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(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 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
175 xprt->xpt_net = get_net(net);
176 strcpy(xprt->xpt_remotebuf, "uninitialized");
178 EXPORT_SYMBOL_GPL(svc_xprt_init);
180 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
181 struct svc_serv *serv,
184 const unsigned short port,
187 struct sockaddr_in sin = {
188 .sin_family = AF_INET,
189 .sin_addr.s_addr = htonl(INADDR_ANY),
190 .sin_port = htons(port),
192 #if IS_ENABLED(CONFIG_IPV6)
193 struct sockaddr_in6 sin6 = {
194 .sin6_family = AF_INET6,
195 .sin6_addr = IN6ADDR_ANY_INIT,
196 .sin6_port = htons(port),
199 struct sockaddr *sap;
204 sap = (struct sockaddr *)&sin;
207 #if IS_ENABLED(CONFIG_IPV6)
209 sap = (struct sockaddr *)&sin6;
214 return ERR_PTR(-EAFNOSUPPORT);
217 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
221 * svc_xprt_received conditionally queues the transport for processing
222 * by another thread. The caller must hold the XPT_BUSY bit and must
223 * not thereafter touch transport data.
225 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
226 * insufficient) data.
228 static void svc_xprt_received(struct svc_xprt *xprt)
230 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
231 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
235 /* As soon as we clear busy, the xprt could be closed and
236 * 'put', so we need a reference to call svc_enqueue_xprt with:
239 smp_mb__before_atomic();
240 clear_bit(XPT_BUSY, &xprt->xpt_flags);
241 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
245 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
247 clear_bit(XPT_TEMP, &new->xpt_flags);
248 spin_lock_bh(&serv->sv_lock);
249 list_add(&new->xpt_list, &serv->sv_permsocks);
250 spin_unlock_bh(&serv->sv_lock);
251 svc_xprt_received(new);
254 static int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
255 struct net *net, const int family,
256 const unsigned short port, int flags)
258 struct svc_xprt_class *xcl;
260 spin_lock(&svc_xprt_class_lock);
261 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
262 struct svc_xprt *newxprt;
263 unsigned short newport;
265 if (strcmp(xprt_name, xcl->xcl_name))
268 if (!try_module_get(xcl->xcl_owner))
271 spin_unlock(&svc_xprt_class_lock);
272 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
273 if (IS_ERR(newxprt)) {
274 module_put(xcl->xcl_owner);
275 return PTR_ERR(newxprt);
277 svc_add_new_perm_xprt(serv, newxprt);
278 newport = svc_xprt_local_port(newxprt);
282 spin_unlock(&svc_xprt_class_lock);
283 /* This errno is exposed to user space. Provide a reasonable
284 * perror msg for a bad transport. */
285 return -EPROTONOSUPPORT;
288 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
289 struct net *net, const int family,
290 const unsigned short port, int flags)
294 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
295 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
296 if (err == -EPROTONOSUPPORT) {
297 request_module("svc%s", xprt_name);
298 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
301 dprintk("svc: transport %s not found, err %d\n",
305 EXPORT_SYMBOL_GPL(svc_create_xprt);
308 * Copy the local and remote xprt addresses to the rqstp structure
310 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
312 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
313 rqstp->rq_addrlen = xprt->xpt_remotelen;
316 * Destination address in request is needed for binding the
317 * source address in RPC replies/callbacks later.
