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.
77 int svc_reg_xprt_class(struct svc_xprt_class *xcl)
79 struct svc_xprt_class *cl;
82 dprintk("svc: Adding svc transport class '%s'\n", xcl->xcl_name);
84 INIT_LIST_HEAD(&xcl->xcl_list);
85 spin_lock(&svc_xprt_class_lock);
86 /* Make sure there isn't already a class with the same name */
87 list_for_each_entry(cl, &svc_xprt_class_list, xcl_list) {
88 if (strcmp(xcl->xcl_name, cl->xcl_name) == 0)
91 list_add_tail(&xcl->xcl_list, &svc_xprt_class_list);
94 spin_unlock(&svc_xprt_class_lock);
97 EXPORT_SYMBOL_GPL(svc_reg_xprt_class);
99 void svc_unreg_xprt_class(struct svc_xprt_class *xcl)
101 dprintk("svc: Removing svc transport class '%s'\n", xcl->xcl_name);
102 spin_lock(&svc_xprt_class_lock);
103 list_del_init(&xcl->xcl_list);
104 spin_unlock(&svc_xprt_class_lock);
106 EXPORT_SYMBOL_GPL(svc_unreg_xprt_class);
109 * svc_print_xprts - Format the transport list for printing
110 * @buf: target buffer for formatted address
111 * @maxlen: length of target buffer
113 * Fills in @buf with a string containing a list of transport names, each name
114 * terminated with '\n'. If the buffer is too small, some entries may be
115 * missing, but it is guaranteed that all lines in the output buffer are
118 * Returns positive length of the filled-in string.
120 int svc_print_xprts(char *buf, int maxlen)
122 struct svc_xprt_class *xcl;
127 spin_lock(&svc_xprt_class_lock);
128 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
131 slen = snprintf(tmpstr, sizeof(tmpstr), "%s %d\n",
132 xcl->xcl_name, xcl->xcl_max_payload);
133 if (slen >= sizeof(tmpstr) || len + slen >= maxlen)
138 spin_unlock(&svc_xprt_class_lock);
144 * svc_xprt_deferred_close - Close a transport
145 * @xprt: transport instance
147 * Used in contexts that need to defer the work of shutting down
148 * the transport to an nfsd thread.
150 void svc_xprt_deferred_close(struct svc_xprt *xprt)
152 if (!test_and_set_bit(XPT_CLOSE, &xprt->xpt_flags))
153 svc_xprt_enqueue(xprt);
155 EXPORT_SYMBOL_GPL(svc_xprt_deferred_close);
157 static void svc_xprt_free(struct kref *kref)
159 struct svc_xprt *xprt =
160 container_of(kref, struct svc_xprt, xpt_ref);
161 struct module *owner = xprt->xpt_class->xcl_owner;
162 if (test_bit(XPT_CACHE_AUTH, &xprt->xpt_flags))
163 svcauth_unix_info_release(xprt);
164 put_cred(xprt->xpt_cred);
165 put_net(xprt->xpt_net);
166 /* See comment on corresponding get in xs_setup_bc_tcp(): */
167 if (xprt->xpt_bc_xprt)
168 xprt_put(xprt->xpt_bc_xprt);
169 if (xprt->xpt_bc_xps)
170 xprt_switch_put(xprt->xpt_bc_xps);
171 trace_svc_xprt_free(xprt);
172 xprt->xpt_ops->xpo_free(xprt);
176 void svc_xprt_put(struct svc_xprt *xprt)
178 kref_put(&xprt->xpt_ref, svc_xprt_free);
180 EXPORT_SYMBOL_GPL(svc_xprt_put);
183 * Called by transport drivers to initialize the transport independent
184 * portion of the transport instance.
186 void svc_xprt_init(struct net *net, struct svc_xprt_class *xcl,
187 struct svc_xprt *xprt, struct svc_serv *serv)
189 memset(xprt, 0, sizeof(*xprt));
190 xprt->xpt_class = xcl;
191 xprt->xpt_ops = xcl->xcl_ops;
192 kref_init(&xprt->xpt_ref);
193 xprt->xpt_server = serv;
194 INIT_LIST_HEAD(&xprt->xpt_list);
195 INIT_LIST_HEAD(&xprt->xpt_ready);
196 INIT_LIST_HEAD(&xprt->xpt_deferred);
197 INIT_LIST_HEAD(&xprt->xpt_users);
198 mutex_init(&xprt->xpt_mutex);
199 spin_lock_init(&xprt->xpt_lock);
200 set_bit(XPT_BUSY, &xprt->xpt_flags);
201 xprt->xpt_net = get_net(net);
202 strcpy(xprt->xpt_remotebuf, "uninitialized");
204 EXPORT_SYMBOL_GPL(svc_xprt_init);
206 static struct svc_xprt *__svc_xpo_create(struct svc_xprt_class *xcl,
207 struct svc_serv *serv,
210 const unsigned short port,
213 struct sockaddr_in sin = {
214 .sin_family = AF_INET,
215 .sin_addr.s_addr = htonl(INADDR_ANY),
216 .sin_port = htons(port),
218 #if IS_ENABLED(CONFIG_IPV6)
219 struct sockaddr_in6 sin6 = {
220 .sin6_family = AF_INET6,
221 .sin6_addr = IN6ADDR_ANY_INIT,
222 .sin6_port = htons(port),
225 struct svc_xprt *xprt;
226 struct sockaddr *sap;
231 sap = (struct sockaddr *)&sin;
234 #if IS_ENABLED(CONFIG_IPV6)
236 sap = (struct sockaddr *)&sin6;
241 return ERR_PTR(-EAFNOSUPPORT);
244 xprt = xcl->xcl_ops->xpo_create(serv, net, sap, len, flags);
246 trace_svc_xprt_create_err(serv->sv_program->pg_name,
247 xcl->xcl_name, sap, xprt);
252 * svc_xprt_received - start next receiver thread
253 * @xprt: controlling transport
255 * The caller must hold the XPT_BUSY bit and must
256 * not thereafter touch transport data.
