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/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/sunrpc/bc_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_deferred);
205 INIT_LIST_HEAD(&xprt->xpt_users);
206 mutex_init(&xprt->xpt_mutex);
207 spin_lock_init(&xprt->xpt_lock);
208 set_bit(XPT_BUSY, &xprt->xpt_flags);
209 xprt->xpt_net = get_net_track(net, &xprt->ns_tracker, GFP_ATOMIC);
210 strcpy(xprt->xpt_remotebuf, "uninitialized");
212 EXPORT_SYMBOL_GPL(svc_xprt_init);
214 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
215 struct svc_serv *serv,
218 const unsigned short port,
221 struct sockaddr_in sin = {
222 .sin_family = AF_INET,
223 .sin_addr.s_addr = htonl(INADDR_ANY),
224 .sin_port = htons(port),
226 #if IS_ENABLED(CONFIG_IPV6)
227 struct sockaddr_in6 sin6 = {
228 .sin6_family = AF_INET6,
229 .sin6_addr = IN6ADDR_ANY_INIT,
230 .sin6_port = htons(port),
233 struct svc_xprt *xprt;
234 struct sockaddr *sap;
239 sap = (struct sockaddr *)&sin;
242 #if IS_ENABLED(CONFIG_IPV6)
244 sap = (struct sockaddr *)&sin6;
249 return ERR_PTR(-EAFNOSUPPORT);
252 xprt = xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
254 trace_svc_xprt_create_err(serv->sv_program->pg_name,
255 xcl->xcl_name, sap, len, xprt);
260 * svc_xprt_received - start next receiver thread
261 * @xprt: controlling transport
263 * The caller must hold the XPT_BUSY bit and must
264 * not thereafter touch transport data.
266 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
267 * insufficient) data.
269 void svc_xprt_received(struct svc_xprt *xprt)
271 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
272 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
276 /* As soon as we clear busy, the xprt could be closed and
277 * 'put', so we need a reference to call svc_xprt_enqueue with:
280 smp_mb__before_atomic();
281 clear_bit(XPT_BUSY, &xprt->xpt_flags);
282 svc_xprt_enqueue(xprt);
285 EXPORT_SYMBOL_GPL(svc_xprt_received);
287 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
289 clear_bit(XPT_TEMP, &new->xpt_flags);
290 spin_lock_bh(&serv->sv_lock);
291 list_add(&new->xpt_list, &serv->sv_permsocks);
292 spin_unlock_bh(&serv->sv_lock);
293 svc_xprt_received(new);
296 static int _svc_xprt_create(struct svc_serv *serv, const char *xprt_name,
297 struct net *net, const int family,
298 const unsigned short port, int flags,
299 const struct cred *cred)
301 struct svc_xprt_class *xcl;
303 spin_lock(&svc_xprt_class_lock);
304 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
305 struct svc_xprt *newxprt;
306 unsigned short newport;
308 if (strcmp(xprt_name, xcl->xcl_name))
311 if (!try_module_get(xcl->xcl_owner))
314 spin_unlock(&svc_xprt_class_lock);
315 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
316 if (IS_ERR(newxprt)) {
317 module_put(xcl->xcl_owner);
318 return PTR_ERR(newxprt);
320 newxprt->xpt_cred = get_cred(cred);
321 svc_add_new_perm_xprt(serv, newxprt);
322 newport = svc_xprt_local_port(newxprt);
326 spin_unlock(&svc_xprt_class_lock);
327 /* This errno is exposed to user space. Provide a reasonable
328 * perror msg for a bad transport. */
329 return -EPROTONOSUPPORT;
333 * svc_xprt_create - Add a new listener to @serv
334 * @serv: target RPC service
335 * @xprt_name: transport class name
336 * @net: network namespace
337 * @family: network address family
338 * @port: listener port
339 * @flags: SVC_SOCK flags
340 * @cred: credential to bind to this transport
343 * %0: New listener added successfully
344 * %-EPROTONOSUPPORT: Requested transport type not supported
346 int svc_xprt_create(struct svc_serv *serv, const char *xprt_name,
347 struct net *net, const int family,
348 const unsigned short port, int flags,
349 const struct cred *cred)
353 err = _svc_xprt_create(serv, xprt_name, net, family, port, flags, cred);
354 if (err == -EPROTONOSUPPORT) {
355 request_module("svc%s", xprt_name);
356 err = _svc_xprt_create(serv, xprt_name, net, family, port, flags, cred);
360 EXPORT_SYMBOL_GPL(svc_xprt_create);
363 * Copy the local and remote xprt addresses to the rqstp structure
365 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
367 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
368 rqstp->rq_addrlen = xprt->xpt_remotelen;
371 * Destination address in request is needed for binding the
372 * source address in RPC replies/callbacks later.
