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);
143 * svc_xprt_deferred_close - Close a transport
144 * @xprt: transport instance
146 * Used in contexts that need to defer the work of shutting down
147 * the transport to an nfsd thread.
149 void svc_xprt_deferred_close(struct svc_xprt *xprt)
151 if (!test_and_set_bit(XPT_CLOSE, &xprt->xpt_flags))
152 svc_xprt_enqueue(xprt);
154 EXPORT_SYMBOL_GPL(svc_xprt_deferred_close);
156 static void svc_xprt_free(struct kref *kref)
158 struct svc_xprt *xprt =
159 container_of(kref, struct svc_xprt, xpt_ref);
160 struct module *owner = xprt->xpt_class->xcl_owner;
161 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
162 svcauth_unix_info_release(xprt);
163 put_cred(xprt->xpt_cred);
164 put_net(xprt->xpt_net);
165 /* See comment on corresponding get in xs_setup_bc_tcp(): */
166 if (xprt->xpt_bc_xprt)
167 xprt_put(xprt->xpt_bc_xprt);
168 if (xprt->xpt_bc_xps)
169 xprt_switch_put(xprt->xpt_bc_xps);
170 trace_svc_xprt_free(xprt);
171 xprt->xpt_ops->xpo_free(xprt);
175 void svc_xprt_put(struct svc_xprt *xprt)
177 kref_put(&xprt->xpt_ref, svc_xprt_free);
179 EXPORT_SYMBOL_GPL(svc_xprt_put);
182 * Called by transport drivers to initialize the transport independent
183 * portion of the transport instance.
185 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
186 struct svc_xprt *xprt, struct svc_serv *serv)
188 memset(xprt, 0, sizeof(*xprt));
189 xprt->xpt_class = xcl;
190 xprt->xpt_ops = xcl->xcl_ops;
191 kref_init(&xprt->xpt_ref);
192 xprt->xpt_server = serv;
193 INIT_LIST_HEAD(&xprt->xpt_list);
194 INIT_LIST_HEAD(&xprt->xpt_ready);
195 INIT_LIST_HEAD(&xprt->xpt_deferred);
196 INIT_LIST_HEAD(&xprt->xpt_users);
197 mutex_init(&xprt->xpt_mutex);
198 spin_lock_init(&xprt->xpt_lock);
199 set_bit(XPT_BUSY, &xprt->xpt_flags);
200 xprt->xpt_net = get_net(net);
201 strcpy(xprt->xpt_remotebuf, "uninitialized");
203 EXPORT_SYMBOL_GPL(svc_xprt_init);
205 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
206 struct svc_serv *serv,
209 const unsigned short port,
212 struct sockaddr_in sin = {
213 .sin_family = AF_INET,
214 .sin_addr.s_addr = htonl(INADDR_ANY),
215 .sin_port = htons(port),
217 #if IS_ENABLED(CONFIG_IPV6)
218 struct sockaddr_in6 sin6 = {
219 .sin6_family = AF_INET6,
220 .sin6_addr = IN6ADDR_ANY_INIT,
221 .sin6_port = htons(port),
224 struct svc_xprt *xprt;
225 struct sockaddr *sap;
230 sap = (struct sockaddr *)&sin;
233 #if IS_ENABLED(CONFIG_IPV6)
235 sap = (struct sockaddr *)&sin6;
240 return ERR_PTR(-EAFNOSUPPORT);
243 xprt = xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
245 trace_svc_xprt_create_err(serv->sv_program->pg_name,
246 xcl->xcl_name, sap, xprt);
251 * svc_xprt_received - start next receiver thread
252 * @xprt: controlling transport
254 * The caller must hold the XPT_BUSY bit and must
255 * not thereafter touch transport data.
257 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
258 * insufficient) data.
