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>
11 #include <linux/slab.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 svc_xprt_class
*xcl
;
107 spin_lock(&svc_xprt_class_lock
);
108 list_for_each_entry(xcl
, &svc_xprt_class_list
, xcl_list
) {
111 sprintf(tmpstr
, "%s %d\n", xcl
->xcl_name
, xcl
->xcl_max_payload
);
112 slen
= strlen(tmpstr
);
113 if (len
+ slen
> maxlen
)
118 spin_unlock(&svc_xprt_class_lock
);
123 static void svc_xprt_free(struct kref
*kref
)
125 struct svc_xprt
*xprt
=
126 container_of(kref
, struct svc_xprt
, xpt_ref
);
127 struct module
*owner
= xprt
->xpt_class
->xcl_owner
;
128 if (test_bit(XPT_CACHE_AUTH
, &xprt
->xpt_flags
))
129 svcauth_unix_info_release(xprt
);
130 put_net(xprt
->xpt_net
);
131 xprt
->xpt_ops
->xpo_free(xprt
);
135 void svc_xprt_put(struct svc_xprt
*xprt
)
137 kref_put(&xprt
->xpt_ref
, svc_xprt_free
);
139 EXPORT_SYMBOL_GPL(svc_xprt_put
);
142 * Called by transport drivers to initialize the transport independent
143 * portion of the transport instance.
145 void svc_xprt_init(struct svc_xprt_class
*xcl
, struct svc_xprt
*xprt
,
146 struct svc_serv
*serv
)
148 memset(xprt
, 0, sizeof(*xprt
));
149 xprt
->xpt_class
= xcl
;
150 xprt
->xpt_ops
= xcl
->xcl_ops
;
151 kref_init(&xprt
->xpt_ref
);
152 xprt
->xpt_server
= serv
;
153 INIT_LIST_HEAD(&xprt
->xpt_list
);
154 INIT_LIST_HEAD(&xprt
->xpt_ready
);
155 INIT_LIST_HEAD(&xprt
->xpt_deferred
);
156 INIT_LIST_HEAD(&xprt
->xpt_users
);
157 mutex_init(&xprt
->xpt_mutex
);
158 spin_lock_init(&xprt
->xpt_lock
);
159 set_bit(XPT_BUSY
, &xprt
->xpt_flags
);
160 rpc_init_wait_queue(&xprt
->xpt_bc_pending
, "xpt_bc_pending");
161 xprt
->xpt_net
= get_net(&init_net
);
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
,
169 const unsigned short port
,
172 struct sockaddr_in sin
= {
173 .sin_family
= AF_INET
,
174 .sin_addr
.s_addr
= htonl(INADDR_ANY
),
175 .sin_port
= htons(port
),
177 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
178 struct sockaddr_in6 sin6
= {
179 .sin6_family
= AF_INET6
,
180 .sin6_addr
= IN6ADDR_ANY_INIT
,
181 .sin6_port
= htons(port
),
183 #endif /* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
184 struct sockaddr
*sap
;
189 sap
= (struct sockaddr
*)&sin
;
192 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
194 sap
= (struct sockaddr
*)&sin6
;
197 #endif /* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
199 return ERR_PTR(-EAFNOSUPPORT
);
202 return xcl
->xcl_ops
->xpo_create(serv
, net
, sap
, len
, flags
);
205 int svc_create_xprt(struct svc_serv
*serv
, const char *xprt_name
,
206 struct net
*net
, const int family
,
207 const unsigned short port
, int flags
)
209 struct svc_xprt_class
*xcl
;
211 dprintk("svc: creating transport %s[%d]\n", xprt_name
, port
);
212 spin_lock(&svc_xprt_class_lock
);
213 list_for_each_entry(xcl
, &svc_xprt_class_list
, xcl_list
) {
214 struct svc_xprt
*newxprt
;
215 unsigned short newport
;
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 newport
= svc_xprt_local_port(newxprt
);
235 clear_bit(XPT_BUSY
, &newxprt
->xpt_flags
);
239 spin_unlock(&svc_xprt_class_lock
);
240 dprintk("svc: transport %s not found\n", xprt_name
);
242 /* This errno is exposed to user space. Provide a reasonable
243 * perror msg for a bad transport. */
244 return -EPROTONOSUPPORT
;
246 EXPORT_SYMBOL_GPL(svc_create_xprt
);
249 * Copy the local and remote xprt addresses to the rqstp structure
251 void svc_xprt_copy_addrs(struct svc_rqst
*rqstp
, struct svc_xprt
*xprt
)
253 struct sockaddr
*sin
;
255 memcpy(&rqstp
->rq_addr
, &xprt
->xpt_remote
, xprt
->xpt_remotelen
);
256 rqstp
->rq_addrlen
= xprt
->xpt_remotelen
;
259 * Destination address in request is needed for binding the
260 * source address in RPC replies/callbacks later.
