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>
12 #include <linux/sunrpc/stats.h>
13 #include <linux/sunrpc/svc_xprt.h>
15 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
17 #define SVC_MAX_WAKING 5
19 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
20 static int svc_deferred_recv(struct svc_rqst *rqstp);
21 static struct cache_deferred_req *svc_defer(struct cache_req *req);
22 static void svc_age_temp_xprts(unsigned long closure);
24 /* apparently the "standard" is that clients close
25 * idle connections after 5 minutes, servers after
27 * http://www.connectathon.org/talks96/nfstcp.pdf
29 static int svc_conn_age_period = 6*60;
31 /* List of registered transport classes */
32 static DEFINE_SPINLOCK(svc_xprt_class_lock);
33 static LIST_HEAD(svc_xprt_class_list);
35 /* SMP locking strategy:
37 * svc_pool->sp_lock protects most of the fields of that pool.
38 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
39 * when both need to be taken (rare), svc_serv->sv_lock is first.
40 * BKL protects svc_serv->sv_nrthread.
41 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
42 * and the ->sk_info_authunix cache.
44 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
45 * enqueued multiply. During normal transport processing this bit
46 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
47 * Providers should not manipulate this bit directly.
49 * Some flags can be set to certain values at any time
50 * providing that certain rules are followed:
53 * - Can be set or cleared at any time.
54 * - After a set, svc_xprt_enqueue must be called to enqueue
55 * the transport for processing.
56 * - After a clear, the transport must be read/accepted.
57 * If this succeeds, it must be set again.
59 * - Can set at any time. It is never cleared.
61 * - Can only be set while XPT_BUSY is held which ensures
62 * that no other thread will be using the transport or will
63 * try to set XPT_DEAD.
66 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
68 struct svc_xprt_class *cl;
71 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
73 INIT_LIST_HEAD(&xcl->xcl_list);
74 spin_lock(&svc_xprt_class_lock);
75 /* Make sure there isn't already a class with the same name */
76 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
77 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
80 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
83 spin_unlock(&svc_xprt_class_lock);
86 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
88 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
90 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
91 spin_lock(&svc_xprt_class_lock);
92 list_del_init(&xcl->xcl_list);
93 spin_unlock(&svc_xprt_class_lock);
95 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
98 * Format the transport list for printing
100 int svc_print_xprts(char *buf, int maxlen)
102 struct list_head *le;
107 spin_lock(&svc_xprt_class_lock);
108 list_for_each(le, &svc_xprt_class_list) {
110 struct svc_xprt_class *xcl =
111 list_entry(le, struct svc_xprt_class, 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 && xprt->xpt_auth_cache != NULL)
132 svcauth_unix_info_release(xprt->xpt_auth_cache);
133 xprt->xpt_ops->xpo_free(xprt);
137 void svc_xprt_put(struct svc_xprt *xprt)
139 kref_put(&xprt->xpt_ref, svc_xprt_free);
141 EXPORT_SYMBOL_GPL(svc_xprt_put);
144 * Called by transport drivers to initialize the transport independent
145 * portion of the transport instance.
