1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/net/sunrpc/svc_xprt.c
5 * Author: Tom Tucker <tom@opengridcomputing.com>
8 #include <linux/sched.h>
9 #include <linux/errno.h>
10 #include <linux/freezer.h>
11 #include <linux/kthread.h>
12 #include <linux/slab.h>
14 #include <linux/sunrpc/addr.h>
15 #include <linux/sunrpc/stats.h>
16 #include <linux/sunrpc/svc_xprt.h>
17 #include <linux/sunrpc/svcsock.h>
18 #include <linux/sunrpc/xprt.h>
19 #include <linux/module.h>
20 #include <linux/netdevice.h>
21 #include <trace/events/sunrpc.h>
23 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
25 static unsigned int svc_rpc_per_connection_limit __read_mostly;
26 module_param(svc_rpc_per_connection_limit, uint, 0644);
29 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
30 static int svc_deferred_recv(struct svc_rqst *rqstp);
31 static struct cache_deferred_req *svc_defer(struct cache_req *req);
32 static void svc_age_temp_xprts(struct timer_list *t);
33 static void svc_delete_xprt(struct svc_xprt *xprt);
35 /* apparently the "standard" is that clients close
36 * idle connections after 5 minutes, servers after
38 * http://www.connectathon.org/talks96/nfstcp.pdf
40 static int svc_conn_age_period = 6*60;
42 /* List of registered transport classes */
43 static DEFINE_SPINLOCK(svc_xprt_class_lock);
44 static LIST_HEAD(svc_xprt_class_list);
46 /* SMP locking strategy:
48 * svc_pool->sp_lock protects most of the fields of that pool.
49 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
50 * when both need to be taken (rare), svc_serv->sv_lock is first.
51 * The "service mutex" protects svc_serv->sv_nrthread.
52 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
53 * and the ->sk_info_authunix cache.
55 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
56 * enqueued multiply. During normal transport processing this bit
57 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
58 * Providers should not manipulate this bit directly.
60 * Some flags can be set to certain values at any time
61 * providing that certain rules are followed:
64 * - Can be set or cleared at any time.
65 * - After a set, svc_xprt_enqueue must be called to enqueue
66 * the transport for processing.
67 * - After a clear, the transport must be read/accepted.
68 * If this succeeds, it must be set again.
70 * - Can set at any time. It is never cleared.
72 * - Can only be set while XPT_BUSY is held which ensures
73 * that no other thread will be using the transport or will
74 * try to set XPT_DEAD.
76 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
78 struct svc_xprt_class *cl;
81 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
83 INIT_LIST_HEAD(&xcl->xcl_list);
84 spin_lock(&svc_xprt_class_lock);
85 /* Make sure there isn't already a class with the same name */
86 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
87 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
90 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
93 spin_unlock(&svc_xprt_class_lock);
96 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
98 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
100 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
101 spin_lock(&svc_xprt_class_lock);
102 list_del_init(&xcl->xcl_list);
103 spin_unlock(&svc_xprt_class_lock);
105 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
108 * Format the transport list for printing
110 int svc_print_xprts(char *buf, int maxlen)
112 struct svc_xprt_class *xcl;
117 spin_lock(&svc_xprt_class_lock);
118 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
121 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
122 slen = strlen(tmpstr);
123 if (len + slen > maxlen)
128 spin_unlock(&svc_xprt_class_lock);
133 static void svc_xprt_free(struct kref *kref)
135 struct svc_xprt *xprt =
136 container_of(kref, struct svc_xprt, xpt_ref);
137 struct module *owner = xprt->xpt_class->xcl_owner;
138 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
139 svcauth_unix_info_release(xprt);
140 put_cred(xprt->xpt_cred);
141 put_net(xprt->xpt_net);
142 /* See comment on corresponding get in xs_setup_bc_tcp(): */
143 if (xprt->xpt_bc_xprt)
144 xprt_put(xprt->xpt_bc_xprt);
145 if (xprt->xpt_bc_xps)
146 xprt_switch_put(xprt->xpt_bc_xps);
147 xprt->xpt_ops->xpo_free(xprt);
151 void svc_xprt_put(struct svc_xprt *xprt)
153 kref_put(&xprt->xpt_ref, svc_xprt_free);
155 EXPORT_SYMBOL_GPL(svc_xprt_put);
158 * Called by transport drivers to initialize the transport independent
159 * portion of the transport instance.
