1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/net/sunrpc/svc.c
5 * High-level RPC service routines
7 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
9 * Multiple threads pools and NUMAisation
10 * Copyright (c) 2006 Silicon Graphics, Inc.
11 * by Greg Banks <gnb@melbourne.sgi.com>
14 #include <linux/linkage.h>
15 #include <linux/sched/signal.h>
16 #include <linux/errno.h>
17 #include <linux/net.h>
20 #include <linux/interrupt.h>
21 #include <linux/module.h>
22 #include <linux/kthread.h>
23 #include <linux/slab.h>
25 #include <linux/sunrpc/types.h>
26 #include <linux/sunrpc/xdr.h>
27 #include <linux/sunrpc/stats.h>
28 #include <linux/sunrpc/svcsock.h>
29 #include <linux/sunrpc/clnt.h>
30 #include <linux/sunrpc/bc_xprt.h>
32 #include <trace/events/sunrpc.h>
36 #define RPCDBG_FACILITY RPCDBG_SVCDSP
38 static void svc_unregister(const struct svc_serv *serv, struct net *net);
40 #define svc_serv_is_pooled(serv) ((serv)->sv_ops->svo_function)
42 #define SVC_POOL_DEFAULT SVC_POOL_GLOBAL
45 * Structure for mapping cpus to pools and vice versa.
46 * Setup once during sunrpc initialisation.
48 struct svc_pool_map svc_pool_map = {
49 .mode = SVC_POOL_DEFAULT
51 EXPORT_SYMBOL_GPL(svc_pool_map);
53 static DEFINE_MUTEX(svc_pool_map_mutex);/* protects svc_pool_map.count only */
56 param_set_pool_mode(const char *val, const struct kernel_param *kp)
58 int *ip = (int *)kp->arg;
59 struct svc_pool_map *m = &svc_pool_map;
62 mutex_lock(&svc_pool_map_mutex);
69 if (!strncmp(val, "auto", 4))
71 else if (!strncmp(val, "global", 6))
72 *ip = SVC_POOL_GLOBAL;
73 else if (!strncmp(val, "percpu", 6))
74 *ip = SVC_POOL_PERCPU;
75 else if (!strncmp(val, "pernode", 7))
76 *ip = SVC_POOL_PERNODE;
81 mutex_unlock(&svc_pool_map_mutex);
86 param_get_pool_mode(char *buf, const struct kernel_param *kp)
88 int *ip = (int *)kp->arg;
93 return strlcpy(buf, "auto\n", 20);
95 return strlcpy(buf, "global\n", 20);
97 return strlcpy(buf, "percpu\n", 20);
98 case SVC_POOL_PERNODE:
99 return strlcpy(buf, "pernode\n", 20);
101 return sprintf(buf, "%d\n", *ip);
105 module_param_call(pool_mode, param_set_pool_mode, param_get_pool_mode,
106 &svc_pool_map.mode, 0644);
109 * Detect best pool mapping mode heuristically,
110 * according to the machine's topology.
113 svc_pool_map_choose_mode(void)
117 if (nr_online_nodes > 1) {
119 * Actually have multiple NUMA nodes,
120 * so split pools on NUMA node boundaries
122 return SVC_POOL_PERNODE;
125 node = first_online_node;
126 if (nr_cpus_node(node) > 2) {
128 * Non-trivial SMP, or CONFIG_NUMA on
129 * non-NUMA hardware, e.g. with a generic
130 * x86_64 kernel on Xeons. In this case we
131 * want to divide the pools on cpu boundaries.
133 return SVC_POOL_PERCPU;
136 /* default: one global pool */
137 return SVC_POOL_GLOBAL;
141 * Allocate the to_pool[] and pool_to[] arrays.
142 * Returns 0 on success or an errno.
145 svc_pool_map_alloc_arrays(struct svc_pool_map *m, unsigned int maxpools)
147 m->to_pool = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
150 m->pool_to = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
164 * Initialise the pool map for SVC_POOL_PERCPU mode.
165 * Returns number of pools or <0 on error.
168 svc_pool_map_init_percpu(struct svc_pool_map *m)
170 unsigned int maxpools = nr_cpu_ids;
171 unsigned int pidx = 0;
175 err = svc_pool_map_alloc_arrays(m, maxpools);
179 for_each_online_cpu(cpu) {
180 BUG_ON(pidx >= maxpools);
181 m->to_pool[cpu] = pidx;
182 m->pool_to[pidx] = cpu;
185 /* cpus brought online later all get mapped to pool0, sorry */
192 * Initialise the pool map for SVC_POOL_PERNODE mode.
193 * Returns number of pools or <0 on error.
196 svc_pool_map_init_pernode(struct svc_pool_map *m)
198 unsigned int maxpools = nr_node_ids;
199 unsigned int pidx = 0;
203 err = svc_pool_map_alloc_arrays(m, maxpools);
207 for_each_node_with_cpus(node) {
208 /* some architectures (e.g. SN2) have cpuless nodes */
209 BUG_ON(pidx > maxpools);
210 m->to_pool[node] = pidx;
211 m->pool_to[pidx] = node;
214 /* nodes brought online later all get mapped to pool0, sorry */
221 * Add a reference to the global map of cpus to pools (and
222 * vice versa). Initialise the map if we're the first user.
223 * Returns the number of pools.
226 svc_pool_map_get(void)
228 struct svc_pool_map *m = &svc_pool_map;
231 mutex_lock(&svc_pool_map_mutex);
234 mutex_unlock(&svc_pool_map_mutex);
238 if (m->mode == SVC_POOL_AUTO)
239 m->mode = svc_pool_map_choose_mode();
242 case SVC_POOL_PERCPU:
243 npools = svc_pool_map_init_percpu(m);
245 case SVC_POOL_PERNODE:
246 npools = svc_pool_map_init_pernode(m);
251 /* default, or memory allocation failure */
253 m->mode = SVC_POOL_GLOBAL;
257 mutex_unlock(&svc_pool_map_mutex);
260 EXPORT_SYMBOL_GPL(svc_pool_map_get);
263 * Drop a reference to the global map of cpus to pools.
