2 * linux/net/sunrpc/svc_xprt.c
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(struct timer_list *t);
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 * Format the transport list for printing
109 int svc_print_xprts(char *buf, int maxlen)
111 struct svc_xprt_class *xcl;
116 spin_lock(&svc_xprt_class_lock);
117 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
120 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
121 slen = strlen(tmpstr);
122 if (len + slen > maxlen)
127 spin_unlock(&svc_xprt_class_lock);
132 static void svc_xprt_free(struct kref *kref)
134 struct svc_xprt *xprt =
135 container_of(kref, struct svc_xprt, xpt_ref);
136 struct module *owner = xprt->xpt_class->xcl_owner;
137 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
138 svcauth_unix_info_release(xprt);
139 put_net(xprt->xpt_net);
140 /* See comment on corresponding get in xs_setup_bc_tcp(): */
141 if (xprt->xpt_bc_xprt)
142 xprt_put(xprt->xpt_bc_xprt);
143 if (xprt->xpt_bc_xps)
144 xprt_switch_put(xprt->xpt_bc_xps);
145 xprt->xpt_ops->xpo_free(xprt);
149 void svc_xprt_put(struct svc_xprt *xprt)
151 kref_put(&xprt->xpt_ref, svc_xprt_free);
153 EXPORT_SYMBOL_GPL(svc_xprt_put);
156 * Called by transport drivers to initialize the transport independent
157 * portion of the transport instance.
159 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
160 struct svc_xprt *xprt, struct svc_serv *serv)
162 memset(xprt, 0, sizeof(*xprt));
163 xprt->xpt_class = xcl;
164 xprt->xpt_ops = xcl->xcl_ops;
165 kref_init(&xprt->xpt_ref);
166 xprt->xpt_server = serv;
167 INIT_LIST_HEAD(&xprt->xpt_list);
168 INIT_LIST_HEAD(&xprt->xpt_ready);
169 INIT_LIST_HEAD(&xprt->xpt_deferred);
170 INIT_LIST_HEAD(&xprt->xpt_users);
171 mutex_init(&xprt->xpt_mutex);
172 spin_lock_init(&xprt->xpt_lock);
173 set_bit(XPT_BUSY, &xprt->xpt_flags);
174 xprt->xpt_net = get_net(net);
175 strcpy(xprt->xpt_remotebuf, "uninitialized");
177 EXPORT_SYMBOL_GPL(svc_xprt_init);
179 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
180 struct svc_serv *serv,
183 const unsigned short port,
186 struct sockaddr_in sin = {
187 .sin_family = AF_INET,
188 .sin_addr.s_addr = htonl(INADDR_ANY),
189 .sin_port = htons(port),
191 #if IS_ENABLED(CONFIG_IPV6)
192 struct sockaddr_in6 sin6 = {
193 .sin6_family = AF_INET6,
194 .sin6_addr = IN6ADDR_ANY_INIT,
195 .sin6_port = htons(port),
198 struct sockaddr *sap;
203 sap = (struct sockaddr *)&sin;
206 #if IS_ENABLED(CONFIG_IPV6)
208 sap = (struct sockaddr *)&sin6;
213 return ERR_PTR(-EAFNOSUPPORT);
216 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
220 * svc_xprt_received conditionally queues the transport for processing
221 * by another thread. The caller must hold the XPT_BUSY bit and must
222 * not thereafter touch transport data.
224 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
225 * insufficient) data.
