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 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
213 clear_bit(XPT_TEMP, &new->xpt_flags);
214 spin_lock_bh(&serv->sv_lock);
215 list_add(&new->xpt_list, &serv->sv_permsocks);
216 spin_unlock_bh(&serv->sv_lock);
217 svc_xprt_received(new);
220 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
221 struct net *net, const int family,
222 const unsigned short port, int flags)
224 struct svc_xprt_class *xcl;
226 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
227 spin_lock(&svc_xprt_class_lock);
228 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
229 struct svc_xprt *newxprt;
230 unsigned short newport;
232 if (strcmp(xprt_name, xcl->xcl_name))
235 if (!try_module_get(xcl->xcl_owner))
238 spin_unlock(&svc_xprt_class_lock);
239 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
240 if (IS_ERR(newxprt)) {
241 module_put(xcl->xcl_owner);
242 return PTR_ERR(newxprt);
244 svc_add_new_perm_xprt(serv, newxprt);
245 newport = svc_xprt_local_port(newxprt);
249 spin_unlock(&svc_xprt_class_lock);
250 dprintk("svc: transport %s not found\n", xprt_name);
252 /* This errno is exposed to user space. Provide a reasonable
253 * perror msg for a bad transport. */
254 return -EPROTONOSUPPORT;
256 EXPORT_SYMBOL_GPL(svc_create_xprt);
259 * Copy the local and remote xprt addresses to the rqstp structure
261 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
263 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
264 rqstp->rq_addrlen = xprt->xpt_remotelen;
267 * Destination address in request is needed for binding the
268 * source address in RPC replies/callbacks later.
270 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
271 rqstp->rq_daddrlen = xprt->xpt_locallen;
273 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
276 * svc_print_addr - Format rq_addr field for printing
277 * @rqstp: svc_rqst struct containing address to print
278 * @buf: target buffer for formatted address
279 * @len: length of target buffer
282 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
284 return __svc_print_addr(svc_addr(rqstp), buf, len);
286 EXPORT_SYMBOL_GPL(svc_print_addr);
289 * Queue up an idle server thread. Must have pool->sp_lock held.
290 * Note: this is really a stack rather than a queue, so that we only
291 * use as many different threads as we need, and the rest don't pollute
294 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
296 list_add(&rqstp->rq_list, &pool->sp_threads);
300 * Dequeue an nfsd thread. Must have pool->sp_lock held.
302 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
304 list_del(&rqstp->rq_list);
307 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
309 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
311 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
312 return xprt->xpt_ops->xpo_has_wspace(xprt);
317 * Queue up a transport with data pending. If there are idle nfsd
318 * processes, wake 'em up.
321 void svc_xprt_enqueue(struct svc_xprt *xprt)
323 struct svc_pool *pool;
324 struct svc_rqst *rqstp;
327 if (!svc_xprt_has_something_to_do(xprt))
331 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
334 spin_lock_bh(&pool->sp_lock);
336 if (!list_empty(&pool->sp_threads) &&
337 !list_empty(&pool->sp_sockets))
340 "threads and transports both waiting??\n");
342 pool->sp_stats.packets++;
344 /* Mark transport as busy. It will remain in this state until
345 * the provider calls svc_xprt_received. We update XPT_BUSY
346 * atomically because it also guards against trying to enqueue
347 * the transport twice.
349 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
350 /* Don't enqueue transport while already enqueued */
351 dprintk("svc: transport %p busy, not enqueued\n", xprt);
355 if (!list_empty(&pool->sp_threads)) {
356 rqstp = list_entry(pool->sp_threads.next,
359 dprintk("svc: transport %p served by daemon %p\n",
361 svc_thread_dequeue(pool, rqstp);
364 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
365 rqstp, rqstp->rq_xprt);
366 rqstp->rq_xprt = xprt;
368 pool->sp_stats.threads_woken++;
369 wake_up(&rqstp->rq_wait);
371 dprintk("svc: transport %p put into queue\n", xprt);
372 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
373 pool->sp_stats.sockets_queued++;
377 spin_unlock_bh(&pool->sp_lock);
379 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
382 * Dequeue the first transport. Must be called with the pool->sp_lock held.
