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);
28 /* apparently the "standard" is that clients close
29 * idle connections after 5 minutes, servers after
31 * http://www.connectathon.org/talks96/nfstcp.pdf
33 static int svc_conn_age_period = 6*60;
35 /* List of registered transport classes */
36 static DEFINE_SPINLOCK(svc_xprt_class_lock);
37 static LIST_HEAD(svc_xprt_class_list);
39 /* SMP locking strategy:
41 * svc_pool->sp_lock protects most of the fields of that pool.
42 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
43 * when both need to be taken (rare), svc_serv->sv_lock is first.
44 * The "service mutex" protects svc_serv->sv_nrthread.
45 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
46 * and the ->sk_info_authunix cache.
48 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
49 * enqueued multiply. During normal transport processing this bit
50 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
51 * Providers should not manipulate this bit directly.
53 * Some flags can be set to certain values at any time
54 * providing that certain rules are followed:
57 * - Can be set or cleared at any time.
58 * - After a set, svc_xprt_enqueue must be called to enqueue
59 * the transport for processing.
60 * - After a clear, the transport must be read/accepted.
61 * If this succeeds, it must be set again.
63 * - Can set at any time. It is never cleared.
65 * - Can only be set while XPT_BUSY is held which ensures
66 * that no other thread will be using the transport or will
67 * try to set XPT_DEAD.
69 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
71 struct svc_xprt_class *cl;
74 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
76 INIT_LIST_HEAD(&xcl->xcl_list);
77 spin_lock(&svc_xprt_class_lock);
78 /* Make sure there isn't already a class with the same name */
79 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
80 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
83 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
86 spin_unlock(&svc_xprt_class_lock);
89 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
91 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
93 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
94 spin_lock(&svc_xprt_class_lock);
95 list_del_init(&xcl->xcl_list);
96 spin_unlock(&svc_xprt_class_lock);
98 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
101 * svc_print_xprts - Format the transport list for printing
102 * @buf: target buffer for formatted address
103 * @maxlen: length of target buffer
105 * Fills in @buf with a string containing a list of transport names, each name
106 * terminated with '\n'. If the buffer is too small, some entries may be
107 * missing, but it is guaranteed that all lines in the output buffer are
110 * Returns positive length of the filled-in string.
112 int svc_print_xprts(char *buf, int maxlen)
114 struct svc_xprt_class *xcl;
119 spin_lock(&svc_xprt_class_lock);
120 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
123 slen = snprintf(tmpstr, sizeof(tmpstr), "%s %d\n",
124 xcl->xcl_name, xcl->xcl_max_payload);
125 if (slen >= sizeof(tmpstr) || len + slen >= maxlen)
130 spin_unlock(&svc_xprt_class_lock);
135 static void svc_xprt_free(struct kref *kref)
137 struct svc_xprt *xprt =
138 container_of(kref, struct svc_xprt, xpt_ref);
139 struct module *owner = xprt->xpt_class->xcl_owner;
140 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
141 svcauth_unix_info_release(xprt);
142 put_net(xprt->xpt_net);
143 /* See comment on corresponding get in xs_setup_bc_tcp(): */
144 if (xprt->xpt_bc_xprt)
145 xprt_put(xprt->xpt_bc_xprt);
146 xprt->xpt_ops->xpo_free(xprt);
150 void svc_xprt_put(struct svc_xprt *xprt)
152 kref_put(&xprt->xpt_ref, svc_xprt_free);
154 EXPORT_SYMBOL_GPL(svc_xprt_put);
157 * Called by transport drivers to initialize the transport independent
158 * portion of the transport instance.