260 void svc_xprt_received(struct svc_xprt *xprt)
262 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
263 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
267 trace_svc_xprt_received(xprt);
269 /* As soon as we clear busy, the xprt could be closed and
270 * 'put', so we need a reference to call svc_enqueue_xprt with:
273 smp_mb__before_atomic();
274 clear_bit(XPT_BUSY, &xprt->xpt_flags);
275 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
278 EXPORT_SYMBOL_GPL(svc_xprt_received);
280 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
282 clear_bit(XPT_TEMP, &new->xpt_flags);
283 spin_lock_bh(&serv->sv_lock);
284 list_add(&new->xpt_list, &serv->sv_permsocks);
285 spin_unlock_bh(&serv->sv_lock);
286 svc_xprt_received(new);
289 static int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
290 struct net *net, const int family,
291 const unsigned short port, int flags,
292 const struct cred *cred)
294 struct svc_xprt_class *xcl;
296 spin_lock(&svc_xprt_class_lock);
297 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
298 struct svc_xprt *newxprt;
299 unsigned short newport;
301 if (strcmp(xprt_name, xcl->xcl_name))
304 if (!try_module_get(xcl->xcl_owner))
307 spin_unlock(&svc_xprt_class_lock);
308 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
309 if (IS_ERR(newxprt)) {
310 module_put(xcl->xcl_owner);
311 return PTR_ERR(newxprt);
313 newxprt->xpt_cred = get_cred(cred);
314 svc_add_new_perm_xprt(serv, newxprt);
315 newport = svc_xprt_local_port(newxprt);
319 spin_unlock(&svc_xprt_class_lock);
320 /* This errno is exposed to user space. Provide a reasonable
321 * perror msg for a bad transport. */
322 return -EPROTONOSUPPORT;
325 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
326 struct net *net, const int family,
327 const unsigned short port, int flags,
328 const struct cred *cred)
332 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags, cred);
333 if (err == -EPROTONOSUPPORT) {
334 request_module("svc%s", xprt_name);
335 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags, cred);
339 EXPORT_SYMBOL_GPL(svc_create_xprt);
342 * Copy the local and remote xprt addresses to the rqstp structure
344 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
346 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
347 rqstp->rq_addrlen = xprt->xpt_remotelen;
350 * Destination address in request is needed for binding the
351 * source address in RPC replies/callbacks later.
353 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
354 rqstp->rq_daddrlen = xprt->xpt_locallen;
356 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
359 * svc_print_addr - Format rq_addr field for printing
360 * @rqstp: svc_rqst struct containing address to print
361 * @buf: target buffer for formatted address
362 * @len: length of target buffer
365 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
367 return __svc_print_addr(svc_addr(rqstp), buf, len);
369 EXPORT_SYMBOL_GPL(svc_print_addr);
371 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
373 unsigned int limit = svc_rpc_per_connection_limit;
374 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
376 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
379 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
381 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
382 if (!svc_xprt_slots_in_range(xprt))
384 atomic_inc(&xprt->xpt_nr_rqsts);
385 set_bit(RQ_DATA, &rqstp->rq_flags);
390 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
392 struct svc_xprt *xprt = rqstp->rq_xprt;
393 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
394 atomic_dec(&xprt->xpt_nr_rqsts);
395 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
396 svc_xprt_enqueue(xprt);
400 static bool svc_xprt_ready(struct svc_xprt *xprt)
402 unsigned long xpt_flags;
405 * If another cpu has recently updated xpt_flags,
406 * sk_sock->flags, xpt_reserved, or xpt_nr_rqsts, we need to
407 * know about it; otherwise it's possible that both that cpu and
408 * this one could call svc_xprt_enqueue() without either
409 * svc_xprt_enqueue() recognizing that the conditions below
410 * are satisfied, and we could stall indefinitely:
413 xpt_flags = READ_ONCE(xprt->xpt_flags);
415 if (xpt_flags & (BIT(XPT_CONN) | BIT(XPT_CLOSE)))
417 if (xpt_flags & (BIT(XPT_DATA) | BIT(XPT_DEFERRED))) {
418 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
419 svc_xprt_slots_in_range(xprt))
421 trace_svc_xprt_no_write_space(xprt);
427 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
429 struct svc_pool *pool;
430 struct svc_rqst *rqstp = NULL;
433 if (!svc_xprt_ready(xprt))
436 /* Mark transport as busy. It will remain in this state until
437 * the provider calls svc_xprt_received. We update XPT_BUSY
438 * atomically because it also guards against trying to enqueue
439 * the transport twice.