258 * Note: XPT_DATA only gets cleared when a read-attempt finds no (or
259 * insufficient) data.
261 void svc_xprt_received(struct svc_xprt *xprt)
263 if (!test_bit(XPT_BUSY, &xprt->xpt_flags)) {
264 WARN_ONCE(1, "xprt=0x%p already busy!", xprt);
268 trace_svc_xprt_received(xprt);
270 /* As soon as we clear busy, the xprt could be closed and
271 * 'put', so we need a reference to call svc_enqueue_xprt with:
274 smp_mb__before_atomic();
275 clear_bit(XPT_BUSY, &xprt->xpt_flags);
276 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
279 EXPORT_SYMBOL_GPL(svc_xprt_received);
281 void svc_add_new_perm_xprt(struct svc_serv *serv, struct svc_xprt *new)
283 clear_bit(XPT_TEMP, &new->xpt_flags);
284 spin_lock_bh(&serv->sv_lock);
285 list_add(&new->xpt_list, &serv->sv_permsocks);
286 spin_unlock_bh(&serv->sv_lock);
287 svc_xprt_received(new);
290 static int _svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
291 struct net *net, const int family,
292 const unsigned short port, int flags,
293 const struct cred *cred)
295 struct svc_xprt_class *xcl;
297 spin_lock(&svc_xprt_class_lock);
298 list_for_each_entry(xcl, &svc_xprt_class_list, xcl_list) {
299 struct svc_xprt *newxprt;
300 unsigned short newport;
302 if (strcmp(xprt_name, xcl->xcl_name))
305 if (!try_module_get(xcl->xcl_owner))
308 spin_unlock(&svc_xprt_class_lock);
309 newxprt = __svc_xpo_create(xcl, serv, net, family, port, flags);
310 if (IS_ERR(newxprt)) {
311 module_put(xcl->xcl_owner);
312 return PTR_ERR(newxprt);
314 newxprt->xpt_cred = get_cred(cred);
315 svc_add_new_perm_xprt(serv, newxprt);
316 newport = svc_xprt_local_port(newxprt);
320 spin_unlock(&svc_xprt_class_lock);
321 /* This errno is exposed to user space. Provide a reasonable
322 * perror msg for a bad transport. */
323 return -EPROTONOSUPPORT;
326 int svc_create_xprt(struct svc_serv *serv, const char *xprt_name,
327 struct net *net, const int family,
328 const unsigned short port, int flags,
329 const struct cred *cred)
333 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags, cred);
334 if (err == -EPROTONOSUPPORT) {
335 request_module("svc%s", xprt_name);
336 err = _svc_create_xprt(serv, xprt_name, net, family, port, flags, cred);
340 EXPORT_SYMBOL_GPL(svc_create_xprt);
343 * Copy the local and remote xprt addresses to the rqstp structure
345 void svc_xprt_copy_addrs(struct svc_rqst *rqstp, struct svc_xprt *xprt)
347 memcpy(&rqstp->rq_addr, &xprt->xpt_remote, xprt->xpt_remotelen);
348 rqstp->rq_addrlen = xprt->xpt_remotelen;
351 * Destination address in request is needed for binding the
352 * source address in RPC replies/callbacks later.
354 memcpy(&rqstp->rq_daddr, &xprt->xpt_local, xprt->xpt_locallen);
355 rqstp->rq_daddrlen = xprt->xpt_locallen;
357 EXPORT_SYMBOL_GPL(svc_xprt_copy_addrs);
360 * svc_print_addr - Format rq_addr field for printing
361 * @rqstp: svc_rqst struct containing address to print
362 * @buf: target buffer for formatted address
363 * @len: length of target buffer
366 char *svc_print_addr(struct svc_rqst *rqstp, char *buf, size_t len)
368 return __svc_print_addr(svc_addr(rqstp), buf, len);
370 EXPORT_SYMBOL_GPL(svc_print_addr);
372 static bool svc_xprt_slots_in_range(struct svc_xprt *xprt)
374 unsigned int limit = svc_rpc_per_connection_limit;
375 int nrqsts = atomic_read(&xprt->xpt_nr_rqsts);
377 return limit == 0 || (nrqsts >= 0 && nrqsts < limit);
380 static bool svc_xprt_reserve_slot(struct svc_rqst *rqstp, struct svc_xprt *xprt)
382 if (!test_bit(RQ_DATA, &rqstp->rq_flags)) {
383 if (!svc_xprt_slots_in_range(xprt))
385 atomic_inc(&xprt->xpt_nr_rqsts);
386 set_bit(RQ_DATA, &rqstp->rq_flags);
391 static void svc_xprt_release_slot(struct svc_rqst *rqstp)
393 struct svc_xprt *xprt = rqstp->rq_xprt;
394 if (test_and_clear_bit(RQ_DATA, &rqstp->rq_flags)) {
395 atomic_dec(&xprt->xpt_nr_rqsts);
396 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
397 svc_xprt_enqueue(xprt);
401 static bool svc_xprt_ready(struct svc_xprt *xprt)
403 unsigned long xpt_flags;
406 * If another cpu has recently updated xpt_flags,
407 * sk_sock->flags, xpt_reserved, or xpt_nr_rqsts, we need to
408 * know about it; otherwise it's possible that both that cpu and
409 * this one could call svc_xprt_enqueue() without either
410 * svc_xprt_enqueue() recognizing that the conditions below
411 * are satisfied, and we could stall indefinitely:
414 xpt_flags = READ_ONCE(xprt->xpt_flags);
416 if (xpt_flags & (BIT(XPT_CONN) | BIT(XPT_CLOSE)))
418 if (xpt_flags & (BIT(XPT_DATA) | BIT(XPT_DEFERRED))) {
419 if (xprt->xpt_ops->xpo_has_wspace(xprt) &&
420 svc_xprt_slots_in_range(xprt))
422 trace_svc_xprt_no_write_space(xprt);
428 void svc_xprt_do_enqueue(struct svc_xprt *xprt)
430 struct svc_pool *pool;
431 struct svc_rqst *rqstp = NULL;
434 if (!svc_xprt_ready(xprt))
437 /* Mark transport as busy. It will remain in this state until
438 * the provider calls svc_xprt_received. We update XPT_BUSY
439 * atomically because it also guards against trying to enqueue
440 * the transport twice.