262 sin
= (struct sockaddr
*)&xprt
->xpt_local
;
263 switch (sin
->sa_family
) {
265 rqstp
->rq_daddr
.addr
= ((struct sockaddr_in
*)sin
)->sin_addr
;
268 rqstp
->rq_daddr
.addr6
= ((struct sockaddr_in6
*)sin
)->sin6_addr
;
272 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs
);
275 * svc_print_addr - Format rq_addr field for printing
276 * @rqstp: svc_rqst struct containing address to print
277 * @buf: target buffer for formatted address
278 * @len: length of target buffer
281 char *svc_print_addr(struct svc_rqst
*rqstp
, char *buf
, size_t len
)
283 return __svc_print_addr(svc_addr(rqstp
), buf
, len
);
285 EXPORT_SYMBOL_GPL(svc_print_addr
);
288 * Queue up an idle server thread. Must have pool->sp_lock held.
289 * Note: this is really a stack rather than a queue, so that we only
290 * use as many different threads as we need, and the rest don't pollute
293 static void svc_thread_enqueue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
295 list_add(&rqstp
->rq_list
, &pool
->sp_threads
);
299 * Dequeue an nfsd thread. Must have pool->sp_lock held.
301 static void svc_thread_dequeue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
303 list_del(&rqstp
->rq_list
);
307 * Queue up a transport with data pending. If there are idle nfsd
308 * processes, wake 'em up.
311 void svc_xprt_enqueue(struct svc_xprt
*xprt
)
313 struct svc_serv
*serv
= xprt
->xpt_server
;
314 struct svc_pool
*pool
;
315 struct svc_rqst
*rqstp
;
318 if (!(xprt
->xpt_flags
&
319 ((1<<XPT_CONN
)|(1<<XPT_DATA
)|(1<<XPT_CLOSE
)|(1<<XPT_DEFERRED
))))
323 pool
= svc_pool_for_cpu(xprt
->xpt_server
, cpu
);
326 spin_lock_bh(&pool
->sp_lock
);
328 if (!list_empty(&pool
->sp_threads
) &&
329 !list_empty(&pool
->sp_sockets
))
332 "threads and transports both waiting??\n");
334 pool
->sp_stats
.packets
++;
336 /* Mark transport as busy. It will remain in this state until
337 * the provider calls svc_xprt_received. We update XPT_BUSY
338 * atomically because it also guards against trying to enqueue
339 * the transport twice.
341 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
342 /* Don't enqueue transport while already enqueued */
343 dprintk("svc: transport %p busy, not enqueued\n", xprt
);
346 BUG_ON(xprt
->xpt_pool
!= NULL
);
347 xprt
->xpt_pool
= pool
;
349 /* Handle pending connection */
350 if (test_bit(XPT_CONN
, &xprt
->xpt_flags
))
353 /* Handle close in-progress */
354 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
))
357 /* Check if we have space to reply to a request */
358 if (!xprt
->xpt_ops
->xpo_has_wspace(xprt
)) {
359 /* Don't enqueue while not enough space for reply */
360 dprintk("svc: no write space, transport %p not enqueued\n",
362 xprt
->xpt_pool
= NULL
;
363 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
368 if (!list_empty(&pool
->sp_threads
)) {
369 rqstp
= list_entry(pool
->sp_threads
.next
,
372 dprintk("svc: transport %p served by daemon %p\n",
374 svc_thread_dequeue(pool
, rqstp
);
377 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
378 rqstp
, rqstp
->rq_xprt
);
379 rqstp
->rq_xprt
= xprt
;
381 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
382 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
383 pool
->sp_stats
.threads_woken
++;
384 BUG_ON(xprt
->xpt_pool
!= pool
);
385 wake_up(&rqstp
->rq_wait
);
387 dprintk("svc: transport %p put into queue\n", xprt
);
388 list_add_tail(&xprt
->xpt_ready
, &pool
->sp_sockets
);
389 pool
->sp_stats
.sockets_queued
++;
390 BUG_ON(xprt
->xpt_pool
!= pool
);
394 spin_unlock_bh(&pool
->sp_lock
);
396 EXPORT_SYMBOL_GPL(svc_xprt_enqueue
);
399 * Dequeue the first transport. Must be called with the pool->sp_lock held.
