1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/net/sunrpc/svc.c
5 * High-level RPC service routines
9 * Multiple threads pools and NUMAisation
10 * Copyright (c) 2006 Silicon Graphics, Inc.
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>
34 #define RPCDBG_FACILITY RPCDBG_SVCDSP
36 static void svc_unregister(const struct svc_serv *serv, struct net *net);
38 #define svc_serv_is_pooled(serv) ((serv)->sv_ops->svo_function)
40 #define SVC_POOL_DEFAULT SVC_POOL_GLOBAL
43 * Structure for mapping cpus to pools and vice versa.
44 * Setup once during sunrpc initialisation.
46 struct svc_pool_map svc_pool_map = {
47 .mode = SVC_POOL_DEFAULT
49 EXPORT_SYMBOL_GPL(svc_pool_map);
51 static DEFINE_MUTEX(svc_pool_map_mutex);/* protects svc_pool_map.count only */
54 param_set_pool_mode(const char *val, const struct kernel_param *kp)
56 int *ip = (int *)kp->arg;
57 struct svc_pool_map *m = &svc_pool_map;
60 mutex_lock(&svc_pool_map_mutex);
67 if (!strncmp(val, "auto", 4))
69 else if (!strncmp(val, "global", 6))
70 *ip = SVC_POOL_GLOBAL;
71 else if (!strncmp(val, "percpu", 6))
72 *ip = SVC_POOL_PERCPU;
73 else if (!strncmp(val, "pernode", 7))
74 *ip = SVC_POOL_PERNODE;
79 mutex_unlock(&svc_pool_map_mutex);
84 param_get_pool_mode(char *buf, const struct kernel_param *kp)
86 int *ip = (int *)kp->arg;
91 return strlcpy(buf, "auto\n", 20);
93 return strlcpy(buf, "global\n", 20);
95 return strlcpy(buf, "percpu\n", 20);
96 case SVC_POOL_PERNODE:
97 return strlcpy(buf, "pernode\n", 20);
99 return sprintf(buf, "%d\n", *ip);
103 module_param_call(pool_mode, param_set_pool_mode, param_get_pool_mode,
104 &svc_pool_map.mode, 0644);
107 * Detect best pool mapping mode heuristically,
108 * according to the machine's topology.
111 svc_pool_map_choose_mode(void)
115 if (nr_online_nodes > 1) {
117 * Actually have multiple NUMA nodes,
118 * so split pools on NUMA node boundaries
120 return SVC_POOL_PERNODE;
123 node = first_online_node;
124 if (nr_cpus_node(node) > 2) {
126 * Non-trivial SMP, or CONFIG_NUMA on
127 * non-NUMA hardware, e.g. with a generic
128 * x86_64 kernel on Xeons. In this case we
129 * want to divide the pools on cpu boundaries.
131 return SVC_POOL_PERCPU;
134 /* default: one global pool */
135 return SVC_POOL_GLOBAL;
139 * Allocate the to_pool[] and pool_to[] arrays.
140 * Returns 0 on success or an errno.
143 svc_pool_map_alloc_arrays(struct svc_pool_map *m, unsigned int maxpools)
145 m->to_pool = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
148 m->pool_to = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
162 * Initialise the pool map for SVC_POOL_PERCPU mode.
163 * Returns number of pools or <0 on error.
166 svc_pool_map_init_percpu(struct svc_pool_map *m)
168 unsigned int maxpools = nr_cpu_ids;
169 unsigned int pidx = 0;
173 err = svc_pool_map_alloc_arrays(m, maxpools);
177 for_each_online_cpu(cpu) {
178 BUG_ON(pidx >= maxpools);
179 m->to_pool[cpu] = pidx;
180 m->pool_to[pidx] = cpu;
183 /* cpus brought online later all get mapped to pool0, sorry */
190 * Initialise the pool map for SVC_POOL_PERNODE mode.
191 * Returns number of pools or <0 on error.
194 svc_pool_map_init_pernode(struct svc_pool_map *m)
196 unsigned int maxpools = nr_node_ids;
197 unsigned int pidx = 0;
201 err = svc_pool_map_alloc_arrays(m, maxpools);
205 for_each_node_with_cpus(node) {
206 /* some architectures (e.g. SN2) have cpuless nodes */
207 BUG_ON(pidx > maxpools);
208 m->to_pool[node] = pidx;
209 m->pool_to[pidx] = node;
212 /* nodes brought online later all get mapped to pool0, sorry */
219 * Add a reference to the global map of cpus to pools (and
220 * vice versa). Initialise the map if we're the first user.
221 * Returns the number of pools.
224 svc_pool_map_get(void)
226 struct svc_pool_map *m = &svc_pool_map;
229 mutex_lock(&svc_pool_map_mutex);
232 mutex_unlock(&svc_pool_map_mutex);
236 if (m->mode == SVC_POOL_AUTO)
237 m->mode = svc_pool_map_choose_mode();
240 case SVC_POOL_PERCPU:
241 npools = svc_pool_map_init_percpu(m);
243 case SVC_POOL_PERNODE:
244 npools = svc_pool_map_init_pernode(m);
249 /* default, or memory allocation failure */
251 m->mode = SVC_POOL_GLOBAL;
255 mutex_unlock(&svc_pool_map_mutex);
258 EXPORT_SYMBOL_GPL(svc_pool_map_get);
261 * Drop a reference to the global map of cpus to pools.
262 * When the last reference is dropped, the map data is
263 * freed; this allows the sysadmin to change the pool
264 * mode using the pool_mode module option without
265 * rebooting or re-loading sunrpc.ko.