384 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
386 struct svc_xprt *xprt;
388 if (list_empty(&pool->sp_sockets))
391 xprt = list_entry(pool->sp_sockets.next,
392 struct svc_xprt, xpt_ready);
393 list_del_init(&xprt->xpt_ready);
395 dprintk("svc: transport %p dequeued, inuse=%d\n",
396 xprt, atomic_read(&xprt->xpt_ref.refcount));
402 * svc_xprt_received conditionally queues the transport for processing
403 * by another thread. The caller must hold the XPT_BUSY bit and must
404 * not thereafter touch transport data.
406 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
407 * insufficient) data.
409 void svc_xprt_received(struct svc_xprt *xprt)
411 BUG_ON(!test_bit(XPT_BUSY, &xprt->xpt_flags));
412 /* As soon as we clear busy, the xprt could be closed and
413 * 'put', so we need a reference to call svc_xprt_enqueue with:
416 clear_bit(XPT_BUSY, &xprt->xpt_flags);
417 svc_xprt_enqueue(xprt);
420 EXPORT_SYMBOL_GPL(svc_xprt_received);
423 * svc_reserve - change the space reserved for the reply to a request.
424 * @rqstp: The request in question
425 * @space: new max space to reserve
427 * Each request reserves some space on the output queue of the transport
428 * to make sure the reply fits. This function reduces that reserved
429 * space to be the amount of space used already, plus @space.
432 void svc_reserve(struct svc_rqst *rqstp, int space)
434 space += rqstp->rq_res.head[0].iov_len;
436 if (space < rqstp->rq_reserved) {
437 struct svc_xprt *xprt = rqstp->rq_xprt;
438 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
439 rqstp->rq_reserved = space;
441 svc_xprt_enqueue(xprt);
444 EXPORT_SYMBOL_GPL(svc_reserve);
446 static void svc_xprt_release(struct svc_rqst *rqstp)
448 struct svc_xprt *xprt = rqstp->rq_xprt;
450 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
452 kfree(rqstp->rq_deferred);
453 rqstp->rq_deferred = NULL;
455 svc_free_res_pages(rqstp);
456 rqstp->rq_res.page_len = 0;
457 rqstp->rq_res.page_base = 0;
459 /* Reset response buffer and release
461 * But first, check that enough space was reserved
462 * for the reply, otherwise we have a bug!
464 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
465 printk(KERN_ERR "RPC request reserved %d but used %d\n",
469 rqstp->rq_res.head[0].iov_len = 0;
470 svc_reserve(rqstp, 0);
471 rqstp->rq_xprt = NULL;
477 * External function to wake up a server waiting for data
478 * This really only makes sense for services like lockd
479 * which have exactly one thread anyway.
481 void svc_wake_up(struct svc_serv *serv)
483 struct svc_rqst *rqstp;
485 struct svc_pool *pool;
487 for (i = 0; i < serv->sv_nrpools; i++) {
488 pool = &serv->sv_pools[i];
490 spin_lock_bh(&pool->sp_lock);
491 if (!list_empty(&pool->sp_threads)) {
492 rqstp = list_entry(pool->sp_threads.next,
495 dprintk("svc: daemon %p woken up.\n", rqstp);
497 svc_thread_dequeue(pool, rqstp);
498 rqstp->rq_xprt = NULL;
500 wake_up(&rqstp->rq_wait);
502 spin_unlock_bh(&pool->sp_lock);
505 EXPORT_SYMBOL_GPL(svc_wake_up);
507 int svc_port_is_privileged(struct sockaddr *sin)
509 switch (sin->sa_family) {
511 return ntohs(((struct sockaddr_in *)sin)->sin_port)
514 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
522 * Make sure that we don't have too many active connections. If we have,
523 * something must be dropped. It's not clear what will happen if we allow
524 * "too many" connections, but when dealing with network-facing software,
525 * we have to code defensively. Here we do that by imposing hard limits.