401 static struct svc_xprt
*svc_xprt_dequeue(struct svc_pool
*pool
)
403 struct svc_xprt
*xprt
;
405 if (list_empty(&pool
->sp_sockets
))
408 xprt
= list_entry(pool
->sp_sockets
.next
,
409 struct svc_xprt
, xpt_ready
);
410 list_del_init(&xprt
->xpt_ready
);
412 dprintk("svc: transport %p dequeued, inuse=%d\n",
413 xprt
, atomic_read(&xprt
->xpt_ref
.refcount
));
419 * svc_xprt_received conditionally queues the transport for processing
420 * by another thread. The caller must hold the XPT_BUSY bit and must
421 * not thereafter touch transport data.
423 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
424 * insufficient) data.
426 void svc_xprt_received(struct svc_xprt
*xprt
)
428 BUG_ON(!test_bit(XPT_BUSY
, &xprt
->xpt_flags
));
429 xprt
->xpt_pool
= NULL
;
430 /* As soon as we clear busy, the xprt could be closed and
431 * 'put', so we need a reference to call svc_xprt_enqueue with:
434 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
435 svc_xprt_enqueue(xprt
);
438 EXPORT_SYMBOL_GPL(svc_xprt_received
);
441 * svc_reserve - change the space reserved for the reply to a request.
442 * @rqstp: The request in question
443 * @space: new max space to reserve
445 * Each request reserves some space on the output queue of the transport
446 * to make sure the reply fits. This function reduces that reserved
447 * space to be the amount of space used already, plus @space.
450 void svc_reserve(struct svc_rqst
*rqstp
, int space
)
452 space
+= rqstp
->rq_res
.head
[0].iov_len
;
454 if (space
< rqstp
->rq_reserved
) {
455 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
456 atomic_sub((rqstp
->rq_reserved
- space
), &xprt
->xpt_reserved
);
457 rqstp
->rq_reserved
= space
;
459 svc_xprt_enqueue(xprt
);
462 EXPORT_SYMBOL_GPL(svc_reserve
);
464 static void svc_xprt_release(struct svc_rqst
*rqstp
)
466 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
468 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
470 kfree(rqstp
->rq_deferred
);
471 rqstp
->rq_deferred
= NULL
;
473 svc_free_res_pages(rqstp
);
474 rqstp
->rq_res
.page_len
= 0;
475 rqstp
->rq_res
.page_base
= 0;
477 /* Reset response buffer and release
479 * But first, check that enough space was reserved
480 * for the reply, otherwise we have a bug!
482 if ((rqstp
->rq_res
.len
) > rqstp
->rq_reserved
)
483 printk(KERN_ERR
"RPC request reserved %d but used %d\n",
487 rqstp
->rq_res
.head
[0].iov_len
= 0;
488 svc_reserve(rqstp
, 0);
489 rqstp
->rq_xprt
= NULL
;
495 * External function to wake up a server waiting for data
496 * This really only makes sense for services like lockd
497 * which have exactly one thread anyway.
499 void svc_wake_up(struct svc_serv
*serv
)
501 struct svc_rqst
*rqstp
;
503 struct svc_pool
*pool
;
505 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
506 pool
= &serv
->sv_pools
[i
];
508 spin_lock_bh(&pool
->sp_lock
);
509 if (!list_empty(&pool
->sp_threads
)) {
510 rqstp
= list_entry(pool
->sp_threads
.next
,
513 dprintk("svc: daemon %p woken up.\n", rqstp
);
515 svc_thread_dequeue(pool, rqstp);
516 rqstp->rq_xprt = NULL;
518 wake_up(&rqstp
->rq_wait
);
520 spin_unlock_bh(&pool
->sp_lock
);
523 EXPORT_SYMBOL_GPL(svc_wake_up
);
525 int svc_port_is_privileged(struct sockaddr
*sin
)
527 switch (sin
->sa_family
) {
529 return ntohs(((struct sockaddr_in
*)sin
)->sin_port
)
532 return ntohs(((struct sockaddr_in6
*)sin
)->sin6_port
)
540 * Make sure that we don't have too many active connections. If we have,
541 * something must be dropped. It's not clear what will happen if we allow
542 * "too many" connections, but when dealing with network-facing software,
543 * we have to code defensively. Here we do that by imposing hard limits.
545 * There's no point in trying to do random drop here for DoS
546 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
547 * attacker can easily beat that.
549 * The only somewhat efficient mechanism would be if drop old
550 * connections from the same IP first. But right now we don't even
551 * record the client IP in svc_sock.
