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://nfsv4bat.org/Documents/ConnectAThon/1996/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 * svc_print_xprts - Format the transport list for printing
109 * @buf: target buffer for formatted address
110 * @maxlen: length of target buffer
112 * Fills in @buf with a string containing a list of transport names, each name
113 * terminated with '\n'. If the buffer is too small, some entries may be
114 * missing, but it is guaranteed that all lines in the output buffer are
117 * Returns positive length of the filled-in string.
119 int svc_print_xprts(char *buf, int maxlen)
121 struct svc_xprt_class *xcl;
126 spin_lock(&svc_xprt_class_lock);
127 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
130 slen = snprintf(tmpstr, sizeof(tmpstr), "%s %d\n",
131 xcl->xcl_name, xcl->xcl_max_payload);
132 if (slen >= sizeof(tmpstr) || len + slen >= maxlen)
137 spin_unlock(&svc_xprt_class_lock);
142 static void svc_xprt_free(struct kref *kref)
144 struct svc_xprt *xprt =
145 container_of(kref, struct svc_xprt, xpt_ref);
146 struct module *owner = xprt->xpt_class->xcl_owner;
147 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
148 svcauth_unix_info_release(xprt);
149 put_cred(xprt->xpt_cred);
150 put_net(xprt->xpt_net);
151 /* See comment on corresponding get in xs_setup_bc_tcp(): */
152 if (xprt->xpt_bc_xprt)
153 xprt_put(xprt->xpt_bc_xprt);
154 if (xprt->xpt_bc_xps)
155 xprt_switch_put(xprt->xpt_bc_xps);
156 trace_svc_xprt_free(xprt);
157 xprt->xpt_ops->xpo_free(xprt);
161 void svc_xprt_put(struct svc_xprt *xprt)
163 kref_put(&xprt->xpt_ref, svc_xprt_free);
165 EXPORT_SYMBOL_GPL(svc_xprt_put);
168 * Called by transport drivers to initialize the transport independent
169 * portion of the transport instance.
171 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
172 struct svc_xprt *xprt, struct svc_serv *serv)
174 memset(xprt, 0, sizeof(*xprt));
175 xprt->xpt_class = xcl;
176 xprt->xpt_ops = xcl->xcl_ops;
177 kref_init(&xprt->xpt_ref);
178 xprt->xpt_server = serv;
179 INIT_LIST_HEAD(&xprt->xpt_list);
180 INIT_LIST_HEAD(&xprt->xpt_ready);
181 INIT_LIST_HEAD(&xprt->xpt_deferred);
182 INIT_LIST_HEAD(&xprt->xpt_users);
183 mutex_init(&xprt->xpt_mutex);
184 spin_lock_init(&xprt->xpt_lock);
185 set_bit(XPT_BUSY, &xprt->xpt_flags);
186 xprt->xpt_net = get_net(net);
187 strcpy(xprt->xpt_remotebuf, "uninitialized");
189 EXPORT_SYMBOL_GPL(svc_xprt_init);
191 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
192 struct svc_serv *serv,
195 const unsigned short port,
198 struct sockaddr_in sin = {
199 .sin_family = AF_INET,
200 .sin_addr.s_addr = htonl(INADDR_ANY),
201 .sin_port = htons(port),
203 #if IS_ENABLED(CONFIG_IPV6)
204 struct sockaddr_in6 sin6 = {
205 .sin6_family = AF_INET6,
206 .sin6_addr = IN6ADDR_ANY_INIT,
207 .sin6_port = htons(port),
210 struct svc_xprt *xprt;
211 struct sockaddr *sap;
216 sap = (struct sockaddr *)&sin;
219 #if IS_ENABLED(CONFIG_IPV6)
221 sap = (struct sockaddr *)&sin6;
226 return ERR_PTR(-EAFNOSUPPORT);
229 xprt = xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
231 trace_svc_xprt_create_err(serv->sv_program->pg_name,
232 xcl->xcl_name, sap, xprt);
237 * svc_xprt_received conditionally queues the transport for processing
238 * by another thread. The caller must hold the XPT_BUSY bit and must
239 * not thereafter touch transport data.
