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 unsigned int svc_rpc_per_connection_limit __read_mostly;
25 module_param(svc_rpc_per_connection_limit, uint, 0644);
28 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
29 static int svc_deferred_recv(struct svc_rqst *rqstp);
30 static struct cache_deferred_req *svc_defer(struct cache_req *req);
31 static void svc_age_temp_xprts(unsigned long closure);
32 static void svc_delete_xprt(struct svc_xprt *xprt);
34 /* apparently the "standard" is that clients close
35 * idle connections after 5 minutes, servers after
37 * http://www.connectathon.org/talks96/nfstcp.pdf
39 static int svc_conn_age_period = 6*60;
41 /* List of registered transport classes */
42 static DEFINE_SPINLOCK(svc_xprt_class_lock);
43 static LIST_HEAD(svc_xprt_class_list);
45 /* SMP locking strategy:
47 * svc_pool->sp_lock protects most of the fields of that pool.
48 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
49 * when both need to be taken (rare), svc_serv->sv_lock is first.
50 * The "service mutex" protects svc_serv->sv_nrthread.
51 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
52 * and the ->sk_info_authunix cache.
54 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
55 * enqueued multiply. During normal transport processing this bit
56 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
57 * Providers should not manipulate this bit directly.
59 * Some flags can be set to certain values at any time
60 * providing that certain rules are followed:
63 * - Can be set or cleared at any time.
64 * - After a set, svc_xprt_enqueue must be called to enqueue
65 * the transport for processing.
66 * - After a clear, the transport must be read/accepted.
67 * If this succeeds, it must be set again.
69 * - Can set at any time. It is never cleared.
71 * - Can only be set while XPT_BUSY is held which ensures
72 * that no other thread will be using the transport or will
73 * try to set XPT_DEAD.
75 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
77 struct svc_xprt_class *cl;
80 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
82 INIT_LIST_HEAD(&xcl->xcl_list);
83 spin_lock(&svc_xprt_class_lock);
84 /* Make sure there isn't already a class with the same name */
85 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
86 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
89 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
92 spin_unlock(&svc_xprt_class_lock);
95 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
97 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
99 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
100 spin_lock(&svc_xprt_class_lock);
101 list_del_init(&xcl->xcl_list);
102 spin_unlock(&svc_xprt_class_lock);
104 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
107 * svc_print_xprts - Format the transport list for printing
108 * @buf: target buffer for formatted address
109 * @maxlen: length of target buffer
111 * Fills in @buf with a string containing a list of transport names, each name
112 * terminated with '\n'. If the buffer is too small, some entries may be
113 * missing, but it is guaranteed that all lines in the output buffer are
116 * Returns positive length of the filled-in string.
118 int svc_print_xprts(char *buf, int maxlen)
120 struct svc_xprt_class *xcl;
125 spin_lock(&svc_xprt_class_lock);
126 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
129 slen = snprintf(tmpstr, sizeof(tmpstr), "%s %d\n",
130 xcl->xcl_name, xcl->xcl_max_payload);
131 if (slen >= sizeof(tmpstr) || len + slen >= maxlen)
136 spin_unlock(&svc_xprt_class_lock);
141 static void svc_xprt_free(struct kref *kref)
143 struct svc_xprt *xprt =
144 container_of(kref, struct svc_xprt, xpt_ref);
145 struct module *owner = xprt->xpt_class->xcl_owner;
146 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
147 svcauth_unix_info_release(xprt);
148 put_net(xprt->xpt_net);
149 /* See comment on corresponding get in xs_setup_bc_tcp(): */
150 if (xprt->xpt_bc_xprt)
151 xprt_put(xprt->xpt_bc_xprt);
152 if (xprt->xpt_bc_xps)
153 xprt_switch_put(xprt->xpt_bc_xps);
154 xprt->xpt_ops->xpo_free(xprt);
158 void svc_xprt_put(struct svc_xprt *xprt)
160 kref_put(&xprt->xpt_ref, svc_xprt_free);
162 EXPORT_SYMBOL_GPL(svc_xprt_put);
165 * Called by transport drivers to initialize the transport independent
166 * portion of the transport instance.