268 svc_pool_map_put(void)
270 struct svc_pool_map *m = &svc_pool_map;
272 mutex_lock(&svc_pool_map_mutex);
282 mutex_unlock(&svc_pool_map_mutex);
284 EXPORT_SYMBOL_GPL(svc_pool_map_put);
286 static int svc_pool_map_get_node(unsigned int pidx)
288 const struct svc_pool_map *m = &svc_pool_map;
291 if (m->mode == SVC_POOL_PERCPU)
292 return cpu_to_node(m->pool_to[pidx]);
293 if (m->mode == SVC_POOL_PERNODE)
294 return m->pool_to[pidx];
299 * Set the given thread's cpus_allowed mask so that it
300 * will only run on cpus in the given pool.
303 svc_pool_map_set_cpumask(struct task_struct *task, unsigned int pidx)
305 struct svc_pool_map *m = &svc_pool_map;
306 unsigned int node = m->pool_to[pidx];
309 * The caller checks for sv_nrpools > 1, which
310 * implies that we've been initialized.
312 WARN_ON_ONCE(m->count == 0);
317 case SVC_POOL_PERCPU:
319 set_cpus_allowed_ptr(task, cpumask_of(node));
322 case SVC_POOL_PERNODE:
324 set_cpus_allowed_ptr(task, cpumask_of_node(node));
331 * Use the mapping mode to choose a pool for a given CPU.
332 * Used when enqueueing an incoming RPC. Always returns
333 * a non-NULL pool pointer.
336 svc_pool_for_cpu(struct svc_serv *serv, int cpu)
338 struct svc_pool_map *m = &svc_pool_map;
339 unsigned int pidx = 0;
342 * An uninitialised map happens in a pure client when
343 * lockd is brought up, so silently treat it the
344 * same as SVC_POOL_GLOBAL.
346 if (svc_serv_is_pooled(serv)) {
348 case SVC_POOL_PERCPU:
349 pidx = m->to_pool[cpu];
351 case SVC_POOL_PERNODE:
352 pidx = m->to_pool[cpu_to_node(cpu)];
356 return &serv->sv_pools[pidx % serv->sv_nrpools];
359 int svc_rpcb_setup(struct svc_serv *serv, struct net *net)
363 err = rpcb_create_local(net);
367 /* Remove any stale portmap registrations */
368 svc_unregister(serv, net);
371 EXPORT_SYMBOL_GPL(svc_rpcb_setup);
373 void svc_rpcb_cleanup(struct svc_serv *serv, struct net *net)
375 svc_unregister(serv, net);
378 EXPORT_SYMBOL_GPL(svc_rpcb_cleanup);
380 static int svc_uses_rpcbind(struct svc_serv *serv)
382 struct svc_program *progp;
385 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
386 for (i = 0; i < progp->pg_nvers; i++) {
387 if (progp->pg_vers[i] == NULL)
389 if (!progp->pg_vers[i]->vs_hidden)
397 int svc_bind(struct svc_serv *serv, struct net *net)
399 if (!svc_uses_rpcbind(serv))
401 return svc_rpcb_setup(serv, net);
403 EXPORT_SYMBOL_GPL(svc_bind);
405 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
407 __svc_init_bc(struct svc_serv *serv)
409 INIT_LIST_HEAD(&serv->sv_cb_list);
410 spin_lock_init(&serv->sv_cb_lock);
411 init_waitqueue_head(&serv->sv_cb_waitq);
415 __svc_init_bc(struct svc_serv *serv)
421 * Create an RPC service
423 static struct svc_serv *
424 __svc_create(struct svc_program *prog, unsigned int bufsize, int npools,
425 const struct svc_serv_ops *ops)
427 struct svc_serv *serv;
429 unsigned int xdrsize;
432 if (!(serv = kzalloc(sizeof(*serv), GFP_KERNEL)))
434 serv->sv_name = prog->pg_name;
435 serv->sv_program = prog;
436 serv->sv_nrthreads = 1;
437 serv->sv_stats = prog->pg_stats;
438 if (bufsize > RPCSVC_MAXPAYLOAD)
439 bufsize = RPCSVC_MAXPAYLOAD;
440 serv->sv_max_payload = bufsize? bufsize : 4096;
441 serv->sv_max_mesg = roundup(serv->sv_max_payload + PAGE_SIZE, PAGE_SIZE);
445 prog->pg_lovers = prog->pg_nvers-1;
446 for (vers=0; vers<prog->pg_nvers ; vers++)
447 if (prog->pg_vers[vers]) {
448 prog->pg_hivers = vers;
449 if (prog->pg_lovers > vers)
450 prog->pg_lovers = vers;
451 if (prog->pg_vers[vers]->vs_xdrsize > xdrsize)
452 xdrsize = prog->pg_vers[vers]->vs_xdrsize;
454 prog = prog->pg_next;
456 serv->sv_xdrsize = xdrsize;
457 INIT_LIST_HEAD(&serv->sv_tempsocks);
458 INIT_LIST_HEAD(&serv->sv_permsocks);
459 timer_setup(&serv->sv_temptimer, NULL, 0);
460 spin_lock_init(&serv->sv_lock);
464 serv->sv_nrpools = npools;
466 kcalloc(serv->sv_nrpools, sizeof(struct svc_pool),
468 if (!serv->sv_pools) {
473 for (i = 0; i < serv->sv_nrpools; i++) {
474 struct svc_pool *pool = &serv->sv_pools[i];
476 dprintk("svc: initialising pool %u for %s\n",
480 INIT_LIST_HEAD(&pool->sp_sockets);
481 INIT_LIST_HEAD(&pool->sp_all_threads);
482 spin_lock_init(&pool->sp_lock);
489 svc_create(struct svc_program *prog, unsigned int bufsize,
490 const struct svc_serv_ops *ops)
492 return __svc_create(prog, bufsize, /*npools*/1, ops);
494 EXPORT_SYMBOL_GPL(svc_create);
497 svc_create_pooled(struct svc_program *prog, unsigned int bufsize,
498 const struct svc_serv_ops *ops)
500 struct svc_serv *serv;
501 unsigned int npools = svc_pool_map_get();
503 serv = __svc_create(prog, bufsize, npools, ops);
511 EXPORT_SYMBOL_GPL(svc_create_pooled);
513 void svc_shutdown_net(struct svc_serv *serv, struct net *net)
515 svc_close_net(serv, net);
517 if (serv->sv_ops->svo_shutdown)
518 serv->sv_ops->svo_shutdown(serv, net);
520 EXPORT_SYMBOL_GPL(svc_shutdown_net);
523 * Destroy an RPC service. Should be called with appropriate locking to
524 * protect the sv_nrthreads, sv_permsocks and sv_tempsocks.
