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>
16 #include <linux/sunrpc/xprt.h>
17 #include <linux/module.h>
19 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
21 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
22 static int svc_deferred_recv(struct svc_rqst *rqstp);
23 static struct cache_deferred_req *svc_defer(struct cache_req *req);
24 static void svc_age_temp_xprts(unsigned long closure);
26 /* apparently the "standard" is that clients close
27 * idle connections after 5 minutes, servers after
29 * http://www.connectathon.org/talks96/nfstcp.pdf
31 static int svc_conn_age_period = 6*60;
33 /* List of registered transport classes */
34 static DEFINE_SPINLOCK(svc_xprt_class_lock);
35 static LIST_HEAD(svc_xprt_class_list);
37 /* SMP locking strategy:
39 * svc_pool->sp_lock protects most of the fields of that pool.
40 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
41 * when both need to be taken (rare), svc_serv->sv_lock is first.
42 * BKL protects svc_serv->sv_nrthread.
43 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
44 * and the ->sk_info_authunix cache.
46 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
47 * enqueued multiply. During normal transport processing this bit
48 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
49 * Providers should not manipulate this bit directly.
51 * Some flags can be set to certain values at any time
52 * providing that certain rules are followed:
55 * - Can be set or cleared at any time.
56 * - After a set, svc_xprt_enqueue must be called to enqueue
57 * the transport for processing.
58 * - After a clear, the transport must be read/accepted.
59 * If this succeeds, it must be set again.
61 * - Can set at any time. It is never cleared.
63 * - Can only be set while XPT_BUSY is held which ensures
64 * that no other thread will be using the transport or will
65 * try to set XPT_DEAD.
68 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
70 struct svc_xprt_class *cl;
73 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
75 INIT_LIST_HEAD(&xcl->xcl_list);
76 spin_lock(&svc_xprt_class_lock);
77 /* Make sure there isn't already a class with the same name */
78 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
79 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
82 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
85 spin_unlock(&svc_xprt_class_lock);
88 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
90 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
92 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
93 spin_lock(&svc_xprt_class_lock);
94 list_del_init(&xcl->xcl_list);
95 spin_unlock(&svc_xprt_class_lock);
97 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
100 * Format the transport list for printing
102 int svc_print_xprts(char *buf, int maxlen)
104 struct svc_xprt_class *xcl;
109 spin_lock(&svc_xprt_class_lock);
110 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
113 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
114 slen = strlen(tmpstr);
115 if (len + slen > maxlen)
120 spin_unlock(&svc_xprt_class_lock);
125 static void svc_xprt_free(struct kref *kref)
127 struct svc_xprt *xprt =
128 container_of(kref, struct svc_xprt, xpt_ref);
129 struct module *owner = xprt->xpt_class->xcl_owner;
130 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
131 svcauth_unix_info_release(xprt);
132 put_net(xprt->xpt_net);
133 /* See comment on corresponding get in xs_setup_bc_tcp(): */
134 if (xprt->xpt_bc_xprt)
135 xprt_put(xprt->xpt_bc_xprt);
136 xprt->xpt_ops->xpo_free(xprt);
140 void svc_xprt_put(struct svc_xprt *xprt)
142 kref_put(&xprt->xpt_ref, svc_xprt_free);
144 EXPORT_SYMBOL_GPL(svc_xprt_put);
147 * Called by transport drivers to initialize the transport independent
148 * portion of the transport instance.
