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/stats.h>
14 #include <linux/sunrpc/svc_xprt.h>
15 #include <linux/sunrpc/svcsock.h>
16 #include <linux/sunrpc/xprt.h>
17 #include <linux/module.h>
19 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
21 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
22 static int svc_deferred_recv(struct svc_rqst *rqstp);
23 static struct cache_deferred_req *svc_defer(struct cache_req *req);
24 static void svc_age_temp_xprts(unsigned long closure);
25 static void svc_delete_xprt(struct svc_xprt *xprt);
26 static void svc_xprt_do_enqueue(struct svc_xprt *xprt);
28 /* apparently the "standard" is that clients close
29 * idle connections after 5 minutes, servers after
31 * http://www.connectathon.org/talks96/nfstcp.pdf
33 static int svc_conn_age_period = 6*60;
35 /* List of registered transport classes */
36 static DEFINE_SPINLOCK(svc_xprt_class_lock);
37 static LIST_HEAD(svc_xprt_class_list);
39 /* SMP locking strategy:
41 * svc_pool->sp_lock protects most of the fields of that pool.
42 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
43 * when both need to be taken (rare), svc_serv->sv_lock is first.
44 * BKL protects svc_serv->sv_nrthread.
45 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
46 * and the ->sk_info_authunix cache.
48 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
49 * enqueued multiply. During normal transport processing this bit
50 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
51 * Providers should not manipulate this bit directly.
53 * Some flags can be set to certain values at any time
54 * providing that certain rules are followed:
57 * - Can be set or cleared at any time.
58 * - After a set, svc_xprt_enqueue must be called to enqueue
59 * the transport for processing.
60 * - After a clear, the transport must be read/accepted.
61 * If this succeeds, it must be set again.
63 * - Can set at any time. It is never cleared.
65 * - Can only be set while XPT_BUSY is held which ensures
66 * that no other thread will be using the transport or will
67 * try to set XPT_DEAD.
70 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
72 struct svc_xprt_class *cl;
75 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
77 INIT_LIST_HEAD(&xcl->xcl_list);
78 spin_lock(&svc_xprt_class_lock);
79 /* Make sure there isn't already a class with the same name */
80 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
81 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
84 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
87 spin_unlock(&svc_xprt_class_lock);
90 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
92 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
94 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
95 spin_lock(&svc_xprt_class_lock);
96 list_del_init(&xcl->xcl_list);
97 spin_unlock(&svc_xprt_class_lock);
99 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
102 * Format the transport list for printing
104 int svc_print_xprts(char *buf, int maxlen)
106 struct svc_xprt_class *xcl;
111 spin_lock(&svc_xprt_class_lock);
112 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
115 sprintf(tmpstr, "%s %d\n", xcl->xcl_name, xcl->xcl_max_payload);
116 slen = strlen(tmpstr);
117 if (len + slen > maxlen)
122 spin_unlock(&svc_xprt_class_lock);
127 static void svc_xprt_free(struct kref *kref)
129 struct svc_xprt *xprt =
130 container_of(kref, struct svc_xprt, xpt_ref);
131 struct module *owner = xprt->xpt_class->xcl_owner;
132 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
133 svcauth_unix_info_release(xprt);
134 put_net(xprt->xpt_net);
135 /* See comment on corresponding get in xs_setup_bc_tcp(): */
136 if (xprt->xpt_bc_xprt)
137 xprt_put(xprt->xpt_bc_xprt);
138 xprt->xpt_ops->xpo_free(xprt);
142 void svc_xprt_put(struct svc_xprt *xprt)
144 kref_put(&xprt->xpt_ref, svc_xprt_free);
146 EXPORT_SYMBOL_GPL(svc_xprt_put);
149 * Called by transport drivers to initialize the transport independent
150 * portion of the transport instance.