527 svc_destroy(struct svc_serv *serv)
529 dprintk("svc: svc_destroy(%s, %d)\n",
530 serv->sv_program->pg_name,
533 if (serv->sv_nrthreads) {
534 if (--(serv->sv_nrthreads) != 0) {
535 svc_sock_update_bufs(serv);
539 printk("svc_destroy: no threads for serv=%p!\n", serv);
541 del_timer_sync(&serv->sv_temptimer);
544 * The last user is gone and thus all sockets have to be destroyed to
545 * the point. Check this.
547 BUG_ON(!list_empty(&serv->sv_permsocks));
548 BUG_ON(!list_empty(&serv->sv_tempsocks));
550 cache_clean_deferred(serv);
552 if (svc_serv_is_pooled(serv))
555 kfree(serv->sv_pools);
558 EXPORT_SYMBOL_GPL(svc_destroy);
561 * Allocate an RPC server's buffer space.
562 * We allocate pages and place them in rq_pages.
565 svc_init_buffer(struct svc_rqst *rqstp, unsigned int size, int node)
567 unsigned int pages, arghi;
569 /* bc_xprt uses fore channel allocated buffers */
570 if (svc_is_backchannel(rqstp))
573 pages = size / PAGE_SIZE + 1; /* extra page as we hold both request and reply.
574 * We assume one is at most one page
577 WARN_ON_ONCE(pages > RPCSVC_MAXPAGES);
578 if (pages > RPCSVC_MAXPAGES)
579 pages = RPCSVC_MAXPAGES;
581 struct page *p = alloc_pages_node(node, GFP_KERNEL, 0);
584 rqstp->rq_pages[arghi++] = p;
591 * Release an RPC server buffer
594 svc_release_buffer(struct svc_rqst *rqstp)
598 for (i = 0; i < ARRAY_SIZE(rqstp->rq_pages); i++)
599 if (rqstp->rq_pages[i])
600 put_page(rqstp->rq_pages[i]);
604 svc_rqst_alloc(struct svc_serv *serv, struct svc_pool *pool, int node)
606 struct svc_rqst *rqstp;
608 rqstp = kzalloc_node(sizeof(*rqstp), GFP_KERNEL, node);
612 __set_bit(RQ_BUSY, &rqstp->rq_flags);
613 spin_lock_init(&rqstp->rq_lock);
614 rqstp->rq_server = serv;
615 rqstp->rq_pool = pool;
617 rqstp->rq_scratch_page = alloc_pages_node(node, GFP_KERNEL, 0);
618 if (!rqstp->rq_scratch_page)
621 rqstp->rq_argp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
625 rqstp->rq_resp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
629 if (!svc_init_buffer(rqstp, serv->sv_max_mesg, node))
634 svc_rqst_free(rqstp);
637 EXPORT_SYMBOL_GPL(svc_rqst_alloc);
640 svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
642 struct svc_rqst *rqstp;
644 rqstp = svc_rqst_alloc(serv, pool, node);
646 return ERR_PTR(-ENOMEM);
648 serv->sv_nrthreads++;
649 spin_lock_bh(&pool->sp_lock);
650 pool->sp_nrthreads++;
651 list_add_rcu(&rqstp->rq_all, &pool->sp_all_threads);
652 spin_unlock_bh(&pool->sp_lock);
655 EXPORT_SYMBOL_GPL(svc_prepare_thread);
658 * Choose a pool in which to create a new thread, for svc_set_num_threads
660 static inline struct svc_pool *
661 choose_pool(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
666 return &serv->sv_pools[(*state)++ % serv->sv_nrpools];
670 * Choose a thread to kill, for svc_set_num_threads
672 static inline struct task_struct *
673 choose_victim(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
676 struct task_struct *task = NULL;
679 spin_lock_bh(&pool->sp_lock);
681 /* choose a pool in round-robin fashion */
682 for (i = 0; i < serv->sv_nrpools; i++) {
683 pool = &serv->sv_pools[--(*state) % serv->sv_nrpools];
684 spin_lock_bh(&pool->sp_lock);
685 if (!list_empty(&pool->sp_all_threads))
687 spin_unlock_bh(&pool->sp_lock);
693 if (!list_empty(&pool->sp_all_threads)) {
694 struct svc_rqst *rqstp;
697 * Remove from the pool->sp_all_threads list
698 * so we don't try to kill it again.
