1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/net/sunrpc/svc.c
5 * High-level RPC service routines
9 * Multiple threads pools and NUMAisation
10 * Copyright (c) 2006 Silicon Graphics, Inc.
14 #include <linux/linkage.h>
15 #include <linux/sched/signal.h>
16 #include <linux/errno.h>
17 #include <linux/net.h>
20 #include <linux/interrupt.h>
21 #include <linux/module.h>
22 #include <linux/kthread.h>
23 #include <linux/slab.h>
25 #include <linux/sunrpc/types.h>
26 #include <linux/sunrpc/xdr.h>
27 #include <linux/sunrpc/stats.h>
28 #include <linux/sunrpc/svcsock.h>
29 #include <linux/sunrpc/clnt.h>
30 #include <linux/sunrpc/bc_xprt.h>
32 #include <trace/events/sunrpc.h>
36 #define RPCDBG_FACILITY RPCDBG_SVCDSP
38 static void svc_unregister(const struct svc_serv *serv, struct net *net);
40 #define SVC_POOL_DEFAULT SVC_POOL_GLOBAL
43 * Mode for mapping cpus to pools.
46 SVC_POOL_AUTO = -1, /* choose one of the others */
47 SVC_POOL_GLOBAL, /* no mapping, just a single global pool
48 * (legacy & UP mode) */
49 SVC_POOL_PERCPU, /* one pool per cpu */
50 SVC_POOL_PERNODE /* one pool per numa node */
54 * Structure for mapping cpus to pools and vice versa.
55 * Setup once during sunrpc initialisation.
59 int count; /* How many svc_servs use us */
60 int mode; /* Note: int not enum to avoid
61 * warnings about "enumeration value
62 * not handled in switch" */
64 unsigned int *pool_to; /* maps pool id to cpu or node */
65 unsigned int *to_pool; /* maps cpu or node to pool id */
68 static struct svc_pool_map svc_pool_map = {
69 .mode = SVC_POOL_DEFAULT
72 static DEFINE_MUTEX(svc_pool_map_mutex);/* protects svc_pool_map.count only */
75 param_set_pool_mode(const char *val, const struct kernel_param *kp)
77 int *ip = (int *)kp->arg;
78 struct svc_pool_map *m = &svc_pool_map;
81 mutex_lock(&svc_pool_map_mutex);
88 if (!strncmp(val, "auto", 4))
90 else if (!strncmp(val, "global", 6))
91 *ip = SVC_POOL_GLOBAL;
92 else if (!strncmp(val, "percpu", 6))
93 *ip = SVC_POOL_PERCPU;
94 else if (!strncmp(val, "pernode", 7))
95 *ip = SVC_POOL_PERNODE;
100 mutex_unlock(&svc_pool_map_mutex);
105 param_get_pool_mode(char *buf, const struct kernel_param *kp)
107 int *ip = (int *)kp->arg;
112 return 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, unsigned int bufsize, int npools,
455 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 kref_init(&serv->sv_refcnt);
467 serv->sv_stats = prog->pg_stats;
468 if (bufsize > RPCSVC_MAXPAYLOAD)
469 bufsize = RPCSVC_MAXPAYLOAD;
470 serv->sv_max_payload = bufsize? bufsize : 4096;
471 serv->sv_max_mesg = roundup(serv->sv_max_payload + PAGE_SIZE, PAGE_SIZE);
472 serv->sv_threadfn = threadfn;
475 prog->pg_lovers = prog->pg_nvers-1;
476 for (vers=0; vers<prog->pg_nvers ; vers++)
477 if (prog->pg_vers[vers]) {
478 prog->pg_hivers = vers;
479 if (prog->pg_lovers > vers)
480 prog->pg_lovers = vers;
481 if (prog->pg_vers[vers]->vs_xdrsize > xdrsize)
482 xdrsize = prog->pg_vers[vers]->vs_xdrsize;
484 prog = prog->pg_next;
486 serv->sv_xdrsize = xdrsize;
487 INIT_LIST_HEAD(&serv->sv_tempsocks);
488 INIT_LIST_HEAD(&serv->sv_permsocks);
489 timer_setup(&serv->sv_temptimer, NULL, 0);
490 spin_lock_init(&serv->sv_lock);
494 serv->sv_nrpools = npools;
496 kcalloc(serv->sv_nrpools, sizeof(struct svc_pool),
498 if (!serv->sv_pools) {
503 for (i = 0; i < serv->sv_nrpools; i++) {
504 struct svc_pool *pool = &serv->sv_pools[i];
506 dprintk("svc: initialising pool %u for %s\n",
510 lwq_init(&pool->sp_xprts);
511 INIT_LIST_HEAD(&pool->sp_all_threads);
512 init_llist_head(&pool->sp_idle_threads);
514 percpu_counter_init(&pool->sp_messages_arrived, 0, GFP_KERNEL);
515 percpu_counter_init(&pool->sp_sockets_queued, 0, GFP_KERNEL);
516 percpu_counter_init(&pool->sp_threads_woken, 0, GFP_KERNEL);
523 * svc_create - Create an RPC service
524 * @prog: the RPC program the new service will handle
525 * @bufsize: maximum message size for @prog
526 * @threadfn: a function to service RPC requests for @prog
528 * Returns an instantiated struct svc_serv object or NULL.
