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_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 sysfs_emit(buf, "auto\n");
113 case SVC_POOL_GLOBAL:
114 return sysfs_emit(buf, "global\n");
115 case SVC_POOL_PERCPU:
116 return sysfs_emit(buf, "percpu\n");
117 case SVC_POOL_PERNODE:
118 return sysfs_emit(buf, "pernode\n");
120 return sysfs_emit(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 * svc_pool_for_cpu - Select pool to run a thread on this cpu
361 * @serv: An RPC service
363 * Use the active CPU and the svc_pool_map's mode setting to
364 * select the svc thread pool to use. Once initialized, the
365 * svc_pool_map does not change.
368 * A pointer to an svc_pool
370 struct svc_pool *svc_pool_for_cpu(struct svc_serv *serv)
372 struct svc_pool_map *m = &svc_pool_map;
373 int cpu = raw_smp_processor_id();
374 unsigned int pidx = 0;
376 if (serv->sv_nrpools <= 1)
377 return serv->sv_pools;
380 case SVC_POOL_PERCPU:
381 pidx = m->to_pool[cpu];
383 case SVC_POOL_PERNODE:
384 pidx = m->to_pool[cpu_to_node(cpu)];
388 return &serv->sv_pools[pidx % serv->sv_nrpools];
391 int svc_rpcb_setup(struct svc_serv *serv, struct net *net)
395 err = rpcb_create_local(net);
399 /* Remove any stale portmap registrations */
400 svc_unregister(serv, net);
403 EXPORT_SYMBOL_GPL(svc_rpcb_setup);
405 void svc_rpcb_cleanup(struct svc_serv *serv, struct net *net)
407 svc_unregister(serv, net);
410 EXPORT_SYMBOL_GPL(svc_rpcb_cleanup);
412 static int svc_uses_rpcbind(struct svc_serv *serv)
414 struct svc_program *progp;
417 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
418 for (i = 0; i < progp->pg_nvers; i++) {
419 if (progp->pg_vers[i] == NULL)
421 if (!progp->pg_vers[i]->vs_hidden)
429 int svc_bind(struct svc_serv *serv, struct net *net)
431 if (!svc_uses_rpcbind(serv))
433 return svc_rpcb_setup(serv, net);
435 EXPORT_SYMBOL_GPL(svc_bind);
437 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
439 __svc_init_bc(struct svc_serv *serv)
441 INIT_LIST_HEAD(&serv->sv_cb_list);
442 spin_lock_init(&serv->sv_cb_lock);
443 init_waitqueue_head(&serv->sv_cb_waitq);
447 __svc_init_bc(struct svc_serv *serv)
453 * Create an RPC service
455 static struct svc_serv *
456 __svc_create(struct svc_program *prog, unsigned int bufsize, int npools,
457 int (*threadfn)(void *data))
459 struct svc_serv *serv;
461 unsigned int xdrsize;
464 if (!(serv = kzalloc(sizeof(*serv), GFP_KERNEL)))
466 serv->sv_name = prog->pg_name;
467 serv->sv_program = prog;
468 kref_init(&serv->sv_refcnt);
469 serv->sv_stats = prog->pg_stats;
470 if (bufsize > RPCSVC_MAXPAYLOAD)
471 bufsize = RPCSVC_MAXPAYLOAD;
472 serv->sv_max_payload = bufsize? bufsize : 4096;
473 serv->sv_max_mesg = roundup(serv->sv_max_payload + PAGE_SIZE, PAGE_SIZE);
474 serv->sv_threadfn = threadfn;
477 prog->pg_lovers = prog->pg_nvers-1;
478 for (vers=0; vers<prog->pg_nvers ; vers++)
479 if (prog->pg_vers[vers]) {
480 prog->pg_hivers = vers;
481 if (prog->pg_lovers > vers)
482 prog->pg_lovers = vers;
483 if (prog->pg_vers[vers]->vs_xdrsize > xdrsize)
484 xdrsize = prog->pg_vers[vers]->vs_xdrsize;
486 prog = prog->pg_next;
488 serv->sv_xdrsize = xdrsize;
489 INIT_LIST_HEAD(&serv->sv_tempsocks);
490 INIT_LIST_HEAD(&serv->sv_permsocks);
491 timer_setup(&serv->sv_temptimer, NULL, 0);
492 spin_lock_init(&serv->sv_lock);
496 serv->sv_nrpools = npools;
498 kcalloc(serv->sv_nrpools, sizeof(struct svc_pool),
500 if (!serv->sv_pools) {
505 for (i = 0; i < serv->sv_nrpools; i++) {
506 struct svc_pool *pool = &serv->sv_pools[i];
508 dprintk("svc: initialising pool %u for %s\n",
512 INIT_LIST_HEAD(&pool->sp_sockets);
513 INIT_LIST_HEAD(&pool->sp_all_threads);
514 spin_lock_init(&pool->sp_lock);
516 percpu_counter_init(&pool->sp_sockets_queued, 0, GFP_KERNEL);
517 percpu_counter_init(&pool->sp_threads_woken, 0, GFP_KERNEL);
524 * svc_create - Create an RPC service
525 * @prog: the RPC program the new service will handle
526 * @bufsize: maximum message size for @prog
527 * @threadfn: a function to service RPC requests for @prog
529 * Returns an instantiated struct svc_serv object or NULL.
