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 * svc_print_xprts - Format the transport list for printing
108 * @buf: target buffer for formatted address
109 * @maxlen: length of target buffer
111 * Fills in @buf with a string containing a list of transport names, each name
112 * terminated with '\n'. If the buffer is too small, some entries may be
113 * missing, but it is guaranteed that all lines in the output buffer are
116 * Returns positive length of the filled-in string.
118 int svc_print_xprts(char *buf, int maxlen)
120 struct svc_xprt_class *xcl;
125 spin_lock(&svc_xprt_class_lock);
126 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
129 slen = snprintf(tmpstr, sizeof(tmpstr), "%s %d\n",
130 xcl->xcl_name, xcl->xcl_max_payload);
131 if (slen >= sizeof(tmpstr) || len + slen >= maxlen)
136 spin_unlock(&svc_xprt_class_lock);
141 static void svc_xprt_free(struct kref *kref)
143 struct svc_xprt *xprt =
144 container_of(kref, struct svc_xprt, xpt_ref);
145 struct module *owner = xprt->xpt_class->xcl_owner;
146 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
147 svcauth_unix_info_release(xprt);
148 put_net(xprt->xpt_net);
149 /* See comment on corresponding get in xs_setup_bc_tcp(): */
150 if (xprt->xpt_bc_xprt)
151 xprt_put(xprt->xpt_bc_xprt);
152 if (xprt->xpt_bc_xps)
153 xprt_switch_put(xprt->xpt_bc_xps);
154 xprt->xpt_ops->xpo_free(xprt);
158 void svc_xprt_put(struct svc_xprt *xprt)
160 kref_put(&xprt->xpt_ref, svc_xprt_free);
162 EXPORT_SYMBOL_GPL(svc_xprt_put);
165 * Called by transport drivers to initialize the transport independent
166 * portion of the transport instance.
168 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
169 struct svc_xprt *xprt, struct svc_serv *serv)
171 memset(xprt, 0, sizeof(*xprt));
172 xprt->xpt_class = xcl;
173 xprt->xpt_ops = xcl->xcl_ops;
174 kref_init(&xprt->xpt_ref);
175 xprt->xpt_server = serv;
176 INIT_LIST_HEAD(&xprt->xpt_list);
177 INIT_LIST_HEAD(&xprt->xpt_ready);
178 INIT_LIST_HEAD(&xprt->xpt_deferred);
179 INIT_LIST_HEAD(&xprt->xpt_users);
180 mutex_init(&xprt->xpt_mutex);
181 spin_lock_init(&xprt->xpt_lock);
182 set_bit(XPT_BUSY, &xprt->xpt_flags);
183 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
184 xprt->xpt_net = get_net(net);
185 strcpy(xprt->xpt_remotebuf, "uninitialized");
187 EXPORT_SYMBOL_GPL(svc_xprt_init);
189 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
190 struct svc_serv *serv,
193 const unsigned short port,
196 struct sockaddr_in sin = {
197 .sin_family = AF_INET,
198 .sin_addr.s_addr = htonl(INADDR_ANY),
199 .sin_port = htons(port),
201 #if IS_ENABLED(CONFIG_IPV6)
202 struct sockaddr_in6 sin6 = {
203 .sin6_family = AF_INET6,
204 .sin6_addr = IN6ADDR_ANY_INIT,
205 .sin6_port = htons(port),
208 struct sockaddr *sap;
213 sap = (struct sockaddr *)&sin;
216 #if IS_ENABLED(CONFIG_IPV6)
218 sap = (struct sockaddr *)&sin6;
223 return ERR_PTR(-EAFNOSUPPORT);
226 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
230 * svc_xprt_received conditionally queues the transport for processing
231 * by another thread. The caller must hold the XPT_BUSY bit and must
232 * not thereafter touch transport data.
234 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
235 * insufficient) data.