152 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
153 struct svc_xprt *xprt, struct svc_serv *serv)
155 memset(xprt, 0, sizeof(*xprt));
156 xprt->xpt_class = xcl;
157 xprt->xpt_ops = xcl->xcl_ops;
158 kref_init(&xprt->xpt_ref);
159 xprt->xpt_server = serv;
160 INIT_LIST_HEAD(&xprt->xpt_list);
161 INIT_LIST_HEAD(&xprt->xpt_ready);
162 INIT_LIST_HEAD(&xprt->xpt_deferred);
163 INIT_LIST_HEAD(&xprt->xpt_users);
164 mutex_init(&xprt->xpt_mutex);
165 spin_lock_init(&xprt->xpt_lock);
166 set_bit(XPT_BUSY, &xprt->xpt_flags);
167 rpc_init_wait_queue(&xprt->xpt_bc_pending, "xpt_bc_pending");
168 xprt->xpt_net = get_net(net);
170 EXPORT_SYMBOL_GPL(svc_xprt_init);
172 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
173 struct svc_serv *serv,
176 const unsigned short port,
179 struct sockaddr_in sin = {
180 .sin_family = AF_INET,
181 .sin_addr.s_addr = htonl(INADDR_ANY),
182 .sin_port = htons(port),
184 #if IS_ENABLED(CONFIG_IPV6)
185 struct sockaddr_in6 sin6 = {
186 .sin6_family = AF_INET6,
187 .sin6_addr = IN6ADDR_ANY_INIT,
188 .sin6_port = htons(port),
191 struct sockaddr *sap;
196 sap = (struct sockaddr *)&sin;
199 #if IS_ENABLED(CONFIG_IPV6)
201 sap = (struct sockaddr *)&sin6;
206 return ERR_PTR(-EAFNOSUPPORT);
209 return xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
213 * svc_xprt_received conditionally queues the transport for processing
214 * by another thread. The caller must hold the XPT_BUSY bit and must
215 * not thereafter touch transport data.
217 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
218 * insufficient) data.
220 static void svc_xprt_received(struct svc_xprt *xprt)
222 WARN_ON_ONCE(!test_bit(XPT_BUSY, &xprt->xpt_flags));
223 if (!test_bit(XPT_BUSY, &xprt->xpt_flags))
225 /* As soon as we clear busy, the xprt could be closed and
226 * 'put', so we need a reference to call svc_xprt_do_enqueue with:
229 smp_mb__before_atomic();
230 clear_bit(XPT_BUSY, &xprt->xpt_flags);
231 svc_xprt_do_enqueue(xprt);
235 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
237 clear_bit(XPT_TEMP, &new->xpt_flags);
238 spin_lock_bh(&serv->sv_lock);
239 list_add(&new->xpt_list, &serv->sv_permsocks);
240 spin_unlock_bh(&serv->sv_lock);
241 svc_xprt_received(new);
244 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
245 struct net *net, const int family,
246 const unsigned short port, int flags)
248 struct svc_xprt_class *xcl;
250 dprintk("svc: creating transport %s[%d]\n", xprt_name, port);
251 spin_lock(&svc_xprt_class_lock);
252 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
253 struct svc_xprt *newxprt;
254 unsigned short newport;
256 if (strcmp(xprt_name, xcl->xcl_name))
259 if (!try_module_get(xcl->xcl_owner))
262 spin_unlock(&svc_xprt_class_lock);
263 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
264 if (IS_ERR(newxprt)) {
265 module_put(xcl->xcl_owner);
266 return PTR_ERR(newxprt);
268 svc_add_new_perm_xprt(serv, newxprt);
269 newport = svc_xprt_local_port(newxprt);
273 spin_unlock(&svc_xprt_class_lock);
274 dprintk("svc: transport %s not found\n", xprt_name);
276 /* This errno is exposed to user space. Provide a reasonable
277 * perror msg for a bad transport. */
278 return -EPROTONOSUPPORT;
280 EXPORT_SYMBOL_GPL(svc_create_xprt);
283 * Copy the local and remote xprt addresses to the rqstp structure
285 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
287 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
288 rqstp->rq_addrlen = xprt->xpt_remotelen;
291 * Destination address in request is needed for binding the
292 * source address in RPC replies/callbacks later.
294 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
295 rqstp->rq_daddrlen = xprt->xpt_locallen;
297 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
300 * svc_print_addr - Format rq_addr field for printing
301 * @rqstp: svc_rqst struct containing address to print
302 * @buf: target buffer for formatted address
303 * @len: length of target buffer
306 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
308 return __svc_print_addr(svc_addr(rqstp), buf, len);
310 EXPORT_SYMBOL_GPL(svc_print_addr);
313 * Queue up an idle server thread. Must have pool->sp_lock held.
314 * Note: this is really a stack rather than a queue, so that we only
315 * use as many different threads as we need, and the rest don't pollute
318 static void svc_thread_enqueue(struct svc_pool *pool, struct svc_rqst *rqstp)
320 list_add(&rqstp->rq_list, &pool->sp_threads);
324 * Dequeue an nfsd thread. Must have pool->sp_lock held.