227 static void svc_xprt_received(struct svc_xprt *xprt)
229 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
230 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
234 /* As soon as we clear busy, the xprt could be closed and
235 * 'put', so we need a reference to call svc_enqueue_xprt with:
238 smp_mb__before_atomic();
239 clear_bit(XPT_BUSY, &xprt->xpt_flags);
240 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
244 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
246 clear_bit(XPT_TEMP, &new->xpt_flags);
247 spin_lock_bh(&serv->sv_lock);
248 list_add(&new->xpt_list, &serv->sv_permsocks);
249 spin_unlock_bh(&serv->sv_lock);
250 svc_xprt_received(new);
253 static int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
254 struct net *net, const int family,
255 const unsigned short port, int flags)
257 struct svc_xprt_class *xcl;
259 spin_lock(&svc_xprt_class_lock);
260 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
261 struct svc_xprt *newxprt;
262 unsigned short newport;
264 if (strcmp(xprt_name, xcl->xcl_name))
267 if (!try_module_get(xcl->xcl_owner))
270 spin_unlock(&svc_xprt_class_lock);
271 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
272 if (IS_ERR(newxprt)) {
273 module_put(xcl->xcl_owner);
274 return PTR_ERR(newxprt);
276 svc_add_new_perm_xprt(serv, newxprt);
277 newport = svc_xprt_local_port(newxprt);
281 spin_unlock(&svc_xprt_class_lock);
282 /* This errno is exposed to user space. Provide a reasonable
283 * perror msg for a bad transport. */
284 return -EPROTONOSUPPORT;
287 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
288 struct net *net, const int family,
289 const unsigned short port, int flags)
293 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
294 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
295 if (err == -EPROTONOSUPPORT) {
296 request_module("svc%s", xprt_name);
297 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
300 dprintk("svc: transport %s not found, err %d\n",
304 EXPORT_SYMBOL_GPL(svc_create_xprt);
307 * Copy the local and remote xprt addresses to the rqstp structure
309 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
311 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
312 rqstp->rq_addrlen = xprt->xpt_remotelen;
315 * Destination address in request is needed for binding the
316 * source address in RPC replies/callbacks later.
318 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
319 rqstp->rq_daddrlen = xprt->xpt_locallen;
321 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
324 * svc_print_addr - Format rq_addr field for printing
325 * @rqstp: svc_rqst struct containing address to print
326 * @buf: target buffer for formatted address
327 * @len: length of target buffer
330 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
332 return __svc_print_addr(svc_addr(rqstp), buf, len);
334 EXPORT_SYMBOL_GPL(svc_print_addr);
336 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
338 unsigned int limit = svc_rpc_per_connection_limit;
339 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
341 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
344 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
346 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
347 if (!svc_xprt_slots_in_range(xprt))
349 atomic_inc(&xprt->xpt_nr_rqsts);
350 set_bit(RQ_DATA, &rqstp->rq_flags);
355 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
357 struct svc_xprt *xprt = rqstp->rq_xprt;
358 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
359 atomic_dec(&xprt->xpt_nr_rqsts);
360 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
361 svc_xprt_enqueue(xprt);
365 static bool svc_xprt_ready(struct svc_xprt *xprt)
367 unsigned long xpt_flags;
370 * If another cpu has recently updated xpt_flags,
371 * sk_sock->flags, xpt_reserved, or xpt_nr_rqsts, we need to
372 * know about it; otherwise it's possible that both that cpu and
373 * this one could call svc_xprt_enqueue() without either
374 * svc_xprt_enqueue() recognizing that the conditions below
375 * are satisfied, and we could stall indefinitely:
378 xpt_flags = READ_ONCE(xprt->xpt_flags);
380 if (xpt_flags & (BIT(XPT_CONN) | BIT(XPT_CLOSE)))
382 if (xpt_flags & (BIT(XPT_DATA) | BIT(XPT_DEFERRED))) {
383 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
384 svc_xprt_slots_in_range(xprt))
386 trace_svc_xprt_no_write_space(xprt);
392 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
394 struct svc_pool *pool;
395 struct svc_rqst *rqstp = NULL;
398 if (!svc_xprt_ready(xprt))
401 /* Mark transport as busy. It will remain in this state until
402 * the provider calls svc_xprt_received. We update XPT_BUSY
403 * atomically because it also guards against trying to enqueue
404 * the transport twice.
406 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
410 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
412 atomic_long_inc(&pool->sp_stats.packets);
414 spin_lock_bh(&pool->sp_lock);
415 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
416 pool->sp_stats.sockets_queued++;
417 spin_unlock_bh(&pool->sp_lock);
419 /* find a thread for this xprt */
421 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
422 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
424 atomic_long_inc(&pool->sp_stats.threads_woken);
425 rqstp->rq_qtime = ktime_get();
426 wake_up_process(rqstp->rq_task);
429 set_bit(SP_CONGESTED, &pool->sp_flags);
434 trace_svc_xprt_do_enqueue(xprt, rqstp);
436 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
439 * Queue up a transport with data pending. If there are idle nfsd
440 * processes, wake 'em up.
