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,
164 unsigned short port, int flags)
166 struct sockaddr_in sin = {
167 .sin_family = AF_INET,
168 .sin_addr.s_addr = htonl(INADDR_ANY),
169 .sin_port = htons(port),
171 struct sockaddr_in6 sin6 = {
172 .sin6_family = AF_INET6,
173 .sin6_addr = IN6ADDR_ANY_INIT,
174 .sin6_port = htons(port),
176 struct sockaddr *sap;
179 switch (serv->sv_family) {
181 sap = (struct sockaddr *)&sin;
185 sap = (struct sockaddr *)&sin6;
189 return ERR_PTR(-EAFNOSUPPORT);
192 return xcl->xcl_ops->xpo_create(serv, sap, len, flags);
195 int svc_create_xprt(struct svc_serv *serv, char *xprt_name, unsigned short port,
198 struct svc_xprt_class *xcl;
200 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
201 spin_lock(&svc_xprt_class_lock);
202 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
203 struct svc_xprt *newxprt;
205 if (strcmp(xprt_name, xcl->xcl_name))
208 if (!try_module_get(xcl->xcl_owner))
211 spin_unlock(&svc_xprt_class_lock);
212 newxprt = __svc_xpo_create(xcl, serv, port, flags);
213 if (IS_ERR(newxprt)) {
214 module_put(xcl->xcl_owner);
215 return PTR_ERR(newxprt);
218 clear_bit(XPT_TEMP, &newxprt->xpt_flags);
219 spin_lock_bh(&serv->sv_lock);
220 list_add(&newxprt->xpt_list, &serv->sv_permsocks);
221 spin_unlock_bh(&serv->sv_lock);
222 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
223 return svc_xprt_local_port(newxprt);
226 spin_unlock(&svc_xprt_class_lock);
227 dprintk("svc: transport %s not found\n", xprt_name);
230 EXPORT_SYMBOL_GPL(svc_create_xprt);
233 * Copy the local and remote xprt addresses to the rqstp structure
235 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
237 struct sockaddr *sin;
239 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
240 rqstp->rq_addrlen = xprt->xpt_remotelen;
243 * Destination address in request is needed for binding the
244 * source address in RPC replies/callbacks later.
246 sin = (struct sockaddr *)&xprt->xpt_local;
247 switch (sin->sa_family) {
249 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
252 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
256 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
259 * svc_print_addr - Format rq_addr field for printing
260 * @rqstp: svc_rqst struct containing address to print
261 * @buf: target buffer for formatted address
262 * @len: length of target buffer
265 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
267 return __svc_print_addr(svc_addr(rqstp), buf, len);
269 EXPORT_SYMBOL_GPL(svc_print_addr);
272 * Queue up an idle server thread. Must have pool->sp_lock held.
273 * Note: this is really a stack rather than a queue, so that we only
274 * use as many different threads as we need, and the rest don't pollute
277 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
279 list_add(&rqstp->rq_list, &pool->sp_threads);
283 * Dequeue an nfsd thread. Must have pool->sp_lock held.
285 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
287 list_del(&rqstp->rq_list);
291 * Queue up a transport with data pending. If there are idle nfsd
292 * processes, wake 'em up.
295 void svc_xprt_enqueue(struct svc_xprt *xprt)
297 struct svc_serv *serv = xprt->xpt_server;
298 struct svc_pool *pool;
299 struct svc_rqst *rqstp;
302 if (!(xprt->xpt_flags &
303 ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
307 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
310 spin_lock_bh(&pool->sp_lock);
312 if (!list_empty(&pool->sp_threads) &&
313 !list_empty(&pool->sp_sockets))
316 "threads and transports both waiting??\n");
318 if (test_bit(XPT_DEAD, &xprt->xpt_flags)) {
319 /* Don't enqueue dead transports */
320 dprintk("svc: transport %p is dead, not enqueued\n", xprt);
324 /* Mark transport as busy. It will remain in this state until
325 * the provider calls svc_xprt_received. We update XPT_BUSY
326 * atomically because it also guards against trying to enqueue
327 * the transport twice.
