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/addr.h>
14 #include <linux/sunrpc/stats.h>
15 #include <linux/sunrpc/svc_xprt.h>
16 #include <linux/sunrpc/svcsock.h>
17 #include <linux/sunrpc/xprt.h>
18 #include <linux/module.h>
19 #include <linux/netdevice.h>
20 #include <trace/events/sunrpc.h>
22 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
24 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
25 static int svc_deferred_recv(struct svc_rqst *rqstp);
26 static struct cache_deferred_req *svc_defer(struct cache_req *req);
27 static void svc_age_temp_xprts(unsigned long closure);
28 static void svc_delete_xprt(struct svc_xprt *xprt);
30 /* apparently the "standard" is that clients close
31 * idle connections after 5 minutes, servers after
33 * http://www.connectathon.org/talks96/nfstcp.pdf
35 static int svc_conn_age_period = 6*60;
37 /* List of registered transport classes */
38 static DEFINE_SPINLOCK(svc_xprt_class_lock);
39 static LIST_HEAD(svc_xprt_class_list);
41 /* SMP locking strategy:
43 * svc_pool->sp_lock protects most of the fields of that pool.
44 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
45 * when both need to be taken (rare), svc_serv->sv_lock is first.
46 * The "service mutex" protects svc_serv->sv_nrthread.
47 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
48 * and the ->sk_info_authunix cache.
50 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
51 * enqueued multiply. During normal transport processing this bit
52 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
53 * Providers should not manipulate this bit directly.
55 * Some flags can be set to certain values at any time
56 * providing that certain rules are followed:
59 * - Can be set or cleared at any time.
60 * - After a set, svc_xprt_enqueue must be called to enqueue
61 * the transport for processing.
62 * - After a clear, the transport must be read/accepted.
63 * If this succeeds, it must be set again.
65 * - Can set at any time. It is never cleared.
67 * - Can only be set while XPT_BUSY is held which ensures
68 * that no other thread will be using the transport or will
69 * try to set XPT_DEAD.
71 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
73 struct svc_xprt_class *cl;
76 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
78 INIT_LIST_HEAD(&xcl->xcl_list);
79 spin_lock(&svc_xprt_class_lock);
80 /* Make sure there isn't already a class with the same name */
81 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
82 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
85 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
88 spin_unlock(&svc_xprt_class_lock);
91 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
93 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
95 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
96 spin_lock(&svc_xprt_class_lock);
97 list_del_init(&xcl->xcl_list);
98 spin_unlock(&svc_xprt_class_lock);
100 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
103 * Format the transport list for printing
105 int svc_print_xprts(char *buf, int maxlen)
107 struct svc_xprt_class *xcl;
112 spin_lock(&svc_xprt_class_lock);
113 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
116 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
117 slen = strlen(tmpstr);
118 if (len + slen > maxlen)
123 spin_unlock(&svc_xprt_class_lock);
128 static void svc_xprt_free(struct kref *kref)
130 struct svc_xprt *xprt =
131 container_of(kref, struct svc_xprt, xpt_ref);
132 struct module *owner = xprt->xpt_class->xcl_owner;
133 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
134 svcauth_unix_info_release(xprt);
135 put_net(xprt->xpt_net);
136 /* See comment on corresponding get in xs_setup_bc_tcp(): */
137 if (xprt->xpt_bc_xprt)
138 xprt_put(xprt->xpt_bc_xprt);
139 if (xprt->xpt_bc_xps)
140 xprt_switch_put(xprt->xpt_bc_xps);
141 xprt->xpt_ops->xpo_free(xprt);
145 void svc_xprt_put(struct svc_xprt *xprt)
147 kref_put(&xprt->xpt_ref, svc_xprt_free);
149 EXPORT_SYMBOL_GPL(svc_xprt_put);
152 * Called by transport drivers to initialize the transport independent
153 * portion of the transport instance.
