2 * linux/net/sunrpc/svc_xprt.c
4 * Author: Tom Tucker <tom@opengridcomputing.com>
7 #include <linux/sched.h>
8 #include <linux/errno.h>
9 #include <linux/freezer.h>
10 #include <linux/kthread.h>
12 #include <linux/sunrpc/stats.h>
13 #include <linux/sunrpc/svc_xprt.h>
15 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
17 #define SVC_MAX_WAKING 5
19 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
20 static int svc_deferred_recv(struct svc_rqst *rqstp);
21 static struct cache_deferred_req *svc_defer(struct cache_req *req);
22 static void svc_age_temp_xprts(unsigned long closure);
24 /* apparently the "standard" is that clients close
25 * idle connections after 5 minutes, servers after
27 * http://www.connectathon.org/talks96/nfstcp.pdf
29 static int svc_conn_age_period = 6*60;
31 /* List of registered transport classes */
32 static DEFINE_SPINLOCK(svc_xprt_class_lock);
33 static LIST_HEAD(svc_xprt_class_list);
35 /* SMP locking strategy:
37 * svc_pool->sp_lock protects most of the fields of that pool.
38 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
39 * when both need to be taken (rare), svc_serv->sv_lock is first.
40 * BKL protects svc_serv->sv_nrthread.
41 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
42 * and the ->sk_info_authunix cache.
44 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
45 * enqueued multiply. During normal transport processing this bit
46 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
47 * Providers should not manipulate this bit directly.
49 * Some flags can be set to certain values at any time
50 * providing that certain rules are followed:
53 * - Can be set or cleared at any time.
54 * - After a set, svc_xprt_enqueue must be called to enqueue
55 * the transport for processing.
56 * - After a clear, the transport must be read/accepted.
57 * If this succeeds, it must be set again.
59 * - Can set at any time. It is never cleared.
61 * - Can only be set while XPT_BUSY is held which ensures
62 * that no other thread will be using the transport or will
63 * try to set XPT_DEAD.
66 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
68 struct svc_xprt_class *cl;
71 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
73 INIT_LIST_HEAD(&xcl->xcl_list);
74 spin_lock(&svc_xprt_class_lock);
75 /* Make sure there isn't already a class with the same name */
76 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
77 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
80 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
83 spin_unlock(&svc_xprt_class_lock);
86 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
88 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
90 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
91 spin_lock(&svc_xprt_class_lock);
92 list_del_init(&xcl->xcl_list);
93 spin_unlock(&svc_xprt_class_lock);
95 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
98 * Format the transport list for printing
100 int svc_print_xprts(char *buf, int maxlen)
102 struct list_head *le;
107 spin_lock(&svc_xprt_class_lock);
108 list_for_each(le, &svc_xprt_class_list) {
110 struct svc_xprt_class *xcl =
111 list_entry(le, struct svc_xprt_class, xcl_list);
113 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
114 slen = strlen(tmpstr);
115 if (len + slen > maxlen)
120 spin_unlock(&svc_xprt_class_lock);
125 static void svc_xprt_free(struct kref *kref)
127 struct svc_xprt *xprt =
128 container_of(kref, struct svc_xprt, xpt_ref);
129 struct module *owner = xprt->xpt_class->xcl_owner;
130 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags)
131 && xprt->xpt_auth_cache != NULL)
132 svcauth_unix_info_release(xprt->xpt_auth_cache);
133 xprt->xpt_ops->xpo_free(xprt);
137 void svc_xprt_put(struct svc_xprt *xprt)
139 kref_put(&xprt->xpt_ref, svc_xprt_free);
141 EXPORT_SYMBOL_GPL(svc_xprt_put);
144 * Called by transport drivers to initialize the transport independent
145 * portion of the transport instance.
