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>
18 #include <trace/events/sunrpc.h>
20 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
22 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
23 static int svc_deferred_recv(struct svc_rqst *rqstp);
24 static struct cache_deferred_req *svc_defer(struct cache_req *req);
25 static void svc_age_temp_xprts(unsigned long closure);
26 static void svc_delete_xprt(struct svc_xprt *xprt);
27 static void svc_xprt_do_enqueue(struct svc_xprt *xprt);
29 /* apparently the "standard" is that clients close
30 * idle connections after 5 minutes, servers after
32 * http://www.connectathon.org/talks96/nfstcp.pdf
34 static int svc_conn_age_period = 6*60;
36 /* List of registered transport classes */
37 static DEFINE_SPINLOCK(svc_xprt_class_lock);
38 static LIST_HEAD(svc_xprt_class_list);
40 /* SMP locking strategy:
42 * svc_pool->sp_lock protects most of the fields of that pool.
43 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
44 * when both need to be taken (rare), svc_serv->sv_lock is first.
45 * BKL protects svc_serv->sv_nrthread.
46 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
47 * and the ->sk_info_authunix cache.
49 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
50 * enqueued multiply. During normal transport processing this bit
51 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
52 * Providers should not manipulate this bit directly.
54 * Some flags can be set to certain values at any time
55 * providing that certain rules are followed:
58 * - Can be set or cleared at any time.
59 * - After a set, svc_xprt_enqueue must be called to enqueue
60 * the transport for processing.
61 * - After a clear, the transport must be read/accepted.
62 * If this succeeds, it must be set again.
64 * - Can set at any time. It is never cleared.
66 * - Can only be set while XPT_BUSY is held which ensures
67 * that no other thread will be using the transport or will
68 * 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 xprt->xpt_ops->xpo_free(xprt);
143 void svc_xprt_put(struct svc_xprt *xprt)
145 kref_put(&xprt->xpt_ref, svc_xprt_free);
147 EXPORT_SYMBOL_GPL(svc_xprt_put);
150 * Called by transport drivers to initialize the transport independent
151 * portion of the transport instance.
153 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
154 struct svc_xprt *xprt, struct svc_serv *serv)
156 memset(xprt, 0, sizeof(*xprt));
157 xprt->xpt_class = xcl;
158 xprt->xpt_ops = xcl->xcl_ops;
159 kref_init(&xprt->xpt_ref);
160 xprt->xpt_server = serv;
161 INIT_LIST_HEAD(&xprt->xpt_list);
162 INIT_LIST_HEAD(&xprt->xpt_ready);
163 INIT_LIST_HEAD(&xprt->xpt_deferred);
164 INIT_LIST_HEAD(&xprt->xpt_users);
165 mutex_init(&xprt->xpt_mutex);
166 spin_lock_init(&xprt->xpt_lock);
167 set_bit(XPT_BUSY, &xprt->xpt_flags);
168 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
169 xprt->xpt_net = get_net(net);
171 EXPORT_SYMBOL_GPL(svc_xprt_init);
173 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
174 struct svc_serv *serv,
177 const unsigned short port,
180 struct sockaddr_in sin = {
181 .sin_family = AF_INET,
182 .sin_addr.s_addr = htonl(INADDR_ANY),
183 .sin_port = htons(port),
185 #if IS_ENABLED(CONFIG_IPV6)
186 struct sockaddr_in6 sin6 = {
187 .sin6_family = AF_INET6,
188 .sin6_addr = IN6ADDR_ANY_INIT,
189 .sin6_port = htons(port),
192 struct sockaddr *sap;
197 sap = (struct sockaddr *)&sin;
200 #if IS_ENABLED(CONFIG_IPV6)
202 sap = (struct sockaddr *)&sin6;
207 return ERR_PTR(-EAFNOSUPPORT);
210 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
214 * svc_xprt_received conditionally queues the transport for processing
215 * by another thread. The caller must hold the XPT_BUSY bit and must
216 * not thereafter touch transport data.
