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 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 * 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);
518 percpu_counter_init(&pool->sp_threads_timedout, 0, GFP_KERNEL);
525 * svc_create - Create an RPC service
526 * @prog: the RPC program the new service will handle
527 * @bufsize: maximum message size for @prog
528 * @threadfn: a function to service RPC requests for @prog
530 * Returns an instantiated struct svc_serv object or NULL.
532 struct svc_serv *svc_create(struct svc_program *prog, unsigned int bufsize,
533 int (*threadfn)(void *data))
535 return __svc_create(prog, bufsize, 1, threadfn);
537 EXPORT_SYMBOL_GPL(svc_create);
540 * svc_create_pooled - Create an RPC service with pooled threads
541 * @prog: the RPC program the new service will handle
542 * @bufsize: maximum message size for @prog
543 * @threadfn: a function to service RPC requests for @prog
545 * Returns an instantiated struct svc_serv object or NULL.
547 struct svc_serv *svc_create_pooled(struct svc_program *prog,
548 unsigned int bufsize,
549 int (*threadfn)(void *data))
551 struct svc_serv *serv;
552 unsigned int npools = svc_pool_map_get();
554 serv = __svc_create(prog, bufsize, npools, threadfn);
559 svc_pool_map_put(npools);
562 EXPORT_SYMBOL_GPL(svc_create_pooled);
565 * Destroy an RPC service. Should be called with appropriate locking to
566 * protect sv_permsocks and sv_tempsocks.
569 svc_destroy(struct kref *ref)
571 struct svc_serv *serv = container_of(ref, struct svc_serv, sv_refcnt);
574 dprintk("svc: svc_destroy(%s)\n", serv->sv_program->pg_name);
575 timer_shutdown_sync(&serv->sv_temptimer);
578 * The last user is gone and thus all sockets have to be destroyed to
579 * the point. Check this.
581 BUG_ON(!list_empty(&serv->sv_permsocks));
582 BUG_ON(!list_empty(&serv->sv_tempsocks));
584 cache_clean_deferred(serv);
586 svc_pool_map_put(serv->sv_nrpools);
588 for (i = 0; i < serv->sv_nrpools; i++) {
589 struct svc_pool *pool = &serv->sv_pools[i];
591 percpu_counter_destroy(&pool->sp_sockets_queued);
592 percpu_counter_destroy(&pool->sp_threads_woken);
593 percpu_counter_destroy(&pool->sp_threads_timedout);
595 kfree(serv->sv_pools);
598 EXPORT_SYMBOL_GPL(svc_destroy);
601 * Allocate an RPC server's buffer space.
602 * We allocate pages and place them in rq_pages.
605 svc_init_buffer(struct svc_rqst *rqstp, unsigned int size, int node)
607 unsigned int pages, arghi;
609 /* bc_xprt uses fore channel allocated buffers */
610 if (svc_is_backchannel(rqstp))
613 pages = size / PAGE_SIZE + 1; /* extra page as we hold both request and reply.
614 * We assume one is at most one page
617 WARN_ON_ONCE(pages > RPCSVC_MAXPAGES);
618 if (pages > RPCSVC_MAXPAGES)
619 pages = RPCSVC_MAXPAGES;
621 struct page *p = alloc_pages_node(node, GFP_KERNEL, 0);
624 rqstp->rq_pages[arghi++] = p;
631 * Release an RPC server buffer
634 svc_release_buffer(struct svc_rqst *rqstp)
638 for (i = 0; i < ARRAY_SIZE(rqstp->rq_pages); i++)
639 if (rqstp->rq_pages[i])
640 put_page(rqstp->rq_pages[i]);
644 svc_rqst_alloc(struct svc_serv *serv, struct svc_pool *pool, int node)
646 struct svc_rqst *rqstp;
648 rqstp = kzalloc_node(sizeof(*rqstp), GFP_KERNEL, node);
652 __set_bit(RQ_BUSY, &rqstp->rq_flags);
653 rqstp->rq_server = serv;
654 rqstp->rq_pool = pool;
656 rqstp->rq_scratch_page = alloc_pages_node(node, GFP_KERNEL, 0);
657 if (!rqstp->rq_scratch_page)
660 rqstp->rq_argp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
664 rqstp->rq_resp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
668 if (!svc_init_buffer(rqstp, serv->sv_max_mesg, node))
673 svc_rqst_free(rqstp);
676 EXPORT_SYMBOL_GPL(svc_rqst_alloc);
