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 * 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 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
175 xprt->xpt_net = get_net(net);
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 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 svc_xprt_enqueue(xprt);
364 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
366 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
368 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED))) {
369 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
370 svc_xprt_slots_in_range(xprt))
372 trace_svc_xprt_no_write_space(xprt);
378 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
380 struct svc_pool *pool;
381 struct svc_rqst *rqstp = NULL;
385 if (!svc_xprt_has_something_to_do(xprt))
388 /* Mark transport as busy. It will remain in this state until
389 * the provider calls svc_xprt_received. We update XPT_BUSY
390 * atomically because it also guards against trying to enqueue
391 * the transport twice.
393 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
394 /* Don't enqueue transport while already enqueued */
395 dprintk("svc: transport %p busy, not enqueued\n", xprt);
400 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
402 atomic_long_inc(&pool->sp_stats.packets);
405 /* find a thread for this xprt */
407 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
408 /* Do a lockless check first */
409 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
413 * Once the xprt has been queued, it can only be dequeued by
414 * the task that intends to service it. All we can do at that
415 * point is to try to wake this thread back up so that it can
419 spin_lock_bh(&rqstp->rq_lock);
420 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags)) {
421 /* already busy, move on... */
422 spin_unlock_bh(&rqstp->rq_lock);
426 /* this one will do */
427 rqstp->rq_xprt = xprt;
429 spin_unlock_bh(&rqstp->rq_lock);
433 atomic_long_inc(&pool->sp_stats.threads_woken);
434 wake_up_process(rqstp->rq_task);
441 * We didn't find an idle thread to use, so we need to queue the xprt.
442 * Do so and then search again. If we find one, we can't hook this one
443 * up to it directly but we can wake the thread up in the hopes that it
444 * will pick it up once it searches for a xprt to service.
448 dprintk("svc: transport %p put into queue\n", xprt);
449 spin_lock_bh(&pool->sp_lock);
450 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
451 pool->sp_stats.sockets_queued++;
452 spin_unlock_bh(&pool->sp_lock);
458 trace_svc_xprt_do_enqueue(xprt, rqstp);
460 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
463 * Queue up a transport with data pending. If there are idle nfsd
464 * processes, wake 'em up.
467 void svc_xprt_enqueue(struct svc_xprt *xprt)
469 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
471 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
473 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
476 * Dequeue the first transport, if there is one.
478 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
480 struct svc_xprt *xprt = NULL;
482 if (list_empty(&pool->sp_sockets))
485 spin_lock_bh(&pool->sp_lock);
486 if (likely(!list_empty(&pool->sp_sockets))) {
487 xprt = list_first_entry(&pool->sp_sockets,
488 struct svc_xprt, xpt_ready);
489 list_del_init(&xprt->xpt_ready);
492 dprintk("svc: transport %p dequeued, inuse=%d\n",
493 xprt, atomic_read(&xprt->xpt_ref.refcount));
495 spin_unlock_bh(&pool->sp_lock);
497 trace_svc_xprt_dequeue(xprt);
502 * svc_reserve - change the space reserved for the reply to a request.
503 * @rqstp: The request in question
504 * @space: new max space to reserve
506 * Each request reserves some space on the output queue of the transport
507 * to make sure the reply fits. This function reduces that reserved
508 * space to be the amount of space used already, plus @space.
511 void svc_reserve(struct svc_rqst *rqstp, int space)
513 space += rqstp->rq_res.head[0].iov_len;
515 if (space < rqstp->rq_reserved) {
516 struct svc_xprt *xprt = rqstp->rq_xprt;
517 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
518 rqstp->rq_reserved = space;
520 svc_xprt_enqueue(xprt);
523 EXPORT_SYMBOL_GPL(svc_reserve);
525 static void svc_xprt_release(struct svc_rqst *rqstp)
527 struct svc_xprt *xprt = rqstp->rq_xprt;
529 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
531 kfree(rqstp->rq_deferred);
532 rqstp->rq_deferred = NULL;
534 svc_free_res_pages(rqstp);
535 rqstp->rq_res.page_len = 0;
536 rqstp->rq_res.page_base = 0;
538 /* Reset response buffer and release
540 * But first, check that enough space was reserved
541 * for the reply, otherwise we have a bug!