264 * When the last reference is dropped, the map data is
265 * freed; this allows the sysadmin to change the pool
266 * mode using the pool_mode module option without
267 * rebooting or re-loading sunrpc.ko.
270 svc_pool_map_put(void)
272 struct svc_pool_map *m = &svc_pool_map;
274 mutex_lock(&svc_pool_map_mutex);
284 mutex_unlock(&svc_pool_map_mutex);
286 EXPORT_SYMBOL_GPL(svc_pool_map_put);
288 static int svc_pool_map_get_node(unsigned int pidx)
290 const struct svc_pool_map *m = &svc_pool_map;
293 if (m->mode == SVC_POOL_PERCPU)
294 return cpu_to_node(m->pool_to[pidx]);
295 if (m->mode == SVC_POOL_PERNODE)
296 return m->pool_to[pidx];
301 * Set the given thread's cpus_allowed mask so that it
302 * will only run on cpus in the given pool.
305 svc_pool_map_set_cpumask(struct task_struct *task, unsigned int pidx)
307 struct svc_pool_map *m = &svc_pool_map;
308 unsigned int node = m->pool_to[pidx];
311 * The caller checks for sv_nrpools > 1, which
312 * implies that we've been initialized.
314 WARN_ON_ONCE(m->count == 0);
319 case SVC_POOL_PERCPU:
321 set_cpus_allowed_ptr(task, cpumask_of(node));
324 case SVC_POOL_PERNODE:
326 set_cpus_allowed_ptr(task, cpumask_of_node(node));
333 * Use the mapping mode to choose a pool for a given CPU.
334 * Used when enqueueing an incoming RPC. Always returns
335 * a non-NULL pool pointer.
338 svc_pool_for_cpu(struct svc_serv *serv, int cpu)
340 struct svc_pool_map *m = &svc_pool_map;
341 unsigned int pidx = 0;
344 * An uninitialised map happens in a pure client when
345 * lockd is brought up, so silently treat it the
346 * same as SVC_POOL_GLOBAL.
348 if (svc_serv_is_pooled(serv)) {
350 case SVC_POOL_PERCPU:
351 pidx = m->to_pool[cpu];
353 case SVC_POOL_PERNODE:
354 pidx = m->to_pool[cpu_to_node(cpu)];
358 return &serv->sv_pools[pidx % serv->sv_nrpools];
361 int svc_rpcb_setup(struct svc_serv *serv, struct net *net)
365 err = rpcb_create_local(net);
369 /* Remove any stale portmap registrations */
370 svc_unregister(serv, net);
373 EXPORT_SYMBOL_GPL(svc_rpcb_setup);
375 void svc_rpcb_cleanup(struct svc_serv *serv, struct net *net)
377 svc_unregister(serv, net);
380 EXPORT_SYMBOL_GPL(svc_rpcb_cleanup);
382 static int svc_uses_rpcbind(struct svc_serv *serv)
384 struct svc_program *progp;
387 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
388 for (i = 0; i < progp->pg_nvers; i++) {
389 if (progp->pg_vers[i] == NULL)
391 if (!progp->pg_vers[i]->vs_hidden)
399 int svc_bind(struct svc_serv *serv, struct net *net)
401 if (!svc_uses_rpcbind(serv))
403 return svc_rpcb_setup(serv, net);
405 EXPORT_SYMBOL_GPL(svc_bind);
407 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
409 __svc_init_bc(struct svc_serv *serv)
411 INIT_LIST_HEAD(&serv->sv_cb_list);
412 spin_lock_init(&serv->sv_cb_lock);
413 init_waitqueue_head(&serv->sv_cb_waitq);
417 __svc_init_bc(struct svc_serv *serv)
423 * Create an RPC service
425 static struct svc_serv *
426 __svc_create(struct svc_program *prog, unsigned int bufsize, int npools,
427 const struct svc_serv_ops *ops)
429 struct svc_serv *serv;
431 unsigned int xdrsize;
434 if (!(serv = kzalloc(sizeof(*serv), GFP_KERNEL)))
436 serv->sv_name = prog->pg_name;
437 serv->sv_program = prog;
438 serv->sv_nrthreads = 1;
439 serv->sv_stats = prog->pg_stats;
440 if (bufsize > RPCSVC_MAXPAYLOAD)
441 bufsize = RPCSVC_MAXPAYLOAD;
442 serv->sv_max_payload = bufsize? bufsize : 4096;
443 serv->sv_max_mesg = roundup(serv->sv_max_payload + PAGE_SIZE, PAGE_SIZE);
447 prog->pg_lovers = prog->pg_nvers-1;
448 for (vers=0; vers<prog->pg_nvers ; vers++)
449 if (prog->pg_vers[vers]) {
450 prog->pg_hivers = vers;
451 if (prog->pg_lovers > vers)
452 prog->pg_lovers = vers;
453 if (prog->pg_vers[vers]->vs_xdrsize > xdrsize)
454 xdrsize = prog->pg_vers[vers]->vs_xdrsize;
456 prog = prog->pg_next;
458 serv->sv_xdrsize = xdrsize;
459 INIT_LIST_HEAD(&serv->sv_tempsocks);
460 INIT_LIST_HEAD(&serv->sv_permsocks);
461 timer_setup(&serv->sv_temptimer, NULL, 0);
462 spin_lock_init(&serv->sv_lock);
466 serv->sv_nrpools = npools;
468 kcalloc(serv->sv_nrpools, sizeof(struct svc_pool),
470 if (!serv->sv_pools) {
475 for (i = 0; i < serv->sv_nrpools; i++) {
476 struct svc_pool *pool = &serv->sv_pools[i];
478 dprintk("svc: initialising pool %u for %s\n",
482 INIT_LIST_HEAD(&pool->sp_sockets);
483 INIT_LIST_HEAD(&pool->sp_all_threads);
484 spin_lock_init(&pool->sp_lock);
491 svc_create(struct svc_program *prog, unsigned int bufsize,
492 const struct svc_serv_ops *ops)
494 return __svc_create(prog, bufsize, /*npools*/1, ops);
496 EXPORT_SYMBOL_GPL(svc_create);
499 svc_create_pooled(struct svc_program *prog, unsigned int bufsize,
500 const struct svc_serv_ops *ops)
502 struct svc_serv *serv;
503 unsigned int npools = svc_pool_map_get();
505 serv = __svc_create(prog, bufsize, npools, ops);
513 EXPORT_SYMBOL_GPL(svc_create_pooled);
515 void svc_shutdown_net(struct svc_serv *serv, struct net *net)
517 svc_close_net(serv, net);
519 if (serv->sv_ops->svo_shutdown)
520 serv->sv_ops->svo_shutdown(serv, net);
522 EXPORT_SYMBOL_GPL(svc_shutdown_net);
525 * Destroy an RPC service. Should be called with appropriate locking to
526 * protect the sv_nrthreads, sv_permsocks and sv_tempsocks.