150 void svc_xprt_init(struct svc_xprt_class *xcl, struct svc_xprt *xprt,
151 struct svc_serv *serv)
153 memset(xprt, 0, sizeof(*xprt));
154 xprt->xpt_class = xcl;
155 xprt->xpt_ops = xcl->xcl_ops;
156 kref_init(&xprt->xpt_ref);
157 xprt->xpt_server = serv;
158 INIT_LIST_HEAD(&xprt->xpt_list);
159 INIT_LIST_HEAD(&xprt->xpt_ready);
160 INIT_LIST_HEAD(&xprt->xpt_deferred);
161 INIT_LIST_HEAD(&xprt->xpt_users);
162 mutex_init(&xprt->xpt_mutex);
163 spin_lock_init(&xprt->xpt_lock);
164 set_bit(XPT_BUSY, &xprt->xpt_flags);
165 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
166 xprt->xpt_net = get_net(&init_net);
168 EXPORT_SYMBOL_GPL(svc_xprt_init);
170 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
171 struct svc_serv *serv,
174 const unsigned short port,
177 struct sockaddr_in sin = {
178 .sin_family = AF_INET,
179 .sin_addr.s_addr = htonl(INADDR_ANY),
180 .sin_port = htons(port),
182 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
183 struct sockaddr_in6 sin6 = {
184 .sin6_family = AF_INET6,
185 .sin6_addr = IN6ADDR_ANY_INIT,
186 .sin6_port = htons(port),
188 #endif /* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
189 struct sockaddr *sap;
194 sap = (struct sockaddr *)&sin;
197 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
199 sap = (struct sockaddr *)&sin6;
202 #endif /* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
204 return ERR_PTR(-EAFNOSUPPORT);
207 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
210 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
211 struct net *net, const int family,
212 const unsigned short port, int flags)
214 struct svc_xprt_class *xcl;
216 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
217 spin_lock(&svc_xprt_class_lock);
218 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
219 struct svc_xprt *newxprt;
220 unsigned short newport;
222 if (strcmp(xprt_name, xcl->xcl_name))
225 if (!try_module_get(xcl->xcl_owner))
228 spin_unlock(&svc_xprt_class_lock);
229 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
230 if (IS_ERR(newxprt)) {
231 module_put(xcl->xcl_owner);
232 return PTR_ERR(newxprt);
235 clear_bit(XPT_TEMP, &newxprt->xpt_flags);
236 spin_lock_bh(&serv->sv_lock);
237 list_add(&newxprt->xpt_list, &serv->sv_permsocks);
238 spin_unlock_bh(&serv->sv_lock);
239 newport = svc_xprt_local_port(newxprt);
240 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
244 spin_unlock(&svc_xprt_class_lock);
245 dprintk("svc: transport %s not found\n", xprt_name);
247 /* This errno is exposed to user space. Provide a reasonable
248 * perror msg for a bad transport. */
249 return -EPROTONOSUPPORT;
251 EXPORT_SYMBOL_GPL(svc_create_xprt);
254 * Copy the local and remote xprt addresses to the rqstp structure
256 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
258 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
259 rqstp->rq_addrlen = xprt->xpt_remotelen;
262 * Destination address in request is needed for binding the
263 * source address in RPC replies/callbacks later.
265 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
266 rqstp->rq_daddrlen = xprt->xpt_locallen;
268 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
271 * svc_print_addr - Format rq_addr field for printing
272 * @rqstp: svc_rqst struct containing address to print
273 * @buf: target buffer for formatted address
274 * @len: length of target buffer
277 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
279 return __svc_print_addr(svc_addr(rqstp), buf, len);
281 EXPORT_SYMBOL_GPL(svc_print_addr);
284 * Queue up an idle server thread. Must have pool->sp_lock held.
285 * Note: this is really a stack rather than a queue, so that we only
286 * use as many different threads as we need, and the rest don't pollute
289 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
291 list_add(&rqstp->rq_list, &pool->sp_threads);
295 * Dequeue an nfsd thread. Must have pool->sp_lock held.
297 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
299 list_del(&rqstp->rq_list);
302 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
304 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
306 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
307 return xprt->xpt_ops->xpo_has_wspace(xprt);
312 * Queue up a transport with data pending. If there are idle nfsd
313 * processes, wake 'em up.
316 void svc_xprt_enqueue(struct svc_xprt *xprt)
318 struct svc_pool *pool;
319 struct svc_rqst *rqstp;
322 if (!svc_xprt_has_something_to_do(xprt))
326 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
329 spin_lock_bh(&pool->sp_lock);
331 if (!list_empty(&pool->sp_threads) &&
332 !list_empty(&pool->sp_sockets))
335 "threads and transports both waiting??\n");
337 pool->sp_stats.packets++;
339 /* Mark transport as busy. It will remain in this state until
340 * the provider calls svc_xprt_received. We update XPT_BUSY
341 * atomically because it also guards against trying to enqueue
342 * the transport twice.