543 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
544 printk(KERN_ERR "RPC request reserved %d but used %d\n",
548 rqstp->rq_res.head[0].iov_len = 0;
549 svc_reserve(rqstp, 0);
550 svc_xprt_release_slot(rqstp);
551 rqstp->rq_xprt = NULL;
556 * Some svc_serv's will have occasional work to do, even when a xprt is not
557 * waiting to be serviced. This function is there to "kick" a task in one of
558 * those services so that it can wake up and do that work. Note that we only
559 * bother with pool 0 as we don't need to wake up more than one thread for
562 void svc_wake_up(struct svc_serv *serv)
564 struct svc_rqst *rqstp;
565 struct svc_pool *pool;
567 pool = &serv->sv_pools[0];
570 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
571 /* skip any that aren't queued */
572 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
575 dprintk("svc: daemon %p woken up.\n", rqstp);
576 wake_up_process(rqstp->rq_task);
577 trace_svc_wake_up(rqstp->rq_task->pid);
582 /* No free entries available */
583 set_bit(SP_TASK_PENDING, &pool->sp_flags);
585 trace_svc_wake_up(0);
587 EXPORT_SYMBOL_GPL(svc_wake_up);
589 int svc_port_is_privileged(struct sockaddr *sin)
591 switch (sin->sa_family) {
593 return ntohs(((struct sockaddr_in *)sin)->sin_port)
596 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
604 * Make sure that we don't have too many active connections. If we have,
605 * something must be dropped. It's not clear what will happen if we allow
606 * "too many" connections, but when dealing with network-facing software,
607 * we have to code defensively. Here we do that by imposing hard limits.
609 * There's no point in trying to do random drop here for DoS
610 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
611 * attacker can easily beat that.
613 * The only somewhat efficient mechanism would be if drop old
614 * connections from the same IP first. But right now we don't even
615 * record the client IP in svc_sock.
617 * single-threaded services that expect a lot of clients will probably
618 * need to set sv_maxconn to override the default value which is based
619 * on the number of threads
621 static void svc_check_conn_limits(struct svc_serv *serv)
623 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
624 (serv->sv_nrthreads+3) * 20;
626 if (serv->sv_tmpcnt > limit) {
627 struct svc_xprt *xprt = NULL;
628 spin_lock_bh(&serv->sv_lock);
629 if (!list_empty(&serv->sv_tempsocks)) {
630 /* Try to help the admin */
631 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
632 serv->sv_name, serv->sv_maxconn ?
633 "max number of connections" :
634 "number of threads");
636 * Always select the oldest connection. It's not fair,
639 xprt = list_entry(serv->sv_tempsocks.prev,
642 set_bit(XPT_CLOSE, &xprt->xpt_flags);
645 spin_unlock_bh(&serv->sv_lock);
648 svc_xprt_enqueue(xprt);
654 static int svc_alloc_arg(struct svc_rqst *rqstp)
656 struct svc_serv *serv = rqstp->rq_server;
661 /* now allocate needed pages. If we get a failure, sleep briefly */
662 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
663 WARN_ON_ONCE(pages >= RPCSVC_MAXPAGES);
664 if (pages >= RPCSVC_MAXPAGES)
665 /* use as many pages as possible */
666 pages = RPCSVC_MAXPAGES - 1;
667 for (i = 0; i < pages ; i++)
668 while (rqstp->rq_pages[i] == NULL) {
669 struct page *p = alloc_page(GFP_KERNEL);
671 set_current_state(TASK_INTERRUPTIBLE);
672 if (signalled() || kthread_should_stop()) {
673 set_current_state(TASK_RUNNING);
676 schedule_timeout(msecs_to_jiffies(500));
678 rqstp->rq_pages[i] = p;
680 rqstp->rq_page_end = &rqstp->rq_pages[i];
681 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
683 /* Make arg->head point to first page and arg->pages point to rest */
684 arg = &rqstp->rq_arg;
685 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
686 arg->head[0].iov_len = PAGE_SIZE;
687 arg->pages = rqstp->rq_pages + 1;
689 /* save at least one page for response */
690 arg->page_len = (pages-2)*PAGE_SIZE;
691 arg->len = (pages-1)*PAGE_SIZE;
692 arg->tail[0].iov_len = 0;
697 rqst_should_sleep(struct svc_rqst *rqstp)
699 struct svc_pool *pool = rqstp->rq_pool;
701 /* did someone call svc_wake_up? */
702 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
705 /* was a socket queued? */
706 if (!list_empty(&pool->sp_sockets))
709 /* are we shutting down? */
710 if (signalled() || kthread_should_stop())
713 /* are we freezing? */
714 if (freezing(current))
720 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
722 struct svc_xprt *xprt;
723 struct svc_pool *pool = rqstp->rq_pool;
726 /* rq_xprt should be clear on entry */
727 WARN_ON_ONCE(rqstp->rq_xprt);
729 /* Normally we will wait up to 5 seconds for any required
730 * cache information to be provided.
