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>
36 #define RPCDBG_FACILITY RPCDBG_SVCDSP
38 static void svc_unregister(const struct svc_serv *serv, struct net *net);
40 #define SVC_POOL_DEFAULT SVC_POOL_GLOBAL
43 * Mode for mapping cpus to pools.
46 SVC_POOL_AUTO = -1, /* choose one of the others */
47 SVC_POOL_GLOBAL, /* no mapping, just a single global pool
48 * (legacy & UP mode) */
49 SVC_POOL_PERCPU, /* one pool per cpu */
50 SVC_POOL_PERNODE /* one pool per numa node */
54 * Structure for mapping cpus to pools and vice versa.
55 * Setup once during sunrpc initialisation.
59 int count; /* How many svc_servs use us */
60 int mode; /* Note: int not enum to avoid
61 * warnings about "enumeration value
62 * not handled in switch" */
64 unsigned int *pool_to; /* maps pool id to cpu or node */
65 unsigned int *to_pool; /* maps cpu or node to pool id */
68 static struct svc_pool_map svc_pool_map = {
69 .mode = SVC_POOL_DEFAULT
72 static DEFINE_MUTEX(svc_pool_map_mutex);/* protects svc_pool_map.count only */
75 param_set_pool_mode(const char *val, const struct kernel_param *kp)
77 int *ip = (int *)kp->arg;
78 struct svc_pool_map *m = &svc_pool_map;
81 mutex_lock(&svc_pool_map_mutex);
88 if (!strncmp(val, "auto", 4))
90 else if (!strncmp(val, "global", 6))
91 *ip = SVC_POOL_GLOBAL;
92 else if (!strncmp(val, "percpu", 6))
93 *ip = SVC_POOL_PERCPU;
94 else if (!strncmp(val, "pernode", 7))
95 *ip = SVC_POOL_PERNODE;
100 mutex_unlock(&svc_pool_map_mutex);
105 param_get_pool_mode(char *buf, const struct kernel_param *kp)
107 int *ip = (int *)kp->arg;
112 return strlcpy(buf, "auto\n", 20);
113 case SVC_POOL_GLOBAL:
114 return strlcpy(buf, "global\n", 20);
115 case SVC_POOL_PERCPU:
116 return strlcpy(buf, "percpu\n", 20);
117 case SVC_POOL_PERNODE:
118 return strlcpy(buf, "pernode\n", 20);
120 return sprintf(buf, "%d\n", *ip);
124 module_param_call(pool_mode, param_set_pool_mode, param_get_pool_mode,
125 &svc_pool_map.mode, 0644);
128 * Detect best pool mapping mode heuristically,
129 * according to the machine's topology.
132 svc_pool_map_choose_mode(void)
136 if (nr_online_nodes > 1) {
138 * Actually have multiple NUMA nodes,
139 * so split pools on NUMA node boundaries
141 return SVC_POOL_PERNODE;
144 node = first_online_node;
145 if (nr_cpus_node(node) > 2) {
147 * Non-trivial SMP, or CONFIG_NUMA on
148 * non-NUMA hardware, e.g. with a generic
149 * x86_64 kernel on Xeons. In this case we
150 * want to divide the pools on cpu boundaries.
152 return SVC_POOL_PERCPU;
155 /* default: one global pool */
156 return SVC_POOL_GLOBAL;
160 * Allocate the to_pool[] and pool_to[] arrays.
161 * Returns 0 on success or an errno.
164 svc_pool_map_alloc_arrays(struct svc_pool_map *m, unsigned int maxpools)
166 m->to_pool = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
169 m->pool_to = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
183 * Initialise the pool map for SVC_POOL_PERCPU mode.
184 * Returns number of pools or <0 on error.
187 svc_pool_map_init_percpu(struct svc_pool_map *m)
189 unsigned int maxpools = nr_cpu_ids;
190 unsigned int pidx = 0;
194 err = svc_pool_map_alloc_arrays(m, maxpools);
198 for_each_online_cpu(cpu) {
199 BUG_ON(pidx >= maxpools);
200 m->to_pool[cpu] = pidx;
201 m->pool_to[pidx] = cpu;
204 /* cpus brought online later all get mapped to pool0, sorry */
211 * Initialise the pool map for SVC_POOL_PERNODE mode.
212 * Returns number of pools or <0 on error.
215 svc_pool_map_init_pernode(struct svc_pool_map *m)
217 unsigned int maxpools = nr_node_ids;
218 unsigned int pidx = 0;
222 err = svc_pool_map_alloc_arrays(m, maxpools);
226 for_each_node_with_cpus(node) {
227 /* some architectures (e.g. SN2) have cpuless nodes */
228 BUG_ON(pidx > maxpools);
229 m->to_pool[node] = pidx;
230 m->pool_to[pidx] = node;
233 /* nodes brought online later all get mapped to pool0, sorry */
240 * Add a reference to the global map of cpus to pools (and
241 * vice versa) if pools are in use.