326 static void svc_thread_dequeue(struct svc_pool *pool, struct svc_rqst *rqstp)
328 list_del(&rqstp->rq_list);
331 static bool svc_xprt_has_something_to_do(struct svc_xprt *xprt)
333 if (xprt->xpt_flags & ((1<<XPT_CONN)|(1<<XPT_CLOSE)))
335 if (xprt->xpt_flags & ((1<<XPT_DATA)|(1<<XPT_DEFERRED)))
336 return xprt->xpt_ops->xpo_has_wspace(xprt);
340 static void svc_xprt_do_enqueue(struct svc_xprt *xprt)
342 struct svc_pool *pool;
343 struct svc_rqst *rqstp;
346 if (!svc_xprt_has_something_to_do(xprt))
350 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
351 spin_lock_bh(&pool->sp_lock);
353 if (!list_empty(&pool->sp_threads) &&
354 !list_empty(&pool->sp_sockets))
357 "threads and transports both waiting??\n");
359 pool->sp_stats.packets++;
361 /* Mark transport as busy. It will remain in this state until
362 * the provider calls svc_xprt_received. We update XPT_BUSY
363 * atomically because it also guards against trying to enqueue
364 * the transport twice.
366 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags)) {
367 /* Don't enqueue transport while already enqueued */
368 dprintk("svc: transport %p busy, not enqueued\n", xprt);
372 if (!list_empty(&pool->sp_threads)) {
373 rqstp = list_entry(pool->sp_threads.next,
376 dprintk("svc: transport %p served by daemon %p\n",
378 svc_thread_dequeue(pool, rqstp);
381 "svc_xprt_enqueue: server %p, rq_xprt=%p!\n",
382 rqstp, rqstp->rq_xprt);
383 /* Note the order of the following 3 lines:
384 * We want to assign xprt to rqstp->rq_xprt only _after_
385 * we've woken up the process, so that we don't race with
386 * the lockless check in svc_get_next_xprt().
389 wake_up_process(rqstp->rq_task);
390 rqstp->rq_xprt = xprt;
391 pool->sp_stats.threads_woken++;
393 dprintk("svc: transport %p put into queue\n", xprt);
394 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
395 pool->sp_stats.sockets_queued++;
399 spin_unlock_bh(&pool->sp_lock);
404 * Queue up a transport with data pending. If there are idle nfsd
405 * processes, wake 'em up.
408 void svc_xprt_enqueue(struct svc_xprt *xprt)
410 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
412 svc_xprt_do_enqueue(xprt);
414 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
417 * Dequeue the first transport. Must be called with the pool->sp_lock held.
419 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
421 struct svc_xprt *xprt;
423 if (list_empty(&pool->sp_sockets))
426 xprt = list_entry(pool->sp_sockets.next,
427 struct svc_xprt, xpt_ready);
428 list_del_init(&xprt->xpt_ready);
430 dprintk("svc: transport %p dequeued, inuse=%d\n",
431 xprt, atomic_read(&xprt->xpt_ref.refcount));
437 * svc_reserve - change the space reserved for the reply to a request.
438 * @rqstp: The request in question
439 * @space: new max space to reserve
441 * Each request reserves some space on the output queue of the transport
442 * to make sure the reply fits. This function reduces that reserved
443 * space to be the amount of space used already, plus @space.
446 void svc_reserve(struct svc_rqst *rqstp, int space)
448 space += rqstp->rq_res.head[0].iov_len;
450 if (space < rqstp->rq_reserved) {
451 struct svc_xprt *xprt = rqstp->rq_xprt;
452 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
453 rqstp->rq_reserved = space;
455 if (xprt->xpt_ops->xpo_adjust_wspace)
456 xprt->xpt_ops->xpo_adjust_wspace(xprt);
457 svc_xprt_enqueue(xprt);
460 EXPORT_SYMBOL_GPL(svc_reserve);
462 static void svc_xprt_release(struct svc_rqst *rqstp)
464 struct svc_xprt *xprt = rqstp->rq_xprt;
466 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
468 kfree(rqstp->rq_deferred);
469 rqstp->rq_deferred = NULL;
471 svc_free_res_pages(rqstp);
472 rqstp->rq_res.page_len = 0;
473 rqstp->rq_res.page_base = 0;
475 /* Reset response buffer and release
477 * But first, check that enough space was reserved
478 * for the reply, otherwise we have a bug!
