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);
212 * svc_xprt_received conditionally queues the transport for processing
213 * by another thread. The caller must hold the XPT_BUSY bit and must
214 * not thereafter touch transport data.
216 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
217 * insufficient) data.
219 static void svc_xprt_received(struct svc_xprt *xprt)
221 WARN_ON_ONCE(!test_bit(XPT_BUSY, &xprt->xpt_flags));
222 if (!test_bit(XPT_BUSY, &xprt->xpt_flags))
224 /* As soon as we clear busy, the xprt could be closed and
225 * 'put', so we need a reference to call svc_xprt_enqueue with:
228 clear_bit(XPT_BUSY, &xprt->xpt_flags);
229 svc_xprt_enqueue(xprt);
233 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
235 clear_bit(XPT_TEMP, &new->xpt_flags);
236 spin_lock_bh(&serv->sv_lock);
237 list_add(&new->xpt_list, &serv->sv_permsocks);
238 spin_unlock_bh(&serv->sv_lock);
239 svc_xprt_received(new);
242 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
243 struct net *net, const int family,
244 const unsigned short port, int flags)
246 struct svc_xprt_class *xcl;
248 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
249 spin_lock(&svc_xprt_class_lock);
250 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
251 struct svc_xprt *newxprt;
252 unsigned short newport;
254 if (strcmp(xprt_name, xcl->xcl_name))
257 if (!try_module_get(xcl->xcl_owner))
260 spin_unlock(&svc_xprt_class_lock);
261 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
262 if (IS_ERR(newxprt)) {
263 module_put(xcl->xcl_owner);
264 return PTR_ERR(newxprt);
266 svc_add_new_perm_xprt(serv, newxprt);
267 newport = svc_xprt_local_port(newxprt);
271 spin_unlock(&svc_xprt_class_lock);
272 dprintk("svc: transport %s not found\n", xprt_name);
274 /* This errno is exposed to user space. Provide a reasonable
275 * perror msg for a bad transport. */
276 return -EPROTONOSUPPORT;
278 EXPORT_SYMBOL_GPL(svc_create_xprt);
281 * Copy the local and remote xprt addresses to the rqstp structure
283 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
285 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
286 rqstp->rq_addrlen = xprt->xpt_remotelen;
289 * Destination address in request is needed for binding the
290 * source address in RPC replies/callbacks later.
292 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
293 rqstp->rq_daddrlen = xprt->xpt_locallen;
295 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
298 * svc_print_addr - Format rq_addr field for printing
299 * @rqstp: svc_rqst struct containing address to print
300 * @buf: target buffer for formatted address
301 * @len: length of target buffer
304 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
306 return __svc_print_addr(svc_addr(rqstp), buf, len);
308 EXPORT_SYMBOL_GPL(svc_print_addr);
311 * Queue up an idle server thread. Must have pool->sp_lock held.
312 * Note: this is really a stack rather than a queue, so that we only
313 * use as many different threads as we need, and the rest don't pollute
316 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
318 list_add(&rqstp->rq_list, &pool->sp_threads);
322 * Dequeue an nfsd thread. Must have pool->sp_lock held.
324 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
326 list_del(&rqstp->rq_list);
329 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
331 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
333 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
334 return xprt->xpt_ops->xpo_has_wspace(xprt);
339 * Queue up a transport with data pending. If there are idle nfsd
340 * processes, wake 'em up.
343 void svc_xprt_enqueue(struct svc_xprt *xprt)
345 struct svc_pool *pool;
346 struct svc_rqst *rqstp;
349 if (!svc_xprt_has_something_to_do(xprt))
353 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
356 spin_lock_bh(&pool->sp_lock);
358 if (!list_empty(&pool->sp_threads) &&
359 !list_empty(&pool->sp_sockets))
362 "threads and transports both waiting??\n");
364 pool->sp_stats.packets++;
366 /* Mark transport as busy. It will remain in this state until
367 * the provider calls svc_xprt_received. We update XPT_BUSY
368 * atomically because it also guards against trying to enqueue
369 * the transport twice.