155 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
156 struct svc_xprt *xprt, struct svc_serv *serv)
158 memset(xprt, 0, sizeof(*xprt));
159 xprt->xpt_class = xcl;
160 xprt->xpt_ops = xcl->xcl_ops;
161 kref_init(&xprt->xpt_ref);
162 xprt->xpt_server = serv;
163 INIT_LIST_HEAD(&xprt->xpt_list);
164 INIT_LIST_HEAD(&xprt->xpt_ready);
165 INIT_LIST_HEAD(&xprt->xpt_deferred);
166 INIT_LIST_HEAD(&xprt->xpt_users);
167 mutex_init(&xprt->xpt_mutex);
168 spin_lock_init(&xprt->xpt_lock);
169 set_bit(XPT_BUSY, &xprt->xpt_flags);
170 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
171 xprt->xpt_net = get_net(net);
173 EXPORT_SYMBOL_GPL(svc_xprt_init);
175 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
176 struct svc_serv *serv,
179 const unsigned short port,
182 struct sockaddr_in sin = {
183 .sin_family = AF_INET,
184 .sin_addr.s_addr = htonl(INADDR_ANY),
185 .sin_port = htons(port),
187 #if IS_ENABLED(CONFIG_IPV6)
188 struct sockaddr_in6 sin6 = {
189 .sin6_family = AF_INET6,
190 .sin6_addr = IN6ADDR_ANY_INIT,
191 .sin6_port = htons(port),
194 struct sockaddr *sap;
199 sap = (struct sockaddr *)&sin;
202 #if IS_ENABLED(CONFIG_IPV6)
204 sap = (struct sockaddr *)&sin6;
209 return ERR_PTR(-EAFNOSUPPORT);
212 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
216 * svc_xprt_received conditionally queues the transport for processing
217 * by another thread. The caller must hold the XPT_BUSY bit and must
218 * not thereafter touch transport data.
220 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
221 * insufficient) data.
223 static void svc_xprt_received(struct svc_xprt *xprt)
225 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
226 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
230 /* As soon as we clear busy, the xprt could be closed and
231 * 'put', so we need a reference to call svc_enqueue_xprt with:
234 smp_mb__before_atomic();
235 clear_bit(XPT_BUSY, &xprt->xpt_flags);
236 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
240 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
242 clear_bit(XPT_TEMP, &new->xpt_flags);
243 spin_lock_bh(&serv->sv_lock);
244 list_add(&new->xpt_list, &serv->sv_permsocks);
245 spin_unlock_bh(&serv->sv_lock);
246 svc_xprt_received(new);
249 int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
250 struct net *net, const int family,
251 const unsigned short port, int flags)
253 struct svc_xprt_class *xcl;
255 spin_lock(&svc_xprt_class_lock);
256 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
257 struct svc_xprt *newxprt;
258 unsigned short newport;
260 if (strcmp(xprt_name, xcl->xcl_name))
263 if (!try_module_get(xcl->xcl_owner))
266 spin_unlock(&svc_xprt_class_lock);
267 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
268 if (IS_ERR(newxprt)) {
269 module_put(xcl->xcl_owner);
270 return PTR_ERR(newxprt);
272 svc_add_new_perm_xprt(serv, newxprt);
273 newport = svc_xprt_local_port(newxprt);
277 spin_unlock(&svc_xprt_class_lock);
278 /* This errno is exposed to user space. Provide a reasonable
279 * perror msg for a bad transport. */
280 return -EPROTONOSUPPORT;
283 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
284 struct net *net, const int family,
285 const unsigned short port, int flags)
289 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
290 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
291 if (err == -EPROTONOSUPPORT) {
292 request_module("svc%s", xprt_name);
293 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
296 dprintk("svc: transport %s not found, err %d\n",
300 EXPORT_SYMBOL_GPL(svc_create_xprt);
303 * Copy the local and remote xprt addresses to the rqstp structure
305 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
307 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
308 rqstp->rq_addrlen = xprt->xpt_remotelen;
311 * Destination address in request is needed for binding the
312 * source address in RPC replies/callbacks later.
314 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
315 rqstp->rq_daddrlen = xprt->xpt_locallen;
317 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
320 * svc_print_addr - Format rq_addr field for printing
321 * @rqstp: svc_rqst struct containing address to print
322 * @buf: target buffer for formatted address
323 * @len: length of target buffer
326 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
328 return __svc_print_addr(svc_addr(rqstp), buf, len);
330 EXPORT_SYMBOL_GPL(svc_print_addr);
332 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
334 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
336 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
337 return xprt->xpt_ops->xpo_has_wspace(xprt);
341 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
343 struct svc_pool *pool;
344 struct svc_rqst *rqstp = NULL;
348 if (!svc_xprt_has_something_to_do(xprt))
351 /* Mark transport as busy. It will remain in this state until
352 * the provider calls svc_xprt_received. We update XPT_BUSY
353 * atomically because it also guards against trying to enqueue
354 * the transport twice.