530 struct svc_serv *svc_create(struct svc_program *prog, unsigned int bufsize,
531 int (*threadfn)(void *data))
533 return __svc_create(prog, bufsize, 1, threadfn);
535 EXPORT_SYMBOL_GPL(svc_create);
538 * svc_create_pooled - Create an RPC service with pooled threads
539 * @prog: the RPC program the new service will handle
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 unsigned int bufsize,
547 int (*threadfn)(void *data))
549 struct svc_serv *serv;
550 unsigned int npools = svc_pool_map_get();
552 serv = __svc_create(prog, bufsize, npools, threadfn);
557 svc_pool_map_put(npools);
560 EXPORT_SYMBOL_GPL(svc_create_pooled);
563 * Destroy an RPC service. Should be called with appropriate locking to
564 * protect sv_permsocks and sv_tempsocks.
567 svc_destroy(struct kref *ref)
569 struct svc_serv *serv = container_of(ref, struct svc_serv, sv_refcnt);
572 dprintk("svc: svc_destroy(%s)\n", serv->sv_program->pg_name);
573 timer_shutdown_sync(&serv->sv_temptimer);
576 * Remaining transports at this point are not expected.
578 WARN_ONCE(!list_empty(&serv->sv_permsocks),
579 "SVC: permsocks remain for %s\n", serv->sv_program->pg_name);
580 WARN_ONCE(!list_empty(&serv->sv_tempsocks),
581 "SVC: tempsocks remain for %s\n", serv->sv_program->pg_name);
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_messages_arrived);
591 percpu_counter_destroy(&pool->sp_sockets_queued);
592 percpu_counter_destroy(&pool->sp_threads_woken);
594 kfree(serv->sv_pools);
597 EXPORT_SYMBOL_GPL(svc_destroy);
600 svc_init_buffer(struct svc_rqst *rqstp, unsigned int size, int node)
602 unsigned long pages, ret;
604 /* bc_xprt uses fore channel allocated buffers */
605 if (svc_is_backchannel(rqstp))
608 pages = size / PAGE_SIZE + 1; /* extra page as we hold both request and reply.
609 * We assume one is at most one page
611 WARN_ON_ONCE(pages > RPCSVC_MAXPAGES);
612 if (pages > RPCSVC_MAXPAGES)
613 pages = RPCSVC_MAXPAGES;
615 ret = alloc_pages_bulk_array_node(GFP_KERNEL, node, pages,
621 * Release an RPC server buffer
624 svc_release_buffer(struct svc_rqst *rqstp)
628 for (i = 0; i < ARRAY_SIZE(rqstp->rq_pages); i++)
629 if (rqstp->rq_pages[i])
630 put_page(rqstp->rq_pages[i]);
634 svc_rqst_alloc(struct svc_serv *serv, struct svc_pool *pool, int node)
636 struct svc_rqst *rqstp;
638 rqstp = kzalloc_node(sizeof(*rqstp), GFP_KERNEL, node);
642 folio_batch_init(&rqstp->rq_fbatch);
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 atomic_inc(&pool->sp_nrthreads);
685 /* Protected by whatever lock the service uses when calling
686 * svc_set_num_threads()
688 list_add_rcu(&rqstp->rq_all, &pool->sp_all_threads);
694 * svc_pool_wake_idle_thread - Awaken an idle thread in @pool
695 * @pool: service thread pool
697 * Can be called from soft IRQ or process context. Finding an idle
698 * service thread and marking it BUSY is atomic with respect to
699 * other calls to svc_pool_wake_idle_thread().
