2 * linux/net/sunrpc/svc_xprt.c
4 * Author: Tom Tucker <tom@opengridcomputing.com>
7 #include <linux/sched.h>
8 #include <linux/errno.h>
9 #include <linux/freezer.h>
10 #include <linux/kthread.h>
12 #include <linux/sunrpc/stats.h>
13 #include <linux/sunrpc/svc_xprt.h>
15 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
17 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
18 static int svc_deferred_recv(struct svc_rqst *rqstp);
19 static struct cache_deferred_req *svc_defer(struct cache_req *req);
20 static void svc_age_temp_xprts(unsigned long closure);
22 /* apparently the "standard" is that clients close
23 * idle connections after 5 minutes, servers after
25 * http://www.connectathon.org/talks96/nfstcp.pdf
27 static int svc_conn_age_period = 6*60;
29 /* List of registered transport classes */
30 static DEFINE_SPINLOCK(svc_xprt_class_lock);
31 static LIST_HEAD(svc_xprt_class_list);
33 /* SMP locking strategy:
35 * svc_pool->sp_lock protects most of the fields of that pool.
36 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
37 * when both need to be taken (rare), svc_serv->sv_lock is first.
38 * BKL protects svc_serv->sv_nrthread.
39 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
40 * and the ->sk_info_authunix cache.
42 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
43 * enqueued multiply. During normal transport processing this bit
44 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
45 * Providers should not manipulate this bit directly.
47 * Some flags can be set to certain values at any time
48 * providing that certain rules are followed:
51 * - Can be set or cleared at any time.
52 * - After a set, svc_xprt_enqueue must be called to enqueue
53 * the transport for processing.
54 * - After a clear, the transport must be read/accepted.
55 * If this succeeds, it must be set again.
57 * - Can set at any time. It is never cleared.
59 * - Can only be set while XPT_BUSY is held which ensures
60 * that no other thread will be using the transport or will
61 * try to set XPT_DEAD.
64 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
66 struct svc_xprt_class *cl;
69 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
71 INIT_LIST_HEAD(&xcl->xcl_list);
72 spin_lock(&svc_xprt_class_lock);
73 /* Make sure there isn't already a class with the same name */
74 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
75 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
78 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
81 spin_unlock(&svc_xprt_class_lock);
84 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
86 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
88 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
89 spin_lock(&svc_xprt_class_lock);
90 list_del_init(&xcl->xcl_list);
91 spin_unlock(&svc_xprt_class_lock);
93 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
96 * Format the transport list for printing
98 int svc_print_xprts(char *buf, int maxlen)
100 struct list_head *le;
105 spin_lock(&svc_xprt_class_lock);
106 list_for_each(le, &svc_xprt_class_list) {
108 struct svc_xprt_class *xcl =
109 list_entry(le, struct svc_xprt_class, 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 && xprt->xpt_auth_cache != NULL)
130 svcauth_unix_info_release(xprt->xpt_auth_cache);
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 mutex_init(&xprt->xpt_mutex);
157 spin_lock_init(&xprt->xpt_lock);
158 set_bit(XPT_BUSY, &xprt->xpt_flags);
160 EXPORT_SYMBOL_GPL(svc_xprt_init);
162 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
163 struct svc_serv *serv,
165 const unsigned short port,
168 struct sockaddr_in sin = {
169 .sin_family = AF_INET,
170 .sin_addr.s_addr = htonl(INADDR_ANY),
171 .sin_port = htons(port),
173 struct sockaddr_in6 sin6 = {
174 .sin6_family = AF_INET6,
175 .sin6_addr = IN6ADDR_ANY_INIT,
176 .sin6_port = htons(port),
178 struct sockaddr *sap;
183 sap = (struct sockaddr *)&sin;
187 sap = (struct sockaddr *)&sin6;
191 return ERR_PTR(-EAFNOSUPPORT);
194 return xcl->xcl_ops->xpo_create(serv, sap, len, flags);
197 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
198 const int family, const unsigned short port,
201 struct svc_xprt_class *xcl;
203 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
204 spin_lock(&svc_xprt_class_lock);
205 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
206 struct svc_xprt *newxprt;
208 if (strcmp(xprt_name, xcl->xcl_name))
211 if (!try_module_get(xcl->xcl_owner))