531 struct svc_serv *svc_create(struct svc_program *prog, unsigned int bufsize,
532 int (*threadfn)(void *data))
534 return __svc_create(prog, bufsize, 1, threadfn);
536 EXPORT_SYMBOL_GPL(svc_create);
539 * svc_create_pooled - Create an RPC service with pooled threads
540 * @prog: the RPC program the new service will handle
541 * @bufsize: maximum message size for @prog
542 * @threadfn: a function to service RPC requests for @prog
544 * Returns an instantiated struct svc_serv object or NULL.
546 struct svc_serv *svc_create_pooled(struct svc_program *prog,
547 unsigned int bufsize,
548 int (*threadfn)(void *data))
550 struct svc_serv *serv;
551 unsigned int npools = svc_pool_map_get();
553 serv = __svc_create(prog, bufsize, npools, threadfn);
558 svc_pool_map_put(npools);
561 EXPORT_SYMBOL_GPL(svc_create_pooled);
564 * Destroy an RPC service. Should be called with appropriate locking to
565 * protect sv_permsocks and sv_tempsocks.
568 svc_destroy(struct kref *ref)
570 struct svc_serv *serv = container_of(ref, struct svc_serv, sv_refcnt);
573 dprintk("svc: svc_destroy(%s)\n", serv->sv_program->pg_name);
574 timer_shutdown_sync(&serv->sv_temptimer);
577 * The last user is gone and thus all sockets have to be destroyed to
578 * the point. Check this.
580 BUG_ON(!list_empty(&serv->sv_permsocks));
581 BUG_ON(!list_empty(&serv->sv_tempsocks));
583 cache_clean_deferred(serv);
585 svc_pool_map_put(serv->sv_nrpools);
587 for (i = 0; i < serv->sv_nrpools; i++) {
588 struct svc_pool *pool = &serv->sv_pools[i];
590 percpu_counter_destroy(&pool->sp_sockets_queued);
591 percpu_counter_destroy(&pool->sp_threads_woken);
593 kfree(serv->sv_pools);
596 EXPORT_SYMBOL_GPL(svc_destroy);
599 svc_init_buffer(struct svc_rqst *rqstp, unsigned int size, int node)
601 unsigned long pages, ret;
603 /* bc_xprt uses fore channel allocated buffers */
604 if (svc_is_backchannel(rqstp))
607 pages = size / PAGE_SIZE + 1; /* extra page as we hold both request and reply.
608 * We assume one is at most one page
610 WARN_ON_ONCE(pages > RPCSVC_MAXPAGES);
611 if (pages > RPCSVC_MAXPAGES)
612 pages = RPCSVC_MAXPAGES;
614 ret = alloc_pages_bulk_array_node(GFP_KERNEL, node, pages,
620 * Release an RPC server buffer
623 svc_release_buffer(struct svc_rqst *rqstp)
627 for (i = 0; i < ARRAY_SIZE(rqstp->rq_pages); i++)
628 if (rqstp->rq_pages[i])
629 put_page(rqstp->rq_pages[i]);
633 svc_rqst_alloc(struct svc_serv *serv, struct svc_pool *pool, int node)
635 struct svc_rqst *rqstp;
637 rqstp = kzalloc_node(sizeof(*rqstp), GFP_KERNEL, node);
641 folio_batch_init(&rqstp->rq_fbatch);
643 __set_bit(RQ_BUSY, &rqstp->rq_flags);
644 rqstp->rq_server = serv;
645 rqstp->rq_pool = pool;
647 rqstp->rq_scratch_page = alloc_pages_node(node, GFP_KERNEL, 0);
648 if (!rqstp->rq_scratch_page)
651 rqstp->rq_argp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
655 rqstp->rq_resp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
659 if (!svc_init_buffer(rqstp, serv->sv_max_mesg, node))
664 svc_rqst_free(rqstp);
667 EXPORT_SYMBOL_GPL(svc_rqst_alloc);
669 static struct svc_rqst *
670 svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
672 struct svc_rqst *rqstp;
674 rqstp = svc_rqst_alloc(serv, pool, node);
676 return ERR_PTR(-ENOMEM);
679 spin_lock_bh(&serv->sv_lock);
680 serv->sv_nrthreads += 1;
681 spin_unlock_bh(&serv->sv_lock);
683 spin_lock_bh(&pool->sp_lock);
684 pool->sp_nrthreads++;
685 list_add_rcu(&rqstp->rq_all, &pool->sp_all_threads);
686 spin_unlock_bh(&pool->sp_lock);
691 * Choose a pool in which to create a new thread, for svc_set_num_threads
693 static inline struct svc_pool *
694 choose_pool(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
699 return &serv->sv_pools[(*state)++ % serv->sv_nrpools];
703 * Choose a thread to kill, for svc_set_num_threads
705 static inline struct task_struct *
706 choose_victim(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
709 struct task_struct *task = NULL;
712 spin_lock_bh(&pool->sp_lock);
714 /* choose a pool in round-robin fashion */
715 for (i = 0; i < serv->sv_nrpools; i++) {
716 pool = &serv->sv_pools[--(*state) % serv->sv_nrpools];
717 spin_lock_bh(&pool->sp_lock);
718 if (!list_empty(&pool->sp_all_threads))
720 spin_unlock_bh(&pool->sp_lock);
726 if (!list_empty(&pool->sp_all_threads)) {
727 struct svc_rqst *rqstp;
730 * Remove from the pool->sp_all_threads list
731 * so we don't try to kill it again.
