2 * linux/net/sunrpc/svc_xprt.c
4 * Author: Tom Tucker <tom@opengridcomputing.com>
7 #include <linux/sched.h>
8 #include <linux/errno.h>
9 #include <linux/freezer.h>
10 #include <linux/kthread.h>
11 #include <linux/slab.h>
13 #include <linux/sunrpc/stats.h>
14 #include <linux/sunrpc/svc_xprt.h>
15 #include <linux/sunrpc/svcsock.h>
17 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
19 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
20 static int svc_deferred_recv(struct svc_rqst *rqstp);
21 static struct cache_deferred_req *svc_defer(struct cache_req *req);
22 static void svc_age_temp_xprts(unsigned long closure);
24 /* apparently the "standard" is that clients close
25 * idle connections after 5 minutes, servers after
27 * http://www.connectathon.org/talks96/nfstcp.pdf
29 static int svc_conn_age_period = 6*60;
31 /* List of registered transport classes */
32 static DEFINE_SPINLOCK(svc_xprt_class_lock);
33 static LIST_HEAD(svc_xprt_class_list);
35 /* SMP locking strategy:
37 * svc_pool->sp_lock protects most of the fields of that pool.
38 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
39 * when both need to be taken (rare), svc_serv->sv_lock is first.
40 * BKL protects svc_serv->sv_nrthread.
41 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
42 * and the ->sk_info_authunix cache.
44 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
45 * enqueued multiply. During normal transport processing this bit
46 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
47 * Providers should not manipulate this bit directly.
49 * Some flags can be set to certain values at any time
50 * providing that certain rules are followed:
53 * - Can be set or cleared at any time.
54 * - After a set, svc_xprt_enqueue must be called to enqueue
55 * the transport for processing.
56 * - After a clear, the transport must be read/accepted.
57 * If this succeeds, it must be set again.
59 * - Can set at any time. It is never cleared.
61 * - Can only be set while XPT_BUSY is held which ensures
62 * that no other thread will be using the transport or will
63 * try to set XPT_DEAD.
66 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
68 struct svc_xprt_class *cl;
71 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
73 INIT_LIST_HEAD(&xcl->xcl_list);
74 spin_lock(&svc_xprt_class_lock);
75 /* Make sure there isn't already a class with the same name */
76 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
77 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
80 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
83 spin_unlock(&svc_xprt_class_lock);
86 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
88 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
90 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
91 spin_lock(&svc_xprt_class_lock);
92 list_del_init(&xcl->xcl_list);
93 spin_unlock(&svc_xprt_class_lock);
95 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
98 * Format the transport list for printing
100 int svc_print_xprts(char *buf, int maxlen)
102 struct svc_xprt_class *xcl;
107 spin_lock(&svc_xprt_class_lock);
108 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
111 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
112 slen = strlen(tmpstr);
113 if (len + slen > maxlen)
118 spin_unlock(&svc_xprt_class_lock);
123 static void svc_xprt_free(struct kref *kref)
125 struct svc_xprt *xprt =
126 container_of(kref, struct svc_xprt, xpt_ref);
127 struct module *owner = xprt->xpt_class->xcl_owner;
128 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
129 svcauth_unix_info_release(xprt);
130 put_net(xprt->xpt_net);
131 xprt->xpt_ops->xpo_free(xprt);
135 void svc_xprt_put(struct svc_xprt *xprt)
137 kref_put(&xprt->xpt_ref, svc_xprt_free);
139 EXPORT_SYMBOL_GPL(svc_xprt_put);
142 * Called by transport drivers to initialize the transport independent
143 * portion of the transport instance.