678 static struct svc_rqst *
679 svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
681 struct svc_rqst *rqstp;
683 rqstp = svc_rqst_alloc(serv, pool, node);
685 return ERR_PTR(-ENOMEM);
688 spin_lock_bh(&serv->sv_lock);
689 serv->sv_nrthreads += 1;
690 spin_unlock_bh(&serv->sv_lock);
692 spin_lock_bh(&pool->sp_lock);
693 pool->sp_nrthreads++;
694 list_add_rcu(&rqstp->rq_all, &pool->sp_all_threads);
695 spin_unlock_bh(&pool->sp_lock);
700 * Choose a pool in which to create a new thread, for svc_set_num_threads
702 static inline struct svc_pool *
703 choose_pool(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
708 return &serv->sv_pools[(*state)++ % serv->sv_nrpools];
712 * Choose a thread to kill, for svc_set_num_threads
714 static inline struct task_struct *
715 choose_victim(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
718 struct task_struct *task = NULL;
721 spin_lock_bh(&pool->sp_lock);
723 /* choose a pool in round-robin fashion */
724 for (i = 0; i < serv->sv_nrpools; i++) {
725 pool = &serv->sv_pools[--(*state) % serv->sv_nrpools];
726 spin_lock_bh(&pool->sp_lock);
727 if (!list_empty(&pool->sp_all_threads))
729 spin_unlock_bh(&pool->sp_lock);
735 if (!list_empty(&pool->sp_all_threads)) {
736 struct svc_rqst *rqstp;
739 * Remove from the pool->sp_all_threads list
740 * so we don't try to kill it again.
742 rqstp = list_entry(pool->sp_all_threads.next, struct svc_rqst, rq_all);
743 set_bit(RQ_VICTIM, &rqstp->rq_flags);
744 list_del_rcu(&rqstp->rq_all);
745 task = rqstp->rq_task;
747 spin_unlock_bh(&pool->sp_lock);
752 /* create new threads */
754 svc_start_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
756 struct svc_rqst *rqstp;
757 struct task_struct *task;
758 struct svc_pool *chosen_pool;
759 unsigned int state = serv->sv_nrthreads-1;
764 chosen_pool = choose_pool(serv, pool, &state);
766 node = svc_pool_map_get_node(chosen_pool->sp_id);
767 rqstp = svc_prepare_thread(serv, chosen_pool, node);
769 return PTR_ERR(rqstp);
771 task = kthread_create_on_node(serv->sv_threadfn, rqstp,
772 node, "%s", serv->sv_name);
774 svc_exit_thread(rqstp);
775 return PTR_ERR(task);
778 rqstp->rq_task = task;
779 if (serv->sv_nrpools > 1)
780 svc_pool_map_set_cpumask(task, chosen_pool->sp_id);
782 svc_sock_update_bufs(serv);
783 wake_up_process(task);
784 } while (nrservs > 0);
790 * Create or destroy enough new threads to make the number
791 * of threads the given number. If `pool' is non-NULL, applies
792 * only to threads in that pool, otherwise round-robins between
793 * all pools. Caller must ensure that mutual exclusion between this and
794 * server startup or shutdown.
797 /* destroy old threads */
799 svc_stop_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
801 struct svc_rqst *rqstp;
802 struct task_struct *task;
803 unsigned int state = serv->sv_nrthreads-1;
805 /* destroy old threads */
807 task = choose_victim(serv, pool, &state);
810 rqstp = kthread_data(task);
811 /* Did we lose a race to svo_function threadfn? */
812 if (kthread_stop(task) == -EINTR)
813 svc_exit_thread(rqstp);
815 } while (nrservs < 0);
820 svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
823 nrservs -= serv->sv_nrthreads;
825 spin_lock_bh(&pool->sp_lock);
826 nrservs -= pool->sp_nrthreads;
827 spin_unlock_bh(&pool->sp_lock);
831 return svc_start_kthreads(serv, pool, nrservs);
833 return svc_stop_kthreads(serv, pool, nrservs);
836 EXPORT_SYMBOL_GPL(svc_set_num_threads);
839 * svc_rqst_replace_page - Replace one page in rq_pages[]
840 * @rqstp: svc_rqst with pages to replace
841 * @page: replacement page
843 * When replacing a page in rq_pages, batch the release of the
844 * replaced pages to avoid hammering the page allocator.