733 rqstp = list_entry(pool->sp_all_threads.next, struct svc_rqst, rq_all);
734 set_bit(RQ_VICTIM, &rqstp->rq_flags);
735 list_del_rcu(&rqstp->rq_all);
736 task = rqstp->rq_task;
738 spin_unlock_bh(&pool->sp_lock);
743 /* create new threads */
745 svc_start_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
747 struct svc_rqst *rqstp;
748 struct task_struct *task;
749 struct svc_pool *chosen_pool;
750 unsigned int state = serv->sv_nrthreads-1;
755 chosen_pool = choose_pool(serv, pool, &state);
757 node = svc_pool_map_get_node(chosen_pool->sp_id);
758 rqstp = svc_prepare_thread(serv, chosen_pool, node);
760 return PTR_ERR(rqstp);
762 task = kthread_create_on_node(serv->sv_threadfn, rqstp,
763 node, "%s", serv->sv_name);
765 svc_exit_thread(rqstp);
766 return PTR_ERR(task);
769 rqstp->rq_task = task;
770 if (serv->sv_nrpools > 1)
771 svc_pool_map_set_cpumask(task, chosen_pool->sp_id);
773 svc_sock_update_bufs(serv);
774 wake_up_process(task);
775 } while (nrservs > 0);
781 * Create or destroy enough new threads to make the number
782 * of threads the given number. If `pool' is non-NULL, applies
783 * only to threads in that pool, otherwise round-robins between
784 * all pools. Caller must ensure that mutual exclusion between this and
785 * server startup or shutdown.
788 /* destroy old threads */
790 svc_stop_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
792 struct svc_rqst *rqstp;
793 struct task_struct *task;
794 unsigned int state = serv->sv_nrthreads-1;
796 /* destroy old threads */
798 task = choose_victim(serv, pool, &state);
801 rqstp = kthread_data(task);
802 /* Did we lose a race to svo_function threadfn? */
803 if (kthread_stop(task) == -EINTR)
804 svc_exit_thread(rqstp);
806 } while (nrservs < 0);
811 svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
814 nrservs -= serv->sv_nrthreads;
816 spin_lock_bh(&pool->sp_lock);
817 nrservs -= pool->sp_nrthreads;
818 spin_unlock_bh(&pool->sp_lock);
822 return svc_start_kthreads(serv, pool, nrservs);
824 return svc_stop_kthreads(serv, pool, nrservs);
827 EXPORT_SYMBOL_GPL(svc_set_num_threads);
830 * svc_rqst_replace_page - Replace one page in rq_pages[]
831 * @rqstp: svc_rqst with pages to replace
832 * @page: replacement page
834 * When replacing a page in rq_pages, batch the release of the
835 * replaced pages to avoid hammering the page allocator.
838 * %true: page replaced
839 * %false: array bounds checking failed
841 bool svc_rqst_replace_page(struct svc_rqst *rqstp, struct page *page)
843 struct page **begin = rqstp->rq_pages;
844 struct page **end = &rqstp->rq_pages[RPCSVC_MAXPAGES];
846 if (unlikely(rqstp->rq_next_page < begin || rqstp->rq_next_page > end)) {
847 trace_svc_replace_page_err(rqstp);
851 if (*rqstp->rq_next_page) {
852 if (!folio_batch_add(&rqstp->rq_fbatch,
853 page_folio(*rqstp->rq_next_page)))
854 __folio_batch_release(&rqstp->rq_fbatch);
858 *(rqstp->rq_next_page++) = page;
861 EXPORT_SYMBOL_GPL(svc_rqst_replace_page);
864 * svc_rqst_release_pages - Release Reply buffer pages
865 * @rqstp: RPC transaction context
867 * Release response pages that might still be in flight after
868 * svc_send, and any spliced filesystem-owned pages.