160 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
161 struct svc_xprt *xprt, struct svc_serv *serv)
163 memset(xprt, 0, sizeof(*xprt));
164 xprt->xpt_class = xcl;
165 xprt->xpt_ops = xcl->xcl_ops;
166 kref_init(&xprt->xpt_ref);
167 xprt->xpt_server = serv;
168 INIT_LIST_HEAD(&xprt->xpt_list);
169 INIT_LIST_HEAD(&xprt->xpt_ready);
170 INIT_LIST_HEAD(&xprt->xpt_deferred);
171 INIT_LIST_HEAD(&xprt->xpt_users);
172 mutex_init(&xprt->xpt_mutex);
173 spin_lock_init(&xprt->xpt_lock);
174 set_bit(XPT_BUSY, &xprt->xpt_flags);
175 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
176 xprt->xpt_net = get_net(net);
178 EXPORT_SYMBOL_GPL(svc_xprt_init);
180 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
181 struct svc_serv *serv,
184 const unsigned short port,
187 struct sockaddr_in sin = {
188 .sin_family = AF_INET,
189 .sin_addr.s_addr = htonl(INADDR_ANY),
190 .sin_port = htons(port),
192 #if IS_ENABLED(CONFIG_IPV6)
193 struct sockaddr_in6 sin6 = {
194 .sin6_family = AF_INET6,
195 .sin6_addr = IN6ADDR_ANY_INIT,
196 .sin6_port = htons(port),
199 struct sockaddr *sap;
204 sap = (struct sockaddr *)&sin;
207 #if IS_ENABLED(CONFIG_IPV6)
209 sap = (struct sockaddr *)&sin6;
214 return ERR_PTR(-EAFNOSUPPORT);
217 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
221 * svc_xprt_received conditionally queues the transport for processing
222 * by another thread. The caller must hold the XPT_BUSY bit and must
223 * not thereafter touch transport data.
225 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
226 * insufficient) data.
228 static void svc_xprt_received(struct svc_xprt *xprt)
230 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
231 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
235 /* As soon as we clear busy, the xprt could be closed and
236 * 'put', so we need a reference to call svc_enqueue_xprt with:
239 smp_mb__before_atomic();
240 clear_bit(XPT_BUSY, &xprt->xpt_flags);
241 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
245 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
247 clear_bit(XPT_TEMP, &new->xpt_flags);
248 spin_lock_bh(&serv->sv_lock);
249 list_add(&new->xpt_list, &serv->sv_permsocks);
250 spin_unlock_bh(&serv->sv_lock);
251 svc_xprt_received(new);
254 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
255 struct net *net, const int family,
256 const unsigned short port, int flags)
258 struct svc_xprt_class *xcl;
260 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
261 spin_lock(&svc_xprt_class_lock);
262 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
263 struct svc_xprt *newxprt;
264 unsigned short newport;
266 if (strcmp(xprt_name, xcl->xcl_name))
269 if (!try_module_get(xcl->xcl_owner))
272 spin_unlock(&svc_xprt_class_lock);
273 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
274 if (IS_ERR(newxprt)) {
275 module_put(xcl->xcl_owner);
276 return PTR_ERR(newxprt);
278 svc_add_new_perm_xprt(serv, newxprt);
279 newport = svc_xprt_local_port(newxprt);
283 spin_unlock(&svc_xprt_class_lock);
284 dprintk("svc: transport %s not found\n", xprt_name);
286 /* This errno is exposed to user space. Provide a reasonable
287 * perror msg for a bad transport. */
288 return -EPROTONOSUPPORT;
290 EXPORT_SYMBOL_GPL(svc_create_xprt);
293 * Copy the local and remote xprt addresses to the rqstp structure
295 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
297 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
298 rqstp->rq_addrlen = xprt->xpt_remotelen;
301 * Destination address in request is needed for binding the
302 * source address in RPC replies/callbacks later.
304 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
305 rqstp->rq_daddrlen = xprt->xpt_locallen;
307 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
310 * svc_print_addr - Format rq_addr field for printing
311 * @rqstp: svc_rqst struct containing address to print
312 * @buf: target buffer for formatted address
313 * @len: length of target buffer
316 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
318 return __svc_print_addr(svc_addr(rqstp), buf, len);
320 EXPORT_SYMBOL_GPL(svc_print_addr);
322 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
324 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
326 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
327 return xprt->xpt_ops->xpo_has_wspace(xprt);
331 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
333 struct svc_pool *pool;
334 struct svc_rqst *rqstp = NULL;
338 if (!svc_xprt_has_something_to_do(xprt))
341 /* Mark transport as busy. It will remain in this state until
342 * the provider calls svc_xprt_received. We update XPT_BUSY
343 * atomically because it also guards against trying to enqueue
344 * the transport twice.