553 * single-threaded services that expect a lot of clients will probably
554 * need to set sv_maxconn to override the default value which is based
555 * on the number of threads
557 static void svc_check_conn_limits(struct svc_serv
*serv
)
559 unsigned int limit
= serv
->sv_maxconn
? serv
->sv_maxconn
:
560 (serv
->sv_nrthreads
+3) * 20;
562 if (serv
->sv_tmpcnt
> limit
) {
563 struct svc_xprt
*xprt
= NULL
;
564 spin_lock_bh(&serv
->sv_lock
);
565 if (!list_empty(&serv
->sv_tempsocks
)) {
566 if (net_ratelimit()) {
567 /* Try to help the admin */
568 printk(KERN_NOTICE
"%s: too many open "
569 "connections, consider increasing %s\n",
570 serv
->sv_name
, serv
->sv_maxconn
?
571 "the max number of connections." :
572 "the number of threads.");
575 * Always select the oldest connection. It's not fair,
578 xprt
= list_entry(serv
->sv_tempsocks
.prev
,
581 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
584 spin_unlock_bh(&serv
->sv_lock
);
587 svc_xprt_enqueue(xprt
);
594 * Receive the next request on any transport. This code is carefully
595 * organised not to touch any cachelines in the shared svc_serv
596 * structure, only cachelines in the local svc_pool.
598 int svc_recv(struct svc_rqst
*rqstp
, long timeout
)
600 struct svc_xprt
*xprt
= NULL
;
601 struct svc_serv
*serv
= rqstp
->rq_server
;
602 struct svc_pool
*pool
= rqstp
->rq_pool
;
606 DECLARE_WAITQUEUE(wait
, current
);
609 dprintk("svc: server %p waiting for data (to = %ld)\n",
614 "svc_recv: service %p, transport not NULL!\n",
616 if (waitqueue_active(&rqstp
->rq_wait
))
618 "svc_recv: service %p, wait queue active!\n",
621 /* now allocate needed pages. If we get a failure, sleep briefly */
622 pages
= (serv
->sv_max_mesg
+ PAGE_SIZE
) / PAGE_SIZE
;
623 for (i
= 0; i
< pages
; i
++)
624 while (rqstp
->rq_pages
[i
] == NULL
) {
625 struct page
*p
= alloc_page(GFP_KERNEL
);
627 set_current_state(TASK_INTERRUPTIBLE
);
628 if (signalled() || kthread_should_stop()) {
629 set_current_state(TASK_RUNNING
);
632 schedule_timeout(msecs_to_jiffies(500));
634 rqstp
->rq_pages
[i
] = p
;
636 rqstp
->rq_pages
[i
++] = NULL
; /* this might be seen in nfs_read_actor */
637 BUG_ON(pages
>= RPCSVC_MAXPAGES
);
639 /* Make arg->head point to first page and arg->pages point to rest */
640 arg
= &rqstp
->rq_arg
;
641 arg
->head
[0].iov_base
= page_address(rqstp
->rq_pages
[0]);
642 arg
->head
[0].iov_len
= PAGE_SIZE
;
643 arg
->pages
= rqstp
->rq_pages
+ 1;
645 /* save at least one page for response */
646 arg
->page_len
= (pages
-2)*PAGE_SIZE
;
647 arg
->len
= (pages
-1)*PAGE_SIZE
;
648 arg
->tail
[0].iov_len
= 0;
652 if (signalled() || kthread_should_stop())
655 /* Normally we will wait up to 5 seconds for any required
656 * cache information to be provided.
658 rqstp
->rq_chandle
.thread_wait
= 5*HZ
;
660 spin_lock_bh(&pool
->sp_lock
);
661 xprt
= svc_xprt_dequeue(pool
);
663 rqstp
->rq_xprt
= xprt
;
665 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
666 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
668 /* As there is a shortage of threads and this request
669 * had to be queued, don't allow the thread to wait so
670 * long for cache updates.
672 rqstp
->rq_chandle
.thread_wait
= 1*HZ
;
674 /* No data pending. Go to sleep */
675 svc_thread_enqueue(pool
, rqstp
);
678 * We have to be able to interrupt this wait
679 * to bring down the daemons ...