732 rqstp->rq_chandle.thread_wait = 5*HZ;
734 xprt = svc_xprt_dequeue(pool);
736 rqstp->rq_xprt = xprt;
738 /* As there is a shortage of threads and this request
739 * had to be queued, don't allow the thread to wait so
740 * long for cache updates.
742 rqstp->rq_chandle.thread_wait = 1*HZ;
743 clear_bit(SP_TASK_PENDING, &pool->sp_flags);
748 * We have to be able to interrupt this wait
749 * to bring down the daemons ...
751 set_current_state(TASK_INTERRUPTIBLE);
752 clear_bit(RQ_BUSY, &rqstp->rq_flags);
755 if (likely(rqst_should_sleep(rqstp)))
756 time_left = schedule_timeout(timeout);
758 __set_current_state(TASK_RUNNING);
762 spin_lock_bh(&rqstp->rq_lock);
763 set_bit(RQ_BUSY, &rqstp->rq_flags);
764 spin_unlock_bh(&rqstp->rq_lock);
766 xprt = rqstp->rq_xprt;
771 atomic_long_inc(&pool->sp_stats.threads_timedout);
773 if (signalled() || kthread_should_stop())
774 return ERR_PTR(-EINTR);
775 return ERR_PTR(-EAGAIN);
778 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
780 spin_lock_bh(&serv->sv_lock);
781 set_bit(XPT_TEMP, &newxpt->xpt_flags);
782 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
784 if (serv->sv_temptimer.function == NULL) {
785 /* setup timer to age temp transports */
786 setup_timer(&serv->sv_temptimer, svc_age_temp_xprts,
787 (unsigned long)serv);
788 mod_timer(&serv->sv_temptimer,
789 jiffies + svc_conn_age_period * HZ);
791 spin_unlock_bh(&serv->sv_lock);
792 svc_xprt_received(newxpt);
795 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
797 struct svc_serv *serv = rqstp->rq_server;
800 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
801 dprintk("svc_recv: found XPT_CLOSE\n");
802 svc_delete_xprt(xprt);
803 /* Leave XPT_BUSY set on the dead xprt: */
806 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
807 struct svc_xprt *newxpt;
809 * We know this module_get will succeed because the
810 * listener holds a reference too
812 __module_get(xprt->xpt_class->xcl_owner);
813 svc_check_conn_limits(xprt->xpt_server);
814 newxpt = xprt->xpt_ops->xpo_accept(xprt);
816 svc_add_new_temp_xprt(serv, newxpt);
818 module_put(xprt->xpt_class->xcl_owner);
819 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
820 /* XPT_DATA|XPT_DEFERRED case: */
821 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
822 rqstp, rqstp->rq_pool->sp_id, xprt,
823 atomic_read(&xprt->xpt_ref.refcount));
824 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
825 if (rqstp->rq_deferred)
826 len = svc_deferred_recv(rqstp);
828 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
829 dprintk("svc: got len=%d\n", len);
830 rqstp->rq_reserved = serv->sv_max_mesg;
831 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
833 /* clear XPT_BUSY: */
834 svc_xprt_received(xprt);
836 trace_svc_handle_xprt(xprt, len);
841 * Receive the next request on any transport. This code is carefully
842 * organised not to touch any cachelines in the shared svc_serv
843 * structure, only cachelines in the local svc_pool.
845 int svc_recv(struct svc_rqst *rqstp, long timeout)
847 struct svc_xprt *xprt = NULL;
848 struct svc_serv *serv = rqstp->rq_server;
851 dprintk("svc: server %p waiting for data (to = %ld)\n",
856 "svc_recv: service %p, transport not NULL!\n",
859 err = svc_alloc_arg(rqstp);
866 if (signalled() || kthread_should_stop())
869 xprt = svc_get_next_xprt(rqstp, timeout);
875 len = svc_handle_xprt(rqstp, xprt);
877 /* No data, incomplete (TCP) read, or accept() */
882 clear_bit(XPT_OLD, &xprt->xpt_flags);
884 if (xprt->xpt_ops->xpo_secure_port(rqstp))
885 set_bit(RQ_SECURE, &rqstp->rq_flags);
887 clear_bit(RQ_SECURE, &rqstp->rq_flags);
888 rqstp->rq_chandle.defer = svc_defer;
889 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
892 serv->sv_stats->netcnt++;
893 trace_svc_recv(rqstp, len);
896 rqstp->rq_res.len = 0;
897 svc_xprt_release(rqstp);
899 trace_svc_recv(rqstp, err);
902 EXPORT_SYMBOL_GPL(svc_recv);
907 void svc_drop(struct svc_rqst *rqstp)
909 trace_svc_drop(rqstp);
910 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
911 svc_xprt_release(rqstp);
913 EXPORT_SYMBOL_GPL(svc_drop);
916 * Return reply to client.