329 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
330 /* Don't enqueue transport while already enqueued */
331 dprintk("svc: transport %p busy, not enqueued\n", xprt);
334 BUG_ON(xprt->xpt_pool != NULL);
335 xprt->xpt_pool = pool;
337 /* Handle pending connection */
338 if (test_bit(XPT_CONN, &xprt->xpt_flags))
341 /* Handle close in-progress */
342 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
345 /* Check if we have space to reply to a request */
346 if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
347 /* Don't enqueue while not enough space for reply */
348 dprintk("svc: no write space, transport %p not enqueued\n",
350 xprt->xpt_pool = NULL;
351 clear_bit(XPT_BUSY, &xprt->xpt_flags);
356 if (!list_empty(&pool->sp_threads)) {
357 rqstp = list_entry(pool->sp_threads.next,
360 dprintk("svc: transport %p served by daemon %p\n",
362 svc_thread_dequeue(pool, rqstp);
365 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
366 rqstp, rqstp->rq_xprt);
367 rqstp->rq_xprt = xprt;
369 rqstp->rq_reserved = serv->sv_max_mesg;
370 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
371 BUG_ON(xprt->xpt_pool != pool);
372 wake_up(&rqstp->rq_wait);
374 dprintk("svc: transport %p put into queue\n", xprt);
375 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
376 BUG_ON(xprt->xpt_pool != pool);
380 spin_unlock_bh(&pool->sp_lock);
382 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
385 * Dequeue the first transport. Must be called with the pool->sp_lock held.
387 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
389 struct svc_xprt *xprt;
391 if (list_empty(&pool->sp_sockets))
394 xprt = list_entry(pool->sp_sockets.next,
395 struct svc_xprt, xpt_ready);
396 list_del_init(&xprt->xpt_ready);
398 dprintk("svc: transport %p dequeued, inuse=%d\n",
399 xprt, atomic_read(&xprt->xpt_ref.refcount));
405 * svc_xprt_received conditionally queues the transport for processing
406 * by another thread. The caller must hold the XPT_BUSY bit and must
407 * not thereafter touch transport data.
409 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
410 * insufficient) data.
412 void svc_xprt_received(struct svc_xprt *xprt)
414 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
415 xprt->xpt_pool = NULL;
416 clear_bit(XPT_BUSY, &xprt->xpt_flags);
417 svc_xprt_enqueue(xprt);
419 EXPORT_SYMBOL_GPL(svc_xprt_received);
422 * svc_reserve - change the space reserved for the reply to a request.
423 * @rqstp: The request in question
424 * @space: new max space to reserve
426 * Each request reserves some space on the output queue of the transport
427 * to make sure the reply fits. This function reduces that reserved
428 * space to be the amount of space used already, plus @space.
431 void svc_reserve(struct svc_rqst *rqstp, int space)
433 space += rqstp->rq_res.head[0].iov_len;
435 if (space < rqstp->rq_reserved) {
436 struct svc_xprt *xprt = rqstp->rq_xprt;
437 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
438 rqstp->rq_reserved = space;
440 svc_xprt_enqueue(xprt);
443 EXPORT_SYMBOL(svc_reserve);
445 static void svc_xprt_release(struct svc_rqst *rqstp)
447 struct svc_xprt *xprt = rqstp->rq_xprt;
449 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
451 svc_free_res_pages(rqstp);
452 rqstp->rq_res.page_len = 0;
453 rqstp->rq_res.page_base = 0;
455 /* Reset response buffer and release
457 * But first, check that enough space was reserved
458 * for the reply, otherwise we have a bug!
460 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
461 printk(KERN_ERR "RPC request reserved %d but used %d\n",
465 rqstp->rq_res.head[0].iov_len = 0;
466 svc_reserve(rqstp, 0);
467 rqstp->rq_xprt = NULL;
473 * External function to wake up a server waiting for data
474 * This really only makes sense for services like lockd
475 * which have exactly one thread anyway.