237 static void svc_xprt_received(struct svc_xprt *xprt)
239 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
240 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
244 /* As soon as we clear busy, the xprt could be closed and
245 * 'put', so we need a reference to call svc_enqueue_xprt with:
248 smp_mb__before_atomic();
249 clear_bit(XPT_BUSY, &xprt->xpt_flags);
250 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
254 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
256 clear_bit(XPT_TEMP, &new->xpt_flags);
257 spin_lock_bh(&serv->sv_lock);
258 list_add(&new->xpt_list, &serv->sv_permsocks);
259 spin_unlock_bh(&serv->sv_lock);
260 svc_xprt_received(new);
263 static int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
264 struct net *net, const int family,
265 const unsigned short port, int flags)
267 struct svc_xprt_class *xcl;
269 spin_lock(&svc_xprt_class_lock);
270 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
271 struct svc_xprt *newxprt;
272 unsigned short newport;
274 if (strcmp(xprt_name, xcl->xcl_name))
277 if (!try_module_get(xcl->xcl_owner))
280 spin_unlock(&svc_xprt_class_lock);
281 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
282 if (IS_ERR(newxprt)) {
283 module_put(xcl->xcl_owner);
284 return PTR_ERR(newxprt);
286 svc_add_new_perm_xprt(serv, newxprt);
287 newport = svc_xprt_local_port(newxprt);
291 spin_unlock(&svc_xprt_class_lock);
292 /* This errno is exposed to user space. Provide a reasonable
293 * perror msg for a bad transport. */
294 return -EPROTONOSUPPORT;
297 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
298 struct net *net, const int family,
299 const unsigned short port, int flags)
303 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
304 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
305 if (err == -EPROTONOSUPPORT) {
306 request_module("svc%s", xprt_name);
307 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags);
310 dprintk("svc: transport %s not found, err %d\n",
314 EXPORT_SYMBOL_GPL(svc_create_xprt);
317 * Copy the local and remote xprt addresses to the rqstp structure
319 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
321 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
322 rqstp->rq_addrlen = xprt->xpt_remotelen;
325 * Destination address in request is needed for binding the
326 * source address in RPC replies/callbacks later.
328 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
329 rqstp->rq_daddrlen = xprt->xpt_locallen;
331 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
334 * svc_print_addr - Format rq_addr field for printing
335 * @rqstp: svc_rqst struct containing address to print
336 * @buf: target buffer for formatted address
337 * @len: length of target buffer
340 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
342 return __svc_print_addr(svc_addr(rqstp), buf, len);
344 EXPORT_SYMBOL_GPL(svc_print_addr);
346 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
348 unsigned int limit = svc_rpc_per_connection_limit;
349 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
351 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
354 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
356 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
357 if (!svc_xprt_slots_in_range(xprt))
359 atomic_inc(&xprt->xpt_nr_rqsts);
360 set_bit(RQ_DATA, &rqstp->rq_flags);
365 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
367 struct svc_xprt *xprt = rqstp->rq_xprt;
368 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
369 atomic_dec(&xprt->xpt_nr_rqsts);
370 svc_xprt_enqueue(xprt);
374 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
376 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
378 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED))) {
379 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
380 svc_xprt_slots_in_range(xprt))
382 trace_svc_xprt_no_write_space(xprt);
388 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
390 struct svc_pool *pool;
391 struct svc_rqst *rqstp = NULL;
394 if (!svc_xprt_has_something_to_do(xprt))
397 /* Mark transport as busy. It will remain in this state until
398 * the provider calls svc_xprt_received. We update XPT_BUSY
399 * atomically because it also guards against trying to enqueue
400 * the transport twice.
402 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
406 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
408 atomic_long_inc(&pool->sp_stats.packets);
410 spin_lock_bh(&pool->sp_lock);
411 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
412 pool->sp_stats.sockets_queued++;
413 spin_unlock_bh(&pool->sp_lock);
415 /* find a thread for this xprt */
417 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
418 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
420 atomic_long_inc(&pool->sp_stats.threads_woken);
421 rqstp->rq_qtime = ktime_get();
422 wake_up_process(rqstp->rq_task);
425 set_bit(SP_CONGESTED, &pool->sp_flags);
430 trace_svc_xprt_do_enqueue(xprt, rqstp);
432 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
435 * Queue up a transport with data pending. If there are idle nfsd
436 * processes, wake 'em up.