344 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
345 /* Don't enqueue transport while already enqueued */
346 dprintk("svc: transport %p busy, not enqueued\n", xprt);
350 if (!list_empty(&pool->sp_threads)) {
351 rqstp = list_entry(pool->sp_threads.next,
354 dprintk("svc: transport %p served by daemon %p\n",
356 svc_thread_dequeue(pool, rqstp);
359 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
360 rqstp, rqstp->rq_xprt);
361 rqstp->rq_xprt = xprt;
363 pool->sp_stats.threads_woken++;
364 wake_up(&rqstp->rq_wait);
366 dprintk("svc: transport %p put into queue\n", xprt);
367 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
368 pool->sp_stats.sockets_queued++;
372 spin_unlock_bh(&pool->sp_lock);
374 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
377 * Dequeue the first transport. Must be called with the pool->sp_lock held.
379 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
381 struct svc_xprt *xprt;
383 if (list_empty(&pool->sp_sockets))
386 xprt = list_entry(pool->sp_sockets.next,
387 struct svc_xprt, xpt_ready);
388 list_del_init(&xprt->xpt_ready);
390 dprintk("svc: transport %p dequeued, inuse=%d\n",
391 xprt, atomic_read(&xprt->xpt_ref.refcount));
397 * svc_xprt_received conditionally queues the transport for processing
398 * by another thread. The caller must hold the XPT_BUSY bit and must
399 * not thereafter touch transport data.
401 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
402 * insufficient) data.
404 void svc_xprt_received(struct svc_xprt *xprt)
406 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
407 /* As soon as we clear busy, the xprt could be closed and
408 * 'put', so we need a reference to call svc_xprt_enqueue with:
411 clear_bit(XPT_BUSY, &xprt->xpt_flags);
412 svc_xprt_enqueue(xprt);
415 EXPORT_SYMBOL_GPL(svc_xprt_received);
418 * svc_reserve - change the space reserved for the reply to a request.
419 * @rqstp: The request in question
420 * @space: new max space to reserve
422 * Each request reserves some space on the output queue of the transport
423 * to make sure the reply fits. This function reduces that reserved
424 * space to be the amount of space used already, plus @space.
427 void svc_reserve(struct svc_rqst *rqstp, int space)
429 space += rqstp->rq_res.head[0].iov_len;
431 if (space < rqstp->rq_reserved) {
432 struct svc_xprt *xprt = rqstp->rq_xprt;
433 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
434 rqstp->rq_reserved = space;
436 svc_xprt_enqueue(xprt);
439 EXPORT_SYMBOL_GPL(svc_reserve);
441 static void svc_xprt_release(struct svc_rqst *rqstp)
443 struct svc_xprt *xprt = rqstp->rq_xprt;
445 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
447 kfree(rqstp->rq_deferred);
448 rqstp->rq_deferred = NULL;
450 svc_free_res_pages(rqstp);
451 rqstp->rq_res.page_len = 0;
452 rqstp->rq_res.page_base = 0;
454 /* Reset response buffer and release
456 * But first, check that enough space was reserved
457 * for the reply, otherwise we have a bug!
459 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
460 printk(KERN_ERR "RPC request reserved %d but used %d\n",
464 rqstp->rq_res.head[0].iov_len = 0;
465 svc_reserve(rqstp, 0);
466 rqstp->rq_xprt = NULL;
472 * External function to wake up a server waiting for data
473 * This really only makes sense for services like lockd
474 * which have exactly one thread anyway.