527 * There's no point in trying to do random drop here for DoS
528 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
529 * attacker can easily beat that.
531 * The only somewhat efficient mechanism would be if drop old
532 * connections from the same IP first. But right now we don't even
533 * record the client IP in svc_sock.
535 * single-threaded services that expect a lot of clients will probably
536 * need to set sv_maxconn to override the default value which is based
537 * on the number of threads
539 static void svc_check_conn_limits(struct svc_serv *serv)
541 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
542 (serv->sv_nrthreads+3) * 20;
544 if (serv->sv_tmpcnt > limit) {
545 struct svc_xprt *xprt = NULL;
546 spin_lock_bh(&serv->sv_lock);
547 if (!list_empty(&serv->sv_tempsocks)) {
548 /* Try to help the admin */
549 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
550 serv->sv_name, serv->sv_maxconn ?
551 "max number of connections" :
552 "number of threads");
554 * Always select the oldest connection. It's not fair,
557 xprt = list_entry(serv->sv_tempsocks.prev,
560 set_bit(XPT_CLOSE, &xprt->xpt_flags);
563 spin_unlock_bh(&serv->sv_lock);
566 svc_xprt_enqueue(xprt);
573 * Receive the next request on any transport. This code is carefully
574 * organised not to touch any cachelines in the shared svc_serv
575 * structure, only cachelines in the local svc_pool.
577 int svc_recv(struct svc_rqst *rqstp, long timeout)
579 struct svc_xprt *xprt = NULL;
580 struct svc_serv *serv = rqstp->rq_server;
581 struct svc_pool *pool = rqstp->rq_pool;
585 DECLARE_WAITQUEUE(wait, current);
588 dprintk("svc: server %p waiting for data (to = %ld)\n",
593 "svc_recv: service %p, transport not NULL!\n",
595 if (waitqueue_active(&rqstp->rq_wait))
597 "svc_recv: service %p, wait queue active!\n",
600 /* now allocate needed pages. If we get a failure, sleep briefly */
601 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
602 BUG_ON(pages >= RPCSVC_MAXPAGES);
603 for (i = 0; i < pages ; i++)
604 while (rqstp->rq_pages[i] == NULL) {
605 struct page *p = alloc_page(GFP_KERNEL);
607 set_current_state(TASK_INTERRUPTIBLE);
608 if (signalled() || kthread_should_stop()) {
609 set_current_state(TASK_RUNNING);
612 schedule_timeout(msecs_to_jiffies(500));
614 rqstp->rq_pages[i] = p;
616 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
618 /* Make arg->head point to first page and arg->pages point to rest */
619 arg = &rqstp->rq_arg;
620 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
621 arg->head[0].iov_len = PAGE_SIZE;
622 arg->pages = rqstp->rq_pages + 1;
624 /* save at least one page for response */
625 arg->page_len = (pages-2)*PAGE_SIZE;
626 arg->len = (pages-1)*PAGE_SIZE;
627 arg->tail[0].iov_len = 0;
631 if (signalled() || kthread_should_stop())
634 /* Normally we will wait up to 5 seconds for any required
635 * cache information to be provided.
637 rqstp->rq_chandle.thread_wait = 5*HZ;
639 spin_lock_bh(&pool->sp_lock);
640 xprt = svc_xprt_dequeue(pool);
642 rqstp->rq_xprt = xprt;
645 /* As there is a shortage of threads and this request
646 * had to be queued, don't allow the thread to wait so
647 * long for cache updates.
649 rqstp->rq_chandle.thread_wait = 1*HZ;
651 /* No data pending. Go to sleep */
652 svc_thread_enqueue(pool, rqstp);
655 * We have to be able to interrupt this wait
656 * to bring down the daemons ...
