for later processing. The request buffer looks
986 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
988 * This code can only handle requests that consist of an xprt-header
991 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
)
993 struct svc_rqst
*rqstp
= container_of(req
, struct svc_rqst
, rq_chandle
);
994 struct svc_deferred_req
*dr
;
996 if (rqstp
->rq_arg
.page_len
|| !rqstp
->rq_usedeferral
)
997 return NULL
; /* if more than a page, give up FIXME */
998 if (rqstp
->rq_deferred
) {
999 dr
= rqstp
->rq_deferred
;
1000 rqstp
->rq_deferred
= NULL
;
1004 /* FIXME maybe discard if size too large */
1005 size
= sizeof(struct svc_deferred_req
) + rqstp
->rq_arg
.len
;
1006 dr
= kmalloc(size
, GFP_KERNEL
);
1010 dr
->handle
.owner
= rqstp
->rq_server
;
1011 dr
->prot
= rqstp
->rq_prot
;
1012 memcpy(&dr
->addr
, &rqstp
->rq_addr
, rqstp
->rq_addrlen
);
1013 dr
->addrlen
= rqstp
->rq_addrlen
;
1014 dr
->daddr
= rqstp
->rq_daddr
;
1015 dr
->argslen
= rqstp
->rq_arg
.len
>> 2;
1016 dr
->xprt_hlen
= rqstp
->rq_xprt_hlen
;
1018 /* back up head to the start of the buffer and copy */
1019 skip
= rqstp
->rq_arg
.len
- rqstp
->rq_arg
.head
[0].iov_len
;
1020 memcpy(dr
->args
, rqstp
->rq_arg
.head
[0].iov_base
- skip
,
1023 svc_xprt_get(rqstp
->rq_xprt
);
1024 dr
->xprt
= rqstp
->rq_xprt
;
1026 dr
->handle
.revisit
= svc_revisit
;
1031 * recv data from a deferred request into an active one
1033 static int svc_deferred_recv(struct svc_rqst
*rqstp
)
1035 struct svc_deferred_req
*dr
= rqstp
->rq_deferred
;
1037 /* setup iov_base past transport header */
1038 rqstp
->rq_arg
.head
[0].iov_base
= dr
->args
+ (dr
->xprt_hlen
>>2);
1039 /* The iov_len does not include the transport header bytes */
1040 rqstp
->rq_arg
.head
[0].iov_len
= (dr
->argslen
<<2) - dr
->xprt_hlen
;
1041 rqstp
->rq_arg
.page_len
= 0;
1042 /* The rq_arg.len includes the transport header bytes */
1043 rqstp
->rq_arg
.len
= dr
->argslen
<<2;
1044 rqstp
->rq_prot
= dr
->prot
;
1045 memcpy(&rqstp
->rq_addr
, &dr
->addr
, dr
->addrlen
);
1046 rqstp
->rq_addrlen
= dr
->addrlen
;
1047 /* Save off transport header len in case we get deferred again */
1048 rqstp
->rq_xprt_hlen
= dr
->xprt_hlen
;
1049 rqstp
->rq_daddr
= dr
->daddr
;
1050 rqstp
->rq_respages
= rqstp
->rq_pages
;
1051 return (dr
->argslen
<<2) - dr
->xprt_hlen
;
1055 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
)
1057 struct svc_deferred_req
*dr
= NULL
;
1059 if (!test_bit(XPT_DEFERRED
, &xprt
->xpt_flags
))
1061 spin_lock(&xprt
->xpt_lock
);
1062 clear_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1063 if (!list_empty(&xprt
->xpt_deferred
)) {
1064 dr
= list_entry(xprt
->xpt_deferred
.next
,
1065 struct svc_deferred_req
,
1067 list_del_init(&dr
->handle
.recent
);
1068 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1070 spin_unlock(&xprt
->xpt_lock
);
1075 * svc_find_xprt - find an RPC transport instance
1076 * @serv: pointer to svc_serv to search
1077 * @xcl_name: C string containing transport's class name
1078 * @af: Address family of transport's local address
1079 * @port: transport's IP port number
1081 * Return the transport instance pointer for the endpoint accepting
1082 * connections/peer traffic from the specified transport class,
1083 * address family and port.
1085 * Specifying 0 for the address family or port is effectively a
1086 * wild-card, and will result in matching the first transport in the
1087 * service's list that has a matching class name.