477 void svc_wake_up(struct svc_serv *serv)
479 struct svc_rqst *rqstp;
481 struct svc_pool *pool;
483 for (i = 0; i < serv->sv_nrpools; i++) {
484 pool = &serv->sv_pools[i];
486 spin_lock_bh(&pool->sp_lock);
487 if (!list_empty(&pool->sp_threads)) {
488 rqstp = list_entry(pool->sp_threads.next,
491 dprintk("svc: daemon %p woken up.\n", rqstp);
493 svc_thread_dequeue(pool, rqstp);
494 rqstp->rq_xprt = NULL;
496 wake_up(&rqstp->rq_wait);
498 spin_unlock_bh(&pool->sp_lock);
501 EXPORT_SYMBOL(svc_wake_up);
503 int svc_port_is_privileged(struct sockaddr *sin)
505 switch (sin->sa_family) {
507 return ntohs(((struct sockaddr_in *)sin)->sin_port)
510 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
518 * Make sure that we don't have too many active connections. If we have,
519 * something must be dropped. It's not clear what will happen if we allow
520 * "too many" connections, but when dealing with network-facing software,
521 * we have to code defensively. Here we do that by imposing hard limits.
523 * There's no point in trying to do random drop here for DoS
524 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
525 * attacker can easily beat that.
527 * The only somewhat efficient mechanism would be if drop old
528 * connections from the same IP first. But right now we don't even
529 * record the client IP in svc_sock.
531 * single-threaded services that expect a lot of clients will probably
532 * need to set sv_maxconn to override the default value which is based
533 * on the number of threads
535 static void svc_check_conn_limits(struct svc_serv *serv)
537 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
538 (serv->sv_nrthreads+3) * 20;
540 if (serv->sv_tmpcnt > limit) {
541 struct svc_xprt *xprt = NULL;
542 spin_lock_bh(&serv->sv_lock);
543 if (!list_empty(&serv->sv_tempsocks)) {
544 if (net_ratelimit()) {
545 /* Try to help the admin */
546 printk(KERN_NOTICE "%s: too many open "
547 "connections, consider increasing %s\n",
548 serv->sv_name, serv->sv_maxconn ?
549 "the max number of connections." :
550 "the number of threads.");
553 * Always select the oldest connection. It's not fair,
556 xprt = list_entry(serv->sv_tempsocks.prev,
559 set_bit(XPT_CLOSE, &xprt->xpt_flags);
562 spin_unlock_bh(&serv->sv_lock);
565 svc_xprt_enqueue(xprt);
572 * Receive the next request on any transport. This code is carefully
573 * organised not to touch any cachelines in the shared svc_serv
574 * structure, only cachelines in the local svc_pool.
576 int svc_recv(struct svc_rqst *rqstp, long timeout)
578 struct svc_xprt *xprt = NULL;
579 struct svc_serv *serv = rqstp->rq_server;
580 struct svc_pool *pool = rqstp->rq_pool;
584 DECLARE_WAITQUEUE(wait, current);
586 dprintk("svc: server %p waiting for data (to = %ld)\n",
591 "svc_recv: service %p, transport not NULL!\n",
593 if (waitqueue_active(&rqstp->rq_wait))
595 "svc_recv: service %p, wait queue active!\n",
598 /* now allocate needed pages. If we get a failure, sleep briefly */
599 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
600 for (i = 0; i < pages ; i++)
601 while (rqstp->rq_pages[i] == NULL) {
602 struct page *p = alloc_page(GFP_KERNEL);
604 set_current_state(TASK_INTERRUPTIBLE);
605 if (signalled() || kthread_should_stop()) {
606 set_current_state(TASK_RUNNING);
609 schedule_timeout(msecs_to_jiffies(500));
611 rqstp->rq_pages[i] = p;
613 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
614 BUG_ON(pages >= RPCSVC_MAXPAGES);
616 /* Make arg->head point to first page and arg->pages point to rest */
617 arg = &rqstp->rq_arg;
618 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
619 arg->head[0].iov_len = PAGE_SIZE;
620 arg->pages = rqstp->rq_pages + 1;
622 /* save at least one page for response */
623 arg->page_len = (pages-2)*PAGE_SIZE;
624 arg->len = (pages-1)*PAGE_SIZE;
625 arg->tail[0].iov_len = 0;
629 if (signalled() || kthread_should_stop())
632 spin_lock_bh(&pool->sp_lock);
633 xprt = svc_xprt_dequeue(pool);
635 rqstp->rq_xprt = xprt;
637 rqstp->rq_reserved = serv->sv_max_mesg;
638 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
640 /* No data pending. Go to sleep */
641 svc_thread_enqueue(pool, rqstp);
644 * We have to be able to interrupt this wait
645 * to bring down the daemons ...