242 * Initialise the map if we're the first user.
243 * Returns the number of pools. If this is '1', no reference
247 svc_pool_map_get(void)
249 struct svc_pool_map *m = &svc_pool_map;
252 mutex_lock(&svc_pool_map_mutex);
255 mutex_unlock(&svc_pool_map_mutex);
256 WARN_ON_ONCE(m->npools <= 1);
260 if (m->mode == SVC_POOL_AUTO)
261 m->mode = svc_pool_map_choose_mode();
264 case SVC_POOL_PERCPU:
265 npools = svc_pool_map_init_percpu(m);
267 case SVC_POOL_PERNODE:
268 npools = svc_pool_map_init_pernode(m);
273 /* default, or memory allocation failure */
275 m->mode = SVC_POOL_GLOBAL;
280 /* service is unpooled, so doesn't hold a reference */
283 mutex_unlock(&svc_pool_map_mutex);
288 * Drop a reference to the global map of cpus to pools, if
289 * pools were in use, i.e. if npools > 1.
290 * When the last reference is dropped, the map data is
291 * freed; this allows the sysadmin to change the pool
292 * mode using the pool_mode module option without
293 * rebooting or re-loading sunrpc.ko.
296 svc_pool_map_put(int npools)
298 struct svc_pool_map *m = &svc_pool_map;
302 mutex_lock(&svc_pool_map_mutex);
312 mutex_unlock(&svc_pool_map_mutex);
315 static int svc_pool_map_get_node(unsigned int pidx)
317 const struct svc_pool_map *m = &svc_pool_map;
320 if (m->mode == SVC_POOL_PERCPU)
321 return cpu_to_node(m->pool_to[pidx]);
322 if (m->mode == SVC_POOL_PERNODE)
323 return m->pool_to[pidx];
328 * Set the given thread's cpus_allowed mask so that it
329 * will only run on cpus in the given pool.
332 svc_pool_map_set_cpumask(struct task_struct *task, unsigned int pidx)
334 struct svc_pool_map *m = &svc_pool_map;
335 unsigned int node = m->pool_to[pidx];
338 * The caller checks for sv_nrpools > 1, which
339 * implies that we've been initialized.
341 WARN_ON_ONCE(m->count == 0);
346 case SVC_POOL_PERCPU:
348 set_cpus_allowed_ptr(task, cpumask_of(node));
351 case SVC_POOL_PERNODE:
353 set_cpus_allowed_ptr(task, cpumask_of_node(node));
360 * Use the mapping mode to choose a pool for a given CPU.
361 * Used when enqueueing an incoming RPC. Always returns
362 * a non-NULL pool pointer.
365 svc_pool_for_cpu(struct svc_serv *serv, int cpu)
367 struct svc_pool_map *m = &svc_pool_map;
368 unsigned int pidx = 0;
370 if (serv->sv_nrpools <= 1)
371 return serv->sv_pools;
374 case SVC_POOL_PERCPU:
375 pidx = m->to_pool[cpu];
377 case SVC_POOL_PERNODE:
378 pidx = m->to_pool[cpu_to_node(cpu)];
382 return &serv->sv_pools[pidx % serv->sv_nrpools];
385 int svc_rpcb_setup(struct svc_serv *serv, struct net *net)
389 err = rpcb_create_local(net);
393 /* Remove any stale portmap registrations */
394 svc_unregister(serv, net);
397 EXPORT_SYMBOL_GPL(svc_rpcb_setup);
399 void svc_rpcb_cleanup(struct svc_serv *serv, struct net *net)
401 svc_unregister(serv, net);
404 EXPORT_SYMBOL_GPL(svc_rpcb_cleanup);
406 static int svc_uses_rpcbind(struct svc_serv *serv)
408 struct svc_program *progp;
411 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
412 for (i = 0; i < progp->pg_nvers; i++) {
413 if (progp->pg_vers[i] == NULL)
415 if (!progp->pg_vers[i]->vs_hidden)
423 int svc_bind(struct svc_serv *serv, struct net *net)
425 if (!svc_uses_rpcbind(serv))
427 return svc_rpcb_setup(serv, net);
429 EXPORT_SYMBOL_GPL(svc_bind);
431 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
433 __svc_init_bc(struct svc_serv *serv)
435 INIT_LIST_HEAD(&serv->sv_cb_list);
436 spin_lock_init(&serv->sv_cb_lock);
437 init_waitqueue_head(&serv->sv_cb_waitq);
441 __svc_init_bc(struct svc_serv *serv)
447 * Create an RPC service
449 static struct svc_serv *
