2 * linux/net/sunrpc/svc_xprt.c
4 * Author: Tom Tucker <tom@opengridcomputing.com>
7 #include <linux/sched.h>
8 #include <linux/errno.h>
9 #include <linux/freezer.h>
10 #include <linux/kthread.h>
11 #include <linux/slab.h>
13 #include <linux/sunrpc/stats.h>
14 #include <linux/sunrpc/svc_xprt.h>
15 #include <linux/sunrpc/svcsock.h>
16 #include <linux/sunrpc/xprt.h>
17 #include <linux/module.h>
18 #include <trace/events/sunrpc.h>
20 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
22 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
23 static int svc_deferred_recv(struct svc_rqst *rqstp);
24 static struct cache_deferred_req *svc_defer(struct cache_req *req);
25 static void svc_age_temp_xprts(unsigned long closure);
26 static void svc_delete_xprt(struct svc_xprt *xprt);
27 static void svc_xprt_do_enqueue(struct svc_xprt *xprt);
29 /* apparently the "standard" is that clients close
30 * idle connections after 5 minutes, servers after
32 * http://www.connectathon.org/talks96/nfstcp.pdf
34 static int svc_conn_age_period = 6*60;
36 /* List of registered transport classes */
37 static DEFINE_SPINLOCK(svc_xprt_class_lock);
38 static LIST_HEAD(svc_xprt_class_list);
40 /* SMP locking strategy:
42 * svc_pool->sp_lock protects most of the fields of that pool.
43 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
44 * when both need to be taken (rare), svc_serv->sv_lock is first.
45 * BKL protects svc_serv->sv_nrthread.
46 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
47 * and the ->sk_info_authunix cache.
49 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
50 * enqueued multiply. During normal transport processing this bit
51 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
52 * Providers should not manipulate this bit directly.
54 * Some flags can be set to certain values at any time
55 * providing that certain rules are followed:
58 * - Can be set or cleared at any time.
59 * - After a set, svc_xprt_enqueue must be called to enqueue
60 * the transport for processing.
61 * - After a clear, the transport must be read/accepted.
62 * If this succeeds, it must be set again.
64 * - Can set at any time. It is never cleared.
66 * - Can only be set while XPT_BUSY is held which ensures
67 * that no other thread will be using the transport or will
68 * try to set XPT_DEAD.
71 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
73 struct svc_xprt_class *cl;
76 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
78 INIT_LIST_HEAD(&xcl->xcl_list);
79 spin_lock(&svc_xprt_class_lock);
80 /* Make sure there isn't already a class with the same name */
81 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
82 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
85 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
88 spin_unlock(&svc_xprt_class_lock);
91 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
93 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
95 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
96 spin_lock(&svc_xprt_class_lock);
97 list_del_init(&xcl->xcl_list);
98 spin_unlock(&svc_xprt_class_lock);
100 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
103 * Format the transport list for printing
105 int svc_print_xprts(char *buf, int maxlen)
107 struct svc_xprt_class *xcl;
112 spin_lock(&svc_xprt_class_lock);
113 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
116 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
117 slen = strlen(tmpstr);
118 if (len + slen > maxlen)
123 spin_unlock(&svc_xprt_class_lock);
128 static void svc_xprt_free(struct kref *kref)
130 struct svc_xprt *xprt =
131 container_of(kref, struct svc_xprt, xpt_ref);
132 struct module *owner = xprt->xpt_class->xcl_owner;
133 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
134 svcauth_unix_info_release(xprt);
135 put_net(xprt->xpt_net);
136 /* See comment on corresponding get in xs_setup_bc_tcp(): */
137 if (xprt->xpt_bc_xprt)
138 xprt_put(xprt->xpt_bc_xprt);
139 xprt->xpt_ops->xpo_free(xprt);
143 void svc_xprt_put(struct svc_xprt *xprt)
145 kref_put(&xprt->xpt_ref, svc_xprt_free);
147 EXPORT_SYMBOL_GPL(svc_xprt_put);
150 * Called by transport drivers to initialize the transport independent
151 * portion of the transport instance.