443 void svc_xprt_enqueue(struct svc_xprt *xprt)
445 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
447 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
449 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
452 * Dequeue the first transport, if there is one.
454 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
456 struct svc_xprt *xprt = NULL;
458 if (list_empty(&pool->sp_sockets))
461 spin_lock_bh(&pool->sp_lock);
462 if (likely(!list_empty(&pool->sp_sockets))) {
463 xprt = list_first_entry(&pool->sp_sockets,
464 struct svc_xprt, xpt_ready);
465 list_del_init(&xprt->xpt_ready);
468 spin_unlock_bh(&pool->sp_lock);
474 * svc_reserve - change the space reserved for the reply to a request.
475 * @rqstp: The request in question
476 * @space: new max space to reserve
478 * Each request reserves some space on the output queue of the transport
479 * to make sure the reply fits. This function reduces that reserved
480 * space to be the amount of space used already, plus @space.
483 void svc_reserve(struct svc_rqst *rqstp, int space)
485 struct svc_xprt *xprt = rqstp->rq_xprt;
487 space += rqstp->rq_res.head[0].iov_len;
489 if (xprt && space < rqstp->rq_reserved) {
490 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
491 rqstp->rq_reserved = space;
492 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
493 svc_xprt_enqueue(xprt);
496 EXPORT_SYMBOL_GPL(svc_reserve);
498 static void svc_xprt_release(struct svc_rqst *rqstp)
500 struct svc_xprt *xprt = rqstp->rq_xprt;
502 xprt->xpt_ops->xpo_release_rqst(rqstp);
504 kfree(rqstp->rq_deferred);
505 rqstp->rq_deferred = NULL;
507 svc_free_res_pages(rqstp);
508 rqstp->rq_res.page_len = 0;
509 rqstp->rq_res.page_base = 0;
511 /* Reset response buffer and release
513 * But first, check that enough space was reserved
514 * for the reply, otherwise we have a bug!
516 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
517 printk(KERN_ERR "RPC request reserved %d but used %d\n",
521 rqstp->rq_res.head[0].iov_len = 0;
522 svc_reserve(rqstp, 0);
523 svc_xprt_release_slot(rqstp);
524 rqstp->rq_xprt = NULL;
529 * Some svc_serv's will have occasional work to do, even when a xprt is not
530 * waiting to be serviced. This function is there to "kick" a task in one of
531 * those services so that it can wake up and do that work. Note that we only
532 * bother with pool 0 as we don't need to wake up more than one thread for
535 void svc_wake_up(struct svc_serv *serv)
537 struct svc_rqst *rqstp;
538 struct svc_pool *pool;
540 pool = &serv->sv_pools[0];
543 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
544 /* skip any that aren't queued */
545 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
548 wake_up_process(rqstp->rq_task);
549 trace_svc_wake_up(rqstp->rq_task->pid);
554 /* No free entries available */
555 set_bit(SP_TASK_PENDING, &pool->sp_flags);
557 trace_svc_wake_up(0);
559 EXPORT_SYMBOL_GPL(svc_wake_up);
561 int svc_port_is_privileged(struct sockaddr *sin)
563 switch (sin->sa_family) {
565 return ntohs(((struct sockaddr_in *)sin)->sin_port)
568 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
576 * Make sure that we don't have too many active connections. If we have,
577 * something must be dropped. It's not clear what will happen if we allow
578 * "too many" connections, but when dealing with network-facing software,
579 * we have to code defensively. Here we do that by imposing hard limits.
581 * There's no point in trying to do random drop here for DoS
582 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
583 * attacker can easily beat that.
585 * The only somewhat efficient mechanism would be if drop old
586 * connections from the same IP first. But right now we don't even
587 * record the client IP in svc_sock.