371 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
372 /* Don't enqueue transport while already enqueued */
373 dprintk("svc: transport %p busy, not enqueued\n", xprt);
377 if (!list_empty(&pool->sp_threads)) {
378 rqstp = list_entry(pool->sp_threads.next,
381 dprintk("svc: transport %p served by daemon %p\n",
383 svc_thread_dequeue(pool, rqstp);
386 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
387 rqstp, rqstp->rq_xprt);
388 rqstp->rq_xprt = xprt;
390 pool->sp_stats.threads_woken++;
391 wake_up(&rqstp->rq_wait);
393 dprintk("svc: transport %p put into queue\n", xprt);
394 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
395 pool->sp_stats.sockets_queued++;
399 spin_unlock_bh(&pool->sp_lock);
401 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
404 * Dequeue the first transport. Must be called with the pool->sp_lock held.
406 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
408 struct svc_xprt *xprt;
410 if (list_empty(&pool->sp_sockets))
413 xprt = list_entry(pool->sp_sockets.next,
414 struct svc_xprt, xpt_ready);
415 list_del_init(&xprt->xpt_ready);
417 dprintk("svc: transport %p dequeued, inuse=%d\n",
418 xprt, atomic_read(&xprt->xpt_ref.refcount));
424 * svc_reserve - change the space reserved for the reply to a request.
425 * @rqstp: The request in question
426 * @space: new max space to reserve
428 * Each request reserves some space on the output queue of the transport
429 * to make sure the reply fits. This function reduces that reserved
430 * space to be the amount of space used already, plus @space.
433 void svc_reserve(struct svc_rqst *rqstp, int space)
435 space += rqstp->rq_res.head[0].iov_len;
437 if (space < rqstp->rq_reserved) {
438 struct svc_xprt *xprt = rqstp->rq_xprt;
439 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
440 rqstp->rq_reserved = space;
442 svc_xprt_enqueue(xprt);
445 EXPORT_SYMBOL_GPL(svc_reserve);
447 static void svc_xprt_release(struct svc_rqst *rqstp)
449 struct svc_xprt *xprt = rqstp->rq_xprt;
451 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
453 kfree(rqstp->rq_deferred);
454 rqstp->rq_deferred = NULL;
456 svc_free_res_pages(rqstp);
457 rqstp->rq_res.page_len = 0;
458 rqstp->rq_res.page_base = 0;
460 /* Reset response buffer and release
462 * But first, check that enough space was reserved
463 * for the reply, otherwise we have a bug!
465 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
466 printk(KERN_ERR "RPC request reserved %d but used %d\n",
470 rqstp->rq_res.head[0].iov_len = 0;
471 svc_reserve(rqstp, 0);
472 rqstp->rq_xprt = NULL;
478 * External function to wake up a server waiting for data
479 * This really only makes sense for services like lockd
480 * which have exactly one thread anyway.
482 void svc_wake_up(struct svc_serv *serv)
484 struct svc_rqst *rqstp;
486 struct svc_pool *pool;
488 for (i = 0; i < serv->sv_nrpools; i++) {
489 pool = &serv->sv_pools[i];
491 spin_lock_bh(&pool->sp_lock);
492 if (!list_empty(&pool->sp_threads)) {
493 rqstp = list_entry(pool->sp_threads.next,
496 dprintk("svc: daemon %p woken up.\n", rqstp);
498 svc_thread_dequeue(pool, rqstp);
499 rqstp->rq_xprt = NULL;
501 wake_up(&rqstp->rq_wait);
503 spin_unlock_bh(&pool->sp_lock);
506 EXPORT_SYMBOL_GPL(svc_wake_up);
508 int svc_port_is_privileged(struct sockaddr *sin)
510 switch (sin->sa_family) {
512 return ntohs(((struct sockaddr_in *)sin)->sin_port)
515 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
523 * Make sure that we don't have too many active connections. If we have,
524 * something must be dropped. It's not clear what will happen if we allow
525 * "too many" connections, but when dealing with network-facing software,
526 * we have to code defensively. Here we do that by imposing hard limits.
528 * There's no point in trying to do random drop here for DoS
529 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
530 * attacker can easily beat that.
