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/sched/mm.h>
10 #include <linux/errno.h>
11 #include <linux/freezer.h>
12 #include <linux/kthread.h>
13 #include <linux/slab.h>
15 #include <linux/sunrpc/addr.h>
16 #include <linux/sunrpc/stats.h>
17 #include <linux/sunrpc/svc_xprt.h>
18 #include <linux/sunrpc/svcsock.h>
19 #include <linux/sunrpc/xprt.h>
20 #include <linux/module.h>
21 #include <linux/netdevice.h>
22 #include <trace/events/sunrpc.h>
24 #define RPCDBG_FACILITY RPCDBG_SVCXPRT
26 static unsigned int svc_rpc_per_connection_limit __read_mostly;
27 module_param(svc_rpc_per_connection_limit, uint, 0644);
30 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt);
31 static int svc_deferred_recv(struct svc_rqst *rqstp);
32 static struct cache_deferred_req *svc_defer(struct cache_req *req);
33 static void svc_age_temp_xprts(struct timer_list *t);
34 static void svc_delete_xprt(struct svc_xprt *xprt);
36 /* apparently the "standard" is that clients close
37 * idle connections after 5 minutes, servers after
39 * http://nfsv4bat.org/Documents/ConnectAThon/1996/nfstcp.pdf
41 static int svc_conn_age_period = 6*60;
43 /* List of registered transport classes */
44 static DEFINE_SPINLOCK(svc_xprt_class_lock);
45 static LIST_HEAD(svc_xprt_class_list);
47 /* SMP locking strategy:
49 * svc_pool->sp_lock protects most of the fields of that pool.
50 * svc_serv->sv_lock protects sv_tempsocks, sv_permsocks, sv_tmpcnt.
51 * when both need to be taken (rare), svc_serv->sv_lock is first.
52 * The "service mutex" protects svc_serv->sv_nrthread.
53 * svc_sock->sk_lock protects the svc_sock->sk_deferred list
54 * and the ->sk_info_authunix cache.
56 * The XPT_BUSY bit in xprt->xpt_flags prevents a transport being
57 * enqueued multiply. During normal transport processing this bit
58 * is set by svc_xprt_enqueue and cleared by svc_xprt_received.
59 * Providers should not manipulate this bit directly.
61 * Some flags can be set to certain values at any time
62 * providing that certain rules are followed:
65 * - Can be set or cleared at any time.
66 * - After a set, svc_xprt_enqueue must be called to enqueue
67 * the transport for processing.
68 * - After a clear, the transport must be read/accepted.
69 * If this succeeds, it must be set again.
71 * - Can set at any time. It is never cleared.
73 * - Can only be set while XPT_BUSY is held which ensures
74 * that no other thread will be using the transport or will
75 * try to set XPT_DEAD.
79 * svc_reg_xprt_class - Register a server-side RPC transport class
80 * @xcl: New transport class to be registered
82 * Returns zero on success; otherwise a negative errno is returned.
84 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
86 struct svc_xprt_class *cl;
89 INIT_LIST_HEAD(&xcl->xcl_list);
90 spin_lock(&svc_xprt_class_lock);
91 /* Make sure there isn't already a class with the same name */
92 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
93 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
96 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
99 spin_unlock(&svc_xprt_class_lock);
102 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
105 * svc_unreg_xprt_class - Unregister a server-side RPC transport class
106 * @xcl: Transport class to be unregistered
109 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
111 spin_lock(&svc_xprt_class_lock);
112 list_del_init(&xcl->xcl_list);
113 spin_unlock(&svc_xprt_class_lock);
115 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
118 * svc_print_xprts - Format the transport list for printing
119 * @buf: target buffer for formatted address
120 * @maxlen: length of target buffer
122 * Fills in @buf with a string containing a list of transport names, each name
123 * terminated with '\n'. If the buffer is too small, some entries may be
124 * missing, but it is guaranteed that all lines in the output buffer are
127 * Returns positive length of the filled-in string.
129 int svc_print_xprts(char *buf, int maxlen)
131 struct svc_xprt_class *xcl;
136 spin_lock(&svc_xprt_class_lock);
137 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
140 slen = snprintf(tmpstr, sizeof(tmpstr), "%s %d\n",
141 xcl->xcl_name, xcl->xcl_max_payload);
142 if (slen >= sizeof(tmpstr) || len + slen >= maxlen)
147 spin_unlock(&svc_xprt_class_lock);
153 * svc_xprt_deferred_close - Close a transport
154 * @xprt: transport instance
156 * Used in contexts that need to defer the work of shutting down
157 * the transport to an nfsd thread.