153 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
154 struct svc_xprt *xprt, struct svc_serv *serv)
156 memset(xprt, 0, sizeof(*xprt));
157 xprt->xpt_class = xcl;
158 xprt->xpt_ops = xcl->xcl_ops;
159 kref_init(&xprt->xpt_ref);
160 xprt->xpt_server = serv;
161 INIT_LIST_HEAD(&xprt->xpt_list);
162 INIT_LIST_HEAD(&xprt->xpt_ready);
163 INIT_LIST_HEAD(&xprt->xpt_deferred);
164 INIT_LIST_HEAD(&xprt->xpt_users);
165 mutex_init(&xprt->xpt_mutex);
166 spin_lock_init(&xprt->xpt_lock);
167 set_bit(XPT_BUSY, &xprt->xpt_flags);
168 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
169 xprt->xpt_net = get_net(net);
171 EXPORT_SYMBOL_GPL(svc_xprt_init);
173 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
174 struct svc_serv *serv,
177 const unsigned short port,
180 struct sockaddr_in sin = {
181 .sin_family = AF_INET,
182 .sin_addr.s_addr = htonl(INADDR_ANY),
183 .sin_port = htons(port),
185 #if IS_ENABLED(CONFIG_IPV6)
186 struct sockaddr_in6 sin6 = {
187 .sin6_family = AF_INET6,
188 .sin6_addr = IN6ADDR_ANY_INIT,
189 .sin6_port = htons(port),
192 struct sockaddr *sap;
197 sap = (struct sockaddr *)&sin;
200 #if IS_ENABLED(CONFIG_IPV6)
202 sap = (struct sockaddr *)&sin6;
207 return ERR_PTR(-EAFNOSUPPORT);
210 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
214 * svc_xprt_received conditionally queues the transport for processing
215 * by another thread. The caller must hold the XPT_BUSY bit and must
216 * not thereafter touch transport data.
218 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
219 * insufficient) data.
221 static void svc_xprt_received(struct svc_xprt *xprt)
223 WARN_ON_ONCE(!test_bit(XPT_BUSY, &xprt->xpt_flags));
224 if (!test_bit(XPT_BUSY, &xprt->xpt_flags))
226 /* As soon as we clear busy, the xprt could be closed and
227 * 'put', so we need a reference to call svc_xprt_do_enqueue with:
230 smp_mb__before_atomic();
231 clear_bit(XPT_BUSY, &xprt->xpt_flags);
232 svc_xprt_do_enqueue(xprt);
236 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
238 clear_bit(XPT_TEMP, &new->xpt_flags);
239 spin_lock_bh(&serv->sv_lock);
240 list_add(&new->xpt_list, &serv->sv_permsocks);
241 spin_unlock_bh(&serv->sv_lock);
242 svc_xprt_received(new);
245 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
246 struct net *net, const int family,
247 const unsigned short port, int flags)
249 struct svc_xprt_class *xcl;
251 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
252 spin_lock(&svc_xprt_class_lock);
253 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
254 struct svc_xprt *newxprt;
255 unsigned short newport;
257 if (strcmp(xprt_name, xcl->xcl_name))
260 if (!try_module_get(xcl->xcl_owner))
263 spin_unlock(&svc_xprt_class_lock);
264 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
265 if (IS_ERR(newxprt)) {
266 module_put(xcl->xcl_owner);
267 return PTR_ERR(newxprt);
269 svc_add_new_perm_xprt(serv, newxprt);
270 newport = svc_xprt_local_port(newxprt);
274 spin_unlock(&svc_xprt_class_lock);
275 dprintk("svc: transport %s not found\n", xprt_name);
277 /* This errno is exposed to user space. Provide a reasonable
278 * perror msg for a bad transport. */
279 return -EPROTONOSUPPORT;
281 EXPORT_SYMBOL_GPL(svc_create_xprt);
284 * Copy the local and remote xprt addresses to the rqstp structure
286 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
288 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
289 rqstp->rq_addrlen = xprt->xpt_remotelen;
292 * Destination address in request is needed for binding the
293 * source address in RPC replies/callbacks later.
295 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
296 rqstp->rq_daddrlen = xprt->xpt_locallen;
298 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
301 * svc_print_addr - Format rq_addr field for printing
302 * @rqstp: svc_rqst struct containing address to print
303 * @buf: target buffer for formatted address
304 * @len: length of target buffer
307 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
309 return __svc_print_addr(svc_addr(rqstp), buf, len);
311 EXPORT_SYMBOL_GPL(svc_print_addr);
314 * Queue up an idle server thread. Must have pool->sp_lock held.