532 * The only somewhat efficient mechanism would be if drop old
533 * connections from the same IP first. But right now we don't even
534 * record the client IP in svc_sock.
536 * single-threaded services that expect a lot of clients will probably
537 * need to set sv_maxconn to override the default value which is based
538 * on the number of threads
540 static void svc_check_conn_limits(struct svc_serv *serv)
542 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
543 (serv->sv_nrthreads+3) * 20;
545 if (serv->sv_tmpcnt > limit) {
546 struct svc_xprt *xprt = NULL;
547 spin_lock_bh(&serv->sv_lock);
548 if (!list_empty(&serv->sv_tempsocks)) {
549 /* Try to help the admin */
550 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
551 serv->sv_name, serv->sv_maxconn ?
552 "max number of connections" :
553 "number of threads");
555 * Always select the oldest connection. It's not fair,
558 xprt = list_entry(serv->sv_tempsocks.prev,
561 set_bit(XPT_CLOSE, &xprt->xpt_flags);
564 spin_unlock_bh(&serv->sv_lock);
567 svc_xprt_enqueue(xprt);
573 int svc_alloc_arg(struct svc_rqst *rqstp)
575 struct svc_serv *serv = rqstp->rq_server;
580 /* now allocate needed pages. If we get a failure, sleep briefly */
581 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
582 BUG_ON(pages >= RPCSVC_MAXPAGES);
583 for (i = 0; i < pages ; i++)
584 while (rqstp->rq_pages[i] == NULL) {
585 struct page *p = alloc_page(GFP_KERNEL);
587 set_current_state(TASK_INTERRUPTIBLE);
588 if (signalled() || kthread_should_stop()) {
589 set_current_state(TASK_RUNNING);
592 schedule_timeout(msecs_to_jiffies(500));
594 rqstp->rq_pages[i] = p;
596 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
598 /* Make arg->head point to first page and arg->pages point to rest */
599 arg = &rqstp->rq_arg;
600 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
601 arg->head[0].iov_len = PAGE_SIZE;
602 arg->pages = rqstp->rq_pages + 1;
604 /* save at least one page for response */
605 arg->page_len = (pages-2)*PAGE_SIZE;
606 arg->len = (pages-1)*PAGE_SIZE;
607 arg->tail[0].iov_len = 0;
611 struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
613 struct svc_xprt *xprt;
614 struct svc_pool *pool = rqstp->rq_pool;
615 DECLARE_WAITQUEUE(wait, current);
618 /* Normally we will wait up to 5 seconds for any required
619 * cache information to be provided.
621 rqstp->rq_chandle.thread_wait = 5*HZ;
623 spin_lock_bh(&pool->sp_lock);
624 xprt = svc_xprt_dequeue(pool);
626 rqstp->rq_xprt = xprt;
629 /* As there is a shortage of threads and this request
630 * had to be queued, don't allow the thread to wait so
631 * long for cache updates.
633 rqstp->rq_chandle.thread_wait = 1*HZ;
635 /* No data pending. Go to sleep */
636 svc_thread_enqueue(pool, rqstp);
639 * We have to be able to interrupt this wait
640 * to bring down the daemons ...