241 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
242 * insufficient) data.
244 static void svc_xprt_received(struct svc_xprt *xprt)
246 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
247 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
251 /* As soon as we clear busy, the xprt could be closed and
252 * 'put', so we need a reference to call svc_enqueue_xprt with:
255 smp_mb__before_atomic();
256 clear_bit(XPT_BUSY, &xprt->xpt_flags);
257 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
261 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
263 clear_bit(XPT_TEMP, &new->xpt_flags);
264 spin_lock_bh(&serv->sv_lock);
265 list_add(&new->xpt_list, &serv->sv_permsocks);
266 spin_unlock_bh(&serv->sv_lock);
267 svc_xprt_received(new);
270 static int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
271 struct net *net, const int family,
272 const unsigned short port, int flags,
273 const struct cred *cred)
275 struct svc_xprt_class *xcl;
277 spin_lock(&svc_xprt_class_lock);
278 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
279 struct svc_xprt *newxprt;
280 unsigned short newport;
282 if (strcmp(xprt_name, xcl->xcl_name))
285 if (!try_module_get(xcl->xcl_owner))
288 spin_unlock(&svc_xprt_class_lock);
289 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
290 if (IS_ERR(newxprt)) {
291 module_put(xcl->xcl_owner);
292 return PTR_ERR(newxprt);
294 newxprt->xpt_cred = get_cred(cred);
295 svc_add_new_perm_xprt(serv, newxprt);
296 newport = svc_xprt_local_port(newxprt);
300 spin_unlock(&svc_xprt_class_lock);
301 /* This errno is exposed to user space. Provide a reasonable
302 * perror msg for a bad transport. */
303 return -EPROTONOSUPPORT;
306 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
307 struct net *net, const int family,
308 const unsigned short port, int flags,
309 const struct cred *cred)
313 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags, cred);
314 if (err == -EPROTONOSUPPORT) {
315 request_module("svc%s", xprt_name);
316 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags, cred);
320 EXPORT_SYMBOL_GPL(svc_create_xprt);
323 * Copy the local and remote xprt addresses to the rqstp structure
325 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
327 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
328 rqstp->rq_addrlen = xprt->xpt_remotelen;
331 * Destination address in request is needed for binding the
332 * source address in RPC replies/callbacks later.
334 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
335 rqstp->rq_daddrlen = xprt->xpt_locallen;
337 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
340 * svc_print_addr - Format rq_addr field for printing
341 * @rqstp: svc_rqst struct containing address to print
342 * @buf: target buffer for formatted address
343 * @len: length of target buffer
346 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
348 return __svc_print_addr(svc_addr(rqstp), buf, len);
350 EXPORT_SYMBOL_GPL(svc_print_addr);
352 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
354 unsigned int limit = svc_rpc_per_connection_limit;
355 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
357 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
360 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
362 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
363 if (!svc_xprt_slots_in_range(xprt))
365 atomic_inc(&xprt->xpt_nr_rqsts);
366 set_bit(RQ_DATA, &rqstp->rq_flags);
371 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
373 struct svc_xprt *xprt = rqstp->rq_xprt;
374 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
375 atomic_dec(&xprt->xpt_nr_rqsts);
376 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
377 svc_xprt_enqueue(xprt);
381 static bool svc_xprt_ready(struct svc_xprt *xprt)
383 unsigned long xpt_flags;
386 * If another cpu has recently updated xpt_flags,
387 * sk_sock->flags, xpt_reserved, or xpt_nr_rqsts, we need to
388 * know about it; otherwise it's possible that both that cpu and
389 * this one could call svc_xprt_enqueue() without either
390 * svc_xprt_enqueue() recognizing that the conditions below
391 * are satisfied, and we could stall indefinitely:
394 xpt_flags = READ_ONCE(xprt->xpt_flags);
396 if (xpt_flags & (BIT(XPT_CONN) | BIT(XPT_CLOSE)))
398 if (xpt_flags & (BIT(XPT_DATA) | BIT(XPT_DEFERRED))) {
399 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
400 svc_xprt_slots_in_range(xprt))
402 trace_svc_xprt_no_write_space(xprt);
408 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
410 struct svc_pool *pool;
411 struct svc_rqst *rqstp = NULL;
414 if (!svc_xprt_ready(xprt))
417 /* Mark transport as busy. It will remain in this state until
418 * the provider calls svc_xprt_received. We update XPT_BUSY
419 * atomically because it also guards against trying to enqueue
420 * the transport twice.