356 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
357 /* Don't enqueue transport while already enqueued */
358 dprintk("svc: transport %p busy, not enqueued\n", xprt);
363 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
365 atomic_long_inc(&pool->sp_stats.packets);
368 /* find a thread for this xprt */
370 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
371 /* Do a lockless check first */
372 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
376 * Once the xprt has been queued, it can only be dequeued by
377 * the task that intends to service it. All we can do at that
378 * point is to try to wake this thread back up so that it can
382 spin_lock_bh(&rqstp->rq_lock);
383 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags)) {
384 /* already busy, move on... */
385 spin_unlock_bh(&rqstp->rq_lock);
389 /* this one will do */
390 rqstp->rq_xprt = xprt;
392 spin_unlock_bh(&rqstp->rq_lock);
396 atomic_long_inc(&pool->sp_stats.threads_woken);
397 wake_up_process(rqstp->rq_task);
404 * We didn't find an idle thread to use, so we need to queue the xprt.
405 * Do so and then search again. If we find one, we can't hook this one
406 * up to it directly but we can wake the thread up in the hopes that it
407 * will pick it up once it searches for a xprt to service.
411 dprintk("svc: transport %p put into queue\n", xprt);
412 spin_lock_bh(&pool->sp_lock);
413 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
414 pool->sp_stats.sockets_queued++;
415 spin_unlock_bh(&pool->sp_lock);
421 trace_svc_xprt_do_enqueue(xprt, rqstp);
423 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
426 * Queue up a transport with data pending. If there are idle nfsd
427 * processes, wake 'em up.
430 void svc_xprt_enqueue(struct svc_xprt *xprt)
432 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
434 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
436 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
439 * Dequeue the first transport, if there is one.
441 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
443 struct svc_xprt *xprt = NULL;
445 if (list_empty(&pool->sp_sockets))
448 spin_lock_bh(&pool->sp_lock);
449 if (likely(!list_empty(&pool->sp_sockets))) {
450 xprt = list_first_entry(&pool->sp_sockets,
451 struct svc_xprt, xpt_ready);
452 list_del_init(&xprt->xpt_ready);
455 dprintk("svc: transport %p dequeued, inuse=%d\n",
456 xprt, atomic_read(&xprt->xpt_ref.refcount));
458 spin_unlock_bh(&pool->sp_lock);
460 trace_svc_xprt_dequeue(xprt);
465 * svc_reserve - change the space reserved for the reply to a request.
466 * @rqstp: The request in question
467 * @space: new max space to reserve
469 * Each request reserves some space on the output queue of the transport
470 * to make sure the reply fits. This function reduces that reserved
471 * space to be the amount of space used already, plus @space.
474 void svc_reserve(struct svc_rqst *rqstp, int space)
476 space += rqstp->rq_res.head[0].iov_len;
478 if (space < rqstp->rq_reserved) {
479 struct svc_xprt *xprt = rqstp->rq_xprt;
480 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
481 rqstp->rq_reserved = space;
483 if (xprt->xpt_ops->xpo_adjust_wspace)
484 xprt->xpt_ops->xpo_adjust_wspace(xprt);
485 svc_xprt_enqueue(xprt);
488 EXPORT_SYMBOL_GPL(svc_reserve);
490 static void svc_xprt_release(struct svc_rqst *rqstp)
492 struct svc_xprt *xprt = rqstp->rq_xprt;
494 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
496 kfree(rqstp->rq_deferred);
497 rqstp->rq_deferred = NULL;
499 svc_free_res_pages(rqstp);
500 rqstp->rq_res.page_len = 0;
501 rqstp->rq_res.page_base = 0;
503 /* Reset response buffer and release
505 * But first, check that enough space was reserved
506 * for the reply, otherwise we have a bug!