147 void svc_xprt_init(struct svc_xprt_class *xcl, struct svc_xprt *xprt,
148 struct svc_serv *serv)
150 memset(xprt, 0, sizeof(*xprt));
151 xprt->xpt_class = xcl;
152 xprt->xpt_ops = xcl->xcl_ops;
153 kref_init(&xprt->xpt_ref);
154 xprt->xpt_server = serv;
155 INIT_LIST_HEAD(&xprt->xpt_list);
156 INIT_LIST_HEAD(&xprt->xpt_ready);
157 INIT_LIST_HEAD(&xprt->xpt_deferred);
158 mutex_init(&xprt->xpt_mutex);
159 spin_lock_init(&xprt->xpt_lock);
160 set_bit(XPT_BUSY, &xprt->xpt_flags);
162 EXPORT_SYMBOL_GPL(svc_xprt_init);
164 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
165 struct svc_serv *serv,
167 const unsigned short port,
170 struct sockaddr_in sin = {
171 .sin_family = AF_INET,
172 .sin_addr.s_addr = htonl(INADDR_ANY),
173 .sin_port = htons(port),
175 struct sockaddr_in6 sin6 = {
176 .sin6_family = AF_INET6,
177 .sin6_addr = IN6ADDR_ANY_INIT,
178 .sin6_port = htons(port),
180 struct sockaddr *sap;
185 sap = (struct sockaddr *)&sin;
189 sap = (struct sockaddr *)&sin6;
193 return ERR_PTR(-EAFNOSUPPORT);
196 return xcl->xcl_ops->xpo_create(serv, sap, len, flags);
199 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
200 const int family, const unsigned short port,
203 struct svc_xprt_class *xcl;
205 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
206 spin_lock(&svc_xprt_class_lock);
207 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
208 struct svc_xprt *newxprt;
210 if (strcmp(xprt_name, xcl->xcl_name))
213 if (!try_module_get(xcl->xcl_owner))
216 spin_unlock(&svc_xprt_class_lock);
217 newxprt = __svc_xpo_create(xcl, serv, family, port, flags);
218 if (IS_ERR(newxprt)) {
219 module_put(xcl->xcl_owner);
220 return PTR_ERR(newxprt);
223 clear_bit(XPT_TEMP, &newxprt->xpt_flags);
224 spin_lock_bh(&serv->sv_lock);
225 list_add(&newxprt->xpt_list, &serv->sv_permsocks);
226 spin_unlock_bh(&serv->sv_lock);
227 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
228 return svc_xprt_local_port(newxprt);
231 spin_unlock(&svc_xprt_class_lock);
232 dprintk("svc: transport %s not found\n", xprt_name);
235 EXPORT_SYMBOL_GPL(svc_create_xprt);
238 * Copy the local and remote xprt addresses to the rqstp structure
240 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
242 struct sockaddr *sin;
244 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
245 rqstp->rq_addrlen = xprt->xpt_remotelen;
248 * Destination address in request is needed for binding the
249 * source address in RPC replies/callbacks later.
251 sin = (struct sockaddr *)&xprt->xpt_local;
252 switch (sin->sa_family) {
254 rqstp->rq_daddr.addr = ((struct sockaddr_in *)sin)->sin_addr;
257 rqstp->rq_daddr.addr6 = ((struct sockaddr_in6 *)sin)->sin6_addr;
261 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
264 * svc_print_addr - Format rq_addr field for printing
265 * @rqstp: svc_rqst struct containing address to print
266 * @buf: target buffer for formatted address
267 * @len: length of target buffer
270 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
272 return __svc_print_addr(svc_addr(rqstp), buf, len);
274 EXPORT_SYMBOL_GPL(svc_print_addr);
277 * Queue up an idle server thread. Must have pool->sp_lock held.
278 * Note: this is really a stack rather than a queue, so that we only
279 * use as many different threads as we need, and the rest don't pollute
282 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
284 list_add(&rqstp->rq_list, &pool->sp_threads);
288 * Dequeue an nfsd thread. Must have pool->sp_lock held.
290 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
292 list_del(&rqstp->rq_list);
296 * Queue up a transport with data pending. If there are idle nfsd
297 * processes, wake 'em up.
300 void svc_xprt_enqueue(struct svc_xprt *xprt)
302 struct svc_serv *serv = xprt->xpt_server;
303 struct svc_pool *pool;
304 struct svc_rqst *rqstp;
308 if (!(xprt->xpt_flags &
309 ((1<<XPT_CONN)|(1<<XPT_DATA)|(1<<XPT_CLOSE)|(1<<XPT_DEFERRED))))
313 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
316 spin_lock_bh(&pool->sp_lock);
318 if (test_bit(XPT_DEAD, &xprt->xpt_flags)) {
319 /* Don't enqueue dead transports */
320 dprintk("svc: transport %p is dead, not enqueued\n", xprt);
324 pool->sp_stats.packets++;
326 /* Mark transport as busy. It will remain in this state until
327 * the provider calls svc_xprt_received. We update XPT_BUSY
328 * atomically because it also guards against trying to enqueue
329 * the transport twice.