422 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
426 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
428 atomic_long_inc(&pool->sp_stats.packets);
430 spin_lock_bh(&pool->sp_lock);
431 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
432 pool->sp_stats.sockets_queued++;
433 spin_unlock_bh(&pool->sp_lock);
435 /* find a thread for this xprt */
437 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
438 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
440 atomic_long_inc(&pool->sp_stats.threads_woken);
441 rqstp->rq_qtime = ktime_get();
442 wake_up_process(rqstp->rq_task);
445 set_bit(SP_CONGESTED, &pool->sp_flags);
450 trace_svc_xprt_do_enqueue(xprt, rqstp);
452 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
455 * Queue up a transport with data pending. If there are idle nfsd
456 * processes, wake 'em up.
459 void svc_xprt_enqueue(struct svc_xprt *xprt)
461 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
463 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
465 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
468 * Dequeue the first transport, if there is one.
470 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
472 struct svc_xprt *xprt = NULL;
474 if (list_empty(&pool->sp_sockets))
477 spin_lock_bh(&pool->sp_lock);
478 if (likely(!list_empty(&pool->sp_sockets))) {
479 xprt = list_first_entry(&pool->sp_sockets,
480 struct svc_xprt, xpt_ready);
481 list_del_init(&xprt->xpt_ready);
484 spin_unlock_bh(&pool->sp_lock);
490 * svc_reserve - change the space reserved for the reply to a request.
491 * @rqstp: The request in question
492 * @space: new max space to reserve
494 * Each request reserves some space on the output queue of the transport
495 * to make sure the reply fits. This function reduces that reserved
496 * space to be the amount of space used already, plus @space.
499 void svc_reserve(struct svc_rqst *rqstp, int space)
501 struct svc_xprt *xprt = rqstp->rq_xprt;
503 space += rqstp->rq_res.head[0].iov_len;
505 if (xprt && space < rqstp->rq_reserved) {
506 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
507 rqstp->rq_reserved = space;
508 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
509 svc_xprt_enqueue(xprt);
512 EXPORT_SYMBOL_GPL(svc_reserve);
514 static void svc_xprt_release(struct svc_rqst *rqstp)
516 struct svc_xprt *xprt = rqstp->rq_xprt;
518 xprt->xpt_ops->xpo_release_rqst(rqstp);
520 kfree(rqstp->rq_deferred);
521 rqstp->rq_deferred = NULL;
523 svc_free_res_pages(rqstp);
524 rqstp->rq_res.page_len = 0;
525 rqstp->rq_res.page_base = 0;
527 /* Reset response buffer and release
529 * But first, check that enough space was reserved
530 * for the reply, otherwise we have a bug!
532 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
533 printk(KERN_ERR "RPC request reserved %d but used %d\n",
537 rqstp->rq_res.head[0].iov_len = 0;
538 svc_reserve(rqstp, 0);
539 svc_xprt_release_slot(rqstp);
540 rqstp->rq_xprt = NULL;
545 * Some svc_serv's will have occasional work to do, even when a xprt is not
546 * waiting to be serviced. This function is there to "kick" a task in one of
547 * those services so that it can wake up and do that work. Note that we only
548 * bother with pool 0 as we don't need to wake up more than one thread for
551 void svc_wake_up(struct svc_serv *serv)
553 struct svc_rqst *rqstp;
554 struct svc_pool *pool;
556 pool = &serv->sv_pools[0];
559 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
560 /* skip any that aren't queued */
561 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
564 wake_up_process(rqstp->rq_task);
565 trace_svc_wake_up(rqstp->rq_task->pid);
570 /* No free entries available */
571 set_bit(SP_TASK_PENDING, &pool->sp_flags);
573 trace_svc_wake_up(0);
575 EXPORT_SYMBOL_GPL(svc_wake_up);
577 int svc_port_is_privileged(struct sockaddr *sin)
579 switch (sin->sa_family) {
581 return ntohs(((struct sockaddr_in *)sin)->sin_port)
584 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
592 * Make sure that we don't have too many active connections. If we have,
593 * something must be dropped. It's not clear what will happen if we allow
594 * "too many" connections, but when dealing with network-facing software,
595 * we have to code defensively. Here we do that by imposing hard limits.