658 set_current_state(TASK_INTERRUPTIBLE);
661 * checking kthread_should_stop() here allows us to avoid
662 * locking and signalling when stopping kthreads that call
663 * svc_recv. If the thread has already been woken up, then
664 * we can exit here without sleeping. If not, then it
665 * it'll be woken up quickly during the schedule_timeout
667 if (kthread_should_stop()) {
668 set_current_state(TASK_RUNNING);
669 spin_unlock_bh(&pool->sp_lock);
673 add_wait_queue(&rqstp->rq_wait, &wait);
674 spin_unlock_bh(&pool->sp_lock);
676 time_left = schedule_timeout(timeout);
680 spin_lock_bh(&pool->sp_lock);
681 remove_wait_queue(&rqstp->rq_wait, &wait);
683 pool->sp_stats.threads_timedout++;
685 xprt = rqstp->rq_xprt;
687 svc_thread_dequeue(pool, rqstp);
688 spin_unlock_bh(&pool->sp_lock);
689 dprintk("svc: server %p, no data yet\n", rqstp);
690 if (signalled() || kthread_should_stop())
696 spin_unlock_bh(&pool->sp_lock);
699 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
700 dprintk("svc_recv: found XPT_CLOSE\n");
701 svc_delete_xprt(xprt);
702 /* Leave XPT_BUSY set on the dead xprt: */
705 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
706 struct svc_xprt *newxpt;
707 newxpt = xprt->xpt_ops->xpo_accept(xprt);
710 * We know this module_get will succeed because the
711 * listener holds a reference too
713 __module_get(newxpt->xpt_class->xcl_owner);
714 svc_check_conn_limits(xprt->xpt_server);
715 spin_lock_bh(&serv->sv_lock);
716 set_bit(XPT_TEMP, &newxpt->xpt_flags);
717 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
719 if (serv->sv_temptimer.function == NULL) {
720 /* setup timer to age temp transports */
721 setup_timer(&serv->sv_temptimer,
723 (unsigned long)serv);
724 mod_timer(&serv->sv_temptimer,
725 jiffies + svc_conn_age_period * HZ);
727 spin_unlock_bh(&serv->sv_lock);
728 svc_xprt_received(newxpt);
730 } else if (xprt->xpt_ops->xpo_has_wspace(xprt)) {
731 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
732 rqstp, pool->sp_id, xprt,
733 atomic_read(&xprt->xpt_ref.refcount));
734 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
735 if (rqstp->rq_deferred)
736 len = svc_deferred_recv(rqstp);
738 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
739 dprintk("svc: got len=%d\n", len);
740 rqstp->rq_reserved = serv->sv_max_mesg;
741 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
743 svc_xprt_received(xprt);
745 /* No data, incomplete (TCP) read, or accept() */
749 clear_bit(XPT_OLD, &xprt->xpt_flags);
751 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
752 rqstp->rq_chandle.defer = svc_defer;
755 serv->sv_stats->netcnt++;
758 rqstp->rq_res.len = 0;
759 svc_xprt_release(rqstp);
762 EXPORT_SYMBOL_GPL(svc_recv);
767 void svc_drop(struct svc_rqst *rqstp)
769 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
770 svc_xprt_release(rqstp);
772 EXPORT_SYMBOL_GPL(svc_drop);
775 * Return reply to client.
777 int svc_send(struct svc_rqst *rqstp)
779 struct svc_xprt *xprt;
783 xprt = rqstp->rq_xprt;
787 /* release the receive skb before sending the reply */
788 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
790 /* calculate over-all length */
792 xb->len = xb->head[0].iov_len +
796 /* Grab mutex to serialize outgoing data. */
797 mutex_lock(&xprt->xpt_mutex);
798 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
799 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
802 len = xprt->xpt_ops->xpo_sendto(rqstp);
803 mutex_unlock(&xprt->xpt_mutex);
804 rpc_wake_up(&xprt->xpt_bc_pending);
805 svc_xprt_release(rqstp);
807 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
813 * Timer function to close old temporary transports, using
814 * a mark-and-sweep algorithm.