642 set_current_state(TASK_INTERRUPTIBLE);
645 * checking kthread_should_stop() here allows us to avoid
646 * locking and signalling when stopping kthreads that call
647 * svc_recv. If the thread has already been woken up, then
648 * we can exit here without sleeping. If not, then it
649 * it'll be woken up quickly during the schedule_timeout
651 if (kthread_should_stop()) {
652 set_current_state(TASK_RUNNING);
653 spin_unlock_bh(&pool->sp_lock);
654 return ERR_PTR(-EINTR);
657 add_wait_queue(&rqstp->rq_wait, &wait);
658 spin_unlock_bh(&pool->sp_lock);
660 time_left = schedule_timeout(timeout);
664 spin_lock_bh(&pool->sp_lock);
665 remove_wait_queue(&rqstp->rq_wait, &wait);
667 pool->sp_stats.threads_timedout++;
669 xprt = rqstp->rq_xprt;
671 svc_thread_dequeue(pool, rqstp);
672 spin_unlock_bh(&pool->sp_lock);
673 dprintk("svc: server %p, no data yet\n", rqstp);
674 if (signalled() || kthread_should_stop())
675 return ERR_PTR(-EINTR);
677 return ERR_PTR(-EAGAIN);
680 spin_unlock_bh(&pool->sp_lock);
684 void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
686 spin_lock_bh(&serv->sv_lock);
687 set_bit(XPT_TEMP, &newxpt->xpt_flags);
688 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
690 if (serv->sv_temptimer.function == NULL) {
691 /* setup timer to age temp transports */
692 setup_timer(&serv->sv_temptimer, svc_age_temp_xprts,
693 (unsigned long)serv);
694 mod_timer(&serv->sv_temptimer,
695 jiffies + svc_conn_age_period * HZ);
697 spin_unlock_bh(&serv->sv_lock);
698 svc_xprt_received(newxpt);
701 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
703 struct svc_serv *serv = rqstp->rq_server;
706 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
707 dprintk("svc_recv: found XPT_CLOSE\n");
708 svc_delete_xprt(xprt);
709 /* Leave XPT_BUSY set on the dead xprt: */
712 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
713 struct svc_xprt *newxpt;
715 * We know this module_get will succeed because the
716 * listener holds a reference too
718 __module_get(xprt->xpt_class->xcl_owner);
719 svc_check_conn_limits(xprt->xpt_server);
720 newxpt = xprt->xpt_ops->xpo_accept(xprt);
722 svc_add_new_temp_xprt(serv, newxpt);
723 } else if (xprt->xpt_ops->xpo_has_wspace(xprt)) {
724 /* XPT_DATA|XPT_DEFERRED case: */
725 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
726 rqstp, rqstp->rq_pool->sp_id, xprt,
727 atomic_read(&xprt->xpt_ref.refcount));
728 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
729 if (rqstp->rq_deferred)
730 len = svc_deferred_recv(rqstp);
732 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
733 dprintk("svc: got len=%d\n", len);
734 rqstp->rq_reserved = serv->sv_max_mesg;
735 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
737 /* clear XPT_BUSY: */
738 svc_xprt_received(xprt);
743 * Receive the next request on any transport. This code is carefully
744 * organised not to touch any cachelines in the shared svc_serv
745 * structure, only cachelines in the local svc_pool.
747 int svc_recv(struct svc_rqst *rqstp, long timeout)
749 struct svc_xprt *xprt = NULL;
750 struct svc_serv *serv = rqstp->rq_server;
753 dprintk("svc: server %p waiting for data (to = %ld)\n",
758 "svc_recv: service %p, transport not NULL!\n",
760 if (waitqueue_active(&rqstp->rq_wait))
762 "svc_recv: service %p, wait queue active!\n",
765 err = svc_alloc_arg(rqstp);
771 if (signalled() || kthread_should_stop())
774 xprt = svc_get_next_xprt(rqstp, timeout);
776 return PTR_ERR(xprt);
778 len = svc_handle_xprt(rqstp, xprt);
780 /* No data, incomplete (TCP) read, or accept() */
784 clear_bit(XPT_OLD, &xprt->xpt_flags);
786 rqstp->rq_secure = svc_port_is_privileged(svc_addr(rqstp));
787 rqstp->rq_chandle.defer = svc_defer;
790 serv->sv_stats->netcnt++;
793 rqstp->rq_res.len = 0;
794 svc_xprt_release(rqstp);
797 EXPORT_SYMBOL_GPL(svc_recv);
802 void svc_drop(struct svc_rqst *rqstp)
804 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
805 svc_xprt_release(rqstp);
807 EXPORT_SYMBOL_GPL(svc_drop);
810 * Return reply to client.