476 void svc_wake_up(struct svc_serv *serv)
478 struct svc_rqst *rqstp;
480 struct svc_pool *pool;
482 for (i = 0; i < serv->sv_nrpools; i++) {
483 pool = &serv->sv_pools[i];
485 spin_lock_bh(&pool->sp_lock);
486 if (!list_empty(&pool->sp_threads)) {
487 rqstp = list_entry(pool->sp_threads.next,
490 dprintk("svc: daemon %p woken up.\n", rqstp);
492 svc_thread_dequeue(pool, rqstp);
493 rqstp->rq_xprt = NULL;
495 wake_up(&rqstp->rq_wait);
497 spin_unlock_bh(&pool->sp_lock);
500 EXPORT_SYMBOL_GPL(svc_wake_up);
502 int svc_port_is_privileged(struct sockaddr *sin)
504 switch (sin->sa_family) {
506 return ntohs(((struct sockaddr_in *)sin)->sin_port)
509 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
517 * Make sure that we don't have too many active connections. If we have,
518 * something must be dropped. It's not clear what will happen if we allow
519 * "too many" connections, but when dealing with network-facing software,
520 * we have to code defensively. Here we do that by imposing hard limits.
522 * There's no point in trying to do random drop here for DoS
523 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
524 * attacker can easily beat that.
526 * The only somewhat efficient mechanism would be if drop old
527 * connections from the same IP first. But right now we don't even
528 * record the client IP in svc_sock.
530 * single-threaded services that expect a lot of clients will probably
531 * need to set sv_maxconn to override the default value which is based
532 * on the number of threads
534 static void svc_check_conn_limits(struct svc_serv *serv)
536 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
537 (serv->sv_nrthreads+3) * 20;
539 if (serv->sv_tmpcnt > limit) {
540 struct svc_xprt *xprt = NULL;
541 spin_lock_bh(&serv->sv_lock);
542 if (!list_empty(&serv->sv_tempsocks)) {
543 if (net_ratelimit()) {
544 /* Try to help the admin */
545 printk(KERN_NOTICE "%s: too many open "
546 "connections, consider increasing %s\n",
547 serv->sv_name, serv->sv_maxconn ?
548 "the max number of connections." :
549 "the number of threads.");
552 * Always select the oldest connection. It's not fair,
555 xprt = list_entry(serv->sv_tempsocks.prev,
558 set_bit(XPT_CLOSE, &xprt->xpt_flags);
561 spin_unlock_bh(&serv->sv_lock);
564 svc_xprt_enqueue(xprt);
571 * Receive the next request on any transport. This code is carefully
572 * organised not to touch any cachelines in the shared svc_serv
573 * structure, only cachelines in the local svc_pool.
575 int svc_recv(struct svc_rqst *rqstp, long timeout)
577 struct svc_xprt *xprt = NULL;
578 struct svc_serv *serv = rqstp->rq_server;
579 struct svc_pool *pool = rqstp->rq_pool;
583 DECLARE_WAITQUEUE(wait, current);
586 dprintk("svc: server %p waiting for data (to = %ld)\n",
591 "svc_recv: service %p, transport not NULL!\n",
593 if (waitqueue_active(&rqstp->rq_wait))
595 "svc_recv: service %p, wait queue active!\n",
598 /* now allocate needed pages. If we get a failure, sleep briefly */
599 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
600 for (i = 0; i < pages ; i++)
601 while (rqstp->rq_pages[i] == NULL) {
602 struct page *p = alloc_page(GFP_KERNEL);
604 set_current_state(TASK_INTERRUPTIBLE);
605 if (signalled() || kthread_should_stop()) {
606 set_current_state(TASK_RUNNING);
609 schedule_timeout(msecs_to_jiffies(500));
611 rqstp->rq_pages[i] = p;
613 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
614 BUG_ON(pages >= RPCSVC_MAXPAGES);
616 /* Make arg->head point to first page and arg->pages point to rest */
617 arg = &rqstp->rq_arg;
618 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
619 arg->head[0].iov_len = PAGE_SIZE;
620 arg->pages = rqstp->rq_pages + 1;
622 /* save at least one page for response */
623 arg->page_len = (pages-2)*PAGE_SIZE;
624 arg->len = (pages-1)*PAGE_SIZE;
625 arg->tail[0].iov_len = 0;
629 if (signalled() || kthread_should_stop())
632 /* Normally we will wait up to 5 seconds for any required
633 * cache information to be provided.