319 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
320 rqstp->rq_daddrlen = xprt->xpt_locallen;
322 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
325 * svc_print_addr - Format rq_addr field for printing
326 * @rqstp: svc_rqst struct containing address to print
327 * @buf: target buffer for formatted address
328 * @len: length of target buffer
331 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
333 return __svc_print_addr(svc_addr(rqstp), buf, len);
335 EXPORT_SYMBOL_GPL(svc_print_addr);
337 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
339 unsigned int limit = svc_rpc_per_connection_limit;
340 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
342 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
345 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
347 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
348 if (!svc_xprt_slots_in_range(xprt))
350 atomic_inc(&xprt->xpt_nr_rqsts);
351 set_bit(RQ_DATA, &rqstp->rq_flags);
356 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
358 struct svc_xprt *xprt = rqstp->rq_xprt;
359 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
360 atomic_dec(&xprt->xpt_nr_rqsts);
361 svc_xprt_enqueue(xprt);
365 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
367 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
369 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED))) {
370 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
371 svc_xprt_slots_in_range(xprt))
373 trace_svc_xprt_no_write_space(xprt);
379 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
381 struct svc_pool *pool;
382 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))
397 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
399 atomic_long_inc(&pool->sp_stats.packets);
401 spin_lock_bh(&pool->sp_lock);
402 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
403 pool->sp_stats.sockets_queued++;
404 spin_unlock_bh(&pool->sp_lock);
406 /* find a thread for this xprt */
408 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
409 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
411 atomic_long_inc(&pool->sp_stats.threads_woken);
412 rqstp->rq_qtime = ktime_get();
413 wake_up_process(rqstp->rq_task);
416 set_bit(SP_CONGESTED, &pool->sp_flags);
421 trace_svc_xprt_do_enqueue(xprt, rqstp);
423 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
426 * Queue up a transport with data pending. If there are idle nfsd
427 * processes, wake 'em up.
430 void svc_xprt_enqueue(struct svc_xprt *xprt)
432 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
434 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
436 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
439 * Dequeue the first transport, if there is one.
441 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
443 struct svc_xprt *xprt = NULL;
445 if (list_empty(&pool->sp_sockets))
448 spin_lock_bh(&pool->sp_lock);
449 if (likely(!list_empty(&pool->sp_sockets))) {
450 xprt = list_first_entry(&pool->sp_sockets,
451 struct svc_xprt, xpt_ready);
452 list_del_init(&xprt->xpt_ready);
455 spin_unlock_bh(&pool->sp_lock);
461 * svc_reserve - change the space reserved for the reply to a request.
462 * @rqstp: The request in question
463 * @space: new max space to reserve
465 * Each request reserves some space on the output queue of the transport
466 * to make sure the reply fits. This function reduces that reserved
467 * space to be the amount of space used already, plus @space.
470 void svc_reserve(struct svc_rqst *rqstp, int space)
472 space += rqstp->rq_res.head[0].iov_len;
474 if (space < rqstp->rq_reserved) {
475 struct svc_xprt *xprt = rqstp->rq_xprt;
476 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
477 rqstp->rq_reserved = space;
479 svc_xprt_enqueue(xprt);
482 EXPORT_SYMBOL_GPL(svc_reserve);
484 static void svc_xprt_release(struct svc_rqst *rqstp)
486 struct svc_xprt *xprt = rqstp->rq_xprt;
488 xprt->xpt_ops->xpo_release_rqst(rqstp);
490 kfree(rqstp->rq_deferred);
491 rqstp->rq_deferred = NULL;
493 svc_free_res_pages(rqstp);
494 rqstp->rq_res.page_len = 0;
495 rqstp->rq_res.page_base = 0;
497 /* Reset response buffer and release
499 * But first, check that enough space was reserved
500 * for the reply, otherwise we have a bug!
502 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
503 printk(KERN_ERR "RPC request reserved %d but used %d\n",
507 rqstp->rq_res.head[0].iov_len = 0;
508 svc_reserve(rqstp, 0);
509 svc_xprt_release_slot(rqstp);
510 rqstp->rq_xprt = NULL;
515 * Some svc_serv's will have occasional work to do, even when a xprt is not
516 * waiting to be serviced. This function is there to "kick" a task in one of
517 * those services so that it can wake up and do that work. Note that we only
518 * bother with pool 0 as we don't need to wake up more than one thread for
521 void svc_wake_up(struct svc_serv *serv)
523 struct svc_rqst *rqstp;
524 struct svc_pool *pool;
526 pool = &serv->sv_pools[0];
529 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
530 /* skip any that aren't queued */
531 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
534 wake_up_process(rqstp->rq_task);
535 trace_svc_wake_up(rqstp->rq_task->pid);
540 /* No free entries available */
541 set_bit(SP_TASK_PENDING, &pool->sp_flags);
543 trace_svc_wake_up(0);
545 EXPORT_SYMBOL_GPL(svc_wake_up);
547 int svc_port_is_privileged(struct sockaddr *sin)
549 switch (sin->sa_family) {
551 return ntohs(((struct sockaddr_in *)sin)->sin_port)
554 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
562 * Make sure that we don't have too many active connections. If we have,
563 * something must be dropped. It's not clear what will happen if we allow
564 * "too many" connections, but when dealing with network-facing software,
565 * we have to code defensively. Here we do that by imposing hard limits.
