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;
216 if (strcmp(xprt_name, xcl->xcl_name))
219 if (!try_module_get(xcl->xcl_owner))
222 spin_unlock(&svc_xprt_class_lock);
223 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
224 if (IS_ERR(newxprt)) {
225 module_put(xcl->xcl_owner);
226 return PTR_ERR(newxprt);
229 clear_bit(XPT_TEMP, &newxprt->xpt_flags);
230 spin_lock_bh(&serv->sv_lock);
231 list_add(&newxprt->xpt_list, &serv->sv_permsocks);
232 spin_unlock_bh(&serv->sv_lock);
233 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
234 return svc_xprt_local_port(newxprt);
237 spin_unlock(&svc_xprt_class_lock);
238 dprintk("svc: transport %s not found\n", xprt_name);
240 /* This errno is exposed to user space. Provide a reasonable
241 * perror msg for a bad transport. */
242 return -EPROTONOSUPPORT;
244 EXPORT_SYMBOL_GPL(svc_create_xprt);
247 * Copy the local and remote xprt addresses to the rqstp structure
249 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
251 struct sockaddr *sin;
253 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
254 rqstp->rq_addrlen = xprt->xpt_remotelen;
257 * Destination address in request is needed for binding the
258 * source address in RPC replies/callbacks later.
260 sin = (struct sockaddr *)&xprt->xpt_local;
261 switch (sin->sa_family) {
263 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
266 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
270 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
273 * svc_print_addr - Format rq_addr field for printing
274 * @rqstp: svc_rqst struct containing address to print
275 * @buf: target buffer for formatted address
276 * @len: length of target buffer
279 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
281 return __svc_print_addr(svc_addr(rqstp), buf, len);
283 EXPORT_SYMBOL_GPL(svc_print_addr);
286 * Queue up an idle server thread. Must have pool->sp_lock held.
287 * Note: this is really a stack rather than a queue, so that we only
288 * use as many different threads as we need, and the rest don't pollute
291 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
293 list_add(&rqstp->rq_list, &pool->sp_threads);
297 * Dequeue an nfsd thread. Must have pool->sp_lock held.
299 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
301 list_del(&rqstp->rq_list);
305 * Queue up a transport with data pending. If there are idle nfsd
306 * processes, wake 'em up.
309 void svc_xprt_enqueue(struct svc_xprt *xprt)
311 struct svc_serv *serv = xprt->xpt_server;
312 struct svc_pool *pool;
313 struct svc_rqst *rqstp;
316 if (!(xprt->xpt_flags &
317 ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
321 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
324 spin_lock_bh(&pool->sp_lock);
326 if (!list_empty(&pool->sp_threads) &&
327 !list_empty(&pool->sp_sockets))
330 "threads and transports both waiting??\n");
332 pool->sp_stats.packets++;
334 /* Mark transport as busy. It will remain in this state until
335 * the provider calls svc_xprt_received. We update XPT_BUSY
336 * atomically because it also guards against trying to enqueue
337 * the transport twice.
339 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
340 /* Don't enqueue transport while already enqueued */
341 dprintk("svc: transport %p busy, not enqueued\n", xprt);
344 BUG_ON(xprt->xpt_pool != NULL);
345 xprt->xpt_pool = pool;
347 /* Handle pending connection */
348 if (test_bit(XPT_CONN, &xprt->xpt_flags))
351 /* Handle close in-progress */
352 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
355 /* Check if we have space to reply to a request */
356 if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
357 /* Don't enqueue while not enough space for reply */
358 dprintk("svc: no write space, transport %p not enqueued\n",
360 xprt->xpt_pool = NULL;
361 clear_bit(XPT_BUSY, &xprt->xpt_flags);
366 if (!list_empty(&pool->sp_threads)) {
367 rqstp = list_entry(pool->sp_threads.next,
370 dprintk("svc: transport %p served by daemon %p\n",
372 svc_thread_dequeue(pool, rqstp);
375 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
376 rqstp, rqstp->rq_xprt);
377 rqstp->rq_xprt = xprt;
379 rqstp->rq_reserved = serv->sv_max_mesg;
380 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
381 pool->sp_stats.threads_woken++;
382 BUG_ON(xprt->xpt_pool != pool);
383 wake_up(&rqstp->rq_wait);
385 dprintk("svc: transport %p put into queue\n", xprt);
386 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
387 pool->sp_stats.sockets_queued++;
388 BUG_ON(xprt->xpt_pool != pool);
392 spin_unlock_bh(&pool->sp_lock);
394 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
397 * Dequeue the first transport. Must be called with the pool->sp_lock held.
