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_cred(xprt->xpt_cred);
140 put_net(xprt->xpt_net);
141 /* See comment on corresponding get in xs_setup_bc_tcp(): */
142 if (xprt->xpt_bc_xprt)
143 xprt_put(xprt->xpt_bc_xprt);
144 if (xprt->xpt_bc_xps)
145 xprt_switch_put(xprt->xpt_bc_xps);
146 xprt->xpt_ops->xpo_free(xprt);
150 void svc_xprt_put(struct svc_xprt *xprt)
152 kref_put(&xprt->xpt_ref, svc_xprt_free);
154 EXPORT_SYMBOL_GPL(svc_xprt_put);
157 * Called by transport drivers to initialize the transport independent
158 * portion of the transport instance.
160 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
161 struct svc_xprt *xprt, struct svc_serv *serv)
163 memset(xprt, 0, sizeof(*xprt));
164 xprt->xpt_class = xcl;
165 xprt->xpt_ops = xcl->xcl_ops;
166 kref_init(&xprt->xpt_ref);
167 xprt->xpt_server = serv;
168 INIT_LIST_HEAD(&xprt->xpt_list);
169 INIT_LIST_HEAD(&xprt->xpt_ready);
170 INIT_LIST_HEAD(&xprt->xpt_deferred);
171 INIT_LIST_HEAD(&xprt->xpt_users);
172 mutex_init(&xprt->xpt_mutex);
173 spin_lock_init(&xprt->xpt_lock);
174 set_bit(XPT_BUSY, &xprt->xpt_flags);
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,
257 const struct cred *cred)
259 struct svc_xprt_class *xcl;
261 spin_lock(&svc_xprt_class_lock);
262 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
263 struct svc_xprt *newxprt;
264 unsigned short newport;
266 if (strcmp(xprt_name, xcl->xcl_name))
269 if (!try_module_get(xcl->xcl_owner))
272 spin_unlock(&svc_xprt_class_lock);
273 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
274 if (IS_ERR(newxprt)) {
275 module_put(xcl->xcl_owner);
276 return PTR_ERR(newxprt);
278 newxprt->xpt_cred = get_cred(cred);
279 svc_add_new_perm_xprt(serv, newxprt);
280 newport = svc_xprt_local_port(newxprt);
284 spin_unlock(&svc_xprt_class_lock);
285 /* This errno is exposed to user space. Provide a reasonable
286 * perror msg for a bad transport. */
287 return -EPROTONOSUPPORT;
290 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
291 struct net *net, const int family,
292 const unsigned short port, int flags,
293 const struct cred *cred)
297 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
298 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags, cred);
299 if (err == -EPROTONOSUPPORT) {
300 request_module("svc%s", xprt_name);
301 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags, cred);
304 dprintk("svc: transport %s not found, err %d\n",
308 EXPORT_SYMBOL_GPL(svc_create_xprt);
311 * Copy the local and remote xprt addresses to the rqstp structure
313 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
315 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
316 rqstp->rq_addrlen = xprt->xpt_remotelen;
319 * Destination address in request is needed for binding the
320 * source address in RPC replies/callbacks later.
322 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
323 rqstp->rq_daddrlen = xprt->xpt_locallen;
325 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
328 * svc_print_addr - Format rq_addr field for printing
329 * @rqstp: svc_rqst struct containing address to print
330 * @buf: target buffer for formatted address
331 * @len: length of target buffer
334 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
336 return __svc_print_addr(svc_addr(rqstp), buf, len);
338 EXPORT_SYMBOL_GPL(svc_print_addr);
340 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
342 unsigned int limit = svc_rpc_per_connection_limit;
343 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
345 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
348 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
350 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
351 if (!svc_xprt_slots_in_range(xprt))
353 atomic_inc(&xprt->xpt_nr_rqsts);
354 set_bit(RQ_DATA, &rqstp->rq_flags);
359 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
361 struct svc_xprt *xprt = rqstp->rq_xprt;
362 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
363 atomic_dec(&xprt->xpt_nr_rqsts);
364 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
365 svc_xprt_enqueue(xprt);
369 static bool svc_xprt_ready(struct svc_xprt *xprt)
371 unsigned long xpt_flags;
374 * If another cpu has recently updated xpt_flags,
375 * sk_sock->flags, xpt_reserved, or xpt_nr_rqsts, we need to
376 * know about it; otherwise it's possible that both that cpu and
377 * this one could call svc_xprt_enqueue() without either
378 * svc_xprt_enqueue() recognizing that the conditions below
379 * are satisfied, and we could stall indefinitely:
382 xpt_flags = READ_ONCE(xprt->xpt_flags);
384 if (xpt_flags & (BIT(XPT_CONN) | BIT(XPT_CLOSE)))
386 if (xpt_flags & (BIT(XPT_DATA) | BIT(XPT_DEFERRED))) {
387 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
388 svc_xprt_slots_in_range(xprt))
390 trace_svc_xprt_no_write_space(xprt);
396 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
398 struct svc_pool *pool;
399 struct svc_rqst *rqstp = NULL;
402 if (!svc_xprt_ready(xprt))
405 /* Mark transport as busy. It will remain in this state until
406 * the provider calls svc_xprt_received. We update XPT_BUSY
407 * atomically because it also guards against trying to enqueue
408 * the transport twice.