529 svc_destroy(struct svc_serv *serv)
531 dprintk("svc: svc_destroy(%s, %d)\n",
532 serv->sv_program->pg_name,
535 if (serv->sv_nrthreads) {
536 if (--(serv->sv_nrthreads) != 0) {
537 svc_sock_update_bufs(serv);
541 printk("svc_destroy: no threads for serv=%p!\n", serv);
543 del_timer_sync(&serv->sv_temptimer);
546 * The last user is gone and thus all sockets have to be destroyed to
547 * the point. Check this.
549 BUG_ON(!list_empty(&serv->sv_permsocks));
550 BUG_ON(!list_empty(&serv->sv_tempsocks));
552 cache_clean_deferred(serv);
554 if (svc_serv_is_pooled(serv))
557 kfree(serv->sv_pools);
560 EXPORT_SYMBOL_GPL(svc_destroy);
563 * Allocate an RPC server's buffer space.
564 * We allocate pages and place them in rq_pages.
567 svc_init_buffer(struct svc_rqst *rqstp, unsigned int size, int node)
569 unsigned int pages, arghi;
571 /* bc_xprt uses fore channel allocated buffers */
572 if (svc_is_backchannel(rqstp))
575 pages = size / PAGE_SIZE + 1; /* extra page as we hold both request and reply.
576 * We assume one is at most one page
579 WARN_ON_ONCE(pages > RPCSVC_MAXPAGES);
580 if (pages > RPCSVC_MAXPAGES)
581 pages = RPCSVC_MAXPAGES;
583 struct page *p = alloc_pages_node(node, GFP_KERNEL, 0);
586 rqstp->rq_pages[arghi++] = p;
593 * Release an RPC server buffer
596 svc_release_buffer(struct svc_rqst *rqstp)
600 for (i = 0; i < ARRAY_SIZE(rqstp->rq_pages); i++)
601 if (rqstp->rq_pages[i])
602 put_page(rqstp->rq_pages[i]);
606 svc_rqst_alloc(struct svc_serv *serv, struct svc_pool *pool, int node)
608 struct svc_rqst *rqstp;
610 rqstp = kzalloc_node(sizeof(*rqstp), GFP_KERNEL, node);
614 __set_bit(RQ_BUSY, &rqstp->rq_flags);
615 spin_lock_init(&rqstp->rq_lock);
616 rqstp->rq_server = serv;
617 rqstp->rq_pool = pool;
619 rqstp->rq_scratch_page = alloc_pages_node(node, GFP_KERNEL, 0);
620 if (!rqstp->rq_scratch_page)
623 rqstp->rq_argp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
627 rqstp->rq_resp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
631 if (!svc_init_buffer(rqstp, serv->sv_max_mesg, node))
636 svc_rqst_free(rqstp);
639 EXPORT_SYMBOL_GPL(svc_rqst_alloc);
642 svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
644 struct svc_rqst *rqstp;
646 rqstp = svc_rqst_alloc(serv, pool, node);
648 return ERR_PTR(-ENOMEM);
650 serv->sv_nrthreads++;
651 spin_lock_bh(&pool->sp_lock);
652 pool->sp_nrthreads++;
653 list_add_rcu(&rqstp->rq_all, &pool->sp_all_threads);
654 spin_unlock_bh(&pool->sp_lock);
657 EXPORT_SYMBOL_GPL(svc_prepare_thread);
660 * Choose a pool in which to create a new thread, for svc_set_num_threads
662 static inline struct svc_pool *
663 choose_pool(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
668 return &serv->sv_pools[(*state)++ % serv->sv_nrpools];
672 * Choose a thread to kill, for svc_set_num_threads
674 static inline struct task_struct *
675 choose_victim(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
678 struct task_struct *task = NULL;
681 spin_lock_bh(&pool->sp_lock);
683 /* choose a pool in round-robin fashion */
684 for (i = 0; i < serv->sv_nrpools; i++) {
685 pool = &serv->sv_pools[--(*state) % serv->sv_nrpools];
686 spin_lock_bh(&pool->sp_lock);
687 if (!list_empty(&pool->sp_all_threads))
689 spin_unlock_bh(&pool->sp_lock);
695 if (!list_empty(&pool->sp_all_threads)) {
696 struct svc_rqst *rqstp;
699 * Remove from the pool->sp_all_threads list
700 * so we don't try to kill it again.