161 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
162 struct svc_xprt *xprt, struct svc_serv *serv)
164 memset(xprt, 0, sizeof(*xprt));
165 xprt->xpt_class = xcl;
166 xprt->xpt_ops = xcl->xcl_ops;
167 kref_init(&xprt->xpt_ref);
168 xprt->xpt_server = serv;
169 INIT_LIST_HEAD(&xprt->xpt_list);
170 INIT_LIST_HEAD(&xprt->xpt_ready);
171 INIT_LIST_HEAD(&xprt->xpt_deferred);
172 INIT_LIST_HEAD(&xprt->xpt_users);
173 mutex_init(&xprt->xpt_mutex);
174 spin_lock_init(&xprt->xpt_lock);
175 set_bit(XPT_BUSY, &xprt->xpt_flags);
176 xprt->xpt_net = get_net(net);
177 strcpy(xprt->xpt_remotebuf, "uninitialized");
179 EXPORT_SYMBOL_GPL(svc_xprt_init);
181 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
182 struct svc_serv *serv,
185 const unsigned short port,
188 struct sockaddr_in sin = {
189 .sin_family = AF_INET,
190 .sin_addr.s_addr = htonl(INADDR_ANY),
191 .sin_port = htons(port),
193 #if IS_ENABLED(CONFIG_IPV6)
194 struct sockaddr_in6 sin6 = {
195 .sin6_family = AF_INET6,
196 .sin6_addr = IN6ADDR_ANY_INIT,
197 .sin6_port = htons(port),
200 struct sockaddr *sap;
205 sap = (struct sockaddr *)&sin;
208 #if IS_ENABLED(CONFIG_IPV6)
210 sap = (struct sockaddr *)&sin6;
215 return ERR_PTR(-EAFNOSUPPORT);
218 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
222 * svc_xprt_received conditionally queues the transport for processing
223 * by another thread. The caller must hold the XPT_BUSY bit and must
224 * not thereafter touch transport data.
226 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
227 * insufficient) data.
229 static void svc_xprt_received(struct svc_xprt *xprt)
231 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
232 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
236 /* As soon as we clear busy, the xprt could be closed and
237 * 'put', so we need a reference to call svc_enqueue_xprt with:
240 smp_mb__before_atomic();
241 clear_bit(XPT_BUSY, &xprt->xpt_flags);
242 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
246 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
248 clear_bit(XPT_TEMP, &new->xpt_flags);
249 spin_lock_bh(&serv->sv_lock);
250 list_add(&new->xpt_list, &serv->sv_permsocks);
251 spin_unlock_bh(&serv->sv_lock);
252 svc_xprt_received(new);
255 static int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
256 struct net *net, const int family,
257 const unsigned short port, int flags,
258 const struct cred *cred)
260 struct svc_xprt_class *xcl;
262 spin_lock(&svc_xprt_class_lock);
263 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
264 struct svc_xprt *newxprt;
265 unsigned short newport;
267 if (strcmp(xprt_name, xcl->xcl_name))
270 if (!try_module_get(xcl->xcl_owner))
273 spin_unlock(&svc_xprt_class_lock);
274 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
275 if (IS_ERR(newxprt)) {
276 module_put(xcl->xcl_owner);
277 return PTR_ERR(newxprt);
279 newxprt->xpt_cred = get_cred(cred);
280 svc_add_new_perm_xprt(serv, newxprt);
281 newport = svc_xprt_local_port(newxprt);
285 spin_unlock(&svc_xprt_class_lock);
286 /* This errno is exposed to user space. Provide a reasonable
287 * perror msg for a bad transport. */
288 return -EPROTONOSUPPORT;
291 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
292 struct net *net, const int family,
293 const unsigned short port, int flags,
294 const struct cred *cred)
298 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
299 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags, cred);
300 if (err == -EPROTONOSUPPORT) {
301 request_module("svc%s", xprt_name);
302 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags, cred);
305 dprintk("svc: transport %s not found, err %d\n",
309 EXPORT_SYMBOL_GPL(svc_create_xprt);
312 * Copy the local and remote xprt addresses to the rqstp structure
314 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
316 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
317 rqstp->rq_addrlen = xprt->xpt_remotelen;
320 * Destination address in request is needed for binding the
321 * source address in RPC replies/callbacks later.