870 void svc_rqst_release_pages(struct svc_rqst *rqstp)
872 int i, count = rqstp->rq_next_page - rqstp->rq_respages;
875 release_pages(rqstp->rq_respages, count);
876 for (i = 0; i < count; i++)
877 rqstp->rq_respages[i] = NULL;
882 * Called from a server thread as it's exiting. Caller must hold the "service
883 * mutex" for the service.
886 svc_rqst_free(struct svc_rqst *rqstp)
888 folio_batch_release(&rqstp->rq_fbatch);
889 svc_release_buffer(rqstp);
890 if (rqstp->rq_scratch_page)
891 put_page(rqstp->rq_scratch_page);
892 kfree(rqstp->rq_resp);
893 kfree(rqstp->rq_argp);
894 kfree(rqstp->rq_auth_data);
895 kfree_rcu(rqstp, rq_rcu_head);
897 EXPORT_SYMBOL_GPL(svc_rqst_free);
900 svc_exit_thread(struct svc_rqst *rqstp)
902 struct svc_serv *serv = rqstp->rq_server;
903 struct svc_pool *pool = rqstp->rq_pool;
905 spin_lock_bh(&pool->sp_lock);
906 pool->sp_nrthreads--;
907 if (!test_and_set_bit(RQ_VICTIM, &rqstp->rq_flags))
908 list_del_rcu(&rqstp->rq_all);
909 spin_unlock_bh(&pool->sp_lock);
911 spin_lock_bh(&serv->sv_lock);
912 serv->sv_nrthreads -= 1;
913 spin_unlock_bh(&serv->sv_lock);
914 svc_sock_update_bufs(serv);
916 svc_rqst_free(rqstp);
920 EXPORT_SYMBOL_GPL(svc_exit_thread);
923 * Register an "inet" protocol family netid with the local
924 * rpcbind daemon via an rpcbind v4 SET request.
926 * No netconfig infrastructure is available in the kernel, so
927 * we map IP_ protocol numbers to netids by hand.
929 * Returns zero on success; a negative errno value is returned
930 * if any error occurs.
932 static int __svc_rpcb_register4(struct net *net, const u32 program,
934 const unsigned short protocol,
935 const unsigned short port)
937 const struct sockaddr_in sin = {
938 .sin_family = AF_INET,
939 .sin_addr.s_addr = htonl(INADDR_ANY),
940 .sin_port = htons(port),
947 netid = RPCBIND_NETID_UDP;
950 netid = RPCBIND_NETID_TCP;
956 error = rpcb_v4_register(net, program, version,
957 (const struct sockaddr *)&sin, netid);
960 * User space didn't support rpcbind v4, so retry this
961 * registration request with the legacy rpcbind v2 protocol.
963 if (error == -EPROTONOSUPPORT)
964 error = rpcb_register(net, program, version, protocol, port);
969 #if IS_ENABLED(CONFIG_IPV6)
971 * Register an "inet6" protocol family netid with the local
972 * rpcbind daemon via an rpcbind v4 SET request.
974 * No netconfig infrastructure is available in the kernel, so
975 * we map IP_ protocol numbers to netids by hand.
977 * Returns zero on success; a negative errno value is returned
978 * if any error occurs.
980 static int __svc_rpcb_register6(struct net *net, const u32 program,
982 const unsigned short protocol,
983 const unsigned short port)
985 const struct sockaddr_in6 sin6 = {
986 .sin6_family = AF_INET6,
987 .sin6_addr = IN6ADDR_ANY_INIT,
988 .sin6_port = htons(port),
995 netid = RPCBIND_NETID_UDP6;
998 netid = RPCBIND_NETID_TCP6;
1001 return -ENOPROTOOPT;
1004 error = rpcb_v4_register(net, program, version,
1005 (const struct sockaddr *)&sin6, netid);
1008 * User space didn't support rpcbind version 4, so we won't
1009 * use a PF_INET6 listener.
1011 if (error == -EPROTONOSUPPORT)
1012 error = -EAFNOSUPPORT;
1016 #endif /* IS_ENABLED(CONFIG_IPV6) */
1019 * Register a kernel RPC service via rpcbind version 4.
1021 * Returns zero on success; a negative errno value is returned
1022 * if any error occurs.