410 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
414 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
416 atomic_long_inc(&pool->sp_stats.packets);
418 spin_lock_bh(&pool->sp_lock);
419 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
420 pool->sp_stats.sockets_queued++;
421 spin_unlock_bh(&pool->sp_lock);
423 /* find a thread for this xprt */
425 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
426 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
428 atomic_long_inc(&pool->sp_stats.threads_woken);
429 rqstp->rq_qtime = ktime_get();
430 wake_up_process(rqstp->rq_task);
433 set_bit(SP_CONGESTED, &pool->sp_flags);
438 trace_svc_xprt_do_enqueue(xprt, rqstp);
440 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
443 * Queue up a transport with data pending. If there are idle nfsd
444 * processes, wake 'em up.
447 void svc_xprt_enqueue(struct svc_xprt *xprt)
449 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
451 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
453 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
456 * Dequeue the first transport, if there is one.
458 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
460 struct svc_xprt *xprt = NULL;
462 if (list_empty(&pool->sp_sockets))
465 spin_lock_bh(&pool->sp_lock);
466 if (likely(!list_empty(&pool->sp_sockets))) {
467 xprt = list_first_entry(&pool->sp_sockets,
468 struct svc_xprt, xpt_ready);
469 list_del_init(&xprt->xpt_ready);
472 spin_unlock_bh(&pool->sp_lock);
478 * svc_reserve - change the space reserved for the reply to a request.
479 * @rqstp: The request in question
480 * @space: new max space to reserve
482 * Each request reserves some space on the output queue of the transport
483 * to make sure the reply fits. This function reduces that reserved
484 * space to be the amount of space used already, plus @space.
487 void svc_reserve(struct svc_rqst *rqstp, int space)
489 struct svc_xprt *xprt = rqstp->rq_xprt;
491 space += rqstp->rq_res.head[0].iov_len;
493 if (xprt && space < rqstp->rq_reserved) {
494 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
495 rqstp->rq_reserved = space;
496 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
497 svc_xprt_enqueue(xprt);
500 EXPORT_SYMBOL_GPL(svc_reserve);
502 static void svc_xprt_release(struct svc_rqst *rqstp)
504 struct svc_xprt *xprt = rqstp->rq_xprt;
506 xprt->xpt_ops->xpo_release_rqst(rqstp);
508 kfree(rqstp->rq_deferred);
509 rqstp->rq_deferred = NULL;
511 svc_free_res_pages(rqstp);
512 rqstp->rq_res.page_len = 0;
513 rqstp->rq_res.page_base = 0;
515 /* Reset response buffer and release
517 * But first, check that enough space was reserved
518 * for the reply, otherwise we have a bug!