508 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
509 printk(KERN_ERR "RPC request reserved %d but used %d\n",
513 rqstp->rq_res.head[0].iov_len = 0;
514 svc_reserve(rqstp, 0);
515 rqstp->rq_xprt = NULL;
521 * Some svc_serv's will have occasional work to do, even when a xprt is not
522 * waiting to be serviced. This function is there to "kick" a task in one of
523 * those services so that it can wake up and do that work. Note that we only
524 * bother with pool 0 as we don't need to wake up more than one thread for
527 void svc_wake_up(struct svc_serv *serv)
529 struct svc_rqst *rqstp;
530 struct svc_pool *pool;
532 pool = &serv->sv_pools[0];
535 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
536 /* skip any that aren't queued */
537 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
540 dprintk("svc: daemon %p woken up.\n", rqstp);
541 wake_up_process(rqstp->rq_task);
542 trace_svc_wake_up(rqstp->rq_task->pid);
547 /* No free entries available */
548 set_bit(SP_TASK_PENDING, &pool->sp_flags);
550 trace_svc_wake_up(0);
552 EXPORT_SYMBOL_GPL(svc_wake_up);
554 int svc_port_is_privileged(struct sockaddr *sin)
556 switch (sin->sa_family) {
558 return ntohs(((struct sockaddr_in *)sin)->sin_port)
561 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
569 * Make sure that we don't have too many active connections. If we have,
570 * something must be dropped. It's not clear what will happen if we allow
571 * "too many" connections, but when dealing with network-facing software,
572 * we have to code defensively. Here we do that by imposing hard limits.
574 * There's no point in trying to do random drop here for DoS
575 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
576 * attacker can easily beat that.
578 * The only somewhat efficient mechanism would be if drop old
579 * connections from the same IP first. But right now we don't even
580 * record the client IP in svc_sock.
582 * single-threaded services that expect a lot of clients will probably
583 * need to set sv_maxconn to override the default value which is based
584 * on the number of threads
586 static void svc_check_conn_limits(struct svc_serv *serv)
588 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
589 (serv->sv_nrthreads+3) * 20;
591 if (serv->sv_tmpcnt > limit) {
592 struct svc_xprt *xprt = NULL;
593 spin_lock_bh(&serv->sv_lock);
594 if (!list_empty(&serv->sv_tempsocks)) {
595 /* Try to help the admin */
596 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
597 serv->sv_name, serv->sv_maxconn ?
598 "max number of connections" :
599 "number of threads");
601 * Always select the oldest connection. It's not fair,
604 xprt = list_entry(serv->sv_tempsocks.prev,
607 set_bit(XPT_CLOSE, &xprt->xpt_flags);
610 spin_unlock_bh(&serv->sv_lock);
613 svc_xprt_enqueue(xprt);
619 static int svc_alloc_arg(struct svc_rqst *rqstp)
621 struct svc_serv *serv = rqstp->rq_server;
626 /* now allocate needed pages. If we get a failure, sleep briefly */
627 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
628 WARN_ON_ONCE(pages >= RPCSVC_MAXPAGES);
629 if (pages >= RPCSVC_MAXPAGES)
630 /* use as many pages as possible */
631 pages = RPCSVC_MAXPAGES - 1;
632 for (i = 0; i < pages ; i++)
633 while (rqstp->rq_pages[i] == NULL) {
634 struct page *p = alloc_page(GFP_KERNEL);
636 set_current_state(TASK_INTERRUPTIBLE);
637 if (signalled() || kthread_should_stop()) {
638 set_current_state(TASK_RUNNING);
641 schedule_timeout(msecs_to_jiffies(500));
643 rqstp->rq_pages[i] = p;
645 rqstp->rq_page_end = &rqstp->rq_pages[i];
646 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
648 /* Make arg->head point to first page and arg->pages point to rest */
649 arg = &rqstp->rq_arg;
650 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
651 arg->head[0].iov_len = PAGE_SIZE;
652 arg->pages = rqstp->rq_pages + 1;
654 /* save at least one page for response */
655 arg->page_len = (pages-2)*PAGE_SIZE;
656 arg->len = (pages-1)*PAGE_SIZE;
657 arg->tail[0].iov_len = 0;
662 rqst_should_sleep(struct svc_rqst *rqstp)
664 struct svc_pool *pool = rqstp->rq_pool;
666 /* did someone call svc_wake_up? */
667 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
670 /* was a socket queued? */
671 if (!list_empty(&pool->sp_sockets))
674 /* are we shutting down? */
675 if (signalled() || kthread_should_stop())
678 /* are we freezing? */
679 if (freezing(current))
685 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
687 struct svc_xprt *xprt;
688 struct svc_pool *pool = rqstp->rq_pool;
691 /* rq_xprt should be clear on entry */
692 WARN_ON_ONCE(rqstp->rq_xprt);
694 /* Normally we will wait up to 5 seconds for any required
695 * cache information to be provided.
