Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/net/sunrpc/svc.c
4 *
5 * High-level RPC service routines
6 *
7 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
8 *
9 * Multiple threads pools and NUMAisation
10 * Copyright (c) 2006 Silicon Graphics, Inc.
11 * by Greg Banks <gnb@melbourne.sgi.com>
12 */
13
14#include <linux/linkage.h>
15#include <linux/sched/signal.h>
16#include <linux/errno.h>
17#include <linux/net.h>
18#include <linux/in.h>
19#include <linux/mm.h>
20#include <linux/interrupt.h>
21#include <linux/module.h>
22#include <linux/kthread.h>
23#include <linux/slab.h>
24
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>
31
32#include <trace/events/sunrpc.h>
33
34#include "fail.h"
35
36#define RPCDBG_FACILITY RPCDBG_SVCDSP
37
38static void svc_unregister(const struct svc_serv *serv, struct net *net);
39
40#define SVC_POOL_DEFAULT SVC_POOL_GLOBAL
41
42/*
43 * Mode for mapping cpus to pools.
44 */
45enum {
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 */
51};
52
53/*
54 * Structure for mapping cpus to pools and vice versa.
55 * Setup once during sunrpc initialisation.
56 */
57
58struct svc_pool_map {
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" */
63 unsigned int npools;
64 unsigned int *pool_to; /* maps pool id to cpu or node */
65 unsigned int *to_pool; /* maps cpu or node to pool id */
66};
67
68static struct svc_pool_map svc_pool_map = {
69 .mode = SVC_POOL_DEFAULT
70};
71
72static DEFINE_MUTEX(svc_pool_map_mutex);/* protects svc_pool_map.count only */
73
74static int
75param_set_pool_mode(const char *val, const struct kernel_param *kp)
76{
77 int *ip = (int *)kp->arg;
78 struct svc_pool_map *m = &svc_pool_map;
79 int err;
80
81 mutex_lock(&svc_pool_map_mutex);
82
83 err = -EBUSY;
84 if (m->count)
85 goto out;
86
87 err = 0;
88 if (!strncmp(val, "auto", 4))
89 *ip = SVC_POOL_AUTO;
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;
96 else
97 err = -EINVAL;
98
99out:
100 mutex_unlock(&svc_pool_map_mutex);
101 return err;
102}
103
104static int
105param_get_pool_mode(char *buf, const struct kernel_param *kp)
106{
107 int *ip = (int *)kp->arg;
108
109 switch (*ip)
110 {
111 case SVC_POOL_AUTO:
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);
119 default:
120 return sprintf(buf, "%d\n", *ip);
121 }
122}
123
124module_param_call(pool_mode, param_set_pool_mode, param_get_pool_mode,
125 &svc_pool_map.mode, 0644);
126
127/*
128 * Detect best pool mapping mode heuristically,
129 * according to the machine's topology.
130 */
131static int
132svc_pool_map_choose_mode(void)
133{
134 unsigned int node;
135
136 if (nr_online_nodes > 1) {
137 /*
138 * Actually have multiple NUMA nodes,
139 * so split pools on NUMA node boundaries
140 */
141 return SVC_POOL_PERNODE;
142 }
143
144 node = first_online_node;
145 if (nr_cpus_node(node) > 2) {
146 /*
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.
151 */
152 return SVC_POOL_PERCPU;
153 }
154
155 /* default: one global pool */
156 return SVC_POOL_GLOBAL;
157}
158
159/*
160 * Allocate the to_pool[] and pool_to[] arrays.
161 * Returns 0 on success or an errno.
162 */
163static int
164svc_pool_map_alloc_arrays(struct svc_pool_map *m, unsigned int maxpools)
165{
166 m->to_pool = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
167 if (!m->to_pool)
168 goto fail;
169 m->pool_to = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
170 if (!m->pool_to)
171 goto fail_free;
172
173 return 0;
174
175fail_free:
176 kfree(m->to_pool);
177 m->to_pool = NULL;
178fail:
179 return -ENOMEM;
180}
181
182/*
183 * Initialise the pool map for SVC_POOL_PERCPU mode.
184 * Returns number of pools or <0 on error.
185 */
186static int
187svc_pool_map_init_percpu(struct svc_pool_map *m)
188{
189 unsigned int maxpools = nr_cpu_ids;
190 unsigned int pidx = 0;
191 unsigned int cpu;
192 int err;
193
194 err = svc_pool_map_alloc_arrays(m, maxpools);
195 if (err)
196 return err;
197
198 for_each_online_cpu(cpu) {
199 BUG_ON(pidx >= maxpools);
200 m->to_pool[cpu] = pidx;
201 m->pool_to[pidx] = cpu;
202 pidx++;
203 }
204 /* cpus brought online later all get mapped to pool0, sorry */
205
206 return pidx;
207};
208
209
210/*
211 * Initialise the pool map for SVC_POOL_PERNODE mode.
212 * Returns number of pools or <0 on error.
213 */
214static int
215svc_pool_map_init_pernode(struct svc_pool_map *m)
216{
217 unsigned int maxpools = nr_node_ids;
218 unsigned int pidx = 0;
219 unsigned int node;
220 int err;
221
222 err = svc_pool_map_alloc_arrays(m, maxpools);
223 if (err)
224 return err;
225
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;
231 pidx++;
232 }
233 /* nodes brought online later all get mapped to pool0, sorry */
234
235 return pidx;
236}
237
238
239/*
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
244 * was taken.
245 */
246static unsigned int
247svc_pool_map_get(void)
248{
249 struct svc_pool_map *m = &svc_pool_map;
250 int npools = -1;
251
252 mutex_lock(&svc_pool_map_mutex);
253
254 if (m->count++) {
255 mutex_unlock(&svc_pool_map_mutex);
256 WARN_ON_ONCE(m->npools <= 1);
257 return m->npools;
258 }
259
260 if (m->mode == SVC_POOL_AUTO)
261 m->mode = svc_pool_map_choose_mode();
262
263 switch (m->mode) {
264 case SVC_POOL_PERCPU:
265 npools = svc_pool_map_init_percpu(m);
266 break;
267 case SVC_POOL_PERNODE:
268 npools = svc_pool_map_init_pernode(m);
269 break;
270 }
271
272 if (npools <= 0) {
273 /* default, or memory allocation failure */
274 npools = 1;
275 m->mode = SVC_POOL_GLOBAL;
276 }
277 m->npools = npools;
278
279 if (npools == 1)
280 /* service is unpooled, so doesn't hold a reference */
281 m->count--;
282
283 mutex_unlock(&svc_pool_map_mutex);
284 return npools;
285}
286
287/*
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.
294 */
295static void
296svc_pool_map_put(int npools)
297{
298 struct svc_pool_map *m = &svc_pool_map;
299
300 if (npools <= 1)
301 return;
302 mutex_lock(&svc_pool_map_mutex);
303
304 if (!--m->count) {
305 kfree(m->to_pool);
306 m->to_pool = NULL;
307 kfree(m->pool_to);
308 m->pool_to = NULL;
309 m->npools = 0;
310 }
311
312 mutex_unlock(&svc_pool_map_mutex);
313}
314
315static int svc_pool_map_get_node(unsigned int pidx)
316{
317 const struct svc_pool_map *m = &svc_pool_map;
318
319 if (m->count) {
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];
324 }
325 return NUMA_NO_NODE;
326}
327/*
328 * Set the given thread's cpus_allowed mask so that it
329 * will only run on cpus in the given pool.
330 */
331static inline void
332svc_pool_map_set_cpumask(struct task_struct *task, unsigned int pidx)
333{
334 struct svc_pool_map *m = &svc_pool_map;
335 unsigned int node = m->pool_to[pidx];
336
337 /*
338 * The caller checks for sv_nrpools > 1, which
339 * implies that we've been initialized.
340 */
341 WARN_ON_ONCE(m->count == 0);
342 if (m->count == 0)
343 return;
344
345 switch (m->mode) {
346 case SVC_POOL_PERCPU:
347 {
348 set_cpus_allowed_ptr(task, cpumask_of(node));
349 break;
350 }
351 case SVC_POOL_PERNODE:
352 {
353 set_cpus_allowed_ptr(task, cpumask_of_node(node));
354 break;
355 }
356 }
357}
358
359/**
360 * svc_pool_for_cpu - Select pool to run a thread on this cpu
361 * @serv: An RPC service
362 *
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.
366 *
367 * Return value:
368 * A pointer to an svc_pool
369 */
370struct svc_pool *svc_pool_for_cpu(struct svc_serv *serv)
371{
372 struct svc_pool_map *m = &svc_pool_map;
373 int cpu = raw_smp_processor_id();
374 unsigned int pidx = 0;
375
376 if (serv->sv_nrpools <= 1)
377 return serv->sv_pools;
378
379 switch (m->mode) {
380 case SVC_POOL_PERCPU:
381 pidx = m->to_pool[cpu];
382 break;
383 case SVC_POOL_PERNODE:
384 pidx = m->to_pool[cpu_to_node(cpu)];
385 break;
386 }
387
388 return &serv->sv_pools[pidx % serv->sv_nrpools];
389}
390
391int svc_rpcb_setup(struct svc_serv *serv, struct net *net)
392{
393 int err;
394
395 err = rpcb_create_local(net);
396 if (err)
397 return err;
398
399 /* Remove any stale portmap registrations */
400 svc_unregister(serv, net);
401 return 0;
402}
403EXPORT_SYMBOL_GPL(svc_rpcb_setup);
404
405void svc_rpcb_cleanup(struct svc_serv *serv, struct net *net)
406{
407 svc_unregister(serv, net);
408 rpcb_put_local(net);
409}
410EXPORT_SYMBOL_GPL(svc_rpcb_cleanup);
411
412static int svc_uses_rpcbind(struct svc_serv *serv)
413{
414 struct svc_program *progp;
415 unsigned int i;
416
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)
420 continue;
421 if (!progp->pg_vers[i]->vs_hidden)
422 return 1;
423 }
424 }
425
426 return 0;
427}
428
429int svc_bind(struct svc_serv *serv, struct net *net)
430{
431 if (!svc_uses_rpcbind(serv))
432 return 0;
433 return svc_rpcb_setup(serv, net);
434}
435EXPORT_SYMBOL_GPL(svc_bind);
436
437#if defined(CONFIG_SUNRPC_BACKCHANNEL)
438static void
439__svc_init_bc(struct svc_serv *serv)
440{
441 INIT_LIST_HEAD(&serv->sv_cb_list);
442 spin_lock_init(&serv->sv_cb_lock);
443 init_waitqueue_head(&serv->sv_cb_waitq);
444}
445#else
446static void
447__svc_init_bc(struct svc_serv *serv)
448{
449}
450#endif
451
452/*
453 * Create an RPC service
454 */
455static struct svc_serv *
456__svc_create(struct svc_program *prog, unsigned int bufsize, int npools,
457 int (*threadfn)(void *data))
458{
459 struct svc_serv *serv;
460 unsigned int vers;
461 unsigned int xdrsize;
462 unsigned int i;
463
464 if (!(serv = kzalloc(sizeof(*serv), GFP_KERNEL)))
465 return NULL;
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;
475 xdrsize = 0;
476 while (prog) {
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;
485 }
486 prog = prog->pg_next;
487 }
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);
493
494 __svc_init_bc(serv);
495
496 serv->sv_nrpools = npools;
497 serv->sv_pools =
498 kcalloc(serv->sv_nrpools, sizeof(struct svc_pool),
499 GFP_KERNEL);
500 if (!serv->sv_pools) {
501 kfree(serv);
502 return NULL;
503 }
504
505 for (i = 0; i < serv->sv_nrpools; i++) {
506 struct svc_pool *pool = &serv->sv_pools[i];
507
508 dprintk("svc: initialising pool %u for %s\n",
509 i, serv->sv_name);
510
511 pool->sp_id = i;
512 INIT_LIST_HEAD(&pool->sp_sockets);
513 INIT_LIST_HEAD(&pool->sp_all_threads);
514 spin_lock_init(&pool->sp_lock);
515 }
516
517 return serv;
518}
519
520/**
521 * svc_create - Create an RPC service
522 * @prog: the RPC program the new service will handle
523 * @bufsize: maximum message size for @prog
524 * @threadfn: a function to service RPC requests for @prog
525 *
526 * Returns an instantiated struct svc_serv object or NULL.
527 */
528struct svc_serv *svc_create(struct svc_program *prog, unsigned int bufsize,
529 int (*threadfn)(void *data))
530{
531 return __svc_create(prog, bufsize, 1, threadfn);
532}
533EXPORT_SYMBOL_GPL(svc_create);
534
535/**
536 * svc_create_pooled - Create an RPC service with pooled threads
537 * @prog: the RPC program the new service will handle
538 * @bufsize: maximum message size for @prog
539 * @threadfn: a function to service RPC requests for @prog
540 *
541 * Returns an instantiated struct svc_serv object or NULL.