159 void svc_xprt_deferred_close(struct svc_xprt *xprt)
161 if (!test_and_set_bit(XPT_CLOSE, &xprt->xpt_flags))
162 svc_xprt_enqueue(xprt);
164 EXPORT_SYMBOL_GPL(svc_xprt_deferred_close);
166 static void svc_xprt_free(struct kref *kref)
168 struct svc_xprt *xprt =
169 container_of(kref, struct svc_xprt, xpt_ref);
170 struct module *owner = xprt->xpt_class->xcl_owner;
171 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
172 svcauth_unix_info_release(xprt);
173 put_cred(xprt->xpt_cred);
174 put_net_track(xprt->xpt_net, &xprt->ns_tracker);
175 /* See comment on corresponding get in xs_setup_bc_tcp(): */
176 if (xprt->xpt_bc_xprt)
177 xprt_put(xprt->xpt_bc_xprt);
178 if (xprt->xpt_bc_xps)
179 xprt_switch_put(xprt->xpt_bc_xps);
180 trace_svc_xprt_free(xprt);
181 xprt->xpt_ops->xpo_free(xprt);
185 void svc_xprt_put(struct svc_xprt *xprt)
187 kref_put(&xprt->xpt_ref, svc_xprt_free);
189 EXPORT_SYMBOL_GPL(svc_xprt_put);
192 * Called by transport drivers to initialize the transport independent
193 * portion of the transport instance.
195 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
196 struct svc_xprt *xprt, struct svc_serv *serv)
198 memset(xprt, 0, sizeof(*xprt));
199 xprt->xpt_class = xcl;
200 xprt->xpt_ops = xcl->xcl_ops;
201 kref_init(&xprt->xpt_ref);
202 xprt->xpt_server = serv;
203 INIT_LIST_HEAD(&xprt->xpt_list);
204 INIT_LIST_HEAD(&xprt->xpt_ready);
205 INIT_LIST_HEAD(&xprt->xpt_deferred);
206 INIT_LIST_HEAD(&xprt->xpt_users);
207 mutex_init(&xprt->xpt_mutex);
208 spin_lock_init(&xprt->xpt_lock);
209 set_bit(XPT_BUSY, &xprt->xpt_flags);
210 xprt->xpt_net = get_net_track(net, &xprt->ns_tracker, GFP_ATOMIC);
211 strcpy(xprt->xpt_remotebuf, "uninitialized");
213 EXPORT_SYMBOL_GPL(svc_xprt_init);
215 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
216 struct svc_serv *serv,
219 const unsigned short port,
222 struct sockaddr_in sin = {
223 .sin_family = AF_INET,
224 .sin_addr.s_addr = htonl(INADDR_ANY),
225 .sin_port = htons(port),
227 #if IS_ENABLED(CONFIG_IPV6)
228 struct sockaddr_in6 sin6 = {
229 .sin6_family = AF_INET6,
230 .sin6_addr = IN6ADDR_ANY_INIT,
231 .sin6_port = htons(port),
234 struct svc_xprt *xprt;
235 struct sockaddr *sap;
240 sap = (struct sockaddr *)&sin;
243 #if IS_ENABLED(CONFIG_IPV6)
245 sap = (struct sockaddr *)&sin6;
250 return ERR_PTR(-EAFNOSUPPORT);
253 xprt = xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
255 trace_svc_xprt_create_err(serv->sv_program->pg_name,
256 xcl->xcl_name, sap, len, xprt);
261 * svc_xprt_received - start next receiver thread
262 * @xprt: controlling transport
264 * The caller must hold the XPT_BUSY bit and must
265 * not thereafter touch transport data.
267 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
268 * insufficient) data.
270 void svc_xprt_received(struct svc_xprt *xprt)
272 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
273 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
277 /* As soon as we clear busy, the xprt could be closed and
278 * 'put', so we need a reference to call svc_xprt_enqueue with:
281 smp_mb__before_atomic();
282 clear_bit(XPT_BUSY, &xprt->xpt_flags);
283 svc_xprt_enqueue(xprt);
286 EXPORT_SYMBOL_GPL(svc_xprt_received);
288 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
290 clear_bit(XPT_TEMP, &new->xpt_flags);
291 spin_lock_bh(&serv->sv_lock);
292 list_add(&new->xpt_list, &serv->sv_permsocks);
293 spin_unlock_bh(&serv->sv_lock);
294 svc_xprt_received(new);
297 static int _svc_xprt_create(struct svc_serv *serv, const char *xprt_name,
298 struct net *net, const int family,
299 const unsigned short port, int flags,
300 const struct cred *cred)
302 struct svc_xprt_class *xcl;
304 spin_lock(&svc_xprt_class_lock);
305 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
306 struct svc_xprt *newxprt;
307 unsigned short newport;
309 if (strcmp(xprt_name, xcl->xcl_name))
312 if (!try_module_get(xcl->xcl_owner))
315 spin_unlock(&svc_xprt_class_lock);
316 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
317 if (IS_ERR(newxprt)) {
318 module_put(xcl->xcl_owner);
319 return PTR_ERR(newxprt);
321 newxprt->xpt_cred = get_cred(cred);
322 svc_add_new_perm_xprt(serv, newxprt);
323 newport = svc_xprt_local_port(newxprt);
327 spin_unlock(&svc_xprt_class_lock);
328 /* This errno is exposed to user space. Provide a reasonable
329 * perror msg for a bad transport. */
330 return -EPROTONOSUPPORT;
334 * svc_xprt_create - Add a new listener to @serv
335 * @serv: target RPC service
336 * @xprt_name: transport class name
337 * @net: network namespace
338 * @family: network address family
339 * @port: listener port
340 * @flags: SVC_SOCK flags
341 * @cred: credential to bind to this transport
344 * %0: New listener added successfully
345 * %-EPROTONOSUPPORT: Requested transport type not supported
347 int svc_xprt_create(struct svc_serv *serv, const char *xprt_name,
348 struct net *net, const int family,
349 const unsigned short port, int flags,
350 const struct cred *cred)
354 err = _svc_xprt_create(serv, xprt_name, net, family, port, flags, cred);
355 if (err == -EPROTONOSUPPORT) {
356 request_module("svc%s", xprt_name);
357 err = _svc_xprt_create(serv, xprt_name, net, family, port, flags, cred);
361 EXPORT_SYMBOL_GPL(svc_xprt_create);
364 * Copy the local and remote xprt addresses to the rqstp structure
366 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
368 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
369 rqstp->rq_addrlen = xprt->xpt_remotelen;
372 * Destination address in request is needed for binding the
373 * source address in RPC replies/callbacks later.