374 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
375 rqstp->rq_daddrlen = xprt->xpt_locallen;
377 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
380 * svc_print_addr - Format rq_addr field for printing
381 * @rqstp: svc_rqst struct containing address to print
382 * @buf: target buffer for formatted address
383 * @len: length of target buffer
386 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
388 return __svc_print_addr(svc_addr(rqstp), buf, len);
390 EXPORT_SYMBOL_GPL(svc_print_addr);
392 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
394 unsigned int limit = svc_rpc_per_connection_limit;
395 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
397 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
400 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
402 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
403 if (!svc_xprt_slots_in_range(xprt))
405 atomic_inc(&xprt->xpt_nr_rqsts);
406 set_bit(RQ_DATA, &rqstp->rq_flags);
411 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
413 struct svc_xprt *xprt = rqstp->rq_xprt;
414 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
415 atomic_dec(&xprt->xpt_nr_rqsts);
416 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
417 svc_xprt_enqueue(xprt);
421 static bool svc_xprt_ready(struct svc_xprt *xprt)
423 unsigned long xpt_flags;
426 * If another cpu has recently updated xpt_flags,
427 * sk_sock->flags, xpt_reserved, or xpt_nr_rqsts, we need to
428 * know about it; otherwise it's possible that both that cpu and
429 * this one could call svc_xprt_enqueue() without either
430 * svc_xprt_enqueue() recognizing that the conditions below
431 * are satisfied, and we could stall indefinitely:
434 xpt_flags = READ_ONCE(xprt->xpt_flags);
436 trace_svc_xprt_enqueue(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;
460 if (!svc_xprt_ready(xprt))
463 /* Mark transport as busy. It will remain in this state until
464 * the provider calls svc_xprt_received. We update XPT_BUSY
465 * atomically because it also guards against trying to enqueue
466 * the transport twice.
468 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
471 pool = svc_pool_for_cpu(xprt->xpt_server);
473 percpu_counter_inc(&pool->sp_sockets_queued);
474 lwq_enqueue(&xprt->xpt_ready, &pool->sp_xprts);
476 svc_pool_wake_idle_thread(pool);
478 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
481 * Dequeue the first transport, if there is one.
483 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
485 struct svc_xprt *xprt = NULL;
487 xprt = lwq_dequeue(&pool->sp_xprts, struct svc_xprt, xpt_ready);
494 * svc_reserve - change the space reserved for the reply to a request.
495 * @rqstp: The request in question
496 * @space: new max space to reserve
498 * Each request reserves some space on the output queue of the transport
499 * to make sure the reply fits. This function reduces that reserved
500 * space to be the amount of space used already, plus @space.
503 void svc_reserve(struct svc_rqst *rqstp, int space)
505 struct svc_xprt *xprt = rqstp->rq_xprt;
507 space += rqstp->rq_res.head[0].iov_len;
509 if (xprt && space < rqstp->rq_reserved) {
510 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
511 rqstp->rq_reserved = space;
512 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
513 svc_xprt_enqueue(xprt);
516 EXPORT_SYMBOL_GPL(svc_reserve);
518 static void free_deferred(struct svc_xprt *xprt, struct svc_deferred_req *dr)
523 xprt->xpt_ops->xpo_release_ctxt(xprt, dr->xprt_ctxt);
527 static void svc_xprt_release(struct svc_rqst *rqstp)
529 struct svc_xprt *xprt = rqstp->rq_xprt;
531 xprt->xpt_ops->xpo_release_ctxt(xprt, rqstp->rq_xprt_ctxt);
532 rqstp->rq_xprt_ctxt = NULL;
534 free_deferred(xprt, rqstp->rq_deferred);
535 rqstp->rq_deferred = NULL;
537 svc_rqst_release_pages(rqstp);
538 rqstp->rq_res.page_len = 0;
539 rqstp->rq_res.page_base = 0;
541 /* Reset response buffer and release
543 * But first, check that enough space was reserved
544 * for the reply, otherwise we have a bug!