331 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
332 /* Don't enqueue transport while already enqueued */
333 dprintk("svc: transport %p busy, not enqueued\n", xprt);
336 BUG_ON(xprt->xpt_pool != NULL);
337 xprt->xpt_pool = pool;
339 /* Handle pending connection */
340 if (test_bit(XPT_CONN, &xprt->xpt_flags))
343 /* Handle close in-progress */
344 if (test_bit(XPT_CLOSE, &xprt->xpt_flags))
347 /* Check if we have space to reply to a request */
348 if (!xprt->xpt_ops->xpo_has_wspace(xprt)) {
349 /* Don't enqueue while not enough space for reply */
350 dprintk("svc: no write space, transport %p not enqueued\n",
352 xprt->xpt_pool = NULL;
353 clear_bit(XPT_BUSY, &xprt->xpt_flags);
358 /* Work out whether threads are available */
359 thread_avail = !list_empty(&pool->sp_threads); /* threads are asleep */
360 if (pool->sp_nwaking >= SVC_MAX_WAKING) {
361 /* too many threads are runnable and trying to wake up */
363 pool->sp_stats.overloads_avoided++;
367 rqstp = list_entry(pool->sp_threads.next,
370 dprintk("svc: transport %p served by daemon %p\n",
372 svc_thread_dequeue(pool, rqstp);
375 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
376 rqstp, rqstp->rq_xprt);
377 rqstp->rq_xprt = xprt;
379 rqstp->rq_reserved = serv->sv_max_mesg;
380 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
381 rqstp->rq_waking = 1;
383 pool->sp_stats.threads_woken++;
384 BUG_ON(xprt->xpt_pool != pool);
385 wake_up(&rqstp->rq_wait);
387 dprintk("svc: transport %p put into queue\n", xprt);
388 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
389 pool->sp_stats.sockets_queued++;
390 BUG_ON(xprt->xpt_pool != pool);
394 spin_unlock_bh(&pool->sp_lock);
396 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
399 * Dequeue the first transport. Must be called with the pool->sp_lock held.
401 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
403 struct svc_xprt *xprt;
405 if (list_empty(&pool->sp_sockets))
408 xprt = list_entry(pool->sp_sockets.next,
409 struct svc_xprt, xpt_ready);
410 list_del_init(&xprt->xpt_ready);
412 dprintk("svc: transport %p dequeued, inuse=%d\n",
413 xprt, atomic_read(&xprt->xpt_ref.refcount));
419 * svc_xprt_received conditionally queues the transport for processing
420 * by another thread. The caller must hold the XPT_BUSY bit and must
421 * not thereafter touch transport data.
423 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
424 * insufficient) data.
426 void svc_xprt_received(struct svc_xprt *xprt)
428 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
429 xprt->xpt_pool = NULL;
430 clear_bit(XPT_BUSY, &xprt->xpt_flags);
431 svc_xprt_enqueue(xprt);
433 EXPORT_SYMBOL_GPL(svc_xprt_received);
436 * svc_reserve - change the space reserved for the reply to a request.
437 * @rqstp: The request in question
438 * @space: new max space to reserve
440 * Each request reserves some space on the output queue of the transport
441 * to make sure the reply fits. This function reduces that reserved
442 * space to be the amount of space used already, plus @space.
445 void svc_reserve(struct svc_rqst *rqstp, int space)
447 space += rqstp->rq_res.head[0].iov_len;
449 if (space < rqstp->rq_reserved) {
450 struct svc_xprt *xprt = rqstp->rq_xprt;
451 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
452 rqstp->rq_reserved = space;
454 svc_xprt_enqueue(xprt);
457 EXPORT_SYMBOL_GPL(svc_reserve);
459 static void svc_xprt_release(struct svc_rqst *rqstp)
461 struct svc_xprt *xprt = rqstp->rq_xprt;
463 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
465 kfree(rqstp->rq_deferred);
466 rqstp->rq_deferred = NULL;
468 svc_free_res_pages(rqstp);
469 rqstp->rq_res.page_len = 0;
470 rqstp->rq_res.page_base = 0;
472 /* Reset response buffer and release
474 * But first, check that enough space was reserved
475 * for the reply, otherwise we have a bug!