700 rqstp = list_entry(pool->sp_all_threads.next, struct svc_rqst, rq_all);
701 set_bit(RQ_VICTIM, &rqstp->rq_flags);
702 list_del_rcu(&rqstp->rq_all);
703 task = rqstp->rq_task;
705 spin_unlock_bh(&pool->sp_lock);
710 /* create new threads */
712 svc_start_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
714 struct svc_rqst *rqstp;
715 struct task_struct *task;
716 struct svc_pool *chosen_pool;
717 unsigned int state = serv->sv_nrthreads-1;
722 chosen_pool = choose_pool(serv, pool, &state);
724 node = svc_pool_map_get_node(chosen_pool->sp_id);
725 rqstp = svc_prepare_thread(serv, chosen_pool, node);
727 return PTR_ERR(rqstp);
729 __module_get(serv->sv_ops->svo_module);
730 task = kthread_create_on_node(serv->sv_ops->svo_function, rqstp,
731 node, "%s", serv->sv_name);
733 module_put(serv->sv_ops->svo_module);
734 svc_exit_thread(rqstp);
735 return PTR_ERR(task);
738 rqstp->rq_task = task;
739 if (serv->sv_nrpools > 1)
740 svc_pool_map_set_cpumask(task, chosen_pool->sp_id);
742 svc_sock_update_bufs(serv);
743 wake_up_process(task);
744 } while (nrservs > 0);
750 /* destroy old threads */
752 svc_signal_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
754 struct task_struct *task;
755 unsigned int state = serv->sv_nrthreads-1;
757 /* destroy old threads */
759 task = choose_victim(serv, pool, &state);
762 send_sig(SIGINT, task, 1);
764 } while (nrservs < 0);
770 * Create or destroy enough new threads to make the number
771 * of threads the given number. If `pool' is non-NULL, applies
772 * only to threads in that pool, otherwise round-robins between
773 * all pools. Caller must ensure that mutual exclusion between this and
774 * server startup or shutdown.
776 * Destroying threads relies on the service threads filling in
777 * rqstp->rq_task, which only the nfs ones do. Assumes the serv
778 * has been created using svc_create_pooled().
780 * Based on code that used to be in nfsd_svc() but tweaked
784 svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
787 /* The -1 assumes caller has done a svc_get() */
788 nrservs -= (serv->sv_nrthreads-1);
790 spin_lock_bh(&pool->sp_lock);
791 nrservs -= pool->sp_nrthreads;
792 spin_unlock_bh(&pool->sp_lock);
796 return svc_start_kthreads(serv, pool, nrservs);
798 return svc_signal_kthreads(serv, pool, nrservs);
801 EXPORT_SYMBOL_GPL(svc_set_num_threads);
803 /* destroy old threads */
805 svc_stop_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
807 struct task_struct *task;
808 unsigned int state = serv->sv_nrthreads-1;
810 /* destroy old threads */
812 task = choose_victim(serv, pool, &state);
817 } while (nrservs < 0);
822 svc_set_num_threads_sync(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
825 /* The -1 assumes caller has done a svc_get() */
826 nrservs -= (serv->sv_nrthreads-1);
828 spin_lock_bh(&pool->sp_lock);
829 nrservs -= pool->sp_nrthreads;
830 spin_unlock_bh(&pool->sp_lock);
834 return svc_start_kthreads(serv, pool, nrservs);
836 return svc_stop_kthreads(serv, pool, nrservs);
839 EXPORT_SYMBOL_GPL(svc_set_num_threads_sync);
842 * Called from a server thread as it's exiting. Caller must hold the "service
843 * mutex" for the service.
846 svc_rqst_free(struct svc_rqst *rqstp)
848 svc_release_buffer(rqstp);
849 if (rqstp->rq_scratch_page)
850 put_page(rqstp->rq_scratch_page);
851 kfree(rqstp->rq_resp);
852 kfree(rqstp->rq_argp);
853 kfree(rqstp->rq_auth_data);
854 kfree_rcu(rqstp, rq_rcu_head);
856 EXPORT_SYMBOL_GPL(svc_rqst_free);
859 svc_exit_thread(struct svc_rqst *rqstp)
861 struct svc_serv *serv = rqstp->rq_server;
862 struct svc_pool *pool = rqstp->rq_pool;
864 spin_lock_bh(&pool->sp_lock);
865 pool->sp_nrthreads--;
866 if (!test_and_set_bit(RQ_VICTIM, &rqstp->rq_flags))
867 list_del_rcu(&rqstp->rq_all);
868 spin_unlock_bh(&pool->sp_lock);
870 svc_rqst_free(rqstp);
872 /* Release the server */
876 EXPORT_SYMBOL_GPL(svc_exit_thread);
879 * Register an "inet" protocol family netid with the local
880 * rpcbind daemon via an rpcbind v4 SET request.
882 * No netconfig infrastructure is available in the kernel, so
883 * we map IP_ protocol numbers to netids by hand.
885 * Returns zero on success; a negative errno value is returned
886 * if any error occurs.
888 static int __svc_rpcb_register4(struct net *net, const u32 program,
890 const unsigned short protocol,
891 const unsigned short port)
893 const struct sockaddr_in sin = {
894 .sin_family = AF_INET,
895 .sin_addr.s_addr = htonl(INADDR_ANY),
896 .sin_port = htons(port),
903 netid = RPCBIND_NETID_UDP;
906 netid = RPCBIND_NETID_TCP;
912 error = rpcb_v4_register(net, program, version,
913 (const struct sockaddr *)&sin, netid);
916 * User space didn't support rpcbind v4, so retry this
917 * registration request with the legacy rpcbind v2 protocol.
919 if (error == -EPROTONOSUPPORT)
920 error = rpcb_register(net, program, version, protocol, port);
925 #if IS_ENABLED(CONFIG_IPV6)
927 * Register an "inet6" protocol family netid with the local
928 * rpcbind daemon via an rpcbind v4 SET request.
930 * No netconfig infrastructure is available in the kernel, so
931 * we map IP_ protocol numbers to netids by hand.
933 * Returns zero on success; a negative errno value is returned
934 * if any error occurs.