375 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
376 rqstp->rq_daddrlen = xprt->xpt_locallen;
378 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
381 * svc_print_addr - Format rq_addr field for printing
382 * @rqstp: svc_rqst struct containing address to print
383 * @buf: target buffer for formatted address
384 * @len: length of target buffer
387 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
389 return __svc_print_addr(svc_addr(rqstp), buf, len);
391 EXPORT_SYMBOL_GPL(svc_print_addr);
393 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
395 unsigned int limit = svc_rpc_per_connection_limit;
396 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
398 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
401 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
403 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
404 if (!svc_xprt_slots_in_range(xprt))
406 atomic_inc(&xprt->xpt_nr_rqsts);
407 set_bit(RQ_DATA, &rqstp->rq_flags);
412 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
414 struct svc_xprt *xprt = rqstp->rq_xprt;
415 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
416 atomic_dec(&xprt->xpt_nr_rqsts);
417 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
418 svc_xprt_enqueue(xprt);
422 static bool svc_xprt_ready(struct svc_xprt *xprt)
424 unsigned long xpt_flags;
427 * If another cpu has recently updated xpt_flags,
428 * sk_sock->flags, xpt_reserved, or xpt_nr_rqsts, we need to
429 * know about it; otherwise it's possible that both that cpu and
430 * this one could call svc_xprt_enqueue() without either
431 * svc_xprt_enqueue() recognizing that the conditions below
432 * are satisfied, and we could stall indefinitely:
435 xpt_flags = READ_ONCE(xprt->xpt_flags);
437 if (xpt_flags & BIT(XPT_BUSY))
439 if (xpt_flags & (BIT(XPT_CONN) | BIT(XPT_CLOSE) | BIT(XPT_HANDSHAKE)))
441 if (xpt_flags & (BIT(XPT_DATA) | BIT(XPT_DEFERRED))) {
442 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
443 svc_xprt_slots_in_range(xprt))
445 trace_svc_xprt_no_write_space(xprt);
452 * svc_xprt_enqueue - Queue a transport on an idle nfsd thread
453 * @xprt: transport with data pending
456 void svc_xprt_enqueue(struct svc_xprt *xprt)
458 struct svc_pool *pool;
459 struct svc_rqst *rqstp = NULL;
461 if (!svc_xprt_ready(xprt))
464 /* Mark transport as busy. It will remain in this state until
465 * the provider calls svc_xprt_received. We update XPT_BUSY
466 * atomically because it also guards against trying to enqueue
467 * the transport twice.
469 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
472 pool = svc_pool_for_cpu(xprt->xpt_server);
474 percpu_counter_inc(&pool->sp_sockets_queued);
475 spin_lock_bh(&pool->sp_lock);
476 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
477 spin_unlock_bh(&pool->sp_lock);
479 /* find a thread for this xprt */
481 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
482 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
484 percpu_counter_inc(&pool->sp_threads_woken);
485 rqstp->rq_qtime = ktime_get();
486 wake_up_process(rqstp->rq_task);
489 set_bit(SP_CONGESTED, &pool->sp_flags);
493 trace_svc_xprt_enqueue(xprt, rqstp);
495 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
498 * Dequeue the first transport, if there is one.
500 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
502 struct svc_xprt *xprt = NULL;
504 if (list_empty(&pool->sp_sockets))
507 spin_lock_bh(&pool->sp_lock);
508 if (likely(!list_empty(&pool->sp_sockets))) {
509 xprt = list_first_entry(&pool->sp_sockets,
510 struct svc_xprt, xpt_ready);
511 list_del_init(&xprt->xpt_ready);
514 spin_unlock_bh(&pool->sp_lock);
520 * svc_reserve - change the space reserved for the reply to a request.