480 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
481 printk(KERN_ERR "RPC request reserved %d but used %d\n",
485 rqstp->rq_res.head[0].iov_len = 0;
486 svc_reserve(rqstp, 0);
487 rqstp->rq_xprt = NULL;
493 * External function to wake up a server waiting for data
494 * This really only makes sense for services like lockd
495 * which have exactly one thread anyway.
497 void svc_wake_up(struct svc_serv *serv)
499 struct svc_rqst *rqstp;
501 struct svc_pool *pool;
503 for (i = 0; i < serv->sv_nrpools; i++) {
504 pool = &serv->sv_pools[i];
506 spin_lock_bh(&pool->sp_lock);
507 if (!list_empty(&pool->sp_threads)) {
508 rqstp = list_entry(pool->sp_threads.next,
511 dprintk("svc: daemon %p woken up.\n", rqstp);
513 svc_thread_dequeue(pool, rqstp);
514 rqstp->rq_xprt = NULL;
516 wake_up_process(rqstp->rq_task);
518 pool->sp_task_pending = 1;
519 spin_unlock_bh(&pool->sp_lock);
522 EXPORT_SYMBOL_GPL(svc_wake_up);
524 int svc_port_is_privileged(struct sockaddr *sin)
526 switch (sin->sa_family) {
528 return ntohs(((struct sockaddr_in *)sin)->sin_port)
531 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
539 * Make sure that we don't have too many active connections. If we have,
540 * something must be dropped. It's not clear what will happen if we allow
541 * "too many" connections, but when dealing with network-facing software,
542 * we have to code defensively. Here we do that by imposing hard limits.
544 * There's no point in trying to do random drop here for DoS
545 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
546 * attacker can easily beat that.
548 * The only somewhat efficient mechanism would be if drop old
549 * connections from the same IP first. But right now we don't even
550 * record the client IP in svc_sock.
552 * single-threaded services that expect a lot of clients will probably
553 * need to set sv_maxconn to override the default value which is based
554 * on the number of threads
556 static void svc_check_conn_limits(struct svc_serv *serv)
558 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
559 (serv->sv_nrthreads+3) * 20;
561 if (serv->sv_tmpcnt > limit) {
562 struct svc_xprt *xprt = NULL;
563 spin_lock_bh(&serv->sv_lock);
564 if (!list_empty(&serv->sv_tempsocks)) {
565 /* Try to help the admin */
566 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
567 serv->sv_name, serv->sv_maxconn ?
568 "max number of connections" :
569 "number of threads");
571 * Always select the oldest connection. It's not fair,
574 xprt = list_entry(serv->sv_tempsocks.prev,
577 set_bit(XPT_CLOSE, &xprt->xpt_flags);
580 spin_unlock_bh(&serv->sv_lock);
583 svc_xprt_enqueue(xprt);
589 static int svc_alloc_arg(struct svc_rqst *rqstp)
591 struct svc_serv *serv = rqstp->rq_server;
596 /* now allocate needed pages. If we get a failure, sleep briefly */
597 pages = (serv->sv_max_mesg + PAGE_SIZE) / PAGE_SIZE;
598 WARN_ON_ONCE(pages >= RPCSVC_MAXPAGES);
599 if (pages >= RPCSVC_MAXPAGES)
600 /* use as many pages as possible */
601 pages = RPCSVC_MAXPAGES - 1;
602 for (i = 0; i < pages ; i++)
603 while (rqstp->rq_pages[i] == NULL) {
604 struct page *p = alloc_page(GFP_KERNEL);
606 set_current_state(TASK_INTERRUPTIBLE);
607 if (signalled() || kthread_should_stop()) {
608 set_current_state(TASK_RUNNING);
611 schedule_timeout(msecs_to_jiffies(500));
613 rqstp->rq_pages[i] = p;
615 rqstp->rq_page_end = &rqstp->rq_pages[i];
616 rqstp->rq_pages[i++] = NULL; /* this might be seen in nfs_read_actor */
618 /* Make arg->head point to first page and arg->pages point to rest */
619 arg = &rqstp->rq_arg;
620 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
621 arg->head[0].iov_len = PAGE_SIZE;
622 arg->pages = rqstp->rq_pages + 1;
624 /* save at least one page for response */
625 arg->page_len = (pages-2)*PAGE_SIZE;
626 arg->len = (pages-1)*PAGE_SIZE;
627 arg->tail[0].iov_len = 0;
631 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
633 struct svc_xprt *xprt;
634 struct svc_pool *pool = rqstp->rq_pool;
637 /* Normally we will wait up to 5 seconds for any required
638 * cache information to be provided.