567 * There's no point in trying to do random drop here for DoS
568 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
569 * attacker can easily beat that.
571 * The only somewhat efficient mechanism would be if drop old
572 * connections from the same IP first. But right now we don't even
573 * record the client IP in svc_sock.
575 * single-threaded services that expect a lot of clients will probably
576 * need to set sv_maxconn to override the default value which is based
577 * on the number of threads
579 static void svc_check_conn_limits(struct svc_serv *serv)
581 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
582 (serv->sv_nrthreads+3) * 20;
584 if (serv->sv_tmpcnt > limit) {
585 struct svc_xprt *xprt = NULL;
586 spin_lock_bh(&serv->sv_lock);
587 if (!list_empty(&serv->sv_tempsocks)) {
588 /* Try to help the admin */
589 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
590 serv->sv_name, serv->sv_maxconn ?
591 "max number of connections" :
592 "number of threads");
594 * Always select the oldest connection. It's not fair,
597 xprt = list_entry(serv->sv_tempsocks.prev,
600 set_bit(XPT_CLOSE, &xprt->xpt_flags);
603 spin_unlock_bh(&serv->sv_lock);
606 svc_xprt_enqueue(xprt);
612 static int svc_alloc_arg(struct svc_rqst *rqstp)
614 struct svc_serv *serv = rqstp->rq_server;
619 /* now allocate needed pages. If we get a failure, sleep briefly */
620 pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
621 if (pages > RPCSVC_MAXPAGES) {
622 pr_warn_once("svc: warning: pages=%u > RPCSVC_MAXPAGES=%lu\n",
623 pages, RPCSVC_MAXPAGES);
624 /* use as many pages as possible */
625 pages = RPCSVC_MAXPAGES;
627 for (i = 0; i < pages ; i++)
628 while (rqstp->rq_pages[i] == NULL) {
629 struct page *p = alloc_page(GFP_KERNEL);
631 set_current_state(TASK_INTERRUPTIBLE);
632 if (signalled() || kthread_should_stop()) {
633 set_current_state(TASK_RUNNING);
636 schedule_timeout(msecs_to_jiffies(500));
638 rqstp->rq_pages[i] = p;
640 rqstp->rq_page_end = &rqstp->rq_pages[i];
641 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
643 /* Make arg->head point to first page and arg->pages point to rest */
644 arg = &rqstp->rq_arg;
645 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
646 arg->head[0].iov_len = PAGE_SIZE;
647 arg->pages = rqstp->rq_pages + 1;
649 /* save at least one page for response */
650 arg->page_len = (pages-2)*PAGE_SIZE;
651 arg->len = (pages-1)*PAGE_SIZE;
652 arg->tail[0].iov_len = 0;
657 rqst_should_sleep(struct svc_rqst *rqstp)
659 struct svc_pool *pool = rqstp->rq_pool;
661 /* did someone call svc_wake_up? */
662 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
665 /* was a socket queued? */
666 if (!list_empty(&pool->sp_sockets))
669 /* are we shutting down? */
670 if (signalled() || kthread_should_stop())
673 /* are we freezing? */
674 if (freezing(current))
680 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
682 struct svc_pool *pool = rqstp->rq_pool;
685 /* rq_xprt should be clear on entry */
686 WARN_ON_ONCE(rqstp->rq_xprt);
688 rqstp->rq_xprt = svc_xprt_dequeue(pool);
693 * We have to be able to interrupt this wait
694 * to bring down the daemons ...
696 set_current_state(TASK_INTERRUPTIBLE);
697 smp_mb__before_atomic();
698 clear_bit(SP_CONGESTED, &pool->sp_flags);
699 clear_bit(RQ_BUSY, &rqstp->rq_flags);
700 smp_mb__after_atomic();
702 if (likely(rqst_should_sleep(rqstp)))
703 time_left = schedule_timeout(timeout);
705 __set_current_state(TASK_RUNNING);
709 set_bit(RQ_BUSY, &rqstp->rq_flags);
710 smp_mb__after_atomic();
711 rqstp->rq_xprt = svc_xprt_dequeue(pool);
716 atomic_long_inc(&pool->sp_stats.threads_timedout);
718 if (signalled() || kthread_should_stop())
719 return ERR_PTR(-EINTR);
720 return ERR_PTR(-EAGAIN);
722 /* Normally we will wait up to 5 seconds for any required
723 * cache information to be provided.
