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 lwq_init(&serv->sv_cb_list);
445 __svc_init_bc(struct svc_serv *serv)
451 * Create an RPC service
453 static struct svc_serv *
454 __svc_create(struct svc_program *prog, struct svc_stat *stats,
455 unsigned int bufsize, int npools, int (*threadfn)(void *data))
457 struct svc_serv *serv;
459 unsigned int xdrsize;
462 if (!(serv = kzalloc(sizeof(*serv), GFP_KERNEL)))
464 serv->sv_name = prog->pg_name;
465 serv->sv_program = prog;
466 serv->sv_stats = stats;
467 if (bufsize > RPCSVC_MAXPAYLOAD)
468 bufsize = RPCSVC_MAXPAYLOAD;
469 serv->sv_max_payload = bufsize? bufsize : 4096;
470 serv->sv_max_mesg = roundup(serv->sv_max_payload + PAGE_SIZE, PAGE_SIZE);
471 serv->sv_threadfn = threadfn;
474 prog->pg_lovers = prog->pg_nvers-1;
475 for (vers=0; vers<prog->pg_nvers ; vers++)
476 if (prog->pg_vers[vers]) {
477 prog->pg_hivers = vers;
478 if (prog->pg_lovers > vers)
479 prog->pg_lovers = vers;
480 if (prog->pg_vers[vers]->vs_xdrsize > xdrsize)
481 xdrsize = prog->pg_vers[vers]->vs_xdrsize;
483 prog = prog->pg_next;
485 serv->sv_xdrsize = xdrsize;
486 INIT_LIST_HEAD(&serv->sv_tempsocks);
487 INIT_LIST_HEAD(&serv->sv_permsocks);
488 timer_setup(&serv->sv_temptimer, NULL, 0);
489 spin_lock_init(&serv->sv_lock);
493 serv->sv_nrpools = npools;
495 kcalloc(serv->sv_nrpools, sizeof(struct svc_pool),
497 if (!serv->sv_pools) {
502 for (i = 0; i < serv->sv_nrpools; i++) {
503 struct svc_pool *pool = &serv->sv_pools[i];
505 dprintk("svc: initialising pool %u for %s\n",
509 lwq_init(&pool->sp_xprts);
510 INIT_LIST_HEAD(&pool->sp_all_threads);
511 init_llist_head(&pool->sp_idle_threads);
513 percpu_counter_init(&pool->sp_messages_arrived, 0, GFP_KERNEL);
514 percpu_counter_init(&pool->sp_sockets_queued, 0, GFP_KERNEL);
515 percpu_counter_init(&pool->sp_threads_woken, 0, GFP_KERNEL);
522 * svc_create - Create an RPC service
523 * @prog: the RPC program the new service will handle
524 * @bufsize: maximum message size for @prog
525 * @threadfn: a function to service RPC requests for @prog
527 * Returns an instantiated struct svc_serv object or NULL.
529 struct svc_serv *svc_create(struct svc_program *prog, unsigned int bufsize,
530 int (*threadfn)(void *data))
532 return __svc_create(prog, NULL, bufsize, 1, threadfn);
534 EXPORT_SYMBOL_GPL(svc_create);
537 * svc_create_pooled - Create an RPC service with pooled threads
538 * @prog: the RPC program the new service will handle
539 * @stats: the stats struct if desired
540 * @bufsize: maximum message size for @prog
541 * @threadfn: a function to service RPC requests for @prog
543 * Returns an instantiated struct svc_serv object or NULL.
545 struct svc_serv *svc_create_pooled(struct svc_program *prog,
546 struct svc_stat *stats,
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, stats, 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 svc_serv **servp)
570 struct svc_serv *serv = *servp;
575 dprintk("svc: svc_destroy(%s)\n", serv->sv_program->pg_name);
576 timer_shutdown_sync(&serv->sv_temptimer);
579 * Remaining transports at this point are not expected.
581 WARN_ONCE(!list_empty(&serv->sv_permsocks),
582 "SVC: permsocks remain for %s\n", serv->sv_program->pg_name);
583 WARN_ONCE(!list_empty(&serv->sv_tempsocks),
584 "SVC: tempsocks remain for %s\n", serv->sv_program->pg_name);
586 cache_clean_deferred(serv);
588 svc_pool_map_put(serv->sv_nrpools);
590 for (i = 0; i < serv->sv_nrpools; i++) {
591 struct svc_pool *pool = &serv->sv_pools[i];
593 percpu_counter_destroy(&pool->sp_messages_arrived);
594 percpu_counter_destroy(&pool->sp_sockets_queued);
595 percpu_counter_destroy(&pool->sp_threads_woken);
597 kfree(serv->sv_pools);
600 EXPORT_SYMBOL_GPL(svc_destroy);
603 svc_init_buffer(struct svc_rqst *rqstp, unsigned int size, int node)
605 unsigned long pages, ret;
607 /* bc_xprt uses fore channel allocated buffers */
608 if (svc_is_backchannel(rqstp))
611 pages = size / PAGE_SIZE + 1; /* extra page as we hold both request and reply.