697 rqstp->rq_chandle.thread_wait = 5*HZ;
699 xprt = svc_xprt_dequeue(pool);
701 rqstp->rq_xprt = xprt;
703 /* As there is a shortage of threads and this request
704 * had to be queued, don't allow the thread to wait so
705 * long for cache updates.
707 rqstp->rq_chandle.thread_wait = 1*HZ;
708 clear_bit(SP_TASK_PENDING, &pool->sp_flags);
713 * We have to be able to interrupt this wait
714 * to bring down the daemons ...
716 set_current_state(TASK_INTERRUPTIBLE);
717 clear_bit(RQ_BUSY, &rqstp->rq_flags);
720 if (likely(rqst_should_sleep(rqstp)))
721 time_left = schedule_timeout(timeout);
723 __set_current_state(TASK_RUNNING);
727 spin_lock_bh(&rqstp->rq_lock);
728 set_bit(RQ_BUSY, &rqstp->rq_flags);
729 spin_unlock_bh(&rqstp->rq_lock);
731 xprt = rqstp->rq_xprt;
736 atomic_long_inc(&pool->sp_stats.threads_timedout);
738 if (signalled() || kthread_should_stop())
739 return ERR_PTR(-EINTR);
740 return ERR_PTR(-EAGAIN);
743 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
745 spin_lock_bh(&serv->sv_lock);
746 set_bit(XPT_TEMP, &newxpt->xpt_flags);
747 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
749 if (serv->sv_temptimer.function == NULL) {
750 /* setup timer to age temp transports */
751 setup_timer(&serv->sv_temptimer, svc_age_temp_xprts,
752 (unsigned long)serv);
753 mod_timer(&serv->sv_temptimer,
754 jiffies + svc_conn_age_period * HZ);
756 spin_unlock_bh(&serv->sv_lock);
757 svc_xprt_received(newxpt);
760 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
762 struct svc_serv *serv = rqstp->rq_server;
765 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
766 dprintk("svc_recv: found XPT_CLOSE\n");
767 svc_delete_xprt(xprt);
768 /* Leave XPT_BUSY set on the dead xprt: */
771 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
772 struct svc_xprt *newxpt;
774 * We know this module_get will succeed because the
775 * listener holds a reference too
777 __module_get(xprt->xpt_class->xcl_owner);
778 svc_check_conn_limits(xprt->xpt_server);
779 newxpt = xprt->xpt_ops->xpo_accept(xprt);
781 svc_add_new_temp_xprt(serv, newxpt);
783 module_put(xprt->xpt_class->xcl_owner);
785 /* XPT_DATA|XPT_DEFERRED case: */
786 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
787 rqstp, rqstp->rq_pool->sp_id, xprt,
788 atomic_read(&xprt->xpt_ref.refcount));
789 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
790 if (rqstp->rq_deferred)
791 len = svc_deferred_recv(rqstp);
793 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
794 dprintk("svc: got len=%d\n", len);
795 rqstp->rq_reserved = serv->sv_max_mesg;
796 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
798 /* clear XPT_BUSY: */
799 svc_xprt_received(xprt);
801 trace_svc_handle_xprt(xprt, len);
806 * Receive the next request on any transport. This code is carefully
807 * organised not to touch any cachelines in the shared svc_serv
808 * structure, only cachelines in the local svc_pool.
810 int svc_recv(struct svc_rqst *rqstp, long timeout)
812 struct svc_xprt *xprt = NULL;
813 struct svc_serv *serv = rqstp->rq_server;
816 dprintk("svc: server %p waiting for data (to = %ld)\n",
821 "svc_recv: service %p, transport not NULL!\n",
824 err = svc_alloc_arg(rqstp);
831 if (signalled() || kthread_should_stop())
834 xprt = svc_get_next_xprt(rqstp, timeout);
840 len = svc_handle_xprt(rqstp, xprt);
842 /* No data, incomplete (TCP) read, or accept() */
847 clear_bit(XPT_OLD, &xprt->xpt_flags);
849 if (xprt->xpt_ops->xpo_secure_port(rqstp))
850 set_bit(RQ_SECURE, &rqstp->rq_flags);
852 clear_bit(RQ_SECURE, &rqstp->rq_flags);
853 rqstp->rq_chandle.defer = svc_defer;
854 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
857 serv->sv_stats->netcnt++;
858 trace_svc_recv(rqstp, len);
861 rqstp->rq_res.len = 0;
862 svc_xprt_release(rqstp);
864 trace_svc_recv(rqstp, err);
867 EXPORT_SYMBOL_GPL(svc_recv);
872 void svc_drop(struct svc_rqst *rqstp)
874 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
875 svc_xprt_release(rqstp);
877 EXPORT_SYMBOL_GPL(svc_drop);
880 * Return reply to client.