597 * There's no point in trying to do random drop here for DoS
598 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
599 * attacker can easily beat that.
601 * The only somewhat efficient mechanism would be if drop old
602 * connections from the same IP first. But right now we don't even
603 * record the client IP in svc_sock.
605 * single-threaded services that expect a lot of clients will probably
606 * need to set sv_maxconn to override the default value which is based
607 * on the number of threads
609 static void svc_check_conn_limits(struct svc_serv *serv)
611 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
612 (serv->sv_nrthreads+3) * 20;
614 if (serv->sv_tmpcnt > limit) {
615 struct svc_xprt *xprt = NULL;
616 spin_lock_bh(&serv->sv_lock);
617 if (!list_empty(&serv->sv_tempsocks)) {
618 /* Try to help the admin */
619 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
620 serv->sv_name, serv->sv_maxconn ?
621 "max number of connections" :
622 "number of threads");
624 * Always select the oldest connection. It's not fair,
627 xprt = list_entry(serv->sv_tempsocks.prev,
630 set_bit(XPT_CLOSE, &xprt->xpt_flags);
633 spin_unlock_bh(&serv->sv_lock);
636 svc_xprt_enqueue(xprt);
642 static int svc_alloc_arg(struct svc_rqst *rqstp)
644 struct svc_serv *serv = rqstp->rq_server;
649 /* now allocate needed pages. If we get a failure, sleep briefly */
650 pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
651 if (pages > RPCSVC_MAXPAGES) {
652 pr_warn_once("svc: warning: pages=%u > RPCSVC_MAXPAGES=%lu\n",
653 pages, RPCSVC_MAXPAGES);
654 /* use as many pages as possible */
655 pages = RPCSVC_MAXPAGES;
657 for (i = 0; i < pages ; i++)
658 while (rqstp->rq_pages[i] == NULL) {
659 struct page *p = alloc_page(GFP_KERNEL);
661 set_current_state(TASK_INTERRUPTIBLE);
662 if (signalled() || kthread_should_stop()) {
663 set_current_state(TASK_RUNNING);
666 schedule_timeout(msecs_to_jiffies(500));
668 rqstp->rq_pages[i] = p;
670 rqstp->rq_page_end = &rqstp->rq_pages[i];
671 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
673 /* Make arg->head point to first page and arg->pages point to rest */
674 arg = &rqstp->rq_arg;
675 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
676 arg->head[0].iov_len = PAGE_SIZE;
677 arg->pages = rqstp->rq_pages + 1;
679 /* save at least one page for response */
680 arg->page_len = (pages-2)*PAGE_SIZE;
681 arg->len = (pages-1)*PAGE_SIZE;
682 arg->tail[0].iov_len = 0;
687 rqst_should_sleep(struct svc_rqst *rqstp)
689 struct svc_pool *pool = rqstp->rq_pool;
691 /* did someone call svc_wake_up? */
692 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
695 /* was a socket queued? */
696 if (!list_empty(&pool->sp_sockets))
699 /* are we shutting down? */
700 if (signalled() || kthread_should_stop())
703 /* are we freezing? */
704 if (freezing(current))
710 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
712 struct svc_pool *pool = rqstp->rq_pool;
715 /* rq_xprt should be clear on entry */
716 WARN_ON_ONCE(rqstp->rq_xprt);
718 rqstp->rq_xprt = svc_xprt_dequeue(pool);
723 * We have to be able to interrupt this wait
724 * to bring down the daemons ...
