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);
25 static void svc_delete_xprt(struct svc_xprt *xprt);
27 /* apparently the "standard" is that clients close
28 * idle connections after 5 minutes, servers after
30 * http://www.connectathon.org/talks96/nfstcp.pdf
32 static int svc_conn_age_period = 6*60;
34 /* List of registered transport classes */
35 static DEFINE_SPINLOCK(svc_xprt_class_lock);
36 static LIST_HEAD(svc_xprt_class_list);
38 /* SMP locking strategy:
40 * svc_pool->sp_lock protects most of the fields of that pool.
41 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
42 * when both need to be taken (rare), svc_serv->sv_lock is first.
43 * BKL protects svc_serv->sv_nrthread.
44 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
45 * and the ->sk_info_authunix cache.
47 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
48 * enqueued multiply. During normal transport processing this bit
49 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
50 * Providers should not manipulate this bit directly.
52 * Some flags can be set to certain values at any time
53 * providing that certain rules are followed:
56 * - Can be set or cleared at any time.
57 * - After a set, svc_xprt_enqueue must be called to enqueue
58 * the transport for processing.
59 * - After a clear, the transport must be read/accepted.
60 * If this succeeds, it must be set again.
62 * - Can set at any time. It is never cleared.
64 * - Can only be set while XPT_BUSY is held which ensures
65 * that no other thread will be using the transport or will
66 * try to set XPT_DEAD.
69 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
71 struct svc_xprt_class *cl;
74 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
76 INIT_LIST_HEAD(&xcl->xcl_list);
77 spin_lock(&svc_xprt_class_lock);
78 /* Make sure there isn't already a class with the same name */
79 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
80 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
83 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
86 spin_unlock(&svc_xprt_class_lock);
89 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
91 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
93 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
94 spin_lock(&svc_xprt_class_lock);
95 list_del_init(&xcl->xcl_list);
96 spin_unlock(&svc_xprt_class_lock);
98 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
101 * Format the transport list for printing
103 int svc_print_xprts(char *buf, int maxlen)
105 struct svc_xprt_class *xcl;
110 spin_lock(&svc_xprt_class_lock);
111 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
114 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
115 slen = strlen(tmpstr);
116 if (len + slen > maxlen)
121 spin_unlock(&svc_xprt_class_lock);
126 static void svc_xprt_free(struct kref *kref)
128 struct svc_xprt *xprt =
129 container_of(kref, struct svc_xprt, xpt_ref);
130 struct module *owner = xprt->xpt_class->xcl_owner;
131 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
132 svcauth_unix_info_release(xprt);
133 put_net(xprt->xpt_net);
134 /* See comment on corresponding get in xs_setup_bc_tcp(): */
135 if (xprt->xpt_bc_xprt)
136 xprt_put(xprt->xpt_bc_xprt);
137 xprt->xpt_ops->xpo_free(xprt);
141 void svc_xprt_put(struct svc_xprt *xprt)
143 kref_put(&xprt->xpt_ref, svc_xprt_free);
145 EXPORT_SYMBOL_GPL(svc_xprt_put);
148 * Called by transport drivers to initialize the transport independent
149 * portion of the transport instance.