589 * single-threaded services that expect a lot of clients will probably
590 * need to set sv_maxconn to override the default value which is based
591 * on the number of threads
593 static void svc_check_conn_limits(struct svc_serv *serv)
595 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
596 (serv->sv_nrthreads+3) * 20;
598 if (serv->sv_tmpcnt > limit) {
599 struct svc_xprt *xprt = NULL;
600 spin_lock_bh(&serv->sv_lock);
601 if (!list_empty(&serv->sv_tempsocks)) {
602 /* Try to help the admin */
603 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
604 serv->sv_name, serv->sv_maxconn ?
605 "max number of connections" :
606 "number of threads");
608 * Always select the oldest connection. It's not fair,
611 xprt = list_entry(serv->sv_tempsocks.prev,
614 set_bit(XPT_CLOSE, &xprt->xpt_flags);
617 spin_unlock_bh(&serv->sv_lock);
620 svc_xprt_enqueue(xprt);
626 static int svc_alloc_arg(struct svc_rqst *rqstp)
628 struct svc_serv *serv = rqstp->rq_server;
633 /* now allocate needed pages. If we get a failure, sleep briefly */
634 pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
635 if (pages > RPCSVC_MAXPAGES) {
636 pr_warn_once("svc: warning: pages=%u > RPCSVC_MAXPAGES=%lu\n",
637 pages, RPCSVC_MAXPAGES);
638 /* use as many pages as possible */
639 pages = RPCSVC_MAXPAGES;
641 for (i = 0; i < pages ; i++)
642 while (rqstp->rq_pages[i] == NULL) {
643 struct page *p = alloc_page(GFP_KERNEL);
645 set_current_state(TASK_INTERRUPTIBLE);
646 if (signalled() || kthread_should_stop()) {
647 set_current_state(TASK_RUNNING);
650 schedule_timeout(msecs_to_jiffies(500));
652 rqstp->rq_pages[i] = p;
654 rqstp->rq_page_end = &rqstp->rq_pages[i];
655 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
657 /* Make arg->head point to first page and arg->pages point to rest */
658 arg = &rqstp->rq_arg;
659 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
660 arg->head[0].iov_len = PAGE_SIZE;
661 arg->pages = rqstp->rq_pages + 1;
663 /* save at least one page for response */
664 arg->page_len = (pages-2)*PAGE_SIZE;
665 arg->len = (pages-1)*PAGE_SIZE;
666 arg->tail[0].iov_len = 0;
671 rqst_should_sleep(struct svc_rqst *rqstp)
673 struct svc_pool *pool = rqstp->rq_pool;
675 /* did someone call svc_wake_up? */
676 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
679 /* was a socket queued? */
680 if (!list_empty(&pool->sp_sockets))
683 /* are we shutting down? */
684 if (signalled() || kthread_should_stop())
687 /* are we freezing? */
688 if (freezing(current))
694 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
696 struct svc_pool *pool = rqstp->rq_pool;
699 /* rq_xprt should be clear on entry */
700 WARN_ON_ONCE(rqstp->rq_xprt);
702 rqstp->rq_xprt = svc_xprt_dequeue(pool);
707 * We have to be able to interrupt this wait
708 * to bring down the daemons ...
710 set_current_state(TASK_INTERRUPTIBLE);
711 smp_mb__before_atomic();
712 clear_bit(SP_CONGESTED, &pool->sp_flags);
713 clear_bit(RQ_BUSY, &rqstp->rq_flags);
714 smp_mb__after_atomic();
716 if (likely(rqst_should_sleep(rqstp)))
717 time_left = schedule_timeout(timeout);
719 __set_current_state(TASK_RUNNING);
723 set_bit(RQ_BUSY, &rqstp->rq_flags);
724 smp_mb__after_atomic();
725 rqstp->rq_xprt = svc_xprt_dequeue(pool);
730 atomic_long_inc(&pool->sp_stats.threads_timedout);
732 if (signalled() || kthread_should_stop())
733 return ERR_PTR(-EINTR);
734 return ERR_PTR(-EAGAIN);
736 /* Normally we will wait up to 5 seconds for any required
737 * cache information to be provided.