147 void svc_xprt_init(struct svc_xprt_class *xcl, struct svc_xprt *xprt,
148 struct svc_serv *serv)
150 memset(xprt, 0, sizeof(*xprt));
151 xprt->xpt_class = xcl;
152 xprt->xpt_ops = xcl->xcl_ops;
153 kref_init(&xprt->xpt_ref);
154 xprt->xpt_server = serv;
155 INIT_LIST_HEAD(&xprt->xpt_list);
156 INIT_LIST_HEAD(&xprt->xpt_ready);
157 INIT_LIST_HEAD(&xprt->xpt_deferred);
158 mutex_init(&xprt->xpt_mutex);
159 spin_lock_init(&xprt->xpt_lock);
160 set_bit(XPT_BUSY, &xprt->xpt_flags);
162 EXPORT_SYMBOL_GPL(svc_xprt_init);
164 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
165 struct svc_serv *serv,
166 unsigned short port, int flags)
168 struct sockaddr_in sin = {
169 .sin_family = AF_INET,
170 .sin_addr.s_addr = htonl(INADDR_ANY),
171 .sin_port = htons(port),
173 struct sockaddr_in6 sin6 = {
174 .sin6_family = AF_INET6,
175 .sin6_addr = IN6ADDR_ANY_INIT,
176 .sin6_port = htons(port),
178 struct sockaddr *sap;
181 switch (serv->sv_family) {
183 sap = (struct sockaddr *)&sin;
187 sap = (struct sockaddr *)&sin6;
191 return ERR_PTR(-EAFNOSUPPORT);
194 return xcl->xcl_ops->xpo_create(serv, sap, len, flags);
197 int svc_create_xprt(struct svc_serv *serv, char *xprt_name, unsigned short port,
200 struct svc_xprt_class *xcl;
202 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
203 spin_lock(&svc_xprt_class_lock);
204 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
205 struct svc_xprt *newxprt;
207 if (strcmp(xprt_name, xcl->xcl_name))
210 if (!try_module_get(xcl->xcl_owner))
213 spin_unlock(&svc_xprt_class_lock);
214 newxprt = __svc_xpo_create(xcl, serv, port, flags);
215 if (IS_ERR(newxprt)) {
216 module_put(xcl->xcl_owner);
217 return PTR_ERR(newxprt);
220 clear_bit(XPT_TEMP, &newxprt->xpt_flags);
221 spin_lock_bh(&serv->sv_lock);
222 list_add(&newxprt->xpt_list, &serv->sv_permsocks);
223 spin_unlock_bh(&serv->sv_lock);
224 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
225 return svc_xprt_local_port(newxprt);
228 spin_unlock(&svc_xprt_class_lock);
229 dprintk("svc: transport %s not found\n", xprt_name);
232 EXPORT_SYMBOL_GPL(svc_create_xprt);
235 * Copy the local and remote xprt addresses to the rqstp structure
237 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
239 struct sockaddr *sin;
241 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
242 rqstp->rq_addrlen = xprt->xpt_remotelen;
245 * Destination address in request is needed for binding the
246 * source address in RPC replies/callbacks later.
248 sin = (struct sockaddr *)&xprt->xpt_local;
249 switch (sin->sa_family) {
251 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
254 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
258 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
261 * svc_print_addr - Format rq_addr field for printing
262 * @rqstp: svc_rqst struct containing address to print
263 * @buf: target buffer for formatted address
264 * @len: length of target buffer
267 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
269 return __svc_print_addr(svc_addr(rqstp), buf, len);
271 EXPORT_SYMBOL_GPL(svc_print_addr);
274 * Queue up an idle server thread. Must have pool->sp_lock held.
275 * Note: this is really a stack rather than a queue, so that we only
276 * use as many different threads as we need, and the rest don't pollute
279 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
281 list_add(&rqstp->rq_list, &pool->sp_threads);
285 * Dequeue an nfsd thread. Must have pool->sp_lock held.
287 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
289 list_del(&rqstp->rq_list);
293 * Queue up a transport with data pending. If there are idle nfsd
294 * processes, wake 'em up.
297 void svc_xprt_enqueue(struct svc_xprt *xprt)
299 struct svc_serv *serv = xprt->xpt_server;
300 struct svc_pool *pool;
301 struct svc_rqst *rqstp;
305 if (!(xprt->xpt_flags &
306 ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
310 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
313 spin_lock_bh(&pool->sp_lock);
315 if (test_bit(XPT_DEAD, &xprt->xpt_flags)) {
316 /* Don't enqueue dead transports */
317 dprintk("svc: transport %p is dead, not enqueued\n", xprt);
321 /* Mark transport as busy. It will remain in this state until
322 * the provider calls svc_xprt_received. We update XPT_BUSY
323 * atomically because it also guards against trying to enqueue
324 * the transport twice.