846 void svc_rqst_replace_page(struct svc_rqst *rqstp, struct page *page)
848 if (*rqstp->rq_next_page) {
849 if (!pagevec_space(&rqstp->rq_pvec))
850 __pagevec_release(&rqstp->rq_pvec);
851 pagevec_add(&rqstp->rq_pvec, *rqstp->rq_next_page);
855 *(rqstp->rq_next_page++) = page;
857 EXPORT_SYMBOL_GPL(svc_rqst_replace_page);
860 * Called from a server thread as it's exiting. Caller must hold the "service
861 * mutex" for the service.
864 svc_rqst_free(struct svc_rqst *rqstp)
866 svc_release_buffer(rqstp);
867 if (rqstp->rq_scratch_page)
868 put_page(rqstp->rq_scratch_page);
869 kfree(rqstp->rq_resp);
870 kfree(rqstp->rq_argp);
871 kfree(rqstp->rq_auth_data);
872 kfree_rcu(rqstp, rq_rcu_head);
874 EXPORT_SYMBOL_GPL(svc_rqst_free);
877 svc_exit_thread(struct svc_rqst *rqstp)
879 struct svc_serv *serv = rqstp->rq_server;
880 struct svc_pool *pool = rqstp->rq_pool;
882 spin_lock_bh(&pool->sp_lock);
883 pool->sp_nrthreads--;
884 if (!test_and_set_bit(RQ_VICTIM, &rqstp->rq_flags))
885 list_del_rcu(&rqstp->rq_all);
886 spin_unlock_bh(&pool->sp_lock);
888 spin_lock_bh(&serv->sv_lock);
889 serv->sv_nrthreads -= 1;
890 spin_unlock_bh(&serv->sv_lock);
891 svc_sock_update_bufs(serv);
893 svc_rqst_free(rqstp);
897 EXPORT_SYMBOL_GPL(svc_exit_thread);
900 * Register an "inet" protocol family netid with the local
901 * rpcbind daemon via an rpcbind v4 SET request.
903 * No netconfig infrastructure is available in the kernel, so
904 * we map IP_ protocol numbers to netids by hand.
906 * Returns zero on success; a negative errno value is returned
907 * if any error occurs.
909 static int __svc_rpcb_register4(struct net *net, const u32 program,
911 const unsigned short protocol,
912 const unsigned short port)
914 const struct sockaddr_in sin = {
915 .sin_family = AF_INET,
916 .sin_addr.s_addr = htonl(INADDR_ANY),
917 .sin_port = htons(port),
924 netid = RPCBIND_NETID_UDP;
927 netid = RPCBIND_NETID_TCP;
933 error = rpcb_v4_register(net, program, version,
934 (const struct sockaddr *)&sin, netid);
937 * User space didn't support rpcbind v4, so retry this
938 * registration request with the legacy rpcbind v2 protocol.
940 if (error == -EPROTONOSUPPORT)
941 error = rpcb_register(net, program, version, protocol, port);
946 #if IS_ENABLED(CONFIG_IPV6)
948 * Register an "inet6" protocol family netid with the local
949 * rpcbind daemon via an rpcbind v4 SET request.
951 * No netconfig infrastructure is available in the kernel, so
952 * we map IP_ protocol numbers to netids by hand.
954 * Returns zero on success; a negative errno value is returned
955 * if any error occurs.
957 static int __svc_rpcb_register6(struct net *net, const u32 program,
959 const unsigned short protocol,
960 const unsigned short port)
962 const struct sockaddr_in6 sin6 = {
963 .sin6_family = AF_INET6,
964 .sin6_addr = IN6ADDR_ANY_INIT,
965 .sin6_port = htons(port),
972 netid = RPCBIND_NETID_UDP6;
975 netid = RPCBIND_NETID_TCP6;
981 error = rpcb_v4_register(net, program, version,
982 (const struct sockaddr *)&sin6, netid);
985 * User space didn't support rpcbind version 4, so we won't
986 * use a PF_INET6 listener.
