2 * linux/net/sunrpc/svc_xprt.c
4 * Author: Tom Tucker <tom@opengridcomputing.com>
7 #include <linux/sched.h>
8 #include <linux/smp_lock.h>
9 #include <linux/errno.h>
10 #include <linux/freezer.h>
11 #include <linux/kthread.h>
12 #include <linux/slab.h>
14 #include <linux/sunrpc/stats.h>
15 #include <linux/sunrpc/svc_xprt.h>
16 #include <linux/sunrpc/svcsock.h>
18 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
20 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
21 static int svc_deferred_recv(struct svc_rqst *rqstp);
22 static struct cache_deferred_req *svc_defer(struct cache_req *req);
23 static void svc_age_temp_xprts(unsigned long closure);
25 /* apparently the "standard" is that clients close
26 * idle connections after 5 minutes, servers after
28 * http://www.connectathon.org/talks96/nfstcp.pdf
30 static int svc_conn_age_period = 6*60;
32 /* List of registered transport classes */
33 static DEFINE_SPINLOCK(svc_xprt_class_lock);
34 static LIST_HEAD(svc_xprt_class_list);
36 /* SMP locking strategy:
38 * svc_pool->sp_lock protects most of the fields of that pool.
39 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
40 * when both need to be taken (rare), svc_serv->sv_lock is first.
41 * BKL protects svc_serv->sv_nrthread.
42 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
43 * and the ->sk_info_authunix cache.
45 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
46 * enqueued multiply. During normal transport processing this bit
47 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
48 * Providers should not manipulate this bit directly.
50 * Some flags can be set to certain values at any time
51 * providing that certain rules are followed:
54 * - Can be set or cleared at any time.
55 * - After a set, svc_xprt_enqueue must be called to enqueue
56 * the transport for processing.
57 * - After a clear, the transport must be read/accepted.
58 * If this succeeds, it must be set again.
60 * - Can set at any time. It is never cleared.
62 * - Can only be set while XPT_BUSY is held which ensures
63 * that no other thread will be using the transport or will
64 * try to set XPT_DEAD.
67 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
69 struct svc_xprt_class *cl;
72 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
74 INIT_LIST_HEAD(&xcl->xcl_list);
75 spin_lock(&svc_xprt_class_lock);
76 /* Make sure there isn't already a class with the same name */
77 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
78 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
81 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
84 spin_unlock(&svc_xprt_class_lock);
87 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
89 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
91 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
92 spin_lock(&svc_xprt_class_lock);
93 list_del_init(&xcl->xcl_list);
94 spin_unlock(&svc_xprt_class_lock);
96 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
99 * Format the transport list for printing
101 int svc_print_xprts(char *buf, int maxlen)
103 struct svc_xprt_class *xcl;
108 spin_lock(&svc_xprt_class_lock);
109 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
112 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
113 slen = strlen(tmpstr);
114 if (len + slen > maxlen)
119 spin_unlock(&svc_xprt_class_lock);
124 static void svc_xprt_free(struct kref *kref)
126 struct svc_xprt *xprt =
127 container_of(kref, struct svc_xprt, xpt_ref);
128 struct module *owner = xprt->xpt_class->xcl_owner;
129 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
130 svcauth_unix_info_release(xprt);
131 put_net(xprt->xpt_net);
132 xprt->xpt_ops->xpo_free(xprt);
136 void svc_xprt_put(struct svc_xprt *xprt)
138 kref_put(&xprt->xpt_ref, svc_xprt_free);
140 EXPORT_SYMBOL_GPL(svc_xprt_put);
143 * Called by transport drivers to initialize the transport independent
144 * portion of the transport instance.