725 if (!test_bit(SP_CONGESTED, &pool->sp_flags))
726 rqstp->rq_chandle.thread_wait = 5*HZ;
728 rqstp->rq_chandle.thread_wait = 1*HZ;
729 trace_svc_xprt_dequeue(rqstp);
730 return rqstp->rq_xprt;
733 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
735 spin_lock_bh(&serv->sv_lock);
736 set_bit(XPT_TEMP, &newxpt->xpt_flags);
737 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
739 if (serv->sv_temptimer.function == NULL) {
740 /* setup timer to age temp transports */
741 serv->sv_temptimer.function = svc_age_temp_xprts;
742 mod_timer(&serv->sv_temptimer,
743 jiffies + svc_conn_age_period * HZ);
745 spin_unlock_bh(&serv->sv_lock);
746 svc_xprt_received(newxpt);
749 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
751 struct svc_serv *serv = rqstp->rq_server;
754 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
755 dprintk("svc_recv: found XPT_CLOSE\n");
756 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
757 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
758 svc_delete_xprt(xprt);
759 /* Leave XPT_BUSY set on the dead xprt: */
762 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
763 struct svc_xprt *newxpt;
765 * We know this module_get will succeed because the
766 * listener holds a reference too
768 __module_get(xprt->xpt_class->xcl_owner);
769 svc_check_conn_limits(xprt->xpt_server);
770 newxpt = xprt->xpt_ops->xpo_accept(xprt);
772 svc_add_new_temp_xprt(serv, newxpt);
774 module_put(xprt->xpt_class->xcl_owner);
775 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
776 /* XPT_DATA|XPT_DEFERRED case: */
777 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
778 rqstp, rqstp->rq_pool->sp_id, xprt,
779 kref_read(&xprt->xpt_ref));
780 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
781 if (rqstp->rq_deferred)
782 len = svc_deferred_recv(rqstp);
784 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
785 rqstp->rq_stime = ktime_get();
786 rqstp->rq_reserved = serv->sv_max_mesg;
787 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
789 /* clear XPT_BUSY: */
790 svc_xprt_received(xprt);
792 trace_svc_handle_xprt(xprt, len);
797 * Receive the next request on any transport. This code is carefully
798 * organised not to touch any cachelines in the shared svc_serv
799 * structure, only cachelines in the local svc_pool.
801 int svc_recv(struct svc_rqst *rqstp, long timeout)
803 struct svc_xprt *xprt = NULL;
804 struct svc_serv *serv = rqstp->rq_server;
807 dprintk("svc: server %p waiting for data (to = %ld)\n",
812 "svc_recv: service %p, transport not NULL!\n",
815 err = svc_alloc_arg(rqstp);
822 if (signalled() || kthread_should_stop())
825 xprt = svc_get_next_xprt(rqstp, timeout);
831 len = svc_handle_xprt(rqstp, xprt);
833 /* No data, incomplete (TCP) read, or accept() */
838 clear_bit(XPT_OLD, &xprt->xpt_flags);
840 xprt->xpt_ops->xpo_secure_port(rqstp);
841 rqstp->rq_chandle.defer = svc_defer;
842 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
845 serv->sv_stats->netcnt++;
846 trace_svc_recv(rqstp, len);
849 rqstp->rq_res.len = 0;
850 svc_xprt_release(rqstp);
854 EXPORT_SYMBOL_GPL(svc_recv);
859 void svc_drop(struct svc_rqst *rqstp)
861 trace_svc_drop(rqstp);
862 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
863 svc_xprt_release(rqstp);
865 EXPORT_SYMBOL_GPL(svc_drop);
868 * Return reply to client.