315 * Note: this is really a stack rather than a queue, so that we only
316 * use as many different threads as we need, and the rest don't pollute
319 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
321 list_add(&rqstp->rq_list, &pool->sp_threads);
325 * Dequeue an nfsd thread. Must have pool->sp_lock held.
327 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
329 list_del(&rqstp->rq_list);
332 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
334 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
336 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
337 return xprt->xpt_ops->xpo_has_wspace(xprt);
341 static void svc_xprt_do_enqueue(struct svc_xprt *xprt)
343 struct svc_pool *pool;
344 struct svc_rqst *rqstp;
347 if (!svc_xprt_has_something_to_do(xprt))
350 /* Mark transport as busy. It will remain in this state until
351 * the provider calls svc_xprt_received. We update XPT_BUSY
352 * atomically because it also guards against trying to enqueue
353 * the transport twice.
355 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
356 /* Don't enqueue transport while already enqueued */
357 dprintk("svc: transport %p busy, not enqueued\n", xprt);
362 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
363 spin_lock_bh(&pool->sp_lock);
365 atomic_long_inc(&pool->sp_stats.packets);
367 if (!list_empty(&pool->sp_threads)) {
368 rqstp = list_entry(pool->sp_threads.next,
371 dprintk("svc: transport %p served by daemon %p\n",
373 svc_thread_dequeue(pool, rqstp);
376 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
377 rqstp, rqstp->rq_xprt);
378 /* Note the order of the following 3 lines:
379 * We want to assign xprt to rqstp->rq_xprt only _after_
380 * we've woken up the process, so that we don't race with
381 * the lockless check in svc_get_next_xprt().
384 wake_up_process(rqstp->rq_task);
385 rqstp->rq_xprt = xprt;
386 atomic_long_inc(&pool->sp_stats.threads_woken);
388 dprintk("svc: transport %p put into queue\n", xprt);
389 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
390 pool->sp_stats.sockets_queued++;
393 spin_unlock_bh(&pool->sp_lock);
398 * Queue up a transport with data pending. If there are idle nfsd
399 * processes, wake 'em up.
402 void svc_xprt_enqueue(struct svc_xprt *xprt)
404 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
406 svc_xprt_do_enqueue(xprt);
408 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
411 * Dequeue the first transport. Must be called with the pool->sp_lock held.
413 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
415 struct svc_xprt *xprt;
417 if (list_empty(&pool->sp_sockets))
420 xprt = list_entry(pool->sp_sockets.next,
421 struct svc_xprt, xpt_ready);
422 list_del_init(&xprt->xpt_ready);
424 dprintk("svc: transport %p dequeued, inuse=%d\n",
425 xprt, atomic_read(&xprt->xpt_ref.refcount));
431 * svc_reserve - change the space reserved for the reply to a request.
432 * @rqstp: The request in question
433 * @space: new max space to reserve
435 * Each request reserves some space on the output queue of the transport
436 * to make sure the reply fits. This function reduces that reserved
437 * space to be the amount of space used already, plus @space.
440 void svc_reserve(struct svc_rqst *rqstp, int space)
442 space += rqstp->rq_res.head[0].iov_len;
444 if (space < rqstp->rq_reserved) {
445 struct svc_xprt *xprt = rqstp->rq_xprt;
446 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
447 rqstp->rq_reserved = space;
449 if (xprt->xpt_ops->xpo_adjust_wspace)
450 xprt->xpt_ops->xpo_adjust_wspace(xprt);
451 svc_xprt_enqueue(xprt);
454 EXPORT_SYMBOL_GPL(svc_reserve);
456 static void svc_xprt_release(struct svc_rqst *rqstp)
458 struct svc_xprt *xprt = rqstp->rq_xprt;
460 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
462 kfree(rqstp->rq_deferred);
463 rqstp->rq_deferred = NULL;
465 svc_free_res_pages(rqstp);
466 rqstp->rq_res.page_len = 0;
467 rqstp->rq_res.page_base = 0;
469 /* Reset response buffer and release
471 * But first, check that enough space was reserved
472 * for the reply, otherwise we have a bug!