521 * @rqstp: The request in question
522 * @space: new max space to reserve
524 * Each request reserves some space on the output queue of the transport
525 * to make sure the reply fits. This function reduces that reserved
526 * space to be the amount of space used already, plus @space.
529 void svc_reserve(struct svc_rqst *rqstp, int space)
531 struct svc_xprt *xprt = rqstp->rq_xprt;
533 space += rqstp->rq_res.head[0].iov_len;
535 if (xprt && space < rqstp->rq_reserved) {
536 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
537 rqstp->rq_reserved = space;
538 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
539 svc_xprt_enqueue(xprt);
542 EXPORT_SYMBOL_GPL(svc_reserve);
544 static void free_deferred(struct svc_xprt *xprt, struct svc_deferred_req *dr)
549 xprt->xpt_ops->xpo_release_ctxt(xprt, dr->xprt_ctxt);
553 static void svc_xprt_release(struct svc_rqst *rqstp)
555 struct svc_xprt *xprt = rqstp->rq_xprt;
557 xprt->xpt_ops->xpo_release_ctxt(xprt, rqstp->rq_xprt_ctxt);
558 rqstp->rq_xprt_ctxt = NULL;
560 free_deferred(xprt, rqstp->rq_deferred);
561 rqstp->rq_deferred = NULL;
563 svc_rqst_release_pages(rqstp);
564 rqstp->rq_res.page_len = 0;
565 rqstp->rq_res.page_base = 0;
567 /* Reset response buffer and release
569 * But first, check that enough space was reserved
570 * for the reply, otherwise we have a bug!
572 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
573 printk(KERN_ERR "RPC request reserved %d but used %d\n",
577 rqstp->rq_res.head[0].iov_len = 0;
578 svc_reserve(rqstp, 0);
579 svc_xprt_release_slot(rqstp);
580 rqstp->rq_xprt = NULL;
585 * Some svc_serv's will have occasional work to do, even when a xprt is not
586 * waiting to be serviced. This function is there to "kick" a task in one of
587 * those services so that it can wake up and do that work. Note that we only
588 * bother with pool 0 as we don't need to wake up more than one thread for
591 void svc_wake_up(struct svc_serv *serv)
593 struct svc_rqst *rqstp;
594 struct svc_pool *pool;
596 pool = &serv->sv_pools[0];
599 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
600 /* skip any that aren't queued */
601 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
604 wake_up_process(rqstp->rq_task);
605 trace_svc_wake_up(rqstp->rq_task->pid);
610 /* No free entries available */
611 set_bit(SP_TASK_PENDING, &pool->sp_flags);
613 trace_svc_wake_up(0);
615 EXPORT_SYMBOL_GPL(svc_wake_up);
617 int svc_port_is_privileged(struct sockaddr *sin)
619 switch (sin->sa_family) {
621 return ntohs(((struct sockaddr_in *)sin)->sin_port)
624 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
632 * Make sure that we don't have too many active connections. If we have,
633 * something must be dropped. It's not clear what will happen if we allow
634 * "too many" connections, but when dealing with network-facing software,
635 * we have to code defensively. Here we do that by imposing hard limits.
637 * There's no point in trying to do random drop here for DoS
638 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
639 * attacker can easily beat that.
641 * The only somewhat efficient mechanism would be if drop old
642 * connections from the same IP first. But right now we don't even
643 * record the client IP in svc_sock.
645 * single-threaded services that expect a lot of clients will probably
646 * need to set sv_maxconn to override the default value which is based
647 * on the number of threads
649 static void svc_check_conn_limits(struct svc_serv *serv)
651 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
652 (serv->sv_nrthreads+3) * 20;
654 if (serv->sv_tmpcnt > limit) {
655 struct svc_xprt *xprt = NULL;
656 spin_lock_bh(&serv->sv_lock);
657 if (!list_empty(&serv->sv_tempsocks)) {
658 /* Try to help the admin */
659 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
660 serv->sv_name, serv->sv_maxconn ?