323 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
324 rqstp->rq_daddrlen = xprt->xpt_locallen;
326 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
329 * svc_print_addr - Format rq_addr field for printing
330 * @rqstp: svc_rqst struct containing address to print
331 * @buf: target buffer for formatted address
332 * @len: length of target buffer
335 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
337 return __svc_print_addr(svc_addr(rqstp), buf, len);
339 EXPORT_SYMBOL_GPL(svc_print_addr);
341 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
343 unsigned int limit = svc_rpc_per_connection_limit;
344 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
346 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
349 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
351 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
352 if (!svc_xprt_slots_in_range(xprt))
354 atomic_inc(&xprt->xpt_nr_rqsts);
355 set_bit(RQ_DATA, &rqstp->rq_flags);
360 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
362 struct svc_xprt *xprt = rqstp->rq_xprt;
363 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
364 atomic_dec(&xprt->xpt_nr_rqsts);
365 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
366 svc_xprt_enqueue(xprt);
370 static bool svc_xprt_ready(struct svc_xprt *xprt)
372 unsigned long xpt_flags;
375 * If another cpu has recently updated xpt_flags,
376 * sk_sock->flags, xpt_reserved, or xpt_nr_rqsts, we need to
377 * know about it; otherwise it's possible that both that cpu and
378 * this one could call svc_xprt_enqueue() without either
379 * svc_xprt_enqueue() recognizing that the conditions below
380 * are satisfied, and we could stall indefinitely:
383 xpt_flags = READ_ONCE(xprt->xpt_flags);
385 if (xpt_flags & (BIT(XPT_CONN) | BIT(XPT_CLOSE)))
387 if (xpt_flags & (BIT(XPT_DATA) | BIT(XPT_DEFERRED))) {
388 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
389 svc_xprt_slots_in_range(xprt))
391 trace_svc_xprt_no_write_space(xprt);
397 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
399 struct svc_pool *pool;
400 struct svc_rqst *rqstp = NULL;
403 if (!svc_xprt_ready(xprt))
406 /* Mark transport as busy. It will remain in this state until
407 * the provider calls svc_xprt_received. We update XPT_BUSY
408 * atomically because it also guards against trying to enqueue
409 * the transport twice.
411 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
415 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
417 atomic_long_inc(&pool->sp_stats.packets);
419 spin_lock_bh(&pool->sp_lock);
420 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
421 pool->sp_stats.sockets_queued++;
422 spin_unlock_bh(&pool->sp_lock);
424 /* find a thread for this xprt */
426 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
427 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
429 atomic_long_inc(&pool->sp_stats.threads_woken);
430 rqstp->rq_qtime = ktime_get();
431 wake_up_process(rqstp->rq_task);
434 set_bit(SP_CONGESTED, &pool->sp_flags);
439 trace_svc_xprt_do_enqueue(xprt, rqstp);
441 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
444 * Queue up a transport with data pending. If there are idle nfsd
445 * processes, wake 'em up.
448 void svc_xprt_enqueue(struct svc_xprt *xprt)
450 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
452 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
454 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
457 * Dequeue the first transport, if there is one.
459 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
461 struct svc_xprt *xprt = NULL;
463 if (list_empty(&pool->sp_sockets))
466 spin_lock_bh(&pool->sp_lock);
467 if (likely(!list_empty(&pool->sp_sockets))) {
468 xprt = list_first_entry(&pool->sp_sockets,
469 struct svc_xprt, xpt_ready);
470 list_del_init(&xprt->xpt_ready);
473 spin_unlock_bh(&pool->sp_lock);
479 * svc_reserve - change the space reserved for the reply to a request.
480 * @rqstp: The request in question
481 * @space: new max space to reserve
483 * Each request reserves some space on the output queue of the transport
484 * to make sure the reply fits. This function reduces that reserved
485 * space to be the amount of space used already, plus @space.
488 void svc_reserve(struct svc_rqst *rqstp, int space)
490 struct svc_xprt *xprt = rqstp->rq_xprt;
492 space += rqstp->rq_res.head[0].iov_len;
494 if (xprt && space < rqstp->rq_reserved) {
495 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
496 rqstp->rq_reserved = space;
497 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
498 svc_xprt_enqueue(xprt);
501 EXPORT_SYMBOL_GPL(svc_reserve);
503 static void svc_xprt_release(struct svc_rqst *rqstp)
505 struct svc_xprt *xprt = rqstp->rq_xprt;
507 xprt->xpt_ops->xpo_release_rqst(rqstp);
509 kfree(rqstp->rq_deferred);
510 rqstp->rq_deferred = NULL;
512 svc_free_res_pages(rqstp);
513 rqstp->rq_res.page_len = 0;
514 rqstp->rq_res.page_base = 0;
516 /* Reset response buffer and release
518 * But first, check that enough space was reserved
519 * for the reply, otherwise we have a bug!