612 * We assume one is at most one page
614 WARN_ON_ONCE(pages > RPCSVC_MAXPAGES);
615 if (pages > RPCSVC_MAXPAGES)
616 pages = RPCSVC_MAXPAGES;
618 ret = alloc_pages_bulk_array_node(GFP_KERNEL, node, pages,
624 * Release an RPC server buffer
627 svc_release_buffer(struct svc_rqst *rqstp)
631 for (i = 0; i < ARRAY_SIZE(rqstp->rq_pages); i++)
632 if (rqstp->rq_pages[i])
633 put_page(rqstp->rq_pages[i]);
637 svc_rqst_alloc(struct svc_serv *serv, struct svc_pool *pool, int node)
639 struct svc_rqst *rqstp;
641 rqstp = kzalloc_node(sizeof(*rqstp), GFP_KERNEL, node);
645 folio_batch_init(&rqstp->rq_fbatch);
647 rqstp->rq_server = serv;
648 rqstp->rq_pool = pool;
650 rqstp->rq_scratch_page = alloc_pages_node(node, GFP_KERNEL, 0);
651 if (!rqstp->rq_scratch_page)
654 rqstp->rq_argp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
658 rqstp->rq_resp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
662 if (!svc_init_buffer(rqstp, serv->sv_max_mesg, node))
667 svc_rqst_free(rqstp);
670 EXPORT_SYMBOL_GPL(svc_rqst_alloc);
672 static struct svc_rqst *
673 svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
675 struct svc_rqst *rqstp;
677 rqstp = svc_rqst_alloc(serv, pool, node);
679 return ERR_PTR(-ENOMEM);
681 spin_lock_bh(&serv->sv_lock);
682 serv->sv_nrthreads += 1;
683 spin_unlock_bh(&serv->sv_lock);
685 atomic_inc(&pool->sp_nrthreads);
687 /* Protected by whatever lock the service uses when calling
688 * svc_set_num_threads()
690 list_add_rcu(&rqstp->rq_all, &pool->sp_all_threads);
696 * svc_pool_wake_idle_thread - Awaken an idle thread in @pool
697 * @pool: service thread pool
699 * Can be called from soft IRQ or process context. Finding an idle
700 * service thread and marking it BUSY is atomic with respect to
701 * other calls to svc_pool_wake_idle_thread().
704 void svc_pool_wake_idle_thread(struct svc_pool *pool)
706 struct svc_rqst *rqstp;
707 struct llist_node *ln;
710 ln = READ_ONCE(pool->sp_idle_threads.first);
712 rqstp = llist_entry(ln, struct svc_rqst, rq_idle);
713 WRITE_ONCE(rqstp->rq_qtime, ktime_get());
714 if (!task_is_running(rqstp->rq_task)) {
715 wake_up_process(rqstp->rq_task);
716 trace_svc_wake_up(rqstp->rq_task->pid);
717 percpu_counter_inc(&pool->sp_threads_woken);
725 EXPORT_SYMBOL_GPL(svc_pool_wake_idle_thread);
727 static struct svc_pool *
728 svc_pool_next(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
730 return pool ? pool : &serv->sv_pools[(*state)++ % serv->sv_nrpools];
733 static struct svc_pool *
734 svc_pool_victim(struct svc_serv *serv, struct svc_pool *target_pool,
737 struct svc_pool *pool;
744 if (atomic_inc_not_zero(&pool->sp_nrthreads))
748 for (i = 0; i < serv->sv_nrpools; i++) {
749 pool = &serv->sv_pools[--(*state) % serv->sv_nrpools];
750 if (atomic_inc_not_zero(&pool->sp_nrthreads))
757 set_bit(SP_VICTIM_REMAINS, &pool->sp_flags);
758 set_bit(SP_NEED_VICTIM, &pool->sp_flags);
759 if (!atomic_dec_and_test(&pool->sp_nrthreads))
761 /* Nothing left in this pool any more */
762 clear_bit(SP_NEED_VICTIM, &pool->sp_flags);
763 clear_bit(SP_VICTIM_REMAINS, &pool->sp_flags);
768 svc_start_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
770 struct svc_rqst *rqstp;
771 struct task_struct *task;
772 struct svc_pool *chosen_pool;
773 unsigned int state = serv->sv_nrthreads-1;
778 chosen_pool = svc_pool_next(serv, pool, &state);
779 node = svc_pool_map_get_node(chosen_pool->sp_id);
781 rqstp = svc_prepare_thread(serv, chosen_pool, node);
783 return PTR_ERR(rqstp);
784 task = kthread_create_on_node(serv->sv_threadfn, rqstp,
785 node, "%s", serv->sv_name);
787 svc_exit_thread(rqstp);
788 return PTR_ERR(task);
791 rqstp->rq_task = task;
792 if (serv->sv_nrpools > 1)
793 svc_pool_map_set_cpumask(task, chosen_pool->sp_id);
795 svc_sock_update_bufs(serv);
796 wake_up_process(task);
797 } while (nrservs > 0);
803 svc_stop_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
805 unsigned int state = serv->sv_nrthreads-1;
806 struct svc_pool *victim;
809 victim = svc_pool_victim(serv, pool, &state);
812 svc_pool_wake_idle_thread(victim);
813 wait_on_bit(&victim->sp_flags, SP_VICTIM_REMAINS,
816 } while (nrservs < 0);
821 * svc_set_num_threads - adjust number of threads per RPC service
822 * @serv: RPC service to adjust
823 * @pool: Specific pool from which to choose threads, or NULL
824 * @nrservs: New number of threads for @serv (0 or less means kill all threads)
826 * Create or destroy threads to make the number of threads for @serv the
827 * given number. If @pool is non-NULL, change only threads in that pool;
828 * otherwise, round-robin between all pools for @serv. @serv's
829 * sv_nrthreads is adjusted for each thread created or destroyed.