218 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
219 * insufficient) data.
221 static void svc_xprt_received(struct svc_xprt *xprt)
223 WARN_ON_ONCE(!test_bit(XPT_BUSY, &xprt->xpt_flags));
224 if (!test_bit(XPT_BUSY, &xprt->xpt_flags))
226 /* As soon as we clear busy, the xprt could be closed and
227 * 'put', so we need a reference to call svc_xprt_do_enqueue with:
230 smp_mb__before_atomic();
231 clear_bit(XPT_BUSY, &xprt->xpt_flags);
232 svc_xprt_do_enqueue(xprt);
236 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
238 clear_bit(XPT_TEMP, &new->xpt_flags);
239 spin_lock_bh(&serv->sv_lock);
240 list_add(&new->xpt_list, &serv->sv_permsocks);
241 spin_unlock_bh(&serv->sv_lock);
242 svc_xprt_received(new);
245 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
246 struct net *net, const int family,
247 const unsigned short port, int flags)
249 struct svc_xprt_class *xcl;
251 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
252 spin_lock(&svc_xprt_class_lock);
253 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
254 struct svc_xprt *newxprt;
255 unsigned short newport;
257 if (strcmp(xprt_name, xcl->xcl_name))
260 if (!try_module_get(xcl->xcl_owner))
263 spin_unlock(&svc_xprt_class_lock);
264 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
265 if (IS_ERR(newxprt)) {
266 module_put(xcl->xcl_owner);
267 return PTR_ERR(newxprt);
269 svc_add_new_perm_xprt(serv, newxprt);
270 newport = svc_xprt_local_port(newxprt);
274 spin_unlock(&svc_xprt_class_lock);
275 dprintk("svc: transport %s not found\n", xprt_name);
277 /* This errno is exposed to user space. Provide a reasonable
278 * perror msg for a bad transport. */
279 return -EPROTONOSUPPORT;
281 EXPORT_SYMBOL_GPL(svc_create_xprt);
284 * Copy the local and remote xprt addresses to the rqstp structure
286 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
288 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
289 rqstp->rq_addrlen = xprt->xpt_remotelen;
292 * Destination address in request is needed for binding the
293 * source address in RPC replies/callbacks later.
295 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
296 rqstp->rq_daddrlen = xprt->xpt_locallen;
298 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
301 * svc_print_addr - Format rq_addr field for printing
302 * @rqstp: svc_rqst struct containing address to print
303 * @buf: target buffer for formatted address
304 * @len: length of target buffer
307 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
309 return __svc_print_addr(svc_addr(rqstp), buf, len);
311 EXPORT_SYMBOL_GPL(svc_print_addr);
313 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
315 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
317 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
318 return xprt->xpt_ops->xpo_has_wspace(xprt);
322 static void svc_xprt_do_enqueue(struct svc_xprt *xprt)
324 struct svc_pool *pool;
325 struct svc_rqst *rqstp;
329 if (!svc_xprt_has_something_to_do(xprt))
332 /* Mark transport as busy. It will remain in this state until
333 * the provider calls svc_xprt_received. We update XPT_BUSY
334 * atomically because it also guards against trying to enqueue
335 * the transport twice.
337 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
338 /* Don't enqueue transport while already enqueued */
339 dprintk("svc: transport %p busy, not enqueued\n", xprt);
344 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
346 atomic_long_inc(&pool->sp_stats.packets);
349 /* find a thread for this xprt */
351 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
352 /* Do a lockless check first */
353 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
357 * Once the xprt has been queued, it can only be dequeued by
358 * the task that intends to service it. All we can do at that
359 * point is to try to wake this thread back up so that it can
363 spin_lock_bh(&rqstp->rq_lock);
364 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags)) {
365 /* already busy, move on... */
366 spin_unlock_bh(&rqstp->rq_lock);
370 /* this one will do */
371 rqstp->rq_xprt = xprt;
373 spin_unlock_bh(&rqstp->rq_lock);
377 atomic_long_inc(&pool->sp_stats.threads_woken);
378 wake_up_process(rqstp->rq_task);
385 * We didn't find an idle thread to use, so we need to queue the xprt.