214 spin_unlock(&svc_xprt_class_lock);
215 newxprt = __svc_xpo_create(xcl, serv, family, port, flags);
216 if (IS_ERR(newxprt)) {
217 module_put(xcl->xcl_owner);
218 return PTR_ERR(newxprt);
221 clear_bit(XPT_TEMP, &newxprt->xpt_flags);
222 spin_lock_bh(&serv->sv_lock);
223 list_add(&newxprt->xpt_list, &serv->sv_permsocks);
224 spin_unlock_bh(&serv->sv_lock);
225 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
226 return svc_xprt_local_port(newxprt);
229 spin_unlock(&svc_xprt_class_lock);
230 dprintk("svc: transport %s not found\n", xprt_name);
233 EXPORT_SYMBOL_GPL(svc_create_xprt);
236 * Copy the local and remote xprt addresses to the rqstp structure
238 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
240 struct sockaddr *sin;
242 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
243 rqstp->rq_addrlen = xprt->xpt_remotelen;
246 * Destination address in request is needed for binding the
247 * source address in RPC replies/callbacks later.
249 sin = (struct sockaddr *)&xprt->xpt_local;
250 switch (sin->sa_family) {
252 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
255 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
259 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
262 * svc_print_addr - Format rq_addr field for printing
263 * @rqstp: svc_rqst struct containing address to print
264 * @buf: target buffer for formatted address
265 * @len: length of target buffer
268 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
270 return __svc_print_addr(svc_addr(rqstp), buf, len);
272 EXPORT_SYMBOL_GPL(svc_print_addr);
275 * Queue up an idle server thread. Must have pool->sp_lock held.
276 * Note: this is really a stack rather than a queue, so that we only
277 * use as many different threads as we need, and the rest don't pollute
280 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
282 list_add(&rqstp->rq_list, &pool->sp_threads);
286 * Dequeue an nfsd thread. Must have pool->sp_lock held.
288 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
290 list_del(&rqstp->rq_list);
294 * Queue up a transport with data pending. If there are idle nfsd
295 * processes, wake 'em up.
298 void svc_xprt_enqueue(struct svc_xprt *xprt)
300 struct svc_serv *serv = xprt->xpt_server;
301 struct svc_pool *pool;
302 struct svc_rqst *rqstp;
305 if (!(xprt->xpt_flags &
306 ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
310 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
313 spin_lock_bh(&pool->sp_lock);
315 if (!list_empty(&pool->sp_threads) &&
316 !list_empty(&pool->sp_sockets))
319 "threads and transports both waiting??\n");
321 if (test_bit(XPT_DEAD, &xprt->xpt_flags)) {
322 /* Don't enqueue dead transports */
323 dprintk("svc: transport %p is dead, not enqueued\n", xprt);
327 /* Mark transport as busy. It will remain in this state until
328 * the provider calls svc_xprt_received. We update XPT_BUSY
329 * atomically because it also guards against trying to enqueue
330 * the transport twice.
332 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
333 /* Don't enqueue transport while already enqueued */
334 dprintk("svc: transport %p busy, not enqueued\n", xprt);
337 BUG_ON(xprt->xpt_pool != NULL);
338 xprt->xpt_pool = pool;
340 /* Handle pending connection */
341 if (test_bit(XPT_CONN, &xprt->xpt_flags))
344 /* Handle close in-progress */
345 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
348 /* Check if we have space to reply to a request */
349 if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
350 /* Don't enqueue while not enough space for reply */
351 dprintk("svc: no write space, transport %p not enqueued\n",
353 xprt->xpt_pool = NULL;
354 clear_bit(XPT_BUSY, &xprt->xpt_flags);
359 if (!list_empty(&pool->sp_threads)) {
360 rqstp = list_entry(pool->sp_threads.next,
363 dprintk("svc: transport %p served by daemon %p\n",
365 svc_thread_dequeue(pool, rqstp);
368 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
369 rqstp, rqstp->rq_xprt);
370 rqstp->rq_xprt = xprt;
372 rqstp->rq_reserved = serv->sv_max_mesg;
373 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
374 BUG_ON(xprt->xpt_pool != pool);
375 wake_up(&rqstp->rq_wait);
377 dprintk("svc: transport %p put into queue\n", xprt);
378 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
379 BUG_ON(xprt->xpt_pool != pool);
383 spin_unlock_bh(&pool->sp_lock);
385 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
388 * Dequeue the first transport. Must be called with the pool->sp_lock held.