442 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
446 pool = svc_pool_for_cpu(xprt->xpt_server, cpu);
448 atomic_long_inc(&pool->sp_stats.packets);
450 spin_lock_bh(&pool->sp_lock);
451 list_add_tail(&xprt->xpt_ready, &pool->sp_sockets);
452 pool->sp_stats.sockets_queued++;
453 spin_unlock_bh(&pool->sp_lock);
455 /* find a thread for this xprt */
457 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
458 if (test_and_set_bit(RQ_BUSY, &rqstp->rq_flags))
460 atomic_long_inc(&pool->sp_stats.threads_woken);
461 rqstp->rq_qtime = ktime_get();
462 wake_up_process(rqstp->rq_task);
465 set_bit(SP_CONGESTED, &pool->sp_flags);
470 trace_svc_xprt_do_enqueue(xprt, rqstp);
472 EXPORT_SYMBOL_GPL(svc_xprt_do_enqueue);
475 * Queue up a transport with data pending. If there are idle nfsd
476 * processes, wake 'em up.
479 void svc_xprt_enqueue(struct svc_xprt *xprt)
481 if (test_bit(XPT_BUSY, &xprt->xpt_flags))
483 xprt->xpt_server->sv_ops->svo_enqueue_xprt(xprt);
485 EXPORT_SYMBOL_GPL(svc_xprt_enqueue);
488 * Dequeue the first transport, if there is one.
490 static struct svc_xprt *svc_xprt_dequeue(struct svc_pool *pool)
492 struct svc_xprt *xprt = NULL;
494 if (list_empty(&pool->sp_sockets))
497 spin_lock_bh(&pool->sp_lock);
498 if (likely(!list_empty(&pool->sp_sockets))) {
499 xprt = list_first_entry(&pool->sp_sockets,
500 struct svc_xprt, xpt_ready);
501 list_del_init(&xprt->xpt_ready);
504 spin_unlock_bh(&pool->sp_lock);
510 * svc_reserve - change the space reserved for the reply to a request.
511 * @rqstp: The request in question
512 * @space: new max space to reserve
514 * Each request reserves some space on the output queue of the transport
515 * to make sure the reply fits. This function reduces that reserved
516 * space to be the amount of space used already, plus @space.
519 void svc_reserve(struct svc_rqst *rqstp, int space)
521 struct svc_xprt *xprt = rqstp->rq_xprt;
523 space += rqstp->rq_res.head[0].iov_len;
525 if (xprt && space < rqstp->rq_reserved) {
526 atomic_sub((rqstp->rq_reserved - space), &xprt->xpt_reserved);
527 rqstp->rq_reserved = space;
528 smp_wmb(); /* See smp_rmb() in svc_xprt_ready() */
529 svc_xprt_enqueue(xprt);
532 EXPORT_SYMBOL_GPL(svc_reserve);
534 static void svc_xprt_release(struct svc_rqst *rqstp)
536 struct svc_xprt *xprt = rqstp->rq_xprt;
538 xprt->xpt_ops->xpo_release_rqst(rqstp);
540 kfree(rqstp->rq_deferred);
541 rqstp->rq_deferred = NULL;
543 pagevec_release(&rqstp->rq_pvec);
544 svc_free_res_pages(rqstp);
545 rqstp->rq_res.page_len = 0;
546 rqstp->rq_res.page_base = 0;
548 /* Reset response buffer and release
550 * But first, check that enough space was reserved
551 * for the reply, otherwise we have a bug!
553 if ((rqstp->rq_res.len) > rqstp->rq_reserved)
554 printk(KERN_ERR "RPC request reserved %d but used %d\n",
558 rqstp->rq_res.head[0].iov_len = 0;
559 svc_reserve(rqstp, 0);
560 svc_xprt_release_slot(rqstp);
561 rqstp->rq_xprt = NULL;
566 * Some svc_serv's will have occasional work to do, even when a xprt is not
567 * waiting to be serviced. This function is there to "kick" a task in one of
568 * those services so that it can wake up and do that work. Note that we only
569 * bother with pool 0 as we don't need to wake up more than one thread for
572 void svc_wake_up(struct svc_serv *serv)
574 struct svc_rqst *rqstp;
575 struct svc_pool *pool;
577 pool = &serv->sv_pools[0];
580 list_for_each_entry_rcu(rqstp, &pool->sp_all_threads, rq_all) {
581 /* skip any that aren't queued */
582 if (test_bit(RQ_BUSY, &rqstp->rq_flags))
585 wake_up_process(rqstp->rq_task);
586 trace_svc_wake_up(rqstp->rq_task->pid);
591 /* No free entries available */
592 set_bit(SP_TASK_PENDING, &pool->sp_flags);
594 trace_svc_wake_up(0);
596 EXPORT_SYMBOL_GPL(svc_wake_up);
598 int svc_port_is_privileged(struct sockaddr *sin)
600 switch (sin->sa_family) {
602 return ntohs(((struct sockaddr_in *)sin)->sin_port)
605 return ntohs(((struct sockaddr_in6 *)sin)->sin6_port)
613 * Make sure that we don't have too many active connections. If we have,
614 * something must be dropped. It's not clear what will happen if we allow
615 * "too many" connections, but when dealing with network-facing software,
616 * we have to code defensively. Here we do that by imposing hard limits.