439 void svc_xprt_enqueue(struct svc_xprt *xprt)
441 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
443 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
445 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
448 * Dequeue the first transport, if there is one.
450 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
452 struct svc_xprt *xprt = NULL;
454 if (list_empty(&pool->sp_sockets))
457 spin_lock_bh(&pool->sp_lock);
458 if (likely(!list_empty(&pool->sp_sockets))) {
459 xprt = list_first_entry(&pool->sp_sockets,
460 struct svc_xprt, xpt_ready);
461 list_del_init(&xprt->xpt_ready);
464 spin_unlock_bh(&pool->sp_lock);
470 * svc_reserve - change the space reserved for the reply to a request.
471 * @rqstp: The request in question
472 * @space: new max space to reserve
474 * Each request reserves some space on the output queue of the transport
475 * to make sure the reply fits. This function reduces that reserved
476 * space to be the amount of space used already, plus @space.
479 void svc_reserve(struct svc_rqst *rqstp, int space)
481 struct svc_xprt *xprt = rqstp->rq_xprt;
483 space += rqstp->rq_res.head[0].iov_len;
485 if (xprt && space < rqstp->rq_reserved) {
486 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
487 rqstp->rq_reserved = space;
489 svc_xprt_enqueue(xprt);
492 EXPORT_SYMBOL_GPL(svc_reserve);
494 static void svc_xprt_release(struct svc_rqst *rqstp)
496 struct svc_xprt *xprt = rqstp->rq_xprt;
498 xprt->xpt_ops->xpo_release_rqst(rqstp);
500 kfree(rqstp->rq_deferred);
501 rqstp->rq_deferred = NULL;
503 svc_free_res_pages(rqstp);
504 rqstp->rq_res.page_len = 0;
505 rqstp->rq_res.page_base = 0;
507 /* Reset response buffer and release
509 * But first, check that enough space was reserved
510 * for the reply, otherwise we have a bug!
512 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
513 printk(KERN_ERR "RPC request reserved %d but used %d\n",
517 rqstp->rq_res.head[0].iov_len = 0;
518 svc_reserve(rqstp, 0);
519 svc_xprt_release_slot(rqstp);
520 rqstp->rq_xprt = NULL;
525 * Some svc_serv's will have occasional work to do, even when a xprt is not
526 * waiting to be serviced. This function is there to "kick" a task in one of
527 * those services so that it can wake up and do that work. Note that we only
528 * bother with pool 0 as we don't need to wake up more than one thread for
531 void svc_wake_up(struct svc_serv *serv)
533 struct svc_rqst *rqstp;
534 struct svc_pool *pool;
536 pool = &serv->sv_pools[0];
539 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
540 /* skip any that aren't queued */
541 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
544 wake_up_process(rqstp->rq_task);
545 trace_svc_wake_up(rqstp->rq_task->pid);
550 /* No free entries available */
551 set_bit(SP_TASK_PENDING, &pool->sp_flags);
553 trace_svc_wake_up(0);
555 EXPORT_SYMBOL_GPL(svc_wake_up);
557 int svc_port_is_privileged(struct sockaddr *sin)
559 switch (sin->sa_family) {
561 return ntohs(((struct sockaddr_in *)sin)->sin_port)
564 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
572 * Make sure that we don't have too many active connections. If we have,
573 * something must be dropped. It's not clear what will happen if we allow
574 * "too many" connections, but when dealing with network-facing software,
575 * we have to code defensively. Here we do that by imposing hard limits.
577 * There's no point in trying to do random drop here for DoS
578 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
579 * attacker can easily beat that.
581 * The only somewhat efficient mechanism would be if drop old
582 * connections from the same IP first. But right now we don't even
583 * record the client IP in svc_sock.
