2 * linux/net/sunrpc/svc_xprt.c
4 * Author: Tom Tucker <tom@opengridcomputing.com>
7 #include <linux/sched.h>
8 #include <linux/smp_lock.h>
9 #include <linux/errno.h>
10 #include <linux/freezer.h>
11 #include <linux/kthread.h>
13 #include <linux/sunrpc/stats.h>
14 #include <linux/sunrpc/svc_xprt.h>
15 #include <linux/sunrpc/svcsock.h>
17 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
19 #define SVC_MAX_WAKING 5
21 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
);
22 static int svc_deferred_recv(struct svc_rqst
*rqstp
);
23 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
);
24 static void svc_age_temp_xprts(unsigned long closure
);
26 /* apparently the "standard" is that clients close
27 * idle connections after 5 minutes, servers after
29 * http://www.connectathon.org/talks96/nfstcp.pdf
31 static int svc_conn_age_period
= 6*60;
33 /* List of registered transport classes */
34 static DEFINE_SPINLOCK(svc_xprt_class_lock
);
35 static LIST_HEAD(svc_xprt_class_list
);
37 /* SMP locking strategy:
39 * svc_pool->sp_lock protects most of the fields of that pool.
40 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
41 * when both need to be taken (rare), svc_serv->sv_lock is first.
42 * BKL protects svc_serv->sv_nrthread.
43 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
44 * and the ->sk_info_authunix cache.
46 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
47 * enqueued multiply. During normal transport processing this bit
48 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
49 * Providers should not manipulate this bit directly.
51 * Some flags can be set to certain values at any time
52 * providing that certain rules are followed:
55 * - Can be set or cleared at any time.
56 * - After a set, svc_xprt_enqueue must be called to enqueue
57 * the transport for processing.
58 * - After a clear, the transport must be read/accepted.
59 * If this succeeds, it must be set again.
61 * - Can set at any time. It is never cleared.
63 * - Can only be set while XPT_BUSY is held which ensures
64 * that no other thread will be using the transport or will
65 * try to set XPT_DEAD.
68 int svc_reg_xprt_class(struct svc_xprt_class
*xcl
)
70 struct svc_xprt_class
*cl
;
73 dprintk("svc: Adding svc transport class '%s'\n", xcl
->xcl_name
);
75 INIT_LIST_HEAD(&xcl
->xcl_list
);
76 spin_lock(&svc_xprt_class_lock
);
77 /* Make sure there isn't already a class with the same name */
78 list_for_each_entry(cl
, &svc_xprt_class_list
, xcl_list
) {
79 if (strcmp(xcl
->xcl_name
, cl
->xcl_name
) == 0)
82 list_add_tail(&xcl
->xcl_list
, &svc_xprt_class_list
);
85 spin_unlock(&svc_xprt_class_lock
);
88 EXPORT_SYMBOL_GPL(svc_reg_xprt_class
);
90 void svc_unreg_xprt_class(struct svc_xprt_class
*xcl
)
92 dprintk("svc: Removing svc transport class '%s'\n", xcl
->xcl_name
);
93 spin_lock(&svc_xprt_class_lock
);
94 list_del_init(&xcl
->xcl_list
);
95 spin_unlock(&svc_xprt_class_lock
);
97 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class
);
100 * Format the transport list for printing
102 int svc_print_xprts(char *buf
, int maxlen
)
104 struct list_head
*le
;
109 spin_lock(&svc_xprt_class_lock
);
110 list_for_each(le
, &svc_xprt_class_list
) {
112 struct svc_xprt_class
*xcl
=
113 list_entry(le
, struct svc_xprt_class
, xcl_list
);
115 sprintf(tmpstr
, "%s %d\n", xcl
->xcl_name
, xcl
->xcl_max_payload
);
116 slen
= strlen(tmpstr
);
117 if (len
+ slen
> maxlen
)
122 spin_unlock(&svc_xprt_class_lock
);
127 static void svc_xprt_free(struct kref
*kref
)
129 struct svc_xprt
*xprt
=
130 container_of(kref
, struct svc_xprt
, xpt_ref
);
131 struct module
*owner
= xprt
->xpt_class
->xcl_owner
;
132 if (test_bit(XPT_CACHE_AUTH
, &xprt
->xpt_flags
)
133 && xprt
->xpt_auth_cache
!= NULL
)
134 svcauth_unix_info_release(xprt
->xpt_auth_cache
);
135 xprt
->xpt_ops
->xpo_free(xprt
);
139 void svc_xprt_put(struct svc_xprt
*xprt
)
141 kref_put(&xprt
->xpt_ref
, svc_xprt_free
);
143 EXPORT_SYMBOL_GPL(svc_xprt_put
);
146 * Called by transport drivers to initialize the transport independent
147 * portion of the transport instance.