702 rqstp = list_entry(pool->sp_all_threads.next, struct svc_rqst, rq_all);
703 set_bit(RQ_VICTIM, &rqstp->rq_flags);
704 list_del_rcu(&rqstp->rq_all);
705 task = rqstp->rq_task;
707 spin_unlock_bh(&pool->sp_lock);
712 /* create new threads */
714 svc_start_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
716 struct svc_rqst *rqstp;
717 struct task_struct *task;
718 struct svc_pool *chosen_pool;
719 unsigned int state = serv->sv_nrthreads-1;
724 chosen_pool = choose_pool(serv, pool, &state);
726 node = svc_pool_map_get_node(chosen_pool->sp_id);
727 rqstp = svc_prepare_thread(serv, chosen_pool, node);
729 return PTR_ERR(rqstp);
731 __module_get(serv->sv_ops->svo_module);
732 task = kthread_create_on_node(serv->sv_ops->svo_function, rqstp,
733 node, "%s", serv->sv_name);
735 module_put(serv->sv_ops->svo_module);
736 svc_exit_thread(rqstp);
737 return PTR_ERR(task);
740 rqstp->rq_task = task;
741 if (serv->sv_nrpools > 1)
742 svc_pool_map_set_cpumask(task, chosen_pool->sp_id);
744 svc_sock_update_bufs(serv);
745 wake_up_process(task);
746 } while (nrservs > 0);
752 /* destroy old threads */
754 svc_signal_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
756 struct task_struct *task;
757 unsigned int state = serv->sv_nrthreads-1;
759 /* destroy old threads */
761 task = choose_victim(serv, pool, &state);
764 send_sig(SIGINT, task, 1);
766 } while (nrservs < 0);
772 * Create or destroy enough new threads to make the number
773 * of threads the given number. If `pool' is non-NULL, applies
774 * only to threads in that pool, otherwise round-robins between
775 * all pools. Caller must ensure that mutual exclusion between this and
776 * server startup or shutdown.
778 * Destroying threads relies on the service threads filling in
779 * rqstp->rq_task, which only the nfs ones do. Assumes the serv
780 * has been created using svc_create_pooled().
782 * Based on code that used to be in nfsd_svc() but tweaked
786 svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
789 /* The -1 assumes caller has done a svc_get() */
790 nrservs -= (serv->sv_nrthreads-1);
792 spin_lock_bh(&pool->sp_lock);
793 nrservs -= pool->sp_nrthreads;
794 spin_unlock_bh(&pool->sp_lock);
798 return svc_start_kthreads(serv, pool, nrservs);
800 return svc_signal_kthreads(serv, pool, nrservs);
803 EXPORT_SYMBOL_GPL(svc_set_num_threads);
805 /* destroy old threads */
807 svc_stop_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
809 struct svc_rqst *rqstp;
810 struct task_struct *task;
811 unsigned int state = serv->sv_nrthreads-1;
813 /* destroy old threads */
815 task = choose_victim(serv, pool, &state);
818 rqstp = kthread_data(task);
819 /* Did we lose a race to svo_function threadfn? */
820 if (kthread_stop(task) == -EINTR)
821 svc_exit_thread(rqstp);
823 } while (nrservs < 0);
828 svc_set_num_threads_sync(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
831 /* The -1 assumes caller has done a svc_get() */
832 nrservs -= (serv->sv_nrthreads-1);
834 spin_lock_bh(&pool->sp_lock);
835 nrservs -= pool->sp_nrthreads;
836 spin_unlock_bh(&pool->sp_lock);
840 return svc_start_kthreads(serv, pool, nrservs);
842 return svc_stop_kthreads(serv, pool, nrservs);
845 EXPORT_SYMBOL_GPL(svc_set_num_threads_sync);
848 * svc_rqst_replace_page - Replace one page in rq_pages[]
849 * @rqstp: svc_rqst with pages to replace
850 * @page: replacement page
852 * When replacing a page in rq_pages, batch the release of the
853 * replaced pages to avoid hammering the page allocator.
855 void svc_rqst_replace_page(struct svc_rqst *rqstp, struct page *page)
857 if (*rqstp->rq_next_page) {
858 if (!pagevec_space(&rqstp->rq_pvec))
859 __pagevec_release(&rqstp->rq_pvec);
860 pagevec_add(&rqstp->rq_pvec, *rqstp->rq_next_page);
864 *(rqstp->rq_next_page++) = page;
866 EXPORT_SYMBOL_GPL(svc_rqst_replace_page);
869 * Called from a server thread as it's exiting. Caller must hold the "service
870 * mutex" for the service.
873 svc_rqst_free(struct svc_rqst *rqstp)
875 svc_release_buffer(rqstp);
876 if (rqstp->rq_scratch_page)
877 put_page(rqstp->rq_scratch_page);
878 kfree(rqstp->rq_resp);
879 kfree(rqstp->rq_argp);
880 kfree(rqstp->rq_auth_data);
881 kfree_rcu(rqstp, rq_rcu_head);
883 EXPORT_SYMBOL_GPL(svc_rqst_free);
886 svc_exit_thread(struct svc_rqst *rqstp)
888 struct svc_serv *serv = rqstp->rq_server;
889 struct svc_pool *pool = rqstp->rq_pool;
891 spin_lock_bh(&pool->sp_lock);
892 pool->sp_nrthreads--;
893 if (!test_and_set_bit(RQ_VICTIM, &rqstp->rq_flags))
894 list_del_rcu(&rqstp->rq_all);
895 spin_unlock_bh(&pool->sp_lock);
897 svc_rqst_free(rqstp);
899 /* Release the server */
903 EXPORT_SYMBOL_GPL(svc_exit_thread);
906 * Register an "inet" protocol family netid with the local
907 * rpcbind daemon via an rpcbind v4 SET request.
909 * No netconfig infrastructure is available in the kernel, so
910 * we map IP_ protocol numbers to netids by hand.
912 * Returns zero on success; a negative errno value is returned
913 * if any error occurs.
915 static int __svc_rpcb_register4(struct net *net, const u32 program,
917 const unsigned short protocol,
918 const unsigned short port)
920 const struct sockaddr_in sin = {
921 .sin_family = AF_INET,
922 .sin_addr.s_addr = htonl(INADDR_ANY),
923 .sin_port = htons(port),
930 netid = RPCBIND_NETID_UDP;
933 netid = RPCBIND_NETID_TCP;
939 error = rpcb_v4_register(net, program, version,
940 (const struct sockaddr *)&sin, netid);
943 * User space didn't support rpcbind v4, so retry this
944 * registration request with the legacy rpcbind v2 protocol.