399 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
401 struct svc_xprt *xprt;
403 if (list_empty(&pool->sp_sockets))
406 xprt = list_entry(pool->sp_sockets.next,
407 struct svc_xprt, xpt_ready);
408 list_del_init(&xprt->xpt_ready);
410 dprintk("svc: transport %p dequeued, inuse=%d\n",
411 xprt, atomic_read(&xprt->xpt_ref.refcount));
417 * svc_xprt_received conditionally queues the transport for processing
418 * by another thread. The caller must hold the XPT_BUSY bit and must
419 * not thereafter touch transport data.
421 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
422 * insufficient) data.
424 void svc_xprt_received(struct svc_xprt *xprt)
426 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
427 xprt->xpt_pool = NULL;
428 clear_bit(XPT_BUSY, &xprt->xpt_flags);
429 svc_xprt_enqueue(xprt);
431 EXPORT_SYMBOL_GPL(svc_xprt_received);
434 * svc_reserve - change the space reserved for the reply to a request.
435 * @rqstp: The request in question
436 * @space: new max space to reserve
438 * Each request reserves some space on the output queue of the transport
439 * to make sure the reply fits. This function reduces that reserved
440 * space to be the amount of space used already, plus @space.
443 void svc_reserve(struct svc_rqst *rqstp, int space)
445 space += rqstp->rq_res.head[0].iov_len;
447 if (space < rqstp->rq_reserved) {
448 struct svc_xprt *xprt = rqstp->rq_xprt;
449 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
450 rqstp->rq_reserved = space;
452 svc_xprt_enqueue(xprt);
455 EXPORT_SYMBOL_GPL(svc_reserve);
457 static void svc_xprt_release(struct svc_rqst *rqstp)
459 struct svc_xprt *xprt = rqstp->rq_xprt;
461 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
463 kfree(rqstp->rq_deferred);
464 rqstp->rq_deferred = NULL;
466 svc_free_res_pages(rqstp);
467 rqstp->rq_res.page_len = 0;
468 rqstp->rq_res.page_base = 0;
470 /* Reset response buffer and release
472 * But first, check that enough space was reserved
473 * for the reply, otherwise we have a bug!
475 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
476 printk(KERN_ERR "RPC request reserved %d but used %d\n",
480 rqstp->rq_res.head[0].iov_len = 0;
481 svc_reserve(rqstp, 0);
482 rqstp->rq_xprt = NULL;
488 * External function to wake up a server waiting for data
489 * This really only makes sense for services like lockd
490 * which have exactly one thread anyway.
492 void svc_wake_up(struct svc_serv *serv)
494 struct svc_rqst *rqstp;
496 struct svc_pool *pool;
498 for (i = 0; i < serv->sv_nrpools; i++) {
499 pool = &serv->sv_pools[i];
501 spin_lock_bh(&pool->sp_lock);
502 if (!list_empty(&pool->sp_threads)) {
503 rqstp = list_entry(pool->sp_threads.next,
506 dprintk("svc: daemon %p woken up.\n", rqstp);
508 svc_thread_dequeue(pool, rqstp);
509 rqstp->rq_xprt = NULL;
511 wake_up(&rqstp->rq_wait);
513 spin_unlock_bh(&pool->sp_lock);
516 EXPORT_SYMBOL_GPL(svc_wake_up);
518 int svc_port_is_privileged(struct sockaddr *sin)
520 switch (sin->sa_family) {
522 return ntohs(((struct sockaddr_in *)sin)->sin_port)
525 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
533 * Make sure that we don't have too many active connections. If we have,
534 * something must be dropped. It's not clear what will happen if we allow
535 * "too many" connections, but when dealing with network-facing software,
536 * we have to code defensively. Here we do that by imposing hard limits.
538 * There's no point in trying to do random drop here for DoS
539 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
540 * attacker can easily beat that.
542 * The only somewhat efficient mechanism would be if drop old
543 * connections from the same IP first. But right now we don't even
544 * record the client IP in svc_sock.