546 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
547 printk(KERN_ERR "RPC request reserved %d but used %d\n",
551 rqstp->rq_res.head[0].iov_len = 0;
552 svc_reserve(rqstp, 0);
553 svc_xprt_release_slot(rqstp);
554 rqstp->rq_xprt = NULL;
559 * svc_wake_up - Wake up a service thread for non-transport work
562 * Some svc_serv's will have occasional work to do, even when a xprt is not
563 * waiting to be serviced. This function is there to "kick" a task in one of
564 * those services so that it can wake up and do that work. Note that we only
565 * bother with pool 0 as we don't need to wake up more than one thread for
568 void svc_wake_up(struct svc_serv *serv)
570 struct svc_pool *pool = &serv->sv_pools[0];
572 set_bit(SP_TASK_PENDING, &pool->sp_flags);
573 svc_pool_wake_idle_thread(pool);
575 EXPORT_SYMBOL_GPL(svc_wake_up);
577 int svc_port_is_privileged(struct sockaddr *sin)
579 switch (sin->sa_family) {
581 return ntohs(((struct sockaddr_in *)sin)->sin_port)
584 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
592 * Make sure that we don't have too many active connections. If we have,
593 * something must be dropped. It's not clear what will happen if we allow
594 * "too many" connections, but when dealing with network-facing software,
595 * we have to code defensively. Here we do that by imposing hard limits.
597 * There's no point in trying to do random drop here for DoS
598 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
599 * attacker can easily beat that.
601 * The only somewhat efficient mechanism would be if drop old
602 * connections from the same IP first. But right now we don't even
603 * record the client IP in svc_sock.
605 * single-threaded services that expect a lot of clients will probably
606 * need to set sv_maxconn to override the default value which is based
607 * on the number of threads
609 static void svc_check_conn_limits(struct svc_serv *serv)
611 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
612 (serv->sv_nrthreads+3) * 20;
614 if (serv->sv_tmpcnt > limit) {
615 struct svc_xprt *xprt = NULL;
616 spin_lock_bh(&serv->sv_lock);
617 if (!list_empty(&serv->sv_tempsocks)) {
618 /* Try to help the admin */
619 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
620 serv->sv_name, serv->sv_maxconn ?
621 "max number of connections" :
622 "number of threads");
624 * Always select the oldest connection. It's not fair,
627 xprt = list_entry(serv->sv_tempsocks.prev,
630 set_bit(XPT_CLOSE, &xprt->xpt_flags);
633 spin_unlock_bh(&serv->sv_lock);
636 svc_xprt_enqueue(xprt);
642 static bool svc_alloc_arg(struct svc_rqst *rqstp)
644 struct svc_serv *serv = rqstp->rq_server;
645 struct xdr_buf *arg = &rqstp->rq_arg;
646 unsigned long pages, filled, ret;
648 pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
649 if (pages > RPCSVC_MAXPAGES) {
650 pr_warn_once("svc: warning: pages=%lu > RPCSVC_MAXPAGES=%lu\n",
651 pages, RPCSVC_MAXPAGES);
652 /* use as many pages as possible */
653 pages = RPCSVC_MAXPAGES;
656 for (filled = 0; filled < pages; filled = ret) {
657 ret = alloc_pages_bulk_array(GFP_KERNEL, pages,
660 /* Made progress, don't sleep yet */
663 set_current_state(TASK_IDLE);
664 if (svc_thread_should_stop(rqstp)) {
665 set_current_state(TASK_RUNNING);
668 trace_svc_alloc_arg_err(pages, ret);
669 memalloc_retry_wait(GFP_KERNEL);
671 rqstp->rq_page_end = &rqstp->rq_pages[pages];
672 rqstp->rq_pages[pages] = NULL; /* this might be seen in nfsd_splice_actor() */
674 /* Make arg->head point to first page and arg->pages point to rest */