520 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
521 printk(KERN_ERR "RPC request reserved %d but used %d\n",
525 rqstp->rq_res.head[0].iov_len = 0;
526 svc_reserve(rqstp, 0);
527 svc_xprt_release_slot(rqstp);
528 rqstp->rq_xprt = NULL;
533 * Some svc_serv's will have occasional work to do, even when a xprt is not
534 * waiting to be serviced. This function is there to "kick" a task in one of
535 * those services so that it can wake up and do that work. Note that we only
536 * bother with pool 0 as we don't need to wake up more than one thread for
539 void svc_wake_up(struct svc_serv *serv)
541 struct svc_rqst *rqstp;
542 struct svc_pool *pool;
544 pool = &serv->sv_pools[0];
547 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
548 /* skip any that aren't queued */
549 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
552 wake_up_process(rqstp->rq_task);
553 trace_svc_wake_up(rqstp->rq_task->pid);
558 /* No free entries available */
559 set_bit(SP_TASK_PENDING, &pool->sp_flags);
561 trace_svc_wake_up(0);
563 EXPORT_SYMBOL_GPL(svc_wake_up);
565 int svc_port_is_privileged(struct sockaddr *sin)
567 switch (sin->sa_family) {
569 return ntohs(((struct sockaddr_in *)sin)->sin_port)
572 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
580 * Make sure that we don't have too many active connections. If we have,
581 * something must be dropped. It's not clear what will happen if we allow
582 * "too many" connections, but when dealing with network-facing software,
583 * we have to code defensively. Here we do that by imposing hard limits.
585 * There's no point in trying to do random drop here for DoS
586 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
587 * attacker can easily beat that.
589 * The only somewhat efficient mechanism would be if drop old
590 * connections from the same IP first. But right now we don't even
591 * record the client IP in svc_sock.
593 * single-threaded services that expect a lot of clients will probably
594 * need to set sv_maxconn to override the default value which is based
595 * on the number of threads
597 static void svc_check_conn_limits(struct svc_serv *serv)
599 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
600 (serv->sv_nrthreads+3) * 20;
602 if (serv->sv_tmpcnt > limit) {
603 struct svc_xprt *xprt = NULL;
604 spin_lock_bh(&serv->sv_lock);
605 if (!list_empty(&serv->sv_tempsocks)) {
606 /* Try to help the admin */
607 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
608 serv->sv_name, serv->sv_maxconn ?
609 "max number of connections" :
610 "number of threads");
612 * Always select the oldest connection. It's not fair,
615 xprt = list_entry(serv->sv_tempsocks.prev,
618 set_bit(XPT_CLOSE, &xprt->xpt_flags);
621 spin_unlock_bh(&serv->sv_lock);
624 svc_xprt_enqueue(xprt);
630 static int svc_alloc_arg(struct svc_rqst *rqstp)
632 struct svc_serv *serv = rqstp->rq_server;
637 /* now allocate needed pages. If we get a failure, sleep briefly */
638 pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
639 if (pages > RPCSVC_MAXPAGES) {
640 pr_warn_once("svc: warning: pages=%u > RPCSVC_MAXPAGES=%lu\n",
641 pages, RPCSVC_MAXPAGES);
642 /* use as many pages as possible */
643 pages = RPCSVC_MAXPAGES;
645 for (i = 0; i < pages ; i++)
646 while (rqstp->rq_pages[i] == NULL) {
647 struct page *p = alloc_page(GFP_KERNEL);
649 set_current_state(TASK_INTERRUPTIBLE);
650 if (signalled() || kthread_should_stop()) {
651 set_current_state(TASK_RUNNING);
654 schedule_timeout(msecs_to_jiffies(500));
656 rqstp->rq_pages[i] = p;
658 rqstp->rq_page_end = &rqstp->rq_pages[i];
659 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
661 /* Make arg->head point to first page and arg->pages point to rest */
662 arg = &rqstp->rq_arg;
663 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
664 arg->head[0].iov_len = PAGE_SIZE;
665 arg->pages = rqstp->rq_pages + 1;
667 /* save at least one page for response */
668 arg->page_len = (pages-2)*PAGE_SIZE;
669 arg->len = (pages-1)*PAGE_SIZE;
670 arg->tail[0].iov_len = 0;
675 rqst_should_sleep(struct svc_rqst *rqstp)
677 struct svc_pool *pool = rqstp->rq_pool;
679 /* did someone call svc_wake_up? */
680 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
683 /* was a socket queued? */
684 if (!list_empty(&pool->sp_sockets))
687 /* are we shutting down? */
688 if (signalled() || kthread_should_stop())
691 /* are we freezing? */
692 if (freezing(current))
698 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
700 struct svc_pool *pool = rqstp->rq_pool;
703 /* rq_xprt should be clear on entry */
704 WARN_ON_ONCE(rqstp->rq_xprt);
706 rqstp->rq_xprt = svc_xprt_dequeue(pool);
711 * We have to be able to interrupt this wait
712 * to bring down the daemons ...