816 static void svc_age_temp_xprts(unsigned long closure)
818 struct svc_serv *serv = (struct svc_serv *)closure;
819 struct svc_xprt *xprt;
820 struct list_head *le, *next;
821 LIST_HEAD(to_be_aged);
823 dprintk("svc_age_temp_xprts\n");
825 if (!spin_trylock_bh(&serv->sv_lock)) {
826 /* busy, try again 1 sec later */
827 dprintk("svc_age_temp_xprts: busy\n");
828 mod_timer(&serv->sv_temptimer, jiffies + HZ);
832 list_for_each_safe(le, next, &serv->sv_tempsocks) {
833 xprt = list_entry(le, struct svc_xprt, xpt_list);
835 /* First time through, just mark it OLD. Second time
836 * through, close it. */
837 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
839 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
840 test_bit(XPT_BUSY, &xprt->xpt_flags))
843 list_move(le, &to_be_aged);
844 set_bit(XPT_CLOSE, &xprt->xpt_flags);
845 set_bit(XPT_DETACHED, &xprt->xpt_flags);
847 spin_unlock_bh(&serv->sv_lock);
849 while (!list_empty(&to_be_aged)) {
850 le = to_be_aged.next;
851 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
853 xprt = list_entry(le, struct svc_xprt, xpt_list);
855 dprintk("queuing xprt %p for closing\n", xprt);
857 /* a thread will dequeue and close it soon */
858 svc_xprt_enqueue(xprt);
862 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
865 static void call_xpt_users(struct svc_xprt *xprt)
867 struct svc_xpt_user *u;
869 spin_lock(&xprt->xpt_lock);
870 while (!list_empty(&xprt->xpt_users)) {
871 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
875 spin_unlock(&xprt->xpt_lock);
879 * Remove a dead transport
881 static void svc_delete_xprt(struct svc_xprt *xprt)
883 struct svc_serv *serv = xprt->xpt_server;
884 struct svc_deferred_req *dr;
886 /* Only do this once */
887 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
890 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
891 xprt->xpt_ops->xpo_detach(xprt);
893 spin_lock_bh(&serv->sv_lock);
894 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
895 list_del_init(&xprt->xpt_list);
896 BUG_ON(!list_empty(&xprt->xpt_ready));
897 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
899 spin_unlock_bh(&serv->sv_lock);
901 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
904 call_xpt_users(xprt);
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 */
915 * We expect svc_close_xprt() to work even when no threads are
916 * running (e.g., while configuring the server before starting
917 * any threads), so if the transport isn't busy, we delete
920 svc_delete_xprt(xprt);
922 EXPORT_SYMBOL_GPL(svc_close_xprt);
924 static void svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
926 struct svc_xprt *xprt;
928 spin_lock(&serv->sv_lock);
929 list_for_each_entry(xprt, xprt_list, xpt_list) {
930 if (xprt->xpt_net != net)
932 set_bit(XPT_CLOSE, &xprt->xpt_flags);
933 set_bit(XPT_BUSY, &xprt->xpt_flags);
935 spin_unlock(&serv->sv_lock);
938 static void svc_clear_pools(struct svc_serv *serv, struct net *net)
940 struct svc_pool *pool;
941 struct svc_xprt *xprt;
942 struct svc_xprt *tmp;
945 for (i = 0; i < serv->sv_nrpools; i++) {
946 pool = &serv->sv_pools[i];
948 spin_lock_bh(&pool->sp_lock);
949 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
950 if (xprt->xpt_net != net)
952 list_del_init(&xprt->xpt_ready);
954 spin_unlock_bh(&pool->sp_lock);
958 static void svc_clear_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
960 struct svc_xprt *xprt;
961 struct svc_xprt *tmp;
964 spin_lock(&serv->sv_lock);
965 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
966 if (xprt->xpt_net != net)
968 list_move(&xprt->xpt_list, &victims);
970 spin_unlock(&serv->sv_lock);
972 list_for_each_entry_safe(xprt, tmp, &victims, xpt_list)
973 svc_delete_xprt(xprt);
976 void svc_close_net(struct svc_serv *serv, struct net *net)
978 svc_close_list(serv, &serv->sv_tempsocks, net);
979 svc_close_list(serv, &serv->sv_permsocks, net);
981 svc_clear_pools(serv, net);
983 * At this point the sp_sockets lists will stay empty, since
984 * svc_xprt_enqueue will not add new entries without taking the
985 * sp_lock and checking XPT_BUSY.
