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>
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 #define SVC_MAX_WAKING 5
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 list_head *le;
109 spin_lock(&svc_xprt_class_lock);
110 list_for_each(le, &svc_xprt_class_list) {
112 struct svc_xprt_class *xcl =
113 list_entry(le, struct svc_xprt_class, xcl_list);
115 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
116 slen = strlen(tmpstr);
117 if (len + slen > maxlen)
122 spin_unlock(&svc_xprt_class_lock);
127 static void svc_xprt_free(struct kref *kref)
129 struct svc_xprt *xprt =
130 container_of(kref, struct svc_xprt, xpt_ref);
131 struct module *owner = xprt->xpt_class->xcl_owner;
132 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags)
133 && xprt->xpt_auth_cache != NULL)
134 svcauth_unix_info_release(xprt->xpt_auth_cache);
135 xprt->xpt_ops->xpo_free(xprt);
139 void svc_xprt_put(struct svc_xprt *xprt)
141 kref_put(&xprt->xpt_ref, svc_xprt_free);
143 EXPORT_SYMBOL_GPL(svc_xprt_put);
146 * Called by transport drivers to initialize the transport independent
147 * portion of the transport instance.
149 void svc_xprt_init(struct svc_xprt_class *xcl, struct svc_xprt *xprt,
150 struct svc_serv *serv)
152 memset(xprt, 0, sizeof(*xprt));
153 xprt->xpt_class = xcl;
154 xprt->xpt_ops = xcl->xcl_ops;
155 kref_init(&xprt->xpt_ref);
156 xprt->xpt_server = serv;
157 INIT_LIST_HEAD(&xprt->xpt_list);
158 INIT_LIST_HEAD(&xprt->xpt_ready);
159 INIT_LIST_HEAD(&xprt->xpt_deferred);
160 mutex_init(&xprt->xpt_mutex);
161 spin_lock_init(&xprt->xpt_lock);
162 set_bit(XPT_BUSY, &xprt->xpt_flags);
163 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
165 EXPORT_SYMBOL_GPL(svc_xprt_init);
167 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
168 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 struct sockaddr_in6 sin6 = {
179 .sin6_family = AF_INET6,
180 .sin6_addr = IN6ADDR_ANY_INIT,
181 .sin6_port = htons(port),
183 struct sockaddr *sap;
188 sap = (struct sockaddr *)&sin;
192 sap = (struct sockaddr *)&sin6;
196 return ERR_PTR(-EAFNOSUPPORT);
199 return xcl->xcl_ops->xpo_create(serv, sap, len, flags);
202 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
203 const int family, const unsigned short port,
206 struct svc_xprt_class *xcl;
208 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
209 spin_lock(&svc_xprt_class_lock);
210 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
211 struct svc_xprt *newxprt;
213 if (strcmp(xprt_name, xcl->xcl_name))
216 if (!try_module_get(xcl->xcl_owner))
219 spin_unlock(&svc_xprt_class_lock);
220 newxprt = __svc_xpo_create(xcl, serv, family, port, flags);
221 if (IS_ERR(newxprt)) {
222 module_put(xcl->xcl_owner);
223 return PTR_ERR(newxprt);
226 clear_bit(XPT_TEMP, &newxprt->xpt_flags);
227 spin_lock_bh(&serv->sv_lock);
228 list_add(&newxprt->xpt_list, &serv->sv_permsocks);
229 spin_unlock_bh(&serv->sv_lock);
230 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
231 return svc_xprt_local_port(newxprt);
234 spin_unlock(&svc_xprt_class_lock);
235 dprintk("svc: transport %s not found\n", xprt_name);
238 EXPORT_SYMBOL_GPL(svc_create_xprt);
241 * Copy the local and remote xprt addresses to the rqstp structure
243 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
245 struct sockaddr *sin;
247 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
248 rqstp->rq_addrlen = xprt->xpt_remotelen;
251 * Destination address in request is needed for binding the
252 * source address in RPC replies/callbacks later.