647 set_current_state(TASK_INTERRUPTIBLE);
650 * checking kthread_should_stop() here allows us to avoid
651 * locking and signalling when stopping kthreads that call
652 * svc_recv. If the thread has already been woken up, then
653 * we can exit here without sleeping. If not, then it
654 * it'll be woken up quickly during the schedule_timeout
656 if (kthread_should_stop()) {
657 set_current_state(TASK_RUNNING);
658 spin_unlock_bh(&pool->sp_lock);
662 add_wait_queue(&rqstp->rq_wait, &wait);
663 spin_unlock_bh(&pool->sp_lock);
665 schedule_timeout(timeout);
669 spin_lock_bh(&pool->sp_lock);
670 remove_wait_queue(&rqstp->rq_wait, &wait);
672 xprt = rqstp->rq_xprt;
674 svc_thread_dequeue(pool, rqstp);
675 spin_unlock_bh(&pool->sp_lock);
676 dprintk("svc: server %p, no data yet\n", rqstp);
677 if (signalled() || kthread_should_stop())
683 spin_unlock_bh(&pool->sp_lock);
686 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
687 dprintk("svc_recv: found XPT_CLOSE\n");
688 svc_delete_xprt(xprt);
689 } else if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
690 struct svc_xprt *newxpt;
691 newxpt = xprt->xpt_ops->xpo_accept(xprt);
694 * We know this module_get will succeed because the
695 * listener holds a reference too
697 __module_get(newxpt->xpt_class->xcl_owner);
698 svc_check_conn_limits(xprt->xpt_server);
699 spin_lock_bh(&serv->sv_lock);
700 set_bit(XPT_TEMP, &newxpt->xpt_flags);
701 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
703 if (serv->sv_temptimer.function == NULL) {
704 /* setup timer to age temp transports */
705 setup_timer(&serv->sv_temptimer,
707 (unsigned long)serv);
708 mod_timer(&serv->sv_temptimer,
709 jiffies + svc_conn_age_period * HZ);
711 spin_unlock_bh(&serv->sv_lock);
712 svc_xprt_received(newxpt);
714 svc_xprt_received(xprt);
716 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
717 rqstp, pool->sp_id, xprt,
718 atomic_read(&xprt->xpt_ref.refcount));
719 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
720 if (rqstp->rq_deferred) {
721 svc_xprt_received(xprt);
722 len = svc_deferred_recv(rqstp);
724 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
725 dprintk("svc: got len=%d\n", len);
728 /* No data, incomplete (TCP) read, or accept() */
729 if (len == 0 || len == -EAGAIN) {
730 rqstp->rq_res.len = 0;
731 svc_xprt_release(rqstp);
734 clear_bit(XPT_OLD, &xprt->xpt_flags);
736 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
737 rqstp->rq_chandle.defer = svc_defer;
740 serv->sv_stats->netcnt++;
743 EXPORT_SYMBOL(svc_recv);
748 void svc_drop(struct svc_rqst *rqstp)
750 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
751 svc_xprt_release(rqstp);
753 EXPORT_SYMBOL(svc_drop);
756 * Return reply to client.
758 int svc_send(struct svc_rqst *rqstp)
760 struct svc_xprt *xprt;
764 xprt = rqstp->rq_xprt;
768 /* release the receive skb before sending the reply */
769 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
771 /* calculate over-all length */
773 xb->len = xb->head[0].iov_len +
777 /* Grab mutex to serialize outgoing data. */
778 mutex_lock(&xprt->xpt_mutex);
779 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
782 len = xprt->xpt_ops->xpo_sendto(rqstp);
783 mutex_unlock(&xprt->xpt_mutex);
784 svc_xprt_release(rqstp);
786 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
792 * Timer function to close old temporary transports, using
793 * a mark-and-sweep algorithm.