831 * Caller must ensure mutual exclusion between this and server startup or
834 * Returns zero on success or a negative errno if an error occurred while
838 svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
841 nrservs -= serv->sv_nrthreads;
843 nrservs -= atomic_read(&pool->sp_nrthreads);
846 return svc_start_kthreads(serv, pool, nrservs);
848 return svc_stop_kthreads(serv, pool, nrservs);
851 EXPORT_SYMBOL_GPL(svc_set_num_threads);
854 * svc_rqst_replace_page - Replace one page in rq_pages[]
855 * @rqstp: svc_rqst with pages to replace
856 * @page: replacement page
858 * When replacing a page in rq_pages, batch the release of the
859 * replaced pages to avoid hammering the page allocator.
862 * %true: page replaced
863 * %false: array bounds checking failed
865 bool svc_rqst_replace_page(struct svc_rqst *rqstp, struct page *page)
867 struct page **begin = rqstp->rq_pages;
868 struct page **end = &rqstp->rq_pages[RPCSVC_MAXPAGES];
870 if (unlikely(rqstp->rq_next_page < begin || rqstp->rq_next_page > end)) {
871 trace_svc_replace_page_err(rqstp);
875 if (*rqstp->rq_next_page) {
876 if (!folio_batch_add(&rqstp->rq_fbatch,
877 page_folio(*rqstp->rq_next_page)))
878 __folio_batch_release(&rqstp->rq_fbatch);
882 *(rqstp->rq_next_page++) = page;
885 EXPORT_SYMBOL_GPL(svc_rqst_replace_page);
888 * svc_rqst_release_pages - Release Reply buffer pages
889 * @rqstp: RPC transaction context
891 * Release response pages that might still be in flight after
892 * svc_send, and any spliced filesystem-owned pages.
894 void svc_rqst_release_pages(struct svc_rqst *rqstp)
896 int i, count = rqstp->rq_next_page - rqstp->rq_respages;
899 release_pages(rqstp->rq_respages, count);
900 for (i = 0; i < count; i++)
901 rqstp->rq_respages[i] = NULL;
906 * Called from a server thread as it's exiting. Caller must hold the "service
907 * mutex" for the service.
910 svc_rqst_free(struct svc_rqst *rqstp)
912 folio_batch_release(&rqstp->rq_fbatch);
913 svc_release_buffer(rqstp);
914 if (rqstp->rq_scratch_page)
915 put_page(rqstp->rq_scratch_page);
916 kfree(rqstp->rq_resp);
917 kfree(rqstp->rq_argp);
918 kfree(rqstp->rq_auth_data);
919 kfree_rcu(rqstp, rq_rcu_head);
921 EXPORT_SYMBOL_GPL(svc_rqst_free);
924 svc_exit_thread(struct svc_rqst *rqstp)
926 struct svc_serv *serv = rqstp->rq_server;
927 struct svc_pool *pool = rqstp->rq_pool;
929 list_del_rcu(&rqstp->rq_all);
931 atomic_dec(&pool->sp_nrthreads);
933 spin_lock_bh(&serv->sv_lock);
934 serv->sv_nrthreads -= 1;
935 spin_unlock_bh(&serv->sv_lock);
936 svc_sock_update_bufs(serv);
938 svc_rqst_free(rqstp);
940 clear_and_wake_up_bit(SP_VICTIM_REMAINS, &pool->sp_flags);
942 EXPORT_SYMBOL_GPL(svc_exit_thread);
945 * Register an "inet" protocol family netid with the local
946 * rpcbind daemon via an rpcbind v4 SET request.