661 "max number of connections" :
662 "number of threads");
664 * Always select the oldest connection. It's not fair,
667 xprt = list_entry(serv->sv_tempsocks.prev,
670 set_bit(XPT_CLOSE, &xprt->xpt_flags);
673 spin_unlock_bh(&serv->sv_lock);
676 svc_xprt_enqueue(xprt);
682 static int svc_alloc_arg(struct svc_rqst *rqstp)
684 struct svc_serv *serv = rqstp->rq_server;
685 struct xdr_buf *arg = &rqstp->rq_arg;
686 unsigned long pages, filled, ret;
688 pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
689 if (pages > RPCSVC_MAXPAGES) {
690 pr_warn_once("svc: warning: pages=%lu > RPCSVC_MAXPAGES=%lu\n",
691 pages, RPCSVC_MAXPAGES);
692 /* use as many pages as possible */
693 pages = RPCSVC_MAXPAGES;
696 for (filled = 0; filled < pages; filled = ret) {
697 ret = alloc_pages_bulk_array_node(GFP_KERNEL,
698 rqstp->rq_pool->sp_id,
699 pages, rqstp->rq_pages);
701 /* Made progress, don't sleep yet */
704 set_current_state(TASK_IDLE);
705 if (kthread_should_stop()) {
706 set_current_state(TASK_RUNNING);
709 trace_svc_alloc_arg_err(pages, ret);
710 memalloc_retry_wait(GFP_KERNEL);
712 rqstp->rq_page_end = &rqstp->rq_pages[pages];
713 rqstp->rq_pages[pages] = NULL; /* this might be seen in nfsd_splice_actor() */
715 /* Make arg->head point to first page and arg->pages point to rest */
716 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
717 arg->head[0].iov_len = PAGE_SIZE;
718 arg->pages = rqstp->rq_pages + 1;
720 /* save at least one page for response */
721 arg->page_len = (pages-2)*PAGE_SIZE;
722 arg->len = (pages-1)*PAGE_SIZE;
723 arg->tail[0].iov_len = 0;
725 rqstp->rq_xid = xdr_zero;
730 rqst_should_sleep(struct svc_rqst *rqstp)
732 struct svc_pool *pool = rqstp->rq_pool;
734 /* did someone call svc_wake_up? */
735 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
738 /* was a socket queued? */
739 if (!list_empty(&pool->sp_sockets))
742 /* are we shutting down? */
743 if (kthread_should_stop())
746 /* are we freezing? */
747 if (freezing(current))
753 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
755 struct svc_pool *pool = rqstp->rq_pool;
758 /* rq_xprt should be clear on entry */
759 WARN_ON_ONCE(rqstp->rq_xprt);
761 rqstp->rq_xprt = svc_xprt_dequeue(pool);
765 set_current_state(TASK_IDLE);
766 smp_mb__before_atomic();
767 clear_bit(SP_CONGESTED, &pool->sp_flags);
768 clear_bit(RQ_BUSY, &rqstp->rq_flags);
769 smp_mb__after_atomic();
771 if (likely(rqst_should_sleep(rqstp)))
772 time_left = schedule_timeout(timeout);
774 __set_current_state(TASK_RUNNING);
778 set_bit(RQ_BUSY, &rqstp->rq_flags);
779 smp_mb__after_atomic();
780 rqstp->rq_xprt = svc_xprt_dequeue(pool);
785 percpu_counter_inc(&pool->sp_threads_timedout);
787 if (kthread_should_stop())
788 return ERR_PTR(-EINTR);
789 return ERR_PTR(-EAGAIN);
791 /* Normally we will wait up to 5 seconds for any required
792 * cache information to be provided.
794 if (!test_bit(SP_CONGESTED, &pool->sp_flags))
795 rqstp->rq_chandle.thread_wait = 5*HZ;
797 rqstp->rq_chandle.thread_wait = 1*HZ;
798 trace_svc_xprt_dequeue(rqstp);
799 return rqstp->rq_xprt;
802 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
804 spin_lock_bh(&serv->sv_lock);
805 set_bit(XPT_TEMP, &newxpt->xpt_flags);
806 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
808 if (serv->sv_temptimer.function == NULL) {
809 /* setup timer to age temp transports */
810 serv->sv_temptimer.function = svc_age_temp_xprts;
811 mod_timer(&serv->sv_temptimer,
812 jiffies + svc_conn_age_period * HZ);
814 spin_unlock_bh(&serv->sv_lock);
815 svc_xprt_received(newxpt);
818 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
820 struct svc_serv *serv = rqstp->rq_server;
823 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
824 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
825 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
826 svc_delete_xprt(xprt);
827 /* Leave XPT_BUSY set on the dead xprt: */
830 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
831 struct svc_xprt *newxpt;
833 * We know this module_get will succeed because the
834 * listener holds a reference too
836 __module_get(xprt->xpt_class->xcl_owner);
837 svc_check_conn_limits(xprt->xpt_server);
838 newxpt = xprt->xpt_ops->xpo_accept(xprt);
840 newxpt->xpt_cred = get_cred(xprt->xpt_cred);
841 svc_add_new_temp_xprt(serv, newxpt);
842 trace_svc_xprt_accept(newxpt, serv->sv_name);
844 module_put(xprt->xpt_class->xcl_owner);
846 svc_xprt_received(xprt);
847 } else if (test_bit(XPT_HANDSHAKE, &xprt->xpt_flags)) {
848 xprt->xpt_ops->xpo_handshake(xprt);
849 svc_xprt_received(xprt);
850 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
851 /* XPT_DATA|XPT_DEFERRED case: */
852 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
853 if (rqstp->rq_deferred)
854 len = svc_deferred_recv(rqstp);
856 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
857 rqstp->rq_reserved = serv->sv_max_mesg;
858 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
860 svc_xprt_received(xprt);
867 * Receive the next request on any transport. This code is carefully
868 * organised not to touch any cachelines in the shared svc_serv
869 * structure, only cachelines in the local svc_pool.