882 int svc_send(struct svc_rqst *rqstp)
884 struct svc_xprt *xprt;
888 xprt = rqstp->rq_xprt;
892 /* release the receive skb before sending the reply */
893 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
895 /* calculate over-all length */
897 xb->len = xb->head[0].iov_len +
901 /* Grab mutex to serialize outgoing data. */
902 mutex_lock(&xprt->xpt_mutex);
903 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
904 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
907 len = xprt->xpt_ops->xpo_sendto(rqstp);
908 mutex_unlock(&xprt->xpt_mutex);
909 rpc_wake_up(&xprt->xpt_bc_pending);
910 svc_xprt_release(rqstp);
912 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
915 trace_svc_send(rqstp, len);
920 * Timer function to close old temporary transports, using
921 * a mark-and-sweep algorithm.
923 static void svc_age_temp_xprts(unsigned long closure)
925 struct svc_serv *serv = (struct svc_serv *)closure;
926 struct svc_xprt *xprt;
927 struct list_head *le, *next;
929 dprintk("svc_age_temp_xprts\n");
931 if (!spin_trylock_bh(&serv->sv_lock)) {
932 /* busy, try again 1 sec later */
933 dprintk("svc_age_temp_xprts: busy\n");
934 mod_timer(&serv->sv_temptimer, jiffies + HZ);
938 list_for_each_safe(le, next, &serv->sv_tempsocks) {
939 xprt = list_entry(le, struct svc_xprt, xpt_list);
941 /* First time through, just mark it OLD. Second time
942 * through, close it. */
943 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
945 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
946 test_bit(XPT_BUSY, &xprt->xpt_flags))
949 set_bit(XPT_CLOSE, &xprt->xpt_flags);
950 dprintk("queuing xprt %p for closing\n", xprt);
952 /* a thread will dequeue and close it soon */
953 svc_xprt_enqueue(xprt);
955 spin_unlock_bh(&serv->sv_lock);
957 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
960 /* Close temporary transports whose xpt_local matches server_addr immediately
961 * instead of waiting for them to be picked up by the timer.
963 * This is meant to be called from a notifier_block that runs when an ip
964 * address is deleted.
966 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
968 struct svc_xprt *xprt;
969 struct svc_sock *svsk;
971 struct list_head *le, *next;
972 LIST_HEAD(to_be_closed);
973 struct linger no_linger = {
978 spin_lock_bh(&serv->sv_lock);
979 list_for_each_safe(le, next, &serv->sv_tempsocks) {
980 xprt = list_entry(le, struct svc_xprt, xpt_list);
981 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
983 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
984 list_move(le, &to_be_closed);
987 spin_unlock_bh(&serv->sv_lock);
989 while (!list_empty(&to_be_closed)) {
990 le = to_be_closed.next;
992 xprt = list_entry(le, struct svc_xprt, xpt_list);
993 dprintk("svc_age_temp_xprts_now: closing %p\n", xprt);
994 svsk = container_of(xprt, struct svc_sock, sk_xprt);
995 sock = svsk->sk_sock;
996 kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER,
997 (char *)&no_linger, sizeof(no_linger));
998 svc_close_xprt(xprt);
1001 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1003 static void call_xpt_users(struct svc_xprt *xprt)
1005 struct svc_xpt_user *u;
1007 spin_lock(&xprt->xpt_lock);
1008 while (!list_empty(&xprt->xpt_users)) {
1009 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1013 spin_unlock(&xprt->xpt_lock);
1017 * Remove a dead transport
1019 static void svc_delete_xprt(struct svc_xprt *xprt)
1021 struct svc_serv *serv = xprt->xpt_server;
1022 struct svc_deferred_req *dr;
1024 /* Only do this once */
1025 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1028 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1029 xprt->xpt_ops->xpo_detach(xprt);
1031 spin_lock_bh(&serv->sv_lock);
1032 list_del_init(&xprt->xpt_list);
1033 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1034 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1036 spin_unlock_bh(&serv->sv_lock);
1038 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1041 call_xpt_users(xprt);
1045 void svc_close_xprt(struct svc_xprt *xprt)
1047 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1048 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1049 /* someone else will have to effect the close */
1052 * We expect svc_close_xprt() to work even when no threads are
1053 * running (e.g., while configuring the server before starting
1054 * any threads), so if the transport isn't busy, we delete
1057 svc_delete_xprt(xprt);
1059 EXPORT_SYMBOL_GPL(svc_close_xprt);
1061 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1063 struct svc_xprt *xprt;
1066 spin_lock(&serv->sv_lock);
1067 list_for_each_entry(xprt, xprt_list, xpt_list) {
1068 if (xprt->xpt_net != net)
1071 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1072 svc_xprt_enqueue(xprt);
1074 spin_unlock(&serv->sv_lock);
1078 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1080 struct svc_pool *pool;
1081 struct svc_xprt *xprt;
1082 struct svc_xprt *tmp;
1085 for (i = 0; i < serv->sv_nrpools; i++) {
1086 pool = &serv->sv_pools[i];