946 if (error == -EPROTONOSUPPORT)
947 error = rpcb_register(net, program, version, protocol, port);
952 #if IS_ENABLED(CONFIG_IPV6)
954 * Register an "inet6" protocol family netid with the local
955 * rpcbind daemon via an rpcbind v4 SET request.
957 * No netconfig infrastructure is available in the kernel, so
958 * we map IP_ protocol numbers to netids by hand.
960 * Returns zero on success; a negative errno value is returned
961 * if any error occurs.
963 static int __svc_rpcb_register6(struct net *net, const u32 program,
965 const unsigned short protocol,
966 const unsigned short port)
968 const struct sockaddr_in6 sin6 = {
969 .sin6_family = AF_INET6,
970 .sin6_addr = IN6ADDR_ANY_INIT,
971 .sin6_port = htons(port),
978 netid = RPCBIND_NETID_UDP6;
981 netid = RPCBIND_NETID_TCP6;
987 error = rpcb_v4_register(net, program, version,
988 (const struct sockaddr *)&sin6, netid);
991 * User space didn't support rpcbind version 4, so we won't
992 * use a PF_INET6 listener.
994 if (error == -EPROTONOSUPPORT)
995 error = -EAFNOSUPPORT;
999 #endif /* IS_ENABLED(CONFIG_IPV6) */
1002 * Register a kernel RPC service via rpcbind version 4.
1004 * Returns zero on success; a negative errno value is returned
1005 * if any error occurs.
1007 static int __svc_register(struct net *net, const char *progname,
1008 const u32 program, const u32 version,
1010 const unsigned short protocol,
1011 const unsigned short port)
1013 int error = -EAFNOSUPPORT;
1017 error = __svc_rpcb_register4(net, program, version,
1020 #if IS_ENABLED(CONFIG_IPV6)
1022 error = __svc_rpcb_register6(net, program, version,
1027 trace_svc_register(progname, version, family, protocol, port, error);
1031 int svc_rpcbind_set_version(struct net *net,
1032 const struct svc_program *progp,
1033 u32 version, int family,
1034 unsigned short proto,
1035 unsigned short port)
1037 return __svc_register(net, progp->pg_name, progp->pg_prog,
1038 version, family, proto, port);
1041 EXPORT_SYMBOL_GPL(svc_rpcbind_set_version);
1043 int svc_generic_rpcbind_set(struct net *net,
1044 const struct svc_program *progp,
1045 u32 version, int family,
1046 unsigned short proto,
1047 unsigned short port)
1049 const struct svc_version *vers = progp->pg_vers[version];
1055 if (vers->vs_hidden) {
1056 trace_svc_noregister(progp->pg_name, version, proto,
1062 * Don't register a UDP port if we need congestion
1065 if (vers->vs_need_cong_ctrl && proto == IPPROTO_UDP)
1068 error = svc_rpcbind_set_version(net, progp, version,
1069 family, proto, port);
1071 return (vers->vs_rpcb_optnl) ? 0 : error;
1073 EXPORT_SYMBOL_GPL(svc_generic_rpcbind_set);
1076 * svc_register - register an RPC service with the local portmapper
1077 * @serv: svc_serv struct for the service to register
1078 * @net: net namespace for the service to register
1079 * @family: protocol family of service's listener socket
1080 * @proto: transport protocol number to advertise
1081 * @port: port to advertise
1083 * Service is registered for any address in the passed-in protocol family
1085 int svc_register(const struct svc_serv *serv, struct net *net,
1086 const int family, const unsigned short proto,
1087 const unsigned short port)
1089 struct svc_program *progp;
1093 WARN_ON_ONCE(proto == 0 && port == 0);
1094 if (proto == 0 && port == 0)
1097 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1098 for (i = 0; i < progp->pg_nvers; i++) {
1100 error = progp->pg_rpcbind_set(net, progp, i,
1101 family, proto, port);
1103 printk(KERN_WARNING "svc: failed to register "
1104 "%sv%u RPC service (errno %d).\n",
1105 progp->pg_name, i, -error);
1115 * If user space is running rpcbind, it should take the v4 UNSET
1116 * and clear everything for this [program, version]. If user space
1117 * is running portmap, it will reject the v4 UNSET, but won't have
1118 * any "inet6" entries anyway. So a PMAP_UNSET should be sufficient
1119 * in this case to clear all existing entries for [program, version].
1121 static void __svc_unregister(struct net *net, const u32 program, const u32 version,
1122 const char *progname)
1126 error = rpcb_v4_register(net, program, version, NULL, "");
1129 * User space didn't support rpcbind v4, so retry this
1130 * request with the legacy rpcbind v2 protocol.
1132 if (error == -EPROTONOSUPPORT)
1133 error = rpcb_register(net, program, version, 0, 0);
1135 trace_svc_unregister(progname, version, error);
1139 * All netids, bind addresses and ports registered for [program, version]
1140 * are removed from the local rpcbind database (if the service is not
1141 * hidden) to make way for a new instance of the service.
1143 * The result of unregistration is reported via dprintk for those who want
1144 * verification of the result, but is otherwise not important.
1146 static void svc_unregister(const struct svc_serv *serv, struct net *net)
1148 struct svc_program *progp;
1149 unsigned long flags;
1152 clear_thread_flag(TIF_SIGPENDING);
1154 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1155 for (i = 0; i < progp->pg_nvers; i++) {
1156 if (progp->pg_vers[i] == NULL)
1158 if (progp->pg_vers[i]->vs_hidden)
1160 __svc_unregister(net, progp->pg_prog, i, progp->pg_name);
1164 spin_lock_irqsave(¤t->sighand->siglock, flags);
1165 recalc_sigpending();
1166 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
1170 * dprintk the given error with the address of the client that caused it.
1172 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
1173 static __printf(2, 3)
1174 void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...)