871 int svc_recv(struct svc_rqst *rqstp, long timeout)
873 struct svc_xprt *xprt = NULL;
874 struct svc_serv *serv = rqstp->rq_server;
877 err = svc_alloc_arg(rqstp);
884 if (kthread_should_stop())
887 xprt = svc_get_next_xprt(rqstp, timeout);
893 len = svc_handle_xprt(rqstp, xprt);
895 /* No data, incomplete (TCP) read, or accept() */
900 trace_svc_xdr_recvfrom(&rqstp->rq_arg);
902 clear_bit(XPT_OLD, &xprt->xpt_flags);
904 rqstp->rq_chandle.defer = svc_defer;
907 serv->sv_stats->netcnt++;
908 rqstp->rq_stime = ktime_get();
911 rqstp->rq_res.len = 0;
912 svc_xprt_release(rqstp);
916 EXPORT_SYMBOL_GPL(svc_recv);
921 void svc_drop(struct svc_rqst *rqstp)
923 trace_svc_drop(rqstp);
924 svc_xprt_release(rqstp);
926 EXPORT_SYMBOL_GPL(svc_drop);
929 * svc_send - Return reply to client
930 * @rqstp: RPC transaction context
933 void svc_send(struct svc_rqst *rqstp)
935 struct svc_xprt *xprt;
939 xprt = rqstp->rq_xprt;
943 /* calculate over-all length */
945 xb->len = xb->head[0].iov_len +
948 trace_svc_xdr_sendto(rqstp->rq_xid, xb);
949 trace_svc_stats_latency(rqstp);
951 status = xprt->xpt_ops->xpo_sendto(rqstp);
953 trace_svc_send(rqstp, status);
954 svc_xprt_release(rqstp);
958 * Timer function to close old temporary transports, using
959 * a mark-and-sweep algorithm.
961 static void svc_age_temp_xprts(struct timer_list *t)
963 struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
964 struct svc_xprt *xprt;
965 struct list_head *le, *next;
967 dprintk("svc_age_temp_xprts\n");
969 if (!spin_trylock_bh(&serv->sv_lock)) {
970 /* busy, try again 1 sec later */
971 dprintk("svc_age_temp_xprts: busy\n");
972 mod_timer(&serv->sv_temptimer, jiffies + HZ);
976 list_for_each_safe(le, next, &serv->sv_tempsocks) {
977 xprt = list_entry(le, struct svc_xprt, xpt_list);
979 /* First time through, just mark it OLD. Second time
980 * through, close it. */
981 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
983 if (kref_read(&xprt->xpt_ref) > 1 ||
984 test_bit(XPT_BUSY, &xprt->xpt_flags))
987 set_bit(XPT_CLOSE, &xprt->xpt_flags);
988 dprintk("queuing xprt %p for closing\n", xprt);
990 /* a thread will dequeue and close it soon */
991 svc_xprt_enqueue(xprt);
993 spin_unlock_bh(&serv->sv_lock);
995 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
998 /* Close temporary transports whose xpt_local matches server_addr immediately
999 * instead of waiting for them to be picked up by the timer.
1001 * This is meant to be called from a notifier_block that runs when an ip
1002 * address is deleted.
1004 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
1006 struct svc_xprt *xprt;
1007 struct list_head *le, *next;
1008 LIST_HEAD(to_be_closed);
1010 spin_lock_bh(&serv->sv_lock);
1011 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1012 xprt = list_entry(le, struct svc_xprt, xpt_list);
1013 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
1014 &xprt->xpt_local)) {
1015 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
1016 list_move(le, &to_be_closed);
1019 spin_unlock_bh(&serv->sv_lock);
1021 while (!list_empty(&to_be_closed)) {
1022 le = to_be_closed.next;
1024 xprt = list_entry(le, struct svc_xprt, xpt_list);
1025 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1026 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
1027 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
1029 svc_xprt_enqueue(xprt);
1032 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1034 static void call_xpt_users(struct svc_xprt *xprt)
1036 struct svc_xpt_user *u;
1038 spin_lock(&xprt->xpt_lock);
1039 while (!list_empty(&xprt->xpt_users)) {
1040 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1041 list_del_init(&u->list);
1044 spin_unlock(&xprt->xpt_lock);
1048 * Remove a dead transport
1050 static void svc_delete_xprt(struct svc_xprt *xprt)
1052 struct svc_serv *serv = xprt->xpt_server;
1053 struct svc_deferred_req *dr;
1055 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1058 trace_svc_xprt_detach(xprt);
1059 xprt->xpt_ops->xpo_detach(xprt);
1060 if (xprt->xpt_bc_xprt)
1061 xprt->xpt_bc_xprt->ops->close(xprt->xpt_bc_xprt);
1063 spin_lock_bh(&serv->sv_lock);
1064 list_del_init(&xprt->xpt_list);
1065 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1066 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1068 spin_unlock_bh(&serv->sv_lock);
1070 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1071 free_deferred(xprt, dr);
1073 call_xpt_users(xprt);
1078 * svc_xprt_close - Close a client connection
1079 * @xprt: transport to disconnect
1082 void svc_xprt_close(struct svc_xprt *xprt)
1084 trace_svc_xprt_close(xprt);
1085 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1086 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1087 /* someone else will have to effect the close */
1090 * We expect svc_close_xprt() to work even when no threads are
1091 * running (e.g., while configuring the server before starting
1092 * any threads), so if the transport isn't busy, we delete
1095 svc_delete_xprt(xprt);
1097 EXPORT_SYMBOL_GPL(svc_xprt_close);
1099 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1101 struct svc_xprt *xprt;
1104 spin_lock_bh(&serv->sv_lock);
1105 list_for_each_entry(xprt, xprt_list, xpt_list) {
1106 if (xprt->xpt_net != net)
1109 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1110 svc_xprt_enqueue(xprt);
1112 spin_unlock_bh(&serv->sv_lock);
1116 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1118 struct svc_pool *pool;
1119 struct svc_xprt *xprt;