145 void svc_xprt_init(struct svc_xprt_class *xcl, struct svc_xprt *xprt,
146 struct svc_serv *serv)
148 memset(xprt, 0, sizeof(*xprt));
149 xprt->xpt_class = xcl;
150 xprt->xpt_ops = xcl->xcl_ops;
151 kref_init(&xprt->xpt_ref);
152 xprt->xpt_server = serv;
153 INIT_LIST_HEAD(&xprt->xpt_list);
154 INIT_LIST_HEAD(&xprt->xpt_ready);
155 INIT_LIST_HEAD(&xprt->xpt_deferred);
156 INIT_LIST_HEAD(&xprt->xpt_users);
157 mutex_init(&xprt->xpt_mutex);
158 spin_lock_init(&xprt->xpt_lock);
159 set_bit(XPT_BUSY, &xprt->xpt_flags);
160 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
161 xprt->xpt_net = get_net(&init_net);
163 EXPORT_SYMBOL_GPL(svc_xprt_init);
165 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
166 struct svc_serv *serv,
169 const unsigned short port,
172 struct sockaddr_in sin = {
173 .sin_family = AF_INET,
174 .sin_addr.s_addr = htonl(INADDR_ANY),
175 .sin_port = htons(port),
177 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
178 struct sockaddr_in6 sin6 = {
179 .sin6_family = AF_INET6,
180 .sin6_addr = IN6ADDR_ANY_INIT,
181 .sin6_port = htons(port),
183 #endif /* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
184 struct sockaddr *sap;
189 sap = (struct sockaddr *)&sin;
192 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
194 sap = (struct sockaddr *)&sin6;
197 #endif /* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
199 return ERR_PTR(-EAFNOSUPPORT);
202 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
205 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
206 struct net *net, const int family,
207 const unsigned short port, int flags)
209 struct svc_xprt_class *xcl;
211 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
212 spin_lock(&svc_xprt_class_lock);
213 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
214 struct svc_xprt *newxprt;
215 unsigned short newport;
217 if (strcmp(xprt_name, xcl->xcl_name))
220 if (!try_module_get(xcl->xcl_owner))
223 spin_unlock(&svc_xprt_class_lock);
224 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
225 if (IS_ERR(newxprt)) {
226 module_put(xcl->xcl_owner);
227 return PTR_ERR(newxprt);
230 clear_bit(XPT_TEMP, &newxprt->xpt_flags);
231 spin_lock_bh(&serv->sv_lock);
232 list_add(&newxprt->xpt_list, &serv->sv_permsocks);
233 spin_unlock_bh(&serv->sv_lock);
234 newport = svc_xprt_local_port(newxprt);
235 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
239 spin_unlock(&svc_xprt_class_lock);
240 dprintk("svc: transport %s not found\n", xprt_name);
242 /* This errno is exposed to user space. Provide a reasonable
243 * perror msg for a bad transport. */
244 return -EPROTONOSUPPORT;
246 EXPORT_SYMBOL_GPL(svc_create_xprt);
249 * Copy the local and remote xprt addresses to the rqstp structure
251 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
253 struct sockaddr *sin;
255 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
256 rqstp->rq_addrlen = xprt->xpt_remotelen;
259 * Destination address in request is needed for binding the
260 * source address in RPC replies/callbacks later.
262 sin = (struct sockaddr *)&xprt->xpt_local;
263 switch (sin->sa_family) {
265 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
268 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
272 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
275 * svc_print_addr - Format rq_addr field for printing
276 * @rqstp: svc_rqst struct containing address to print
277 * @buf: target buffer for formatted address
278 * @len: length of target buffer
281 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
283 return __svc_print_addr(svc_addr(rqstp), buf, len);
285 EXPORT_SYMBOL_GPL(svc_print_addr);
288 * Queue up an idle server thread. Must have pool->sp_lock held.
289 * Note: this is really a stack rather than a queue, so that we only
290 * use as many different threads as we need, and the rest don't pollute
293 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
295 list_add(&rqstp->rq_list, &pool->sp_threads);
299 * Dequeue an nfsd thread. Must have pool->sp_lock held.
301 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
303 list_del(&rqstp->rq_list);
306 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
308 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
310 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
311 return xprt->xpt_ops->xpo_has_wspace(xprt);
316 * Queue up a transport with data pending. If there are idle nfsd
317 * processes, wake 'em up.
320 void svc_xprt_enqueue(struct svc_xprt *xprt)
322 struct svc_serv *serv = xprt->xpt_server;
323 struct svc_pool *pool;
324 struct svc_rqst *rqstp;
327 if (!svc_xprt_has_something_to_do(xprt))
331 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
334 spin_lock_bh(&pool->sp_lock);
336 if (!list_empty(&pool->sp_threads) &&
337 !list_empty(&pool->sp_sockets))
340 "threads and transports both waiting??\n");
342 pool->sp_stats.packets++;
344 /* Mark transport as busy. It will remain in this state until
345 * the provider calls svc_xprt_received. We update XPT_BUSY
346 * atomically because it also guards against trying to enqueue
347 * the transport twice.