948 * No netconfig infrastructure is available in the kernel, so
949 * we map IP_ protocol numbers to netids by hand.
951 * Returns zero on success; a negative errno value is returned
952 * if any error occurs.
954 static int __svc_rpcb_register4(struct net *net, const u32 program,
956 const unsigned short protocol,
957 const unsigned short port)
959 const struct sockaddr_in sin = {
960 .sin_family = AF_INET,
961 .sin_addr.s_addr = htonl(INADDR_ANY),
962 .sin_port = htons(port),
969 netid = RPCBIND_NETID_UDP;
972 netid = RPCBIND_NETID_TCP;
978 error = rpcb_v4_register(net, program, version,
979 (const struct sockaddr *)&sin, netid);
982 * User space didn't support rpcbind v4, so retry this
983 * registration request with the legacy rpcbind v2 protocol.
985 if (error == -EPROTONOSUPPORT)
986 error = rpcb_register(net, program, version, protocol, port);
991 #if IS_ENABLED(CONFIG_IPV6)
993 * Register an "inet6" protocol family netid with the local
994 * rpcbind daemon via an rpcbind v4 SET request.
996 * No netconfig infrastructure is available in the kernel, so
997 * we map IP_ protocol numbers to netids by hand.
999 * Returns zero on success; a negative errno value is returned
1000 * if any error occurs.
1002 static int __svc_rpcb_register6(struct net *net, const u32 program,
1004 const unsigned short protocol,
1005 const unsigned short port)
1007 const struct sockaddr_in6 sin6 = {
1008 .sin6_family = AF_INET6,
1009 .sin6_addr = IN6ADDR_ANY_INIT,
1010 .sin6_port = htons(port),
1017 netid = RPCBIND_NETID_UDP6;
1020 netid = RPCBIND_NETID_TCP6;
1023 return -ENOPROTOOPT;
1026 error = rpcb_v4_register(net, program, version,
1027 (const struct sockaddr *)&sin6, netid);
1030 * User space didn't support rpcbind version 4, so we won't
1031 * use a PF_INET6 listener.
1033 if (error == -EPROTONOSUPPORT)
1034 error = -EAFNOSUPPORT;
1038 #endif /* IS_ENABLED(CONFIG_IPV6) */
1041 * Register a kernel RPC service via rpcbind version 4.
1043 * Returns zero on success; a negative errno value is returned
1044 * if any error occurs.
1046 static int __svc_register(struct net *net, const char *progname,
1047 const u32 program, const u32 version,
1049 const unsigned short protocol,
1050 const unsigned short port)
1052 int error = -EAFNOSUPPORT;
1056 error = __svc_rpcb_register4(net, program, version,
1059 #if IS_ENABLED(CONFIG_IPV6)
1061 error = __svc_rpcb_register6(net, program, version,
1066 trace_svc_register(progname, version, family, protocol, port, error);
1070 int svc_rpcbind_set_version(struct net *net,
1071 const struct svc_program *progp,
1072 u32 version, int family,
1073 unsigned short proto,
1074 unsigned short port)
1076 return __svc_register(net, progp->pg_name, progp->pg_prog,
1077 version, family, proto, port);
1080 EXPORT_SYMBOL_GPL(svc_rpcbind_set_version);
1082 int svc_generic_rpcbind_set(struct net *net,
1083 const struct svc_program *progp,
1084 u32 version, int family,
1085 unsigned short proto,
1086 unsigned short port)
1088 const struct svc_version *vers = progp->pg_vers[version];
1094 if (vers->vs_hidden) {
1095 trace_svc_noregister(progp->pg_name, version, proto,
1101 * Don't register a UDP port if we need congestion
1104 if (vers->vs_need_cong_ctrl && proto == IPPROTO_UDP)
1107 error = svc_rpcbind_set_version(net, progp, version,
1108 family, proto, port);
1110 return (vers->vs_rpcb_optnl) ? 0 : error;
1112 EXPORT_SYMBOL_GPL(svc_generic_rpcbind_set);
1115 * svc_register - register an RPC service with the local portmapper
1116 * @serv: svc_serv struct for the service to register
1117 * @net: net namespace for the service to register
1118 * @family: protocol family of service's listener socket
1119 * @proto: transport protocol number to advertise
1120 * @port: port to advertise
1122 * Service is registered for any address in the passed-in protocol family
1124 int svc_register(const struct svc_serv *serv, struct net *net,
1125 const int family, const unsigned short proto,
1126 const unsigned short port)
1128 struct svc_program *progp;
1132 WARN_ON_ONCE(proto == 0 && port == 0);
1133 if (proto == 0 && port == 0)
1136 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1137 for (i = 0; i < progp->pg_nvers; i++) {
1139 error = progp->pg_rpcbind_set(net, progp, i,
1140 family, proto, port);
1142 printk(KERN_WARNING "svc: failed to register "
1143 "%sv%u RPC service (errno %d).\n",
1144 progp->pg_name, i, -error);
1154 * If user space is running rpcbind, it should take the v4 UNSET
1155 * and clear everything for this [program, version]. If user space
1156 * is running portmap, it will reject the v4 UNSET, but won't have
1157 * any "inet6" entries anyway. So a PMAP_UNSET should be sufficient
1158 * in this case to clear all existing entries for [program, version].