702 void svc_pool_wake_idle_thread(struct svc_pool *pool)
704 struct svc_rqst *rqstp;
705 struct llist_node *ln;
708 ln = READ_ONCE(pool->sp_idle_threads.first);
710 rqstp = llist_entry(ln, struct svc_rqst, rq_idle);
711 WRITE_ONCE(rqstp->rq_qtime, ktime_get());
712 if (!task_is_running(rqstp->rq_task)) {
713 wake_up_process(rqstp->rq_task);
714 trace_svc_wake_up(rqstp->rq_task->pid);
715 percpu_counter_inc(&pool->sp_threads_woken);
723 EXPORT_SYMBOL_GPL(svc_pool_wake_idle_thread);
725 static struct svc_pool *
726 svc_pool_next(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
728 return pool ? pool : &serv->sv_pools[(*state)++ % serv->sv_nrpools];
731 static struct svc_pool *
732 svc_pool_victim(struct svc_serv *serv, struct svc_pool *target_pool,
735 struct svc_pool *pool;
742 if (atomic_inc_not_zero(&pool->sp_nrthreads))
746 for (i = 0; i < serv->sv_nrpools; i++) {
747 pool = &serv->sv_pools[--(*state) % serv->sv_nrpools];
748 if (atomic_inc_not_zero(&pool->sp_nrthreads))
755 set_bit(SP_VICTIM_REMAINS, &pool->sp_flags);
756 set_bit(SP_NEED_VICTIM, &pool->sp_flags);
757 if (!atomic_dec_and_test(&pool->sp_nrthreads))
759 /* Nothing left in this pool any more */
760 clear_bit(SP_NEED_VICTIM, &pool->sp_flags);
761 clear_bit(SP_VICTIM_REMAINS, &pool->sp_flags);
766 svc_start_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
768 struct svc_rqst *rqstp;
769 struct task_struct *task;
770 struct svc_pool *chosen_pool;
771 unsigned int state = serv->sv_nrthreads-1;
776 chosen_pool = svc_pool_next(serv, pool, &state);
777 node = svc_pool_map_get_node(chosen_pool->sp_id);
779 rqstp = svc_prepare_thread(serv, chosen_pool, node);
781 return PTR_ERR(rqstp);
782 task = kthread_create_on_node(serv->sv_threadfn, rqstp,
783 node, "%s", serv->sv_name);
785 svc_exit_thread(rqstp);
786 return PTR_ERR(task);
789 rqstp->rq_task = task;
790 if (serv->sv_nrpools > 1)
791 svc_pool_map_set_cpumask(task, chosen_pool->sp_id);
793 svc_sock_update_bufs(serv);
794 wake_up_process(task);
795 } while (nrservs > 0);
801 svc_stop_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
803 unsigned int state = serv->sv_nrthreads-1;
804 struct svc_pool *victim;
807 victim = svc_pool_victim(serv, pool, &state);
810 svc_pool_wake_idle_thread(victim);
811 wait_on_bit(&victim->sp_flags, SP_VICTIM_REMAINS,
814 } while (nrservs < 0);
819 * svc_set_num_threads - adjust number of threads per RPC service
820 * @serv: RPC service to adjust
821 * @pool: Specific pool from which to choose threads, or NULL
822 * @nrservs: New number of threads for @serv (0 or less means kill all threads)
824 * Create or destroy threads to make the number of threads for @serv the
825 * given number. If @pool is non-NULL, change only threads in that pool;
826 * otherwise, round-robin between all pools for @serv. @serv's
827 * sv_nrthreads is adjusted for each thread created or destroyed.
829 * Caller must ensure mutual exclusion between this and server startup or
832 * Returns zero on success or a negative errno if an error occurred while
836 svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
839 nrservs -= serv->sv_nrthreads;
841 nrservs -= atomic_read(&pool->sp_nrthreads);
844 return svc_start_kthreads(serv, pool, nrservs);
846 return svc_stop_kthreads(serv, pool, nrservs);
849 EXPORT_SYMBOL_GPL(svc_set_num_threads);
852 * svc_rqst_replace_page - Replace one page in rq_pages[]
853 * @rqstp: svc_rqst with pages to replace
854 * @page: replacement page
856 * When replacing a page in rq_pages, batch the release of the
857 * replaced pages to avoid hammering the page allocator.
860 * %true: page replaced
861 * %false: array bounds checking failed
863 bool svc_rqst_replace_page(struct svc_rqst *rqstp, struct page *page)
865 struct page **begin = rqstp->rq_pages;
866 struct page **end = &rqstp->rq_pages[RPCSVC_MAXPAGES];
868 if (unlikely(rqstp->rq_next_page < begin || rqstp->rq_next_page > end)) {
869 trace_svc_replace_page_err(rqstp);
873 if (*rqstp->rq_next_page) {
874 if (!folio_batch_add(&rqstp->rq_fbatch,
875 page_folio(*rqstp->rq_next_page)))
876 __folio_batch_release(&rqstp->rq_fbatch);
880 *(rqstp->rq_next_page++) = page;
883 EXPORT_SYMBOL_GPL(svc_rqst_replace_page);
886 * svc_rqst_release_pages - Release Reply buffer pages
887 * @rqstp: RPC transaction context
889 * Release response pages that might still be in flight after
890 * svc_send, and any spliced filesystem-owned pages.
892 void svc_rqst_release_pages(struct svc_rqst *rqstp)
894 int i, count = rqstp->rq_next_page - rqstp->rq_respages;
897 release_pages(rqstp->rq_respages, count);
898 for (i = 0; i < count; i++)
899 rqstp->rq_respages[i] = NULL;
904 * Called from a server thread as it's exiting. Caller must hold the "service
905 * mutex" for the service.