151 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
152 struct svc_xprt *xprt, struct svc_serv *serv)
154 memset(xprt, 0, sizeof(*xprt));
155 xprt->xpt_class = xcl;
156 xprt->xpt_ops = xcl->xcl_ops;
157 kref_init(&xprt->xpt_ref);
158 xprt->xpt_server = serv;
159 INIT_LIST_HEAD(&xprt->xpt_list);
160 INIT_LIST_HEAD(&xprt->xpt_ready);
161 INIT_LIST_HEAD(&xprt->xpt_deferred);
162 INIT_LIST_HEAD(&xprt->xpt_users);
163 mutex_init(&xprt->xpt_mutex);
164 spin_lock_init(&xprt->xpt_lock);
165 set_bit(XPT_BUSY, &xprt->xpt_flags);
166 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
167 xprt->xpt_net = get_net(net);
169 EXPORT_SYMBOL_GPL(svc_xprt_init);
171 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
172 struct svc_serv *serv,
175 const unsigned short port,
178 struct sockaddr_in sin = {
179 .sin_family = AF_INET,
180 .sin_addr.s_addr = htonl(INADDR_ANY),
181 .sin_port = htons(port),
183 #if IS_ENABLED(CONFIG_IPV6)
184 struct sockaddr_in6 sin6 = {
185 .sin6_family = AF_INET6,
186 .sin6_addr = IN6ADDR_ANY_INIT,
187 .sin6_port = htons(port),
190 struct sockaddr *sap;
195 sap = (struct sockaddr *)&sin;
198 #if IS_ENABLED(CONFIG_IPV6)
200 sap = (struct sockaddr *)&sin6;
205 return ERR_PTR(-EAFNOSUPPORT);
208 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
211 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
212 struct net *net, const int family,
213 const unsigned short port, int flags)
215 struct svc_xprt_class *xcl;
217 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
218 spin_lock(&svc_xprt_class_lock);
219 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
220 struct svc_xprt *newxprt;
221 unsigned short newport;
223 if (strcmp(xprt_name, xcl->xcl_name))
226 if (!try_module_get(xcl->xcl_owner))
229 spin_unlock(&svc_xprt_class_lock);
230 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
231 if (IS_ERR(newxprt)) {
232 module_put(xcl->xcl_owner);
233 return PTR_ERR(newxprt);
236 clear_bit(XPT_TEMP, &newxprt->xpt_flags);
237 spin_lock_bh(&serv->sv_lock);
238 list_add(&newxprt->xpt_list, &serv->sv_permsocks);
239 spin_unlock_bh(&serv->sv_lock);
240 newport = svc_xprt_local_port(newxprt);
241 clear_bit(XPT_BUSY, &newxprt->xpt_flags);
245 spin_unlock(&svc_xprt_class_lock);
246 dprintk("svc: transport %s not found\n", xprt_name);
248 /* This errno is exposed to user space. Provide a reasonable
249 * perror msg for a bad transport. */
250 return -EPROTONOSUPPORT;
252 EXPORT_SYMBOL_GPL(svc_create_xprt);
255 * Copy the local and remote xprt addresses to the rqstp structure
257 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
259 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
260 rqstp->rq_addrlen = xprt->xpt_remotelen;
263 * Destination address in request is needed for binding the
264 * source address in RPC replies/callbacks later.
266 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
267 rqstp->rq_daddrlen = xprt->xpt_locallen;
269 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
272 * svc_print_addr - Format rq_addr field for printing
273 * @rqstp: svc_rqst struct containing address to print
274 * @buf: target buffer for formatted address
275 * @len: length of target buffer
278 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
280 return __svc_print_addr(svc_addr(rqstp), buf, len);
282 EXPORT_SYMBOL_GPL(svc_print_addr);
285 * Queue up an idle server thread. Must have pool->sp_lock held.
286 * Note: this is really a stack rather than a queue, so that we only
287 * use as many different threads as we need, and the rest don't pollute
290 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
292 list_add(&rqstp->rq_list, &pool->sp_threads);
296 * Dequeue an nfsd thread. Must have pool->sp_lock held.
298 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
300 list_del(&rqstp->rq_list);
303 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
305 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
307 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
308 return xprt->xpt_ops->xpo_has_wspace(xprt);
313 * Queue up a transport with data pending. If there are idle nfsd
314 * processes, wake 'em up.
317 void svc_xprt_enqueue(struct svc_xprt *xprt)
319 struct svc_serv *serv = xprt->xpt_server;
320 struct svc_pool *pool;
321 struct svc_rqst *rqstp;
324 if (!svc_xprt_has_something_to_do(xprt))
328 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
331 spin_lock_bh(&pool->sp_lock);
333 if (!list_empty(&pool->sp_threads) &&
334 !list_empty(&pool->sp_sockets))
337 "threads and transports both waiting??\n");
339 pool->sp_stats.packets++;
341 /* Mark transport as busy. It will remain in this state until
342 * the provider calls svc_xprt_received. We update XPT_BUSY
343 * atomically because it also guards against trying to enqueue
344 * the transport twice.