390 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
392 struct svc_xprt *xprt;
394 if (list_empty(&pool->sp_sockets))
397 xprt = list_entry(pool->sp_sockets.next,
398 struct svc_xprt, xpt_ready);
399 list_del_init(&xprt->xpt_ready);
401 dprintk("svc: transport %p dequeued, inuse=%d\n",
402 xprt, atomic_read(&xprt->xpt_ref.refcount));
408 * svc_xprt_received conditionally queues the transport for processing
409 * by another thread. The caller must hold the XPT_BUSY bit and must
410 * not thereafter touch transport data.
412 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
413 * insufficient) data.
415 void svc_xprt_received(struct svc_xprt *xprt)
417 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
418 xprt->xpt_pool = NULL;
419 clear_bit(XPT_BUSY, &xprt->xpt_flags);
420 svc_xprt_enqueue(xprt);
422 EXPORT_SYMBOL_GPL(svc_xprt_received);
425 * svc_reserve - change the space reserved for the reply to a request.
426 * @rqstp: The request in question
427 * @space: new max space to reserve
429 * Each request reserves some space on the output queue of the transport
430 * to make sure the reply fits. This function reduces that reserved
431 * space to be the amount of space used already, plus @space.
434 void svc_reserve(struct svc_rqst *rqstp, int space)
436 space += rqstp->rq_res.head[0].iov_len;
438 if (space < rqstp->rq_reserved) {
439 struct svc_xprt *xprt = rqstp->rq_xprt;
440 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
441 rqstp->rq_reserved = space;
443 svc_xprt_enqueue(xprt);
446 EXPORT_SYMBOL_GPL(svc_reserve);
448 static void svc_xprt_release(struct svc_rqst *rqstp)
450 struct svc_xprt *xprt = rqstp->rq_xprt;
452 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
454 kfree(rqstp->rq_deferred);
455 rqstp->rq_deferred = NULL;
457 svc_free_res_pages(rqstp);
458 rqstp->rq_res.page_len = 0;
459 rqstp->rq_res.page_base = 0;
461 /* Reset response buffer and release
463 * But first, check that enough space was reserved
464 * for the reply, otherwise we have a bug!
466 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
467 printk(KERN_ERR "RPC request reserved %d but used %d\n",
471 rqstp->rq_res.head[0].iov_len = 0;
472 svc_reserve(rqstp, 0);
473 rqstp->rq_xprt = NULL;
479 * External function to wake up a server waiting for data
480 * This really only makes sense for services like lockd
481 * which have exactly one thread anyway.
483 void svc_wake_up(struct svc_serv *serv)
485 struct svc_rqst *rqstp;
487 struct svc_pool *pool;
489 for (i = 0; i < serv->sv_nrpools; i++) {
490 pool = &serv->sv_pools[i];
492 spin_lock_bh(&pool->sp_lock);
493 if (!list_empty(&pool->sp_threads)) {
494 rqstp = list_entry(pool->sp_threads.next,
497 dprintk("svc: daemon %p woken up.\n", rqstp);
499 svc_thread_dequeue(pool, rqstp);
500 rqstp->rq_xprt = NULL;
502 wake_up(&rqstp->rq_wait);
504 spin_unlock_bh(&pool->sp_lock);
507 EXPORT_SYMBOL_GPL(svc_wake_up);
509 int svc_port_is_privileged(struct sockaddr *sin)
511 switch (sin->sa_family) {
513 return ntohs(((struct sockaddr_in *)sin)->sin_port)
516 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
524 * Make sure that we don't have too many active connections. If we have,
525 * something must be dropped. It's not clear what will happen if we allow
526 * "too many" connections, but when dealing with network-facing software,
527 * we have to code defensively. Here we do that by imposing hard limits.