675 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
676 arg->head[0].iov_len = PAGE_SIZE;
677 arg->pages = rqstp->rq_pages + 1;
679 /* save at least one page for response */
680 arg->page_len = (pages-2)*PAGE_SIZE;
681 arg->len = (pages-1)*PAGE_SIZE;
682 arg->tail[0].iov_len = 0;
684 rqstp->rq_xid = xdr_zero;
689 svc_thread_should_sleep(struct svc_rqst *rqstp)
691 struct svc_pool *pool = rqstp->rq_pool;
693 /* did someone call svc_wake_up? */
694 if (test_bit(SP_TASK_PENDING, &pool->sp_flags))
697 /* was a socket queued? */
698 if (!lwq_empty(&pool->sp_xprts))
701 /* are we shutting down? */
702 if (svc_thread_should_stop(rqstp))
705 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
706 if (svc_is_backchannel(rqstp)) {
707 if (!lwq_empty(&rqstp->rq_server->sv_cb_list))
715 static void svc_thread_wait_for_work(struct svc_rqst *rqstp)
717 struct svc_pool *pool = rqstp->rq_pool;
719 if (svc_thread_should_sleep(rqstp)) {
720 set_current_state(TASK_IDLE | TASK_FREEZABLE);
721 llist_add(&rqstp->rq_idle, &pool->sp_idle_threads);
722 if (likely(svc_thread_should_sleep(rqstp)))
725 while (!llist_del_first_this(&pool->sp_idle_threads,
727 /* Work just became available. This thread can only
728 * handle it after removing rqstp from the idle
729 * list. If that attempt failed, some other thread
730 * must have queued itself after finding no
731 * work to do, so that thread has taken responsibly
732 * for this new work. This thread can safely sleep
736 set_current_state(TASK_IDLE | TASK_FREEZABLE);
738 __set_current_state(TASK_RUNNING);
745 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
747 spin_lock_bh(&serv->sv_lock);
748 set_bit(XPT_TEMP, &newxpt->xpt_flags);
749 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
751 if (serv->sv_temptimer.function == NULL) {
752 /* setup timer to age temp transports */
753 serv->sv_temptimer.function = svc_age_temp_xprts;
754 mod_timer(&serv->sv_temptimer,
755 jiffies + svc_conn_age_period * HZ);
757 spin_unlock_bh(&serv->sv_lock);
758 svc_xprt_received(newxpt);
761 static void svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
763 struct svc_serv *serv = rqstp->rq_server;
766 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
767 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
768 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
769 svc_delete_xprt(xprt);
770 /* Leave XPT_BUSY set on the dead xprt: */
773 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
774 struct svc_xprt *newxpt;
776 * We know this module_get will succeed because the
777 * listener holds a reference too
779 __module_get(xprt->xpt_class->xcl_owner);
780 svc_check_conn_limits(xprt->xpt_server);
781 newxpt = xprt->xpt_ops->xpo_accept(xprt);
783 newxpt->xpt_cred = get_cred(xprt->xpt_cred);
784 svc_add_new_temp_xprt(serv, newxpt);
785 trace_svc_xprt_accept(newxpt, serv->sv_name);
787 module_put(xprt->xpt_class->xcl_owner);
789 svc_xprt_received(xprt);
790 } else if (test_bit(XPT_HANDSHAKE, &xprt->xpt_flags)) {
791 xprt->xpt_ops->xpo_handshake(xprt);
792 svc_xprt_received(xprt);
793 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
794 /* XPT_DATA|XPT_DEFERRED case: */
795 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
796 if (rqstp->rq_deferred)
797 len = svc_deferred_recv(rqstp);
799 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
800 rqstp->rq_reserved = serv->sv_max_mesg;
801 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
805 trace_svc_xdr_recvfrom(&rqstp->rq_arg);
807 clear_bit(XPT_OLD, &xprt->xpt_flags);
809 rqstp->rq_chandle.defer = svc_defer;