1160 static void __svc_unregister(struct net *net, const u32 program, const u32 version,
1161 const char *progname)
1165 error = rpcb_v4_register(net, program, version, NULL, "");
1168 * User space didn't support rpcbind v4, so retry this
1169 * request with the legacy rpcbind v2 protocol.
1171 if (error == -EPROTONOSUPPORT)
1172 error = rpcb_register(net, program, version, 0, 0);
1174 trace_svc_unregister(progname, version, error);
1178 * All netids, bind addresses and ports registered for [program, version]
1179 * are removed from the local rpcbind database (if the service is not
1180 * hidden) to make way for a new instance of the service.
1182 * The result of unregistration is reported via dprintk for those who want
1183 * verification of the result, but is otherwise not important.
1185 static void svc_unregister(const struct svc_serv *serv, struct net *net)
1187 struct sighand_struct *sighand;
1188 struct svc_program *progp;
1189 unsigned long flags;
1192 clear_thread_flag(TIF_SIGPENDING);
1194 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1195 for (i = 0; i < progp->pg_nvers; i++) {
1196 if (progp->pg_vers[i] == NULL)
1198 if (progp->pg_vers[i]->vs_hidden)
1200 __svc_unregister(net, progp->pg_prog, i, progp->pg_name);
1205 sighand = rcu_dereference(current->sighand);
1206 spin_lock_irqsave(&sighand->siglock, flags);
1207 recalc_sigpending();
1208 spin_unlock_irqrestore(&sighand->siglock, flags);
1213 * dprintk the given error with the address of the client that caused it.
1215 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
1216 static __printf(2, 3)
1217 void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...)
1219 struct va_format vaf;
1221 char buf[RPC_MAX_ADDRBUFLEN];
1223 va_start(args, fmt);
1228 dprintk("svc: %s: %pV", svc_print_addr(rqstp, buf, sizeof(buf)), &vaf);
1233 static __printf(2,3) void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...) {}
1237 svc_generic_init_request(struct svc_rqst *rqstp,
1238 const struct svc_program *progp,
1239 struct svc_process_info *ret)
1241 const struct svc_version *versp = NULL; /* compiler food */
1242 const struct svc_procedure *procp = NULL;
1244 if (rqstp->rq_vers >= progp->pg_nvers )
1246 versp = progp->pg_vers[rqstp->rq_vers];
1251 * Some protocol versions (namely NFSv4) require some form of
1252 * congestion control. (See RFC 7530 section 3.1 paragraph 2)
1253 * In other words, UDP is not allowed. We mark those when setting
1254 * up the svc_xprt, and verify that here.
1256 * The spec is not very clear about what error should be returned
1257 * when someone tries to access a server that is listening on UDP
1258 * for lower versions. RPC_PROG_MISMATCH seems to be the closest
1261 if (versp->vs_need_cong_ctrl && rqstp->rq_xprt &&
1262 !test_bit(XPT_CONG_CTRL, &rqstp->rq_xprt->xpt_flags))
1265 if (rqstp->rq_proc >= versp->vs_nproc)
1267 rqstp->rq_procinfo = procp = &versp->vs_proc[rqstp->rq_proc];
1271 /* Initialize storage for argp and resp */
1272 memset(rqstp->rq_argp, 0, procp->pc_argzero);
1273 memset(rqstp->rq_resp, 0, procp->pc_ressize);
1275 /* Bump per-procedure stats counter */
1276 this_cpu_inc(versp->vs_count[rqstp->rq_proc]);
1278 ret->dispatch = versp->vs_dispatch;
1281 ret->mismatch.lovers = progp->pg_lovers;
1282 ret->mismatch.hivers = progp->pg_hivers;
1283 return rpc_prog_mismatch;
1285 return rpc_proc_unavail;
1287 EXPORT_SYMBOL_GPL(svc_generic_init_request);
1290 * Common routine for processing the RPC request.