936 static int __svc_rpcb_register6(struct net *net, const u32 program,
938 const unsigned short protocol,
939 const unsigned short port)
941 const struct sockaddr_in6 sin6 = {
942 .sin6_family = AF_INET6,
943 .sin6_addr = IN6ADDR_ANY_INIT,
944 .sin6_port = htons(port),
951 netid = RPCBIND_NETID_UDP6;
954 netid = RPCBIND_NETID_TCP6;
960 error = rpcb_v4_register(net, program, version,
961 (const struct sockaddr *)&sin6, netid);
964 * User space didn't support rpcbind version 4, so we won't
965 * use a PF_INET6 listener.
967 if (error == -EPROTONOSUPPORT)
968 error = -EAFNOSUPPORT;
972 #endif /* IS_ENABLED(CONFIG_IPV6) */
975 * Register a kernel RPC service via rpcbind version 4.
977 * Returns zero on success; a negative errno value is returned
978 * if any error occurs.
980 static int __svc_register(struct net *net, const char *progname,
981 const u32 program, const u32 version,
983 const unsigned short protocol,
984 const unsigned short port)
986 int error = -EAFNOSUPPORT;
990 error = __svc_rpcb_register4(net, program, version,
993 #if IS_ENABLED(CONFIG_IPV6)
995 error = __svc_rpcb_register6(net, program, version,
1000 trace_svc_register(progname, version, protocol, port, family, error);
1004 int svc_rpcbind_set_version(struct net *net,
1005 const struct svc_program *progp,
1006 u32 version, int family,
1007 unsigned short proto,
1008 unsigned short port)
1010 return __svc_register(net, progp->pg_name, progp->pg_prog,
1011 version, family, proto, port);
1014 EXPORT_SYMBOL_GPL(svc_rpcbind_set_version);
1016 int svc_generic_rpcbind_set(struct net *net,
1017 const struct svc_program *progp,
1018 u32 version, int family,
1019 unsigned short proto,
1020 unsigned short port)
1022 const struct svc_version *vers = progp->pg_vers[version];
1028 if (vers->vs_hidden) {
1029 trace_svc_noregister(progp->pg_name, version, proto,
1035 * Don't register a UDP port if we need congestion
1038 if (vers->vs_need_cong_ctrl && proto == IPPROTO_UDP)
1041 error = svc_rpcbind_set_version(net, progp, version,
1042 family, proto, port);
1044 return (vers->vs_rpcb_optnl) ? 0 : error;
1046 EXPORT_SYMBOL_GPL(svc_generic_rpcbind_set);
1049 * svc_register - register an RPC service with the local portmapper
1050 * @serv: svc_serv struct for the service to register
1051 * @net: net namespace for the service to register
1052 * @family: protocol family of service's listener socket
1053 * @proto: transport protocol number to advertise
1054 * @port: port to advertise
1056 * Service is registered for any address in the passed-in protocol family
1058 int svc_register(const struct svc_serv *serv, struct net *net,
1059 const int family, const unsigned short proto,
1060 const unsigned short port)
1062 struct svc_program *progp;
1066 WARN_ON_ONCE(proto == 0 && port == 0);
1067 if (proto == 0 && port == 0)
1070 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1071 for (i = 0; i < progp->pg_nvers; i++) {
1073 error = progp->pg_rpcbind_set(net, progp, i,
1074 family, proto, port);
1076 printk(KERN_WARNING "svc: failed to register "
1077 "%sv%u RPC service (errno %d).\n",
1078 progp->pg_name, i, -error);
1088 * If user space is running rpcbind, it should take the v4 UNSET
1089 * and clear everything for this [program, version]. If user space
1090 * is running portmap, it will reject the v4 UNSET, but won't have
1091 * any "inet6" entries anyway. So a PMAP_UNSET should be sufficient
1092 * in this case to clear all existing entries for [program, version].
1094 static void __svc_unregister(struct net *net, const u32 program, const u32 version,
1095 const char *progname)
1099 error = rpcb_v4_register(net, program, version, NULL, "");
1102 * User space didn't support rpcbind v4, so retry this
1103 * request with the legacy rpcbind v2 protocol.
1105 if (error == -EPROTONOSUPPORT)
1106 error = rpcb_register(net, program, version, 0, 0);
1108 trace_svc_unregister(progname, version, error);
1112 * All netids, bind addresses and ports registered for [program, version]
1113 * are removed from the local rpcbind database (if the service is not
1114 * hidden) to make way for a new instance of the service.
1116 * The result of unregistration is reported via dprintk for those who want
1117 * verification of the result, but is otherwise not important.
1119 static void svc_unregister(const struct svc_serv *serv, struct net *net)
1121 struct svc_program *progp;
1122 unsigned long flags;
1125 clear_thread_flag(TIF_SIGPENDING);
1127 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1128 for (i = 0; i < progp->pg_nvers; i++) {
1129 if (progp->pg_vers[i] == NULL)
1131 if (progp->pg_vers[i]->vs_hidden)
1133 __svc_unregister(net, progp->pg_prog, i, progp->pg_name);
1137 spin_lock_irqsave(¤t->sighand->siglock, flags);
1138 recalc_sigpending();
1139 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
1143 * dprintk the given error with the address of the client that caused it.
1145 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
1146 static __printf(2, 3)
1147 void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...)
1149 struct va_format vaf;
1151 char buf[RPC_MAX_ADDRBUFLEN];
1153 va_start(args, fmt);
1158 dprintk("svc: %s: %pV", svc_print_addr(rqstp, buf, sizeof(buf)), &vaf);
1163 static __printf(2,3) void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...) {}
1167 svc_return_autherr(struct svc_rqst *rqstp, __be32 auth_err)
1169 set_bit(RQ_AUTHERR, &rqstp->rq_flags);
1172 EXPORT_SYMBOL_GPL(svc_return_autherr);
1175 svc_get_autherr(struct svc_rqst *rqstp, __be32 *statp)
1177 if (test_and_clear_bit(RQ_AUTHERR, &rqstp->rq_flags))
1183 svc_generic_dispatch(struct svc_rqst *rqstp, __be32 *statp)
1185 struct kvec *argv = &rqstp->rq_arg.head[0];
1186 struct kvec *resv = &rqstp->rq_res.head[0];
1187 const struct svc_procedure *procp = rqstp->rq_procinfo;
1191 * XXX: why do we ignore the return value?