918 int svc_send(struct svc_rqst *rqstp)
920 struct svc_xprt *xprt;
924 xprt = rqstp->rq_xprt;
928 /* release the receive skb before sending the reply */
929 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
931 /* calculate over-all length */
933 xb->len = xb->head[0].iov_len +
937 /* Grab mutex to serialize outgoing data. */
938 mutex_lock(&xprt->xpt_mutex);
939 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
940 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
943 len = xprt->xpt_ops->xpo_sendto(rqstp);
944 mutex_unlock(&xprt->xpt_mutex);
945 rpc_wake_up(&xprt->xpt_bc_pending);
946 svc_xprt_release(rqstp);
948 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
951 trace_svc_send(rqstp, len);
956 * Timer function to close old temporary transports, using
957 * a mark-and-sweep algorithm.
959 static void svc_age_temp_xprts(unsigned long closure)
961 struct svc_serv *serv = (struct svc_serv *)closure;
962 struct svc_xprt *xprt;
963 struct list_head *le, *next;
965 dprintk("svc_age_temp_xprts\n");
967 if (!spin_trylock_bh(&serv->sv_lock)) {
968 /* busy, try again 1 sec later */
969 dprintk("svc_age_temp_xprts: busy\n");
970 mod_timer(&serv->sv_temptimer, jiffies + HZ);
974 list_for_each_safe(le, next, &serv->sv_tempsocks) {
975 xprt = list_entry(le, struct svc_xprt, xpt_list);
977 /* First time through, just mark it OLD. Second time
978 * through, close it. */
979 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
981 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
982 test_bit(XPT_BUSY, &xprt->xpt_flags))
985 set_bit(XPT_CLOSE, &xprt->xpt_flags);
986 dprintk("queuing xprt %p for closing\n", xprt);
988 /* a thread will dequeue and close it soon */
989 svc_xprt_enqueue(xprt);
991 spin_unlock_bh(&serv->sv_lock);
993 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
996 /* Close temporary transports whose xpt_local matches server_addr immediately
997 * instead of waiting for them to be picked up by the timer.
999 * This is meant to be called from a notifier_block that runs when an ip
1000 * address is deleted.
1002 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
1004 struct svc_xprt *xprt;
1005 struct list_head *le, *next;
1006 LIST_HEAD(to_be_closed);
1008 spin_lock_bh(&serv->sv_lock);
1009 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1010 xprt = list_entry(le, struct svc_xprt, xpt_list);
1011 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
1012 &xprt->xpt_local)) {
1013 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
1014 list_move(le, &to_be_closed);
1017 spin_unlock_bh(&serv->sv_lock);
1019 while (!list_empty(&to_be_closed)) {
1020 le = to_be_closed.next;
1022 xprt = list_entry(le, struct svc_xprt, xpt_list);
1023 dprintk("svc_age_temp_xprts_now: closing %p\n", xprt);
1024 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
1025 svc_close_xprt(xprt);
1028 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1030 static void call_xpt_users(struct svc_xprt *xprt)
1032 struct svc_xpt_user *u;
1034 spin_lock(&xprt->xpt_lock);
1035 while (!list_empty(&xprt->xpt_users)) {
1036 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1040 spin_unlock(&xprt->xpt_lock);
1044 * Remove a dead transport
1046 static void svc_delete_xprt(struct svc_xprt *xprt)
1048 struct svc_serv *serv = xprt->xpt_server;
1049 struct svc_deferred_req *dr;
1051 /* Only do this once */
1052 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1055 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1056 xprt->xpt_ops->xpo_detach(xprt);
1058 spin_lock_bh(&serv->sv_lock);
1059 list_del_init(&xprt->xpt_list);
1060 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1061 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1063 spin_unlock_bh(&serv->sv_lock);
1065 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1068 call_xpt_users(xprt);
1072 void svc_close_xprt(struct svc_xprt *xprt)
1074 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1075 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1076 /* someone else will have to effect the close */
1079 * We expect svc_close_xprt() to work even when no threads are
1080 * running (e.g., while configuring the server before starting
1081 * any threads), so if the transport isn't busy, we delete
1084 svc_delete_xprt(xprt);
1086 EXPORT_SYMBOL_GPL(svc_close_xprt);
1088 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1090 struct svc_xprt *xprt;
1093 spin_lock(&serv->sv_lock);
1094 list_for_each_entry(xprt, xprt_list, xpt_list) {
1095 if (xprt->xpt_net != net)
1098 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1099 svc_xprt_enqueue(xprt);
1101 spin_unlock(&serv->sv_lock);
1105 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1107 struct svc_pool *pool;
1108 struct svc_xprt *xprt;
1109 struct svc_xprt *tmp;
1112 for (i = 0; i < serv->sv_nrpools; i++) {
1113 pool = &serv->sv_pools[i];
1115 spin_lock_bh(&pool->sp_lock);
1116 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1117 if (xprt->xpt_net != net)