349 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
350 /* Don't enqueue transport while already enqueued */
351 dprintk("svc: transport %p busy, not enqueued\n", xprt);
354 BUG_ON(xprt->xpt_pool != NULL);
355 xprt->xpt_pool = pool;
357 if (!list_empty(&pool->sp_threads)) {
358 rqstp = list_entry(pool->sp_threads.next,
361 dprintk("svc: transport %p served by daemon %p\n",
363 svc_thread_dequeue(pool, rqstp);
366 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
367 rqstp, rqstp->rq_xprt);
368 rqstp->rq_xprt = xprt;
370 rqstp->rq_reserved = serv->sv_max_mesg;
371 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
372 pool->sp_stats.threads_woken++;
373 BUG_ON(xprt->xpt_pool != pool);
374 wake_up(&rqstp->rq_wait);
376 dprintk("svc: transport %p put into queue\n", xprt);
377 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
378 pool->sp_stats.sockets_queued++;
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 /* As soon as we clear busy, the xprt could be closed and
420 * 'put', so we need a reference to call svc_xprt_enqueue with:
423 clear_bit(XPT_BUSY, &xprt->xpt_flags);
424 svc_xprt_enqueue(xprt);
427 EXPORT_SYMBOL_GPL(svc_xprt_received);
430 * svc_reserve - change the space reserved for the reply to a request.
431 * @rqstp: The request in question
432 * @space: new max space to reserve
434 * Each request reserves some space on the output queue of the transport
435 * to make sure the reply fits. This function reduces that reserved
436 * space to be the amount of space used already, plus @space.
439 void svc_reserve(struct svc_rqst *rqstp, int space)
441 space += rqstp->rq_res.head[0].iov_len;
443 if (space < rqstp->rq_reserved) {
444 struct svc_xprt *xprt = rqstp->rq_xprt;
445 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
446 rqstp->rq_reserved = space;
448 svc_xprt_enqueue(xprt);
451 EXPORT_SYMBOL_GPL(svc_reserve);
453 static void svc_xprt_release(struct svc_rqst *rqstp)
455 struct svc_xprt *xprt = rqstp->rq_xprt;
457 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
459 kfree(rqstp->rq_deferred);
460 rqstp->rq_deferred = NULL;
462 svc_free_res_pages(rqstp);
463 rqstp->rq_res.page_len = 0;
464 rqstp->rq_res.page_base = 0;
466 /* Reset response buffer and release
468 * But first, check that enough space was reserved
469 * for the reply, otherwise we have a bug!
471 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
472 printk(KERN_ERR "RPC request reserved %d but used %d\n",
476 rqstp->rq_res.head[0].iov_len = 0;
477 svc_reserve(rqstp, 0);
478 rqstp->rq_xprt = NULL;
484 * External function to wake up a server waiting for data
485 * This really only makes sense for services like lockd
486 * which have exactly one thread anyway.
488 void svc_wake_up(struct svc_serv *serv)
490 struct svc_rqst *rqstp;
492 struct svc_pool *pool;
494 for (i = 0; i < serv->sv_nrpools; i++) {
495 pool = &serv->sv_pools[i];
497 spin_lock_bh(&pool->sp_lock);
498 if (!list_empty(&pool->sp_threads)) {
499 rqstp = list_entry(pool->sp_threads.next,
502 dprintk("svc: daemon %p woken up.\n", rqstp);
504 svc_thread_dequeue(pool, rqstp);
505 rqstp->rq_xprt = NULL;
507 wake_up(&rqstp->rq_wait);
509 spin_unlock_bh(&pool->sp_lock);
512 EXPORT_SYMBOL_GPL(svc_wake_up);
514 int svc_port_is_privileged(struct sockaddr *sin)
516 switch (sin->sa_family) {
518 return ntohs(((struct sockaddr_in *)sin)->sin_port)
521 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
529 * Make sure that we don't have too many active connections. If we have,
530 * something must be dropped. It's not clear what will happen if we allow
531 * "too many" connections, but when dealing with network-facing software,
532 * we have to code defensively. Here we do that by imposing hard limits.
534 * There's no point in trying to do random drop here for DoS
535 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
536 * attacker can easily beat that.
538 * The only somewhat efficient mechanism would be if drop old
539 * connections from the same IP first. But right now we don't even
540 * record the client IP in svc_sock.