542 */
543struct svc_serv *svc_create_pooled(struct svc_program *prog,
544 unsigned int bufsize,
545 int (*threadfn)(void *data))
546{
547 struct svc_serv *serv;
548 unsigned int npools = svc_pool_map_get();
549
550 serv = __svc_create(prog, bufsize, npools, threadfn);
551 if (!serv)
552 goto out_err;
553 return serv;
554out_err:
555 svc_pool_map_put(npools);
556 return NULL;
557}
558EXPORT_SYMBOL_GPL(svc_create_pooled);
559
560/*
561 * Destroy an RPC service. Should be called with appropriate locking to
562 * protect sv_permsocks and sv_tempsocks.
563 */
564void
565svc_destroy(struct kref *ref)
566{
567 struct svc_serv *serv = container_of(ref, struct svc_serv, sv_refcnt);
568
569 dprintk("svc: svc_destroy(%s)\n", serv->sv_program->pg_name);
570 timer_shutdown_sync(&serv->sv_temptimer);
571
572 /*
573 * The last user is gone and thus all sockets have to be destroyed to
574 * the point. Check this.
575 */
576 BUG_ON(!list_empty(&serv->sv_permsocks));
577 BUG_ON(!list_empty(&serv->sv_tempsocks));
578
579 cache_clean_deferred(serv);
580
581 svc_pool_map_put(serv->sv_nrpools);
582
583 kfree(serv->sv_pools);
584 kfree(serv);
585}
586EXPORT_SYMBOL_GPL(svc_destroy);
587
588/*
589 * Allocate an RPC server's buffer space.
590 * We allocate pages and place them in rq_pages.
591 */
592static int
593svc_init_buffer(struct svc_rqst *rqstp, unsigned int size, int node)
594{
595 unsigned int pages, arghi;
596
597 /* bc_xprt uses fore channel allocated buffers */
598 if (svc_is_backchannel(rqstp))
599 return 1;
600
601 pages = size / PAGE_SIZE + 1; /* extra page as we hold both request and reply.
602 * We assume one is at most one page
603 */
604 arghi = 0;
605 WARN_ON_ONCE(pages > RPCSVC_MAXPAGES);
606 if (pages > RPCSVC_MAXPAGES)
607 pages = RPCSVC_MAXPAGES;
608 while (pages) {
609 struct page *p = alloc_pages_node(node, GFP_KERNEL, 0);
610 if (!p)
611 break;
612 rqstp->rq_pages[arghi++] = p;
613 pages--;
614 }
615 return pages == 0;
616}
617
618/*
619 * Release an RPC server buffer
620 */
621static void
622svc_release_buffer(struct svc_rqst *rqstp)
623{
624 unsigned int i;
625
626 for (i = 0; i < ARRAY_SIZE(rqstp->rq_pages); i++)
627 if (rqstp->rq_pages[i])
628 put_page(rqstp->rq_pages[i]);
629}
630
631struct svc_rqst *
632svc_rqst_alloc(struct svc_serv *serv, struct svc_pool *pool, int node)
633{
634 struct svc_rqst *rqstp;
635
636 rqstp = kzalloc_node(sizeof(*rqstp), GFP_KERNEL, node);
637 if (!rqstp)
638 return rqstp;
639
640 __set_bit(RQ_BUSY, &rqstp->rq_flags);
641 rqstp->rq_server = serv;
642 rqstp->rq_pool = pool;
643
644 rqstp->rq_scratch_page = alloc_pages_node(node, GFP_KERNEL, 0);
645 if (!rqstp->rq_scratch_page)
646 goto out_enomem;
647
648 rqstp->rq_argp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
649 if (!rqstp->rq_argp)
650 goto out_enomem;
651
652 rqstp->rq_resp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
653 if (!rqstp->rq_resp)
654 goto out_enomem;
655
656 if (!svc_init_buffer(rqstp, serv->sv_max_mesg, node))
657 goto out_enomem;
658
659 return rqstp;
660out_enomem:
661 svc_rqst_free(rqstp);
662 return NULL;
663}
664EXPORT_SYMBOL_GPL(svc_rqst_alloc);
665
666static struct svc_rqst *
667svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
668{
669 struct svc_rqst *rqstp;
670
671 rqstp = svc_rqst_alloc(serv, pool, node);
672 if (!rqstp)
673 return ERR_PTR(-ENOMEM);
674
675 svc_get(serv);
676 spin_lock_bh(&serv->sv_lock);
677 serv->sv_nrthreads += 1;
678 spin_unlock_bh(&serv->sv_lock);
679
680 spin_lock_bh(&pool->sp_lock);
681 pool->sp_nrthreads++;
682 list_add_rcu(&rqstp->rq_all, &pool->sp_all_threads);
683 spin_unlock_bh(&pool->sp_lock);
684 return rqstp;
685}
686
687/*
688 * Choose a pool in which to create a new thread, for svc_set_num_threads
689 */
690static inline struct svc_pool *
691choose_pool(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
692{
693 if (pool != NULL)
694 return pool;
695
696 return &serv->sv_pools[(*state)++ % serv->sv_nrpools];
697}
698
699/*
700 * Choose a thread to kill, for svc_set_num_threads
701 */
702static inline struct task_struct *
703choose_victim(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
704{
705 unsigned int i;
706 struct task_struct *task = NULL;
707
708 if (pool != NULL) {
709 spin_lock_bh(&pool->sp_lock);
710 } else {
711 /* choose a pool in round-robin fashion */
712 for (i = 0; i < serv->sv_nrpools; i++) {
713 pool = &serv->sv_pools[--(*state) % serv->sv_nrpools];
714 spin_lock_bh(&pool->sp_lock);
715 if (!list_empty(&pool->sp_all_threads))
716 goto found_pool;
717 spin_unlock_bh(&pool->sp_lock);
718 }
719 return NULL;
720 }
721
722found_pool:
723 if (!list_empty(&pool->sp_all_threads)) {
724 struct svc_rqst *rqstp;
725
726 /*
727 * Remove from the pool->sp_all_threads list
728 * so we don't try to kill it again.
729 */
730 rqstp = list_entry(pool->sp_all_threads.next, struct svc_rqst, rq_all);
731 set_bit(RQ_VICTIM, &rqstp->rq_flags);
732 list_del_rcu(&rqstp->rq_all);
733 task = rqstp->rq_task;
734 }
735 spin_unlock_bh(&pool->sp_lock);
736
737 return task;
738}
739
740/* create new threads */
741static int
742svc_start_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
743{
744 struct svc_rqst *rqstp;
745 struct task_struct *task;
746 struct svc_pool *chosen_pool;
747 unsigned int state = serv->sv_nrthreads-1;
748 int node;
749
750 do {
751 nrservs--;
752 chosen_pool = choose_pool(serv, pool, &state);
753
754 node = svc_pool_map_get_node(chosen_pool->sp_id);
755 rqstp = svc_prepare_thread(serv, chosen_pool, node);
756 if (IS_ERR(rqstp))
757 return PTR_ERR(rqstp);
758
759 task = kthread_create_on_node(serv->sv_threadfn, rqstp,
760 node, "%s", serv->sv_name);
761 if (IS_ERR(task)) {
762 svc_exit_thread(rqstp);
763 return PTR_ERR(task);
764 }
765
766 rqstp->rq_task = task;
767 if (serv->sv_nrpools > 1)
768 svc_pool_map_set_cpumask(task, chosen_pool->sp_id);
769
770 svc_sock_update_bufs(serv);
771 wake_up_process(task);
772 } while (nrservs > 0);
773
774 return 0;
775}
776
777/*
778 * Create or destroy enough new threads to make the number
779 * of threads the given number. If `pool' is non-NULL, applies
780 * only to threads in that pool, otherwise round-robins between
781 * all pools. Caller must ensure that mutual exclusion between this and
782 * server startup or shutdown.
783 */
784
785/* destroy old threads */
786static int
787svc_stop_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
788{
789 struct task_struct *task;
790 unsigned int state = serv->sv_nrthreads-1;
791
792 /* destroy old threads */
793 do {
794 task = choose_victim(serv, pool, &state);
795 if (task == NULL)
796 break;
797 kthread_stop(task);
798 nrservs++;
799 } while (nrservs < 0);
800 return 0;
801}
802
803int
804svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
805{
806 if (pool == NULL) {
807 nrservs -= serv->sv_nrthreads;
808 } else {
809 spin_lock_bh(&pool->sp_lock);
810 nrservs -= pool->sp_nrthreads;
811 spin_unlock_bh(&pool->sp_lock);
812 }
813
814 if (nrservs > 0)
815 return svc_start_kthreads(serv, pool, nrservs);
816 if (nrservs < 0)
817 return svc_stop_kthreads(serv, pool, nrservs);
818 return 0;
819}
820EXPORT_SYMBOL_GPL(svc_set_num_threads);
821
822/**
823 * svc_rqst_replace_page - Replace one page in rq_pages[]
824 * @rqstp: svc_rqst with pages to replace
825 * @page: replacement page
826 *
827 * When replacing a page in rq_pages, batch the release of the
828 * replaced pages to avoid hammering the page allocator.
829 */
830void svc_rqst_replace_page(struct svc_rqst *rqstp, struct page *page)
831{
832 if (*rqstp->rq_next_page) {
833 if (!pagevec_space(&rqstp->rq_pvec))
834 __pagevec_release(&rqstp->rq_pvec);
835 pagevec_add(&rqstp->rq_pvec, *rqstp->rq_next_page);
836 }
837
838 get_page(page);
839 *(rqstp->rq_next_page++) = page;
840}
841EXPORT_SYMBOL_GPL(svc_rqst_replace_page);
842
843/*
844 * Called from a server thread as it's exiting. Caller must hold the "service
845 * mutex" for the service.
846 */
847void
848svc_rqst_free(struct svc_rqst *rqstp)
849{
850 svc_release_buffer(rqstp);
851 if (rqstp->rq_scratch_page)
852 put_page(rqstp->rq_scratch_page);
853 kfree(rqstp->rq_resp);
854 kfree(rqstp->rq_argp);
855 kfree(rqstp->rq_auth_data);
856 kfree_rcu(rqstp, rq_rcu_head);
857}
858EXPORT_SYMBOL_GPL(svc_rqst_free);
859
860void
861svc_exit_thread(struct svc_rqst *rqstp)
862{
863 struct svc_serv *serv = rqstp->rq_server;
864 struct svc_pool *pool = rqstp->rq_pool;
865
866 spin_lock_bh(&pool->sp_lock);
867 pool->sp_nrthreads--;
868 if (!test_and_set_bit(RQ_VICTIM, &rqstp->rq_flags))
869 list_del_rcu(&rqstp->rq_all);
870 spin_unlock_bh(&pool->sp_lock);
871
872 spin_lock_bh(&serv->sv_lock);
873 serv->sv_nrthreads -= 1;
874 spin_unlock_bh(&serv->sv_lock);
875 svc_sock_update_bufs(serv);
876
877 svc_rqst_free(rqstp);
878
879 svc_put(serv);
880}
881EXPORT_SYMBOL_GPL(svc_exit_thread);
882
883/*
884 * Register an "inet" protocol family netid with the local
885 * rpcbind daemon via an rpcbind v4 SET request.
886 *
887 * No netconfig infrastructure is available in the kernel, so
888 * we map IP_ protocol numbers to netids by hand.
889 *
890 * Returns zero on success; a negative errno value is returned
891 * if any error occurs.
892 */
893static int __svc_rpcb_register4(struct net *net, const u32 program,
894 const u32 version,
895 const unsigned short protocol,
896 const unsigned short port)
897{
898 const struct sockaddr_in sin = {
899 .sin_family = AF_INET,
900 .sin_addr.s_addr = htonl(INADDR_ANY),
901 .sin_port = htons(port),
902 };
903 const char *netid;
904 int error;
905
906 switch (protocol) {
907 case IPPROTO_UDP:
908 netid = RPCBIND_NETID_UDP;
909 break;
910 case IPPROTO_TCP:
911 netid = RPCBIND_NETID_TCP;
912 break;
913 default:
914 return -ENOPROTOOPT;
915 }
916
917 error = rpcb_v4_register(net, program, version,
918 (const struct sockaddr *)&sin, netid);
919
920 /*
921 * User space didn't support rpcbind v4, so retry this
922 * registration request with the legacy rpcbind v2 protocol.
923 */
924 if (error == -EPROTONOSUPPORT)
925 error = rpcb_register(net, program, version, protocol, port);
926
927 return error;
928}
929
930#if IS_ENABLED(CONFIG_IPV6)
931/*
932 * Register an "inet6" protocol family netid with the local
933 * rpcbind daemon via an rpcbind v4 SET request.
934 *
935 * No netconfig infrastructure is available in the kernel, so
936 * we map IP_ protocol numbers to netids by hand.
937 *
938 * Returns zero on success; a negative errno value is returned
939 * if any error occurs.
940 */
941static int __svc_rpcb_register6(struct net *net, const u32 program,
942 const u32 version,
943 const unsigned short protocol,
944 const unsigned short port)
945{
946 const struct sockaddr_in6 sin6 = {
947 .sin6_family = AF_INET6,
948 .sin6_addr = IN6ADDR_ANY_INIT,
949 .sin6_port = htons(port),
950 };
951 const char *netid;
952 int error;
953
954 switch (protocol) {
955 case IPPROTO_UDP:
956 netid = RPCBIND_NETID_UDP6;
957 break;
958 case IPPROTO_TCP:
959 netid = RPCBIND_NETID_TCP6;
960 break;
961 default:
962 return -ENOPROTOOPT;
963 }
964
965 error = rpcb_v4_register(net, program, version,
966 (const struct sockaddr *)&sin6, netid);
967
968 /*
969 * User space didn't support rpcbind version 4, so we won't
970 * use a PF_INET6 listener.