812 int svc_send(struct svc_rqst *rqstp)
814 struct svc_xprt *xprt;
818 xprt = rqstp->rq_xprt;
822 /* release the receive skb before sending the reply */
823 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
825 /* calculate over-all length */
827 xb->len = xb->head[0].iov_len +
831 /* Grab mutex to serialize outgoing data. */
832 mutex_lock(&xprt->xpt_mutex);
833 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
834 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
837 len = xprt->xpt_ops->xpo_sendto(rqstp);
838 mutex_unlock(&xprt->xpt_mutex);
839 rpc_wake_up(&xprt->xpt_bc_pending);
840 svc_xprt_release(rqstp);
842 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
848 * Timer function to close old temporary transports, using
849 * a mark-and-sweep algorithm.
851 static void svc_age_temp_xprts(unsigned long closure)
853 struct svc_serv *serv = (struct svc_serv *)closure;
854 struct svc_xprt *xprt;
855 struct list_head *le, *next;
856 LIST_HEAD(to_be_aged);
858 dprintk("svc_age_temp_xprts\n");
860 if (!spin_trylock_bh(&serv->sv_lock)) {
861 /* busy, try again 1 sec later */
862 dprintk("svc_age_temp_xprts: busy\n");
863 mod_timer(&serv->sv_temptimer, jiffies + HZ);
867 list_for_each_safe(le, next, &serv->sv_tempsocks) {
868 xprt = list_entry(le, struct svc_xprt, xpt_list);
870 /* First time through, just mark it OLD. Second time
871 * through, close it. */
872 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
874 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
875 test_bit(XPT_BUSY, &xprt->xpt_flags))
878 list_move(le, &to_be_aged);
879 set_bit(XPT_CLOSE, &xprt->xpt_flags);
880 set_bit(XPT_DETACHED, &xprt->xpt_flags);
882 spin_unlock_bh(&serv->sv_lock);
884 while (!list_empty(&to_be_aged)) {
885 le = to_be_aged.next;
886 /* fiddling the xpt_list node is safe 'cos we're XPT_DETACHED */
888 xprt = list_entry(le, struct svc_xprt, xpt_list);
890 dprintk("queuing xprt %p for closing\n", xprt);
892 /* a thread will dequeue and close it soon */
893 svc_xprt_enqueue(xprt);
897 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
900 static void call_xpt_users(struct svc_xprt *xprt)
902 struct svc_xpt_user *u;
904 spin_lock(&xprt->xpt_lock);
905 while (!list_empty(&xprt->xpt_users)) {
906 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
910 spin_unlock(&xprt->xpt_lock);
914 * Remove a dead transport
916 static void svc_delete_xprt(struct svc_xprt *xprt)
918 struct svc_serv *serv = xprt->xpt_server;
919 struct svc_deferred_req *dr;
921 /* Only do this once */
922 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
925 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
926 xprt->xpt_ops->xpo_detach(xprt);
928 spin_lock_bh(&serv->sv_lock);
929 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
930 list_del_init(&xprt->xpt_list);
931 BUG_ON(!list_empty(&xprt->xpt_ready));
932 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
934 spin_unlock_bh(&serv->sv_lock);
936 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
939 call_xpt_users(xprt);
943 void svc_close_xprt(struct svc_xprt *xprt)
945 set_bit(XPT_CLOSE, &xprt->xpt_flags);
946 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
947 /* someone else will have to effect the close */
950 * We expect svc_close_xprt() to work even when no threads are
951 * running (e.g., while configuring the server before starting
952 * any threads), so if the transport isn't busy, we delete
955 svc_delete_xprt(xprt);
957 EXPORT_SYMBOL_GPL(svc_close_xprt);
959 static void svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
961 struct svc_xprt *xprt;
963 spin_lock(&serv->sv_lock);
964 list_for_each_entry(xprt, xprt_list, xpt_list) {
965 if (xprt->xpt_net != net)
967 set_bit(XPT_CLOSE, &xprt->xpt_flags);
968 set_bit(XPT_BUSY, &xprt->xpt_flags);
970 spin_unlock(&serv->sv_lock);
973 static void svc_clear_pools(struct svc_serv *serv, struct net *net)
975 struct svc_pool *pool;
976 struct svc_xprt *xprt;
977 struct svc_xprt *tmp;
980 for (i = 0; i < serv->sv_nrpools; i++) {
981 pool = &serv->sv_pools[i];
983 spin_lock_bh(&pool->sp_lock);
984 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
985 if (xprt->xpt_net != net)
987 list_del_init(&xprt->xpt_ready);
989 spin_unlock_bh(&pool->sp_lock);
993 static void svc_clear_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
995 struct svc_xprt *xprt;
996 struct svc_xprt *tmp;
999 spin_lock(&serv->sv_lock);
1000 list_for_each_entry_safe(xprt, tmp, xprt_list, xpt_list) {
1001 if (xprt->xpt_net != net)
1003 list_move(&xprt->xpt_list, &victims);
1005 spin_unlock(&serv->sv_lock);
1007 list_for_each_entry_safe(xprt, tmp, &victims, xpt_list)
1008 svc_delete_xprt(xprt);
1011 void svc_close_net(struct svc_serv *serv, struct net *net)
1013 svc_close_list(serv, &serv->sv_tempsocks, net);
1014 svc_close_list(serv, &serv->sv_permsocks, net);
1016 svc_clear_pools(serv, net);
1018 * At this point the sp_sockets lists will stay empty, since
1019 * svc_xprt_enqueue will not add new entries without taking the
1020 * sp_lock and checking XPT_BUSY.