908 svc_rqst_free(struct svc_rqst *rqstp)
910 folio_batch_release(&rqstp->rq_fbatch);
911 svc_release_buffer(rqstp);
912 if (rqstp->rq_scratch_page)
913 put_page(rqstp->rq_scratch_page);
914 kfree(rqstp->rq_resp);
915 kfree(rqstp->rq_argp);
916 kfree(rqstp->rq_auth_data);
917 kfree_rcu(rqstp, rq_rcu_head);
919 EXPORT_SYMBOL_GPL(svc_rqst_free);
922 svc_exit_thread(struct svc_rqst *rqstp)
924 struct svc_serv *serv = rqstp->rq_server;
925 struct svc_pool *pool = rqstp->rq_pool;
927 list_del_rcu(&rqstp->rq_all);
929 atomic_dec(&pool->sp_nrthreads);
931 spin_lock_bh(&serv->sv_lock);
932 serv->sv_nrthreads -= 1;
933 spin_unlock_bh(&serv->sv_lock);
934 svc_sock_update_bufs(serv);
936 svc_rqst_free(rqstp);
939 /* That svc_put() cannot be the last, because the thread
940 * waiting for SP_VICTIM_REMAINS to clear must hold
941 * a reference. So it is still safe to access pool.
943 clear_and_wake_up_bit(SP_VICTIM_REMAINS, &pool->sp_flags);
945 EXPORT_SYMBOL_GPL(svc_exit_thread);
948 * Register an "inet" 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_register4(struct net *net, const u32 program,
959 const unsigned short protocol,
960 const unsigned short port)
962 const struct sockaddr_in sin = {
963 .sin_family = AF_INET,
964 .sin_addr.s_addr = htonl(INADDR_ANY),
965 .sin_port = htons(port),
972 netid = RPCBIND_NETID_UDP;
975 netid = RPCBIND_NETID_TCP;
981 error = rpcb_v4_register(net, program, version,
982 (const struct sockaddr *)&sin, netid);
985 * User space didn't support rpcbind v4, so retry this
986 * registration request with the legacy rpcbind v2 protocol.
988 if (error == -EPROTONOSUPPORT)
989 error = rpcb_register(net, program, version, protocol, port);
994 #if IS_ENABLED(CONFIG_IPV6)
996 * Register an "inet6" protocol family netid with the local
997 * rpcbind daemon via an rpcbind v4 SET request.
999 * No netconfig infrastructure is available in the kernel, so
1000 * we map IP_ protocol numbers to netids by hand.
1002 * Returns zero on success; a negative errno value is returned
1003 * if any error occurs.
1005 static int __svc_rpcb_register6(struct net *net, const u32 program,
1007 const unsigned short protocol,
1008 const unsigned short port)
1010 const struct sockaddr_in6 sin6 = {
1011 .sin6_family = AF_INET6,
1012 .sin6_addr = IN6ADDR_ANY_INIT,
1013 .sin6_port = htons(port),
1020 netid = RPCBIND_NETID_UDP6;
1023 netid = RPCBIND_NETID_TCP6;
1026 return -ENOPROTOOPT;
1029 error = rpcb_v4_register(net, program, version,
1030 (const struct sockaddr *)&sin6, netid);
1033 * User space didn't support rpcbind version 4, so we won't
1034 * use a PF_INET6 listener.
1036 if (error == -EPROTONOSUPPORT)
1037 error = -EAFNOSUPPORT;
1041 #endif /* IS_ENABLED(CONFIG_IPV6) */
1044 * Register a kernel RPC service via rpcbind version 4.
1046 * Returns zero on success; a negative errno value is returned
1047 * if any error occurs.