618 * There's no point in trying to do random drop here for DoS
619 * prevention. The NFS clients does 1 reconnect in 15 seconds. An
620 * attacker can easily beat that.
622 * The only somewhat efficient mechanism would be if drop old
623 * connections from the same IP first. But right now we don't even
624 * record the client IP in svc_sock.
626 * single-threaded services that expect a lot of clients will probably
627 * need to set sv_maxconn to override the default value which is based
628 * on the number of threads
630 static void svc_check_conn_limits(struct svc_serv *serv)
632 unsigned int limit = serv->sv_maxconn ? serv->sv_maxconn :
633 (serv->sv_nrthreads+3) * 20;
635 if (serv->sv_tmpcnt > limit) {
636 struct svc_xprt *xprt = NULL;
637 spin_lock_bh(&serv->sv_lock);
638 if (!list_empty(&serv->sv_tempsocks)) {
639 /* Try to help the admin */
640 net_notice_ratelimited("%s: too many open connections, consider increasing the %s\n",
641 serv->sv_name, serv->sv_maxconn ?
642 "max number of connections" :
643 "number of threads");
645 * Always select the oldest connection. It's not fair,
648 xprt = list_entry(serv->sv_tempsocks.prev,
651 set_bit(XPT_CLOSE, &xprt->xpt_flags);
654 spin_unlock_bh(&serv->sv_lock);
657 svc_xprt_enqueue(xprt);
663 static int svc_alloc_arg(struct svc_rqst *rqstp)
665 struct svc_serv *serv = rqstp->rq_server;
666 struct xdr_buf *arg = &rqstp->rq_arg;
667 unsigned long pages, filled, ret;
669 pagevec_init(&rqstp->rq_pvec);
671 pages = (serv->sv_max_mesg + 2 * PAGE_SIZE) >> PAGE_SHIFT;
672 if (pages > RPCSVC_MAXPAGES) {
673 pr_warn_once("svc: warning: pages=%lu > RPCSVC_MAXPAGES=%lu\n",
674 pages, RPCSVC_MAXPAGES);
675 /* use as many pages as possible */
676 pages = RPCSVC_MAXPAGES;
679 for (filled = 0; filled < pages; filled = ret) {
680 ret = alloc_pages_bulk_array(GFP_KERNEL, pages,
683 /* Made progress, don't sleep yet */
686 set_current_state(TASK_INTERRUPTIBLE);
687 if (signalled() || kthread_should_stop()) {
688 set_current_state(TASK_RUNNING);
691 trace_svc_alloc_arg_err(pages);
692 memalloc_retry_wait(GFP_KERNEL);
694 rqstp->rq_page_end = &rqstp->rq_pages[pages];
695 rqstp->rq_pages[pages] = NULL; /* this might be seen in nfsd_splice_actor() */
697 /* Make arg->head point to first page and arg->pages point to rest */
698 arg->head[0].iov_base = page_address(rqstp->rq_pages[0]);
699 arg->head[0].iov_len = PAGE_SIZE;
700 arg->pages = rqstp->rq_pages + 1;
702 /* save at least one page for response */
703 arg->page_len = (pages-2)*PAGE_SIZE;
704 arg->len = (pages-1)*PAGE_SIZE;
705 arg->tail[0].iov_len = 0;
710 rqst_should_sleep(struct svc_rqst *rqstp)
712 struct svc_pool *pool = rqstp->rq_pool;
714 /* did someone call svc_wake_up? */
715 if (test_and_clear_bit(SP_TASK_PENDING, &pool->sp_flags))
718 /* was a socket queued? */
719 if (!list_empty(&pool->sp_sockets))
722 /* are we shutting down? */
723 if (signalled() || kthread_should_stop())
726 /* are we freezing? */
727 if (freezing(current))
733 static struct svc_xprt *svc_get_next_xprt(struct svc_rqst *rqstp, long timeout)
735 struct svc_pool *pool = rqstp->rq_pool;
738 /* rq_xprt should be clear on entry */
739 WARN_ON_ONCE(rqstp->rq_xprt);
741 rqstp->rq_xprt = svc_xprt_dequeue(pool);
746 * We have to be able to interrupt this wait
747 * to bring down the daemons ...
749 set_current_state(TASK_INTERRUPTIBLE);
750 smp_mb__before_atomic();
751 clear_bit(SP_CONGESTED, &pool->sp_flags);
752 clear_bit(RQ_BUSY, &rqstp->rq_flags);
753 smp_mb__after_atomic();
755 if (likely(rqst_should_sleep(rqstp)))
756 time_left = schedule_timeout(timeout);
758 __set_current_state(TASK_RUNNING);
762 set_bit(RQ_BUSY, &rqstp->rq_flags);
763 smp_mb__after_atomic();
764 rqstp->rq_xprt = svc_xprt_dequeue(pool);
769 atomic_long_inc(&pool->sp_stats.threads_timedout);
771 if (signalled() || kthread_should_stop())
772 return ERR_PTR(-EINTR);
773 return ERR_PTR(-EAGAIN);
775 /* Normally we will wait up to 5 seconds for any required
776 * cache information to be provided.