987 svc_clear_list(serv, &serv->sv_tempsocks, net);
988 svc_clear_list(serv, &serv->sv_permsocks, net);
992 * Handle defer and revisit of requests
995 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
997 struct svc_deferred_req *dr =
998 container_of(dreq, struct svc_deferred_req, handle);
999 struct svc_xprt *xprt = dr->xprt;
1001 spin_lock(&xprt->xpt_lock);
1002 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1003 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1004 spin_unlock(&xprt->xpt_lock);
1005 dprintk("revisit canceled\n");
1010 dprintk("revisit queued\n");
1012 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1013 spin_unlock(&xprt->xpt_lock);
1014 svc_xprt_enqueue(xprt);
1019 * Save the request off for later processing. The request buffer looks
1022 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1024 * This code can only handle requests that consist of an xprt-header
1027 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1029 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1030 struct svc_deferred_req *dr;
1032 if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
1033 return NULL; /* if more than a page, give up FIXME */
1034 if (rqstp->rq_deferred) {
1035 dr = rqstp->rq_deferred;
1036 rqstp->rq_deferred = NULL;
1040 /* FIXME maybe discard if size too large */
1041 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1042 dr = kmalloc(size, GFP_KERNEL);
1046 dr->handle.owner = rqstp->rq_server;
1047 dr->prot = rqstp->rq_prot;
1048 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1049 dr->addrlen = rqstp->rq_addrlen;
1050 dr->daddr = rqstp->rq_daddr;
1051 dr->argslen = rqstp->rq_arg.len >> 2;
1052 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1054 /* back up head to the start of the buffer and copy */
1055 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1056 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1059 svc_xprt_get(rqstp->rq_xprt);
1060 dr->xprt = rqstp->rq_xprt;
1061 rqstp->rq_dropme = true;
1063 dr->handle.revisit = svc_revisit;
1068 * recv data from a deferred request into an active one
1070 static int svc_deferred_recv(struct svc_rqst *rqstp)
1072 struct svc_deferred_req *dr = rqstp->rq_deferred;
1074 /* setup iov_base past transport header */
1075 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1076 /* The iov_len does not include the transport header bytes */
1077 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1078 rqstp->rq_arg.page_len = 0;
1079 /* The rq_arg.len includes the transport header bytes */
1080 rqstp->rq_arg.len = dr->argslen<<2;
1081 rqstp->rq_prot = dr->prot;
1082 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1083 rqstp->rq_addrlen = dr->addrlen;
1084 /* Save off transport header len in case we get deferred again */
1085 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1086 rqstp->rq_daddr = dr->daddr;
1087 rqstp->rq_respages = rqstp->rq_pages;
1088 return (dr->argslen<<2) - dr->xprt_hlen;
1092 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1094 struct svc_deferred_req *dr = NULL;
1096 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1098 spin_lock(&xprt->xpt_lock);
1099 if (!list_empty(&xprt->xpt_deferred)) {
1100 dr = list_entry(xprt->xpt_deferred.next,
1101 struct svc_deferred_req,
1103 list_del_init(&dr->handle.recent);
1105 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1106 spin_unlock(&xprt->xpt_lock);
1111 * svc_find_xprt - find an RPC transport instance
1112 * @serv: pointer to svc_serv to search
1113 * @xcl_name: C string containing transport's class name
1114 * @net: owner net pointer
1115 * @af: Address family of transport's local address
1116 * @port: transport's IP port number
1118 * Return the transport instance pointer for the endpoint accepting
1119 * connections/peer traffic from the specified transport class,
1120 * address family and port.