477 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
478 printk(KERN_ERR "RPC request reserved %d but used %d\n",
482 rqstp->rq_res.head[0].iov_len = 0;
483 svc_reserve(rqstp, 0);
484 rqstp->rq_xprt = NULL;
490 * External function to wake up a server waiting for data
491 * This really only makes sense for services like lockd
492 * which have exactly one thread anyway.
494 void svc_wake_up(struct svc_serv *serv)
496 struct svc_rqst *rqstp;
498 struct svc_pool *pool;
500 for (i = 0; i < serv->sv_nrpools; i++) {
501 pool = &serv->sv_pools[i];
503 spin_lock_bh(&pool->sp_lock);
504 if (!list_empty(&pool->sp_threads)) {
505 rqstp = list_entry(pool->sp_threads.next,
508 dprintk("svc: daemon %p woken up.\n", rqstp);
510 svc_thread_dequeue(pool, rqstp);
511 rqstp->rq_xprt = NULL;
513 wake_up(&rqstp->rq_wait);
515 spin_unlock_bh(&pool->sp_lock);
518 EXPORT_SYMBOL_GPL(svc_wake_up);
520 int svc_port_is_privileged(struct sockaddr *sin)
522 switch (sin->sa_family) {
524 return ntohs(((struct sockaddr_in *)sin)->sin_port)
527 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
535 * Make sure that we don't have too many active connections. If we have,
536 * something must be dropped. It's not clear what will happen if we allow
537 * "too many" connections, but when dealing with network-facing software,
538 * we have to code defensively. Here we do that by imposing hard limits.
540 * There's no point in trying to do random drop here for DoS
541 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
542 * attacker can easily beat that.
544 * The only somewhat efficient mechanism would be if drop old
545 * connections from the same IP first. But right now we don't even
546 * record the client IP in svc_sock.
548 * single-threaded services that expect a lot of clients will probably
549 * need to set sv_maxconn to override the default value which is based
550 * on the number of threads
552 static void svc_check_conn_limits(struct svc_serv *serv)
554 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
555 (serv->sv_nrthreads+3) * 20;
557 if (serv->sv_tmpcnt > limit) {
558 struct svc_xprt *xprt = NULL;
559 spin_lock_bh(&serv->sv_lock);
560 if (!list_empty(&serv->sv_tempsocks)) {
561 if (net_ratelimit()) {
562 /* Try to help the admin */
563 printk(KERN_NOTICE "%s: too many open "
564 "connections, consider increasing %s\n",
565 serv->sv_name, serv->sv_maxconn ?
566 "the max number of connections." :
567 "the number of threads.");
570 * Always select the oldest connection. It's not fair,
573 xprt = list_entry(serv->sv_tempsocks.prev,
576 set_bit(XPT_CLOSE, &xprt->xpt_flags);
579 spin_unlock_bh(&serv->sv_lock);
582 svc_xprt_enqueue(xprt);
589 * Receive the next request on any transport. This code is carefully
590 * organised not to touch any cachelines in the shared svc_serv
591 * structure, only cachelines in the local svc_pool.