542 * single-threaded services that expect a lot of clients will probably
543 * need to set sv_maxconn to override the default value which is based
544 * on the number of threads
546 static void svc_check_conn_limits(struct svc_serv *serv)
548 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
549 (serv->sv_nrthreads+3) * 20;
551 if (serv->sv_tmpcnt > limit) {
552 struct svc_xprt *xprt = NULL;
553 spin_lock_bh(&serv->sv_lock);
554 if (!list_empty(&serv->sv_tempsocks)) {
555 if (net_ratelimit()) {
556 /* Try to help the admin */
557 printk(KERN_NOTICE "%s: too many open "
558 "connections, consider increasing %s\n",
559 serv->sv_name, serv->sv_maxconn ?
560 "the max number of connections." :
561 "the number of threads.");
564 * Always select the oldest connection. It's not fair,
567 xprt = list_entry(serv->sv_tempsocks.prev,
570 set_bit(XPT_CLOSE, &xprt->xpt_flags);
573 spin_unlock_bh(&serv->sv_lock);
576 svc_xprt_enqueue(xprt);
583 * Receive the next request on any transport. This code is carefully
584 * organised not to touch any cachelines in the shared svc_serv
585 * structure, only cachelines in the local svc_pool.
587 int svc_recv(struct svc_rqst *rqstp, long timeout)
589 struct svc_xprt *xprt = NULL;
590 struct svc_serv *serv = rqstp->rq_server;
591 struct svc_pool *pool = rqstp->rq_pool;
595 DECLARE_WAITQUEUE(wait, current);
598 dprintk("svc: server %p waiting for data (to = %ld)\n",
603 "svc_recv: service %p, transport not NULL!\n",
605 if (waitqueue_active(&rqstp->rq_wait))
607 "svc_recv: service %p, wait queue active!\n",
610 /* now allocate needed pages. If we get a failure, sleep briefly */
611 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
612 for (i = 0; i < pages ; i++)
613 while (rqstp->rq_pages[i] == NULL) {
614 struct page *p = alloc_page(GFP_KERNEL);
616 set_current_state(TASK_INTERRUPTIBLE);
617 if (signalled() || kthread_should_stop()) {
618 set_current_state(TASK_RUNNING);
621 schedule_timeout(msecs_to_jiffies(500));
623 rqstp->rq_pages[i] = p;
625 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
626 BUG_ON(pages >= RPCSVC_MAXPAGES);
628 /* Make arg->head point to first page and arg->pages point to rest */
629 arg = &rqstp->rq_arg;
630 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
631 arg->head[0].iov_len = PAGE_SIZE;
632 arg->pages = rqstp->rq_pages + 1;
634 /* save at least one page for response */
635 arg->page_len = (pages-2)*PAGE_SIZE;
636 arg->len = (pages-1)*PAGE_SIZE;
637 arg->tail[0].iov_len = 0;
641 if (signalled() || kthread_should_stop())
644 /* Normally we will wait up to 5 seconds for any required
645 * cache information to be provided.
647 rqstp->rq_chandle.thread_wait = 5*HZ;
649 spin_lock_bh(&pool->sp_lock);
650 xprt = svc_xprt_dequeue(pool);
652 rqstp->rq_xprt = xprt;
654 rqstp->rq_reserved = serv->sv_max_mesg;
655 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
657 /* As there is a shortage of threads and this request
658 * had to be queued, don't allow the thread to wait so
659 * long for cache updates.
661 rqstp->rq_chandle.thread_wait = 1*HZ;
663 /* No data pending. Go to sleep */
664 svc_thread_enqueue(pool, rqstp);
667 * We have to be able to interrupt this wait
668 * to bring down the daemons ...