739 if (!test_bit(SP_CONGESTED, &pool->sp_flags))
740 rqstp->rq_chandle.thread_wait = 5*HZ;
742 rqstp->rq_chandle.thread_wait = 1*HZ;
743 trace_svc_xprt_dequeue(rqstp);
744 return rqstp->rq_xprt;
747 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
749 spin_lock_bh(&serv->sv_lock);
750 set_bit(XPT_TEMP, &newxpt->xpt_flags);
751 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
753 if (serv->sv_temptimer.function == NULL) {
754 /* setup timer to age temp transports */
755 serv->sv_temptimer.function = svc_age_temp_xprts;
756 mod_timer(&serv->sv_temptimer,
757 jiffies + svc_conn_age_period * HZ);
759 spin_unlock_bh(&serv->sv_lock);
760 svc_xprt_received(newxpt);
763 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
765 struct svc_serv *serv = rqstp->rq_server;
768 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
769 dprintk("svc_recv: found XPT_CLOSE\n");
770 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
771 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
772 svc_delete_xprt(xprt);
773 /* Leave XPT_BUSY set on the dead xprt: */
776 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
777 struct svc_xprt *newxpt;
779 * We know this module_get will succeed because the
780 * listener holds a reference too
782 __module_get(xprt->xpt_class->xcl_owner);
783 svc_check_conn_limits(xprt->xpt_server);
784 newxpt = xprt->xpt_ops->xpo_accept(xprt);
786 svc_add_new_temp_xprt(serv, newxpt);
788 module_put(xprt->xpt_class->xcl_owner);
789 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
790 /* XPT_DATA|XPT_DEFERRED case: */
791 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
792 rqstp, rqstp->rq_pool->sp_id, xprt,
793 kref_read(&xprt->xpt_ref));
794 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
795 if (rqstp->rq_deferred)
796 len = svc_deferred_recv(rqstp);
798 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
799 rqstp->rq_stime = ktime_get();
800 rqstp->rq_reserved = serv->sv_max_mesg;
801 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
803 /* clear XPT_BUSY: */
804 svc_xprt_received(xprt);
806 trace_svc_handle_xprt(xprt, len);
811 * Receive the next request on any transport. This code is carefully
812 * organised not to touch any cachelines in the shared svc_serv
813 * structure, only cachelines in the local svc_pool.
815 int svc_recv(struct svc_rqst *rqstp, long timeout)
817 struct svc_xprt *xprt = NULL;
818 struct svc_serv *serv = rqstp->rq_server;
821 dprintk("svc: server %p waiting for data (to = %ld)\n",
826 "svc_recv: service %p, transport not NULL!\n",
829 err = svc_alloc_arg(rqstp);
836 if (signalled() || kthread_should_stop())
839 xprt = svc_get_next_xprt(rqstp, timeout);
845 len = svc_handle_xprt(rqstp, xprt);
847 /* No data, incomplete (TCP) read, or accept() */
852 clear_bit(XPT_OLD, &xprt->xpt_flags);
854 xprt->xpt_ops->xpo_secure_port(rqstp);
855 rqstp->rq_chandle.defer = svc_defer;
856 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
859 serv->sv_stats->netcnt++;
860 trace_svc_recv(rqstp, len);
863 rqstp->rq_res.len = 0;
864 svc_xprt_release(rqstp);
868 EXPORT_SYMBOL_GPL(svc_recv);
873 void svc_drop(struct svc_rqst *rqstp)
875 trace_svc_drop(rqstp);
876 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
877 svc_xprt_release(rqstp);
879 EXPORT_SYMBOL_GPL(svc_drop);
882 * Return reply to client.
884 int svc_send(struct svc_rqst *rqstp)
886 struct svc_xprt *xprt;
890 xprt = rqstp->rq_xprt;
894 /* release the receive skb before sending the reply */
895 xprt->xpt_ops->xpo_release_rqst(rqstp);
897 /* calculate over-all length */
899 xb->len = xb->head[0].iov_len +
903 /* Grab mutex to serialize outgoing data. */
904 mutex_lock(&xprt->xpt_mutex);
905 trace_svc_stats_latency(rqstp);
906 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
907 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
910 len = xprt->xpt_ops->xpo_sendto(rqstp);
911 mutex_unlock(&xprt->xpt_mutex);
912 trace_svc_send(rqstp, len);
913 svc_xprt_release(rqstp);
915 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
922 * Timer function to close old temporary transports, using
923 * a mark-and-sweep algorithm.