593 int svc_recv(struct svc_rqst *rqstp, long timeout)
595 struct svc_xprt *xprt = NULL;
596 struct svc_serv *serv = rqstp->rq_server;
597 struct svc_pool *pool = rqstp->rq_pool;
601 DECLARE_WAITQUEUE(wait, current);
604 dprintk("svc: server %p waiting for data (to = %ld)\n",
609 "svc_recv: service %p, transport not NULL!\n",
611 if (waitqueue_active(&rqstp->rq_wait))
613 "svc_recv: service %p, wait queue active!\n",
616 /* now allocate needed pages. If we get a failure, sleep briefly */
617 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
618 for (i = 0; i < pages ; i++)
619 while (rqstp->rq_pages[i] == NULL) {
620 struct page *p = alloc_page(GFP_KERNEL);
622 set_current_state(TASK_INTERRUPTIBLE);
623 if (signalled() || kthread_should_stop()) {
624 set_current_state(TASK_RUNNING);
627 schedule_timeout(msecs_to_jiffies(500));
629 rqstp->rq_pages[i] = p;
631 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
632 BUG_ON(pages >= RPCSVC_MAXPAGES);
634 /* Make arg->head point to first page and arg->pages point to rest */
635 arg = &rqstp->rq_arg;
636 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
637 arg->head[0].iov_len = PAGE_SIZE;
638 arg->pages = rqstp->rq_pages + 1;
640 /* save at least one page for response */
641 arg->page_len = (pages-2)*PAGE_SIZE;
642 arg->len = (pages-1)*PAGE_SIZE;
643 arg->tail[0].iov_len = 0;
647 if (signalled() || kthread_should_stop())
650 spin_lock_bh(&pool->sp_lock);
651 if (rqstp->rq_waking) {
652 rqstp->rq_waking = 0;
654 BUG_ON(pool->sp_nwaking < 0);
656 xprt = svc_xprt_dequeue(pool);
658 rqstp->rq_xprt = xprt;
660 rqstp->rq_reserved = serv->sv_max_mesg;
661 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
663 /* No data pending. Go to sleep */
664 svc_thread_enqueue(pool, rqstp);
667 * We have to be able to interrupt this wait
668 * to bring down the daemons ...
670 set_current_state(TASK_INTERRUPTIBLE);
673 * checking kthread_should_stop() here allows us to avoid
674 * locking and signalling when stopping kthreads that call
675 * svc_recv. If the thread has already been woken up, then
676 * we can exit here without sleeping. If not, then it
677 * it'll be woken up quickly during the schedule_timeout
679 if (kthread_should_stop()) {
680 set_current_state(TASK_RUNNING);
681 spin_unlock_bh(&pool->sp_lock);
685 add_wait_queue(&rqstp->rq_wait, &wait);
686 spin_unlock_bh(&pool->sp_lock);
688 time_left = schedule_timeout(timeout);
692 spin_lock_bh(&pool->sp_lock);
693 remove_wait_queue(&rqstp->rq_wait, &wait);
695 pool->sp_stats.threads_timedout++;
697 xprt = rqstp->rq_xprt;
699 svc_thread_dequeue(pool, rqstp);
700 spin_unlock_bh(&pool->sp_lock);
701 dprintk("svc: server %p, no data yet\n", rqstp);
702 if (signalled() || kthread_should_stop())
708 spin_unlock_bh(&pool->sp_lock);
711 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
712 dprintk("svc_recv: found XPT_CLOSE\n");
713 svc_delete_xprt(xprt);
714 } else if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
715 struct svc_xprt *newxpt;
716 newxpt = xprt->xpt_ops->xpo_accept(xprt);
719 * We know this module_get will succeed because the
720 * listener holds a reference too
722 __module_get(newxpt->xpt_class->xcl_owner);
723 svc_check_conn_limits(xprt->xpt_server);
724 spin_lock_bh(&serv->sv_lock);
725 set_bit(XPT_TEMP, &newxpt->xpt_flags);
726 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
728 if (serv->sv_temptimer.function == NULL) {
729 /* setup timer to age temp transports */
730 setup_timer(&serv->sv_temptimer,
732 (unsigned long)serv);
733 mod_timer(&serv->sv_temptimer,
734 jiffies + svc_conn_age_period * HZ);
736 spin_unlock_bh(&serv->sv_lock);
737 svc_xprt_received(newxpt);
739 svc_xprt_received(xprt);
741 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
742 rqstp, pool->sp_id, xprt,
743 atomic_read(&xprt->xpt_ref.refcount));
744 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
745 if (rqstp->rq_deferred) {
746 svc_xprt_received(xprt);
747 len = svc_deferred_recv(rqstp);
749 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
750 dprintk("svc: got len=%d\n", len);
753 /* No data, incomplete (TCP) read, or accept() */
754 if (len == 0 || len == -EAGAIN) {
755 rqstp->rq_res.len = 0;
756 svc_xprt_release(rqstp);
759 clear_bit(XPT_OLD, &xprt->xpt_flags);
761 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
762 rqstp->rq_chandle.defer = svc_defer;
765 serv->sv_stats->netcnt++;
768 EXPORT_SYMBOL_GPL(svc_recv);
773 void svc_drop(struct svc_rqst *rqstp)
775 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
776 svc_xprt_release(rqstp);
778 EXPORT_SYMBOL_GPL(svc_drop);
781 * Return reply to client.