585 * single-threaded services that expect a lot of clients will probably
586 * need to set sv_maxconn to override the default value which is based
587 * on the number of threads
589 static void svc_check_conn_limits(struct svc_serv *serv)
591 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
592 (serv->sv_nrthreads+3) * 20;
594 if (serv->sv_tmpcnt > limit) {
595 struct svc_xprt *xprt = NULL;
596 spin_lock_bh(&serv->sv_lock);
597 if (!list_empty(&serv->sv_tempsocks)) {
598 /* Try to help the admin */
599 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
600 serv->sv_name, serv->sv_maxconn ?
601 "max number of connections" :
602 "number of threads");
604 * Always select the oldest connection. It's not fair,
607 xprt = list_entry(serv->sv_tempsocks.prev,
610 set_bit(XPT_CLOSE, &xprt->xpt_flags);
613 spin_unlock_bh(&serv->sv_lock);
616 svc_xprt_enqueue(xprt);
622 static int svc_alloc_arg(struct svc_rqst *rqstp)
624 struct svc_serv *serv = rqstp->rq_server;
629 /* now allocate needed pages. If we get a failure, sleep briefly */
630 pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
631 if (pages > RPCSVC_MAXPAGES) {
632 pr_warn_once("svc: warning: pages=%u > RPCSVC_MAXPAGES=%lu\n",
633 pages, RPCSVC_MAXPAGES);
634 /* use as many pages as possible */
635 pages = RPCSVC_MAXPAGES;
637 for (i = 0; i < pages ; i++)
638 while (rqstp->rq_pages[i] == NULL) {
639 struct page *p = alloc_page(GFP_KERNEL);
641 set_current_state(TASK_INTERRUPTIBLE);
642 if (signalled() || kthread_should_stop()) {
643 set_current_state(TASK_RUNNING);
646 schedule_timeout(msecs_to_jiffies(500));
648 rqstp->rq_pages[i] = p;
650 rqstp->rq_page_end = &rqstp->rq_pages[i];
651 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
653 /* Make arg->head point to first page and arg->pages point to rest */
654 arg = &rqstp->rq_arg;
655 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
656 arg->head[0].iov_len = PAGE_SIZE;
657 arg->pages = rqstp->rq_pages + 1;
659 /* save at least one page for response */
660 arg->page_len = (pages-2)*PAGE_SIZE;
661 arg->len = (pages-1)*PAGE_SIZE;
662 arg->tail[0].iov_len = 0;
667 rqst_should_sleep(struct svc_rqst *rqstp)
669 struct svc_pool *pool = rqstp->rq_pool;
671 /* did someone call svc_wake_up? */
672 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
675 /* was a socket queued? */
676 if (!list_empty(&pool->sp_sockets))
679 /* are we shutting down? */
680 if (signalled() || kthread_should_stop())
683 /* are we freezing? */
684 if (freezing(current))
690 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
692 struct svc_pool *pool = rqstp->rq_pool;
695 /* rq_xprt should be clear on entry */
696 WARN_ON_ONCE(rqstp->rq_xprt);
698 rqstp->rq_xprt = svc_xprt_dequeue(pool);
703 * We have to be able to interrupt this wait
704 * to bring down the daemons ...
706 set_current_state(TASK_INTERRUPTIBLE);
707 smp_mb__before_atomic();
708 clear_bit(SP_CONGESTED, &pool->sp_flags);
709 clear_bit(RQ_BUSY, &rqstp->rq_flags);
710 smp_mb__after_atomic();
712 if (likely(rqst_should_sleep(rqstp)))
713 time_left = schedule_timeout(timeout);
715 __set_current_state(TASK_RUNNING);
719 set_bit(RQ_BUSY, &rqstp->rq_flags);
720 smp_mb__after_atomic();
721 rqstp->rq_xprt = svc_xprt_dequeue(pool);
726 atomic_long_inc(&pool->sp_stats.threads_timedout);
728 if (signalled() || kthread_should_stop())
729 return ERR_PTR(-EINTR);
730 return ERR_PTR(-EAGAIN);
732 /* Normally we will wait up to 5 seconds for any required
733 * cache information to be provided.