971 */
972 if (error == -EPROTONOSUPPORT)
973 error = -EAFNOSUPPORT;
974
975 return error;
976}
977#endif /* IS_ENABLED(CONFIG_IPV6) */
978
979/*
980 * Register a kernel RPC service via rpcbind version 4.
981 *
982 * Returns zero on success; a negative errno value is returned
983 * if any error occurs.
984 */
985static int __svc_register(struct net *net, const char *progname,
986 const u32 program, const u32 version,
987 const int family,
988 const unsigned short protocol,
989 const unsigned short port)
990{
991 int error = -EAFNOSUPPORT;
992
993 switch (family) {
994 case PF_INET:
995 error = __svc_rpcb_register4(net, program, version,
996 protocol, port);
997 break;
998#if IS_ENABLED(CONFIG_IPV6)
999 case PF_INET6:
1000 error = __svc_rpcb_register6(net, program, version,
1001 protocol, port);
1002#endif
1003 }
1004
1005 trace_svc_register(progname, version, protocol, port, family, error);
1006 return error;
1007}
1008
1009int svc_rpcbind_set_version(struct net *net,
1010 const struct svc_program *progp,
1011 u32 version, int family,
1012 unsigned short proto,
1013 unsigned short port)
1014{
1015 return __svc_register(net, progp->pg_name, progp->pg_prog,
1016 version, family, proto, port);
1017
1018}
1019EXPORT_SYMBOL_GPL(svc_rpcbind_set_version);
1020
1021int svc_generic_rpcbind_set(struct net *net,
1022 const struct svc_program *progp,
1023 u32 version, int family,
1024 unsigned short proto,
1025 unsigned short port)
1026{
1027 const struct svc_version *vers = progp->pg_vers[version];
1028 int error;
1029
1030 if (vers == NULL)
1031 return 0;
1032
1033 if (vers->vs_hidden) {
1034 trace_svc_noregister(progp->pg_name, version, proto,
1035 port, family, 0);
1036 return 0;
1037 }
1038
1039 /*
1040 * Don't register a UDP port if we need congestion
1041 * control.
1042 */
1043 if (vers->vs_need_cong_ctrl && proto == IPPROTO_UDP)
1044 return 0;
1045
1046 error = svc_rpcbind_set_version(net, progp, version,
1047 family, proto, port);
1048
1049 return (vers->vs_rpcb_optnl) ? 0 : error;
1050}
1051EXPORT_SYMBOL_GPL(svc_generic_rpcbind_set);
1052
1053/**
1054 * svc_register - register an RPC service with the local portmapper
1055 * @serv: svc_serv struct for the service to register
1056 * @net: net namespace for the service to register
1057 * @family: protocol family of service's listener socket
1058 * @proto: transport protocol number to advertise
1059 * @port: port to advertise
1060 *
1061 * Service is registered for any address in the passed-in protocol family
1062 */
1063int svc_register(const struct svc_serv *serv, struct net *net,
1064 const int family, const unsigned short proto,
1065 const unsigned short port)
1066{
1067 struct svc_program *progp;
1068 unsigned int i;
1069 int error = 0;
1070
1071 WARN_ON_ONCE(proto == 0 && port == 0);
1072 if (proto == 0 && port == 0)
1073 return -EINVAL;
1074
1075 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1076 for (i = 0; i < progp->pg_nvers; i++) {
1077
1078 error = progp->pg_rpcbind_set(net, progp, i,
1079 family, proto, port);
1080 if (error < 0) {
1081 printk(KERN_WARNING "svc: failed to register "
1082 "%sv%u RPC service (errno %d).\n",
1083 progp->pg_name, i, -error);
1084 break;
1085 }
1086 }
1087 }
1088
1089 return error;
1090}
1091
1092/*
1093 * If user space is running rpcbind, it should take the v4 UNSET
1094 * and clear everything for this [program, version]. If user space
1095 * is running portmap, it will reject the v4 UNSET, but won't have
1096 * any "inet6" entries anyway. So a PMAP_UNSET should be sufficient
1097 * in this case to clear all existing entries for [program, version].
1098 */
1099static void __svc_unregister(struct net *net, const u32 program, const u32 version,
1100 const char *progname)
1101{
1102 int error;
1103
1104 error = rpcb_v4_register(net, program, version, NULL, "");
1105
1106 /*
1107 * User space didn't support rpcbind v4, so retry this
1108 * request with the legacy rpcbind v2 protocol.
1109 */
1110 if (error == -EPROTONOSUPPORT)
1111 error = rpcb_register(net, program, version, 0, 0);
1112
1113 trace_svc_unregister(progname, version, error);
1114}
1115
1116/*
1117 * All netids, bind addresses and ports registered for [program, version]
1118 * are removed from the local rpcbind database (if the service is not
1119 * hidden) to make way for a new instance of the service.
1120 *
1121 * The result of unregistration is reported via dprintk for those who want
1122 * verification of the result, but is otherwise not important.
1123 */
1124static void svc_unregister(const struct svc_serv *serv, struct net *net)
1125{
1126 struct svc_program *progp;
1127 unsigned long flags;
1128 unsigned int i;
1129
1130 clear_thread_flag(TIF_SIGPENDING);
1131
1132 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1133 for (i = 0; i < progp->pg_nvers; i++) {
1134 if (progp->pg_vers[i] == NULL)
1135 continue;
1136 if (progp->pg_vers[i]->vs_hidden)
1137 continue;
1138 __svc_unregister(net, progp->pg_prog, i, progp->pg_name);
1139 }
1140 }
1141
1142 spin_lock_irqsave(¤t->sighand->siglock, flags);
1143 recalc_sigpending();
1144 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
1145}
1146
1147/*
1148 * dprintk the given error with the address of the client that caused it.
1149 */
1150#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
1151static __printf(2, 3)
1152void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...)
1153{
1154 struct va_format vaf;
1155 va_list args;
1156 char buf[RPC_MAX_ADDRBUFLEN];
1157
1158 va_start(args, fmt);
1159
1160 vaf.fmt = fmt;
1161 vaf.va = &args;
1162
1163 dprintk("svc: %s: %pV", svc_print_addr(rqstp, buf, sizeof(buf)), &vaf);
1164
1165 va_end(args);
1166}
1167#else
1168static __printf(2,3) void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...) {}
1169#endif
1170
1171__be32
1172svc_generic_init_request(struct svc_rqst *rqstp,
1173 const struct svc_program *progp,
1174 struct svc_process_info *ret)
1175{
1176 const struct svc_version *versp = NULL; /* compiler food */
1177 const struct svc_procedure *procp = NULL;
1178
1179 if (rqstp->rq_vers >= progp->pg_nvers )
1180 goto err_bad_vers;
1181 versp = progp->pg_vers[rqstp->rq_vers];
1182 if (!versp)
1183 goto err_bad_vers;
1184
1185 /*
1186 * Some protocol versions (namely NFSv4) require some form of
1187 * congestion control. (See RFC 7530 section 3.1 paragraph 2)
1188 * In other words, UDP is not allowed. We mark those when setting
1189 * up the svc_xprt, and verify that here.
1190 *
1191 * The spec is not very clear about what error should be returned
1192 * when someone tries to access a server that is listening on UDP
1193 * for lower versions. RPC_PROG_MISMATCH seems to be the closest
1194 * fit.
1195 */
1196 if (versp->vs_need_cong_ctrl && rqstp->rq_xprt &&
1197 !test_bit(XPT_CONG_CTRL, &rqstp->rq_xprt->xpt_flags))
1198 goto err_bad_vers;
1199
1200 if (rqstp->rq_proc >= versp->vs_nproc)
1201 goto err_bad_proc;
1202 rqstp->rq_procinfo = procp = &versp->vs_proc[rqstp->rq_proc];
1203 if (!procp)
1204 goto err_bad_proc;
1205
1206 /* Initialize storage for argp and resp */
1207 memset(rqstp->rq_argp, 0, procp->pc_argzero);
1208 memset(rqstp->rq_resp, 0, procp->pc_ressize);
1209
1210 /* Bump per-procedure stats counter */
1211 versp->vs_count[rqstp->rq_proc]++;
1212
1213 ret->dispatch = versp->vs_dispatch;
1214 return rpc_success;
1215err_bad_vers:
1216 ret->mismatch.lovers = progp->pg_lovers;
1217 ret->mismatch.hivers = progp->pg_hivers;
1218 return rpc_prog_mismatch;
1219err_bad_proc:
1220 return rpc_proc_unavail;
1221}
1222EXPORT_SYMBOL_GPL(svc_generic_init_request);
1223
1224/*
1225 * Common routine for processing the RPC request.
1226 */
1227static int
1228svc_process_common(struct svc_rqst *rqstp, struct kvec *argv, struct kvec *resv)
1229{
1230 struct svc_program *progp;
1231 const struct svc_procedure *procp = NULL;
1232 struct svc_serv *serv = rqstp->rq_server;
1233 struct svc_process_info process;
1234 __be32 *statp;
1235 u32 prog, vers;
1236 __be32 rpc_stat;
1237 int auth_res, rc;
1238 __be32 *reply_statp;
1239
1240 rpc_stat = rpc_success;
1241
1242 if (argv->iov_len < 6*4)
1243 goto err_short_len;
1244
1245 /* Will be turned off by GSS integrity and privacy services */
1246 set_bit(RQ_SPLICE_OK, &rqstp->rq_flags);
1247 /* Will be turned off only when NFSv4 Sessions are used */
1248 set_bit(RQ_USEDEFERRAL, &rqstp->rq_flags);
1249 clear_bit(RQ_DROPME, &rqstp->rq_flags);
1250
1251 svc_putu32(resv, rqstp->rq_xid);
1252
1253 vers = svc_getnl(argv);
1254
1255 /* First words of reply: */
1256 svc_putnl(resv, 1); /* REPLY */
1257
1258 if (vers != 2) /* RPC version number */
1259 goto err_bad_rpc;
1260
1261 /* Save position in case we later decide to reject: */
1262 reply_statp = resv->iov_base + resv->iov_len;
1263
1264 svc_putnl(resv, 0); /* ACCEPT */
1265
1266 rqstp->rq_prog = prog = svc_getnl(argv); /* program number */
1267 rqstp->rq_vers = svc_getnl(argv); /* version number */
1268 rqstp->rq_proc = svc_getnl(argv); /* procedure number */
1269
1270 for (progp = serv->sv_program; progp; progp = progp->pg_next)
1271 if (prog == progp->pg_prog)
1272 break;
1273
1274 /*
1275 * Decode auth data, and add verifier to reply buffer.
1276 * We do this before anything else in order to get a decent
1277 * auth verifier.