546 * single-threaded services that expect a lot of clients will probably
547 * need to set sv_maxconn to override the default value which is based
548 * on the number of threads
550 static void svc_check_conn_limits(struct svc_serv *serv)
552 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
553 (serv->sv_nrthreads+3) * 20;
555 if (serv->sv_tmpcnt > limit) {
556 struct svc_xprt *xprt = NULL;
557 spin_lock_bh(&serv->sv_lock);
558 if (!list_empty(&serv->sv_tempsocks)) {
559 if (net_ratelimit()) {
560 /* Try to help the admin */
561 printk(KERN_NOTICE "%s: too many open "
562 "connections, consider increasing %s\n",
563 serv->sv_name, serv->sv_maxconn ?
564 "the max number of connections." :
565 "the number of threads.");
568 * Always select the oldest connection. It's not fair,
571 xprt = list_entry(serv->sv_tempsocks.prev,
574 set_bit(XPT_CLOSE, &xprt->xpt_flags);
577 spin_unlock_bh(&serv->sv_lock);
580 svc_xprt_enqueue(xprt);
587 * Receive the next request on any transport. This code is carefully
588 * organised not to touch any cachelines in the shared svc_serv
589 * structure, only cachelines in the local svc_pool.
591 int svc_recv(struct svc_rqst *rqstp, long timeout)
593 struct svc_xprt *xprt = NULL;
594 struct svc_serv *serv = rqstp->rq_server;
595 struct svc_pool *pool = rqstp->rq_pool;
599 DECLARE_WAITQUEUE(wait, current);
602 dprintk("svc: server %p waiting for data (to = %ld)\n",
607 "svc_recv: service %p, transport not NULL!\n",
609 if (waitqueue_active(&rqstp->rq_wait))
611 "svc_recv: service %p, wait queue active!\n",
614 /* now allocate needed pages. If we get a failure, sleep briefly */
615 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
616 for (i = 0; i < pages ; i++)
617 while (rqstp->rq_pages[i] == NULL) {
618 struct page *p = alloc_page(GFP_KERNEL);
620 set_current_state(TASK_INTERRUPTIBLE);
621 if (signalled() || kthread_should_stop()) {
622 set_current_state(TASK_RUNNING);
625 schedule_timeout(msecs_to_jiffies(500));
627 rqstp->rq_pages[i] = p;
629 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
630 BUG_ON(pages >= RPCSVC_MAXPAGES);
632 /* Make arg->head point to first page and arg->pages point to rest */
633 arg = &rqstp->rq_arg;
634 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
635 arg->head[0].iov_len = PAGE_SIZE;
636 arg->pages = rqstp->rq_pages + 1;
638 /* save at least one page for response */
639 arg->page_len = (pages-2)*PAGE_SIZE;
640 arg->len = (pages-1)*PAGE_SIZE;
641 arg->tail[0].iov_len = 0;
645 if (signalled() || kthread_should_stop())
648 /* Normally we will wait up to 5 seconds for any required
649 * cache information to be provided.
651 rqstp->rq_chandle.thread_wait = 5*HZ;
653 spin_lock_bh(&pool->sp_lock);
654 xprt = svc_xprt_dequeue(pool);
656 rqstp->rq_xprt = xprt;
658 rqstp->rq_reserved = serv->sv_max_mesg;
659 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
661 /* As there is a shortage of threads and this request
662 * had to be queued, don't allow the thread to wait so
663 * long for cache updates.
665 rqstp->rq_chandle.thread_wait = 1*HZ;
667 /* No data pending. Go to sleep */
668 svc_thread_enqueue(pool, rqstp);
671 * We have to be able to interrupt this wait
672 * to bring down the daemons ...