812 serv->sv_stats->netcnt++;
813 percpu_counter_inc(&rqstp->rq_pool->sp_messages_arrived);
814 rqstp->rq_stime = ktime_get();
817 svc_xprt_received(xprt);
820 rqstp->rq_res.len = 0;
821 svc_xprt_release(rqstp);
824 static void svc_thread_wake_next(struct svc_rqst *rqstp)
826 if (!svc_thread_should_sleep(rqstp))
827 /* More work pending after I dequeued some,
828 * wake another worker
830 svc_pool_wake_idle_thread(rqstp->rq_pool);
834 * svc_recv - Receive and process the next request on any transport
835 * @rqstp: an idle RPC service thread
837 * This code is carefully organised not to touch any cachelines in
838 * the shared svc_serv structure, only cachelines in the local
841 void svc_recv(struct svc_rqst *rqstp)
843 struct svc_pool *pool = rqstp->rq_pool;
845 if (!svc_alloc_arg(rqstp))
848 svc_thread_wait_for_work(rqstp);
850 clear_bit(SP_TASK_PENDING, &pool->sp_flags);
852 if (svc_thread_should_stop(rqstp)) {
853 svc_thread_wake_next(rqstp);
857 rqstp->rq_xprt = svc_xprt_dequeue(pool);
858 if (rqstp->rq_xprt) {
859 struct svc_xprt *xprt = rqstp->rq_xprt;
861 svc_thread_wake_next(rqstp);
862 /* Normally we will wait up to 5 seconds for any required
863 * cache information to be provided. When there are no
864 * idle threads, we reduce the wait time.
866 if (pool->sp_idle_threads.first)
867 rqstp->rq_chandle.thread_wait = 5 * HZ;
869 rqstp->rq_chandle.thread_wait = 1 * HZ;
871 trace_svc_xprt_dequeue(rqstp);
872 svc_handle_xprt(rqstp, xprt);
875 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
876 if (svc_is_backchannel(rqstp)) {
877 struct svc_serv *serv = rqstp->rq_server;
878 struct rpc_rqst *req;
880 req = lwq_dequeue(&serv->sv_cb_list,
881 struct rpc_rqst, rq_bc_list);
883 svc_thread_wake_next(rqstp);
884 svc_process_bc(req, rqstp);
889 EXPORT_SYMBOL_GPL(svc_recv);
894 void svc_drop(struct svc_rqst *rqstp)
896 trace_svc_drop(rqstp);
898 EXPORT_SYMBOL_GPL(svc_drop);
901 * svc_send - Return reply to client
902 * @rqstp: RPC transaction context
905 void svc_send(struct svc_rqst *rqstp)
907 struct svc_xprt *xprt;
911 xprt = rqstp->rq_xprt;
913 /* calculate over-all length */
915 xb->len = xb->head[0].iov_len +
918 trace_svc_xdr_sendto(rqstp->rq_xid, xb);
919 trace_svc_stats_latency(rqstp);
921 status = xprt->xpt_ops->xpo_sendto(rqstp);
923 trace_svc_send(rqstp, status);
927 * Timer function to close old temporary transports, using
928 * a mark-and-sweep algorithm.
930 static void svc_age_temp_xprts(struct timer_list *t)
932 struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
933 struct svc_xprt *xprt;
934 struct list_head *le, *next;
936 dprintk("svc_age_temp_xprts\n");
938 if (!spin_trylock_bh(&serv->sv_lock)) {
939 /* busy, try again 1 sec later */
940 dprintk("svc_age_temp_xprts: busy\n");
941 mod_timer(&serv->sv_temptimer, jiffies + HZ);
945 list_for_each_safe(le, next, &serv->sv_tempsocks) {
946 xprt = list_entry(le, struct svc_xprt, xpt_list);
948 /* First time through, just mark it OLD. Second time
949 * through, close it. */
950 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
952 if (kref_read(&xprt->xpt_ref) > 1 ||
953 test_bit(XPT_BUSY, &xprt->xpt_flags))
956 set_bit(XPT_CLOSE, &xprt->xpt_flags);
957 dprintk("queuing xprt %p for closing\n", xprt);
959 /* a thread will dequeue and close it soon */
960 svc_xprt_enqueue(xprt);
962 spin_unlock_bh(&serv->sv_lock);
964 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
967 /* Close temporary transports whose xpt_local matches server_addr immediately
968 * instead of waiting for them to be picked up by the timer.