149 void svc_xprt_init(struct svc_xprt_class
*xcl
, struct svc_xprt
*xprt
,
150 struct svc_serv
*serv
)
152 memset(xprt
, 0, sizeof(*xprt
));
153 xprt
->xpt_class
= xcl
;
154 xprt
->xpt_ops
= xcl
->xcl_ops
;
155 kref_init(&xprt
->xpt_ref
);
156 xprt
->xpt_server
= serv
;
157 INIT_LIST_HEAD(&xprt
->xpt_list
);
158 INIT_LIST_HEAD(&xprt
->xpt_ready
);
159 INIT_LIST_HEAD(&xprt
->xpt_deferred
);
160 mutex_init(&xprt
->xpt_mutex
);
161 spin_lock_init(&xprt
->xpt_lock
);
162 set_bit(XPT_BUSY
, &xprt
->xpt_flags
);
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
,
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 struct sockaddr_in6 sin6
= {
178 .sin6_family
= AF_INET6
,
179 .sin6_addr
= IN6ADDR_ANY_INIT
,
180 .sin6_port
= htons(port
),
182 struct sockaddr
*sap
;
187 sap
= (struct sockaddr
*)&sin
;
191 sap
= (struct sockaddr
*)&sin6
;
195 return ERR_PTR(-EAFNOSUPPORT
);
198 return xcl
->xcl_ops
->xpo_create(serv
, sap
, len
, flags
);
201 int svc_create_xprt(struct svc_serv
*serv
, const char *xprt_name
,
202 const int family
, const unsigned short port
,
205 struct svc_xprt_class
*xcl
;
207 dprintk("svc: creating transport %s[%d]\n", xprt_name
, port
);
208 spin_lock(&svc_xprt_class_lock
);
209 list_for_each_entry(xcl
, &svc_xprt_class_list
, xcl_list
) {
210 struct svc_xprt
*newxprt
;
212 if (strcmp(xprt_name
, xcl
->xcl_name
))
215 if (!try_module_get(xcl
->xcl_owner
))
218 spin_unlock(&svc_xprt_class_lock
);
219 newxprt
= __svc_xpo_create(xcl
, serv
, family
, port
, flags
);
220 if (IS_ERR(newxprt
)) {
221 module_put(xcl
->xcl_owner
);
222 return PTR_ERR(newxprt
);
225 clear_bit(XPT_TEMP
, &newxprt
->xpt_flags
);
226 spin_lock_bh(&serv
->sv_lock
);
227 list_add(&newxprt
->xpt_list
, &serv
->sv_permsocks
);
228 spin_unlock_bh(&serv
->sv_lock
);
229 clear_bit(XPT_BUSY
, &newxprt
->xpt_flags
);
230 return svc_xprt_local_port(newxprt
);
233 spin_unlock(&svc_xprt_class_lock
);
234 dprintk("svc: transport %s not found\n", xprt_name
);
237 EXPORT_SYMBOL_GPL(svc_create_xprt
);
240 * Copy the local and remote xprt addresses to the rqstp structure
242 void svc_xprt_copy_addrs(struct svc_rqst
*rqstp
, struct svc_xprt
*xprt
)
244 struct sockaddr
*sin
;
246 memcpy(&rqstp
->rq_addr
, &xprt
->xpt_remote
, xprt
->xpt_remotelen
);
247 rqstp
->rq_addrlen
= xprt
->xpt_remotelen
;
250 * Destination address in request is needed for binding the
251 * source address in RPC replies/callbacks later.
253 sin
= (struct sockaddr
*)&xprt
->xpt_local
;
254 switch (sin
->sa_family
) {
256 rqstp
->rq_daddr
.addr
= ((struct sockaddr_in
*)sin
)->sin_addr
;
259 rqstp
->rq_daddr
.addr6
= ((struct sockaddr_in6
*)sin
)->sin6_addr
;
263 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs
);
266 * svc_print_addr - Format rq_addr field for printing
267 * @rqstp: svc_rqst struct containing address to print
268 * @buf: target buffer for formatted address
269 * @len: length of target buffer
272 char *svc_print_addr(struct svc_rqst
*rqstp
, char *buf
, size_t len
)
274 return __svc_print_addr(svc_addr(rqstp
), buf
, len
);
276 EXPORT_SYMBOL_GPL(svc_print_addr
);
279 * Queue up an idle server thread. Must have pool->sp_lock held.
280 * Note: this is really a stack rather than a queue, so that we only
281 * use as many different threads as we need, and the rest don't pollute
284 static void svc_thread_enqueue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
286 list_add(&rqstp
->rq_list
, &pool
->sp_threads
);
290 * Dequeue an nfsd thread. Must have pool->sp_lock held.
292 static void svc_thread_dequeue(struct svc_pool
*pool
, struct svc_rqst
*rqstp
)
294 list_del(&rqstp
->rq_list
);
298 * Queue up a transport with data pending. If there are idle nfsd
299 * processes, wake 'em up.