326 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
327 /* Don't enqueue transport while already enqueued */
328 dprintk("svc: transport %p busy, not enqueued\n", xprt);
331 BUG_ON(xprt->xpt_pool != NULL);
332 xprt->xpt_pool = pool;
334 /* Handle pending connection */
335 if (test_bit(XPT_CONN, &xprt->xpt_flags))
338 /* Handle close in-progress */
339 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
342 /* Check if we have space to reply to a request */
343 if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
344 /* Don't enqueue while not enough space for reply */
345 dprintk("svc: no write space, transport %p not enqueued\n",
347 xprt->xpt_pool = NULL;
348 clear_bit(XPT_BUSY, &xprt->xpt_flags);
353 /* Work out whether threads are available */
354 thread_avail = !list_empty(&pool->sp_threads); /* threads are asleep */
355 if (pool->sp_nwaking >= SVC_MAX_WAKING) {
356 /* too many threads are runnable and trying to wake up */
361 rqstp = list_entry(pool->sp_threads.next,
364 dprintk("svc: transport %p served by daemon %p\n",
366 svc_thread_dequeue(pool, rqstp);
369 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
370 rqstp, rqstp->rq_xprt);
371 rqstp->rq_xprt = xprt;
373 rqstp->rq_reserved = serv->sv_max_mesg;
374 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
375 rqstp->rq_waking = 1;
377 BUG_ON(xprt->xpt_pool != pool);
378 wake_up(&rqstp->rq_wait);
380 dprintk("svc: transport %p put into queue\n", xprt);
381 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
382 BUG_ON(xprt->xpt_pool != pool);
386 spin_unlock_bh(&pool->sp_lock);
388 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
391 * Dequeue the first transport. Must be called with the pool->sp_lock held.
393 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
395 struct svc_xprt *xprt;
397 if (list_empty(&pool->sp_sockets))
400 xprt = list_entry(pool->sp_sockets.next,
401 struct svc_xprt, xpt_ready);
402 list_del_init(&xprt->xpt_ready);
404 dprintk("svc: transport %p dequeued, inuse=%d\n",
405 xprt, atomic_read(&xprt->xpt_ref.refcount));
411 * svc_xprt_received conditionally queues the transport for processing
412 * by another thread. The caller must hold the XPT_BUSY bit and must
413 * not thereafter touch transport data.
415 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
416 * insufficient) data.
418 void svc_xprt_received(struct svc_xprt *xprt)
420 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
421 xprt->xpt_pool = NULL;
422 clear_bit(XPT_BUSY, &xprt->xpt_flags);
423 svc_xprt_enqueue(xprt);
425 EXPORT_SYMBOL_GPL(svc_xprt_received);
428 * svc_reserve - change the space reserved for the reply to a request.
429 * @rqstp: The request in question
430 * @space: new max space to reserve
432 * Each request reserves some space on the output queue of the transport
433 * to make sure the reply fits. This function reduces that reserved
434 * space to be the amount of space used already, plus @space.
437 void svc_reserve(struct svc_rqst *rqstp, int space)
439 space += rqstp->rq_res.head[0].iov_len;
441 if (space < rqstp->rq_reserved) {
442 struct svc_xprt *xprt = rqstp->rq_xprt;
443 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
444 rqstp->rq_reserved = space;
446 svc_xprt_enqueue(xprt);
449 EXPORT_SYMBOL_GPL(svc_reserve);
451 static void svc_xprt_release(struct svc_rqst *rqstp)
453 struct svc_xprt *xprt = rqstp->rq_xprt;
455 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
457 kfree(rqstp->rq_deferred);
458 rqstp->rq_deferred = NULL;
460 svc_free_res_pages(rqstp);
461 rqstp->rq_res.page_len = 0;
462 rqstp->rq_res.page_base = 0;
464 /* Reset response buffer and release
466 * But first, check that enough space was reserved
467 * for the reply, otherwise we have a bug!
469 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
470 printk(KERN_ERR "RPC request reserved %d but used %d\n",
474 rqstp->rq_res.head[0].iov_len = 0;
475 svc_reserve(rqstp, 0);
476 rqstp->rq_xprt = NULL;
482 * External function to wake up a server waiting for data
483 * This really only makes sense for services like lockd
484 * which have exactly one thread anyway.