1024 static int __svc_register(struct net *net, const char *progname,
1025 const u32 program, const u32 version,
1027 const unsigned short protocol,
1028 const unsigned short port)
1030 int error = -EAFNOSUPPORT;
1034 error = __svc_rpcb_register4(net, program, version,
1037 #if IS_ENABLED(CONFIG_IPV6)
1039 error = __svc_rpcb_register6(net, program, version,
1044 trace_svc_register(progname, version, family, protocol, port, error);
1048 int svc_rpcbind_set_version(struct net *net,
1049 const struct svc_program *progp,
1050 u32 version, int family,
1051 unsigned short proto,
1052 unsigned short port)
1054 return __svc_register(net, progp->pg_name, progp->pg_prog,
1055 version, family, proto, port);
1058 EXPORT_SYMBOL_GPL(svc_rpcbind_set_version);
1060 int svc_generic_rpcbind_set(struct net *net,
1061 const struct svc_program *progp,
1062 u32 version, int family,
1063 unsigned short proto,
1064 unsigned short port)
1066 const struct svc_version *vers = progp->pg_vers[version];
1072 if (vers->vs_hidden) {
1073 trace_svc_noregister(progp->pg_name, version, proto,
1079 * Don't register a UDP port if we need congestion
1082 if (vers->vs_need_cong_ctrl && proto == IPPROTO_UDP)
1085 error = svc_rpcbind_set_version(net, progp, version,
1086 family, proto, port);
1088 return (vers->vs_rpcb_optnl) ? 0 : error;
1090 EXPORT_SYMBOL_GPL(svc_generic_rpcbind_set);
1093 * svc_register - register an RPC service with the local portmapper
1094 * @serv: svc_serv struct for the service to register
1095 * @net: net namespace for the service to register
1096 * @family: protocol family of service's listener socket
1097 * @proto: transport protocol number to advertise
1098 * @port: port to advertise
1100 * Service is registered for any address in the passed-in protocol family
1102 int svc_register(const struct svc_serv *serv, struct net *net,
1103 const int family, const unsigned short proto,
1104 const unsigned short port)
1106 struct svc_program *progp;
1110 WARN_ON_ONCE(proto == 0 && port == 0);
1111 if (proto == 0 && port == 0)
1114 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1115 for (i = 0; i < progp->pg_nvers; i++) {
1117 error = progp->pg_rpcbind_set(net, progp, i,
1118 family, proto, port);
1120 printk(KERN_WARNING "svc: failed to register "
1121 "%sv%u RPC service (errno %d).\n",
1122 progp->pg_name, i, -error);
1132 * If user space is running rpcbind, it should take the v4 UNSET
1133 * and clear everything for this [program, version]. If user space
1134 * is running portmap, it will reject the v4 UNSET, but won't have
1135 * any "inet6" entries anyway. So a PMAP_UNSET should be sufficient
1136 * in this case to clear all existing entries for [program, version].
1138 static void __svc_unregister(struct net *net, const u32 program, const u32 version,
1139 const char *progname)
1143 error = rpcb_v4_register(net, program, version, NULL, "");
1146 * User space didn't support rpcbind v4, so retry this
1147 * request with the legacy rpcbind v2 protocol.
1149 if (error == -EPROTONOSUPPORT)
1150 error = rpcb_register(net, program, version, 0, 0);
1152 trace_svc_unregister(progname, version, error);
1156 * All netids, bind addresses and ports registered for [program, version]
1157 * are removed from the local rpcbind database (if the service is not
1158 * hidden) to make way for a new instance of the service.
1160 * The result of unregistration is reported via dprintk for those who want
1161 * verification of the result, but is otherwise not important.
1163 static void svc_unregister(const struct svc_serv *serv, struct net *net)
1165 struct sighand_struct *sighand;
1166 struct svc_program *progp;
1167 unsigned long flags;
1170 clear_thread_flag(TIF_SIGPENDING);
1172 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1173 for (i = 0; i < progp->pg_nvers; i++) {
1174 if (progp->pg_vers[i] == NULL)
1176 if (progp->pg_vers[i]->vs_hidden)
1178 __svc_unregister(net, progp->pg_prog, i, progp->pg_name);
1183 sighand = rcu_dereference(current->sighand);
1184 spin_lock_irqsave(&sighand->siglock, flags);
1185 recalc_sigpending();
1186 spin_unlock_irqrestore(&sighand->siglock, flags);
1191 * dprintk the given error with the address of the client that caused it.
1193 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
1194 static __printf(2, 3)
1195 void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...)