681 set_current_state(TASK_INTERRUPTIBLE
);
684 * checking kthread_should_stop() here allows us to avoid
685 * locking and signalling when stopping kthreads that call
686 * svc_recv. If the thread has already been woken up, then
687 * we can exit here without sleeping. If not, then it
688 * it'll be woken up quickly during the schedule_timeout
690 if (kthread_should_stop()) {
691 set_current_state(TASK_RUNNING
);
692 spin_unlock_bh(&pool
->sp_lock
);
696 add_wait_queue(&rqstp
->rq_wait
, &wait
);
697 spin_unlock_bh(&pool
->sp_lock
);
699 time_left
= schedule_timeout(timeout
);
703 spin_lock_bh(&pool
->sp_lock
);
704 remove_wait_queue(&rqstp
->rq_wait
, &wait
);
706 pool
->sp_stats
.threads_timedout
++;
708 xprt
= rqstp
->rq_xprt
;
710 svc_thread_dequeue(pool
, rqstp
);
711 spin_unlock_bh(&pool
->sp_lock
);
712 dprintk("svc: server %p, no data yet\n", rqstp
);
713 if (signalled() || kthread_should_stop())
719 spin_unlock_bh(&pool
->sp_lock
);
722 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
)) {
723 dprintk("svc_recv: found XPT_CLOSE\n");
724 svc_delete_xprt(xprt
);
725 } else if (test_bit(XPT_LISTENER
, &xprt
->xpt_flags
)) {
726 struct svc_xprt
*newxpt
;
727 newxpt
= xprt
->xpt_ops
->xpo_accept(xprt
);
730 * We know this module_get will succeed because the
731 * listener holds a reference too
733 __module_get(newxpt
->xpt_class
->xcl_owner
);
734 svc_check_conn_limits(xprt
->xpt_server
);
735 spin_lock_bh(&serv
->sv_lock
);
736 set_bit(XPT_TEMP
, &newxpt
->xpt_flags
);
737 list_add(&newxpt
->xpt_list
, &serv
->sv_tempsocks
);
739 if (serv
->sv_temptimer
.function
== NULL
) {
740 /* setup timer to age temp transports */
741 setup_timer(&serv
->sv_temptimer
,
743 (unsigned long)serv
);
744 mod_timer(&serv
->sv_temptimer
,
745 jiffies
+ svc_conn_age_period
* HZ
);
747 spin_unlock_bh(&serv
->sv_lock
);
748 svc_xprt_received(newxpt
);
750 svc_xprt_received(xprt
);
752 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
753 rqstp
, pool
->sp_id
, xprt
,
754 atomic_read(&xprt
->xpt_ref
.refcount
));
755 rqstp
->rq_deferred
= svc_deferred_dequeue(xprt
);
756 if (rqstp
->rq_deferred
) {
757 svc_xprt_received(xprt
);
758 len
= svc_deferred_recv(rqstp
);
760 len
= xprt
->xpt_ops
->xpo_recvfrom(rqstp
);
761 svc_xprt_received(xprt
);
763 dprintk("svc: got len=%d\n", len
);
766 /* No data, incomplete (TCP) read, or accept() */
767 if (len
== 0 || len
== -EAGAIN
) {
768 rqstp
->rq_res
.len
= 0;
769 svc_xprt_release(rqstp
);
772 clear_bit(XPT_OLD
, &xprt
->xpt_flags
);
774 rqstp
->rq_secure
= svc_port_is_privileged(svc_addr(rqstp
));
775 rqstp
->rq_chandle
.defer
= svc_defer
;
778 serv
->sv_stats
->netcnt
++;
781 EXPORT_SYMBOL_GPL(svc_recv
);
786 void svc_drop(struct svc_rqst
*rqstp
)
788 dprintk("svc: xprt %p dropped request\n", rqstp
->rq_xprt
);
789 svc_xprt_release(rqstp
);
791 EXPORT_SYMBOL_GPL(svc_drop
);
794 * Return reply to client.
796 int svc_send(struct svc_rqst
*rqstp
)
798 struct svc_xprt
*xprt
;
802 xprt
= rqstp
->rq_xprt
;
806 /* release the receive skb before sending the reply */
807 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
809 /* calculate over-all length */
811 xb
->len
= xb
->head
[0].iov_len
+
815 /* Grab mutex to serialize outgoing data. */
816 mutex_lock(&xprt
->xpt_mutex
);
817 if (test_bit(XPT_DEAD
, &xprt
->xpt_flags
))
820 len
= xprt
->xpt_ops
->xpo_sendto(rqstp
);
821 mutex_unlock(&xprt
->xpt_mutex
);
822 rpc_wake_up(&xprt
->xpt_bc_pending
);
823 svc_xprt_release(rqstp
);
825 if (len
== -ECONNREFUSED
|| len
== -ENOTCONN
|| len
== -EAGAIN
)
831 * Timer function to close old temporary transports, using
832 * a mark-and-sweep algorithm.