168 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
169 struct svc_xprt *xprt, struct svc_serv *serv)
171 memset(xprt, 0, sizeof(*xprt));
172 xprt->xpt_class = xcl;
173 xprt->xpt_ops = xcl->xcl_ops;
174 kref_init(&xprt->xpt_ref);
175 xprt->xpt_server = serv;
176 INIT_LIST_HEAD(&xprt->xpt_list);
177 INIT_LIST_HEAD(&xprt->xpt_ready);
178 INIT_LIST_HEAD(&xprt->xpt_deferred);
179 INIT_LIST_HEAD(&xprt->xpt_users);
180 mutex_init(&xprt->xpt_mutex);
181 spin_lock_init(&xprt->xpt_lock);
182 set_bit(XPT_BUSY, &xprt->xpt_flags);
183 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
184 xprt->xpt_net = get_net(net);
186 EXPORT_SYMBOL_GPL(svc_xprt_init);
188 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
189 struct svc_serv *serv,
192 const unsigned short port,
195 struct sockaddr_in sin = {
196 .sin_family = AF_INET,
197 .sin_addr.s_addr = htonl(INADDR_ANY),
198 .sin_port = htons(port),
200 #if IS_ENABLED(CONFIG_IPV6)
201 struct sockaddr_in6 sin6 = {
202 .sin6_family = AF_INET6,
203 .sin6_addr = IN6ADDR_ANY_INIT,
204 .sin6_port = htons(port),
207 struct sockaddr *sap;
212 sap = (struct sockaddr *)&sin;
215 #if IS_ENABLED(CONFIG_IPV6)
217 sap = (struct sockaddr *)&sin6;
222 return ERR_PTR(-EAFNOSUPPORT);
225 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
229 * svc_xprt_received conditionally queues the transport for processing
230 * by another thread. The caller must hold the XPT_BUSY bit and must
231 * not thereafter touch transport data.
233 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
234 * insufficient) data.
236 static void svc_xprt_received(struct svc_xprt *xprt)
238 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
239 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
243 /* As soon as we clear busy, the xprt could be closed and
244 * 'put', so we need a reference to call svc_enqueue_xprt with:
247 smp_mb__before_atomic();
248 clear_bit(XPT_BUSY, &xprt->xpt_flags);
249 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
253 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
255 clear_bit(XPT_TEMP, &new->xpt_flags);
256 spin_lock_bh(&serv->sv_lock);
257 list_add(&new->xpt_list, &serv->sv_permsocks);
258 spin_unlock_bh(&serv->sv_lock);
259 svc_xprt_received(new);
262 int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
263 struct net *net, const int family,
264 const unsigned short port, int flags)
266 struct svc_xprt_class *xcl;
268 spin_lock(&svc_xprt_class_lock);
269 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
270 struct svc_xprt *newxprt;
271 unsigned short newport;
273 if (strcmp(xprt_name, xcl->xcl_name))
276 if (!try_module_get(xcl->xcl_owner))
279 spin_unlock(&svc_xprt_class_lock);
280 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
281 if (IS_ERR(newxprt)) {
282 module_put(xcl->xcl_owner);
283 return PTR_ERR(newxprt);
285 svc_add_new_perm_xprt(serv, newxprt);
286 newport = svc_xprt_local_port(newxprt);
290 spin_unlock(&svc_xprt_class_lock);
291 /* This errno is exposed to user space. Provide a reasonable
292 * perror msg for a bad transport. */
293 return -EPROTONOSUPPORT;
296 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
297 struct net *net, const int family,
298 const unsigned short port, int flags)
302 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
303 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
304 if (err == -EPROTONOSUPPORT) {
305 request_module("svc%s", xprt_name);
306 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
309 dprintk("svc: transport %s not found, err %d\n",
313 EXPORT_SYMBOL_GPL(svc_create_xprt);
316 * Copy the local and remote xprt addresses to the rqstp structure
318 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
320 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
321 rqstp->rq_addrlen = xprt->xpt_remotelen;
324 * Destination address in request is needed for binding the
325 * source address in RPC replies/callbacks later.
327 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
328 rqstp->rq_daddrlen = xprt->xpt_locallen;
330 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
333 * svc_print_addr - Format rq_addr field for printing
334 * @rqstp: svc_rqst struct containing address to print
335 * @buf: target buffer for formatted address
336 * @len: length of target buffer
339 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
341 return __svc_print_addr(svc_addr(rqstp), buf, len);
343 EXPORT_SYMBOL_GPL(svc_print_addr);
345 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
347 unsigned int limit = svc_rpc_per_connection_limit;
348 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
350 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
353 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
355 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
356 if (!svc_xprt_slots_in_range(xprt))
358 atomic_inc(&xprt->xpt_nr_rqsts);
359 set_bit(RQ_DATA, &rqstp->rq_flags);
364 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
366 struct svc_xprt *xprt = rqstp->rq_xprt;
367 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
368 atomic_dec(&xprt->xpt_nr_rqsts);
369 svc_xprt_enqueue(xprt);
373 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
375 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
377 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED))) {
378 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
379 svc_xprt_slots_in_range(xprt))
381 trace_svc_xprt_no_write_space(xprt);
387 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
389 struct svc_pool *pool;
390 struct svc_rqst *rqstp = NULL;
394 if (!svc_xprt_has_something_to_do(xprt))
397 /* Mark transport as busy. It will remain in this state until
398 * the provider calls svc_xprt_received. We update XPT_BUSY
399 * atomically because it also guards against trying to enqueue
400 * the transport twice.