486 void svc_wake_up(struct svc_serv *serv)
488 struct svc_rqst *rqstp;
490 struct svc_pool *pool;
492 for (i = 0; i < serv->sv_nrpools; i++) {
493 pool = &serv->sv_pools[i];
495 spin_lock_bh(&pool->sp_lock);
496 if (!list_empty(&pool->sp_threads)) {
497 rqstp = list_entry(pool->sp_threads.next,
500 dprintk("svc: daemon %p woken up.\n", rqstp);
502 svc_thread_dequeue(pool, rqstp);
503 rqstp->rq_xprt = NULL;
505 wake_up(&rqstp->rq_wait);
507 spin_unlock_bh(&pool->sp_lock);
510 EXPORT_SYMBOL_GPL(svc_wake_up);
512 int svc_port_is_privileged(struct sockaddr *sin)
514 switch (sin->sa_family) {
516 return ntohs(((struct sockaddr_in *)sin)->sin_port)
519 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
527 * Make sure that we don't have too many active connections. If we have,
528 * something must be dropped. It's not clear what will happen if we allow
529 * "too many" connections, but when dealing with network-facing software,
530 * we have to code defensively. Here we do that by imposing hard limits.
532 * There's no point in trying to do random drop here for DoS
533 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
534 * attacker can easily beat that.
536 * The only somewhat efficient mechanism would be if drop old
537 * connections from the same IP first. But right now we don't even
538 * record the client IP in svc_sock.
540 * single-threaded services that expect a lot of clients will probably
541 * need to set sv_maxconn to override the default value which is based
542 * on the number of threads
544 static void svc_check_conn_limits(struct svc_serv *serv)
546 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
547 (serv->sv_nrthreads+3) * 20;
549 if (serv->sv_tmpcnt > limit) {
550 struct svc_xprt *xprt = NULL;
551 spin_lock_bh(&serv->sv_lock);
552 if (!list_empty(&serv->sv_tempsocks)) {
553 if (net_ratelimit()) {
554 /* Try to help the admin */
555 printk(KERN_NOTICE "%s: too many open "
556 "connections, consider increasing %s\n",
557 serv->sv_name, serv->sv_maxconn ?
558 "the max number of connections." :
559 "the number of threads.");
562 * Always select the oldest connection. It's not fair,
565 xprt = list_entry(serv->sv_tempsocks.prev,
568 set_bit(XPT_CLOSE, &xprt->xpt_flags);
571 spin_unlock_bh(&serv->sv_lock);
574 svc_xprt_enqueue(xprt);
581 * Receive the next request on any transport. This code is carefully
582 * organised not to touch any cachelines in the shared svc_serv
583 * structure, only cachelines in the local svc_pool.
585 int svc_recv(struct svc_rqst *rqstp, long timeout)
587 struct svc_xprt *xprt = NULL;
588 struct svc_serv *serv = rqstp->rq_server;
589 struct svc_pool *pool = rqstp->rq_pool;
593 DECLARE_WAITQUEUE(wait, current);
595 dprintk("svc: server %p waiting for data (to = %ld)\n",
600 "svc_recv: service %p, transport not NULL!\n",
602 if (waitqueue_active(&rqstp->rq_wait))
604 "svc_recv: service %p, wait queue active!\n",
607 /* now allocate needed pages. If we get a failure, sleep briefly */
608 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
609 for (i = 0; i < pages ; i++)
610 while (rqstp->rq_pages[i] == NULL) {
611 struct page *p = alloc_page(GFP_KERNEL);
613 set_current_state(TASK_INTERRUPTIBLE);
614 if (signalled() || kthread_should_stop()) {
615 set_current_state(TASK_RUNNING);
618 schedule_timeout(msecs_to_jiffies(500));
620 rqstp->rq_pages[i] = p;
622 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
623 BUG_ON(pages >= RPCSVC_MAXPAGES);
625 /* Make arg->head point to first page and arg->pages point to rest */
626 arg = &rqstp->rq_arg;
627 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
628 arg->head[0].iov_len = PAGE_SIZE;
629 arg->pages = rqstp->rq_pages + 1;
631 /* save at least one page for response */
632 arg->page_len = (pages-2)*PAGE_SIZE;
633 arg->len = (pages-1)*PAGE_SIZE;
634 arg->tail[0].iov_len = 0;
638 if (signalled() || kthread_should_stop())
641 spin_lock_bh(&pool->sp_lock);
642 if (rqstp->rq_waking) {
643 rqstp->rq_waking = 0;
645 BUG_ON(pool->sp_nwaking < 0);
647 xprt = svc_xprt_dequeue(pool);
649 rqstp->rq_xprt = xprt;
651 rqstp->rq_reserved = serv->sv_max_mesg;
652 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
654 /* No data pending. Go to sleep */
655 svc_thread_enqueue(pool, rqstp);
658 * We have to be able to interrupt this wait
659 * to bring down the daemons ...