1088 spin_lock_bh(&pool->sp_lock);
1089 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1090 if (xprt->xpt_net != net)
1092 list_del_init(&xprt->xpt_ready);
1093 spin_unlock_bh(&pool->sp_lock);
1096 spin_unlock_bh(&pool->sp_lock);
1101 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1103 struct svc_xprt *xprt;
1105 while ((xprt = svc_dequeue_net(serv, net))) {
1106 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1107 svc_delete_xprt(xprt);
1112 * Server threads may still be running (especially in the case where the
1113 * service is still running in other network namespaces).
1115 * So we shut down sockets the same way we would on a running server, by
1116 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1117 * the close. In the case there are no such other threads,
1118 * threads running, svc_clean_up_xprts() does a simple version of a
1119 * server's main event loop, and in the case where there are other
1120 * threads, we may need to wait a little while and then check again to
1121 * see if they're done.
1123 void svc_close_net(struct svc_serv *serv, struct net *net)
1127 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1128 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1130 svc_clean_up_xprts(serv, net);
1136 * Handle defer and revisit of requests
1139 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1141 struct svc_deferred_req *dr =
1142 container_of(dreq, struct svc_deferred_req, handle);
1143 struct svc_xprt *xprt = dr->xprt;
1145 spin_lock(&xprt->xpt_lock);
1146 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1147 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1148 spin_unlock(&xprt->xpt_lock);
1149 dprintk("revisit canceled\n");
1154 dprintk("revisit queued\n");
1156 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1157 spin_unlock(&xprt->xpt_lock);
1158 svc_xprt_enqueue(xprt);
1163 * Save the request off for later processing. The request buffer looks
1166 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1168 * This code can only handle requests that consist of an xprt-header
1171 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1173 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1174 struct svc_deferred_req *dr;
1176 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1177 return NULL; /* if more than a page, give up FIXME */
1178 if (rqstp->rq_deferred) {
1179 dr = rqstp->rq_deferred;
1180 rqstp->rq_deferred = NULL;
1184 /* FIXME maybe discard if size too large */
1185 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1186 dr = kmalloc(size, GFP_KERNEL);
1190 dr->handle.owner = rqstp->rq_server;
1191 dr->prot = rqstp->rq_prot;
1192 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1193 dr->addrlen = rqstp->rq_addrlen;
1194 dr->daddr = rqstp->rq_daddr;
1195 dr->argslen = rqstp->rq_arg.len >> 2;
1196 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1198 /* back up head to the start of the buffer and copy */
1199 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1200 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1203 svc_xprt_get(rqstp->rq_xprt);
1204 dr->xprt = rqstp->rq_xprt;
1205 set_bit(RQ_DROPME, &rqstp->rq_flags);
1207 dr->handle.revisit = svc_revisit;
1212 * recv data from a deferred request into an active one
1214 static int svc_deferred_recv(struct svc_rqst *rqstp)
1216 struct svc_deferred_req *dr = rqstp->rq_deferred;
1218 /* setup iov_base past transport header */
1219 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1220 /* The iov_len does not include the transport header bytes */
1221 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1222 rqstp->rq_arg.page_len = 0;
1223 /* The rq_arg.len includes the transport header bytes */
1224 rqstp->rq_arg.len = dr->argslen<<2;
1225 rqstp->rq_prot = dr->prot;
1226 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1227 rqstp->rq_addrlen = dr->addrlen;
1228 /* Save off transport header len in case we get deferred again */
1229 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1230 rqstp->rq_daddr = dr->daddr;
1231 rqstp->rq_respages = rqstp->rq_pages;
1232 return (dr->argslen<<2) - dr->xprt_hlen;
1236 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1238 struct svc_deferred_req *dr = NULL;
1240 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1242 spin_lock(&xprt->xpt_lock);
1243 if (!list_empty(&xprt->xpt_deferred)) {
1244 dr = list_entry(xprt->xpt_deferred.next,
1245 struct svc_deferred_req,
1247 list_del_init(&dr->handle.recent);
1249 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1250 spin_unlock(&xprt->xpt_lock);
1255 * svc_find_xprt - find an RPC transport instance
1256 * @serv: pointer to svc_serv to search
1257 * @xcl_name: C string containing transport's class name
1258 * @net: owner net pointer
1259 * @af: Address family of transport's local address
1260 * @port: transport's IP port number
1262 * Return the transport instance pointer for the endpoint accepting
1263 * connections/peer traffic from the specified transport class,
1264 * address family and port.