146 void svc_xprt_init(struct svc_xprt_class *xcl, struct svc_xprt *xprt,
147 struct svc_serv *serv)
149 memset(xprt, 0, sizeof(*xprt));
150 xprt->xpt_class = xcl;
151 xprt->xpt_ops = xcl->xcl_ops;
152 kref_init(&xprt->xpt_ref);
153 xprt->xpt_server = serv;
154 INIT_LIST_HEAD(&xprt->xpt_list);
155 INIT_LIST_HEAD(&xprt->xpt_ready);
156 INIT_LIST_HEAD(&xprt->xpt_deferred);
157 INIT_LIST_HEAD(&xprt->xpt_users);
158 mutex_init(&xprt->xpt_mutex);
159 spin_lock_init(&xprt->xpt_lock);
160 set_bit(XPT_BUSY, &xprt->xpt_flags);
161 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
162 xprt->xpt_net = get_net(&init_net);
164 EXPORT_SYMBOL_GPL(svc_xprt_init);
166 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
167 struct svc_serv *serv,
170 const unsigned short port,
173 struct sockaddr_in sin = {
174 .sin_family = AF_INET,
175 .sin_addr.s_addr = htonl(INADDR_ANY),
176 .sin_port = htons(port),
178 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
179 struct sockaddr_in6 sin6 = {
180 .sin6_family = AF_INET6,
181 .sin6_addr = IN6ADDR_ANY_INIT,
182 .sin6_port = htons(port),
184 #endif /* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
185 struct sockaddr *sap;
190 sap = (struct sockaddr *)&sin;
193 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
195 sap = (struct sockaddr *)&sin6;
198 #endif /* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
200 return ERR_PTR(-EAFNOSUPPORT);
203 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
206 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
207 struct net *net, const int family,
208 const unsigned short port, int flags)
210 struct svc_xprt_class *xcl;
212 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
213 spin_lock(&svc_xprt_class_lock);
214 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
215 struct svc_xprt *newxprt;
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 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
235 return svc_xprt_local_port(newxprt);
238 spin_unlock(&svc_xprt_class_lock);
239 dprintk("svc: transport %s not found\n", xprt_name);
241 /* This errno is exposed to user space. Provide a reasonable
242 * perror msg for a bad transport. */
243 return -EPROTONOSUPPORT;
245 EXPORT_SYMBOL_GPL(svc_create_xprt);
248 * Copy the local and remote xprt addresses to the rqstp structure
250 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
252 struct sockaddr *sin;
254 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
255 rqstp->rq_addrlen = xprt->xpt_remotelen;
258 * Destination address in request is needed for binding the
259 * source address in RPC replies/callbacks later.
261 sin = (struct sockaddr *)&xprt->xpt_local;
262 switch (sin->sa_family) {
264 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
267 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
271 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
274 * svc_print_addr - Format rq_addr field for printing
275 * @rqstp: svc_rqst struct containing address to print
276 * @buf: target buffer for formatted address
277 * @len: length of target buffer
280 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
282 return __svc_print_addr(svc_addr(rqstp), buf, len);
284 EXPORT_SYMBOL_GPL(svc_print_addr);
287 * Queue up an idle server thread. Must have pool->sp_lock held.
288 * Note: this is really a stack rather than a queue, so that we only
289 * use as many different threads as we need, and the rest don't pollute
292 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
294 list_add(&rqstp->rq_list, &pool->sp_threads);
298 * Dequeue an nfsd thread. Must have pool->sp_lock held.
300 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
302 list_del(&rqstp->rq_list);
305 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
307 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
309 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
310 return xprt->xpt_ops->xpo_has_wspace(xprt);
315 * Queue up a transport with data pending. If there are idle nfsd
316 * processes, wake 'em up.
319 void svc_xprt_enqueue(struct svc_xprt *xprt)
321 struct svc_serv *serv = xprt->xpt_server;
322 struct svc_pool *pool;
323 struct svc_rqst *rqstp;
326 if (!svc_xprt_has_something_to_do(xprt))
330 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
333 spin_lock_bh(&pool->sp_lock);
335 if (!list_empty(&pool->sp_threads) &&
336 !list_empty(&pool->sp_sockets))
339 "threads and transports both waiting??\n");
341 pool->sp_stats.packets++;
343 /* Mark transport as busy. It will remain in this state until
344 * the provider calls svc_xprt_received. We update XPT_BUSY
345 * atomically because it also guards against trying to enqueue
346 * the transport twice.
348 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
349 /* Don't enqueue transport while already enqueued */
350 dprintk("svc: transport %p busy, not enqueued\n", xprt);
353 BUG_ON(xprt->xpt_pool != NULL);
354 xprt->xpt_pool = pool;
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 pool->sp_stats.threads_woken++;
372 BUG_ON(xprt->xpt_pool != pool);
373 wake_up(&rqstp->rq_wait);
375 dprintk("svc: transport %p put into queue\n", xprt);
376 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
377 pool->sp_stats.sockets_queued++;
378 BUG_ON(xprt->xpt_pool != pool);
382 spin_unlock_bh(&pool->sp_lock);
384 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
387 * Dequeue the first transport. Must be called with the pool->sp_lock held.