254 sin = (struct sockaddr *)&xprt->xpt_local;
255 switch (sin->sa_family) {
257 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
260 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
264 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
267 * svc_print_addr - Format rq_addr field for printing
268 * @rqstp: svc_rqst struct containing address to print
269 * @buf: target buffer for formatted address
270 * @len: length of target buffer
273 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
275 return __svc_print_addr(svc_addr(rqstp), buf, len);
277 EXPORT_SYMBOL_GPL(svc_print_addr);
280 * Queue up an idle server thread. Must have pool->sp_lock held.
281 * Note: this is really a stack rather than a queue, so that we only
282 * use as many different threads as we need, and the rest don't pollute
285 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
287 list_add(&rqstp->rq_list, &pool->sp_threads);
291 * Dequeue an nfsd thread. Must have pool->sp_lock held.
293 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
295 list_del(&rqstp->rq_list);
299 * Queue up a transport with data pending. If there are idle nfsd
300 * processes, wake 'em up.
303 void svc_xprt_enqueue(struct svc_xprt *xprt)
305 struct svc_serv *serv = xprt->xpt_server;
306 struct svc_pool *pool;
307 struct svc_rqst *rqstp;
311 if (!(xprt->xpt_flags &
312 ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
316 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
319 spin_lock_bh(&pool->sp_lock);
321 if (test_bit(XPT_DEAD, &xprt->xpt_flags)) {
322 /* Don't enqueue dead transports */
323 dprintk("svc: transport %p is dead, not enqueued\n", xprt);
327 pool->sp_stats.packets++;
329 /* Mark transport as busy. It will remain in this state until
330 * the provider calls svc_xprt_received. We update XPT_BUSY
331 * atomically because it also guards against trying to enqueue
332 * the transport twice.
334 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
335 /* Don't enqueue transport while already enqueued */
336 dprintk("svc: transport %p busy, not enqueued\n", xprt);
339 BUG_ON(xprt->xpt_pool != NULL);
340 xprt->xpt_pool = pool;
342 /* Handle pending connection */
343 if (test_bit(XPT_CONN, &xprt->xpt_flags))
346 /* Handle close in-progress */
347 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
350 /* Check if we have space to reply to a request */
351 if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
352 /* Don't enqueue while not enough space for reply */
353 dprintk("svc: no write space, transport %p not enqueued\n",
355 xprt->xpt_pool = NULL;
356 clear_bit(XPT_BUSY, &xprt->xpt_flags);
361 /* Work out whether threads are available */
362 thread_avail = !list_empty(&pool->sp_threads); /* threads are asleep */
363 if (pool->sp_nwaking >= SVC_MAX_WAKING) {
364 /* too many threads are runnable and trying to wake up */
366 pool->sp_stats.overloads_avoided++;
370 rqstp = list_entry(pool->sp_threads.next,
373 dprintk("svc: transport %p served by daemon %p\n",
375 svc_thread_dequeue(pool, rqstp);
378 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
379 rqstp, rqstp->rq_xprt);
380 rqstp->rq_xprt = xprt;
382 rqstp->rq_reserved = serv->sv_max_mesg;
383 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
384 rqstp->rq_waking = 1;
386 pool->sp_stats.threads_woken++;
387 BUG_ON(xprt->xpt_pool != pool);
388 wake_up(&rqstp->rq_wait);
390 dprintk("svc: transport %p put into queue\n", xprt);
391 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
392 pool->sp_stats.sockets_queued++;
393 BUG_ON(xprt->xpt_pool != pool);
397 spin_unlock_bh(&pool->sp_lock);
399 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
402 * Dequeue the first transport. Must be called with the pool->sp_lock held.
404 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
406 struct svc_xprt *xprt;
408 if (list_empty(&pool->sp_sockets))
411 xprt = list_entry(pool->sp_sockets.next,
412 struct svc_xprt, xpt_ready);
413 list_del_init(&xprt->xpt_ready);
415 dprintk("svc: transport %p dequeued, inuse=%d\n",
416 xprt, atomic_read(&xprt->xpt_ref.refcount));
422 * svc_xprt_received conditionally queues the transport for processing
423 * by another thread. The caller must hold the XPT_BUSY bit and must
424 * not thereafter touch transport data.
426 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
427 * insufficient) data.