521 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
522 printk(KERN_ERR "RPC request reserved %d but used %d\n",
526 rqstp->rq_res.head[0].iov_len = 0;
527 svc_reserve(rqstp, 0);
528 svc_xprt_release_slot(rqstp);
529 rqstp->rq_xprt = NULL;
534 * Some svc_serv's will have occasional work to do, even when a xprt is not
535 * waiting to be serviced. This function is there to "kick" a task in one of
536 * those services so that it can wake up and do that work. Note that we only
537 * bother with pool 0 as we don't need to wake up more than one thread for
540 void svc_wake_up(struct svc_serv *serv)
542 struct svc_rqst *rqstp;
543 struct svc_pool *pool;
545 pool = &serv->sv_pools[0];
548 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
549 /* skip any that aren't queued */
550 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
553 wake_up_process(rqstp->rq_task);
554 trace_svc_wake_up(rqstp->rq_task->pid);
559 /* No free entries available */
560 set_bit(SP_TASK_PENDING, &pool->sp_flags);
562 trace_svc_wake_up(0);
564 EXPORT_SYMBOL_GPL(svc_wake_up);
566 int svc_port_is_privileged(struct sockaddr *sin)
568 switch (sin->sa_family) {
570 return ntohs(((struct sockaddr_in *)sin)->sin_port)
573 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
581 * Make sure that we don't have too many active connections. If we have,
582 * something must be dropped. It's not clear what will happen if we allow
583 * "too many" connections, but when dealing with network-facing software,
584 * we have to code defensively. Here we do that by imposing hard limits.
586 * There's no point in trying to do random drop here for DoS
587 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
588 * attacker can easily beat that.
590 * The only somewhat efficient mechanism would be if drop old
591 * connections from the same IP first. But right now we don't even
592 * record the client IP in svc_sock.
594 * single-threaded services that expect a lot of clients will probably
595 * need to set sv_maxconn to override the default value which is based
596 * on the number of threads
598 static void svc_check_conn_limits(struct svc_serv *serv)
600 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
601 (serv->sv_nrthreads+3) * 20;
603 if (serv->sv_tmpcnt > limit) {
604 struct svc_xprt *xprt = NULL;
605 spin_lock_bh(&serv->sv_lock);
606 if (!list_empty(&serv->sv_tempsocks)) {
607 /* Try to help the admin */
608 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
609 serv->sv_name, serv->sv_maxconn ?
610 "max number of connections" :
611 "number of threads");
613 * Always select the oldest connection. It's not fair,
616 xprt = list_entry(serv->sv_tempsocks.prev,
619 set_bit(XPT_CLOSE, &xprt->xpt_flags);
622 spin_unlock_bh(&serv->sv_lock);
625 svc_xprt_enqueue(xprt);
631 static int svc_alloc_arg(struct svc_rqst *rqstp)
633 struct svc_serv *serv = rqstp->rq_server;
638 /* now allocate needed pages. If we get a failure, sleep briefly */
639 pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
640 if (pages > RPCSVC_MAXPAGES) {
641 pr_warn_once("svc: warning: pages=%u > RPCSVC_MAXPAGES=%lu\n",
642 pages, RPCSVC_MAXPAGES);
643 /* use as many pages as possible */
644 pages = RPCSVC_MAXPAGES;
646 for (i = 0; i < pages ; i++)
647 while (rqstp->rq_pages[i] == NULL) {
648 struct page *p = alloc_page(GFP_KERNEL);
650 set_current_state(TASK_INTERRUPTIBLE);
651 if (signalled() || kthread_should_stop()) {
652 set_current_state(TASK_RUNNING);
655 schedule_timeout(msecs_to_jiffies(500));
657 rqstp->rq_pages[i] = p;
659 rqstp->rq_page_end = &rqstp->rq_pages[i];
660 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
662 /* Make arg->head point to first page and arg->pages point to rest */
663 arg = &rqstp->rq_arg;
664 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
665 arg->head[0].iov_len = PAGE_SIZE;
666 arg->pages = rqstp->rq_pages + 1;
668 /* save at least one page for response */
669 arg->page_len = (pages-2)*PAGE_SIZE;
670 arg->len = (pages-1)*PAGE_SIZE;
671 arg->tail[0].iov_len = 0;
676 rqst_should_sleep(struct svc_rqst *rqstp)
678 struct svc_pool *pool = rqstp->rq_pool;
680 /* did someone call svc_wake_up? */
681 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
684 /* was a socket queued? */
685 if (!list_empty(&pool->sp_sockets))
688 /* are we shutting down? */
689 if (signalled() || kthread_should_stop())
692 /* are we freezing? */
693 if (freezing(current))
699 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
701 struct svc_pool *pool = rqstp->rq_pool;
704 /* rq_xprt should be clear on entry */
705 WARN_ON_ONCE(rqstp->rq_xprt);
707 rqstp->rq_xprt = svc_xprt_dequeue(pool);
712 * We have to be able to interrupt this wait
713 * to bring down the daemons ...