726 set_current_state(TASK_INTERRUPTIBLE);
727 smp_mb__before_atomic();
728 clear_bit(SP_CONGESTED, &pool->sp_flags);
729 clear_bit(RQ_BUSY, &rqstp->rq_flags);
730 smp_mb__after_atomic();
732 if (likely(rqst_should_sleep(rqstp)))
733 time_left = schedule_timeout(timeout);
735 __set_current_state(TASK_RUNNING);
739 set_bit(RQ_BUSY, &rqstp->rq_flags);
740 smp_mb__after_atomic();
741 rqstp->rq_xprt = svc_xprt_dequeue(pool);
746 atomic_long_inc(&pool->sp_stats.threads_timedout);
748 if (signalled() || kthread_should_stop())
749 return ERR_PTR(-EINTR);
750 return ERR_PTR(-EAGAIN);
752 /* Normally we will wait up to 5 seconds for any required
753 * cache information to be provided.
755 if (!test_bit(SP_CONGESTED, &pool->sp_flags))
756 rqstp->rq_chandle.thread_wait = 5*HZ;
758 rqstp->rq_chandle.thread_wait = 1*HZ;
759 trace_svc_xprt_dequeue(rqstp);
760 return rqstp->rq_xprt;
763 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
765 spin_lock_bh(&serv->sv_lock);
766 set_bit(XPT_TEMP, &newxpt->xpt_flags);
767 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
769 if (serv->sv_temptimer.function == NULL) {
770 /* setup timer to age temp transports */
771 serv->sv_temptimer.function = svc_age_temp_xprts;
772 mod_timer(&serv->sv_temptimer,
773 jiffies + svc_conn_age_period * HZ);
775 spin_unlock_bh(&serv->sv_lock);
776 svc_xprt_received(newxpt);
779 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
781 struct svc_serv *serv = rqstp->rq_server;
784 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
785 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
786 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
787 svc_delete_xprt(xprt);
788 /* Leave XPT_BUSY set on the dead xprt: */
791 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
792 struct svc_xprt *newxpt;
794 * We know this module_get will succeed because the
795 * listener holds a reference too
797 __module_get(xprt->xpt_class->xcl_owner);
798 svc_check_conn_limits(xprt->xpt_server);
799 newxpt = xprt->xpt_ops->xpo_accept(xprt);
801 newxpt->xpt_cred = get_cred(xprt->xpt_cred);
802 svc_add_new_temp_xprt(serv, newxpt);
803 trace_svc_xprt_accept(newxpt, serv->sv_name);
805 module_put(xprt->xpt_class->xcl_owner);
806 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
807 /* XPT_DATA|XPT_DEFERRED case: */
808 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
809 rqstp, rqstp->rq_pool->sp_id, xprt,
810 kref_read(&xprt->xpt_ref));
811 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
812 if (rqstp->rq_deferred)
813 len = svc_deferred_recv(rqstp);
815 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
817 trace_svc_xdr_recvfrom(rqstp, &rqstp->rq_arg);
818 rqstp->rq_stime = ktime_get();
819 rqstp->rq_reserved = serv->sv_max_mesg;
820 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
822 /* clear XPT_BUSY: */
823 svc_xprt_received(xprt);
825 trace_svc_handle_xprt(xprt, len);
830 * Receive the next request on any transport. This code is carefully
831 * organised not to touch any cachelines in the shared svc_serv
832 * structure, only cachelines in the local svc_pool.
834 int svc_recv(struct svc_rqst *rqstp, long timeout)
836 struct svc_xprt *xprt = NULL;
837 struct svc_serv *serv = rqstp->rq_server;
840 err = svc_alloc_arg(rqstp);
847 if (signalled() || kthread_should_stop())
850 xprt = svc_get_next_xprt(rqstp, timeout);
856 len = svc_handle_xprt(rqstp, xprt);
858 /* No data, incomplete (TCP) read, or accept() */
863 clear_bit(XPT_OLD, &xprt->xpt_flags);
865 xprt->xpt_ops->xpo_secure_port(rqstp);
866 rqstp->rq_chandle.defer = svc_defer;
867 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
870 serv->sv_stats->netcnt++;
871 trace_svc_recv(rqstp, len);
874 rqstp->rq_res.len = 0;
875 svc_xprt_release(rqstp);
879 EXPORT_SYMBOL_GPL(svc_recv);
884 void svc_drop(struct svc_rqst *rqstp)
886 trace_svc_drop(rqstp);
887 svc_xprt_release(rqstp);
889 EXPORT_SYMBOL_GPL(svc_drop);
892 * Return reply to client.