386 * Do so and then search again. If we find one, we can't hook this one
387 * up to it directly but we can wake the thread up in the hopes that it
388 * will pick it up once it searches for a xprt to service.
392 dprintk("svc: transport %p put into queue\n", xprt);
393 spin_lock_bh(&pool->sp_lock);
394 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
395 pool->sp_stats.sockets_queued++;
396 spin_unlock_bh(&pool->sp_lock);
403 * Queue up a transport with data pending. If there are idle nfsd
404 * processes, wake 'em up.
407 void svc_xprt_enqueue(struct svc_xprt *xprt)
409 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
411 svc_xprt_do_enqueue(xprt);
413 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
416 * Dequeue the first transport, if there is one.
418 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
420 struct svc_xprt *xprt = NULL;
422 if (list_empty(&pool->sp_sockets))
425 spin_lock_bh(&pool->sp_lock);
426 if (likely(!list_empty(&pool->sp_sockets))) {
427 xprt = list_first_entry(&pool->sp_sockets,
428 struct svc_xprt, xpt_ready);
429 list_del_init(&xprt->xpt_ready);
432 dprintk("svc: transport %p dequeued, inuse=%d\n",
433 xprt, atomic_read(&xprt->xpt_ref.refcount));
435 spin_unlock_bh(&pool->sp_lock);
441 * svc_reserve - change the space reserved for the reply to a request.
442 * @rqstp: The request in question
443 * @space: new max space to reserve
445 * Each request reserves some space on the output queue of the transport
446 * to make sure the reply fits. This function reduces that reserved
447 * space to be the amount of space used already, plus @space.
450 void svc_reserve(struct svc_rqst *rqstp, int space)
452 space += rqstp->rq_res.head[0].iov_len;
454 if (space < rqstp->rq_reserved) {
455 struct svc_xprt *xprt = rqstp->rq_xprt;
456 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
457 rqstp->rq_reserved = space;
459 if (xprt->xpt_ops->xpo_adjust_wspace)
460 xprt->xpt_ops->xpo_adjust_wspace(xprt);
461 svc_xprt_enqueue(xprt);
464 EXPORT_SYMBOL_GPL(svc_reserve);
466 static void svc_xprt_release(struct svc_rqst *rqstp)
468 struct svc_xprt *xprt = rqstp->rq_xprt;
470 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
472 kfree(rqstp->rq_deferred);
473 rqstp->rq_deferred = NULL;
475 svc_free_res_pages(rqstp);
476 rqstp->rq_res.page_len = 0;
477 rqstp->rq_res.page_base = 0;
479 /* Reset response buffer and release
481 * But first, check that enough space was reserved
482 * for the reply, otherwise we have a bug!
484 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
485 printk(KERN_ERR "RPC request reserved %d but used %d\n",
489 rqstp->rq_res.head[0].iov_len = 0;
490 svc_reserve(rqstp, 0);
491 rqstp->rq_xprt = NULL;
497 * Some svc_serv's will have occasional work to do, even when a xprt is not
498 * waiting to be serviced. This function is there to "kick" a task in one of
499 * those services so that it can wake up and do that work. Note that we only
500 * bother with pool 0 as we don't need to wake up more than one thread for
503 void svc_wake_up(struct svc_serv *serv)
505 struct svc_rqst *rqstp;
506 struct svc_pool *pool;
508 pool = &serv->sv_pools[0];
511 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
512 /* skip any that aren't queued */
513 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
516 dprintk("svc: daemon %p woken up.\n", rqstp);
517 wake_up_process(rqstp->rq_task);
522 /* No free entries available */
523 set_bit(SP_TASK_PENDING, &pool->sp_flags);
526 EXPORT_SYMBOL_GPL(svc_wake_up);
528 int svc_port_is_privileged(struct sockaddr *sin)
530 switch (sin->sa_family) {
532 return ntohs(((struct sockaddr_in *)sin)->sin_port)
535 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
543 * Make sure that we don't have too many active connections. If we have,
544 * something must be dropped. It's not clear what will happen if we allow
545 * "too many" connections, but when dealing with network-facing software,
546 * we have to code defensively. Here we do that by imposing hard limits.