1278 */
1279 auth_res = svc_authenticate(rqstp);
1280 /* Also give the program a chance to reject this call: */
1281 if (auth_res == SVC_OK && progp)
1282 auth_res = progp->pg_authenticate(rqstp);
1283 trace_svc_authenticate(rqstp, auth_res);
1284 switch (auth_res) {
1285 case SVC_OK:
1286 break;
1287 case SVC_GARBAGE:
1288 goto err_garbage;
1289 case SVC_SYSERR:
1290 rpc_stat = rpc_system_err;
1291 goto err_bad;
1292 case SVC_DENIED:
1293 goto err_bad_auth;
1294 case SVC_CLOSE:
1295 goto close;
1296 case SVC_DROP:
1297 goto dropit;
1298 case SVC_COMPLETE:
1299 goto sendit;
1300 }
1301
1302 if (progp == NULL)
1303 goto err_bad_prog;
1304
1305 rpc_stat = progp->pg_init_request(rqstp, progp, &process);
1306 switch (rpc_stat) {
1307 case rpc_success:
1308 break;
1309 case rpc_prog_unavail:
1310 goto err_bad_prog;
1311 case rpc_prog_mismatch:
1312 goto err_bad_vers;
1313 case rpc_proc_unavail:
1314 goto err_bad_proc;
1315 }
1316
1317 procp = rqstp->rq_procinfo;
1318 /* Should this check go into the dispatcher? */
1319 if (!procp || !procp->pc_func)
1320 goto err_bad_proc;
1321
1322 /* Syntactic check complete */
1323 serv->sv_stats->rpccnt++;
1324 trace_svc_process(rqstp, progp->pg_name);
1325
1326 /* Build the reply header. */
1327 statp = resv->iov_base +resv->iov_len;
1328 svc_putnl(resv, RPC_SUCCESS);
1329
1330 /* un-reserve some of the out-queue now that we have a
1331 * better idea of reply size
1332 */
1333 if (procp->pc_xdrressize)
1334 svc_reserve_auth(rqstp, procp->pc_xdrressize<<2);
1335
1336 /* Call the function that processes the request. */
1337 rc = process.dispatch(rqstp, statp);
1338 if (procp->pc_release)
1339 procp->pc_release(rqstp);
1340 if (!rc)
1341 goto dropit;
1342 if (rqstp->rq_auth_stat != rpc_auth_ok)
1343 goto err_bad_auth;
1344
1345 /* Check RPC status result */
1346 if (*statp != rpc_success)
1347 resv->iov_len = ((void*)statp) - resv->iov_base + 4;
1348
1349 if (procp->pc_encode == NULL)
1350 goto dropit;
1351
1352 sendit:
1353 if (svc_authorise(rqstp))
1354 goto close_xprt;
1355 return 1; /* Caller can now send it */
1356
1357 dropit:
1358 svc_authorise(rqstp); /* doesn't hurt to call this twice */
1359 dprintk("svc: svc_process dropit\n");
1360 return 0;
1361
1362 close:
1363 svc_authorise(rqstp);
1364close_xprt:
1365 if (rqstp->rq_xprt && test_bit(XPT_TEMP, &rqstp->rq_xprt->xpt_flags))
1366 svc_xprt_close(rqstp->rq_xprt);
1367 dprintk("svc: svc_process close\n");
1368 return 0;
1369
1370err_short_len:
1371 svc_printk(rqstp, "short len %zd, dropping request\n",
1372 argv->iov_len);
1373 goto close_xprt;
1374
1375err_bad_rpc:
1376 serv->sv_stats->rpcbadfmt++;
1377 svc_putnl(resv, 1); /* REJECT */
1378 svc_putnl(resv, 0); /* RPC_MISMATCH */
1379 svc_putnl(resv, 2); /* Only RPCv2 supported */
1380 svc_putnl(resv, 2);
1381 goto sendit;
1382
1383err_bad_auth:
1384 dprintk("svc: authentication failed (%d)\n",
1385 be32_to_cpu(rqstp->rq_auth_stat));
1386 serv->sv_stats->rpcbadauth++;
1387 /* Restore write pointer to location of accept status: */
1388 xdr_ressize_check(rqstp, reply_statp);
1389 svc_putnl(resv, 1); /* REJECT */
1390 svc_putnl(resv, 1); /* AUTH_ERROR */
1391 svc_putu32(resv, rqstp->rq_auth_stat); /* status */
1392 goto sendit;
1393
1394err_bad_prog:
1395 dprintk("svc: unknown program %d\n", prog);
1396 serv->sv_stats->rpcbadfmt++;
1397 svc_putnl(resv, RPC_PROG_UNAVAIL);
1398 goto sendit;
1399
1400err_bad_vers:
1401 svc_printk(rqstp, "unknown version (%d for prog %d, %s)\n",
1402 rqstp->rq_vers, rqstp->rq_prog, progp->pg_name);
1403
1404 serv->sv_stats->rpcbadfmt++;
1405 svc_putnl(resv, RPC_PROG_MISMATCH);
1406 svc_putnl(resv, process.mismatch.lovers);
1407 svc_putnl(resv, process.mismatch.hivers);
1408 goto sendit;
1409
1410err_bad_proc:
1411 svc_printk(rqstp, "unknown procedure (%d)\n", rqstp->rq_proc);
1412
1413 serv->sv_stats->rpcbadfmt++;
1414 svc_putnl(resv, RPC_PROC_UNAVAIL);
1415 goto sendit;
1416
1417err_garbage:
1418 svc_printk(rqstp, "failed to decode args\n");
1419
1420 rpc_stat = rpc_garbage_args;
1421err_bad:
1422 serv->sv_stats->rpcbadfmt++;
1423 svc_putnl(resv, ntohl(rpc_stat));
1424 goto sendit;
1425}
1426
1427/*
1428 * Process the RPC request.
1429 */
1430int
1431svc_process(struct svc_rqst *rqstp)
1432{
1433 struct kvec *argv = &rqstp->rq_arg.head[0];
1434 struct kvec *resv = &rqstp->rq_res.head[0];
1435 __be32 dir;
1436
1437#if IS_ENABLED(CONFIG_FAIL_SUNRPC)
1438 if (!fail_sunrpc.ignore_server_disconnect &&
1439 should_fail(&fail_sunrpc.attr, 1))
1440 svc_xprt_deferred_close(rqstp->rq_xprt);
1441#endif
1442
1443 /*
1444 * Setup response xdr_buf.
1445 * Initially it has just one page
1446 */
1447 rqstp->rq_next_page = &rqstp->rq_respages[1];
1448 resv->iov_base = page_address(rqstp->rq_respages[0]);
1449 resv->iov_len = 0;
1450 rqstp->rq_res.pages = rqstp->rq_next_page;
1451 rqstp->rq_res.len = 0;
1452 rqstp->rq_res.page_base = 0;
1453 rqstp->rq_res.page_len = 0;
1454 rqstp->rq_res.buflen = PAGE_SIZE;
1455 rqstp->rq_res.tail[0].iov_base = NULL;
1456 rqstp->rq_res.tail[0].iov_len = 0;
1457
1458 dir = svc_getu32(argv);
1459 if (dir != rpc_call)
1460 goto out_baddir;
1461 if (!svc_process_common(rqstp, argv, resv))
1462 goto out_drop;
1463 return svc_send(rqstp);
1464
1465out_baddir:
1466 svc_printk(rqstp, "bad direction 0x%08x, dropping request\n",
1467 be32_to_cpu(dir));
1468 rqstp->rq_server->sv_stats->rpcbadfmt++;
1469out_drop:
1470 svc_drop(rqstp);
1471 return 0;
1472}
1473EXPORT_SYMBOL_GPL(svc_process);
1474
1475#if defined(CONFIG_SUNRPC_BACKCHANNEL)
1476/*
1477 * Process a backchannel RPC request that arrived over an existing
1478 * outbound connection
1479 */
1480int
1481bc_svc_process(struct svc_serv *serv, struct rpc_rqst *req,
1482 struct svc_rqst *rqstp)
1483{
1484 struct kvec *argv = &rqstp->rq_arg.head[0];
1485 struct kvec *resv = &rqstp->rq_res.head[0];
1486 struct rpc_task *task;
1487 int proc_error;
1488 int error;
1489
1490 dprintk("svc: %s(%p)\n", __func__, req);
1491
1492 /* Build the svc_rqst used by the common processing routine */
1493 rqstp->rq_xid = req->rq_xid;
1494 rqstp->rq_prot = req->rq_xprt->prot;
1495 rqstp->rq_server = serv;
1496 rqstp->rq_bc_net = req->rq_xprt->xprt_net;
1497
1498 rqstp->rq_addrlen = sizeof(req->rq_xprt->addr);
1499 memcpy(&rqstp->rq_addr, &req->rq_xprt->addr, rqstp->rq_addrlen);
1500 memcpy(&rqstp->rq_arg, &req->rq_rcv_buf, sizeof(rqstp->rq_arg));
1501 memcpy(&rqstp->rq_res, &req->rq_snd_buf, sizeof(rqstp->rq_res));
1502
1503 /* Adjust the argument buffer length */
1504 rqstp->rq_arg.len = req->rq_private_buf.len;
1505 if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len) {
1506 rqstp->rq_arg.head[0].iov_len = rqstp->rq_arg.len;
1507 rqstp->rq_arg.page_len = 0;
1508 } else if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len +
1509 rqstp->rq_arg.page_len)
1510 rqstp->rq_arg.page_len = rqstp->rq_arg.len -
1511 rqstp->rq_arg.head[0].iov_len;
1512 else
1513 rqstp->rq_arg.len = rqstp->rq_arg.head[0].iov_len +
1514 rqstp->rq_arg.page_len;
1515
1516 /* reset result send buffer "put" position */
1517 resv->iov_len = 0;
1518
1519 /*
1520 * Skip the next two words because they've already been
1521 * processed in the transport
1522 */
1523 svc_getu32(argv); /* XID */
1524 svc_getnl(argv); /* CALLDIR */
1525
1526 /* Parse and execute the bc call */
1527 proc_error = svc_process_common(rqstp, argv, resv);
1528
1529 atomic_dec(&req->rq_xprt->bc_slot_count);
1530 if (!proc_error) {
1531 /* Processing error: drop the request */
1532 xprt_free_bc_request(req);
1533 error = -EINVAL;
1534 goto out;
1535 }
1536 /* Finally, send the reply synchronously */
1537 memcpy(&req->rq_snd_buf, &rqstp->rq_res, sizeof(req->rq_snd_buf));
1538 task = rpc_run_bc_task(req);
1539 if (IS_ERR(task)) {
1540 error = PTR_ERR(task);
1541 goto out;
1542 }
1543
1544 WARN_ON_ONCE(atomic_read(&task->tk_count) != 1);
1545 error = task->tk_status;
1546 rpc_put_task(task);
1547
1548out:
1549 dprintk("svc: %s(), error=%d\n", __func__, error);
1550 return error;
1551}
1552EXPORT_SYMBOL_GPL(bc_svc_process);
1553#endif /* CONFIG_SUNRPC_BACKCHANNEL */
1554
1555/**
1556 * svc_max_payload - Return transport-specific limit on the RPC payload
1557 * @rqstp: RPC transaction context
1558 *
1559 * Returns the maximum number of payload bytes the current transport
1560 * allows.
1561 */
1562u32 svc_max_payload(const struct svc_rqst *rqstp)
1563{
1564 u32 max = rqstp->rq_xprt->xpt_class->xcl_max_payload;
1565
1566 if (rqstp->rq_server->sv_max_payload < max)
1567 max = rqstp->rq_server->sv_max_payload;
1568 return max;
1569}
1570EXPORT_SYMBOL_GPL(svc_max_payload);
1571
1572/**
1573 * svc_proc_name - Return RPC procedure name in string form
1574 * @rqstp: svc_rqst to operate on
1575 *
1576 * Return value:
1577 * Pointer to a NUL-terminated string
1578 */
1579const char *svc_proc_name(const struct svc_rqst *rqstp)
1580{
1581 if (rqstp && rqstp->rq_procinfo)
1582 return rqstp->rq_procinfo->pc_name;
1583 return "unknown";
1584}
1585
1586
1587/**
1588 * svc_encode_result_payload - mark a range of bytes as a result payload
1589 * @rqstp: svc_rqst to operate on
1590 * @offset: payload's byte offset in rqstp->rq_res
1591 * @length: size of payload, in bytes
1592 *
1593 * Returns zero on success, or a negative errno if a permanent
1594 * error occurred.
1595 */
1596int svc_encode_result_payload(struct svc_rqst *rqstp, unsigned int offset,
1597 unsigned int length)
1598{
1599 return rqstp->rq_xprt->xpt_ops->xpo_result_payload(rqstp, offset,
1600 length);
1601}
1602EXPORT_SYMBOL_GPL(svc_encode_result_payload);
1603
1604/**
1605 * svc_fill_write_vector - Construct data argument for VFS write call
1606 * @rqstp: svc_rqst to operate on
1607 * @payload: xdr_buf containing only the write data payload
1608 *
1609 * Fills in rqstp::rq_vec, and returns the number of elements.
1610 */
1611unsigned int svc_fill_write_vector(struct svc_rqst *rqstp,
1612 struct xdr_buf *payload)
1613{
1614 struct page **pages = payload->pages;
1615 struct kvec *first = payload->head;
1616 struct kvec *vec = rqstp->rq_vec;
1617 size_t total = payload->len;
1618 unsigned int i;
1619
1620 /* Some types of transport can present the write payload
1621 * entirely in rq_arg.pages. In this case, @first is empty.
1622 */
1623 i = 0;
1624 if (first->iov_len) {
1625 vec[i].iov_base = first->iov_base;
1626 vec[i].iov_len = min_t(size_t, total, first->iov_len);
1627 total -= vec[i].iov_len;
1628 ++i;
1629 }
1630
1631 while (total) {
1632 vec[i].iov_base = page_address(*pages);
1633 vec[i].iov_len = min_t(size_t, total, PAGE_SIZE);
1634 total -= vec[i].iov_len;
1635 ++i;
1636 ++pages;
1637 }
1638
1639 WARN_ON_ONCE(i > ARRAY_SIZE(rqstp->rq_vec));
1640 return i;
1641}
1642EXPORT_SYMBOL_GPL(svc_fill_write_vector);
1643
1644/**
1645 * svc_fill_symlink_pathname - Construct pathname argument for VFS symlink call
1646 * @rqstp: svc_rqst to operate on
1647 * @first: buffer containing first section of pathname
1648 * @p: buffer containing remaining section of pathname
1649 * @total: total length of the pathname argument
1650 *
1651 * The VFS symlink API demands a NUL-terminated pathname in mapped memory.
1652 * Returns pointer to a NUL-terminated string, or an ERR_PTR. Caller must free
1653 * the returned string.
1654 */
1655char *svc_fill_symlink_pathname(struct svc_rqst *rqstp, struct kvec *first,
1656 void *p, size_t total)
1657{
1658 size_t len, remaining;
1659 char *result, *dst;
1660
1661 result = kmalloc(total + 1, GFP_KERNEL);
1662 if (!result)
1663 return ERR_PTR(-ESERVERFAULT);
1664
1665 dst = result;
1666 remaining = total;
1667
1668 len = min_t(size_t, total, first->iov_len);
1669 if (len) {
1670 memcpy(dst, first->iov_base, len);
1671 dst += len;
1672 remaining -= len;
1673 }
1674
1675 if (remaining) {
1676 len = min_t(size_t, remaining, PAGE_SIZE);
1677 memcpy(dst, p, len);
1678 dst += len;
1679 }
1680
1681 *dst = '\0';
1682
1683 /* Sanity check: Linux doesn't allow the pathname argument to
1684 * contain a NUL byte.