346 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
347 /* Don't enqueue transport while already enqueued */
348 dprintk("svc: transport %p busy, not enqueued\n", xprt);
352 if (!list_empty(&pool->sp_threads)) {
353 rqstp = list_entry(pool->sp_threads.next,
356 dprintk("svc: transport %p served by daemon %p\n",
358 svc_thread_dequeue(pool, rqstp);
361 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
362 rqstp, rqstp->rq_xprt);
363 rqstp->rq_xprt = xprt;
365 rqstp->rq_reserved = serv->sv_max_mesg;
366 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
367 pool->sp_stats.threads_woken++;
368 wake_up(&rqstp->rq_wait);
370 dprintk("svc: transport %p put into queue\n", xprt);
371 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
372 pool->sp_stats.sockets_queued++;
376 spin_unlock_bh(&pool->sp_lock);
378 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
381 * Dequeue the first transport. Must be called with the pool->sp_lock held.
383 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
385 struct svc_xprt *xprt;
387 if (list_empty(&pool->sp_sockets))
390 xprt = list_entry(pool->sp_sockets.next,
391 struct svc_xprt, xpt_ready);
392 list_del_init(&xprt->xpt_ready);
394 dprintk("svc: transport %p dequeued, inuse=%d\n",
395 xprt, atomic_read(&xprt->xpt_ref.refcount));
401 * svc_xprt_received conditionally queues the transport for processing
402 * by another thread. The caller must hold the XPT_BUSY bit and must
403 * not thereafter touch transport data.
405 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
406 * insufficient) data.
408 void svc_xprt_received(struct svc_xprt *xprt)
410 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
411 /* As soon as we clear busy, the xprt could be closed and
412 * 'put', so we need a reference to call svc_xprt_enqueue with:
415 clear_bit(XPT_BUSY, &xprt->xpt_flags);
416 svc_xprt_enqueue(xprt);
419 EXPORT_SYMBOL_GPL(svc_xprt_received);
422 * svc_reserve - change the space reserved for the reply to a request.
423 * @rqstp: The request in question
424 * @space: new max space to reserve
426 * Each request reserves some space on the output queue of the transport
427 * to make sure the reply fits. This function reduces that reserved
428 * space to be the amount of space used already, plus @space.
431 void svc_reserve(struct svc_rqst *rqstp, int space)
433 space += rqstp->rq_res.head[0].iov_len;
435 if (space < rqstp->rq_reserved) {
436 struct svc_xprt *xprt = rqstp->rq_xprt;
437 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
438 rqstp->rq_reserved = space;
440 svc_xprt_enqueue(xprt);
443 EXPORT_SYMBOL_GPL(svc_reserve);
445 static void svc_xprt_release(struct svc_rqst *rqstp)
447 struct svc_xprt *xprt = rqstp->rq_xprt;
449 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
451 kfree(rqstp->rq_deferred);
452 rqstp->rq_deferred = NULL;
454 svc_free_res_pages(rqstp);
455 rqstp->rq_res.page_len = 0;
456 rqstp->rq_res.page_base = 0;
458 /* Reset response buffer and release
460 * But first, check that enough space was reserved
461 * for the reply, otherwise we have a bug!
463 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
464 printk(KERN_ERR "RPC request reserved %d but used %d\n",
468 rqstp->rq_res.head[0].iov_len = 0;
469 svc_reserve(rqstp, 0);
470 rqstp->rq_xprt = NULL;
476 * External function to wake up a server waiting for data
477 * This really only makes sense for services like lockd
478 * which have exactly one thread anyway.
480 void svc_wake_up(struct svc_serv *serv)
482 struct svc_rqst *rqstp;
484 struct svc_pool *pool;
486 for (i = 0; i < serv->sv_nrpools; i++) {
487 pool = &serv->sv_pools[i];
489 spin_lock_bh(&pool->sp_lock);
490 if (!list_empty(&pool->sp_threads)) {
491 rqstp = list_entry(pool->sp_threads.next,
494 dprintk("svc: daemon %p woken up.\n", rqstp);
496 svc_thread_dequeue(pool, rqstp);
497 rqstp->rq_xprt = NULL;
499 wake_up(&rqstp->rq_wait);
501 spin_unlock_bh(&pool->sp_lock);
504 EXPORT_SYMBOL_GPL(svc_wake_up);
506 int svc_port_is_privileged(struct sockaddr *sin)
508 switch (sin->sa_family) {
510 return ntohs(((struct sockaddr_in *)sin)->sin_port)
513 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
521 * Make sure that we don't have too many active connections. If we have,
522 * something must be dropped. It's not clear what will happen if we allow
523 * "too many" connections, but when dealing with network-facing software,
524 * we have to code defensively. Here we do that by imposing hard limits.