640 rqstp->rq_chandle.thread_wait = 5*HZ;
642 spin_lock_bh(&pool->sp_lock);
643 xprt = svc_xprt_dequeue(pool);
645 rqstp->rq_xprt = xprt;
648 /* As there is a shortage of threads and this request
649 * had to be queued, don't allow the thread to wait so
650 * long for cache updates.
652 rqstp->rq_chandle.thread_wait = 1*HZ;
653 pool->sp_task_pending = 0;
655 if (pool->sp_task_pending) {
656 pool->sp_task_pending = 0;
657 xprt = ERR_PTR(-EAGAIN);
661 * We have to be able to interrupt this wait
662 * to bring down the daemons ...
664 set_current_state(TASK_INTERRUPTIBLE);
666 /* No data pending. Go to sleep */
667 svc_thread_enqueue(pool, rqstp);
670 * checking kthread_should_stop() here allows us to avoid
671 * locking and signalling when stopping kthreads that call
672 * svc_recv. If the thread has already been woken up, then
673 * we can exit here without sleeping. If not, then it
674 * it'll be woken up quickly during the schedule_timeout
676 if (kthread_should_stop()) {
677 set_current_state(TASK_RUNNING);
678 xprt = ERR_PTR(-EINTR);
682 spin_unlock_bh(&pool->sp_lock);
684 time_left = schedule_timeout(timeout);
685 __set_current_state(TASK_RUNNING);
689 xprt = rqstp->rq_xprt;
693 spin_lock_bh(&pool->sp_lock);
695 pool->sp_stats.threads_timedout++;
697 xprt = rqstp->rq_xprt;
699 svc_thread_dequeue(pool, rqstp);
700 spin_unlock_bh(&pool->sp_lock);
701 dprintk("svc: server %p, no data yet\n", rqstp);
702 if (signalled() || kthread_should_stop())
703 return ERR_PTR(-EINTR);
705 return ERR_PTR(-EAGAIN);
709 spin_unlock_bh(&pool->sp_lock);
713 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
715 spin_lock_bh(&serv->sv_lock);
716 set_bit(XPT_TEMP, &newxpt->xpt_flags);
717 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
719 if (serv->sv_temptimer.function == NULL) {
720 /* setup timer to age temp transports */
721 setup_timer(&serv->sv_temptimer, svc_age_temp_xprts,
722 (unsigned long)serv);
723 mod_timer(&serv->sv_temptimer,
724 jiffies + svc_conn_age_period * HZ);
726 spin_unlock_bh(&serv->sv_lock);
727 svc_xprt_received(newxpt);
730 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
732 struct svc_serv *serv = rqstp->rq_server;
735 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
736 dprintk("svc_recv: found XPT_CLOSE\n");
737 svc_delete_xprt(xprt);
738 /* Leave XPT_BUSY set on the dead xprt: */
741 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
742 struct svc_xprt *newxpt;
744 * We know this module_get will succeed because the
745 * listener holds a reference too
747 __module_get(xprt->xpt_class->xcl_owner);
748 svc_check_conn_limits(xprt->xpt_server);
749 newxpt = xprt->xpt_ops->xpo_accept(xprt);
751 svc_add_new_temp_xprt(serv, newxpt);
753 module_put(xprt->xpt_class->xcl_owner);
755 /* XPT_DATA|XPT_DEFERRED case: */
756 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
757 rqstp, rqstp->rq_pool->sp_id, xprt,
758 atomic_read(&xprt->xpt_ref.refcount));
759 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
760 if (rqstp->rq_deferred)
761 len = svc_deferred_recv(rqstp);
763 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
764 dprintk("svc: got len=%d\n", len);
765 rqstp->rq_reserved = serv->sv_max_mesg;
766 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
768 /* clear XPT_BUSY: */
769 svc_xprt_received(xprt);
774 * Receive the next request on any transport. This code is carefully
775 * organised not to touch any cachelines in the shared svc_serv
776 * structure, only cachelines in the local svc_pool.