674 set_current_state(TASK_INTERRUPTIBLE);
677 * checking kthread_should_stop() here allows us to avoid
678 * locking and signalling when stopping kthreads that call
679 * svc_recv. If the thread has already been woken up, then
680 * we can exit here without sleeping. If not, then it
681 * it'll be woken up quickly during the schedule_timeout
683 if (kthread_should_stop()) {
684 set_current_state(TASK_RUNNING);
685 spin_unlock_bh(&pool->sp_lock);
689 add_wait_queue(&rqstp->rq_wait, &wait);
690 spin_unlock_bh(&pool->sp_lock);
692 time_left = schedule_timeout(timeout);
696 spin_lock_bh(&pool->sp_lock);
697 remove_wait_queue(&rqstp->rq_wait, &wait);
699 pool->sp_stats.threads_timedout++;
701 xprt = rqstp->rq_xprt;
703 svc_thread_dequeue(pool, rqstp);
704 spin_unlock_bh(&pool->sp_lock);
705 dprintk("svc: server %p, no data yet\n", rqstp);
706 if (signalled() || kthread_should_stop())
712 spin_unlock_bh(&pool->sp_lock);
715 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
716 dprintk("svc_recv: found XPT_CLOSE\n");
717 svc_delete_xprt(xprt);
718 } else if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
719 struct svc_xprt *newxpt;
720 newxpt = xprt->xpt_ops->xpo_accept(xprt);
723 * We know this module_get will succeed because the
724 * listener holds a reference too
726 __module_get(newxpt->xpt_class->xcl_owner);
727 svc_check_conn_limits(xprt->xpt_server);
728 spin_lock_bh(&serv->sv_lock);
729 set_bit(XPT_TEMP, &newxpt->xpt_flags);
730 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
732 if (serv->sv_temptimer.function == NULL) {
733 /* setup timer to age temp transports */
734 setup_timer(&serv->sv_temptimer,
736 (unsigned long)serv);
737 mod_timer(&serv->sv_temptimer,
738 jiffies + svc_conn_age_period * HZ);
740 spin_unlock_bh(&serv->sv_lock);
741 svc_xprt_received(newxpt);
743 svc_xprt_received(xprt);
745 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
746 rqstp, pool->sp_id, xprt,
747 atomic_read(&xprt->xpt_ref.refcount));
748 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
749 if (rqstp->rq_deferred) {
750 svc_xprt_received(xprt);
751 len = svc_deferred_recv(rqstp);
753 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
754 svc_xprt_received(xprt);
756 dprintk("svc: got len=%d\n", len);
759 /* No data, incomplete (TCP) read, or accept() */
760 if (len == 0 || len == -EAGAIN) {
761 rqstp->rq_res.len = 0;
762 svc_xprt_release(rqstp);
765 clear_bit(XPT_OLD, &xprt->xpt_flags);
767 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
768 rqstp->rq_chandle.defer = svc_defer;
771 serv->sv_stats->netcnt++;
774 EXPORT_SYMBOL_GPL(svc_recv);
779 void svc_drop(struct svc_rqst *rqstp)
781 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
782 svc_xprt_release(rqstp);
784 EXPORT_SYMBOL_GPL(svc_drop);
787 * Return reply to client.
789 int svc_send(struct svc_rqst *rqstp)
791 struct svc_xprt *xprt;
795 xprt = rqstp->rq_xprt;
799 /* release the receive skb before sending the reply */
800 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
802 /* calculate over-all length */
804 xb->len = xb->head[0].iov_len +
808 /* Grab mutex to serialize outgoing data. */
809 mutex_lock(&xprt->xpt_mutex);
810 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
813 len = xprt->xpt_ops->xpo_sendto(rqstp);
814 mutex_unlock(&xprt->xpt_mutex);
815 rpc_wake_up(&xprt->xpt_bc_pending);
816 svc_xprt_release(rqstp);
818 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
824 * Timer function to close old temporary transports, using
825 * a mark-and-sweep algorithm.
827 static void svc_age_temp_xprts(unsigned long closure)
829 struct svc_serv *serv = (struct svc_serv *)closure;
830 struct svc_xprt *xprt;
831 struct list_head *le, *next;
832 LIST_HEAD(to_be_aged);
834 dprintk("svc_age_temp_xprts\n");
836 if (!spin_trylock_bh(&serv->sv_lock)) {
837 /* busy, try again 1 sec later */
838 dprintk("svc_age_temp_xprts: busy\n");
839 mod_timer(&serv->sv_temptimer, jiffies + HZ);
843 list_for_each_safe(le, next, &serv->sv_tempsocks) {
844 xprt = list_entry(le, struct svc_xprt, xpt_list);
846 /* First time through, just mark it OLD. Second time
847 * through, close it. */
848 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
850 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
851 test_bit(XPT_BUSY, &xprt->xpt_flags))
854 list_move(le, &to_be_aged);
855 set_bit(XPT_CLOSE, &xprt->xpt_flags);
856 set_bit(XPT_DETACHED, &xprt->xpt_flags);
858 spin_unlock_bh(&serv->sv_lock);
860 while (!list_empty(&to_be_aged)) {
861 le = to_be_aged.next;
862 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
864 xprt = list_entry(le, struct svc_xprt, xpt_list);
866 dprintk("queuing xprt %p for closing\n", xprt);
868 /* a thread will dequeue and close it soon */
869 svc_xprt_enqueue(xprt);
873 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
876 static void call_xpt_users(struct svc_xprt *xprt)
878 struct svc_xpt_user *u;
880 spin_lock(&xprt->xpt_lock);
881 while (!list_empty(&xprt->xpt_users)) {
882 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
886 spin_unlock(&xprt->xpt_lock);
890 * Remove a dead transport
892 void svc_delete_xprt(struct svc_xprt *xprt)
894 struct svc_serv *serv = xprt->xpt_server;
895 struct svc_deferred_req *dr;
897 /* Only do this once */
898 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
901 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
902 xprt->xpt_ops->xpo_detach(xprt);
904 spin_lock_bh(&serv->sv_lock);
905 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
906 list_del_init(&xprt->xpt_list);
908 * We used to delete the transport from whichever list
909 * it's sk_xprt.xpt_ready node was on, but we don't actually
910 * need to. This is because the only time we're called
911 * while still attached to a queue, the queue itself
912 * is about to be destroyed (in svc_destroy).