1193 if (procp->pc_decode &&
1194 !procp->pc_decode(rqstp, argv->iov_base)) {
1195 *statp = rpc_garbage_args;
1199 *statp = procp->pc_func(rqstp);
1201 if (*statp == rpc_drop_reply ||
1202 test_bit(RQ_DROPME, &rqstp->rq_flags))
1205 if (test_bit(RQ_AUTHERR, &rqstp->rq_flags))
1208 if (*statp != rpc_success)
1212 if (procp->pc_encode &&
1213 !procp->pc_encode(rqstp, resv->iov_base + resv->iov_len)) {
1214 dprintk("svc: failed to encode reply\n");
1215 /* serv->sv_stats->rpcsystemerr++; */
1216 *statp = rpc_system_err;
1222 svc_generic_init_request(struct svc_rqst *rqstp,
1223 const struct svc_program *progp,
1224 struct svc_process_info *ret)
1226 const struct svc_version *versp = NULL; /* compiler food */
1227 const struct svc_procedure *procp = NULL;
1229 if (rqstp->rq_vers >= progp->pg_nvers )
1231 versp = progp->pg_vers[rqstp->rq_vers];
1236 * Some protocol versions (namely NFSv4) require some form of
1237 * congestion control. (See RFC 7530 section 3.1 paragraph 2)
1238 * In other words, UDP is not allowed. We mark those when setting
1239 * up the svc_xprt, and verify that here.
1241 * The spec is not very clear about what error should be returned
1242 * when someone tries to access a server that is listening on UDP
1243 * for lower versions. RPC_PROG_MISMATCH seems to be the closest
1246 if (versp->vs_need_cong_ctrl && rqstp->rq_xprt &&
1247 !test_bit(XPT_CONG_CTRL, &rqstp->rq_xprt->xpt_flags))
1250 if (rqstp->rq_proc >= versp->vs_nproc)
1252 rqstp->rq_procinfo = procp = &versp->vs_proc[rqstp->rq_proc];
1256 /* Initialize storage for argp and resp */
1257 memset(rqstp->rq_argp, 0, procp->pc_argsize);
1258 memset(rqstp->rq_resp, 0, procp->pc_ressize);
1260 /* Bump per-procedure stats counter */
1261 versp->vs_count[rqstp->rq_proc]++;
1263 ret->dispatch = versp->vs_dispatch;
1266 ret->mismatch.lovers = progp->pg_lovers;
1267 ret->mismatch.hivers = progp->pg_hivers;
1268 return rpc_prog_mismatch;
1270 return rpc_proc_unavail;
1272 EXPORT_SYMBOL_GPL(svc_generic_init_request);
1275 * Common routine for processing the RPC request.
1278 svc_process_common(struct svc_rqst *rqstp, struct kvec *argv, struct kvec *resv)
1280 struct svc_program *progp;
1281 const struct svc_procedure *procp = NULL;
1282 struct svc_serv *serv = rqstp->rq_server;
1283 struct svc_process_info process;
1286 __be32 auth_stat, rpc_stat;
1288 __be32 *reply_statp;
1290 rpc_stat = rpc_success;
1292 if (argv->iov_len < 6*4)
1295 /* Will be turned off by GSS integrity and privacy services */
1296 set_bit(RQ_SPLICE_OK, &rqstp->rq_flags);
1297 /* Will be turned off only when NFSv4 Sessions are used */
1298 set_bit(RQ_USEDEFERRAL, &rqstp->rq_flags);
1299 clear_bit(RQ_DROPME, &rqstp->rq_flags);
1301 svc_putu32(resv, rqstp->rq_xid);
1303 vers = svc_getnl(argv);
1305 /* First words of reply: */
1306 svc_putnl(resv, 1); /* REPLY */
1308 if (vers != 2) /* RPC version number */
1311 /* Save position in case we later decide to reject: */
1312 reply_statp = resv->iov_base + resv->iov_len;
1314 svc_putnl(resv, 0); /* ACCEPT */
1316 rqstp->rq_prog = prog = svc_getnl(argv); /* program number */
1317 rqstp->rq_vers = svc_getnl(argv); /* version number */
1318 rqstp->rq_proc = svc_getnl(argv); /* procedure number */
1320 for (progp = serv->sv_program; progp; progp = progp->pg_next)
1321 if (prog == progp->pg_prog)
1325 * Decode auth data, and add verifier to reply buffer.