474 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
475 printk(KERN_ERR "RPC request reserved %d but used %d\n",
479 rqstp->rq_res.head[0].iov_len = 0;
480 svc_reserve(rqstp, 0);
481 rqstp->rq_xprt = NULL;
487 * Some svc_serv's will have occasional work to do, even when a xprt is not
488 * waiting to be serviced. This function is there to "kick" a task in one of
489 * those services so that it can wake up and do that work. Note that we only
490 * bother with pool 0 as we don't need to wake up more than one thread for
493 void svc_wake_up(struct svc_serv *serv)
495 struct svc_rqst *rqstp;
496 struct svc_pool *pool;
498 pool = &serv->sv_pools[0];
500 spin_lock_bh(&pool->sp_lock);
501 if (!list_empty(&pool->sp_threads)) {
502 rqstp = list_entry(pool->sp_threads.next,
505 dprintk("svc: daemon %p woken up.\n", rqstp);
506 wake_up_process(rqstp->rq_task);
508 set_bit(SP_TASK_PENDING, &pool->sp_flags);
509 spin_unlock_bh(&pool->sp_lock);
511 EXPORT_SYMBOL_GPL(svc_wake_up);
513 int svc_port_is_privileged(struct sockaddr *sin)
515 switch (sin->sa_family) {
517 return ntohs(((struct sockaddr_in *)sin)->sin_port)
520 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
528 * Make sure that we don't have too many active connections. If we have,
529 * something must be dropped. It's not clear what will happen if we allow
530 * "too many" connections, but when dealing with network-facing software,
531 * we have to code defensively. Here we do that by imposing hard limits.
533 * There's no point in trying to do random drop here for DoS
534 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
535 * attacker can easily beat that.
537 * The only somewhat efficient mechanism would be if drop old
538 * connections from the same IP first. But right now we don't even
539 * record the client IP in svc_sock.
541 * single-threaded services that expect a lot of clients will probably
542 * need to set sv_maxconn to override the default value which is based
543 * on the number of threads
545 static void svc_check_conn_limits(struct svc_serv *serv)
547 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
548 (serv->sv_nrthreads+3) * 20;
550 if (serv->sv_tmpcnt > limit) {
551 struct svc_xprt *xprt = NULL;
552 spin_lock_bh(&serv->sv_lock);
553 if (!list_empty(&serv->sv_tempsocks)) {
554 /* Try to help the admin */
555 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
556 serv->sv_name, serv->sv_maxconn ?
557 "max number of connections" :
558 "number of threads");
560 * Always select the oldest connection. It's not fair,
563 xprt = list_entry(serv->sv_tempsocks.prev,
566 set_bit(XPT_CLOSE, &xprt->xpt_flags);
569 spin_unlock_bh(&serv->sv_lock);
572 svc_xprt_enqueue(xprt);
578 static int svc_alloc_arg(struct svc_rqst *rqstp)
580 struct svc_serv *serv = rqstp->rq_server;
585 /* now allocate needed pages. If we get a failure, sleep briefly */
586 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
587 WARN_ON_ONCE(pages >= RPCSVC_MAXPAGES);
588 if (pages >= RPCSVC_MAXPAGES)
589 /* use as many pages as possible */
590 pages = RPCSVC_MAXPAGES - 1;
591 for (i = 0; i < pages ; i++)
592 while (rqstp->rq_pages[i] == NULL) {
593 struct page *p = alloc_page(GFP_KERNEL);
595 set_current_state(TASK_INTERRUPTIBLE);
596 if (signalled() || kthread_should_stop()) {
597 set_current_state(TASK_RUNNING);
600 schedule_timeout(msecs_to_jiffies(500));
602 rqstp->rq_pages[i] = p;
604 rqstp->rq_page_end = &rqstp->rq_pages[i];
605 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
607 /* Make arg->head point to first page and arg->pages point to rest */
608 arg = &rqstp->rq_arg;
609 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
610 arg->head[0].iov_len = PAGE_SIZE;
611 arg->pages = rqstp->rq_pages + 1;
613 /* save at least one page for response */
614 arg->page_len = (pages-2)*PAGE_SIZE;
615 arg->len = (pages-1)*PAGE_SIZE;
616 arg->tail[0].iov_len = 0;
620 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
622 struct svc_xprt *xprt;
623 struct svc_pool *pool = rqstp->rq_pool;
626 /* Normally we will wait up to 5 seconds for any required
627 * cache information to be provided.