715 set_current_state(TASK_INTERRUPTIBLE);
716 smp_mb__before_atomic();
717 clear_bit(SP_CONGESTED, &pool->sp_flags);
718 clear_bit(RQ_BUSY, &rqstp->rq_flags);
719 smp_mb__after_atomic();
721 if (likely(rqst_should_sleep(rqstp)))
722 time_left = schedule_timeout(timeout);
724 __set_current_state(TASK_RUNNING);
728 set_bit(RQ_BUSY, &rqstp->rq_flags);
729 smp_mb__after_atomic();
730 rqstp->rq_xprt = svc_xprt_dequeue(pool);
735 atomic_long_inc(&pool->sp_stats.threads_timedout);
737 if (signalled() || kthread_should_stop())
738 return ERR_PTR(-EINTR);
739 return ERR_PTR(-EAGAIN);
741 /* Normally we will wait up to 5 seconds for any required
742 * cache information to be provided.
744 if (!test_bit(SP_CONGESTED, &pool->sp_flags))
745 rqstp->rq_chandle.thread_wait = 5*HZ;
747 rqstp->rq_chandle.thread_wait = 1*HZ;
748 trace_svc_xprt_dequeue(rqstp);
749 return rqstp->rq_xprt;
752 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
754 spin_lock_bh(&serv->sv_lock);
755 set_bit(XPT_TEMP, &newxpt->xpt_flags);
756 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
758 if (serv->sv_temptimer.function == NULL) {
759 /* setup timer to age temp transports */
760 serv->sv_temptimer.function = svc_age_temp_xprts;
761 mod_timer(&serv->sv_temptimer,
762 jiffies + svc_conn_age_period * HZ);
764 spin_unlock_bh(&serv->sv_lock);
765 svc_xprt_received(newxpt);
768 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
770 struct svc_serv *serv = rqstp->rq_server;
773 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
774 dprintk("svc_recv: found XPT_CLOSE\n");
775 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
776 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
777 svc_delete_xprt(xprt);
778 /* Leave XPT_BUSY set on the dead xprt: */
781 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
782 struct svc_xprt *newxpt;
784 * We know this module_get will succeed because the
785 * listener holds a reference too
787 __module_get(xprt->xpt_class->xcl_owner);
788 svc_check_conn_limits(xprt->xpt_server);
789 newxpt = xprt->xpt_ops->xpo_accept(xprt);
791 newxpt->xpt_cred = get_cred(xprt->xpt_cred);
792 svc_add_new_temp_xprt(serv, newxpt);
794 module_put(xprt->xpt_class->xcl_owner);
795 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
796 /* XPT_DATA|XPT_DEFERRED case: */
797 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
798 rqstp, rqstp->rq_pool->sp_id, xprt,
799 kref_read(&xprt->xpt_ref));
800 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
801 if (rqstp->rq_deferred)
802 len = svc_deferred_recv(rqstp);
804 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
805 rqstp->rq_stime = ktime_get();
806 rqstp->rq_reserved = serv->sv_max_mesg;
807 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
809 /* clear XPT_BUSY: */
810 svc_xprt_received(xprt);
812 trace_svc_handle_xprt(xprt, len);
817 * Receive the next request on any transport. This code is carefully
818 * organised not to touch any cachelines in the shared svc_serv
819 * structure, only cachelines in the local svc_pool.