302 void svc_xprt_enqueue(struct svc_xprt
*xprt
)
304 struct svc_serv
*serv
= xprt
->xpt_server
;
305 struct svc_pool
*pool
;
306 struct svc_rqst
*rqstp
;
310 if (!(xprt
->xpt_flags
&
311 ((1<<XPT_CONN
)|(1<<XPT_DATA
)|(1<<XPT_CLOSE
)|(1<<XPT_DEFERRED
))))
315 pool
= svc_pool_for_cpu(xprt
->xpt_server
, cpu
);
318 spin_lock_bh(&pool
->sp_lock
);
320 if (test_bit(XPT_DEAD
, &xprt
->xpt_flags
)) {
321 /* Don't enqueue dead transports */
322 dprintk("svc: transport %p is dead, not enqueued\n", xprt
);
326 pool
->sp_stats
.packets
++;
328 /* Mark transport as busy. It will remain in this state until
329 * the provider calls svc_xprt_received. We update XPT_BUSY
330 * atomically because it also guards against trying to enqueue
331 * the transport twice.
333 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
334 /* Don't enqueue transport while already enqueued */
335 dprintk("svc: transport %p busy, not enqueued\n", xprt
);
338 BUG_ON(xprt
->xpt_pool
!= NULL
);
339 xprt
->xpt_pool
= pool
;
341 /* Handle pending connection */
342 if (test_bit(XPT_CONN
, &xprt
->xpt_flags
))
345 /* Handle close in-progress */
346 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
))
349 /* Check if we have space to reply to a request */
350 if (!xprt
->xpt_ops
->xpo_has_wspace(xprt
)) {
351 /* Don't enqueue while not enough space for reply */
352 dprintk("svc: no write space, transport %p not enqueued\n",
354 xprt
->xpt_pool
= NULL
;
355 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
360 /* Work out whether threads are available */
361 thread_avail
= !list_empty(&pool
->sp_threads
); /* threads are asleep */
362 if (pool
->sp_nwaking
>= SVC_MAX_WAKING
) {
363 /* too many threads are runnable and trying to wake up */
365 pool
->sp_stats
.overloads_avoided
++;
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 rqstp
->rq_waking
= 1;
385 pool
->sp_stats
.threads_woken
++;
386 BUG_ON(xprt
->xpt_pool
!= pool
);
387 wake_up(&rqstp
->rq_wait
);
389 dprintk("svc: transport %p put into queue\n", xprt
);
390 list_add_tail(&xprt
->xpt_ready
, &pool
->sp_sockets
);
391 pool
->sp_stats
.sockets_queued
++;
392 BUG_ON(xprt
->xpt_pool
!= pool
);
396 spin_unlock_bh(&pool
->sp_lock
);
398 EXPORT_SYMBOL_GPL(svc_xprt_enqueue
);
401 * Dequeue the first transport. Must be called with the pool->sp_lock held.
403 static struct svc_xprt
*svc_xprt_dequeue(struct svc_pool
*pool
)
405 struct svc_xprt
*xprt
;
407 if (list_empty(&pool
->sp_sockets
))
410 xprt
= list_entry(pool
->sp_sockets
.next
,
411 struct svc_xprt
, xpt_ready
);
412 list_del_init(&xprt
->xpt_ready
);
414 dprintk("svc: transport %p dequeued, inuse=%d\n",
415 xprt
, atomic_read(&xprt
->xpt_ref
.refcount
));
421 * svc_xprt_received conditionally queues the transport for processing
422 * by another thread. The caller must hold the XPT_BUSY bit and must
423 * not thereafter touch transport data.
425 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
426 * insufficient) data.
428 void svc_xprt_received(struct svc_xprt
*xprt
)
430 BUG_ON(!test_bit(XPT_BUSY
, &xprt
->xpt_flags
));
431 xprt
->xpt_pool
= NULL
;
432 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
433 svc_xprt_enqueue(xprt
);
435 EXPORT_SYMBOL_GPL(svc_xprt_received
);
438 * svc_reserve - change the space reserved for the reply to a request.
439 * @rqstp: The request in question
440 * @space: new max space to reserve
442 * Each request reserves some space on the output queue of the transport
443 * to make sure the reply fits. This function reduces that reserved
444 * space to be the amount of space used already, plus @space.
447 void svc_reserve(struct svc_rqst
*rqstp
, int space
)
449 space
+= rqstp
->rq_res
.head
[0].iov_len
;
451 if (space
< rqstp
->rq_reserved
) {
452 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
453 atomic_sub((rqstp
->rq_reserved
- space
), &xprt
->xpt_reserved
);
454 rqstp
->rq_reserved
= space
;
456 svc_xprt_enqueue(xprt
);
459 EXPORT_SYMBOL_GPL(svc_reserve
);
461 static void svc_xprt_release(struct svc_rqst
*rqstp
)
463 struct svc_xprt
*xprt
= rqstp
->rq_xprt
;
465 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
467 kfree(rqstp
->rq_deferred
);
468 rqstp
->rq_deferred
= NULL
;
470 svc_free_res_pages(rqstp
);
471 rqstp
->rq_res
.page_len
= 0;
472 rqstp
->rq_res
.page_base
= 0;
474 /* Reset response buffer and release
476 * But first, check that enough space was reserved
477 * for the reply, otherwise we have a bug!