526 * There's no point in trying to do random drop here for DoS
527 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
528 * attacker can easily beat that.
530 * The only somewhat efficient mechanism would be if drop old
531 * connections from the same IP first. But right now we don't even
532 * record the client IP in svc_sock.
534 * single-threaded services that expect a lot of clients will probably
535 * need to set sv_maxconn to override the default value which is based
536 * on the number of threads
538 static void svc_check_conn_limits(struct svc_serv *serv)
540 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
541 (serv->sv_nrthreads+3) * 20;
543 if (serv->sv_tmpcnt > limit) {
544 struct svc_xprt *xprt = NULL;
545 spin_lock_bh(&serv->sv_lock);
546 if (!list_empty(&serv->sv_tempsocks)) {
547 /* Try to help the admin */
548 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
549 serv->sv_name, serv->sv_maxconn ?
550 "max number of connections" :
551 "number of threads");
553 * Always select the oldest connection. It's not fair,
556 xprt = list_entry(serv->sv_tempsocks.prev,
559 set_bit(XPT_CLOSE, &xprt->xpt_flags);
562 spin_unlock_bh(&serv->sv_lock);
565 svc_xprt_enqueue(xprt);
572 * Receive the next request on any transport. This code is carefully
573 * organised not to touch any cachelines in the shared svc_serv
574 * structure, only cachelines in the local svc_pool.
576 int svc_recv(struct svc_rqst *rqstp, long timeout)
578 struct svc_xprt *xprt = NULL;
579 struct svc_serv *serv = rqstp->rq_server;
580 struct svc_pool *pool = rqstp->rq_pool;
584 DECLARE_WAITQUEUE(wait, current);
587 dprintk("svc: server %p waiting for data (to = %ld)\n",
592 "svc_recv: service %p, transport not NULL!\n",
594 if (waitqueue_active(&rqstp->rq_wait))
596 "svc_recv: service %p, wait queue active!\n",
599 /* now allocate needed pages. If we get a failure, sleep briefly */
600 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
601 BUG_ON(pages >= RPCSVC_MAXPAGES);
602 for (i = 0; i < pages ; i++)
603 while (rqstp->rq_pages[i] == NULL) {
604 struct page *p = alloc_page(GFP_KERNEL);
606 set_current_state(TASK_INTERRUPTIBLE);
607 if (signalled() || kthread_should_stop()) {
608 set_current_state(TASK_RUNNING);
611 schedule_timeout(msecs_to_jiffies(500));
613 rqstp->rq_pages[i] = p;
615 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
617 /* Make arg->head point to first page and arg->pages point to rest */
618 arg = &rqstp->rq_arg;
619 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
620 arg->head[0].iov_len = PAGE_SIZE;
621 arg->pages = rqstp->rq_pages + 1;
623 /* save at least one page for response */
624 arg->page_len = (pages-2)*PAGE_SIZE;
625 arg->len = (pages-1)*PAGE_SIZE;
626 arg->tail[0].iov_len = 0;
630 if (signalled() || kthread_should_stop())
633 /* Normally we will wait up to 5 seconds for any required
634 * cache information to be provided.
636 rqstp->rq_chandle.thread_wait = 5*HZ;
638 spin_lock_bh(&pool->sp_lock);
639 xprt = svc_xprt_dequeue(pool);
641 rqstp->rq_xprt = xprt;
643 rqstp->rq_reserved = serv->sv_max_mesg;
644 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
646 /* As there is a shortage of threads and this request
647 * had to be queued, don't allow the thread to wait so
648 * long for cache updates.