429 void svc_xprt_received(struct svc_xprt *xprt)
431 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
432 xprt->xpt_pool = NULL;
433 clear_bit(XPT_BUSY, &xprt->xpt_flags);
434 svc_xprt_enqueue(xprt);
436 EXPORT_SYMBOL_GPL(svc_xprt_received);
439 * svc_reserve - change the space reserved for the reply to a request.
440 * @rqstp: The request in question
441 * @space: new max space to reserve
443 * Each request reserves some space on the output queue of the transport
444 * to make sure the reply fits. This function reduces that reserved
445 * space to be the amount of space used already, plus @space.
448 void svc_reserve(struct svc_rqst *rqstp, int space)
450 space += rqstp->rq_res.head[0].iov_len;
452 if (space < rqstp->rq_reserved) {
453 struct svc_xprt *xprt = rqstp->rq_xprt;
454 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
455 rqstp->rq_reserved = space;
457 svc_xprt_enqueue(xprt);
460 EXPORT_SYMBOL_GPL(svc_reserve);
462 static void svc_xprt_release(struct svc_rqst *rqstp)
464 struct svc_xprt *xprt = rqstp->rq_xprt;
466 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
468 kfree(rqstp->rq_deferred);
469 rqstp->rq_deferred = NULL;
471 svc_free_res_pages(rqstp);
472 rqstp->rq_res.page_len = 0;
473 rqstp->rq_res.page_base = 0;
475 /* Reset response buffer and release
477 * But first, check that enough space was reserved
478 * for the reply, otherwise we have a bug!
480 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
481 printk(KERN_ERR "RPC request reserved %d but used %d\n",
485 rqstp->rq_res.head[0].iov_len = 0;
486 svc_reserve(rqstp, 0);
487 rqstp->rq_xprt = NULL;
493 * External function to wake up a server waiting for data
494 * This really only makes sense for services like lockd
495 * which have exactly one thread anyway.
497 void svc_wake_up(struct svc_serv *serv)
499 struct svc_rqst *rqstp;
501 struct svc_pool *pool;
503 for (i = 0; i < serv->sv_nrpools; i++) {
504 pool = &serv->sv_pools[i];
506 spin_lock_bh(&pool->sp_lock);
507 if (!list_empty(&pool->sp_threads)) {
508 rqstp = list_entry(pool->sp_threads.next,
511 dprintk("svc: daemon %p woken up.\n", rqstp);
513 svc_thread_dequeue(pool, rqstp);
514 rqstp->rq_xprt = NULL;
516 wake_up(&rqstp->rq_wait);
518 spin_unlock_bh(&pool->sp_lock);
521 EXPORT_SYMBOL_GPL(svc_wake_up);
523 int svc_port_is_privileged(struct sockaddr *sin)
525 switch (sin->sa_family) {
527 return ntohs(((struct sockaddr_in *)sin)->sin_port)
530 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
538 * Make sure that we don't have too many active connections. If we have,
539 * something must be dropped. It's not clear what will happen if we allow
540 * "too many" connections, but when dealing with network-facing software,
541 * we have to code defensively. Here we do that by imposing hard limits.
543 * There's no point in trying to do random drop here for DoS
544 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
545 * attacker can easily beat that.
547 * The only somewhat efficient mechanism would be if drop old
548 * connections from the same IP first. But right now we don't even
549 * record the client IP in svc_sock.
551 * single-threaded services that expect a lot of clients will probably
552 * need to set sv_maxconn to override the default value which is based
553 * on the number of threads
555 static void svc_check_conn_limits(struct svc_serv *serv)
557 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
558 (serv->sv_nrthreads+3) * 20;
560 if (serv->sv_tmpcnt > limit) {
561 struct svc_xprt *xprt = NULL;
562 spin_lock_bh(&serv->sv_lock);
563 if (!list_empty(&serv->sv_tempsocks)) {
564 if (net_ratelimit()) {
565 /* Try to help the admin */
566 printk(KERN_NOTICE "%s: too many open "
567 "connections, consider increasing %s\n",
568 serv->sv_name, serv->sv_maxconn ?