661 set_current_state(TASK_INTERRUPTIBLE);
664 * checking kthread_should_stop() here allows us to avoid
665 * locking and signalling when stopping kthreads that call
666 * svc_recv. If the thread has already been woken up, then
667 * we can exit here without sleeping. If not, then it
668 * it'll be woken up quickly during the schedule_timeout
670 if (kthread_should_stop()) {
671 set_current_state(TASK_RUNNING);
672 spin_unlock_bh(&pool->sp_lock);
676 add_wait_queue(&rqstp->rq_wait, &wait);
677 spin_unlock_bh(&pool->sp_lock);
679 schedule_timeout(timeout);
683 spin_lock_bh(&pool->sp_lock);
684 remove_wait_queue(&rqstp->rq_wait, &wait);
686 xprt = rqstp->rq_xprt;
688 svc_thread_dequeue(pool, rqstp);
689 spin_unlock_bh(&pool->sp_lock);
690 dprintk("svc: server %p, no data yet\n", rqstp);
691 if (signalled() || kthread_should_stop())
697 spin_unlock_bh(&pool->sp_lock);
700 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
701 dprintk("svc_recv: found XPT_CLOSE\n");
702 svc_delete_xprt(xprt);
703 } else if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
704 struct svc_xprt *newxpt;
705 newxpt = xprt->xpt_ops->xpo_accept(xprt);
708 * We know this module_get will succeed because the
709 * listener holds a reference too
711 __module_get(newxpt->xpt_class->xcl_owner);
712 svc_check_conn_limits(xprt->xpt_server);
713 spin_lock_bh(&serv->sv_lock);
714 set_bit(XPT_TEMP, &newxpt->xpt_flags);
715 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
717 if (serv->sv_temptimer.function == NULL) {
718 /* setup timer to age temp transports */
719 setup_timer(&serv->sv_temptimer,
721 (unsigned long)serv);
722 mod_timer(&serv->sv_temptimer,
723 jiffies + svc_conn_age_period * HZ);
725 spin_unlock_bh(&serv->sv_lock);
726 svc_xprt_received(newxpt);
728 svc_xprt_received(xprt);
730 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
731 rqstp, pool->sp_id, xprt,
732 atomic_read(&xprt->xpt_ref.refcount));
733 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
734 if (rqstp->rq_deferred) {
735 svc_xprt_received(xprt);
736 len = svc_deferred_recv(rqstp);
738 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
739 dprintk("svc: got len=%d\n", len);
742 /* No data, incomplete (TCP) read, or accept() */
743 if (len == 0 || len == -EAGAIN) {
744 rqstp->rq_res.len = 0;
745 svc_xprt_release(rqstp);
748 clear_bit(XPT_OLD, &xprt->xpt_flags);
750 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
751 rqstp->rq_chandle.defer = svc_defer;
754 serv->sv_stats->netcnt++;
757 EXPORT_SYMBOL_GPL(svc_recv);
762 void svc_drop(struct svc_rqst *rqstp)
764 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
765 svc_xprt_release(rqstp);
767 EXPORT_SYMBOL_GPL(svc_drop);
770 * Return reply to client.
772 int svc_send(struct svc_rqst *rqstp)
774 struct svc_xprt *xprt;
778 xprt = rqstp->rq_xprt;
782 /* release the receive skb before sending the reply */
783 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
785 /* calculate over-all length */
787 xb->len = xb->head[0].iov_len +
791 /* Grab mutex to serialize outgoing data. */
792 mutex_lock(&xprt->xpt_mutex);
793 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
796 len = xprt->xpt_ops->xpo_sendto(rqstp);
797 mutex_unlock(&xprt->xpt_mutex);
798 svc_xprt_release(rqstp);
800 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
806 * Timer function to close old temporary transports, using
807 * a mark-and-sweep algorithm.