346 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
347 /* Don't enqueue transport while already enqueued */
348 dprintk("svc: transport %p busy, not enqueued\n", xprt);
353 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
355 atomic_long_inc(&pool->sp_stats.packets);
358 /* find a thread for this xprt */
360 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
361 /* Do a lockless check first */
362 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
366 * Once the xprt has been queued, it can only be dequeued by
367 * the task that intends to service it. All we can do at that
368 * point is to try to wake this thread back up so that it can
372 spin_lock_bh(&rqstp->rq_lock);
373 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags)) {
374 /* already busy, move on... */
375 spin_unlock_bh(&rqstp->rq_lock);
379 /* this one will do */
380 rqstp->rq_xprt = xprt;
382 spin_unlock_bh(&rqstp->rq_lock);
386 atomic_long_inc(&pool->sp_stats.threads_woken);
387 wake_up_process(rqstp->rq_task);
394 * We didn't find an idle thread to use, so we need to queue the xprt.
395 * Do so and then search again. If we find one, we can't hook this one
396 * up to it directly but we can wake the thread up in the hopes that it
397 * will pick it up once it searches for a xprt to service.
401 dprintk("svc: transport %p put into queue\n", xprt);
402 spin_lock_bh(&pool->sp_lock);
403 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
404 pool->sp_stats.sockets_queued++;
405 spin_unlock_bh(&pool->sp_lock);
411 trace_svc_xprt_do_enqueue(xprt, rqstp);
413 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
416 * Queue up a transport with data pending. If there are idle nfsd
417 * processes, wake 'em up.
420 void svc_xprt_enqueue(struct svc_xprt *xprt)
422 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
424 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
426 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
429 * Dequeue the first transport, if there is one.
431 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
433 struct svc_xprt *xprt = NULL;
435 if (list_empty(&pool->sp_sockets))
438 spin_lock_bh(&pool->sp_lock);
439 if (likely(!list_empty(&pool->sp_sockets))) {
440 xprt = list_first_entry(&pool->sp_sockets,
441 struct svc_xprt, xpt_ready);
442 list_del_init(&xprt->xpt_ready);
445 dprintk("svc: transport %p dequeued, inuse=%d\n",
446 xprt, atomic_read(&xprt->xpt_ref.refcount));
448 spin_unlock_bh(&pool->sp_lock);
450 trace_svc_xprt_dequeue(xprt);
455 * svc_reserve - change the space reserved for the reply to a request.
456 * @rqstp: The request in question
457 * @space: new max space to reserve
459 * Each request reserves some space on the output queue of the transport
460 * to make sure the reply fits. This function reduces that reserved
461 * space to be the amount of space used already, plus @space.
464 void svc_reserve(struct svc_rqst *rqstp, int space)
466 struct svc_xprt *xprt = rqstp->rq_xprt;
468 space += rqstp->rq_res.head[0].iov_len;
470 if (xprt && space < rqstp->rq_reserved) {
471 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
472 rqstp->rq_reserved = space;
474 if (xprt->xpt_ops->xpo_adjust_wspace)
475 xprt->xpt_ops->xpo_adjust_wspace(xprt);
476 svc_xprt_enqueue(xprt);
479 EXPORT_SYMBOL_GPL(svc_reserve);
481 static void svc_xprt_release(struct svc_rqst *rqstp)
483 struct svc_xprt *xprt = rqstp->rq_xprt;
485 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
487 kfree(rqstp->rq_deferred);
488 rqstp->rq_deferred = NULL;
490 svc_free_res_pages(rqstp);
491 rqstp->rq_res.page_len = 0;
492 rqstp->rq_res.page_base = 0;
494 /* Reset response buffer and release
496 * But first, check that enough space was reserved
497 * for the reply, otherwise we have a bug!