1197 struct va_format vaf;
1199 char buf[RPC_MAX_ADDRBUFLEN];
1201 va_start(args, fmt);
1206 dprintk("svc: %s: %pV", svc_print_addr(rqstp, buf, sizeof(buf)), &vaf);
1211 static __printf(2,3) void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...) {}
1215 svc_generic_init_request(struct svc_rqst *rqstp,
1216 const struct svc_program *progp,
1217 struct svc_process_info *ret)
1219 const struct svc_version *versp = NULL; /* compiler food */
1220 const struct svc_procedure *procp = NULL;
1222 if (rqstp->rq_vers >= progp->pg_nvers )
1224 versp = progp->pg_vers[rqstp->rq_vers];
1229 * Some protocol versions (namely NFSv4) require some form of
1230 * congestion control. (See RFC 7530 section 3.1 paragraph 2)
1231 * In other words, UDP is not allowed. We mark those when setting
1232 * up the svc_xprt, and verify that here.
1234 * The spec is not very clear about what error should be returned
1235 * when someone tries to access a server that is listening on UDP
1236 * for lower versions. RPC_PROG_MISMATCH seems to be the closest
1239 if (versp->vs_need_cong_ctrl && rqstp->rq_xprt &&
1240 !test_bit(XPT_CONG_CTRL, &rqstp->rq_xprt->xpt_flags))
1243 if (rqstp->rq_proc >= versp->vs_nproc)
1245 rqstp->rq_procinfo = procp = &versp->vs_proc[rqstp->rq_proc];
1249 /* Initialize storage for argp and resp */
1250 memset(rqstp->rq_argp, 0, procp->pc_argzero);
1251 memset(rqstp->rq_resp, 0, procp->pc_ressize);
1253 /* Bump per-procedure stats counter */
1254 this_cpu_inc(versp->vs_count[rqstp->rq_proc]);
1256 ret->dispatch = versp->vs_dispatch;
1259 ret->mismatch.lovers = progp->pg_lovers;
1260 ret->mismatch.hivers = progp->pg_hivers;
1261 return rpc_prog_mismatch;
1263 return rpc_proc_unavail;
1265 EXPORT_SYMBOL_GPL(svc_generic_init_request);
1268 * Common routine for processing the RPC request.
1271 svc_process_common(struct svc_rqst *rqstp)
1273 struct xdr_stream *xdr = &rqstp->rq_res_stream;
1274 struct svc_program *progp;
1275 const struct svc_procedure *procp = NULL;
1276 struct svc_serv *serv = rqstp->rq_server;
1277 struct svc_process_info process;
1279 unsigned int aoffset;
1282 /* Will be turned off by GSS integrity and privacy services */
1283 set_bit(RQ_SPLICE_OK, &rqstp->rq_flags);
1284 /* Will be turned off only when NFSv4 Sessions are used */
1285 set_bit(RQ_USEDEFERRAL, &rqstp->rq_flags);
1286 clear_bit(RQ_DROPME, &rqstp->rq_flags);
1288 /* Construct the first words of the reply: */
1289 svcxdr_init_encode(rqstp);
1290 xdr_stream_encode_be32(xdr, rqstp->rq_xid);
1291 xdr_stream_encode_be32(xdr, rpc_reply);
1293 p = xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 4);
1296 if (*p++ != cpu_to_be32(RPC_VERSION))
1299 xdr_stream_encode_be32(xdr, rpc_msg_accepted);
1301 rqstp->rq_prog = be32_to_cpup(p++);
1302 rqstp->rq_vers = be32_to_cpup(p++);
1303 rqstp->rq_proc = be32_to_cpup(p);
1305 for (progp = serv->sv_program; progp; progp = progp->pg_next)
1306 if (rqstp->rq_prog == progp->pg_prog)
1310 * Decode auth data, and add verifier to reply buffer.