569 "the max number of connections." :
570 "the number of threads.");
573 * Always select the oldest connection. It's not fair,
576 xprt = list_entry(serv->sv_tempsocks.prev,
579 set_bit(XPT_CLOSE, &xprt->xpt_flags);
582 spin_unlock_bh(&serv->sv_lock);
585 svc_xprt_enqueue(xprt);
592 * Receive the next request on any transport. This code is carefully
593 * organised not to touch any cachelines in the shared svc_serv
594 * structure, only cachelines in the local svc_pool.
596 int svc_recv(struct svc_rqst *rqstp, long timeout)
598 struct svc_xprt *xprt = NULL;
599 struct svc_serv *serv = rqstp->rq_server;
600 struct svc_pool *pool = rqstp->rq_pool;
604 DECLARE_WAITQUEUE(wait, current);
607 dprintk("svc: server %p waiting for data (to = %ld)\n",
612 "svc_recv: service %p, transport not NULL!\n",
614 if (waitqueue_active(&rqstp->rq_wait))
616 "svc_recv: service %p, wait queue active!\n",
619 /* now allocate needed pages. If we get a failure, sleep briefly */
620 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
621 for (i = 0; i < pages ; i++)
622 while (rqstp->rq_pages[i] == NULL) {
623 struct page *p = alloc_page(GFP_KERNEL);
625 set_current_state(TASK_INTERRUPTIBLE);
626 if (signalled() || kthread_should_stop()) {
627 set_current_state(TASK_RUNNING);
630 schedule_timeout(msecs_to_jiffies(500));
632 rqstp->rq_pages[i] = p;
634 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
635 BUG_ON(pages >= RPCSVC_MAXPAGES);
637 /* Make arg->head point to first page and arg->pages point to rest */
638 arg = &rqstp->rq_arg;
639 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
640 arg->head[0].iov_len = PAGE_SIZE;
641 arg->pages = rqstp->rq_pages + 1;
643 /* save at least one page for response */
644 arg->page_len = (pages-2)*PAGE_SIZE;
645 arg->len = (pages-1)*PAGE_SIZE;
646 arg->tail[0].iov_len = 0;
650 if (signalled() || kthread_should_stop())
653 spin_lock_bh(&pool->sp_lock);
654 if (rqstp->rq_waking) {
655 rqstp->rq_waking = 0;
657 BUG_ON(pool->sp_nwaking < 0);
659 xprt = svc_xprt_dequeue(pool);
661 rqstp->rq_xprt = xprt;
663 rqstp->rq_reserved = serv->sv_max_mesg;
664 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
666 /* No data pending. Go to sleep */
667 svc_thread_enqueue(pool, rqstp);
670 * We have to be able to interrupt this wait
671 * to bring down the daemons ...
673 set_current_state(TASK_INTERRUPTIBLE);
676 * checking kthread_should_stop() here allows us to avoid
677 * locking and signalling when stopping kthreads that call
678 * svc_recv. If the thread has already been woken up, then
679 * we can exit here without sleeping. If not, then it
680 * it'll be woken up quickly during the schedule_timeout
682 if (kthread_should_stop()) {
683 set_current_state(TASK_RUNNING);
684 spin_unlock_bh(&pool->sp_lock);
688 add_wait_queue(&rqstp->rq_wait, &wait);
689 spin_unlock_bh(&pool->sp_lock);
691 time_left = schedule_timeout(timeout);
695 spin_lock_bh(&pool->sp_lock);
696 remove_wait_queue(&rqstp->rq_wait, &wait);
698 pool->sp_stats.threads_timedout++;
700 xprt = rqstp->rq_xprt;
702 svc_thread_dequeue(pool, rqstp);
703 spin_unlock_bh(&pool->sp_lock);
704 dprintk("svc: server %p, no data yet\n", rqstp);
705 if (signalled() || kthread_should_stop())
711 spin_unlock_bh(&pool->sp_lock);
714 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
715 dprintk("svc_recv: found XPT_CLOSE\n");
716 svc_delete_xprt(xprt);
717 } else if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
718 struct svc_xprt *newxpt;
719 newxpt = xprt->xpt_ops->xpo_accept(xprt);
722 * We know this module_get will succeed because the
723 * listener holds a reference too
725 __module_get(newxpt->xpt_class->xcl_owner);
726 svc_check_conn_limits(xprt->xpt_server);
727 spin_lock_bh(&serv->sv_lock);
728 set_bit(XPT_TEMP, &newxpt->xpt_flags);
729 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
731 if (serv->sv_temptimer.function == NULL) {
732 /* setup timer to age temp transports */