548 * There's no point in trying to do random drop here for DoS
549 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
550 * attacker can easily beat that.
552 * The only somewhat efficient mechanism would be if drop old
553 * connections from the same IP first. But right now we don't even
554 * record the client IP in svc_sock.
556 * single-threaded services that expect a lot of clients will probably
557 * need to set sv_maxconn to override the default value which is based
558 * on the number of threads
560 static void svc_check_conn_limits(struct svc_serv *serv)
562 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
563 (serv->sv_nrthreads+3) * 20;
565 if (serv->sv_tmpcnt > limit) {
566 struct svc_xprt *xprt = NULL;
567 spin_lock_bh(&serv->sv_lock);
568 if (!list_empty(&serv->sv_tempsocks)) {
569 /* Try to help the admin */
570 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
571 serv->sv_name, serv->sv_maxconn ?
572 "max number of connections" :
573 "number of threads");
575 * Always select the oldest connection. It's not fair,
578 xprt = list_entry(serv->sv_tempsocks.prev,
581 set_bit(XPT_CLOSE, &xprt->xpt_flags);
584 spin_unlock_bh(&serv->sv_lock);
587 svc_xprt_enqueue(xprt);
593 static int svc_alloc_arg(struct svc_rqst *rqstp)
595 struct svc_serv *serv = rqstp->rq_server;
600 /* now allocate needed pages. If we get a failure, sleep briefly */
601 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
602 WARN_ON_ONCE(pages >= RPCSVC_MAXPAGES);
603 if (pages >= RPCSVC_MAXPAGES)
604 /* use as many pages as possible */
605 pages = RPCSVC_MAXPAGES - 1;
606 for (i = 0; i < pages ; i++)
607 while (rqstp->rq_pages[i] == NULL) {
608 struct page *p = alloc_page(GFP_KERNEL);
610 set_current_state(TASK_INTERRUPTIBLE);
611 if (signalled() || kthread_should_stop()) {
612 set_current_state(TASK_RUNNING);
615 schedule_timeout(msecs_to_jiffies(500));
617 rqstp->rq_pages[i] = p;
619 rqstp->rq_page_end = &rqstp->rq_pages[i];
620 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
622 /* Make arg->head point to first page and arg->pages point to rest */
623 arg = &rqstp->rq_arg;
624 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
625 arg->head[0].iov_len = PAGE_SIZE;
626 arg->pages = rqstp->rq_pages + 1;
628 /* save at least one page for response */
629 arg->page_len = (pages-2)*PAGE_SIZE;
630 arg->len = (pages-1)*PAGE_SIZE;
631 arg->tail[0].iov_len = 0;
636 rqst_should_sleep(struct svc_rqst *rqstp)
638 struct svc_pool *pool = rqstp->rq_pool;
640 /* did someone call svc_wake_up? */
641 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
644 /* was a socket queued? */
645 if (!list_empty(&pool->sp_sockets))
648 /* are we shutting down? */
649 if (signalled() || kthread_should_stop())
652 /* are we freezing? */
653 if (freezing(current))
659 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
661 struct svc_xprt *xprt;
662 struct svc_pool *pool = rqstp->rq_pool;
665 /* rq_xprt should be clear on entry */
666 WARN_ON_ONCE(rqstp->rq_xprt);
668 /* Normally we will wait up to 5 seconds for any required
669 * cache information to be provided.
671 rqstp->rq_chandle.thread_wait = 5*HZ;
673 xprt = svc_xprt_dequeue(pool);
675 rqstp->rq_xprt = xprt;
677 /* As there is a shortage of threads and this request
678 * had to be queued, don't allow the thread to wait so
679 * long for cache updates.