402 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
403 /* Don't enqueue transport while already enqueued */
404 dprintk("svc: transport %p busy, not enqueued\n", xprt);
409 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
411 atomic_long_inc(&pool->sp_stats.packets);
414 /* find a thread for this xprt */
416 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
417 /* Do a lockless check first */
418 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
422 * Once the xprt has been queued, it can only be dequeued by
423 * the task that intends to service it. All we can do at that
424 * point is to try to wake this thread back up so that it can
428 spin_lock_bh(&rqstp->rq_lock);
429 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags)) {
430 /* already busy, move on... */
431 spin_unlock_bh(&rqstp->rq_lock);
435 /* this one will do */
436 rqstp->rq_xprt = xprt;
438 spin_unlock_bh(&rqstp->rq_lock);
442 atomic_long_inc(&pool->sp_stats.threads_woken);
443 wake_up_process(rqstp->rq_task);
450 * We didn't find an idle thread to use, so we need to queue the xprt.
451 * Do so and then search again. If we find one, we can't hook this one
452 * up to it directly but we can wake the thread up in the hopes that it
453 * will pick it up once it searches for a xprt to service.
457 dprintk("svc: transport %p put into queue\n", xprt);
458 spin_lock_bh(&pool->sp_lock);
459 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
460 pool->sp_stats.sockets_queued++;
461 spin_unlock_bh(&pool->sp_lock);
467 trace_svc_xprt_do_enqueue(xprt, rqstp);
469 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
472 * Queue up a transport with data pending. If there are idle nfsd
473 * processes, wake 'em up.
476 void svc_xprt_enqueue(struct svc_xprt *xprt)
478 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
480 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
482 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
485 * Dequeue the first transport, if there is one.
487 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
489 struct svc_xprt *xprt = NULL;
491 if (list_empty(&pool->sp_sockets))
494 spin_lock_bh(&pool->sp_lock);
495 if (likely(!list_empty(&pool->sp_sockets))) {
496 xprt = list_first_entry(&pool->sp_sockets,
497 struct svc_xprt, xpt_ready);
498 list_del_init(&xprt->xpt_ready);
501 dprintk("svc: transport %p dequeued, inuse=%d\n",
502 xprt, atomic_read(&xprt->xpt_ref.refcount));
504 spin_unlock_bh(&pool->sp_lock);
506 trace_svc_xprt_dequeue(xprt);
511 * svc_reserve - change the space reserved for the reply to a request.
512 * @rqstp: The request in question
513 * @space: new max space to reserve
515 * Each request reserves some space on the output queue of the transport
516 * to make sure the reply fits. This function reduces that reserved
517 * space to be the amount of space used already, plus @space.
520 void svc_reserve(struct svc_rqst *rqstp, int space)
522 struct svc_xprt *xprt = rqstp->rq_xprt;
524 space += rqstp->rq_res.head[0].iov_len;
526 if (xprt && space < rqstp->rq_reserved) {
527 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
528 rqstp->rq_reserved = space;
530 svc_xprt_enqueue(xprt);
533 EXPORT_SYMBOL_GPL(svc_reserve);
535 static void svc_xprt_release(struct svc_rqst *rqstp)
537 struct svc_xprt *xprt = rqstp->rq_xprt;
539 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
541 kfree(rqstp->rq_deferred);
542 rqstp->rq_deferred = NULL;
544 svc_free_res_pages(rqstp);
545 rqstp->rq_res.page_len = 0;
546 rqstp->rq_res.page_base = 0;
548 /* Reset response buffer and release
550 * But first, check that enough space was reserved
551 * for the reply, otherwise we have a bug!