795 static void svc_age_temp_xprts(unsigned long closure)
797 struct svc_serv *serv = (struct svc_serv *)closure;
798 struct svc_xprt *xprt;
799 struct list_head *le, *next;
800 LIST_HEAD(to_be_aged);
802 dprintk("svc_age_temp_xprts\n");
804 if (!spin_trylock_bh(&serv->sv_lock)) {
805 /* busy, try again 1 sec later */
806 dprintk("svc_age_temp_xprts: busy\n");
807 mod_timer(&serv->sv_temptimer, jiffies + HZ);
811 list_for_each_safe(le, next, &serv->sv_tempsocks) {
812 xprt = list_entry(le, struct svc_xprt, xpt_list);
814 /* First time through, just mark it OLD. Second time
815 * through, close it. */
816 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
818 if (atomic_read(&xprt->xpt_ref.refcount) > 1
819 || test_bit(XPT_BUSY, &xprt->xpt_flags))
822 list_move(le, &to_be_aged);
823 set_bit(XPT_CLOSE, &xprt->xpt_flags);
824 set_bit(XPT_DETACHED, &xprt->xpt_flags);
826 spin_unlock_bh(&serv->sv_lock);
828 while (!list_empty(&to_be_aged)) {
829 le = to_be_aged.next;
830 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
832 xprt = list_entry(le, struct svc_xprt, xpt_list);
834 dprintk("queuing xprt %p for closing\n", xprt);
836 /* a thread will dequeue and close it soon */
837 svc_xprt_enqueue(xprt);
841 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
845 * Remove a dead transport
847 void svc_delete_xprt(struct svc_xprt *xprt)
849 struct svc_serv *serv = xprt->xpt_server;
851 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
852 xprt->xpt_ops->xpo_detach(xprt);
854 spin_lock_bh(&serv->sv_lock);
855 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
856 list_del_init(&xprt->xpt_list);
858 * We used to delete the transport from whichever list
859 * it's sk_xprt.xpt_ready node was on, but we don't actually
860 * need to. This is because the only time we're called
861 * while still attached to a queue, the queue itself
862 * is about to be destroyed (in svc_destroy).
864 if (!test_and_set_bit(XPT_DEAD, &xprt->xpt_flags)) {
865 BUG_ON(atomic_read(&xprt->xpt_ref.refcount) < 2);
866 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
870 spin_unlock_bh(&serv->sv_lock);
873 void svc_close_xprt(struct svc_xprt *xprt)
875 set_bit(XPT_CLOSE, &xprt->xpt_flags);
876 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
877 /* someone else will have to effect the close */
881 svc_delete_xprt(xprt);
882 clear_bit(XPT_BUSY, &xprt->xpt_flags);
885 EXPORT_SYMBOL_GPL(svc_close_xprt);
887 void svc_close_all(struct list_head *xprt_list)
889 struct svc_xprt *xprt;
890 struct svc_xprt *tmp;
892 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
893 set_bit(XPT_CLOSE, &xprt->xpt_flags);
894 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
895 /* Waiting to be processed, but no threads left,
896 * So just remove it from the waiting list
898 list_del_init(&xprt->xpt_ready);
899 clear_bit(XPT_BUSY, &xprt->xpt_flags);
901 svc_close_xprt(xprt);
906 * Handle defer and revisit of requests
909 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
911 struct svc_deferred_req *dr =
912 container_of(dreq, struct svc_deferred_req, handle);
913 struct svc_xprt *xprt = dr->xprt;
920 dprintk("revisit queued\n");
922 spin_lock(&xprt->xpt_lock);
923 list_add(&dr->handle.recent, &xprt->xpt_deferred);
924 spin_unlock(&xprt->xpt_lock);
925 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
926 svc_xprt_enqueue(xprt);
931 * Save the request off for later processing. The request buffer looks
934 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
936 * This code can only handle requests that consist of an xprt-header
939 static struct cache_deferred_req *svc_defer(struct cache_req *req)
941 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
942 struct svc_deferred_req *dr;
944 if (rqstp->rq_arg.page_len)
945 return NULL; /* if more than a page, give up FIXME */
946 if (rqstp->rq_deferred) {
947 dr = rqstp->rq_deferred;