681 rqstp->rq_chandle.thread_wait = 1*HZ;
682 clear_bit(SP_TASK_PENDING, &pool->sp_flags);
687 * We have to be able to interrupt this wait
688 * to bring down the daemons ...
690 set_current_state(TASK_INTERRUPTIBLE);
691 clear_bit(RQ_BUSY, &rqstp->rq_flags);
694 if (likely(rqst_should_sleep(rqstp)))
695 time_left = schedule_timeout(timeout);
697 __set_current_state(TASK_RUNNING);
701 spin_lock_bh(&rqstp->rq_lock);
702 set_bit(RQ_BUSY, &rqstp->rq_flags);
703 spin_unlock_bh(&rqstp->rq_lock);
705 xprt = rqstp->rq_xprt;
710 atomic_long_inc(&pool->sp_stats.threads_timedout);
712 if (signalled() || kthread_should_stop())
713 return ERR_PTR(-EINTR);
714 return ERR_PTR(-EAGAIN);
717 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
719 spin_lock_bh(&serv->sv_lock);
720 set_bit(XPT_TEMP, &newxpt->xpt_flags);
721 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
723 if (serv->sv_temptimer.function == NULL) {
724 /* setup timer to age temp transports */
725 setup_timer(&serv->sv_temptimer, svc_age_temp_xprts,
726 (unsigned long)serv);
727 mod_timer(&serv->sv_temptimer,
728 jiffies + svc_conn_age_period * HZ);
730 spin_unlock_bh(&serv->sv_lock);
731 svc_xprt_received(newxpt);
734 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
736 struct svc_serv *serv = rqstp->rq_server;
739 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
740 dprintk("svc_recv: found XPT_CLOSE\n");
741 svc_delete_xprt(xprt);
742 /* Leave XPT_BUSY set on the dead xprt: */
745 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
746 struct svc_xprt *newxpt;
748 * We know this module_get will succeed because the
749 * listener holds a reference too
751 __module_get(xprt->xpt_class->xcl_owner);
752 svc_check_conn_limits(xprt->xpt_server);
753 newxpt = xprt->xpt_ops->xpo_accept(xprt);
755 svc_add_new_temp_xprt(serv, newxpt);
757 module_put(xprt->xpt_class->xcl_owner);
759 /* XPT_DATA|XPT_DEFERRED case: */
760 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
761 rqstp, rqstp->rq_pool->sp_id, xprt,
762 atomic_read(&xprt->xpt_ref.refcount));
763 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
764 if (rqstp->rq_deferred)
765 len = svc_deferred_recv(rqstp);
767 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
768 dprintk("svc: got len=%d\n", len);
769 rqstp->rq_reserved = serv->sv_max_mesg;
770 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
772 /* clear XPT_BUSY: */
773 svc_xprt_received(xprt);
778 * Receive the next request on any transport. This code is carefully
779 * organised not to touch any cachelines in the shared svc_serv
780 * structure, only cachelines in the local svc_pool.
782 int svc_recv(struct svc_rqst *rqstp, long timeout)
784 struct svc_xprt *xprt = NULL;
785 struct svc_serv *serv = rqstp->rq_server;
788 dprintk("svc: server %p waiting for data (to = %ld)\n",
793 "svc_recv: service %p, transport not NULL!\n",
796 err = svc_alloc_arg(rqstp);
803 if (signalled() || kthread_should_stop())
806 xprt = svc_get_next_xprt(rqstp, timeout);
812 len = svc_handle_xprt(rqstp, xprt);
814 /* No data, incomplete (TCP) read, or accept() */
819 clear_bit(XPT_OLD, &xprt->xpt_flags);
821 if (xprt->xpt_ops->xpo_secure_port(rqstp))
822 set_bit(RQ_SECURE, &rqstp->rq_flags);
824 clear_bit(RQ_SECURE, &rqstp->rq_flags);
825 rqstp->rq_chandle.defer = svc_defer;
826 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
829 serv->sv_stats->netcnt++;
830 trace_svc_recv(rqstp, len);
833 rqstp->rq_res.len = 0;
834 svc_xprt_release(rqstp);
836 trace_svc_recv(rqstp, err);
839 EXPORT_SYMBOL_GPL(svc_recv);
844 void svc_drop(struct svc_rqst *rqstp)
846 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
847 svc_xprt_release(rqstp);
849 EXPORT_SYMBOL_GPL(svc_drop);
852 * Return reply to client.