778 int svc_recv(struct svc_rqst *rqstp, long timeout)
780 struct svc_xprt *xprt = NULL;
781 struct svc_serv *serv = rqstp->rq_server;
784 dprintk("svc: server %p waiting for data (to = %ld)\n",
789 "svc_recv: service %p, transport not NULL!\n",
792 /* Make sure the task pointer is set! */
793 if (WARN_ON_ONCE(!rqstp->rq_task))
794 rqstp->rq_task = current_task;
796 err = svc_alloc_arg(rqstp);
802 if (signalled() || kthread_should_stop())
805 xprt = svc_get_next_xprt(rqstp, timeout);
807 return PTR_ERR(xprt);
809 len = svc_handle_xprt(rqstp, xprt);
811 /* No data, incomplete (TCP) read, or accept() */
815 clear_bit(XPT_OLD, &xprt->xpt_flags);
817 rqstp->rq_secure = xprt->xpt_ops->xpo_secure_port(rqstp);
818 rqstp->rq_chandle.defer = svc_defer;
821 serv->sv_stats->netcnt++;
824 rqstp->rq_res.len = 0;
825 svc_xprt_release(rqstp);
828 EXPORT_SYMBOL_GPL(svc_recv);
833 void svc_drop(struct svc_rqst *rqstp)
835 dprintk("svc: xprt %p dropped request\n", rqstp->rq_xprt);
836 svc_xprt_release(rqstp);
838 EXPORT_SYMBOL_GPL(svc_drop);
841 * Return reply to client.
843 int svc_send(struct svc_rqst *rqstp)
845 struct svc_xprt *xprt;
849 xprt = rqstp->rq_xprt;
853 /* release the receive skb before sending the reply */
854 rqstp->rq_xprt->xpt_ops->xpo_release_rqst(rqstp);
856 /* calculate over-all length */
858 xb->len = xb->head[0].iov_len +
862 /* Grab mutex to serialize outgoing data. */
863 mutex_lock(&xprt->xpt_mutex);
864 if (test_bit(XPT_DEAD, &xprt->xpt_flags)
865 || test_bit(XPT_CLOSE, &xprt->xpt_flags))
868 len = xprt->xpt_ops->xpo_sendto(rqstp);
869 mutex_unlock(&xprt->xpt_mutex);
870 rpc_wake_up(&xprt->xpt_bc_pending);
871 svc_xprt_release(rqstp);
873 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
879 * Timer function to close old temporary transports, using
880 * a mark-and-sweep algorithm.
882 static void svc_age_temp_xprts(unsigned long closure)
884 struct svc_serv *serv = (struct svc_serv *)closure;
885 struct svc_xprt *xprt;
886 struct list_head *le, *next;
888 dprintk("svc_age_temp_xprts\n");
890 if (!spin_trylock_bh(&serv->sv_lock)) {
891 /* busy, try again 1 sec later */
892 dprintk("svc_age_temp_xprts: busy\n");
893 mod_timer(&serv->sv_temptimer, jiffies + HZ);
897 list_for_each_safe(le, next, &serv->sv_tempsocks) {
898 xprt = list_entry(le, struct svc_xprt, xpt_list);
900 /* First time through, just mark it OLD. Second time
901 * through, close it. */
902 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
904 if (atomic_read(&xprt->xpt_ref.refcount) > 1 ||
905 test_bit(XPT_BUSY, &xprt->xpt_flags))
908 set_bit(XPT_CLOSE, &xprt->xpt_flags);
909 set_bit(XPT_DETACHED, &xprt->xpt_flags);
910 dprintk("queuing xprt %p for closing\n", xprt);
912 /* a thread will dequeue and close it soon */
913 svc_xprt_enqueue(xprt);
915 spin_unlock_bh(&serv->sv_lock);
917 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
920 static void call_xpt_users(struct svc_xprt *xprt)
922 struct svc_xpt_user *u;
924 spin_lock(&xprt->xpt_lock);
925 while (!list_empty(&xprt->xpt_users)) {
926 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
930 spin_unlock(&xprt->xpt_lock);
934 * Remove a dead transport
936 static void svc_delete_xprt(struct svc_xprt *xprt)
938 struct svc_serv *serv = xprt->xpt_server;
939 struct svc_deferred_req *dr;
941 /* Only do this once */
942 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
945 dprintk("svc: svc_delete_xprt(%p)\n", xprt);
946 xprt->xpt_ops->xpo_detach(xprt);
948 spin_lock_bh(&serv->sv_lock);
949 if (!test_and_set_bit(XPT_DETACHED, &xprt->xpt_flags))
950 list_del_init(&xprt->xpt_list);
951 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
952 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
954 spin_unlock_bh(&serv->sv_lock);
956 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
959 call_xpt_users(xprt);
963 void svc_close_xprt(struct svc_xprt *xprt)
965 set_bit(XPT_CLOSE, &xprt->xpt_flags);
966 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
967 /* someone else will have to effect the close */
970 * We expect svc_close_xprt() to work even when no threads are
971 * running (e.g., while configuring the server before starting
972 * any threads), so if the transport isn't busy, we delete
975 svc_delete_xprt(xprt);
977 EXPORT_SYMBOL_GPL(svc_close_xprt);
979 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
981 struct svc_xprt *xprt;
984 spin_lock(&serv->sv_lock);
985 list_for_each_entry(xprt, xprt_list, xpt_list) {
986 if (xprt->xpt_net != net)
989 set_bit(XPT_CLOSE, &xprt->xpt_flags);
990 svc_xprt_enqueue(xprt);
992 spin_unlock(&serv->sv_lock);
996 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
998 struct svc_pool *pool;
999 struct svc_xprt *xprt;
1000 struct svc_xprt *tmp;
1003 for (i = 0; i < serv->sv_nrpools; i++) {
1004 pool = &serv->sv_pools[i];
1006 spin_lock_bh(&pool->sp_lock);
1007 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1008 if (xprt->xpt_net != net)