1293 svc_process_common(struct svc_rqst *rqstp)
1295 struct xdr_stream *xdr = &rqstp->rq_res_stream;
1296 struct svc_program *progp;
1297 const struct svc_procedure *procp = NULL;
1298 struct svc_serv *serv = rqstp->rq_server;
1299 struct svc_process_info process;
1300 enum svc_auth_status auth_res;
1301 unsigned int aoffset;
1305 /* Will be turned off only when NFSv4 Sessions are used */
1306 set_bit(RQ_USEDEFERRAL, &rqstp->rq_flags);
1307 clear_bit(RQ_DROPME, &rqstp->rq_flags);
1309 /* Construct the first words of the reply: */
1310 svcxdr_init_encode(rqstp);
1311 xdr_stream_encode_be32(xdr, rqstp->rq_xid);
1312 xdr_stream_encode_be32(xdr, rpc_reply);
1314 p = xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 4);
1317 if (*p++ != cpu_to_be32(RPC_VERSION))
1320 xdr_stream_encode_be32(xdr, rpc_msg_accepted);
1322 rqstp->rq_prog = be32_to_cpup(p++);
1323 rqstp->rq_vers = be32_to_cpup(p++);
1324 rqstp->rq_proc = be32_to_cpup(p);
1326 for (progp = serv->sv_program; progp; progp = progp->pg_next)
1327 if (rqstp->rq_prog == progp->pg_prog)
1331 * Decode auth data, and add verifier to reply buffer.
1332 * We do this before anything else in order to get a decent
1335 auth_res = svc_authenticate(rqstp);
1336 /* Also give the program a chance to reject this call: */
1337 if (auth_res == SVC_OK && progp)
1338 auth_res = progp->pg_authenticate(rqstp);
1339 trace_svc_authenticate(rqstp, auth_res);
1344 goto err_garbage_args;
1346 goto err_system_err;
1356 pr_warn_once("Unexpected svc_auth_status (%d)\n", auth_res);
1357 goto err_system_err;
1363 switch (progp->pg_init_request(rqstp, progp, &process)) {
1366 case rpc_prog_unavail:
1368 case rpc_prog_mismatch:
1370 case rpc_proc_unavail:
1374 procp = rqstp->rq_procinfo;
1375 /* Should this check go into the dispatcher? */
1376 if (!procp || !procp->pc_func)
1379 /* Syntactic check complete */
1381 serv->sv_stats->rpccnt++;
1382 trace_svc_process(rqstp, progp->pg_name);
1384 aoffset = xdr_stream_pos(xdr);
1386 /* un-reserve some of the out-queue now that we have a
1387 * better idea of reply size
1389 if (procp->pc_xdrressize)
1390 svc_reserve_auth(rqstp, procp->pc_xdrressize<<2);
1392 /* Call the function that processes the request. */
1393 rc = process.dispatch(rqstp);
1394 if (procp->pc_release)
1395 procp->pc_release(rqstp);
1396 xdr_finish_decode(xdr);
1400 if (rqstp->rq_auth_stat != rpc_auth_ok)
1403 if (*rqstp->rq_accept_statp != rpc_success)
1404 xdr_truncate_encode(xdr, aoffset);
1406 if (procp->pc_encode == NULL)
1410 if (svc_authorise(rqstp))
1412 return 1; /* Caller can now send it */
1415 svc_authorise(rqstp); /* doesn't hurt to call this twice */
1416 dprintk("svc: svc_process dropit\n");
1420 svc_authorise(rqstp);
1422 if (rqstp->rq_xprt && test_bit(XPT_TEMP, &rqstp->rq_xprt->xpt_flags))
1423 svc_xprt_close(rqstp->rq_xprt);
1424 dprintk("svc: svc_process close\n");
1428 svc_printk(rqstp, "short len %u, dropping request\n",
1434 serv->sv_stats->rpcbadfmt++;
1435 xdr_stream_encode_u32(xdr, RPC_MSG_DENIED);
1436 xdr_stream_encode_u32(xdr, RPC_MISMATCH);
1437 /* Only RPCv2 supported */
1438 xdr_stream_encode_u32(xdr, RPC_VERSION);
1439 xdr_stream_encode_u32(xdr, RPC_VERSION);
1440 return 1; /* don't wrap */
1443 dprintk("svc: authentication failed (%d)\n",
1444 be32_to_cpu(rqstp->rq_auth_stat));
1446 serv->sv_stats->rpcbadauth++;
1447 /* Restore write pointer to location of reply status: */
1448 xdr_truncate_encode(xdr, XDR_UNIT * 2);
1449 xdr_stream_encode_u32(xdr, RPC_MSG_DENIED);
1450 xdr_stream_encode_u32(xdr, RPC_AUTH_ERROR);
1451 xdr_stream_encode_be32(xdr, rqstp->rq_auth_stat);
1455 dprintk("svc: unknown program %d\n", rqstp->rq_prog);
1457 serv->sv_stats->rpcbadfmt++;
1458 *rqstp->rq_accept_statp = rpc_prog_unavail;
1462 svc_printk(rqstp, "unknown version (%d for prog %d, %s)\n",
1463 rqstp->rq_vers, rqstp->rq_prog, progp->pg_name);
1466 serv->sv_stats->rpcbadfmt++;
1467 *rqstp->rq_accept_statp = rpc_prog_mismatch;
1470 * svc_authenticate() has already added the verifier and
1471 * advanced the stream just past rq_accept_statp.