479 if ((rqstp
->rq_res
.len
) > rqstp
->rq_reserved
)
480 printk(KERN_ERR
"RPC request reserved %d but used %d\n",
484 rqstp
->rq_res
.head
[0].iov_len
= 0;
485 svc_reserve(rqstp
, 0);
486 rqstp
->rq_xprt
= NULL
;
492 * External function to wake up a server waiting for data
493 * This really only makes sense for services like lockd
494 * which have exactly one thread anyway.
496 void svc_wake_up(struct svc_serv
*serv
)
498 struct svc_rqst
*rqstp
;
500 struct svc_pool
*pool
;
502 for (i
= 0; i
< serv
->sv_nrpools
; i
++) {
503 pool
= &serv
->sv_pools
[i
];
505 spin_lock_bh(&pool
->sp_lock
);
506 if (!list_empty(&pool
->sp_threads
)) {
507 rqstp
= list_entry(pool
->sp_threads
.next
,
510 dprintk("svc: daemon %p woken up.\n", rqstp
);
512 svc_thread_dequeue(pool, rqstp);
513 rqstp->rq_xprt = NULL;
515 wake_up(&rqstp
->rq_wait
);
517 spin_unlock_bh(&pool
->sp_lock
);
520 EXPORT_SYMBOL_GPL(svc_wake_up
);
522 int svc_port_is_privileged(struct sockaddr
*sin
)
524 switch (sin
->sa_family
) {
526 return ntohs(((struct sockaddr_in
*)sin
)->sin_port
)
529 return ntohs(((struct sockaddr_in6
*)sin
)->sin6_port
)
537 * Make sure that we don't have too many active connections. If we have,
538 * something must be dropped. It's not clear what will happen if we allow
539 * "too many" connections, but when dealing with network-facing software,
540 * we have to code defensively. Here we do that by imposing hard limits.
542 * There's no point in trying to do random drop here for DoS
543 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
544 * attacker can easily beat that.
546 * The only somewhat efficient mechanism would be if drop old
547 * connections from the same IP first. But right now we don't even
548 * record the client IP in svc_sock.
550 * single-threaded services that expect a lot of clients will probably
551 * need to set sv_maxconn to override the default value which is based
552 * on the number of threads
554 static void svc_check_conn_limits(struct svc_serv
*serv
)
556 unsigned int limit
= serv
->sv_maxconn
? serv
->sv_maxconn
:
557 (serv
->sv_nrthreads
+3) * 20;
559 if (serv
->sv_tmpcnt
> limit
) {
560 struct svc_xprt
*xprt
= NULL
;
561 spin_lock_bh(&serv
->sv_lock
);
562 if (!list_empty(&serv
->sv_tempsocks
)) {
563 if (net_ratelimit()) {
564 /* Try to help the admin */
565 printk(KERN_NOTICE
"%s: too many open "
566 "connections, consider increasing %s\n",
567 serv
->sv_name
, serv
->sv_maxconn
?
568 "the max number of connections." :
569 "the number of threads.");
572 * Always select the oldest connection. It's not fair,
575 xprt
= list_entry(serv
->sv_tempsocks
.prev
,
578 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
581 spin_unlock_bh(&serv
->sv_lock
);
584 svc_xprt_enqueue(xprt
);
591 * Receive the next request on any transport. This code is carefully
592 * organised not to touch any cachelines in the shared svc_serv
593 * structure, only cachelines in the local svc_pool.
595 int svc_recv(struct svc_rqst
*rqstp
, long timeout
)
597 struct svc_xprt
*xprt
= NULL
;
598 struct svc_serv
*serv
= rqstp
->rq_server
;
599 struct svc_pool
*pool
= rqstp
->rq_pool
;
603 DECLARE_WAITQUEUE(wait
, current
);
606 dprintk("svc: server %p waiting for data (to = %ld)\n",
611 "svc_recv: service %p, transport not NULL!\n",
613 if (waitqueue_active(&rqstp
->rq_wait
))
615 "svc_recv: service %p, wait queue active!\n",
618 /* now allocate needed pages. If we get a failure, sleep briefly */
619 pages
= (serv
->sv_max_mesg
+ PAGE_SIZE
) / PAGE_SIZE
;
620 for (i
= 0; i
< pages
; i
++)
621 while (rqstp
->rq_pages
[i
] == NULL
) {
622 struct page
*p
= alloc_page(GFP_KERNEL
);
624 set_current_state(TASK_INTERRUPTIBLE
);
625 if (signalled() || kthread_should_stop()) {
626 set_current_state(TASK_RUNNING
);
629 schedule_timeout(msecs_to_jiffies(500));
631 rqstp
->rq_pages
[i
] = p
;
633 rqstp
->rq_pages
[i
++] = NULL
; /* this might be seen in nfs_read_actor */
634 BUG_ON(pages
>= RPCSVC_MAXPAGES
);
636 /* Make arg->head point to first page and arg->pages point to rest */
637 arg
= &rqstp
->rq_arg
;
638 arg
->head
[0].iov_base
= page_address(rqstp
->rq_pages
[0]);
639 arg
->head
[0].iov_len
= PAGE_SIZE
;
640 arg
->pages
= rqstp
->rq_pages
+ 1;
642 /* save at least one page for response */
643 arg
->page_len
= (pages
-2)*PAGE_SIZE
;
644 arg
->len
= (pages
-1)*PAGE_SIZE
;
645 arg
->tail
[0].iov_len
= 0;
649 if (signalled() || kthread_should_stop())
652 spin_lock_bh(&pool
->sp_lock
);
653 if (rqstp
->rq_waking
) {
654 rqstp
->rq_waking
= 0;
656 BUG_ON(pool
->sp_nwaking
< 0);
658 xprt
= svc_xprt_dequeue(pool
);
660 rqstp
->rq_xprt
= xprt
;
662 rqstp
->rq_reserved
= serv
->sv_max_mesg
;
663 atomic_add(rqstp
->rq_reserved
, &xprt
->xpt_reserved
);
665 /* No data pending. Go to sleep */
666 svc_thread_enqueue(pool
, rqstp
);
669 * We have to be able to interrupt this wait
670 * to bring down the daemons ...