389 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
391 struct svc_xprt *xprt;
393 if (list_empty(&pool->sp_sockets))
396 xprt = list_entry(pool->sp_sockets.next,
397 struct svc_xprt, xpt_ready);
398 list_del_init(&xprt->xpt_ready);
400 dprintk("svc: transport %p dequeued, inuse=%d\n",
401 xprt, atomic_read(&xprt->xpt_ref.refcount));
407 * svc_xprt_received conditionally queues the transport for processing
408 * by another thread. The caller must hold the XPT_BUSY bit and must
409 * not thereafter touch transport data.
411 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
412 * insufficient) data.
414 void svc_xprt_received(struct svc_xprt *xprt)
416 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
417 xprt->xpt_pool = NULL;
418 clear_bit(XPT_BUSY, &xprt->xpt_flags);
419 svc_xprt_enqueue(xprt);
421 EXPORT_SYMBOL_GPL(svc_xprt_received);
424 * svc_reserve - change the space reserved for the reply to a request.
425 * @rqstp: The request in question
426 * @space: new max space to reserve
428 * Each request reserves some space on the output queue of the transport
429 * to make sure the reply fits. This function reduces that reserved
430 * space to be the amount of space used already, plus @space.
433 void svc_reserve(struct svc_rqst *rqstp, int space)
435 space += rqstp->rq_res.head[0].iov_len;
437 if (space < rqstp->rq_reserved) {
438 struct svc_xprt *xprt = rqstp->rq_xprt;
439 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
440 rqstp->rq_reserved = space;
442 svc_xprt_enqueue(xprt);
445 EXPORT_SYMBOL_GPL(svc_reserve);
447 static void svc_xprt_release(struct svc_rqst *rqstp)
449 struct svc_xprt *xprt = rqstp->rq_xprt;
451 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
453 kfree(rqstp->rq_deferred);
454 rqstp->rq_deferred = NULL;
456 svc_free_res_pages(rqstp);
457 rqstp->rq_res.page_len = 0;
458 rqstp->rq_res.page_base = 0;
460 /* Reset response buffer and release
462 * But first, check that enough space was reserved
463 * for the reply, otherwise we have a bug!
465 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
466 printk(KERN_ERR "RPC request reserved %d but used %d\n",
470 rqstp->rq_res.head[0].iov_len = 0;
471 svc_reserve(rqstp, 0);
472 rqstp->rq_xprt = NULL;
478 * External function to wake up a server waiting for data
479 * This really only makes sense for services like lockd
480 * which have exactly one thread anyway.
482 void svc_wake_up(struct svc_serv *serv)
484 struct svc_rqst *rqstp;
486 struct svc_pool *pool;
488 for (i = 0; i < serv->sv_nrpools; i++) {
489 pool = &serv->sv_pools[i];
491 spin_lock_bh(&pool->sp_lock);
492 if (!list_empty(&pool->sp_threads)) {
493 rqstp = list_entry(pool->sp_threads.next,
496 dprintk("svc: daemon %p woken up.\n", rqstp);
498 svc_thread_dequeue(pool, rqstp);
499 rqstp->rq_xprt = NULL;
501 wake_up(&rqstp->rq_wait);
503 spin_unlock_bh(&pool->sp_lock);
506 EXPORT_SYMBOL_GPL(svc_wake_up);
508 int svc_port_is_privileged(struct sockaddr *sin)
510 switch (sin->sa_family) {
512 return ntohs(((struct sockaddr_in *)sin)->sin_port)
515 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
523 * Make sure that we don't have too many active connections. If we have,
524 * something must be dropped. It's not clear what will happen if we allow
525 * "too many" connections, but when dealing with network-facing software,
526 * we have to code defensively. Here we do that by imposing hard limits.
528 * There's no point in trying to do random drop here for DoS
529 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
530 * attacker can easily beat that.
532 * The only somewhat efficient mechanism would be if drop old
533 * connections from the same IP first. But right now we don't even
534 * record the client IP in svc_sock.
536 * single-threaded services that expect a lot of clients will probably
537 * need to set sv_maxconn to override the default value which is based
538 * on the number of threads
540 static void svc_check_conn_limits(struct svc_serv *serv)
542 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
543 (serv->sv_nrthreads+3) * 20;
545 if (serv->sv_tmpcnt > limit) {
546 struct svc_xprt *xprt = NULL;
547 spin_lock_bh(&serv->sv_lock);
548 if (!list_empty(&serv->sv_tempsocks)) {
549 if (net_ratelimit()) {
550 /* Try to help the admin */
551 printk(KERN_NOTICE "%s: too many open "
552 "connections, consider increasing %s\n",
553 serv->sv_name, serv->sv_maxconn ?