1326 * We do this before anything else in order to get a decent
1329 auth_res = svc_authenticate(rqstp, &auth_stat);
1330 /* Also give the program a chance to reject this call: */
1331 if (auth_res == SVC_OK && progp) {
1332 auth_stat = rpc_autherr_badcred;
1333 auth_res = progp->pg_authenticate(rqstp);
1335 if (auth_res != SVC_OK)
1336 trace_svc_authenticate(rqstp, auth_res, auth_stat);
1343 rpc_stat = rpc_system_err;
1358 rpc_stat = progp->pg_init_request(rqstp, progp, &process);
1362 case rpc_prog_unavail:
1364 case rpc_prog_mismatch:
1366 case rpc_proc_unavail:
1370 procp = rqstp->rq_procinfo;
1371 /* Should this check go into the dispatcher? */
1372 if (!procp || !procp->pc_func)
1375 /* Syntactic check complete */
1376 serv->sv_stats->rpccnt++;
1377 trace_svc_process(rqstp, progp->pg_name);
1379 /* Build the reply header. */
1380 statp = resv->iov_base +resv->iov_len;
1381 svc_putnl(resv, RPC_SUCCESS);
1383 /* un-reserve some of the out-queue now that we have a
1384 * better idea of reply size
1386 if (procp->pc_xdrressize)
1387 svc_reserve_auth(rqstp, procp->pc_xdrressize<<2);
1389 /* Call the function that processes the request. */
1390 if (!process.dispatch) {
1391 if (!svc_generic_dispatch(rqstp, statp))
1392 goto release_dropit;
1393 if (*statp == rpc_garbage_args)
1395 auth_stat = svc_get_autherr(rqstp, statp);
1396 if (auth_stat != rpc_auth_ok)
1397 goto err_release_bad_auth;
1399 dprintk("svc: calling dispatcher\n");
1400 if (!process.dispatch(rqstp, statp))
1401 goto release_dropit; /* Release reply info */
1404 /* Check RPC status result */
1405 if (*statp != rpc_success)
1406 resv->iov_len = ((void*)statp) - resv->iov_base + 4;
1408 /* Release reply info */
1409 if (procp->pc_release)
1410 procp->pc_release(rqstp);
1412 if (procp->pc_encode == NULL)
1416 if (svc_authorise(rqstp))
1418 return 1; /* Caller can now send it */
1421 if (procp->pc_release)
1422 procp->pc_release(rqstp);
1424 svc_authorise(rqstp); /* doesn't hurt to call this twice */
1425 dprintk("svc: svc_process dropit\n");
1429 svc_authorise(rqstp);
1431 if (rqstp->rq_xprt && test_bit(XPT_TEMP, &rqstp->rq_xprt->xpt_flags))
1432 svc_close_xprt(rqstp->rq_xprt);
1433 dprintk("svc: svc_process close\n");
1437 svc_printk(rqstp, "short len %zd, dropping request\n",
1442 serv->sv_stats->rpcbadfmt++;
1443 svc_putnl(resv, 1); /* REJECT */
1444 svc_putnl(resv, 0); /* RPC_MISMATCH */
1445 svc_putnl(resv, 2); /* Only RPCv2 supported */
1449 err_release_bad_auth:
1450 if (procp->pc_release)
1451 procp->pc_release(rqstp);
1453 dprintk("svc: authentication failed (%d)\n", ntohl(auth_stat));
1454 serv->sv_stats->rpcbadauth++;
1455 /* Restore write pointer to location of accept status: */
1456 xdr_ressize_check(rqstp, reply_statp);
1457 svc_putnl(resv, 1); /* REJECT */
1458 svc_putnl(resv, 1); /* AUTH_ERROR */
1459 svc_putnl(resv, ntohl(auth_stat)); /* status */
1463 dprintk("svc: unknown program %d\n", prog);
1464 serv->sv_stats->rpcbadfmt++;
1465 svc_putnl(resv, RPC_PROG_UNAVAIL);
1469 svc_printk(rqstp, "unknown version (%d for prog %d, %s)\n",
1470 rqstp->rq_vers, rqstp->rq_prog, progp->pg_name);
1472 serv->sv_stats->rpcbadfmt++;
1473 svc_putnl(resv, RPC_PROG_MISMATCH);
1474 svc_putnl(resv, process.mismatch.lovers);
1475 svc_putnl(resv, process.mismatch.hivers);
1479 svc_printk(rqstp, "unknown procedure (%d)\n", rqstp->rq_proc);
1481 serv->sv_stats->rpcbadfmt++;
1482 svc_putnl(resv, RPC_PROC_UNAVAIL);
1486 svc_printk(rqstp, "failed to decode args\n");
1488 rpc_stat = rpc_garbage_args;
1490 serv->sv_stats->rpcbadfmt++;
1491 svc_putnl(resv, ntohl(rpc_stat));
1496 * Process the RPC request.
1499 svc_process(struct svc_rqst *rqstp)
1501 struct kvec *argv = &rqstp->rq_arg.head[0];
1502 struct kvec *resv = &rqstp->rq_res.head[0];
1503 struct svc_serv *serv = rqstp->rq_server;
1507 * Setup response xdr_buf.