809 static void svc_age_temp_xprts(unsigned long closure)
811 struct svc_serv *serv = (struct svc_serv *)closure;
812 struct svc_xprt *xprt;
813 struct list_head *le, *next;
814 LIST_HEAD(to_be_aged);
816 dprintk("svc_age_temp_xprts\n");
818 if (!spin_trylock_bh(&serv->sv_lock)) {
819 /* busy, try again 1 sec later */
820 dprintk("svc_age_temp_xprts: busy\n");
821 mod_timer(&serv->sv_temptimer, jiffies + HZ);
825 list_for_each_safe(le, next, &serv->sv_tempsocks) {
826 xprt = list_entry(le, struct svc_xprt, xpt_list);
828 /* First time through, just mark it OLD. Second time
829 * through, close it. */
830 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
832 if (atomic_read(&xprt->xpt_ref.refcount) > 1
833 || test_bit(XPT_BUSY, &xprt->xpt_flags))
836 list_move(le, &to_be_aged);
837 set_bit(XPT_CLOSE, &xprt->xpt_flags);
838 set_bit(XPT_DETACHED, &xprt->xpt_flags);
840 spin_unlock_bh(&serv->sv_lock);
842 while (!list_empty(&to_be_aged)) {
843 le = to_be_aged.next;
844 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
846 xprt = list_entry(le, struct svc_xprt, xpt_list);
848 dprintk("queuing xprt %p for closing\n", xprt);
850 /* a thread will dequeue and close it soon */
851 svc_xprt_enqueue(xprt);
855 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
859 * Remove a dead transport
861 void svc_delete_xprt(struct svc_xprt *xprt)
863 struct svc_serv *serv = xprt->xpt_server;
864 struct svc_deferred_req *dr;
866 /* Only do this once */
867 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
870 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
871 xprt->xpt_ops->xpo_detach(xprt);
873 spin_lock_bh(&serv->sv_lock);
874 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
875 list_del_init(&xprt->xpt_list);
877 * We used to delete the transport from whichever list
878 * it's sk_xprt.xpt_ready node was on, but we don't actually
879 * need to. This is because the only time we're called
880 * while still attached to a queue, the queue itself
881 * is about to be destroyed (in svc_destroy).
883 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
886 for (dr = svc_deferred_dequeue(xprt); dr;
887 dr = svc_deferred_dequeue(xprt)) {
893 spin_unlock_bh(&serv->sv_lock);
896 void svc_close_xprt(struct svc_xprt *xprt)
898 set_bit(XPT_CLOSE, &xprt->xpt_flags);
899 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
900 /* someone else will have to effect the close */
904 svc_delete_xprt(xprt);
905 clear_bit(XPT_BUSY, &xprt->xpt_flags);
908 EXPORT_SYMBOL_GPL(svc_close_xprt);
910 void svc_close_all(struct list_head *xprt_list)
912 struct svc_xprt *xprt;
913 struct svc_xprt *tmp;
915 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
916 set_bit(XPT_CLOSE, &xprt->xpt_flags);
917 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
918 /* Waiting to be processed, but no threads left,
919 * So just remove it from the waiting list
921 list_del_init(&xprt->xpt_ready);
922 clear_bit(XPT_BUSY, &xprt->xpt_flags);
924 svc_close_xprt(xprt);
929 * Handle defer and revisit of requests
932 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
934 struct svc_deferred_req *dr =
935 container_of(dreq, struct svc_deferred_req, handle);
936 struct svc_xprt *xprt = dr->xprt;
938 spin_lock(&xprt->xpt_lock);
939 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
940 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
941 spin_unlock(&xprt->xpt_lock);
942 dprintk("revisit canceled\n");
947 dprintk("revisit queued\n");
949 list_add(&dr->handle.recent, &xprt->xpt_deferred);
950 spin_unlock(&xprt->xpt_lock);
951 svc_xprt_enqueue(xprt);
956 * Save the request off for later processing. The request buffer looks
959 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
961 * This code can only handle requests that consist of an xprt-header
964 static struct cache_deferred_req *svc_defer(struct cache_req *req)
966 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
967 struct svc_deferred_req *dr;
969 if (rqstp->rq_arg.page_len)