441 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
445 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
447 atomic_long_inc(&pool->sp_stats.packets);
449 spin_lock_bh(&pool->sp_lock);
450 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
451 pool->sp_stats.sockets_queued++;
452 spin_unlock_bh(&pool->sp_lock);
454 /* find a thread for this xprt */
456 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
457 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
459 atomic_long_inc(&pool->sp_stats.threads_woken);
460 rqstp->rq_qtime = ktime_get();
461 wake_up_process(rqstp->rq_task);
464 set_bit(SP_CONGESTED, &pool->sp_flags);
469 trace_svc_xprt_do_enqueue(xprt, rqstp);
471 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
474 * Queue up a transport with data pending. If there are idle nfsd
475 * processes, wake 'em up.
478 void svc_xprt_enqueue(struct svc_xprt *xprt)
480 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
482 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
484 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
487 * Dequeue the first transport, if there is one.
489 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
491 struct svc_xprt *xprt = NULL;
493 if (list_empty(&pool->sp_sockets))
496 spin_lock_bh(&pool->sp_lock);
497 if (likely(!list_empty(&pool->sp_sockets))) {
498 xprt = list_first_entry(&pool->sp_sockets,
499 struct svc_xprt, xpt_ready);
500 list_del_init(&xprt->xpt_ready);
503 spin_unlock_bh(&pool->sp_lock);
509 * svc_reserve - change the space reserved for the reply to a request.
510 * @rqstp: The request in question
511 * @space: new max space to reserve
513 * Each request reserves some space on the output queue of the transport
514 * to make sure the reply fits. This function reduces that reserved
515 * space to be the amount of space used already, plus @space.
518 void svc_reserve(struct svc_rqst *rqstp, int space)
520 struct svc_xprt *xprt = rqstp->rq_xprt;
522 space += rqstp->rq_res.head[0].iov_len;
524 if (xprt && space < rqstp->rq_reserved) {
525 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
526 rqstp->rq_reserved = space;
527 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
528 svc_xprt_enqueue(xprt);
531 EXPORT_SYMBOL_GPL(svc_reserve);
533 static void svc_xprt_release(struct svc_rqst *rqstp)
535 struct svc_xprt *xprt = rqstp->rq_xprt;
537 xprt->xpt_ops->xpo_release_rqst(rqstp);
539 kfree(rqstp->rq_deferred);
540 rqstp->rq_deferred = NULL;
542 pagevec_release(&rqstp->rq_pvec);
543 svc_free_res_pages(rqstp);
544 rqstp->rq_res.page_len = 0;
545 rqstp->rq_res.page_base = 0;
547 /* Reset response buffer and release
549 * But first, check that enough space was reserved
550 * for the reply, otherwise we have a bug!
552 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
553 printk(KERN_ERR "RPC request reserved %d but used %d\n",
557 rqstp->rq_res.head[0].iov_len = 0;
558 svc_reserve(rqstp, 0);
559 svc_xprt_release_slot(rqstp);
560 rqstp->rq_xprt = NULL;
565 * Some svc_serv's will have occasional work to do, even when a xprt is not
566 * waiting to be serviced. This function is there to "kick" a task in one of
567 * those services so that it can wake up and do that work. Note that we only
568 * bother with pool 0 as we don't need to wake up more than one thread for
571 void svc_wake_up(struct svc_serv *serv)
573 struct svc_rqst *rqstp;
574 struct svc_pool *pool;
576 pool = &serv->sv_pools[0];
579 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
580 /* skip any that aren't queued */
581 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
584 wake_up_process(rqstp->rq_task);
585 trace_svc_wake_up(rqstp->rq_task->pid);
590 /* No free entries available */
591 set_bit(SP_TASK_PENDING, &pool->sp_flags);
593 trace_svc_wake_up(0);
595 EXPORT_SYMBOL_GPL(svc_wake_up);
597 int svc_port_is_privileged(struct sockaddr *sin)
599 switch (sin->sa_family) {
601 return ntohs(((struct sockaddr_in *)sin)->sin_port)
604 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
612 * Make sure that we don't have too many active connections. If we have,
613 * something must be dropped. It's not clear what will happen if we allow
614 * "too many" connections, but when dealing with network-facing software,
615 * we have to code defensively. Here we do that by imposing hard limits.