1508 * Initially it has just one page
1510 rqstp->rq_next_page = &rqstp->rq_respages[1];
1511 resv->iov_base = page_address(rqstp->rq_respages[0]);
1513 rqstp->rq_res.pages = rqstp->rq_respages + 1;
1514 rqstp->rq_res.len = 0;
1515 rqstp->rq_res.page_base = 0;
1516 rqstp->rq_res.page_len = 0;
1517 rqstp->rq_res.buflen = PAGE_SIZE;
1518 rqstp->rq_res.tail[0].iov_base = NULL;
1519 rqstp->rq_res.tail[0].iov_len = 0;
1521 dir = svc_getnl(argv);
1523 /* direction != CALL */
1524 svc_printk(rqstp, "bad direction %d, dropping request\n", dir);
1525 serv->sv_stats->rpcbadfmt++;
1529 /* Returns 1 for send, 0 for drop */
1530 if (likely(svc_process_common(rqstp, argv, resv)))
1531 return svc_send(rqstp);
1537 EXPORT_SYMBOL_GPL(svc_process);
1539 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1541 * Process a backchannel RPC request that arrived over an existing
1542 * outbound connection
1545 bc_svc_process(struct svc_serv *serv, struct rpc_rqst *req,
1546 struct svc_rqst *rqstp)
1548 struct kvec *argv = &rqstp->rq_arg.head[0];
1549 struct kvec *resv = &rqstp->rq_res.head[0];
1550 struct rpc_task *task;
1554 dprintk("svc: %s(%p)\n", __func__, req);
1556 /* Build the svc_rqst used by the common processing routine */
1557 rqstp->rq_xid = req->rq_xid;
1558 rqstp->rq_prot = req->rq_xprt->prot;
1559 rqstp->rq_server = serv;
1560 rqstp->rq_bc_net = req->rq_xprt->xprt_net;
1562 rqstp->rq_addrlen = sizeof(req->rq_xprt->addr);
1563 memcpy(&rqstp->rq_addr, &req->rq_xprt->addr, rqstp->rq_addrlen);
1564 memcpy(&rqstp->rq_arg, &req->rq_rcv_buf, sizeof(rqstp->rq_arg));
1565 memcpy(&rqstp->rq_res, &req->rq_snd_buf, sizeof(rqstp->rq_res));
1567 /* Adjust the argument buffer length */
1568 rqstp->rq_arg.len = req->rq_private_buf.len;
1569 if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len) {
1570 rqstp->rq_arg.head[0].iov_len = rqstp->rq_arg.len;
1571 rqstp->rq_arg.page_len = 0;
1572 } else if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len +
1573 rqstp->rq_arg.page_len)
1574 rqstp->rq_arg.page_len = rqstp->rq_arg.len -
1575 rqstp->rq_arg.head[0].iov_len;
1577 rqstp->rq_arg.len = rqstp->rq_arg.head[0].iov_len +
1578 rqstp->rq_arg.page_len;
1580 /* reset result send buffer "put" position */
1584 * Skip the next two words because they've already been
1585 * processed in the transport
1587 svc_getu32(argv); /* XID */
1588 svc_getnl(argv); /* CALLDIR */
1590 /* Parse and execute the bc call */
1591 proc_error = svc_process_common(rqstp, argv, resv);
1593 atomic_dec(&req->rq_xprt->bc_slot_count);
1595 /* Processing error: drop the request */
1596 xprt_free_bc_request(req);
1600 /* Finally, send the reply synchronously */
1601 memcpy(&req->rq_snd_buf, &rqstp->rq_res, sizeof(req->rq_snd_buf));
1602 task = rpc_run_bc_task(req);
1604 error = PTR_ERR(task);
1608 WARN_ON_ONCE(atomic_read(&task->tk_count) != 1);
1609 error = task->tk_status;
1613 dprintk("svc: %s(), error=%d\n", __func__, error);
1616 EXPORT_SYMBOL_GPL(bc_svc_process);
1617 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1620 * Return (transport-specific) limit on the rpc payload.
1622 u32 svc_max_payload(const struct svc_rqst *rqstp)
1624 u32 max = rqstp->rq_xprt->xpt_class->xcl_max_payload;
1626 if (rqstp->rq_server->sv_max_payload < max)
1627 max = rqstp->rq_server->sv_max_payload;
1630 EXPORT_SYMBOL_GPL(svc_max_payload);
1633 * svc_encode_result_payload - mark a range of bytes as a result payload
1634 * @rqstp: svc_rqst to operate on
1635 * @offset: payload's byte offset in rqstp->rq_res
1636 * @length: size of payload, in bytes
1638 * Returns zero on success, or a negative errno if a permanent
1641 int svc_encode_result_payload(struct svc_rqst *rqstp, unsigned int offset,
1642 unsigned int length)
1644 return rqstp->rq_xprt->xpt_ops->xpo_result_payload(rqstp, offset,
1647 EXPORT_SYMBOL_GPL(svc_encode_result_payload);
1650 * svc_fill_write_vector - Construct data argument for VFS write call
1651 * @rqstp: svc_rqst to operate on
1652 * @pages: list of pages containing data payload
1653 * @first: buffer containing first section of write payload
1654 * @total: total number of bytes of write payload
1656 * Fills in rqstp::rq_vec, and returns the number of elements.
1658 unsigned int svc_fill_write_vector(struct svc_rqst *rqstp, struct page **pages,
1659 struct kvec *first, size_t total)
1661 struct kvec *vec = rqstp->rq_vec;
1664 /* Some types of transport can present the write payload
1665 * entirely in rq_arg.pages. In this case, @first is empty.
1668 if (first->iov_len) {
1669 vec[i].iov_base = first->iov_base;
1670 vec[i].iov_len = min_t(size_t, total, first->iov_len);
1671 total -= vec[i].iov_len;
1676 vec[i].iov_base = page_address(*pages);
1677 vec[i].iov_len = min_t(size_t, total, PAGE_SIZE);
1678 total -= vec[i].iov_len;
1683 WARN_ON_ONCE(i > ARRAY_SIZE(rqstp->rq_vec));
1686 EXPORT_SYMBOL_GPL(svc_fill_write_vector);
1689 * svc_fill_symlink_pathname - Construct pathname argument for VFS symlink call
1690 * @rqstp: svc_rqst to operate on
1691 * @first: buffer containing first section of pathname
1692 * @p: buffer containing remaining section of pathname
1693 * @total: total length of the pathname argument
1695 * The VFS symlink API demands a NUL-terminated pathname in mapped memory.
1696 * Returns pointer to a NUL-terminated string, or an ERR_PTR. Caller must free
1697 * the returned string.
1699 char *svc_fill_symlink_pathname(struct svc_rqst *rqstp, struct kvec *first,
1700 void *p, size_t total)
1702 size_t len, remaining;
1705 result = kmalloc(total + 1, GFP_KERNEL);
1707 return ERR_PTR(-ESERVERFAULT);
1712 len = min_t(size_t, total, first->iov_len);
1714 memcpy(dst, first->iov_base, len);
1720 len = min_t(size_t, remaining, PAGE_SIZE);
1721 memcpy(dst, p, len);
1727 /* Sanity check: Linux doesn't allow the pathname argument to
1728 * contain a NUL byte.
1730 if (strlen(result) != total) {
1732 return ERR_PTR(-EINVAL);
1736 EXPORT_SYMBOL_GPL(svc_fill_symlink_pathname);