1049 static int __svc_register(struct net *net, const char *progname,
1050 const u32 program, const u32 version,
1052 const unsigned short protocol,
1053 const unsigned short port)
1055 int error = -EAFNOSUPPORT;
1059 error = __svc_rpcb_register4(net, program, version,
1062 #if IS_ENABLED(CONFIG_IPV6)
1064 error = __svc_rpcb_register6(net, program, version,
1069 trace_svc_register(progname, version, family, protocol, port, error);
1073 int svc_rpcbind_set_version(struct net *net,
1074 const struct svc_program *progp,
1075 u32 version, int family,
1076 unsigned short proto,
1077 unsigned short port)
1079 return __svc_register(net, progp->pg_name, progp->pg_prog,
1080 version, family, proto, port);
1083 EXPORT_SYMBOL_GPL(svc_rpcbind_set_version);
1085 int svc_generic_rpcbind_set(struct net *net,
1086 const struct svc_program *progp,
1087 u32 version, int family,
1088 unsigned short proto,
1089 unsigned short port)
1091 const struct svc_version *vers = progp->pg_vers[version];
1097 if (vers->vs_hidden) {
1098 trace_svc_noregister(progp->pg_name, version, proto,
1104 * Don't register a UDP port if we need congestion
1107 if (vers->vs_need_cong_ctrl && proto == IPPROTO_UDP)
1110 error = svc_rpcbind_set_version(net, progp, version,
1111 family, proto, port);
1113 return (vers->vs_rpcb_optnl) ? 0 : error;
1115 EXPORT_SYMBOL_GPL(svc_generic_rpcbind_set);
1118 * svc_register - register an RPC service with the local portmapper
1119 * @serv: svc_serv struct for the service to register
1120 * @net: net namespace for the service to register
1121 * @family: protocol family of service's listener socket
1122 * @proto: transport protocol number to advertise
1123 * @port: port to advertise
1125 * Service is registered for any address in the passed-in protocol family
1127 int svc_register(const struct svc_serv *serv, struct net *net,
1128 const int family, const unsigned short proto,
1129 const unsigned short port)
1131 struct svc_program *progp;
1135 WARN_ON_ONCE(proto == 0 && port == 0);
1136 if (proto == 0 && port == 0)
1139 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1140 for (i = 0; i < progp->pg_nvers; i++) {
1142 error = progp->pg_rpcbind_set(net, progp, i,
1143 family, proto, port);
1145 printk(KERN_WARNING "svc: failed to register "
1146 "%sv%u RPC service (errno %d).\n",
1147 progp->pg_name, i, -error);
1157 * If user space is running rpcbind, it should take the v4 UNSET
1158 * and clear everything for this [program, version]. If user space
1159 * is running portmap, it will reject the v4 UNSET, but won't have
1160 * any "inet6" entries anyway. So a PMAP_UNSET should be sufficient
1161 * in this case to clear all existing entries for [program, version].
1163 static void __svc_unregister(struct net *net, const u32 program, const u32 version,
1164 const char *progname)
1168 error = rpcb_v4_register(net, program, version, NULL, "");
1171 * User space didn't support rpcbind v4, so retry this
1172 * request with the legacy rpcbind v2 protocol.
1174 if (error == -EPROTONOSUPPORT)
1175 error = rpcb_register(net, program, version, 0, 0);
1177 trace_svc_unregister(progname, version, error);
1181 * All netids, bind addresses and ports registered for [program, version]
1182 * are removed from the local rpcbind database (if the service is not
1183 * hidden) to make way for a new instance of the service.
1185 * The result of unregistration is reported via dprintk for those who want
1186 * verification of the result, but is otherwise not important.
1188 static void svc_unregister(const struct svc_serv *serv, struct net *net)
1190 struct sighand_struct *sighand;
1191 struct svc_program *progp;
1192 unsigned long flags;
1195 clear_thread_flag(TIF_SIGPENDING);
1197 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1198 for (i = 0; i < progp->pg_nvers; i++) {
1199 if (progp->pg_vers[i] == NULL)
1201 if (progp->pg_vers[i]->vs_hidden)
1203 __svc_unregister(net, progp->pg_prog, i, progp->pg_name);
1208 sighand = rcu_dereference(current->sighand);
1209 spin_lock_irqsave(&sighand->siglock, flags);
1210 recalc_sigpending();
1211 spin_unlock_irqrestore(&sighand->siglock, flags);
1216 * dprintk the given error with the address of the client that caused it.
1218 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
1219 static __printf(2, 3)
1220 void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...)
1222 struct va_format vaf;
1224 char buf[RPC_MAX_ADDRBUFLEN];
1226 va_start(args, fmt);
1231 dprintk("svc: %s: %pV", svc_print_addr(rqstp, buf, sizeof(buf)), &vaf);
1236 static __printf(2,3) void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...) {}
1240 svc_generic_init_request(struct svc_rqst *rqstp,
1241 const struct svc_program *progp,
1242 struct svc_process_info *ret)
1244 const struct svc_version *versp = NULL; /* compiler food */
1245 const struct svc_procedure *procp = NULL;
1247 if (rqstp->rq_vers >= progp->pg_nvers )
1249 versp = progp->pg_vers[rqstp->rq_vers];
1254 * Some protocol versions (namely NFSv4) require some form of
1255 * congestion control. (See RFC 7530 section 3.1 paragraph 2)
1256 * In other words, UDP is not allowed. We mark those when setting
1257 * up the svc_xprt, and verify that here.