778 if (!test_bit(SP_CONGESTED, &pool->sp_flags))
779 rqstp->rq_chandle.thread_wait = 5*HZ;
781 rqstp->rq_chandle.thread_wait = 1*HZ;
782 trace_svc_xprt_dequeue(rqstp);
783 return rqstp->rq_xprt;
786 static void svc_add_new_temp_xprt(struct svc_serv *serv, struct svc_xprt *newxpt)
788 spin_lock_bh(&serv->sv_lock);
789 set_bit(XPT_TEMP, &newxpt->xpt_flags);
790 list_add(&newxpt->xpt_list, &serv->sv_tempsocks);
792 if (serv->sv_temptimer.function == NULL) {
793 /* setup timer to age temp transports */
794 serv->sv_temptimer.function = svc_age_temp_xprts;
795 mod_timer(&serv->sv_temptimer,
796 jiffies + svc_conn_age_period * HZ);
798 spin_unlock_bh(&serv->sv_lock);
799 svc_xprt_received(newxpt);
802 static int svc_handle_xprt(struct svc_rqst *rqstp, struct svc_xprt *xprt)
804 struct svc_serv *serv = rqstp->rq_server;
807 if (test_bit(XPT_CLOSE, &xprt->xpt_flags)) {
808 if (test_and_clear_bit(XPT_KILL_TEMP, &xprt->xpt_flags))
809 xprt->xpt_ops->xpo_kill_temp_xprt(xprt);
810 svc_delete_xprt(xprt);
811 /* Leave XPT_BUSY set on the dead xprt: */
814 if (test_bit(XPT_LISTENER, &xprt->xpt_flags)) {
815 struct svc_xprt *newxpt;
817 * We know this module_get will succeed because the
818 * listener holds a reference too
820 __module_get(xprt->xpt_class->xcl_owner);
821 svc_check_conn_limits(xprt->xpt_server);
822 newxpt = xprt->xpt_ops->xpo_accept(xprt);
824 newxpt->xpt_cred = get_cred(xprt->xpt_cred);
825 svc_add_new_temp_xprt(serv, newxpt);
826 trace_svc_xprt_accept(newxpt, serv->sv_name);
828 module_put(xprt->xpt_class->xcl_owner);
830 svc_xprt_received(xprt);
831 } else if (svc_xprt_reserve_slot(rqstp, xprt)) {
832 /* XPT_DATA|XPT_DEFERRED case: */
833 dprintk("svc: server %p, pool %u, transport %p, inuse=%d\n",
834 rqstp, rqstp->rq_pool->sp_id, xprt,
835 kref_read(&xprt->xpt_ref));
836 rqstp->rq_deferred = svc_deferred_dequeue(xprt);
837 if (rqstp->rq_deferred)
838 len = svc_deferred_recv(rqstp);
840 len = xprt->xpt_ops->xpo_recvfrom(rqstp);
841 rqstp->rq_stime = ktime_get();
842 rqstp->rq_reserved = serv->sv_max_mesg;
843 atomic_add(rqstp->rq_reserved, &xprt->xpt_reserved);
845 svc_xprt_received(xprt);
847 trace_svc_handle_xprt(xprt, len);
852 * Receive the next request on any transport. This code is carefully
853 * organised not to touch any cachelines in the shared svc_serv
854 * structure, only cachelines in the local svc_pool.
856 int svc_recv(struct svc_rqst *rqstp, long timeout)
858 struct svc_xprt *xprt = NULL;
859 struct svc_serv *serv = rqstp->rq_server;
862 err = svc_alloc_arg(rqstp);
869 if (signalled() || kthread_should_stop())
872 xprt = svc_get_next_xprt(rqstp, timeout);
878 len = svc_handle_xprt(rqstp, xprt);
880 /* No data, incomplete (TCP) read, or accept() */
884 trace_svc_xdr_recvfrom(&rqstp->rq_arg);
886 clear_bit(XPT_OLD, &xprt->xpt_flags);
888 xprt->xpt_ops->xpo_secure_port(rqstp);
889 rqstp->rq_chandle.defer = svc_defer;
890 rqstp->rq_xid = svc_getu32(&rqstp->rq_arg.head[0]);
893 serv->sv_stats->netcnt++;
896 rqstp->rq_res.len = 0;
897 svc_xprt_release(rqstp);
901 EXPORT_SYMBOL_GPL(svc_recv);
906 void svc_drop(struct svc_rqst *rqstp)
908 trace_svc_drop(rqstp);
909 svc_xprt_release(rqstp);
911 EXPORT_SYMBOL_GPL(svc_drop);
914 * Return reply to client.
916 int svc_send(struct svc_rqst *rqstp)
918 struct svc_xprt *xprt;
922 xprt = rqstp->rq_xprt;
926 /* calculate over-all length */
928 xb->len = xb->head[0].iov_len +
931 trace_svc_xdr_sendto(rqstp->rq_xid, xb);
932 trace_svc_stats_latency(rqstp);
934 len = xprt->xpt_ops->xpo_sendto(rqstp);
936 trace_svc_send(rqstp, len);
937 svc_xprt_release(rqstp);
939 if (len == -ECONNREFUSED || len == -ENOTCONN || len == -EAGAIN)
946 * Timer function to close old temporary transports, using
947 * a mark-and-sweep algorithm.