617 * There's no point in trying to do random drop here for DoS
618 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
619 * attacker can easily beat that.
621 * The only somewhat efficient mechanism would be if drop old
622 * connections from the same IP first. But right now we don't even
623 * record the client IP in svc_sock.
625 * single-threaded services that expect a lot of clients will probably
626 * need to set sv_maxconn to override the default value which is based
627 * on the number of threads
629 static void svc_check_conn_limits(struct svc_serv *serv)
631 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
632 (serv->sv_nrthreads+3) * 20;
634 if (serv->sv_tmpcnt > limit) {
635 struct svc_xprt *xprt = NULL;
636 spin_lock_bh(&serv->sv_lock);
637 if (!list_empty(&serv->sv_tempsocks)) {
638 /* Try to help the admin */
639 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
640 serv->sv_name, serv->sv_maxconn ?
641 "max number of connections" :
642 "number of threads");
644 * Always select the oldest connection. It's not fair,
647 xprt = list_entry(serv->sv_tempsocks.prev,
650 set_bit(XPT_CLOSE, &xprt->xpt_flags);
653 spin_unlock_bh(&serv->sv_lock);
656 svc_xprt_enqueue(xprt);
662 static int svc_alloc_arg(struct svc_rqst *rqstp)
664 struct svc_serv *serv = rqstp->rq_server;
665 struct xdr_buf *arg = &rqstp->rq_arg;
666 unsigned long pages, filled, ret;
668 pagevec_init(&rqstp->rq_pvec);
670 pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
671 if (pages > RPCSVC_MAXPAGES) {
672 pr_warn_once("svc: warning: pages=%lu > RPCSVC_MAXPAGES=%lu\n",
673 pages, RPCSVC_MAXPAGES);
674 /* use as many pages as possible */
675 pages = RPCSVC_MAXPAGES;
678 for (filled = 0; filled < pages; filled = ret) {
679 ret = alloc_pages_bulk_array(GFP_KERNEL, pages,
682 /* Made progress, don't sleep yet */
685 set_current_state(TASK_INTERRUPTIBLE);
686 if (signalled() || kthread_should_stop()) {
687 set_current_state(TASK_RUNNING);
690 schedule_timeout(msecs_to_jiffies(500));
692 rqstp->rq_page_end = &rqstp->rq_pages[pages];
693 rqstp->rq_pages[pages] = NULL; /* this might be seen in nfsd_splice_actor() */
695 /* Make arg->head point to first page and arg->pages point to rest */
696 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
697 arg->head[0].iov_len = PAGE_SIZE;
698 arg->pages = rqstp->rq_pages + 1;
700 /* save at least one page for response */
701 arg->page_len = (pages-2)*PAGE_SIZE;
702 arg->len = (pages-1)*PAGE_SIZE;
703 arg->tail[0].iov_len = 0;
708 rqst_should_sleep(struct svc_rqst *rqstp)
710 struct svc_pool *pool = rqstp->rq_pool;
712 /* did someone call svc_wake_up? */
713 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
716 /* was a socket queued? */
717 if (!list_empty(&pool->sp_sockets))
720 /* are we shutting down? */
721 if (signalled() || kthread_should_stop())
724 /* are we freezing? */
725 if (freezing(current))
731 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
733 struct svc_pool *pool = rqstp->rq_pool;
736 /* rq_xprt should be clear on entry */
737 WARN_ON_ONCE(rqstp->rq_xprt);
739 rqstp->rq_xprt = svc_xprt_dequeue(pool);
744 * We have to be able to interrupt this wait
745 * to bring down the daemons ...
747 set_current_state(TASK_INTERRUPTIBLE);
748 smp_mb__before_atomic();
749 clear_bit(SP_CONGESTED, &pool->sp_flags);
750 clear_bit(RQ_BUSY, &rqstp->rq_flags);
751 smp_mb__after_atomic();
753 if (likely(rqst_should_sleep(rqstp)))
754 time_left = schedule_timeout(timeout);
756 __set_current_state(TASK_RUNNING);
760 set_bit(RQ_BUSY, &rqstp->rq_flags);
761 smp_mb__after_atomic();
762 rqstp->rq_xprt = svc_xprt_dequeue(pool);
767 atomic_long_inc(&pool->sp_stats.threads_timedout);
769 if (signalled() || kthread_should_stop())
770 return ERR_PTR(-EINTR);
771 return ERR_PTR(-EAGAIN);
773 /* Normally we will wait up to 5 seconds for any required
774 * cache information to be provided.