1685 */
1686 if (strlen(result) != total) {
1687 kfree(result);
1688 return ERR_PTR(-EINVAL);
1689 }
1690 return result;
1691}
1692EXPORT_SYMBOL_GPL(svc_fill_symlink_pathname);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/net/sunrpc/svc.c
4 *
5 * High-level RPC service routines
6 *
7 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
8 *
9 * Multiple threads pools and NUMAisation
10 * Copyright (c) 2006 Silicon Graphics, Inc.
11 * by Greg Banks <gnb@melbourne.sgi.com>
12 */
13
14#include <linux/linkage.h>
15#include <linux/sched/signal.h>
16#include <linux/errno.h>
17#include <linux/net.h>
18#include <linux/in.h>
19#include <linux/mm.h>
20#include <linux/interrupt.h>
21#include <linux/module.h>
22#include <linux/kthread.h>
23#include <linux/slab.h>
24
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>
31
32#include <trace/events/sunrpc.h>
33
34#include "fail.h"
35
36#define RPCDBG_FACILITY RPCDBG_SVCDSP
37
38static void svc_unregister(const struct svc_serv *serv, struct net *net);
39
40#define SVC_POOL_DEFAULT SVC_POOL_GLOBAL
41
42/*
43 * Mode for mapping cpus to pools.
44 */
45enum {
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 */
51};
52
53/*
54 * Structure for mapping cpus to pools and vice versa.
55 * Setup once during sunrpc initialisation.
56 */
57
58struct svc_pool_map {
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" */
63 unsigned int npools;
64 unsigned int *pool_to; /* maps pool id to cpu or node */
65 unsigned int *to_pool; /* maps cpu or node to pool id */
66};
67
68static struct svc_pool_map svc_pool_map = {
69 .mode = SVC_POOL_DEFAULT
70};
71
72static DEFINE_MUTEX(svc_pool_map_mutex);/* protects svc_pool_map.count only */
73
74static int
75param_set_pool_mode(const char *val, const struct kernel_param *kp)
76{
77 int *ip = (int *)kp->arg;
78 struct svc_pool_map *m = &svc_pool_map;
79 int err;
80
81 mutex_lock(&svc_pool_map_mutex);
82
83 err = -EBUSY;
84 if (m->count)
85 goto out;
86
87 err = 0;
88 if (!strncmp(val, "auto", 4))
89 *ip = SVC_POOL_AUTO;
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;
96 else
97 err = -EINVAL;
98
99out:
100 mutex_unlock(&svc_pool_map_mutex);
101 return err;
102}
103
104static int
105param_get_pool_mode(char *buf, const struct kernel_param *kp)
106{
107 int *ip = (int *)kp->arg;
108
109 switch (*ip)
110 {
111 case SVC_POOL_AUTO:
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");
119 default:
120 return sysfs_emit(buf, "%d\n", *ip);
121 }
122}
123
124module_param_call(pool_mode, param_set_pool_mode, param_get_pool_mode,
125 &svc_pool_map.mode, 0644);
126
127/*
128 * Detect best pool mapping mode heuristically,
129 * according to the machine's topology.
130 */
131static int
132svc_pool_map_choose_mode(void)
133{
134 unsigned int node;
135
136 if (nr_online_nodes > 1) {
137 /*
138 * Actually have multiple NUMA nodes,
139 * so split pools on NUMA node boundaries
140 */
141 return SVC_POOL_PERNODE;
142 }
143
144 node = first_online_node;
145 if (nr_cpus_node(node) > 2) {
146 /*
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.
151 */
152 return SVC_POOL_PERCPU;
153 }
154
155 /* default: one global pool */
156 return SVC_POOL_GLOBAL;
157}
158
159/*
160 * Allocate the to_pool[] and pool_to[] arrays.
161 * Returns 0 on success or an errno.
162 */
163static int
164svc_pool_map_alloc_arrays(struct svc_pool_map *m, unsigned int maxpools)
165{
166 m->to_pool = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
167 if (!m->to_pool)
168 goto fail;
169 m->pool_to = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
170 if (!m->pool_to)
171 goto fail_free;
172
173 return 0;
174
175fail_free:
176 kfree(m->to_pool);
177 m->to_pool = NULL;
178fail:
179 return -ENOMEM;
180}
181
182/*
183 * Initialise the pool map for SVC_POOL_PERCPU mode.
184 * Returns number of pools or <0 on error.
185 */
186static int
187svc_pool_map_init_percpu(struct svc_pool_map *m)
188{
189 unsigned int maxpools = nr_cpu_ids;
190 unsigned int pidx = 0;
191 unsigned int cpu;
192 int err;
193
194 err = svc_pool_map_alloc_arrays(m, maxpools);
195 if (err)
196 return err;
197
198 for_each_online_cpu(cpu) {
199 BUG_ON(pidx >= maxpools);
200 m->to_pool[cpu] = pidx;
201 m->pool_to[pidx] = cpu;
202 pidx++;
203 }
204 /* cpus brought online later all get mapped to pool0, sorry */
205
206 return pidx;
207};
208
209
210/*
211 * Initialise the pool map for SVC_POOL_PERNODE mode.
212 * Returns number of pools or <0 on error.
213 */
214static int
215svc_pool_map_init_pernode(struct svc_pool_map *m)
216{
217 unsigned int maxpools = nr_node_ids;
218 unsigned int pidx = 0;
219 unsigned int node;
220 int err;
221
222 err = svc_pool_map_alloc_arrays(m, maxpools);
223 if (err)
224 return err;
225
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;
231 pidx++;
232 }
233 /* nodes brought online later all get mapped to pool0, sorry */
234
235 return pidx;
236}
237
238
239/*
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
244 * was taken.
245 */
246static unsigned int
247svc_pool_map_get(void)
248{
249 struct svc_pool_map *m = &svc_pool_map;
250 int npools = -1;
251
252 mutex_lock(&svc_pool_map_mutex);
253
254 if (m->count++) {
255 mutex_unlock(&svc_pool_map_mutex);
256 WARN_ON_ONCE(m->npools <= 1);
257 return m->npools;
258 }
259
260 if (m->mode == SVC_POOL_AUTO)
261 m->mode = svc_pool_map_choose_mode();
262
263 switch (m->mode) {
264 case SVC_POOL_PERCPU:
265 npools = svc_pool_map_init_percpu(m);
266 break;
267 case SVC_POOL_PERNODE:
268 npools = svc_pool_map_init_pernode(m);
269 break;
270 }
271
272 if (npools <= 0) {
273 /* default, or memory allocation failure */
274 npools = 1;
275 m->mode = SVC_POOL_GLOBAL;
276 }
277 m->npools = npools;
278
279 if (npools == 1)
280 /* service is unpooled, so doesn't hold a reference */
281 m->count--;
282
283 mutex_unlock(&svc_pool_map_mutex);
284 return npools;
285}
286
287/*
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.
294 */
295static void
296svc_pool_map_put(int npools)
297{
298 struct svc_pool_map *m = &svc_pool_map;
299
300 if (npools <= 1)
301 return;
302 mutex_lock(&svc_pool_map_mutex);
303
304 if (!--m->count) {
305 kfree(m->to_pool);
306 m->to_pool = NULL;
307 kfree(m->pool_to);
308 m->pool_to = NULL;
309 m->npools = 0;
310 }
311
312 mutex_unlock(&svc_pool_map_mutex);
313}
314
315static int svc_pool_map_get_node(unsigned int pidx)
316{
317 const struct svc_pool_map *m = &svc_pool_map;
318
319 if (m->count) {
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];
324 }
325 return NUMA_NO_NODE;
326}
327/*
328 * Set the given thread's cpus_allowed mask so that it
329 * will only run on cpus in the given pool.
330 */
331static inline void
332svc_pool_map_set_cpumask(struct task_struct *task, unsigned int pidx)
333{
334 struct svc_pool_map *m = &svc_pool_map;
335 unsigned int node = m->pool_to[pidx];
336
337 /*
338 * The caller checks for sv_nrpools > 1, which
339 * implies that we've been initialized.
340 */
341 WARN_ON_ONCE(m->count == 0);
342 if (m->count == 0)
343 return;
344
345 switch (m->mode) {
346 case SVC_POOL_PERCPU:
347 {
348 set_cpus_allowed_ptr(task, cpumask_of(node));
349 break;
350 }
351 case SVC_POOL_PERNODE:
352 {
353 set_cpus_allowed_ptr(task, cpumask_of_node(node));
354 break;
355 }
356 }
357}
358
359/**
360 * svc_pool_for_cpu - Select pool to run a thread on this cpu
361 * @serv: An RPC service
362 *
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.
366 *
367 * Return value:
368 * A pointer to an svc_pool
369 */
370struct svc_pool *svc_pool_for_cpu(struct svc_serv *serv)
371{
372 struct svc_pool_map *m = &svc_pool_map;
373 int cpu = raw_smp_processor_id();
374 unsigned int pidx = 0;
375
376 if (serv->sv_nrpools <= 1)
377 return serv->sv_pools;
378
379 switch (m->mode) {
380 case SVC_POOL_PERCPU:
381 pidx = m->to_pool[cpu];
382 break;
383 case SVC_POOL_PERNODE:
384 pidx = m->to_pool[cpu_to_node(cpu)];
385 break;
386 }
387
388 return &serv->sv_pools[pidx % serv->sv_nrpools];
389}
390
391int svc_rpcb_setup(struct svc_serv *serv, struct net *net)
392{
393 int err;
394
395 err = rpcb_create_local(net);
396 if (err)
397 return err;
398
399 /* Remove any stale portmap registrations */
400 svc_unregister(serv, net);
401 return 0;
402}
403EXPORT_SYMBOL_GPL(svc_rpcb_setup);
404
405void svc_rpcb_cleanup(struct svc_serv *serv, struct net *net)
406{
407 svc_unregister(serv, net);
408 rpcb_put_local(net);
409}
410EXPORT_SYMBOL_GPL(svc_rpcb_cleanup);
411
412static int svc_uses_rpcbind(struct svc_serv *serv)
413{
414 struct svc_program *progp;
415 unsigned int i;
416
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)
420 continue;
421 if (!progp->pg_vers[i]->vs_hidden)
422 return 1;
423 }
424 }
425
426 return 0;
427}
428
429int svc_bind(struct svc_serv *serv, struct net *net)
430{
431 if (!svc_uses_rpcbind(serv))
432 return 0;
433 return svc_rpcb_setup(serv, net);
434}
435EXPORT_SYMBOL_GPL(svc_bind);
436
437#if defined(CONFIG_SUNRPC_BACKCHANNEL)
438static void
439__svc_init_bc(struct svc_serv *serv)
440{
441 lwq_init(&serv->sv_cb_list);
442}
443#else
444static void
445__svc_init_bc(struct svc_serv *serv)
446{
447}
448#endif
449
450/*
451 * Create an RPC service
452 */
453static struct svc_serv *
454__svc_create(struct svc_program *prog, struct svc_stat *stats,
455 unsigned int bufsize, int npools, int (*threadfn)(void *data))
456{
457 struct svc_serv *serv;
458 unsigned int vers;
459 unsigned int xdrsize;
460 unsigned int i;
461
462 if (!(serv = kzalloc(sizeof(*serv), GFP_KERNEL)))
463 return NULL;
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;
472 xdrsize = 0;
473 while (prog) {
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;
482 }
483 prog = prog->pg_next;
484 }
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);
490
491 __svc_init_bc(serv);
492
493 serv->sv_nrpools = npools;
494 serv->sv_pools =
495 kcalloc(serv->sv_nrpools, sizeof(struct svc_pool),
496 GFP_KERNEL);
497 if (!serv->sv_pools) {
498 kfree(serv);
499 return NULL;
500 }
501
502 for (i = 0; i < serv->sv_nrpools; i++) {
503 struct svc_pool *pool = &serv->sv_pools[i];
504
505 dprintk("svc: initialising pool %u for %s\n",
506 i, serv->sv_name);
507
508 pool->sp_id = i;
509 lwq_init(&pool->sp_xprts);
510 INIT_LIST_HEAD(&pool->sp_all_threads);
511 init_llist_head(&pool->sp_idle_threads);
512
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);
516 }
517
518 return serv;
519}
520
521/**
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
526 *
527 * Returns an instantiated struct svc_serv object or NULL.
528 */
529struct svc_serv *svc_create(struct svc_program *prog, unsigned int bufsize,
530 int (*threadfn)(void *data))
531{
532 return __svc_create(prog, NULL, bufsize, 1, threadfn);
533}
534EXPORT_SYMBOL_GPL(svc_create);
535
536/**
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
542 *
543 * Returns an instantiated struct svc_serv object or NULL.