970 * This is meant to be called from a notifier_block that runs when an ip
971 * address is deleted.
973 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
975 struct svc_xprt *xprt;
976 struct list_head *le, *next;
977 LIST_HEAD(to_be_closed);
979 spin_lock_bh(&serv->sv_lock);
980 list_for_each_safe(le, next, &serv->sv_tempsocks) {
981 xprt = list_entry(le, struct svc_xprt, xpt_list);
982 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
984 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
985 list_move(le, &to_be_closed);
988 spin_unlock_bh(&serv->sv_lock);
990 while (!list_empty(&to_be_closed)) {
991 le = to_be_closed.next;
993 xprt = list_entry(le, struct svc_xprt, xpt_list);
994 set_bit(XPT_CLOSE, &xprt->xpt_flags);
995 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
996 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
998 svc_xprt_enqueue(xprt);
1001 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1003 static void call_xpt_users(struct svc_xprt *xprt)
1005 struct svc_xpt_user *u;
1007 spin_lock(&xprt->xpt_lock);
1008 while (!list_empty(&xprt->xpt_users)) {
1009 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1010 list_del_init(&u->list);
1013 spin_unlock(&xprt->xpt_lock);
1017 * Remove a dead transport
1019 static void svc_delete_xprt(struct svc_xprt *xprt)
1021 struct svc_serv *serv = xprt->xpt_server;
1022 struct svc_deferred_req *dr;
1024 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1027 trace_svc_xprt_detach(xprt);
1028 xprt->xpt_ops->xpo_detach(xprt);
1029 if (xprt->xpt_bc_xprt)
1030 xprt->xpt_bc_xprt->ops->close(xprt->xpt_bc_xprt);
1032 spin_lock_bh(&serv->sv_lock);
1033 list_del_init(&xprt->xpt_list);
1034 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1036 spin_unlock_bh(&serv->sv_lock);
1038 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1039 free_deferred(xprt, dr);
1041 call_xpt_users(xprt);
1046 * svc_xprt_close - Close a client connection
1047 * @xprt: transport to disconnect
1050 void svc_xprt_close(struct svc_xprt *xprt)
1052 trace_svc_xprt_close(xprt);
1053 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1054 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1055 /* someone else will have to effect the close */
1058 * We expect svc_close_xprt() to work even when no threads are
1059 * running (e.g., while configuring the server before starting
1060 * any threads), so if the transport isn't busy, we delete
1063 svc_delete_xprt(xprt);
1065 EXPORT_SYMBOL_GPL(svc_xprt_close);
1067 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1069 struct svc_xprt *xprt;
1072 spin_lock_bh(&serv->sv_lock);
1073 list_for_each_entry(xprt, xprt_list, xpt_list) {
1074 if (xprt->xpt_net != net)
1077 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1078 svc_xprt_enqueue(xprt);
1080 spin_unlock_bh(&serv->sv_lock);
1084 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1086 struct svc_xprt *xprt;
1089 for (i = 0; i < serv->sv_nrpools; i++) {
1090 struct svc_pool *pool = &serv->sv_pools[i];
1091 struct llist_node *q, **t1, *t2;
1093 q = lwq_dequeue_all(&pool->sp_xprts);
1094 lwq_for_each_safe(xprt, t1, t2, &q, xpt_ready) {
1095 if (xprt->xpt_net == net) {
1096 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1097 svc_delete_xprt(xprt);
1103 lwq_enqueue_batch(q, &pool->sp_xprts);
1108 * svc_xprt_destroy_all - Destroy transports associated with @serv
1109 * @serv: RPC service to be shut down
1110 * @net: target network namespace
1112 * Server threads may still be running (especially in the case where the
1113 * service is still running in other network namespaces).
1115 * So we shut down sockets the same way we would on a running server, by
1116 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1117 * the close. In the case there are no such other threads,
1118 * threads running, svc_clean_up_xprts() does a simple version of a
1119 * server's main event loop, and in the case where there are other
1120 * threads, we may need to wait a little while and then check again to
1121 * see if they're done.