499 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
500 printk(KERN_ERR "RPC request reserved %d but used %d\n",
504 rqstp->rq_res.head[0].iov_len = 0;
505 svc_reserve(rqstp, 0);
506 rqstp->rq_xprt = NULL;
512 * Some svc_serv's will have occasional work to do, even when a xprt is not
513 * waiting to be serviced. This function is there to "kick" a task in one of
514 * those services so that it can wake up and do that work. Note that we only
515 * bother with pool 0 as we don't need to wake up more than one thread for
518 void svc_wake_up(struct svc_serv *serv)
520 struct svc_rqst *rqstp;
521 struct svc_pool *pool;
523 pool = &serv->sv_pools[0];
526 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
527 /* skip any that aren't queued */
528 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
531 dprintk("svc: daemon %p woken up.\n", rqstp);
532 wake_up_process(rqstp->rq_task);
533 trace_svc_wake_up(rqstp->rq_task->pid);
538 /* No free entries available */
539 set_bit(SP_TASK_PENDING, &pool->sp_flags);
541 trace_svc_wake_up(0);
543 EXPORT_SYMBOL_GPL(svc_wake_up);
545 int svc_port_is_privileged(struct sockaddr *sin)
547 switch (sin->sa_family) {
549 return ntohs(((struct sockaddr_in *)sin)->sin_port)
552 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
560 * Make sure that we don't have too many active connections. If we have,
561 * something must be dropped. It's not clear what will happen if we allow
562 * "too many" connections, but when dealing with network-facing software,
563 * we have to code defensively. Here we do that by imposing hard limits.
565 * There's no point in trying to do random drop here for DoS
566 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
567 * attacker can easily beat that.
569 * The only somewhat efficient mechanism would be if drop old
570 * connections from the same IP first. But right now we don't even
571 * record the client IP in svc_sock.
573 * single-threaded services that expect a lot of clients will probably
574 * need to set sv_maxconn to override the default value which is based
575 * on the number of threads
577 static void svc_check_conn_limits(struct svc_serv *serv)
579 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
580 (serv->sv_nrthreads+3) * 20;
582 if (serv->sv_tmpcnt > limit) {
583 struct svc_xprt *xprt = NULL;
584 spin_lock_bh(&serv->sv_lock);
585 if (!list_empty(&serv->sv_tempsocks)) {
586 /* Try to help the admin */
587 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
588 serv->sv_name, serv->sv_maxconn ?
589 "max number of connections" :
590 "number of threads");
592 * Always select the oldest connection. It's not fair,
595 xprt = list_entry(serv->sv_tempsocks.prev,
598 set_bit(XPT_CLOSE, &xprt->xpt_flags);
601 spin_unlock_bh(&serv->sv_lock);
604 svc_xprt_enqueue(xprt);
610 static int svc_alloc_arg(struct svc_rqst *rqstp)
612 struct svc_serv *serv = rqstp->rq_server;
617 /* now allocate needed pages. If we get a failure, sleep briefly */
618 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
619 WARN_ON_ONCE(pages >= RPCSVC_MAXPAGES);
620 if (pages >= RPCSVC_MAXPAGES)
621 /* use as many pages as possible */
622 pages = RPCSVC_MAXPAGES - 1;
623 for (i = 0; i < pages ; i++)
624 while (rqstp->rq_pages[i] == NULL) {
625 struct page *p = alloc_page(GFP_KERNEL);
627 set_current_state(TASK_INTERRUPTIBLE);
628 if (signalled() || kthread_should_stop()) {
629 set_current_state(TASK_RUNNING);
632 schedule_timeout(msecs_to_jiffies(500));
634 rqstp->rq_pages[i] = p;
636 rqstp->rq_page_end = &rqstp->rq_pages[i];
637 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
639 /* Make arg->head point to first page and arg->pages point to rest */
640 arg = &rqstp->rq_arg;
641 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
642 arg->head[0].iov_len = PAGE_SIZE;
643 arg->pages = rqstp->rq_pages + 1;
645 /* save at least one page for response */
646 arg->page_len = (pages-2)*PAGE_SIZE;
647 arg->len = (pages-1)*PAGE_SIZE;
648 arg->tail[0].iov_len = 0;
653 rqst_should_sleep(struct svc_rqst *rqstp)
655 struct svc_pool *pool = rqstp->rq_pool;
657 /* did someone call svc_wake_up? */
658 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
661 /* was a socket queued? */
662 if (!list_empty(&pool->sp_sockets))
665 /* are we shutting down? */
666 if (signalled() || kthread_should_stop())
669 /* are we freezing? */
670 if (freezing(current))
676 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
678 struct svc_xprt *xprt;
679 struct svc_pool *pool = rqstp->rq_pool;
682 /* rq_xprt should be clear on entry */
683 WARN_ON_ONCE(rqstp->rq_xprt);
685 /* Normally we will wait up to 5 seconds for any required
686 * cache information to be provided.