544 */
545struct svc_serv *svc_create_pooled(struct svc_program *prog,
546 struct svc_stat *stats,
547 unsigned int bufsize,
548 int (*threadfn)(void *data))
549{
550 struct svc_serv *serv;
551 unsigned int npools = svc_pool_map_get();
552
553 serv = __svc_create(prog, stats, bufsize, npools, threadfn);
554 if (!serv)
555 goto out_err;
556 return serv;
557out_err:
558 svc_pool_map_put(npools);
559 return NULL;
560}
561EXPORT_SYMBOL_GPL(svc_create_pooled);
562
563/*
564 * Destroy an RPC service. Should be called with appropriate locking to
565 * protect sv_permsocks and sv_tempsocks.
566 */
567void
568svc_destroy(struct svc_serv **servp)
569{
570 struct svc_serv *serv = *servp;
571 unsigned int i;
572
573 *servp = NULL;
574
575 dprintk("svc: svc_destroy(%s)\n", serv->sv_program->pg_name);
576 timer_shutdown_sync(&serv->sv_temptimer);
577
578 /*
579 * Remaining transports at this point are not expected.
580 */
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);
585
586 cache_clean_deferred(serv);
587
588 svc_pool_map_put(serv->sv_nrpools);
589
590 for (i = 0; i < serv->sv_nrpools; i++) {
591 struct svc_pool *pool = &serv->sv_pools[i];
592
593 percpu_counter_destroy(&pool->sp_messages_arrived);
594 percpu_counter_destroy(&pool->sp_sockets_queued);
595 percpu_counter_destroy(&pool->sp_threads_woken);
596 }
597 kfree(serv->sv_pools);
598 kfree(serv);
599}
600EXPORT_SYMBOL_GPL(svc_destroy);
601
602static bool
603svc_init_buffer(struct svc_rqst *rqstp, unsigned int size, int node)
604{
605 unsigned long pages, ret;
606
607 /* bc_xprt uses fore channel allocated buffers */
608 if (svc_is_backchannel(rqstp))
609 return true;
610
611 pages = size / PAGE_SIZE + 1; /* extra page as we hold both request and reply.
612 * We assume one is at most one page
613 */
614 WARN_ON_ONCE(pages > RPCSVC_MAXPAGES);
615 if (pages > RPCSVC_MAXPAGES)
616 pages = RPCSVC_MAXPAGES;
617
618 ret = alloc_pages_bulk_array_node(GFP_KERNEL, node, pages,
619 rqstp->rq_pages);
620 return ret == pages;
621}
622
623/*
624 * Release an RPC server buffer
625 */
626static void
627svc_release_buffer(struct svc_rqst *rqstp)
628{
629 unsigned int i;
630
631 for (i = 0; i < ARRAY_SIZE(rqstp->rq_pages); i++)
632 if (rqstp->rq_pages[i])
633 put_page(rqstp->rq_pages[i]);
634}
635
636struct svc_rqst *
637svc_rqst_alloc(struct svc_serv *serv, struct svc_pool *pool, int node)
638{
639 struct svc_rqst *rqstp;
640
641 rqstp = kzalloc_node(sizeof(*rqstp), GFP_KERNEL, node);
642 if (!rqstp)
643 return rqstp;
644
645 folio_batch_init(&rqstp->rq_fbatch);
646
647 rqstp->rq_server = serv;
648 rqstp->rq_pool = pool;
649
650 rqstp->rq_scratch_page = alloc_pages_node(node, GFP_KERNEL, 0);
651 if (!rqstp->rq_scratch_page)
652 goto out_enomem;
653
654 rqstp->rq_argp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
655 if (!rqstp->rq_argp)
656 goto out_enomem;
657
658 rqstp->rq_resp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
659 if (!rqstp->rq_resp)
660 goto out_enomem;
661
662 if (!svc_init_buffer(rqstp, serv->sv_max_mesg, node))
663 goto out_enomem;
664
665 return rqstp;
666out_enomem:
667 svc_rqst_free(rqstp);
668 return NULL;
669}
670EXPORT_SYMBOL_GPL(svc_rqst_alloc);
671
672static struct svc_rqst *
673svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
674{
675 struct svc_rqst *rqstp;
676
677 rqstp = svc_rqst_alloc(serv, pool, node);
678 if (!rqstp)
679 return ERR_PTR(-ENOMEM);
680
681 spin_lock_bh(&serv->sv_lock);
682 serv->sv_nrthreads += 1;
683 spin_unlock_bh(&serv->sv_lock);
684
685 atomic_inc(&pool->sp_nrthreads);
686
687 /* Protected by whatever lock the service uses when calling
688 * svc_set_num_threads()
689 */
690 list_add_rcu(&rqstp->rq_all, &pool->sp_all_threads);
691
692 return rqstp;
693}
694
695/**
696 * svc_pool_wake_idle_thread - Awaken an idle thread in @pool
697 * @pool: service thread pool
698 *
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().
702 *
703 */
704void svc_pool_wake_idle_thread(struct svc_pool *pool)
705{
706 struct svc_rqst *rqstp;
707 struct llist_node *ln;
708
709 rcu_read_lock();
710 ln = READ_ONCE(pool->sp_idle_threads.first);
711 if (ln) {
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);
718 }
719 rcu_read_unlock();
720 return;
721 }
722 rcu_read_unlock();
723
724}
725EXPORT_SYMBOL_GPL(svc_pool_wake_idle_thread);
726
727static struct svc_pool *
728svc_pool_next(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
729{
730 return pool ? pool : &serv->sv_pools[(*state)++ % serv->sv_nrpools];
731}
732
733static struct svc_pool *
734svc_pool_victim(struct svc_serv *serv, struct svc_pool *target_pool,
735 unsigned int *state)
736{
737 struct svc_pool *pool;
738 unsigned int i;
739
740retry:
741 pool = target_pool;
742
743 if (pool != NULL) {
744 if (atomic_inc_not_zero(&pool->sp_nrthreads))
745 goto found_pool;
746 return NULL;
747 } else {
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))
751 goto found_pool;
752 }
753 return NULL;
754 }
755
756found_pool:
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))
760 return pool;
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);
764 goto retry;
765}
766
767static int
768svc_start_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
769{
770 struct svc_rqst *rqstp;
771 struct task_struct *task;
772 struct svc_pool *chosen_pool;
773 unsigned int state = serv->sv_nrthreads-1;
774 int node;
775
776 do {
777 nrservs--;
778 chosen_pool = svc_pool_next(serv, pool, &state);
779 node = svc_pool_map_get_node(chosen_pool->sp_id);
780
781 rqstp = svc_prepare_thread(serv, chosen_pool, node);
782 if (IS_ERR(rqstp))
783 return PTR_ERR(rqstp);
784 task = kthread_create_on_node(serv->sv_threadfn, rqstp,
785 node, "%s", serv->sv_name);
786 if (IS_ERR(task)) {
787 svc_exit_thread(rqstp);
788 return PTR_ERR(task);
789 }
790
791 rqstp->rq_task = task;
792 if (serv->sv_nrpools > 1)
793 svc_pool_map_set_cpumask(task, chosen_pool->sp_id);
794
795 svc_sock_update_bufs(serv);
796 wake_up_process(task);
797 } while (nrservs > 0);
798
799 return 0;
800}
801
802static int
803svc_stop_kthreads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
804{
805 unsigned int state = serv->sv_nrthreads-1;
806 struct svc_pool *victim;
807
808 do {
809 victim = svc_pool_victim(serv, pool, &state);
810 if (!victim)
811 break;
812 svc_pool_wake_idle_thread(victim);
813 wait_on_bit(&victim->sp_flags, SP_VICTIM_REMAINS,
814 TASK_IDLE);
815 nrservs++;
816 } while (nrservs < 0);
817 return 0;
818}
819
820/**
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)
825 *
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.
830 *
831 * Caller must ensure mutual exclusion between this and server startup or
832 * shutdown.
833 *
834 * Returns zero on success or a negative errno if an error occurred while
835 * starting a thread.
836 */
837int
838svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
839{
840 if (!pool)
841 nrservs -= serv->sv_nrthreads;
842 else
843 nrservs -= atomic_read(&pool->sp_nrthreads);
844
845 if (nrservs > 0)
846 return svc_start_kthreads(serv, pool, nrservs);
847 if (nrservs < 0)
848 return svc_stop_kthreads(serv, pool, nrservs);
849 return 0;
850}
851EXPORT_SYMBOL_GPL(svc_set_num_threads);
852
853/**
854 * svc_rqst_replace_page - Replace one page in rq_pages[]
855 * @rqstp: svc_rqst with pages to replace
856 * @page: replacement page
857 *
858 * When replacing a page in rq_pages, batch the release of the
859 * replaced pages to avoid hammering the page allocator.
860 *
861 * Return values:
862 * %true: page replaced
863 * %false: array bounds checking failed
864 */
865bool svc_rqst_replace_page(struct svc_rqst *rqstp, struct page *page)
866{
867 struct page **begin = rqstp->rq_pages;
868 struct page **end = &rqstp->rq_pages[RPCSVC_MAXPAGES];
869
870 if (unlikely(rqstp->rq_next_page < begin || rqstp->rq_next_page > end)) {
871 trace_svc_replace_page_err(rqstp);
872 return false;
873 }
874
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);
879 }
880
881 get_page(page);
882 *(rqstp->rq_next_page++) = page;
883 return true;
884}
885EXPORT_SYMBOL_GPL(svc_rqst_replace_page);
886
887/**
888 * svc_rqst_release_pages - Release Reply buffer pages
889 * @rqstp: RPC transaction context
890 *
891 * Release response pages that might still be in flight after
892 * svc_send, and any spliced filesystem-owned pages.
893 */
894void svc_rqst_release_pages(struct svc_rqst *rqstp)
895{
896 int i, count = rqstp->rq_next_page - rqstp->rq_respages;
897
898 if (count) {
899 release_pages(rqstp->rq_respages, count);
900 for (i = 0; i < count; i++)
901 rqstp->rq_respages[i] = NULL;
902 }
903}
904
905/*
906 * Called from a server thread as it's exiting. Caller must hold the "service
907 * mutex" for the service.
908 */
909void
910svc_rqst_free(struct svc_rqst *rqstp)
911{
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);
920}
921EXPORT_SYMBOL_GPL(svc_rqst_free);
922
923void
924svc_exit_thread(struct svc_rqst *rqstp)
925{
926 struct svc_serv *serv = rqstp->rq_server;
927 struct svc_pool *pool = rqstp->rq_pool;
928
929 list_del_rcu(&rqstp->rq_all);
930
931 atomic_dec(&pool->sp_nrthreads);
932
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);
937
938 svc_rqst_free(rqstp);
939
940 clear_and_wake_up_bit(SP_VICTIM_REMAINS, &pool->sp_flags);
941}
942EXPORT_SYMBOL_GPL(svc_exit_thread);
943
944/*
945 * Register an "inet" protocol family netid with the local
946 * rpcbind daemon via an rpcbind v4 SET request.
947 *
948 * No netconfig infrastructure is available in the kernel, so
949 * we map IP_ protocol numbers to netids by hand.
950 *
951 * Returns zero on success; a negative errno value is returned
952 * if any error occurs.
953 */
954static int __svc_rpcb_register4(struct net *net, const u32 program,
955 const u32 version,
956 const unsigned short protocol,
957 const unsigned short port)
958{
959 const struct sockaddr_in sin = {
960 .sin_family = AF_INET,
961 .sin_addr.s_addr = htonl(INADDR_ANY),
962 .sin_port = htons(port),
963 };
964 const char *netid;
965 int error;
966
967 switch (protocol) {
968 case IPPROTO_UDP:
969 netid = RPCBIND_NETID_UDP;
970 break;
971 case IPPROTO_TCP:
972 netid = RPCBIND_NETID_TCP;
973 break;
974 default:
975 return -ENOPROTOOPT;
976 }
977
978 error = rpcb_v4_register(net, program, version,
979 (const struct sockaddr *)&sin, netid);
980
981 /*
982 * User space didn't support rpcbind v4, so retry this
983 * registration request with the legacy rpcbind v2 protocol.
984 */
985 if (error == -EPROTONOSUPPORT)
986 error = rpcb_register(net, program, version, protocol, port);
987
988 return error;
989}
990
991#if IS_ENABLED(CONFIG_IPV6)
992/*
993 * Register an "inet6" protocol family netid with the local
994 * rpcbind daemon via an rpcbind v4 SET request.
995 *
996 * No netconfig infrastructure is available in the kernel, so
997 * we map IP_ protocol numbers to netids by hand.
998 *
999 * Returns zero on success; a negative errno value is returned
1000 * if any error occurs.
1001 */
1002static int __svc_rpcb_register6(struct net *net, const u32 program,
1003 const u32 version,
1004 const unsigned short protocol,
1005 const unsigned short port)
1006{
1007 const struct sockaddr_in6 sin6 = {
1008 .sin6_family = AF_INET6,
1009 .sin6_addr = IN6ADDR_ANY_INIT,
1010 .sin6_port = htons(port),
1011 };
1012 const char *netid;
1013 int error;
1014
1015 switch (protocol) {
1016 case IPPROTO_UDP:
1017 netid = RPCBIND_NETID_UDP6;
1018 break;
1019 case IPPROTO_TCP:
1020 netid = RPCBIND_NETID_TCP6;
1021 break;
1022 default:
1023 return -ENOPROTOOPT;
1024 }
1025
1026 error = rpcb_v4_register(net, program, version,
1027 (const struct sockaddr *)&sin6, netid);
1028
1029 /*
1030 * User space didn't support rpcbind version 4, so we won't
1031 * use a PF_INET6 listener.