1120 struct svc_xprt *tmp;
1123 for (i = 0; i < serv->sv_nrpools; i++) {
1124 pool = &serv->sv_pools[i];
1126 spin_lock_bh(&pool->sp_lock);
1127 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1128 if (xprt->xpt_net != net)
1130 list_del_init(&xprt->xpt_ready);
1131 spin_unlock_bh(&pool->sp_lock);
1134 spin_unlock_bh(&pool->sp_lock);
1139 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1141 struct svc_xprt *xprt;
1143 while ((xprt = svc_dequeue_net(serv, net))) {
1144 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1145 svc_delete_xprt(xprt);
1150 * svc_xprt_destroy_all - Destroy transports associated with @serv
1151 * @serv: RPC service to be shut down
1152 * @net: target network namespace
1154 * Server threads may still be running (especially in the case where the
1155 * service is still running in other network namespaces).
1157 * So we shut down sockets the same way we would on a running server, by
1158 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1159 * the close. In the case there are no such other threads,
1160 * threads running, svc_clean_up_xprts() does a simple version of a
1161 * server's main event loop, and in the case where there are other
1162 * threads, we may need to wait a little while and then check again to
1163 * see if they're done.
1165 void svc_xprt_destroy_all(struct svc_serv *serv, struct net *net)
1169 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1170 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1172 svc_clean_up_xprts(serv, net);
1176 EXPORT_SYMBOL_GPL(svc_xprt_destroy_all);
1179 * Handle defer and revisit of requests
1182 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1184 struct svc_deferred_req *dr =
1185 container_of(dreq, struct svc_deferred_req, handle);
1186 struct svc_xprt *xprt = dr->xprt;
1188 spin_lock(&xprt->xpt_lock);
1189 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1190 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1191 spin_unlock(&xprt->xpt_lock);
1192 trace_svc_defer_drop(dr);
1193 free_deferred(xprt, dr);
1198 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1199 spin_unlock(&xprt->xpt_lock);
1200 trace_svc_defer_queue(dr);
1201 svc_xprt_enqueue(xprt);
1206 * Save the request off for later processing. The request buffer looks
1209 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1211 * This code can only handle requests that consist of an xprt-header
1214 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1216 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1217 struct svc_deferred_req *dr;
1219 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1220 return NULL; /* if more than a page, give up FIXME */
1221 if (rqstp->rq_deferred) {
1222 dr = rqstp->rq_deferred;
1223 rqstp->rq_deferred = NULL;
1227 /* FIXME maybe discard if size too large */
1228 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1229 dr = kmalloc(size, GFP_KERNEL);
1233 dr->handle.owner = rqstp->rq_server;
1234 dr->prot = rqstp->rq_prot;
1235 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1236 dr->addrlen = rqstp->rq_addrlen;
1237 dr->daddr = rqstp->rq_daddr;
1238 dr->argslen = rqstp->rq_arg.len >> 2;
1240 /* back up head to the start of the buffer and copy */
1241 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1242 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1245 dr->xprt_ctxt = rqstp->rq_xprt_ctxt;
1246 rqstp->rq_xprt_ctxt = NULL;
1247 trace_svc_defer(rqstp);
1248 svc_xprt_get(rqstp->rq_xprt);
1249 dr->xprt = rqstp->rq_xprt;
1250 set_bit(RQ_DROPME, &rqstp->rq_flags);
1252 dr->handle.revisit = svc_revisit;
1257 * recv data from a deferred request into an active one
1259 static noinline int svc_deferred_recv(struct svc_rqst *rqstp)
1261 struct svc_deferred_req *dr = rqstp->rq_deferred;
1263 trace_svc_defer_recv(dr);
1265 /* setup iov_base past transport header */
1266 rqstp->rq_arg.head[0].iov_base = dr->args;
1267 /* The iov_len does not include the transport header bytes */
1268 rqstp->rq_arg.head[0].iov_len = dr->argslen << 2;
1269 rqstp->rq_arg.page_len = 0;
1270 /* The rq_arg.len includes the transport header bytes */
1271 rqstp->rq_arg.len = dr->argslen << 2;
1272 rqstp->rq_prot = dr->prot;
1273 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1274 rqstp->rq_addrlen = dr->addrlen;
1275 /* Save off transport header len in case we get deferred again */
1276 rqstp->rq_daddr = dr->daddr;
1277 rqstp->rq_respages = rqstp->rq_pages;
1278 rqstp->rq_xprt_ctxt = dr->xprt_ctxt;
1280 dr->xprt_ctxt = NULL;
1281 svc_xprt_received(rqstp->rq_xprt);
1282 return dr->argslen << 2;
1286 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1288 struct svc_deferred_req *dr = NULL;
1290 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1292 spin_lock(&xprt->xpt_lock);
1293 if (!list_empty(&xprt->xpt_deferred)) {
1294 dr = list_entry(xprt->xpt_deferred.next,
1295 struct svc_deferred_req,
1297 list_del_init(&dr->handle.recent);
1299 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1300 spin_unlock(&xprt->xpt_lock);
1305 * svc_find_xprt - find an RPC transport instance
1306 * @serv: pointer to svc_serv to search
1307 * @xcl_name: C string containing transport's class name
1308 * @net: owner net pointer
1309 * @af: Address family of transport's local address
1310 * @port: transport's IP port number
1312 * Return the transport instance pointer for the endpoint accepting
1313 * connections/peer traffic from the specified transport class,
1314 * address family and port.