776 if (!test_bit(SP_CONGESTED, &pool->sp_flags))
777 rqstp->rq_chandle.thread_wait = 5*HZ;
779 rqstp->rq_chandle.thread_wait = 1*HZ;
780 trace_svc_xprt_dequeue(rqstp);
781 return rqstp->rq_xprt;
784 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
786 spin_lock_bh(&serv->sv_lock);
787 set_bit(XPT_TEMP, &newxpt->xpt_flags);
788 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
790 if (serv->sv_temptimer.function == NULL) {
791 /* setup timer to age temp transports */
792 serv->sv_temptimer.function = svc_age_temp_xprts;
793 mod_timer(&serv->sv_temptimer,
794 jiffies + svc_conn_age_period * HZ);
796 spin_unlock_bh(&serv->sv_lock);
797 svc_xprt_received(newxpt);
800 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
802 struct svc_serv *serv = rqstp->rq_server;
805 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
806 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
807 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
808 svc_delete_xprt(xprt);
809 /* Leave XPT_BUSY set on the dead xprt: */
812 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
813 struct svc_xprt *newxpt;
815 * We know this module_get will succeed because the
816 * listener holds a reference too
818 __module_get(xprt->xpt_class->xcl_owner);
819 svc_check_conn_limits(xprt->xpt_server);
820 newxpt = xprt->xpt_ops->xpo_accept(xprt);
822 newxpt->xpt_cred = get_cred(xprt->xpt_cred);
823 svc_add_new_temp_xprt(serv, newxpt);
824 trace_svc_xprt_accept(newxpt, serv->sv_name);
826 module_put(xprt->xpt_class->xcl_owner);
828 svc_xprt_received(xprt);
829 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
830 /* XPT_DATA|XPT_DEFERRED case: */
831 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
832 rqstp, rqstp->rq_pool->sp_id, xprt,
833 kref_read(&xprt->xpt_ref));
834 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
835 if (rqstp->rq_deferred)
836 len = svc_deferred_recv(rqstp);
838 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
839 rqstp->rq_stime = ktime_get();
840 rqstp->rq_reserved = serv->sv_max_mesg;
841 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
843 svc_xprt_received(xprt);
845 trace_svc_handle_xprt(xprt, len);
850 * Receive the next request on any transport. This code is carefully
851 * organised not to touch any cachelines in the shared svc_serv
852 * structure, only cachelines in the local svc_pool.
854 int svc_recv(struct svc_rqst *rqstp, long timeout)
856 struct svc_xprt *xprt = NULL;
857 struct svc_serv *serv = rqstp->rq_server;
860 err = svc_alloc_arg(rqstp);
867 if (signalled() || kthread_should_stop())
870 xprt = svc_get_next_xprt(rqstp, timeout);
876 len = svc_handle_xprt(rqstp, xprt);
878 /* No data, incomplete (TCP) read, or accept() */
882 trace_svc_xdr_recvfrom(&rqstp->rq_arg);
884 clear_bit(XPT_OLD, &xprt->xpt_flags);
886 xprt->xpt_ops->xpo_secure_port(rqstp);
887 rqstp->rq_chandle.defer = svc_defer;
888 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
891 serv->sv_stats->netcnt++;
894 rqstp->rq_res.len = 0;
895 svc_xprt_release(rqstp);
899 EXPORT_SYMBOL_GPL(svc_recv);
904 void svc_drop(struct svc_rqst *rqstp)
906 trace_svc_drop(rqstp);
907 svc_xprt_release(rqstp);
909 EXPORT_SYMBOL_GPL(svc_drop);
912 * Return reply to client.