988 if (error == -EPROTONOSUPPORT)
989 error = -EAFNOSUPPORT;
993 #endif /* IS_ENABLED(CONFIG_IPV6) */
996 * Register a kernel RPC service via rpcbind version 4.
998 * Returns zero on success; a negative errno value is returned
999 * if any error occurs.
1001 static int __svc_register(struct net *net, const char *progname,
1002 const u32 program, const u32 version,
1004 const unsigned short protocol,
1005 const unsigned short port)
1007 int error = -EAFNOSUPPORT;
1011 error = __svc_rpcb_register4(net, program, version,
1014 #if IS_ENABLED(CONFIG_IPV6)
1016 error = __svc_rpcb_register6(net, program, version,
1021 trace_svc_register(progname, version, protocol, port, family, error);
1025 int svc_rpcbind_set_version(struct net *net,
1026 const struct svc_program *progp,
1027 u32 version, int family,
1028 unsigned short proto,
1029 unsigned short port)
1031 return __svc_register(net, progp->pg_name, progp->pg_prog,
1032 version, family, proto, port);
1035 EXPORT_SYMBOL_GPL(svc_rpcbind_set_version);
1037 int svc_generic_rpcbind_set(struct net *net,
1038 const struct svc_program *progp,
1039 u32 version, int family,
1040 unsigned short proto,
1041 unsigned short port)
1043 const struct svc_version *vers = progp->pg_vers[version];
1049 if (vers->vs_hidden) {
1050 trace_svc_noregister(progp->pg_name, version, proto,
1056 * Don't register a UDP port if we need congestion
1059 if (vers->vs_need_cong_ctrl && proto == IPPROTO_UDP)
1062 error = svc_rpcbind_set_version(net, progp, version,
1063 family, proto, port);
1065 return (vers->vs_rpcb_optnl) ? 0 : error;
1067 EXPORT_SYMBOL_GPL(svc_generic_rpcbind_set);
1070 * svc_register - register an RPC service with the local portmapper
1071 * @serv: svc_serv struct for the service to register
1072 * @net: net namespace for the service to register
1073 * @family: protocol family of service's listener socket
1074 * @proto: transport protocol number to advertise
1075 * @port: port to advertise
1077 * Service is registered for any address in the passed-in protocol family
1079 int svc_register(const struct svc_serv *serv, struct net *net,
1080 const int family, const unsigned short proto,
1081 const unsigned short port)
1083 struct svc_program *progp;
1087 WARN_ON_ONCE(proto == 0 && port == 0);
1088 if (proto == 0 && port == 0)
1091 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1092 for (i = 0; i < progp->pg_nvers; i++) {
1094 error = progp->pg_rpcbind_set(net, progp, i,
1095 family, proto, port);
1097 printk(KERN_WARNING "svc: failed to register "
1098 "%sv%u RPC service (errno %d).\n",
1099 progp->pg_name, i, -error);
1109 * If user space is running rpcbind, it should take the v4 UNSET
1110 * and clear everything for this [program, version]. If user space
1111 * is running portmap, it will reject the v4 UNSET, but won't have
1112 * any "inet6" entries anyway. So a PMAP_UNSET should be sufficient
1113 * in this case to clear all existing entries for [program, version].
1115 static void __svc_unregister(struct net *net, const u32 program, const u32 version,
1116 const char *progname)
1120 error = rpcb_v4_register(net, program, version, NULL, "");
1123 * User space didn't support rpcbind v4, so retry this
1124 * request with the legacy rpcbind v2 protocol.
1126 if (error == -EPROTONOSUPPORT)
1127 error = rpcb_register(net, program, version, 0, 0);
1129 trace_svc_unregister(progname, version, error);
1133 * All netids, bind addresses and ports registered for [program, version]
1134 * are removed from the local rpcbind database (if the service is not
1135 * hidden) to make way for a new instance of the service.
1137 * The result of unregistration is reported via dprintk for those who want
1138 * verification of the result, but is otherwise not important.
1140 static void svc_unregister(const struct svc_serv *serv, struct net *net)
1142 struct svc_program *progp;
1143 unsigned long flags;
1146 clear_thread_flag(TIF_SIGPENDING);
1148 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1149 for (i = 0; i < progp->pg_nvers; i++) {
1150 if (progp->pg_vers[i] == NULL)
1152 if (progp->pg_vers[i]->vs_hidden)
1154 __svc_unregister(net, progp->pg_prog, i, progp->pg_name);
1158 spin_lock_irqsave(¤t->sighand->siglock, flags);
1159 recalc_sigpending();
1160 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
1164 * dprintk the given error with the address of the client that caused it.