629 rqstp->rq_chandle.thread_wait = 5*HZ;
631 spin_lock_bh(&pool->sp_lock);
632 xprt = svc_xprt_dequeue(pool);
634 rqstp->rq_xprt = xprt;
637 /* As there is a shortage of threads and this request
638 * had to be queued, don't allow the thread to wait so
639 * long for cache updates.
641 rqstp->rq_chandle.thread_wait = 1*HZ;
642 clear_bit(SP_TASK_PENDING, &pool->sp_flags);
644 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags)) {
645 xprt = ERR_PTR(-EAGAIN);
649 * We have to be able to interrupt this wait
650 * to bring down the daemons ...
652 set_current_state(TASK_INTERRUPTIBLE);
654 /* No data pending. Go to sleep */
655 svc_thread_enqueue(pool, rqstp);
656 spin_unlock_bh(&pool->sp_lock);
658 if (!(signalled() || kthread_should_stop())) {
659 time_left = schedule_timeout(timeout);
660 __set_current_state(TASK_RUNNING);
664 xprt = rqstp->rq_xprt;
668 __set_current_state(TASK_RUNNING);
670 spin_lock_bh(&pool->sp_lock);
672 atomic_long_inc(&pool->sp_stats.threads_timedout);
674 xprt = rqstp->rq_xprt;
676 svc_thread_dequeue(pool, rqstp);
677 spin_unlock_bh(&pool->sp_lock);
678 dprintk("svc: server %p, no data yet\n", rqstp);
679 if (signalled() || kthread_should_stop())
680 return ERR_PTR(-EINTR);
682 return ERR_PTR(-EAGAIN);
686 spin_unlock_bh(&pool->sp_lock);
690 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
692 spin_lock_bh(&serv->sv_lock);
693 set_bit(XPT_TEMP, &newxpt->xpt_flags);
694 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
696 if (serv->sv_temptimer.function == NULL) {
697 /* setup timer to age temp transports */
698 setup_timer(&serv->sv_temptimer, svc_age_temp_xprts,
699 (unsigned long)serv);
700 mod_timer(&serv->sv_temptimer,
701 jiffies + svc_conn_age_period * HZ);
703 spin_unlock_bh(&serv->sv_lock);
704 svc_xprt_received(newxpt);
707 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
709 struct svc_serv *serv = rqstp->rq_server;
712 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
713 dprintk("svc_recv: found XPT_CLOSE\n");
714 svc_delete_xprt(xprt);
715 /* Leave XPT_BUSY set on the dead xprt: */
718 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
719 struct svc_xprt *newxpt;
721 * We know this module_get will succeed because the
722 * listener holds a reference too
724 __module_get(xprt->xpt_class->xcl_owner);
725 svc_check_conn_limits(xprt->xpt_server);
726 newxpt = xprt->xpt_ops->xpo_accept(xprt);
728 svc_add_new_temp_xprt(serv, newxpt);
730 module_put(xprt->xpt_class->xcl_owner);
732 /* XPT_DATA|XPT_DEFERRED case: */
733 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
734 rqstp, rqstp->rq_pool->sp_id, xprt,
735 atomic_read(&xprt->xpt_ref.refcount));
736 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
737 if (rqstp->rq_deferred)
738 len = svc_deferred_recv(rqstp);
740 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
741 dprintk("svc: got len=%d\n", len);
742 rqstp->rq_reserved = serv->sv_max_mesg;
743 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
745 /* clear XPT_BUSY: */
746 svc_xprt_received(xprt);
751 * Receive the next request on any transport. This code is carefully
752 * organised not to touch any cachelines in the shared svc_serv
753 * structure, only cachelines in the local svc_pool.