672 set_current_state(TASK_INTERRUPTIBLE
);
675 * checking kthread_should_stop() here allows us to avoid
676 * locking and signalling when stopping kthreads that call
677 * svc_recv. If the thread has already been woken up, then
678 * we can exit here without sleeping. If not, then it
679 * it'll be woken up quickly during the schedule_timeout
681 if (kthread_should_stop()) {
682 set_current_state(TASK_RUNNING
);
683 spin_unlock_bh(&pool
->sp_lock
);
687 add_wait_queue(&rqstp
->rq_wait
, &wait
);
688 spin_unlock_bh(&pool
->sp_lock
);
690 time_left
= schedule_timeout(timeout
);
694 spin_lock_bh(&pool
->sp_lock
);
695 remove_wait_queue(&rqstp
->rq_wait
, &wait
);
697 pool
->sp_stats
.threads_timedout
++;
699 xprt
= rqstp
->rq_xprt
;
701 svc_thread_dequeue(pool
, rqstp
);
702 spin_unlock_bh(&pool
->sp_lock
);
703 dprintk("svc: server %p, no data yet\n", rqstp
);
704 if (signalled() || kthread_should_stop())
710 spin_unlock_bh(&pool
->sp_lock
);
713 if (test_bit(XPT_LISTENER
, &xprt
->xpt_flags
)) {
714 struct svc_xprt
*newxpt
;
715 newxpt
= xprt
->xpt_ops
->xpo_accept(xprt
);
718 * We know this module_get will succeed because the
719 * listener holds a reference too
721 __module_get(newxpt
->xpt_class
->xcl_owner
);
722 svc_check_conn_limits(xprt
->xpt_server
);
723 spin_lock_bh(&serv
->sv_lock
);
724 set_bit(XPT_TEMP
, &newxpt
->xpt_flags
);
725 list_add(&newxpt
->xpt_list
, &serv
->sv_tempsocks
);
727 if (serv
->sv_temptimer
.function
== NULL
) {
728 /* setup timer to age temp transports */
729 setup_timer(&serv
->sv_temptimer
,
731 (unsigned long)serv
);
732 mod_timer(&serv
->sv_temptimer
,
733 jiffies
+ svc_conn_age_period
* HZ
);
735 spin_unlock_bh(&serv
->sv_lock
);
736 svc_xprt_received(newxpt
);
738 svc_xprt_received(xprt
);
739 } else if (!test_bit(XPT_CLOSE
, &xprt
->xpt_flags
)) {
740 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
741 rqstp
, pool
->sp_id
, xprt
,
742 atomic_read(&xprt
->xpt_ref
.refcount
));
743 rqstp
->rq_deferred
= svc_deferred_dequeue(xprt
);
744 if (rqstp
->rq_deferred
) {
745 svc_xprt_received(xprt
);
746 len
= svc_deferred_recv(rqstp
);
748 len
= xprt
->xpt_ops
->xpo_recvfrom(rqstp
);
749 dprintk("svc: got len=%d\n", len
);
752 if (test_bit(XPT_CLOSE
, &xprt
->xpt_flags
)) {
753 dprintk("svc_recv: found XPT_CLOSE\n");
754 svc_delete_xprt(xprt
);
757 /* No data, incomplete (TCP) read, or accept() */
758 if (len
== 0 || len
== -EAGAIN
) {
759 rqstp
->rq_res
.len
= 0;
760 svc_xprt_release(rqstp
);
763 clear_bit(XPT_OLD
, &xprt
->xpt_flags
);
765 rqstp
->rq_secure
= svc_port_is_privileged(svc_addr(rqstp
));
766 rqstp
->rq_chandle
.defer
= svc_defer
;
769 serv
->sv_stats
->netcnt
++;
772 EXPORT_SYMBOL_GPL(svc_recv
);
777 void svc_drop(struct svc_rqst
*rqstp
)
779 dprintk("svc: xprt %p dropped request\n", rqstp
->rq_xprt
);
780 svc_xprt_release(rqstp
);
782 EXPORT_SYMBOL_GPL(svc_drop
);
785 * Return reply to client.