783 int svc_send(struct svc_rqst *rqstp)
785 struct svc_xprt *xprt;
789 xprt = rqstp->rq_xprt;
793 /* release the receive skb before sending the reply */
794 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
796 /* calculate over-all length */
798 xb->len = xb->head[0].iov_len +
802 /* Grab mutex to serialize outgoing data. */
803 mutex_lock(&xprt->xpt_mutex);
804 if (test_bit(XPT_DEAD, &xprt->xpt_flags))
807 len = xprt->xpt_ops->xpo_sendto(rqstp);
808 mutex_unlock(&xprt->xpt_mutex);
809 svc_xprt_release(rqstp);
811 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
817 * Timer function to close old temporary transports, using
818 * a mark-and-sweep algorithm.
820 static void svc_age_temp_xprts(unsigned long closure)
822 struct svc_serv *serv = (struct svc_serv *)closure;
823 struct svc_xprt *xprt;
824 struct list_head *le, *next;
825 LIST_HEAD(to_be_aged);
827 dprintk("svc_age_temp_xprts\n");
829 if (!spin_trylock_bh(&serv->sv_lock)) {
830 /* busy, try again 1 sec later */
831 dprintk("svc_age_temp_xprts: busy\n");
832 mod_timer(&serv->sv_temptimer, jiffies + HZ);
836 list_for_each_safe(le, next, &serv->sv_tempsocks) {
837 xprt = list_entry(le, struct svc_xprt, xpt_list);
839 /* First time through, just mark it OLD. Second time
840 * through, close it. */
841 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
843 if (atomic_read(&xprt->xpt_ref.refcount) > 1
844 || test_bit(XPT_BUSY, &xprt->xpt_flags))
847 list_move(le, &to_be_aged);
848 set_bit(XPT_CLOSE, &xprt->xpt_flags);
849 set_bit(XPT_DETACHED, &xprt->xpt_flags);
851 spin_unlock_bh(&serv->sv_lock);
853 while (!list_empty(&to_be_aged)) {
854 le = to_be_aged.next;
855 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
857 xprt = list_entry(le, struct svc_xprt, xpt_list);
859 dprintk("queuing xprt %p for closing\n", xprt);
861 /* a thread will dequeue and close it soon */
862 svc_xprt_enqueue(xprt);
866 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
870 * Remove a dead transport
872 void svc_delete_xprt(struct svc_xprt *xprt)
874 struct svc_serv *serv = xprt->xpt_server;
875 struct svc_deferred_req *dr;
877 /* Only do this once */
878 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
881 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
882 xprt->xpt_ops->xpo_detach(xprt);
884 spin_lock_bh(&serv->sv_lock);
885 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
886 list_del_init(&xprt->xpt_list);
888 * We used to delete the transport from whichever list
889 * it's sk_xprt.xpt_ready node was on, but we don't actually
890 * need to. This is because the only time we're called
891 * while still attached to a queue, the queue itself
892 * is about to be destroyed (in svc_destroy).