735 if (!test_bit(SP_CONGESTED, &pool->sp_flags))
736 rqstp->rq_chandle.thread_wait = 5*HZ;
738 rqstp->rq_chandle.thread_wait = 1*HZ;
739 trace_svc_xprt_dequeue(rqstp);
740 return rqstp->rq_xprt;
743 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
745 spin_lock_bh(&serv->sv_lock);
746 set_bit(XPT_TEMP, &newxpt->xpt_flags);
747 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
749 if (serv->sv_temptimer.function == NULL) {
750 /* setup timer to age temp transports */
751 serv->sv_temptimer.function = svc_age_temp_xprts;
752 mod_timer(&serv->sv_temptimer,
753 jiffies + svc_conn_age_period * HZ);
755 spin_unlock_bh(&serv->sv_lock);
756 svc_xprt_received(newxpt);
759 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
761 struct svc_serv *serv = rqstp->rq_server;
764 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
765 dprintk("svc_recv: found XPT_CLOSE\n");
766 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
767 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
768 svc_delete_xprt(xprt);
769 /* Leave XPT_BUSY set on the dead xprt: */
772 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
773 struct svc_xprt *newxpt;
775 * We know this module_get will succeed because the
776 * listener holds a reference too
778 __module_get(xprt->xpt_class->xcl_owner);
779 svc_check_conn_limits(xprt->xpt_server);
780 newxpt = xprt->xpt_ops->xpo_accept(xprt);
782 svc_add_new_temp_xprt(serv, newxpt);
784 module_put(xprt->xpt_class->xcl_owner);
785 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
786 /* XPT_DATA|XPT_DEFERRED case: */
787 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
788 rqstp, rqstp->rq_pool->sp_id, xprt,
789 kref_read(&xprt->xpt_ref));
790 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
791 if (rqstp->rq_deferred)
792 len = svc_deferred_recv(rqstp);
794 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
795 rqstp->rq_stime = ktime_get();
796 rqstp->rq_reserved = serv->sv_max_mesg;
797 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
799 /* clear XPT_BUSY: */
800 svc_xprt_received(xprt);
802 trace_svc_handle_xprt(xprt, len);
807 * Receive the next request on any transport. This code is carefully
808 * organised not to touch any cachelines in the shared svc_serv
809 * structure, only cachelines in the local svc_pool.
811 int svc_recv(struct svc_rqst *rqstp, long timeout)
813 struct svc_xprt *xprt = NULL;
814 struct svc_serv *serv = rqstp->rq_server;
817 dprintk("svc: server %p waiting for data (to = %ld)\n",
822 "svc_recv: service %p, transport not NULL!\n",
825 err = svc_alloc_arg(rqstp);
832 if (signalled() || kthread_should_stop())
835 xprt = svc_get_next_xprt(rqstp, timeout);
841 len = svc_handle_xprt(rqstp, xprt);
843 /* No data, incomplete (TCP) read, or accept() */
848 clear_bit(XPT_OLD, &xprt->xpt_flags);
850 xprt->xpt_ops->xpo_secure_port(rqstp);
851 rqstp->rq_chandle.defer = svc_defer;
852 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
855 serv->sv_stats->netcnt++;
856 trace_svc_recv(rqstp, len);
859 rqstp->rq_res.len = 0;
860 svc_xprt_release(rqstp);
864 EXPORT_SYMBOL_GPL(svc_recv);
869 void svc_drop(struct svc_rqst *rqstp)
871 trace_svc_drop(rqstp);
872 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
873 svc_xprt_release(rqstp);
875 EXPORT_SYMBOL_GPL(svc_drop);
878 * Return reply to client.
880 int svc_send(struct svc_rqst *rqstp)
882 struct svc_xprt *xprt;
886 xprt = rqstp->rq_xprt;
890 /* calculate over-all length */
892 xb->len = xb->head[0].iov_len +
896 /* Grab mutex to serialize outgoing data. */
897 mutex_lock(&xprt->xpt_mutex);
898 trace_svc_stats_latency(rqstp);
899 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
900 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
903 len = xprt->xpt_ops->xpo_sendto(rqstp);
904 mutex_unlock(&xprt->xpt_mutex);
905 rpc_wake_up(&xprt->xpt_bc_pending);
906 trace_svc_send(rqstp, len);
907 svc_xprt_release(rqstp);
909 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
916 * Timer function to close old temporary transports, using
917 * a mark-and-sweep algorithm.