554 "the max number of connections." :
555 "the number of threads.");
558 * Always select the oldest connection. It's not fair,
561 xprt = list_entry(serv->sv_tempsocks.prev,
564 set_bit(XPT_CLOSE, &xprt->xpt_flags);
567 spin_unlock_bh(&serv->sv_lock);
570 svc_xprt_enqueue(xprt);
577 * Receive the next request on any transport. This code is carefully
578 * organised not to touch any cachelines in the shared svc_serv
579 * structure, only cachelines in the local svc_pool.
581 int svc_recv(struct svc_rqst *rqstp, long timeout)
583 struct svc_xprt *xprt = NULL;
584 struct svc_serv *serv = rqstp->rq_server;
585 struct svc_pool *pool = rqstp->rq_pool;
589 DECLARE_WAITQUEUE(wait, current);
592 dprintk("svc: server %p waiting for data (to = %ld)\n",
597 "svc_recv: service %p, transport not NULL!\n",
599 if (waitqueue_active(&rqstp->rq_wait))
601 "svc_recv: service %p, wait queue active!\n",
604 /* now allocate needed pages. If we get a failure, sleep briefly */
605 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
606 for (i = 0; i < pages ; i++)
607 while (rqstp->rq_pages[i] == NULL) {
608 struct page *p = alloc_page(GFP_KERNEL);
610 set_current_state(TASK_INTERRUPTIBLE);
611 if (signalled() || kthread_should_stop()) {
612 set_current_state(TASK_RUNNING);
615 schedule_timeout(msecs_to_jiffies(500));
617 rqstp->rq_pages[i] = p;
619 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
620 BUG_ON(pages >= RPCSVC_MAXPAGES);
622 /* Make arg->head point to first page and arg->pages point to rest */
623 arg = &rqstp->rq_arg;
624 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
625 arg->head[0].iov_len = PAGE_SIZE;
626 arg->pages = rqstp->rq_pages + 1;
628 /* save at least one page for response */
629 arg->page_len = (pages-2)*PAGE_SIZE;
630 arg->len = (pages-1)*PAGE_SIZE;
631 arg->tail[0].iov_len = 0;
635 if (signalled() || kthread_should_stop())
638 /* Normally we will wait up to 5 seconds for any required
639 * cache information to be provided.
641 rqstp->rq_chandle.thread_wait = 5*HZ;
643 spin_lock_bh(&pool->sp_lock);
644 xprt = svc_xprt_dequeue(pool);
646 rqstp->rq_xprt = xprt;
648 rqstp->rq_reserved = serv->sv_max_mesg;
649 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
651 /* As there is a shortage of threads and this request
652 * had to be queued, don't allow the thread to wait so
653 * long for cache updates.
655 rqstp->rq_chandle.thread_wait = 1*HZ;
657 /* No data pending. Go to sleep */
658 svc_thread_enqueue(pool, rqstp);
661 * We have to be able to interrupt this wait
662 * to bring down the daemons ...