650 rqstp->rq_chandle.thread_wait = 1*HZ;
652 /* No data pending. Go to sleep */
653 svc_thread_enqueue(pool, rqstp);
656 * We have to be able to interrupt this wait
657 * to bring down the daemons ...
659 set_current_state(TASK_INTERRUPTIBLE);
662 * checking kthread_should_stop() here allows us to avoid
663 * locking and signalling when stopping kthreads that call
664 * svc_recv. If the thread has already been woken up, then
665 * we can exit here without sleeping. If not, then it
666 * it'll be woken up quickly during the schedule_timeout
668 if (kthread_should_stop()) {
669 set_current_state(TASK_RUNNING);
670 spin_unlock_bh(&pool->sp_lock);
674 add_wait_queue(&rqstp->rq_wait, &wait);
675 spin_unlock_bh(&pool->sp_lock);
677 time_left = schedule_timeout(timeout);
681 spin_lock_bh(&pool->sp_lock);
682 remove_wait_queue(&rqstp->rq_wait, &wait);
684 pool->sp_stats.threads_timedout++;
686 xprt = rqstp->rq_xprt;
688 svc_thread_dequeue(pool, rqstp);
689 spin_unlock_bh(&pool->sp_lock);
690 dprintk("svc: server %p, no data yet\n", rqstp);
691 if (signalled() || kthread_should_stop())
697 spin_unlock_bh(&pool->sp_lock);
700 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
701 dprintk("svc_recv: found XPT_CLOSE\n");
702 svc_delete_xprt(xprt);
703 /* Leave XPT_BUSY set on the dead xprt: */
706 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
707 struct svc_xprt *newxpt;
708 newxpt = xprt->xpt_ops->xpo_accept(xprt);
711 * We know this module_get will succeed because the
712 * listener holds a reference too
714 __module_get(newxpt->xpt_class->xcl_owner);
715 svc_check_conn_limits(xprt->xpt_server);
716 spin_lock_bh(&serv->sv_lock);
717 set_bit(XPT_TEMP, &newxpt->xpt_flags);
718 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
720 if (serv->sv_temptimer.function == NULL) {
721 /* setup timer to age temp transports */
722 setup_timer(&serv->sv_temptimer,
724 (unsigned long)serv);
725 mod_timer(&serv->sv_temptimer,
726 jiffies + svc_conn_age_period * HZ);
728 spin_unlock_bh(&serv->sv_lock);
729 svc_xprt_received(newxpt);
731 } else if (xprt->xpt_ops->xpo_has_wspace(xprt)) {
732 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
733 rqstp, pool->sp_id, xprt,
734 atomic_read(&xprt->xpt_ref.refcount));
735 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
736 if (rqstp->rq_deferred)
737 len = svc_deferred_recv(rqstp);
739 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
740 dprintk("svc: got len=%d\n", len);
742 svc_xprt_received(xprt);
744 /* No data, incomplete (TCP) read, or accept() */
745 if (len == 0 || len == -EAGAIN)
748 clear_bit(XPT_OLD, &xprt->xpt_flags);
750 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
751 rqstp->rq_chandle.defer = svc_defer;
754 serv->sv_stats->netcnt++;
757 rqstp->rq_res.len = 0;
758 svc_xprt_release(rqstp);
761 EXPORT_SYMBOL_GPL(svc_recv);
766 void svc_drop(struct svc_rqst *rqstp)
768 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
769 svc_xprt_release(rqstp);
771 EXPORT_SYMBOL_GPL(svc_drop);
774 * Return reply to client.