1123 void svc_xprt_destroy_all(struct svc_serv *serv, struct net *net)
1127 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1128 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1130 svc_clean_up_xprts(serv, net);
1134 EXPORT_SYMBOL_GPL(svc_xprt_destroy_all);
1137 * Handle defer and revisit of requests
1140 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1142 struct svc_deferred_req *dr =
1143 container_of(dreq, struct svc_deferred_req, handle);
1144 struct svc_xprt *xprt = dr->xprt;
1146 spin_lock(&xprt->xpt_lock);
1147 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1148 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1149 spin_unlock(&xprt->xpt_lock);
1150 trace_svc_defer_drop(dr);
1151 free_deferred(xprt, dr);
1156 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1157 spin_unlock(&xprt->xpt_lock);
1158 trace_svc_defer_queue(dr);
1159 svc_xprt_enqueue(xprt);
1164 * Save the request off for later processing. The request buffer looks
1167 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1169 * This code can only handle requests that consist of an xprt-header
1172 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1174 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1175 struct svc_deferred_req *dr;
1177 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1178 return NULL; /* if more than a page, give up FIXME */
1179 if (rqstp->rq_deferred) {
1180 dr = rqstp->rq_deferred;
1181 rqstp->rq_deferred = NULL;
1185 /* FIXME maybe discard if size too large */
1186 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1187 dr = kmalloc(size, GFP_KERNEL);
1191 dr->handle.owner = rqstp->rq_server;
1192 dr->prot = rqstp->rq_prot;
1193 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1194 dr->addrlen = rqstp->rq_addrlen;
1195 dr->daddr = rqstp->rq_daddr;
1196 dr->argslen = rqstp->rq_arg.len >> 2;
1198 /* back up head to the start of the buffer and copy */
1199 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1200 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1203 dr->xprt_ctxt = rqstp->rq_xprt_ctxt;
1204 rqstp->rq_xprt_ctxt = NULL;
1205 trace_svc_defer(rqstp);
1206 svc_xprt_get(rqstp->rq_xprt);
1207 dr->xprt = rqstp->rq_xprt;
1208 set_bit(RQ_DROPME, &rqstp->rq_flags);
1210 dr->handle.revisit = svc_revisit;
1215 * recv data from a deferred request into an active one
1217 static noinline int svc_deferred_recv(struct svc_rqst *rqstp)
1219 struct svc_deferred_req *dr = rqstp->rq_deferred;
1221 trace_svc_defer_recv(dr);
1223 /* setup iov_base past transport header */
1224 rqstp->rq_arg.head[0].iov_base = dr->args;
1225 /* The iov_len does not include the transport header bytes */
1226 rqstp->rq_arg.head[0].iov_len = dr->argslen << 2;
1227 rqstp->rq_arg.page_len = 0;
1228 /* The rq_arg.len includes the transport header bytes */
1229 rqstp->rq_arg.len = dr->argslen << 2;
1230 rqstp->rq_prot = dr->prot;
1231 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1232 rqstp->rq_addrlen = dr->addrlen;
1233 /* Save off transport header len in case we get deferred again */
1234 rqstp->rq_daddr = dr->daddr;
1235 rqstp->rq_respages = rqstp->rq_pages;
1236 rqstp->rq_xprt_ctxt = dr->xprt_ctxt;
1238 dr->xprt_ctxt = NULL;
1239 svc_xprt_received(rqstp->rq_xprt);
1240 return dr->argslen << 2;
1244 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1246 struct svc_deferred_req *dr = NULL;
1248 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1250 spin_lock(&xprt->xpt_lock);
1251 if (!list_empty(&xprt->xpt_deferred)) {
1252 dr = list_entry(xprt->xpt_deferred.next,
1253 struct svc_deferred_req,
1255 list_del_init(&dr->handle.recent);
1257 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1258 spin_unlock(&xprt->xpt_lock);
1263 * svc_find_xprt - find an RPC transport instance
1264 * @serv: pointer to svc_serv to search
1265 * @xcl_name: C string containing transport's class name
1266 * @net: owner net pointer
1267 * @af: Address family of transport's local address
1268 * @port: transport's IP port number
1270 * Return the transport instance pointer for the endpoint accepting
1271 * connections/peer traffic from the specified transport class,
1272 * address family and port.