854 int svc_send(struct svc_rqst *rqstp)
856 struct svc_xprt *xprt;
860 xprt = rqstp->rq_xprt;
864 /* release the receive skb before sending the reply */
865 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
867 /* calculate over-all length */
869 xb->len = xb->head[0].iov_len +
873 /* Grab mutex to serialize outgoing data. */
874 mutex_lock(&xprt->xpt_mutex);
875 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
876 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
879 len = xprt->xpt_ops->xpo_sendto(rqstp);
880 mutex_unlock(&xprt->xpt_mutex);
881 rpc_wake_up(&xprt->xpt_bc_pending);
882 svc_xprt_release(rqstp);
884 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
887 trace_svc_send(rqstp, len);
892 * Timer function to close old temporary transports, using
893 * a mark-and-sweep algorithm.
895 static void svc_age_temp_xprts(unsigned long closure)
897 struct svc_serv *serv = (struct svc_serv *)closure;
898 struct svc_xprt *xprt;
899 struct list_head *le, *next;
901 dprintk("svc_age_temp_xprts\n");
903 if (!spin_trylock_bh(&serv->sv_lock)) {
904 /* busy, try again 1 sec later */
905 dprintk("svc_age_temp_xprts: busy\n");
906 mod_timer(&serv->sv_temptimer, jiffies + HZ);
910 list_for_each_safe(le, next, &serv->sv_tempsocks) {
911 xprt = list_entry(le, struct svc_xprt, xpt_list);
913 /* First time through, just mark it OLD. Second time
914 * through, close it. */
915 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
917 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
918 test_bit(XPT_BUSY, &xprt->xpt_flags))
921 set_bit(XPT_CLOSE, &xprt->xpt_flags);
922 dprintk("queuing xprt %p for closing\n", xprt);
924 /* a thread will dequeue and close it soon */
925 svc_xprt_enqueue(xprt);
927 spin_unlock_bh(&serv->sv_lock);
929 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
932 static void call_xpt_users(struct svc_xprt *xprt)
934 struct svc_xpt_user *u;
936 spin_lock(&xprt->xpt_lock);
937 while (!list_empty(&xprt->xpt_users)) {
938 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
942 spin_unlock(&xprt->xpt_lock);
946 * Remove a dead transport
948 static void svc_delete_xprt(struct svc_xprt *xprt)
950 struct svc_serv *serv = xprt->xpt_server;
951 struct svc_deferred_req *dr;
953 /* Only do this once */
954 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
957 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
958 xprt->xpt_ops->xpo_detach(xprt);
960 spin_lock_bh(&serv->sv_lock);
961 list_del_init(&xprt->xpt_list);
962 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
963 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
965 spin_unlock_bh(&serv->sv_lock);
967 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
970 call_xpt_users(xprt);
974 void svc_close_xprt(struct svc_xprt *xprt)
976 set_bit(XPT_CLOSE, &xprt->xpt_flags);
977 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
978 /* someone else will have to effect the close */
981 * We expect svc_close_xprt() to work even when no threads are
982 * running (e.g., while configuring the server before starting
983 * any threads), so if the transport isn't busy, we delete
986 svc_delete_xprt(xprt);
988 EXPORT_SYMBOL_GPL(svc_close_xprt);
990 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
992 struct svc_xprt *xprt;
995 spin_lock(&serv->sv_lock);
996 list_for_each_entry(xprt, xprt_list, xpt_list) {
997 if (xprt->xpt_net != net)
1000 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1001 svc_xprt_enqueue(xprt);
1003 spin_unlock(&serv->sv_lock);
1007 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1009 struct svc_pool *pool;
1010 struct svc_xprt *xprt;
1011 struct svc_xprt *tmp;
1014 for (i = 0; i < serv->sv_nrpools; i++) {
1015 pool = &serv->sv_pools[i];
1017 spin_lock_bh(&pool->sp_lock);
1018 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1019 if (xprt->xpt_net != net)
1021 list_del_init(&xprt->xpt_ready);
1022 spin_unlock_bh(&pool->sp_lock);
1025 spin_unlock_bh(&pool->sp_lock);
1030 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1032 struct svc_xprt *xprt;
1034 while ((xprt = svc_dequeue_net(serv, net))) {
1035 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1036 svc_delete_xprt(xprt);
1041 * Server threads may still be running (especially in the case where the
1042 * service is still running in other network namespaces).