948 rqstp->rq_deferred = NULL;
952 /* FIXME maybe discard if size too large */
953 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
954 dr = kmalloc(size, GFP_KERNEL);
958 dr->handle.owner = rqstp->rq_server;
959 dr->prot = rqstp->rq_prot;
960 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
961 dr->addrlen = rqstp->rq_addrlen;
962 dr->daddr = rqstp->rq_daddr;
963 dr->argslen = rqstp->rq_arg.len >> 2;
964 dr->xprt_hlen = rqstp->rq_xprt_hlen;
966 /* back up head to the start of the buffer and copy */
967 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
968 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
971 svc_xprt_get(rqstp->rq_xprt);
972 dr->xprt = rqstp->rq_xprt;
974 dr->handle.revisit = svc_revisit;
979 * recv data from a deferred request into an active one
981 static int svc_deferred_recv(struct svc_rqst *rqstp)
983 struct svc_deferred_req *dr = rqstp->rq_deferred;
985 /* setup iov_base past transport header */
986 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
987 /* The iov_len does not include the transport header bytes */
988 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
989 rqstp->rq_arg.page_len = 0;
990 /* The rq_arg.len includes the transport header bytes */
991 rqstp->rq_arg.len = dr->argslen<<2;
992 rqstp->rq_prot = dr->prot;
993 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
994 rqstp->rq_addrlen = dr->addrlen;
995 /* Save off transport header len in case we get deferred again */
996 rqstp->rq_xprt_hlen = dr->xprt_hlen;
997 rqstp->rq_daddr = dr->daddr;
998 rqstp->rq_respages = rqstp->rq_pages;
999 return (dr->argslen<<2) - dr->xprt_hlen;
1003 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1005 struct svc_deferred_req *dr = NULL;
1007 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1009 spin_lock(&xprt->xpt_lock);
1010 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1011 if (!list_empty(&xprt->xpt_deferred)) {
1012 dr = list_entry(xprt->xpt_deferred.next,
1013 struct svc_deferred_req,
1015 list_del_init(&dr->handle.recent);
1016 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1018 spin_unlock(&xprt->xpt_lock);
1023 * Return the transport instance pointer for the endpoint accepting
1024 * connections/peer traffic from the specified transport class,
1025 * address family and port.
1027 * Specifying 0 for the address family or port is effectively a
1028 * wild-card, and will result in matching the first transport in the
1029 * service's list that has a matching class name.
1031 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, char *xcl_name,
1034 struct svc_xprt *xprt;
1035 struct svc_xprt *found = NULL;
1037 /* Sanity check the args */
1038 if (!serv || !xcl_name)
1041 spin_lock_bh(&serv->sv_lock);
1042 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1043 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1045 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1047 if (port && port != svc_xprt_local_port(xprt))
1053 spin_unlock_bh(&serv->sv_lock);
1056 EXPORT_SYMBOL_GPL(svc_find_xprt);
1059 * Format a buffer with a list of the active transports. A zero for
1060 * the buflen parameter disables target buffer overflow checking.
1062 int svc_xprt_names(struct svc_serv *serv, char *buf, int buflen)
1064 struct svc_xprt *xprt;
1069 /* Sanity check args */
1073 spin_lock_bh(&serv->sv_lock);
1074 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1075 len = snprintf(xprt_str, sizeof(xprt_str),
1076 "%s %d\n", xprt->xpt_class->xcl_name,
1077 svc_xprt_local_port(xprt));
1078 /* If the string was truncated, replace with error string */
1079 if (len >= sizeof(xprt_str))
1080 strcpy(xprt_str, "name-too-long\n");
1081 /* Don't overflow buffer */
1082 len = strlen(xprt_str);
1083 if (buflen && (len + totlen >= buflen))
1085 strcpy(buf+totlen, xprt_str);
1088 spin_unlock_bh(&serv->sv_lock);
1091 EXPORT_SYMBOL_GPL(svc_xprt_names);