635 rqstp->rq_chandle.thread_wait = 5*HZ;
637 spin_lock_bh(&pool->sp_lock);
638 xprt = svc_xprt_dequeue(pool);
640 rqstp->rq_xprt = xprt;
643 /* As there is a shortage of threads and this request
644 * had to be queued, don't allow the thread to wait so
645 * long for cache updates.
647 rqstp->rq_chandle.thread_wait = 1*HZ;
649 /* No data pending. Go to sleep */
650 svc_thread_enqueue(pool, rqstp);
653 * We have to be able to interrupt this wait
654 * to bring down the daemons ...
656 set_current_state(TASK_INTERRUPTIBLE);
659 * checking kthread_should_stop() here allows us to avoid
660 * locking and signalling when stopping kthreads that call
661 * svc_recv. If the thread has already been woken up, then
662 * we can exit here without sleeping. If not, then it
663 * it'll be woken up quickly during the schedule_timeout
665 if (kthread_should_stop()) {
666 set_current_state(TASK_RUNNING);
667 spin_unlock_bh(&pool->sp_lock);
671 add_wait_queue(&rqstp->rq_wait, &wait);
672 spin_unlock_bh(&pool->sp_lock);
674 time_left = schedule_timeout(timeout);
678 spin_lock_bh(&pool->sp_lock);
679 remove_wait_queue(&rqstp->rq_wait, &wait);
681 pool->sp_stats.threads_timedout++;
683 xprt = rqstp->rq_xprt;
685 svc_thread_dequeue(pool, rqstp);
686 spin_unlock_bh(&pool->sp_lock);
687 dprintk("svc: server %p, no data yet\n", rqstp);
688 if (signalled() || kthread_should_stop())
694 spin_unlock_bh(&pool->sp_lock);
697 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
698 dprintk("svc_recv: found XPT_CLOSE\n");
699 svc_delete_xprt(xprt);
700 /* Leave XPT_BUSY set on the dead xprt: */
703 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
704 struct svc_xprt *newxpt;
705 newxpt = xprt->xpt_ops->xpo_accept(xprt);
708 * We know this module_get will succeed because the
709 * listener holds a reference too
711 __module_get(newxpt->xpt_class->xcl_owner);
712 svc_check_conn_limits(xprt->xpt_server);
713 spin_lock_bh(&serv->sv_lock);
714 set_bit(XPT_TEMP, &newxpt->xpt_flags);
715 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
717 if (serv->sv_temptimer.function == NULL) {
718 /* setup timer to age temp transports */
719 setup_timer(&serv->sv_temptimer,
721 (unsigned long)serv);
722 mod_timer(&serv->sv_temptimer,
723 jiffies + svc_conn_age_period * HZ);
725 spin_unlock_bh(&serv->sv_lock);
726 svc_xprt_received(newxpt);
728 } else if (xprt->xpt_ops->xpo_has_wspace(xprt)) {
729 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
730 rqstp, pool->sp_id, xprt,
731 atomic_read(&xprt->xpt_ref.refcount));
732 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
733 if (rqstp->rq_deferred)
734 len = svc_deferred_recv(rqstp);
736 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
737 dprintk("svc: got len=%d\n", len);
738 rqstp->rq_reserved = serv->sv_max_mesg;
739 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
741 svc_xprt_received(xprt);
743 /* No data, incomplete (TCP) read, or accept() */
744 if (len == 0 || len == -EAGAIN)
747 clear_bit(XPT_OLD, &xprt->xpt_flags);
749 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
750 rqstp->rq_chandle.defer = svc_defer;
753 serv->sv_stats->netcnt++;
756 rqstp->rq_res.len = 0;
757 svc_xprt_release(rqstp);
760 EXPORT_SYMBOL_GPL(svc_recv);
765 void svc_drop(struct svc_rqst *rqstp)
767 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
768 svc_xprt_release(rqstp);
770 EXPORT_SYMBOL_GPL(svc_drop);
773 * Return reply to client.