1010 list_del_init(&xprt->xpt_ready);
1011 spin_unlock_bh(&pool->sp_lock);
1014 spin_unlock_bh(&pool->sp_lock);
1019 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1021 struct svc_xprt *xprt;
1023 while ((xprt = svc_dequeue_net(serv, net))) {
1024 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1025 svc_delete_xprt(xprt);
1030 * Server threads may still be running (especially in the case where the
1031 * service is still running in other network namespaces).
1033 * So we shut down sockets the same way we would on a running server, by
1034 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1035 * the close. In the case there are no such other threads,
1036 * threads running, svc_clean_up_xprts() does a simple version of a
1037 * server's main event loop, and in the case where there are other
1038 * threads, we may need to wait a little while and then check again to
1039 * see if they're done.
1041 void svc_close_net(struct svc_serv *serv, struct net *net)
1045 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1046 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1048 svc_clean_up_xprts(serv, net);
1054 * Handle defer and revisit of requests
1057 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1059 struct svc_deferred_req *dr =
1060 container_of(dreq, struct svc_deferred_req, handle);
1061 struct svc_xprt *xprt = dr->xprt;
1063 spin_lock(&xprt->xpt_lock);
1064 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1065 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1066 spin_unlock(&xprt->xpt_lock);
1067 dprintk("revisit canceled\n");
1072 dprintk("revisit queued\n");
1074 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1075 spin_unlock(&xprt->xpt_lock);
1076 svc_xprt_enqueue(xprt);
1081 * Save the request off for later processing. The request buffer looks
1084 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1086 * This code can only handle requests that consist of an xprt-header
1089 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1091 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1092 struct svc_deferred_req *dr;
1094 if (rqstp->rq_arg.page_len || !rqstp->rq_usedeferral)
1095 return NULL; /* if more than a page, give up FIXME */
1096 if (rqstp->rq_deferred) {
1097 dr = rqstp->rq_deferred;
1098 rqstp->rq_deferred = NULL;
1102 /* FIXME maybe discard if size too large */
1103 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1104 dr = kmalloc(size, GFP_KERNEL);
1108 dr->handle.owner = rqstp->rq_server;
1109 dr->prot = rqstp->rq_prot;
1110 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1111 dr->addrlen = rqstp->rq_addrlen;
1112 dr->daddr = rqstp->rq_daddr;
1113 dr->argslen = rqstp->rq_arg.len >> 2;
1114 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1116 /* back up head to the start of the buffer and copy */
1117 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1118 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1121 svc_xprt_get(rqstp->rq_xprt);
1122 dr->xprt = rqstp->rq_xprt;
1123 rqstp->rq_dropme = true;
1125 dr->handle.revisit = svc_revisit;
1130 * recv data from a deferred request into an active one
1132 static int svc_deferred_recv(struct svc_rqst *rqstp)
1134 struct svc_deferred_req *dr = rqstp->rq_deferred;
1136 /* setup iov_base past transport header */
1137 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1138 /* The iov_len does not include the transport header bytes */
1139 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1140 rqstp->rq_arg.page_len = 0;
1141 /* The rq_arg.len includes the transport header bytes */
1142 rqstp->rq_arg.len = dr->argslen<<2;
1143 rqstp->rq_prot = dr->prot;
1144 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1145 rqstp->rq_addrlen = dr->addrlen;
1146 /* Save off transport header len in case we get deferred again */
1147 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1148 rqstp->rq_daddr = dr->daddr;
1149 rqstp->rq_respages = rqstp->rq_pages;
1150 return (dr->argslen<<2) - dr->xprt_hlen;
1154 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1156 struct svc_deferred_req *dr = NULL;
1158 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1160 spin_lock(&xprt->xpt_lock);
1161 if (!list_empty(&xprt->xpt_deferred)) {
1162 dr = list_entry(xprt->xpt_deferred.next,
1163 struct svc_deferred_req,
1165 list_del_init(&dr->handle.recent);
1167 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1168 spin_unlock(&xprt->xpt_lock);
1173 * svc_find_xprt - find an RPC transport instance
1174 * @serv: pointer to svc_serv to search
1175 * @xcl_name: C string containing transport's class name
1176 * @net: owner net pointer
1177 * @af: Address family of transport's local address
1178 * @port: transport's IP port number
1180 * Return the transport instance pointer for the endpoint accepting
1181 * connections/peer traffic from the specified transport class,
1182 * address family and port.