688 rqstp->rq_chandle.thread_wait = 5*HZ;
690 xprt = svc_xprt_dequeue(pool);
692 rqstp->rq_xprt = xprt;
694 /* As there is a shortage of threads and this request
695 * had to be queued, don't allow the thread to wait so
696 * long for cache updates.
698 rqstp->rq_chandle.thread_wait = 1*HZ;
699 clear_bit(SP_TASK_PENDING, &pool->sp_flags);
704 * We have to be able to interrupt this wait
705 * to bring down the daemons ...
707 set_current_state(TASK_INTERRUPTIBLE);
708 clear_bit(RQ_BUSY, &rqstp->rq_flags);
711 if (likely(rqst_should_sleep(rqstp)))
712 time_left = schedule_timeout(timeout);
714 __set_current_state(TASK_RUNNING);
718 spin_lock_bh(&rqstp->rq_lock);
719 set_bit(RQ_BUSY, &rqstp->rq_flags);
720 spin_unlock_bh(&rqstp->rq_lock);
722 xprt = rqstp->rq_xprt;
727 atomic_long_inc(&pool->sp_stats.threads_timedout);
729 if (signalled() || kthread_should_stop())
730 return ERR_PTR(-EINTR);
731 return ERR_PTR(-EAGAIN);
734 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
736 spin_lock_bh(&serv->sv_lock);
737 set_bit(XPT_TEMP, &newxpt->xpt_flags);
738 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
740 if (serv->sv_temptimer.function == NULL) {
741 /* setup timer to age temp transports */
742 setup_timer(&serv->sv_temptimer, svc_age_temp_xprts,
743 (unsigned long)serv);
744 mod_timer(&serv->sv_temptimer,
745 jiffies + svc_conn_age_period * HZ);
747 spin_unlock_bh(&serv->sv_lock);
748 svc_xprt_received(newxpt);
751 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
753 struct svc_serv *serv = rqstp->rq_server;
756 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
757 dprintk("svc_recv: found XPT_CLOSE\n");
758 svc_delete_xprt(xprt);
759 /* Leave XPT_BUSY set on the dead xprt: */
762 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
763 struct svc_xprt *newxpt;
765 * We know this module_get will succeed because the
766 * listener holds a reference too
768 __module_get(xprt->xpt_class->xcl_owner);
769 svc_check_conn_limits(xprt->xpt_server);
770 newxpt = xprt->xpt_ops->xpo_accept(xprt);
772 svc_add_new_temp_xprt(serv, newxpt);
774 module_put(xprt->xpt_class->xcl_owner);
776 /* XPT_DATA|XPT_DEFERRED case: */
777 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
778 rqstp, rqstp->rq_pool->sp_id, xprt,
779 atomic_read(&xprt->xpt_ref.refcount));
780 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
781 if (rqstp->rq_deferred)
782 len = svc_deferred_recv(rqstp);
784 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
785 dprintk("svc: got len=%d\n", len);
786 rqstp->rq_reserved = serv->sv_max_mesg;
787 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
789 /* clear XPT_BUSY: */
790 svc_xprt_received(xprt);
792 trace_svc_handle_xprt(xprt, len);
797 * Receive the next request on any transport. This code is carefully
798 * organised not to touch any cachelines in the shared svc_serv
799 * structure, only cachelines in the local svc_pool.