1259 * The spec is not very clear about what error should be returned
1260 * when someone tries to access a server that is listening on UDP
1261 * for lower versions. RPC_PROG_MISMATCH seems to be the closest
1264 if (versp->vs_need_cong_ctrl && rqstp->rq_xprt &&
1265 !test_bit(XPT_CONG_CTRL, &rqstp->rq_xprt->xpt_flags))
1268 if (rqstp->rq_proc >= versp->vs_nproc)
1270 rqstp->rq_procinfo = procp = &versp->vs_proc[rqstp->rq_proc];
1274 /* Initialize storage for argp and resp */
1275 memset(rqstp->rq_argp, 0, procp->pc_argzero);
1276 memset(rqstp->rq_resp, 0, procp->pc_ressize);
1278 /* Bump per-procedure stats counter */
1279 this_cpu_inc(versp->vs_count[rqstp->rq_proc]);
1281 ret->dispatch = versp->vs_dispatch;
1284 ret->mismatch.lovers = progp->pg_lovers;
1285 ret->mismatch.hivers = progp->pg_hivers;
1286 return rpc_prog_mismatch;
1288 return rpc_proc_unavail;
1290 EXPORT_SYMBOL_GPL(svc_generic_init_request);
1293 * Common routine for processing the RPC request.
1296 svc_process_common(struct svc_rqst *rqstp)
1298 struct xdr_stream *xdr = &rqstp->rq_res_stream;
1299 struct svc_program *progp;
1300 const struct svc_procedure *procp = NULL;
1301 struct svc_serv *serv = rqstp->rq_server;
1302 struct svc_process_info process;
1303 enum svc_auth_status auth_res;
1304 unsigned int aoffset;
1308 /* Will be turned off by GSS integrity and privacy services */
1309 set_bit(RQ_SPLICE_OK, &rqstp->rq_flags);
1310 /* Will be turned off only when NFSv4 Sessions are used */
1311 set_bit(RQ_USEDEFERRAL, &rqstp->rq_flags);
1312 clear_bit(RQ_DROPME, &rqstp->rq_flags);
1314 /* Construct the first words of the reply: */
1315 svcxdr_init_encode(rqstp);
1316 xdr_stream_encode_be32(xdr, rqstp->rq_xid);
1317 xdr_stream_encode_be32(xdr, rpc_reply);
1319 p = xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 4);
1322 if (*p++ != cpu_to_be32(RPC_VERSION))
1325 xdr_stream_encode_be32(xdr, rpc_msg_accepted);
1327 rqstp->rq_prog = be32_to_cpup(p++);
1328 rqstp->rq_vers = be32_to_cpup(p++);
1329 rqstp->rq_proc = be32_to_cpup(p);
1331 for (progp = serv->sv_program; progp; progp = progp->pg_next)
1332 if (rqstp->rq_prog == progp->pg_prog)
1336 * Decode auth data, and add verifier to reply buffer.
1337 * We do this before anything else in order to get a decent
1340 auth_res = svc_authenticate(rqstp);
1341 /* Also give the program a chance to reject this call: */
1342 if (auth_res == SVC_OK && progp)
1343 auth_res = progp->pg_authenticate(rqstp);
1344 trace_svc_authenticate(rqstp, auth_res);
1349 goto err_garbage_args;
1351 goto err_system_err;
1361 pr_warn_once("Unexpected svc_auth_status (%d)\n", auth_res);
1362 goto err_system_err;
1368 switch (progp->pg_init_request(rqstp, progp, &process)) {
1371 case rpc_prog_unavail:
1373 case rpc_prog_mismatch:
1375 case rpc_proc_unavail:
1379 procp = rqstp->rq_procinfo;
1380 /* Should this check go into the dispatcher? */
1381 if (!procp || !procp->pc_func)
1384 /* Syntactic check complete */
1385 serv->sv_stats->rpccnt++;
1386 trace_svc_process(rqstp, progp->pg_name);
1388 aoffset = xdr_stream_pos(xdr);
1390 /* un-reserve some of the out-queue now that we have a
1391 * better idea of reply size
1393 if (procp->pc_xdrressize)
1394 svc_reserve_auth(rqstp, procp->pc_xdrressize<<2);
1396 /* Call the function that processes the request. */
1397 rc = process.dispatch(rqstp);
1398 if (procp->pc_release)
1399 procp->pc_release(rqstp);
1400 xdr_finish_decode(xdr);
1404 if (rqstp->rq_auth_stat != rpc_auth_ok)
1407 if (*rqstp->rq_accept_statp != rpc_success)
1408 xdr_truncate_encode(xdr, aoffset);
1410 if (procp->pc_encode == NULL)
1414 if (svc_authorise(rqstp))
1416 return 1; /* Caller can now send it */
1419 svc_authorise(rqstp); /* doesn't hurt to call this twice */
1420 dprintk("svc: svc_process dropit\n");
1424 svc_authorise(rqstp);
1426 if (rqstp->rq_xprt && test_bit(XPT_TEMP, &rqstp->rq_xprt->xpt_flags))
1427 svc_xprt_close(rqstp->rq_xprt);