870 int svc_send(struct svc_rqst *rqstp)
872 struct svc_xprt *xprt;
876 xprt = rqstp->rq_xprt;
880 /* release the receive skb before sending the reply */
881 xprt->xpt_ops->xpo_release_rqst(rqstp);
883 /* calculate over-all length */
885 xb->len = xb->head[0].iov_len +
889 /* Grab mutex to serialize outgoing data. */
890 mutex_lock(&xprt->xpt_mutex);
891 trace_svc_stats_latency(rqstp);
892 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
893 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
896 len = xprt->xpt_ops->xpo_sendto(rqstp);
897 mutex_unlock(&xprt->xpt_mutex);
898 rpc_wake_up(&xprt->xpt_bc_pending);
899 trace_svc_send(rqstp, len);
900 svc_xprt_release(rqstp);
902 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
909 * Timer function to close old temporary transports, using
910 * a mark-and-sweep algorithm.
912 static void svc_age_temp_xprts(struct timer_list *t)
914 struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
915 struct svc_xprt *xprt;
916 struct list_head *le, *next;
918 dprintk("svc_age_temp_xprts\n");
920 if (!spin_trylock_bh(&serv->sv_lock)) {
921 /* busy, try again 1 sec later */
922 dprintk("svc_age_temp_xprts: busy\n");
923 mod_timer(&serv->sv_temptimer, jiffies + HZ);
927 list_for_each_safe(le, next, &serv->sv_tempsocks) {
928 xprt = list_entry(le, struct svc_xprt, xpt_list);
930 /* First time through, just mark it OLD. Second time
931 * through, close it. */
932 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
934 if (kref_read(&xprt->xpt_ref) > 1 ||
935 test_bit(XPT_BUSY, &xprt->xpt_flags))
938 set_bit(XPT_CLOSE, &xprt->xpt_flags);
939 dprintk("queuing xprt %p for closing\n", xprt);
941 /* a thread will dequeue and close it soon */
942 svc_xprt_enqueue(xprt);
944 spin_unlock_bh(&serv->sv_lock);
946 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
949 /* Close temporary transports whose xpt_local matches server_addr immediately
950 * instead of waiting for them to be picked up by the timer.
952 * This is meant to be called from a notifier_block that runs when an ip
953 * address is deleted.
955 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
957 struct svc_xprt *xprt;
958 struct list_head *le, *next;
959 LIST_HEAD(to_be_closed);
961 spin_lock_bh(&serv->sv_lock);
962 list_for_each_safe(le, next, &serv->sv_tempsocks) {
963 xprt = list_entry(le, struct svc_xprt, xpt_list);
964 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
966 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
967 list_move(le, &to_be_closed);
970 spin_unlock_bh(&serv->sv_lock);
972 while (!list_empty(&to_be_closed)) {
973 le = to_be_closed.next;
975 xprt = list_entry(le, struct svc_xprt, xpt_list);
976 set_bit(XPT_CLOSE, &xprt->xpt_flags);
977 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
978 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
980 svc_xprt_enqueue(xprt);
983 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
985 static void call_xpt_users(struct svc_xprt *xprt)
987 struct svc_xpt_user *u;
989 spin_lock(&xprt->xpt_lock);
990 while (!list_empty(&xprt->xpt_users)) {
991 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
995 spin_unlock(&xprt->xpt_lock);
999 * Remove a dead transport
1001 static void svc_delete_xprt(struct svc_xprt *xprt)
1003 struct svc_serv *serv = xprt->xpt_server;
1004 struct svc_deferred_req *dr;
1006 /* Only do this once */
1007 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1010 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1011 xprt->xpt_ops->xpo_detach(xprt);
1013 spin_lock_bh(&serv->sv_lock);
1014 list_del_init(&xprt->xpt_list);
1015 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1016 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1018 spin_unlock_bh(&serv->sv_lock);
1020 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1023 call_xpt_users(xprt);
1027 void svc_close_xprt(struct svc_xprt *xprt)
1029 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1030 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1031 /* someone else will have to effect the close */
1034 * We expect svc_close_xprt() to work even when no threads are
1035 * running (e.g., while configuring the server before starting
1036 * any threads), so if the transport isn't busy, we delete
1039 svc_delete_xprt(xprt);
1041 EXPORT_SYMBOL_GPL(svc_close_xprt);
1043 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1045 struct svc_xprt *xprt;
1048 spin_lock(&serv->sv_lock);
1049 list_for_each_entry(xprt, xprt_list, xpt_list) {
1050 if (xprt->xpt_net != net)
1053 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1054 svc_xprt_enqueue(xprt);
1056 spin_unlock(&serv->sv_lock);
1060 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1062 struct svc_pool *pool;
1063 struct svc_xprt *xprt;
1064 struct svc_xprt *tmp;
1067 for (i = 0; i < serv->sv_nrpools; i++) {
1068 pool = &serv->sv_pools[i];
1070 spin_lock_bh(&pool->sp_lock);
1071 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1072 if (xprt->xpt_net != net)
1074 list_del_init(&xprt->xpt_ready);
1075 spin_unlock_bh(&pool->sp_lock);
1078 spin_unlock_bh(&pool->sp_lock);
1083 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1085 struct svc_xprt *xprt;
1087 while ((xprt = svc_dequeue_net(serv, net))) {
1088 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1089 svc_delete_xprt(xprt);
1094 * Server threads may still be running (especially in the case where the
1095 * service is still running in other network namespaces).