529 * There's no point in trying to do random drop here for DoS
530 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
531 * attacker can easily beat that.
533 * The only somewhat efficient mechanism would be if drop old
534 * connections from the same IP first. But right now we don't even
535 * record the client IP in svc_sock.
537 * single-threaded services that expect a lot of clients will probably
538 * need to set sv_maxconn to override the default value which is based
539 * on the number of threads
541 static void svc_check_conn_limits(struct svc_serv *serv)
543 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
544 (serv->sv_nrthreads+3) * 20;
546 if (serv->sv_tmpcnt > limit) {
547 struct svc_xprt *xprt = NULL;
548 spin_lock_bh(&serv->sv_lock);
549 if (!list_empty(&serv->sv_tempsocks)) {
550 if (net_ratelimit()) {
551 /* Try to help the admin */
552 printk(KERN_NOTICE "%s: too many open "
553 "connections, consider increasing %s\n",
554 serv->sv_name, serv->sv_maxconn ?
555 "the max number of connections." :
556 "the number of threads.");
559 * Always select the oldest connection. It's not fair,
562 xprt = list_entry(serv->sv_tempsocks.prev,
565 set_bit(XPT_CLOSE, &xprt->xpt_flags);
568 spin_unlock_bh(&serv->sv_lock);
571 svc_xprt_enqueue(xprt);
578 * Receive the next request on any transport. This code is carefully
579 * organised not to touch any cachelines in the shared svc_serv
580 * structure, only cachelines in the local svc_pool.
582 int svc_recv(struct svc_rqst *rqstp, long timeout)
584 struct svc_xprt *xprt = NULL;
585 struct svc_serv *serv = rqstp->rq_server;
586 struct svc_pool *pool = rqstp->rq_pool;
590 DECLARE_WAITQUEUE(wait, current);
592 dprintk("svc: server %p waiting for data (to = %ld)\n",
597 "svc_recv: service %p, transport not NULL!\n",
599 if (waitqueue_active(&rqstp->rq_wait))
601 "svc_recv: service %p, wait queue active!\n",
604 /* now allocate needed pages. If we get a failure, sleep briefly */
605 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
606 for (i = 0; i < pages ; i++)
607 while (rqstp->rq_pages[i] == NULL) {
608 struct page *p = alloc_page(GFP_KERNEL);
610 set_current_state(TASK_INTERRUPTIBLE);
611 if (signalled() || kthread_should_stop()) {
612 set_current_state(TASK_RUNNING);
615 schedule_timeout(msecs_to_jiffies(500));
617 rqstp->rq_pages[i] = p;
619 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
620 BUG_ON(pages >= RPCSVC_MAXPAGES);
622 /* Make arg->head point to first page and arg->pages point to rest */
623 arg = &rqstp->rq_arg;
624 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
625 arg->head[0].iov_len = PAGE_SIZE;
626 arg->pages = rqstp->rq_pages + 1;
628 /* save at least one page for response */
629 arg->page_len = (pages-2)*PAGE_SIZE;
630 arg->len = (pages-1)*PAGE_SIZE;
631 arg->tail[0].iov_len = 0;
635 if (signalled() || kthread_should_stop())
638 spin_lock_bh(&pool->sp_lock);
639 xprt = svc_xprt_dequeue(pool);
641 rqstp->rq_xprt = xprt;
643 rqstp->rq_reserved = serv->sv_max_mesg;
644 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
646 /* No data pending. Go to sleep */
647 svc_thread_enqueue(pool, rqstp);
650 * We have to be able to interrupt this wait
651 * to bring down the daemons ...