450 __svc_create(struct svc_program *prog, unsigned int bufsize, int npools,
451 int (*threadfn)(void *data))
453 struct svc_serv *serv;
455 unsigned int xdrsize;
458 if (!(serv = kzalloc(sizeof(*serv), GFP_KERNEL)))
460 serv->sv_name = prog->pg_name;
461 serv->sv_program = prog;
462 kref_init(&serv->sv_refcnt);
463 serv->sv_stats = prog->pg_stats;
464 if (bufsize > RPCSVC_MAXPAYLOAD)
465 bufsize = RPCSVC_MAXPAYLOAD;
466 serv->sv_max_payload = bufsize? bufsize : 4096;
467 serv->sv_max_mesg = roundup(serv->sv_max_payload + PAGE_SIZE, PAGE_SIZE);
468 serv->sv_threadfn = threadfn;
471 prog->pg_lovers = prog->pg_nvers-1;
472 for (vers=0; vers<prog->pg_nvers ; vers++)
473 if (prog->pg_vers[vers]) {
474 prog->pg_hivers = vers;
475 if (prog->pg_lovers > vers)
476 prog->pg_lovers = vers;
477 if (prog->pg_vers[vers]->vs_xdrsize > xdrsize)
478 xdrsize = prog->pg_vers[vers]->vs_xdrsize;
480 prog = prog->pg_next;
482 serv->sv_xdrsize = xdrsize;
483 INIT_LIST_HEAD(&serv->sv_tempsocks);
484 INIT_LIST_HEAD(&serv->sv_permsocks);
485 timer_setup(&serv->sv_temptimer, NULL, 0);
486 spin_lock_init(&serv->sv_lock);
490 serv->sv_nrpools = npools;
492 kcalloc(serv->sv_nrpools, sizeof(struct svc_pool),
494 if (!serv->sv_pools) {
499 for (i = 0; i < serv->sv_nrpools; i++) {
500 struct svc_pool *pool = &serv->sv_pools[i];
502 dprintk("svc: initialising pool %u for %s\n",
506 INIT_LIST_HEAD(&pool->sp_sockets);
507 INIT_LIST_HEAD(&pool->sp_all_threads);
508 spin_lock_init(&pool->sp_lock);
515 * svc_create - Create an RPC service
516 * @prog: the RPC program the new service will handle
517 * @bufsize: maximum message size for @prog
518 * @threadfn: a function to service RPC requests for @prog
520 * Returns an instantiated struct svc_serv object or NULL.
522 struct svc_serv *svc_create(struct svc_program *prog, unsigned int bufsize,
523 int (*threadfn)(void *data))
525 return __svc_create(prog, bufsize, 1, threadfn);
527 EXPORT_SYMBOL_GPL(svc_create);
530 * svc_create_pooled - Create an RPC service with pooled threads
531 * @prog: the RPC program the new service will handle
532 * @bufsize: maximum message size for @prog
533 * @threadfn: a function to service RPC requests for @prog
535 * Returns an instantiated struct svc_serv object or NULL.
537 struct svc_serv *svc_create_pooled(struct svc_program *prog,
538 unsigned int bufsize,
539 int (*threadfn)(void *data))
541 struct svc_serv *serv;
542 unsigned int npools = svc_pool_map_get();
544 serv = __svc_create(prog, bufsize, npools, threadfn);
549 svc_pool_map_put(npools);
552 EXPORT_SYMBOL_GPL(svc_create_pooled);
555 * Destroy an RPC service. Should be called with appropriate locking to
556 * protect sv_permsocks and sv_tempsocks.
559 svc_destroy(struct kref *ref)
561 struct svc_serv *serv = container_of(ref, struct svc_serv, sv_refcnt);
563 dprintk("svc: svc_destroy(%s)\n", serv->sv_program->pg_name);
564 del_timer_sync(&serv->sv_temptimer);
567 * The last user is gone and thus all sockets have to be destroyed to
568 * the point. Check this.
570 BUG_ON(!list_empty(&serv->sv_permsocks));
571 BUG_ON(!list_empty(&serv->sv_tempsocks));
573 cache_clean_deferred(serv);
575 svc_pool_map_put(serv->sv_nrpools);
577 kfree(serv->sv_pools);
580 EXPORT_SYMBOL_GPL(svc_destroy);
583 * Allocate an RPC server's buffer space.
584 * We allocate pages and place them in rq_pages.
587 svc_init_buffer(struct svc_rqst *rqstp, unsigned int size, int node)
589 unsigned int pages, arghi;
591 /* bc_xprt uses fore channel allocated buffers */
592 if (svc_is_backchannel(rqstp))
595 pages = size / PAGE_SIZE + 1; /* extra page as we hold both request and reply.