925 static void svc_age_temp_xprts(struct timer_list *t)
927 struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
928 struct svc_xprt *xprt;
929 struct list_head *le, *next;
931 dprintk("svc_age_temp_xprts\n");
933 if (!spin_trylock_bh(&serv->sv_lock)) {
934 /* busy, try again 1 sec later */
935 dprintk("svc_age_temp_xprts: busy\n");
936 mod_timer(&serv->sv_temptimer, jiffies + HZ);
940 list_for_each_safe(le, next, &serv->sv_tempsocks) {
941 xprt = list_entry(le, struct svc_xprt, xpt_list);
943 /* First time through, just mark it OLD. Second time
944 * through, close it. */
945 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
947 if (kref_read(&xprt->xpt_ref) > 1 ||
948 test_bit(XPT_BUSY, &xprt->xpt_flags))
951 set_bit(XPT_CLOSE, &xprt->xpt_flags);
952 dprintk("queuing xprt %p for closing\n", xprt);
954 /* a thread will dequeue and close it soon */
955 svc_xprt_enqueue(xprt);
957 spin_unlock_bh(&serv->sv_lock);
959 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
962 /* Close temporary transports whose xpt_local matches server_addr immediately
963 * instead of waiting for them to be picked up by the timer.
965 * This is meant to be called from a notifier_block that runs when an ip
966 * address is deleted.
968 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
970 struct svc_xprt *xprt;
971 struct list_head *le, *next;
972 LIST_HEAD(to_be_closed);
974 spin_lock_bh(&serv->sv_lock);
975 list_for_each_safe(le, next, &serv->sv_tempsocks) {
976 xprt = list_entry(le, struct svc_xprt, xpt_list);
977 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
979 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
980 list_move(le, &to_be_closed);
983 spin_unlock_bh(&serv->sv_lock);
985 while (!list_empty(&to_be_closed)) {
986 le = to_be_closed.next;
988 xprt = list_entry(le, struct svc_xprt, xpt_list);
989 set_bit(XPT_CLOSE, &xprt->xpt_flags);
990 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
991 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
993 svc_xprt_enqueue(xprt);
996 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
998 static void call_xpt_users(struct svc_xprt *xprt)
1000 struct svc_xpt_user *u;
1002 spin_lock(&xprt->xpt_lock);
1003 while (!list_empty(&xprt->xpt_users)) {
1004 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1005 list_del_init(&u->list);
1008 spin_unlock(&xprt->xpt_lock);
1012 * Remove a dead transport
1014 static void svc_delete_xprt(struct svc_xprt *xprt)
1016 struct svc_serv *serv = xprt->xpt_server;
1017 struct svc_deferred_req *dr;
1019 /* Only do this once */
1020 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1023 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1024 xprt->xpt_ops->xpo_detach(xprt);
1026 spin_lock_bh(&serv->sv_lock);
1027 list_del_init(&xprt->xpt_list);
1028 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1029 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1031 spin_unlock_bh(&serv->sv_lock);
1033 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1036 call_xpt_users(xprt);
1040 void svc_close_xprt(struct svc_xprt *xprt)
1042 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1043 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1044 /* someone else will have to effect the close */
1047 * We expect svc_close_xprt() to work even when no threads are
1048 * running (e.g., while configuring the server before starting
1049 * any threads), so if the transport isn't busy, we delete
1052 svc_delete_xprt(xprt);
1054 EXPORT_SYMBOL_GPL(svc_close_xprt);
1056 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1058 struct svc_xprt *xprt;
1061 spin_lock(&serv->sv_lock);
1062 list_for_each_entry(xprt, xprt_list, xpt_list) {
1063 if (xprt->xpt_net != net)
1066 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1067 svc_xprt_enqueue(xprt);
1069 spin_unlock(&serv->sv_lock);
1073 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1075 struct svc_pool *pool;
1076 struct svc_xprt *xprt;
1077 struct svc_xprt *tmp;
1080 for (i = 0; i < serv->sv_nrpools; i++) {
1081 pool = &serv->sv_pools[i];
1083 spin_lock_bh(&pool->sp_lock);
1084 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1085 if (xprt->xpt_net != net)
1087 list_del_init(&xprt->xpt_ready);
1088 spin_unlock_bh(&pool->sp_lock);
1091 spin_unlock_bh(&pool->sp_lock);
1096 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1098 struct svc_xprt *xprt;
1100 while ((xprt = svc_dequeue_net(serv, net))) {
1101 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1102 svc_delete_xprt(xprt);
1107 * Server threads may still be running (especially in the case where the
1108 * service is still running in other network namespaces).