653 set_current_state(TASK_INTERRUPTIBLE);
656 * checking kthread_should_stop() here allows us to avoid
657 * locking and signalling when stopping kthreads that call
658 * svc_recv. If the thread has already been woken up, then
659 * we can exit here without sleeping. If not, then it
660 * it'll be woken up quickly during the schedule_timeout
662 if (kthread_should_stop()) {
663 set_current_state(TASK_RUNNING);
664 spin_unlock_bh(&pool->sp_lock);
668 add_wait_queue(&rqstp->rq_wait, &wait);
669 spin_unlock_bh(&pool->sp_lock);
671 schedule_timeout(timeout);
675 spin_lock_bh(&pool->sp_lock);
676 remove_wait_queue(&rqstp->rq_wait, &wait);
678 xprt = rqstp->rq_xprt;
680 svc_thread_dequeue(pool, rqstp);
681 spin_unlock_bh(&pool->sp_lock);
682 dprintk("svc: server %p, no data yet\n", rqstp);
683 if (signalled() || kthread_should_stop())
689 spin_unlock_bh(&pool->sp_lock);
692 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
693 dprintk("svc_recv: found XPT_CLOSE\n");
694 svc_delete_xprt(xprt);
695 } else if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
696 struct svc_xprt *newxpt;
697 newxpt = xprt->xpt_ops->xpo_accept(xprt);
700 * We know this module_get will succeed because the
701 * listener holds a reference too
703 __module_get(newxpt->xpt_class->xcl_owner);
704 svc_check_conn_limits(xprt->xpt_server);
705 spin_lock_bh(&serv->sv_lock);
706 set_bit(XPT_TEMP, &newxpt->xpt_flags);
707 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
709 if (serv->sv_temptimer.function == NULL) {
710 /* setup timer to age temp transports */
711 setup_timer(&serv->sv_temptimer,
713 (unsigned long)serv);
714 mod_timer(&serv->sv_temptimer,
715 jiffies + svc_conn_age_period * HZ);
717 spin_unlock_bh(&serv->sv_lock);
718 svc_xprt_received(newxpt);
720 svc_xprt_received(xprt);
722 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
723 rqstp, pool->sp_id, xprt,
724 atomic_read(&xprt->xpt_ref.refcount));
725 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
726 if (rqstp->rq_deferred) {
727 svc_xprt_received(xprt);
728 len = svc_deferred_recv(rqstp);
730 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
731 dprintk("svc: got len=%d\n", len);
734 /* No data, incomplete (TCP) read, or accept() */
735 if (len == 0 || len == -EAGAIN) {
736 rqstp->rq_res.len = 0;
737 svc_xprt_release(rqstp);
740 clear_bit(XPT_OLD, &xprt->xpt_flags);
742 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
743 rqstp->rq_chandle.defer = svc_defer;
746 serv->sv_stats->netcnt++;
749 EXPORT_SYMBOL_GPL(svc_recv);
754 void svc_drop(struct svc_rqst *rqstp)
756 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
757 svc_xprt_release(rqstp);
759 EXPORT_SYMBOL_GPL(svc_drop);
762 * Return reply to client.
764 int svc_send(struct svc_rqst *rqstp)
766 struct svc_xprt *xprt;
770 xprt = rqstp->rq_xprt;
774 /* release the receive skb before sending the reply */
775 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
777 /* calculate over-all length */
779 xb->len = xb->head[0].iov_len +
783 /* Grab mutex to serialize outgoing data. */
784 mutex_lock(&xprt->xpt_mutex);
785 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
788 len = xprt->xpt_ops->xpo_sendto(rqstp);
789 mutex_unlock(&xprt->xpt_mutex);
790 svc_xprt_release(rqstp);
792 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
798 * Timer function to close old temporary transports, using
799 * a mark-and-sweep algorithm.