914 int svc_send(struct svc_rqst *rqstp)
916 struct svc_xprt *xprt;
920 xprt = rqstp->rq_xprt;
924 /* calculate over-all length */
926 xb->len = xb->head[0].iov_len +
929 trace_svc_xdr_sendto(rqstp->rq_xid, xb);
930 trace_svc_stats_latency(rqstp);
932 len = xprt->xpt_ops->xpo_sendto(rqstp);
934 trace_svc_send(rqstp, len);
935 svc_xprt_release(rqstp);
937 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
944 * Timer function to close old temporary transports, using
945 * a mark-and-sweep algorithm.
947 static void svc_age_temp_xprts(struct timer_list *t)
949 struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
950 struct svc_xprt *xprt;
951 struct list_head *le, *next;
953 dprintk("svc_age_temp_xprts\n");
955 if (!spin_trylock_bh(&serv->sv_lock)) {
956 /* busy, try again 1 sec later */
957 dprintk("svc_age_temp_xprts: busy\n");
958 mod_timer(&serv->sv_temptimer, jiffies + HZ);
962 list_for_each_safe(le, next, &serv->sv_tempsocks) {
963 xprt = list_entry(le, struct svc_xprt, xpt_list);
965 /* First time through, just mark it OLD. Second time
966 * through, close it. */
967 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
969 if (kref_read(&xprt->xpt_ref) > 1 ||
970 test_bit(XPT_BUSY, &xprt->xpt_flags))
973 set_bit(XPT_CLOSE, &xprt->xpt_flags);
974 dprintk("queuing xprt %p for closing\n", xprt);
976 /* a thread will dequeue and close it soon */
977 svc_xprt_enqueue(xprt);
979 spin_unlock_bh(&serv->sv_lock);
981 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
984 /* Close temporary transports whose xpt_local matches server_addr immediately
985 * instead of waiting for them to be picked up by the timer.
987 * This is meant to be called from a notifier_block that runs when an ip
988 * address is deleted.
990 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
992 struct svc_xprt *xprt;
993 struct list_head *le, *next;
994 LIST_HEAD(to_be_closed);
996 spin_lock_bh(&serv->sv_lock);
997 list_for_each_safe(le, next, &serv->sv_tempsocks) {
998 xprt = list_entry(le, struct svc_xprt, xpt_list);
999 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
1000 &xprt->xpt_local)) {
1001 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
1002 list_move(le, &to_be_closed);
1005 spin_unlock_bh(&serv->sv_lock);
1007 while (!list_empty(&to_be_closed)) {
1008 le = to_be_closed.next;
1010 xprt = list_entry(le, struct svc_xprt, xpt_list);
1011 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1012 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
1013 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
1015 svc_xprt_enqueue(xprt);
1018 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1020 static void call_xpt_users(struct svc_xprt *xprt)
1022 struct svc_xpt_user *u;
1024 spin_lock(&xprt->xpt_lock);
1025 while (!list_empty(&xprt->xpt_users)) {
1026 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1027 list_del_init(&u->list);
1030 spin_unlock(&xprt->xpt_lock);
1034 * Remove a dead transport
1036 static void svc_delete_xprt(struct svc_xprt *xprt)
1038 struct svc_serv *serv = xprt->xpt_server;
1039 struct svc_deferred_req *dr;
1041 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1044 trace_svc_xprt_detach(xprt);
1045 xprt->xpt_ops->xpo_detach(xprt);
1046 if (xprt->xpt_bc_xprt)
1047 xprt->xpt_bc_xprt->ops->close(xprt->xpt_bc_xprt);
1049 spin_lock_bh(&serv->sv_lock);
1050 list_del_init(&xprt->xpt_list);
1051 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1052 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1054 spin_unlock_bh(&serv->sv_lock);
1056 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1059 call_xpt_users(xprt);
1063 void svc_close_xprt(struct svc_xprt *xprt)
1065 trace_svc_xprt_close(xprt);
1066 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1067 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1068 /* someone else will have to effect the close */
1071 * We expect svc_close_xprt() to work even when no threads are
1072 * running (e.g., while configuring the server before starting
1073 * any threads), so if the transport isn't busy, we delete
1076 svc_delete_xprt(xprt);
1078 EXPORT_SYMBOL_GPL(svc_close_xprt);
1080 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1082 struct svc_xprt *xprt;
1085 spin_lock_bh(&serv->sv_lock);
1086 list_for_each_entry(xprt, xprt_list, xpt_list) {
1087 if (xprt->xpt_net != net)
1090 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1091 svc_xprt_enqueue(xprt);
1093 spin_unlock_bh(&serv->sv_lock);
1097 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1099 struct svc_pool *pool;
1100 struct svc_xprt *xprt;
1101 struct svc_xprt *tmp;
1104 for (i = 0; i < serv->sv_nrpools; i++) {
1105 pool = &serv->sv_pools[i];
1107 spin_lock_bh(&pool->sp_lock);
1108 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1109 if (xprt->xpt_net != net)
1111 list_del_init(&xprt->xpt_ready);
1112 spin_unlock_bh(&pool->sp_lock);
1115 spin_unlock_bh(&pool->sp_lock);
1120 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1122 struct svc_xprt *xprt;
1124 while ((xprt = svc_dequeue_net(serv, net))) {
1125 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1126 svc_delete_xprt(xprt);
1131 * Server threads may still be running (especially in the case where the
1132 * service is still running in other network namespaces).