1311 * We do this before anything else in order to get a decent
1314 auth_res = svc_authenticate(rqstp);
1315 /* Also give the program a chance to reject this call: */
1316 if (auth_res == SVC_OK && progp)
1317 auth_res = progp->pg_authenticate(rqstp);
1318 trace_svc_authenticate(rqstp, auth_res);
1323 goto err_garbage_args;
1325 goto err_system_err;
1339 switch (progp->pg_init_request(rqstp, progp, &process)) {
1342 case rpc_prog_unavail:
1344 case rpc_prog_mismatch:
1346 case rpc_proc_unavail:
1350 procp = rqstp->rq_procinfo;
1351 /* Should this check go into the dispatcher? */
1352 if (!procp || !procp->pc_func)
1355 /* Syntactic check complete */
1356 serv->sv_stats->rpccnt++;
1357 trace_svc_process(rqstp, progp->pg_name);
1359 aoffset = xdr_stream_pos(xdr);
1361 /* un-reserve some of the out-queue now that we have a
1362 * better idea of reply size
1364 if (procp->pc_xdrressize)
1365 svc_reserve_auth(rqstp, procp->pc_xdrressize<<2);
1367 /* Call the function that processes the request. */
1368 rc = process.dispatch(rqstp);
1369 if (procp->pc_release)
1370 procp->pc_release(rqstp);
1373 if (rqstp->rq_auth_stat != rpc_auth_ok)
1376 if (*rqstp->rq_accept_statp != rpc_success)
1377 xdr_truncate_encode(xdr, aoffset);
1379 if (procp->pc_encode == NULL)
1383 if (svc_authorise(rqstp))
1385 return 1; /* Caller can now send it */
1388 svc_authorise(rqstp); /* doesn't hurt to call this twice */
1389 dprintk("svc: svc_process dropit\n");
1393 svc_authorise(rqstp);
1395 if (rqstp->rq_xprt && test_bit(XPT_TEMP, &rqstp->rq_xprt->xpt_flags))
1396 svc_xprt_close(rqstp->rq_xprt);
1397 dprintk("svc: svc_process close\n");
1401 svc_printk(rqstp, "short len %u, dropping request\n",
1406 serv->sv_stats->rpcbadfmt++;
1407 xdr_stream_encode_u32(xdr, RPC_MSG_DENIED);
1408 xdr_stream_encode_u32(xdr, RPC_MISMATCH);
1409 /* Only RPCv2 supported */
1410 xdr_stream_encode_u32(xdr, RPC_VERSION);
1411 xdr_stream_encode_u32(xdr, RPC_VERSION);
1412 return 1; /* don't wrap */
1415 dprintk("svc: authentication failed (%d)\n",
1416 be32_to_cpu(rqstp->rq_auth_stat));
1417 serv->sv_stats->rpcbadauth++;
1418 /* Restore write pointer to location of reply status: */
1419 xdr_truncate_encode(xdr, XDR_UNIT * 2);
1420 xdr_stream_encode_u32(xdr, RPC_MSG_DENIED);
1421 xdr_stream_encode_u32(xdr, RPC_AUTH_ERROR);
1422 xdr_stream_encode_be32(xdr, rqstp->rq_auth_stat);
1426 dprintk("svc: unknown program %d\n", rqstp->rq_prog);
1427 serv->sv_stats->rpcbadfmt++;
1428 *rqstp->rq_accept_statp = rpc_prog_unavail;
1432 svc_printk(rqstp, "unknown version (%d for prog %d, %s)\n",
1433 rqstp->rq_vers, rqstp->rq_prog, progp->pg_name);
1435 serv->sv_stats->rpcbadfmt++;
1436 *rqstp->rq_accept_statp = rpc_prog_mismatch;
1439 * svc_authenticate() has already added the verifier and
1440 * advanced the stream just past rq_accept_statp.
1442 xdr_stream_encode_u32(xdr, process.mismatch.lovers);
1443 xdr_stream_encode_u32(xdr, process.mismatch.hivers);
1447 svc_printk(rqstp, "unknown procedure (%d)\n", rqstp->rq_proc);
1449 serv->sv_stats->rpcbadfmt++;
1450 *rqstp->rq_accept_statp = rpc_proc_unavail;
1454 svc_printk(rqstp, "failed to decode RPC header\n");
1456 serv->sv_stats->rpcbadfmt++;
1457 *rqstp->rq_accept_statp = rpc_garbage_args;
1461 serv->sv_stats->rpcbadfmt++;
1462 *rqstp->rq_accept_statp = rpc_system_err;
1467 * svc_process - Execute one RPC transaction
1468 * @rqstp: RPC transaction context
1471 void svc_process(struct svc_rqst *rqstp)
1473 struct kvec *resv = &rqstp->rq_res.head[0];
1476 #if IS_ENABLED(CONFIG_FAIL_SUNRPC)
1477 if (!fail_sunrpc.ignore_server_disconnect &&
1478 should_fail(&fail_sunrpc.attr, 1))
1479 svc_xprt_deferred_close(rqstp->rq_xprt);
1483 * Setup response xdr_buf.