801 int svc_recv(struct svc_rqst *rqstp, long timeout)
803 struct svc_xprt *xprt = NULL;
804 struct svc_serv *serv = rqstp->rq_server;
807 dprintk("svc: server %p waiting for data (to = %ld)\n",
812 "svc_recv: service %p, transport not NULL!\n",
815 err = svc_alloc_arg(rqstp);
822 if (signalled() || kthread_should_stop())
825 xprt = svc_get_next_xprt(rqstp, timeout);
831 len = svc_handle_xprt(rqstp, xprt);
833 /* No data, incomplete (TCP) read, or accept() */
838 clear_bit(XPT_OLD, &xprt->xpt_flags);
840 if (xprt->xpt_ops->xpo_secure_port(rqstp))
841 set_bit(RQ_SECURE, &rqstp->rq_flags);
843 clear_bit(RQ_SECURE, &rqstp->rq_flags);
844 rqstp->rq_chandle.defer = svc_defer;
845 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
848 serv->sv_stats->netcnt++;
849 trace_svc_recv(rqstp, len);
852 rqstp->rq_res.len = 0;
853 svc_xprt_release(rqstp);
855 trace_svc_recv(rqstp, err);
858 EXPORT_SYMBOL_GPL(svc_recv);
863 void svc_drop(struct svc_rqst *rqstp)
865 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
866 svc_xprt_release(rqstp);
868 EXPORT_SYMBOL_GPL(svc_drop);
871 * Return reply to client.
873 int svc_send(struct svc_rqst *rqstp)
875 struct svc_xprt *xprt;
879 xprt = rqstp->rq_xprt;
883 /* release the receive skb before sending the reply */
884 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
886 /* calculate over-all length */
888 xb->len = xb->head[0].iov_len +
892 /* Grab mutex to serialize outgoing data. */
893 mutex_lock(&xprt->xpt_mutex);
894 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
895 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
898 len = xprt->xpt_ops->xpo_sendto(rqstp);
899 mutex_unlock(&xprt->xpt_mutex);
900 rpc_wake_up(&xprt->xpt_bc_pending);
901 svc_xprt_release(rqstp);
903 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
906 trace_svc_send(rqstp, len);
911 * Timer function to close old temporary transports, using
912 * a mark-and-sweep algorithm.
914 static void svc_age_temp_xprts(unsigned long closure)
916 struct svc_serv *serv = (struct svc_serv *)closure;
917 struct svc_xprt *xprt;
918 struct list_head *le, *next;
920 dprintk("svc_age_temp_xprts\n");
922 if (!spin_trylock_bh(&serv->sv_lock)) {
923 /* busy, try again 1 sec later */
924 dprintk("svc_age_temp_xprts: busy\n");
925 mod_timer(&serv->sv_temptimer, jiffies + HZ);
929 list_for_each_safe(le, next, &serv->sv_tempsocks) {
930 xprt = list_entry(le, struct svc_xprt, xpt_list);
932 /* First time through, just mark it OLD. Second time
933 * through, close it. */
934 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
936 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
937 test_bit(XPT_BUSY, &xprt->xpt_flags))
940 set_bit(XPT_CLOSE, &xprt->xpt_flags);
941 dprintk("queuing xprt %p for closing\n", xprt);
943 /* a thread will dequeue and close it soon */
944 svc_xprt_enqueue(xprt);
946 spin_unlock_bh(&serv->sv_lock);
948 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
951 static void call_xpt_users(struct svc_xprt *xprt)
953 struct svc_xpt_user *u;
955 spin_lock(&xprt->xpt_lock);
956 while (!list_empty(&xprt->xpt_users)) {
957 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
958 list_del_init(&u->list);
961 spin_unlock(&xprt->xpt_lock);
965 * Remove a dead transport
967 static void svc_delete_xprt(struct svc_xprt *xprt)
969 struct svc_serv *serv = xprt->xpt_server;
970 struct svc_deferred_req *dr;
972 /* Only do this once */
973 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
976 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
977 xprt->xpt_ops->xpo_detach(xprt);
979 spin_lock_bh(&serv->sv_lock);
980 list_del_init(&xprt->xpt_list);
981 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
982 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
984 spin_unlock_bh(&serv->sv_lock);
986 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
989 call_xpt_users(xprt);
993 void svc_close_xprt(struct svc_xprt *xprt)
995 set_bit(XPT_CLOSE, &xprt->xpt_flags);
996 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
997 /* someone else will have to effect the close */
1000 * We expect svc_close_xprt() to work even when no threads are
1001 * running (e.g., while configuring the server before starting
1002 * any threads), so if the transport isn't busy, we delete
1005 svc_delete_xprt(xprt);
1007 EXPORT_SYMBOL_GPL(svc_close_xprt);
1009 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1011 struct svc_xprt *xprt;
1014 spin_lock_bh(&serv->sv_lock);
1015 list_for_each_entry(xprt, xprt_list, xpt_list) {
1016 if (xprt->xpt_net != net)
1019 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1020 svc_xprt_enqueue(xprt);
1022 spin_unlock_bh(&serv->sv_lock);
1026 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1028 struct svc_pool *pool;
1029 struct svc_xprt *xprt;
1030 struct svc_xprt *tmp;
1033 for (i = 0; i < serv->sv_nrpools; i++) {
1034 pool = &serv->sv_pools[i];
1036 spin_lock_bh(&pool->sp_lock);
1037 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1038 if (xprt->xpt_net != net)
1040 list_del_init(&xprt->xpt_ready);
1041 spin_unlock_bh(&pool->sp_lock);
1044 spin_unlock_bh(&pool->sp_lock);
1049 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1051 struct svc_xprt *xprt;
1053 while ((xprt = svc_dequeue_net(serv, net))) {
1054 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1055 svc_delete_xprt(xprt);
1060 * Server threads may still be running (especially in the case where the
1061 * service is still running in other network namespaces).