1134 * So we shut down sockets the same way we would on a running server, by
1135 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1136 * the close. In the case there are no such other threads,
1137 * threads running, svc_clean_up_xprts() does a simple version of a
1138 * server's main event loop, and in the case where there are other
1139 * threads, we may need to wait a little while and then check again to
1140 * see if they're done.
1142 void svc_close_net(struct svc_serv *serv, struct net *net)
1146 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1147 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1149 svc_clean_up_xprts(serv, net);
1155 * Handle defer and revisit of requests
1158 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1160 struct svc_deferred_req *dr =
1161 container_of(dreq, struct svc_deferred_req, handle);
1162 struct svc_xprt *xprt = dr->xprt;
1164 spin_lock(&xprt->xpt_lock);
1165 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1166 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1167 spin_unlock(&xprt->xpt_lock);
1168 trace_svc_defer_drop(dr);
1174 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1175 spin_unlock(&xprt->xpt_lock);
1176 trace_svc_defer_queue(dr);
1177 svc_xprt_enqueue(xprt);
1182 * Save the request off for later processing. The request buffer looks
1185 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1187 * This code can only handle requests that consist of an xprt-header
1190 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1192 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1193 struct svc_deferred_req *dr;
1195 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1196 return NULL; /* if more than a page, give up FIXME */
1197 if (rqstp->rq_deferred) {
1198 dr = rqstp->rq_deferred;
1199 rqstp->rq_deferred = NULL;
1203 /* FIXME maybe discard if size too large */
1204 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1205 dr = kmalloc(size, GFP_KERNEL);
1209 dr->handle.owner = rqstp->rq_server;
1210 dr->prot = rqstp->rq_prot;
1211 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1212 dr->addrlen = rqstp->rq_addrlen;
1213 dr->daddr = rqstp->rq_daddr;
1214 dr->argslen = rqstp->rq_arg.len >> 2;
1215 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1217 /* back up head to the start of the buffer and copy */
1218 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1219 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1222 trace_svc_defer(rqstp);
1223 svc_xprt_get(rqstp->rq_xprt);
1224 dr->xprt = rqstp->rq_xprt;
1225 set_bit(RQ_DROPME, &rqstp->rq_flags);
1227 dr->handle.revisit = svc_revisit;
1232 * recv data from a deferred request into an active one
1234 static noinline int svc_deferred_recv(struct svc_rqst *rqstp)
1236 struct svc_deferred_req *dr = rqstp->rq_deferred;
1238 trace_svc_defer_recv(dr);
1240 /* setup iov_base past transport header */
1241 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1242 /* The iov_len does not include the transport header bytes */
1243 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1244 rqstp->rq_arg.page_len = 0;
1245 /* The rq_arg.len includes the transport header bytes */
1246 rqstp->rq_arg.len = dr->argslen<<2;
1247 rqstp->rq_prot = dr->prot;
1248 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1249 rqstp->rq_addrlen = dr->addrlen;
1250 /* Save off transport header len in case we get deferred again */
1251 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1252 rqstp->rq_daddr = dr->daddr;
1253 rqstp->rq_respages = rqstp->rq_pages;
1254 svc_xprt_received(rqstp->rq_xprt);
1255 return (dr->argslen<<2) - dr->xprt_hlen;
1259 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1261 struct svc_deferred_req *dr = NULL;
1263 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1265 spin_lock(&xprt->xpt_lock);
1266 if (!list_empty(&xprt->xpt_deferred)) {
1267 dr = list_entry(xprt->xpt_deferred.next,
1268 struct svc_deferred_req,
1270 list_del_init(&dr->handle.recent);
1272 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1273 spin_unlock(&xprt->xpt_lock);
1278 * svc_find_xprt - find an RPC transport instance
1279 * @serv: pointer to svc_serv to search
1280 * @xcl_name: C string containing transport's class name
1281 * @net: owner net pointer
1282 * @af: Address family of transport's local address
1283 * @port: transport's IP port number
1285 * Return the transport instance pointer for the endpoint accepting
1286 * connections/peer traffic from the specified transport class,
1287 * address family and port.