1266 * Specifying 0 for the address family or port is effectively a
1267 * wild-card, and will result in matching the first transport in the
1268 * service's list that has a matching class name.
1270 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1271 struct net *net, const sa_family_t af,
1272 const unsigned short port)
1274 struct svc_xprt *xprt;
1275 struct svc_xprt *found = NULL;
1277 /* Sanity check the args */
1278 if (serv == NULL || xcl_name == NULL)
1281 spin_lock_bh(&serv->sv_lock);
1282 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1283 if (xprt->xpt_net != net)
1285 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1287 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1289 if (port != 0 && port != svc_xprt_local_port(xprt))
1295 spin_unlock_bh(&serv->sv_lock);
1298 EXPORT_SYMBOL_GPL(svc_find_xprt);
1300 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1301 char *pos, int remaining)
1305 len = snprintf(pos, remaining, "%s %u\n",
1306 xprt->xpt_class->xcl_name,
1307 svc_xprt_local_port(xprt));
1308 if (len >= remaining)
1309 return -ENAMETOOLONG;
1314 * svc_xprt_names - format a buffer with a list of transport names
1315 * @serv: pointer to an RPC service
1316 * @buf: pointer to a buffer to be filled in
1317 * @buflen: length of buffer to be filled in
1319 * Fills in @buf with a string containing a list of transport names,
1320 * each name terminated with '\n'.
1322 * Returns positive length of the filled-in string on success; otherwise
1323 * a negative errno value is returned if an error occurs.
1325 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1327 struct svc_xprt *xprt;
1331 /* Sanity check args */
1335 spin_lock_bh(&serv->sv_lock);
1339 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1340 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1352 spin_unlock_bh(&serv->sv_lock);
1355 EXPORT_SYMBOL_GPL(svc_xprt_names);
1358 /*----------------------------------------------------------------------------*/
1360 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1362 unsigned int pidx = (unsigned int)*pos;
1363 struct svc_serv *serv = m->private;
1365 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1368 return SEQ_START_TOKEN;
1369 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1372 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1374 struct svc_pool *pool = p;
1375 struct svc_serv *serv = m->private;
1377 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1379 if (p == SEQ_START_TOKEN) {
1380 pool = &serv->sv_pools[0];
1382 unsigned int pidx = (pool - &serv->sv_pools[0]);
1383 if (pidx < serv->sv_nrpools-1)
1384 pool = &serv->sv_pools[pidx+1];
1392 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1396 static int svc_pool_stats_show(struct seq_file *m, void *p)
1398 struct svc_pool *pool = p;
1400 if (p == SEQ_START_TOKEN) {
1401 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1405 seq_printf(m, "%u %lu %lu %lu %lu\n",
1407 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1408 pool->sp_stats.sockets_queued,
1409 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1410 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1415 static const struct seq_operations svc_pool_stats_seq_ops = {
1416 .start = svc_pool_stats_start,
1417 .next = svc_pool_stats_next,
1418 .stop = svc_pool_stats_stop,
1419 .show = svc_pool_stats_show,
1422 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1426 err = seq_open(file, &svc_pool_stats_seq_ops);
1428 ((struct seq_file *) file->private_data)->private = serv;
1431 EXPORT_SYMBOL(svc_pool_stats_open);
1433 /*----------------------------------------------------------------------------*/