914 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
916 spin_unlock_bh(&serv->sv_lock);
918 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
921 call_xpt_users(xprt);
925 void svc_close_xprt(struct svc_xprt *xprt)
927 set_bit(XPT_CLOSE, &xprt->xpt_flags);
928 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
929 /* someone else will have to effect the close */
932 svc_delete_xprt(xprt);
934 EXPORT_SYMBOL_GPL(svc_close_xprt);
936 void svc_close_all(struct list_head *xprt_list)
938 struct svc_xprt *xprt;
939 struct svc_xprt *tmp;
941 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
942 set_bit(XPT_CLOSE, &xprt->xpt_flags);
943 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
944 /* Waiting to be processed, but no threads left,
945 * So just remove it from the waiting list
947 list_del_init(&xprt->xpt_ready);
948 clear_bit(XPT_BUSY, &xprt->xpt_flags);
950 svc_close_xprt(xprt);
955 * Handle defer and revisit of requests
958 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
960 struct svc_deferred_req *dr =
961 container_of(dreq, struct svc_deferred_req, handle);
962 struct svc_xprt *xprt = dr->xprt;
964 spin_lock(&xprt->xpt_lock);
965 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
966 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
967 spin_unlock(&xprt->xpt_lock);
968 dprintk("revisit canceled\n");
973 dprintk("revisit queued\n");
975 list_add(&dr->handle.recent, &xprt->xpt_deferred);
976 spin_unlock(&xprt->xpt_lock);
977 svc_xprt_enqueue(xprt);
982 * Save the request off for later processing. The request buffer looks
985 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
987 * This code can only handle requests that consist of an xprt-header
990 static struct cache_deferred_req *svc_defer(struct cache_req *req)
992 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
993 struct svc_deferred_req *dr;
995 if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
996 return NULL; /* if more than a page, give up FIXME */
997 if (rqstp->rq_deferred) {
998 dr = rqstp->rq_deferred;
999 rqstp->rq_deferred = NULL;
1003 /* FIXME maybe discard if size too large */
1004 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1005 dr = kmalloc(size, GFP_KERNEL);
1009 dr->handle.owner = rqstp->rq_server;
1010 dr->prot = rqstp->rq_prot;
1011 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1012 dr->addrlen = rqstp->rq_addrlen;
1013 dr->daddr = rqstp->rq_daddr;
1014 dr->argslen = rqstp->rq_arg.len >> 2;
1015 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1017 /* back up head to the start of the buffer and copy */
1018 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1019 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1022 svc_xprt_get(rqstp->rq_xprt);
1023 dr->xprt = rqstp->rq_xprt;
1025 dr->handle.revisit = svc_revisit;
1030 * recv data from a deferred request into an active one
1032 static int svc_deferred_recv(struct svc_rqst *rqstp)
1034 struct svc_deferred_req *dr = rqstp->rq_deferred;
1036 /* setup iov_base past transport header */
1037 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1038 /* The iov_len does not include the transport header bytes */
1039 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1040 rqstp->rq_arg.page_len = 0;
1041 /* The rq_arg.len includes the transport header bytes */
1042 rqstp->rq_arg.len = dr->argslen<<2;
1043 rqstp->rq_prot = dr->prot;
1044 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1045 rqstp->rq_addrlen = dr->addrlen;
1046 /* Save off transport header len in case we get deferred again */
1047 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1048 rqstp->rq_daddr = dr->daddr;
1049 rqstp->rq_respages = rqstp->rq_pages;
1050 return (dr->argslen<<2) - dr->xprt_hlen;
1054 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1056 struct svc_deferred_req *dr = NULL;
1058 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1060 spin_lock(&xprt->xpt_lock);
1061 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1062 if (!list_empty(&xprt->xpt_deferred)) {
1063 dr = list_entry(xprt->xpt_deferred.next,
1064 struct svc_deferred_req,
1066 list_del_init(&dr->handle.recent);
1067 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1069 spin_unlock(&xprt->xpt_lock);
1074 * svc_find_xprt - find an RPC transport instance
1075 * @serv: pointer to svc_serv to search
1076 * @xcl_name: C string containing transport's class name
1077 * @af: Address family of transport's local address
1078 * @port: transport's IP port number
1080 * Return the transport instance pointer for the endpoint accepting
1081 * connections/peer traffic from the specified transport class,
1082 * address family and port.