733 setup_timer(&serv->sv_temptimer,
735 (unsigned long)serv);
736 mod_timer(&serv->sv_temptimer,
737 jiffies + svc_conn_age_period * HZ);
739 spin_unlock_bh(&serv->sv_lock);
740 svc_xprt_received(newxpt);
742 svc_xprt_received(xprt);
744 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
745 rqstp, pool->sp_id, xprt,
746 atomic_read(&xprt->xpt_ref.refcount));
747 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
748 if (rqstp->rq_deferred) {
749 svc_xprt_received(xprt);
750 len = svc_deferred_recv(rqstp);
752 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
753 dprintk("svc: got len=%d\n", len);
756 /* No data, incomplete (TCP) read, or accept() */
757 if (len == 0 || len == -EAGAIN) {
758 rqstp->rq_res.len = 0;
759 svc_xprt_release(rqstp);
762 clear_bit(XPT_OLD, &xprt->xpt_flags);
764 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
765 rqstp->rq_chandle.defer = svc_defer;
768 serv->sv_stats->netcnt++;
771 EXPORT_SYMBOL_GPL(svc_recv);
776 void svc_drop(struct svc_rqst *rqstp)
778 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
779 svc_xprt_release(rqstp);
781 EXPORT_SYMBOL_GPL(svc_drop);
784 * Return reply to client.
786 int svc_send(struct svc_rqst *rqstp)
788 struct svc_xprt *xprt;
792 xprt = rqstp->rq_xprt;
796 /* release the receive skb before sending the reply */
797 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
799 /* calculate over-all length */
801 xb->len = xb->head[0].iov_len +
805 /* Grab mutex to serialize outgoing data. */
806 mutex_lock(&xprt->xpt_mutex);
807 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
810 len = xprt->xpt_ops->xpo_sendto(rqstp);
811 mutex_unlock(&xprt->xpt_mutex);
812 rpc_wake_up(&xprt->xpt_bc_pending);
813 svc_xprt_release(rqstp);
815 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
821 * Timer function to close old temporary transports, using
822 * a mark-and-sweep algorithm.
824 static void svc_age_temp_xprts(unsigned long closure)
826 struct svc_serv *serv = (struct svc_serv *)closure;
827 struct svc_xprt *xprt;
828 struct list_head *le, *next;
829 LIST_HEAD(to_be_aged);
831 dprintk("svc_age_temp_xprts\n");
833 if (!spin_trylock_bh(&serv->sv_lock)) {
834 /* busy, try again 1 sec later */
835 dprintk("svc_age_temp_xprts: busy\n");
836 mod_timer(&serv->sv_temptimer, jiffies + HZ);
840 list_for_each_safe(le, next, &serv->sv_tempsocks) {
841 xprt = list_entry(le, struct svc_xprt, xpt_list);
843 /* First time through, just mark it OLD. Second time
844 * through, close it. */
845 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
847 if (atomic_read(&xprt->xpt_ref.refcount) > 1
848 || test_bit(XPT_BUSY, &xprt->xpt_flags))
851 list_move(le, &to_be_aged);
852 set_bit(XPT_CLOSE, &xprt->xpt_flags);
853 set_bit(XPT_DETACHED, &xprt->xpt_flags);
855 spin_unlock_bh(&serv->sv_lock);
857 while (!list_empty(&to_be_aged)) {
858 le = to_be_aged.next;
859 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
861 xprt = list_entry(le, struct svc_xprt, xpt_list);
863 dprintk("queuing xprt %p for closing\n", xprt);
865 /* a thread will dequeue and close it soon */
866 svc_xprt_enqueue(xprt);
870 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
874 * Remove a dead transport
876 void svc_delete_xprt(struct svc_xprt *xprt)
878 struct svc_serv *serv = xprt->xpt_server;
879 struct svc_deferred_req *dr;
881 /* Only do this once */
882 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
885 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
886 xprt->xpt_ops->xpo_detach(xprt);
888 spin_lock_bh(&serv->sv_lock);
889 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
890 list_del_init(&xprt->xpt_list);
892 * We used to delete the transport from whichever list
893 * it's sk_xprt.xpt_ready node was on, but we don't actually
894 * need to. This is because the only time we're called
895 * while still attached to a queue, the queue itself
896 * is about to be destroyed (in svc_destroy).