755 int svc_recv(struct svc_rqst *rqstp, long timeout)
757 struct svc_xprt *xprt = NULL;
758 struct svc_serv *serv = rqstp->rq_server;
761 dprintk("svc: server %p waiting for data (to = %ld)\n",
766 "svc_recv: service %p, transport not NULL!\n",
769 err = svc_alloc_arg(rqstp);
776 if (signalled() || kthread_should_stop())
779 xprt = svc_get_next_xprt(rqstp, timeout);
785 len = svc_handle_xprt(rqstp, xprt);
787 /* No data, incomplete (TCP) read, or accept() */
792 clear_bit(XPT_OLD, &xprt->xpt_flags);
794 if (xprt->xpt_ops->xpo_secure_port(rqstp))
795 set_bit(RQ_SECURE, &rqstp->rq_flags);
797 clear_bit(RQ_SECURE, &rqstp->rq_flags);
798 rqstp->rq_chandle.defer = svc_defer;
799 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
802 serv->sv_stats->netcnt++;
803 trace_svc_recv(rqstp, len);
806 rqstp->rq_res.len = 0;
807 svc_xprt_release(rqstp);
809 trace_svc_recv(rqstp, err);
812 EXPORT_SYMBOL_GPL(svc_recv);
817 void svc_drop(struct svc_rqst *rqstp)
819 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
820 svc_xprt_release(rqstp);
822 EXPORT_SYMBOL_GPL(svc_drop);
825 * Return reply to client.
827 int svc_send(struct svc_rqst *rqstp)
829 struct svc_xprt *xprt;
833 xprt = rqstp->rq_xprt;
837 /* release the receive skb before sending the reply */
838 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
840 /* calculate over-all length */
842 xb->len = xb->head[0].iov_len +
846 /* Grab mutex to serialize outgoing data. */
847 mutex_lock(&xprt->xpt_mutex);
848 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
849 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
852 len = xprt->xpt_ops->xpo_sendto(rqstp);
853 mutex_unlock(&xprt->xpt_mutex);
854 rpc_wake_up(&xprt->xpt_bc_pending);
855 svc_xprt_release(rqstp);
857 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
860 trace_svc_send(rqstp, len);
865 * Timer function to close old temporary transports, using
866 * a mark-and-sweep algorithm.
868 static void svc_age_temp_xprts(unsigned long closure)
870 struct svc_serv *serv = (struct svc_serv *)closure;
871 struct svc_xprt *xprt;
872 struct list_head *le, *next;
874 dprintk("svc_age_temp_xprts\n");
876 if (!spin_trylock_bh(&serv->sv_lock)) {
877 /* busy, try again 1 sec later */
878 dprintk("svc_age_temp_xprts: busy\n");
879 mod_timer(&serv->sv_temptimer, jiffies + HZ);
883 list_for_each_safe(le, next, &serv->sv_tempsocks) {
884 xprt = list_entry(le, struct svc_xprt, xpt_list);
886 /* First time through, just mark it OLD. Second time
887 * through, close it. */
888 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
890 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
891 test_bit(XPT_BUSY, &xprt->xpt_flags))
894 set_bit(XPT_CLOSE, &xprt->xpt_flags);
895 dprintk("queuing xprt %p for closing\n", xprt);
897 /* a thread will dequeue and close it soon */
898 svc_xprt_enqueue(xprt);
900 spin_unlock_bh(&serv->sv_lock);
902 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
905 static void call_xpt_users(struct svc_xprt *xprt)
907 struct svc_xpt_user *u;
909 spin_lock(&xprt->xpt_lock);
910 while (!list_empty(&xprt->xpt_users)) {
911 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
915 spin_unlock(&xprt->xpt_lock);
919 * Remove a dead transport
921 static void svc_delete_xprt(struct svc_xprt *xprt)
923 struct svc_serv *serv = xprt->xpt_server;
924 struct svc_deferred_req *dr;
926 /* Only do this once */
927 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
930 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
931 xprt->xpt_ops->xpo_detach(xprt);
933 spin_lock_bh(&serv->sv_lock);
934 list_del_init(&xprt->xpt_list);
935 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
936 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
938 spin_unlock_bh(&serv->sv_lock);
940 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
943 call_xpt_users(xprt);
947 void svc_close_xprt(struct svc_xprt *xprt)
949 set_bit(XPT_CLOSE, &xprt->xpt_flags);
950 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
951 /* someone else will have to effect the close */
954 * We expect svc_close_xprt() to work even when no threads are
955 * running (e.g., while configuring the server before starting
956 * any threads), so if the transport isn't busy, we delete
959 svc_delete_xprt(xprt);
961 EXPORT_SYMBOL_GPL(svc_close_xprt);
963 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
965 struct svc_xprt *xprt;
968 spin_lock(&serv->sv_lock);
969 list_for_each_entry(xprt, xprt_list, xpt_list) {
970 if (xprt->xpt_net != net)
973 set_bit(XPT_CLOSE, &xprt->xpt_flags);
974 svc_xprt_enqueue(xprt);
976 spin_unlock(&serv->sv_lock);
980 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
982 struct svc_pool *pool;
983 struct svc_xprt *xprt;
984 struct svc_xprt *tmp;
987 for (i = 0; i < serv->sv_nrpools; i++) {
988 pool = &serv->sv_pools[i];
990 spin_lock_bh(&pool->sp_lock);
991 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
992 if (xprt->xpt_net != net)
994 list_del_init(&xprt->xpt_ready);
995 spin_unlock_bh(&pool->sp_lock);
998 spin_unlock_bh(&pool->sp_lock);
1003 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1005 struct svc_xprt *xprt;
1007 while ((xprt = svc_dequeue_net(serv, net))) {
1008 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1009 svc_delete_xprt(xprt);
1014 * Server threads may still be running (especially in the case where the
1015 * service is still running in other network namespaces).