894 int svc_send(struct svc_rqst *rqstp)
896 struct svc_xprt *xprt;
900 xprt = rqstp->rq_xprt;
904 /* calculate over-all length */
906 xb->len = xb->head[0].iov_len +
909 trace_svc_xdr_sendto(rqstp, xb);
910 trace_svc_stats_latency(rqstp);
912 len = xprt->xpt_ops->xpo_sendto(rqstp);
914 trace_svc_send(rqstp, len);
915 svc_xprt_release(rqstp);
917 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
924 * Timer function to close old temporary transports, using
925 * a mark-and-sweep algorithm.
927 static void svc_age_temp_xprts(struct timer_list *t)
929 struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
930 struct svc_xprt *xprt;
931 struct list_head *le, *next;
933 dprintk("svc_age_temp_xprts\n");
935 if (!spin_trylock_bh(&serv->sv_lock)) {
936 /* busy, try again 1 sec later */
937 dprintk("svc_age_temp_xprts: busy\n");
938 mod_timer(&serv->sv_temptimer, jiffies + HZ);
942 list_for_each_safe(le, next, &serv->sv_tempsocks) {
943 xprt = list_entry(le, struct svc_xprt, xpt_list);
945 /* First time through, just mark it OLD. Second time
946 * through, close it. */
947 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
949 if (kref_read(&xprt->xpt_ref) > 1 ||
950 test_bit(XPT_BUSY, &xprt->xpt_flags))
953 set_bit(XPT_CLOSE, &xprt->xpt_flags);
954 dprintk("queuing xprt %p for closing\n", xprt);
956 /* a thread will dequeue and close it soon */
957 svc_xprt_enqueue(xprt);
959 spin_unlock_bh(&serv->sv_lock);
961 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
964 /* Close temporary transports whose xpt_local matches server_addr immediately
965 * instead of waiting for them to be picked up by the timer.
967 * This is meant to be called from a notifier_block that runs when an ip
968 * address is deleted.
970 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
972 struct svc_xprt *xprt;
973 struct list_head *le, *next;
974 LIST_HEAD(to_be_closed);
976 spin_lock_bh(&serv->sv_lock);
977 list_for_each_safe(le, next, &serv->sv_tempsocks) {
978 xprt = list_entry(le, struct svc_xprt, xpt_list);
979 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
981 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
982 list_move(le, &to_be_closed);
985 spin_unlock_bh(&serv->sv_lock);
987 while (!list_empty(&to_be_closed)) {
988 le = to_be_closed.next;
990 xprt = list_entry(le, struct svc_xprt, xpt_list);
991 set_bit(XPT_CLOSE, &xprt->xpt_flags);
992 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
993 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
995 svc_xprt_enqueue(xprt);
998 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1000 static void call_xpt_users(struct svc_xprt *xprt)
1002 struct svc_xpt_user *u;
1004 spin_lock(&xprt->xpt_lock);
1005 while (!list_empty(&xprt->xpt_users)) {
1006 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1007 list_del_init(&u->list);
1010 spin_unlock(&xprt->xpt_lock);
1014 * Remove a dead transport
1016 static void svc_delete_xprt(struct svc_xprt *xprt)
1018 struct svc_serv *serv = xprt->xpt_server;
1019 struct svc_deferred_req *dr;
1021 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1024 trace_svc_xprt_detach(xprt);
1025 xprt->xpt_ops->xpo_detach(xprt);
1026 if (xprt->xpt_bc_xprt)
1027 xprt->xpt_bc_xprt->ops->close(xprt->xpt_bc_xprt);
1029 spin_lock_bh(&serv->sv_lock);
1030 list_del_init(&xprt->xpt_list);
1031 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1032 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1034 spin_unlock_bh(&serv->sv_lock);
1036 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1039 call_xpt_users(xprt);
1043 void svc_close_xprt(struct svc_xprt *xprt)
1045 trace_svc_xprt_close(xprt);
1046 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1047 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1048 /* someone else will have to effect the close */
1051 * We expect svc_close_xprt() to work even when no threads are
1052 * running (e.g., while configuring the server before starting
1053 * any threads), so if the transport isn't busy, we delete
1056 svc_delete_xprt(xprt);
1058 EXPORT_SYMBOL_GPL(svc_close_xprt);
1060 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1062 struct svc_xprt *xprt;
1065 spin_lock(&serv->sv_lock);
1066 list_for_each_entry(xprt, xprt_list, xpt_list) {
1067 if (xprt->xpt_net != net)
1070 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1071 svc_xprt_enqueue(xprt);