1097 * So we shut down sockets the same way we would on a running server, by
1098 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1099 * the close. In the case there are no such other threads,
1100 * threads running, svc_clean_up_xprts() does a simple version of a
1101 * server's main event loop, and in the case where there are other
1102 * threads, we may need to wait a little while and then check again to
1103 * see if they're done.
1105 void svc_close_net(struct svc_serv *serv, struct net *net)
1109 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1110 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1112 svc_clean_up_xprts(serv, net);
1118 * Handle defer and revisit of requests
1121 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1123 struct svc_deferred_req *dr =
1124 container_of(dreq, struct svc_deferred_req, handle);
1125 struct svc_xprt *xprt = dr->xprt;
1127 spin_lock(&xprt->xpt_lock);
1128 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1129 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1130 spin_unlock(&xprt->xpt_lock);
1131 dprintk("revisit canceled\n");
1133 trace_svc_drop_deferred(dr);
1137 dprintk("revisit queued\n");
1139 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1140 spin_unlock(&xprt->xpt_lock);
1141 svc_xprt_enqueue(xprt);
1146 * Save the request off for later processing. The request buffer looks
1149 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1151 * This code can only handle requests that consist of an xprt-header
1154 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1156 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1157 struct svc_deferred_req *dr;
1159 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1160 return NULL; /* if more than a page, give up FIXME */
1161 if (rqstp->rq_deferred) {
1162 dr = rqstp->rq_deferred;
1163 rqstp->rq_deferred = NULL;
1167 /* FIXME maybe discard if size too large */
1168 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1169 dr = kmalloc(size, GFP_KERNEL);
1173 dr->handle.owner = rqstp->rq_server;
1174 dr->prot = rqstp->rq_prot;
1175 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1176 dr->addrlen = rqstp->rq_addrlen;
1177 dr->daddr = rqstp->rq_daddr;
1178 dr->argslen = rqstp->rq_arg.len >> 2;
1179 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1181 /* back up head to the start of the buffer and copy */
1182 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1183 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1186 svc_xprt_get(rqstp->rq_xprt);
1187 dr->xprt = rqstp->rq_xprt;
1188 set_bit(RQ_DROPME, &rqstp->rq_flags);
1190 dr->handle.revisit = svc_revisit;
1191 trace_svc_defer(rqstp);
1196 * recv data from a deferred request into an active one
1198 static int svc_deferred_recv(struct svc_rqst *rqstp)
1200 struct svc_deferred_req *dr = rqstp->rq_deferred;
1202 /* setup iov_base past transport header */
1203 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1204 /* The iov_len does not include the transport header bytes */
1205 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1206 rqstp->rq_arg.page_len = 0;
1207 /* The rq_arg.len includes the transport header bytes */
1208 rqstp->rq_arg.len = dr->argslen<<2;
1209 rqstp->rq_prot = dr->prot;
1210 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1211 rqstp->rq_addrlen = dr->addrlen;
1212 /* Save off transport header len in case we get deferred again */
1213 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1214 rqstp->rq_daddr = dr->daddr;
1215 rqstp->rq_respages = rqstp->rq_pages;
1216 return (dr->argslen<<2) - dr->xprt_hlen;
1220 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1222 struct svc_deferred_req *dr = NULL;
1224 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1226 spin_lock(&xprt->xpt_lock);
1227 if (!list_empty(&xprt->xpt_deferred)) {
1228 dr = list_entry(xprt->xpt_deferred.next,
1229 struct svc_deferred_req,
1231 list_del_init(&dr->handle.recent);
1232 trace_svc_revisit_deferred(dr);
1234 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1235 spin_unlock(&xprt->xpt_lock);
1240 * svc_find_xprt - find an RPC transport instance
1241 * @serv: pointer to svc_serv to search
1242 * @xcl_name: C string containing transport's class name
1243 * @net: owner net pointer
1244 * @af: Address family of transport's local address
1245 * @port: transport's IP port number
1247 * Return the transport instance pointer for the endpoint accepting
1248 * connections/peer traffic from the specified transport class,
1249 * address family and port.