1119 list_del_init(&xprt->xpt_ready);
1120 spin_unlock_bh(&pool->sp_lock);
1123 spin_unlock_bh(&pool->sp_lock);
1128 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1130 struct svc_xprt *xprt;
1132 while ((xprt = svc_dequeue_net(serv, net))) {
1133 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1134 svc_delete_xprt(xprt);
1139 * Server threads may still be running (especially in the case where the
1140 * service is still running in other network namespaces).
1142 * So we shut down sockets the same way we would on a running server, by
1143 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1144 * the close. In the case there are no such other threads,
1145 * threads running, svc_clean_up_xprts() does a simple version of a
1146 * server's main event loop, and in the case where there are other
1147 * threads, we may need to wait a little while and then check again to
1148 * see if they're done.
1150 void svc_close_net(struct svc_serv *serv, struct net *net)
1154 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1155 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1157 svc_clean_up_xprts(serv, net);
1163 * Handle defer and revisit of requests
1166 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1168 struct svc_deferred_req *dr =
1169 container_of(dreq, struct svc_deferred_req, handle);
1170 struct svc_xprt *xprt = dr->xprt;
1172 spin_lock(&xprt->xpt_lock);
1173 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1174 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1175 spin_unlock(&xprt->xpt_lock);
1176 dprintk("revisit canceled\n");
1178 trace_svc_drop_deferred(dr);
1182 dprintk("revisit queued\n");
1184 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1185 spin_unlock(&xprt->xpt_lock);
1186 svc_xprt_enqueue(xprt);
1191 * Save the request off for later processing. The request buffer looks
1194 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1196 * This code can only handle requests that consist of an xprt-header
1199 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1201 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1202 struct svc_deferred_req *dr;
1204 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1205 return NULL; /* if more than a page, give up FIXME */
1206 if (rqstp->rq_deferred) {
1207 dr = rqstp->rq_deferred;
1208 rqstp->rq_deferred = NULL;
1212 /* FIXME maybe discard if size too large */
1213 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1214 dr = kmalloc(size, GFP_KERNEL);
1218 dr->handle.owner = rqstp->rq_server;
1219 dr->prot = rqstp->rq_prot;
1220 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1221 dr->addrlen = rqstp->rq_addrlen;
1222 dr->daddr = rqstp->rq_daddr;
1223 dr->argslen = rqstp->rq_arg.len >> 2;
1224 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1226 /* back up head to the start of the buffer and copy */
1227 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1228 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1231 svc_xprt_get(rqstp->rq_xprt);
1232 dr->xprt = rqstp->rq_xprt;
1233 set_bit(RQ_DROPME, &rqstp->rq_flags);
1235 dr->handle.revisit = svc_revisit;
1236 trace_svc_defer(rqstp);
1241 * recv data from a deferred request into an active one
1243 static int svc_deferred_recv(struct svc_rqst *rqstp)
1245 struct svc_deferred_req *dr = rqstp->rq_deferred;
1247 /* setup iov_base past transport header */
1248 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1249 /* The iov_len does not include the transport header bytes */
1250 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1251 rqstp->rq_arg.page_len = 0;
1252 /* The rq_arg.len includes the transport header bytes */
1253 rqstp->rq_arg.len = dr->argslen<<2;
1254 rqstp->rq_prot = dr->prot;
1255 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1256 rqstp->rq_addrlen = dr->addrlen;
1257 /* Save off transport header len in case we get deferred again */
1258 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1259 rqstp->rq_daddr = dr->daddr;
1260 rqstp->rq_respages = rqstp->rq_pages;
1261 return (dr->argslen<<2) - dr->xprt_hlen;
1265 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1267 struct svc_deferred_req *dr = NULL;
1269 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1271 spin_lock(&xprt->xpt_lock);
1272 if (!list_empty(&xprt->xpt_deferred)) {
1273 dr = list_entry(xprt->xpt_deferred.next,
1274 struct svc_deferred_req,
1276 list_del_init(&dr->handle.recent);
1277 trace_svc_revisit_deferred(dr);
1279 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1280 spin_unlock(&xprt->xpt_lock);
1285 * svc_find_xprt - find an RPC transport instance
1286 * @serv: pointer to svc_serv to search
1287 * @xcl_name: C string containing transport's class name
1288 * @net: owner net pointer
1289 * @af: Address family of transport's local address
1290 * @port: transport's IP port number
1292 * Return the transport instance pointer for the endpoint accepting
1293 * connections/peer traffic from the specified transport class,
1294 * address family and port.