834 static void svc_age_temp_xprts(unsigned long closure
)
836 struct svc_serv
*serv
= (struct svc_serv
*)closure
;
837 struct svc_xprt
*xprt
;
838 struct list_head
*le
, *next
;
839 LIST_HEAD(to_be_aged
);
841 dprintk("svc_age_temp_xprts\n");
843 if (!spin_trylock_bh(&serv
->sv_lock
)) {
844 /* busy, try again 1 sec later */
845 dprintk("svc_age_temp_xprts: busy\n");
846 mod_timer(&serv
->sv_temptimer
, jiffies
+ HZ
);
850 list_for_each_safe(le
, next
, &serv
->sv_tempsocks
) {
851 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
853 /* First time through, just mark it OLD. Second time
854 * through, close it. */
855 if (!test_and_set_bit(XPT_OLD
, &xprt
->xpt_flags
))
857 if (atomic_read(&xprt
->xpt_ref
.refcount
) > 1 ||
858 test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
861 list_move(le
, &to_be_aged
);
862 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
863 set_bit(XPT_DETACHED
, &xprt
->xpt_flags
);
865 spin_unlock_bh(&serv
->sv_lock
);
867 while (!list_empty(&to_be_aged
)) {
868 le
= to_be_aged
.next
;
869 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
871 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
873 dprintk("queuing xprt %p for closing\n", xprt
);
875 /* a thread will dequeue and close it soon */
876 svc_xprt_enqueue(xprt
);
880 mod_timer(&serv
->sv_temptimer
, jiffies
+ svc_conn_age_period
* HZ
);
883 static void call_xpt_users(struct svc_xprt
*xprt
)
885 struct svc_xpt_user
*u
;
887 spin_lock(&xprt
->xpt_lock
);
888 while (!list_empty(&xprt
->xpt_users
)) {
889 u
= list_first_entry(&xprt
->xpt_users
, struct svc_xpt_user
, list
);
893 spin_unlock(&xprt
->xpt_lock
);
897 * Remove a dead transport
899 void svc_delete_xprt(struct svc_xprt
*xprt
)
901 struct svc_serv
*serv
= xprt
->xpt_server
;
902 struct svc_deferred_req
*dr
;
904 /* Only do this once */
905 if (test_and_set_bit(XPT_DEAD
, &xprt
->xpt_flags
))
908 dprintk("svc: svc_delete_xprt(%p)\n", xprt
);
909 xprt
->xpt_ops
->xpo_detach(xprt
);
911 spin_lock_bh(&serv
->sv_lock
);
912 if (!test_and_set_bit(XPT_DETACHED
, &xprt
->xpt_flags
))
913 list_del_init(&xprt
->xpt_list
);
915 * We used to delete the transport from whichever list
916 * it's sk_xprt.xpt_ready node was on, but we don't actually
917 * need to. This is because the only time we're called
918 * while still attached to a queue, the queue itself
919 * is about to be destroyed (in svc_destroy).
921 if (test_bit(XPT_TEMP
, &xprt
->xpt_flags
))
923 spin_unlock_bh(&serv
->sv_lock
);
925 while ((dr
= svc_deferred_dequeue(xprt
)) != NULL
)
928 call_xpt_users(xprt
);
932 void svc_close_xprt(struct svc_xprt
*xprt
)
934 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
935 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
))
936 /* someone else will have to effect the close */
939 svc_delete_xprt(xprt
);
941 EXPORT_SYMBOL_GPL(svc_close_xprt
);
943 void svc_close_all(struct list_head
*xprt_list
)
945 struct svc_xprt
*xprt
;
946 struct svc_xprt
*tmp
;
948 list_for_each_entry_safe(xprt
, tmp
, xprt_list
, xpt_list
) {
949 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
950 if (test_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
951 /* Waiting to be processed, but no threads left,
952 * So just remove it from the waiting list
954 list_del_init(&xprt
->xpt_ready
);
955 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
957 svc_close_xprt(xprt
);
962 * Handle defer and revisit of requests
965 static void svc_revisit(struct cache_deferred_req
*dreq
, int too_many
)
967 struct svc_deferred_req
*dr
=
968 container_of(dreq
, struct svc_deferred_req
, handle
);
969 struct svc_xprt
*xprt
= dr
->xprt
;
971 spin_lock(&xprt
->xpt_lock