664 set_current_state(TASK_INTERRUPTIBLE);
667 * checking kthread_should_stop() here allows us to avoid
668 * locking and signalling when stopping kthreads that call
669 * svc_recv. If the thread has already been woken up, then
670 * we can exit here without sleeping. If not, then it
671 * it'll be woken up quickly during the schedule_timeout
673 if (kthread_should_stop()) {
674 set_current_state(TASK_RUNNING);
675 spin_unlock_bh(&pool->sp_lock);
679 add_wait_queue(&rqstp->rq_wait, &wait);
680 spin_unlock_bh(&pool->sp_lock);
682 time_left = schedule_timeout(timeout);
686 spin_lock_bh(&pool->sp_lock);
687 remove_wait_queue(&rqstp->rq_wait, &wait);
689 pool->sp_stats.threads_timedout++;
691 xprt = rqstp->rq_xprt;
693 svc_thread_dequeue(pool, rqstp);
694 spin_unlock_bh(&pool->sp_lock);
695 dprintk("svc: server %p, no data yet\n", rqstp);
696 if (signalled() || kthread_should_stop())
702 spin_unlock_bh(&pool->sp_lock);
705 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
706 dprintk("svc_recv: found XPT_CLOSE\n");
707 svc_delete_xprt(xprt);
708 /* Leave XPT_BUSY set on the dead xprt: */
711 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
712 struct svc_xprt *newxpt;
713 newxpt = xprt->xpt_ops->xpo_accept(xprt);
716 * We know this module_get will succeed because the
717 * listener holds a reference too
719 __module_get(newxpt->xpt_class->xcl_owner);
720 svc_check_conn_limits(xprt->xpt_server);
721 spin_lock_bh(&serv->sv_lock);
722 set_bit(XPT_TEMP, &newxpt->xpt_flags);
723 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
725 if (serv->sv_temptimer.function == NULL) {
726 /* setup timer to age temp transports */
727 setup_timer(&serv->sv_temptimer,
729 (unsigned long)serv);
730 mod_timer(&serv->sv_temptimer,
731 jiffies + svc_conn_age_period * HZ);
733 spin_unlock_bh(&serv->sv_lock);
734 svc_xprt_received(newxpt);
736 } else if (xprt->xpt_ops->xpo_has_wspace(xprt)) {
737 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
738 rqstp, pool->sp_id, xprt,
739 atomic_read(&xprt->xpt_ref.refcount));
740 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
741 if (rqstp->rq_deferred)
742 len = svc_deferred_recv(rqstp);
744 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
745 dprintk("svc: got len=%d\n", len);
747 svc_xprt_received(xprt);
749 /* No data, incomplete (TCP) read, or accept() */
750 if (len == 0 || len == -EAGAIN)
753 clear_bit(XPT_OLD, &xprt->xpt_flags);
755 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
756 rqstp->rq_chandle.defer = svc_defer;
759 serv->sv_stats->netcnt++;
762 rqstp->rq_res.len = 0;
763 svc_xprt_release(rqstp);
766 EXPORT_SYMBOL_GPL(svc_recv);
771 void svc_drop(struct svc_rqst *rqstp)
773 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
774 svc_xprt_release(rqstp);
776 EXPORT_SYMBOL_GPL(svc_drop);
779 * Return reply to client.
781 int svc_send(struct svc_rqst *rqstp)
783 struct svc_xprt *xprt;
787 xprt = rqstp->rq_xprt;
791 /* release the receive skb before sending the reply */
792 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
794 /* calculate over-all length */
796 xb->len = xb->head[0].iov_len +
800 /* Grab mutex to serialize outgoing data. */
801 mutex_lock(&xprt->xpt_mutex);
802 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
805 len = xprt->xpt_ops->xpo_sendto(rqstp);
806 mutex_unlock(&xprt->xpt_mutex);
807 rpc_wake_up(&xprt->xpt_bc_pending);
808 svc_xprt_release(rqstp);
810 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
816 * Timer function to close old temporary transports, using
817 * a mark-and-sweep algorithm.
819 static void svc_age_temp_xprts(unsigned long closure)
821 struct svc_serv *serv = (struct svc_serv *)closure;
822 struct svc_xprt *xprt;
823 struct list_head *le, *next;
824 LIST_HEAD(to_be_aged);
826 dprintk("svc_age_temp_xprts\n");
828 if (!spin_trylock_bh(&serv->sv_lock)) {
829 /* busy, try again 1 sec later */
830 dprintk("svc_age_temp_xprts: busy\n");
831 mod_timer(&serv->sv_temptimer, jiffies + HZ);
835 list_for_each_safe(le, next, &serv->sv_tempsocks) {
836 xprt = list_entry(le, struct svc_xprt, xpt_list);
838 /* First time through, just mark it OLD. Second time
839 * through, close it. */
840 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
842 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
843 test_bit(XPT_BUSY, &xprt->xpt_flags))
846 list_move(le, &to_be_aged);
847 set_bit(XPT_CLOSE, &xprt->xpt_flags);
848 set_bit(XPT_DETACHED, &xprt->xpt_flags);
850 spin_unlock_bh(&serv->sv_lock);
852 while (!list_empty(&to_be_aged)) {
853 le = to_be_aged.next;
854 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
856 xprt = list_entry(le, struct svc_xprt, xpt_list);
858 dprintk("queuing xprt %p for closing\n", xprt);
860 /* a thread will dequeue and close it soon */
861 svc_xprt_enqueue(xprt);
865 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
868 static void call_xpt_users(struct svc_xprt *xprt)
870 struct svc_xpt_user *u;
872 spin_lock(&xprt->xpt_lock);
873 while (!list_empty(&xprt->xpt_users)) {
874 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
878 spin_unlock(&xprt->xpt_lock);
882 * Remove a dead transport
884 void svc_delete_xprt(struct svc_xprt *xprt)
886 struct svc_serv *serv = xprt->xpt_server;
887 struct svc_deferred_req *dr;
889 /* Only do this once */
890 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
893 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
894 xprt->xpt_ops->xpo_detach(xprt);
896 spin_lock_bh(&serv->sv_lock);
897 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
898 list_del_init(&xprt->xpt_list);
900 * We used to delete the transport from whichever list
901 * it's sk_xprt.xpt_ready node was on, but we don't actually
902 * need to. This is because the only time we're called
903 * while still attached to a queue, the queue itself
904 * is about to be destroyed (in svc_destroy).