596 * We assume one is at most one page
599 WARN_ON_ONCE(pages > RPCSVC_MAXPAGES);
600 if (pages > RPCSVC_MAXPAGES)
601 pages = RPCSVC_MAXPAGES;
603 struct page *p = alloc_pages_node(node, GFP_KERNEL, 0);
606 rqstp->rq_pages[arghi++] = p;
613 * Release an RPC server buffer
616 svc_release_buffer(struct svc_rqst *rqstp)
620 for (i = 0; i < ARRAY_SIZE(rqstp->rq_pages); i++)
621 if (rqstp->rq_pages[i])
622 put_page(rqstp->rq_pages[i]);
626 svc_rqst_alloc(struct svc_serv *serv, struct svc_pool *pool, int node)
628 struct svc_rqst *rqstp;
630 rqstp = kzalloc_node(sizeof(*rqstp), GFP_KERNEL, node);
634 __set_bit(RQ_BUSY, &rqstp->rq_flags);
635 spin_lock_init(&rqstp->rq_lock);
636 rqstp->rq_server = serv;
637 rqstp->rq_pool = pool;
639 rqstp->rq_scratch_page = alloc_pages_node(node, GFP_KERNEL, 0);
640 if (!rqstp->rq_scratch_page)
643 rqstp->rq_argp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
647 rqstp->rq_resp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
651 if (!svc_init_buffer(rqstp, serv->sv_max_mesg, node))
656 svc_rqst_free(rqstp);
659 EXPORT_SYMBOL_GPL(svc_rqst_alloc);
661 static struct svc_rqst *
662 svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
664 struct svc_rqst *rqstp;
666 rqstp = svc_rqst_alloc(serv, pool, node);
668 return ERR_PTR(-ENOMEM);
671 spin_lock_bh(&serv->sv_lock);
672 serv->sv_nrthreads += 1;
673 spin_unlock_bh(&serv->sv_lock);
675 spin_lock_bh(&pool->sp_lock);
676 pool->sp_nrthreads++;
677 list_add_rcu(&rqstp->rq_all, &pool->sp_all_threads);
678 spin_unlock_bh(&pool->sp_lock);
683 * Choose a pool in which to create a new thread, for svc_set_num_threads
685 static inline struct svc_pool *
686 choose_pool(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
691 return &serv->sv_pools[(*state)++ % serv->sv_nrpools];
695 * Choose a thread to kill, for svc_set_num_threads
697 static inline struct task_struct *
698 choose_victim(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
701 struct task_struct *task = NULL;
704 spin_lock_bh(&pool->sp_lock);
706 /* choose a pool in round-robin fashion */
707 for (i = 0; i < serv->sv_nrpools; i++) {
708 pool = &serv->sv_pools[--(*state) % serv->sv_nrpools];
709 spin_lock_bh(&pool->sp_lock);
710 if (!list_empty(&pool->sp_all_threads))
712 spin_unlock_bh(&pool->sp_lock);
718 if (!list_empty(&pool->sp_all_threads)) {
719 struct svc_rqst *rqstp;
722 * Remove from the pool->sp_all_threads list
723 * so we don't try to kill it again.
725 rqstp = list_entry(pool->sp_all_threads.next, struct svc_rqst, rq_all);
726 set_bit(RQ_VICTIM, &rqstp->rq_flags);
727 list_del_rcu(&rqstp->rq_all);
728 task = rqstp->rq_task;
730 spin_unlock_bh(&pool->sp_lock);
735 /* create new threads */
737 svc_start_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
739 struct svc_rqst *rqstp;
740 struct task_struct *task;
741 struct svc_pool *chosen_pool;
742 unsigned int state = serv->sv_nrthreads-1;
747 chosen_pool = choose_pool(serv, pool, &state);
749 node = svc_pool_map_get_node(chosen_pool->sp_id);
750 rqstp = svc_prepare_thread(serv, chosen_pool, node);
752 return PTR_ERR(rqstp);
754 task = kthread_create_on_node(serv->sv_threadfn, rqstp,
755 node, "%s", serv->sv_name);
757 svc_exit_thread(rqstp);
758 return PTR_ERR(task);
761 rqstp->rq_task = task;
762 if (serv->sv_nrpools > 1)
763 svc_pool_map_set_cpumask(task, chosen_pool->sp_id);
765 svc_sock_update_bufs(serv);
766 wake_up_process(task);
767 } while (nrservs > 0);
773 * Create or destroy enough new threads to make the number
774 * of threads the given number. If `pool' is non-NULL, applies
775 * only to threads in that pool, otherwise round-robins between
776 * all pools. Caller must ensure that mutual exclusion between this and
777 * server startup or shutdown.