714 set_current_state(TASK_INTERRUPTIBLE);
715 smp_mb__before_atomic();
716 clear_bit(SP_CONGESTED, &pool->sp_flags);
717 clear_bit(RQ_BUSY, &rqstp->rq_flags);
718 smp_mb__after_atomic();
720 if (likely(rqst_should_sleep(rqstp)))
721 time_left = schedule_timeout(timeout);
723 __set_current_state(TASK_RUNNING);
727 set_bit(RQ_BUSY, &rqstp->rq_flags);
728 smp_mb__after_atomic();
729 rqstp->rq_xprt = svc_xprt_dequeue(pool);
734 atomic_long_inc(&pool->sp_stats.threads_timedout);
736 if (signalled() || kthread_should_stop())
737 return ERR_PTR(-EINTR);
738 return ERR_PTR(-EAGAIN);
740 /* Normally we will wait up to 5 seconds for any required
741 * cache information to be provided.
743 if (!test_bit(SP_CONGESTED, &pool->sp_flags))
744 rqstp->rq_chandle.thread_wait = 5*HZ;
746 rqstp->rq_chandle.thread_wait = 1*HZ;
747 trace_svc_xprt_dequeue(rqstp);
748 return rqstp->rq_xprt;
751 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
753 spin_lock_bh(&serv->sv_lock);
754 set_bit(XPT_TEMP, &newxpt->xpt_flags);
755 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
757 if (serv->sv_temptimer.function == NULL) {
758 /* setup timer to age temp transports */
759 serv->sv_temptimer.function = svc_age_temp_xprts;
760 mod_timer(&serv->sv_temptimer,
761 jiffies + svc_conn_age_period * HZ);
763 spin_unlock_bh(&serv->sv_lock);
764 svc_xprt_received(newxpt);
767 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
769 struct svc_serv *serv = rqstp->rq_server;
772 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
773 dprintk("svc_recv: found XPT_CLOSE\n");
774 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
775 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
776 svc_delete_xprt(xprt);
777 /* Leave XPT_BUSY set on the dead xprt: */
780 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
781 struct svc_xprt *newxpt;
783 * We know this module_get will succeed because the
784 * listener holds a reference too
786 __module_get(xprt->xpt_class->xcl_owner);
787 svc_check_conn_limits(xprt->xpt_server);
788 newxpt = xprt->xpt_ops->xpo_accept(xprt);
790 newxpt->xpt_cred = get_cred(xprt->xpt_cred);
791 svc_add_new_temp_xprt(serv, newxpt);
793 module_put(xprt->xpt_class->xcl_owner);
794 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
795 /* XPT_DATA|XPT_DEFERRED case: */
796 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
797 rqstp, rqstp->rq_pool->sp_id, xprt,
798 kref_read(&xprt->xpt_ref));
799 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
800 if (rqstp->rq_deferred)
801 len = svc_deferred_recv(rqstp);
803 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
804 rqstp->rq_stime = ktime_get();
805 rqstp->rq_reserved = serv->sv_max_mesg;
806 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
808 /* clear XPT_BUSY: */
809 svc_xprt_received(xprt);
811 trace_svc_handle_xprt(xprt, len);
816 * Receive the next request on any transport. This code is carefully
817 * organised not to touch any cachelines in the shared svc_serv
818 * structure, only cachelines in the local svc_pool.
820 int svc_recv(struct svc_rqst *rqstp, long timeout)
822 struct svc_xprt *xprt = NULL;
823 struct svc_serv *serv = rqstp->rq_server;
826 dprintk("svc: server %p waiting for data (to = %ld)\n",
831 "svc_recv: service %p, transport not NULL!\n",
834 err = svc_alloc_arg(rqstp);
841 if (signalled() || kthread_should_stop())
844 xprt = svc_get_next_xprt(rqstp, timeout);
850 len = svc_handle_xprt(rqstp, xprt);
852 /* No data, incomplete (TCP) read, or accept() */
857 clear_bit(XPT_OLD, &xprt->xpt_flags);
859 xprt->xpt_ops->xpo_secure_port(rqstp);
860 rqstp->rq_chandle.defer = svc_defer;
861 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
864 serv->sv_stats->netcnt++;
865 trace_svc_recv(rqstp, len);
868 rqstp->rq_res.len = 0;
869 svc_xprt_release(rqstp);
873 EXPORT_SYMBOL_GPL(svc_recv);
878 void svc_drop(struct svc_rqst *rqstp)
880 trace_svc_drop(rqstp);
881 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
882 svc_xprt_release(rqstp);
884 EXPORT_SYMBOL_GPL(svc_drop);
887 * Return reply to client.