787 int svc_send(struct svc_rqst
*rqstp
)
789 struct svc_xprt
*xprt
;
793 xprt
= rqstp
->rq_xprt
;
797 /* release the receive skb before sending the reply */
798 rqstp
->rq_xprt
->xpt_ops
->xpo_release_rqst(rqstp
);
800 /* calculate over-all length */
802 xb
->len
= xb
->head
[0].iov_len
+
806 /* Grab mutex to serialize outgoing data. */
807 mutex_lock(&xprt
->xpt_mutex
);
808 if (test_bit(XPT_DEAD
, &xprt
->xpt_flags
))
811 len
= xprt
->xpt_ops
->xpo_sendto(rqstp
);
812 mutex_unlock(&xprt
->xpt_mutex
);
813 svc_xprt_release(rqstp
);
815 if (len
== -ECONNREFUSED
|| len
== -ENOTCONN
|| len
== -EAGAIN
)
821 * Timer function to close old temporary transports, using
822 * a mark-and-sweep algorithm.
824 static void svc_age_temp_xprts(unsigned long closure
)
826 struct svc_serv
*serv
= (struct svc_serv
*)closure
;
827 struct svc_xprt
*xprt
;
828 struct list_head
*le
, *next
;
829 LIST_HEAD(to_be_aged
);
831 dprintk("svc_age_temp_xprts\n");
833 if (!spin_trylock_bh(&serv
->sv_lock
)) {
834 /* busy, try again 1 sec later */
835 dprintk("svc_age_temp_xprts: busy\n");
836 mod_timer(&serv
->sv_temptimer
, jiffies
+ HZ
);
840 list_for_each_safe(le
, next
, &serv
->sv_tempsocks
) {
841 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
843 /* First time through, just mark it OLD. Second time
844 * through, close it. */
845 if (!test_and_set_bit(XPT_OLD
, &xprt
->xpt_flags
))
847 if (atomic_read(&xprt
->xpt_ref
.refcount
) > 1
848 || test_bit(XPT_BUSY
, &xprt
->xpt_flags
))
851 list_move(le
, &to_be_aged
);
852 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
853 set_bit(XPT_DETACHED
, &xprt
->xpt_flags
);
855 spin_unlock_bh(&serv
->sv_lock
);
857 while (!list_empty(&to_be_aged
)) {
858 le
= to_be_aged
.next
;
859 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
861 xprt
= list_entry(le
, struct svc_xprt
, xpt_list
);
863 dprintk("queuing xprt %p for closing\n", xprt
);
865 /* a thread will dequeue and close it soon */
866 svc_xprt_enqueue(xprt
);
870 mod_timer(&serv
->sv_temptimer
, jiffies
+ svc_conn_age_period
* HZ
);
874 * Remove a dead transport
876 void svc_delete_xprt(struct svc_xprt
*xprt
)
878 struct svc_serv
*serv
= xprt
->xpt_server
;
879 struct svc_deferred_req
*dr
;
881 /* Only do this once */
882 if (test_and_set_bit(XPT_DEAD
, &xprt
->xpt_flags
))
885 dprintk("svc: svc_delete_xprt(%p)\n", xprt
);
886 xprt
->xpt_ops
->xpo_detach(xprt
);
888 spin_lock_bh(&serv
->sv_lock
);
889 if (!test_and_set_bit(XPT_DETACHED
, &xprt
->xpt_flags
))
890 list_del_init(&xprt
->xpt_list
);
892 * We used to delete the transport from whichever list
893 * it's sk_xprt.xpt_ready node was on, but we don't actually
894 * need to. This is because the only time we're called
895 * while still attached to a queue, the queue itself
896 * is about to be destroyed (in svc_destroy).