970 return NULL; /* if more than a page, give up FIXME */
971 if (rqstp->rq_deferred) {
972 dr = rqstp->rq_deferred;
973 rqstp->rq_deferred = NULL;
977 /* FIXME maybe discard if size too large */
978 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
979 dr = kmalloc(size, GFP_KERNEL);
983 dr->handle.owner = rqstp->rq_server;
984 dr->prot = rqstp->rq_prot;
985 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
986 dr->addrlen = rqstp->rq_addrlen;
987 dr->daddr = rqstp->rq_daddr;
988 dr->argslen = rqstp->rq_arg.len >> 2;
989 dr->xprt_hlen = rqstp->rq_xprt_hlen;
991 /* back up head to the start of the buffer and copy */
992 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
993 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
996 svc_xprt_get(rqstp->rq_xprt);
997 dr->xprt = rqstp->rq_xprt;
999 dr->handle.revisit = svc_revisit;
1004 * recv data from a deferred request into an active one
1006 static int svc_deferred_recv(struct svc_rqst *rqstp)
1008 struct svc_deferred_req *dr = rqstp->rq_deferred;
1010 /* setup iov_base past transport header */
1011 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1012 /* The iov_len does not include the transport header bytes */
1013 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1014 rqstp->rq_arg.page_len = 0;
1015 /* The rq_arg.len includes the transport header bytes */
1016 rqstp->rq_arg.len = dr->argslen<<2;
1017 rqstp->rq_prot = dr->prot;
1018 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1019 rqstp->rq_addrlen = dr->addrlen;
1020 /* Save off transport header len in case we get deferred again */
1021 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1022 rqstp->rq_daddr = dr->daddr;
1023 rqstp->rq_respages = rqstp->rq_pages;
1024 return (dr->argslen<<2) - dr->xprt_hlen;
1028 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1030 struct svc_deferred_req *dr = NULL;
1032 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1034 spin_lock(&xprt->xpt_lock);
1035 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1036 if (!list_empty(&xprt->xpt_deferred)) {
1037 dr = list_entry(xprt->xpt_deferred.next,
1038 struct svc_deferred_req,
1040 list_del_init(&dr->handle.recent);
1041 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1043 spin_unlock(&xprt->xpt_lock);
1048 * Return the transport instance pointer for the endpoint accepting
1049 * connections/peer traffic from the specified transport class,
1050 * address family and port.
1052 * Specifying 0 for the address family or port is effectively a
1053 * wild-card, and will result in matching the first transport in the
1054 * service's list that has a matching class name.
1056 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, char *xcl_name,
1059 struct svc_xprt *xprt;
1060 struct svc_xprt *found = NULL;
1062 /* Sanity check the args */
1063 if (!serv || !xcl_name)
1066 spin_lock_bh(&serv->sv_lock);
1067 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1068 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1070 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1072 if (port && port != svc_xprt_local_port(xprt))
1078 spin_unlock_bh(&serv->sv_lock);
1081 EXPORT_SYMBOL_GPL(svc_find_xprt);
1084 * Format a buffer with a list of the active transports. A zero for
1085 * the buflen parameter disables target buffer overflow checking.
1087 int svc_xprt_names(struct svc_serv *serv, char *buf, int buflen)
1089 struct svc_xprt *xprt;
1094 /* Sanity check args */
1098 spin_lock_bh(&serv->sv_lock);
1099 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1100 len = snprintf(xprt_str, sizeof(xprt_str),
1101 "%s %d\n", xprt->xpt_class->xcl_name,
1102 svc_xprt_local_port(xprt));
1103 /* If the string was truncated, replace with error string */
1104 if (len >= sizeof(xprt_str))
1105 strcpy(xprt_str, "name-too-long\n");
1106 /* Don't overflow buffer */
1107 len = strlen(xprt_str);
1108 if (buflen && (len + totlen >= buflen))
1110 strcpy(buf+totlen, xprt_str);
1113 spin_unlock_bh(&serv->sv_lock);
1116 EXPORT_SYMBOL_GPL(svc_xprt_names);