);
972 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
973 if (too_many
|| test_bit(XPT_DEAD
, &xprt
->xpt_flags
)) {
974 spin_unlock(&xprt
->xpt_lock
);
975 dprintk("revisit canceled\n");
980 dprintk("revisit queued\n");
982 list_add(&dr
->handle
.recent
, &xprt
->xpt_deferred
);
983 spin_unlock(&xprt
->xpt_lock
);
984 svc_xprt_enqueue(xprt
);
989 * Save the request off for later processing. The request buffer looks
992 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
994 * This code can only handle requests that consist of an xprt-header
997 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
)
999 struct svc_rqst
*rqstp
= container_of(req
, struct svc_rqst
, rq_chandle
);
1000 struct svc_deferred_req
*dr
;
1002 if (rqstp
->rq_arg
.page_len
|| !rqstp
->rq_usedeferral
)
1003 return NULL
; /* if more than a page, give up FIXME */
1004 if (rqstp
->rq_deferred
) {
1005 dr
= rqstp
->rq_deferred
;
1006 rqstp
->rq_deferred
= NULL
;
1010 /* FIXME maybe discard if size too large */
1011 size
= sizeof(struct svc_deferred_req
) + rqstp
->rq_arg
.len
;
1012 dr
= kmalloc(size
, GFP_KERNEL
);
1016 dr
->handle
.owner
= rqstp
->rq_server
;
1017 dr
->prot
= rqstp
->rq_prot
;
1018 memcpy(&dr
->addr
, &rqstp
->rq_addr
, rqstp
->rq_addrlen
);
1019 dr
->addrlen
= rqstp
->rq_addrlen
;
1020 dr
->daddr
= rqstp
->rq_daddr
;
1021 dr
->argslen
= rqstp
->rq_arg
.len
>> 2;
1022 dr
->xprt_hlen
= rqstp
->rq_xprt_hlen
;
1024 /* back up head to the start of the buffer and copy */
1025 skip
= rqstp
->rq_arg
.len
- rqstp
->rq_arg
.head
[0].iov_len
;
1026 memcpy(dr
->args
, rqstp
->rq_arg
.head
[0].iov_base
- skip
,
1029 svc_xprt_get(rqstp
->rq_xprt
);
1030 dr
->xprt
= rqstp
->rq_xprt
;
1032 dr
->handle
.revisit
= svc_revisit
;
1037 * recv data from a deferred request into an active one
1039 static int svc_deferred_recv(struct svc_rqst
*rqstp
)
1041 struct svc_deferred_req
*dr
= rqstp
->rq_deferred
;
1043 /* setup iov_base past transport header */
1044 rqstp
->rq_arg
.head
[0].iov_base
= dr
->args
+ (dr
->xprt_hlen
>>2);
1045 /* The iov_len does not include the transport header bytes */
1046 rqstp
->rq_arg
.head
[0].iov_len
= (dr
->argslen
<<2) - dr
->xprt_hlen
;
1047 rqstp
->rq_arg
.page_len
= 0;
1048 /* The rq_arg.len includes the transport header bytes */
1049 rqstp
->rq_arg
.len
= dr
->argslen
<<2;
1050 rqstp
->rq_prot
= dr
->prot
;
1051 memcpy(&rqstp
->rq_addr
, &dr
->addr
, dr
->addrlen
);
1052 rqstp
->rq_addrlen
= dr
->addrlen
;
1053 /* Save off transport header len in case we get deferred again */
1054 rqstp
->rq_xprt_hlen
= dr
->xprt_hlen
;
1055 rqstp
->rq_daddr
= dr
->daddr
;
1056 rqstp
->rq_respages
= rqstp
->rq_pages
;
1057 return (dr
->argslen
<<2) - dr
->xprt_hlen
;
1061 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
)
1063 struct svc_deferred_req
*dr
= NULL
;
1065 if (!test_bit(XPT_DEFERRED
, &xprt
->xpt_flags
))
1067 spin_lock(&xprt
->xpt_lock
);
1068 clear_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1069 if (!list_empty(&xprt
->xpt_deferred
)) {
1070 dr
= list_entry(xprt
->xpt_deferred
.next
,
1071 struct svc_deferred_req
,
1073 list_del_init(&dr
->handle
.recent
);
1074 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1076 spin_unlock(&xprt
->xpt_lock
);
1081 * svc_find_xprt - find an RPC transport instance
1082 * @serv: pointer to svc_serv to search
1083 * @xcl_name: C string containing transport's class name
1084 * @af: Address family of transport's local address
1085 * @port: transport's IP port number
1087 * Return the transport instance pointer for the endpoint accepting
1088 * connections/peer traffic from the specified transport class,
1089 * address family and port.