949 static void svc_age_temp_xprts(struct timer_list *t)
951 struct svc_serv *serv = from_timer(serv, t, sv_temptimer);
952 struct svc_xprt *xprt;
953 struct list_head *le, *next;
955 dprintk("svc_age_temp_xprts\n");
957 if (!spin_trylock_bh(&serv->sv_lock)) {
958 /* busy, try again 1 sec later */
959 dprintk("svc_age_temp_xprts: busy\n");
960 mod_timer(&serv->sv_temptimer, jiffies + HZ);
964 list_for_each_safe(le, next, &serv->sv_tempsocks) {
965 xprt = list_entry(le, struct svc_xprt, xpt_list);
967 /* First time through, just mark it OLD. Second time
968 * through, close it. */
969 if (!test_and_set_bit(XPT_OLD, &xprt->xpt_flags))
971 if (kref_read(&xprt->xpt_ref) > 1 ||
972 test_bit(XPT_BUSY, &xprt->xpt_flags))
975 set_bit(XPT_CLOSE, &xprt->xpt_flags);
976 dprintk("queuing xprt %p for closing\n", xprt);
978 /* a thread will dequeue and close it soon */
979 svc_xprt_enqueue(xprt);
981 spin_unlock_bh(&serv->sv_lock);
983 mod_timer(&serv->sv_temptimer, jiffies + svc_conn_age_period * HZ);
986 /* Close temporary transports whose xpt_local matches server_addr immediately
987 * instead of waiting for them to be picked up by the timer.
989 * This is meant to be called from a notifier_block that runs when an ip
990 * address is deleted.
992 void svc_age_temp_xprts_now(struct svc_serv *serv, struct sockaddr *server_addr)
994 struct svc_xprt *xprt;
995 struct list_head *le, *next;
996 LIST_HEAD(to_be_closed);
998 spin_lock_bh(&serv->sv_lock);
999 list_for_each_safe(le, next, &serv->sv_tempsocks) {
1000 xprt = list_entry(le, struct svc_xprt, xpt_list);
1001 if (rpc_cmp_addr(server_addr, (struct sockaddr *)
1002 &xprt->xpt_local)) {
1003 dprintk("svc_age_temp_xprts_now: found %p\n", xprt);
1004 list_move(le, &to_be_closed);
1007 spin_unlock_bh(&serv->sv_lock);
1009 while (!list_empty(&to_be_closed)) {
1010 le = to_be_closed.next;
1012 xprt = list_entry(le, struct svc_xprt, xpt_list);
1013 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1014 set_bit(XPT_KILL_TEMP, &xprt->xpt_flags);
1015 dprintk("svc_age_temp_xprts_now: queuing xprt %p for closing\n",
1017 svc_xprt_enqueue(xprt);
1020 EXPORT_SYMBOL_GPL(svc_age_temp_xprts_now);
1022 static void call_xpt_users(struct svc_xprt *xprt)
1024 struct svc_xpt_user *u;
1026 spin_lock(&xprt->xpt_lock);
1027 while (!list_empty(&xprt->xpt_users)) {
1028 u = list_first_entry(&xprt->xpt_users, struct svc_xpt_user, list);
1029 list_del_init(&u->list);
1032 spin_unlock(&xprt->xpt_lock);
1036 * Remove a dead transport
1038 static void svc_delete_xprt(struct svc_xprt *xprt)
1040 struct svc_serv *serv = xprt->xpt_server;
1041 struct svc_deferred_req *dr;
1043 if (test_and_set_bit(XPT_DEAD, &xprt->xpt_flags))
1046 trace_svc_xprt_detach(xprt);
1047 xprt->xpt_ops->xpo_detach(xprt);
1048 if (xprt->xpt_bc_xprt)
1049 xprt->xpt_bc_xprt->ops->close(xprt->xpt_bc_xprt);
1051 spin_lock_bh(&serv->sv_lock);
1052 list_del_init(&xprt->xpt_list);
1053 WARN_ON_ONCE(!list_empty(&xprt->xpt_ready));
1054 if (test_bit(XPT_TEMP, &xprt->xpt_flags))
1056 spin_unlock_bh(&serv->sv_lock);
1058 while ((dr = svc_deferred_dequeue(xprt)) != NULL)
1061 call_xpt_users(xprt);
1065 void svc_close_xprt(struct svc_xprt *xprt)
1067 trace_svc_xprt_close(xprt);
1068 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1069 if (test_and_set_bit(XPT_BUSY, &xprt->xpt_flags))
1070 /* someone else will have to effect the close */
1073 * We expect svc_close_xprt() to work even when no threads are
1074 * running (e.g., while configuring the server before starting
1075 * any threads), so if the transport isn't busy, we delete
1078 svc_delete_xprt(xprt);
1080 EXPORT_SYMBOL_GPL(svc_close_xprt);
1082 static int svc_close_list(struct svc_serv *serv, struct list_head *xprt_list, struct net *net)
1084 struct svc_xprt *xprt;
1087 spin_lock_bh(&serv->sv_lock);
1088 list_for_each_entry(xprt, xprt_list, xpt_list) {
1089 if (xprt->xpt_net != net)
1092 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1093 svc_xprt_enqueue(xprt);
1095 spin_unlock_bh(&serv->sv_lock);
1099 static struct svc_xprt *svc_dequeue_net(struct svc_serv *serv, struct net *net)
1101 struct svc_pool *pool;
1102 struct svc_xprt *xprt;
1103 struct svc_xprt *tmp;
1106 for (i = 0; i < serv->sv_nrpools; i++) {
1107 pool = &serv->sv_pools[i];
1109 spin_lock_bh(&pool->sp_lock);
1110 list_for_each_entry_safe(xprt, tmp, &pool->sp_sockets, xpt_ready) {
1111 if (xprt->xpt_net != net)
1113 list_del_init(&xprt->xpt_ready);
1114 spin_unlock_bh(&pool->sp_lock);
1117 spin_unlock_bh(&pool->sp_lock);
1122 static void svc_clean_up_xprts(struct svc_serv *serv, struct net *net)
1124 struct svc_xprt *xprt;
1126 while ((xprt = svc_dequeue_net(serv, net))) {
1127 set_bit(XPT_CLOSE, &xprt->xpt_flags);
1128 svc_delete_xprt(xprt);
1133 * Server threads may still be running (especially in the case where the
1134 * service is still running in other network namespaces).