1089 struct svc_xprt
*svc_find_xprt(struct svc_serv
*serv
, const char *xcl_name
,
1090 const sa_family_t af
, const unsigned short port
)
1092 struct svc_xprt
*xprt
;
1093 struct svc_xprt
*found
= NULL
;
1095 /* Sanity check the args */
1096 if (serv
== NULL
|| xcl_name
== NULL
)
1099 spin_lock_bh(&serv
->sv_lock
);
1100 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1101 if (strcmp(xprt
->xpt_class
->xcl_name
, xcl_name
))
1103 if (af
!= AF_UNSPEC
&& af
!= xprt
->xpt_local
.ss_family
)
1105 if (port
!= 0 && port
!= svc_xprt_local_port(xprt
))
1111 spin_unlock_bh(&serv
->sv_lock
);
1114 EXPORT_SYMBOL_GPL(svc_find_xprt
);
1116 static int svc_one_xprt_name(const struct svc_xprt
*xprt
,
1117 char *pos
, int remaining
)
1121 len
= snprintf(pos
, remaining
, "%s %u\n",
1122 xprt
->xpt_class
->xcl_name
,
1123 svc_xprt_local_port(xprt
));
1124 if (len
>= remaining
)
1125 return -ENAMETOOLONG
;
1130 * svc_xprt_names - format a buffer with a list of transport names
1131 * @serv: pointer to an RPC service
1132 * @buf: pointer to a buffer to be filled in
1133 * @buflen: length of buffer to be filled in
1135 * Fills in @buf with a string containing a list of transport names,
1136 * each name terminated with '\n'.
1138 * Returns positive length of the filled-in string on success; otherwise
1139 * a negative errno value is returned if an error occurs.
1141 int svc_xprt_names(struct svc_serv
*serv
, char *buf
, const int buflen
)
1143 struct svc_xprt
*xprt
;
1147 /* Sanity check args */
1151 spin_lock_bh(&serv
->sv_lock
);
1155 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1156 len
= svc_one_xprt_name(xprt
, pos
, buflen
- totlen
);
1168 spin_unlock_bh(&serv
->sv_lock
);
1171 EXPORT_SYMBOL_GPL(svc_xprt_names
);
1174 /*----------------------------------------------------------------------------*/
1176 static void *svc_pool_stats_start(struct seq_file
*m
, loff_t
*pos
)
1178 unsigned int pidx
= (unsigned int)*pos
;
1179 struct svc_serv
*serv
= m
->private;
1181 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx
);
1184 return SEQ_START_TOKEN
;
1185 return (pidx
> serv
->sv_nrpools
? NULL
: &serv
->sv_pools
[pidx
-1]);
1188 static void *svc_pool_stats_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
1190 struct svc_pool
*pool
= p
;
1191 struct svc_serv
*serv
= m
->private;
1193 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos
);
1195 if (p
== SEQ_START_TOKEN
) {
1196 pool
= &serv
->sv_pools
[0];
1198 unsigned int pidx
= (pool
- &serv
->sv_pools
[0]);
1199 if (pidx
< serv
->sv_nrpools
-1)
1200 pool
= &serv
->sv_pools
[pidx
+1];
1208 static void svc_pool_stats_stop(struct seq_file
*m
, void *p
)
1212 static int svc_pool_stats_show(struct seq_file
*m
, void *p
)
1214 struct svc_pool
*pool
= p
;
1216 if (p
== SEQ_START_TOKEN
) {
1217 seq_puts(m
, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1221 seq_printf(m
, "%u %lu %lu %lu %lu\n",
1223 pool
->sp_stats
.packets
,
1224 pool
->sp_stats
.sockets_queued
,
1225 pool
->sp_stats
.threads_woken
,
1226 pool
->sp_stats
.threads_timedout
);
1231 static const struct seq_operations svc_pool_stats_seq_ops
= {
1232 .start
= svc_pool_stats_start
,
1233 .next
= svc_pool_stats_next
,
1234 .stop
= svc_pool_stats_stop
,
1235 .show
= svc_pool_stats_show
,
1238 int svc_pool_stats_open(struct svc_serv
*serv
, struct file
*file
)
1242 err
= seq_open(file
, &svc_pool_stats_seq_ops
);
1244 ((struct seq_file
*) file
->private_data
)->private = serv
;
1247 EXPORT_SYMBOL(svc_pool_stats_open
);
1249 /*----------------------------------------------------------------------------*/