670 set_current_state(TASK_INTERRUPTIBLE);
673 * checking kthread_should_stop() here allows us to avoid
674 * locking and signalling when stopping kthreads that call
675 * svc_recv. If the thread has already been woken up, then
676 * we can exit here without sleeping. If not, then it
677 * it'll be woken up quickly during the schedule_timeout
679 if (kthread_should_stop()) {
680 set_current_state(TASK_RUNNING);
681 spin_unlock_bh(&pool->sp_lock);
685 add_wait_queue(&rqstp->rq_wait, &wait);
686 spin_unlock_bh(&pool->sp_lock);
688 time_left = schedule_timeout(timeout);
692 spin_lock_bh(&pool->sp_lock);
693 remove_wait_queue(&rqstp->rq_wait, &wait);
695 pool->sp_stats.threads_timedout++;
697 xprt = rqstp->rq_xprt;
699 svc_thread_dequeue(pool, rqstp);
700 spin_unlock_bh(&pool->sp_lock);
701 dprintk("svc: server %p, no data yet\n", rqstp);
702 if (signalled() || kthread_should_stop())
708 spin_unlock_bh(&pool->sp_lock);
711 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
712 dprintk("svc_recv: found XPT_CLOSE\n");
713 svc_delete_xprt(xprt);
714 /* Leave XPT_BUSY set on the dead xprt: */
717 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
718 struct svc_xprt *newxpt;
719 newxpt = xprt->xpt_ops->xpo_accept(xprt);
722 * We know this module_get will succeed because the
723 * listener holds a reference too
725 __module_get(newxpt->xpt_class->xcl_owner);
726 svc_check_conn_limits(xprt->xpt_server);
727 spin_lock_bh(&serv->sv_lock);
728 set_bit(XPT_TEMP, &newxpt->xpt_flags);
729 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
731 if (serv->sv_temptimer.function == NULL) {
732 /* setup timer to age temp transports */
733 setup_timer(&serv->sv_temptimer,
735 (unsigned long)serv);
736 mod_timer(&serv->sv_temptimer,
737 jiffies + svc_conn_age_period * HZ);
739 spin_unlock_bh(&serv->sv_lock);
740 svc_xprt_received(newxpt);
742 } else if (xprt->xpt_ops->xpo_has_wspace(xprt)) {
743 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
744 rqstp, pool->sp_id, xprt,
745 atomic_read(&xprt->xpt_ref.refcount));
746 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
747 if (rqstp->rq_deferred)
748 len = svc_deferred_recv(rqstp);
750 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
751 dprintk("svc: got len=%d\n", len);
753 svc_xprt_received(xprt);
755 /* No data, incomplete (TCP) read, or accept() */
756 if (len == 0 || len == -EAGAIN)
759 clear_bit(XPT_OLD, &xprt->xpt_flags);
761 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
762 rqstp->rq_chandle.defer = svc_defer;
765 serv->sv_stats->netcnt++;
768 rqstp->rq_res.len = 0;
769 svc_xprt_release(rqstp);
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 rpc_wake_up(&xprt->xpt_bc_pending);
814 svc_xprt_release(rqstp);
816 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
822 * Timer function to close old temporary transports, using
823 * a mark-and-sweep algorithm.
825 static void svc_age_temp_xprts(unsigned long closure)
827 struct svc_serv *serv = (struct svc_serv *)closure;
828 struct svc_xprt *xprt;
829 struct list_head *le, *next;
830 LIST_HEAD(to_be_aged);
832 dprintk("svc_age_temp_xprts\n");
834 if (!spin_trylock_bh(&serv->sv_lock)) {
835 /* busy, try again 1 sec later */
836 dprintk("svc_age_temp_xprts: busy\n");
837 mod_timer(&serv->sv_temptimer, jiffies + HZ);
841 list_for_each_safe(le, next, &serv->sv_tempsocks) {
842 xprt = list_entry(le, struct svc_xprt, xpt_list);
844 /* First time through, just mark it OLD. Second time
845 * through, close it. */
846 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
848 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
849 test_bit(XPT_BUSY, &xprt->xpt_flags))
852 list_move(le, &to_be_aged);
853 set_bit(XPT_CLOSE, &xprt->xpt_flags);
854 set_bit(XPT_DETACHED, &xprt->xpt_flags);
856 spin_unlock_bh(&serv->sv_lock);
858 while (!list_empty(&to_be_aged)) {
859 le = to_be_aged.next;
860 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
862 xprt = list_entry(le, struct svc_xprt, xpt_list);
864 dprintk("queuing xprt %p for closing\n", xprt);
866 /* a thread will dequeue and close it soon */
867 svc_xprt_enqueue(xprt);
871 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
874 static void call_xpt_users(struct svc_xprt *xprt)
876 struct svc_xpt_user *u;
878 spin_lock(&xprt->xpt_lock);
879 while (!list_empty(&xprt->xpt_users)) {
880 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
884 spin_unlock(&xprt->xpt_lock);
888 * Remove a dead transport
890 void svc_delete_xprt(struct svc_xprt *xprt)
892 struct svc_serv *serv = xprt->xpt_server;
893 struct svc_deferred_req *dr;
895 /* Only do this once */
896 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
899 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
900 xprt->xpt_ops->xpo_detach(xprt);
902 spin_lock_bh(&serv->sv_lock);
903 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
904 list_del_init(&xprt->xpt_list);
906 * We used to delete the transport from whichever list
907 * it's sk_xprt.xpt_ready node was on, but we don't actually
908 * need to. This is because the only time we're called
909 * while still attached to a queue, the queue itself
910 * is about to be destroyed (in svc_destroy).