889 int svc_send(struct svc_rqst *rqstp)
891 struct svc_xprt *xprt;
895 xprt = rqstp->rq_xprt;
899 /* release the receive skb before sending the reply */
900 xprt->xpt_ops->xpo_release_rqst(rqstp);
902 /* calculate over-all length */
904 xb->len = xb->head[0].iov_len +
908 /* Grab mutex to serialize outgoing data. */
909 mutex_lock(&xprt->xpt_mutex);
910 trace_svc_stats_latency(rqstp);
911 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
912 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
915 len = xprt->xpt_ops->xpo_sendto(rqstp);
916 mutex_unlock(&xprt->xpt_mutex);
917 trace_svc_send(rqstp, len);
918 svc_xprt_release(rqstp);
920 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
927 * Timer function to close old temporary transports, using
928 * a mark-and-sweep algorithm.
930 static void svc_age_temp_xprts(struct timer_list *t)
932 struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
933 struct svc_xprt *xprt;
934 struct list_head *le, *next;
936 dprintk("svc_age_temp_xprts\n");
938 if (!spin_trylock_bh(&serv->sv_lock)) {
939 /* busy, try again 1 sec later */
940 dprintk("svc_age_temp_xprts: busy\n");
941 mod_timer(&serv->sv_temptimer, jiffies + HZ);
945 list_for_each_safe(le, next, &serv->sv_tempsocks) {
946 xprt = list_entry(le, struct svc_xprt, xpt_list);
948 /* First time through, just mark it OLD. Second time
949 * through, close it. */
950 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
952 if (kref_read(&xprt->xpt_ref) > 1 ||
953 test_bit(XPT_BUSY, &xprt->xpt_flags))
956 set_bit(XPT_CLOSE, &xprt->xpt_flags);
957 dprintk("queuing xprt %p for closing\n", xprt);
959 /* a thread will dequeue and close it soon */
960 svc_xprt_enqueue(xprt);
962 spin_unlock_bh(&serv->sv_lock);
964 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
967 /* Close temporary transports whose xpt_local matches server_addr immediately
968 * instead of waiting for them to be picked up by the timer.
970 * This is meant to be called from a notifier_block that runs when an ip
971 * address is deleted.
973 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
975 struct svc_xprt *xprt;
976 struct list_head *le, *next;
977 LIST_HEAD(to_be_closed);
979 spin_lock_bh(&serv->sv_lock);
980 list_for_each_safe(le, next, &serv->sv_tempsocks) {
981 xprt = list_entry(le, struct svc_xprt, xpt_list);
982 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
984 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
985 list_move(le, &to_be_closed);
988 spin_unlock_bh(&serv->sv_lock);
990 while (!list_empty(&to_be_closed)) {
991 le = to_be_closed.next;
993 xprt = list_entry(le, struct svc_xprt, xpt_list);
994 set_bit(XPT_CLOSE, &xprt->xpt_flags);
995 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
996 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
998 svc_xprt_enqueue(xprt);
1001 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1003 static void call_xpt_users(struct svc_xprt *xprt)
1005 struct svc_xpt_user *u;
1007 spin_lock(&xprt->xpt_lock);
1008 while (!list_empty(&xprt->xpt_users)) {
1009 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1010 list_del_init(&u->list);
1013 spin_unlock(&xprt->xpt_lock);
1017 * Remove a dead transport
1019 static void svc_delete_xprt(struct svc_xprt *xprt)
1021 struct svc_serv *serv = xprt->xpt_server;
1022 struct svc_deferred_req *dr;
1024 /* Only do this once */
1025 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1028 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
1029 xprt->xpt_ops->xpo_detach(xprt);
1031 spin_lock_bh(&serv->sv_lock);
1032 list_del_init(&xprt->xpt_list);
1033 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1034 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1036 spin_unlock_bh(&serv->sv_lock);
1038 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1041 call_xpt_users(xprt);
1045 void svc_close_xprt(struct svc_xprt *xprt)
1047 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1048 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1049 /* someone else will have to effect the close */
1052 * We expect svc_close_xprt() to work even when no threads are
1053 * running (e.g., while configuring the server before starting
1054 * any threads), so if the transport isn't busy, we delete
1057 svc_delete_xprt(xprt);
1059 EXPORT_SYMBOL_GPL(svc_close_xprt);
1061 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1063 struct svc_xprt *xprt;
1066 spin_lock(&serv->sv_lock);
1067 list_for_each_entry(xprt, xprt_list, xpt_list) {
1068 if (xprt->xpt_net != net)
1071 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1072 svc_xprt_enqueue(xprt);
1074 spin_unlock(&serv->sv_lock);
1078 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1080 struct svc_pool *pool;
1081 struct svc_xprt *xprt;
1082 struct svc_xprt *tmp;
1085 for (i = 0; i < serv->sv_nrpools; i++) {
1086 pool = &serv->sv_pools[i];
1088 spin_lock_bh(&pool->sp_lock);
1089 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1090 if (xprt->xpt_net != net)