898 if (test_bit(XPT_TEMP
, &xprt
->xpt_flags
))
901 for (dr
= svc_deferred_dequeue(xprt
); dr
;
902 dr
= svc_deferred_dequeue(xprt
)) {
908 spin_unlock_bh(&serv
->sv_lock
);
911 void svc_close_xprt(struct svc_xprt
*xprt
)
913 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
914 if (test_and_set_bit(XPT_BUSY
, &xprt
->xpt_flags
))
915 /* someone else will have to effect the close */
919 svc_delete_xprt(xprt
);
920 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
923 EXPORT_SYMBOL_GPL(svc_close_xprt
);
925 void svc_close_all(struct list_head
*xprt_list
)
927 struct svc_xprt
*xprt
;
928 struct svc_xprt
*tmp
;
930 list_for_each_entry_safe(xprt
, tmp
, xprt_list
, xpt_list
) {
931 set_bit(XPT_CLOSE
, &xprt
->xpt_flags
);
932 if (test_bit(XPT_BUSY
, &xprt
->xpt_flags
)) {
933 /* Waiting to be processed, but no threads left,
934 * So just remove it from the waiting list
936 list_del_init(&xprt
->xpt_ready
);
937 clear_bit(XPT_BUSY
, &xprt
->xpt_flags
);
939 svc_close_xprt(xprt
);
944 * Handle defer and revisit of requests
947 static void svc_revisit(struct cache_deferred_req
*dreq
, int too_many
)
949 struct svc_deferred_req
*dr
=
950 container_of(dreq
, struct svc_deferred_req
, handle
);
951 struct svc_xprt
*xprt
= dr
->xprt
;
953 spin_lock(&xprt
->xpt_lock
);
954 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
955 if (too_many
|| test_bit(XPT_DEAD
, &xprt
->xpt_flags
)) {
956 spin_unlock(&xprt
->xpt_lock
);
957 dprintk("revisit canceled\n");
962 dprintk("revisit queued\n");
964 list_add(&dr
->handle
.recent
, &xprt
->xpt_deferred
);
965 spin_unlock(&xprt
->xpt_lock
);
966 svc_xprt_enqueue(xprt
);
971 * Save the request off for later processing. The request buffer looks
974 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
976 * This code can only handle requests that consist of an xprt-header
979 static struct cache_deferred_req
*svc_defer(struct cache_req
*req
)
981 struct svc_rqst
*rqstp
= container_of(req
, struct svc_rqst
, rq_chandle
);
982 struct svc_deferred_req
*dr
;
984 if (rqstp
->rq_arg
.page_len
|| !rqstp
->rq_usedeferral
)
985 return NULL
; /* if more than a page, give up FIXME */
986 if (rqstp
->rq_deferred
) {
987 dr
= rqstp
->rq_deferred
;
988 rqstp
->rq_deferred
= NULL
;
992 /* FIXME maybe discard if size too large */
993 size
= sizeof(struct svc_deferred_req
) + rqstp
->rq_arg
.len
;
994 dr
= kmalloc(size
, GFP_KERNEL
);
998 dr
->handle
.owner
= rqstp
->rq_server
;
999 dr
->prot
= rqstp
->rq_prot
;
1000 memcpy(&dr
->addr
, &rqstp
->rq_addr
, rqstp
->rq_addrlen
);
1001 dr
->addrlen
= rqstp
->rq_addrlen
;
1002 dr
->daddr
= rqstp
->rq_daddr
;
1003 dr
->argslen
= rqstp
->rq_arg
.len
>> 2;
1004 dr
->xprt_hlen
= rqstp
->rq_xprt_hlen
;
1006 /* back up head to the start of the buffer and copy */
1007 skip
= rqstp
->rq_arg
.len
- rqstp
->rq_arg
.head
[0].iov_len
;
1008 memcpy(dr
->args
, rqstp
->rq_arg
.head
[0].iov_base
- skip
,
1011 svc_xprt_get(rqstp
->rq_xprt
);
1012 dr
->xprt
= rqstp
->rq_xprt
;
1014 dr
->handle
.revisit
= svc_revisit
;
1019 * recv data from a deferred request into an active one
1021 static int svc_deferred_recv(struct svc_rqst
*rqstp
)
1023 struct svc_deferred_req
*dr
= rqstp
->rq_deferred
;
1025 /* setup iov_base past transport header */
1026 rqstp
->rq_arg
.head
[0].iov_base
= dr
->args
+ (dr
->xprt_hlen
>>2);
1027 /* The iov_len does not include the transport header bytes */
1028 rqstp
->rq_arg
.head
[0].iov_len
= (dr
->argslen
<<2) - dr
->xprt_hlen
;
1029 rqstp
->rq_arg
.page_len
= 0;
1030 /* The rq_arg.len includes the transport header bytes */
1031 rqstp
->rq_arg
.len
= dr
->argslen
<<2;
1032 rqstp
->rq_prot
= dr
->prot
;
1033 memcpy(&rqstp
->rq_addr
, &dr
->addr
, dr
->addrlen
);
1034 rqstp
->rq_addrlen
= dr
->addrlen
;
1035 /* Save off transport header len in case we get deferred again */
1036 rqstp
->rq_xprt_hlen
= dr
->xprt_hlen
;
1037 rqstp
->rq_daddr
= dr
->daddr
;
1038 rqstp
->rq_respages
= rqstp
->rq_pages
;
1039 return (dr
->argslen
<<2) - dr
->xprt_hlen
;
1043 static struct svc_deferred_req
*svc_deferred_dequeue(struct svc_xprt
*xprt
)
1045 struct svc_deferred_req
*dr
= NULL
;
1047 if (!test_bit(XPT_DEFERRED
, &xprt
->xpt_flags
))
1049 spin_lock(&xprt
->xpt_lock
);
1050 clear_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1051 if (!list_empty(&xprt
->xpt_deferred
)) {
1052 dr
= list_entry(xprt
->xpt_deferred
.next
,
1053 struct svc_deferred_req
,
1055 list_del_init(&dr
->handle
.recent
);
1056 set_bit(XPT_DEFERRED
, &xprt
->xpt_flags
);
1058 spin_unlock(&xprt
->xpt_lock
);
1063 * svc_find_xprt - find an RPC transport instance
1064 * @serv: pointer to svc_serv to search
1065 * @xcl_name: C string containing transport's class name
1066 * @af: Address family of transport's local address
1067 * @port: transport's IP port number
1069 * Return the transport instance pointer for the endpoint accepting
1070 * connections/peer traffic from the specified transport class,
1071 * address family and port.