1289 * Specifying 0 for the address family or port is effectively a
1290 * wild-card, and will result in matching the first transport in the
1291 * service's list that has a matching class name.
1293 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1294 struct net *net, const sa_family_t af,
1295 const unsigned short port)
1297 struct svc_xprt *xprt;
1298 struct svc_xprt *found = NULL;
1300 /* Sanity check the args */
1301 if (serv == NULL || xcl_name == NULL)
1304 spin_lock_bh(&serv->sv_lock);
1305 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1306 if (xprt->xpt_net != net)
1308 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1310 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1312 if (port != 0 && port != svc_xprt_local_port(xprt))
1318 spin_unlock_bh(&serv->sv_lock);
1321 EXPORT_SYMBOL_GPL(svc_find_xprt);
1323 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1324 char *pos, int remaining)
1328 len = snprintf(pos, remaining, "%s %u\n",
1329 xprt->xpt_class->xcl_name,
1330 svc_xprt_local_port(xprt));
1331 if (len >= remaining)
1332 return -ENAMETOOLONG;
1337 * svc_xprt_names - format a buffer with a list of transport names
1338 * @serv: pointer to an RPC service
1339 * @buf: pointer to a buffer to be filled in
1340 * @buflen: length of buffer to be filled in
1342 * Fills in @buf with a string containing a list of transport names,
1343 * each name terminated with '\n'.
1345 * Returns positive length of the filled-in string on success; otherwise
1346 * a negative errno value is returned if an error occurs.
1348 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1350 struct svc_xprt *xprt;
1354 /* Sanity check args */
1358 spin_lock_bh(&serv->sv_lock);
1362 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1363 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1375 spin_unlock_bh(&serv->sv_lock);
1378 EXPORT_SYMBOL_GPL(svc_xprt_names);
1381 /*----------------------------------------------------------------------------*/
1383 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1385 unsigned int pidx = (unsigned int)*pos;
1386 struct svc_serv *serv = m->private;
1388 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1391 return SEQ_START_TOKEN;
1392 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1395 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1397 struct svc_pool *pool = p;
1398 struct svc_serv *serv = m->private;
1400 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1402 if (p == SEQ_START_TOKEN) {
1403 pool = &serv->sv_pools[0];
1405 unsigned int pidx = (pool - &serv->sv_pools[0]);
1406 if (pidx < serv->sv_nrpools-1)
1407 pool = &serv->sv_pools[pidx+1];
1415 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1419 static int svc_pool_stats_show(struct seq_file *m, void *p)
1421 struct svc_pool *pool = p;
1423 if (p == SEQ_START_TOKEN) {
1424 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1428 seq_printf(m, "%u %lu %lu %lu %lu\n",
1430 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1431 pool->sp_stats.sockets_queued,
1432 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1433 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1438 static const struct seq_operations svc_pool_stats_seq_ops = {
1439 .start = svc_pool_stats_start,
1440 .next = svc_pool_stats_next,
1441 .stop = svc_pool_stats_stop,
1442 .show = svc_pool_stats_show,
1445 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1449 err = seq_open(file, &svc_pool_stats_seq_ops);
1451 ((struct seq_file *) file->private_data)->private = serv;
1454 EXPORT_SYMBOL(svc_pool_stats_open);
1456 /*----------------------------------------------------------------------------*/