1092 list_del_init(&xprt->xpt_ready);
1093 spin_unlock_bh(&pool->sp_lock);
1096 spin_unlock_bh(&pool->sp_lock);
1101 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1103 struct svc_xprt *xprt;
1105 while ((xprt = svc_dequeue_net(serv, net))) {
1106 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1107 svc_delete_xprt(xprt);
1112 * Server threads may still be running (especially in the case where the
1113 * service is still running in other network namespaces).
1115 * So we shut down sockets the same way we would on a running server, by
1116 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1117 * the close. In the case there are no such other threads,
1118 * threads running, svc_clean_up_xprts() does a simple version of a
1119 * server's main event loop, and in the case where there are other
1120 * threads, we may need to wait a little while and then check again to
1121 * see if they're done.
1123 void svc_close_net(struct svc_serv *serv, struct net *net)
1127 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1128 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1130 svc_clean_up_xprts(serv, net);
1136 * Handle defer and revisit of requests
1139 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1141 struct svc_deferred_req *dr =
1142 container_of(dreq, struct svc_deferred_req, handle);
1143 struct svc_xprt *xprt = dr->xprt;
1145 spin_lock(&xprt->xpt_lock);
1146 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1147 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1148 spin_unlock(&xprt->xpt_lock);
1149 dprintk("revisit canceled\n");
1151 trace_svc_drop_deferred(dr);
1155 dprintk("revisit queued\n");
1157 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1158 spin_unlock(&xprt->xpt_lock);
1159 svc_xprt_enqueue(xprt);
1164 * Save the request off for later processing. The request buffer looks
1167 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1169 * This code can only handle requests that consist of an xprt-header
1172 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1174 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1175 struct svc_deferred_req *dr;
1177 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1178 return NULL; /* if more than a page, give up FIXME */
1179 if (rqstp->rq_deferred) {
1180 dr = rqstp->rq_deferred;
1181 rqstp->rq_deferred = NULL;
1185 /* FIXME maybe discard if size too large */
1186 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1187 dr = kmalloc(size, GFP_KERNEL);
1191 dr->handle.owner = rqstp->rq_server;
1192 dr->prot = rqstp->rq_prot;
1193 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1194 dr->addrlen = rqstp->rq_addrlen;
1195 dr->daddr = rqstp->rq_daddr;
1196 dr->argslen = rqstp->rq_arg.len >> 2;
1197 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1199 /* back up head to the start of the buffer and copy */
1200 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1201 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1204 svc_xprt_get(rqstp->rq_xprt);
1205 dr->xprt = rqstp->rq_xprt;
1206 set_bit(RQ_DROPME, &rqstp->rq_flags);
1208 dr->handle.revisit = svc_revisit;
1209 trace_svc_defer(rqstp);
1214 * recv data from a deferred request into an active one
1216 static int svc_deferred_recv(struct svc_rqst *rqstp)
1218 struct svc_deferred_req *dr = rqstp->rq_deferred;
1220 /* setup iov_base past transport header */
1221 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1222 /* The iov_len does not include the transport header bytes */
1223 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1224 rqstp->rq_arg.page_len = 0;
1225 /* The rq_arg.len includes the transport header bytes */
1226 rqstp->rq_arg.len = dr->argslen<<2;
1227 rqstp->rq_prot = dr->prot;
1228 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1229 rqstp->rq_addrlen = dr->addrlen;
1230 /* Save off transport header len in case we get deferred again */
1231 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1232 rqstp->rq_daddr = dr->daddr;
1233 rqstp->rq_respages = rqstp->rq_pages;
1234 return (dr->argslen<<2) - dr->xprt_hlen;
1238 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1240 struct svc_deferred_req *dr = NULL;
1242 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1244 spin_lock(&xprt->xpt_lock);
1245 if (!list_empty(&xprt->xpt_deferred)) {
1246 dr = list_entry(xprt->xpt_deferred.next,
1247 struct svc_deferred_req,
1249 list_del_init(&dr->handle.recent);
1250 trace_svc_revisit_deferred(dr);
1252 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1253 spin_unlock(&xprt->xpt_lock);
1258 * svc_find_xprt - find an RPC transport instance
1259 * @serv: pointer to svc_serv to search
1260 * @xcl_name: C string containing transport's class name
1261 * @net: owner net pointer
1262 * @af: Address family of transport's local address
1263 * @port: transport's IP port number
1265 * Return the transport instance pointer for the endpoint accepting
1266 * connections/peer traffic from the specified transport class,
1267 * address family and port.