1063 * So we shut down sockets the same way we would on a running server, by
1064 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1065 * the close. In the case there are no such other threads,
1066 * threads running, svc_clean_up_xprts() does a simple version of a
1067 * server's main event loop, and in the case where there are other
1068 * threads, we may need to wait a little while and then check again to
1069 * see if they're done.
1071 void svc_close_net(struct svc_serv *serv, struct net *net)
1075 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1076 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1078 svc_clean_up_xprts(serv, net);
1084 * Handle defer and revisit of requests
1087 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1089 struct svc_deferred_req *dr =
1090 container_of(dreq, struct svc_deferred_req, handle);
1091 struct svc_xprt *xprt = dr->xprt;
1093 spin_lock(&xprt->xpt_lock);
1094 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1095 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1096 spin_unlock(&xprt->xpt_lock);
1097 dprintk("revisit canceled\n");
1102 dprintk("revisit queued\n");
1104 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1105 spin_unlock(&xprt->xpt_lock);
1106 svc_xprt_enqueue(xprt);
1111 * Save the request off for later processing. The request buffer looks
1114 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1116 * This code can only handle requests that consist of an xprt-header
1119 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1121 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1122 struct svc_deferred_req *dr;
1124 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1125 return NULL; /* if more than a page, give up FIXME */
1126 if (rqstp->rq_deferred) {
1127 dr = rqstp->rq_deferred;
1128 rqstp->rq_deferred = NULL;
1132 /* FIXME maybe discard if size too large */
1133 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1134 dr = kmalloc(size, GFP_KERNEL);
1138 dr->handle.owner = rqstp->rq_server;
1139 dr->prot = rqstp->rq_prot;
1140 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1141 dr->addrlen = rqstp->rq_addrlen;
1142 dr->daddr = rqstp->rq_daddr;
1143 dr->argslen = rqstp->rq_arg.len >> 2;
1144 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1146 /* back up head to the start of the buffer and copy */
1147 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1148 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1151 svc_xprt_get(rqstp->rq_xprt);
1152 dr->xprt = rqstp->rq_xprt;
1153 set_bit(RQ_DROPME, &rqstp->rq_flags);
1155 dr->handle.revisit = svc_revisit;
1160 * recv data from a deferred request into an active one
1162 static int svc_deferred_recv(struct svc_rqst *rqstp)
1164 struct svc_deferred_req *dr = rqstp->rq_deferred;
1166 /* setup iov_base past transport header */
1167 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1168 /* The iov_len does not include the transport header bytes */
1169 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1170 rqstp->rq_arg.page_len = 0;
1171 /* The rq_arg.len includes the transport header bytes */
1172 rqstp->rq_arg.len = dr->argslen<<2;
1173 rqstp->rq_prot = dr->prot;
1174 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1175 rqstp->rq_addrlen = dr->addrlen;
1176 /* Save off transport header len in case we get deferred again */
1177 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1178 rqstp->rq_daddr = dr->daddr;
1179 rqstp->rq_respages = rqstp->rq_pages;
1180 return (dr->argslen<<2) - dr->xprt_hlen;
1184 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1186 struct svc_deferred_req *dr = NULL;
1188 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1190 spin_lock(&xprt->xpt_lock);
1191 if (!list_empty(&xprt->xpt_deferred)) {
1192 dr = list_entry(xprt->xpt_deferred.next,
1193 struct svc_deferred_req,
1195 list_del_init(&dr->handle.recent);
1197 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1198 spin_unlock(&xprt->xpt_lock);
1203 * svc_find_xprt - find an RPC transport instance
1204 * @serv: pointer to svc_serv to search
1205 * @xcl_name: C string containing transport's class name
1206 * @net: owner net pointer
1207 * @af: Address family of transport's local address
1208 * @port: transport's IP port number
1210 * Return the transport instance pointer for the endpoint accepting
1211 * connections/peer traffic from the specified transport class,
1212 * address family and port.