1251 * Specifying 0 for the address family or port is effectively a
1252 * wild-card, and will result in matching the first transport in the
1253 * service's list that has a matching class name.
1255 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1256 struct net *net, const sa_family_t af,
1257 const unsigned short port)
1259 struct svc_xprt *xprt;
1260 struct svc_xprt *found = NULL;
1262 /* Sanity check the args */
1263 if (serv == NULL || xcl_name == NULL)
1266 spin_lock_bh(&serv->sv_lock);
1267 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1268 if (xprt->xpt_net != net)
1270 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1272 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1274 if (port != 0 && port != svc_xprt_local_port(xprt))
1280 spin_unlock_bh(&serv->sv_lock);
1283 EXPORT_SYMBOL_GPL(svc_find_xprt);
1285 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1286 char *pos, int remaining)
1290 len = snprintf(pos, remaining, "%s %u\n",
1291 xprt->xpt_class->xcl_name,
1292 svc_xprt_local_port(xprt));
1293 if (len >= remaining)
1294 return -ENAMETOOLONG;
1299 * svc_xprt_names - format a buffer with a list of transport names
1300 * @serv: pointer to an RPC service
1301 * @buf: pointer to a buffer to be filled in
1302 * @buflen: length of buffer to be filled in
1304 * Fills in @buf with a string containing a list of transport names,
1305 * each name terminated with '\n'.
1307 * Returns positive length of the filled-in string on success; otherwise
1308 * a negative errno value is returned if an error occurs.
1310 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1312 struct svc_xprt *xprt;
1316 /* Sanity check args */
1320 spin_lock_bh(&serv->sv_lock);
1324 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1325 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1337 spin_unlock_bh(&serv->sv_lock);
1340 EXPORT_SYMBOL_GPL(svc_xprt_names);
1343 /*----------------------------------------------------------------------------*/
1345 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1347 unsigned int pidx = (unsigned int)*pos;
1348 struct svc_serv *serv = m->private;
1350 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1353 return SEQ_START_TOKEN;
1354 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1357 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1359 struct svc_pool *pool = p;
1360 struct svc_serv *serv = m->private;
1362 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1364 if (p == SEQ_START_TOKEN) {
1365 pool = &serv->sv_pools[0];
1367 unsigned int pidx = (pool - &serv->sv_pools[0]);
1368 if (pidx < serv->sv_nrpools-1)
1369 pool = &serv->sv_pools[pidx+1];
1377 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1381 static int svc_pool_stats_show(struct seq_file *m, void *p)
1383 struct svc_pool *pool = p;
1385 if (p == SEQ_START_TOKEN) {
1386 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1390 seq_printf(m, "%u %lu %lu %lu %lu\n",
1392 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1393 pool->sp_stats.sockets_queued,
1394 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1395 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1400 static const struct seq_operations svc_pool_stats_seq_ops = {
1401 .start = svc_pool_stats_start,
1402 .next = svc_pool_stats_next,
1403 .stop = svc_pool_stats_stop,
1404 .show = svc_pool_stats_show,
1407 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1411 err = seq_open(file, &svc_pool_stats_seq_ops);
1413 ((struct seq_file *) file->private_data)->private = serv;
1416 EXPORT_SYMBOL(svc_pool_stats_open);
1418 /*----------------------------------------------------------------------------*/