1122 * Specifying 0 for the address family or port is effectively a
1123 * wild-card, and will result in matching the first transport in the
1124 * service's list that has a matching class name.
1126 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1127 struct net *net, const sa_family_t af,
1128 const unsigned short port)
1130 struct svc_xprt *xprt;
1131 struct svc_xprt *found = NULL;
1133 /* Sanity check the args */
1134 if (serv == NULL || xcl_name == NULL)
1137 spin_lock_bh(&serv->sv_lock);
1138 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1139 if (xprt->xpt_net != net)
1141 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1143 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1145 if (port != 0 && port != svc_xprt_local_port(xprt))
1151 spin_unlock_bh(&serv->sv_lock);
1154 EXPORT_SYMBOL_GPL(svc_find_xprt);
1156 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1157 char *pos, int remaining)
1161 len = snprintf(pos, remaining, "%s %u\n",
1162 xprt->xpt_class->xcl_name,
1163 svc_xprt_local_port(xprt));
1164 if (len >= remaining)
1165 return -ENAMETOOLONG;
1170 * svc_xprt_names - format a buffer with a list of transport names
1171 * @serv: pointer to an RPC service
1172 * @buf: pointer to a buffer to be filled in
1173 * @buflen: length of buffer to be filled in
1175 * Fills in @buf with a string containing a list of transport names,
1176 * each name terminated with '\n'.
1178 * Returns positive length of the filled-in string on success; otherwise
1179 * a negative errno value is returned if an error occurs.
1181 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1183 struct svc_xprt *xprt;
1187 /* Sanity check args */
1191 spin_lock_bh(&serv->sv_lock);
1195 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1196 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1208 spin_unlock_bh(&serv->sv_lock);
1211 EXPORT_SYMBOL_GPL(svc_xprt_names);
1214 /*----------------------------------------------------------------------------*/
1216 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1218 unsigned int pidx = (unsigned int)*pos;
1219 struct svc_serv *serv = m->private;
1221 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1224 return SEQ_START_TOKEN;
1225 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1228 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1230 struct svc_pool *pool = p;
1231 struct svc_serv *serv = m->private;
1233 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1235 if (p == SEQ_START_TOKEN) {
1236 pool = &serv->sv_pools[0];
1238 unsigned int pidx = (pool - &serv->sv_pools[0]);
1239 if (pidx < serv->sv_nrpools-1)
1240 pool = &serv->sv_pools[pidx+1];
1248 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1252 static int svc_pool_stats_show(struct seq_file *m, void *p)
1254 struct svc_pool *pool = p;
1256 if (p == SEQ_START_TOKEN) {
1257 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1261 seq_printf(m, "%u %lu %lu %lu %lu\n",
1263 pool->sp_stats.packets,
1264 pool->sp_stats.sockets_queued,
1265 pool->sp_stats.threads_woken,
1266 pool->sp_stats.threads_timedout);
1271 static const struct seq_operations svc_pool_stats_seq_ops = {
1272 .start = svc_pool_stats_start,
1273 .next = svc_pool_stats_next,
1274 .stop = svc_pool_stats_stop,
1275 .show = svc_pool_stats_show,
1278 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1282 err = seq_open(file, &svc_pool_stats_seq_ops);
1284 ((struct seq_file *) file->private_data)->private = serv;
1287 EXPORT_SYMBOL(svc_pool_stats_open);
1289 /*----------------------------------------------------------------------------*/