1316 * Specifying 0 for the address family or port is effectively a
1317 * wild-card, and will result in matching the first transport in the
1318 * service's list that has a matching class name.
1320 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1321 struct net *net, const sa_family_t af,
1322 const unsigned short port)
1324 struct svc_xprt *xprt;
1325 struct svc_xprt *found = NULL;
1327 /* Sanity check the args */
1328 if (serv == NULL || xcl_name == NULL)
1331 spin_lock_bh(&serv->sv_lock);
1332 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1333 if (xprt->xpt_net != net)
1335 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1337 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1339 if (port != 0 && port != svc_xprt_local_port(xprt))
1345 spin_unlock_bh(&serv->sv_lock);
1348 EXPORT_SYMBOL_GPL(svc_find_xprt);
1350 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1351 char *pos, int remaining)
1355 len = snprintf(pos, remaining, "%s %u\n",
1356 xprt->xpt_class->xcl_name,
1357 svc_xprt_local_port(xprt));
1358 if (len >= remaining)
1359 return -ENAMETOOLONG;
1364 * svc_xprt_names - format a buffer with a list of transport names
1365 * @serv: pointer to an RPC service
1366 * @buf: pointer to a buffer to be filled in
1367 * @buflen: length of buffer to be filled in
1369 * Fills in @buf with a string containing a list of transport names,
1370 * each name terminated with '\n'.
1372 * Returns positive length of the filled-in string on success; otherwise
1373 * a negative errno value is returned if an error occurs.
1375 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1377 struct svc_xprt *xprt;
1381 /* Sanity check args */
1385 spin_lock_bh(&serv->sv_lock);
1389 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1390 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1402 spin_unlock_bh(&serv->sv_lock);
1405 EXPORT_SYMBOL_GPL(svc_xprt_names);
1408 /*----------------------------------------------------------------------------*/
1410 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1412 unsigned int pidx = (unsigned int)*pos;
1413 struct svc_serv *serv = m->private;
1415 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1418 return SEQ_START_TOKEN;
1419 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1422 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1424 struct svc_pool *pool = p;
1425 struct svc_serv *serv = m->private;
1427 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1429 if (p == SEQ_START_TOKEN) {
1430 pool = &serv->sv_pools[0];
1432 unsigned int pidx = (pool - &serv->sv_pools[0]);
1433 if (pidx < serv->sv_nrpools-1)
1434 pool = &serv->sv_pools[pidx+1];
1442 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1446 static int svc_pool_stats_show(struct seq_file *m, void *p)
1448 struct svc_pool *pool = p;
1450 if (p == SEQ_START_TOKEN) {
1451 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1455 seq_printf(m, "%u %llu %llu %llu %llu\n",
1457 percpu_counter_sum_positive(&pool->sp_sockets_queued),
1458 percpu_counter_sum_positive(&pool->sp_sockets_queued),
1459 percpu_counter_sum_positive(&pool->sp_threads_woken),
1460 percpu_counter_sum_positive(&pool->sp_threads_timedout));
1465 static const struct seq_operations svc_pool_stats_seq_ops = {
1466 .start = svc_pool_stats_start,
1467 .next = svc_pool_stats_next,
1468 .stop = svc_pool_stats_stop,
1469 .show = svc_pool_stats_show,
1472 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1476 err = seq_open(file, &svc_pool_stats_seq_ops);
1478 ((struct seq_file *) file->private_data)->private = serv;
1481 EXPORT_SYMBOL(svc_pool_stats_open);
1483 /*----------------------------------------------------------------------------*/