821 int svc_recv(struct svc_rqst *rqstp, long timeout)
823 struct svc_xprt *xprt = NULL;
824 struct svc_serv *serv = rqstp->rq_server;
827 dprintk("svc: server %p waiting for data (to = %ld)\n",
832 "svc_recv: service %p, transport not NULL!\n",
835 err = svc_alloc_arg(rqstp);
842 if (signalled() || kthread_should_stop())
845 xprt = svc_get_next_xprt(rqstp, timeout);
851 len = svc_handle_xprt(rqstp, xprt);
853 /* No data, incomplete (TCP) read, or accept() */
858 clear_bit(XPT_OLD, &xprt->xpt_flags);
860 xprt->xpt_ops->xpo_secure_port(rqstp);
861 rqstp->rq_chandle.defer = svc_defer;
862 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
865 serv->sv_stats->netcnt++;
866 trace_svc_recv(rqstp, len);
869 rqstp->rq_res.len = 0;
870 svc_xprt_release(rqstp);
874 EXPORT_SYMBOL_GPL(svc_recv);
879 void svc_drop(struct svc_rqst *rqstp)
881 trace_svc_drop(rqstp);
882 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
883 svc_xprt_release(rqstp);
885 EXPORT_SYMBOL_GPL(svc_drop);
888 * Return reply to client.
890 int svc_send(struct svc_rqst *rqstp)
892 struct svc_xprt *xprt;
896 xprt = rqstp->rq_xprt;
900 /* release the receive skb before sending the reply */
901 xprt->xpt_ops->xpo_release_rqst(rqstp);
903 /* calculate over-all length */
905 xb->len = xb->head[0].iov_len +
909 /* Grab mutex to serialize outgoing data. */
910 mutex_lock(&xprt->xpt_mutex);
911 trace_svc_stats_latency(rqstp);
912 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
913 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
916 len = xprt->xpt_ops->xpo_sendto(rqstp);
917 mutex_unlock(&xprt->xpt_mutex);
918 trace_svc_send(rqstp, len);
919 svc_xprt_release(rqstp);
921 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
928 * Timer function to close old temporary transports, using
929 * a mark-and-sweep algorithm.
931 static void svc_age_temp_xprts(struct timer_list *t)
933 struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
934 struct svc_xprt *xprt;
935 struct list_head *le, *next;
937 dprintk("svc_age_temp_xprts\n");
939 if (!spin_trylock_bh(&serv->sv_lock)) {
940 /* busy, try again 1 sec later */
941 dprintk("svc_age_temp_xprts: busy\n");
942 mod_timer(&serv->sv_temptimer, jiffies + HZ);
946 list_for_each_safe(le, next, &serv->sv_tempsocks) {
947 xprt = list_entry(le, struct svc_xprt, xpt_list);
949 /* First time through, just mark it OLD. Second time
950 * through, close it. */
951 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
953 if (kref_read(&xprt->xpt_ref) > 1 ||
954 test_bit(XPT_BUSY, &xprt->xpt_flags))
957 set_bit(XPT_CLOSE, &xprt->xpt_flags);
958 dprintk("queuing xprt %p for closing\n", xprt);
960 /* a thread will dequeue and close it soon */
961 svc_xprt_enqueue(xprt);
963 spin_unlock_bh(&serv->sv_lock);
965 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
968 /* Close temporary transports whose xpt_local matches server_addr immediately
969 * instead of waiting for them to be picked up by the timer.
971 * This is meant to be called from a notifier_block that runs when an ip
972 * address is deleted.
974 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
976 struct svc_xprt *xprt;
977 struct list_head *le, *next;
978 LIST_HEAD(to_be_closed);
980 spin_lock_bh(&serv->sv_lock);
981 list_for_each_safe(le, next, &serv->sv_tempsocks) {
982 xprt = list_entry(le, struct svc_xprt, xpt_list);
983 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
985 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
986 list_move(le, &to_be_closed);
989 spin_unlock_bh(&serv->sv_lock);
991 while (!list_empty(&to_be_closed)) {
992 le = to_be_closed.next;
994 xprt = list_entry(le, struct svc_xprt, xpt_list);
995 set_bit(XPT_CLOSE, &xprt->xpt_flags);
996 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
997 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
999 svc_xprt_enqueue(xprt);
1002 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1004 static void call_xpt_users(struct svc_xprt *xprt)
1006 struct svc_xpt_user *u;
1008 spin_lock(&xprt->xpt_lock);
1009 while (!list_empty(&xprt->xpt_users)) {
1010 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1011 list_del_init(&u->list);
1014 spin_unlock(&xprt->xpt_lock);
1018 * Remove a dead transport
1020 static void svc_delete_xprt(struct svc_xprt *xprt)
1022 struct svc_serv *serv = xprt->xpt_server;
1023 struct svc_deferred_req *dr;
1025 /* Only do this once */
1026 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1029 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1030 xprt->xpt_ops->xpo_detach(xprt);
1032 spin_lock_bh(&serv->sv_lock);
1033 list_del_init(&xprt->xpt_list);
1034 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1035 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1037 spin_unlock_bh(&serv->sv_lock);
1039 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1042 call_xpt_users(xprt);
1046 void svc_close_xprt(struct svc_xprt *xprt)
1048 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1049 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1050 /* someone else will have to effect the close */
1053 * We expect svc_close_xprt() to work even when no threads are
1054 * running (e.g., while configuring the server before starting
1055 * any threads), so if the transport isn't busy, we delete
1058 svc_delete_xprt(xprt);
1060 EXPORT_SYMBOL_GPL(svc_close_xprt);
1062 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1064 struct svc_xprt *xprt;
1067 spin_lock(&serv->sv_lock);
1068 list_for_each_entry(xprt, xprt_list, xpt_list) {