1022 svc_clear_list(serv, &serv->sv_tempsocks, net);
1023 svc_clear_list(serv, &serv->sv_permsocks, net);
1027 * Handle defer and revisit of requests
1030 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1032 struct svc_deferred_req *dr =
1033 container_of(dreq, struct svc_deferred_req, handle);
1034 struct svc_xprt *xprt = dr->xprt;
1036 spin_lock(&xprt->xpt_lock);
1037 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1038 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1039 spin_unlock(&xprt->xpt_lock);
1040 dprintk("revisit canceled\n");
1045 dprintk("revisit queued\n");
1047 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1048 spin_unlock(&xprt->xpt_lock);
1049 svc_xprt_enqueue(xprt);
1054 * Save the request off for later processing. The request buffer looks
1057 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1059 * This code can only handle requests that consist of an xprt-header
1062 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1064 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1065 struct svc_deferred_req *dr;
1067 if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
1068 return NULL; /* if more than a page, give up FIXME */
1069 if (rqstp->rq_deferred) {
1070 dr = rqstp->rq_deferred;
1071 rqstp->rq_deferred = NULL;
1075 /* FIXME maybe discard if size too large */
1076 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1077 dr = kmalloc(size, GFP_KERNEL);
1081 dr->handle.owner = rqstp->rq_server;
1082 dr->prot = rqstp->rq_prot;
1083 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1084 dr->addrlen = rqstp->rq_addrlen;
1085 dr->daddr = rqstp->rq_daddr;
1086 dr->argslen = rqstp->rq_arg.len >> 2;
1087 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1089 /* back up head to the start of the buffer and copy */
1090 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1091 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1094 svc_xprt_get(rqstp->rq_xprt);
1095 dr->xprt = rqstp->rq_xprt;
1096 rqstp->rq_dropme = true;
1098 dr->handle.revisit = svc_revisit;
1103 * recv data from a deferred request into an active one
1105 static int svc_deferred_recv(struct svc_rqst *rqstp)
1107 struct svc_deferred_req *dr = rqstp->rq_deferred;
1109 /* setup iov_base past transport header */
1110 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1111 /* The iov_len does not include the transport header bytes */
1112 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1113 rqstp->rq_arg.page_len = 0;
1114 /* The rq_arg.len includes the transport header bytes */
1115 rqstp->rq_arg.len = dr->argslen<<2;
1116 rqstp->rq_prot = dr->prot;
1117 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1118 rqstp->rq_addrlen = dr->addrlen;
1119 /* Save off transport header len in case we get deferred again */
1120 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1121 rqstp->rq_daddr = dr->daddr;
1122 rqstp->rq_respages = rqstp->rq_pages;
1123 return (dr->argslen<<2) - dr->xprt_hlen;
1127 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1129 struct svc_deferred_req *dr = NULL;
1131 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1133 spin_lock(&xprt->xpt_lock);
1134 if (!list_empty(&xprt->xpt_deferred)) {
1135 dr = list_entry(xprt->xpt_deferred.next,
1136 struct svc_deferred_req,
1138 list_del_init(&dr->handle.recent);
1140 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1141 spin_unlock(&xprt->xpt_lock);
1146 * svc_find_xprt - find an RPC transport instance
1147 * @serv: pointer to svc_serv to search
1148 * @xcl_name: C string containing transport's class name
1149 * @net: owner net pointer
1150 * @af: Address family of transport's local address
1151 * @port: transport's IP port number
1153 * Return the transport instance pointer for the endpoint accepting
1154 * connections/peer traffic from the specified transport class,
1155 * address family and port.