1428 dprintk("svc: svc_process close\n");
1432 svc_printk(rqstp, "short len %u, dropping request\n",
1437 serv->sv_stats->rpcbadfmt++;
1438 xdr_stream_encode_u32(xdr, RPC_MSG_DENIED);
1439 xdr_stream_encode_u32(xdr, RPC_MISMATCH);
1440 /* Only RPCv2 supported */
1441 xdr_stream_encode_u32(xdr, RPC_VERSION);
1442 xdr_stream_encode_u32(xdr, RPC_VERSION);
1443 return 1; /* don't wrap */
1446 dprintk("svc: authentication failed (%d)\n",
1447 be32_to_cpu(rqstp->rq_auth_stat));
1448 serv->sv_stats->rpcbadauth++;
1449 /* Restore write pointer to location of reply status: */
1450 xdr_truncate_encode(xdr, XDR_UNIT * 2);
1451 xdr_stream_encode_u32(xdr, RPC_MSG_DENIED);
1452 xdr_stream_encode_u32(xdr, RPC_AUTH_ERROR);
1453 xdr_stream_encode_be32(xdr, rqstp->rq_auth_stat);
1457 dprintk("svc: unknown program %d\n", rqstp->rq_prog);
1458 serv->sv_stats->rpcbadfmt++;
1459 *rqstp->rq_accept_statp = rpc_prog_unavail;
1463 svc_printk(rqstp, "unknown version (%d for prog %d, %s)\n",
1464 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);
1480 serv->sv_stats->rpcbadfmt++;
1481 *rqstp->rq_accept_statp = rpc_proc_unavail;
1485 svc_printk(rqstp, "failed to decode RPC header\n");
1487 serv->sv_stats->rpcbadfmt++;
1488 *rqstp->rq_accept_statp = rpc_garbage_args;
1492 serv->sv_stats->rpcbadfmt++;
1493 *rqstp->rq_accept_statp = rpc_system_err;
1498 * svc_process - Execute one RPC transaction
1499 * @rqstp: RPC transaction context
1502 void svc_process(struct svc_rqst *rqstp)
1504 struct kvec *resv = &rqstp->rq_res.head[0];
1507 #if IS_ENABLED(CONFIG_FAIL_SUNRPC)
1508 if (!fail_sunrpc.ignore_server_disconnect &&
1509 should_fail(&fail_sunrpc.attr, 1))
1510 svc_xprt_deferred_close(rqstp->rq_xprt);
1514 * Setup response xdr_buf.
1515 * Initially it has just one page
1517 rqstp->rq_next_page = &rqstp->rq_respages[1];
1518 resv->iov_base = page_address(rqstp->rq_respages[0]);
1520 rqstp->rq_res.pages = rqstp->rq_next_page;
1521 rqstp->rq_res.len = 0;
1522 rqstp->rq_res.page_base = 0;
1523 rqstp->rq_res.page_len = 0;
1524 rqstp->rq_res.buflen = PAGE_SIZE;
1525 rqstp->rq_res.tail[0].iov_base = NULL;
1526 rqstp->rq_res.tail[0].iov_len = 0;
1528 svcxdr_init_decode(rqstp);
1529 p = xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 2);
1532 rqstp->rq_xid = *p++;
1533 if (unlikely(*p != rpc_call))
1536 if (!svc_process_common(rqstp))
1542 svc_printk(rqstp, "bad direction 0x%08x, dropping request\n",
1544 rqstp->rq_server->sv_stats->rpcbadfmt++;
1549 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1551 * svc_process_bc - process a reverse-direction RPC request
1552 * @req: RPC request to be used for client-side processing
1553 * @rqstp: server-side execution context
1556 void svc_process_bc(struct rpc_rqst *req, struct svc_rqst *rqstp)
1558 struct rpc_task *task;
1561 /* Build the svc_rqst used by the common processing routine */
1562 rqstp->rq_xid = req->rq_xid;
1563 rqstp->rq_prot = req->rq_xprt->prot;
1564 rqstp->rq_bc_net = req->rq_xprt->xprt_net;
1566 rqstp->rq_addrlen = sizeof(req->rq_xprt->addr);
1567 memcpy(&rqstp->rq_addr, &req->rq_xprt->addr, rqstp->rq_addrlen);
1568 memcpy(&rqstp->rq_arg, &req->rq_rcv_buf, sizeof(rqstp->rq_arg));
1569 memcpy(&rqstp->rq_res, &req->rq_snd_buf, sizeof(rqstp->rq_res));
1571 /* Adjust the argument buffer length */
1572 rqstp->rq_arg.len = req->rq_private_buf.len;
1573 if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len) {
1574 rqstp->rq_arg.head[0].iov_len = rqstp->rq_arg.len;
1575 rqstp->rq_arg.page_len = 0;
1576 } else if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len +
1577 rqstp->rq_arg.page_len)
1578 rqstp->rq_arg.page_len = rqstp->rq_arg.len -
1579 rqstp->rq_arg.head[0].iov_len;
1581 rqstp->rq_arg.len = rqstp->rq_arg.head[0].iov_len +
1582 rqstp->rq_arg.page_len;
1584 /* Reset the response buffer */
1585 rqstp->rq_res.head[0].iov_len = 0;
1588 * Skip the XID and calldir fields because they've already
1589 * been processed by the caller.