1073 spin_unlock(&serv->sv_lock);
1077 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1079 struct svc_pool *pool;
1080 struct svc_xprt *xprt;
1081 struct svc_xprt *tmp;
1084 for (i = 0; i < serv->sv_nrpools; i++) {
1085 pool = &serv->sv_pools[i];
1087 spin_lock_bh(&pool->sp_lock);
1088 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1089 if (xprt->xpt_net != net)
1091 list_del_init(&xprt->xpt_ready);
1092 spin_unlock_bh(&pool->sp_lock);
1095 spin_unlock_bh(&pool->sp_lock);
1100 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1102 struct svc_xprt *xprt;
1104 while ((xprt = svc_dequeue_net(serv, net))) {
1105 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1106 svc_delete_xprt(xprt);
1111 * Server threads may still be running (especially in the case where the
1112 * service is still running in other network namespaces).
1114 * So we shut down sockets the same way we would on a running server, by
1115 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1116 * the close. In the case there are no such other threads,
1117 * threads running, svc_clean_up_xprts() does a simple version of a
1118 * server's main event loop, and in the case where there are other
1119 * threads, we may need to wait a little while and then check again to
1120 * see if they're done.
1122 void svc_close_net(struct svc_serv *serv, struct net *net)
1126 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1127 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1129 svc_clean_up_xprts(serv, net);
1135 * Handle defer and revisit of requests
1138 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1140 struct svc_deferred_req *dr =
1141 container_of(dreq, struct svc_deferred_req, handle);
1142 struct svc_xprt *xprt = dr->xprt;
1144 spin_lock(&xprt->xpt_lock);
1145 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1146 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1147 spin_unlock(&xprt->xpt_lock);
1148 trace_svc_defer_drop(dr);
1154 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1155 spin_unlock(&xprt->xpt_lock);
1156 trace_svc_defer_queue(dr);
1157 svc_xprt_enqueue(xprt);
1162 * Save the request off for later processing. The request buffer looks
1165 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1167 * This code can only handle requests that consist of an xprt-header
1170 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1172 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1173 struct svc_deferred_req *dr;
1175 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1176 return NULL; /* if more than a page, give up FIXME */
1177 if (rqstp->rq_deferred) {
1178 dr = rqstp->rq_deferred;
1179 rqstp->rq_deferred = NULL;
1183 /* FIXME maybe discard if size too large */
1184 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1185 dr = kmalloc(size, GFP_KERNEL);
1189 dr->handle.owner = rqstp->rq_server;
1190 dr->prot = rqstp->rq_prot;
1191 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1192 dr->addrlen = rqstp->rq_addrlen;
1193 dr->daddr = rqstp->rq_daddr;
1194 dr->argslen = rqstp->rq_arg.len >> 2;
1195 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1197 /* back up head to the start of the buffer and copy */
1198 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1199 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1202 trace_svc_defer(rqstp);
1203 svc_xprt_get(rqstp->rq_xprt);
1204 dr->xprt = rqstp->rq_xprt;
1205 set_bit(RQ_DROPME, &rqstp->rq_flags);
1207 dr->handle.revisit = svc_revisit;
1212 * recv data from a deferred request into an active one
1214 static noinline int svc_deferred_recv(struct svc_rqst *rqstp)
1216 struct svc_deferred_req *dr = rqstp->rq_deferred;
1218 trace_svc_defer_recv(dr);
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);
1251 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1252 spin_unlock(&xprt->xpt_lock);
1257 * svc_find_xprt - find an RPC transport instance
1258 * @serv: pointer to svc_serv to search
1259 * @xcl_name: C string containing transport's class name
1260 * @net: owner net pointer
1261 * @af: Address family of transport's local address
1262 * @port: transport's IP port number
1264 * Return the transport instance pointer for the endpoint accepting
1265 * connections/peer traffic from the specified transport class,
1266 * address family and port.