1176 struct va_format vaf;
1178 char buf[RPC_MAX_ADDRBUFLEN];
1180 va_start(args, fmt);
1185 dprintk("svc: %s: %pV", svc_print_addr(rqstp, buf, sizeof(buf)), &vaf);
1190 static __printf(2,3) void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...) {}
1194 svc_generic_dispatch(struct svc_rqst *rqstp, __be32 *statp)
1196 struct kvec *argv = &rqstp->rq_arg.head[0];
1197 struct kvec *resv = &rqstp->rq_res.head[0];
1198 const struct svc_procedure *procp = rqstp->rq_procinfo;
1202 * XXX: why do we ignore the return value?
1204 if (procp->pc_decode &&
1205 !procp->pc_decode(rqstp, argv->iov_base)) {
1206 *statp = rpc_garbage_args;
1210 *statp = procp->pc_func(rqstp);
1212 if (*statp == rpc_drop_reply ||
1213 test_bit(RQ_DROPME, &rqstp->rq_flags))
1216 if (rqstp->rq_auth_stat != rpc_auth_ok)
1219 if (*statp != rpc_success)
1223 if (procp->pc_encode &&
1224 !procp->pc_encode(rqstp, resv->iov_base + resv->iov_len)) {
1225 dprintk("svc: failed to encode reply\n");
1226 /* serv->sv_stats->rpcsystemerr++; */
1227 *statp = rpc_system_err;
1233 svc_generic_init_request(struct svc_rqst *rqstp,
1234 const struct svc_program *progp,
1235 struct svc_process_info *ret)
1237 const struct svc_version *versp = NULL; /* compiler food */
1238 const struct svc_procedure *procp = NULL;
1240 if (rqstp->rq_vers >= progp->pg_nvers )
1242 versp = progp->pg_vers[rqstp->rq_vers];
1247 * Some protocol versions (namely NFSv4) require some form of
1248 * congestion control. (See RFC 7530 section 3.1 paragraph 2)
1249 * In other words, UDP is not allowed. We mark those when setting
1250 * up the svc_xprt, and verify that here.
1252 * The spec is not very clear about what error should be returned
1253 * when someone tries to access a server that is listening on UDP
1254 * for lower versions. RPC_PROG_MISMATCH seems to be the closest
1257 if (versp->vs_need_cong_ctrl && rqstp->rq_xprt &&
1258 !test_bit(XPT_CONG_CTRL, &rqstp->rq_xprt->xpt_flags))
1261 if (rqstp->rq_proc >= versp->vs_nproc)
1263 rqstp->rq_procinfo = procp = &versp->vs_proc[rqstp->rq_proc];
1267 /* Initialize storage for argp and resp */
1268 memset(rqstp->rq_argp, 0, procp->pc_argsize);
1269 memset(rqstp->rq_resp, 0, procp->pc_ressize);
1271 /* Bump per-procedure stats counter */
1272 versp->vs_count[rqstp->rq_proc]++;
1274 ret->dispatch = versp->vs_dispatch;
1277 ret->mismatch.lovers = progp->pg_lovers;
1278 ret->mismatch.hivers = progp->pg_hivers;
1279 return rpc_prog_mismatch;
1281 return rpc_proc_unavail;
1283 EXPORT_SYMBOL_GPL(svc_generic_init_request);
1286 * Common routine for processing the RPC request.
1289 svc_process_common(struct svc_rqst *rqstp, struct kvec *argv, struct kvec *resv)
1291 struct svc_program *progp;
1292 const struct svc_procedure *procp = NULL;
1293 struct svc_serv *serv = rqstp->rq_server;
1294 struct svc_process_info process;
1299 __be32 *reply_statp;
1301 rpc_stat = rpc_success;
1303 if (argv->iov_len < 6*4)
1306 /* Will be turned off by GSS integrity and privacy services */
1307 set_bit(RQ_SPLICE_OK, &rqstp->rq_flags);
1308 /* Will be turned off only when NFSv4 Sessions are used */
1309 set_bit(RQ_USEDEFERRAL, &rqstp->rq_flags);
1310 clear_bit(RQ_DROPME, &rqstp->rq_flags);
1312 svc_putu32(resv, rqstp->rq_xid);
1314 vers = svc_getnl(argv);
1316 /* First words of reply: */
1317 svc_putnl(resv, 1); /* REPLY */
1319 if (vers != 2) /* RPC version number */
1322 /* Save position in case we later decide to reject: */
1323 reply_statp = resv->iov_base + resv->iov_len;
1325 svc_putnl(resv, 0); /* ACCEPT */
1327 rqstp->rq_prog = prog = svc_getnl(argv); /* program number */
1328 rqstp->rq_vers = svc_getnl(argv); /* version number */
1329 rqstp->rq_proc = svc_getnl(argv); /* procedure number */
1331 for (progp = serv->sv_program; progp; progp = progp->pg_next)
1332 if (prog == progp->pg_prog)
1336 * Decode auth data, and add verifier to reply buffer.