1073 * Specifying 0 for the address family or port is effectively a
1074 * wild-card, and will result in matching the first transport in the
1075 * service's list that has a matching class name.
1077 struct svc_xprt
*svc_find_xprt(struct svc_serv
*serv
, const char *xcl_name
,
1078 const sa_family_t af
, const unsigned short port
)
1080 struct svc_xprt
*xprt
;
1081 struct svc_xprt
*found
= NULL
;
1083 /* Sanity check the args */
1084 if (serv
== NULL
|| xcl_name
== NULL
)
1087 spin_lock_bh(&serv
->sv_lock
);
1088 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1089 if (strcmp(xprt
->xpt_class
->xcl_name
, xcl_name
))
1091 if (af
!= AF_UNSPEC
&& af
!= xprt
->xpt_local
.ss_family
)
1093 if (port
!= 0 && port
!= svc_xprt_local_port(xprt
))
1099 spin_unlock_bh(&serv
->sv_lock
);
1102 EXPORT_SYMBOL_GPL(svc_find_xprt
);
1104 static int svc_one_xprt_name(const struct svc_xprt
*xprt
,
1105 char *pos
, int remaining
)
1109 len
= snprintf(pos
, remaining
, "%s %u\n",
1110 xprt
->xpt_class
->xcl_name
,
1111 svc_xprt_local_port(xprt
));
1112 if (len
>= remaining
)
1113 return -ENAMETOOLONG
;
1118 * svc_xprt_names - format a buffer with a list of transport names
1119 * @serv: pointer to an RPC service
1120 * @buf: pointer to a buffer to be filled in
1121 * @buflen: length of buffer to be filled in
1123 * Fills in @buf with a string containing a list of transport names,
1124 * each name terminated with '\n'.
1126 * Returns positive length of the filled-in string on success; otherwise
1127 * a negative errno value is returned if an error occurs.
1129 int svc_xprt_names(struct svc_serv
*serv
, char *buf
, const int buflen
)
1131 struct svc_xprt
*xprt
;
1135 /* Sanity check args */
1139 spin_lock_bh(&serv
->sv_lock
);
1143 list_for_each_entry(xprt
, &serv
->sv_permsocks
, xpt_list
) {
1144 len
= svc_one_xprt_name(xprt
, pos
, buflen
- totlen
);
1156 spin_unlock_bh(&serv
->sv_lock
);
1159 EXPORT_SYMBOL_GPL(svc_xprt_names
);
1162 /*----------------------------------------------------------------------------*/
1164 static void *svc_pool_stats_start(struct seq_file
*m
, loff_t
*pos
)
1166 unsigned int pidx
= (unsigned int)*pos
;
1167 struct svc_serv
*serv
= m
->private;
1169 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx
);
1172 return SEQ_START_TOKEN
;
1173 return (pidx
> serv
->sv_nrpools
? NULL
: &serv
->sv_pools
[pidx
-1]);
1176 static void *svc_pool_stats_next(struct seq_file
*m
, void *p
, loff_t
*pos
)
1178 struct svc_pool
*pool
= p
;
1179 struct svc_serv
*serv
= m
->private;
1181 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos
);
1183 if (p
== SEQ_START_TOKEN
) {
1184 pool
= &serv
->sv_pools
[0];
1186 unsigned int pidx
= (pool
- &serv
->sv_pools
[0]);
1187 if (pidx
< serv
->sv_nrpools
-1)
1188 pool
= &serv
->sv_pools
[pidx
+1];
1196 static void svc_pool_stats_stop(struct seq_file
*m
, void *p
)
1200 static int svc_pool_stats_show(struct seq_file
*m
, void *p
)
1202 struct svc_pool
*pool
= p
;
1204 if (p
== SEQ_START_TOKEN
) {
1205 seq_puts(m
, "# pool packets-arrived sockets-enqueued threads-woken overloads-avoided threads-timedout\n");
1209 seq_printf(m
, "%u %lu %lu %lu %lu %lu\n",
1211 pool
->sp_stats
.packets
,
1212 pool
->sp_stats
.sockets_queued
,
1213 pool
->sp_stats
.threads_woken
,
1214 pool
->sp_stats
.overloads_avoided
,
1215 pool
->sp_stats
.threads_timedout
);
1220 static const struct seq_operations svc_pool_stats_seq_ops
= {
1221 .start
= svc_pool_stats_start
,
1222 .next
= svc_pool_stats_next
,
1223 .stop
= svc_pool_stats_stop
,
1224 .show
= svc_pool_stats_show
,
1227 int svc_pool_stats_open(struct svc_serv
*serv
, struct file
*file
)
1231 err
= seq_open(file
, &svc_pool_stats_seq_ops
);
1233 ((struct seq_file
*) file
->private_data
)->private = serv
;
1236 EXPORT_SYMBOL(svc_pool_stats_open
);
1238 /*----------------------------------------------------------------------------*/