1166 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
1167 static __printf(2, 3)
1168 void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...)
1170 struct va_format vaf;
1172 char buf[RPC_MAX_ADDRBUFLEN];
1174 va_start(args, fmt);
1179 dprintk("svc: %s: %pV", svc_print_addr(rqstp, buf, sizeof(buf)), &vaf);
1184 static __printf(2,3) void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...) {}
1188 svc_generic_init_request(struct svc_rqst *rqstp,
1189 const struct svc_program *progp,
1190 struct svc_process_info *ret)
1192 const struct svc_version *versp = NULL; /* compiler food */
1193 const struct svc_procedure *procp = NULL;
1195 if (rqstp->rq_vers >= progp->pg_nvers )
1197 versp = progp->pg_vers[rqstp->rq_vers];
1202 * Some protocol versions (namely NFSv4) require some form of
1203 * congestion control. (See RFC 7530 section 3.1 paragraph 2)
1204 * In other words, UDP is not allowed. We mark those when setting
1205 * up the svc_xprt, and verify that here.
1207 * The spec is not very clear about what error should be returned
1208 * when someone tries to access a server that is listening on UDP
1209 * for lower versions. RPC_PROG_MISMATCH seems to be the closest
1212 if (versp->vs_need_cong_ctrl && rqstp->rq_xprt &&
1213 !test_bit(XPT_CONG_CTRL, &rqstp->rq_xprt->xpt_flags))
1216 if (rqstp->rq_proc >= versp->vs_nproc)
1218 rqstp->rq_procinfo = procp = &versp->vs_proc[rqstp->rq_proc];
1222 /* Initialize storage for argp and resp */
1223 memset(rqstp->rq_argp, 0, procp->pc_argzero);
1224 memset(rqstp->rq_resp, 0, procp->pc_ressize);
1226 /* Bump per-procedure stats counter */
1227 this_cpu_inc(versp->vs_count[rqstp->rq_proc]);
1229 ret->dispatch = versp->vs_dispatch;
1232 ret->mismatch.lovers = progp->pg_lovers;
1233 ret->mismatch.hivers = progp->pg_hivers;
1234 return rpc_prog_mismatch;
1236 return rpc_proc_unavail;
1238 EXPORT_SYMBOL_GPL(svc_generic_init_request);
1241 * Common routine for processing the RPC request.
1244 svc_process_common(struct svc_rqst *rqstp)
1246 struct xdr_stream *xdr = &rqstp->rq_res_stream;
1247 struct svc_program *progp;
1248 const struct svc_procedure *procp = NULL;
1249 struct svc_serv *serv = rqstp->rq_server;
1250 struct svc_process_info process;
1252 unsigned int aoffset;
1255 /* Will be turned off by GSS integrity and privacy services */
1256 set_bit(RQ_SPLICE_OK, &rqstp->rq_flags);
1257 /* Will be turned off only when NFSv4 Sessions are used */
1258 set_bit(RQ_USEDEFERRAL, &rqstp->rq_flags);
1259 clear_bit(RQ_DROPME, &rqstp->rq_flags);
1261 /* Construct the first words of the reply: */
1262 svcxdr_init_encode(rqstp);
1263 xdr_stream_encode_be32(xdr, rqstp->rq_xid);
1264 xdr_stream_encode_be32(xdr, rpc_reply);
1266 p = xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 4);
1269 if (*p++ != cpu_to_be32(RPC_VERSION))
1272 xdr_stream_encode_be32(xdr, rpc_msg_accepted);
1274 rqstp->rq_prog = be32_to_cpup(p++);
1275 rqstp->rq_vers = be32_to_cpup(p++);
1276 rqstp->rq_proc = be32_to_cpup(p);
1278 for (progp = serv->sv_program; progp; progp = progp->pg_next)
1279 if (rqstp->rq_prog == progp->pg_prog)
1283 * Decode auth data, and add verifier to reply buffer.