1337 * We do this before anything else in order to get a decent
1340 auth_res = svc_authenticate(rqstp);
1341 /* Also give the program a chance to reject this call: */
1342 if (auth_res == SVC_OK && progp)
1343 auth_res = progp->pg_authenticate(rqstp);
1344 if (auth_res != SVC_OK)
1345 trace_svc_authenticate(rqstp, auth_res);
1352 rpc_stat = rpc_system_err;
1367 rpc_stat = progp->pg_init_request(rqstp, progp, &process);
1371 case rpc_prog_unavail:
1373 case rpc_prog_mismatch:
1375 case rpc_proc_unavail:
1379 procp = rqstp->rq_procinfo;
1380 /* Should this check go into the dispatcher? */
1381 if (!procp || !procp->pc_func)
1384 /* Syntactic check complete */
1385 serv->sv_stats->rpccnt++;
1386 trace_svc_process(rqstp, progp->pg_name);
1388 /* Build the reply header. */
1389 statp = resv->iov_base +resv->iov_len;
1390 svc_putnl(resv, RPC_SUCCESS);
1392 /* un-reserve some of the out-queue now that we have a
1393 * better idea of reply size
1395 if (procp->pc_xdrressize)
1396 svc_reserve_auth(rqstp, procp->pc_xdrressize<<2);
1398 /* Call the function that processes the request. */
1399 if (!process.dispatch) {
1400 if (!svc_generic_dispatch(rqstp, statp))
1401 goto release_dropit;
1402 if (*statp == rpc_garbage_args)
1405 dprintk("svc: calling dispatcher\n");
1406 if (!process.dispatch(rqstp, statp))
1407 goto release_dropit; /* Release reply info */
1410 if (rqstp->rq_auth_stat != rpc_auth_ok)
1411 goto err_release_bad_auth;
1413 /* Check RPC status result */
1414 if (*statp != rpc_success)
1415 resv->iov_len = ((void*)statp) - resv->iov_base + 4;
1417 /* Release reply info */
1418 if (procp->pc_release)
1419 procp->pc_release(rqstp);
1421 if (procp->pc_encode == NULL)
1425 if (svc_authorise(rqstp))
1427 return 1; /* Caller can now send it */
1430 if (procp->pc_release)
1431 procp->pc_release(rqstp);
1433 svc_authorise(rqstp); /* doesn't hurt to call this twice */
1434 dprintk("svc: svc_process dropit\n");
1438 svc_authorise(rqstp);
1440 if (rqstp->rq_xprt && test_bit(XPT_TEMP, &rqstp->rq_xprt->xpt_flags))
1441 svc_close_xprt(rqstp->rq_xprt);
1442 dprintk("svc: svc_process close\n");
1446 svc_printk(rqstp, "short len %zd, dropping request\n",
1451 serv->sv_stats->rpcbadfmt++;
1452 svc_putnl(resv, 1); /* REJECT */
1453 svc_putnl(resv, 0); /* RPC_MISMATCH */
1454 svc_putnl(resv, 2); /* Only RPCv2 supported */
1458 err_release_bad_auth:
1459 if (procp->pc_release)
1460 procp->pc_release(rqstp);
1462 dprintk("svc: authentication failed (%d)\n",
1463 be32_to_cpu(rqstp->rq_auth_stat));
1464 serv->sv_stats->rpcbadauth++;
1465 /* Restore write pointer to location of accept status: */
1466 xdr_ressize_check(rqstp, reply_statp);
1467 svc_putnl(resv, 1); /* REJECT */
1468 svc_putnl(resv, 1); /* AUTH_ERROR */
1469 svc_putu32(resv, rqstp->rq_auth_stat); /* status */
1473 dprintk("svc: unknown program %d\n", prog);
1474 serv->sv_stats->rpcbadfmt++;
1475 svc_putnl(resv, RPC_PROG_UNAVAIL);
1479 svc_printk(rqstp, "unknown version (%d for prog %d, %s)\n",
1480 rqstp->rq_vers, rqstp->rq_prog, progp->pg_name);
1482 serv->sv_stats->rpcbadfmt++;
1483 svc_putnl(resv, RPC_PROG_MISMATCH);
1484 svc_putnl(resv, process.mismatch.lovers);
1485 svc_putnl(resv, process.mismatch.hivers);
1489 svc_printk(rqstp, "unknown procedure (%d)\n", rqstp->rq_proc);
1491 serv->sv_stats->rpcbadfmt++;
1492 svc_putnl(resv, RPC_PROC_UNAVAIL);
1496 svc_printk(rqstp, "failed to decode args\n");
1498 rpc_stat = rpc_garbage_args;
1500 serv->sv_stats->rpcbadfmt++;
1501 svc_putnl(resv, ntohl(rpc_stat));
1506 * Process the RPC request.
1509 svc_process(struct svc_rqst *rqstp)
1511 struct kvec *argv = &rqstp->rq_arg.head[0];
1512 struct kvec *resv = &rqstp->rq_res.head[0];
1513 struct svc_serv *serv = rqstp->rq_server;
1516 #if IS_ENABLED(CONFIG_FAIL_SUNRPC)
1517 if (!fail_sunrpc.ignore_server_disconnect &&
1518 should_fail(&fail_sunrpc.attr, 1))
1519 svc_xprt_deferred_close(rqstp->rq_xprt);
1523 * Setup response xdr_buf.
1524 * Initially it has just one page
1526 rqstp->rq_next_page = &rqstp->rq_respages[1];
1527 resv->iov_base = page_address(rqstp->rq_respages[0]);
1529 rqstp->rq_res.pages = rqstp->rq_respages + 1;
1530 rqstp->rq_res.len = 0;
1531 rqstp->rq_res.page_base = 0;
1532 rqstp->rq_res.page_len = 0;
1533 rqstp->rq_res.buflen = PAGE_SIZE;
1534 rqstp->rq_res.tail[0].iov_base = NULL;
1535 rqstp->rq_res.tail[0].iov_len = 0;
1537 dir = svc_getnl(argv);
1539 /* direction != CALL */
1540 svc_printk(rqstp, "bad direction %d, dropping request\n", dir);
1541 serv->sv_stats->rpcbadfmt++;
1545 /* Returns 1 for send, 0 for drop */
1546 if (likely(svc_process_common(rqstp, argv, resv)))
1547 return svc_send(rqstp);
1553 EXPORT_SYMBOL_GPL(svc_process);
1555 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1557 * Process a backchannel RPC request that arrived over an existing
1558 * outbound connection
1561 bc_svc_process(struct svc_serv *serv, struct rpc_rqst *req,
1562 struct svc_rqst *rqstp)
1564 struct kvec *argv = &rqstp->rq_arg.head[0];
1565 struct kvec *resv = &rqstp->rq_res.head[0];
1566 struct rpc_task *task;