1044 * So we shut down sockets the same way we would on a running server, by
1045 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1046 * the close. In the case there are no such other threads,
1047 * threads running, svc_clean_up_xprts() does a simple version of a
1048 * server's main event loop, and in the case where there are other
1049 * threads, we may need to wait a little while and then check again to
1050 * see if they're done.
1052 void svc_close_net(struct svc_serv *serv, struct net *net)
1056 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1057 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1059 svc_clean_up_xprts(serv, net);
1065 * Handle defer and revisit of requests
1068 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1070 struct svc_deferred_req *dr =
1071 container_of(dreq, struct svc_deferred_req, handle);
1072 struct svc_xprt *xprt = dr->xprt;
1074 spin_lock(&xprt->xpt_lock);
1075 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1076 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1077 spin_unlock(&xprt->xpt_lock);
1078 dprintk("revisit canceled\n");
1083 dprintk("revisit queued\n");
1085 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1086 spin_unlock(&xprt->xpt_lock);
1087 svc_xprt_enqueue(xprt);
1092 * Save the request off for later processing. The request buffer looks
1095 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1097 * This code can only handle requests that consist of an xprt-header
1100 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1102 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1103 struct svc_deferred_req *dr;
1105 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1106 return NULL; /* if more than a page, give up FIXME */
1107 if (rqstp->rq_deferred) {
1108 dr = rqstp->rq_deferred;
1109 rqstp->rq_deferred = NULL;
1113 /* FIXME maybe discard if size too large */
1114 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1115 dr = kmalloc(size, GFP_KERNEL);
1119 dr->handle.owner = rqstp->rq_server;
1120 dr->prot = rqstp->rq_prot;
1121 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1122 dr->addrlen = rqstp->rq_addrlen;
1123 dr->daddr = rqstp->rq_daddr;
1124 dr->argslen = rqstp->rq_arg.len >> 2;
1125 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1127 /* back up head to the start of the buffer and copy */
1128 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1129 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1132 svc_xprt_get(rqstp->rq_xprt);
1133 dr->xprt = rqstp->rq_xprt;
1134 set_bit(RQ_DROPME, &rqstp->rq_flags);
1136 dr->handle.revisit = svc_revisit;
1141 * recv data from a deferred request into an active one
1143 static int svc_deferred_recv(struct svc_rqst *rqstp)
1145 struct svc_deferred_req *dr = rqstp->rq_deferred;
1147 /* setup iov_base past transport header */
1148 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1149 /* The iov_len does not include the transport header bytes */
1150 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1151 rqstp->rq_arg.page_len = 0;
1152 /* The rq_arg.len includes the transport header bytes */
1153 rqstp->rq_arg.len = dr->argslen<<2;
1154 rqstp->rq_prot = dr->prot;
1155 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1156 rqstp->rq_addrlen = dr->addrlen;
1157 /* Save off transport header len in case we get deferred again */
1158 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1159 rqstp->rq_daddr = dr->daddr;
1160 rqstp->rq_respages = rqstp->rq_pages;
1161 return (dr->argslen<<2) - dr->xprt_hlen;
1165 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1167 struct svc_deferred_req *dr = NULL;
1169 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1171 spin_lock(&xprt->xpt_lock);
1172 if (!list_empty(&xprt->xpt_deferred)) {
1173 dr = list_entry(xprt->xpt_deferred.next,
1174 struct svc_deferred_req,
1176 list_del_init(&dr->handle.recent);
1178 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1179 spin_unlock(&xprt->xpt_lock);
1184 * svc_find_xprt - find an RPC transport instance
1185 * @serv: pointer to svc_serv to search
1186 * @xcl_name: C string containing transport's class name
1187 * @net: owner net pointer
1188 * @af: Address family of transport's local address
1189 * @port: transport's IP port number
1191 * Return the transport instance pointer for the endpoint accepting
1192 * connections/peer traffic from the specified transport class,
1193 * address family and port.