1017 * So we shut down sockets the same way we would on a running server, by
1018 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1019 * the close. In the case there are no such other threads,
1020 * threads running, svc_clean_up_xprts() does a simple version of a
1021 * server's main event loop, and in the case where there are other
1022 * threads, we may need to wait a little while and then check again to
1023 * see if they're done.
1025 void svc_close_net(struct svc_serv *serv, struct net *net)
1029 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1030 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1032 svc_clean_up_xprts(serv, net);
1038 * Handle defer and revisit of requests
1041 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1043 struct svc_deferred_req *dr =
1044 container_of(dreq, struct svc_deferred_req, handle);
1045 struct svc_xprt *xprt = dr->xprt;
1047 spin_lock(&xprt->xpt_lock);
1048 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1049 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1050 spin_unlock(&xprt->xpt_lock);
1051 dprintk("revisit canceled\n");
1056 dprintk("revisit queued\n");
1058 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1059 spin_unlock(&xprt->xpt_lock);
1060 svc_xprt_enqueue(xprt);
1065 * Save the request off for later processing. The request buffer looks
1068 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1070 * This code can only handle requests that consist of an xprt-header
1073 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1075 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1076 struct svc_deferred_req *dr;
1078 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1079 return NULL; /* if more than a page, give up FIXME */
1080 if (rqstp->rq_deferred) {
1081 dr = rqstp->rq_deferred;
1082 rqstp->rq_deferred = NULL;
1086 /* FIXME maybe discard if size too large */
1087 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1088 dr = kmalloc(size, GFP_KERNEL);
1092 dr->handle.owner = rqstp->rq_server;
1093 dr->prot = rqstp->rq_prot;
1094 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1095 dr->addrlen = rqstp->rq_addrlen;
1096 dr->daddr = rqstp->rq_daddr;
1097 dr->argslen = rqstp->rq_arg.len >> 2;
1098 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1100 /* back up head to the start of the buffer and copy */
1101 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1102 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1105 svc_xprt_get(rqstp->rq_xprt);
1106 dr->xprt = rqstp->rq_xprt;
1107 set_bit(RQ_DROPME, &rqstp->rq_flags);
1109 dr->handle.revisit = svc_revisit;
1114 * recv data from a deferred request into an active one
1116 static int svc_deferred_recv(struct svc_rqst *rqstp)
1118 struct svc_deferred_req *dr = rqstp->rq_deferred;
1120 /* setup iov_base past transport header */
1121 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1122 /* The iov_len does not include the transport header bytes */
1123 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1124 rqstp->rq_arg.page_len = 0;
1125 /* The rq_arg.len includes the transport header bytes */
1126 rqstp->rq_arg.len = dr->argslen<<2;
1127 rqstp->rq_prot = dr->prot;
1128 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1129 rqstp->rq_addrlen = dr->addrlen;
1130 /* Save off transport header len in case we get deferred again */
1131 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1132 rqstp->rq_daddr = dr->daddr;
1133 rqstp->rq_respages = rqstp->rq_pages;
1134 return (dr->argslen<<2) - dr->xprt_hlen;
1138 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1140 struct svc_deferred_req *dr = NULL;
1142 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1144 spin_lock(&xprt->xpt_lock);
1145 if (!list_empty(&xprt->xpt_deferred)) {
1146 dr = list_entry(xprt->xpt_deferred.next,
1147 struct svc_deferred_req,
1149 list_del_init(&dr->handle.recent);
1151 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1152 spin_unlock(&xprt->xpt_lock);
1157 * svc_find_xprt - find an RPC transport instance
1158 * @serv: pointer to svc_serv to search
1159 * @xcl_name: C string containing transport's class name
1160 * @net: owner net pointer
1161 * @af: Address family of transport's local address
1162 * @port: transport's IP port number
1164 * Return the transport instance pointer for the endpoint accepting
1165 * connections/peer traffic from the specified transport class,
1166 * address family and port.