898 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
901 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
905 spin_unlock_bh(&serv->sv_lock);
908 void svc_close_xprt(struct svc_xprt *xprt)
910 set_bit(XPT_CLOSE, &xprt->xpt_flags);
911 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
912 /* someone else will have to effect the close */
916 svc_delete_xprt(xprt);
917 clear_bit(XPT_BUSY, &xprt->xpt_flags);
920 EXPORT_SYMBOL_GPL(svc_close_xprt);
922 void svc_close_all(struct list_head *xprt_list)
924 struct svc_xprt *xprt;
925 struct svc_xprt *tmp;
927 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
928 set_bit(XPT_CLOSE, &xprt->xpt_flags);
929 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
930 /* Waiting to be processed, but no threads left,
931 * So just remove it from the waiting list
933 list_del_init(&xprt->xpt_ready);
934 clear_bit(XPT_BUSY, &xprt->xpt_flags);
936 svc_close_xprt(xprt);
941 * Handle defer and revisit of requests
944 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
946 struct svc_deferred_req *dr =
947 container_of(dreq, struct svc_deferred_req, handle);
948 struct svc_xprt *xprt = dr->xprt;
950 spin_lock(&xprt->xpt_lock);
951 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
952 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
953 spin_unlock(&xprt->xpt_lock);
954 dprintk("revisit canceled\n");
959 dprintk("revisit queued\n");
961 list_add(&dr->handle.recent, &xprt->xpt_deferred);
962 spin_unlock(&xprt->xpt_lock);
963 svc_xprt_enqueue(xprt);
968 * Save the request off for later processing. The request buffer looks
971 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
973 * This code can only handle requests that consist of an xprt-header
976 static struct cache_deferred_req *svc_defer(struct cache_req *req)
978 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
979 struct svc_deferred_req *dr;
981 if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
982 return NULL; /* if more than a page, give up FIXME */
983 if (rqstp->rq_deferred) {
984 dr = rqstp->rq_deferred;
985 rqstp->rq_deferred = NULL;
989 /* FIXME maybe discard if size too large */
990 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
991 dr = kmalloc(size, GFP_KERNEL);
995 dr->handle.owner = rqstp->rq_server;
996 dr->prot = rqstp->rq_prot;
997 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
998 dr->addrlen = rqstp->rq_addrlen;
999 dr->daddr = rqstp->rq_daddr;
1000 dr->argslen = rqstp->rq_arg.len >> 2;
1001 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1003 /* back up head to the start of the buffer and copy */
1004 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1005 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1008 svc_xprt_get(rqstp->rq_xprt);
1009 dr->xprt = rqstp->rq_xprt;
1011 dr->handle.revisit = svc_revisit;
1016 * recv data from a deferred request into an active one
1018 static int svc_deferred_recv(struct svc_rqst *rqstp)
1020 struct svc_deferred_req *dr = rqstp->rq_deferred;
1022 /* setup iov_base past transport header */
1023 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1024 /* The iov_len does not include the transport header bytes */
1025 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1026 rqstp->rq_arg.page_len = 0;
1027 /* The rq_arg.len includes the transport header bytes */
1028 rqstp->rq_arg.len = dr->argslen<<2;
1029 rqstp->rq_prot = dr->prot;
1030 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1031 rqstp->rq_addrlen = dr->addrlen;
1032 /* Save off transport header len in case we get deferred again */
1033 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1034 rqstp->rq_daddr = dr->daddr;
1035 rqstp->rq_respages = rqstp->rq_pages;
1036 return (dr->argslen<<2) - dr->xprt_hlen;
1040 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1042 struct svc_deferred_req *dr = NULL;
1044 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1046 spin_lock(&xprt->xpt_lock);
1047 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1048 if (!list_empty(&xprt->xpt_deferred)) {
1049 dr = list_entry(xprt->xpt_deferred.next,
1050 struct svc_deferred_req,
1052 list_del_init(&dr->handle.recent);
1053 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1055 spin_unlock(&xprt->xpt_lock);
1060 * svc_find_xprt - find an RPC transport instance
1061 * @serv: pointer to svc_serv to search
1062 * @xcl_name: C string containing transport's class name
1063 * @af: Address family of transport's local address
1064 * @port: transport's IP port number
1066 * Return the transport instance pointer for the endpoint accepting
1067 * connections/peer traffic from the specified transport class,
1068 * address family and port.