919 static void svc_age_temp_xprts(struct timer_list *t)
921 struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
922 struct svc_xprt *xprt;
923 struct list_head *le, *next;
925 dprintk("svc_age_temp_xprts\n");
927 if (!spin_trylock_bh(&serv->sv_lock)) {
928 /* busy, try again 1 sec later */
929 dprintk("svc_age_temp_xprts: busy\n");
930 mod_timer(&serv->sv_temptimer, jiffies + HZ);
934 list_for_each_safe(le, next, &serv->sv_tempsocks) {
935 xprt = list_entry(le, struct svc_xprt, xpt_list);
937 /* First time through, just mark it OLD. Second time
938 * through, close it. */
939 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
941 if (kref_read(&xprt->xpt_ref) > 1 ||
942 test_bit(XPT_BUSY, &xprt->xpt_flags))
945 set_bit(XPT_CLOSE, &xprt->xpt_flags);
946 dprintk("queuing xprt %p for closing\n", xprt);
948 /* a thread will dequeue and close it soon */
949 svc_xprt_enqueue(xprt);
951 spin_unlock_bh(&serv->sv_lock);
953 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
956 /* Close temporary transports whose xpt_local matches server_addr immediately
957 * instead of waiting for them to be picked up by the timer.
959 * This is meant to be called from a notifier_block that runs when an ip
960 * address is deleted.
962 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
964 struct svc_xprt *xprt;
965 struct list_head *le, *next;
966 LIST_HEAD(to_be_closed);
968 spin_lock_bh(&serv->sv_lock);
969 list_for_each_safe(le, next, &serv->sv_tempsocks) {
970 xprt = list_entry(le, struct svc_xprt, xpt_list);
971 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
973 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
974 list_move(le, &to_be_closed);
977 spin_unlock_bh(&serv->sv_lock);
979 while (!list_empty(&to_be_closed)) {
980 le = to_be_closed.next;
982 xprt = list_entry(le, struct svc_xprt, xpt_list);
983 set_bit(XPT_CLOSE, &xprt->xpt_flags);
984 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
985 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
987 svc_xprt_enqueue(xprt);
990 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
992 static void call_xpt_users(struct svc_xprt *xprt)
994 struct svc_xpt_user *u;
996 spin_lock(&xprt->xpt_lock);
997 while (!list_empty(&xprt->xpt_users)) {
998 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
999 list_del_init(&u->list);
1002 spin_unlock(&xprt->xpt_lock);
1006 * Remove a dead transport
1008 static void svc_delete_xprt(struct svc_xprt *xprt)
1010 struct svc_serv *serv = xprt->xpt_server;
1011 struct svc_deferred_req *dr;
1013 /* Only do this once */
1014 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1017 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1018 xprt->xpt_ops->xpo_detach(xprt);
1020 spin_lock_bh(&serv->sv_lock);
1021 list_del_init(&xprt->xpt_list);
1022 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1023 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1025 spin_unlock_bh(&serv->sv_lock);
1027 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1030 call_xpt_users(xprt);
1034 void svc_close_xprt(struct svc_xprt *xprt)
1036 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1037 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1038 /* someone else will have to effect the close */
1041 * We expect svc_close_xprt() to work even when no threads are
1042 * running (e.g., while configuring the server before starting
1043 * any threads), so if the transport isn't busy, we delete
1046 svc_delete_xprt(xprt);
1048 EXPORT_SYMBOL_GPL(svc_close_xprt);
1050 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1052 struct svc_xprt *xprt;
1055 spin_lock_bh(&serv->sv_lock);
1056 list_for_each_entry(xprt, xprt_list, xpt_list) {
1057 if (xprt->xpt_net != net)
1060 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1061 svc_xprt_enqueue(xprt);
1063 spin_unlock_bh(&serv->sv_lock);
1067 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1069 struct svc_pool *pool;
1070 struct svc_xprt *xprt;
1071 struct svc_xprt *tmp;
1074 for (i = 0; i < serv->sv_nrpools; i++) {
1075 pool = &serv->sv_pools[i];
1077 spin_lock_bh(&pool->sp_lock);
1078 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1079 if (xprt->xpt_net != net)
1081 list_del_init(&xprt->xpt_ready);
1082 spin_unlock_bh(&pool->sp_lock);
1085 spin_unlock_bh(&pool->sp_lock);
1090 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1092 struct svc_xprt *xprt;
1094 while ((xprt = svc_dequeue_net(serv, net))) {
1095 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1096 svc_delete_xprt(xprt);
1101 * Server threads may still be running (especially in the case where the
1102 * service is still running in other network namespaces).