906 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
908 spin_unlock_bh(&serv->sv_lock);
910 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
913 call_xpt_users(xprt);
917 void svc_close_xprt(struct svc_xprt *xprt)
919 set_bit(XPT_CLOSE, &xprt->xpt_flags);
920 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
921 /* someone else will have to effect the close */
924 * We expect svc_close_xprt() to work even when no threads are
925 * running (e.g., while configuring the server before starting
926 * any threads), so if the transport isn't busy, we delete
929 svc_delete_xprt(xprt);
931 EXPORT_SYMBOL_GPL(svc_close_xprt);
933 void svc_close_all(struct list_head *xprt_list)
935 struct svc_xprt *xprt;
936 struct svc_xprt *tmp;
939 * The server is shutting down, and no more threads are running.
940 * svc_xprt_enqueue() might still be running, but at worst it
941 * will re-add the xprt to sp_sockets, which will soon get
942 * freed. So we don't bother with any more locking, and don't
943 * leave the close to the (nonexistent) server threads:
945 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
946 set_bit(XPT_CLOSE, &xprt->xpt_flags);
947 svc_delete_xprt(xprt);
952 * Handle defer and revisit of requests
955 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
957 struct svc_deferred_req *dr =
958 container_of(dreq, struct svc_deferred_req, handle);
959 struct svc_xprt *xprt = dr->xprt;
961 spin_lock(&xprt->xpt_lock);
962 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
963 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
964 spin_unlock(&xprt->xpt_lock);
965 dprintk("revisit canceled\n");
970 dprintk("revisit queued\n");
972 list_add(&dr->handle.recent, &xprt->xpt_deferred);
973 spin_unlock(&xprt->xpt_lock);
974 svc_xprt_enqueue(xprt);
979 * Save the request off for later processing. The request buffer looks
982 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
984 * This code can only handle requests that consist of an xprt-header
987 static struct cache_deferred_req *svc_defer(struct cache_req *req)
989 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
990 struct svc_deferred_req *dr;
992 if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
993 return NULL; /* if more than a page, give up FIXME */
994 if (rqstp->rq_deferred) {
995 dr = rqstp->rq_deferred;
996 rqstp->rq_deferred = NULL;
1000 /* FIXME maybe discard if size too large */
1001 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1002 dr = kmalloc(size, GFP_KERNEL);
1006 dr->handle.owner = rqstp->rq_server;
1007 dr->prot = rqstp->rq_prot;
1008 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1009 dr->addrlen = rqstp->rq_addrlen;
1010 dr->daddr = rqstp->rq_daddr;
1011 dr->argslen = rqstp->rq_arg.len >> 2;
1012 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1014 /* back up head to the start of the buffer and copy */
1015 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1016 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1019 svc_xprt_get(rqstp->rq_xprt);
1020 dr->xprt = rqstp->rq_xprt;
1022 dr->handle.revisit = svc_revisit;
1027 * recv data from a deferred request into an active one
1029 static int svc_deferred_recv(struct svc_rqst *rqstp)
1031 struct svc_deferred_req *dr = rqstp->rq_deferred;
1033 /* setup iov_base past transport header */
1034 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1035 /* The iov_len does not include the transport header bytes */
1036 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1037 rqstp->rq_arg.page_len = 0;
1038 /* The rq_arg.len includes the transport header bytes */
1039 rqstp->rq_arg.len = dr->argslen<<2;
1040 rqstp->rq_prot = dr->prot;
1041 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1042 rqstp->rq_addrlen = dr->addrlen;
1043 /* Save off transport header len in case we get deferred again */
1044 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1045 rqstp->rq_daddr = dr->daddr;
1046 rqstp->rq_respages = rqstp->rq_pages;
1047 return (dr->argslen<<2) - dr->xprt_hlen;
1051 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1053 struct svc_deferred_req *dr = NULL;
1055 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1057 spin_lock(&xprt->xpt_lock);
1058 if (!list_empty(&xprt->xpt_deferred)) {
1059 dr = list_entry(xprt->xpt_deferred.next,
1060 struct svc_deferred_req,
1062 list_del_init(&dr->handle.recent);
1064 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1065 spin_unlock(&xprt->xpt_lock);
1070 * svc_find_xprt - find an RPC transport instance
1071 * @serv: pointer to svc_serv to search
1072 * @xcl_name: C string containing transport's class name
1073 * @af: Address family of transport's local address
1074 * @port: transport's IP port number
1076 * Return the transport instance pointer for the endpoint accepting
1077 * connections/peer traffic from the specified transport class,
1078 * address family and port.