775 int svc_send(struct svc_rqst *rqstp)
777 struct svc_xprt *xprt;
781 xprt = rqstp->rq_xprt;
785 /* release the receive skb before sending the reply */
786 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
788 /* calculate over-all length */
790 xb->len = xb->head[0].iov_len +
794 /* Grab mutex to serialize outgoing data. */
795 mutex_lock(&xprt->xpt_mutex);
796 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
797 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
800 len = xprt->xpt_ops->xpo_sendto(rqstp);
801 mutex_unlock(&xprt->xpt_mutex);
802 rpc_wake_up(&xprt->xpt_bc_pending);
803 svc_xprt_release(rqstp);
805 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
811 * Timer function to close old temporary transports, using
812 * a mark-and-sweep algorithm.
814 static void svc_age_temp_xprts(unsigned long closure)
816 struct svc_serv *serv = (struct svc_serv *)closure;
817 struct svc_xprt *xprt;
818 struct list_head *le, *next;
819 LIST_HEAD(to_be_aged);
821 dprintk("svc_age_temp_xprts\n");
823 if (!spin_trylock_bh(&serv->sv_lock)) {
824 /* busy, try again 1 sec later */
825 dprintk("svc_age_temp_xprts: busy\n");
826 mod_timer(&serv->sv_temptimer, jiffies + HZ);
830 list_for_each_safe(le, next, &serv->sv_tempsocks) {
831 xprt = list_entry(le, struct svc_xprt, xpt_list);
833 /* First time through, just mark it OLD. Second time
834 * through, close it. */
835 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
837 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
838 test_bit(XPT_BUSY, &xprt->xpt_flags))
841 list_move(le, &to_be_aged);
842 set_bit(XPT_CLOSE, &xprt->xpt_flags);
843 set_bit(XPT_DETACHED, &xprt->xpt_flags);
845 spin_unlock_bh(&serv->sv_lock);
847 while (!list_empty(&to_be_aged)) {
848 le = to_be_aged.next;
849 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
851 xprt = list_entry(le, struct svc_xprt, xpt_list);
853 dprintk("queuing xprt %p for closing\n", xprt);
855 /* a thread will dequeue and close it soon */
856 svc_xprt_enqueue(xprt);
860 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
863 static void call_xpt_users(struct svc_xprt *xprt)
865 struct svc_xpt_user *u;
867 spin_lock(&xprt->xpt_lock);
868 while (!list_empty(&xprt->xpt_users)) {
869 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
873 spin_unlock(&xprt->xpt_lock);
877 * Remove a dead transport
879 void svc_delete_xprt(struct svc_xprt *xprt)
881 struct svc_serv *serv = xprt->xpt_server;
882 struct svc_deferred_req *dr;
884 /* Only do this once */
885 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
888 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
889 xprt->xpt_ops->xpo_detach(xprt);
891 spin_lock_bh(&serv->sv_lock);
892 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
893 list_del_init(&xprt->xpt_list);
894 BUG_ON(!list_empty(&xprt->xpt_ready));
895 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
897 spin_unlock_bh(&serv->sv_lock);
899 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
902 call_xpt_users(xprt);
906 void svc_close_xprt(struct svc_xprt *xprt)
908 set_bit(XPT_CLOSE, &xprt->xpt_flags);
909 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
910 /* someone else will have to effect the close */
913 * We expect svc_close_xprt() to work even when no threads are
914 * running (e.g., while configuring the server before starting
915 * any threads), so if the transport isn't busy, we delete
918 svc_delete_xprt(xprt);
920 EXPORT_SYMBOL_GPL(svc_close_xprt);
922 static void svc_close_list(struct list_head *xprt_list)
924 struct svc_xprt *xprt;
926 list_for_each_entry(xprt, xprt_list, xpt_list) {
927 set_bit(XPT_CLOSE, &xprt->xpt_flags);
928 set_bit(XPT_BUSY, &xprt->xpt_flags);
932 void svc_close_all(struct svc_serv *serv)
934 struct svc_pool *pool;
935 struct svc_xprt *xprt;
936 struct svc_xprt *tmp;
939 svc_close_list(&serv->sv_tempsocks);
940 svc_close_list(&serv->sv_permsocks);
942 for (i = 0; i < serv->sv_nrpools; i++) {
943 pool = &serv->sv_pools[i];
945 spin_lock_bh(&pool->sp_lock);
946 while (!list_empty(&pool->sp_sockets)) {
947 xprt = list_first_entry(&pool->sp_sockets, struct svc_xprt, xpt_ready);
948 list_del_init(&xprt->xpt_ready);
950 spin_unlock_bh(&pool->sp_lock);
953 * At this point the sp_sockets lists will stay empty, since
954 * svc_enqueue will not add new entries without taking the
955 * sp_lock and checking XPT_BUSY.