1195 * Specifying 0 for the address family or port is effectively a
1196 * wild-card, and will result in matching the first transport in the
1197 * service's list that has a matching class name.
1199 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1200 struct net *net, const sa_family_t af,
1201 const unsigned short port)
1203 struct svc_xprt *xprt;
1204 struct svc_xprt *found = NULL;
1206 /* Sanity check the args */
1207 if (serv == NULL || xcl_name == NULL)
1210 spin_lock_bh(&serv->sv_lock);
1211 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1212 if (xprt->xpt_net != net)
1214 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1216 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1218 if (port != 0 && port != svc_xprt_local_port(xprt))
1224 spin_unlock_bh(&serv->sv_lock);
1227 EXPORT_SYMBOL_GPL(svc_find_xprt);
1229 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1230 char *pos, int remaining)
1234 len = snprintf(pos, remaining, "%s %u\n",
1235 xprt->xpt_class->xcl_name,
1236 svc_xprt_local_port(xprt));
1237 if (len >= remaining)
1238 return -ENAMETOOLONG;
1243 * svc_xprt_names - format a buffer with a list of transport names
1244 * @serv: pointer to an RPC service
1245 * @buf: pointer to a buffer to be filled in
1246 * @buflen: length of buffer to be filled in
1248 * Fills in @buf with a string containing a list of transport names,
1249 * each name terminated with '\n'.
1251 * Returns positive length of the filled-in string on success; otherwise
1252 * a negative errno value is returned if an error occurs.
1254 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1256 struct svc_xprt *xprt;
1260 /* Sanity check args */
1264 spin_lock_bh(&serv->sv_lock);
1268 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1269 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1281 spin_unlock_bh(&serv->sv_lock);
1284 EXPORT_SYMBOL_GPL(svc_xprt_names);
1287 /*----------------------------------------------------------------------------*/
1289 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1291 unsigned int pidx = (unsigned int)*pos;
1292 struct svc_serv *serv = m->private;
1294 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1297 return SEQ_START_TOKEN;
1298 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1301 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1303 struct svc_pool *pool = p;
1304 struct svc_serv *serv = m->private;
1306 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1308 if (p == SEQ_START_TOKEN) {
1309 pool = &serv->sv_pools[0];
1311 unsigned int pidx = (pool - &serv->sv_pools[0]);
1312 if (pidx < serv->sv_nrpools-1)
1313 pool = &serv->sv_pools[pidx+1];
1321 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1325 static int svc_pool_stats_show(struct seq_file *m, void *p)
1327 struct svc_pool *pool = p;
1329 if (p == SEQ_START_TOKEN) {
1330 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1334 seq_printf(m, "%u %lu %lu %lu %lu\n",
1336 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1337 pool->sp_stats.sockets_queued,
1338 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1339 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1344 static const struct seq_operations svc_pool_stats_seq_ops = {
1345 .start = svc_pool_stats_start,
1346 .next = svc_pool_stats_next,
1347 .stop = svc_pool_stats_stop,
1348 .show = svc_pool_stats_show,
1351 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1355 err = seq_open(file, &svc_pool_stats_seq_ops);
1357 ((struct seq_file *) file->private_data)->private = serv;
1360 EXPORT_SYMBOL(svc_pool_stats_open);
1362 /*----------------------------------------------------------------------------*/