801 static void svc_age_temp_xprts(unsigned long closure)
803 struct svc_serv *serv = (struct svc_serv *)closure;
804 struct svc_xprt *xprt;
805 struct list_head *le, *next;
806 LIST_HEAD(to_be_aged);
808 dprintk("svc_age_temp_xprts\n");
810 if (!spin_trylock_bh(&serv->sv_lock)) {
811 /* busy, try again 1 sec later */
812 dprintk("svc_age_temp_xprts: busy\n");
813 mod_timer(&serv->sv_temptimer, jiffies + HZ);
817 list_for_each_safe(le, next, &serv->sv_tempsocks) {
818 xprt = list_entry(le, struct svc_xprt, xpt_list);
820 /* First time through, just mark it OLD. Second time
821 * through, close it. */
822 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
824 if (atomic_read(&xprt->xpt_ref.refcount) > 1
825 || test_bit(XPT_BUSY, &xprt->xpt_flags))
828 list_move(le, &to_be_aged);
829 set_bit(XPT_CLOSE, &xprt->xpt_flags);
830 set_bit(XPT_DETACHED, &xprt->xpt_flags);
832 spin_unlock_bh(&serv->sv_lock);
834 while (!list_empty(&to_be_aged)) {
835 le = to_be_aged.next;
836 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
838 xprt = list_entry(le, struct svc_xprt, xpt_list);
840 dprintk("queuing xprt %p for closing\n", xprt);
842 /* a thread will dequeue and close it soon */
843 svc_xprt_enqueue(xprt);
847 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
851 * Remove a dead transport
853 void svc_delete_xprt(struct svc_xprt *xprt)
855 struct svc_serv *serv = xprt->xpt_server;
856 struct svc_deferred_req *dr;
858 /* Only do this once */
859 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
862 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
863 xprt->xpt_ops->xpo_detach(xprt);
865 spin_lock_bh(&serv->sv_lock);
866 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
867 list_del_init(&xprt->xpt_list);
869 * We used to delete the transport from whichever list
870 * it's sk_xprt.xpt_ready node was on, but we don't actually
871 * need to. This is because the only time we're called
872 * while still attached to a queue, the queue itself
873 * is about to be destroyed (in svc_destroy).
875 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
878 for (dr = svc_deferred_dequeue(xprt); dr;
879 dr = svc_deferred_dequeue(xprt)) {
885 spin_unlock_bh(&serv->sv_lock);
888 void svc_close_xprt(struct svc_xprt *xprt)
890 set_bit(XPT_CLOSE, &xprt->xpt_flags);
891 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
892 /* someone else will have to effect the close */
896 svc_delete_xprt(xprt);
897 clear_bit(XPT_BUSY, &xprt->xpt_flags);
900 EXPORT_SYMBOL_GPL(svc_close_xprt);
902 void svc_close_all(struct list_head *xprt_list)
904 struct svc_xprt *xprt;
905 struct svc_xprt *tmp;
907 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
908 set_bit(XPT_CLOSE, &xprt->xpt_flags);
909 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
910 /* Waiting to be processed, but no threads left,
911 * So just remove it from the waiting list
913 list_del_init(&xprt->xpt_ready);
914 clear_bit(XPT_BUSY, &xprt->xpt_flags);
916 svc_close_xprt(xprt);
921 * Handle defer and revisit of requests
924 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
926 struct svc_deferred_req *dr =
927 container_of(dreq, struct svc_deferred_req, handle);
928 struct svc_xprt *xprt = dr->xprt;
930 spin_lock(&xprt->xpt_lock);
931 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
932 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
933 spin_unlock(&xprt->xpt_lock);
934 dprintk("revisit canceled\n");
939 dprintk("revisit queued\n");
941 list_add(&dr->handle.recent, &xprt->xpt_deferred);
942 spin_unlock(&xprt->xpt_lock);
943 svc_xprt_enqueue(xprt);
948 * Save the request off for later processing. The request buffer looks
951 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
953 * This code can only handle requests that consist of an xprt-header
956 static struct cache_deferred_req *svc_defer(struct cache_req *req)
958 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
959 struct svc_deferred_req *dr;
961 if (rqstp->rq_arg.page_len)
962 return NULL; /* if more than a page, give up FIXME */
963 if (rqstp->rq_deferred) {
964 dr = rqstp->rq_deferred;
965 rqstp->rq_deferred = NULL;
969 /* FIXME maybe discard if size too large */