776 int svc_send(struct svc_rqst *rqstp)
778 struct svc_xprt *xprt;
782 xprt = rqstp->rq_xprt;
786 /* release the receive skb before sending the reply */
787 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
789 /* calculate over-all length */
791 xb->len = xb->head[0].iov_len +
795 /* Grab mutex to serialize outgoing data. */
796 mutex_lock(&xprt->xpt_mutex);
797 if (test_bit(XPT_DEAD, &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 static 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, struct net *net)
924 struct svc_xprt *xprt;
926 list_for_each_entry(xprt, xprt_list, xpt_list) {
927 if (xprt->xpt_net != net)
929 set_bit(XPT_CLOSE, &xprt->xpt_flags);
930 set_bit(XPT_BUSY, &xprt->xpt_flags);
934 static void svc_clear_pools(struct svc_serv *serv, struct net *net)
936 struct svc_pool *pool;
937 struct svc_xprt *xprt;
938 struct svc_xprt *tmp;
941 for (i = 0; i < serv->sv_nrpools; i++) {
942 pool = &serv->sv_pools[i];
944 spin_lock_bh(&pool->sp_lock);
945 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
946 if (xprt->xpt_net != net)
948 list_del_init(&xprt->xpt_ready);
950 spin_unlock_bh(&pool->sp_lock);
954 static void svc_clear_list(struct list_head *xprt_list, struct net *net)
956 struct svc_xprt *xprt;
957 struct svc_xprt *tmp;
959 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
960 if (xprt->xpt_net != net)
962 svc_delete_xprt(xprt);
964 list_for_each_entry(xprt, xprt_list, xpt_list)
965 BUG_ON(xprt->xpt_net == net);
968 void svc_close_net(struct svc_serv *serv, struct net *net)
970 svc_close_list(&serv->sv_tempsocks, net);
971 svc_close_list(&serv->sv_permsocks, net);
973 svc_clear_pools(serv, net);
975 * At this point the sp_sockets lists will stay empty, since
976 * svc_xprt_enqueue will not add new entries without taking the
977 * sp_lock and checking XPT_BUSY.
979 svc_clear_list(&serv->sv_tempsocks, net);
980 svc_clear_list(&serv->sv_permsocks, net);
984 * Handle defer and revisit of requests
987 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
989 struct svc_deferred_req *dr =
990 container_of(dreq, struct svc_deferred_req, handle);
991 struct svc_xprt *xprt = dr->xprt;
993 spin_lock(&xprt->xpt_lock);
994 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
995 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
996 spin_unlock(&xprt->xpt_lock);
997 dprintk("revisit canceled\n");
1002 dprintk("revisit queued\n");
1004 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1005 spin_unlock(&xprt->xpt_lock);
1006 svc_xprt_enqueue(xprt);
1011 * Save the request off for later processing. The request buffer looks
1014 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1016 * This code can only handle requests that consist of an xprt-header
1019 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1021 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1022 struct svc_deferred_req *dr;
1024 if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
1025 return NULL; /* if more than a page, give up FIXME */
1026 if (rqstp->rq_deferred) {
1027 dr = rqstp->rq_deferred;
1028 rqstp->rq_deferred = NULL;
1032 /* FIXME maybe discard if size too large */
1033 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1034 dr = kmalloc(size, GFP_KERNEL);
1038 dr->handle.owner = rqstp->rq_server;
1039 dr->prot = rqstp->rq_prot;
1040 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1041 dr->addrlen = rqstp->rq_addrlen;
1042 dr->daddr = rqstp->rq_daddr;
1043 dr->argslen = rqstp->rq_arg.len >> 2;
1044 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1046 /* back up head to the start of the buffer and copy */
1047 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1048 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1051 svc_xprt_get(rqstp->rq_xprt);
1052 dr->xprt = rqstp->rq_xprt;
1053 rqstp->rq_dropme = true;
1055 dr->handle.revisit = svc_revisit;
1060 * recv data from a deferred request into an active one
1062 static int svc_deferred_recv(struct svc_rqst *rqstp)
1064 struct svc_deferred_req *dr = rqstp->rq_deferred;
1066 /* setup iov_base past transport header */
1067 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1068 /* The iov_len does not include the transport header bytes */
1069 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1070 rqstp->rq_arg.page_len = 0;
1071 /* The rq_arg.len includes the transport header bytes */
1072 rqstp->rq_arg.len = dr->argslen<<2;
1073 rqstp->rq_prot = dr->prot;
1074 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1075 rqstp->rq_addrlen = dr->addrlen;
1076 /* Save off transport header len in case we get deferred again */
1077 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1078 rqstp->rq_daddr = dr->daddr;
1079 rqstp->rq_respages = rqstp->rq_pages;
1080 return (dr->argslen<<2) - dr->xprt_hlen;
1084 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1086 struct svc_deferred_req *dr = NULL;
1088 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1090 spin_lock(&xprt->xpt_lock);
1091 if (!list_empty(&xprt->xpt_deferred)) {
1092 dr = list_entry(xprt->xpt_deferred.next,
1093 struct svc_deferred_req,
1095 list_del_init(&dr->handle.recent);
1097 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1098 spin_unlock(&xprt->xpt_lock);
1103 * svc_find_xprt - find an RPC transport instance
1104 * @serv: pointer to svc_serv to search
1105 * @xcl_name: C string containing transport's class name
1106 * @net: owner net pointer
1107 * @af: Address family of transport's local address
1108 * @port: transport's IP port number
1110 * Return the transport instance pointer for the endpoint accepting
1111 * connections/peer traffic from the specified transport class,
1112 * address family and port.