1214 * Specifying 0 for the address family or port is effectively a
1215 * wild-card, and will result in matching the first transport in the
1216 * service's list that has a matching class name.
1218 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1219 struct net *net, const sa_family_t af,
1220 const unsigned short port)
1222 struct svc_xprt *xprt;
1223 struct svc_xprt *found = NULL;
1225 /* Sanity check the args */
1226 if (serv == NULL || xcl_name == NULL)
1229 spin_lock_bh(&serv->sv_lock);
1230 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1231 if (xprt->xpt_net != net)
1233 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1235 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1237 if (port != 0 && port != svc_xprt_local_port(xprt))
1243 spin_unlock_bh(&serv->sv_lock);
1246 EXPORT_SYMBOL_GPL(svc_find_xprt);
1248 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1249 char *pos, int remaining)
1253 len = snprintf(pos, remaining, "%s %u\n",
1254 xprt->xpt_class->xcl_name,
1255 svc_xprt_local_port(xprt));
1256 if (len >= remaining)
1257 return -ENAMETOOLONG;
1262 * svc_xprt_names - format a buffer with a list of transport names
1263 * @serv: pointer to an RPC service
1264 * @buf: pointer to a buffer to be filled in
1265 * @buflen: length of buffer to be filled in
1267 * Fills in @buf with a string containing a list of transport names,
1268 * each name terminated with '\n'.
1270 * Returns positive length of the filled-in string on success; otherwise
1271 * a negative errno value is returned if an error occurs.
1273 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1275 struct svc_xprt *xprt;
1279 /* Sanity check args */
1283 spin_lock_bh(&serv->sv_lock);
1287 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1288 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1300 spin_unlock_bh(&serv->sv_lock);
1303 EXPORT_SYMBOL_GPL(svc_xprt_names);
1306 /*----------------------------------------------------------------------------*/
1308 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1310 unsigned int pidx = (unsigned int)*pos;
1311 struct svc_serv *serv = m->private;
1313 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1316 return SEQ_START_TOKEN;
1317 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1320 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1322 struct svc_pool *pool = p;
1323 struct svc_serv *serv = m->private;
1325 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1327 if (p == SEQ_START_TOKEN) {
1328 pool = &serv->sv_pools[0];
1330 unsigned int pidx = (pool - &serv->sv_pools[0]);
1331 if (pidx < serv->sv_nrpools-1)
1332 pool = &serv->sv_pools[pidx+1];
1340 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1344 static int svc_pool_stats_show(struct seq_file *m, void *p)
1346 struct svc_pool *pool = p;
1348 if (p == SEQ_START_TOKEN) {
1349 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1353 seq_printf(m, "%u %lu %lu %lu %lu\n",
1355 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1356 pool->sp_stats.sockets_queued,
1357 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1358 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1363 static const struct seq_operations svc_pool_stats_seq_ops = {
1364 .start = svc_pool_stats_start,
1365 .next = svc_pool_stats_next,
1366 .stop = svc_pool_stats_stop,
1367 .show = svc_pool_stats_show,
1370 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1374 err = seq_open(file, &svc_pool_stats_seq_ops);
1376 ((struct seq_file *) file->private_data)->private = serv;
1379 EXPORT_SYMBOL(svc_pool_stats_open);
1381 /*----------------------------------------------------------------------------*/