1104 * So we shut down sockets the same way we would on a running server, by
1105 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1106 * the close. In the case there are no such other threads,
1107 * threads running, svc_clean_up_xprts() does a simple version of a
1108 * server's main event loop, and in the case where there are other
1109 * threads, we may need to wait a little while and then check again to
1110 * see if they're done.
1112 void svc_close_net(struct svc_serv *serv, struct net *net)
1116 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1117 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1119 svc_clean_up_xprts(serv, net);
1125 * Handle defer and revisit of requests
1128 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1130 struct svc_deferred_req *dr =
1131 container_of(dreq, struct svc_deferred_req, handle);
1132 struct svc_xprt *xprt = dr->xprt;
1134 spin_lock(&xprt->xpt_lock);
1135 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1136 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1137 spin_unlock(&xprt->xpt_lock);
1138 dprintk("revisit canceled\n");
1140 trace_svc_drop_deferred(dr);
1144 dprintk("revisit queued\n");
1146 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1147 spin_unlock(&xprt->xpt_lock);
1148 svc_xprt_enqueue(xprt);
1153 * Save the request off for later processing. The request buffer looks
1156 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1158 * This code can only handle requests that consist of an xprt-header
1161 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1163 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1164 struct svc_deferred_req *dr;
1166 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1167 return NULL; /* if more than a page, give up FIXME */
1168 if (rqstp->rq_deferred) {
1169 dr = rqstp->rq_deferred;
1170 rqstp->rq_deferred = NULL;
1174 /* FIXME maybe discard if size too large */
1175 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1176 dr = kmalloc(size, GFP_KERNEL);
1180 dr->handle.owner = rqstp->rq_server;
1181 dr->prot = rqstp->rq_prot;
1182 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1183 dr->addrlen = rqstp->rq_addrlen;
1184 dr->daddr = rqstp->rq_daddr;
1185 dr->argslen = rqstp->rq_arg.len >> 2;
1186 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1188 /* back up head to the start of the buffer and copy */
1189 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1190 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1193 svc_xprt_get(rqstp->rq_xprt);
1194 dr->xprt = rqstp->rq_xprt;
1195 set_bit(RQ_DROPME, &rqstp->rq_flags);
1197 dr->handle.revisit = svc_revisit;
1198 trace_svc_defer(rqstp);
1203 * recv data from a deferred request into an active one
1205 static int svc_deferred_recv(struct svc_rqst *rqstp)
1207 struct svc_deferred_req *dr = rqstp->rq_deferred;
1209 /* setup iov_base past transport header */
1210 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1211 /* The iov_len does not include the transport header bytes */
1212 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1213 rqstp->rq_arg.page_len = 0;
1214 /* The rq_arg.len includes the transport header bytes */
1215 rqstp->rq_arg.len = dr->argslen<<2;
1216 rqstp->rq_prot = dr->prot;
1217 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1218 rqstp->rq_addrlen = dr->addrlen;
1219 /* Save off transport header len in case we get deferred again */
1220 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1221 rqstp->rq_daddr = dr->daddr;
1222 rqstp->rq_respages = rqstp->rq_pages;
1223 return (dr->argslen<<2) - dr->xprt_hlen;
1227 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1229 struct svc_deferred_req *dr = NULL;
1231 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1233 spin_lock(&xprt->xpt_lock);
1234 if (!list_empty(&xprt->xpt_deferred)) {
1235 dr = list_entry(xprt->xpt_deferred.next,
1236 struct svc_deferred_req,
1238 list_del_init(&dr->handle.recent);
1239 trace_svc_revisit_deferred(dr);
1241 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1242 spin_unlock(&xprt->xpt_lock);
1247 * svc_find_xprt - find an RPC transport instance
1248 * @serv: pointer to svc_serv to search
1249 * @xcl_name: C string containing transport's class name
1250 * @net: owner net pointer
1251 * @af: Address family of transport's local address
1252 * @port: transport's IP port number
1254 * Return the transport instance pointer for the endpoint accepting
1255 * connections/peer traffic from the specified transport class,
1256 * address family and port.