780 /* destroy old threads */
782 svc_stop_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
784 struct task_struct *task;
785 unsigned int state = serv->sv_nrthreads-1;
787 /* destroy old threads */
789 task = choose_victim(serv, pool, &state);
794 } while (nrservs < 0);
799 svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
802 nrservs -= serv->sv_nrthreads;
804 spin_lock_bh(&pool->sp_lock);
805 nrservs -= pool->sp_nrthreads;
806 spin_unlock_bh(&pool->sp_lock);
810 return svc_start_kthreads(serv, pool, nrservs);
812 return svc_stop_kthreads(serv, pool, nrservs);
815 EXPORT_SYMBOL_GPL(svc_set_num_threads);
818 * svc_rqst_replace_page - Replace one page in rq_pages[]
819 * @rqstp: svc_rqst with pages to replace
820 * @page: replacement page
822 * When replacing a page in rq_pages, batch the release of the
823 * replaced pages to avoid hammering the page allocator.
825 void svc_rqst_replace_page(struct svc_rqst *rqstp, struct page *page)
827 if (*rqstp->rq_next_page) {
828 if (!pagevec_space(&rqstp->rq_pvec))
829 __pagevec_release(&rqstp->rq_pvec);
830 pagevec_add(&rqstp->rq_pvec, *rqstp->rq_next_page);
834 *(rqstp->rq_next_page++) = page;
836 EXPORT_SYMBOL_GPL(svc_rqst_replace_page);
839 * Called from a server thread as it's exiting. Caller must hold the "service
840 * mutex" for the service.
843 svc_rqst_free(struct svc_rqst *rqstp)
845 svc_release_buffer(rqstp);
846 if (rqstp->rq_scratch_page)
847 put_page(rqstp->rq_scratch_page);
848 kfree(rqstp->rq_resp);
849 kfree(rqstp->rq_argp);
850 kfree(rqstp->rq_auth_data);
851 kfree_rcu(rqstp, rq_rcu_head);
853 EXPORT_SYMBOL_GPL(svc_rqst_free);
856 svc_exit_thread(struct svc_rqst *rqstp)
858 struct svc_serv *serv = rqstp->rq_server;
859 struct svc_pool *pool = rqstp->rq_pool;
861 spin_lock_bh(&pool->sp_lock);
862 pool->sp_nrthreads--;
863 if (!test_and_set_bit(RQ_VICTIM, &rqstp->rq_flags))
864 list_del_rcu(&rqstp->rq_all);
865 spin_unlock_bh(&pool->sp_lock);
867 spin_lock_bh(&serv->sv_lock);
868 serv->sv_nrthreads -= 1;
869 spin_unlock_bh(&serv->sv_lock);
870 svc_sock_update_bufs(serv);
872 svc_rqst_free(rqstp);
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_generic_init_request(struct svc_rqst *rqstp,
1168 const struct svc_program *progp,
1169 struct svc_process_info *ret)
1171 const struct svc_version *versp = NULL; /* compiler food */
1172 const struct svc_procedure *procp = NULL;
1174 if (rqstp->rq_vers >= progp->pg_nvers )
1176 versp = progp->pg_vers[rqstp->rq_vers];
1181 * Some protocol versions (namely NFSv4) require some form of
1182 * congestion control. (See RFC 7530 section 3.1 paragraph 2)
1183 * In other words, UDP is not allowed. We mark those when setting
1184 * up the svc_xprt, and verify that here.
1186 * The spec is not very clear about what error should be returned
1187 * when someone tries to access a server that is listening on UDP
1188 * for lower versions. RPC_PROG_MISMATCH seems to be the closest
1191 if (versp->vs_need_cong_ctrl && rqstp->rq_xprt &&
1192 !test_bit(XPT_CONG_CTRL, &rqstp->rq_xprt->xpt_flags))
1195 if (rqstp->rq_proc >= versp->vs_nproc)
1197 rqstp->rq_procinfo = procp = &versp->vs_proc[rqstp->rq_proc];
1201 /* Initialize storage for argp and resp */
1202 memset(rqstp->rq_argp, 0, procp->pc_argsize);
1203 memset(rqstp->rq_resp, 0, procp->pc_ressize);
1205 /* Bump per-procedure stats counter */
1206 versp->vs_count[rqstp->rq_proc]++;
1208 ret->dispatch = versp->vs_dispatch;
1211 ret->mismatch.lovers = progp->pg_lovers;
1212 ret->mismatch.hivers = progp->pg_hivers;
1213 return rpc_prog_mismatch;
1215 return rpc_proc_unavail;
1217 EXPORT_SYMBOL_GPL(svc_generic_init_request);
1220 * Common routine for processing the RPC request.