1032 */
1033 if (error == -EPROTONOSUPPORT)
1034 error = -EAFNOSUPPORT;
1035
1036 return error;
1037}
1038#endif /* IS_ENABLED(CONFIG_IPV6) */
1039
1040/*
1041 * Register a kernel RPC service via rpcbind version 4.
1042 *
1043 * Returns zero on success; a negative errno value is returned
1044 * if any error occurs.
1045 */
1046static int __svc_register(struct net *net, const char *progname,
1047 const u32 program, const u32 version,
1048 const int family,
1049 const unsigned short protocol,
1050 const unsigned short port)
1051{
1052 int error = -EAFNOSUPPORT;
1053
1054 switch (family) {
1055 case PF_INET:
1056 error = __svc_rpcb_register4(net, program, version,
1057 protocol, port);
1058 break;
1059#if IS_ENABLED(CONFIG_IPV6)
1060 case PF_INET6:
1061 error = __svc_rpcb_register6(net, program, version,
1062 protocol, port);
1063#endif
1064 }
1065
1066 trace_svc_register(progname, version, family, protocol, port, error);
1067 return error;
1068}
1069
1070int 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)
1075{
1076 return __svc_register(net, progp->pg_name, progp->pg_prog,
1077 version, family, proto, port);
1078
1079}
1080EXPORT_SYMBOL_GPL(svc_rpcbind_set_version);
1081
1082int 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)
1087{
1088 const struct svc_version *vers = progp->pg_vers[version];
1089 int error;
1090
1091 if (vers == NULL)
1092 return 0;
1093
1094 if (vers->vs_hidden) {
1095 trace_svc_noregister(progp->pg_name, version, proto,
1096 port, family, 0);
1097 return 0;
1098 }
1099
1100 /*
1101 * Don't register a UDP port if we need congestion
1102 * control.
1103 */
1104 if (vers->vs_need_cong_ctrl && proto == IPPROTO_UDP)
1105 return 0;
1106
1107 error = svc_rpcbind_set_version(net, progp, version,
1108 family, proto, port);
1109
1110 return (vers->vs_rpcb_optnl) ? 0 : error;
1111}
1112EXPORT_SYMBOL_GPL(svc_generic_rpcbind_set);
1113
1114/**
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
1121 *
1122 * Service is registered for any address in the passed-in protocol family
1123 */
1124int svc_register(const struct svc_serv *serv, struct net *net,
1125 const int family, const unsigned short proto,
1126 const unsigned short port)
1127{
1128 struct svc_program *progp;
1129 unsigned int i;
1130 int error = 0;
1131
1132 WARN_ON_ONCE(proto == 0 && port == 0);
1133 if (proto == 0 && port == 0)
1134 return -EINVAL;
1135
1136 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1137 for (i = 0; i < progp->pg_nvers; i++) {
1138
1139 error = progp->pg_rpcbind_set(net, progp, i,
1140 family, proto, port);
1141 if (error < 0) {
1142 printk(KERN_WARNING "svc: failed to register "
1143 "%sv%u RPC service (errno %d).\n",
1144 progp->pg_name, i, -error);
1145 break;
1146 }
1147 }
1148 }
1149
1150 return error;
1151}
1152
1153/*
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].
1159 */
1160static void __svc_unregister(struct net *net, const u32 program, const u32 version,
1161 const char *progname)
1162{
1163 int error;
1164
1165 error = rpcb_v4_register(net, program, version, NULL, "");
1166
1167 /*
1168 * User space didn't support rpcbind v4, so retry this
1169 * request with the legacy rpcbind v2 protocol.
1170 */
1171 if (error == -EPROTONOSUPPORT)
1172 error = rpcb_register(net, program, version, 0, 0);
1173
1174 trace_svc_unregister(progname, version, error);
1175}
1176
1177/*
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.
1181 *
1182 * The result of unregistration is reported via dprintk for those who want
1183 * verification of the result, but is otherwise not important.
1184 */
1185static void svc_unregister(const struct svc_serv *serv, struct net *net)
1186{
1187 struct sighand_struct *sighand;
1188 struct svc_program *progp;
1189 unsigned long flags;
1190 unsigned int i;
1191
1192 clear_thread_flag(TIF_SIGPENDING);
1193
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)
1197 continue;
1198 if (progp->pg_vers[i]->vs_hidden)
1199 continue;
1200 __svc_unregister(net, progp->pg_prog, i, progp->pg_name);
1201 }
1202 }
1203
1204 rcu_read_lock();
1205 sighand = rcu_dereference(current->sighand);
1206 spin_lock_irqsave(&sighand->siglock, flags);
1207 recalc_sigpending();
1208 spin_unlock_irqrestore(&sighand->siglock, flags);
1209 rcu_read_unlock();
1210}
1211
1212/*
1213 * dprintk the given error with the address of the client that caused it.
1214 */
1215#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
1216static __printf(2, 3)
1217void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...)
1218{
1219 struct va_format vaf;
1220 va_list args;
1221 char buf[RPC_MAX_ADDRBUFLEN];
1222
1223 va_start(args, fmt);
1224
1225 vaf.fmt = fmt;
1226 vaf.va = &args;
1227
1228 dprintk("svc: %s: %pV", svc_print_addr(rqstp, buf, sizeof(buf)), &vaf);
1229
1230 va_end(args);
1231}
1232#else
1233static __printf(2,3) void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...) {}
1234#endif
1235
1236__be32
1237svc_generic_init_request(struct svc_rqst *rqstp,
1238 const struct svc_program *progp,
1239 struct svc_process_info *ret)
1240{
1241 const struct svc_version *versp = NULL; /* compiler food */
1242 const struct svc_procedure *procp = NULL;
1243
1244 if (rqstp->rq_vers >= progp->pg_nvers )
1245 goto err_bad_vers;
1246 versp = progp->pg_vers[rqstp->rq_vers];
1247 if (!versp)
1248 goto err_bad_vers;
1249
1250 /*
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.
1255 *
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
1259 * fit.
1260 */
1261 if (versp->vs_need_cong_ctrl && rqstp->rq_xprt &&
1262 !test_bit(XPT_CONG_CTRL, &rqstp->rq_xprt->xpt_flags))
1263 goto err_bad_vers;
1264
1265 if (rqstp->rq_proc >= versp->vs_nproc)
1266 goto err_bad_proc;
1267 rqstp->rq_procinfo = procp = &versp->vs_proc[rqstp->rq_proc];
1268
1269 /* Initialize storage for argp and resp */
1270 memset(rqstp->rq_argp, 0, procp->pc_argzero);
1271 memset(rqstp->rq_resp, 0, procp->pc_ressize);
1272
1273 /* Bump per-procedure stats counter */
1274 this_cpu_inc(versp->vs_count[rqstp->rq_proc]);
1275
1276 ret->dispatch = versp->vs_dispatch;
1277 return rpc_success;
1278err_bad_vers:
1279 ret->mismatch.lovers = progp->pg_lovers;
1280 ret->mismatch.hivers = progp->pg_hivers;
1281 return rpc_prog_mismatch;
1282err_bad_proc:
1283 return rpc_proc_unavail;
1284}
1285EXPORT_SYMBOL_GPL(svc_generic_init_request);
1286
1287/*
1288 * Common routine for processing the RPC request.
1289 */
1290static int
1291svc_process_common(struct svc_rqst *rqstp)
1292{
1293 struct xdr_stream *xdr = &rqstp->rq_res_stream;
1294 struct svc_program *progp;
1295 const struct svc_procedure *procp = NULL;
1296 struct svc_serv *serv = rqstp->rq_server;
1297 struct svc_process_info process;
1298 enum svc_auth_status auth_res;
1299 unsigned int aoffset;
1300 int rc;
1301 __be32 *p;
1302
1303 /* Will be turned off only when NFSv4 Sessions are used */
1304 set_bit(RQ_USEDEFERRAL, &rqstp->rq_flags);
1305 clear_bit(RQ_DROPME, &rqstp->rq_flags);
1306
1307 /* Construct the first words of the reply: */
1308 svcxdr_init_encode(rqstp);
1309 xdr_stream_encode_be32(xdr, rqstp->rq_xid);
1310 xdr_stream_encode_be32(xdr, rpc_reply);
1311
1312 p = xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 4);
1313 if (unlikely(!p))
1314 goto err_short_len;
1315 if (*p++ != cpu_to_be32(RPC_VERSION))
1316 goto err_bad_rpc;
1317
1318 xdr_stream_encode_be32(xdr, rpc_msg_accepted);
1319
1320 rqstp->rq_prog = be32_to_cpup(p++);
1321 rqstp->rq_vers = be32_to_cpup(p++);
1322 rqstp->rq_proc = be32_to_cpup(p);
1323
1324 for (progp = serv->sv_program; progp; progp = progp->pg_next)
1325 if (rqstp->rq_prog == progp->pg_prog)
1326 break;
1327
1328 /*
1329 * Decode auth data, and add verifier to reply buffer.
1330 * We do this before anything else in order to get a decent
1331 * auth verifier.
1332 */
1333 auth_res = svc_authenticate(rqstp);
1334 /* Also give the program a chance to reject this call: */
1335 if (auth_res == SVC_OK && progp)
1336 auth_res = progp->pg_authenticate(rqstp);
1337 trace_svc_authenticate(rqstp, auth_res);
1338 switch (auth_res) {
1339 case SVC_OK:
1340 break;
1341 case SVC_GARBAGE:
1342 goto err_garbage_args;
1343 case SVC_SYSERR:
1344 goto err_system_err;
1345 case SVC_DENIED:
1346 goto err_bad_auth;
1347 case SVC_CLOSE:
1348 goto close;
1349 case SVC_DROP:
1350 goto dropit;
1351 case SVC_COMPLETE:
1352 goto sendit;
1353 default:
1354 pr_warn_once("Unexpected svc_auth_status (%d)\n", auth_res);
1355 goto err_system_err;
1356 }
1357
1358 if (progp == NULL)
1359 goto err_bad_prog;
1360
1361 switch (progp->pg_init_request(rqstp, progp, &process)) {
1362 case rpc_success:
1363 break;
1364 case rpc_prog_unavail:
1365 goto err_bad_prog;
1366 case rpc_prog_mismatch:
1367 goto err_bad_vers;
1368 case rpc_proc_unavail:
1369 goto err_bad_proc;
1370 }
1371
1372 procp = rqstp->rq_procinfo;
1373 /* Should this check go into the dispatcher? */
1374 if (!procp || !procp->pc_func)
1375 goto err_bad_proc;
1376
1377 /* Syntactic check complete */
1378 if (serv->sv_stats)
1379 serv->sv_stats->rpccnt++;
1380 trace_svc_process(rqstp, progp->pg_name);
1381
1382 aoffset = xdr_stream_pos(xdr);
1383
1384 /* un-reserve some of the out-queue now that we have a
1385 * better idea of reply size
1386 */
1387 if (procp->pc_xdrressize)
1388 svc_reserve_auth(rqstp, procp->pc_xdrressize<<2);
1389
1390 /* Call the function that processes the request. */
1391 rc = process.dispatch(rqstp);
1392 if (procp->pc_release)
1393 procp->pc_release(rqstp);
1394 xdr_finish_decode(xdr);
1395
1396 if (!rc)
1397 goto dropit;
1398 if (rqstp->rq_auth_stat != rpc_auth_ok)
1399 goto err_bad_auth;
1400
1401 if (*rqstp->rq_accept_statp != rpc_success)
1402 xdr_truncate_encode(xdr, aoffset);
1403
1404 if (procp->pc_encode == NULL)
1405 goto dropit;
1406
1407 sendit:
1408 if (svc_authorise(rqstp))
1409 goto close_xprt;
1410 return 1; /* Caller can now send it */
1411
1412 dropit:
1413 svc_authorise(rqstp); /* doesn't hurt to call this twice */
1414 dprintk("svc: svc_process dropit\n");
1415 return 0;
1416
1417 close:
1418 svc_authorise(rqstp);
1419close_xprt:
1420 if (rqstp->rq_xprt && test_bit(XPT_TEMP, &rqstp->rq_xprt->xpt_flags))
1421 svc_xprt_close(rqstp->rq_xprt);
1422 dprintk("svc: svc_process close\n");
1423 return 0;
1424
1425err_short_len:
1426 svc_printk(rqstp, "short len %u, dropping request\n",
1427 rqstp->rq_arg.len);
1428 goto close_xprt;
1429
1430err_bad_rpc:
1431 if (serv->sv_stats)
1432 serv->sv_stats->rpcbadfmt++;
1433 xdr_stream_encode_u32(xdr, RPC_MSG_DENIED);
1434 xdr_stream_encode_u32(xdr, RPC_MISMATCH);
1435 /* Only RPCv2 supported */
1436 xdr_stream_encode_u32(xdr, RPC_VERSION);
1437 xdr_stream_encode_u32(xdr, RPC_VERSION);
1438 return 1; /* don't wrap */
1439
1440err_bad_auth:
1441 dprintk("svc: authentication failed (%d)\n",
1442 be32_to_cpu(rqstp->rq_auth_stat));
1443 if (serv->sv_stats)
1444 serv->sv_stats->rpcbadauth++;
1445 /* Restore write pointer to location of reply status: */
1446 xdr_truncate_encode(xdr, XDR_UNIT * 2);
1447 xdr_stream_encode_u32(xdr, RPC_MSG_DENIED);
1448 xdr_stream_encode_u32(xdr, RPC_AUTH_ERROR);
1449 xdr_stream_encode_be32(xdr, rqstp->rq_auth_stat);
1450 goto sendit;
1451
1452err_bad_prog:
1453 dprintk("svc: unknown program %d\n", rqstp->rq_prog);
1454 if (serv->sv_stats)
1455 serv->sv_stats->rpcbadfmt++;
1456 *rqstp->rq_accept_statp = rpc_prog_unavail;
1457 goto sendit;
1458
1459err_bad_vers:
1460 svc_printk(rqstp, "unknown version (%d for prog %d, %s)\n",
1461 rqstp->rq_vers, rqstp->rq_prog, progp->pg_name);
1462
1463 if (serv->sv_stats)
1464 serv->sv_stats->rpcbadfmt++;
1465 *rqstp->rq_accept_statp = rpc_prog_mismatch;
1466
1467 /*
1468 * svc_authenticate() has already added the verifier and
1469 * advanced the stream just past rq_accept_statp.