1284 * We do this before anything else in order to get a decent
1287 auth_res = svc_authenticate(rqstp);
1288 /* Also give the program a chance to reject this call: */
1289 if (auth_res == SVC_OK && progp)
1290 auth_res = progp->pg_authenticate(rqstp);
1291 trace_svc_authenticate(rqstp, auth_res);
1296 goto err_garbage_args;
1298 goto err_system_err;
1312 switch (progp->pg_init_request(rqstp, progp, &process)) {
1315 case rpc_prog_unavail:
1317 case rpc_prog_mismatch:
1319 case rpc_proc_unavail:
1323 procp = rqstp->rq_procinfo;
1324 /* Should this check go into the dispatcher? */
1325 if (!procp || !procp->pc_func)
1328 /* Syntactic check complete */
1329 serv->sv_stats->rpccnt++;
1330 trace_svc_process(rqstp, progp->pg_name);
1332 aoffset = xdr_stream_pos(xdr);
1334 /* un-reserve some of the out-queue now that we have a
1335 * better idea of reply size
1337 if (procp->pc_xdrressize)
1338 svc_reserve_auth(rqstp, procp->pc_xdrressize<<2);
1340 /* Call the function that processes the request. */
1341 rc = process.dispatch(rqstp);
1342 if (procp->pc_release)
1343 procp->pc_release(rqstp);
1346 if (rqstp->rq_auth_stat != rpc_auth_ok)
1349 if (*rqstp->rq_accept_statp != rpc_success)
1350 xdr_truncate_encode(xdr, aoffset);
1352 if (procp->pc_encode == NULL)
1356 if (svc_authorise(rqstp))
1358 return 1; /* Caller can now send it */
1361 svc_authorise(rqstp); /* doesn't hurt to call this twice */
1362 dprintk("svc: svc_process dropit\n");
1366 svc_authorise(rqstp);
1368 if (rqstp->rq_xprt && test_bit(XPT_TEMP, &rqstp->rq_xprt->xpt_flags))
1369 svc_xprt_close(rqstp->rq_xprt);
1370 dprintk("svc: svc_process close\n");
1374 svc_printk(rqstp, "short len %u, dropping request\n",
1379 serv->sv_stats->rpcbadfmt++;
1380 xdr_stream_encode_u32(xdr, RPC_MSG_DENIED);
1381 xdr_stream_encode_u32(xdr, RPC_MISMATCH);
1382 /* Only RPCv2 supported */
1383 xdr_stream_encode_u32(xdr, RPC_VERSION);
1384 xdr_stream_encode_u32(xdr, RPC_VERSION);
1388 dprintk("svc: authentication failed (%d)\n",
1389 be32_to_cpu(rqstp->rq_auth_stat));
1390 serv->sv_stats->rpcbadauth++;
1391 /* Restore write pointer to location of reply status: */
1392 xdr_truncate_encode(xdr, XDR_UNIT * 2);
1393 xdr_stream_encode_u32(xdr, RPC_MSG_DENIED);
1394 xdr_stream_encode_u32(xdr, RPC_AUTH_ERROR);
1395 xdr_stream_encode_be32(xdr, rqstp->rq_auth_stat);
1399 dprintk("svc: unknown program %d\n", rqstp->rq_prog);
1400 serv->sv_stats->rpcbadfmt++;
1401 xdr_stream_encode_u32(xdr, RPC_PROG_UNAVAIL);
1405 svc_printk(rqstp, "unknown version (%d for prog %d, %s)\n",
1406 rqstp->rq_vers, rqstp->rq_prog, progp->pg_name);
1408 serv->sv_stats->rpcbadfmt++;
1409 xdr_stream_encode_u32(xdr, RPC_PROG_MISMATCH);
1410 xdr_stream_encode_u32(xdr, process.mismatch.lovers);
1411 xdr_stream_encode_u32(xdr, process.mismatch.hivers);
1415 svc_printk(rqstp, "unknown procedure (%d)\n", rqstp->rq_proc);
1417 serv->sv_stats->rpcbadfmt++;
1418 xdr_stream_encode_u32(xdr, RPC_PROC_UNAVAIL);
1422 svc_printk(rqstp, "failed to decode RPC header\n");
1424 serv->sv_stats->rpcbadfmt++;
1425 xdr_stream_encode_u32(xdr, RPC_GARBAGE_ARGS);
1429 serv->sv_stats->rpcbadfmt++;
1430 xdr_stream_encode_u32(xdr, RPC_SYSTEM_ERR);
1435 * Process the RPC request.