1274 * Specifying 0 for the address family or port is effectively a
1275 * wild-card, and will result in matching the first transport in the
1276 * service's list that has a matching class name.
1278 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1279 struct net *net, const sa_family_t af,
1280 const unsigned short port)
1282 struct svc_xprt *xprt;
1283 struct svc_xprt *found = NULL;
1285 /* Sanity check the args */
1286 if (serv == NULL || xcl_name == NULL)
1289 spin_lock_bh(&serv->sv_lock);
1290 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1291 if (xprt->xpt_net != net)
1293 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1295 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1297 if (port != 0 && port != svc_xprt_local_port(xprt))
1303 spin_unlock_bh(&serv->sv_lock);
1306 EXPORT_SYMBOL_GPL(svc_find_xprt);
1308 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1309 char *pos, int remaining)
1313 len = snprintf(pos, remaining, "%s %u\n",
1314 xprt->xpt_class->xcl_name,
1315 svc_xprt_local_port(xprt));
1316 if (len >= remaining)
1317 return -ENAMETOOLONG;
1322 * svc_xprt_names - format a buffer with a list of transport names
1323 * @serv: pointer to an RPC service
1324 * @buf: pointer to a buffer to be filled in
1325 * @buflen: length of buffer to be filled in
1327 * Fills in @buf with a string containing a list of transport names,
1328 * each name terminated with '\n'.
1330 * Returns positive length of the filled-in string on success; otherwise
1331 * a negative errno value is returned if an error occurs.
1333 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1335 struct svc_xprt *xprt;
1339 /* Sanity check args */
1343 spin_lock_bh(&serv->sv_lock);
1347 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1348 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1360 spin_unlock_bh(&serv->sv_lock);
1363 EXPORT_SYMBOL_GPL(svc_xprt_names);
1365 /*----------------------------------------------------------------------------*/
1367 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1369 unsigned int pidx = (unsigned int)*pos;
1370 struct svc_info *si = m->private;
1372 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1374 mutex_lock(si->mutex);
1377 return SEQ_START_TOKEN;
1380 return pidx > si->serv->sv_nrpools ? NULL
1381 : &si->serv->sv_pools[pidx - 1];
1384 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1386 struct svc_pool *pool = p;
1387 struct svc_info *si = m->private;
1388 struct svc_serv *serv = si->serv;
1390 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1394 } else if (p == SEQ_START_TOKEN) {
1395 pool = &serv->sv_pools[0];
1397 unsigned int pidx = (pool - &serv->sv_pools[0]);
1398 if (pidx < serv->sv_nrpools-1)
1399 pool = &serv->sv_pools[pidx+1];
1407 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1409 struct svc_info *si = m->private;
1411 mutex_unlock(si->mutex);
1414 static int svc_pool_stats_show(struct seq_file *m, void *p)
1416 struct svc_pool *pool = p;
1418 if (p == SEQ_START_TOKEN) {
1419 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1423 seq_printf(m, "%u %llu %llu %llu 0\n",
1425 percpu_counter_sum_positive(&pool->sp_messages_arrived),
1426 percpu_counter_sum_positive(&pool->sp_sockets_queued),
1427 percpu_counter_sum_positive(&pool->sp_threads_woken));
1432 static const struct seq_operations svc_pool_stats_seq_ops = {
1433 .start = svc_pool_stats_start,
1434 .next = svc_pool_stats_next,
1435 .stop = svc_pool_stats_stop,
1436 .show = svc_pool_stats_show,
1439 int svc_pool_stats_open(struct svc_info *info, struct file *file)
1441 struct seq_file *seq;
1444 err = seq_open(file, &svc_pool_stats_seq_ops);
1447 seq = file->private_data;
1448 seq->private = info;
1452 EXPORT_SYMBOL(svc_pool_stats_open);
1454 /*----------------------------------------------------------------------------*/