1269 * Specifying 0 for the address family or port is effectively a
1270 * wild-card, and will result in matching the first transport in the
1271 * service's list that has a matching class name.
1273 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1274 struct net *net, const sa_family_t af,
1275 const unsigned short port)
1277 struct svc_xprt *xprt;
1278 struct svc_xprt *found = NULL;
1280 /* Sanity check the args */
1281 if (serv == NULL || xcl_name == NULL)
1284 spin_lock_bh(&serv->sv_lock);
1285 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1286 if (xprt->xpt_net != net)
1288 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1290 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1292 if (port != 0 && port != svc_xprt_local_port(xprt))
1298 spin_unlock_bh(&serv->sv_lock);
1301 EXPORT_SYMBOL_GPL(svc_find_xprt);
1303 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1304 char *pos, int remaining)
1308 len = snprintf(pos, remaining, "%s %u\n",
1309 xprt->xpt_class->xcl_name,
1310 svc_xprt_local_port(xprt));
1311 if (len >= remaining)
1312 return -ENAMETOOLONG;
1317 * svc_xprt_names - format a buffer with a list of transport names
1318 * @serv: pointer to an RPC service
1319 * @buf: pointer to a buffer to be filled in
1320 * @buflen: length of buffer to be filled in
1322 * Fills in @buf with a string containing a list of transport names,
1323 * each name terminated with '\n'.
1325 * Returns positive length of the filled-in string on success; otherwise
1326 * a negative errno value is returned if an error occurs.
1328 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1330 struct svc_xprt *xprt;
1334 /* Sanity check args */
1338 spin_lock_bh(&serv->sv_lock);
1342 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1343 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1355 spin_unlock_bh(&serv->sv_lock);
1358 EXPORT_SYMBOL_GPL(svc_xprt_names);
1361 /*----------------------------------------------------------------------------*/
1363 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1365 unsigned int pidx = (unsigned int)*pos;
1366 struct svc_serv *serv = m->private;
1368 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1371 return SEQ_START_TOKEN;
1372 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1375 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1377 struct svc_pool *pool = p;
1378 struct svc_serv *serv = m->private;
1380 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1382 if (p == SEQ_START_TOKEN) {
1383 pool = &serv->sv_pools[0];
1385 unsigned int pidx = (pool - &serv->sv_pools[0]);
1386 if (pidx < serv->sv_nrpools-1)
1387 pool = &serv->sv_pools[pidx+1];
1395 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1399 static int svc_pool_stats_show(struct seq_file *m, void *p)
1401 struct svc_pool *pool = p;
1403 if (p == SEQ_START_TOKEN) {
1404 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1408 seq_printf(m, "%u %lu %lu %lu %lu\n",
1410 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1411 pool->sp_stats.sockets_queued,
1412 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1413 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1418 static const struct seq_operations svc_pool_stats_seq_ops = {
1419 .start = svc_pool_stats_start,
1420 .next = svc_pool_stats_next,
1421 .stop = svc_pool_stats_stop,
1422 .show = svc_pool_stats_show,
1425 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1429 err = seq_open(file, &svc_pool_stats_seq_ops);
1431 ((struct seq_file *) file->private_data)->private = serv;
1434 EXPORT_SYMBOL(svc_pool_stats_open);
1436 /*----------------------------------------------------------------------------*/