553 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
554 printk(KERN_ERR "RPC request reserved %d but used %d\n",
558 rqstp->rq_res.head[0].iov_len = 0;
559 svc_reserve(rqstp, 0);
560 svc_xprt_release_slot(rqstp);
561 rqstp->rq_xprt = NULL;
566 * Some svc_serv's will have occasional work to do, even when a xprt is not
567 * waiting to be serviced. This function is there to "kick" a task in one of
568 * those services so that it can wake up and do that work. Note that we only
569 * bother with pool 0 as we don't need to wake up more than one thread for
572 void svc_wake_up(struct svc_serv *serv)
574 struct svc_rqst *rqstp;
575 struct svc_pool *pool;
577 pool = &serv->sv_pools[0];
580 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
581 /* skip any that aren't queued */
582 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
585 dprintk("svc: daemon %p woken up.\n", rqstp);
586 wake_up_process(rqstp->rq_task);
587 trace_svc_wake_up(rqstp->rq_task->pid);
592 /* No free entries available */
593 set_bit(SP_TASK_PENDING, &pool->sp_flags);
595 trace_svc_wake_up(0);
597 EXPORT_SYMBOL_GPL(svc_wake_up);
599 int svc_port_is_privileged(struct sockaddr *sin)
601 switch (sin->sa_family) {
603 return ntohs(((struct sockaddr_in *)sin)->sin_port)
606 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
614 * Make sure that we don't have too many active connections. If we have,
615 * something must be dropped. It's not clear what will happen if we allow
616 * "too many" connections, but when dealing with network-facing software,
617 * we have to code defensively. Here we do that by imposing hard limits.
619 * There's no point in trying to do random drop here for DoS
620 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
621 * attacker can easily beat that.
623 * The only somewhat efficient mechanism would be if drop old
624 * connections from the same IP first. But right now we don't even
625 * record the client IP in svc_sock.
627 * single-threaded services that expect a lot of clients will probably
628 * need to set sv_maxconn to override the default value which is based
629 * on the number of threads
631 static void svc_check_conn_limits(struct svc_serv *serv)
633 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
634 (serv->sv_nrthreads+3) * 20;
636 if (serv->sv_tmpcnt > limit) {
637 struct svc_xprt *xprt = NULL;
638 spin_lock_bh(&serv->sv_lock);
639 if (!list_empty(&serv->sv_tempsocks)) {
640 /* Try to help the admin */
641 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
642 serv->sv_name, serv->sv_maxconn ?
643 "max number of connections" :
644 "number of threads");
646 * Always select the oldest connection. It's not fair,
649 xprt = list_entry(serv->sv_tempsocks.prev,
652 set_bit(XPT_CLOSE, &xprt->xpt_flags);
655 spin_unlock_bh(&serv->sv_lock);
658 svc_xprt_enqueue(xprt);
664 static int svc_alloc_arg(struct svc_rqst *rqstp)
666 struct svc_serv *serv = rqstp->rq_server;
671 /* now allocate needed pages. If we get a failure, sleep briefly */
672 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
673 WARN_ON_ONCE(pages >= RPCSVC_MAXPAGES);
674 if (pages >= RPCSVC_MAXPAGES)
675 /* use as many pages as possible */
676 pages = RPCSVC_MAXPAGES - 1;
677 for (i = 0; i < pages ; i++)
678 while (rqstp->rq_pages[i] == NULL) {
679 struct page *p = alloc_page(GFP_KERNEL);
681 set_current_state(TASK_INTERRUPTIBLE);
682 if (signalled() || kthread_should_stop()) {
683 set_current_state(TASK_RUNNING);
686 schedule_timeout(msecs_to_jiffies(500));
688 rqstp->rq_pages[i] = p;
690 rqstp->rq_page_end = &rqstp->rq_pages[i];
691 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
693 /* Make arg->head point to first page and arg->pages point to rest */
694 arg = &rqstp->rq_arg;
695 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
696 arg->head[0].iov_len = PAGE_SIZE;
697 arg->pages = rqstp->rq_pages + 1;
699 /* save at least one page for response */
700 arg->page_len = (pages-2)*PAGE_SIZE;
701 arg->len = (pages-1)*PAGE_SIZE;
702 arg->tail[0].iov_len = 0;
707 rqst_should_sleep(struct svc_rqst *rqstp)
709 struct svc_pool *pool = rqstp->rq_pool;
711 /* did someone call svc_wake_up? */
712 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
715 /* was a socket queued? */
716 if (!list_empty(&pool->sp_sockets))
719 /* are we shutting down? */
720 if (signalled() || kthread_should_stop())
723 /* are we freezing? */
724 if (freezing(current))
730 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
732 struct svc_xprt *xprt;
733 struct svc_pool *pool = rqstp->rq_pool;
736 /* rq_xprt should be clear on entry */
737 WARN_ON_ONCE(rqstp->rq_xprt);
739 /* Normally we will wait up to 5 seconds for any required
740 * cache information to be provided.