1223 svc_process_common(struct svc_rqst *rqstp, struct kvec *argv, struct kvec *resv)
1225 struct svc_program *progp;
1226 const struct svc_procedure *procp = NULL;
1227 struct svc_serv *serv = rqstp->rq_server;
1228 struct svc_process_info process;
1233 __be32 *reply_statp;
1235 rpc_stat = rpc_success;
1237 if (argv->iov_len < 6*4)
1240 /* Will be turned off by GSS integrity and privacy services */
1241 set_bit(RQ_SPLICE_OK, &rqstp->rq_flags);
1242 /* Will be turned off only when NFSv4 Sessions are used */
1243 set_bit(RQ_USEDEFERRAL, &rqstp->rq_flags);
1244 clear_bit(RQ_DROPME, &rqstp->rq_flags);
1246 svc_putu32(resv, rqstp->rq_xid);
1248 vers = svc_getnl(argv);
1250 /* First words of reply: */
1251 svc_putnl(resv, 1); /* REPLY */
1253 if (vers != 2) /* RPC version number */
1256 /* Save position in case we later decide to reject: */
1257 reply_statp = resv->iov_base + resv->iov_len;
1259 svc_putnl(resv, 0); /* ACCEPT */
1261 rqstp->rq_prog = prog = svc_getnl(argv); /* program number */
1262 rqstp->rq_vers = svc_getnl(argv); /* version number */
1263 rqstp->rq_proc = svc_getnl(argv); /* procedure number */
1265 for (progp = serv->sv_program; progp; progp = progp->pg_next)
1266 if (prog == progp->pg_prog)
1270 * Decode auth data, and add verifier to reply buffer.
1271 * We do this before anything else in order to get a decent
1274 auth_res = svc_authenticate(rqstp);
1275 /* Also give the program a chance to reject this call: */
1276 if (auth_res == SVC_OK && progp)
1277 auth_res = progp->pg_authenticate(rqstp);
1278 if (auth_res != SVC_OK)
1279 trace_svc_authenticate(rqstp, auth_res);
1286 rpc_stat = rpc_system_err;
1301 rpc_stat = progp->pg_init_request(rqstp, progp, &process);
1305 case rpc_prog_unavail:
1307 case rpc_prog_mismatch:
1309 case rpc_proc_unavail:
1313 procp = rqstp->rq_procinfo;
1314 /* Should this check go into the dispatcher? */
1315 if (!procp || !procp->pc_func)
1318 /* Syntactic check complete */
1319 serv->sv_stats->rpccnt++;
1320 trace_svc_process(rqstp, progp->pg_name);
1322 /* Build the reply header. */
1323 statp = resv->iov_base +resv->iov_len;
1324 svc_putnl(resv, RPC_SUCCESS);
1326 /* un-reserve some of the out-queue now that we have a
1327 * better idea of reply size
1329 if (procp->pc_xdrressize)
1330 svc_reserve_auth(rqstp, procp->pc_xdrressize<<2);
1332 /* Call the function that processes the request. */
1333 rc = process.dispatch(rqstp, statp);
1334 if (procp->pc_release)
1335 procp->pc_release(rqstp);
1338 if (rqstp->rq_auth_stat != rpc_auth_ok)
1341 /* Check RPC status result */
1342 if (*statp != rpc_success)
1343 resv->iov_len = ((void*)statp) - resv->iov_base + 4;
1345 if (procp->pc_encode == NULL)
1349 if (svc_authorise(rqstp))
1351 return 1; /* Caller can now send it */
1354 svc_authorise(rqstp); /* doesn't hurt to call this twice */
1355 dprintk("svc: svc_process dropit\n");
1359 svc_authorise(rqstp);
1361 if (rqstp->rq_xprt && test_bit(XPT_TEMP, &rqstp->rq_xprt->xpt_flags))
1362 svc_xprt_close(rqstp->rq_xprt);
1363 dprintk("svc: svc_process close\n");
1367 svc_printk(rqstp, "short len %zd, dropping request\n",
1372 serv->sv_stats->rpcbadfmt++;
1373 svc_putnl(resv, 1); /* REJECT */
1374 svc_putnl(resv, 0); /* RPC_MISMATCH */
1375 svc_putnl(resv, 2); /* Only RPCv2 supported */
1380 dprintk("svc: authentication failed (%d)\n",
1381 be32_to_cpu(rqstp->rq_auth_stat));
1382 serv->sv_stats->rpcbadauth++;
1383 /* Restore write pointer to location of accept status: */
1384 xdr_ressize_check(rqstp, reply_statp);
1385 svc_putnl(resv, 1); /* REJECT */
1386 svc_putnl(resv, 1); /* AUTH_ERROR */
1387 svc_putu32(resv, rqstp->rq_auth_stat); /* status */
1391 dprintk("svc: unknown program %d\n", prog);
1392 serv->sv_stats->rpcbadfmt++;
1393 svc_putnl(resv, RPC_PROG_UNAVAIL);
1397 svc_printk(rqstp, "unknown version (%d for prog %d, %s)\n",
1398 rqstp->rq_vers, rqstp->rq_prog, progp->pg_name);
1400 serv->sv_stats->rpcbadfmt++;
1401 svc_putnl(resv, RPC_PROG_MISMATCH);
1402 svc_putnl(resv, process.mismatch.lovers);
1403 svc_putnl(resv, process.mismatch.hivers);
1407 svc_printk(rqstp, "unknown procedure (%d)\n", rqstp->rq_proc);
1409 serv->sv_stats->rpcbadfmt++;
1410 svc_putnl(resv, RPC_PROC_UNAVAIL);
1414 svc_printk(rqstp, "failed to decode args\n");
1416 rpc_stat = rpc_garbage_args;
1418 serv->sv_stats->rpcbadfmt++;
1419 svc_putnl(resv, ntohl(rpc_stat));
1424 * Process the RPC request.