1110 * So we shut down sockets the same way we would on a running server, by
1111 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1112 * the close. In the case there are no such other threads,
1113 * threads running, svc_clean_up_xprts() does a simple version of a
1114 * server's main event loop, and in the case where there are other
1115 * threads, we may need to wait a little while and then check again to
1116 * see if they're done.
1118 void svc_close_net(struct svc_serv *serv, struct net *net)
1122 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1123 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1125 svc_clean_up_xprts(serv, net);
1131 * Handle defer and revisit of requests
1134 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1136 struct svc_deferred_req *dr =
1137 container_of(dreq, struct svc_deferred_req, handle);
1138 struct svc_xprt *xprt = dr->xprt;
1140 spin_lock(&xprt->xpt_lock);
1141 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1142 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1143 spin_unlock(&xprt->xpt_lock);
1144 dprintk("revisit canceled\n");
1146 trace_svc_drop_deferred(dr);
1150 dprintk("revisit queued\n");
1152 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1153 spin_unlock(&xprt->xpt_lock);
1154 svc_xprt_enqueue(xprt);
1159 * Save the request off for later processing. The request buffer looks
1162 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1164 * This code can only handle requests that consist of an xprt-header
1167 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1169 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1170 struct svc_deferred_req *dr;
1172 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1173 return NULL; /* if more than a page, give up FIXME */
1174 if (rqstp->rq_deferred) {
1175 dr = rqstp->rq_deferred;
1176 rqstp->rq_deferred = NULL;
1180 /* FIXME maybe discard if size too large */
1181 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1182 dr = kmalloc(size, GFP_KERNEL);
1186 dr->handle.owner = rqstp->rq_server;
1187 dr->prot = rqstp->rq_prot;
1188 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1189 dr->addrlen = rqstp->rq_addrlen;
1190 dr->daddr = rqstp->rq_daddr;
1191 dr->argslen = rqstp->rq_arg.len >> 2;
1192 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1194 /* back up head to the start of the buffer and copy */
1195 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1196 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1199 svc_xprt_get(rqstp->rq_xprt);
1200 dr->xprt = rqstp->rq_xprt;
1201 set_bit(RQ_DROPME, &rqstp->rq_flags);
1203 dr->handle.revisit = svc_revisit;
1204 trace_svc_defer(rqstp);
1209 * recv data from a deferred request into an active one
1211 static int svc_deferred_recv(struct svc_rqst *rqstp)
1213 struct svc_deferred_req *dr = rqstp->rq_deferred;
1215 /* setup iov_base past transport header */
1216 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1217 /* The iov_len does not include the transport header bytes */
1218 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1219 rqstp->rq_arg.page_len = 0;
1220 /* The rq_arg.len includes the transport header bytes */
1221 rqstp->rq_arg.len = dr->argslen<<2;
1222 rqstp->rq_prot = dr->prot;
1223 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1224 rqstp->rq_addrlen = dr->addrlen;
1225 /* Save off transport header len in case we get deferred again */
1226 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1227 rqstp->rq_daddr = dr->daddr;
1228 rqstp->rq_respages = rqstp->rq_pages;
1229 return (dr->argslen<<2) - dr->xprt_hlen;
1233 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1235 struct svc_deferred_req *dr = NULL;
1237 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1239 spin_lock(&xprt->xpt_lock);
1240 if (!list_empty(&xprt->xpt_deferred)) {
1241 dr = list_entry(xprt->xpt_deferred.next,
1242 struct svc_deferred_req,
1244 list_del_init(&dr->handle.recent);
1245 trace_svc_revisit_deferred(dr);
1247 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1248 spin_unlock(&xprt->xpt_lock);
1253 * svc_find_xprt - find an RPC transport instance
1254 * @serv: pointer to svc_serv to search
1255 * @xcl_name: C string containing transport's class name
1256 * @net: owner net pointer
1257 * @af: Address family of transport's local address
1258 * @port: transport's IP port number
1260 * Return the transport instance pointer for the endpoint accepting
1261 * connections/peer traffic from the specified transport class,
1262 * address family and port.