912 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
914 spin_unlock_bh(&serv->sv_lock);
916 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
919 call_xpt_users(xprt);
923 void svc_close_xprt(struct svc_xprt *xprt)
925 set_bit(XPT_CLOSE, &xprt->xpt_flags);
926 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
927 /* someone else will have to effect the close */
930 * We expect svc_close_xprt() to work even when no threads are
931 * running (e.g., while configuring the server before starting
932 * any threads), so if the transport isn't busy, we delete
935 svc_delete_xprt(xprt);
937 EXPORT_SYMBOL_GPL(svc_close_xprt);
939 void svc_close_all(struct list_head *xprt_list)
941 struct svc_xprt *xprt;
942 struct svc_xprt *tmp;
945 * The server is shutting down, and no more threads are running.
946 * svc_xprt_enqueue() might still be running, but at worst it
947 * will re-add the xprt to sp_sockets, which will soon get
948 * freed. So we don't bother with any more locking, and don't
949 * leave the close to the (nonexistent) server threads:
951 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
952 set_bit(XPT_CLOSE, &xprt->xpt_flags);
953 svc_delete_xprt(xprt);
958 * Handle defer and revisit of requests
961 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
963 struct svc_deferred_req *dr =
964 container_of(dreq, struct svc_deferred_req, handle);
965 struct svc_xprt *xprt = dr->xprt;
967 spin_lock(&xprt->xpt_lock);
968 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
969 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
970 spin_unlock(&xprt->xpt_lock);
971 dprintk("revisit canceled\n");
976 dprintk("revisit queued\n");
978 list_add(&dr->handle.recent, &xprt->xpt_deferred);
979 spin_unlock(&xprt->xpt_lock);
980 svc_xprt_enqueue(xprt);
985 * Save the request off for later processing. The request buffer looks
988 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
990 * This code can only handle requests that consist of an xprt-header
993 static struct cache_deferred_req *svc_defer(struct cache_req *req)
995 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
996 struct svc_deferred_req *dr;
998 if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
999 return NULL; /* if more than a page, give up FIXME */
1000 if (rqstp->rq_deferred) {
1001 dr = rqstp->rq_deferred;
1002 rqstp->rq_deferred = NULL;
1006 /* FIXME maybe discard if size too large */
1007 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1008 dr = kmalloc(size, GFP_KERNEL);
1012 dr->handle.owner = rqstp->rq_server;
1013 dr->prot = rqstp->rq_prot;
1014 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1015 dr->addrlen = rqstp->rq_addrlen;
1016 dr->daddr = rqstp->rq_daddr;
1017 dr->argslen = rqstp->rq_arg.len >> 2;
1018 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1020 /* back up head to the start of the buffer and copy */
1021 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1022 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1025 svc_xprt_get(rqstp->rq_xprt);
1026 dr->xprt = rqstp->rq_xprt;
1028 dr->handle.revisit = svc_revisit;
1033 * recv data from a deferred request into an active one
1035 static int svc_deferred_recv(struct svc_rqst *rqstp)
1037 struct svc_deferred_req *dr = rqstp->rq_deferred;
1039 /* setup iov_base past transport header */
1040 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1041 /* The iov_len does not include the transport header bytes */
1042 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1043 rqstp->rq_arg.page_len = 0;
1044 /* The rq_arg.len includes the transport header bytes */
1045 rqstp->rq_arg.len = dr->argslen<<2;
1046 rqstp->rq_prot = dr->prot;
1047 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1048 rqstp->rq_addrlen = dr->addrlen;
1049 /* Save off transport header len in case we get deferred again */
1050 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1051 rqstp->rq_daddr = dr->daddr;
1052 rqstp->rq_respages = rqstp->rq_pages;
1053 return (dr->argslen<<2) - dr->xprt_hlen;
1057 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1059 struct svc_deferred_req *dr = NULL;
1061 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1063 spin_lock(&xprt->xpt_lock);
1064 if (!list_empty(&xprt->xpt_deferred)) {
1065 dr = list_entry(xprt->xpt_deferred.next,
1066 struct svc_deferred_req,
1068 list_del_init(&dr->handle.recent);
1070 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1071 spin_unlock(&xprt->xpt_lock);
1076 * svc_find_xprt - find an RPC transport instance
1077 * @serv: pointer to svc_serv to search
1078 * @xcl_name: C string containing transport's class name
1079 * @af: Address family of transport's local address
1080 * @port: transport's IP port number
1082 * Return the transport instance pointer for the endpoint accepting
1083 * connections/peer traffic from the specified transport class,
1084 * address family and port.