1427 svc_process(struct svc_rqst *rqstp)
1429 struct kvec *argv = &rqstp->rq_arg.head[0];
1430 struct kvec *resv = &rqstp->rq_res.head[0];
1431 struct svc_serv *serv = rqstp->rq_server;
1434 #if IS_ENABLED(CONFIG_FAIL_SUNRPC)
1435 if (!fail_sunrpc.ignore_server_disconnect &&
1436 should_fail(&fail_sunrpc.attr, 1))
1437 svc_xprt_deferred_close(rqstp->rq_xprt);
1441 * Setup response xdr_buf.
1442 * Initially it has just one page
1444 rqstp->rq_next_page = &rqstp->rq_respages[1];
1445 resv->iov_base = page_address(rqstp->rq_respages[0]);
1447 rqstp->rq_res.pages = rqstp->rq_respages + 1;
1448 rqstp->rq_res.len = 0;
1449 rqstp->rq_res.page_base = 0;
1450 rqstp->rq_res.page_len = 0;
1451 rqstp->rq_res.buflen = PAGE_SIZE;
1452 rqstp->rq_res.tail[0].iov_base = NULL;
1453 rqstp->rq_res.tail[0].iov_len = 0;
1455 dir = svc_getnl(argv);
1457 /* direction != CALL */
1458 svc_printk(rqstp, "bad direction %d, dropping request\n", dir);
1459 serv->sv_stats->rpcbadfmt++;
1463 /* Returns 1 for send, 0 for drop */
1464 if (likely(svc_process_common(rqstp, argv, resv)))
1465 return svc_send(rqstp);
1471 EXPORT_SYMBOL_GPL(svc_process);
1473 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1475 * Process a backchannel RPC request that arrived over an existing
1476 * outbound connection
1479 bc_svc_process(struct svc_serv *serv, struct rpc_rqst *req,
1480 struct svc_rqst *rqstp)
1482 struct kvec *argv = &rqstp->rq_arg.head[0];
1483 struct kvec *resv = &rqstp->rq_res.head[0];
1484 struct rpc_task *task;
1488 dprintk("svc: %s(%p)\n", __func__, req);
1490 /* Build the svc_rqst used by the common processing routine */
1491 rqstp->rq_xid = req->rq_xid;
1492 rqstp->rq_prot = req->rq_xprt->prot;
1493 rqstp->rq_server = serv;
1494 rqstp->rq_bc_net = req->rq_xprt->xprt_net;
1496 rqstp->rq_addrlen = sizeof(req->rq_xprt->addr);
1497 memcpy(&rqstp->rq_addr, &req->rq_xprt->addr, rqstp->rq_addrlen);
1498 memcpy(&rqstp->rq_arg, &req->rq_rcv_buf, sizeof(rqstp->rq_arg));
1499 memcpy(&rqstp->rq_res, &req->rq_snd_buf, sizeof(rqstp->rq_res));
1501 /* Adjust the argument buffer length */
1502 rqstp->rq_arg.len = req->rq_private_buf.len;
1503 if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len) {
1504 rqstp->rq_arg.head[0].iov_len = rqstp->rq_arg.len;
1505 rqstp->rq_arg.page_len = 0;
1506 } else if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len +
1507 rqstp->rq_arg.page_len)
1508 rqstp->rq_arg.page_len = rqstp->rq_arg.len -
1509 rqstp->rq_arg.head[0].iov_len;
1511 rqstp->rq_arg.len = rqstp->rq_arg.head[0].iov_len +
1512 rqstp->rq_arg.page_len;
1514 /* reset result send buffer "put" position */
1518 * Skip the next two words because they've already been
1519 * processed in the transport
1521 svc_getu32(argv); /* XID */
1522 svc_getnl(argv); /* CALLDIR */
1524 /* Parse and execute the bc call */
1525 proc_error = svc_process_common(rqstp, argv, resv);
1527 atomic_dec(&req->rq_xprt->bc_slot_count);
1529 /* Processing error: drop the request */
1530 xprt_free_bc_request(req);
1534 /* Finally, send the reply synchronously */
1535 memcpy(&req->rq_snd_buf, &rqstp->rq_res, sizeof(req->rq_snd_buf));
1536 task = rpc_run_bc_task(req);
1538 error = PTR_ERR(task);
1542 WARN_ON_ONCE(atomic_read(&task->tk_count) != 1);
1543 error = task->tk_status;
1547 dprintk("svc: %s(), error=%d\n", __func__, error);
1550 EXPORT_SYMBOL_GPL(bc_svc_process);
1551 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1554 * Return (transport-specific) limit on the rpc payload.