1157 * Specifying 0 for the address family or port is effectively a
1158 * wild-card, and will result in matching the first transport in the
1159 * service's list that has a matching class name.
1161 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1162 struct net *net, const sa_family_t af,
1163 const unsigned short port)
1165 struct svc_xprt *xprt;
1166 struct svc_xprt *found = NULL;
1168 /* Sanity check the args */
1169 if (serv == NULL || xcl_name == NULL)
1172 spin_lock_bh(&serv->sv_lock);
1173 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1174 if (xprt->xpt_net != net)
1176 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1178 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1180 if (port != 0 && port != svc_xprt_local_port(xprt))
1186 spin_unlock_bh(&serv->sv_lock);
1189 EXPORT_SYMBOL_GPL(svc_find_xprt);
1191 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1192 char *pos, int remaining)
1196 len = snprintf(pos, remaining, "%s %u\n",
1197 xprt->xpt_class->xcl_name,
1198 svc_xprt_local_port(xprt));
1199 if (len >= remaining)
1200 return -ENAMETOOLONG;
1205 * svc_xprt_names - format a buffer with a list of transport names
1206 * @serv: pointer to an RPC service
1207 * @buf: pointer to a buffer to be filled in
1208 * @buflen: length of buffer to be filled in
1210 * Fills in @buf with a string containing a list of transport names,
1211 * each name terminated with '\n'.
1213 * Returns positive length of the filled-in string on success; otherwise
1214 * a negative errno value is returned if an error occurs.
1216 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1218 struct svc_xprt *xprt;
1222 /* Sanity check args */
1226 spin_lock_bh(&serv->sv_lock);
1230 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1231 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1243 spin_unlock_bh(&serv->sv_lock);
1246 EXPORT_SYMBOL_GPL(svc_xprt_names);
1249 /*----------------------------------------------------------------------------*/
1251 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1253 unsigned int pidx = (unsigned int)*pos;
1254 struct svc_serv *serv = m->private;
1256 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1259 return SEQ_START_TOKEN;
1260 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1263 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1265 struct svc_pool *pool = p;
1266 struct svc_serv *serv = m->private;
1268 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1270 if (p == SEQ_START_TOKEN) {
1271 pool = &serv->sv_pools[0];
1273 unsigned int pidx = (pool - &serv->sv_pools[0]);
1274 if (pidx < serv->sv_nrpools-1)
1275 pool = &serv->sv_pools[pidx+1];
1283 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1287 static int svc_pool_stats_show(struct seq_file *m, void *p)
1289 struct svc_pool *pool = p;
1291 if (p == SEQ_START_TOKEN) {
1292 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1296 seq_printf(m, "%u %lu %lu %lu %lu\n",
1298 pool->sp_stats.packets,
1299 pool->sp_stats.sockets_queued,
1300 pool->sp_stats.threads_woken,
1301 pool->sp_stats.threads_timedout);
1306 static const struct seq_operations svc_pool_stats_seq_ops = {
1307 .start = svc_pool_stats_start,
1308 .next = svc_pool_stats_next,
1309 .stop = svc_pool_stats_stop,
1310 .show = svc_pool_stats_show,
1313 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1317 err = seq_open(file, &svc_pool_stats_seq_ops);
1319 ((struct seq_file *) file->private_data)->private = serv;
1322 EXPORT_SYMBOL(svc_pool_stats_open);
1324 /*----------------------------------------------------------------------------*/