970 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
971 dr = kmalloc(size, GFP_KERNEL);
975 dr->handle.owner = rqstp->rq_server;
976 dr->prot = rqstp->rq_prot;
977 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
978 dr->addrlen = rqstp->rq_addrlen;
979 dr->daddr = rqstp->rq_daddr;
980 dr->argslen = rqstp->rq_arg.len >> 2;
981 dr->xprt_hlen = rqstp->rq_xprt_hlen;
983 /* back up head to the start of the buffer and copy */
984 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
985 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
988 svc_xprt_get(rqstp->rq_xprt);
989 dr->xprt = rqstp->rq_xprt;
991 dr->handle.revisit = svc_revisit;
996 * recv data from a deferred request into an active one
998 static int svc_deferred_recv(struct svc_rqst *rqstp)
1000 struct svc_deferred_req *dr = rqstp->rq_deferred;
1002 /* setup iov_base past transport header */
1003 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1004 /* The iov_len does not include the transport header bytes */
1005 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1006 rqstp->rq_arg.page_len = 0;
1007 /* The rq_arg.len includes the transport header bytes */
1008 rqstp->rq_arg.len = dr->argslen<<2;
1009 rqstp->rq_prot = dr->prot;
1010 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1011 rqstp->rq_addrlen = dr->addrlen;
1012 /* Save off transport header len in case we get deferred again */
1013 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1014 rqstp->rq_daddr = dr->daddr;
1015 rqstp->rq_respages = rqstp->rq_pages;
1016 return (dr->argslen<<2) - dr->xprt_hlen;
1020 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1022 struct svc_deferred_req *dr = NULL;
1024 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1026 spin_lock(&xprt->xpt_lock);
1027 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1028 if (!list_empty(&xprt->xpt_deferred)) {
1029 dr = list_entry(xprt->xpt_deferred.next,
1030 struct svc_deferred_req,
1032 list_del_init(&dr->handle.recent);
1033 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1035 spin_unlock(&xprt->xpt_lock);
1040 * svc_find_xprt - find an RPC transport instance
1041 * @serv: pointer to svc_serv to search
1042 * @xcl_name: C string containing transport's class name
1043 * @af: Address family of transport's local address
1044 * @port: transport's IP port number
1046 * Return the transport instance pointer for the endpoint accepting
1047 * connections/peer traffic from the specified transport class,
1048 * address family and port.
1050 * Specifying 0 for the address family or port is effectively a
1051 * wild-card, and will result in matching the first transport in the
1052 * service's list that has a matching class name.
1054 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1055 const sa_family_t af, const unsigned short port)
1057 struct svc_xprt *xprt;
1058 struct svc_xprt *found = NULL;
1060 /* Sanity check the args */
1061 if (serv == NULL || xcl_name == NULL)
1064 spin_lock_bh(&serv->sv_lock);
1065 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1066 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1068 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1070 if (port != 0 && port != svc_xprt_local_port(xprt))
1076 spin_unlock_bh(&serv->sv_lock);
1079 EXPORT_SYMBOL_GPL(svc_find_xprt);
1082 * Format a buffer with a list of the active transports. A zero for
1083 * the buflen parameter disables target buffer overflow checking.
1085 int svc_xprt_names(struct svc_serv *serv, char *buf, int buflen)
1087 struct svc_xprt *xprt;
1092 /* Sanity check args */
1096 spin_lock_bh(&serv->sv_lock);
1097 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1098 len = snprintf(xprt_str, sizeof(xprt_str),
1099 "%s %d\n", xprt->xpt_class->xcl_name,
1100 svc_xprt_local_port(xprt));
1101 /* If the string was truncated, replace with error string */
1102 if (len >= sizeof(xprt_str))
1103 strcpy(xprt_str, "name-too-long\n");
1104 /* Don't overflow buffer */
1105 len = strlen(xprt_str);
1106 if (buflen && (len + totlen >= buflen))
1108 strcpy(buf+totlen, xprt_str);
1111 spin_unlock_bh(&serv->sv_lock);
1114 EXPORT_SYMBOL_GPL(svc_xprt_names);