1484 * Initially it has just one page
1486 rqstp->rq_next_page = &rqstp->rq_respages[1];
1487 resv->iov_base = page_address(rqstp->rq_respages[0]);
1489 rqstp->rq_res.pages = rqstp->rq_next_page;
1490 rqstp->rq_res.len = 0;
1491 rqstp->rq_res.page_base = 0;
1492 rqstp->rq_res.page_len = 0;
1493 rqstp->rq_res.buflen = PAGE_SIZE;
1494 rqstp->rq_res.tail[0].iov_base = NULL;
1495 rqstp->rq_res.tail[0].iov_len = 0;
1497 svcxdr_init_decode(rqstp);
1498 p = xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 2);
1501 rqstp->rq_xid = *p++;
1502 if (unlikely(*p != rpc_call))
1505 if (!svc_process_common(rqstp))
1511 svc_printk(rqstp, "bad direction 0x%08x, dropping request\n",
1513 rqstp->rq_server->sv_stats->rpcbadfmt++;
1518 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1520 * Process a backchannel RPC request that arrived over an existing
1521 * outbound connection
1524 bc_svc_process(struct svc_serv *serv, struct rpc_rqst *req,
1525 struct svc_rqst *rqstp)
1527 struct rpc_task *task;
1531 dprintk("svc: %s(%p)\n", __func__, req);
1533 /* Build the svc_rqst used by the common processing routine */
1534 rqstp->rq_xid = req->rq_xid;
1535 rqstp->rq_prot = req->rq_xprt->prot;
1536 rqstp->rq_server = serv;
1537 rqstp->rq_bc_net = req->rq_xprt->xprt_net;
1539 rqstp->rq_addrlen = sizeof(req->rq_xprt->addr);
1540 memcpy(&rqstp->rq_addr, &req->rq_xprt->addr, rqstp->rq_addrlen);
1541 memcpy(&rqstp->rq_arg, &req->rq_rcv_buf, sizeof(rqstp->rq_arg));
1542 memcpy(&rqstp->rq_res, &req->rq_snd_buf, sizeof(rqstp->rq_res));
1544 /* Adjust the argument buffer length */
1545 rqstp->rq_arg.len = req->rq_private_buf.len;
1546 if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len) {
1547 rqstp->rq_arg.head[0].iov_len = rqstp->rq_arg.len;
1548 rqstp->rq_arg.page_len = 0;
1549 } else if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len +
1550 rqstp->rq_arg.page_len)
1551 rqstp->rq_arg.page_len = rqstp->rq_arg.len -
1552 rqstp->rq_arg.head[0].iov_len;
1554 rqstp->rq_arg.len = rqstp->rq_arg.head[0].iov_len +
1555 rqstp->rq_arg.page_len;
1557 /* Reset the response buffer */
1558 rqstp->rq_res.head[0].iov_len = 0;
1561 * Skip the XID and calldir fields because they've already
1562 * been processed by the caller.
1564 svcxdr_init_decode(rqstp);
1565 if (!xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 2)) {
1570 /* Parse and execute the bc call */
1571 proc_error = svc_process_common(rqstp);
1573 atomic_dec(&req->rq_xprt->bc_slot_count);
1575 /* Processing error: drop the request */
1576 xprt_free_bc_request(req);
1580 /* Finally, send the reply synchronously */
1581 memcpy(&req->rq_snd_buf, &rqstp->rq_res, sizeof(req->rq_snd_buf));
1582 task = rpc_run_bc_task(req);
1584 error = PTR_ERR(task);
1588 WARN_ON_ONCE(atomic_read(&task->tk_count) != 1);
1589 error = task->tk_status;
1593 dprintk("svc: %s(), error=%d\n", __func__, error);
1596 EXPORT_SYMBOL_GPL(bc_svc_process);
1597 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1600 * svc_max_payload - Return transport-specific limit on the RPC payload
1601 * @rqstp: RPC transaction context
1603 * Returns the maximum number of payload bytes the current transport
1606 u32 svc_max_payload(const struct svc_rqst *rqstp)
1608 u32 max = rqstp->rq_xprt->xpt_class->xcl_max_payload;
1610 if (rqstp->rq_server->sv_max_payload < max)
1611 max = rqstp->rq_server->sv_max_payload;
1614 EXPORT_SYMBOL_GPL(svc_max_payload);
1617 * svc_proc_name - Return RPC procedure name in string form
1618 * @rqstp: svc_rqst to operate on
1621 * Pointer to a NUL-terminated string
1623 const char *svc_proc_name(const struct svc_rqst *rqstp)
1625 if (rqstp && rqstp->rq_procinfo)
1626 return rqstp->rq_procinfo->pc_name;
1632 * svc_encode_result_payload - mark a range of bytes as a result payload
1633 * @rqstp: svc_rqst to operate on
1634 * @offset: payload's byte offset in rqstp->rq_res
1635 * @length: size of payload, in bytes
1637 * Returns zero on success, or a negative errno if a permanent
1640 int svc_encode_result_payload(struct svc_rqst *rqstp, unsigned int offset,
1641 unsigned int length)
1643 return rqstp->rq_xprt->xpt_ops->xpo_result_payload(rqstp, offset,
1646 EXPORT_SYMBOL_GPL(svc_encode_result_payload);
1649 * svc_fill_write_vector - Construct data argument for VFS write call
1650 * @rqstp: svc_rqst to operate on
1651 * @payload: xdr_buf containing only the write data payload
1653 * Fills in rqstp::rq_vec, and returns the number of elements.
1655 unsigned int svc_fill_write_vector(struct svc_rqst *rqstp,
1656 struct xdr_buf *payload)
1658 struct page **pages = payload->pages;
1659 struct kvec *first = payload->head;
1660 struct kvec *vec = rqstp->rq_vec;
1661 size_t total = payload->len;
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);