1070 * Specifying 0 for the address family or port is effectively a
1071 * wild-card, and will result in matching the first transport in the
1072 * service's list that has a matching class name.
1074 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1075 const sa_family_t af, const unsigned short port)
1077 struct svc_xprt *xprt;
1078 struct svc_xprt *found = NULL;
1080 /* Sanity check the args */
1081 if (serv == NULL || xcl_name == NULL)
1084 spin_lock_bh(&serv->sv_lock);
1085 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1086 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1088 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1090 if (port != 0 && port != svc_xprt_local_port(xprt))
1096 spin_unlock_bh(&serv->sv_lock);
1099 EXPORT_SYMBOL_GPL(svc_find_xprt);
1101 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1102 char *pos, int remaining)
1106 len = snprintf(pos, remaining, "%s %u\n",
1107 xprt->xpt_class->xcl_name,
1108 svc_xprt_local_port(xprt));
1109 if (len >= remaining)
1110 return -ENAMETOOLONG;
1115 * svc_xprt_names - format a buffer with a list of transport names
1116 * @serv: pointer to an RPC service
1117 * @buf: pointer to a buffer to be filled in
1118 * @buflen: length of buffer to be filled in
1120 * Fills in @buf with a string containing a list of transport names,
1121 * each name terminated with '\n'.
1123 * Returns positive length of the filled-in string on success; otherwise
1124 * a negative errno value is returned if an error occurs.
1126 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1128 struct svc_xprt *xprt;
1132 /* Sanity check args */
1136 spin_lock_bh(&serv->sv_lock);
1140 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1141 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1153 spin_unlock_bh(&serv->sv_lock);
1156 EXPORT_SYMBOL_GPL(svc_xprt_names);
1159 /*----------------------------------------------------------------------------*/
1161 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1163 unsigned int pidx = (unsigned int)*pos;
1164 struct svc_serv *serv = m->private;
1166 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1169 return SEQ_START_TOKEN;
1170 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1173 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1175 struct svc_pool *pool = p;
1176 struct svc_serv *serv = m->private;
1178 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1180 if (p == SEQ_START_TOKEN) {
1181 pool = &serv->sv_pools[0];
1183 unsigned int pidx = (pool - &serv->sv_pools[0]);
1184 if (pidx < serv->sv_nrpools-1)
1185 pool = &serv->sv_pools[pidx+1];
1193 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1197 static int svc_pool_stats_show(struct seq_file *m, void *p)
1199 struct svc_pool *pool = p;
1201 if (p == SEQ_START_TOKEN) {
1202 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken overloads-avoided threads-timedout\n");
1206 seq_printf(m, "%u %lu %lu %lu %lu %lu\n",
1208 pool->sp_stats.packets,
1209 pool->sp_stats.sockets_queued,
1210 pool->sp_stats.threads_woken,
1211 pool->sp_stats.overloads_avoided,
1212 pool->sp_stats.threads_timedout);
1217 static const struct seq_operations svc_pool_stats_seq_ops = {
1218 .start = svc_pool_stats_start,
1219 .next = svc_pool_stats_next,
1220 .stop = svc_pool_stats_stop,
1221 .show = svc_pool_stats_show,
1224 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1228 err = seq_open(file, &svc_pool_stats_seq_ops);
1230 ((struct seq_file *) file->private_data)->private = serv;
1233 EXPORT_SYMBOL(svc_pool_stats_open);
1235 /*----------------------------------------------------------------------------*/