742 rqstp->rq_chandle.thread_wait = 5*HZ;
744 xprt = svc_xprt_dequeue(pool);
746 rqstp->rq_xprt = xprt;
748 /* As there is a shortage of threads and this request
749 * had to be queued, don't allow the thread to wait so
750 * long for cache updates.
752 rqstp->rq_chandle.thread_wait = 1*HZ;
753 clear_bit(SP_TASK_PENDING, &pool->sp_flags);
758 * We have to be able to interrupt this wait
759 * to bring down the daemons ...
761 set_current_state(TASK_INTERRUPTIBLE);
762 clear_bit(RQ_BUSY, &rqstp->rq_flags);
765 if (likely(rqst_should_sleep(rqstp)))
766 time_left = schedule_timeout(timeout);
768 __set_current_state(TASK_RUNNING);
772 spin_lock_bh(&rqstp->rq_lock);
773 set_bit(RQ_BUSY, &rqstp->rq_flags);
774 spin_unlock_bh(&rqstp->rq_lock);
776 xprt = rqstp->rq_xprt;
781 atomic_long_inc(&pool->sp_stats.threads_timedout);
783 if (signalled() || kthread_should_stop())
784 return ERR_PTR(-EINTR);
785 return ERR_PTR(-EAGAIN);
788 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
790 spin_lock_bh(&serv->sv_lock);
791 set_bit(XPT_TEMP, &newxpt->xpt_flags);
792 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
794 if (serv->sv_temptimer.function == NULL) {
795 /* setup timer to age temp transports */
796 setup_timer(&serv->sv_temptimer, svc_age_temp_xprts,
797 (unsigned long)serv);
798 mod_timer(&serv->sv_temptimer,
799 jiffies + svc_conn_age_period * HZ);
801 spin_unlock_bh(&serv->sv_lock);
802 svc_xprt_received(newxpt);
805 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
807 struct svc_serv *serv = rqstp->rq_server;
810 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
811 dprintk("svc_recv: found XPT_CLOSE\n");
812 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
813 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
814 svc_delete_xprt(xprt);
815 /* Leave XPT_BUSY set on the dead xprt: */
818 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
819 struct svc_xprt *newxpt;
821 * We know this module_get will succeed because the
822 * listener holds a reference too
824 __module_get(xprt->xpt_class->xcl_owner);
825 svc_check_conn_limits(xprt->xpt_server);
826 newxpt = xprt->xpt_ops->xpo_accept(xprt);
828 svc_add_new_temp_xprt(serv, newxpt);
830 module_put(xprt->xpt_class->xcl_owner);
831 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
832 /* XPT_DATA|XPT_DEFERRED case: */
833 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
834 rqstp, rqstp->rq_pool->sp_id, xprt,
835 atomic_read(&xprt->xpt_ref.refcount));
836 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
837 if (rqstp->rq_deferred)
838 len = svc_deferred_recv(rqstp);
840 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
841 dprintk("svc: got len=%d\n", len);
842 rqstp->rq_reserved = serv->sv_max_mesg;
843 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
845 /* clear XPT_BUSY: */
846 svc_xprt_received(xprt);
848 trace_svc_handle_xprt(xprt, len);
853 * Receive the next request on any transport. This code is carefully
854 * organised not to touch any cachelines in the shared svc_serv
855 * structure, only cachelines in the local svc_pool.
857 int svc_recv(struct svc_rqst *rqstp, long timeout)
859 struct svc_xprt *xprt = NULL;
860 struct svc_serv *serv = rqstp->rq_server;
863 dprintk("svc: server %p waiting for data (to = %ld)\n",
868 "svc_recv: service %p, transport not NULL!\n",
871 err = svc_alloc_arg(rqstp);
878 if (signalled() || kthread_should_stop())
881 xprt = svc_get_next_xprt(rqstp, timeout);
887 len = svc_handle_xprt(rqstp, xprt);
889 /* No data, incomplete (TCP) read, or accept() */
894 clear_bit(XPT_OLD, &xprt->xpt_flags);
896 if (xprt->xpt_ops->xpo_secure_port(rqstp))
897 set_bit(RQ_SECURE, &rqstp->rq_flags);
899 clear_bit(RQ_SECURE, &rqstp->rq_flags);
900 rqstp->rq_chandle.defer = svc_defer;
901 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
904 serv->sv_stats->netcnt++;
905 trace_svc_recv(rqstp, len);
908 rqstp->rq_res.len = 0;
909 svc_xprt_release(rqstp);
911 trace_svc_recv(rqstp, err);
914 EXPORT_SYMBOL_GPL(svc_recv);
919 void svc_drop(struct svc_rqst *rqstp)
921 trace_svc_drop(rqstp);
922 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
923 svc_xprt_release(rqstp);
925 EXPORT_SYMBOL_GPL(svc_drop);
928 * Return reply to client.