1591 svcxdr_init_decode(rqstp);
1592 if (!xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 2))
1595 /* Parse and execute the bc call */
1596 proc_error = svc_process_common(rqstp);
1598 atomic_dec(&req->rq_xprt->bc_slot_count);
1600 /* Processing error: drop the request */
1601 xprt_free_bc_request(req);
1604 /* Finally, send the reply synchronously */
1605 memcpy(&req->rq_snd_buf, &rqstp->rq_res, sizeof(req->rq_snd_buf));
1606 task = rpc_run_bc_task(req);
1610 WARN_ON_ONCE(atomic_read(&task->tk_count) != 1);
1613 EXPORT_SYMBOL_GPL(svc_process_bc);
1614 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1617 * svc_max_payload - Return transport-specific limit on the RPC payload
1618 * @rqstp: RPC transaction context
1620 * Returns the maximum number of payload bytes the current transport
1623 u32 svc_max_payload(const struct svc_rqst *rqstp)
1625 u32 max = rqstp->rq_xprt->xpt_class->xcl_max_payload;
1627 if (rqstp->rq_server->sv_max_payload < max)
1628 max = rqstp->rq_server->sv_max_payload;
1631 EXPORT_SYMBOL_GPL(svc_max_payload);
1634 * svc_proc_name - Return RPC procedure name in string form
1635 * @rqstp: svc_rqst to operate on
1638 * Pointer to a NUL-terminated string
1640 const char *svc_proc_name(const struct svc_rqst *rqstp)
1642 if (rqstp && rqstp->rq_procinfo)
1643 return rqstp->rq_procinfo->pc_name;
1649 * svc_encode_result_payload - mark a range of bytes as a result payload
1650 * @rqstp: svc_rqst to operate on
1651 * @offset: payload's byte offset in rqstp->rq_res
1652 * @length: size of payload, in bytes
1654 * Returns zero on success, or a negative errno if a permanent
1657 int svc_encode_result_payload(struct svc_rqst *rqstp, unsigned int offset,
1658 unsigned int length)
1660 return rqstp->rq_xprt->xpt_ops->xpo_result_payload(rqstp, offset,
1663 EXPORT_SYMBOL_GPL(svc_encode_result_payload);
1666 * svc_fill_write_vector - Construct data argument for VFS write call
1667 * @rqstp: svc_rqst to operate on
1668 * @payload: xdr_buf containing only the write data payload
1670 * Fills in rqstp::rq_vec, and returns the number of elements.
1672 unsigned int svc_fill_write_vector(struct svc_rqst *rqstp,
1673 struct xdr_buf *payload)
1675 struct page **pages = payload->pages;
1676 struct kvec *first = payload->head;
1677 struct kvec *vec = rqstp->rq_vec;
1678 size_t total = payload->len;
1681 /* Some types of transport can present the write payload
1682 * entirely in rq_arg.pages. In this case, @first is empty.
1685 if (first->iov_len) {
1686 vec[i].iov_base = first->iov_base;
1687 vec[i].iov_len = min_t(size_t, total, first->iov_len);
1688 total -= vec[i].iov_len;
1693 vec[i].iov_base = page_address(*pages);
1694 vec[i].iov_len = min_t(size_t, total, PAGE_SIZE);
1695 total -= vec[i].iov_len;
1700 WARN_ON_ONCE(i > ARRAY_SIZE(rqstp->rq_vec));
1703 EXPORT_SYMBOL_GPL(svc_fill_write_vector);
1706 * svc_fill_symlink_pathname - Construct pathname argument for VFS symlink call
1707 * @rqstp: svc_rqst to operate on
1708 * @first: buffer containing first section of pathname
1709 * @p: buffer containing remaining section of pathname
1710 * @total: total length of the pathname argument
1712 * The VFS symlink API demands a NUL-terminated pathname in mapped memory.
1713 * Returns pointer to a NUL-terminated string, or an ERR_PTR. Caller must free
1714 * the returned string.
1716 char *svc_fill_symlink_pathname(struct svc_rqst *rqstp, struct kvec *first,
1717 void *p, size_t total)
1719 size_t len, remaining;
1722 result = kmalloc(total + 1, GFP_KERNEL);
1724 return ERR_PTR(-ESERVERFAULT);
1729 len = min_t(size_t, total, first->iov_len);
1731 memcpy(dst, first->iov_base, len);
1737 len = min_t(size_t, remaining, PAGE_SIZE);
1738 memcpy(dst, p, len);
1744 /* Sanity check: Linux doesn't allow the pathname argument to
1745 * contain a NUL byte.
1747 if (strlen(result) != total) {
1749 return ERR_PTR(-EINVAL);
1753 EXPORT_SYMBOL_GPL(svc_fill_symlink_pathname);