1258 * Specifying 0 for the address family or port is effectively a
1259 * wild-card, and will result in matching the first transport in the
1260 * service's list that has a matching class name.
1262 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1263 struct net *net, const sa_family_t af,
1264 const unsigned short port)
1266 struct svc_xprt *xprt;
1267 struct svc_xprt *found = NULL;
1269 /* Sanity check the args */
1270 if (serv == NULL || xcl_name == NULL)
1273 spin_lock_bh(&serv->sv_lock);
1274 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1275 if (xprt->xpt_net != net)
1277 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1279 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1281 if (port != 0 && port != svc_xprt_local_port(xprt))
1287 spin_unlock_bh(&serv->sv_lock);
1290 EXPORT_SYMBOL_GPL(svc_find_xprt);
1292 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1293 char *pos, int remaining)
1297 len = snprintf(pos, remaining, "%s %u\n",
1298 xprt->xpt_class->xcl_name,
1299 svc_xprt_local_port(xprt));
1300 if (len >= remaining)
1301 return -ENAMETOOLONG;
1306 * svc_xprt_names - format a buffer with a list of transport names
1307 * @serv: pointer to an RPC service
1308 * @buf: pointer to a buffer to be filled in
1309 * @buflen: length of buffer to be filled in
1311 * Fills in @buf with a string containing a list of transport names,
1312 * each name terminated with '\n'.
1314 * Returns positive length of the filled-in string on success; otherwise
1315 * a negative errno value is returned if an error occurs.
1317 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1319 struct svc_xprt *xprt;
1323 /* Sanity check args */
1327 spin_lock_bh(&serv->sv_lock);
1331 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1332 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1344 spin_unlock_bh(&serv->sv_lock);
1347 EXPORT_SYMBOL_GPL(svc_xprt_names);
1350 /*----------------------------------------------------------------------------*/
1352 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1354 unsigned int pidx = (unsigned int)*pos;
1355 struct svc_serv *serv = m->private;
1357 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1360 return SEQ_START_TOKEN;
1361 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1364 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1366 struct svc_pool *pool = p;
1367 struct svc_serv *serv = m->private;
1369 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1371 if (p == SEQ_START_TOKEN) {
1372 pool = &serv->sv_pools[0];
1374 unsigned int pidx = (pool - &serv->sv_pools[0]);
1375 if (pidx < serv->sv_nrpools-1)
1376 pool = &serv->sv_pools[pidx+1];
1384 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1388 static int svc_pool_stats_show(struct seq_file *m, void *p)
1390 struct svc_pool *pool = p;
1392 if (p == SEQ_START_TOKEN) {
1393 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1397 seq_printf(m, "%u %lu %lu %lu %lu\n",
1399 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1400 pool->sp_stats.sockets_queued,
1401 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1402 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1407 static const struct seq_operations svc_pool_stats_seq_ops = {
1408 .start = svc_pool_stats_start,
1409 .next = svc_pool_stats_next,
1410 .stop = svc_pool_stats_stop,
1411 .show = svc_pool_stats_show,
1414 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1418 err = seq_open(file, &svc_pool_stats_seq_ops);
1420 ((struct seq_file *) file->private_data)->private = serv;
1423 EXPORT_SYMBOL(svc_pool_stats_open);
1425 /*----------------------------------------------------------------------------*/