1296 * Specifying 0 for the address family or port is effectively a
1297 * wild-card, and will result in matching the first transport in the
1298 * service's list that has a matching class name.
1300 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1301 struct net *net, const sa_family_t af,
1302 const unsigned short port)
1304 struct svc_xprt *xprt;
1305 struct svc_xprt *found = NULL;
1307 /* Sanity check the args */
1308 if (serv == NULL || xcl_name == NULL)
1311 spin_lock_bh(&serv->sv_lock);
1312 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1313 if (xprt->xpt_net != net)
1315 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1317 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1319 if (port != 0 && port != svc_xprt_local_port(xprt))
1325 spin_unlock_bh(&serv->sv_lock);
1328 EXPORT_SYMBOL_GPL(svc_find_xprt);
1330 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1331 char *pos, int remaining)
1335 len = snprintf(pos, remaining, "%s %u\n",
1336 xprt->xpt_class->xcl_name,
1337 svc_xprt_local_port(xprt));
1338 if (len >= remaining)
1339 return -ENAMETOOLONG;
1344 * svc_xprt_names - format a buffer with a list of transport names
1345 * @serv: pointer to an RPC service
1346 * @buf: pointer to a buffer to be filled in
1347 * @buflen: length of buffer to be filled in
1349 * Fills in @buf with a string containing a list of transport names,
1350 * each name terminated with '\n'.
1352 * Returns positive length of the filled-in string on success; otherwise
1353 * a negative errno value is returned if an error occurs.
1355 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1357 struct svc_xprt *xprt;
1361 /* Sanity check args */
1365 spin_lock_bh(&serv->sv_lock);
1369 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1370 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1382 spin_unlock_bh(&serv->sv_lock);
1385 EXPORT_SYMBOL_GPL(svc_xprt_names);
1388 /*----------------------------------------------------------------------------*/
1390 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1392 unsigned int pidx = (unsigned int)*pos;
1393 struct svc_serv *serv = m->private;
1395 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1398 return SEQ_START_TOKEN;
1399 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1402 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1404 struct svc_pool *pool = p;
1405 struct svc_serv *serv = m->private;
1407 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1409 if (p == SEQ_START_TOKEN) {
1410 pool = &serv->sv_pools[0];
1412 unsigned int pidx = (pool - &serv->sv_pools[0]);
1413 if (pidx < serv->sv_nrpools-1)
1414 pool = &serv->sv_pools[pidx+1];
1422 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1426 static int svc_pool_stats_show(struct seq_file *m, void *p)
1428 struct svc_pool *pool = p;
1430 if (p == SEQ_START_TOKEN) {
1431 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1435 seq_printf(m, "%u %lu %lu %lu %lu\n",
1437 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1438 pool->sp_stats.sockets_queued,
1439 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1440 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1445 static const struct seq_operations svc_pool_stats_seq_ops = {
1446 .start = svc_pool_stats_start,
1447 .next = svc_pool_stats_next,
1448 .stop = svc_pool_stats_stop,
1449 .show = svc_pool_stats_show,
1452 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1456 err = seq_open(file, &svc_pool_stats_seq_ops);
1458 ((struct seq_file *) file->private_data)->private = serv;
1461 EXPORT_SYMBOL(svc_pool_stats_open);
1463 /*----------------------------------------------------------------------------*/