1136 * So we shut down sockets the same way we would on a running server, by
1137 * setting XPT_CLOSE, enqueuing, and letting a thread pick it up to do
1138 * the close. In the case there are no such other threads,
1139 * threads running, svc_clean_up_xprts() does a simple version of a
1140 * server's main event loop, and in the case where there are other
1141 * threads, we may need to wait a little while and then check again to
1142 * see if they're done.
1144 void svc_close_net(struct svc_serv *serv, struct net *net)
1148 while (svc_close_list(serv, &serv->sv_permsocks, net) +
1149 svc_close_list(serv, &serv->sv_tempsocks, net)) {
1151 svc_clean_up_xprts(serv, net);
1157 * Handle defer and revisit of requests
1160 static void svc_revisit(struct cache_deferred_req *dreq, int too_many)
1162 struct svc_deferred_req *dr =
1163 container_of(dreq, struct svc_deferred_req, handle);
1164 struct svc_xprt *xprt = dr->xprt;
1166 spin_lock(&xprt->xpt_lock);
1167 set_bit(XPT_DEFERRED, &xprt->xpt_flags);
1168 if (too_many || test_bit(XPT_DEAD, &xprt->xpt_flags)) {
1169 spin_unlock(&xprt->xpt_lock);
1170 trace_svc_defer_drop(dr);
1176 list_add(&dr->handle.recent, &xprt->xpt_deferred);
1177 spin_unlock(&xprt->xpt_lock);
1178 trace_svc_defer_queue(dr);
1179 svc_xprt_enqueue(xprt);
1184 * Save the request off for later processing. The request buffer looks
1187 * <xprt-header><rpc-header><rpc-pagelist><rpc-tail>
1189 * This code can only handle requests that consist of an xprt-header
1192 static struct cache_deferred_req *svc_defer(struct cache_req *req)
1194 struct svc_rqst *rqstp = container_of(req, struct svc_rqst, rq_chandle);
1195 struct svc_deferred_req *dr;
1197 if (rqstp->rq_arg.page_len || !test_bit(RQ_USEDEFERRAL, &rqstp->rq_flags))
1198 return NULL; /* if more than a page, give up FIXME */
1199 if (rqstp->rq_deferred) {
1200 dr = rqstp->rq_deferred;
1201 rqstp->rq_deferred = NULL;
1205 /* FIXME maybe discard if size too large */
1206 size = sizeof(struct svc_deferred_req) + rqstp->rq_arg.len;
1207 dr = kmalloc(size, GFP_KERNEL);
1211 dr->handle.owner = rqstp->rq_server;
1212 dr->prot = rqstp->rq_prot;
1213 memcpy(&dr->addr, &rqstp->rq_addr, rqstp->rq_addrlen);
1214 dr->addrlen = rqstp->rq_addrlen;
1215 dr->daddr = rqstp->rq_daddr;
1216 dr->argslen = rqstp->rq_arg.len >> 2;
1217 dr->xprt_hlen = rqstp->rq_xprt_hlen;
1219 /* back up head to the start of the buffer and copy */
1220 skip = rqstp->rq_arg.len - rqstp->rq_arg.head[0].iov_len;
1221 memcpy(dr->args, rqstp->rq_arg.head[0].iov_base - skip,
1224 trace_svc_defer(rqstp);
1225 svc_xprt_get(rqstp->rq_xprt);
1226 dr->xprt = rqstp->rq_xprt;
1227 set_bit(RQ_DROPME, &rqstp->rq_flags);
1229 dr->handle.revisit = svc_revisit;
1234 * recv data from a deferred request into an active one
1236 static noinline int svc_deferred_recv(struct svc_rqst *rqstp)
1238 struct svc_deferred_req *dr = rqstp->rq_deferred;
1240 trace_svc_defer_recv(dr);
1242 /* setup iov_base past transport header */
1243 rqstp->rq_arg.head[0].iov_base = dr->args + (dr->xprt_hlen>>2);
1244 /* The iov_len does not include the transport header bytes */
1245 rqstp->rq_arg.head[0].iov_len = (dr->argslen<<2) - dr->xprt_hlen;
1246 rqstp->rq_arg.page_len = 0;
1247 /* The rq_arg.len includes the transport header bytes */
1248 rqstp->rq_arg.len = dr->argslen<<2;
1249 rqstp->rq_prot = dr->prot;
1250 memcpy(&rqstp->rq_addr, &dr->addr, dr->addrlen);
1251 rqstp->rq_addrlen = dr->addrlen;
1252 /* Save off transport header len in case we get deferred again */
1253 rqstp->rq_xprt_hlen = dr->xprt_hlen;
1254 rqstp->rq_daddr = dr->daddr;
1255 rqstp->rq_respages = rqstp->rq_pages;
1256 svc_xprt_received(rqstp->rq_xprt);
1257 return (dr->argslen<<2) - dr->xprt_hlen;
1261 static struct svc_deferred_req *svc_deferred_dequeue(struct svc_xprt *xprt)
1263 struct svc_deferred_req *dr = NULL;
1265 if (!test_bit(XPT_DEFERRED, &xprt->xpt_flags))
1267 spin_lock(&xprt->xpt_lock);
1268 if (!list_empty(&xprt->xpt_deferred)) {
1269 dr = list_entry(xprt->xpt_deferred.next,
1270 struct svc_deferred_req,
1272 list_del_init(&dr->handle.recent);
1274 clear_bit(XPT_DEFERRED, &xprt->xpt_flags);
1275 spin_unlock(&xprt->xpt_lock);
1280 * svc_find_xprt - find an RPC transport instance
1281 * @serv: pointer to svc_serv to search
1282 * @xcl_name: C string containing transport's class name
1283 * @net: owner net pointer
1284 * @af: Address family of transport's local address
1285 * @port: transport's IP port number
1287 * Return the transport instance pointer for the endpoint accepting
1288 * connections/peer traffic from the specified transport class,
1289 * address family and port.