1473 xdr_stream_encode_u32(xdr, process.mismatch.lovers);
1474 xdr_stream_encode_u32(xdr, process.mismatch.hivers);
1478 svc_printk(rqstp, "unknown procedure (%d)\n", rqstp->rq_proc);
1481 serv->sv_stats->rpcbadfmt++;
1482 *rqstp->rq_accept_statp = rpc_proc_unavail;
1486 svc_printk(rqstp, "failed to decode RPC header\n");
1489 serv->sv_stats->rpcbadfmt++;
1490 *rqstp->rq_accept_statp = rpc_garbage_args;
1495 serv->sv_stats->rpcbadfmt++;
1496 *rqstp->rq_accept_statp = rpc_system_err;
1501 * svc_process - Execute one RPC transaction
1502 * @rqstp: RPC transaction context
1505 void svc_process(struct svc_rqst *rqstp)
1507 struct kvec *resv = &rqstp->rq_res.head[0];
1510 #if IS_ENABLED(CONFIG_FAIL_SUNRPC)
1511 if (!fail_sunrpc.ignore_server_disconnect &&
1512 should_fail(&fail_sunrpc.attr, 1))
1513 svc_xprt_deferred_close(rqstp->rq_xprt);
1517 * Setup response xdr_buf.
1518 * Initially it has just one page
1520 rqstp->rq_next_page = &rqstp->rq_respages[1];
1521 resv->iov_base = page_address(rqstp->rq_respages[0]);
1523 rqstp->rq_res.pages = rqstp->rq_next_page;
1524 rqstp->rq_res.len = 0;
1525 rqstp->rq_res.page_base = 0;
1526 rqstp->rq_res.page_len = 0;
1527 rqstp->rq_res.buflen = PAGE_SIZE;
1528 rqstp->rq_res.tail[0].iov_base = NULL;
1529 rqstp->rq_res.tail[0].iov_len = 0;
1531 svcxdr_init_decode(rqstp);
1532 p = xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 2);
1535 rqstp->rq_xid = *p++;
1536 if (unlikely(*p != rpc_call))
1539 if (!svc_process_common(rqstp))
1545 svc_printk(rqstp, "bad direction 0x%08x, dropping request\n",
1547 if (rqstp->rq_server->sv_stats)
1548 rqstp->rq_server->sv_stats->rpcbadfmt++;
1553 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1555 * svc_process_bc - process a reverse-direction RPC request
1556 * @req: RPC request to be used for client-side processing
1557 * @rqstp: server-side execution context
1560 void svc_process_bc(struct rpc_rqst *req, struct svc_rqst *rqstp)
1562 struct rpc_task *task;
1564 struct rpc_timeout timeout;
1566 /* Build the svc_rqst used by the common processing routine */
1567 rqstp->rq_xid = req->rq_xid;
1568 rqstp->rq_prot = req->rq_xprt->prot;
1569 rqstp->rq_bc_net = req->rq_xprt->xprt_net;
1571 rqstp->rq_addrlen = sizeof(req->rq_xprt->addr);
1572 memcpy(&rqstp->rq_addr, &req->rq_xprt->addr, rqstp->rq_addrlen);
1573 memcpy(&rqstp->rq_arg, &req->rq_rcv_buf, sizeof(rqstp->rq_arg));
1574 memcpy(&rqstp->rq_res, &req->rq_snd_buf, sizeof(rqstp->rq_res));
1576 /* Adjust the argument buffer length */
1577 rqstp->rq_arg.len = req->rq_private_buf.len;
1578 if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len) {
1579 rqstp->rq_arg.head[0].iov_len = rqstp->rq_arg.len;
1580 rqstp->rq_arg.page_len = 0;
1581 } else if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len +
1582 rqstp->rq_arg.page_len)
1583 rqstp->rq_arg.page_len = rqstp->rq_arg.len -
1584 rqstp->rq_arg.head[0].iov_len;
1586 rqstp->rq_arg.len = rqstp->rq_arg.head[0].iov_len +
1587 rqstp->rq_arg.page_len;
1589 /* Reset the response buffer */
1590 rqstp->rq_res.head[0].iov_len = 0;
1593 * Skip the XID and calldir fields because they've already
1594 * been processed by the caller.