1168 * Specifying 0 for the address family or port is effectively a
1169 * wild-card, and will result in matching the first transport in the
1170 * service's list that has a matching class name.
1172 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1173 struct net *net, const sa_family_t af,
1174 const unsigned short port)
1176 struct svc_xprt *xprt;
1177 struct svc_xprt *found = NULL;
1179 /* Sanity check the args */
1180 if (serv == NULL || xcl_name == NULL)
1183 spin_lock_bh(&serv->sv_lock);
1184 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1185 if (xprt->xpt_net != net)
1187 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1189 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1191 if (port != 0 && port != svc_xprt_local_port(xprt))
1197 spin_unlock_bh(&serv->sv_lock);
1200 EXPORT_SYMBOL_GPL(svc_find_xprt);
1202 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1203 char *pos, int remaining)
1207 len = snprintf(pos, remaining, "%s %u\n",
1208 xprt->xpt_class->xcl_name,
1209 svc_xprt_local_port(xprt));
1210 if (len >= remaining)
1211 return -ENAMETOOLONG;
1216 * svc_xprt_names - format a buffer with a list of transport names
1217 * @serv: pointer to an RPC service
1218 * @buf: pointer to a buffer to be filled in
1219 * @buflen: length of buffer to be filled in
1221 * Fills in @buf with a string containing a list of transport names,
1222 * each name terminated with '\n'.
1224 * Returns positive length of the filled-in string on success; otherwise
1225 * a negative errno value is returned if an error occurs.
1227 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1229 struct svc_xprt *xprt;
1233 /* Sanity check args */
1237 spin_lock_bh(&serv->sv_lock);
1241 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1242 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1254 spin_unlock_bh(&serv->sv_lock);
1257 EXPORT_SYMBOL_GPL(svc_xprt_names);
1260 /*----------------------------------------------------------------------------*/
1262 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1264 unsigned int pidx = (unsigned int)*pos;
1265 struct svc_serv *serv = m->private;
1267 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1270 return SEQ_START_TOKEN;
1271 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1274 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1276 struct svc_pool *pool = p;
1277 struct svc_serv *serv = m->private;
1279 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1281 if (p == SEQ_START_TOKEN) {
1282 pool = &serv->sv_pools[0];
1284 unsigned int pidx = (pool - &serv->sv_pools[0]);
1285 if (pidx < serv->sv_nrpools-1)
1286 pool = &serv->sv_pools[pidx+1];
1294 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1298 static int svc_pool_stats_show(struct seq_file *m, void *p)
1300 struct svc_pool *pool = p;
1302 if (p == SEQ_START_TOKEN) {
1303 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1307 seq_printf(m, "%u %lu %lu %lu %lu\n",
1309 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1310 pool->sp_stats.sockets_queued,
1311 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1312 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1317 static const struct seq_operations svc_pool_stats_seq_ops = {
1318 .start = svc_pool_stats_start,
1319 .next = svc_pool_stats_next,
1320 .stop = svc_pool_stats_stop,
1321 .show = svc_pool_stats_show,
1324 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1328 err = seq_open(file, &svc_pool_stats_seq_ops);
1330 ((struct seq_file *) file->private_data)->private = serv;
1333 EXPORT_SYMBOL(svc_pool_stats_open);
1335 /*----------------------------------------------------------------------------*/