1114 * Specifying 0 for the address family or port is effectively a
1115 * wild-card, and will result in matching the first transport in the
1116 * service's list that has a matching class name.
1118 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1119 struct net *net, const sa_family_t af,
1120 const unsigned short port)
1122 struct svc_xprt *xprt;
1123 struct svc_xprt *found = NULL;
1125 /* Sanity check the args */
1126 if (serv == NULL || xcl_name == NULL)
1129 spin_lock_bh(&serv->sv_lock);
1130 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1131 if (xprt->xpt_net != net)
1133 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1135 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1137 if (port != 0 && port != svc_xprt_local_port(xprt))
1143 spin_unlock_bh(&serv->sv_lock);
1146 EXPORT_SYMBOL_GPL(svc_find_xprt);
1148 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1149 char *pos, int remaining)
1153 len = snprintf(pos, remaining, "%s %u\n",
1154 xprt->xpt_class->xcl_name,
1155 svc_xprt_local_port(xprt));
1156 if (len >= remaining)
1157 return -ENAMETOOLONG;
1162 * svc_xprt_names - format a buffer with a list of transport names
1163 * @serv: pointer to an RPC service
1164 * @buf: pointer to a buffer to be filled in
1165 * @buflen: length of buffer to be filled in
1167 * Fills in @buf with a string containing a list of transport names,
1168 * each name terminated with '\n'.
1170 * Returns positive length of the filled-in string on success; otherwise
1171 * a negative errno value is returned if an error occurs.
1173 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1175 struct svc_xprt *xprt;
1179 /* Sanity check args */
1183 spin_lock_bh(&serv->sv_lock);
1187 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1188 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1200 spin_unlock_bh(&serv->sv_lock);
1203 EXPORT_SYMBOL_GPL(svc_xprt_names);
1206 /*----------------------------------------------------------------------------*/
1208 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1210 unsigned int pidx = (unsigned int)*pos;
1211 struct svc_serv *serv = m->private;
1213 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1216 return SEQ_START_TOKEN;
1217 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1220 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1222 struct svc_pool *pool = p;
1223 struct svc_serv *serv = m->private;
1225 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1227 if (p == SEQ_START_TOKEN) {
1228 pool = &serv->sv_pools[0];
1230 unsigned int pidx = (pool - &serv->sv_pools[0]);
1231 if (pidx < serv->sv_nrpools-1)
1232 pool = &serv->sv_pools[pidx+1];
1240 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1244 static int svc_pool_stats_show(struct seq_file *m, void *p)
1246 struct svc_pool *pool = p;
1248 if (p == SEQ_START_TOKEN) {
1249 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1253 seq_printf(m, "%u %lu %lu %lu %lu\n",
1255 pool->sp_stats.packets,
1256 pool->sp_stats.sockets_queued,
1257 pool->sp_stats.threads_woken,
1258 pool->sp_stats.threads_timedout);
1263 static const struct seq_operations svc_pool_stats_seq_ops = {
1264 .start = svc_pool_stats_start,
1265 .next = svc_pool_stats_next,
1266 .stop = svc_pool_stats_stop,
1267 .show = svc_pool_stats_show,
1270 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1274 err = seq_open(file, &svc_pool_stats_seq_ops);
1276 ((struct seq_file *) file->private_data)->private = serv;
1279 EXPORT_SYMBOL(svc_pool_stats_open);
1281 /*----------------------------------------------------------------------------*/