1184 * Specifying 0 for the address family or port is effectively a
1185 * wild-card, and will result in matching the first transport in the
1186 * service's list that has a matching class name.
1188 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1189 struct net *net, const sa_family_t af,
1190 const unsigned short port)
1192 struct svc_xprt *xprt;
1193 struct svc_xprt *found = NULL;
1195 /* Sanity check the args */
1196 if (serv == NULL || xcl_name == NULL)
1199 spin_lock_bh(&serv->sv_lock);
1200 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1201 if (xprt->xpt_net != net)
1203 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1205 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1207 if (port != 0 && port != svc_xprt_local_port(xprt))
1213 spin_unlock_bh(&serv->sv_lock);
1216 EXPORT_SYMBOL_GPL(svc_find_xprt);
1218 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1219 char *pos, int remaining)
1223 len = snprintf(pos, remaining, "%s %u\n",
1224 xprt->xpt_class->xcl_name,
1225 svc_xprt_local_port(xprt));
1226 if (len >= remaining)
1227 return -ENAMETOOLONG;
1232 * svc_xprt_names - format a buffer with a list of transport names
1233 * @serv: pointer to an RPC service
1234 * @buf: pointer to a buffer to be filled in
1235 * @buflen: length of buffer to be filled in
1237 * Fills in @buf with a string containing a list of transport names,
1238 * each name terminated with '\n'.
1240 * Returns positive length of the filled-in string on success; otherwise
1241 * a negative errno value is returned if an error occurs.
1243 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1245 struct svc_xprt *xprt;
1249 /* Sanity check args */
1253 spin_lock_bh(&serv->sv_lock);
1257 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1258 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1270 spin_unlock_bh(&serv->sv_lock);
1273 EXPORT_SYMBOL_GPL(svc_xprt_names);
1276 /*----------------------------------------------------------------------------*/
1278 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1280 unsigned int pidx = (unsigned int)*pos;
1281 struct svc_serv *serv = m->private;
1283 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1286 return SEQ_START_TOKEN;
1287 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1290 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1292 struct svc_pool *pool = p;
1293 struct svc_serv *serv = m->private;
1295 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1297 if (p == SEQ_START_TOKEN) {
1298 pool = &serv->sv_pools[0];
1300 unsigned int pidx = (pool - &serv->sv_pools[0]);
1301 if (pidx < serv->sv_nrpools-1)
1302 pool = &serv->sv_pools[pidx+1];
1310 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1314 static int svc_pool_stats_show(struct seq_file *m, void *p)
1316 struct svc_pool *pool = p;
1318 if (p == SEQ_START_TOKEN) {
1319 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1323 seq_printf(m, "%u %lu %lu %lu %lu\n",
1325 pool->sp_stats.packets,
1326 pool->sp_stats.sockets_queued,
1327 pool->sp_stats.threads_woken,
1328 pool->sp_stats.threads_timedout);
1333 static const struct seq_operations svc_pool_stats_seq_ops = {
1334 .start = svc_pool_stats_start,
1335 .next = svc_pool_stats_next,
1336 .stop = svc_pool_stats_stop,
1337 .show = svc_pool_stats_show,
1340 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1344 err = seq_open(file, &svc_pool_stats_seq_ops);
1346 ((struct seq_file *) file->private_data)->private = serv;
1349 EXPORT_SYMBOL(svc_pool_stats_open);
1351 /*----------------------------------------------------------------------------*/