957 list_for_each_entry_safe(xprt, tmp, &serv->sv_tempsocks, xpt_list)
958 svc_delete_xprt(xprt);
959 list_for_each_entry_safe(xprt, tmp, &serv->sv_permsocks, xpt_list)
960 svc_delete_xprt(xprt);
962 BUG_ON(!list_empty(&serv->sv_permsocks));
963 BUG_ON(!list_empty(&serv->sv_tempsocks));
967 * Handle defer and revisit of requests
970 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
972 struct svc_deferred_req *dr =
973 container_of(dreq, struct svc_deferred_req, handle);
974 struct svc_xprt *xprt = dr->xprt;
976 spin_lock(&xprt->xpt_lock);
977 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
978 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
979 spin_unlock(&xprt->xpt_lock);
980 dprintk("revisit canceled\n");
985 dprintk("revisit queued\n");
987 list_add(&dr->handle.recent, &xprt->xpt_deferred);
988 spin_unlock(&xprt->xpt_lock);
989 svc_xprt_enqueue(xprt);
994 * Save the request off for later processing. The request buffer looks
997 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
999 * This code can only handle requests that consist of an xprt-header
1002 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1004 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1005 struct svc_deferred_req *dr;
1007 if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
1008 return NULL; /* if more than a page, give up FIXME */
1009 if (rqstp->rq_deferred) {
1010 dr = rqstp->rq_deferred;
1011 rqstp->rq_deferred = NULL;
1015 /* FIXME maybe discard if size too large */
1016 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1017 dr = kmalloc(size, GFP_KERNEL);
1021 dr->handle.owner = rqstp->rq_server;
1022 dr->prot = rqstp->rq_prot;
1023 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1024 dr->addrlen = rqstp->rq_addrlen;
1025 dr->daddr = rqstp->rq_daddr;
1026 dr->argslen = rqstp->rq_arg.len >> 2;
1027 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1029 /* back up head to the start of the buffer and copy */
1030 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1031 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1034 svc_xprt_get(rqstp->rq_xprt);
1035 dr->xprt = rqstp->rq_xprt;
1036 rqstp->rq_dropme = true;
1038 dr->handle.revisit = svc_revisit;
1043 * recv data from a deferred request into an active one
1045 static int svc_deferred_recv(struct svc_rqst *rqstp)
1047 struct svc_deferred_req *dr = rqstp->rq_deferred;
1049 /* setup iov_base past transport header */
1050 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1051 /* The iov_len does not include the transport header bytes */
1052 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1053 rqstp->rq_arg.page_len = 0;
1054 /* The rq_arg.len includes the transport header bytes */
1055 rqstp->rq_arg.len = dr->argslen<<2;
1056 rqstp->rq_prot = dr->prot;
1057 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1058 rqstp->rq_addrlen = dr->addrlen;
1059 /* Save off transport header len in case we get deferred again */
1060 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1061 rqstp->rq_daddr = dr->daddr;
1062 rqstp->rq_respages = rqstp->rq_pages;
1063 return (dr->argslen<<2) - dr->xprt_hlen;
1067 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1069 struct svc_deferred_req *dr = NULL;
1071 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1073 spin_lock(&xprt->xpt_lock);
1074 if (!list_empty(&xprt->xpt_deferred)) {
1075 dr = list_entry(xprt->xpt_deferred.next,
1076 struct svc_deferred_req,
1078 list_del_init(&dr->handle.recent);
1080 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1081 spin_unlock(&xprt->xpt_lock);
1086 * svc_find_xprt - find an RPC transport instance
1087 * @serv: pointer to svc_serv to search
1088 * @xcl_name: C string containing transport's class name
1089 * @af: Address family of transport's local address
1090 * @port: transport's IP port number
1092 * Return the transport instance pointer for the endpoint accepting
1093 * connections/peer traffic from the specified transport class,
1094 * address family and port.