1086 * Specifying 0 for the address family or port is effectively a
1087 * wild-card, and will result in matching the first transport in the
1088 * service's list that has a matching class name.
1090 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1091 const sa_family_t af, const unsigned short port)
1093 struct svc_xprt *xprt;
1094 struct svc_xprt *found = NULL;
1096 /* Sanity check the args */
1097 if (serv == NULL || xcl_name == NULL)
1100 spin_lock_bh(&serv->sv_lock);
1101 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1102 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1104 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1106 if (port != 0 && port != svc_xprt_local_port(xprt))
1112 spin_unlock_bh(&serv->sv_lock);
1115 EXPORT_SYMBOL_GPL(svc_find_xprt);
1117 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1118 char *pos, int remaining)
1122 len = snprintf(pos, remaining, "%s %u\n",
1123 xprt->xpt_class->xcl_name,
1124 svc_xprt_local_port(xprt));
1125 if (len >= remaining)
1126 return -ENAMETOOLONG;
1131 * svc_xprt_names - format a buffer with a list of transport names
1132 * @serv: pointer to an RPC service
1133 * @buf: pointer to a buffer to be filled in
1134 * @buflen: length of buffer to be filled in
1136 * Fills in @buf with a string containing a list of transport names,
1137 * each name terminated with '\n'.
1139 * Returns positive length of the filled-in string on success; otherwise
1140 * a negative errno value is returned if an error occurs.
1142 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1144 struct svc_xprt *xprt;
1148 /* Sanity check args */
1152 spin_lock_bh(&serv->sv_lock);
1156 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1157 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1169 spin_unlock_bh(&serv->sv_lock);
1172 EXPORT_SYMBOL_GPL(svc_xprt_names);
1175 /*----------------------------------------------------------------------------*/
1177 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1179 unsigned int pidx = (unsigned int)*pos;
1180 struct svc_serv *serv = m->private;
1182 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1185 return SEQ_START_TOKEN;
1186 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1189 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1191 struct svc_pool *pool = p;
1192 struct svc_serv *serv = m->private;
1194 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1196 if (p == SEQ_START_TOKEN) {
1197 pool = &serv->sv_pools[0];
1199 unsigned int pidx = (pool - &serv->sv_pools[0]);
1200 if (pidx < serv->sv_nrpools-1)
1201 pool = &serv->sv_pools[pidx+1];
1209 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1213 static int svc_pool_stats_show(struct seq_file *m, void *p)
1215 struct svc_pool *pool = p;
1217 if (p == SEQ_START_TOKEN) {
1218 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1222 seq_printf(m, "%u %lu %lu %lu %lu\n",
1224 pool->sp_stats.packets,
1225 pool->sp_stats.sockets_queued,
1226 pool->sp_stats.threads_woken,
1227 pool->sp_stats.threads_timedout);
1232 static const struct seq_operations svc_pool_stats_seq_ops = {
1233 .start = svc_pool_stats_start,
1234 .next = svc_pool_stats_next,
1235 .stop = svc_pool_stats_stop,
1236 .show = svc_pool_stats_show,
1239 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1243 err = seq_open(file, &svc_pool_stats_seq_ops);
1245 ((struct seq_file *) file->private_data)->private = serv;
1248 EXPORT_SYMBOL(svc_pool_stats_open);
1250 /*----------------------------------------------------------------------------*/