894 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
897 for (dr = svc_deferred_dequeue(xprt); dr;
898 dr = svc_deferred_dequeue(xprt)) {
904 spin_unlock_bh(&serv->sv_lock);
907 void svc_close_xprt(struct svc_xprt *xprt)
909 set_bit(XPT_CLOSE, &xprt->xpt_flags);
910 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
911 /* someone else will have to effect the close */
915 svc_delete_xprt(xprt);
916 clear_bit(XPT_BUSY, &xprt->xpt_flags);
919 EXPORT_SYMBOL_GPL(svc_close_xprt);
921 void svc_close_all(struct list_head *xprt_list)
923 struct svc_xprt *xprt;
924 struct svc_xprt *tmp;
926 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
927 set_bit(XPT_CLOSE, &xprt->xpt_flags);
928 if (test_bit(XPT_BUSY, &xprt->xpt_flags)) {
929 /* Waiting to be processed, but no threads left,
930 * So just remove it from the waiting list
932 list_del_init(&xprt->xpt_ready);
933 clear_bit(XPT_BUSY, &xprt->xpt_flags);
935 svc_close_xprt(xprt);
940 * Handle defer and revisit of requests
943 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
945 struct svc_deferred_req *dr =
946 container_of(dreq, struct svc_deferred_req, handle);
947 struct svc_xprt *xprt = dr->xprt;
949 spin_lock(&xprt->xpt_lock);
950 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
951 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
952 spin_unlock(&xprt->xpt_lock);
953 dprintk("revisit canceled\n");
958 dprintk("revisit queued\n");
960 list_add(&dr->handle.recent, &xprt->xpt_deferred);
961 spin_unlock(&xprt->xpt_lock);
962 svc_xprt_enqueue(xprt);
967 * Save the request off for later processing. The request buffer looks
970 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
972 * This code can only handle requests that consist of an xprt-header
975 static struct cache_deferred_req *svc_defer(struct cache_req *req)
977 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
978 struct svc_deferred_req *dr;
980 if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
981 return NULL; /* if more than a page, give up FIXME */
982 if (rqstp->rq_deferred) {
983 dr = rqstp->rq_deferred;
984 rqstp->rq_deferred = NULL;
988 /* FIXME maybe discard if size too large */
989 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
990 dr = kmalloc(size, GFP_KERNEL);
994 dr->handle.owner = rqstp->rq_server;
995 dr->prot = rqstp->rq_prot;
996 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
997 dr->addrlen = rqstp->rq_addrlen;
998 dr->daddr = rqstp->rq_daddr;
999 dr->argslen = rqstp->rq_arg.len >> 2;
1000 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1002 /* back up head to the start of the buffer and copy */
1003 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1004 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1007 svc_xprt_get(rqstp->rq_xprt);
1008 dr->xprt = rqstp->rq_xprt;
1010 dr->handle.revisit = svc_revisit;
1015 * recv data from a deferred request into an active one
1017 static int svc_deferred_recv(struct svc_rqst *rqstp)
1019 struct svc_deferred_req *dr = rqstp->rq_deferred;
1021 /* setup iov_base past transport header */
1022 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1023 /* The iov_len does not include the transport header bytes */
1024 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1025 rqstp->rq_arg.page_len = 0;
1026 /* The rq_arg.len includes the transport header bytes */
1027 rqstp->rq_arg.len = dr->argslen<<2;
1028 rqstp->rq_prot = dr->prot;
1029 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1030 rqstp->rq_addrlen = dr->addrlen;
1031 /* Save off transport header len in case we get deferred again */
1032 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1033 rqstp->rq_daddr = dr->daddr;
1034 rqstp->rq_respages = rqstp->rq_pages;
1035 return (dr->argslen<<2) - dr->xprt_hlen;
1039 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1041 struct svc_deferred_req *dr = NULL;
1043 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1045 spin_lock(&xprt->xpt_lock);
1046 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1047 if (!list_empty(&xprt->xpt_deferred)) {
1048 dr = list_entry(xprt->xpt_deferred.next,
1049 struct svc_deferred_req,
1051 list_del_init(&dr->handle.recent);
1052 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1054 spin_unlock(&xprt->xpt_lock);
1059 * svc_find_xprt - find an RPC transport instance
1060 * @serv: pointer to svc_serv to search
1061 * @xcl_name: C string containing transport's class name
1062 * @af: Address family of transport's local address
1063 * @port: transport's IP port number
1065 * Return the transport instance pointer for the endpoint accepting
1066 * connections/peer traffic from the specified transport class,
1067 * address family and port.