1091 * Specifying 0 for the address family or port is effectively a
1092 * wild-card, and will result in matching the first transport in the
1093 * service's list that has a matching class name.
1095 struct svc_xprt
*svc_find_xprt(struct svc_serv
*serv
, const char *xcl_name
,
1096 const sa_family_t af
, const unsigned short port
)
1098 struct svc_xprt
*xprt
;
1099 struct svc_xprt
*found
= NULL
;
1101 /* Sanity check the args */
1102 if (serv
== NULL
|| xcl_name
== NULL
)
1105 spin_lock_bh(&serv
->sv_lock
);
1106 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1107 if (strcmp(xprt
->xpt_class
->xcl_name
, xcl_name
))
1109 if (af
!= AF_UNSPEC
&& af
!= xprt
->xpt_local
.ss_family
)
1111 if (port
!= 0 && port
!= svc_xprt_local_port(xprt
))
1117 spin_unlock_bh(&serv
->sv_lock
);
1120 EXPORT_SYMBOL_GPL(svc_find_xprt
);
1122 static int svc_one_xprt_name(const struct svc_xprt
*xprt
,
1123 char *pos
, int remaining
)
1127 len
= snprintf(pos
, remaining
, "%s %u\n",
1128 xprt
->xpt_class
->xcl_name
,
1129 svc_xprt_local_port(xprt
));
1130 if (len
>= remaining
)
1131 return -ENAMETOOLONG
;
1136 * svc_xprt_names - format a buffer with a list of transport names
1137 * @serv: pointer to an RPC service
1138 * @buf: pointer to a buffer to be filled in
1139 * @buflen: length of buffer to be filled in
1141 * Fills in @buf with a string containing a list of transport names,
1142 * each name terminated with '\n'.
1144 * Returns positive length of the filled-in string on success; otherwise
1145 * a negative errno value is returned if an error occurs.
1147 int svc_xprt_names(struct svc_serv
*serv
, char *buf
, const int buflen
)
1149 struct svc_xprt
*xprt
;
1153 /* Sanity check args */
1157 spin_lock_bh(&serv
->sv_lock
);
1161 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1162 len
= svc_one_xprt_name(xprt
, pos
, buflen
- totlen
);
1174 spin_unlock_bh(&serv
->sv_lock
);
1177 EXPORT_SYMBOL_GPL(svc_xprt_names
);
1180 /*----------------------------------------------------------------------------*/
1182 static void *svc_pool_stats_start(struct seq_file
*m
, loff_t
*pos
)
1184 unsigned int pidx
= (unsigned int)*pos
;
1185 struct svc_serv
*serv
= m
->private;
1187 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx
);
1190 return SEQ_START_TOKEN
;
1191 return (pidx
> serv
->sv_nrpools
? NULL
: &serv
->sv_pools
[pidx
-1]);
1194 static void *svc_pool_stats_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
1196 struct svc_pool
*pool
= p
;
1197 struct svc_serv
*serv
= m
->private;
1199 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos
);
1201 if (p
== SEQ_START_TOKEN
) {
1202 pool
= &serv
->sv_pools
[0];
1204 unsigned int pidx
= (pool
- &serv
->sv_pools
[0]);
1205 if (pidx
< serv
->sv_nrpools
-1)
1206 pool
= &serv
->sv_pools
[pidx
+1];
1214 static void svc_pool_stats_stop(struct seq_file
*m
, void *p
)
1218 static int svc_pool_stats_show(struct seq_file
*m
, void *p
)
1220 struct svc_pool
*pool
= p
;
1222 if (p
== SEQ_START_TOKEN
) {
1223 seq_puts(m
, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1227 seq_printf(m
, "%u %lu %lu %lu %lu\n",
1229 pool
->sp_stats
.packets
,
1230 pool
->sp_stats
.sockets_queued
,
1231 pool
->sp_stats
.threads_woken
,
1232 pool
->sp_stats
.threads_timedout
);
1237 static const struct seq_operations svc_pool_stats_seq_ops
= {
1238 .start
= svc_pool_stats_start
,
1239 .next
= svc_pool_stats_next
,
1240 .stop
= svc_pool_stats_stop
,
1241 .show
= svc_pool_stats_show
,
1244 int svc_pool_stats_open(struct svc_serv
*serv
, struct file
*file
)
1248 err
= seq_open(file
, &svc_pool_stats_seq_ops
);
1250 ((struct seq_file
*) file
->private_data
)->private = serv
;
1253 EXPORT_SYMBOL(svc_pool_stats_open
);
1255 /*----------------------------------------------------------------------------*/