1084 * Specifying 0 for the address family or port is effectively a
1085 * wild-card, and will result in matching the first transport in the
1086 * service's list that has a matching class name.
1088 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1089 const sa_family_t af, const unsigned short port)
1091 struct svc_xprt *xprt;
1092 struct svc_xprt *found = NULL;
1094 /* Sanity check the args */
1095 if (serv == NULL || xcl_name == NULL)
1098 spin_lock_bh(&serv->sv_lock);
1099 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1100 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1102 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1104 if (port != 0 && port != svc_xprt_local_port(xprt))
1110 spin_unlock_bh(&serv->sv_lock);
1113 EXPORT_SYMBOL_GPL(svc_find_xprt);
1115 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1116 char *pos, int remaining)
1120 len = snprintf(pos, remaining, "%s %u\n",
1121 xprt->xpt_class->xcl_name,
1122 svc_xprt_local_port(xprt));
1123 if (len >= remaining)
1124 return -ENAMETOOLONG;
1129 * svc_xprt_names - format a buffer with a list of transport names
1130 * @serv: pointer to an RPC service
1131 * @buf: pointer to a buffer to be filled in
1132 * @buflen: length of buffer to be filled in
1134 * Fills in @buf with a string containing a list of transport names,
1135 * each name terminated with '\n'.
1137 * Returns positive length of the filled-in string on success; otherwise
1138 * a negative errno value is returned if an error occurs.
1140 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1142 struct svc_xprt *xprt;
1146 /* Sanity check args */
1150 spin_lock_bh(&serv->sv_lock);
1154 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1155 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1167 spin_unlock_bh(&serv->sv_lock);
1170 EXPORT_SYMBOL_GPL(svc_xprt_names);
1173 /*----------------------------------------------------------------------------*/
1175 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1177 unsigned int pidx = (unsigned int)*pos;
1178 struct svc_serv *serv = m->private;
1180 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1183 return SEQ_START_TOKEN;
1184 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1187 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1189 struct svc_pool *pool = p;
1190 struct svc_serv *serv = m->private;
1192 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1194 if (p == SEQ_START_TOKEN) {
1195 pool = &serv->sv_pools[0];
1197 unsigned int pidx = (pool - &serv->sv_pools[0]);
1198 if (pidx < serv->sv_nrpools-1)
1199 pool = &serv->sv_pools[pidx+1];
1207 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1211 static int svc_pool_stats_show(struct seq_file *m, void *p)
1213 struct svc_pool *pool = p;
1215 if (p == SEQ_START_TOKEN) {
1216 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1220 seq_printf(m, "%u %lu %lu %lu %lu\n",
1222 pool->sp_stats.packets,
1223 pool->sp_stats.sockets_queued,
1224 pool->sp_stats.threads_woken,
1225 pool->sp_stats.threads_timedout);
1230 static const struct seq_operations svc_pool_stats_seq_ops = {
1231 .start = svc_pool_stats_start,
1232 .next = svc_pool_stats_next,
1233 .stop = svc_pool_stats_stop,
1234 .show = svc_pool_stats_show,
1237 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1241 err = seq_open(file, &svc_pool_stats_seq_ops);
1243 ((struct seq_file *) file->private_data)->private = serv;
1246 EXPORT_SYMBOL(svc_pool_stats_open);
1248 /*----------------------------------------------------------------------------*/