1556 u32 svc_max_payload(const struct svc_rqst *rqstp)
1558 u32 max = rqstp->rq_xprt->xpt_class->xcl_max_payload;
1560 if (rqstp->rq_server->sv_max_payload < max)
1561 max = rqstp->rq_server->sv_max_payload;
1564 EXPORT_SYMBOL_GPL(svc_max_payload);
1567 * svc_proc_name - Return RPC procedure name in string form
1568 * @rqstp: svc_rqst to operate on
1571 * Pointer to a NUL-terminated string
1573 const char *svc_proc_name(const struct svc_rqst *rqstp)
1575 if (rqstp && rqstp->rq_procinfo)
1576 return rqstp->rq_procinfo->pc_name;
1582 * svc_encode_result_payload - mark a range of bytes as a result payload
1583 * @rqstp: svc_rqst to operate on
1584 * @offset: payload's byte offset in rqstp->rq_res
1585 * @length: size of payload, in bytes
1587 * Returns zero on success, or a negative errno if a permanent
1590 int svc_encode_result_payload(struct svc_rqst *rqstp, unsigned int offset,
1591 unsigned int length)
1593 return rqstp->rq_xprt->xpt_ops->xpo_result_payload(rqstp, offset,
1596 EXPORT_SYMBOL_GPL(svc_encode_result_payload);
1599 * svc_fill_write_vector - Construct data argument for VFS write call
1600 * @rqstp: svc_rqst to operate on
1601 * @payload: xdr_buf containing only the write data payload
1603 * Fills in rqstp::rq_vec, and returns the number of elements.
1605 unsigned int svc_fill_write_vector(struct svc_rqst *rqstp,
1606 struct xdr_buf *payload)
1608 struct page **pages = payload->pages;
1609 struct kvec *first = payload->head;
1610 struct kvec *vec = rqstp->rq_vec;
1611 size_t total = payload->len;
1614 /* Some types of transport can present the write payload
1615 * entirely in rq_arg.pages. In this case, @first is empty.
1618 if (first->iov_len) {
1619 vec[i].iov_base = first->iov_base;
1620 vec[i].iov_len = min_t(size_t, total, first->iov_len);
1621 total -= vec[i].iov_len;
1626 vec[i].iov_base = page_address(*pages);
1627 vec[i].iov_len = min_t(size_t, total, PAGE_SIZE);
1628 total -= vec[i].iov_len;
1633 WARN_ON_ONCE(i > ARRAY_SIZE(rqstp->rq_vec));
1636 EXPORT_SYMBOL_GPL(svc_fill_write_vector);
1639 * svc_fill_symlink_pathname - Construct pathname argument for VFS symlink call
1640 * @rqstp: svc_rqst to operate on
1641 * @first: buffer containing first section of pathname
1642 * @p: buffer containing remaining section of pathname
1643 * @total: total length of the pathname argument
1645 * The VFS symlink API demands a NUL-terminated pathname in mapped memory.
1646 * Returns pointer to a NUL-terminated string, or an ERR_PTR. Caller must free
1647 * the returned string.
1649 char *svc_fill_symlink_pathname(struct svc_rqst *rqstp, struct kvec *first,
1650 void *p, size_t total)
1652 size_t len, remaining;
1655 result = kmalloc(total + 1, GFP_KERNEL);
1657 return ERR_PTR(-ESERVERFAULT);
1662 len = min_t(size_t, total, first->iov_len);
1664 memcpy(dst, first->iov_base, len);
1670 len = min_t(size_t, remaining, PAGE_SIZE);
1671 memcpy(dst, p, len);
1677 /* Sanity check: Linux doesn't allow the pathname argument to
1678 * contain a NUL byte.
1680 if (strlen(result) != total) {
1682 return ERR_PTR(-EINVAL);
1686 EXPORT_SYMBOL_GPL(svc_fill_symlink_pathname);