1291 * Specifying 0 for the address family or port is effectively a
1292 * wild-card, and will result in matching the first transport in the
1293 * service's list that has a matching class name.
1295 struct svc_xprt *svc_find_xprt(struct svc_serv *serv, const char *xcl_name,
1296 struct net *net, const sa_family_t af,
1297 const unsigned short port)
1299 struct svc_xprt *xprt;
1300 struct svc_xprt *found = NULL;
1302 /* Sanity check the args */
1303 if (serv == NULL || xcl_name == NULL)
1306 spin_lock_bh(&serv->sv_lock);
1307 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1308 if (xprt->xpt_net != net)
1310 if (strcmp(xprt->xpt_class->xcl_name, xcl_name))
1312 if (af != AF_UNSPEC && af != xprt->xpt_local.ss_family)
1314 if (port != 0 && port != svc_xprt_local_port(xprt))
1320 spin_unlock_bh(&serv->sv_lock);
1323 EXPORT_SYMBOL_GPL(svc_find_xprt);
1325 static int svc_one_xprt_name(const struct svc_xprt *xprt,
1326 char *pos, int remaining)
1330 len = snprintf(pos, remaining, "%s %u\n",
1331 xprt->xpt_class->xcl_name,
1332 svc_xprt_local_port(xprt));
1333 if (len >= remaining)
1334 return -ENAMETOOLONG;
1339 * svc_xprt_names - format a buffer with a list of transport names
1340 * @serv: pointer to an RPC service
1341 * @buf: pointer to a buffer to be filled in
1342 * @buflen: length of buffer to be filled in
1344 * Fills in @buf with a string containing a list of transport names,
1345 * each name terminated with '\n'.
1347 * Returns positive length of the filled-in string on success; otherwise
1348 * a negative errno value is returned if an error occurs.
1350 int svc_xprt_names(struct svc_serv *serv, char *buf, const int buflen)
1352 struct svc_xprt *xprt;
1356 /* Sanity check args */
1360 spin_lock_bh(&serv->sv_lock);
1364 list_for_each_entry(xprt, &serv->sv_permsocks, xpt_list) {
1365 len = svc_one_xprt_name(xprt, pos, buflen - totlen);
1377 spin_unlock_bh(&serv->sv_lock);
1380 EXPORT_SYMBOL_GPL(svc_xprt_names);
1383 /*----------------------------------------------------------------------------*/
1385 static void *svc_pool_stats_start(struct seq_file *m, loff_t *pos)
1387 unsigned int pidx = (unsigned int)*pos;
1388 struct svc_serv *serv = m->private;
1390 dprintk("svc_pool_stats_start, *pidx=%u\n", pidx);
1393 return SEQ_START_TOKEN;
1394 return (pidx > serv->sv_nrpools ? NULL : &serv->sv_pools[pidx-1]);
1397 static void *svc_pool_stats_next(struct seq_file *m, void *p, loff_t *pos)
1399 struct svc_pool *pool = p;
1400 struct svc_serv *serv = m->private;
1402 dprintk("svc_pool_stats_next, *pos=%llu\n", *pos);
1404 if (p == SEQ_START_TOKEN) {
1405 pool = &serv->sv_pools[0];
1407 unsigned int pidx = (pool - &serv->sv_pools[0]);
1408 if (pidx < serv->sv_nrpools-1)
1409 pool = &serv->sv_pools[pidx+1];
1417 static void svc_pool_stats_stop(struct seq_file *m, void *p)
1421 static int svc_pool_stats_show(struct seq_file *m, void *p)
1423 struct svc_pool *pool = p;
1425 if (p == SEQ_START_TOKEN) {
1426 seq_puts(m, "# pool packets-arrived sockets-enqueued threads-woken threads-timedout\n");
1430 seq_printf(m, "%u %lu %lu %lu %lu\n",
1432 (unsigned long)atomic_long_read(&pool->sp_stats.packets),
1433 pool->sp_stats.sockets_queued,
1434 (unsigned long)atomic_long_read(&pool->sp_stats.threads_woken),
1435 (unsigned long)atomic_long_read(&pool->sp_stats.threads_timedout));
1440 static const struct seq_operations svc_pool_stats_seq_ops = {
1441 .start = svc_pool_stats_start,
1442 .next = svc_pool_stats_next,
1443 .stop = svc_pool_stats_stop,
1444 .show = svc_pool_stats_show,
1447 int svc_pool_stats_open(struct svc_serv *serv, struct file *file)
1451 err = seq_open(file, &svc_pool_stats_seq_ops);
1453 ((struct seq_file *) file->private_data)->private = serv;
1456 EXPORT_SYMBOL(svc_pool_stats_open);
1458 /*----------------------------------------------------------------------------*/