1096 * Specifying 0 for the address family or port is effectively a
1097 * wild-card, and will result in matching the first transport in the
1098 * service's list that has a matching class name.
1100 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1101 const sa_family_t af, const unsigned short port)
1103 struct svc_xprt *xprt;
1104 struct svc_xprt *found = NULL;
1106 /* Sanity check the args */
1107 if (serv == NULL || xcl_name == NULL)
1110 spin_lock_bh(&serv->sv_lock);
1111 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1112 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1114 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1116 if (port != 0 && port != svc_xprt_local_port(xprt))
1122 spin_unlock_bh(&serv->sv_lock);
1125 EXPORT_SYMBOL_GPL(svc_find_xprt);
1127 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1128 char *pos, int remaining)
1132 len = snprintf(pos, remaining, "%s %u\n",
1133 xprt->xpt_class->xcl_name,
1134 svc_xprt_local_port(xprt));
1135 if (len >= remaining)
1136 return -ENAMETOOLONG;
1141 * svc_xprt_names - format a buffer with a list of transport names
1142 * @serv: pointer to an RPC service
1143 * @buf: pointer to a buffer to be filled in
1144 * @buflen: length of buffer to be filled in
1146 * Fills in @buf with a string containing a list of transport names,
1147 * each name terminated with '\n'.
1149 * Returns positive length of the filled-in string on success; otherwise
1150 * a negative errno value is returned if an error occurs.
1152 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1154 struct svc_xprt *xprt;
1158 /* Sanity check args */
1162 spin_lock_bh(&serv->sv_lock);
1166 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1167 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1179 spin_unlock_bh(&serv->sv_lock);
1182 EXPORT_SYMBOL_GPL(svc_xprt_names);
1185 /*----------------------------------------------------------------------------*/
1187 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1189 unsigned int pidx = (unsigned int)*pos;
1190 struct svc_serv *serv = m->private;
1192 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1195 return SEQ_START_TOKEN;
1196 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1199 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1201 struct svc_pool *pool = p;
1202 struct svc_serv *serv = m->private;
1204 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1206 if (p == SEQ_START_TOKEN) {
1207 pool = &serv->sv_pools[0];
1209 unsigned int pidx = (pool - &serv->sv_pools[0]);
1210 if (pidx < serv->sv_nrpools-1)
1211 pool = &serv->sv_pools[pidx+1];
1219 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1223 static int svc_pool_stats_show(struct seq_file *m, void *p)
1225 struct svc_pool *pool = p;
1227 if (p == SEQ_START_TOKEN) {
1228 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1232 seq_printf(m, "%u %lu %lu %lu %lu\n",
1234 pool->sp_stats.packets,
1235 pool->sp_stats.sockets_queued,
1236 pool->sp_stats.threads_woken,
1237 pool->sp_stats.threads_timedout);
1242 static const struct seq_operations svc_pool_stats_seq_ops = {
1243 .start = svc_pool_stats_start,
1244 .next = svc_pool_stats_next,
1245 .stop = svc_pool_stats_stop,
1246 .show = svc_pool_stats_show,
1249 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1253 err = seq_open(file, &svc_pool_stats_seq_ops);
1255 ((struct seq_file *) file->private_data)->private = serv;
1258 EXPORT_SYMBOL(svc_pool_stats_open);
1260 /*----------------------------------------------------------------------------*/