1069 if (xprt->xpt_net != net)
1072 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1073 svc_xprt_enqueue(xprt);
1075 spin_unlock(&serv->sv_lock);
1079 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1081 struct svc_pool *pool;
1082 struct svc_xprt *xprt;
1083 struct svc_xprt *tmp;
1086 for (i = 0; i < serv->sv_nrpools; i++) {
1087 pool = &serv->sv_pools[i];
1089 spin_lock_bh(&pool->sp_lock);
1090 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1091 if (xprt->xpt_net != net)
1093 list_del_init(&xprt->xpt_ready);
1094 spin_unlock_bh(&pool->sp_lock);
1097 spin_unlock_bh(&pool->sp_lock);
1102 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1104 struct svc_xprt *xprt;
1106 while ((xprt = svc_dequeue_net(serv, net))) {
1107 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1108 svc_delete_xprt(xprt);
1113 * Server threads may still be running (especially in the case where the
1114 * service is still running in other network namespaces).
1116 * So we shut down sockets the same way we would on a running server, by
1117 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1118 * the close. In the case there are no such other threads,
1119 * threads running, svc_clean_up_xprts() does a simple version of a
1120 * server's main event loop, and in the case where there are other
1121 * threads, we may need to wait a little while and then check again to
1122 * see if they're done.
1124 void svc_close_net(struct svc_serv *serv, struct net *net)
1128 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1129 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1131 svc_clean_up_xprts(serv, net);
1137 * Handle defer and revisit of requests
1140 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1142 struct svc_deferred_req *dr =
1143 container_of(dreq, struct svc_deferred_req, handle);
1144 struct svc_xprt *xprt = dr->xprt;
1146 spin_lock(&xprt->xpt_lock);
1147 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1148 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1149 spin_unlock(&xprt->xpt_lock);
1150 dprintk("revisit canceled\n");
1152 trace_svc_drop_deferred(dr);
1156 dprintk("revisit queued\n");
1158 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1159 spin_unlock(&xprt->xpt_lock);
1160 svc_xprt_enqueue(xprt);
1165 * Save the request off for later processing. The request buffer looks
1168 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1170 * This code can only handle requests that consist of an xprt-header
1173 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1175 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1176 struct svc_deferred_req *dr;
1178 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1179 return NULL; /* if more than a page, give up FIXME */
1180 if (rqstp->rq_deferred) {
1181 dr = rqstp->rq_deferred;
1182 rqstp->rq_deferred = NULL;
1186 /* FIXME maybe discard if size too large */
1187 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1188 dr = kmalloc(size, GFP_KERNEL);
1192 dr->handle.owner = rqstp->rq_server;
1193 dr->prot = rqstp->rq_prot;
1194 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1195 dr->addrlen = rqstp->rq_addrlen;
1196 dr->daddr = rqstp->rq_daddr;
1197 dr->argslen = rqstp->rq_arg.len >> 2;
1198 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1200 /* back up head to the start of the buffer and copy */
1201 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1202 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1205 svc_xprt_get(rqstp->rq_xprt);
1206 dr->xprt = rqstp->rq_xprt;
1207 set_bit(RQ_DROPME, &rqstp->rq_flags);
1209 dr->handle.revisit = svc_revisit;
1210 trace_svc_defer(rqstp);
1215 * recv data from a deferred request into an active one
1217 static int svc_deferred_recv(struct svc_rqst *rqstp)
1219 struct svc_deferred_req *dr = rqstp->rq_deferred;
1221 /* setup iov_base past transport header */
1222 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1223 /* The iov_len does not include the transport header bytes */
1224 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1225 rqstp->rq_arg.page_len = 0;
1226 /* The rq_arg.len includes the transport header bytes */
1227 rqstp->rq_arg.len = dr->argslen<<2;
1228 rqstp->rq_prot = dr->prot;
1229 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1230 rqstp->rq_addrlen = dr->addrlen;
1231 /* Save off transport header len in case we get deferred again */
1232 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1233 rqstp->rq_daddr = dr->daddr;
1234 rqstp->rq_respages = rqstp->rq_pages;
1235 return (dr->argslen<<2) - dr->xprt_hlen;
1239 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1241 struct svc_deferred_req *dr = NULL;
1243 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1245 spin_lock(&xprt->xpt_lock);
1246 if (!list_empty(&xprt->xpt_deferred)) {
1247 dr = list_entry(xprt->xpt_deferred.next,
1248 struct svc_deferred_req,
1250 list_del_init(&dr->handle.recent);
1251 trace_svc_revisit_deferred(dr);
1253 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1254 spin_unlock(&xprt->xpt_lock);
1259 * svc_find_xprt - find an RPC transport instance
1260 * @serv: pointer to svc_serv to search
1261 * @xcl_name: C string containing transport's class name
1262 * @net: owner net pointer
1263 * @af: Address family of transport's local address
1264 * @port: transport's IP port number
1266 * Return the transport instance pointer for the endpoint accepting
1267 * connections/peer traffic from the specified transport class,
1268 * address family and port.