1080 * Specifying 0 for the address family or port is effectively a
1081 * wild-card, and will result in matching the first transport in the
1082 * service's list that has a matching class name.
1084 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1085 const sa_family_t af, const unsigned short port)
1087 struct svc_xprt *xprt;
1088 struct svc_xprt *found = NULL;
1090 /* Sanity check the args */
1091 if (serv == NULL || xcl_name == NULL)
1094 spin_lock_bh(&serv->sv_lock);
1095 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1096 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1098 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1100 if (port != 0 && port != svc_xprt_local_port(xprt))
1106 spin_unlock_bh(&serv->sv_lock);
1109 EXPORT_SYMBOL_GPL(svc_find_xprt);
1111 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1112 char *pos, int remaining)
1116 len = snprintf(pos, remaining, "%s %u\n",
1117 xprt->xpt_class->xcl_name,
1118 svc_xprt_local_port(xprt));
1119 if (len >= remaining)
1120 return -ENAMETOOLONG;
1125 * svc_xprt_names - format a buffer with a list of transport names
1126 * @serv: pointer to an RPC service
1127 * @buf: pointer to a buffer to be filled in
1128 * @buflen: length of buffer to be filled in
1130 * Fills in @buf with a string containing a list of transport names,
1131 * each name terminated with '\n'.
1133 * Returns positive length of the filled-in string on success; otherwise
1134 * a negative errno value is returned if an error occurs.
1136 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1138 struct svc_xprt *xprt;
1142 /* Sanity check args */
1146 spin_lock_bh(&serv->sv_lock);
1150 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1151 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1163 spin_unlock_bh(&serv->sv_lock);
1166 EXPORT_SYMBOL_GPL(svc_xprt_names);
1169 /*----------------------------------------------------------------------------*/
1171 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1173 unsigned int pidx = (unsigned int)*pos;
1174 struct svc_serv *serv = m->private;
1176 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1179 return SEQ_START_TOKEN;
1180 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1183 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1185 struct svc_pool *pool = p;
1186 struct svc_serv *serv = m->private;
1188 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1190 if (p == SEQ_START_TOKEN) {
1191 pool = &serv->sv_pools[0];
1193 unsigned int pidx = (pool - &serv->sv_pools[0]);
1194 if (pidx < serv->sv_nrpools-1)
1195 pool = &serv->sv_pools[pidx+1];
1203 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1207 static int svc_pool_stats_show(struct seq_file *m, void *p)
1209 struct svc_pool *pool = p;
1211 if (p == SEQ_START_TOKEN) {
1212 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1216 seq_printf(m, "%u %lu %lu %lu %lu\n",
1218 pool->sp_stats.packets,
1219 pool->sp_stats.sockets_queued,
1220 pool->sp_stats.threads_woken,
1221 pool->sp_stats.threads_timedout);
1226 static const struct seq_operations svc_pool_stats_seq_ops = {
1227 .start = svc_pool_stats_start,
1228 .next = svc_pool_stats_next,
1229 .stop = svc_pool_stats_stop,
1230 .show = svc_pool_stats_show,
1233 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1237 err = seq_open(file, &svc_pool_stats_seq_ops);
1239 ((struct seq_file *) file->private_data)->private = serv;
1242 EXPORT_SYMBOL(svc_pool_stats_open);
1244 /*----------------------------------------------------------------------------*/