1268 * Specifying 0 for the address family or port is effectively a
1269 * wild-card, and will result in matching the first transport in the
1270 * service's list that has a matching class name.
1272 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1273 struct net *net, const sa_family_t af,
1274 const unsigned short port)
1276 struct svc_xprt *xprt;
1277 struct svc_xprt *found = NULL;
1279 /* Sanity check the args */
1280 if (serv == NULL || xcl_name == NULL)
1283 spin_lock_bh(&serv->sv_lock);
1284 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1285 if (xprt->xpt_net != net)
1287 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1289 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1291 if (port != 0 && port != svc_xprt_local_port(xprt))
1297 spin_unlock_bh(&serv->sv_lock);
1300 EXPORT_SYMBOL_GPL(svc_find_xprt);
1302 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1303 char *pos, int remaining)
1307 len = snprintf(pos, remaining, "%s %u\n",
1308 xprt->xpt_class->xcl_name,
1309 svc_xprt_local_port(xprt));
1310 if (len >= remaining)
1311 return -ENAMETOOLONG;
1316 * svc_xprt_names - format a buffer with a list of transport names
1317 * @serv: pointer to an RPC service
1318 * @buf: pointer to a buffer to be filled in
1319 * @buflen: length of buffer to be filled in
1321 * Fills in @buf with a string containing a list of transport names,
1322 * each name terminated with '\n'.
1324 * Returns positive length of the filled-in string on success; otherwise
1325 * a negative errno value is returned if an error occurs.
1327 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1329 struct svc_xprt *xprt;
1333 /* Sanity check args */
1337 spin_lock_bh(&serv->sv_lock);
1341 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1342 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1354 spin_unlock_bh(&serv->sv_lock);
1357 EXPORT_SYMBOL_GPL(svc_xprt_names);
1360 /*----------------------------------------------------------------------------*/
1362 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1364 unsigned int pidx = (unsigned int)*pos;
1365 struct svc_serv *serv = m->private;
1367 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1370 return SEQ_START_TOKEN;
1371 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1374 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1376 struct svc_pool *pool = p;
1377 struct svc_serv *serv = m->private;
1379 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1381 if (p == SEQ_START_TOKEN) {
1382 pool = &serv->sv_pools[0];
1384 unsigned int pidx = (pool - &serv->sv_pools[0]);
1385 if (pidx < serv->sv_nrpools-1)
1386 pool = &serv->sv_pools[pidx+1];
1394 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1398 static int svc_pool_stats_show(struct seq_file *m, void *p)
1400 struct svc_pool *pool = p;
1402 if (p == SEQ_START_TOKEN) {
1403 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1407 seq_printf(m, "%u %lu %lu %lu %lu\n",
1409 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1410 pool->sp_stats.sockets_queued,
1411 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1412 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1417 static const struct seq_operations svc_pool_stats_seq_ops = {
1418 .start = svc_pool_stats_start,
1419 .next = svc_pool_stats_next,
1420 .stop = svc_pool_stats_stop,
1421 .show = svc_pool_stats_show,
1424 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1428 err = seq_open(file, &svc_pool_stats_seq_ops);
1430 ((struct seq_file *) file->private_data)->private = serv;
1433 EXPORT_SYMBOL(svc_pool_stats_open);
1435 /*----------------------------------------------------------------------------*/