1270 * Specifying 0 for the address family or port is effectively a
1271 * wild-card, and will result in matching the first transport in the
1272 * service's list that has a matching class name.
1274 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1275 struct net *net, const sa_family_t af,
1276 const unsigned short port)
1278 struct svc_xprt *xprt;
1279 struct svc_xprt *found = NULL;
1281 /* Sanity check the args */
1282 if (serv == NULL || xcl_name == NULL)
1285 spin_lock_bh(&serv->sv_lock);
1286 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1287 if (xprt->xpt_net != net)
1289 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1291 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1293 if (port != 0 && port != svc_xprt_local_port(xprt))
1299 spin_unlock_bh(&serv->sv_lock);
1302 EXPORT_SYMBOL_GPL(svc_find_xprt);
1304 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1305 char *pos, int remaining)
1309 len = snprintf(pos, remaining, "%s %u\n",
1310 xprt->xpt_class->xcl_name,
1311 svc_xprt_local_port(xprt));
1312 if (len >= remaining)
1313 return -ENAMETOOLONG;
1318 * svc_xprt_names - format a buffer with a list of transport names
1319 * @serv: pointer to an RPC service
1320 * @buf: pointer to a buffer to be filled in
1321 * @buflen: length of buffer to be filled in
1323 * Fills in @buf with a string containing a list of transport names,
1324 * each name terminated with '\n'.
1326 * Returns positive length of the filled-in string on success; otherwise
1327 * a negative errno value is returned if an error occurs.
1329 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1331 struct svc_xprt *xprt;
1335 /* Sanity check args */
1339 spin_lock_bh(&serv->sv_lock);
1343 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1344 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1356 spin_unlock_bh(&serv->sv_lock);
1359 EXPORT_SYMBOL_GPL(svc_xprt_names);
1362 /*----------------------------------------------------------------------------*/
1364 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1366 unsigned int pidx = (unsigned int)*pos;
1367 struct svc_serv *serv = m->private;
1369 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1372 return SEQ_START_TOKEN;
1373 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1376 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1378 struct svc_pool *pool = p;
1379 struct svc_serv *serv = m->private;
1381 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1383 if (p == SEQ_START_TOKEN) {
1384 pool = &serv->sv_pools[0];
1386 unsigned int pidx = (pool - &serv->sv_pools[0]);
1387 if (pidx < serv->sv_nrpools-1)
1388 pool = &serv->sv_pools[pidx+1];
1396 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1400 static int svc_pool_stats_show(struct seq_file *m, void *p)
1402 struct svc_pool *pool = p;
1404 if (p == SEQ_START_TOKEN) {
1405 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1409 seq_printf(m, "%u %lu %lu %lu %lu\n",
1411 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1412 pool->sp_stats.sockets_queued,
1413 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1414 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1419 static const struct seq_operations svc_pool_stats_seq_ops = {
1420 .start = svc_pool_stats_start,
1421 .next = svc_pool_stats_next,
1422 .stop = svc_pool_stats_stop,
1423 .show = svc_pool_stats_show,
1426 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1430 err = seq_open(file, &svc_pool_stats_seq_ops);
1432 ((struct seq_file *) file->private_data)->private = serv;
1435 EXPORT_SYMBOL(svc_pool_stats_open);
1437 /*----------------------------------------------------------------------------*/