1438 svc_process(struct svc_rqst *rqstp)
1440 struct kvec *resv = &rqstp->rq_res.head[0];
1443 #if IS_ENABLED(CONFIG_FAIL_SUNRPC)
1444 if (!fail_sunrpc.ignore_server_disconnect &&
1445 should_fail(&fail_sunrpc.attr, 1))
1446 svc_xprt_deferred_close(rqstp->rq_xprt);
1450 * Setup response xdr_buf.
1451 * Initially it has just one page
1453 rqstp->rq_next_page = &rqstp->rq_respages[1];
1454 resv->iov_base = page_address(rqstp->rq_respages[0]);
1456 rqstp->rq_res.pages = rqstp->rq_next_page;
1457 rqstp->rq_res.len = 0;
1458 rqstp->rq_res.page_base = 0;
1459 rqstp->rq_res.page_len = 0;
1460 rqstp->rq_res.buflen = PAGE_SIZE;
1461 rqstp->rq_res.tail[0].iov_base = NULL;
1462 rqstp->rq_res.tail[0].iov_len = 0;
1464 svcxdr_init_decode(rqstp);
1465 p = xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 2);
1468 rqstp->rq_xid = *p++;
1469 if (unlikely(*p != rpc_call))
1472 if (!svc_process_common(rqstp))
1474 return svc_send(rqstp);
1477 svc_printk(rqstp, "bad direction 0x%08x, dropping request\n",
1479 rqstp->rq_server->sv_stats->rpcbadfmt++;
1484 EXPORT_SYMBOL_GPL(svc_process);
1486 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1488 * Process a backchannel RPC request that arrived over an existing
1489 * outbound connection
1492 bc_svc_process(struct svc_serv *serv, struct rpc_rqst *req,
1493 struct svc_rqst *rqstp)
1495 struct rpc_task *task;
1499 dprintk("svc: %s(%p)\n", __func__, req);
1501 /* Build the svc_rqst used by the common processing routine */
1502 rqstp->rq_xid = req->rq_xid;
1503 rqstp->rq_prot = req->rq_xprt->prot;
1504 rqstp->rq_server = serv;
1505 rqstp->rq_bc_net = req->rq_xprt->xprt_net;
1507 rqstp->rq_addrlen = sizeof(req->rq_xprt->addr);
1508 memcpy(&rqstp->rq_addr, &req->rq_xprt->addr, rqstp->rq_addrlen);
1509 memcpy(&rqstp->rq_arg, &req->rq_rcv_buf, sizeof(rqstp->rq_arg));
1510 memcpy(&rqstp->rq_res, &req->rq_snd_buf, sizeof(rqstp->rq_res));
1512 /* Adjust the argument buffer length */
1513 rqstp->rq_arg.len = req->rq_private_buf.len;
1514 if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len) {
1515 rqstp->rq_arg.head[0].iov_len = rqstp->rq_arg.len;
1516 rqstp->rq_arg.page_len = 0;
1517 } else if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len +
1518 rqstp->rq_arg.page_len)
1519 rqstp->rq_arg.page_len = rqstp->rq_arg.len -
1520 rqstp->rq_arg.head[0].iov_len;
1522 rqstp->rq_arg.len = rqstp->rq_arg.head[0].iov_len +
1523 rqstp->rq_arg.page_len;
1525 /* Reset the response buffer */
1526 rqstp->rq_res.head[0].iov_len = 0;
1529 * Skip the XID and calldir fields because they've already
1530 * been processed by the caller.