1264 * Specifying 0 for the address family or port is effectively a
1265 * wild-card, and will result in matching the first transport in the
1266 * service's list that has a matching class name.
1268 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1269 struct net *net, const sa_family_t af,
1270 const unsigned short port)
1272 struct svc_xprt *xprt;
1273 struct svc_xprt *found = NULL;
1275 /* Sanity check the args */
1276 if (serv == NULL || xcl_name == NULL)
1279 spin_lock_bh(&serv->sv_lock);
1280 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1281 if (xprt->xpt_net != net)
1283 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1285 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1287 if (port != 0 && port != svc_xprt_local_port(xprt))
1293 spin_unlock_bh(&serv->sv_lock);
1296 EXPORT_SYMBOL_GPL(svc_find_xprt);
1298 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1299 char *pos, int remaining)
1303 len = snprintf(pos, remaining, "%s %u\n",
1304 xprt->xpt_class->xcl_name,
1305 svc_xprt_local_port(xprt));
1306 if (len >= remaining)
1307 return -ENAMETOOLONG;
1312 * svc_xprt_names - format a buffer with a list of transport names
1313 * @serv: pointer to an RPC service
1314 * @buf: pointer to a buffer to be filled in
1315 * @buflen: length of buffer to be filled in
1317 * Fills in @buf with a string containing a list of transport names,
1318 * each name terminated with '\n'.
1320 * Returns positive length of the filled-in string on success; otherwise
1321 * a negative errno value is returned if an error occurs.
1323 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1325 struct svc_xprt *xprt;
1329 /* Sanity check args */
1333 spin_lock_bh(&serv->sv_lock);
1337 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1338 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1350 spin_unlock_bh(&serv->sv_lock);
1353 EXPORT_SYMBOL_GPL(svc_xprt_names);
1356 /*----------------------------------------------------------------------------*/
1358 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1360 unsigned int pidx = (unsigned int)*pos;
1361 struct svc_serv *serv = m->private;
1363 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1366 return SEQ_START_TOKEN;
1367 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1370 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1372 struct svc_pool *pool = p;
1373 struct svc_serv *serv = m->private;
1375 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1377 if (p == SEQ_START_TOKEN) {
1378 pool = &serv->sv_pools[0];
1380 unsigned int pidx = (pool - &serv->sv_pools[0]);
1381 if (pidx < serv->sv_nrpools-1)
1382 pool = &serv->sv_pools[pidx+1];
1390 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1394 static int svc_pool_stats_show(struct seq_file *m, void *p)
1396 struct svc_pool *pool = p;
1398 if (p == SEQ_START_TOKEN) {
1399 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1403 seq_printf(m, "%u %lu %lu %lu %lu\n",
1405 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1406 pool->sp_stats.sockets_queued,
1407 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1408 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1413 static const struct seq_operations svc_pool_stats_seq_ops = {
1414 .start = svc_pool_stats_start,
1415 .next = svc_pool_stats_next,
1416 .stop = svc_pool_stats_stop,
1417 .show = svc_pool_stats_show,
1420 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1424 err = seq_open(file, &svc_pool_stats_seq_ops);
1426 ((struct seq_file *) file->private_data)->private = serv;
1429 EXPORT_SYMBOL(svc_pool_stats_open);
1431 /*----------------------------------------------------------------------------*/