1470 */
1471 xdr_stream_encode_u32(xdr, process.mismatch.lovers);
1472 xdr_stream_encode_u32(xdr, process.mismatch.hivers);
1473 goto sendit;
1474
1475err_bad_proc:
1476 svc_printk(rqstp, "unknown procedure (%d)\n", rqstp->rq_proc);
1477
1478 if (serv->sv_stats)
1479 serv->sv_stats->rpcbadfmt++;
1480 *rqstp->rq_accept_statp = rpc_proc_unavail;
1481 goto sendit;
1482
1483err_garbage_args:
1484 svc_printk(rqstp, "failed to decode RPC header\n");
1485
1486 if (serv->sv_stats)
1487 serv->sv_stats->rpcbadfmt++;
1488 *rqstp->rq_accept_statp = rpc_garbage_args;
1489 goto sendit;
1490
1491err_system_err:
1492 if (serv->sv_stats)
1493 serv->sv_stats->rpcbadfmt++;
1494 *rqstp->rq_accept_statp = rpc_system_err;
1495 goto sendit;
1496}
1497
1498/**
1499 * svc_process - Execute one RPC transaction
1500 * @rqstp: RPC transaction context
1501 *
1502 */
1503void svc_process(struct svc_rqst *rqstp)
1504{
1505 struct kvec *resv = &rqstp->rq_res.head[0];
1506 __be32 *p;
1507
1508#if IS_ENABLED(CONFIG_FAIL_SUNRPC)
1509 if (!fail_sunrpc.ignore_server_disconnect &&
1510 should_fail(&fail_sunrpc.attr, 1))
1511 svc_xprt_deferred_close(rqstp->rq_xprt);
1512#endif
1513
1514 /*
1515 * Setup response xdr_buf.
1516 * Initially it has just one page
1517 */
1518 rqstp->rq_next_page = &rqstp->rq_respages[1];
1519 resv->iov_base = page_address(rqstp->rq_respages[0]);
1520 resv->iov_len = 0;
1521 rqstp->rq_res.pages = rqstp->rq_next_page;
1522 rqstp->rq_res.len = 0;
1523 rqstp->rq_res.page_base = 0;
1524 rqstp->rq_res.page_len = 0;
1525 rqstp->rq_res.buflen = PAGE_SIZE;
1526 rqstp->rq_res.tail[0].iov_base = NULL;
1527 rqstp->rq_res.tail[0].iov_len = 0;
1528
1529 svcxdr_init_decode(rqstp);
1530 p = xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 2);
1531 if (unlikely(!p))
1532 goto out_drop;
1533 rqstp->rq_xid = *p++;
1534 if (unlikely(*p != rpc_call))
1535 goto out_baddir;
1536
1537 if (!svc_process_common(rqstp))
1538 goto out_drop;
1539 svc_send(rqstp);
1540 return;
1541
1542out_baddir:
1543 svc_printk(rqstp, "bad direction 0x%08x, dropping request\n",
1544 be32_to_cpu(*p));
1545 if (rqstp->rq_server->sv_stats)
1546 rqstp->rq_server->sv_stats->rpcbadfmt++;
1547out_drop:
1548 svc_drop(rqstp);
1549}
1550
1551#if defined(CONFIG_SUNRPC_BACKCHANNEL)
1552/**
1553 * svc_process_bc - process a reverse-direction RPC request
1554 * @req: RPC request to be used for client-side processing
1555 * @rqstp: server-side execution context
1556 *
1557 */
1558void svc_process_bc(struct rpc_rqst *req, struct svc_rqst *rqstp)
1559{
1560 struct rpc_task *task;
1561 int proc_error;
1562 struct rpc_timeout timeout;
1563
1564 /* Build the svc_rqst used by the common processing routine */
1565 rqstp->rq_xid = req->rq_xid;
1566 rqstp->rq_prot = req->rq_xprt->prot;
1567 rqstp->rq_bc_net = req->rq_xprt->xprt_net;
1568
1569 rqstp->rq_addrlen = sizeof(req->rq_xprt->addr);
1570 memcpy(&rqstp->rq_addr, &req->rq_xprt->addr, rqstp->rq_addrlen);
1571 memcpy(&rqstp->rq_arg, &req->rq_rcv_buf, sizeof(rqstp->rq_arg));
1572 memcpy(&rqstp->rq_res, &req->rq_snd_buf, sizeof(rqstp->rq_res));
1573
1574 /* Adjust the argument buffer length */
1575 rqstp->rq_arg.len = req->rq_private_buf.len;
1576 if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len) {
1577 rqstp->rq_arg.head[0].iov_len = rqstp->rq_arg.len;
1578 rqstp->rq_arg.page_len = 0;
1579 } else if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len +
1580 rqstp->rq_arg.page_len)
1581 rqstp->rq_arg.page_len = rqstp->rq_arg.len -
1582 rqstp->rq_arg.head[0].iov_len;
1583 else
1584 rqstp->rq_arg.len = rqstp->rq_arg.head[0].iov_len +
1585 rqstp->rq_arg.page_len;
1586
1587 /* Reset the response buffer */
1588 rqstp->rq_res.head[0].iov_len = 0;
1589
1590 /*
1591 * Skip the XID and calldir fields because they've already
1592 * been processed by the caller.
1593 */
1594 svcxdr_init_decode(rqstp);
1595 if (!xdr_inline_decode(&rqstp->rq_arg_stream, XDR_UNIT * 2))
1596 return;
1597
1598 /* Parse and execute the bc call */
1599 proc_error = svc_process_common(rqstp);
1600
1601 atomic_dec(&req->rq_xprt->bc_slot_count);
1602 if (!proc_error) {
1603 /* Processing error: drop the request */
1604 xprt_free_bc_request(req);
1605 return;
1606 }
1607 /* Finally, send the reply synchronously */
1608 if (rqstp->bc_to_initval > 0) {
1609 timeout.to_initval = rqstp->bc_to_initval;
1610 timeout.to_retries = rqstp->bc_to_retries;
1611 } else {
1612 timeout.to_initval = req->rq_xprt->timeout->to_initval;
1613 timeout.to_retries = req->rq_xprt->timeout->to_retries;
1614 }
1615 memcpy(&req->rq_snd_buf, &rqstp->rq_res, sizeof(req->rq_snd_buf));
1616 task = rpc_run_bc_task(req, &timeout);
1617
1618 if (IS_ERR(task))
1619 return;
1620
1621 WARN_ON_ONCE(atomic_read(&task->tk_count) != 1);
1622 rpc_put_task(task);
1623}
1624#endif /* CONFIG_SUNRPC_BACKCHANNEL */
1625
1626/**
1627 * svc_max_payload - Return transport-specific limit on the RPC payload
1628 * @rqstp: RPC transaction context
1629 *
1630 * Returns the maximum number of payload bytes the current transport
1631 * allows.
1632 */
1633u32 svc_max_payload(const struct svc_rqst *rqstp)
1634{
1635 u32 max = rqstp->rq_xprt->xpt_class->xcl_max_payload;
1636
1637 if (rqstp->rq_server->sv_max_payload < max)
1638 max = rqstp->rq_server->sv_max_payload;
1639 return max;
1640}
1641EXPORT_SYMBOL_GPL(svc_max_payload);
1642
1643/**
1644 * svc_proc_name - Return RPC procedure name in string form
1645 * @rqstp: svc_rqst to operate on
1646 *
1647 * Return value:
1648 * Pointer to a NUL-terminated string
1649 */
1650const char *svc_proc_name(const struct svc_rqst *rqstp)
1651{
1652 if (rqstp && rqstp->rq_procinfo)
1653 return rqstp->rq_procinfo->pc_name;
1654 return "unknown";
1655}
1656
1657
1658/**
1659 * svc_encode_result_payload - mark a range of bytes as a result payload
1660 * @rqstp: svc_rqst to operate on
1661 * @offset: payload's byte offset in rqstp->rq_res
1662 * @length: size of payload, in bytes
1663 *
1664 * Returns zero on success, or a negative errno if a permanent
1665 * error occurred.
1666 */
1667int svc_encode_result_payload(struct svc_rqst *rqstp, unsigned int offset,
1668 unsigned int length)
1669{
1670 return rqstp->rq_xprt->xpt_ops->xpo_result_payload(rqstp, offset,
1671 length);
1672}
1673EXPORT_SYMBOL_GPL(svc_encode_result_payload);
1674
1675/**
1676 * svc_fill_write_vector - Construct data argument for VFS write call
1677 * @rqstp: svc_rqst to operate on
1678 * @payload: xdr_buf containing only the write data payload
1679 *
1680 * Fills in rqstp::rq_vec, and returns the number of elements.
1681 */
1682unsigned int svc_fill_write_vector(struct svc_rqst *rqstp,
1683 struct xdr_buf *payload)
1684{
1685 struct page **pages = payload->pages;
1686 struct kvec *first = payload->head;
1687 struct kvec *vec = rqstp->rq_vec;
1688 size_t total = payload->len;
1689 unsigned int i;
1690
1691 /* Some types of transport can present the write payload
1692 * entirely in rq_arg.pages. In this case, @first is empty.
1693 */
1694 i = 0;
1695 if (first->iov_len) {
1696 vec[i].iov_base = first->iov_base;
1697 vec[i].iov_len = min_t(size_t, total, first->iov_len);
1698 total -= vec[i].iov_len;
1699 ++i;
1700 }
1701
1702 while (total) {
1703 vec[i].iov_base = page_address(*pages);
1704 vec[i].iov_len = min_t(size_t, total, PAGE_SIZE);
1705 total -= vec[i].iov_len;
1706 ++i;
1707 ++pages;
1708 }
1709
1710 WARN_ON_ONCE(i > ARRAY_SIZE(rqstp->rq_vec));
1711 return i;
1712}
1713EXPORT_SYMBOL_GPL(svc_fill_write_vector);
1714
1715/**
1716 * svc_fill_symlink_pathname - Construct pathname argument for VFS symlink call
1717 * @rqstp: svc_rqst to operate on
1718 * @first: buffer containing first section of pathname
1719 * @p: buffer containing remaining section of pathname
1720 * @total: total length of the pathname argument
1721 *
1722 * The VFS symlink API demands a NUL-terminated pathname in mapped memory.
1723 * Returns pointer to a NUL-terminated string, or an ERR_PTR. Caller must free
1724 * the returned string.
1725 */
1726char *svc_fill_symlink_pathname(struct svc_rqst *rqstp, struct kvec *first,
1727 void *p, size_t total)
1728{
1729 size_t len, remaining;
1730 char *result, *dst;
1731
1732 result = kmalloc(total + 1, GFP_KERNEL);
1733 if (!result)
1734 return ERR_PTR(-ESERVERFAULT);
1735
1736 dst = result;
1737 remaining = total;
1738
1739 len = min_t(size_t, total, first->iov_len);
1740 if (len) {
1741 memcpy(dst, first->iov_base, len);
1742 dst += len;
1743 remaining -= len;
1744 }
1745
1746 if (remaining) {
1747 len = min_t(size_t, remaining, PAGE_SIZE);
1748 memcpy(dst, p, len);
1749 dst += len;
1750 }
1751
1752 *dst = '\0';
1753
1754 /* Sanity check: Linux doesn't allow the pathname argument to
1755 * contain a NUL byte.
1756 */
1757 if (strlen(result) != total) {
1758 kfree(result);
1759 return ERR_PTR(-EINVAL);
1760 }
1761 return result;
1762}
1763EXPORT_SYMBOL_GPL(svc_fill_symlink_pathname);