930 int svc_send(struct svc_rqst *rqstp)
932 struct svc_xprt *xprt;
936 xprt = rqstp->rq_xprt;
940 /* release the receive skb before sending the reply */
941 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
943 /* calculate over-all length */
945 xb->len = xb->head[0].iov_len +
949 /* Grab mutex to serialize outgoing data. */
950 mutex_lock(&xprt->xpt_mutex);
951 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
952 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
955 len = xprt->xpt_ops->xpo_sendto(rqstp);
956 mutex_unlock(&xprt->xpt_mutex);
957 rpc_wake_up(&xprt->xpt_bc_pending);
958 svc_xprt_release(rqstp);
960 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
963 trace_svc_send(rqstp, len);
968 * Timer function to close old temporary transports, using
969 * a mark-and-sweep algorithm.
971 static void svc_age_temp_xprts(unsigned long closure)
973 struct svc_serv *serv = (struct svc_serv *)closure;
974 struct svc_xprt *xprt;
975 struct list_head *le, *next;
977 dprintk("svc_age_temp_xprts\n");
979 if (!spin_trylock_bh(&serv->sv_lock)) {
980 /* busy, try again 1 sec later */
981 dprintk("svc_age_temp_xprts: busy\n");
982 mod_timer(&serv->sv_temptimer, jiffies + HZ);
986 list_for_each_safe(le, next, &serv->sv_tempsocks) {
987 xprt = list_entry(le, struct svc_xprt, xpt_list);
989 /* First time through, just mark it OLD. Second time
990 * through, close it. */
991 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
993 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
994 test_bit(XPT_BUSY, &xprt->xpt_flags))
997 set_bit(XPT_CLOSE, &xprt->xpt_flags);
998 dprintk("queuing xprt %p for closing\n", xprt);
1000 /* a thread will dequeue and close it soon */
1001 svc_xprt_enqueue(xprt);
1003 spin_unlock_bh(&serv->sv_lock);
1005 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
1008 /* Close temporary transports whose xpt_local matches server_addr immediately
1009 * instead of waiting for them to be picked up by the timer.
1011 * This is meant to be called from a notifier_block that runs when an ip
1012 * address is deleted.
1014 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
1016 struct svc_xprt *xprt;
1017 struct list_head *le, *next;
1018 LIST_HEAD(to_be_closed);
1020 spin_lock_bh(&serv->sv_lock);
1021 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1022 xprt = list_entry(le, struct svc_xprt, xpt_list);
1023 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
1024 &xprt->xpt_local)) {
1025 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
1026 list_move(le, &to_be_closed);
1029 spin_unlock_bh(&serv->sv_lock);
1031 while (!list_empty(&to_be_closed)) {
1032 le = to_be_closed.next;
1034 xprt = list_entry(le, struct svc_xprt, xpt_list);
1035 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1036 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
1037 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
1039 svc_xprt_enqueue(xprt);
1042 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1044 static void call_xpt_users(struct svc_xprt *xprt)
1046 struct svc_xpt_user *u;
1048 spin_lock(&xprt->xpt_lock);
1049 while (!list_empty(&xprt->xpt_users)) {
1050 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1051 list_del_init(&u->list);
1054 spin_unlock(&xprt->xpt_lock);
1058 * Remove a dead transport
1060 static void svc_delete_xprt(struct svc_xprt *xprt)
1062 struct svc_serv *serv = xprt->xpt_server;
1063 struct svc_deferred_req *dr;
1065 /* Only do this once */
1066 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1069 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1070 xprt->xpt_ops->xpo_detach(xprt);
1072 spin_lock_bh(&serv->sv_lock);
1073 list_del_init(&xprt->xpt_list);
1074 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1075 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1077 spin_unlock_bh(&serv->sv_lock);
1079 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1082 call_xpt_users(xprt);
1086 void svc_close_xprt(struct svc_xprt *xprt)
1088 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1089 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1090 /* someone else will have to effect the close */
1093 * We expect svc_close_xprt() to work even when no threads are
1094 * running (e.g., while configuring the server before starting
1095 * any threads), so if the transport isn't busy, we delete
1098 svc_delete_xprt(xprt);
1100 EXPORT_SYMBOL_GPL(svc_close_xprt);
1102 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1104 struct svc_xprt *xprt;
1107 spin_lock_bh(&serv->sv_lock);
1108 list_for_each_entry(xprt, xprt_list, xpt_list) {
1109 if (xprt->xpt_net != net)
1112 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1113 svc_xprt_enqueue(xprt);
1115 spin_unlock_bh(&serv->sv_lock);
1119 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1121 struct svc_pool *pool;
1122 struct svc_xprt *xprt;
1123 struct svc_xprt *tmp;
1126 for (i = 0; i < serv->sv_nrpools; i++) {
1127 pool = &serv->sv_pools[i];
1129 spin_lock_bh(&pool->sp_lock);
1130 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1131 if (xprt->xpt_net != net)