1069 * Specifying 0 for the address family or port is effectively a
1070 * wild-card, and will result in matching the first transport in the
1071 * service's list that has a matching class name.
1073 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1074 const sa_family_t af, const unsigned short port)
1076 struct svc_xprt *xprt;
1077 struct svc_xprt *found = NULL;
1079 /* Sanity check the args */
1080 if (serv == NULL || xcl_name == NULL)
1083 spin_lock_bh(&serv->sv_lock);
1084 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1085 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1087 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1089 if (port != 0 && port != svc_xprt_local_port(xprt))
1095 spin_unlock_bh(&serv->sv_lock);
1098 EXPORT_SYMBOL_GPL(svc_find_xprt);
1101 * Format a buffer with a list of the active transports. A zero for
1102 * the buflen parameter disables target buffer overflow checking.
1104 int svc_xprt_names(struct svc_serv *serv, char *buf, int buflen)
1106 struct svc_xprt *xprt;
1111 /* Sanity check args */
1115 spin_lock_bh(&serv->sv_lock);
1116 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1117 len = snprintf(xprt_str, sizeof(xprt_str),
1118 "%s %d\n", xprt->xpt_class->xcl_name,
1119 svc_xprt_local_port(xprt));
1120 /* If the string was truncated, replace with error string */
1121 if (len >= sizeof(xprt_str))
1122 strcpy(xprt_str, "name-too-long\n");
1123 /* Don't overflow buffer */
1124 len = strlen(xprt_str);
1125 if (buflen && (len + totlen >= buflen))
1127 strcpy(buf+totlen, xprt_str);
1130 spin_unlock_bh(&serv->sv_lock);
1133 EXPORT_SYMBOL_GPL(svc_xprt_names);
1136 /*----------------------------------------------------------------------------*/
1138 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1140 unsigned int pidx = (unsigned int)*pos;
1141 struct svc_serv *serv = m->private;
1143 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1146 /* bump up the pseudo refcount while traversing */
1151 return SEQ_START_TOKEN;
1152 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1155 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1157 struct svc_pool *pool = p;
1158 struct svc_serv *serv = m->private;
1160 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1162 if (p == SEQ_START_TOKEN) {
1163 pool = &serv->sv_pools[0];
1165 unsigned int pidx = (pool - &serv->sv_pools[0]);
1166 if (pidx < serv->sv_nrpools-1)
1167 pool = &serv->sv_pools[pidx+1];
1175 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1177 struct svc_serv *serv = m->private;
1180 /* this function really, really should have been called svc_put() */
1185 static int svc_pool_stats_show(struct seq_file *m, void *p)
1187 struct svc_pool *pool = p;
1189 if (p == SEQ_START_TOKEN) {
1190 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken overloads-avoided threads-timedout\n");
1194 seq_printf(m, "%u %lu %lu %lu %lu %lu\n",
1196 pool->sp_stats.packets,
1197 pool->sp_stats.sockets_queued,
1198 pool->sp_stats.threads_woken,
1199 pool->sp_stats.overloads_avoided,
1200 pool->sp_stats.threads_timedout);
1205 static const struct seq_operations svc_pool_stats_seq_ops = {
1206 .start = svc_pool_stats_start,
1207 .next = svc_pool_stats_next,
1208 .stop = svc_pool_stats_stop,
1209 .show = svc_pool_stats_show,
1212 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1216 err = seq_open(file, &svc_pool_stats_seq_ops);
1218 ((struct seq_file *) file->private_data)->private = serv;
1221 EXPORT_SYMBOL(svc_pool_stats_open);
1223 /*----------------------------------------------------------------------------*/