1532 svcxdr_init_decode(rqstp);
1533 if (!xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 2)) {
1538 /* Parse and execute the bc call */
1539 proc_error = svc_process_common(rqstp);
1541 atomic_dec(&req->rq_xprt->bc_slot_count);
1543 /* Processing error: drop the request */
1544 xprt_free_bc_request(req);
1548 /* Finally, send the reply synchronously */
1549 memcpy(&req->rq_snd_buf, &rqstp->rq_res, sizeof(req->rq_snd_buf));
1550 task = rpc_run_bc_task(req);
1552 error = PTR_ERR(task);
1556 WARN_ON_ONCE(atomic_read(&task->tk_count) != 1);
1557 error = task->tk_status;
1561 dprintk("svc: %s(), error=%d\n", __func__, error);
1564 EXPORT_SYMBOL_GPL(bc_svc_process);
1565 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1568 * svc_max_payload - Return transport-specific limit on the RPC payload
1569 * @rqstp: RPC transaction context
1571 * Returns the maximum number of payload bytes the current transport
1574 u32 svc_max_payload(const struct svc_rqst *rqstp)
1576 u32 max = rqstp->rq_xprt->xpt_class->xcl_max_payload;
1578 if (rqstp->rq_server->sv_max_payload < max)
1579 max = rqstp->rq_server->sv_max_payload;
1582 EXPORT_SYMBOL_GPL(svc_max_payload);
1585 * svc_proc_name - Return RPC procedure name in string form
1586 * @rqstp: svc_rqst to operate on
1589 * Pointer to a NUL-terminated string
1591 const char *svc_proc_name(const struct svc_rqst *rqstp)
1593 if (rqstp && rqstp->rq_procinfo)
1594 return rqstp->rq_procinfo->pc_name;
1600 * svc_encode_result_payload - mark a range of bytes as a result payload
1601 * @rqstp: svc_rqst to operate on
1602 * @offset: payload's byte offset in rqstp->rq_res
1603 * @length: size of payload, in bytes
1605 * Returns zero on success, or a negative errno if a permanent
1608 int svc_encode_result_payload(struct svc_rqst *rqstp, unsigned int offset,
1609 unsigned int length)
1611 return rqstp->rq_xprt->xpt_ops->xpo_result_payload(rqstp, offset,
1614 EXPORT_SYMBOL_GPL(svc_encode_result_payload);
1617 * svc_fill_write_vector - Construct data argument for VFS write call
1618 * @rqstp: svc_rqst to operate on
1619 * @payload: xdr_buf containing only the write data payload
1621 * Fills in rqstp::rq_vec, and returns the number of elements.
1623 unsigned int svc_fill_write_vector(struct svc_rqst *rqstp,
1624 struct xdr_buf *payload)
1626 struct page **pages = payload->pages;
1627 struct kvec *first = payload->head;
1628 struct kvec *vec = rqstp->rq_vec;
1629 size_t total = payload->len;
1632 /* Some types of transport can present the write payload
1633 * entirely in rq_arg.pages. In this case, @first is empty.
1636 if (first->iov_len) {
1637 vec[i].iov_base = first->iov_base;
1638 vec[i].iov_len = min_t(size_t, total, first->iov_len);
1639 total -= vec[i].iov_len;
1644 vec[i].iov_base = page_address(*pages);
1645 vec[i].iov_len = min_t(size_t, total, PAGE_SIZE);
1646 total -= vec[i].iov_len;
1651 WARN_ON_ONCE(i > ARRAY_SIZE(rqstp->rq_vec));
1654 EXPORT_SYMBOL_GPL(svc_fill_write_vector);
1657 * svc_fill_symlink_pathname - Construct pathname argument for VFS symlink call
1658 * @rqstp: svc_rqst to operate on
1659 * @first: buffer containing first section of pathname
1660 * @p: buffer containing remaining section of pathname
1661 * @total: total length of the pathname argument
1663 * The VFS symlink API demands a NUL-terminated pathname in mapped memory.
1664 * Returns pointer to a NUL-terminated string, or an ERR_PTR. Caller must free
1665 * the returned string.
1667 char *svc_fill_symlink_pathname(struct svc_rqst *rqstp, struct kvec *first,
1668 void *p, size_t total)
1670 size_t len, remaining;
1673 result = kmalloc(total + 1, GFP_KERNEL);
1675 return ERR_PTR(-ESERVERFAULT);
1680 len = min_t(size_t, total, first->iov_len);
1682 memcpy(dst, first->iov_base, len);
1688 len = min_t(size_t, remaining, PAGE_SIZE);
1689 memcpy(dst, p, len);
1695 /* Sanity check: Linux doesn't allow the pathname argument to
1696 * contain a NUL byte.
1698 if (strlen(result) != total) {
1700 return ERR_PTR(-EINVAL);
1704 EXPORT_SYMBOL_GPL(svc_fill_symlink_pathname);