1133 list_del_init(&xprt->xpt_ready);
1134 spin_unlock_bh(&pool->sp_lock);
1137 spin_unlock_bh(&pool->sp_lock);
1142 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1144 struct svc_xprt *xprt;
1146 while ((xprt = svc_dequeue_net(serv, net))) {
1147 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1148 svc_delete_xprt(xprt);
1153 * Server threads may still be running (especially in the case where the
1154 * service is still running in other network namespaces).
1156 * So we shut down sockets the same way we would on a running server, by
1157 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1158 * the close. In the case there are no such other threads,
1159 * threads running, svc_clean_up_xprts() does a simple version of a
1160 * server's main event loop, and in the case where there are other
1161 * threads, we may need to wait a little while and then check again to
1162 * see if they're done.
1164 void svc_close_net(struct svc_serv *serv, struct net *net)
1168 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1169 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1171 svc_clean_up_xprts(serv, net);
1177 * Handle defer and revisit of requests
1180 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1182 struct svc_deferred_req *dr =
1183 container_of(dreq, struct svc_deferred_req, handle);
1184 struct svc_xprt *xprt = dr->xprt;
1186 spin_lock(&xprt->xpt_lock);
1187 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1188 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1189 spin_unlock(&xprt->xpt_lock);
1190 dprintk("revisit canceled\n");
1192 trace_svc_drop_deferred(dr);
1196 dprintk("revisit queued\n");
1198 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1199 spin_unlock(&xprt->xpt_lock);
1200 svc_xprt_enqueue(xprt);
1205 * Save the request off for later processing. The request buffer looks
1208 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1210 * This code can only handle requests that consist of an xprt-header
1213 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1215 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1216 struct svc_deferred_req *dr;
1218 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1219 return NULL; /* if more than a page, give up FIXME */
1220 if (rqstp->rq_deferred) {
1221 dr = rqstp->rq_deferred;
1222 rqstp->rq_deferred = NULL;
1226 /* FIXME maybe discard if size too large */
1227 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1228 dr = kmalloc(size, GFP_KERNEL);
1232 dr->handle.owner = rqstp->rq_server;
1233 dr->prot = rqstp->rq_prot;
1234 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1235 dr->addrlen = rqstp->rq_addrlen;
1236 dr->daddr = rqstp->rq_daddr;
1237 dr->argslen = rqstp->rq_arg.len >> 2;
1238 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1240 /* back up head to the start of the buffer and copy */
1241 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1242 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1245 svc_xprt_get(rqstp->rq_xprt);
1246 dr->xprt = rqstp->rq_xprt;
1247 set_bit(RQ_DROPME, &rqstp->rq_flags);
1249 dr->handle.revisit = svc_revisit;
1250 trace_svc_defer(rqstp);
1255 * recv data from a deferred request into an active one
1257 static int svc_deferred_recv(struct svc_rqst *rqstp)
1259 struct svc_deferred_req *dr = rqstp->rq_deferred;
1261 /* setup iov_base past transport header */
1262 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1263 /* The iov_len does not include the transport header bytes */
1264 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1265 rqstp->rq_arg.page_len = 0;
1266 /* The rq_arg.len includes the transport header bytes */
1267 rqstp->rq_arg.len = dr->argslen<<2;
1268 rqstp->rq_prot = dr->prot;
1269 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1270 rqstp->rq_addrlen = dr->addrlen;
1271 /* Save off transport header len in case we get deferred again */
1272 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1273 rqstp->rq_daddr = dr->daddr;
1274 rqstp->rq_respages = rqstp->rq_pages;
1275 return (dr->argslen<<2) - dr->xprt_hlen;
1279 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1281 struct svc_deferred_req *dr = NULL;
1283 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1285 spin_lock(&xprt->xpt_lock);
1286 if (!list_empty(&xprt->xpt_deferred)) {
1287 dr = list_entry(xprt->xpt_deferred.next,
1288 struct svc_deferred_req,
1290 list_del_init(&dr->handle.recent);
1291 trace_svc_revisit_deferred(dr);
1293 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1294 spin_unlock(&xprt->xpt_lock);
1299 * svc_find_xprt - find an RPC transport instance
1300 * @serv: pointer to svc_serv to search
1301 * @xcl_name: C string containing transport's class name
1302 * @net: owner net pointer
1303 * @af: Address family of transport's local address
1304 * @port: transport's IP port number
1306 * Return the transport instance pointer for the endpoint accepting
1307 * connections/peer traffic from the specified transport class,
1308 * address family and port.