1570 dprintk("svc: %s(%p)\n", __func__, req);
1572 /* Build the svc_rqst used by the common processing routine */
1573 rqstp->rq_xid = req->rq_xid;
1574 rqstp->rq_prot = req->rq_xprt->prot;
1575 rqstp->rq_server = serv;
1576 rqstp->rq_bc_net = req->rq_xprt->xprt_net;
1578 rqstp->rq_addrlen = sizeof(req->rq_xprt->addr);
1579 memcpy(&rqstp->rq_addr, &req->rq_xprt->addr, rqstp->rq_addrlen);
1580 memcpy(&rqstp->rq_arg, &req->rq_rcv_buf, sizeof(rqstp->rq_arg));
1581 memcpy(&rqstp->rq_res, &req->rq_snd_buf, sizeof(rqstp->rq_res));
1583 /* Adjust the argument buffer length */
1584 rqstp->rq_arg.len = req->rq_private_buf.len;
1585 if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len) {
1586 rqstp->rq_arg.head[0].iov_len = rqstp->rq_arg.len;
1587 rqstp->rq_arg.page_len = 0;
1588 } else if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len +
1589 rqstp->rq_arg.page_len)
1590 rqstp->rq_arg.page_len = rqstp->rq_arg.len -
1591 rqstp->rq_arg.head[0].iov_len;
1593 rqstp->rq_arg.len = rqstp->rq_arg.head[0].iov_len +
1594 rqstp->rq_arg.page_len;
1596 /* reset result send buffer "put" position */
1600 * Skip the next two words because they've already been
1601 * processed in the transport
1603 svc_getu32(argv); /* XID */
1604 svc_getnl(argv); /* CALLDIR */
1606 /* Parse and execute the bc call */
1607 proc_error = svc_process_common(rqstp, argv, resv);
1609 atomic_dec(&req->rq_xprt->bc_slot_count);
1611 /* Processing error: drop the request */
1612 xprt_free_bc_request(req);
1616 /* Finally, send the reply synchronously */
1617 memcpy(&req->rq_snd_buf, &rqstp->rq_res, sizeof(req->rq_snd_buf));
1618 task = rpc_run_bc_task(req);
1620 error = PTR_ERR(task);
1624 WARN_ON_ONCE(atomic_read(&task->tk_count) != 1);
1625 error = task->tk_status;
1629 dprintk("svc: %s(), error=%d\n", __func__, error);
1632 EXPORT_SYMBOL_GPL(bc_svc_process);
1633 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1636 * Return (transport-specific) limit on the rpc payload.
1638 u32 svc_max_payload(const struct svc_rqst *rqstp)
1640 u32 max = rqstp->rq_xprt->xpt_class->xcl_max_payload;
1642 if (rqstp->rq_server->sv_max_payload < max)
1643 max = rqstp->rq_server->sv_max_payload;
1646 EXPORT_SYMBOL_GPL(svc_max_payload);
1649 * svc_proc_name - Return RPC procedure name in string form
1650 * @rqstp: svc_rqst to operate on
1653 * Pointer to a NUL-terminated string
1655 const char *svc_proc_name(const struct svc_rqst *rqstp)
1657 if (rqstp && rqstp->rq_procinfo)
1658 return rqstp->rq_procinfo->pc_name;
1664 * svc_encode_result_payload - mark a range of bytes as a result payload
1665 * @rqstp: svc_rqst to operate on
1666 * @offset: payload's byte offset in rqstp->rq_res
1667 * @length: size of payload, in bytes
1669 * Returns zero on success, or a negative errno if a permanent
1672 int svc_encode_result_payload(struct svc_rqst *rqstp, unsigned int offset,
1673 unsigned int length)
1675 return rqstp->rq_xprt->xpt_ops->xpo_result_payload(rqstp, offset,
1678 EXPORT_SYMBOL_GPL(svc_encode_result_payload);
1681 * svc_fill_write_vector - Construct data argument for VFS write call
1682 * @rqstp: svc_rqst to operate on
1683 * @payload: xdr_buf containing only the write data payload
1685 * Fills in rqstp::rq_vec, and returns the number of elements.
1687 unsigned int svc_fill_write_vector(struct svc_rqst *rqstp,
1688 struct xdr_buf *payload)
1690 struct page **pages = payload->pages;
1691 struct kvec *first = payload->head;
1692 struct kvec *vec = rqstp->rq_vec;
1693 size_t total = payload->len;
1696 /* Some types of transport can present the write payload
1697 * entirely in rq_arg.pages. In this case, @first is empty.
1700 if (first->iov_len) {
1701 vec[i].iov_base = first->iov_base;
1702 vec[i].iov_len = min_t(size_t, total, first->iov_len);
1703 total -= vec[i].iov_len;
1708 vec[i].iov_base = page_address(*pages);
1709 vec[i].iov_len = min_t(size_t, total, PAGE_SIZE);
1710 total -= vec[i].iov_len;
1715 WARN_ON_ONCE(i > ARRAY_SIZE(rqstp->rq_vec));
1718 EXPORT_SYMBOL_GPL(svc_fill_write_vector);
1721 * svc_fill_symlink_pathname - Construct pathname argument for VFS symlink call
1722 * @rqstp: svc_rqst to operate on
1723 * @first: buffer containing first section of pathname
1724 * @p: buffer containing remaining section of pathname
1725 * @total: total length of the pathname argument
1727 * The VFS symlink API demands a NUL-terminated pathname in mapped memory.
1728 * Returns pointer to a NUL-terminated string, or an ERR_PTR. Caller must free
1729 * the returned string.
1731 char *svc_fill_symlink_pathname(struct svc_rqst *rqstp, struct kvec *first,
1732 void *p, size_t total)
1734 size_t len, remaining;
1737 result = kmalloc(total + 1, GFP_KERNEL);
1739 return ERR_PTR(-ESERVERFAULT);
1744 len = min_t(size_t, total, first->iov_len);
1746 memcpy(dst, first->iov_base, len);
1752 len = min_t(size_t, remaining, PAGE_SIZE);
1753 memcpy(dst, p, len);
1759 /* Sanity check: Linux doesn't allow the pathname argument to
1760 * contain a NUL byte.
1762 if (strlen(result) != total) {
1764 return ERR_PTR(-EINVAL);
1768 EXPORT_SYMBOL_GPL(svc_fill_symlink_pathname);