1596 svcxdr_init_decode(rqstp);
1597 if (!xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 2))
1600 /* Parse and execute the bc call */
1601 proc_error = svc_process_common(rqstp);
1603 atomic_dec(&req->rq_xprt->bc_slot_count);
1605 /* Processing error: drop the request */
1606 xprt_free_bc_request(req);
1609 /* Finally, send the reply synchronously */
1610 if (rqstp->bc_to_initval > 0) {
1611 timeout.to_initval = rqstp->bc_to_initval;
1612 timeout.to_retries = rqstp->bc_to_retries;
1614 timeout.to_initval = req->rq_xprt->timeout->to_initval;
1615 timeout.to_retries = req->rq_xprt->timeout->to_retries;
1617 memcpy(&req->rq_snd_buf, &rqstp->rq_res, sizeof(req->rq_snd_buf));
1618 task = rpc_run_bc_task(req, &timeout);
1623 WARN_ON_ONCE(atomic_read(&task->tk_count) != 1);
1626 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1629 * svc_max_payload - Return transport-specific limit on the RPC payload
1630 * @rqstp: RPC transaction context
1632 * Returns the maximum number of payload bytes the current transport
1635 u32 svc_max_payload(const struct svc_rqst *rqstp)
1637 u32 max = rqstp->rq_xprt->xpt_class->xcl_max_payload;
1639 if (rqstp->rq_server->sv_max_payload < max)
1640 max = rqstp->rq_server->sv_max_payload;
1643 EXPORT_SYMBOL_GPL(svc_max_payload);
1646 * svc_proc_name - Return RPC procedure name in string form
1647 * @rqstp: svc_rqst to operate on
1650 * Pointer to a NUL-terminated string
1652 const char *svc_proc_name(const struct svc_rqst *rqstp)
1654 if (rqstp && rqstp->rq_procinfo)
1655 return rqstp->rq_procinfo->pc_name;
1661 * svc_encode_result_payload - mark a range of bytes as a result payload
1662 * @rqstp: svc_rqst to operate on
1663 * @offset: payload's byte offset in rqstp->rq_res
1664 * @length: size of payload, in bytes
1666 * Returns zero on success, or a negative errno if a permanent
1669 int svc_encode_result_payload(struct svc_rqst *rqstp, unsigned int offset,
1670 unsigned int length)
1672 return rqstp->rq_xprt->xpt_ops->xpo_result_payload(rqstp, offset,
1675 EXPORT_SYMBOL_GPL(svc_encode_result_payload);
1678 * svc_fill_write_vector - Construct data argument for VFS write call
1679 * @rqstp: svc_rqst to operate on
1680 * @payload: xdr_buf containing only the write data payload
1682 * Fills in rqstp::rq_vec, and returns the number of elements.
1684 unsigned int svc_fill_write_vector(struct svc_rqst *rqstp,
1685 struct xdr_buf *payload)
1687 struct page **pages = payload->pages;
1688 struct kvec *first = payload->head;
1689 struct kvec *vec = rqstp->rq_vec;
1690 size_t total = payload->len;
1693 /* Some types of transport can present the write payload
1694 * entirely in rq_arg.pages. In this case, @first is empty.
1697 if (first->iov_len) {
1698 vec[i].iov_base = first->iov_base;
1699 vec[i].iov_len = min_t(size_t, total, first->iov_len);
1700 total -= vec[i].iov_len;
1705 vec[i].iov_base = page_address(*pages);
1706 vec[i].iov_len = min_t(size_t, total, PAGE_SIZE);
1707 total -= vec[i].iov_len;
1712 WARN_ON_ONCE(i > ARRAY_SIZE(rqstp->rq_vec));
1715 EXPORT_SYMBOL_GPL(svc_fill_write_vector);
1718 * svc_fill_symlink_pathname - Construct pathname argument for VFS symlink call
1719 * @rqstp: svc_rqst to operate on
1720 * @first: buffer containing first section of pathname
1721 * @p: buffer containing remaining section of pathname
1722 * @total: total length of the pathname argument
1724 * The VFS symlink API demands a NUL-terminated pathname in mapped memory.
1725 * Returns pointer to a NUL-terminated string, or an ERR_PTR. Caller must free
1726 * the returned string.
1728 char *svc_fill_symlink_pathname(struct svc_rqst *rqstp, struct kvec *first,
1729 void *p, size_t total)
1731 size_t len, remaining;
1734 result = kmalloc(total + 1, GFP_KERNEL);
1736 return ERR_PTR(-ESERVERFAULT);
1741 len = min_t(size_t, total, first->iov_len);
1743 memcpy(dst, first->iov_base, len);
1749 len = min_t(size_t, remaining, PAGE_SIZE);
1750 memcpy(dst, p, len);
1756 /* Sanity check: Linux doesn't allow the pathname argument to
1757 * contain a NUL byte.
1759 if (strlen(result) != total) {
1761 return ERR_PTR(-EINVAL);
1765 EXPORT_SYMBOL_GPL(svc_fill_symlink_pathname);