1310 * Specifying 0 for the address family or port is effectively a
1311 * wild-card, and will result in matching the first transport in the
1312 * service's list that has a matching class name.
1314 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1315 struct net *net, const sa_family_t af,
1316 const unsigned short port)
1318 struct svc_xprt *xprt;
1319 struct svc_xprt *found = NULL;
1321 /* Sanity check the args */
1322 if (serv == NULL || xcl_name == NULL)
1325 spin_lock_bh(&serv->sv_lock);
1326 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1327 if (xprt->xpt_net != net)
1329 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1331 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1333 if (port != 0 && port != svc_xprt_local_port(xprt))
1339 spin_unlock_bh(&serv->sv_lock);
1342 EXPORT_SYMBOL_GPL(svc_find_xprt);
1344 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1345 char *pos, int remaining)
1349 len = snprintf(pos, remaining, "%s %u\n",
1350 xprt->xpt_class->xcl_name,
1351 svc_xprt_local_port(xprt));
1352 if (len >= remaining)
1353 return -ENAMETOOLONG;
1358 * svc_xprt_names - format a buffer with a list of transport names
1359 * @serv: pointer to an RPC service
1360 * @buf: pointer to a buffer to be filled in
1361 * @buflen: length of buffer to be filled in
1363 * Fills in @buf with a string containing a list of transport names,
1364 * each name terminated with '\n'.
1366 * Returns positive length of the filled-in string on success; otherwise
1367 * a negative errno value is returned if an error occurs.
1369 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1371 struct svc_xprt *xprt;
1375 /* Sanity check args */
1379 spin_lock_bh(&serv->sv_lock);
1383 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1384 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1396 spin_unlock_bh(&serv->sv_lock);
1399 EXPORT_SYMBOL_GPL(svc_xprt_names);
1402 /*----------------------------------------------------------------------------*/
1404 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1406 unsigned int pidx = (unsigned int)*pos;
1407 struct svc_serv *serv = m->private;
1409 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1412 return SEQ_START_TOKEN;
1413 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1416 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1418 struct svc_pool *pool = p;
1419 struct svc_serv *serv = m->private;
1421 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1423 if (p == SEQ_START_TOKEN) {
1424 pool = &serv->sv_pools[0];
1426 unsigned int pidx = (pool - &serv->sv_pools[0]);
1427 if (pidx < serv->sv_nrpools-1)
1428 pool = &serv->sv_pools[pidx+1];
1436 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1440 static int svc_pool_stats_show(struct seq_file *m, void *p)
1442 struct svc_pool *pool = p;
1444 if (p == SEQ_START_TOKEN) {
1445 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1449 seq_printf(m, "%u %lu %lu %lu %lu\n",
1451 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1452 pool->sp_stats.sockets_queued,
1453 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1454 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1459 static const struct seq_operations svc_pool_stats_seq_ops = {
1460 .start = svc_pool_stats_start,
1461 .next = svc_pool_stats_next,
1462 .stop = svc_pool_stats_stop,
1463 .show = svc_pool_stats_show,
1466 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1470 err = seq_open(file, &svc_pool_stats_seq_ops);
1472 ((struct seq_file *) file->private_data)->private = serv;
1475 EXPORT_SYMBOL(svc_pool_stats_open);
1477 /*----------------------------------------------------------------------------*/